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 #define DEBUG_TYPE "globalopt"
17 #include "llvm/Transforms/IPO.h"
18 #include "llvm/CallingConv.h"
19 #include "llvm/Constants.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/Instructions.h"
22 #include "llvm/IntrinsicInst.h"
23 #include "llvm/Module.h"
24 #include "llvm/Operator.h"
25 #include "llvm/Pass.h"
26 #include "llvm/Analysis/ConstantFolding.h"
27 #include "llvm/Analysis/MemoryBuiltins.h"
28 #include "llvm/Target/TargetData.h"
29 #include "llvm/Target/TargetLibraryInfo.h"
30 #include "llvm/Support/CallSite.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/GetElementPtrTypeIterator.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/ADT/DenseMap.h"
37 #include "llvm/ADT/SmallPtrSet.h"
38 #include "llvm/ADT/SmallVector.h"
39 #include "llvm/ADT/Statistic.h"
40 #include "llvm/ADT/STLExtras.h"
44 STATISTIC(NumMarked , "Number of globals marked constant");
45 STATISTIC(NumUnnamed , "Number of globals marked unnamed_addr");
46 STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
47 STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
48 STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
49 STATISTIC(NumDeleted , "Number of globals deleted");
50 STATISTIC(NumFnDeleted , "Number of functions deleted");
51 STATISTIC(NumGlobUses , "Number of global uses devirtualized");
52 STATISTIC(NumLocalized , "Number of globals localized");
53 STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans");
54 STATISTIC(NumFastCallFns , "Number of functions converted to fastcc");
55 STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
56 STATISTIC(NumNestRemoved , "Number of nest attributes removed");
57 STATISTIC(NumAliasesResolved, "Number of global aliases resolved");
58 STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated");
59 STATISTIC(NumCXXDtorsRemoved, "Number of global C++ destructors removed");
63 struct GlobalOpt : public ModulePass {
64 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
65 AU.addRequired<TargetLibraryInfo>();
67 static char ID; // Pass identification, replacement for typeid
68 GlobalOpt() : ModulePass(ID) {
69 initializeGlobalOptPass(*PassRegistry::getPassRegistry());
72 bool runOnModule(Module &M);
75 GlobalVariable *FindGlobalCtors(Module &M);
76 bool OptimizeFunctions(Module &M);
77 bool OptimizeGlobalVars(Module &M);
78 bool OptimizeGlobalAliases(Module &M);
79 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
80 bool ProcessGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
81 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI,
82 const SmallPtrSet<const PHINode*, 16> &PHIUsers,
83 const GlobalStatus &GS);
84 bool OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn);
88 char GlobalOpt::ID = 0;
89 INITIALIZE_PASS_BEGIN(GlobalOpt, "globalopt",
90 "Global Variable Optimizer", false, false)
91 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
92 INITIALIZE_PASS_END(GlobalOpt, "globalopt",
93 "Global Variable Optimizer", false, false)
95 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
99 /// GlobalStatus - As we analyze each global, keep track of some information
100 /// about it. If we find out that the address of the global is taken, none of
101 /// this info will be accurate.
102 struct GlobalStatus {
103 /// isCompared - True if the global's address is used in a comparison.
106 /// isLoaded - True if the global is ever loaded. If the global isn't ever
107 /// loaded it can be deleted.
110 /// StoredType - Keep track of what stores to the global look like.
113 /// NotStored - There is no store to this global. It can thus be marked
117 /// isInitializerStored - This global is stored to, but the only thing
118 /// stored is the constant it was initialized with. This is only tracked
119 /// for scalar globals.
122 /// isStoredOnce - This global is stored to, but only its initializer and
123 /// one other value is ever stored to it. If this global isStoredOnce, we
124 /// track the value stored to it in StoredOnceValue below. This is only
125 /// tracked for scalar globals.
128 /// isStored - This global is stored to by multiple values or something else
129 /// that we cannot track.
133 /// StoredOnceValue - If only one value (besides the initializer constant) is
134 /// ever stored to this global, keep track of what value it is.
135 Value *StoredOnceValue;
137 /// AccessingFunction/HasMultipleAccessingFunctions - These start out
138 /// null/false. When the first accessing function is noticed, it is recorded.
139 /// When a second different accessing function is noticed,
140 /// HasMultipleAccessingFunctions is set to true.
141 const Function *AccessingFunction;
142 bool HasMultipleAccessingFunctions;
144 /// HasNonInstructionUser - Set to true if this global has a user that is not
145 /// an instruction (e.g. a constant expr or GV initializer).
146 bool HasNonInstructionUser;
148 /// HasPHIUser - Set to true if this global has a user that is a PHI node.
151 /// AtomicOrdering - Set to the strongest atomic ordering requirement.
152 AtomicOrdering Ordering;
154 GlobalStatus() : isCompared(false), isLoaded(false), StoredType(NotStored),
155 StoredOnceValue(0), AccessingFunction(0),
156 HasMultipleAccessingFunctions(false),
157 HasNonInstructionUser(false), HasPHIUser(false),
158 Ordering(NotAtomic) {}
163 /// StrongerOrdering - Return the stronger of the two ordering. If the two
164 /// orderings are acquire and release, then return AcquireRelease.
166 static AtomicOrdering StrongerOrdering(AtomicOrdering X, AtomicOrdering Y) {
167 if (X == Acquire && Y == Release) return AcquireRelease;
168 if (Y == Acquire && X == Release) return AcquireRelease;
169 return (AtomicOrdering)std::max(X, Y);
172 /// SafeToDestroyConstant - It is safe to destroy a constant iff it is only used
173 /// by constants itself. Note that constants cannot be cyclic, so this test is
174 /// pretty easy to implement recursively.
176 static bool SafeToDestroyConstant(const Constant *C) {
177 if (isa<GlobalValue>(C)) return false;
179 for (Value::const_use_iterator UI = C->use_begin(), E = C->use_end(); UI != E;
181 if (const Constant *CU = dyn_cast<Constant>(*UI)) {
182 if (!SafeToDestroyConstant(CU)) return false;
189 /// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
190 /// structure. If the global has its address taken, return true to indicate we
191 /// can't do anything with it.
193 static bool AnalyzeGlobal(const Value *V, GlobalStatus &GS,
194 SmallPtrSet<const PHINode*, 16> &PHIUsers) {
195 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;
198 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
199 GS.HasNonInstructionUser = true;
201 // If the result of the constantexpr isn't pointer type, then we won't
202 // know to expect it in various places. Just reject early.
203 if (!isa<PointerType>(CE->getType())) return true;
205 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
206 } else if (const Instruction *I = dyn_cast<Instruction>(U)) {
207 if (!GS.HasMultipleAccessingFunctions) {
208 const Function *F = I->getParent()->getParent();
209 if (GS.AccessingFunction == 0)
210 GS.AccessingFunction = F;
211 else if (GS.AccessingFunction != F)
212 GS.HasMultipleAccessingFunctions = true;
214 if (const LoadInst *LI = dyn_cast<LoadInst>(I)) {
216 // Don't hack on volatile loads.
217 if (LI->isVolatile()) return true;
218 GS.Ordering = StrongerOrdering(GS.Ordering, LI->getOrdering());
219 } else if (const StoreInst *SI = dyn_cast<StoreInst>(I)) {
220 // Don't allow a store OF the address, only stores TO the address.
221 if (SI->getOperand(0) == V) return true;
223 // Don't hack on volatile stores.
224 if (SI->isVolatile()) return true;
225 GS.Ordering = StrongerOrdering(GS.Ordering, SI->getOrdering());
227 // If this is a direct store to the global (i.e., the global is a scalar
228 // value, not an aggregate), keep more specific information about
230 if (GS.StoredType != GlobalStatus::isStored) {
231 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(
232 SI->getOperand(1))) {
233 Value *StoredVal = SI->getOperand(0);
234 if (StoredVal == GV->getInitializer()) {
235 if (GS.StoredType < GlobalStatus::isInitializerStored)
236 GS.StoredType = GlobalStatus::isInitializerStored;
237 } else if (isa<LoadInst>(StoredVal) &&
238 cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
239 if (GS.StoredType < GlobalStatus::isInitializerStored)
240 GS.StoredType = GlobalStatus::isInitializerStored;
241 } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
242 GS.StoredType = GlobalStatus::isStoredOnce;
243 GS.StoredOnceValue = StoredVal;
244 } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
245 GS.StoredOnceValue == StoredVal) {
248 GS.StoredType = GlobalStatus::isStored;
251 GS.StoredType = GlobalStatus::isStored;
254 } else if (isa<GetElementPtrInst>(I)) {
255 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
256 } else if (isa<SelectInst>(I)) {
257 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
258 } else if (const PHINode *PN = dyn_cast<PHINode>(I)) {
259 // PHI nodes we can check just like select or GEP instructions, but we
260 // have to be careful about infinite recursion.
261 if (PHIUsers.insert(PN)) // Not already visited.
262 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
263 GS.HasPHIUser = true;
264 } else if (isa<CmpInst>(I)) {
265 GS.isCompared = true;
266 } else if (const MemTransferInst *MTI = dyn_cast<MemTransferInst>(I)) {
267 if (MTI->isVolatile()) return true;
268 if (MTI->getArgOperand(0) == V)
269 GS.StoredType = GlobalStatus::isStored;
270 if (MTI->getArgOperand(1) == V)
272 } else if (const MemSetInst *MSI = dyn_cast<MemSetInst>(I)) {
273 assert(MSI->getArgOperand(0) == V && "Memset only takes one pointer!");
274 if (MSI->isVolatile()) return true;
275 GS.StoredType = GlobalStatus::isStored;
277 return true; // Any other non-load instruction might take address!
279 } else if (const Constant *C = dyn_cast<Constant>(U)) {
280 GS.HasNonInstructionUser = true;
281 // We might have a dead and dangling constant hanging off of here.
282 if (!SafeToDestroyConstant(C))
285 GS.HasNonInstructionUser = true;
286 // Otherwise must be some other user.
294 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
295 /// users of the global, cleaning up the obvious ones. This is largely just a
296 /// quick scan over the use list to clean up the easy and obvious cruft. This
297 /// returns true if it made a change.
298 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
299 bool Changed = false;
300 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
303 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
305 // Replace the load with the initializer.
306 LI->replaceAllUsesWith(Init);
307 LI->eraseFromParent();
310 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
311 // Store must be unreachable or storing Init into the global.
312 SI->eraseFromParent();
314 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
315 if (CE->getOpcode() == Instruction::GetElementPtr) {
316 Constant *SubInit = 0;
318 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
319 Changed |= CleanupConstantGlobalUsers(CE, SubInit);
320 } else if (CE->getOpcode() == Instruction::BitCast &&
321 CE->getType()->isPointerTy()) {
322 // Pointer cast, delete any stores and memsets to the global.
323 Changed |= CleanupConstantGlobalUsers(CE, 0);
326 if (CE->use_empty()) {
327 CE->destroyConstant();
330 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
331 // Do not transform "gepinst (gep constexpr (GV))" here, because forming
332 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
333 // and will invalidate our notion of what Init is.
334 Constant *SubInit = 0;
335 if (!isa<ConstantExpr>(GEP->getOperand(0))) {
336 // FIXME: use TargetData/TargetLibraryInfo for smarter constant folding.
338 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
339 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
340 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
342 Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
344 if (GEP->use_empty()) {
345 GEP->eraseFromParent();
348 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
349 if (MI->getRawDest() == V) {
350 MI->eraseFromParent();
354 } else if (Constant *C = dyn_cast<Constant>(U)) {
355 // If we have a chain of dead constantexprs or other things dangling from
356 // us, and if they are all dead, nuke them without remorse.
357 if (SafeToDestroyConstant(C)) {
358 C->destroyConstant();
359 // This could have invalidated UI, start over from scratch.
360 CleanupConstantGlobalUsers(V, Init);
368 /// isSafeSROAElementUse - Return true if the specified instruction is a safe
369 /// user of a derived expression from a global that we want to SROA.
370 static bool isSafeSROAElementUse(Value *V) {
371 // We might have a dead and dangling constant hanging off of here.
372 if (Constant *C = dyn_cast<Constant>(V))
373 return SafeToDestroyConstant(C);
375 Instruction *I = dyn_cast<Instruction>(V);
376 if (!I) return false;
379 if (isa<LoadInst>(I)) return true;
381 // Stores *to* the pointer are ok.
382 if (StoreInst *SI = dyn_cast<StoreInst>(I))
383 return SI->getOperand(0) != V;
385 // Otherwise, it must be a GEP.
386 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I);
387 if (GEPI == 0) return false;
389 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
390 !cast<Constant>(GEPI->getOperand(1))->isNullValue())
393 for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end();
395 if (!isSafeSROAElementUse(*I))
401 /// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value.
402 /// Look at it and its uses and decide whether it is safe to SROA this global.
404 static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
405 // The user of the global must be a GEP Inst or a ConstantExpr GEP.
406 if (!isa<GetElementPtrInst>(U) &&
407 (!isa<ConstantExpr>(U) ||
408 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
411 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
412 // don't like < 3 operand CE's, and we don't like non-constant integer
413 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some
415 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
416 !cast<Constant>(U->getOperand(1))->isNullValue() ||
417 !isa<ConstantInt>(U->getOperand(2)))
420 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
421 ++GEPI; // Skip over the pointer index.
423 // If this is a use of an array allocation, do a bit more checking for sanity.
424 if (ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
425 uint64_t NumElements = AT->getNumElements();
426 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
428 // Check to make sure that index falls within the array. If not,
429 // something funny is going on, so we won't do the optimization.
431 if (Idx->getZExtValue() >= NumElements)
434 // We cannot scalar repl this level of the array unless any array
435 // sub-indices are in-range constants. In particular, consider:
436 // A[0][i]. We cannot know that the user isn't doing invalid things like
437 // allowing i to index an out-of-range subscript that accesses A[1].
439 // Scalar replacing *just* the outer index of the array is probably not
440 // going to be a win anyway, so just give up.
441 for (++GEPI; // Skip array index.
444 uint64_t NumElements;
445 if (ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
446 NumElements = SubArrayTy->getNumElements();
447 else if (VectorType *SubVectorTy = dyn_cast<VectorType>(*GEPI))
448 NumElements = SubVectorTy->getNumElements();
450 assert((*GEPI)->isStructTy() &&
451 "Indexed GEP type is not array, vector, or struct!");
455 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
456 if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
461 for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I)
462 if (!isSafeSROAElementUse(*I))
467 /// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
468 /// is safe for us to perform this transformation.
470 static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
471 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
473 if (!IsUserOfGlobalSafeForSRA(*UI, GV))
480 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
481 /// variable. This opens the door for other optimizations by exposing the
482 /// behavior of the program in a more fine-grained way. We have determined that
483 /// this transformation is safe already. We return the first global variable we
484 /// insert so that the caller can reprocess it.
485 static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) {
486 // Make sure this global only has simple uses that we can SRA.
487 if (!GlobalUsersSafeToSRA(GV))
490 assert(GV->hasLocalLinkage() && !GV->isConstant());
491 Constant *Init = GV->getInitializer();
492 Type *Ty = Init->getType();
494 std::vector<GlobalVariable*> NewGlobals;
495 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
497 // Get the alignment of the global, either explicit or target-specific.
498 unsigned StartAlignment = GV->getAlignment();
499 if (StartAlignment == 0)
500 StartAlignment = TD.getABITypeAlignment(GV->getType());
502 if (StructType *STy = dyn_cast<StructType>(Ty)) {
503 NewGlobals.reserve(STy->getNumElements());
504 const StructLayout &Layout = *TD.getStructLayout(STy);
505 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
506 Constant *In = Init->getAggregateElement(i);
507 assert(In && "Couldn't get element of initializer?");
508 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
509 GlobalVariable::InternalLinkage,
510 In, GV->getName()+"."+Twine(i),
512 GV->getType()->getAddressSpace());
513 Globals.insert(GV, NGV);
514 NewGlobals.push_back(NGV);
516 // Calculate the known alignment of the field. If the original aggregate
517 // had 256 byte alignment for example, something might depend on that:
518 // propagate info to each field.
519 uint64_t FieldOffset = Layout.getElementOffset(i);
520 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset);
521 if (NewAlign > TD.getABITypeAlignment(STy->getElementType(i)))
522 NGV->setAlignment(NewAlign);
524 } else if (SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
525 unsigned NumElements = 0;
526 if (ArrayType *ATy = dyn_cast<ArrayType>(STy))
527 NumElements = ATy->getNumElements();
529 NumElements = cast<VectorType>(STy)->getNumElements();
531 if (NumElements > 16 && GV->hasNUsesOrMore(16))
532 return 0; // It's not worth it.
533 NewGlobals.reserve(NumElements);
535 uint64_t EltSize = TD.getTypeAllocSize(STy->getElementType());
536 unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType());
537 for (unsigned i = 0, e = NumElements; i != e; ++i) {
538 Constant *In = Init->getAggregateElement(i);
539 assert(In && "Couldn't get element of initializer?");
541 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
542 GlobalVariable::InternalLinkage,
543 In, GV->getName()+"."+Twine(i),
545 GV->getType()->getAddressSpace());
546 Globals.insert(GV, NGV);
547 NewGlobals.push_back(NGV);
549 // Calculate the known alignment of the field. If the original aggregate
550 // had 256 byte alignment for example, something might depend on that:
551 // propagate info to each field.
552 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
553 if (NewAlign > EltAlign)
554 NGV->setAlignment(NewAlign);
558 if (NewGlobals.empty())
561 DEBUG(dbgs() << "PERFORMING GLOBAL SRA ON: " << *GV);
563 Constant *NullInt =Constant::getNullValue(Type::getInt32Ty(GV->getContext()));
565 // Loop over all of the uses of the global, replacing the constantexpr geps,
566 // with smaller constantexpr geps or direct references.
567 while (!GV->use_empty()) {
568 User *GEP = GV->use_back();
569 assert(((isa<ConstantExpr>(GEP) &&
570 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
571 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
573 // Ignore the 1th operand, which has to be zero or else the program is quite
574 // broken (undefined). Get the 2nd operand, which is the structure or array
576 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
577 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
579 Value *NewPtr = NewGlobals[Val];
581 // Form a shorter GEP if needed.
582 if (GEP->getNumOperands() > 3) {
583 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
584 SmallVector<Constant*, 8> Idxs;
585 Idxs.push_back(NullInt);
586 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
587 Idxs.push_back(CE->getOperand(i));
588 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr), Idxs);
590 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
591 SmallVector<Value*, 8> Idxs;
592 Idxs.push_back(NullInt);
593 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
594 Idxs.push_back(GEPI->getOperand(i));
595 NewPtr = GetElementPtrInst::Create(NewPtr, Idxs,
596 GEPI->getName()+"."+Twine(Val),GEPI);
599 GEP->replaceAllUsesWith(NewPtr);
601 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
602 GEPI->eraseFromParent();
604 cast<ConstantExpr>(GEP)->destroyConstant();
607 // Delete the old global, now that it is dead.
611 // Loop over the new globals array deleting any globals that are obviously
612 // dead. This can arise due to scalarization of a structure or an array that
613 // has elements that are dead.
614 unsigned FirstGlobal = 0;
615 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
616 if (NewGlobals[i]->use_empty()) {
617 Globals.erase(NewGlobals[i]);
618 if (FirstGlobal == i) ++FirstGlobal;
621 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
624 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
625 /// value will trap if the value is dynamically null. PHIs keeps track of any
626 /// phi nodes we've seen to avoid reprocessing them.
627 static bool AllUsesOfValueWillTrapIfNull(const Value *V,
628 SmallPtrSet<const PHINode*, 8> &PHIs) {
629 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;
633 if (isa<LoadInst>(U)) {
635 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
636 if (SI->getOperand(0) == V) {
637 //cerr << "NONTRAPPING USE: " << *U;
638 return false; // Storing the value.
640 } else if (const CallInst *CI = dyn_cast<CallInst>(U)) {
641 if (CI->getCalledValue() != V) {
642 //cerr << "NONTRAPPING USE: " << *U;
643 return false; // Not calling the ptr
645 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(U)) {
646 if (II->getCalledValue() != V) {
647 //cerr << "NONTRAPPING USE: " << *U;
648 return false; // Not calling the ptr
650 } else if (const BitCastInst *CI = dyn_cast<BitCastInst>(U)) {
651 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
652 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
653 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
654 } else if (const PHINode *PN = dyn_cast<PHINode>(U)) {
655 // If we've already seen this phi node, ignore it, it has already been
657 if (PHIs.insert(PN) && !AllUsesOfValueWillTrapIfNull(PN, PHIs))
659 } else if (isa<ICmpInst>(U) &&
660 isa<ConstantPointerNull>(UI->getOperand(1))) {
661 // Ignore icmp X, null
663 //cerr << "NONTRAPPING USE: " << *U;
670 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
671 /// from GV will trap if the loaded value is null. Note that this also permits
672 /// comparisons of the loaded value against null, as a special case.
673 static bool AllUsesOfLoadedValueWillTrapIfNull(const GlobalVariable *GV) {
674 for (Value::const_use_iterator UI = GV->use_begin(), E = GV->use_end();
678 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
679 SmallPtrSet<const PHINode*, 8> PHIs;
680 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
682 } else if (isa<StoreInst>(U)) {
683 // Ignore stores to the global.
685 // We don't know or understand this user, bail out.
686 //cerr << "UNKNOWN USER OF GLOBAL!: " << *U;
693 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
694 bool Changed = false;
695 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
696 Instruction *I = cast<Instruction>(*UI++);
697 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
698 LI->setOperand(0, NewV);
700 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
701 if (SI->getOperand(1) == V) {
702 SI->setOperand(1, NewV);
705 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
707 if (CS.getCalledValue() == V) {
708 // Calling through the pointer! Turn into a direct call, but be careful
709 // that the pointer is not also being passed as an argument.
710 CS.setCalledFunction(NewV);
712 bool PassedAsArg = false;
713 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
714 if (CS.getArgument(i) == V) {
716 CS.setArgument(i, NewV);
720 // Being passed as an argument also. Be careful to not invalidate UI!
724 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
725 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
726 ConstantExpr::getCast(CI->getOpcode(),
727 NewV, CI->getType()));
728 if (CI->use_empty()) {
730 CI->eraseFromParent();
732 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
733 // Should handle GEP here.
734 SmallVector<Constant*, 8> Idxs;
735 Idxs.reserve(GEPI->getNumOperands()-1);
736 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
738 if (Constant *C = dyn_cast<Constant>(*i))
742 if (Idxs.size() == GEPI->getNumOperands()-1)
743 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
744 ConstantExpr::getGetElementPtr(NewV, Idxs));
745 if (GEPI->use_empty()) {
747 GEPI->eraseFromParent();
756 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
757 /// value stored into it. If there are uses of the loaded value that would trap
758 /// if the loaded value is dynamically null, then we know that they cannot be
759 /// reachable with a null optimize away the load.
760 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) {
761 bool Changed = false;
763 // Keep track of whether we are able to remove all the uses of the global
764 // other than the store that defines it.
765 bool AllNonStoreUsesGone = true;
767 // Replace all uses of loads with uses of uses of the stored value.
768 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end(); GUI != E;){
769 User *GlobalUser = *GUI++;
770 if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) {
771 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
772 // If we were able to delete all uses of the loads
773 if (LI->use_empty()) {
774 LI->eraseFromParent();
777 AllNonStoreUsesGone = false;
779 } else if (isa<StoreInst>(GlobalUser)) {
780 // Ignore the store that stores "LV" to the global.
781 assert(GlobalUser->getOperand(1) == GV &&
782 "Must be storing *to* the global");
784 AllNonStoreUsesGone = false;
786 // If we get here we could have other crazy uses that are transitively
788 assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) ||
789 isa<ConstantExpr>(GlobalUser) || isa<CmpInst>(GlobalUser)) &&
790 "Only expect load and stores!");
795 DEBUG(dbgs() << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV);
799 // If we nuked all of the loads, then none of the stores are needed either,
800 // nor is the global.
801 if (AllNonStoreUsesGone) {
802 DEBUG(dbgs() << " *** GLOBAL NOW DEAD!\n");
803 CleanupConstantGlobalUsers(GV, 0);
804 if (GV->use_empty()) {
805 GV->eraseFromParent();
813 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
814 /// instructions that are foldable.
815 static void ConstantPropUsersOf(Value *V) {
816 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
817 if (Instruction *I = dyn_cast<Instruction>(*UI++))
818 // FIXME: use TargetData/TargetLibraryInfo for smarter constant folding.
819 if (Constant *NewC = ConstantFoldInstruction(I)) {
820 I->replaceAllUsesWith(NewC);
822 // Advance UI to the next non-I use to avoid invalidating it!
823 // Instructions could multiply use V.
824 while (UI != E && *UI == I)
826 I->eraseFromParent();
830 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
831 /// variable, and transforms the program as if it always contained the result of
832 /// the specified malloc. Because it is always the result of the specified
833 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
834 /// malloc into a global, and any loads of GV as uses of the new global.
835 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
838 ConstantInt *NElements,
840 DEBUG(errs() << "PROMOTING GLOBAL: " << *GV << " CALL = " << *CI << '\n');
843 if (NElements->getZExtValue() == 1)
844 GlobalType = AllocTy;
846 // If we have an array allocation, the global variable is of an array.
847 GlobalType = ArrayType::get(AllocTy, NElements->getZExtValue());
849 // Create the new global variable. The contents of the malloc'd memory is
850 // undefined, so initialize with an undef value.
851 GlobalVariable *NewGV = new GlobalVariable(*GV->getParent(),
853 GlobalValue::InternalLinkage,
854 UndefValue::get(GlobalType),
855 GV->getName()+".body",
857 GV->isThreadLocal());
859 // If there are bitcast users of the malloc (which is typical, usually we have
860 // a malloc + bitcast) then replace them with uses of the new global. Update
861 // other users to use the global as well.
862 BitCastInst *TheBC = 0;
863 while (!CI->use_empty()) {
864 Instruction *User = cast<Instruction>(CI->use_back());
865 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
866 if (BCI->getType() == NewGV->getType()) {
867 BCI->replaceAllUsesWith(NewGV);
868 BCI->eraseFromParent();
870 BCI->setOperand(0, NewGV);
874 TheBC = new BitCastInst(NewGV, CI->getType(), "newgv", CI);
875 User->replaceUsesOfWith(CI, TheBC);
879 Constant *RepValue = NewGV;
880 if (NewGV->getType() != GV->getType()->getElementType())
881 RepValue = ConstantExpr::getBitCast(RepValue,
882 GV->getType()->getElementType());
884 // If there is a comparison against null, we will insert a global bool to
885 // keep track of whether the global was initialized yet or not.
886 GlobalVariable *InitBool =
887 new GlobalVariable(Type::getInt1Ty(GV->getContext()), false,
888 GlobalValue::InternalLinkage,
889 ConstantInt::getFalse(GV->getContext()),
890 GV->getName()+".init", GV->isThreadLocal());
891 bool InitBoolUsed = false;
893 // Loop over all uses of GV, processing them in turn.
894 while (!GV->use_empty()) {
895 if (StoreInst *SI = dyn_cast<StoreInst>(GV->use_back())) {
896 // The global is initialized when the store to it occurs.
897 new StoreInst(ConstantInt::getTrue(GV->getContext()), InitBool, false, 0,
898 SI->getOrdering(), SI->getSynchScope(), SI);
899 SI->eraseFromParent();
903 LoadInst *LI = cast<LoadInst>(GV->use_back());
904 while (!LI->use_empty()) {
905 Use &LoadUse = LI->use_begin().getUse();
906 if (!isa<ICmpInst>(LoadUse.getUser())) {
911 ICmpInst *ICI = cast<ICmpInst>(LoadUse.getUser());
912 // Replace the cmp X, 0 with a use of the bool value.
913 // Sink the load to where the compare was, if atomic rules allow us to.
914 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", false, 0,
915 LI->getOrdering(), LI->getSynchScope(),
916 LI->isUnordered() ? (Instruction*)ICI : LI);
918 switch (ICI->getPredicate()) {
919 default: llvm_unreachable("Unknown ICmp Predicate!");
920 case ICmpInst::ICMP_ULT:
921 case ICmpInst::ICMP_SLT: // X < null -> always false
922 LV = ConstantInt::getFalse(GV->getContext());
924 case ICmpInst::ICMP_ULE:
925 case ICmpInst::ICMP_SLE:
926 case ICmpInst::ICMP_EQ:
927 LV = BinaryOperator::CreateNot(LV, "notinit", ICI);
929 case ICmpInst::ICMP_NE:
930 case ICmpInst::ICMP_UGE:
931 case ICmpInst::ICMP_SGE:
932 case ICmpInst::ICMP_UGT:
933 case ICmpInst::ICMP_SGT:
936 ICI->replaceAllUsesWith(LV);
937 ICI->eraseFromParent();
939 LI->eraseFromParent();
942 // If the initialization boolean was used, insert it, otherwise delete it.
944 while (!InitBool->use_empty()) // Delete initializations
945 cast<StoreInst>(InitBool->use_back())->eraseFromParent();
948 GV->getParent()->getGlobalList().insert(GV, InitBool);
950 // Now the GV is dead, nuke it and the malloc..
951 GV->eraseFromParent();
952 CI->eraseFromParent();
954 // To further other optimizations, loop over all users of NewGV and try to
955 // constant prop them. This will promote GEP instructions with constant
956 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
957 ConstantPropUsersOf(NewGV);
958 if (RepValue != NewGV)
959 ConstantPropUsersOf(RepValue);
964 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
965 /// to make sure that there are no complex uses of V. We permit simple things
966 /// like dereferencing the pointer, but not storing through the address, unless
967 /// it is to the specified global.
968 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(const Instruction *V,
969 const GlobalVariable *GV,
970 SmallPtrSet<const PHINode*, 8> &PHIs) {
971 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end();
973 const Instruction *Inst = cast<Instruction>(*UI);
975 if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) {
976 continue; // Fine, ignore.
979 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
980 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
981 return false; // Storing the pointer itself... bad.
982 continue; // Otherwise, storing through it, or storing into GV... fine.
985 // Must index into the array and into the struct.
986 if (isa<GetElementPtrInst>(Inst) && Inst->getNumOperands() >= 3) {
987 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs))
992 if (const PHINode *PN = dyn_cast<PHINode>(Inst)) {
993 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI
996 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
1001 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) {
1002 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs))
1012 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
1013 /// somewhere. Transform all uses of the allocation into loads from the
1014 /// global and uses of the resultant pointer. Further, delete the store into
1015 /// GV. This assumes that these value pass the
1016 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
1017 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
1018 GlobalVariable *GV) {
1019 while (!Alloc->use_empty()) {
1020 Instruction *U = cast<Instruction>(*Alloc->use_begin());
1021 Instruction *InsertPt = U;
1022 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1023 // If this is the store of the allocation into the global, remove it.
1024 if (SI->getOperand(1) == GV) {
1025 SI->eraseFromParent();
1028 } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
1029 // Insert the load in the corresponding predecessor, not right before the
1031 InsertPt = PN->getIncomingBlock(Alloc->use_begin())->getTerminator();
1032 } else if (isa<BitCastInst>(U)) {
1033 // Must be bitcast between the malloc and store to initialize the global.
1034 ReplaceUsesOfMallocWithGlobal(U, GV);
1035 U->eraseFromParent();
1037 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
1038 // If this is a "GEP bitcast" and the user is a store to the global, then
1039 // just process it as a bitcast.
1040 if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse())
1041 if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->use_back()))
1042 if (SI->getOperand(1) == GV) {
1043 // Must be bitcast GEP between the malloc and store to initialize
1045 ReplaceUsesOfMallocWithGlobal(GEPI, GV);
1046 GEPI->eraseFromParent();
1051 // Insert a load from the global, and use it instead of the malloc.
1052 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
1053 U->replaceUsesOfWith(Alloc, NL);
1057 /// LoadUsesSimpleEnoughForHeapSRA - Verify that all uses of V (a load, or a phi
1058 /// of a load) are simple enough to perform heap SRA on. This permits GEP's
1059 /// that index through the array and struct field, icmps of null, and PHIs.
1060 static bool LoadUsesSimpleEnoughForHeapSRA(const Value *V,
1061 SmallPtrSet<const PHINode*, 32> &LoadUsingPHIs,
1062 SmallPtrSet<const PHINode*, 32> &LoadUsingPHIsPerLoad) {
1063 // We permit two users of the load: setcc comparing against the null
1064 // pointer, and a getelementptr of a specific form.
1065 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;
1067 const Instruction *User = cast<Instruction>(*UI);
1069 // Comparison against null is ok.
1070 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(User)) {
1071 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1076 // getelementptr is also ok, but only a simple form.
1077 if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1078 // Must index into the array and into the struct.
1079 if (GEPI->getNumOperands() < 3)
1082 // Otherwise the GEP is ok.
1086 if (const PHINode *PN = dyn_cast<PHINode>(User)) {
1087 if (!LoadUsingPHIsPerLoad.insert(PN))
1088 // This means some phi nodes are dependent on each other.
1089 // Avoid infinite looping!
1091 if (!LoadUsingPHIs.insert(PN))
1092 // If we have already analyzed this PHI, then it is safe.
1095 // Make sure all uses of the PHI are simple enough to transform.
1096 if (!LoadUsesSimpleEnoughForHeapSRA(PN,
1097 LoadUsingPHIs, LoadUsingPHIsPerLoad))
1103 // Otherwise we don't know what this is, not ok.
1111 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
1112 /// GV are simple enough to perform HeapSRA, return true.
1113 static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(const GlobalVariable *GV,
1114 Instruction *StoredVal) {
1115 SmallPtrSet<const PHINode*, 32> LoadUsingPHIs;
1116 SmallPtrSet<const PHINode*, 32> LoadUsingPHIsPerLoad;
1117 for (Value::const_use_iterator UI = GV->use_begin(), E = GV->use_end();
1119 if (const LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
1120 if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs,
1121 LoadUsingPHIsPerLoad))
1123 LoadUsingPHIsPerLoad.clear();
1126 // If we reach here, we know that all uses of the loads and transitive uses
1127 // (through PHI nodes) are simple enough to transform. However, we don't know
1128 // that all inputs the to the PHI nodes are in the same equivalence sets.
1129 // Check to verify that all operands of the PHIs are either PHIS that can be
1130 // transformed, loads from GV, or MI itself.
1131 for (SmallPtrSet<const PHINode*, 32>::const_iterator I = LoadUsingPHIs.begin()
1132 , E = LoadUsingPHIs.end(); I != E; ++I) {
1133 const PHINode *PN = *I;
1134 for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) {
1135 Value *InVal = PN->getIncomingValue(op);
1137 // PHI of the stored value itself is ok.
1138 if (InVal == StoredVal) continue;
1140 if (const PHINode *InPN = dyn_cast<PHINode>(InVal)) {
1141 // One of the PHIs in our set is (optimistically) ok.
1142 if (LoadUsingPHIs.count(InPN))
1147 // Load from GV is ok.
1148 if (const LoadInst *LI = dyn_cast<LoadInst>(InVal))
1149 if (LI->getOperand(0) == GV)
1154 // Anything else is rejected.
1162 static Value *GetHeapSROAValue(Value *V, unsigned FieldNo,
1163 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1164 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1165 std::vector<Value*> &FieldVals = InsertedScalarizedValues[V];
1167 if (FieldNo >= FieldVals.size())
1168 FieldVals.resize(FieldNo+1);
1170 // If we already have this value, just reuse the previously scalarized
1172 if (Value *FieldVal = FieldVals[FieldNo])
1175 // Depending on what instruction this is, we have several cases.
1177 if (LoadInst *LI = dyn_cast<LoadInst>(V)) {
1178 // This is a scalarized version of the load from the global. Just create
1179 // a new Load of the scalarized global.
1180 Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo,
1181 InsertedScalarizedValues,
1183 LI->getName()+".f"+Twine(FieldNo), LI);
1184 } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1185 // PN's type is pointer to struct. Make a new PHI of pointer to struct
1188 cast<StructType>(cast<PointerType>(PN->getType())->getElementType());
1191 PHINode::Create(PointerType::getUnqual(ST->getElementType(FieldNo)),
1192 PN->getNumIncomingValues(),
1193 PN->getName()+".f"+Twine(FieldNo), PN);
1195 PHIsToRewrite.push_back(std::make_pair(PN, FieldNo));
1197 llvm_unreachable("Unknown usable value");
1200 return FieldVals[FieldNo] = Result;
1203 /// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1204 /// the load, rewrite the derived value to use the HeapSRoA'd load.
1205 static void RewriteHeapSROALoadUser(Instruction *LoadUser,
1206 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1207 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1208 // If this is a comparison against null, handle it.
1209 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1210 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1211 // If we have a setcc of the loaded pointer, we can use a setcc of any
1213 Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0,
1214 InsertedScalarizedValues, PHIsToRewrite);
1216 Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr,
1217 Constant::getNullValue(NPtr->getType()),
1219 SCI->replaceAllUsesWith(New);
1220 SCI->eraseFromParent();
1224 // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...'
1225 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1226 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1227 && "Unexpected GEPI!");
1229 // Load the pointer for this field.
1230 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1231 Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo,
1232 InsertedScalarizedValues, PHIsToRewrite);
1234 // Create the new GEP idx vector.
1235 SmallVector<Value*, 8> GEPIdx;
1236 GEPIdx.push_back(GEPI->getOperand(1));
1237 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1239 Value *NGEPI = GetElementPtrInst::Create(NewPtr, GEPIdx,
1240 GEPI->getName(), GEPI);
1241 GEPI->replaceAllUsesWith(NGEPI);
1242 GEPI->eraseFromParent();
1246 // Recursively transform the users of PHI nodes. This will lazily create the
1247 // PHIs that are needed for individual elements. Keep track of what PHIs we
1248 // see in InsertedScalarizedValues so that we don't get infinite loops (very
1249 // antisocial). If the PHI is already in InsertedScalarizedValues, it has
1250 // already been seen first by another load, so its uses have already been
1252 PHINode *PN = cast<PHINode>(LoadUser);
1253 if (!InsertedScalarizedValues.insert(std::make_pair(PN,
1254 std::vector<Value*>())).second)
1257 // If this is the first time we've seen this PHI, recursively process all
1259 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); UI != E; ) {
1260 Instruction *User = cast<Instruction>(*UI++);
1261 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1265 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1266 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1267 /// use FieldGlobals instead. All uses of loaded values satisfy
1268 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA.
1269 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1270 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1271 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1272 for (Value::use_iterator UI = Load->use_begin(), E = Load->use_end();
1274 Instruction *User = cast<Instruction>(*UI++);
1275 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1278 if (Load->use_empty()) {
1279 Load->eraseFromParent();
1280 InsertedScalarizedValues.erase(Load);
1284 /// PerformHeapAllocSRoA - CI is an allocation of an array of structures. Break
1285 /// it up into multiple allocations of arrays of the fields.
1286 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI,
1287 Value *NElems, TargetData *TD) {
1288 DEBUG(dbgs() << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *CI << '\n');
1289 Type *MAT = getMallocAllocatedType(CI);
1290 StructType *STy = cast<StructType>(MAT);
1292 // There is guaranteed to be at least one use of the malloc (storing
1293 // it into GV). If there are other uses, change them to be uses of
1294 // the global to simplify later code. This also deletes the store
1296 ReplaceUsesOfMallocWithGlobal(CI, GV);
1298 // Okay, at this point, there are no users of the malloc. Insert N
1299 // new mallocs at the same place as CI, and N globals.
1300 std::vector<Value*> FieldGlobals;
1301 std::vector<Value*> FieldMallocs;
1303 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1304 Type *FieldTy = STy->getElementType(FieldNo);
1305 PointerType *PFieldTy = PointerType::getUnqual(FieldTy);
1307 GlobalVariable *NGV =
1308 new GlobalVariable(*GV->getParent(),
1309 PFieldTy, false, GlobalValue::InternalLinkage,
1310 Constant::getNullValue(PFieldTy),
1311 GV->getName() + ".f" + Twine(FieldNo), GV,
1312 GV->isThreadLocal());
1313 FieldGlobals.push_back(NGV);
1315 unsigned TypeSize = TD->getTypeAllocSize(FieldTy);
1316 if (StructType *ST = dyn_cast<StructType>(FieldTy))
1317 TypeSize = TD->getStructLayout(ST)->getSizeInBytes();
1318 Type *IntPtrTy = TD->getIntPtrType(CI->getContext());
1319 Value *NMI = CallInst::CreateMalloc(CI, IntPtrTy, FieldTy,
1320 ConstantInt::get(IntPtrTy, TypeSize),
1322 CI->getName() + ".f" + Twine(FieldNo));
1323 FieldMallocs.push_back(NMI);
1324 new StoreInst(NMI, NGV, CI);
1327 // The tricky aspect of this transformation is handling the case when malloc
1328 // fails. In the original code, malloc failing would set the result pointer
1329 // of malloc to null. In this case, some mallocs could succeed and others
1330 // could fail. As such, we emit code that looks like this:
1331 // F0 = malloc(field0)
1332 // F1 = malloc(field1)
1333 // F2 = malloc(field2)
1334 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1335 // if (F0) { free(F0); F0 = 0; }
1336 // if (F1) { free(F1); F1 = 0; }
1337 // if (F2) { free(F2); F2 = 0; }
1339 // The malloc can also fail if its argument is too large.
1340 Constant *ConstantZero = ConstantInt::get(CI->getArgOperand(0)->getType(), 0);
1341 Value *RunningOr = new ICmpInst(CI, ICmpInst::ICMP_SLT, CI->getArgOperand(0),
1342 ConstantZero, "isneg");
1343 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1344 Value *Cond = new ICmpInst(CI, ICmpInst::ICMP_EQ, FieldMallocs[i],
1345 Constant::getNullValue(FieldMallocs[i]->getType()),
1347 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", CI);
1350 // Split the basic block at the old malloc.
1351 BasicBlock *OrigBB = CI->getParent();
1352 BasicBlock *ContBB = OrigBB->splitBasicBlock(CI, "malloc_cont");
1354 // Create the block to check the first condition. Put all these blocks at the
1355 // end of the function as they are unlikely to be executed.
1356 BasicBlock *NullPtrBlock = BasicBlock::Create(OrigBB->getContext(),
1358 OrigBB->getParent());
1360 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1361 // branch on RunningOr.
1362 OrigBB->getTerminator()->eraseFromParent();
1363 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1365 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1366 // pointer, because some may be null while others are not.
1367 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1368 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1369 Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
1370 Constant::getNullValue(GVVal->getType()));
1371 BasicBlock *FreeBlock = BasicBlock::Create(Cmp->getContext(), "free_it",
1372 OrigBB->getParent());
1373 BasicBlock *NextBlock = BasicBlock::Create(Cmp->getContext(), "next",
1374 OrigBB->getParent());
1375 Instruction *BI = BranchInst::Create(FreeBlock, NextBlock,
1378 // Fill in FreeBlock.
1379 CallInst::CreateFree(GVVal, BI);
1380 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1382 BranchInst::Create(NextBlock, FreeBlock);
1384 NullPtrBlock = NextBlock;
1387 BranchInst::Create(ContBB, NullPtrBlock);
1389 // CI is no longer needed, remove it.
1390 CI->eraseFromParent();
1392 /// InsertedScalarizedLoads - As we process loads, if we can't immediately
1393 /// update all uses of the load, keep track of what scalarized loads are
1394 /// inserted for a given load.
1395 DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues;
1396 InsertedScalarizedValues[GV] = FieldGlobals;
1398 std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite;
1400 // Okay, the malloc site is completely handled. All of the uses of GV are now
1401 // loads, and all uses of those loads are simple. Rewrite them to use loads
1402 // of the per-field globals instead.
1403 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;) {
1404 Instruction *User = cast<Instruction>(*UI++);
1406 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1407 RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite);
1411 // Must be a store of null.
1412 StoreInst *SI = cast<StoreInst>(User);
1413 assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
1414 "Unexpected heap-sra user!");
1416 // Insert a store of null into each global.
1417 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1418 PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType());
1419 Constant *Null = Constant::getNullValue(PT->getElementType());
1420 new StoreInst(Null, FieldGlobals[i], SI);
1422 // Erase the original store.
1423 SI->eraseFromParent();
1426 // While we have PHIs that are interesting to rewrite, do it.
1427 while (!PHIsToRewrite.empty()) {
1428 PHINode *PN = PHIsToRewrite.back().first;
1429 unsigned FieldNo = PHIsToRewrite.back().second;
1430 PHIsToRewrite.pop_back();
1431 PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]);
1432 assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi");
1434 // Add all the incoming values. This can materialize more phis.
1435 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1436 Value *InVal = PN->getIncomingValue(i);
1437 InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues,
1439 FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
1443 // Drop all inter-phi links and any loads that made it this far.
1444 for (DenseMap<Value*, std::vector<Value*> >::iterator
1445 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1447 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1448 PN->dropAllReferences();
1449 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1450 LI->dropAllReferences();
1453 // Delete all the phis and loads now that inter-references are dead.
1454 for (DenseMap<Value*, std::vector<Value*> >::iterator
1455 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1457 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1458 PN->eraseFromParent();
1459 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1460 LI->eraseFromParent();
1463 // The old global is now dead, remove it.
1464 GV->eraseFromParent();
1467 return cast<GlobalVariable>(FieldGlobals[0]);
1470 /// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1471 /// pointer global variable with a single value stored it that is a malloc or
1473 static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
1476 AtomicOrdering Ordering,
1477 Module::global_iterator &GVI,
1482 // If this is a malloc of an abstract type, don't touch it.
1483 if (!AllocTy->isSized())
1486 // We can't optimize this global unless all uses of it are *known* to be
1487 // of the malloc value, not of the null initializer value (consider a use
1488 // that compares the global's value against zero to see if the malloc has
1489 // been reached). To do this, we check to see if all uses of the global
1490 // would trap if the global were null: this proves that they must all
1491 // happen after the malloc.
1492 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1495 // We can't optimize this if the malloc itself is used in a complex way,
1496 // for example, being stored into multiple globals. This allows the
1497 // malloc to be stored into the specified global, loaded icmp'd, and
1498 // GEP'd. These are all things we could transform to using the global
1500 SmallPtrSet<const PHINode*, 8> PHIs;
1501 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV, PHIs))
1504 // If we have a global that is only initialized with a fixed size malloc,
1505 // transform the program to use global memory instead of malloc'd memory.
1506 // This eliminates dynamic allocation, avoids an indirection accessing the
1507 // data, and exposes the resultant global to further GlobalOpt.
1508 // We cannot optimize the malloc if we cannot determine malloc array size.
1509 Value *NElems = getMallocArraySize(CI, TD, true);
1513 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
1514 // Restrict this transformation to only working on small allocations
1515 // (2048 bytes currently), as we don't want to introduce a 16M global or
1517 if (NElements->getZExtValue() * TD->getTypeAllocSize(AllocTy) < 2048) {
1518 GVI = OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, TD);
1522 // If the allocation is an array of structures, consider transforming this
1523 // into multiple malloc'd arrays, one for each field. This is basically
1524 // SRoA for malloc'd memory.
1526 if (Ordering != NotAtomic)
1529 // If this is an allocation of a fixed size array of structs, analyze as a
1530 // variable size array. malloc [100 x struct],1 -> malloc struct, 100
1531 if (NElems == ConstantInt::get(CI->getArgOperand(0)->getType(), 1))
1532 if (ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
1533 AllocTy = AT->getElementType();
1535 StructType *AllocSTy = dyn_cast<StructType>(AllocTy);
1539 // This the structure has an unreasonable number of fields, leave it
1541 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
1542 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, CI)) {
1544 // If this is a fixed size array, transform the Malloc to be an alloc of
1545 // structs. malloc [100 x struct],1 -> malloc struct, 100
1546 if (ArrayType *AT = dyn_cast<ArrayType>(getMallocAllocatedType(CI))) {
1547 Type *IntPtrTy = TD->getIntPtrType(CI->getContext());
1548 unsigned TypeSize = TD->getStructLayout(AllocSTy)->getSizeInBytes();
1549 Value *AllocSize = ConstantInt::get(IntPtrTy, TypeSize);
1550 Value *NumElements = ConstantInt::get(IntPtrTy, AT->getNumElements());
1551 Instruction *Malloc = CallInst::CreateMalloc(CI, IntPtrTy, AllocSTy,
1552 AllocSize, NumElements,
1554 Instruction *Cast = new BitCastInst(Malloc, CI->getType(), "tmp", CI);
1555 CI->replaceAllUsesWith(Cast);
1556 CI->eraseFromParent();
1557 CI = dyn_cast<BitCastInst>(Malloc) ?
1558 extractMallocCallFromBitCast(Malloc) : cast<CallInst>(Malloc);
1561 GVI = PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, TD, true), TD);
1568 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1569 // that only one value (besides its initializer) is ever stored to the global.
1570 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1571 AtomicOrdering Ordering,
1572 Module::global_iterator &GVI,
1574 // Ignore no-op GEPs and bitcasts.
1575 StoredOnceVal = StoredOnceVal->stripPointerCasts();
1577 // If we are dealing with a pointer global that is initialized to null and
1578 // only has one (non-null) value stored into it, then we can optimize any
1579 // users of the loaded value (often calls and loads) that would trap if the
1581 if (GV->getInitializer()->getType()->isPointerTy() &&
1582 GV->getInitializer()->isNullValue()) {
1583 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1584 if (GV->getInitializer()->getType() != SOVC->getType())
1585 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1587 // Optimize away any trapping uses of the loaded value.
1588 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1590 } else if (CallInst *CI = extractMallocCall(StoredOnceVal)) {
1591 Type *MallocType = getMallocAllocatedType(CI);
1592 if (MallocType && TryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType,
1601 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1602 /// two values ever stored into GV are its initializer and OtherVal. See if we
1603 /// can shrink the global into a boolean and select between the two values
1604 /// whenever it is used. This exposes the values to other scalar optimizations.
1605 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1606 Type *GVElType = GV->getType()->getElementType();
1608 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1609 // an FP value, pointer or vector, don't do this optimization because a select
1610 // between them is very expensive and unlikely to lead to later
1611 // simplification. In these cases, we typically end up with "cond ? v1 : v2"
1612 // where v1 and v2 both require constant pool loads, a big loss.
1613 if (GVElType == Type::getInt1Ty(GV->getContext()) ||
1614 GVElType->isFloatingPointTy() ||
1615 GVElType->isPointerTy() || GVElType->isVectorTy())
1618 // Walk the use list of the global seeing if all the uses are load or store.
1619 // If there is anything else, bail out.
1620 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I){
1622 if (!isa<LoadInst>(U) && !isa<StoreInst>(U))
1626 DEBUG(dbgs() << " *** SHRINKING TO BOOL: " << *GV);
1628 // Create the new global, initializing it to false.
1629 GlobalVariable *NewGV = new GlobalVariable(Type::getInt1Ty(GV->getContext()),
1631 GlobalValue::InternalLinkage,
1632 ConstantInt::getFalse(GV->getContext()),
1634 GV->isThreadLocal());
1635 GV->getParent()->getGlobalList().insert(GV, NewGV);
1637 Constant *InitVal = GV->getInitializer();
1638 assert(InitVal->getType() != Type::getInt1Ty(GV->getContext()) &&
1639 "No reason to shrink to bool!");
1641 // If initialized to zero and storing one into the global, we can use a cast
1642 // instead of a select to synthesize the desired value.
1643 bool IsOneZero = false;
1644 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1645 IsOneZero = InitVal->isNullValue() && CI->isOne();
1647 while (!GV->use_empty()) {
1648 Instruction *UI = cast<Instruction>(GV->use_back());
1649 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1650 // Change the store into a boolean store.
1651 bool StoringOther = SI->getOperand(0) == OtherVal;
1652 // Only do this if we weren't storing a loaded value.
1654 if (StoringOther || SI->getOperand(0) == InitVal)
1655 StoreVal = ConstantInt::get(Type::getInt1Ty(GV->getContext()),
1658 // Otherwise, we are storing a previously loaded copy. To do this,
1659 // change the copy from copying the original value to just copying the
1661 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1663 // If we've already replaced the input, StoredVal will be a cast or
1664 // select instruction. If not, it will be a load of the original
1666 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1667 assert(LI->getOperand(0) == GV && "Not a copy!");
1668 // Insert a new load, to preserve the saved value.
1669 StoreVal = new LoadInst(NewGV, LI->getName()+".b", false, 0,
1670 LI->getOrdering(), LI->getSynchScope(), LI);
1672 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1673 "This is not a form that we understand!");
1674 StoreVal = StoredVal->getOperand(0);
1675 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1678 new StoreInst(StoreVal, NewGV, false, 0,
1679 SI->getOrdering(), SI->getSynchScope(), SI);
1681 // Change the load into a load of bool then a select.
1682 LoadInst *LI = cast<LoadInst>(UI);
1683 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", false, 0,
1684 LI->getOrdering(), LI->getSynchScope(), LI);
1687 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1689 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1691 LI->replaceAllUsesWith(NSI);
1693 UI->eraseFromParent();
1696 GV->eraseFromParent();
1701 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1702 /// it if possible. If we make a change, return true.
1703 bool GlobalOpt::ProcessGlobal(GlobalVariable *GV,
1704 Module::global_iterator &GVI) {
1705 if (!GV->hasLocalLinkage())
1708 // Do more involved optimizations if the global is internal.
1709 GV->removeDeadConstantUsers();
1711 if (GV->use_empty()) {
1712 DEBUG(dbgs() << "GLOBAL DEAD: " << *GV);
1713 GV->eraseFromParent();
1718 SmallPtrSet<const PHINode*, 16> PHIUsers;
1721 if (AnalyzeGlobal(GV, GS, PHIUsers))
1724 if (!GS.isCompared && !GV->hasUnnamedAddr()) {
1725 GV->setUnnamedAddr(true);
1729 if (GV->isConstant() || !GV->hasInitializer())
1732 return ProcessInternalGlobal(GV, GVI, PHIUsers, GS);
1735 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1736 /// it if possible. If we make a change, return true.
1737 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1738 Module::global_iterator &GVI,
1739 const SmallPtrSet<const PHINode*, 16> &PHIUsers,
1740 const GlobalStatus &GS) {
1741 // If this is a first class global and has only one accessing function
1742 // and this function is main (which we know is not recursive we can make
1743 // this global a local variable) we replace the global with a local alloca
1744 // in this function.
1746 // NOTE: It doesn't make sense to promote non single-value types since we
1747 // are just replacing static memory to stack memory.
1749 // If the global is in different address space, don't bring it to stack.
1750 if (!GS.HasMultipleAccessingFunctions &&
1751 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1752 GV->getType()->getElementType()->isSingleValueType() &&
1753 GS.AccessingFunction->getName() == "main" &&
1754 GS.AccessingFunction->hasExternalLinkage() &&
1755 GV->getType()->getAddressSpace() == 0) {
1756 DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV);
1757 Instruction &FirstI = const_cast<Instruction&>(*GS.AccessingFunction
1758 ->getEntryBlock().begin());
1759 Type *ElemTy = GV->getType()->getElementType();
1760 // FIXME: Pass Global's alignment when globals have alignment
1761 AllocaInst *Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), &FirstI);
1762 if (!isa<UndefValue>(GV->getInitializer()))
1763 new StoreInst(GV->getInitializer(), Alloca, &FirstI);
1765 GV->replaceAllUsesWith(Alloca);
1766 GV->eraseFromParent();
1771 // If the global is never loaded (but may be stored to), it is dead.
1774 DEBUG(dbgs() << "GLOBAL NEVER LOADED: " << *GV);
1776 // Delete any stores we can find to the global. We may not be able to
1777 // make it completely dead though.
1778 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1780 // If the global is dead now, delete it.
1781 if (GV->use_empty()) {
1782 GV->eraseFromParent();
1788 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1789 DEBUG(dbgs() << "MARKING CONSTANT: " << *GV);
1790 GV->setConstant(true);
1792 // Clean up any obviously simplifiable users now.
1793 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1795 // If the global is dead now, just nuke it.
1796 if (GV->use_empty()) {
1797 DEBUG(dbgs() << " *** Marking constant allowed us to simplify "
1798 << "all users and delete global!\n");
1799 GV->eraseFromParent();
1805 } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
1806 if (TargetData *TD = getAnalysisIfAvailable<TargetData>())
1807 if (GlobalVariable *FirstNewGV = SRAGlobal(GV, *TD)) {
1808 GVI = FirstNewGV; // Don't skip the newly produced globals!
1811 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1812 // If the initial value for the global was an undef value, and if only
1813 // one other value was stored into it, we can just change the
1814 // initializer to be the stored value, then delete all stores to the
1815 // global. This allows us to mark it constant.
1816 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1817 if (isa<UndefValue>(GV->getInitializer())) {
1818 // Change the initial value here.
1819 GV->setInitializer(SOVConstant);
1821 // Clean up any obviously simplifiable users now.
1822 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1824 if (GV->use_empty()) {
1825 DEBUG(dbgs() << " *** Substituting initializer allowed us to "
1826 << "simplify all users and delete global!\n");
1827 GV->eraseFromParent();
1836 // Try to optimize globals based on the knowledge that only one value
1837 // (besides its initializer) is ever stored to the global.
1838 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GS.Ordering, GVI,
1839 getAnalysisIfAvailable<TargetData>()))
1842 // Otherwise, if the global was not a boolean, we can shrink it to be a
1844 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1845 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1854 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1855 /// function, changing them to FastCC.
1856 static void ChangeCalleesToFastCall(Function *F) {
1857 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1858 CallSite User(cast<Instruction>(*UI));
1859 User.setCallingConv(CallingConv::Fast);
1863 static AttrListPtr StripNest(const AttrListPtr &Attrs) {
1864 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1865 if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0)
1868 // There can be only one.
1869 return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest);
1875 static void RemoveNestAttribute(Function *F) {
1876 F->setAttributes(StripNest(F->getAttributes()));
1877 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1878 CallSite User(cast<Instruction>(*UI));
1879 User.setAttributes(StripNest(User.getAttributes()));
1883 bool GlobalOpt::OptimizeFunctions(Module &M) {
1884 bool Changed = false;
1885 // Optimize functions.
1886 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1888 // Functions without names cannot be referenced outside this module.
1889 if (!F->hasName() && !F->isDeclaration())
1890 F->setLinkage(GlobalValue::InternalLinkage);
1891 F->removeDeadConstantUsers();
1892 if (F->isDefTriviallyDead()) {
1893 F->eraseFromParent();
1896 } else if (F->hasLocalLinkage()) {
1897 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1898 !F->hasAddressTaken()) {
1899 // If this function has C calling conventions, is not a varargs
1900 // function, and is only called directly, promote it to use the Fast
1901 // calling convention.
1902 F->setCallingConv(CallingConv::Fast);
1903 ChangeCalleesToFastCall(F);
1908 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
1909 !F->hasAddressTaken()) {
1910 // The function is not used by a trampoline intrinsic, so it is safe
1911 // to remove the 'nest' attribute.
1912 RemoveNestAttribute(F);
1921 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1922 bool Changed = false;
1923 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1925 GlobalVariable *GV = GVI++;
1926 // Global variables without names cannot be referenced outside this module.
1927 if (!GV->hasName() && !GV->isDeclaration())
1928 GV->setLinkage(GlobalValue::InternalLinkage);
1929 // Simplify the initializer.
1930 if (GV->hasInitializer())
1931 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GV->getInitializer())) {
1932 TargetData *TD = getAnalysisIfAvailable<TargetData>();
1933 TargetLibraryInfo *TLI = &getAnalysis<TargetLibraryInfo>();
1934 Constant *New = ConstantFoldConstantExpression(CE, TD, TLI);
1935 if (New && New != CE)
1936 GV->setInitializer(New);
1939 Changed |= ProcessGlobal(GV, GVI);
1944 /// FindGlobalCtors - Find the llvm.global_ctors list, verifying that all
1945 /// initializers have an init priority of 65535.
1946 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1947 GlobalVariable *GV = M.getGlobalVariable("llvm.global_ctors");
1948 if (GV == 0) return 0;
1950 // Verify that the initializer is simple enough for us to handle. We are
1951 // only allowed to optimize the initializer if it is unique.
1952 if (!GV->hasUniqueInitializer()) return 0;
1954 if (isa<ConstantAggregateZero>(GV->getInitializer()))
1956 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1958 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
1959 if (isa<ConstantAggregateZero>(*i))
1961 ConstantStruct *CS = cast<ConstantStruct>(*i);
1962 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1965 // Must have a function or null ptr.
1966 if (!isa<Function>(CS->getOperand(1)))
1969 // Init priority must be standard.
1970 ConstantInt *CI = cast<ConstantInt>(CS->getOperand(0));
1971 if (CI->getZExtValue() != 65535)
1978 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1979 /// return a list of the functions and null terminator as a vector.
1980 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1981 if (GV->getInitializer()->isNullValue())
1982 return std::vector<Function*>();
1983 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1984 std::vector<Function*> Result;
1985 Result.reserve(CA->getNumOperands());
1986 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
1987 ConstantStruct *CS = cast<ConstantStruct>(*i);
1988 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1993 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1994 /// specified array, returning the new global to use.
1995 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1996 const std::vector<Function*> &Ctors) {
1997 // If we made a change, reassemble the initializer list.
1998 Constant *CSVals[2];
1999 CSVals[0] = ConstantInt::get(Type::getInt32Ty(GCL->getContext()), 65535);
2002 StructType *StructTy =
2004 cast<ArrayType>(GCL->getType()->getElementType())->getElementType());
2006 // Create the new init list.
2007 std::vector<Constant*> CAList;
2008 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
2010 CSVals[1] = Ctors[i];
2012 Type *FTy = FunctionType::get(Type::getVoidTy(GCL->getContext()),
2014 PointerType *PFTy = PointerType::getUnqual(FTy);
2015 CSVals[1] = Constant::getNullValue(PFTy);
2016 CSVals[0] = ConstantInt::get(Type::getInt32Ty(GCL->getContext()),
2019 CAList.push_back(ConstantStruct::get(StructTy, CSVals));
2022 // Create the array initializer.
2023 Constant *CA = ConstantArray::get(ArrayType::get(StructTy,
2024 CAList.size()), CAList);
2026 // If we didn't change the number of elements, don't create a new GV.
2027 if (CA->getType() == GCL->getInitializer()->getType()) {
2028 GCL->setInitializer(CA);
2032 // Create the new global and insert it next to the existing list.
2033 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
2034 GCL->getLinkage(), CA, "",
2035 GCL->isThreadLocal());
2036 GCL->getParent()->getGlobalList().insert(GCL, NGV);
2039 // Nuke the old list, replacing any uses with the new one.
2040 if (!GCL->use_empty()) {
2042 if (V->getType() != GCL->getType())
2043 V = ConstantExpr::getBitCast(V, GCL->getType());
2044 GCL->replaceAllUsesWith(V);
2046 GCL->eraseFromParent();
2055 static Constant *getVal(DenseMap<Value*, Constant*> &ComputedValues, Value *V) {
2056 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
2057 Constant *R = ComputedValues[V];
2058 assert(R && "Reference to an uncomputed value!");
2063 isSimpleEnoughValueToCommit(Constant *C,
2064 SmallPtrSet<Constant*, 8> &SimpleConstants,
2065 const TargetData *TD);
2068 /// isSimpleEnoughValueToCommit - Return true if the specified constant can be
2069 /// handled by the code generator. We don't want to generate something like:
2070 /// void *X = &X/42;
2071 /// because the code generator doesn't have a relocation that can handle that.
2073 /// This function should be called if C was not found (but just got inserted)
2074 /// in SimpleConstants to avoid having to rescan the same constants all the
2076 static bool isSimpleEnoughValueToCommitHelper(Constant *C,
2077 SmallPtrSet<Constant*, 8> &SimpleConstants,
2078 const TargetData *TD) {
2079 // Simple integer, undef, constant aggregate zero, global addresses, etc are
2081 if (C->getNumOperands() == 0 || isa<BlockAddress>(C) ||
2082 isa<GlobalValue>(C))
2085 // Aggregate values are safe if all their elements are.
2086 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C) ||
2087 isa<ConstantVector>(C)) {
2088 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) {
2089 Constant *Op = cast<Constant>(C->getOperand(i));
2090 if (!isSimpleEnoughValueToCommit(Op, SimpleConstants, TD))
2096 // We don't know exactly what relocations are allowed in constant expressions,
2097 // so we allow &global+constantoffset, which is safe and uniformly supported
2099 ConstantExpr *CE = cast<ConstantExpr>(C);
2100 switch (CE->getOpcode()) {
2101 case Instruction::BitCast:
2102 // Bitcast is fine if the casted value is fine.
2103 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, TD);
2105 case Instruction::IntToPtr:
2106 case Instruction::PtrToInt:
2107 // int <=> ptr is fine if the int type is the same size as the
2109 if (!TD || TD->getTypeSizeInBits(CE->getType()) !=
2110 TD->getTypeSizeInBits(CE->getOperand(0)->getType()))
2112 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, TD);
2114 // GEP is fine if it is simple + constant offset.
2115 case Instruction::GetElementPtr:
2116 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
2117 if (!isa<ConstantInt>(CE->getOperand(i)))
2119 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, TD);
2121 case Instruction::Add:
2122 // We allow simple+cst.
2123 if (!isa<ConstantInt>(CE->getOperand(1)))
2125 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, TD);
2131 isSimpleEnoughValueToCommit(Constant *C,
2132 SmallPtrSet<Constant*, 8> &SimpleConstants,
2133 const TargetData *TD) {
2134 // If we already checked this constant, we win.
2135 if (!SimpleConstants.insert(C)) return true;
2136 // Check the constant.
2137 return isSimpleEnoughValueToCommitHelper(C, SimpleConstants, TD);
2141 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
2142 /// enough for us to understand. In particular, if it is a cast to anything
2143 /// other than from one pointer type to another pointer type, we punt.
2144 /// We basically just support direct accesses to globals and GEP's of
2145 /// globals. This should be kept up to date with CommitValueTo.
2146 static bool isSimpleEnoughPointerToCommit(Constant *C) {
2147 // Conservatively, avoid aggregate types. This is because we don't
2148 // want to worry about them partially overlapping other stores.
2149 if (!cast<PointerType>(C->getType())->getElementType()->isSingleValueType())
2152 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
2153 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2154 // external globals.
2155 return GV->hasUniqueInitializer();
2157 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2158 // Handle a constantexpr gep.
2159 if (CE->getOpcode() == Instruction::GetElementPtr &&
2160 isa<GlobalVariable>(CE->getOperand(0)) &&
2161 cast<GEPOperator>(CE)->isInBounds()) {
2162 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2163 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2164 // external globals.
2165 if (!GV->hasUniqueInitializer())
2168 // The first index must be zero.
2169 ConstantInt *CI = dyn_cast<ConstantInt>(*llvm::next(CE->op_begin()));
2170 if (!CI || !CI->isZero()) return false;
2172 // The remaining indices must be compile-time known integers within the
2173 // notional bounds of the corresponding static array types.
2174 if (!CE->isGEPWithNoNotionalOverIndexing())
2177 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2179 // A constantexpr bitcast from a pointer to another pointer is a no-op,
2180 // and we know how to evaluate it by moving the bitcast from the pointer
2181 // operand to the value operand.
2182 } else if (CE->getOpcode() == Instruction::BitCast &&
2183 isa<GlobalVariable>(CE->getOperand(0))) {
2184 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2185 // external globals.
2186 return cast<GlobalVariable>(CE->getOperand(0))->hasUniqueInitializer();
2193 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
2194 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
2195 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
2196 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
2197 ConstantExpr *Addr, unsigned OpNo) {
2198 // Base case of the recursion.
2199 if (OpNo == Addr->getNumOperands()) {
2200 assert(Val->getType() == Init->getType() && "Type mismatch!");
2204 SmallVector<Constant*, 32> Elts;
2205 if (StructType *STy = dyn_cast<StructType>(Init->getType())) {
2206 // Break up the constant into its elements.
2207 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2208 Elts.push_back(Init->getAggregateElement(i));
2210 // Replace the element that we are supposed to.
2211 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
2212 unsigned Idx = CU->getZExtValue();
2213 assert(Idx < STy->getNumElements() && "Struct index out of range!");
2214 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
2216 // Return the modified struct.
2217 return ConstantStruct::get(STy, Elts);
2220 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
2221 SequentialType *InitTy = cast<SequentialType>(Init->getType());
2224 if (ArrayType *ATy = dyn_cast<ArrayType>(InitTy))
2225 NumElts = ATy->getNumElements();
2227 NumElts = InitTy->getVectorNumElements();
2229 // Break up the array into elements.
2230 for (uint64_t i = 0, e = NumElts; i != e; ++i)
2231 Elts.push_back(Init->getAggregateElement(i));
2233 assert(CI->getZExtValue() < NumElts);
2234 Elts[CI->getZExtValue()] =
2235 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
2237 if (Init->getType()->isArrayTy())
2238 return ConstantArray::get(cast<ArrayType>(InitTy), Elts);
2239 return ConstantVector::get(Elts);
2242 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
2243 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
2244 static void CommitValueTo(Constant *Val, Constant *Addr) {
2245 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
2246 assert(GV->hasInitializer());
2247 GV->setInitializer(Val);
2251 ConstantExpr *CE = cast<ConstantExpr>(Addr);
2252 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2253 GV->setInitializer(EvaluateStoreInto(GV->getInitializer(), Val, CE, 2));
2256 /// ComputeLoadResult - Return the value that would be computed by a load from
2257 /// P after the stores reflected by 'memory' have been performed. If we can't
2258 /// decide, return null.
2259 static Constant *ComputeLoadResult(Constant *P,
2260 const DenseMap<Constant*, Constant*> &Memory) {
2261 // If this memory location has been recently stored, use the stored value: it
2262 // is the most up-to-date.
2263 DenseMap<Constant*, Constant*>::const_iterator I = Memory.find(P);
2264 if (I != Memory.end()) return I->second;
2267 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
2268 if (GV->hasDefinitiveInitializer())
2269 return GV->getInitializer();
2273 // Handle a constantexpr getelementptr.
2274 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
2275 if (CE->getOpcode() == Instruction::GetElementPtr &&
2276 isa<GlobalVariable>(CE->getOperand(0))) {
2277 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2278 if (GV->hasDefinitiveInitializer())
2279 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2282 return 0; // don't know how to evaluate.
2285 /// EvaluateFunction - Evaluate a call to function F, returning true if
2286 /// successful, false if we can't evaluate it. ActualArgs contains the formal
2287 /// arguments for the function.
2288 static bool EvaluateFunction(Function *F, Constant *&RetVal,
2289 const SmallVectorImpl<Constant*> &ActualArgs,
2290 std::vector<Function*> &CallStack,
2291 DenseMap<Constant*, Constant*> &MutatedMemory,
2292 std::vector<GlobalVariable*> &AllocaTmps,
2293 SmallPtrSet<Constant*, 8> &SimpleConstants,
2294 const TargetData *TD,
2295 const TargetLibraryInfo *TLI) {
2296 // Check to see if this function is already executing (recursion). If so,
2297 // bail out. TODO: we might want to accept limited recursion.
2298 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
2301 CallStack.push_back(F);
2303 /// Values - As we compute SSA register values, we store their contents here.
2304 DenseMap<Value*, Constant*> Values;
2306 // Initialize arguments to the incoming values specified.
2308 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
2310 Values[AI] = ActualArgs[ArgNo];
2312 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
2313 /// we can only evaluate any one basic block at most once. This set keeps
2314 /// track of what we have executed so we can detect recursive cases etc.
2315 SmallPtrSet<BasicBlock*, 32> ExecutedBlocks;
2317 // CurInst - The current instruction we're evaluating.
2318 BasicBlock::iterator CurInst = F->begin()->begin();
2320 // This is the main evaluation loop.
2322 Constant *InstResult = 0;
2324 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
2325 if (!SI->isSimple()) return false; // no volatile/atomic accesses.
2326 Constant *Ptr = getVal(Values, SI->getOperand(1));
2327 if (!isSimpleEnoughPointerToCommit(Ptr))
2328 // If this is too complex for us to commit, reject it.
2331 Constant *Val = getVal(Values, SI->getOperand(0));
2333 // If this might be too difficult for the backend to handle (e.g. the addr
2334 // of one global variable divided by another) then we can't commit it.
2335 if (!isSimpleEnoughValueToCommit(Val, SimpleConstants, TD))
2338 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
2339 if (CE->getOpcode() == Instruction::BitCast) {
2340 // If we're evaluating a store through a bitcast, then we need
2341 // to pull the bitcast off the pointer type and push it onto the
2343 Ptr = CE->getOperand(0);
2345 Type *NewTy=cast<PointerType>(Ptr->getType())->getElementType();
2347 // In order to push the bitcast onto the stored value, a bitcast
2348 // from NewTy to Val's type must be legal. If it's not, we can try
2349 // introspecting NewTy to find a legal conversion.
2350 while (!Val->getType()->canLosslesslyBitCastTo(NewTy)) {
2351 // If NewTy is a struct, we can convert the pointer to the struct
2352 // into a pointer to its first member.
2353 // FIXME: This could be extended to support arrays as well.
2354 if (StructType *STy = dyn_cast<StructType>(NewTy)) {
2355 NewTy = STy->getTypeAtIndex(0U);
2357 IntegerType *IdxTy =IntegerType::get(NewTy->getContext(), 32);
2358 Constant *IdxZero = ConstantInt::get(IdxTy, 0, false);
2359 Constant * const IdxList[] = {IdxZero, IdxZero};
2361 Ptr = ConstantExpr::getGetElementPtr(Ptr, IdxList);
2363 // If we can't improve the situation by introspecting NewTy,
2364 // we have to give up.
2370 // If we found compatible types, go ahead and push the bitcast
2371 // onto the stored value.
2372 Val = ConstantExpr::getBitCast(Val, NewTy);
2375 MutatedMemory[Ptr] = Val;
2376 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2377 InstResult = ConstantExpr::get(BO->getOpcode(),
2378 getVal(Values, BO->getOperand(0)),
2379 getVal(Values, BO->getOperand(1)));
2380 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2381 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2382 getVal(Values, CI->getOperand(0)),
2383 getVal(Values, CI->getOperand(1)));
2384 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2385 InstResult = ConstantExpr::getCast(CI->getOpcode(),
2386 getVal(Values, CI->getOperand(0)),
2388 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2389 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
2390 getVal(Values, SI->getOperand(1)),
2391 getVal(Values, SI->getOperand(2)));
2392 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2393 Constant *P = getVal(Values, GEP->getOperand(0));
2394 SmallVector<Constant*, 8> GEPOps;
2395 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2397 GEPOps.push_back(getVal(Values, *i));
2399 ConstantExpr::getGetElementPtr(P, GEPOps,
2400 cast<GEPOperator>(GEP)->isInBounds());
2401 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2402 if (!LI->isSimple()) return false; // no volatile/atomic accesses.
2403 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
2405 if (InstResult == 0) return false; // Could not evaluate load.
2406 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2407 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs.
2408 Type *Ty = AI->getType()->getElementType();
2409 AllocaTmps.push_back(new GlobalVariable(Ty, false,
2410 GlobalValue::InternalLinkage,
2411 UndefValue::get(Ty),
2413 InstResult = AllocaTmps.back();
2414 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
2416 // Debug info can safely be ignored here.
2417 if (isa<DbgInfoIntrinsic>(CI)) {
2422 // Cannot handle inline asm.
2423 if (isa<InlineAsm>(CI->getCalledValue())) return false;
2425 if (MemSetInst *MSI = dyn_cast<MemSetInst>(CI)) {
2426 if (MSI->isVolatile()) return false;
2427 Constant *Ptr = getVal(Values, MSI->getDest());
2428 Constant *Val = getVal(Values, MSI->getValue());
2429 Constant *DestVal = ComputeLoadResult(getVal(Values, Ptr),
2431 if (Val->isNullValue() && DestVal && DestVal->isNullValue()) {
2432 // This memset is a no-op.
2439 // Resolve function pointers.
2440 Function *Callee = dyn_cast<Function>(getVal(Values,
2441 CI->getCalledValue()));
2442 if (!Callee) return false; // Cannot resolve.
2444 SmallVector<Constant*, 8> Formals;
2446 for (User::op_iterator i = CS.arg_begin(), e = CS.arg_end();
2448 Formals.push_back(getVal(Values, *i));
2450 if (Callee->isDeclaration()) {
2451 // If this is a function we can constant fold, do it.
2452 if (Constant *C = ConstantFoldCall(Callee, Formals, TLI)) {
2458 if (Callee->getFunctionType()->isVarArg())
2462 // Execute the call, if successful, use the return value.
2463 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
2464 MutatedMemory, AllocaTmps, SimpleConstants, TD,
2467 InstResult = RetVal;
2469 } else if (isa<TerminatorInst>(CurInst)) {
2470 BasicBlock *NewBB = 0;
2471 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2472 if (BI->isUnconditional()) {
2473 NewBB = BI->getSuccessor(0);
2476 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
2477 if (!Cond) return false; // Cannot determine.
2479 NewBB = BI->getSuccessor(!Cond->getZExtValue());
2481 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2483 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
2484 if (!Val) return false; // Cannot determine.
2485 unsigned ValTISucc = SI->resolveSuccessorIndex(SI->findCaseValue(Val));
2486 NewBB = SI->getSuccessor(ValTISucc);
2487 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(CurInst)) {
2488 Value *Val = getVal(Values, IBI->getAddress())->stripPointerCasts();
2489 if (BlockAddress *BA = dyn_cast<BlockAddress>(Val))
2490 NewBB = BA->getBasicBlock();
2492 return false; // Cannot determine.
2493 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
2494 if (RI->getNumOperands())
2495 RetVal = getVal(Values, RI->getOperand(0));
2497 CallStack.pop_back(); // return from fn.
2498 return true; // We succeeded at evaluating this ctor!
2500 // invoke, unwind, resume, unreachable.
2501 return false; // Cannot handle this terminator.
2504 // Okay, we succeeded in evaluating this control flow. See if we have
2505 // executed the new block before. If so, we have a looping function,
2506 // which we cannot evaluate in reasonable time.
2507 if (!ExecutedBlocks.insert(NewBB))
2508 return false; // looped!
2510 // Okay, we have never been in this block before. Check to see if there
2511 // are any PHI nodes. If so, evaluate them with information about where
2513 BasicBlock *OldBB = CurInst->getParent();
2514 CurInst = NewBB->begin();
2516 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2517 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
2519 // Do NOT increment CurInst. We know that the terminator had no value.
2522 // Did not know how to evaluate this!
2526 if (!CurInst->use_empty()) {
2527 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(InstResult))
2528 InstResult = ConstantFoldConstantExpression(CE, TD, TLI);
2530 Values[CurInst] = InstResult;
2533 // Advance program counter.
2538 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2539 /// we can. Return true if we can, false otherwise.
2540 static bool EvaluateStaticConstructor(Function *F, const TargetData *TD,
2541 const TargetLibraryInfo *TLI) {
2542 /// MutatedMemory - For each store we execute, we update this map. Loads
2543 /// check this to get the most up-to-date value. If evaluation is successful,
2544 /// this state is committed to the process.
2545 DenseMap<Constant*, Constant*> MutatedMemory;
2547 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2548 /// to represent its body. This vector is needed so we can delete the
2549 /// temporary globals when we are done.
2550 std::vector<GlobalVariable*> AllocaTmps;
2552 /// CallStack - This is used to detect recursion. In pathological situations
2553 /// we could hit exponential behavior, but at least there is nothing
2555 std::vector<Function*> CallStack;
2557 /// SimpleConstants - These are constants we have checked and know to be
2558 /// simple enough to live in a static initializer of a global.
2559 SmallPtrSet<Constant*, 8> SimpleConstants;
2561 // Call the function.
2562 Constant *RetValDummy;
2563 bool EvalSuccess = EvaluateFunction(F, RetValDummy,
2564 SmallVector<Constant*, 0>(), CallStack,
2565 MutatedMemory, AllocaTmps,
2566 SimpleConstants, TD, TLI);
2569 // We succeeded at evaluation: commit the result.
2570 DEBUG(dbgs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2571 << F->getName() << "' to " << MutatedMemory.size()
2573 for (DenseMap<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2574 E = MutatedMemory.end(); I != E; ++I)
2575 CommitValueTo(I->second, I->first);
2578 // At this point, we are done interpreting. If we created any 'alloca'
2579 // temporaries, release them now.
2580 while (!AllocaTmps.empty()) {
2581 GlobalVariable *Tmp = AllocaTmps.back();
2582 AllocaTmps.pop_back();
2584 // If there are still users of the alloca, the program is doing something
2585 // silly, e.g. storing the address of the alloca somewhere and using it
2586 // later. Since this is undefined, we'll just make it be null.
2587 if (!Tmp->use_empty())
2588 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2595 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2596 /// Return true if anything changed.
2597 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2598 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2599 bool MadeChange = false;
2600 if (Ctors.empty()) return false;
2602 const TargetData *TD = getAnalysisIfAvailable<TargetData>();
2603 const TargetLibraryInfo *TLI = &getAnalysis<TargetLibraryInfo>();
2605 // Loop over global ctors, optimizing them when we can.
2606 for (unsigned i = 0; i != Ctors.size(); ++i) {
2607 Function *F = Ctors[i];
2608 // Found a null terminator in the middle of the list, prune off the rest of
2611 if (i != Ctors.size()-1) {
2618 // We cannot simplify external ctor functions.
2619 if (F->empty()) continue;
2621 // If we can evaluate the ctor at compile time, do.
2622 if (EvaluateStaticConstructor(F, TD, TLI)) {
2623 Ctors.erase(Ctors.begin()+i);
2626 ++NumCtorsEvaluated;
2631 if (!MadeChange) return false;
2633 GCL = InstallGlobalCtors(GCL, Ctors);
2637 bool GlobalOpt::OptimizeGlobalAliases(Module &M) {
2638 bool Changed = false;
2640 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end();
2642 Module::alias_iterator J = I++;
2643 // Aliases without names cannot be referenced outside this module.
2644 if (!J->hasName() && !J->isDeclaration())
2645 J->setLinkage(GlobalValue::InternalLinkage);
2646 // If the aliasee may change at link time, nothing can be done - bail out.
2647 if (J->mayBeOverridden())
2650 Constant *Aliasee = J->getAliasee();
2651 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2652 Target->removeDeadConstantUsers();
2653 bool hasOneUse = Target->hasOneUse() && Aliasee->hasOneUse();
2655 // Make all users of the alias use the aliasee instead.
2656 if (!J->use_empty()) {
2657 J->replaceAllUsesWith(Aliasee);
2658 ++NumAliasesResolved;
2662 // If the alias is externally visible, we may still be able to simplify it.
2663 if (!J->hasLocalLinkage()) {
2664 // If the aliasee has internal linkage, give it the name and linkage
2665 // of the alias, and delete the alias. This turns:
2666 // define internal ... @f(...)
2667 // @a = alias ... @f
2669 // define ... @a(...)
2670 if (!Target->hasLocalLinkage())
2673 // Do not perform the transform if multiple aliases potentially target the
2674 // aliasee. This check also ensures that it is safe to replace the section
2675 // and other attributes of the aliasee with those of the alias.
2679 // Give the aliasee the name, linkage and other attributes of the alias.
2680 Target->takeName(J);
2681 Target->setLinkage(J->getLinkage());
2682 Target->GlobalValue::copyAttributesFrom(J);
2685 // Delete the alias.
2686 M.getAliasList().erase(J);
2687 ++NumAliasesRemoved;
2694 static Function *FindCXAAtExit(Module &M) {
2695 Function *Fn = M.getFunction("__cxa_atexit");
2700 FunctionType *FTy = Fn->getFunctionType();
2702 // Checking that the function has the right return type, the right number of
2703 // parameters and that they all have pointer types should be enough.
2704 if (!FTy->getReturnType()->isIntegerTy() ||
2705 FTy->getNumParams() != 3 ||
2706 !FTy->getParamType(0)->isPointerTy() ||
2707 !FTy->getParamType(1)->isPointerTy() ||
2708 !FTy->getParamType(2)->isPointerTy())
2714 /// cxxDtorIsEmpty - Returns whether the given function is an empty C++
2715 /// destructor and can therefore be eliminated.
2716 /// Note that we assume that other optimization passes have already simplified
2717 /// the code so we only look for a function with a single basic block, where
2718 /// the only allowed instructions are 'ret' or 'call' to empty C++ dtor.
2719 static bool cxxDtorIsEmpty(const Function &Fn,
2720 SmallPtrSet<const Function *, 8> &CalledFunctions) {
2721 // FIXME: We could eliminate C++ destructors if they're readonly/readnone and
2722 // nounwind, but that doesn't seem worth doing.
2723 if (Fn.isDeclaration())
2726 if (++Fn.begin() != Fn.end())
2729 const BasicBlock &EntryBlock = Fn.getEntryBlock();
2730 for (BasicBlock::const_iterator I = EntryBlock.begin(), E = EntryBlock.end();
2732 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
2733 // Ignore debug intrinsics.
2734 if (isa<DbgInfoIntrinsic>(CI))
2737 const Function *CalledFn = CI->getCalledFunction();
2742 SmallPtrSet<const Function *, 8> NewCalledFunctions(CalledFunctions);
2744 // Don't treat recursive functions as empty.
2745 if (!NewCalledFunctions.insert(CalledFn))
2748 if (!cxxDtorIsEmpty(*CalledFn, NewCalledFunctions))
2750 } else if (isa<ReturnInst>(*I))
2759 bool GlobalOpt::OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) {
2760 /// Itanium C++ ABI p3.3.5:
2762 /// After constructing a global (or local static) object, that will require
2763 /// destruction on exit, a termination function is registered as follows:
2765 /// extern "C" int __cxa_atexit ( void (*f)(void *), void *p, void *d );
2767 /// This registration, e.g. __cxa_atexit(f,p,d), is intended to cause the
2768 /// call f(p) when DSO d is unloaded, before all such termination calls
2769 /// registered before this one. It returns zero if registration is
2770 /// successful, nonzero on failure.
2772 // This pass will look for calls to __cxa_atexit where the function is trivial
2774 bool Changed = false;
2776 for (Function::use_iterator I = CXAAtExitFn->use_begin(),
2777 E = CXAAtExitFn->use_end(); I != E;) {
2778 // We're only interested in calls. Theoretically, we could handle invoke
2779 // instructions as well, but neither llvm-gcc nor clang generate invokes
2781 CallInst *CI = dyn_cast<CallInst>(*I++);
2786 dyn_cast<Function>(CI->getArgOperand(0)->stripPointerCasts());
2790 SmallPtrSet<const Function *, 8> CalledFunctions;
2791 if (!cxxDtorIsEmpty(*DtorFn, CalledFunctions))
2794 // Just remove the call.
2795 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
2796 CI->eraseFromParent();
2798 ++NumCXXDtorsRemoved;
2806 bool GlobalOpt::runOnModule(Module &M) {
2807 bool Changed = false;
2809 // Try to find the llvm.globalctors list.
2810 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2812 Function *CXAAtExitFn = FindCXAAtExit(M);
2814 bool LocalChange = true;
2815 while (LocalChange) {
2816 LocalChange = false;
2818 // Delete functions that are trivially dead, ccc -> fastcc
2819 LocalChange |= OptimizeFunctions(M);
2821 // Optimize global_ctors list.
2823 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2825 // Optimize non-address-taken globals.
2826 LocalChange |= OptimizeGlobalVars(M);
2828 // Resolve aliases, when possible.
2829 LocalChange |= OptimizeGlobalAliases(M);
2831 // Try to remove trivial global destructors.
2833 LocalChange |= OptimizeEmptyGlobalCXXDtors(CXAAtExitFn);
2835 Changed |= LocalChange;
2838 // TODO: Move all global ctors functions to the end of the module for code