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"
36 Statistic<> NumMarked ("globalopt", "Number of globals marked constant");
37 Statistic<> NumSRA ("globalopt", "Number of aggregate globals broken "
39 Statistic<> NumHeapSRA ("globalopt", "Number of heap objects SRA'd");
40 Statistic<> NumSubstitute("globalopt",
41 "Number of globals with initializers stored into them");
42 Statistic<> NumDeleted ("globalopt", "Number of globals deleted");
43 Statistic<> NumFnDeleted("globalopt", "Number of functions deleted");
44 Statistic<> NumGlobUses ("globalopt", "Number of global uses devirtualized");
45 Statistic<> NumLocalized("globalopt", "Number of globals localized");
46 Statistic<> NumShrunkToBool("globalopt",
47 "Number of global vars shrunk to booleans");
48 Statistic<> NumFastCallFns("globalopt",
49 "Number of functions converted to fastcc");
50 Statistic<> NumCtorsEvaluated("globalopt","Number of static ctors evaluated");
52 struct GlobalOpt : public ModulePass {
53 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
54 AU.addRequired<TargetData>();
57 bool runOnModule(Module &M);
60 GlobalVariable *FindGlobalCtors(Module &M);
61 bool OptimizeFunctions(Module &M);
62 bool OptimizeGlobalVars(Module &M);
63 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
64 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
67 RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer");
70 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
72 /// GlobalStatus - As we analyze each global, keep track of some information
73 /// about it. If we find out that the address of the global is taken, none of
74 /// this info will be accurate.
76 /// isLoaded - True if the global is ever loaded. If the global isn't ever
77 /// loaded it can be deleted.
80 /// StoredType - Keep track of what stores to the global look like.
83 /// NotStored - There is no store to this global. It can thus be marked
87 /// isInitializerStored - This global is stored to, but the only thing
88 /// stored is the constant it was initialized with. This is only tracked
89 /// for scalar globals.
92 /// isStoredOnce - This global is stored to, but only its initializer and
93 /// one other value is ever stored to it. If this global isStoredOnce, we
94 /// track the value stored to it in StoredOnceValue below. This is only
95 /// tracked for scalar globals.
98 /// isStored - This global is stored to by multiple values or something else
99 /// that we cannot track.
103 /// StoredOnceValue - If only one value (besides the initializer constant) is
104 /// ever stored to this global, keep track of what value it is.
105 Value *StoredOnceValue;
107 // AccessingFunction/HasMultipleAccessingFunctions - These start out
108 // null/false. When the first accessing function is noticed, it is recorded.
109 // When a second different accessing function is noticed,
110 // HasMultipleAccessingFunctions is set to true.
111 Function *AccessingFunction;
112 bool HasMultipleAccessingFunctions;
114 // HasNonInstructionUser - Set to true if this global has a user that is not
115 // an instruction (e.g. a constant expr or GV initializer).
116 bool HasNonInstructionUser;
118 /// isNotSuitableForSRA - Keep track of whether any SRA preventing users of
119 /// the global exist. Such users include GEP instruction with variable
120 /// indexes, and non-gep/load/store users like constant expr casts.
121 bool isNotSuitableForSRA;
123 GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0),
124 AccessingFunction(0), HasMultipleAccessingFunctions(false),
125 HasNonInstructionUser(false), isNotSuitableForSRA(false) {}
130 /// ConstantIsDead - Return true if the specified constant is (transitively)
131 /// dead. The constant may be used by other constants (e.g. constant arrays and
132 /// constant exprs) as long as they are dead, but it cannot be used by anything
134 static bool ConstantIsDead(Constant *C) {
135 if (isa<GlobalValue>(C)) return false;
137 for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI)
138 if (Constant *CU = dyn_cast<Constant>(*UI)) {
139 if (!ConstantIsDead(CU)) return false;
146 /// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
147 /// structure. If the global has its address taken, return true to indicate we
148 /// can't do anything with it.
150 static bool AnalyzeGlobal(Value *V, GlobalStatus &GS,
151 std::set<PHINode*> &PHIUsers) {
152 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
153 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
154 GS.HasNonInstructionUser = true;
156 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
157 if (CE->getOpcode() != Instruction::GetElementPtr)
158 GS.isNotSuitableForSRA = true;
159 else if (!GS.isNotSuitableForSRA) {
160 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
161 // don't like < 3 operand CE's, and we don't like non-constant integer
163 if (CE->getNumOperands() < 3 || !CE->getOperand(1)->isNullValue())
164 GS.isNotSuitableForSRA = true;
166 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
167 if (!isa<ConstantInt>(CE->getOperand(i))) {
168 GS.isNotSuitableForSRA = true;
174 } else if (Instruction *I = dyn_cast<Instruction>(*UI)) {
175 if (!GS.HasMultipleAccessingFunctions) {
176 Function *F = I->getParent()->getParent();
177 if (GS.AccessingFunction == 0)
178 GS.AccessingFunction = F;
179 else if (GS.AccessingFunction != F)
180 GS.HasMultipleAccessingFunctions = true;
182 if (isa<LoadInst>(I)) {
184 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
185 // Don't allow a store OF the address, only stores TO the address.
186 if (SI->getOperand(0) == V) return true;
188 // If this is a direct store to the global (i.e., the global is a scalar
189 // value, not an aggregate), keep more specific information about
191 if (GS.StoredType != GlobalStatus::isStored)
192 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){
193 Value *StoredVal = SI->getOperand(0);
194 if (StoredVal == GV->getInitializer()) {
195 if (GS.StoredType < GlobalStatus::isInitializerStored)
196 GS.StoredType = GlobalStatus::isInitializerStored;
197 } else if (isa<LoadInst>(StoredVal) &&
198 cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
200 if (GS.StoredType < GlobalStatus::isInitializerStored)
201 GS.StoredType = GlobalStatus::isInitializerStored;
202 } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
203 GS.StoredType = GlobalStatus::isStoredOnce;
204 GS.StoredOnceValue = StoredVal;
205 } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
206 GS.StoredOnceValue == StoredVal) {
209 GS.StoredType = GlobalStatus::isStored;
212 GS.StoredType = GlobalStatus::isStored;
214 } else if (isa<GetElementPtrInst>(I)) {
215 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
217 // If the first two indices are constants, this can be SRA'd.
218 if (isa<GlobalVariable>(I->getOperand(0))) {
219 if (I->getNumOperands() < 3 || !isa<Constant>(I->getOperand(1)) ||
220 !cast<Constant>(I->getOperand(1))->isNullValue() ||
221 !isa<ConstantInt>(I->getOperand(2)))
222 GS.isNotSuitableForSRA = true;
223 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(I->getOperand(0))){
224 if (CE->getOpcode() != Instruction::GetElementPtr ||
225 CE->getNumOperands() < 3 || I->getNumOperands() < 2 ||
226 !isa<Constant>(I->getOperand(0)) ||
227 !cast<Constant>(I->getOperand(0))->isNullValue())
228 GS.isNotSuitableForSRA = true;
230 GS.isNotSuitableForSRA = true;
232 } else if (isa<SelectInst>(I)) {
233 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
234 GS.isNotSuitableForSRA = true;
235 } else if (PHINode *PN = dyn_cast<PHINode>(I)) {
236 // PHI nodes we can check just like select or GEP instructions, but we
237 // have to be careful about infinite recursion.
238 if (PHIUsers.insert(PN).second) // Not already visited.
239 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
240 GS.isNotSuitableForSRA = true;
241 } else if (isa<SetCondInst>(I)) {
242 GS.isNotSuitableForSRA = true;
243 } else if (isa<MemCpyInst>(I) || isa<MemMoveInst>(I)) {
244 if (I->getOperand(1) == V)
245 GS.StoredType = GlobalStatus::isStored;
246 if (I->getOperand(2) == V)
248 GS.isNotSuitableForSRA = true;
249 } else if (isa<MemSetInst>(I)) {
250 assert(I->getOperand(1) == V && "Memset only takes one pointer!");
251 GS.StoredType = GlobalStatus::isStored;
252 GS.isNotSuitableForSRA = true;
254 return true; // Any other non-load instruction might take address!
256 } else if (Constant *C = dyn_cast<Constant>(*UI)) {
257 GS.HasNonInstructionUser = true;
258 // We might have a dead and dangling constant hanging off of here.
259 if (!ConstantIsDead(C))
262 GS.HasNonInstructionUser = true;
263 // Otherwise must be some other user.
270 static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) {
271 ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
273 unsigned IdxV = (unsigned)CI->getRawValue();
275 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
276 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
277 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
278 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
279 } else if (ConstantPacked *CP = dyn_cast<ConstantPacked>(Agg)) {
280 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
281 } else if (isa<ConstantAggregateZero>(Agg)) {
282 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
283 if (IdxV < STy->getNumElements())
284 return Constant::getNullValue(STy->getElementType(IdxV));
285 } else if (const SequentialType *STy =
286 dyn_cast<SequentialType>(Agg->getType())) {
287 return Constant::getNullValue(STy->getElementType());
289 } else if (isa<UndefValue>(Agg)) {
290 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
291 if (IdxV < STy->getNumElements())
292 return UndefValue::get(STy->getElementType(IdxV));
293 } else if (const SequentialType *STy =
294 dyn_cast<SequentialType>(Agg->getType())) {
295 return UndefValue::get(STy->getElementType());
302 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
303 /// users of the global, cleaning up the obvious ones. This is largely just a
304 /// quick scan over the use list to clean up the easy and obvious cruft. This
305 /// returns true if it made a change.
306 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
307 bool Changed = false;
308 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
311 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
313 // Replace the load with the initializer.
314 LI->replaceAllUsesWith(Init);
315 LI->eraseFromParent();
318 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
319 // Store must be unreachable or storing Init into the global.
320 SI->eraseFromParent();
322 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
323 if (CE->getOpcode() == Instruction::GetElementPtr) {
324 Constant *SubInit = 0;
326 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
327 Changed |= CleanupConstantGlobalUsers(CE, SubInit);
328 } else if (CE->getOpcode() == Instruction::Cast &&
329 isa<PointerType>(CE->getType())) {
330 // Pointer cast, delete any stores and memsets to the global.
331 Changed |= CleanupConstantGlobalUsers(CE, 0);
334 if (CE->use_empty()) {
335 CE->destroyConstant();
338 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
339 Constant *SubInit = 0;
341 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
342 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
343 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
344 Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
346 if (GEP->use_empty()) {
347 GEP->eraseFromParent();
350 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
351 if (MI->getRawDest() == V) {
352 MI->eraseFromParent();
356 } else if (Constant *C = dyn_cast<Constant>(U)) {
357 // If we have a chain of dead constantexprs or other things dangling from
358 // us, and if they are all dead, nuke them without remorse.
359 if (ConstantIsDead(C)) {
360 C->destroyConstant();
361 // This could have invalidated UI, start over from scratch.
362 CleanupConstantGlobalUsers(V, Init);
370 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
371 /// variable. This opens the door for other optimizations by exposing the
372 /// behavior of the program in a more fine-grained way. We have determined that
373 /// this transformation is safe already. We return the first global variable we
374 /// insert so that the caller can reprocess it.
375 static GlobalVariable *SRAGlobal(GlobalVariable *GV) {
376 assert(GV->hasInternalLinkage() && !GV->isConstant());
377 Constant *Init = GV->getInitializer();
378 const Type *Ty = Init->getType();
380 std::vector<GlobalVariable*> NewGlobals;
381 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
383 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
384 NewGlobals.reserve(STy->getNumElements());
385 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
386 Constant *In = getAggregateConstantElement(Init,
387 ConstantUInt::get(Type::UIntTy, i));
388 assert(In && "Couldn't get element of initializer?");
389 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
390 GlobalVariable::InternalLinkage,
391 In, GV->getName()+"."+utostr(i));
392 Globals.insert(GV, NGV);
393 NewGlobals.push_back(NGV);
395 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
396 unsigned NumElements = 0;
397 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
398 NumElements = ATy->getNumElements();
399 else if (const PackedType *PTy = dyn_cast<PackedType>(STy))
400 NumElements = PTy->getNumElements();
402 assert(0 && "Unknown aggregate sequential type!");
404 if (NumElements > 16 && GV->hasNUsesOrMore(16))
405 return 0; // It's not worth it.
406 NewGlobals.reserve(NumElements);
407 for (unsigned i = 0, e = NumElements; i != e; ++i) {
408 Constant *In = getAggregateConstantElement(Init,
409 ConstantUInt::get(Type::UIntTy, i));
410 assert(In && "Couldn't get element of initializer?");
412 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
413 GlobalVariable::InternalLinkage,
414 In, GV->getName()+"."+utostr(i));
415 Globals.insert(GV, NGV);
416 NewGlobals.push_back(NGV);
420 if (NewGlobals.empty())
423 DEBUG(std::cerr << "PERFORMING GLOBAL SRA ON: " << *GV);
425 Constant *NullInt = Constant::getNullValue(Type::IntTy);
427 // Loop over all of the uses of the global, replacing the constantexpr geps,
428 // with smaller constantexpr geps or direct references.
429 while (!GV->use_empty()) {
430 User *GEP = GV->use_back();
431 assert(((isa<ConstantExpr>(GEP) &&
432 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
433 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
435 // Ignore the 1th operand, which has to be zero or else the program is quite
436 // broken (undefined). Get the 2nd operand, which is the structure or array
439 (unsigned)cast<ConstantInt>(GEP->getOperand(2))->getRawValue();
440 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
442 Value *NewPtr = NewGlobals[Val];
444 // Form a shorter GEP if needed.
445 if (GEP->getNumOperands() > 3)
446 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
447 std::vector<Constant*> Idxs;
448 Idxs.push_back(NullInt);
449 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
450 Idxs.push_back(CE->getOperand(i));
451 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr), Idxs);
453 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
454 std::vector<Value*> Idxs;
455 Idxs.push_back(NullInt);
456 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
457 Idxs.push_back(GEPI->getOperand(i));
458 NewPtr = new GetElementPtrInst(NewPtr, Idxs,
459 GEPI->getName()+"."+utostr(Val), GEPI);
461 GEP->replaceAllUsesWith(NewPtr);
463 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
464 GEPI->eraseFromParent();
466 cast<ConstantExpr>(GEP)->destroyConstant();
469 // Delete the old global, now that it is dead.
473 // Loop over the new globals array deleting any globals that are obviously
474 // dead. This can arise due to scalarization of a structure or an array that
475 // has elements that are dead.
476 unsigned FirstGlobal = 0;
477 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
478 if (NewGlobals[i]->use_empty()) {
479 Globals.erase(NewGlobals[i]);
480 if (FirstGlobal == i) ++FirstGlobal;
483 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
486 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
487 /// value will trap if the value is dynamically null.
488 static bool AllUsesOfValueWillTrapIfNull(Value *V) {
489 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
490 if (isa<LoadInst>(*UI)) {
492 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
493 if (SI->getOperand(0) == V) {
494 //std::cerr << "NONTRAPPING USE: " << **UI;
495 return false; // Storing the value.
497 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
498 if (CI->getOperand(0) != V) {
499 //std::cerr << "NONTRAPPING USE: " << **UI;
500 return false; // Not calling the ptr
502 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
503 if (II->getOperand(0) != V) {
504 //std::cerr << "NONTRAPPING USE: " << **UI;
505 return false; // Not calling the ptr
507 } else if (CastInst *CI = dyn_cast<CastInst>(*UI)) {
508 if (!AllUsesOfValueWillTrapIfNull(CI)) return false;
509 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
510 if (!AllUsesOfValueWillTrapIfNull(GEPI)) return false;
511 } else if (isa<SetCondInst>(*UI) &&
512 isa<ConstantPointerNull>(UI->getOperand(1))) {
513 // Ignore setcc X, null
515 //std::cerr << "NONTRAPPING USE: " << **UI;
521 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
522 /// from GV will trap if the loaded value is null. Note that this also permits
523 /// comparisons of the loaded value against null, as a special case.
524 static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) {
525 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI)
526 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
527 if (!AllUsesOfValueWillTrapIfNull(LI))
529 } else if (isa<StoreInst>(*UI)) {
530 // Ignore stores to the global.
532 // We don't know or understand this user, bail out.
533 //std::cerr << "UNKNOWN USER OF GLOBAL!: " << **UI;
540 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
541 bool Changed = false;
542 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
543 Instruction *I = cast<Instruction>(*UI++);
544 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
545 LI->setOperand(0, NewV);
547 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
548 if (SI->getOperand(1) == V) {
549 SI->setOperand(1, NewV);
552 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
553 if (I->getOperand(0) == V) {
554 // Calling through the pointer! Turn into a direct call, but be careful
555 // that the pointer is not also being passed as an argument.
556 I->setOperand(0, NewV);
558 bool PassedAsArg = false;
559 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
560 if (I->getOperand(i) == V) {
562 I->setOperand(i, NewV);
566 // Being passed as an argument also. Be careful to not invalidate UI!
570 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
571 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
572 ConstantExpr::getCast(NewV, CI->getType()));
573 if (CI->use_empty()) {
575 CI->eraseFromParent();
577 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
578 // Should handle GEP here.
579 std::vector<Constant*> Indices;
580 Indices.reserve(GEPI->getNumOperands()-1);
581 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
582 if (Constant *C = dyn_cast<Constant>(GEPI->getOperand(i)))
583 Indices.push_back(C);
586 if (Indices.size() == GEPI->getNumOperands()-1)
587 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
588 ConstantExpr::getGetElementPtr(NewV, Indices));
589 if (GEPI->use_empty()) {
591 GEPI->eraseFromParent();
600 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
601 /// value stored into it. If there are uses of the loaded value that would trap
602 /// if the loaded value is dynamically null, then we know that they cannot be
603 /// reachable with a null optimize away the load.
604 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) {
605 std::vector<LoadInst*> Loads;
606 bool Changed = false;
608 // Replace all uses of loads with uses of uses of the stored value.
609 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end();
611 if (LoadInst *LI = dyn_cast<LoadInst>(*GUI)) {
613 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
615 assert(isa<StoreInst>(*GUI) && "Only expect load and stores!");
619 DEBUG(std::cerr << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV);
623 // Delete all of the loads we can, keeping track of whether we nuked them all!
624 bool AllLoadsGone = true;
625 while (!Loads.empty()) {
626 LoadInst *L = Loads.back();
627 if (L->use_empty()) {
628 L->eraseFromParent();
631 AllLoadsGone = false;
636 // If we nuked all of the loads, then none of the stores are needed either,
637 // nor is the global.
639 DEBUG(std::cerr << " *** GLOBAL NOW DEAD!\n");
640 CleanupConstantGlobalUsers(GV, 0);
641 if (GV->use_empty()) {
642 GV->eraseFromParent();
650 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
651 /// instructions that are foldable.
652 static void ConstantPropUsersOf(Value *V) {
653 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
654 if (Instruction *I = dyn_cast<Instruction>(*UI++))
655 if (Constant *NewC = ConstantFoldInstruction(I)) {
656 I->replaceAllUsesWith(NewC);
658 // Advance UI to the next non-I use to avoid invalidating it!
659 // Instructions could multiply use V.
660 while (UI != E && *UI == I)
662 I->eraseFromParent();
666 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
667 /// variable, and transforms the program as if it always contained the result of
668 /// the specified malloc. Because it is always the result of the specified
669 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
670 /// malloc into a global, and any laods of GV as uses of the new global.
671 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
673 DEBUG(std::cerr << "PROMOTING MALLOC GLOBAL: " << *GV << " MALLOC = " <<*MI);
674 ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize());
676 if (NElements->getRawValue() != 1) {
677 // If we have an array allocation, transform it to a single element
678 // allocation to make the code below simpler.
679 Type *NewTy = ArrayType::get(MI->getAllocatedType(),
680 (unsigned)NElements->getRawValue());
682 new MallocInst(NewTy, Constant::getNullValue(Type::UIntTy),
683 MI->getAlignment(), MI->getName(), MI);
684 std::vector<Value*> Indices;
685 Indices.push_back(Constant::getNullValue(Type::IntTy));
686 Indices.push_back(Indices[0]);
687 Value *NewGEP = new GetElementPtrInst(NewMI, Indices,
688 NewMI->getName()+".el0", MI);
689 MI->replaceAllUsesWith(NewGEP);
690 MI->eraseFromParent();
694 // Create the new global variable. The contents of the malloc'd memory is
695 // undefined, so initialize with an undef value.
696 Constant *Init = UndefValue::get(MI->getAllocatedType());
697 GlobalVariable *NewGV = new GlobalVariable(MI->getAllocatedType(), false,
698 GlobalValue::InternalLinkage, Init,
699 GV->getName()+".body");
700 GV->getParent()->getGlobalList().insert(GV, NewGV);
702 // Anything that used the malloc now uses the global directly.
703 MI->replaceAllUsesWith(NewGV);
705 Constant *RepValue = NewGV;
706 if (NewGV->getType() != GV->getType()->getElementType())
707 RepValue = ConstantExpr::getCast(RepValue, GV->getType()->getElementType());
709 // If there is a comparison against null, we will insert a global bool to
710 // keep track of whether the global was initialized yet or not.
711 GlobalVariable *InitBool =
712 new GlobalVariable(Type::BoolTy, false, GlobalValue::InternalLinkage,
713 ConstantBool::getFalse(), GV->getName()+".init");
714 bool InitBoolUsed = false;
716 // Loop over all uses of GV, processing them in turn.
717 std::vector<StoreInst*> Stores;
718 while (!GV->use_empty())
719 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
720 while (!LI->use_empty()) {
721 Use &LoadUse = LI->use_begin().getUse();
722 if (!isa<SetCondInst>(LoadUse.getUser()))
725 // Replace the setcc X, 0 with a use of the bool value.
726 SetCondInst *SCI = cast<SetCondInst>(LoadUse.getUser());
727 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", SCI);
729 switch (SCI->getOpcode()) {
730 default: assert(0 && "Unknown opcode!");
731 case Instruction::SetLT:
732 LV = ConstantBool::getFalse(); // X < null -> always false
734 case Instruction::SetEQ:
735 case Instruction::SetLE:
736 LV = BinaryOperator::createNot(LV, "notinit", SCI);
738 case Instruction::SetNE:
739 case Instruction::SetGE:
740 case Instruction::SetGT:
743 SCI->replaceAllUsesWith(LV);
744 SCI->eraseFromParent();
747 LI->eraseFromParent();
749 StoreInst *SI = cast<StoreInst>(GV->use_back());
750 // The global is initialized when the store to it occurs.
751 new StoreInst(ConstantBool::getTrue(), InitBool, SI);
752 SI->eraseFromParent();
755 // If the initialization boolean was used, insert it, otherwise delete it.
757 while (!InitBool->use_empty()) // Delete initializations
758 cast<Instruction>(InitBool->use_back())->eraseFromParent();
761 GV->getParent()->getGlobalList().insert(GV, InitBool);
764 // Now the GV is dead, nuke it and the malloc.
765 GV->eraseFromParent();
766 MI->eraseFromParent();
768 // To further other optimizations, loop over all users of NewGV and try to
769 // constant prop them. This will promote GEP instructions with constant
770 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
771 ConstantPropUsersOf(NewGV);
772 if (RepValue != NewGV)
773 ConstantPropUsersOf(RepValue);
778 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
779 /// to make sure that there are no complex uses of V. We permit simple things
780 /// like dereferencing the pointer, but not storing through the address, unless
781 /// it is to the specified global.
782 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V,
783 GlobalVariable *GV) {
784 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI)
785 if (isa<LoadInst>(*UI) || isa<SetCondInst>(*UI)) {
787 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
788 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
789 return false; // Storing the pointer itself... bad.
790 // Otherwise, storing through it, or storing into GV... fine.
791 } else if (isa<GetElementPtrInst>(*UI) || isa<SelectInst>(*UI)) {
792 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(cast<Instruction>(*UI),GV))
800 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
801 /// somewhere. Transform all uses of the allocation into loads from the
802 /// global and uses of the resultant pointer. Further, delete the store into
803 /// GV. This assumes that these value pass the
804 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
805 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
806 GlobalVariable *GV) {
807 while (!Alloc->use_empty()) {
808 Instruction *U = Alloc->use_back();
809 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
810 // If this is the store of the allocation into the global, remove it.
811 if (SI->getOperand(1) == GV) {
812 SI->eraseFromParent();
817 // Insert a load from the global, and use it instead of the malloc.
818 Value *NL = new LoadInst(GV, GV->getName()+".val", U);
819 U->replaceUsesOfWith(Alloc, NL);
823 /// GlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
824 /// GV are simple enough to perform HeapSRA, return true.
825 static bool GlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV) {
826 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;
828 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
829 // We permit two users of the load: setcc comparing against the null
830 // pointer, and a getelementptr of a specific form.
831 for (Value::use_iterator UI = LI->use_begin(), E = LI->use_end(); UI != E;
833 // Comparison against null is ok.
834 if (SetCondInst *SCI = dyn_cast<SetCondInst>(*UI)) {
835 if (!isa<ConstantPointerNull>(SCI->getOperand(1)))
840 // getelementptr is also ok, but only a simple form.
841 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI);
842 if (!GEPI) return false;
844 // Must index into the array and into the struct.
845 if (GEPI->getNumOperands() < 3)
848 // Otherwise the GEP is ok.
855 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
856 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
857 /// use FieldGlobals instead. All uses of loaded values satisfy
858 /// GlobalLoadUsesSimpleEnoughForHeapSRA.
859 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Ptr,
860 const std::vector<GlobalVariable*> &FieldGlobals) {
861 std::vector<Value *> InsertedLoadsForPtr;
862 //InsertedLoadsForPtr.resize(FieldGlobals.size());
863 while (!Ptr->use_empty()) {
864 Instruction *User = Ptr->use_back();
866 // If this is a comparison against null, handle it.
867 if (SetCondInst *SCI = dyn_cast<SetCondInst>(User)) {
868 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
869 // If we have a setcc of the loaded pointer, we can use a setcc of any
872 if (InsertedLoadsForPtr.empty()) {
873 NPtr = new LoadInst(FieldGlobals[0], Ptr->getName()+".f0", Ptr);
874 InsertedLoadsForPtr.push_back(Ptr);
876 NPtr = InsertedLoadsForPtr.back();
879 Value *New = new SetCondInst(SCI->getOpcode(), NPtr,
880 Constant::getNullValue(NPtr->getType()),
881 SCI->getName(), SCI);
882 SCI->replaceAllUsesWith(New);
883 SCI->eraseFromParent();
887 // Otherwise, this should be: 'getelementptr Ptr, Idx, uint FieldNo ...'
888 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
889 assert(GEPI->getNumOperands() >= 3 && isa<ConstantUInt>(GEPI->getOperand(2))
890 && "Unexpected GEPI!");
892 // Load the pointer for this field.
893 unsigned FieldNo = cast<ConstantUInt>(GEPI->getOperand(2))->getValue();
894 if (InsertedLoadsForPtr.size() <= FieldNo)
895 InsertedLoadsForPtr.resize(FieldNo+1);
896 if (InsertedLoadsForPtr[FieldNo] == 0)
897 InsertedLoadsForPtr[FieldNo] = new LoadInst(FieldGlobals[FieldNo],
898 Ptr->getName()+".f" +
899 utostr(FieldNo), Ptr);
900 Value *NewPtr = InsertedLoadsForPtr[FieldNo];
902 // Create the new GEP idx vector.
903 std::vector<Value*> GEPIdx;
904 GEPIdx.push_back(GEPI->getOperand(1));
905 GEPIdx.insert(GEPIdx.end(), GEPI->op_begin()+3, GEPI->op_end());
907 Value *NGEPI = new GetElementPtrInst(NewPtr, GEPIdx, GEPI->getName(), GEPI);
908 GEPI->replaceAllUsesWith(NGEPI);
909 GEPI->eraseFromParent();
913 /// PerformHeapAllocSRoA - MI is an allocation of an array of structures. Break
914 /// it up into multiple allocations of arrays of the fields.
915 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){
916 DEBUG(std::cerr << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *MI);
917 const StructType *STy = cast<StructType>(MI->getAllocatedType());
919 // There is guaranteed to be at least one use of the malloc (storing
920 // it into GV). If there are other uses, change them to be uses of
921 // the global to simplify later code. This also deletes the store
923 ReplaceUsesOfMallocWithGlobal(MI, GV);
925 // Okay, at this point, there are no users of the malloc. Insert N
926 // new mallocs at the same place as MI, and N globals.
927 std::vector<GlobalVariable*> FieldGlobals;
928 std::vector<MallocInst*> FieldMallocs;
930 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
931 const Type *FieldTy = STy->getElementType(FieldNo);
932 const Type *PFieldTy = PointerType::get(FieldTy);
934 GlobalVariable *NGV =
935 new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage,
936 Constant::getNullValue(PFieldTy),
937 GV->getName() + ".f" + utostr(FieldNo), GV);
938 FieldGlobals.push_back(NGV);
940 MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(),
941 MI->getName() + ".f" + utostr(FieldNo),MI);
942 FieldMallocs.push_back(NMI);
943 new StoreInst(NMI, NGV, MI);
946 // The tricky aspect of this transformation is handling the case when malloc
947 // fails. In the original code, malloc failing would set the result pointer
948 // of malloc to null. In this case, some mallocs could succeed and others
949 // could fail. As such, we emit code that looks like this:
950 // F0 = malloc(field0)
951 // F1 = malloc(field1)
952 // F2 = malloc(field2)
953 // if (F0 == 0 || F1 == 0 || F2 == 0) {
954 // if (F0) { free(F0); F0 = 0; }
955 // if (F1) { free(F1); F1 = 0; }
956 // if (F2) { free(F2); F2 = 0; }
958 Value *RunningOr = 0;
959 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
960 Value *Cond = new SetCondInst(Instruction::SetEQ, FieldMallocs[i],
961 Constant::getNullValue(FieldMallocs[i]->getType()),
964 RunningOr = Cond; // First seteq
966 RunningOr = BinaryOperator::createOr(RunningOr, Cond, "tmp", MI);
969 // Split the basic block at the old malloc.
970 BasicBlock *OrigBB = MI->getParent();
971 BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont");
973 // Create the block to check the first condition. Put all these blocks at the
974 // end of the function as they are unlikely to be executed.
975 BasicBlock *NullPtrBlock = new BasicBlock("malloc_ret_null",
976 OrigBB->getParent());
978 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
979 // branch on RunningOr.
980 OrigBB->getTerminator()->eraseFromParent();
981 new BranchInst(NullPtrBlock, ContBB, RunningOr, OrigBB);
983 // Within the NullPtrBlock, we need to emit a comparison and branch for each
984 // pointer, because some may be null while others are not.
985 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
986 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
987 Value *Cmp = new SetCondInst(Instruction::SetNE, GVVal,
988 Constant::getNullValue(GVVal->getType()),
989 "tmp", NullPtrBlock);
990 BasicBlock *FreeBlock = new BasicBlock("free_it", OrigBB->getParent());
991 BasicBlock *NextBlock = new BasicBlock("next", OrigBB->getParent());
992 new BranchInst(FreeBlock, NextBlock, Cmp, NullPtrBlock);
994 // Fill in FreeBlock.
995 new FreeInst(GVVal, FreeBlock);
996 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
998 new BranchInst(NextBlock, FreeBlock);
1000 NullPtrBlock = NextBlock;
1003 new BranchInst(ContBB, NullPtrBlock);
1006 // MI is no longer needed, remove it.
1007 MI->eraseFromParent();
1010 // Okay, the malloc site is completely handled. All of the uses of GV are now
1011 // loads, and all uses of those loads are simple. Rewrite them to use loads
1012 // of the per-field globals instead.
1013 while (!GV->use_empty()) {
1014 LoadInst *LI = cast<LoadInst>(GV->use_back());
1015 RewriteUsesOfLoadForHeapSRoA(LI, FieldGlobals);
1016 LI->eraseFromParent();
1019 // The old global is now dead, remove it.
1020 GV->eraseFromParent();
1023 return FieldGlobals[0];
1027 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1028 // that only one value (besides its initializer) is ever stored to the global.
1029 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1030 Module::global_iterator &GVI,
1032 if (CastInst *CI = dyn_cast<CastInst>(StoredOnceVal))
1033 StoredOnceVal = CI->getOperand(0);
1034 else if (GetElementPtrInst *GEPI =dyn_cast<GetElementPtrInst>(StoredOnceVal)){
1035 // "getelementptr Ptr, 0, 0, 0" is really just a cast.
1036 bool IsJustACast = true;
1037 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
1038 if (!isa<Constant>(GEPI->getOperand(i)) ||
1039 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
1040 IsJustACast = false;
1044 StoredOnceVal = GEPI->getOperand(0);
1047 // If we are dealing with a pointer global that is initialized to null and
1048 // only has one (non-null) value stored into it, then we can optimize any
1049 // users of the loaded value (often calls and loads) that would trap if the
1051 if (isa<PointerType>(GV->getInitializer()->getType()) &&
1052 GV->getInitializer()->isNullValue()) {
1053 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1054 if (GV->getInitializer()->getType() != SOVC->getType())
1055 SOVC = ConstantExpr::getCast(SOVC, GV->getInitializer()->getType());
1057 // Optimize away any trapping uses of the loaded value.
1058 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1060 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) {
1061 // If this is a malloc of an abstract type, don't touch it.
1062 if (!MI->getAllocatedType()->isSized())
1065 // We can't optimize this global unless all uses of it are *known* to be
1066 // of the malloc value, not of the null initializer value (consider a use
1067 // that compares the global's value against zero to see if the malloc has
1068 // been reached). To do this, we check to see if all uses of the global
1069 // would trap if the global were null: this proves that they must all
1070 // happen after the malloc.
1071 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1074 // We can't optimize this if the malloc itself is used in a complex way,
1075 // for example, being stored into multiple globals. This allows the
1076 // malloc to be stored into the specified global, loaded setcc'd, and
1077 // GEP'd. These are all things we could transform to using the global
1079 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV))
1083 // If we have a global that is only initialized with a fixed size malloc,
1084 // transform the program to use global memory instead of malloc'd memory.
1085 // This eliminates dynamic allocation, avoids an indirection accessing the
1086 // data, and exposes the resultant global to further GlobalOpt.
1087 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) {
1088 // Restrict this transformation to only working on small allocations
1089 // (2048 bytes currently), as we don't want to introduce a 16M global or
1091 if (NElements->getRawValue()*
1092 TD.getTypeSize(MI->getAllocatedType()) < 2048) {
1093 GVI = OptimizeGlobalAddressOfMalloc(GV, MI);
1098 // If the allocation is an array of structures, consider transforming this
1099 // into multiple malloc'd arrays, one for each field. This is basically
1100 // SRoA for malloc'd memory.
1101 if (const StructType *AllocTy =
1102 dyn_cast<StructType>(MI->getAllocatedType())) {
1103 // This the structure has an unreasonable number of fields, leave it
1105 if (AllocTy->getNumElements() <= 16 && AllocTy->getNumElements() > 0 &&
1106 GlobalLoadUsesSimpleEnoughForHeapSRA(GV)) {
1107 GVI = PerformHeapAllocSRoA(GV, MI);
1117 /// ShrinkGlobalToBoolean - At this point, we have learned that the only two
1118 /// values ever stored into GV are its initializer and OtherVal.
1119 static void ShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1120 // Create the new global, initializing it to false.
1121 GlobalVariable *NewGV = new GlobalVariable(Type::BoolTy, false,
1122 GlobalValue::InternalLinkage, ConstantBool::getFalse(),
1123 GV->getName()+".b");
1124 GV->getParent()->getGlobalList().insert(GV, NewGV);
1126 Constant *InitVal = GV->getInitializer();
1127 assert(InitVal->getType() != Type::BoolTy && "No reason to shrink to bool!");
1129 // If initialized to zero and storing one into the global, we can use a cast
1130 // instead of a select to synthesize the desired value.
1131 bool IsOneZero = false;
1132 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1133 IsOneZero = InitVal->isNullValue() && CI->equalsInt(1);
1135 while (!GV->use_empty()) {
1136 Instruction *UI = cast<Instruction>(GV->use_back());
1137 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1138 // Change the store into a boolean store.
1139 bool StoringOther = SI->getOperand(0) == OtherVal;
1140 // Only do this if we weren't storing a loaded value.
1142 if (StoringOther || SI->getOperand(0) == InitVal)
1143 StoreVal = ConstantBool::get(StoringOther);
1145 // Otherwise, we are storing a previously loaded copy. To do this,
1146 // change the copy from copying the original value to just copying the
1148 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1150 // If we're already replaced the input, StoredVal will be a cast or
1151 // select instruction. If not, it will be a load of the original
1153 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1154 assert(LI->getOperand(0) == GV && "Not a copy!");
1155 // Insert a new load, to preserve the saved value.
1156 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1158 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1159 "This is not a form that we understand!");
1160 StoreVal = StoredVal->getOperand(0);
1161 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1164 new StoreInst(StoreVal, NewGV, SI);
1165 } else if (!UI->use_empty()) {
1166 // Change the load into a load of bool then a select.
1167 LoadInst *LI = cast<LoadInst>(UI);
1169 std::string Name = LI->getName(); LI->setName("");
1170 LoadInst *NLI = new LoadInst(NewGV, Name+".b", LI);
1173 NSI = new CastInst(NLI, LI->getType(), Name, LI);
1175 NSI = new SelectInst(NLI, OtherVal, InitVal, Name, LI);
1176 LI->replaceAllUsesWith(NSI);
1178 UI->eraseFromParent();
1181 GV->eraseFromParent();
1185 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1186 /// it if possible. If we make a change, return true.
1187 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1188 Module::global_iterator &GVI) {
1189 std::set<PHINode*> PHIUsers;
1191 GV->removeDeadConstantUsers();
1193 if (GV->use_empty()) {
1194 DEBUG(std::cerr << "GLOBAL DEAD: " << *GV);
1195 GV->eraseFromParent();
1200 if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
1202 std::cerr << "Global: " << *GV;
1203 std::cerr << " isLoaded = " << GS.isLoaded << "\n";
1204 std::cerr << " StoredType = ";
1205 switch (GS.StoredType) {
1206 case GlobalStatus::NotStored: std::cerr << "NEVER STORED\n"; break;
1207 case GlobalStatus::isInitializerStored: std::cerr << "INIT STORED\n"; break;
1208 case GlobalStatus::isStoredOnce: std::cerr << "STORED ONCE\n"; break;
1209 case GlobalStatus::isStored: std::cerr << "stored\n"; break;
1211 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
1212 std::cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n";
1213 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
1214 std::cerr << " AccessingFunction = " << GS.AccessingFunction->getName()
1216 std::cerr << " HasMultipleAccessingFunctions = "
1217 << GS.HasMultipleAccessingFunctions << "\n";
1218 std::cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n";
1219 std::cerr << " isNotSuitableForSRA = " << GS.isNotSuitableForSRA << "\n";
1223 // If this is a first class global and has only one accessing function
1224 // and this function is main (which we know is not recursive we can make
1225 // this global a local variable) we replace the global with a local alloca
1226 // in this function.
1228 // NOTE: It doesn't make sense to promote non first class types since we
1229 // are just replacing static memory to stack memory.
1230 if (!GS.HasMultipleAccessingFunctions &&
1231 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1232 GV->getType()->getElementType()->isFirstClassType() &&
1233 GS.AccessingFunction->getName() == "main" &&
1234 GS.AccessingFunction->hasExternalLinkage()) {
1235 DEBUG(std::cerr << "LOCALIZING GLOBAL: " << *GV);
1236 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin();
1237 const Type* ElemTy = GV->getType()->getElementType();
1238 // FIXME: Pass Global's alignment when globals have alignment
1239 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI);
1240 if (!isa<UndefValue>(GV->getInitializer()))
1241 new StoreInst(GV->getInitializer(), Alloca, FirstI);
1243 GV->replaceAllUsesWith(Alloca);
1244 GV->eraseFromParent();
1249 // If the global is never loaded (but may be stored to), it is dead.
1252 DEBUG(std::cerr << "GLOBAL NEVER LOADED: " << *GV);
1254 // Delete any stores we can find to the global. We may not be able to
1255 // make it completely dead though.
1256 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1258 // If the global is dead now, delete it.
1259 if (GV->use_empty()) {
1260 GV->eraseFromParent();
1266 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1267 DEBUG(std::cerr << "MARKING CONSTANT: " << *GV);
1268 GV->setConstant(true);
1270 // Clean up any obviously simplifiable users now.
1271 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1273 // If the global is dead now, just nuke it.
1274 if (GV->use_empty()) {
1275 DEBUG(std::cerr << " *** Marking constant allowed us to simplify "
1276 "all users and delete global!\n");
1277 GV->eraseFromParent();
1283 } else if (!GS.isNotSuitableForSRA &&
1284 !GV->getInitializer()->getType()->isFirstClassType()) {
1285 if (GlobalVariable *FirstNewGV = SRAGlobal(GV)) {
1286 GVI = FirstNewGV; // Don't skip the newly produced globals!
1289 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1290 // If the initial value for the global was an undef value, and if only
1291 // one other value was stored into it, we can just change the
1292 // initializer to be an undef value, then delete all stores to the
1293 // global. This allows us to mark it constant.
1294 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1295 if (isa<UndefValue>(GV->getInitializer())) {
1296 // Change the initial value here.
1297 GV->setInitializer(SOVConstant);
1299 // Clean up any obviously simplifiable users now.
1300 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1302 if (GV->use_empty()) {
1303 DEBUG(std::cerr << " *** Substituting initializer allowed us to "
1304 "simplify all users and delete global!\n");
1305 GV->eraseFromParent();
1314 // Try to optimize globals based on the knowledge that only one value
1315 // (besides its initializer) is ever stored to the global.
1316 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1317 getAnalysis<TargetData>()))
1320 // Otherwise, if the global was not a boolean, we can shrink it to be a
1322 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1323 if (GV->getType()->getElementType() != Type::BoolTy &&
1324 !GV->getType()->getElementType()->isFloatingPoint()) {
1325 DEBUG(std::cerr << " *** SHRINKING TO BOOL: " << *GV);
1326 ShrinkGlobalToBoolean(GV, SOVConstant);
1335 /// OnlyCalledDirectly - Return true if the specified function is only called
1336 /// directly. In other words, its address is never taken.
1337 static bool OnlyCalledDirectly(Function *F) {
1338 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1339 Instruction *User = dyn_cast<Instruction>(*UI);
1340 if (!User) return false;
1341 if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false;
1343 // See if the function address is passed as an argument.
1344 for (unsigned i = 1, e = User->getNumOperands(); i != e; ++i)
1345 if (User->getOperand(i) == F) return false;
1350 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1351 /// function, changing them to FastCC.
1352 static void ChangeCalleesToFastCall(Function *F) {
1353 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1354 Instruction *User = cast<Instruction>(*UI);
1355 if (CallInst *CI = dyn_cast<CallInst>(User))
1356 CI->setCallingConv(CallingConv::Fast);
1358 cast<InvokeInst>(User)->setCallingConv(CallingConv::Fast);
1362 bool GlobalOpt::OptimizeFunctions(Module &M) {
1363 bool Changed = false;
1364 // Optimize functions.
1365 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1367 F->removeDeadConstantUsers();
1368 if (F->use_empty() && (F->hasInternalLinkage() ||
1369 F->hasLinkOnceLinkage())) {
1370 M.getFunctionList().erase(F);
1373 } else if (F->hasInternalLinkage() &&
1374 F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1375 OnlyCalledDirectly(F)) {
1376 // If this function has C calling conventions, is not a varargs
1377 // function, and is only called directly, promote it to use the Fast
1378 // calling convention.
1379 F->setCallingConv(CallingConv::Fast);
1380 ChangeCalleesToFastCall(F);
1388 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1389 bool Changed = false;
1390 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1392 GlobalVariable *GV = GVI++;
1393 if (!GV->isConstant() && GV->hasInternalLinkage() &&
1394 GV->hasInitializer())
1395 Changed |= ProcessInternalGlobal(GV, GVI);
1400 /// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1401 /// initializers have an init priority of 65535.
1402 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1403 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1405 if (I->getName() == "llvm.global_ctors") {
1406 // Found it, verify it's an array of { int, void()* }.
1407 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1409 const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1410 if (!STy || STy->getNumElements() != 2 ||
1411 STy->getElementType(0) != Type::IntTy) return 0;
1412 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1413 if (!PFTy) return 0;
1414 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1415 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() ||
1416 FTy->getNumParams() != 0)
1419 // Verify that the initializer is simple enough for us to handle.
1420 if (!I->hasInitializer()) return 0;
1421 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1423 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1424 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(CA->getOperand(i))) {
1425 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1428 // Must have a function or null ptr.
1429 if (!isa<Function>(CS->getOperand(1)))
1432 // Init priority must be standard.
1433 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1434 if (!CI || CI->getRawValue() != 65535)
1445 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1446 /// return a list of the functions and null terminator as a vector.
1447 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1448 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1449 std::vector<Function*> Result;
1450 Result.reserve(CA->getNumOperands());
1451 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) {
1452 ConstantStruct *CS = cast<ConstantStruct>(CA->getOperand(i));
1453 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1458 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1459 /// specified array, returning the new global to use.
1460 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1461 const std::vector<Function*> &Ctors) {
1462 // If we made a change, reassemble the initializer list.
1463 std::vector<Constant*> CSVals;
1464 CSVals.push_back(ConstantSInt::get(Type::IntTy, 65535));
1465 CSVals.push_back(0);
1467 // Create the new init list.
1468 std::vector<Constant*> CAList;
1469 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1471 CSVals[1] = Ctors[i];
1473 const Type *FTy = FunctionType::get(Type::VoidTy,
1474 std::vector<const Type*>(), false);
1475 const PointerType *PFTy = PointerType::get(FTy);
1476 CSVals[1] = Constant::getNullValue(PFTy);
1477 CSVals[0] = ConstantSInt::get(Type::IntTy, 2147483647);
1479 CAList.push_back(ConstantStruct::get(CSVals));
1482 // Create the array initializer.
1483 const Type *StructTy =
1484 cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
1485 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()),
1488 // If we didn't change the number of elements, don't create a new GV.
1489 if (CA->getType() == GCL->getInitializer()->getType()) {
1490 GCL->setInitializer(CA);
1494 // Create the new global and insert it next to the existing list.
1495 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
1496 GCL->getLinkage(), CA,
1499 GCL->getParent()->getGlobalList().insert(GCL, NGV);
1501 // Nuke the old list, replacing any uses with the new one.
1502 if (!GCL->use_empty()) {
1504 if (V->getType() != GCL->getType())
1505 V = ConstantExpr::getCast(V, GCL->getType());
1506 GCL->replaceAllUsesWith(V);
1508 GCL->eraseFromParent();
1517 static Constant *getVal(std::map<Value*, Constant*> &ComputedValues,
1519 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
1520 Constant *R = ComputedValues[V];
1521 assert(R && "Reference to an uncomputed value!");
1525 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
1526 /// enough for us to understand. In particular, if it is a cast of something,
1527 /// we punt. We basically just support direct accesses to globals and GEP's of
1528 /// globals. This should be kept up to date with CommitValueTo.
1529 static bool isSimpleEnoughPointerToCommit(Constant *C) {
1530 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
1531 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1532 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1533 return !GV->isExternal(); // reject external globals.
1535 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
1536 // Handle a constantexpr gep.
1537 if (CE->getOpcode() == Instruction::GetElementPtr &&
1538 isa<GlobalVariable>(CE->getOperand(0))) {
1539 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1540 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1541 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1542 return GV->hasInitializer() &&
1543 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1548 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
1549 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
1550 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
1551 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
1552 ConstantExpr *Addr, unsigned OpNo) {
1553 // Base case of the recursion.
1554 if (OpNo == Addr->getNumOperands()) {
1555 assert(Val->getType() == Init->getType() && "Type mismatch!");
1559 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
1560 std::vector<Constant*> Elts;
1562 // Break up the constant into its elements.
1563 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
1564 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i)
1565 Elts.push_back(CS->getOperand(i));
1566 } else if (isa<ConstantAggregateZero>(Init)) {
1567 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1568 Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
1569 } else if (isa<UndefValue>(Init)) {
1570 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1571 Elts.push_back(UndefValue::get(STy->getElementType(i)));
1573 assert(0 && "This code is out of sync with "
1574 " ConstantFoldLoadThroughGEPConstantExpr");
1577 // Replace the element that we are supposed to.
1578 ConstantUInt *CU = cast<ConstantUInt>(Addr->getOperand(OpNo));
1579 assert(CU->getValue() < STy->getNumElements() &&
1580 "Struct index out of range!");
1581 unsigned Idx = (unsigned)CU->getValue();
1582 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
1584 // Return the modified struct.
1585 return ConstantStruct::get(Elts);
1587 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
1588 const ArrayType *ATy = cast<ArrayType>(Init->getType());
1590 // Break up the array into elements.
1591 std::vector<Constant*> Elts;
1592 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
1593 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1594 Elts.push_back(CA->getOperand(i));
1595 } else if (isa<ConstantAggregateZero>(Init)) {
1596 Constant *Elt = Constant::getNullValue(ATy->getElementType());
1597 Elts.assign(ATy->getNumElements(), Elt);
1598 } else if (isa<UndefValue>(Init)) {
1599 Constant *Elt = UndefValue::get(ATy->getElementType());
1600 Elts.assign(ATy->getNumElements(), Elt);
1602 assert(0 && "This code is out of sync with "
1603 " ConstantFoldLoadThroughGEPConstantExpr");
1606 assert((uint64_t)CI->getRawValue() < ATy->getNumElements());
1607 Elts[(uint64_t)CI->getRawValue()] =
1608 EvaluateStoreInto(Elts[(uint64_t)CI->getRawValue()], Val, Addr, OpNo+1);
1609 return ConstantArray::get(ATy, Elts);
1613 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
1614 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
1615 static void CommitValueTo(Constant *Val, Constant *Addr) {
1616 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
1617 assert(GV->hasInitializer());
1618 GV->setInitializer(Val);
1622 ConstantExpr *CE = cast<ConstantExpr>(Addr);
1623 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1625 Constant *Init = GV->getInitializer();
1626 Init = EvaluateStoreInto(Init, Val, CE, 2);
1627 GV->setInitializer(Init);
1630 /// ComputeLoadResult - Return the value that would be computed by a load from
1631 /// P after the stores reflected by 'memory' have been performed. If we can't
1632 /// decide, return null.
1633 static Constant *ComputeLoadResult(Constant *P,
1634 const std::map<Constant*, Constant*> &Memory) {
1635 // If this memory location has been recently stored, use the stored value: it
1636 // is the most up-to-date.
1637 std::map<Constant*, Constant*>::const_iterator I = Memory.find(P);
1638 if (I != Memory.end()) return I->second;
1641 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
1642 if (GV->hasInitializer())
1643 return GV->getInitializer();
1647 // Handle a constantexpr getelementptr.
1648 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
1649 if (CE->getOpcode() == Instruction::GetElementPtr &&
1650 isa<GlobalVariable>(CE->getOperand(0))) {
1651 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1652 if (GV->hasInitializer())
1653 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1656 return 0; // don't know how to evaluate.
1659 /// EvaluateFunction - Evaluate a call to function F, returning true if
1660 /// successful, false if we can't evaluate it. ActualArgs contains the formal
1661 /// arguments for the function.
1662 static bool EvaluateFunction(Function *F, Constant *&RetVal,
1663 const std::vector<Constant*> &ActualArgs,
1664 std::vector<Function*> &CallStack,
1665 std::map<Constant*, Constant*> &MutatedMemory,
1666 std::vector<GlobalVariable*> &AllocaTmps) {
1667 // Check to see if this function is already executing (recursion). If so,
1668 // bail out. TODO: we might want to accept limited recursion.
1669 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
1672 CallStack.push_back(F);
1674 /// Values - As we compute SSA register values, we store their contents here.
1675 std::map<Value*, Constant*> Values;
1677 // Initialize arguments to the incoming values specified.
1679 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
1681 Values[AI] = ActualArgs[ArgNo];
1683 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
1684 /// we can only evaluate any one basic block at most once. This set keeps
1685 /// track of what we have executed so we can detect recursive cases etc.
1686 std::set<BasicBlock*> ExecutedBlocks;
1688 // CurInst - The current instruction we're evaluating.
1689 BasicBlock::iterator CurInst = F->begin()->begin();
1691 // This is the main evaluation loop.
1693 Constant *InstResult = 0;
1695 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
1696 if (SI->isVolatile()) return false; // no volatile accesses.
1697 Constant *Ptr = getVal(Values, SI->getOperand(1));
1698 if (!isSimpleEnoughPointerToCommit(Ptr))
1699 // If this is too complex for us to commit, reject it.
1701 Constant *Val = getVal(Values, SI->getOperand(0));
1702 MutatedMemory[Ptr] = Val;
1703 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
1704 InstResult = ConstantExpr::get(BO->getOpcode(),
1705 getVal(Values, BO->getOperand(0)),
1706 getVal(Values, BO->getOperand(1)));
1707 } else if (ShiftInst *SI = dyn_cast<ShiftInst>(CurInst)) {
1708 InstResult = ConstantExpr::get(SI->getOpcode(),
1709 getVal(Values, SI->getOperand(0)),
1710 getVal(Values, SI->getOperand(1)));
1711 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
1712 InstResult = ConstantExpr::getCast(getVal(Values, CI->getOperand(0)),
1714 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
1715 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
1716 getVal(Values, SI->getOperand(1)),
1717 getVal(Values, SI->getOperand(2)));
1718 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
1719 Constant *P = getVal(Values, GEP->getOperand(0));
1720 std::vector<Constant*> GEPOps;
1721 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
1722 GEPOps.push_back(getVal(Values, GEP->getOperand(i)));
1723 InstResult = ConstantExpr::getGetElementPtr(P, GEPOps);
1724 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
1725 if (LI->isVolatile()) return false; // no volatile accesses.
1726 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
1728 if (InstResult == 0) return false; // Could not evaluate load.
1729 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
1730 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs.
1731 const Type *Ty = AI->getType()->getElementType();
1732 AllocaTmps.push_back(new GlobalVariable(Ty, false,
1733 GlobalValue::InternalLinkage,
1734 UndefValue::get(Ty),
1736 InstResult = AllocaTmps.back();
1737 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
1738 // Cannot handle inline asm.
1739 if (isa<InlineAsm>(CI->getOperand(0))) return false;
1741 // Resolve function pointers.
1742 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0)));
1743 if (!Callee) return false; // Cannot resolve.
1745 std::vector<Constant*> Formals;
1746 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
1747 Formals.push_back(getVal(Values, CI->getOperand(i)));
1749 if (Callee->isExternal()) {
1750 // If this is a function we can constant fold, do it.
1751 if (Constant *C = ConstantFoldCall(Callee, Formals)) {
1757 if (Callee->getFunctionType()->isVarArg())
1762 // Execute the call, if successful, use the return value.
1763 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
1764 MutatedMemory, AllocaTmps))
1766 InstResult = RetVal;
1768 } else if (TerminatorInst *TI = dyn_cast<TerminatorInst>(CurInst)) {
1769 BasicBlock *NewBB = 0;
1770 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
1771 if (BI->isUnconditional()) {
1772 NewBB = BI->getSuccessor(0);
1774 ConstantBool *Cond =
1775 dyn_cast<ConstantBool>(getVal(Values, BI->getCondition()));
1776 if (!Cond) return false; // Cannot determine.
1777 NewBB = BI->getSuccessor(!Cond->getValue());
1779 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
1781 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
1782 if (!Val) return false; // Cannot determine.
1783 NewBB = SI->getSuccessor(SI->findCaseValue(Val));
1784 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
1785 if (RI->getNumOperands())
1786 RetVal = getVal(Values, RI->getOperand(0));
1788 CallStack.pop_back(); // return from fn.
1789 return true; // We succeeded at evaluating this ctor!
1791 // invoke, unwind, unreachable.
1792 return false; // Cannot handle this terminator.
1795 // Okay, we succeeded in evaluating this control flow. See if we have
1796 // executed the new block before. If so, we have a looping function,
1797 // which we cannot evaluate in reasonable time.
1798 if (!ExecutedBlocks.insert(NewBB).second)
1799 return false; // looped!
1801 // Okay, we have never been in this block before. Check to see if there
1802 // are any PHI nodes. If so, evaluate them with information about where
1804 BasicBlock *OldBB = CurInst->getParent();
1805 CurInst = NewBB->begin();
1807 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
1808 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
1810 // Do NOT increment CurInst. We know that the terminator had no value.
1813 // Did not know how to evaluate this!
1817 if (!CurInst->use_empty())
1818 Values[CurInst] = InstResult;
1820 // Advance program counter.
1825 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
1826 /// we can. Return true if we can, false otherwise.
1827 static bool EvaluateStaticConstructor(Function *F) {
1828 /// MutatedMemory - For each store we execute, we update this map. Loads
1829 /// check this to get the most up-to-date value. If evaluation is successful,
1830 /// this state is committed to the process.
1831 std::map<Constant*, Constant*> MutatedMemory;
1833 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
1834 /// to represent its body. This vector is needed so we can delete the
1835 /// temporary globals when we are done.
1836 std::vector<GlobalVariable*> AllocaTmps;
1838 /// CallStack - This is used to detect recursion. In pathological situations
1839 /// we could hit exponential behavior, but at least there is nothing
1841 std::vector<Function*> CallStack;
1843 // Call the function.
1844 Constant *RetValDummy;
1845 bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(),
1846 CallStack, MutatedMemory, AllocaTmps);
1848 // We succeeded at evaluation: commit the result.
1849 DEBUG(std::cerr << "FULLY EVALUATED GLOBAL CTOR FUNCTION '" <<
1850 F->getName() << "' to " << MutatedMemory.size() << " stores.\n");
1851 for (std::map<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
1852 E = MutatedMemory.end(); I != E; ++I)
1853 CommitValueTo(I->second, I->first);
1856 // At this point, we are done interpreting. If we created any 'alloca'
1857 // temporaries, release them now.
1858 while (!AllocaTmps.empty()) {
1859 GlobalVariable *Tmp = AllocaTmps.back();
1860 AllocaTmps.pop_back();
1862 // If there are still users of the alloca, the program is doing something
1863 // silly, e.g. storing the address of the alloca somewhere and using it
1864 // later. Since this is undefined, we'll just make it be null.
1865 if (!Tmp->use_empty())
1866 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
1875 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
1876 /// Return true if anything changed.
1877 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
1878 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
1879 bool MadeChange = false;
1880 if (Ctors.empty()) return false;
1882 // Loop over global ctors, optimizing them when we can.
1883 for (unsigned i = 0; i != Ctors.size(); ++i) {
1884 Function *F = Ctors[i];
1885 // Found a null terminator in the middle of the list, prune off the rest of
1888 if (i != Ctors.size()-1) {
1895 // We cannot simplify external ctor functions.
1896 if (F->empty()) continue;
1898 // If we can evaluate the ctor at compile time, do.
1899 if (EvaluateStaticConstructor(F)) {
1900 Ctors.erase(Ctors.begin()+i);
1903 ++NumCtorsEvaluated;
1908 if (!MadeChange) return false;
1910 GCL = InstallGlobalCtors(GCL, Ctors);
1915 bool GlobalOpt::runOnModule(Module &M) {
1916 bool Changed = false;
1918 // Try to find the llvm.globalctors list.
1919 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
1921 bool LocalChange = true;
1922 while (LocalChange) {
1923 LocalChange = false;
1925 // Delete functions that are trivially dead, ccc -> fastcc
1926 LocalChange |= OptimizeFunctions(M);
1928 // Optimize global_ctors list.
1930 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
1932 // Optimize non-address-taken globals.
1933 LocalChange |= OptimizeGlobalVars(M);
1934 Changed |= LocalChange;
1937 // TODO: Move all global ctors functions to the end of the module for code