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/GlobalOpt.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/ADT/Twine.h"
23 #include "llvm/ADT/iterator_range.h"
24 #include "llvm/Analysis/BlockFrequencyInfo.h"
25 #include "llvm/Analysis/ConstantFolding.h"
26 #include "llvm/Analysis/MemoryBuiltins.h"
27 #include "llvm/Analysis/TargetLibraryInfo.h"
28 #include "llvm/Analysis/TargetTransformInfo.h"
29 #include "llvm/Transforms/Utils/Local.h"
30 #include "llvm/BinaryFormat/Dwarf.h"
31 #include "llvm/IR/Attributes.h"
32 #include "llvm/IR/BasicBlock.h"
33 #include "llvm/IR/CallSite.h"
34 #include "llvm/IR/CallingConv.h"
35 #include "llvm/IR/Constant.h"
36 #include "llvm/IR/Constants.h"
37 #include "llvm/IR/DataLayout.h"
38 #include "llvm/IR/DebugInfoMetadata.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/Dominators.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/GetElementPtrTypeIterator.h"
43 #include "llvm/IR/GlobalAlias.h"
44 #include "llvm/IR/GlobalValue.h"
45 #include "llvm/IR/GlobalVariable.h"
46 #include "llvm/IR/InstrTypes.h"
47 #include "llvm/IR/Instruction.h"
48 #include "llvm/IR/Instructions.h"
49 #include "llvm/IR/IntrinsicInst.h"
50 #include "llvm/IR/Module.h"
51 #include "llvm/IR/Operator.h"
52 #include "llvm/IR/Type.h"
53 #include "llvm/IR/Use.h"
54 #include "llvm/IR/User.h"
55 #include "llvm/IR/Value.h"
56 #include "llvm/IR/ValueHandle.h"
57 #include "llvm/Pass.h"
58 #include "llvm/Support/AtomicOrdering.h"
59 #include "llvm/Support/Casting.h"
60 #include "llvm/Support/CommandLine.h"
61 #include "llvm/Support/Debug.h"
62 #include "llvm/Support/ErrorHandling.h"
63 #include "llvm/Support/MathExtras.h"
64 #include "llvm/Support/raw_ostream.h"
65 #include "llvm/Transforms/IPO.h"
66 #include "llvm/Transforms/Utils/CtorUtils.h"
67 #include "llvm/Transforms/Utils/Evaluator.h"
68 #include "llvm/Transforms/Utils/GlobalStatus.h"
76 #define DEBUG_TYPE "globalopt"
78 STATISTIC(NumMarked , "Number of globals marked constant");
79 STATISTIC(NumUnnamed , "Number of globals marked unnamed_addr");
80 STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
81 STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
82 STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
83 STATISTIC(NumDeleted , "Number of globals deleted");
84 STATISTIC(NumGlobUses , "Number of global uses devirtualized");
85 STATISTIC(NumLocalized , "Number of globals localized");
86 STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans");
87 STATISTIC(NumFastCallFns , "Number of functions converted to fastcc");
88 STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
89 STATISTIC(NumNestRemoved , "Number of nest attributes removed");
90 STATISTIC(NumAliasesResolved, "Number of global aliases resolved");
91 STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated");
92 STATISTIC(NumCXXDtorsRemoved, "Number of global C++ destructors removed");
93 STATISTIC(NumInternalFunc, "Number of internal functions");
94 STATISTIC(NumColdCC, "Number of functions marked coldcc");
97 EnableColdCCStressTest("enable-coldcc-stress-test",
98 cl::desc("Enable stress test of coldcc by adding "
99 "calling conv to all internal functions."),
100 cl::init(false), cl::Hidden);
102 static cl::opt<int> ColdCCRelFreq(
103 "coldcc-rel-freq", cl::Hidden, cl::init(2), cl::ZeroOrMore,
105 "Maximum block frequency, expressed as a percentage of caller's "
106 "entry frequency, for a call site to be considered cold for enabling"
109 /// Is this global variable possibly used by a leak checker as a root? If so,
110 /// we might not really want to eliminate the stores to it.
111 static bool isLeakCheckerRoot(GlobalVariable *GV) {
112 // A global variable is a root if it is a pointer, or could plausibly contain
113 // a pointer. There are two challenges; one is that we could have a struct
114 // the has an inner member which is a pointer. We recurse through the type to
115 // detect these (up to a point). The other is that we may actually be a union
116 // of a pointer and another type, and so our LLVM type is an integer which
117 // gets converted into a pointer, or our type is an [i8 x #] with a pointer
118 // potentially contained here.
120 if (GV->hasPrivateLinkage())
123 SmallVector<Type *, 4> Types;
124 Types.push_back(GV->getValueType());
128 Type *Ty = Types.pop_back_val();
129 switch (Ty->getTypeID()) {
131 case Type::PointerTyID: return true;
132 case Type::ArrayTyID:
133 case Type::VectorTyID: {
134 SequentialType *STy = cast<SequentialType>(Ty);
135 Types.push_back(STy->getElementType());
138 case Type::StructTyID: {
139 StructType *STy = cast<StructType>(Ty);
140 if (STy->isOpaque()) return true;
141 for (StructType::element_iterator I = STy->element_begin(),
142 E = STy->element_end(); I != E; ++I) {
144 if (isa<PointerType>(InnerTy)) return true;
145 if (isa<CompositeType>(InnerTy))
146 Types.push_back(InnerTy);
151 if (--Limit == 0) return true;
152 } while (!Types.empty());
156 /// Given a value that is stored to a global but never read, determine whether
157 /// it's safe to remove the store and the chain of computation that feeds the
159 static bool IsSafeComputationToRemove(Value *V, const TargetLibraryInfo *TLI) {
161 if (isa<Constant>(V))
165 if (isa<LoadInst>(V) || isa<InvokeInst>(V) || isa<Argument>(V) ||
168 if (isAllocationFn(V, TLI))
171 Instruction *I = cast<Instruction>(V);
172 if (I->mayHaveSideEffects())
174 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
175 if (!GEP->hasAllConstantIndices())
177 } else if (I->getNumOperands() != 1) {
181 V = I->getOperand(0);
185 /// This GV is a pointer root. Loop over all users of the global and clean up
186 /// any that obviously don't assign the global a value that isn't dynamically
188 static bool CleanupPointerRootUsers(GlobalVariable *GV,
189 const TargetLibraryInfo *TLI) {
190 // A brief explanation of leak checkers. The goal is to find bugs where
191 // pointers are forgotten, causing an accumulating growth in memory
192 // usage over time. The common strategy for leak checkers is to whitelist the
193 // memory pointed to by globals at exit. This is popular because it also
194 // solves another problem where the main thread of a C++ program may shut down
195 // before other threads that are still expecting to use those globals. To
196 // handle that case, we expect the program may create a singleton and never
199 bool Changed = false;
201 // If Dead[n].first is the only use of a malloc result, we can delete its
202 // chain of computation and the store to the global in Dead[n].second.
203 SmallVector<std::pair<Instruction *, Instruction *>, 32> Dead;
205 // Constants can't be pointers to dynamically allocated memory.
206 for (Value::user_iterator UI = GV->user_begin(), E = GV->user_end();
209 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
210 Value *V = SI->getValueOperand();
211 if (isa<Constant>(V)) {
213 SI->eraseFromParent();
214 } else if (Instruction *I = dyn_cast<Instruction>(V)) {
216 Dead.push_back(std::make_pair(I, SI));
218 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(U)) {
219 if (isa<Constant>(MSI->getValue())) {
221 MSI->eraseFromParent();
222 } else if (Instruction *I = dyn_cast<Instruction>(MSI->getValue())) {
224 Dead.push_back(std::make_pair(I, MSI));
226 } else if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(U)) {
227 GlobalVariable *MemSrc = dyn_cast<GlobalVariable>(MTI->getSource());
228 if (MemSrc && MemSrc->isConstant()) {
230 MTI->eraseFromParent();
231 } else if (Instruction *I = dyn_cast<Instruction>(MemSrc)) {
233 Dead.push_back(std::make_pair(I, MTI));
235 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
236 if (CE->use_empty()) {
237 CE->destroyConstant();
240 } else if (Constant *C = dyn_cast<Constant>(U)) {
241 if (isSafeToDestroyConstant(C)) {
242 C->destroyConstant();
243 // This could have invalidated UI, start over from scratch.
245 CleanupPointerRootUsers(GV, TLI);
251 for (int i = 0, e = Dead.size(); i != e; ++i) {
252 if (IsSafeComputationToRemove(Dead[i].first, TLI)) {
253 Dead[i].second->eraseFromParent();
254 Instruction *I = Dead[i].first;
256 if (isAllocationFn(I, TLI))
258 Instruction *J = dyn_cast<Instruction>(I->getOperand(0));
261 I->eraseFromParent();
264 I->eraseFromParent();
271 /// We just marked GV constant. Loop over all users of the global, cleaning up
272 /// the obvious ones. This is largely just a quick scan over the use list to
273 /// clean up the easy and obvious cruft. This returns true if it made a change.
274 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init,
275 const DataLayout &DL,
276 TargetLibraryInfo *TLI) {
277 bool Changed = false;
278 // Note that we need to use a weak value handle for the worklist items. When
279 // we delete a constant array, we may also be holding pointer to one of its
280 // elements (or an element of one of its elements if we're dealing with an
281 // array of arrays) in the worklist.
282 SmallVector<WeakTrackingVH, 8> WorkList(V->user_begin(), V->user_end());
283 while (!WorkList.empty()) {
284 Value *UV = WorkList.pop_back_val();
288 User *U = cast<User>(UV);
290 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
292 // Replace the load with the initializer.
293 LI->replaceAllUsesWith(Init);
294 LI->eraseFromParent();
297 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
298 // Store must be unreachable or storing Init into the global.
299 SI->eraseFromParent();
301 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
302 if (CE->getOpcode() == Instruction::GetElementPtr) {
303 Constant *SubInit = nullptr;
305 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
306 Changed |= CleanupConstantGlobalUsers(CE, SubInit, DL, TLI);
307 } else if ((CE->getOpcode() == Instruction::BitCast &&
308 CE->getType()->isPointerTy()) ||
309 CE->getOpcode() == Instruction::AddrSpaceCast) {
310 // Pointer cast, delete any stores and memsets to the global.
311 Changed |= CleanupConstantGlobalUsers(CE, nullptr, DL, TLI);
314 if (CE->use_empty()) {
315 CE->destroyConstant();
318 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
319 // Do not transform "gepinst (gep constexpr (GV))" here, because forming
320 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
321 // and will invalidate our notion of what Init is.
322 Constant *SubInit = nullptr;
323 if (!isa<ConstantExpr>(GEP->getOperand(0))) {
324 ConstantExpr *CE = dyn_cast_or_null<ConstantExpr>(
325 ConstantFoldInstruction(GEP, DL, TLI));
326 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
327 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
329 // If the initializer is an all-null value and we have an inbounds GEP,
330 // we already know what the result of any load from that GEP is.
331 // TODO: Handle splats.
332 if (Init && isa<ConstantAggregateZero>(Init) && GEP->isInBounds())
333 SubInit = Constant::getNullValue(GEP->getResultElementType());
335 Changed |= CleanupConstantGlobalUsers(GEP, SubInit, DL, TLI);
337 if (GEP->use_empty()) {
338 GEP->eraseFromParent();
341 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
342 if (MI->getRawDest() == V) {
343 MI->eraseFromParent();
347 } else if (Constant *C = dyn_cast<Constant>(U)) {
348 // If we have a chain of dead constantexprs or other things dangling from
349 // us, and if they are all dead, nuke them without remorse.
350 if (isSafeToDestroyConstant(C)) {
351 C->destroyConstant();
352 CleanupConstantGlobalUsers(V, Init, DL, TLI);
360 static bool isSafeSROAElementUse(Value *V);
362 /// Return true if the specified GEP is a safe user of a derived
363 /// expression from a global that we want to SROA.
364 static bool isSafeSROAGEP(User *U) {
365 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
366 // don't like < 3 operand CE's, and we don't like non-constant integer
367 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some
369 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
370 !cast<Constant>(U->getOperand(1))->isNullValue())
373 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
374 ++GEPI; // Skip over the pointer index.
376 // For all other level we require that the indices are constant and inrange.
377 // In particular, consider: A[0][i]. We cannot know that the user isn't doing
378 // invalid things like allowing i to index an out-of-range subscript that
379 // accesses A[1]. This can also happen between different members of a struct
381 for (; GEPI != E; ++GEPI) {
385 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
386 if (!IdxVal || (GEPI.isBoundedSequential() &&
387 IdxVal->getZExtValue() >= GEPI.getSequentialNumElements()))
391 return llvm::all_of(U->users(),
392 [](User *UU) { return isSafeSROAElementUse(UU); });
395 /// Return true if the specified instruction is a safe user of a derived
396 /// expression from a global that we want to SROA.
397 static bool isSafeSROAElementUse(Value *V) {
398 // We might have a dead and dangling constant hanging off of here.
399 if (Constant *C = dyn_cast<Constant>(V))
400 return isSafeToDestroyConstant(C);
402 Instruction *I = dyn_cast<Instruction>(V);
403 if (!I) return false;
406 if (isa<LoadInst>(I)) return true;
408 // Stores *to* the pointer are ok.
409 if (StoreInst *SI = dyn_cast<StoreInst>(I))
410 return SI->getOperand(0) != V;
412 // Otherwise, it must be a GEP. Check it and its users are safe to SRA.
413 return isa<GetElementPtrInst>(I) && isSafeSROAGEP(I);
416 /// Look at all uses of the global and decide whether it is safe for us to
417 /// perform this transformation.
418 static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
419 for (User *U : GV->users()) {
420 // The user of the global must be a GEP Inst or a ConstantExpr GEP.
421 if (!isa<GetElementPtrInst>(U) &&
422 (!isa<ConstantExpr>(U) ||
423 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
426 // Check the gep and it's users are safe to SRA
427 if (!isSafeSROAGEP(U))
434 /// Copy over the debug info for a variable to its SRA replacements.
435 static void transferSRADebugInfo(GlobalVariable *GV, GlobalVariable *NGV,
436 uint64_t FragmentOffsetInBits,
437 uint64_t FragmentSizeInBits,
438 unsigned NumElements) {
439 SmallVector<DIGlobalVariableExpression *, 1> GVs;
440 GV->getDebugInfo(GVs);
441 for (auto *GVE : GVs) {
442 DIVariable *Var = GVE->getVariable();
443 DIExpression *Expr = GVE->getExpression();
444 if (NumElements > 1) {
445 if (auto E = DIExpression::createFragmentExpression(
446 Expr, FragmentOffsetInBits, FragmentSizeInBits))
451 auto *NGVE = DIGlobalVariableExpression::get(GVE->getContext(), Var, Expr);
452 NGV->addDebugInfo(NGVE);
456 /// Perform scalar replacement of aggregates on the specified global variable.
457 /// This opens the door for other optimizations by exposing the behavior of the
458 /// program in a more fine-grained way. We have determined that this
459 /// transformation is safe already. We return the first global variable we
460 /// insert so that the caller can reprocess it.
461 static GlobalVariable *SRAGlobal(GlobalVariable *GV, const DataLayout &DL) {
462 // Make sure this global only has simple uses that we can SRA.
463 if (!GlobalUsersSafeToSRA(GV))
466 assert(GV->hasLocalLinkage());
467 Constant *Init = GV->getInitializer();
468 Type *Ty = Init->getType();
470 std::vector<GlobalVariable *> NewGlobals;
471 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
473 // Get the alignment of the global, either explicit or target-specific.
474 unsigned StartAlignment = GV->getAlignment();
475 if (StartAlignment == 0)
476 StartAlignment = DL.getABITypeAlignment(GV->getType());
478 if (StructType *STy = dyn_cast<StructType>(Ty)) {
479 unsigned NumElements = STy->getNumElements();
480 NewGlobals.reserve(NumElements);
481 const StructLayout &Layout = *DL.getStructLayout(STy);
482 for (unsigned i = 0, e = NumElements; i != e; ++i) {
483 Constant *In = Init->getAggregateElement(i);
484 assert(In && "Couldn't get element of initializer?");
485 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
486 GlobalVariable::InternalLinkage,
487 In, GV->getName()+"."+Twine(i),
488 GV->getThreadLocalMode(),
489 GV->getType()->getAddressSpace());
490 NGV->setExternallyInitialized(GV->isExternallyInitialized());
491 NGV->copyAttributesFrom(GV);
492 Globals.push_back(NGV);
493 NewGlobals.push_back(NGV);
495 // Calculate the known alignment of the field. If the original aggregate
496 // had 256 byte alignment for example, something might depend on that:
497 // propagate info to each field.
498 uint64_t FieldOffset = Layout.getElementOffset(i);
499 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset);
500 if (NewAlign > DL.getABITypeAlignment(STy->getElementType(i)))
501 NGV->setAlignment(NewAlign);
503 // Copy over the debug info for the variable.
504 uint64_t Size = DL.getTypeAllocSizeInBits(NGV->getValueType());
505 uint64_t FragmentOffsetInBits = Layout.getElementOffsetInBits(i);
506 transferSRADebugInfo(GV, NGV, FragmentOffsetInBits, Size, NumElements);
508 } else if (SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
509 unsigned NumElements = STy->getNumElements();
510 if (NumElements > 16 && GV->hasNUsesOrMore(16))
511 return nullptr; // It's not worth it.
512 NewGlobals.reserve(NumElements);
513 auto ElTy = STy->getElementType();
514 uint64_t EltSize = DL.getTypeAllocSize(ElTy);
515 unsigned EltAlign = DL.getABITypeAlignment(ElTy);
516 uint64_t FragmentSizeInBits = DL.getTypeAllocSizeInBits(ElTy);
517 for (unsigned i = 0, e = NumElements; i != e; ++i) {
518 Constant *In = Init->getAggregateElement(i);
519 assert(In && "Couldn't get element of initializer?");
521 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
522 GlobalVariable::InternalLinkage,
523 In, GV->getName()+"."+Twine(i),
524 GV->getThreadLocalMode(),
525 GV->getType()->getAddressSpace());
526 NGV->setExternallyInitialized(GV->isExternallyInitialized());
527 NGV->copyAttributesFrom(GV);
528 Globals.push_back(NGV);
529 NewGlobals.push_back(NGV);
531 // Calculate the known alignment of the field. If the original aggregate
532 // had 256 byte alignment for example, something might depend on that:
533 // propagate info to each field.
534 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
535 if (NewAlign > EltAlign)
536 NGV->setAlignment(NewAlign);
537 transferSRADebugInfo(GV, NGV, FragmentSizeInBits * i, FragmentSizeInBits,
542 if (NewGlobals.empty())
545 LLVM_DEBUG(dbgs() << "PERFORMING GLOBAL SRA ON: " << *GV << "\n");
547 Constant *NullInt =Constant::getNullValue(Type::getInt32Ty(GV->getContext()));
549 // Loop over all of the uses of the global, replacing the constantexpr geps,
550 // with smaller constantexpr geps or direct references.
551 while (!GV->use_empty()) {
552 User *GEP = GV->user_back();
553 assert(((isa<ConstantExpr>(GEP) &&
554 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
555 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
557 // Ignore the 1th operand, which has to be zero or else the program is quite
558 // broken (undefined). Get the 2nd operand, which is the structure or array
560 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
561 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
563 Value *NewPtr = NewGlobals[Val];
564 Type *NewTy = NewGlobals[Val]->getValueType();
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));
574 ConstantExpr::getGetElementPtr(NewTy, cast<Constant>(NewPtr), Idxs);
576 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
577 SmallVector<Value*, 8> Idxs;
578 Idxs.push_back(NullInt);
579 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
580 Idxs.push_back(GEPI->getOperand(i));
581 NewPtr = GetElementPtrInst::Create(
582 NewTy, NewPtr, Idxs, GEPI->getName() + "." + Twine(Val), GEPI);
585 GEP->replaceAllUsesWith(NewPtr);
587 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
588 GEPI->eraseFromParent();
590 cast<ConstantExpr>(GEP)->destroyConstant();
593 // Delete the old global, now that it is dead.
597 // Loop over the new globals array deleting any globals that are obviously
598 // dead. This can arise due to scalarization of a structure or an array that
599 // has elements that are dead.
600 unsigned FirstGlobal = 0;
601 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
602 if (NewGlobals[i]->use_empty()) {
603 Globals.erase(NewGlobals[i]);
604 if (FirstGlobal == i) ++FirstGlobal;
607 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : nullptr;
610 /// Return true if all users of the specified value will trap if the value is
611 /// dynamically null. PHIs keeps track of any phi nodes we've seen to avoid
612 /// reprocessing them.
613 static bool AllUsesOfValueWillTrapIfNull(const Value *V,
614 SmallPtrSetImpl<const PHINode*> &PHIs) {
615 for (const User *U : V->users()) {
616 if (const Instruction *I = dyn_cast<Instruction>(U)) {
617 // If null pointer is considered valid, then all uses are non-trapping.
618 // Non address-space 0 globals have already been pruned by the caller.
619 if (NullPointerIsDefined(I->getFunction()))
622 if (isa<LoadInst>(U)) {
624 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
625 if (SI->getOperand(0) == V) {
626 //cerr << "NONTRAPPING USE: " << *U;
627 return false; // Storing the value.
629 } else if (const CallInst *CI = dyn_cast<CallInst>(U)) {
630 if (CI->getCalledValue() != V) {
631 //cerr << "NONTRAPPING USE: " << *U;
632 return false; // Not calling the ptr
634 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(U)) {
635 if (II->getCalledValue() != V) {
636 //cerr << "NONTRAPPING USE: " << *U;
637 return false; // Not calling the ptr
639 } else if (const BitCastInst *CI = dyn_cast<BitCastInst>(U)) {
640 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
641 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
642 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
643 } else if (const PHINode *PN = dyn_cast<PHINode>(U)) {
644 // If we've already seen this phi node, ignore it, it has already been
646 if (PHIs.insert(PN).second && !AllUsesOfValueWillTrapIfNull(PN, PHIs))
648 } else if (isa<ICmpInst>(U) &&
649 isa<ConstantPointerNull>(U->getOperand(1))) {
650 // Ignore icmp X, null
652 //cerr << "NONTRAPPING USE: " << *U;
659 /// Return true if all uses of any loads from GV will trap if the loaded value
660 /// is null. Note that this also permits comparisons of the loaded value
661 /// against null, as a special case.
662 static bool AllUsesOfLoadedValueWillTrapIfNull(const GlobalVariable *GV) {
663 for (const User *U : GV->users())
664 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
665 SmallPtrSet<const PHINode*, 8> PHIs;
666 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
668 } else if (isa<StoreInst>(U)) {
669 // Ignore stores to the global.
671 // We don't know or understand this user, bail out.
672 //cerr << "UNKNOWN USER OF GLOBAL!: " << *U;
678 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
679 bool Changed = false;
680 for (auto UI = V->user_begin(), E = V->user_end(); UI != E; ) {
681 Instruction *I = cast<Instruction>(*UI++);
682 // Uses are non-trapping if null pointer is considered valid.
683 // Non address-space 0 globals are already pruned by the caller.
684 if (NullPointerIsDefined(I->getFunction()))
686 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
687 LI->setOperand(0, NewV);
689 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
690 if (SI->getOperand(1) == V) {
691 SI->setOperand(1, NewV);
694 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
696 if (CS.getCalledValue() == V) {
697 // Calling through the pointer! Turn into a direct call, but be careful
698 // that the pointer is not also being passed as an argument.
699 CS.setCalledFunction(NewV);
701 bool PassedAsArg = false;
702 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
703 if (CS.getArgument(i) == V) {
705 CS.setArgument(i, NewV);
709 // Being passed as an argument also. Be careful to not invalidate UI!
710 UI = V->user_begin();
713 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
714 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
715 ConstantExpr::getCast(CI->getOpcode(),
716 NewV, CI->getType()));
717 if (CI->use_empty()) {
719 CI->eraseFromParent();
721 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
722 // Should handle GEP here.
723 SmallVector<Constant*, 8> Idxs;
724 Idxs.reserve(GEPI->getNumOperands()-1);
725 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
727 if (Constant *C = dyn_cast<Constant>(*i))
731 if (Idxs.size() == GEPI->getNumOperands()-1)
732 Changed |= OptimizeAwayTrappingUsesOfValue(
733 GEPI, ConstantExpr::getGetElementPtr(nullptr, NewV, Idxs));
734 if (GEPI->use_empty()) {
736 GEPI->eraseFromParent();
744 /// The specified global has only one non-null value stored into it. If there
745 /// are uses of the loaded value that would trap if the loaded value is
746 /// dynamically null, then we know that they cannot be reachable with a null
747 /// optimize away the load.
748 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV,
749 const DataLayout &DL,
750 TargetLibraryInfo *TLI) {
751 bool Changed = false;
753 // Keep track of whether we are able to remove all the uses of the global
754 // other than the store that defines it.
755 bool AllNonStoreUsesGone = true;
757 // Replace all uses of loads with uses of uses of the stored value.
758 for (Value::user_iterator GUI = GV->user_begin(), E = GV->user_end(); GUI != E;){
759 User *GlobalUser = *GUI++;
760 if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) {
761 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
762 // If we were able to delete all uses of the loads
763 if (LI->use_empty()) {
764 LI->eraseFromParent();
767 AllNonStoreUsesGone = false;
769 } else if (isa<StoreInst>(GlobalUser)) {
770 // Ignore the store that stores "LV" to the global.
771 assert(GlobalUser->getOperand(1) == GV &&
772 "Must be storing *to* the global");
774 AllNonStoreUsesGone = false;
776 // If we get here we could have other crazy uses that are transitively
778 assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) ||
779 isa<ConstantExpr>(GlobalUser) || isa<CmpInst>(GlobalUser) ||
780 isa<BitCastInst>(GlobalUser) ||
781 isa<GetElementPtrInst>(GlobalUser)) &&
782 "Only expect load and stores!");
787 LLVM_DEBUG(dbgs() << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV
792 // If we nuked all of the loads, then none of the stores are needed either,
793 // nor is the global.
794 if (AllNonStoreUsesGone) {
795 if (isLeakCheckerRoot(GV)) {
796 Changed |= CleanupPointerRootUsers(GV, TLI);
799 CleanupConstantGlobalUsers(GV, nullptr, DL, TLI);
801 if (GV->use_empty()) {
802 LLVM_DEBUG(dbgs() << " *** GLOBAL NOW DEAD!\n");
804 GV->eraseFromParent();
811 /// Walk the use list of V, constant folding all of the instructions that are
813 static void ConstantPropUsersOf(Value *V, const DataLayout &DL,
814 TargetLibraryInfo *TLI) {
815 for (Value::user_iterator UI = V->user_begin(), E = V->user_end(); UI != E; )
816 if (Instruction *I = dyn_cast<Instruction>(*UI++))
817 if (Constant *NewC = ConstantFoldInstruction(I, DL, TLI)) {
818 I->replaceAllUsesWith(NewC);
820 // Advance UI to the next non-I use to avoid invalidating it!
821 // Instructions could multiply use V.
822 while (UI != E && *UI == I)
824 if (isInstructionTriviallyDead(I, TLI))
825 I->eraseFromParent();
829 /// This function takes the specified global variable, and transforms the
830 /// program as if it always contained the result of the specified malloc.
831 /// Because it is always the result of the specified malloc, there is no reason
832 /// to actually DO the malloc. Instead, turn the malloc into a global, and any
833 /// loads of GV as uses of the new global.
834 static GlobalVariable *
835 OptimizeGlobalAddressOfMalloc(GlobalVariable *GV, CallInst *CI, Type *AllocTy,
836 ConstantInt *NElements, const DataLayout &DL,
837 TargetLibraryInfo *TLI) {
838 LLVM_DEBUG(errs() << "PROMOTING GLOBAL: " << *GV << " CALL = " << *CI
842 if (NElements->getZExtValue() == 1)
843 GlobalType = AllocTy;
845 // If we have an array allocation, the global variable is of an array.
846 GlobalType = ArrayType::get(AllocTy, NElements->getZExtValue());
848 // Create the new global variable. The contents of the malloc'd memory is
849 // undefined, so initialize with an undef value.
850 GlobalVariable *NewGV = new GlobalVariable(
851 *GV->getParent(), GlobalType, false, GlobalValue::InternalLinkage,
852 UndefValue::get(GlobalType), GV->getName() + ".body", nullptr,
853 GV->getThreadLocalMode());
855 // If there are bitcast users of the malloc (which is typical, usually we have
856 // a malloc + bitcast) then replace them with uses of the new global. Update
857 // other users to use the global as well.
858 BitCastInst *TheBC = nullptr;
859 while (!CI->use_empty()) {
860 Instruction *User = cast<Instruction>(CI->user_back());
861 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
862 if (BCI->getType() == NewGV->getType()) {
863 BCI->replaceAllUsesWith(NewGV);
864 BCI->eraseFromParent();
866 BCI->setOperand(0, NewGV);
870 TheBC = new BitCastInst(NewGV, CI->getType(), "newgv", CI);
871 User->replaceUsesOfWith(CI, TheBC);
875 Constant *RepValue = NewGV;
876 if (NewGV->getType() != GV->getValueType())
877 RepValue = ConstantExpr::getBitCast(RepValue, GV->getValueType());
879 // If there is a comparison against null, we will insert a global bool to
880 // keep track of whether the global was initialized yet or not.
881 GlobalVariable *InitBool =
882 new GlobalVariable(Type::getInt1Ty(GV->getContext()), false,
883 GlobalValue::InternalLinkage,
884 ConstantInt::getFalse(GV->getContext()),
885 GV->getName()+".init", GV->getThreadLocalMode());
886 bool InitBoolUsed = false;
888 // Loop over all uses of GV, processing them in turn.
889 while (!GV->use_empty()) {
890 if (StoreInst *SI = dyn_cast<StoreInst>(GV->user_back())) {
891 // The global is initialized when the store to it occurs.
892 new StoreInst(ConstantInt::getTrue(GV->getContext()), InitBool, false, 0,
893 SI->getOrdering(), SI->getSyncScopeID(), SI);
894 SI->eraseFromParent();
898 LoadInst *LI = cast<LoadInst>(GV->user_back());
899 while (!LI->use_empty()) {
900 Use &LoadUse = *LI->use_begin();
901 ICmpInst *ICI = dyn_cast<ICmpInst>(LoadUse.getUser());
907 // Replace the cmp X, 0 with a use of the bool value.
908 // Sink the load to where the compare was, if atomic rules allow us to.
909 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", false, 0,
910 LI->getOrdering(), LI->getSyncScopeID(),
911 LI->isUnordered() ? (Instruction*)ICI : LI);
913 switch (ICI->getPredicate()) {
914 default: llvm_unreachable("Unknown ICmp Predicate!");
915 case ICmpInst::ICMP_ULT:
916 case ICmpInst::ICMP_SLT: // X < null -> always false
917 LV = ConstantInt::getFalse(GV->getContext());
919 case ICmpInst::ICMP_ULE:
920 case ICmpInst::ICMP_SLE:
921 case ICmpInst::ICMP_EQ:
922 LV = BinaryOperator::CreateNot(LV, "notinit", ICI);
924 case ICmpInst::ICMP_NE:
925 case ICmpInst::ICMP_UGE:
926 case ICmpInst::ICMP_SGE:
927 case ICmpInst::ICMP_UGT:
928 case ICmpInst::ICMP_SGT:
931 ICI->replaceAllUsesWith(LV);
932 ICI->eraseFromParent();
934 LI->eraseFromParent();
937 // If the initialization boolean was used, insert it, otherwise delete it.
939 while (!InitBool->use_empty()) // Delete initializations
940 cast<StoreInst>(InitBool->user_back())->eraseFromParent();
943 GV->getParent()->getGlobalList().insert(GV->getIterator(), InitBool);
945 // Now the GV is dead, nuke it and the malloc..
946 GV->eraseFromParent();
947 CI->eraseFromParent();
949 // To further other optimizations, loop over all users of NewGV and try to
950 // constant prop them. This will promote GEP instructions with constant
951 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
952 ConstantPropUsersOf(NewGV, DL, TLI);
953 if (RepValue != NewGV)
954 ConstantPropUsersOf(RepValue, DL, TLI);
959 /// Scan the use-list of V checking to make sure that there are no complex uses
960 /// of V. We permit simple things like dereferencing the pointer, but not
961 /// storing through the address, unless it is to the specified global.
962 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(const Instruction *V,
963 const GlobalVariable *GV,
964 SmallPtrSetImpl<const PHINode*> &PHIs) {
965 for (const User *U : V->users()) {
966 const Instruction *Inst = cast<Instruction>(U);
968 if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) {
969 continue; // Fine, ignore.
972 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
973 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
974 return false; // Storing the pointer itself... bad.
975 continue; // Otherwise, storing through it, or storing into GV... fine.
978 // Must index into the array and into the struct.
979 if (isa<GetElementPtrInst>(Inst) && Inst->getNumOperands() >= 3) {
980 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs))
985 if (const PHINode *PN = dyn_cast<PHINode>(Inst)) {
986 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI
988 if (PHIs.insert(PN).second)
989 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
994 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) {
995 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs))
1005 /// The Alloc pointer is stored into GV somewhere. Transform all uses of the
1006 /// allocation into loads from the global and uses of the resultant pointer.
1007 /// Further, delete the store into GV. This assumes that these value pass the
1008 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
1009 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
1010 GlobalVariable *GV) {
1011 while (!Alloc->use_empty()) {
1012 Instruction *U = cast<Instruction>(*Alloc->user_begin());
1013 Instruction *InsertPt = U;
1014 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1015 // If this is the store of the allocation into the global, remove it.
1016 if (SI->getOperand(1) == GV) {
1017 SI->eraseFromParent();
1020 } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
1021 // Insert the load in the corresponding predecessor, not right before the
1023 InsertPt = PN->getIncomingBlock(*Alloc->use_begin())->getTerminator();
1024 } else if (isa<BitCastInst>(U)) {
1025 // Must be bitcast between the malloc and store to initialize the global.
1026 ReplaceUsesOfMallocWithGlobal(U, GV);
1027 U->eraseFromParent();
1029 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
1030 // If this is a "GEP bitcast" and the user is a store to the global, then
1031 // just process it as a bitcast.
1032 if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse())
1033 if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->user_back()))
1034 if (SI->getOperand(1) == GV) {
1035 // Must be bitcast GEP between the malloc and store to initialize
1037 ReplaceUsesOfMallocWithGlobal(GEPI, GV);
1038 GEPI->eraseFromParent();
1043 // Insert a load from the global, and use it instead of the malloc.
1044 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
1045 U->replaceUsesOfWith(Alloc, NL);
1049 /// Verify that all uses of V (a load, or a phi of a load) are simple enough to
1050 /// perform heap SRA on. This permits GEP's that index through the array and
1051 /// struct field, icmps of null, and PHIs.
1052 static bool LoadUsesSimpleEnoughForHeapSRA(const Value *V,
1053 SmallPtrSetImpl<const PHINode*> &LoadUsingPHIs,
1054 SmallPtrSetImpl<const PHINode*> &LoadUsingPHIsPerLoad) {
1055 // We permit two users of the load: setcc comparing against the null
1056 // pointer, and a getelementptr of a specific form.
1057 for (const User *U : V->users()) {
1058 const Instruction *UI = cast<Instruction>(U);
1060 // Comparison against null is ok.
1061 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UI)) {
1062 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1067 // getelementptr is also ok, but only a simple form.
1068 if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(UI)) {
1069 // Must index into the array and into the struct.
1070 if (GEPI->getNumOperands() < 3)
1073 // Otherwise the GEP is ok.
1077 if (const PHINode *PN = dyn_cast<PHINode>(UI)) {
1078 if (!LoadUsingPHIsPerLoad.insert(PN).second)
1079 // This means some phi nodes are dependent on each other.
1080 // Avoid infinite looping!
1082 if (!LoadUsingPHIs.insert(PN).second)
1083 // If we have already analyzed this PHI, then it is safe.
1086 // Make sure all uses of the PHI are simple enough to transform.
1087 if (!LoadUsesSimpleEnoughForHeapSRA(PN,
1088 LoadUsingPHIs, LoadUsingPHIsPerLoad))
1094 // Otherwise we don't know what this is, not ok.
1101 /// If all users of values loaded from GV are simple enough to perform HeapSRA,
1103 static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(const GlobalVariable *GV,
1104 Instruction *StoredVal) {
1105 SmallPtrSet<const PHINode*, 32> LoadUsingPHIs;
1106 SmallPtrSet<const PHINode*, 32> LoadUsingPHIsPerLoad;
1107 for (const User *U : GV->users())
1108 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
1109 if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs,
1110 LoadUsingPHIsPerLoad))
1112 LoadUsingPHIsPerLoad.clear();
1115 // If we reach here, we know that all uses of the loads and transitive uses
1116 // (through PHI nodes) are simple enough to transform. However, we don't know
1117 // that all inputs the to the PHI nodes are in the same equivalence sets.
1118 // Check to verify that all operands of the PHIs are either PHIS that can be
1119 // transformed, loads from GV, or MI itself.
1120 for (const PHINode *PN : LoadUsingPHIs) {
1121 for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) {
1122 Value *InVal = PN->getIncomingValue(op);
1124 // PHI of the stored value itself is ok.
1125 if (InVal == StoredVal) continue;
1127 if (const PHINode *InPN = dyn_cast<PHINode>(InVal)) {
1128 // One of the PHIs in our set is (optimistically) ok.
1129 if (LoadUsingPHIs.count(InPN))
1134 // Load from GV is ok.
1135 if (const LoadInst *LI = dyn_cast<LoadInst>(InVal))
1136 if (LI->getOperand(0) == GV)
1141 // Anything else is rejected.
1149 static Value *GetHeapSROAValue(Value *V, unsigned FieldNo,
1150 DenseMap<Value *, std::vector<Value *>> &InsertedScalarizedValues,
1151 std::vector<std::pair<PHINode *, unsigned>> &PHIsToRewrite) {
1152 std::vector<Value *> &FieldVals = InsertedScalarizedValues[V];
1154 if (FieldNo >= FieldVals.size())
1155 FieldVals.resize(FieldNo+1);
1157 // If we already have this value, just reuse the previously scalarized
1159 if (Value *FieldVal = FieldVals[FieldNo])
1162 // Depending on what instruction this is, we have several cases.
1164 if (LoadInst *LI = dyn_cast<LoadInst>(V)) {
1165 // This is a scalarized version of the load from the global. Just create
1166 // a new Load of the scalarized global.
1167 Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo,
1168 InsertedScalarizedValues,
1170 LI->getName()+".f"+Twine(FieldNo), LI);
1172 PHINode *PN = cast<PHINode>(V);
1173 // PN's type is pointer to struct. Make a new PHI of pointer to struct
1176 PointerType *PTy = cast<PointerType>(PN->getType());
1177 StructType *ST = cast<StructType>(PTy->getElementType());
1179 unsigned AS = PTy->getAddressSpace();
1181 PHINode::Create(PointerType::get(ST->getElementType(FieldNo), AS),
1182 PN->getNumIncomingValues(),
1183 PN->getName()+".f"+Twine(FieldNo), PN);
1185 PHIsToRewrite.push_back(std::make_pair(PN, FieldNo));
1188 return FieldVals[FieldNo] = Result;
1191 /// Given a load instruction and a value derived from the load, rewrite the
1192 /// derived value to use the HeapSRoA'd load.
1193 static void RewriteHeapSROALoadUser(Instruction *LoadUser,
1194 DenseMap<Value *, std::vector<Value *>> &InsertedScalarizedValues,
1195 std::vector<std::pair<PHINode *, unsigned>> &PHIsToRewrite) {
1196 // If this is a comparison against null, handle it.
1197 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1198 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1199 // If we have a setcc of the loaded pointer, we can use a setcc of any
1201 Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0,
1202 InsertedScalarizedValues, PHIsToRewrite);
1204 Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr,
1205 Constant::getNullValue(NPtr->getType()),
1207 SCI->replaceAllUsesWith(New);
1208 SCI->eraseFromParent();
1212 // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...'
1213 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1214 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1215 && "Unexpected GEPI!");
1217 // Load the pointer for this field.
1218 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1219 Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo,
1220 InsertedScalarizedValues, PHIsToRewrite);
1222 // Create the new GEP idx vector.
1223 SmallVector<Value*, 8> GEPIdx;
1224 GEPIdx.push_back(GEPI->getOperand(1));
1225 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1227 Value *NGEPI = GetElementPtrInst::Create(GEPI->getResultElementType(), NewPtr, GEPIdx,
1228 GEPI->getName(), GEPI);
1229 GEPI->replaceAllUsesWith(NGEPI);
1230 GEPI->eraseFromParent();
1234 // Recursively transform the users of PHI nodes. This will lazily create the
1235 // PHIs that are needed for individual elements. Keep track of what PHIs we
1236 // see in InsertedScalarizedValues so that we don't get infinite loops (very
1237 // antisocial). If the PHI is already in InsertedScalarizedValues, it has
1238 // already been seen first by another load, so its uses have already been
1240 PHINode *PN = cast<PHINode>(LoadUser);
1241 if (!InsertedScalarizedValues.insert(std::make_pair(PN,
1242 std::vector<Value *>())).second)
1245 // If this is the first time we've seen this PHI, recursively process all
1247 for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
1248 Instruction *User = cast<Instruction>(*UI++);
1249 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1253 /// We are performing Heap SRoA on a global. Ptr is a value loaded from the
1254 /// global. Eliminate all uses of Ptr, making them use FieldGlobals instead.
1255 /// All uses of loaded values satisfy 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 /// CI is an allocation of an array of structures. Break it up into multiple
1271 /// 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 LLVM_DEBUG(dbgs() << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *CI
1277 Type *MAT = getMallocAllocatedType(CI, TLI);
1278 StructType *STy = cast<StructType>(MAT);
1280 // There is guaranteed to be at least one use of the malloc (storing
1281 // it into GV). If there are other uses, change them to be uses of
1282 // the global to simplify later code. This also deletes the store
1284 ReplaceUsesOfMallocWithGlobal(CI, GV);
1286 // Okay, at this point, there are no users of the malloc. Insert N
1287 // new mallocs at the same place as CI, and N globals.
1288 std::vector<Value *> FieldGlobals;
1289 std::vector<Value *> FieldMallocs;
1291 SmallVector<OperandBundleDef, 1> OpBundles;
1292 CI->getOperandBundlesAsDefs(OpBundles);
1294 unsigned AS = GV->getType()->getPointerAddressSpace();
1295 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1296 Type *FieldTy = STy->getElementType(FieldNo);
1297 PointerType *PFieldTy = PointerType::get(FieldTy, AS);
1299 GlobalVariable *NGV = new GlobalVariable(
1300 *GV->getParent(), PFieldTy, false, GlobalValue::InternalLinkage,
1301 Constant::getNullValue(PFieldTy), GV->getName() + ".f" + Twine(FieldNo),
1302 nullptr, GV->getThreadLocalMode());
1303 NGV->copyAttributesFrom(GV);
1304 FieldGlobals.push_back(NGV);
1306 unsigned TypeSize = DL.getTypeAllocSize(FieldTy);
1307 if (StructType *ST = dyn_cast<StructType>(FieldTy))
1308 TypeSize = DL.getStructLayout(ST)->getSizeInBytes();
1309 Type *IntPtrTy = DL.getIntPtrType(CI->getType());
1310 Value *NMI = CallInst::CreateMalloc(CI, IntPtrTy, FieldTy,
1311 ConstantInt::get(IntPtrTy, TypeSize),
1312 NElems, OpBundles, nullptr,
1313 CI->getName() + ".f" + Twine(FieldNo));
1314 FieldMallocs.push_back(NMI);
1315 new StoreInst(NMI, NGV, CI);
1318 // The tricky aspect of this transformation is handling the case when malloc
1319 // fails. In the original code, malloc failing would set the result pointer
1320 // of malloc to null. In this case, some mallocs could succeed and others
1321 // could fail. As such, we emit code that looks like this:
1322 // F0 = malloc(field0)
1323 // F1 = malloc(field1)
1324 // F2 = malloc(field2)
1325 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1326 // if (F0) { free(F0); F0 = 0; }
1327 // if (F1) { free(F1); F1 = 0; }
1328 // if (F2) { free(F2); F2 = 0; }
1330 // The malloc can also fail if its argument is too large.
1331 Constant *ConstantZero = ConstantInt::get(CI->getArgOperand(0)->getType(), 0);
1332 Value *RunningOr = new ICmpInst(CI, ICmpInst::ICMP_SLT, CI->getArgOperand(0),
1333 ConstantZero, "isneg");
1334 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1335 Value *Cond = new ICmpInst(CI, ICmpInst::ICMP_EQ, FieldMallocs[i],
1336 Constant::getNullValue(FieldMallocs[i]->getType()),
1338 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", CI);
1341 // Split the basic block at the old malloc.
1342 BasicBlock *OrigBB = CI->getParent();
1343 BasicBlock *ContBB =
1344 OrigBB->splitBasicBlock(CI->getIterator(), "malloc_cont");
1346 // Create the block to check the first condition. Put all these blocks at the
1347 // end of the function as they are unlikely to be executed.
1348 BasicBlock *NullPtrBlock = BasicBlock::Create(OrigBB->getContext(),
1350 OrigBB->getParent());
1352 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1353 // branch on RunningOr.
1354 OrigBB->getTerminator()->eraseFromParent();
1355 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1357 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1358 // pointer, because some may be null while others are not.
1359 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1360 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1361 Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
1362 Constant::getNullValue(GVVal->getType()));
1363 BasicBlock *FreeBlock = BasicBlock::Create(Cmp->getContext(), "free_it",
1364 OrigBB->getParent());
1365 BasicBlock *NextBlock = BasicBlock::Create(Cmp->getContext(), "next",
1366 OrigBB->getParent());
1367 Instruction *BI = BranchInst::Create(FreeBlock, NextBlock,
1370 // Fill in FreeBlock.
1371 CallInst::CreateFree(GVVal, OpBundles, BI);
1372 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1374 BranchInst::Create(NextBlock, FreeBlock);
1376 NullPtrBlock = NextBlock;
1379 BranchInst::Create(ContBB, NullPtrBlock);
1381 // CI is no longer needed, remove it.
1382 CI->eraseFromParent();
1384 /// As we process loads, if we can't immediately update all uses of the load,
1385 /// keep track of what scalarized loads are inserted for a given load.
1386 DenseMap<Value *, std::vector<Value *>> InsertedScalarizedValues;
1387 InsertedScalarizedValues[GV] = FieldGlobals;
1389 std::vector<std::pair<PHINode *, unsigned>> PHIsToRewrite;
1391 // Okay, the malloc site is completely handled. All of the uses of GV are now
1392 // loads, and all uses of those loads are simple. Rewrite them to use loads
1393 // of the per-field globals instead.
1394 for (auto UI = GV->user_begin(), E = GV->user_end(); UI != E;) {
1395 Instruction *User = cast<Instruction>(*UI++);
1397 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1398 RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite);
1402 // Must be a store of null.
1403 StoreInst *SI = cast<StoreInst>(User);
1404 assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
1405 "Unexpected heap-sra user!");
1407 // Insert a store of null into each global.
1408 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1409 Type *ValTy = cast<GlobalValue>(FieldGlobals[i])->getValueType();
1410 Constant *Null = Constant::getNullValue(ValTy);
1411 new StoreInst(Null, FieldGlobals[i], SI);
1413 // Erase the original store.
1414 SI->eraseFromParent();
1417 // While we have PHIs that are interesting to rewrite, do it.
1418 while (!PHIsToRewrite.empty()) {
1419 PHINode *PN = PHIsToRewrite.back().first;
1420 unsigned FieldNo = PHIsToRewrite.back().second;
1421 PHIsToRewrite.pop_back();
1422 PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]);
1423 assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi");
1425 // Add all the incoming values. This can materialize more phis.
1426 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1427 Value *InVal = PN->getIncomingValue(i);
1428 InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues,
1430 FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
1434 // Drop all inter-phi links and any loads that made it this far.
1435 for (DenseMap<Value *, std::vector<Value *>>::iterator
1436 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1438 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1439 PN->dropAllReferences();
1440 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1441 LI->dropAllReferences();
1444 // Delete all the phis and loads now that inter-references are dead.
1445 for (DenseMap<Value *, std::vector<Value *>>::iterator
1446 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1448 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1449 PN->eraseFromParent();
1450 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1451 LI->eraseFromParent();
1454 // The old global is now dead, remove it.
1455 GV->eraseFromParent();
1458 return cast<GlobalVariable>(FieldGlobals[0]);
1461 /// This function is called when we see a pointer global variable with a single
1462 /// value stored it that is a malloc or cast of malloc.
1463 static bool tryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, CallInst *CI,
1465 AtomicOrdering Ordering,
1466 const DataLayout &DL,
1467 TargetLibraryInfo *TLI) {
1468 // If this is a malloc of an abstract type, don't touch it.
1469 if (!AllocTy->isSized())
1472 // We can't optimize this global unless all uses of it are *known* to be
1473 // of the malloc value, not of the null initializer value (consider a use
1474 // that compares the global's value against zero to see if the malloc has
1475 // been reached). To do this, we check to see if all uses of the global
1476 // would trap if the global were null: this proves that they must all
1477 // happen after the malloc.
1478 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1481 // We can't optimize this if the malloc itself is used in a complex way,
1482 // for example, being stored into multiple globals. This allows the
1483 // malloc to be stored into the specified global, loaded icmp'd, and
1484 // GEP'd. These are all things we could transform to using the global
1486 SmallPtrSet<const PHINode*, 8> PHIs;
1487 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV, PHIs))
1490 // If we have a global that is only initialized with a fixed size malloc,
1491 // transform the program to use global memory instead of malloc'd memory.
1492 // This eliminates dynamic allocation, avoids an indirection accessing the
1493 // data, and exposes the resultant global to further GlobalOpt.
1494 // We cannot optimize the malloc if we cannot determine malloc array size.
1495 Value *NElems = getMallocArraySize(CI, DL, TLI, true);
1499 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
1500 // Restrict this transformation to only working on small allocations
1501 // (2048 bytes currently), as we don't want to introduce a 16M global or
1503 if (NElements->getZExtValue() * DL.getTypeAllocSize(AllocTy) < 2048) {
1504 OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, DL, TLI);
1508 // If the allocation is an array of structures, consider transforming this
1509 // into multiple malloc'd arrays, one for each field. This is basically
1510 // SRoA for malloc'd memory.
1512 if (Ordering != AtomicOrdering::NotAtomic)
1515 // If this is an allocation of a fixed size array of structs, analyze as a
1516 // variable size array. malloc [100 x struct],1 -> malloc struct, 100
1517 if (NElems == ConstantInt::get(CI->getArgOperand(0)->getType(), 1))
1518 if (ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
1519 AllocTy = AT->getElementType();
1521 StructType *AllocSTy = dyn_cast<StructType>(AllocTy);
1525 // This the structure has an unreasonable number of fields, leave it
1527 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
1528 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, CI)) {
1530 // If this is a fixed size array, transform the Malloc to be an alloc of
1531 // structs. malloc [100 x struct],1 -> malloc struct, 100
1532 if (ArrayType *AT = dyn_cast<ArrayType>(getMallocAllocatedType(CI, TLI))) {
1533 Type *IntPtrTy = DL.getIntPtrType(CI->getType());
1534 unsigned TypeSize = DL.getStructLayout(AllocSTy)->getSizeInBytes();
1535 Value *AllocSize = ConstantInt::get(IntPtrTy, TypeSize);
1536 Value *NumElements = ConstantInt::get(IntPtrTy, AT->getNumElements());
1537 SmallVector<OperandBundleDef, 1> OpBundles;
1538 CI->getOperandBundlesAsDefs(OpBundles);
1539 Instruction *Malloc =
1540 CallInst::CreateMalloc(CI, IntPtrTy, AllocSTy, AllocSize, NumElements,
1541 OpBundles, nullptr, CI->getName());
1542 Instruction *Cast = new BitCastInst(Malloc, CI->getType(), "tmp", CI);
1543 CI->replaceAllUsesWith(Cast);
1544 CI->eraseFromParent();
1545 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Malloc))
1546 CI = cast<CallInst>(BCI->getOperand(0));
1548 CI = cast<CallInst>(Malloc);
1551 PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, DL, TLI, true), DL,
1559 // Try to optimize globals based on the knowledge that only one value (besides
1560 // its initializer) is ever stored to the global.
1561 static bool optimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1562 AtomicOrdering Ordering,
1563 const DataLayout &DL,
1564 TargetLibraryInfo *TLI) {
1565 // Ignore no-op GEPs and bitcasts.
1566 StoredOnceVal = StoredOnceVal->stripPointerCasts();
1568 // If we are dealing with a pointer global that is initialized to null and
1569 // only has one (non-null) value stored into it, then we can optimize any
1570 // users of the loaded value (often calls and loads) that would trap if the
1572 if (GV->getInitializer()->getType()->isPointerTy() &&
1573 GV->getInitializer()->isNullValue() &&
1574 !NullPointerIsDefined(
1576 GV->getInitializer()->getType()->getPointerAddressSpace())) {
1577 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1578 if (GV->getInitializer()->getType() != SOVC->getType())
1579 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1581 // Optimize away any trapping uses of the loaded value.
1582 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC, DL, TLI))
1584 } else if (CallInst *CI = extractMallocCall(StoredOnceVal, TLI)) {
1585 Type *MallocType = getMallocAllocatedType(CI, TLI);
1586 if (MallocType && tryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType,
1595 /// At this point, we have learned that the only two values ever stored into GV
1596 /// are its initializer and OtherVal. See if we can shrink the global into a
1597 /// boolean and select between the two values whenever it is used. This exposes
1598 /// the values to other scalar optimizations.
1599 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1600 Type *GVElType = GV->getValueType();
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 LLVM_DEBUG(dbgs() << " *** SHRINKING TO BOOL: " << *GV << "\n");
1620 // Create the new global, initializing it to false.
1621 GlobalVariable *NewGV = new GlobalVariable(Type::getInt1Ty(GV->getContext()),
1623 GlobalValue::InternalLinkage,
1624 ConstantInt::getFalse(GV->getContext()),
1626 GV->getThreadLocalMode(),
1627 GV->getType()->getAddressSpace());
1628 NewGV->copyAttributesFrom(GV);
1629 GV->getParent()->getGlobalList().insert(GV->getIterator(), NewGV);
1631 Constant *InitVal = GV->getInitializer();
1632 assert(InitVal->getType() != Type::getInt1Ty(GV->getContext()) &&
1633 "No reason to shrink to bool!");
1635 SmallVector<DIGlobalVariableExpression *, 1> GVs;
1636 GV->getDebugInfo(GVs);
1638 // If initialized to zero and storing one into the global, we can use a cast
1639 // instead of a select to synthesize the desired value.
1640 bool IsOneZero = false;
1641 bool EmitOneOrZero = true;
1642 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)){
1643 IsOneZero = InitVal->isNullValue() && CI->isOne();
1645 if (ConstantInt *CIInit = dyn_cast<ConstantInt>(GV->getInitializer())){
1646 uint64_t ValInit = CIInit->getZExtValue();
1647 uint64_t ValOther = CI->getZExtValue();
1648 uint64_t ValMinus = ValOther - ValInit;
1650 for(auto *GVe : GVs){
1651 DIGlobalVariable *DGV = GVe->getVariable();
1652 DIExpression *E = GVe->getExpression();
1654 // It is expected that the address of global optimized variable is on
1655 // top of the stack. After optimization, value of that variable will
1656 // be ether 0 for initial value or 1 for other value. The following
1657 // expression should return constant integer value depending on the
1658 // value at global object address:
1659 // val * (ValOther - ValInit) + ValInit:
1660 // DW_OP_deref DW_OP_constu <ValMinus>
1661 // DW_OP_mul DW_OP_constu <ValInit> DW_OP_plus DW_OP_stack_value
1662 SmallVector<uint64_t, 12> Ops = {
1663 dwarf::DW_OP_deref, dwarf::DW_OP_constu, ValMinus,
1664 dwarf::DW_OP_mul, dwarf::DW_OP_constu, ValInit,
1666 E = DIExpression::prependOpcodes(E, Ops, DIExpression::WithStackValue);
1667 DIGlobalVariableExpression *DGVE =
1668 DIGlobalVariableExpression::get(NewGV->getContext(), DGV, E);
1669 NewGV->addDebugInfo(DGVE);
1671 EmitOneOrZero = false;
1675 if (EmitOneOrZero) {
1676 // FIXME: This will only emit address for debugger on which will
1677 // be written only 0 or 1.
1679 NewGV->addDebugInfo(GV);
1682 while (!GV->use_empty()) {
1683 Instruction *UI = cast<Instruction>(GV->user_back());
1684 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1685 // Change the store into a boolean store.
1686 bool StoringOther = SI->getOperand(0) == OtherVal;
1687 // Only do this if we weren't storing a loaded value.
1689 if (StoringOther || SI->getOperand(0) == InitVal) {
1690 StoreVal = ConstantInt::get(Type::getInt1Ty(GV->getContext()),
1693 // Otherwise, we are storing a previously loaded copy. To do this,
1694 // change the copy from copying the original value to just copying the
1696 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1698 // If we've already replaced the input, StoredVal will be a cast or
1699 // select instruction. If not, it will be a load of the original
1701 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1702 assert(LI->getOperand(0) == GV && "Not a copy!");
1703 // Insert a new load, to preserve the saved value.
1704 StoreVal = new LoadInst(NewGV, LI->getName()+".b", false, 0,
1705 LI->getOrdering(), LI->getSyncScopeID(), LI);
1707 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1708 "This is not a form that we understand!");
1709 StoreVal = StoredVal->getOperand(0);
1710 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1713 new StoreInst(StoreVal, NewGV, false, 0,
1714 SI->getOrdering(), SI->getSyncScopeID(), SI);
1716 // Change the load into a load of bool then a select.
1717 LoadInst *LI = cast<LoadInst>(UI);
1718 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", false, 0,
1719 LI->getOrdering(), LI->getSyncScopeID(), LI);
1722 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1724 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1726 LI->replaceAllUsesWith(NSI);
1728 UI->eraseFromParent();
1731 // Retain the name of the old global variable. People who are debugging their
1732 // programs may expect these variables to be named the same.
1733 NewGV->takeName(GV);
1734 GV->eraseFromParent();
1738 static bool deleteIfDead(
1739 GlobalValue &GV, SmallPtrSetImpl<const Comdat *> &NotDiscardableComdats) {
1740 GV.removeDeadConstantUsers();
1742 if (!GV.isDiscardableIfUnused() && !GV.isDeclaration())
1745 if (const Comdat *C = GV.getComdat())
1746 if (!GV.hasLocalLinkage() && NotDiscardableComdats.count(C))
1750 if (auto *F = dyn_cast<Function>(&GV))
1751 Dead = (F->isDeclaration() && F->use_empty()) || F->isDefTriviallyDead();
1753 Dead = GV.use_empty();
1757 LLVM_DEBUG(dbgs() << "GLOBAL DEAD: " << GV << "\n");
1758 GV.eraseFromParent();
1763 static bool isPointerValueDeadOnEntryToFunction(
1764 const Function *F, GlobalValue *GV,
1765 function_ref<DominatorTree &(Function &)> LookupDomTree) {
1766 // Find all uses of GV. We expect them all to be in F, and if we can't
1767 // identify any of the uses we bail out.
1769 // On each of these uses, identify if the memory that GV points to is
1770 // used/required/live at the start of the function. If it is not, for example
1771 // if the first thing the function does is store to the GV, the GV can
1772 // possibly be demoted.
1774 // We don't do an exhaustive search for memory operations - simply look
1775 // through bitcasts as they're quite common and benign.
1776 const DataLayout &DL = GV->getParent()->getDataLayout();
1777 SmallVector<LoadInst *, 4> Loads;
1778 SmallVector<StoreInst *, 4> Stores;
1779 for (auto *U : GV->users()) {
1780 if (Operator::getOpcode(U) == Instruction::BitCast) {
1781 for (auto *UU : U->users()) {
1782 if (auto *LI = dyn_cast<LoadInst>(UU))
1783 Loads.push_back(LI);
1784 else if (auto *SI = dyn_cast<StoreInst>(UU))
1785 Stores.push_back(SI);
1792 Instruction *I = dyn_cast<Instruction>(U);
1795 assert(I->getParent()->getParent() == F);
1797 if (auto *LI = dyn_cast<LoadInst>(I))
1798 Loads.push_back(LI);
1799 else if (auto *SI = dyn_cast<StoreInst>(I))
1800 Stores.push_back(SI);
1805 // We have identified all uses of GV into loads and stores. Now check if all
1806 // of them are known not to depend on the value of the global at the function
1807 // entry point. We do this by ensuring that every load is dominated by at
1809 auto &DT = LookupDomTree(*const_cast<Function *>(F));
1811 // The below check is quadratic. Check we're not going to do too many tests.
1812 // FIXME: Even though this will always have worst-case quadratic time, we
1813 // could put effort into minimizing the average time by putting stores that
1814 // have been shown to dominate at least one load at the beginning of the
1815 // Stores array, making subsequent dominance checks more likely to succeed
1818 // The threshold here is fairly large because global->local demotion is a
1819 // very powerful optimization should it fire.
1820 const unsigned Threshold = 100;
1821 if (Loads.size() * Stores.size() > Threshold)
1824 for (auto *L : Loads) {
1825 auto *LTy = L->getType();
1826 if (none_of(Stores, [&](const StoreInst *S) {
1827 auto *STy = S->getValueOperand()->getType();
1828 // The load is only dominated by the store if DomTree says so
1829 // and the number of bits loaded in L is less than or equal to
1830 // the number of bits stored in S.
1831 return DT.dominates(S, L) &&
1832 DL.getTypeStoreSize(LTy) <= DL.getTypeStoreSize(STy);
1836 // All loads have known dependences inside F, so the global can be localized.
1840 /// C may have non-instruction users. Can all of those users be turned into
1842 static bool allNonInstructionUsersCanBeMadeInstructions(Constant *C) {
1843 // We don't do this exhaustively. The most common pattern that we really need
1844 // to care about is a constant GEP or constant bitcast - so just looking
1845 // through one single ConstantExpr.
1847 // The set of constants that this function returns true for must be able to be
1848 // handled by makeAllConstantUsesInstructions.
1849 for (auto *U : C->users()) {
1850 if (isa<Instruction>(U))
1852 if (!isa<ConstantExpr>(U))
1853 // Non instruction, non-constantexpr user; cannot convert this.
1855 for (auto *UU : U->users())
1856 if (!isa<Instruction>(UU))
1857 // A constantexpr used by another constant. We don't try and recurse any
1858 // further but just bail out at this point.
1865 /// C may have non-instruction users, and
1866 /// allNonInstructionUsersCanBeMadeInstructions has returned true. Convert the
1867 /// non-instruction users to instructions.
1868 static void makeAllConstantUsesInstructions(Constant *C) {
1869 SmallVector<ConstantExpr*,4> Users;
1870 for (auto *U : C->users()) {
1871 if (isa<ConstantExpr>(U))
1872 Users.push_back(cast<ConstantExpr>(U));
1874 // We should never get here; allNonInstructionUsersCanBeMadeInstructions
1875 // should not have returned true for C.
1877 isa<Instruction>(U) &&
1878 "Can't transform non-constantexpr non-instruction to instruction!");
1881 SmallVector<Value*,4> UUsers;
1882 for (auto *U : Users) {
1884 for (auto *UU : U->users())
1885 UUsers.push_back(UU);
1886 for (auto *UU : UUsers) {
1887 Instruction *UI = cast<Instruction>(UU);
1888 Instruction *NewU = U->getAsInstruction();
1889 NewU->insertBefore(UI);
1890 UI->replaceUsesOfWith(U, NewU);
1892 // We've replaced all the uses, so destroy the constant. (destroyConstant
1893 // will update value handles and metadata.)
1894 U->destroyConstant();
1898 /// Analyze the specified global variable and optimize
1899 /// it if possible. If we make a change, return true.
1900 static bool processInternalGlobal(
1901 GlobalVariable *GV, const GlobalStatus &GS, TargetLibraryInfo *TLI,
1902 function_ref<DominatorTree &(Function &)> LookupDomTree) {
1903 auto &DL = GV->getParent()->getDataLayout();
1904 // If this is a first class global and has only one accessing function and
1905 // this function is non-recursive, we replace the global with a local alloca
1906 // in this function.
1908 // NOTE: It doesn't make sense to promote non-single-value types since we
1909 // are just replacing static memory to stack memory.
1911 // If the global is in different address space, don't bring it to stack.
1912 if (!GS.HasMultipleAccessingFunctions &&
1913 GS.AccessingFunction &&
1914 GV->getValueType()->isSingleValueType() &&
1915 GV->getType()->getAddressSpace() == 0 &&
1916 !GV->isExternallyInitialized() &&
1917 allNonInstructionUsersCanBeMadeInstructions(GV) &&
1918 GS.AccessingFunction->doesNotRecurse() &&
1919 isPointerValueDeadOnEntryToFunction(GS.AccessingFunction, GV,
1921 const DataLayout &DL = GV->getParent()->getDataLayout();
1923 LLVM_DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV << "\n");
1924 Instruction &FirstI = const_cast<Instruction&>(*GS.AccessingFunction
1925 ->getEntryBlock().begin());
1926 Type *ElemTy = GV->getValueType();
1927 // FIXME: Pass Global's alignment when globals have alignment
1928 AllocaInst *Alloca = new AllocaInst(ElemTy, DL.getAllocaAddrSpace(), nullptr,
1929 GV->getName(), &FirstI);
1930 if (!isa<UndefValue>(GV->getInitializer()))
1931 new StoreInst(GV->getInitializer(), Alloca, &FirstI);
1933 makeAllConstantUsesInstructions(GV);
1935 GV->replaceAllUsesWith(Alloca);
1936 GV->eraseFromParent();
1941 // If the global is never loaded (but may be stored to), it is dead.
1944 LLVM_DEBUG(dbgs() << "GLOBAL NEVER LOADED: " << *GV << "\n");
1947 if (isLeakCheckerRoot(GV)) {
1948 // Delete any constant stores to the global.
1949 Changed = CleanupPointerRootUsers(GV, TLI);
1951 // Delete any stores we can find to the global. We may not be able to
1952 // make it completely dead though.
1953 Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, TLI);
1956 // If the global is dead now, delete it.
1957 if (GV->use_empty()) {
1958 GV->eraseFromParent();
1965 if (GS.StoredType <= GlobalStatus::InitializerStored) {
1966 LLVM_DEBUG(dbgs() << "MARKING CONSTANT: " << *GV << "\n");
1967 GV->setConstant(true);
1969 // Clean up any obviously simplifiable users now.
1970 CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, TLI);
1972 // If the global is dead now, just nuke it.
1973 if (GV->use_empty()) {
1974 LLVM_DEBUG(dbgs() << " *** Marking constant allowed us to simplify "
1975 << "all users and delete global!\n");
1976 GV->eraseFromParent();
1981 // Fall through to the next check; see if we can optimize further.
1984 if (!GV->getInitializer()->getType()->isSingleValueType()) {
1985 const DataLayout &DL = GV->getParent()->getDataLayout();
1986 if (SRAGlobal(GV, DL))
1989 if (GS.StoredType == GlobalStatus::StoredOnce && GS.StoredOnceValue) {
1990 // If the initial value for the global was an undef value, and if only
1991 // one other value was stored into it, we can just change the
1992 // initializer to be the stored value, then delete all stores to the
1993 // global. This allows us to mark it constant.
1994 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1995 if (isa<UndefValue>(GV->getInitializer())) {
1996 // Change the initial value here.
1997 GV->setInitializer(SOVConstant);
1999 // Clean up any obviously simplifiable users now.
2000 CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, TLI);
2002 if (GV->use_empty()) {
2003 LLVM_DEBUG(dbgs() << " *** Substituting initializer allowed us to "
2004 << "simplify all users and delete global!\n");
2005 GV->eraseFromParent();
2012 // Try to optimize globals based on the knowledge that only one value
2013 // (besides its initializer) is ever stored to the global.
2014 if (optimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GS.Ordering, DL, TLI))
2017 // Otherwise, if the global was not a boolean, we can shrink it to be a
2019 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue)) {
2020 if (GS.Ordering == AtomicOrdering::NotAtomic) {
2021 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
2032 /// Analyze the specified global variable and optimize it if possible. If we
2033 /// make a change, return true.
2035 processGlobal(GlobalValue &GV, TargetLibraryInfo *TLI,
2036 function_ref<DominatorTree &(Function &)> LookupDomTree) {
2037 if (GV.getName().startswith("llvm."))
2042 if (GlobalStatus::analyzeGlobal(&GV, GS))
2045 bool Changed = false;
2046 if (!GS.IsCompared && !GV.hasGlobalUnnamedAddr()) {
2047 auto NewUnnamedAddr = GV.hasLocalLinkage() ? GlobalValue::UnnamedAddr::Global
2048 : GlobalValue::UnnamedAddr::Local;
2049 if (NewUnnamedAddr != GV.getUnnamedAddr()) {
2050 GV.setUnnamedAddr(NewUnnamedAddr);
2056 // Do more involved optimizations if the global is internal.
2057 if (!GV.hasLocalLinkage())
2060 auto *GVar = dyn_cast<GlobalVariable>(&GV);
2064 if (GVar->isConstant() || !GVar->hasInitializer())
2067 return processInternalGlobal(GVar, GS, TLI, LookupDomTree) || Changed;
2070 /// Walk all of the direct calls of the specified function, changing them to
2072 static void ChangeCalleesToFastCall(Function *F) {
2073 for (User *U : F->users()) {
2074 if (isa<BlockAddress>(U))
2076 CallSite CS(cast<Instruction>(U));
2077 CS.setCallingConv(CallingConv::Fast);
2081 static AttributeList StripNest(LLVMContext &C, AttributeList Attrs) {
2082 // There can be at most one attribute set with a nest attribute.
2084 if (Attrs.hasAttrSomewhere(Attribute::Nest, &NestIndex))
2085 return Attrs.removeAttribute(C, NestIndex, Attribute::Nest);
2089 static void RemoveNestAttribute(Function *F) {
2090 F->setAttributes(StripNest(F->getContext(), F->getAttributes()));
2091 for (User *U : F->users()) {
2092 if (isa<BlockAddress>(U))
2094 CallSite CS(cast<Instruction>(U));
2095 CS.setAttributes(StripNest(F->getContext(), CS.getAttributes()));
2099 /// Return true if this is a calling convention that we'd like to change. The
2100 /// idea here is that we don't want to mess with the convention if the user
2101 /// explicitly requested something with performance implications like coldcc,
2102 /// GHC, or anyregcc.
2103 static bool hasChangeableCC(Function *F) {
2104 CallingConv::ID CC = F->getCallingConv();
2106 // FIXME: Is it worth transforming x86_stdcallcc and x86_fastcallcc?
2107 if (CC != CallingConv::C && CC != CallingConv::X86_ThisCall)
2110 // FIXME: Change CC for the whole chain of musttail calls when possible.
2112 // Can't change CC of the function that either has musttail calls, or is a
2113 // musttail callee itself
2114 for (User *U : F->users()) {
2115 if (isa<BlockAddress>(U))
2117 CallInst* CI = dyn_cast<CallInst>(U);
2121 if (CI->isMustTailCall())
2125 for (BasicBlock &BB : *F)
2126 if (BB.getTerminatingMustTailCall())
2132 /// Return true if the block containing the call site has a BlockFrequency of
2133 /// less than ColdCCRelFreq% of the entry block.
2134 static bool isColdCallSite(CallSite CS, BlockFrequencyInfo &CallerBFI) {
2135 const BranchProbability ColdProb(ColdCCRelFreq, 100);
2136 auto CallSiteBB = CS.getInstruction()->getParent();
2137 auto CallSiteFreq = CallerBFI.getBlockFreq(CallSiteBB);
2138 auto CallerEntryFreq =
2139 CallerBFI.getBlockFreq(&(CS.getCaller()->getEntryBlock()));
2140 return CallSiteFreq < CallerEntryFreq * ColdProb;
2143 // This function checks if the input function F is cold at all call sites. It
2144 // also looks each call site's containing function, returning false if the
2145 // caller function contains other non cold calls. The input vector AllCallsCold
2146 // contains a list of functions that only have call sites in cold blocks.
2148 isValidCandidateForColdCC(Function &F,
2149 function_ref<BlockFrequencyInfo &(Function &)> GetBFI,
2150 const std::vector<Function *> &AllCallsCold) {
2155 for (User *U : F.users()) {
2156 if (isa<BlockAddress>(U))
2159 CallSite CS(cast<Instruction>(U));
2160 Function *CallerFunc = CS.getInstruction()->getParent()->getParent();
2161 BlockFrequencyInfo &CallerBFI = GetBFI(*CallerFunc);
2162 if (!isColdCallSite(CS, CallerBFI))
2164 auto It = std::find(AllCallsCold.begin(), AllCallsCold.end(), CallerFunc);
2165 if (It == AllCallsCold.end())
2171 static void changeCallSitesToColdCC(Function *F) {
2172 for (User *U : F->users()) {
2173 if (isa<BlockAddress>(U))
2175 CallSite CS(cast<Instruction>(U));
2176 CS.setCallingConv(CallingConv::Cold);
2180 // This function iterates over all the call instructions in the input Function
2181 // and checks that all call sites are in cold blocks and are allowed to use the
2182 // coldcc calling convention.
2184 hasOnlyColdCalls(Function &F,
2185 function_ref<BlockFrequencyInfo &(Function &)> GetBFI) {
2186 for (BasicBlock &BB : F) {
2187 for (Instruction &I : BB) {
2188 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
2189 CallSite CS(cast<Instruction>(CI));
2190 // Skip over isline asm instructions since they aren't function calls.
2191 if (CI->isInlineAsm())
2193 Function *CalledFn = CI->getCalledFunction();
2196 if (!CalledFn->hasLocalLinkage())
2198 // Skip over instrinsics since they won't remain as function calls.
2199 if (CalledFn->getIntrinsicID() != Intrinsic::not_intrinsic)
2201 // Check if it's valid to use coldcc calling convention.
2202 if (!hasChangeableCC(CalledFn) || CalledFn->isVarArg() ||
2203 CalledFn->hasAddressTaken())
2205 BlockFrequencyInfo &CallerBFI = GetBFI(F);
2206 if (!isColdCallSite(CS, CallerBFI))
2215 OptimizeFunctions(Module &M, TargetLibraryInfo *TLI,
2216 function_ref<TargetTransformInfo &(Function &)> GetTTI,
2217 function_ref<BlockFrequencyInfo &(Function &)> GetBFI,
2218 function_ref<DominatorTree &(Function &)> LookupDomTree,
2219 SmallPtrSetImpl<const Comdat *> &NotDiscardableComdats) {
2221 bool Changed = false;
2223 std::vector<Function *> AllCallsCold;
2224 for (Module::iterator FI = M.begin(), E = M.end(); FI != E;) {
2225 Function *F = &*FI++;
2226 if (hasOnlyColdCalls(*F, GetBFI))
2227 AllCallsCold.push_back(F);
2230 // Optimize functions.
2231 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
2232 Function *F = &*FI++;
2234 // Don't perform global opt pass on naked functions; we don't want fast
2235 // calling conventions for naked functions.
2236 if (F->hasFnAttribute(Attribute::Naked))
2239 // Functions without names cannot be referenced outside this module.
2240 if (!F->hasName() && !F->isDeclaration() && !F->hasLocalLinkage())
2241 F->setLinkage(GlobalValue::InternalLinkage);
2243 if (deleteIfDead(*F, NotDiscardableComdats)) {
2248 // LLVM's definition of dominance allows instructions that are cyclic
2249 // in unreachable blocks, e.g.:
2250 // %pat = select i1 %condition, @global, i16* %pat
2251 // because any instruction dominates an instruction in a block that's
2252 // not reachable from entry.
2253 // So, remove unreachable blocks from the function, because a) there's
2254 // no point in analyzing them and b) GlobalOpt should otherwise grow
2255 // some more complicated logic to break these cycles.
2256 // Removing unreachable blocks might invalidate the dominator so we
2258 if (!F->isDeclaration()) {
2259 if (removeUnreachableBlocks(*F)) {
2260 auto &DT = LookupDomTree(*F);
2266 Changed |= processGlobal(*F, TLI, LookupDomTree);
2268 if (!F->hasLocalLinkage())
2271 if (hasChangeableCC(F) && !F->isVarArg() && !F->hasAddressTaken()) {
2273 TargetTransformInfo &TTI = GetTTI(*F);
2274 // Change the calling convention to coldcc if either stress testing is
2275 // enabled or the target would like to use coldcc on functions which are
2276 // cold at all call sites and the callers contain no other non coldcc
2278 if (EnableColdCCStressTest ||
2279 (isValidCandidateForColdCC(*F, GetBFI, AllCallsCold) &&
2280 TTI.useColdCCForColdCall(*F))) {
2281 F->setCallingConv(CallingConv::Cold);
2282 changeCallSitesToColdCC(F);
2288 if (hasChangeableCC(F) && !F->isVarArg() &&
2289 !F->hasAddressTaken()) {
2290 // If this function has a calling convention worth changing, is not a
2291 // varargs function, and is only called directly, promote it to use the
2292 // Fast calling convention.
2293 F->setCallingConv(CallingConv::Fast);
2294 ChangeCalleesToFastCall(F);
2299 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
2300 !F->hasAddressTaken()) {
2301 // The function is not used by a trampoline intrinsic, so it is safe
2302 // to remove the 'nest' attribute.
2303 RemoveNestAttribute(F);
2312 OptimizeGlobalVars(Module &M, TargetLibraryInfo *TLI,
2313 function_ref<DominatorTree &(Function &)> LookupDomTree,
2314 SmallPtrSetImpl<const Comdat *> &NotDiscardableComdats) {
2315 bool Changed = false;
2317 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
2319 GlobalVariable *GV = &*GVI++;
2320 // Global variables without names cannot be referenced outside this module.
2321 if (!GV->hasName() && !GV->isDeclaration() && !GV->hasLocalLinkage())
2322 GV->setLinkage(GlobalValue::InternalLinkage);
2323 // Simplify the initializer.
2324 if (GV->hasInitializer())
2325 if (auto *C = dyn_cast<Constant>(GV->getInitializer())) {
2326 auto &DL = M.getDataLayout();
2327 Constant *New = ConstantFoldConstant(C, DL, TLI);
2328 if (New && New != C)
2329 GV->setInitializer(New);
2332 if (deleteIfDead(*GV, NotDiscardableComdats)) {
2337 Changed |= processGlobal(*GV, TLI, LookupDomTree);
2342 /// Evaluate a piece of a constantexpr store into a global initializer. This
2343 /// returns 'Init' modified to reflect 'Val' stored into it. At this point, the
2344 /// GEP operands of Addr [0, OpNo) have been stepped into.
2345 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
2346 ConstantExpr *Addr, unsigned OpNo) {
2347 // Base case of the recursion.
2348 if (OpNo == Addr->getNumOperands()) {
2349 assert(Val->getType() == Init->getType() && "Type mismatch!");
2353 SmallVector<Constant*, 32> Elts;
2354 if (StructType *STy = dyn_cast<StructType>(Init->getType())) {
2355 // Break up the constant into its elements.
2356 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2357 Elts.push_back(Init->getAggregateElement(i));
2359 // Replace the element that we are supposed to.
2360 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
2361 unsigned Idx = CU->getZExtValue();
2362 assert(Idx < STy->getNumElements() && "Struct index out of range!");
2363 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
2365 // Return the modified struct.
2366 return ConstantStruct::get(STy, Elts);
2369 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
2370 SequentialType *InitTy = cast<SequentialType>(Init->getType());
2371 uint64_t NumElts = InitTy->getNumElements();
2373 // Break up the array into elements.
2374 for (uint64_t i = 0, e = NumElts; i != e; ++i)
2375 Elts.push_back(Init->getAggregateElement(i));
2377 assert(CI->getZExtValue() < NumElts);
2378 Elts[CI->getZExtValue()] =
2379 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
2381 if (Init->getType()->isArrayTy())
2382 return ConstantArray::get(cast<ArrayType>(InitTy), Elts);
2383 return ConstantVector::get(Elts);
2386 /// We have decided that Addr (which satisfies the predicate
2387 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
2388 static void CommitValueTo(Constant *Val, Constant *Addr) {
2389 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
2390 assert(GV->hasInitializer());
2391 GV->setInitializer(Val);
2395 ConstantExpr *CE = cast<ConstantExpr>(Addr);
2396 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2397 GV->setInitializer(EvaluateStoreInto(GV->getInitializer(), Val, CE, 2));
2400 /// Given a map of address -> value, where addresses are expected to be some form
2401 /// of either a global or a constant GEP, set the initializer for the address to
2402 /// be the value. This performs mostly the same function as CommitValueTo()
2403 /// and EvaluateStoreInto() but is optimized to be more efficient for the common
2404 /// case where the set of addresses are GEPs sharing the same underlying global,
2405 /// processing the GEPs in batches rather than individually.
2407 /// To give an example, consider the following C++ code adapted from the clang
2408 /// regression tests:
2412 /// S(int a) : n(a) {}
2415 /// template<typename T>
2424 /// The global static constructor for 'e' will need to initialize 'r' and 'p' of
2425 /// the outer struct, while also initializing the inner 'q' structs 'n' and 'm'
2426 /// members. This batch algorithm will simply use general CommitValueTo() method
2427 /// to handle the complex nested S struct initialization of 'q', before
2428 /// processing the outermost members in a single batch. Using CommitValueTo() to
2429 /// handle member in the outer struct is inefficient when the struct/array is
2430 /// very large as we end up creating and destroy constant arrays for each
2432 /// For the above case, we expect the following IR to be generated:
2434 /// %struct.U = type { %struct.S*, %struct.S, %struct.U* }
2435 /// %struct.S = type { i32, i32 }
2436 /// @e = global %struct.U { %struct.S* gep inbounds (%struct.U, %struct.U* @e,
2438 /// %struct.S { i32 42, i32 84 }, %struct.U* @e }
2439 /// The %struct.S { i32 42, i32 84 } inner initializer is treated as a complex
2440 /// constant expression, while the other two elements of @e are "simple".
2441 static void BatchCommitValueTo(const DenseMap<Constant*, Constant*> &Mem) {
2442 SmallVector<std::pair<GlobalVariable*, Constant*>, 32> GVs;
2443 SmallVector<std::pair<ConstantExpr*, Constant*>, 32> ComplexCEs;
2444 SmallVector<std::pair<ConstantExpr*, Constant*>, 32> SimpleCEs;
2445 SimpleCEs.reserve(Mem.size());
2447 for (const auto &I : Mem) {
2448 if (auto *GV = dyn_cast<GlobalVariable>(I.first)) {
2449 GVs.push_back(std::make_pair(GV, I.second));
2451 ConstantExpr *GEP = cast<ConstantExpr>(I.first);
2452 // We don't handle the deeply recursive case using the batch method.
2453 if (GEP->getNumOperands() > 3)
2454 ComplexCEs.push_back(std::make_pair(GEP, I.second));
2456 SimpleCEs.push_back(std::make_pair(GEP, I.second));
2460 // The algorithm below doesn't handle cases like nested structs, so use the
2461 // slower fully general method if we have to.
2462 for (auto ComplexCE : ComplexCEs)
2463 CommitValueTo(ComplexCE.second, ComplexCE.first);
2465 for (auto GVPair : GVs) {
2466 assert(GVPair.first->hasInitializer());
2467 GVPair.first->setInitializer(GVPair.second);
2470 if (SimpleCEs.empty())
2473 // We cache a single global's initializer elements in the case where the
2474 // subsequent address/val pair uses the same one. This avoids throwing away and
2475 // rebuilding the constant struct/vector/array just because one element is
2476 // modified at a time.
2477 SmallVector<Constant *, 32> Elts;
2478 Elts.reserve(SimpleCEs.size());
2479 GlobalVariable *CurrentGV = nullptr;
2481 auto commitAndSetupCache = [&](GlobalVariable *GV, bool Update) {
2482 Constant *Init = GV->getInitializer();
2483 Type *Ty = Init->getType();
2486 assert(CurrentGV && "Expected a GV to commit to!");
2487 Type *CurrentInitTy = CurrentGV->getInitializer()->getType();
2488 // We have a valid cache that needs to be committed.
2489 if (StructType *STy = dyn_cast<StructType>(CurrentInitTy))
2490 CurrentGV->setInitializer(ConstantStruct::get(STy, Elts));
2491 else if (ArrayType *ArrTy = dyn_cast<ArrayType>(CurrentInitTy))
2492 CurrentGV->setInitializer(ConstantArray::get(ArrTy, Elts));
2494 CurrentGV->setInitializer(ConstantVector::get(Elts));
2496 if (CurrentGV == GV)
2498 // Need to clear and set up cache for new initializer.
2502 if (auto *STy = dyn_cast<StructType>(Ty))
2503 NumElts = STy->getNumElements();
2505 NumElts = cast<SequentialType>(Ty)->getNumElements();
2506 for (unsigned i = 0, e = NumElts; i != e; ++i)
2507 Elts.push_back(Init->getAggregateElement(i));
2511 for (auto CEPair : SimpleCEs) {
2512 ConstantExpr *GEP = CEPair.first;
2513 Constant *Val = CEPair.second;
2515 GlobalVariable *GV = cast<GlobalVariable>(GEP->getOperand(0));
2516 commitAndSetupCache(GV, GV != CurrentGV);
2517 ConstantInt *CI = cast<ConstantInt>(GEP->getOperand(2));
2518 Elts[CI->getZExtValue()] = Val;
2520 // The last initializer in the list needs to be committed, others
2521 // will be committed on a new initializer being processed.
2522 commitAndSetupCache(CurrentGV, true);
2525 /// Evaluate static constructors in the function, if we can. Return true if we
2526 /// can, false otherwise.
2527 static bool EvaluateStaticConstructor(Function *F, const DataLayout &DL,
2528 TargetLibraryInfo *TLI) {
2529 // Call the function.
2530 Evaluator Eval(DL, TLI);
2531 Constant *RetValDummy;
2532 bool EvalSuccess = Eval.EvaluateFunction(F, RetValDummy,
2533 SmallVector<Constant*, 0>());
2536 ++NumCtorsEvaluated;
2538 // We succeeded at evaluation: commit the result.
2539 LLVM_DEBUG(dbgs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2540 << F->getName() << "' to "
2541 << Eval.getMutatedMemory().size() << " stores.\n");
2542 BatchCommitValueTo(Eval.getMutatedMemory());
2543 for (GlobalVariable *GV : Eval.getInvariants())
2544 GV->setConstant(true);
2550 static int compareNames(Constant *const *A, Constant *const *B) {
2551 Value *AStripped = (*A)->stripPointerCastsNoFollowAliases();
2552 Value *BStripped = (*B)->stripPointerCastsNoFollowAliases();
2553 return AStripped->getName().compare(BStripped->getName());
2556 static void setUsedInitializer(GlobalVariable &V,
2557 const SmallPtrSetImpl<GlobalValue *> &Init) {
2559 V.eraseFromParent();
2563 // Type of pointer to the array of pointers.
2564 PointerType *Int8PtrTy = Type::getInt8PtrTy(V.getContext(), 0);
2566 SmallVector<Constant *, 8> UsedArray;
2567 for (GlobalValue *GV : Init) {
2569 = ConstantExpr::getPointerBitCastOrAddrSpaceCast(GV, Int8PtrTy);
2570 UsedArray.push_back(Cast);
2572 // Sort to get deterministic order.
2573 array_pod_sort(UsedArray.begin(), UsedArray.end(), compareNames);
2574 ArrayType *ATy = ArrayType::get(Int8PtrTy, UsedArray.size());
2576 Module *M = V.getParent();
2577 V.removeFromParent();
2578 GlobalVariable *NV =
2579 new GlobalVariable(*M, ATy, false, GlobalValue::AppendingLinkage,
2580 ConstantArray::get(ATy, UsedArray), "");
2582 NV->setSection("llvm.metadata");
2588 /// An easy to access representation of llvm.used and llvm.compiler.used.
2590 SmallPtrSet<GlobalValue *, 8> Used;
2591 SmallPtrSet<GlobalValue *, 8> CompilerUsed;
2592 GlobalVariable *UsedV;
2593 GlobalVariable *CompilerUsedV;
2596 LLVMUsed(Module &M) {
2597 UsedV = collectUsedGlobalVariables(M, Used, false);
2598 CompilerUsedV = collectUsedGlobalVariables(M, CompilerUsed, true);
2601 using iterator = SmallPtrSet<GlobalValue *, 8>::iterator;
2602 using used_iterator_range = iterator_range<iterator>;
2604 iterator usedBegin() { return Used.begin(); }
2605 iterator usedEnd() { return Used.end(); }
2607 used_iterator_range used() {
2608 return used_iterator_range(usedBegin(), usedEnd());
2611 iterator compilerUsedBegin() { return CompilerUsed.begin(); }
2612 iterator compilerUsedEnd() { return CompilerUsed.end(); }
2614 used_iterator_range compilerUsed() {
2615 return used_iterator_range(compilerUsedBegin(), compilerUsedEnd());
2618 bool usedCount(GlobalValue *GV) const { return Used.count(GV); }
2620 bool compilerUsedCount(GlobalValue *GV) const {
2621 return CompilerUsed.count(GV);
2624 bool usedErase(GlobalValue *GV) { return Used.erase(GV); }
2625 bool compilerUsedErase(GlobalValue *GV) { return CompilerUsed.erase(GV); }
2626 bool usedInsert(GlobalValue *GV) { return Used.insert(GV).second; }
2628 bool compilerUsedInsert(GlobalValue *GV) {
2629 return CompilerUsed.insert(GV).second;
2632 void syncVariablesAndSets() {
2634 setUsedInitializer(*UsedV, Used);
2636 setUsedInitializer(*CompilerUsedV, CompilerUsed);
2640 } // end anonymous namespace
2642 static bool hasUseOtherThanLLVMUsed(GlobalAlias &GA, const LLVMUsed &U) {
2643 if (GA.use_empty()) // No use at all.
2646 assert((!U.usedCount(&GA) || !U.compilerUsedCount(&GA)) &&
2647 "We should have removed the duplicated "
2648 "element from llvm.compiler.used");
2649 if (!GA.hasOneUse())
2650 // Strictly more than one use. So at least one is not in llvm.used and
2651 // llvm.compiler.used.
2654 // Exactly one use. Check if it is in llvm.used or llvm.compiler.used.
2655 return !U.usedCount(&GA) && !U.compilerUsedCount(&GA);
2658 static bool hasMoreThanOneUseOtherThanLLVMUsed(GlobalValue &V,
2659 const LLVMUsed &U) {
2661 assert((!U.usedCount(&V) || !U.compilerUsedCount(&V)) &&
2662 "We should have removed the duplicated "
2663 "element from llvm.compiler.used");
2664 if (U.usedCount(&V) || U.compilerUsedCount(&V))
2666 return V.hasNUsesOrMore(N);
2669 static bool mayHaveOtherReferences(GlobalAlias &GA, const LLVMUsed &U) {
2670 if (!GA.hasLocalLinkage())
2673 return U.usedCount(&GA) || U.compilerUsedCount(&GA);
2676 static bool hasUsesToReplace(GlobalAlias &GA, const LLVMUsed &U,
2677 bool &RenameTarget) {
2678 RenameTarget = false;
2680 if (hasUseOtherThanLLVMUsed(GA, U))
2683 // If the alias is externally visible, we may still be able to simplify it.
2684 if (!mayHaveOtherReferences(GA, U))
2687 // If the aliasee has internal linkage, give it the name and linkage
2688 // of the alias, and delete the alias. This turns:
2689 // define internal ... @f(...)
2690 // @a = alias ... @f
2692 // define ... @a(...)
2693 Constant *Aliasee = GA.getAliasee();
2694 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2695 if (!Target->hasLocalLinkage())
2698 // Do not perform the transform if multiple aliases potentially target the
2699 // aliasee. This check also ensures that it is safe to replace the section
2700 // and other attributes of the aliasee with those of the alias.
2701 if (hasMoreThanOneUseOtherThanLLVMUsed(*Target, U))
2704 RenameTarget = true;
2709 OptimizeGlobalAliases(Module &M,
2710 SmallPtrSetImpl<const Comdat *> &NotDiscardableComdats) {
2711 bool Changed = false;
2714 for (GlobalValue *GV : Used.used())
2715 Used.compilerUsedErase(GV);
2717 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end();
2719 GlobalAlias *J = &*I++;
2721 // Aliases without names cannot be referenced outside this module.
2722 if (!J->hasName() && !J->isDeclaration() && !J->hasLocalLinkage())
2723 J->setLinkage(GlobalValue::InternalLinkage);
2725 if (deleteIfDead(*J, NotDiscardableComdats)) {
2730 // If the alias can change at link time, nothing can be done - bail out.
2731 if (J->isInterposable())
2734 Constant *Aliasee = J->getAliasee();
2735 GlobalValue *Target = dyn_cast<GlobalValue>(Aliasee->stripPointerCasts());
2736 // We can't trivially replace the alias with the aliasee if the aliasee is
2737 // non-trivial in some way.
2738 // TODO: Try to handle non-zero GEPs of local aliasees.
2741 Target->removeDeadConstantUsers();
2743 // Make all users of the alias use the aliasee instead.
2745 if (!hasUsesToReplace(*J, Used, RenameTarget))
2748 J->replaceAllUsesWith(ConstantExpr::getBitCast(Aliasee, J->getType()));
2749 ++NumAliasesResolved;
2753 // Give the aliasee the name, linkage and other attributes of the alias.
2754 Target->takeName(&*J);
2755 Target->setLinkage(J->getLinkage());
2756 Target->setDSOLocal(J->isDSOLocal());
2757 Target->setVisibility(J->getVisibility());
2758 Target->setDLLStorageClass(J->getDLLStorageClass());
2760 if (Used.usedErase(&*J))
2761 Used.usedInsert(Target);
2763 if (Used.compilerUsedErase(&*J))
2764 Used.compilerUsedInsert(Target);
2765 } else if (mayHaveOtherReferences(*J, Used))
2768 // Delete the alias.
2769 M.getAliasList().erase(J);
2770 ++NumAliasesRemoved;
2774 Used.syncVariablesAndSets();
2779 static Function *FindCXAAtExit(Module &M, TargetLibraryInfo *TLI) {
2780 LibFunc F = LibFunc_cxa_atexit;
2784 Function *Fn = M.getFunction(TLI->getName(F));
2788 // Make sure that the function has the correct prototype.
2789 if (!TLI->getLibFunc(*Fn, F) || F != LibFunc_cxa_atexit)
2795 /// Returns whether the given function is an empty C++ destructor and can
2796 /// therefore be eliminated.
2797 /// Note that we assume that other optimization passes have already simplified
2798 /// the code so we only look for a function with a single basic block, where
2799 /// the only allowed instructions are 'ret', 'call' to an empty C++ dtor and
2800 /// other side-effect free instructions.
2801 static bool cxxDtorIsEmpty(const Function &Fn,
2802 SmallPtrSet<const Function *, 8> &CalledFunctions) {
2803 // FIXME: We could eliminate C++ destructors if they're readonly/readnone and
2804 // nounwind, but that doesn't seem worth doing.
2805 if (Fn.isDeclaration())
2808 if (++Fn.begin() != Fn.end())
2811 const BasicBlock &EntryBlock = Fn.getEntryBlock();
2812 for (BasicBlock::const_iterator I = EntryBlock.begin(), E = EntryBlock.end();
2814 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
2815 // Ignore debug intrinsics.
2816 if (isa<DbgInfoIntrinsic>(CI))
2819 const Function *CalledFn = CI->getCalledFunction();
2824 SmallPtrSet<const Function *, 8> NewCalledFunctions(CalledFunctions);
2826 // Don't treat recursive functions as empty.
2827 if (!NewCalledFunctions.insert(CalledFn).second)
2830 if (!cxxDtorIsEmpty(*CalledFn, NewCalledFunctions))
2832 } else if (isa<ReturnInst>(*I))
2833 return true; // We're done.
2834 else if (I->mayHaveSideEffects())
2835 return false; // Destructor with side effects, bail.
2841 static bool OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) {
2842 /// Itanium C++ ABI p3.3.5:
2844 /// After constructing a global (or local static) object, that will require
2845 /// destruction on exit, a termination function is registered as follows:
2847 /// extern "C" int __cxa_atexit ( void (*f)(void *), void *p, void *d );
2849 /// This registration, e.g. __cxa_atexit(f,p,d), is intended to cause the
2850 /// call f(p) when DSO d is unloaded, before all such termination calls
2851 /// registered before this one. It returns zero if registration is
2852 /// successful, nonzero on failure.
2854 // This pass will look for calls to __cxa_atexit where the function is trivial
2856 bool Changed = false;
2858 for (auto I = CXAAtExitFn->user_begin(), E = CXAAtExitFn->user_end();
2860 // We're only interested in calls. Theoretically, we could handle invoke
2861 // instructions as well, but neither llvm-gcc nor clang generate invokes
2863 CallInst *CI = dyn_cast<CallInst>(*I++);
2868 dyn_cast<Function>(CI->getArgOperand(0)->stripPointerCasts());
2872 SmallPtrSet<const Function *, 8> CalledFunctions;
2873 if (!cxxDtorIsEmpty(*DtorFn, CalledFunctions))
2876 // Just remove the call.
2877 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
2878 CI->eraseFromParent();
2880 ++NumCXXDtorsRemoved;
2888 static bool optimizeGlobalsInModule(
2889 Module &M, const DataLayout &DL, TargetLibraryInfo *TLI,
2890 function_ref<TargetTransformInfo &(Function &)> GetTTI,
2891 function_ref<BlockFrequencyInfo &(Function &)> GetBFI,
2892 function_ref<DominatorTree &(Function &)> LookupDomTree) {
2893 SmallPtrSet<const Comdat *, 8> NotDiscardableComdats;
2894 bool Changed = false;
2895 bool LocalChange = true;
2896 while (LocalChange) {
2897 LocalChange = false;
2899 NotDiscardableComdats.clear();
2900 for (const GlobalVariable &GV : M.globals())
2901 if (const Comdat *C = GV.getComdat())
2902 if (!GV.isDiscardableIfUnused() || !GV.use_empty())
2903 NotDiscardableComdats.insert(C);
2904 for (Function &F : M)
2905 if (const Comdat *C = F.getComdat())
2906 if (!F.isDefTriviallyDead())
2907 NotDiscardableComdats.insert(C);
2908 for (GlobalAlias &GA : M.aliases())
2909 if (const Comdat *C = GA.getComdat())
2910 if (!GA.isDiscardableIfUnused() || !GA.use_empty())
2911 NotDiscardableComdats.insert(C);
2913 // Delete functions that are trivially dead, ccc -> fastcc
2914 LocalChange |= OptimizeFunctions(M, TLI, GetTTI, GetBFI, LookupDomTree,
2915 NotDiscardableComdats);
2917 // Optimize global_ctors list.
2918 LocalChange |= optimizeGlobalCtorsList(M, [&](Function *F) {
2919 return EvaluateStaticConstructor(F, DL, TLI);
2922 // Optimize non-address-taken globals.
2923 LocalChange |= OptimizeGlobalVars(M, TLI, LookupDomTree,
2924 NotDiscardableComdats);
2926 // Resolve aliases, when possible.
2927 LocalChange |= OptimizeGlobalAliases(M, NotDiscardableComdats);
2929 // Try to remove trivial global destructors if they are not removed
2931 Function *CXAAtExitFn = FindCXAAtExit(M, TLI);
2933 LocalChange |= OptimizeEmptyGlobalCXXDtors(CXAAtExitFn);
2935 Changed |= LocalChange;
2938 // TODO: Move all global ctors functions to the end of the module for code
2944 PreservedAnalyses GlobalOptPass::run(Module &M, ModuleAnalysisManager &AM) {
2945 auto &DL = M.getDataLayout();
2946 auto &TLI = AM.getResult<TargetLibraryAnalysis>(M);
2948 AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
2949 auto LookupDomTree = [&FAM](Function &F) -> DominatorTree &{
2950 return FAM.getResult<DominatorTreeAnalysis>(F);
2952 auto GetTTI = [&FAM](Function &F) -> TargetTransformInfo & {
2953 return FAM.getResult<TargetIRAnalysis>(F);
2956 auto GetBFI = [&FAM](Function &F) -> BlockFrequencyInfo & {
2957 return FAM.getResult<BlockFrequencyAnalysis>(F);
2960 if (!optimizeGlobalsInModule(M, DL, &TLI, GetTTI, GetBFI, LookupDomTree))
2961 return PreservedAnalyses::all();
2962 return PreservedAnalyses::none();
2967 struct GlobalOptLegacyPass : public ModulePass {
2968 static char ID; // Pass identification, replacement for typeid
2970 GlobalOptLegacyPass() : ModulePass(ID) {
2971 initializeGlobalOptLegacyPassPass(*PassRegistry::getPassRegistry());
2974 bool runOnModule(Module &M) override {
2978 auto &DL = M.getDataLayout();
2979 auto *TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
2980 auto LookupDomTree = [this](Function &F) -> DominatorTree & {
2981 return this->getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
2983 auto GetTTI = [this](Function &F) -> TargetTransformInfo & {
2984 return this->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2987 auto GetBFI = [this](Function &F) -> BlockFrequencyInfo & {
2988 return this->getAnalysis<BlockFrequencyInfoWrapperPass>(F).getBFI();
2991 return optimizeGlobalsInModule(M, DL, TLI, GetTTI, GetBFI, LookupDomTree);
2994 void getAnalysisUsage(AnalysisUsage &AU) const override {
2995 AU.addRequired<TargetLibraryInfoWrapperPass>();
2996 AU.addRequired<TargetTransformInfoWrapperPass>();
2997 AU.addRequired<DominatorTreeWrapperPass>();
2998 AU.addRequired<BlockFrequencyInfoWrapperPass>();
3002 } // end anonymous namespace
3004 char GlobalOptLegacyPass::ID = 0;
3006 INITIALIZE_PASS_BEGIN(GlobalOptLegacyPass, "globalopt",
3007 "Global Variable Optimizer", false, false)
3008 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
3009 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
3010 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
3011 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
3012 INITIALIZE_PASS_END(GlobalOptLegacyPass, "globalopt",
3013 "Global Variable Optimizer", false, false)
3015 ModulePass *llvm::createGlobalOptimizerPass() {
3016 return new GlobalOptLegacyPass();