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/SmallVector.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/ConstantFolding.h"
23 #include "llvm/Analysis/MemoryBuiltins.h"
24 #include "llvm/IR/CallSite.h"
25 #include "llvm/IR/CallingConv.h"
26 #include "llvm/IR/Constants.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/DerivedTypes.h"
29 #include "llvm/IR/GetElementPtrTypeIterator.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/Module.h"
33 #include "llvm/IR/Operator.h"
34 #include "llvm/IR/ValueHandle.h"
35 #include "llvm/Pass.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/ErrorHandling.h"
38 #include "llvm/Support/MathExtras.h"
39 #include "llvm/Support/raw_ostream.h"
40 #include "llvm/Target/TargetLibraryInfo.h"
41 #include "llvm/Transforms/Utils/GlobalStatus.h"
42 #include "llvm/Transforms/Utils/ModuleUtils.h"
46 #define DEBUG_TYPE "globalopt"
48 STATISTIC(NumMarked , "Number of globals marked constant");
49 STATISTIC(NumUnnamed , "Number of globals marked unnamed_addr");
50 STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
51 STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
52 STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
53 STATISTIC(NumDeleted , "Number of globals deleted");
54 STATISTIC(NumFnDeleted , "Number of functions deleted");
55 STATISTIC(NumGlobUses , "Number of global uses devirtualized");
56 STATISTIC(NumLocalized , "Number of globals localized");
57 STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans");
58 STATISTIC(NumFastCallFns , "Number of functions converted to fastcc");
59 STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
60 STATISTIC(NumNestRemoved , "Number of nest attributes removed");
61 STATISTIC(NumAliasesResolved, "Number of global aliases resolved");
62 STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated");
63 STATISTIC(NumCXXDtorsRemoved, "Number of global C++ destructors removed");
66 struct GlobalOpt : public ModulePass {
67 void getAnalysisUsage(AnalysisUsage &AU) const override {
68 AU.addRequired<TargetLibraryInfo>();
70 static char ID; // Pass identification, replacement for typeid
71 GlobalOpt() : ModulePass(ID) {
72 initializeGlobalOptPass(*PassRegistry::getPassRegistry());
75 bool runOnModule(Module &M) override;
78 GlobalVariable *FindGlobalCtors(Module &M);
79 bool OptimizeFunctions(Module &M);
80 bool OptimizeGlobalVars(Module &M);
81 bool OptimizeGlobalAliases(Module &M);
82 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
83 bool ProcessGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
84 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI,
85 const GlobalStatus &GS);
86 bool OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn);
89 TargetLibraryInfo *TLI;
93 char GlobalOpt::ID = 0;
94 INITIALIZE_PASS_BEGIN(GlobalOpt, "globalopt",
95 "Global Variable Optimizer", false, false)
96 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
97 INITIALIZE_PASS_END(GlobalOpt, "globalopt",
98 "Global Variable Optimizer", false, false)
100 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
102 /// isLeakCheckerRoot - Is this global variable possibly used by a leak checker
103 /// as a root? If so, we might not really want to eliminate the stores to it.
104 static bool isLeakCheckerRoot(GlobalVariable *GV) {
105 // A global variable is a root if it is a pointer, or could plausibly contain
106 // a pointer. There are two challenges; one is that we could have a struct
107 // the has an inner member which is a pointer. We recurse through the type to
108 // detect these (up to a point). The other is that we may actually be a union
109 // of a pointer and another type, and so our LLVM type is an integer which
110 // gets converted into a pointer, or our type is an [i8 x #] with a pointer
111 // potentially contained here.
113 if (GV->hasPrivateLinkage())
116 SmallVector<Type *, 4> Types;
117 Types.push_back(cast<PointerType>(GV->getType())->getElementType());
121 Type *Ty = Types.pop_back_val();
122 switch (Ty->getTypeID()) {
124 case Type::PointerTyID: return true;
125 case Type::ArrayTyID:
126 case Type::VectorTyID: {
127 SequentialType *STy = cast<SequentialType>(Ty);
128 Types.push_back(STy->getElementType());
131 case Type::StructTyID: {
132 StructType *STy = cast<StructType>(Ty);
133 if (STy->isOpaque()) return true;
134 for (StructType::element_iterator I = STy->element_begin(),
135 E = STy->element_end(); I != E; ++I) {
137 if (isa<PointerType>(InnerTy)) return true;
138 if (isa<CompositeType>(InnerTy))
139 Types.push_back(InnerTy);
144 if (--Limit == 0) return true;
145 } while (!Types.empty());
149 /// Given a value that is stored to a global but never read, determine whether
150 /// it's safe to remove the store and the chain of computation that feeds the
152 static bool IsSafeComputationToRemove(Value *V, const TargetLibraryInfo *TLI) {
154 if (isa<Constant>(V))
158 if (isa<LoadInst>(V) || isa<InvokeInst>(V) || isa<Argument>(V) ||
161 if (isAllocationFn(V, TLI))
164 Instruction *I = cast<Instruction>(V);
165 if (I->mayHaveSideEffects())
167 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
168 if (!GEP->hasAllConstantIndices())
170 } else if (I->getNumOperands() != 1) {
174 V = I->getOperand(0);
178 /// CleanupPointerRootUsers - This GV is a pointer root. Loop over all users
179 /// of the global and clean up any that obviously don't assign the global a
180 /// value that isn't dynamically allocated.
182 static bool CleanupPointerRootUsers(GlobalVariable *GV,
183 const TargetLibraryInfo *TLI) {
184 // A brief explanation of leak checkers. The goal is to find bugs where
185 // pointers are forgotten, causing an accumulating growth in memory
186 // usage over time. The common strategy for leak checkers is to whitelist the
187 // memory pointed to by globals at exit. This is popular because it also
188 // solves another problem where the main thread of a C++ program may shut down
189 // before other threads that are still expecting to use those globals. To
190 // handle that case, we expect the program may create a singleton and never
193 bool Changed = false;
195 // If Dead[n].first is the only use of a malloc result, we can delete its
196 // chain of computation and the store to the global in Dead[n].second.
197 SmallVector<std::pair<Instruction *, Instruction *>, 32> Dead;
199 // Constants can't be pointers to dynamically allocated memory.
200 for (Value::user_iterator UI = GV->user_begin(), E = GV->user_end();
203 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
204 Value *V = SI->getValueOperand();
205 if (isa<Constant>(V)) {
207 SI->eraseFromParent();
208 } else if (Instruction *I = dyn_cast<Instruction>(V)) {
210 Dead.push_back(std::make_pair(I, SI));
212 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(U)) {
213 if (isa<Constant>(MSI->getValue())) {
215 MSI->eraseFromParent();
216 } else if (Instruction *I = dyn_cast<Instruction>(MSI->getValue())) {
218 Dead.push_back(std::make_pair(I, MSI));
220 } else if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(U)) {
221 GlobalVariable *MemSrc = dyn_cast<GlobalVariable>(MTI->getSource());
222 if (MemSrc && MemSrc->isConstant()) {
224 MTI->eraseFromParent();
225 } else if (Instruction *I = dyn_cast<Instruction>(MemSrc)) {
227 Dead.push_back(std::make_pair(I, MTI));
229 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
230 if (CE->use_empty()) {
231 CE->destroyConstant();
234 } else if (Constant *C = dyn_cast<Constant>(U)) {
235 if (isSafeToDestroyConstant(C)) {
236 C->destroyConstant();
237 // This could have invalidated UI, start over from scratch.
239 CleanupPointerRootUsers(GV, TLI);
245 for (int i = 0, e = Dead.size(); i != e; ++i) {
246 if (IsSafeComputationToRemove(Dead[i].first, TLI)) {
247 Dead[i].second->eraseFromParent();
248 Instruction *I = Dead[i].first;
250 if (isAllocationFn(I, TLI))
252 Instruction *J = dyn_cast<Instruction>(I->getOperand(0));
255 I->eraseFromParent();
258 I->eraseFromParent();
265 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
266 /// users of the global, cleaning up the obvious ones. This is largely just a
267 /// quick scan over the use list to clean up the easy and obvious cruft. This
268 /// returns true if it made a change.
269 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init,
270 const DataLayout *DL,
271 TargetLibraryInfo *TLI) {
272 bool Changed = false;
273 // Note that we need to use a weak value handle for the worklist items. When
274 // we delete a constant array, we may also be holding pointer to one of its
275 // elements (or an element of one of its elements if we're dealing with an
276 // array of arrays) in the worklist.
277 SmallVector<WeakVH, 8> WorkList(V->user_begin(), V->user_end());
278 while (!WorkList.empty()) {
279 Value *UV = WorkList.pop_back_val();
283 User *U = cast<User>(UV);
285 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
287 // Replace the load with the initializer.
288 LI->replaceAllUsesWith(Init);
289 LI->eraseFromParent();
292 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
293 // Store must be unreachable or storing Init into the global.
294 SI->eraseFromParent();
296 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
297 if (CE->getOpcode() == Instruction::GetElementPtr) {
298 Constant *SubInit = 0;
300 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
301 Changed |= CleanupConstantGlobalUsers(CE, SubInit, DL, TLI);
302 } else if ((CE->getOpcode() == Instruction::BitCast &&
303 CE->getType()->isPointerTy()) ||
304 CE->getOpcode() == Instruction::AddrSpaceCast) {
305 // Pointer cast, delete any stores and memsets to the global.
306 Changed |= CleanupConstantGlobalUsers(CE, 0, DL, TLI);
309 if (CE->use_empty()) {
310 CE->destroyConstant();
313 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
314 // Do not transform "gepinst (gep constexpr (GV))" here, because forming
315 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
316 // and will invalidate our notion of what Init is.
317 Constant *SubInit = 0;
318 if (!isa<ConstantExpr>(GEP->getOperand(0))) {
320 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP, DL, TLI));
321 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
322 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
324 // If the initializer is an all-null value and we have an inbounds GEP,
325 // we already know what the result of any load from that GEP is.
326 // TODO: Handle splats.
327 if (Init && isa<ConstantAggregateZero>(Init) && GEP->isInBounds())
328 SubInit = Constant::getNullValue(GEP->getType()->getElementType());
330 Changed |= CleanupConstantGlobalUsers(GEP, SubInit, DL, TLI);
332 if (GEP->use_empty()) {
333 GEP->eraseFromParent();
336 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
337 if (MI->getRawDest() == V) {
338 MI->eraseFromParent();
342 } else if (Constant *C = dyn_cast<Constant>(U)) {
343 // If we have a chain of dead constantexprs or other things dangling from
344 // us, and if they are all dead, nuke them without remorse.
345 if (isSafeToDestroyConstant(C)) {
346 C->destroyConstant();
347 CleanupConstantGlobalUsers(V, Init, DL, TLI);
355 /// isSafeSROAElementUse - Return true if the specified instruction is a safe
356 /// user of a derived expression from a global that we want to SROA.
357 static bool isSafeSROAElementUse(Value *V) {
358 // We might have a dead and dangling constant hanging off of here.
359 if (Constant *C = dyn_cast<Constant>(V))
360 return isSafeToDestroyConstant(C);
362 Instruction *I = dyn_cast<Instruction>(V);
363 if (!I) return false;
366 if (isa<LoadInst>(I)) return true;
368 // Stores *to* the pointer are ok.
369 if (StoreInst *SI = dyn_cast<StoreInst>(I))
370 return SI->getOperand(0) != V;
372 // Otherwise, it must be a GEP.
373 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I);
374 if (GEPI == 0) return false;
376 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
377 !cast<Constant>(GEPI->getOperand(1))->isNullValue())
380 for (User *U : GEPI->users())
381 if (!isSafeSROAElementUse(U))
387 /// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value.
388 /// Look at it and its uses and decide whether it is safe to SROA this global.
390 static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
391 // The user of the global must be a GEP Inst or a ConstantExpr GEP.
392 if (!isa<GetElementPtrInst>(U) &&
393 (!isa<ConstantExpr>(U) ||
394 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
397 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
398 // don't like < 3 operand CE's, and we don't like non-constant integer
399 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some
401 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
402 !cast<Constant>(U->getOperand(1))->isNullValue() ||
403 !isa<ConstantInt>(U->getOperand(2)))
406 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
407 ++GEPI; // Skip over the pointer index.
409 // If this is a use of an array allocation, do a bit more checking for sanity.
410 if (ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
411 uint64_t NumElements = AT->getNumElements();
412 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
414 // Check to make sure that index falls within the array. If not,
415 // something funny is going on, so we won't do the optimization.
417 if (Idx->getZExtValue() >= NumElements)
420 // We cannot scalar repl this level of the array unless any array
421 // sub-indices are in-range constants. In particular, consider:
422 // A[0][i]. We cannot know that the user isn't doing invalid things like
423 // allowing i to index an out-of-range subscript that accesses A[1].
425 // Scalar replacing *just* the outer index of the array is probably not
426 // going to be a win anyway, so just give up.
427 for (++GEPI; // Skip array index.
430 uint64_t NumElements;
431 if (ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
432 NumElements = SubArrayTy->getNumElements();
433 else if (VectorType *SubVectorTy = dyn_cast<VectorType>(*GEPI))
434 NumElements = SubVectorTy->getNumElements();
436 assert((*GEPI)->isStructTy() &&
437 "Indexed GEP type is not array, vector, or struct!");
441 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
442 if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
447 for (User *UU : U->users())
448 if (!isSafeSROAElementUse(UU))
454 /// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
455 /// is safe for us to perform this transformation.
457 static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
458 for (User *U : GV->users())
459 if (!IsUserOfGlobalSafeForSRA(U, GV))
466 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
467 /// variable. This opens the door for other optimizations by exposing the
468 /// behavior of the program in a more fine-grained way. We have determined that
469 /// this transformation is safe already. We return the first global variable we
470 /// insert so that the caller can reprocess it.
471 static GlobalVariable *SRAGlobal(GlobalVariable *GV, const DataLayout &DL) {
472 // Make sure this global only has simple uses that we can SRA.
473 if (!GlobalUsersSafeToSRA(GV))
476 assert(GV->hasLocalLinkage() && !GV->isConstant());
477 Constant *Init = GV->getInitializer();
478 Type *Ty = Init->getType();
480 std::vector<GlobalVariable*> NewGlobals;
481 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
483 // Get the alignment of the global, either explicit or target-specific.
484 unsigned StartAlignment = GV->getAlignment();
485 if (StartAlignment == 0)
486 StartAlignment = DL.getABITypeAlignment(GV->getType());
488 if (StructType *STy = dyn_cast<StructType>(Ty)) {
489 NewGlobals.reserve(STy->getNumElements());
490 const StructLayout &Layout = *DL.getStructLayout(STy);
491 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
492 Constant *In = Init->getAggregateElement(i);
493 assert(In && "Couldn't get element of initializer?");
494 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
495 GlobalVariable::InternalLinkage,
496 In, GV->getName()+"."+Twine(i),
497 GV->getThreadLocalMode(),
498 GV->getType()->getAddressSpace());
499 Globals.insert(GV, NGV);
500 NewGlobals.push_back(NGV);
502 // Calculate the known alignment of the field. If the original aggregate
503 // had 256 byte alignment for example, something might depend on that:
504 // propagate info to each field.
505 uint64_t FieldOffset = Layout.getElementOffset(i);
506 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset);
507 if (NewAlign > DL.getABITypeAlignment(STy->getElementType(i)))
508 NGV->setAlignment(NewAlign);
510 } else if (SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
511 unsigned NumElements = 0;
512 if (ArrayType *ATy = dyn_cast<ArrayType>(STy))
513 NumElements = ATy->getNumElements();
515 NumElements = cast<VectorType>(STy)->getNumElements();
517 if (NumElements > 16 && GV->hasNUsesOrMore(16))
518 return 0; // It's not worth it.
519 NewGlobals.reserve(NumElements);
521 uint64_t EltSize = DL.getTypeAllocSize(STy->getElementType());
522 unsigned EltAlign = DL.getABITypeAlignment(STy->getElementType());
523 for (unsigned i = 0, e = NumElements; i != e; ++i) {
524 Constant *In = Init->getAggregateElement(i);
525 assert(In && "Couldn't get element of initializer?");
527 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
528 GlobalVariable::InternalLinkage,
529 In, GV->getName()+"."+Twine(i),
530 GV->getThreadLocalMode(),
531 GV->getType()->getAddressSpace());
532 Globals.insert(GV, NGV);
533 NewGlobals.push_back(NGV);
535 // Calculate the known alignment of the field. If the original aggregate
536 // had 256 byte alignment for example, something might depend on that:
537 // propagate info to each field.
538 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
539 if (NewAlign > EltAlign)
540 NGV->setAlignment(NewAlign);
544 if (NewGlobals.empty())
547 DEBUG(dbgs() << "PERFORMING GLOBAL SRA ON: " << *GV);
549 Constant *NullInt =Constant::getNullValue(Type::getInt32Ty(GV->getContext()));
551 // Loop over all of the uses of the global, replacing the constantexpr geps,
552 // with smaller constantexpr geps or direct references.
553 while (!GV->use_empty()) {
554 User *GEP = GV->user_back();
555 assert(((isa<ConstantExpr>(GEP) &&
556 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
557 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
559 // Ignore the 1th operand, which has to be zero or else the program is quite
560 // broken (undefined). Get the 2nd operand, which is the structure or array
562 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
563 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
565 Value *NewPtr = NewGlobals[Val];
567 // Form a shorter GEP if needed.
568 if (GEP->getNumOperands() > 3) {
569 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
570 SmallVector<Constant*, 8> Idxs;
571 Idxs.push_back(NullInt);
572 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
573 Idxs.push_back(CE->getOperand(i));
574 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr), Idxs);
576 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
577 SmallVector<Value*, 8> Idxs;
578 Idxs.push_back(NullInt);
579 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
580 Idxs.push_back(GEPI->getOperand(i));
581 NewPtr = GetElementPtrInst::Create(NewPtr, Idxs,
582 GEPI->getName()+"."+Twine(Val),GEPI);
585 GEP->replaceAllUsesWith(NewPtr);
587 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
588 GEPI->eraseFromParent();
590 cast<ConstantExpr>(GEP)->destroyConstant();
593 // Delete the old global, now that it is dead.
597 // Loop over the new globals array deleting any globals that are obviously
598 // dead. This can arise due to scalarization of a structure or an array that
599 // has elements that are dead.
600 unsigned FirstGlobal = 0;
601 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
602 if (NewGlobals[i]->use_empty()) {
603 Globals.erase(NewGlobals[i]);
604 if (FirstGlobal == i) ++FirstGlobal;
607 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
610 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
611 /// value will trap if the value is dynamically null. PHIs keeps track of any
612 /// phi nodes we've seen to avoid reprocessing them.
613 static bool AllUsesOfValueWillTrapIfNull(const Value *V,
614 SmallPtrSet<const PHINode*, 8> &PHIs) {
615 for (const User *U : V->users())
616 if (isa<LoadInst>(U)) {
618 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
619 if (SI->getOperand(0) == V) {
620 //cerr << "NONTRAPPING USE: " << *U;
621 return false; // Storing the value.
623 } else if (const CallInst *CI = dyn_cast<CallInst>(U)) {
624 if (CI->getCalledValue() != V) {
625 //cerr << "NONTRAPPING USE: " << *U;
626 return false; // Not calling the ptr
628 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(U)) {
629 if (II->getCalledValue() != V) {
630 //cerr << "NONTRAPPING USE: " << *U;
631 return false; // Not calling the ptr
633 } else if (const BitCastInst *CI = dyn_cast<BitCastInst>(U)) {
634 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
635 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
636 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
637 } else if (const PHINode *PN = dyn_cast<PHINode>(U)) {
638 // If we've already seen this phi node, ignore it, it has already been
640 if (PHIs.insert(PN) && !AllUsesOfValueWillTrapIfNull(PN, PHIs))
642 } else if (isa<ICmpInst>(U) &&
643 isa<ConstantPointerNull>(U->getOperand(1))) {
644 // Ignore icmp X, null
646 //cerr << "NONTRAPPING USE: " << *U;
653 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
654 /// from GV will trap if the loaded value is null. Note that this also permits
655 /// comparisons of the loaded value against null, as a special case.
656 static bool AllUsesOfLoadedValueWillTrapIfNull(const GlobalVariable *GV) {
657 for (const User *U : GV->users())
658 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
659 SmallPtrSet<const PHINode*, 8> PHIs;
660 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
662 } else if (isa<StoreInst>(U)) {
663 // Ignore stores to the global.
665 // We don't know or understand this user, bail out.
666 //cerr << "UNKNOWN USER OF GLOBAL!: " << *U;
672 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
673 bool Changed = false;
674 for (auto UI = V->user_begin(), E = V->user_end(); UI != E; ) {
675 Instruction *I = cast<Instruction>(*UI++);
676 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
677 LI->setOperand(0, NewV);
679 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
680 if (SI->getOperand(1) == V) {
681 SI->setOperand(1, NewV);
684 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
686 if (CS.getCalledValue() == V) {
687 // Calling through the pointer! Turn into a direct call, but be careful
688 // that the pointer is not also being passed as an argument.
689 CS.setCalledFunction(NewV);
691 bool PassedAsArg = false;
692 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
693 if (CS.getArgument(i) == V) {
695 CS.setArgument(i, NewV);
699 // Being passed as an argument also. Be careful to not invalidate UI!
700 UI = V->user_begin();
703 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
704 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
705 ConstantExpr::getCast(CI->getOpcode(),
706 NewV, CI->getType()));
707 if (CI->use_empty()) {
709 CI->eraseFromParent();
711 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
712 // Should handle GEP here.
713 SmallVector<Constant*, 8> Idxs;
714 Idxs.reserve(GEPI->getNumOperands()-1);
715 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
717 if (Constant *C = dyn_cast<Constant>(*i))
721 if (Idxs.size() == GEPI->getNumOperands()-1)
722 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
723 ConstantExpr::getGetElementPtr(NewV, Idxs));
724 if (GEPI->use_empty()) {
726 GEPI->eraseFromParent();
735 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
736 /// value stored into it. If there are uses of the loaded value that would trap
737 /// if the loaded value is dynamically null, then we know that they cannot be
738 /// reachable with a null optimize away the load.
739 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV,
740 const DataLayout *DL,
741 TargetLibraryInfo *TLI) {
742 bool Changed = false;
744 // Keep track of whether we are able to remove all the uses of the global
745 // other than the store that defines it.
746 bool AllNonStoreUsesGone = true;
748 // Replace all uses of loads with uses of uses of the stored value.
749 for (Value::user_iterator GUI = GV->user_begin(), E = GV->user_end(); GUI != E;){
750 User *GlobalUser = *GUI++;
751 if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) {
752 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
753 // If we were able to delete all uses of the loads
754 if (LI->use_empty()) {
755 LI->eraseFromParent();
758 AllNonStoreUsesGone = false;
760 } else if (isa<StoreInst>(GlobalUser)) {
761 // Ignore the store that stores "LV" to the global.
762 assert(GlobalUser->getOperand(1) == GV &&
763 "Must be storing *to* the global");
765 AllNonStoreUsesGone = false;
767 // If we get here we could have other crazy uses that are transitively
769 assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) ||
770 isa<ConstantExpr>(GlobalUser) || isa<CmpInst>(GlobalUser) ||
771 isa<BitCastInst>(GlobalUser) ||
772 isa<GetElementPtrInst>(GlobalUser)) &&
773 "Only expect load and stores!");
778 DEBUG(dbgs() << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV);
782 // If we nuked all of the loads, then none of the stores are needed either,
783 // nor is the global.
784 if (AllNonStoreUsesGone) {
785 if (isLeakCheckerRoot(GV)) {
786 Changed |= CleanupPointerRootUsers(GV, TLI);
789 CleanupConstantGlobalUsers(GV, 0, DL, TLI);
791 if (GV->use_empty()) {
792 DEBUG(dbgs() << " *** GLOBAL NOW DEAD!\n");
794 GV->eraseFromParent();
801 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
802 /// instructions that are foldable.
803 static void ConstantPropUsersOf(Value *V, const DataLayout *DL,
804 TargetLibraryInfo *TLI) {
805 for (Value::user_iterator UI = V->user_begin(), E = V->user_end(); UI != E; )
806 if (Instruction *I = dyn_cast<Instruction>(*UI++))
807 if (Constant *NewC = ConstantFoldInstruction(I, DL, TLI)) {
808 I->replaceAllUsesWith(NewC);
810 // Advance UI to the next non-I use to avoid invalidating it!
811 // Instructions could multiply use V.
812 while (UI != E && *UI == I)
814 I->eraseFromParent();
818 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
819 /// variable, and transforms the program as if it always contained the result of
820 /// the specified malloc. Because it is always the result of the specified
821 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
822 /// malloc into a global, and any loads of GV as uses of the new global.
823 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
826 ConstantInt *NElements,
827 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 = 0;
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 SmallPtrSet<const PHINode*, 8> &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
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 SmallPtrSet<const PHINode*, 32> &LoadUsingPHIs,
1050 SmallPtrSet<const PHINode*, 32> &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))
1075 // This means some phi nodes are dependent on each other.
1076 // Avoid infinite looping!
1078 if (!LoadUsingPHIs.insert(PN))
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 (SmallPtrSet<const PHINode*, 32>::const_iterator I = LoadUsingPHIs.begin()
1118 , E = LoadUsingPHIs.end(); I != E; ++I) {
1119 const PHINode *PN = *I;
1120 for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) {
1121 Value *InVal = PN->getIncomingValue(op);
1123 // PHI of the stored value itself is ok.
1124 if (InVal == StoredVal) continue;
1126 if (const PHINode *InPN = dyn_cast<PHINode>(InVal)) {
1127 // One of the PHIs in our set is (optimistically) ok.
1128 if (LoadUsingPHIs.count(InPN))
1133 // Load from GV is ok.
1134 if (const LoadInst *LI = dyn_cast<LoadInst>(InVal))
1135 if (LI->getOperand(0) == GV)
1140 // Anything else is rejected.
1148 static Value *GetHeapSROAValue(Value *V, unsigned FieldNo,
1149 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1150 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1151 std::vector<Value*> &FieldVals = InsertedScalarizedValues[V];
1153 if (FieldNo >= FieldVals.size())
1154 FieldVals.resize(FieldNo+1);
1156 // If we already have this value, just reuse the previously scalarized
1158 if (Value *FieldVal = FieldVals[FieldNo])
1161 // Depending on what instruction this is, we have several cases.
1163 if (LoadInst *LI = dyn_cast<LoadInst>(V)) {
1164 // This is a scalarized version of the load from the global. Just create
1165 // a new Load of the scalarized global.
1166 Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo,
1167 InsertedScalarizedValues,
1169 LI->getName()+".f"+Twine(FieldNo), LI);
1170 } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1171 // PN's type is pointer to struct. Make a new PHI of pointer to struct
1174 PointerType *PTy = cast<PointerType>(PN->getType());
1175 StructType *ST = cast<StructType>(PTy->getElementType());
1177 unsigned AS = PTy->getAddressSpace();
1179 PHINode::Create(PointerType::get(ST->getElementType(FieldNo), AS),
1180 PN->getNumIncomingValues(),
1181 PN->getName()+".f"+Twine(FieldNo), PN);
1183 PHIsToRewrite.push_back(std::make_pair(PN, FieldNo));
1185 llvm_unreachable("Unknown usable value");
1188 return FieldVals[FieldNo] = Result;
1191 /// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1192 /// the load, rewrite the derived value to use the HeapSRoA'd load.
1193 static void RewriteHeapSROALoadUser(Instruction *LoadUser,
1194 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1195 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1196 // If this is a comparison against null, handle it.
1197 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1198 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1199 // If we have a setcc of the loaded pointer, we can use a setcc of any
1201 Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0,
1202 InsertedScalarizedValues, PHIsToRewrite);
1204 Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr,
1205 Constant::getNullValue(NPtr->getType()),
1207 SCI->replaceAllUsesWith(New);
1208 SCI->eraseFromParent();
1212 // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...'
1213 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1214 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1215 && "Unexpected GEPI!");
1217 // Load the pointer for this field.
1218 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1219 Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo,
1220 InsertedScalarizedValues, PHIsToRewrite);
1222 // Create the new GEP idx vector.
1223 SmallVector<Value*, 8> GEPIdx;
1224 GEPIdx.push_back(GEPI->getOperand(1));
1225 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1227 Value *NGEPI = GetElementPtrInst::Create(NewPtr, GEPIdx,
1228 GEPI->getName(), GEPI);
1229 GEPI->replaceAllUsesWith(NGEPI);
1230 GEPI->eraseFromParent();
1234 // Recursively transform the users of PHI nodes. This will lazily create the
1235 // PHIs that are needed for individual elements. Keep track of what PHIs we
1236 // see in InsertedScalarizedValues so that we don't get infinite loops (very
1237 // antisocial). If the PHI is already in InsertedScalarizedValues, it has
1238 // already been seen first by another load, so its uses have already been
1240 PHINode *PN = cast<PHINode>(LoadUser);
1241 if (!InsertedScalarizedValues.insert(std::make_pair(PN,
1242 std::vector<Value*>())).second)
1245 // If this is the first time we've seen this PHI, recursively process all
1247 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
1248 Instruction *User = cast<Instruction>(*UI++);
1249 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1253 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1254 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1255 /// use FieldGlobals instead. All uses of loaded values satisfy
1256 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA.
1257 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1258 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1259 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1260 for (auto UI = Load->user_begin(), E = Load->user_end(); UI != E;) {
1261 Instruction *User = cast<Instruction>(*UI++);
1262 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1265 if (Load->use_empty()) {
1266 Load->eraseFromParent();
1267 InsertedScalarizedValues.erase(Load);
1271 /// PerformHeapAllocSRoA - CI is an allocation of an array of structures. Break
1272 /// it up into multiple allocations of arrays of the fields.
1273 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI,
1274 Value *NElems, const DataLayout *DL,
1275 const TargetLibraryInfo *TLI) {
1276 DEBUG(dbgs() << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *CI << '\n');
1277 Type *MAT = getMallocAllocatedType(CI, TLI);
1278 StructType *STy = cast<StructType>(MAT);
1280 // There is guaranteed to be at least one use of the malloc (storing
1281 // it into GV). If there are other uses, change them to be uses of
1282 // the global to simplify later code. This also deletes the store
1284 ReplaceUsesOfMallocWithGlobal(CI, GV);
1286 // Okay, at this point, there are no users of the malloc. Insert N
1287 // new mallocs at the same place as CI, and N globals.
1288 std::vector<Value*> FieldGlobals;
1289 std::vector<Value*> FieldMallocs;
1291 unsigned AS = GV->getType()->getPointerAddressSpace();
1292 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1293 Type *FieldTy = STy->getElementType(FieldNo);
1294 PointerType *PFieldTy = PointerType::get(FieldTy, AS);
1296 GlobalVariable *NGV =
1297 new GlobalVariable(*GV->getParent(),
1298 PFieldTy, false, GlobalValue::InternalLinkage,
1299 Constant::getNullValue(PFieldTy),
1300 GV->getName() + ".f" + Twine(FieldNo), GV,
1301 GV->getThreadLocalMode());
1302 FieldGlobals.push_back(NGV);
1304 unsigned TypeSize = DL->getTypeAllocSize(FieldTy);
1305 if (StructType *ST = dyn_cast<StructType>(FieldTy))
1306 TypeSize = DL->getStructLayout(ST)->getSizeInBytes();
1307 Type *IntPtrTy = DL->getIntPtrType(CI->getType());
1308 Value *NMI = CallInst::CreateMalloc(CI, IntPtrTy, FieldTy,
1309 ConstantInt::get(IntPtrTy, TypeSize),
1311 CI->getName() + ".f" + Twine(FieldNo));
1312 FieldMallocs.push_back(NMI);
1313 new StoreInst(NMI, NGV, CI);
1316 // The tricky aspect of this transformation is handling the case when malloc
1317 // fails. In the original code, malloc failing would set the result pointer
1318 // of malloc to null. In this case, some mallocs could succeed and others
1319 // could fail. As such, we emit code that looks like this:
1320 // F0 = malloc(field0)
1321 // F1 = malloc(field1)
1322 // F2 = malloc(field2)
1323 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1324 // if (F0) { free(F0); F0 = 0; }
1325 // if (F1) { free(F1); F1 = 0; }
1326 // if (F2) { free(F2); F2 = 0; }
1328 // The malloc can also fail if its argument is too large.
1329 Constant *ConstantZero = ConstantInt::get(CI->getArgOperand(0)->getType(), 0);
1330 Value *RunningOr = new ICmpInst(CI, ICmpInst::ICMP_SLT, CI->getArgOperand(0),
1331 ConstantZero, "isneg");
1332 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1333 Value *Cond = new ICmpInst(CI, ICmpInst::ICMP_EQ, FieldMallocs[i],
1334 Constant::getNullValue(FieldMallocs[i]->getType()),
1336 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", CI);
1339 // Split the basic block at the old malloc.
1340 BasicBlock *OrigBB = CI->getParent();
1341 BasicBlock *ContBB = OrigBB->splitBasicBlock(CI, "malloc_cont");
1343 // Create the block to check the first condition. Put all these blocks at the
1344 // end of the function as they are unlikely to be executed.
1345 BasicBlock *NullPtrBlock = BasicBlock::Create(OrigBB->getContext(),
1347 OrigBB->getParent());
1349 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1350 // branch on RunningOr.
1351 OrigBB->getTerminator()->eraseFromParent();
1352 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1354 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1355 // pointer, because some may be null while others are not.
1356 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1357 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1358 Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
1359 Constant::getNullValue(GVVal->getType()));
1360 BasicBlock *FreeBlock = BasicBlock::Create(Cmp->getContext(), "free_it",
1361 OrigBB->getParent());
1362 BasicBlock *NextBlock = BasicBlock::Create(Cmp->getContext(), "next",
1363 OrigBB->getParent());
1364 Instruction *BI = BranchInst::Create(FreeBlock, NextBlock,
1367 // Fill in FreeBlock.
1368 CallInst::CreateFree(GVVal, BI);
1369 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1371 BranchInst::Create(NextBlock, FreeBlock);
1373 NullPtrBlock = NextBlock;
1376 BranchInst::Create(ContBB, NullPtrBlock);
1378 // CI is no longer needed, remove it.
1379 CI->eraseFromParent();
1381 /// InsertedScalarizedLoads - As we process loads, if we can't immediately
1382 /// update all uses of the load, keep track of what scalarized loads are
1383 /// inserted for a given load.
1384 DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues;
1385 InsertedScalarizedValues[GV] = FieldGlobals;
1387 std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite;
1389 // Okay, the malloc site is completely handled. All of the uses of GV are now
1390 // loads, and all uses of those loads are simple. Rewrite them to use loads
1391 // of the per-field globals instead.
1392 for (auto UI = GV->user_begin(), E = GV->user_end(); UI != E;) {
1393 Instruction *User = cast<Instruction>(*UI++);
1395 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1396 RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite);
1400 // Must be a store of null.
1401 StoreInst *SI = cast<StoreInst>(User);
1402 assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
1403 "Unexpected heap-sra user!");
1405 // Insert a store of null into each global.
1406 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1407 PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType());
1408 Constant *Null = Constant::getNullValue(PT->getElementType());
1409 new StoreInst(Null, FieldGlobals[i], SI);
1411 // Erase the original store.
1412 SI->eraseFromParent();
1415 // While we have PHIs that are interesting to rewrite, do it.
1416 while (!PHIsToRewrite.empty()) {
1417 PHINode *PN = PHIsToRewrite.back().first;
1418 unsigned FieldNo = PHIsToRewrite.back().second;
1419 PHIsToRewrite.pop_back();
1420 PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]);
1421 assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi");
1423 // Add all the incoming values. This can materialize more phis.
1424 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1425 Value *InVal = PN->getIncomingValue(i);
1426 InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues,
1428 FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
1432 // Drop all inter-phi links and any loads that made it this far.
1433 for (DenseMap<Value*, std::vector<Value*> >::iterator
1434 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1436 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1437 PN->dropAllReferences();
1438 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1439 LI->dropAllReferences();
1442 // Delete all the phis and loads now that inter-references are dead.
1443 for (DenseMap<Value*, std::vector<Value*> >::iterator
1444 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1446 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1447 PN->eraseFromParent();
1448 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1449 LI->eraseFromParent();
1452 // The old global is now dead, remove it.
1453 GV->eraseFromParent();
1456 return cast<GlobalVariable>(FieldGlobals[0]);
1459 /// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1460 /// pointer global variable with a single value stored it that is a malloc or
1462 static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
1465 AtomicOrdering Ordering,
1466 Module::global_iterator &GVI,
1467 const DataLayout *DL,
1468 TargetLibraryInfo *TLI) {
1472 // If this is a malloc of an abstract type, don't touch it.
1473 if (!AllocTy->isSized())
1476 // We can't optimize this global unless all uses of it are *known* to be
1477 // of the malloc value, not of the null initializer value (consider a use
1478 // that compares the global's value against zero to see if the malloc has
1479 // been reached). To do this, we check to see if all uses of the global
1480 // would trap if the global were null: this proves that they must all
1481 // happen after the malloc.
1482 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1485 // We can't optimize this if the malloc itself is used in a complex way,
1486 // for example, being stored into multiple globals. This allows the
1487 // malloc to be stored into the specified global, loaded icmp'd, and
1488 // GEP'd. These are all things we could transform to using the global
1490 SmallPtrSet<const PHINode*, 8> PHIs;
1491 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV, PHIs))
1494 // If we have a global that is only initialized with a fixed size malloc,
1495 // transform the program to use global memory instead of malloc'd memory.
1496 // This eliminates dynamic allocation, avoids an indirection accessing the
1497 // data, and exposes the resultant global to further GlobalOpt.
1498 // We cannot optimize the malloc if we cannot determine malloc array size.
1499 Value *NElems = getMallocArraySize(CI, DL, TLI, true);
1503 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
1504 // Restrict this transformation to only working on small allocations
1505 // (2048 bytes currently), as we don't want to introduce a 16M global or
1507 if (NElements->getZExtValue() * DL->getTypeAllocSize(AllocTy) < 2048) {
1508 GVI = OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, DL, TLI);
1512 // If the allocation is an array of structures, consider transforming this
1513 // into multiple malloc'd arrays, one for each field. This is basically
1514 // SRoA for malloc'd memory.
1516 if (Ordering != NotAtomic)
1519 // If this is an allocation of a fixed size array of structs, analyze as a
1520 // variable size array. malloc [100 x struct],1 -> malloc struct, 100
1521 if (NElems == ConstantInt::get(CI->getArgOperand(0)->getType(), 1))
1522 if (ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
1523 AllocTy = AT->getElementType();
1525 StructType *AllocSTy = dyn_cast<StructType>(AllocTy);
1529 // This the structure has an unreasonable number of fields, leave it
1531 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
1532 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, CI)) {
1534 // If this is a fixed size array, transform the Malloc to be an alloc of
1535 // structs. malloc [100 x struct],1 -> malloc struct, 100
1536 if (ArrayType *AT = dyn_cast<ArrayType>(getMallocAllocatedType(CI, TLI))) {
1537 Type *IntPtrTy = DL->getIntPtrType(CI->getType());
1538 unsigned TypeSize = DL->getStructLayout(AllocSTy)->getSizeInBytes();
1539 Value *AllocSize = ConstantInt::get(IntPtrTy, TypeSize);
1540 Value *NumElements = ConstantInt::get(IntPtrTy, AT->getNumElements());
1541 Instruction *Malloc = CallInst::CreateMalloc(CI, IntPtrTy, AllocSTy,
1542 AllocSize, NumElements,
1544 Instruction *Cast = new BitCastInst(Malloc, CI->getType(), "tmp", CI);
1545 CI->replaceAllUsesWith(Cast);
1546 CI->eraseFromParent();
1547 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Malloc))
1548 CI = cast<CallInst>(BCI->getOperand(0));
1550 CI = cast<CallInst>(Malloc);
1553 GVI = PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, DL, TLI, true),
1561 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1562 // that only one value (besides its initializer) is ever stored to the global.
1563 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1564 AtomicOrdering Ordering,
1565 Module::global_iterator &GVI,
1566 const DataLayout *DL,
1567 TargetLibraryInfo *TLI) {
1568 // Ignore no-op GEPs and bitcasts.
1569 StoredOnceVal = StoredOnceVal->stripPointerCasts();
1571 // If we are dealing with a pointer global that is initialized to null and
1572 // only has one (non-null) value stored into it, then we can optimize any
1573 // users of the loaded value (often calls and loads) that would trap if the
1575 if (GV->getInitializer()->getType()->isPointerTy() &&
1576 GV->getInitializer()->isNullValue()) {
1577 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1578 if (GV->getInitializer()->getType() != SOVC->getType())
1579 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1581 // Optimize away any trapping uses of the loaded value.
1582 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC, DL, TLI))
1584 } else if (CallInst *CI = extractMallocCall(StoredOnceVal, TLI)) {
1585 Type *MallocType = getMallocAllocatedType(CI, TLI);
1587 TryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType, Ordering, GVI,
1596 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1597 /// two values ever stored into GV are its initializer and OtherVal. See if we
1598 /// can shrink the global into a boolean and select between the two values
1599 /// whenever it is used. This exposes the values to other scalar optimizations.
1600 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1601 Type *GVElType = GV->getType()->getElementType();
1603 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1604 // an FP value, pointer or vector, don't do this optimization because a select
1605 // between them is very expensive and unlikely to lead to later
1606 // simplification. In these cases, we typically end up with "cond ? v1 : v2"
1607 // where v1 and v2 both require constant pool loads, a big loss.
1608 if (GVElType == Type::getInt1Ty(GV->getContext()) ||
1609 GVElType->isFloatingPointTy() ||
1610 GVElType->isPointerTy() || GVElType->isVectorTy())
1613 // Walk the use list of the global seeing if all the uses are load or store.
1614 // If there is anything else, bail out.
1615 for (User *U : GV->users())
1616 if (!isa<LoadInst>(U) && !isa<StoreInst>(U))
1619 DEBUG(dbgs() << " *** SHRINKING TO BOOL: " << *GV);
1621 // Create the new global, initializing it to false.
1622 GlobalVariable *NewGV = new GlobalVariable(Type::getInt1Ty(GV->getContext()),
1624 GlobalValue::InternalLinkage,
1625 ConstantInt::getFalse(GV->getContext()),
1627 GV->getThreadLocalMode(),
1628 GV->getType()->getAddressSpace());
1629 GV->getParent()->getGlobalList().insert(GV, NewGV);
1631 Constant *InitVal = GV->getInitializer();
1632 assert(InitVal->getType() != Type::getInt1Ty(GV->getContext()) &&
1633 "No reason to shrink to bool!");
1635 // If initialized to zero and storing one into the global, we can use a cast
1636 // instead of a select to synthesize the desired value.
1637 bool IsOneZero = false;
1638 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1639 IsOneZero = InitVal->isNullValue() && CI->isOne();
1641 while (!GV->use_empty()) {
1642 Instruction *UI = cast<Instruction>(GV->user_back());
1643 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1644 // Change the store into a boolean store.
1645 bool StoringOther = SI->getOperand(0) == OtherVal;
1646 // Only do this if we weren't storing a loaded value.
1648 if (StoringOther || SI->getOperand(0) == InitVal) {
1649 StoreVal = ConstantInt::get(Type::getInt1Ty(GV->getContext()),
1652 // Otherwise, we are storing a previously loaded copy. To do this,
1653 // change the copy from copying the original value to just copying the
1655 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1657 // If we've already replaced the input, StoredVal will be a cast or
1658 // select instruction. If not, it will be a load of the original
1660 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1661 assert(LI->getOperand(0) == GV && "Not a copy!");
1662 // Insert a new load, to preserve the saved value.
1663 StoreVal = new LoadInst(NewGV, LI->getName()+".b", false, 0,
1664 LI->getOrdering(), LI->getSynchScope(), LI);
1666 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1667 "This is not a form that we understand!");
1668 StoreVal = StoredVal->getOperand(0);
1669 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1672 new StoreInst(StoreVal, NewGV, false, 0,
1673 SI->getOrdering(), SI->getSynchScope(), SI);
1675 // Change the load into a load of bool then a select.
1676 LoadInst *LI = cast<LoadInst>(UI);
1677 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", false, 0,
1678 LI->getOrdering(), LI->getSynchScope(), LI);
1681 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1683 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1685 LI->replaceAllUsesWith(NSI);
1687 UI->eraseFromParent();
1690 // Retain the name of the old global variable. People who are debugging their
1691 // programs may expect these variables to be named the same.
1692 NewGV->takeName(GV);
1693 GV->eraseFromParent();
1698 /// ProcessGlobal - Analyze the specified global variable and optimize it if
1699 /// possible. If we make a change, return true.
1700 bool GlobalOpt::ProcessGlobal(GlobalVariable *GV,
1701 Module::global_iterator &GVI) {
1702 if (!GV->isDiscardableIfUnused())
1705 // Do more involved optimizations if the global is internal.
1706 GV->removeDeadConstantUsers();
1708 if (GV->use_empty()) {
1709 DEBUG(dbgs() << "GLOBAL DEAD: " << *GV);
1710 GV->eraseFromParent();
1715 if (!GV->hasLocalLinkage())
1720 if (GlobalStatus::analyzeGlobal(GV, GS))
1723 if (!GS.IsCompared && !GV->hasUnnamedAddr()) {
1724 GV->setUnnamedAddr(true);
1728 if (GV->isConstant() || !GV->hasInitializer())
1731 return ProcessInternalGlobal(GV, GVI, GS);
1734 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1735 /// it if possible. If we make a change, return true.
1736 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1737 Module::global_iterator &GVI,
1738 const GlobalStatus &GS) {
1739 // If this is a first class global and has only one accessing function
1740 // and this function is main (which we know is not recursive), we replace
1741 // the global with a local alloca in this function.
1743 // NOTE: It doesn't make sense to promote non-single-value types since we
1744 // are just replacing static memory to stack memory.
1746 // If the global is in different address space, don't bring it to stack.
1747 if (!GS.HasMultipleAccessingFunctions &&
1748 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1749 GV->getType()->getElementType()->isSingleValueType() &&
1750 GS.AccessingFunction->getName() == "main" &&
1751 GS.AccessingFunction->hasExternalLinkage() &&
1752 GV->getType()->getAddressSpace() == 0) {
1753 DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV);
1754 Instruction &FirstI = const_cast<Instruction&>(*GS.AccessingFunction
1755 ->getEntryBlock().begin());
1756 Type *ElemTy = GV->getType()->getElementType();
1757 // FIXME: Pass Global's alignment when globals have alignment
1758 AllocaInst *Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), &FirstI);
1759 if (!isa<UndefValue>(GV->getInitializer()))
1760 new StoreInst(GV->getInitializer(), Alloca, &FirstI);
1762 GV->replaceAllUsesWith(Alloca);
1763 GV->eraseFromParent();
1768 // If the global is never loaded (but may be stored to), it is dead.
1771 DEBUG(dbgs() << "GLOBAL NEVER LOADED: " << *GV);
1774 if (isLeakCheckerRoot(GV)) {
1775 // Delete any constant stores to the global.
1776 Changed = CleanupPointerRootUsers(GV, TLI);
1778 // Delete any stores we can find to the global. We may not be able to
1779 // make it completely dead though.
1780 Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, TLI);
1783 // If the global is dead now, delete it.
1784 if (GV->use_empty()) {
1785 GV->eraseFromParent();
1791 } else if (GS.StoredType <= GlobalStatus::InitializerStored) {
1792 DEBUG(dbgs() << "MARKING CONSTANT: " << *GV << "\n");
1793 GV->setConstant(true);
1795 // Clean up any obviously simplifiable users now.
1796 CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, TLI);
1798 // If the global is dead now, just nuke it.
1799 if (GV->use_empty()) {
1800 DEBUG(dbgs() << " *** Marking constant allowed us to simplify "
1801 << "all users and delete global!\n");
1802 GV->eraseFromParent();
1808 } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
1809 if (DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>()) {
1810 const DataLayout &DL = DLP->getDataLayout();
1811 if (GlobalVariable *FirstNewGV = SRAGlobal(GV, DL)) {
1812 GVI = FirstNewGV; // Don't skip the newly produced globals!
1816 } else if (GS.StoredType == GlobalStatus::StoredOnce) {
1817 // If the initial value for the global was an undef value, and if only
1818 // one other value was stored into it, we can just change the
1819 // initializer to be the stored value, then delete all stores to the
1820 // global. This allows us to mark it constant.
1821 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1822 if (isa<UndefValue>(GV->getInitializer())) {
1823 // Change the initial value here.
1824 GV->setInitializer(SOVConstant);
1826 // Clean up any obviously simplifiable users now.
1827 CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, TLI);
1829 if (GV->use_empty()) {
1830 DEBUG(dbgs() << " *** Substituting initializer allowed us to "
1831 << "simplify all users and delete global!\n");
1832 GV->eraseFromParent();
1841 // Try to optimize globals based on the knowledge that only one value
1842 // (besides its initializer) is ever stored to the global.
1843 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GS.Ordering, GVI,
1847 // Otherwise, if the global was not a boolean, we can shrink it to be a
1849 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) {
1850 if (GS.Ordering == NotAtomic) {
1851 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1862 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1863 /// function, changing them to FastCC.
1864 static void ChangeCalleesToFastCall(Function *F) {
1865 for (User *U : F->users()) {
1866 if (isa<BlockAddress>(U))
1868 CallSite CS(cast<Instruction>(U));
1869 CS.setCallingConv(CallingConv::Fast);
1873 static AttributeSet StripNest(LLVMContext &C, const AttributeSet &Attrs) {
1874 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1875 unsigned Index = Attrs.getSlotIndex(i);
1876 if (!Attrs.getSlotAttributes(i).hasAttribute(Index, Attribute::Nest))
1879 // There can be only one.
1880 return Attrs.removeAttribute(C, Index, Attribute::Nest);
1886 static void RemoveNestAttribute(Function *F) {
1887 F->setAttributes(StripNest(F->getContext(), F->getAttributes()));
1888 for (User *U : F->users()) {
1889 if (isa<BlockAddress>(U))
1891 CallSite CS(cast<Instruction>(U));
1892 CS.setAttributes(StripNest(F->getContext(), CS.getAttributes()));
1896 /// Return true if this is a calling convention that we'd like to change. The
1897 /// idea here is that we don't want to mess with the convention if the user
1898 /// explicitly requested something with performance implications like coldcc,
1899 /// GHC, or anyregcc.
1900 static bool isProfitableToMakeFastCC(Function *F) {
1901 CallingConv::ID CC = F->getCallingConv();
1902 // FIXME: Is it worth transforming x86_stdcallcc and x86_fastcallcc?
1903 return CC == CallingConv::C || CC == CallingConv::X86_ThisCall;
1906 bool GlobalOpt::OptimizeFunctions(Module &M) {
1907 bool Changed = false;
1908 // Optimize functions.
1909 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1911 // Functions without names cannot be referenced outside this module.
1912 if (!F->hasName() && !F->isDeclaration())
1913 F->setLinkage(GlobalValue::InternalLinkage);
1914 F->removeDeadConstantUsers();
1915 if (F->isDefTriviallyDead()) {
1916 F->eraseFromParent();
1919 } else if (F->hasLocalLinkage()) {
1920 if (isProfitableToMakeFastCC(F) && !F->isVarArg() &&
1921 !F->hasAddressTaken()) {
1922 // If this function has a calling convention worth changing, is not a
1923 // varargs function, and is only called directly, promote it to use the
1924 // Fast calling convention.
1925 F->setCallingConv(CallingConv::Fast);
1926 ChangeCalleesToFastCall(F);
1931 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
1932 !F->hasAddressTaken()) {
1933 // The function is not used by a trampoline intrinsic, so it is safe
1934 // to remove the 'nest' attribute.
1935 RemoveNestAttribute(F);
1944 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1945 bool Changed = false;
1946 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1948 GlobalVariable *GV = GVI++;
1949 // Global variables without names cannot be referenced outside this module.
1950 if (!GV->hasName() && !GV->isDeclaration())
1951 GV->setLinkage(GlobalValue::InternalLinkage);
1952 // Simplify the initializer.
1953 if (GV->hasInitializer())
1954 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GV->getInitializer())) {
1955 Constant *New = ConstantFoldConstantExpression(CE, DL, TLI);
1956 if (New && New != CE)
1957 GV->setInitializer(New);
1960 Changed |= ProcessGlobal(GV, GVI);
1965 /// FindGlobalCtors - Find the llvm.global_ctors list, verifying that all
1966 /// initializers have an init priority of 65535.
1967 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1968 GlobalVariable *GV = M.getGlobalVariable("llvm.global_ctors");
1969 if (GV == 0) return 0;
1971 // Verify that the initializer is simple enough for us to handle. We are
1972 // only allowed to optimize the initializer if it is unique.
1973 if (!GV->hasUniqueInitializer()) return 0;
1975 if (isa<ConstantAggregateZero>(GV->getInitializer()))
1977 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1979 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
1980 if (isa<ConstantAggregateZero>(*i))
1982 ConstantStruct *CS = cast<ConstantStruct>(*i);
1983 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1986 // Must have a function or null ptr.
1987 if (!isa<Function>(CS->getOperand(1)))
1990 // Init priority must be standard.
1991 ConstantInt *CI = cast<ConstantInt>(CS->getOperand(0));
1992 if (CI->getZExtValue() != 65535)
1999 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
2000 /// return a list of the functions and null terminator as a vector.
2001 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
2002 if (GV->getInitializer()->isNullValue())
2003 return std::vector<Function*>();
2004 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
2005 std::vector<Function*> Result;
2006 Result.reserve(CA->getNumOperands());
2007 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
2008 ConstantStruct *CS = cast<ConstantStruct>(*i);
2009 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
2014 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
2015 /// specified array, returning the new global to use.
2016 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
2017 const std::vector<Function*> &Ctors) {
2018 // If we made a change, reassemble the initializer list.
2019 Constant *CSVals[2];
2020 CSVals[0] = ConstantInt::get(Type::getInt32Ty(GCL->getContext()), 65535);
2023 StructType *StructTy =
2024 cast<StructType>(GCL->getType()->getElementType()->getArrayElementType());
2026 // Create the new init list.
2027 std::vector<Constant*> CAList;
2028 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
2030 CSVals[1] = Ctors[i];
2032 Type *FTy = FunctionType::get(Type::getVoidTy(GCL->getContext()),
2034 PointerType *PFTy = PointerType::getUnqual(FTy);
2035 CSVals[1] = Constant::getNullValue(PFTy);
2036 CSVals[0] = ConstantInt::get(Type::getInt32Ty(GCL->getContext()),
2039 CAList.push_back(ConstantStruct::get(StructTy, CSVals));
2042 // Create the array initializer.
2043 Constant *CA = ConstantArray::get(ArrayType::get(StructTy,
2044 CAList.size()), CAList);
2046 // If we didn't change the number of elements, don't create a new GV.
2047 if (CA->getType() == GCL->getInitializer()->getType()) {
2048 GCL->setInitializer(CA);
2052 // Create the new global and insert it next to the existing list.
2053 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
2054 GCL->getLinkage(), CA, "",
2055 GCL->getThreadLocalMode());
2056 GCL->getParent()->getGlobalList().insert(GCL, NGV);
2059 // Nuke the old list, replacing any uses with the new one.
2060 if (!GCL->use_empty()) {
2062 if (V->getType() != GCL->getType())
2063 V = ConstantExpr::getBitCast(V, GCL->getType());
2064 GCL->replaceAllUsesWith(V);
2066 GCL->eraseFromParent();
2076 isSimpleEnoughValueToCommit(Constant *C,
2077 SmallPtrSet<Constant*, 8> &SimpleConstants,
2078 const DataLayout *DL);
2081 /// isSimpleEnoughValueToCommit - Return true if the specified constant can be
2082 /// handled by the code generator. We don't want to generate something like:
2083 /// void *X = &X/42;
2084 /// because the code generator doesn't have a relocation that can handle that.
2086 /// This function should be called if C was not found (but just got inserted)
2087 /// in SimpleConstants to avoid having to rescan the same constants all the
2089 static bool isSimpleEnoughValueToCommitHelper(Constant *C,
2090 SmallPtrSet<Constant*, 8> &SimpleConstants,
2091 const DataLayout *DL) {
2092 // Simple integer, undef, constant aggregate zero, global addresses, etc are
2094 if (C->getNumOperands() == 0 || isa<BlockAddress>(C) ||
2095 isa<GlobalValue>(C))
2098 // Aggregate values are safe if all their elements are.
2099 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C) ||
2100 isa<ConstantVector>(C)) {
2101 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) {
2102 Constant *Op = cast<Constant>(C->getOperand(i));
2103 if (!isSimpleEnoughValueToCommit(Op, SimpleConstants, DL))
2109 // We don't know exactly what relocations are allowed in constant expressions,
2110 // so we allow &global+constantoffset, which is safe and uniformly supported
2112 ConstantExpr *CE = cast<ConstantExpr>(C);
2113 switch (CE->getOpcode()) {
2114 case Instruction::BitCast:
2115 // Bitcast is fine if the casted value is fine.
2116 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL);
2118 case Instruction::IntToPtr:
2119 case Instruction::PtrToInt:
2120 // int <=> ptr is fine if the int type is the same size as the
2122 if (!DL || DL->getTypeSizeInBits(CE->getType()) !=
2123 DL->getTypeSizeInBits(CE->getOperand(0)->getType()))
2125 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL);
2127 // GEP is fine if it is simple + constant offset.
2128 case Instruction::GetElementPtr:
2129 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
2130 if (!isa<ConstantInt>(CE->getOperand(i)))
2132 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL);
2134 case Instruction::Add:
2135 // We allow simple+cst.
2136 if (!isa<ConstantInt>(CE->getOperand(1)))
2138 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL);
2144 isSimpleEnoughValueToCommit(Constant *C,
2145 SmallPtrSet<Constant*, 8> &SimpleConstants,
2146 const DataLayout *DL) {
2147 // If we already checked this constant, we win.
2148 if (!SimpleConstants.insert(C)) return true;
2149 // Check the constant.
2150 return isSimpleEnoughValueToCommitHelper(C, SimpleConstants, DL);
2154 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
2155 /// enough for us to understand. In particular, if it is a cast to anything
2156 /// other than from one pointer type to another pointer type, we punt.
2157 /// We basically just support direct accesses to globals and GEP's of
2158 /// globals. This should be kept up to date with CommitValueTo.
2159 static bool isSimpleEnoughPointerToCommit(Constant *C) {
2160 // Conservatively, avoid aggregate types. This is because we don't
2161 // want to worry about them partially overlapping other stores.
2162 if (!cast<PointerType>(C->getType())->getElementType()->isSingleValueType())
2165 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
2166 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2167 // external globals.
2168 return GV->hasUniqueInitializer();
2170 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2171 // Handle a constantexpr gep.
2172 if (CE->getOpcode() == Instruction::GetElementPtr &&
2173 isa<GlobalVariable>(CE->getOperand(0)) &&
2174 cast<GEPOperator>(CE)->isInBounds()) {
2175 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2176 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2177 // external globals.
2178 if (!GV->hasUniqueInitializer())
2181 // The first index must be zero.
2182 ConstantInt *CI = dyn_cast<ConstantInt>(*std::next(CE->op_begin()));
2183 if (!CI || !CI->isZero()) return false;
2185 // The remaining indices must be compile-time known integers within the
2186 // notional bounds of the corresponding static array types.
2187 if (!CE->isGEPWithNoNotionalOverIndexing())
2190 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2192 // A constantexpr bitcast from a pointer to another pointer is a no-op,
2193 // and we know how to evaluate it by moving the bitcast from the pointer
2194 // operand to the value operand.
2195 } else if (CE->getOpcode() == Instruction::BitCast &&
2196 isa<GlobalVariable>(CE->getOperand(0))) {
2197 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2198 // external globals.
2199 return cast<GlobalVariable>(CE->getOperand(0))->hasUniqueInitializer();
2206 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
2207 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
2208 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
2209 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
2210 ConstantExpr *Addr, unsigned OpNo) {
2211 // Base case of the recursion.
2212 if (OpNo == Addr->getNumOperands()) {
2213 assert(Val->getType() == Init->getType() && "Type mismatch!");
2217 SmallVector<Constant*, 32> Elts;
2218 if (StructType *STy = dyn_cast<StructType>(Init->getType())) {
2219 // Break up the constant into its elements.
2220 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2221 Elts.push_back(Init->getAggregateElement(i));
2223 // Replace the element that we are supposed to.
2224 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
2225 unsigned Idx = CU->getZExtValue();
2226 assert(Idx < STy->getNumElements() && "Struct index out of range!");
2227 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
2229 // Return the modified struct.
2230 return ConstantStruct::get(STy, Elts);
2233 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
2234 SequentialType *InitTy = cast<SequentialType>(Init->getType());
2237 if (ArrayType *ATy = dyn_cast<ArrayType>(InitTy))
2238 NumElts = ATy->getNumElements();
2240 NumElts = InitTy->getVectorNumElements();
2242 // Break up the array into elements.
2243 for (uint64_t i = 0, e = NumElts; i != e; ++i)
2244 Elts.push_back(Init->getAggregateElement(i));
2246 assert(CI->getZExtValue() < NumElts);
2247 Elts[CI->getZExtValue()] =
2248 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
2250 if (Init->getType()->isArrayTy())
2251 return ConstantArray::get(cast<ArrayType>(InitTy), Elts);
2252 return ConstantVector::get(Elts);
2255 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
2256 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
2257 static void CommitValueTo(Constant *Val, Constant *Addr) {
2258 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
2259 assert(GV->hasInitializer());
2260 GV->setInitializer(Val);
2264 ConstantExpr *CE = cast<ConstantExpr>(Addr);
2265 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2266 GV->setInitializer(EvaluateStoreInto(GV->getInitializer(), Val, CE, 2));
2271 /// Evaluator - This class evaluates LLVM IR, producing the Constant
2272 /// representing each SSA instruction. Changes to global variables are stored
2273 /// in a mapping that can be iterated over after the evaluation is complete.
2274 /// Once an evaluation call fails, the evaluation object should not be reused.
2277 Evaluator(const DataLayout *DL, const TargetLibraryInfo *TLI)
2278 : DL(DL), TLI(TLI) {
2279 ValueStack.push_back(make_unique<DenseMap<Value*, Constant*>>());
2283 for (auto &Tmp : AllocaTmps)
2284 // If there are still users of the alloca, the program is doing something
2285 // silly, e.g. storing the address of the alloca somewhere and using it
2286 // later. Since this is undefined, we'll just make it be null.
2287 if (!Tmp->use_empty())
2288 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2291 /// EvaluateFunction - Evaluate a call to function F, returning true if
2292 /// successful, false if we can't evaluate it. ActualArgs contains the formal
2293 /// arguments for the function.
2294 bool EvaluateFunction(Function *F, Constant *&RetVal,
2295 const SmallVectorImpl<Constant*> &ActualArgs);
2297 /// EvaluateBlock - Evaluate all instructions in block BB, returning true if
2298 /// successful, false if we can't evaluate it. NewBB returns the next BB that
2299 /// control flows into, or null upon return.
2300 bool EvaluateBlock(BasicBlock::iterator CurInst, BasicBlock *&NextBB);
2302 Constant *getVal(Value *V) {
2303 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
2304 Constant *R = ValueStack.back()->lookup(V);
2305 assert(R && "Reference to an uncomputed value!");
2309 void setVal(Value *V, Constant *C) {
2310 (*ValueStack.back())[V] = C;
2313 const DenseMap<Constant*, Constant*> &getMutatedMemory() const {
2314 return MutatedMemory;
2317 const SmallPtrSet<GlobalVariable*, 8> &getInvariants() const {
2322 Constant *ComputeLoadResult(Constant *P);
2324 /// ValueStack - As we compute SSA register values, we store their contents
2325 /// here. The back of the vector contains the current function and the stack
2326 /// contains the values in the calling frames.
2327 SmallVector<std::unique_ptr<DenseMap<Value*, Constant*>>, 4> ValueStack;
2329 /// CallStack - This is used to detect recursion. In pathological situations
2330 /// we could hit exponential behavior, but at least there is nothing
2332 SmallVector<Function*, 4> CallStack;
2334 /// MutatedMemory - For each store we execute, we update this map. Loads
2335 /// check this to get the most up-to-date value. If evaluation is successful,
2336 /// this state is committed to the process.
2337 DenseMap<Constant*, Constant*> MutatedMemory;
2339 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2340 /// to represent its body. This vector is needed so we can delete the
2341 /// temporary globals when we are done.
2342 SmallVector<std::unique_ptr<GlobalVariable>, 32> AllocaTmps;
2344 /// Invariants - These global variables have been marked invariant by the
2345 /// static constructor.
2346 SmallPtrSet<GlobalVariable*, 8> Invariants;
2348 /// SimpleConstants - These are constants we have checked and know to be
2349 /// simple enough to live in a static initializer of a global.
2350 SmallPtrSet<Constant*, 8> SimpleConstants;
2352 const DataLayout *DL;
2353 const TargetLibraryInfo *TLI;
2356 } // anonymous namespace
2358 /// ComputeLoadResult - Return the value that would be computed by a load from
2359 /// P after the stores reflected by 'memory' have been performed. If we can't
2360 /// decide, return null.
2361 Constant *Evaluator::ComputeLoadResult(Constant *P) {
2362 // If this memory location has been recently stored, use the stored value: it
2363 // is the most up-to-date.
2364 DenseMap<Constant*, Constant*>::const_iterator I = MutatedMemory.find(P);
2365 if (I != MutatedMemory.end()) return I->second;
2368 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
2369 if (GV->hasDefinitiveInitializer())
2370 return GV->getInitializer();
2374 // Handle a constantexpr getelementptr.
2375 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
2376 if (CE->getOpcode() == Instruction::GetElementPtr &&
2377 isa<GlobalVariable>(CE->getOperand(0))) {
2378 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2379 if (GV->hasDefinitiveInitializer())
2380 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2383 return 0; // don't know how to evaluate.
2386 /// EvaluateBlock - Evaluate all instructions in block BB, returning true if
2387 /// successful, false if we can't evaluate it. NewBB returns the next BB that
2388 /// control flows into, or null upon return.
2389 bool Evaluator::EvaluateBlock(BasicBlock::iterator CurInst,
2390 BasicBlock *&NextBB) {
2391 // This is the main evaluation loop.
2393 Constant *InstResult = 0;
2395 DEBUG(dbgs() << "Evaluating Instruction: " << *CurInst << "\n");
2397 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
2398 if (!SI->isSimple()) {
2399 DEBUG(dbgs() << "Store is not simple! Can not evaluate.\n");
2400 return false; // no volatile/atomic accesses.
2402 Constant *Ptr = getVal(SI->getOperand(1));
2403 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
2404 DEBUG(dbgs() << "Folding constant ptr expression: " << *Ptr);
2405 Ptr = ConstantFoldConstantExpression(CE, DL, TLI);
2406 DEBUG(dbgs() << "; To: " << *Ptr << "\n");
2408 if (!isSimpleEnoughPointerToCommit(Ptr)) {
2409 // If this is too complex for us to commit, reject it.
2410 DEBUG(dbgs() << "Pointer is too complex for us to evaluate store.");
2414 Constant *Val = getVal(SI->getOperand(0));
2416 // If this might be too difficult for the backend to handle (e.g. the addr
2417 // of one global variable divided by another) then we can't commit it.
2418 if (!isSimpleEnoughValueToCommit(Val, SimpleConstants, DL)) {
2419 DEBUG(dbgs() << "Store value is too complex to evaluate store. " << *Val
2424 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
2425 if (CE->getOpcode() == Instruction::BitCast) {
2426 DEBUG(dbgs() << "Attempting to resolve bitcast on constant ptr.\n");
2427 // If we're evaluating a store through a bitcast, then we need
2428 // to pull the bitcast off the pointer type and push it onto the
2430 Ptr = CE->getOperand(0);
2432 Type *NewTy = cast<PointerType>(Ptr->getType())->getElementType();
2434 // In order to push the bitcast onto the stored value, a bitcast
2435 // from NewTy to Val's type must be legal. If it's not, we can try
2436 // introspecting NewTy to find a legal conversion.
2437 while (!Val->getType()->canLosslesslyBitCastTo(NewTy)) {
2438 // If NewTy is a struct, we can convert the pointer to the struct
2439 // into a pointer to its first member.
2440 // FIXME: This could be extended to support arrays as well.
2441 if (StructType *STy = dyn_cast<StructType>(NewTy)) {
2442 NewTy = STy->getTypeAtIndex(0U);
2444 IntegerType *IdxTy = IntegerType::get(NewTy->getContext(), 32);
2445 Constant *IdxZero = ConstantInt::get(IdxTy, 0, false);
2446 Constant * const IdxList[] = {IdxZero, IdxZero};
2448 Ptr = ConstantExpr::getGetElementPtr(Ptr, IdxList);
2449 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
2450 Ptr = ConstantFoldConstantExpression(CE, DL, TLI);
2452 // If we can't improve the situation by introspecting NewTy,
2453 // we have to give up.
2455 DEBUG(dbgs() << "Failed to bitcast constant ptr, can not "
2461 // If we found compatible types, go ahead and push the bitcast
2462 // onto the stored value.
2463 Val = ConstantExpr::getBitCast(Val, NewTy);
2465 DEBUG(dbgs() << "Evaluated bitcast: " << *Val << "\n");
2469 MutatedMemory[Ptr] = Val;
2470 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2471 InstResult = ConstantExpr::get(BO->getOpcode(),
2472 getVal(BO->getOperand(0)),
2473 getVal(BO->getOperand(1)));
2474 DEBUG(dbgs() << "Found a BinaryOperator! Simplifying: " << *InstResult
2476 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2477 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2478 getVal(CI->getOperand(0)),
2479 getVal(CI->getOperand(1)));
2480 DEBUG(dbgs() << "Found a CmpInst! Simplifying: " << *InstResult
2482 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2483 InstResult = ConstantExpr::getCast(CI->getOpcode(),
2484 getVal(CI->getOperand(0)),
2486 DEBUG(dbgs() << "Found a Cast! Simplifying: " << *InstResult
2488 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2489 InstResult = ConstantExpr::getSelect(getVal(SI->getOperand(0)),
2490 getVal(SI->getOperand(1)),
2491 getVal(SI->getOperand(2)));
2492 DEBUG(dbgs() << "Found a Select! Simplifying: " << *InstResult
2494 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2495 Constant *P = getVal(GEP->getOperand(0));
2496 SmallVector<Constant*, 8> GEPOps;
2497 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2499 GEPOps.push_back(getVal(*i));
2501 ConstantExpr::getGetElementPtr(P, GEPOps,
2502 cast<GEPOperator>(GEP)->isInBounds());
2503 DEBUG(dbgs() << "Found a GEP! Simplifying: " << *InstResult
2505 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2507 if (!LI->isSimple()) {
2508 DEBUG(dbgs() << "Found a Load! Not a simple load, can not evaluate.\n");
2509 return false; // no volatile/atomic accesses.
2512 Constant *Ptr = getVal(LI->getOperand(0));
2513 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
2514 Ptr = ConstantFoldConstantExpression(CE, DL, TLI);
2515 DEBUG(dbgs() << "Found a constant pointer expression, constant "
2516 "folding: " << *Ptr << "\n");
2518 InstResult = ComputeLoadResult(Ptr);
2519 if (InstResult == 0) {
2520 DEBUG(dbgs() << "Failed to compute load result. Can not evaluate load."
2522 return false; // Could not evaluate load.
2525 DEBUG(dbgs() << "Evaluated load: " << *InstResult << "\n");
2526 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2527 if (AI->isArrayAllocation()) {
2528 DEBUG(dbgs() << "Found an array alloca. Can not evaluate.\n");
2529 return false; // Cannot handle array allocs.
2531 Type *Ty = AI->getType()->getElementType();
2532 AllocaTmps.push_back(
2533 make_unique<GlobalVariable>(Ty, false, GlobalValue::InternalLinkage,
2534 UndefValue::get(Ty), AI->getName()));
2535 InstResult = AllocaTmps.back().get();
2536 DEBUG(dbgs() << "Found an alloca. Result: " << *InstResult << "\n");
2537 } else if (isa<CallInst>(CurInst) || isa<InvokeInst>(CurInst)) {
2538 CallSite CS(CurInst);
2540 // Debug info can safely be ignored here.
2541 if (isa<DbgInfoIntrinsic>(CS.getInstruction())) {
2542 DEBUG(dbgs() << "Ignoring debug info.\n");
2547 // Cannot handle inline asm.
2548 if (isa<InlineAsm>(CS.getCalledValue())) {
2549 DEBUG(dbgs() << "Found inline asm, can not evaluate.\n");
2553 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) {
2554 if (MemSetInst *MSI = dyn_cast<MemSetInst>(II)) {
2555 if (MSI->isVolatile()) {
2556 DEBUG(dbgs() << "Can not optimize a volatile memset " <<
2560 Constant *Ptr = getVal(MSI->getDest());
2561 Constant *Val = getVal(MSI->getValue());
2562 Constant *DestVal = ComputeLoadResult(getVal(Ptr));
2563 if (Val->isNullValue() && DestVal && DestVal->isNullValue()) {
2564 // This memset is a no-op.
2565 DEBUG(dbgs() << "Ignoring no-op memset.\n");
2571 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
2572 II->getIntrinsicID() == Intrinsic::lifetime_end) {
2573 DEBUG(dbgs() << "Ignoring lifetime intrinsic.\n");
2578 if (II->getIntrinsicID() == Intrinsic::invariant_start) {
2579 // We don't insert an entry into Values, as it doesn't have a
2580 // meaningful return value.
2581 if (!II->use_empty()) {
2582 DEBUG(dbgs() << "Found unused invariant_start. Can't evaluate.\n");
2585 ConstantInt *Size = cast<ConstantInt>(II->getArgOperand(0));
2586 Value *PtrArg = getVal(II->getArgOperand(1));
2587 Value *Ptr = PtrArg->stripPointerCasts();
2588 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) {
2589 Type *ElemTy = cast<PointerType>(GV->getType())->getElementType();
2590 if (DL && !Size->isAllOnesValue() &&
2591 Size->getValue().getLimitedValue() >=
2592 DL->getTypeStoreSize(ElemTy)) {
2593 Invariants.insert(GV);
2594 DEBUG(dbgs() << "Found a global var that is an invariant: " << *GV
2597 DEBUG(dbgs() << "Found a global var, but can not treat it as an "
2601 // Continue even if we do nothing.
2606 DEBUG(dbgs() << "Unknown intrinsic. Can not evaluate.\n");
2610 // Resolve function pointers.
2611 Function *Callee = dyn_cast<Function>(getVal(CS.getCalledValue()));
2612 if (!Callee || Callee->mayBeOverridden()) {
2613 DEBUG(dbgs() << "Can not resolve function pointer.\n");
2614 return false; // Cannot resolve.
2617 SmallVector<Constant*, 8> Formals;
2618 for (User::op_iterator i = CS.arg_begin(), e = CS.arg_end(); i != e; ++i)
2619 Formals.push_back(getVal(*i));
2621 if (Callee->isDeclaration()) {
2622 // If this is a function we can constant fold, do it.
2623 if (Constant *C = ConstantFoldCall(Callee, Formals, TLI)) {
2625 DEBUG(dbgs() << "Constant folded function call. Result: " <<
2626 *InstResult << "\n");
2628 DEBUG(dbgs() << "Can not constant fold function call.\n");
2632 if (Callee->getFunctionType()->isVarArg()) {
2633 DEBUG(dbgs() << "Can not constant fold vararg function call.\n");
2637 Constant *RetVal = 0;
2638 // Execute the call, if successful, use the return value.
2639 ValueStack.push_back(make_unique<DenseMap<Value *, Constant *>>());
2640 if (!EvaluateFunction(Callee, RetVal, Formals)) {
2641 DEBUG(dbgs() << "Failed to evaluate function.\n");
2644 ValueStack.pop_back();
2645 InstResult = RetVal;
2647 if (InstResult != NULL) {
2648 DEBUG(dbgs() << "Successfully evaluated function. Result: " <<
2649 InstResult << "\n\n");
2651 DEBUG(dbgs() << "Successfully evaluated function. Result: 0\n\n");
2654 } else if (isa<TerminatorInst>(CurInst)) {
2655 DEBUG(dbgs() << "Found a terminator instruction.\n");
2657 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2658 if (BI->isUnconditional()) {
2659 NextBB = BI->getSuccessor(0);
2662 dyn_cast<ConstantInt>(getVal(BI->getCondition()));
2663 if (!Cond) return false; // Cannot determine.
2665 NextBB = BI->getSuccessor(!Cond->getZExtValue());
2667 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2669 dyn_cast<ConstantInt>(getVal(SI->getCondition()));
2670 if (!Val) return false; // Cannot determine.
2671 NextBB = SI->findCaseValue(Val).getCaseSuccessor();
2672 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(CurInst)) {
2673 Value *Val = getVal(IBI->getAddress())->stripPointerCasts();
2674 if (BlockAddress *BA = dyn_cast<BlockAddress>(Val))
2675 NextBB = BA->getBasicBlock();
2677 return false; // Cannot determine.
2678 } else if (isa<ReturnInst>(CurInst)) {
2681 // invoke, unwind, resume, unreachable.
2682 DEBUG(dbgs() << "Can not handle terminator.");
2683 return false; // Cannot handle this terminator.
2686 // We succeeded at evaluating this block!
2687 DEBUG(dbgs() << "Successfully evaluated block.\n");
2690 // Did not know how to evaluate this!
2691 DEBUG(dbgs() << "Failed to evaluate block due to unhandled instruction."
2696 if (!CurInst->use_empty()) {
2697 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(InstResult))
2698 InstResult = ConstantFoldConstantExpression(CE, DL, TLI);
2700 setVal(CurInst, InstResult);
2703 // If we just processed an invoke, we finished evaluating the block.
2704 if (InvokeInst *II = dyn_cast<InvokeInst>(CurInst)) {
2705 NextBB = II->getNormalDest();
2706 DEBUG(dbgs() << "Found an invoke instruction. Finished Block.\n\n");
2710 // Advance program counter.
2715 /// EvaluateFunction - Evaluate a call to function F, returning true if
2716 /// successful, false if we can't evaluate it. ActualArgs contains the formal
2717 /// arguments for the function.
2718 bool Evaluator::EvaluateFunction(Function *F, Constant *&RetVal,
2719 const SmallVectorImpl<Constant*> &ActualArgs) {
2720 // Check to see if this function is already executing (recursion). If so,
2721 // bail out. TODO: we might want to accept limited recursion.
2722 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
2725 CallStack.push_back(F);
2727 // Initialize arguments to the incoming values specified.
2729 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
2731 setVal(AI, ActualArgs[ArgNo]);
2733 // ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
2734 // we can only evaluate any one basic block at most once. This set keeps
2735 // track of what we have executed so we can detect recursive cases etc.
2736 SmallPtrSet<BasicBlock*, 32> ExecutedBlocks;
2738 // CurBB - The current basic block we're evaluating.
2739 BasicBlock *CurBB = F->begin();
2741 BasicBlock::iterator CurInst = CurBB->begin();
2744 BasicBlock *NextBB = 0; // Initialized to avoid compiler warnings.
2745 DEBUG(dbgs() << "Trying to evaluate BB: " << *CurBB << "\n");
2747 if (!EvaluateBlock(CurInst, NextBB))
2751 // Successfully running until there's no next block means that we found
2752 // the return. Fill it the return value and pop the call stack.
2753 ReturnInst *RI = cast<ReturnInst>(CurBB->getTerminator());
2754 if (RI->getNumOperands())
2755 RetVal = getVal(RI->getOperand(0));
2756 CallStack.pop_back();
2760 // Okay, we succeeded in evaluating this control flow. See if we have
2761 // executed the new block before. If so, we have a looping function,
2762 // which we cannot evaluate in reasonable time.
2763 if (!ExecutedBlocks.insert(NextBB))
2764 return false; // looped!
2766 // Okay, we have never been in this block before. Check to see if there
2767 // are any PHI nodes. If so, evaluate them with information about where
2770 for (CurInst = NextBB->begin();
2771 (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2772 setVal(PN, getVal(PN->getIncomingValueForBlock(CurBB)));
2774 // Advance to the next block.
2779 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2780 /// we can. Return true if we can, false otherwise.
2781 static bool EvaluateStaticConstructor(Function *F, const DataLayout *DL,
2782 const TargetLibraryInfo *TLI) {
2783 // Call the function.
2784 Evaluator Eval(DL, TLI);
2785 Constant *RetValDummy;
2786 bool EvalSuccess = Eval.EvaluateFunction(F, RetValDummy,
2787 SmallVector<Constant*, 0>());
2790 // We succeeded at evaluation: commit the result.
2791 DEBUG(dbgs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2792 << F->getName() << "' to " << Eval.getMutatedMemory().size()
2794 for (DenseMap<Constant*, Constant*>::const_iterator I =
2795 Eval.getMutatedMemory().begin(), E = Eval.getMutatedMemory().end();
2797 CommitValueTo(I->second, I->first);
2798 for (SmallPtrSet<GlobalVariable*, 8>::const_iterator I =
2799 Eval.getInvariants().begin(), E = Eval.getInvariants().end();
2801 (*I)->setConstant(true);
2807 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2808 /// Return true if anything changed.
2809 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2810 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2811 bool MadeChange = false;
2812 if (Ctors.empty()) return false;
2814 // Loop over global ctors, optimizing them when we can.
2815 for (unsigned i = 0; i != Ctors.size(); ++i) {
2816 Function *F = Ctors[i];
2817 // Found a null terminator in the middle of the list, prune off the rest of
2820 if (i != Ctors.size()-1) {
2826 DEBUG(dbgs() << "Optimizing Global Constructor: " << *F << "\n");
2828 // We cannot simplify external ctor functions.
2829 if (F->empty()) continue;
2831 // If we can evaluate the ctor at compile time, do.
2832 if (EvaluateStaticConstructor(F, DL, TLI)) {
2833 Ctors.erase(Ctors.begin()+i);
2836 ++NumCtorsEvaluated;
2841 if (!MadeChange) return false;
2843 GCL = InstallGlobalCtors(GCL, Ctors);
2847 static int compareNames(Constant *const *A, Constant *const *B) {
2848 return (*A)->getName().compare((*B)->getName());
2851 static void setUsedInitializer(GlobalVariable &V,
2852 SmallPtrSet<GlobalValue *, 8> Init) {
2854 V.eraseFromParent();
2858 // Type of pointer to the array of pointers.
2859 PointerType *Int8PtrTy = Type::getInt8PtrTy(V.getContext(), 0);
2861 SmallVector<llvm::Constant *, 8> UsedArray;
2862 for (SmallPtrSet<GlobalValue *, 8>::iterator I = Init.begin(), E = Init.end();
2865 = ConstantExpr::getPointerBitCastOrAddrSpaceCast(*I, Int8PtrTy);
2866 UsedArray.push_back(Cast);
2868 // Sort to get deterministic order.
2869 array_pod_sort(UsedArray.begin(), UsedArray.end(), compareNames);
2870 ArrayType *ATy = ArrayType::get(Int8PtrTy, UsedArray.size());
2872 Module *M = V.getParent();
2873 V.removeFromParent();
2874 GlobalVariable *NV =
2875 new GlobalVariable(*M, ATy, false, llvm::GlobalValue::AppendingLinkage,
2876 llvm::ConstantArray::get(ATy, UsedArray), "");
2878 NV->setSection("llvm.metadata");
2883 /// \brief An easy to access representation of llvm.used and llvm.compiler.used.
2885 SmallPtrSet<GlobalValue *, 8> Used;
2886 SmallPtrSet<GlobalValue *, 8> CompilerUsed;
2887 GlobalVariable *UsedV;
2888 GlobalVariable *CompilerUsedV;
2891 LLVMUsed(Module &M) {
2892 UsedV = collectUsedGlobalVariables(M, Used, false);
2893 CompilerUsedV = collectUsedGlobalVariables(M, CompilerUsed, true);
2895 typedef SmallPtrSet<GlobalValue *, 8>::iterator iterator;
2896 iterator usedBegin() { return Used.begin(); }
2897 iterator usedEnd() { return Used.end(); }
2898 iterator compilerUsedBegin() { return CompilerUsed.begin(); }
2899 iterator compilerUsedEnd() { return CompilerUsed.end(); }
2900 bool usedCount(GlobalValue *GV) const { return Used.count(GV); }
2901 bool compilerUsedCount(GlobalValue *GV) const {
2902 return CompilerUsed.count(GV);
2904 bool usedErase(GlobalValue *GV) { return Used.erase(GV); }
2905 bool compilerUsedErase(GlobalValue *GV) { return CompilerUsed.erase(GV); }
2906 bool usedInsert(GlobalValue *GV) { return Used.insert(GV); }
2907 bool compilerUsedInsert(GlobalValue *GV) { return CompilerUsed.insert(GV); }
2909 void syncVariablesAndSets() {
2911 setUsedInitializer(*UsedV, Used);
2913 setUsedInitializer(*CompilerUsedV, CompilerUsed);
2918 static bool hasUseOtherThanLLVMUsed(GlobalAlias &GA, const LLVMUsed &U) {
2919 if (GA.use_empty()) // No use at all.
2922 assert((!U.usedCount(&GA) || !U.compilerUsedCount(&GA)) &&
2923 "We should have removed the duplicated "
2924 "element from llvm.compiler.used");
2925 if (!GA.hasOneUse())
2926 // Strictly more than one use. So at least one is not in llvm.used and
2927 // llvm.compiler.used.
2930 // Exactly one use. Check if it is in llvm.used or llvm.compiler.used.
2931 return !U.usedCount(&GA) && !U.compilerUsedCount(&GA);
2934 static bool hasMoreThanOneUseOtherThanLLVMUsed(GlobalValue &V,
2935 const LLVMUsed &U) {
2937 assert((!U.usedCount(&V) || !U.compilerUsedCount(&V)) &&
2938 "We should have removed the duplicated "
2939 "element from llvm.compiler.used");
2940 if (U.usedCount(&V) || U.compilerUsedCount(&V))
2942 return V.hasNUsesOrMore(N);
2945 static bool mayHaveOtherReferences(GlobalAlias &GA, const LLVMUsed &U) {
2946 if (!GA.hasLocalLinkage())
2949 return U.usedCount(&GA) || U.compilerUsedCount(&GA);
2952 static bool hasUsesToReplace(GlobalAlias &GA, LLVMUsed &U, bool &RenameTarget) {
2953 RenameTarget = false;
2955 if (hasUseOtherThanLLVMUsed(GA, U))
2958 // If the alias is externally visible, we may still be able to simplify it.
2959 if (!mayHaveOtherReferences(GA, U))
2962 // If the aliasee has internal linkage, give it the name and linkage
2963 // of the alias, and delete the alias. This turns:
2964 // define internal ... @f(...)
2965 // @a = alias ... @f
2967 // define ... @a(...)
2968 Constant *Aliasee = GA.getAliasee();
2969 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2970 if (!Target->hasLocalLinkage())
2973 // Do not perform the transform if multiple aliases potentially target the
2974 // aliasee. This check also ensures that it is safe to replace the section
2975 // and other attributes of the aliasee with those of the alias.
2976 if (hasMoreThanOneUseOtherThanLLVMUsed(*Target, U))
2979 RenameTarget = true;
2983 bool GlobalOpt::OptimizeGlobalAliases(Module &M) {
2984 bool Changed = false;
2987 for (SmallPtrSet<GlobalValue *, 8>::iterator I = Used.usedBegin(),
2990 Used.compilerUsedErase(*I);
2992 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end();
2994 Module::alias_iterator J = I++;
2995 // Aliases without names cannot be referenced outside this module.
2996 if (!J->hasName() && !J->isDeclaration())
2997 J->setLinkage(GlobalValue::InternalLinkage);
2998 // If the aliasee may change at link time, nothing can be done - bail out.
2999 if (J->mayBeOverridden())
3002 Constant *Aliasee = J->getAliasee();
3003 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
3004 Target->removeDeadConstantUsers();
3006 // Make all users of the alias use the aliasee instead.
3008 if (!hasUsesToReplace(*J, Used, RenameTarget))
3011 J->replaceAllUsesWith(Aliasee);
3012 ++NumAliasesResolved;
3016 // Give the aliasee the name, linkage and other attributes of the alias.
3017 Target->takeName(J);
3018 Target->setLinkage(J->getLinkage());
3019 Target->setVisibility(J->getVisibility());
3020 Target->setDLLStorageClass(J->getDLLStorageClass());
3022 if (Used.usedErase(J))
3023 Used.usedInsert(Target);
3025 if (Used.compilerUsedErase(J))
3026 Used.compilerUsedInsert(Target);
3027 } else if (mayHaveOtherReferences(*J, Used))
3030 // Delete the alias.
3031 M.getAliasList().erase(J);
3032 ++NumAliasesRemoved;
3036 Used.syncVariablesAndSets();
3041 static Function *FindCXAAtExit(Module &M, TargetLibraryInfo *TLI) {
3042 if (!TLI->has(LibFunc::cxa_atexit))
3045 Function *Fn = M.getFunction(TLI->getName(LibFunc::cxa_atexit));
3050 FunctionType *FTy = Fn->getFunctionType();
3052 // Checking that the function has the right return type, the right number of
3053 // parameters and that they all have pointer types should be enough.
3054 if (!FTy->getReturnType()->isIntegerTy() ||
3055 FTy->getNumParams() != 3 ||
3056 !FTy->getParamType(0)->isPointerTy() ||
3057 !FTy->getParamType(1)->isPointerTy() ||
3058 !FTy->getParamType(2)->isPointerTy())
3064 /// cxxDtorIsEmpty - Returns whether the given function is an empty C++
3065 /// destructor and can therefore be eliminated.
3066 /// Note that we assume that other optimization passes have already simplified
3067 /// the code so we only look for a function with a single basic block, where
3068 /// the only allowed instructions are 'ret', 'call' to an empty C++ dtor and
3069 /// other side-effect free instructions.
3070 static bool cxxDtorIsEmpty(const Function &Fn,
3071 SmallPtrSet<const Function *, 8> &CalledFunctions) {
3072 // FIXME: We could eliminate C++ destructors if they're readonly/readnone and
3073 // nounwind, but that doesn't seem worth doing.
3074 if (Fn.isDeclaration())
3077 if (++Fn.begin() != Fn.end())
3080 const BasicBlock &EntryBlock = Fn.getEntryBlock();
3081 for (BasicBlock::const_iterator I = EntryBlock.begin(), E = EntryBlock.end();
3083 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
3084 // Ignore debug intrinsics.
3085 if (isa<DbgInfoIntrinsic>(CI))
3088 const Function *CalledFn = CI->getCalledFunction();
3093 SmallPtrSet<const Function *, 8> NewCalledFunctions(CalledFunctions);
3095 // Don't treat recursive functions as empty.
3096 if (!NewCalledFunctions.insert(CalledFn))
3099 if (!cxxDtorIsEmpty(*CalledFn, NewCalledFunctions))
3101 } else if (isa<ReturnInst>(*I))
3102 return true; // We're done.
3103 else if (I->mayHaveSideEffects())
3104 return false; // Destructor with side effects, bail.
3110 bool GlobalOpt::OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) {
3111 /// Itanium C++ ABI p3.3.5:
3113 /// After constructing a global (or local static) object, that will require
3114 /// destruction on exit, a termination function is registered as follows:
3116 /// extern "C" int __cxa_atexit ( void (*f)(void *), void *p, void *d );
3118 /// This registration, e.g. __cxa_atexit(f,p,d), is intended to cause the
3119 /// call f(p) when DSO d is unloaded, before all such termination calls
3120 /// registered before this one. It returns zero if registration is
3121 /// successful, nonzero on failure.
3123 // This pass will look for calls to __cxa_atexit where the function is trivial
3125 bool Changed = false;
3127 for (auto I = CXAAtExitFn->user_begin(), E = CXAAtExitFn->user_end();
3129 // We're only interested in calls. Theoretically, we could handle invoke
3130 // instructions as well, but neither llvm-gcc nor clang generate invokes
3132 CallInst *CI = dyn_cast<CallInst>(*I++);
3137 dyn_cast<Function>(CI->getArgOperand(0)->stripPointerCasts());
3141 SmallPtrSet<const Function *, 8> CalledFunctions;
3142 if (!cxxDtorIsEmpty(*DtorFn, CalledFunctions))
3145 // Just remove the call.
3146 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
3147 CI->eraseFromParent();
3149 ++NumCXXDtorsRemoved;
3157 bool GlobalOpt::runOnModule(Module &M) {
3158 bool Changed = false;
3160 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
3161 DL = DLP ? &DLP->getDataLayout() : 0;
3162 TLI = &getAnalysis<TargetLibraryInfo>();
3164 // Try to find the llvm.globalctors list.
3165 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
3167 bool LocalChange = true;
3168 while (LocalChange) {
3169 LocalChange = false;
3171 // Delete functions that are trivially dead, ccc -> fastcc
3172 LocalChange |= OptimizeFunctions(M);
3174 // Optimize global_ctors list.
3176 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
3178 // Optimize non-address-taken globals.
3179 LocalChange |= OptimizeGlobalVars(M);
3181 // Resolve aliases, when possible.
3182 LocalChange |= OptimizeGlobalAliases(M);
3184 // Try to remove trivial global destructors if they are not removed
3186 Function *CXAAtExitFn = FindCXAAtExit(M, TLI);
3188 LocalChange |= OptimizeEmptyGlobalCXXDtors(CXAAtExitFn);
3190 Changed |= LocalChange;
3193 // TODO: Move all global ctors functions to the end of the module for code