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/LLVMContext.h"
24 #include "llvm/Module.h"
25 #include "llvm/Pass.h"
26 #include "llvm/Analysis/ConstantFolding.h"
27 #include "llvm/Analysis/MemoryBuiltins.h"
28 #include "llvm/Target/TargetData.h"
29 #include "llvm/Support/CallSite.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/GetElementPtrTypeIterator.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/ADT/DenseMap.h"
36 #include "llvm/ADT/SmallPtrSet.h"
37 #include "llvm/ADT/SmallVector.h"
38 #include "llvm/ADT/Statistic.h"
39 #include "llvm/ADT/STLExtras.h"
43 STATISTIC(NumMarked , "Number of globals marked constant");
44 STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
45 STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
46 STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
47 STATISTIC(NumDeleted , "Number of globals deleted");
48 STATISTIC(NumFnDeleted , "Number of functions deleted");
49 STATISTIC(NumGlobUses , "Number of global uses devirtualized");
50 STATISTIC(NumLocalized , "Number of globals localized");
51 STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans");
52 STATISTIC(NumFastCallFns , "Number of functions converted to fastcc");
53 STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
54 STATISTIC(NumNestRemoved , "Number of nest attributes removed");
55 STATISTIC(NumAliasesResolved, "Number of global aliases resolved");
56 STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated");
59 struct GlobalOpt : public ModulePass {
60 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
62 static char ID; // Pass identification, replacement for typeid
63 GlobalOpt() : ModulePass(&ID) {}
65 bool runOnModule(Module &M);
68 GlobalVariable *FindGlobalCtors(Module &M);
69 bool OptimizeFunctions(Module &M);
70 bool OptimizeGlobalVars(Module &M);
71 bool OptimizeGlobalAliases(Module &M);
72 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
73 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
77 char GlobalOpt::ID = 0;
78 static RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer");
80 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
84 /// GlobalStatus - As we analyze each global, keep track of some information
85 /// about it. If we find out that the address of the global is taken, none of
86 /// this info will be accurate.
88 /// isLoaded - True if the global is ever loaded. If the global isn't ever
89 /// loaded it can be deleted.
92 /// StoredType - Keep track of what stores to the global look like.
95 /// NotStored - There is no store to this global. It can thus be marked
99 /// isInitializerStored - This global is stored to, but the only thing
100 /// stored is the constant it was initialized with. This is only tracked
101 /// for scalar globals.
104 /// isStoredOnce - This global is stored to, but only its initializer and
105 /// one other value is ever stored to it. If this global isStoredOnce, we
106 /// track the value stored to it in StoredOnceValue below. This is only
107 /// tracked for scalar globals.
110 /// isStored - This global is stored to by multiple values or something else
111 /// that we cannot track.
115 /// StoredOnceValue - If only one value (besides the initializer constant) is
116 /// ever stored to this global, keep track of what value it is.
117 Value *StoredOnceValue;
119 /// AccessingFunction/HasMultipleAccessingFunctions - These start out
120 /// null/false. When the first accessing function is noticed, it is recorded.
121 /// When a second different accessing function is noticed,
122 /// HasMultipleAccessingFunctions is set to true.
123 Function *AccessingFunction;
124 bool HasMultipleAccessingFunctions;
126 /// HasNonInstructionUser - Set to true if this global has a user that is not
127 /// an instruction (e.g. a constant expr or GV initializer).
128 bool HasNonInstructionUser;
130 /// HasPHIUser - Set to true if this global has a user that is a PHI node.
133 GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0),
134 AccessingFunction(0), HasMultipleAccessingFunctions(false),
135 HasNonInstructionUser(false), HasPHIUser(false) {}
140 // SafeToDestroyConstant - It is safe to destroy a constant iff it is only used
141 // by constants itself. Note that constants cannot be cyclic, so this test is
142 // pretty easy to implement recursively.
144 static bool SafeToDestroyConstant(Constant *C) {
145 if (isa<GlobalValue>(C)) return false;
147 for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI)
148 if (Constant *CU = dyn_cast<Constant>(*UI)) {
149 if (!SafeToDestroyConstant(CU)) return false;
156 /// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
157 /// structure. If the global has its address taken, return true to indicate we
158 /// can't do anything with it.
160 static bool AnalyzeGlobal(Value *V, GlobalStatus &GS,
161 SmallPtrSet<PHINode*, 16> &PHIUsers) {
162 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
163 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
164 GS.HasNonInstructionUser = true;
166 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
168 } else if (Instruction *I = dyn_cast<Instruction>(*UI)) {
169 if (!GS.HasMultipleAccessingFunctions) {
170 Function *F = I->getParent()->getParent();
171 if (GS.AccessingFunction == 0)
172 GS.AccessingFunction = F;
173 else if (GS.AccessingFunction != F)
174 GS.HasMultipleAccessingFunctions = true;
176 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
178 if (LI->isVolatile()) return true; // Don't hack on volatile loads.
179 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
180 // Don't allow a store OF the address, only stores TO the address.
181 if (SI->getOperand(0) == V) return true;
183 if (SI->isVolatile()) return true; // Don't hack on volatile stores.
185 // If this is a direct store to the global (i.e., the global is a scalar
186 // value, not an aggregate), keep more specific information about
188 if (GS.StoredType != GlobalStatus::isStored) {
189 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){
190 Value *StoredVal = SI->getOperand(0);
191 if (StoredVal == GV->getInitializer()) {
192 if (GS.StoredType < GlobalStatus::isInitializerStored)
193 GS.StoredType = GlobalStatus::isInitializerStored;
194 } else if (isa<LoadInst>(StoredVal) &&
195 cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
197 if (GS.StoredType < GlobalStatus::isInitializerStored)
198 GS.StoredType = GlobalStatus::isInitializerStored;
199 } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
200 GS.StoredType = GlobalStatus::isStoredOnce;
201 GS.StoredOnceValue = StoredVal;
202 } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
203 GS.StoredOnceValue == StoredVal) {
206 GS.StoredType = GlobalStatus::isStored;
209 GS.StoredType = GlobalStatus::isStored;
212 } else if (isa<GetElementPtrInst>(I)) {
213 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
214 } else if (isa<SelectInst>(I)) {
215 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
216 } else if (PHINode *PN = dyn_cast<PHINode>(I)) {
217 // PHI nodes we can check just like select or GEP instructions, but we
218 // have to be careful about infinite recursion.
219 if (PHIUsers.insert(PN)) // Not already visited.
220 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
221 GS.HasPHIUser = true;
222 } else if (isa<CmpInst>(I)) {
223 } else if (isa<MemTransferInst>(I)) {
224 if (I->getOperand(1) == V)
225 GS.StoredType = GlobalStatus::isStored;
226 if (I->getOperand(2) == V)
228 } else if (isa<MemSetInst>(I)) {
229 assert(I->getOperand(1) == V && "Memset only takes one pointer!");
230 GS.StoredType = GlobalStatus::isStored;
232 return true; // Any other non-load instruction might take address!
234 } else if (Constant *C = dyn_cast<Constant>(*UI)) {
235 GS.HasNonInstructionUser = true;
236 // We might have a dead and dangling constant hanging off of here.
237 if (!SafeToDestroyConstant(C))
240 GS.HasNonInstructionUser = true;
241 // Otherwise must be some other user.
248 static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx,
249 LLVMContext &Context) {
250 ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
252 unsigned IdxV = CI->getZExtValue();
254 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
255 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
256 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
257 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
258 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) {
259 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
260 } else if (isa<ConstantAggregateZero>(Agg)) {
261 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
262 if (IdxV < STy->getNumElements())
263 return Constant::getNullValue(STy->getElementType(IdxV));
264 } else if (const SequentialType *STy =
265 dyn_cast<SequentialType>(Agg->getType())) {
266 return Constant::getNullValue(STy->getElementType());
268 } else if (isa<UndefValue>(Agg)) {
269 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
270 if (IdxV < STy->getNumElements())
271 return UndefValue::get(STy->getElementType(IdxV));
272 } else if (const SequentialType *STy =
273 dyn_cast<SequentialType>(Agg->getType())) {
274 return UndefValue::get(STy->getElementType());
281 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
282 /// users of the global, cleaning up the obvious ones. This is largely just a
283 /// quick scan over the use list to clean up the easy and obvious cruft. This
284 /// returns true if it made a change.
285 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init,
286 LLVMContext &Context) {
287 bool Changed = false;
288 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
291 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
293 // Replace the load with the initializer.
294 LI->replaceAllUsesWith(Init);
295 LI->eraseFromParent();
298 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
299 // Store must be unreachable or storing Init into the global.
300 SI->eraseFromParent();
302 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
303 if (CE->getOpcode() == Instruction::GetElementPtr) {
304 Constant *SubInit = 0;
306 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
307 Changed |= CleanupConstantGlobalUsers(CE, SubInit, Context);
308 } else if (CE->getOpcode() == Instruction::BitCast &&
309 isa<PointerType>(CE->getType())) {
310 // Pointer cast, delete any stores and memsets to the global.
311 Changed |= CleanupConstantGlobalUsers(CE, 0, Context);
314 if (CE->use_empty()) {
315 CE->destroyConstant();
318 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
319 // Do not transform "gepinst (gep constexpr (GV))" here, because forming
320 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
321 // and will invalidate our notion of what Init is.
322 Constant *SubInit = 0;
323 if (!isa<ConstantExpr>(GEP->getOperand(0))) {
325 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP, Context));
326 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
327 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
329 Changed |= CleanupConstantGlobalUsers(GEP, SubInit, Context);
331 if (GEP->use_empty()) {
332 GEP->eraseFromParent();
335 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
336 if (MI->getRawDest() == V) {
337 MI->eraseFromParent();
341 } else if (Constant *C = dyn_cast<Constant>(U)) {
342 // If we have a chain of dead constantexprs or other things dangling from
343 // us, and if they are all dead, nuke them without remorse.
344 if (SafeToDestroyConstant(C)) {
345 C->destroyConstant();
346 // This could have invalidated UI, start over from scratch.
347 CleanupConstantGlobalUsers(V, Init, Context);
355 /// isSafeSROAElementUse - Return true if the specified instruction is a safe
356 /// user of a derived expression from a global that we want to SROA.
357 static bool isSafeSROAElementUse(Value *V) {
358 // We might have a dead and dangling constant hanging off of here.
359 if (Constant *C = dyn_cast<Constant>(V))
360 return SafeToDestroyConstant(C);
362 Instruction *I = dyn_cast<Instruction>(V);
363 if (!I) return false;
366 if (isa<LoadInst>(I)) return true;
368 // Stores *to* the pointer are ok.
369 if (StoreInst *SI = dyn_cast<StoreInst>(I))
370 return SI->getOperand(0) != V;
372 // Otherwise, it must be a GEP.
373 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I);
374 if (GEPI == 0) return false;
376 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
377 !cast<Constant>(GEPI->getOperand(1))->isNullValue())
380 for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end();
382 if (!isSafeSROAElementUse(*I))
388 /// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value.
389 /// Look at it and its uses and decide whether it is safe to SROA this global.
391 static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
392 // The user of the global must be a GEP Inst or a ConstantExpr GEP.
393 if (!isa<GetElementPtrInst>(U) &&
394 (!isa<ConstantExpr>(U) ||
395 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
398 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
399 // don't like < 3 operand CE's, and we don't like non-constant integer
400 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some
402 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
403 !cast<Constant>(U->getOperand(1))->isNullValue() ||
404 !isa<ConstantInt>(U->getOperand(2)))
407 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
408 ++GEPI; // Skip over the pointer index.
410 // If this is a use of an array allocation, do a bit more checking for sanity.
411 if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
412 uint64_t NumElements = AT->getNumElements();
413 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
415 // Check to make sure that index falls within the array. If not,
416 // something funny is going on, so we won't do the optimization.
418 if (Idx->getZExtValue() >= NumElements)
421 // We cannot scalar repl this level of the array unless any array
422 // sub-indices are in-range constants. In particular, consider:
423 // A[0][i]. We cannot know that the user isn't doing invalid things like
424 // allowing i to index an out-of-range subscript that accesses A[1].
426 // Scalar replacing *just* the outer index of the array is probably not
427 // going to be a win anyway, so just give up.
428 for (++GEPI; // Skip array index.
431 uint64_t NumElements;
432 if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
433 NumElements = SubArrayTy->getNumElements();
434 else if (const VectorType *SubVectorTy = dyn_cast<VectorType>(*GEPI))
435 NumElements = SubVectorTy->getNumElements();
437 assert(isa<StructType>(*GEPI) &&
438 "Indexed GEP type is not array, vector, or struct!");
442 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
443 if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
448 for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I)
449 if (!isSafeSROAElementUse(*I))
454 /// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
455 /// is safe for us to perform this transformation.
457 static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
458 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
460 if (!IsUserOfGlobalSafeForSRA(*UI, GV))
467 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
468 /// variable. This opens the door for other optimizations by exposing the
469 /// behavior of the program in a more fine-grained way. We have determined that
470 /// this transformation is safe already. We return the first global variable we
471 /// insert so that the caller can reprocess it.
472 static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD,
473 LLVMContext &Context) {
474 // Make sure this global only has simple uses that we can SRA.
475 if (!GlobalUsersSafeToSRA(GV))
478 assert(GV->hasLocalLinkage() && !GV->isConstant());
479 Constant *Init = GV->getInitializer();
480 const Type *Ty = Init->getType();
482 std::vector<GlobalVariable*> NewGlobals;
483 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
485 // Get the alignment of the global, either explicit or target-specific.
486 unsigned StartAlignment = GV->getAlignment();
487 if (StartAlignment == 0)
488 StartAlignment = TD.getABITypeAlignment(GV->getType());
490 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
491 NewGlobals.reserve(STy->getNumElements());
492 const StructLayout &Layout = *TD.getStructLayout(STy);
493 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
494 Constant *In = getAggregateConstantElement(Init,
495 ConstantInt::get(Type::getInt32Ty(Context), i),
497 assert(In && "Couldn't get element of initializer?");
498 GlobalVariable *NGV = new GlobalVariable(Context,
499 STy->getElementType(i), false,
500 GlobalVariable::InternalLinkage,
501 In, GV->getName()+"."+Twine(i),
503 GV->getType()->getAddressSpace());
504 Globals.insert(GV, NGV);
505 NewGlobals.push_back(NGV);
507 // Calculate the known alignment of the field. If the original aggregate
508 // had 256 byte alignment for example, something might depend on that:
509 // propagate info to each field.
510 uint64_t FieldOffset = Layout.getElementOffset(i);
511 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset);
512 if (NewAlign > TD.getABITypeAlignment(STy->getElementType(i)))
513 NGV->setAlignment(NewAlign);
515 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
516 unsigned NumElements = 0;
517 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
518 NumElements = ATy->getNumElements();
520 NumElements = cast<VectorType>(STy)->getNumElements();
522 if (NumElements > 16 && GV->hasNUsesOrMore(16))
523 return 0; // It's not worth it.
524 NewGlobals.reserve(NumElements);
526 uint64_t EltSize = TD.getTypeAllocSize(STy->getElementType());
527 unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType());
528 for (unsigned i = 0, e = NumElements; i != e; ++i) {
529 Constant *In = getAggregateConstantElement(Init,
530 ConstantInt::get(Type::getInt32Ty(Context), i),
532 assert(In && "Couldn't get element of initializer?");
534 GlobalVariable *NGV = new GlobalVariable(Context,
535 STy->getElementType(), false,
536 GlobalVariable::InternalLinkage,
537 In, GV->getName()+"."+Twine(i),
539 GV->getType()->getAddressSpace());
540 Globals.insert(GV, NGV);
541 NewGlobals.push_back(NGV);
543 // Calculate the known alignment of the field. If the original aggregate
544 // had 256 byte alignment for example, something might depend on that:
545 // propagate info to each field.
546 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
547 if (NewAlign > EltAlign)
548 NGV->setAlignment(NewAlign);
552 if (NewGlobals.empty())
555 DEBUG(errs() << "PERFORMING GLOBAL SRA ON: " << *GV);
557 Constant *NullInt = Constant::getNullValue(Type::getInt32Ty(Context));
559 // Loop over all of the uses of the global, replacing the constantexpr geps,
560 // with smaller constantexpr geps or direct references.
561 while (!GV->use_empty()) {
562 User *GEP = GV->use_back();
563 assert(((isa<ConstantExpr>(GEP) &&
564 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
565 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
567 // Ignore the 1th operand, which has to be zero or else the program is quite
568 // broken (undefined). Get the 2nd operand, which is the structure or array
570 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
571 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
573 Value *NewPtr = NewGlobals[Val];
575 // Form a shorter GEP if needed.
576 if (GEP->getNumOperands() > 3) {
577 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
578 SmallVector<Constant*, 8> Idxs;
579 Idxs.push_back(NullInt);
580 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
581 Idxs.push_back(CE->getOperand(i));
582 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr),
583 &Idxs[0], Idxs.size());
585 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
586 SmallVector<Value*, 8> Idxs;
587 Idxs.push_back(NullInt);
588 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
589 Idxs.push_back(GEPI->getOperand(i));
590 NewPtr = GetElementPtrInst::Create(NewPtr, Idxs.begin(), Idxs.end(),
591 GEPI->getName()+"."+Twine(Val),GEPI);
594 GEP->replaceAllUsesWith(NewPtr);
596 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
597 GEPI->eraseFromParent();
599 cast<ConstantExpr>(GEP)->destroyConstant();
602 // Delete the old global, now that it is dead.
606 // Loop over the new globals array deleting any globals that are obviously
607 // dead. This can arise due to scalarization of a structure or an array that
608 // has elements that are dead.
609 unsigned FirstGlobal = 0;
610 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
611 if (NewGlobals[i]->use_empty()) {
612 Globals.erase(NewGlobals[i]);
613 if (FirstGlobal == i) ++FirstGlobal;
616 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
619 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
620 /// value will trap if the value is dynamically null. PHIs keeps track of any
621 /// phi nodes we've seen to avoid reprocessing them.
622 static bool AllUsesOfValueWillTrapIfNull(Value *V,
623 SmallPtrSet<PHINode*, 8> &PHIs) {
624 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
625 if (isa<LoadInst>(*UI)) {
627 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
628 if (SI->getOperand(0) == V) {
629 //cerr << "NONTRAPPING USE: " << **UI;
630 return false; // Storing the value.
632 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
633 if (CI->getOperand(0) != V) {
634 //cerr << "NONTRAPPING USE: " << **UI;
635 return false; // Not calling the ptr
637 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
638 if (II->getOperand(0) != V) {
639 //cerr << "NONTRAPPING USE: " << **UI;
640 return false; // Not calling the ptr
642 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(*UI)) {
643 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
644 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
645 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
646 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
647 // If we've already seen this phi node, ignore it, it has already been
650 return AllUsesOfValueWillTrapIfNull(PN, PHIs);
651 } else if (isa<ICmpInst>(*UI) &&
652 isa<ConstantPointerNull>(UI->getOperand(1))) {
653 // Ignore setcc X, null
655 //cerr << "NONTRAPPING USE: " << **UI;
661 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
662 /// from GV will trap if the loaded value is null. Note that this also permits
663 /// comparisons of the loaded value against null, as a special case.
664 static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) {
665 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI)
666 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
667 SmallPtrSet<PHINode*, 8> PHIs;
668 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
670 } else if (isa<StoreInst>(*UI)) {
671 // Ignore stores to the global.
673 // We don't know or understand this user, bail out.
674 //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI;
681 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV,
682 LLVMContext &Context) {
683 bool Changed = false;
684 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
685 Instruction *I = cast<Instruction>(*UI++);
686 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
687 LI->setOperand(0, NewV);
689 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
690 if (SI->getOperand(1) == V) {
691 SI->setOperand(1, NewV);
694 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
695 if (I->getOperand(0) == V) {
696 // Calling through the pointer! Turn into a direct call, but be careful
697 // that the pointer is not also being passed as an argument.
698 I->setOperand(0, NewV);
700 bool PassedAsArg = false;
701 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
702 if (I->getOperand(i) == V) {
704 I->setOperand(i, NewV);
708 // Being passed as an argument also. Be careful to not invalidate UI!
712 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
713 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
714 ConstantExpr::getCast(CI->getOpcode(),
715 NewV, CI->getType()), Context);
716 if (CI->use_empty()) {
718 CI->eraseFromParent();
720 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
721 // Should handle GEP here.
722 SmallVector<Constant*, 8> Idxs;
723 Idxs.reserve(GEPI->getNumOperands()-1);
724 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
726 if (Constant *C = dyn_cast<Constant>(*i))
730 if (Idxs.size() == GEPI->getNumOperands()-1)
731 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
732 ConstantExpr::getGetElementPtr(NewV, &Idxs[0],
733 Idxs.size()), Context);
734 if (GEPI->use_empty()) {
736 GEPI->eraseFromParent();
745 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
746 /// value stored into it. If there are uses of the loaded value that would trap
747 /// if the loaded value is dynamically null, then we know that they cannot be
748 /// reachable with a null optimize away the load.
749 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV,
750 LLVMContext &Context) {
751 bool Changed = false;
753 // Keep track of whether we are able to remove all the uses of the global
754 // other than the store that defines it.
755 bool AllNonStoreUsesGone = true;
757 // Replace all uses of loads with uses of uses of the stored value.
758 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end(); GUI != E;){
759 User *GlobalUser = *GUI++;
760 if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) {
761 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV, Context);
762 // If we were able to delete all uses of the loads
763 if (LI->use_empty()) {
764 LI->eraseFromParent();
767 AllNonStoreUsesGone = false;
769 } else if (isa<StoreInst>(GlobalUser)) {
770 // Ignore the store that stores "LV" to the global.
771 assert(GlobalUser->getOperand(1) == GV &&
772 "Must be storing *to* the global");
774 AllNonStoreUsesGone = false;
776 // If we get here we could have other crazy uses that are transitively
778 assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) ||
779 isa<ConstantExpr>(GlobalUser)) && "Only expect load and stores!");
784 DEBUG(errs() << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV);
788 // If we nuked all of the loads, then none of the stores are needed either,
789 // nor is the global.
790 if (AllNonStoreUsesGone) {
791 DEBUG(errs() << " *** GLOBAL NOW DEAD!\n");
792 CleanupConstantGlobalUsers(GV, 0, Context);
793 if (GV->use_empty()) {
794 GV->eraseFromParent();
802 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
803 /// instructions that are foldable.
804 static void ConstantPropUsersOf(Value *V, LLVMContext &Context) {
805 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
806 if (Instruction *I = dyn_cast<Instruction>(*UI++))
807 if (Constant *NewC = ConstantFoldInstruction(I, Context)) {
808 I->replaceAllUsesWith(NewC);
810 // Advance UI to the next non-I use to avoid invalidating it!
811 // Instructions could multiply use V.
812 while (UI != E && *UI == I)
814 I->eraseFromParent();
818 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
819 /// variable, and transforms the program as if it always contained the result of
820 /// the specified malloc. Because it is always the result of the specified
821 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
822 /// malloc into a global, and any loads of GV as uses of the new global.
823 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
827 LLVMContext &Context,
829 DEBUG(errs() << "PROMOTING MALLOC GLOBAL: " << *GV
830 << " CALL = " << *CI << " BCI = " << *BCI << '\n');
832 const Type *IntPtrTy = TD->getIntPtrType(Context);
834 ConstantInt *NElements = cast<ConstantInt>(NElems);
835 if (NElements->getZExtValue() != 1) {
836 // If we have an array allocation, transform it to a single element
837 // allocation to make the code below simpler.
838 Type *NewTy = ArrayType::get(getMallocAllocatedType(CI),
839 NElements->getZExtValue());
840 Value* NewM = CallInst::CreateMalloc(CI, IntPtrTy, NewTy);
841 Instruction* NewMI = cast<Instruction>(NewM);
843 Indices[0] = Indices[1] = Constant::getNullValue(IntPtrTy);
844 Value *NewGEP = GetElementPtrInst::Create(NewMI, Indices, Indices + 2,
845 NewMI->getName()+".el0", CI);
846 BCI->replaceAllUsesWith(NewGEP);
847 BCI->eraseFromParent();
848 CI->eraseFromParent();
849 BCI = cast<BitCastInst>(NewMI);
850 CI = extractMallocCallFromBitCast(NewMI);
853 // Create the new global variable. The contents of the malloc'd memory is
854 // undefined, so initialize with an undef value.
855 const Type *MAT = getMallocAllocatedType(CI);
856 Constant *Init = UndefValue::get(MAT);
857 GlobalVariable *NewGV = new GlobalVariable(*GV->getParent(),
859 GlobalValue::InternalLinkage, Init,
860 GV->getName()+".body",
862 GV->isThreadLocal());
864 // Anything that used the malloc now uses the global directly.
865 BCI->replaceAllUsesWith(NewGV);
867 Constant *RepValue = NewGV;
868 if (NewGV->getType() != GV->getType()->getElementType())
869 RepValue = ConstantExpr::getBitCast(RepValue,
870 GV->getType()->getElementType());
872 // If there is a comparison against null, we will insert a global bool to
873 // keep track of whether the global was initialized yet or not.
874 GlobalVariable *InitBool =
875 new GlobalVariable(Context, Type::getInt1Ty(Context), false,
876 GlobalValue::InternalLinkage,
877 ConstantInt::getFalse(Context), GV->getName()+".init",
878 GV->isThreadLocal());
879 bool InitBoolUsed = false;
881 // Loop over all uses of GV, processing them in turn.
882 std::vector<StoreInst*> Stores;
883 while (!GV->use_empty())
884 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
885 while (!LI->use_empty()) {
886 Use &LoadUse = LI->use_begin().getUse();
887 if (!isa<ICmpInst>(LoadUse.getUser()))
890 ICmpInst *ICI = cast<ICmpInst>(LoadUse.getUser());
891 // Replace the cmp X, 0 with a use of the bool value.
892 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", ICI);
894 switch (ICI->getPredicate()) {
895 default: llvm_unreachable("Unknown ICmp Predicate!");
896 case ICmpInst::ICMP_ULT:
897 case ICmpInst::ICMP_SLT:
898 LV = ConstantInt::getFalse(Context); // X < null -> always false
900 case ICmpInst::ICMP_ULE:
901 case ICmpInst::ICMP_SLE:
902 case ICmpInst::ICMP_EQ:
903 LV = BinaryOperator::CreateNot(LV, "notinit", ICI);
905 case ICmpInst::ICMP_NE:
906 case ICmpInst::ICMP_UGE:
907 case ICmpInst::ICMP_SGE:
908 case ICmpInst::ICMP_UGT:
909 case ICmpInst::ICMP_SGT:
912 ICI->replaceAllUsesWith(LV);
913 ICI->eraseFromParent();
916 LI->eraseFromParent();
918 StoreInst *SI = cast<StoreInst>(GV->use_back());
919 // The global is initialized when the store to it occurs.
920 new StoreInst(ConstantInt::getTrue(Context), InitBool, SI);
921 SI->eraseFromParent();
924 // If the initialization boolean was used, insert it, otherwise delete it.
926 while (!InitBool->use_empty()) // Delete initializations
927 cast<Instruction>(InitBool->use_back())->eraseFromParent();
930 GV->getParent()->getGlobalList().insert(GV, InitBool);
933 // Now the GV is dead, nuke it and the malloc.
934 GV->eraseFromParent();
935 BCI->eraseFromParent();
936 CI->eraseFromParent();
938 // To further other optimizations, loop over all users of NewGV and try to
939 // constant prop them. This will promote GEP instructions with constant
940 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
941 ConstantPropUsersOf(NewGV, Context);
942 if (RepValue != NewGV)
943 ConstantPropUsersOf(RepValue, Context);
948 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
949 /// to make sure that there are no complex uses of V. We permit simple things
950 /// like dereferencing the pointer, but not storing through the address, unless
951 /// it is to the specified global.
952 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V,
954 SmallPtrSet<PHINode*, 8> &PHIs) {
955 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
956 Instruction *Inst = cast<Instruction>(*UI);
958 if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) {
959 continue; // Fine, ignore.
962 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
963 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
964 return false; // Storing the pointer itself... bad.
965 continue; // Otherwise, storing through it, or storing into GV... fine.
968 if (isa<GetElementPtrInst>(Inst)) {
969 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs))
974 if (PHINode *PN = dyn_cast<PHINode>(Inst)) {
975 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI
978 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
983 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) {
984 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs))
994 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
995 /// somewhere. Transform all uses of the allocation into loads from the
996 /// global and uses of the resultant pointer. Further, delete the store into
997 /// GV. This assumes that these value pass the
998 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
999 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
1000 GlobalVariable *GV) {
1001 while (!Alloc->use_empty()) {
1002 Instruction *U = cast<Instruction>(*Alloc->use_begin());
1003 Instruction *InsertPt = U;
1004 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1005 // If this is the store of the allocation into the global, remove it.
1006 if (SI->getOperand(1) == GV) {
1007 SI->eraseFromParent();
1010 } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
1011 // Insert the load in the corresponding predecessor, not right before the
1013 InsertPt = PN->getIncomingBlock(Alloc->use_begin())->getTerminator();
1014 } else if (isa<BitCastInst>(U)) {
1015 // Must be bitcast between the malloc and store to initialize the global.
1016 ReplaceUsesOfMallocWithGlobal(U, GV);
1017 U->eraseFromParent();
1019 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
1020 // If this is a "GEP bitcast" and the user is a store to the global, then
1021 // just process it as a bitcast.
1022 if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse())
1023 if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->use_back()))
1024 if (SI->getOperand(1) == GV) {
1025 // Must be bitcast GEP between the malloc and store to initialize
1027 ReplaceUsesOfMallocWithGlobal(GEPI, GV);
1028 GEPI->eraseFromParent();
1033 // Insert a load from the global, and use it instead of the malloc.
1034 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
1035 U->replaceUsesOfWith(Alloc, NL);
1039 /// LoadUsesSimpleEnoughForHeapSRA - Verify that all uses of V (a load, or a phi
1040 /// of a load) are simple enough to perform heap SRA on. This permits GEP's
1041 /// that index through the array and struct field, icmps of null, and PHIs.
1042 static bool LoadUsesSimpleEnoughForHeapSRA(Value *V,
1043 SmallPtrSet<PHINode*, 32> &LoadUsingPHIs,
1044 SmallPtrSet<PHINode*, 32> &LoadUsingPHIsPerLoad) {
1045 // We permit two users of the load: setcc comparing against the null
1046 // pointer, and a getelementptr of a specific form.
1047 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
1048 Instruction *User = cast<Instruction>(*UI);
1050 // Comparison against null is ok.
1051 if (ICmpInst *ICI = dyn_cast<ICmpInst>(User)) {
1052 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1057 // getelementptr is also ok, but only a simple form.
1058 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1059 // Must index into the array and into the struct.
1060 if (GEPI->getNumOperands() < 3)
1063 // Otherwise the GEP is ok.
1067 if (PHINode *PN = dyn_cast<PHINode>(User)) {
1068 if (!LoadUsingPHIsPerLoad.insert(PN))
1069 // This means some phi nodes are dependent on each other.
1070 // Avoid infinite looping!
1072 if (!LoadUsingPHIs.insert(PN))
1073 // If we have already analyzed this PHI, then it is safe.
1076 // Make sure all uses of the PHI are simple enough to transform.
1077 if (!LoadUsesSimpleEnoughForHeapSRA(PN,
1078 LoadUsingPHIs, LoadUsingPHIsPerLoad))
1084 // Otherwise we don't know what this is, not ok.
1092 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
1093 /// GV are simple enough to perform HeapSRA, return true.
1094 static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV,
1095 Instruction *StoredVal) {
1096 SmallPtrSet<PHINode*, 32> LoadUsingPHIs;
1097 SmallPtrSet<PHINode*, 32> LoadUsingPHIsPerLoad;
1098 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;
1100 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
1101 if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs,
1102 LoadUsingPHIsPerLoad))
1104 LoadUsingPHIsPerLoad.clear();
1107 // If we reach here, we know that all uses of the loads and transitive uses
1108 // (through PHI nodes) are simple enough to transform. However, we don't know
1109 // that all inputs the to the PHI nodes are in the same equivalence sets.
1110 // Check to verify that all operands of the PHIs are either PHIS that can be
1111 // transformed, loads from GV, or MI itself.
1112 for (SmallPtrSet<PHINode*, 32>::iterator I = LoadUsingPHIs.begin(),
1113 E = LoadUsingPHIs.end(); I != E; ++I) {
1115 for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) {
1116 Value *InVal = PN->getIncomingValue(op);
1118 // PHI of the stored value itself is ok.
1119 if (InVal == StoredVal) continue;
1121 if (PHINode *InPN = dyn_cast<PHINode>(InVal)) {
1122 // One of the PHIs in our set is (optimistically) ok.
1123 if (LoadUsingPHIs.count(InPN))
1128 // Load from GV is ok.
1129 if (LoadInst *LI = dyn_cast<LoadInst>(InVal))
1130 if (LI->getOperand(0) == GV)
1135 // Anything else is rejected.
1143 static Value *GetHeapSROAValue(Value *V, unsigned FieldNo,
1144 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1145 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite,
1146 LLVMContext &Context) {
1147 std::vector<Value*> &FieldVals = InsertedScalarizedValues[V];
1149 if (FieldNo >= FieldVals.size())
1150 FieldVals.resize(FieldNo+1);
1152 // If we already have this value, just reuse the previously scalarized
1154 if (Value *FieldVal = FieldVals[FieldNo])
1157 // Depending on what instruction this is, we have several cases.
1159 if (LoadInst *LI = dyn_cast<LoadInst>(V)) {
1160 // This is a scalarized version of the load from the global. Just create
1161 // a new Load of the scalarized global.
1162 Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo,
1163 InsertedScalarizedValues,
1164 PHIsToRewrite, Context),
1165 LI->getName()+".f"+Twine(FieldNo), LI);
1166 } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1167 // PN's type is pointer to struct. Make a new PHI of pointer to struct
1169 const StructType *ST =
1170 cast<StructType>(cast<PointerType>(PN->getType())->getElementType());
1173 PHINode::Create(PointerType::getUnqual(ST->getElementType(FieldNo)),
1174 PN->getName()+".f"+Twine(FieldNo), PN);
1175 PHIsToRewrite.push_back(std::make_pair(PN, FieldNo));
1177 llvm_unreachable("Unknown usable value");
1181 return FieldVals[FieldNo] = Result;
1184 /// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1185 /// the load, rewrite the derived value to use the HeapSRoA'd load.
1186 static void RewriteHeapSROALoadUser(Instruction *LoadUser,
1187 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1188 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite,
1189 LLVMContext &Context) {
1190 // If this is a comparison against null, handle it.
1191 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1192 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1193 // If we have a setcc of the loaded pointer, we can use a setcc of any
1195 Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0,
1196 InsertedScalarizedValues, PHIsToRewrite,
1199 Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr,
1200 Constant::getNullValue(NPtr->getType()),
1202 SCI->replaceAllUsesWith(New);
1203 SCI->eraseFromParent();
1207 // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...'
1208 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1209 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1210 && "Unexpected GEPI!");
1212 // Load the pointer for this field.
1213 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1214 Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo,
1215 InsertedScalarizedValues, PHIsToRewrite,
1218 // Create the new GEP idx vector.
1219 SmallVector<Value*, 8> GEPIdx;
1220 GEPIdx.push_back(GEPI->getOperand(1));
1221 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1223 Value *NGEPI = GetElementPtrInst::Create(NewPtr,
1224 GEPIdx.begin(), GEPIdx.end(),
1225 GEPI->getName(), GEPI);
1226 GEPI->replaceAllUsesWith(NGEPI);
1227 GEPI->eraseFromParent();
1231 // Recursively transform the users of PHI nodes. This will lazily create the
1232 // PHIs that are needed for individual elements. Keep track of what PHIs we
1233 // see in InsertedScalarizedValues so that we don't get infinite loops (very
1234 // antisocial). If the PHI is already in InsertedScalarizedValues, it has
1235 // already been seen first by another load, so its uses have already been
1237 PHINode *PN = cast<PHINode>(LoadUser);
1239 DenseMap<Value*, std::vector<Value*> >::iterator InsertPos;
1240 tie(InsertPos, Inserted) =
1241 InsertedScalarizedValues.insert(std::make_pair(PN, std::vector<Value*>()));
1242 if (!Inserted) return;
1244 // If this is the first time we've seen this PHI, recursively process all
1246 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); UI != E; ) {
1247 Instruction *User = cast<Instruction>(*UI++);
1248 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite,
1253 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1254 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1255 /// use FieldGlobals instead. All uses of loaded values satisfy
1256 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA.
1257 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1258 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1259 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite,
1260 LLVMContext &Context) {
1261 for (Value::use_iterator UI = Load->use_begin(), E = Load->use_end();
1263 Instruction *User = cast<Instruction>(*UI++);
1264 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite,
1268 if (Load->use_empty()) {
1269 Load->eraseFromParent();
1270 InsertedScalarizedValues.erase(Load);
1274 /// PerformHeapAllocSRoA - CI is an allocation of an array of structures. Break
1275 /// it up into multiple allocations of arrays of the fields.
1276 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV,
1277 CallInst *CI, BitCastInst* BCI,
1279 LLVMContext &Context,
1281 DEBUG(errs() << "SROA HEAP ALLOC: " << *GV << " MALLOC CALL = " << *CI
1282 << " BITCAST = " << *BCI << '\n');
1283 const Type* MAT = getMallocAllocatedType(CI);
1284 const StructType *STy = cast<StructType>(MAT);
1286 // There is guaranteed to be at least one use of the malloc (storing
1287 // it into GV). If there are other uses, change them to be uses of
1288 // the global to simplify later code. This also deletes the store
1290 ReplaceUsesOfMallocWithGlobal(BCI, GV);
1292 // Okay, at this point, there are no users of the malloc. Insert N
1293 // new mallocs at the same place as CI, and N globals.
1294 std::vector<Value*> FieldGlobals;
1295 std::vector<Value*> FieldMallocs;
1297 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1298 const Type *FieldTy = STy->getElementType(FieldNo);
1299 const PointerType *PFieldTy = PointerType::getUnqual(FieldTy);
1301 GlobalVariable *NGV =
1302 new GlobalVariable(*GV->getParent(),
1303 PFieldTy, false, GlobalValue::InternalLinkage,
1304 Constant::getNullValue(PFieldTy),
1305 GV->getName() + ".f" + Twine(FieldNo), GV,
1306 GV->isThreadLocal());
1307 FieldGlobals.push_back(NGV);
1309 Value *NMI = CallInst::CreateMalloc(CI, TD->getIntPtrType(Context),
1311 BCI->getName() + ".f" + Twine(FieldNo));
1312 FieldMallocs.push_back(NMI);
1313 new StoreInst(NMI, NGV, BCI);
1316 // The tricky aspect of this transformation is handling the case when malloc
1317 // fails. In the original code, malloc failing would set the result pointer
1318 // of malloc to null. In this case, some mallocs could succeed and others
1319 // could fail. As such, we emit code that looks like this:
1320 // F0 = malloc(field0)
1321 // F1 = malloc(field1)
1322 // F2 = malloc(field2)
1323 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1324 // if (F0) { free(F0); F0 = 0; }
1325 // if (F1) { free(F1); F1 = 0; }
1326 // if (F2) { free(F2); F2 = 0; }
1328 Value *RunningOr = 0;
1329 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1330 Value *Cond = new ICmpInst(BCI, ICmpInst::ICMP_EQ, FieldMallocs[i],
1331 Constant::getNullValue(FieldMallocs[i]->getType()),
1334 RunningOr = Cond; // First seteq
1336 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", BCI);
1339 // Split the basic block at the old malloc.
1340 BasicBlock *OrigBB = BCI->getParent();
1341 BasicBlock *ContBB = OrigBB->splitBasicBlock(BCI, "malloc_cont");
1343 // Create the block to check the first condition. Put all these blocks at the
1344 // end of the function as they are unlikely to be executed.
1345 BasicBlock *NullPtrBlock = BasicBlock::Create(Context, "malloc_ret_null",
1346 OrigBB->getParent());
1348 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1349 // branch on RunningOr.
1350 OrigBB->getTerminator()->eraseFromParent();
1351 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1353 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1354 // pointer, because some may be null while others are not.
1355 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1356 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1357 Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
1358 Constant::getNullValue(GVVal->getType()),
1360 BasicBlock *FreeBlock = BasicBlock::Create(Context, "free_it",
1361 OrigBB->getParent());
1362 BasicBlock *NextBlock = BasicBlock::Create(Context, "next",
1363 OrigBB->getParent());
1364 Instruction *BI = BranchInst::Create(FreeBlock, NextBlock,
1367 // Fill in FreeBlock.
1368 CallInst::CreateFree(GVVal, BI);
1369 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1371 BranchInst::Create(NextBlock, FreeBlock);
1373 NullPtrBlock = NextBlock;
1376 BranchInst::Create(ContBB, NullPtrBlock);
1378 // CI and BCI are no longer needed, remove them.
1379 BCI->eraseFromParent();
1380 CI->eraseFromParent();
1382 /// InsertedScalarizedLoads - As we process loads, if we can't immediately
1383 /// update all uses of the load, keep track of what scalarized loads are
1384 /// inserted for a given load.
1385 DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues;
1386 InsertedScalarizedValues[GV] = FieldGlobals;
1388 std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite;
1390 // Okay, the malloc site is completely handled. All of the uses of GV are now
1391 // loads, and all uses of those loads are simple. Rewrite them to use loads
1392 // of the per-field globals instead.
1393 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;) {
1394 Instruction *User = cast<Instruction>(*UI++);
1396 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1397 RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite,
1402 // Must be a store of null.
1403 StoreInst *SI = cast<StoreInst>(User);
1404 assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
1405 "Unexpected heap-sra user!");
1407 // Insert a store of null into each global.
1408 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1409 const PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType());
1410 Constant *Null = Constant::getNullValue(PT->getElementType());
1411 new StoreInst(Null, FieldGlobals[i], SI);
1413 // Erase the original store.
1414 SI->eraseFromParent();
1417 // While we have PHIs that are interesting to rewrite, do it.
1418 while (!PHIsToRewrite.empty()) {
1419 PHINode *PN = PHIsToRewrite.back().first;
1420 unsigned FieldNo = PHIsToRewrite.back().second;
1421 PHIsToRewrite.pop_back();
1422 PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]);
1423 assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi");
1425 // Add all the incoming values. This can materialize more phis.
1426 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1427 Value *InVal = PN->getIncomingValue(i);
1428 InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues,
1429 PHIsToRewrite, Context);
1430 FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
1434 // Drop all inter-phi links and any loads that made it this far.
1435 for (DenseMap<Value*, std::vector<Value*> >::iterator
1436 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1438 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1439 PN->dropAllReferences();
1440 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1441 LI->dropAllReferences();
1444 // Delete all the phis and loads now that inter-references are dead.
1445 for (DenseMap<Value*, std::vector<Value*> >::iterator
1446 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1448 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1449 PN->eraseFromParent();
1450 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1451 LI->eraseFromParent();
1454 // The old global is now dead, remove it.
1455 GV->eraseFromParent();
1458 return cast<GlobalVariable>(FieldGlobals[0]);
1461 /// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1462 /// pointer global variable with a single value stored it that is a malloc or
1464 static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
1467 Module::global_iterator &GVI,
1469 LLVMContext &Context) {
1470 // If we can't figure out the type being malloced, then we can't optimize.
1471 const Type *AllocTy = getMallocAllocatedType(CI);
1474 // If this is a malloc of an abstract type, don't touch it.
1475 if (!AllocTy->isSized())
1478 // We can't optimize this global unless all uses of it are *known* to be
1479 // of the malloc value, not of the null initializer value (consider a use
1480 // that compares the global's value against zero to see if the malloc has
1481 // been reached). To do this, we check to see if all uses of the global
1482 // would trap if the global were null: this proves that they must all
1483 // happen after the malloc.
1484 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1487 // We can't optimize this if the malloc itself is used in a complex way,
1488 // for example, being stored into multiple globals. This allows the
1489 // malloc to be stored into the specified global, loaded setcc'd, and
1490 // GEP'd. These are all things we could transform to using the global
1493 SmallPtrSet<PHINode*, 8> PHIs;
1494 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs))
1498 // If we have a global that is only initialized with a fixed size malloc,
1499 // transform the program to use global memory instead of malloc'd memory.
1500 // This eliminates dynamic allocation, avoids an indirection accessing the
1501 // data, and exposes the resultant global to further GlobalOpt.
1502 Value *NElems = getMallocArraySize(CI, Context, TD);
1503 // We cannot optimize the malloc if we cannot determine malloc array size.
1505 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
1506 // Restrict this transformation to only working on small allocations
1507 // (2048 bytes currently), as we don't want to introduce a 16M global or
1510 NElements->getZExtValue() * TD->getTypeAllocSize(AllocTy) < 2048) {
1511 GVI = OptimizeGlobalAddressOfMalloc(GV, CI, BCI, NElems, Context, TD);
1515 // If the allocation is an array of structures, consider transforming this
1516 // into multiple malloc'd arrays, one for each field. This is basically
1517 // SRoA for malloc'd memory.
1519 // If this is an allocation of a fixed size array of structs, analyze as a
1520 // variable size array. malloc [100 x struct],1 -> malloc struct, 100
1521 if (NElems == ConstantInt::get(CI->getOperand(1)->getType(), 1))
1522 if (const ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
1523 AllocTy = AT->getElementType();
1525 if (const StructType *AllocSTy = dyn_cast<StructType>(AllocTy)) {
1526 // This the structure has an unreasonable number of fields, leave it
1528 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
1529 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, BCI)) {
1531 // If this is a fixed size array, transform the Malloc to be an alloc of
1532 // structs. malloc [100 x struct],1 -> malloc struct, 100
1533 if (const ArrayType *AT =
1534 dyn_cast<ArrayType>(getMallocAllocatedType(CI))) {
1535 Value* NumElements = ConstantInt::get(Type::getInt32Ty(Context),
1536 AT->getNumElements());
1537 Value* NewMI = CallInst::CreateMalloc(CI, TD->getIntPtrType(Context),
1538 AllocSTy, NumElements,
1540 Value *Cast = new BitCastInst(NewMI, getMallocType(CI), "tmp", CI);
1541 BCI->replaceAllUsesWith(Cast);
1542 BCI->eraseFromParent();
1543 CI->eraseFromParent();
1544 BCI = cast<BitCastInst>(NewMI);
1545 CI = extractMallocCallFromBitCast(NewMI);
1548 GVI = PerformHeapAllocSRoA(GV, CI, BCI, NElems, Context, TD);
1557 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1558 // that only one value (besides its initializer) is ever stored to the global.
1559 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1560 Module::global_iterator &GVI,
1561 TargetData *TD, LLVMContext &Context) {
1562 // Ignore no-op GEPs and bitcasts.
1563 StoredOnceVal = StoredOnceVal->stripPointerCasts();
1565 // If we are dealing with a pointer global that is initialized to null and
1566 // only has one (non-null) value stored into it, then we can optimize any
1567 // users of the loaded value (often calls and loads) that would trap if the
1569 if (isa<PointerType>(GV->getInitializer()->getType()) &&
1570 GV->getInitializer()->isNullValue()) {
1571 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1572 if (GV->getInitializer()->getType() != SOVC->getType())
1574 ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1576 // Optimize away any trapping uses of the loaded value.
1577 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC, Context))
1579 } else if (CallInst *CI = extractMallocCall(StoredOnceVal)) {
1580 if (getMallocAllocatedType(CI)) {
1581 BitCastInst* BCI = NULL;
1582 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
1584 BCI = dyn_cast<BitCastInst>(cast<Instruction>(*UI++));
1586 TryToOptimizeStoreOfMallocToGlobal(GV, CI, BCI, GVI, TD, Context))
1595 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1596 /// two values ever stored into GV are its initializer and OtherVal. See if we
1597 /// can shrink the global into a boolean and select between the two values
1598 /// whenever it is used. This exposes the values to other scalar optimizations.
1599 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal,
1600 LLVMContext &Context) {
1601 const Type *GVElType = GV->getType()->getElementType();
1603 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1604 // an FP value, pointer or vector, don't do this optimization because a select
1605 // between them is very expensive and unlikely to lead to later
1606 // simplification. In these cases, we typically end up with "cond ? v1 : v2"
1607 // where v1 and v2 both require constant pool loads, a big loss.
1608 if (GVElType == Type::getInt1Ty(Context) || GVElType->isFloatingPoint() ||
1609 isa<PointerType>(GVElType) || isa<VectorType>(GVElType))
1612 // Walk the use list of the global seeing if all the uses are load or store.
1613 // If there is anything else, bail out.
1614 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I)
1615 if (!isa<LoadInst>(I) && !isa<StoreInst>(I))
1618 DEBUG(errs() << " *** SHRINKING TO BOOL: " << *GV);
1620 // Create the new global, initializing it to false.
1621 GlobalVariable *NewGV = new GlobalVariable(Context,
1622 Type::getInt1Ty(Context), false,
1623 GlobalValue::InternalLinkage, ConstantInt::getFalse(Context),
1625 GV->isThreadLocal());
1626 GV->getParent()->getGlobalList().insert(GV, NewGV);
1628 Constant *InitVal = GV->getInitializer();
1629 assert(InitVal->getType() != Type::getInt1Ty(Context) &&
1630 "No reason to shrink to bool!");
1632 // If initialized to zero and storing one into the global, we can use a cast
1633 // instead of a select to synthesize the desired value.
1634 bool IsOneZero = false;
1635 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1636 IsOneZero = InitVal->isNullValue() && CI->isOne();
1638 while (!GV->use_empty()) {
1639 Instruction *UI = cast<Instruction>(GV->use_back());
1640 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1641 // Change the store into a boolean store.
1642 bool StoringOther = SI->getOperand(0) == OtherVal;
1643 // Only do this if we weren't storing a loaded value.
1645 if (StoringOther || SI->getOperand(0) == InitVal)
1646 StoreVal = ConstantInt::get(Type::getInt1Ty(Context), StoringOther);
1648 // Otherwise, we are storing a previously loaded copy. To do this,
1649 // change the copy from copying the original value to just copying the
1651 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1653 // If we're already replaced the input, StoredVal will be a cast or
1654 // select instruction. If not, it will be a load of the original
1656 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1657 assert(LI->getOperand(0) == GV && "Not a copy!");
1658 // Insert a new load, to preserve the saved value.
1659 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1661 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1662 "This is not a form that we understand!");
1663 StoreVal = StoredVal->getOperand(0);
1664 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1667 new StoreInst(StoreVal, NewGV, SI);
1669 // Change the load into a load of bool then a select.
1670 LoadInst *LI = cast<LoadInst>(UI);
1671 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1674 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1676 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1678 LI->replaceAllUsesWith(NSI);
1680 UI->eraseFromParent();
1683 GV->eraseFromParent();
1688 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1689 /// it if possible. If we make a change, return true.
1690 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1691 Module::global_iterator &GVI) {
1692 SmallPtrSet<PHINode*, 16> PHIUsers;
1694 GV->removeDeadConstantUsers();
1696 if (GV->use_empty()) {
1697 DEBUG(errs() << "GLOBAL DEAD: " << *GV);
1698 GV->eraseFromParent();
1703 if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
1705 cerr << "Global: " << *GV;
1706 cerr << " isLoaded = " << GS.isLoaded << "\n";
1707 cerr << " StoredType = ";
1708 switch (GS.StoredType) {
1709 case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break;
1710 case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break;
1711 case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break;
1712 case GlobalStatus::isStored: cerr << "stored\n"; break;
1714 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
1715 cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n";
1716 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
1717 cerr << " AccessingFunction = " << GS.AccessingFunction->getName()
1719 cerr << " HasMultipleAccessingFunctions = "
1720 << GS.HasMultipleAccessingFunctions << "\n";
1721 cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n";
1725 // If this is a first class global and has only one accessing function
1726 // and this function is main (which we know is not recursive we can make
1727 // this global a local variable) we replace the global with a local alloca
1728 // in this function.
1730 // NOTE: It doesn't make sense to promote non single-value types since we
1731 // are just replacing static memory to stack memory.
1733 // If the global is in different address space, don't bring it to stack.
1734 if (!GS.HasMultipleAccessingFunctions &&
1735 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1736 GV->getType()->getElementType()->isSingleValueType() &&
1737 GS.AccessingFunction->getName() == "main" &&
1738 GS.AccessingFunction->hasExternalLinkage() &&
1739 GV->getType()->getAddressSpace() == 0) {
1740 DEBUG(errs() << "LOCALIZING GLOBAL: " << *GV);
1741 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin();
1742 const Type* ElemTy = GV->getType()->getElementType();
1743 // FIXME: Pass Global's alignment when globals have alignment
1744 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI);
1745 if (!isa<UndefValue>(GV->getInitializer()))
1746 new StoreInst(GV->getInitializer(), Alloca, FirstI);
1748 GV->replaceAllUsesWith(Alloca);
1749 GV->eraseFromParent();
1754 // If the global is never loaded (but may be stored to), it is dead.
1757 DEBUG(errs() << "GLOBAL NEVER LOADED: " << *GV);
1759 // Delete any stores we can find to the global. We may not be able to
1760 // make it completely dead though.
1761 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer(),
1764 // If the global is dead now, delete it.
1765 if (GV->use_empty()) {
1766 GV->eraseFromParent();
1772 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1773 DEBUG(errs() << "MARKING CONSTANT: " << *GV);
1774 GV->setConstant(true);
1776 // Clean up any obviously simplifiable users now.
1777 CleanupConstantGlobalUsers(GV, GV->getInitializer(), GV->getContext());
1779 // If the global is dead now, just nuke it.
1780 if (GV->use_empty()) {
1781 DEBUG(errs() << " *** Marking constant allowed us to simplify "
1782 << "all users and delete global!\n");
1783 GV->eraseFromParent();
1789 } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
1790 if (TargetData *TD = getAnalysisIfAvailable<TargetData>())
1791 if (GlobalVariable *FirstNewGV = SRAGlobal(GV, *TD,
1792 GV->getContext())) {
1793 GVI = FirstNewGV; // Don't skip the newly produced globals!
1796 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1797 // If the initial value for the global was an undef value, and if only
1798 // one other value was stored into it, we can just change the
1799 // initializer to be the stored value, then delete all stores to the
1800 // global. This allows us to mark it constant.
1801 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1802 if (isa<UndefValue>(GV->getInitializer())) {
1803 // Change the initial value here.
1804 GV->setInitializer(SOVConstant);
1806 // Clean up any obviously simplifiable users now.
1807 CleanupConstantGlobalUsers(GV, GV->getInitializer(),
1810 if (GV->use_empty()) {
1811 DEBUG(errs() << " *** Substituting initializer allowed us to "
1812 << "simplify all users and delete global!\n");
1813 GV->eraseFromParent();
1822 // Try to optimize globals based on the knowledge that only one value
1823 // (besides its initializer) is ever stored to the global.
1824 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1825 getAnalysisIfAvailable<TargetData>(),
1829 // Otherwise, if the global was not a boolean, we can shrink it to be a
1831 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1832 if (TryToShrinkGlobalToBoolean(GV, SOVConstant, GV->getContext())) {
1841 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1842 /// function, changing them to FastCC.
1843 static void ChangeCalleesToFastCall(Function *F) {
1844 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1845 CallSite User(cast<Instruction>(*UI));
1846 User.setCallingConv(CallingConv::Fast);
1850 static AttrListPtr StripNest(const AttrListPtr &Attrs) {
1851 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1852 if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0)
1855 // There can be only one.
1856 return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest);
1862 static void RemoveNestAttribute(Function *F) {
1863 F->setAttributes(StripNest(F->getAttributes()));
1864 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1865 CallSite User(cast<Instruction>(*UI));
1866 User.setAttributes(StripNest(User.getAttributes()));
1870 bool GlobalOpt::OptimizeFunctions(Module &M) {
1871 bool Changed = false;
1872 // Optimize functions.
1873 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1875 // Functions without names cannot be referenced outside this module.
1876 if (!F->hasName() && !F->isDeclaration())
1877 F->setLinkage(GlobalValue::InternalLinkage);
1878 F->removeDeadConstantUsers();
1879 if (F->use_empty() && (F->hasLocalLinkage() || F->hasLinkOnceLinkage())) {
1880 F->eraseFromParent();
1883 } else if (F->hasLocalLinkage()) {
1884 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1885 !F->hasAddressTaken()) {
1886 // If this function has C calling conventions, is not a varargs
1887 // function, and is only called directly, promote it to use the Fast
1888 // calling convention.
1889 F->setCallingConv(CallingConv::Fast);
1890 ChangeCalleesToFastCall(F);
1895 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
1896 !F->hasAddressTaken()) {
1897 // The function is not used by a trampoline intrinsic, so it is safe
1898 // to remove the 'nest' attribute.
1899 RemoveNestAttribute(F);
1908 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1909 bool Changed = false;
1910 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1912 GlobalVariable *GV = GVI++;
1913 // Global variables without names cannot be referenced outside this module.
1914 if (!GV->hasName() && !GV->isDeclaration())
1915 GV->setLinkage(GlobalValue::InternalLinkage);
1916 if (!GV->isConstant() && GV->hasLocalLinkage() &&
1917 GV->hasInitializer())
1918 Changed |= ProcessInternalGlobal(GV, GVI);
1923 /// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1924 /// initializers have an init priority of 65535.
1925 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1926 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1928 if (I->getName() == "llvm.global_ctors") {
1929 // Found it, verify it's an array of { int, void()* }.
1930 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1932 const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1933 if (!STy || STy->getNumElements() != 2 ||
1934 STy->getElementType(0) != Type::getInt32Ty(M.getContext())) return 0;
1935 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1936 if (!PFTy) return 0;
1937 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1938 if (!FTy || FTy->getReturnType() != Type::getVoidTy(M.getContext()) ||
1939 FTy->isVarArg() || FTy->getNumParams() != 0)
1942 // Verify that the initializer is simple enough for us to handle.
1943 if (!I->hasDefinitiveInitializer()) return 0;
1944 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1946 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
1947 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(*i)) {
1948 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1951 // Must have a function or null ptr.
1952 if (!isa<Function>(CS->getOperand(1)))
1955 // Init priority must be standard.
1956 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1957 if (!CI || CI->getZExtValue() != 65535)
1968 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1969 /// return a list of the functions and null terminator as a vector.
1970 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1971 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1972 std::vector<Function*> Result;
1973 Result.reserve(CA->getNumOperands());
1974 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
1975 ConstantStruct *CS = cast<ConstantStruct>(*i);
1976 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1981 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1982 /// specified array, returning the new global to use.
1983 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1984 const std::vector<Function*> &Ctors,
1985 LLVMContext &Context) {
1986 // If we made a change, reassemble the initializer list.
1987 std::vector<Constant*> CSVals;
1988 CSVals.push_back(ConstantInt::get(Type::getInt32Ty(Context), 65535));
1989 CSVals.push_back(0);
1991 // Create the new init list.
1992 std::vector<Constant*> CAList;
1993 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1995 CSVals[1] = Ctors[i];
1997 const Type *FTy = FunctionType::get(Type::getVoidTy(Context), false);
1998 const PointerType *PFTy = PointerType::getUnqual(FTy);
1999 CSVals[1] = Constant::getNullValue(PFTy);
2000 CSVals[0] = ConstantInt::get(Type::getInt32Ty(Context), 2147483647);
2002 CAList.push_back(ConstantStruct::get(Context, CSVals, false));
2005 // Create the array initializer.
2006 const Type *StructTy =
2007 cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
2008 Constant *CA = ConstantArray::get(ArrayType::get(StructTy,
2009 CAList.size()), CAList);
2011 // If we didn't change the number of elements, don't create a new GV.
2012 if (CA->getType() == GCL->getInitializer()->getType()) {
2013 GCL->setInitializer(CA);
2017 // Create the new global and insert it next to the existing list.
2018 GlobalVariable *NGV = new GlobalVariable(Context, CA->getType(),
2020 GCL->getLinkage(), CA, "",
2021 GCL->isThreadLocal());
2022 GCL->getParent()->getGlobalList().insert(GCL, NGV);
2025 // Nuke the old list, replacing any uses with the new one.
2026 if (!GCL->use_empty()) {
2028 if (V->getType() != GCL->getType())
2029 V = ConstantExpr::getBitCast(V, GCL->getType());
2030 GCL->replaceAllUsesWith(V);
2032 GCL->eraseFromParent();
2041 static Constant *getVal(DenseMap<Value*, Constant*> &ComputedValues,
2043 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
2044 Constant *R = ComputedValues[V];
2045 assert(R && "Reference to an uncomputed value!");
2049 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
2050 /// enough for us to understand. In particular, if it is a cast of something,
2051 /// we punt. We basically just support direct accesses to globals and GEP's of
2052 /// globals. This should be kept up to date with CommitValueTo.
2053 static bool isSimpleEnoughPointerToCommit(Constant *C, LLVMContext &Context) {
2054 // Conservatively, avoid aggregate types. This is because we don't
2055 // want to worry about them partially overlapping other stores.
2056 if (!cast<PointerType>(C->getType())->getElementType()->isSingleValueType())
2059 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
2060 // Do not allow weak/linkonce/dllimport/dllexport linkage or
2061 // external globals.
2062 return GV->hasDefinitiveInitializer();
2064 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2065 // Handle a constantexpr gep.
2066 if (CE->getOpcode() == Instruction::GetElementPtr &&
2067 isa<GlobalVariable>(CE->getOperand(0)) &&
2068 cast<GEPOperator>(CE)->isInBounds()) {
2069 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2070 // Do not allow weak/linkonce/dllimport/dllexport linkage or
2071 // external globals.
2072 if (!GV->hasDefinitiveInitializer())
2075 // The first index must be zero.
2076 ConstantInt *CI = dyn_cast<ConstantInt>(*next(CE->op_begin()));
2077 if (!CI || !CI->isZero()) return false;
2079 // The remaining indices must be compile-time known integers within the
2080 // notional bounds of the corresponding static array types.
2081 if (!CE->isGEPWithNoNotionalOverIndexing())
2084 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2089 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
2090 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
2091 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
2092 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
2093 ConstantExpr *Addr, unsigned OpNo,
2094 LLVMContext &Context) {
2095 // Base case of the recursion.
2096 if (OpNo == Addr->getNumOperands()) {
2097 assert(Val->getType() == Init->getType() && "Type mismatch!");
2101 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
2102 std::vector<Constant*> Elts;
2104 // Break up the constant into its elements.
2105 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
2106 for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i)
2107 Elts.push_back(cast<Constant>(*i));
2108 } else if (isa<ConstantAggregateZero>(Init)) {
2109 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2110 Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
2111 } else if (isa<UndefValue>(Init)) {
2112 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2113 Elts.push_back(UndefValue::get(STy->getElementType(i)));
2115 llvm_unreachable("This code is out of sync with "
2116 " ConstantFoldLoadThroughGEPConstantExpr");
2119 // Replace the element that we are supposed to.
2120 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
2121 unsigned Idx = CU->getZExtValue();
2122 assert(Idx < STy->getNumElements() && "Struct index out of range!");
2123 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1, Context);
2125 // Return the modified struct.
2126 return ConstantStruct::get(Context, &Elts[0], Elts.size(), STy->isPacked());
2128 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
2129 const ArrayType *ATy = cast<ArrayType>(Init->getType());
2131 // Break up the array into elements.
2132 std::vector<Constant*> Elts;
2133 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
2134 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
2135 Elts.push_back(cast<Constant>(*i));
2136 } else if (isa<ConstantAggregateZero>(Init)) {
2137 Constant *Elt = Constant::getNullValue(ATy->getElementType());
2138 Elts.assign(ATy->getNumElements(), Elt);
2139 } else if (isa<UndefValue>(Init)) {
2140 Constant *Elt = UndefValue::get(ATy->getElementType());
2141 Elts.assign(ATy->getNumElements(), Elt);
2143 llvm_unreachable("This code is out of sync with "
2144 " ConstantFoldLoadThroughGEPConstantExpr");
2147 assert(CI->getZExtValue() < ATy->getNumElements());
2148 Elts[CI->getZExtValue()] =
2149 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1, Context);
2150 return ConstantArray::get(ATy, Elts);
2154 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
2155 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
2156 static void CommitValueTo(Constant *Val, Constant *Addr,
2157 LLVMContext &Context) {
2158 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
2159 assert(GV->hasInitializer());
2160 GV->setInitializer(Val);
2164 ConstantExpr *CE = cast<ConstantExpr>(Addr);
2165 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2167 Constant *Init = GV->getInitializer();
2168 Init = EvaluateStoreInto(Init, Val, CE, 2, Context);
2169 GV->setInitializer(Init);
2172 /// ComputeLoadResult - Return the value that would be computed by a load from
2173 /// P after the stores reflected by 'memory' have been performed. If we can't
2174 /// decide, return null.
2175 static Constant *ComputeLoadResult(Constant *P,
2176 const DenseMap<Constant*, Constant*> &Memory,
2177 LLVMContext &Context) {
2178 // If this memory location has been recently stored, use the stored value: it
2179 // is the most up-to-date.
2180 DenseMap<Constant*, Constant*>::const_iterator I = Memory.find(P);
2181 if (I != Memory.end()) return I->second;
2184 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
2185 if (GV->hasDefinitiveInitializer())
2186 return GV->getInitializer();
2190 // Handle a constantexpr getelementptr.
2191 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
2192 if (CE->getOpcode() == Instruction::GetElementPtr &&
2193 isa<GlobalVariable>(CE->getOperand(0))) {
2194 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2195 if (GV->hasDefinitiveInitializer())
2196 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2199 return 0; // don't know how to evaluate.
2202 /// EvaluateFunction - Evaluate a call to function F, returning true if
2203 /// successful, false if we can't evaluate it. ActualArgs contains the formal
2204 /// arguments for the function.
2205 static bool EvaluateFunction(Function *F, Constant *&RetVal,
2206 const SmallVectorImpl<Constant*> &ActualArgs,
2207 std::vector<Function*> &CallStack,
2208 DenseMap<Constant*, Constant*> &MutatedMemory,
2209 std::vector<GlobalVariable*> &AllocaTmps) {
2210 // Check to see if this function is already executing (recursion). If so,
2211 // bail out. TODO: we might want to accept limited recursion.
2212 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
2215 LLVMContext &Context = F->getContext();
2217 CallStack.push_back(F);
2219 /// Values - As we compute SSA register values, we store their contents here.
2220 DenseMap<Value*, Constant*> Values;
2222 // Initialize arguments to the incoming values specified.
2224 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
2226 Values[AI] = ActualArgs[ArgNo];
2228 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
2229 /// we can only evaluate any one basic block at most once. This set keeps
2230 /// track of what we have executed so we can detect recursive cases etc.
2231 SmallPtrSet<BasicBlock*, 32> ExecutedBlocks;
2233 // CurInst - The current instruction we're evaluating.
2234 BasicBlock::iterator CurInst = F->begin()->begin();
2236 // This is the main evaluation loop.
2238 Constant *InstResult = 0;
2240 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
2241 if (SI->isVolatile()) return false; // no volatile accesses.
2242 Constant *Ptr = getVal(Values, SI->getOperand(1));
2243 if (!isSimpleEnoughPointerToCommit(Ptr, Context))
2244 // If this is too complex for us to commit, reject it.
2246 Constant *Val = getVal(Values, SI->getOperand(0));
2247 MutatedMemory[Ptr] = Val;
2248 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2249 InstResult = ConstantExpr::get(BO->getOpcode(),
2250 getVal(Values, BO->getOperand(0)),
2251 getVal(Values, BO->getOperand(1)));
2252 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2253 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2254 getVal(Values, CI->getOperand(0)),
2255 getVal(Values, CI->getOperand(1)));
2256 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2257 InstResult = ConstantExpr::getCast(CI->getOpcode(),
2258 getVal(Values, CI->getOperand(0)),
2260 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2262 ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
2263 getVal(Values, SI->getOperand(1)),
2264 getVal(Values, SI->getOperand(2)));
2265 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2266 Constant *P = getVal(Values, GEP->getOperand(0));
2267 SmallVector<Constant*, 8> GEPOps;
2268 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2270 GEPOps.push_back(getVal(Values, *i));
2271 InstResult = cast<GEPOperator>(GEP)->isInBounds() ?
2272 ConstantExpr::getInBoundsGetElementPtr(P, &GEPOps[0], GEPOps.size()) :
2273 ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size());
2274 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2275 if (LI->isVolatile()) return false; // no volatile accesses.
2276 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
2277 MutatedMemory, Context);
2278 if (InstResult == 0) return false; // Could not evaluate load.
2279 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2280 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs.
2281 const Type *Ty = AI->getType()->getElementType();
2282 AllocaTmps.push_back(new GlobalVariable(Context, Ty, false,
2283 GlobalValue::InternalLinkage,
2284 UndefValue::get(Ty),
2286 InstResult = AllocaTmps.back();
2287 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
2289 // Debug info can safely be ignored here.
2290 if (isa<DbgInfoIntrinsic>(CI)) {
2295 // Cannot handle inline asm.
2296 if (isa<InlineAsm>(CI->getOperand(0))) return false;
2298 // Resolve function pointers.
2299 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0)));
2300 if (!Callee) return false; // Cannot resolve.
2302 SmallVector<Constant*, 8> Formals;
2303 for (User::op_iterator i = CI->op_begin() + 1, e = CI->op_end();
2305 Formals.push_back(getVal(Values, *i));
2307 if (Callee->isDeclaration()) {
2308 // If this is a function we can constant fold, do it.
2309 if (Constant *C = ConstantFoldCall(Callee, Formals.data(),
2316 if (Callee->getFunctionType()->isVarArg())
2320 // Execute the call, if successful, use the return value.
2321 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
2322 MutatedMemory, AllocaTmps))
2324 InstResult = RetVal;
2326 } else if (isa<TerminatorInst>(CurInst)) {
2327 BasicBlock *NewBB = 0;
2328 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2329 if (BI->isUnconditional()) {
2330 NewBB = BI->getSuccessor(0);
2333 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
2334 if (!Cond) return false; // Cannot determine.
2336 NewBB = BI->getSuccessor(!Cond->getZExtValue());
2338 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2340 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
2341 if (!Val) return false; // Cannot determine.
2342 NewBB = SI->getSuccessor(SI->findCaseValue(Val));
2343 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(CurInst)) {
2344 Value *Val = getVal(Values, IBI->getAddress())->stripPointerCasts();
2345 if (BlockAddress *BA = dyn_cast<BlockAddress>(Val))
2346 NewBB = BA->getBasicBlock();
2348 return false; // Cannot determine.
2349 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
2350 if (RI->getNumOperands())
2351 RetVal = getVal(Values, RI->getOperand(0));
2353 CallStack.pop_back(); // return from fn.
2354 return true; // We succeeded at evaluating this ctor!
2356 // invoke, unwind, unreachable.
2357 return false; // Cannot handle this terminator.
2360 // Okay, we succeeded in evaluating this control flow. See if we have
2361 // executed the new block before. If so, we have a looping function,
2362 // which we cannot evaluate in reasonable time.
2363 if (!ExecutedBlocks.insert(NewBB))
2364 return false; // looped!
2366 // Okay, we have never been in this block before. Check to see if there
2367 // are any PHI nodes. If so, evaluate them with information about where
2369 BasicBlock *OldBB = CurInst->getParent();
2370 CurInst = NewBB->begin();
2372 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2373 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
2375 // Do NOT increment CurInst. We know that the terminator had no value.
2378 // Did not know how to evaluate this!
2382 if (!CurInst->use_empty())
2383 Values[CurInst] = InstResult;
2385 // Advance program counter.
2390 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2391 /// we can. Return true if we can, false otherwise.
2392 static bool EvaluateStaticConstructor(Function *F) {
2393 /// MutatedMemory - For each store we execute, we update this map. Loads
2394 /// check this to get the most up-to-date value. If evaluation is successful,
2395 /// this state is committed to the process.
2396 DenseMap<Constant*, Constant*> MutatedMemory;
2398 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2399 /// to represent its body. This vector is needed so we can delete the
2400 /// temporary globals when we are done.
2401 std::vector<GlobalVariable*> AllocaTmps;
2403 /// CallStack - This is used to detect recursion. In pathological situations
2404 /// we could hit exponential behavior, but at least there is nothing
2406 std::vector<Function*> CallStack;
2408 // Call the function.
2409 Constant *RetValDummy;
2410 bool EvalSuccess = EvaluateFunction(F, RetValDummy,
2411 SmallVector<Constant*, 0>(), CallStack,
2412 MutatedMemory, AllocaTmps);
2414 // We succeeded at evaluation: commit the result.
2415 DEBUG(errs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2416 << F->getName() << "' to " << MutatedMemory.size()
2418 for (DenseMap<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2419 E = MutatedMemory.end(); I != E; ++I)
2420 CommitValueTo(I->second, I->first, F->getContext());
2423 // At this point, we are done interpreting. If we created any 'alloca'
2424 // temporaries, release them now.
2425 while (!AllocaTmps.empty()) {
2426 GlobalVariable *Tmp = AllocaTmps.back();
2427 AllocaTmps.pop_back();
2429 // If there are still users of the alloca, the program is doing something
2430 // silly, e.g. storing the address of the alloca somewhere and using it
2431 // later. Since this is undefined, we'll just make it be null.
2432 if (!Tmp->use_empty())
2433 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2442 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2443 /// Return true if anything changed.
2444 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2445 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2446 bool MadeChange = false;
2447 if (Ctors.empty()) return false;
2449 // Loop over global ctors, optimizing them when we can.
2450 for (unsigned i = 0; i != Ctors.size(); ++i) {
2451 Function *F = Ctors[i];
2452 // Found a null terminator in the middle of the list, prune off the rest of
2455 if (i != Ctors.size()-1) {
2462 // We cannot simplify external ctor functions.
2463 if (F->empty()) continue;
2465 // If we can evaluate the ctor at compile time, do.
2466 if (EvaluateStaticConstructor(F)) {
2467 Ctors.erase(Ctors.begin()+i);
2470 ++NumCtorsEvaluated;
2475 if (!MadeChange) return false;
2477 GCL = InstallGlobalCtors(GCL, Ctors, GCL->getContext());
2481 bool GlobalOpt::OptimizeGlobalAliases(Module &M) {
2482 bool Changed = false;
2484 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end();
2486 Module::alias_iterator J = I++;
2487 // Aliases without names cannot be referenced outside this module.
2488 if (!J->hasName() && !J->isDeclaration())
2489 J->setLinkage(GlobalValue::InternalLinkage);
2490 // If the aliasee may change at link time, nothing can be done - bail out.
2491 if (J->mayBeOverridden())
2494 Constant *Aliasee = J->getAliasee();
2495 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2496 Target->removeDeadConstantUsers();
2497 bool hasOneUse = Target->hasOneUse() && Aliasee->hasOneUse();
2499 // Make all users of the alias use the aliasee instead.
2500 if (!J->use_empty()) {
2501 J->replaceAllUsesWith(Aliasee);
2502 ++NumAliasesResolved;
2506 // If the aliasee has internal linkage, give it the name and linkage
2507 // of the alias, and delete the alias. This turns:
2508 // define internal ... @f(...)
2509 // @a = alias ... @f
2511 // define ... @a(...)
2512 if (!Target->hasLocalLinkage())
2515 // The transform is only useful if the alias does not have internal linkage.
2516 if (J->hasLocalLinkage())
2519 // Do not perform the transform if multiple aliases potentially target the
2520 // aliasee. This check also ensures that it is safe to replace the section
2521 // and other attributes of the aliasee with those of the alias.
2525 // Give the aliasee the name, linkage and other attributes of the alias.
2526 Target->takeName(J);
2527 Target->setLinkage(J->getLinkage());
2528 Target->GlobalValue::copyAttributesFrom(J);
2530 // Delete the alias.
2531 M.getAliasList().erase(J);
2532 ++NumAliasesRemoved;
2539 bool GlobalOpt::runOnModule(Module &M) {
2540 bool Changed = false;
2542 // Try to find the llvm.globalctors list.
2543 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2545 bool LocalChange = true;
2546 while (LocalChange) {
2547 LocalChange = false;
2549 // Delete functions that are trivially dead, ccc -> fastcc
2550 LocalChange |= OptimizeFunctions(M);
2552 // Optimize global_ctors list.
2554 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2556 // Optimize non-address-taken globals.
2557 LocalChange |= OptimizeGlobalVars(M);
2559 // Resolve aliases, when possible.
2560 LocalChange |= OptimizeGlobalAliases(M);
2561 Changed |= LocalChange;
2564 // TODO: Move all global ctors functions to the end of the module for code