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"
34 STATISTIC(NumMarked , "Number of globals marked constant");
35 STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
36 STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
37 STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
38 STATISTIC(NumDeleted , "Number of globals deleted");
39 STATISTIC(NumFnDeleted , "Number of functions deleted");
40 STATISTIC(NumGlobUses , "Number of global uses devirtualized");
41 STATISTIC(NumLocalized , "Number of globals localized");
42 STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans");
43 STATISTIC(NumFastCallFns , "Number of functions converted to fastcc");
44 STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
47 struct GlobalOpt : public ModulePass {
48 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
49 AU.addRequired<TargetData>();
52 bool runOnModule(Module &M);
55 GlobalVariable *FindGlobalCtors(Module &M);
56 bool OptimizeFunctions(Module &M);
57 bool OptimizeGlobalVars(Module &M);
58 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
59 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
62 RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer");
65 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
67 /// GlobalStatus - As we analyze each global, keep track of some information
68 /// about it. If we find out that the address of the global is taken, none of
69 /// this info will be accurate.
71 /// isLoaded - True if the global is ever loaded. If the global isn't ever
72 /// loaded it can be deleted.
75 /// StoredType - Keep track of what stores to the global look like.
78 /// NotStored - There is no store to this global. It can thus be marked
82 /// isInitializerStored - This global is stored to, but the only thing
83 /// stored is the constant it was initialized with. This is only tracked
84 /// for scalar globals.
87 /// isStoredOnce - This global is stored to, but only its initializer and
88 /// one other value is ever stored to it. If this global isStoredOnce, we
89 /// track the value stored to it in StoredOnceValue below. This is only
90 /// tracked for scalar globals.
93 /// isStored - This global is stored to by multiple values or something else
94 /// that we cannot track.
98 /// StoredOnceValue - If only one value (besides the initializer constant) is
99 /// ever stored to this global, keep track of what value it is.
100 Value *StoredOnceValue;
102 /// AccessingFunction/HasMultipleAccessingFunctions - These start out
103 /// null/false. When the first accessing function is noticed, it is recorded.
104 /// When a second different accessing function is noticed,
105 /// HasMultipleAccessingFunctions is set to true.
106 Function *AccessingFunction;
107 bool HasMultipleAccessingFunctions;
109 /// HasNonInstructionUser - Set to true if this global has a user that is not
110 /// an instruction (e.g. a constant expr or GV initializer).
111 bool HasNonInstructionUser;
113 /// HasPHIUser - Set to true if this global has a user that is a PHI node.
116 /// isNotSuitableForSRA - Keep track of whether any SRA preventing users of
117 /// the global exist. Such users include GEP instruction with variable
118 /// indexes, and non-gep/load/store users like constant expr casts.
119 bool isNotSuitableForSRA;
121 GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0),
122 AccessingFunction(0), HasMultipleAccessingFunctions(false),
123 HasNonInstructionUser(false), HasPHIUser(false),
124 isNotSuitableForSRA(false) {}
129 /// ConstantIsDead - Return true if the specified constant is (transitively)
130 /// dead. The constant may be used by other constants (e.g. constant arrays and
131 /// constant exprs) as long as they are dead, but it cannot be used by anything
133 static bool ConstantIsDead(Constant *C) {
134 if (isa<GlobalValue>(C)) return false;
136 for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI)
137 if (Constant *CU = dyn_cast<Constant>(*UI)) {
138 if (!ConstantIsDead(CU)) return false;
145 /// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
146 /// structure. If the global has its address taken, return true to indicate we
147 /// can't do anything with it.
149 static bool AnalyzeGlobal(Value *V, GlobalStatus &GS,
150 std::set<PHINode*> &PHIUsers) {
151 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
152 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
153 GS.HasNonInstructionUser = true;
155 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
156 if (CE->getOpcode() != Instruction::GetElementPtr)
157 GS.isNotSuitableForSRA = true;
158 else if (!GS.isNotSuitableForSRA) {
159 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
160 // don't like < 3 operand CE's, and we don't like non-constant integer
162 if (CE->getNumOperands() < 3 || !CE->getOperand(1)->isNullValue())
163 GS.isNotSuitableForSRA = true;
165 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
166 if (!isa<ConstantInt>(CE->getOperand(i))) {
167 GS.isNotSuitableForSRA = true;
173 } else if (Instruction *I = dyn_cast<Instruction>(*UI)) {
174 if (!GS.HasMultipleAccessingFunctions) {
175 Function *F = I->getParent()->getParent();
176 if (GS.AccessingFunction == 0)
177 GS.AccessingFunction = F;
178 else if (GS.AccessingFunction != F)
179 GS.HasMultipleAccessingFunctions = true;
181 if (isa<LoadInst>(I)) {
183 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
184 // Don't allow a store OF the address, only stores TO the address.
185 if (SI->getOperand(0) == V) return true;
187 // If this is a direct store to the global (i.e., the global is a scalar
188 // value, not an aggregate), keep more specific information about
190 if (GS.StoredType != GlobalStatus::isStored)
191 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){
192 Value *StoredVal = SI->getOperand(0);
193 if (StoredVal == GV->getInitializer()) {
194 if (GS.StoredType < GlobalStatus::isInitializerStored)
195 GS.StoredType = GlobalStatus::isInitializerStored;
196 } else if (isa<LoadInst>(StoredVal) &&
197 cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
199 if (GS.StoredType < GlobalStatus::isInitializerStored)
200 GS.StoredType = GlobalStatus::isInitializerStored;
201 } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
202 GS.StoredType = GlobalStatus::isStoredOnce;
203 GS.StoredOnceValue = StoredVal;
204 } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
205 GS.StoredOnceValue == StoredVal) {
208 GS.StoredType = GlobalStatus::isStored;
211 GS.StoredType = GlobalStatus::isStored;
213 } else if (isa<GetElementPtrInst>(I)) {
214 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
216 // If the first two indices are constants, this can be SRA'd.
217 if (isa<GlobalVariable>(I->getOperand(0))) {
218 if (I->getNumOperands() < 3 || !isa<Constant>(I->getOperand(1)) ||
219 !cast<Constant>(I->getOperand(1))->isNullValue() ||
220 !isa<ConstantInt>(I->getOperand(2)))
221 GS.isNotSuitableForSRA = true;
222 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(I->getOperand(0))){
223 if (CE->getOpcode() != Instruction::GetElementPtr ||
224 CE->getNumOperands() < 3 || I->getNumOperands() < 2 ||
225 !isa<Constant>(I->getOperand(0)) ||
226 !cast<Constant>(I->getOperand(0))->isNullValue())
227 GS.isNotSuitableForSRA = true;
229 GS.isNotSuitableForSRA = true;
231 } else if (isa<SelectInst>(I)) {
232 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
233 GS.isNotSuitableForSRA = true;
234 } else if (PHINode *PN = dyn_cast<PHINode>(I)) {
235 // PHI nodes we can check just like select or GEP instructions, but we
236 // have to be careful about infinite recursion.
237 if (PHIUsers.insert(PN).second) // Not already visited.
238 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
239 GS.isNotSuitableForSRA = true;
240 GS.HasPHIUser = true;
241 } else if (isa<CmpInst>(I)) {
242 GS.isNotSuitableForSRA = true;
243 } else if (isa<MemCpyInst>(I) || isa<MemMoveInst>(I)) {
244 if (I->getOperand(1) == V)
245 GS.StoredType = GlobalStatus::isStored;
246 if (I->getOperand(2) == V)
248 GS.isNotSuitableForSRA = true;
249 } else if (isa<MemSetInst>(I)) {
250 assert(I->getOperand(1) == V && "Memset only takes one pointer!");
251 GS.StoredType = GlobalStatus::isStored;
252 GS.isNotSuitableForSRA = true;
254 return true; // Any other non-load instruction might take address!
256 } else if (Constant *C = dyn_cast<Constant>(*UI)) {
257 GS.HasNonInstructionUser = true;
258 // We might have a dead and dangling constant hanging off of here.
259 if (!ConstantIsDead(C))
262 GS.HasNonInstructionUser = true;
263 // Otherwise must be some other user.
270 static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) {
271 ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
273 unsigned IdxV = CI->getZExtValue();
275 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
276 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
277 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
278 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
279 } else if (ConstantPacked *CP = dyn_cast<ConstantPacked>(Agg)) {
280 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
281 } else if (isa<ConstantAggregateZero>(Agg)) {
282 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
283 if (IdxV < STy->getNumElements())
284 return Constant::getNullValue(STy->getElementType(IdxV));
285 } else if (const SequentialType *STy =
286 dyn_cast<SequentialType>(Agg->getType())) {
287 return Constant::getNullValue(STy->getElementType());
289 } else if (isa<UndefValue>(Agg)) {
290 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
291 if (IdxV < STy->getNumElements())
292 return UndefValue::get(STy->getElementType(IdxV));
293 } else if (const SequentialType *STy =
294 dyn_cast<SequentialType>(Agg->getType())) {
295 return UndefValue::get(STy->getElementType());
302 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
303 /// users of the global, cleaning up the obvious ones. This is largely just a
304 /// quick scan over the use list to clean up the easy and obvious cruft. This
305 /// returns true if it made a change.
306 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
307 bool Changed = false;
308 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
311 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
313 // Replace the load with the initializer.
314 LI->replaceAllUsesWith(Init);
315 LI->eraseFromParent();
318 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
319 // Store must be unreachable or storing Init into the global.
320 SI->eraseFromParent();
322 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
323 if (CE->getOpcode() == Instruction::GetElementPtr) {
324 Constant *SubInit = 0;
326 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
327 Changed |= CleanupConstantGlobalUsers(CE, SubInit);
328 } else if (CE->getOpcode() == Instruction::BitCast &&
329 isa<PointerType>(CE->getType())) {
330 // Pointer cast, delete any stores and memsets to the global.
331 Changed |= CleanupConstantGlobalUsers(CE, 0);
334 if (CE->use_empty()) {
335 CE->destroyConstant();
338 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
339 Constant *SubInit = 0;
341 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
342 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
343 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
344 Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
346 if (GEP->use_empty()) {
347 GEP->eraseFromParent();
350 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
351 if (MI->getRawDest() == V) {
352 MI->eraseFromParent();
356 } else if (Constant *C = dyn_cast<Constant>(U)) {
357 // If we have a chain of dead constantexprs or other things dangling from
358 // us, and if they are all dead, nuke them without remorse.
359 if (ConstantIsDead(C)) {
360 C->destroyConstant();
361 // This could have invalidated UI, start over from scratch.
362 CleanupConstantGlobalUsers(V, Init);
370 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
371 /// variable. This opens the door for other optimizations by exposing the
372 /// behavior of the program in a more fine-grained way. We have determined that
373 /// this transformation is safe already. We return the first global variable we
374 /// insert so that the caller can reprocess it.
375 static GlobalVariable *SRAGlobal(GlobalVariable *GV) {
376 assert(GV->hasInternalLinkage() && !GV->isConstant());
377 Constant *Init = GV->getInitializer();
378 const Type *Ty = Init->getType();
380 std::vector<GlobalVariable*> NewGlobals;
381 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
383 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
384 NewGlobals.reserve(STy->getNumElements());
385 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
386 Constant *In = getAggregateConstantElement(Init,
387 ConstantInt::get(Type::Int32Ty, i));
388 assert(In && "Couldn't get element of initializer?");
389 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
390 GlobalVariable::InternalLinkage,
391 In, GV->getName()+"."+utostr(i));
392 Globals.insert(GV, NGV);
393 NewGlobals.push_back(NGV);
395 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
396 unsigned NumElements = 0;
397 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
398 NumElements = ATy->getNumElements();
399 else if (const PackedType *PTy = dyn_cast<PackedType>(STy))
400 NumElements = PTy->getNumElements();
402 assert(0 && "Unknown aggregate sequential type!");
404 if (NumElements > 16 && GV->hasNUsesOrMore(16))
405 return 0; // It's not worth it.
406 NewGlobals.reserve(NumElements);
407 for (unsigned i = 0, e = NumElements; i != e; ++i) {
408 Constant *In = getAggregateConstantElement(Init,
409 ConstantInt::get(Type::Int32Ty, i));
410 assert(In && "Couldn't get element of initializer?");
412 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
413 GlobalVariable::InternalLinkage,
414 In, GV->getName()+"."+utostr(i));
415 Globals.insert(GV, NGV);
416 NewGlobals.push_back(NGV);
420 if (NewGlobals.empty())
423 DOUT << "PERFORMING GLOBAL SRA ON: " << *GV;
425 Constant *NullInt = Constant::getNullValue(Type::Int32Ty);
427 // Loop over all of the uses of the global, replacing the constantexpr geps,
428 // with smaller constantexpr geps or direct references.
429 while (!GV->use_empty()) {
430 User *GEP = GV->use_back();
431 assert(((isa<ConstantExpr>(GEP) &&
432 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
433 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
435 // Ignore the 1th operand, which has to be zero or else the program is quite
436 // broken (undefined). Get the 2nd operand, which is the structure or array
438 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
439 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
441 Value *NewPtr = NewGlobals[Val];
443 // Form a shorter GEP if needed.
444 if (GEP->getNumOperands() > 3)
445 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
446 std::vector<Constant*> Idxs;
447 Idxs.push_back(NullInt);
448 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
449 Idxs.push_back(CE->getOperand(i));
450 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr), Idxs);
452 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
453 std::vector<Value*> Idxs;
454 Idxs.push_back(NullInt);
455 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
456 Idxs.push_back(GEPI->getOperand(i));
457 NewPtr = new GetElementPtrInst(NewPtr, Idxs,
458 GEPI->getName()+"."+utostr(Val), GEPI);
460 GEP->replaceAllUsesWith(NewPtr);
462 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
463 GEPI->eraseFromParent();
465 cast<ConstantExpr>(GEP)->destroyConstant();
468 // Delete the old global, now that it is dead.
472 // Loop over the new globals array deleting any globals that are obviously
473 // dead. This can arise due to scalarization of a structure or an array that
474 // has elements that are dead.
475 unsigned FirstGlobal = 0;
476 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
477 if (NewGlobals[i]->use_empty()) {
478 Globals.erase(NewGlobals[i]);
479 if (FirstGlobal == i) ++FirstGlobal;
482 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
485 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
486 /// value will trap if the value is dynamically null.
487 static bool AllUsesOfValueWillTrapIfNull(Value *V) {
488 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
489 if (isa<LoadInst>(*UI)) {
491 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
492 if (SI->getOperand(0) == V) {
493 //cerr << "NONTRAPPING USE: " << **UI;
494 return false; // Storing the value.
496 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
497 if (CI->getOperand(0) != V) {
498 //cerr << "NONTRAPPING USE: " << **UI;
499 return false; // Not calling the ptr
501 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
502 if (II->getOperand(0) != V) {
503 //cerr << "NONTRAPPING USE: " << **UI;
504 return false; // Not calling the ptr
506 } else if (CastInst *CI = dyn_cast<CastInst>(*UI)) {
507 if (!AllUsesOfValueWillTrapIfNull(CI)) return false;
508 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
509 if (!AllUsesOfValueWillTrapIfNull(GEPI)) return false;
510 } else if (isa<ICmpInst>(*UI) &&
511 isa<ConstantPointerNull>(UI->getOperand(1))) {
512 // Ignore setcc X, null
514 //cerr << "NONTRAPPING USE: " << **UI;
520 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
521 /// from GV will trap if the loaded value is null. Note that this also permits
522 /// comparisons of the loaded value against null, as a special case.
523 static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) {
524 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI)
525 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
526 if (!AllUsesOfValueWillTrapIfNull(LI))
528 } else if (isa<StoreInst>(*UI)) {
529 // Ignore stores to the global.
531 // We don't know or understand this user, bail out.
532 //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI;
539 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
540 bool Changed = false;
541 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
542 Instruction *I = cast<Instruction>(*UI++);
543 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
544 LI->setOperand(0, NewV);
546 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
547 if (SI->getOperand(1) == V) {
548 SI->setOperand(1, NewV);
551 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
552 if (I->getOperand(0) == V) {
553 // Calling through the pointer! Turn into a direct call, but be careful
554 // that the pointer is not also being passed as an argument.
555 I->setOperand(0, NewV);
557 bool PassedAsArg = false;
558 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
559 if (I->getOperand(i) == V) {
561 I->setOperand(i, NewV);
565 // Being passed as an argument also. Be careful to not invalidate UI!
569 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
570 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
571 ConstantExpr::getCast(CI->getOpcode(),
572 NewV, CI->getType()));
573 if (CI->use_empty()) {
575 CI->eraseFromParent();
577 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
578 // Should handle GEP here.
579 std::vector<Constant*> Indices;
580 Indices.reserve(GEPI->getNumOperands()-1);
581 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
582 if (Constant *C = dyn_cast<Constant>(GEPI->getOperand(i)))
583 Indices.push_back(C);
586 if (Indices.size() == GEPI->getNumOperands()-1)
587 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
588 ConstantExpr::getGetElementPtr(NewV, Indices));
589 if (GEPI->use_empty()) {
591 GEPI->eraseFromParent();
600 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
601 /// value stored into it. If there are uses of the loaded value that would trap
602 /// if the loaded value is dynamically null, then we know that they cannot be
603 /// reachable with a null optimize away the load.
604 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) {
605 std::vector<LoadInst*> Loads;
606 bool Changed = false;
608 // Replace all uses of loads with uses of uses of the stored value.
609 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end();
611 if (LoadInst *LI = dyn_cast<LoadInst>(*GUI)) {
613 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
615 assert(isa<StoreInst>(*GUI) && "Only expect load and stores!");
619 DOUT << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV;
623 // Delete all of the loads we can, keeping track of whether we nuked them all!
624 bool AllLoadsGone = true;
625 while (!Loads.empty()) {
626 LoadInst *L = Loads.back();
627 if (L->use_empty()) {
628 L->eraseFromParent();
631 AllLoadsGone = false;
636 // If we nuked all of the loads, then none of the stores are needed either,
637 // nor is the global.
639 DOUT << " *** GLOBAL NOW DEAD!\n";
640 CleanupConstantGlobalUsers(GV, 0);
641 if (GV->use_empty()) {
642 GV->eraseFromParent();
650 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
651 /// instructions that are foldable.
652 static void ConstantPropUsersOf(Value *V) {
653 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
654 if (Instruction *I = dyn_cast<Instruction>(*UI++))
655 if (Constant *NewC = ConstantFoldInstruction(I)) {
656 I->replaceAllUsesWith(NewC);
658 // Advance UI to the next non-I use to avoid invalidating it!
659 // Instructions could multiply use V.
660 while (UI != E && *UI == I)
662 I->eraseFromParent();
666 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
667 /// variable, and transforms the program as if it always contained the result of
668 /// the specified malloc. Because it is always the result of the specified
669 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
670 /// malloc into a global, and any loads of GV as uses of the new global.
671 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
673 DOUT << "PROMOTING MALLOC GLOBAL: " << *GV << " MALLOC = " << *MI;
674 ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize());
676 if (NElements->getZExtValue() != 1) {
677 // If we have an array allocation, transform it to a single element
678 // allocation to make the code below simpler.
679 Type *NewTy = ArrayType::get(MI->getAllocatedType(),
680 NElements->getZExtValue());
682 new MallocInst(NewTy, Constant::getNullValue(Type::Int32Ty),
683 MI->getAlignment(), MI->getName(), MI);
684 std::vector<Value*> Indices;
685 Indices.push_back(Constant::getNullValue(Type::Int32Ty));
686 Indices.push_back(Indices[0]);
687 Value *NewGEP = new GetElementPtrInst(NewMI, Indices,
688 NewMI->getName()+".el0", MI);
689 MI->replaceAllUsesWith(NewGEP);
690 MI->eraseFromParent();
694 // Create the new global variable. The contents of the malloc'd memory is
695 // undefined, so initialize with an undef value.
696 Constant *Init = UndefValue::get(MI->getAllocatedType());
697 GlobalVariable *NewGV = new GlobalVariable(MI->getAllocatedType(), false,
698 GlobalValue::InternalLinkage, Init,
699 GV->getName()+".body");
700 GV->getParent()->getGlobalList().insert(GV, NewGV);
702 // Anything that used the malloc now uses the global directly.
703 MI->replaceAllUsesWith(NewGV);
705 Constant *RepValue = NewGV;
706 if (NewGV->getType() != GV->getType()->getElementType())
707 RepValue = ConstantExpr::getBitCast(RepValue,
708 GV->getType()->getElementType());
710 // If there is a comparison against null, we will insert a global bool to
711 // keep track of whether the global was initialized yet or not.
712 GlobalVariable *InitBool =
713 new GlobalVariable(Type::BoolTy, false, GlobalValue::InternalLinkage,
714 ConstantInt::getFalse(), GV->getName()+".init");
715 bool InitBoolUsed = false;
717 // Loop over all uses of GV, processing them in turn.
718 std::vector<StoreInst*> Stores;
719 while (!GV->use_empty())
720 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
721 while (!LI->use_empty()) {
722 Use &LoadUse = LI->use_begin().getUse();
723 if (!isa<ICmpInst>(LoadUse.getUser()))
726 ICmpInst *CI = cast<ICmpInst>(LoadUse.getUser());
727 // Replace the cmp X, 0 with a use of the bool value.
728 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", CI);
730 switch (CI->getPredicate()) {
731 default: assert(0 && "Unknown ICmp Predicate!");
732 case ICmpInst::ICMP_ULT:
733 case ICmpInst::ICMP_SLT:
734 LV = ConstantInt::getFalse(); // X < null -> always false
736 case ICmpInst::ICMP_ULE:
737 case ICmpInst::ICMP_SLE:
738 case ICmpInst::ICMP_EQ:
739 LV = BinaryOperator::createNot(LV, "notinit", CI);
741 case ICmpInst::ICMP_NE:
742 case ICmpInst::ICMP_UGE:
743 case ICmpInst::ICMP_SGE:
744 case ICmpInst::ICMP_UGT:
745 case ICmpInst::ICMP_SGT:
748 CI->replaceAllUsesWith(LV);
749 CI->eraseFromParent();
752 LI->eraseFromParent();
754 StoreInst *SI = cast<StoreInst>(GV->use_back());
755 // The global is initialized when the store to it occurs.
756 new StoreInst(ConstantInt::getTrue(), InitBool, SI);
757 SI->eraseFromParent();
760 // If the initialization boolean was used, insert it, otherwise delete it.
762 while (!InitBool->use_empty()) // Delete initializations
763 cast<Instruction>(InitBool->use_back())->eraseFromParent();
766 GV->getParent()->getGlobalList().insert(GV, InitBool);
769 // Now the GV is dead, nuke it and the malloc.
770 GV->eraseFromParent();
771 MI->eraseFromParent();
773 // To further other optimizations, loop over all users of NewGV and try to
774 // constant prop them. This will promote GEP instructions with constant
775 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
776 ConstantPropUsersOf(NewGV);
777 if (RepValue != NewGV)
778 ConstantPropUsersOf(RepValue);
783 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
784 /// to make sure that there are no complex uses of V. We permit simple things
785 /// like dereferencing the pointer, but not storing through the address, unless
786 /// it is to the specified global.
787 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V,
788 GlobalVariable *GV) {
789 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI)
790 if (isa<LoadInst>(*UI) || isa<CmpInst>(*UI)) {
792 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
793 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
794 return false; // Storing the pointer itself... bad.
795 // Otherwise, storing through it, or storing into GV... fine.
796 } else if (isa<GetElementPtrInst>(*UI) || isa<SelectInst>(*UI)) {
797 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(cast<Instruction>(*UI),GV))
805 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
806 /// somewhere. Transform all uses of the allocation into loads from the
807 /// global and uses of the resultant pointer. Further, delete the store into
808 /// GV. This assumes that these value pass the
809 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
810 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
811 GlobalVariable *GV) {
812 while (!Alloc->use_empty()) {
813 Instruction *U = Alloc->use_back();
814 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
815 // If this is the store of the allocation into the global, remove it.
816 if (SI->getOperand(1) == GV) {
817 SI->eraseFromParent();
822 // Insert a load from the global, and use it instead of the malloc.
823 Value *NL = new LoadInst(GV, GV->getName()+".val", U);
824 U->replaceUsesOfWith(Alloc, NL);
828 /// GlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
829 /// GV are simple enough to perform HeapSRA, return true.
830 static bool GlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV) {
831 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;
833 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
834 // We permit two users of the load: setcc comparing against the null
835 // pointer, and a getelementptr of a specific form.
836 for (Value::use_iterator UI = LI->use_begin(), E = LI->use_end(); UI != E;
838 // Comparison against null is ok.
839 if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) {
840 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
845 // getelementptr is also ok, but only a simple form.
846 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI);
847 if (!GEPI) return false;
849 // Must index into the array and into the struct.
850 if (GEPI->getNumOperands() < 3)
853 // Otherwise the GEP is ok.
860 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
861 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
862 /// use FieldGlobals instead. All uses of loaded values satisfy
863 /// GlobalLoadUsesSimpleEnoughForHeapSRA.
864 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Ptr,
865 const std::vector<GlobalVariable*> &FieldGlobals) {
866 std::vector<Value *> InsertedLoadsForPtr;
867 //InsertedLoadsForPtr.resize(FieldGlobals.size());
868 while (!Ptr->use_empty()) {
869 Instruction *User = Ptr->use_back();
871 // If this is a comparison against null, handle it.
872 if (ICmpInst *SCI = dyn_cast<ICmpInst>(User)) {
873 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
874 // If we have a setcc of the loaded pointer, we can use a setcc of any
877 if (InsertedLoadsForPtr.empty()) {
878 NPtr = new LoadInst(FieldGlobals[0], Ptr->getName()+".f0", Ptr);
879 InsertedLoadsForPtr.push_back(Ptr);
881 NPtr = InsertedLoadsForPtr.back();
884 Value *New = new ICmpInst(SCI->getPredicate(), NPtr,
885 Constant::getNullValue(NPtr->getType()),
886 SCI->getName(), SCI);
887 SCI->replaceAllUsesWith(New);
888 SCI->eraseFromParent();
892 // Otherwise, this should be: 'getelementptr Ptr, Idx, uint FieldNo ...'
893 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
894 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
895 && "Unexpected GEPI!");
897 // Load the pointer for this field.
898 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
899 if (InsertedLoadsForPtr.size() <= FieldNo)
900 InsertedLoadsForPtr.resize(FieldNo+1);
901 if (InsertedLoadsForPtr[FieldNo] == 0)
902 InsertedLoadsForPtr[FieldNo] = new LoadInst(FieldGlobals[FieldNo],
903 Ptr->getName()+".f" +
904 utostr(FieldNo), Ptr);
905 Value *NewPtr = InsertedLoadsForPtr[FieldNo];
907 // Create the new GEP idx vector.
908 std::vector<Value*> GEPIdx;
909 GEPIdx.push_back(GEPI->getOperand(1));
910 GEPIdx.insert(GEPIdx.end(), GEPI->op_begin()+3, GEPI->op_end());
912 Value *NGEPI = new GetElementPtrInst(NewPtr, GEPIdx, GEPI->getName(), GEPI);
913 GEPI->replaceAllUsesWith(NGEPI);
914 GEPI->eraseFromParent();
918 /// PerformHeapAllocSRoA - MI is an allocation of an array of structures. Break
919 /// it up into multiple allocations of arrays of the fields.
920 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){
921 DOUT << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *MI;
922 const StructType *STy = cast<StructType>(MI->getAllocatedType());
924 // There is guaranteed to be at least one use of the malloc (storing
925 // it into GV). If there are other uses, change them to be uses of
926 // the global to simplify later code. This also deletes the store
928 ReplaceUsesOfMallocWithGlobal(MI, GV);
930 // Okay, at this point, there are no users of the malloc. Insert N
931 // new mallocs at the same place as MI, and N globals.
932 std::vector<GlobalVariable*> FieldGlobals;
933 std::vector<MallocInst*> FieldMallocs;
935 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
936 const Type *FieldTy = STy->getElementType(FieldNo);
937 const Type *PFieldTy = PointerType::get(FieldTy);
939 GlobalVariable *NGV =
940 new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage,
941 Constant::getNullValue(PFieldTy),
942 GV->getName() + ".f" + utostr(FieldNo), GV);
943 FieldGlobals.push_back(NGV);
945 MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(),
946 MI->getName() + ".f" + utostr(FieldNo),MI);
947 FieldMallocs.push_back(NMI);
948 new StoreInst(NMI, NGV, MI);
951 // The tricky aspect of this transformation is handling the case when malloc
952 // fails. In the original code, malloc failing would set the result pointer
953 // of malloc to null. In this case, some mallocs could succeed and others
954 // could fail. As such, we emit code that looks like this:
955 // F0 = malloc(field0)
956 // F1 = malloc(field1)
957 // F2 = malloc(field2)
958 // if (F0 == 0 || F1 == 0 || F2 == 0) {
959 // if (F0) { free(F0); F0 = 0; }
960 // if (F1) { free(F1); F1 = 0; }
961 // if (F2) { free(F2); F2 = 0; }
963 Value *RunningOr = 0;
964 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
965 Value *Cond = new ICmpInst(ICmpInst::ICMP_EQ, FieldMallocs[i],
966 Constant::getNullValue(FieldMallocs[i]->getType()),
969 RunningOr = Cond; // First seteq
971 RunningOr = BinaryOperator::createOr(RunningOr, Cond, "tmp", MI);
974 // Split the basic block at the old malloc.
975 BasicBlock *OrigBB = MI->getParent();
976 BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont");
978 // Create the block to check the first condition. Put all these blocks at the
979 // end of the function as they are unlikely to be executed.
980 BasicBlock *NullPtrBlock = new BasicBlock("malloc_ret_null",
981 OrigBB->getParent());
983 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
984 // branch on RunningOr.
985 OrigBB->getTerminator()->eraseFromParent();
986 new BranchInst(NullPtrBlock, ContBB, RunningOr, OrigBB);
988 // Within the NullPtrBlock, we need to emit a comparison and branch for each
989 // pointer, because some may be null while others are not.
990 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
991 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
992 Value *Cmp = new ICmpInst(ICmpInst::ICMP_NE, GVVal,
993 Constant::getNullValue(GVVal->getType()),
994 "tmp", NullPtrBlock);
995 BasicBlock *FreeBlock = new BasicBlock("free_it", OrigBB->getParent());
996 BasicBlock *NextBlock = new BasicBlock("next", OrigBB->getParent());
997 new BranchInst(FreeBlock, NextBlock, Cmp, NullPtrBlock);
999 // Fill in FreeBlock.
1000 new FreeInst(GVVal, FreeBlock);
1001 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1003 new BranchInst(NextBlock, FreeBlock);
1005 NullPtrBlock = NextBlock;
1008 new BranchInst(ContBB, NullPtrBlock);
1011 // MI is no longer needed, remove it.
1012 MI->eraseFromParent();
1015 // Okay, the malloc site is completely handled. All of the uses of GV are now
1016 // loads, and all uses of those loads are simple. Rewrite them to use loads
1017 // of the per-field globals instead.
1018 while (!GV->use_empty()) {
1019 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
1020 RewriteUsesOfLoadForHeapSRoA(LI, FieldGlobals);
1021 LI->eraseFromParent();
1023 // Must be a store of null.
1024 StoreInst *SI = cast<StoreInst>(GV->use_back());
1025 assert(isa<Constant>(SI->getOperand(0)) &&
1026 cast<Constant>(SI->getOperand(0))->isNullValue() &&
1027 "Unexpected heap-sra user!");
1029 // Insert a store of null into each global.
1030 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1032 Constant::getNullValue(FieldGlobals[i]->getType()->getElementType());
1033 new StoreInst(Null, FieldGlobals[i], SI);
1035 // Erase the original store.
1036 SI->eraseFromParent();
1040 // The old global is now dead, remove it.
1041 GV->eraseFromParent();
1044 return FieldGlobals[0];
1048 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1049 // that only one value (besides its initializer) is ever stored to the global.
1050 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1051 Module::global_iterator &GVI,
1053 if (CastInst *CI = dyn_cast<CastInst>(StoredOnceVal))
1054 StoredOnceVal = CI->getOperand(0);
1055 else if (GetElementPtrInst *GEPI =dyn_cast<GetElementPtrInst>(StoredOnceVal)){
1056 // "getelementptr Ptr, 0, 0, 0" is really just a cast.
1057 bool IsJustACast = true;
1058 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
1059 if (!isa<Constant>(GEPI->getOperand(i)) ||
1060 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
1061 IsJustACast = false;
1065 StoredOnceVal = GEPI->getOperand(0);
1068 // If we are dealing with a pointer global that is initialized to null and
1069 // only has one (non-null) value stored into it, then we can optimize any
1070 // users of the loaded value (often calls and loads) that would trap if the
1072 if (isa<PointerType>(GV->getInitializer()->getType()) &&
1073 GV->getInitializer()->isNullValue()) {
1074 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1075 if (GV->getInitializer()->getType() != SOVC->getType())
1076 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1078 // Optimize away any trapping uses of the loaded value.
1079 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1081 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) {
1082 // If this is a malloc of an abstract type, don't touch it.
1083 if (!MI->getAllocatedType()->isSized())
1086 // We can't optimize this global unless all uses of it are *known* to be
1087 // of the malloc value, not of the null initializer value (consider a use
1088 // that compares the global's value against zero to see if the malloc has
1089 // been reached). To do this, we check to see if all uses of the global
1090 // would trap if the global were null: this proves that they must all
1091 // happen after the malloc.
1092 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1095 // We can't optimize this if the malloc itself is used in a complex way,
1096 // for example, being stored into multiple globals. This allows the
1097 // malloc to be stored into the specified global, loaded setcc'd, and
1098 // GEP'd. These are all things we could transform to using the global
1100 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV))
1104 // If we have a global that is only initialized with a fixed size malloc,
1105 // transform the program to use global memory instead of malloc'd memory.
1106 // This eliminates dynamic allocation, avoids an indirection accessing the
1107 // data, and exposes the resultant global to further GlobalOpt.
1108 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) {
1109 // Restrict this transformation to only working on small allocations
1110 // (2048 bytes currently), as we don't want to introduce a 16M global or
1112 if (NElements->getZExtValue()*
1113 TD.getTypeSize(MI->getAllocatedType()) < 2048) {
1114 GVI = OptimizeGlobalAddressOfMalloc(GV, MI);
1119 // If the allocation is an array of structures, consider transforming this
1120 // into multiple malloc'd arrays, one for each field. This is basically
1121 // SRoA for malloc'd memory.
1122 if (const StructType *AllocTy =
1123 dyn_cast<StructType>(MI->getAllocatedType())) {
1124 // This the structure has an unreasonable number of fields, leave it
1126 if (AllocTy->getNumElements() <= 16 && AllocTy->getNumElements() > 0 &&
1127 GlobalLoadUsesSimpleEnoughForHeapSRA(GV)) {
1128 GVI = PerformHeapAllocSRoA(GV, MI);
1138 /// ShrinkGlobalToBoolean - At this point, we have learned that the only two
1139 /// values ever stored into GV are its initializer and OtherVal.
1140 static void ShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1141 // Create the new global, initializing it to false.
1142 GlobalVariable *NewGV = new GlobalVariable(Type::BoolTy, false,
1143 GlobalValue::InternalLinkage, ConstantInt::getFalse(),
1144 GV->getName()+".b");
1145 GV->getParent()->getGlobalList().insert(GV, NewGV);
1147 Constant *InitVal = GV->getInitializer();
1148 assert(InitVal->getType() != Type::BoolTy && "No reason to shrink to bool!");
1150 // If initialized to zero and storing one into the global, we can use a cast
1151 // instead of a select to synthesize the desired value.
1152 bool IsOneZero = false;
1153 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1154 IsOneZero = InitVal->isNullValue() && CI->equalsInt(1);
1156 while (!GV->use_empty()) {
1157 Instruction *UI = cast<Instruction>(GV->use_back());
1158 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1159 // Change the store into a boolean store.
1160 bool StoringOther = SI->getOperand(0) == OtherVal;
1161 // Only do this if we weren't storing a loaded value.
1163 if (StoringOther || SI->getOperand(0) == InitVal)
1164 StoreVal = ConstantInt::get(StoringOther);
1166 // Otherwise, we are storing a previously loaded copy. To do this,
1167 // change the copy from copying the original value to just copying the
1169 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1171 // If we're already replaced the input, StoredVal will be a cast or
1172 // select instruction. If not, it will be a load of the original
1174 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1175 assert(LI->getOperand(0) == GV && "Not a copy!");
1176 // Insert a new load, to preserve the saved value.
1177 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1179 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1180 "This is not a form that we understand!");
1181 StoreVal = StoredVal->getOperand(0);
1182 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1185 new StoreInst(StoreVal, NewGV, SI);
1186 } else if (!UI->use_empty()) {
1187 // Change the load into a load of bool then a select.
1188 LoadInst *LI = cast<LoadInst>(UI);
1190 std::string Name = LI->getName(); LI->setName("");
1191 LoadInst *NLI = new LoadInst(NewGV, Name+".b", LI);
1194 NSI = new ZExtInst(NLI, LI->getType(), Name, LI);
1196 NSI = new SelectInst(NLI, OtherVal, InitVal, Name, LI);
1197 LI->replaceAllUsesWith(NSI);
1199 UI->eraseFromParent();
1202 GV->eraseFromParent();
1206 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1207 /// it if possible. If we make a change, return true.
1208 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1209 Module::global_iterator &GVI) {
1210 std::set<PHINode*> PHIUsers;
1212 GV->removeDeadConstantUsers();
1214 if (GV->use_empty()) {
1215 DOUT << "GLOBAL DEAD: " << *GV;
1216 GV->eraseFromParent();
1221 if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
1223 cerr << "Global: " << *GV;
1224 cerr << " isLoaded = " << GS.isLoaded << "\n";
1225 cerr << " StoredType = ";
1226 switch (GS.StoredType) {
1227 case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break;
1228 case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break;
1229 case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break;
1230 case GlobalStatus::isStored: cerr << "stored\n"; break;
1232 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
1233 cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n";
1234 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
1235 cerr << " AccessingFunction = " << GS.AccessingFunction->getName()
1237 cerr << " HasMultipleAccessingFunctions = "
1238 << GS.HasMultipleAccessingFunctions << "\n";
1239 cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n";
1240 cerr << " isNotSuitableForSRA = " << GS.isNotSuitableForSRA << "\n";
1244 // If this is a first class global and has only one accessing function
1245 // and this function is main (which we know is not recursive we can make
1246 // this global a local variable) we replace the global with a local alloca
1247 // in this function.
1249 // NOTE: It doesn't make sense to promote non first class types since we
1250 // are just replacing static memory to stack memory.
1251 if (!GS.HasMultipleAccessingFunctions &&
1252 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1253 GV->getType()->getElementType()->isFirstClassType() &&
1254 GS.AccessingFunction->getName() == "main" &&
1255 GS.AccessingFunction->hasExternalLinkage()) {
1256 DOUT << "LOCALIZING GLOBAL: " << *GV;
1257 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin();
1258 const Type* ElemTy = GV->getType()->getElementType();
1259 // FIXME: Pass Global's alignment when globals have alignment
1260 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI);
1261 if (!isa<UndefValue>(GV->getInitializer()))
1262 new StoreInst(GV->getInitializer(), Alloca, FirstI);
1264 GV->replaceAllUsesWith(Alloca);
1265 GV->eraseFromParent();
1270 // If the global is never loaded (but may be stored to), it is dead.
1273 DOUT << "GLOBAL NEVER LOADED: " << *GV;
1275 // Delete any stores we can find to the global. We may not be able to
1276 // make it completely dead though.
1277 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1279 // If the global is dead now, delete it.
1280 if (GV->use_empty()) {
1281 GV->eraseFromParent();
1287 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1288 DOUT << "MARKING CONSTANT: " << *GV;
1289 GV->setConstant(true);
1291 // Clean up any obviously simplifiable users now.
1292 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1294 // If the global is dead now, just nuke it.
1295 if (GV->use_empty()) {
1296 DOUT << " *** Marking constant allowed us to simplify "
1297 << "all users and delete global!\n";
1298 GV->eraseFromParent();
1304 } else if (!GS.isNotSuitableForSRA &&
1305 !GV->getInitializer()->getType()->isFirstClassType()) {
1306 if (GlobalVariable *FirstNewGV = SRAGlobal(GV)) {
1307 GVI = FirstNewGV; // Don't skip the newly produced globals!
1310 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1311 // If the initial value for the global was an undef value, and if only
1312 // one other value was stored into it, we can just change the
1313 // initializer to be an undef value, then delete all stores to the
1314 // global. This allows us to mark it constant.
1315 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1316 if (isa<UndefValue>(GV->getInitializer())) {
1317 // Change the initial value here.
1318 GV->setInitializer(SOVConstant);
1320 // Clean up any obviously simplifiable users now.
1321 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1323 if (GV->use_empty()) {
1324 DOUT << " *** Substituting initializer allowed us to "
1325 << "simplify all users and delete global!\n";
1326 GV->eraseFromParent();
1335 // Try to optimize globals based on the knowledge that only one value
1336 // (besides its initializer) is ever stored to the global.
1337 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1338 getAnalysis<TargetData>()))
1341 // Otherwise, if the global was not a boolean, we can shrink it to be a
1343 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1344 if (GV->getType()->getElementType() != Type::BoolTy &&
1345 !GV->getType()->getElementType()->isFloatingPoint() &&
1347 DOUT << " *** SHRINKING TO BOOL: " << *GV;
1348 ShrinkGlobalToBoolean(GV, SOVConstant);
1357 /// OnlyCalledDirectly - Return true if the specified function is only called
1358 /// directly. In other words, its address is never taken.
1359 static bool OnlyCalledDirectly(Function *F) {
1360 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1361 Instruction *User = dyn_cast<Instruction>(*UI);
1362 if (!User) return false;
1363 if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false;
1365 // See if the function address is passed as an argument.
1366 for (unsigned i = 1, e = User->getNumOperands(); i != e; ++i)
1367 if (User->getOperand(i) == F) return false;
1372 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1373 /// function, changing them to FastCC.
1374 static void ChangeCalleesToFastCall(Function *F) {
1375 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1376 Instruction *User = cast<Instruction>(*UI);
1377 if (CallInst *CI = dyn_cast<CallInst>(User))
1378 CI->setCallingConv(CallingConv::Fast);
1380 cast<InvokeInst>(User)->setCallingConv(CallingConv::Fast);
1384 bool GlobalOpt::OptimizeFunctions(Module &M) {
1385 bool Changed = false;
1386 // Optimize functions.
1387 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1389 F->removeDeadConstantUsers();
1390 if (F->use_empty() && (F->hasInternalLinkage() ||
1391 F->hasLinkOnceLinkage())) {
1392 M.getFunctionList().erase(F);
1395 } else if (F->hasInternalLinkage() &&
1396 F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1397 OnlyCalledDirectly(F)) {
1398 // If this function has C calling conventions, is not a varargs
1399 // function, and is only called directly, promote it to use the Fast
1400 // calling convention.
1401 F->setCallingConv(CallingConv::Fast);
1402 ChangeCalleesToFastCall(F);
1410 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1411 bool Changed = false;
1412 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1414 GlobalVariable *GV = GVI++;
1415 if (!GV->isConstant() && GV->hasInternalLinkage() &&
1416 GV->hasInitializer())
1417 Changed |= ProcessInternalGlobal(GV, GVI);
1422 /// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1423 /// initializers have an init priority of 65535.
1424 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1425 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1427 if (I->getName() == "llvm.global_ctors") {
1428 // Found it, verify it's an array of { int, void()* }.
1429 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1431 const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1432 if (!STy || STy->getNumElements() != 2 ||
1433 STy->getElementType(0) != Type::Int32Ty) return 0;
1434 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1435 if (!PFTy) return 0;
1436 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1437 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() ||
1438 FTy->getNumParams() != 0)
1441 // Verify that the initializer is simple enough for us to handle.
1442 if (!I->hasInitializer()) return 0;
1443 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1445 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1446 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(CA->getOperand(i))) {
1447 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1450 // Must have a function or null ptr.
1451 if (!isa<Function>(CS->getOperand(1)))
1454 // Init priority must be standard.
1455 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1456 if (!CI || CI->getZExtValue() != 65535)
1467 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1468 /// return a list of the functions and null terminator as a vector.
1469 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1470 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1471 std::vector<Function*> Result;
1472 Result.reserve(CA->getNumOperands());
1473 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) {
1474 ConstantStruct *CS = cast<ConstantStruct>(CA->getOperand(i));
1475 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1480 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1481 /// specified array, returning the new global to use.
1482 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1483 const std::vector<Function*> &Ctors) {
1484 // If we made a change, reassemble the initializer list.
1485 std::vector<Constant*> CSVals;
1486 CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535));
1487 CSVals.push_back(0);
1489 // Create the new init list.
1490 std::vector<Constant*> CAList;
1491 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1493 CSVals[1] = Ctors[i];
1495 const Type *FTy = FunctionType::get(Type::VoidTy,
1496 std::vector<const Type*>(), false);
1497 const PointerType *PFTy = PointerType::get(FTy);
1498 CSVals[1] = Constant::getNullValue(PFTy);
1499 CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647);
1501 CAList.push_back(ConstantStruct::get(CSVals));
1504 // Create the array initializer.
1505 const Type *StructTy =
1506 cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
1507 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()),
1510 // If we didn't change the number of elements, don't create a new GV.
1511 if (CA->getType() == GCL->getInitializer()->getType()) {
1512 GCL->setInitializer(CA);
1516 // Create the new global and insert it next to the existing list.
1517 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
1518 GCL->getLinkage(), CA,
1521 GCL->getParent()->getGlobalList().insert(GCL, NGV);
1523 // Nuke the old list, replacing any uses with the new one.
1524 if (!GCL->use_empty()) {
1526 if (V->getType() != GCL->getType())
1527 V = ConstantExpr::getBitCast(V, GCL->getType());
1528 GCL->replaceAllUsesWith(V);
1530 GCL->eraseFromParent();
1539 static Constant *getVal(std::map<Value*, Constant*> &ComputedValues,
1541 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
1542 Constant *R = ComputedValues[V];
1543 assert(R && "Reference to an uncomputed value!");
1547 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
1548 /// enough for us to understand. In particular, if it is a cast of something,
1549 /// we punt. We basically just support direct accesses to globals and GEP's of
1550 /// globals. This should be kept up to date with CommitValueTo.
1551 static bool isSimpleEnoughPointerToCommit(Constant *C) {
1552 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
1553 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1554 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1555 return !GV->isExternal(); // reject external globals.
1557 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
1558 // Handle a constantexpr gep.
1559 if (CE->getOpcode() == Instruction::GetElementPtr &&
1560 isa<GlobalVariable>(CE->getOperand(0))) {
1561 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1562 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1563 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1564 return GV->hasInitializer() &&
1565 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1570 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
1571 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
1572 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
1573 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
1574 ConstantExpr *Addr, unsigned OpNo) {
1575 // Base case of the recursion.
1576 if (OpNo == Addr->getNumOperands()) {
1577 assert(Val->getType() == Init->getType() && "Type mismatch!");
1581 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
1582 std::vector<Constant*> Elts;
1584 // Break up the constant into its elements.
1585 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
1586 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i)
1587 Elts.push_back(CS->getOperand(i));
1588 } else if (isa<ConstantAggregateZero>(Init)) {
1589 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1590 Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
1591 } else if (isa<UndefValue>(Init)) {
1592 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1593 Elts.push_back(UndefValue::get(STy->getElementType(i)));
1595 assert(0 && "This code is out of sync with "
1596 " ConstantFoldLoadThroughGEPConstantExpr");
1599 // Replace the element that we are supposed to.
1600 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
1601 unsigned Idx = CU->getZExtValue();
1602 assert(Idx < STy->getNumElements() && "Struct index out of range!");
1603 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
1605 // Return the modified struct.
1606 return ConstantStruct::get(Elts);
1608 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
1609 const ArrayType *ATy = cast<ArrayType>(Init->getType());
1611 // Break up the array into elements.
1612 std::vector<Constant*> Elts;
1613 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
1614 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1615 Elts.push_back(CA->getOperand(i));
1616 } else if (isa<ConstantAggregateZero>(Init)) {
1617 Constant *Elt = Constant::getNullValue(ATy->getElementType());
1618 Elts.assign(ATy->getNumElements(), Elt);
1619 } else if (isa<UndefValue>(Init)) {
1620 Constant *Elt = UndefValue::get(ATy->getElementType());
1621 Elts.assign(ATy->getNumElements(), Elt);
1623 assert(0 && "This code is out of sync with "
1624 " ConstantFoldLoadThroughGEPConstantExpr");
1627 assert(CI->getZExtValue() < ATy->getNumElements());
1628 Elts[CI->getZExtValue()] =
1629 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
1630 return ConstantArray::get(ATy, Elts);
1634 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
1635 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
1636 static void CommitValueTo(Constant *Val, Constant *Addr) {
1637 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
1638 assert(GV->hasInitializer());
1639 GV->setInitializer(Val);
1643 ConstantExpr *CE = cast<ConstantExpr>(Addr);
1644 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1646 Constant *Init = GV->getInitializer();
1647 Init = EvaluateStoreInto(Init, Val, CE, 2);
1648 GV->setInitializer(Init);
1651 /// ComputeLoadResult - Return the value that would be computed by a load from
1652 /// P after the stores reflected by 'memory' have been performed. If we can't
1653 /// decide, return null.
1654 static Constant *ComputeLoadResult(Constant *P,
1655 const std::map<Constant*, Constant*> &Memory) {
1656 // If this memory location has been recently stored, use the stored value: it
1657 // is the most up-to-date.
1658 std::map<Constant*, Constant*>::const_iterator I = Memory.find(P);
1659 if (I != Memory.end()) return I->second;
1662 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
1663 if (GV->hasInitializer())
1664 return GV->getInitializer();
1668 // Handle a constantexpr getelementptr.
1669 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
1670 if (CE->getOpcode() == Instruction::GetElementPtr &&
1671 isa<GlobalVariable>(CE->getOperand(0))) {
1672 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1673 if (GV->hasInitializer())
1674 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1677 return 0; // don't know how to evaluate.
1680 /// EvaluateFunction - Evaluate a call to function F, returning true if
1681 /// successful, false if we can't evaluate it. ActualArgs contains the formal
1682 /// arguments for the function.
1683 static bool EvaluateFunction(Function *F, Constant *&RetVal,
1684 const std::vector<Constant*> &ActualArgs,
1685 std::vector<Function*> &CallStack,
1686 std::map<Constant*, Constant*> &MutatedMemory,
1687 std::vector<GlobalVariable*> &AllocaTmps) {
1688 // Check to see if this function is already executing (recursion). If so,
1689 // bail out. TODO: we might want to accept limited recursion.
1690 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
1693 CallStack.push_back(F);
1695 /// Values - As we compute SSA register values, we store their contents here.
1696 std::map<Value*, Constant*> Values;
1698 // Initialize arguments to the incoming values specified.
1700 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
1702 Values[AI] = ActualArgs[ArgNo];
1704 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
1705 /// we can only evaluate any one basic block at most once. This set keeps
1706 /// track of what we have executed so we can detect recursive cases etc.
1707 std::set<BasicBlock*> ExecutedBlocks;
1709 // CurInst - The current instruction we're evaluating.
1710 BasicBlock::iterator CurInst = F->begin()->begin();
1712 // This is the main evaluation loop.
1714 Constant *InstResult = 0;
1716 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
1717 if (SI->isVolatile()) return false; // no volatile accesses.
1718 Constant *Ptr = getVal(Values, SI->getOperand(1));
1719 if (!isSimpleEnoughPointerToCommit(Ptr))
1720 // If this is too complex for us to commit, reject it.
1722 Constant *Val = getVal(Values, SI->getOperand(0));
1723 MutatedMemory[Ptr] = Val;
1724 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
1725 InstResult = ConstantExpr::get(BO->getOpcode(),
1726 getVal(Values, BO->getOperand(0)),
1727 getVal(Values, BO->getOperand(1)));
1728 } else if (ShiftInst *SI = dyn_cast<ShiftInst>(CurInst)) {
1729 InstResult = ConstantExpr::get(SI->getOpcode(),
1730 getVal(Values, SI->getOperand(0)),
1731 getVal(Values, SI->getOperand(1)));
1732 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
1733 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
1734 getVal(Values, CI->getOperand(0)),
1735 getVal(Values, CI->getOperand(1)));
1736 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
1737 InstResult = ConstantExpr::getCast(CI->getOpcode(),
1738 getVal(Values, CI->getOperand(0)),
1740 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
1741 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
1742 getVal(Values, SI->getOperand(1)),
1743 getVal(Values, SI->getOperand(2)));
1744 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
1745 Constant *P = getVal(Values, GEP->getOperand(0));
1746 std::vector<Constant*> GEPOps;
1747 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
1748 GEPOps.push_back(getVal(Values, GEP->getOperand(i)));
1749 InstResult = ConstantExpr::getGetElementPtr(P, GEPOps);
1750 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
1751 if (LI->isVolatile()) return false; // no volatile accesses.
1752 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
1754 if (InstResult == 0) return false; // Could not evaluate load.
1755 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
1756 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs.
1757 const Type *Ty = AI->getType()->getElementType();
1758 AllocaTmps.push_back(new GlobalVariable(Ty, false,
1759 GlobalValue::InternalLinkage,
1760 UndefValue::get(Ty),
1762 InstResult = AllocaTmps.back();
1763 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
1764 // Cannot handle inline asm.
1765 if (isa<InlineAsm>(CI->getOperand(0))) return false;
1767 // Resolve function pointers.
1768 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0)));
1769 if (!Callee) return false; // Cannot resolve.
1771 std::vector<Constant*> Formals;
1772 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
1773 Formals.push_back(getVal(Values, CI->getOperand(i)));
1775 if (Callee->isExternal()) {
1776 // If this is a function we can constant fold, do it.
1777 if (Constant *C = ConstantFoldCall(Callee, Formals)) {
1783 if (Callee->getFunctionType()->isVarArg())
1788 // Execute the call, if successful, use the return value.
1789 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
1790 MutatedMemory, AllocaTmps))
1792 InstResult = RetVal;
1794 } else if (isa<TerminatorInst>(CurInst)) {
1795 BasicBlock *NewBB = 0;
1796 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
1797 if (BI->isUnconditional()) {
1798 NewBB = BI->getSuccessor(0);
1801 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
1803 // Cannot determine.
1804 if (!Cond || Cond->getType() != Type::BoolTy)
1806 NewBB = BI->getSuccessor(!Cond->getBoolValue());
1808 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
1810 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
1811 if (!Val) return false; // Cannot determine.
1812 NewBB = SI->getSuccessor(SI->findCaseValue(Val));
1813 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
1814 if (RI->getNumOperands())
1815 RetVal = getVal(Values, RI->getOperand(0));
1817 CallStack.pop_back(); // return from fn.
1818 return true; // We succeeded at evaluating this ctor!
1820 // invoke, unwind, unreachable.
1821 return false; // Cannot handle this terminator.
1824 // Okay, we succeeded in evaluating this control flow. See if we have
1825 // executed the new block before. If so, we have a looping function,
1826 // which we cannot evaluate in reasonable time.
1827 if (!ExecutedBlocks.insert(NewBB).second)
1828 return false; // looped!
1830 // Okay, we have never been in this block before. Check to see if there
1831 // are any PHI nodes. If so, evaluate them with information about where
1833 BasicBlock *OldBB = CurInst->getParent();
1834 CurInst = NewBB->begin();
1836 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
1837 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
1839 // Do NOT increment CurInst. We know that the terminator had no value.
1842 // Did not know how to evaluate this!
1846 if (!CurInst->use_empty())
1847 Values[CurInst] = InstResult;
1849 // Advance program counter.
1854 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
1855 /// we can. Return true if we can, false otherwise.
1856 static bool EvaluateStaticConstructor(Function *F) {
1857 /// MutatedMemory - For each store we execute, we update this map. Loads
1858 /// check this to get the most up-to-date value. If evaluation is successful,
1859 /// this state is committed to the process.
1860 std::map<Constant*, Constant*> MutatedMemory;
1862 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
1863 /// to represent its body. This vector is needed so we can delete the
1864 /// temporary globals when we are done.
1865 std::vector<GlobalVariable*> AllocaTmps;
1867 /// CallStack - This is used to detect recursion. In pathological situations
1868 /// we could hit exponential behavior, but at least there is nothing
1870 std::vector<Function*> CallStack;
1872 // Call the function.
1873 Constant *RetValDummy;
1874 bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(),
1875 CallStack, MutatedMemory, AllocaTmps);
1877 // We succeeded at evaluation: commit the result.
1878 DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
1879 << F->getName() << "' to " << MutatedMemory.size()
1881 for (std::map<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
1882 E = MutatedMemory.end(); I != E; ++I)
1883 CommitValueTo(I->second, I->first);
1886 // At this point, we are done interpreting. If we created any 'alloca'
1887 // temporaries, release them now.
1888 while (!AllocaTmps.empty()) {
1889 GlobalVariable *Tmp = AllocaTmps.back();
1890 AllocaTmps.pop_back();
1892 // If there are still users of the alloca, the program is doing something
1893 // silly, e.g. storing the address of the alloca somewhere and using it
1894 // later. Since this is undefined, we'll just make it be null.
1895 if (!Tmp->use_empty())
1896 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
1905 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
1906 /// Return true if anything changed.
1907 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
1908 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
1909 bool MadeChange = false;
1910 if (Ctors.empty()) return false;
1912 // Loop over global ctors, optimizing them when we can.
1913 for (unsigned i = 0; i != Ctors.size(); ++i) {
1914 Function *F = Ctors[i];
1915 // Found a null terminator in the middle of the list, prune off the rest of
1918 if (i != Ctors.size()-1) {
1925 // We cannot simplify external ctor functions.
1926 if (F->empty()) continue;
1928 // If we can evaluate the ctor at compile time, do.
1929 if (EvaluateStaticConstructor(F)) {
1930 Ctors.erase(Ctors.begin()+i);
1933 ++NumCtorsEvaluated;
1938 if (!MadeChange) return false;
1940 GCL = InstallGlobalCtors(GCL, Ctors);
1945 bool GlobalOpt::runOnModule(Module &M) {
1946 bool Changed = false;
1948 // Try to find the llvm.globalctors list.
1949 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
1951 bool LocalChange = true;
1952 while (LocalChange) {
1953 LocalChange = false;
1955 // Delete functions that are trivially dead, ccc -> fastcc
1956 LocalChange |= OptimizeFunctions(M);
1958 // Optimize global_ctors list.
1960 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
1962 // Optimize non-address-taken globals.
1963 LocalChange |= OptimizeGlobalVars(M);
1964 Changed |= LocalChange;
1967 // TODO: Move all global ctors functions to the end of the module for code