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/Analysis/ConstantFolding.h"
26 #include "llvm/Target/TargetData.h"
27 #include "llvm/Support/Compiler.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/ADT/StringExtras.h"
36 STATISTIC(NumMarked , "Number of globals marked constant");
37 STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
38 STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
39 STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
40 STATISTIC(NumDeleted , "Number of globals deleted");
41 STATISTIC(NumFnDeleted , "Number of functions deleted");
42 STATISTIC(NumGlobUses , "Number of global uses devirtualized");
43 STATISTIC(NumLocalized , "Number of globals localized");
44 STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans");
45 STATISTIC(NumFastCallFns , "Number of functions converted to fastcc");
46 STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
49 struct VISIBILITY_HIDDEN GlobalOpt : public ModulePass {
50 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
51 AU.addRequired<TargetData>();
54 bool runOnModule(Module &M);
57 GlobalVariable *FindGlobalCtors(Module &M);
58 bool OptimizeFunctions(Module &M);
59 bool OptimizeGlobalVars(Module &M);
60 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
61 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
64 RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer");
67 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
69 /// GlobalStatus - As we analyze each global, keep track of some information
70 /// about it. If we find out that the address of the global is taken, none of
71 /// this info will be accurate.
72 struct VISIBILITY_HIDDEN GlobalStatus {
73 /// isLoaded - True if the global is ever loaded. If the global isn't ever
74 /// loaded it can be deleted.
77 /// StoredType - Keep track of what stores to the global look like.
80 /// NotStored - There is no store to this global. It can thus be marked
84 /// isInitializerStored - This global is stored to, but the only thing
85 /// stored is the constant it was initialized with. This is only tracked
86 /// for scalar globals.
89 /// isStoredOnce - This global is stored to, but only its initializer and
90 /// one other value is ever stored to it. If this global isStoredOnce, we
91 /// track the value stored to it in StoredOnceValue below. This is only
92 /// tracked for scalar globals.
95 /// isStored - This global is stored to by multiple values or something else
96 /// that we cannot track.
100 /// StoredOnceValue - If only one value (besides the initializer constant) is
101 /// ever stored to this global, keep track of what value it is.
102 Value *StoredOnceValue;
104 /// AccessingFunction/HasMultipleAccessingFunctions - These start out
105 /// null/false. When the first accessing function is noticed, it is recorded.
106 /// When a second different accessing function is noticed,
107 /// HasMultipleAccessingFunctions is set to true.
108 Function *AccessingFunction;
109 bool HasMultipleAccessingFunctions;
111 /// HasNonInstructionUser - Set to true if this global has a user that is not
112 /// an instruction (e.g. a constant expr or GV initializer).
113 bool HasNonInstructionUser;
115 /// HasPHIUser - Set to true if this global has a user that is a PHI node.
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), HasPHIUser(false),
126 isNotSuitableForSRA(false) {}
131 /// ConstantIsDead - Return true if the specified constant is (transitively)
132 /// dead. The constant may be used by other constants (e.g. constant arrays and
133 /// constant exprs) as long as they are dead, but it cannot be used by anything
135 static bool ConstantIsDead(Constant *C) {
136 if (isa<GlobalValue>(C)) return false;
138 for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI)
139 if (Constant *CU = dyn_cast<Constant>(*UI)) {
140 if (!ConstantIsDead(CU)) return false;
147 /// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
148 /// structure. If the global has its address taken, return true to indicate we
149 /// can't do anything with it.
151 static bool AnalyzeGlobal(Value *V, GlobalStatus &GS,
152 std::set<PHINode*> &PHIUsers) {
153 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
154 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
155 GS.HasNonInstructionUser = true;
157 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
158 if (CE->getOpcode() != Instruction::GetElementPtr)
159 GS.isNotSuitableForSRA = true;
160 else if (!GS.isNotSuitableForSRA) {
161 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
162 // don't like < 3 operand CE's, and we don't like non-constant integer
164 if (CE->getNumOperands() < 3 || !CE->getOperand(1)->isNullValue())
165 GS.isNotSuitableForSRA = true;
167 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
168 if (!isa<ConstantInt>(CE->getOperand(i))) {
169 GS.isNotSuitableForSRA = true;
175 } else if (Instruction *I = dyn_cast<Instruction>(*UI)) {
176 if (!GS.HasMultipleAccessingFunctions) {
177 Function *F = I->getParent()->getParent();
178 if (GS.AccessingFunction == 0)
179 GS.AccessingFunction = F;
180 else if (GS.AccessingFunction != F)
181 GS.HasMultipleAccessingFunctions = true;
183 if (isa<LoadInst>(I)) {
185 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
186 // Don't allow a store OF the address, only stores TO the address.
187 if (SI->getOperand(0) == V) return true;
189 // If this is a direct store to the global (i.e., the global is a scalar
190 // value, not an aggregate), keep more specific information about
192 if (GS.StoredType != GlobalStatus::isStored)
193 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){
194 Value *StoredVal = SI->getOperand(0);
195 if (StoredVal == GV->getInitializer()) {
196 if (GS.StoredType < GlobalStatus::isInitializerStored)
197 GS.StoredType = GlobalStatus::isInitializerStored;
198 } else if (isa<LoadInst>(StoredVal) &&
199 cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
201 if (GS.StoredType < GlobalStatus::isInitializerStored)
202 GS.StoredType = GlobalStatus::isInitializerStored;
203 } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
204 GS.StoredType = GlobalStatus::isStoredOnce;
205 GS.StoredOnceValue = StoredVal;
206 } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
207 GS.StoredOnceValue == StoredVal) {
210 GS.StoredType = GlobalStatus::isStored;
213 GS.StoredType = GlobalStatus::isStored;
215 } else if (isa<GetElementPtrInst>(I)) {
216 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
218 // If the first two indices are constants, this can be SRA'd.
219 if (isa<GlobalVariable>(I->getOperand(0))) {
220 if (I->getNumOperands() < 3 || !isa<Constant>(I->getOperand(1)) ||
221 !cast<Constant>(I->getOperand(1))->isNullValue() ||
222 !isa<ConstantInt>(I->getOperand(2)))
223 GS.isNotSuitableForSRA = true;
224 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(I->getOperand(0))){
225 if (CE->getOpcode() != Instruction::GetElementPtr ||
226 CE->getNumOperands() < 3 || I->getNumOperands() < 2 ||
227 !isa<Constant>(I->getOperand(0)) ||
228 !cast<Constant>(I->getOperand(0))->isNullValue())
229 GS.isNotSuitableForSRA = true;
231 GS.isNotSuitableForSRA = true;
233 } else if (isa<SelectInst>(I)) {
234 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
235 GS.isNotSuitableForSRA = true;
236 } else if (PHINode *PN = dyn_cast<PHINode>(I)) {
237 // PHI nodes we can check just like select or GEP instructions, but we
238 // have to be careful about infinite recursion.
239 if (PHIUsers.insert(PN).second) // Not already visited.
240 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
241 GS.isNotSuitableForSRA = true;
242 GS.HasPHIUser = true;
243 } else if (isa<CmpInst>(I)) {
244 GS.isNotSuitableForSRA = true;
245 } else if (isa<MemCpyInst>(I) || isa<MemMoveInst>(I)) {
246 if (I->getOperand(1) == V)
247 GS.StoredType = GlobalStatus::isStored;
248 if (I->getOperand(2) == V)
250 GS.isNotSuitableForSRA = true;
251 } else if (isa<MemSetInst>(I)) {
252 assert(I->getOperand(1) == V && "Memset only takes one pointer!");
253 GS.StoredType = GlobalStatus::isStored;
254 GS.isNotSuitableForSRA = true;
256 return true; // Any other non-load instruction might take address!
258 } else if (Constant *C = dyn_cast<Constant>(*UI)) {
259 GS.HasNonInstructionUser = true;
260 // We might have a dead and dangling constant hanging off of here.
261 if (!ConstantIsDead(C))
264 GS.HasNonInstructionUser = true;
265 // Otherwise must be some other user.
272 static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) {
273 ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
275 unsigned IdxV = CI->getZExtValue();
277 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
278 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
279 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
280 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
281 } else if (ConstantPacked *CP = dyn_cast<ConstantPacked>(Agg)) {
282 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
283 } else if (isa<ConstantAggregateZero>(Agg)) {
284 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
285 if (IdxV < STy->getNumElements())
286 return Constant::getNullValue(STy->getElementType(IdxV));
287 } else if (const SequentialType *STy =
288 dyn_cast<SequentialType>(Agg->getType())) {
289 return Constant::getNullValue(STy->getElementType());
291 } else if (isa<UndefValue>(Agg)) {
292 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
293 if (IdxV < STy->getNumElements())
294 return UndefValue::get(STy->getElementType(IdxV));
295 } else if (const SequentialType *STy =
296 dyn_cast<SequentialType>(Agg->getType())) {
297 return UndefValue::get(STy->getElementType());
304 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
305 /// users of the global, cleaning up the obvious ones. This is largely just a
306 /// quick scan over the use list to clean up the easy and obvious cruft. This
307 /// returns true if it made a change.
308 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
309 bool Changed = false;
310 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
313 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
315 // Replace the load with the initializer.
316 LI->replaceAllUsesWith(Init);
317 LI->eraseFromParent();
320 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
321 // Store must be unreachable or storing Init into the global.
322 SI->eraseFromParent();
324 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
325 if (CE->getOpcode() == Instruction::GetElementPtr) {
326 Constant *SubInit = 0;
328 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
329 Changed |= CleanupConstantGlobalUsers(CE, SubInit);
330 } else if (CE->getOpcode() == Instruction::BitCast &&
331 isa<PointerType>(CE->getType())) {
332 // Pointer cast, delete any stores and memsets to the global.
333 Changed |= CleanupConstantGlobalUsers(CE, 0);
336 if (CE->use_empty()) {
337 CE->destroyConstant();
340 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
341 Constant *SubInit = 0;
343 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
344 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
345 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
346 Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
348 if (GEP->use_empty()) {
349 GEP->eraseFromParent();
352 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
353 if (MI->getRawDest() == V) {
354 MI->eraseFromParent();
358 } else if (Constant *C = dyn_cast<Constant>(U)) {
359 // If we have a chain of dead constantexprs or other things dangling from
360 // us, and if they are all dead, nuke them without remorse.
361 if (ConstantIsDead(C)) {
362 C->destroyConstant();
363 // This could have invalidated UI, start over from scratch.
364 CleanupConstantGlobalUsers(V, Init);
372 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
373 /// variable. This opens the door for other optimizations by exposing the
374 /// behavior of the program in a more fine-grained way. We have determined that
375 /// this transformation is safe already. We return the first global variable we
376 /// insert so that the caller can reprocess it.
377 static GlobalVariable *SRAGlobal(GlobalVariable *GV) {
378 assert(GV->hasInternalLinkage() && !GV->isConstant());
379 Constant *Init = GV->getInitializer();
380 const Type *Ty = Init->getType();
382 std::vector<GlobalVariable*> NewGlobals;
383 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
385 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
386 NewGlobals.reserve(STy->getNumElements());
387 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
388 Constant *In = getAggregateConstantElement(Init,
389 ConstantInt::get(Type::Int32Ty, i));
390 assert(In && "Couldn't get element of initializer?");
391 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
392 GlobalVariable::InternalLinkage,
393 In, GV->getName()+"."+utostr(i));
394 Globals.insert(GV, NGV);
395 NewGlobals.push_back(NGV);
397 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
398 unsigned NumElements = 0;
399 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
400 NumElements = ATy->getNumElements();
401 else if (const PackedType *PTy = dyn_cast<PackedType>(STy))
402 NumElements = PTy->getNumElements();
404 assert(0 && "Unknown aggregate sequential type!");
406 if (NumElements > 16 && GV->hasNUsesOrMore(16))
407 return 0; // It's not worth it.
408 NewGlobals.reserve(NumElements);
409 for (unsigned i = 0, e = NumElements; i != e; ++i) {
410 Constant *In = getAggregateConstantElement(Init,
411 ConstantInt::get(Type::Int32Ty, i));
412 assert(In && "Couldn't get element of initializer?");
414 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
415 GlobalVariable::InternalLinkage,
416 In, GV->getName()+"."+utostr(i));
417 Globals.insert(GV, NGV);
418 NewGlobals.push_back(NGV);
422 if (NewGlobals.empty())
425 DOUT << "PERFORMING GLOBAL SRA ON: " << *GV;
427 Constant *NullInt = Constant::getNullValue(Type::Int32Ty);
429 // Loop over all of the uses of the global, replacing the constantexpr geps,
430 // with smaller constantexpr geps or direct references.
431 while (!GV->use_empty()) {
432 User *GEP = GV->use_back();
433 assert(((isa<ConstantExpr>(GEP) &&
434 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
435 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
437 // Ignore the 1th operand, which has to be zero or else the program is quite
438 // broken (undefined). Get the 2nd operand, which is the structure or array
440 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
441 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
443 Value *NewPtr = NewGlobals[Val];
445 // Form a shorter GEP if needed.
446 if (GEP->getNumOperands() > 3)
447 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
448 SmallVector<Constant*, 8> Idxs;
449 Idxs.push_back(NullInt);
450 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
451 Idxs.push_back(CE->getOperand(i));
452 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr),
453 &Idxs[0], Idxs.size());
455 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
456 SmallVector<Value*, 8> Idxs;
457 Idxs.push_back(NullInt);
458 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
459 Idxs.push_back(GEPI->getOperand(i));
460 NewPtr = new GetElementPtrInst(NewPtr, &Idxs[0], Idxs.size(),
461 GEPI->getName()+"."+utostr(Val), GEPI);
463 GEP->replaceAllUsesWith(NewPtr);
465 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
466 GEPI->eraseFromParent();
468 cast<ConstantExpr>(GEP)->destroyConstant();
471 // Delete the old global, now that it is dead.
475 // Loop over the new globals array deleting any globals that are obviously
476 // dead. This can arise due to scalarization of a structure or an array that
477 // has elements that are dead.
478 unsigned FirstGlobal = 0;
479 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
480 if (NewGlobals[i]->use_empty()) {
481 Globals.erase(NewGlobals[i]);
482 if (FirstGlobal == i) ++FirstGlobal;
485 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
488 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
489 /// value will trap if the value is dynamically null.
490 static bool AllUsesOfValueWillTrapIfNull(Value *V) {
491 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
492 if (isa<LoadInst>(*UI)) {
494 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
495 if (SI->getOperand(0) == V) {
496 //cerr << "NONTRAPPING USE: " << **UI;
497 return false; // Storing the value.
499 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
500 if (CI->getOperand(0) != V) {
501 //cerr << "NONTRAPPING USE: " << **UI;
502 return false; // Not calling the ptr
504 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
505 if (II->getOperand(0) != V) {
506 //cerr << "NONTRAPPING USE: " << **UI;
507 return false; // Not calling the ptr
509 } else if (CastInst *CI = dyn_cast<CastInst>(*UI)) {
510 if (!AllUsesOfValueWillTrapIfNull(CI)) return false;
511 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
512 if (!AllUsesOfValueWillTrapIfNull(GEPI)) return false;
513 } else if (isa<ICmpInst>(*UI) &&
514 isa<ConstantPointerNull>(UI->getOperand(1))) {
515 // Ignore setcc X, null
517 //cerr << "NONTRAPPING USE: " << **UI;
523 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
524 /// from GV will trap if the loaded value is null. Note that this also permits
525 /// comparisons of the loaded value against null, as a special case.
526 static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) {
527 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI)
528 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
529 if (!AllUsesOfValueWillTrapIfNull(LI))
531 } else if (isa<StoreInst>(*UI)) {
532 // Ignore stores to the global.
534 // We don't know or understand this user, bail out.
535 //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI;
542 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
543 bool Changed = false;
544 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
545 Instruction *I = cast<Instruction>(*UI++);
546 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
547 LI->setOperand(0, NewV);
549 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
550 if (SI->getOperand(1) == V) {
551 SI->setOperand(1, NewV);
554 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
555 if (I->getOperand(0) == V) {
556 // Calling through the pointer! Turn into a direct call, but be careful
557 // that the pointer is not also being passed as an argument.
558 I->setOperand(0, NewV);
560 bool PassedAsArg = false;
561 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
562 if (I->getOperand(i) == V) {
564 I->setOperand(i, NewV);
568 // Being passed as an argument also. Be careful to not invalidate UI!
572 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
573 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
574 ConstantExpr::getCast(CI->getOpcode(),
575 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 SmallVector<Constant*, 8> Idxs;
583 Idxs.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)))
589 if (Idxs.size() == GEPI->getNumOperands()-1)
590 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
591 ConstantExpr::getGetElementPtr(NewV, &Idxs[0],
593 if (GEPI->use_empty()) {
595 GEPI->eraseFromParent();
604 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
605 /// value stored into it. If there are uses of the loaded value that would trap
606 /// if the loaded value is dynamically null, then we know that they cannot be
607 /// reachable with a null optimize away the load.
608 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) {
609 std::vector<LoadInst*> Loads;
610 bool Changed = false;
612 // Replace all uses of loads with uses of uses of the stored value.
613 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end();
615 if (LoadInst *LI = dyn_cast<LoadInst>(*GUI)) {
617 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
619 assert(isa<StoreInst>(*GUI) && "Only expect load and stores!");
623 DOUT << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV;
627 // Delete all of the loads we can, keeping track of whether we nuked them all!
628 bool AllLoadsGone = true;
629 while (!Loads.empty()) {
630 LoadInst *L = Loads.back();
631 if (L->use_empty()) {
632 L->eraseFromParent();
635 AllLoadsGone = false;
640 // If we nuked all of the loads, then none of the stores are needed either,
641 // nor is the global.
643 DOUT << " *** GLOBAL NOW DEAD!\n";
644 CleanupConstantGlobalUsers(GV, 0);
645 if (GV->use_empty()) {
646 GV->eraseFromParent();
654 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
655 /// instructions that are foldable.
656 static void ConstantPropUsersOf(Value *V) {
657 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
658 if (Instruction *I = dyn_cast<Instruction>(*UI++))
659 if (Constant *NewC = ConstantFoldInstruction(I)) {
660 I->replaceAllUsesWith(NewC);
662 // Advance UI to the next non-I use to avoid invalidating it!
663 // Instructions could multiply use V.
664 while (UI != E && *UI == I)
666 I->eraseFromParent();
670 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
671 /// variable, and transforms the program as if it always contained the result of
672 /// the specified malloc. Because it is always the result of the specified
673 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
674 /// malloc into a global, and any loads of GV as uses of the new global.
675 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
677 DOUT << "PROMOTING MALLOC GLOBAL: " << *GV << " MALLOC = " << *MI;
678 ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize());
680 if (NElements->getZExtValue() != 1) {
681 // If we have an array allocation, transform it to a single element
682 // allocation to make the code below simpler.
683 Type *NewTy = ArrayType::get(MI->getAllocatedType(),
684 NElements->getZExtValue());
686 new MallocInst(NewTy, Constant::getNullValue(Type::Int32Ty),
687 MI->getAlignment(), MI->getName(), MI);
689 Indices[0] = Indices[1] = Constant::getNullValue(Type::Int32Ty);
690 Value *NewGEP = new GetElementPtrInst(NewMI, Indices, 2,
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::getBitCast(RepValue,
711 GV->getType()->getElementType());
713 // If there is a comparison against null, we will insert a global bool to
714 // keep track of whether the global was initialized yet or not.
715 GlobalVariable *InitBool =
716 new GlobalVariable(Type::Int1Ty, false, GlobalValue::InternalLinkage,
717 ConstantInt::getFalse(), GV->getName()+".init");
718 bool InitBoolUsed = false;
720 // Loop over all uses of GV, processing them in turn.
721 std::vector<StoreInst*> Stores;
722 while (!GV->use_empty())
723 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
724 while (!LI->use_empty()) {
725 Use &LoadUse = LI->use_begin().getUse();
726 if (!isa<ICmpInst>(LoadUse.getUser()))
729 ICmpInst *CI = cast<ICmpInst>(LoadUse.getUser());
730 // Replace the cmp X, 0 with a use of the bool value.
731 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", CI);
733 switch (CI->getPredicate()) {
734 default: assert(0 && "Unknown ICmp Predicate!");
735 case ICmpInst::ICMP_ULT:
736 case ICmpInst::ICMP_SLT:
737 LV = ConstantInt::getFalse(); // X < null -> always false
739 case ICmpInst::ICMP_ULE:
740 case ICmpInst::ICMP_SLE:
741 case ICmpInst::ICMP_EQ:
742 LV = BinaryOperator::createNot(LV, "notinit", CI);
744 case ICmpInst::ICMP_NE:
745 case ICmpInst::ICMP_UGE:
746 case ICmpInst::ICMP_SGE:
747 case ICmpInst::ICMP_UGT:
748 case ICmpInst::ICMP_SGT:
751 CI->replaceAllUsesWith(LV);
752 CI->eraseFromParent();
755 LI->eraseFromParent();
757 StoreInst *SI = cast<StoreInst>(GV->use_back());
758 // The global is initialized when the store to it occurs.
759 new StoreInst(ConstantInt::getTrue(), InitBool, SI);
760 SI->eraseFromParent();
763 // If the initialization boolean was used, insert it, otherwise delete it.
765 while (!InitBool->use_empty()) // Delete initializations
766 cast<Instruction>(InitBool->use_back())->eraseFromParent();
769 GV->getParent()->getGlobalList().insert(GV, InitBool);
772 // Now the GV is dead, nuke it and the malloc.
773 GV->eraseFromParent();
774 MI->eraseFromParent();
776 // To further other optimizations, loop over all users of NewGV and try to
777 // constant prop them. This will promote GEP instructions with constant
778 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
779 ConstantPropUsersOf(NewGV);
780 if (RepValue != NewGV)
781 ConstantPropUsersOf(RepValue);
786 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
787 /// to make sure that there are no complex uses of V. We permit simple things
788 /// like dereferencing the pointer, but not storing through the address, unless
789 /// it is to the specified global.
790 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V,
791 GlobalVariable *GV) {
792 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI)
793 if (isa<LoadInst>(*UI) || isa<CmpInst>(*UI)) {
795 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
796 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
797 return false; // Storing the pointer itself... bad.
798 // Otherwise, storing through it, or storing into GV... fine.
799 } else if (isa<GetElementPtrInst>(*UI) || isa<SelectInst>(*UI)) {
800 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(cast<Instruction>(*UI),GV))
808 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
809 /// somewhere. Transform all uses of the allocation into loads from the
810 /// global and uses of the resultant pointer. Further, delete the store into
811 /// GV. This assumes that these value pass the
812 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
813 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
814 GlobalVariable *GV) {
815 while (!Alloc->use_empty()) {
816 Instruction *U = Alloc->use_back();
817 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
818 // If this is the store of the allocation into the global, remove it.
819 if (SI->getOperand(1) == GV) {
820 SI->eraseFromParent();
825 // Insert a load from the global, and use it instead of the malloc.
826 Value *NL = new LoadInst(GV, GV->getName()+".val", U);
827 U->replaceUsesOfWith(Alloc, NL);
831 /// GlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
832 /// GV are simple enough to perform HeapSRA, return true.
833 static bool GlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV) {
834 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;
836 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
837 // We permit two users of the load: setcc comparing against the null
838 // pointer, and a getelementptr of a specific form.
839 for (Value::use_iterator UI = LI->use_begin(), E = LI->use_end(); UI != E;
841 // Comparison against null is ok.
842 if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) {
843 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
848 // getelementptr is also ok, but only a simple form.
849 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI);
850 if (!GEPI) return false;
852 // Must index into the array and into the struct.
853 if (GEPI->getNumOperands() < 3)
856 // Otherwise the GEP is ok.
863 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
864 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
865 /// use FieldGlobals instead. All uses of loaded values satisfy
866 /// GlobalLoadUsesSimpleEnoughForHeapSRA.
867 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Ptr,
868 const std::vector<GlobalVariable*> &FieldGlobals) {
869 std::vector<Value *> InsertedLoadsForPtr;
870 //InsertedLoadsForPtr.resize(FieldGlobals.size());
871 while (!Ptr->use_empty()) {
872 Instruction *User = Ptr->use_back();
874 // If this is a comparison against null, handle it.
875 if (ICmpInst *SCI = dyn_cast<ICmpInst>(User)) {
876 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
877 // If we have a setcc of the loaded pointer, we can use a setcc of any
880 if (InsertedLoadsForPtr.empty()) {
881 NPtr = new LoadInst(FieldGlobals[0], Ptr->getName()+".f0", Ptr);
882 InsertedLoadsForPtr.push_back(Ptr);
884 NPtr = InsertedLoadsForPtr.back();
887 Value *New = new ICmpInst(SCI->getPredicate(), NPtr,
888 Constant::getNullValue(NPtr->getType()),
889 SCI->getName(), SCI);
890 SCI->replaceAllUsesWith(New);
891 SCI->eraseFromParent();
895 // Otherwise, this should be: 'getelementptr Ptr, Idx, uint FieldNo ...'
896 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
897 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
898 && "Unexpected GEPI!");
900 // Load the pointer for this field.
901 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
902 if (InsertedLoadsForPtr.size() <= FieldNo)
903 InsertedLoadsForPtr.resize(FieldNo+1);
904 if (InsertedLoadsForPtr[FieldNo] == 0)
905 InsertedLoadsForPtr[FieldNo] = new LoadInst(FieldGlobals[FieldNo],
906 Ptr->getName()+".f" +
907 utostr(FieldNo), Ptr);
908 Value *NewPtr = InsertedLoadsForPtr[FieldNo];
910 // Create the new GEP idx vector.
911 std::vector<Value*> GEPIdx;
912 GEPIdx.push_back(GEPI->getOperand(1));
913 GEPIdx.insert(GEPIdx.end(), GEPI->op_begin()+3, GEPI->op_end());
915 Value *NGEPI = new GetElementPtrInst(NewPtr, GEPIdx, GEPI->getName(), GEPI);
916 GEPI->replaceAllUsesWith(NGEPI);
917 GEPI->eraseFromParent();
921 /// PerformHeapAllocSRoA - MI is an allocation of an array of structures. Break
922 /// it up into multiple allocations of arrays of the fields.
923 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){
924 DOUT << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *MI;
925 const StructType *STy = cast<StructType>(MI->getAllocatedType());
927 // There is guaranteed to be at least one use of the malloc (storing
928 // it into GV). If there are other uses, change them to be uses of
929 // the global to simplify later code. This also deletes the store
931 ReplaceUsesOfMallocWithGlobal(MI, GV);
933 // Okay, at this point, there are no users of the malloc. Insert N
934 // new mallocs at the same place as MI, and N globals.
935 std::vector<GlobalVariable*> FieldGlobals;
936 std::vector<MallocInst*> FieldMallocs;
938 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
939 const Type *FieldTy = STy->getElementType(FieldNo);
940 const Type *PFieldTy = PointerType::get(FieldTy);
942 GlobalVariable *NGV =
943 new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage,
944 Constant::getNullValue(PFieldTy),
945 GV->getName() + ".f" + utostr(FieldNo), GV);
946 FieldGlobals.push_back(NGV);
948 MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(),
949 MI->getName() + ".f" + utostr(FieldNo),MI);
950 FieldMallocs.push_back(NMI);
951 new StoreInst(NMI, NGV, MI);
954 // The tricky aspect of this transformation is handling the case when malloc
955 // fails. In the original code, malloc failing would set the result pointer
956 // of malloc to null. In this case, some mallocs could succeed and others
957 // could fail. As such, we emit code that looks like this:
958 // F0 = malloc(field0)
959 // F1 = malloc(field1)
960 // F2 = malloc(field2)
961 // if (F0 == 0 || F1 == 0 || F2 == 0) {
962 // if (F0) { free(F0); F0 = 0; }
963 // if (F1) { free(F1); F1 = 0; }
964 // if (F2) { free(F2); F2 = 0; }
966 Value *RunningOr = 0;
967 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
968 Value *Cond = new ICmpInst(ICmpInst::ICMP_EQ, FieldMallocs[i],
969 Constant::getNullValue(FieldMallocs[i]->getType()),
972 RunningOr = Cond; // First seteq
974 RunningOr = BinaryOperator::createOr(RunningOr, Cond, "tmp", MI);
977 // Split the basic block at the old malloc.
978 BasicBlock *OrigBB = MI->getParent();
979 BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont");
981 // Create the block to check the first condition. Put all these blocks at the
982 // end of the function as they are unlikely to be executed.
983 BasicBlock *NullPtrBlock = new BasicBlock("malloc_ret_null",
984 OrigBB->getParent());
986 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
987 // branch on RunningOr.
988 OrigBB->getTerminator()->eraseFromParent();
989 new BranchInst(NullPtrBlock, ContBB, RunningOr, OrigBB);
991 // Within the NullPtrBlock, we need to emit a comparison and branch for each
992 // pointer, because some may be null while others are not.
993 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
994 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
995 Value *Cmp = new ICmpInst(ICmpInst::ICMP_NE, GVVal,
996 Constant::getNullValue(GVVal->getType()),
997 "tmp", NullPtrBlock);
998 BasicBlock *FreeBlock = new BasicBlock("free_it", OrigBB->getParent());
999 BasicBlock *NextBlock = new BasicBlock("next", OrigBB->getParent());
1000 new BranchInst(FreeBlock, NextBlock, Cmp, NullPtrBlock);
1002 // Fill in FreeBlock.
1003 new FreeInst(GVVal, FreeBlock);
1004 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1006 new BranchInst(NextBlock, FreeBlock);
1008 NullPtrBlock = NextBlock;
1011 new BranchInst(ContBB, NullPtrBlock);
1014 // MI is no longer needed, remove it.
1015 MI->eraseFromParent();
1018 // Okay, the malloc site is completely handled. All of the uses of GV are now
1019 // loads, and all uses of those loads are simple. Rewrite them to use loads
1020 // of the per-field globals instead.
1021 while (!GV->use_empty()) {
1022 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
1023 RewriteUsesOfLoadForHeapSRoA(LI, FieldGlobals);
1024 LI->eraseFromParent();
1026 // Must be a store of null.
1027 StoreInst *SI = cast<StoreInst>(GV->use_back());
1028 assert(isa<Constant>(SI->getOperand(0)) &&
1029 cast<Constant>(SI->getOperand(0))->isNullValue() &&
1030 "Unexpected heap-sra user!");
1032 // Insert a store of null into each global.
1033 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1035 Constant::getNullValue(FieldGlobals[i]->getType()->getElementType());
1036 new StoreInst(Null, FieldGlobals[i], SI);
1038 // Erase the original store.
1039 SI->eraseFromParent();
1043 // The old global is now dead, remove it.
1044 GV->eraseFromParent();
1047 return FieldGlobals[0];
1051 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1052 // that only one value (besides its initializer) is ever stored to the global.
1053 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1054 Module::global_iterator &GVI,
1056 if (CastInst *CI = dyn_cast<CastInst>(StoredOnceVal))
1057 StoredOnceVal = CI->getOperand(0);
1058 else if (GetElementPtrInst *GEPI =dyn_cast<GetElementPtrInst>(StoredOnceVal)){
1059 // "getelementptr Ptr, 0, 0, 0" is really just a cast.
1060 bool IsJustACast = true;
1061 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
1062 if (!isa<Constant>(GEPI->getOperand(i)) ||
1063 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
1064 IsJustACast = false;
1068 StoredOnceVal = GEPI->getOperand(0);
1071 // If we are dealing with a pointer global that is initialized to null and
1072 // only has one (non-null) value stored into it, then we can optimize any
1073 // users of the loaded value (often calls and loads) that would trap if the
1075 if (isa<PointerType>(GV->getInitializer()->getType()) &&
1076 GV->getInitializer()->isNullValue()) {
1077 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1078 if (GV->getInitializer()->getType() != SOVC->getType())
1079 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1081 // Optimize away any trapping uses of the loaded value.
1082 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1084 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) {
1085 // If this is a malloc of an abstract type, don't touch it.
1086 if (!MI->getAllocatedType()->isSized())
1089 // We can't optimize this global unless all uses of it are *known* to be
1090 // of the malloc value, not of the null initializer value (consider a use
1091 // that compares the global's value against zero to see if the malloc has
1092 // been reached). To do this, we check to see if all uses of the global
1093 // would trap if the global were null: this proves that they must all
1094 // happen after the malloc.
1095 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1098 // We can't optimize this if the malloc itself is used in a complex way,
1099 // for example, being stored into multiple globals. This allows the
1100 // malloc to be stored into the specified global, loaded setcc'd, and
1101 // GEP'd. These are all things we could transform to using the global
1103 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV))
1107 // If we have a global that is only initialized with a fixed size malloc,
1108 // transform the program to use global memory instead of malloc'd memory.
1109 // This eliminates dynamic allocation, avoids an indirection accessing the
1110 // data, and exposes the resultant global to further GlobalOpt.
1111 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) {
1112 // Restrict this transformation to only working on small allocations
1113 // (2048 bytes currently), as we don't want to introduce a 16M global or
1115 if (NElements->getZExtValue()*
1116 TD.getTypeSize(MI->getAllocatedType()) < 2048) {
1117 GVI = OptimizeGlobalAddressOfMalloc(GV, MI);
1122 // If the allocation is an array of structures, consider transforming this
1123 // into multiple malloc'd arrays, one for each field. This is basically
1124 // SRoA for malloc'd memory.
1125 if (const StructType *AllocTy =
1126 dyn_cast<StructType>(MI->getAllocatedType())) {
1127 // This the structure has an unreasonable number of fields, leave it
1129 if (AllocTy->getNumElements() <= 16 && AllocTy->getNumElements() > 0 &&
1130 GlobalLoadUsesSimpleEnoughForHeapSRA(GV)) {
1131 GVI = PerformHeapAllocSRoA(GV, MI);
1141 /// ShrinkGlobalToBoolean - At this point, we have learned that the only two
1142 /// values ever stored into GV are its initializer and OtherVal.
1143 static void ShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1144 // Create the new global, initializing it to false.
1145 GlobalVariable *NewGV = new GlobalVariable(Type::Int1Ty, false,
1146 GlobalValue::InternalLinkage, ConstantInt::getFalse(),
1147 GV->getName()+".b");
1148 GV->getParent()->getGlobalList().insert(GV, NewGV);
1150 Constant *InitVal = GV->getInitializer();
1151 assert(InitVal->getType() != Type::Int1Ty && "No reason to shrink to bool!");
1153 // If initialized to zero and storing one into the global, we can use a cast
1154 // instead of a select to synthesize the desired value.
1155 bool IsOneZero = false;
1156 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1157 IsOneZero = InitVal->isNullValue() && CI->equalsInt(1);
1159 while (!GV->use_empty()) {
1160 Instruction *UI = cast<Instruction>(GV->use_back());
1161 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1162 // Change the store into a boolean store.
1163 bool StoringOther = SI->getOperand(0) == OtherVal;
1164 // Only do this if we weren't storing a loaded value.
1166 if (StoringOther || SI->getOperand(0) == InitVal)
1167 StoreVal = ConstantInt::get(Type::Int1Ty, StoringOther);
1169 // Otherwise, we are storing a previously loaded copy. To do this,
1170 // change the copy from copying the original value to just copying the
1172 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1174 // If we're already replaced the input, StoredVal will be a cast or
1175 // select instruction. If not, it will be a load of the original
1177 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1178 assert(LI->getOperand(0) == GV && "Not a copy!");
1179 // Insert a new load, to preserve the saved value.
1180 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1182 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1183 "This is not a form that we understand!");
1184 StoreVal = StoredVal->getOperand(0);
1185 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1188 new StoreInst(StoreVal, NewGV, SI);
1189 } else if (!UI->use_empty()) {
1190 // Change the load into a load of bool then a select.
1191 LoadInst *LI = cast<LoadInst>(UI);
1192 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1195 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1197 NSI = new SelectInst(NLI, OtherVal, InitVal, "", LI);
1199 LI->replaceAllUsesWith(NSI);
1201 UI->eraseFromParent();
1204 GV->eraseFromParent();
1208 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1209 /// it if possible. If we make a change, return true.
1210 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1211 Module::global_iterator &GVI) {
1212 std::set<PHINode*> PHIUsers;
1214 GV->removeDeadConstantUsers();
1216 if (GV->use_empty()) {
1217 DOUT << "GLOBAL DEAD: " << *GV;
1218 GV->eraseFromParent();
1223 if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
1225 cerr << "Global: " << *GV;
1226 cerr << " isLoaded = " << GS.isLoaded << "\n";
1227 cerr << " StoredType = ";
1228 switch (GS.StoredType) {
1229 case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break;
1230 case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break;
1231 case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break;
1232 case GlobalStatus::isStored: cerr << "stored\n"; break;
1234 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
1235 cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n";
1236 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
1237 cerr << " AccessingFunction = " << GS.AccessingFunction->getName()
1239 cerr << " HasMultipleAccessingFunctions = "
1240 << GS.HasMultipleAccessingFunctions << "\n";
1241 cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n";
1242 cerr << " isNotSuitableForSRA = " << GS.isNotSuitableForSRA << "\n";
1246 // If this is a first class global and has only one accessing function
1247 // and this function is main (which we know is not recursive we can make
1248 // this global a local variable) we replace the global with a local alloca
1249 // in this function.
1251 // NOTE: It doesn't make sense to promote non first class types since we
1252 // are just replacing static memory to stack memory.
1253 if (!GS.HasMultipleAccessingFunctions &&
1254 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1255 GV->getType()->getElementType()->isFirstClassType() &&
1256 GS.AccessingFunction->getName() == "main" &&
1257 GS.AccessingFunction->hasExternalLinkage()) {
1258 DOUT << "LOCALIZING GLOBAL: " << *GV;
1259 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin();
1260 const Type* ElemTy = GV->getType()->getElementType();
1261 // FIXME: Pass Global's alignment when globals have alignment
1262 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI);
1263 if (!isa<UndefValue>(GV->getInitializer()))
1264 new StoreInst(GV->getInitializer(), Alloca, FirstI);
1266 GV->replaceAllUsesWith(Alloca);
1267 GV->eraseFromParent();
1272 // If the global is never loaded (but may be stored to), it is dead.
1275 DOUT << "GLOBAL NEVER LOADED: " << *GV;
1277 // Delete any stores we can find to the global. We may not be able to
1278 // make it completely dead though.
1279 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1281 // If the global is dead now, delete it.
1282 if (GV->use_empty()) {
1283 GV->eraseFromParent();
1289 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1290 DOUT << "MARKING CONSTANT: " << *GV;
1291 GV->setConstant(true);
1293 // Clean up any obviously simplifiable users now.
1294 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1296 // If the global is dead now, just nuke it.
1297 if (GV->use_empty()) {
1298 DOUT << " *** Marking constant allowed us to simplify "
1299 << "all users and delete global!\n";
1300 GV->eraseFromParent();
1306 } else if (!GS.isNotSuitableForSRA &&
1307 !GV->getInitializer()->getType()->isFirstClassType()) {
1308 if (GlobalVariable *FirstNewGV = SRAGlobal(GV)) {
1309 GVI = FirstNewGV; // Don't skip the newly produced globals!
1312 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1313 // If the initial value for the global was an undef value, and if only
1314 // one other value was stored into it, we can just change the
1315 // initializer to be an undef value, then delete all stores to the
1316 // global. This allows us to mark it constant.
1317 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1318 if (isa<UndefValue>(GV->getInitializer())) {
1319 // Change the initial value here.
1320 GV->setInitializer(SOVConstant);
1322 // Clean up any obviously simplifiable users now.
1323 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1325 if (GV->use_empty()) {
1326 DOUT << " *** Substituting initializer allowed us to "
1327 << "simplify all users and delete global!\n";
1328 GV->eraseFromParent();
1337 // Try to optimize globals based on the knowledge that only one value
1338 // (besides its initializer) is ever stored to the global.
1339 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1340 getAnalysis<TargetData>()))
1343 // Otherwise, if the global was not a boolean, we can shrink it to be a
1345 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1346 if (GV->getType()->getElementType() != Type::Int1Ty &&
1347 !GV->getType()->getElementType()->isFloatingPoint() &&
1348 !isa<PackedType>(GV->getType()->getElementType()) &&
1350 DOUT << " *** SHRINKING TO BOOL: " << *GV;
1351 ShrinkGlobalToBoolean(GV, SOVConstant);
1360 /// OnlyCalledDirectly - Return true if the specified function is only called
1361 /// directly. In other words, its address is never taken.
1362 static bool OnlyCalledDirectly(Function *F) {
1363 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1364 Instruction *User = dyn_cast<Instruction>(*UI);
1365 if (!User) return false;
1366 if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false;
1368 // See if the function address is passed as an argument.
1369 for (unsigned i = 1, e = User->getNumOperands(); i != e; ++i)
1370 if (User->getOperand(i) == F) return false;
1375 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1376 /// function, changing them to FastCC.
1377 static void ChangeCalleesToFastCall(Function *F) {
1378 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1379 Instruction *User = cast<Instruction>(*UI);
1380 if (CallInst *CI = dyn_cast<CallInst>(User))
1381 CI->setCallingConv(CallingConv::Fast);
1383 cast<InvokeInst>(User)->setCallingConv(CallingConv::Fast);
1387 bool GlobalOpt::OptimizeFunctions(Module &M) {
1388 bool Changed = false;
1389 // Optimize functions.
1390 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1392 F->removeDeadConstantUsers();
1393 if (F->use_empty() && (F->hasInternalLinkage() ||
1394 F->hasLinkOnceLinkage())) {
1395 M.getFunctionList().erase(F);
1398 } else if (F->hasInternalLinkage() &&
1399 F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1400 OnlyCalledDirectly(F)) {
1401 // If this function has C calling conventions, is not a varargs
1402 // function, and is only called directly, promote it to use the Fast
1403 // calling convention.
1404 F->setCallingConv(CallingConv::Fast);
1405 ChangeCalleesToFastCall(F);
1413 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1414 bool Changed = false;
1415 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1417 GlobalVariable *GV = GVI++;
1418 if (!GV->isConstant() && GV->hasInternalLinkage() &&
1419 GV->hasInitializer())
1420 Changed |= ProcessInternalGlobal(GV, GVI);
1425 /// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1426 /// initializers have an init priority of 65535.
1427 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1428 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1430 if (I->getName() == "llvm.global_ctors") {
1431 // Found it, verify it's an array of { int, void()* }.
1432 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1434 const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1435 if (!STy || STy->getNumElements() != 2 ||
1436 STy->getElementType(0) != Type::Int32Ty) return 0;
1437 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1438 if (!PFTy) return 0;
1439 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1440 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() ||
1441 FTy->getNumParams() != 0)
1444 // Verify that the initializer is simple enough for us to handle.
1445 if (!I->hasInitializer()) return 0;
1446 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1448 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1449 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(CA->getOperand(i))) {
1450 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1453 // Must have a function or null ptr.
1454 if (!isa<Function>(CS->getOperand(1)))
1457 // Init priority must be standard.
1458 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1459 if (!CI || CI->getZExtValue() != 65535)
1470 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1471 /// return a list of the functions and null terminator as a vector.
1472 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1473 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1474 std::vector<Function*> Result;
1475 Result.reserve(CA->getNumOperands());
1476 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) {
1477 ConstantStruct *CS = cast<ConstantStruct>(CA->getOperand(i));
1478 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1483 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1484 /// specified array, returning the new global to use.
1485 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1486 const std::vector<Function*> &Ctors) {
1487 // If we made a change, reassemble the initializer list.
1488 std::vector<Constant*> CSVals;
1489 CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535));
1490 CSVals.push_back(0);
1492 // Create the new init list.
1493 std::vector<Constant*> CAList;
1494 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1496 CSVals[1] = Ctors[i];
1498 const Type *FTy = FunctionType::get(Type::VoidTy,
1499 std::vector<const Type*>(), false);
1500 const PointerType *PFTy = PointerType::get(FTy);
1501 CSVals[1] = Constant::getNullValue(PFTy);
1502 CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647);
1504 CAList.push_back(ConstantStruct::get(CSVals));
1507 // Create the array initializer.
1508 const Type *StructTy =
1509 cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
1510 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()),
1513 // If we didn't change the number of elements, don't create a new GV.
1514 if (CA->getType() == GCL->getInitializer()->getType()) {
1515 GCL->setInitializer(CA);
1519 // Create the new global and insert it next to the existing list.
1520 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
1521 GCL->getLinkage(), CA);
1522 GCL->getParent()->getGlobalList().insert(GCL, NGV);
1525 // Nuke the old list, replacing any uses with the new one.
1526 if (!GCL->use_empty()) {
1528 if (V->getType() != GCL->getType())
1529 V = ConstantExpr::getBitCast(V, GCL->getType());
1530 GCL->replaceAllUsesWith(V);
1532 GCL->eraseFromParent();
1541 static Constant *getVal(std::map<Value*, Constant*> &ComputedValues,
1543 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
1544 Constant *R = ComputedValues[V];
1545 assert(R && "Reference to an uncomputed value!");
1549 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
1550 /// enough for us to understand. In particular, if it is a cast of something,
1551 /// we punt. We basically just support direct accesses to globals and GEP's of
1552 /// globals. This should be kept up to date with CommitValueTo.
1553 static bool isSimpleEnoughPointerToCommit(Constant *C) {
1554 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
1555 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1556 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1557 return !GV->isDeclaration(); // reject external globals.
1559 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
1560 // Handle a constantexpr gep.
1561 if (CE->getOpcode() == Instruction::GetElementPtr &&
1562 isa<GlobalVariable>(CE->getOperand(0))) {
1563 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1564 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1565 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1566 return GV->hasInitializer() &&
1567 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1572 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
1573 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
1574 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
1575 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
1576 ConstantExpr *Addr, unsigned OpNo) {
1577 // Base case of the recursion.
1578 if (OpNo == Addr->getNumOperands()) {
1579 assert(Val->getType() == Init->getType() && "Type mismatch!");
1583 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
1584 std::vector<Constant*> Elts;
1586 // Break up the constant into its elements.
1587 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
1588 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i)
1589 Elts.push_back(CS->getOperand(i));
1590 } else if (isa<ConstantAggregateZero>(Init)) {
1591 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1592 Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
1593 } else if (isa<UndefValue>(Init)) {
1594 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1595 Elts.push_back(UndefValue::get(STy->getElementType(i)));
1597 assert(0 && "This code is out of sync with "
1598 " ConstantFoldLoadThroughGEPConstantExpr");
1601 // Replace the element that we are supposed to.
1602 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
1603 unsigned Idx = CU->getZExtValue();
1604 assert(Idx < STy->getNumElements() && "Struct index out of range!");
1605 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
1607 // Return the modified struct.
1608 return ConstantStruct::get(Elts);
1610 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
1611 const ArrayType *ATy = cast<ArrayType>(Init->getType());
1613 // Break up the array into elements.
1614 std::vector<Constant*> Elts;
1615 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
1616 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1617 Elts.push_back(CA->getOperand(i));
1618 } else if (isa<ConstantAggregateZero>(Init)) {
1619 Constant *Elt = Constant::getNullValue(ATy->getElementType());
1620 Elts.assign(ATy->getNumElements(), Elt);
1621 } else if (isa<UndefValue>(Init)) {
1622 Constant *Elt = UndefValue::get(ATy->getElementType());
1623 Elts.assign(ATy->getNumElements(), Elt);
1625 assert(0 && "This code is out of sync with "
1626 " ConstantFoldLoadThroughGEPConstantExpr");
1629 assert(CI->getZExtValue() < ATy->getNumElements());
1630 Elts[CI->getZExtValue()] =
1631 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
1632 return ConstantArray::get(ATy, Elts);
1636 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
1637 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
1638 static void CommitValueTo(Constant *Val, Constant *Addr) {
1639 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
1640 assert(GV->hasInitializer());
1641 GV->setInitializer(Val);
1645 ConstantExpr *CE = cast<ConstantExpr>(Addr);
1646 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1648 Constant *Init = GV->getInitializer();
1649 Init = EvaluateStoreInto(Init, Val, CE, 2);
1650 GV->setInitializer(Init);
1653 /// ComputeLoadResult - Return the value that would be computed by a load from
1654 /// P after the stores reflected by 'memory' have been performed. If we can't
1655 /// decide, return null.
1656 static Constant *ComputeLoadResult(Constant *P,
1657 const std::map<Constant*, Constant*> &Memory) {
1658 // If this memory location has been recently stored, use the stored value: it
1659 // is the most up-to-date.
1660 std::map<Constant*, Constant*>::const_iterator I = Memory.find(P);
1661 if (I != Memory.end()) return I->second;
1664 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
1665 if (GV->hasInitializer())
1666 return GV->getInitializer();
1670 // Handle a constantexpr getelementptr.
1671 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
1672 if (CE->getOpcode() == Instruction::GetElementPtr &&
1673 isa<GlobalVariable>(CE->getOperand(0))) {
1674 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1675 if (GV->hasInitializer())
1676 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1679 return 0; // don't know how to evaluate.
1682 /// EvaluateFunction - Evaluate a call to function F, returning true if
1683 /// successful, false if we can't evaluate it. ActualArgs contains the formal
1684 /// arguments for the function.
1685 static bool EvaluateFunction(Function *F, Constant *&RetVal,
1686 const std::vector<Constant*> &ActualArgs,
1687 std::vector<Function*> &CallStack,
1688 std::map<Constant*, Constant*> &MutatedMemory,
1689 std::vector<GlobalVariable*> &AllocaTmps) {
1690 // Check to see if this function is already executing (recursion). If so,
1691 // bail out. TODO: we might want to accept limited recursion.
1692 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
1695 CallStack.push_back(F);
1697 /// Values - As we compute SSA register values, we store their contents here.
1698 std::map<Value*, Constant*> Values;
1700 // Initialize arguments to the incoming values specified.
1702 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
1704 Values[AI] = ActualArgs[ArgNo];
1706 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
1707 /// we can only evaluate any one basic block at most once. This set keeps
1708 /// track of what we have executed so we can detect recursive cases etc.
1709 std::set<BasicBlock*> ExecutedBlocks;
1711 // CurInst - The current instruction we're evaluating.
1712 BasicBlock::iterator CurInst = F->begin()->begin();
1714 // This is the main evaluation loop.
1716 Constant *InstResult = 0;
1718 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
1719 if (SI->isVolatile()) return false; // no volatile accesses.
1720 Constant *Ptr = getVal(Values, SI->getOperand(1));
1721 if (!isSimpleEnoughPointerToCommit(Ptr))
1722 // If this is too complex for us to commit, reject it.
1724 Constant *Val = getVal(Values, SI->getOperand(0));
1725 MutatedMemory[Ptr] = Val;
1726 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
1727 InstResult = ConstantExpr::get(BO->getOpcode(),
1728 getVal(Values, BO->getOperand(0)),
1729 getVal(Values, BO->getOperand(1)));
1730 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
1731 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
1732 getVal(Values, CI->getOperand(0)),
1733 getVal(Values, CI->getOperand(1)));
1734 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
1735 InstResult = ConstantExpr::getCast(CI->getOpcode(),
1736 getVal(Values, CI->getOperand(0)),
1738 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
1739 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
1740 getVal(Values, SI->getOperand(1)),
1741 getVal(Values, SI->getOperand(2)));
1742 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
1743 Constant *P = getVal(Values, GEP->getOperand(0));
1744 SmallVector<Constant*, 8> GEPOps;
1745 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
1746 GEPOps.push_back(getVal(Values, GEP->getOperand(i)));
1747 InstResult = ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size());
1748 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
1749 if (LI->isVolatile()) return false; // no volatile accesses.
1750 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
1752 if (InstResult == 0) return false; // Could not evaluate load.
1753 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
1754 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs.
1755 const Type *Ty = AI->getType()->getElementType();
1756 AllocaTmps.push_back(new GlobalVariable(Ty, false,
1757 GlobalValue::InternalLinkage,
1758 UndefValue::get(Ty),
1760 InstResult = AllocaTmps.back();
1761 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
1762 // Cannot handle inline asm.
1763 if (isa<InlineAsm>(CI->getOperand(0))) return false;
1765 // Resolve function pointers.
1766 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0)));
1767 if (!Callee) return false; // Cannot resolve.
1769 std::vector<Constant*> Formals;
1770 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
1771 Formals.push_back(getVal(Values, CI->getOperand(i)));
1773 if (Callee->isDeclaration()) {
1774 // If this is a function we can constant fold, do it.
1775 if (Constant *C = ConstantFoldCall(Callee, &Formals[0],
1782 if (Callee->getFunctionType()->isVarArg())
1787 // Execute the call, if successful, use the return value.
1788 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
1789 MutatedMemory, AllocaTmps))
1791 InstResult = RetVal;
1793 } else if (isa<TerminatorInst>(CurInst)) {
1794 BasicBlock *NewBB = 0;
1795 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
1796 if (BI->isUnconditional()) {
1797 NewBB = BI->getSuccessor(0);
1800 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
1801 if (!Cond) return false; // Cannot determine.
1803 NewBB = BI->getSuccessor(!Cond->getZExtValue());
1805 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
1807 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
1808 if (!Val) return false; // Cannot determine.
1809 NewBB = SI->getSuccessor(SI->findCaseValue(Val));
1810 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
1811 if (RI->getNumOperands())
1812 RetVal = getVal(Values, RI->getOperand(0));
1814 CallStack.pop_back(); // return from fn.
1815 return true; // We succeeded at evaluating this ctor!
1817 // invoke, unwind, unreachable.
1818 return false; // Cannot handle this terminator.
1821 // Okay, we succeeded in evaluating this control flow. See if we have
1822 // executed the new block before. If so, we have a looping function,
1823 // which we cannot evaluate in reasonable time.
1824 if (!ExecutedBlocks.insert(NewBB).second)
1825 return false; // looped!
1827 // Okay, we have never been in this block before. Check to see if there
1828 // are any PHI nodes. If so, evaluate them with information about where
1830 BasicBlock *OldBB = CurInst->getParent();
1831 CurInst = NewBB->begin();
1833 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
1834 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
1836 // Do NOT increment CurInst. We know that the terminator had no value.
1839 // Did not know how to evaluate this!
1843 if (!CurInst->use_empty())
1844 Values[CurInst] = InstResult;
1846 // Advance program counter.
1851 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
1852 /// we can. Return true if we can, false otherwise.
1853 static bool EvaluateStaticConstructor(Function *F) {
1854 /// MutatedMemory - For each store we execute, we update this map. Loads
1855 /// check this to get the most up-to-date value. If evaluation is successful,
1856 /// this state is committed to the process.
1857 std::map<Constant*, Constant*> MutatedMemory;
1859 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
1860 /// to represent its body. This vector is needed so we can delete the
1861 /// temporary globals when we are done.
1862 std::vector<GlobalVariable*> AllocaTmps;
1864 /// CallStack - This is used to detect recursion. In pathological situations
1865 /// we could hit exponential behavior, but at least there is nothing
1867 std::vector<Function*> CallStack;
1869 // Call the function.
1870 Constant *RetValDummy;
1871 bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(),
1872 CallStack, MutatedMemory, AllocaTmps);
1874 // We succeeded at evaluation: commit the result.
1875 DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
1876 << F->getName() << "' to " << MutatedMemory.size()
1878 for (std::map<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
1879 E = MutatedMemory.end(); I != E; ++I)
1880 CommitValueTo(I->second, I->first);
1883 // At this point, we are done interpreting. If we created any 'alloca'
1884 // temporaries, release them now.
1885 while (!AllocaTmps.empty()) {
1886 GlobalVariable *Tmp = AllocaTmps.back();
1887 AllocaTmps.pop_back();
1889 // If there are still users of the alloca, the program is doing something
1890 // silly, e.g. storing the address of the alloca somewhere and using it
1891 // later. Since this is undefined, we'll just make it be null.
1892 if (!Tmp->use_empty())
1893 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
1902 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
1903 /// Return true if anything changed.
1904 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
1905 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
1906 bool MadeChange = false;
1907 if (Ctors.empty()) return false;
1909 // Loop over global ctors, optimizing them when we can.
1910 for (unsigned i = 0; i != Ctors.size(); ++i) {
1911 Function *F = Ctors[i];
1912 // Found a null terminator in the middle of the list, prune off the rest of
1915 if (i != Ctors.size()-1) {
1922 // We cannot simplify external ctor functions.
1923 if (F->empty()) continue;
1925 // If we can evaluate the ctor at compile time, do.
1926 if (EvaluateStaticConstructor(F)) {
1927 Ctors.erase(Ctors.begin()+i);
1930 ++NumCtorsEvaluated;
1935 if (!MadeChange) return false;
1937 GCL = InstallGlobalCtors(GCL, Ctors);
1942 bool GlobalOpt::runOnModule(Module &M) {
1943 bool Changed = false;
1945 // Try to find the llvm.globalctors list.
1946 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
1948 bool LocalChange = true;
1949 while (LocalChange) {
1950 LocalChange = false;
1952 // Delete functions that are trivially dead, ccc -> fastcc
1953 LocalChange |= OptimizeFunctions(M);
1955 // Optimize global_ctors list.
1957 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
1959 // Optimize non-address-taken globals.
1960 LocalChange |= OptimizeGlobalVars(M);
1961 Changed |= LocalChange;
1964 // TODO: Move all global ctors functions to the end of the module for code