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<> NumHeapSRA ("globalopt", "Number of heap objects SRA'd");
39 Statistic<> NumSubstitute("globalopt",
40 "Number of globals with initializers stored into them");
41 Statistic<> NumDeleted ("globalopt", "Number of globals deleted");
42 Statistic<> NumFnDeleted("globalopt", "Number of functions deleted");
43 Statistic<> NumGlobUses ("globalopt", "Number of global uses devirtualized");
44 Statistic<> NumLocalized("globalopt", "Number of globals localized");
45 Statistic<> NumShrunkToBool("globalopt",
46 "Number of global vars shrunk to booleans");
47 Statistic<> NumFastCallFns("globalopt",
48 "Number of functions converted to fastcc");
49 Statistic<> NumCtorsEvaluated("globalopt","Number of static ctors evaluated");
51 struct GlobalOpt : public ModulePass {
52 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
53 AU.addRequired<TargetData>();
56 bool runOnModule(Module &M);
59 GlobalVariable *FindGlobalCtors(Module &M);
60 bool OptimizeFunctions(Module &M);
61 bool OptimizeGlobalVars(Module &M);
62 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
63 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
66 RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer");
69 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
71 /// GlobalStatus - As we analyze each global, keep track of some information
72 /// about it. If we find out that the address of the global is taken, none of
73 /// this info will be accurate.
75 /// isLoaded - True if the global is ever loaded. If the global isn't ever
76 /// loaded it can be deleted.
79 /// StoredType - Keep track of what stores to the global look like.
82 /// NotStored - There is no store to this global. It can thus be marked
86 /// isInitializerStored - This global is stored to, but the only thing
87 /// stored is the constant it was initialized with. This is only tracked
88 /// for scalar globals.
91 /// isStoredOnce - This global is stored to, but only its initializer and
92 /// one other value is ever stored to it. If this global isStoredOnce, we
93 /// track the value stored to it in StoredOnceValue below. This is only
94 /// tracked for scalar globals.
97 /// isStored - This global is stored to by multiple values or something else
98 /// that we cannot track.
102 /// StoredOnceValue - If only one value (besides the initializer constant) is
103 /// ever stored to this global, keep track of what value it is.
104 Value *StoredOnceValue;
106 /// AccessingFunction/HasMultipleAccessingFunctions - These start out
107 /// null/false. When the first accessing function is noticed, it is recorded.
108 /// When a second different accessing function is noticed,
109 /// HasMultipleAccessingFunctions is set to true.
110 Function *AccessingFunction;
111 bool HasMultipleAccessingFunctions;
113 /// HasNonInstructionUser - Set to true if this global has a user that is not
114 /// an instruction (e.g. a constant expr or GV initializer).
115 bool HasNonInstructionUser;
117 /// HasPHIUser - Set to true if this global has a user that is a PHI node.
120 /// isNotSuitableForSRA - Keep track of whether any SRA preventing users of
121 /// the global exist. Such users include GEP instruction with variable
122 /// indexes, and non-gep/load/store users like constant expr casts.
123 bool isNotSuitableForSRA;
125 GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0),
126 AccessingFunction(0), HasMultipleAccessingFunctions(false),
127 HasNonInstructionUser(false), HasPHIUser(false),
128 isNotSuitableForSRA(false) {}
133 /// ConstantIsDead - Return true if the specified constant is (transitively)
134 /// dead. The constant may be used by other constants (e.g. constant arrays and
135 /// constant exprs) as long as they are dead, but it cannot be used by anything
137 static bool ConstantIsDead(Constant *C) {
138 if (isa<GlobalValue>(C)) return false;
140 for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI)
141 if (Constant *CU = dyn_cast<Constant>(*UI)) {
142 if (!ConstantIsDead(CU)) return false;
149 /// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
150 /// structure. If the global has its address taken, return true to indicate we
151 /// can't do anything with it.
153 static bool AnalyzeGlobal(Value *V, GlobalStatus &GS,
154 std::set<PHINode*> &PHIUsers) {
155 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
156 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
157 GS.HasNonInstructionUser = true;
159 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
160 if (CE->getOpcode() != Instruction::GetElementPtr)
161 GS.isNotSuitableForSRA = true;
162 else if (!GS.isNotSuitableForSRA) {
163 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
164 // don't like < 3 operand CE's, and we don't like non-constant integer
166 if (CE->getNumOperands() < 3 || !CE->getOperand(1)->isNullValue())
167 GS.isNotSuitableForSRA = true;
169 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
170 if (!isa<ConstantInt>(CE->getOperand(i))) {
171 GS.isNotSuitableForSRA = true;
177 } else if (Instruction *I = dyn_cast<Instruction>(*UI)) {
178 if (!GS.HasMultipleAccessingFunctions) {
179 Function *F = I->getParent()->getParent();
180 if (GS.AccessingFunction == 0)
181 GS.AccessingFunction = F;
182 else if (GS.AccessingFunction != F)
183 GS.HasMultipleAccessingFunctions = true;
185 if (isa<LoadInst>(I)) {
187 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
188 // Don't allow a store OF the address, only stores TO the address.
189 if (SI->getOperand(0) == V) return true;
191 // If this is a direct store to the global (i.e., the global is a scalar
192 // value, not an aggregate), keep more specific information about
194 if (GS.StoredType != GlobalStatus::isStored)
195 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){
196 Value *StoredVal = SI->getOperand(0);
197 if (StoredVal == GV->getInitializer()) {
198 if (GS.StoredType < GlobalStatus::isInitializerStored)
199 GS.StoredType = GlobalStatus::isInitializerStored;
200 } else if (isa<LoadInst>(StoredVal) &&
201 cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
203 if (GS.StoredType < GlobalStatus::isInitializerStored)
204 GS.StoredType = GlobalStatus::isInitializerStored;
205 } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
206 GS.StoredType = GlobalStatus::isStoredOnce;
207 GS.StoredOnceValue = StoredVal;
208 } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
209 GS.StoredOnceValue == StoredVal) {
212 GS.StoredType = GlobalStatus::isStored;
215 GS.StoredType = GlobalStatus::isStored;
217 } else if (isa<GetElementPtrInst>(I)) {
218 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
220 // If the first two indices are constants, this can be SRA'd.
221 if (isa<GlobalVariable>(I->getOperand(0))) {
222 if (I->getNumOperands() < 3 || !isa<Constant>(I->getOperand(1)) ||
223 !cast<Constant>(I->getOperand(1))->isNullValue() ||
224 !isa<ConstantInt>(I->getOperand(2)))
225 GS.isNotSuitableForSRA = true;
226 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(I->getOperand(0))){
227 if (CE->getOpcode() != Instruction::GetElementPtr ||
228 CE->getNumOperands() < 3 || I->getNumOperands() < 2 ||
229 !isa<Constant>(I->getOperand(0)) ||
230 !cast<Constant>(I->getOperand(0))->isNullValue())
231 GS.isNotSuitableForSRA = true;
233 GS.isNotSuitableForSRA = true;
235 } else if (isa<SelectInst>(I)) {
236 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
237 GS.isNotSuitableForSRA = true;
238 } else if (PHINode *PN = dyn_cast<PHINode>(I)) {
239 // PHI nodes we can check just like select or GEP instructions, but we
240 // have to be careful about infinite recursion.
241 if (PHIUsers.insert(PN).second) // Not already visited.
242 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
243 GS.isNotSuitableForSRA = true;
244 GS.HasPHIUser = true;
245 } else if (isa<SetCondInst>(I)) {
246 GS.isNotSuitableForSRA = true;
247 } else if (isa<MemCpyInst>(I) || isa<MemMoveInst>(I)) {
248 if (I->getOperand(1) == V)
249 GS.StoredType = GlobalStatus::isStored;
250 if (I->getOperand(2) == V)
252 GS.isNotSuitableForSRA = true;
253 } else if (isa<MemSetInst>(I)) {
254 assert(I->getOperand(1) == V && "Memset only takes one pointer!");
255 GS.StoredType = GlobalStatus::isStored;
256 GS.isNotSuitableForSRA = true;
258 return true; // Any other non-load instruction might take address!
260 } else if (Constant *C = dyn_cast<Constant>(*UI)) {
261 GS.HasNonInstructionUser = true;
262 // We might have a dead and dangling constant hanging off of here.
263 if (!ConstantIsDead(C))
266 GS.HasNonInstructionUser = true;
267 // Otherwise must be some other user.
274 static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) {
275 ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
277 unsigned IdxV = CI->getZExtValue();
279 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
280 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
281 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
282 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
283 } else if (ConstantPacked *CP = dyn_cast<ConstantPacked>(Agg)) {
284 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
285 } else if (isa<ConstantAggregateZero>(Agg)) {
286 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
287 if (IdxV < STy->getNumElements())
288 return Constant::getNullValue(STy->getElementType(IdxV));
289 } else if (const SequentialType *STy =
290 dyn_cast<SequentialType>(Agg->getType())) {
291 return Constant::getNullValue(STy->getElementType());
293 } else if (isa<UndefValue>(Agg)) {
294 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
295 if (IdxV < STy->getNumElements())
296 return UndefValue::get(STy->getElementType(IdxV));
297 } else if (const SequentialType *STy =
298 dyn_cast<SequentialType>(Agg->getType())) {
299 return UndefValue::get(STy->getElementType());
306 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
307 /// users of the global, cleaning up the obvious ones. This is largely just a
308 /// quick scan over the use list to clean up the easy and obvious cruft. This
309 /// returns true if it made a change.
310 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
311 bool Changed = false;
312 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
315 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
317 // Replace the load with the initializer.
318 LI->replaceAllUsesWith(Init);
319 LI->eraseFromParent();
322 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
323 // Store must be unreachable or storing Init into the global.
324 SI->eraseFromParent();
326 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
327 if (CE->getOpcode() == Instruction::GetElementPtr) {
328 Constant *SubInit = 0;
330 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
331 Changed |= CleanupConstantGlobalUsers(CE, SubInit);
332 } else if (CE->getOpcode() == Instruction::Cast &&
333 isa<PointerType>(CE->getType())) {
334 // Pointer cast, delete any stores and memsets to the global.
335 Changed |= CleanupConstantGlobalUsers(CE, 0);
338 if (CE->use_empty()) {
339 CE->destroyConstant();
342 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
343 Constant *SubInit = 0;
345 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
346 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
347 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
348 Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
350 if (GEP->use_empty()) {
351 GEP->eraseFromParent();
354 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
355 if (MI->getRawDest() == V) {
356 MI->eraseFromParent();
360 } else if (Constant *C = dyn_cast<Constant>(U)) {
361 // If we have a chain of dead constantexprs or other things dangling from
362 // us, and if they are all dead, nuke them without remorse.
363 if (ConstantIsDead(C)) {
364 C->destroyConstant();
365 // This could have invalidated UI, start over from scratch.
366 CleanupConstantGlobalUsers(V, Init);
374 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
375 /// variable. This opens the door for other optimizations by exposing the
376 /// behavior of the program in a more fine-grained way. We have determined that
377 /// this transformation is safe already. We return the first global variable we
378 /// insert so that the caller can reprocess it.
379 static GlobalVariable *SRAGlobal(GlobalVariable *GV) {
380 assert(GV->hasInternalLinkage() && !GV->isConstant());
381 Constant *Init = GV->getInitializer();
382 const Type *Ty = Init->getType();
384 std::vector<GlobalVariable*> NewGlobals;
385 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
387 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
388 NewGlobals.reserve(STy->getNumElements());
389 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
390 Constant *In = getAggregateConstantElement(Init,
391 ConstantInt::get(Type::UIntTy, i));
392 assert(In && "Couldn't get element of initializer?");
393 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
394 GlobalVariable::InternalLinkage,
395 In, GV->getName()+"."+utostr(i));
396 Globals.insert(GV, NGV);
397 NewGlobals.push_back(NGV);
399 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
400 unsigned NumElements = 0;
401 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
402 NumElements = ATy->getNumElements();
403 else if (const PackedType *PTy = dyn_cast<PackedType>(STy))
404 NumElements = PTy->getNumElements();
406 assert(0 && "Unknown aggregate sequential type!");
408 if (NumElements > 16 && GV->hasNUsesOrMore(16))
409 return 0; // It's not worth it.
410 NewGlobals.reserve(NumElements);
411 for (unsigned i = 0, e = NumElements; i != e; ++i) {
412 Constant *In = getAggregateConstantElement(Init,
413 ConstantInt::get(Type::UIntTy, i));
414 assert(In && "Couldn't get element of initializer?");
416 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
417 GlobalVariable::InternalLinkage,
418 In, GV->getName()+"."+utostr(i));
419 Globals.insert(GV, NGV);
420 NewGlobals.push_back(NGV);
424 if (NewGlobals.empty())
427 DOUT << "PERFORMING GLOBAL SRA ON: " << *GV;
429 Constant *NullInt = Constant::getNullValue(Type::IntTy);
431 // Loop over all of the uses of the global, replacing the constantexpr geps,
432 // with smaller constantexpr geps or direct references.
433 while (!GV->use_empty()) {
434 User *GEP = GV->use_back();
435 assert(((isa<ConstantExpr>(GEP) &&
436 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
437 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
439 // Ignore the 1th operand, which has to be zero or else the program is quite
440 // broken (undefined). Get the 2nd operand, which is the structure or array
442 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
443 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
445 Value *NewPtr = NewGlobals[Val];
447 // Form a shorter GEP if needed.
448 if (GEP->getNumOperands() > 3)
449 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
450 std::vector<Constant*> Idxs;
451 Idxs.push_back(NullInt);
452 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
453 Idxs.push_back(CE->getOperand(i));
454 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr), Idxs);
456 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
457 std::vector<Value*> Idxs;
458 Idxs.push_back(NullInt);
459 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
460 Idxs.push_back(GEPI->getOperand(i));
461 NewPtr = new GetElementPtrInst(NewPtr, Idxs,
462 GEPI->getName()+"."+utostr(Val), GEPI);
464 GEP->replaceAllUsesWith(NewPtr);
466 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
467 GEPI->eraseFromParent();
469 cast<ConstantExpr>(GEP)->destroyConstant();
472 // Delete the old global, now that it is dead.
476 // Loop over the new globals array deleting any globals that are obviously
477 // dead. This can arise due to scalarization of a structure or an array that
478 // has elements that are dead.
479 unsigned FirstGlobal = 0;
480 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
481 if (NewGlobals[i]->use_empty()) {
482 Globals.erase(NewGlobals[i]);
483 if (FirstGlobal == i) ++FirstGlobal;
486 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
489 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
490 /// value will trap if the value is dynamically null.
491 static bool AllUsesOfValueWillTrapIfNull(Value *V) {
492 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
493 if (isa<LoadInst>(*UI)) {
495 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
496 if (SI->getOperand(0) == V) {
497 //llvm_cerr << "NONTRAPPING USE: " << **UI;
498 return false; // Storing the value.
500 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
501 if (CI->getOperand(0) != V) {
502 //llvm_cerr << "NONTRAPPING USE: " << **UI;
503 return false; // Not calling the ptr
505 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
506 if (II->getOperand(0) != V) {
507 //llvm_cerr << "NONTRAPPING USE: " << **UI;
508 return false; // Not calling the ptr
510 } else if (CastInst *CI = dyn_cast<CastInst>(*UI)) {
511 if (!AllUsesOfValueWillTrapIfNull(CI)) return false;
512 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
513 if (!AllUsesOfValueWillTrapIfNull(GEPI)) return false;
514 } else if (isa<SetCondInst>(*UI) &&
515 isa<ConstantPointerNull>(UI->getOperand(1))) {
516 // Ignore setcc X, null
518 //llvm_cerr << "NONTRAPPING USE: " << **UI;
524 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
525 /// from GV will trap if the loaded value is null. Note that this also permits
526 /// comparisons of the loaded value against null, as a special case.
527 static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) {
528 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI)
529 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
530 if (!AllUsesOfValueWillTrapIfNull(LI))
532 } else if (isa<StoreInst>(*UI)) {
533 // Ignore stores to the global.
535 // We don't know or understand this user, bail out.
536 //llvm_cerr << "UNKNOWN USER OF GLOBAL!: " << **UI;
543 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
544 bool Changed = false;
545 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
546 Instruction *I = cast<Instruction>(*UI++);
547 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
548 LI->setOperand(0, NewV);
550 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
551 if (SI->getOperand(1) == V) {
552 SI->setOperand(1, NewV);
555 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
556 if (I->getOperand(0) == V) {
557 // Calling through the pointer! Turn into a direct call, but be careful
558 // that the pointer is not also being passed as an argument.
559 I->setOperand(0, NewV);
561 bool PassedAsArg = false;
562 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
563 if (I->getOperand(i) == V) {
565 I->setOperand(i, NewV);
569 // Being passed as an argument also. Be careful to not invalidate UI!
573 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
574 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
575 ConstantExpr::getCast(NewV, CI->getType()));
576 if (CI->use_empty()) {
578 CI->eraseFromParent();
580 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
581 // Should handle GEP here.
582 std::vector<Constant*> Indices;
583 Indices.reserve(GEPI->getNumOperands()-1);
584 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
585 if (Constant *C = dyn_cast<Constant>(GEPI->getOperand(i)))
586 Indices.push_back(C);
589 if (Indices.size() == GEPI->getNumOperands()-1)
590 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
591 ConstantExpr::getGetElementPtr(NewV, Indices));
592 if (GEPI->use_empty()) {
594 GEPI->eraseFromParent();
603 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
604 /// value stored into it. If there are uses of the loaded value that would trap
605 /// if the loaded value is dynamically null, then we know that they cannot be
606 /// reachable with a null optimize away the load.
607 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) {
608 std::vector<LoadInst*> Loads;
609 bool Changed = false;
611 // Replace all uses of loads with uses of uses of the stored value.
612 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end();
614 if (LoadInst *LI = dyn_cast<LoadInst>(*GUI)) {
616 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
618 assert(isa<StoreInst>(*GUI) && "Only expect load and stores!");
622 DOUT << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV;
626 // Delete all of the loads we can, keeping track of whether we nuked them all!
627 bool AllLoadsGone = true;
628 while (!Loads.empty()) {
629 LoadInst *L = Loads.back();
630 if (L->use_empty()) {
631 L->eraseFromParent();
634 AllLoadsGone = false;
639 // If we nuked all of the loads, then none of the stores are needed either,
640 // nor is the global.
642 DOUT << " *** GLOBAL NOW DEAD!\n";
643 CleanupConstantGlobalUsers(GV, 0);
644 if (GV->use_empty()) {
645 GV->eraseFromParent();
653 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
654 /// instructions that are foldable.
655 static void ConstantPropUsersOf(Value *V) {
656 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
657 if (Instruction *I = dyn_cast<Instruction>(*UI++))
658 if (Constant *NewC = ConstantFoldInstruction(I)) {
659 I->replaceAllUsesWith(NewC);
661 // Advance UI to the next non-I use to avoid invalidating it!
662 // Instructions could multiply use V.
663 while (UI != E && *UI == I)
665 I->eraseFromParent();
669 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
670 /// variable, and transforms the program as if it always contained the result of
671 /// the specified malloc. Because it is always the result of the specified
672 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
673 /// malloc into a global, and any laods of GV as uses of the new global.
674 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
676 DOUT << "PROMOTING MALLOC GLOBAL: " << *GV << " MALLOC = " << *MI;
677 ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize());
679 if (NElements->getZExtValue() != 1) {
680 // If we have an array allocation, transform it to a single element
681 // allocation to make the code below simpler.
682 Type *NewTy = ArrayType::get(MI->getAllocatedType(),
683 NElements->getZExtValue());
685 new MallocInst(NewTy, Constant::getNullValue(Type::UIntTy),
686 MI->getAlignment(), MI->getName(), MI);
687 std::vector<Value*> Indices;
688 Indices.push_back(Constant::getNullValue(Type::IntTy));
689 Indices.push_back(Indices[0]);
690 Value *NewGEP = new GetElementPtrInst(NewMI, Indices,
691 NewMI->getName()+".el0", MI);
692 MI->replaceAllUsesWith(NewGEP);
693 MI->eraseFromParent();
697 // Create the new global variable. The contents of the malloc'd memory is
698 // undefined, so initialize with an undef value.
699 Constant *Init = UndefValue::get(MI->getAllocatedType());
700 GlobalVariable *NewGV = new GlobalVariable(MI->getAllocatedType(), false,
701 GlobalValue::InternalLinkage, Init,
702 GV->getName()+".body");
703 GV->getParent()->getGlobalList().insert(GV, NewGV);
705 // Anything that used the malloc now uses the global directly.
706 MI->replaceAllUsesWith(NewGV);
708 Constant *RepValue = NewGV;
709 if (NewGV->getType() != GV->getType()->getElementType())
710 RepValue = ConstantExpr::getCast(RepValue, GV->getType()->getElementType());
712 // If there is a comparison against null, we will insert a global bool to
713 // keep track of whether the global was initialized yet or not.
714 GlobalVariable *InitBool =
715 new GlobalVariable(Type::BoolTy, false, GlobalValue::InternalLinkage,
716 ConstantBool::getFalse(), GV->getName()+".init");
717 bool InitBoolUsed = false;
719 // Loop over all uses of GV, processing them in turn.
720 std::vector<StoreInst*> Stores;
721 while (!GV->use_empty())
722 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
723 while (!LI->use_empty()) {
724 Use &LoadUse = LI->use_begin().getUse();
725 if (!isa<SetCondInst>(LoadUse.getUser()))
728 // Replace the setcc X, 0 with a use of the bool value.
729 SetCondInst *SCI = cast<SetCondInst>(LoadUse.getUser());
730 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", SCI);
732 switch (SCI->getOpcode()) {
733 default: assert(0 && "Unknown opcode!");
734 case Instruction::SetLT:
735 LV = ConstantBool::getFalse(); // X < null -> always false
737 case Instruction::SetEQ:
738 case Instruction::SetLE:
739 LV = BinaryOperator::createNot(LV, "notinit", SCI);
741 case Instruction::SetNE:
742 case Instruction::SetGE:
743 case Instruction::SetGT:
746 SCI->replaceAllUsesWith(LV);
747 SCI->eraseFromParent();
750 LI->eraseFromParent();
752 StoreInst *SI = cast<StoreInst>(GV->use_back());
753 // The global is initialized when the store to it occurs.
754 new StoreInst(ConstantBool::getTrue(), InitBool, SI);
755 SI->eraseFromParent();
758 // If the initialization boolean was used, insert it, otherwise delete it.
760 while (!InitBool->use_empty()) // Delete initializations
761 cast<Instruction>(InitBool->use_back())->eraseFromParent();
764 GV->getParent()->getGlobalList().insert(GV, InitBool);
767 // Now the GV is dead, nuke it and the malloc.
768 GV->eraseFromParent();
769 MI->eraseFromParent();
771 // To further other optimizations, loop over all users of NewGV and try to
772 // constant prop them. This will promote GEP instructions with constant
773 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
774 ConstantPropUsersOf(NewGV);
775 if (RepValue != NewGV)
776 ConstantPropUsersOf(RepValue);
781 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
782 /// to make sure that there are no complex uses of V. We permit simple things
783 /// like dereferencing the pointer, but not storing through the address, unless
784 /// it is to the specified global.
785 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V,
786 GlobalVariable *GV) {
787 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI)
788 if (isa<LoadInst>(*UI) || isa<SetCondInst>(*UI)) {
790 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
791 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
792 return false; // Storing the pointer itself... bad.
793 // Otherwise, storing through it, or storing into GV... fine.
794 } else if (isa<GetElementPtrInst>(*UI) || isa<SelectInst>(*UI)) {
795 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(cast<Instruction>(*UI),GV))
803 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
804 /// somewhere. Transform all uses of the allocation into loads from the
805 /// global and uses of the resultant pointer. Further, delete the store into
806 /// GV. This assumes that these value pass the
807 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
808 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
809 GlobalVariable *GV) {
810 while (!Alloc->use_empty()) {
811 Instruction *U = Alloc->use_back();
812 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
813 // If this is the store of the allocation into the global, remove it.
814 if (SI->getOperand(1) == GV) {
815 SI->eraseFromParent();
820 // Insert a load from the global, and use it instead of the malloc.
821 Value *NL = new LoadInst(GV, GV->getName()+".val", U);
822 U->replaceUsesOfWith(Alloc, NL);
826 /// GlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
827 /// GV are simple enough to perform HeapSRA, return true.
828 static bool GlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV) {
829 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;
831 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
832 // We permit two users of the load: setcc comparing against the null
833 // pointer, and a getelementptr of a specific form.
834 for (Value::use_iterator UI = LI->use_begin(), E = LI->use_end(); UI != E;
836 // Comparison against null is ok.
837 if (SetCondInst *SCI = dyn_cast<SetCondInst>(*UI)) {
838 if (!isa<ConstantPointerNull>(SCI->getOperand(1)))
843 // getelementptr is also ok, but only a simple form.
844 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI);
845 if (!GEPI) return false;
847 // Must index into the array and into the struct.
848 if (GEPI->getNumOperands() < 3)
851 // Otherwise the GEP is ok.
858 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
859 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
860 /// use FieldGlobals instead. All uses of loaded values satisfy
861 /// GlobalLoadUsesSimpleEnoughForHeapSRA.
862 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Ptr,
863 const std::vector<GlobalVariable*> &FieldGlobals) {
864 std::vector<Value *> InsertedLoadsForPtr;
865 //InsertedLoadsForPtr.resize(FieldGlobals.size());
866 while (!Ptr->use_empty()) {
867 Instruction *User = Ptr->use_back();
869 // If this is a comparison against null, handle it.
870 if (SetCondInst *SCI = dyn_cast<SetCondInst>(User)) {
871 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
872 // If we have a setcc of the loaded pointer, we can use a setcc of any
875 if (InsertedLoadsForPtr.empty()) {
876 NPtr = new LoadInst(FieldGlobals[0], Ptr->getName()+".f0", Ptr);
877 InsertedLoadsForPtr.push_back(Ptr);
879 NPtr = InsertedLoadsForPtr.back();
882 Value *New = new SetCondInst(SCI->getOpcode(), NPtr,
883 Constant::getNullValue(NPtr->getType()),
884 SCI->getName(), SCI);
885 SCI->replaceAllUsesWith(New);
886 SCI->eraseFromParent();
890 // Otherwise, this should be: 'getelementptr Ptr, Idx, uint FieldNo ...'
891 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
892 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
893 && GEPI->getOperand(2)->getType()->isUnsigned()
894 && "Unexpected GEPI!");
896 // Load the pointer for this field.
897 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
898 if (InsertedLoadsForPtr.size() <= FieldNo)
899 InsertedLoadsForPtr.resize(FieldNo+1);
900 if (InsertedLoadsForPtr[FieldNo] == 0)
901 InsertedLoadsForPtr[FieldNo] = new LoadInst(FieldGlobals[FieldNo],
902 Ptr->getName()+".f" +
903 utostr(FieldNo), Ptr);
904 Value *NewPtr = InsertedLoadsForPtr[FieldNo];
906 // Create the new GEP idx vector.
907 std::vector<Value*> GEPIdx;
908 GEPIdx.push_back(GEPI->getOperand(1));
909 GEPIdx.insert(GEPIdx.end(), GEPI->op_begin()+3, GEPI->op_end());
911 Value *NGEPI = new GetElementPtrInst(NewPtr, GEPIdx, GEPI->getName(), GEPI);
912 GEPI->replaceAllUsesWith(NGEPI);
913 GEPI->eraseFromParent();
917 /// PerformHeapAllocSRoA - MI is an allocation of an array of structures. Break
918 /// it up into multiple allocations of arrays of the fields.
919 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){
920 DOUT << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *MI;
921 const StructType *STy = cast<StructType>(MI->getAllocatedType());
923 // There is guaranteed to be at least one use of the malloc (storing
924 // it into GV). If there are other uses, change them to be uses of
925 // the global to simplify later code. This also deletes the store
927 ReplaceUsesOfMallocWithGlobal(MI, GV);
929 // Okay, at this point, there are no users of the malloc. Insert N
930 // new mallocs at the same place as MI, and N globals.
931 std::vector<GlobalVariable*> FieldGlobals;
932 std::vector<MallocInst*> FieldMallocs;
934 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
935 const Type *FieldTy = STy->getElementType(FieldNo);
936 const Type *PFieldTy = PointerType::get(FieldTy);
938 GlobalVariable *NGV =
939 new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage,
940 Constant::getNullValue(PFieldTy),
941 GV->getName() + ".f" + utostr(FieldNo), GV);
942 FieldGlobals.push_back(NGV);
944 MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(),
945 MI->getName() + ".f" + utostr(FieldNo),MI);
946 FieldMallocs.push_back(NMI);
947 new StoreInst(NMI, NGV, MI);
950 // The tricky aspect of this transformation is handling the case when malloc
951 // fails. In the original code, malloc failing would set the result pointer
952 // of malloc to null. In this case, some mallocs could succeed and others
953 // could fail. As such, we emit code that looks like this:
954 // F0 = malloc(field0)
955 // F1 = malloc(field1)
956 // F2 = malloc(field2)
957 // if (F0 == 0 || F1 == 0 || F2 == 0) {
958 // if (F0) { free(F0); F0 = 0; }
959 // if (F1) { free(F1); F1 = 0; }
960 // if (F2) { free(F2); F2 = 0; }
962 Value *RunningOr = 0;
963 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
964 Value *Cond = new SetCondInst(Instruction::SetEQ, FieldMallocs[i],
965 Constant::getNullValue(FieldMallocs[i]->getType()),
968 RunningOr = Cond; // First seteq
970 RunningOr = BinaryOperator::createOr(RunningOr, Cond, "tmp", MI);
973 // Split the basic block at the old malloc.
974 BasicBlock *OrigBB = MI->getParent();
975 BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont");
977 // Create the block to check the first condition. Put all these blocks at the
978 // end of the function as they are unlikely to be executed.
979 BasicBlock *NullPtrBlock = new BasicBlock("malloc_ret_null",
980 OrigBB->getParent());
982 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
983 // branch on RunningOr.
984 OrigBB->getTerminator()->eraseFromParent();
985 new BranchInst(NullPtrBlock, ContBB, RunningOr, OrigBB);
987 // Within the NullPtrBlock, we need to emit a comparison and branch for each
988 // pointer, because some may be null while others are not.
989 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
990 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
991 Value *Cmp = new SetCondInst(Instruction::SetNE, GVVal,
992 Constant::getNullValue(GVVal->getType()),
993 "tmp", NullPtrBlock);
994 BasicBlock *FreeBlock = new BasicBlock("free_it", OrigBB->getParent());
995 BasicBlock *NextBlock = new BasicBlock("next", OrigBB->getParent());
996 new BranchInst(FreeBlock, NextBlock, Cmp, NullPtrBlock);
998 // Fill in FreeBlock.
999 new FreeInst(GVVal, FreeBlock);
1000 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1002 new BranchInst(NextBlock, FreeBlock);
1004 NullPtrBlock = NextBlock;
1007 new BranchInst(ContBB, NullPtrBlock);
1010 // MI is no longer needed, remove it.
1011 MI->eraseFromParent();
1014 // Okay, the malloc site is completely handled. All of the uses of GV are now
1015 // loads, and all uses of those loads are simple. Rewrite them to use loads
1016 // of the per-field globals instead.
1017 while (!GV->use_empty()) {
1018 LoadInst *LI = cast<LoadInst>(GV->use_back());
1019 RewriteUsesOfLoadForHeapSRoA(LI, FieldGlobals);
1020 LI->eraseFromParent();
1023 // The old global is now dead, remove it.
1024 GV->eraseFromParent();
1027 return FieldGlobals[0];
1031 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1032 // that only one value (besides its initializer) is ever stored to the global.
1033 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1034 Module::global_iterator &GVI,
1036 if (CastInst *CI = dyn_cast<CastInst>(StoredOnceVal))
1037 StoredOnceVal = CI->getOperand(0);
1038 else if (GetElementPtrInst *GEPI =dyn_cast<GetElementPtrInst>(StoredOnceVal)){
1039 // "getelementptr Ptr, 0, 0, 0" is really just a cast.
1040 bool IsJustACast = true;
1041 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
1042 if (!isa<Constant>(GEPI->getOperand(i)) ||
1043 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
1044 IsJustACast = false;
1048 StoredOnceVal = GEPI->getOperand(0);
1051 // If we are dealing with a pointer global that is initialized to null and
1052 // only has one (non-null) value stored into it, then we can optimize any
1053 // users of the loaded value (often calls and loads) that would trap if the
1055 if (isa<PointerType>(GV->getInitializer()->getType()) &&
1056 GV->getInitializer()->isNullValue()) {
1057 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1058 if (GV->getInitializer()->getType() != SOVC->getType())
1059 SOVC = ConstantExpr::getCast(SOVC, GV->getInitializer()->getType());
1061 // Optimize away any trapping uses of the loaded value.
1062 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1064 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) {
1065 // If this is a malloc of an abstract type, don't touch it.
1066 if (!MI->getAllocatedType()->isSized())
1069 // We can't optimize this global unless all uses of it are *known* to be
1070 // of the malloc value, not of the null initializer value (consider a use
1071 // that compares the global's value against zero to see if the malloc has
1072 // been reached). To do this, we check to see if all uses of the global
1073 // would trap if the global were null: this proves that they must all
1074 // happen after the malloc.
1075 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1078 // We can't optimize this if the malloc itself is used in a complex way,
1079 // for example, being stored into multiple globals. This allows the
1080 // malloc to be stored into the specified global, loaded setcc'd, and
1081 // GEP'd. These are all things we could transform to using the global
1083 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV))
1087 // If we have a global that is only initialized with a fixed size malloc,
1088 // transform the program to use global memory instead of malloc'd memory.
1089 // This eliminates dynamic allocation, avoids an indirection accessing the
1090 // data, and exposes the resultant global to further GlobalOpt.
1091 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) {
1092 // Restrict this transformation to only working on small allocations
1093 // (2048 bytes currently), as we don't want to introduce a 16M global or
1095 if (NElements->getZExtValue()*
1096 TD.getTypeSize(MI->getAllocatedType()) < 2048) {
1097 GVI = OptimizeGlobalAddressOfMalloc(GV, MI);
1102 // If the allocation is an array of structures, consider transforming this
1103 // into multiple malloc'd arrays, one for each field. This is basically
1104 // SRoA for malloc'd memory.
1105 if (const StructType *AllocTy =
1106 dyn_cast<StructType>(MI->getAllocatedType())) {
1107 // This the structure has an unreasonable number of fields, leave it
1109 if (AllocTy->getNumElements() <= 16 && AllocTy->getNumElements() > 0 &&
1110 GlobalLoadUsesSimpleEnoughForHeapSRA(GV)) {
1111 GVI = PerformHeapAllocSRoA(GV, MI);
1121 /// ShrinkGlobalToBoolean - At this point, we have learned that the only two
1122 /// values ever stored into GV are its initializer and OtherVal.
1123 static void ShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1124 // Create the new global, initializing it to false.
1125 GlobalVariable *NewGV = new GlobalVariable(Type::BoolTy, false,
1126 GlobalValue::InternalLinkage, ConstantBool::getFalse(),
1127 GV->getName()+".b");
1128 GV->getParent()->getGlobalList().insert(GV, NewGV);
1130 Constant *InitVal = GV->getInitializer();
1131 assert(InitVal->getType() != Type::BoolTy && "No reason to shrink to bool!");
1133 // If initialized to zero and storing one into the global, we can use a cast
1134 // instead of a select to synthesize the desired value.
1135 bool IsOneZero = false;
1136 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1137 IsOneZero = InitVal->isNullValue() && CI->equalsInt(1);
1139 while (!GV->use_empty()) {
1140 Instruction *UI = cast<Instruction>(GV->use_back());
1141 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1142 // Change the store into a boolean store.
1143 bool StoringOther = SI->getOperand(0) == OtherVal;
1144 // Only do this if we weren't storing a loaded value.
1146 if (StoringOther || SI->getOperand(0) == InitVal)
1147 StoreVal = ConstantBool::get(StoringOther);
1149 // Otherwise, we are storing a previously loaded copy. To do this,
1150 // change the copy from copying the original value to just copying the
1152 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1154 // If we're already replaced the input, StoredVal will be a cast or
1155 // select instruction. If not, it will be a load of the original
1157 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1158 assert(LI->getOperand(0) == GV && "Not a copy!");
1159 // Insert a new load, to preserve the saved value.
1160 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1162 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1163 "This is not a form that we understand!");
1164 StoreVal = StoredVal->getOperand(0);
1165 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1168 new StoreInst(StoreVal, NewGV, SI);
1169 } else if (!UI->use_empty()) {
1170 // Change the load into a load of bool then a select.
1171 LoadInst *LI = cast<LoadInst>(UI);
1173 std::string Name = LI->getName(); LI->setName("");
1174 LoadInst *NLI = new LoadInst(NewGV, Name+".b", LI);
1177 NSI = new CastInst(NLI, LI->getType(), Name, LI);
1179 NSI = new SelectInst(NLI, OtherVal, InitVal, Name, LI);
1180 LI->replaceAllUsesWith(NSI);
1182 UI->eraseFromParent();
1185 GV->eraseFromParent();
1189 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1190 /// it if possible. If we make a change, return true.
1191 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1192 Module::global_iterator &GVI) {
1193 std::set<PHINode*> PHIUsers;
1195 GV->removeDeadConstantUsers();
1197 if (GV->use_empty()) {
1198 DOUT << "GLOBAL DEAD: " << *GV;
1199 GV->eraseFromParent();
1204 if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
1206 llvm_cerr << "Global: " << *GV;
1207 llvm_cerr << " isLoaded = " << GS.isLoaded << "\n";
1208 llvm_cerr << " StoredType = ";
1209 switch (GS.StoredType) {
1210 case GlobalStatus::NotStored: llvm_cerr << "NEVER STORED\n"; break;
1211 case GlobalStatus::isInitializerStored: llvm_cerr << "INIT STORED\n"; break;
1212 case GlobalStatus::isStoredOnce: llvm_cerr << "STORED ONCE\n"; break;
1213 case GlobalStatus::isStored: llvm_cerr << "stored\n"; break;
1215 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
1216 llvm_cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n";
1217 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
1218 llvm_cerr << " AccessingFunction = " << GS.AccessingFunction->getName()
1220 llvm_cerr << " HasMultipleAccessingFunctions = "
1221 << GS.HasMultipleAccessingFunctions << "\n";
1222 llvm_cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n";
1223 llvm_cerr << " isNotSuitableForSRA = " << GS.isNotSuitableForSRA << "\n";
1227 // If this is a first class global and has only one accessing function
1228 // and this function is main (which we know is not recursive we can make
1229 // this global a local variable) we replace the global with a local alloca
1230 // in this function.
1232 // NOTE: It doesn't make sense to promote non first class types since we
1233 // are just replacing static memory to stack memory.
1234 if (!GS.HasMultipleAccessingFunctions &&
1235 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1236 GV->getType()->getElementType()->isFirstClassType() &&
1237 GS.AccessingFunction->getName() == "main" &&
1238 GS.AccessingFunction->hasExternalLinkage()) {
1239 DOUT << "LOCALIZING GLOBAL: " << *GV;
1240 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin();
1241 const Type* ElemTy = GV->getType()->getElementType();
1242 // FIXME: Pass Global's alignment when globals have alignment
1243 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI);
1244 if (!isa<UndefValue>(GV->getInitializer()))
1245 new StoreInst(GV->getInitializer(), Alloca, FirstI);
1247 GV->replaceAllUsesWith(Alloca);
1248 GV->eraseFromParent();
1253 // If the global is never loaded (but may be stored to), it is dead.
1256 DOUT << "GLOBAL NEVER LOADED: " << *GV;
1258 // Delete any stores we can find to the global. We may not be able to
1259 // make it completely dead though.
1260 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1262 // If the global is dead now, delete it.
1263 if (GV->use_empty()) {
1264 GV->eraseFromParent();
1270 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1271 DOUT << "MARKING CONSTANT: " << *GV;
1272 GV->setConstant(true);
1274 // Clean up any obviously simplifiable users now.
1275 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1277 // If the global is dead now, just nuke it.
1278 if (GV->use_empty()) {
1279 DOUT << " *** Marking constant allowed us to simplify "
1280 << "all users and delete global!\n";
1281 GV->eraseFromParent();
1287 } else if (!GS.isNotSuitableForSRA &&
1288 !GV->getInitializer()->getType()->isFirstClassType()) {
1289 if (GlobalVariable *FirstNewGV = SRAGlobal(GV)) {
1290 GVI = FirstNewGV; // Don't skip the newly produced globals!
1293 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1294 // If the initial value for the global was an undef value, and if only
1295 // one other value was stored into it, we can just change the
1296 // initializer to be an undef value, then delete all stores to the
1297 // global. This allows us to mark it constant.
1298 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1299 if (isa<UndefValue>(GV->getInitializer())) {
1300 // Change the initial value here.
1301 GV->setInitializer(SOVConstant);
1303 // Clean up any obviously simplifiable users now.
1304 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1306 if (GV->use_empty()) {
1307 DOUT << " *** Substituting initializer allowed us to "
1308 << "simplify all users and delete global!\n";
1309 GV->eraseFromParent();
1318 // Try to optimize globals based on the knowledge that only one value
1319 // (besides its initializer) is ever stored to the global.
1320 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1321 getAnalysis<TargetData>()))
1324 // Otherwise, if the global was not a boolean, we can shrink it to be a
1326 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1327 if (GV->getType()->getElementType() != Type::BoolTy &&
1328 !GV->getType()->getElementType()->isFloatingPoint() &&
1330 DOUT << " *** SHRINKING TO BOOL: " << *GV;
1331 ShrinkGlobalToBoolean(GV, SOVConstant);
1340 /// OnlyCalledDirectly - Return true if the specified function is only called
1341 /// directly. In other words, its address is never taken.
1342 static bool OnlyCalledDirectly(Function *F) {
1343 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1344 Instruction *User = dyn_cast<Instruction>(*UI);
1345 if (!User) return false;
1346 if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false;
1348 // See if the function address is passed as an argument.
1349 for (unsigned i = 1, e = User->getNumOperands(); i != e; ++i)
1350 if (User->getOperand(i) == F) return false;
1355 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1356 /// function, changing them to FastCC.
1357 static void ChangeCalleesToFastCall(Function *F) {
1358 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1359 Instruction *User = cast<Instruction>(*UI);
1360 if (CallInst *CI = dyn_cast<CallInst>(User))
1361 CI->setCallingConv(CallingConv::Fast);
1363 cast<InvokeInst>(User)->setCallingConv(CallingConv::Fast);
1367 bool GlobalOpt::OptimizeFunctions(Module &M) {
1368 bool Changed = false;
1369 // Optimize functions.
1370 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1372 F->removeDeadConstantUsers();
1373 if (F->use_empty() && (F->hasInternalLinkage() ||
1374 F->hasLinkOnceLinkage())) {
1375 M.getFunctionList().erase(F);
1378 } else if (F->hasInternalLinkage() &&
1379 F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1380 OnlyCalledDirectly(F)) {
1381 // If this function has C calling conventions, is not a varargs
1382 // function, and is only called directly, promote it to use the Fast
1383 // calling convention.
1384 F->setCallingConv(CallingConv::Fast);
1385 ChangeCalleesToFastCall(F);
1393 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1394 bool Changed = false;
1395 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1397 GlobalVariable *GV = GVI++;
1398 if (!GV->isConstant() && GV->hasInternalLinkage() &&
1399 GV->hasInitializer())
1400 Changed |= ProcessInternalGlobal(GV, GVI);
1405 /// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1406 /// initializers have an init priority of 65535.
1407 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1408 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1410 if (I->getName() == "llvm.global_ctors") {
1411 // Found it, verify it's an array of { int, void()* }.
1412 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1414 const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1415 if (!STy || STy->getNumElements() != 2 ||
1416 STy->getElementType(0) != Type::IntTy) return 0;
1417 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1418 if (!PFTy) return 0;
1419 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1420 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() ||
1421 FTy->getNumParams() != 0)
1424 // Verify that the initializer is simple enough for us to handle.
1425 if (!I->hasInitializer()) return 0;
1426 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1428 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1429 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(CA->getOperand(i))) {
1430 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1433 // Must have a function or null ptr.
1434 if (!isa<Function>(CS->getOperand(1)))
1437 // Init priority must be standard.
1438 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1439 if (!CI || CI->getZExtValue() != 65535)
1450 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1451 /// return a list of the functions and null terminator as a vector.
1452 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1453 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1454 std::vector<Function*> Result;
1455 Result.reserve(CA->getNumOperands());
1456 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) {
1457 ConstantStruct *CS = cast<ConstantStruct>(CA->getOperand(i));
1458 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1463 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1464 /// specified array, returning the new global to use.
1465 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1466 const std::vector<Function*> &Ctors) {
1467 // If we made a change, reassemble the initializer list.
1468 std::vector<Constant*> CSVals;
1469 CSVals.push_back(ConstantInt::get(Type::IntTy, 65535));
1470 CSVals.push_back(0);
1472 // Create the new init list.
1473 std::vector<Constant*> CAList;
1474 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1476 CSVals[1] = Ctors[i];
1478 const Type *FTy = FunctionType::get(Type::VoidTy,
1479 std::vector<const Type*>(), false);
1480 const PointerType *PFTy = PointerType::get(FTy);
1481 CSVals[1] = Constant::getNullValue(PFTy);
1482 CSVals[0] = ConstantInt::get(Type::IntTy, 2147483647);
1484 CAList.push_back(ConstantStruct::get(CSVals));
1487 // Create the array initializer.
1488 const Type *StructTy =
1489 cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
1490 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()),
1493 // If we didn't change the number of elements, don't create a new GV.
1494 if (CA->getType() == GCL->getInitializer()->getType()) {
1495 GCL->setInitializer(CA);
1499 // Create the new global and insert it next to the existing list.
1500 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
1501 GCL->getLinkage(), CA,
1504 GCL->getParent()->getGlobalList().insert(GCL, NGV);
1506 // Nuke the old list, replacing any uses with the new one.
1507 if (!GCL->use_empty()) {
1509 if (V->getType() != GCL->getType())
1510 V = ConstantExpr::getCast(V, GCL->getType());
1511 GCL->replaceAllUsesWith(V);
1513 GCL->eraseFromParent();
1522 static Constant *getVal(std::map<Value*, Constant*> &ComputedValues,
1524 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
1525 Constant *R = ComputedValues[V];
1526 assert(R && "Reference to an uncomputed value!");
1530 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
1531 /// enough for us to understand. In particular, if it is a cast of something,
1532 /// we punt. We basically just support direct accesses to globals and GEP's of
1533 /// globals. This should be kept up to date with CommitValueTo.
1534 static bool isSimpleEnoughPointerToCommit(Constant *C) {
1535 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
1536 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1537 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1538 return !GV->isExternal(); // reject external globals.
1540 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
1541 // Handle a constantexpr gep.
1542 if (CE->getOpcode() == Instruction::GetElementPtr &&
1543 isa<GlobalVariable>(CE->getOperand(0))) {
1544 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1545 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1546 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1547 return GV->hasInitializer() &&
1548 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1553 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
1554 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
1555 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
1556 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
1557 ConstantExpr *Addr, unsigned OpNo) {
1558 // Base case of the recursion.
1559 if (OpNo == Addr->getNumOperands()) {
1560 assert(Val->getType() == Init->getType() && "Type mismatch!");
1564 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
1565 std::vector<Constant*> Elts;
1567 // Break up the constant into its elements.
1568 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
1569 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i)
1570 Elts.push_back(CS->getOperand(i));
1571 } else if (isa<ConstantAggregateZero>(Init)) {
1572 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1573 Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
1574 } else if (isa<UndefValue>(Init)) {
1575 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1576 Elts.push_back(UndefValue::get(STy->getElementType(i)));
1578 assert(0 && "This code is out of sync with "
1579 " ConstantFoldLoadThroughGEPConstantExpr");
1582 // Replace the element that we are supposed to.
1583 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
1584 unsigned Idx = CU->getZExtValue();
1585 assert(Idx < STy->getNumElements() && "Struct index out of range!");
1586 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
1588 // Return the modified struct.
1589 return ConstantStruct::get(Elts);
1591 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
1592 const ArrayType *ATy = cast<ArrayType>(Init->getType());
1594 // Break up the array into elements.
1595 std::vector<Constant*> Elts;
1596 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
1597 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1598 Elts.push_back(CA->getOperand(i));
1599 } else if (isa<ConstantAggregateZero>(Init)) {
1600 Constant *Elt = Constant::getNullValue(ATy->getElementType());
1601 Elts.assign(ATy->getNumElements(), Elt);
1602 } else if (isa<UndefValue>(Init)) {
1603 Constant *Elt = UndefValue::get(ATy->getElementType());
1604 Elts.assign(ATy->getNumElements(), Elt);
1606 assert(0 && "This code is out of sync with "
1607 " ConstantFoldLoadThroughGEPConstantExpr");
1610 assert(CI->getZExtValue() < ATy->getNumElements());
1611 Elts[CI->getZExtValue()] =
1612 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
1613 return ConstantArray::get(ATy, Elts);
1617 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
1618 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
1619 static void CommitValueTo(Constant *Val, Constant *Addr) {
1620 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
1621 assert(GV->hasInitializer());
1622 GV->setInitializer(Val);
1626 ConstantExpr *CE = cast<ConstantExpr>(Addr);
1627 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1629 Constant *Init = GV->getInitializer();
1630 Init = EvaluateStoreInto(Init, Val, CE, 2);
1631 GV->setInitializer(Init);
1634 /// ComputeLoadResult - Return the value that would be computed by a load from
1635 /// P after the stores reflected by 'memory' have been performed. If we can't
1636 /// decide, return null.
1637 static Constant *ComputeLoadResult(Constant *P,
1638 const std::map<Constant*, Constant*> &Memory) {
1639 // If this memory location has been recently stored, use the stored value: it
1640 // is the most up-to-date.
1641 std::map<Constant*, Constant*>::const_iterator I = Memory.find(P);
1642 if (I != Memory.end()) return I->second;
1645 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
1646 if (GV->hasInitializer())
1647 return GV->getInitializer();
1651 // Handle a constantexpr getelementptr.
1652 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
1653 if (CE->getOpcode() == Instruction::GetElementPtr &&
1654 isa<GlobalVariable>(CE->getOperand(0))) {
1655 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1656 if (GV->hasInitializer())
1657 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1660 return 0; // don't know how to evaluate.
1663 /// EvaluateFunction - Evaluate a call to function F, returning true if
1664 /// successful, false if we can't evaluate it. ActualArgs contains the formal
1665 /// arguments for the function.
1666 static bool EvaluateFunction(Function *F, Constant *&RetVal,
1667 const std::vector<Constant*> &ActualArgs,
1668 std::vector<Function*> &CallStack,
1669 std::map<Constant*, Constant*> &MutatedMemory,
1670 std::vector<GlobalVariable*> &AllocaTmps) {
1671 // Check to see if this function is already executing (recursion). If so,
1672 // bail out. TODO: we might want to accept limited recursion.
1673 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
1676 CallStack.push_back(F);
1678 /// Values - As we compute SSA register values, we store their contents here.
1679 std::map<Value*, Constant*> Values;
1681 // Initialize arguments to the incoming values specified.
1683 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
1685 Values[AI] = ActualArgs[ArgNo];
1687 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
1688 /// we can only evaluate any one basic block at most once. This set keeps
1689 /// track of what we have executed so we can detect recursive cases etc.
1690 std::set<BasicBlock*> ExecutedBlocks;
1692 // CurInst - The current instruction we're evaluating.
1693 BasicBlock::iterator CurInst = F->begin()->begin();
1695 // This is the main evaluation loop.
1697 Constant *InstResult = 0;
1699 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
1700 if (SI->isVolatile()) return false; // no volatile accesses.
1701 Constant *Ptr = getVal(Values, SI->getOperand(1));
1702 if (!isSimpleEnoughPointerToCommit(Ptr))
1703 // If this is too complex for us to commit, reject it.
1705 Constant *Val = getVal(Values, SI->getOperand(0));
1706 MutatedMemory[Ptr] = Val;
1707 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
1708 InstResult = ConstantExpr::get(BO->getOpcode(),
1709 getVal(Values, BO->getOperand(0)),
1710 getVal(Values, BO->getOperand(1)));
1711 } else if (ShiftInst *SI = dyn_cast<ShiftInst>(CurInst)) {
1712 InstResult = ConstantExpr::get(SI->getOpcode(),
1713 getVal(Values, SI->getOperand(0)),
1714 getVal(Values, SI->getOperand(1)));
1715 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
1716 InstResult = ConstantExpr::getCast(getVal(Values, CI->getOperand(0)),
1718 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
1719 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
1720 getVal(Values, SI->getOperand(1)),
1721 getVal(Values, SI->getOperand(2)));
1722 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
1723 Constant *P = getVal(Values, GEP->getOperand(0));
1724 std::vector<Constant*> GEPOps;
1725 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
1726 GEPOps.push_back(getVal(Values, GEP->getOperand(i)));
1727 InstResult = ConstantExpr::getGetElementPtr(P, GEPOps);
1728 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
1729 if (LI->isVolatile()) return false; // no volatile accesses.
1730 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
1732 if (InstResult == 0) return false; // Could not evaluate load.
1733 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
1734 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs.
1735 const Type *Ty = AI->getType()->getElementType();
1736 AllocaTmps.push_back(new GlobalVariable(Ty, false,
1737 GlobalValue::InternalLinkage,
1738 UndefValue::get(Ty),
1740 InstResult = AllocaTmps.back();
1741 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
1742 // Cannot handle inline asm.
1743 if (isa<InlineAsm>(CI->getOperand(0))) return false;
1745 // Resolve function pointers.
1746 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0)));
1747 if (!Callee) return false; // Cannot resolve.
1749 std::vector<Constant*> Formals;
1750 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
1751 Formals.push_back(getVal(Values, CI->getOperand(i)));
1753 if (Callee->isExternal()) {
1754 // If this is a function we can constant fold, do it.
1755 if (Constant *C = ConstantFoldCall(Callee, Formals)) {
1761 if (Callee->getFunctionType()->isVarArg())
1766 // Execute the call, if successful, use the return value.
1767 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
1768 MutatedMemory, AllocaTmps))
1770 InstResult = RetVal;
1772 } else if (isa<TerminatorInst>(CurInst)) {
1773 BasicBlock *NewBB = 0;
1774 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
1775 if (BI->isUnconditional()) {
1776 NewBB = BI->getSuccessor(0);
1778 ConstantBool *Cond =
1779 dyn_cast<ConstantBool>(getVal(Values, BI->getCondition()));
1780 if (!Cond) return false; // Cannot determine.
1781 NewBB = BI->getSuccessor(!Cond->getValue());
1783 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
1785 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
1786 if (!Val) return false; // Cannot determine.
1787 NewBB = SI->getSuccessor(SI->findCaseValue(Val));
1788 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
1789 if (RI->getNumOperands())
1790 RetVal = getVal(Values, RI->getOperand(0));
1792 CallStack.pop_back(); // return from fn.
1793 return true; // We succeeded at evaluating this ctor!
1795 // invoke, unwind, unreachable.
1796 return false; // Cannot handle this terminator.
1799 // Okay, we succeeded in evaluating this control flow. See if we have
1800 // executed the new block before. If so, we have a looping function,
1801 // which we cannot evaluate in reasonable time.
1802 if (!ExecutedBlocks.insert(NewBB).second)
1803 return false; // looped!
1805 // Okay, we have never been in this block before. Check to see if there
1806 // are any PHI nodes. If so, evaluate them with information about where
1808 BasicBlock *OldBB = CurInst->getParent();
1809 CurInst = NewBB->begin();
1811 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
1812 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
1814 // Do NOT increment CurInst. We know that the terminator had no value.
1817 // Did not know how to evaluate this!
1821 if (!CurInst->use_empty())
1822 Values[CurInst] = InstResult;
1824 // Advance program counter.
1829 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
1830 /// we can. Return true if we can, false otherwise.
1831 static bool EvaluateStaticConstructor(Function *F) {
1832 /// MutatedMemory - For each store we execute, we update this map. Loads
1833 /// check this to get the most up-to-date value. If evaluation is successful,
1834 /// this state is committed to the process.
1835 std::map<Constant*, Constant*> MutatedMemory;
1837 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
1838 /// to represent its body. This vector is needed so we can delete the
1839 /// temporary globals when we are done.
1840 std::vector<GlobalVariable*> AllocaTmps;
1842 /// CallStack - This is used to detect recursion. In pathological situations
1843 /// we could hit exponential behavior, but at least there is nothing
1845 std::vector<Function*> CallStack;
1847 // Call the function.
1848 Constant *RetValDummy;
1849 bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(),
1850 CallStack, MutatedMemory, AllocaTmps);
1852 // We succeeded at evaluation: commit the result.
1853 DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
1854 << F->getName() << "' to " << MutatedMemory.size()
1856 for (std::map<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
1857 E = MutatedMemory.end(); I != E; ++I)
1858 CommitValueTo(I->second, I->first);
1861 // At this point, we are done interpreting. If we created any 'alloca'
1862 // temporaries, release them now.
1863 while (!AllocaTmps.empty()) {
1864 GlobalVariable *Tmp = AllocaTmps.back();
1865 AllocaTmps.pop_back();
1867 // If there are still users of the alloca, the program is doing something
1868 // silly, e.g. storing the address of the alloca somewhere and using it
1869 // later. Since this is undefined, we'll just make it be null.
1870 if (!Tmp->use_empty())
1871 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
1880 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
1881 /// Return true if anything changed.
1882 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
1883 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
1884 bool MadeChange = false;
1885 if (Ctors.empty()) return false;
1887 // Loop over global ctors, optimizing them when we can.
1888 for (unsigned i = 0; i != Ctors.size(); ++i) {
1889 Function *F = Ctors[i];
1890 // Found a null terminator in the middle of the list, prune off the rest of
1893 if (i != Ctors.size()-1) {
1900 // We cannot simplify external ctor functions.
1901 if (F->empty()) continue;
1903 // If we can evaluate the ctor at compile time, do.
1904 if (EvaluateStaticConstructor(F)) {
1905 Ctors.erase(Ctors.begin()+i);
1908 ++NumCtorsEvaluated;
1913 if (!MadeChange) return false;
1915 GCL = InstallGlobalCtors(GCL, Ctors);
1920 bool GlobalOpt::runOnModule(Module &M) {
1921 bool Changed = false;
1923 // Try to find the llvm.globalctors list.
1924 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
1926 bool LocalChange = true;
1927 while (LocalChange) {
1928 LocalChange = false;
1930 // Delete functions that are trivially dead, ccc -> fastcc
1931 LocalChange |= OptimizeFunctions(M);
1933 // Optimize global_ctors list.
1935 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
1937 // Optimize non-address-taken globals.
1938 LocalChange |= OptimizeGlobalVars(M);
1939 Changed |= LocalChange;
1942 // TODO: Move all global ctors functions to the end of the module for code