1 //===- GlobalOpt.cpp - Optimize Global Variables --------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source 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/Support/Debug.h"
26 #include "llvm/Target/TargetData.h"
27 #include "llvm/Transforms/Utils/Local.h"
28 #include "llvm/ADT/Statistic.h"
29 #include "llvm/ADT/StringExtras.h"
35 Statistic<> NumMarked ("globalopt", "Number of globals marked constant");
36 Statistic<> NumSRA ("globalopt", "Number of aggregate globals broken "
38 Statistic<> NumSubstitute("globalopt",
39 "Number of globals with initializers stored into them");
40 Statistic<> NumDeleted ("globalopt", "Number of globals deleted");
41 Statistic<> NumFnDeleted("globalopt", "Number of functions deleted");
42 Statistic<> NumGlobUses ("globalopt", "Number of global uses devirtualized");
43 Statistic<> NumLocalized("globalopt", "Number of globals localized");
44 Statistic<> NumShrunkToBool("globalopt",
45 "Number of global vars shrunk to booleans");
46 Statistic<> NumFastCallFns("globalopt",
47 "Number of functions converted to fastcc");
48 Statistic<> NumCtorsEvaluated("globalopt","Number of static ctors evaluated");
50 struct GlobalOpt : public ModulePass {
51 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
52 AU.addRequired<TargetData>();
55 bool runOnModule(Module &M);
58 GlobalVariable *FindGlobalCtors(Module &M);
59 bool OptimizeFunctions(Module &M);
60 bool OptimizeGlobalVars(Module &M);
61 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
62 bool ProcessInternalGlobal(GlobalVariable *GV, Module::global_iterator &GVI);
65 RegisterOpt<GlobalOpt> X("globalopt", "Global Variable Optimizer");
68 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
70 /// GlobalStatus - As we analyze each global, keep track of some information
71 /// about it. If we find out that the address of the global is taken, none of
72 /// this info will be accurate.
74 /// isLoaded - True if the global is ever loaded. If the global isn't ever
75 /// loaded it can be deleted.
78 /// StoredType - Keep track of what stores to the global look like.
81 /// NotStored - There is no store to this global. It can thus be marked
85 /// isInitializerStored - This global is stored to, but the only thing
86 /// stored is the constant it was initialized with. This is only tracked
87 /// for scalar globals.
90 /// isStoredOnce - This global is stored to, but only its initializer and
91 /// one other value is ever stored to it. If this global isStoredOnce, we
92 /// track the value stored to it in StoredOnceValue below. This is only
93 /// tracked for scalar globals.
96 /// isStored - This global is stored to by multiple values or something else
97 /// that we cannot track.
101 /// StoredOnceValue - If only one value (besides the initializer constant) is
102 /// ever stored to this global, keep track of what value it is.
103 Value *StoredOnceValue;
105 // AccessingFunction/HasMultipleAccessingFunctions - These start out
106 // null/false. When the first accessing function is noticed, it is recorded.
107 // When a second different accessing function is noticed,
108 // HasMultipleAccessingFunctions is set to true.
109 Function *AccessingFunction;
110 bool HasMultipleAccessingFunctions;
112 // HasNonInstructionUser - Set to true if this global has a user that is not
113 // an instruction (e.g. a constant expr or GV initializer).
114 bool HasNonInstructionUser;
116 /// isNotSuitableForSRA - Keep track of whether any SRA preventing users of
117 /// the global exist. Such users include GEP instruction with variable
118 /// indexes, and non-gep/load/store users like constant expr casts.
119 bool isNotSuitableForSRA;
121 GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0),
122 AccessingFunction(0), HasMultipleAccessingFunctions(false),
123 HasNonInstructionUser(false), isNotSuitableForSRA(false) {}
128 /// ConstantIsDead - Return true if the specified constant is (transitively)
129 /// dead. The constant may be used by other constants (e.g. constant arrays and
130 /// constant exprs) as long as they are dead, but it cannot be used by anything
132 static bool ConstantIsDead(Constant *C) {
133 if (isa<GlobalValue>(C)) return false;
135 for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI)
136 if (Constant *CU = dyn_cast<Constant>(*UI)) {
137 if (!ConstantIsDead(CU)) return false;
144 /// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
145 /// structure. If the global has its address taken, return true to indicate we
146 /// can't do anything with it.
148 static bool AnalyzeGlobal(Value *V, GlobalStatus &GS,
149 std::set<PHINode*> &PHIUsers) {
150 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
151 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
152 GS.HasNonInstructionUser = true;
154 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
155 if (CE->getOpcode() != Instruction::GetElementPtr)
156 GS.isNotSuitableForSRA = true;
157 else if (!GS.isNotSuitableForSRA) {
158 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
159 // don't like < 3 operand CE's, and we don't like non-constant integer
161 if (CE->getNumOperands() < 3 || !CE->getOperand(1)->isNullValue())
162 GS.isNotSuitableForSRA = true;
164 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
165 if (!isa<ConstantInt>(CE->getOperand(i))) {
166 GS.isNotSuitableForSRA = true;
172 } else if (Instruction *I = dyn_cast<Instruction>(*UI)) {
173 if (!GS.HasMultipleAccessingFunctions) {
174 Function *F = I->getParent()->getParent();
175 if (GS.AccessingFunction == 0)
176 GS.AccessingFunction = F;
177 else if (GS.AccessingFunction != F)
178 GS.HasMultipleAccessingFunctions = true;
180 if (isa<LoadInst>(I)) {
182 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
183 // Don't allow a store OF the address, only stores TO the address.
184 if (SI->getOperand(0) == V) return true;
186 // If this is a direct store to the global (i.e., the global is a scalar
187 // value, not an aggregate), keep more specific information about
189 if (GS.StoredType != GlobalStatus::isStored)
190 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){
191 Value *StoredVal = SI->getOperand(0);
192 if (StoredVal == GV->getInitializer()) {
193 if (GS.StoredType < GlobalStatus::isInitializerStored)
194 GS.StoredType = GlobalStatus::isInitializerStored;
195 } else if (isa<LoadInst>(StoredVal) &&
196 cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
198 if (GS.StoredType < GlobalStatus::isInitializerStored)
199 GS.StoredType = GlobalStatus::isInitializerStored;
200 } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
201 GS.StoredType = GlobalStatus::isStoredOnce;
202 GS.StoredOnceValue = StoredVal;
203 } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
204 GS.StoredOnceValue == StoredVal) {
207 GS.StoredType = GlobalStatus::isStored;
210 GS.StoredType = GlobalStatus::isStored;
212 } else if (isa<GetElementPtrInst>(I)) {
213 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
215 // If the first two indices are constants, this can be SRA'd.
216 if (isa<GlobalVariable>(I->getOperand(0))) {
217 if (I->getNumOperands() < 3 || !isa<Constant>(I->getOperand(1)) ||
218 !cast<Constant>(I->getOperand(1))->isNullValue() ||
219 !isa<ConstantInt>(I->getOperand(2)))
220 GS.isNotSuitableForSRA = true;
221 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(I->getOperand(0))){
222 if (CE->getOpcode() != Instruction::GetElementPtr ||
223 CE->getNumOperands() < 3 || I->getNumOperands() < 2 ||
224 !isa<Constant>(I->getOperand(0)) ||
225 !cast<Constant>(I->getOperand(0))->isNullValue())
226 GS.isNotSuitableForSRA = true;
228 GS.isNotSuitableForSRA = true;
230 } else if (isa<SelectInst>(I)) {
231 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
232 GS.isNotSuitableForSRA = true;
233 } else if (PHINode *PN = dyn_cast<PHINode>(I)) {
234 // PHI nodes we can check just like select or GEP instructions, but we
235 // have to be careful about infinite recursion.
236 if (PHIUsers.insert(PN).second) // Not already visited.
237 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
238 GS.isNotSuitableForSRA = true;
239 } else if (isa<SetCondInst>(I)) {
240 GS.isNotSuitableForSRA = true;
241 } else if (isa<MemCpyInst>(I) || isa<MemMoveInst>(I)) {
242 if (I->getOperand(1) == V)
243 GS.StoredType = GlobalStatus::isStored;
244 if (I->getOperand(2) == V)
246 GS.isNotSuitableForSRA = true;
247 } else if (isa<MemSetInst>(I)) {
248 assert(I->getOperand(1) == V && "Memset only takes one pointer!");
249 GS.StoredType = GlobalStatus::isStored;
250 GS.isNotSuitableForSRA = true;
252 return true; // Any other non-load instruction might take address!
254 } else if (Constant *C = dyn_cast<Constant>(*UI)) {
255 GS.HasNonInstructionUser = true;
256 // We might have a dead and dangling constant hanging off of here.
257 if (!ConstantIsDead(C))
260 GS.HasNonInstructionUser = true;
261 // Otherwise must be some other user.
268 static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) {
269 ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
271 unsigned IdxV = (unsigned)CI->getRawValue();
273 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
274 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
275 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
276 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
277 } else if (ConstantPacked *CP = dyn_cast<ConstantPacked>(Agg)) {
278 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
279 } else if (isa<ConstantAggregateZero>(Agg)) {
280 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
281 if (IdxV < STy->getNumElements())
282 return Constant::getNullValue(STy->getElementType(IdxV));
283 } else if (const SequentialType *STy =
284 dyn_cast<SequentialType>(Agg->getType())) {
285 return Constant::getNullValue(STy->getElementType());
287 } else if (isa<UndefValue>(Agg)) {
288 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
289 if (IdxV < STy->getNumElements())
290 return UndefValue::get(STy->getElementType(IdxV));
291 } else if (const SequentialType *STy =
292 dyn_cast<SequentialType>(Agg->getType())) {
293 return UndefValue::get(STy->getElementType());
300 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
301 /// users of the global, cleaning up the obvious ones. This is largely just a
302 /// quick scan over the use list to clean up the easy and obvious cruft. This
303 /// returns true if it made a change.
304 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
305 bool Changed = false;
306 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
309 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
311 // Replace the load with the initializer.
312 LI->replaceAllUsesWith(Init);
313 LI->eraseFromParent();
316 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
317 // Store must be unreachable or storing Init into the global.
318 SI->eraseFromParent();
320 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
321 if (CE->getOpcode() == Instruction::GetElementPtr) {
322 Constant *SubInit = 0;
324 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
325 Changed |= CleanupConstantGlobalUsers(CE, SubInit);
326 } else if (CE->getOpcode() == Instruction::Cast &&
327 isa<PointerType>(CE->getType())) {
328 // Pointer cast, delete any stores and memsets to the global.
329 Changed |= CleanupConstantGlobalUsers(CE, 0);
332 if (CE->use_empty()) {
333 CE->destroyConstant();
336 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
337 Constant *SubInit = 0;
339 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
340 if (CE && CE->getOpcode() == Instruction::GetElementPtr)
341 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
342 Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
344 if (GEP->use_empty()) {
345 GEP->eraseFromParent();
348 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
349 if (MI->getRawDest() == V) {
350 MI->eraseFromParent();
354 } else if (Constant *C = dyn_cast<Constant>(U)) {
355 // If we have a chain of dead constantexprs or other things dangling from
356 // us, and if they are all dead, nuke them without remorse.
357 if (ConstantIsDead(C)) {
358 C->destroyConstant();
359 // This could have invalidated UI, start over from scratch.
360 CleanupConstantGlobalUsers(V, Init);
368 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
369 /// variable. This opens the door for other optimizations by exposing the
370 /// behavior of the program in a more fine-grained way. We have determined that
371 /// this transformation is safe already. We return the first global variable we
372 /// insert so that the caller can reprocess it.
373 static GlobalVariable *SRAGlobal(GlobalVariable *GV) {
374 assert(GV->hasInternalLinkage() && !GV->isConstant());
375 Constant *Init = GV->getInitializer();
376 const Type *Ty = Init->getType();
378 std::vector<GlobalVariable*> NewGlobals;
379 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
381 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
382 NewGlobals.reserve(STy->getNumElements());
383 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
384 Constant *In = getAggregateConstantElement(Init,
385 ConstantUInt::get(Type::UIntTy, i));
386 assert(In && "Couldn't get element of initializer?");
387 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
388 GlobalVariable::InternalLinkage,
389 In, GV->getName()+"."+utostr(i));
390 Globals.insert(GV, NGV);
391 NewGlobals.push_back(NGV);
393 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
394 unsigned NumElements = 0;
395 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
396 NumElements = ATy->getNumElements();
397 else if (const PackedType *PTy = dyn_cast<PackedType>(STy))
398 NumElements = PTy->getNumElements();
400 assert(0 && "Unknown aggregate sequential type!");
402 if (NumElements > 16 && GV->hasNUsesOrMore(16))
403 return 0; // It's not worth it.
404 NewGlobals.reserve(NumElements);
405 for (unsigned i = 0, e = NumElements; i != e; ++i) {
406 Constant *In = getAggregateConstantElement(Init,
407 ConstantUInt::get(Type::UIntTy, i));
408 assert(In && "Couldn't get element of initializer?");
410 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
411 GlobalVariable::InternalLinkage,
412 In, GV->getName()+"."+utostr(i));
413 Globals.insert(GV, NGV);
414 NewGlobals.push_back(NGV);
418 if (NewGlobals.empty())
421 DEBUG(std::cerr << "PERFORMING GLOBAL SRA ON: " << *GV);
423 Constant *NullInt = Constant::getNullValue(Type::IntTy);
425 // Loop over all of the uses of the global, replacing the constantexpr geps,
426 // with smaller constantexpr geps or direct references.
427 while (!GV->use_empty()) {
428 User *GEP = GV->use_back();
429 assert(((isa<ConstantExpr>(GEP) &&
430 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
431 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
433 // Ignore the 1th operand, which has to be zero or else the program is quite
434 // broken (undefined). Get the 2nd operand, which is the structure or array
437 (unsigned)cast<ConstantInt>(GEP->getOperand(2))->getRawValue();
438 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
440 Value *NewPtr = NewGlobals[Val];
442 // Form a shorter GEP if needed.
443 if (GEP->getNumOperands() > 3)
444 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
445 std::vector<Constant*> Idxs;
446 Idxs.push_back(NullInt);
447 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
448 Idxs.push_back(CE->getOperand(i));
449 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr), Idxs);
451 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
452 std::vector<Value*> Idxs;
453 Idxs.push_back(NullInt);
454 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
455 Idxs.push_back(GEPI->getOperand(i));
456 NewPtr = new GetElementPtrInst(NewPtr, Idxs,
457 GEPI->getName()+"."+utostr(Val), GEPI);
459 GEP->replaceAllUsesWith(NewPtr);
461 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
462 GEPI->eraseFromParent();
464 cast<ConstantExpr>(GEP)->destroyConstant();
467 // Delete the old global, now that it is dead.
471 // Loop over the new globals array deleting any globals that are obviously
472 // dead. This can arise due to scalarization of a structure or an array that
473 // has elements that are dead.
474 unsigned FirstGlobal = 0;
475 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
476 if (NewGlobals[i]->use_empty()) {
477 Globals.erase(NewGlobals[i]);
478 if (FirstGlobal == i) ++FirstGlobal;
481 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
484 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
485 /// value will trap if the value is dynamically null.
486 static bool AllUsesOfValueWillTrapIfNull(Value *V) {
487 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
488 if (isa<LoadInst>(*UI)) {
490 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
491 if (SI->getOperand(0) == V) {
492 //std::cerr << "NONTRAPPING USE: " << **UI;
493 return false; // Storing the value.
495 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
496 if (CI->getOperand(0) != V) {
497 //std::cerr << "NONTRAPPING USE: " << **UI;
498 return false; // Not calling the ptr
500 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
501 if (II->getOperand(0) != V) {
502 //std::cerr << "NONTRAPPING USE: " << **UI;
503 return false; // Not calling the ptr
505 } else if (CastInst *CI = dyn_cast<CastInst>(*UI)) {
506 if (!AllUsesOfValueWillTrapIfNull(CI)) return false;
507 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
508 if (!AllUsesOfValueWillTrapIfNull(GEPI)) return false;
509 } else if (isa<SetCondInst>(*UI) &&
510 isa<ConstantPointerNull>(UI->getOperand(1))) {
511 // Ignore setcc X, null
513 //std::cerr << "NONTRAPPING USE: " << **UI;
519 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
520 /// from GV will trap if the loaded value is null. Note that this also permits
521 /// comparisons of the loaded value against null, as a special case.
522 static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) {
523 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI)
524 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
525 if (!AllUsesOfValueWillTrapIfNull(LI))
527 } else if (isa<StoreInst>(*UI)) {
528 // Ignore stores to the global.
530 // We don't know or understand this user, bail out.
531 //std::cerr << "UNKNOWN USER OF GLOBAL!: " << **UI;
538 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
539 bool Changed = false;
540 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
541 Instruction *I = cast<Instruction>(*UI++);
542 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
543 LI->setOperand(0, NewV);
545 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
546 if (SI->getOperand(1) == V) {
547 SI->setOperand(1, NewV);
550 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
551 if (I->getOperand(0) == V) {
552 // Calling through the pointer! Turn into a direct call, but be careful
553 // that the pointer is not also being passed as an argument.
554 I->setOperand(0, NewV);
556 bool PassedAsArg = false;
557 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
558 if (I->getOperand(i) == V) {
560 I->setOperand(i, NewV);
564 // Being passed as an argument also. Be careful to not invalidate UI!
568 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
569 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
570 ConstantExpr::getCast(NewV, CI->getType()));
571 if (CI->use_empty()) {
573 CI->eraseFromParent();
575 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
576 // Should handle GEP here.
577 std::vector<Constant*> Indices;
578 Indices.reserve(GEPI->getNumOperands()-1);
579 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
580 if (Constant *C = dyn_cast<Constant>(GEPI->getOperand(i)))
581 Indices.push_back(C);
584 if (Indices.size() == GEPI->getNumOperands()-1)
585 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
586 ConstantExpr::getGetElementPtr(NewV, Indices));
587 if (GEPI->use_empty()) {
589 GEPI->eraseFromParent();
598 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
599 /// value stored into it. If there are uses of the loaded value that would trap
600 /// if the loaded value is dynamically null, then we know that they cannot be
601 /// reachable with a null optimize away the load.
602 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) {
603 std::vector<LoadInst*> Loads;
604 bool Changed = false;
606 // Replace all uses of loads with uses of uses of the stored value.
607 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end();
609 if (LoadInst *LI = dyn_cast<LoadInst>(*GUI)) {
611 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
613 assert(isa<StoreInst>(*GUI) && "Only expect load and stores!");
617 DEBUG(std::cerr << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV);
621 // Delete all of the loads we can, keeping track of whether we nuked them all!
622 bool AllLoadsGone = true;
623 while (!Loads.empty()) {
624 LoadInst *L = Loads.back();
625 if (L->use_empty()) {
626 L->eraseFromParent();
629 AllLoadsGone = false;
634 // If we nuked all of the loads, then none of the stores are needed either,
635 // nor is the global.
637 DEBUG(std::cerr << " *** GLOBAL NOW DEAD!\n");
638 CleanupConstantGlobalUsers(GV, 0);
639 if (GV->use_empty()) {
640 GV->eraseFromParent();
648 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
649 /// instructions that are foldable.
650 static void ConstantPropUsersOf(Value *V) {
651 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
652 if (Instruction *I = dyn_cast<Instruction>(*UI++))
653 if (Constant *NewC = ConstantFoldInstruction(I)) {
654 I->replaceAllUsesWith(NewC);
656 // Advance UI to the next non-I use to avoid invalidating it!
657 // Instructions could multiply use V.
658 while (UI != E && *UI == I)
660 I->eraseFromParent();
664 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
665 /// variable, and transforms the program as if it always contained the result of
666 /// the specified malloc. Because it is always the result of the specified
667 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
668 /// malloc into a global, and any laods of GV as uses of the new global.
669 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
671 DEBUG(std::cerr << "PROMOTING MALLOC GLOBAL: " << *GV << " MALLOC = " <<*MI);
672 ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize());
674 if (NElements->getRawValue() != 1) {
675 // If we have an array allocation, transform it to a single element
676 // allocation to make the code below simpler.
677 Type *NewTy = ArrayType::get(MI->getAllocatedType(),
678 (unsigned)NElements->getRawValue());
680 new MallocInst(NewTy, Constant::getNullValue(Type::UIntTy),
682 std::vector<Value*> Indices;
683 Indices.push_back(Constant::getNullValue(Type::IntTy));
684 Indices.push_back(Indices[0]);
685 Value *NewGEP = new GetElementPtrInst(NewMI, Indices,
686 NewMI->getName()+".el0", MI);
687 MI->replaceAllUsesWith(NewGEP);
688 MI->eraseFromParent();
692 // Create the new global variable. The contents of the malloc'd memory is
693 // undefined, so initialize with an undef value.
694 Constant *Init = UndefValue::get(MI->getAllocatedType());
695 GlobalVariable *NewGV = new GlobalVariable(MI->getAllocatedType(), false,
696 GlobalValue::InternalLinkage, Init,
697 GV->getName()+".body");
698 GV->getParent()->getGlobalList().insert(GV, NewGV);
700 // Anything that used the malloc now uses the global directly.
701 MI->replaceAllUsesWith(NewGV);
703 Constant *RepValue = NewGV;
704 if (NewGV->getType() != GV->getType()->getElementType())
705 RepValue = ConstantExpr::getCast(RepValue, GV->getType()->getElementType());
707 // If there is a comparison against null, we will insert a global bool to
708 // keep track of whether the global was initialized yet or not.
709 GlobalVariable *InitBool =
710 new GlobalVariable(Type::BoolTy, false, GlobalValue::InternalLinkage,
711 ConstantBool::False, GV->getName()+".init");
712 bool InitBoolUsed = false;
714 // Loop over all uses of GV, processing them in turn.
715 std::vector<StoreInst*> Stores;
716 while (!GV->use_empty())
717 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
718 while (!LI->use_empty()) {
719 Use &LoadUse = LI->use_begin().getUse();
720 if (!isa<SetCondInst>(LoadUse.getUser()))
723 // Replace the setcc X, 0 with a use of the bool value.
724 SetCondInst *SCI = cast<SetCondInst>(LoadUse.getUser());
725 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", SCI);
727 switch (SCI->getOpcode()) {
728 default: assert(0 && "Unknown opcode!");
729 case Instruction::SetLT:
730 LV = ConstantBool::False; // X < null -> always false
732 case Instruction::SetEQ:
733 case Instruction::SetLE:
734 LV = BinaryOperator::createNot(LV, "notinit", SCI);
736 case Instruction::SetNE:
737 case Instruction::SetGE:
738 case Instruction::SetGT:
741 SCI->replaceAllUsesWith(LV);
742 SCI->eraseFromParent();
745 LI->eraseFromParent();
747 StoreInst *SI = cast<StoreInst>(GV->use_back());
748 // The global is initialized when the store to it occurs.
749 new StoreInst(ConstantBool::True, InitBool, SI);
750 SI->eraseFromParent();
753 // If the initialization boolean was used, insert it, otherwise delete it.
755 while (!InitBool->use_empty()) // Delete initializations
756 cast<Instruction>(InitBool->use_back())->eraseFromParent();
759 GV->getParent()->getGlobalList().insert(GV, InitBool);
762 // Now the GV is dead, nuke it and the malloc.
763 GV->eraseFromParent();
764 MI->eraseFromParent();
766 // To further other optimizations, loop over all users of NewGV and try to
767 // constant prop them. This will promote GEP instructions with constant
768 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
769 ConstantPropUsersOf(NewGV);
770 if (RepValue != NewGV)
771 ConstantPropUsersOf(RepValue);
776 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
777 /// to make sure that there are no complex uses of V. We permit simple things
778 /// like dereferencing the pointer, but not storing through the address, unless
779 /// it is to the specified global.
780 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V,
781 GlobalVariable *GV) {
782 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI)
783 if (isa<LoadInst>(*UI) || isa<SetCondInst>(*UI)) {
785 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
786 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
787 return false; // Storing the pointer itself... bad.
788 // Otherwise, storing through it, or storing into GV... fine.
789 } else if (isa<GetElementPtrInst>(*UI) || isa<SelectInst>(*UI)) {
790 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(cast<Instruction>(*UI),GV))
799 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
800 // that only one value (besides its initializer) is ever stored to the global.
801 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
802 Module::global_iterator &GVI, TargetData &TD) {
803 if (CastInst *CI = dyn_cast<CastInst>(StoredOnceVal))
804 StoredOnceVal = CI->getOperand(0);
805 else if (GetElementPtrInst *GEPI =dyn_cast<GetElementPtrInst>(StoredOnceVal)){
806 // "getelementptr Ptr, 0, 0, 0" is really just a cast.
807 bool IsJustACast = true;
808 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
809 if (!isa<Constant>(GEPI->getOperand(i)) ||
810 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
815 StoredOnceVal = GEPI->getOperand(0);
818 // If we are dealing with a pointer global that is initialized to null and
819 // only has one (non-null) value stored into it, then we can optimize any
820 // users of the loaded value (often calls and loads) that would trap if the
822 if (isa<PointerType>(GV->getInitializer()->getType()) &&
823 GV->getInitializer()->isNullValue()) {
824 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
825 if (GV->getInitializer()->getType() != SOVC->getType())
826 SOVC = ConstantExpr::getCast(SOVC, GV->getInitializer()->getType());
828 // Optimize away any trapping uses of the loaded value.
829 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
831 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) {
832 // If we have a global that is only initialized with a fixed size malloc,
833 // and if all users of the malloc trap, and if the malloc'd address is not
834 // put anywhere else, transform the program to use global memory instead
835 // of malloc'd memory. This eliminates dynamic allocation (good) and
836 // exposes the resultant global to further GlobalOpt (even better). Note
837 // that we restrict this transformation to only working on small
838 // allocations (2048 bytes currently), as we don't want to introduce a 16M
839 // global or something.
840 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize()))
841 if (MI->getAllocatedType()->isSized() &&
842 NElements->getRawValue()*
843 TD.getTypeSize(MI->getAllocatedType()) < 2048 &&
844 AllUsesOfLoadedValueWillTrapIfNull(GV) &&
845 ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV)) {
846 GVI = OptimizeGlobalAddressOfMalloc(GV, MI);
855 /// ShrinkGlobalToBoolean - At this point, we have learned that the only two
856 /// values ever stored into GV are its initializer and OtherVal.
857 static void ShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
858 // Create the new global, initializing it to false.
859 GlobalVariable *NewGV = new GlobalVariable(Type::BoolTy, false,
860 GlobalValue::InternalLinkage, ConstantBool::False, GV->getName()+".b");
861 GV->getParent()->getGlobalList().insert(GV, NewGV);
863 Constant *InitVal = GV->getInitializer();
864 assert(InitVal->getType() != Type::BoolTy && "No reason to shrink to bool!");
866 // If initialized to zero and storing one into the global, we can use a cast
867 // instead of a select to synthesize the desired value.
868 bool IsOneZero = false;
869 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
870 IsOneZero = InitVal->isNullValue() && CI->equalsInt(1);
872 while (!GV->use_empty()) {
873 Instruction *UI = cast<Instruction>(GV->use_back());
874 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
875 // Change the store into a boolean store.
876 bool StoringOther = SI->getOperand(0) == OtherVal;
877 // Only do this if we weren't storing a loaded value.
879 if (StoringOther || SI->getOperand(0) == InitVal)
880 StoreVal = ConstantBool::get(StoringOther);
882 // Otherwise, we are storing a previously loaded copy. To do this,
883 // change the copy from copying the original value to just copying the
885 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
887 // If we're already replaced the input, StoredVal will be a cast or
888 // select instruction. If not, it will be a load of the original
890 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
891 assert(LI->getOperand(0) == GV && "Not a copy!");
892 // Insert a new load, to preserve the saved value.
893 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
895 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
896 "This is not a form that we understand!");
897 StoreVal = StoredVal->getOperand(0);
898 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
901 new StoreInst(StoreVal, NewGV, SI);
902 } else if (!UI->use_empty()) {
903 // Change the load into a load of bool then a select.
904 LoadInst *LI = cast<LoadInst>(UI);
906 std::string Name = LI->getName(); LI->setName("");
907 LoadInst *NLI = new LoadInst(NewGV, Name+".b", LI);
910 NSI = new CastInst(NLI, LI->getType(), Name, LI);
912 NSI = new SelectInst(NLI, OtherVal, InitVal, Name, LI);
913 LI->replaceAllUsesWith(NSI);
915 UI->eraseFromParent();
918 GV->eraseFromParent();
922 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
923 /// it if possible. If we make a change, return true.
924 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
925 Module::global_iterator &GVI) {
926 std::set<PHINode*> PHIUsers;
928 GV->removeDeadConstantUsers();
930 if (GV->use_empty()) {
931 DEBUG(std::cerr << "GLOBAL DEAD: " << *GV);
932 GV->eraseFromParent();
937 if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
938 // If this is a first class global and has only one accessing function
939 // and this function is main (which we know is not recursive we can make
940 // this global a local variable) we replace the global with a local alloca
943 // NOTE: It doesn't make sense to promote non first class types since we
944 // are just replacing static memory to stack memory.
945 if (!GS.HasMultipleAccessingFunctions &&
946 GS.AccessingFunction && !GS.HasNonInstructionUser &&
947 GV->getType()->getElementType()->isFirstClassType() &&
948 GS.AccessingFunction->getName() == "main" &&
949 GS.AccessingFunction->hasExternalLinkage()) {
950 DEBUG(std::cerr << "LOCALIZING GLOBAL: " << *GV);
951 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin();
952 const Type* ElemTy = GV->getType()->getElementType();
953 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI);
954 if (!isa<UndefValue>(GV->getInitializer()))
955 new StoreInst(GV->getInitializer(), Alloca, FirstI);
957 GV->replaceAllUsesWith(Alloca);
958 GV->eraseFromParent();
962 // If the global is never loaded (but may be stored to), it is dead.
965 DEBUG(std::cerr << "GLOBAL NEVER LOADED: " << *GV);
967 // Delete any stores we can find to the global. We may not be able to
968 // make it completely dead though.
969 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
971 // If the global is dead now, delete it.
972 if (GV->use_empty()) {
973 GV->eraseFromParent();
979 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
980 DEBUG(std::cerr << "MARKING CONSTANT: " << *GV);
981 GV->setConstant(true);
983 // Clean up any obviously simplifiable users now.
984 CleanupConstantGlobalUsers(GV, GV->getInitializer());
986 // If the global is dead now, just nuke it.
987 if (GV->use_empty()) {
988 DEBUG(std::cerr << " *** Marking constant allowed us to simplify "
989 "all users and delete global!\n");
990 GV->eraseFromParent();
996 } else if (!GS.isNotSuitableForSRA &&
997 !GV->getInitializer()->getType()->isFirstClassType()) {
998 if (GlobalVariable *FirstNewGV = SRAGlobal(GV)) {
999 GVI = FirstNewGV; // Don't skip the newly produced globals!
1002 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1003 // If the initial value for the global was an undef value, and if only
1004 // one other value was stored into it, we can just change the
1005 // initializer to be an undef value, then delete all stores to the
1006 // global. This allows us to mark it constant.
1007 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1008 if (isa<UndefValue>(GV->getInitializer())) {
1009 // Change the initial value here.
1010 GV->setInitializer(SOVConstant);
1012 // Clean up any obviously simplifiable users now.
1013 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1015 if (GV->use_empty()) {
1016 DEBUG(std::cerr << " *** Substituting initializer allowed us to "
1017 "simplify all users and delete global!\n");
1018 GV->eraseFromParent();
1027 // Try to optimize globals based on the knowledge that only one value
1028 // (besides its initializer) is ever stored to the global.
1029 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1030 getAnalysis<TargetData>()))
1033 // Otherwise, if the global was not a boolean, we can shrink it to be a
1035 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1036 if (GV->getType()->getElementType() != Type::BoolTy &&
1037 !GV->getType()->getElementType()->isFloatingPoint()) {
1038 DEBUG(std::cerr << " *** SHRINKING TO BOOL: " << *GV);
1039 ShrinkGlobalToBoolean(GV, SOVConstant);
1048 /// OnlyCalledDirectly - Return true if the specified function is only called
1049 /// directly. In other words, its address is never taken.
1050 static bool OnlyCalledDirectly(Function *F) {
1051 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1052 Instruction *User = dyn_cast<Instruction>(*UI);
1053 if (!User) return false;
1054 if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false;
1056 // See if the function address is passed as an argument.
1057 for (unsigned i = 1, e = User->getNumOperands(); i != e; ++i)
1058 if (User->getOperand(i) == F) return false;
1063 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1064 /// function, changing them to FastCC.
1065 static void ChangeCalleesToFastCall(Function *F) {
1066 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1067 Instruction *User = cast<Instruction>(*UI);
1068 if (CallInst *CI = dyn_cast<CallInst>(User))
1069 CI->setCallingConv(CallingConv::Fast);
1071 cast<InvokeInst>(User)->setCallingConv(CallingConv::Fast);
1075 bool GlobalOpt::OptimizeFunctions(Module &M) {
1076 bool Changed = false;
1077 // Optimize functions.
1078 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1080 F->removeDeadConstantUsers();
1081 if (F->use_empty() && (F->hasInternalLinkage() ||
1082 F->hasLinkOnceLinkage())) {
1083 M.getFunctionList().erase(F);
1086 } else if (F->hasInternalLinkage() &&
1087 F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1088 OnlyCalledDirectly(F)) {
1089 // If this function has C calling conventions, is not a varargs
1090 // function, and is only called directly, promote it to use the Fast
1091 // calling convention.
1092 F->setCallingConv(CallingConv::Fast);
1093 ChangeCalleesToFastCall(F);
1101 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1102 bool Changed = false;
1103 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1105 GlobalVariable *GV = GVI++;
1106 if (!GV->isConstant() && GV->hasInternalLinkage() &&
1107 GV->hasInitializer())
1108 Changed |= ProcessInternalGlobal(GV, GVI);
1113 /// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1114 /// initializers have an init priority of 65535.
1115 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1116 for (Module::giterator I = M.global_begin(), E = M.global_end(); I != E; ++I)
1117 if (I->getName() == "llvm.global_ctors") {
1118 // Found it, verify it's an array of { int, void()* }.
1119 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1121 const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1122 if (!STy || STy->getNumElements() != 2 ||
1123 STy->getElementType(0) != Type::IntTy) return 0;
1124 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1125 if (!PFTy) return 0;
1126 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1127 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() ||
1128 FTy->getNumParams() != 0)
1131 // Verify that the initializer is simple enough for us to handle.
1132 if (!I->hasInitializer()) return 0;
1133 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1135 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1136 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(CA->getOperand(i))) {
1137 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1140 // Must have a function or null ptr.
1141 if (!isa<Function>(CS->getOperand(1)))
1144 // Init priority must be standard.
1145 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1146 if (!CI || CI->getRawValue() != 65535)
1157 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1158 /// return a list of the functions and null terminator as a vector.
1159 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1160 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1161 std::vector<Function*> Result;
1162 Result.reserve(CA->getNumOperands());
1163 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) {
1164 ConstantStruct *CS = cast<ConstantStruct>(CA->getOperand(i));
1165 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1170 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1171 /// specified array, returning the new global to use.
1172 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1173 const std::vector<Function*> &Ctors) {
1174 // If we made a change, reassemble the initializer list.
1175 std::vector<Constant*> CSVals;
1176 CSVals.push_back(ConstantSInt::get(Type::IntTy, 65535));
1177 CSVals.push_back(0);
1179 // Create the new init list.
1180 std::vector<Constant*> CAList;
1181 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1183 CSVals[1] = Ctors[i];
1185 const Type *FTy = FunctionType::get(Type::VoidTy,
1186 std::vector<const Type*>(), false);
1187 const PointerType *PFTy = PointerType::get(FTy);
1188 CSVals[1] = Constant::getNullValue(PFTy);
1189 CSVals[0] = ConstantSInt::get(Type::IntTy, 2147483647);
1191 CAList.push_back(ConstantStruct::get(CSVals));
1194 // Create the array initializer.
1195 const Type *StructTy =
1196 cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
1197 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()),
1200 // If we didn't change the number of elements, don't create a new GV.
1201 if (CA->getType() == GCL->getInitializer()->getType()) {
1202 GCL->setInitializer(CA);
1206 // Create the new global and insert it next to the existing list.
1207 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
1208 GCL->getLinkage(), CA,
1211 GCL->getParent()->getGlobalList().insert(GCL, NGV);
1213 // Nuke the old list, replacing any uses with the new one.
1214 if (!GCL->use_empty()) {
1216 if (V->getType() != GCL->getType())
1217 V = ConstantExpr::getCast(V, GCL->getType());
1218 GCL->replaceAllUsesWith(V);
1220 GCL->eraseFromParent();
1229 static Constant *getVal(std::map<Value*, Constant*> &ComputedValues,
1231 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
1232 Constant *R = ComputedValues[V];
1233 assert(R && "Reference to an uncomputed value!");
1237 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
1238 /// enough for us to understand. In particular, if it is a cast of something,
1239 /// we punt. We basically just support direct accesses to globals and GEP's of
1240 /// globals. This should be kept up to date with CommitValueTo.
1241 static bool isSimpleEnoughPointerToCommit(Constant *C) {
1242 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
1243 return !GV->isExternal(); // reject external globals.
1244 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
1245 // Handle a constantexpr gep.
1246 if (CE->getOpcode() == Instruction::GetElementPtr &&
1247 isa<GlobalVariable>(CE->getOperand(0))) {
1248 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1249 return GV->hasInitializer() &&
1250 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1255 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
1256 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
1257 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
1258 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
1259 ConstantExpr *Addr, unsigned OpNo) {
1260 // Base case of the recursion.
1261 if (OpNo == Addr->getNumOperands()) {
1262 assert(Val->getType() == Init->getType() && "Type mismatch!");
1266 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
1267 std::vector<Constant*> Elts;
1269 // Break up the constant into its elements.
1270 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
1271 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i)
1272 Elts.push_back(CS->getOperand(i));
1273 } else if (isa<ConstantAggregateZero>(Init)) {
1274 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1275 Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
1276 } else if (isa<UndefValue>(Init)) {
1277 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1278 Elts.push_back(UndefValue::get(STy->getElementType(i)));
1280 assert(0 && "This code is out of sync with "
1281 " ConstantFoldLoadThroughGEPConstantExpr");
1284 // Replace the element that we are supposed to.
1285 ConstantUInt *CU = cast<ConstantUInt>(Addr->getOperand(OpNo));
1286 assert(CU->getValue() < STy->getNumElements() &&
1287 "Struct index out of range!");
1288 unsigned Idx = (unsigned)CU->getValue();
1289 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
1291 // Return the modified struct.
1292 return ConstantStruct::get(Elts);
1294 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
1295 const ArrayType *ATy = cast<ArrayType>(Init->getType());
1297 // Break up the array into elements.
1298 std::vector<Constant*> Elts;
1299 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
1300 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1301 Elts.push_back(CA->getOperand(i));
1302 } else if (isa<ConstantAggregateZero>(Init)) {
1303 Constant *Elt = Constant::getNullValue(ATy->getElementType());
1304 Elts.assign(ATy->getNumElements(), Elt);
1305 } else if (isa<UndefValue>(Init)) {
1306 Constant *Elt = UndefValue::get(ATy->getElementType());
1307 Elts.assign(ATy->getNumElements(), Elt);
1309 assert(0 && "This code is out of sync with "
1310 " ConstantFoldLoadThroughGEPConstantExpr");
1313 assert((uint64_t)CI->getRawValue() < ATy->getNumElements());
1314 Elts[(uint64_t)CI->getRawValue()] =
1315 EvaluateStoreInto(Elts[(uint64_t)CI->getRawValue()], Val, Addr, OpNo+1);
1316 return ConstantArray::get(ATy, Elts);
1320 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
1321 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
1322 static void CommitValueTo(Constant *Val, Constant *Addr) {
1323 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
1324 assert(GV->hasInitializer());
1325 GV->setInitializer(Val);
1329 ConstantExpr *CE = cast<ConstantExpr>(Addr);
1330 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1332 Constant *Init = GV->getInitializer();
1333 Init = EvaluateStoreInto(Init, Val, CE, 2);
1334 GV->setInitializer(Init);
1337 /// ComputeLoadResult - Return the value that would be computed by a load from
1338 /// P after the stores reflected by 'memory' have been performed. If we can't
1339 /// decide, return null.
1340 static Constant *ComputeLoadResult(Constant *P,
1341 const std::map<Constant*, Constant*> &Memory) {
1342 // If this memory location has been recently stored, use the stored value: it
1343 // is the most up-to-date.
1344 std::map<Constant*, Constant*>::const_iterator I = Memory.find(P);
1345 if (I != Memory.end()) return I->second;
1348 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
1349 if (GV->hasInitializer())
1350 return GV->getInitializer();
1354 // Handle a constantexpr getelementptr.
1355 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
1356 if (CE->getOpcode() == Instruction::GetElementPtr &&
1357 isa<GlobalVariable>(CE->getOperand(0))) {
1358 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1359 if (GV->hasInitializer())
1360 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1363 return 0; // don't know how to evaluate.
1366 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
1367 /// we can. Return true if we can, false otherwise.
1368 static bool EvaluateStaticConstructor(Function *F) {
1369 /// Values - As we compute SSA register values, we store their contents here.
1370 std::map<Value*, Constant*> Values;
1372 /// MutatedMemory - For each store we execute, we update this map. Loads
1373 /// check this to get the most up-to-date value. If evaluation is successful,
1374 /// this state is committed to the process.
1375 std::map<Constant*, Constant*> MutatedMemory;
1377 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
1378 /// we can only evaluate any one basic block at most once. This set keeps
1379 /// track of what we have executed so we can detect recursive cases etc.
1380 std::set<BasicBlock*> ExecutedBlocks;
1382 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
1383 /// to represent its body. This allows us to delete the temporary globals
1384 /// when we are done.
1385 std::vector<GlobalVariable*> AllocaTmps;
1387 // CurInst - The current instruction we're evaluating.
1388 BasicBlock::iterator CurInst = F->begin()->begin();
1389 ExecutedBlocks.insert(F->begin());
1391 bool EvaluationSuccessful = false;
1393 // This is the main evaluation loop.
1395 Constant *InstResult = 0;
1397 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
1398 if (SI->isVolatile()) break; // no volatile accesses.
1399 Constant *Ptr = getVal(Values, SI->getOperand(1));
1400 if (!isSimpleEnoughPointerToCommit(Ptr))
1401 // If this is too complex for us to commit, reject it.
1403 Constant *Val = getVal(Values, SI->getOperand(0));
1404 MutatedMemory[Ptr] = Val;
1405 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
1406 InstResult = ConstantExpr::get(BO->getOpcode(),
1407 getVal(Values, BO->getOperand(0)),
1408 getVal(Values, BO->getOperand(1)));
1409 } else if (ShiftInst *SI = dyn_cast<ShiftInst>(CurInst)) {
1410 InstResult = ConstantExpr::get(SI->getOpcode(),
1411 getVal(Values, SI->getOperand(0)),
1412 getVal(Values, SI->getOperand(1)));
1413 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
1414 InstResult = ConstantExpr::getCast(getVal(Values, CI->getOperand(0)),
1416 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
1417 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
1418 getVal(Values, SI->getOperand(1)),
1419 getVal(Values, SI->getOperand(2)));
1420 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
1421 Constant *P = getVal(Values, GEP->getOperand(0));
1422 std::vector<Constant*> GEPOps;
1423 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
1424 GEPOps.push_back(getVal(Values, GEP->getOperand(i)));
1425 InstResult = ConstantExpr::getGetElementPtr(P, GEPOps);
1426 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
1427 if (LI->isVolatile()) break; // no volatile accesses.
1428 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
1430 if (InstResult == 0) break; // Could not evaluate load.
1431 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
1432 if (AI->isArrayAllocation()) break; // Cannot handle array allocs.
1433 const Type *Ty = AI->getType()->getElementType();
1434 AllocaTmps.push_back(new GlobalVariable(Ty, false,
1435 GlobalValue::InternalLinkage,
1436 UndefValue::get(Ty),
1438 InstResult = AllocaTmps.back();
1439 } else if (TerminatorInst *TI = dyn_cast<TerminatorInst>(CurInst)) {
1440 BasicBlock *NewBB = 0;
1441 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
1442 if (BI->isUnconditional()) {
1443 NewBB = BI->getSuccessor(0);
1445 ConstantBool *Cond =
1446 dyn_cast<ConstantBool>(getVal(Values, BI->getCondition()));
1447 if (!Cond) break; // Cannot determine.
1448 NewBB = BI->getSuccessor(!Cond->getValue());
1450 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
1452 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
1453 if (!Val) break; // Cannot determine.
1454 NewBB = SI->getSuccessor(SI->findCaseValue(Val));
1455 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
1456 assert(RI->getNumOperands() == 0);
1457 EvaluationSuccessful = true;
1458 break; // We succeeded at evaluating this ctor!
1460 // unwind, unreachable.
1461 break; // Cannot handle this terminator.
1464 // Okay, we succeeded in evaluating this control flow. See if we have
1465 // executed the new block before. If so, we have a looping or recursive
1466 // function, which we cannot evaluate in reasonable time.
1467 if (!ExecutedBlocks.insert(NewBB).second)
1468 break; // Recursed/looped!
1470 // Okay, we have never been in this block before. Check to see if there
1471 // are any PHI nodes. If so, evaluate them with information about where
1473 BasicBlock *OldBB = CurInst->getParent();
1474 CurInst = NewBB->begin();
1476 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
1477 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
1479 // Do NOT increment CurInst. We know that the terminator had no value.
1482 // TODO: use ConstantFoldCall for function calls.
1484 // Did not know how to evaluate this!
1488 if (!CurInst->use_empty())
1489 Values[CurInst] = InstResult;
1491 // Advance program counter.
1495 if (EvaluationSuccessful) {
1496 // We succeeded at evaluation: commit the result.
1497 DEBUG(std::cerr << "FULLY EVALUATED GLOBAL CTOR FUNCTION '" <<
1498 F->getName() << "'\n");
1499 for (std::map<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
1500 E = MutatedMemory.end(); I != E; ++I)
1501 CommitValueTo(I->second, I->first);
1504 // At this point, we are done interpreting. If we created any 'alloca'
1505 // temporaries, release them now.
1506 while (!AllocaTmps.empty()) {
1507 GlobalVariable *Tmp = AllocaTmps.back();
1508 AllocaTmps.pop_back();
1510 // If there are still users of the alloca, the program is doing something
1511 // silly, e.g. storing the address of the alloca somewhere and using it
1512 // later. Since this is undefined, we'll just make it be null.
1513 if (!Tmp->use_empty())
1514 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
1518 return EvaluationSuccessful;
1522 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
1523 /// Return true if anything changed.
1524 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
1525 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
1526 bool MadeChange = false;
1527 if (Ctors.empty()) return false;
1529 // Loop over global ctors, optimizing them when we can.
1530 for (unsigned i = 0; i != Ctors.size(); ++i) {
1531 Function *F = Ctors[i];
1532 // Found a null terminator in the middle of the list, prune off the rest of
1535 if (i != Ctors.size()-1) {
1542 // We cannot simplify external ctor functions.
1543 if (F->empty()) continue;
1545 // If we can evaluate the ctor at compile time, do.
1546 if (EvaluateStaticConstructor(F)) {
1547 Ctors.erase(Ctors.begin()+i);
1550 ++NumCtorsEvaluated;
1555 if (!MadeChange) return false;
1557 GCL = InstallGlobalCtors(GCL, Ctors);
1562 bool GlobalOpt::runOnModule(Module &M) {
1563 bool Changed = false;
1565 // Try to find the llvm.globalctors list.
1566 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
1568 bool LocalChange = true;
1569 while (LocalChange) {
1570 LocalChange = false;
1572 // Delete functions that are trivially dead, ccc -> fastcc
1573 LocalChange |= OptimizeFunctions(M);
1575 // Optimize global_ctors list.
1577 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
1579 // Optimize non-address-taken globals.
1580 LocalChange |= OptimizeGlobalVars(M);
1581 Changed |= LocalChange;
1584 // TODO: Move all global ctors functions to the end of the module for code