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 Globals.insert(GV, NGV);
501 NewGlobals.push_back(NGV);
503 // Calculate the known alignment of the field. If the original aggregate
504 // had 256 byte alignment for example, something might depend on that:
505 // propagate info to each field.
506 uint64_t FieldOffset = Layout.getElementOffset(i);
507 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset);
508 if (NewAlign > DL.getABITypeAlignment(STy->getElementType(i)))
509 NGV->setAlignment(NewAlign);
511 } else if (SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
512 unsigned NumElements = 0;
513 if (ArrayType *ATy = dyn_cast<ArrayType>(STy))
514 NumElements = ATy->getNumElements();
516 NumElements = cast<VectorType>(STy)->getNumElements();
518 if (NumElements > 16 && GV->hasNUsesOrMore(16))
519 return nullptr; // It's not worth it.
520 NewGlobals.reserve(NumElements);
522 uint64_t EltSize = DL.getTypeAllocSize(STy->getElementType());
523 unsigned EltAlign = DL.getABITypeAlignment(STy->getElementType());
524 for (unsigned i = 0, e = NumElements; i != e; ++i) {
525 Constant *In = Init->getAggregateElement(i);
526 assert(In && "Couldn't get element of initializer?");
528 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
529 GlobalVariable::InternalLinkage,
530 In, GV->getName()+"."+Twine(i),
531 GV->getThreadLocalMode(),
532 GV->getType()->getAddressSpace());
533 Globals.insert(GV, NGV);
534 NewGlobals.push_back(NGV);
536 // Calculate the known alignment of the field. If the original aggregate
537 // had 256 byte alignment for example, something might depend on that:
538 // propagate info to each field.
539 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
540 if (NewAlign > EltAlign)
541 NGV->setAlignment(NewAlign);
545 if (NewGlobals.empty())
548 DEBUG(dbgs() << "PERFORMING GLOBAL SRA ON: " << *GV);
550 Constant *NullInt =Constant::getNullValue(Type::getInt32Ty(GV->getContext()));
552 // Loop over all of the uses of the global, replacing the constantexpr geps,
553 // with smaller constantexpr geps or direct references.
554 while (!GV->use_empty()) {
555 User *GEP = GV->user_back();
556 assert(((isa<ConstantExpr>(GEP) &&
557 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
558 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
560 // Ignore the 1th operand, which has to be zero or else the program is quite
561 // broken (undefined). Get the 2nd operand, which is the structure or array
563 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
564 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
566 Value *NewPtr = NewGlobals[Val];
568 // Form a shorter GEP if needed.
569 if (GEP->getNumOperands() > 3) {
570 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
571 SmallVector<Constant*, 8> Idxs;
572 Idxs.push_back(NullInt);
573 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
574 Idxs.push_back(CE->getOperand(i));
575 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr), Idxs);
577 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
578 SmallVector<Value*, 8> Idxs;
579 Idxs.push_back(NullInt);
580 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
581 Idxs.push_back(GEPI->getOperand(i));
582 NewPtr = GetElementPtrInst::Create(
583 NewPtr->getType()->getPointerElementType(), NewPtr, Idxs,
584 GEPI->getName() + "." + Twine(Val), GEPI);
587 GEP->replaceAllUsesWith(NewPtr);
589 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
590 GEPI->eraseFromParent();
592 cast<ConstantExpr>(GEP)->destroyConstant();
595 // Delete the old global, now that it is dead.
599 // Loop over the new globals array deleting any globals that are obviously
600 // dead. This can arise due to scalarization of a structure or an array that
601 // has elements that are dead.
602 unsigned FirstGlobal = 0;
603 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
604 if (NewGlobals[i]->use_empty()) {
605 Globals.erase(NewGlobals[i]);
606 if (FirstGlobal == i) ++FirstGlobal;
609 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : nullptr;
612 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
613 /// value will trap if the value is dynamically null. PHIs keeps track of any
614 /// phi nodes we've seen to avoid reprocessing them.
615 static bool AllUsesOfValueWillTrapIfNull(const Value *V,
616 SmallPtrSetImpl<const PHINode*> &PHIs) {
617 for (const User *U : V->users())
618 if (isa<LoadInst>(U)) {
620 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
621 if (SI->getOperand(0) == V) {
622 //cerr << "NONTRAPPING USE: " << *U;
623 return false; // Storing the value.
625 } else if (const CallInst *CI = dyn_cast<CallInst>(U)) {
626 if (CI->getCalledValue() != V) {
627 //cerr << "NONTRAPPING USE: " << *U;
628 return false; // Not calling the ptr
630 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(U)) {
631 if (II->getCalledValue() != V) {
632 //cerr << "NONTRAPPING USE: " << *U;
633 return false; // Not calling the ptr
635 } else if (const BitCastInst *CI = dyn_cast<BitCastInst>(U)) {
636 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
637 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
638 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
639 } else if (const PHINode *PN = dyn_cast<PHINode>(U)) {
640 // If we've already seen this phi node, ignore it, it has already been
642 if (PHIs.insert(PN).second && !AllUsesOfValueWillTrapIfNull(PN, PHIs))
644 } else if (isa<ICmpInst>(U) &&
645 isa<ConstantPointerNull>(U->getOperand(1))) {
646 // Ignore icmp X, null
648 //cerr << "NONTRAPPING USE: " << *U;
655 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
656 /// from GV will trap if the loaded value is null. Note that this also permits
657 /// comparisons of the loaded value against null, as a special case.
658 static bool AllUsesOfLoadedValueWillTrapIfNull(const GlobalVariable *GV) {
659 for (const User *U : GV->users())
660 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
661 SmallPtrSet<const PHINode*, 8> PHIs;
662 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
664 } else if (isa<StoreInst>(U)) {
665 // Ignore stores to the global.
667 // We don't know or understand this user, bail out.
668 //cerr << "UNKNOWN USER OF GLOBAL!: " << *U;
674 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
675 bool Changed = false;
676 for (auto UI = V->user_begin(), E = V->user_end(); UI != E; ) {
677 Instruction *I = cast<Instruction>(*UI++);
678 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
679 LI->setOperand(0, NewV);
681 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
682 if (SI->getOperand(1) == V) {
683 SI->setOperand(1, NewV);
686 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
688 if (CS.getCalledValue() == V) {
689 // Calling through the pointer! Turn into a direct call, but be careful
690 // that the pointer is not also being passed as an argument.
691 CS.setCalledFunction(NewV);
693 bool PassedAsArg = false;
694 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
695 if (CS.getArgument(i) == V) {
697 CS.setArgument(i, NewV);
701 // Being passed as an argument also. Be careful to not invalidate UI!
702 UI = V->user_begin();
705 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
706 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
707 ConstantExpr::getCast(CI->getOpcode(),
708 NewV, CI->getType()));
709 if (CI->use_empty()) {
711 CI->eraseFromParent();
713 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
714 // Should handle GEP here.
715 SmallVector<Constant*, 8> Idxs;
716 Idxs.reserve(GEPI->getNumOperands()-1);
717 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
719 if (Constant *C = dyn_cast<Constant>(*i))
723 if (Idxs.size() == GEPI->getNumOperands()-1)
724 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
725 ConstantExpr::getGetElementPtr(NewV, Idxs));
726 if (GEPI->use_empty()) {
728 GEPI->eraseFromParent();
737 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
738 /// value stored into it. If there are uses of the loaded value that would trap
739 /// if the loaded value is dynamically null, then we know that they cannot be
740 /// reachable with a null optimize away the load.
741 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV,
742 const DataLayout &DL,
743 TargetLibraryInfo *TLI) {
744 bool Changed = false;
746 // Keep track of whether we are able to remove all the uses of the global
747 // other than the store that defines it.
748 bool AllNonStoreUsesGone = true;
750 // Replace all uses of loads with uses of uses of the stored value.
751 for (Value::user_iterator GUI = GV->user_begin(), E = GV->user_end(); GUI != E;){
752 User *GlobalUser = *GUI++;
753 if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) {
754 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
755 // If we were able to delete all uses of the loads
756 if (LI->use_empty()) {
757 LI->eraseFromParent();
760 AllNonStoreUsesGone = false;
762 } else if (isa<StoreInst>(GlobalUser)) {
763 // Ignore the store that stores "LV" to the global.
764 assert(GlobalUser->getOperand(1) == GV &&
765 "Must be storing *to* the global");
767 AllNonStoreUsesGone = false;
769 // If we get here we could have other crazy uses that are transitively
771 assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) ||
772 isa<ConstantExpr>(GlobalUser) || isa<CmpInst>(GlobalUser) ||
773 isa<BitCastInst>(GlobalUser) ||
774 isa<GetElementPtrInst>(GlobalUser)) &&
775 "Only expect load and stores!");
780 DEBUG(dbgs() << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV);
784 // If we nuked all of the loads, then none of the stores are needed either,
785 // nor is the global.
786 if (AllNonStoreUsesGone) {
787 if (isLeakCheckerRoot(GV)) {
788 Changed |= CleanupPointerRootUsers(GV, TLI);
791 CleanupConstantGlobalUsers(GV, nullptr, DL, TLI);
793 if (GV->use_empty()) {
794 DEBUG(dbgs() << " *** GLOBAL NOW DEAD!\n");
796 GV->eraseFromParent();
803 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
804 /// instructions that are foldable.
805 static void ConstantPropUsersOf(Value *V, const DataLayout &DL,
806 TargetLibraryInfo *TLI) {
807 for (Value::user_iterator UI = V->user_begin(), E = V->user_end(); UI != E; )
808 if (Instruction *I = dyn_cast<Instruction>(*UI++))
809 if (Constant *NewC = ConstantFoldInstruction(I, DL, TLI)) {
810 I->replaceAllUsesWith(NewC);
812 // Advance UI to the next non-I use to avoid invalidating it!
813 // Instructions could multiply use V.
814 while (UI != E && *UI == I)
816 I->eraseFromParent();
820 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
821 /// variable, and transforms the program as if it always contained the result of
822 /// the specified malloc. Because it is always the result of the specified
823 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
824 /// malloc into a global, and any loads of GV as uses of the new global.
825 static GlobalVariable *
826 OptimizeGlobalAddressOfMalloc(GlobalVariable *GV, CallInst *CI, Type *AllocTy,
827 ConstantInt *NElements, const DataLayout &DL,
828 TargetLibraryInfo *TLI) {
829 DEBUG(errs() << "PROMOTING GLOBAL: " << *GV << " CALL = " << *CI << '\n');
832 if (NElements->getZExtValue() == 1)
833 GlobalType = AllocTy;
835 // If we have an array allocation, the global variable is of an array.
836 GlobalType = ArrayType::get(AllocTy, NElements->getZExtValue());
838 // Create the new global variable. The contents of the malloc'd memory is
839 // undefined, so initialize with an undef value.
840 GlobalVariable *NewGV = new GlobalVariable(*GV->getParent(),
842 GlobalValue::InternalLinkage,
843 UndefValue::get(GlobalType),
844 GV->getName()+".body",
846 GV->getThreadLocalMode());
848 // If there are bitcast users of the malloc (which is typical, usually we have
849 // a malloc + bitcast) then replace them with uses of the new global. Update
850 // other users to use the global as well.
851 BitCastInst *TheBC = nullptr;
852 while (!CI->use_empty()) {
853 Instruction *User = cast<Instruction>(CI->user_back());
854 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
855 if (BCI->getType() == NewGV->getType()) {
856 BCI->replaceAllUsesWith(NewGV);
857 BCI->eraseFromParent();
859 BCI->setOperand(0, NewGV);
863 TheBC = new BitCastInst(NewGV, CI->getType(), "newgv", CI);
864 User->replaceUsesOfWith(CI, TheBC);
868 Constant *RepValue = NewGV;
869 if (NewGV->getType() != GV->getType()->getElementType())
870 RepValue = ConstantExpr::getBitCast(RepValue,
871 GV->getType()->getElementType());
873 // If there is a comparison against null, we will insert a global bool to
874 // keep track of whether the global was initialized yet or not.
875 GlobalVariable *InitBool =
876 new GlobalVariable(Type::getInt1Ty(GV->getContext()), false,
877 GlobalValue::InternalLinkage,
878 ConstantInt::getFalse(GV->getContext()),
879 GV->getName()+".init", GV->getThreadLocalMode());
880 bool InitBoolUsed = false;
882 // Loop over all uses of GV, processing them in turn.
883 while (!GV->use_empty()) {
884 if (StoreInst *SI = dyn_cast<StoreInst>(GV->user_back())) {
885 // The global is initialized when the store to it occurs.
886 new StoreInst(ConstantInt::getTrue(GV->getContext()), InitBool, false, 0,
887 SI->getOrdering(), SI->getSynchScope(), SI);
888 SI->eraseFromParent();
892 LoadInst *LI = cast<LoadInst>(GV->user_back());
893 while (!LI->use_empty()) {
894 Use &LoadUse = *LI->use_begin();
895 ICmpInst *ICI = dyn_cast<ICmpInst>(LoadUse.getUser());
901 // Replace the cmp X, 0 with a use of the bool value.
902 // Sink the load to where the compare was, if atomic rules allow us to.
903 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", false, 0,
904 LI->getOrdering(), LI->getSynchScope(),
905 LI->isUnordered() ? (Instruction*)ICI : LI);
907 switch (ICI->getPredicate()) {
908 default: llvm_unreachable("Unknown ICmp Predicate!");
909 case ICmpInst::ICMP_ULT:
910 case ICmpInst::ICMP_SLT: // X < null -> always false
911 LV = ConstantInt::getFalse(GV->getContext());
913 case ICmpInst::ICMP_ULE:
914 case ICmpInst::ICMP_SLE:
915 case ICmpInst::ICMP_EQ:
916 LV = BinaryOperator::CreateNot(LV, "notinit", ICI);
918 case ICmpInst::ICMP_NE:
919 case ICmpInst::ICMP_UGE:
920 case ICmpInst::ICMP_SGE:
921 case ICmpInst::ICMP_UGT:
922 case ICmpInst::ICMP_SGT:
925 ICI->replaceAllUsesWith(LV);
926 ICI->eraseFromParent();
928 LI->eraseFromParent();
931 // If the initialization boolean was used, insert it, otherwise delete it.
933 while (!InitBool->use_empty()) // Delete initializations
934 cast<StoreInst>(InitBool->user_back())->eraseFromParent();
937 GV->getParent()->getGlobalList().insert(GV, InitBool);
939 // Now the GV is dead, nuke it and the malloc..
940 GV->eraseFromParent();
941 CI->eraseFromParent();
943 // To further other optimizations, loop over all users of NewGV and try to
944 // constant prop them. This will promote GEP instructions with constant
945 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
946 ConstantPropUsersOf(NewGV, DL, TLI);
947 if (RepValue != NewGV)
948 ConstantPropUsersOf(RepValue, DL, TLI);
953 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
954 /// to make sure that there are no complex uses of V. We permit simple things
955 /// like dereferencing the pointer, but not storing through the address, unless
956 /// it is to the specified global.
957 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(const Instruction *V,
958 const GlobalVariable *GV,
959 SmallPtrSetImpl<const PHINode*> &PHIs) {
960 for (const User *U : V->users()) {
961 const Instruction *Inst = cast<Instruction>(U);
963 if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) {
964 continue; // Fine, ignore.
967 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
968 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
969 return false; // Storing the pointer itself... bad.
970 continue; // Otherwise, storing through it, or storing into GV... fine.
973 // Must index into the array and into the struct.
974 if (isa<GetElementPtrInst>(Inst) && Inst->getNumOperands() >= 3) {
975 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs))
980 if (const PHINode *PN = dyn_cast<PHINode>(Inst)) {
981 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI
983 if (PHIs.insert(PN).second)
984 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
989 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) {
990 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs))
1000 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
1001 /// somewhere. Transform all uses of the allocation into loads from the
1002 /// global and uses of the resultant pointer. Further, delete the store into
1003 /// GV. This assumes that these value pass the
1004 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
1005 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
1006 GlobalVariable *GV) {
1007 while (!Alloc->use_empty()) {
1008 Instruction *U = cast<Instruction>(*Alloc->user_begin());
1009 Instruction *InsertPt = U;
1010 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1011 // If this is the store of the allocation into the global, remove it.
1012 if (SI->getOperand(1) == GV) {
1013 SI->eraseFromParent();
1016 } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
1017 // Insert the load in the corresponding predecessor, not right before the
1019 InsertPt = PN->getIncomingBlock(*Alloc->use_begin())->getTerminator();
1020 } else if (isa<BitCastInst>(U)) {
1021 // Must be bitcast between the malloc and store to initialize the global.
1022 ReplaceUsesOfMallocWithGlobal(U, GV);
1023 U->eraseFromParent();
1025 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
1026 // If this is a "GEP bitcast" and the user is a store to the global, then
1027 // just process it as a bitcast.
1028 if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse())
1029 if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->user_back()))
1030 if (SI->getOperand(1) == GV) {
1031 // Must be bitcast GEP between the malloc and store to initialize
1033 ReplaceUsesOfMallocWithGlobal(GEPI, GV);
1034 GEPI->eraseFromParent();
1039 // Insert a load from the global, and use it instead of the malloc.
1040 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
1041 U->replaceUsesOfWith(Alloc, NL);
1045 /// LoadUsesSimpleEnoughForHeapSRA - Verify that all uses of V (a load, or a phi
1046 /// of a load) are simple enough to perform heap SRA on. This permits GEP's
1047 /// that index through the array and struct field, icmps of null, and PHIs.
1048 static bool LoadUsesSimpleEnoughForHeapSRA(const Value *V,
1049 SmallPtrSetImpl<const PHINode*> &LoadUsingPHIs,
1050 SmallPtrSetImpl<const PHINode*> &LoadUsingPHIsPerLoad) {
1051 // We permit two users of the load: setcc comparing against the null
1052 // pointer, and a getelementptr of a specific form.
1053 for (const User *U : V->users()) {
1054 const Instruction *UI = cast<Instruction>(U);
1056 // Comparison against null is ok.
1057 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UI)) {
1058 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1063 // getelementptr is also ok, but only a simple form.
1064 if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(UI)) {
1065 // Must index into the array and into the struct.
1066 if (GEPI->getNumOperands() < 3)
1069 // Otherwise the GEP is ok.
1073 if (const PHINode *PN = dyn_cast<PHINode>(UI)) {
1074 if (!LoadUsingPHIsPerLoad.insert(PN).second)
1075 // This means some phi nodes are dependent on each other.
1076 // Avoid infinite looping!
1078 if (!LoadUsingPHIs.insert(PN).second)
1079 // If we have already analyzed this PHI, then it is safe.
1082 // Make sure all uses of the PHI are simple enough to transform.
1083 if (!LoadUsesSimpleEnoughForHeapSRA(PN,
1084 LoadUsingPHIs, LoadUsingPHIsPerLoad))
1090 // Otherwise we don't know what this is, not ok.
1098 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
1099 /// GV are simple enough to perform HeapSRA, return true.
1100 static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(const GlobalVariable *GV,
1101 Instruction *StoredVal) {
1102 SmallPtrSet<const PHINode*, 32> LoadUsingPHIs;
1103 SmallPtrSet<const PHINode*, 32> LoadUsingPHIsPerLoad;
1104 for (const User *U : GV->users())
1105 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
1106 if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs,
1107 LoadUsingPHIsPerLoad))
1109 LoadUsingPHIsPerLoad.clear();
1112 // If we reach here, we know that all uses of the loads and transitive uses
1113 // (through PHI nodes) are simple enough to transform. However, we don't know
1114 // that all inputs the to the PHI nodes are in the same equivalence sets.
1115 // Check to verify that all operands of the PHIs are either PHIS that can be
1116 // transformed, loads from GV, or MI itself.
1117 for (const PHINode *PN : LoadUsingPHIs) {
1118 for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) {
1119 Value *InVal = PN->getIncomingValue(op);
1121 // PHI of the stored value itself is ok.
1122 if (InVal == StoredVal) continue;
1124 if (const PHINode *InPN = dyn_cast<PHINode>(InVal)) {
1125 // One of the PHIs in our set is (optimistically) ok.
1126 if (LoadUsingPHIs.count(InPN))
1131 // Load from GV is ok.
1132 if (const LoadInst *LI = dyn_cast<LoadInst>(InVal))
1133 if (LI->getOperand(0) == GV)
1138 // Anything else is rejected.
1146 static Value *GetHeapSROAValue(Value *V, unsigned FieldNo,
1147 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1148 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1149 std::vector<Value*> &FieldVals = InsertedScalarizedValues[V];
1151 if (FieldNo >= FieldVals.size())
1152 FieldVals.resize(FieldNo+1);
1154 // If we already have this value, just reuse the previously scalarized
1156 if (Value *FieldVal = FieldVals[FieldNo])
1159 // Depending on what instruction this is, we have several cases.
1161 if (LoadInst *LI = dyn_cast<LoadInst>(V)) {
1162 // This is a scalarized version of the load from the global. Just create
1163 // a new Load of the scalarized global.
1164 Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo,
1165 InsertedScalarizedValues,
1167 LI->getName()+".f"+Twine(FieldNo), LI);
1169 PHINode *PN = cast<PHINode>(V);
1170 // PN's type is pointer to struct. Make a new PHI of pointer to struct
1173 PointerType *PTy = cast<PointerType>(PN->getType());
1174 StructType *ST = cast<StructType>(PTy->getElementType());
1176 unsigned AS = PTy->getAddressSpace();
1178 PHINode::Create(PointerType::get(ST->getElementType(FieldNo), AS),
1179 PN->getNumIncomingValues(),
1180 PN->getName()+".f"+Twine(FieldNo), PN);
1182 PHIsToRewrite.push_back(std::make_pair(PN, FieldNo));
1185 return FieldVals[FieldNo] = Result;
1188 /// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1189 /// the load, rewrite the derived value to use the HeapSRoA'd load.
1190 static void RewriteHeapSROALoadUser(Instruction *LoadUser,
1191 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1192 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1193 // If this is a comparison against null, handle it.
1194 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1195 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1196 // If we have a setcc of the loaded pointer, we can use a setcc of any
1198 Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0,
1199 InsertedScalarizedValues, PHIsToRewrite);
1201 Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr,
1202 Constant::getNullValue(NPtr->getType()),
1204 SCI->replaceAllUsesWith(New);
1205 SCI->eraseFromParent();
1209 // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...'
1210 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1211 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1212 && "Unexpected GEPI!");
1214 // Load the pointer for this field.
1215 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1216 Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo,
1217 InsertedScalarizedValues, PHIsToRewrite);
1219 // Create the new GEP idx vector.
1220 SmallVector<Value*, 8> GEPIdx;
1221 GEPIdx.push_back(GEPI->getOperand(1));
1222 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1224 Value *NGEPI = GetElementPtrInst::Create(GEPI->getResultElementType(), NewPtr, GEPIdx,
1225 GEPI->getName(), GEPI);
1226 GEPI->replaceAllUsesWith(NGEPI);
1227 GEPI->eraseFromParent();
1231 // Recursively transform the users of PHI nodes. This will lazily create the
1232 // PHIs that are needed for individual elements. Keep track of what PHIs we
1233 // see in InsertedScalarizedValues so that we don't get infinite loops (very
1234 // antisocial). If the PHI is already in InsertedScalarizedValues, it has
1235 // already been seen first by another load, so its uses have already been
1237 PHINode *PN = cast<PHINode>(LoadUser);
1238 if (!InsertedScalarizedValues.insert(std::make_pair(PN,
1239 std::vector<Value*>())).second)
1242 // If this is the first time we've seen this PHI, recursively process all
1244 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
1245 Instruction *User = cast<Instruction>(*UI++);
1246 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1250 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1251 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1252 /// use FieldGlobals instead. All uses of loaded values satisfy
1253 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA.
1254 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1255 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1256 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1257 for (auto UI = Load->user_begin(), E = Load->user_end(); UI != E;) {
1258 Instruction *User = cast<Instruction>(*UI++);
1259 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1262 if (Load->use_empty()) {
1263 Load->eraseFromParent();
1264 InsertedScalarizedValues.erase(Load);
1268 /// PerformHeapAllocSRoA - CI is an allocation of an array of structures. Break
1269 /// it up into multiple allocations of arrays of the fields.
1270 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI,
1271 Value *NElems, const DataLayout &DL,
1272 const TargetLibraryInfo *TLI) {
1273 DEBUG(dbgs() << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *CI << '\n');
1274 Type *MAT = getMallocAllocatedType(CI, TLI);
1275 StructType *STy = cast<StructType>(MAT);
1277 // There is guaranteed to be at least one use of the malloc (storing
1278 // it into GV). If there are other uses, change them to be uses of
1279 // the global to simplify later code. This also deletes the store
1281 ReplaceUsesOfMallocWithGlobal(CI, GV);
1283 // Okay, at this point, there are no users of the malloc. Insert N
1284 // new mallocs at the same place as CI, and N globals.
1285 std::vector<Value*> FieldGlobals;
1286 std::vector<Value*> FieldMallocs;
1288 unsigned AS = GV->getType()->getPointerAddressSpace();
1289 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1290 Type *FieldTy = STy->getElementType(FieldNo);
1291 PointerType *PFieldTy = PointerType::get(FieldTy, AS);
1293 GlobalVariable *NGV =
1294 new GlobalVariable(*GV->getParent(),
1295 PFieldTy, false, GlobalValue::InternalLinkage,
1296 Constant::getNullValue(PFieldTy),
1297 GV->getName() + ".f" + Twine(FieldNo), GV,
1298 GV->getThreadLocalMode());
1299 FieldGlobals.push_back(NGV);
1301 unsigned TypeSize = DL.getTypeAllocSize(FieldTy);
1302 if (StructType *ST = dyn_cast<StructType>(FieldTy))
1303 TypeSize = DL.getStructLayout(ST)->getSizeInBytes();
1304 Type *IntPtrTy = DL.getIntPtrType(CI->getType());
1305 Value *NMI = CallInst::CreateMalloc(CI, IntPtrTy, FieldTy,
1306 ConstantInt::get(IntPtrTy, TypeSize),
1308 CI->getName() + ".f" + Twine(FieldNo));
1309 FieldMallocs.push_back(NMI);
1310 new StoreInst(NMI, NGV, CI);
1313 // The tricky aspect of this transformation is handling the case when malloc
1314 // fails. In the original code, malloc failing would set the result pointer
1315 // of malloc to null. In this case, some mallocs could succeed and others
1316 // could fail. As such, we emit code that looks like this:
1317 // F0 = malloc(field0)
1318 // F1 = malloc(field1)
1319 // F2 = malloc(field2)
1320 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1321 // if (F0) { free(F0); F0 = 0; }
1322 // if (F1) { free(F1); F1 = 0; }
1323 // if (F2) { free(F2); F2 = 0; }
1325 // The malloc can also fail if its argument is too large.
1326 Constant *ConstantZero = ConstantInt::get(CI->getArgOperand(0)->getType(), 0);
1327 Value *RunningOr = new ICmpInst(CI, ICmpInst::ICMP_SLT, CI->getArgOperand(0),
1328 ConstantZero, "isneg");
1329 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1330 Value *Cond = new ICmpInst(CI, ICmpInst::ICMP_EQ, FieldMallocs[i],
1331 Constant::getNullValue(FieldMallocs[i]->getType()),
1333 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", CI);
1336 // Split the basic block at the old malloc.
1337 BasicBlock *OrigBB = CI->getParent();
1338 BasicBlock *ContBB = OrigBB->splitBasicBlock(CI, "malloc_cont");
1340 // Create the block to check the first condition. Put all these blocks at the
1341 // end of the function as they are unlikely to be executed.
1342 BasicBlock *NullPtrBlock = BasicBlock::Create(OrigBB->getContext(),
1344 OrigBB->getParent());
1346 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1347 // branch on RunningOr.
1348 OrigBB->getTerminator()->eraseFromParent();
1349 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1351 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1352 // pointer, because some may be null while others are not.
1353 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1354 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1355 Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
1356 Constant::getNullValue(GVVal->getType()));
1357 BasicBlock *FreeBlock = BasicBlock::Create(Cmp->getContext(), "free_it",
1358 OrigBB->getParent());
1359 BasicBlock *NextBlock = BasicBlock::Create(Cmp->getContext(), "next",
1360 OrigBB->getParent());
1361 Instruction *BI = BranchInst::Create(FreeBlock, NextBlock,
1364 // Fill in FreeBlock.
1365 CallInst::CreateFree(GVVal, BI);
1366 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1368 BranchInst::Create(NextBlock, FreeBlock);
1370 NullPtrBlock = NextBlock;
1373 BranchInst::Create(ContBB, NullPtrBlock);
1375 // CI is no longer needed, remove it.
1376 CI->eraseFromParent();
1378 /// InsertedScalarizedLoads - As we process loads, if we can't immediately
1379 /// update all uses of the load, keep track of what scalarized loads are
1380 /// inserted for a given load.
1381 DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues;
1382 InsertedScalarizedValues[GV] = FieldGlobals;
1384 std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite;
1386 // Okay, the malloc site is completely handled. All of the uses of GV are now
1387 // loads, and all uses of those loads are simple. Rewrite them to use loads
1388 // of the per-field globals instead.
1389 for (auto UI = GV->user_begin(), E = GV->user_end(); UI != E;) {
1390 Instruction *User = cast<Instruction>(*UI++);
1392 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1393 RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite);
1397 // Must be a store of null.
1398 StoreInst *SI = cast<StoreInst>(User);
1399 assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
1400 "Unexpected heap-sra user!");
1402 // Insert a store of null into each global.
1403 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1404 PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType());
1405 Constant *Null = Constant::getNullValue(PT->getElementType());
1406 new StoreInst(Null, FieldGlobals[i], SI);
1408 // Erase the original store.
1409 SI->eraseFromParent();
1412 // While we have PHIs that are interesting to rewrite, do it.
1413 while (!PHIsToRewrite.empty()) {
1414 PHINode *PN = PHIsToRewrite.back().first;
1415 unsigned FieldNo = PHIsToRewrite.back().second;
1416 PHIsToRewrite.pop_back();
1417 PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]);
1418 assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi");
1420 // Add all the incoming values. This can materialize more phis.
1421 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1422 Value *InVal = PN->getIncomingValue(i);
1423 InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues,
1425 FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
1429 // Drop all inter-phi links and any loads that made it this far.
1430 for (DenseMap<Value*, std::vector<Value*> >::iterator
1431 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1433 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1434 PN->dropAllReferences();
1435 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1436 LI->dropAllReferences();
1439 // Delete all the phis and loads now that inter-references are dead.
1440 for (DenseMap<Value*, std::vector<Value*> >::iterator
1441 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1443 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1444 PN->eraseFromParent();
1445 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1446 LI->eraseFromParent();
1449 // The old global is now dead, remove it.
1450 GV->eraseFromParent();
1453 return cast<GlobalVariable>(FieldGlobals[0]);
1456 /// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1457 /// pointer global variable with a single value stored it that is a malloc or
1459 static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, CallInst *CI,
1461 AtomicOrdering Ordering,
1462 Module::global_iterator &GVI,
1463 const DataLayout &DL,
1464 TargetLibraryInfo *TLI) {
1465 // If this is a malloc of an abstract type, don't touch it.
1466 if (!AllocTy->isSized())
1469 // We can't optimize this global unless all uses of it are *known* to be
1470 // of the malloc value, not of the null initializer value (consider a use
1471 // that compares the global's value against zero to see if the malloc has
1472 // been reached). To do this, we check to see if all uses of the global
1473 // would trap if the global were null: this proves that they must all
1474 // happen after the malloc.
1475 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1478 // We can't optimize this if the malloc itself is used in a complex way,
1479 // for example, being stored into multiple globals. This allows the
1480 // malloc to be stored into the specified global, loaded icmp'd, and
1481 // GEP'd. These are all things we could transform to using the global
1483 SmallPtrSet<const PHINode*, 8> PHIs;
1484 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV, PHIs))
1487 // If we have a global that is only initialized with a fixed size malloc,
1488 // transform the program to use global memory instead of malloc'd memory.
1489 // This eliminates dynamic allocation, avoids an indirection accessing the
1490 // data, and exposes the resultant global to further GlobalOpt.
1491 // We cannot optimize the malloc if we cannot determine malloc array size.
1492 Value *NElems = getMallocArraySize(CI, DL, TLI, true);
1496 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
1497 // Restrict this transformation to only working on small allocations
1498 // (2048 bytes currently), as we don't want to introduce a 16M global or
1500 if (NElements->getZExtValue() * DL.getTypeAllocSize(AllocTy) < 2048) {
1501 GVI = OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, DL, TLI);
1505 // If the allocation is an array of structures, consider transforming this
1506 // into multiple malloc'd arrays, one for each field. This is basically
1507 // SRoA for malloc'd memory.
1509 if (Ordering != NotAtomic)
1512 // If this is an allocation of a fixed size array of structs, analyze as a
1513 // variable size array. malloc [100 x struct],1 -> malloc struct, 100
1514 if (NElems == ConstantInt::get(CI->getArgOperand(0)->getType(), 1))
1515 if (ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
1516 AllocTy = AT->getElementType();
1518 StructType *AllocSTy = dyn_cast<StructType>(AllocTy);
1522 // This the structure has an unreasonable number of fields, leave it
1524 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
1525 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, CI)) {
1527 // If this is a fixed size array, transform the Malloc to be an alloc of
1528 // structs. malloc [100 x struct],1 -> malloc struct, 100
1529 if (ArrayType *AT = dyn_cast<ArrayType>(getMallocAllocatedType(CI, TLI))) {
1530 Type *IntPtrTy = DL.getIntPtrType(CI->getType());
1531 unsigned TypeSize = DL.getStructLayout(AllocSTy)->getSizeInBytes();
1532 Value *AllocSize = ConstantInt::get(IntPtrTy, TypeSize);
1533 Value *NumElements = ConstantInt::get(IntPtrTy, AT->getNumElements());
1534 Instruction *Malloc = CallInst::CreateMalloc(CI, IntPtrTy, AllocSTy,
1535 AllocSize, NumElements,
1536 nullptr, CI->getName());
1537 Instruction *Cast = new BitCastInst(Malloc, CI->getType(), "tmp", CI);
1538 CI->replaceAllUsesWith(Cast);
1539 CI->eraseFromParent();
1540 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Malloc))
1541 CI = cast<CallInst>(BCI->getOperand(0));
1543 CI = cast<CallInst>(Malloc);
1546 GVI = PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, DL, TLI, true),
1554 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1555 // that only one value (besides its initializer) is ever stored to the global.
1556 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1557 AtomicOrdering Ordering,
1558 Module::global_iterator &GVI,
1559 const DataLayout &DL,
1560 TargetLibraryInfo *TLI) {
1561 // Ignore no-op GEPs and bitcasts.
1562 StoredOnceVal = StoredOnceVal->stripPointerCasts();
1564 // If we are dealing with a pointer global that is initialized to null and
1565 // only has one (non-null) value stored into it, then we can optimize any
1566 // users of the loaded value (often calls and loads) that would trap if the
1568 if (GV->getInitializer()->getType()->isPointerTy() &&
1569 GV->getInitializer()->isNullValue()) {
1570 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1571 if (GV->getInitializer()->getType() != SOVC->getType())
1572 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1574 // Optimize away any trapping uses of the loaded value.
1575 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC, DL, TLI))
1577 } else if (CallInst *CI = extractMallocCall(StoredOnceVal, TLI)) {
1578 Type *MallocType = getMallocAllocatedType(CI, TLI);
1580 TryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType, Ordering, GVI,
1589 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1590 /// two values ever stored into GV are its initializer and OtherVal. See if we
1591 /// can shrink the global into a boolean and select between the two values
1592 /// whenever it is used. This exposes the values to other scalar optimizations.
1593 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1594 Type *GVElType = GV->getType()->getElementType();
1596 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1597 // an FP value, pointer or vector, don't do this optimization because a select
1598 // between them is very expensive and unlikely to lead to later
1599 // simplification. In these cases, we typically end up with "cond ? v1 : v2"
1600 // where v1 and v2 both require constant pool loads, a big loss.
1601 if (GVElType == Type::getInt1Ty(GV->getContext()) ||
1602 GVElType->isFloatingPointTy() ||
1603 GVElType->isPointerTy() || GVElType->isVectorTy())
1606 // Walk the use list of the global seeing if all the uses are load or store.
1607 // If there is anything else, bail out.
1608 for (User *U : GV->users())
1609 if (!isa<LoadInst>(U) && !isa<StoreInst>(U))
1612 DEBUG(dbgs() << " *** SHRINKING TO BOOL: " << *GV);
1614 // Create the new global, initializing it to false.
1615 GlobalVariable *NewGV = new GlobalVariable(Type::getInt1Ty(GV->getContext()),
1617 GlobalValue::InternalLinkage,
1618 ConstantInt::getFalse(GV->getContext()),
1620 GV->getThreadLocalMode(),
1621 GV->getType()->getAddressSpace());
1622 GV->getParent()->getGlobalList().insert(GV, NewGV);
1624 Constant *InitVal = GV->getInitializer();
1625 assert(InitVal->getType() != Type::getInt1Ty(GV->getContext()) &&
1626 "No reason to shrink to bool!");
1628 // If initialized to zero and storing one into the global, we can use a cast
1629 // instead of a select to synthesize the desired value.
1630 bool IsOneZero = false;
1631 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1632 IsOneZero = InitVal->isNullValue() && CI->isOne();
1634 while (!GV->use_empty()) {
1635 Instruction *UI = cast<Instruction>(GV->user_back());
1636 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1637 // Change the store into a boolean store.
1638 bool StoringOther = SI->getOperand(0) == OtherVal;
1639 // Only do this if we weren't storing a loaded value.
1641 if (StoringOther || SI->getOperand(0) == InitVal) {
1642 StoreVal = ConstantInt::get(Type::getInt1Ty(GV->getContext()),
1645 // Otherwise, we are storing a previously loaded copy. To do this,
1646 // change the copy from copying the original value to just copying the
1648 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1650 // If we've already replaced the input, StoredVal will be a cast or
1651 // select instruction. If not, it will be a load of the original
1653 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1654 assert(LI->getOperand(0) == GV && "Not a copy!");
1655 // Insert a new load, to preserve the saved value.
1656 StoreVal = new LoadInst(NewGV, LI->getName()+".b", false, 0,
1657 LI->getOrdering(), LI->getSynchScope(), LI);
1659 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1660 "This is not a form that we understand!");
1661 StoreVal = StoredVal->getOperand(0);
1662 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1665 new StoreInst(StoreVal, NewGV, false, 0,
1666 SI->getOrdering(), SI->getSynchScope(), SI);
1668 // Change the load into a load of bool then a select.
1669 LoadInst *LI = cast<LoadInst>(UI);
1670 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", false, 0,
1671 LI->getOrdering(), LI->getSynchScope(), LI);
1674 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1676 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1678 LI->replaceAllUsesWith(NSI);
1680 UI->eraseFromParent();
1683 // Retain the name of the old global variable. People who are debugging their
1684 // programs may expect these variables to be named the same.
1685 NewGV->takeName(GV);
1686 GV->eraseFromParent();
1691 /// ProcessGlobal - Analyze the specified global variable and optimize it if
1692 /// possible. If we make a change, return true.
1693 bool GlobalOpt::ProcessGlobal(GlobalVariable *GV,
1694 Module::global_iterator &GVI) {
1695 // Do more involved optimizations if the global is internal.
1696 GV->removeDeadConstantUsers();
1698 if (GV->use_empty()) {
1699 DEBUG(dbgs() << "GLOBAL DEAD: " << *GV);
1700 GV->eraseFromParent();
1705 if (!GV->hasLocalLinkage())
1710 if (GlobalStatus::analyzeGlobal(GV, GS))
1713 if (!GS.IsCompared && !GV->hasUnnamedAddr()) {
1714 GV->setUnnamedAddr(true);
1718 if (GV->isConstant() || !GV->hasInitializer())
1721 return ProcessInternalGlobal(GV, GVI, GS);
1724 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1725 /// it if possible. If we make a change, return true.
1726 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1727 Module::global_iterator &GVI,
1728 const GlobalStatus &GS) {
1729 auto &DL = GV->getParent()->getDataLayout();
1730 // If this is a first class global and has only one accessing function
1731 // and this function is main (which we know is not recursive), we replace
1732 // the global with a local alloca in this function.
1734 // NOTE: It doesn't make sense to promote non-single-value types since we
1735 // are just replacing static memory to stack memory.
1737 // If the global is in different address space, don't bring it to stack.
1738 if (!GS.HasMultipleAccessingFunctions &&
1739 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1740 GV->getType()->getElementType()->isSingleValueType() &&
1741 GS.AccessingFunction->getName() == "main" &&
1742 GS.AccessingFunction->hasExternalLinkage() &&
1743 GV->getType()->getAddressSpace() == 0) {
1744 DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV);
1745 Instruction &FirstI = const_cast<Instruction&>(*GS.AccessingFunction
1746 ->getEntryBlock().begin());
1747 Type *ElemTy = GV->getType()->getElementType();
1748 // FIXME: Pass Global's alignment when globals have alignment
1749 AllocaInst *Alloca = new AllocaInst(ElemTy, nullptr,
1750 GV->getName(), &FirstI);
1751 if (!isa<UndefValue>(GV->getInitializer()))
1752 new StoreInst(GV->getInitializer(), Alloca, &FirstI);
1754 GV->replaceAllUsesWith(Alloca);
1755 GV->eraseFromParent();
1760 // If the global is never loaded (but may be stored to), it is dead.
1763 DEBUG(dbgs() << "GLOBAL NEVER LOADED: " << *GV);
1766 if (isLeakCheckerRoot(GV)) {
1767 // Delete any constant stores to the global.
1768 Changed = CleanupPointerRootUsers(GV, TLI);
1770 // Delete any stores we can find to the global. We may not be able to
1771 // make it completely dead though.
1772 Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, TLI);
1775 // If the global is dead now, delete it.
1776 if (GV->use_empty()) {
1777 GV->eraseFromParent();
1783 } else if (GS.StoredType <= GlobalStatus::InitializerStored) {
1784 DEBUG(dbgs() << "MARKING CONSTANT: " << *GV << "\n");
1785 GV->setConstant(true);
1787 // Clean up any obviously simplifiable users now.
1788 CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, TLI);
1790 // If the global is dead now, just nuke it.
1791 if (GV->use_empty()) {
1792 DEBUG(dbgs() << " *** Marking constant allowed us to simplify "
1793 << "all users and delete global!\n");
1794 GV->eraseFromParent();
1800 } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
1801 const DataLayout &DL = GV->getParent()->getDataLayout();
1802 if (GlobalVariable *FirstNewGV = SRAGlobal(GV, DL)) {
1803 GVI = FirstNewGV; // Don't skip the newly produced globals!
1806 } else if (GS.StoredType == GlobalStatus::StoredOnce) {
1807 // If the initial value for the global was an undef value, and if only
1808 // one other value was stored into it, we can just change the
1809 // initializer to be the stored value, then delete all stores to the
1810 // global. This allows us to mark it constant.
1811 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1812 if (isa<UndefValue>(GV->getInitializer())) {
1813 // Change the initial value here.
1814 GV->setInitializer(SOVConstant);
1816 // Clean up any obviously simplifiable users now.
1817 CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, TLI);
1819 if (GV->use_empty()) {
1820 DEBUG(dbgs() << " *** Substituting initializer allowed us to "
1821 << "simplify all users and delete global!\n");
1822 GV->eraseFromParent();
1831 // Try to optimize globals based on the knowledge that only one value
1832 // (besides its initializer) is ever stored to the global.
1833 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GS.Ordering, GVI,
1837 // Otherwise, if the global was not a boolean, we can shrink it to be a
1839 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) {
1840 if (GS.Ordering == NotAtomic) {
1841 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1852 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1853 /// function, changing them to FastCC.
1854 static void ChangeCalleesToFastCall(Function *F) {
1855 for (User *U : F->users()) {
1856 if (isa<BlockAddress>(U))
1858 CallSite CS(cast<Instruction>(U));
1859 CS.setCallingConv(CallingConv::Fast);
1863 static AttributeSet StripNest(LLVMContext &C, const AttributeSet &Attrs) {
1864 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1865 unsigned Index = Attrs.getSlotIndex(i);
1866 if (!Attrs.getSlotAttributes(i).hasAttribute(Index, Attribute::Nest))
1869 // There can be only one.
1870 return Attrs.removeAttribute(C, Index, Attribute::Nest);
1876 static void RemoveNestAttribute(Function *F) {
1877 F->setAttributes(StripNest(F->getContext(), F->getAttributes()));
1878 for (User *U : F->users()) {
1879 if (isa<BlockAddress>(U))
1881 CallSite CS(cast<Instruction>(U));
1882 CS.setAttributes(StripNest(F->getContext(), CS.getAttributes()));
1886 /// Return true if this is a calling convention that we'd like to change. The
1887 /// idea here is that we don't want to mess with the convention if the user
1888 /// explicitly requested something with performance implications like coldcc,
1889 /// GHC, or anyregcc.
1890 static bool isProfitableToMakeFastCC(Function *F) {
1891 CallingConv::ID CC = F->getCallingConv();
1892 // FIXME: Is it worth transforming x86_stdcallcc and x86_fastcallcc?
1893 return CC == CallingConv::C || CC == CallingConv::X86_ThisCall;
1896 bool GlobalOpt::OptimizeFunctions(Module &M) {
1897 bool Changed = false;
1898 // Optimize functions.
1899 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1901 // Functions without names cannot be referenced outside this module.
1902 if (!F->hasName() && !F->isDeclaration() && !F->hasLocalLinkage())
1903 F->setLinkage(GlobalValue::InternalLinkage);
1905 const Comdat *C = F->getComdat();
1906 bool inComdat = C && NotDiscardableComdats.count(C);
1907 F->removeDeadConstantUsers();
1908 if ((!inComdat || F->hasLocalLinkage()) && F->isDefTriviallyDead()) {
1909 F->eraseFromParent();
1912 } else if (F->hasLocalLinkage()) {
1913 if (isProfitableToMakeFastCC(F) && !F->isVarArg() &&
1914 !F->hasAddressTaken()) {
1915 // If this function has a calling convention worth changing, is not a
1916 // varargs function, and is only called directly, promote it to use the
1917 // Fast calling convention.
1918 F->setCallingConv(CallingConv::Fast);
1919 ChangeCalleesToFastCall(F);
1924 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
1925 !F->hasAddressTaken()) {
1926 // The function is not used by a trampoline intrinsic, so it is safe
1927 // to remove the 'nest' attribute.
1928 RemoveNestAttribute(F);
1937 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1938 bool Changed = false;
1940 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1942 GlobalVariable *GV = GVI++;
1943 // Global variables without names cannot be referenced outside this module.
1944 if (!GV->hasName() && !GV->isDeclaration() && !GV->hasLocalLinkage())
1945 GV->setLinkage(GlobalValue::InternalLinkage);
1946 // Simplify the initializer.
1947 if (GV->hasInitializer())
1948 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GV->getInitializer())) {
1949 auto &DL = M.getDataLayout();
1950 Constant *New = ConstantFoldConstantExpression(CE, DL, TLI);
1951 if (New && New != CE)
1952 GV->setInitializer(New);
1955 if (GV->isDiscardableIfUnused()) {
1956 if (const Comdat *C = GV->getComdat())
1957 if (NotDiscardableComdats.count(C) && !GV->hasLocalLinkage())
1959 Changed |= ProcessGlobal(GV, GVI);
1966 isSimpleEnoughValueToCommit(Constant *C,
1967 SmallPtrSetImpl<Constant *> &SimpleConstants,
1968 const DataLayout &DL);
1970 /// isSimpleEnoughValueToCommit - Return true if the specified constant can be
1971 /// handled by the code generator. We don't want to generate something like:
1972 /// void *X = &X/42;
1973 /// because the code generator doesn't have a relocation that can handle that.
1975 /// This function should be called if C was not found (but just got inserted)
1976 /// in SimpleConstants to avoid having to rescan the same constants all the
1979 isSimpleEnoughValueToCommitHelper(Constant *C,
1980 SmallPtrSetImpl<Constant *> &SimpleConstants,
1981 const DataLayout &DL) {
1982 // Simple global addresses are supported, do not allow dllimport or
1983 // thread-local globals.
1984 if (auto *GV = dyn_cast<GlobalValue>(C))
1985 return !GV->hasDLLImportStorageClass() && !GV->isThreadLocal();
1987 // Simple integer, undef, constant aggregate zero, etc are all supported.
1988 if (C->getNumOperands() == 0 || isa<BlockAddress>(C))
1991 // Aggregate values are safe if all their elements are.
1992 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C) ||
1993 isa<ConstantVector>(C)) {
1994 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) {
1995 Constant *Op = cast<Constant>(C->getOperand(i));
1996 if (!isSimpleEnoughValueToCommit(Op, SimpleConstants, DL))
2002 // We don't know exactly what relocations are allowed in constant expressions,
2003 // so we allow &global+constantoffset, which is safe and uniformly supported
2005 ConstantExpr *CE = cast<ConstantExpr>(C);
2006 switch (CE->getOpcode()) {
2007 case Instruction::BitCast:
2008 // Bitcast is fine if the casted value is fine.
2009 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL);
2011 case Instruction::IntToPtr:
2012 case Instruction::PtrToInt:
2013 // int <=> ptr is fine if the int type is the same size as the
2015 if (DL.getTypeSizeInBits(CE->getType()) !=
2016 DL.getTypeSizeInBits(CE->getOperand(0)->getType()))
2018 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL);
2020 // GEP is fine if it is simple + constant offset.
2021 case Instruction::GetElementPtr:
2022 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
2023 if (!isa<ConstantInt>(CE->getOperand(i)))
2025 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL);
2027 case Instruction::Add:
2028 // We allow simple+cst.
2029 if (!isa<ConstantInt>(CE->getOperand(1)))
2031 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL);
2037 isSimpleEnoughValueToCommit(Constant *C,
2038 SmallPtrSetImpl<Constant *> &SimpleConstants,
2039 const DataLayout &DL) {
2040 // If we already checked this constant, we win.
2041 if (!SimpleConstants.insert(C).second)
2043 // Check the constant.
2044 return isSimpleEnoughValueToCommitHelper(C, SimpleConstants, DL);
2048 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
2049 /// enough for us to understand. In particular, if it is a cast to anything
2050 /// other than from one pointer type to another pointer type, we punt.
2051 /// We basically just support direct accesses to globals and GEP's of
2052 /// globals. This should be kept up to date with CommitValueTo.
2053 static bool isSimpleEnoughPointerToCommit(Constant *C) {
2054 // Conservatively, avoid aggregate types. This is because we don't
2055 // want to worry about them partially overlapping other stores.
2056 if (!cast<PointerType>(C->getType())->getElementType()->isSingleValueType())
2059 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
2060 // Do not allow weak/*_odr/linkonce linkage or external globals.
2061 return GV->hasUniqueInitializer();
2063 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2064 // Handle a constantexpr gep.
2065 if (CE->getOpcode() == Instruction::GetElementPtr &&
2066 isa<GlobalVariable>(CE->getOperand(0)) &&
2067 cast<GEPOperator>(CE)->isInBounds()) {
2068 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2069 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2070 // external globals.
2071 if (!GV->hasUniqueInitializer())
2074 // The first index must be zero.
2075 ConstantInt *CI = dyn_cast<ConstantInt>(*std::next(CE->op_begin()));
2076 if (!CI || !CI->isZero()) return false;
2078 // The remaining indices must be compile-time known integers within the
2079 // notional bounds of the corresponding static array types.
2080 if (!CE->isGEPWithNoNotionalOverIndexing())
2083 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2085 // A constantexpr bitcast from a pointer to another pointer is a no-op,
2086 // and we know how to evaluate it by moving the bitcast from the pointer
2087 // operand to the value operand.
2088 } else if (CE->getOpcode() == Instruction::BitCast &&
2089 isa<GlobalVariable>(CE->getOperand(0))) {
2090 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2091 // external globals.
2092 return cast<GlobalVariable>(CE->getOperand(0))->hasUniqueInitializer();
2099 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
2100 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
2101 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
2102 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
2103 ConstantExpr *Addr, unsigned OpNo) {
2104 // Base case of the recursion.
2105 if (OpNo == Addr->getNumOperands()) {
2106 assert(Val->getType() == Init->getType() && "Type mismatch!");
2110 SmallVector<Constant*, 32> Elts;
2111 if (StructType *STy = dyn_cast<StructType>(Init->getType())) {
2112 // Break up the constant into its elements.
2113 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2114 Elts.push_back(Init->getAggregateElement(i));
2116 // Replace the element that we are supposed to.
2117 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
2118 unsigned Idx = CU->getZExtValue();
2119 assert(Idx < STy->getNumElements() && "Struct index out of range!");
2120 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
2122 // Return the modified struct.
2123 return ConstantStruct::get(STy, Elts);
2126 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
2127 SequentialType *InitTy = cast<SequentialType>(Init->getType());
2130 if (ArrayType *ATy = dyn_cast<ArrayType>(InitTy))
2131 NumElts = ATy->getNumElements();
2133 NumElts = InitTy->getVectorNumElements();
2135 // Break up the array into elements.
2136 for (uint64_t i = 0, e = NumElts; i != e; ++i)
2137 Elts.push_back(Init->getAggregateElement(i));
2139 assert(CI->getZExtValue() < NumElts);
2140 Elts[CI->getZExtValue()] =
2141 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
2143 if (Init->getType()->isArrayTy())
2144 return ConstantArray::get(cast<ArrayType>(InitTy), Elts);
2145 return ConstantVector::get(Elts);
2148 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
2149 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
2150 static void CommitValueTo(Constant *Val, Constant *Addr) {
2151 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
2152 assert(GV->hasInitializer());
2153 GV->setInitializer(Val);
2157 ConstantExpr *CE = cast<ConstantExpr>(Addr);
2158 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2159 GV->setInitializer(EvaluateStoreInto(GV->getInitializer(), Val, CE, 2));
2164 /// Evaluator - This class evaluates LLVM IR, producing the Constant
2165 /// representing each SSA instruction. Changes to global variables are stored
2166 /// in a mapping that can be iterated over after the evaluation is complete.
2167 /// Once an evaluation call fails, the evaluation object should not be reused.
2170 Evaluator(const DataLayout &DL, const TargetLibraryInfo *TLI)
2171 : DL(DL), TLI(TLI) {
2172 ValueStack.emplace_back();
2176 for (auto &Tmp : AllocaTmps)
2177 // If there are still users of the alloca, the program is doing something
2178 // silly, e.g. storing the address of the alloca somewhere and using it
2179 // later. Since this is undefined, we'll just make it be null.
2180 if (!Tmp->use_empty())
2181 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2184 /// EvaluateFunction - Evaluate a call to function F, returning true if
2185 /// successful, false if we can't evaluate it. ActualArgs contains the formal
2186 /// arguments for the function.
2187 bool EvaluateFunction(Function *F, Constant *&RetVal,
2188 const SmallVectorImpl<Constant*> &ActualArgs);
2190 /// EvaluateBlock - Evaluate all instructions in block BB, returning true if
2191 /// successful, false if we can't evaluate it. NewBB returns the next BB that
2192 /// control flows into, or null upon return.
2193 bool EvaluateBlock(BasicBlock::iterator CurInst, BasicBlock *&NextBB);
2195 Constant *getVal(Value *V) {
2196 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
2197 Constant *R = ValueStack.back().lookup(V);
2198 assert(R && "Reference to an uncomputed value!");
2202 void setVal(Value *V, Constant *C) {
2203 ValueStack.back()[V] = C;
2206 const DenseMap<Constant*, Constant*> &getMutatedMemory() const {
2207 return MutatedMemory;
2210 const SmallPtrSetImpl<GlobalVariable*> &getInvariants() const {
2215 Constant *ComputeLoadResult(Constant *P);
2217 /// ValueStack - As we compute SSA register values, we store their contents
2218 /// here. The back of the deque contains the current function and the stack
2219 /// contains the values in the calling frames.
2220 std::deque<DenseMap<Value*, Constant*>> ValueStack;
2222 /// CallStack - This is used to detect recursion. In pathological situations
2223 /// we could hit exponential behavior, but at least there is nothing
2225 SmallVector<Function*, 4> CallStack;
2227 /// MutatedMemory - For each store we execute, we update this map. Loads
2228 /// check this to get the most up-to-date value. If evaluation is successful,
2229 /// this state is committed to the process.
2230 DenseMap<Constant*, Constant*> MutatedMemory;
2232 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2233 /// to represent its body. This vector is needed so we can delete the
2234 /// temporary globals when we are done.
2235 SmallVector<std::unique_ptr<GlobalVariable>, 32> AllocaTmps;
2237 /// Invariants - These global variables have been marked invariant by the
2238 /// static constructor.
2239 SmallPtrSet<GlobalVariable*, 8> Invariants;
2241 /// SimpleConstants - These are constants we have checked and know to be
2242 /// simple enough to live in a static initializer of a global.
2243 SmallPtrSet<Constant*, 8> SimpleConstants;
2245 const DataLayout &DL;
2246 const TargetLibraryInfo *TLI;
2249 } // anonymous namespace
2251 /// ComputeLoadResult - Return the value that would be computed by a load from
2252 /// P after the stores reflected by 'memory' have been performed. If we can't
2253 /// decide, return null.
2254 Constant *Evaluator::ComputeLoadResult(Constant *P) {
2255 // If this memory location has been recently stored, use the stored value: it
2256 // is the most up-to-date.
2257 DenseMap<Constant*, Constant*>::const_iterator I = MutatedMemory.find(P);
2258 if (I != MutatedMemory.end()) return I->second;
2261 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
2262 if (GV->hasDefinitiveInitializer())
2263 return GV->getInitializer();
2267 // Handle a constantexpr getelementptr.
2268 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
2269 if (CE->getOpcode() == Instruction::GetElementPtr &&
2270 isa<GlobalVariable>(CE->getOperand(0))) {
2271 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2272 if (GV->hasDefinitiveInitializer())
2273 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2276 return nullptr; // don't know how to evaluate.
2279 /// EvaluateBlock - Evaluate all instructions in block BB, returning true if
2280 /// successful, false if we can't evaluate it. NewBB returns the next BB that
2281 /// control flows into, or null upon return.
2282 bool Evaluator::EvaluateBlock(BasicBlock::iterator CurInst,
2283 BasicBlock *&NextBB) {
2284 // This is the main evaluation loop.
2286 Constant *InstResult = nullptr;
2288 DEBUG(dbgs() << "Evaluating Instruction: " << *CurInst << "\n");
2290 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
2291 if (!SI->isSimple()) {
2292 DEBUG(dbgs() << "Store is not simple! Can not evaluate.\n");
2293 return false; // no volatile/atomic accesses.
2295 Constant *Ptr = getVal(SI->getOperand(1));
2296 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
2297 DEBUG(dbgs() << "Folding constant ptr expression: " << *Ptr);
2298 Ptr = ConstantFoldConstantExpression(CE, DL, TLI);
2299 DEBUG(dbgs() << "; To: " << *Ptr << "\n");
2301 if (!isSimpleEnoughPointerToCommit(Ptr)) {
2302 // If this is too complex for us to commit, reject it.
2303 DEBUG(dbgs() << "Pointer is too complex for us to evaluate store.");
2307 Constant *Val = getVal(SI->getOperand(0));
2309 // If this might be too difficult for the backend to handle (e.g. the addr
2310 // of one global variable divided by another) then we can't commit it.
2311 if (!isSimpleEnoughValueToCommit(Val, SimpleConstants, DL)) {
2312 DEBUG(dbgs() << "Store value is too complex to evaluate store. " << *Val
2317 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
2318 if (CE->getOpcode() == Instruction::BitCast) {
2319 DEBUG(dbgs() << "Attempting to resolve bitcast on constant ptr.\n");
2320 // If we're evaluating a store through a bitcast, then we need
2321 // to pull the bitcast off the pointer type and push it onto the
2323 Ptr = CE->getOperand(0);
2325 Type *NewTy = cast<PointerType>(Ptr->getType())->getElementType();
2327 // In order to push the bitcast onto the stored value, a bitcast
2328 // from NewTy to Val's type must be legal. If it's not, we can try
2329 // introspecting NewTy to find a legal conversion.
2330 while (!Val->getType()->canLosslesslyBitCastTo(NewTy)) {
2331 // If NewTy is a struct, we can convert the pointer to the struct
2332 // into a pointer to its first member.
2333 // FIXME: This could be extended to support arrays as well.
2334 if (StructType *STy = dyn_cast<StructType>(NewTy)) {
2335 NewTy = STy->getTypeAtIndex(0U);
2337 IntegerType *IdxTy = IntegerType::get(NewTy->getContext(), 32);
2338 Constant *IdxZero = ConstantInt::get(IdxTy, 0, false);
2339 Constant * const IdxList[] = {IdxZero, IdxZero};
2341 Ptr = ConstantExpr::getGetElementPtr(Ptr, IdxList);
2342 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
2343 Ptr = ConstantFoldConstantExpression(CE, DL, TLI);
2345 // If we can't improve the situation by introspecting NewTy,
2346 // we have to give up.
2348 DEBUG(dbgs() << "Failed to bitcast constant ptr, can not "
2354 // If we found compatible types, go ahead and push the bitcast
2355 // onto the stored value.
2356 Val = ConstantExpr::getBitCast(Val, NewTy);
2358 DEBUG(dbgs() << "Evaluated bitcast: " << *Val << "\n");
2362 MutatedMemory[Ptr] = Val;
2363 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2364 InstResult = ConstantExpr::get(BO->getOpcode(),
2365 getVal(BO->getOperand(0)),
2366 getVal(BO->getOperand(1)));
2367 DEBUG(dbgs() << "Found a BinaryOperator! Simplifying: " << *InstResult
2369 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2370 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2371 getVal(CI->getOperand(0)),
2372 getVal(CI->getOperand(1)));
2373 DEBUG(dbgs() << "Found a CmpInst! Simplifying: " << *InstResult
2375 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2376 InstResult = ConstantExpr::getCast(CI->getOpcode(),
2377 getVal(CI->getOperand(0)),
2379 DEBUG(dbgs() << "Found a Cast! Simplifying: " << *InstResult
2381 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2382 InstResult = ConstantExpr::getSelect(getVal(SI->getOperand(0)),
2383 getVal(SI->getOperand(1)),
2384 getVal(SI->getOperand(2)));
2385 DEBUG(dbgs() << "Found a Select! Simplifying: " << *InstResult
2387 } else if (auto *EVI = dyn_cast<ExtractValueInst>(CurInst)) {
2388 InstResult = ConstantExpr::getExtractValue(
2389 getVal(EVI->getAggregateOperand()), EVI->getIndices());
2390 DEBUG(dbgs() << "Found an ExtractValueInst! Simplifying: " << *InstResult
2392 } else if (auto *IVI = dyn_cast<InsertValueInst>(CurInst)) {
2393 InstResult = ConstantExpr::getInsertValue(
2394 getVal(IVI->getAggregateOperand()),
2395 getVal(IVI->getInsertedValueOperand()), IVI->getIndices());
2396 DEBUG(dbgs() << "Found an InsertValueInst! Simplifying: " << *InstResult
2398 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2399 Constant *P = getVal(GEP->getOperand(0));
2400 SmallVector<Constant*, 8> GEPOps;
2401 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2403 GEPOps.push_back(getVal(*i));
2405 ConstantExpr::getGetElementPtr(P, GEPOps,
2406 cast<GEPOperator>(GEP)->isInBounds());
2407 DEBUG(dbgs() << "Found a GEP! Simplifying: " << *InstResult
2409 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2411 if (!LI->isSimple()) {
2412 DEBUG(dbgs() << "Found a Load! Not a simple load, can not evaluate.\n");
2413 return false; // no volatile/atomic accesses.
2416 Constant *Ptr = getVal(LI->getOperand(0));
2417 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
2418 Ptr = ConstantFoldConstantExpression(CE, DL, TLI);
2419 DEBUG(dbgs() << "Found a constant pointer expression, constant "
2420 "folding: " << *Ptr << "\n");
2422 InstResult = ComputeLoadResult(Ptr);
2424 DEBUG(dbgs() << "Failed to compute load result. Can not evaluate load."
2426 return false; // Could not evaluate load.
2429 DEBUG(dbgs() << "Evaluated load: " << *InstResult << "\n");
2430 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2431 if (AI->isArrayAllocation()) {
2432 DEBUG(dbgs() << "Found an array alloca. Can not evaluate.\n");
2433 return false; // Cannot handle array allocs.
2435 Type *Ty = AI->getType()->getElementType();
2436 AllocaTmps.push_back(
2437 make_unique<GlobalVariable>(Ty, false, GlobalValue::InternalLinkage,
2438 UndefValue::get(Ty), AI->getName()));
2439 InstResult = AllocaTmps.back().get();
2440 DEBUG(dbgs() << "Found an alloca. Result: " << *InstResult << "\n");
2441 } else if (isa<CallInst>(CurInst) || isa<InvokeInst>(CurInst)) {
2442 CallSite CS(CurInst);
2444 // Debug info can safely be ignored here.
2445 if (isa<DbgInfoIntrinsic>(CS.getInstruction())) {
2446 DEBUG(dbgs() << "Ignoring debug info.\n");
2451 // Cannot handle inline asm.
2452 if (isa<InlineAsm>(CS.getCalledValue())) {
2453 DEBUG(dbgs() << "Found inline asm, can not evaluate.\n");
2457 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) {
2458 if (MemSetInst *MSI = dyn_cast<MemSetInst>(II)) {
2459 if (MSI->isVolatile()) {
2460 DEBUG(dbgs() << "Can not optimize a volatile memset " <<
2464 Constant *Ptr = getVal(MSI->getDest());
2465 Constant *Val = getVal(MSI->getValue());
2466 Constant *DestVal = ComputeLoadResult(getVal(Ptr));
2467 if (Val->isNullValue() && DestVal && DestVal->isNullValue()) {
2468 // This memset is a no-op.
2469 DEBUG(dbgs() << "Ignoring no-op memset.\n");
2475 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
2476 II->getIntrinsicID() == Intrinsic::lifetime_end) {
2477 DEBUG(dbgs() << "Ignoring lifetime intrinsic.\n");
2482 if (II->getIntrinsicID() == Intrinsic::invariant_start) {
2483 // We don't insert an entry into Values, as it doesn't have a
2484 // meaningful return value.
2485 if (!II->use_empty()) {
2486 DEBUG(dbgs() << "Found unused invariant_start. Can't evaluate.\n");
2489 ConstantInt *Size = cast<ConstantInt>(II->getArgOperand(0));
2490 Value *PtrArg = getVal(II->getArgOperand(1));
2491 Value *Ptr = PtrArg->stripPointerCasts();
2492 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) {
2493 Type *ElemTy = cast<PointerType>(GV->getType())->getElementType();
2494 if (!Size->isAllOnesValue() &&
2495 Size->getValue().getLimitedValue() >=
2496 DL.getTypeStoreSize(ElemTy)) {
2497 Invariants.insert(GV);
2498 DEBUG(dbgs() << "Found a global var that is an invariant: " << *GV
2501 DEBUG(dbgs() << "Found a global var, but can not treat it as an "
2505 // Continue even if we do nothing.
2510 DEBUG(dbgs() << "Unknown intrinsic. Can not evaluate.\n");
2514 // Resolve function pointers.
2515 Function *Callee = dyn_cast<Function>(getVal(CS.getCalledValue()));
2516 if (!Callee || Callee->mayBeOverridden()) {
2517 DEBUG(dbgs() << "Can not resolve function pointer.\n");
2518 return false; // Cannot resolve.
2521 SmallVector<Constant*, 8> Formals;
2522 for (User::op_iterator i = CS.arg_begin(), e = CS.arg_end(); i != e; ++i)
2523 Formals.push_back(getVal(*i));
2525 if (Callee->isDeclaration()) {
2526 // If this is a function we can constant fold, do it.
2527 if (Constant *C = ConstantFoldCall(Callee, Formals, TLI)) {
2529 DEBUG(dbgs() << "Constant folded function call. Result: " <<
2530 *InstResult << "\n");
2532 DEBUG(dbgs() << "Can not constant fold function call.\n");
2536 if (Callee->getFunctionType()->isVarArg()) {
2537 DEBUG(dbgs() << "Can not constant fold vararg function call.\n");
2541 Constant *RetVal = nullptr;
2542 // Execute the call, if successful, use the return value.
2543 ValueStack.emplace_back();
2544 if (!EvaluateFunction(Callee, RetVal, Formals)) {
2545 DEBUG(dbgs() << "Failed to evaluate function.\n");
2548 ValueStack.pop_back();
2549 InstResult = RetVal;
2552 DEBUG(dbgs() << "Successfully evaluated function. Result: " <<
2553 InstResult << "\n\n");
2555 DEBUG(dbgs() << "Successfully evaluated function. Result: 0\n\n");
2558 } else if (isa<TerminatorInst>(CurInst)) {
2559 DEBUG(dbgs() << "Found a terminator instruction.\n");
2561 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2562 if (BI->isUnconditional()) {
2563 NextBB = BI->getSuccessor(0);
2566 dyn_cast<ConstantInt>(getVal(BI->getCondition()));
2567 if (!Cond) return false; // Cannot determine.
2569 NextBB = BI->getSuccessor(!Cond->getZExtValue());
2571 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2573 dyn_cast<ConstantInt>(getVal(SI->getCondition()));
2574 if (!Val) return false; // Cannot determine.
2575 NextBB = SI->findCaseValue(Val).getCaseSuccessor();
2576 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(CurInst)) {
2577 Value *Val = getVal(IBI->getAddress())->stripPointerCasts();
2578 if (BlockAddress *BA = dyn_cast<BlockAddress>(Val))
2579 NextBB = BA->getBasicBlock();
2581 return false; // Cannot determine.
2582 } else if (isa<ReturnInst>(CurInst)) {
2585 // invoke, unwind, resume, unreachable.
2586 DEBUG(dbgs() << "Can not handle terminator.");
2587 return false; // Cannot handle this terminator.
2590 // We succeeded at evaluating this block!
2591 DEBUG(dbgs() << "Successfully evaluated block.\n");
2594 // Did not know how to evaluate this!
2595 DEBUG(dbgs() << "Failed to evaluate block due to unhandled instruction."
2600 if (!CurInst->use_empty()) {
2601 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(InstResult))
2602 InstResult = ConstantFoldConstantExpression(CE, DL, TLI);
2604 setVal(CurInst, InstResult);
2607 // If we just processed an invoke, we finished evaluating the block.
2608 if (InvokeInst *II = dyn_cast<InvokeInst>(CurInst)) {
2609 NextBB = II->getNormalDest();
2610 DEBUG(dbgs() << "Found an invoke instruction. Finished Block.\n\n");
2614 // Advance program counter.
2619 /// EvaluateFunction - Evaluate a call to function F, returning true if
2620 /// successful, false if we can't evaluate it. ActualArgs contains the formal
2621 /// arguments for the function.
2622 bool Evaluator::EvaluateFunction(Function *F, Constant *&RetVal,
2623 const SmallVectorImpl<Constant*> &ActualArgs) {
2624 // Check to see if this function is already executing (recursion). If so,
2625 // bail out. TODO: we might want to accept limited recursion.
2626 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
2629 CallStack.push_back(F);
2631 // Initialize arguments to the incoming values specified.
2633 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
2635 setVal(AI, ActualArgs[ArgNo]);
2637 // ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
2638 // we can only evaluate any one basic block at most once. This set keeps
2639 // track of what we have executed so we can detect recursive cases etc.
2640 SmallPtrSet<BasicBlock*, 32> ExecutedBlocks;
2642 // CurBB - The current basic block we're evaluating.
2643 BasicBlock *CurBB = F->begin();
2645 BasicBlock::iterator CurInst = CurBB->begin();
2648 BasicBlock *NextBB = nullptr; // Initialized to avoid compiler warnings.
2649 DEBUG(dbgs() << "Trying to evaluate BB: " << *CurBB << "\n");
2651 if (!EvaluateBlock(CurInst, NextBB))
2655 // Successfully running until there's no next block means that we found
2656 // the return. Fill it the return value and pop the call stack.
2657 ReturnInst *RI = cast<ReturnInst>(CurBB->getTerminator());
2658 if (RI->getNumOperands())
2659 RetVal = getVal(RI->getOperand(0));
2660 CallStack.pop_back();
2664 // Okay, we succeeded in evaluating this control flow. See if we have
2665 // executed the new block before. If so, we have a looping function,
2666 // which we cannot evaluate in reasonable time.
2667 if (!ExecutedBlocks.insert(NextBB).second)
2668 return false; // looped!
2670 // Okay, we have never been in this block before. Check to see if there
2671 // are any PHI nodes. If so, evaluate them with information about where
2673 PHINode *PN = nullptr;
2674 for (CurInst = NextBB->begin();
2675 (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2676 setVal(PN, getVal(PN->getIncomingValueForBlock(CurBB)));
2678 // Advance to the next block.
2683 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2684 /// we can. Return true if we can, false otherwise.
2685 static bool EvaluateStaticConstructor(Function *F, const DataLayout &DL,
2686 const TargetLibraryInfo *TLI) {
2687 // Call the function.
2688 Evaluator Eval(DL, TLI);
2689 Constant *RetValDummy;
2690 bool EvalSuccess = Eval.EvaluateFunction(F, RetValDummy,
2691 SmallVector<Constant*, 0>());
2694 ++NumCtorsEvaluated;
2696 // We succeeded at evaluation: commit the result.
2697 DEBUG(dbgs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2698 << F->getName() << "' to " << Eval.getMutatedMemory().size()
2700 for (DenseMap<Constant*, Constant*>::const_iterator I =
2701 Eval.getMutatedMemory().begin(), E = Eval.getMutatedMemory().end();
2703 CommitValueTo(I->second, I->first);
2704 for (GlobalVariable *GV : Eval.getInvariants())
2705 GV->setConstant(true);
2711 static int compareNames(Constant *const *A, Constant *const *B) {
2712 return (*A)->getName().compare((*B)->getName());
2715 static void setUsedInitializer(GlobalVariable &V,
2716 const SmallPtrSet<GlobalValue *, 8> &Init) {
2718 V.eraseFromParent();
2722 // Type of pointer to the array of pointers.
2723 PointerType *Int8PtrTy = Type::getInt8PtrTy(V.getContext(), 0);
2725 SmallVector<llvm::Constant *, 8> UsedArray;
2726 for (GlobalValue *GV : Init) {
2728 = ConstantExpr::getPointerBitCastOrAddrSpaceCast(GV, Int8PtrTy);
2729 UsedArray.push_back(Cast);
2731 // Sort to get deterministic order.
2732 array_pod_sort(UsedArray.begin(), UsedArray.end(), compareNames);
2733 ArrayType *ATy = ArrayType::get(Int8PtrTy, UsedArray.size());
2735 Module *M = V.getParent();
2736 V.removeFromParent();
2737 GlobalVariable *NV =
2738 new GlobalVariable(*M, ATy, false, llvm::GlobalValue::AppendingLinkage,
2739 llvm::ConstantArray::get(ATy, UsedArray), "");
2741 NV->setSection("llvm.metadata");
2746 /// \brief An easy to access representation of llvm.used and llvm.compiler.used.
2748 SmallPtrSet<GlobalValue *, 8> Used;
2749 SmallPtrSet<GlobalValue *, 8> CompilerUsed;
2750 GlobalVariable *UsedV;
2751 GlobalVariable *CompilerUsedV;
2754 LLVMUsed(Module &M) {
2755 UsedV = collectUsedGlobalVariables(M, Used, false);
2756 CompilerUsedV = collectUsedGlobalVariables(M, CompilerUsed, true);
2758 typedef SmallPtrSet<GlobalValue *, 8>::iterator iterator;
2759 typedef iterator_range<iterator> used_iterator_range;
2760 iterator usedBegin() { return Used.begin(); }
2761 iterator usedEnd() { return Used.end(); }
2762 used_iterator_range used() {
2763 return used_iterator_range(usedBegin(), usedEnd());
2765 iterator compilerUsedBegin() { return CompilerUsed.begin(); }
2766 iterator compilerUsedEnd() { return CompilerUsed.end(); }
2767 used_iterator_range compilerUsed() {
2768 return used_iterator_range(compilerUsedBegin(), compilerUsedEnd());
2770 bool usedCount(GlobalValue *GV) const { return Used.count(GV); }
2771 bool compilerUsedCount(GlobalValue *GV) const {
2772 return CompilerUsed.count(GV);
2774 bool usedErase(GlobalValue *GV) { return Used.erase(GV); }
2775 bool compilerUsedErase(GlobalValue *GV) { return CompilerUsed.erase(GV); }
2776 bool usedInsert(GlobalValue *GV) { return Used.insert(GV).second; }
2777 bool compilerUsedInsert(GlobalValue *GV) {
2778 return CompilerUsed.insert(GV).second;
2781 void syncVariablesAndSets() {
2783 setUsedInitializer(*UsedV, Used);
2785 setUsedInitializer(*CompilerUsedV, CompilerUsed);
2790 static bool hasUseOtherThanLLVMUsed(GlobalAlias &GA, const LLVMUsed &U) {
2791 if (GA.use_empty()) // No use at all.
2794 assert((!U.usedCount(&GA) || !U.compilerUsedCount(&GA)) &&
2795 "We should have removed the duplicated "
2796 "element from llvm.compiler.used");
2797 if (!GA.hasOneUse())
2798 // Strictly more than one use. So at least one is not in llvm.used and
2799 // llvm.compiler.used.
2802 // Exactly one use. Check if it is in llvm.used or llvm.compiler.used.
2803 return !U.usedCount(&GA) && !U.compilerUsedCount(&GA);
2806 static bool hasMoreThanOneUseOtherThanLLVMUsed(GlobalValue &V,
2807 const LLVMUsed &U) {
2809 assert((!U.usedCount(&V) || !U.compilerUsedCount(&V)) &&
2810 "We should have removed the duplicated "
2811 "element from llvm.compiler.used");
2812 if (U.usedCount(&V) || U.compilerUsedCount(&V))
2814 return V.hasNUsesOrMore(N);
2817 static bool mayHaveOtherReferences(GlobalAlias &GA, const LLVMUsed &U) {
2818 if (!GA.hasLocalLinkage())
2821 return U.usedCount(&GA) || U.compilerUsedCount(&GA);
2824 static bool hasUsesToReplace(GlobalAlias &GA, const LLVMUsed &U,
2825 bool &RenameTarget) {
2826 RenameTarget = false;
2828 if (hasUseOtherThanLLVMUsed(GA, U))
2831 // If the alias is externally visible, we may still be able to simplify it.
2832 if (!mayHaveOtherReferences(GA, U))
2835 // If the aliasee has internal linkage, give it the name and linkage
2836 // of the alias, and delete the alias. This turns:
2837 // define internal ... @f(...)
2838 // @a = alias ... @f
2840 // define ... @a(...)
2841 Constant *Aliasee = GA.getAliasee();
2842 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2843 if (!Target->hasLocalLinkage())
2846 // Do not perform the transform if multiple aliases potentially target the
2847 // aliasee. This check also ensures that it is safe to replace the section
2848 // and other attributes of the aliasee with those of the alias.
2849 if (hasMoreThanOneUseOtherThanLLVMUsed(*Target, U))
2852 RenameTarget = true;
2856 bool GlobalOpt::OptimizeGlobalAliases(Module &M) {
2857 bool Changed = false;
2860 for (GlobalValue *GV : Used.used())
2861 Used.compilerUsedErase(GV);
2863 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end();
2865 Module::alias_iterator J = I++;
2866 // Aliases without names cannot be referenced outside this module.
2867 if (!J->hasName() && !J->isDeclaration() && !J->hasLocalLinkage())
2868 J->setLinkage(GlobalValue::InternalLinkage);
2869 // If the aliasee may change at link time, nothing can be done - bail out.
2870 if (J->mayBeOverridden())
2873 Constant *Aliasee = J->getAliasee();
2874 GlobalValue *Target = dyn_cast<GlobalValue>(Aliasee->stripPointerCasts());
2875 // We can't trivially replace the alias with the aliasee if the aliasee is
2876 // non-trivial in some way.
2877 // TODO: Try to handle non-zero GEPs of local aliasees.
2880 Target->removeDeadConstantUsers();
2882 // Make all users of the alias use the aliasee instead.
2884 if (!hasUsesToReplace(*J, Used, RenameTarget))
2887 J->replaceAllUsesWith(ConstantExpr::getBitCast(Aliasee, J->getType()));
2888 ++NumAliasesResolved;
2892 // Give the aliasee the name, linkage and other attributes of the alias.
2893 Target->takeName(J);
2894 Target->setLinkage(J->getLinkage());
2895 Target->setVisibility(J->getVisibility());
2896 Target->setDLLStorageClass(J->getDLLStorageClass());
2898 if (Used.usedErase(J))
2899 Used.usedInsert(Target);
2901 if (Used.compilerUsedErase(J))
2902 Used.compilerUsedInsert(Target);
2903 } else if (mayHaveOtherReferences(*J, Used))
2906 // Delete the alias.
2907 M.getAliasList().erase(J);
2908 ++NumAliasesRemoved;
2912 Used.syncVariablesAndSets();
2917 static Function *FindCXAAtExit(Module &M, TargetLibraryInfo *TLI) {
2918 if (!TLI->has(LibFunc::cxa_atexit))
2921 Function *Fn = M.getFunction(TLI->getName(LibFunc::cxa_atexit));
2926 FunctionType *FTy = Fn->getFunctionType();
2928 // Checking that the function has the right return type, the right number of
2929 // parameters and that they all have pointer types should be enough.
2930 if (!FTy->getReturnType()->isIntegerTy() ||
2931 FTy->getNumParams() != 3 ||
2932 !FTy->getParamType(0)->isPointerTy() ||
2933 !FTy->getParamType(1)->isPointerTy() ||
2934 !FTy->getParamType(2)->isPointerTy())
2940 /// cxxDtorIsEmpty - Returns whether the given function is an empty C++
2941 /// destructor and can therefore be eliminated.
2942 /// Note that we assume that other optimization passes have already simplified
2943 /// the code so we only look for a function with a single basic block, where
2944 /// the only allowed instructions are 'ret', 'call' to an empty C++ dtor and
2945 /// other side-effect free instructions.
2946 static bool cxxDtorIsEmpty(const Function &Fn,
2947 SmallPtrSet<const Function *, 8> &CalledFunctions) {
2948 // FIXME: We could eliminate C++ destructors if they're readonly/readnone and
2949 // nounwind, but that doesn't seem worth doing.
2950 if (Fn.isDeclaration())
2953 if (++Fn.begin() != Fn.end())
2956 const BasicBlock &EntryBlock = Fn.getEntryBlock();
2957 for (BasicBlock::const_iterator I = EntryBlock.begin(), E = EntryBlock.end();
2959 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
2960 // Ignore debug intrinsics.
2961 if (isa<DbgInfoIntrinsic>(CI))
2964 const Function *CalledFn = CI->getCalledFunction();
2969 SmallPtrSet<const Function *, 8> NewCalledFunctions(CalledFunctions);
2971 // Don't treat recursive functions as empty.
2972 if (!NewCalledFunctions.insert(CalledFn).second)
2975 if (!cxxDtorIsEmpty(*CalledFn, NewCalledFunctions))
2977 } else if (isa<ReturnInst>(*I))
2978 return true; // We're done.
2979 else if (I->mayHaveSideEffects())
2980 return false; // Destructor with side effects, bail.
2986 bool GlobalOpt::OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) {
2987 /// Itanium C++ ABI p3.3.5:
2989 /// After constructing a global (or local static) object, that will require
2990 /// destruction on exit, a termination function is registered as follows:
2992 /// extern "C" int __cxa_atexit ( void (*f)(void *), void *p, void *d );
2994 /// This registration, e.g. __cxa_atexit(f,p,d), is intended to cause the
2995 /// call f(p) when DSO d is unloaded, before all such termination calls
2996 /// registered before this one. It returns zero if registration is
2997 /// successful, nonzero on failure.
2999 // This pass will look for calls to __cxa_atexit where the function is trivial
3001 bool Changed = false;
3003 for (auto I = CXAAtExitFn->user_begin(), E = CXAAtExitFn->user_end();
3005 // We're only interested in calls. Theoretically, we could handle invoke
3006 // instructions as well, but neither llvm-gcc nor clang generate invokes
3008 CallInst *CI = dyn_cast<CallInst>(*I++);
3013 dyn_cast<Function>(CI->getArgOperand(0)->stripPointerCasts());
3017 SmallPtrSet<const Function *, 8> CalledFunctions;
3018 if (!cxxDtorIsEmpty(*DtorFn, CalledFunctions))
3021 // Just remove the call.
3022 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
3023 CI->eraseFromParent();
3025 ++NumCXXDtorsRemoved;
3033 bool GlobalOpt::runOnModule(Module &M) {
3034 bool Changed = false;
3036 auto &DL = M.getDataLayout();
3037 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
3039 bool LocalChange = true;
3040 while (LocalChange) {
3041 LocalChange = false;
3043 NotDiscardableComdats.clear();
3044 for (const GlobalVariable &GV : M.globals())
3045 if (const Comdat *C = GV.getComdat())
3046 if (!GV.isDiscardableIfUnused() || !GV.use_empty())
3047 NotDiscardableComdats.insert(C);
3048 for (Function &F : M)
3049 if (const Comdat *C = F.getComdat())
3050 if (!F.isDefTriviallyDead())
3051 NotDiscardableComdats.insert(C);
3052 for (GlobalAlias &GA : M.aliases())
3053 if (const Comdat *C = GA.getComdat())
3054 if (!GA.isDiscardableIfUnused() || !GA.use_empty())
3055 NotDiscardableComdats.insert(C);
3057 // Delete functions that are trivially dead, ccc -> fastcc
3058 LocalChange |= OptimizeFunctions(M);
3060 // Optimize global_ctors list.
3061 LocalChange |= optimizeGlobalCtorsList(M, [&](Function *F) {
3062 return EvaluateStaticConstructor(F, DL, TLI);
3065 // Optimize non-address-taken globals.
3066 LocalChange |= OptimizeGlobalVars(M);
3068 // Resolve aliases, when possible.
3069 LocalChange |= OptimizeGlobalAliases(M);
3071 // Try to remove trivial global destructors if they are not removed
3073 Function *CXAAtExitFn = FindCXAAtExit(M, TLI);
3075 LocalChange |= OptimizeEmptyGlobalCXXDtors(CXAAtExitFn);
3077 Changed |= LocalChange;
3080 // TODO: Move all global ctors functions to the end of the module for code