1 //===- GlobalOpt.cpp - Optimize Global Variables --------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass transforms simple global variables that never have their address
11 // taken. If obviously true, it marks read/write globals as constant, deletes
12 // variables only stored to, etc.
14 //===----------------------------------------------------------------------===//
16 #define DEBUG_TYPE "globalopt"
17 #include "llvm/Transforms/IPO.h"
18 #include "llvm/CallingConv.h"
19 #include "llvm/Constants.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/Instructions.h"
22 #include "llvm/IntrinsicInst.h"
23 #include "llvm/Module.h"
24 #include "llvm/Pass.h"
25 #include "llvm/Analysis/ConstantFolding.h"
26 #include "llvm/Target/TargetData.h"
27 #include "llvm/Support/CallSite.h"
28 #include "llvm/Support/Compiler.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/GetElementPtrTypeIterator.h"
31 #include "llvm/Support/MathExtras.h"
32 #include "llvm/ADT/DenseMap.h"
33 #include "llvm/ADT/SmallPtrSet.h"
34 #include "llvm/ADT/SmallVector.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/ADT/StringExtras.h"
40 STATISTIC(NumMarked , "Number of globals marked constant");
41 STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
42 STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
43 STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
44 STATISTIC(NumDeleted , "Number of globals deleted");
45 STATISTIC(NumFnDeleted , "Number of functions deleted");
46 STATISTIC(NumGlobUses , "Number of global uses devirtualized");
47 STATISTIC(NumLocalized , "Number of globals localized");
48 STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans");
49 STATISTIC(NumFastCallFns , "Number of functions converted to fastcc");
50 STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
51 STATISTIC(NumNestRemoved , "Number of nest attributes removed");
54 struct VISIBILITY_HIDDEN GlobalOpt : public ModulePass {
55 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
56 AU.addRequired<TargetData>();
58 static char ID; // Pass identification, replacement for typeid
59 GlobalOpt() : ModulePass(&ID) {}
61 bool runOnModule(Module &M);
64 GlobalVariable *FindGlobalCtors(Module &M);
65 bool OptimizeFunctions(Module &M);
66 bool OptimizeGlobalVars(Module &M);
67 bool ResolveAliases(Module &M);
68 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
69 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
73 char GlobalOpt::ID = 0;
74 static RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer");
76 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
80 /// GlobalStatus - As we analyze each global, keep track of some information
81 /// about it. If we find out that the address of the global is taken, none of
82 /// this info will be accurate.
83 struct VISIBILITY_HIDDEN GlobalStatus {
84 /// isLoaded - True if the global is ever loaded. If the global isn't ever
85 /// loaded it can be deleted.
88 /// StoredType - Keep track of what stores to the global look like.
91 /// NotStored - There is no store to this global. It can thus be marked
95 /// isInitializerStored - This global is stored to, but the only thing
96 /// stored is the constant it was initialized with. This is only tracked
97 /// for scalar globals.
100 /// isStoredOnce - This global is stored to, but only its initializer and
101 /// one other value is ever stored to it. If this global isStoredOnce, we
102 /// track the value stored to it in StoredOnceValue below. This is only
103 /// tracked for scalar globals.
106 /// isStored - This global is stored to by multiple values or something else
107 /// that we cannot track.
111 /// StoredOnceValue - If only one value (besides the initializer constant) is
112 /// ever stored to this global, keep track of what value it is.
113 Value *StoredOnceValue;
115 /// AccessingFunction/HasMultipleAccessingFunctions - These start out
116 /// null/false. When the first accessing function is noticed, it is recorded.
117 /// When a second different accessing function is noticed,
118 /// HasMultipleAccessingFunctions is set to true.
119 Function *AccessingFunction;
120 bool HasMultipleAccessingFunctions;
122 /// HasNonInstructionUser - Set to true if this global has a user that is not
123 /// an instruction (e.g. a constant expr or GV initializer).
124 bool HasNonInstructionUser;
126 /// HasPHIUser - Set to true if this global has a user that is a PHI node.
129 GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0),
130 AccessingFunction(0), HasMultipleAccessingFunctions(false),
131 HasNonInstructionUser(false), HasPHIUser(false) {}
136 /// ConstantIsDead - Return true if the specified constant is (transitively)
137 /// dead. The constant may be used by other constants (e.g. constant arrays and
138 /// constant exprs) as long as they are dead, but it cannot be used by anything
140 static bool ConstantIsDead(Constant *C) {
141 if (isa<GlobalValue>(C)) return false;
143 for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI)
144 if (Constant *CU = dyn_cast<Constant>(*UI)) {
145 if (!ConstantIsDead(CU)) return false;
152 /// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
153 /// structure. If the global has its address taken, return true to indicate we
154 /// can't do anything with it.
156 static bool AnalyzeGlobal(Value *V, GlobalStatus &GS,
157 SmallPtrSet<PHINode*, 16> &PHIUsers) {
158 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
159 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
160 GS.HasNonInstructionUser = true;
162 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
164 } else if (Instruction *I = dyn_cast<Instruction>(*UI)) {
165 if (!GS.HasMultipleAccessingFunctions) {
166 Function *F = I->getParent()->getParent();
167 if (GS.AccessingFunction == 0)
168 GS.AccessingFunction = F;
169 else if (GS.AccessingFunction != F)
170 GS.HasMultipleAccessingFunctions = true;
172 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
174 if (LI->isVolatile()) return true; // Don't hack on volatile loads.
175 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
176 // Don't allow a store OF the address, only stores TO the address.
177 if (SI->getOperand(0) == V) return true;
179 if (SI->isVolatile()) return true; // Don't hack on volatile stores.
181 // If this is a direct store to the global (i.e., the global is a scalar
182 // value, not an aggregate), keep more specific information about
184 if (GS.StoredType != GlobalStatus::isStored) {
185 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){
186 Value *StoredVal = SI->getOperand(0);
187 if (StoredVal == GV->getInitializer()) {
188 if (GS.StoredType < GlobalStatus::isInitializerStored)
189 GS.StoredType = GlobalStatus::isInitializerStored;
190 } else if (isa<LoadInst>(StoredVal) &&
191 cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
193 if (GS.StoredType < GlobalStatus::isInitializerStored)
194 GS.StoredType = GlobalStatus::isInitializerStored;
195 } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
196 GS.StoredType = GlobalStatus::isStoredOnce;
197 GS.StoredOnceValue = StoredVal;
198 } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
199 GS.StoredOnceValue == StoredVal) {
202 GS.StoredType = GlobalStatus::isStored;
205 GS.StoredType = GlobalStatus::isStored;
208 } else if (isa<GetElementPtrInst>(I)) {
209 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
210 } else if (isa<SelectInst>(I)) {
211 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
212 } else if (PHINode *PN = dyn_cast<PHINode>(I)) {
213 // PHI nodes we can check just like select or GEP instructions, but we
214 // have to be careful about infinite recursion.
215 if (PHIUsers.insert(PN)) // Not already visited.
216 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
217 GS.HasPHIUser = true;
218 } else if (isa<CmpInst>(I)) {
219 } else if (isa<MemCpyInst>(I) || isa<MemMoveInst>(I)) {
220 if (I->getOperand(1) == V)
221 GS.StoredType = GlobalStatus::isStored;
222 if (I->getOperand(2) == V)
224 } else if (isa<MemSetInst>(I)) {
225 assert(I->getOperand(1) == V && "Memset only takes one pointer!");
226 GS.StoredType = GlobalStatus::isStored;
228 return true; // Any other non-load instruction might take address!
230 } else if (Constant *C = dyn_cast<Constant>(*UI)) {
231 GS.HasNonInstructionUser = true;
232 // We might have a dead and dangling constant hanging off of here.
233 if (!ConstantIsDead(C))
236 GS.HasNonInstructionUser = true;
237 // Otherwise must be some other user.
244 static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) {
245 ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
247 unsigned IdxV = CI->getZExtValue();
249 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
250 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
251 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
252 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
253 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) {
254 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
255 } else if (isa<ConstantAggregateZero>(Agg)) {
256 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
257 if (IdxV < STy->getNumElements())
258 return Constant::getNullValue(STy->getElementType(IdxV));
259 } else if (const SequentialType *STy =
260 dyn_cast<SequentialType>(Agg->getType())) {
261 return Constant::getNullValue(STy->getElementType());
263 } else if (isa<UndefValue>(Agg)) {
264 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
265 if (IdxV < STy->getNumElements())
266 return UndefValue::get(STy->getElementType(IdxV));
267 } else if (const SequentialType *STy =
268 dyn_cast<SequentialType>(Agg->getType())) {
269 return UndefValue::get(STy->getElementType());
276 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
277 /// users of the global, cleaning up the obvious ones. This is largely just a
278 /// quick scan over the use list to clean up the easy and obvious cruft. This
279 /// returns true if it made a change.
280 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
281 bool Changed = false;
282 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
285 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
287 // Replace the load with the initializer.
288 LI->replaceAllUsesWith(Init);
289 LI->eraseFromParent();
292 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
293 // Store must be unreachable or storing Init into the global.
294 SI->eraseFromParent();
296 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
297 if (CE->getOpcode() == Instruction::GetElementPtr) {
298 Constant *SubInit = 0;
300 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
301 Changed |= CleanupConstantGlobalUsers(CE, SubInit);
302 } else if (CE->getOpcode() == Instruction::BitCast &&
303 isa<PointerType>(CE->getType())) {
304 // Pointer cast, delete any stores and memsets to the global.
305 Changed |= CleanupConstantGlobalUsers(CE, 0);
308 if (CE->use_empty()) {
309 CE->destroyConstant();
312 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
313 // Do not transform "gepinst (gep constexpr (GV))" here, because forming
314 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
315 // and will invalidate our notion of what Init is.
316 Constant *SubInit = 0;
317 if (!isa<ConstantExpr>(GEP->getOperand(0))) {
319 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
320 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
321 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
323 Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
325 if (GEP->use_empty()) {
326 GEP->eraseFromParent();
329 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
330 if (MI->getRawDest() == V) {
331 MI->eraseFromParent();
335 } else if (Constant *C = dyn_cast<Constant>(U)) {
336 // If we have a chain of dead constantexprs or other things dangling from
337 // us, and if they are all dead, nuke them without remorse.
338 if (ConstantIsDead(C)) {
339 C->destroyConstant();
340 // This could have invalidated UI, start over from scratch.
341 CleanupConstantGlobalUsers(V, Init);
349 /// isSafeSROAElementUse - Return true if the specified instruction is a safe
350 /// user of a derived expression from a global that we want to SROA.
351 static bool isSafeSROAElementUse(Value *V) {
352 // We might have a dead and dangling constant hanging off of here.
353 if (Constant *C = dyn_cast<Constant>(V))
354 return ConstantIsDead(C);
356 Instruction *I = dyn_cast<Instruction>(V);
357 if (!I) return false;
360 if (isa<LoadInst>(I)) return true;
362 // Stores *to* the pointer are ok.
363 if (StoreInst *SI = dyn_cast<StoreInst>(I))
364 return SI->getOperand(0) != V;
366 // Otherwise, it must be a GEP.
367 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I);
368 if (GEPI == 0) return false;
370 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
371 !cast<Constant>(GEPI->getOperand(1))->isNullValue())
374 for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end();
376 if (!isSafeSROAElementUse(*I))
382 /// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value.
383 /// Look at it and its uses and decide whether it is safe to SROA this global.
385 static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
386 // The user of the global must be a GEP Inst or a ConstantExpr GEP.
387 if (!isa<GetElementPtrInst>(U) &&
388 (!isa<ConstantExpr>(U) ||
389 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
392 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
393 // don't like < 3 operand CE's, and we don't like non-constant integer
394 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some
396 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
397 !cast<Constant>(U->getOperand(1))->isNullValue() ||
398 !isa<ConstantInt>(U->getOperand(2)))
401 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
402 ++GEPI; // Skip over the pointer index.
404 // If this is a use of an array allocation, do a bit more checking for sanity.
405 if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
406 uint64_t NumElements = AT->getNumElements();
407 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
409 // Check to make sure that index falls within the array. If not,
410 // something funny is going on, so we won't do the optimization.
412 if (Idx->getZExtValue() >= NumElements)
415 // We cannot scalar repl this level of the array unless any array
416 // sub-indices are in-range constants. In particular, consider:
417 // A[0][i]. We cannot know that the user isn't doing invalid things like
418 // allowing i to index an out-of-range subscript that accesses A[1].
420 // Scalar replacing *just* the outer index of the array is probably not
421 // going to be a win anyway, so just give up.
422 for (++GEPI; // Skip array index.
423 GEPI != E && (isa<ArrayType>(*GEPI) || isa<VectorType>(*GEPI));
425 uint64_t NumElements;
426 if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
427 NumElements = SubArrayTy->getNumElements();
429 NumElements = cast<VectorType>(*GEPI)->getNumElements();
431 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
432 if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
437 for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I)
438 if (!isSafeSROAElementUse(*I))
443 /// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
444 /// is safe for us to perform this transformation.
446 static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
447 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
449 if (!IsUserOfGlobalSafeForSRA(*UI, GV))
456 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
457 /// variable. This opens the door for other optimizations by exposing the
458 /// behavior of the program in a more fine-grained way. We have determined that
459 /// this transformation is safe already. We return the first global variable we
460 /// insert so that the caller can reprocess it.
461 static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) {
462 // Make sure this global only has simple uses that we can SRA.
463 if (!GlobalUsersSafeToSRA(GV))
466 assert(GV->hasInternalLinkage() && !GV->isConstant());
467 Constant *Init = GV->getInitializer();
468 const Type *Ty = Init->getType();
470 std::vector<GlobalVariable*> NewGlobals;
471 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
473 // Get the alignment of the global, either explicit or target-specific.
474 unsigned StartAlignment = GV->getAlignment();
475 if (StartAlignment == 0)
476 StartAlignment = TD.getABITypeAlignment(GV->getType());
478 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
479 NewGlobals.reserve(STy->getNumElements());
480 const StructLayout &Layout = *TD.getStructLayout(STy);
481 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
482 Constant *In = getAggregateConstantElement(Init,
483 ConstantInt::get(Type::Int32Ty, i));
484 assert(In && "Couldn't get element of initializer?");
485 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
486 GlobalVariable::InternalLinkage,
487 In, GV->getName()+"."+utostr(i),
490 GV->getType()->getAddressSpace());
491 Globals.insert(GV, NGV);
492 NewGlobals.push_back(NGV);
494 // Calculate the known alignment of the field. If the original aggregate
495 // had 256 byte alignment for example, something might depend on that:
496 // propagate info to each field.
497 uint64_t FieldOffset = Layout.getElementOffset(i);
498 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset);
499 if (NewAlign > TD.getABITypeAlignment(STy->getElementType(i)))
500 NGV->setAlignment(NewAlign);
502 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
503 unsigned NumElements = 0;
504 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
505 NumElements = ATy->getNumElements();
507 NumElements = cast<VectorType>(STy)->getNumElements();
509 if (NumElements > 16 && GV->hasNUsesOrMore(16))
510 return 0; // It's not worth it.
511 NewGlobals.reserve(NumElements);
513 uint64_t EltSize = TD.getABITypeSize(STy->getElementType());
514 unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType());
515 for (unsigned i = 0, e = NumElements; i != e; ++i) {
516 Constant *In = getAggregateConstantElement(Init,
517 ConstantInt::get(Type::Int32Ty, i));
518 assert(In && "Couldn't get element of initializer?");
520 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
521 GlobalVariable::InternalLinkage,
522 In, GV->getName()+"."+utostr(i),
525 GV->getType()->getAddressSpace());
526 Globals.insert(GV, NGV);
527 NewGlobals.push_back(NGV);
529 // Calculate the known alignment of the field. If the original aggregate
530 // had 256 byte alignment for example, something might depend on that:
531 // propagate info to each field.
532 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
533 if (NewAlign > EltAlign)
534 NGV->setAlignment(NewAlign);
538 if (NewGlobals.empty())
541 DOUT << "PERFORMING GLOBAL SRA ON: " << *GV;
543 Constant *NullInt = Constant::getNullValue(Type::Int32Ty);
545 // Loop over all of the uses of the global, replacing the constantexpr geps,
546 // with smaller constantexpr geps or direct references.
547 while (!GV->use_empty()) {
548 User *GEP = GV->use_back();
549 assert(((isa<ConstantExpr>(GEP) &&
550 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
551 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
553 // Ignore the 1th operand, which has to be zero or else the program is quite
554 // broken (undefined). Get the 2nd operand, which is the structure or array
556 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
557 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
559 Value *NewPtr = NewGlobals[Val];
561 // Form a shorter GEP if needed.
562 if (GEP->getNumOperands() > 3) {
563 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
564 SmallVector<Constant*, 8> Idxs;
565 Idxs.push_back(NullInt);
566 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
567 Idxs.push_back(CE->getOperand(i));
568 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr),
569 &Idxs[0], Idxs.size());
571 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
572 SmallVector<Value*, 8> Idxs;
573 Idxs.push_back(NullInt);
574 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
575 Idxs.push_back(GEPI->getOperand(i));
576 NewPtr = GetElementPtrInst::Create(NewPtr, Idxs.begin(), Idxs.end(),
577 GEPI->getName()+"."+utostr(Val), GEPI);
580 GEP->replaceAllUsesWith(NewPtr);
582 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
583 GEPI->eraseFromParent();
585 cast<ConstantExpr>(GEP)->destroyConstant();
588 // Delete the old global, now that it is dead.
592 // Loop over the new globals array deleting any globals that are obviously
593 // dead. This can arise due to scalarization of a structure or an array that
594 // has elements that are dead.
595 unsigned FirstGlobal = 0;
596 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
597 if (NewGlobals[i]->use_empty()) {
598 Globals.erase(NewGlobals[i]);
599 if (FirstGlobal == i) ++FirstGlobal;
602 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
605 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
606 /// value will trap if the value is dynamically null. PHIs keeps track of any
607 /// phi nodes we've seen to avoid reprocessing them.
608 static bool AllUsesOfValueWillTrapIfNull(Value *V,
609 SmallPtrSet<PHINode*, 8> &PHIs) {
610 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
611 if (isa<LoadInst>(*UI)) {
613 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
614 if (SI->getOperand(0) == V) {
615 //cerr << "NONTRAPPING USE: " << **UI;
616 return false; // Storing the value.
618 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
619 if (CI->getOperand(0) != V) {
620 //cerr << "NONTRAPPING USE: " << **UI;
621 return false; // Not calling the ptr
623 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
624 if (II->getOperand(0) != V) {
625 //cerr << "NONTRAPPING USE: " << **UI;
626 return false; // Not calling the ptr
628 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(*UI)) {
629 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
630 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
631 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
632 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
633 // If we've already seen this phi node, ignore it, it has already been
636 return AllUsesOfValueWillTrapIfNull(PN, PHIs);
637 } else if (isa<ICmpInst>(*UI) &&
638 isa<ConstantPointerNull>(UI->getOperand(1))) {
639 // Ignore setcc X, null
641 //cerr << "NONTRAPPING USE: " << **UI;
647 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
648 /// from GV will trap if the loaded value is null. Note that this also permits
649 /// comparisons of the loaded value against null, as a special case.
650 static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) {
651 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI)
652 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
653 SmallPtrSet<PHINode*, 8> PHIs;
654 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
656 } else if (isa<StoreInst>(*UI)) {
657 // Ignore stores to the global.
659 // We don't know or understand this user, bail out.
660 //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI;
667 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
668 bool Changed = false;
669 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
670 Instruction *I = cast<Instruction>(*UI++);
671 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
672 LI->setOperand(0, NewV);
674 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
675 if (SI->getOperand(1) == V) {
676 SI->setOperand(1, NewV);
679 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
680 if (I->getOperand(0) == V) {
681 // Calling through the pointer! Turn into a direct call, but be careful
682 // that the pointer is not also being passed as an argument.
683 I->setOperand(0, NewV);
685 bool PassedAsArg = false;
686 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
687 if (I->getOperand(i) == V) {
689 I->setOperand(i, NewV);
693 // Being passed as an argument also. Be careful to not invalidate UI!
697 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
698 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
699 ConstantExpr::getCast(CI->getOpcode(),
700 NewV, CI->getType()));
701 if (CI->use_empty()) {
703 CI->eraseFromParent();
705 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
706 // Should handle GEP here.
707 SmallVector<Constant*, 8> Idxs;
708 Idxs.reserve(GEPI->getNumOperands()-1);
709 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
711 if (Constant *C = dyn_cast<Constant>(*i))
715 if (Idxs.size() == GEPI->getNumOperands()-1)
716 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
717 ConstantExpr::getGetElementPtr(NewV, &Idxs[0],
719 if (GEPI->use_empty()) {
721 GEPI->eraseFromParent();
730 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
731 /// value stored into it. If there are uses of the loaded value that would trap
732 /// if the loaded value is dynamically null, then we know that they cannot be
733 /// reachable with a null optimize away the load.
734 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) {
735 std::vector<LoadInst*> Loads;
736 bool Changed = false;
738 // Replace all uses of loads with uses of uses of the stored value.
739 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end();
741 if (LoadInst *LI = dyn_cast<LoadInst>(*GUI)) {
743 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
745 // If we get here we could have stores, selects, or phi nodes whose values
747 assert((isa<StoreInst>(*GUI) || isa<PHINode>(*GUI) ||
748 isa<SelectInst>(*GUI) || isa<ConstantExpr>(*GUI)) &&
749 "Only expect load and stores!");
753 DOUT << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV;
757 // Delete all of the loads we can, keeping track of whether we nuked them all!
758 bool AllLoadsGone = true;
759 while (!Loads.empty()) {
760 LoadInst *L = Loads.back();
761 if (L->use_empty()) {
762 L->eraseFromParent();
765 AllLoadsGone = false;
770 // If we nuked all of the loads, then none of the stores are needed either,
771 // nor is the global.
773 DOUT << " *** GLOBAL NOW DEAD!\n";
774 CleanupConstantGlobalUsers(GV, 0);
775 if (GV->use_empty()) {
776 GV->eraseFromParent();
784 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
785 /// instructions that are foldable.
786 static void ConstantPropUsersOf(Value *V) {
787 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
788 if (Instruction *I = dyn_cast<Instruction>(*UI++))
789 if (Constant *NewC = ConstantFoldInstruction(I)) {
790 I->replaceAllUsesWith(NewC);
792 // Advance UI to the next non-I use to avoid invalidating it!
793 // Instructions could multiply use V.
794 while (UI != E && *UI == I)
796 I->eraseFromParent();
800 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
801 /// variable, and transforms the program as if it always contained the result of
802 /// the specified malloc. Because it is always the result of the specified
803 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
804 /// malloc into a global, and any loads of GV as uses of the new global.
805 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
807 DOUT << "PROMOTING MALLOC GLOBAL: " << *GV << " MALLOC = " << *MI;
808 ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize());
810 if (NElements->getZExtValue() != 1) {
811 // If we have an array allocation, transform it to a single element
812 // allocation to make the code below simpler.
813 Type *NewTy = ArrayType::get(MI->getAllocatedType(),
814 NElements->getZExtValue());
816 new MallocInst(NewTy, Constant::getNullValue(Type::Int32Ty),
817 MI->getAlignment(), MI->getName(), MI);
819 Indices[0] = Indices[1] = Constant::getNullValue(Type::Int32Ty);
820 Value *NewGEP = GetElementPtrInst::Create(NewMI, Indices, Indices + 2,
821 NewMI->getName()+".el0", MI);
822 MI->replaceAllUsesWith(NewGEP);
823 MI->eraseFromParent();
827 // Create the new global variable. The contents of the malloc'd memory is
828 // undefined, so initialize with an undef value.
829 Constant *Init = UndefValue::get(MI->getAllocatedType());
830 GlobalVariable *NewGV = new GlobalVariable(MI->getAllocatedType(), false,
831 GlobalValue::InternalLinkage, Init,
832 GV->getName()+".body",
834 GV->isThreadLocal());
835 // FIXME: This new global should have the alignment returned by malloc. Code
836 // could depend on malloc returning large alignment (on the mac, 16 bytes) but
837 // this would only guarantee some lower alignment.
838 GV->getParent()->getGlobalList().insert(GV, NewGV);
840 // Anything that used the malloc now uses the global directly.
841 MI->replaceAllUsesWith(NewGV);
843 Constant *RepValue = NewGV;
844 if (NewGV->getType() != GV->getType()->getElementType())
845 RepValue = ConstantExpr::getBitCast(RepValue,
846 GV->getType()->getElementType());
848 // If there is a comparison against null, we will insert a global bool to
849 // keep track of whether the global was initialized yet or not.
850 GlobalVariable *InitBool =
851 new GlobalVariable(Type::Int1Ty, false, GlobalValue::InternalLinkage,
852 ConstantInt::getFalse(), GV->getName()+".init",
853 (Module *)NULL, GV->isThreadLocal());
854 bool InitBoolUsed = false;
856 // Loop over all uses of GV, processing them in turn.
857 std::vector<StoreInst*> Stores;
858 while (!GV->use_empty())
859 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
860 while (!LI->use_empty()) {
861 Use &LoadUse = LI->use_begin().getUse();
862 if (!isa<ICmpInst>(LoadUse.getUser()))
865 ICmpInst *CI = cast<ICmpInst>(LoadUse.getUser());
866 // Replace the cmp X, 0 with a use of the bool value.
867 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", CI);
869 switch (CI->getPredicate()) {
870 default: assert(0 && "Unknown ICmp Predicate!");
871 case ICmpInst::ICMP_ULT:
872 case ICmpInst::ICMP_SLT:
873 LV = ConstantInt::getFalse(); // X < null -> always false
875 case ICmpInst::ICMP_ULE:
876 case ICmpInst::ICMP_SLE:
877 case ICmpInst::ICMP_EQ:
878 LV = BinaryOperator::CreateNot(LV, "notinit", CI);
880 case ICmpInst::ICMP_NE:
881 case ICmpInst::ICMP_UGE:
882 case ICmpInst::ICMP_SGE:
883 case ICmpInst::ICMP_UGT:
884 case ICmpInst::ICMP_SGT:
887 CI->replaceAllUsesWith(LV);
888 CI->eraseFromParent();
891 LI->eraseFromParent();
893 StoreInst *SI = cast<StoreInst>(GV->use_back());
894 // The global is initialized when the store to it occurs.
895 new StoreInst(ConstantInt::getTrue(), InitBool, SI);
896 SI->eraseFromParent();
899 // If the initialization boolean was used, insert it, otherwise delete it.
901 while (!InitBool->use_empty()) // Delete initializations
902 cast<Instruction>(InitBool->use_back())->eraseFromParent();
905 GV->getParent()->getGlobalList().insert(GV, InitBool);
908 // Now the GV is dead, nuke it and the malloc.
909 GV->eraseFromParent();
910 MI->eraseFromParent();
912 // To further other optimizations, loop over all users of NewGV and try to
913 // constant prop them. This will promote GEP instructions with constant
914 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
915 ConstantPropUsersOf(NewGV);
916 if (RepValue != NewGV)
917 ConstantPropUsersOf(RepValue);
922 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
923 /// to make sure that there are no complex uses of V. We permit simple things
924 /// like dereferencing the pointer, but not storing through the address, unless
925 /// it is to the specified global.
926 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V,
928 SmallPtrSet<PHINode*, 8> &PHIs) {
929 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
930 Instruction *Inst = dyn_cast<Instruction>(*UI);
931 if (Inst == 0) return false;
933 if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) {
934 continue; // Fine, ignore.
937 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
938 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
939 return false; // Storing the pointer itself... bad.
940 continue; // Otherwise, storing through it, or storing into GV... fine.
943 if (isa<GetElementPtrInst>(Inst)) {
944 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs))
949 if (PHINode *PN = dyn_cast<PHINode>(Inst)) {
950 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI
953 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
958 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) {
959 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs))
969 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
970 /// somewhere. Transform all uses of the allocation into loads from the
971 /// global and uses of the resultant pointer. Further, delete the store into
972 /// GV. This assumes that these value pass the
973 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
974 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
975 GlobalVariable *GV) {
976 while (!Alloc->use_empty()) {
977 Instruction *U = cast<Instruction>(*Alloc->use_begin());
978 Instruction *InsertPt = U;
979 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
980 // If this is the store of the allocation into the global, remove it.
981 if (SI->getOperand(1) == GV) {
982 SI->eraseFromParent();
985 } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
986 // Insert the load in the corresponding predecessor, not right before the
988 unsigned PredNo = Alloc->use_begin().getOperandNo()/2;
989 InsertPt = PN->getIncomingBlock(PredNo)->getTerminator();
990 } else if (isa<BitCastInst>(U)) {
991 // Must be bitcast between the malloc and store to initialize the global.
992 ReplaceUsesOfMallocWithGlobal(U, GV);
993 U->eraseFromParent();
995 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
996 // If this is a "GEP bitcast" and the user is a store to the global, then
997 // just process it as a bitcast.
998 if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse())
999 if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->use_back()))
1000 if (SI->getOperand(1) == GV) {
1001 // Must be bitcast GEP between the malloc and store to initialize
1003 ReplaceUsesOfMallocWithGlobal(GEPI, GV);
1004 GEPI->eraseFromParent();
1009 // Insert a load from the global, and use it instead of the malloc.
1010 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
1011 U->replaceUsesOfWith(Alloc, NL);
1015 /// LoadUsesSimpleEnoughForHeapSRA - Verify that all uses of V (a load, or a phi
1016 /// of a load) are simple enough to perform heap SRA on. This permits GEP's
1017 /// that index through the array and struct field, icmps of null, and PHIs.
1018 static bool LoadUsesSimpleEnoughForHeapSRA(Value *V, MallocInst *MI,
1019 SmallPtrSet<PHINode*, 32> &AnalyzedLoadUsingPHIs) {
1020 // We permit two users of the load: setcc comparing against the null
1021 // pointer, and a getelementptr of a specific form.
1022 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
1023 Instruction *User = cast<Instruction>(*UI);
1025 // Comparison against null is ok.
1026 if (ICmpInst *ICI = dyn_cast<ICmpInst>(User)) {
1027 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1032 // getelementptr is also ok, but only a simple form.
1033 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1034 // Must index into the array and into the struct.
1035 if (GEPI->getNumOperands() < 3)
1038 // Otherwise the GEP is ok.
1042 if (PHINode *PN = dyn_cast<PHINode>(User)) {
1043 // If we have already recursively analyzed this PHI, then it is safe.
1044 if (AnalyzedLoadUsingPHIs.insert(PN))
1047 // We have a phi of a load from the global. We can only handle this
1048 // if the other PHI'd values are actually the same. In this case,
1049 // the rewriter will just drop the phi entirely.
1050 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1051 Value *IV = PN->getIncomingValue(i);
1052 if (IV == V) continue; // Trivial the same.
1054 // If the phi'd value is from the malloc that initializes the value,
1056 if (IV == MI) continue;
1058 // Otherwise, we don't know what it is.
1062 if (!LoadUsesSimpleEnoughForHeapSRA(PN, MI, AnalyzedLoadUsingPHIs))
1068 // Otherwise we don't know what this is, not ok.
1076 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
1077 /// GV are simple enough to perform HeapSRA, return true.
1078 static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV,
1080 SmallPtrSet<PHINode*, 32> AnalyzedLoadUsingPHIs;
1081 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;
1083 if (LoadInst *LI = dyn_cast<LoadInst>(*UI))
1084 if (!LoadUsesSimpleEnoughForHeapSRA(LI, MI, AnalyzedLoadUsingPHIs))
1089 /// GetHeapSROALoad - Return the load for the specified field of the HeapSROA'd
1090 /// value, lazily creating it on demand.
1091 static Value *GetHeapSROALoad(Instruction *Load, unsigned FieldNo,
1092 const std::vector<GlobalVariable*> &FieldGlobals,
1093 std::vector<Value *> &InsertedLoadsForPtr) {
1094 if (InsertedLoadsForPtr.size() <= FieldNo)
1095 InsertedLoadsForPtr.resize(FieldNo+1);
1096 if (InsertedLoadsForPtr[FieldNo] == 0)
1097 InsertedLoadsForPtr[FieldNo] = new LoadInst(FieldGlobals[FieldNo],
1098 Load->getName()+".f" +
1099 utostr(FieldNo), Load);
1100 return InsertedLoadsForPtr[FieldNo];
1103 /// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1104 /// the load, rewrite the derived value to use the HeapSRoA'd load.
1105 static void RewriteHeapSROALoadUser(LoadInst *Load, Instruction *LoadUser,
1106 const std::vector<GlobalVariable*> &FieldGlobals,
1107 std::vector<Value *> &InsertedLoadsForPtr) {
1108 // If this is a comparison against null, handle it.
1109 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1110 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1111 // If we have a setcc of the loaded pointer, we can use a setcc of any
1114 if (InsertedLoadsForPtr.empty()) {
1115 NPtr = GetHeapSROALoad(Load, 0, FieldGlobals, InsertedLoadsForPtr);
1117 NPtr = InsertedLoadsForPtr.back();
1120 Value *New = new ICmpInst(SCI->getPredicate(), NPtr,
1121 Constant::getNullValue(NPtr->getType()),
1122 SCI->getName(), SCI);
1123 SCI->replaceAllUsesWith(New);
1124 SCI->eraseFromParent();
1128 // Handle 'getelementptr Ptr, Idx, uint FieldNo ...'
1129 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1130 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1131 && "Unexpected GEPI!");
1133 // Load the pointer for this field.
1134 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1135 Value *NewPtr = GetHeapSROALoad(Load, FieldNo,
1136 FieldGlobals, InsertedLoadsForPtr);
1138 // Create the new GEP idx vector.
1139 SmallVector<Value*, 8> GEPIdx;
1140 GEPIdx.push_back(GEPI->getOperand(1));
1141 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1143 Value *NGEPI = GetElementPtrInst::Create(NewPtr,
1144 GEPIdx.begin(), GEPIdx.end(),
1145 GEPI->getName(), GEPI);
1146 GEPI->replaceAllUsesWith(NGEPI);
1147 GEPI->eraseFromParent();
1151 // Handle PHI nodes. PHI nodes must be merging in the same values, plus
1152 // potentially the original malloc. Insert phi nodes for each field, then
1153 // process uses of the PHI.
1154 PHINode *PN = cast<PHINode>(LoadUser);
1155 std::vector<Value *> PHIsForField;
1156 PHIsForField.resize(FieldGlobals.size());
1157 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1158 Value *LoadV = GetHeapSROALoad(Load, i, FieldGlobals, InsertedLoadsForPtr);
1160 PHINode *FieldPN = PHINode::Create(LoadV->getType(),
1161 PN->getName()+"."+utostr(i), PN);
1162 // Fill in the predecessor values.
1163 for (unsigned pred = 0, e = PN->getNumIncomingValues(); pred != e; ++pred) {
1164 // Each predecessor either uses the load or the original malloc.
1165 Value *InVal = PN->getIncomingValue(pred);
1166 BasicBlock *BB = PN->getIncomingBlock(pred);
1168 if (isa<MallocInst>(InVal)) {
1169 // Insert a reload from the global in the predecessor.
1170 NewVal = GetHeapSROALoad(BB->getTerminator(), i, FieldGlobals,
1173 NewVal = InsertedLoadsForPtr[i];
1175 FieldPN->addIncoming(NewVal, BB);
1177 PHIsForField[i] = FieldPN;
1180 // Since PHIsForField specifies a phi for every input value, the lazy inserter
1181 // will never insert a load.
1182 while (!PN->use_empty())
1183 RewriteHeapSROALoadUser(Load, PN->use_back(), FieldGlobals, PHIsForField);
1184 PN->eraseFromParent();
1187 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1188 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1189 /// use FieldGlobals instead. All uses of loaded values satisfy
1190 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA.
1191 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1192 const std::vector<GlobalVariable*> &FieldGlobals) {
1193 std::vector<Value *> InsertedLoadsForPtr;
1194 //InsertedLoadsForPtr.resize(FieldGlobals.size());
1195 while (!Load->use_empty())
1196 RewriteHeapSROALoadUser(Load, Load->use_back(),
1197 FieldGlobals, InsertedLoadsForPtr);
1200 /// PerformHeapAllocSRoA - MI is an allocation of an array of structures. Break
1201 /// it up into multiple allocations of arrays of the fields.
1202 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){
1203 DOUT << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *MI;
1204 const StructType *STy = cast<StructType>(MI->getAllocatedType());
1206 // There is guaranteed to be at least one use of the malloc (storing
1207 // it into GV). If there are other uses, change them to be uses of
1208 // the global to simplify later code. This also deletes the store
1210 ReplaceUsesOfMallocWithGlobal(MI, GV);
1212 // Okay, at this point, there are no users of the malloc. Insert N
1213 // new mallocs at the same place as MI, and N globals.
1214 std::vector<GlobalVariable*> FieldGlobals;
1215 std::vector<MallocInst*> FieldMallocs;
1217 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1218 const Type *FieldTy = STy->getElementType(FieldNo);
1219 const Type *PFieldTy = PointerType::getUnqual(FieldTy);
1221 GlobalVariable *NGV =
1222 new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage,
1223 Constant::getNullValue(PFieldTy),
1224 GV->getName() + ".f" + utostr(FieldNo), GV,
1225 GV->isThreadLocal());
1226 FieldGlobals.push_back(NGV);
1228 MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(),
1229 MI->getName() + ".f" + utostr(FieldNo),MI);
1230 FieldMallocs.push_back(NMI);
1231 new StoreInst(NMI, NGV, MI);
1234 // The tricky aspect of this transformation is handling the case when malloc
1235 // fails. In the original code, malloc failing would set the result pointer
1236 // of malloc to null. In this case, some mallocs could succeed and others
1237 // could fail. As such, we emit code that looks like this:
1238 // F0 = malloc(field0)
1239 // F1 = malloc(field1)
1240 // F2 = malloc(field2)
1241 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1242 // if (F0) { free(F0); F0 = 0; }
1243 // if (F1) { free(F1); F1 = 0; }
1244 // if (F2) { free(F2); F2 = 0; }
1246 Value *RunningOr = 0;
1247 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1248 Value *Cond = new ICmpInst(ICmpInst::ICMP_EQ, FieldMallocs[i],
1249 Constant::getNullValue(FieldMallocs[i]->getType()),
1252 RunningOr = Cond; // First seteq
1254 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", MI);
1257 // Split the basic block at the old malloc.
1258 BasicBlock *OrigBB = MI->getParent();
1259 BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont");
1261 // Create the block to check the first condition. Put all these blocks at the
1262 // end of the function as they are unlikely to be executed.
1263 BasicBlock *NullPtrBlock = BasicBlock::Create("malloc_ret_null",
1264 OrigBB->getParent());
1266 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1267 // branch on RunningOr.
1268 OrigBB->getTerminator()->eraseFromParent();
1269 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1271 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1272 // pointer, because some may be null while others are not.
1273 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1274 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1275 Value *Cmp = new ICmpInst(ICmpInst::ICMP_NE, GVVal,
1276 Constant::getNullValue(GVVal->getType()),
1277 "tmp", NullPtrBlock);
1278 BasicBlock *FreeBlock = BasicBlock::Create("free_it", OrigBB->getParent());
1279 BasicBlock *NextBlock = BasicBlock::Create("next", OrigBB->getParent());
1280 BranchInst::Create(FreeBlock, NextBlock, Cmp, NullPtrBlock);
1282 // Fill in FreeBlock.
1283 new FreeInst(GVVal, FreeBlock);
1284 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1286 BranchInst::Create(NextBlock, FreeBlock);
1288 NullPtrBlock = NextBlock;
1291 BranchInst::Create(ContBB, NullPtrBlock);
1293 // MI is no longer needed, remove it.
1294 MI->eraseFromParent();
1297 // Okay, the malloc site is completely handled. All of the uses of GV are now
1298 // loads, and all uses of those loads are simple. Rewrite them to use loads
1299 // of the per-field globals instead.
1300 while (!GV->use_empty()) {
1301 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
1302 RewriteUsesOfLoadForHeapSRoA(LI, FieldGlobals);
1303 LI->eraseFromParent();
1305 // Must be a store of null.
1306 StoreInst *SI = cast<StoreInst>(GV->use_back());
1307 assert(isa<Constant>(SI->getOperand(0)) &&
1308 cast<Constant>(SI->getOperand(0))->isNullValue() &&
1309 "Unexpected heap-sra user!");
1311 // Insert a store of null into each global.
1312 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1314 Constant::getNullValue(FieldGlobals[i]->getType()->getElementType());
1315 new StoreInst(Null, FieldGlobals[i], SI);
1317 // Erase the original store.
1318 SI->eraseFromParent();
1322 // The old global is now dead, remove it.
1323 GV->eraseFromParent();
1326 return FieldGlobals[0];
1329 /// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1330 /// pointer global variable with a single value stored it that is a malloc or
1332 static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
1334 Module::global_iterator &GVI,
1336 // If this is a malloc of an abstract type, don't touch it.
1337 if (!MI->getAllocatedType()->isSized())
1340 // We can't optimize this global unless all uses of it are *known* to be
1341 // of the malloc value, not of the null initializer value (consider a use
1342 // that compares the global's value against zero to see if the malloc has
1343 // been reached). To do this, we check to see if all uses of the global
1344 // would trap if the global were null: this proves that they must all
1345 // happen after the malloc.
1346 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1349 // We can't optimize this if the malloc itself is used in a complex way,
1350 // for example, being stored into multiple globals. This allows the
1351 // malloc to be stored into the specified global, loaded setcc'd, and
1352 // GEP'd. These are all things we could transform to using the global
1355 SmallPtrSet<PHINode*, 8> PHIs;
1356 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV, PHIs))
1361 // If we have a global that is only initialized with a fixed size malloc,
1362 // transform the program to use global memory instead of malloc'd memory.
1363 // This eliminates dynamic allocation, avoids an indirection accessing the
1364 // data, and exposes the resultant global to further GlobalOpt.
1365 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) {
1366 // Restrict this transformation to only working on small allocations
1367 // (2048 bytes currently), as we don't want to introduce a 16M global or
1369 if (NElements->getZExtValue()*
1370 TD.getABITypeSize(MI->getAllocatedType()) < 2048) {
1371 GVI = OptimizeGlobalAddressOfMalloc(GV, MI);
1376 // If the allocation is an array of structures, consider transforming this
1377 // into multiple malloc'd arrays, one for each field. This is basically
1378 // SRoA for malloc'd memory.
1379 const Type *AllocTy = MI->getAllocatedType();
1381 // If this is an allocation of a fixed size array of structs, analyze as a
1382 // variable size array. malloc [100 x struct],1 -> malloc struct, 100
1383 if (!MI->isArrayAllocation())
1384 if (const ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
1385 AllocTy = AT->getElementType();
1387 if (const StructType *AllocSTy = dyn_cast<StructType>(AllocTy)) {
1388 // This the structure has an unreasonable number of fields, leave it
1390 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
1391 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, MI)) {
1393 // If this is a fixed size array, transform the Malloc to be an alloc of
1394 // structs. malloc [100 x struct],1 -> malloc struct, 100
1395 if (const ArrayType *AT = dyn_cast<ArrayType>(MI->getAllocatedType())) {
1397 new MallocInst(AllocSTy,
1398 ConstantInt::get(Type::Int32Ty, AT->getNumElements()),
1400 NewMI->takeName(MI);
1401 Value *Cast = new BitCastInst(NewMI, MI->getType(), "tmp", MI);
1402 MI->replaceAllUsesWith(Cast);
1403 MI->eraseFromParent();
1407 GVI = PerformHeapAllocSRoA(GV, MI);
1415 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1416 // that only one value (besides its initializer) is ever stored to the global.
1417 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1418 Module::global_iterator &GVI,
1420 // Ignore no-op GEPs and bitcasts.
1421 StoredOnceVal = StoredOnceVal->stripPointerCasts();
1423 // If we are dealing with a pointer global that is initialized to null and
1424 // only has one (non-null) value stored into it, then we can optimize any
1425 // users of the loaded value (often calls and loads) that would trap if the
1427 if (isa<PointerType>(GV->getInitializer()->getType()) &&
1428 GV->getInitializer()->isNullValue()) {
1429 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1430 if (GV->getInitializer()->getType() != SOVC->getType())
1431 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1433 // Optimize away any trapping uses of the loaded value.
1434 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1436 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) {
1437 if (TryToOptimizeStoreOfMallocToGlobal(GV, MI, GVI, TD))
1445 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1446 /// two values ever stored into GV are its initializer and OtherVal. See if we
1447 /// can shrink the global into a boolean and select between the two values
1448 /// whenever it is used. This exposes the values to other scalar optimizations.
1449 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1450 const Type *GVElType = GV->getType()->getElementType();
1452 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1453 // an FP value or vector, don't do this optimization because a select between
1454 // them is very expensive and unlikely to lead to later simplification.
1455 if (GVElType == Type::Int1Ty || GVElType->isFloatingPoint() ||
1456 isa<VectorType>(GVElType))
1459 // Walk the use list of the global seeing if all the uses are load or store.
1460 // If there is anything else, bail out.
1461 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I)
1462 if (!isa<LoadInst>(I) && !isa<StoreInst>(I))
1465 DOUT << " *** SHRINKING TO BOOL: " << *GV;
1467 // Create the new global, initializing it to false.
1468 GlobalVariable *NewGV = new GlobalVariable(Type::Int1Ty, false,
1469 GlobalValue::InternalLinkage, ConstantInt::getFalse(),
1472 GV->isThreadLocal());
1473 GV->getParent()->getGlobalList().insert(GV, NewGV);
1475 Constant *InitVal = GV->getInitializer();
1476 assert(InitVal->getType() != Type::Int1Ty && "No reason to shrink to bool!");
1478 // If initialized to zero and storing one into the global, we can use a cast
1479 // instead of a select to synthesize the desired value.
1480 bool IsOneZero = false;
1481 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1482 IsOneZero = InitVal->isNullValue() && CI->isOne();
1484 while (!GV->use_empty()) {
1485 Instruction *UI = cast<Instruction>(GV->use_back());
1486 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1487 // Change the store into a boolean store.
1488 bool StoringOther = SI->getOperand(0) == OtherVal;
1489 // Only do this if we weren't storing a loaded value.
1491 if (StoringOther || SI->getOperand(0) == InitVal)
1492 StoreVal = ConstantInt::get(Type::Int1Ty, StoringOther);
1494 // Otherwise, we are storing a previously loaded copy. To do this,
1495 // change the copy from copying the original value to just copying the
1497 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1499 // If we're already replaced the input, StoredVal will be a cast or
1500 // select instruction. If not, it will be a load of the original
1502 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1503 assert(LI->getOperand(0) == GV && "Not a copy!");
1504 // Insert a new load, to preserve the saved value.
1505 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1507 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1508 "This is not a form that we understand!");
1509 StoreVal = StoredVal->getOperand(0);
1510 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1513 new StoreInst(StoreVal, NewGV, SI);
1515 // Change the load into a load of bool then a select.
1516 LoadInst *LI = cast<LoadInst>(UI);
1517 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1520 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1522 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1524 LI->replaceAllUsesWith(NSI);
1526 UI->eraseFromParent();
1529 GV->eraseFromParent();
1534 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1535 /// it if possible. If we make a change, return true.
1536 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1537 Module::global_iterator &GVI) {
1538 SmallPtrSet<PHINode*, 16> PHIUsers;
1540 GV->removeDeadConstantUsers();
1542 if (GV->use_empty()) {
1543 DOUT << "GLOBAL DEAD: " << *GV;
1544 GV->eraseFromParent();
1549 if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
1551 cerr << "Global: " << *GV;
1552 cerr << " isLoaded = " << GS.isLoaded << "\n";
1553 cerr << " StoredType = ";
1554 switch (GS.StoredType) {
1555 case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break;
1556 case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break;
1557 case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break;
1558 case GlobalStatus::isStored: cerr << "stored\n"; break;
1560 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
1561 cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n";
1562 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
1563 cerr << " AccessingFunction = " << GS.AccessingFunction->getName()
1565 cerr << " HasMultipleAccessingFunctions = "
1566 << GS.HasMultipleAccessingFunctions << "\n";
1567 cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n";
1571 // If this is a first class global and has only one accessing function
1572 // and this function is main (which we know is not recursive we can make
1573 // this global a local variable) we replace the global with a local alloca
1574 // in this function.
1576 // NOTE: It doesn't make sense to promote non single-value types since we
1577 // are just replacing static memory to stack memory.
1578 if (!GS.HasMultipleAccessingFunctions &&
1579 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1580 GV->getType()->getElementType()->isSingleValueType() &&
1581 GS.AccessingFunction->getName() == "main" &&
1582 GS.AccessingFunction->hasExternalLinkage()) {
1583 DOUT << "LOCALIZING GLOBAL: " << *GV;
1584 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin();
1585 const Type* ElemTy = GV->getType()->getElementType();
1586 // FIXME: Pass Global's alignment when globals have alignment
1587 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI);
1588 if (!isa<UndefValue>(GV->getInitializer()))
1589 new StoreInst(GV->getInitializer(), Alloca, FirstI);
1591 GV->replaceAllUsesWith(Alloca);
1592 GV->eraseFromParent();
1597 // If the global is never loaded (but may be stored to), it is dead.
1600 DOUT << "GLOBAL NEVER LOADED: " << *GV;
1602 // Delete any stores we can find to the global. We may not be able to
1603 // make it completely dead though.
1604 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1606 // If the global is dead now, delete it.
1607 if (GV->use_empty()) {
1608 GV->eraseFromParent();
1614 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1615 DOUT << "MARKING CONSTANT: " << *GV;
1616 GV->setConstant(true);
1618 // Clean up any obviously simplifiable users now.
1619 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1621 // If the global is dead now, just nuke it.
1622 if (GV->use_empty()) {
1623 DOUT << " *** Marking constant allowed us to simplify "
1624 << "all users and delete global!\n";
1625 GV->eraseFromParent();
1631 } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
1632 if (GlobalVariable *FirstNewGV = SRAGlobal(GV,
1633 getAnalysis<TargetData>())) {
1634 GVI = FirstNewGV; // Don't skip the newly produced globals!
1637 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1638 // If the initial value for the global was an undef value, and if only
1639 // one other value was stored into it, we can just change the
1640 // initializer to be an undef value, then delete all stores to the
1641 // global. This allows us to mark it constant.
1642 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1643 if (isa<UndefValue>(GV->getInitializer())) {
1644 // Change the initial value here.
1645 GV->setInitializer(SOVConstant);
1647 // Clean up any obviously simplifiable users now.
1648 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1650 if (GV->use_empty()) {
1651 DOUT << " *** Substituting initializer allowed us to "
1652 << "simplify all users and delete global!\n";
1653 GV->eraseFromParent();
1662 // Try to optimize globals based on the knowledge that only one value
1663 // (besides its initializer) is ever stored to the global.
1664 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1665 getAnalysis<TargetData>()))
1668 // Otherwise, if the global was not a boolean, we can shrink it to be a
1670 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1671 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1680 /// OnlyCalledDirectly - Return true if the specified function is only called
1681 /// directly. In other words, its address is never taken.
1682 static bool OnlyCalledDirectly(Function *F) {
1683 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1684 Instruction *User = dyn_cast<Instruction>(*UI);
1685 if (!User) return false;
1686 if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false;
1688 // See if the function address is passed as an argument.
1689 for (User::op_iterator i = User->op_begin() + 1, e = User->op_end();
1691 if (*i == F) return false;
1696 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1697 /// function, changing them to FastCC.
1698 static void ChangeCalleesToFastCall(Function *F) {
1699 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1700 CallSite User(cast<Instruction>(*UI));
1701 User.setCallingConv(CallingConv::Fast);
1705 static AttrListPtr StripNest(const AttrListPtr &Attrs) {
1706 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1707 if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0)
1710 // There can be only one.
1711 return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest);
1717 static void RemoveNestAttribute(Function *F) {
1718 F->setAttributes(StripNest(F->getAttributes()));
1719 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1720 CallSite User(cast<Instruction>(*UI));
1721 User.setAttributes(StripNest(User.getAttributes()));
1725 bool GlobalOpt::OptimizeFunctions(Module &M) {
1726 bool Changed = false;
1727 // Optimize functions.
1728 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1730 F->removeDeadConstantUsers();
1731 if (F->use_empty() && (F->hasInternalLinkage() ||
1732 F->hasLinkOnceLinkage())) {
1733 M.getFunctionList().erase(F);
1736 } else if (F->hasInternalLinkage()) {
1737 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1738 OnlyCalledDirectly(F)) {
1739 // If this function has C calling conventions, is not a varargs
1740 // function, and is only called directly, promote it to use the Fast
1741 // calling convention.
1742 F->setCallingConv(CallingConv::Fast);
1743 ChangeCalleesToFastCall(F);
1748 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
1749 OnlyCalledDirectly(F)) {
1750 // The function is not used by a trampoline intrinsic, so it is safe
1751 // to remove the 'nest' attribute.
1752 RemoveNestAttribute(F);
1761 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1762 bool Changed = false;
1763 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1765 GlobalVariable *GV = GVI++;
1766 if (!GV->isConstant() && GV->hasInternalLinkage() &&
1767 GV->hasInitializer())
1768 Changed |= ProcessInternalGlobal(GV, GVI);
1773 /// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1774 /// initializers have an init priority of 65535.
1775 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1776 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1778 if (I->getName() == "llvm.global_ctors") {
1779 // Found it, verify it's an array of { int, void()* }.
1780 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1782 const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1783 if (!STy || STy->getNumElements() != 2 ||
1784 STy->getElementType(0) != Type::Int32Ty) return 0;
1785 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1786 if (!PFTy) return 0;
1787 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1788 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() ||
1789 FTy->getNumParams() != 0)
1792 // Verify that the initializer is simple enough for us to handle.
1793 if (!I->hasInitializer()) return 0;
1794 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1796 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
1797 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(*i)) {
1798 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1801 // Must have a function or null ptr.
1802 if (!isa<Function>(CS->getOperand(1)))
1805 // Init priority must be standard.
1806 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1807 if (!CI || CI->getZExtValue() != 65535)
1818 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1819 /// return a list of the functions and null terminator as a vector.
1820 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1821 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1822 std::vector<Function*> Result;
1823 Result.reserve(CA->getNumOperands());
1824 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
1825 ConstantStruct *CS = cast<ConstantStruct>(*i);
1826 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1831 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1832 /// specified array, returning the new global to use.
1833 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1834 const std::vector<Function*> &Ctors) {
1835 // If we made a change, reassemble the initializer list.
1836 std::vector<Constant*> CSVals;
1837 CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535));
1838 CSVals.push_back(0);
1840 // Create the new init list.
1841 std::vector<Constant*> CAList;
1842 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1844 CSVals[1] = Ctors[i];
1846 const Type *FTy = FunctionType::get(Type::VoidTy,
1847 std::vector<const Type*>(), false);
1848 const PointerType *PFTy = PointerType::getUnqual(FTy);
1849 CSVals[1] = Constant::getNullValue(PFTy);
1850 CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647);
1852 CAList.push_back(ConstantStruct::get(CSVals));
1855 // Create the array initializer.
1856 const Type *StructTy =
1857 cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
1858 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()),
1861 // If we didn't change the number of elements, don't create a new GV.
1862 if (CA->getType() == GCL->getInitializer()->getType()) {
1863 GCL->setInitializer(CA);
1867 // Create the new global and insert it next to the existing list.
1868 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
1869 GCL->getLinkage(), CA, "",
1871 GCL->isThreadLocal());
1872 GCL->getParent()->getGlobalList().insert(GCL, NGV);
1875 // Nuke the old list, replacing any uses with the new one.
1876 if (!GCL->use_empty()) {
1878 if (V->getType() != GCL->getType())
1879 V = ConstantExpr::getBitCast(V, GCL->getType());
1880 GCL->replaceAllUsesWith(V);
1882 GCL->eraseFromParent();
1891 static Constant *getVal(DenseMap<Value*, Constant*> &ComputedValues,
1893 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
1894 Constant *R = ComputedValues[V];
1895 assert(R && "Reference to an uncomputed value!");
1899 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
1900 /// enough for us to understand. In particular, if it is a cast of something,
1901 /// we punt. We basically just support direct accesses to globals and GEP's of
1902 /// globals. This should be kept up to date with CommitValueTo.
1903 static bool isSimpleEnoughPointerToCommit(Constant *C) {
1904 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
1905 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1906 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1907 return !GV->isDeclaration(); // reject external globals.
1909 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
1910 // Handle a constantexpr gep.
1911 if (CE->getOpcode() == Instruction::GetElementPtr &&
1912 isa<GlobalVariable>(CE->getOperand(0))) {
1913 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1914 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1915 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1916 return GV->hasInitializer() &&
1917 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1922 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
1923 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
1924 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
1925 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
1926 ConstantExpr *Addr, unsigned OpNo) {
1927 // Base case of the recursion.
1928 if (OpNo == Addr->getNumOperands()) {
1929 assert(Val->getType() == Init->getType() && "Type mismatch!");
1933 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
1934 std::vector<Constant*> Elts;
1936 // Break up the constant into its elements.
1937 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
1938 for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i)
1939 Elts.push_back(cast<Constant>(*i));
1940 } else if (isa<ConstantAggregateZero>(Init)) {
1941 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1942 Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
1943 } else if (isa<UndefValue>(Init)) {
1944 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1945 Elts.push_back(UndefValue::get(STy->getElementType(i)));
1947 assert(0 && "This code is out of sync with "
1948 " ConstantFoldLoadThroughGEPConstantExpr");
1951 // Replace the element that we are supposed to.
1952 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
1953 unsigned Idx = CU->getZExtValue();
1954 assert(Idx < STy->getNumElements() && "Struct index out of range!");
1955 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
1957 // Return the modified struct.
1958 return ConstantStruct::get(&Elts[0], Elts.size(), STy->isPacked());
1960 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
1961 const ArrayType *ATy = cast<ArrayType>(Init->getType());
1963 // Break up the array into elements.
1964 std::vector<Constant*> Elts;
1965 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
1966 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
1967 Elts.push_back(cast<Constant>(*i));
1968 } else if (isa<ConstantAggregateZero>(Init)) {
1969 Constant *Elt = Constant::getNullValue(ATy->getElementType());
1970 Elts.assign(ATy->getNumElements(), Elt);
1971 } else if (isa<UndefValue>(Init)) {
1972 Constant *Elt = UndefValue::get(ATy->getElementType());
1973 Elts.assign(ATy->getNumElements(), Elt);
1975 assert(0 && "This code is out of sync with "
1976 " ConstantFoldLoadThroughGEPConstantExpr");
1979 assert(CI->getZExtValue() < ATy->getNumElements());
1980 Elts[CI->getZExtValue()] =
1981 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
1982 return ConstantArray::get(ATy, Elts);
1986 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
1987 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
1988 static void CommitValueTo(Constant *Val, Constant *Addr) {
1989 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
1990 assert(GV->hasInitializer());
1991 GV->setInitializer(Val);
1995 ConstantExpr *CE = cast<ConstantExpr>(Addr);
1996 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1998 Constant *Init = GV->getInitializer();
1999 Init = EvaluateStoreInto(Init, Val, CE, 2);
2000 GV->setInitializer(Init);
2003 /// ComputeLoadResult - Return the value that would be computed by a load from
2004 /// P after the stores reflected by 'memory' have been performed. If we can't
2005 /// decide, return null.
2006 static Constant *ComputeLoadResult(Constant *P,
2007 const DenseMap<Constant*, Constant*> &Memory) {
2008 // If this memory location has been recently stored, use the stored value: it
2009 // is the most up-to-date.
2010 DenseMap<Constant*, Constant*>::const_iterator I = Memory.find(P);
2011 if (I != Memory.end()) return I->second;
2014 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
2015 if (GV->hasInitializer())
2016 return GV->getInitializer();
2020 // Handle a constantexpr getelementptr.
2021 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
2022 if (CE->getOpcode() == Instruction::GetElementPtr &&
2023 isa<GlobalVariable>(CE->getOperand(0))) {
2024 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2025 if (GV->hasInitializer())
2026 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2029 return 0; // don't know how to evaluate.
2032 /// EvaluateFunction - Evaluate a call to function F, returning true if
2033 /// successful, false if we can't evaluate it. ActualArgs contains the formal
2034 /// arguments for the function.
2035 static bool EvaluateFunction(Function *F, Constant *&RetVal,
2036 const std::vector<Constant*> &ActualArgs,
2037 std::vector<Function*> &CallStack,
2038 DenseMap<Constant*, Constant*> &MutatedMemory,
2039 std::vector<GlobalVariable*> &AllocaTmps) {
2040 // Check to see if this function is already executing (recursion). If so,
2041 // bail out. TODO: we might want to accept limited recursion.
2042 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
2045 CallStack.push_back(F);
2047 /// Values - As we compute SSA register values, we store their contents here.
2048 DenseMap<Value*, Constant*> Values;
2050 // Initialize arguments to the incoming values specified.
2052 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
2054 Values[AI] = ActualArgs[ArgNo];
2056 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
2057 /// we can only evaluate any one basic block at most once. This set keeps
2058 /// track of what we have executed so we can detect recursive cases etc.
2059 SmallPtrSet<BasicBlock*, 32> ExecutedBlocks;
2061 // CurInst - The current instruction we're evaluating.
2062 BasicBlock::iterator CurInst = F->begin()->begin();
2064 // This is the main evaluation loop.
2066 Constant *InstResult = 0;
2068 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
2069 if (SI->isVolatile()) return false; // no volatile accesses.
2070 Constant *Ptr = getVal(Values, SI->getOperand(1));
2071 if (!isSimpleEnoughPointerToCommit(Ptr))
2072 // If this is too complex for us to commit, reject it.
2074 Constant *Val = getVal(Values, SI->getOperand(0));
2075 MutatedMemory[Ptr] = Val;
2076 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2077 InstResult = ConstantExpr::get(BO->getOpcode(),
2078 getVal(Values, BO->getOperand(0)),
2079 getVal(Values, BO->getOperand(1)));
2080 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2081 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2082 getVal(Values, CI->getOperand(0)),
2083 getVal(Values, CI->getOperand(1)));
2084 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2085 InstResult = ConstantExpr::getCast(CI->getOpcode(),
2086 getVal(Values, CI->getOperand(0)),
2088 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2089 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
2090 getVal(Values, SI->getOperand(1)),
2091 getVal(Values, SI->getOperand(2)));
2092 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2093 Constant *P = getVal(Values, GEP->getOperand(0));
2094 SmallVector<Constant*, 8> GEPOps;
2095 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2097 GEPOps.push_back(getVal(Values, *i));
2098 InstResult = ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size());
2099 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2100 if (LI->isVolatile()) return false; // no volatile accesses.
2101 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
2103 if (InstResult == 0) return false; // Could not evaluate load.
2104 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2105 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs.
2106 const Type *Ty = AI->getType()->getElementType();
2107 AllocaTmps.push_back(new GlobalVariable(Ty, false,
2108 GlobalValue::InternalLinkage,
2109 UndefValue::get(Ty),
2111 InstResult = AllocaTmps.back();
2112 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
2113 // Cannot handle inline asm.
2114 if (isa<InlineAsm>(CI->getOperand(0))) return false;
2116 // Resolve function pointers.
2117 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0)));
2118 if (!Callee) return false; // Cannot resolve.
2120 std::vector<Constant*> Formals;
2121 for (User::op_iterator i = CI->op_begin() + 1, e = CI->op_end();
2123 Formals.push_back(getVal(Values, *i));
2125 if (Callee->isDeclaration()) {
2126 // If this is a function we can constant fold, do it.
2127 if (Constant *C = ConstantFoldCall(Callee, &Formals[0],
2134 if (Callee->getFunctionType()->isVarArg())
2139 // Execute the call, if successful, use the return value.
2140 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
2141 MutatedMemory, AllocaTmps))
2143 InstResult = RetVal;
2145 } else if (isa<TerminatorInst>(CurInst)) {
2146 BasicBlock *NewBB = 0;
2147 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2148 if (BI->isUnconditional()) {
2149 NewBB = BI->getSuccessor(0);
2152 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
2153 if (!Cond) return false; // Cannot determine.
2155 NewBB = BI->getSuccessor(!Cond->getZExtValue());
2157 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2159 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
2160 if (!Val) return false; // Cannot determine.
2161 NewBB = SI->getSuccessor(SI->findCaseValue(Val));
2162 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
2163 if (RI->getNumOperands())
2164 RetVal = getVal(Values, RI->getOperand(0));
2166 CallStack.pop_back(); // return from fn.
2167 return true; // We succeeded at evaluating this ctor!
2169 // invoke, unwind, unreachable.
2170 return false; // Cannot handle this terminator.
2173 // Okay, we succeeded in evaluating this control flow. See if we have
2174 // executed the new block before. If so, we have a looping function,
2175 // which we cannot evaluate in reasonable time.
2176 if (!ExecutedBlocks.insert(NewBB))
2177 return false; // looped!
2179 // Okay, we have never been in this block before. Check to see if there
2180 // are any PHI nodes. If so, evaluate them with information about where
2182 BasicBlock *OldBB = CurInst->getParent();
2183 CurInst = NewBB->begin();
2185 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2186 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
2188 // Do NOT increment CurInst. We know that the terminator had no value.
2191 // Did not know how to evaluate this!
2195 if (!CurInst->use_empty())
2196 Values[CurInst] = InstResult;
2198 // Advance program counter.
2203 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2204 /// we can. Return true if we can, false otherwise.
2205 static bool EvaluateStaticConstructor(Function *F) {
2206 /// MutatedMemory - For each store we execute, we update this map. Loads
2207 /// check this to get the most up-to-date value. If evaluation is successful,
2208 /// this state is committed to the process.
2209 DenseMap<Constant*, Constant*> MutatedMemory;
2211 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2212 /// to represent its body. This vector is needed so we can delete the
2213 /// temporary globals when we are done.
2214 std::vector<GlobalVariable*> AllocaTmps;
2216 /// CallStack - This is used to detect recursion. In pathological situations
2217 /// we could hit exponential behavior, but at least there is nothing
2219 std::vector<Function*> CallStack;
2221 // Call the function.
2222 Constant *RetValDummy;
2223 bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(),
2224 CallStack, MutatedMemory, AllocaTmps);
2226 // We succeeded at evaluation: commit the result.
2227 DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2228 << F->getName() << "' to " << MutatedMemory.size()
2230 for (DenseMap<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2231 E = MutatedMemory.end(); I != E; ++I)
2232 CommitValueTo(I->second, I->first);
2235 // At this point, we are done interpreting. If we created any 'alloca'
2236 // temporaries, release them now.
2237 while (!AllocaTmps.empty()) {
2238 GlobalVariable *Tmp = AllocaTmps.back();
2239 AllocaTmps.pop_back();
2241 // If there are still users of the alloca, the program is doing something
2242 // silly, e.g. storing the address of the alloca somewhere and using it
2243 // later. Since this is undefined, we'll just make it be null.
2244 if (!Tmp->use_empty())
2245 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2254 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2255 /// Return true if anything changed.
2256 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2257 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2258 bool MadeChange = false;
2259 if (Ctors.empty()) return false;
2261 // Loop over global ctors, optimizing them when we can.
2262 for (unsigned i = 0; i != Ctors.size(); ++i) {
2263 Function *F = Ctors[i];
2264 // Found a null terminator in the middle of the list, prune off the rest of
2267 if (i != Ctors.size()-1) {
2274 // We cannot simplify external ctor functions.
2275 if (F->empty()) continue;
2277 // If we can evaluate the ctor at compile time, do.
2278 if (EvaluateStaticConstructor(F)) {
2279 Ctors.erase(Ctors.begin()+i);
2282 ++NumCtorsEvaluated;
2287 if (!MadeChange) return false;
2289 GCL = InstallGlobalCtors(GCL, Ctors);
2293 bool GlobalOpt::ResolveAliases(Module &M) {
2294 bool Changed = false;
2296 for (Module::alias_iterator I = M.alias_begin(),
2297 E = M.alias_end(); I != E; ++I) {
2301 if (const GlobalValue *GV = I->resolveAliasedGlobal())
2303 I->replaceAllUsesWith(const_cast<GlobalValue*>(GV));
2311 bool GlobalOpt::runOnModule(Module &M) {
2312 bool Changed = false;
2314 // Try to find the llvm.globalctors list.
2315 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2317 bool LocalChange = true;
2318 while (LocalChange) {
2319 LocalChange = false;
2321 // Delete functions that are trivially dead, ccc -> fastcc
2322 LocalChange |= OptimizeFunctions(M);
2324 // Optimize global_ctors list.
2326 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2328 // Optimize non-address-taken globals.
2329 LocalChange |= OptimizeGlobalVars(M);
2331 // Resolve aliases, when possible.
2332 LocalChange |= ResolveAliases(M);
2333 Changed |= LocalChange;
2336 // TODO: Move all global ctors functions to the end of the module for code