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/CallSite.h"
28 #include "llvm/Support/Compiler.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/GetElementPtrTypeIterator.h"
31 #include "llvm/Support/MathExtras.h"
32 #include "llvm/ADT/SmallPtrSet.h"
33 #include "llvm/ADT/SmallVector.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/ADT/StringExtras.h"
41 STATISTIC(NumMarked , "Number of globals marked constant");
42 STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
43 STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
44 STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
45 STATISTIC(NumDeleted , "Number of globals deleted");
46 STATISTIC(NumFnDeleted , "Number of functions deleted");
47 STATISTIC(NumGlobUses , "Number of global uses devirtualized");
48 STATISTIC(NumLocalized , "Number of globals localized");
49 STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans");
50 STATISTIC(NumFastCallFns , "Number of functions converted to fastcc");
51 STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
52 STATISTIC(NumNestRemoved , "Number of nest attributes removed");
55 struct VISIBILITY_HIDDEN GlobalOpt : public ModulePass {
56 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
57 AU.addRequired<TargetData>();
59 static char ID; // Pass identification, replacement for typeid
60 GlobalOpt() : ModulePass(&ID) {}
62 bool runOnModule(Module &M);
65 GlobalVariable *FindGlobalCtors(Module &M);
66 bool OptimizeFunctions(Module &M);
67 bool OptimizeGlobalVars(Module &M);
68 bool ResolveAliases(Module &M);
69 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
70 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
74 char GlobalOpt::ID = 0;
75 static RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer");
77 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
81 /// GlobalStatus - As we analyze each global, keep track of some information
82 /// about it. If we find out that the address of the global is taken, none of
83 /// this info will be accurate.
84 struct VISIBILITY_HIDDEN GlobalStatus {
85 /// isLoaded - True if the global is ever loaded. If the global isn't ever
86 /// loaded it can be deleted.
89 /// StoredType - Keep track of what stores to the global look like.
92 /// NotStored - There is no store to this global. It can thus be marked
96 /// isInitializerStored - This global is stored to, but the only thing
97 /// stored is the constant it was initialized with. This is only tracked
98 /// for scalar globals.
101 /// isStoredOnce - This global is stored to, but only its initializer and
102 /// one other value is ever stored to it. If this global isStoredOnce, we
103 /// track the value stored to it in StoredOnceValue below. This is only
104 /// tracked for scalar globals.
107 /// isStored - This global is stored to by multiple values or something else
108 /// that we cannot track.
112 /// StoredOnceValue - If only one value (besides the initializer constant) is
113 /// ever stored to this global, keep track of what value it is.
114 Value *StoredOnceValue;
116 /// AccessingFunction/HasMultipleAccessingFunctions - These start out
117 /// null/false. When the first accessing function is noticed, it is recorded.
118 /// When a second different accessing function is noticed,
119 /// HasMultipleAccessingFunctions is set to true.
120 Function *AccessingFunction;
121 bool HasMultipleAccessingFunctions;
123 /// HasNonInstructionUser - Set to true if this global has a user that is not
124 /// an instruction (e.g. a constant expr or GV initializer).
125 bool HasNonInstructionUser;
127 /// HasPHIUser - Set to true if this global has a user that is a PHI node.
130 GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0),
131 AccessingFunction(0), HasMultipleAccessingFunctions(false),
132 HasNonInstructionUser(false), HasPHIUser(false) {}
137 /// ConstantIsDead - Return true if the specified constant is (transitively)
138 /// dead. The constant may be used by other constants (e.g. constant arrays and
139 /// constant exprs) as long as they are dead, but it cannot be used by anything
141 static bool ConstantIsDead(Constant *C) {
142 if (isa<GlobalValue>(C)) return false;
144 for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI)
145 if (Constant *CU = dyn_cast<Constant>(*UI)) {
146 if (!ConstantIsDead(CU)) return false;
153 /// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
154 /// structure. If the global has its address taken, return true to indicate we
155 /// can't do anything with it.
157 static bool AnalyzeGlobal(Value *V, GlobalStatus &GS,
158 std::set<PHINode*> &PHIUsers) {
159 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
160 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
161 GS.HasNonInstructionUser = true;
163 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
165 } else if (Instruction *I = dyn_cast<Instruction>(*UI)) {
166 if (!GS.HasMultipleAccessingFunctions) {
167 Function *F = I->getParent()->getParent();
168 if (GS.AccessingFunction == 0)
169 GS.AccessingFunction = F;
170 else if (GS.AccessingFunction != F)
171 GS.HasMultipleAccessingFunctions = true;
173 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
175 if (LI->isVolatile()) return true; // Don't hack on volatile loads.
176 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
177 // Don't allow a store OF the address, only stores TO the address.
178 if (SI->getOperand(0) == V) return true;
180 if (SI->isVolatile()) return true; // Don't hack on volatile stores.
182 // If this is a direct store to the global (i.e., the global is a scalar
183 // value, not an aggregate), keep more specific information about
185 if (GS.StoredType != GlobalStatus::isStored) {
186 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){
187 Value *StoredVal = SI->getOperand(0);
188 if (StoredVal == GV->getInitializer()) {
189 if (GS.StoredType < GlobalStatus::isInitializerStored)
190 GS.StoredType = GlobalStatus::isInitializerStored;
191 } else if (isa<LoadInst>(StoredVal) &&
192 cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
194 if (GS.StoredType < GlobalStatus::isInitializerStored)
195 GS.StoredType = GlobalStatus::isInitializerStored;
196 } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
197 GS.StoredType = GlobalStatus::isStoredOnce;
198 GS.StoredOnceValue = StoredVal;
199 } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
200 GS.StoredOnceValue == StoredVal) {
203 GS.StoredType = GlobalStatus::isStored;
206 GS.StoredType = GlobalStatus::isStored;
209 } else if (isa<GetElementPtrInst>(I)) {
210 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
211 } else if (isa<SelectInst>(I)) {
212 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
213 } else if (PHINode *PN = dyn_cast<PHINode>(I)) {
214 // PHI nodes we can check just like select or GEP instructions, but we
215 // have to be careful about infinite recursion.
216 if (PHIUsers.insert(PN).second) // Not already visited.
217 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
218 GS.HasPHIUser = true;
219 } else if (isa<CmpInst>(I)) {
220 } else if (isa<MemCpyInst>(I) || isa<MemMoveInst>(I)) {
221 if (I->getOperand(1) == V)
222 GS.StoredType = GlobalStatus::isStored;
223 if (I->getOperand(2) == V)
225 } else if (isa<MemSetInst>(I)) {
226 assert(I->getOperand(1) == V && "Memset only takes one pointer!");
227 GS.StoredType = GlobalStatus::isStored;
229 return true; // Any other non-load instruction might take address!
231 } else if (Constant *C = dyn_cast<Constant>(*UI)) {
232 GS.HasNonInstructionUser = true;
233 // We might have a dead and dangling constant hanging off of here.
234 if (!ConstantIsDead(C))
237 GS.HasNonInstructionUser = true;
238 // Otherwise must be some other user.
245 static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) {
246 ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
248 unsigned IdxV = CI->getZExtValue();
250 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
251 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
252 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
253 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
254 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) {
255 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
256 } else if (isa<ConstantAggregateZero>(Agg)) {
257 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
258 if (IdxV < STy->getNumElements())
259 return Constant::getNullValue(STy->getElementType(IdxV));
260 } else if (const SequentialType *STy =
261 dyn_cast<SequentialType>(Agg->getType())) {
262 return Constant::getNullValue(STy->getElementType());
264 } else if (isa<UndefValue>(Agg)) {
265 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
266 if (IdxV < STy->getNumElements())
267 return UndefValue::get(STy->getElementType(IdxV));
268 } else if (const SequentialType *STy =
269 dyn_cast<SequentialType>(Agg->getType())) {
270 return UndefValue::get(STy->getElementType());
277 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
278 /// users of the global, cleaning up the obvious ones. This is largely just a
279 /// quick scan over the use list to clean up the easy and obvious cruft. This
280 /// returns true if it made a change.
281 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
282 bool Changed = false;
283 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
286 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
288 // Replace the load with the initializer.
289 LI->replaceAllUsesWith(Init);
290 LI->eraseFromParent();
293 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
294 // Store must be unreachable or storing Init into the global.
295 SI->eraseFromParent();
297 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
298 if (CE->getOpcode() == Instruction::GetElementPtr) {
299 Constant *SubInit = 0;
301 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
302 Changed |= CleanupConstantGlobalUsers(CE, SubInit);
303 } else if (CE->getOpcode() == Instruction::BitCast &&
304 isa<PointerType>(CE->getType())) {
305 // Pointer cast, delete any stores and memsets to the global.
306 Changed |= CleanupConstantGlobalUsers(CE, 0);
309 if (CE->use_empty()) {
310 CE->destroyConstant();
313 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
314 // Do not transform "gepinst (gep constexpr (GV))" here, because forming
315 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
316 // and will invalidate our notion of what Init is.
317 Constant *SubInit = 0;
318 if (!isa<ConstantExpr>(GEP->getOperand(0))) {
320 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
321 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
322 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
324 Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
326 if (GEP->use_empty()) {
327 GEP->eraseFromParent();
330 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
331 if (MI->getRawDest() == V) {
332 MI->eraseFromParent();
336 } else if (Constant *C = dyn_cast<Constant>(U)) {
337 // If we have a chain of dead constantexprs or other things dangling from
338 // us, and if they are all dead, nuke them without remorse.
339 if (ConstantIsDead(C)) {
340 C->destroyConstant();
341 // This could have invalidated UI, start over from scratch.
342 CleanupConstantGlobalUsers(V, Init);
350 /// isSafeSROAElementUse - Return true if the specified instruction is a safe
351 /// user of a derived expression from a global that we want to SROA.
352 static bool isSafeSROAElementUse(Value *V) {
353 // We might have a dead and dangling constant hanging off of here.
354 if (Constant *C = dyn_cast<Constant>(V))
355 return ConstantIsDead(C);
357 Instruction *I = dyn_cast<Instruction>(V);
358 if (!I) return false;
361 if (isa<LoadInst>(I)) return true;
363 // Stores *to* the pointer are ok.
364 if (StoreInst *SI = dyn_cast<StoreInst>(I))
365 return SI->getOperand(0) != V;
367 // Otherwise, it must be a GEP.
368 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I);
369 if (GEPI == 0) return false;
371 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
372 !cast<Constant>(GEPI->getOperand(1))->isNullValue())
375 for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end();
377 if (!isSafeSROAElementUse(*I))
383 /// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value.
384 /// Look at it and its uses and decide whether it is safe to SROA this global.
386 static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
387 // The user of the global must be a GEP Inst or a ConstantExpr GEP.
388 if (!isa<GetElementPtrInst>(U) &&
389 (!isa<ConstantExpr>(U) ||
390 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
393 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
394 // don't like < 3 operand CE's, and we don't like non-constant integer
395 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some
397 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
398 !cast<Constant>(U->getOperand(1))->isNullValue() ||
399 !isa<ConstantInt>(U->getOperand(2)))
402 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
403 ++GEPI; // Skip over the pointer index.
405 // If this is a use of an array allocation, do a bit more checking for sanity.
406 if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
407 uint64_t NumElements = AT->getNumElements();
408 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
410 // Check to make sure that index falls within the array. If not,
411 // something funny is going on, so we won't do the optimization.
413 if (Idx->getZExtValue() >= NumElements)
416 // We cannot scalar repl this level of the array unless any array
417 // sub-indices are in-range constants. In particular, consider:
418 // A[0][i]. We cannot know that the user isn't doing invalid things like
419 // allowing i to index an out-of-range subscript that accesses A[1].
421 // Scalar replacing *just* the outer index of the array is probably not
422 // going to be a win anyway, so just give up.
423 for (++GEPI; // Skip array index.
424 GEPI != E && (isa<ArrayType>(*GEPI) || isa<VectorType>(*GEPI));
426 uint64_t NumElements;
427 if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
428 NumElements = SubArrayTy->getNumElements();
430 NumElements = cast<VectorType>(*GEPI)->getNumElements();
432 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
433 if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
438 for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I)
439 if (!isSafeSROAElementUse(*I))
444 /// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
445 /// is safe for us to perform this transformation.
447 static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
448 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
450 if (!IsUserOfGlobalSafeForSRA(*UI, GV))
457 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
458 /// variable. This opens the door for other optimizations by exposing the
459 /// behavior of the program in a more fine-grained way. We have determined that
460 /// this transformation is safe already. We return the first global variable we
461 /// insert so that the caller can reprocess it.
462 static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) {
463 // Make sure this global only has simple uses that we can SRA.
464 if (!GlobalUsersSafeToSRA(GV))
467 assert(GV->hasInternalLinkage() && !GV->isConstant());
468 Constant *Init = GV->getInitializer();
469 const Type *Ty = Init->getType();
471 std::vector<GlobalVariable*> NewGlobals;
472 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
474 // Get the alignment of the global, either explicit or target-specific.
475 unsigned StartAlignment = GV->getAlignment();
476 if (StartAlignment == 0)
477 StartAlignment = TD.getABITypeAlignment(GV->getType());
479 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
480 NewGlobals.reserve(STy->getNumElements());
481 const StructLayout &Layout = *TD.getStructLayout(STy);
482 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
483 Constant *In = getAggregateConstantElement(Init,
484 ConstantInt::get(Type::Int32Ty, i));
485 assert(In && "Couldn't get element of initializer?");
486 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
487 GlobalVariable::InternalLinkage,
488 In, GV->getName()+"."+utostr(i),
491 GV->getType()->getAddressSpace());
492 Globals.insert(GV, NGV);
493 NewGlobals.push_back(NGV);
495 // Calculate the known alignment of the field. If the original aggregate
496 // had 256 byte alignment for example, something might depend on that:
497 // propagate info to each field.
498 uint64_t FieldOffset = Layout.getElementOffset(i);
499 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset);
500 if (NewAlign > TD.getABITypeAlignment(STy->getElementType(i)))
501 NGV->setAlignment(NewAlign);
503 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
504 unsigned NumElements = 0;
505 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
506 NumElements = ATy->getNumElements();
508 NumElements = cast<VectorType>(STy)->getNumElements();
510 if (NumElements > 16 && GV->hasNUsesOrMore(16))
511 return 0; // It's not worth it.
512 NewGlobals.reserve(NumElements);
514 uint64_t EltSize = TD.getABITypeSize(STy->getElementType());
515 unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType());
516 for (unsigned i = 0, e = NumElements; i != e; ++i) {
517 Constant *In = getAggregateConstantElement(Init,
518 ConstantInt::get(Type::Int32Ty, i));
519 assert(In && "Couldn't get element of initializer?");
521 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
522 GlobalVariable::InternalLinkage,
523 In, GV->getName()+"."+utostr(i),
526 GV->getType()->getAddressSpace());
527 Globals.insert(GV, NGV);
528 NewGlobals.push_back(NGV);
530 // Calculate the known alignment of the field. If the original aggregate
531 // had 256 byte alignment for example, something might depend on that:
532 // propagate info to each field.
533 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
534 if (NewAlign > EltAlign)
535 NGV->setAlignment(NewAlign);
539 if (NewGlobals.empty())
542 DOUT << "PERFORMING GLOBAL SRA ON: " << *GV;
544 Constant *NullInt = Constant::getNullValue(Type::Int32Ty);
546 // Loop over all of the uses of the global, replacing the constantexpr geps,
547 // with smaller constantexpr geps or direct references.
548 while (!GV->use_empty()) {
549 User *GEP = GV->use_back();
550 assert(((isa<ConstantExpr>(GEP) &&
551 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
552 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
554 // Ignore the 1th operand, which has to be zero or else the program is quite
555 // broken (undefined). Get the 2nd operand, which is the structure or array
557 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
558 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
560 Value *NewPtr = NewGlobals[Val];
562 // Form a shorter GEP if needed.
563 if (GEP->getNumOperands() > 3) {
564 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
565 SmallVector<Constant*, 8> Idxs;
566 Idxs.push_back(NullInt);
567 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
568 Idxs.push_back(CE->getOperand(i));
569 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr),
570 &Idxs[0], Idxs.size());
572 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
573 SmallVector<Value*, 8> Idxs;
574 Idxs.push_back(NullInt);
575 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
576 Idxs.push_back(GEPI->getOperand(i));
577 NewPtr = GetElementPtrInst::Create(NewPtr, Idxs.begin(), Idxs.end(),
578 GEPI->getName()+"."+utostr(Val), GEPI);
581 GEP->replaceAllUsesWith(NewPtr);
583 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
584 GEPI->eraseFromParent();
586 cast<ConstantExpr>(GEP)->destroyConstant();
589 // Delete the old global, now that it is dead.
593 // Loop over the new globals array deleting any globals that are obviously
594 // dead. This can arise due to scalarization of a structure or an array that
595 // has elements that are dead.
596 unsigned FirstGlobal = 0;
597 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
598 if (NewGlobals[i]->use_empty()) {
599 Globals.erase(NewGlobals[i]);
600 if (FirstGlobal == i) ++FirstGlobal;
603 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
606 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
607 /// value will trap if the value is dynamically null. PHIs keeps track of any
608 /// phi nodes we've seen to avoid reprocessing them.
609 static bool AllUsesOfValueWillTrapIfNull(Value *V,
610 SmallPtrSet<PHINode*, 8> &PHIs) {
611 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
612 if (isa<LoadInst>(*UI)) {
614 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
615 if (SI->getOperand(0) == V) {
616 //cerr << "NONTRAPPING USE: " << **UI;
617 return false; // Storing the value.
619 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
620 if (CI->getOperand(0) != V) {
621 //cerr << "NONTRAPPING USE: " << **UI;
622 return false; // Not calling the ptr
624 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
625 if (II->getOperand(0) != V) {
626 //cerr << "NONTRAPPING USE: " << **UI;
627 return false; // Not calling the ptr
629 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(*UI)) {
630 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
631 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
632 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
633 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
634 // If we've already seen this phi node, ignore it, it has already been
637 return AllUsesOfValueWillTrapIfNull(PN, PHIs);
638 } else if (isa<ICmpInst>(*UI) &&
639 isa<ConstantPointerNull>(UI->getOperand(1))) {
640 // Ignore setcc X, null
642 //cerr << "NONTRAPPING USE: " << **UI;
648 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
649 /// from GV will trap if the loaded value is null. Note that this also permits
650 /// comparisons of the loaded value against null, as a special case.
651 static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) {
652 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI)
653 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
654 SmallPtrSet<PHINode*, 8> PHIs;
655 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
657 } else if (isa<StoreInst>(*UI)) {
658 // Ignore stores to the global.
660 // We don't know or understand this user, bail out.
661 //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI;
668 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
669 bool Changed = false;
670 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
671 Instruction *I = cast<Instruction>(*UI++);
672 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
673 LI->setOperand(0, NewV);
675 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
676 if (SI->getOperand(1) == V) {
677 SI->setOperand(1, NewV);
680 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
681 if (I->getOperand(0) == V) {
682 // Calling through the pointer! Turn into a direct call, but be careful
683 // that the pointer is not also being passed as an argument.
684 I->setOperand(0, NewV);
686 bool PassedAsArg = false;
687 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
688 if (I->getOperand(i) == V) {
690 I->setOperand(i, NewV);
694 // Being passed as an argument also. Be careful to not invalidate UI!
698 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
699 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
700 ConstantExpr::getCast(CI->getOpcode(),
701 NewV, CI->getType()));
702 if (CI->use_empty()) {
704 CI->eraseFromParent();
706 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
707 // Should handle GEP here.
708 SmallVector<Constant*, 8> Idxs;
709 Idxs.reserve(GEPI->getNumOperands()-1);
710 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
712 if (Constant *C = dyn_cast<Constant>(*i))
716 if (Idxs.size() == GEPI->getNumOperands()-1)
717 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
718 ConstantExpr::getGetElementPtr(NewV, &Idxs[0],
720 if (GEPI->use_empty()) {
722 GEPI->eraseFromParent();
731 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
732 /// value stored into it. If there are uses of the loaded value that would trap
733 /// if the loaded value is dynamically null, then we know that they cannot be
734 /// reachable with a null optimize away the load.
735 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) {
736 std::vector<LoadInst*> Loads;
737 bool Changed = false;
739 // Replace all uses of loads with uses of uses of the stored value.
740 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end();
742 if (LoadInst *LI = dyn_cast<LoadInst>(*GUI)) {
744 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
746 // If we get here we could have stores, selects, or phi nodes whose values
748 assert((isa<StoreInst>(*GUI) || isa<PHINode>(*GUI) ||
749 isa<SelectInst>(*GUI) || isa<ConstantExpr>(*GUI)) &&
750 "Only expect load and stores!");
754 DOUT << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV;
758 // Delete all of the loads we can, keeping track of whether we nuked them all!
759 bool AllLoadsGone = true;
760 while (!Loads.empty()) {
761 LoadInst *L = Loads.back();
762 if (L->use_empty()) {
763 L->eraseFromParent();
766 AllLoadsGone = false;
771 // If we nuked all of the loads, then none of the stores are needed either,
772 // nor is the global.
774 DOUT << " *** GLOBAL NOW DEAD!\n";
775 CleanupConstantGlobalUsers(GV, 0);
776 if (GV->use_empty()) {
777 GV->eraseFromParent();
785 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
786 /// instructions that are foldable.
787 static void ConstantPropUsersOf(Value *V) {
788 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
789 if (Instruction *I = dyn_cast<Instruction>(*UI++))
790 if (Constant *NewC = ConstantFoldInstruction(I)) {
791 I->replaceAllUsesWith(NewC);
793 // Advance UI to the next non-I use to avoid invalidating it!
794 // Instructions could multiply use V.
795 while (UI != E && *UI == I)
797 I->eraseFromParent();
801 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
802 /// variable, and transforms the program as if it always contained the result of
803 /// the specified malloc. Because it is always the result of the specified
804 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
805 /// malloc into a global, and any loads of GV as uses of the new global.
806 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
808 DOUT << "PROMOTING MALLOC GLOBAL: " << *GV << " MALLOC = " << *MI;
809 ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize());
811 if (NElements->getZExtValue() != 1) {
812 // If we have an array allocation, transform it to a single element
813 // allocation to make the code below simpler.
814 Type *NewTy = ArrayType::get(MI->getAllocatedType(),
815 NElements->getZExtValue());
817 new MallocInst(NewTy, Constant::getNullValue(Type::Int32Ty),
818 MI->getAlignment(), MI->getName(), MI);
820 Indices[0] = Indices[1] = Constant::getNullValue(Type::Int32Ty);
821 Value *NewGEP = GetElementPtrInst::Create(NewMI, Indices, Indices + 2,
822 NewMI->getName()+".el0", MI);
823 MI->replaceAllUsesWith(NewGEP);
824 MI->eraseFromParent();
828 // Create the new global variable. The contents of the malloc'd memory is
829 // undefined, so initialize with an undef value.
830 Constant *Init = UndefValue::get(MI->getAllocatedType());
831 GlobalVariable *NewGV = new GlobalVariable(MI->getAllocatedType(), false,
832 GlobalValue::InternalLinkage, Init,
833 GV->getName()+".body",
835 GV->isThreadLocal());
836 // FIXME: This new global should have the alignment returned by malloc. Code
837 // could depend on malloc returning large alignment (on the mac, 16 bytes) but
838 // this would only guarantee some lower alignment.
839 GV->getParent()->getGlobalList().insert(GV, NewGV);
841 // Anything that used the malloc now uses the global directly.
842 MI->replaceAllUsesWith(NewGV);
844 Constant *RepValue = NewGV;
845 if (NewGV->getType() != GV->getType()->getElementType())
846 RepValue = ConstantExpr::getBitCast(RepValue,
847 GV->getType()->getElementType());
849 // If there is a comparison against null, we will insert a global bool to
850 // keep track of whether the global was initialized yet or not.
851 GlobalVariable *InitBool =
852 new GlobalVariable(Type::Int1Ty, false, GlobalValue::InternalLinkage,
853 ConstantInt::getFalse(), GV->getName()+".init",
854 (Module *)NULL, GV->isThreadLocal());
855 bool InitBoolUsed = false;
857 // Loop over all uses of GV, processing them in turn.
858 std::vector<StoreInst*> Stores;
859 while (!GV->use_empty())
860 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
861 while (!LI->use_empty()) {
862 Use &LoadUse = LI->use_begin().getUse();
863 if (!isa<ICmpInst>(LoadUse.getUser()))
866 ICmpInst *CI = cast<ICmpInst>(LoadUse.getUser());
867 // Replace the cmp X, 0 with a use of the bool value.
868 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", CI);
870 switch (CI->getPredicate()) {
871 default: assert(0 && "Unknown ICmp Predicate!");
872 case ICmpInst::ICMP_ULT:
873 case ICmpInst::ICMP_SLT:
874 LV = ConstantInt::getFalse(); // X < null -> always false
876 case ICmpInst::ICMP_ULE:
877 case ICmpInst::ICMP_SLE:
878 case ICmpInst::ICMP_EQ:
879 LV = BinaryOperator::CreateNot(LV, "notinit", CI);
881 case ICmpInst::ICMP_NE:
882 case ICmpInst::ICMP_UGE:
883 case ICmpInst::ICMP_SGE:
884 case ICmpInst::ICMP_UGT:
885 case ICmpInst::ICMP_SGT:
888 CI->replaceAllUsesWith(LV);
889 CI->eraseFromParent();
892 LI->eraseFromParent();
894 StoreInst *SI = cast<StoreInst>(GV->use_back());
895 // The global is initialized when the store to it occurs.
896 new StoreInst(ConstantInt::getTrue(), InitBool, SI);
897 SI->eraseFromParent();
900 // If the initialization boolean was used, insert it, otherwise delete it.
902 while (!InitBool->use_empty()) // Delete initializations
903 cast<Instruction>(InitBool->use_back())->eraseFromParent();
906 GV->getParent()->getGlobalList().insert(GV, InitBool);
909 // Now the GV is dead, nuke it and the malloc.
910 GV->eraseFromParent();
911 MI->eraseFromParent();
913 // To further other optimizations, loop over all users of NewGV and try to
914 // constant prop them. This will promote GEP instructions with constant
915 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
916 ConstantPropUsersOf(NewGV);
917 if (RepValue != NewGV)
918 ConstantPropUsersOf(RepValue);
923 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
924 /// to make sure that there are no complex uses of V. We permit simple things
925 /// like dereferencing the pointer, but not storing through the address, unless
926 /// it is to the specified global.
927 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V,
929 SmallPtrSet<PHINode*, 8> &PHIs) {
930 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
931 if (isa<LoadInst>(*UI) || isa<CmpInst>(*UI)) {
933 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
934 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
935 return false; // Storing the pointer itself... bad.
936 // Otherwise, storing through it, or storing into GV... fine.
937 } else if (isa<GetElementPtrInst>(*UI)) {
938 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(cast<Instruction>(*UI),
941 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
942 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI
945 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
953 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
954 /// somewhere. Transform all uses of the allocation into loads from the
955 /// global and uses of the resultant pointer. Further, delete the store into
956 /// GV. This assumes that these value pass the
957 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
958 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
959 GlobalVariable *GV) {
960 while (!Alloc->use_empty()) {
961 Instruction *U = cast<Instruction>(*Alloc->use_begin());
962 Instruction *InsertPt = U;
963 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
964 // If this is the store of the allocation into the global, remove it.
965 if (SI->getOperand(1) == GV) {
966 SI->eraseFromParent();
969 } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
970 // Insert the load in the corresponding predecessor, not right before the
972 unsigned PredNo = Alloc->use_begin().getOperandNo()/2;
973 InsertPt = PN->getIncomingBlock(PredNo)->getTerminator();
976 // Insert a load from the global, and use it instead of the malloc.
977 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
978 U->replaceUsesOfWith(Alloc, NL);
982 /// GlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
983 /// GV are simple enough to perform HeapSRA, return true.
984 static bool GlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV,
986 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;
988 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
989 // We permit two users of the load: setcc comparing against the null
990 // pointer, and a getelementptr of a specific form.
991 for (Value::use_iterator UI = LI->use_begin(), E = LI->use_end();
993 // Comparison against null is ok.
994 if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) {
995 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1000 // getelementptr is also ok, but only a simple form.
1001 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
1002 // Must index into the array and into the struct.
1003 if (GEPI->getNumOperands() < 3)
1006 // Otherwise the GEP is ok.
1010 if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
1011 // We have a phi of a load from the global. We can only handle this
1012 // if the other PHI'd values are actually the same. In this case,
1013 // the rewriter will just drop the phi entirely.
1014 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1015 Value *IV = PN->getIncomingValue(i);
1016 if (IV == LI) continue; // Trivial the same.
1018 // If the phi'd value is from the malloc that initializes the value,
1020 if (IV == MI) continue;
1022 // Otherwise, we don't know what it is.
1028 // Otherwise we don't know what this is, not ok.
1035 /// GetHeapSROALoad - Return the load for the specified field of the HeapSROA'd
1036 /// value, lazily creating it on demand.
1037 static Value *GetHeapSROALoad(Instruction *Load, unsigned FieldNo,
1038 const std::vector<GlobalVariable*> &FieldGlobals,
1039 std::vector<Value *> &InsertedLoadsForPtr) {
1040 if (InsertedLoadsForPtr.size() <= FieldNo)
1041 InsertedLoadsForPtr.resize(FieldNo+1);
1042 if (InsertedLoadsForPtr[FieldNo] == 0)
1043 InsertedLoadsForPtr[FieldNo] = new LoadInst(FieldGlobals[FieldNo],
1044 Load->getName()+".f" +
1045 utostr(FieldNo), Load);
1046 return InsertedLoadsForPtr[FieldNo];
1049 /// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1050 /// the load, rewrite the derived value to use the HeapSRoA'd load.
1051 static void RewriteHeapSROALoadUser(LoadInst *Load, Instruction *LoadUser,
1052 const std::vector<GlobalVariable*> &FieldGlobals,
1053 std::vector<Value *> &InsertedLoadsForPtr) {
1054 // If this is a comparison against null, handle it.
1055 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1056 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1057 // If we have a setcc of the loaded pointer, we can use a setcc of any
1060 if (InsertedLoadsForPtr.empty()) {
1061 NPtr = GetHeapSROALoad(Load, 0, FieldGlobals, InsertedLoadsForPtr);
1063 NPtr = InsertedLoadsForPtr.back();
1066 Value *New = new ICmpInst(SCI->getPredicate(), NPtr,
1067 Constant::getNullValue(NPtr->getType()),
1068 SCI->getName(), SCI);
1069 SCI->replaceAllUsesWith(New);
1070 SCI->eraseFromParent();
1074 // Handle 'getelementptr Ptr, Idx, uint FieldNo ...'
1075 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1076 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1077 && "Unexpected GEPI!");
1079 // Load the pointer for this field.
1080 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1081 Value *NewPtr = GetHeapSROALoad(Load, FieldNo,
1082 FieldGlobals, InsertedLoadsForPtr);
1084 // Create the new GEP idx vector.
1085 SmallVector<Value*, 8> GEPIdx;
1086 GEPIdx.push_back(GEPI->getOperand(1));
1087 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1089 Value *NGEPI = GetElementPtrInst::Create(NewPtr,
1090 GEPIdx.begin(), GEPIdx.end(),
1091 GEPI->getName(), GEPI);
1092 GEPI->replaceAllUsesWith(NGEPI);
1093 GEPI->eraseFromParent();
1097 // Handle PHI nodes. PHI nodes must be merging in the same values, plus
1098 // potentially the original malloc. Insert phi nodes for each field, then
1099 // process uses of the PHI.
1100 PHINode *PN = cast<PHINode>(LoadUser);
1101 std::vector<Value *> PHIsForField;
1102 PHIsForField.resize(FieldGlobals.size());
1103 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1104 Value *LoadV = GetHeapSROALoad(Load, i, FieldGlobals, InsertedLoadsForPtr);
1106 PHINode *FieldPN = PHINode::Create(LoadV->getType(),
1107 PN->getName()+"."+utostr(i), PN);
1108 // Fill in the predecessor values.
1109 for (unsigned pred = 0, e = PN->getNumIncomingValues(); pred != e; ++pred) {
1110 // Each predecessor either uses the load or the original malloc.
1111 Value *InVal = PN->getIncomingValue(pred);
1112 BasicBlock *BB = PN->getIncomingBlock(pred);
1114 if (isa<MallocInst>(InVal)) {
1115 // Insert a reload from the global in the predecessor.
1116 NewVal = GetHeapSROALoad(BB->getTerminator(), i, FieldGlobals,
1119 NewVal = InsertedLoadsForPtr[i];
1121 FieldPN->addIncoming(NewVal, BB);
1123 PHIsForField[i] = FieldPN;
1126 // Since PHIsForField specifies a phi for every input value, the lazy inserter
1127 // will never insert a load.
1128 while (!PN->use_empty())
1129 RewriteHeapSROALoadUser(Load, PN->use_back(), FieldGlobals, PHIsForField);
1130 PN->eraseFromParent();
1133 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1134 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1135 /// use FieldGlobals instead. All uses of loaded values satisfy
1136 /// GlobalLoadUsesSimpleEnoughForHeapSRA.
1137 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1138 const std::vector<GlobalVariable*> &FieldGlobals) {
1139 std::vector<Value *> InsertedLoadsForPtr;
1140 //InsertedLoadsForPtr.resize(FieldGlobals.size());
1141 while (!Load->use_empty())
1142 RewriteHeapSROALoadUser(Load, Load->use_back(),
1143 FieldGlobals, InsertedLoadsForPtr);
1146 /// PerformHeapAllocSRoA - MI is an allocation of an array of structures. Break
1147 /// it up into multiple allocations of arrays of the fields.
1148 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){
1149 DOUT << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *MI;
1150 const StructType *STy = cast<StructType>(MI->getAllocatedType());
1152 // There is guaranteed to be at least one use of the malloc (storing
1153 // it into GV). If there are other uses, change them to be uses of
1154 // the global to simplify later code. This also deletes the store
1156 ReplaceUsesOfMallocWithGlobal(MI, GV);
1158 // Okay, at this point, there are no users of the malloc. Insert N
1159 // new mallocs at the same place as MI, and N globals.
1160 std::vector<GlobalVariable*> FieldGlobals;
1161 std::vector<MallocInst*> FieldMallocs;
1163 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1164 const Type *FieldTy = STy->getElementType(FieldNo);
1165 const Type *PFieldTy = PointerType::getUnqual(FieldTy);
1167 GlobalVariable *NGV =
1168 new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage,
1169 Constant::getNullValue(PFieldTy),
1170 GV->getName() + ".f" + utostr(FieldNo), GV,
1171 GV->isThreadLocal());
1172 FieldGlobals.push_back(NGV);
1174 MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(),
1175 MI->getName() + ".f" + utostr(FieldNo),MI);
1176 FieldMallocs.push_back(NMI);
1177 new StoreInst(NMI, NGV, MI);
1180 // The tricky aspect of this transformation is handling the case when malloc
1181 // fails. In the original code, malloc failing would set the result pointer
1182 // of malloc to null. In this case, some mallocs could succeed and others
1183 // could fail. As such, we emit code that looks like this:
1184 // F0 = malloc(field0)
1185 // F1 = malloc(field1)
1186 // F2 = malloc(field2)
1187 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1188 // if (F0) { free(F0); F0 = 0; }
1189 // if (F1) { free(F1); F1 = 0; }
1190 // if (F2) { free(F2); F2 = 0; }
1192 Value *RunningOr = 0;
1193 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1194 Value *Cond = new ICmpInst(ICmpInst::ICMP_EQ, FieldMallocs[i],
1195 Constant::getNullValue(FieldMallocs[i]->getType()),
1198 RunningOr = Cond; // First seteq
1200 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", MI);
1203 // Split the basic block at the old malloc.
1204 BasicBlock *OrigBB = MI->getParent();
1205 BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont");
1207 // Create the block to check the first condition. Put all these blocks at the
1208 // end of the function as they are unlikely to be executed.
1209 BasicBlock *NullPtrBlock = BasicBlock::Create("malloc_ret_null",
1210 OrigBB->getParent());
1212 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1213 // branch on RunningOr.
1214 OrigBB->getTerminator()->eraseFromParent();
1215 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1217 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1218 // pointer, because some may be null while others are not.
1219 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1220 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1221 Value *Cmp = new ICmpInst(ICmpInst::ICMP_NE, GVVal,
1222 Constant::getNullValue(GVVal->getType()),
1223 "tmp", NullPtrBlock);
1224 BasicBlock *FreeBlock = BasicBlock::Create("free_it", OrigBB->getParent());
1225 BasicBlock *NextBlock = BasicBlock::Create("next", OrigBB->getParent());
1226 BranchInst::Create(FreeBlock, NextBlock, Cmp, NullPtrBlock);
1228 // Fill in FreeBlock.
1229 new FreeInst(GVVal, FreeBlock);
1230 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1232 BranchInst::Create(NextBlock, FreeBlock);
1234 NullPtrBlock = NextBlock;
1237 BranchInst::Create(ContBB, NullPtrBlock);
1239 // MI is no longer needed, remove it.
1240 MI->eraseFromParent();
1243 // Okay, the malloc site is completely handled. All of the uses of GV are now
1244 // loads, and all uses of those loads are simple. Rewrite them to use loads
1245 // of the per-field globals instead.
1246 while (!GV->use_empty()) {
1247 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
1248 RewriteUsesOfLoadForHeapSRoA(LI, FieldGlobals);
1249 LI->eraseFromParent();
1251 // Must be a store of null.
1252 StoreInst *SI = cast<StoreInst>(GV->use_back());
1253 assert(isa<Constant>(SI->getOperand(0)) &&
1254 cast<Constant>(SI->getOperand(0))->isNullValue() &&
1255 "Unexpected heap-sra user!");
1257 // Insert a store of null into each global.
1258 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1260 Constant::getNullValue(FieldGlobals[i]->getType()->getElementType());
1261 new StoreInst(Null, FieldGlobals[i], SI);
1263 // Erase the original store.
1264 SI->eraseFromParent();
1268 // The old global is now dead, remove it.
1269 GV->eraseFromParent();
1272 return FieldGlobals[0];
1275 /// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1276 /// pointer global variable with a single value stored it that is a malloc or
1278 static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
1280 Module::global_iterator &GVI,
1282 // If this is a malloc of an abstract type, don't touch it.
1283 if (!MI->getAllocatedType()->isSized())
1286 // We can't optimize this global unless all uses of it are *known* to be
1287 // of the malloc value, not of the null initializer value (consider a use
1288 // that compares the global's value against zero to see if the malloc has
1289 // been reached). To do this, we check to see if all uses of the global
1290 // would trap if the global were null: this proves that they must all
1291 // happen after the malloc.
1292 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1295 // We can't optimize this if the malloc itself is used in a complex way,
1296 // for example, being stored into multiple globals. This allows the
1297 // malloc to be stored into the specified global, loaded setcc'd, and
1298 // GEP'd. These are all things we could transform to using the global
1301 SmallPtrSet<PHINode*, 8> PHIs;
1302 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV, PHIs))
1307 // If we have a global that is only initialized with a fixed size malloc,
1308 // transform the program to use global memory instead of malloc'd memory.
1309 // This eliminates dynamic allocation, avoids an indirection accessing the
1310 // data, and exposes the resultant global to further GlobalOpt.
1311 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) {
1312 // Restrict this transformation to only working on small allocations
1313 // (2048 bytes currently), as we don't want to introduce a 16M global or
1315 if (NElements->getZExtValue()*
1316 TD.getABITypeSize(MI->getAllocatedType()) < 2048) {
1317 GVI = OptimizeGlobalAddressOfMalloc(GV, MI);
1322 // If the allocation is an array of structures, consider transforming this
1323 // into multiple malloc'd arrays, one for each field. This is basically
1324 // SRoA for malloc'd memory.
1325 if (const StructType *AllocTy =
1326 dyn_cast<StructType>(MI->getAllocatedType())) {
1327 // This the structure has an unreasonable number of fields, leave it
1329 if (AllocTy->getNumElements() <= 16 && AllocTy->getNumElements() > 0 &&
1330 GlobalLoadUsesSimpleEnoughForHeapSRA(GV, MI)) {
1331 GVI = PerformHeapAllocSRoA(GV, MI);
1339 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1340 // that only one value (besides its initializer) is ever stored to the global.
1341 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1342 Module::global_iterator &GVI,
1344 if (CastInst *CI = dyn_cast<CastInst>(StoredOnceVal))
1345 StoredOnceVal = CI->getOperand(0);
1346 else if (GetElementPtrInst *GEPI =dyn_cast<GetElementPtrInst>(StoredOnceVal)){
1347 // "getelementptr Ptr, 0, 0, 0" is really just a cast.
1348 if (GEPI->hasAllZeroIndices())
1349 StoredOnceVal = GEPI->getOperand(0);
1352 // If we are dealing with a pointer global that is initialized to null and
1353 // only has one (non-null) value stored into it, then we can optimize any
1354 // users of the loaded value (often calls and loads) that would trap if the
1356 if (isa<PointerType>(GV->getInitializer()->getType()) &&
1357 GV->getInitializer()->isNullValue()) {
1358 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1359 if (GV->getInitializer()->getType() != SOVC->getType())
1360 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1362 // Optimize away any trapping uses of the loaded value.
1363 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1365 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) {
1366 if (TryToOptimizeStoreOfMallocToGlobal(GV, MI, GVI, TD))
1374 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1375 /// two values ever stored into GV are its initializer and OtherVal. See if we
1376 /// can shrink the global into a boolean and select between the two values
1377 /// whenever it is used. This exposes the values to other scalar optimizations.
1378 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1379 const Type *GVElType = GV->getType()->getElementType();
1381 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1382 // an FP value or vector, don't do this optimization because a select between
1383 // them is very expensive and unlikely to lead to later simplification.
1384 if (GVElType == Type::Int1Ty || GVElType->isFloatingPoint() ||
1385 isa<VectorType>(GVElType))
1388 // Walk the use list of the global seeing if all the uses are load or store.
1389 // If there is anything else, bail out.
1390 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I)
1391 if (!isa<LoadInst>(I) && !isa<StoreInst>(I))
1394 DOUT << " *** SHRINKING TO BOOL: " << *GV;
1396 // Create the new global, initializing it to false.
1397 GlobalVariable *NewGV = new GlobalVariable(Type::Int1Ty, false,
1398 GlobalValue::InternalLinkage, ConstantInt::getFalse(),
1401 GV->isThreadLocal());
1402 GV->getParent()->getGlobalList().insert(GV, NewGV);
1404 Constant *InitVal = GV->getInitializer();
1405 assert(InitVal->getType() != Type::Int1Ty && "No reason to shrink to bool!");
1407 // If initialized to zero and storing one into the global, we can use a cast
1408 // instead of a select to synthesize the desired value.
1409 bool IsOneZero = false;
1410 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1411 IsOneZero = InitVal->isNullValue() && CI->isOne();
1413 while (!GV->use_empty()) {
1414 Instruction *UI = cast<Instruction>(GV->use_back());
1415 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1416 // Change the store into a boolean store.
1417 bool StoringOther = SI->getOperand(0) == OtherVal;
1418 // Only do this if we weren't storing a loaded value.
1420 if (StoringOther || SI->getOperand(0) == InitVal)
1421 StoreVal = ConstantInt::get(Type::Int1Ty, StoringOther);
1423 // Otherwise, we are storing a previously loaded copy. To do this,
1424 // change the copy from copying the original value to just copying the
1426 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1428 // If we're already replaced the input, StoredVal will be a cast or
1429 // select instruction. If not, it will be a load of the original
1431 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1432 assert(LI->getOperand(0) == GV && "Not a copy!");
1433 // Insert a new load, to preserve the saved value.
1434 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1436 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1437 "This is not a form that we understand!");
1438 StoreVal = StoredVal->getOperand(0);
1439 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1442 new StoreInst(StoreVal, NewGV, SI);
1444 // Change the load into a load of bool then a select.
1445 LoadInst *LI = cast<LoadInst>(UI);
1446 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1449 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1451 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1453 LI->replaceAllUsesWith(NSI);
1455 UI->eraseFromParent();
1458 GV->eraseFromParent();
1463 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1464 /// it if possible. If we make a change, return true.
1465 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1466 Module::global_iterator &GVI) {
1467 std::set<PHINode*> PHIUsers;
1469 GV->removeDeadConstantUsers();
1471 if (GV->use_empty()) {
1472 DOUT << "GLOBAL DEAD: " << *GV;
1473 GV->eraseFromParent();
1478 if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
1480 cerr << "Global: " << *GV;
1481 cerr << " isLoaded = " << GS.isLoaded << "\n";
1482 cerr << " StoredType = ";
1483 switch (GS.StoredType) {
1484 case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break;
1485 case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break;
1486 case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break;
1487 case GlobalStatus::isStored: cerr << "stored\n"; break;
1489 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
1490 cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n";
1491 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
1492 cerr << " AccessingFunction = " << GS.AccessingFunction->getName()
1494 cerr << " HasMultipleAccessingFunctions = "
1495 << GS.HasMultipleAccessingFunctions << "\n";
1496 cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n";
1500 // If this is a first class global and has only one accessing function
1501 // and this function is main (which we know is not recursive we can make
1502 // this global a local variable) we replace the global with a local alloca
1503 // in this function.
1505 // NOTE: It doesn't make sense to promote non single-value types since we
1506 // are just replacing static memory to stack memory.
1507 if (!GS.HasMultipleAccessingFunctions &&
1508 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1509 GV->getType()->getElementType()->isSingleValueType() &&
1510 GS.AccessingFunction->getName() == "main" &&
1511 GS.AccessingFunction->hasExternalLinkage()) {
1512 DOUT << "LOCALIZING GLOBAL: " << *GV;
1513 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin();
1514 const Type* ElemTy = GV->getType()->getElementType();
1515 // FIXME: Pass Global's alignment when globals have alignment
1516 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI);
1517 if (!isa<UndefValue>(GV->getInitializer()))
1518 new StoreInst(GV->getInitializer(), Alloca, FirstI);
1520 GV->replaceAllUsesWith(Alloca);
1521 GV->eraseFromParent();
1526 // If the global is never loaded (but may be stored to), it is dead.
1529 DOUT << "GLOBAL NEVER LOADED: " << *GV;
1531 // Delete any stores we can find to the global. We may not be able to
1532 // make it completely dead though.
1533 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1535 // If the global is dead now, delete it.
1536 if (GV->use_empty()) {
1537 GV->eraseFromParent();
1543 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1544 DOUT << "MARKING CONSTANT: " << *GV;
1545 GV->setConstant(true);
1547 // Clean up any obviously simplifiable users now.
1548 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1550 // If the global is dead now, just nuke it.
1551 if (GV->use_empty()) {
1552 DOUT << " *** Marking constant allowed us to simplify "
1553 << "all users and delete global!\n";
1554 GV->eraseFromParent();
1560 } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
1561 if (GlobalVariable *FirstNewGV = SRAGlobal(GV,
1562 getAnalysis<TargetData>())) {
1563 GVI = FirstNewGV; // Don't skip the newly produced globals!
1566 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1567 // If the initial value for the global was an undef value, and if only
1568 // one other value was stored into it, we can just change the
1569 // initializer to be an undef value, then delete all stores to the
1570 // global. This allows us to mark it constant.
1571 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1572 if (isa<UndefValue>(GV->getInitializer())) {
1573 // Change the initial value here.
1574 GV->setInitializer(SOVConstant);
1576 // Clean up any obviously simplifiable users now.
1577 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1579 if (GV->use_empty()) {
1580 DOUT << " *** Substituting initializer allowed us to "
1581 << "simplify all users and delete global!\n";
1582 GV->eraseFromParent();
1591 // Try to optimize globals based on the knowledge that only one value
1592 // (besides its initializer) is ever stored to the global.
1593 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1594 getAnalysis<TargetData>()))
1597 // Otherwise, if the global was not a boolean, we can shrink it to be a
1599 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1600 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1609 /// OnlyCalledDirectly - Return true if the specified function is only called
1610 /// directly. In other words, its address is never taken.
1611 static bool OnlyCalledDirectly(Function *F) {
1612 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1613 Instruction *User = dyn_cast<Instruction>(*UI);
1614 if (!User) return false;
1615 if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false;
1617 // See if the function address is passed as an argument.
1618 for (User::op_iterator i = User->op_begin() + 1, e = User->op_end();
1620 if (*i == F) return false;
1625 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1626 /// function, changing them to FastCC.
1627 static void ChangeCalleesToFastCall(Function *F) {
1628 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1629 CallSite User(cast<Instruction>(*UI));
1630 User.setCallingConv(CallingConv::Fast);
1634 static AttrListPtr StripNest(const AttrListPtr &Attrs) {
1635 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1636 if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0)
1639 // There can be only one.
1640 return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest);
1646 static void RemoveNestAttribute(Function *F) {
1647 F->setAttributes(StripNest(F->getAttributes()));
1648 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1649 CallSite User(cast<Instruction>(*UI));
1650 User.setAttributes(StripNest(User.getAttributes()));
1654 bool GlobalOpt::OptimizeFunctions(Module &M) {
1655 bool Changed = false;
1656 // Optimize functions.
1657 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1659 F->removeDeadConstantUsers();
1660 if (F->use_empty() && (F->hasInternalLinkage() ||
1661 F->hasLinkOnceLinkage())) {
1662 M.getFunctionList().erase(F);
1665 } else if (F->hasInternalLinkage()) {
1666 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1667 OnlyCalledDirectly(F)) {
1668 // If this function has C calling conventions, is not a varargs
1669 // function, and is only called directly, promote it to use the Fast
1670 // calling convention.
1671 F->setCallingConv(CallingConv::Fast);
1672 ChangeCalleesToFastCall(F);
1677 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
1678 OnlyCalledDirectly(F)) {
1679 // The function is not used by a trampoline intrinsic, so it is safe
1680 // to remove the 'nest' attribute.
1681 RemoveNestAttribute(F);
1690 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1691 bool Changed = false;
1692 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1694 GlobalVariable *GV = GVI++;
1695 if (!GV->isConstant() && GV->hasInternalLinkage() &&
1696 GV->hasInitializer())
1697 Changed |= ProcessInternalGlobal(GV, GVI);
1702 /// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1703 /// initializers have an init priority of 65535.
1704 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1705 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1707 if (I->getName() == "llvm.global_ctors") {
1708 // Found it, verify it's an array of { int, void()* }.
1709 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1711 const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1712 if (!STy || STy->getNumElements() != 2 ||
1713 STy->getElementType(0) != Type::Int32Ty) return 0;
1714 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1715 if (!PFTy) return 0;
1716 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1717 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() ||
1718 FTy->getNumParams() != 0)
1721 // Verify that the initializer is simple enough for us to handle.
1722 if (!I->hasInitializer()) return 0;
1723 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1725 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
1726 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(*i)) {
1727 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1730 // Must have a function or null ptr.
1731 if (!isa<Function>(CS->getOperand(1)))
1734 // Init priority must be standard.
1735 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1736 if (!CI || CI->getZExtValue() != 65535)
1747 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1748 /// return a list of the functions and null terminator as a vector.
1749 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1750 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1751 std::vector<Function*> Result;
1752 Result.reserve(CA->getNumOperands());
1753 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
1754 ConstantStruct *CS = cast<ConstantStruct>(*i);
1755 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1760 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1761 /// specified array, returning the new global to use.
1762 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1763 const std::vector<Function*> &Ctors) {
1764 // If we made a change, reassemble the initializer list.
1765 std::vector<Constant*> CSVals;
1766 CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535));
1767 CSVals.push_back(0);
1769 // Create the new init list.
1770 std::vector<Constant*> CAList;
1771 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1773 CSVals[1] = Ctors[i];
1775 const Type *FTy = FunctionType::get(Type::VoidTy,
1776 std::vector<const Type*>(), false);
1777 const PointerType *PFTy = PointerType::getUnqual(FTy);
1778 CSVals[1] = Constant::getNullValue(PFTy);
1779 CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647);
1781 CAList.push_back(ConstantStruct::get(CSVals));
1784 // Create the array initializer.
1785 const Type *StructTy =
1786 cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
1787 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()),
1790 // If we didn't change the number of elements, don't create a new GV.
1791 if (CA->getType() == GCL->getInitializer()->getType()) {
1792 GCL->setInitializer(CA);
1796 // Create the new global and insert it next to the existing list.
1797 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
1798 GCL->getLinkage(), CA, "",
1800 GCL->isThreadLocal());
1801 GCL->getParent()->getGlobalList().insert(GCL, NGV);
1804 // Nuke the old list, replacing any uses with the new one.
1805 if (!GCL->use_empty()) {
1807 if (V->getType() != GCL->getType())
1808 V = ConstantExpr::getBitCast(V, GCL->getType());
1809 GCL->replaceAllUsesWith(V);
1811 GCL->eraseFromParent();
1820 static Constant *getVal(std::map<Value*, Constant*> &ComputedValues,
1822 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
1823 Constant *R = ComputedValues[V];
1824 assert(R && "Reference to an uncomputed value!");
1828 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
1829 /// enough for us to understand. In particular, if it is a cast of something,
1830 /// we punt. We basically just support direct accesses to globals and GEP's of
1831 /// globals. This should be kept up to date with CommitValueTo.
1832 static bool isSimpleEnoughPointerToCommit(Constant *C) {
1833 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
1834 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1835 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1836 return !GV->isDeclaration(); // reject external globals.
1838 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
1839 // Handle a constantexpr gep.
1840 if (CE->getOpcode() == Instruction::GetElementPtr &&
1841 isa<GlobalVariable>(CE->getOperand(0))) {
1842 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1843 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1844 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1845 return GV->hasInitializer() &&
1846 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1851 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
1852 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
1853 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
1854 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
1855 ConstantExpr *Addr, unsigned OpNo) {
1856 // Base case of the recursion.
1857 if (OpNo == Addr->getNumOperands()) {
1858 assert(Val->getType() == Init->getType() && "Type mismatch!");
1862 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
1863 std::vector<Constant*> Elts;
1865 // Break up the constant into its elements.
1866 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
1867 for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i)
1868 Elts.push_back(cast<Constant>(*i));
1869 } else if (isa<ConstantAggregateZero>(Init)) {
1870 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1871 Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
1872 } else if (isa<UndefValue>(Init)) {
1873 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1874 Elts.push_back(UndefValue::get(STy->getElementType(i)));
1876 assert(0 && "This code is out of sync with "
1877 " ConstantFoldLoadThroughGEPConstantExpr");
1880 // Replace the element that we are supposed to.
1881 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
1882 unsigned Idx = CU->getZExtValue();
1883 assert(Idx < STy->getNumElements() && "Struct index out of range!");
1884 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
1886 // Return the modified struct.
1887 return ConstantStruct::get(&Elts[0], Elts.size(), STy->isPacked());
1889 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
1890 const ArrayType *ATy = cast<ArrayType>(Init->getType());
1892 // Break up the array into elements.
1893 std::vector<Constant*> Elts;
1894 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
1895 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
1896 Elts.push_back(cast<Constant>(*i));
1897 } else if (isa<ConstantAggregateZero>(Init)) {
1898 Constant *Elt = Constant::getNullValue(ATy->getElementType());
1899 Elts.assign(ATy->getNumElements(), Elt);
1900 } else if (isa<UndefValue>(Init)) {
1901 Constant *Elt = UndefValue::get(ATy->getElementType());
1902 Elts.assign(ATy->getNumElements(), Elt);
1904 assert(0 && "This code is out of sync with "
1905 " ConstantFoldLoadThroughGEPConstantExpr");
1908 assert(CI->getZExtValue() < ATy->getNumElements());
1909 Elts[CI->getZExtValue()] =
1910 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
1911 return ConstantArray::get(ATy, Elts);
1915 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
1916 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
1917 static void CommitValueTo(Constant *Val, Constant *Addr) {
1918 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
1919 assert(GV->hasInitializer());
1920 GV->setInitializer(Val);
1924 ConstantExpr *CE = cast<ConstantExpr>(Addr);
1925 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1927 Constant *Init = GV->getInitializer();
1928 Init = EvaluateStoreInto(Init, Val, CE, 2);
1929 GV->setInitializer(Init);
1932 /// ComputeLoadResult - Return the value that would be computed by a load from
1933 /// P after the stores reflected by 'memory' have been performed. If we can't
1934 /// decide, return null.
1935 static Constant *ComputeLoadResult(Constant *P,
1936 const std::map<Constant*, Constant*> &Memory) {
1937 // If this memory location has been recently stored, use the stored value: it
1938 // is the most up-to-date.
1939 std::map<Constant*, Constant*>::const_iterator I = Memory.find(P);
1940 if (I != Memory.end()) return I->second;
1943 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
1944 if (GV->hasInitializer())
1945 return GV->getInitializer();
1949 // Handle a constantexpr getelementptr.
1950 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
1951 if (CE->getOpcode() == Instruction::GetElementPtr &&
1952 isa<GlobalVariable>(CE->getOperand(0))) {
1953 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1954 if (GV->hasInitializer())
1955 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1958 return 0; // don't know how to evaluate.
1961 /// EvaluateFunction - Evaluate a call to function F, returning true if
1962 /// successful, false if we can't evaluate it. ActualArgs contains the formal
1963 /// arguments for the function.
1964 static bool EvaluateFunction(Function *F, Constant *&RetVal,
1965 const std::vector<Constant*> &ActualArgs,
1966 std::vector<Function*> &CallStack,
1967 std::map<Constant*, Constant*> &MutatedMemory,
1968 std::vector<GlobalVariable*> &AllocaTmps) {
1969 // Check to see if this function is already executing (recursion). If so,
1970 // bail out. TODO: we might want to accept limited recursion.
1971 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
1974 CallStack.push_back(F);
1976 /// Values - As we compute SSA register values, we store their contents here.
1977 std::map<Value*, Constant*> Values;
1979 // Initialize arguments to the incoming values specified.
1981 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
1983 Values[AI] = ActualArgs[ArgNo];
1985 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
1986 /// we can only evaluate any one basic block at most once. This set keeps
1987 /// track of what we have executed so we can detect recursive cases etc.
1988 std::set<BasicBlock*> ExecutedBlocks;
1990 // CurInst - The current instruction we're evaluating.
1991 BasicBlock::iterator CurInst = F->begin()->begin();
1993 // This is the main evaluation loop.
1995 Constant *InstResult = 0;
1997 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
1998 if (SI->isVolatile()) return false; // no volatile accesses.
1999 Constant *Ptr = getVal(Values, SI->getOperand(1));
2000 if (!isSimpleEnoughPointerToCommit(Ptr))
2001 // If this is too complex for us to commit, reject it.
2003 Constant *Val = getVal(Values, SI->getOperand(0));
2004 MutatedMemory[Ptr] = Val;
2005 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2006 InstResult = ConstantExpr::get(BO->getOpcode(),
2007 getVal(Values, BO->getOperand(0)),
2008 getVal(Values, BO->getOperand(1)));
2009 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2010 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2011 getVal(Values, CI->getOperand(0)),
2012 getVal(Values, CI->getOperand(1)));
2013 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2014 InstResult = ConstantExpr::getCast(CI->getOpcode(),
2015 getVal(Values, CI->getOperand(0)),
2017 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2018 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
2019 getVal(Values, SI->getOperand(1)),
2020 getVal(Values, SI->getOperand(2)));
2021 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2022 Constant *P = getVal(Values, GEP->getOperand(0));
2023 SmallVector<Constant*, 8> GEPOps;
2024 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2026 GEPOps.push_back(getVal(Values, *i));
2027 InstResult = ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size());
2028 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2029 if (LI->isVolatile()) return false; // no volatile accesses.
2030 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
2032 if (InstResult == 0) return false; // Could not evaluate load.
2033 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2034 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs.
2035 const Type *Ty = AI->getType()->getElementType();
2036 AllocaTmps.push_back(new GlobalVariable(Ty, false,
2037 GlobalValue::InternalLinkage,
2038 UndefValue::get(Ty),
2040 InstResult = AllocaTmps.back();
2041 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
2042 // Cannot handle inline asm.
2043 if (isa<InlineAsm>(CI->getOperand(0))) return false;
2045 // Resolve function pointers.
2046 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0)));
2047 if (!Callee) return false; // Cannot resolve.
2049 std::vector<Constant*> Formals;
2050 for (User::op_iterator i = CI->op_begin() + 1, e = CI->op_end();
2052 Formals.push_back(getVal(Values, *i));
2054 if (Callee->isDeclaration()) {
2055 // If this is a function we can constant fold, do it.
2056 if (Constant *C = ConstantFoldCall(Callee, &Formals[0],
2063 if (Callee->getFunctionType()->isVarArg())
2068 // Execute the call, if successful, use the return value.
2069 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
2070 MutatedMemory, AllocaTmps))
2072 InstResult = RetVal;
2074 } else if (isa<TerminatorInst>(CurInst)) {
2075 BasicBlock *NewBB = 0;
2076 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2077 if (BI->isUnconditional()) {
2078 NewBB = BI->getSuccessor(0);
2081 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
2082 if (!Cond) return false; // Cannot determine.
2084 NewBB = BI->getSuccessor(!Cond->getZExtValue());
2086 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2088 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
2089 if (!Val) return false; // Cannot determine.
2090 NewBB = SI->getSuccessor(SI->findCaseValue(Val));
2091 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
2092 if (RI->getNumOperands())
2093 RetVal = getVal(Values, RI->getOperand(0));
2095 CallStack.pop_back(); // return from fn.
2096 return true; // We succeeded at evaluating this ctor!
2098 // invoke, unwind, unreachable.
2099 return false; // Cannot handle this terminator.
2102 // Okay, we succeeded in evaluating this control flow. See if we have
2103 // executed the new block before. If so, we have a looping function,
2104 // which we cannot evaluate in reasonable time.
2105 if (!ExecutedBlocks.insert(NewBB).second)
2106 return false; // looped!
2108 // Okay, we have never been in this block before. Check to see if there
2109 // are any PHI nodes. If so, evaluate them with information about where
2111 BasicBlock *OldBB = CurInst->getParent();
2112 CurInst = NewBB->begin();
2114 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2115 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
2117 // Do NOT increment CurInst. We know that the terminator had no value.
2120 // Did not know how to evaluate this!
2124 if (!CurInst->use_empty())
2125 Values[CurInst] = InstResult;
2127 // Advance program counter.
2132 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2133 /// we can. Return true if we can, false otherwise.
2134 static bool EvaluateStaticConstructor(Function *F) {
2135 /// MutatedMemory - For each store we execute, we update this map. Loads
2136 /// check this to get the most up-to-date value. If evaluation is successful,
2137 /// this state is committed to the process.
2138 std::map<Constant*, Constant*> MutatedMemory;
2140 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2141 /// to represent its body. This vector is needed so we can delete the
2142 /// temporary globals when we are done.
2143 std::vector<GlobalVariable*> AllocaTmps;
2145 /// CallStack - This is used to detect recursion. In pathological situations
2146 /// we could hit exponential behavior, but at least there is nothing
2148 std::vector<Function*> CallStack;
2150 // Call the function.
2151 Constant *RetValDummy;
2152 bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(),
2153 CallStack, MutatedMemory, AllocaTmps);
2155 // We succeeded at evaluation: commit the result.
2156 DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2157 << F->getName() << "' to " << MutatedMemory.size()
2159 for (std::map<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2160 E = MutatedMemory.end(); I != E; ++I)
2161 CommitValueTo(I->second, I->first);
2164 // At this point, we are done interpreting. If we created any 'alloca'
2165 // temporaries, release them now.
2166 while (!AllocaTmps.empty()) {
2167 GlobalVariable *Tmp = AllocaTmps.back();
2168 AllocaTmps.pop_back();
2170 // If there are still users of the alloca, the program is doing something
2171 // silly, e.g. storing the address of the alloca somewhere and using it
2172 // later. Since this is undefined, we'll just make it be null.
2173 if (!Tmp->use_empty())
2174 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2183 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2184 /// Return true if anything changed.
2185 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2186 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2187 bool MadeChange = false;
2188 if (Ctors.empty()) return false;
2190 // Loop over global ctors, optimizing them when we can.
2191 for (unsigned i = 0; i != Ctors.size(); ++i) {
2192 Function *F = Ctors[i];
2193 // Found a null terminator in the middle of the list, prune off the rest of
2196 if (i != Ctors.size()-1) {
2203 // We cannot simplify external ctor functions.
2204 if (F->empty()) continue;
2206 // If we can evaluate the ctor at compile time, do.
2207 if (EvaluateStaticConstructor(F)) {
2208 Ctors.erase(Ctors.begin()+i);
2211 ++NumCtorsEvaluated;
2216 if (!MadeChange) return false;
2218 GCL = InstallGlobalCtors(GCL, Ctors);
2222 bool GlobalOpt::ResolveAliases(Module &M) {
2223 bool Changed = false;
2225 for (Module::alias_iterator I = M.alias_begin(),
2226 E = M.alias_end(); I != E; ++I) {
2230 if (const GlobalValue *GV = I->resolveAliasedGlobal())
2232 I->replaceAllUsesWith(const_cast<GlobalValue*>(GV));
2240 bool GlobalOpt::runOnModule(Module &M) {
2241 bool Changed = false;
2243 // Try to find the llvm.globalctors list.
2244 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2246 bool LocalChange = true;
2247 while (LocalChange) {
2248 LocalChange = false;
2250 // Delete functions that are trivially dead, ccc -> fastcc
2251 LocalChange |= OptimizeFunctions(M);
2253 // Optimize global_ctors list.
2255 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2257 // Optimize non-address-taken globals.
2258 LocalChange |= OptimizeGlobalVars(M);
2260 // Resolve aliases, when possible.
2261 LocalChange |= ResolveAliases(M);
2262 Changed |= LocalChange;
2265 // TODO: Move all global ctors functions to the end of the module for code