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/ADT/SmallPtrSet.h"
32 #include "llvm/ADT/SmallVector.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/ADT/StringExtras.h"
39 STATISTIC(NumMarked , "Number of globals marked constant");
40 STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
41 STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
42 STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
43 STATISTIC(NumDeleted , "Number of globals deleted");
44 STATISTIC(NumFnDeleted , "Number of functions deleted");
45 STATISTIC(NumGlobUses , "Number of global uses devirtualized");
46 STATISTIC(NumLocalized , "Number of globals localized");
47 STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans");
48 STATISTIC(NumFastCallFns , "Number of functions converted to fastcc");
49 STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
50 STATISTIC(NumNestRemoved , "Number of nest attributes removed");
53 struct VISIBILITY_HIDDEN GlobalOpt : public ModulePass {
54 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
55 AU.addRequired<TargetData>();
57 static char ID; // Pass identification, replacement for typeid
58 GlobalOpt() : ModulePass((intptr_t)&ID) {}
60 bool runOnModule(Module &M);
63 GlobalVariable *FindGlobalCtors(Module &M);
64 bool OptimizeFunctions(Module &M);
65 bool OptimizeGlobalVars(Module &M);
66 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
67 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
70 char GlobalOpt::ID = 0;
71 RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer");
74 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
76 /// GlobalStatus - As we analyze each global, keep track of some information
77 /// about it. If we find out that the address of the global is taken, none of
78 /// this info will be accurate.
79 struct VISIBILITY_HIDDEN GlobalStatus {
80 /// isLoaded - True if the global is ever loaded. If the global isn't ever
81 /// loaded it can be deleted.
84 /// StoredType - Keep track of what stores to the global look like.
87 /// NotStored - There is no store to this global. It can thus be marked
91 /// isInitializerStored - This global is stored to, but the only thing
92 /// stored is the constant it was initialized with. This is only tracked
93 /// for scalar globals.
96 /// isStoredOnce - This global is stored to, but only its initializer and
97 /// one other value is ever stored to it. If this global isStoredOnce, we
98 /// track the value stored to it in StoredOnceValue below. This is only
99 /// tracked for scalar globals.
102 /// isStored - This global is stored to by multiple values or something else
103 /// that we cannot track.
107 /// StoredOnceValue - If only one value (besides the initializer constant) is
108 /// ever stored to this global, keep track of what value it is.
109 Value *StoredOnceValue;
111 /// AccessingFunction/HasMultipleAccessingFunctions - These start out
112 /// null/false. When the first accessing function is noticed, it is recorded.
113 /// When a second different accessing function is noticed,
114 /// HasMultipleAccessingFunctions is set to true.
115 Function *AccessingFunction;
116 bool HasMultipleAccessingFunctions;
118 /// HasNonInstructionUser - Set to true if this global has a user that is not
119 /// an instruction (e.g. a constant expr or GV initializer).
120 bool HasNonInstructionUser;
122 /// HasPHIUser - Set to true if this global has a user that is a PHI node.
125 GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0),
126 AccessingFunction(0), HasMultipleAccessingFunctions(false),
127 HasNonInstructionUser(false), HasPHIUser(false) {}
132 /// ConstantIsDead - Return true if the specified constant is (transitively)
133 /// dead. The constant may be used by other constants (e.g. constant arrays and
134 /// constant exprs) as long as they are dead, but it cannot be used by anything
136 static bool ConstantIsDead(Constant *C) {
137 if (isa<GlobalValue>(C)) return false;
139 for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI)
140 if (Constant *CU = dyn_cast<Constant>(*UI)) {
141 if (!ConstantIsDead(CU)) return false;
148 /// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
149 /// structure. If the global has its address taken, return true to indicate we
150 /// can't do anything with it.
152 static bool AnalyzeGlobal(Value *V, GlobalStatus &GS,
153 std::set<PHINode*> &PHIUsers) {
154 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
155 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
156 GS.HasNonInstructionUser = true;
158 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
160 } else if (Instruction *I = dyn_cast<Instruction>(*UI)) {
161 if (!GS.HasMultipleAccessingFunctions) {
162 Function *F = I->getParent()->getParent();
163 if (GS.AccessingFunction == 0)
164 GS.AccessingFunction = F;
165 else if (GS.AccessingFunction != F)
166 GS.HasMultipleAccessingFunctions = true;
168 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
170 if (LI->isVolatile()) return true; // Don't hack on volatile loads.
171 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
172 // Don't allow a store OF the address, only stores TO the address.
173 if (SI->getOperand(0) == V) return true;
175 if (SI->isVolatile()) return true; // Don't hack on volatile stores.
177 // If this is a direct store to the global (i.e., the global is a scalar
178 // value, not an aggregate), keep more specific information about
180 if (GS.StoredType != GlobalStatus::isStored) {
181 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){
182 Value *StoredVal = SI->getOperand(0);
183 if (StoredVal == GV->getInitializer()) {
184 if (GS.StoredType < GlobalStatus::isInitializerStored)
185 GS.StoredType = GlobalStatus::isInitializerStored;
186 } else if (isa<LoadInst>(StoredVal) &&
187 cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
189 if (GS.StoredType < GlobalStatus::isInitializerStored)
190 GS.StoredType = GlobalStatus::isInitializerStored;
191 } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
192 GS.StoredType = GlobalStatus::isStoredOnce;
193 GS.StoredOnceValue = StoredVal;
194 } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
195 GS.StoredOnceValue == StoredVal) {
198 GS.StoredType = GlobalStatus::isStored;
201 GS.StoredType = GlobalStatus::isStored;
204 } else if (isa<GetElementPtrInst>(I)) {
205 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
206 } else if (isa<SelectInst>(I)) {
207 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
208 } else if (PHINode *PN = dyn_cast<PHINode>(I)) {
209 // PHI nodes we can check just like select or GEP instructions, but we
210 // have to be careful about infinite recursion.
211 if (PHIUsers.insert(PN).second) // Not already visited.
212 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
213 GS.HasPHIUser = true;
214 } else if (isa<CmpInst>(I)) {
215 } else if (isa<MemCpyInst>(I) || isa<MemMoveInst>(I)) {
216 if (I->getOperand(1) == V)
217 GS.StoredType = GlobalStatus::isStored;
218 if (I->getOperand(2) == V)
220 } else if (isa<MemSetInst>(I)) {
221 assert(I->getOperand(1) == V && "Memset only takes one pointer!");
222 GS.StoredType = GlobalStatus::isStored;
224 return true; // Any other non-load instruction might take address!
226 } else if (Constant *C = dyn_cast<Constant>(*UI)) {
227 GS.HasNonInstructionUser = true;
228 // We might have a dead and dangling constant hanging off of here.
229 if (!ConstantIsDead(C))
232 GS.HasNonInstructionUser = true;
233 // Otherwise must be some other user.
240 static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) {
241 ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
243 unsigned IdxV = CI->getZExtValue();
245 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
246 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
247 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
248 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
249 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) {
250 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
251 } else if (isa<ConstantAggregateZero>(Agg)) {
252 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
253 if (IdxV < STy->getNumElements())
254 return Constant::getNullValue(STy->getElementType(IdxV));
255 } else if (const SequentialType *STy =
256 dyn_cast<SequentialType>(Agg->getType())) {
257 return Constant::getNullValue(STy->getElementType());
259 } else if (isa<UndefValue>(Agg)) {
260 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
261 if (IdxV < STy->getNumElements())
262 return UndefValue::get(STy->getElementType(IdxV));
263 } else if (const SequentialType *STy =
264 dyn_cast<SequentialType>(Agg->getType())) {
265 return UndefValue::get(STy->getElementType());
272 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
273 /// users of the global, cleaning up the obvious ones. This is largely just a
274 /// quick scan over the use list to clean up the easy and obvious cruft. This
275 /// returns true if it made a change.
276 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
277 bool Changed = false;
278 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
281 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
283 // Replace the load with the initializer.
284 LI->replaceAllUsesWith(Init);
285 LI->eraseFromParent();
288 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
289 // Store must be unreachable or storing Init into the global.
290 SI->eraseFromParent();
292 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
293 if (CE->getOpcode() == Instruction::GetElementPtr) {
294 Constant *SubInit = 0;
296 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
297 Changed |= CleanupConstantGlobalUsers(CE, SubInit);
298 } else if (CE->getOpcode() == Instruction::BitCast &&
299 isa<PointerType>(CE->getType())) {
300 // Pointer cast, delete any stores and memsets to the global.
301 Changed |= CleanupConstantGlobalUsers(CE, 0);
304 if (CE->use_empty()) {
305 CE->destroyConstant();
308 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
309 // Do not transform "gepinst (gep constexpr (GV))" here, because forming
310 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
311 // and will invalidate our notion of what Init is.
312 Constant *SubInit = 0;
313 if (!isa<ConstantExpr>(GEP->getOperand(0))) {
315 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
316 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
317 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
319 Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
321 if (GEP->use_empty()) {
322 GEP->eraseFromParent();
325 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
326 if (MI->getRawDest() == V) {
327 MI->eraseFromParent();
331 } else if (Constant *C = dyn_cast<Constant>(U)) {
332 // If we have a chain of dead constantexprs or other things dangling from
333 // us, and if they are all dead, nuke them without remorse.
334 if (ConstantIsDead(C)) {
335 C->destroyConstant();
336 // This could have invalidated UI, start over from scratch.
337 CleanupConstantGlobalUsers(V, Init);
345 /// isSafeSROAElementUse - Return true if the specified instruction is a safe
346 /// user of a derived expression from a global that we want to SROA.
347 static bool isSafeSROAElementUse(Value *V) {
348 // We might have a dead and dangling constant hanging off of here.
349 if (Constant *C = dyn_cast<Constant>(V))
350 return ConstantIsDead(C);
352 Instruction *I = dyn_cast<Instruction>(V);
353 if (!I) return false;
356 if (isa<LoadInst>(I)) return true;
358 // Stores *to* the pointer are ok.
359 if (StoreInst *SI = dyn_cast<StoreInst>(I))
360 return SI->getOperand(0) != V;
362 // Otherwise, it must be a GEP.
363 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I);
364 if (GEPI == 0) return false;
366 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
367 !cast<Constant>(GEPI->getOperand(1))->isNullValue())
370 for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end();
372 if (!isSafeSROAElementUse(*I))
378 /// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value.
379 /// Look at it and its uses and decide whether it is safe to SROA this global.
381 static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
382 // The user of the global must be a GEP Inst or a ConstantExpr GEP.
383 if (!isa<GetElementPtrInst>(U) &&
384 (!isa<ConstantExpr>(U) ||
385 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
388 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
389 // don't like < 3 operand CE's, and we don't like non-constant integer
390 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some
392 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
393 !cast<Constant>(U->getOperand(1))->isNullValue() ||
394 !isa<ConstantInt>(U->getOperand(2)))
397 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
398 ++GEPI; // Skip over the pointer index.
400 // If this is a use of an array allocation, do a bit more checking for sanity.
401 if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
402 uint64_t NumElements = AT->getNumElements();
403 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
405 // Check to make sure that index falls within the array. If not,
406 // something funny is going on, so we won't do the optimization.
408 if (Idx->getZExtValue() >= NumElements)
411 // We cannot scalar repl this level of the array unless any array
412 // sub-indices are in-range constants. In particular, consider:
413 // A[0][i]. We cannot know that the user isn't doing invalid things like
414 // allowing i to index an out-of-range subscript that accesses A[1].
416 // Scalar replacing *just* the outer index of the array is probably not
417 // going to be a win anyway, so just give up.
418 for (++GEPI; // Skip array index.
419 GEPI != E && (isa<ArrayType>(*GEPI) || isa<VectorType>(*GEPI));
421 uint64_t NumElements;
422 if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
423 NumElements = SubArrayTy->getNumElements();
425 NumElements = cast<VectorType>(*GEPI)->getNumElements();
427 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
428 if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
433 for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I)
434 if (!isSafeSROAElementUse(*I))
439 /// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
440 /// is safe for us to perform this transformation.
442 static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
443 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
445 if (!IsUserOfGlobalSafeForSRA(*UI, GV))
452 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
453 /// variable. This opens the door for other optimizations by exposing the
454 /// behavior of the program in a more fine-grained way. We have determined that
455 /// this transformation is safe already. We return the first global variable we
456 /// insert so that the caller can reprocess it.
457 static GlobalVariable *SRAGlobal(GlobalVariable *GV) {
458 // Make sure this global only has simple uses that we can SRA.
459 if (!GlobalUsersSafeToSRA(GV))
462 assert(GV->hasInternalLinkage() && !GV->isConstant());
463 Constant *Init = GV->getInitializer();
464 const Type *Ty = Init->getType();
466 std::vector<GlobalVariable*> NewGlobals;
467 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
469 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
470 NewGlobals.reserve(STy->getNumElements());
471 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
472 Constant *In = getAggregateConstantElement(Init,
473 ConstantInt::get(Type::Int32Ty, i));
474 assert(In && "Couldn't get element of initializer?");
475 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
476 GlobalVariable::InternalLinkage,
477 In, GV->getName()+"."+utostr(i),
479 GV->isThreadLocal());
480 Globals.insert(GV, NGV);
481 NewGlobals.push_back(NGV);
483 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
484 unsigned NumElements = 0;
485 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
486 NumElements = ATy->getNumElements();
487 else if (const VectorType *PTy = dyn_cast<VectorType>(STy))
488 NumElements = PTy->getNumElements();
490 assert(0 && "Unknown aggregate sequential type!");
492 if (NumElements > 16 && GV->hasNUsesOrMore(16))
493 return 0; // It's not worth it.
494 NewGlobals.reserve(NumElements);
495 for (unsigned i = 0, e = NumElements; i != e; ++i) {
496 Constant *In = getAggregateConstantElement(Init,
497 ConstantInt::get(Type::Int32Ty, i));
498 assert(In && "Couldn't get element of initializer?");
500 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
501 GlobalVariable::InternalLinkage,
502 In, GV->getName()+"."+utostr(i),
504 GV->isThreadLocal());
505 Globals.insert(GV, NGV);
506 NewGlobals.push_back(NGV);
510 if (NewGlobals.empty())
513 DOUT << "PERFORMING GLOBAL SRA ON: " << *GV;
515 Constant *NullInt = Constant::getNullValue(Type::Int32Ty);
517 // Loop over all of the uses of the global, replacing the constantexpr geps,
518 // with smaller constantexpr geps or direct references.
519 while (!GV->use_empty()) {
520 User *GEP = GV->use_back();
521 assert(((isa<ConstantExpr>(GEP) &&
522 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
523 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
525 // Ignore the 1th operand, which has to be zero or else the program is quite
526 // broken (undefined). Get the 2nd operand, which is the structure or array
528 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
529 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
531 Value *NewPtr = NewGlobals[Val];
533 // Form a shorter GEP if needed.
534 if (GEP->getNumOperands() > 3) {
535 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
536 SmallVector<Constant*, 8> Idxs;
537 Idxs.push_back(NullInt);
538 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
539 Idxs.push_back(CE->getOperand(i));
540 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr),
541 &Idxs[0], Idxs.size());
543 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
544 SmallVector<Value*, 8> Idxs;
545 Idxs.push_back(NullInt);
546 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
547 Idxs.push_back(GEPI->getOperand(i));
548 NewPtr = new GetElementPtrInst(NewPtr, Idxs.begin(), Idxs.end(),
549 GEPI->getName()+"."+utostr(Val), GEPI);
552 GEP->replaceAllUsesWith(NewPtr);
554 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
555 GEPI->eraseFromParent();
557 cast<ConstantExpr>(GEP)->destroyConstant();
560 // Delete the old global, now that it is dead.
564 // Loop over the new globals array deleting any globals that are obviously
565 // dead. This can arise due to scalarization of a structure or an array that
566 // has elements that are dead.
567 unsigned FirstGlobal = 0;
568 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
569 if (NewGlobals[i]->use_empty()) {
570 Globals.erase(NewGlobals[i]);
571 if (FirstGlobal == i) ++FirstGlobal;
574 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
577 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
578 /// value will trap if the value is dynamically null. PHIs keeps track of any
579 /// phi nodes we've seen to avoid reprocessing them.
580 static bool AllUsesOfValueWillTrapIfNull(Value *V,
581 SmallPtrSet<PHINode*, 8> &PHIs) {
582 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
583 if (isa<LoadInst>(*UI)) {
585 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
586 if (SI->getOperand(0) == V) {
587 //cerr << "NONTRAPPING USE: " << **UI;
588 return false; // Storing the value.
590 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
591 if (CI->getOperand(0) != V) {
592 //cerr << "NONTRAPPING USE: " << **UI;
593 return false; // Not calling the ptr
595 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
596 if (II->getOperand(0) != V) {
597 //cerr << "NONTRAPPING USE: " << **UI;
598 return false; // Not calling the ptr
600 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(*UI)) {
601 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
602 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
603 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
604 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
605 // If we've already seen this phi node, ignore it, it has already been
608 return AllUsesOfValueWillTrapIfNull(PN, PHIs);
609 } else if (isa<ICmpInst>(*UI) &&
610 isa<ConstantPointerNull>(UI->getOperand(1))) {
611 // Ignore setcc X, null
613 //cerr << "NONTRAPPING USE: " << **UI;
619 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
620 /// from GV will trap if the loaded value is null. Note that this also permits
621 /// comparisons of the loaded value against null, as a special case.
622 static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) {
623 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI)
624 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
625 SmallPtrSet<PHINode*, 8> PHIs;
626 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
628 } else if (isa<StoreInst>(*UI)) {
629 // Ignore stores to the global.
631 // We don't know or understand this user, bail out.
632 //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI;
639 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
640 bool Changed = false;
641 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
642 Instruction *I = cast<Instruction>(*UI++);
643 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
644 LI->setOperand(0, NewV);
646 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
647 if (SI->getOperand(1) == V) {
648 SI->setOperand(1, NewV);
651 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
652 if (I->getOperand(0) == V) {
653 // Calling through the pointer! Turn into a direct call, but be careful
654 // that the pointer is not also being passed as an argument.
655 I->setOperand(0, NewV);
657 bool PassedAsArg = false;
658 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
659 if (I->getOperand(i) == V) {
661 I->setOperand(i, NewV);
665 // Being passed as an argument also. Be careful to not invalidate UI!
669 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
670 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
671 ConstantExpr::getCast(CI->getOpcode(),
672 NewV, CI->getType()));
673 if (CI->use_empty()) {
675 CI->eraseFromParent();
677 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
678 // Should handle GEP here.
679 SmallVector<Constant*, 8> Idxs;
680 Idxs.reserve(GEPI->getNumOperands()-1);
681 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
682 if (Constant *C = dyn_cast<Constant>(GEPI->getOperand(i)))
686 if (Idxs.size() == GEPI->getNumOperands()-1)
687 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
688 ConstantExpr::getGetElementPtr(NewV, &Idxs[0],
690 if (GEPI->use_empty()) {
692 GEPI->eraseFromParent();
701 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
702 /// value stored into it. If there are uses of the loaded value that would trap
703 /// if the loaded value is dynamically null, then we know that they cannot be
704 /// reachable with a null optimize away the load.
705 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) {
706 std::vector<LoadInst*> Loads;
707 bool Changed = false;
709 // Replace all uses of loads with uses of uses of the stored value.
710 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end();
712 if (LoadInst *LI = dyn_cast<LoadInst>(*GUI)) {
714 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
716 // If we get here we could have stores, selects, or phi nodes whose values
718 assert((isa<StoreInst>(*GUI) || isa<PHINode>(*GUI) ||
719 isa<SelectInst>(*GUI) || isa<ConstantExpr>(*GUI)) &&
720 "Only expect load and stores!");
724 DOUT << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV;
728 // Delete all of the loads we can, keeping track of whether we nuked them all!
729 bool AllLoadsGone = true;
730 while (!Loads.empty()) {
731 LoadInst *L = Loads.back();
732 if (L->use_empty()) {
733 L->eraseFromParent();
736 AllLoadsGone = false;
741 // If we nuked all of the loads, then none of the stores are needed either,
742 // nor is the global.
744 DOUT << " *** GLOBAL NOW DEAD!\n";
745 CleanupConstantGlobalUsers(GV, 0);
746 if (GV->use_empty()) {
747 GV->eraseFromParent();
755 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
756 /// instructions that are foldable.
757 static void ConstantPropUsersOf(Value *V) {
758 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
759 if (Instruction *I = dyn_cast<Instruction>(*UI++))
760 if (Constant *NewC = ConstantFoldInstruction(I)) {
761 I->replaceAllUsesWith(NewC);
763 // Advance UI to the next non-I use to avoid invalidating it!
764 // Instructions could multiply use V.
765 while (UI != E && *UI == I)
767 I->eraseFromParent();
771 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
772 /// variable, and transforms the program as if it always contained the result of
773 /// the specified malloc. Because it is always the result of the specified
774 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
775 /// malloc into a global, and any loads of GV as uses of the new global.
776 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
778 DOUT << "PROMOTING MALLOC GLOBAL: " << *GV << " MALLOC = " << *MI;
779 ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize());
781 if (NElements->getZExtValue() != 1) {
782 // If we have an array allocation, transform it to a single element
783 // allocation to make the code below simpler.
784 Type *NewTy = ArrayType::get(MI->getAllocatedType(),
785 NElements->getZExtValue());
787 new MallocInst(NewTy, Constant::getNullValue(Type::Int32Ty),
788 MI->getAlignment(), MI->getName(), MI);
790 Indices[0] = Indices[1] = Constant::getNullValue(Type::Int32Ty);
791 Value *NewGEP = new GetElementPtrInst(NewMI, Indices, Indices + 2,
792 NewMI->getName()+".el0", MI);
793 MI->replaceAllUsesWith(NewGEP);
794 MI->eraseFromParent();
798 // Create the new global variable. The contents of the malloc'd memory is
799 // undefined, so initialize with an undef value.
800 Constant *Init = UndefValue::get(MI->getAllocatedType());
801 GlobalVariable *NewGV = new GlobalVariable(MI->getAllocatedType(), false,
802 GlobalValue::InternalLinkage, Init,
803 GV->getName()+".body",
805 GV->isThreadLocal());
806 GV->getParent()->getGlobalList().insert(GV, NewGV);
808 // Anything that used the malloc now uses the global directly.
809 MI->replaceAllUsesWith(NewGV);
811 Constant *RepValue = NewGV;
812 if (NewGV->getType() != GV->getType()->getElementType())
813 RepValue = ConstantExpr::getBitCast(RepValue,
814 GV->getType()->getElementType());
816 // If there is a comparison against null, we will insert a global bool to
817 // keep track of whether the global was initialized yet or not.
818 GlobalVariable *InitBool =
819 new GlobalVariable(Type::Int1Ty, false, GlobalValue::InternalLinkage,
820 ConstantInt::getFalse(), GV->getName()+".init",
821 (Module *)NULL, GV->isThreadLocal());
822 bool InitBoolUsed = false;
824 // Loop over all uses of GV, processing them in turn.
825 std::vector<StoreInst*> Stores;
826 while (!GV->use_empty())
827 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
828 while (!LI->use_empty()) {
829 Use &LoadUse = LI->use_begin().getUse();
830 if (!isa<ICmpInst>(LoadUse.getUser()))
833 ICmpInst *CI = cast<ICmpInst>(LoadUse.getUser());
834 // Replace the cmp X, 0 with a use of the bool value.
835 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", CI);
837 switch (CI->getPredicate()) {
838 default: assert(0 && "Unknown ICmp Predicate!");
839 case ICmpInst::ICMP_ULT:
840 case ICmpInst::ICMP_SLT:
841 LV = ConstantInt::getFalse(); // X < null -> always false
843 case ICmpInst::ICMP_ULE:
844 case ICmpInst::ICMP_SLE:
845 case ICmpInst::ICMP_EQ:
846 LV = BinaryOperator::createNot(LV, "notinit", CI);
848 case ICmpInst::ICMP_NE:
849 case ICmpInst::ICMP_UGE:
850 case ICmpInst::ICMP_SGE:
851 case ICmpInst::ICMP_UGT:
852 case ICmpInst::ICMP_SGT:
855 CI->replaceAllUsesWith(LV);
856 CI->eraseFromParent();
859 LI->eraseFromParent();
861 StoreInst *SI = cast<StoreInst>(GV->use_back());
862 // The global is initialized when the store to it occurs.
863 new StoreInst(ConstantInt::getTrue(), InitBool, SI);
864 SI->eraseFromParent();
867 // If the initialization boolean was used, insert it, otherwise delete it.
869 while (!InitBool->use_empty()) // Delete initializations
870 cast<Instruction>(InitBool->use_back())->eraseFromParent();
873 GV->getParent()->getGlobalList().insert(GV, InitBool);
876 // Now the GV is dead, nuke it and the malloc.
877 GV->eraseFromParent();
878 MI->eraseFromParent();
880 // To further other optimizations, loop over all users of NewGV and try to
881 // constant prop them. This will promote GEP instructions with constant
882 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
883 ConstantPropUsersOf(NewGV);
884 if (RepValue != NewGV)
885 ConstantPropUsersOf(RepValue);
890 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
891 /// to make sure that there are no complex uses of V. We permit simple things
892 /// like dereferencing the pointer, but not storing through the address, unless
893 /// it is to the specified global.
894 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V,
896 SmallPtrSet<PHINode*, 8> &PHIs) {
897 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
898 if (isa<LoadInst>(*UI) || isa<CmpInst>(*UI)) {
900 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
901 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
902 return false; // Storing the pointer itself... bad.
903 // Otherwise, storing through it, or storing into GV... fine.
904 } else if (isa<GetElementPtrInst>(*UI)) {
905 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(cast<Instruction>(*UI),
908 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
909 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI
912 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
920 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
921 /// somewhere. Transform all uses of the allocation into loads from the
922 /// global and uses of the resultant pointer. Further, delete the store into
923 /// GV. This assumes that these value pass the
924 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
925 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
926 GlobalVariable *GV) {
927 while (!Alloc->use_empty()) {
928 Instruction *U = cast<Instruction>(*Alloc->use_begin());
929 Instruction *InsertPt = U;
930 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
931 // If this is the store of the allocation into the global, remove it.
932 if (SI->getOperand(1) == GV) {
933 SI->eraseFromParent();
936 } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
937 // Insert the load in the corresponding predecessor, not right before the
939 unsigned PredNo = Alloc->use_begin().getOperandNo()/2;
940 InsertPt = PN->getIncomingBlock(PredNo)->getTerminator();
943 // Insert a load from the global, and use it instead of the malloc.
944 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
945 U->replaceUsesOfWith(Alloc, NL);
949 /// GlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
950 /// GV are simple enough to perform HeapSRA, return true.
951 static bool GlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV,
953 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;
955 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
956 // We permit two users of the load: setcc comparing against the null
957 // pointer, and a getelementptr of a specific form.
958 for (Value::use_iterator UI = LI->use_begin(), E = LI->use_end(); UI != E;
960 // Comparison against null is ok.
961 if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) {
962 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
967 // getelementptr is also ok, but only a simple form.
968 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
969 // Must index into the array and into the struct.
970 if (GEPI->getNumOperands() < 3)
973 // Otherwise the GEP is ok.
977 if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
978 // We have a phi of a load from the global. We can only handle this
979 // if the other PHI'd values are actually the same. In this case,
980 // the rewriter will just drop the phi entirely.
981 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
982 Value *IV = PN->getIncomingValue(i);
983 if (IV == LI) continue; // Trivial the same.
985 // If the phi'd value is from the malloc that initializes the value,
987 if (IV == MI) continue;
989 // Otherwise, we don't know what it is.
995 // Otherwise we don't know what this is, not ok.
1002 /// GetHeapSROALoad - Return the load for the specified field of the HeapSROA'd
1003 /// value, lazily creating it on demand.
1004 static Value *GetHeapSROALoad(Instruction *Load, unsigned FieldNo,
1005 const std::vector<GlobalVariable*> &FieldGlobals,
1006 std::vector<Value *> &InsertedLoadsForPtr) {
1007 if (InsertedLoadsForPtr.size() <= FieldNo)
1008 InsertedLoadsForPtr.resize(FieldNo+1);
1009 if (InsertedLoadsForPtr[FieldNo] == 0)
1010 InsertedLoadsForPtr[FieldNo] = new LoadInst(FieldGlobals[FieldNo],
1011 Load->getName()+".f" +
1012 utostr(FieldNo), Load);
1013 return InsertedLoadsForPtr[FieldNo];
1016 /// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1017 /// the load, rewrite the derived value to use the HeapSRoA'd load.
1018 static void RewriteHeapSROALoadUser(LoadInst *Load, Instruction *LoadUser,
1019 const std::vector<GlobalVariable*> &FieldGlobals,
1020 std::vector<Value *> &InsertedLoadsForPtr) {
1021 // If this is a comparison against null, handle it.
1022 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1023 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1024 // If we have a setcc of the loaded pointer, we can use a setcc of any
1027 if (InsertedLoadsForPtr.empty()) {
1028 NPtr = GetHeapSROALoad(Load, 0, FieldGlobals, InsertedLoadsForPtr);
1030 NPtr = InsertedLoadsForPtr.back();
1033 Value *New = new ICmpInst(SCI->getPredicate(), NPtr,
1034 Constant::getNullValue(NPtr->getType()),
1035 SCI->getName(), SCI);
1036 SCI->replaceAllUsesWith(New);
1037 SCI->eraseFromParent();
1041 // Handle 'getelementptr Ptr, Idx, uint FieldNo ...'
1042 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1043 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1044 && "Unexpected GEPI!");
1046 // Load the pointer for this field.
1047 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1048 Value *NewPtr = GetHeapSROALoad(Load, FieldNo,
1049 FieldGlobals, InsertedLoadsForPtr);
1051 // Create the new GEP idx vector.
1052 SmallVector<Value*, 8> GEPIdx;
1053 GEPIdx.push_back(GEPI->getOperand(1));
1054 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1056 Value *NGEPI = new GetElementPtrInst(NewPtr, GEPIdx.begin(), GEPIdx.end(),
1057 GEPI->getName(), GEPI);
1058 GEPI->replaceAllUsesWith(NGEPI);
1059 GEPI->eraseFromParent();
1063 // Handle PHI nodes. PHI nodes must be merging in the same values, plus
1064 // potentially the original malloc. Insert phi nodes for each field, then
1065 // process uses of the PHI.
1066 PHINode *PN = cast<PHINode>(LoadUser);
1067 std::vector<Value *> PHIsForField;
1068 PHIsForField.resize(FieldGlobals.size());
1069 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1070 Value *LoadV = GetHeapSROALoad(Load, i, FieldGlobals, InsertedLoadsForPtr);
1072 PHINode *FieldPN = new PHINode(LoadV->getType(),
1073 PN->getName()+"."+utostr(i), PN);
1074 // Fill in the predecessor values.
1075 for (unsigned pred = 0, e = PN->getNumIncomingValues(); pred != e; ++pred) {
1076 // Each predecessor either uses the load or the original malloc.
1077 Value *InVal = PN->getIncomingValue(pred);
1078 BasicBlock *BB = PN->getIncomingBlock(pred);
1080 if (isa<MallocInst>(InVal)) {
1081 // Insert a reload from the global in the predecessor.
1082 NewVal = GetHeapSROALoad(BB->getTerminator(), i, FieldGlobals,
1085 NewVal = InsertedLoadsForPtr[i];
1087 FieldPN->addIncoming(NewVal, BB);
1089 PHIsForField[i] = FieldPN;
1092 // Since PHIsForField specifies a phi for every input value, the lazy inserter
1093 // will never insert a load.
1094 while (!PN->use_empty())
1095 RewriteHeapSROALoadUser(Load, PN->use_back(), FieldGlobals, PHIsForField);
1096 PN->eraseFromParent();
1099 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1100 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1101 /// use FieldGlobals instead. All uses of loaded values satisfy
1102 /// GlobalLoadUsesSimpleEnoughForHeapSRA.
1103 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1104 const std::vector<GlobalVariable*> &FieldGlobals) {
1105 std::vector<Value *> InsertedLoadsForPtr;
1106 //InsertedLoadsForPtr.resize(FieldGlobals.size());
1107 while (!Load->use_empty())
1108 RewriteHeapSROALoadUser(Load, Load->use_back(),
1109 FieldGlobals, InsertedLoadsForPtr);
1112 /// PerformHeapAllocSRoA - MI is an allocation of an array of structures. Break
1113 /// it up into multiple allocations of arrays of the fields.
1114 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){
1115 DOUT << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *MI;
1116 const StructType *STy = cast<StructType>(MI->getAllocatedType());
1118 // There is guaranteed to be at least one use of the malloc (storing
1119 // it into GV). If there are other uses, change them to be uses of
1120 // the global to simplify later code. This also deletes the store
1122 ReplaceUsesOfMallocWithGlobal(MI, GV);
1124 // Okay, at this point, there are no users of the malloc. Insert N
1125 // new mallocs at the same place as MI, and N globals.
1126 std::vector<GlobalVariable*> FieldGlobals;
1127 std::vector<MallocInst*> FieldMallocs;
1129 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1130 const Type *FieldTy = STy->getElementType(FieldNo);
1131 const Type *PFieldTy = PointerType::getUnqual(FieldTy);
1133 GlobalVariable *NGV =
1134 new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage,
1135 Constant::getNullValue(PFieldTy),
1136 GV->getName() + ".f" + utostr(FieldNo), GV,
1137 GV->isThreadLocal());
1138 FieldGlobals.push_back(NGV);
1140 MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(),
1141 MI->getName() + ".f" + utostr(FieldNo),MI);
1142 FieldMallocs.push_back(NMI);
1143 new StoreInst(NMI, NGV, MI);
1146 // The tricky aspect of this transformation is handling the case when malloc
1147 // fails. In the original code, malloc failing would set the result pointer
1148 // of malloc to null. In this case, some mallocs could succeed and others
1149 // could fail. As such, we emit code that looks like this:
1150 // F0 = malloc(field0)
1151 // F1 = malloc(field1)
1152 // F2 = malloc(field2)
1153 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1154 // if (F0) { free(F0); F0 = 0; }
1155 // if (F1) { free(F1); F1 = 0; }
1156 // if (F2) { free(F2); F2 = 0; }
1158 Value *RunningOr = 0;
1159 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1160 Value *Cond = new ICmpInst(ICmpInst::ICMP_EQ, FieldMallocs[i],
1161 Constant::getNullValue(FieldMallocs[i]->getType()),
1164 RunningOr = Cond; // First seteq
1166 RunningOr = BinaryOperator::createOr(RunningOr, Cond, "tmp", MI);
1169 // Split the basic block at the old malloc.
1170 BasicBlock *OrigBB = MI->getParent();
1171 BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont");
1173 // Create the block to check the first condition. Put all these blocks at the
1174 // end of the function as they are unlikely to be executed.
1175 BasicBlock *NullPtrBlock = new BasicBlock("malloc_ret_null",
1176 OrigBB->getParent());
1178 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1179 // branch on RunningOr.
1180 OrigBB->getTerminator()->eraseFromParent();
1181 new BranchInst(NullPtrBlock, ContBB, RunningOr, OrigBB);
1183 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1184 // pointer, because some may be null while others are not.
1185 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1186 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1187 Value *Cmp = new ICmpInst(ICmpInst::ICMP_NE, GVVal,
1188 Constant::getNullValue(GVVal->getType()),
1189 "tmp", NullPtrBlock);
1190 BasicBlock *FreeBlock = new BasicBlock("free_it", OrigBB->getParent());
1191 BasicBlock *NextBlock = new BasicBlock("next", OrigBB->getParent());
1192 new BranchInst(FreeBlock, NextBlock, Cmp, NullPtrBlock);
1194 // Fill in FreeBlock.
1195 new FreeInst(GVVal, FreeBlock);
1196 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1198 new BranchInst(NextBlock, FreeBlock);
1200 NullPtrBlock = NextBlock;
1203 new BranchInst(ContBB, NullPtrBlock);
1206 // MI is no longer needed, remove it.
1207 MI->eraseFromParent();
1210 // Okay, the malloc site is completely handled. All of the uses of GV are now
1211 // loads, and all uses of those loads are simple. Rewrite them to use loads
1212 // of the per-field globals instead.
1213 while (!GV->use_empty()) {
1214 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
1215 RewriteUsesOfLoadForHeapSRoA(LI, FieldGlobals);
1216 LI->eraseFromParent();
1218 // Must be a store of null.
1219 StoreInst *SI = cast<StoreInst>(GV->use_back());
1220 assert(isa<Constant>(SI->getOperand(0)) &&
1221 cast<Constant>(SI->getOperand(0))->isNullValue() &&
1222 "Unexpected heap-sra user!");
1224 // Insert a store of null into each global.
1225 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1227 Constant::getNullValue(FieldGlobals[i]->getType()->getElementType());
1228 new StoreInst(Null, FieldGlobals[i], SI);
1230 // Erase the original store.
1231 SI->eraseFromParent();
1235 // The old global is now dead, remove it.
1236 GV->eraseFromParent();
1239 return FieldGlobals[0];
1243 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1244 // that only one value (besides its initializer) is ever stored to the global.
1245 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1246 Module::global_iterator &GVI,
1248 if (CastInst *CI = dyn_cast<CastInst>(StoredOnceVal))
1249 StoredOnceVal = CI->getOperand(0);
1250 else if (GetElementPtrInst *GEPI =dyn_cast<GetElementPtrInst>(StoredOnceVal)){
1251 // "getelementptr Ptr, 0, 0, 0" is really just a cast.
1252 bool IsJustACast = true;
1253 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
1254 if (!isa<Constant>(GEPI->getOperand(i)) ||
1255 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
1256 IsJustACast = false;
1260 StoredOnceVal = GEPI->getOperand(0);
1263 // If we are dealing with a pointer global that is initialized to null and
1264 // only has one (non-null) value stored into it, then we can optimize any
1265 // users of the loaded value (often calls and loads) that would trap if the
1267 if (isa<PointerType>(GV->getInitializer()->getType()) &&
1268 GV->getInitializer()->isNullValue()) {
1269 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1270 if (GV->getInitializer()->getType() != SOVC->getType())
1271 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1273 // Optimize away any trapping uses of the loaded value.
1274 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1276 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) {
1277 // If this is a malloc of an abstract type, don't touch it.
1278 if (!MI->getAllocatedType()->isSized())
1281 // We can't optimize this global unless all uses of it are *known* to be
1282 // of the malloc value, not of the null initializer value (consider a use
1283 // that compares the global's value against zero to see if the malloc has
1284 // been reached). To do this, we check to see if all uses of the global
1285 // would trap if the global were null: this proves that they must all
1286 // happen after the malloc.
1287 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1290 // We can't optimize this if the malloc itself is used in a complex way,
1291 // for example, being stored into multiple globals. This allows the
1292 // malloc to be stored into the specified global, loaded setcc'd, and
1293 // GEP'd. These are all things we could transform to using the global
1296 SmallPtrSet<PHINode*, 8> PHIs;
1297 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV, PHIs))
1302 // If we have a global that is only initialized with a fixed size malloc,
1303 // transform the program to use global memory instead of malloc'd memory.
1304 // This eliminates dynamic allocation, avoids an indirection accessing the
1305 // data, and exposes the resultant global to further GlobalOpt.
1306 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) {
1307 // Restrict this transformation to only working on small allocations
1308 // (2048 bytes currently), as we don't want to introduce a 16M global or
1310 if (NElements->getZExtValue()*
1311 TD.getABITypeSize(MI->getAllocatedType()) < 2048) {
1312 GVI = OptimizeGlobalAddressOfMalloc(GV, MI);
1317 // If the allocation is an array of structures, consider transforming this
1318 // into multiple malloc'd arrays, one for each field. This is basically
1319 // SRoA for malloc'd memory.
1320 if (const StructType *AllocTy =
1321 dyn_cast<StructType>(MI->getAllocatedType())) {
1322 // This the structure has an unreasonable number of fields, leave it
1324 if (AllocTy->getNumElements() <= 16 && AllocTy->getNumElements() > 0 &&
1325 GlobalLoadUsesSimpleEnoughForHeapSRA(GV, MI)) {
1326 GVI = PerformHeapAllocSRoA(GV, MI);
1336 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1337 /// two values ever stored into GV are its initializer and OtherVal. See if we
1338 /// can shrink the global into a boolean and select between the two values
1339 /// whenever it is used. This exposes the values to other scalar optimizations.
1340 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1341 const Type *GVElType = GV->getType()->getElementType();
1343 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1344 // an FP value or vector, don't do this optimization because a select between
1345 // them is very expensive and unlikely to lead to later simplification.
1346 if (GVElType == Type::Int1Ty || GVElType->isFloatingPoint() ||
1347 isa<VectorType>(GVElType))
1350 // Walk the use list of the global seeing if all the uses are load or store.
1351 // If there is anything else, bail out.
1352 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I)
1353 if (!isa<LoadInst>(I) && !isa<StoreInst>(I))
1356 DOUT << " *** SHRINKING TO BOOL: " << *GV;
1358 // Create the new global, initializing it to false.
1359 GlobalVariable *NewGV = new GlobalVariable(Type::Int1Ty, false,
1360 GlobalValue::InternalLinkage, ConstantInt::getFalse(),
1363 GV->isThreadLocal());
1364 GV->getParent()->getGlobalList().insert(GV, NewGV);
1366 Constant *InitVal = GV->getInitializer();
1367 assert(InitVal->getType() != Type::Int1Ty && "No reason to shrink to bool!");
1369 // If initialized to zero and storing one into the global, we can use a cast
1370 // instead of a select to synthesize the desired value.
1371 bool IsOneZero = false;
1372 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1373 IsOneZero = InitVal->isNullValue() && CI->isOne();
1375 while (!GV->use_empty()) {
1376 Instruction *UI = cast<Instruction>(GV->use_back());
1377 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1378 // Change the store into a boolean store.
1379 bool StoringOther = SI->getOperand(0) == OtherVal;
1380 // Only do this if we weren't storing a loaded value.
1382 if (StoringOther || SI->getOperand(0) == InitVal)
1383 StoreVal = ConstantInt::get(Type::Int1Ty, StoringOther);
1385 // Otherwise, we are storing a previously loaded copy. To do this,
1386 // change the copy from copying the original value to just copying the
1388 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1390 // If we're already replaced the input, StoredVal will be a cast or
1391 // select instruction. If not, it will be a load of the original
1393 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1394 assert(LI->getOperand(0) == GV && "Not a copy!");
1395 // Insert a new load, to preserve the saved value.
1396 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1398 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1399 "This is not a form that we understand!");
1400 StoreVal = StoredVal->getOperand(0);
1401 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1404 new StoreInst(StoreVal, NewGV, SI);
1406 // Change the load into a load of bool then a select.
1407 LoadInst *LI = cast<LoadInst>(UI);
1408 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1411 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1413 NSI = new SelectInst(NLI, OtherVal, InitVal, "", LI);
1415 LI->replaceAllUsesWith(NSI);
1417 UI->eraseFromParent();
1420 GV->eraseFromParent();
1425 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1426 /// it if possible. If we make a change, return true.
1427 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1428 Module::global_iterator &GVI) {
1429 std::set<PHINode*> PHIUsers;
1431 GV->removeDeadConstantUsers();
1433 if (GV->use_empty()) {
1434 DOUT << "GLOBAL DEAD: " << *GV;
1435 GV->eraseFromParent();
1440 if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
1442 cerr << "Global: " << *GV;
1443 cerr << " isLoaded = " << GS.isLoaded << "\n";
1444 cerr << " StoredType = ";
1445 switch (GS.StoredType) {
1446 case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break;
1447 case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break;
1448 case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break;
1449 case GlobalStatus::isStored: cerr << "stored\n"; break;
1451 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
1452 cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n";
1453 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
1454 cerr << " AccessingFunction = " << GS.AccessingFunction->getName()
1456 cerr << " HasMultipleAccessingFunctions = "
1457 << GS.HasMultipleAccessingFunctions << "\n";
1458 cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n";
1462 // If this is a first class global and has only one accessing function
1463 // and this function is main (which we know is not recursive we can make
1464 // this global a local variable) we replace the global with a local alloca
1465 // in this function.
1467 // NOTE: It doesn't make sense to promote non first class types since we
1468 // are just replacing static memory to stack memory.
1469 if (!GS.HasMultipleAccessingFunctions &&
1470 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1471 GV->getType()->getElementType()->isFirstClassType() &&
1472 GS.AccessingFunction->getName() == "main" &&
1473 GS.AccessingFunction->hasExternalLinkage()) {
1474 DOUT << "LOCALIZING GLOBAL: " << *GV;
1475 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin();
1476 const Type* ElemTy = GV->getType()->getElementType();
1477 // FIXME: Pass Global's alignment when globals have alignment
1478 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI);
1479 if (!isa<UndefValue>(GV->getInitializer()))
1480 new StoreInst(GV->getInitializer(), Alloca, FirstI);
1482 GV->replaceAllUsesWith(Alloca);
1483 GV->eraseFromParent();
1488 // If the global is never loaded (but may be stored to), it is dead.
1491 DOUT << "GLOBAL NEVER LOADED: " << *GV;
1493 // Delete any stores we can find to the global. We may not be able to
1494 // make it completely dead though.
1495 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1497 // If the global is dead now, delete it.
1498 if (GV->use_empty()) {
1499 GV->eraseFromParent();
1505 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1506 DOUT << "MARKING CONSTANT: " << *GV;
1507 GV->setConstant(true);
1509 // Clean up any obviously simplifiable users now.
1510 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1512 // If the global is dead now, just nuke it.
1513 if (GV->use_empty()) {
1514 DOUT << " *** Marking constant allowed us to simplify "
1515 << "all users and delete global!\n";
1516 GV->eraseFromParent();
1522 } else if (!GV->getInitializer()->getType()->isFirstClassType()) {
1523 if (GlobalVariable *FirstNewGV = SRAGlobal(GV)) {
1524 GVI = FirstNewGV; // Don't skip the newly produced globals!
1527 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1528 // If the initial value for the global was an undef value, and if only
1529 // one other value was stored into it, we can just change the
1530 // initializer to be an undef value, then delete all stores to the
1531 // global. This allows us to mark it constant.
1532 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1533 if (isa<UndefValue>(GV->getInitializer())) {
1534 // Change the initial value here.
1535 GV->setInitializer(SOVConstant);
1537 // Clean up any obviously simplifiable users now.
1538 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1540 if (GV->use_empty()) {
1541 DOUT << " *** Substituting initializer allowed us to "
1542 << "simplify all users and delete global!\n";
1543 GV->eraseFromParent();
1552 // Try to optimize globals based on the knowledge that only one value
1553 // (besides its initializer) is ever stored to the global.
1554 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1555 getAnalysis<TargetData>()))
1558 // Otherwise, if the global was not a boolean, we can shrink it to be a
1560 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1561 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1570 /// OnlyCalledDirectly - Return true if the specified function is only called
1571 /// directly. In other words, its address is never taken.
1572 static bool OnlyCalledDirectly(Function *F) {
1573 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1574 Instruction *User = dyn_cast<Instruction>(*UI);
1575 if (!User) return false;
1576 if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false;
1578 // See if the function address is passed as an argument.
1579 for (unsigned i = 1, e = User->getNumOperands(); i != e; ++i)
1580 if (User->getOperand(i) == F) return false;
1585 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1586 /// function, changing them to FastCC.
1587 static void ChangeCalleesToFastCall(Function *F) {
1588 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1589 CallSite User(cast<Instruction>(*UI));
1590 User.setCallingConv(CallingConv::Fast);
1594 static PAListPtr StripNest(const PAListPtr &Attrs) {
1595 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1596 if ((Attrs.getSlot(i).Attrs & ParamAttr::Nest) == 0)
1599 // There can be only one.
1600 return Attrs.removeAttr(Attrs.getSlot(i).Index, ParamAttr::Nest);
1606 static void RemoveNestAttribute(Function *F) {
1607 F->setParamAttrs(StripNest(F->getParamAttrs()));
1608 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1609 CallSite User(cast<Instruction>(*UI));
1610 User.setParamAttrs(StripNest(User.getParamAttrs()));
1614 bool GlobalOpt::OptimizeFunctions(Module &M) {
1615 bool Changed = false;
1616 // Optimize functions.
1617 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1619 F->removeDeadConstantUsers();
1620 if (F->use_empty() && (F->hasInternalLinkage() ||
1621 F->hasLinkOnceLinkage())) {
1622 M.getFunctionList().erase(F);
1625 } else if (F->hasInternalLinkage()) {
1626 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1627 OnlyCalledDirectly(F)) {
1628 // If this function has C calling conventions, is not a varargs
1629 // function, and is only called directly, promote it to use the Fast
1630 // calling convention.
1631 F->setCallingConv(CallingConv::Fast);
1632 ChangeCalleesToFastCall(F);
1637 if (F->getParamAttrs().hasAttrSomewhere(ParamAttr::Nest) &&
1638 OnlyCalledDirectly(F)) {
1639 // The function is not used by a trampoline intrinsic, so it is safe
1640 // to remove the 'nest' attribute.
1641 RemoveNestAttribute(F);
1650 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1651 bool Changed = false;
1652 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1654 GlobalVariable *GV = GVI++;
1655 if (!GV->isConstant() && GV->hasInternalLinkage() &&
1656 GV->hasInitializer())
1657 Changed |= ProcessInternalGlobal(GV, GVI);
1662 /// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1663 /// initializers have an init priority of 65535.
1664 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1665 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1667 if (I->getName() == "llvm.global_ctors") {
1668 // Found it, verify it's an array of { int, void()* }.
1669 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1671 const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1672 if (!STy || STy->getNumElements() != 2 ||
1673 STy->getElementType(0) != Type::Int32Ty) return 0;
1674 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1675 if (!PFTy) return 0;
1676 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1677 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() ||
1678 FTy->getNumParams() != 0)
1681 // Verify that the initializer is simple enough for us to handle.
1682 if (!I->hasInitializer()) return 0;
1683 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1685 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1686 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(CA->getOperand(i))) {
1687 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1690 // Must have a function or null ptr.
1691 if (!isa<Function>(CS->getOperand(1)))
1694 // Init priority must be standard.
1695 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1696 if (!CI || CI->getZExtValue() != 65535)
1707 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1708 /// return a list of the functions and null terminator as a vector.
1709 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1710 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1711 std::vector<Function*> Result;
1712 Result.reserve(CA->getNumOperands());
1713 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) {
1714 ConstantStruct *CS = cast<ConstantStruct>(CA->getOperand(i));
1715 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1720 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1721 /// specified array, returning the new global to use.
1722 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1723 const std::vector<Function*> &Ctors) {
1724 // If we made a change, reassemble the initializer list.
1725 std::vector<Constant*> CSVals;
1726 CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535));
1727 CSVals.push_back(0);
1729 // Create the new init list.
1730 std::vector<Constant*> CAList;
1731 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1733 CSVals[1] = Ctors[i];
1735 const Type *FTy = FunctionType::get(Type::VoidTy,
1736 std::vector<const Type*>(), false);
1737 const PointerType *PFTy = PointerType::getUnqual(FTy);
1738 CSVals[1] = Constant::getNullValue(PFTy);
1739 CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647);
1741 CAList.push_back(ConstantStruct::get(CSVals));
1744 // Create the array initializer.
1745 const Type *StructTy =
1746 cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
1747 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()),
1750 // If we didn't change the number of elements, don't create a new GV.
1751 if (CA->getType() == GCL->getInitializer()->getType()) {
1752 GCL->setInitializer(CA);
1756 // Create the new global and insert it next to the existing list.
1757 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
1758 GCL->getLinkage(), CA, "",
1760 GCL->isThreadLocal());
1761 GCL->getParent()->getGlobalList().insert(GCL, NGV);
1764 // Nuke the old list, replacing any uses with the new one.
1765 if (!GCL->use_empty()) {
1767 if (V->getType() != GCL->getType())
1768 V = ConstantExpr::getBitCast(V, GCL->getType());
1769 GCL->replaceAllUsesWith(V);
1771 GCL->eraseFromParent();
1780 static Constant *getVal(std::map<Value*, Constant*> &ComputedValues,
1782 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
1783 Constant *R = ComputedValues[V];
1784 assert(R && "Reference to an uncomputed value!");
1788 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
1789 /// enough for us to understand. In particular, if it is a cast of something,
1790 /// we punt. We basically just support direct accesses to globals and GEP's of
1791 /// globals. This should be kept up to date with CommitValueTo.
1792 static bool isSimpleEnoughPointerToCommit(Constant *C) {
1793 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
1794 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1795 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1796 return !GV->isDeclaration(); // reject external globals.
1798 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
1799 // Handle a constantexpr gep.
1800 if (CE->getOpcode() == Instruction::GetElementPtr &&
1801 isa<GlobalVariable>(CE->getOperand(0))) {
1802 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1803 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1804 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1805 return GV->hasInitializer() &&
1806 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1811 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
1812 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
1813 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
1814 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
1815 ConstantExpr *Addr, unsigned OpNo) {
1816 // Base case of the recursion.
1817 if (OpNo == Addr->getNumOperands()) {
1818 assert(Val->getType() == Init->getType() && "Type mismatch!");
1822 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
1823 std::vector<Constant*> Elts;
1825 // Break up the constant into its elements.
1826 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
1827 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i)
1828 Elts.push_back(CS->getOperand(i));
1829 } else if (isa<ConstantAggregateZero>(Init)) {
1830 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1831 Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
1832 } else if (isa<UndefValue>(Init)) {
1833 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1834 Elts.push_back(UndefValue::get(STy->getElementType(i)));
1836 assert(0 && "This code is out of sync with "
1837 " ConstantFoldLoadThroughGEPConstantExpr");
1840 // Replace the element that we are supposed to.
1841 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
1842 unsigned Idx = CU->getZExtValue();
1843 assert(Idx < STy->getNumElements() && "Struct index out of range!");
1844 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
1846 // Return the modified struct.
1847 return ConstantStruct::get(&Elts[0], Elts.size(), STy->isPacked());
1849 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
1850 const ArrayType *ATy = cast<ArrayType>(Init->getType());
1852 // Break up the array into elements.
1853 std::vector<Constant*> Elts;
1854 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
1855 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1856 Elts.push_back(CA->getOperand(i));
1857 } else if (isa<ConstantAggregateZero>(Init)) {
1858 Constant *Elt = Constant::getNullValue(ATy->getElementType());
1859 Elts.assign(ATy->getNumElements(), Elt);
1860 } else if (isa<UndefValue>(Init)) {
1861 Constant *Elt = UndefValue::get(ATy->getElementType());
1862 Elts.assign(ATy->getNumElements(), Elt);
1864 assert(0 && "This code is out of sync with "
1865 " ConstantFoldLoadThroughGEPConstantExpr");
1868 assert(CI->getZExtValue() < ATy->getNumElements());
1869 Elts[CI->getZExtValue()] =
1870 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
1871 return ConstantArray::get(ATy, Elts);
1875 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
1876 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
1877 static void CommitValueTo(Constant *Val, Constant *Addr) {
1878 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
1879 assert(GV->hasInitializer());
1880 GV->setInitializer(Val);
1884 ConstantExpr *CE = cast<ConstantExpr>(Addr);
1885 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1887 Constant *Init = GV->getInitializer();
1888 Init = EvaluateStoreInto(Init, Val, CE, 2);
1889 GV->setInitializer(Init);
1892 /// ComputeLoadResult - Return the value that would be computed by a load from
1893 /// P after the stores reflected by 'memory' have been performed. If we can't
1894 /// decide, return null.
1895 static Constant *ComputeLoadResult(Constant *P,
1896 const std::map<Constant*, Constant*> &Memory) {
1897 // If this memory location has been recently stored, use the stored value: it
1898 // is the most up-to-date.
1899 std::map<Constant*, Constant*>::const_iterator I = Memory.find(P);
1900 if (I != Memory.end()) return I->second;
1903 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
1904 if (GV->hasInitializer())
1905 return GV->getInitializer();
1909 // Handle a constantexpr getelementptr.
1910 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
1911 if (CE->getOpcode() == Instruction::GetElementPtr &&
1912 isa<GlobalVariable>(CE->getOperand(0))) {
1913 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1914 if (GV->hasInitializer())
1915 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1918 return 0; // don't know how to evaluate.
1921 /// EvaluateFunction - Evaluate a call to function F, returning true if
1922 /// successful, false if we can't evaluate it. ActualArgs contains the formal
1923 /// arguments for the function.
1924 static bool EvaluateFunction(Function *F, Constant *&RetVal,
1925 const std::vector<Constant*> &ActualArgs,
1926 std::vector<Function*> &CallStack,
1927 std::map<Constant*, Constant*> &MutatedMemory,
1928 std::vector<GlobalVariable*> &AllocaTmps) {
1929 // Check to see if this function is already executing (recursion). If so,
1930 // bail out. TODO: we might want to accept limited recursion.
1931 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
1934 CallStack.push_back(F);
1936 /// Values - As we compute SSA register values, we store their contents here.
1937 std::map<Value*, Constant*> Values;
1939 // Initialize arguments to the incoming values specified.
1941 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
1943 Values[AI] = ActualArgs[ArgNo];
1945 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
1946 /// we can only evaluate any one basic block at most once. This set keeps
1947 /// track of what we have executed so we can detect recursive cases etc.
1948 std::set<BasicBlock*> ExecutedBlocks;
1950 // CurInst - The current instruction we're evaluating.
1951 BasicBlock::iterator CurInst = F->begin()->begin();
1953 // This is the main evaluation loop.
1955 Constant *InstResult = 0;
1957 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
1958 if (SI->isVolatile()) return false; // no volatile accesses.
1959 Constant *Ptr = getVal(Values, SI->getOperand(1));
1960 if (!isSimpleEnoughPointerToCommit(Ptr))
1961 // If this is too complex for us to commit, reject it.
1963 Constant *Val = getVal(Values, SI->getOperand(0));
1964 MutatedMemory[Ptr] = Val;
1965 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
1966 InstResult = ConstantExpr::get(BO->getOpcode(),
1967 getVal(Values, BO->getOperand(0)),
1968 getVal(Values, BO->getOperand(1)));
1969 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
1970 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
1971 getVal(Values, CI->getOperand(0)),
1972 getVal(Values, CI->getOperand(1)));
1973 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
1974 InstResult = ConstantExpr::getCast(CI->getOpcode(),
1975 getVal(Values, CI->getOperand(0)),
1977 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
1978 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
1979 getVal(Values, SI->getOperand(1)),
1980 getVal(Values, SI->getOperand(2)));
1981 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
1982 Constant *P = getVal(Values, GEP->getOperand(0));
1983 SmallVector<Constant*, 8> GEPOps;
1984 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
1985 GEPOps.push_back(getVal(Values, GEP->getOperand(i)));
1986 InstResult = ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size());
1987 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
1988 if (LI->isVolatile()) return false; // no volatile accesses.
1989 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
1991 if (InstResult == 0) return false; // Could not evaluate load.
1992 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
1993 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs.
1994 const Type *Ty = AI->getType()->getElementType();
1995 AllocaTmps.push_back(new GlobalVariable(Ty, false,
1996 GlobalValue::InternalLinkage,
1997 UndefValue::get(Ty),
1999 InstResult = AllocaTmps.back();
2000 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
2001 // Cannot handle inline asm.
2002 if (isa<InlineAsm>(CI->getOperand(0))) return false;
2004 // Resolve function pointers.
2005 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0)));
2006 if (!Callee) return false; // Cannot resolve.
2008 std::vector<Constant*> Formals;
2009 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
2010 Formals.push_back(getVal(Values, CI->getOperand(i)));
2012 if (Callee->isDeclaration()) {
2013 // If this is a function we can constant fold, do it.
2014 if (Constant *C = ConstantFoldCall(Callee, &Formals[0],
2021 if (Callee->getFunctionType()->isVarArg())
2026 // Execute the call, if successful, use the return value.
2027 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
2028 MutatedMemory, AllocaTmps))
2030 InstResult = RetVal;
2032 } else if (isa<TerminatorInst>(CurInst)) {
2033 BasicBlock *NewBB = 0;
2034 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2035 if (BI->isUnconditional()) {
2036 NewBB = BI->getSuccessor(0);
2039 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
2040 if (!Cond) return false; // Cannot determine.
2042 NewBB = BI->getSuccessor(!Cond->getZExtValue());
2044 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2046 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
2047 if (!Val) return false; // Cannot determine.
2048 NewBB = SI->getSuccessor(SI->findCaseValue(Val));
2049 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
2050 if (RI->getNumOperands())
2051 RetVal = getVal(Values, RI->getOperand(0));
2053 CallStack.pop_back(); // return from fn.
2054 return true; // We succeeded at evaluating this ctor!
2056 // invoke, unwind, unreachable.
2057 return false; // Cannot handle this terminator.
2060 // Okay, we succeeded in evaluating this control flow. See if we have
2061 // executed the new block before. If so, we have a looping function,
2062 // which we cannot evaluate in reasonable time.
2063 if (!ExecutedBlocks.insert(NewBB).second)
2064 return false; // looped!
2066 // Okay, we have never been in this block before. Check to see if there
2067 // are any PHI nodes. If so, evaluate them with information about where
2069 BasicBlock *OldBB = CurInst->getParent();
2070 CurInst = NewBB->begin();
2072 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2073 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
2075 // Do NOT increment CurInst. We know that the terminator had no value.
2078 // Did not know how to evaluate this!
2082 if (!CurInst->use_empty())
2083 Values[CurInst] = InstResult;
2085 // Advance program counter.
2090 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2091 /// we can. Return true if we can, false otherwise.
2092 static bool EvaluateStaticConstructor(Function *F) {
2093 /// MutatedMemory - For each store we execute, we update this map. Loads
2094 /// check this to get the most up-to-date value. If evaluation is successful,
2095 /// this state is committed to the process.
2096 std::map<Constant*, Constant*> MutatedMemory;
2098 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2099 /// to represent its body. This vector is needed so we can delete the
2100 /// temporary globals when we are done.
2101 std::vector<GlobalVariable*> AllocaTmps;
2103 /// CallStack - This is used to detect recursion. In pathological situations
2104 /// we could hit exponential behavior, but at least there is nothing
2106 std::vector<Function*> CallStack;
2108 // Call the function.
2109 Constant *RetValDummy;
2110 bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(),
2111 CallStack, MutatedMemory, AllocaTmps);
2113 // We succeeded at evaluation: commit the result.
2114 DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2115 << F->getName() << "' to " << MutatedMemory.size()
2117 for (std::map<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2118 E = MutatedMemory.end(); I != E; ++I)
2119 CommitValueTo(I->second, I->first);
2122 // At this point, we are done interpreting. If we created any 'alloca'
2123 // temporaries, release them now.
2124 while (!AllocaTmps.empty()) {
2125 GlobalVariable *Tmp = AllocaTmps.back();
2126 AllocaTmps.pop_back();
2128 // If there are still users of the alloca, the program is doing something
2129 // silly, e.g. storing the address of the alloca somewhere and using it
2130 // later. Since this is undefined, we'll just make it be null.
2131 if (!Tmp->use_empty())
2132 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2141 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2142 /// Return true if anything changed.
2143 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2144 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2145 bool MadeChange = false;
2146 if (Ctors.empty()) return false;
2148 // Loop over global ctors, optimizing them when we can.
2149 for (unsigned i = 0; i != Ctors.size(); ++i) {
2150 Function *F = Ctors[i];
2151 // Found a null terminator in the middle of the list, prune off the rest of
2154 if (i != Ctors.size()-1) {
2161 // We cannot simplify external ctor functions.
2162 if (F->empty()) continue;
2164 // If we can evaluate the ctor at compile time, do.
2165 if (EvaluateStaticConstructor(F)) {
2166 Ctors.erase(Ctors.begin()+i);
2169 ++NumCtorsEvaluated;
2174 if (!MadeChange) return false;
2176 GCL = InstallGlobalCtors(GCL, Ctors);
2181 bool GlobalOpt::runOnModule(Module &M) {
2182 bool Changed = false;
2184 // Try to find the llvm.globalctors list.
2185 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2187 bool LocalChange = true;
2188 while (LocalChange) {
2189 LocalChange = false;
2191 // Delete functions that are trivially dead, ccc -> fastcc
2192 LocalChange |= OptimizeFunctions(M);
2194 // Optimize global_ctors list.
2196 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2198 // Optimize non-address-taken globals.
2199 LocalChange |= OptimizeGlobalVars(M);
2200 Changed |= LocalChange;
2203 // TODO: Move all global ctors functions to the end of the module for code