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 #include "llvm/Transforms/IPO.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/ConstantFolding.h"
24 #include "llvm/Analysis/MemoryBuiltins.h"
25 #include "llvm/IR/CallSite.h"
26 #include "llvm/IR/CallingConv.h"
27 #include "llvm/IR/Constants.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/DerivedTypes.h"
30 #include "llvm/IR/GetElementPtrTypeIterator.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/IntrinsicInst.h"
33 #include "llvm/IR/Module.h"
34 #include "llvm/IR/Operator.h"
35 #include "llvm/IR/ValueHandle.h"
36 #include "llvm/Pass.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/ErrorHandling.h"
39 #include "llvm/Support/MathExtras.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include "llvm/Target/TargetLibraryInfo.h"
42 #include "llvm/Transforms/Utils/CtorUtils.h"
43 #include "llvm/Transforms/Utils/GlobalStatus.h"
44 #include "llvm/Transforms/Utils/ModuleUtils.h"
49 #define DEBUG_TYPE "globalopt"
51 STATISTIC(NumMarked , "Number of globals marked constant");
52 STATISTIC(NumUnnamed , "Number of globals marked unnamed_addr");
53 STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
54 STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
55 STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
56 STATISTIC(NumDeleted , "Number of globals deleted");
57 STATISTIC(NumFnDeleted , "Number of functions deleted");
58 STATISTIC(NumGlobUses , "Number of global uses devirtualized");
59 STATISTIC(NumLocalized , "Number of globals localized");
60 STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans");
61 STATISTIC(NumFastCallFns , "Number of functions converted to fastcc");
62 STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
63 STATISTIC(NumNestRemoved , "Number of nest attributes removed");
64 STATISTIC(NumAliasesResolved, "Number of global aliases resolved");
65 STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated");
66 STATISTIC(NumCXXDtorsRemoved, "Number of global C++ destructors removed");
69 struct GlobalOpt : public ModulePass {
70 void getAnalysisUsage(AnalysisUsage &AU) const override {
71 AU.addRequired<TargetLibraryInfo>();
73 static char ID; // Pass identification, replacement for typeid
74 GlobalOpt() : ModulePass(ID) {
75 initializeGlobalOptPass(*PassRegistry::getPassRegistry());
78 bool runOnModule(Module &M) override;
81 bool OptimizeFunctions(Module &M);
82 bool OptimizeGlobalVars(Module &M);
83 bool OptimizeGlobalAliases(Module &M);
84 bool ProcessGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
85 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI,
86 const GlobalStatus &GS);
87 bool OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn);
90 TargetLibraryInfo *TLI;
94 char GlobalOpt::ID = 0;
95 INITIALIZE_PASS_BEGIN(GlobalOpt, "globalopt",
96 "Global Variable Optimizer", false, false)
97 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
98 INITIALIZE_PASS_END(GlobalOpt, "globalopt",
99 "Global Variable Optimizer", false, false)
101 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
103 /// isLeakCheckerRoot - Is this global variable possibly used by a leak checker
104 /// as a root? If so, we might not really want to eliminate the stores to it.
105 static bool isLeakCheckerRoot(GlobalVariable *GV) {
106 // A global variable is a root if it is a pointer, or could plausibly contain
107 // a pointer. There are two challenges; one is that we could have a struct
108 // the has an inner member which is a pointer. We recurse through the type to
109 // detect these (up to a point). The other is that we may actually be a union
110 // of a pointer and another type, and so our LLVM type is an integer which
111 // gets converted into a pointer, or our type is an [i8 x #] with a pointer
112 // potentially contained here.
114 if (GV->hasPrivateLinkage())
117 SmallVector<Type *, 4> Types;
118 Types.push_back(cast<PointerType>(GV->getType())->getElementType());
122 Type *Ty = Types.pop_back_val();
123 switch (Ty->getTypeID()) {
125 case Type::PointerTyID: return true;
126 case Type::ArrayTyID:
127 case Type::VectorTyID: {
128 SequentialType *STy = cast<SequentialType>(Ty);
129 Types.push_back(STy->getElementType());
132 case Type::StructTyID: {
133 StructType *STy = cast<StructType>(Ty);
134 if (STy->isOpaque()) return true;
135 for (StructType::element_iterator I = STy->element_begin(),
136 E = STy->element_end(); I != E; ++I) {
138 if (isa<PointerType>(InnerTy)) return true;
139 if (isa<CompositeType>(InnerTy))
140 Types.push_back(InnerTy);
145 if (--Limit == 0) return true;
146 } while (!Types.empty());
150 /// Given a value that is stored to a global but never read, determine whether
151 /// it's safe to remove the store and the chain of computation that feeds the
153 static bool IsSafeComputationToRemove(Value *V, const TargetLibraryInfo *TLI) {
155 if (isa<Constant>(V))
159 if (isa<LoadInst>(V) || isa<InvokeInst>(V) || isa<Argument>(V) ||
162 if (isAllocationFn(V, TLI))
165 Instruction *I = cast<Instruction>(V);
166 if (I->mayHaveSideEffects())
168 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
169 if (!GEP->hasAllConstantIndices())
171 } else if (I->getNumOperands() != 1) {
175 V = I->getOperand(0);
179 /// CleanupPointerRootUsers - This GV is a pointer root. Loop over all users
180 /// of the global and clean up any that obviously don't assign the global a
181 /// value that isn't dynamically allocated.
183 static bool CleanupPointerRootUsers(GlobalVariable *GV,
184 const TargetLibraryInfo *TLI) {
185 // A brief explanation of leak checkers. The goal is to find bugs where
186 // pointers are forgotten, causing an accumulating growth in memory
187 // usage over time. The common strategy for leak checkers is to whitelist the
188 // memory pointed to by globals at exit. This is popular because it also
189 // solves another problem where the main thread of a C++ program may shut down
190 // before other threads that are still expecting to use those globals. To
191 // handle that case, we expect the program may create a singleton and never
194 bool Changed = false;
196 // If Dead[n].first is the only use of a malloc result, we can delete its
197 // chain of computation and the store to the global in Dead[n].second.
198 SmallVector<std::pair<Instruction *, Instruction *>, 32> Dead;
200 // Constants can't be pointers to dynamically allocated memory.
201 for (Value::user_iterator UI = GV->user_begin(), E = GV->user_end();
204 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
205 Value *V = SI->getValueOperand();
206 if (isa<Constant>(V)) {
208 SI->eraseFromParent();
209 } else if (Instruction *I = dyn_cast<Instruction>(V)) {
211 Dead.push_back(std::make_pair(I, SI));
213 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(U)) {
214 if (isa<Constant>(MSI->getValue())) {
216 MSI->eraseFromParent();
217 } else if (Instruction *I = dyn_cast<Instruction>(MSI->getValue())) {
219 Dead.push_back(std::make_pair(I, MSI));
221 } else if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(U)) {
222 GlobalVariable *MemSrc = dyn_cast<GlobalVariable>(MTI->getSource());
223 if (MemSrc && MemSrc->isConstant()) {
225 MTI->eraseFromParent();
226 } else if (Instruction *I = dyn_cast<Instruction>(MemSrc)) {
228 Dead.push_back(std::make_pair(I, MTI));
230 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
231 if (CE->use_empty()) {
232 CE->destroyConstant();
235 } else if (Constant *C = dyn_cast<Constant>(U)) {
236 if (isSafeToDestroyConstant(C)) {
237 C->destroyConstant();
238 // This could have invalidated UI, start over from scratch.
240 CleanupPointerRootUsers(GV, TLI);
246 for (int i = 0, e = Dead.size(); i != e; ++i) {
247 if (IsSafeComputationToRemove(Dead[i].first, TLI)) {
248 Dead[i].second->eraseFromParent();
249 Instruction *I = Dead[i].first;
251 if (isAllocationFn(I, TLI))
253 Instruction *J = dyn_cast<Instruction>(I->getOperand(0));
256 I->eraseFromParent();
259 I->eraseFromParent();
266 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
267 /// users of the global, cleaning up the obvious ones. This is largely just a
268 /// quick scan over the use list to clean up the easy and obvious cruft. This
269 /// returns true if it made a change.
270 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init,
271 const DataLayout *DL,
272 TargetLibraryInfo *TLI) {
273 bool Changed = false;
274 // Note that we need to use a weak value handle for the worklist items. When
275 // we delete a constant array, we may also be holding pointer to one of its
276 // elements (or an element of one of its elements if we're dealing with an
277 // array of arrays) in the worklist.
278 SmallVector<WeakVH, 8> WorkList(V->user_begin(), V->user_end());
279 while (!WorkList.empty()) {
280 Value *UV = WorkList.pop_back_val();
284 User *U = cast<User>(UV);
286 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
288 // Replace the load with the initializer.
289 LI->replaceAllUsesWith(Init);
290 LI->eraseFromParent();
293 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
294 // Store must be unreachable or storing Init into the global.
295 SI->eraseFromParent();
297 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
298 if (CE->getOpcode() == Instruction::GetElementPtr) {
299 Constant *SubInit = nullptr;
301 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
302 Changed |= CleanupConstantGlobalUsers(CE, SubInit, DL, TLI);
303 } else if ((CE->getOpcode() == Instruction::BitCast &&
304 CE->getType()->isPointerTy()) ||
305 CE->getOpcode() == Instruction::AddrSpaceCast) {
306 // Pointer cast, delete any stores and memsets to the global.
307 Changed |= CleanupConstantGlobalUsers(CE, nullptr, DL, TLI);
310 if (CE->use_empty()) {
311 CE->destroyConstant();
314 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
315 // Do not transform "gepinst (gep constexpr (GV))" here, because forming
316 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
317 // and will invalidate our notion of what Init is.
318 Constant *SubInit = nullptr;
319 if (!isa<ConstantExpr>(GEP->getOperand(0))) {
321 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP, DL, TLI));
322 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
323 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
325 // If the initializer is an all-null value and we have an inbounds GEP,
326 // we already know what the result of any load from that GEP is.
327 // TODO: Handle splats.
328 if (Init && isa<ConstantAggregateZero>(Init) && GEP->isInBounds())
329 SubInit = Constant::getNullValue(GEP->getType()->getElementType());
331 Changed |= CleanupConstantGlobalUsers(GEP, SubInit, DL, TLI);
333 if (GEP->use_empty()) {
334 GEP->eraseFromParent();
337 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
338 if (MI->getRawDest() == V) {
339 MI->eraseFromParent();
343 } else if (Constant *C = dyn_cast<Constant>(U)) {
344 // If we have a chain of dead constantexprs or other things dangling from
345 // us, and if they are all dead, nuke them without remorse.
346 if (isSafeToDestroyConstant(C)) {
347 C->destroyConstant();
348 CleanupConstantGlobalUsers(V, Init, DL, TLI);
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 isSafeToDestroyConstant(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) return false;
377 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
378 !cast<Constant>(GEPI->getOperand(1))->isNullValue())
381 for (User *U : GEPI->users())
382 if (!isSafeSROAElementUse(U))
388 /// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value.
389 /// Look at it and its uses and decide whether it is safe to SROA this global.
391 static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
392 // The user of the global must be a GEP Inst or a ConstantExpr GEP.
393 if (!isa<GetElementPtrInst>(U) &&
394 (!isa<ConstantExpr>(U) ||
395 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
398 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
399 // don't like < 3 operand CE's, and we don't like non-constant integer
400 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some
402 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
403 !cast<Constant>(U->getOperand(1))->isNullValue() ||
404 !isa<ConstantInt>(U->getOperand(2)))
407 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
408 ++GEPI; // Skip over the pointer index.
410 // If this is a use of an array allocation, do a bit more checking for sanity.
411 if (ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
412 uint64_t NumElements = AT->getNumElements();
413 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
415 // Check to make sure that index falls within the array. If not,
416 // something funny is going on, so we won't do the optimization.
418 if (Idx->getZExtValue() >= NumElements)
421 // We cannot scalar repl this level of the array unless any array
422 // sub-indices are in-range constants. In particular, consider:
423 // A[0][i]. We cannot know that the user isn't doing invalid things like
424 // allowing i to index an out-of-range subscript that accesses A[1].
426 // Scalar replacing *just* the outer index of the array is probably not
427 // going to be a win anyway, so just give up.
428 for (++GEPI; // Skip array index.
431 uint64_t NumElements;
432 if (ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
433 NumElements = SubArrayTy->getNumElements();
434 else if (VectorType *SubVectorTy = dyn_cast<VectorType>(*GEPI))
435 NumElements = SubVectorTy->getNumElements();
437 assert((*GEPI)->isStructTy() &&
438 "Indexed GEP type is not array, vector, or struct!");
442 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
443 if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
448 for (User *UU : U->users())
449 if (!isSafeSROAElementUse(UU))
455 /// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
456 /// is safe for us to perform this transformation.
458 static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
459 for (User *U : GV->users())
460 if (!IsUserOfGlobalSafeForSRA(U, GV))
467 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
468 /// variable. This opens the door for other optimizations by exposing the
469 /// behavior of the program in a more fine-grained way. We have determined that
470 /// this transformation is safe already. We return the first global variable we
471 /// insert so that the caller can reprocess it.
472 static GlobalVariable *SRAGlobal(GlobalVariable *GV, const DataLayout &DL) {
473 // Make sure this global only has simple uses that we can SRA.
474 if (!GlobalUsersSafeToSRA(GV))
477 assert(GV->hasLocalLinkage() && !GV->isConstant());
478 Constant *Init = GV->getInitializer();
479 Type *Ty = Init->getType();
481 std::vector<GlobalVariable*> NewGlobals;
482 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
484 // Get the alignment of the global, either explicit or target-specific.
485 unsigned StartAlignment = GV->getAlignment();
486 if (StartAlignment == 0)
487 StartAlignment = DL.getABITypeAlignment(GV->getType());
489 if (StructType *STy = dyn_cast<StructType>(Ty)) {
490 NewGlobals.reserve(STy->getNumElements());
491 const StructLayout &Layout = *DL.getStructLayout(STy);
492 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
493 Constant *In = Init->getAggregateElement(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),
498 GV->getThreadLocalMode(),
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 > DL.getABITypeAlignment(STy->getElementType(i)))
509 NGV->setAlignment(NewAlign);
511 } else if (SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
512 unsigned NumElements = 0;
513 if (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 nullptr; // It's not worth it.
520 NewGlobals.reserve(NumElements);
522 uint64_t EltSize = DL.getTypeAllocSize(STy->getElementType());
523 unsigned EltAlign = DL.getABITypeAlignment(STy->getElementType());
524 for (unsigned i = 0, e = NumElements; i != e; ++i) {
525 Constant *In = Init->getAggregateElement(i);
526 assert(In && "Couldn't get element of initializer?");
528 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
529 GlobalVariable::InternalLinkage,
530 In, GV->getName()+"."+Twine(i),
531 GV->getThreadLocalMode(),
532 GV->getType()->getAddressSpace());
533 Globals.insert(GV, NGV);
534 NewGlobals.push_back(NGV);
536 // Calculate the known alignment of the field. If the original aggregate
537 // had 256 byte alignment for example, something might depend on that:
538 // propagate info to each field.
539 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
540 if (NewAlign > EltAlign)
541 NGV->setAlignment(NewAlign);
545 if (NewGlobals.empty())
548 DEBUG(dbgs() << "PERFORMING GLOBAL SRA ON: " << *GV);
550 Constant *NullInt =Constant::getNullValue(Type::getInt32Ty(GV->getContext()));
552 // Loop over all of the uses of the global, replacing the constantexpr geps,
553 // with smaller constantexpr geps or direct references.
554 while (!GV->use_empty()) {
555 User *GEP = GV->user_back();
556 assert(((isa<ConstantExpr>(GEP) &&
557 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
558 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
560 // Ignore the 1th operand, which has to be zero or else the program is quite
561 // broken (undefined). Get the 2nd operand, which is the structure or array
563 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
564 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
566 Value *NewPtr = NewGlobals[Val];
568 // Form a shorter GEP if needed.
569 if (GEP->getNumOperands() > 3) {
570 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
571 SmallVector<Constant*, 8> Idxs;
572 Idxs.push_back(NullInt);
573 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
574 Idxs.push_back(CE->getOperand(i));
575 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr), Idxs);
577 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
578 SmallVector<Value*, 8> Idxs;
579 Idxs.push_back(NullInt);
580 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
581 Idxs.push_back(GEPI->getOperand(i));
582 NewPtr = GetElementPtrInst::Create(NewPtr, Idxs,
583 GEPI->getName()+"."+Twine(Val),GEPI);
586 GEP->replaceAllUsesWith(NewPtr);
588 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
589 GEPI->eraseFromParent();
591 cast<ConstantExpr>(GEP)->destroyConstant();
594 // Delete the old global, now that it is dead.
598 // Loop over the new globals array deleting any globals that are obviously
599 // dead. This can arise due to scalarization of a structure or an array that
600 // has elements that are dead.
601 unsigned FirstGlobal = 0;
602 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
603 if (NewGlobals[i]->use_empty()) {
604 Globals.erase(NewGlobals[i]);
605 if (FirstGlobal == i) ++FirstGlobal;
608 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : nullptr;
611 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
612 /// value will trap if the value is dynamically null. PHIs keeps track of any
613 /// phi nodes we've seen to avoid reprocessing them.
614 static bool AllUsesOfValueWillTrapIfNull(const Value *V,
615 SmallPtrSetImpl<const PHINode*> &PHIs) {
616 for (const User *U : V->users())
617 if (isa<LoadInst>(U)) {
619 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
620 if (SI->getOperand(0) == V) {
621 //cerr << "NONTRAPPING USE: " << *U;
622 return false; // Storing the value.
624 } else if (const CallInst *CI = dyn_cast<CallInst>(U)) {
625 if (CI->getCalledValue() != V) {
626 //cerr << "NONTRAPPING USE: " << *U;
627 return false; // Not calling the ptr
629 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(U)) {
630 if (II->getCalledValue() != V) {
631 //cerr << "NONTRAPPING USE: " << *U;
632 return false; // Not calling the ptr
634 } else if (const BitCastInst *CI = dyn_cast<BitCastInst>(U)) {
635 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
636 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
637 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
638 } else if (const PHINode *PN = dyn_cast<PHINode>(U)) {
639 // If we've already seen this phi node, ignore it, it has already been
641 if (PHIs.insert(PN) && !AllUsesOfValueWillTrapIfNull(PN, PHIs))
643 } else if (isa<ICmpInst>(U) &&
644 isa<ConstantPointerNull>(U->getOperand(1))) {
645 // Ignore icmp X, null
647 //cerr << "NONTRAPPING USE: " << *U;
654 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
655 /// from GV will trap if the loaded value is null. Note that this also permits
656 /// comparisons of the loaded value against null, as a special case.
657 static bool AllUsesOfLoadedValueWillTrapIfNull(const GlobalVariable *GV) {
658 for (const User *U : GV->users())
659 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
660 SmallPtrSet<const PHINode*, 8> PHIs;
661 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
663 } else if (isa<StoreInst>(U)) {
664 // Ignore stores to the global.
666 // We don't know or understand this user, bail out.
667 //cerr << "UNKNOWN USER OF GLOBAL!: " << *U;
673 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
674 bool Changed = false;
675 for (auto UI = V->user_begin(), E = V->user_end(); UI != E; ) {
676 Instruction *I = cast<Instruction>(*UI++);
677 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
678 LI->setOperand(0, NewV);
680 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
681 if (SI->getOperand(1) == V) {
682 SI->setOperand(1, NewV);
685 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
687 if (CS.getCalledValue() == V) {
688 // Calling through the pointer! Turn into a direct call, but be careful
689 // that the pointer is not also being passed as an argument.
690 CS.setCalledFunction(NewV);
692 bool PassedAsArg = false;
693 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
694 if (CS.getArgument(i) == V) {
696 CS.setArgument(i, NewV);
700 // Being passed as an argument also. Be careful to not invalidate UI!
701 UI = V->user_begin();
704 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
705 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
706 ConstantExpr::getCast(CI->getOpcode(),
707 NewV, CI->getType()));
708 if (CI->use_empty()) {
710 CI->eraseFromParent();
712 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
713 // Should handle GEP here.
714 SmallVector<Constant*, 8> Idxs;
715 Idxs.reserve(GEPI->getNumOperands()-1);
716 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
718 if (Constant *C = dyn_cast<Constant>(*i))
722 if (Idxs.size() == GEPI->getNumOperands()-1)
723 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
724 ConstantExpr::getGetElementPtr(NewV, Idxs));
725 if (GEPI->use_empty()) {
727 GEPI->eraseFromParent();
736 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
737 /// value stored into it. If there are uses of the loaded value that would trap
738 /// if the loaded value is dynamically null, then we know that they cannot be
739 /// reachable with a null optimize away the load.
740 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV,
741 const DataLayout *DL,
742 TargetLibraryInfo *TLI) {
743 bool Changed = false;
745 // Keep track of whether we are able to remove all the uses of the global
746 // other than the store that defines it.
747 bool AllNonStoreUsesGone = true;
749 // Replace all uses of loads with uses of uses of the stored value.
750 for (Value::user_iterator GUI = GV->user_begin(), E = GV->user_end(); GUI != E;){
751 User *GlobalUser = *GUI++;
752 if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) {
753 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
754 // If we were able to delete all uses of the loads
755 if (LI->use_empty()) {
756 LI->eraseFromParent();
759 AllNonStoreUsesGone = false;
761 } else if (isa<StoreInst>(GlobalUser)) {
762 // Ignore the store that stores "LV" to the global.
763 assert(GlobalUser->getOperand(1) == GV &&
764 "Must be storing *to* the global");
766 AllNonStoreUsesGone = false;
768 // If we get here we could have other crazy uses that are transitively
770 assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) ||
771 isa<ConstantExpr>(GlobalUser) || isa<CmpInst>(GlobalUser) ||
772 isa<BitCastInst>(GlobalUser) ||
773 isa<GetElementPtrInst>(GlobalUser)) &&
774 "Only expect load and stores!");
779 DEBUG(dbgs() << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV);
783 // If we nuked all of the loads, then none of the stores are needed either,
784 // nor is the global.
785 if (AllNonStoreUsesGone) {
786 if (isLeakCheckerRoot(GV)) {
787 Changed |= CleanupPointerRootUsers(GV, TLI);
790 CleanupConstantGlobalUsers(GV, nullptr, DL, TLI);
792 if (GV->use_empty()) {
793 DEBUG(dbgs() << " *** GLOBAL NOW DEAD!\n");
795 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, const DataLayout *DL,
805 TargetLibraryInfo *TLI) {
806 for (Value::user_iterator UI = V->user_begin(), E = V->user_end(); UI != E; )
807 if (Instruction *I = dyn_cast<Instruction>(*UI++))
808 if (Constant *NewC = ConstantFoldInstruction(I, DL, TLI)) {
809 I->replaceAllUsesWith(NewC);
811 // Advance UI to the next non-I use to avoid invalidating it!
812 // Instructions could multiply use V.
813 while (UI != E && *UI == I)
815 I->eraseFromParent();
819 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
820 /// variable, and transforms the program as if it always contained the result of
821 /// the specified malloc. Because it is always the result of the specified
822 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
823 /// malloc into a global, and any loads of GV as uses of the new global.
824 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
827 ConstantInt *NElements,
828 const DataLayout *DL,
829 TargetLibraryInfo *TLI) {
830 DEBUG(errs() << "PROMOTING GLOBAL: " << *GV << " CALL = " << *CI << '\n');
833 if (NElements->getZExtValue() == 1)
834 GlobalType = AllocTy;
836 // If we have an array allocation, the global variable is of an array.
837 GlobalType = ArrayType::get(AllocTy, NElements->getZExtValue());
839 // Create the new global variable. The contents of the malloc'd memory is
840 // undefined, so initialize with an undef value.
841 GlobalVariable *NewGV = new GlobalVariable(*GV->getParent(),
843 GlobalValue::InternalLinkage,
844 UndefValue::get(GlobalType),
845 GV->getName()+".body",
847 GV->getThreadLocalMode());
849 // If there are bitcast users of the malloc (which is typical, usually we have
850 // a malloc + bitcast) then replace them with uses of the new global. Update
851 // other users to use the global as well.
852 BitCastInst *TheBC = nullptr;
853 while (!CI->use_empty()) {
854 Instruction *User = cast<Instruction>(CI->user_back());
855 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
856 if (BCI->getType() == NewGV->getType()) {
857 BCI->replaceAllUsesWith(NewGV);
858 BCI->eraseFromParent();
860 BCI->setOperand(0, NewGV);
864 TheBC = new BitCastInst(NewGV, CI->getType(), "newgv", CI);
865 User->replaceUsesOfWith(CI, TheBC);
869 Constant *RepValue = NewGV;
870 if (NewGV->getType() != GV->getType()->getElementType())
871 RepValue = ConstantExpr::getBitCast(RepValue,
872 GV->getType()->getElementType());
874 // If there is a comparison against null, we will insert a global bool to
875 // keep track of whether the global was initialized yet or not.
876 GlobalVariable *InitBool =
877 new GlobalVariable(Type::getInt1Ty(GV->getContext()), false,
878 GlobalValue::InternalLinkage,
879 ConstantInt::getFalse(GV->getContext()),
880 GV->getName()+".init", GV->getThreadLocalMode());
881 bool InitBoolUsed = false;
883 // Loop over all uses of GV, processing them in turn.
884 while (!GV->use_empty()) {
885 if (StoreInst *SI = dyn_cast<StoreInst>(GV->user_back())) {
886 // The global is initialized when the store to it occurs.
887 new StoreInst(ConstantInt::getTrue(GV->getContext()), InitBool, false, 0,
888 SI->getOrdering(), SI->getSynchScope(), SI);
889 SI->eraseFromParent();
893 LoadInst *LI = cast<LoadInst>(GV->user_back());
894 while (!LI->use_empty()) {
895 Use &LoadUse = *LI->use_begin();
896 ICmpInst *ICI = dyn_cast<ICmpInst>(LoadUse.getUser());
902 // Replace the cmp X, 0 with a use of the bool value.
903 // Sink the load to where the compare was, if atomic rules allow us to.
904 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", false, 0,
905 LI->getOrdering(), LI->getSynchScope(),
906 LI->isUnordered() ? (Instruction*)ICI : LI);
908 switch (ICI->getPredicate()) {
909 default: llvm_unreachable("Unknown ICmp Predicate!");
910 case ICmpInst::ICMP_ULT:
911 case ICmpInst::ICMP_SLT: // X < null -> always false
912 LV = ConstantInt::getFalse(GV->getContext());
914 case ICmpInst::ICMP_ULE:
915 case ICmpInst::ICMP_SLE:
916 case ICmpInst::ICMP_EQ:
917 LV = BinaryOperator::CreateNot(LV, "notinit", ICI);
919 case ICmpInst::ICMP_NE:
920 case ICmpInst::ICMP_UGE:
921 case ICmpInst::ICMP_SGE:
922 case ICmpInst::ICMP_UGT:
923 case ICmpInst::ICMP_SGT:
926 ICI->replaceAllUsesWith(LV);
927 ICI->eraseFromParent();
929 LI->eraseFromParent();
932 // If the initialization boolean was used, insert it, otherwise delete it.
934 while (!InitBool->use_empty()) // Delete initializations
935 cast<StoreInst>(InitBool->user_back())->eraseFromParent();
938 GV->getParent()->getGlobalList().insert(GV, InitBool);
940 // Now the GV is dead, nuke it and the malloc..
941 GV->eraseFromParent();
942 CI->eraseFromParent();
944 // To further other optimizations, loop over all users of NewGV and try to
945 // constant prop them. This will promote GEP instructions with constant
946 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
947 ConstantPropUsersOf(NewGV, DL, TLI);
948 if (RepValue != NewGV)
949 ConstantPropUsersOf(RepValue, DL, TLI);
954 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
955 /// to make sure that there are no complex uses of V. We permit simple things
956 /// like dereferencing the pointer, but not storing through the address, unless
957 /// it is to the specified global.
958 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(const Instruction *V,
959 const GlobalVariable *GV,
960 SmallPtrSetImpl<const PHINode*> &PHIs) {
961 for (const User *U : V->users()) {
962 const Instruction *Inst = cast<Instruction>(U);
964 if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) {
965 continue; // Fine, ignore.
968 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
969 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
970 return false; // Storing the pointer itself... bad.
971 continue; // Otherwise, storing through it, or storing into GV... fine.
974 // Must index into the array and into the struct.
975 if (isa<GetElementPtrInst>(Inst) && Inst->getNumOperands() >= 3) {
976 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs))
981 if (const PHINode *PN = dyn_cast<PHINode>(Inst)) {
982 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI
985 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
990 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) {
991 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs))
1001 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
1002 /// somewhere. Transform all uses of the allocation into loads from the
1003 /// global and uses of the resultant pointer. Further, delete the store into
1004 /// GV. This assumes that these value pass the
1005 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
1006 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
1007 GlobalVariable *GV) {
1008 while (!Alloc->use_empty()) {
1009 Instruction *U = cast<Instruction>(*Alloc->user_begin());
1010 Instruction *InsertPt = U;
1011 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1012 // If this is the store of the allocation into the global, remove it.
1013 if (SI->getOperand(1) == GV) {
1014 SI->eraseFromParent();
1017 } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
1018 // Insert the load in the corresponding predecessor, not right before the
1020 InsertPt = PN->getIncomingBlock(*Alloc->use_begin())->getTerminator();
1021 } else if (isa<BitCastInst>(U)) {
1022 // Must be bitcast between the malloc and store to initialize the global.
1023 ReplaceUsesOfMallocWithGlobal(U, GV);
1024 U->eraseFromParent();
1026 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
1027 // If this is a "GEP bitcast" and the user is a store to the global, then
1028 // just process it as a bitcast.
1029 if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse())
1030 if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->user_back()))
1031 if (SI->getOperand(1) == GV) {
1032 // Must be bitcast GEP between the malloc and store to initialize
1034 ReplaceUsesOfMallocWithGlobal(GEPI, GV);
1035 GEPI->eraseFromParent();
1040 // Insert a load from the global, and use it instead of the malloc.
1041 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
1042 U->replaceUsesOfWith(Alloc, NL);
1046 /// LoadUsesSimpleEnoughForHeapSRA - Verify that all uses of V (a load, or a phi
1047 /// of a load) are simple enough to perform heap SRA on. This permits GEP's
1048 /// that index through the array and struct field, icmps of null, and PHIs.
1049 static bool LoadUsesSimpleEnoughForHeapSRA(const Value *V,
1050 SmallPtrSetImpl<const PHINode*> &LoadUsingPHIs,
1051 SmallPtrSetImpl<const PHINode*> &LoadUsingPHIsPerLoad) {
1052 // We permit two users of the load: setcc comparing against the null
1053 // pointer, and a getelementptr of a specific form.
1054 for (const User *U : V->users()) {
1055 const Instruction *UI = cast<Instruction>(U);
1057 // Comparison against null is ok.
1058 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UI)) {
1059 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1064 // getelementptr is also ok, but only a simple form.
1065 if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(UI)) {
1066 // Must index into the array and into the struct.
1067 if (GEPI->getNumOperands() < 3)
1070 // Otherwise the GEP is ok.
1074 if (const PHINode *PN = dyn_cast<PHINode>(UI)) {
1075 if (!LoadUsingPHIsPerLoad.insert(PN))
1076 // This means some phi nodes are dependent on each other.
1077 // Avoid infinite looping!
1079 if (!LoadUsingPHIs.insert(PN))
1080 // If we have already analyzed this PHI, then it is safe.
1083 // Make sure all uses of the PHI are simple enough to transform.
1084 if (!LoadUsesSimpleEnoughForHeapSRA(PN,
1085 LoadUsingPHIs, LoadUsingPHIsPerLoad))
1091 // Otherwise we don't know what this is, not ok.
1099 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
1100 /// GV are simple enough to perform HeapSRA, return true.
1101 static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(const GlobalVariable *GV,
1102 Instruction *StoredVal) {
1103 SmallPtrSet<const PHINode*, 32> LoadUsingPHIs;
1104 SmallPtrSet<const PHINode*, 32> LoadUsingPHIsPerLoad;
1105 for (const User *U : GV->users())
1106 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
1107 if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs,
1108 LoadUsingPHIsPerLoad))
1110 LoadUsingPHIsPerLoad.clear();
1113 // If we reach here, we know that all uses of the loads and transitive uses
1114 // (through PHI nodes) are simple enough to transform. However, we don't know
1115 // that all inputs the to the PHI nodes are in the same equivalence sets.
1116 // Check to verify that all operands of the PHIs are either PHIS that can be
1117 // transformed, loads from GV, or MI itself.
1118 for (const PHINode *PN : LoadUsingPHIs) {
1119 for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) {
1120 Value *InVal = PN->getIncomingValue(op);
1122 // PHI of the stored value itself is ok.
1123 if (InVal == StoredVal) continue;
1125 if (const PHINode *InPN = dyn_cast<PHINode>(InVal)) {
1126 // One of the PHIs in our set is (optimistically) ok.
1127 if (LoadUsingPHIs.count(InPN))
1132 // Load from GV is ok.
1133 if (const LoadInst *LI = dyn_cast<LoadInst>(InVal))
1134 if (LI->getOperand(0) == GV)
1139 // Anything else is rejected.
1147 static Value *GetHeapSROAValue(Value *V, unsigned FieldNo,
1148 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1149 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1150 std::vector<Value*> &FieldVals = InsertedScalarizedValues[V];
1152 if (FieldNo >= FieldVals.size())
1153 FieldVals.resize(FieldNo+1);
1155 // If we already have this value, just reuse the previously scalarized
1157 if (Value *FieldVal = FieldVals[FieldNo])
1160 // Depending on what instruction this is, we have several cases.
1162 if (LoadInst *LI = dyn_cast<LoadInst>(V)) {
1163 // This is a scalarized version of the load from the global. Just create
1164 // a new Load of the scalarized global.
1165 Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo,
1166 InsertedScalarizedValues,
1168 LI->getName()+".f"+Twine(FieldNo), LI);
1169 } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1170 // PN's type is pointer to struct. Make a new PHI of pointer to struct
1173 PointerType *PTy = cast<PointerType>(PN->getType());
1174 StructType *ST = cast<StructType>(PTy->getElementType());
1176 unsigned AS = PTy->getAddressSpace();
1178 PHINode::Create(PointerType::get(ST->getElementType(FieldNo), AS),
1179 PN->getNumIncomingValues(),
1180 PN->getName()+".f"+Twine(FieldNo), PN);
1182 PHIsToRewrite.push_back(std::make_pair(PN, FieldNo));
1184 llvm_unreachable("Unknown usable value");
1187 return FieldVals[FieldNo] = Result;
1190 /// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1191 /// the load, rewrite the derived value to use the HeapSRoA'd load.
1192 static void RewriteHeapSROALoadUser(Instruction *LoadUser,
1193 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1194 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1195 // If this is a comparison against null, handle it.
1196 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1197 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1198 // If we have a setcc of the loaded pointer, we can use a setcc of any
1200 Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0,
1201 InsertedScalarizedValues, PHIsToRewrite);
1203 Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr,
1204 Constant::getNullValue(NPtr->getType()),
1206 SCI->replaceAllUsesWith(New);
1207 SCI->eraseFromParent();
1211 // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...'
1212 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1213 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1214 && "Unexpected GEPI!");
1216 // Load the pointer for this field.
1217 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1218 Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo,
1219 InsertedScalarizedValues, PHIsToRewrite);
1221 // Create the new GEP idx vector.
1222 SmallVector<Value*, 8> GEPIdx;
1223 GEPIdx.push_back(GEPI->getOperand(1));
1224 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1226 Value *NGEPI = GetElementPtrInst::Create(NewPtr, GEPIdx,
1227 GEPI->getName(), GEPI);
1228 GEPI->replaceAllUsesWith(NGEPI);
1229 GEPI->eraseFromParent();
1233 // Recursively transform the users of PHI nodes. This will lazily create the
1234 // PHIs that are needed for individual elements. Keep track of what PHIs we
1235 // see in InsertedScalarizedValues so that we don't get infinite loops (very
1236 // antisocial). If the PHI is already in InsertedScalarizedValues, it has
1237 // already been seen first by another load, so its uses have already been
1239 PHINode *PN = cast<PHINode>(LoadUser);
1240 if (!InsertedScalarizedValues.insert(std::make_pair(PN,
1241 std::vector<Value*>())).second)
1244 // If this is the first time we've seen this PHI, recursively process all
1246 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
1247 Instruction *User = cast<Instruction>(*UI++);
1248 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1252 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1253 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1254 /// use FieldGlobals instead. All uses of loaded values satisfy
1255 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA.
1256 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1257 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1258 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1259 for (auto UI = Load->user_begin(), E = Load->user_end(); UI != E;) {
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, const DataLayout *DL,
1274 const TargetLibraryInfo *TLI) {
1275 DEBUG(dbgs() << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *CI << '\n');
1276 Type *MAT = getMallocAllocatedType(CI, TLI);
1277 StructType *STy = cast<StructType>(MAT);
1279 // There is guaranteed to be at least one use of the malloc (storing
1280 // it into GV). If there are other uses, change them to be uses of
1281 // the global to simplify later code. This also deletes the store
1283 ReplaceUsesOfMallocWithGlobal(CI, GV);
1285 // Okay, at this point, there are no users of the malloc. Insert N
1286 // new mallocs at the same place as CI, and N globals.
1287 std::vector<Value*> FieldGlobals;
1288 std::vector<Value*> FieldMallocs;
1290 unsigned AS = GV->getType()->getPointerAddressSpace();
1291 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1292 Type *FieldTy = STy->getElementType(FieldNo);
1293 PointerType *PFieldTy = PointerType::get(FieldTy, AS);
1295 GlobalVariable *NGV =
1296 new GlobalVariable(*GV->getParent(),
1297 PFieldTy, false, GlobalValue::InternalLinkage,
1298 Constant::getNullValue(PFieldTy),
1299 GV->getName() + ".f" + Twine(FieldNo), GV,
1300 GV->getThreadLocalMode());
1301 FieldGlobals.push_back(NGV);
1303 unsigned TypeSize = DL->getTypeAllocSize(FieldTy);
1304 if (StructType *ST = dyn_cast<StructType>(FieldTy))
1305 TypeSize = DL->getStructLayout(ST)->getSizeInBytes();
1306 Type *IntPtrTy = DL->getIntPtrType(CI->getType());
1307 Value *NMI = CallInst::CreateMalloc(CI, IntPtrTy, FieldTy,
1308 ConstantInt::get(IntPtrTy, TypeSize),
1310 CI->getName() + ".f" + Twine(FieldNo));
1311 FieldMallocs.push_back(NMI);
1312 new StoreInst(NMI, NGV, CI);
1315 // The tricky aspect of this transformation is handling the case when malloc
1316 // fails. In the original code, malloc failing would set the result pointer
1317 // of malloc to null. In this case, some mallocs could succeed and others
1318 // could fail. As such, we emit code that looks like this:
1319 // F0 = malloc(field0)
1320 // F1 = malloc(field1)
1321 // F2 = malloc(field2)
1322 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1323 // if (F0) { free(F0); F0 = 0; }
1324 // if (F1) { free(F1); F1 = 0; }
1325 // if (F2) { free(F2); F2 = 0; }
1327 // The malloc can also fail if its argument is too large.
1328 Constant *ConstantZero = ConstantInt::get(CI->getArgOperand(0)->getType(), 0);
1329 Value *RunningOr = new ICmpInst(CI, ICmpInst::ICMP_SLT, CI->getArgOperand(0),
1330 ConstantZero, "isneg");
1331 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1332 Value *Cond = new ICmpInst(CI, ICmpInst::ICMP_EQ, FieldMallocs[i],
1333 Constant::getNullValue(FieldMallocs[i]->getType()),
1335 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", CI);
1338 // Split the basic block at the old malloc.
1339 BasicBlock *OrigBB = CI->getParent();
1340 BasicBlock *ContBB = OrigBB->splitBasicBlock(CI, "malloc_cont");
1342 // Create the block to check the first condition. Put all these blocks at the
1343 // end of the function as they are unlikely to be executed.
1344 BasicBlock *NullPtrBlock = BasicBlock::Create(OrigBB->getContext(),
1346 OrigBB->getParent());
1348 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1349 // branch on RunningOr.
1350 OrigBB->getTerminator()->eraseFromParent();
1351 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1353 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1354 // pointer, because some may be null while others are not.
1355 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1356 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1357 Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
1358 Constant::getNullValue(GVVal->getType()));
1359 BasicBlock *FreeBlock = BasicBlock::Create(Cmp->getContext(), "free_it",
1360 OrigBB->getParent());
1361 BasicBlock *NextBlock = BasicBlock::Create(Cmp->getContext(), "next",
1362 OrigBB->getParent());
1363 Instruction *BI = BranchInst::Create(FreeBlock, NextBlock,
1366 // Fill in FreeBlock.
1367 CallInst::CreateFree(GVVal, BI);
1368 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1370 BranchInst::Create(NextBlock, FreeBlock);
1372 NullPtrBlock = NextBlock;
1375 BranchInst::Create(ContBB, NullPtrBlock);
1377 // CI is no longer needed, remove it.
1378 CI->eraseFromParent();
1380 /// InsertedScalarizedLoads - As we process loads, if we can't immediately
1381 /// update all uses of the load, keep track of what scalarized loads are
1382 /// inserted for a given load.
1383 DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues;
1384 InsertedScalarizedValues[GV] = FieldGlobals;
1386 std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite;
1388 // Okay, the malloc site is completely handled. All of the uses of GV are now
1389 // loads, and all uses of those loads are simple. Rewrite them to use loads
1390 // of the per-field globals instead.
1391 for (auto UI = GV->user_begin(), E = GV->user_end(); UI != E;) {
1392 Instruction *User = cast<Instruction>(*UI++);
1394 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1395 RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite);
1399 // Must be a store of null.
1400 StoreInst *SI = cast<StoreInst>(User);
1401 assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
1402 "Unexpected heap-sra user!");
1404 // Insert a store of null into each global.
1405 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1406 PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType());
1407 Constant *Null = Constant::getNullValue(PT->getElementType());
1408 new StoreInst(Null, FieldGlobals[i], SI);
1410 // Erase the original store.
1411 SI->eraseFromParent();
1414 // While we have PHIs that are interesting to rewrite, do it.
1415 while (!PHIsToRewrite.empty()) {
1416 PHINode *PN = PHIsToRewrite.back().first;
1417 unsigned FieldNo = PHIsToRewrite.back().second;
1418 PHIsToRewrite.pop_back();
1419 PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]);
1420 assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi");
1422 // Add all the incoming values. This can materialize more phis.
1423 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1424 Value *InVal = PN->getIncomingValue(i);
1425 InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues,
1427 FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
1431 // Drop all inter-phi links and any loads that made it this far.
1432 for (DenseMap<Value*, std::vector<Value*> >::iterator
1433 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1435 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1436 PN->dropAllReferences();
1437 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1438 LI->dropAllReferences();
1441 // Delete all the phis and loads now that inter-references are dead.
1442 for (DenseMap<Value*, std::vector<Value*> >::iterator
1443 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1445 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1446 PN->eraseFromParent();
1447 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1448 LI->eraseFromParent();
1451 // The old global is now dead, remove it.
1452 GV->eraseFromParent();
1455 return cast<GlobalVariable>(FieldGlobals[0]);
1458 /// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1459 /// pointer global variable with a single value stored it that is a malloc or
1461 static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
1464 AtomicOrdering Ordering,
1465 Module::global_iterator &GVI,
1466 const DataLayout *DL,
1467 TargetLibraryInfo *TLI) {
1471 // If this is a malloc of an abstract type, don't touch it.
1472 if (!AllocTy->isSized())
1475 // We can't optimize this global unless all uses of it are *known* to be
1476 // of the malloc value, not of the null initializer value (consider a use
1477 // that compares the global's value against zero to see if the malloc has
1478 // been reached). To do this, we check to see if all uses of the global
1479 // would trap if the global were null: this proves that they must all
1480 // happen after the malloc.
1481 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1484 // We can't optimize this if the malloc itself is used in a complex way,
1485 // for example, being stored into multiple globals. This allows the
1486 // malloc to be stored into the specified global, loaded icmp'd, and
1487 // GEP'd. These are all things we could transform to using the global
1489 SmallPtrSet<const PHINode*, 8> PHIs;
1490 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV, PHIs))
1493 // If we have a global that is only initialized with a fixed size malloc,
1494 // transform the program to use global memory instead of malloc'd memory.
1495 // This eliminates dynamic allocation, avoids an indirection accessing the
1496 // data, and exposes the resultant global to further GlobalOpt.
1497 // We cannot optimize the malloc if we cannot determine malloc array size.
1498 Value *NElems = getMallocArraySize(CI, DL, TLI, true);
1502 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
1503 // Restrict this transformation to only working on small allocations
1504 // (2048 bytes currently), as we don't want to introduce a 16M global or
1506 if (NElements->getZExtValue() * DL->getTypeAllocSize(AllocTy) < 2048) {
1507 GVI = OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, DL, TLI);
1511 // If the allocation is an array of structures, consider transforming this
1512 // into multiple malloc'd arrays, one for each field. This is basically
1513 // SRoA for malloc'd memory.
1515 if (Ordering != NotAtomic)
1518 // If this is an allocation of a fixed size array of structs, analyze as a
1519 // variable size array. malloc [100 x struct],1 -> malloc struct, 100
1520 if (NElems == ConstantInt::get(CI->getArgOperand(0)->getType(), 1))
1521 if (ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
1522 AllocTy = AT->getElementType();
1524 StructType *AllocSTy = dyn_cast<StructType>(AllocTy);
1528 // This the structure has an unreasonable number of fields, leave it
1530 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
1531 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, CI)) {
1533 // If this is a fixed size array, transform the Malloc to be an alloc of
1534 // structs. malloc [100 x struct],1 -> malloc struct, 100
1535 if (ArrayType *AT = dyn_cast<ArrayType>(getMallocAllocatedType(CI, TLI))) {
1536 Type *IntPtrTy = DL->getIntPtrType(CI->getType());
1537 unsigned TypeSize = DL->getStructLayout(AllocSTy)->getSizeInBytes();
1538 Value *AllocSize = ConstantInt::get(IntPtrTy, TypeSize);
1539 Value *NumElements = ConstantInt::get(IntPtrTy, AT->getNumElements());
1540 Instruction *Malloc = CallInst::CreateMalloc(CI, IntPtrTy, AllocSTy,
1541 AllocSize, NumElements,
1542 nullptr, CI->getName());
1543 Instruction *Cast = new BitCastInst(Malloc, CI->getType(), "tmp", CI);
1544 CI->replaceAllUsesWith(Cast);
1545 CI->eraseFromParent();
1546 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Malloc))
1547 CI = cast<CallInst>(BCI->getOperand(0));
1549 CI = cast<CallInst>(Malloc);
1552 GVI = PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, DL, TLI, true),
1560 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1561 // that only one value (besides its initializer) is ever stored to the global.
1562 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1563 AtomicOrdering Ordering,
1564 Module::global_iterator &GVI,
1565 const DataLayout *DL,
1566 TargetLibraryInfo *TLI) {
1567 // Ignore no-op GEPs and bitcasts.
1568 StoredOnceVal = StoredOnceVal->stripPointerCasts();
1570 // If we are dealing with a pointer global that is initialized to null and
1571 // only has one (non-null) value stored into it, then we can optimize any
1572 // users of the loaded value (often calls and loads) that would trap if the
1574 if (GV->getInitializer()->getType()->isPointerTy() &&
1575 GV->getInitializer()->isNullValue()) {
1576 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1577 if (GV->getInitializer()->getType() != SOVC->getType())
1578 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1580 // Optimize away any trapping uses of the loaded value.
1581 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC, DL, TLI))
1583 } else if (CallInst *CI = extractMallocCall(StoredOnceVal, TLI)) {
1584 Type *MallocType = getMallocAllocatedType(CI, TLI);
1586 TryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType, Ordering, GVI,
1595 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1596 /// two values ever stored into GV are its initializer and OtherVal. See if we
1597 /// can shrink the global into a boolean and select between the two values
1598 /// whenever it is used. This exposes the values to other scalar optimizations.
1599 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1600 Type *GVElType = GV->getType()->getElementType();
1602 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1603 // an FP value, pointer or vector, don't do this optimization because a select
1604 // between them is very expensive and unlikely to lead to later
1605 // simplification. In these cases, we typically end up with "cond ? v1 : v2"
1606 // where v1 and v2 both require constant pool loads, a big loss.
1607 if (GVElType == Type::getInt1Ty(GV->getContext()) ||
1608 GVElType->isFloatingPointTy() ||
1609 GVElType->isPointerTy() || GVElType->isVectorTy())
1612 // Walk the use list of the global seeing if all the uses are load or store.
1613 // If there is anything else, bail out.
1614 for (User *U : GV->users())
1615 if (!isa<LoadInst>(U) && !isa<StoreInst>(U))
1618 DEBUG(dbgs() << " *** SHRINKING TO BOOL: " << *GV);
1620 // Create the new global, initializing it to false.
1621 GlobalVariable *NewGV = new GlobalVariable(Type::getInt1Ty(GV->getContext()),
1623 GlobalValue::InternalLinkage,
1624 ConstantInt::getFalse(GV->getContext()),
1626 GV->getThreadLocalMode(),
1627 GV->getType()->getAddressSpace());
1628 GV->getParent()->getGlobalList().insert(GV, NewGV);
1630 Constant *InitVal = GV->getInitializer();
1631 assert(InitVal->getType() != Type::getInt1Ty(GV->getContext()) &&
1632 "No reason to shrink to bool!");
1634 // If initialized to zero and storing one into the global, we can use a cast
1635 // instead of a select to synthesize the desired value.
1636 bool IsOneZero = false;
1637 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1638 IsOneZero = InitVal->isNullValue() && CI->isOne();
1640 while (!GV->use_empty()) {
1641 Instruction *UI = cast<Instruction>(GV->user_back());
1642 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1643 // Change the store into a boolean store.
1644 bool StoringOther = SI->getOperand(0) == OtherVal;
1645 // Only do this if we weren't storing a loaded value.
1647 if (StoringOther || SI->getOperand(0) == InitVal) {
1648 StoreVal = ConstantInt::get(Type::getInt1Ty(GV->getContext()),
1651 // Otherwise, we are storing a previously loaded copy. To do this,
1652 // change the copy from copying the original value to just copying the
1654 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1656 // If we've already replaced the input, StoredVal will be a cast or
1657 // select instruction. If not, it will be a load of the original
1659 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1660 assert(LI->getOperand(0) == GV && "Not a copy!");
1661 // Insert a new load, to preserve the saved value.
1662 StoreVal = new LoadInst(NewGV, LI->getName()+".b", false, 0,
1663 LI->getOrdering(), LI->getSynchScope(), LI);
1665 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1666 "This is not a form that we understand!");
1667 StoreVal = StoredVal->getOperand(0);
1668 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1671 new StoreInst(StoreVal, NewGV, false, 0,
1672 SI->getOrdering(), SI->getSynchScope(), SI);
1674 // Change the load into a load of bool then a select.
1675 LoadInst *LI = cast<LoadInst>(UI);
1676 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", false, 0,
1677 LI->getOrdering(), LI->getSynchScope(), LI);
1680 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1682 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1684 LI->replaceAllUsesWith(NSI);
1686 UI->eraseFromParent();
1689 // Retain the name of the old global variable. People who are debugging their
1690 // programs may expect these variables to be named the same.
1691 NewGV->takeName(GV);
1692 GV->eraseFromParent();
1697 /// ProcessGlobal - Analyze the specified global variable and optimize it if
1698 /// possible. If we make a change, return true.
1699 bool GlobalOpt::ProcessGlobal(GlobalVariable *GV,
1700 Module::global_iterator &GVI) {
1701 // Do more involved optimizations if the global is internal.
1702 GV->removeDeadConstantUsers();
1704 if (GV->use_empty()) {
1705 DEBUG(dbgs() << "GLOBAL DEAD: " << *GV);
1706 GV->eraseFromParent();
1711 if (!GV->hasLocalLinkage())
1716 if (GlobalStatus::analyzeGlobal(GV, GS))
1719 if (!GS.IsCompared && !GV->hasUnnamedAddr()) {
1720 GV->setUnnamedAddr(true);
1724 if (GV->isConstant() || !GV->hasInitializer())
1727 return ProcessInternalGlobal(GV, GVI, GS);
1730 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1731 /// it if possible. If we make a change, return true.
1732 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1733 Module::global_iterator &GVI,
1734 const GlobalStatus &GS) {
1735 // If this is a first class global and has only one accessing function
1736 // and this function is main (which we know is not recursive), we replace
1737 // the global with a local alloca in this function.
1739 // NOTE: It doesn't make sense to promote non-single-value types since we
1740 // are just replacing static memory to stack memory.
1742 // If the global is in different address space, don't bring it to stack.
1743 if (!GS.HasMultipleAccessingFunctions &&
1744 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1745 GV->getType()->getElementType()->isSingleValueType() &&
1746 GS.AccessingFunction->getName() == "main" &&
1747 GS.AccessingFunction->hasExternalLinkage() &&
1748 GV->getType()->getAddressSpace() == 0) {
1749 DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV);
1750 Instruction &FirstI = const_cast<Instruction&>(*GS.AccessingFunction
1751 ->getEntryBlock().begin());
1752 Type *ElemTy = GV->getType()->getElementType();
1753 // FIXME: Pass Global's alignment when globals have alignment
1754 AllocaInst *Alloca = new AllocaInst(ElemTy, nullptr,
1755 GV->getName(), &FirstI);
1756 if (!isa<UndefValue>(GV->getInitializer()))
1757 new StoreInst(GV->getInitializer(), Alloca, &FirstI);
1759 GV->replaceAllUsesWith(Alloca);
1760 GV->eraseFromParent();
1765 // If the global is never loaded (but may be stored to), it is dead.
1768 DEBUG(dbgs() << "GLOBAL NEVER LOADED: " << *GV);
1771 if (isLeakCheckerRoot(GV)) {
1772 // Delete any constant stores to the global.
1773 Changed = CleanupPointerRootUsers(GV, TLI);
1775 // Delete any stores we can find to the global. We may not be able to
1776 // make it completely dead though.
1777 Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, TLI);
1780 // If the global is dead now, delete it.
1781 if (GV->use_empty()) {
1782 GV->eraseFromParent();
1788 } else if (GS.StoredType <= GlobalStatus::InitializerStored) {
1789 DEBUG(dbgs() << "MARKING CONSTANT: " << *GV << "\n");
1790 GV->setConstant(true);
1792 // Clean up any obviously simplifiable users now.
1793 CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, TLI);
1795 // If the global is dead now, just nuke it.
1796 if (GV->use_empty()) {
1797 DEBUG(dbgs() << " *** Marking constant allowed us to simplify "
1798 << "all users and delete global!\n");
1799 GV->eraseFromParent();
1805 } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
1806 if (DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>()) {
1807 const DataLayout &DL = DLP->getDataLayout();
1808 if (GlobalVariable *FirstNewGV = SRAGlobal(GV, DL)) {
1809 GVI = FirstNewGV; // Don't skip the newly produced globals!
1813 } else if (GS.StoredType == GlobalStatus::StoredOnce) {
1814 // If the initial value for the global was an undef value, and if only
1815 // one other value was stored into it, we can just change the
1816 // initializer to be the stored value, then delete all stores to the
1817 // global. This allows us to mark it constant.
1818 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1819 if (isa<UndefValue>(GV->getInitializer())) {
1820 // Change the initial value here.
1821 GV->setInitializer(SOVConstant);
1823 // Clean up any obviously simplifiable users now.
1824 CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, TLI);
1826 if (GV->use_empty()) {
1827 DEBUG(dbgs() << " *** Substituting initializer allowed us to "
1828 << "simplify all users and delete global!\n");
1829 GV->eraseFromParent();
1838 // Try to optimize globals based on the knowledge that only one value
1839 // (besides its initializer) is ever stored to the global.
1840 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GS.Ordering, GVI,
1844 // Otherwise, if the global was not a boolean, we can shrink it to be a
1846 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) {
1847 if (GS.Ordering == NotAtomic) {
1848 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1859 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1860 /// function, changing them to FastCC.
1861 static void ChangeCalleesToFastCall(Function *F) {
1862 for (User *U : F->users()) {
1863 if (isa<BlockAddress>(U))
1865 CallSite CS(cast<Instruction>(U));
1866 CS.setCallingConv(CallingConv::Fast);
1870 static AttributeSet StripNest(LLVMContext &C, const AttributeSet &Attrs) {
1871 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1872 unsigned Index = Attrs.getSlotIndex(i);
1873 if (!Attrs.getSlotAttributes(i).hasAttribute(Index, Attribute::Nest))
1876 // There can be only one.
1877 return Attrs.removeAttribute(C, Index, Attribute::Nest);
1883 static void RemoveNestAttribute(Function *F) {
1884 F->setAttributes(StripNest(F->getContext(), F->getAttributes()));
1885 for (User *U : F->users()) {
1886 if (isa<BlockAddress>(U))
1888 CallSite CS(cast<Instruction>(U));
1889 CS.setAttributes(StripNest(F->getContext(), CS.getAttributes()));
1893 /// Return true if this is a calling convention that we'd like to change. The
1894 /// idea here is that we don't want to mess with the convention if the user
1895 /// explicitly requested something with performance implications like coldcc,
1896 /// GHC, or anyregcc.
1897 static bool isProfitableToMakeFastCC(Function *F) {
1898 CallingConv::ID CC = F->getCallingConv();
1899 // FIXME: Is it worth transforming x86_stdcallcc and x86_fastcallcc?
1900 return CC == CallingConv::C || CC == CallingConv::X86_ThisCall;
1903 bool GlobalOpt::OptimizeFunctions(Module &M) {
1904 bool Changed = false;
1905 // Optimize functions.
1906 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1908 // Functions without names cannot be referenced outside this module.
1909 if (!F->hasName() && !F->isDeclaration() && !F->hasLocalLinkage())
1910 F->setLinkage(GlobalValue::InternalLinkage);
1911 F->removeDeadConstantUsers();
1912 if (F->isDefTriviallyDead()) {
1913 F->eraseFromParent();
1916 } else if (F->hasLocalLinkage()) {
1917 if (isProfitableToMakeFastCC(F) && !F->isVarArg() &&
1918 !F->hasAddressTaken()) {
1919 // If this function has a calling convention worth changing, is not a
1920 // varargs function, and is only called directly, promote it to use the
1921 // Fast calling convention.
1922 F->setCallingConv(CallingConv::Fast);
1923 ChangeCalleesToFastCall(F);
1928 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
1929 !F->hasAddressTaken()) {
1930 // The function is not used by a trampoline intrinsic, so it is safe
1931 // to remove the 'nest' attribute.
1932 RemoveNestAttribute(F);
1941 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1942 bool Changed = false;
1944 SmallSet<const Comdat *, 8> NotDiscardableComdats;
1945 for (const GlobalVariable &GV : M.globals())
1946 if (const Comdat *C = GV.getComdat())
1947 if (!GV.isDiscardableIfUnused())
1948 NotDiscardableComdats.insert(C);
1950 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1952 GlobalVariable *GV = GVI++;
1953 // Global variables without names cannot be referenced outside this module.
1954 if (!GV->hasName() && !GV->isDeclaration() && !GV->hasLocalLinkage())
1955 GV->setLinkage(GlobalValue::InternalLinkage);
1956 // Simplify the initializer.
1957 if (GV->hasInitializer())
1958 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GV->getInitializer())) {
1959 Constant *New = ConstantFoldConstantExpression(CE, DL, TLI);
1960 if (New && New != CE)
1961 GV->setInitializer(New);
1964 if (GV->isDiscardableIfUnused()) {
1965 if (const Comdat *C = GV->getComdat())
1966 if (NotDiscardableComdats.count(C))
1968 Changed |= ProcessGlobal(GV, GVI);
1975 isSimpleEnoughValueToCommit(Constant *C,
1976 SmallPtrSetImpl<Constant*> &SimpleConstants,
1977 const DataLayout *DL);
1980 /// isSimpleEnoughValueToCommit - Return true if the specified constant can be
1981 /// handled by the code generator. We don't want to generate something like:
1982 /// void *X = &X/42;
1983 /// because the code generator doesn't have a relocation that can handle that.
1985 /// This function should be called if C was not found (but just got inserted)
1986 /// in SimpleConstants to avoid having to rescan the same constants all the
1988 static bool isSimpleEnoughValueToCommitHelper(Constant *C,
1989 SmallPtrSetImpl<Constant*> &SimpleConstants,
1990 const DataLayout *DL) {
1991 // Simple global addresses are supported, do not allow dllimport or
1992 // thread-local globals.
1993 if (auto *GV = dyn_cast<GlobalValue>(C))
1994 return !GV->hasDLLImportStorageClass() && !GV->isThreadLocal();
1996 // Simple integer, undef, constant aggregate zero, etc are all supported.
1997 if (C->getNumOperands() == 0 || isa<BlockAddress>(C))
2000 // Aggregate values are safe if all their elements are.
2001 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C) ||
2002 isa<ConstantVector>(C)) {
2003 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) {
2004 Constant *Op = cast<Constant>(C->getOperand(i));
2005 if (!isSimpleEnoughValueToCommit(Op, SimpleConstants, DL))
2011 // We don't know exactly what relocations are allowed in constant expressions,
2012 // so we allow &global+constantoffset, which is safe and uniformly supported
2014 ConstantExpr *CE = cast<ConstantExpr>(C);
2015 switch (CE->getOpcode()) {
2016 case Instruction::BitCast:
2017 // Bitcast is fine if the casted value is fine.
2018 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL);
2020 case Instruction::IntToPtr:
2021 case Instruction::PtrToInt:
2022 // int <=> ptr is fine if the int type is the same size as the
2024 if (!DL || DL->getTypeSizeInBits(CE->getType()) !=
2025 DL->getTypeSizeInBits(CE->getOperand(0)->getType()))
2027 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL);
2029 // GEP is fine if it is simple + constant offset.
2030 case Instruction::GetElementPtr:
2031 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
2032 if (!isa<ConstantInt>(CE->getOperand(i)))
2034 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL);
2036 case Instruction::Add:
2037 // We allow simple+cst.
2038 if (!isa<ConstantInt>(CE->getOperand(1)))
2040 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL);
2046 isSimpleEnoughValueToCommit(Constant *C,
2047 SmallPtrSetImpl<Constant*> &SimpleConstants,
2048 const DataLayout *DL) {
2049 // If we already checked this constant, we win.
2050 if (!SimpleConstants.insert(C)) return true;
2051 // Check the constant.
2052 return isSimpleEnoughValueToCommitHelper(C, SimpleConstants, DL);
2056 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
2057 /// enough for us to understand. In particular, if it is a cast to anything
2058 /// other than from one pointer type to another pointer type, we punt.
2059 /// We basically just support direct accesses to globals and GEP's of
2060 /// globals. This should be kept up to date with CommitValueTo.
2061 static bool isSimpleEnoughPointerToCommit(Constant *C) {
2062 // Conservatively, avoid aggregate types. This is because we don't
2063 // want to worry about them partially overlapping other stores.
2064 if (!cast<PointerType>(C->getType())->getElementType()->isSingleValueType())
2067 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
2068 // Do not allow weak/*_odr/linkonce linkage or external globals.
2069 return GV->hasUniqueInitializer();
2071 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2072 // Handle a constantexpr gep.
2073 if (CE->getOpcode() == Instruction::GetElementPtr &&
2074 isa<GlobalVariable>(CE->getOperand(0)) &&
2075 cast<GEPOperator>(CE)->isInBounds()) {
2076 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2077 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2078 // external globals.
2079 if (!GV->hasUniqueInitializer())
2082 // The first index must be zero.
2083 ConstantInt *CI = dyn_cast<ConstantInt>(*std::next(CE->op_begin()));
2084 if (!CI || !CI->isZero()) return false;
2086 // The remaining indices must be compile-time known integers within the
2087 // notional bounds of the corresponding static array types.
2088 if (!CE->isGEPWithNoNotionalOverIndexing())
2091 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2093 // A constantexpr bitcast from a pointer to another pointer is a no-op,
2094 // and we know how to evaluate it by moving the bitcast from the pointer
2095 // operand to the value operand.
2096 } else if (CE->getOpcode() == Instruction::BitCast &&
2097 isa<GlobalVariable>(CE->getOperand(0))) {
2098 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2099 // external globals.
2100 return cast<GlobalVariable>(CE->getOperand(0))->hasUniqueInitializer();
2107 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
2108 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
2109 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
2110 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
2111 ConstantExpr *Addr, unsigned OpNo) {
2112 // Base case of the recursion.
2113 if (OpNo == Addr->getNumOperands()) {
2114 assert(Val->getType() == Init->getType() && "Type mismatch!");
2118 SmallVector<Constant*, 32> Elts;
2119 if (StructType *STy = dyn_cast<StructType>(Init->getType())) {
2120 // Break up the constant into its elements.
2121 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2122 Elts.push_back(Init->getAggregateElement(i));
2124 // Replace the element that we are supposed to.
2125 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
2126 unsigned Idx = CU->getZExtValue();
2127 assert(Idx < STy->getNumElements() && "Struct index out of range!");
2128 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
2130 // Return the modified struct.
2131 return ConstantStruct::get(STy, Elts);
2134 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
2135 SequentialType *InitTy = cast<SequentialType>(Init->getType());
2138 if (ArrayType *ATy = dyn_cast<ArrayType>(InitTy))
2139 NumElts = ATy->getNumElements();
2141 NumElts = InitTy->getVectorNumElements();
2143 // Break up the array into elements.
2144 for (uint64_t i = 0, e = NumElts; i != e; ++i)
2145 Elts.push_back(Init->getAggregateElement(i));
2147 assert(CI->getZExtValue() < NumElts);
2148 Elts[CI->getZExtValue()] =
2149 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
2151 if (Init->getType()->isArrayTy())
2152 return ConstantArray::get(cast<ArrayType>(InitTy), Elts);
2153 return ConstantVector::get(Elts);
2156 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
2157 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
2158 static void CommitValueTo(Constant *Val, Constant *Addr) {
2159 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
2160 assert(GV->hasInitializer());
2161 GV->setInitializer(Val);
2165 ConstantExpr *CE = cast<ConstantExpr>(Addr);
2166 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2167 GV->setInitializer(EvaluateStoreInto(GV->getInitializer(), Val, CE, 2));
2172 /// Evaluator - This class evaluates LLVM IR, producing the Constant
2173 /// representing each SSA instruction. Changes to global variables are stored
2174 /// in a mapping that can be iterated over after the evaluation is complete.
2175 /// Once an evaluation call fails, the evaluation object should not be reused.
2178 Evaluator(const DataLayout *DL, const TargetLibraryInfo *TLI)
2179 : DL(DL), TLI(TLI) {
2180 ValueStack.emplace_back();
2184 for (auto &Tmp : AllocaTmps)
2185 // If there are still users of the alloca, the program is doing something
2186 // silly, e.g. storing the address of the alloca somewhere and using it
2187 // later. Since this is undefined, we'll just make it be null.
2188 if (!Tmp->use_empty())
2189 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2192 /// EvaluateFunction - Evaluate a call to function F, returning true if
2193 /// successful, false if we can't evaluate it. ActualArgs contains the formal
2194 /// arguments for the function.
2195 bool EvaluateFunction(Function *F, Constant *&RetVal,
2196 const SmallVectorImpl<Constant*> &ActualArgs);
2198 /// EvaluateBlock - Evaluate all instructions in block BB, returning true if
2199 /// successful, false if we can't evaluate it. NewBB returns the next BB that
2200 /// control flows into, or null upon return.
2201 bool EvaluateBlock(BasicBlock::iterator CurInst, BasicBlock *&NextBB);
2203 Constant *getVal(Value *V) {
2204 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
2205 Constant *R = ValueStack.back().lookup(V);
2206 assert(R && "Reference to an uncomputed value!");
2210 void setVal(Value *V, Constant *C) {
2211 ValueStack.back()[V] = C;
2214 const DenseMap<Constant*, Constant*> &getMutatedMemory() const {
2215 return MutatedMemory;
2218 const SmallPtrSetImpl<GlobalVariable*> &getInvariants() const {
2223 Constant *ComputeLoadResult(Constant *P);
2225 /// ValueStack - As we compute SSA register values, we store their contents
2226 /// here. The back of the deque contains the current function and the stack
2227 /// contains the values in the calling frames.
2228 std::deque<DenseMap<Value*, Constant*>> ValueStack;
2230 /// CallStack - This is used to detect recursion. In pathological situations
2231 /// we could hit exponential behavior, but at least there is nothing
2233 SmallVector<Function*, 4> CallStack;
2235 /// MutatedMemory - For each store we execute, we update this map. Loads
2236 /// check this to get the most up-to-date value. If evaluation is successful,
2237 /// this state is committed to the process.
2238 DenseMap<Constant*, Constant*> MutatedMemory;
2240 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2241 /// to represent its body. This vector is needed so we can delete the
2242 /// temporary globals when we are done.
2243 SmallVector<std::unique_ptr<GlobalVariable>, 32> AllocaTmps;
2245 /// Invariants - These global variables have been marked invariant by the
2246 /// static constructor.
2247 SmallPtrSet<GlobalVariable*, 8> Invariants;
2249 /// SimpleConstants - These are constants we have checked and know to be
2250 /// simple enough to live in a static initializer of a global.
2251 SmallPtrSet<Constant*, 8> SimpleConstants;
2253 const DataLayout *DL;
2254 const TargetLibraryInfo *TLI;
2257 } // anonymous namespace
2259 /// ComputeLoadResult - Return the value that would be computed by a load from
2260 /// P after the stores reflected by 'memory' have been performed. If we can't
2261 /// decide, return null.
2262 Constant *Evaluator::ComputeLoadResult(Constant *P) {
2263 // If this memory location has been recently stored, use the stored value: it
2264 // is the most up-to-date.
2265 DenseMap<Constant*, Constant*>::const_iterator I = MutatedMemory.find(P);
2266 if (I != MutatedMemory.end()) return I->second;
2269 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
2270 if (GV->hasDefinitiveInitializer())
2271 return GV->getInitializer();
2275 // Handle a constantexpr getelementptr.
2276 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
2277 if (CE->getOpcode() == Instruction::GetElementPtr &&
2278 isa<GlobalVariable>(CE->getOperand(0))) {
2279 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2280 if (GV->hasDefinitiveInitializer())
2281 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2284 return nullptr; // don't know how to evaluate.
2287 /// EvaluateBlock - Evaluate all instructions in block BB, returning true if
2288 /// successful, false if we can't evaluate it. NewBB returns the next BB that
2289 /// control flows into, or null upon return.
2290 bool Evaluator::EvaluateBlock(BasicBlock::iterator CurInst,
2291 BasicBlock *&NextBB) {
2292 // This is the main evaluation loop.
2294 Constant *InstResult = nullptr;
2296 DEBUG(dbgs() << "Evaluating Instruction: " << *CurInst << "\n");
2298 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
2299 if (!SI->isSimple()) {
2300 DEBUG(dbgs() << "Store is not simple! Can not evaluate.\n");
2301 return false; // no volatile/atomic accesses.
2303 Constant *Ptr = getVal(SI->getOperand(1));
2304 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
2305 DEBUG(dbgs() << "Folding constant ptr expression: " << *Ptr);
2306 Ptr = ConstantFoldConstantExpression(CE, DL, TLI);
2307 DEBUG(dbgs() << "; To: " << *Ptr << "\n");
2309 if (!isSimpleEnoughPointerToCommit(Ptr)) {
2310 // If this is too complex for us to commit, reject it.
2311 DEBUG(dbgs() << "Pointer is too complex for us to evaluate store.");
2315 Constant *Val = getVal(SI->getOperand(0));
2317 // If this might be too difficult for the backend to handle (e.g. the addr
2318 // of one global variable divided by another) then we can't commit it.
2319 if (!isSimpleEnoughValueToCommit(Val, SimpleConstants, DL)) {
2320 DEBUG(dbgs() << "Store value is too complex to evaluate store. " << *Val
2325 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
2326 if (CE->getOpcode() == Instruction::BitCast) {
2327 DEBUG(dbgs() << "Attempting to resolve bitcast on constant ptr.\n");
2328 // If we're evaluating a store through a bitcast, then we need
2329 // to pull the bitcast off the pointer type and push it onto the
2331 Ptr = CE->getOperand(0);
2333 Type *NewTy = cast<PointerType>(Ptr->getType())->getElementType();
2335 // In order to push the bitcast onto the stored value, a bitcast
2336 // from NewTy to Val's type must be legal. If it's not, we can try
2337 // introspecting NewTy to find a legal conversion.
2338 while (!Val->getType()->canLosslesslyBitCastTo(NewTy)) {
2339 // If NewTy is a struct, we can convert the pointer to the struct
2340 // into a pointer to its first member.
2341 // FIXME: This could be extended to support arrays as well.
2342 if (StructType *STy = dyn_cast<StructType>(NewTy)) {
2343 NewTy = STy->getTypeAtIndex(0U);
2345 IntegerType *IdxTy = IntegerType::get(NewTy->getContext(), 32);
2346 Constant *IdxZero = ConstantInt::get(IdxTy, 0, false);
2347 Constant * const IdxList[] = {IdxZero, IdxZero};
2349 Ptr = ConstantExpr::getGetElementPtr(Ptr, IdxList);
2350 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
2351 Ptr = ConstantFoldConstantExpression(CE, DL, TLI);
2353 // If we can't improve the situation by introspecting NewTy,
2354 // we have to give up.
2356 DEBUG(dbgs() << "Failed to bitcast constant ptr, can not "
2362 // If we found compatible types, go ahead and push the bitcast
2363 // onto the stored value.
2364 Val = ConstantExpr::getBitCast(Val, NewTy);
2366 DEBUG(dbgs() << "Evaluated bitcast: " << *Val << "\n");
2370 MutatedMemory[Ptr] = Val;
2371 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2372 InstResult = ConstantExpr::get(BO->getOpcode(),
2373 getVal(BO->getOperand(0)),
2374 getVal(BO->getOperand(1)));
2375 DEBUG(dbgs() << "Found a BinaryOperator! Simplifying: " << *InstResult
2377 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2378 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2379 getVal(CI->getOperand(0)),
2380 getVal(CI->getOperand(1)));
2381 DEBUG(dbgs() << "Found a CmpInst! Simplifying: " << *InstResult
2383 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2384 InstResult = ConstantExpr::getCast(CI->getOpcode(),
2385 getVal(CI->getOperand(0)),
2387 DEBUG(dbgs() << "Found a Cast! Simplifying: " << *InstResult
2389 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2390 InstResult = ConstantExpr::getSelect(getVal(SI->getOperand(0)),
2391 getVal(SI->getOperand(1)),
2392 getVal(SI->getOperand(2)));
2393 DEBUG(dbgs() << "Found a Select! Simplifying: " << *InstResult
2395 } else if (auto *EVI = dyn_cast<ExtractValueInst>(CurInst)) {
2396 InstResult = ConstantExpr::getExtractValue(
2397 getVal(EVI->getAggregateOperand()), EVI->getIndices());
2398 DEBUG(dbgs() << "Found an ExtractValueInst! Simplifying: " << *InstResult
2400 } else if (auto *IVI = dyn_cast<InsertValueInst>(CurInst)) {
2401 InstResult = ConstantExpr::getInsertValue(
2402 getVal(IVI->getAggregateOperand()),
2403 getVal(IVI->getInsertedValueOperand()), IVI->getIndices());
2404 DEBUG(dbgs() << "Found an InsertValueInst! Simplifying: " << *InstResult
2406 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2407 Constant *P = getVal(GEP->getOperand(0));
2408 SmallVector<Constant*, 8> GEPOps;
2409 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2411 GEPOps.push_back(getVal(*i));
2413 ConstantExpr::getGetElementPtr(P, GEPOps,
2414 cast<GEPOperator>(GEP)->isInBounds());
2415 DEBUG(dbgs() << "Found a GEP! Simplifying: " << *InstResult
2417 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2419 if (!LI->isSimple()) {
2420 DEBUG(dbgs() << "Found a Load! Not a simple load, can not evaluate.\n");
2421 return false; // no volatile/atomic accesses.
2424 Constant *Ptr = getVal(LI->getOperand(0));
2425 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
2426 Ptr = ConstantFoldConstantExpression(CE, DL, TLI);
2427 DEBUG(dbgs() << "Found a constant pointer expression, constant "
2428 "folding: " << *Ptr << "\n");
2430 InstResult = ComputeLoadResult(Ptr);
2432 DEBUG(dbgs() << "Failed to compute load result. Can not evaluate load."
2434 return false; // Could not evaluate load.
2437 DEBUG(dbgs() << "Evaluated load: " << *InstResult << "\n");
2438 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2439 if (AI->isArrayAllocation()) {
2440 DEBUG(dbgs() << "Found an array alloca. Can not evaluate.\n");
2441 return false; // Cannot handle array allocs.
2443 Type *Ty = AI->getType()->getElementType();
2444 AllocaTmps.push_back(
2445 make_unique<GlobalVariable>(Ty, false, GlobalValue::InternalLinkage,
2446 UndefValue::get(Ty), AI->getName()));
2447 InstResult = AllocaTmps.back().get();
2448 DEBUG(dbgs() << "Found an alloca. Result: " << *InstResult << "\n");
2449 } else if (isa<CallInst>(CurInst) || isa<InvokeInst>(CurInst)) {
2450 CallSite CS(CurInst);
2452 // Debug info can safely be ignored here.
2453 if (isa<DbgInfoIntrinsic>(CS.getInstruction())) {
2454 DEBUG(dbgs() << "Ignoring debug info.\n");
2459 // Cannot handle inline asm.
2460 if (isa<InlineAsm>(CS.getCalledValue())) {
2461 DEBUG(dbgs() << "Found inline asm, can not evaluate.\n");
2465 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) {
2466 if (MemSetInst *MSI = dyn_cast<MemSetInst>(II)) {
2467 if (MSI->isVolatile()) {
2468 DEBUG(dbgs() << "Can not optimize a volatile memset " <<
2472 Constant *Ptr = getVal(MSI->getDest());
2473 Constant *Val = getVal(MSI->getValue());
2474 Constant *DestVal = ComputeLoadResult(getVal(Ptr));
2475 if (Val->isNullValue() && DestVal && DestVal->isNullValue()) {
2476 // This memset is a no-op.
2477 DEBUG(dbgs() << "Ignoring no-op memset.\n");
2483 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
2484 II->getIntrinsicID() == Intrinsic::lifetime_end) {
2485 DEBUG(dbgs() << "Ignoring lifetime intrinsic.\n");
2490 if (II->getIntrinsicID() == Intrinsic::invariant_start) {
2491 // We don't insert an entry into Values, as it doesn't have a
2492 // meaningful return value.
2493 if (!II->use_empty()) {
2494 DEBUG(dbgs() << "Found unused invariant_start. Can't evaluate.\n");
2497 ConstantInt *Size = cast<ConstantInt>(II->getArgOperand(0));
2498 Value *PtrArg = getVal(II->getArgOperand(1));
2499 Value *Ptr = PtrArg->stripPointerCasts();
2500 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) {
2501 Type *ElemTy = cast<PointerType>(GV->getType())->getElementType();
2502 if (DL && !Size->isAllOnesValue() &&
2503 Size->getValue().getLimitedValue() >=
2504 DL->getTypeStoreSize(ElemTy)) {
2505 Invariants.insert(GV);
2506 DEBUG(dbgs() << "Found a global var that is an invariant: " << *GV
2509 DEBUG(dbgs() << "Found a global var, but can not treat it as an "
2513 // Continue even if we do nothing.
2518 DEBUG(dbgs() << "Unknown intrinsic. Can not evaluate.\n");
2522 // Resolve function pointers.
2523 Function *Callee = dyn_cast<Function>(getVal(CS.getCalledValue()));
2524 if (!Callee || Callee->mayBeOverridden()) {
2525 DEBUG(dbgs() << "Can not resolve function pointer.\n");
2526 return false; // Cannot resolve.
2529 SmallVector<Constant*, 8> Formals;
2530 for (User::op_iterator i = CS.arg_begin(), e = CS.arg_end(); i != e; ++i)
2531 Formals.push_back(getVal(*i));
2533 if (Callee->isDeclaration()) {
2534 // If this is a function we can constant fold, do it.
2535 if (Constant *C = ConstantFoldCall(Callee, Formals, TLI)) {
2537 DEBUG(dbgs() << "Constant folded function call. Result: " <<
2538 *InstResult << "\n");
2540 DEBUG(dbgs() << "Can not constant fold function call.\n");
2544 if (Callee->getFunctionType()->isVarArg()) {
2545 DEBUG(dbgs() << "Can not constant fold vararg function call.\n");
2549 Constant *RetVal = nullptr;
2550 // Execute the call, if successful, use the return value.
2551 ValueStack.emplace_back();
2552 if (!EvaluateFunction(Callee, RetVal, Formals)) {
2553 DEBUG(dbgs() << "Failed to evaluate function.\n");
2556 ValueStack.pop_back();
2557 InstResult = RetVal;
2560 DEBUG(dbgs() << "Successfully evaluated function. Result: " <<
2561 InstResult << "\n\n");
2563 DEBUG(dbgs() << "Successfully evaluated function. Result: 0\n\n");
2566 } else if (isa<TerminatorInst>(CurInst)) {
2567 DEBUG(dbgs() << "Found a terminator instruction.\n");
2569 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2570 if (BI->isUnconditional()) {
2571 NextBB = BI->getSuccessor(0);
2574 dyn_cast<ConstantInt>(getVal(BI->getCondition()));
2575 if (!Cond) return false; // Cannot determine.
2577 NextBB = BI->getSuccessor(!Cond->getZExtValue());
2579 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2581 dyn_cast<ConstantInt>(getVal(SI->getCondition()));
2582 if (!Val) return false; // Cannot determine.
2583 NextBB = SI->findCaseValue(Val).getCaseSuccessor();
2584 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(CurInst)) {
2585 Value *Val = getVal(IBI->getAddress())->stripPointerCasts();
2586 if (BlockAddress *BA = dyn_cast<BlockAddress>(Val))
2587 NextBB = BA->getBasicBlock();
2589 return false; // Cannot determine.
2590 } else if (isa<ReturnInst>(CurInst)) {
2593 // invoke, unwind, resume, unreachable.
2594 DEBUG(dbgs() << "Can not handle terminator.");
2595 return false; // Cannot handle this terminator.
2598 // We succeeded at evaluating this block!
2599 DEBUG(dbgs() << "Successfully evaluated block.\n");
2602 // Did not know how to evaluate this!
2603 DEBUG(dbgs() << "Failed to evaluate block due to unhandled instruction."
2608 if (!CurInst->use_empty()) {
2609 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(InstResult))
2610 InstResult = ConstantFoldConstantExpression(CE, DL, TLI);
2612 setVal(CurInst, InstResult);
2615 // If we just processed an invoke, we finished evaluating the block.
2616 if (InvokeInst *II = dyn_cast<InvokeInst>(CurInst)) {
2617 NextBB = II->getNormalDest();
2618 DEBUG(dbgs() << "Found an invoke instruction. Finished Block.\n\n");
2622 // Advance program counter.
2627 /// EvaluateFunction - Evaluate a call to function F, returning true if
2628 /// successful, false if we can't evaluate it. ActualArgs contains the formal
2629 /// arguments for the function.
2630 bool Evaluator::EvaluateFunction(Function *F, Constant *&RetVal,
2631 const SmallVectorImpl<Constant*> &ActualArgs) {
2632 // Check to see if this function is already executing (recursion). If so,
2633 // bail out. TODO: we might want to accept limited recursion.
2634 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
2637 CallStack.push_back(F);
2639 // Initialize arguments to the incoming values specified.
2641 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
2643 setVal(AI, ActualArgs[ArgNo]);
2645 // ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
2646 // we can only evaluate any one basic block at most once. This set keeps
2647 // track of what we have executed so we can detect recursive cases etc.
2648 SmallPtrSet<BasicBlock*, 32> ExecutedBlocks;
2650 // CurBB - The current basic block we're evaluating.
2651 BasicBlock *CurBB = F->begin();
2653 BasicBlock::iterator CurInst = CurBB->begin();
2656 BasicBlock *NextBB = nullptr; // Initialized to avoid compiler warnings.
2657 DEBUG(dbgs() << "Trying to evaluate BB: " << *CurBB << "\n");
2659 if (!EvaluateBlock(CurInst, NextBB))
2663 // Successfully running until there's no next block means that we found
2664 // the return. Fill it the return value and pop the call stack.
2665 ReturnInst *RI = cast<ReturnInst>(CurBB->getTerminator());
2666 if (RI->getNumOperands())
2667 RetVal = getVal(RI->getOperand(0));
2668 CallStack.pop_back();
2672 // Okay, we succeeded in evaluating this control flow. See if we have
2673 // executed the new block before. If so, we have a looping function,
2674 // which we cannot evaluate in reasonable time.
2675 if (!ExecutedBlocks.insert(NextBB))
2676 return false; // looped!
2678 // Okay, we have never been in this block before. Check to see if there
2679 // are any PHI nodes. If so, evaluate them with information about where
2681 PHINode *PN = nullptr;
2682 for (CurInst = NextBB->begin();
2683 (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2684 setVal(PN, getVal(PN->getIncomingValueForBlock(CurBB)));
2686 // Advance to the next block.
2691 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2692 /// we can. Return true if we can, false otherwise.
2693 static bool EvaluateStaticConstructor(Function *F, const DataLayout *DL,
2694 const TargetLibraryInfo *TLI) {
2695 // Call the function.
2696 Evaluator Eval(DL, TLI);
2697 Constant *RetValDummy;
2698 bool EvalSuccess = Eval.EvaluateFunction(F, RetValDummy,
2699 SmallVector<Constant*, 0>());
2702 ++NumCtorsEvaluated;
2704 // We succeeded at evaluation: commit the result.
2705 DEBUG(dbgs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2706 << F->getName() << "' to " << Eval.getMutatedMemory().size()
2708 for (DenseMap<Constant*, Constant*>::const_iterator I =
2709 Eval.getMutatedMemory().begin(), E = Eval.getMutatedMemory().end();
2711 CommitValueTo(I->second, I->first);
2712 for (GlobalVariable *GV : Eval.getInvariants())
2713 GV->setConstant(true);
2719 static int compareNames(Constant *const *A, Constant *const *B) {
2720 return (*A)->getName().compare((*B)->getName());
2723 static void setUsedInitializer(GlobalVariable &V,
2724 const SmallPtrSet<GlobalValue *, 8> &Init) {
2726 V.eraseFromParent();
2730 // Type of pointer to the array of pointers.
2731 PointerType *Int8PtrTy = Type::getInt8PtrTy(V.getContext(), 0);
2733 SmallVector<llvm::Constant *, 8> UsedArray;
2734 for (GlobalValue *GV : Init) {
2736 = ConstantExpr::getPointerBitCastOrAddrSpaceCast(GV, Int8PtrTy);
2737 UsedArray.push_back(Cast);
2739 // Sort to get deterministic order.
2740 array_pod_sort(UsedArray.begin(), UsedArray.end(), compareNames);
2741 ArrayType *ATy = ArrayType::get(Int8PtrTy, UsedArray.size());
2743 Module *M = V.getParent();
2744 V.removeFromParent();
2745 GlobalVariable *NV =
2746 new GlobalVariable(*M, ATy, false, llvm::GlobalValue::AppendingLinkage,
2747 llvm::ConstantArray::get(ATy, UsedArray), "");
2749 NV->setSection("llvm.metadata");
2754 /// \brief An easy to access representation of llvm.used and llvm.compiler.used.
2756 SmallPtrSet<GlobalValue *, 8> Used;
2757 SmallPtrSet<GlobalValue *, 8> CompilerUsed;
2758 GlobalVariable *UsedV;
2759 GlobalVariable *CompilerUsedV;
2762 LLVMUsed(Module &M) {
2763 UsedV = collectUsedGlobalVariables(M, Used, false);
2764 CompilerUsedV = collectUsedGlobalVariables(M, CompilerUsed, true);
2766 typedef SmallPtrSet<GlobalValue *, 8>::iterator iterator;
2767 typedef iterator_range<iterator> used_iterator_range;
2768 iterator usedBegin() { return Used.begin(); }
2769 iterator usedEnd() { return Used.end(); }
2770 used_iterator_range used() {
2771 return used_iterator_range(usedBegin(), usedEnd());
2773 iterator compilerUsedBegin() { return CompilerUsed.begin(); }
2774 iterator compilerUsedEnd() { return CompilerUsed.end(); }
2775 used_iterator_range compilerUsed() {
2776 return used_iterator_range(compilerUsedBegin(), compilerUsedEnd());
2778 bool usedCount(GlobalValue *GV) const { return Used.count(GV); }
2779 bool compilerUsedCount(GlobalValue *GV) const {
2780 return CompilerUsed.count(GV);
2782 bool usedErase(GlobalValue *GV) { return Used.erase(GV); }
2783 bool compilerUsedErase(GlobalValue *GV) { return CompilerUsed.erase(GV); }
2784 bool usedInsert(GlobalValue *GV) { return Used.insert(GV); }
2785 bool compilerUsedInsert(GlobalValue *GV) { return CompilerUsed.insert(GV); }
2787 void syncVariablesAndSets() {
2789 setUsedInitializer(*UsedV, Used);
2791 setUsedInitializer(*CompilerUsedV, CompilerUsed);
2796 static bool hasUseOtherThanLLVMUsed(GlobalAlias &GA, const LLVMUsed &U) {
2797 if (GA.use_empty()) // No use at all.
2800 assert((!U.usedCount(&GA) || !U.compilerUsedCount(&GA)) &&
2801 "We should have removed the duplicated "
2802 "element from llvm.compiler.used");
2803 if (!GA.hasOneUse())
2804 // Strictly more than one use. So at least one is not in llvm.used and
2805 // llvm.compiler.used.
2808 // Exactly one use. Check if it is in llvm.used or llvm.compiler.used.
2809 return !U.usedCount(&GA) && !U.compilerUsedCount(&GA);
2812 static bool hasMoreThanOneUseOtherThanLLVMUsed(GlobalValue &V,
2813 const LLVMUsed &U) {
2815 assert((!U.usedCount(&V) || !U.compilerUsedCount(&V)) &&
2816 "We should have removed the duplicated "
2817 "element from llvm.compiler.used");
2818 if (U.usedCount(&V) || U.compilerUsedCount(&V))
2820 return V.hasNUsesOrMore(N);
2823 static bool mayHaveOtherReferences(GlobalAlias &GA, const LLVMUsed &U) {
2824 if (!GA.hasLocalLinkage())
2827 return U.usedCount(&GA) || U.compilerUsedCount(&GA);
2830 static bool hasUsesToReplace(GlobalAlias &GA, const LLVMUsed &U,
2831 bool &RenameTarget) {
2832 RenameTarget = false;
2834 if (hasUseOtherThanLLVMUsed(GA, U))
2837 // If the alias is externally visible, we may still be able to simplify it.
2838 if (!mayHaveOtherReferences(GA, U))
2841 // If the aliasee has internal linkage, give it the name and linkage
2842 // of the alias, and delete the alias. This turns:
2843 // define internal ... @f(...)
2844 // @a = alias ... @f
2846 // define ... @a(...)
2847 Constant *Aliasee = GA.getAliasee();
2848 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2849 if (!Target->hasLocalLinkage())
2852 // Do not perform the transform if multiple aliases potentially target the
2853 // aliasee. This check also ensures that it is safe to replace the section
2854 // and other attributes of the aliasee with those of the alias.
2855 if (hasMoreThanOneUseOtherThanLLVMUsed(*Target, U))
2858 RenameTarget = true;
2862 bool GlobalOpt::OptimizeGlobalAliases(Module &M) {
2863 bool Changed = false;
2866 for (GlobalValue *GV : Used.used())
2867 Used.compilerUsedErase(GV);
2869 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end();
2871 Module::alias_iterator J = I++;
2872 // Aliases without names cannot be referenced outside this module.
2873 if (!J->hasName() && !J->isDeclaration() && !J->hasLocalLinkage())
2874 J->setLinkage(GlobalValue::InternalLinkage);
2875 // If the aliasee may change at link time, nothing can be done - bail out.
2876 if (J->mayBeOverridden())
2879 Constant *Aliasee = J->getAliasee();
2880 GlobalValue *Target = dyn_cast<GlobalValue>(Aliasee->stripPointerCasts());
2881 // We can't trivially replace the alias with the aliasee if the aliasee is
2882 // non-trivial in some way.
2883 // TODO: Try to handle non-zero GEPs of local aliasees.
2886 Target->removeDeadConstantUsers();
2888 // Make all users of the alias use the aliasee instead.
2890 if (!hasUsesToReplace(*J, Used, RenameTarget))
2893 J->replaceAllUsesWith(ConstantExpr::getBitCast(Aliasee, J->getType()));
2894 ++NumAliasesResolved;
2898 // Give the aliasee the name, linkage and other attributes of the alias.
2899 Target->takeName(J);
2900 Target->setLinkage(J->getLinkage());
2901 Target->setVisibility(J->getVisibility());
2902 Target->setDLLStorageClass(J->getDLLStorageClass());
2904 if (Used.usedErase(J))
2905 Used.usedInsert(Target);
2907 if (Used.compilerUsedErase(J))
2908 Used.compilerUsedInsert(Target);
2909 } else if (mayHaveOtherReferences(*J, Used))
2912 // Delete the alias.
2913 M.getAliasList().erase(J);
2914 ++NumAliasesRemoved;
2918 Used.syncVariablesAndSets();
2923 static Function *FindCXAAtExit(Module &M, TargetLibraryInfo *TLI) {
2924 if (!TLI->has(LibFunc::cxa_atexit))
2927 Function *Fn = M.getFunction(TLI->getName(LibFunc::cxa_atexit));
2932 FunctionType *FTy = Fn->getFunctionType();
2934 // Checking that the function has the right return type, the right number of
2935 // parameters and that they all have pointer types should be enough.
2936 if (!FTy->getReturnType()->isIntegerTy() ||
2937 FTy->getNumParams() != 3 ||
2938 !FTy->getParamType(0)->isPointerTy() ||
2939 !FTy->getParamType(1)->isPointerTy() ||
2940 !FTy->getParamType(2)->isPointerTy())
2946 /// cxxDtorIsEmpty - Returns whether the given function is an empty C++
2947 /// destructor and can therefore be eliminated.
2948 /// Note that we assume that other optimization passes have already simplified
2949 /// the code so we only look for a function with a single basic block, where
2950 /// the only allowed instructions are 'ret', 'call' to an empty C++ dtor and
2951 /// other side-effect free instructions.
2952 static bool cxxDtorIsEmpty(const Function &Fn,
2953 SmallPtrSet<const Function *, 8> &CalledFunctions) {
2954 // FIXME: We could eliminate C++ destructors if they're readonly/readnone and
2955 // nounwind, but that doesn't seem worth doing.
2956 if (Fn.isDeclaration())
2959 if (++Fn.begin() != Fn.end())
2962 const BasicBlock &EntryBlock = Fn.getEntryBlock();
2963 for (BasicBlock::const_iterator I = EntryBlock.begin(), E = EntryBlock.end();
2965 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
2966 // Ignore debug intrinsics.
2967 if (isa<DbgInfoIntrinsic>(CI))
2970 const Function *CalledFn = CI->getCalledFunction();
2975 SmallPtrSet<const Function *, 8> NewCalledFunctions(CalledFunctions);
2977 // Don't treat recursive functions as empty.
2978 if (!NewCalledFunctions.insert(CalledFn))
2981 if (!cxxDtorIsEmpty(*CalledFn, NewCalledFunctions))
2983 } else if (isa<ReturnInst>(*I))
2984 return true; // We're done.
2985 else if (I->mayHaveSideEffects())
2986 return false; // Destructor with side effects, bail.
2992 bool GlobalOpt::OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) {
2993 /// Itanium C++ ABI p3.3.5:
2995 /// After constructing a global (or local static) object, that will require
2996 /// destruction on exit, a termination function is registered as follows:
2998 /// extern "C" int __cxa_atexit ( void (*f)(void *), void *p, void *d );
3000 /// This registration, e.g. __cxa_atexit(f,p,d), is intended to cause the
3001 /// call f(p) when DSO d is unloaded, before all such termination calls
3002 /// registered before this one. It returns zero if registration is
3003 /// successful, nonzero on failure.
3005 // This pass will look for calls to __cxa_atexit where the function is trivial
3007 bool Changed = false;
3009 for (auto I = CXAAtExitFn->user_begin(), E = CXAAtExitFn->user_end();
3011 // We're only interested in calls. Theoretically, we could handle invoke
3012 // instructions as well, but neither llvm-gcc nor clang generate invokes
3014 CallInst *CI = dyn_cast<CallInst>(*I++);
3019 dyn_cast<Function>(CI->getArgOperand(0)->stripPointerCasts());
3023 SmallPtrSet<const Function *, 8> CalledFunctions;
3024 if (!cxxDtorIsEmpty(*DtorFn, CalledFunctions))
3027 // Just remove the call.
3028 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
3029 CI->eraseFromParent();
3031 ++NumCXXDtorsRemoved;
3039 bool GlobalOpt::runOnModule(Module &M) {
3040 bool Changed = false;
3042 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
3043 DL = DLP ? &DLP->getDataLayout() : nullptr;
3044 TLI = &getAnalysis<TargetLibraryInfo>();
3046 bool LocalChange = true;
3047 while (LocalChange) {
3048 LocalChange = false;
3050 // Delete functions that are trivially dead, ccc -> fastcc
3051 LocalChange |= OptimizeFunctions(M);
3053 // Optimize global_ctors list.
3054 LocalChange |= optimizeGlobalCtorsList(M, [&](Function *F) {
3055 return EvaluateStaticConstructor(F, DL, TLI);
3058 // Optimize non-address-taken globals.
3059 LocalChange |= OptimizeGlobalVars(M);
3061 // Resolve aliases, when possible.
3062 LocalChange |= OptimizeGlobalAliases(M);
3064 // Try to remove trivial global destructors if they are not removed
3066 Function *CXAAtExitFn = FindCXAAtExit(M, TLI);
3068 LocalChange |= OptimizeEmptyGlobalCXXDtors(CXAAtExitFn);
3070 Changed |= LocalChange;
3073 // TODO: Move all global ctors functions to the end of the module for code