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/CtorUtils.h"
42 #include "llvm/Transforms/Utils/GlobalStatus.h"
43 #include "llvm/Transforms/Utils/ModuleUtils.h"
47 #define DEBUG_TYPE "globalopt"
49 STATISTIC(NumMarked , "Number of globals marked constant");
50 STATISTIC(NumUnnamed , "Number of globals marked unnamed_addr");
51 STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
52 STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
53 STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
54 STATISTIC(NumDeleted , "Number of globals deleted");
55 STATISTIC(NumFnDeleted , "Number of functions deleted");
56 STATISTIC(NumGlobUses , "Number of global uses devirtualized");
57 STATISTIC(NumLocalized , "Number of globals localized");
58 STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans");
59 STATISTIC(NumFastCallFns , "Number of functions converted to fastcc");
60 STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
61 STATISTIC(NumNestRemoved , "Number of nest attributes removed");
62 STATISTIC(NumAliasesResolved, "Number of global aliases resolved");
63 STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated");
64 STATISTIC(NumCXXDtorsRemoved, "Number of global C++ destructors removed");
67 struct GlobalOpt : public ModulePass {
68 void getAnalysisUsage(AnalysisUsage &AU) const override {
69 AU.addRequired<TargetLibraryInfo>();
71 static char ID; // Pass identification, replacement for typeid
72 GlobalOpt() : ModulePass(ID) {
73 initializeGlobalOptPass(*PassRegistry::getPassRegistry());
76 bool runOnModule(Module &M) override;
79 bool OptimizeFunctions(Module &M);
80 bool OptimizeGlobalVars(Module &M);
81 bool OptimizeGlobalAliases(Module &M);
82 bool ProcessGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
83 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI,
84 const GlobalStatus &GS);
85 bool OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn);
88 TargetLibraryInfo *TLI;
92 char GlobalOpt::ID = 0;
93 INITIALIZE_PASS_BEGIN(GlobalOpt, "globalopt",
94 "Global Variable Optimizer", false, false)
95 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
96 INITIALIZE_PASS_END(GlobalOpt, "globalopt",
97 "Global Variable Optimizer", false, false)
99 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
101 /// isLeakCheckerRoot - Is this global variable possibly used by a leak checker
102 /// as a root? If so, we might not really want to eliminate the stores to it.
103 static bool isLeakCheckerRoot(GlobalVariable *GV) {
104 // A global variable is a root if it is a pointer, or could plausibly contain
105 // a pointer. There are two challenges; one is that we could have a struct
106 // the has an inner member which is a pointer. We recurse through the type to
107 // detect these (up to a point). The other is that we may actually be a union
108 // of a pointer and another type, and so our LLVM type is an integer which
109 // gets converted into a pointer, or our type is an [i8 x #] with a pointer
110 // potentially contained here.
112 if (GV->hasPrivateLinkage())
115 SmallVector<Type *, 4> Types;
116 Types.push_back(cast<PointerType>(GV->getType())->getElementType());
120 Type *Ty = Types.pop_back_val();
121 switch (Ty->getTypeID()) {
123 case Type::PointerTyID: return true;
124 case Type::ArrayTyID:
125 case Type::VectorTyID: {
126 SequentialType *STy = cast<SequentialType>(Ty);
127 Types.push_back(STy->getElementType());
130 case Type::StructTyID: {
131 StructType *STy = cast<StructType>(Ty);
132 if (STy->isOpaque()) return true;
133 for (StructType::element_iterator I = STy->element_begin(),
134 E = STy->element_end(); I != E; ++I) {
136 if (isa<PointerType>(InnerTy)) return true;
137 if (isa<CompositeType>(InnerTy))
138 Types.push_back(InnerTy);
143 if (--Limit == 0) return true;
144 } while (!Types.empty());
148 /// Given a value that is stored to a global but never read, determine whether
149 /// it's safe to remove the store and the chain of computation that feeds the
151 static bool IsSafeComputationToRemove(Value *V, const TargetLibraryInfo *TLI) {
153 if (isa<Constant>(V))
157 if (isa<LoadInst>(V) || isa<InvokeInst>(V) || isa<Argument>(V) ||
160 if (isAllocationFn(V, TLI))
163 Instruction *I = cast<Instruction>(V);
164 if (I->mayHaveSideEffects())
166 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
167 if (!GEP->hasAllConstantIndices())
169 } else if (I->getNumOperands() != 1) {
173 V = I->getOperand(0);
177 /// CleanupPointerRootUsers - This GV is a pointer root. Loop over all users
178 /// of the global and clean up any that obviously don't assign the global a
179 /// value that isn't dynamically allocated.
181 static bool CleanupPointerRootUsers(GlobalVariable *GV,
182 const TargetLibraryInfo *TLI) {
183 // A brief explanation of leak checkers. The goal is to find bugs where
184 // pointers are forgotten, causing an accumulating growth in memory
185 // usage over time. The common strategy for leak checkers is to whitelist the
186 // memory pointed to by globals at exit. This is popular because it also
187 // solves another problem where the main thread of a C++ program may shut down
188 // before other threads that are still expecting to use those globals. To
189 // handle that case, we expect the program may create a singleton and never
192 bool Changed = false;
194 // If Dead[n].first is the only use of a malloc result, we can delete its
195 // chain of computation and the store to the global in Dead[n].second.
196 SmallVector<std::pair<Instruction *, Instruction *>, 32> Dead;
198 // Constants can't be pointers to dynamically allocated memory.
199 for (Value::user_iterator UI = GV->user_begin(), E = GV->user_end();
202 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
203 Value *V = SI->getValueOperand();
204 if (isa<Constant>(V)) {
206 SI->eraseFromParent();
207 } else if (Instruction *I = dyn_cast<Instruction>(V)) {
209 Dead.push_back(std::make_pair(I, SI));
211 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(U)) {
212 if (isa<Constant>(MSI->getValue())) {
214 MSI->eraseFromParent();
215 } else if (Instruction *I = dyn_cast<Instruction>(MSI->getValue())) {
217 Dead.push_back(std::make_pair(I, MSI));
219 } else if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(U)) {
220 GlobalVariable *MemSrc = dyn_cast<GlobalVariable>(MTI->getSource());
221 if (MemSrc && MemSrc->isConstant()) {
223 MTI->eraseFromParent();
224 } else if (Instruction *I = dyn_cast<Instruction>(MemSrc)) {
226 Dead.push_back(std::make_pair(I, MTI));
228 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
229 if (CE->use_empty()) {
230 CE->destroyConstant();
233 } else if (Constant *C = dyn_cast<Constant>(U)) {
234 if (isSafeToDestroyConstant(C)) {
235 C->destroyConstant();
236 // This could have invalidated UI, start over from scratch.
238 CleanupPointerRootUsers(GV, TLI);
244 for (int i = 0, e = Dead.size(); i != e; ++i) {
245 if (IsSafeComputationToRemove(Dead[i].first, TLI)) {
246 Dead[i].second->eraseFromParent();
247 Instruction *I = Dead[i].first;
249 if (isAllocationFn(I, TLI))
251 Instruction *J = dyn_cast<Instruction>(I->getOperand(0));
254 I->eraseFromParent();
257 I->eraseFromParent();
264 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
265 /// users of the global, cleaning up the obvious ones. This is largely just a
266 /// quick scan over the use list to clean up the easy and obvious cruft. This
267 /// returns true if it made a change.
268 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init,
269 const DataLayout *DL,
270 TargetLibraryInfo *TLI) {
271 bool Changed = false;
272 // Note that we need to use a weak value handle for the worklist items. When
273 // we delete a constant array, we may also be holding pointer to one of its
274 // elements (or an element of one of its elements if we're dealing with an
275 // array of arrays) in the worklist.
276 SmallVector<WeakVH, 8> WorkList(V->user_begin(), V->user_end());
277 while (!WorkList.empty()) {
278 Value *UV = WorkList.pop_back_val();
282 User *U = cast<User>(UV);
284 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
286 // Replace the load with the initializer.
287 LI->replaceAllUsesWith(Init);
288 LI->eraseFromParent();
291 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
292 // Store must be unreachable or storing Init into the global.
293 SI->eraseFromParent();
295 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
296 if (CE->getOpcode() == Instruction::GetElementPtr) {
297 Constant *SubInit = nullptr;
299 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
300 Changed |= CleanupConstantGlobalUsers(CE, SubInit, DL, TLI);
301 } else if ((CE->getOpcode() == Instruction::BitCast &&
302 CE->getType()->isPointerTy()) ||
303 CE->getOpcode() == Instruction::AddrSpaceCast) {
304 // Pointer cast, delete any stores and memsets to the global.
305 Changed |= CleanupConstantGlobalUsers(CE, nullptr, DL, TLI);
308 if (CE->use_empty()) {
309 CE->destroyConstant();
312 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
313 // Do not transform "gepinst (gep constexpr (GV))" here, because forming
314 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
315 // and will invalidate our notion of what Init is.
316 Constant *SubInit = nullptr;
317 if (!isa<ConstantExpr>(GEP->getOperand(0))) {
319 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP, DL, TLI));
320 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
321 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
323 // If the initializer is an all-null value and we have an inbounds GEP,
324 // we already know what the result of any load from that GEP is.
325 // TODO: Handle splats.
326 if (Init && isa<ConstantAggregateZero>(Init) && GEP->isInBounds())
327 SubInit = Constant::getNullValue(GEP->getType()->getElementType());
329 Changed |= CleanupConstantGlobalUsers(GEP, SubInit, DL, TLI);
331 if (GEP->use_empty()) {
332 GEP->eraseFromParent();
335 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
336 if (MI->getRawDest() == V) {
337 MI->eraseFromParent();
341 } else if (Constant *C = dyn_cast<Constant>(U)) {
342 // If we have a chain of dead constantexprs or other things dangling from
343 // us, and if they are all dead, nuke them without remorse.
344 if (isSafeToDestroyConstant(C)) {
345 C->destroyConstant();
346 CleanupConstantGlobalUsers(V, Init, DL, TLI);
354 /// isSafeSROAElementUse - Return true if the specified instruction is a safe
355 /// user of a derived expression from a global that we want to SROA.
356 static bool isSafeSROAElementUse(Value *V) {
357 // We might have a dead and dangling constant hanging off of here.
358 if (Constant *C = dyn_cast<Constant>(V))
359 return isSafeToDestroyConstant(C);
361 Instruction *I = dyn_cast<Instruction>(V);
362 if (!I) return false;
365 if (isa<LoadInst>(I)) return true;
367 // Stores *to* the pointer are ok.
368 if (StoreInst *SI = dyn_cast<StoreInst>(I))
369 return SI->getOperand(0) != V;
371 // Otherwise, it must be a GEP.
372 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I);
373 if (!GEPI) return false;
375 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
376 !cast<Constant>(GEPI->getOperand(1))->isNullValue())
379 for (User *U : GEPI->users())
380 if (!isSafeSROAElementUse(U))
386 /// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value.
387 /// Look at it and its uses and decide whether it is safe to SROA this global.
389 static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
390 // The user of the global must be a GEP Inst or a ConstantExpr GEP.
391 if (!isa<GetElementPtrInst>(U) &&
392 (!isa<ConstantExpr>(U) ||
393 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
396 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
397 // don't like < 3 operand CE's, and we don't like non-constant integer
398 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some
400 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
401 !cast<Constant>(U->getOperand(1))->isNullValue() ||
402 !isa<ConstantInt>(U->getOperand(2)))
405 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
406 ++GEPI; // Skip over the pointer index.
408 // If this is a use of an array allocation, do a bit more checking for sanity.
409 if (ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
410 uint64_t NumElements = AT->getNumElements();
411 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
413 // Check to make sure that index falls within the array. If not,
414 // something funny is going on, so we won't do the optimization.
416 if (Idx->getZExtValue() >= NumElements)
419 // We cannot scalar repl this level of the array unless any array
420 // sub-indices are in-range constants. In particular, consider:
421 // A[0][i]. We cannot know that the user isn't doing invalid things like
422 // allowing i to index an out-of-range subscript that accesses A[1].
424 // Scalar replacing *just* the outer index of the array is probably not
425 // going to be a win anyway, so just give up.
426 for (++GEPI; // Skip array index.
429 uint64_t NumElements;
430 if (ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
431 NumElements = SubArrayTy->getNumElements();
432 else if (VectorType *SubVectorTy = dyn_cast<VectorType>(*GEPI))
433 NumElements = SubVectorTy->getNumElements();
435 assert((*GEPI)->isStructTy() &&
436 "Indexed GEP type is not array, vector, or struct!");
440 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
441 if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
446 for (User *UU : U->users())
447 if (!isSafeSROAElementUse(UU))
453 /// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
454 /// is safe for us to perform this transformation.
456 static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
457 for (User *U : GV->users())
458 if (!IsUserOfGlobalSafeForSRA(U, GV))
465 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
466 /// variable. This opens the door for other optimizations by exposing the
467 /// behavior of the program in a more fine-grained way. We have determined that
468 /// this transformation is safe already. We return the first global variable we
469 /// insert so that the caller can reprocess it.
470 static GlobalVariable *SRAGlobal(GlobalVariable *GV, const DataLayout &DL) {
471 // Make sure this global only has simple uses that we can SRA.
472 if (!GlobalUsersSafeToSRA(GV))
475 assert(GV->hasLocalLinkage() && !GV->isConstant());
476 Constant *Init = GV->getInitializer();
477 Type *Ty = Init->getType();
479 std::vector<GlobalVariable*> NewGlobals;
480 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
482 // Get the alignment of the global, either explicit or target-specific.
483 unsigned StartAlignment = GV->getAlignment();
484 if (StartAlignment == 0)
485 StartAlignment = DL.getABITypeAlignment(GV->getType());
487 if (StructType *STy = dyn_cast<StructType>(Ty)) {
488 NewGlobals.reserve(STy->getNumElements());
489 const StructLayout &Layout = *DL.getStructLayout(STy);
490 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
491 Constant *In = Init->getAggregateElement(i);
492 assert(In && "Couldn't get element of initializer?");
493 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
494 GlobalVariable::InternalLinkage,
495 In, GV->getName()+"."+Twine(i),
496 GV->getThreadLocalMode(),
497 GV->getType()->getAddressSpace());
498 Globals.insert(GV, NGV);
499 NewGlobals.push_back(NGV);
501 // Calculate the known alignment of the field. If the original aggregate
502 // had 256 byte alignment for example, something might depend on that:
503 // propagate info to each field.
504 uint64_t FieldOffset = Layout.getElementOffset(i);
505 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset);
506 if (NewAlign > DL.getABITypeAlignment(STy->getElementType(i)))
507 NGV->setAlignment(NewAlign);
509 } else if (SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
510 unsigned NumElements = 0;
511 if (ArrayType *ATy = dyn_cast<ArrayType>(STy))
512 NumElements = ATy->getNumElements();
514 NumElements = cast<VectorType>(STy)->getNumElements();
516 if (NumElements > 16 && GV->hasNUsesOrMore(16))
517 return nullptr; // It's not worth it.
518 NewGlobals.reserve(NumElements);
520 uint64_t EltSize = DL.getTypeAllocSize(STy->getElementType());
521 unsigned EltAlign = DL.getABITypeAlignment(STy->getElementType());
522 for (unsigned i = 0, e = NumElements; i != e; ++i) {
523 Constant *In = Init->getAggregateElement(i);
524 assert(In && "Couldn't get element of initializer?");
526 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
527 GlobalVariable::InternalLinkage,
528 In, GV->getName()+"."+Twine(i),
529 GV->getThreadLocalMode(),
530 GV->getType()->getAddressSpace());
531 Globals.insert(GV, NGV);
532 NewGlobals.push_back(NGV);
534 // Calculate the known alignment of the field. If the original aggregate
535 // had 256 byte alignment for example, something might depend on that:
536 // propagate info to each field.
537 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
538 if (NewAlign > EltAlign)
539 NGV->setAlignment(NewAlign);
543 if (NewGlobals.empty())
546 DEBUG(dbgs() << "PERFORMING GLOBAL SRA ON: " << *GV);
548 Constant *NullInt =Constant::getNullValue(Type::getInt32Ty(GV->getContext()));
550 // Loop over all of the uses of the global, replacing the constantexpr geps,
551 // with smaller constantexpr geps or direct references.
552 while (!GV->use_empty()) {
553 User *GEP = GV->user_back();
554 assert(((isa<ConstantExpr>(GEP) &&
555 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
556 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
558 // Ignore the 1th operand, which has to be zero or else the program is quite
559 // broken (undefined). Get the 2nd operand, which is the structure or array
561 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
562 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
564 Value *NewPtr = NewGlobals[Val];
566 // Form a shorter GEP if needed.
567 if (GEP->getNumOperands() > 3) {
568 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
569 SmallVector<Constant*, 8> Idxs;
570 Idxs.push_back(NullInt);
571 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
572 Idxs.push_back(CE->getOperand(i));
573 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr), Idxs);
575 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
576 SmallVector<Value*, 8> Idxs;
577 Idxs.push_back(NullInt);
578 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
579 Idxs.push_back(GEPI->getOperand(i));
580 NewPtr = GetElementPtrInst::Create(NewPtr, Idxs,
581 GEPI->getName()+"."+Twine(Val),GEPI);
584 GEP->replaceAllUsesWith(NewPtr);
586 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
587 GEPI->eraseFromParent();
589 cast<ConstantExpr>(GEP)->destroyConstant();
592 // Delete the old global, now that it is dead.
596 // Loop over the new globals array deleting any globals that are obviously
597 // dead. This can arise due to scalarization of a structure or an array that
598 // has elements that are dead.
599 unsigned FirstGlobal = 0;
600 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
601 if (NewGlobals[i]->use_empty()) {
602 Globals.erase(NewGlobals[i]);
603 if (FirstGlobal == i) ++FirstGlobal;
606 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : nullptr;
609 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
610 /// value will trap if the value is dynamically null. PHIs keeps track of any
611 /// phi nodes we've seen to avoid reprocessing them.
612 static bool AllUsesOfValueWillTrapIfNull(const Value *V,
613 SmallPtrSet<const PHINode*, 8> &PHIs) {
614 for (const User *U : V->users())
615 if (isa<LoadInst>(U)) {
617 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
618 if (SI->getOperand(0) == V) {
619 //cerr << "NONTRAPPING USE: " << *U;
620 return false; // Storing the value.
622 } else if (const CallInst *CI = dyn_cast<CallInst>(U)) {
623 if (CI->getCalledValue() != V) {
624 //cerr << "NONTRAPPING USE: " << *U;
625 return false; // Not calling the ptr
627 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(U)) {
628 if (II->getCalledValue() != V) {
629 //cerr << "NONTRAPPING USE: " << *U;
630 return false; // Not calling the ptr
632 } else if (const BitCastInst *CI = dyn_cast<BitCastInst>(U)) {
633 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
634 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
635 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
636 } else if (const PHINode *PN = dyn_cast<PHINode>(U)) {
637 // If we've already seen this phi node, ignore it, it has already been
639 if (PHIs.insert(PN) && !AllUsesOfValueWillTrapIfNull(PN, PHIs))
641 } else if (isa<ICmpInst>(U) &&
642 isa<ConstantPointerNull>(U->getOperand(1))) {
643 // Ignore icmp X, null
645 //cerr << "NONTRAPPING USE: " << *U;
652 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
653 /// from GV will trap if the loaded value is null. Note that this also permits
654 /// comparisons of the loaded value against null, as a special case.
655 static bool AllUsesOfLoadedValueWillTrapIfNull(const GlobalVariable *GV) {
656 for (const User *U : GV->users())
657 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
658 SmallPtrSet<const PHINode*, 8> PHIs;
659 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
661 } else if (isa<StoreInst>(U)) {
662 // Ignore stores to the global.
664 // We don't know or understand this user, bail out.
665 //cerr << "UNKNOWN USER OF GLOBAL!: " << *U;
671 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
672 bool Changed = false;
673 for (auto UI = V->user_begin(), E = V->user_end(); UI != E; ) {
674 Instruction *I = cast<Instruction>(*UI++);
675 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
676 LI->setOperand(0, NewV);
678 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
679 if (SI->getOperand(1) == V) {
680 SI->setOperand(1, NewV);
683 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
685 if (CS.getCalledValue() == V) {
686 // Calling through the pointer! Turn into a direct call, but be careful
687 // that the pointer is not also being passed as an argument.
688 CS.setCalledFunction(NewV);
690 bool PassedAsArg = false;
691 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
692 if (CS.getArgument(i) == V) {
694 CS.setArgument(i, NewV);
698 // Being passed as an argument also. Be careful to not invalidate UI!
699 UI = V->user_begin();
702 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
703 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
704 ConstantExpr::getCast(CI->getOpcode(),
705 NewV, CI->getType()));
706 if (CI->use_empty()) {
708 CI->eraseFromParent();
710 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
711 // Should handle GEP here.
712 SmallVector<Constant*, 8> Idxs;
713 Idxs.reserve(GEPI->getNumOperands()-1);
714 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
716 if (Constant *C = dyn_cast<Constant>(*i))
720 if (Idxs.size() == GEPI->getNumOperands()-1)
721 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
722 ConstantExpr::getGetElementPtr(NewV, Idxs));
723 if (GEPI->use_empty()) {
725 GEPI->eraseFromParent();
734 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
735 /// value stored into it. If there are uses of the loaded value that would trap
736 /// if the loaded value is dynamically null, then we know that they cannot be
737 /// reachable with a null optimize away the load.
738 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV,
739 const DataLayout *DL,
740 TargetLibraryInfo *TLI) {
741 bool Changed = false;
743 // Keep track of whether we are able to remove all the uses of the global
744 // other than the store that defines it.
745 bool AllNonStoreUsesGone = true;
747 // Replace all uses of loads with uses of uses of the stored value.
748 for (Value::user_iterator GUI = GV->user_begin(), E = GV->user_end(); GUI != E;){
749 User *GlobalUser = *GUI++;
750 if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) {
751 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
752 // If we were able to delete all uses of the loads
753 if (LI->use_empty()) {
754 LI->eraseFromParent();
757 AllNonStoreUsesGone = false;
759 } else if (isa<StoreInst>(GlobalUser)) {
760 // Ignore the store that stores "LV" to the global.
761 assert(GlobalUser->getOperand(1) == GV &&
762 "Must be storing *to* the global");
764 AllNonStoreUsesGone = false;
766 // If we get here we could have other crazy uses that are transitively
768 assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) ||
769 isa<ConstantExpr>(GlobalUser) || isa<CmpInst>(GlobalUser) ||
770 isa<BitCastInst>(GlobalUser) ||
771 isa<GetElementPtrInst>(GlobalUser)) &&
772 "Only expect load and stores!");
777 DEBUG(dbgs() << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV);
781 // If we nuked all of the loads, then none of the stores are needed either,
782 // nor is the global.
783 if (AllNonStoreUsesGone) {
784 if (isLeakCheckerRoot(GV)) {
785 Changed |= CleanupPointerRootUsers(GV, TLI);
788 CleanupConstantGlobalUsers(GV, nullptr, DL, TLI);
790 if (GV->use_empty()) {
791 DEBUG(dbgs() << " *** GLOBAL NOW DEAD!\n");
793 GV->eraseFromParent();
800 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
801 /// instructions that are foldable.
802 static void ConstantPropUsersOf(Value *V, const DataLayout *DL,
803 TargetLibraryInfo *TLI) {
804 for (Value::user_iterator UI = V->user_begin(), E = V->user_end(); UI != E; )
805 if (Instruction *I = dyn_cast<Instruction>(*UI++))
806 if (Constant *NewC = ConstantFoldInstruction(I, DL, TLI)) {
807 I->replaceAllUsesWith(NewC);
809 // Advance UI to the next non-I use to avoid invalidating it!
810 // Instructions could multiply use V.
811 while (UI != E && *UI == I)
813 I->eraseFromParent();
817 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
818 /// variable, and transforms the program as if it always contained the result of
819 /// the specified malloc. Because it is always the result of the specified
820 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
821 /// malloc into a global, and any loads of GV as uses of the new global.
822 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
825 ConstantInt *NElements,
826 const DataLayout *DL,
827 TargetLibraryInfo *TLI) {
828 DEBUG(errs() << "PROMOTING GLOBAL: " << *GV << " CALL = " << *CI << '\n');
831 if (NElements->getZExtValue() == 1)
832 GlobalType = AllocTy;
834 // If we have an array allocation, the global variable is of an array.
835 GlobalType = ArrayType::get(AllocTy, NElements->getZExtValue());
837 // Create the new global variable. The contents of the malloc'd memory is
838 // undefined, so initialize with an undef value.
839 GlobalVariable *NewGV = new GlobalVariable(*GV->getParent(),
841 GlobalValue::InternalLinkage,
842 UndefValue::get(GlobalType),
843 GV->getName()+".body",
845 GV->getThreadLocalMode());
847 // If there are bitcast users of the malloc (which is typical, usually we have
848 // a malloc + bitcast) then replace them with uses of the new global. Update
849 // other users to use the global as well.
850 BitCastInst *TheBC = nullptr;
851 while (!CI->use_empty()) {
852 Instruction *User = cast<Instruction>(CI->user_back());
853 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
854 if (BCI->getType() == NewGV->getType()) {
855 BCI->replaceAllUsesWith(NewGV);
856 BCI->eraseFromParent();
858 BCI->setOperand(0, NewGV);
862 TheBC = new BitCastInst(NewGV, CI->getType(), "newgv", CI);
863 User->replaceUsesOfWith(CI, TheBC);
867 Constant *RepValue = NewGV;
868 if (NewGV->getType() != GV->getType()->getElementType())
869 RepValue = ConstantExpr::getBitCast(RepValue,
870 GV->getType()->getElementType());
872 // If there is a comparison against null, we will insert a global bool to
873 // keep track of whether the global was initialized yet or not.
874 GlobalVariable *InitBool =
875 new GlobalVariable(Type::getInt1Ty(GV->getContext()), false,
876 GlobalValue::InternalLinkage,
877 ConstantInt::getFalse(GV->getContext()),
878 GV->getName()+".init", GV->getThreadLocalMode());
879 bool InitBoolUsed = false;
881 // Loop over all uses of GV, processing them in turn.
882 while (!GV->use_empty()) {
883 if (StoreInst *SI = dyn_cast<StoreInst>(GV->user_back())) {
884 // The global is initialized when the store to it occurs.
885 new StoreInst(ConstantInt::getTrue(GV->getContext()), InitBool, false, 0,
886 SI->getOrdering(), SI->getSynchScope(), SI);
887 SI->eraseFromParent();
891 LoadInst *LI = cast<LoadInst>(GV->user_back());
892 while (!LI->use_empty()) {
893 Use &LoadUse = *LI->use_begin();
894 ICmpInst *ICI = dyn_cast<ICmpInst>(LoadUse.getUser());
900 // Replace the cmp X, 0 with a use of the bool value.
901 // Sink the load to where the compare was, if atomic rules allow us to.
902 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", false, 0,
903 LI->getOrdering(), LI->getSynchScope(),
904 LI->isUnordered() ? (Instruction*)ICI : LI);
906 switch (ICI->getPredicate()) {
907 default: llvm_unreachable("Unknown ICmp Predicate!");
908 case ICmpInst::ICMP_ULT:
909 case ICmpInst::ICMP_SLT: // X < null -> always false
910 LV = ConstantInt::getFalse(GV->getContext());
912 case ICmpInst::ICMP_ULE:
913 case ICmpInst::ICMP_SLE:
914 case ICmpInst::ICMP_EQ:
915 LV = BinaryOperator::CreateNot(LV, "notinit", ICI);
917 case ICmpInst::ICMP_NE:
918 case ICmpInst::ICMP_UGE:
919 case ICmpInst::ICMP_SGE:
920 case ICmpInst::ICMP_UGT:
921 case ICmpInst::ICMP_SGT:
924 ICI->replaceAllUsesWith(LV);
925 ICI->eraseFromParent();
927 LI->eraseFromParent();
930 // If the initialization boolean was used, insert it, otherwise delete it.
932 while (!InitBool->use_empty()) // Delete initializations
933 cast<StoreInst>(InitBool->user_back())->eraseFromParent();
936 GV->getParent()->getGlobalList().insert(GV, InitBool);
938 // Now the GV is dead, nuke it and the malloc..
939 GV->eraseFromParent();
940 CI->eraseFromParent();
942 // To further other optimizations, loop over all users of NewGV and try to
943 // constant prop them. This will promote GEP instructions with constant
944 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
945 ConstantPropUsersOf(NewGV, DL, TLI);
946 if (RepValue != NewGV)
947 ConstantPropUsersOf(RepValue, DL, TLI);
952 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
953 /// to make sure that there are no complex uses of V. We permit simple things
954 /// like dereferencing the pointer, but not storing through the address, unless
955 /// it is to the specified global.
956 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(const Instruction *V,
957 const GlobalVariable *GV,
958 SmallPtrSet<const PHINode*, 8> &PHIs) {
959 for (const User *U : V->users()) {
960 const Instruction *Inst = cast<Instruction>(U);
962 if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) {
963 continue; // Fine, ignore.
966 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
967 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
968 return false; // Storing the pointer itself... bad.
969 continue; // Otherwise, storing through it, or storing into GV... fine.
972 // Must index into the array and into the struct.
973 if (isa<GetElementPtrInst>(Inst) && Inst->getNumOperands() >= 3) {
974 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs))
979 if (const PHINode *PN = dyn_cast<PHINode>(Inst)) {
980 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI
983 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
988 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) {
989 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs))
999 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
1000 /// somewhere. Transform all uses of the allocation into loads from the
1001 /// global and uses of the resultant pointer. Further, delete the store into
1002 /// GV. This assumes that these value pass the
1003 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
1004 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
1005 GlobalVariable *GV) {
1006 while (!Alloc->use_empty()) {
1007 Instruction *U = cast<Instruction>(*Alloc->user_begin());
1008 Instruction *InsertPt = U;
1009 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1010 // If this is the store of the allocation into the global, remove it.
1011 if (SI->getOperand(1) == GV) {
1012 SI->eraseFromParent();
1015 } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
1016 // Insert the load in the corresponding predecessor, not right before the
1018 InsertPt = PN->getIncomingBlock(*Alloc->use_begin())->getTerminator();
1019 } else if (isa<BitCastInst>(U)) {
1020 // Must be bitcast between the malloc and store to initialize the global.
1021 ReplaceUsesOfMallocWithGlobal(U, GV);
1022 U->eraseFromParent();
1024 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
1025 // If this is a "GEP bitcast" and the user is a store to the global, then
1026 // just process it as a bitcast.
1027 if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse())
1028 if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->user_back()))
1029 if (SI->getOperand(1) == GV) {
1030 // Must be bitcast GEP between the malloc and store to initialize
1032 ReplaceUsesOfMallocWithGlobal(GEPI, GV);
1033 GEPI->eraseFromParent();
1038 // Insert a load from the global, and use it instead of the malloc.
1039 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
1040 U->replaceUsesOfWith(Alloc, NL);
1044 /// LoadUsesSimpleEnoughForHeapSRA - Verify that all uses of V (a load, or a phi
1045 /// of a load) are simple enough to perform heap SRA on. This permits GEP's
1046 /// that index through the array and struct field, icmps of null, and PHIs.
1047 static bool LoadUsesSimpleEnoughForHeapSRA(const Value *V,
1048 SmallPtrSet<const PHINode*, 32> &LoadUsingPHIs,
1049 SmallPtrSet<const PHINode*, 32> &LoadUsingPHIsPerLoad) {
1050 // We permit two users of the load: setcc comparing against the null
1051 // pointer, and a getelementptr of a specific form.
1052 for (const User *U : V->users()) {
1053 const Instruction *UI = cast<Instruction>(U);
1055 // Comparison against null is ok.
1056 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UI)) {
1057 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1062 // getelementptr is also ok, but only a simple form.
1063 if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(UI)) {
1064 // Must index into the array and into the struct.
1065 if (GEPI->getNumOperands() < 3)
1068 // Otherwise the GEP is ok.
1072 if (const PHINode *PN = dyn_cast<PHINode>(UI)) {
1073 if (!LoadUsingPHIsPerLoad.insert(PN))
1074 // This means some phi nodes are dependent on each other.
1075 // Avoid infinite looping!
1077 if (!LoadUsingPHIs.insert(PN))
1078 // If we have already analyzed this PHI, then it is safe.
1081 // Make sure all uses of the PHI are simple enough to transform.
1082 if (!LoadUsesSimpleEnoughForHeapSRA(PN,
1083 LoadUsingPHIs, LoadUsingPHIsPerLoad))
1089 // Otherwise we don't know what this is, not ok.
1097 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
1098 /// GV are simple enough to perform HeapSRA, return true.
1099 static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(const GlobalVariable *GV,
1100 Instruction *StoredVal) {
1101 SmallPtrSet<const PHINode*, 32> LoadUsingPHIs;
1102 SmallPtrSet<const PHINode*, 32> LoadUsingPHIsPerLoad;
1103 for (const User *U : GV->users())
1104 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
1105 if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs,
1106 LoadUsingPHIsPerLoad))
1108 LoadUsingPHIsPerLoad.clear();
1111 // If we reach here, we know that all uses of the loads and transitive uses
1112 // (through PHI nodes) are simple enough to transform. However, we don't know
1113 // that all inputs the to the PHI nodes are in the same equivalence sets.
1114 // Check to verify that all operands of the PHIs are either PHIS that can be
1115 // transformed, loads from GV, or MI itself.
1116 for (SmallPtrSet<const PHINode*, 32>::const_iterator I = LoadUsingPHIs.begin()
1117 , E = LoadUsingPHIs.end(); I != E; ++I) {
1118 const PHINode *PN = *I;
1119 for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) {
1120 Value *InVal = PN->getIncomingValue(op);
1122 // PHI of the stored value itself is ok.
1123 if (InVal == StoredVal) continue;
1125 if (const PHINode *InPN = dyn_cast<PHINode>(InVal)) {
1126 // One of the PHIs in our set is (optimistically) ok.
1127 if (LoadUsingPHIs.count(InPN))
1132 // Load from GV is ok.
1133 if (const LoadInst *LI = dyn_cast<LoadInst>(InVal))
1134 if (LI->getOperand(0) == GV)
1139 // Anything else is rejected.
1147 static Value *GetHeapSROAValue(Value *V, unsigned FieldNo,
1148 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1149 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1150 std::vector<Value*> &FieldVals = InsertedScalarizedValues[V];
1152 if (FieldNo >= FieldVals.size())
1153 FieldVals.resize(FieldNo+1);
1155 // If we already have this value, just reuse the previously scalarized
1157 if (Value *FieldVal = FieldVals[FieldNo])
1160 // Depending on what instruction this is, we have several cases.
1162 if (LoadInst *LI = dyn_cast<LoadInst>(V)) {
1163 // This is a scalarized version of the load from the global. Just create
1164 // a new Load of the scalarized global.
1165 Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo,
1166 InsertedScalarizedValues,
1168 LI->getName()+".f"+Twine(FieldNo), LI);
1169 } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1170 // PN's type is pointer to struct. Make a new PHI of pointer to struct
1173 PointerType *PTy = cast<PointerType>(PN->getType());
1174 StructType *ST = cast<StructType>(PTy->getElementType());
1176 unsigned AS = PTy->getAddressSpace();
1178 PHINode::Create(PointerType::get(ST->getElementType(FieldNo), AS),
1179 PN->getNumIncomingValues(),
1180 PN->getName()+".f"+Twine(FieldNo), PN);
1182 PHIsToRewrite.push_back(std::make_pair(PN, FieldNo));
1184 llvm_unreachable("Unknown usable value");
1187 return FieldVals[FieldNo] = Result;
1190 /// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1191 /// the load, rewrite the derived value to use the HeapSRoA'd load.
1192 static void RewriteHeapSROALoadUser(Instruction *LoadUser,
1193 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1194 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1195 // If this is a comparison against null, handle it.
1196 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1197 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1198 // If we have a setcc of the loaded pointer, we can use a setcc of any
1200 Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0,
1201 InsertedScalarizedValues, PHIsToRewrite);
1203 Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr,
1204 Constant::getNullValue(NPtr->getType()),
1206 SCI->replaceAllUsesWith(New);
1207 SCI->eraseFromParent();
1211 // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...'
1212 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1213 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1214 && "Unexpected GEPI!");
1216 // Load the pointer for this field.
1217 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1218 Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo,
1219 InsertedScalarizedValues, PHIsToRewrite);
1221 // Create the new GEP idx vector.
1222 SmallVector<Value*, 8> GEPIdx;
1223 GEPIdx.push_back(GEPI->getOperand(1));
1224 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1226 Value *NGEPI = GetElementPtrInst::Create(NewPtr, GEPIdx,
1227 GEPI->getName(), GEPI);
1228 GEPI->replaceAllUsesWith(NGEPI);
1229 GEPI->eraseFromParent();
1233 // Recursively transform the users of PHI nodes. This will lazily create the
1234 // PHIs that are needed for individual elements. Keep track of what PHIs we
1235 // see in InsertedScalarizedValues so that we don't get infinite loops (very
1236 // antisocial). If the PHI is already in InsertedScalarizedValues, it has
1237 // already been seen first by another load, so its uses have already been
1239 PHINode *PN = cast<PHINode>(LoadUser);
1240 if (!InsertedScalarizedValues.insert(std::make_pair(PN,
1241 std::vector<Value*>())).second)
1244 // If this is the first time we've seen this PHI, recursively process all
1246 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
1247 Instruction *User = cast<Instruction>(*UI++);
1248 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1252 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1253 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1254 /// use FieldGlobals instead. All uses of loaded values satisfy
1255 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA.
1256 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1257 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1258 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1259 for (auto UI = Load->user_begin(), E = Load->user_end(); UI != E;) {
1260 Instruction *User = cast<Instruction>(*UI++);
1261 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1264 if (Load->use_empty()) {
1265 Load->eraseFromParent();
1266 InsertedScalarizedValues.erase(Load);
1270 /// PerformHeapAllocSRoA - CI is an allocation of an array of structures. Break
1271 /// it up into multiple allocations of arrays of the fields.
1272 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI,
1273 Value *NElems, const DataLayout *DL,
1274 const TargetLibraryInfo *TLI) {
1275 DEBUG(dbgs() << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *CI << '\n');
1276 Type *MAT = getMallocAllocatedType(CI, TLI);
1277 StructType *STy = cast<StructType>(MAT);
1279 // There is guaranteed to be at least one use of the malloc (storing
1280 // it into GV). If there are other uses, change them to be uses of
1281 // the global to simplify later code. This also deletes the store
1283 ReplaceUsesOfMallocWithGlobal(CI, GV);
1285 // Okay, at this point, there are no users of the malloc. Insert N
1286 // new mallocs at the same place as CI, and N globals.
1287 std::vector<Value*> FieldGlobals;
1288 std::vector<Value*> FieldMallocs;
1290 unsigned AS = GV->getType()->getPointerAddressSpace();
1291 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1292 Type *FieldTy = STy->getElementType(FieldNo);
1293 PointerType *PFieldTy = PointerType::get(FieldTy, AS);
1295 GlobalVariable *NGV =
1296 new GlobalVariable(*GV->getParent(),
1297 PFieldTy, false, GlobalValue::InternalLinkage,
1298 Constant::getNullValue(PFieldTy),
1299 GV->getName() + ".f" + Twine(FieldNo), GV,
1300 GV->getThreadLocalMode());
1301 FieldGlobals.push_back(NGV);
1303 unsigned TypeSize = DL->getTypeAllocSize(FieldTy);
1304 if (StructType *ST = dyn_cast<StructType>(FieldTy))
1305 TypeSize = DL->getStructLayout(ST)->getSizeInBytes();
1306 Type *IntPtrTy = DL->getIntPtrType(CI->getType());
1307 Value *NMI = CallInst::CreateMalloc(CI, IntPtrTy, FieldTy,
1308 ConstantInt::get(IntPtrTy, TypeSize),
1310 CI->getName() + ".f" + Twine(FieldNo));
1311 FieldMallocs.push_back(NMI);
1312 new StoreInst(NMI, NGV, CI);
1315 // The tricky aspect of this transformation is handling the case when malloc
1316 // fails. In the original code, malloc failing would set the result pointer
1317 // of malloc to null. In this case, some mallocs could succeed and others
1318 // could fail. As such, we emit code that looks like this:
1319 // F0 = malloc(field0)
1320 // F1 = malloc(field1)
1321 // F2 = malloc(field2)
1322 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1323 // if (F0) { free(F0); F0 = 0; }
1324 // if (F1) { free(F1); F1 = 0; }
1325 // if (F2) { free(F2); F2 = 0; }
1327 // The malloc can also fail if its argument is too large.
1328 Constant *ConstantZero = ConstantInt::get(CI->getArgOperand(0)->getType(), 0);
1329 Value *RunningOr = new ICmpInst(CI, ICmpInst::ICMP_SLT, CI->getArgOperand(0),
1330 ConstantZero, "isneg");
1331 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1332 Value *Cond = new ICmpInst(CI, ICmpInst::ICMP_EQ, FieldMallocs[i],
1333 Constant::getNullValue(FieldMallocs[i]->getType()),
1335 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", CI);
1338 // Split the basic block at the old malloc.
1339 BasicBlock *OrigBB = CI->getParent();
1340 BasicBlock *ContBB = OrigBB->splitBasicBlock(CI, "malloc_cont");
1342 // Create the block to check the first condition. Put all these blocks at the
1343 // end of the function as they are unlikely to be executed.
1344 BasicBlock *NullPtrBlock = BasicBlock::Create(OrigBB->getContext(),
1346 OrigBB->getParent());
1348 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1349 // branch on RunningOr.
1350 OrigBB->getTerminator()->eraseFromParent();
1351 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1353 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1354 // pointer, because some may be null while others are not.
1355 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1356 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1357 Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
1358 Constant::getNullValue(GVVal->getType()));
1359 BasicBlock *FreeBlock = BasicBlock::Create(Cmp->getContext(), "free_it",
1360 OrigBB->getParent());
1361 BasicBlock *NextBlock = BasicBlock::Create(Cmp->getContext(), "next",
1362 OrigBB->getParent());
1363 Instruction *BI = BranchInst::Create(FreeBlock, NextBlock,
1366 // Fill in FreeBlock.
1367 CallInst::CreateFree(GVVal, BI);
1368 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1370 BranchInst::Create(NextBlock, FreeBlock);
1372 NullPtrBlock = NextBlock;
1375 BranchInst::Create(ContBB, NullPtrBlock);
1377 // CI is no longer needed, remove it.
1378 CI->eraseFromParent();
1380 /// InsertedScalarizedLoads - As we process loads, if we can't immediately
1381 /// update all uses of the load, keep track of what scalarized loads are
1382 /// inserted for a given load.
1383 DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues;
1384 InsertedScalarizedValues[GV] = FieldGlobals;
1386 std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite;
1388 // Okay, the malloc site is completely handled. All of the uses of GV are now
1389 // loads, and all uses of those loads are simple. Rewrite them to use loads
1390 // of the per-field globals instead.
1391 for (auto UI = GV->user_begin(), E = GV->user_end(); UI != E;) {
1392 Instruction *User = cast<Instruction>(*UI++);
1394 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1395 RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite);
1399 // Must be a store of null.
1400 StoreInst *SI = cast<StoreInst>(User);
1401 assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
1402 "Unexpected heap-sra user!");
1404 // Insert a store of null into each global.
1405 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1406 PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType());
1407 Constant *Null = Constant::getNullValue(PT->getElementType());
1408 new StoreInst(Null, FieldGlobals[i], SI);
1410 // Erase the original store.
1411 SI->eraseFromParent();
1414 // While we have PHIs that are interesting to rewrite, do it.
1415 while (!PHIsToRewrite.empty()) {
1416 PHINode *PN = PHIsToRewrite.back().first;
1417 unsigned FieldNo = PHIsToRewrite.back().second;
1418 PHIsToRewrite.pop_back();
1419 PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]);
1420 assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi");
1422 // Add all the incoming values. This can materialize more phis.
1423 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1424 Value *InVal = PN->getIncomingValue(i);
1425 InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues,
1427 FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
1431 // Drop all inter-phi links and any loads that made it this far.
1432 for (DenseMap<Value*, std::vector<Value*> >::iterator
1433 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1435 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1436 PN->dropAllReferences();
1437 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1438 LI->dropAllReferences();
1441 // Delete all the phis and loads now that inter-references are dead.
1442 for (DenseMap<Value*, std::vector<Value*> >::iterator
1443 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1445 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1446 PN->eraseFromParent();
1447 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1448 LI->eraseFromParent();
1451 // The old global is now dead, remove it.
1452 GV->eraseFromParent();
1455 return cast<GlobalVariable>(FieldGlobals[0]);
1458 /// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1459 /// pointer global variable with a single value stored it that is a malloc or
1461 static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
1464 AtomicOrdering Ordering,
1465 Module::global_iterator &GVI,
1466 const DataLayout *DL,
1467 TargetLibraryInfo *TLI) {
1471 // If this is a malloc of an abstract type, don't touch it.
1472 if (!AllocTy->isSized())
1475 // We can't optimize this global unless all uses of it are *known* to be
1476 // of the malloc value, not of the null initializer value (consider a use
1477 // that compares the global's value against zero to see if the malloc has
1478 // been reached). To do this, we check to see if all uses of the global
1479 // would trap if the global were null: this proves that they must all
1480 // happen after the malloc.
1481 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1484 // We can't optimize this if the malloc itself is used in a complex way,
1485 // for example, being stored into multiple globals. This allows the
1486 // malloc to be stored into the specified global, loaded icmp'd, and
1487 // GEP'd. These are all things we could transform to using the global
1489 SmallPtrSet<const PHINode*, 8> PHIs;
1490 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV, PHIs))
1493 // If we have a global that is only initialized with a fixed size malloc,
1494 // transform the program to use global memory instead of malloc'd memory.
1495 // This eliminates dynamic allocation, avoids an indirection accessing the
1496 // data, and exposes the resultant global to further GlobalOpt.
1497 // We cannot optimize the malloc if we cannot determine malloc array size.
1498 Value *NElems = getMallocArraySize(CI, DL, TLI, true);
1502 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
1503 // Restrict this transformation to only working on small allocations
1504 // (2048 bytes currently), as we don't want to introduce a 16M global or
1506 if (NElements->getZExtValue() * DL->getTypeAllocSize(AllocTy) < 2048) {
1507 GVI = OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, DL, TLI);
1511 // If the allocation is an array of structures, consider transforming this
1512 // into multiple malloc'd arrays, one for each field. This is basically
1513 // SRoA for malloc'd memory.
1515 if (Ordering != NotAtomic)
1518 // If this is an allocation of a fixed size array of structs, analyze as a
1519 // variable size array. malloc [100 x struct],1 -> malloc struct, 100
1520 if (NElems == ConstantInt::get(CI->getArgOperand(0)->getType(), 1))
1521 if (ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
1522 AllocTy = AT->getElementType();
1524 StructType *AllocSTy = dyn_cast<StructType>(AllocTy);
1528 // This the structure has an unreasonable number of fields, leave it
1530 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
1531 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, CI)) {
1533 // If this is a fixed size array, transform the Malloc to be an alloc of
1534 // structs. malloc [100 x struct],1 -> malloc struct, 100
1535 if (ArrayType *AT = dyn_cast<ArrayType>(getMallocAllocatedType(CI, TLI))) {
1536 Type *IntPtrTy = DL->getIntPtrType(CI->getType());
1537 unsigned TypeSize = DL->getStructLayout(AllocSTy)->getSizeInBytes();
1538 Value *AllocSize = ConstantInt::get(IntPtrTy, TypeSize);
1539 Value *NumElements = ConstantInt::get(IntPtrTy, AT->getNumElements());
1540 Instruction *Malloc = CallInst::CreateMalloc(CI, IntPtrTy, AllocSTy,
1541 AllocSize, NumElements,
1542 nullptr, CI->getName());
1543 Instruction *Cast = new BitCastInst(Malloc, CI->getType(), "tmp", CI);
1544 CI->replaceAllUsesWith(Cast);
1545 CI->eraseFromParent();
1546 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Malloc))
1547 CI = cast<CallInst>(BCI->getOperand(0));
1549 CI = cast<CallInst>(Malloc);
1552 GVI = PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, DL, TLI, true),
1560 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1561 // that only one value (besides its initializer) is ever stored to the global.
1562 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1563 AtomicOrdering Ordering,
1564 Module::global_iterator &GVI,
1565 const DataLayout *DL,
1566 TargetLibraryInfo *TLI) {
1567 // Ignore no-op GEPs and bitcasts.
1568 StoredOnceVal = StoredOnceVal->stripPointerCasts();
1570 // If we are dealing with a pointer global that is initialized to null and
1571 // only has one (non-null) value stored into it, then we can optimize any
1572 // users of the loaded value (often calls and loads) that would trap if the
1574 if (GV->getInitializer()->getType()->isPointerTy() &&
1575 GV->getInitializer()->isNullValue()) {
1576 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1577 if (GV->getInitializer()->getType() != SOVC->getType())
1578 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1580 // Optimize away any trapping uses of the loaded value.
1581 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC, DL, TLI))
1583 } else if (CallInst *CI = extractMallocCall(StoredOnceVal, TLI)) {
1584 Type *MallocType = getMallocAllocatedType(CI, TLI);
1586 TryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType, Ordering, GVI,
1595 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1596 /// two values ever stored into GV are its initializer and OtherVal. See if we
1597 /// can shrink the global into a boolean and select between the two values
1598 /// whenever it is used. This exposes the values to other scalar optimizations.
1599 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1600 Type *GVElType = GV->getType()->getElementType();
1602 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1603 // an FP value, pointer or vector, don't do this optimization because a select
1604 // between them is very expensive and unlikely to lead to later
1605 // simplification. In these cases, we typically end up with "cond ? v1 : v2"
1606 // where v1 and v2 both require constant pool loads, a big loss.
1607 if (GVElType == Type::getInt1Ty(GV->getContext()) ||
1608 GVElType->isFloatingPointTy() ||
1609 GVElType->isPointerTy() || GVElType->isVectorTy())
1612 // Walk the use list of the global seeing if all the uses are load or store.
1613 // If there is anything else, bail out.
1614 for (User *U : GV->users())
1615 if (!isa<LoadInst>(U) && !isa<StoreInst>(U))
1618 DEBUG(dbgs() << " *** SHRINKING TO BOOL: " << *GV);
1620 // Create the new global, initializing it to false.
1621 GlobalVariable *NewGV = new GlobalVariable(Type::getInt1Ty(GV->getContext()),
1623 GlobalValue::InternalLinkage,
1624 ConstantInt::getFalse(GV->getContext()),
1626 GV->getThreadLocalMode(),
1627 GV->getType()->getAddressSpace());
1628 GV->getParent()->getGlobalList().insert(GV, NewGV);
1630 Constant *InitVal = GV->getInitializer();
1631 assert(InitVal->getType() != Type::getInt1Ty(GV->getContext()) &&
1632 "No reason to shrink to bool!");
1634 // If initialized to zero and storing one into the global, we can use a cast
1635 // instead of a select to synthesize the desired value.
1636 bool IsOneZero = false;
1637 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1638 IsOneZero = InitVal->isNullValue() && CI->isOne();
1640 while (!GV->use_empty()) {
1641 Instruction *UI = cast<Instruction>(GV->user_back());
1642 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1643 // Change the store into a boolean store.
1644 bool StoringOther = SI->getOperand(0) == OtherVal;
1645 // Only do this if we weren't storing a loaded value.
1647 if (StoringOther || SI->getOperand(0) == InitVal) {
1648 StoreVal = ConstantInt::get(Type::getInt1Ty(GV->getContext()),
1651 // Otherwise, we are storing a previously loaded copy. To do this,
1652 // change the copy from copying the original value to just copying the
1654 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1656 // If we've already replaced the input, StoredVal will be a cast or
1657 // select instruction. If not, it will be a load of the original
1659 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1660 assert(LI->getOperand(0) == GV && "Not a copy!");
1661 // Insert a new load, to preserve the saved value.
1662 StoreVal = new LoadInst(NewGV, LI->getName()+".b", false, 0,
1663 LI->getOrdering(), LI->getSynchScope(), LI);
1665 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1666 "This is not a form that we understand!");
1667 StoreVal = StoredVal->getOperand(0);
1668 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1671 new StoreInst(StoreVal, NewGV, false, 0,
1672 SI->getOrdering(), SI->getSynchScope(), SI);
1674 // Change the load into a load of bool then a select.
1675 LoadInst *LI = cast<LoadInst>(UI);
1676 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", false, 0,
1677 LI->getOrdering(), LI->getSynchScope(), LI);
1680 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1682 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1684 LI->replaceAllUsesWith(NSI);
1686 UI->eraseFromParent();
1689 // Retain the name of the old global variable. People who are debugging their
1690 // programs may expect these variables to be named the same.
1691 NewGV->takeName(GV);
1692 GV->eraseFromParent();
1697 /// ProcessGlobal - Analyze the specified global variable and optimize it if
1698 /// possible. If we make a change, return true.
1699 bool GlobalOpt::ProcessGlobal(GlobalVariable *GV,
1700 Module::global_iterator &GVI) {
1701 if (!GV->isDiscardableIfUnused())
1704 // Do more involved optimizations if the global is internal.
1705 GV->removeDeadConstantUsers();
1707 if (GV->use_empty()) {
1708 DEBUG(dbgs() << "GLOBAL DEAD: " << *GV);
1709 GV->eraseFromParent();
1714 if (!GV->hasLocalLinkage())
1719 if (GlobalStatus::analyzeGlobal(GV, GS))
1722 if (!GS.IsCompared && !GV->hasUnnamedAddr()) {
1723 GV->setUnnamedAddr(true);
1727 if (GV->isConstant() || !GV->hasInitializer())
1730 return ProcessInternalGlobal(GV, GVI, GS);
1733 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1734 /// it if possible. If we make a change, return true.
1735 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1736 Module::global_iterator &GVI,
1737 const GlobalStatus &GS) {
1738 // If this is a first class global and has only one accessing function
1739 // and this function is main (which we know is not recursive), we replace
1740 // the global with a local alloca in this function.
1742 // NOTE: It doesn't make sense to promote non-single-value types since we
1743 // are just replacing static memory to stack memory.
1745 // If the global is in different address space, don't bring it to stack.
1746 if (!GS.HasMultipleAccessingFunctions &&
1747 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1748 GV->getType()->getElementType()->isSingleValueType() &&
1749 GS.AccessingFunction->getName() == "main" &&
1750 GS.AccessingFunction->hasExternalLinkage() &&
1751 GV->getType()->getAddressSpace() == 0) {
1752 DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV);
1753 Instruction &FirstI = const_cast<Instruction&>(*GS.AccessingFunction
1754 ->getEntryBlock().begin());
1755 Type *ElemTy = GV->getType()->getElementType();
1756 // FIXME: Pass Global's alignment when globals have alignment
1757 AllocaInst *Alloca = new AllocaInst(ElemTy, nullptr,
1758 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);
1966 isSimpleEnoughValueToCommit(Constant *C,
1967 SmallPtrSet<Constant*, 8> &SimpleConstants,
1968 const DataLayout *DL);
1971 /// isSimpleEnoughValueToCommit - Return true if the specified constant can be
1972 /// handled by the code generator. We don't want to generate something like:
1973 /// void *X = &X/42;
1974 /// because the code generator doesn't have a relocation that can handle that.
1976 /// This function should be called if C was not found (but just got inserted)
1977 /// in SimpleConstants to avoid having to rescan the same constants all the
1979 static bool isSimpleEnoughValueToCommitHelper(Constant *C,
1980 SmallPtrSet<Constant*, 8> &SimpleConstants,
1981 const DataLayout *DL) {
1982 // Simple integer, undef, constant aggregate zero, global addresses, etc are
1984 if (C->getNumOperands() == 0 || isa<BlockAddress>(C) ||
1985 isa<GlobalValue>(C))
1988 // Aggregate values are safe if all their elements are.
1989 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C) ||
1990 isa<ConstantVector>(C)) {
1991 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) {
1992 Constant *Op = cast<Constant>(C->getOperand(i));
1993 if (!isSimpleEnoughValueToCommit(Op, SimpleConstants, DL))
1999 // We don't know exactly what relocations are allowed in constant expressions,
2000 // so we allow &global+constantoffset, which is safe and uniformly supported
2002 ConstantExpr *CE = cast<ConstantExpr>(C);
2003 switch (CE->getOpcode()) {
2004 case Instruction::BitCast:
2005 // Bitcast is fine if the casted value is fine.
2006 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL);
2008 case Instruction::IntToPtr:
2009 case Instruction::PtrToInt:
2010 // int <=> ptr is fine if the int type is the same size as the
2012 if (!DL || DL->getTypeSizeInBits(CE->getType()) !=
2013 DL->getTypeSizeInBits(CE->getOperand(0)->getType()))
2015 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL);
2017 // GEP is fine if it is simple + constant offset.
2018 case Instruction::GetElementPtr:
2019 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
2020 if (!isa<ConstantInt>(CE->getOperand(i)))
2022 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL);
2024 case Instruction::Add:
2025 // We allow simple+cst.
2026 if (!isa<ConstantInt>(CE->getOperand(1)))
2028 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL);
2034 isSimpleEnoughValueToCommit(Constant *C,
2035 SmallPtrSet<Constant*, 8> &SimpleConstants,
2036 const DataLayout *DL) {
2037 // If we already checked this constant, we win.
2038 if (!SimpleConstants.insert(C)) return true;
2039 // Check the constant.
2040 return isSimpleEnoughValueToCommitHelper(C, SimpleConstants, DL);
2044 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
2045 /// enough for us to understand. In particular, if it is a cast to anything
2046 /// other than from one pointer type to another pointer type, we punt.
2047 /// We basically just support direct accesses to globals and GEP's of
2048 /// globals. This should be kept up to date with CommitValueTo.
2049 static bool isSimpleEnoughPointerToCommit(Constant *C) {
2050 // Conservatively, avoid aggregate types. This is because we don't
2051 // want to worry about them partially overlapping other stores.
2052 if (!cast<PointerType>(C->getType())->getElementType()->isSingleValueType())
2055 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
2056 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2057 // external globals.
2058 return GV->hasUniqueInitializer();
2060 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2061 // Handle a constantexpr gep.
2062 if (CE->getOpcode() == Instruction::GetElementPtr &&
2063 isa<GlobalVariable>(CE->getOperand(0)) &&
2064 cast<GEPOperator>(CE)->isInBounds()) {
2065 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2066 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2067 // external globals.
2068 if (!GV->hasUniqueInitializer())
2071 // The first index must be zero.
2072 ConstantInt *CI = dyn_cast<ConstantInt>(*std::next(CE->op_begin()));
2073 if (!CI || !CI->isZero()) return false;
2075 // The remaining indices must be compile-time known integers within the
2076 // notional bounds of the corresponding static array types.
2077 if (!CE->isGEPWithNoNotionalOverIndexing())
2080 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2082 // A constantexpr bitcast from a pointer to another pointer is a no-op,
2083 // and we know how to evaluate it by moving the bitcast from the pointer
2084 // operand to the value operand.
2085 } else if (CE->getOpcode() == Instruction::BitCast &&
2086 isa<GlobalVariable>(CE->getOperand(0))) {
2087 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2088 // external globals.
2089 return cast<GlobalVariable>(CE->getOperand(0))->hasUniqueInitializer();
2096 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
2097 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
2098 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
2099 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
2100 ConstantExpr *Addr, unsigned OpNo) {
2101 // Base case of the recursion.
2102 if (OpNo == Addr->getNumOperands()) {
2103 assert(Val->getType() == Init->getType() && "Type mismatch!");
2107 SmallVector<Constant*, 32> Elts;
2108 if (StructType *STy = dyn_cast<StructType>(Init->getType())) {
2109 // Break up the constant into its elements.
2110 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2111 Elts.push_back(Init->getAggregateElement(i));
2113 // Replace the element that we are supposed to.
2114 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
2115 unsigned Idx = CU->getZExtValue();
2116 assert(Idx < STy->getNumElements() && "Struct index out of range!");
2117 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
2119 // Return the modified struct.
2120 return ConstantStruct::get(STy, Elts);
2123 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
2124 SequentialType *InitTy = cast<SequentialType>(Init->getType());
2127 if (ArrayType *ATy = dyn_cast<ArrayType>(InitTy))
2128 NumElts = ATy->getNumElements();
2130 NumElts = InitTy->getVectorNumElements();
2132 // Break up the array into elements.
2133 for (uint64_t i = 0, e = NumElts; i != e; ++i)
2134 Elts.push_back(Init->getAggregateElement(i));
2136 assert(CI->getZExtValue() < NumElts);
2137 Elts[CI->getZExtValue()] =
2138 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
2140 if (Init->getType()->isArrayTy())
2141 return ConstantArray::get(cast<ArrayType>(InitTy), Elts);
2142 return ConstantVector::get(Elts);
2145 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
2146 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
2147 static void CommitValueTo(Constant *Val, Constant *Addr) {
2148 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
2149 assert(GV->hasInitializer());
2150 GV->setInitializer(Val);
2154 ConstantExpr *CE = cast<ConstantExpr>(Addr);
2155 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2156 GV->setInitializer(EvaluateStoreInto(GV->getInitializer(), Val, CE, 2));
2161 /// Evaluator - This class evaluates LLVM IR, producing the Constant
2162 /// representing each SSA instruction. Changes to global variables are stored
2163 /// in a mapping that can be iterated over after the evaluation is complete.
2164 /// Once an evaluation call fails, the evaluation object should not be reused.
2167 Evaluator(const DataLayout *DL, const TargetLibraryInfo *TLI)
2168 : DL(DL), TLI(TLI) {
2169 ValueStack.push_back(make_unique<DenseMap<Value*, Constant*>>());
2173 for (auto &Tmp : AllocaTmps)
2174 // If there are still users of the alloca, the program is doing something
2175 // silly, e.g. storing the address of the alloca somewhere and using it
2176 // later. Since this is undefined, we'll just make it be null.
2177 if (!Tmp->use_empty())
2178 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2181 /// EvaluateFunction - Evaluate a call to function F, returning true if
2182 /// successful, false if we can't evaluate it. ActualArgs contains the formal
2183 /// arguments for the function.
2184 bool EvaluateFunction(Function *F, Constant *&RetVal,
2185 const SmallVectorImpl<Constant*> &ActualArgs);
2187 /// EvaluateBlock - Evaluate all instructions in block BB, returning true if
2188 /// successful, false if we can't evaluate it. NewBB returns the next BB that
2189 /// control flows into, or null upon return.
2190 bool EvaluateBlock(BasicBlock::iterator CurInst, BasicBlock *&NextBB);
2192 Constant *getVal(Value *V) {
2193 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
2194 Constant *R = ValueStack.back()->lookup(V);
2195 assert(R && "Reference to an uncomputed value!");
2199 void setVal(Value *V, Constant *C) {
2200 (*ValueStack.back())[V] = C;
2203 const DenseMap<Constant*, Constant*> &getMutatedMemory() const {
2204 return MutatedMemory;
2207 const SmallPtrSet<GlobalVariable*, 8> &getInvariants() const {
2212 Constant *ComputeLoadResult(Constant *P);
2214 /// ValueStack - As we compute SSA register values, we store their contents
2215 /// here. The back of the vector contains the current function and the stack
2216 /// contains the values in the calling frames.
2217 SmallVector<std::unique_ptr<DenseMap<Value*, Constant*>>, 4> ValueStack;
2219 /// CallStack - This is used to detect recursion. In pathological situations
2220 /// we could hit exponential behavior, but at least there is nothing
2222 SmallVector<Function*, 4> CallStack;
2224 /// MutatedMemory - For each store we execute, we update this map. Loads
2225 /// check this to get the most up-to-date value. If evaluation is successful,
2226 /// this state is committed to the process.
2227 DenseMap<Constant*, Constant*> MutatedMemory;
2229 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2230 /// to represent its body. This vector is needed so we can delete the
2231 /// temporary globals when we are done.
2232 SmallVector<std::unique_ptr<GlobalVariable>, 32> AllocaTmps;
2234 /// Invariants - These global variables have been marked invariant by the
2235 /// static constructor.
2236 SmallPtrSet<GlobalVariable*, 8> Invariants;
2238 /// SimpleConstants - These are constants we have checked and know to be
2239 /// simple enough to live in a static initializer of a global.
2240 SmallPtrSet<Constant*, 8> SimpleConstants;
2242 const DataLayout *DL;
2243 const TargetLibraryInfo *TLI;
2246 } // anonymous namespace
2248 /// ComputeLoadResult - Return the value that would be computed by a load from
2249 /// P after the stores reflected by 'memory' have been performed. If we can't
2250 /// decide, return null.
2251 Constant *Evaluator::ComputeLoadResult(Constant *P) {
2252 // If this memory location has been recently stored, use the stored value: it
2253 // is the most up-to-date.
2254 DenseMap<Constant*, Constant*>::const_iterator I = MutatedMemory.find(P);
2255 if (I != MutatedMemory.end()) return I->second;
2258 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
2259 if (GV->hasDefinitiveInitializer())
2260 return GV->getInitializer();
2264 // Handle a constantexpr getelementptr.
2265 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
2266 if (CE->getOpcode() == Instruction::GetElementPtr &&
2267 isa<GlobalVariable>(CE->getOperand(0))) {
2268 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2269 if (GV->hasDefinitiveInitializer())
2270 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2273 return nullptr; // don't know how to evaluate.
2276 /// EvaluateBlock - Evaluate all instructions in block BB, returning true if
2277 /// successful, false if we can't evaluate it. NewBB returns the next BB that
2278 /// control flows into, or null upon return.
2279 bool Evaluator::EvaluateBlock(BasicBlock::iterator CurInst,
2280 BasicBlock *&NextBB) {
2281 // This is the main evaluation loop.
2283 Constant *InstResult = nullptr;
2285 DEBUG(dbgs() << "Evaluating Instruction: " << *CurInst << "\n");
2287 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
2288 if (!SI->isSimple()) {
2289 DEBUG(dbgs() << "Store is not simple! Can not evaluate.\n");
2290 return false; // no volatile/atomic accesses.
2292 Constant *Ptr = getVal(SI->getOperand(1));
2293 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
2294 DEBUG(dbgs() << "Folding constant ptr expression: " << *Ptr);
2295 Ptr = ConstantFoldConstantExpression(CE, DL, TLI);
2296 DEBUG(dbgs() << "; To: " << *Ptr << "\n");
2298 if (!isSimpleEnoughPointerToCommit(Ptr)) {
2299 // If this is too complex for us to commit, reject it.
2300 DEBUG(dbgs() << "Pointer is too complex for us to evaluate store.");
2304 Constant *Val = getVal(SI->getOperand(0));
2306 // If this might be too difficult for the backend to handle (e.g. the addr
2307 // of one global variable divided by another) then we can't commit it.
2308 if (!isSimpleEnoughValueToCommit(Val, SimpleConstants, DL)) {
2309 DEBUG(dbgs() << "Store value is too complex to evaluate store. " << *Val
2314 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
2315 if (CE->getOpcode() == Instruction::BitCast) {
2316 DEBUG(dbgs() << "Attempting to resolve bitcast on constant ptr.\n");
2317 // If we're evaluating a store through a bitcast, then we need
2318 // to pull the bitcast off the pointer type and push it onto the
2320 Ptr = CE->getOperand(0);
2322 Type *NewTy = cast<PointerType>(Ptr->getType())->getElementType();
2324 // In order to push the bitcast onto the stored value, a bitcast
2325 // from NewTy to Val's type must be legal. If it's not, we can try
2326 // introspecting NewTy to find a legal conversion.
2327 while (!Val->getType()->canLosslesslyBitCastTo(NewTy)) {
2328 // If NewTy is a struct, we can convert the pointer to the struct
2329 // into a pointer to its first member.
2330 // FIXME: This could be extended to support arrays as well.
2331 if (StructType *STy = dyn_cast<StructType>(NewTy)) {
2332 NewTy = STy->getTypeAtIndex(0U);
2334 IntegerType *IdxTy = IntegerType::get(NewTy->getContext(), 32);
2335 Constant *IdxZero = ConstantInt::get(IdxTy, 0, false);
2336 Constant * const IdxList[] = {IdxZero, IdxZero};
2338 Ptr = ConstantExpr::getGetElementPtr(Ptr, IdxList);
2339 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
2340 Ptr = ConstantFoldConstantExpression(CE, DL, TLI);
2342 // If we can't improve the situation by introspecting NewTy,
2343 // we have to give up.
2345 DEBUG(dbgs() << "Failed to bitcast constant ptr, can not "
2351 // If we found compatible types, go ahead and push the bitcast
2352 // onto the stored value.
2353 Val = ConstantExpr::getBitCast(Val, NewTy);
2355 DEBUG(dbgs() << "Evaluated bitcast: " << *Val << "\n");
2359 MutatedMemory[Ptr] = Val;
2360 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2361 InstResult = ConstantExpr::get(BO->getOpcode(),
2362 getVal(BO->getOperand(0)),
2363 getVal(BO->getOperand(1)));
2364 DEBUG(dbgs() << "Found a BinaryOperator! Simplifying: " << *InstResult
2366 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2367 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2368 getVal(CI->getOperand(0)),
2369 getVal(CI->getOperand(1)));
2370 DEBUG(dbgs() << "Found a CmpInst! Simplifying: " << *InstResult
2372 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2373 InstResult = ConstantExpr::getCast(CI->getOpcode(),
2374 getVal(CI->getOperand(0)),
2376 DEBUG(dbgs() << "Found a Cast! Simplifying: " << *InstResult
2378 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2379 InstResult = ConstantExpr::getSelect(getVal(SI->getOperand(0)),
2380 getVal(SI->getOperand(1)),
2381 getVal(SI->getOperand(2)));
2382 DEBUG(dbgs() << "Found a Select! Simplifying: " << *InstResult
2384 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2385 Constant *P = getVal(GEP->getOperand(0));
2386 SmallVector<Constant*, 8> GEPOps;
2387 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2389 GEPOps.push_back(getVal(*i));
2391 ConstantExpr::getGetElementPtr(P, GEPOps,
2392 cast<GEPOperator>(GEP)->isInBounds());
2393 DEBUG(dbgs() << "Found a GEP! Simplifying: " << *InstResult
2395 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2397 if (!LI->isSimple()) {
2398 DEBUG(dbgs() << "Found a Load! Not a simple load, can not evaluate.\n");
2399 return false; // no volatile/atomic accesses.
2402 Constant *Ptr = getVal(LI->getOperand(0));
2403 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
2404 Ptr = ConstantFoldConstantExpression(CE, DL, TLI);
2405 DEBUG(dbgs() << "Found a constant pointer expression, constant "
2406 "folding: " << *Ptr << "\n");
2408 InstResult = ComputeLoadResult(Ptr);
2410 DEBUG(dbgs() << "Failed to compute load result. Can not evaluate load."
2412 return false; // Could not evaluate load.
2415 DEBUG(dbgs() << "Evaluated load: " << *InstResult << "\n");
2416 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2417 if (AI->isArrayAllocation()) {
2418 DEBUG(dbgs() << "Found an array alloca. Can not evaluate.\n");
2419 return false; // Cannot handle array allocs.
2421 Type *Ty = AI->getType()->getElementType();
2422 AllocaTmps.push_back(
2423 make_unique<GlobalVariable>(Ty, false, GlobalValue::InternalLinkage,
2424 UndefValue::get(Ty), AI->getName()));
2425 InstResult = AllocaTmps.back().get();
2426 DEBUG(dbgs() << "Found an alloca. Result: " << *InstResult << "\n");
2427 } else if (isa<CallInst>(CurInst) || isa<InvokeInst>(CurInst)) {
2428 CallSite CS(CurInst);
2430 // Debug info can safely be ignored here.
2431 if (isa<DbgInfoIntrinsic>(CS.getInstruction())) {
2432 DEBUG(dbgs() << "Ignoring debug info.\n");
2437 // Cannot handle inline asm.
2438 if (isa<InlineAsm>(CS.getCalledValue())) {
2439 DEBUG(dbgs() << "Found inline asm, can not evaluate.\n");
2443 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) {
2444 if (MemSetInst *MSI = dyn_cast<MemSetInst>(II)) {
2445 if (MSI->isVolatile()) {
2446 DEBUG(dbgs() << "Can not optimize a volatile memset " <<
2450 Constant *Ptr = getVal(MSI->getDest());
2451 Constant *Val = getVal(MSI->getValue());
2452 Constant *DestVal = ComputeLoadResult(getVal(Ptr));
2453 if (Val->isNullValue() && DestVal && DestVal->isNullValue()) {
2454 // This memset is a no-op.
2455 DEBUG(dbgs() << "Ignoring no-op memset.\n");
2461 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
2462 II->getIntrinsicID() == Intrinsic::lifetime_end) {
2463 DEBUG(dbgs() << "Ignoring lifetime intrinsic.\n");
2468 if (II->getIntrinsicID() == Intrinsic::invariant_start) {
2469 // We don't insert an entry into Values, as it doesn't have a
2470 // meaningful return value.
2471 if (!II->use_empty()) {
2472 DEBUG(dbgs() << "Found unused invariant_start. Can't evaluate.\n");
2475 ConstantInt *Size = cast<ConstantInt>(II->getArgOperand(0));
2476 Value *PtrArg = getVal(II->getArgOperand(1));
2477 Value *Ptr = PtrArg->stripPointerCasts();
2478 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) {
2479 Type *ElemTy = cast<PointerType>(GV->getType())->getElementType();
2480 if (DL && !Size->isAllOnesValue() &&
2481 Size->getValue().getLimitedValue() >=
2482 DL->getTypeStoreSize(ElemTy)) {
2483 Invariants.insert(GV);
2484 DEBUG(dbgs() << "Found a global var that is an invariant: " << *GV
2487 DEBUG(dbgs() << "Found a global var, but can not treat it as an "
2491 // Continue even if we do nothing.
2496 DEBUG(dbgs() << "Unknown intrinsic. Can not evaluate.\n");
2500 // Resolve function pointers.
2501 Function *Callee = dyn_cast<Function>(getVal(CS.getCalledValue()));
2502 if (!Callee || Callee->mayBeOverridden()) {
2503 DEBUG(dbgs() << "Can not resolve function pointer.\n");
2504 return false; // Cannot resolve.
2507 SmallVector<Constant*, 8> Formals;
2508 for (User::op_iterator i = CS.arg_begin(), e = CS.arg_end(); i != e; ++i)
2509 Formals.push_back(getVal(*i));
2511 if (Callee->isDeclaration()) {
2512 // If this is a function we can constant fold, do it.
2513 if (Constant *C = ConstantFoldCall(Callee, Formals, TLI)) {
2515 DEBUG(dbgs() << "Constant folded function call. Result: " <<
2516 *InstResult << "\n");
2518 DEBUG(dbgs() << "Can not constant fold function call.\n");
2522 if (Callee->getFunctionType()->isVarArg()) {
2523 DEBUG(dbgs() << "Can not constant fold vararg function call.\n");
2527 Constant *RetVal = nullptr;
2528 // Execute the call, if successful, use the return value.
2529 ValueStack.push_back(make_unique<DenseMap<Value *, Constant *>>());
2530 if (!EvaluateFunction(Callee, RetVal, Formals)) {
2531 DEBUG(dbgs() << "Failed to evaluate function.\n");
2534 ValueStack.pop_back();
2535 InstResult = RetVal;
2538 DEBUG(dbgs() << "Successfully evaluated function. Result: " <<
2539 InstResult << "\n\n");
2541 DEBUG(dbgs() << "Successfully evaluated function. Result: 0\n\n");
2544 } else if (isa<TerminatorInst>(CurInst)) {
2545 DEBUG(dbgs() << "Found a terminator instruction.\n");
2547 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2548 if (BI->isUnconditional()) {
2549 NextBB = BI->getSuccessor(0);
2552 dyn_cast<ConstantInt>(getVal(BI->getCondition()));
2553 if (!Cond) return false; // Cannot determine.
2555 NextBB = BI->getSuccessor(!Cond->getZExtValue());
2557 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2559 dyn_cast<ConstantInt>(getVal(SI->getCondition()));
2560 if (!Val) return false; // Cannot determine.
2561 NextBB = SI->findCaseValue(Val).getCaseSuccessor();
2562 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(CurInst)) {
2563 Value *Val = getVal(IBI->getAddress())->stripPointerCasts();
2564 if (BlockAddress *BA = dyn_cast<BlockAddress>(Val))
2565 NextBB = BA->getBasicBlock();
2567 return false; // Cannot determine.
2568 } else if (isa<ReturnInst>(CurInst)) {
2571 // invoke, unwind, resume, unreachable.
2572 DEBUG(dbgs() << "Can not handle terminator.");
2573 return false; // Cannot handle this terminator.
2576 // We succeeded at evaluating this block!
2577 DEBUG(dbgs() << "Successfully evaluated block.\n");
2580 // Did not know how to evaluate this!
2581 DEBUG(dbgs() << "Failed to evaluate block due to unhandled instruction."
2586 if (!CurInst->use_empty()) {
2587 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(InstResult))
2588 InstResult = ConstantFoldConstantExpression(CE, DL, TLI);
2590 setVal(CurInst, InstResult);
2593 // If we just processed an invoke, we finished evaluating the block.
2594 if (InvokeInst *II = dyn_cast<InvokeInst>(CurInst)) {
2595 NextBB = II->getNormalDest();
2596 DEBUG(dbgs() << "Found an invoke instruction. Finished Block.\n\n");
2600 // Advance program counter.
2605 /// EvaluateFunction - Evaluate a call to function F, returning true if
2606 /// successful, false if we can't evaluate it. ActualArgs contains the formal
2607 /// arguments for the function.
2608 bool Evaluator::EvaluateFunction(Function *F, Constant *&RetVal,
2609 const SmallVectorImpl<Constant*> &ActualArgs) {
2610 // Check to see if this function is already executing (recursion). If so,
2611 // bail out. TODO: we might want to accept limited recursion.
2612 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
2615 CallStack.push_back(F);
2617 // Initialize arguments to the incoming values specified.
2619 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
2621 setVal(AI, ActualArgs[ArgNo]);
2623 // ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
2624 // we can only evaluate any one basic block at most once. This set keeps
2625 // track of what we have executed so we can detect recursive cases etc.
2626 SmallPtrSet<BasicBlock*, 32> ExecutedBlocks;
2628 // CurBB - The current basic block we're evaluating.
2629 BasicBlock *CurBB = F->begin();
2631 BasicBlock::iterator CurInst = CurBB->begin();
2634 BasicBlock *NextBB = nullptr; // Initialized to avoid compiler warnings.
2635 DEBUG(dbgs() << "Trying to evaluate BB: " << *CurBB << "\n");
2637 if (!EvaluateBlock(CurInst, NextBB))
2641 // Successfully running until there's no next block means that we found
2642 // the return. Fill it the return value and pop the call stack.
2643 ReturnInst *RI = cast<ReturnInst>(CurBB->getTerminator());
2644 if (RI->getNumOperands())
2645 RetVal = getVal(RI->getOperand(0));
2646 CallStack.pop_back();
2650 // Okay, we succeeded in evaluating this control flow. See if we have
2651 // executed the new block before. If so, we have a looping function,
2652 // which we cannot evaluate in reasonable time.
2653 if (!ExecutedBlocks.insert(NextBB))
2654 return false; // looped!
2656 // Okay, we have never been in this block before. Check to see if there
2657 // are any PHI nodes. If so, evaluate them with information about where
2659 PHINode *PN = nullptr;
2660 for (CurInst = NextBB->begin();
2661 (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2662 setVal(PN, getVal(PN->getIncomingValueForBlock(CurBB)));
2664 // Advance to the next block.
2669 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2670 /// we can. Return true if we can, false otherwise.
2671 static bool EvaluateStaticConstructor(Function *F, const DataLayout *DL,
2672 const TargetLibraryInfo *TLI) {
2673 // Call the function.
2674 Evaluator Eval(DL, TLI);
2675 Constant *RetValDummy;
2676 bool EvalSuccess = Eval.EvaluateFunction(F, RetValDummy,
2677 SmallVector<Constant*, 0>());
2680 ++NumCtorsEvaluated;
2682 // We succeeded at evaluation: commit the result.
2683 DEBUG(dbgs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2684 << F->getName() << "' to " << Eval.getMutatedMemory().size()
2686 for (DenseMap<Constant*, Constant*>::const_iterator I =
2687 Eval.getMutatedMemory().begin(), E = Eval.getMutatedMemory().end();
2689 CommitValueTo(I->second, I->first);
2690 for (SmallPtrSet<GlobalVariable*, 8>::const_iterator I =
2691 Eval.getInvariants().begin(), E = Eval.getInvariants().end();
2693 (*I)->setConstant(true);
2699 static int compareNames(Constant *const *A, Constant *const *B) {
2700 return (*A)->getName().compare((*B)->getName());
2703 static void setUsedInitializer(GlobalVariable &V,
2704 SmallPtrSet<GlobalValue *, 8> Init) {
2706 V.eraseFromParent();
2710 // Type of pointer to the array of pointers.
2711 PointerType *Int8PtrTy = Type::getInt8PtrTy(V.getContext(), 0);
2713 SmallVector<llvm::Constant *, 8> UsedArray;
2714 for (SmallPtrSet<GlobalValue *, 8>::iterator I = Init.begin(), E = Init.end();
2717 = ConstantExpr::getPointerBitCastOrAddrSpaceCast(*I, Int8PtrTy);
2718 UsedArray.push_back(Cast);
2720 // Sort to get deterministic order.
2721 array_pod_sort(UsedArray.begin(), UsedArray.end(), compareNames);
2722 ArrayType *ATy = ArrayType::get(Int8PtrTy, UsedArray.size());
2724 Module *M = V.getParent();
2725 V.removeFromParent();
2726 GlobalVariable *NV =
2727 new GlobalVariable(*M, ATy, false, llvm::GlobalValue::AppendingLinkage,
2728 llvm::ConstantArray::get(ATy, UsedArray), "");
2730 NV->setSection("llvm.metadata");
2735 /// \brief An easy to access representation of llvm.used and llvm.compiler.used.
2737 SmallPtrSet<GlobalValue *, 8> Used;
2738 SmallPtrSet<GlobalValue *, 8> CompilerUsed;
2739 GlobalVariable *UsedV;
2740 GlobalVariable *CompilerUsedV;
2743 LLVMUsed(Module &M) {
2744 UsedV = collectUsedGlobalVariables(M, Used, false);
2745 CompilerUsedV = collectUsedGlobalVariables(M, CompilerUsed, true);
2747 typedef SmallPtrSet<GlobalValue *, 8>::iterator iterator;
2748 iterator usedBegin() { return Used.begin(); }
2749 iterator usedEnd() { return Used.end(); }
2750 iterator compilerUsedBegin() { return CompilerUsed.begin(); }
2751 iterator compilerUsedEnd() { return CompilerUsed.end(); }
2752 bool usedCount(GlobalValue *GV) const { return Used.count(GV); }
2753 bool compilerUsedCount(GlobalValue *GV) const {
2754 return CompilerUsed.count(GV);
2756 bool usedErase(GlobalValue *GV) { return Used.erase(GV); }
2757 bool compilerUsedErase(GlobalValue *GV) { return CompilerUsed.erase(GV); }
2758 bool usedInsert(GlobalValue *GV) { return Used.insert(GV); }
2759 bool compilerUsedInsert(GlobalValue *GV) { return CompilerUsed.insert(GV); }
2761 void syncVariablesAndSets() {
2763 setUsedInitializer(*UsedV, Used);
2765 setUsedInitializer(*CompilerUsedV, CompilerUsed);
2770 static bool hasUseOtherThanLLVMUsed(GlobalAlias &GA, const LLVMUsed &U) {
2771 if (GA.use_empty()) // No use at all.
2774 assert((!U.usedCount(&GA) || !U.compilerUsedCount(&GA)) &&
2775 "We should have removed the duplicated "
2776 "element from llvm.compiler.used");
2777 if (!GA.hasOneUse())
2778 // Strictly more than one use. So at least one is not in llvm.used and
2779 // llvm.compiler.used.
2782 // Exactly one use. Check if it is in llvm.used or llvm.compiler.used.
2783 return !U.usedCount(&GA) && !U.compilerUsedCount(&GA);
2786 static bool hasMoreThanOneUseOtherThanLLVMUsed(GlobalValue &V,
2787 const LLVMUsed &U) {
2789 assert((!U.usedCount(&V) || !U.compilerUsedCount(&V)) &&
2790 "We should have removed the duplicated "
2791 "element from llvm.compiler.used");
2792 if (U.usedCount(&V) || U.compilerUsedCount(&V))
2794 return V.hasNUsesOrMore(N);
2797 static bool mayHaveOtherReferences(GlobalAlias &GA, const LLVMUsed &U) {
2798 if (!GA.hasLocalLinkage())
2801 return U.usedCount(&GA) || U.compilerUsedCount(&GA);
2804 static bool hasUsesToReplace(GlobalAlias &GA, LLVMUsed &U, bool &RenameTarget) {
2805 RenameTarget = false;
2807 if (hasUseOtherThanLLVMUsed(GA, U))
2810 // If the alias is externally visible, we may still be able to simplify it.
2811 if (!mayHaveOtherReferences(GA, U))
2814 // If the aliasee has internal linkage, give it the name and linkage
2815 // of the alias, and delete the alias. This turns:
2816 // define internal ... @f(...)
2817 // @a = alias ... @f
2819 // define ... @a(...)
2820 Constant *Aliasee = GA.getAliasee();
2821 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2822 if (!Target->hasLocalLinkage())
2825 // Do not perform the transform if multiple aliases potentially target the
2826 // aliasee. This check also ensures that it is safe to replace the section
2827 // and other attributes of the aliasee with those of the alias.
2828 if (hasMoreThanOneUseOtherThanLLVMUsed(*Target, U))
2831 RenameTarget = true;
2835 bool GlobalOpt::OptimizeGlobalAliases(Module &M) {
2836 bool Changed = false;
2839 for (SmallPtrSet<GlobalValue *, 8>::iterator I = Used.usedBegin(),
2842 Used.compilerUsedErase(*I);
2844 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end();
2846 Module::alias_iterator J = I++;
2847 // Aliases without names cannot be referenced outside this module.
2848 if (!J->hasName() && !J->isDeclaration())
2849 J->setLinkage(GlobalValue::InternalLinkage);
2850 // If the aliasee may change at link time, nothing can be done - bail out.
2851 if (J->mayBeOverridden())
2854 Constant *Aliasee = J->getAliasee();
2855 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2856 Target->removeDeadConstantUsers();
2858 // Make all users of the alias use the aliasee instead.
2860 if (!hasUsesToReplace(*J, Used, RenameTarget))
2863 J->replaceAllUsesWith(Aliasee);
2864 ++NumAliasesResolved;
2868 // Give the aliasee the name, linkage and other attributes of the alias.
2869 Target->takeName(J);
2870 Target->setLinkage(J->getLinkage());
2871 Target->setVisibility(J->getVisibility());
2872 Target->setDLLStorageClass(J->getDLLStorageClass());
2874 if (Used.usedErase(J))
2875 Used.usedInsert(Target);
2877 if (Used.compilerUsedErase(J))
2878 Used.compilerUsedInsert(Target);
2879 } else if (mayHaveOtherReferences(*J, Used))
2882 // Delete the alias.
2883 M.getAliasList().erase(J);
2884 ++NumAliasesRemoved;
2888 Used.syncVariablesAndSets();
2893 static Function *FindCXAAtExit(Module &M, TargetLibraryInfo *TLI) {
2894 if (!TLI->has(LibFunc::cxa_atexit))
2897 Function *Fn = M.getFunction(TLI->getName(LibFunc::cxa_atexit));
2902 FunctionType *FTy = Fn->getFunctionType();
2904 // Checking that the function has the right return type, the right number of
2905 // parameters and that they all have pointer types should be enough.
2906 if (!FTy->getReturnType()->isIntegerTy() ||
2907 FTy->getNumParams() != 3 ||
2908 !FTy->getParamType(0)->isPointerTy() ||
2909 !FTy->getParamType(1)->isPointerTy() ||
2910 !FTy->getParamType(2)->isPointerTy())
2916 /// cxxDtorIsEmpty - Returns whether the given function is an empty C++
2917 /// destructor and can therefore be eliminated.
2918 /// Note that we assume that other optimization passes have already simplified
2919 /// the code so we only look for a function with a single basic block, where
2920 /// the only allowed instructions are 'ret', 'call' to an empty C++ dtor and
2921 /// other side-effect free instructions.
2922 static bool cxxDtorIsEmpty(const Function &Fn,
2923 SmallPtrSet<const Function *, 8> &CalledFunctions) {
2924 // FIXME: We could eliminate C++ destructors if they're readonly/readnone and
2925 // nounwind, but that doesn't seem worth doing.
2926 if (Fn.isDeclaration())
2929 if (++Fn.begin() != Fn.end())
2932 const BasicBlock &EntryBlock = Fn.getEntryBlock();
2933 for (BasicBlock::const_iterator I = EntryBlock.begin(), E = EntryBlock.end();
2935 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
2936 // Ignore debug intrinsics.
2937 if (isa<DbgInfoIntrinsic>(CI))
2940 const Function *CalledFn = CI->getCalledFunction();
2945 SmallPtrSet<const Function *, 8> NewCalledFunctions(CalledFunctions);
2947 // Don't treat recursive functions as empty.
2948 if (!NewCalledFunctions.insert(CalledFn))
2951 if (!cxxDtorIsEmpty(*CalledFn, NewCalledFunctions))
2953 } else if (isa<ReturnInst>(*I))
2954 return true; // We're done.
2955 else if (I->mayHaveSideEffects())
2956 return false; // Destructor with side effects, bail.
2962 bool GlobalOpt::OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) {
2963 /// Itanium C++ ABI p3.3.5:
2965 /// After constructing a global (or local static) object, that will require
2966 /// destruction on exit, a termination function is registered as follows:
2968 /// extern "C" int __cxa_atexit ( void (*f)(void *), void *p, void *d );
2970 /// This registration, e.g. __cxa_atexit(f,p,d), is intended to cause the
2971 /// call f(p) when DSO d is unloaded, before all such termination calls
2972 /// registered before this one. It returns zero if registration is
2973 /// successful, nonzero on failure.
2975 // This pass will look for calls to __cxa_atexit where the function is trivial
2977 bool Changed = false;
2979 for (auto I = CXAAtExitFn->user_begin(), E = CXAAtExitFn->user_end();
2981 // We're only interested in calls. Theoretically, we could handle invoke
2982 // instructions as well, but neither llvm-gcc nor clang generate invokes
2984 CallInst *CI = dyn_cast<CallInst>(*I++);
2989 dyn_cast<Function>(CI->getArgOperand(0)->stripPointerCasts());
2993 SmallPtrSet<const Function *, 8> CalledFunctions;
2994 if (!cxxDtorIsEmpty(*DtorFn, CalledFunctions))
2997 // Just remove the call.
2998 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
2999 CI->eraseFromParent();
3001 ++NumCXXDtorsRemoved;
3009 bool GlobalOpt::runOnModule(Module &M) {
3010 bool Changed = false;
3012 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
3013 DL = DLP ? &DLP->getDataLayout() : nullptr;
3014 TLI = &getAnalysis<TargetLibraryInfo>();
3016 bool LocalChange = true;
3017 while (LocalChange) {
3018 LocalChange = false;
3020 // Delete functions that are trivially dead, ccc -> fastcc
3021 LocalChange |= OptimizeFunctions(M);
3023 // Optimize global_ctors list.
3024 LocalChange |= optimizeGlobalCtorsList(M, [](void *C, Function *F) -> bool {
3025 GlobalOpt *self = static_cast<GlobalOpt *>(C);
3026 return EvaluateStaticConstructor(F, self->DL, self->TLI);
3029 // Optimize non-address-taken globals.
3030 LocalChange |= OptimizeGlobalVars(M);
3032 // Resolve aliases, when possible.
3033 LocalChange |= OptimizeGlobalAliases(M);
3035 // Try to remove trivial global destructors if they are not removed
3037 Function *CXAAtExitFn = FindCXAAtExit(M, TLI);
3039 LocalChange |= OptimizeEmptyGlobalCXXDtors(CXAAtExitFn);
3041 Changed |= LocalChange;
3044 // TODO: Move all global ctors functions to the end of the module for code