1 //===- GlobalOpt.cpp - Optimize Global Variables --------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // 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/SmallPtrSet.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/ADT/StringExtras.h"
37 STATISTIC(NumMarked , "Number of globals marked constant");
38 STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
39 STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
40 STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
41 STATISTIC(NumDeleted , "Number of globals deleted");
42 STATISTIC(NumFnDeleted , "Number of functions deleted");
43 STATISTIC(NumGlobUses , "Number of global uses devirtualized");
44 STATISTIC(NumLocalized , "Number of globals localized");
45 STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans");
46 STATISTIC(NumFastCallFns , "Number of functions converted to fastcc");
47 STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
50 struct VISIBILITY_HIDDEN GlobalOpt : public ModulePass {
51 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
52 AU.addRequired<TargetData>();
54 static char ID; // Pass identification, replacement for typeid
55 GlobalOpt() : ModulePass((intptr_t)&ID) {}
57 bool runOnModule(Module &M);
60 GlobalVariable *FindGlobalCtors(Module &M);
61 bool OptimizeFunctions(Module &M);
62 bool OptimizeGlobalVars(Module &M);
63 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
64 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
67 char GlobalOpt::ID = 0;
68 RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer");
71 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
73 /// GlobalStatus - As we analyze each global, keep track of some information
74 /// about it. If we find out that the address of the global is taken, none of
75 /// this info will be accurate.
76 struct VISIBILITY_HIDDEN GlobalStatus {
77 /// isLoaded - True if the global is ever loaded. If the global isn't ever
78 /// loaded it can be deleted.
81 /// StoredType - Keep track of what stores to the global look like.
84 /// NotStored - There is no store to this global. It can thus be marked
88 /// isInitializerStored - This global is stored to, but the only thing
89 /// stored is the constant it was initialized with. This is only tracked
90 /// for scalar globals.
93 /// isStoredOnce - This global is stored to, but only its initializer and
94 /// one other value is ever stored to it. If this global isStoredOnce, we
95 /// track the value stored to it in StoredOnceValue below. This is only
96 /// tracked for scalar globals.
99 /// isStored - This global is stored to by multiple values or something else
100 /// that we cannot track.
104 /// StoredOnceValue - If only one value (besides the initializer constant) is
105 /// ever stored to this global, keep track of what value it is.
106 Value *StoredOnceValue;
108 /// AccessingFunction/HasMultipleAccessingFunctions - These start out
109 /// null/false. When the first accessing function is noticed, it is recorded.
110 /// When a second different accessing function is noticed,
111 /// HasMultipleAccessingFunctions is set to true.
112 Function *AccessingFunction;
113 bool HasMultipleAccessingFunctions;
115 /// HasNonInstructionUser - Set to true if this global has a user that is not
116 /// an instruction (e.g. a constant expr or GV initializer).
117 bool HasNonInstructionUser;
119 /// HasPHIUser - Set to true if this global has a user that is a PHI node.
122 /// isNotSuitableForSRA - Keep track of whether any SRA preventing users of
123 /// the global exist. Such users include GEP instruction with variable
124 /// indexes, and non-gep/load/store users like constant expr casts.
125 bool isNotSuitableForSRA;
127 GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0),
128 AccessingFunction(0), HasMultipleAccessingFunctions(false),
129 HasNonInstructionUser(false), HasPHIUser(false),
130 isNotSuitableForSRA(false) {}
135 /// ConstantIsDead - Return true if the specified constant is (transitively)
136 /// dead. The constant may be used by other constants (e.g. constant arrays and
137 /// constant exprs) as long as they are dead, but it cannot be used by anything
139 static bool ConstantIsDead(Constant *C) {
140 if (isa<GlobalValue>(C)) return false;
142 for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI)
143 if (Constant *CU = dyn_cast<Constant>(*UI)) {
144 if (!ConstantIsDead(CU)) return false;
151 /// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
152 /// structure. If the global has its address taken, return true to indicate we
153 /// can't do anything with it.
155 static bool AnalyzeGlobal(Value *V, GlobalStatus &GS,
156 std::set<PHINode*> &PHIUsers) {
157 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
158 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
159 GS.HasNonInstructionUser = true;
161 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
162 if (CE->getOpcode() != Instruction::GetElementPtr)
163 GS.isNotSuitableForSRA = true;
164 else if (!GS.isNotSuitableForSRA) {
165 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
166 // don't like < 3 operand CE's, and we don't like non-constant integer
168 if (CE->getNumOperands() < 3 || !CE->getOperand(1)->isNullValue())
169 GS.isNotSuitableForSRA = true;
171 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
172 if (!isa<ConstantInt>(CE->getOperand(i))) {
173 GS.isNotSuitableForSRA = true;
179 } else if (Instruction *I = dyn_cast<Instruction>(*UI)) {
180 if (!GS.HasMultipleAccessingFunctions) {
181 Function *F = I->getParent()->getParent();
182 if (GS.AccessingFunction == 0)
183 GS.AccessingFunction = F;
184 else if (GS.AccessingFunction != F)
185 GS.HasMultipleAccessingFunctions = true;
187 if (isa<LoadInst>(I)) {
189 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
190 // Don't allow a store OF the address, only stores TO the address.
191 if (SI->getOperand(0) == V) return true;
193 // If this is a direct store to the global (i.e., the global is a scalar
194 // value, not an aggregate), keep more specific information about
196 if (GS.StoredType != GlobalStatus::isStored)
197 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){
198 Value *StoredVal = SI->getOperand(0);
199 if (StoredVal == GV->getInitializer()) {
200 if (GS.StoredType < GlobalStatus::isInitializerStored)
201 GS.StoredType = GlobalStatus::isInitializerStored;
202 } else if (isa<LoadInst>(StoredVal) &&
203 cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
205 if (GS.StoredType < GlobalStatus::isInitializerStored)
206 GS.StoredType = GlobalStatus::isInitializerStored;
207 } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
208 GS.StoredType = GlobalStatus::isStoredOnce;
209 GS.StoredOnceValue = StoredVal;
210 } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
211 GS.StoredOnceValue == StoredVal) {
214 GS.StoredType = GlobalStatus::isStored;
217 GS.StoredType = GlobalStatus::isStored;
219 } else if (isa<GetElementPtrInst>(I)) {
220 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
222 // If the first two indices are constants, this can be SRA'd.
223 if (isa<GlobalVariable>(I->getOperand(0))) {
224 if (I->getNumOperands() < 3 || !isa<Constant>(I->getOperand(1)) ||
225 !cast<Constant>(I->getOperand(1))->isNullValue() ||
226 !isa<ConstantInt>(I->getOperand(2)))
227 GS.isNotSuitableForSRA = true;
228 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(I->getOperand(0))){
229 if (CE->getOpcode() != Instruction::GetElementPtr ||
230 CE->getNumOperands() < 3 || I->getNumOperands() < 2 ||
231 !isa<Constant>(I->getOperand(0)) ||
232 !cast<Constant>(I->getOperand(0))->isNullValue())
233 GS.isNotSuitableForSRA = true;
235 GS.isNotSuitableForSRA = true;
237 } else if (isa<SelectInst>(I)) {
238 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
239 GS.isNotSuitableForSRA = true;
240 } else if (PHINode *PN = dyn_cast<PHINode>(I)) {
241 // PHI nodes we can check just like select or GEP instructions, but we
242 // have to be careful about infinite recursion.
243 if (PHIUsers.insert(PN).second) // Not already visited.
244 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
245 GS.isNotSuitableForSRA = true;
246 GS.HasPHIUser = true;
247 } else if (isa<CmpInst>(I)) {
248 GS.isNotSuitableForSRA = true;
249 } else if (isa<MemCpyInst>(I) || isa<MemMoveInst>(I)) {
250 if (I->getOperand(1) == V)
251 GS.StoredType = GlobalStatus::isStored;
252 if (I->getOperand(2) == V)
254 GS.isNotSuitableForSRA = true;
255 } else if (isa<MemSetInst>(I)) {
256 assert(I->getOperand(1) == V && "Memset only takes one pointer!");
257 GS.StoredType = GlobalStatus::isStored;
258 GS.isNotSuitableForSRA = true;
260 return true; // Any other non-load instruction might take address!
262 } else if (Constant *C = dyn_cast<Constant>(*UI)) {
263 GS.HasNonInstructionUser = true;
264 // We might have a dead and dangling constant hanging off of here.
265 if (!ConstantIsDead(C))
268 GS.HasNonInstructionUser = true;
269 // Otherwise must be some other user.
276 static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) {
277 ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
279 unsigned IdxV = CI->getZExtValue();
281 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
282 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
283 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
284 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
285 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) {
286 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
287 } else if (isa<ConstantAggregateZero>(Agg)) {
288 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
289 if (IdxV < STy->getNumElements())
290 return Constant::getNullValue(STy->getElementType(IdxV));
291 } else if (const SequentialType *STy =
292 dyn_cast<SequentialType>(Agg->getType())) {
293 return Constant::getNullValue(STy->getElementType());
295 } else if (isa<UndefValue>(Agg)) {
296 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
297 if (IdxV < STy->getNumElements())
298 return UndefValue::get(STy->getElementType(IdxV));
299 } else if (const SequentialType *STy =
300 dyn_cast<SequentialType>(Agg->getType())) {
301 return UndefValue::get(STy->getElementType());
308 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
309 /// users of the global, cleaning up the obvious ones. This is largely just a
310 /// quick scan over the use list to clean up the easy and obvious cruft. This
311 /// returns true if it made a change.
312 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
313 bool Changed = false;
314 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
317 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
319 // Replace the load with the initializer.
320 LI->replaceAllUsesWith(Init);
321 LI->eraseFromParent();
324 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
325 // Store must be unreachable or storing Init into the global.
326 SI->eraseFromParent();
328 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
329 if (CE->getOpcode() == Instruction::GetElementPtr) {
330 Constant *SubInit = 0;
332 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
333 Changed |= CleanupConstantGlobalUsers(CE, SubInit);
334 } else if (CE->getOpcode() == Instruction::BitCast &&
335 isa<PointerType>(CE->getType())) {
336 // Pointer cast, delete any stores and memsets to the global.
337 Changed |= CleanupConstantGlobalUsers(CE, 0);
340 if (CE->use_empty()) {
341 CE->destroyConstant();
344 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
345 // Do not transform "gepinst (gep constexpr (GV))" here, because forming
346 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
347 // and will invalidate our notion of what Init is.
348 Constant *SubInit = 0;
349 if (!isa<ConstantExpr>(GEP->getOperand(0))) {
351 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
352 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
353 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
355 Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
357 if (GEP->use_empty()) {
358 GEP->eraseFromParent();
361 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
362 if (MI->getRawDest() == V) {
363 MI->eraseFromParent();
367 } else if (Constant *C = dyn_cast<Constant>(U)) {
368 // If we have a chain of dead constantexprs or other things dangling from
369 // us, and if they are all dead, nuke them without remorse.
370 if (ConstantIsDead(C)) {
371 C->destroyConstant();
372 // This could have invalidated UI, start over from scratch.
373 CleanupConstantGlobalUsers(V, Init);
382 /// UsersSafeToSRA - Look at all uses of the global and decide whether it is
383 /// safe for us to perform this transformation.
385 static bool UsersSafeToSRA(Value *V) {
386 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
387 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
388 if (CE->getOpcode() != Instruction::GetElementPtr)
391 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
392 // don't like < 3 operand CE's, and we don't like non-constant integer
394 if (CE->getNumOperands() < 3 || !CE->getOperand(1)->isNullValue())
397 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
398 if (!isa<ConstantInt>(CE->getOperand(i)))
401 if (!UsersSafeToSRA(CE)) return false;
405 if (Instruction *I = dyn_cast<Instruction>(*UI)) {
406 if (isa<LoadInst>(I)) continue;
408 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
409 // Don't allow a store OF the address, only stores TO the address.
410 if (SI->getOperand(0) == V) return false;
414 if (isa<GetElementPtrInst>(I)) {
415 if (!UsersSafeToSRA(I)) return false;
417 // If the first two indices are constants, this can be SRA'd.
418 if (isa<GlobalVariable>(I->getOperand(0))) {
419 if (I->getNumOperands() < 3 || !isa<Constant>(I->getOperand(1)) ||
420 !cast<Constant>(I->getOperand(1))->isNullValue() ||
421 !isa<ConstantInt>(I->getOperand(2)))
426 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(I->getOperand(0))){
427 if (CE->getOpcode() != Instruction::GetElementPtr ||
428 CE->getNumOperands() < 3 || I->getNumOperands() < 2 ||
429 !isa<Constant>(I->getOperand(0)) ||
430 !cast<Constant>(I->getOperand(0))->isNullValue())
436 return false; // Any other instruction is not safe.
438 if (Constant *C = dyn_cast<Constant>(*UI)) {
439 // We might have a dead and dangling constant hanging off of here.
440 if (!ConstantIsDead(C))
444 // Otherwise must be some other user.
451 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
452 /// variable. This opens the door for other optimizations by exposing the
453 /// behavior of the program in a more fine-grained way. We have determined that
454 /// this transformation is safe already. We return the first global variable we
455 /// insert so that the caller can reprocess it.
456 static GlobalVariable *SRAGlobal(GlobalVariable *GV) {
457 // Make sure this global only has simple uses that we can SRA.
458 if (!UsersSafeToSRA(GV))
461 assert(GV->hasInternalLinkage() && !GV->isConstant());
462 Constant *Init = GV->getInitializer();
463 const Type *Ty = Init->getType();
465 std::vector<GlobalVariable*> NewGlobals;
466 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
468 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
469 NewGlobals.reserve(STy->getNumElements());
470 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
471 Constant *In = getAggregateConstantElement(Init,
472 ConstantInt::get(Type::Int32Ty, i));
473 assert(In && "Couldn't get element of initializer?");
474 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
475 GlobalVariable::InternalLinkage,
476 In, GV->getName()+"."+utostr(i),
478 GV->isThreadLocal());
479 Globals.insert(GV, NGV);
480 NewGlobals.push_back(NGV);
482 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
483 unsigned NumElements = 0;
484 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
485 NumElements = ATy->getNumElements();
486 else if (const VectorType *PTy = dyn_cast<VectorType>(STy))
487 NumElements = PTy->getNumElements();
489 assert(0 && "Unknown aggregate sequential type!");
491 if (NumElements > 16 && GV->hasNUsesOrMore(16))
492 return 0; // It's not worth it.
493 NewGlobals.reserve(NumElements);
494 for (unsigned i = 0, e = NumElements; i != e; ++i) {
495 Constant *In = getAggregateConstantElement(Init,
496 ConstantInt::get(Type::Int32Ty, i));
497 assert(In && "Couldn't get element of initializer?");
499 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
500 GlobalVariable::InternalLinkage,
501 In, GV->getName()+"."+utostr(i),
503 GV->isThreadLocal());
504 Globals.insert(GV, NGV);
505 NewGlobals.push_back(NGV);
509 if (NewGlobals.empty())
512 DOUT << "PERFORMING GLOBAL SRA ON: " << *GV;
514 Constant *NullInt = Constant::getNullValue(Type::Int32Ty);
516 // Loop over all of the uses of the global, replacing the constantexpr geps,
517 // with smaller constantexpr geps or direct references.
518 while (!GV->use_empty()) {
519 User *GEP = GV->use_back();
520 assert(((isa<ConstantExpr>(GEP) &&
521 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
522 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
524 // Ignore the 1th operand, which has to be zero or else the program is quite
525 // broken (undefined). Get the 2nd operand, which is the structure or array
527 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
528 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
530 Value *NewPtr = NewGlobals[Val];
532 // Form a shorter GEP if needed.
533 if (GEP->getNumOperands() > 3)
534 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
535 SmallVector<Constant*, 8> Idxs;
536 Idxs.push_back(NullInt);
537 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
538 Idxs.push_back(CE->getOperand(i));
539 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr),
540 &Idxs[0], Idxs.size());
542 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
543 SmallVector<Value*, 8> Idxs;
544 Idxs.push_back(NullInt);
545 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
546 Idxs.push_back(GEPI->getOperand(i));
547 NewPtr = new GetElementPtrInst(NewPtr, Idxs.begin(), Idxs.end(),
548 GEPI->getName()+"."+utostr(Val), GEPI);
550 GEP->replaceAllUsesWith(NewPtr);
552 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
553 GEPI->eraseFromParent();
555 cast<ConstantExpr>(GEP)->destroyConstant();
558 // Delete the old global, now that it is dead.
562 // Loop over the new globals array deleting any globals that are obviously
563 // dead. This can arise due to scalarization of a structure or an array that
564 // has elements that are dead.
565 unsigned FirstGlobal = 0;
566 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
567 if (NewGlobals[i]->use_empty()) {
568 Globals.erase(NewGlobals[i]);
569 if (FirstGlobal == i) ++FirstGlobal;
572 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
575 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
576 /// value will trap if the value is dynamically null. PHIs keeps track of any
577 /// phi nodes we've seen to avoid reprocessing them.
578 static bool AllUsesOfValueWillTrapIfNull(Value *V,
579 SmallPtrSet<PHINode*, 8> &PHIs) {
580 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
581 if (isa<LoadInst>(*UI)) {
583 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
584 if (SI->getOperand(0) == V) {
585 //cerr << "NONTRAPPING USE: " << **UI;
586 return false; // Storing the value.
588 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
589 if (CI->getOperand(0) != V) {
590 //cerr << "NONTRAPPING USE: " << **UI;
591 return false; // Not calling the ptr
593 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
594 if (II->getOperand(0) != V) {
595 //cerr << "NONTRAPPING USE: " << **UI;
596 return false; // Not calling the ptr
598 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(*UI)) {
599 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
600 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
601 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
602 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
603 // If we've already seen this phi node, ignore it, it has already been
606 return AllUsesOfValueWillTrapIfNull(PN, PHIs);
607 } else if (isa<ICmpInst>(*UI) &&
608 isa<ConstantPointerNull>(UI->getOperand(1))) {
609 // Ignore setcc X, null
611 //cerr << "NONTRAPPING USE: " << **UI;
617 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
618 /// from GV will trap if the loaded value is null. Note that this also permits
619 /// comparisons of the loaded value against null, as a special case.
620 static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) {
621 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI)
622 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
623 SmallPtrSet<PHINode*, 8> PHIs;
624 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
626 } else if (isa<StoreInst>(*UI)) {
627 // Ignore stores to the global.
629 // We don't know or understand this user, bail out.
630 //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI;
637 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
638 bool Changed = false;
639 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
640 Instruction *I = cast<Instruction>(*UI++);
641 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
642 LI->setOperand(0, NewV);
644 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
645 if (SI->getOperand(1) == V) {
646 SI->setOperand(1, NewV);
649 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
650 if (I->getOperand(0) == V) {
651 // Calling through the pointer! Turn into a direct call, but be careful
652 // that the pointer is not also being passed as an argument.
653 I->setOperand(0, NewV);
655 bool PassedAsArg = false;
656 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
657 if (I->getOperand(i) == V) {
659 I->setOperand(i, NewV);
663 // Being passed as an argument also. Be careful to not invalidate UI!
667 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
668 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
669 ConstantExpr::getCast(CI->getOpcode(),
670 NewV, CI->getType()));
671 if (CI->use_empty()) {
673 CI->eraseFromParent();
675 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
676 // Should handle GEP here.
677 SmallVector<Constant*, 8> Idxs;
678 Idxs.reserve(GEPI->getNumOperands()-1);
679 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
680 if (Constant *C = dyn_cast<Constant>(GEPI->getOperand(i)))
684 if (Idxs.size() == GEPI->getNumOperands()-1)
685 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
686 ConstantExpr::getGetElementPtr(NewV, &Idxs[0],
688 if (GEPI->use_empty()) {
690 GEPI->eraseFromParent();
699 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
700 /// value stored into it. If there are uses of the loaded value that would trap
701 /// if the loaded value is dynamically null, then we know that they cannot be
702 /// reachable with a null optimize away the load.
703 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) {
704 std::vector<LoadInst*> Loads;
705 bool Changed = false;
707 // Replace all uses of loads with uses of uses of the stored value.
708 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end();
710 if (LoadInst *LI = dyn_cast<LoadInst>(*GUI)) {
712 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
714 // If we get here we could have stores, selects, or phi nodes whose values
716 assert((isa<StoreInst>(*GUI) || isa<PHINode>(*GUI) ||
717 isa<SelectInst>(*GUI) || isa<ConstantExpr>(*GUI)) &&
718 "Only expect load and stores!");
722 DOUT << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV;
726 // Delete all of the loads we can, keeping track of whether we nuked them all!
727 bool AllLoadsGone = true;
728 while (!Loads.empty()) {
729 LoadInst *L = Loads.back();
730 if (L->use_empty()) {
731 L->eraseFromParent();
734 AllLoadsGone = false;
739 // If we nuked all of the loads, then none of the stores are needed either,
740 // nor is the global.
742 DOUT << " *** GLOBAL NOW DEAD!\n";
743 CleanupConstantGlobalUsers(GV, 0);
744 if (GV->use_empty()) {
745 GV->eraseFromParent();
753 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
754 /// instructions that are foldable.
755 static void ConstantPropUsersOf(Value *V) {
756 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
757 if (Instruction *I = dyn_cast<Instruction>(*UI++))
758 if (Constant *NewC = ConstantFoldInstruction(I)) {
759 I->replaceAllUsesWith(NewC);
761 // Advance UI to the next non-I use to avoid invalidating it!
762 // Instructions could multiply use V.
763 while (UI != E && *UI == I)
765 I->eraseFromParent();
769 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
770 /// variable, and transforms the program as if it always contained the result of
771 /// the specified malloc. Because it is always the result of the specified
772 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
773 /// malloc into a global, and any loads of GV as uses of the new global.
774 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
776 DOUT << "PROMOTING MALLOC GLOBAL: " << *GV << " MALLOC = " << *MI;
777 ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize());
779 if (NElements->getZExtValue() != 1) {
780 // If we have an array allocation, transform it to a single element
781 // allocation to make the code below simpler.
782 Type *NewTy = ArrayType::get(MI->getAllocatedType(),
783 NElements->getZExtValue());
785 new MallocInst(NewTy, Constant::getNullValue(Type::Int32Ty),
786 MI->getAlignment(), MI->getName(), MI);
788 Indices[0] = Indices[1] = Constant::getNullValue(Type::Int32Ty);
789 Value *NewGEP = new GetElementPtrInst(NewMI, Indices, Indices + 2,
790 NewMI->getName()+".el0", MI);
791 MI->replaceAllUsesWith(NewGEP);
792 MI->eraseFromParent();
796 // Create the new global variable. The contents of the malloc'd memory is
797 // undefined, so initialize with an undef value.
798 Constant *Init = UndefValue::get(MI->getAllocatedType());
799 GlobalVariable *NewGV = new GlobalVariable(MI->getAllocatedType(), false,
800 GlobalValue::InternalLinkage, Init,
801 GV->getName()+".body",
803 GV->isThreadLocal());
804 GV->getParent()->getGlobalList().insert(GV, NewGV);
806 // Anything that used the malloc now uses the global directly.
807 MI->replaceAllUsesWith(NewGV);
809 Constant *RepValue = NewGV;
810 if (NewGV->getType() != GV->getType()->getElementType())
811 RepValue = ConstantExpr::getBitCast(RepValue,
812 GV->getType()->getElementType());
814 // If there is a comparison against null, we will insert a global bool to
815 // keep track of whether the global was initialized yet or not.
816 GlobalVariable *InitBool =
817 new GlobalVariable(Type::Int1Ty, false, GlobalValue::InternalLinkage,
818 ConstantInt::getFalse(), GV->getName()+".init",
819 (Module *)NULL, GV->isThreadLocal());
820 bool InitBoolUsed = false;
822 // Loop over all uses of GV, processing them in turn.
823 std::vector<StoreInst*> Stores;
824 while (!GV->use_empty())
825 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
826 while (!LI->use_empty()) {
827 Use &LoadUse = LI->use_begin().getUse();
828 if (!isa<ICmpInst>(LoadUse.getUser()))
831 ICmpInst *CI = cast<ICmpInst>(LoadUse.getUser());
832 // Replace the cmp X, 0 with a use of the bool value.
833 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", CI);
835 switch (CI->getPredicate()) {
836 default: assert(0 && "Unknown ICmp Predicate!");
837 case ICmpInst::ICMP_ULT:
838 case ICmpInst::ICMP_SLT:
839 LV = ConstantInt::getFalse(); // X < null -> always false
841 case ICmpInst::ICMP_ULE:
842 case ICmpInst::ICMP_SLE:
843 case ICmpInst::ICMP_EQ:
844 LV = BinaryOperator::createNot(LV, "notinit", CI);
846 case ICmpInst::ICMP_NE:
847 case ICmpInst::ICMP_UGE:
848 case ICmpInst::ICMP_SGE:
849 case ICmpInst::ICMP_UGT:
850 case ICmpInst::ICMP_SGT:
853 CI->replaceAllUsesWith(LV);
854 CI->eraseFromParent();
857 LI->eraseFromParent();
859 StoreInst *SI = cast<StoreInst>(GV->use_back());
860 // The global is initialized when the store to it occurs.
861 new StoreInst(ConstantInt::getTrue(), InitBool, SI);
862 SI->eraseFromParent();
865 // If the initialization boolean was used, insert it, otherwise delete it.
867 while (!InitBool->use_empty()) // Delete initializations
868 cast<Instruction>(InitBool->use_back())->eraseFromParent();
871 GV->getParent()->getGlobalList().insert(GV, InitBool);
874 // Now the GV is dead, nuke it and the malloc.
875 GV->eraseFromParent();
876 MI->eraseFromParent();
878 // To further other optimizations, loop over all users of NewGV and try to
879 // constant prop them. This will promote GEP instructions with constant
880 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
881 ConstantPropUsersOf(NewGV);
882 if (RepValue != NewGV)
883 ConstantPropUsersOf(RepValue);
888 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
889 /// to make sure that there are no complex uses of V. We permit simple things
890 /// like dereferencing the pointer, but not storing through the address, unless
891 /// it is to the specified global.
892 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V,
894 SmallPtrSet<PHINode*, 8> &PHIs) {
895 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
896 if (isa<LoadInst>(*UI) || isa<CmpInst>(*UI)) {
898 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
899 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
900 return false; // Storing the pointer itself... bad.
901 // Otherwise, storing through it, or storing into GV... fine.
902 } else if (isa<GetElementPtrInst>(*UI)) {
903 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(cast<Instruction>(*UI),
906 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
907 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI
910 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
918 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
919 /// somewhere. Transform all uses of the allocation into loads from the
920 /// global and uses of the resultant pointer. Further, delete the store into
921 /// GV. This assumes that these value pass the
922 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
923 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
924 GlobalVariable *GV) {
925 while (!Alloc->use_empty()) {
926 Instruction *U = cast<Instruction>(*Alloc->use_begin());
927 Instruction *InsertPt = U;
928 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
929 // If this is the store of the allocation into the global, remove it.
930 if (SI->getOperand(1) == GV) {
931 SI->eraseFromParent();
934 } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
935 // Insert the load in the corresponding predecessor, not right before the
937 unsigned PredNo = Alloc->use_begin().getOperandNo()/2;
938 InsertPt = PN->getIncomingBlock(PredNo)->getTerminator();
941 // Insert a load from the global, and use it instead of the malloc.
942 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
943 U->replaceUsesOfWith(Alloc, NL);
947 /// GlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
948 /// GV are simple enough to perform HeapSRA, return true.
949 static bool GlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV,
951 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;
953 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
954 // We permit two users of the load: setcc comparing against the null
955 // pointer, and a getelementptr of a specific form.
956 for (Value::use_iterator UI = LI->use_begin(), E = LI->use_end(); UI != E;
958 // Comparison against null is ok.
959 if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) {
960 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
965 // getelementptr is also ok, but only a simple form.
966 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
967 // Must index into the array and into the struct.
968 if (GEPI->getNumOperands() < 3)
971 // Otherwise the GEP is ok.
975 if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
976 // We have a phi of a load from the global. We can only handle this
977 // if the other PHI'd values are actually the same. In this case,
978 // the rewriter will just drop the phi entirely.
979 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
980 Value *IV = PN->getIncomingValue(i);
981 if (IV == LI) continue; // Trivial the same.
983 // If the phi'd value is from the malloc that initializes the value,
985 if (IV == MI) continue;
987 // Otherwise, we don't know what it is.
993 // Otherwise we don't know what this is, not ok.
1000 /// GetHeapSROALoad - Return the load for the specified field of the HeapSROA'd
1001 /// value, lazily creating it on demand.
1002 static Value *GetHeapSROALoad(Instruction *Load, unsigned FieldNo,
1003 const std::vector<GlobalVariable*> &FieldGlobals,
1004 std::vector<Value *> &InsertedLoadsForPtr) {
1005 if (InsertedLoadsForPtr.size() <= FieldNo)
1006 InsertedLoadsForPtr.resize(FieldNo+1);
1007 if (InsertedLoadsForPtr[FieldNo] == 0)
1008 InsertedLoadsForPtr[FieldNo] = new LoadInst(FieldGlobals[FieldNo],
1009 Load->getName()+".f" +
1010 utostr(FieldNo), Load);
1011 return InsertedLoadsForPtr[FieldNo];
1014 /// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1015 /// the load, rewrite the derived value to use the HeapSRoA'd load.
1016 static void RewriteHeapSROALoadUser(LoadInst *Load, Instruction *LoadUser,
1017 const std::vector<GlobalVariable*> &FieldGlobals,
1018 std::vector<Value *> &InsertedLoadsForPtr) {
1019 // If this is a comparison against null, handle it.
1020 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1021 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1022 // If we have a setcc of the loaded pointer, we can use a setcc of any
1025 if (InsertedLoadsForPtr.empty()) {
1026 NPtr = GetHeapSROALoad(Load, 0, FieldGlobals, InsertedLoadsForPtr);
1028 NPtr = InsertedLoadsForPtr.back();
1031 Value *New = new ICmpInst(SCI->getPredicate(), NPtr,
1032 Constant::getNullValue(NPtr->getType()),
1033 SCI->getName(), SCI);
1034 SCI->replaceAllUsesWith(New);
1035 SCI->eraseFromParent();
1039 // Handle 'getelementptr Ptr, Idx, uint FieldNo ...'
1040 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1041 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1042 && "Unexpected GEPI!");
1044 // Load the pointer for this field.
1045 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1046 Value *NewPtr = GetHeapSROALoad(Load, FieldNo,
1047 FieldGlobals, InsertedLoadsForPtr);
1049 // Create the new GEP idx vector.
1050 SmallVector<Value*, 8> GEPIdx;
1051 GEPIdx.push_back(GEPI->getOperand(1));
1052 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1054 Value *NGEPI = new GetElementPtrInst(NewPtr, GEPIdx.begin(), GEPIdx.end(),
1055 GEPI->getName(), GEPI);
1056 GEPI->replaceAllUsesWith(NGEPI);
1057 GEPI->eraseFromParent();
1061 // Handle PHI nodes. PHI nodes must be merging in the same values, plus
1062 // potentially the original malloc. Insert phi nodes for each field, then
1063 // process uses of the PHI.
1064 PHINode *PN = cast<PHINode>(LoadUser);
1065 std::vector<Value *> PHIsForField;
1066 PHIsForField.resize(FieldGlobals.size());
1067 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1068 Value *LoadV = GetHeapSROALoad(Load, i, FieldGlobals, InsertedLoadsForPtr);
1070 PHINode *FieldPN = new PHINode(LoadV->getType(),
1071 PN->getName()+"."+utostr(i), PN);
1072 // Fill in the predecessor values.
1073 for (unsigned pred = 0, e = PN->getNumIncomingValues(); pred != e; ++pred) {
1074 // Each predecessor either uses the load or the original malloc.
1075 Value *InVal = PN->getIncomingValue(pred);
1076 BasicBlock *BB = PN->getIncomingBlock(pred);
1078 if (isa<MallocInst>(InVal)) {
1079 // Insert a reload from the global in the predecessor.
1080 NewVal = GetHeapSROALoad(BB->getTerminator(), i, FieldGlobals,
1083 NewVal = InsertedLoadsForPtr[i];
1085 FieldPN->addIncoming(NewVal, BB);
1087 PHIsForField[i] = FieldPN;
1090 // Since PHIsForField specifies a phi for every input value, the lazy inserter
1091 // will never insert a load.
1092 while (!PN->use_empty())
1093 RewriteHeapSROALoadUser(Load, PN->use_back(), FieldGlobals, PHIsForField);
1094 PN->eraseFromParent();
1097 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1098 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1099 /// use FieldGlobals instead. All uses of loaded values satisfy
1100 /// GlobalLoadUsesSimpleEnoughForHeapSRA.
1101 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1102 const std::vector<GlobalVariable*> &FieldGlobals) {
1103 std::vector<Value *> InsertedLoadsForPtr;
1104 //InsertedLoadsForPtr.resize(FieldGlobals.size());
1105 while (!Load->use_empty())
1106 RewriteHeapSROALoadUser(Load, Load->use_back(),
1107 FieldGlobals, InsertedLoadsForPtr);
1110 /// PerformHeapAllocSRoA - MI is an allocation of an array of structures. Break
1111 /// it up into multiple allocations of arrays of the fields.
1112 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){
1113 DOUT << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *MI;
1114 const StructType *STy = cast<StructType>(MI->getAllocatedType());
1116 // There is guaranteed to be at least one use of the malloc (storing
1117 // it into GV). If there are other uses, change them to be uses of
1118 // the global to simplify later code. This also deletes the store
1120 ReplaceUsesOfMallocWithGlobal(MI, GV);
1122 // Okay, at this point, there are no users of the malloc. Insert N
1123 // new mallocs at the same place as MI, and N globals.
1124 std::vector<GlobalVariable*> FieldGlobals;
1125 std::vector<MallocInst*> FieldMallocs;
1127 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1128 const Type *FieldTy = STy->getElementType(FieldNo);
1129 const Type *PFieldTy = PointerType::getUnqual(FieldTy);
1131 GlobalVariable *NGV =
1132 new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage,
1133 Constant::getNullValue(PFieldTy),
1134 GV->getName() + ".f" + utostr(FieldNo), GV,
1135 GV->isThreadLocal());
1136 FieldGlobals.push_back(NGV);
1138 MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(),
1139 MI->getName() + ".f" + utostr(FieldNo),MI);
1140 FieldMallocs.push_back(NMI);
1141 new StoreInst(NMI, NGV, MI);
1144 // The tricky aspect of this transformation is handling the case when malloc
1145 // fails. In the original code, malloc failing would set the result pointer
1146 // of malloc to null. In this case, some mallocs could succeed and others
1147 // could fail. As such, we emit code that looks like this:
1148 // F0 = malloc(field0)
1149 // F1 = malloc(field1)
1150 // F2 = malloc(field2)
1151 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1152 // if (F0) { free(F0); F0 = 0; }
1153 // if (F1) { free(F1); F1 = 0; }
1154 // if (F2) { free(F2); F2 = 0; }
1156 Value *RunningOr = 0;
1157 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1158 Value *Cond = new ICmpInst(ICmpInst::ICMP_EQ, FieldMallocs[i],
1159 Constant::getNullValue(FieldMallocs[i]->getType()),
1162 RunningOr = Cond; // First seteq
1164 RunningOr = BinaryOperator::createOr(RunningOr, Cond, "tmp", MI);
1167 // Split the basic block at the old malloc.
1168 BasicBlock *OrigBB = MI->getParent();
1169 BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont");
1171 // Create the block to check the first condition. Put all these blocks at the
1172 // end of the function as they are unlikely to be executed.
1173 BasicBlock *NullPtrBlock = new BasicBlock("malloc_ret_null",
1174 OrigBB->getParent());
1176 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1177 // branch on RunningOr.
1178 OrigBB->getTerminator()->eraseFromParent();
1179 new BranchInst(NullPtrBlock, ContBB, RunningOr, OrigBB);
1181 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1182 // pointer, because some may be null while others are not.
1183 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1184 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1185 Value *Cmp = new ICmpInst(ICmpInst::ICMP_NE, GVVal,
1186 Constant::getNullValue(GVVal->getType()),
1187 "tmp", NullPtrBlock);
1188 BasicBlock *FreeBlock = new BasicBlock("free_it", OrigBB->getParent());
1189 BasicBlock *NextBlock = new BasicBlock("next", OrigBB->getParent());
1190 new BranchInst(FreeBlock, NextBlock, Cmp, NullPtrBlock);
1192 // Fill in FreeBlock.
1193 new FreeInst(GVVal, FreeBlock);
1194 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1196 new BranchInst(NextBlock, FreeBlock);
1198 NullPtrBlock = NextBlock;
1201 new BranchInst(ContBB, NullPtrBlock);
1204 // MI is no longer needed, remove it.
1205 MI->eraseFromParent();
1208 // Okay, the malloc site is completely handled. All of the uses of GV are now
1209 // loads, and all uses of those loads are simple. Rewrite them to use loads
1210 // of the per-field globals instead.
1211 while (!GV->use_empty()) {
1212 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
1213 RewriteUsesOfLoadForHeapSRoA(LI, FieldGlobals);
1214 LI->eraseFromParent();
1216 // Must be a store of null.
1217 StoreInst *SI = cast<StoreInst>(GV->use_back());
1218 assert(isa<Constant>(SI->getOperand(0)) &&
1219 cast<Constant>(SI->getOperand(0))->isNullValue() &&
1220 "Unexpected heap-sra user!");
1222 // Insert a store of null into each global.
1223 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1225 Constant::getNullValue(FieldGlobals[i]->getType()->getElementType());
1226 new StoreInst(Null, FieldGlobals[i], SI);
1228 // Erase the original store.
1229 SI->eraseFromParent();
1233 // The old global is now dead, remove it.
1234 GV->eraseFromParent();
1237 return FieldGlobals[0];
1241 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1242 // that only one value (besides its initializer) is ever stored to the global.
1243 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1244 Module::global_iterator &GVI,
1246 if (CastInst *CI = dyn_cast<CastInst>(StoredOnceVal))
1247 StoredOnceVal = CI->getOperand(0);
1248 else if (GetElementPtrInst *GEPI =dyn_cast<GetElementPtrInst>(StoredOnceVal)){
1249 // "getelementptr Ptr, 0, 0, 0" is really just a cast.
1250 bool IsJustACast = true;
1251 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
1252 if (!isa<Constant>(GEPI->getOperand(i)) ||
1253 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
1254 IsJustACast = false;
1258 StoredOnceVal = GEPI->getOperand(0);
1261 // If we are dealing with a pointer global that is initialized to null and
1262 // only has one (non-null) value stored into it, then we can optimize any
1263 // users of the loaded value (often calls and loads) that would trap if the
1265 if (isa<PointerType>(GV->getInitializer()->getType()) &&
1266 GV->getInitializer()->isNullValue()) {
1267 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1268 if (GV->getInitializer()->getType() != SOVC->getType())
1269 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1271 // Optimize away any trapping uses of the loaded value.
1272 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1274 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) {
1275 // If this is a malloc of an abstract type, don't touch it.
1276 if (!MI->getAllocatedType()->isSized())
1279 // We can't optimize this global unless all uses of it are *known* to be
1280 // of the malloc value, not of the null initializer value (consider a use
1281 // that compares the global's value against zero to see if the malloc has
1282 // been reached). To do this, we check to see if all uses of the global
1283 // would trap if the global were null: this proves that they must all
1284 // happen after the malloc.
1285 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1288 // We can't optimize this if the malloc itself is used in a complex way,
1289 // for example, being stored into multiple globals. This allows the
1290 // malloc to be stored into the specified global, loaded setcc'd, and
1291 // GEP'd. These are all things we could transform to using the global
1294 SmallPtrSet<PHINode*, 8> PHIs;
1295 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV, PHIs))
1300 // If we have a global that is only initialized with a fixed size malloc,
1301 // transform the program to use global memory instead of malloc'd memory.
1302 // This eliminates dynamic allocation, avoids an indirection accessing the
1303 // data, and exposes the resultant global to further GlobalOpt.
1304 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) {
1305 // Restrict this transformation to only working on small allocations
1306 // (2048 bytes currently), as we don't want to introduce a 16M global or
1308 if (NElements->getZExtValue()*
1309 TD.getABITypeSize(MI->getAllocatedType()) < 2048) {
1310 GVI = OptimizeGlobalAddressOfMalloc(GV, MI);
1315 // If the allocation is an array of structures, consider transforming this
1316 // into multiple malloc'd arrays, one for each field. This is basically
1317 // SRoA for malloc'd memory.
1318 if (const StructType *AllocTy =
1319 dyn_cast<StructType>(MI->getAllocatedType())) {
1320 // This the structure has an unreasonable number of fields, leave it
1322 if (AllocTy->getNumElements() <= 16 && AllocTy->getNumElements() > 0 &&
1323 GlobalLoadUsesSimpleEnoughForHeapSRA(GV, MI)) {
1324 GVI = PerformHeapAllocSRoA(GV, MI);
1334 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1335 /// two values ever stored into GV are its initializer and OtherVal. See if we
1336 /// can shrink the global into a boolean and select between the two values
1337 /// whenever it is used. This exposes the values to other scalar optimizations.
1338 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1339 const Type *GVElType = GV->getType()->getElementType();
1341 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1342 // an FP value or vector, don't do this optimization because a select between
1343 // them is very expensive and unlikely to lead to later simplification.
1344 if (GVElType == Type::Int1Ty || GVElType->isFloatingPoint() ||
1345 isa<VectorType>(GVElType))
1348 // Walk the use list of the global seeing if all the uses are load or store.
1349 // If there is anything else, bail out.
1350 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I)
1351 if (!isa<LoadInst>(I) && !isa<StoreInst>(I))
1354 DOUT << " *** SHRINKING TO BOOL: " << *GV;
1356 // Create the new global, initializing it to false.
1357 GlobalVariable *NewGV = new GlobalVariable(Type::Int1Ty, false,
1358 GlobalValue::InternalLinkage, ConstantInt::getFalse(),
1361 GV->isThreadLocal());
1362 GV->getParent()->getGlobalList().insert(GV, NewGV);
1364 Constant *InitVal = GV->getInitializer();
1365 assert(InitVal->getType() != Type::Int1Ty && "No reason to shrink to bool!");
1367 // If initialized to zero and storing one into the global, we can use a cast
1368 // instead of a select to synthesize the desired value.
1369 bool IsOneZero = false;
1370 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1371 IsOneZero = InitVal->isNullValue() && CI->isOne();
1373 while (!GV->use_empty()) {
1374 Instruction *UI = cast<Instruction>(GV->use_back());
1375 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1376 // Change the store into a boolean store.
1377 bool StoringOther = SI->getOperand(0) == OtherVal;
1378 // Only do this if we weren't storing a loaded value.
1380 if (StoringOther || SI->getOperand(0) == InitVal)
1381 StoreVal = ConstantInt::get(Type::Int1Ty, StoringOther);
1383 // Otherwise, we are storing a previously loaded copy. To do this,
1384 // change the copy from copying the original value to just copying the
1386 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1388 // If we're already replaced the input, StoredVal will be a cast or
1389 // select instruction. If not, it will be a load of the original
1391 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1392 assert(LI->getOperand(0) == GV && "Not a copy!");
1393 // Insert a new load, to preserve the saved value.
1394 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1396 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1397 "This is not a form that we understand!");
1398 StoreVal = StoredVal->getOperand(0);
1399 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1402 new StoreInst(StoreVal, NewGV, SI);
1404 // Change the load into a load of bool then a select.
1405 LoadInst *LI = cast<LoadInst>(UI);
1406 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1409 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1411 NSI = new SelectInst(NLI, OtherVal, InitVal, "", LI);
1413 LI->replaceAllUsesWith(NSI);
1415 UI->eraseFromParent();
1418 GV->eraseFromParent();
1423 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1424 /// it if possible. If we make a change, return true.
1425 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1426 Module::global_iterator &GVI) {
1427 std::set<PHINode*> PHIUsers;
1429 GV->removeDeadConstantUsers();
1431 if (GV->use_empty()) {
1432 DOUT << "GLOBAL DEAD: " << *GV;
1433 GV->eraseFromParent();
1438 if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
1440 cerr << "Global: " << *GV;
1441 cerr << " isLoaded = " << GS.isLoaded << "\n";
1442 cerr << " StoredType = ";
1443 switch (GS.StoredType) {
1444 case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break;
1445 case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break;
1446 case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break;
1447 case GlobalStatus::isStored: cerr << "stored\n"; break;
1449 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
1450 cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n";
1451 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
1452 cerr << " AccessingFunction = " << GS.AccessingFunction->getName()
1454 cerr << " HasMultipleAccessingFunctions = "
1455 << GS.HasMultipleAccessingFunctions << "\n";
1456 cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n";
1457 cerr << " isNotSuitableForSRA = " << GS.isNotSuitableForSRA << "\n";
1461 // If this is a first class global and has only one accessing function
1462 // and this function is main (which we know is not recursive we can make
1463 // this global a local variable) we replace the global with a local alloca
1464 // in this function.
1466 // NOTE: It doesn't make sense to promote non first class types since we
1467 // are just replacing static memory to stack memory.
1468 if (!GS.HasMultipleAccessingFunctions &&
1469 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1470 GV->getType()->getElementType()->isFirstClassType() &&
1471 GS.AccessingFunction->getName() == "main" &&
1472 GS.AccessingFunction->hasExternalLinkage()) {
1473 DOUT << "LOCALIZING GLOBAL: " << *GV;
1474 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin();
1475 const Type* ElemTy = GV->getType()->getElementType();
1476 // FIXME: Pass Global's alignment when globals have alignment
1477 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI);
1478 if (!isa<UndefValue>(GV->getInitializer()))
1479 new StoreInst(GV->getInitializer(), Alloca, FirstI);
1481 GV->replaceAllUsesWith(Alloca);
1482 GV->eraseFromParent();
1487 // If the global is never loaded (but may be stored to), it is dead.
1490 DOUT << "GLOBAL NEVER LOADED: " << *GV;
1492 // Delete any stores we can find to the global. We may not be able to
1493 // make it completely dead though.
1494 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1496 // If the global is dead now, delete it.
1497 if (GV->use_empty()) {
1498 GV->eraseFromParent();
1504 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1505 DOUT << "MARKING CONSTANT: " << *GV;
1506 GV->setConstant(true);
1508 // Clean up any obviously simplifiable users now.
1509 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1511 // If the global is dead now, just nuke it.
1512 if (GV->use_empty()) {
1513 DOUT << " *** Marking constant allowed us to simplify "
1514 << "all users and delete global!\n";
1515 GV->eraseFromParent();
1521 } else if (!GV->getInitializer()->getType()->isFirstClassType()) {
1522 if (GlobalVariable *FirstNewGV = SRAGlobal(GV)) {
1523 GVI = FirstNewGV; // Don't skip the newly produced globals!
1526 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1527 // If the initial value for the global was an undef value, and if only
1528 // one other value was stored into it, we can just change the
1529 // initializer to be an undef value, then delete all stores to the
1530 // global. This allows us to mark it constant.
1531 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1532 if (isa<UndefValue>(GV->getInitializer())) {
1533 // Change the initial value here.
1534 GV->setInitializer(SOVConstant);
1536 // Clean up any obviously simplifiable users now.
1537 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1539 if (GV->use_empty()) {
1540 DOUT << " *** Substituting initializer allowed us to "
1541 << "simplify all users and delete global!\n";
1542 GV->eraseFromParent();
1551 // Try to optimize globals based on the knowledge that only one value
1552 // (besides its initializer) is ever stored to the global.
1553 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1554 getAnalysis<TargetData>()))
1557 // Otherwise, if the global was not a boolean, we can shrink it to be a
1559 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1560 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1569 /// OnlyCalledDirectly - Return true if the specified function is only called
1570 /// directly. In other words, its address is never taken.
1571 static bool OnlyCalledDirectly(Function *F) {
1572 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1573 Instruction *User = dyn_cast<Instruction>(*UI);
1574 if (!User) return false;
1575 if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false;
1577 // See if the function address is passed as an argument.
1578 for (unsigned i = 1, e = User->getNumOperands(); i != e; ++i)
1579 if (User->getOperand(i) == F) return false;
1584 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1585 /// function, changing them to FastCC.
1586 static void ChangeCalleesToFastCall(Function *F) {
1587 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1588 Instruction *User = cast<Instruction>(*UI);
1589 if (CallInst *CI = dyn_cast<CallInst>(User))
1590 CI->setCallingConv(CallingConv::Fast);
1592 cast<InvokeInst>(User)->setCallingConv(CallingConv::Fast);
1596 bool GlobalOpt::OptimizeFunctions(Module &M) {
1597 bool Changed = false;
1598 // Optimize functions.
1599 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1601 F->removeDeadConstantUsers();
1602 if (F->use_empty() && (F->hasInternalLinkage() ||
1603 F->hasLinkOnceLinkage())) {
1604 M.getFunctionList().erase(F);
1607 } else if (F->hasInternalLinkage() &&
1608 F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1609 OnlyCalledDirectly(F)) {
1610 // If this function has C calling conventions, is not a varargs
1611 // function, and is only called directly, promote it to use the Fast
1612 // calling convention.
1613 F->setCallingConv(CallingConv::Fast);
1614 ChangeCalleesToFastCall(F);
1622 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1623 bool Changed = false;
1624 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1626 GlobalVariable *GV = GVI++;
1627 if (!GV->isConstant() && GV->hasInternalLinkage() &&
1628 GV->hasInitializer())
1629 Changed |= ProcessInternalGlobal(GV, GVI);
1634 /// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1635 /// initializers have an init priority of 65535.
1636 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1637 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1639 if (I->getName() == "llvm.global_ctors") {
1640 // Found it, verify it's an array of { int, void()* }.
1641 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1643 const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1644 if (!STy || STy->getNumElements() != 2 ||
1645 STy->getElementType(0) != Type::Int32Ty) return 0;
1646 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1647 if (!PFTy) return 0;
1648 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1649 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() ||
1650 FTy->getNumParams() != 0)
1653 // Verify that the initializer is simple enough for us to handle.
1654 if (!I->hasInitializer()) return 0;
1655 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1657 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1658 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(CA->getOperand(i))) {
1659 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1662 // Must have a function or null ptr.
1663 if (!isa<Function>(CS->getOperand(1)))
1666 // Init priority must be standard.
1667 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1668 if (!CI || CI->getZExtValue() != 65535)
1679 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1680 /// return a list of the functions and null terminator as a vector.
1681 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1682 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1683 std::vector<Function*> Result;
1684 Result.reserve(CA->getNumOperands());
1685 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) {
1686 ConstantStruct *CS = cast<ConstantStruct>(CA->getOperand(i));
1687 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1692 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1693 /// specified array, returning the new global to use.
1694 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1695 const std::vector<Function*> &Ctors) {
1696 // If we made a change, reassemble the initializer list.
1697 std::vector<Constant*> CSVals;
1698 CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535));
1699 CSVals.push_back(0);
1701 // Create the new init list.
1702 std::vector<Constant*> CAList;
1703 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1705 CSVals[1] = Ctors[i];
1707 const Type *FTy = FunctionType::get(Type::VoidTy,
1708 std::vector<const Type*>(), false);
1709 const PointerType *PFTy = PointerType::getUnqual(FTy);
1710 CSVals[1] = Constant::getNullValue(PFTy);
1711 CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647);
1713 CAList.push_back(ConstantStruct::get(CSVals));
1716 // Create the array initializer.
1717 const Type *StructTy =
1718 cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
1719 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()),
1722 // If we didn't change the number of elements, don't create a new GV.
1723 if (CA->getType() == GCL->getInitializer()->getType()) {
1724 GCL->setInitializer(CA);
1728 // Create the new global and insert it next to the existing list.
1729 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
1730 GCL->getLinkage(), CA, "",
1732 GCL->isThreadLocal());
1733 GCL->getParent()->getGlobalList().insert(GCL, NGV);
1736 // Nuke the old list, replacing any uses with the new one.
1737 if (!GCL->use_empty()) {
1739 if (V->getType() != GCL->getType())
1740 V = ConstantExpr::getBitCast(V, GCL->getType());
1741 GCL->replaceAllUsesWith(V);
1743 GCL->eraseFromParent();
1752 static Constant *getVal(std::map<Value*, Constant*> &ComputedValues,
1754 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
1755 Constant *R = ComputedValues[V];
1756 assert(R && "Reference to an uncomputed value!");
1760 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
1761 /// enough for us to understand. In particular, if it is a cast of something,
1762 /// we punt. We basically just support direct accesses to globals and GEP's of
1763 /// globals. This should be kept up to date with CommitValueTo.
1764 static bool isSimpleEnoughPointerToCommit(Constant *C) {
1765 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
1766 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1767 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1768 return !GV->isDeclaration(); // reject external globals.
1770 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
1771 // Handle a constantexpr gep.
1772 if (CE->getOpcode() == Instruction::GetElementPtr &&
1773 isa<GlobalVariable>(CE->getOperand(0))) {
1774 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1775 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1776 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1777 return GV->hasInitializer() &&
1778 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1783 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
1784 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
1785 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
1786 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
1787 ConstantExpr *Addr, unsigned OpNo) {
1788 // Base case of the recursion.
1789 if (OpNo == Addr->getNumOperands()) {
1790 assert(Val->getType() == Init->getType() && "Type mismatch!");
1794 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
1795 std::vector<Constant*> Elts;
1797 // Break up the constant into its elements.
1798 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
1799 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i)
1800 Elts.push_back(CS->getOperand(i));
1801 } else if (isa<ConstantAggregateZero>(Init)) {
1802 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1803 Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
1804 } else if (isa<UndefValue>(Init)) {
1805 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1806 Elts.push_back(UndefValue::get(STy->getElementType(i)));
1808 assert(0 && "This code is out of sync with "
1809 " ConstantFoldLoadThroughGEPConstantExpr");
1812 // Replace the element that we are supposed to.
1813 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
1814 unsigned Idx = CU->getZExtValue();
1815 assert(Idx < STy->getNumElements() && "Struct index out of range!");
1816 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
1818 // Return the modified struct.
1819 return ConstantStruct::get(&Elts[0], Elts.size(), STy->isPacked());
1821 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
1822 const ArrayType *ATy = cast<ArrayType>(Init->getType());
1824 // Break up the array into elements.
1825 std::vector<Constant*> Elts;
1826 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
1827 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1828 Elts.push_back(CA->getOperand(i));
1829 } else if (isa<ConstantAggregateZero>(Init)) {
1830 Constant *Elt = Constant::getNullValue(ATy->getElementType());
1831 Elts.assign(ATy->getNumElements(), Elt);
1832 } else if (isa<UndefValue>(Init)) {
1833 Constant *Elt = UndefValue::get(ATy->getElementType());
1834 Elts.assign(ATy->getNumElements(), Elt);
1836 assert(0 && "This code is out of sync with "
1837 " ConstantFoldLoadThroughGEPConstantExpr");
1840 assert(CI->getZExtValue() < ATy->getNumElements());
1841 Elts[CI->getZExtValue()] =
1842 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
1843 return ConstantArray::get(ATy, Elts);
1847 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
1848 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
1849 static void CommitValueTo(Constant *Val, Constant *Addr) {
1850 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
1851 assert(GV->hasInitializer());
1852 GV->setInitializer(Val);
1856 ConstantExpr *CE = cast<ConstantExpr>(Addr);
1857 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1859 Constant *Init = GV->getInitializer();
1860 Init = EvaluateStoreInto(Init, Val, CE, 2);
1861 GV->setInitializer(Init);
1864 /// ComputeLoadResult - Return the value that would be computed by a load from
1865 /// P after the stores reflected by 'memory' have been performed. If we can't
1866 /// decide, return null.
1867 static Constant *ComputeLoadResult(Constant *P,
1868 const std::map<Constant*, Constant*> &Memory) {
1869 // If this memory location has been recently stored, use the stored value: it
1870 // is the most up-to-date.
1871 std::map<Constant*, Constant*>::const_iterator I = Memory.find(P);
1872 if (I != Memory.end()) return I->second;
1875 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
1876 if (GV->hasInitializer())
1877 return GV->getInitializer();
1881 // Handle a constantexpr getelementptr.
1882 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
1883 if (CE->getOpcode() == Instruction::GetElementPtr &&
1884 isa<GlobalVariable>(CE->getOperand(0))) {
1885 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1886 if (GV->hasInitializer())
1887 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1890 return 0; // don't know how to evaluate.
1893 /// EvaluateFunction - Evaluate a call to function F, returning true if
1894 /// successful, false if we can't evaluate it. ActualArgs contains the formal
1895 /// arguments for the function.
1896 static bool EvaluateFunction(Function *F, Constant *&RetVal,
1897 const std::vector<Constant*> &ActualArgs,
1898 std::vector<Function*> &CallStack,
1899 std::map<Constant*, Constant*> &MutatedMemory,
1900 std::vector<GlobalVariable*> &AllocaTmps) {
1901 // Check to see if this function is already executing (recursion). If so,
1902 // bail out. TODO: we might want to accept limited recursion.
1903 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
1906 CallStack.push_back(F);
1908 /// Values - As we compute SSA register values, we store their contents here.
1909 std::map<Value*, Constant*> Values;
1911 // Initialize arguments to the incoming values specified.
1913 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
1915 Values[AI] = ActualArgs[ArgNo];
1917 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
1918 /// we can only evaluate any one basic block at most once. This set keeps
1919 /// track of what we have executed so we can detect recursive cases etc.
1920 std::set<BasicBlock*> ExecutedBlocks;
1922 // CurInst - The current instruction we're evaluating.
1923 BasicBlock::iterator CurInst = F->begin()->begin();
1925 // This is the main evaluation loop.
1927 Constant *InstResult = 0;
1929 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
1930 if (SI->isVolatile()) return false; // no volatile accesses.
1931 Constant *Ptr = getVal(Values, SI->getOperand(1));
1932 if (!isSimpleEnoughPointerToCommit(Ptr))
1933 // If this is too complex for us to commit, reject it.
1935 Constant *Val = getVal(Values, SI->getOperand(0));
1936 MutatedMemory[Ptr] = Val;
1937 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
1938 InstResult = ConstantExpr::get(BO->getOpcode(),
1939 getVal(Values, BO->getOperand(0)),
1940 getVal(Values, BO->getOperand(1)));
1941 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
1942 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
1943 getVal(Values, CI->getOperand(0)),
1944 getVal(Values, CI->getOperand(1)));
1945 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
1946 InstResult = ConstantExpr::getCast(CI->getOpcode(),
1947 getVal(Values, CI->getOperand(0)),
1949 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
1950 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
1951 getVal(Values, SI->getOperand(1)),
1952 getVal(Values, SI->getOperand(2)));
1953 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
1954 Constant *P = getVal(Values, GEP->getOperand(0));
1955 SmallVector<Constant*, 8> GEPOps;
1956 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
1957 GEPOps.push_back(getVal(Values, GEP->getOperand(i)));
1958 InstResult = ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size());
1959 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
1960 if (LI->isVolatile()) return false; // no volatile accesses.
1961 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
1963 if (InstResult == 0) return false; // Could not evaluate load.
1964 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
1965 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs.
1966 const Type *Ty = AI->getType()->getElementType();
1967 AllocaTmps.push_back(new GlobalVariable(Ty, false,
1968 GlobalValue::InternalLinkage,
1969 UndefValue::get(Ty),
1971 InstResult = AllocaTmps.back();
1972 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
1973 // Cannot handle inline asm.
1974 if (isa<InlineAsm>(CI->getOperand(0))) return false;
1976 // Resolve function pointers.
1977 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0)));
1978 if (!Callee) return false; // Cannot resolve.
1980 std::vector<Constant*> Formals;
1981 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
1982 Formals.push_back(getVal(Values, CI->getOperand(i)));
1984 if (Callee->isDeclaration()) {
1985 // If this is a function we can constant fold, do it.
1986 if (Constant *C = ConstantFoldCall(Callee, &Formals[0],
1993 if (Callee->getFunctionType()->isVarArg())
1998 // Execute the call, if successful, use the return value.
1999 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
2000 MutatedMemory, AllocaTmps))
2002 InstResult = RetVal;
2004 } else if (isa<TerminatorInst>(CurInst)) {
2005 BasicBlock *NewBB = 0;
2006 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2007 if (BI->isUnconditional()) {
2008 NewBB = BI->getSuccessor(0);
2011 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
2012 if (!Cond) return false; // Cannot determine.
2014 NewBB = BI->getSuccessor(!Cond->getZExtValue());
2016 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2018 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
2019 if (!Val) return false; // Cannot determine.
2020 NewBB = SI->getSuccessor(SI->findCaseValue(Val));
2021 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
2022 if (RI->getNumOperands())
2023 RetVal = getVal(Values, RI->getOperand(0));
2025 CallStack.pop_back(); // return from fn.
2026 return true; // We succeeded at evaluating this ctor!
2028 // invoke, unwind, unreachable.
2029 return false; // Cannot handle this terminator.
2032 // Okay, we succeeded in evaluating this control flow. See if we have
2033 // executed the new block before. If so, we have a looping function,
2034 // which we cannot evaluate in reasonable time.
2035 if (!ExecutedBlocks.insert(NewBB).second)
2036 return false; // looped!
2038 // Okay, we have never been in this block before. Check to see if there
2039 // are any PHI nodes. If so, evaluate them with information about where
2041 BasicBlock *OldBB = CurInst->getParent();
2042 CurInst = NewBB->begin();
2044 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2045 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
2047 // Do NOT increment CurInst. We know that the terminator had no value.
2050 // Did not know how to evaluate this!
2054 if (!CurInst->use_empty())
2055 Values[CurInst] = InstResult;
2057 // Advance program counter.
2062 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2063 /// we can. Return true if we can, false otherwise.
2064 static bool EvaluateStaticConstructor(Function *F) {
2065 /// MutatedMemory - For each store we execute, we update this map. Loads
2066 /// check this to get the most up-to-date value. If evaluation is successful,
2067 /// this state is committed to the process.
2068 std::map<Constant*, Constant*> MutatedMemory;
2070 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2071 /// to represent its body. This vector is needed so we can delete the
2072 /// temporary globals when we are done.
2073 std::vector<GlobalVariable*> AllocaTmps;
2075 /// CallStack - This is used to detect recursion. In pathological situations
2076 /// we could hit exponential behavior, but at least there is nothing
2078 std::vector<Function*> CallStack;
2080 // Call the function.
2081 Constant *RetValDummy;
2082 bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(),
2083 CallStack, MutatedMemory, AllocaTmps);
2085 // We succeeded at evaluation: commit the result.
2086 DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2087 << F->getName() << "' to " << MutatedMemory.size()
2089 for (std::map<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2090 E = MutatedMemory.end(); I != E; ++I)
2091 CommitValueTo(I->second, I->first);
2094 // At this point, we are done interpreting. If we created any 'alloca'
2095 // temporaries, release them now.
2096 while (!AllocaTmps.empty()) {
2097 GlobalVariable *Tmp = AllocaTmps.back();
2098 AllocaTmps.pop_back();
2100 // If there are still users of the alloca, the program is doing something
2101 // silly, e.g. storing the address of the alloca somewhere and using it
2102 // later. Since this is undefined, we'll just make it be null.
2103 if (!Tmp->use_empty())
2104 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2113 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2114 /// Return true if anything changed.
2115 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2116 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2117 bool MadeChange = false;
2118 if (Ctors.empty()) return false;
2120 // Loop over global ctors, optimizing them when we can.
2121 for (unsigned i = 0; i != Ctors.size(); ++i) {
2122 Function *F = Ctors[i];
2123 // Found a null terminator in the middle of the list, prune off the rest of
2126 if (i != Ctors.size()-1) {
2133 // We cannot simplify external ctor functions.
2134 if (F->empty()) continue;
2136 // If we can evaluate the ctor at compile time, do.
2137 if (EvaluateStaticConstructor(F)) {
2138 Ctors.erase(Ctors.begin()+i);
2141 ++NumCtorsEvaluated;
2146 if (!MadeChange) return false;
2148 GCL = InstallGlobalCtors(GCL, Ctors);
2153 bool GlobalOpt::runOnModule(Module &M) {
2154 bool Changed = false;
2156 // Try to find the llvm.globalctors list.
2157 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2159 bool LocalChange = true;
2160 while (LocalChange) {
2161 LocalChange = false;
2163 // Delete functions that are trivially dead, ccc -> fastcc
2164 LocalChange |= OptimizeFunctions(M);
2166 // Optimize global_ctors list.
2168 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2170 // Optimize non-address-taken globals.
2171 LocalChange |= OptimizeGlobalVars(M);
2172 Changed |= LocalChange;
2175 // TODO: Move all global ctors functions to the end of the module for code