1 //===- GlobalOpt.cpp - Optimize Global Variables --------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass transforms simple global variables that never have their address
11 // taken. If obviously true, it marks read/write globals as constant, deletes
12 // variables only stored to, etc.
14 //===----------------------------------------------------------------------===//
16 #define DEBUG_TYPE "globalopt"
17 #include "llvm/Transforms/IPO.h"
18 #include "llvm/CallingConv.h"
19 #include "llvm/Constants.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/Instructions.h"
22 #include "llvm/IntrinsicInst.h"
23 #include "llvm/Module.h"
24 #include "llvm/Pass.h"
25 #include "llvm/Analysis/ConstantFolding.h"
26 #include "llvm/Analysis/MemoryBuiltins.h"
27 #include "llvm/Target/TargetData.h"
28 #include "llvm/Support/CallSite.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/ErrorHandling.h"
31 #include "llvm/Support/GetElementPtrTypeIterator.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/ADT/DenseMap.h"
35 #include "llvm/ADT/SmallPtrSet.h"
36 #include "llvm/ADT/SmallVector.h"
37 #include "llvm/ADT/Statistic.h"
38 #include "llvm/ADT/STLExtras.h"
42 STATISTIC(NumMarked , "Number of globals marked constant");
43 STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
44 STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
45 STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
46 STATISTIC(NumDeleted , "Number of globals deleted");
47 STATISTIC(NumFnDeleted , "Number of functions deleted");
48 STATISTIC(NumGlobUses , "Number of global uses devirtualized");
49 STATISTIC(NumLocalized , "Number of globals localized");
50 STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans");
51 STATISTIC(NumFastCallFns , "Number of functions converted to fastcc");
52 STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
53 STATISTIC(NumNestRemoved , "Number of nest attributes removed");
54 STATISTIC(NumAliasesResolved, "Number of global aliases resolved");
55 STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated");
58 struct GlobalOpt : public ModulePass {
59 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
61 static char ID; // Pass identification, replacement for typeid
62 GlobalOpt() : ModulePass(&ID) {}
64 bool runOnModule(Module &M);
67 GlobalVariable *FindGlobalCtors(Module &M);
68 bool OptimizeFunctions(Module &M);
69 bool OptimizeGlobalVars(Module &M);
70 bool OptimizeGlobalAliases(Module &M);
71 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
72 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
76 char GlobalOpt::ID = 0;
77 static RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer");
79 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
83 /// GlobalStatus - As we analyze each global, keep track of some information
84 /// about it. If we find out that the address of the global is taken, none of
85 /// this info will be accurate.
87 /// isLoaded - True if the global is ever loaded. If the global isn't ever
88 /// loaded it can be deleted.
91 /// StoredType - Keep track of what stores to the global look like.
94 /// NotStored - There is no store to this global. It can thus be marked
98 /// isInitializerStored - This global is stored to, but the only thing
99 /// stored is the constant it was initialized with. This is only tracked
100 /// for scalar globals.
103 /// isStoredOnce - This global is stored to, but only its initializer and
104 /// one other value is ever stored to it. If this global isStoredOnce, we
105 /// track the value stored to it in StoredOnceValue below. This is only
106 /// tracked for scalar globals.
109 /// isStored - This global is stored to by multiple values or something else
110 /// that we cannot track.
114 /// StoredOnceValue - If only one value (besides the initializer constant) is
115 /// ever stored to this global, keep track of what value it is.
116 Value *StoredOnceValue;
118 /// AccessingFunction/HasMultipleAccessingFunctions - These start out
119 /// null/false. When the first accessing function is noticed, it is recorded.
120 /// When a second different accessing function is noticed,
121 /// HasMultipleAccessingFunctions is set to true.
122 const Function *AccessingFunction;
123 bool HasMultipleAccessingFunctions;
125 /// HasNonInstructionUser - Set to true if this global has a user that is not
126 /// an instruction (e.g. a constant expr or GV initializer).
127 bool HasNonInstructionUser;
129 /// HasPHIUser - Set to true if this global has a user that is a PHI node.
132 GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0),
133 AccessingFunction(0), HasMultipleAccessingFunctions(false),
134 HasNonInstructionUser(false), HasPHIUser(false) {}
139 // SafeToDestroyConstant - It is safe to destroy a constant iff it is only used
140 // by constants itself. Note that constants cannot be cyclic, so this test is
141 // pretty easy to implement recursively.
143 static bool SafeToDestroyConstant(const Constant *C) {
144 if (isa<GlobalValue>(C)) return false;
146 for (Value::const_use_iterator UI = C->use_begin(), E = C->use_end(); UI != E;
148 if (const Constant *CU = dyn_cast<Constant>(*UI)) {
149 if (!SafeToDestroyConstant(CU)) return false;
156 /// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
157 /// structure. If the global has its address taken, return true to indicate we
158 /// can't do anything with it.
160 static bool AnalyzeGlobal(const Value *V, GlobalStatus &GS,
161 SmallPtrSet<const PHINode*, 16> &PHIUsers) {
162 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;
164 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
165 GS.HasNonInstructionUser = true;
167 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
169 } else if (const Instruction *I = dyn_cast<Instruction>(*UI)) {
170 if (!GS.HasMultipleAccessingFunctions) {
171 const 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 (const LoadInst *LI = dyn_cast<LoadInst>(I)) {
179 if (LI->isVolatile()) return true; // Don't hack on volatile loads.
180 } else if (const 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 (const GlobalVariable *GV = dyn_cast<GlobalVariable>(
191 SI->getOperand(1))) {
192 Value *StoredVal = SI->getOperand(0);
193 if (StoredVal == GV->getInitializer()) {
194 if (GS.StoredType < GlobalStatus::isInitializerStored)
195 GS.StoredType = GlobalStatus::isInitializerStored;
196 } else if (isa<LoadInst>(StoredVal) &&
197 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 (const 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 (const 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 ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
252 unsigned IdxV = CI->getZExtValue();
254 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
255 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
256 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
257 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
258 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) {
259 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
260 } else if (isa<ConstantAggregateZero>(Agg)) {
261 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
262 if (IdxV < STy->getNumElements())
263 return Constant::getNullValue(STy->getElementType(IdxV));
264 } else if (const SequentialType *STy =
265 dyn_cast<SequentialType>(Agg->getType())) {
266 return Constant::getNullValue(STy->getElementType());
268 } else if (isa<UndefValue>(Agg)) {
269 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
270 if (IdxV < STy->getNumElements())
271 return UndefValue::get(STy->getElementType(IdxV));
272 } else if (const SequentialType *STy =
273 dyn_cast<SequentialType>(Agg->getType())) {
274 return UndefValue::get(STy->getElementType());
281 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
282 /// users of the global, cleaning up the obvious ones. This is largely just a
283 /// quick scan over the use list to clean up the easy and obvious cruft. This
284 /// returns true if it made a change.
285 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
286 bool Changed = false;
287 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
290 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
292 // Replace the load with the initializer.
293 LI->replaceAllUsesWith(Init);
294 LI->eraseFromParent();
297 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
298 // Store must be unreachable or storing Init into the global.
299 SI->eraseFromParent();
301 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
302 if (CE->getOpcode() == Instruction::GetElementPtr) {
303 Constant *SubInit = 0;
305 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
306 Changed |= CleanupConstantGlobalUsers(CE, SubInit);
307 } else if (CE->getOpcode() == Instruction::BitCast &&
308 CE->getType()->isPointerTy()) {
309 // Pointer cast, delete any stores and memsets to the global.
310 Changed |= CleanupConstantGlobalUsers(CE, 0);
313 if (CE->use_empty()) {
314 CE->destroyConstant();
317 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
318 // Do not transform "gepinst (gep constexpr (GV))" here, because forming
319 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
320 // and will invalidate our notion of what Init is.
321 Constant *SubInit = 0;
322 if (!isa<ConstantExpr>(GEP->getOperand(0))) {
324 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
325 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
326 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
328 Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
330 if (GEP->use_empty()) {
331 GEP->eraseFromParent();
334 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
335 if (MI->getRawDest() == V) {
336 MI->eraseFromParent();
340 } else if (Constant *C = dyn_cast<Constant>(U)) {
341 // If we have a chain of dead constantexprs or other things dangling from
342 // us, and if they are all dead, nuke them without remorse.
343 if (SafeToDestroyConstant(C)) {
344 C->destroyConstant();
345 // This could have invalidated UI, start over from scratch.
346 CleanupConstantGlobalUsers(V, Init);
354 /// isSafeSROAElementUse - Return true if the specified instruction is a safe
355 /// user of a derived expression from a global that we want to SROA.
356 static bool isSafeSROAElementUse(Value *V) {
357 // We might have a dead and dangling constant hanging off of here.
358 if (Constant *C = dyn_cast<Constant>(V))
359 return SafeToDestroyConstant(C);
361 Instruction *I = dyn_cast<Instruction>(V);
362 if (!I) return false;
365 if (isa<LoadInst>(I)) return true;
367 // Stores *to* the pointer are ok.
368 if (StoreInst *SI = dyn_cast<StoreInst>(I))
369 return SI->getOperand(0) != V;
371 // Otherwise, it must be a GEP.
372 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I);
373 if (GEPI == 0) return false;
375 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
376 !cast<Constant>(GEPI->getOperand(1))->isNullValue())
379 for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end();
381 if (!isSafeSROAElementUse(*I))
387 /// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value.
388 /// Look at it and its uses and decide whether it is safe to SROA this global.
390 static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
391 // The user of the global must be a GEP Inst or a ConstantExpr GEP.
392 if (!isa<GetElementPtrInst>(U) &&
393 (!isa<ConstantExpr>(U) ||
394 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
397 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
398 // don't like < 3 operand CE's, and we don't like non-constant integer
399 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some
401 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
402 !cast<Constant>(U->getOperand(1))->isNullValue() ||
403 !isa<ConstantInt>(U->getOperand(2)))
406 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
407 ++GEPI; // Skip over the pointer index.
409 // If this is a use of an array allocation, do a bit more checking for sanity.
410 if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
411 uint64_t NumElements = AT->getNumElements();
412 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
414 // Check to make sure that index falls within the array. If not,
415 // something funny is going on, so we won't do the optimization.
417 if (Idx->getZExtValue() >= NumElements)
420 // We cannot scalar repl this level of the array unless any array
421 // sub-indices are in-range constants. In particular, consider:
422 // A[0][i]. We cannot know that the user isn't doing invalid things like
423 // allowing i to index an out-of-range subscript that accesses A[1].
425 // Scalar replacing *just* the outer index of the array is probably not
426 // going to be a win anyway, so just give up.
427 for (++GEPI; // Skip array index.
430 uint64_t NumElements;
431 if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
432 NumElements = SubArrayTy->getNumElements();
433 else if (const VectorType *SubVectorTy = dyn_cast<VectorType>(*GEPI))
434 NumElements = SubVectorTy->getNumElements();
436 assert((*GEPI)->isStructTy() &&
437 "Indexed GEP type is not array, vector, or struct!");
441 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
442 if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
447 for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I)
448 if (!isSafeSROAElementUse(*I))
453 /// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
454 /// is safe for us to perform this transformation.
456 static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
457 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
459 if (!IsUserOfGlobalSafeForSRA(*UI, GV))
466 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
467 /// variable. This opens the door for other optimizations by exposing the
468 /// behavior of the program in a more fine-grained way. We have determined that
469 /// this transformation is safe already. We return the first global variable we
470 /// insert so that the caller can reprocess it.
471 static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) {
472 // Make sure this global only has simple uses that we can SRA.
473 if (!GlobalUsersSafeToSRA(GV))
476 assert(GV->hasLocalLinkage() && !GV->isConstant());
477 Constant *Init = GV->getInitializer();
478 const Type *Ty = Init->getType();
480 std::vector<GlobalVariable*> NewGlobals;
481 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
483 // Get the alignment of the global, either explicit or target-specific.
484 unsigned StartAlignment = GV->getAlignment();
485 if (StartAlignment == 0)
486 StartAlignment = TD.getABITypeAlignment(GV->getType());
488 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
489 NewGlobals.reserve(STy->getNumElements());
490 const StructLayout &Layout = *TD.getStructLayout(STy);
491 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
492 Constant *In = getAggregateConstantElement(Init,
493 ConstantInt::get(Type::getInt32Ty(STy->getContext()), i));
494 assert(In && "Couldn't get element of initializer?");
495 GlobalVariable *NGV = new GlobalVariable(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::getInt32Ty(Init->getContext()), i));
527 assert(In && "Couldn't get element of initializer?");
529 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
530 GlobalVariable::InternalLinkage,
531 In, GV->getName()+"."+Twine(i),
533 GV->getType()->getAddressSpace());
534 Globals.insert(GV, NGV);
535 NewGlobals.push_back(NGV);
537 // Calculate the known alignment of the field. If the original aggregate
538 // had 256 byte alignment for example, something might depend on that:
539 // propagate info to each field.
540 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
541 if (NewAlign > EltAlign)
542 NGV->setAlignment(NewAlign);
546 if (NewGlobals.empty())
549 DEBUG(dbgs() << "PERFORMING GLOBAL SRA ON: " << *GV);
551 Constant *NullInt =Constant::getNullValue(Type::getInt32Ty(GV->getContext()));
553 // Loop over all of the uses of the global, replacing the constantexpr geps,
554 // with smaller constantexpr geps or direct references.
555 while (!GV->use_empty()) {
556 User *GEP = GV->use_back();
557 assert(((isa<ConstantExpr>(GEP) &&
558 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
559 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
561 // Ignore the 1th operand, which has to be zero or else the program is quite
562 // broken (undefined). Get the 2nd operand, which is the structure or array
564 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
565 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
567 Value *NewPtr = NewGlobals[Val];
569 // Form a shorter GEP if needed.
570 if (GEP->getNumOperands() > 3) {
571 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
572 SmallVector<Constant*, 8> Idxs;
573 Idxs.push_back(NullInt);
574 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
575 Idxs.push_back(CE->getOperand(i));
576 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr),
577 &Idxs[0], Idxs.size());
579 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
580 SmallVector<Value*, 8> Idxs;
581 Idxs.push_back(NullInt);
582 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
583 Idxs.push_back(GEPI->getOperand(i));
584 NewPtr = GetElementPtrInst::Create(NewPtr, Idxs.begin(), Idxs.end(),
585 GEPI->getName()+"."+Twine(Val),GEPI);
588 GEP->replaceAllUsesWith(NewPtr);
590 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
591 GEPI->eraseFromParent();
593 cast<ConstantExpr>(GEP)->destroyConstant();
596 // Delete the old global, now that it is dead.
600 // Loop over the new globals array deleting any globals that are obviously
601 // dead. This can arise due to scalarization of a structure or an array that
602 // has elements that are dead.
603 unsigned FirstGlobal = 0;
604 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
605 if (NewGlobals[i]->use_empty()) {
606 Globals.erase(NewGlobals[i]);
607 if (FirstGlobal == i) ++FirstGlobal;
610 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
613 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
614 /// value will trap if the value is dynamically null. PHIs keeps track of any
615 /// phi nodes we've seen to avoid reprocessing them.
616 static bool AllUsesOfValueWillTrapIfNull(const Value *V,
617 SmallPtrSet<const PHINode*, 8> &PHIs) {
618 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;
622 if (isa<LoadInst>(U)) {
624 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
625 if (SI->getOperand(0) == V) {
626 //cerr << "NONTRAPPING USE: " << *U;
627 return false; // Storing the value.
629 } else if (const CallInst *CI = dyn_cast<CallInst>(U)) {
630 if (CI->getCalledValue() != V) {
631 //cerr << "NONTRAPPING USE: " << *U;
632 return false; // Not calling the ptr
634 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(U)) {
635 if (II->getCalledValue() != V) {
636 //cerr << "NONTRAPPING USE: " << *U;
637 return false; // Not calling the ptr
639 } else if (const BitCastInst *CI = dyn_cast<BitCastInst>(U)) {
640 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
641 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
642 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
643 } else if (const PHINode *PN = dyn_cast<PHINode>(U)) {
644 // If we've already seen this phi node, ignore it, it has already been
646 if (PHIs.insert(PN) && !AllUsesOfValueWillTrapIfNull(PN, PHIs))
648 } else if (isa<ICmpInst>(U) &&
649 isa<ConstantPointerNull>(UI->getOperand(1))) {
650 // Ignore icmp X, null
652 //cerr << "NONTRAPPING USE: " << *U;
659 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
660 /// from GV will trap if the loaded value is null. Note that this also permits
661 /// comparisons of the loaded value against null, as a special case.
662 static bool AllUsesOfLoadedValueWillTrapIfNull(const GlobalVariable *GV) {
663 for (Value::const_use_iterator UI = GV->use_begin(), E = GV->use_end();
667 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
668 SmallPtrSet<const PHINode*, 8> PHIs;
669 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
671 } else if (isa<StoreInst>(U)) {
672 // Ignore stores to the global.
674 // We don't know or understand this user, bail out.
675 //cerr << "UNKNOWN USER OF GLOBAL!: " << *U;
682 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
683 bool Changed = false;
684 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
685 Instruction *I = cast<Instruction>(*UI++);
686 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
687 LI->setOperand(0, NewV);
689 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
690 if (SI->getOperand(1) == V) {
691 SI->setOperand(1, NewV);
694 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
696 if (CS.getCalledValue() == 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 CS.setCalledFunction(NewV);
701 bool PassedAsArg = false;
702 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
703 if (CS.getArgument(i) == V) {
705 CS.setArgument(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()));
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],
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 bool Changed = false;
753 // Keep track of whether we are able to remove all the uses of the global
754 // other than the store that defines it.
755 bool AllNonStoreUsesGone = true;
757 // Replace all uses of loads with uses of uses of the stored value.
758 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end(); GUI != E;){
759 User *GlobalUser = *GUI++;
760 if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) {
761 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
762 // If we were able to delete all uses of the loads
763 if (LI->use_empty()) {
764 LI->eraseFromParent();
767 AllNonStoreUsesGone = false;
769 } else if (isa<StoreInst>(GlobalUser)) {
770 // Ignore the store that stores "LV" to the global.
771 assert(GlobalUser->getOperand(1) == GV &&
772 "Must be storing *to* the global");
774 AllNonStoreUsesGone = false;
776 // If we get here we could have other crazy uses that are transitively
778 assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) ||
779 isa<ConstantExpr>(GlobalUser)) && "Only expect load and stores!");
784 DEBUG(dbgs() << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV);
788 // If we nuked all of the loads, then none of the stores are needed either,
789 // nor is the global.
790 if (AllNonStoreUsesGone) {
791 DEBUG(dbgs() << " *** GLOBAL NOW DEAD!\n");
792 CleanupConstantGlobalUsers(GV, 0);
793 if (GV->use_empty()) {
794 GV->eraseFromParent();
802 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
803 /// instructions that are foldable.
804 static void ConstantPropUsersOf(Value *V) {
805 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
806 if (Instruction *I = dyn_cast<Instruction>(*UI++))
807 if (Constant *NewC = ConstantFoldInstruction(I)) {
808 I->replaceAllUsesWith(NewC);
810 // Advance UI to the next non-I use to avoid invalidating it!
811 // Instructions could multiply use V.
812 while (UI != E && *UI == I)
814 I->eraseFromParent();
818 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
819 /// variable, and transforms the program as if it always contained the result of
820 /// the specified malloc. Because it is always the result of the specified
821 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
822 /// malloc into a global, and any loads of GV as uses of the new global.
823 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
826 ConstantInt *NElements,
828 DEBUG(errs() << "PROMOTING GLOBAL: " << *GV << " CALL = " << *CI << '\n');
830 const Type *GlobalType;
831 if (NElements->getZExtValue() == 1)
832 GlobalType = AllocTy;
834 // If we have an array allocation, the global variable is of an array.
835 GlobalType = ArrayType::get(AllocTy, NElements->getZExtValue());
837 // Create the new global variable. The contents of the malloc'd memory is
838 // undefined, so initialize with an undef value.
839 GlobalVariable *NewGV = new GlobalVariable(*GV->getParent(),
841 GlobalValue::InternalLinkage,
842 UndefValue::get(GlobalType),
843 GV->getName()+".body",
845 GV->isThreadLocal());
847 // If there are bitcast users of the malloc (which is typical, usually we have
848 // a malloc + bitcast) then replace them with uses of the new global. Update
849 // other users to use the global as well.
850 BitCastInst *TheBC = 0;
851 while (!CI->use_empty()) {
852 Instruction *User = cast<Instruction>(CI->use_back());
853 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
854 if (BCI->getType() == NewGV->getType()) {
855 BCI->replaceAllUsesWith(NewGV);
856 BCI->eraseFromParent();
858 BCI->setOperand(0, NewGV);
862 TheBC = new BitCastInst(NewGV, CI->getType(), "newgv", CI);
863 User->replaceUsesOfWith(CI, TheBC);
867 Constant *RepValue = NewGV;
868 if (NewGV->getType() != GV->getType()->getElementType())
869 RepValue = ConstantExpr::getBitCast(RepValue,
870 GV->getType()->getElementType());
872 // If there is a comparison against null, we will insert a global bool to
873 // keep track of whether the global was initialized yet or not.
874 GlobalVariable *InitBool =
875 new GlobalVariable(Type::getInt1Ty(GV->getContext()), false,
876 GlobalValue::InternalLinkage,
877 ConstantInt::getFalse(GV->getContext()),
878 GV->getName()+".init", GV->isThreadLocal());
879 bool InitBoolUsed = false;
881 // Loop over all uses of GV, processing them in turn.
882 while (!GV->use_empty()) {
883 if (StoreInst *SI = dyn_cast<StoreInst>(GV->use_back())) {
884 // The global is initialized when the store to it occurs.
885 new StoreInst(ConstantInt::getTrue(GV->getContext()), InitBool, SI);
886 SI->eraseFromParent();
890 LoadInst *LI = cast<LoadInst>(GV->use_back());
891 while (!LI->use_empty()) {
892 Use &LoadUse = LI->use_begin().getUse();
893 if (!isa<ICmpInst>(LoadUse.getUser())) {
898 ICmpInst *ICI = cast<ICmpInst>(LoadUse.getUser());
899 // Replace the cmp X, 0 with a use of the bool value.
900 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", ICI);
902 switch (ICI->getPredicate()) {
903 default: llvm_unreachable("Unknown ICmp Predicate!");
904 case ICmpInst::ICMP_ULT:
905 case ICmpInst::ICMP_SLT: // X < null -> always false
906 LV = ConstantInt::getFalse(GV->getContext());
908 case ICmpInst::ICMP_ULE:
909 case ICmpInst::ICMP_SLE:
910 case ICmpInst::ICMP_EQ:
911 LV = BinaryOperator::CreateNot(LV, "notinit", ICI);
913 case ICmpInst::ICMP_NE:
914 case ICmpInst::ICMP_UGE:
915 case ICmpInst::ICMP_SGE:
916 case ICmpInst::ICMP_UGT:
917 case ICmpInst::ICMP_SGT:
920 ICI->replaceAllUsesWith(LV);
921 ICI->eraseFromParent();
923 LI->eraseFromParent();
926 // If the initialization boolean was used, insert it, otherwise delete it.
928 while (!InitBool->use_empty()) // Delete initializations
929 cast<StoreInst>(InitBool->use_back())->eraseFromParent();
932 GV->getParent()->getGlobalList().insert(GV, InitBool);
934 // Now the GV is dead, nuke it and the malloc..
935 GV->eraseFromParent();
936 CI->eraseFromParent();
938 // To further other optimizations, loop over all users of NewGV and try to
939 // constant prop them. This will promote GEP instructions with constant
940 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
941 ConstantPropUsersOf(NewGV);
942 if (RepValue != NewGV)
943 ConstantPropUsersOf(RepValue);
948 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
949 /// to make sure that there are no complex uses of V. We permit simple things
950 /// like dereferencing the pointer, but not storing through the address, unless
951 /// it is to the specified global.
952 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(const Instruction *V,
953 const GlobalVariable *GV,
954 SmallPtrSet<const PHINode*, 8> &PHIs) {
955 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end();
957 const Instruction *Inst = cast<Instruction>(*UI);
959 if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) {
960 continue; // Fine, ignore.
963 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
964 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
965 return false; // Storing the pointer itself... bad.
966 continue; // Otherwise, storing through it, or storing into GV... fine.
969 // Must index into the array and into the struct.
970 if (isa<GetElementPtrInst>(Inst) && Inst->getNumOperands() >= 3) {
971 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs))
976 if (const 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 (const 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(const Value *V,
1045 SmallPtrSet<const PHINode*, 32> &LoadUsingPHIs,
1046 SmallPtrSet<const 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::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;
1051 const Instruction *User = cast<Instruction>(*UI);
1053 // Comparison against null is ok.
1054 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(User)) {
1055 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1060 // getelementptr is also ok, but only a simple form.
1061 if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1062 // Must index into the array and into the struct.
1063 if (GEPI->getNumOperands() < 3)
1066 // Otherwise the GEP is ok.
1070 if (const PHINode *PN = dyn_cast<PHINode>(User)) {
1071 if (!LoadUsingPHIsPerLoad.insert(PN))
1072 // This means some phi nodes are dependent on each other.
1073 // Avoid infinite looping!
1075 if (!LoadUsingPHIs.insert(PN))
1076 // If we have already analyzed this PHI, then it is safe.
1079 // Make sure all uses of the PHI are simple enough to transform.
1080 if (!LoadUsesSimpleEnoughForHeapSRA(PN,
1081 LoadUsingPHIs, LoadUsingPHIsPerLoad))
1087 // Otherwise we don't know what this is, not ok.
1095 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
1096 /// GV are simple enough to perform HeapSRA, return true.
1097 static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(const GlobalVariable *GV,
1098 Instruction *StoredVal) {
1099 SmallPtrSet<const PHINode*, 32> LoadUsingPHIs;
1100 SmallPtrSet<const PHINode*, 32> LoadUsingPHIsPerLoad;
1101 for (Value::const_use_iterator UI = GV->use_begin(), E = GV->use_end();
1103 if (const LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
1104 if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs,
1105 LoadUsingPHIsPerLoad))
1107 LoadUsingPHIsPerLoad.clear();
1110 // If we reach here, we know that all uses of the loads and transitive uses
1111 // (through PHI nodes) are simple enough to transform. However, we don't know
1112 // that all inputs the to the PHI nodes are in the same equivalence sets.
1113 // Check to verify that all operands of the PHIs are either PHIS that can be
1114 // transformed, loads from GV, or MI itself.
1115 for (SmallPtrSet<const PHINode*, 32>::const_iterator I = LoadUsingPHIs.begin()
1116 , E = LoadUsingPHIs.end(); I != E; ++I) {
1117 const PHINode *PN = *I;
1118 for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) {
1119 Value *InVal = PN->getIncomingValue(op);
1121 // PHI of the stored value itself is ok.
1122 if (InVal == StoredVal) continue;
1124 if (const PHINode *InPN = dyn_cast<PHINode>(InVal)) {
1125 // One of the PHIs in our set is (optimistically) ok.
1126 if (LoadUsingPHIs.count(InPN))
1131 // Load from GV is ok.
1132 if (const LoadInst *LI = dyn_cast<LoadInst>(InVal))
1133 if (LI->getOperand(0) == GV)
1138 // Anything else is rejected.
1146 static Value *GetHeapSROAValue(Value *V, unsigned FieldNo,
1147 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1148 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
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,
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 // If this is a comparison against null, handle it.
1192 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1193 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1194 // If we have a setcc of the loaded pointer, we can use a setcc of any
1196 Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0,
1197 InsertedScalarizedValues, PHIsToRewrite);
1199 Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr,
1200 Constant::getNullValue(NPtr->getType()),
1202 SCI->replaceAllUsesWith(New);
1203 SCI->eraseFromParent();
1207 // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...'
1208 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1209 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1210 && "Unexpected GEPI!");
1212 // Load the pointer for this field.
1213 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1214 Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo,
1215 InsertedScalarizedValues, PHIsToRewrite);
1217 // Create the new GEP idx vector.
1218 SmallVector<Value*, 8> GEPIdx;
1219 GEPIdx.push_back(GEPI->getOperand(1));
1220 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1222 Value *NGEPI = GetElementPtrInst::Create(NewPtr,
1223 GEPIdx.begin(), GEPIdx.end(),
1224 GEPI->getName(), GEPI);
1225 GEPI->replaceAllUsesWith(NGEPI);
1226 GEPI->eraseFromParent();
1230 // Recursively transform the users of PHI nodes. This will lazily create the
1231 // PHIs that are needed for individual elements. Keep track of what PHIs we
1232 // see in InsertedScalarizedValues so that we don't get infinite loops (very
1233 // antisocial). If the PHI is already in InsertedScalarizedValues, it has
1234 // already been seen first by another load, so its uses have already been
1236 PHINode *PN = cast<PHINode>(LoadUser);
1238 DenseMap<Value*, std::vector<Value*> >::iterator InsertPos;
1239 tie(InsertPos, Inserted) =
1240 InsertedScalarizedValues.insert(std::make_pair(PN, std::vector<Value*>()));
1241 if (!Inserted) return;
1243 // If this is the first time we've seen this PHI, recursively process all
1245 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); UI != E; ) {
1246 Instruction *User = cast<Instruction>(*UI++);
1247 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1251 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1252 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1253 /// use FieldGlobals instead. All uses of loaded values satisfy
1254 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA.
1255 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1256 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1257 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1258 for (Value::use_iterator UI = Load->use_begin(), E = Load->use_end();
1260 Instruction *User = cast<Instruction>(*UI++);
1261 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1264 if (Load->use_empty()) {
1265 Load->eraseFromParent();
1266 InsertedScalarizedValues.erase(Load);
1270 /// PerformHeapAllocSRoA - CI is an allocation of an array of structures. Break
1271 /// it up into multiple allocations of arrays of the fields.
1272 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI,
1273 Value* NElems, TargetData *TD) {
1274 DEBUG(dbgs() << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *CI << '\n');
1275 const Type* MAT = getMallocAllocatedType(CI);
1276 const StructType *STy = cast<StructType>(MAT);
1278 // There is guaranteed to be at least one use of the malloc (storing
1279 // it into GV). If there are other uses, change them to be uses of
1280 // the global to simplify later code. This also deletes the store
1282 ReplaceUsesOfMallocWithGlobal(CI, GV);
1284 // Okay, at this point, there are no users of the malloc. Insert N
1285 // new mallocs at the same place as CI, and N globals.
1286 std::vector<Value*> FieldGlobals;
1287 std::vector<Value*> FieldMallocs;
1289 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1290 const Type *FieldTy = STy->getElementType(FieldNo);
1291 const PointerType *PFieldTy = PointerType::getUnqual(FieldTy);
1293 GlobalVariable *NGV =
1294 new GlobalVariable(*GV->getParent(),
1295 PFieldTy, false, GlobalValue::InternalLinkage,
1296 Constant::getNullValue(PFieldTy),
1297 GV->getName() + ".f" + Twine(FieldNo), GV,
1298 GV->isThreadLocal());
1299 FieldGlobals.push_back(NGV);
1301 unsigned TypeSize = TD->getTypeAllocSize(FieldTy);
1302 if (const StructType *ST = dyn_cast<StructType>(FieldTy))
1303 TypeSize = TD->getStructLayout(ST)->getSizeInBytes();
1304 const Type *IntPtrTy = TD->getIntPtrType(CI->getContext());
1305 Value *NMI = CallInst::CreateMalloc(CI, IntPtrTy, FieldTy,
1306 ConstantInt::get(IntPtrTy, TypeSize),
1308 CI->getName() + ".f" + Twine(FieldNo));
1309 FieldMallocs.push_back(NMI);
1310 new StoreInst(NMI, NGV, CI);
1313 // The tricky aspect of this transformation is handling the case when malloc
1314 // fails. In the original code, malloc failing would set the result pointer
1315 // of malloc to null. In this case, some mallocs could succeed and others
1316 // could fail. As such, we emit code that looks like this:
1317 // F0 = malloc(field0)
1318 // F1 = malloc(field1)
1319 // F2 = malloc(field2)
1320 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1321 // if (F0) { free(F0); F0 = 0; }
1322 // if (F1) { free(F1); F1 = 0; }
1323 // if (F2) { free(F2); F2 = 0; }
1325 // The malloc can also fail if its argument is too large.
1326 Constant *ConstantZero = ConstantInt::get(CI->getOperand(1)->getType(), 0);
1327 Value *RunningOr = new ICmpInst(CI, ICmpInst::ICMP_SLT, CI->getOperand(1),
1328 ConstantZero, "isneg");
1329 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1330 Value *Cond = new ICmpInst(CI, ICmpInst::ICMP_EQ, FieldMallocs[i],
1331 Constant::getNullValue(FieldMallocs[i]->getType()),
1333 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", CI);
1336 // Split the basic block at the old malloc.
1337 BasicBlock *OrigBB = CI->getParent();
1338 BasicBlock *ContBB = OrigBB->splitBasicBlock(CI, "malloc_cont");
1340 // Create the block to check the first condition. Put all these blocks at the
1341 // end of the function as they are unlikely to be executed.
1342 BasicBlock *NullPtrBlock = BasicBlock::Create(OrigBB->getContext(),
1344 OrigBB->getParent());
1346 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1347 // branch on RunningOr.
1348 OrigBB->getTerminator()->eraseFromParent();
1349 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1351 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1352 // pointer, because some may be null while others are not.
1353 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1354 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1355 Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
1356 Constant::getNullValue(GVVal->getType()),
1358 BasicBlock *FreeBlock = BasicBlock::Create(Cmp->getContext(), "free_it",
1359 OrigBB->getParent());
1360 BasicBlock *NextBlock = BasicBlock::Create(Cmp->getContext(), "next",
1361 OrigBB->getParent());
1362 Instruction *BI = BranchInst::Create(FreeBlock, NextBlock,
1365 // Fill in FreeBlock.
1366 CallInst::CreateFree(GVVal, BI);
1367 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1369 BranchInst::Create(NextBlock, FreeBlock);
1371 NullPtrBlock = NextBlock;
1374 BranchInst::Create(ContBB, NullPtrBlock);
1376 // CI is no longer needed, remove it.
1377 CI->eraseFromParent();
1379 /// InsertedScalarizedLoads - As we process loads, if we can't immediately
1380 /// update all uses of the load, keep track of what scalarized loads are
1381 /// inserted for a given load.
1382 DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues;
1383 InsertedScalarizedValues[GV] = FieldGlobals;
1385 std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite;
1387 // Okay, the malloc site is completely handled. All of the uses of GV are now
1388 // loads, and all uses of those loads are simple. Rewrite them to use loads
1389 // of the per-field globals instead.
1390 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;) {
1391 Instruction *User = cast<Instruction>(*UI++);
1393 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1394 RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite);
1398 // Must be a store of null.
1399 StoreInst *SI = cast<StoreInst>(User);
1400 assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
1401 "Unexpected heap-sra user!");
1403 // Insert a store of null into each global.
1404 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1405 const PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType());
1406 Constant *Null = Constant::getNullValue(PT->getElementType());
1407 new StoreInst(Null, FieldGlobals[i], SI);
1409 // Erase the original store.
1410 SI->eraseFromParent();
1413 // While we have PHIs that are interesting to rewrite, do it.
1414 while (!PHIsToRewrite.empty()) {
1415 PHINode *PN = PHIsToRewrite.back().first;
1416 unsigned FieldNo = PHIsToRewrite.back().second;
1417 PHIsToRewrite.pop_back();
1418 PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]);
1419 assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi");
1421 // Add all the incoming values. This can materialize more phis.
1422 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1423 Value *InVal = PN->getIncomingValue(i);
1424 InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues,
1426 FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
1430 // Drop all inter-phi links and any loads that made it this far.
1431 for (DenseMap<Value*, std::vector<Value*> >::iterator
1432 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1434 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1435 PN->dropAllReferences();
1436 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1437 LI->dropAllReferences();
1440 // Delete all the phis and loads now that inter-references are dead.
1441 for (DenseMap<Value*, std::vector<Value*> >::iterator
1442 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1444 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1445 PN->eraseFromParent();
1446 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1447 LI->eraseFromParent();
1450 // The old global is now dead, remove it.
1451 GV->eraseFromParent();
1454 return cast<GlobalVariable>(FieldGlobals[0]);
1457 /// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1458 /// pointer global variable with a single value stored it that is a malloc or
1460 static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
1462 const Type *AllocTy,
1463 Module::global_iterator &GVI,
1468 // If this is a malloc of an abstract type, don't touch it.
1469 if (!AllocTy->isSized())
1472 // We can't optimize this global unless all uses of it are *known* to be
1473 // of the malloc value, not of the null initializer value (consider a use
1474 // that compares the global's value against zero to see if the malloc has
1475 // been reached). To do this, we check to see if all uses of the global
1476 // would trap if the global were null: this proves that they must all
1477 // happen after the malloc.
1478 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1481 // We can't optimize this if the malloc itself is used in a complex way,
1482 // for example, being stored into multiple globals. This allows the
1483 // malloc to be stored into the specified global, loaded setcc'd, and
1484 // GEP'd. These are all things we could transform to using the global
1486 SmallPtrSet<const PHINode*, 8> PHIs;
1487 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV, PHIs))
1490 // If we have a global that is only initialized with a fixed size malloc,
1491 // transform the program to use global memory instead of malloc'd memory.
1492 // This eliminates dynamic allocation, avoids an indirection accessing the
1493 // data, and exposes the resultant global to further GlobalOpt.
1494 // We cannot optimize the malloc if we cannot determine malloc array size.
1495 Value *NElems = getMallocArraySize(CI, TD, true);
1499 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
1500 // Restrict this transformation to only working on small allocations
1501 // (2048 bytes currently), as we don't want to introduce a 16M global or
1503 if (NElements->getZExtValue() * TD->getTypeAllocSize(AllocTy) < 2048) {
1504 GVI = OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, TD);
1508 // If the allocation is an array of structures, consider transforming this
1509 // into multiple malloc'd arrays, one for each field. This is basically
1510 // SRoA for malloc'd memory.
1512 // If this is an allocation of a fixed size array of structs, analyze as a
1513 // variable size array. malloc [100 x struct],1 -> malloc struct, 100
1514 if (NElems == ConstantInt::get(CI->getOperand(1)->getType(), 1))
1515 if (const ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
1516 AllocTy = AT->getElementType();
1518 const StructType *AllocSTy = dyn_cast<StructType>(AllocTy);
1522 // This the structure has an unreasonable number of fields, leave it
1524 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
1525 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, CI)) {
1527 // If this is a fixed size array, transform the Malloc to be an alloc of
1528 // structs. malloc [100 x struct],1 -> malloc struct, 100
1529 if (const ArrayType *AT = dyn_cast<ArrayType>(getMallocAllocatedType(CI))) {
1530 const Type *IntPtrTy = TD->getIntPtrType(CI->getContext());
1531 unsigned TypeSize = TD->getStructLayout(AllocSTy)->getSizeInBytes();
1532 Value *AllocSize = ConstantInt::get(IntPtrTy, TypeSize);
1533 Value *NumElements = ConstantInt::get(IntPtrTy, AT->getNumElements());
1534 Instruction *Malloc = CallInst::CreateMalloc(CI, IntPtrTy, AllocSTy,
1535 AllocSize, NumElements,
1537 Instruction *Cast = new BitCastInst(Malloc, CI->getType(), "tmp", CI);
1538 CI->replaceAllUsesWith(Cast);
1539 CI->eraseFromParent();
1540 CI = dyn_cast<BitCastInst>(Malloc) ?
1541 extractMallocCallFromBitCast(Malloc) : cast<CallInst>(Malloc);
1544 GVI = PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, TD, true),TD);
1551 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1552 // that only one value (besides its initializer) is ever stored to the global.
1553 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1554 Module::global_iterator &GVI,
1556 // Ignore no-op GEPs and bitcasts.
1557 StoredOnceVal = StoredOnceVal->stripPointerCasts();
1559 // If we are dealing with a pointer global that is initialized to null and
1560 // only has one (non-null) value stored into it, then we can optimize any
1561 // users of the loaded value (often calls and loads) that would trap if the
1563 if (GV->getInitializer()->getType()->isPointerTy() &&
1564 GV->getInitializer()->isNullValue()) {
1565 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1566 if (GV->getInitializer()->getType() != SOVC->getType())
1568 ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1570 // Optimize away any trapping uses of the loaded value.
1571 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1573 } else if (CallInst *CI = extractMallocCall(StoredOnceVal)) {
1574 const Type* MallocType = getMallocAllocatedType(CI);
1575 if (MallocType && TryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType,
1584 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1585 /// two values ever stored into GV are its initializer and OtherVal. See if we
1586 /// can shrink the global into a boolean and select between the two values
1587 /// whenever it is used. This exposes the values to other scalar optimizations.
1588 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1589 const Type *GVElType = GV->getType()->getElementType();
1591 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1592 // an FP value, pointer or vector, don't do this optimization because a select
1593 // between them is very expensive and unlikely to lead to later
1594 // simplification. In these cases, we typically end up with "cond ? v1 : v2"
1595 // where v1 and v2 both require constant pool loads, a big loss.
1596 if (GVElType == Type::getInt1Ty(GV->getContext()) ||
1597 GVElType->isFloatingPointTy() ||
1598 GVElType->isPointerTy() || GVElType->isVectorTy())
1601 // Walk the use list of the global seeing if all the uses are load or store.
1602 // If there is anything else, bail out.
1603 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I)
1604 if (!isa<LoadInst>(I) && !isa<StoreInst>(I))
1607 DEBUG(dbgs() << " *** SHRINKING TO BOOL: " << *GV);
1609 // Create the new global, initializing it to false.
1610 GlobalVariable *NewGV = new GlobalVariable(Type::getInt1Ty(GV->getContext()),
1612 GlobalValue::InternalLinkage,
1613 ConstantInt::getFalse(GV->getContext()),
1615 GV->isThreadLocal());
1616 GV->getParent()->getGlobalList().insert(GV, NewGV);
1618 Constant *InitVal = GV->getInitializer();
1619 assert(InitVal->getType() != Type::getInt1Ty(GV->getContext()) &&
1620 "No reason to shrink to bool!");
1622 // If initialized to zero and storing one into the global, we can use a cast
1623 // instead of a select to synthesize the desired value.
1624 bool IsOneZero = false;
1625 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1626 IsOneZero = InitVal->isNullValue() && CI->isOne();
1628 while (!GV->use_empty()) {
1629 Instruction *UI = cast<Instruction>(GV->use_back());
1630 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1631 // Change the store into a boolean store.
1632 bool StoringOther = SI->getOperand(0) == OtherVal;
1633 // Only do this if we weren't storing a loaded value.
1635 if (StoringOther || SI->getOperand(0) == InitVal)
1636 StoreVal = ConstantInt::get(Type::getInt1Ty(GV->getContext()),
1639 // Otherwise, we are storing a previously loaded copy. To do this,
1640 // change the copy from copying the original value to just copying the
1642 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1644 // If we're already replaced the input, StoredVal will be a cast or
1645 // select instruction. If not, it will be a load of the original
1647 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1648 assert(LI->getOperand(0) == GV && "Not a copy!");
1649 // Insert a new load, to preserve the saved value.
1650 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1652 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1653 "This is not a form that we understand!");
1654 StoreVal = StoredVal->getOperand(0);
1655 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1658 new StoreInst(StoreVal, NewGV, SI);
1660 // Change the load into a load of bool then a select.
1661 LoadInst *LI = cast<LoadInst>(UI);
1662 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1665 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1667 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1669 LI->replaceAllUsesWith(NSI);
1671 UI->eraseFromParent();
1674 GV->eraseFromParent();
1679 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1680 /// it if possible. If we make a change, return true.
1681 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1682 Module::global_iterator &GVI) {
1683 SmallPtrSet<const PHINode*, 16> PHIUsers;
1685 GV->removeDeadConstantUsers();
1687 if (GV->use_empty()) {
1688 DEBUG(dbgs() << "GLOBAL DEAD: " << *GV);
1689 GV->eraseFromParent();
1694 if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
1696 DEBUG(dbgs() << "Global: " << *GV);
1697 DEBUG(dbgs() << " isLoaded = " << GS.isLoaded << "\n");
1698 DEBUG(dbgs() << " StoredType = ");
1699 switch (GS.StoredType) {
1700 case GlobalStatus::NotStored: DEBUG(dbgs() << "NEVER STORED\n"); break;
1701 case GlobalStatus::isInitializerStored: DEBUG(dbgs() << "INIT STORED\n");
1703 case GlobalStatus::isStoredOnce: DEBUG(dbgs() << "STORED ONCE\n"); break;
1704 case GlobalStatus::isStored: DEBUG(dbgs() << "stored\n"); break;
1706 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
1707 DEBUG(dbgs() << " StoredOnceValue = " << *GS.StoredOnceValue << "\n");
1708 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
1709 DEBUG(dbgs() << " AccessingFunction = "
1710 << GS.AccessingFunction->getName() << "\n");
1711 DEBUG(dbgs() << " HasMultipleAccessingFunctions = "
1712 << GS.HasMultipleAccessingFunctions << "\n");
1713 DEBUG(dbgs() << " HasNonInstructionUser = "
1714 << GS.HasNonInstructionUser<<"\n");
1715 DEBUG(dbgs() << "\n");
1718 // If this is a first class global and has only one accessing function
1719 // and this function is main (which we know is not recursive we can make
1720 // this global a local variable) we replace the global with a local alloca
1721 // in this function.
1723 // NOTE: It doesn't make sense to promote non single-value types since we
1724 // are just replacing static memory to stack memory.
1726 // If the global is in different address space, don't bring it to stack.
1727 if (!GS.HasMultipleAccessingFunctions &&
1728 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1729 GV->getType()->getElementType()->isSingleValueType() &&
1730 GS.AccessingFunction->getName() == "main" &&
1731 GS.AccessingFunction->hasExternalLinkage() &&
1732 GV->getType()->getAddressSpace() == 0) {
1733 DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV);
1734 Instruction& FirstI = const_cast<Instruction&>(*GS.AccessingFunction
1735 ->getEntryBlock().begin());
1736 const Type* ElemTy = GV->getType()->getElementType();
1737 // FIXME: Pass Global's alignment when globals have alignment
1738 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), &FirstI);
1739 if (!isa<UndefValue>(GV->getInitializer()))
1740 new StoreInst(GV->getInitializer(), Alloca, &FirstI);
1742 GV->replaceAllUsesWith(Alloca);
1743 GV->eraseFromParent();
1748 // If the global is never loaded (but may be stored to), it is dead.
1751 DEBUG(dbgs() << "GLOBAL NEVER LOADED: " << *GV);
1753 // Delete any stores we can find to the global. We may not be able to
1754 // make it completely dead though.
1755 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1757 // If the global is dead now, delete it.
1758 if (GV->use_empty()) {
1759 GV->eraseFromParent();
1765 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1766 DEBUG(dbgs() << "MARKING CONSTANT: " << *GV);
1767 GV->setConstant(true);
1769 // Clean up any obviously simplifiable users now.
1770 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1772 // If the global is dead now, just nuke it.
1773 if (GV->use_empty()) {
1774 DEBUG(dbgs() << " *** Marking constant allowed us to simplify "
1775 << "all users and delete global!\n");
1776 GV->eraseFromParent();
1782 } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
1783 if (TargetData *TD = getAnalysisIfAvailable<TargetData>())
1784 if (GlobalVariable *FirstNewGV = SRAGlobal(GV, *TD)) {
1785 GVI = FirstNewGV; // Don't skip the newly produced globals!
1788 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1789 // If the initial value for the global was an undef value, and if only
1790 // one other value was stored into it, we can just change the
1791 // initializer to be the stored value, then delete all stores to the
1792 // global. This allows us to mark it constant.
1793 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1794 if (isa<UndefValue>(GV->getInitializer())) {
1795 // Change the initial value here.
1796 GV->setInitializer(SOVConstant);
1798 // Clean up any obviously simplifiable users now.
1799 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1801 if (GV->use_empty()) {
1802 DEBUG(dbgs() << " *** Substituting initializer allowed us to "
1803 << "simplify all users and delete global!\n");
1804 GV->eraseFromParent();
1813 // Try to optimize globals based on the knowledge that only one value
1814 // (besides its initializer) is ever stored to the global.
1815 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1816 getAnalysisIfAvailable<TargetData>()))
1819 // Otherwise, if the global was not a boolean, we can shrink it to be a
1821 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1822 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1831 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1832 /// function, changing them to FastCC.
1833 static void ChangeCalleesToFastCall(Function *F) {
1834 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1835 CallSite User(cast<Instruction>(*UI));
1836 User.setCallingConv(CallingConv::Fast);
1840 static AttrListPtr StripNest(const AttrListPtr &Attrs) {
1841 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1842 if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0)
1845 // There can be only one.
1846 return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest);
1852 static void RemoveNestAttribute(Function *F) {
1853 F->setAttributes(StripNest(F->getAttributes()));
1854 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1855 CallSite User(cast<Instruction>(*UI));
1856 User.setAttributes(StripNest(User.getAttributes()));
1860 bool GlobalOpt::OptimizeFunctions(Module &M) {
1861 bool Changed = false;
1862 // Optimize functions.
1863 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1865 // Functions without names cannot be referenced outside this module.
1866 if (!F->hasName() && !F->isDeclaration())
1867 F->setLinkage(GlobalValue::InternalLinkage);
1868 F->removeDeadConstantUsers();
1869 if (F->use_empty() && (F->hasLocalLinkage() || F->hasLinkOnceLinkage())) {
1870 F->eraseFromParent();
1873 } else if (F->hasLocalLinkage()) {
1874 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1875 !F->hasAddressTaken()) {
1876 // If this function has C calling conventions, is not a varargs
1877 // function, and is only called directly, promote it to use the Fast
1878 // calling convention.
1879 F->setCallingConv(CallingConv::Fast);
1880 ChangeCalleesToFastCall(F);
1885 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
1886 !F->hasAddressTaken()) {
1887 // The function is not used by a trampoline intrinsic, so it is safe
1888 // to remove the 'nest' attribute.
1889 RemoveNestAttribute(F);
1898 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1899 bool Changed = false;
1900 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1902 GlobalVariable *GV = GVI++;
1903 // Global variables without names cannot be referenced outside this module.
1904 if (!GV->hasName() && !GV->isDeclaration())
1905 GV->setLinkage(GlobalValue::InternalLinkage);
1906 // Simplify the initializer.
1907 if (GV->hasInitializer())
1908 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GV->getInitializer())) {
1909 TargetData *TD = getAnalysisIfAvailable<TargetData>();
1910 Constant *New = ConstantFoldConstantExpression(CE, TD);
1911 if (New && New != CE)
1912 GV->setInitializer(New);
1914 // Do more involved optimizations if the global is internal.
1915 if (!GV->isConstant() && GV->hasLocalLinkage() &&
1916 GV->hasInitializer())
1917 Changed |= ProcessInternalGlobal(GV, GVI);
1922 /// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1923 /// initializers have an init priority of 65535.
1924 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1925 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1927 if (I->getName() == "llvm.global_ctors") {
1928 // Found it, verify it's an array of { int, void()* }.
1929 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1931 const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1932 if (!STy || STy->getNumElements() != 2 ||
1933 !STy->getElementType(0)->isIntegerTy(32)) return 0;
1934 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1935 if (!PFTy) return 0;
1936 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1937 if (!FTy || !FTy->getReturnType()->isVoidTy() ||
1938 FTy->isVarArg() || FTy->getNumParams() != 0)
1941 // Verify that the initializer is simple enough for us to handle.
1942 if (!I->hasDefinitiveInitializer()) return 0;
1943 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1945 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
1946 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(*i)) {
1947 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1950 // Must have a function or null ptr.
1951 if (!isa<Function>(CS->getOperand(1)))
1954 // Init priority must be standard.
1955 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1956 if (!CI || CI->getZExtValue() != 65535)
1967 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1968 /// return a list of the functions and null terminator as a vector.
1969 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1970 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1971 std::vector<Function*> Result;
1972 Result.reserve(CA->getNumOperands());
1973 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
1974 ConstantStruct *CS = cast<ConstantStruct>(*i);
1975 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1980 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1981 /// specified array, returning the new global to use.
1982 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1983 const std::vector<Function*> &Ctors) {
1984 // If we made a change, reassemble the initializer list.
1985 std::vector<Constant*> CSVals;
1986 CSVals.push_back(ConstantInt::get(Type::getInt32Ty(GCL->getContext()),65535));
1987 CSVals.push_back(0);
1989 // Create the new init list.
1990 std::vector<Constant*> CAList;
1991 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1993 CSVals[1] = Ctors[i];
1995 const Type *FTy = FunctionType::get(Type::getVoidTy(GCL->getContext()),
1997 const PointerType *PFTy = PointerType::getUnqual(FTy);
1998 CSVals[1] = Constant::getNullValue(PFTy);
1999 CSVals[0] = ConstantInt::get(Type::getInt32Ty(GCL->getContext()),
2002 CAList.push_back(ConstantStruct::get(GCL->getContext(), CSVals, false));
2005 // Create the array initializer.
2006 const Type *StructTy =
2007 cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
2008 Constant *CA = ConstantArray::get(ArrayType::get(StructTy,
2009 CAList.size()), CAList);
2011 // If we didn't change the number of elements, don't create a new GV.
2012 if (CA->getType() == GCL->getInitializer()->getType()) {
2013 GCL->setInitializer(CA);
2017 // Create the new global and insert it next to the existing list.
2018 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
2019 GCL->getLinkage(), CA, "",
2020 GCL->isThreadLocal());
2021 GCL->getParent()->getGlobalList().insert(GCL, NGV);
2024 // Nuke the old list, replacing any uses with the new one.
2025 if (!GCL->use_empty()) {
2027 if (V->getType() != GCL->getType())
2028 V = ConstantExpr::getBitCast(V, GCL->getType());
2029 GCL->replaceAllUsesWith(V);
2031 GCL->eraseFromParent();
2040 static Constant *getVal(DenseMap<Value*, Constant*> &ComputedValues,
2042 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
2043 Constant *R = ComputedValues[V];
2044 assert(R && "Reference to an uncomputed value!");
2048 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
2049 /// enough for us to understand. In particular, if it is a cast of something,
2050 /// we punt. We basically just support direct accesses to globals and GEP's of
2051 /// globals. This should be kept up to date with CommitValueTo.
2052 static bool isSimpleEnoughPointerToCommit(Constant *C) {
2053 // Conservatively, avoid aggregate types. This is because we don't
2054 // want to worry about them partially overlapping other stores.
2055 if (!cast<PointerType>(C->getType())->getElementType()->isSingleValueType())
2058 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
2059 // Do not allow weak/linkonce/dllimport/dllexport linkage or
2060 // external globals.
2061 return GV->hasDefinitiveInitializer();
2063 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2064 // Handle a constantexpr gep.
2065 if (CE->getOpcode() == Instruction::GetElementPtr &&
2066 isa<GlobalVariable>(CE->getOperand(0)) &&
2067 cast<GEPOperator>(CE)->isInBounds()) {
2068 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2069 // Do not allow weak/linkonce/dllimport/dllexport linkage or
2070 // external globals.
2071 if (!GV->hasDefinitiveInitializer())
2074 // The first index must be zero.
2075 ConstantInt *CI = dyn_cast<ConstantInt>(*next(CE->op_begin()));
2076 if (!CI || !CI->isZero()) return false;
2078 // The remaining indices must be compile-time known integers within the
2079 // notional bounds of the corresponding static array types.
2080 if (!CE->isGEPWithNoNotionalOverIndexing())
2083 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2088 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
2089 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
2090 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
2091 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
2092 ConstantExpr *Addr, unsigned OpNo) {
2093 // Base case of the recursion.
2094 if (OpNo == Addr->getNumOperands()) {
2095 assert(Val->getType() == Init->getType() && "Type mismatch!");
2099 std::vector<Constant*> Elts;
2100 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
2102 // Break up the constant into its elements.
2103 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
2104 for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i)
2105 Elts.push_back(cast<Constant>(*i));
2106 } else if (isa<ConstantAggregateZero>(Init)) {
2107 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2108 Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
2109 } else if (isa<UndefValue>(Init)) {
2110 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2111 Elts.push_back(UndefValue::get(STy->getElementType(i)));
2113 llvm_unreachable("This code is out of sync with "
2114 " ConstantFoldLoadThroughGEPConstantExpr");
2117 // Replace the element that we are supposed to.
2118 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
2119 unsigned Idx = CU->getZExtValue();
2120 assert(Idx < STy->getNumElements() && "Struct index out of range!");
2121 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
2123 // Return the modified struct.
2124 return ConstantStruct::get(Init->getContext(), &Elts[0], Elts.size(),
2127 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
2128 const SequentialType *InitTy = cast<SequentialType>(Init->getType());
2131 if (const ArrayType *ATy = dyn_cast<ArrayType>(InitTy))
2132 NumElts = ATy->getNumElements();
2134 NumElts = cast<VectorType>(InitTy)->getNumElements();
2137 // Break up the array into elements.
2138 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
2139 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
2140 Elts.push_back(cast<Constant>(*i));
2141 } else if (ConstantVector *CV = dyn_cast<ConstantVector>(Init)) {
2142 for (User::op_iterator i = CV->op_begin(), e = CV->op_end(); i != e; ++i)
2143 Elts.push_back(cast<Constant>(*i));
2144 } else if (isa<ConstantAggregateZero>(Init)) {
2145 Elts.assign(NumElts, Constant::getNullValue(InitTy->getElementType()));
2147 assert(isa<UndefValue>(Init) && "This code is out of sync with "
2148 " ConstantFoldLoadThroughGEPConstantExpr");
2149 Elts.assign(NumElts, UndefValue::get(InitTy->getElementType()));
2152 assert(CI->getZExtValue() < NumElts);
2153 Elts[CI->getZExtValue()] =
2154 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
2156 if (Init->getType()->isArrayTy())
2157 return ConstantArray::get(cast<ArrayType>(InitTy), Elts);
2159 return ConstantVector::get(&Elts[0], Elts.size());
2163 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
2164 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
2165 static void CommitValueTo(Constant *Val, Constant *Addr) {
2166 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
2167 assert(GV->hasInitializer());
2168 GV->setInitializer(Val);
2172 ConstantExpr *CE = cast<ConstantExpr>(Addr);
2173 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2174 GV->setInitializer(EvaluateStoreInto(GV->getInitializer(), Val, CE, 2));
2177 /// ComputeLoadResult - Return the value that would be computed by a load from
2178 /// P after the stores reflected by 'memory' have been performed. If we can't
2179 /// decide, return null.
2180 static Constant *ComputeLoadResult(Constant *P,
2181 const DenseMap<Constant*, Constant*> &Memory) {
2182 // If this memory location has been recently stored, use the stored value: it
2183 // is the most up-to-date.
2184 DenseMap<Constant*, Constant*>::const_iterator I = Memory.find(P);
2185 if (I != Memory.end()) return I->second;
2188 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
2189 if (GV->hasDefinitiveInitializer())
2190 return GV->getInitializer();
2194 // Handle a constantexpr getelementptr.
2195 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
2196 if (CE->getOpcode() == Instruction::GetElementPtr &&
2197 isa<GlobalVariable>(CE->getOperand(0))) {
2198 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2199 if (GV->hasDefinitiveInitializer())
2200 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2203 return 0; // don't know how to evaluate.
2206 /// EvaluateFunction - Evaluate a call to function F, returning true if
2207 /// successful, false if we can't evaluate it. ActualArgs contains the formal
2208 /// arguments for the function.
2209 static bool EvaluateFunction(Function *F, Constant *&RetVal,
2210 const SmallVectorImpl<Constant*> &ActualArgs,
2211 std::vector<Function*> &CallStack,
2212 DenseMap<Constant*, Constant*> &MutatedMemory,
2213 std::vector<GlobalVariable*> &AllocaTmps) {
2214 // Check to see if this function is already executing (recursion). If so,
2215 // bail out. TODO: we might want to accept limited recursion.
2216 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
2219 CallStack.push_back(F);
2221 /// Values - As we compute SSA register values, we store their contents here.
2222 DenseMap<Value*, Constant*> Values;
2224 // Initialize arguments to the incoming values specified.
2226 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
2228 Values[AI] = ActualArgs[ArgNo];
2230 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
2231 /// we can only evaluate any one basic block at most once. This set keeps
2232 /// track of what we have executed so we can detect recursive cases etc.
2233 SmallPtrSet<BasicBlock*, 32> ExecutedBlocks;
2235 // CurInst - The current instruction we're evaluating.
2236 BasicBlock::iterator CurInst = F->begin()->begin();
2238 // This is the main evaluation loop.
2240 Constant *InstResult = 0;
2242 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
2243 if (SI->isVolatile()) return false; // no volatile accesses.
2244 Constant *Ptr = getVal(Values, SI->getOperand(1));
2245 if (!isSimpleEnoughPointerToCommit(Ptr))
2246 // If this is too complex for us to commit, reject it.
2248 Constant *Val = getVal(Values, SI->getOperand(0));
2249 MutatedMemory[Ptr] = Val;
2250 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2251 InstResult = ConstantExpr::get(BO->getOpcode(),
2252 getVal(Values, BO->getOperand(0)),
2253 getVal(Values, BO->getOperand(1)));
2254 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2255 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2256 getVal(Values, CI->getOperand(0)),
2257 getVal(Values, CI->getOperand(1)));
2258 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2259 InstResult = ConstantExpr::getCast(CI->getOpcode(),
2260 getVal(Values, CI->getOperand(0)),
2262 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2264 ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
2265 getVal(Values, SI->getOperand(1)),
2266 getVal(Values, SI->getOperand(2)));
2267 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2268 Constant *P = getVal(Values, GEP->getOperand(0));
2269 SmallVector<Constant*, 8> GEPOps;
2270 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2272 GEPOps.push_back(getVal(Values, *i));
2273 InstResult = cast<GEPOperator>(GEP)->isInBounds() ?
2274 ConstantExpr::getInBoundsGetElementPtr(P, &GEPOps[0], GEPOps.size()) :
2275 ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size());
2276 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2277 if (LI->isVolatile()) return false; // no volatile accesses.
2278 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
2280 if (InstResult == 0) return false; // Could not evaluate load.
2281 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2282 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs.
2283 const Type *Ty = AI->getType()->getElementType();
2284 AllocaTmps.push_back(new GlobalVariable(Ty, false,
2285 GlobalValue::InternalLinkage,
2286 UndefValue::get(Ty),
2288 InstResult = AllocaTmps.back();
2289 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
2291 // Debug info can safely be ignored here.
2292 if (isa<DbgInfoIntrinsic>(CI)) {
2297 // Cannot handle inline asm.
2298 if (isa<InlineAsm>(CI->getCalledValue())) return false;
2300 // Resolve function pointers.
2301 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getCalledValue()));
2302 if (!Callee) return false; // Cannot resolve.
2304 SmallVector<Constant*, 8> Formals;
2305 for (User::op_iterator i = CI->op_begin() + 1, e = CI->op_end();
2307 Formals.push_back(getVal(Values, *i));
2309 if (Callee->isDeclaration()) {
2310 // If this is a function we can constant fold, do it.
2311 if (Constant *C = ConstantFoldCall(Callee, Formals.data(),
2318 if (Callee->getFunctionType()->isVarArg())
2322 // Execute the call, if successful, use the return value.
2323 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
2324 MutatedMemory, AllocaTmps))
2326 InstResult = RetVal;
2328 } else if (isa<TerminatorInst>(CurInst)) {
2329 BasicBlock *NewBB = 0;
2330 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2331 if (BI->isUnconditional()) {
2332 NewBB = BI->getSuccessor(0);
2335 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
2336 if (!Cond) return false; // Cannot determine.
2338 NewBB = BI->getSuccessor(!Cond->getZExtValue());
2340 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2342 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
2343 if (!Val) return false; // Cannot determine.
2344 NewBB = SI->getSuccessor(SI->findCaseValue(Val));
2345 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(CurInst)) {
2346 Value *Val = getVal(Values, IBI->getAddress())->stripPointerCasts();
2347 if (BlockAddress *BA = dyn_cast<BlockAddress>(Val))
2348 NewBB = BA->getBasicBlock();
2350 return false; // Cannot determine.
2351 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
2352 if (RI->getNumOperands())
2353 RetVal = getVal(Values, RI->getOperand(0));
2355 CallStack.pop_back(); // return from fn.
2356 return true; // We succeeded at evaluating this ctor!
2358 // invoke, unwind, unreachable.
2359 return false; // Cannot handle this terminator.
2362 // Okay, we succeeded in evaluating this control flow. See if we have
2363 // executed the new block before. If so, we have a looping function,
2364 // which we cannot evaluate in reasonable time.
2365 if (!ExecutedBlocks.insert(NewBB))
2366 return false; // looped!
2368 // Okay, we have never been in this block before. Check to see if there
2369 // are any PHI nodes. If so, evaluate them with information about where
2371 BasicBlock *OldBB = CurInst->getParent();
2372 CurInst = NewBB->begin();
2374 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2375 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
2377 // Do NOT increment CurInst. We know that the terminator had no value.
2380 // Did not know how to evaluate this!
2384 if (!CurInst->use_empty())
2385 Values[CurInst] = InstResult;
2387 // Advance program counter.
2392 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2393 /// we can. Return true if we can, false otherwise.
2394 static bool EvaluateStaticConstructor(Function *F) {
2395 /// MutatedMemory - For each store we execute, we update this map. Loads
2396 /// check this to get the most up-to-date value. If evaluation is successful,
2397 /// this state is committed to the process.
2398 DenseMap<Constant*, Constant*> MutatedMemory;
2400 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2401 /// to represent its body. This vector is needed so we can delete the
2402 /// temporary globals when we are done.
2403 std::vector<GlobalVariable*> AllocaTmps;
2405 /// CallStack - This is used to detect recursion. In pathological situations
2406 /// we could hit exponential behavior, but at least there is nothing
2408 std::vector<Function*> CallStack;
2410 // Call the function.
2411 Constant *RetValDummy;
2412 bool EvalSuccess = EvaluateFunction(F, RetValDummy,
2413 SmallVector<Constant*, 0>(), CallStack,
2414 MutatedMemory, AllocaTmps);
2416 // We succeeded at evaluation: commit the result.
2417 DEBUG(dbgs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2418 << F->getName() << "' to " << MutatedMemory.size()
2420 for (DenseMap<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2421 E = MutatedMemory.end(); I != E; ++I)
2422 CommitValueTo(I->second, I->first);
2425 // At this point, we are done interpreting. If we created any 'alloca'
2426 // temporaries, release them now.
2427 while (!AllocaTmps.empty()) {
2428 GlobalVariable *Tmp = AllocaTmps.back();
2429 AllocaTmps.pop_back();
2431 // If there are still users of the alloca, the program is doing something
2432 // silly, e.g. storing the address of the alloca somewhere and using it
2433 // later. Since this is undefined, we'll just make it be null.
2434 if (!Tmp->use_empty())
2435 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2444 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2445 /// Return true if anything changed.
2446 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2447 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2448 bool MadeChange = false;
2449 if (Ctors.empty()) return false;
2451 // Loop over global ctors, optimizing them when we can.
2452 for (unsigned i = 0; i != Ctors.size(); ++i) {
2453 Function *F = Ctors[i];
2454 // Found a null terminator in the middle of the list, prune off the rest of
2457 if (i != Ctors.size()-1) {
2464 // We cannot simplify external ctor functions.
2465 if (F->empty()) continue;
2467 // If we can evaluate the ctor at compile time, do.
2468 if (EvaluateStaticConstructor(F)) {
2469 Ctors.erase(Ctors.begin()+i);
2472 ++NumCtorsEvaluated;
2477 if (!MadeChange) return false;
2479 GCL = InstallGlobalCtors(GCL, Ctors);
2483 bool GlobalOpt::OptimizeGlobalAliases(Module &M) {
2484 bool Changed = false;
2486 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end();
2488 Module::alias_iterator J = I++;
2489 // Aliases without names cannot be referenced outside this module.
2490 if (!J->hasName() && !J->isDeclaration())
2491 J->setLinkage(GlobalValue::InternalLinkage);
2492 // If the aliasee may change at link time, nothing can be done - bail out.
2493 if (J->mayBeOverridden())
2496 Constant *Aliasee = J->getAliasee();
2497 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2498 Target->removeDeadConstantUsers();
2499 bool hasOneUse = Target->hasOneUse() && Aliasee->hasOneUse();
2501 // Make all users of the alias use the aliasee instead.
2502 if (!J->use_empty()) {
2503 J->replaceAllUsesWith(Aliasee);
2504 ++NumAliasesResolved;
2508 // If the alias is externally visible, we may still be able to simplify it.
2509 if (!J->hasLocalLinkage()) {
2510 // If the aliasee has internal linkage, give it the name and linkage
2511 // of the alias, and delete the alias. This turns:
2512 // define internal ... @f(...)
2513 // @a = alias ... @f
2515 // define ... @a(...)
2516 if (!Target->hasLocalLinkage())
2519 // Do not perform the transform if multiple aliases potentially target the
2520 // aliasee. This check also ensures that it is safe to replace the section
2521 // and other attributes of the aliasee with those of the alias.
2525 // Give the aliasee the name, linkage and other attributes of the alias.
2526 Target->takeName(J);
2527 Target->setLinkage(J->getLinkage());
2528 Target->GlobalValue::copyAttributesFrom(J);
2531 // Delete the alias.
2532 M.getAliasList().erase(J);
2533 ++NumAliasesRemoved;
2540 bool GlobalOpt::runOnModule(Module &M) {
2541 bool Changed = false;
2543 // Try to find the llvm.globalctors list.
2544 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2546 bool LocalChange = true;
2547 while (LocalChange) {
2548 LocalChange = false;
2550 // Delete functions that are trivially dead, ccc -> fastcc
2551 LocalChange |= OptimizeFunctions(M);
2553 // Optimize global_ctors list.
2555 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2557 // Optimize non-address-taken globals.
2558 LocalChange |= OptimizeGlobalVars(M);
2560 // Resolve aliases, when possible.
2561 LocalChange |= OptimizeGlobalAliases(M);
2562 Changed |= LocalChange;
2565 // TODO: Move all global ctors functions to the end of the module for code