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/ParameterAttributes.h"
25 #include "llvm/Pass.h"
26 #include "llvm/Analysis/ConstantFolding.h"
27 #include "llvm/Target/TargetData.h"
28 #include "llvm/Support/CallSite.h"
29 #include "llvm/Support/Compiler.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/GetElementPtrTypeIterator.h"
32 #include "llvm/ADT/SmallPtrSet.h"
33 #include "llvm/ADT/SmallVector.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/ADT/StringExtras.h"
40 STATISTIC(NumMarked , "Number of globals marked constant");
41 STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
42 STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
43 STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
44 STATISTIC(NumDeleted , "Number of globals deleted");
45 STATISTIC(NumFnDeleted , "Number of functions deleted");
46 STATISTIC(NumGlobUses , "Number of global uses devirtualized");
47 STATISTIC(NumLocalized , "Number of globals localized");
48 STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans");
49 STATISTIC(NumFastCallFns , "Number of functions converted to fastcc");
50 STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
51 STATISTIC(NumNestRemoved , "Number of nest attributes removed");
54 struct VISIBILITY_HIDDEN GlobalOpt : public ModulePass {
55 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
56 AU.addRequired<TargetData>();
58 static char ID; // Pass identification, replacement for typeid
59 GlobalOpt() : ModulePass((intptr_t)&ID) {}
61 bool runOnModule(Module &M);
64 GlobalVariable *FindGlobalCtors(Module &M);
65 bool OptimizeFunctions(Module &M);
66 bool OptimizeGlobalVars(Module &M);
67 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
68 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
71 char GlobalOpt::ID = 0;
72 RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer");
75 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
77 /// GlobalStatus - As we analyze each global, keep track of some information
78 /// about it. If we find out that the address of the global is taken, none of
79 /// this info will be accurate.
80 struct VISIBILITY_HIDDEN GlobalStatus {
81 /// isLoaded - True if the global is ever loaded. If the global isn't ever
82 /// loaded it can be deleted.
85 /// StoredType - Keep track of what stores to the global look like.
88 /// NotStored - There is no store to this global. It can thus be marked
92 /// isInitializerStored - This global is stored to, but the only thing
93 /// stored is the constant it was initialized with. This is only tracked
94 /// for scalar globals.
97 /// isStoredOnce - This global is stored to, but only its initializer and
98 /// one other value is ever stored to it. If this global isStoredOnce, we
99 /// track the value stored to it in StoredOnceValue below. This is only
100 /// tracked for scalar globals.
103 /// isStored - This global is stored to by multiple values or something else
104 /// that we cannot track.
108 /// StoredOnceValue - If only one value (besides the initializer constant) is
109 /// ever stored to this global, keep track of what value it is.
110 Value *StoredOnceValue;
112 /// AccessingFunction/HasMultipleAccessingFunctions - These start out
113 /// null/false. When the first accessing function is noticed, it is recorded.
114 /// When a second different accessing function is noticed,
115 /// HasMultipleAccessingFunctions is set to true.
116 Function *AccessingFunction;
117 bool HasMultipleAccessingFunctions;
119 /// HasNonInstructionUser - Set to true if this global has a user that is not
120 /// an instruction (e.g. a constant expr or GV initializer).
121 bool HasNonInstructionUser;
123 /// HasPHIUser - Set to true if this global has a user that is a PHI node.
126 GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0),
127 AccessingFunction(0), HasMultipleAccessingFunctions(false),
128 HasNonInstructionUser(false), HasPHIUser(false) {}
133 /// ConstantIsDead - Return true if the specified constant is (transitively)
134 /// dead. The constant may be used by other constants (e.g. constant arrays and
135 /// constant exprs) as long as they are dead, but it cannot be used by anything
137 static bool ConstantIsDead(Constant *C) {
138 if (isa<GlobalValue>(C)) return false;
140 for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI)
141 if (Constant *CU = dyn_cast<Constant>(*UI)) {
142 if (!ConstantIsDead(CU)) return false;
149 /// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
150 /// structure. If the global has its address taken, return true to indicate we
151 /// can't do anything with it.
153 static bool AnalyzeGlobal(Value *V, GlobalStatus &GS,
154 std::set<PHINode*> &PHIUsers) {
155 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
156 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
157 GS.HasNonInstructionUser = true;
159 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
161 } else if (Instruction *I = dyn_cast<Instruction>(*UI)) {
162 if (!GS.HasMultipleAccessingFunctions) {
163 Function *F = I->getParent()->getParent();
164 if (GS.AccessingFunction == 0)
165 GS.AccessingFunction = F;
166 else if (GS.AccessingFunction != F)
167 GS.HasMultipleAccessingFunctions = true;
169 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
171 if (LI->isVolatile()) return true; // Don't hack on volatile loads.
172 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
173 // Don't allow a store OF the address, only stores TO the address.
174 if (SI->getOperand(0) == V) return true;
176 if (SI->isVolatile()) return true; // Don't hack on volatile stores.
178 // If this is a direct store to the global (i.e., the global is a scalar
179 // value, not an aggregate), keep more specific information about
181 if (GS.StoredType != GlobalStatus::isStored)
182 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){
183 Value *StoredVal = SI->getOperand(0);
184 if (StoredVal == GV->getInitializer()) {
185 if (GS.StoredType < GlobalStatus::isInitializerStored)
186 GS.StoredType = GlobalStatus::isInitializerStored;
187 } else if (isa<LoadInst>(StoredVal) &&
188 cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
190 if (GS.StoredType < GlobalStatus::isInitializerStored)
191 GS.StoredType = GlobalStatus::isInitializerStored;
192 } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
193 GS.StoredType = GlobalStatus::isStoredOnce;
194 GS.StoredOnceValue = StoredVal;
195 } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
196 GS.StoredOnceValue == StoredVal) {
199 GS.StoredType = GlobalStatus::isStored;
202 GS.StoredType = GlobalStatus::isStored;
204 } else if (isa<GetElementPtrInst>(I)) {
205 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
206 } else if (isa<SelectInst>(I)) {
207 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
208 } else if (PHINode *PN = dyn_cast<PHINode>(I)) {
209 // PHI nodes we can check just like select or GEP instructions, but we
210 // have to be careful about infinite recursion.
211 if (PHIUsers.insert(PN).second) // Not already visited.
212 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
213 GS.HasPHIUser = true;
214 } else if (isa<CmpInst>(I)) {
215 } else if (isa<MemCpyInst>(I) || isa<MemMoveInst>(I)) {
216 if (I->getOperand(1) == V)
217 GS.StoredType = GlobalStatus::isStored;
218 if (I->getOperand(2) == V)
220 } else if (isa<MemSetInst>(I)) {
221 assert(I->getOperand(1) == V && "Memset only takes one pointer!");
222 GS.StoredType = GlobalStatus::isStored;
224 return true; // Any other non-load instruction might take address!
226 } else if (Constant *C = dyn_cast<Constant>(*UI)) {
227 GS.HasNonInstructionUser = true;
228 // We might have a dead and dangling constant hanging off of here.
229 if (!ConstantIsDead(C))
232 GS.HasNonInstructionUser = true;
233 // Otherwise must be some other user.
240 static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) {
241 ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
243 unsigned IdxV = CI->getZExtValue();
245 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
246 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
247 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
248 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
249 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) {
250 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
251 } else if (isa<ConstantAggregateZero>(Agg)) {
252 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
253 if (IdxV < STy->getNumElements())
254 return Constant::getNullValue(STy->getElementType(IdxV));
255 } else if (const SequentialType *STy =
256 dyn_cast<SequentialType>(Agg->getType())) {
257 return Constant::getNullValue(STy->getElementType());
259 } else if (isa<UndefValue>(Agg)) {
260 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
261 if (IdxV < STy->getNumElements())
262 return UndefValue::get(STy->getElementType(IdxV));
263 } else if (const SequentialType *STy =
264 dyn_cast<SequentialType>(Agg->getType())) {
265 return UndefValue::get(STy->getElementType());
272 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
273 /// users of the global, cleaning up the obvious ones. This is largely just a
274 /// quick scan over the use list to clean up the easy and obvious cruft. This
275 /// returns true if it made a change.
276 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
277 bool Changed = false;
278 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
281 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
283 // Replace the load with the initializer.
284 LI->replaceAllUsesWith(Init);
285 LI->eraseFromParent();
288 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
289 // Store must be unreachable or storing Init into the global.
290 SI->eraseFromParent();
292 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
293 if (CE->getOpcode() == Instruction::GetElementPtr) {
294 Constant *SubInit = 0;
296 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
297 Changed |= CleanupConstantGlobalUsers(CE, SubInit);
298 } else if (CE->getOpcode() == Instruction::BitCast &&
299 isa<PointerType>(CE->getType())) {
300 // Pointer cast, delete any stores and memsets to the global.
301 Changed |= CleanupConstantGlobalUsers(CE, 0);
304 if (CE->use_empty()) {
305 CE->destroyConstant();
308 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
309 // Do not transform "gepinst (gep constexpr (GV))" here, because forming
310 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
311 // and will invalidate our notion of what Init is.
312 Constant *SubInit = 0;
313 if (!isa<ConstantExpr>(GEP->getOperand(0))) {
315 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
316 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
317 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
319 Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
321 if (GEP->use_empty()) {
322 GEP->eraseFromParent();
325 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
326 if (MI->getRawDest() == V) {
327 MI->eraseFromParent();
331 } else if (Constant *C = dyn_cast<Constant>(U)) {
332 // If we have a chain of dead constantexprs or other things dangling from
333 // us, and if they are all dead, nuke them without remorse.
334 if (ConstantIsDead(C)) {
335 C->destroyConstant();
336 // This could have invalidated UI, start over from scratch.
337 CleanupConstantGlobalUsers(V, Init);
345 /// isSafeSROAElementUse - Return true if the specified instruction is a safe
346 /// user of a derived expression from a global that we want to SROA.
347 static bool isSafeSROAElementUse(Value *V) {
348 // We might have a dead and dangling constant hanging off of here.
349 if (Constant *C = dyn_cast<Constant>(V))
350 return ConstantIsDead(C);
352 Instruction *I = dyn_cast<Instruction>(V);
353 if (!I) return false;
356 if (isa<LoadInst>(I)) return true;
358 // Stores *to* the pointer are ok.
359 if (StoreInst *SI = dyn_cast<StoreInst>(I))
360 return SI->getOperand(0) != V;
362 // Otherwise, it must be a GEP.
363 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I);
364 if (GEPI == 0) return false;
366 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
367 !cast<Constant>(GEPI->getOperand(1))->isNullValue())
370 for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end();
372 if (!isSafeSROAElementUse(*I))
378 /// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value.
379 /// Look at it and its uses and decide whether it is safe to SROA this global.
381 static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
382 // The user of the global must be a GEP Inst or a ConstantExpr GEP.
383 if (!isa<GetElementPtrInst>(U) &&
384 (!isa<ConstantExpr>(U) ||
385 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
388 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
389 // don't like < 3 operand CE's, and we don't like non-constant integer
390 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some
392 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
393 !cast<Constant>(U->getOperand(1))->isNullValue() ||
394 !isa<ConstantInt>(U->getOperand(2)))
397 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
398 ++GEPI; // Skip over the pointer index.
400 // If this is a use of an array allocation, do a bit more checking for sanity.
401 if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
402 uint64_t NumElements = AT->getNumElements();
403 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
405 // Check to make sure that index falls within the array. If not,
406 // something funny is going on, so we won't do the optimization.
408 if (Idx->getZExtValue() >= NumElements)
411 // We cannot scalar repl this level of the array unless any array
412 // sub-indices are in-range constants. In particular, consider:
413 // A[0][i]. We cannot know that the user isn't doing invalid things like
414 // allowing i to index an out-of-range subscript that accesses A[1].
416 // Scalar replacing *just* the outer index of the array is probably not
417 // going to be a win anyway, so just give up.
418 for (++GEPI; // Skip array index.
419 GEPI != E && (isa<ArrayType>(*GEPI) || isa<VectorType>(*GEPI));
421 uint64_t NumElements;
422 if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
423 NumElements = SubArrayTy->getNumElements();
425 NumElements = cast<VectorType>(*GEPI)->getNumElements();
427 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
428 if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
433 for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I)
434 if (!isSafeSROAElementUse(*I))
439 /// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
440 /// is safe for us to perform this transformation.
442 static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
443 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
445 if (!IsUserOfGlobalSafeForSRA(*UI, GV))
452 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
453 /// variable. This opens the door for other optimizations by exposing the
454 /// behavior of the program in a more fine-grained way. We have determined that
455 /// this transformation is safe already. We return the first global variable we
456 /// insert so that the caller can reprocess it.
457 static GlobalVariable *SRAGlobal(GlobalVariable *GV) {
458 // Make sure this global only has simple uses that we can SRA.
459 if (!GlobalUsersSafeToSRA(GV))
462 assert(GV->hasInternalLinkage() && !GV->isConstant());
463 Constant *Init = GV->getInitializer();
464 const Type *Ty = Init->getType();
466 std::vector<GlobalVariable*> NewGlobals;
467 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
469 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
470 NewGlobals.reserve(STy->getNumElements());
471 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
472 Constant *In = getAggregateConstantElement(Init,
473 ConstantInt::get(Type::Int32Ty, i));
474 assert(In && "Couldn't get element of initializer?");
475 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
476 GlobalVariable::InternalLinkage,
477 In, GV->getName()+"."+utostr(i),
479 GV->isThreadLocal());
480 Globals.insert(GV, NGV);
481 NewGlobals.push_back(NGV);
483 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
484 unsigned NumElements = 0;
485 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
486 NumElements = ATy->getNumElements();
487 else if (const VectorType *PTy = dyn_cast<VectorType>(STy))
488 NumElements = PTy->getNumElements();
490 assert(0 && "Unknown aggregate sequential type!");
492 if (NumElements > 16 && GV->hasNUsesOrMore(16))
493 return 0; // It's not worth it.
494 NewGlobals.reserve(NumElements);
495 for (unsigned i = 0, e = NumElements; i != e; ++i) {
496 Constant *In = getAggregateConstantElement(Init,
497 ConstantInt::get(Type::Int32Ty, i));
498 assert(In && "Couldn't get element of initializer?");
500 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
501 GlobalVariable::InternalLinkage,
502 In, GV->getName()+"."+utostr(i),
504 GV->isThreadLocal());
505 Globals.insert(GV, NGV);
506 NewGlobals.push_back(NGV);
510 if (NewGlobals.empty())
513 DOUT << "PERFORMING GLOBAL SRA ON: " << *GV;
515 Constant *NullInt = Constant::getNullValue(Type::Int32Ty);
517 // Loop over all of the uses of the global, replacing the constantexpr geps,
518 // with smaller constantexpr geps or direct references.
519 while (!GV->use_empty()) {
520 User *GEP = GV->use_back();
521 assert(((isa<ConstantExpr>(GEP) &&
522 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
523 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
525 // Ignore the 1th operand, which has to be zero or else the program is quite
526 // broken (undefined). Get the 2nd operand, which is the structure or array
528 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
529 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
531 Value *NewPtr = NewGlobals[Val];
533 // Form a shorter GEP if needed.
534 if (GEP->getNumOperands() > 3)
535 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
536 SmallVector<Constant*, 8> Idxs;
537 Idxs.push_back(NullInt);
538 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
539 Idxs.push_back(CE->getOperand(i));
540 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr),
541 &Idxs[0], Idxs.size());
543 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
544 SmallVector<Value*, 8> Idxs;
545 Idxs.push_back(NullInt);
546 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
547 Idxs.push_back(GEPI->getOperand(i));
548 NewPtr = new GetElementPtrInst(NewPtr, Idxs.begin(), Idxs.end(),
549 GEPI->getName()+"."+utostr(Val), GEPI);
551 GEP->replaceAllUsesWith(NewPtr);
553 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
554 GEPI->eraseFromParent();
556 cast<ConstantExpr>(GEP)->destroyConstant();
559 // Delete the old global, now that it is dead.
563 // Loop over the new globals array deleting any globals that are obviously
564 // dead. This can arise due to scalarization of a structure or an array that
565 // has elements that are dead.
566 unsigned FirstGlobal = 0;
567 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
568 if (NewGlobals[i]->use_empty()) {
569 Globals.erase(NewGlobals[i]);
570 if (FirstGlobal == i) ++FirstGlobal;
573 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
576 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
577 /// value will trap if the value is dynamically null. PHIs keeps track of any
578 /// phi nodes we've seen to avoid reprocessing them.
579 static bool AllUsesOfValueWillTrapIfNull(Value *V,
580 SmallPtrSet<PHINode*, 8> &PHIs) {
581 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
582 if (isa<LoadInst>(*UI)) {
584 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
585 if (SI->getOperand(0) == V) {
586 //cerr << "NONTRAPPING USE: " << **UI;
587 return false; // Storing the value.
589 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
590 if (CI->getOperand(0) != V) {
591 //cerr << "NONTRAPPING USE: " << **UI;
592 return false; // Not calling the ptr
594 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
595 if (II->getOperand(0) != V) {
596 //cerr << "NONTRAPPING USE: " << **UI;
597 return false; // Not calling the ptr
599 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(*UI)) {
600 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
601 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
602 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
603 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
604 // If we've already seen this phi node, ignore it, it has already been
607 return AllUsesOfValueWillTrapIfNull(PN, PHIs);
608 } else if (isa<ICmpInst>(*UI) &&
609 isa<ConstantPointerNull>(UI->getOperand(1))) {
610 // Ignore setcc X, null
612 //cerr << "NONTRAPPING USE: " << **UI;
618 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
619 /// from GV will trap if the loaded value is null. Note that this also permits
620 /// comparisons of the loaded value against null, as a special case.
621 static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) {
622 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI)
623 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
624 SmallPtrSet<PHINode*, 8> PHIs;
625 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
627 } else if (isa<StoreInst>(*UI)) {
628 // Ignore stores to the global.
630 // We don't know or understand this user, bail out.
631 //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI;
638 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
639 bool Changed = false;
640 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
641 Instruction *I = cast<Instruction>(*UI++);
642 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
643 LI->setOperand(0, NewV);
645 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
646 if (SI->getOperand(1) == V) {
647 SI->setOperand(1, NewV);
650 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
651 if (I->getOperand(0) == V) {
652 // Calling through the pointer! Turn into a direct call, but be careful
653 // that the pointer is not also being passed as an argument.
654 I->setOperand(0, NewV);
656 bool PassedAsArg = false;
657 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
658 if (I->getOperand(i) == V) {
660 I->setOperand(i, NewV);
664 // Being passed as an argument also. Be careful to not invalidate UI!
668 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
669 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
670 ConstantExpr::getCast(CI->getOpcode(),
671 NewV, CI->getType()));
672 if (CI->use_empty()) {
674 CI->eraseFromParent();
676 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
677 // Should handle GEP here.
678 SmallVector<Constant*, 8> Idxs;
679 Idxs.reserve(GEPI->getNumOperands()-1);
680 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
681 if (Constant *C = dyn_cast<Constant>(GEPI->getOperand(i)))
685 if (Idxs.size() == GEPI->getNumOperands()-1)
686 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
687 ConstantExpr::getGetElementPtr(NewV, &Idxs[0],
689 if (GEPI->use_empty()) {
691 GEPI->eraseFromParent();
700 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
701 /// value stored into it. If there are uses of the loaded value that would trap
702 /// if the loaded value is dynamically null, then we know that they cannot be
703 /// reachable with a null optimize away the load.
704 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) {
705 std::vector<LoadInst*> Loads;
706 bool Changed = false;
708 // Replace all uses of loads with uses of uses of the stored value.
709 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end();
711 if (LoadInst *LI = dyn_cast<LoadInst>(*GUI)) {
713 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
715 // If we get here we could have stores, selects, or phi nodes whose values
717 assert((isa<StoreInst>(*GUI) || isa<PHINode>(*GUI) ||
718 isa<SelectInst>(*GUI) || isa<ConstantExpr>(*GUI)) &&
719 "Only expect load and stores!");
723 DOUT << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV;
727 // Delete all of the loads we can, keeping track of whether we nuked them all!
728 bool AllLoadsGone = true;
729 while (!Loads.empty()) {
730 LoadInst *L = Loads.back();
731 if (L->use_empty()) {
732 L->eraseFromParent();
735 AllLoadsGone = false;
740 // If we nuked all of the loads, then none of the stores are needed either,
741 // nor is the global.
743 DOUT << " *** GLOBAL NOW DEAD!\n";
744 CleanupConstantGlobalUsers(GV, 0);
745 if (GV->use_empty()) {
746 GV->eraseFromParent();
754 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
755 /// instructions that are foldable.
756 static void ConstantPropUsersOf(Value *V) {
757 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
758 if (Instruction *I = dyn_cast<Instruction>(*UI++))
759 if (Constant *NewC = ConstantFoldInstruction(I)) {
760 I->replaceAllUsesWith(NewC);
762 // Advance UI to the next non-I use to avoid invalidating it!
763 // Instructions could multiply use V.
764 while (UI != E && *UI == I)
766 I->eraseFromParent();
770 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
771 /// variable, and transforms the program as if it always contained the result of
772 /// the specified malloc. Because it is always the result of the specified
773 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
774 /// malloc into a global, and any loads of GV as uses of the new global.
775 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
777 DOUT << "PROMOTING MALLOC GLOBAL: " << *GV << " MALLOC = " << *MI;
778 ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize());
780 if (NElements->getZExtValue() != 1) {
781 // If we have an array allocation, transform it to a single element
782 // allocation to make the code below simpler.
783 Type *NewTy = ArrayType::get(MI->getAllocatedType(),
784 NElements->getZExtValue());
786 new MallocInst(NewTy, Constant::getNullValue(Type::Int32Ty),
787 MI->getAlignment(), MI->getName(), MI);
789 Indices[0] = Indices[1] = Constant::getNullValue(Type::Int32Ty);
790 Value *NewGEP = new GetElementPtrInst(NewMI, Indices, Indices + 2,
791 NewMI->getName()+".el0", MI);
792 MI->replaceAllUsesWith(NewGEP);
793 MI->eraseFromParent();
797 // Create the new global variable. The contents of the malloc'd memory is
798 // undefined, so initialize with an undef value.
799 Constant *Init = UndefValue::get(MI->getAllocatedType());
800 GlobalVariable *NewGV = new GlobalVariable(MI->getAllocatedType(), false,
801 GlobalValue::InternalLinkage, Init,
802 GV->getName()+".body",
804 GV->isThreadLocal());
805 GV->getParent()->getGlobalList().insert(GV, NewGV);
807 // Anything that used the malloc now uses the global directly.
808 MI->replaceAllUsesWith(NewGV);
810 Constant *RepValue = NewGV;
811 if (NewGV->getType() != GV->getType()->getElementType())
812 RepValue = ConstantExpr::getBitCast(RepValue,
813 GV->getType()->getElementType());
815 // If there is a comparison against null, we will insert a global bool to
816 // keep track of whether the global was initialized yet or not.
817 GlobalVariable *InitBool =
818 new GlobalVariable(Type::Int1Ty, false, GlobalValue::InternalLinkage,
819 ConstantInt::getFalse(), GV->getName()+".init",
820 (Module *)NULL, GV->isThreadLocal());
821 bool InitBoolUsed = false;
823 // Loop over all uses of GV, processing them in turn.
824 std::vector<StoreInst*> Stores;
825 while (!GV->use_empty())
826 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
827 while (!LI->use_empty()) {
828 Use &LoadUse = LI->use_begin().getUse();
829 if (!isa<ICmpInst>(LoadUse.getUser()))
832 ICmpInst *CI = cast<ICmpInst>(LoadUse.getUser());
833 // Replace the cmp X, 0 with a use of the bool value.
834 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", CI);
836 switch (CI->getPredicate()) {
837 default: assert(0 && "Unknown ICmp Predicate!");
838 case ICmpInst::ICMP_ULT:
839 case ICmpInst::ICMP_SLT:
840 LV = ConstantInt::getFalse(); // X < null -> always false
842 case ICmpInst::ICMP_ULE:
843 case ICmpInst::ICMP_SLE:
844 case ICmpInst::ICMP_EQ:
845 LV = BinaryOperator::createNot(LV, "notinit", CI);
847 case ICmpInst::ICMP_NE:
848 case ICmpInst::ICMP_UGE:
849 case ICmpInst::ICMP_SGE:
850 case ICmpInst::ICMP_UGT:
851 case ICmpInst::ICMP_SGT:
854 CI->replaceAllUsesWith(LV);
855 CI->eraseFromParent();
858 LI->eraseFromParent();
860 StoreInst *SI = cast<StoreInst>(GV->use_back());
861 // The global is initialized when the store to it occurs.
862 new StoreInst(ConstantInt::getTrue(), InitBool, SI);
863 SI->eraseFromParent();
866 // If the initialization boolean was used, insert it, otherwise delete it.
868 while (!InitBool->use_empty()) // Delete initializations
869 cast<Instruction>(InitBool->use_back())->eraseFromParent();
872 GV->getParent()->getGlobalList().insert(GV, InitBool);
875 // Now the GV is dead, nuke it and the malloc.
876 GV->eraseFromParent();
877 MI->eraseFromParent();
879 // To further other optimizations, loop over all users of NewGV and try to
880 // constant prop them. This will promote GEP instructions with constant
881 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
882 ConstantPropUsersOf(NewGV);
883 if (RepValue != NewGV)
884 ConstantPropUsersOf(RepValue);
889 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
890 /// to make sure that there are no complex uses of V. We permit simple things
891 /// like dereferencing the pointer, but not storing through the address, unless
892 /// it is to the specified global.
893 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V,
895 SmallPtrSet<PHINode*, 8> &PHIs) {
896 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
897 if (isa<LoadInst>(*UI) || isa<CmpInst>(*UI)) {
899 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
900 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
901 return false; // Storing the pointer itself... bad.
902 // Otherwise, storing through it, or storing into GV... fine.
903 } else if (isa<GetElementPtrInst>(*UI)) {
904 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(cast<Instruction>(*UI),
907 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
908 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI
911 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
919 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
920 /// somewhere. Transform all uses of the allocation into loads from the
921 /// global and uses of the resultant pointer. Further, delete the store into
922 /// GV. This assumes that these value pass the
923 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
924 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
925 GlobalVariable *GV) {
926 while (!Alloc->use_empty()) {
927 Instruction *U = cast<Instruction>(*Alloc->use_begin());
928 Instruction *InsertPt = U;
929 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
930 // If this is the store of the allocation into the global, remove it.
931 if (SI->getOperand(1) == GV) {
932 SI->eraseFromParent();
935 } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
936 // Insert the load in the corresponding predecessor, not right before the
938 unsigned PredNo = Alloc->use_begin().getOperandNo()/2;
939 InsertPt = PN->getIncomingBlock(PredNo)->getTerminator();
942 // Insert a load from the global, and use it instead of the malloc.
943 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
944 U->replaceUsesOfWith(Alloc, NL);
948 /// GlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
949 /// GV are simple enough to perform HeapSRA, return true.
950 static bool GlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV,
952 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;
954 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
955 // We permit two users of the load: setcc comparing against the null
956 // pointer, and a getelementptr of a specific form.
957 for (Value::use_iterator UI = LI->use_begin(), E = LI->use_end(); UI != E;
959 // Comparison against null is ok.
960 if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) {
961 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
966 // getelementptr is also ok, but only a simple form.
967 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
968 // Must index into the array and into the struct.
969 if (GEPI->getNumOperands() < 3)
972 // Otherwise the GEP is ok.
976 if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
977 // We have a phi of a load from the global. We can only handle this
978 // if the other PHI'd values are actually the same. In this case,
979 // the rewriter will just drop the phi entirely.
980 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
981 Value *IV = PN->getIncomingValue(i);
982 if (IV == LI) continue; // Trivial the same.
984 // If the phi'd value is from the malloc that initializes the value,
986 if (IV == MI) continue;
988 // Otherwise, we don't know what it is.
994 // Otherwise we don't know what this is, not ok.
1001 /// GetHeapSROALoad - Return the load for the specified field of the HeapSROA'd
1002 /// value, lazily creating it on demand.
1003 static Value *GetHeapSROALoad(Instruction *Load, unsigned FieldNo,
1004 const std::vector<GlobalVariable*> &FieldGlobals,
1005 std::vector<Value *> &InsertedLoadsForPtr) {
1006 if (InsertedLoadsForPtr.size() <= FieldNo)
1007 InsertedLoadsForPtr.resize(FieldNo+1);
1008 if (InsertedLoadsForPtr[FieldNo] == 0)
1009 InsertedLoadsForPtr[FieldNo] = new LoadInst(FieldGlobals[FieldNo],
1010 Load->getName()+".f" +
1011 utostr(FieldNo), Load);
1012 return InsertedLoadsForPtr[FieldNo];
1015 /// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1016 /// the load, rewrite the derived value to use the HeapSRoA'd load.
1017 static void RewriteHeapSROALoadUser(LoadInst *Load, Instruction *LoadUser,
1018 const std::vector<GlobalVariable*> &FieldGlobals,
1019 std::vector<Value *> &InsertedLoadsForPtr) {
1020 // If this is a comparison against null, handle it.
1021 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1022 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1023 // If we have a setcc of the loaded pointer, we can use a setcc of any
1026 if (InsertedLoadsForPtr.empty()) {
1027 NPtr = GetHeapSROALoad(Load, 0, FieldGlobals, InsertedLoadsForPtr);
1029 NPtr = InsertedLoadsForPtr.back();
1032 Value *New = new ICmpInst(SCI->getPredicate(), NPtr,
1033 Constant::getNullValue(NPtr->getType()),
1034 SCI->getName(), SCI);
1035 SCI->replaceAllUsesWith(New);
1036 SCI->eraseFromParent();
1040 // Handle 'getelementptr Ptr, Idx, uint FieldNo ...'
1041 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1042 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1043 && "Unexpected GEPI!");
1045 // Load the pointer for this field.
1046 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1047 Value *NewPtr = GetHeapSROALoad(Load, FieldNo,
1048 FieldGlobals, InsertedLoadsForPtr);
1050 // Create the new GEP idx vector.
1051 SmallVector<Value*, 8> GEPIdx;
1052 GEPIdx.push_back(GEPI->getOperand(1));
1053 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1055 Value *NGEPI = new GetElementPtrInst(NewPtr, GEPIdx.begin(), GEPIdx.end(),
1056 GEPI->getName(), GEPI);
1057 GEPI->replaceAllUsesWith(NGEPI);
1058 GEPI->eraseFromParent();
1062 // Handle PHI nodes. PHI nodes must be merging in the same values, plus
1063 // potentially the original malloc. Insert phi nodes for each field, then
1064 // process uses of the PHI.
1065 PHINode *PN = cast<PHINode>(LoadUser);
1066 std::vector<Value *> PHIsForField;
1067 PHIsForField.resize(FieldGlobals.size());
1068 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1069 Value *LoadV = GetHeapSROALoad(Load, i, FieldGlobals, InsertedLoadsForPtr);
1071 PHINode *FieldPN = new PHINode(LoadV->getType(),
1072 PN->getName()+"."+utostr(i), PN);
1073 // Fill in the predecessor values.
1074 for (unsigned pred = 0, e = PN->getNumIncomingValues(); pred != e; ++pred) {
1075 // Each predecessor either uses the load or the original malloc.
1076 Value *InVal = PN->getIncomingValue(pred);
1077 BasicBlock *BB = PN->getIncomingBlock(pred);
1079 if (isa<MallocInst>(InVal)) {
1080 // Insert a reload from the global in the predecessor.
1081 NewVal = GetHeapSROALoad(BB->getTerminator(), i, FieldGlobals,
1084 NewVal = InsertedLoadsForPtr[i];
1086 FieldPN->addIncoming(NewVal, BB);
1088 PHIsForField[i] = FieldPN;
1091 // Since PHIsForField specifies a phi for every input value, the lazy inserter
1092 // will never insert a load.
1093 while (!PN->use_empty())
1094 RewriteHeapSROALoadUser(Load, PN->use_back(), FieldGlobals, PHIsForField);
1095 PN->eraseFromParent();
1098 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1099 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1100 /// use FieldGlobals instead. All uses of loaded values satisfy
1101 /// GlobalLoadUsesSimpleEnoughForHeapSRA.
1102 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1103 const std::vector<GlobalVariable*> &FieldGlobals) {
1104 std::vector<Value *> InsertedLoadsForPtr;
1105 //InsertedLoadsForPtr.resize(FieldGlobals.size());
1106 while (!Load->use_empty())
1107 RewriteHeapSROALoadUser(Load, Load->use_back(),
1108 FieldGlobals, InsertedLoadsForPtr);
1111 /// PerformHeapAllocSRoA - MI is an allocation of an array of structures. Break
1112 /// it up into multiple allocations of arrays of the fields.
1113 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){
1114 DOUT << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *MI;
1115 const StructType *STy = cast<StructType>(MI->getAllocatedType());
1117 // There is guaranteed to be at least one use of the malloc (storing
1118 // it into GV). If there are other uses, change them to be uses of
1119 // the global to simplify later code. This also deletes the store
1121 ReplaceUsesOfMallocWithGlobal(MI, GV);
1123 // Okay, at this point, there are no users of the malloc. Insert N
1124 // new mallocs at the same place as MI, and N globals.
1125 std::vector<GlobalVariable*> FieldGlobals;
1126 std::vector<MallocInst*> FieldMallocs;
1128 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1129 const Type *FieldTy = STy->getElementType(FieldNo);
1130 const Type *PFieldTy = PointerType::getUnqual(FieldTy);
1132 GlobalVariable *NGV =
1133 new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage,
1134 Constant::getNullValue(PFieldTy),
1135 GV->getName() + ".f" + utostr(FieldNo), GV,
1136 GV->isThreadLocal());
1137 FieldGlobals.push_back(NGV);
1139 MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(),
1140 MI->getName() + ".f" + utostr(FieldNo),MI);
1141 FieldMallocs.push_back(NMI);
1142 new StoreInst(NMI, NGV, MI);
1145 // The tricky aspect of this transformation is handling the case when malloc
1146 // fails. In the original code, malloc failing would set the result pointer
1147 // of malloc to null. In this case, some mallocs could succeed and others
1148 // could fail. As such, we emit code that looks like this:
1149 // F0 = malloc(field0)
1150 // F1 = malloc(field1)
1151 // F2 = malloc(field2)
1152 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1153 // if (F0) { free(F0); F0 = 0; }
1154 // if (F1) { free(F1); F1 = 0; }
1155 // if (F2) { free(F2); F2 = 0; }
1157 Value *RunningOr = 0;
1158 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1159 Value *Cond = new ICmpInst(ICmpInst::ICMP_EQ, FieldMallocs[i],
1160 Constant::getNullValue(FieldMallocs[i]->getType()),
1163 RunningOr = Cond; // First seteq
1165 RunningOr = BinaryOperator::createOr(RunningOr, Cond, "tmp", MI);
1168 // Split the basic block at the old malloc.
1169 BasicBlock *OrigBB = MI->getParent();
1170 BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont");
1172 // Create the block to check the first condition. Put all these blocks at the
1173 // end of the function as they are unlikely to be executed.
1174 BasicBlock *NullPtrBlock = new BasicBlock("malloc_ret_null",
1175 OrigBB->getParent());
1177 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1178 // branch on RunningOr.
1179 OrigBB->getTerminator()->eraseFromParent();
1180 new BranchInst(NullPtrBlock, ContBB, RunningOr, OrigBB);
1182 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1183 // pointer, because some may be null while others are not.
1184 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1185 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1186 Value *Cmp = new ICmpInst(ICmpInst::ICMP_NE, GVVal,
1187 Constant::getNullValue(GVVal->getType()),
1188 "tmp", NullPtrBlock);
1189 BasicBlock *FreeBlock = new BasicBlock("free_it", OrigBB->getParent());
1190 BasicBlock *NextBlock = new BasicBlock("next", OrigBB->getParent());
1191 new BranchInst(FreeBlock, NextBlock, Cmp, NullPtrBlock);
1193 // Fill in FreeBlock.
1194 new FreeInst(GVVal, FreeBlock);
1195 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1197 new BranchInst(NextBlock, FreeBlock);
1199 NullPtrBlock = NextBlock;
1202 new BranchInst(ContBB, NullPtrBlock);
1205 // MI is no longer needed, remove it.
1206 MI->eraseFromParent();
1209 // Okay, the malloc site is completely handled. All of the uses of GV are now
1210 // loads, and all uses of those loads are simple. Rewrite them to use loads
1211 // of the per-field globals instead.
1212 while (!GV->use_empty()) {
1213 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
1214 RewriteUsesOfLoadForHeapSRoA(LI, FieldGlobals);
1215 LI->eraseFromParent();
1217 // Must be a store of null.
1218 StoreInst *SI = cast<StoreInst>(GV->use_back());
1219 assert(isa<Constant>(SI->getOperand(0)) &&
1220 cast<Constant>(SI->getOperand(0))->isNullValue() &&
1221 "Unexpected heap-sra user!");
1223 // Insert a store of null into each global.
1224 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1226 Constant::getNullValue(FieldGlobals[i]->getType()->getElementType());
1227 new StoreInst(Null, FieldGlobals[i], SI);
1229 // Erase the original store.
1230 SI->eraseFromParent();
1234 // The old global is now dead, remove it.
1235 GV->eraseFromParent();
1238 return FieldGlobals[0];
1242 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1243 // that only one value (besides its initializer) is ever stored to the global.
1244 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1245 Module::global_iterator &GVI,
1247 if (CastInst *CI = dyn_cast<CastInst>(StoredOnceVal))
1248 StoredOnceVal = CI->getOperand(0);
1249 else if (GetElementPtrInst *GEPI =dyn_cast<GetElementPtrInst>(StoredOnceVal)){
1250 // "getelementptr Ptr, 0, 0, 0" is really just a cast.
1251 bool IsJustACast = true;
1252 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
1253 if (!isa<Constant>(GEPI->getOperand(i)) ||
1254 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
1255 IsJustACast = false;
1259 StoredOnceVal = GEPI->getOperand(0);
1262 // If we are dealing with a pointer global that is initialized to null and
1263 // only has one (non-null) value stored into it, then we can optimize any
1264 // users of the loaded value (often calls and loads) that would trap if the
1266 if (isa<PointerType>(GV->getInitializer()->getType()) &&
1267 GV->getInitializer()->isNullValue()) {
1268 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1269 if (GV->getInitializer()->getType() != SOVC->getType())
1270 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1272 // Optimize away any trapping uses of the loaded value.
1273 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1275 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) {
1276 // If this is a malloc of an abstract type, don't touch it.
1277 if (!MI->getAllocatedType()->isSized())
1280 // We can't optimize this global unless all uses of it are *known* to be
1281 // of the malloc value, not of the null initializer value (consider a use
1282 // that compares the global's value against zero to see if the malloc has
1283 // been reached). To do this, we check to see if all uses of the global
1284 // would trap if the global were null: this proves that they must all
1285 // happen after the malloc.
1286 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1289 // We can't optimize this if the malloc itself is used in a complex way,
1290 // for example, being stored into multiple globals. This allows the
1291 // malloc to be stored into the specified global, loaded setcc'd, and
1292 // GEP'd. These are all things we could transform to using the global
1295 SmallPtrSet<PHINode*, 8> PHIs;
1296 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV, PHIs))
1301 // If we have a global that is only initialized with a fixed size malloc,
1302 // transform the program to use global memory instead of malloc'd memory.
1303 // This eliminates dynamic allocation, avoids an indirection accessing the
1304 // data, and exposes the resultant global to further GlobalOpt.
1305 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) {
1306 // Restrict this transformation to only working on small allocations
1307 // (2048 bytes currently), as we don't want to introduce a 16M global or
1309 if (NElements->getZExtValue()*
1310 TD.getABITypeSize(MI->getAllocatedType()) < 2048) {
1311 GVI = OptimizeGlobalAddressOfMalloc(GV, MI);
1316 // If the allocation is an array of structures, consider transforming this
1317 // into multiple malloc'd arrays, one for each field. This is basically
1318 // SRoA for malloc'd memory.
1319 if (const StructType *AllocTy =
1320 dyn_cast<StructType>(MI->getAllocatedType())) {
1321 // This the structure has an unreasonable number of fields, leave it
1323 if (AllocTy->getNumElements() <= 16 && AllocTy->getNumElements() > 0 &&
1324 GlobalLoadUsesSimpleEnoughForHeapSRA(GV, MI)) {
1325 GVI = PerformHeapAllocSRoA(GV, MI);
1335 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1336 /// two values ever stored into GV are its initializer and OtherVal. See if we
1337 /// can shrink the global into a boolean and select between the two values
1338 /// whenever it is used. This exposes the values to other scalar optimizations.
1339 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1340 const Type *GVElType = GV->getType()->getElementType();
1342 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1343 // an FP value or vector, don't do this optimization because a select between
1344 // them is very expensive and unlikely to lead to later simplification.
1345 if (GVElType == Type::Int1Ty || GVElType->isFloatingPoint() ||
1346 isa<VectorType>(GVElType))
1349 // Walk the use list of the global seeing if all the uses are load or store.
1350 // If there is anything else, bail out.
1351 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I)
1352 if (!isa<LoadInst>(I) && !isa<StoreInst>(I))
1355 DOUT << " *** SHRINKING TO BOOL: " << *GV;
1357 // Create the new global, initializing it to false.
1358 GlobalVariable *NewGV = new GlobalVariable(Type::Int1Ty, false,
1359 GlobalValue::InternalLinkage, ConstantInt::getFalse(),
1362 GV->isThreadLocal());
1363 GV->getParent()->getGlobalList().insert(GV, NewGV);
1365 Constant *InitVal = GV->getInitializer();
1366 assert(InitVal->getType() != Type::Int1Ty && "No reason to shrink to bool!");
1368 // If initialized to zero and storing one into the global, we can use a cast
1369 // instead of a select to synthesize the desired value.
1370 bool IsOneZero = false;
1371 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1372 IsOneZero = InitVal->isNullValue() && CI->isOne();
1374 while (!GV->use_empty()) {
1375 Instruction *UI = cast<Instruction>(GV->use_back());
1376 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1377 // Change the store into a boolean store.
1378 bool StoringOther = SI->getOperand(0) == OtherVal;
1379 // Only do this if we weren't storing a loaded value.
1381 if (StoringOther || SI->getOperand(0) == InitVal)
1382 StoreVal = ConstantInt::get(Type::Int1Ty, StoringOther);
1384 // Otherwise, we are storing a previously loaded copy. To do this,
1385 // change the copy from copying the original value to just copying the
1387 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1389 // If we're already replaced the input, StoredVal will be a cast or
1390 // select instruction. If not, it will be a load of the original
1392 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1393 assert(LI->getOperand(0) == GV && "Not a copy!");
1394 // Insert a new load, to preserve the saved value.
1395 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1397 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1398 "This is not a form that we understand!");
1399 StoreVal = StoredVal->getOperand(0);
1400 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1403 new StoreInst(StoreVal, NewGV, SI);
1405 // Change the load into a load of bool then a select.
1406 LoadInst *LI = cast<LoadInst>(UI);
1407 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1410 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1412 NSI = new SelectInst(NLI, OtherVal, InitVal, "", LI);
1414 LI->replaceAllUsesWith(NSI);
1416 UI->eraseFromParent();
1419 GV->eraseFromParent();
1424 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1425 /// it if possible. If we make a change, return true.
1426 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1427 Module::global_iterator &GVI) {
1428 std::set<PHINode*> PHIUsers;
1430 GV->removeDeadConstantUsers();
1432 if (GV->use_empty()) {
1433 DOUT << "GLOBAL DEAD: " << *GV;
1434 GV->eraseFromParent();
1439 if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
1441 cerr << "Global: " << *GV;
1442 cerr << " isLoaded = " << GS.isLoaded << "\n";
1443 cerr << " StoredType = ";
1444 switch (GS.StoredType) {
1445 case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break;
1446 case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break;
1447 case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break;
1448 case GlobalStatus::isStored: cerr << "stored\n"; break;
1450 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
1451 cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n";
1452 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
1453 cerr << " AccessingFunction = " << GS.AccessingFunction->getName()
1455 cerr << " HasMultipleAccessingFunctions = "
1456 << GS.HasMultipleAccessingFunctions << "\n";
1457 cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n";
1461 // If this is a first class global and has only one accessing function
1462 // and this function is main (which we know is not recursive we can make
1463 // this global a local variable) we replace the global with a local alloca
1464 // in this function.
1466 // NOTE: It doesn't make sense to promote non first class types since we
1467 // are just replacing static memory to stack memory.
1468 if (!GS.HasMultipleAccessingFunctions &&
1469 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1470 GV->getType()->getElementType()->isFirstClassType() &&
1471 GS.AccessingFunction->getName() == "main" &&
1472 GS.AccessingFunction->hasExternalLinkage()) {
1473 DOUT << "LOCALIZING GLOBAL: " << *GV;
1474 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin();
1475 const Type* ElemTy = GV->getType()->getElementType();
1476 // FIXME: Pass Global's alignment when globals have alignment
1477 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI);
1478 if (!isa<UndefValue>(GV->getInitializer()))
1479 new StoreInst(GV->getInitializer(), Alloca, FirstI);
1481 GV->replaceAllUsesWith(Alloca);
1482 GV->eraseFromParent();
1487 // If the global is never loaded (but may be stored to), it is dead.
1490 DOUT << "GLOBAL NEVER LOADED: " << *GV;
1492 // Delete any stores we can find to the global. We may not be able to
1493 // make it completely dead though.
1494 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1496 // If the global is dead now, delete it.
1497 if (GV->use_empty()) {
1498 GV->eraseFromParent();
1504 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1505 DOUT << "MARKING CONSTANT: " << *GV;
1506 GV->setConstant(true);
1508 // Clean up any obviously simplifiable users now.
1509 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1511 // If the global is dead now, just nuke it.
1512 if (GV->use_empty()) {
1513 DOUT << " *** Marking constant allowed us to simplify "
1514 << "all users and delete global!\n";
1515 GV->eraseFromParent();
1521 } else if (!GV->getInitializer()->getType()->isFirstClassType()) {
1522 if (GlobalVariable *FirstNewGV = SRAGlobal(GV)) {
1523 GVI = FirstNewGV; // Don't skip the newly produced globals!
1526 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1527 // If the initial value for the global was an undef value, and if only
1528 // one other value was stored into it, we can just change the
1529 // initializer to be an undef value, then delete all stores to the
1530 // global. This allows us to mark it constant.
1531 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1532 if (isa<UndefValue>(GV->getInitializer())) {
1533 // Change the initial value here.
1534 GV->setInitializer(SOVConstant);
1536 // Clean up any obviously simplifiable users now.
1537 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1539 if (GV->use_empty()) {
1540 DOUT << " *** Substituting initializer allowed us to "
1541 << "simplify all users and delete global!\n";
1542 GV->eraseFromParent();
1551 // Try to optimize globals based on the knowledge that only one value
1552 // (besides its initializer) is ever stored to the global.
1553 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1554 getAnalysis<TargetData>()))
1557 // Otherwise, if the global was not a boolean, we can shrink it to be a
1559 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1560 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1569 /// OnlyCalledDirectly - Return true if the specified function is only called
1570 /// directly. In other words, its address is never taken.
1571 static bool OnlyCalledDirectly(Function *F) {
1572 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1573 Instruction *User = dyn_cast<Instruction>(*UI);
1574 if (!User) return false;
1575 if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false;
1577 // See if the function address is passed as an argument.
1578 for (unsigned i = 1, e = User->getNumOperands(); i != e; ++i)
1579 if (User->getOperand(i) == F) return false;
1584 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1585 /// function, changing them to FastCC.
1586 static void ChangeCalleesToFastCall(Function *F) {
1587 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1588 CallSite User(cast<Instruction>(*UI));
1589 User.setCallingConv(CallingConv::Fast);
1593 static const ParamAttrsList *StripNest(const ParamAttrsList *Attrs) {
1597 for (unsigned i = 0, e = Attrs->size(); i != e; ++i) {
1598 if ((Attrs->getParamAttrsAtIndex(i) & ParamAttr::Nest) == 0)
1601 Attrs = ParamAttrsList::excludeAttrs(Attrs, Attrs->getParamIndex(i),
1603 // There can be only one.
1610 static void RemoveNestAttribute(Function *F) {
1611 F->setParamAttrs(StripNest(F->getParamAttrs()));
1612 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1613 CallSite User(cast<Instruction>(*UI));
1614 User.setParamAttrs(StripNest(User.getParamAttrs()));
1618 bool GlobalOpt::OptimizeFunctions(Module &M) {
1619 bool Changed = false;
1620 // Optimize functions.
1621 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1623 F->removeDeadConstantUsers();
1624 if (F->use_empty() && (F->hasInternalLinkage() ||
1625 F->hasLinkOnceLinkage())) {
1626 M.getFunctionList().erase(F);
1629 } else if (F->hasInternalLinkage()) {
1630 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1631 OnlyCalledDirectly(F)) {
1632 // If this function has C calling conventions, is not a varargs
1633 // function, and is only called directly, promote it to use the Fast
1634 // calling convention.
1635 F->setCallingConv(CallingConv::Fast);
1636 ChangeCalleesToFastCall(F);
1641 if (F->getParamAttrs() &&
1642 F->getParamAttrs()->hasAttrSomewhere(ParamAttr::Nest) &&
1643 OnlyCalledDirectly(F)) {
1644 // The function is not used by a trampoline intrinsic, so it is safe
1645 // to remove the 'nest' attribute.
1646 RemoveNestAttribute(F);
1655 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1656 bool Changed = false;
1657 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1659 GlobalVariable *GV = GVI++;
1660 if (!GV->isConstant() && GV->hasInternalLinkage() &&
1661 GV->hasInitializer())
1662 Changed |= ProcessInternalGlobal(GV, GVI);
1667 /// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1668 /// initializers have an init priority of 65535.
1669 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1670 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1672 if (I->getName() == "llvm.global_ctors") {
1673 // Found it, verify it's an array of { int, void()* }.
1674 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1676 const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1677 if (!STy || STy->getNumElements() != 2 ||
1678 STy->getElementType(0) != Type::Int32Ty) return 0;
1679 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1680 if (!PFTy) return 0;
1681 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1682 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() ||
1683 FTy->getNumParams() != 0)
1686 // Verify that the initializer is simple enough for us to handle.
1687 if (!I->hasInitializer()) return 0;
1688 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1690 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1691 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(CA->getOperand(i))) {
1692 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1695 // Must have a function or null ptr.
1696 if (!isa<Function>(CS->getOperand(1)))
1699 // Init priority must be standard.
1700 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1701 if (!CI || CI->getZExtValue() != 65535)
1712 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1713 /// return a list of the functions and null terminator as a vector.
1714 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1715 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1716 std::vector<Function*> Result;
1717 Result.reserve(CA->getNumOperands());
1718 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) {
1719 ConstantStruct *CS = cast<ConstantStruct>(CA->getOperand(i));
1720 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1725 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1726 /// specified array, returning the new global to use.
1727 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1728 const std::vector<Function*> &Ctors) {
1729 // If we made a change, reassemble the initializer list.
1730 std::vector<Constant*> CSVals;
1731 CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535));
1732 CSVals.push_back(0);
1734 // Create the new init list.
1735 std::vector<Constant*> CAList;
1736 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1738 CSVals[1] = Ctors[i];
1740 const Type *FTy = FunctionType::get(Type::VoidTy,
1741 std::vector<const Type*>(), false);
1742 const PointerType *PFTy = PointerType::getUnqual(FTy);
1743 CSVals[1] = Constant::getNullValue(PFTy);
1744 CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647);
1746 CAList.push_back(ConstantStruct::get(CSVals));
1749 // Create the array initializer.
1750 const Type *StructTy =
1751 cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
1752 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()),
1755 // If we didn't change the number of elements, don't create a new GV.
1756 if (CA->getType() == GCL->getInitializer()->getType()) {
1757 GCL->setInitializer(CA);
1761 // Create the new global and insert it next to the existing list.
1762 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
1763 GCL->getLinkage(), CA, "",
1765 GCL->isThreadLocal());
1766 GCL->getParent()->getGlobalList().insert(GCL, NGV);
1769 // Nuke the old list, replacing any uses with the new one.
1770 if (!GCL->use_empty()) {
1772 if (V->getType() != GCL->getType())
1773 V = ConstantExpr::getBitCast(V, GCL->getType());
1774 GCL->replaceAllUsesWith(V);
1776 GCL->eraseFromParent();
1785 static Constant *getVal(std::map<Value*, Constant*> &ComputedValues,
1787 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
1788 Constant *R = ComputedValues[V];
1789 assert(R && "Reference to an uncomputed value!");
1793 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
1794 /// enough for us to understand. In particular, if it is a cast of something,
1795 /// we punt. We basically just support direct accesses to globals and GEP's of
1796 /// globals. This should be kept up to date with CommitValueTo.
1797 static bool isSimpleEnoughPointerToCommit(Constant *C) {
1798 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
1799 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1800 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1801 return !GV->isDeclaration(); // reject external globals.
1803 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
1804 // Handle a constantexpr gep.
1805 if (CE->getOpcode() == Instruction::GetElementPtr &&
1806 isa<GlobalVariable>(CE->getOperand(0))) {
1807 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1808 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1809 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1810 return GV->hasInitializer() &&
1811 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1816 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
1817 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
1818 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
1819 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
1820 ConstantExpr *Addr, unsigned OpNo) {
1821 // Base case of the recursion.
1822 if (OpNo == Addr->getNumOperands()) {
1823 assert(Val->getType() == Init->getType() && "Type mismatch!");
1827 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
1828 std::vector<Constant*> Elts;
1830 // Break up the constant into its elements.
1831 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
1832 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i)
1833 Elts.push_back(CS->getOperand(i));
1834 } else if (isa<ConstantAggregateZero>(Init)) {
1835 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1836 Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
1837 } else if (isa<UndefValue>(Init)) {
1838 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1839 Elts.push_back(UndefValue::get(STy->getElementType(i)));
1841 assert(0 && "This code is out of sync with "
1842 " ConstantFoldLoadThroughGEPConstantExpr");
1845 // Replace the element that we are supposed to.
1846 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
1847 unsigned Idx = CU->getZExtValue();
1848 assert(Idx < STy->getNumElements() && "Struct index out of range!");
1849 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
1851 // Return the modified struct.
1852 return ConstantStruct::get(&Elts[0], Elts.size(), STy->isPacked());
1854 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
1855 const ArrayType *ATy = cast<ArrayType>(Init->getType());
1857 // Break up the array into elements.
1858 std::vector<Constant*> Elts;
1859 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
1860 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1861 Elts.push_back(CA->getOperand(i));
1862 } else if (isa<ConstantAggregateZero>(Init)) {
1863 Constant *Elt = Constant::getNullValue(ATy->getElementType());
1864 Elts.assign(ATy->getNumElements(), Elt);
1865 } else if (isa<UndefValue>(Init)) {
1866 Constant *Elt = UndefValue::get(ATy->getElementType());
1867 Elts.assign(ATy->getNumElements(), Elt);
1869 assert(0 && "This code is out of sync with "
1870 " ConstantFoldLoadThroughGEPConstantExpr");
1873 assert(CI->getZExtValue() < ATy->getNumElements());
1874 Elts[CI->getZExtValue()] =
1875 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
1876 return ConstantArray::get(ATy, Elts);
1880 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
1881 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
1882 static void CommitValueTo(Constant *Val, Constant *Addr) {
1883 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
1884 assert(GV->hasInitializer());
1885 GV->setInitializer(Val);
1889 ConstantExpr *CE = cast<ConstantExpr>(Addr);
1890 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1892 Constant *Init = GV->getInitializer();
1893 Init = EvaluateStoreInto(Init, Val, CE, 2);
1894 GV->setInitializer(Init);
1897 /// ComputeLoadResult - Return the value that would be computed by a load from
1898 /// P after the stores reflected by 'memory' have been performed. If we can't
1899 /// decide, return null.
1900 static Constant *ComputeLoadResult(Constant *P,
1901 const std::map<Constant*, Constant*> &Memory) {
1902 // If this memory location has been recently stored, use the stored value: it
1903 // is the most up-to-date.
1904 std::map<Constant*, Constant*>::const_iterator I = Memory.find(P);
1905 if (I != Memory.end()) return I->second;
1908 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
1909 if (GV->hasInitializer())
1910 return GV->getInitializer();
1914 // Handle a constantexpr getelementptr.
1915 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
1916 if (CE->getOpcode() == Instruction::GetElementPtr &&
1917 isa<GlobalVariable>(CE->getOperand(0))) {
1918 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1919 if (GV->hasInitializer())
1920 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1923 return 0; // don't know how to evaluate.
1926 /// EvaluateFunction - Evaluate a call to function F, returning true if
1927 /// successful, false if we can't evaluate it. ActualArgs contains the formal
1928 /// arguments for the function.
1929 static bool EvaluateFunction(Function *F, Constant *&RetVal,
1930 const std::vector<Constant*> &ActualArgs,
1931 std::vector<Function*> &CallStack,
1932 std::map<Constant*, Constant*> &MutatedMemory,
1933 std::vector<GlobalVariable*> &AllocaTmps) {
1934 // Check to see if this function is already executing (recursion). If so,
1935 // bail out. TODO: we might want to accept limited recursion.
1936 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
1939 CallStack.push_back(F);
1941 /// Values - As we compute SSA register values, we store their contents here.
1942 std::map<Value*, Constant*> Values;
1944 // Initialize arguments to the incoming values specified.
1946 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
1948 Values[AI] = ActualArgs[ArgNo];
1950 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
1951 /// we can only evaluate any one basic block at most once. This set keeps
1952 /// track of what we have executed so we can detect recursive cases etc.
1953 std::set<BasicBlock*> ExecutedBlocks;
1955 // CurInst - The current instruction we're evaluating.
1956 BasicBlock::iterator CurInst = F->begin()->begin();
1958 // This is the main evaluation loop.
1960 Constant *InstResult = 0;
1962 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
1963 if (SI->isVolatile()) return false; // no volatile accesses.
1964 Constant *Ptr = getVal(Values, SI->getOperand(1));
1965 if (!isSimpleEnoughPointerToCommit(Ptr))
1966 // If this is too complex for us to commit, reject it.
1968 Constant *Val = getVal(Values, SI->getOperand(0));
1969 MutatedMemory[Ptr] = Val;
1970 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
1971 InstResult = ConstantExpr::get(BO->getOpcode(),
1972 getVal(Values, BO->getOperand(0)),
1973 getVal(Values, BO->getOperand(1)));
1974 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
1975 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
1976 getVal(Values, CI->getOperand(0)),
1977 getVal(Values, CI->getOperand(1)));
1978 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
1979 InstResult = ConstantExpr::getCast(CI->getOpcode(),
1980 getVal(Values, CI->getOperand(0)),
1982 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
1983 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
1984 getVal(Values, SI->getOperand(1)),
1985 getVal(Values, SI->getOperand(2)));
1986 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
1987 Constant *P = getVal(Values, GEP->getOperand(0));
1988 SmallVector<Constant*, 8> GEPOps;
1989 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
1990 GEPOps.push_back(getVal(Values, GEP->getOperand(i)));
1991 InstResult = ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size());
1992 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
1993 if (LI->isVolatile()) return false; // no volatile accesses.
1994 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
1996 if (InstResult == 0) return false; // Could not evaluate load.
1997 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
1998 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs.
1999 const Type *Ty = AI->getType()->getElementType();
2000 AllocaTmps.push_back(new GlobalVariable(Ty, false,
2001 GlobalValue::InternalLinkage,
2002 UndefValue::get(Ty),
2004 InstResult = AllocaTmps.back();
2005 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
2006 // Cannot handle inline asm.
2007 if (isa<InlineAsm>(CI->getOperand(0))) return false;
2009 // Resolve function pointers.
2010 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0)));
2011 if (!Callee) return false; // Cannot resolve.
2013 std::vector<Constant*> Formals;
2014 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
2015 Formals.push_back(getVal(Values, CI->getOperand(i)));
2017 if (Callee->isDeclaration()) {
2018 // If this is a function we can constant fold, do it.
2019 if (Constant *C = ConstantFoldCall(Callee, &Formals[0],
2026 if (Callee->getFunctionType()->isVarArg())
2031 // Execute the call, if successful, use the return value.
2032 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
2033 MutatedMemory, AllocaTmps))
2035 InstResult = RetVal;
2037 } else if (isa<TerminatorInst>(CurInst)) {
2038 BasicBlock *NewBB = 0;
2039 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2040 if (BI->isUnconditional()) {
2041 NewBB = BI->getSuccessor(0);
2044 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
2045 if (!Cond) return false; // Cannot determine.
2047 NewBB = BI->getSuccessor(!Cond->getZExtValue());
2049 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2051 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
2052 if (!Val) return false; // Cannot determine.
2053 NewBB = SI->getSuccessor(SI->findCaseValue(Val));
2054 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
2055 if (RI->getNumOperands())
2056 RetVal = getVal(Values, RI->getOperand(0));
2058 CallStack.pop_back(); // return from fn.
2059 return true; // We succeeded at evaluating this ctor!
2061 // invoke, unwind, unreachable.
2062 return false; // Cannot handle this terminator.
2065 // Okay, we succeeded in evaluating this control flow. See if we have
2066 // executed the new block before. If so, we have a looping function,
2067 // which we cannot evaluate in reasonable time.
2068 if (!ExecutedBlocks.insert(NewBB).second)
2069 return false; // looped!
2071 // Okay, we have never been in this block before. Check to see if there
2072 // are any PHI nodes. If so, evaluate them with information about where
2074 BasicBlock *OldBB = CurInst->getParent();
2075 CurInst = NewBB->begin();
2077 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2078 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
2080 // Do NOT increment CurInst. We know that the terminator had no value.
2083 // Did not know how to evaluate this!
2087 if (!CurInst->use_empty())
2088 Values[CurInst] = InstResult;
2090 // Advance program counter.
2095 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2096 /// we can. Return true if we can, false otherwise.
2097 static bool EvaluateStaticConstructor(Function *F) {
2098 /// MutatedMemory - For each store we execute, we update this map. Loads
2099 /// check this to get the most up-to-date value. If evaluation is successful,
2100 /// this state is committed to the process.
2101 std::map<Constant*, Constant*> MutatedMemory;
2103 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2104 /// to represent its body. This vector is needed so we can delete the
2105 /// temporary globals when we are done.
2106 std::vector<GlobalVariable*> AllocaTmps;
2108 /// CallStack - This is used to detect recursion. In pathological situations
2109 /// we could hit exponential behavior, but at least there is nothing
2111 std::vector<Function*> CallStack;
2113 // Call the function.
2114 Constant *RetValDummy;
2115 bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(),
2116 CallStack, MutatedMemory, AllocaTmps);
2118 // We succeeded at evaluation: commit the result.
2119 DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2120 << F->getName() << "' to " << MutatedMemory.size()
2122 for (std::map<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2123 E = MutatedMemory.end(); I != E; ++I)
2124 CommitValueTo(I->second, I->first);
2127 // At this point, we are done interpreting. If we created any 'alloca'
2128 // temporaries, release them now.
2129 while (!AllocaTmps.empty()) {
2130 GlobalVariable *Tmp = AllocaTmps.back();
2131 AllocaTmps.pop_back();
2133 // If there are still users of the alloca, the program is doing something
2134 // silly, e.g. storing the address of the alloca somewhere and using it
2135 // later. Since this is undefined, we'll just make it be null.
2136 if (!Tmp->use_empty())
2137 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2146 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2147 /// Return true if anything changed.
2148 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2149 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2150 bool MadeChange = false;
2151 if (Ctors.empty()) return false;
2153 // Loop over global ctors, optimizing them when we can.
2154 for (unsigned i = 0; i != Ctors.size(); ++i) {
2155 Function *F = Ctors[i];
2156 // Found a null terminator in the middle of the list, prune off the rest of
2159 if (i != Ctors.size()-1) {
2166 // We cannot simplify external ctor functions.
2167 if (F->empty()) continue;
2169 // If we can evaluate the ctor at compile time, do.
2170 if (EvaluateStaticConstructor(F)) {
2171 Ctors.erase(Ctors.begin()+i);
2174 ++NumCtorsEvaluated;
2179 if (!MadeChange) return false;
2181 GCL = InstallGlobalCtors(GCL, Ctors);
2186 bool GlobalOpt::runOnModule(Module &M) {
2187 bool Changed = false;
2189 // Try to find the llvm.globalctors list.
2190 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2192 bool LocalChange = true;
2193 while (LocalChange) {
2194 LocalChange = false;
2196 // Delete functions that are trivially dead, ccc -> fastcc
2197 LocalChange |= OptimizeFunctions(M);
2199 // Optimize global_ctors list.
2201 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2203 // Optimize non-address-taken globals.
2204 LocalChange |= OptimizeGlobalVars(M);
2205 Changed |= LocalChange;
2208 // TODO: Move all global ctors functions to the end of the module for code