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/Pass.h"
25 #include "llvm/Analysis/ConstantFolding.h"
26 #include "llvm/Target/TargetData.h"
27 #include "llvm/Support/Compiler.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/GetElementPtrTypeIterator.h"
30 #include "llvm/ADT/SmallPtrSet.h"
31 #include "llvm/ADT/SmallVector.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/ADT/StringExtras.h"
38 STATISTIC(NumMarked , "Number of globals marked constant");
39 STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
40 STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
41 STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
42 STATISTIC(NumDeleted , "Number of globals deleted");
43 STATISTIC(NumFnDeleted , "Number of functions deleted");
44 STATISTIC(NumGlobUses , "Number of global uses devirtualized");
45 STATISTIC(NumLocalized , "Number of globals localized");
46 STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans");
47 STATISTIC(NumFastCallFns , "Number of functions converted to fastcc");
48 STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
51 struct VISIBILITY_HIDDEN GlobalOpt : public ModulePass {
52 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
53 AU.addRequired<TargetData>();
55 static char ID; // Pass identification, replacement for typeid
56 GlobalOpt() : ModulePass((intptr_t)&ID) {}
58 bool runOnModule(Module &M);
61 GlobalVariable *FindGlobalCtors(Module &M);
62 bool OptimizeFunctions(Module &M);
63 bool OptimizeGlobalVars(Module &M);
64 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
65 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
68 char GlobalOpt::ID = 0;
69 RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer");
72 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
74 /// GlobalStatus - As we analyze each global, keep track of some information
75 /// about it. If we find out that the address of the global is taken, none of
76 /// this info will be accurate.
77 struct VISIBILITY_HIDDEN GlobalStatus {
78 /// isLoaded - True if the global is ever loaded. If the global isn't ever
79 /// loaded it can be deleted.
82 /// StoredType - Keep track of what stores to the global look like.
85 /// NotStored - There is no store to this global. It can thus be marked
89 /// isInitializerStored - This global is stored to, but the only thing
90 /// stored is the constant it was initialized with. This is only tracked
91 /// for scalar globals.
94 /// isStoredOnce - This global is stored to, but only its initializer and
95 /// one other value is ever stored to it. If this global isStoredOnce, we
96 /// track the value stored to it in StoredOnceValue below. This is only
97 /// tracked for scalar globals.
100 /// isStored - This global is stored to by multiple values or something else
101 /// that we cannot track.
105 /// StoredOnceValue - If only one value (besides the initializer constant) is
106 /// ever stored to this global, keep track of what value it is.
107 Value *StoredOnceValue;
109 /// AccessingFunction/HasMultipleAccessingFunctions - These start out
110 /// null/false. When the first accessing function is noticed, it is recorded.
111 /// When a second different accessing function is noticed,
112 /// HasMultipleAccessingFunctions is set to true.
113 Function *AccessingFunction;
114 bool HasMultipleAccessingFunctions;
116 /// HasNonInstructionUser - Set to true if this global has a user that is not
117 /// an instruction (e.g. a constant expr or GV initializer).
118 bool HasNonInstructionUser;
120 /// HasPHIUser - Set to true if this global has a user that is a PHI node.
123 GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0),
124 AccessingFunction(0), HasMultipleAccessingFunctions(false),
125 HasNonInstructionUser(false), HasPHIUser(false) {}
130 /// ConstantIsDead - Return true if the specified constant is (transitively)
131 /// dead. The constant may be used by other constants (e.g. constant arrays and
132 /// constant exprs) as long as they are dead, but it cannot be used by anything
134 static bool ConstantIsDead(Constant *C) {
135 if (isa<GlobalValue>(C)) return false;
137 for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI)
138 if (Constant *CU = dyn_cast<Constant>(*UI)) {
139 if (!ConstantIsDead(CU)) return false;
146 /// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
147 /// structure. If the global has its address taken, return true to indicate we
148 /// can't do anything with it.
150 static bool AnalyzeGlobal(Value *V, GlobalStatus &GS,
151 std::set<PHINode*> &PHIUsers) {
152 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
153 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
154 GS.HasNonInstructionUser = true;
156 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
158 } else if (Instruction *I = dyn_cast<Instruction>(*UI)) {
159 if (!GS.HasMultipleAccessingFunctions) {
160 Function *F = I->getParent()->getParent();
161 if (GS.AccessingFunction == 0)
162 GS.AccessingFunction = F;
163 else if (GS.AccessingFunction != F)
164 GS.HasMultipleAccessingFunctions = true;
166 if (isa<LoadInst>(I)) {
168 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
169 // Don't allow a store OF the address, only stores TO the address.
170 if (SI->getOperand(0) == V) return true;
172 // If this is a direct store to the global (i.e., the global is a scalar
173 // value, not an aggregate), keep more specific information about
175 if (GS.StoredType != GlobalStatus::isStored)
176 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){
177 Value *StoredVal = SI->getOperand(0);
178 if (StoredVal == GV->getInitializer()) {
179 if (GS.StoredType < GlobalStatus::isInitializerStored)
180 GS.StoredType = GlobalStatus::isInitializerStored;
181 } else if (isa<LoadInst>(StoredVal) &&
182 cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
184 if (GS.StoredType < GlobalStatus::isInitializerStored)
185 GS.StoredType = GlobalStatus::isInitializerStored;
186 } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
187 GS.StoredType = GlobalStatus::isStoredOnce;
188 GS.StoredOnceValue = StoredVal;
189 } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
190 GS.StoredOnceValue == StoredVal) {
193 GS.StoredType = GlobalStatus::isStored;
196 GS.StoredType = GlobalStatus::isStored;
198 } else if (isa<GetElementPtrInst>(I)) {
199 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
200 } else if (isa<SelectInst>(I)) {
201 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
202 } else if (PHINode *PN = dyn_cast<PHINode>(I)) {
203 // PHI nodes we can check just like select or GEP instructions, but we
204 // have to be careful about infinite recursion.
205 if (PHIUsers.insert(PN).second) // Not already visited.
206 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
207 GS.HasPHIUser = true;
208 } else if (isa<CmpInst>(I)) {
209 } else if (isa<MemCpyInst>(I) || isa<MemMoveInst>(I)) {
210 if (I->getOperand(1) == V)
211 GS.StoredType = GlobalStatus::isStored;
212 if (I->getOperand(2) == V)
214 } else if (isa<MemSetInst>(I)) {
215 assert(I->getOperand(1) == V && "Memset only takes one pointer!");
216 GS.StoredType = GlobalStatus::isStored;
218 return true; // Any other non-load instruction might take address!
220 } else if (Constant *C = dyn_cast<Constant>(*UI)) {
221 GS.HasNonInstructionUser = true;
222 // We might have a dead and dangling constant hanging off of here.
223 if (!ConstantIsDead(C))
226 GS.HasNonInstructionUser = true;
227 // Otherwise must be some other user.
234 static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) {
235 ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
237 unsigned IdxV = CI->getZExtValue();
239 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
240 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
241 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
242 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
243 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) {
244 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
245 } else if (isa<ConstantAggregateZero>(Agg)) {
246 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
247 if (IdxV < STy->getNumElements())
248 return Constant::getNullValue(STy->getElementType(IdxV));
249 } else if (const SequentialType *STy =
250 dyn_cast<SequentialType>(Agg->getType())) {
251 return Constant::getNullValue(STy->getElementType());
253 } else if (isa<UndefValue>(Agg)) {
254 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
255 if (IdxV < STy->getNumElements())
256 return UndefValue::get(STy->getElementType(IdxV));
257 } else if (const SequentialType *STy =
258 dyn_cast<SequentialType>(Agg->getType())) {
259 return UndefValue::get(STy->getElementType());
266 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
267 /// users of the global, cleaning up the obvious ones. This is largely just a
268 /// quick scan over the use list to clean up the easy and obvious cruft. This
269 /// returns true if it made a change.
270 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
271 bool Changed = false;
272 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
275 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
277 // Replace the load with the initializer.
278 LI->replaceAllUsesWith(Init);
279 LI->eraseFromParent();
282 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
283 // Store must be unreachable or storing Init into the global.
284 SI->eraseFromParent();
286 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
287 if (CE->getOpcode() == Instruction::GetElementPtr) {
288 Constant *SubInit = 0;
290 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
291 Changed |= CleanupConstantGlobalUsers(CE, SubInit);
292 } else if (CE->getOpcode() == Instruction::BitCast &&
293 isa<PointerType>(CE->getType())) {
294 // Pointer cast, delete any stores and memsets to the global.
295 Changed |= CleanupConstantGlobalUsers(CE, 0);
298 if (CE->use_empty()) {
299 CE->destroyConstant();
302 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
303 // Do not transform "gepinst (gep constexpr (GV))" here, because forming
304 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
305 // and will invalidate our notion of what Init is.
306 Constant *SubInit = 0;
307 if (!isa<ConstantExpr>(GEP->getOperand(0))) {
309 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
310 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
311 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
313 Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
315 if (GEP->use_empty()) {
316 GEP->eraseFromParent();
319 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
320 if (MI->getRawDest() == V) {
321 MI->eraseFromParent();
325 } else if (Constant *C = dyn_cast<Constant>(U)) {
326 // If we have a chain of dead constantexprs or other things dangling from
327 // us, and if they are all dead, nuke them without remorse.
328 if (ConstantIsDead(C)) {
329 C->destroyConstant();
330 // This could have invalidated UI, start over from scratch.
331 CleanupConstantGlobalUsers(V, Init);
339 /// isSafeSROAElementUse - Return true if the specified instruction is a safe
340 /// user of a derived expression from a global that we want to SROA.
341 static bool isSafeSROAElementUse(Value *V) {
342 // We might have a dead and dangling constant hanging off of here.
343 if (Constant *C = dyn_cast<Constant>(V))
344 return ConstantIsDead(C);
346 Instruction *I = dyn_cast<Instruction>(V);
347 if (!I) return false;
350 if (isa<LoadInst>(I)) return true;
352 // Stores *to* the pointer are ok.
353 if (StoreInst *SI = dyn_cast<StoreInst>(I))
354 return SI->getOperand(0) != V;
356 // Otherwise, it must be a GEP.
357 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I);
358 if (GEPI == 0) return false;
360 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
361 !cast<Constant>(GEPI->getOperand(1))->isNullValue())
364 for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end();
366 if (!isSafeSROAElementUse(*I))
372 /// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value.
373 /// Look at it and its uses and decide whether it is safe to SROA this global.
375 static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
376 // The user of the global must be a GEP Inst or a ConstantExpr GEP.
377 if (!isa<GetElementPtrInst>(U) &&
378 (!isa<ConstantExpr>(U) ||
379 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
382 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
383 // don't like < 3 operand CE's, and we don't like non-constant integer
384 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some
386 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
387 !cast<Constant>(U->getOperand(1))->isNullValue() ||
388 !isa<ConstantInt>(U->getOperand(2)))
391 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
392 ++GEPI; // Skip over the pointer index.
394 // If this is a use of an array allocation, do a bit more checking for sanity.
395 if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
396 uint64_t NumElements = AT->getNumElements();
397 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
399 // Check to make sure that index falls within the array. If not,
400 // something funny is going on, so we won't do the optimization.
402 if (Idx->getZExtValue() >= NumElements)
405 // We cannot scalar repl this level of the array unless any array
406 // sub-indices are in-range constants. In particular, consider:
407 // A[0][i]. We cannot know that the user isn't doing invalid things like
408 // allowing i to index an out-of-range subscript that accesses A[1].
410 // Scalar replacing *just* the outer index of the array is probably not
411 // going to be a win anyway, so just give up.
412 for (++GEPI; // Skip array index.
413 GEPI != E && (isa<ArrayType>(*GEPI) || isa<VectorType>(*GEPI));
415 uint64_t NumElements;
416 if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
417 NumElements = SubArrayTy->getNumElements();
419 NumElements = cast<VectorType>(*GEPI)->getNumElements();
421 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
422 if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
427 for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I)
428 if (!isSafeSROAElementUse(*I))
433 /// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
434 /// is safe for us to perform this transformation.
436 static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
437 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
439 if (!IsUserOfGlobalSafeForSRA(*UI, GV))
446 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
447 /// variable. This opens the door for other optimizations by exposing the
448 /// behavior of the program in a more fine-grained way. We have determined that
449 /// this transformation is safe already. We return the first global variable we
450 /// insert so that the caller can reprocess it.
451 static GlobalVariable *SRAGlobal(GlobalVariable *GV) {
452 // Make sure this global only has simple uses that we can SRA.
453 if (!GlobalUsersSafeToSRA(GV))
456 assert(GV->hasInternalLinkage() && !GV->isConstant());
457 Constant *Init = GV->getInitializer();
458 const Type *Ty = Init->getType();
460 std::vector<GlobalVariable*> NewGlobals;
461 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
463 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
464 NewGlobals.reserve(STy->getNumElements());
465 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
466 Constant *In = getAggregateConstantElement(Init,
467 ConstantInt::get(Type::Int32Ty, i));
468 assert(In && "Couldn't get element of initializer?");
469 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
470 GlobalVariable::InternalLinkage,
471 In, GV->getName()+"."+utostr(i),
473 GV->isThreadLocal());
474 Globals.insert(GV, NGV);
475 NewGlobals.push_back(NGV);
477 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
478 unsigned NumElements = 0;
479 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
480 NumElements = ATy->getNumElements();
481 else if (const VectorType *PTy = dyn_cast<VectorType>(STy))
482 NumElements = PTy->getNumElements();
484 assert(0 && "Unknown aggregate sequential type!");
486 if (NumElements > 16 && GV->hasNUsesOrMore(16))
487 return 0; // It's not worth it.
488 NewGlobals.reserve(NumElements);
489 for (unsigned i = 0, e = NumElements; i != e; ++i) {
490 Constant *In = getAggregateConstantElement(Init,
491 ConstantInt::get(Type::Int32Ty, i));
492 assert(In && "Couldn't get element of initializer?");
494 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
495 GlobalVariable::InternalLinkage,
496 In, GV->getName()+"."+utostr(i),
498 GV->isThreadLocal());
499 Globals.insert(GV, NGV);
500 NewGlobals.push_back(NGV);
504 if (NewGlobals.empty())
507 DOUT << "PERFORMING GLOBAL SRA ON: " << *GV;
509 Constant *NullInt = Constant::getNullValue(Type::Int32Ty);
511 // Loop over all of the uses of the global, replacing the constantexpr geps,
512 // with smaller constantexpr geps or direct references.
513 while (!GV->use_empty()) {
514 User *GEP = GV->use_back();
515 assert(((isa<ConstantExpr>(GEP) &&
516 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
517 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
519 // Ignore the 1th operand, which has to be zero or else the program is quite
520 // broken (undefined). Get the 2nd operand, which is the structure or array
522 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
523 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
525 Value *NewPtr = NewGlobals[Val];
527 // Form a shorter GEP if needed.
528 if (GEP->getNumOperands() > 3)
529 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
530 SmallVector<Constant*, 8> Idxs;
531 Idxs.push_back(NullInt);
532 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
533 Idxs.push_back(CE->getOperand(i));
534 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr),
535 &Idxs[0], Idxs.size());
537 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
538 SmallVector<Value*, 8> Idxs;
539 Idxs.push_back(NullInt);
540 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
541 Idxs.push_back(GEPI->getOperand(i));
542 NewPtr = new GetElementPtrInst(NewPtr, Idxs.begin(), Idxs.end(),
543 GEPI->getName()+"."+utostr(Val), GEPI);
545 GEP->replaceAllUsesWith(NewPtr);
547 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
548 GEPI->eraseFromParent();
550 cast<ConstantExpr>(GEP)->destroyConstant();
553 // Delete the old global, now that it is dead.
557 // Loop over the new globals array deleting any globals that are obviously
558 // dead. This can arise due to scalarization of a structure or an array that
559 // has elements that are dead.
560 unsigned FirstGlobal = 0;
561 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
562 if (NewGlobals[i]->use_empty()) {
563 Globals.erase(NewGlobals[i]);
564 if (FirstGlobal == i) ++FirstGlobal;
567 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
570 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
571 /// value will trap if the value is dynamically null. PHIs keeps track of any
572 /// phi nodes we've seen to avoid reprocessing them.
573 static bool AllUsesOfValueWillTrapIfNull(Value *V,
574 SmallPtrSet<PHINode*, 8> &PHIs) {
575 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
576 if (isa<LoadInst>(*UI)) {
578 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
579 if (SI->getOperand(0) == V) {
580 //cerr << "NONTRAPPING USE: " << **UI;
581 return false; // Storing the value.
583 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
584 if (CI->getOperand(0) != V) {
585 //cerr << "NONTRAPPING USE: " << **UI;
586 return false; // Not calling the ptr
588 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
589 if (II->getOperand(0) != V) {
590 //cerr << "NONTRAPPING USE: " << **UI;
591 return false; // Not calling the ptr
593 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(*UI)) {
594 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
595 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
596 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
597 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
598 // If we've already seen this phi node, ignore it, it has already been
601 return AllUsesOfValueWillTrapIfNull(PN, PHIs);
602 } else if (isa<ICmpInst>(*UI) &&
603 isa<ConstantPointerNull>(UI->getOperand(1))) {
604 // Ignore setcc X, null
606 //cerr << "NONTRAPPING USE: " << **UI;
612 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
613 /// from GV will trap if the loaded value is null. Note that this also permits
614 /// comparisons of the loaded value against null, as a special case.
615 static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) {
616 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI)
617 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
618 SmallPtrSet<PHINode*, 8> PHIs;
619 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
621 } else if (isa<StoreInst>(*UI)) {
622 // Ignore stores to the global.
624 // We don't know or understand this user, bail out.
625 //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI;
632 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
633 bool Changed = false;
634 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
635 Instruction *I = cast<Instruction>(*UI++);
636 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
637 LI->setOperand(0, NewV);
639 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
640 if (SI->getOperand(1) == V) {
641 SI->setOperand(1, NewV);
644 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
645 if (I->getOperand(0) == V) {
646 // Calling through the pointer! Turn into a direct call, but be careful
647 // that the pointer is not also being passed as an argument.
648 I->setOperand(0, NewV);
650 bool PassedAsArg = false;
651 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
652 if (I->getOperand(i) == V) {
654 I->setOperand(i, NewV);
658 // Being passed as an argument also. Be careful to not invalidate UI!
662 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
663 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
664 ConstantExpr::getCast(CI->getOpcode(),
665 NewV, CI->getType()));
666 if (CI->use_empty()) {
668 CI->eraseFromParent();
670 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
671 // Should handle GEP here.
672 SmallVector<Constant*, 8> Idxs;
673 Idxs.reserve(GEPI->getNumOperands()-1);
674 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
675 if (Constant *C = dyn_cast<Constant>(GEPI->getOperand(i)))
679 if (Idxs.size() == GEPI->getNumOperands()-1)
680 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
681 ConstantExpr::getGetElementPtr(NewV, &Idxs[0],
683 if (GEPI->use_empty()) {
685 GEPI->eraseFromParent();
694 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
695 /// value stored into it. If there are uses of the loaded value that would trap
696 /// if the loaded value is dynamically null, then we know that they cannot be
697 /// reachable with a null optimize away the load.
698 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) {
699 std::vector<LoadInst*> Loads;
700 bool Changed = false;
702 // Replace all uses of loads with uses of uses of the stored value.
703 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end();
705 if (LoadInst *LI = dyn_cast<LoadInst>(*GUI)) {
707 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
709 // If we get here we could have stores, selects, or phi nodes whose values
711 assert((isa<StoreInst>(*GUI) || isa<PHINode>(*GUI) ||
712 isa<SelectInst>(*GUI) || isa<ConstantExpr>(*GUI)) &&
713 "Only expect load and stores!");
717 DOUT << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV;
721 // Delete all of the loads we can, keeping track of whether we nuked them all!
722 bool AllLoadsGone = true;
723 while (!Loads.empty()) {
724 LoadInst *L = Loads.back();
725 if (L->use_empty()) {
726 L->eraseFromParent();
729 AllLoadsGone = false;
734 // If we nuked all of the loads, then none of the stores are needed either,
735 // nor is the global.
737 DOUT << " *** GLOBAL NOW DEAD!\n";
738 CleanupConstantGlobalUsers(GV, 0);
739 if (GV->use_empty()) {
740 GV->eraseFromParent();
748 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
749 /// instructions that are foldable.
750 static void ConstantPropUsersOf(Value *V) {
751 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
752 if (Instruction *I = dyn_cast<Instruction>(*UI++))
753 if (Constant *NewC = ConstantFoldInstruction(I)) {
754 I->replaceAllUsesWith(NewC);
756 // Advance UI to the next non-I use to avoid invalidating it!
757 // Instructions could multiply use V.
758 while (UI != E && *UI == I)
760 I->eraseFromParent();
764 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
765 /// variable, and transforms the program as if it always contained the result of
766 /// the specified malloc. Because it is always the result of the specified
767 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
768 /// malloc into a global, and any loads of GV as uses of the new global.
769 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
771 DOUT << "PROMOTING MALLOC GLOBAL: " << *GV << " MALLOC = " << *MI;
772 ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize());
774 if (NElements->getZExtValue() != 1) {
775 // If we have an array allocation, transform it to a single element
776 // allocation to make the code below simpler.
777 Type *NewTy = ArrayType::get(MI->getAllocatedType(),
778 NElements->getZExtValue());
780 new MallocInst(NewTy, Constant::getNullValue(Type::Int32Ty),
781 MI->getAlignment(), MI->getName(), MI);
783 Indices[0] = Indices[1] = Constant::getNullValue(Type::Int32Ty);
784 Value *NewGEP = new GetElementPtrInst(NewMI, Indices, Indices + 2,
785 NewMI->getName()+".el0", MI);
786 MI->replaceAllUsesWith(NewGEP);
787 MI->eraseFromParent();
791 // Create the new global variable. The contents of the malloc'd memory is
792 // undefined, so initialize with an undef value.
793 Constant *Init = UndefValue::get(MI->getAllocatedType());
794 GlobalVariable *NewGV = new GlobalVariable(MI->getAllocatedType(), false,
795 GlobalValue::InternalLinkage, Init,
796 GV->getName()+".body",
798 GV->isThreadLocal());
799 GV->getParent()->getGlobalList().insert(GV, NewGV);
801 // Anything that used the malloc now uses the global directly.
802 MI->replaceAllUsesWith(NewGV);
804 Constant *RepValue = NewGV;
805 if (NewGV->getType() != GV->getType()->getElementType())
806 RepValue = ConstantExpr::getBitCast(RepValue,
807 GV->getType()->getElementType());
809 // If there is a comparison against null, we will insert a global bool to
810 // keep track of whether the global was initialized yet or not.
811 GlobalVariable *InitBool =
812 new GlobalVariable(Type::Int1Ty, false, GlobalValue::InternalLinkage,
813 ConstantInt::getFalse(), GV->getName()+".init",
814 (Module *)NULL, GV->isThreadLocal());
815 bool InitBoolUsed = false;
817 // Loop over all uses of GV, processing them in turn.
818 std::vector<StoreInst*> Stores;
819 while (!GV->use_empty())
820 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
821 while (!LI->use_empty()) {
822 Use &LoadUse = LI->use_begin().getUse();
823 if (!isa<ICmpInst>(LoadUse.getUser()))
826 ICmpInst *CI = cast<ICmpInst>(LoadUse.getUser());
827 // Replace the cmp X, 0 with a use of the bool value.
828 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", CI);
830 switch (CI->getPredicate()) {
831 default: assert(0 && "Unknown ICmp Predicate!");
832 case ICmpInst::ICMP_ULT:
833 case ICmpInst::ICMP_SLT:
834 LV = ConstantInt::getFalse(); // X < null -> always false
836 case ICmpInst::ICMP_ULE:
837 case ICmpInst::ICMP_SLE:
838 case ICmpInst::ICMP_EQ:
839 LV = BinaryOperator::createNot(LV, "notinit", CI);
841 case ICmpInst::ICMP_NE:
842 case ICmpInst::ICMP_UGE:
843 case ICmpInst::ICMP_SGE:
844 case ICmpInst::ICMP_UGT:
845 case ICmpInst::ICMP_SGT:
848 CI->replaceAllUsesWith(LV);
849 CI->eraseFromParent();
852 LI->eraseFromParent();
854 StoreInst *SI = cast<StoreInst>(GV->use_back());
855 // The global is initialized when the store to it occurs.
856 new StoreInst(ConstantInt::getTrue(), InitBool, SI);
857 SI->eraseFromParent();
860 // If the initialization boolean was used, insert it, otherwise delete it.
862 while (!InitBool->use_empty()) // Delete initializations
863 cast<Instruction>(InitBool->use_back())->eraseFromParent();
866 GV->getParent()->getGlobalList().insert(GV, InitBool);
869 // Now the GV is dead, nuke it and the malloc.
870 GV->eraseFromParent();
871 MI->eraseFromParent();
873 // To further other optimizations, loop over all users of NewGV and try to
874 // constant prop them. This will promote GEP instructions with constant
875 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
876 ConstantPropUsersOf(NewGV);
877 if (RepValue != NewGV)
878 ConstantPropUsersOf(RepValue);
883 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
884 /// to make sure that there are no complex uses of V. We permit simple things
885 /// like dereferencing the pointer, but not storing through the address, unless
886 /// it is to the specified global.
887 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V,
889 SmallPtrSet<PHINode*, 8> &PHIs) {
890 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
891 if (isa<LoadInst>(*UI) || isa<CmpInst>(*UI)) {
893 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
894 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
895 return false; // Storing the pointer itself... bad.
896 // Otherwise, storing through it, or storing into GV... fine.
897 } else if (isa<GetElementPtrInst>(*UI)) {
898 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(cast<Instruction>(*UI),
901 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
902 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI
905 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
913 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
914 /// somewhere. Transform all uses of the allocation into loads from the
915 /// global and uses of the resultant pointer. Further, delete the store into
916 /// GV. This assumes that these value pass the
917 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
918 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
919 GlobalVariable *GV) {
920 while (!Alloc->use_empty()) {
921 Instruction *U = cast<Instruction>(*Alloc->use_begin());
922 Instruction *InsertPt = U;
923 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
924 // If this is the store of the allocation into the global, remove it.
925 if (SI->getOperand(1) == GV) {
926 SI->eraseFromParent();
929 } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
930 // Insert the load in the corresponding predecessor, not right before the
932 unsigned PredNo = Alloc->use_begin().getOperandNo()/2;
933 InsertPt = PN->getIncomingBlock(PredNo)->getTerminator();
936 // Insert a load from the global, and use it instead of the malloc.
937 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
938 U->replaceUsesOfWith(Alloc, NL);
942 /// GlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
943 /// GV are simple enough to perform HeapSRA, return true.
944 static bool GlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV,
946 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;
948 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
949 // We permit two users of the load: setcc comparing against the null
950 // pointer, and a getelementptr of a specific form.
951 for (Value::use_iterator UI = LI->use_begin(), E = LI->use_end(); UI != E;
953 // Comparison against null is ok.
954 if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) {
955 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
960 // getelementptr is also ok, but only a simple form.
961 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
962 // Must index into the array and into the struct.
963 if (GEPI->getNumOperands() < 3)
966 // Otherwise the GEP is ok.
970 if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
971 // We have a phi of a load from the global. We can only handle this
972 // if the other PHI'd values are actually the same. In this case,
973 // the rewriter will just drop the phi entirely.
974 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
975 Value *IV = PN->getIncomingValue(i);
976 if (IV == LI) continue; // Trivial the same.
978 // If the phi'd value is from the malloc that initializes the value,
980 if (IV == MI) continue;
982 // Otherwise, we don't know what it is.
988 // Otherwise we don't know what this is, not ok.
995 /// GetHeapSROALoad - Return the load for the specified field of the HeapSROA'd
996 /// value, lazily creating it on demand.
997 static Value *GetHeapSROALoad(Instruction *Load, unsigned FieldNo,
998 const std::vector<GlobalVariable*> &FieldGlobals,
999 std::vector<Value *> &InsertedLoadsForPtr) {
1000 if (InsertedLoadsForPtr.size() <= FieldNo)
1001 InsertedLoadsForPtr.resize(FieldNo+1);
1002 if (InsertedLoadsForPtr[FieldNo] == 0)
1003 InsertedLoadsForPtr[FieldNo] = new LoadInst(FieldGlobals[FieldNo],
1004 Load->getName()+".f" +
1005 utostr(FieldNo), Load);
1006 return InsertedLoadsForPtr[FieldNo];
1009 /// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1010 /// the load, rewrite the derived value to use the HeapSRoA'd load.
1011 static void RewriteHeapSROALoadUser(LoadInst *Load, Instruction *LoadUser,
1012 const std::vector<GlobalVariable*> &FieldGlobals,
1013 std::vector<Value *> &InsertedLoadsForPtr) {
1014 // If this is a comparison against null, handle it.
1015 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1016 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1017 // If we have a setcc of the loaded pointer, we can use a setcc of any
1020 if (InsertedLoadsForPtr.empty()) {
1021 NPtr = GetHeapSROALoad(Load, 0, FieldGlobals, InsertedLoadsForPtr);
1023 NPtr = InsertedLoadsForPtr.back();
1026 Value *New = new ICmpInst(SCI->getPredicate(), NPtr,
1027 Constant::getNullValue(NPtr->getType()),
1028 SCI->getName(), SCI);
1029 SCI->replaceAllUsesWith(New);
1030 SCI->eraseFromParent();
1034 // Handle 'getelementptr Ptr, Idx, uint FieldNo ...'
1035 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1036 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1037 && "Unexpected GEPI!");
1039 // Load the pointer for this field.
1040 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1041 Value *NewPtr = GetHeapSROALoad(Load, FieldNo,
1042 FieldGlobals, InsertedLoadsForPtr);
1044 // Create the new GEP idx vector.
1045 SmallVector<Value*, 8> GEPIdx;
1046 GEPIdx.push_back(GEPI->getOperand(1));
1047 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1049 Value *NGEPI = new GetElementPtrInst(NewPtr, GEPIdx.begin(), GEPIdx.end(),
1050 GEPI->getName(), GEPI);
1051 GEPI->replaceAllUsesWith(NGEPI);
1052 GEPI->eraseFromParent();
1056 // Handle PHI nodes. PHI nodes must be merging in the same values, plus
1057 // potentially the original malloc. Insert phi nodes for each field, then
1058 // process uses of the PHI.
1059 PHINode *PN = cast<PHINode>(LoadUser);
1060 std::vector<Value *> PHIsForField;
1061 PHIsForField.resize(FieldGlobals.size());
1062 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1063 Value *LoadV = GetHeapSROALoad(Load, i, FieldGlobals, InsertedLoadsForPtr);
1065 PHINode *FieldPN = new PHINode(LoadV->getType(),
1066 PN->getName()+"."+utostr(i), PN);
1067 // Fill in the predecessor values.
1068 for (unsigned pred = 0, e = PN->getNumIncomingValues(); pred != e; ++pred) {
1069 // Each predecessor either uses the load or the original malloc.
1070 Value *InVal = PN->getIncomingValue(pred);
1071 BasicBlock *BB = PN->getIncomingBlock(pred);
1073 if (isa<MallocInst>(InVal)) {
1074 // Insert a reload from the global in the predecessor.
1075 NewVal = GetHeapSROALoad(BB->getTerminator(), i, FieldGlobals,
1078 NewVal = InsertedLoadsForPtr[i];
1080 FieldPN->addIncoming(NewVal, BB);
1082 PHIsForField[i] = FieldPN;
1085 // Since PHIsForField specifies a phi for every input value, the lazy inserter
1086 // will never insert a load.
1087 while (!PN->use_empty())
1088 RewriteHeapSROALoadUser(Load, PN->use_back(), FieldGlobals, PHIsForField);
1089 PN->eraseFromParent();
1092 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1093 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1094 /// use FieldGlobals instead. All uses of loaded values satisfy
1095 /// GlobalLoadUsesSimpleEnoughForHeapSRA.
1096 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1097 const std::vector<GlobalVariable*> &FieldGlobals) {
1098 std::vector<Value *> InsertedLoadsForPtr;
1099 //InsertedLoadsForPtr.resize(FieldGlobals.size());
1100 while (!Load->use_empty())
1101 RewriteHeapSROALoadUser(Load, Load->use_back(),
1102 FieldGlobals, InsertedLoadsForPtr);
1105 /// PerformHeapAllocSRoA - MI is an allocation of an array of structures. Break
1106 /// it up into multiple allocations of arrays of the fields.
1107 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){
1108 DOUT << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *MI;
1109 const StructType *STy = cast<StructType>(MI->getAllocatedType());
1111 // There is guaranteed to be at least one use of the malloc (storing
1112 // it into GV). If there are other uses, change them to be uses of
1113 // the global to simplify later code. This also deletes the store
1115 ReplaceUsesOfMallocWithGlobal(MI, GV);
1117 // Okay, at this point, there are no users of the malloc. Insert N
1118 // new mallocs at the same place as MI, and N globals.
1119 std::vector<GlobalVariable*> FieldGlobals;
1120 std::vector<MallocInst*> FieldMallocs;
1122 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1123 const Type *FieldTy = STy->getElementType(FieldNo);
1124 const Type *PFieldTy = PointerType::getUnqual(FieldTy);
1126 GlobalVariable *NGV =
1127 new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage,
1128 Constant::getNullValue(PFieldTy),
1129 GV->getName() + ".f" + utostr(FieldNo), GV,
1130 GV->isThreadLocal());
1131 FieldGlobals.push_back(NGV);
1133 MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(),
1134 MI->getName() + ".f" + utostr(FieldNo),MI);
1135 FieldMallocs.push_back(NMI);
1136 new StoreInst(NMI, NGV, MI);
1139 // The tricky aspect of this transformation is handling the case when malloc
1140 // fails. In the original code, malloc failing would set the result pointer
1141 // of malloc to null. In this case, some mallocs could succeed and others
1142 // could fail. As such, we emit code that looks like this:
1143 // F0 = malloc(field0)
1144 // F1 = malloc(field1)
1145 // F2 = malloc(field2)
1146 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1147 // if (F0) { free(F0); F0 = 0; }
1148 // if (F1) { free(F1); F1 = 0; }
1149 // if (F2) { free(F2); F2 = 0; }
1151 Value *RunningOr = 0;
1152 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1153 Value *Cond = new ICmpInst(ICmpInst::ICMP_EQ, FieldMallocs[i],
1154 Constant::getNullValue(FieldMallocs[i]->getType()),
1157 RunningOr = Cond; // First seteq
1159 RunningOr = BinaryOperator::createOr(RunningOr, Cond, "tmp", MI);
1162 // Split the basic block at the old malloc.
1163 BasicBlock *OrigBB = MI->getParent();
1164 BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont");
1166 // Create the block to check the first condition. Put all these blocks at the
1167 // end of the function as they are unlikely to be executed.
1168 BasicBlock *NullPtrBlock = new BasicBlock("malloc_ret_null",
1169 OrigBB->getParent());
1171 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1172 // branch on RunningOr.
1173 OrigBB->getTerminator()->eraseFromParent();
1174 new BranchInst(NullPtrBlock, ContBB, RunningOr, OrigBB);
1176 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1177 // pointer, because some may be null while others are not.
1178 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1179 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1180 Value *Cmp = new ICmpInst(ICmpInst::ICMP_NE, GVVal,
1181 Constant::getNullValue(GVVal->getType()),
1182 "tmp", NullPtrBlock);
1183 BasicBlock *FreeBlock = new BasicBlock("free_it", OrigBB->getParent());
1184 BasicBlock *NextBlock = new BasicBlock("next", OrigBB->getParent());
1185 new BranchInst(FreeBlock, NextBlock, Cmp, NullPtrBlock);
1187 // Fill in FreeBlock.
1188 new FreeInst(GVVal, FreeBlock);
1189 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1191 new BranchInst(NextBlock, FreeBlock);
1193 NullPtrBlock = NextBlock;
1196 new BranchInst(ContBB, NullPtrBlock);
1199 // MI is no longer needed, remove it.
1200 MI->eraseFromParent();
1203 // Okay, the malloc site is completely handled. All of the uses of GV are now
1204 // loads, and all uses of those loads are simple. Rewrite them to use loads
1205 // of the per-field globals instead.
1206 while (!GV->use_empty()) {
1207 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
1208 RewriteUsesOfLoadForHeapSRoA(LI, FieldGlobals);
1209 LI->eraseFromParent();
1211 // Must be a store of null.
1212 StoreInst *SI = cast<StoreInst>(GV->use_back());
1213 assert(isa<Constant>(SI->getOperand(0)) &&
1214 cast<Constant>(SI->getOperand(0))->isNullValue() &&
1215 "Unexpected heap-sra user!");
1217 // Insert a store of null into each global.
1218 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1220 Constant::getNullValue(FieldGlobals[i]->getType()->getElementType());
1221 new StoreInst(Null, FieldGlobals[i], SI);
1223 // Erase the original store.
1224 SI->eraseFromParent();
1228 // The old global is now dead, remove it.
1229 GV->eraseFromParent();
1232 return FieldGlobals[0];
1236 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1237 // that only one value (besides its initializer) is ever stored to the global.
1238 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1239 Module::global_iterator &GVI,
1241 if (CastInst *CI = dyn_cast<CastInst>(StoredOnceVal))
1242 StoredOnceVal = CI->getOperand(0);
1243 else if (GetElementPtrInst *GEPI =dyn_cast<GetElementPtrInst>(StoredOnceVal)){
1244 // "getelementptr Ptr, 0, 0, 0" is really just a cast.
1245 bool IsJustACast = true;
1246 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
1247 if (!isa<Constant>(GEPI->getOperand(i)) ||
1248 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
1249 IsJustACast = false;
1253 StoredOnceVal = GEPI->getOperand(0);
1256 // If we are dealing with a pointer global that is initialized to null and
1257 // only has one (non-null) value stored into it, then we can optimize any
1258 // users of the loaded value (often calls and loads) that would trap if the
1260 if (isa<PointerType>(GV->getInitializer()->getType()) &&
1261 GV->getInitializer()->isNullValue()) {
1262 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1263 if (GV->getInitializer()->getType() != SOVC->getType())
1264 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1266 // Optimize away any trapping uses of the loaded value.
1267 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1269 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) {
1270 // If this is a malloc of an abstract type, don't touch it.
1271 if (!MI->getAllocatedType()->isSized())
1274 // We can't optimize this global unless all uses of it are *known* to be
1275 // of the malloc value, not of the null initializer value (consider a use
1276 // that compares the global's value against zero to see if the malloc has
1277 // been reached). To do this, we check to see if all uses of the global
1278 // would trap if the global were null: this proves that they must all
1279 // happen after the malloc.
1280 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1283 // We can't optimize this if the malloc itself is used in a complex way,
1284 // for example, being stored into multiple globals. This allows the
1285 // malloc to be stored into the specified global, loaded setcc'd, and
1286 // GEP'd. These are all things we could transform to using the global
1289 SmallPtrSet<PHINode*, 8> PHIs;
1290 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV, PHIs))
1295 // If we have a global that is only initialized with a fixed size malloc,
1296 // transform the program to use global memory instead of malloc'd memory.
1297 // This eliminates dynamic allocation, avoids an indirection accessing the
1298 // data, and exposes the resultant global to further GlobalOpt.
1299 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) {
1300 // Restrict this transformation to only working on small allocations
1301 // (2048 bytes currently), as we don't want to introduce a 16M global or
1303 if (NElements->getZExtValue()*
1304 TD.getABITypeSize(MI->getAllocatedType()) < 2048) {
1305 GVI = OptimizeGlobalAddressOfMalloc(GV, MI);
1310 // If the allocation is an array of structures, consider transforming this
1311 // into multiple malloc'd arrays, one for each field. This is basically
1312 // SRoA for malloc'd memory.
1313 if (const StructType *AllocTy =
1314 dyn_cast<StructType>(MI->getAllocatedType())) {
1315 // This the structure has an unreasonable number of fields, leave it
1317 if (AllocTy->getNumElements() <= 16 && AllocTy->getNumElements() > 0 &&
1318 GlobalLoadUsesSimpleEnoughForHeapSRA(GV, MI)) {
1319 GVI = PerformHeapAllocSRoA(GV, MI);
1329 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1330 /// two values ever stored into GV are its initializer and OtherVal. See if we
1331 /// can shrink the global into a boolean and select between the two values
1332 /// whenever it is used. This exposes the values to other scalar optimizations.
1333 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1334 const Type *GVElType = GV->getType()->getElementType();
1336 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1337 // an FP value or vector, don't do this optimization because a select between
1338 // them is very expensive and unlikely to lead to later simplification.
1339 if (GVElType == Type::Int1Ty || GVElType->isFloatingPoint() ||
1340 isa<VectorType>(GVElType))
1343 // Walk the use list of the global seeing if all the uses are load or store.
1344 // If there is anything else, bail out.
1345 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I)
1346 if (!isa<LoadInst>(I) && !isa<StoreInst>(I))
1349 DOUT << " *** SHRINKING TO BOOL: " << *GV;
1351 // Create the new global, initializing it to false.
1352 GlobalVariable *NewGV = new GlobalVariable(Type::Int1Ty, false,
1353 GlobalValue::InternalLinkage, ConstantInt::getFalse(),
1356 GV->isThreadLocal());
1357 GV->getParent()->getGlobalList().insert(GV, NewGV);
1359 Constant *InitVal = GV->getInitializer();
1360 assert(InitVal->getType() != Type::Int1Ty && "No reason to shrink to bool!");
1362 // If initialized to zero and storing one into the global, we can use a cast
1363 // instead of a select to synthesize the desired value.
1364 bool IsOneZero = false;
1365 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1366 IsOneZero = InitVal->isNullValue() && CI->isOne();
1368 while (!GV->use_empty()) {
1369 Instruction *UI = cast<Instruction>(GV->use_back());
1370 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1371 // Change the store into a boolean store.
1372 bool StoringOther = SI->getOperand(0) == OtherVal;
1373 // Only do this if we weren't storing a loaded value.
1375 if (StoringOther || SI->getOperand(0) == InitVal)
1376 StoreVal = ConstantInt::get(Type::Int1Ty, StoringOther);
1378 // Otherwise, we are storing a previously loaded copy. To do this,
1379 // change the copy from copying the original value to just copying the
1381 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1383 // If we're already replaced the input, StoredVal will be a cast or
1384 // select instruction. If not, it will be a load of the original
1386 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1387 assert(LI->getOperand(0) == GV && "Not a copy!");
1388 // Insert a new load, to preserve the saved value.
1389 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1391 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1392 "This is not a form that we understand!");
1393 StoreVal = StoredVal->getOperand(0);
1394 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1397 new StoreInst(StoreVal, NewGV, SI);
1399 // Change the load into a load of bool then a select.
1400 LoadInst *LI = cast<LoadInst>(UI);
1401 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1404 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1406 NSI = new SelectInst(NLI, OtherVal, InitVal, "", LI);
1408 LI->replaceAllUsesWith(NSI);
1410 UI->eraseFromParent();
1413 GV->eraseFromParent();
1418 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1419 /// it if possible. If we make a change, return true.
1420 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1421 Module::global_iterator &GVI) {
1422 std::set<PHINode*> PHIUsers;
1424 GV->removeDeadConstantUsers();
1426 if (GV->use_empty()) {
1427 DOUT << "GLOBAL DEAD: " << *GV;
1428 GV->eraseFromParent();
1433 if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
1435 cerr << "Global: " << *GV;
1436 cerr << " isLoaded = " << GS.isLoaded << "\n";
1437 cerr << " StoredType = ";
1438 switch (GS.StoredType) {
1439 case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break;
1440 case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break;
1441 case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break;
1442 case GlobalStatus::isStored: cerr << "stored\n"; break;
1444 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
1445 cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n";
1446 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
1447 cerr << " AccessingFunction = " << GS.AccessingFunction->getName()
1449 cerr << " HasMultipleAccessingFunctions = "
1450 << GS.HasMultipleAccessingFunctions << "\n";
1451 cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n";
1455 // If this is a first class global and has only one accessing function
1456 // and this function is main (which we know is not recursive we can make
1457 // this global a local variable) we replace the global with a local alloca
1458 // in this function.
1460 // NOTE: It doesn't make sense to promote non first class types since we
1461 // are just replacing static memory to stack memory.
1462 if (!GS.HasMultipleAccessingFunctions &&
1463 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1464 GV->getType()->getElementType()->isFirstClassType() &&
1465 GS.AccessingFunction->getName() == "main" &&
1466 GS.AccessingFunction->hasExternalLinkage()) {
1467 DOUT << "LOCALIZING GLOBAL: " << *GV;
1468 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin();
1469 const Type* ElemTy = GV->getType()->getElementType();
1470 // FIXME: Pass Global's alignment when globals have alignment
1471 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI);
1472 if (!isa<UndefValue>(GV->getInitializer()))
1473 new StoreInst(GV->getInitializer(), Alloca, FirstI);
1475 GV->replaceAllUsesWith(Alloca);
1476 GV->eraseFromParent();
1481 // If the global is never loaded (but may be stored to), it is dead.
1484 DOUT << "GLOBAL NEVER LOADED: " << *GV;
1486 // Delete any stores we can find to the global. We may not be able to
1487 // make it completely dead though.
1488 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1490 // If the global is dead now, delete it.
1491 if (GV->use_empty()) {
1492 GV->eraseFromParent();
1498 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1499 DOUT << "MARKING CONSTANT: " << *GV;
1500 GV->setConstant(true);
1502 // Clean up any obviously simplifiable users now.
1503 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1505 // If the global is dead now, just nuke it.
1506 if (GV->use_empty()) {
1507 DOUT << " *** Marking constant allowed us to simplify "
1508 << "all users and delete global!\n";
1509 GV->eraseFromParent();
1515 } else if (!GV->getInitializer()->getType()->isFirstClassType()) {
1516 if (GlobalVariable *FirstNewGV = SRAGlobal(GV)) {
1517 GVI = FirstNewGV; // Don't skip the newly produced globals!
1520 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1521 // If the initial value for the global was an undef value, and if only
1522 // one other value was stored into it, we can just change the
1523 // initializer to be an undef value, then delete all stores to the
1524 // global. This allows us to mark it constant.
1525 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1526 if (isa<UndefValue>(GV->getInitializer())) {
1527 // Change the initial value here.
1528 GV->setInitializer(SOVConstant);
1530 // Clean up any obviously simplifiable users now.
1531 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1533 if (GV->use_empty()) {
1534 DOUT << " *** Substituting initializer allowed us to "
1535 << "simplify all users and delete global!\n";
1536 GV->eraseFromParent();
1545 // Try to optimize globals based on the knowledge that only one value
1546 // (besides its initializer) is ever stored to the global.
1547 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1548 getAnalysis<TargetData>()))
1551 // Otherwise, if the global was not a boolean, we can shrink it to be a
1553 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1554 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1563 /// OnlyCalledDirectly - Return true if the specified function is only called
1564 /// directly. In other words, its address is never taken.
1565 static bool OnlyCalledDirectly(Function *F) {
1566 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1567 Instruction *User = dyn_cast<Instruction>(*UI);
1568 if (!User) return false;
1569 if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false;
1571 // See if the function address is passed as an argument.
1572 for (unsigned i = 1, e = User->getNumOperands(); i != e; ++i)
1573 if (User->getOperand(i) == F) return false;
1578 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1579 /// function, changing them to FastCC.
1580 static void ChangeCalleesToFastCall(Function *F) {
1581 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1582 Instruction *User = cast<Instruction>(*UI);
1583 if (CallInst *CI = dyn_cast<CallInst>(User))
1584 CI->setCallingConv(CallingConv::Fast);
1586 cast<InvokeInst>(User)->setCallingConv(CallingConv::Fast);
1590 bool GlobalOpt::OptimizeFunctions(Module &M) {
1591 bool Changed = false;
1592 // Optimize functions.
1593 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1595 F->removeDeadConstantUsers();
1596 if (F->use_empty() && (F->hasInternalLinkage() ||
1597 F->hasLinkOnceLinkage())) {
1598 M.getFunctionList().erase(F);
1601 } else if (F->hasInternalLinkage() &&
1602 F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1603 OnlyCalledDirectly(F)) {
1604 // If this function has C calling conventions, is not a varargs
1605 // function, and is only called directly, promote it to use the Fast
1606 // calling convention.
1607 F->setCallingConv(CallingConv::Fast);
1608 ChangeCalleesToFastCall(F);
1616 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1617 bool Changed = false;
1618 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1620 GlobalVariable *GV = GVI++;
1621 if (!GV->isConstant() && GV->hasInternalLinkage() &&
1622 GV->hasInitializer())
1623 Changed |= ProcessInternalGlobal(GV, GVI);
1628 /// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1629 /// initializers have an init priority of 65535.
1630 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1631 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1633 if (I->getName() == "llvm.global_ctors") {
1634 // Found it, verify it's an array of { int, void()* }.
1635 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1637 const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1638 if (!STy || STy->getNumElements() != 2 ||
1639 STy->getElementType(0) != Type::Int32Ty) return 0;
1640 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1641 if (!PFTy) return 0;
1642 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1643 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() ||
1644 FTy->getNumParams() != 0)
1647 // Verify that the initializer is simple enough for us to handle.
1648 if (!I->hasInitializer()) return 0;
1649 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1651 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1652 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(CA->getOperand(i))) {
1653 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1656 // Must have a function or null ptr.
1657 if (!isa<Function>(CS->getOperand(1)))
1660 // Init priority must be standard.
1661 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1662 if (!CI || CI->getZExtValue() != 65535)
1673 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1674 /// return a list of the functions and null terminator as a vector.
1675 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1676 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1677 std::vector<Function*> Result;
1678 Result.reserve(CA->getNumOperands());
1679 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) {
1680 ConstantStruct *CS = cast<ConstantStruct>(CA->getOperand(i));
1681 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1686 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1687 /// specified array, returning the new global to use.
1688 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1689 const std::vector<Function*> &Ctors) {
1690 // If we made a change, reassemble the initializer list.
1691 std::vector<Constant*> CSVals;
1692 CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535));
1693 CSVals.push_back(0);
1695 // Create the new init list.
1696 std::vector<Constant*> CAList;
1697 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1699 CSVals[1] = Ctors[i];
1701 const Type *FTy = FunctionType::get(Type::VoidTy,
1702 std::vector<const Type*>(), false);
1703 const PointerType *PFTy = PointerType::getUnqual(FTy);
1704 CSVals[1] = Constant::getNullValue(PFTy);
1705 CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647);
1707 CAList.push_back(ConstantStruct::get(CSVals));
1710 // Create the array initializer.
1711 const Type *StructTy =
1712 cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
1713 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()),
1716 // If we didn't change the number of elements, don't create a new GV.
1717 if (CA->getType() == GCL->getInitializer()->getType()) {
1718 GCL->setInitializer(CA);
1722 // Create the new global and insert it next to the existing list.
1723 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
1724 GCL->getLinkage(), CA, "",
1726 GCL->isThreadLocal());
1727 GCL->getParent()->getGlobalList().insert(GCL, NGV);
1730 // Nuke the old list, replacing any uses with the new one.
1731 if (!GCL->use_empty()) {
1733 if (V->getType() != GCL->getType())
1734 V = ConstantExpr::getBitCast(V, GCL->getType());
1735 GCL->replaceAllUsesWith(V);
1737 GCL->eraseFromParent();
1746 static Constant *getVal(std::map<Value*, Constant*> &ComputedValues,
1748 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
1749 Constant *R = ComputedValues[V];
1750 assert(R && "Reference to an uncomputed value!");
1754 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
1755 /// enough for us to understand. In particular, if it is a cast of something,
1756 /// we punt. We basically just support direct accesses to globals and GEP's of
1757 /// globals. This should be kept up to date with CommitValueTo.
1758 static bool isSimpleEnoughPointerToCommit(Constant *C) {
1759 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
1760 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1761 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1762 return !GV->isDeclaration(); // reject external globals.
1764 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
1765 // Handle a constantexpr gep.
1766 if (CE->getOpcode() == Instruction::GetElementPtr &&
1767 isa<GlobalVariable>(CE->getOperand(0))) {
1768 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1769 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1770 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1771 return GV->hasInitializer() &&
1772 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1777 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
1778 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
1779 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
1780 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
1781 ConstantExpr *Addr, unsigned OpNo) {
1782 // Base case of the recursion.
1783 if (OpNo == Addr->getNumOperands()) {
1784 assert(Val->getType() == Init->getType() && "Type mismatch!");
1788 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
1789 std::vector<Constant*> Elts;
1791 // Break up the constant into its elements.
1792 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
1793 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i)
1794 Elts.push_back(CS->getOperand(i));
1795 } else if (isa<ConstantAggregateZero>(Init)) {
1796 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1797 Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
1798 } else if (isa<UndefValue>(Init)) {
1799 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1800 Elts.push_back(UndefValue::get(STy->getElementType(i)));
1802 assert(0 && "This code is out of sync with "
1803 " ConstantFoldLoadThroughGEPConstantExpr");
1806 // Replace the element that we are supposed to.
1807 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
1808 unsigned Idx = CU->getZExtValue();
1809 assert(Idx < STy->getNumElements() && "Struct index out of range!");
1810 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
1812 // Return the modified struct.
1813 return ConstantStruct::get(&Elts[0], Elts.size(), STy->isPacked());
1815 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
1816 const ArrayType *ATy = cast<ArrayType>(Init->getType());
1818 // Break up the array into elements.
1819 std::vector<Constant*> Elts;
1820 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
1821 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1822 Elts.push_back(CA->getOperand(i));
1823 } else if (isa<ConstantAggregateZero>(Init)) {
1824 Constant *Elt = Constant::getNullValue(ATy->getElementType());
1825 Elts.assign(ATy->getNumElements(), Elt);
1826 } else if (isa<UndefValue>(Init)) {
1827 Constant *Elt = UndefValue::get(ATy->getElementType());
1828 Elts.assign(ATy->getNumElements(), Elt);
1830 assert(0 && "This code is out of sync with "
1831 " ConstantFoldLoadThroughGEPConstantExpr");
1834 assert(CI->getZExtValue() < ATy->getNumElements());
1835 Elts[CI->getZExtValue()] =
1836 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
1837 return ConstantArray::get(ATy, Elts);
1841 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
1842 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
1843 static void CommitValueTo(Constant *Val, Constant *Addr) {
1844 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
1845 assert(GV->hasInitializer());
1846 GV->setInitializer(Val);
1850 ConstantExpr *CE = cast<ConstantExpr>(Addr);
1851 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1853 Constant *Init = GV->getInitializer();
1854 Init = EvaluateStoreInto(Init, Val, CE, 2);
1855 GV->setInitializer(Init);
1858 /// ComputeLoadResult - Return the value that would be computed by a load from
1859 /// P after the stores reflected by 'memory' have been performed. If we can't
1860 /// decide, return null.
1861 static Constant *ComputeLoadResult(Constant *P,
1862 const std::map<Constant*, Constant*> &Memory) {
1863 // If this memory location has been recently stored, use the stored value: it
1864 // is the most up-to-date.
1865 std::map<Constant*, Constant*>::const_iterator I = Memory.find(P);
1866 if (I != Memory.end()) return I->second;
1869 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
1870 if (GV->hasInitializer())
1871 return GV->getInitializer();
1875 // Handle a constantexpr getelementptr.
1876 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
1877 if (CE->getOpcode() == Instruction::GetElementPtr &&
1878 isa<GlobalVariable>(CE->getOperand(0))) {
1879 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1880 if (GV->hasInitializer())
1881 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1884 return 0; // don't know how to evaluate.
1887 /// EvaluateFunction - Evaluate a call to function F, returning true if
1888 /// successful, false if we can't evaluate it. ActualArgs contains the formal
1889 /// arguments for the function.
1890 static bool EvaluateFunction(Function *F, Constant *&RetVal,
1891 const std::vector<Constant*> &ActualArgs,
1892 std::vector<Function*> &CallStack,
1893 std::map<Constant*, Constant*> &MutatedMemory,
1894 std::vector<GlobalVariable*> &AllocaTmps) {
1895 // Check to see if this function is already executing (recursion). If so,
1896 // bail out. TODO: we might want to accept limited recursion.
1897 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
1900 CallStack.push_back(F);
1902 /// Values - As we compute SSA register values, we store their contents here.
1903 std::map<Value*, Constant*> Values;
1905 // Initialize arguments to the incoming values specified.
1907 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
1909 Values[AI] = ActualArgs[ArgNo];
1911 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
1912 /// we can only evaluate any one basic block at most once. This set keeps
1913 /// track of what we have executed so we can detect recursive cases etc.
1914 std::set<BasicBlock*> ExecutedBlocks;
1916 // CurInst - The current instruction we're evaluating.
1917 BasicBlock::iterator CurInst = F->begin()->begin();
1919 // This is the main evaluation loop.
1921 Constant *InstResult = 0;
1923 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
1924 if (SI->isVolatile()) return false; // no volatile accesses.
1925 Constant *Ptr = getVal(Values, SI->getOperand(1));
1926 if (!isSimpleEnoughPointerToCommit(Ptr))
1927 // If this is too complex for us to commit, reject it.
1929 Constant *Val = getVal(Values, SI->getOperand(0));
1930 MutatedMemory[Ptr] = Val;
1931 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
1932 InstResult = ConstantExpr::get(BO->getOpcode(),
1933 getVal(Values, BO->getOperand(0)),
1934 getVal(Values, BO->getOperand(1)));
1935 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
1936 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
1937 getVal(Values, CI->getOperand(0)),
1938 getVal(Values, CI->getOperand(1)));
1939 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
1940 InstResult = ConstantExpr::getCast(CI->getOpcode(),
1941 getVal(Values, CI->getOperand(0)),
1943 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
1944 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
1945 getVal(Values, SI->getOperand(1)),
1946 getVal(Values, SI->getOperand(2)));
1947 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
1948 Constant *P = getVal(Values, GEP->getOperand(0));
1949 SmallVector<Constant*, 8> GEPOps;
1950 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
1951 GEPOps.push_back(getVal(Values, GEP->getOperand(i)));
1952 InstResult = ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size());
1953 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
1954 if (LI->isVolatile()) return false; // no volatile accesses.
1955 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
1957 if (InstResult == 0) return false; // Could not evaluate load.
1958 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
1959 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs.
1960 const Type *Ty = AI->getType()->getElementType();
1961 AllocaTmps.push_back(new GlobalVariable(Ty, false,
1962 GlobalValue::InternalLinkage,
1963 UndefValue::get(Ty),
1965 InstResult = AllocaTmps.back();
1966 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
1967 // Cannot handle inline asm.
1968 if (isa<InlineAsm>(CI->getOperand(0))) return false;
1970 // Resolve function pointers.
1971 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0)));
1972 if (!Callee) return false; // Cannot resolve.
1974 std::vector<Constant*> Formals;
1975 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
1976 Formals.push_back(getVal(Values, CI->getOperand(i)));
1978 if (Callee->isDeclaration()) {
1979 // If this is a function we can constant fold, do it.
1980 if (Constant *C = ConstantFoldCall(Callee, &Formals[0],
1987 if (Callee->getFunctionType()->isVarArg())
1992 // Execute the call, if successful, use the return value.
1993 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
1994 MutatedMemory, AllocaTmps))
1996 InstResult = RetVal;
1998 } else if (isa<TerminatorInst>(CurInst)) {
1999 BasicBlock *NewBB = 0;
2000 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2001 if (BI->isUnconditional()) {
2002 NewBB = BI->getSuccessor(0);
2005 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
2006 if (!Cond) return false; // Cannot determine.
2008 NewBB = BI->getSuccessor(!Cond->getZExtValue());
2010 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2012 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
2013 if (!Val) return false; // Cannot determine.
2014 NewBB = SI->getSuccessor(SI->findCaseValue(Val));
2015 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
2016 if (RI->getNumOperands())
2017 RetVal = getVal(Values, RI->getOperand(0));
2019 CallStack.pop_back(); // return from fn.
2020 return true; // We succeeded at evaluating this ctor!
2022 // invoke, unwind, unreachable.
2023 return false; // Cannot handle this terminator.
2026 // Okay, we succeeded in evaluating this control flow. See if we have
2027 // executed the new block before. If so, we have a looping function,
2028 // which we cannot evaluate in reasonable time.
2029 if (!ExecutedBlocks.insert(NewBB).second)
2030 return false; // looped!
2032 // Okay, we have never been in this block before. Check to see if there
2033 // are any PHI nodes. If so, evaluate them with information about where
2035 BasicBlock *OldBB = CurInst->getParent();
2036 CurInst = NewBB->begin();
2038 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2039 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
2041 // Do NOT increment CurInst. We know that the terminator had no value.
2044 // Did not know how to evaluate this!
2048 if (!CurInst->use_empty())
2049 Values[CurInst] = InstResult;
2051 // Advance program counter.
2056 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2057 /// we can. Return true if we can, false otherwise.
2058 static bool EvaluateStaticConstructor(Function *F) {
2059 /// MutatedMemory - For each store we execute, we update this map. Loads
2060 /// check this to get the most up-to-date value. If evaluation is successful,
2061 /// this state is committed to the process.
2062 std::map<Constant*, Constant*> MutatedMemory;
2064 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2065 /// to represent its body. This vector is needed so we can delete the
2066 /// temporary globals when we are done.
2067 std::vector<GlobalVariable*> AllocaTmps;
2069 /// CallStack - This is used to detect recursion. In pathological situations
2070 /// we could hit exponential behavior, but at least there is nothing
2072 std::vector<Function*> CallStack;
2074 // Call the function.
2075 Constant *RetValDummy;
2076 bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(),
2077 CallStack, MutatedMemory, AllocaTmps);
2079 // We succeeded at evaluation: commit the result.
2080 DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2081 << F->getName() << "' to " << MutatedMemory.size()
2083 for (std::map<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2084 E = MutatedMemory.end(); I != E; ++I)
2085 CommitValueTo(I->second, I->first);
2088 // At this point, we are done interpreting. If we created any 'alloca'
2089 // temporaries, release them now.
2090 while (!AllocaTmps.empty()) {
2091 GlobalVariable *Tmp = AllocaTmps.back();
2092 AllocaTmps.pop_back();
2094 // If there are still users of the alloca, the program is doing something
2095 // silly, e.g. storing the address of the alloca somewhere and using it
2096 // later. Since this is undefined, we'll just make it be null.
2097 if (!Tmp->use_empty())
2098 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2107 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2108 /// Return true if anything changed.
2109 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2110 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2111 bool MadeChange = false;
2112 if (Ctors.empty()) return false;
2114 // Loop over global ctors, optimizing them when we can.
2115 for (unsigned i = 0; i != Ctors.size(); ++i) {
2116 Function *F = Ctors[i];
2117 // Found a null terminator in the middle of the list, prune off the rest of
2120 if (i != Ctors.size()-1) {
2127 // We cannot simplify external ctor functions.
2128 if (F->empty()) continue;
2130 // If we can evaluate the ctor at compile time, do.
2131 if (EvaluateStaticConstructor(F)) {
2132 Ctors.erase(Ctors.begin()+i);
2135 ++NumCtorsEvaluated;
2140 if (!MadeChange) return false;
2142 GCL = InstallGlobalCtors(GCL, Ctors);
2147 bool GlobalOpt::runOnModule(Module &M) {
2148 bool Changed = false;
2150 // Try to find the llvm.globalctors list.
2151 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2153 bool LocalChange = true;
2154 while (LocalChange) {
2155 LocalChange = false;
2157 // Delete functions that are trivially dead, ccc -> fastcc
2158 LocalChange |= OptimizeFunctions(M);
2160 // Optimize global_ctors list.
2162 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2164 // Optimize non-address-taken globals.
2165 LocalChange |= OptimizeGlobalVars(M);
2166 Changed |= LocalChange;
2169 // TODO: Move all global ctors functions to the end of the module for code