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/Analysis/TargetLibraryInfo.h"
26 #include "llvm/IR/CallSite.h"
27 #include "llvm/IR/CallingConv.h"
28 #include "llvm/IR/Constants.h"
29 #include "llvm/IR/DataLayout.h"
30 #include "llvm/IR/DerivedTypes.h"
31 #include "llvm/IR/GetElementPtrTypeIterator.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/Module.h"
35 #include "llvm/IR/Operator.h"
36 #include "llvm/IR/ValueHandle.h"
37 #include "llvm/Pass.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/ErrorHandling.h"
40 #include "llvm/Support/MathExtras.h"
41 #include "llvm/Support/raw_ostream.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<TargetLibraryInfoWrapperPass>();
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);
89 TargetLibraryInfo *TLI;
90 SmallSet<const Comdat *, 8> NotDiscardableComdats;
94 char GlobalOpt::ID = 0;
95 INITIALIZE_PASS_BEGIN(GlobalOpt, "globalopt",
96 "Global Variable Optimizer", false, false)
97 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
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))) {
320 ConstantExpr *CE = dyn_cast_or_null<ConstantExpr>(
321 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 NGV->setExternallyInitialized(GV->isExternallyInitialized());
501 Globals.insert(GV->getIterator(), NGV);
502 NewGlobals.push_back(NGV);
504 // Calculate the known alignment of the field. If the original aggregate
505 // had 256 byte alignment for example, something might depend on that:
506 // propagate info to each field.
507 uint64_t FieldOffset = Layout.getElementOffset(i);
508 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset);
509 if (NewAlign > DL.getABITypeAlignment(STy->getElementType(i)))
510 NGV->setAlignment(NewAlign);
512 } else if (SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
513 unsigned NumElements = 0;
514 if (ArrayType *ATy = dyn_cast<ArrayType>(STy))
515 NumElements = ATy->getNumElements();
517 NumElements = cast<VectorType>(STy)->getNumElements();
519 if (NumElements > 16 && GV->hasNUsesOrMore(16))
520 return nullptr; // It's not worth it.
521 NewGlobals.reserve(NumElements);
523 uint64_t EltSize = DL.getTypeAllocSize(STy->getElementType());
524 unsigned EltAlign = DL.getABITypeAlignment(STy->getElementType());
525 for (unsigned i = 0, e = NumElements; i != e; ++i) {
526 Constant *In = Init->getAggregateElement(i);
527 assert(In && "Couldn't get element of initializer?");
529 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
530 GlobalVariable::InternalLinkage,
531 In, GV->getName()+"."+Twine(i),
532 GV->getThreadLocalMode(),
533 GV->getType()->getAddressSpace());
534 NGV->setExternallyInitialized(GV->isExternallyInitialized());
535 Globals.insert(GV->getIterator(), NGV);
536 NewGlobals.push_back(NGV);
538 // Calculate the known alignment of the field. If the original aggregate
539 // had 256 byte alignment for example, something might depend on that:
540 // propagate info to each field.
541 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
542 if (NewAlign > EltAlign)
543 NGV->setAlignment(NewAlign);
547 if (NewGlobals.empty())
550 DEBUG(dbgs() << "PERFORMING GLOBAL SRA ON: " << *GV << "\n");
552 Constant *NullInt =Constant::getNullValue(Type::getInt32Ty(GV->getContext()));
554 // Loop over all of the uses of the global, replacing the constantexpr geps,
555 // with smaller constantexpr geps or direct references.
556 while (!GV->use_empty()) {
557 User *GEP = GV->user_back();
558 assert(((isa<ConstantExpr>(GEP) &&
559 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
560 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
562 // Ignore the 1th operand, which has to be zero or else the program is quite
563 // broken (undefined). Get the 2nd operand, which is the structure or array
565 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
566 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
568 Value *NewPtr = NewGlobals[Val];
569 Type *NewTy = NewGlobals[Val]->getValueType();
571 // Form a shorter GEP if needed.
572 if (GEP->getNumOperands() > 3) {
573 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
574 SmallVector<Constant*, 8> Idxs;
575 Idxs.push_back(NullInt);
576 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
577 Idxs.push_back(CE->getOperand(i));
579 ConstantExpr::getGetElementPtr(NewTy, cast<Constant>(NewPtr), Idxs);
581 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
582 SmallVector<Value*, 8> Idxs;
583 Idxs.push_back(NullInt);
584 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
585 Idxs.push_back(GEPI->getOperand(i));
586 NewPtr = GetElementPtrInst::Create(
587 NewTy, NewPtr, Idxs, GEPI->getName() + "." + Twine(Val), GEPI);
590 GEP->replaceAllUsesWith(NewPtr);
592 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
593 GEPI->eraseFromParent();
595 cast<ConstantExpr>(GEP)->destroyConstant();
598 // Delete the old global, now that it is dead.
602 // Loop over the new globals array deleting any globals that are obviously
603 // dead. This can arise due to scalarization of a structure or an array that
604 // has elements that are dead.
605 unsigned FirstGlobal = 0;
606 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
607 if (NewGlobals[i]->use_empty()) {
608 Globals.erase(NewGlobals[i]);
609 if (FirstGlobal == i) ++FirstGlobal;
612 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : nullptr;
615 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
616 /// value will trap if the value is dynamically null. PHIs keeps track of any
617 /// phi nodes we've seen to avoid reprocessing them.
618 static bool AllUsesOfValueWillTrapIfNull(const Value *V,
619 SmallPtrSetImpl<const PHINode*> &PHIs) {
620 for (const User *U : V->users())
621 if (isa<LoadInst>(U)) {
623 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
624 if (SI->getOperand(0) == V) {
625 //cerr << "NONTRAPPING USE: " << *U;
626 return false; // Storing the value.
628 } else if (const CallInst *CI = dyn_cast<CallInst>(U)) {
629 if (CI->getCalledValue() != V) {
630 //cerr << "NONTRAPPING USE: " << *U;
631 return false; // Not calling the ptr
633 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(U)) {
634 if (II->getCalledValue() != V) {
635 //cerr << "NONTRAPPING USE: " << *U;
636 return false; // Not calling the ptr
638 } else if (const BitCastInst *CI = dyn_cast<BitCastInst>(U)) {
639 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
640 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
641 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
642 } else if (const PHINode *PN = dyn_cast<PHINode>(U)) {
643 // If we've already seen this phi node, ignore it, it has already been
645 if (PHIs.insert(PN).second && !AllUsesOfValueWillTrapIfNull(PN, PHIs))
647 } else if (isa<ICmpInst>(U) &&
648 isa<ConstantPointerNull>(U->getOperand(1))) {
649 // Ignore icmp X, null
651 //cerr << "NONTRAPPING USE: " << *U;
658 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
659 /// from GV will trap if the loaded value is null. Note that this also permits
660 /// comparisons of the loaded value against null, as a special case.
661 static bool AllUsesOfLoadedValueWillTrapIfNull(const GlobalVariable *GV) {
662 for (const User *U : GV->users())
663 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
664 SmallPtrSet<const PHINode*, 8> PHIs;
665 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
667 } else if (isa<StoreInst>(U)) {
668 // Ignore stores to the global.
670 // We don't know or understand this user, bail out.
671 //cerr << "UNKNOWN USER OF GLOBAL!: " << *U;
677 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
678 bool Changed = false;
679 for (auto UI = V->user_begin(), E = V->user_end(); UI != E; ) {
680 Instruction *I = cast<Instruction>(*UI++);
681 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
682 LI->setOperand(0, NewV);
684 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
685 if (SI->getOperand(1) == V) {
686 SI->setOperand(1, NewV);
689 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
691 if (CS.getCalledValue() == V) {
692 // Calling through the pointer! Turn into a direct call, but be careful
693 // that the pointer is not also being passed as an argument.
694 CS.setCalledFunction(NewV);
696 bool PassedAsArg = false;
697 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
698 if (CS.getArgument(i) == V) {
700 CS.setArgument(i, NewV);
704 // Being passed as an argument also. Be careful to not invalidate UI!
705 UI = V->user_begin();
708 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
709 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
710 ConstantExpr::getCast(CI->getOpcode(),
711 NewV, CI->getType()));
712 if (CI->use_empty()) {
714 CI->eraseFromParent();
716 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
717 // Should handle GEP here.
718 SmallVector<Constant*, 8> Idxs;
719 Idxs.reserve(GEPI->getNumOperands()-1);
720 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
722 if (Constant *C = dyn_cast<Constant>(*i))
726 if (Idxs.size() == GEPI->getNumOperands()-1)
727 Changed |= OptimizeAwayTrappingUsesOfValue(
728 GEPI, ConstantExpr::getGetElementPtr(nullptr, NewV, Idxs));
729 if (GEPI->use_empty()) {
731 GEPI->eraseFromParent();
740 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
741 /// value stored into it. If there are uses of the loaded value that would trap
742 /// if the loaded value is dynamically null, then we know that they cannot be
743 /// reachable with a null optimize away the load.
744 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV,
745 const DataLayout &DL,
746 TargetLibraryInfo *TLI) {
747 bool Changed = false;
749 // Keep track of whether we are able to remove all the uses of the global
750 // other than the store that defines it.
751 bool AllNonStoreUsesGone = true;
753 // Replace all uses of loads with uses of uses of the stored value.
754 for (Value::user_iterator GUI = GV->user_begin(), E = GV->user_end(); GUI != E;){
755 User *GlobalUser = *GUI++;
756 if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) {
757 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
758 // If we were able to delete all uses of the loads
759 if (LI->use_empty()) {
760 LI->eraseFromParent();
763 AllNonStoreUsesGone = false;
765 } else if (isa<StoreInst>(GlobalUser)) {
766 // Ignore the store that stores "LV" to the global.
767 assert(GlobalUser->getOperand(1) == GV &&
768 "Must be storing *to* the global");
770 AllNonStoreUsesGone = false;
772 // If we get here we could have other crazy uses that are transitively
774 assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) ||
775 isa<ConstantExpr>(GlobalUser) || isa<CmpInst>(GlobalUser) ||
776 isa<BitCastInst>(GlobalUser) ||
777 isa<GetElementPtrInst>(GlobalUser)) &&
778 "Only expect load and stores!");
783 DEBUG(dbgs() << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV << "\n");
787 // If we nuked all of the loads, then none of the stores are needed either,
788 // nor is the global.
789 if (AllNonStoreUsesGone) {
790 if (isLeakCheckerRoot(GV)) {
791 Changed |= CleanupPointerRootUsers(GV, TLI);
794 CleanupConstantGlobalUsers(GV, nullptr, DL, TLI);
796 if (GV->use_empty()) {
797 DEBUG(dbgs() << " *** GLOBAL NOW DEAD!\n");
799 GV->eraseFromParent();
806 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
807 /// instructions that are foldable.
808 static void ConstantPropUsersOf(Value *V, const DataLayout &DL,
809 TargetLibraryInfo *TLI) {
810 for (Value::user_iterator UI = V->user_begin(), E = V->user_end(); UI != E; )
811 if (Instruction *I = dyn_cast<Instruction>(*UI++))
812 if (Constant *NewC = ConstantFoldInstruction(I, DL, TLI)) {
813 I->replaceAllUsesWith(NewC);
815 // Advance UI to the next non-I use to avoid invalidating it!
816 // Instructions could multiply use V.
817 while (UI != E && *UI == I)
819 I->eraseFromParent();
823 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
824 /// variable, and transforms the program as if it always contained the result of
825 /// the specified malloc. Because it is always the result of the specified
826 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
827 /// malloc into a global, and any loads of GV as uses of the new global.
828 static GlobalVariable *
829 OptimizeGlobalAddressOfMalloc(GlobalVariable *GV, CallInst *CI, Type *AllocTy,
830 ConstantInt *NElements, const DataLayout &DL,
831 TargetLibraryInfo *TLI) {
832 DEBUG(errs() << "PROMOTING GLOBAL: " << *GV << " CALL = " << *CI << '\n');
835 if (NElements->getZExtValue() == 1)
836 GlobalType = AllocTy;
838 // If we have an array allocation, the global variable is of an array.
839 GlobalType = ArrayType::get(AllocTy, NElements->getZExtValue());
841 // Create the new global variable. The contents of the malloc'd memory is
842 // undefined, so initialize with an undef value.
843 GlobalVariable *NewGV = new GlobalVariable(*GV->getParent(),
845 GlobalValue::InternalLinkage,
846 UndefValue::get(GlobalType),
847 GV->getName()+".body",
849 GV->getThreadLocalMode());
851 // If there are bitcast users of the malloc (which is typical, usually we have
852 // a malloc + bitcast) then replace them with uses of the new global. Update
853 // other users to use the global as well.
854 BitCastInst *TheBC = nullptr;
855 while (!CI->use_empty()) {
856 Instruction *User = cast<Instruction>(CI->user_back());
857 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
858 if (BCI->getType() == NewGV->getType()) {
859 BCI->replaceAllUsesWith(NewGV);
860 BCI->eraseFromParent();
862 BCI->setOperand(0, NewGV);
866 TheBC = new BitCastInst(NewGV, CI->getType(), "newgv", CI);
867 User->replaceUsesOfWith(CI, TheBC);
871 Constant *RepValue = NewGV;
872 if (NewGV->getType() != GV->getType()->getElementType())
873 RepValue = ConstantExpr::getBitCast(RepValue,
874 GV->getType()->getElementType());
876 // If there is a comparison against null, we will insert a global bool to
877 // keep track of whether the global was initialized yet or not.
878 GlobalVariable *InitBool =
879 new GlobalVariable(Type::getInt1Ty(GV->getContext()), false,
880 GlobalValue::InternalLinkage,
881 ConstantInt::getFalse(GV->getContext()),
882 GV->getName()+".init", GV->getThreadLocalMode());
883 bool InitBoolUsed = false;
885 // Loop over all uses of GV, processing them in turn.
886 while (!GV->use_empty()) {
887 if (StoreInst *SI = dyn_cast<StoreInst>(GV->user_back())) {
888 // The global is initialized when the store to it occurs.
889 new StoreInst(ConstantInt::getTrue(GV->getContext()), InitBool, false, 0,
890 SI->getOrdering(), SI->getSynchScope(), SI);
891 SI->eraseFromParent();
895 LoadInst *LI = cast<LoadInst>(GV->user_back());
896 while (!LI->use_empty()) {
897 Use &LoadUse = *LI->use_begin();
898 ICmpInst *ICI = dyn_cast<ICmpInst>(LoadUse.getUser());
904 // Replace the cmp X, 0 with a use of the bool value.
905 // Sink the load to where the compare was, if atomic rules allow us to.
906 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", false, 0,
907 LI->getOrdering(), LI->getSynchScope(),
908 LI->isUnordered() ? (Instruction*)ICI : LI);
910 switch (ICI->getPredicate()) {
911 default: llvm_unreachable("Unknown ICmp Predicate!");
912 case ICmpInst::ICMP_ULT:
913 case ICmpInst::ICMP_SLT: // X < null -> always false
914 LV = ConstantInt::getFalse(GV->getContext());
916 case ICmpInst::ICMP_ULE:
917 case ICmpInst::ICMP_SLE:
918 case ICmpInst::ICMP_EQ:
919 LV = BinaryOperator::CreateNot(LV, "notinit", ICI);
921 case ICmpInst::ICMP_NE:
922 case ICmpInst::ICMP_UGE:
923 case ICmpInst::ICMP_SGE:
924 case ICmpInst::ICMP_UGT:
925 case ICmpInst::ICMP_SGT:
928 ICI->replaceAllUsesWith(LV);
929 ICI->eraseFromParent();
931 LI->eraseFromParent();
934 // If the initialization boolean was used, insert it, otherwise delete it.
936 while (!InitBool->use_empty()) // Delete initializations
937 cast<StoreInst>(InitBool->user_back())->eraseFromParent();
940 GV->getParent()->getGlobalList().insert(GV->getIterator(), InitBool);
942 // Now the GV is dead, nuke it and the malloc..
943 GV->eraseFromParent();
944 CI->eraseFromParent();
946 // To further other optimizations, loop over all users of NewGV and try to
947 // constant prop them. This will promote GEP instructions with constant
948 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
949 ConstantPropUsersOf(NewGV, DL, TLI);
950 if (RepValue != NewGV)
951 ConstantPropUsersOf(RepValue, DL, TLI);
956 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
957 /// to make sure that there are no complex uses of V. We permit simple things
958 /// like dereferencing the pointer, but not storing through the address, unless
959 /// it is to the specified global.
960 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(const Instruction *V,
961 const GlobalVariable *GV,
962 SmallPtrSetImpl<const PHINode*> &PHIs) {
963 for (const User *U : V->users()) {
964 const Instruction *Inst = cast<Instruction>(U);
966 if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) {
967 continue; // Fine, ignore.
970 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
971 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
972 return false; // Storing the pointer itself... bad.
973 continue; // Otherwise, storing through it, or storing into GV... fine.
976 // Must index into the array and into the struct.
977 if (isa<GetElementPtrInst>(Inst) && Inst->getNumOperands() >= 3) {
978 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs))
983 if (const PHINode *PN = dyn_cast<PHINode>(Inst)) {
984 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI
986 if (PHIs.insert(PN).second)
987 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
992 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) {
993 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs))
1003 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
1004 /// somewhere. Transform all uses of the allocation into loads from the
1005 /// global and uses of the resultant pointer. Further, delete the store into
1006 /// GV. This assumes that these value pass the
1007 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
1008 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
1009 GlobalVariable *GV) {
1010 while (!Alloc->use_empty()) {
1011 Instruction *U = cast<Instruction>(*Alloc->user_begin());
1012 Instruction *InsertPt = U;
1013 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1014 // If this is the store of the allocation into the global, remove it.
1015 if (SI->getOperand(1) == GV) {
1016 SI->eraseFromParent();
1019 } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
1020 // Insert the load in the corresponding predecessor, not right before the
1022 InsertPt = PN->getIncomingBlock(*Alloc->use_begin())->getTerminator();
1023 } else if (isa<BitCastInst>(U)) {
1024 // Must be bitcast between the malloc and store to initialize the global.
1025 ReplaceUsesOfMallocWithGlobal(U, GV);
1026 U->eraseFromParent();
1028 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
1029 // If this is a "GEP bitcast" and the user is a store to the global, then
1030 // just process it as a bitcast.
1031 if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse())
1032 if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->user_back()))
1033 if (SI->getOperand(1) == GV) {
1034 // Must be bitcast GEP between the malloc and store to initialize
1036 ReplaceUsesOfMallocWithGlobal(GEPI, GV);
1037 GEPI->eraseFromParent();
1042 // Insert a load from the global, and use it instead of the malloc.
1043 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
1044 U->replaceUsesOfWith(Alloc, NL);
1048 /// LoadUsesSimpleEnoughForHeapSRA - Verify that all uses of V (a load, or a phi
1049 /// of a load) are simple enough to perform heap SRA on. This permits GEP's
1050 /// that index through the array and struct field, icmps of null, and PHIs.
1051 static bool LoadUsesSimpleEnoughForHeapSRA(const Value *V,
1052 SmallPtrSetImpl<const PHINode*> &LoadUsingPHIs,
1053 SmallPtrSetImpl<const PHINode*> &LoadUsingPHIsPerLoad) {
1054 // We permit two users of the load: setcc comparing against the null
1055 // pointer, and a getelementptr of a specific form.
1056 for (const User *U : V->users()) {
1057 const Instruction *UI = cast<Instruction>(U);
1059 // Comparison against null is ok.
1060 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UI)) {
1061 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1066 // getelementptr is also ok, but only a simple form.
1067 if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(UI)) {
1068 // Must index into the array and into the struct.
1069 if (GEPI->getNumOperands() < 3)
1072 // Otherwise the GEP is ok.
1076 if (const PHINode *PN = dyn_cast<PHINode>(UI)) {
1077 if (!LoadUsingPHIsPerLoad.insert(PN).second)
1078 // This means some phi nodes are dependent on each other.
1079 // Avoid infinite looping!
1081 if (!LoadUsingPHIs.insert(PN).second)
1082 // If we have already analyzed this PHI, then it is safe.
1085 // Make sure all uses of the PHI are simple enough to transform.
1086 if (!LoadUsesSimpleEnoughForHeapSRA(PN,
1087 LoadUsingPHIs, LoadUsingPHIsPerLoad))
1093 // Otherwise we don't know what this is, not ok.
1101 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
1102 /// GV are simple enough to perform HeapSRA, return true.
1103 static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(const GlobalVariable *GV,
1104 Instruction *StoredVal) {
1105 SmallPtrSet<const PHINode*, 32> LoadUsingPHIs;
1106 SmallPtrSet<const PHINode*, 32> LoadUsingPHIsPerLoad;
1107 for (const User *U : GV->users())
1108 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
1109 if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs,
1110 LoadUsingPHIsPerLoad))
1112 LoadUsingPHIsPerLoad.clear();
1115 // If we reach here, we know that all uses of the loads and transitive uses
1116 // (through PHI nodes) are simple enough to transform. However, we don't know
1117 // that all inputs the to the PHI nodes are in the same equivalence sets.
1118 // Check to verify that all operands of the PHIs are either PHIS that can be
1119 // transformed, loads from GV, or MI itself.
1120 for (const PHINode *PN : LoadUsingPHIs) {
1121 for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) {
1122 Value *InVal = PN->getIncomingValue(op);
1124 // PHI of the stored value itself is ok.
1125 if (InVal == StoredVal) continue;
1127 if (const PHINode *InPN = dyn_cast<PHINode>(InVal)) {
1128 // One of the PHIs in our set is (optimistically) ok.
1129 if (LoadUsingPHIs.count(InPN))
1134 // Load from GV is ok.
1135 if (const LoadInst *LI = dyn_cast<LoadInst>(InVal))
1136 if (LI->getOperand(0) == GV)
1141 // Anything else is rejected.
1149 static Value *GetHeapSROAValue(Value *V, unsigned FieldNo,
1150 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1151 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1152 std::vector<Value*> &FieldVals = InsertedScalarizedValues[V];
1154 if (FieldNo >= FieldVals.size())
1155 FieldVals.resize(FieldNo+1);
1157 // If we already have this value, just reuse the previously scalarized
1159 if (Value *FieldVal = FieldVals[FieldNo])
1162 // Depending on what instruction this is, we have several cases.
1164 if (LoadInst *LI = dyn_cast<LoadInst>(V)) {
1165 // This is a scalarized version of the load from the global. Just create
1166 // a new Load of the scalarized global.
1167 Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo,
1168 InsertedScalarizedValues,
1170 LI->getName()+".f"+Twine(FieldNo), LI);
1172 PHINode *PN = cast<PHINode>(V);
1173 // PN's type is pointer to struct. Make a new PHI of pointer to struct
1176 PointerType *PTy = cast<PointerType>(PN->getType());
1177 StructType *ST = cast<StructType>(PTy->getElementType());
1179 unsigned AS = PTy->getAddressSpace();
1181 PHINode::Create(PointerType::get(ST->getElementType(FieldNo), AS),
1182 PN->getNumIncomingValues(),
1183 PN->getName()+".f"+Twine(FieldNo), PN);
1185 PHIsToRewrite.push_back(std::make_pair(PN, FieldNo));
1188 return FieldVals[FieldNo] = Result;
1191 /// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1192 /// the load, rewrite the derived value to use the HeapSRoA'd load.
1193 static void RewriteHeapSROALoadUser(Instruction *LoadUser,
1194 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1195 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1196 // If this is a comparison against null, handle it.
1197 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1198 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1199 // If we have a setcc of the loaded pointer, we can use a setcc of any
1201 Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0,
1202 InsertedScalarizedValues, PHIsToRewrite);
1204 Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr,
1205 Constant::getNullValue(NPtr->getType()),
1207 SCI->replaceAllUsesWith(New);
1208 SCI->eraseFromParent();
1212 // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...'
1213 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1214 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1215 && "Unexpected GEPI!");
1217 // Load the pointer for this field.
1218 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1219 Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo,
1220 InsertedScalarizedValues, PHIsToRewrite);
1222 // Create the new GEP idx vector.
1223 SmallVector<Value*, 8> GEPIdx;
1224 GEPIdx.push_back(GEPI->getOperand(1));
1225 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1227 Value *NGEPI = GetElementPtrInst::Create(GEPI->getResultElementType(), NewPtr, GEPIdx,
1228 GEPI->getName(), GEPI);
1229 GEPI->replaceAllUsesWith(NGEPI);
1230 GEPI->eraseFromParent();
1234 // Recursively transform the users of PHI nodes. This will lazily create the
1235 // PHIs that are needed for individual elements. Keep track of what PHIs we
1236 // see in InsertedScalarizedValues so that we don't get infinite loops (very
1237 // antisocial). If the PHI is already in InsertedScalarizedValues, it has
1238 // already been seen first by another load, so its uses have already been
1240 PHINode *PN = cast<PHINode>(LoadUser);
1241 if (!InsertedScalarizedValues.insert(std::make_pair(PN,
1242 std::vector<Value*>())).second)
1245 // If this is the first time we've seen this PHI, recursively process all
1247 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
1248 Instruction *User = cast<Instruction>(*UI++);
1249 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1253 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1254 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1255 /// use FieldGlobals instead. All uses of loaded values satisfy
1256 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA.
1257 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1258 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1259 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1260 for (auto UI = Load->user_begin(), E = Load->user_end(); UI != E;) {
1261 Instruction *User = cast<Instruction>(*UI++);
1262 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1265 if (Load->use_empty()) {
1266 Load->eraseFromParent();
1267 InsertedScalarizedValues.erase(Load);
1271 /// PerformHeapAllocSRoA - CI is an allocation of an array of structures. Break
1272 /// it up into multiple allocations of arrays of the fields.
1273 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI,
1274 Value *NElems, const DataLayout &DL,
1275 const TargetLibraryInfo *TLI) {
1276 DEBUG(dbgs() << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *CI << '\n');
1277 Type *MAT = getMallocAllocatedType(CI, TLI);
1278 StructType *STy = cast<StructType>(MAT);
1280 // There is guaranteed to be at least one use of the malloc (storing
1281 // it into GV). If there are other uses, change them to be uses of
1282 // the global to simplify later code. This also deletes the store
1284 ReplaceUsesOfMallocWithGlobal(CI, GV);
1286 // Okay, at this point, there are no users of the malloc. Insert N
1287 // new mallocs at the same place as CI, and N globals.
1288 std::vector<Value*> FieldGlobals;
1289 std::vector<Value*> FieldMallocs;
1291 unsigned AS = GV->getType()->getPointerAddressSpace();
1292 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1293 Type *FieldTy = STy->getElementType(FieldNo);
1294 PointerType *PFieldTy = PointerType::get(FieldTy, AS);
1296 GlobalVariable *NGV =
1297 new GlobalVariable(*GV->getParent(),
1298 PFieldTy, false, GlobalValue::InternalLinkage,
1299 Constant::getNullValue(PFieldTy),
1300 GV->getName() + ".f" + Twine(FieldNo), GV,
1301 GV->getThreadLocalMode());
1302 FieldGlobals.push_back(NGV);
1304 unsigned TypeSize = DL.getTypeAllocSize(FieldTy);
1305 if (StructType *ST = dyn_cast<StructType>(FieldTy))
1306 TypeSize = DL.getStructLayout(ST)->getSizeInBytes();
1307 Type *IntPtrTy = DL.getIntPtrType(CI->getType());
1308 Value *NMI = CallInst::CreateMalloc(CI, IntPtrTy, FieldTy,
1309 ConstantInt::get(IntPtrTy, TypeSize),
1311 CI->getName() + ".f" + Twine(FieldNo));
1312 FieldMallocs.push_back(NMI);
1313 new StoreInst(NMI, NGV, CI);
1316 // The tricky aspect of this transformation is handling the case when malloc
1317 // fails. In the original code, malloc failing would set the result pointer
1318 // of malloc to null. In this case, some mallocs could succeed and others
1319 // could fail. As such, we emit code that looks like this:
1320 // F0 = malloc(field0)
1321 // F1 = malloc(field1)
1322 // F2 = malloc(field2)
1323 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1324 // if (F0) { free(F0); F0 = 0; }
1325 // if (F1) { free(F1); F1 = 0; }
1326 // if (F2) { free(F2); F2 = 0; }
1328 // The malloc can also fail if its argument is too large.
1329 Constant *ConstantZero = ConstantInt::get(CI->getArgOperand(0)->getType(), 0);
1330 Value *RunningOr = new ICmpInst(CI, ICmpInst::ICMP_SLT, CI->getArgOperand(0),
1331 ConstantZero, "isneg");
1332 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1333 Value *Cond = new ICmpInst(CI, ICmpInst::ICMP_EQ, FieldMallocs[i],
1334 Constant::getNullValue(FieldMallocs[i]->getType()),
1336 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", CI);
1339 // Split the basic block at the old malloc.
1340 BasicBlock *OrigBB = CI->getParent();
1341 BasicBlock *ContBB =
1342 OrigBB->splitBasicBlock(CI->getIterator(), "malloc_cont");
1344 // Create the block to check the first condition. Put all these blocks at the
1345 // end of the function as they are unlikely to be executed.
1346 BasicBlock *NullPtrBlock = BasicBlock::Create(OrigBB->getContext(),
1348 OrigBB->getParent());
1350 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1351 // branch on RunningOr.
1352 OrigBB->getTerminator()->eraseFromParent();
1353 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1355 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1356 // pointer, because some may be null while others are not.
1357 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1358 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1359 Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
1360 Constant::getNullValue(GVVal->getType()));
1361 BasicBlock *FreeBlock = BasicBlock::Create(Cmp->getContext(), "free_it",
1362 OrigBB->getParent());
1363 BasicBlock *NextBlock = BasicBlock::Create(Cmp->getContext(), "next",
1364 OrigBB->getParent());
1365 Instruction *BI = BranchInst::Create(FreeBlock, NextBlock,
1368 // Fill in FreeBlock.
1369 CallInst::CreateFree(GVVal, BI);
1370 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1372 BranchInst::Create(NextBlock, FreeBlock);
1374 NullPtrBlock = NextBlock;
1377 BranchInst::Create(ContBB, NullPtrBlock);
1379 // CI is no longer needed, remove it.
1380 CI->eraseFromParent();
1382 /// InsertedScalarizedLoads - As we process loads, if we can't immediately
1383 /// update all uses of the load, keep track of what scalarized loads are
1384 /// inserted for a given load.
1385 DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues;
1386 InsertedScalarizedValues[GV] = FieldGlobals;
1388 std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite;
1390 // Okay, the malloc site is completely handled. All of the uses of GV are now
1391 // loads, and all uses of those loads are simple. Rewrite them to use loads
1392 // of the per-field globals instead.
1393 for (auto UI = GV->user_begin(), E = GV->user_end(); UI != E;) {
1394 Instruction *User = cast<Instruction>(*UI++);
1396 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1397 RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite);
1401 // Must be a store of null.
1402 StoreInst *SI = cast<StoreInst>(User);
1403 assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
1404 "Unexpected heap-sra user!");
1406 // Insert a store of null into each global.
1407 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1408 PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType());
1409 Constant *Null = Constant::getNullValue(PT->getElementType());
1410 new StoreInst(Null, FieldGlobals[i], SI);
1412 // Erase the original store.
1413 SI->eraseFromParent();
1416 // While we have PHIs that are interesting to rewrite, do it.
1417 while (!PHIsToRewrite.empty()) {
1418 PHINode *PN = PHIsToRewrite.back().first;
1419 unsigned FieldNo = PHIsToRewrite.back().second;
1420 PHIsToRewrite.pop_back();
1421 PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]);
1422 assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi");
1424 // Add all the incoming values. This can materialize more phis.
1425 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1426 Value *InVal = PN->getIncomingValue(i);
1427 InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues,
1429 FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
1433 // Drop all inter-phi links and any loads that made it this far.
1434 for (DenseMap<Value*, std::vector<Value*> >::iterator
1435 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1437 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1438 PN->dropAllReferences();
1439 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1440 LI->dropAllReferences();
1443 // Delete all the phis and loads now that inter-references are dead.
1444 for (DenseMap<Value*, std::vector<Value*> >::iterator
1445 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1447 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1448 PN->eraseFromParent();
1449 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1450 LI->eraseFromParent();
1453 // The old global is now dead, remove it.
1454 GV->eraseFromParent();
1457 return cast<GlobalVariable>(FieldGlobals[0]);
1460 /// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1461 /// pointer global variable with a single value stored it that is a malloc or
1463 static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, CallInst *CI,
1465 AtomicOrdering Ordering,
1466 Module::global_iterator &GVI,
1467 const DataLayout &DL,
1468 TargetLibraryInfo *TLI) {
1469 // If this is a malloc of an abstract type, don't touch it.
1470 if (!AllocTy->isSized())
1473 // We can't optimize this global unless all uses of it are *known* to be
1474 // of the malloc value, not of the null initializer value (consider a use
1475 // that compares the global's value against zero to see if the malloc has
1476 // been reached). To do this, we check to see if all uses of the global
1477 // would trap if the global were null: this proves that they must all
1478 // happen after the malloc.
1479 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1482 // We can't optimize this if the malloc itself is used in a complex way,
1483 // for example, being stored into multiple globals. This allows the
1484 // malloc to be stored into the specified global, loaded icmp'd, and
1485 // GEP'd. These are all things we could transform to using the global
1487 SmallPtrSet<const PHINode*, 8> PHIs;
1488 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV, PHIs))
1491 // If we have a global that is only initialized with a fixed size malloc,
1492 // transform the program to use global memory instead of malloc'd memory.
1493 // This eliminates dynamic allocation, avoids an indirection accessing the
1494 // data, and exposes the resultant global to further GlobalOpt.
1495 // We cannot optimize the malloc if we cannot determine malloc array size.
1496 Value *NElems = getMallocArraySize(CI, DL, TLI, true);
1500 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
1501 // Restrict this transformation to only working on small allocations
1502 // (2048 bytes currently), as we don't want to introduce a 16M global or
1504 if (NElements->getZExtValue() * DL.getTypeAllocSize(AllocTy) < 2048) {
1505 GVI = OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, DL, TLI)
1510 // If the allocation is an array of structures, consider transforming this
1511 // into multiple malloc'd arrays, one for each field. This is basically
1512 // SRoA for malloc'd memory.
1514 if (Ordering != NotAtomic)
1517 // If this is an allocation of a fixed size array of structs, analyze as a
1518 // variable size array. malloc [100 x struct],1 -> malloc struct, 100
1519 if (NElems == ConstantInt::get(CI->getArgOperand(0)->getType(), 1))
1520 if (ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
1521 AllocTy = AT->getElementType();
1523 StructType *AllocSTy = dyn_cast<StructType>(AllocTy);
1527 // This the structure has an unreasonable number of fields, leave it
1529 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
1530 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, CI)) {
1532 // If this is a fixed size array, transform the Malloc to be an alloc of
1533 // structs. malloc [100 x struct],1 -> malloc struct, 100
1534 if (ArrayType *AT = dyn_cast<ArrayType>(getMallocAllocatedType(CI, TLI))) {
1535 Type *IntPtrTy = DL.getIntPtrType(CI->getType());
1536 unsigned TypeSize = DL.getStructLayout(AllocSTy)->getSizeInBytes();
1537 Value *AllocSize = ConstantInt::get(IntPtrTy, TypeSize);
1538 Value *NumElements = ConstantInt::get(IntPtrTy, AT->getNumElements());
1539 Instruction *Malloc = CallInst::CreateMalloc(CI, IntPtrTy, AllocSTy,
1540 AllocSize, NumElements,
1541 nullptr, CI->getName());
1542 Instruction *Cast = new BitCastInst(Malloc, CI->getType(), "tmp", CI);
1543 CI->replaceAllUsesWith(Cast);
1544 CI->eraseFromParent();
1545 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Malloc))
1546 CI = cast<CallInst>(BCI->getOperand(0));
1548 CI = cast<CallInst>(Malloc);
1551 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 << "\n");
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->getIterator(), 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 << "\n");
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 auto &DL = GV->getParent()->getDataLayout();
1736 // If this is a first class global and has only one accessing function
1737 // and this function is main (which we know is not recursive), we replace
1738 // the global with a local alloca in this function.
1740 // NOTE: It doesn't make sense to promote non-single-value types since we
1741 // are just replacing static memory to stack memory.
1743 // If the global is in different address space, don't bring it to stack.
1744 if (!GS.HasMultipleAccessingFunctions &&
1745 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1746 GV->getType()->getElementType()->isSingleValueType() &&
1747 GS.AccessingFunction->getName() == "main" &&
1748 GS.AccessingFunction->hasExternalLinkage() &&
1749 GV->getType()->getAddressSpace() == 0) {
1750 DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV);
1751 Instruction &FirstI = const_cast<Instruction&>(*GS.AccessingFunction
1752 ->getEntryBlock().begin());
1753 Type *ElemTy = GV->getType()->getElementType();
1754 // FIXME: Pass Global's alignment when globals have alignment
1755 AllocaInst *Alloca = new AllocaInst(ElemTy, nullptr,
1756 GV->getName(), &FirstI);
1757 if (!isa<UndefValue>(GV->getInitializer()))
1758 new StoreInst(GV->getInitializer(), Alloca, &FirstI);
1760 GV->replaceAllUsesWith(Alloca);
1761 GV->eraseFromParent();
1766 // If the global is never loaded (but may be stored to), it is dead.
1769 DEBUG(dbgs() << "GLOBAL NEVER LOADED: " << *GV);
1772 if (isLeakCheckerRoot(GV)) {
1773 // Delete any constant stores to the global.
1774 Changed = CleanupPointerRootUsers(GV, TLI);
1776 // Delete any stores we can find to the global. We may not be able to
1777 // make it completely dead though.
1778 Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, TLI);
1781 // If the global is dead now, delete it.
1782 if (GV->use_empty()) {
1783 GV->eraseFromParent();
1789 } else if (GS.StoredType <= GlobalStatus::InitializerStored) {
1790 DEBUG(dbgs() << "MARKING CONSTANT: " << *GV << "\n");
1791 GV->setConstant(true);
1793 // Clean up any obviously simplifiable users now.
1794 CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, TLI);
1796 // If the global is dead now, just nuke it.
1797 if (GV->use_empty()) {
1798 DEBUG(dbgs() << " *** Marking constant allowed us to simplify "
1799 << "all users and delete global!\n");
1800 GV->eraseFromParent();
1806 } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
1807 const DataLayout &DL = GV->getParent()->getDataLayout();
1808 if (GlobalVariable *FirstNewGV = SRAGlobal(GV, DL)) {
1809 GVI = FirstNewGV->getIterator(); // Don't skip the newly produced globals!
1812 } else if (GS.StoredType == GlobalStatus::StoredOnce && GS.StoredOnceValue) {
1813 // If the initial value for the global was an undef value, and if only
1814 // one other value was stored into it, we can just change the
1815 // initializer to be the stored value, then delete all stores to the
1816 // global. This allows us to mark it constant.
1817 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1818 if (isa<UndefValue>(GV->getInitializer())) {
1819 // Change the initial value here.
1820 GV->setInitializer(SOVConstant);
1822 // Clean up any obviously simplifiable users now.
1823 CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, TLI);
1825 if (GV->use_empty()) {
1826 DEBUG(dbgs() << " *** Substituting initializer allowed us to "
1827 << "simplify all users and delete global!\n");
1828 GV->eraseFromParent();
1831 GVI = GV->getIterator();
1837 // Try to optimize globals based on the knowledge that only one value
1838 // (besides its initializer) is ever stored to the global.
1839 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GS.Ordering, GVI,
1843 // Otherwise, if the global was not a boolean, we can shrink it to be a
1845 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) {
1846 if (GS.Ordering == NotAtomic) {
1847 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1858 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1859 /// function, changing them to FastCC.
1860 static void ChangeCalleesToFastCall(Function *F) {
1861 for (User *U : F->users()) {
1862 if (isa<BlockAddress>(U))
1864 CallSite CS(cast<Instruction>(U));
1865 CS.setCallingConv(CallingConv::Fast);
1869 static AttributeSet StripNest(LLVMContext &C, const AttributeSet &Attrs) {
1870 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1871 unsigned Index = Attrs.getSlotIndex(i);
1872 if (!Attrs.getSlotAttributes(i).hasAttribute(Index, Attribute::Nest))
1875 // There can be only one.
1876 return Attrs.removeAttribute(C, Index, Attribute::Nest);
1882 static void RemoveNestAttribute(Function *F) {
1883 F->setAttributes(StripNest(F->getContext(), F->getAttributes()));
1884 for (User *U : F->users()) {
1885 if (isa<BlockAddress>(U))
1887 CallSite CS(cast<Instruction>(U));
1888 CS.setAttributes(StripNest(F->getContext(), CS.getAttributes()));
1892 /// Return true if this is a calling convention that we'd like to change. The
1893 /// idea here is that we don't want to mess with the convention if the user
1894 /// explicitly requested something with performance implications like coldcc,
1895 /// GHC, or anyregcc.
1896 static bool isProfitableToMakeFastCC(Function *F) {
1897 CallingConv::ID CC = F->getCallingConv();
1898 // FIXME: Is it worth transforming x86_stdcallcc and x86_fastcallcc?
1899 return CC == CallingConv::C || CC == CallingConv::X86_ThisCall;
1902 bool GlobalOpt::OptimizeFunctions(Module &M) {
1903 bool Changed = false;
1904 // Optimize functions.
1905 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1906 Function *F = &*FI++;
1907 // Functions without names cannot be referenced outside this module.
1908 if (!F->hasName() && !F->isDeclaration() && !F->hasLocalLinkage())
1909 F->setLinkage(GlobalValue::InternalLinkage);
1911 const Comdat *C = F->getComdat();
1912 bool inComdat = C && NotDiscardableComdats.count(C);
1913 F->removeDeadConstantUsers();
1914 if ((!inComdat || F->hasLocalLinkage()) && F->isDefTriviallyDead()) {
1915 F->eraseFromParent();
1918 } else if (F->hasLocalLinkage()) {
1919 if (isProfitableToMakeFastCC(F) && !F->isVarArg() &&
1920 !F->hasAddressTaken()) {
1921 // If this function has a calling convention worth changing, is not a
1922 // varargs function, and is only called directly, promote it to use the
1923 // Fast calling convention.
1924 F->setCallingConv(CallingConv::Fast);
1925 ChangeCalleesToFastCall(F);
1930 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
1931 !F->hasAddressTaken()) {
1932 // The function is not used by a trampoline intrinsic, so it is safe
1933 // to remove the 'nest' attribute.
1934 RemoveNestAttribute(F);
1943 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1944 bool Changed = false;
1946 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1948 GlobalVariable *GV = &*GVI++;
1949 // Global variables without names cannot be referenced outside this module.
1950 if (!GV->hasName() && !GV->isDeclaration() && !GV->hasLocalLinkage())
1951 GV->setLinkage(GlobalValue::InternalLinkage);
1952 // Simplify the initializer.
1953 if (GV->hasInitializer())
1954 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GV->getInitializer())) {
1955 auto &DL = M.getDataLayout();
1956 Constant *New = ConstantFoldConstantExpression(CE, DL, TLI);
1957 if (New && New != CE)
1958 GV->setInitializer(New);
1961 if (GV->isDiscardableIfUnused()) {
1962 if (const Comdat *C = GV->getComdat())
1963 if (NotDiscardableComdats.count(C) && !GV->hasLocalLinkage())
1965 Changed |= ProcessGlobal(GV, GVI);
1972 isSimpleEnoughValueToCommit(Constant *C,
1973 SmallPtrSetImpl<Constant *> &SimpleConstants,
1974 const DataLayout &DL);
1976 /// isSimpleEnoughValueToCommit - Return true if the specified constant can be
1977 /// handled by the code generator. We don't want to generate something like:
1978 /// void *X = &X/42;
1979 /// because the code generator doesn't have a relocation that can handle that.
1981 /// This function should be called if C was not found (but just got inserted)
1982 /// in SimpleConstants to avoid having to rescan the same constants all the
1985 isSimpleEnoughValueToCommitHelper(Constant *C,
1986 SmallPtrSetImpl<Constant *> &SimpleConstants,
1987 const DataLayout &DL) {
1988 // Simple global addresses are supported, do not allow dllimport or
1989 // thread-local globals.
1990 if (auto *GV = dyn_cast<GlobalValue>(C))
1991 return !GV->hasDLLImportStorageClass() && !GV->isThreadLocal();
1993 // Simple integer, undef, constant aggregate zero, etc are all supported.
1994 if (C->getNumOperands() == 0 || isa<BlockAddress>(C))
1997 // Aggregate values are safe if all their elements are.
1998 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C) ||
1999 isa<ConstantVector>(C)) {
2000 for (Value *Op : C->operands())
2001 if (!isSimpleEnoughValueToCommit(cast<Constant>(Op), SimpleConstants, DL))
2006 // We don't know exactly what relocations are allowed in constant expressions,
2007 // so we allow &global+constantoffset, which is safe and uniformly supported
2009 ConstantExpr *CE = cast<ConstantExpr>(C);
2010 switch (CE->getOpcode()) {
2011 case Instruction::BitCast:
2012 // Bitcast is fine if the casted value is fine.
2013 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL);
2015 case Instruction::IntToPtr:
2016 case Instruction::PtrToInt:
2017 // int <=> ptr is fine if the int type is the same size as the
2019 if (DL.getTypeSizeInBits(CE->getType()) !=
2020 DL.getTypeSizeInBits(CE->getOperand(0)->getType()))
2022 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL);
2024 // GEP is fine if it is simple + constant offset.
2025 case Instruction::GetElementPtr:
2026 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
2027 if (!isa<ConstantInt>(CE->getOperand(i)))
2029 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL);
2031 case Instruction::Add:
2032 // We allow simple+cst.
2033 if (!isa<ConstantInt>(CE->getOperand(1)))
2035 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL);
2041 isSimpleEnoughValueToCommit(Constant *C,
2042 SmallPtrSetImpl<Constant *> &SimpleConstants,
2043 const DataLayout &DL) {
2044 // If we already checked this constant, we win.
2045 if (!SimpleConstants.insert(C).second)
2047 // Check the constant.
2048 return isSimpleEnoughValueToCommitHelper(C, SimpleConstants, DL);
2052 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
2053 /// enough for us to understand. In particular, if it is a cast to anything
2054 /// other than from one pointer type to another pointer type, we punt.
2055 /// We basically just support direct accesses to globals and GEP's of
2056 /// globals. This should be kept up to date with CommitValueTo.
2057 static bool isSimpleEnoughPointerToCommit(Constant *C) {
2058 // Conservatively, avoid aggregate types. This is because we don't
2059 // want to worry about them partially overlapping other stores.
2060 if (!cast<PointerType>(C->getType())->getElementType()->isSingleValueType())
2063 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
2064 // Do not allow weak/*_odr/linkonce linkage or external globals.
2065 return GV->hasUniqueInitializer();
2067 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2068 // Handle a constantexpr gep.
2069 if (CE->getOpcode() == Instruction::GetElementPtr &&
2070 isa<GlobalVariable>(CE->getOperand(0)) &&
2071 cast<GEPOperator>(CE)->isInBounds()) {
2072 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2073 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2074 // external globals.
2075 if (!GV->hasUniqueInitializer())
2078 // The first index must be zero.
2079 ConstantInt *CI = dyn_cast<ConstantInt>(*std::next(CE->op_begin()));
2080 if (!CI || !CI->isZero()) return false;
2082 // The remaining indices must be compile-time known integers within the
2083 // notional bounds of the corresponding static array types.
2084 if (!CE->isGEPWithNoNotionalOverIndexing())
2087 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2089 // A constantexpr bitcast from a pointer to another pointer is a no-op,
2090 // and we know how to evaluate it by moving the bitcast from the pointer
2091 // operand to the value operand.
2092 } else if (CE->getOpcode() == Instruction::BitCast &&
2093 isa<GlobalVariable>(CE->getOperand(0))) {
2094 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2095 // external globals.
2096 return cast<GlobalVariable>(CE->getOperand(0))->hasUniqueInitializer();
2103 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
2104 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
2105 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
2106 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
2107 ConstantExpr *Addr, unsigned OpNo) {
2108 // Base case of the recursion.
2109 if (OpNo == Addr->getNumOperands()) {
2110 assert(Val->getType() == Init->getType() && "Type mismatch!");
2114 SmallVector<Constant*, 32> Elts;
2115 if (StructType *STy = dyn_cast<StructType>(Init->getType())) {
2116 // Break up the constant into its elements.
2117 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2118 Elts.push_back(Init->getAggregateElement(i));
2120 // Replace the element that we are supposed to.
2121 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
2122 unsigned Idx = CU->getZExtValue();
2123 assert(Idx < STy->getNumElements() && "Struct index out of range!");
2124 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
2126 // Return the modified struct.
2127 return ConstantStruct::get(STy, Elts);
2130 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
2131 SequentialType *InitTy = cast<SequentialType>(Init->getType());
2134 if (ArrayType *ATy = dyn_cast<ArrayType>(InitTy))
2135 NumElts = ATy->getNumElements();
2137 NumElts = InitTy->getVectorNumElements();
2139 // Break up the array into elements.
2140 for (uint64_t i = 0, e = NumElts; i != e; ++i)
2141 Elts.push_back(Init->getAggregateElement(i));
2143 assert(CI->getZExtValue() < NumElts);
2144 Elts[CI->getZExtValue()] =
2145 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
2147 if (Init->getType()->isArrayTy())
2148 return ConstantArray::get(cast<ArrayType>(InitTy), Elts);
2149 return ConstantVector::get(Elts);
2152 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
2153 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
2154 static void CommitValueTo(Constant *Val, Constant *Addr) {
2155 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
2156 assert(GV->hasInitializer());
2157 GV->setInitializer(Val);
2161 ConstantExpr *CE = cast<ConstantExpr>(Addr);
2162 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2163 GV->setInitializer(EvaluateStoreInto(GV->getInitializer(), Val, CE, 2));
2168 /// Evaluator - This class evaluates LLVM IR, producing the Constant
2169 /// representing each SSA instruction. Changes to global variables are stored
2170 /// in a mapping that can be iterated over after the evaluation is complete.
2171 /// Once an evaluation call fails, the evaluation object should not be reused.
2174 Evaluator(const DataLayout &DL, const TargetLibraryInfo *TLI)
2175 : DL(DL), TLI(TLI) {
2176 ValueStack.emplace_back();
2180 for (auto &Tmp : AllocaTmps)
2181 // If there are still users of the alloca, the program is doing something
2182 // silly, e.g. storing the address of the alloca somewhere and using it
2183 // later. Since this is undefined, we'll just make it be null.
2184 if (!Tmp->use_empty())
2185 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2188 /// EvaluateFunction - Evaluate a call to function F, returning true if
2189 /// successful, false if we can't evaluate it. ActualArgs contains the formal
2190 /// arguments for the function.
2191 bool EvaluateFunction(Function *F, Constant *&RetVal,
2192 const SmallVectorImpl<Constant*> &ActualArgs);
2194 /// EvaluateBlock - Evaluate all instructions in block BB, returning true if
2195 /// successful, false if we can't evaluate it. NewBB returns the next BB that
2196 /// control flows into, or null upon return.
2197 bool EvaluateBlock(BasicBlock::iterator CurInst, BasicBlock *&NextBB);
2199 Constant *getVal(Value *V) {
2200 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
2201 Constant *R = ValueStack.back().lookup(V);
2202 assert(R && "Reference to an uncomputed value!");
2206 void setVal(Value *V, Constant *C) {
2207 ValueStack.back()[V] = C;
2210 const DenseMap<Constant*, Constant*> &getMutatedMemory() const {
2211 return MutatedMemory;
2214 const SmallPtrSetImpl<GlobalVariable*> &getInvariants() const {
2219 Constant *ComputeLoadResult(Constant *P);
2221 /// ValueStack - As we compute SSA register values, we store their contents
2222 /// here. The back of the deque contains the current function and the stack
2223 /// contains the values in the calling frames.
2224 std::deque<DenseMap<Value*, Constant*>> ValueStack;
2226 /// CallStack - This is used to detect recursion. In pathological situations
2227 /// we could hit exponential behavior, but at least there is nothing
2229 SmallVector<Function*, 4> CallStack;
2231 /// MutatedMemory - For each store we execute, we update this map. Loads
2232 /// check this to get the most up-to-date value. If evaluation is successful,
2233 /// this state is committed to the process.
2234 DenseMap<Constant*, Constant*> MutatedMemory;
2236 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2237 /// to represent its body. This vector is needed so we can delete the
2238 /// temporary globals when we are done.
2239 SmallVector<std::unique_ptr<GlobalVariable>, 32> AllocaTmps;
2241 /// Invariants - These global variables have been marked invariant by the
2242 /// static constructor.
2243 SmallPtrSet<GlobalVariable*, 8> Invariants;
2245 /// SimpleConstants - These are constants we have checked and know to be
2246 /// simple enough to live in a static initializer of a global.
2247 SmallPtrSet<Constant*, 8> SimpleConstants;
2249 const DataLayout &DL;
2250 const TargetLibraryInfo *TLI;
2253 } // anonymous namespace
2255 /// ComputeLoadResult - Return the value that would be computed by a load from
2256 /// P after the stores reflected by 'memory' have been performed. If we can't
2257 /// decide, return null.
2258 Constant *Evaluator::ComputeLoadResult(Constant *P) {
2259 // If this memory location has been recently stored, use the stored value: it
2260 // is the most up-to-date.
2261 DenseMap<Constant*, Constant*>::const_iterator I = MutatedMemory.find(P);
2262 if (I != MutatedMemory.end()) return I->second;
2265 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
2266 if (GV->hasDefinitiveInitializer())
2267 return GV->getInitializer();
2271 // Handle a constantexpr getelementptr.
2272 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
2273 if (CE->getOpcode() == Instruction::GetElementPtr &&
2274 isa<GlobalVariable>(CE->getOperand(0))) {
2275 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2276 if (GV->hasDefinitiveInitializer())
2277 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2280 return nullptr; // don't know how to evaluate.
2283 /// EvaluateBlock - Evaluate all instructions in block BB, returning true if
2284 /// successful, false if we can't evaluate it. NewBB returns the next BB that
2285 /// control flows into, or null upon return.
2286 bool Evaluator::EvaluateBlock(BasicBlock::iterator CurInst,
2287 BasicBlock *&NextBB) {
2288 // This is the main evaluation loop.
2290 Constant *InstResult = nullptr;
2292 DEBUG(dbgs() << "Evaluating Instruction: " << *CurInst << "\n");
2294 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
2295 if (!SI->isSimple()) {
2296 DEBUG(dbgs() << "Store is not simple! Can not evaluate.\n");
2297 return false; // no volatile/atomic accesses.
2299 Constant *Ptr = getVal(SI->getOperand(1));
2300 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
2301 DEBUG(dbgs() << "Folding constant ptr expression: " << *Ptr);
2302 Ptr = ConstantFoldConstantExpression(CE, DL, TLI);
2303 DEBUG(dbgs() << "; To: " << *Ptr << "\n");
2305 if (!isSimpleEnoughPointerToCommit(Ptr)) {
2306 // If this is too complex for us to commit, reject it.
2307 DEBUG(dbgs() << "Pointer is too complex for us to evaluate store.");
2311 Constant *Val = getVal(SI->getOperand(0));
2313 // If this might be too difficult for the backend to handle (e.g. the addr
2314 // of one global variable divided by another) then we can't commit it.
2315 if (!isSimpleEnoughValueToCommit(Val, SimpleConstants, DL)) {
2316 DEBUG(dbgs() << "Store value is too complex to evaluate store. " << *Val
2321 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
2322 if (CE->getOpcode() == Instruction::BitCast) {
2323 DEBUG(dbgs() << "Attempting to resolve bitcast on constant ptr.\n");
2324 // If we're evaluating a store through a bitcast, then we need
2325 // to pull the bitcast off the pointer type and push it onto the
2327 Ptr = CE->getOperand(0);
2329 Type *NewTy = cast<PointerType>(Ptr->getType())->getElementType();
2331 // In order to push the bitcast onto the stored value, a bitcast
2332 // from NewTy to Val's type must be legal. If it's not, we can try
2333 // introspecting NewTy to find a legal conversion.
2334 while (!Val->getType()->canLosslesslyBitCastTo(NewTy)) {
2335 // If NewTy is a struct, we can convert the pointer to the struct
2336 // into a pointer to its first member.
2337 // FIXME: This could be extended to support arrays as well.
2338 if (StructType *STy = dyn_cast<StructType>(NewTy)) {
2339 NewTy = STy->getTypeAtIndex(0U);
2341 IntegerType *IdxTy = IntegerType::get(NewTy->getContext(), 32);
2342 Constant *IdxZero = ConstantInt::get(IdxTy, 0, false);
2343 Constant * const IdxList[] = {IdxZero, IdxZero};
2345 Ptr = ConstantExpr::getGetElementPtr(nullptr, Ptr, IdxList);
2346 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
2347 Ptr = ConstantFoldConstantExpression(CE, DL, TLI);
2349 // If we can't improve the situation by introspecting NewTy,
2350 // we have to give up.
2352 DEBUG(dbgs() << "Failed to bitcast constant ptr, can not "
2358 // If we found compatible types, go ahead and push the bitcast
2359 // onto the stored value.
2360 Val = ConstantExpr::getBitCast(Val, NewTy);
2362 DEBUG(dbgs() << "Evaluated bitcast: " << *Val << "\n");
2366 MutatedMemory[Ptr] = Val;
2367 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2368 InstResult = ConstantExpr::get(BO->getOpcode(),
2369 getVal(BO->getOperand(0)),
2370 getVal(BO->getOperand(1)));
2371 DEBUG(dbgs() << "Found a BinaryOperator! Simplifying: " << *InstResult
2373 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2374 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2375 getVal(CI->getOperand(0)),
2376 getVal(CI->getOperand(1)));
2377 DEBUG(dbgs() << "Found a CmpInst! Simplifying: " << *InstResult
2379 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2380 InstResult = ConstantExpr::getCast(CI->getOpcode(),
2381 getVal(CI->getOperand(0)),
2383 DEBUG(dbgs() << "Found a Cast! Simplifying: " << *InstResult
2385 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2386 InstResult = ConstantExpr::getSelect(getVal(SI->getOperand(0)),
2387 getVal(SI->getOperand(1)),
2388 getVal(SI->getOperand(2)));
2389 DEBUG(dbgs() << "Found a Select! Simplifying: " << *InstResult
2391 } else if (auto *EVI = dyn_cast<ExtractValueInst>(CurInst)) {
2392 InstResult = ConstantExpr::getExtractValue(
2393 getVal(EVI->getAggregateOperand()), EVI->getIndices());
2394 DEBUG(dbgs() << "Found an ExtractValueInst! Simplifying: " << *InstResult
2396 } else if (auto *IVI = dyn_cast<InsertValueInst>(CurInst)) {
2397 InstResult = ConstantExpr::getInsertValue(
2398 getVal(IVI->getAggregateOperand()),
2399 getVal(IVI->getInsertedValueOperand()), IVI->getIndices());
2400 DEBUG(dbgs() << "Found an InsertValueInst! Simplifying: " << *InstResult
2402 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2403 Constant *P = getVal(GEP->getOperand(0));
2404 SmallVector<Constant*, 8> GEPOps;
2405 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2407 GEPOps.push_back(getVal(*i));
2409 ConstantExpr::getGetElementPtr(GEP->getSourceElementType(), P, GEPOps,
2410 cast<GEPOperator>(GEP)->isInBounds());
2411 DEBUG(dbgs() << "Found a GEP! Simplifying: " << *InstResult
2413 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2415 if (!LI->isSimple()) {
2416 DEBUG(dbgs() << "Found a Load! Not a simple load, can not evaluate.\n");
2417 return false; // no volatile/atomic accesses.
2420 Constant *Ptr = getVal(LI->getOperand(0));
2421 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
2422 Ptr = ConstantFoldConstantExpression(CE, DL, TLI);
2423 DEBUG(dbgs() << "Found a constant pointer expression, constant "
2424 "folding: " << *Ptr << "\n");
2426 InstResult = ComputeLoadResult(Ptr);
2428 DEBUG(dbgs() << "Failed to compute load result. Can not evaluate load."
2430 return false; // Could not evaluate load.
2433 DEBUG(dbgs() << "Evaluated load: " << *InstResult << "\n");
2434 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2435 if (AI->isArrayAllocation()) {
2436 DEBUG(dbgs() << "Found an array alloca. Can not evaluate.\n");
2437 return false; // Cannot handle array allocs.
2439 Type *Ty = AI->getType()->getElementType();
2440 AllocaTmps.push_back(
2441 make_unique<GlobalVariable>(Ty, false, GlobalValue::InternalLinkage,
2442 UndefValue::get(Ty), AI->getName()));
2443 InstResult = AllocaTmps.back().get();
2444 DEBUG(dbgs() << "Found an alloca. Result: " << *InstResult << "\n");
2445 } else if (isa<CallInst>(CurInst) || isa<InvokeInst>(CurInst)) {
2446 CallSite CS(&*CurInst);
2448 // Debug info can safely be ignored here.
2449 if (isa<DbgInfoIntrinsic>(CS.getInstruction())) {
2450 DEBUG(dbgs() << "Ignoring debug info.\n");
2455 // Cannot handle inline asm.
2456 if (isa<InlineAsm>(CS.getCalledValue())) {
2457 DEBUG(dbgs() << "Found inline asm, can not evaluate.\n");
2461 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) {
2462 if (MemSetInst *MSI = dyn_cast<MemSetInst>(II)) {
2463 if (MSI->isVolatile()) {
2464 DEBUG(dbgs() << "Can not optimize a volatile memset " <<
2468 Constant *Ptr = getVal(MSI->getDest());
2469 Constant *Val = getVal(MSI->getValue());
2470 Constant *DestVal = ComputeLoadResult(getVal(Ptr));
2471 if (Val->isNullValue() && DestVal && DestVal->isNullValue()) {
2472 // This memset is a no-op.
2473 DEBUG(dbgs() << "Ignoring no-op memset.\n");
2479 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
2480 II->getIntrinsicID() == Intrinsic::lifetime_end) {
2481 DEBUG(dbgs() << "Ignoring lifetime intrinsic.\n");
2486 if (II->getIntrinsicID() == Intrinsic::invariant_start) {
2487 // We don't insert an entry into Values, as it doesn't have a
2488 // meaningful return value.
2489 if (!II->use_empty()) {
2490 DEBUG(dbgs() << "Found unused invariant_start. Can't evaluate.\n");
2493 ConstantInt *Size = cast<ConstantInt>(II->getArgOperand(0));
2494 Value *PtrArg = getVal(II->getArgOperand(1));
2495 Value *Ptr = PtrArg->stripPointerCasts();
2496 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) {
2497 Type *ElemTy = cast<PointerType>(GV->getType())->getElementType();
2498 if (!Size->isAllOnesValue() &&
2499 Size->getValue().getLimitedValue() >=
2500 DL.getTypeStoreSize(ElemTy)) {
2501 Invariants.insert(GV);
2502 DEBUG(dbgs() << "Found a global var that is an invariant: " << *GV
2505 DEBUG(dbgs() << "Found a global var, but can not treat it as an "
2509 // Continue even if we do nothing.
2512 } else if (II->getIntrinsicID() == Intrinsic::assume) {
2513 DEBUG(dbgs() << "Skipping assume intrinsic.\n");
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->front();
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).second)
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)->stripPointerCasts()->getName().compare(
2721 (*B)->stripPointerCasts()->getName());
2724 static void setUsedInitializer(GlobalVariable &V,
2725 const SmallPtrSet<GlobalValue *, 8> &Init) {
2727 V.eraseFromParent();
2731 // Type of pointer to the array of pointers.
2732 PointerType *Int8PtrTy = Type::getInt8PtrTy(V.getContext(), 0);
2734 SmallVector<llvm::Constant *, 8> UsedArray;
2735 for (GlobalValue *GV : Init) {
2737 = ConstantExpr::getPointerBitCastOrAddrSpaceCast(GV, Int8PtrTy);
2738 UsedArray.push_back(Cast);
2740 // Sort to get deterministic order.
2741 array_pod_sort(UsedArray.begin(), UsedArray.end(), compareNames);
2742 ArrayType *ATy = ArrayType::get(Int8PtrTy, UsedArray.size());
2744 Module *M = V.getParent();
2745 V.removeFromParent();
2746 GlobalVariable *NV =
2747 new GlobalVariable(*M, ATy, false, llvm::GlobalValue::AppendingLinkage,
2748 llvm::ConstantArray::get(ATy, UsedArray), "");
2750 NV->setSection("llvm.metadata");
2755 /// \brief An easy to access representation of llvm.used and llvm.compiler.used.
2757 SmallPtrSet<GlobalValue *, 8> Used;
2758 SmallPtrSet<GlobalValue *, 8> CompilerUsed;
2759 GlobalVariable *UsedV;
2760 GlobalVariable *CompilerUsedV;
2763 LLVMUsed(Module &M) {
2764 UsedV = collectUsedGlobalVariables(M, Used, false);
2765 CompilerUsedV = collectUsedGlobalVariables(M, CompilerUsed, true);
2767 typedef SmallPtrSet<GlobalValue *, 8>::iterator iterator;
2768 typedef iterator_range<iterator> used_iterator_range;
2769 iterator usedBegin() { return Used.begin(); }
2770 iterator usedEnd() { return Used.end(); }
2771 used_iterator_range used() {
2772 return used_iterator_range(usedBegin(), usedEnd());
2774 iterator compilerUsedBegin() { return CompilerUsed.begin(); }
2775 iterator compilerUsedEnd() { return CompilerUsed.end(); }
2776 used_iterator_range compilerUsed() {
2777 return used_iterator_range(compilerUsedBegin(), compilerUsedEnd());
2779 bool usedCount(GlobalValue *GV) const { return Used.count(GV); }
2780 bool compilerUsedCount(GlobalValue *GV) const {
2781 return CompilerUsed.count(GV);
2783 bool usedErase(GlobalValue *GV) { return Used.erase(GV); }
2784 bool compilerUsedErase(GlobalValue *GV) { return CompilerUsed.erase(GV); }
2785 bool usedInsert(GlobalValue *GV) { return Used.insert(GV).second; }
2786 bool compilerUsedInsert(GlobalValue *GV) {
2787 return CompilerUsed.insert(GV).second;
2790 void syncVariablesAndSets() {
2792 setUsedInitializer(*UsedV, Used);
2794 setUsedInitializer(*CompilerUsedV, CompilerUsed);
2799 static bool hasUseOtherThanLLVMUsed(GlobalAlias &GA, const LLVMUsed &U) {
2800 if (GA.use_empty()) // No use at all.
2803 assert((!U.usedCount(&GA) || !U.compilerUsedCount(&GA)) &&
2804 "We should have removed the duplicated "
2805 "element from llvm.compiler.used");
2806 if (!GA.hasOneUse())
2807 // Strictly more than one use. So at least one is not in llvm.used and
2808 // llvm.compiler.used.
2811 // Exactly one use. Check if it is in llvm.used or llvm.compiler.used.
2812 return !U.usedCount(&GA) && !U.compilerUsedCount(&GA);
2815 static bool hasMoreThanOneUseOtherThanLLVMUsed(GlobalValue &V,
2816 const LLVMUsed &U) {
2818 assert((!U.usedCount(&V) || !U.compilerUsedCount(&V)) &&
2819 "We should have removed the duplicated "
2820 "element from llvm.compiler.used");
2821 if (U.usedCount(&V) || U.compilerUsedCount(&V))
2823 return V.hasNUsesOrMore(N);
2826 static bool mayHaveOtherReferences(GlobalAlias &GA, const LLVMUsed &U) {
2827 if (!GA.hasLocalLinkage())
2830 return U.usedCount(&GA) || U.compilerUsedCount(&GA);
2833 static bool hasUsesToReplace(GlobalAlias &GA, const LLVMUsed &U,
2834 bool &RenameTarget) {
2835 RenameTarget = false;
2837 if (hasUseOtherThanLLVMUsed(GA, U))
2840 // If the alias is externally visible, we may still be able to simplify it.
2841 if (!mayHaveOtherReferences(GA, U))
2844 // If the aliasee has internal linkage, give it the name and linkage
2845 // of the alias, and delete the alias. This turns:
2846 // define internal ... @f(...)
2847 // @a = alias ... @f
2849 // define ... @a(...)
2850 Constant *Aliasee = GA.getAliasee();
2851 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2852 if (!Target->hasLocalLinkage())
2855 // Do not perform the transform if multiple aliases potentially target the
2856 // aliasee. This check also ensures that it is safe to replace the section
2857 // and other attributes of the aliasee with those of the alias.
2858 if (hasMoreThanOneUseOtherThanLLVMUsed(*Target, U))
2861 RenameTarget = true;
2865 bool GlobalOpt::OptimizeGlobalAliases(Module &M) {
2866 bool Changed = false;
2869 for (GlobalValue *GV : Used.used())
2870 Used.compilerUsedErase(GV);
2872 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end();
2874 Module::alias_iterator J = I++;
2875 // Aliases without names cannot be referenced outside this module.
2876 if (!J->hasName() && !J->isDeclaration() && !J->hasLocalLinkage())
2877 J->setLinkage(GlobalValue::InternalLinkage);
2878 // If the aliasee may change at link time, nothing can be done - bail out.
2879 if (J->mayBeOverridden())
2882 Constant *Aliasee = J->getAliasee();
2883 GlobalValue *Target = dyn_cast<GlobalValue>(Aliasee->stripPointerCasts());
2884 // We can't trivially replace the alias with the aliasee if the aliasee is
2885 // non-trivial in some way.
2886 // TODO: Try to handle non-zero GEPs of local aliasees.
2889 Target->removeDeadConstantUsers();
2891 // Make all users of the alias use the aliasee instead.
2893 if (!hasUsesToReplace(*J, Used, RenameTarget))
2896 J->replaceAllUsesWith(ConstantExpr::getBitCast(Aliasee, J->getType()));
2897 ++NumAliasesResolved;
2901 // Give the aliasee the name, linkage and other attributes of the alias.
2902 Target->takeName(&*J);
2903 Target->setLinkage(J->getLinkage());
2904 Target->setVisibility(J->getVisibility());
2905 Target->setDLLStorageClass(J->getDLLStorageClass());
2907 if (Used.usedErase(&*J))
2908 Used.usedInsert(Target);
2910 if (Used.compilerUsedErase(&*J))
2911 Used.compilerUsedInsert(Target);
2912 } else if (mayHaveOtherReferences(*J, Used))
2915 // Delete the alias.
2916 M.getAliasList().erase(J);
2917 ++NumAliasesRemoved;
2921 Used.syncVariablesAndSets();
2926 static Function *FindCXAAtExit(Module &M, TargetLibraryInfo *TLI) {
2927 if (!TLI->has(LibFunc::cxa_atexit))
2930 Function *Fn = M.getFunction(TLI->getName(LibFunc::cxa_atexit));
2935 FunctionType *FTy = Fn->getFunctionType();
2937 // Checking that the function has the right return type, the right number of
2938 // parameters and that they all have pointer types should be enough.
2939 if (!FTy->getReturnType()->isIntegerTy() ||
2940 FTy->getNumParams() != 3 ||
2941 !FTy->getParamType(0)->isPointerTy() ||
2942 !FTy->getParamType(1)->isPointerTy() ||
2943 !FTy->getParamType(2)->isPointerTy())
2949 /// cxxDtorIsEmpty - Returns whether the given function is an empty C++
2950 /// destructor and can therefore be eliminated.
2951 /// Note that we assume that other optimization passes have already simplified
2952 /// the code so we only look for a function with a single basic block, where
2953 /// the only allowed instructions are 'ret', 'call' to an empty C++ dtor and
2954 /// other side-effect free instructions.
2955 static bool cxxDtorIsEmpty(const Function &Fn,
2956 SmallPtrSet<const Function *, 8> &CalledFunctions) {
2957 // FIXME: We could eliminate C++ destructors if they're readonly/readnone and
2958 // nounwind, but that doesn't seem worth doing.
2959 if (Fn.isDeclaration())
2962 if (++Fn.begin() != Fn.end())
2965 const BasicBlock &EntryBlock = Fn.getEntryBlock();
2966 for (BasicBlock::const_iterator I = EntryBlock.begin(), E = EntryBlock.end();
2968 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
2969 // Ignore debug intrinsics.
2970 if (isa<DbgInfoIntrinsic>(CI))
2973 const Function *CalledFn = CI->getCalledFunction();
2978 SmallPtrSet<const Function *, 8> NewCalledFunctions(CalledFunctions);
2980 // Don't treat recursive functions as empty.
2981 if (!NewCalledFunctions.insert(CalledFn).second)
2984 if (!cxxDtorIsEmpty(*CalledFn, NewCalledFunctions))
2986 } else if (isa<ReturnInst>(*I))
2987 return true; // We're done.
2988 else if (I->mayHaveSideEffects())
2989 return false; // Destructor with side effects, bail.
2995 bool GlobalOpt::OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) {
2996 /// Itanium C++ ABI p3.3.5:
2998 /// After constructing a global (or local static) object, that will require
2999 /// destruction on exit, a termination function is registered as follows:
3001 /// extern "C" int __cxa_atexit ( void (*f)(void *), void *p, void *d );
3003 /// This registration, e.g. __cxa_atexit(f,p,d), is intended to cause the
3004 /// call f(p) when DSO d is unloaded, before all such termination calls
3005 /// registered before this one. It returns zero if registration is
3006 /// successful, nonzero on failure.
3008 // This pass will look for calls to __cxa_atexit where the function is trivial
3010 bool Changed = false;
3012 for (auto I = CXAAtExitFn->user_begin(), E = CXAAtExitFn->user_end();
3014 // We're only interested in calls. Theoretically, we could handle invoke
3015 // instructions as well, but neither llvm-gcc nor clang generate invokes
3017 CallInst *CI = dyn_cast<CallInst>(*I++);
3022 dyn_cast<Function>(CI->getArgOperand(0)->stripPointerCasts());
3026 SmallPtrSet<const Function *, 8> CalledFunctions;
3027 if (!cxxDtorIsEmpty(*DtorFn, CalledFunctions))
3030 // Just remove the call.
3031 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
3032 CI->eraseFromParent();
3034 ++NumCXXDtorsRemoved;
3042 bool GlobalOpt::runOnModule(Module &M) {
3043 bool Changed = false;
3045 auto &DL = M.getDataLayout();
3046 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
3048 bool LocalChange = true;
3049 while (LocalChange) {
3050 LocalChange = false;
3052 NotDiscardableComdats.clear();
3053 for (const GlobalVariable &GV : M.globals())
3054 if (const Comdat *C = GV.getComdat())
3055 if (!GV.isDiscardableIfUnused() || !GV.use_empty())
3056 NotDiscardableComdats.insert(C);
3057 for (Function &F : M)
3058 if (const Comdat *C = F.getComdat())
3059 if (!F.isDefTriviallyDead())
3060 NotDiscardableComdats.insert(C);
3061 for (GlobalAlias &GA : M.aliases())
3062 if (const Comdat *C = GA.getComdat())
3063 if (!GA.isDiscardableIfUnused() || !GA.use_empty())
3064 NotDiscardableComdats.insert(C);
3066 // Delete functions that are trivially dead, ccc -> fastcc
3067 LocalChange |= OptimizeFunctions(M);
3069 // Optimize global_ctors list.
3070 LocalChange |= optimizeGlobalCtorsList(M, [&](Function *F) {
3071 return EvaluateStaticConstructor(F, DL, TLI);
3074 // Optimize non-address-taken globals.
3075 LocalChange |= OptimizeGlobalVars(M);
3077 // Resolve aliases, when possible.
3078 LocalChange |= OptimizeGlobalAliases(M);
3080 // Try to remove trivial global destructors if they are not removed
3082 Function *CXAAtExitFn = FindCXAAtExit(M, TLI);
3084 LocalChange |= OptimizeEmptyGlobalCXXDtors(CXAAtExitFn);
3086 Changed |= LocalChange;
3089 // TODO: Move all global ctors functions to the end of the module for code