1 //===- GlobalOpt.cpp - Optimize Global Variables --------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass transforms simple global variables that never have their address
11 // taken. If obviously true, it marks read/write globals as constant, deletes
12 // variables only stored to, etc.
14 //===----------------------------------------------------------------------===//
16 #define DEBUG_TYPE "globalopt"
17 #include "llvm/Transforms/IPO.h"
18 #include "llvm/CallingConv.h"
19 #include "llvm/Constants.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/Instructions.h"
22 #include "llvm/IntrinsicInst.h"
23 #include "llvm/Module.h"
24 #include "llvm/Pass.h"
25 #include "llvm/Analysis/ConstantFolding.h"
26 #include "llvm/Target/TargetData.h"
27 #include "llvm/Support/CallSite.h"
28 #include "llvm/Support/Compiler.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/GetElementPtrTypeIterator.h"
31 #include "llvm/Support/MathExtras.h"
32 #include "llvm/ADT/SmallPtrSet.h"
33 #include "llvm/ADT/SmallVector.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/ADT/StringExtras.h"
41 STATISTIC(NumMarked , "Number of globals marked constant");
42 STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
43 STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
44 STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
45 STATISTIC(NumDeleted , "Number of globals deleted");
46 STATISTIC(NumFnDeleted , "Number of functions deleted");
47 STATISTIC(NumGlobUses , "Number of global uses devirtualized");
48 STATISTIC(NumLocalized , "Number of globals localized");
49 STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans");
50 STATISTIC(NumFastCallFns , "Number of functions converted to fastcc");
51 STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
52 STATISTIC(NumNestRemoved , "Number of nest attributes removed");
55 struct VISIBILITY_HIDDEN GlobalOpt : public ModulePass {
56 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
57 AU.addRequired<TargetData>();
59 static char ID; // Pass identification, replacement for typeid
60 GlobalOpt() : ModulePass((intptr_t)&ID) {}
62 bool runOnModule(Module &M);
65 GlobalVariable *FindGlobalCtors(Module &M);
66 bool OptimizeFunctions(Module &M);
67 bool OptimizeGlobalVars(Module &M);
68 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
69 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
72 char GlobalOpt::ID = 0;
73 RegisterPass<GlobalOpt> X("globalopt", "Global Variable Optimizer");
76 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
78 /// GlobalStatus - As we analyze each global, keep track of some information
79 /// about it. If we find out that the address of the global is taken, none of
80 /// this info will be accurate.
81 struct VISIBILITY_HIDDEN GlobalStatus {
82 /// isLoaded - True if the global is ever loaded. If the global isn't ever
83 /// loaded it can be deleted.
86 /// StoredType - Keep track of what stores to the global look like.
89 /// NotStored - There is no store to this global. It can thus be marked
93 /// isInitializerStored - This global is stored to, but the only thing
94 /// stored is the constant it was initialized with. This is only tracked
95 /// for scalar globals.
98 /// isStoredOnce - This global is stored to, but only its initializer and
99 /// one other value is ever stored to it. If this global isStoredOnce, we
100 /// track the value stored to it in StoredOnceValue below. This is only
101 /// tracked for scalar globals.
104 /// isStored - This global is stored to by multiple values or something else
105 /// that we cannot track.
109 /// StoredOnceValue - If only one value (besides the initializer constant) is
110 /// ever stored to this global, keep track of what value it is.
111 Value *StoredOnceValue;
113 /// AccessingFunction/HasMultipleAccessingFunctions - These start out
114 /// null/false. When the first accessing function is noticed, it is recorded.
115 /// When a second different accessing function is noticed,
116 /// HasMultipleAccessingFunctions is set to true.
117 Function *AccessingFunction;
118 bool HasMultipleAccessingFunctions;
120 /// HasNonInstructionUser - Set to true if this global has a user that is not
121 /// an instruction (e.g. a constant expr or GV initializer).
122 bool HasNonInstructionUser;
124 /// HasPHIUser - Set to true if this global has a user that is a PHI node.
127 GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0),
128 AccessingFunction(0), HasMultipleAccessingFunctions(false),
129 HasNonInstructionUser(false), HasPHIUser(false) {}
134 /// ConstantIsDead - Return true if the specified constant is (transitively)
135 /// dead. The constant may be used by other constants (e.g. constant arrays and
136 /// constant exprs) as long as they are dead, but it cannot be used by anything
138 static bool ConstantIsDead(Constant *C) {
139 if (isa<GlobalValue>(C)) return false;
141 for (Value::use_iterator UI = C->use_begin(), E = C->use_end(); UI != E; ++UI)
142 if (Constant *CU = dyn_cast<Constant>(*UI)) {
143 if (!ConstantIsDead(CU)) return false;
150 /// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
151 /// structure. If the global has its address taken, return true to indicate we
152 /// can't do anything with it.
154 static bool AnalyzeGlobal(Value *V, GlobalStatus &GS,
155 std::set<PHINode*> &PHIUsers) {
156 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
157 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(*UI)) {
158 GS.HasNonInstructionUser = true;
160 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
162 } else if (Instruction *I = dyn_cast<Instruction>(*UI)) {
163 if (!GS.HasMultipleAccessingFunctions) {
164 Function *F = I->getParent()->getParent();
165 if (GS.AccessingFunction == 0)
166 GS.AccessingFunction = F;
167 else if (GS.AccessingFunction != F)
168 GS.HasMultipleAccessingFunctions = true;
170 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
172 if (LI->isVolatile()) return true; // Don't hack on volatile loads.
173 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
174 // Don't allow a store OF the address, only stores TO the address.
175 if (SI->getOperand(0) == V) return true;
177 if (SI->isVolatile()) return true; // Don't hack on volatile stores.
179 // If this is a direct store to the global (i.e., the global is a scalar
180 // value, not an aggregate), keep more specific information about
182 if (GS.StoredType != GlobalStatus::isStored) {
183 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(SI->getOperand(1))){
184 Value *StoredVal = SI->getOperand(0);
185 if (StoredVal == GV->getInitializer()) {
186 if (GS.StoredType < GlobalStatus::isInitializerStored)
187 GS.StoredType = GlobalStatus::isInitializerStored;
188 } else if (isa<LoadInst>(StoredVal) &&
189 cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
191 if (GS.StoredType < GlobalStatus::isInitializerStored)
192 GS.StoredType = GlobalStatus::isInitializerStored;
193 } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
194 GS.StoredType = GlobalStatus::isStoredOnce;
195 GS.StoredOnceValue = StoredVal;
196 } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
197 GS.StoredOnceValue == StoredVal) {
200 GS.StoredType = GlobalStatus::isStored;
203 GS.StoredType = GlobalStatus::isStored;
206 } else if (isa<GetElementPtrInst>(I)) {
207 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
208 } else if (isa<SelectInst>(I)) {
209 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
210 } else if (PHINode *PN = dyn_cast<PHINode>(I)) {
211 // PHI nodes we can check just like select or GEP instructions, but we
212 // have to be careful about infinite recursion.
213 if (PHIUsers.insert(PN).second) // Not already visited.
214 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
215 GS.HasPHIUser = true;
216 } else if (isa<CmpInst>(I)) {
217 } else if (isa<MemCpyInst>(I) || isa<MemMoveInst>(I)) {
218 if (I->getOperand(1) == V)
219 GS.StoredType = GlobalStatus::isStored;
220 if (I->getOperand(2) == V)
222 } else if (isa<MemSetInst>(I)) {
223 assert(I->getOperand(1) == V && "Memset only takes one pointer!");
224 GS.StoredType = GlobalStatus::isStored;
226 return true; // Any other non-load instruction might take address!
228 } else if (Constant *C = dyn_cast<Constant>(*UI)) {
229 GS.HasNonInstructionUser = true;
230 // We might have a dead and dangling constant hanging off of here.
231 if (!ConstantIsDead(C))
234 GS.HasNonInstructionUser = true;
235 // Otherwise must be some other user.
242 static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) {
243 ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
245 unsigned IdxV = CI->getZExtValue();
247 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
248 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
249 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
250 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
251 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) {
252 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
253 } else if (isa<ConstantAggregateZero>(Agg)) {
254 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
255 if (IdxV < STy->getNumElements())
256 return Constant::getNullValue(STy->getElementType(IdxV));
257 } else if (const SequentialType *STy =
258 dyn_cast<SequentialType>(Agg->getType())) {
259 return Constant::getNullValue(STy->getElementType());
261 } else if (isa<UndefValue>(Agg)) {
262 if (const StructType *STy = dyn_cast<StructType>(Agg->getType())) {
263 if (IdxV < STy->getNumElements())
264 return UndefValue::get(STy->getElementType(IdxV));
265 } else if (const SequentialType *STy =
266 dyn_cast<SequentialType>(Agg->getType())) {
267 return UndefValue::get(STy->getElementType());
274 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
275 /// users of the global, cleaning up the obvious ones. This is largely just a
276 /// quick scan over the use list to clean up the easy and obvious cruft. This
277 /// returns true if it made a change.
278 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
279 bool Changed = false;
280 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
283 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
285 // Replace the load with the initializer.
286 LI->replaceAllUsesWith(Init);
287 LI->eraseFromParent();
290 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
291 // Store must be unreachable or storing Init into the global.
292 SI->eraseFromParent();
294 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
295 if (CE->getOpcode() == Instruction::GetElementPtr) {
296 Constant *SubInit = 0;
298 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
299 Changed |= CleanupConstantGlobalUsers(CE, SubInit);
300 } else if (CE->getOpcode() == Instruction::BitCast &&
301 isa<PointerType>(CE->getType())) {
302 // Pointer cast, delete any stores and memsets to the global.
303 Changed |= CleanupConstantGlobalUsers(CE, 0);
306 if (CE->use_empty()) {
307 CE->destroyConstant();
310 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
311 // Do not transform "gepinst (gep constexpr (GV))" here, because forming
312 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
313 // and will invalidate our notion of what Init is.
314 Constant *SubInit = 0;
315 if (!isa<ConstantExpr>(GEP->getOperand(0))) {
317 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
318 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
319 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
321 Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
323 if (GEP->use_empty()) {
324 GEP->eraseFromParent();
327 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
328 if (MI->getRawDest() == V) {
329 MI->eraseFromParent();
333 } else if (Constant *C = dyn_cast<Constant>(U)) {
334 // If we have a chain of dead constantexprs or other things dangling from
335 // us, and if they are all dead, nuke them without remorse.
336 if (ConstantIsDead(C)) {
337 C->destroyConstant();
338 // This could have invalidated UI, start over from scratch.
339 CleanupConstantGlobalUsers(V, Init);
347 /// isSafeSROAElementUse - Return true if the specified instruction is a safe
348 /// user of a derived expression from a global that we want to SROA.
349 static bool isSafeSROAElementUse(Value *V) {
350 // We might have a dead and dangling constant hanging off of here.
351 if (Constant *C = dyn_cast<Constant>(V))
352 return ConstantIsDead(C);
354 Instruction *I = dyn_cast<Instruction>(V);
355 if (!I) return false;
358 if (isa<LoadInst>(I)) return true;
360 // Stores *to* the pointer are ok.
361 if (StoreInst *SI = dyn_cast<StoreInst>(I))
362 return SI->getOperand(0) != V;
364 // Otherwise, it must be a GEP.
365 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I);
366 if (GEPI == 0) return false;
368 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
369 !cast<Constant>(GEPI->getOperand(1))->isNullValue())
372 for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end();
374 if (!isSafeSROAElementUse(*I))
380 /// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value.
381 /// Look at it and its uses and decide whether it is safe to SROA this global.
383 static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
384 // The user of the global must be a GEP Inst or a ConstantExpr GEP.
385 if (!isa<GetElementPtrInst>(U) &&
386 (!isa<ConstantExpr>(U) ||
387 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
390 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
391 // don't like < 3 operand CE's, and we don't like non-constant integer
392 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some
394 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
395 !cast<Constant>(U->getOperand(1))->isNullValue() ||
396 !isa<ConstantInt>(U->getOperand(2)))
399 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
400 ++GEPI; // Skip over the pointer index.
402 // If this is a use of an array allocation, do a bit more checking for sanity.
403 if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
404 uint64_t NumElements = AT->getNumElements();
405 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
407 // Check to make sure that index falls within the array. If not,
408 // something funny is going on, so we won't do the optimization.
410 if (Idx->getZExtValue() >= NumElements)
413 // We cannot scalar repl this level of the array unless any array
414 // sub-indices are in-range constants. In particular, consider:
415 // A[0][i]. We cannot know that the user isn't doing invalid things like
416 // allowing i to index an out-of-range subscript that accesses A[1].
418 // Scalar replacing *just* the outer index of the array is probably not
419 // going to be a win anyway, so just give up.
420 for (++GEPI; // Skip array index.
421 GEPI != E && (isa<ArrayType>(*GEPI) || isa<VectorType>(*GEPI));
423 uint64_t NumElements;
424 if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
425 NumElements = SubArrayTy->getNumElements();
427 NumElements = cast<VectorType>(*GEPI)->getNumElements();
429 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
430 if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
435 for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I)
436 if (!isSafeSROAElementUse(*I))
441 /// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
442 /// is safe for us to perform this transformation.
444 static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
445 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
447 if (!IsUserOfGlobalSafeForSRA(*UI, GV))
454 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
455 /// variable. This opens the door for other optimizations by exposing the
456 /// behavior of the program in a more fine-grained way. We have determined that
457 /// this transformation is safe already. We return the first global variable we
458 /// insert so that the caller can reprocess it.
459 static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) {
460 // Make sure this global only has simple uses that we can SRA.
461 if (!GlobalUsersSafeToSRA(GV))
464 assert(GV->hasInternalLinkage() && !GV->isConstant());
465 Constant *Init = GV->getInitializer();
466 const Type *Ty = Init->getType();
468 std::vector<GlobalVariable*> NewGlobals;
469 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
471 // Get the alignment of the global, either explicit or target-specific.
472 unsigned StartAlignment = GV->getAlignment();
473 if (StartAlignment == 0)
474 StartAlignment = TD.getABITypeAlignment(GV->getType());
476 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
477 NewGlobals.reserve(STy->getNumElements());
478 const StructLayout &Layout = *TD.getStructLayout(STy);
479 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
480 Constant *In = getAggregateConstantElement(Init,
481 ConstantInt::get(Type::Int32Ty, i));
482 assert(In && "Couldn't get element of initializer?");
483 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
484 GlobalVariable::InternalLinkage,
485 In, GV->getName()+"."+utostr(i),
487 GV->isThreadLocal());
488 Globals.insert(GV, NGV);
489 NewGlobals.push_back(NGV);
491 // Calculate the known alignment of the field. If the original aggregate
492 // had 256 byte alignment for example, something might depend on that:
493 // propagate info to each field.
494 uint64_t FieldOffset = Layout.getElementOffset(i);
495 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset);
496 if (NewAlign > TD.getABITypeAlignment(STy->getElementType(i)))
497 NGV->setAlignment(NewAlign);
499 } else if (const SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
500 unsigned NumElements = 0;
501 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy))
502 NumElements = ATy->getNumElements();
504 NumElements = cast<VectorType>(STy)->getNumElements();
506 if (NumElements > 16 && GV->hasNUsesOrMore(16))
507 return 0; // It's not worth it.
508 NewGlobals.reserve(NumElements);
510 uint64_t EltSize = TD.getABITypeSize(STy->getElementType());
511 unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType());
512 for (unsigned i = 0, e = NumElements; i != e; ++i) {
513 Constant *In = getAggregateConstantElement(Init,
514 ConstantInt::get(Type::Int32Ty, i));
515 assert(In && "Couldn't get element of initializer?");
517 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
518 GlobalVariable::InternalLinkage,
519 In, GV->getName()+"."+utostr(i),
521 GV->isThreadLocal());
522 Globals.insert(GV, NGV);
523 NewGlobals.push_back(NGV);
525 // Calculate the known alignment of the field. If the original aggregate
526 // had 256 byte alignment for example, something might depend on that:
527 // propagate info to each field.
528 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
529 if (NewAlign > EltAlign)
530 NGV->setAlignment(NewAlign);
534 if (NewGlobals.empty())
537 DOUT << "PERFORMING GLOBAL SRA ON: " << *GV;
539 Constant *NullInt = Constant::getNullValue(Type::Int32Ty);
541 // Loop over all of the uses of the global, replacing the constantexpr geps,
542 // with smaller constantexpr geps or direct references.
543 while (!GV->use_empty()) {
544 User *GEP = GV->use_back();
545 assert(((isa<ConstantExpr>(GEP) &&
546 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
547 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
549 // Ignore the 1th operand, which has to be zero or else the program is quite
550 // broken (undefined). Get the 2nd operand, which is the structure or array
552 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
553 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
555 Value *NewPtr = NewGlobals[Val];
557 // Form a shorter GEP if needed.
558 if (GEP->getNumOperands() > 3) {
559 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
560 SmallVector<Constant*, 8> Idxs;
561 Idxs.push_back(NullInt);
562 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
563 Idxs.push_back(CE->getOperand(i));
564 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr),
565 &Idxs[0], Idxs.size());
567 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
568 SmallVector<Value*, 8> Idxs;
569 Idxs.push_back(NullInt);
570 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
571 Idxs.push_back(GEPI->getOperand(i));
572 NewPtr = GetElementPtrInst::Create(NewPtr, Idxs.begin(), Idxs.end(),
573 GEPI->getName()+"."+utostr(Val), GEPI);
576 GEP->replaceAllUsesWith(NewPtr);
578 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
579 GEPI->eraseFromParent();
581 cast<ConstantExpr>(GEP)->destroyConstant();
584 // Delete the old global, now that it is dead.
588 // Loop over the new globals array deleting any globals that are obviously
589 // dead. This can arise due to scalarization of a structure or an array that
590 // has elements that are dead.
591 unsigned FirstGlobal = 0;
592 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
593 if (NewGlobals[i]->use_empty()) {
594 Globals.erase(NewGlobals[i]);
595 if (FirstGlobal == i) ++FirstGlobal;
598 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
601 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
602 /// value will trap if the value is dynamically null. PHIs keeps track of any
603 /// phi nodes we've seen to avoid reprocessing them.
604 static bool AllUsesOfValueWillTrapIfNull(Value *V,
605 SmallPtrSet<PHINode*, 8> &PHIs) {
606 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
607 if (isa<LoadInst>(*UI)) {
609 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
610 if (SI->getOperand(0) == V) {
611 //cerr << "NONTRAPPING USE: " << **UI;
612 return false; // Storing the value.
614 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
615 if (CI->getOperand(0) != V) {
616 //cerr << "NONTRAPPING USE: " << **UI;
617 return false; // Not calling the ptr
619 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
620 if (II->getOperand(0) != V) {
621 //cerr << "NONTRAPPING USE: " << **UI;
622 return false; // Not calling the ptr
624 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(*UI)) {
625 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
626 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
627 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
628 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
629 // If we've already seen this phi node, ignore it, it has already been
632 return AllUsesOfValueWillTrapIfNull(PN, PHIs);
633 } else if (isa<ICmpInst>(*UI) &&
634 isa<ConstantPointerNull>(UI->getOperand(1))) {
635 // Ignore setcc X, null
637 //cerr << "NONTRAPPING USE: " << **UI;
643 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
644 /// from GV will trap if the loaded value is null. Note that this also permits
645 /// comparisons of the loaded value against null, as a special case.
646 static bool AllUsesOfLoadedValueWillTrapIfNull(GlobalVariable *GV) {
647 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI!=E; ++UI)
648 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
649 SmallPtrSet<PHINode*, 8> PHIs;
650 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
652 } else if (isa<StoreInst>(*UI)) {
653 // Ignore stores to the global.
655 // We don't know or understand this user, bail out.
656 //cerr << "UNKNOWN USER OF GLOBAL!: " << **UI;
663 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
664 bool Changed = false;
665 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
666 Instruction *I = cast<Instruction>(*UI++);
667 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
668 LI->setOperand(0, NewV);
670 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
671 if (SI->getOperand(1) == V) {
672 SI->setOperand(1, NewV);
675 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
676 if (I->getOperand(0) == V) {
677 // Calling through the pointer! Turn into a direct call, but be careful
678 // that the pointer is not also being passed as an argument.
679 I->setOperand(0, NewV);
681 bool PassedAsArg = false;
682 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i)
683 if (I->getOperand(i) == V) {
685 I->setOperand(i, NewV);
689 // Being passed as an argument also. Be careful to not invalidate UI!
693 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
694 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
695 ConstantExpr::getCast(CI->getOpcode(),
696 NewV, CI->getType()));
697 if (CI->use_empty()) {
699 CI->eraseFromParent();
701 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
702 // Should handle GEP here.
703 SmallVector<Constant*, 8> Idxs;
704 Idxs.reserve(GEPI->getNumOperands()-1);
705 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
706 if (Constant *C = dyn_cast<Constant>(GEPI->getOperand(i)))
710 if (Idxs.size() == GEPI->getNumOperands()-1)
711 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
712 ConstantExpr::getGetElementPtr(NewV, &Idxs[0],
714 if (GEPI->use_empty()) {
716 GEPI->eraseFromParent();
725 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
726 /// value stored into it. If there are uses of the loaded value that would trap
727 /// if the loaded value is dynamically null, then we know that they cannot be
728 /// reachable with a null optimize away the load.
729 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) {
730 std::vector<LoadInst*> Loads;
731 bool Changed = false;
733 // Replace all uses of loads with uses of uses of the stored value.
734 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end();
736 if (LoadInst *LI = dyn_cast<LoadInst>(*GUI)) {
738 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
740 // If we get here we could have stores, selects, or phi nodes whose values
742 assert((isa<StoreInst>(*GUI) || isa<PHINode>(*GUI) ||
743 isa<SelectInst>(*GUI) || isa<ConstantExpr>(*GUI)) &&
744 "Only expect load and stores!");
748 DOUT << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV;
752 // Delete all of the loads we can, keeping track of whether we nuked them all!
753 bool AllLoadsGone = true;
754 while (!Loads.empty()) {
755 LoadInst *L = Loads.back();
756 if (L->use_empty()) {
757 L->eraseFromParent();
760 AllLoadsGone = false;
765 // If we nuked all of the loads, then none of the stores are needed either,
766 // nor is the global.
768 DOUT << " *** GLOBAL NOW DEAD!\n";
769 CleanupConstantGlobalUsers(GV, 0);
770 if (GV->use_empty()) {
771 GV->eraseFromParent();
779 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
780 /// instructions that are foldable.
781 static void ConstantPropUsersOf(Value *V) {
782 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
783 if (Instruction *I = dyn_cast<Instruction>(*UI++))
784 if (Constant *NewC = ConstantFoldInstruction(I)) {
785 I->replaceAllUsesWith(NewC);
787 // Advance UI to the next non-I use to avoid invalidating it!
788 // Instructions could multiply use V.
789 while (UI != E && *UI == I)
791 I->eraseFromParent();
795 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
796 /// variable, and transforms the program as if it always contained the result of
797 /// the specified malloc. Because it is always the result of the specified
798 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
799 /// malloc into a global, and any loads of GV as uses of the new global.
800 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
802 DOUT << "PROMOTING MALLOC GLOBAL: " << *GV << " MALLOC = " << *MI;
803 ConstantInt *NElements = cast<ConstantInt>(MI->getArraySize());
805 if (NElements->getZExtValue() != 1) {
806 // If we have an array allocation, transform it to a single element
807 // allocation to make the code below simpler.
808 Type *NewTy = ArrayType::get(MI->getAllocatedType(),
809 NElements->getZExtValue());
811 new MallocInst(NewTy, Constant::getNullValue(Type::Int32Ty),
812 MI->getAlignment(), MI->getName(), MI);
814 Indices[0] = Indices[1] = Constant::getNullValue(Type::Int32Ty);
815 Value *NewGEP = GetElementPtrInst::Create(NewMI, Indices, Indices + 2,
816 NewMI->getName()+".el0", MI);
817 MI->replaceAllUsesWith(NewGEP);
818 MI->eraseFromParent();
822 // Create the new global variable. The contents of the malloc'd memory is
823 // undefined, so initialize with an undef value.
824 Constant *Init = UndefValue::get(MI->getAllocatedType());
825 GlobalVariable *NewGV = new GlobalVariable(MI->getAllocatedType(), false,
826 GlobalValue::InternalLinkage, Init,
827 GV->getName()+".body",
829 GV->isThreadLocal());
830 // FIXME: This new global should have the alignment returned by malloc. Code
831 // could depend on malloc returning large alignment (on the mac, 16 bytes) but
832 // this would only guarantee some lower alignment.
833 GV->getParent()->getGlobalList().insert(GV, NewGV);
835 // Anything that used the malloc now uses the global directly.
836 MI->replaceAllUsesWith(NewGV);
838 Constant *RepValue = NewGV;
839 if (NewGV->getType() != GV->getType()->getElementType())
840 RepValue = ConstantExpr::getBitCast(RepValue,
841 GV->getType()->getElementType());
843 // If there is a comparison against null, we will insert a global bool to
844 // keep track of whether the global was initialized yet or not.
845 GlobalVariable *InitBool =
846 new GlobalVariable(Type::Int1Ty, false, GlobalValue::InternalLinkage,
847 ConstantInt::getFalse(), GV->getName()+".init",
848 (Module *)NULL, GV->isThreadLocal());
849 bool InitBoolUsed = false;
851 // Loop over all uses of GV, processing them in turn.
852 std::vector<StoreInst*> Stores;
853 while (!GV->use_empty())
854 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
855 while (!LI->use_empty()) {
856 Use &LoadUse = LI->use_begin().getUse();
857 if (!isa<ICmpInst>(LoadUse.getUser()))
860 ICmpInst *CI = cast<ICmpInst>(LoadUse.getUser());
861 // Replace the cmp X, 0 with a use of the bool value.
862 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", CI);
864 switch (CI->getPredicate()) {
865 default: assert(0 && "Unknown ICmp Predicate!");
866 case ICmpInst::ICMP_ULT:
867 case ICmpInst::ICMP_SLT:
868 LV = ConstantInt::getFalse(); // X < null -> always false
870 case ICmpInst::ICMP_ULE:
871 case ICmpInst::ICMP_SLE:
872 case ICmpInst::ICMP_EQ:
873 LV = BinaryOperator::createNot(LV, "notinit", CI);
875 case ICmpInst::ICMP_NE:
876 case ICmpInst::ICMP_UGE:
877 case ICmpInst::ICMP_SGE:
878 case ICmpInst::ICMP_UGT:
879 case ICmpInst::ICMP_SGT:
882 CI->replaceAllUsesWith(LV);
883 CI->eraseFromParent();
886 LI->eraseFromParent();
888 StoreInst *SI = cast<StoreInst>(GV->use_back());
889 // The global is initialized when the store to it occurs.
890 new StoreInst(ConstantInt::getTrue(), InitBool, SI);
891 SI->eraseFromParent();
894 // If the initialization boolean was used, insert it, otherwise delete it.
896 while (!InitBool->use_empty()) // Delete initializations
897 cast<Instruction>(InitBool->use_back())->eraseFromParent();
900 GV->getParent()->getGlobalList().insert(GV, InitBool);
903 // Now the GV is dead, nuke it and the malloc.
904 GV->eraseFromParent();
905 MI->eraseFromParent();
907 // To further other optimizations, loop over all users of NewGV and try to
908 // constant prop them. This will promote GEP instructions with constant
909 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
910 ConstantPropUsersOf(NewGV);
911 if (RepValue != NewGV)
912 ConstantPropUsersOf(RepValue);
917 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
918 /// to make sure that there are no complex uses of V. We permit simple things
919 /// like dereferencing the pointer, but not storing through the address, unless
920 /// it is to the specified global.
921 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Instruction *V,
923 SmallPtrSet<PHINode*, 8> &PHIs) {
924 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
925 if (isa<LoadInst>(*UI) || isa<CmpInst>(*UI)) {
927 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
928 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
929 return false; // Storing the pointer itself... bad.
930 // Otherwise, storing through it, or storing into GV... fine.
931 } else if (isa<GetElementPtrInst>(*UI)) {
932 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(cast<Instruction>(*UI),
935 } else if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
936 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI
939 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
947 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
948 /// somewhere. Transform all uses of the allocation into loads from the
949 /// global and uses of the resultant pointer. Further, delete the store into
950 /// GV. This assumes that these value pass the
951 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
952 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
953 GlobalVariable *GV) {
954 while (!Alloc->use_empty()) {
955 Instruction *U = cast<Instruction>(*Alloc->use_begin());
956 Instruction *InsertPt = U;
957 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
958 // If this is the store of the allocation into the global, remove it.
959 if (SI->getOperand(1) == GV) {
960 SI->eraseFromParent();
963 } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
964 // Insert the load in the corresponding predecessor, not right before the
966 unsigned PredNo = Alloc->use_begin().getOperandNo()/2;
967 InsertPt = PN->getIncomingBlock(PredNo)->getTerminator();
970 // Insert a load from the global, and use it instead of the malloc.
971 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
972 U->replaceUsesOfWith(Alloc, NL);
976 /// GlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
977 /// GV are simple enough to perform HeapSRA, return true.
978 static bool GlobalLoadUsesSimpleEnoughForHeapSRA(GlobalVariable *GV,
980 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;
982 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
983 // We permit two users of the load: setcc comparing against the null
984 // pointer, and a getelementptr of a specific form.
985 for (Value::use_iterator UI = LI->use_begin(), E = LI->use_end(); UI != E;
987 // Comparison against null is ok.
988 if (ICmpInst *ICI = dyn_cast<ICmpInst>(*UI)) {
989 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
994 // getelementptr is also ok, but only a simple form.
995 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) {
996 // Must index into the array and into the struct.
997 if (GEPI->getNumOperands() < 3)
1000 // Otherwise the GEP is ok.
1004 if (PHINode *PN = dyn_cast<PHINode>(*UI)) {
1005 // We have a phi of a load from the global. We can only handle this
1006 // if the other PHI'd values are actually the same. In this case,
1007 // the rewriter will just drop the phi entirely.
1008 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1009 Value *IV = PN->getIncomingValue(i);
1010 if (IV == LI) continue; // Trivial the same.
1012 // If the phi'd value is from the malloc that initializes the value,
1014 if (IV == MI) continue;
1016 // Otherwise, we don't know what it is.
1022 // Otherwise we don't know what this is, not ok.
1029 /// GetHeapSROALoad - Return the load for the specified field of the HeapSROA'd
1030 /// value, lazily creating it on demand.
1031 static Value *GetHeapSROALoad(Instruction *Load, unsigned FieldNo,
1032 const std::vector<GlobalVariable*> &FieldGlobals,
1033 std::vector<Value *> &InsertedLoadsForPtr) {
1034 if (InsertedLoadsForPtr.size() <= FieldNo)
1035 InsertedLoadsForPtr.resize(FieldNo+1);
1036 if (InsertedLoadsForPtr[FieldNo] == 0)
1037 InsertedLoadsForPtr[FieldNo] = new LoadInst(FieldGlobals[FieldNo],
1038 Load->getName()+".f" +
1039 utostr(FieldNo), Load);
1040 return InsertedLoadsForPtr[FieldNo];
1043 /// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1044 /// the load, rewrite the derived value to use the HeapSRoA'd load.
1045 static void RewriteHeapSROALoadUser(LoadInst *Load, Instruction *LoadUser,
1046 const std::vector<GlobalVariable*> &FieldGlobals,
1047 std::vector<Value *> &InsertedLoadsForPtr) {
1048 // If this is a comparison against null, handle it.
1049 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1050 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1051 // If we have a setcc of the loaded pointer, we can use a setcc of any
1054 if (InsertedLoadsForPtr.empty()) {
1055 NPtr = GetHeapSROALoad(Load, 0, FieldGlobals, InsertedLoadsForPtr);
1057 NPtr = InsertedLoadsForPtr.back();
1060 Value *New = new ICmpInst(SCI->getPredicate(), NPtr,
1061 Constant::getNullValue(NPtr->getType()),
1062 SCI->getName(), SCI);
1063 SCI->replaceAllUsesWith(New);
1064 SCI->eraseFromParent();
1068 // Handle 'getelementptr Ptr, Idx, uint FieldNo ...'
1069 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1070 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1071 && "Unexpected GEPI!");
1073 // Load the pointer for this field.
1074 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1075 Value *NewPtr = GetHeapSROALoad(Load, FieldNo,
1076 FieldGlobals, InsertedLoadsForPtr);
1078 // Create the new GEP idx vector.
1079 SmallVector<Value*, 8> GEPIdx;
1080 GEPIdx.push_back(GEPI->getOperand(1));
1081 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1083 Value *NGEPI = GetElementPtrInst::Create(NewPtr, GEPIdx.begin(), GEPIdx.end(),
1084 GEPI->getName(), GEPI);
1085 GEPI->replaceAllUsesWith(NGEPI);
1086 GEPI->eraseFromParent();
1090 // Handle PHI nodes. PHI nodes must be merging in the same values, plus
1091 // potentially the original malloc. Insert phi nodes for each field, then
1092 // process uses of the PHI.
1093 PHINode *PN = cast<PHINode>(LoadUser);
1094 std::vector<Value *> PHIsForField;
1095 PHIsForField.resize(FieldGlobals.size());
1096 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1097 Value *LoadV = GetHeapSROALoad(Load, i, FieldGlobals, InsertedLoadsForPtr);
1099 PHINode *FieldPN = PHINode::Create(LoadV->getType(),
1100 PN->getName()+"."+utostr(i), PN);
1101 // Fill in the predecessor values.
1102 for (unsigned pred = 0, e = PN->getNumIncomingValues(); pred != e; ++pred) {
1103 // Each predecessor either uses the load or the original malloc.
1104 Value *InVal = PN->getIncomingValue(pred);
1105 BasicBlock *BB = PN->getIncomingBlock(pred);
1107 if (isa<MallocInst>(InVal)) {
1108 // Insert a reload from the global in the predecessor.
1109 NewVal = GetHeapSROALoad(BB->getTerminator(), i, FieldGlobals,
1112 NewVal = InsertedLoadsForPtr[i];
1114 FieldPN->addIncoming(NewVal, BB);
1116 PHIsForField[i] = FieldPN;
1119 // Since PHIsForField specifies a phi for every input value, the lazy inserter
1120 // will never insert a load.
1121 while (!PN->use_empty())
1122 RewriteHeapSROALoadUser(Load, PN->use_back(), FieldGlobals, PHIsForField);
1123 PN->eraseFromParent();
1126 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1127 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1128 /// use FieldGlobals instead. All uses of loaded values satisfy
1129 /// GlobalLoadUsesSimpleEnoughForHeapSRA.
1130 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1131 const std::vector<GlobalVariable*> &FieldGlobals) {
1132 std::vector<Value *> InsertedLoadsForPtr;
1133 //InsertedLoadsForPtr.resize(FieldGlobals.size());
1134 while (!Load->use_empty())
1135 RewriteHeapSROALoadUser(Load, Load->use_back(),
1136 FieldGlobals, InsertedLoadsForPtr);
1139 /// PerformHeapAllocSRoA - MI is an allocation of an array of structures. Break
1140 /// it up into multiple allocations of arrays of the fields.
1141 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, MallocInst *MI){
1142 DOUT << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *MI;
1143 const StructType *STy = cast<StructType>(MI->getAllocatedType());
1145 // There is guaranteed to be at least one use of the malloc (storing
1146 // it into GV). If there are other uses, change them to be uses of
1147 // the global to simplify later code. This also deletes the store
1149 ReplaceUsesOfMallocWithGlobal(MI, GV);
1151 // Okay, at this point, there are no users of the malloc. Insert N
1152 // new mallocs at the same place as MI, and N globals.
1153 std::vector<GlobalVariable*> FieldGlobals;
1154 std::vector<MallocInst*> FieldMallocs;
1156 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1157 const Type *FieldTy = STy->getElementType(FieldNo);
1158 const Type *PFieldTy = PointerType::getUnqual(FieldTy);
1160 GlobalVariable *NGV =
1161 new GlobalVariable(PFieldTy, false, GlobalValue::InternalLinkage,
1162 Constant::getNullValue(PFieldTy),
1163 GV->getName() + ".f" + utostr(FieldNo), GV,
1164 GV->isThreadLocal());
1165 FieldGlobals.push_back(NGV);
1167 MallocInst *NMI = new MallocInst(FieldTy, MI->getArraySize(),
1168 MI->getName() + ".f" + utostr(FieldNo),MI);
1169 FieldMallocs.push_back(NMI);
1170 new StoreInst(NMI, NGV, MI);
1173 // The tricky aspect of this transformation is handling the case when malloc
1174 // fails. In the original code, malloc failing would set the result pointer
1175 // of malloc to null. In this case, some mallocs could succeed and others
1176 // could fail. As such, we emit code that looks like this:
1177 // F0 = malloc(field0)
1178 // F1 = malloc(field1)
1179 // F2 = malloc(field2)
1180 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1181 // if (F0) { free(F0); F0 = 0; }
1182 // if (F1) { free(F1); F1 = 0; }
1183 // if (F2) { free(F2); F2 = 0; }
1185 Value *RunningOr = 0;
1186 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1187 Value *Cond = new ICmpInst(ICmpInst::ICMP_EQ, FieldMallocs[i],
1188 Constant::getNullValue(FieldMallocs[i]->getType()),
1191 RunningOr = Cond; // First seteq
1193 RunningOr = BinaryOperator::createOr(RunningOr, Cond, "tmp", MI);
1196 // Split the basic block at the old malloc.
1197 BasicBlock *OrigBB = MI->getParent();
1198 BasicBlock *ContBB = OrigBB->splitBasicBlock(MI, "malloc_cont");
1200 // Create the block to check the first condition. Put all these blocks at the
1201 // end of the function as they are unlikely to be executed.
1202 BasicBlock *NullPtrBlock = BasicBlock::Create("malloc_ret_null",
1203 OrigBB->getParent());
1205 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1206 // branch on RunningOr.
1207 OrigBB->getTerminator()->eraseFromParent();
1208 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1210 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1211 // pointer, because some may be null while others are not.
1212 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1213 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1214 Value *Cmp = new ICmpInst(ICmpInst::ICMP_NE, GVVal,
1215 Constant::getNullValue(GVVal->getType()),
1216 "tmp", NullPtrBlock);
1217 BasicBlock *FreeBlock = BasicBlock::Create("free_it", OrigBB->getParent());
1218 BasicBlock *NextBlock = BasicBlock::Create("next", OrigBB->getParent());
1219 BranchInst::Create(FreeBlock, NextBlock, Cmp, NullPtrBlock);
1221 // Fill in FreeBlock.
1222 new FreeInst(GVVal, FreeBlock);
1223 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1225 BranchInst::Create(NextBlock, FreeBlock);
1227 NullPtrBlock = NextBlock;
1230 BranchInst::Create(ContBB, NullPtrBlock);
1232 // MI is no longer needed, remove it.
1233 MI->eraseFromParent();
1236 // Okay, the malloc site is completely handled. All of the uses of GV are now
1237 // loads, and all uses of those loads are simple. Rewrite them to use loads
1238 // of the per-field globals instead.
1239 while (!GV->use_empty()) {
1240 if (LoadInst *LI = dyn_cast<LoadInst>(GV->use_back())) {
1241 RewriteUsesOfLoadForHeapSRoA(LI, FieldGlobals);
1242 LI->eraseFromParent();
1244 // Must be a store of null.
1245 StoreInst *SI = cast<StoreInst>(GV->use_back());
1246 assert(isa<Constant>(SI->getOperand(0)) &&
1247 cast<Constant>(SI->getOperand(0))->isNullValue() &&
1248 "Unexpected heap-sra user!");
1250 // Insert a store of null into each global.
1251 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1253 Constant::getNullValue(FieldGlobals[i]->getType()->getElementType());
1254 new StoreInst(Null, FieldGlobals[i], SI);
1256 // Erase the original store.
1257 SI->eraseFromParent();
1261 // The old global is now dead, remove it.
1262 GV->eraseFromParent();
1265 return FieldGlobals[0];
1269 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1270 // that only one value (besides its initializer) is ever stored to the global.
1271 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1272 Module::global_iterator &GVI,
1274 if (CastInst *CI = dyn_cast<CastInst>(StoredOnceVal))
1275 StoredOnceVal = CI->getOperand(0);
1276 else if (GetElementPtrInst *GEPI =dyn_cast<GetElementPtrInst>(StoredOnceVal)){
1277 // "getelementptr Ptr, 0, 0, 0" is really just a cast.
1278 bool IsJustACast = true;
1279 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
1280 if (!isa<Constant>(GEPI->getOperand(i)) ||
1281 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
1282 IsJustACast = false;
1286 StoredOnceVal = GEPI->getOperand(0);
1289 // If we are dealing with a pointer global that is initialized to null and
1290 // only has one (non-null) value stored into it, then we can optimize any
1291 // users of the loaded value (often calls and loads) that would trap if the
1293 if (isa<PointerType>(GV->getInitializer()->getType()) &&
1294 GV->getInitializer()->isNullValue()) {
1295 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1296 if (GV->getInitializer()->getType() != SOVC->getType())
1297 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1299 // Optimize away any trapping uses of the loaded value.
1300 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1302 } else if (MallocInst *MI = dyn_cast<MallocInst>(StoredOnceVal)) {
1303 // If this is a malloc of an abstract type, don't touch it.
1304 if (!MI->getAllocatedType()->isSized())
1307 // We can't optimize this global unless all uses of it are *known* to be
1308 // of the malloc value, not of the null initializer value (consider a use
1309 // that compares the global's value against zero to see if the malloc has
1310 // been reached). To do this, we check to see if all uses of the global
1311 // would trap if the global were null: this proves that they must all
1312 // happen after the malloc.
1313 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1316 // We can't optimize this if the malloc itself is used in a complex way,
1317 // for example, being stored into multiple globals. This allows the
1318 // malloc to be stored into the specified global, loaded setcc'd, and
1319 // GEP'd. These are all things we could transform to using the global
1322 SmallPtrSet<PHINode*, 8> PHIs;
1323 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(MI, GV, PHIs))
1328 // If we have a global that is only initialized with a fixed size malloc,
1329 // transform the program to use global memory instead of malloc'd memory.
1330 // This eliminates dynamic allocation, avoids an indirection accessing the
1331 // data, and exposes the resultant global to further GlobalOpt.
1332 if (ConstantInt *NElements = dyn_cast<ConstantInt>(MI->getArraySize())) {
1333 // Restrict this transformation to only working on small allocations
1334 // (2048 bytes currently), as we don't want to introduce a 16M global or
1336 if (NElements->getZExtValue()*
1337 TD.getABITypeSize(MI->getAllocatedType()) < 2048) {
1338 GVI = OptimizeGlobalAddressOfMalloc(GV, MI);
1343 // If the allocation is an array of structures, consider transforming this
1344 // into multiple malloc'd arrays, one for each field. This is basically
1345 // SRoA for malloc'd memory.
1346 if (const StructType *AllocTy =
1347 dyn_cast<StructType>(MI->getAllocatedType())) {
1348 // This the structure has an unreasonable number of fields, leave it
1350 if (AllocTy->getNumElements() <= 16 && AllocTy->getNumElements() > 0 &&
1351 GlobalLoadUsesSimpleEnoughForHeapSRA(GV, MI)) {
1352 GVI = PerformHeapAllocSRoA(GV, MI);
1362 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1363 /// two values ever stored into GV are its initializer and OtherVal. See if we
1364 /// can shrink the global into a boolean and select between the two values
1365 /// whenever it is used. This exposes the values to other scalar optimizations.
1366 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1367 const Type *GVElType = GV->getType()->getElementType();
1369 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1370 // an FP value or vector, don't do this optimization because a select between
1371 // them is very expensive and unlikely to lead to later simplification.
1372 if (GVElType == Type::Int1Ty || GVElType->isFloatingPoint() ||
1373 isa<VectorType>(GVElType))
1376 // Walk the use list of the global seeing if all the uses are load or store.
1377 // If there is anything else, bail out.
1378 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I)
1379 if (!isa<LoadInst>(I) && !isa<StoreInst>(I))
1382 DOUT << " *** SHRINKING TO BOOL: " << *GV;
1384 // Create the new global, initializing it to false.
1385 GlobalVariable *NewGV = new GlobalVariable(Type::Int1Ty, false,
1386 GlobalValue::InternalLinkage, ConstantInt::getFalse(),
1389 GV->isThreadLocal());
1390 GV->getParent()->getGlobalList().insert(GV, NewGV);
1392 Constant *InitVal = GV->getInitializer();
1393 assert(InitVal->getType() != Type::Int1Ty && "No reason to shrink to bool!");
1395 // If initialized to zero and storing one into the global, we can use a cast
1396 // instead of a select to synthesize the desired value.
1397 bool IsOneZero = false;
1398 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1399 IsOneZero = InitVal->isNullValue() && CI->isOne();
1401 while (!GV->use_empty()) {
1402 Instruction *UI = cast<Instruction>(GV->use_back());
1403 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1404 // Change the store into a boolean store.
1405 bool StoringOther = SI->getOperand(0) == OtherVal;
1406 // Only do this if we weren't storing a loaded value.
1408 if (StoringOther || SI->getOperand(0) == InitVal)
1409 StoreVal = ConstantInt::get(Type::Int1Ty, StoringOther);
1411 // Otherwise, we are storing a previously loaded copy. To do this,
1412 // change the copy from copying the original value to just copying the
1414 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1416 // If we're already replaced the input, StoredVal will be a cast or
1417 // select instruction. If not, it will be a load of the original
1419 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1420 assert(LI->getOperand(0) == GV && "Not a copy!");
1421 // Insert a new load, to preserve the saved value.
1422 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1424 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1425 "This is not a form that we understand!");
1426 StoreVal = StoredVal->getOperand(0);
1427 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1430 new StoreInst(StoreVal, NewGV, SI);
1432 // Change the load into a load of bool then a select.
1433 LoadInst *LI = cast<LoadInst>(UI);
1434 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1437 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1439 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1441 LI->replaceAllUsesWith(NSI);
1443 UI->eraseFromParent();
1446 GV->eraseFromParent();
1451 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1452 /// it if possible. If we make a change, return true.
1453 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1454 Module::global_iterator &GVI) {
1455 std::set<PHINode*> PHIUsers;
1457 GV->removeDeadConstantUsers();
1459 if (GV->use_empty()) {
1460 DOUT << "GLOBAL DEAD: " << *GV;
1461 GV->eraseFromParent();
1466 if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
1468 cerr << "Global: " << *GV;
1469 cerr << " isLoaded = " << GS.isLoaded << "\n";
1470 cerr << " StoredType = ";
1471 switch (GS.StoredType) {
1472 case GlobalStatus::NotStored: cerr << "NEVER STORED\n"; break;
1473 case GlobalStatus::isInitializerStored: cerr << "INIT STORED\n"; break;
1474 case GlobalStatus::isStoredOnce: cerr << "STORED ONCE\n"; break;
1475 case GlobalStatus::isStored: cerr << "stored\n"; break;
1477 if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
1478 cerr << " StoredOnceValue = " << *GS.StoredOnceValue << "\n";
1479 if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
1480 cerr << " AccessingFunction = " << GS.AccessingFunction->getName()
1482 cerr << " HasMultipleAccessingFunctions = "
1483 << GS.HasMultipleAccessingFunctions << "\n";
1484 cerr << " HasNonInstructionUser = " << GS.HasNonInstructionUser<<"\n";
1488 // If this is a first class global and has only one accessing function
1489 // and this function is main (which we know is not recursive we can make
1490 // this global a local variable) we replace the global with a local alloca
1491 // in this function.
1493 // NOTE: It doesn't make sense to promote non first class types since we
1494 // are just replacing static memory to stack memory.
1495 if (!GS.HasMultipleAccessingFunctions &&
1496 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1497 GV->getType()->getElementType()->isFirstClassType() &&
1498 GS.AccessingFunction->getName() == "main" &&
1499 GS.AccessingFunction->hasExternalLinkage()) {
1500 DOUT << "LOCALIZING GLOBAL: " << *GV;
1501 Instruction* FirstI = GS.AccessingFunction->getEntryBlock().begin();
1502 const Type* ElemTy = GV->getType()->getElementType();
1503 // FIXME: Pass Global's alignment when globals have alignment
1504 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), FirstI);
1505 if (!isa<UndefValue>(GV->getInitializer()))
1506 new StoreInst(GV->getInitializer(), Alloca, FirstI);
1508 GV->replaceAllUsesWith(Alloca);
1509 GV->eraseFromParent();
1514 // If the global is never loaded (but may be stored to), it is dead.
1517 DOUT << "GLOBAL NEVER LOADED: " << *GV;
1519 // Delete any stores we can find to the global. We may not be able to
1520 // make it completely dead though.
1521 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1523 // If the global is dead now, delete it.
1524 if (GV->use_empty()) {
1525 GV->eraseFromParent();
1531 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1532 DOUT << "MARKING CONSTANT: " << *GV;
1533 GV->setConstant(true);
1535 // Clean up any obviously simplifiable users now.
1536 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1538 // If the global is dead now, just nuke it.
1539 if (GV->use_empty()) {
1540 DOUT << " *** Marking constant allowed us to simplify "
1541 << "all users and delete global!\n";
1542 GV->eraseFromParent();
1548 } else if (!GV->getInitializer()->getType()->isFirstClassType()) {
1549 if (GlobalVariable *FirstNewGV = SRAGlobal(GV,
1550 getAnalysis<TargetData>())) {
1551 GVI = FirstNewGV; // Don't skip the newly produced globals!
1554 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1555 // If the initial value for the global was an undef value, and if only
1556 // one other value was stored into it, we can just change the
1557 // initializer to be an undef value, then delete all stores to the
1558 // global. This allows us to mark it constant.
1559 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1560 if (isa<UndefValue>(GV->getInitializer())) {
1561 // Change the initial value here.
1562 GV->setInitializer(SOVConstant);
1564 // Clean up any obviously simplifiable users now.
1565 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1567 if (GV->use_empty()) {
1568 DOUT << " *** Substituting initializer allowed us to "
1569 << "simplify all users and delete global!\n";
1570 GV->eraseFromParent();
1579 // Try to optimize globals based on the knowledge that only one value
1580 // (besides its initializer) is ever stored to the global.
1581 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1582 getAnalysis<TargetData>()))
1585 // Otherwise, if the global was not a boolean, we can shrink it to be a
1587 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1588 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1597 /// OnlyCalledDirectly - Return true if the specified function is only called
1598 /// directly. In other words, its address is never taken.
1599 static bool OnlyCalledDirectly(Function *F) {
1600 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1601 Instruction *User = dyn_cast<Instruction>(*UI);
1602 if (!User) return false;
1603 if (!isa<CallInst>(User) && !isa<InvokeInst>(User)) return false;
1605 // See if the function address is passed as an argument.
1606 for (unsigned i = 1, e = User->getNumOperands(); i != e; ++i)
1607 if (User->getOperand(i) == F) return false;
1612 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1613 /// function, changing them to FastCC.
1614 static void ChangeCalleesToFastCall(Function *F) {
1615 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1616 CallSite User(cast<Instruction>(*UI));
1617 User.setCallingConv(CallingConv::Fast);
1621 static PAListPtr StripNest(const PAListPtr &Attrs) {
1622 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1623 if ((Attrs.getSlot(i).Attrs & ParamAttr::Nest) == 0)
1626 // There can be only one.
1627 return Attrs.removeAttr(Attrs.getSlot(i).Index, ParamAttr::Nest);
1633 static void RemoveNestAttribute(Function *F) {
1634 F->setParamAttrs(StripNest(F->getParamAttrs()));
1635 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1636 CallSite User(cast<Instruction>(*UI));
1637 User.setParamAttrs(StripNest(User.getParamAttrs()));
1641 bool GlobalOpt::OptimizeFunctions(Module &M) {
1642 bool Changed = false;
1643 // Optimize functions.
1644 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1646 F->removeDeadConstantUsers();
1647 if (F->use_empty() && (F->hasInternalLinkage() ||
1648 F->hasLinkOnceLinkage())) {
1649 M.getFunctionList().erase(F);
1652 } else if (F->hasInternalLinkage()) {
1653 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1654 OnlyCalledDirectly(F)) {
1655 // If this function has C calling conventions, is not a varargs
1656 // function, and is only called directly, promote it to use the Fast
1657 // calling convention.
1658 F->setCallingConv(CallingConv::Fast);
1659 ChangeCalleesToFastCall(F);
1664 if (F->getParamAttrs().hasAttrSomewhere(ParamAttr::Nest) &&
1665 OnlyCalledDirectly(F)) {
1666 // The function is not used by a trampoline intrinsic, so it is safe
1667 // to remove the 'nest' attribute.
1668 RemoveNestAttribute(F);
1677 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1678 bool Changed = false;
1679 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1681 GlobalVariable *GV = GVI++;
1682 if (!GV->isConstant() && GV->hasInternalLinkage() &&
1683 GV->hasInitializer())
1684 Changed |= ProcessInternalGlobal(GV, GVI);
1689 /// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
1690 /// initializers have an init priority of 65535.
1691 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1692 for (Module::global_iterator I = M.global_begin(), E = M.global_end();
1694 if (I->getName() == "llvm.global_ctors") {
1695 // Found it, verify it's an array of { int, void()* }.
1696 const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
1698 const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1699 if (!STy || STy->getNumElements() != 2 ||
1700 STy->getElementType(0) != Type::Int32Ty) return 0;
1701 const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
1702 if (!PFTy) return 0;
1703 const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
1704 if (!FTy || FTy->getReturnType() != Type::VoidTy || FTy->isVarArg() ||
1705 FTy->getNumParams() != 0)
1708 // Verify that the initializer is simple enough for us to handle.
1709 if (!I->hasInitializer()) return 0;
1710 ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
1712 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1713 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(CA->getOperand(i))) {
1714 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1717 // Must have a function or null ptr.
1718 if (!isa<Function>(CS->getOperand(1)))
1721 // Init priority must be standard.
1722 ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
1723 if (!CI || CI->getZExtValue() != 65535)
1734 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1735 /// return a list of the functions and null terminator as a vector.
1736 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1737 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1738 std::vector<Function*> Result;
1739 Result.reserve(CA->getNumOperands());
1740 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) {
1741 ConstantStruct *CS = cast<ConstantStruct>(CA->getOperand(i));
1742 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1747 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1748 /// specified array, returning the new global to use.
1749 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1750 const std::vector<Function*> &Ctors) {
1751 // If we made a change, reassemble the initializer list.
1752 std::vector<Constant*> CSVals;
1753 CSVals.push_back(ConstantInt::get(Type::Int32Ty, 65535));
1754 CSVals.push_back(0);
1756 // Create the new init list.
1757 std::vector<Constant*> CAList;
1758 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
1760 CSVals[1] = Ctors[i];
1762 const Type *FTy = FunctionType::get(Type::VoidTy,
1763 std::vector<const Type*>(), false);
1764 const PointerType *PFTy = PointerType::getUnqual(FTy);
1765 CSVals[1] = Constant::getNullValue(PFTy);
1766 CSVals[0] = ConstantInt::get(Type::Int32Ty, 2147483647);
1768 CAList.push_back(ConstantStruct::get(CSVals));
1771 // Create the array initializer.
1772 const Type *StructTy =
1773 cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
1774 Constant *CA = ConstantArray::get(ArrayType::get(StructTy, CAList.size()),
1777 // If we didn't change the number of elements, don't create a new GV.
1778 if (CA->getType() == GCL->getInitializer()->getType()) {
1779 GCL->setInitializer(CA);
1783 // Create the new global and insert it next to the existing list.
1784 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
1785 GCL->getLinkage(), CA, "",
1787 GCL->isThreadLocal());
1788 GCL->getParent()->getGlobalList().insert(GCL, NGV);
1791 // Nuke the old list, replacing any uses with the new one.
1792 if (!GCL->use_empty()) {
1794 if (V->getType() != GCL->getType())
1795 V = ConstantExpr::getBitCast(V, GCL->getType());
1796 GCL->replaceAllUsesWith(V);
1798 GCL->eraseFromParent();
1807 static Constant *getVal(std::map<Value*, Constant*> &ComputedValues,
1809 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
1810 Constant *R = ComputedValues[V];
1811 assert(R && "Reference to an uncomputed value!");
1815 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
1816 /// enough for us to understand. In particular, if it is a cast of something,
1817 /// we punt. We basically just support direct accesses to globals and GEP's of
1818 /// globals. This should be kept up to date with CommitValueTo.
1819 static bool isSimpleEnoughPointerToCommit(Constant *C) {
1820 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
1821 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1822 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1823 return !GV->isDeclaration(); // reject external globals.
1825 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
1826 // Handle a constantexpr gep.
1827 if (CE->getOpcode() == Instruction::GetElementPtr &&
1828 isa<GlobalVariable>(CE->getOperand(0))) {
1829 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1830 if (!GV->hasExternalLinkage() && !GV->hasInternalLinkage())
1831 return false; // do not allow weak/linkonce/dllimport/dllexport linkage.
1832 return GV->hasInitializer() &&
1833 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1838 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
1839 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
1840 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
1841 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
1842 ConstantExpr *Addr, unsigned OpNo) {
1843 // Base case of the recursion.
1844 if (OpNo == Addr->getNumOperands()) {
1845 assert(Val->getType() == Init->getType() && "Type mismatch!");
1849 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
1850 std::vector<Constant*> Elts;
1852 // Break up the constant into its elements.
1853 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
1854 for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i)
1855 Elts.push_back(CS->getOperand(i));
1856 } else if (isa<ConstantAggregateZero>(Init)) {
1857 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1858 Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
1859 } else if (isa<UndefValue>(Init)) {
1860 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1861 Elts.push_back(UndefValue::get(STy->getElementType(i)));
1863 assert(0 && "This code is out of sync with "
1864 " ConstantFoldLoadThroughGEPConstantExpr");
1867 // Replace the element that we are supposed to.
1868 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
1869 unsigned Idx = CU->getZExtValue();
1870 assert(Idx < STy->getNumElements() && "Struct index out of range!");
1871 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
1873 // Return the modified struct.
1874 return ConstantStruct::get(&Elts[0], Elts.size(), STy->isPacked());
1876 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
1877 const ArrayType *ATy = cast<ArrayType>(Init->getType());
1879 // Break up the array into elements.
1880 std::vector<Constant*> Elts;
1881 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
1882 for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i)
1883 Elts.push_back(CA->getOperand(i));
1884 } else if (isa<ConstantAggregateZero>(Init)) {
1885 Constant *Elt = Constant::getNullValue(ATy->getElementType());
1886 Elts.assign(ATy->getNumElements(), Elt);
1887 } else if (isa<UndefValue>(Init)) {
1888 Constant *Elt = UndefValue::get(ATy->getElementType());
1889 Elts.assign(ATy->getNumElements(), Elt);
1891 assert(0 && "This code is out of sync with "
1892 " ConstantFoldLoadThroughGEPConstantExpr");
1895 assert(CI->getZExtValue() < ATy->getNumElements());
1896 Elts[CI->getZExtValue()] =
1897 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
1898 return ConstantArray::get(ATy, Elts);
1902 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
1903 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
1904 static void CommitValueTo(Constant *Val, Constant *Addr) {
1905 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
1906 assert(GV->hasInitializer());
1907 GV->setInitializer(Val);
1911 ConstantExpr *CE = cast<ConstantExpr>(Addr);
1912 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1914 Constant *Init = GV->getInitializer();
1915 Init = EvaluateStoreInto(Init, Val, CE, 2);
1916 GV->setInitializer(Init);
1919 /// ComputeLoadResult - Return the value that would be computed by a load from
1920 /// P after the stores reflected by 'memory' have been performed. If we can't
1921 /// decide, return null.
1922 static Constant *ComputeLoadResult(Constant *P,
1923 const std::map<Constant*, Constant*> &Memory) {
1924 // If this memory location has been recently stored, use the stored value: it
1925 // is the most up-to-date.
1926 std::map<Constant*, Constant*>::const_iterator I = Memory.find(P);
1927 if (I != Memory.end()) return I->second;
1930 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
1931 if (GV->hasInitializer())
1932 return GV->getInitializer();
1936 // Handle a constantexpr getelementptr.
1937 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
1938 if (CE->getOpcode() == Instruction::GetElementPtr &&
1939 isa<GlobalVariable>(CE->getOperand(0))) {
1940 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
1941 if (GV->hasInitializer())
1942 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
1945 return 0; // don't know how to evaluate.
1948 /// EvaluateFunction - Evaluate a call to function F, returning true if
1949 /// successful, false if we can't evaluate it. ActualArgs contains the formal
1950 /// arguments for the function.
1951 static bool EvaluateFunction(Function *F, Constant *&RetVal,
1952 const std::vector<Constant*> &ActualArgs,
1953 std::vector<Function*> &CallStack,
1954 std::map<Constant*, Constant*> &MutatedMemory,
1955 std::vector<GlobalVariable*> &AllocaTmps) {
1956 // Check to see if this function is already executing (recursion). If so,
1957 // bail out. TODO: we might want to accept limited recursion.
1958 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
1961 CallStack.push_back(F);
1963 /// Values - As we compute SSA register values, we store their contents here.
1964 std::map<Value*, Constant*> Values;
1966 // Initialize arguments to the incoming values specified.
1968 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
1970 Values[AI] = ActualArgs[ArgNo];
1972 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
1973 /// we can only evaluate any one basic block at most once. This set keeps
1974 /// track of what we have executed so we can detect recursive cases etc.
1975 std::set<BasicBlock*> ExecutedBlocks;
1977 // CurInst - The current instruction we're evaluating.
1978 BasicBlock::iterator CurInst = F->begin()->begin();
1980 // This is the main evaluation loop.
1982 Constant *InstResult = 0;
1984 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
1985 if (SI->isVolatile()) return false; // no volatile accesses.
1986 Constant *Ptr = getVal(Values, SI->getOperand(1));
1987 if (!isSimpleEnoughPointerToCommit(Ptr))
1988 // If this is too complex for us to commit, reject it.
1990 Constant *Val = getVal(Values, SI->getOperand(0));
1991 MutatedMemory[Ptr] = Val;
1992 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
1993 InstResult = ConstantExpr::get(BO->getOpcode(),
1994 getVal(Values, BO->getOperand(0)),
1995 getVal(Values, BO->getOperand(1)));
1996 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
1997 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
1998 getVal(Values, CI->getOperand(0)),
1999 getVal(Values, CI->getOperand(1)));
2000 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2001 InstResult = ConstantExpr::getCast(CI->getOpcode(),
2002 getVal(Values, CI->getOperand(0)),
2004 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2005 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
2006 getVal(Values, SI->getOperand(1)),
2007 getVal(Values, SI->getOperand(2)));
2008 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2009 Constant *P = getVal(Values, GEP->getOperand(0));
2010 SmallVector<Constant*, 8> GEPOps;
2011 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
2012 GEPOps.push_back(getVal(Values, GEP->getOperand(i)));
2013 InstResult = ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size());
2014 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2015 if (LI->isVolatile()) return false; // no volatile accesses.
2016 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
2018 if (InstResult == 0) return false; // Could not evaluate load.
2019 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2020 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs.
2021 const Type *Ty = AI->getType()->getElementType();
2022 AllocaTmps.push_back(new GlobalVariable(Ty, false,
2023 GlobalValue::InternalLinkage,
2024 UndefValue::get(Ty),
2026 InstResult = AllocaTmps.back();
2027 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
2028 // Cannot handle inline asm.
2029 if (isa<InlineAsm>(CI->getOperand(0))) return false;
2031 // Resolve function pointers.
2032 Function *Callee = dyn_cast<Function>(getVal(Values, CI->getOperand(0)));
2033 if (!Callee) return false; // Cannot resolve.
2035 std::vector<Constant*> Formals;
2036 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
2037 Formals.push_back(getVal(Values, CI->getOperand(i)));
2039 if (Callee->isDeclaration()) {
2040 // If this is a function we can constant fold, do it.
2041 if (Constant *C = ConstantFoldCall(Callee, &Formals[0],
2048 if (Callee->getFunctionType()->isVarArg())
2053 // Execute the call, if successful, use the return value.
2054 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
2055 MutatedMemory, AllocaTmps))
2057 InstResult = RetVal;
2059 } else if (isa<TerminatorInst>(CurInst)) {
2060 BasicBlock *NewBB = 0;
2061 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2062 if (BI->isUnconditional()) {
2063 NewBB = BI->getSuccessor(0);
2066 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
2067 if (!Cond) return false; // Cannot determine.
2069 NewBB = BI->getSuccessor(!Cond->getZExtValue());
2071 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2073 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
2074 if (!Val) return false; // Cannot determine.
2075 NewBB = SI->getSuccessor(SI->findCaseValue(Val));
2076 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
2077 if (RI->getNumOperands())
2078 RetVal = getVal(Values, RI->getOperand(0));
2080 CallStack.pop_back(); // return from fn.
2081 return true; // We succeeded at evaluating this ctor!
2083 // invoke, unwind, unreachable.
2084 return false; // Cannot handle this terminator.
2087 // Okay, we succeeded in evaluating this control flow. See if we have
2088 // executed the new block before. If so, we have a looping function,
2089 // which we cannot evaluate in reasonable time.
2090 if (!ExecutedBlocks.insert(NewBB).second)
2091 return false; // looped!
2093 // Okay, we have never been in this block before. Check to see if there
2094 // are any PHI nodes. If so, evaluate them with information about where
2096 BasicBlock *OldBB = CurInst->getParent();
2097 CurInst = NewBB->begin();
2099 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2100 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
2102 // Do NOT increment CurInst. We know that the terminator had no value.
2105 // Did not know how to evaluate this!
2109 if (!CurInst->use_empty())
2110 Values[CurInst] = InstResult;
2112 // Advance program counter.
2117 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2118 /// we can. Return true if we can, false otherwise.
2119 static bool EvaluateStaticConstructor(Function *F) {
2120 /// MutatedMemory - For each store we execute, we update this map. Loads
2121 /// check this to get the most up-to-date value. If evaluation is successful,
2122 /// this state is committed to the process.
2123 std::map<Constant*, Constant*> MutatedMemory;
2125 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2126 /// to represent its body. This vector is needed so we can delete the
2127 /// temporary globals when we are done.
2128 std::vector<GlobalVariable*> AllocaTmps;
2130 /// CallStack - This is used to detect recursion. In pathological situations
2131 /// we could hit exponential behavior, but at least there is nothing
2133 std::vector<Function*> CallStack;
2135 // Call the function.
2136 Constant *RetValDummy;
2137 bool EvalSuccess = EvaluateFunction(F, RetValDummy, std::vector<Constant*>(),
2138 CallStack, MutatedMemory, AllocaTmps);
2140 // We succeeded at evaluation: commit the result.
2141 DOUT << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2142 << F->getName() << "' to " << MutatedMemory.size()
2144 for (std::map<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2145 E = MutatedMemory.end(); I != E; ++I)
2146 CommitValueTo(I->second, I->first);
2149 // At this point, we are done interpreting. If we created any 'alloca'
2150 // temporaries, release them now.
2151 while (!AllocaTmps.empty()) {
2152 GlobalVariable *Tmp = AllocaTmps.back();
2153 AllocaTmps.pop_back();
2155 // If there are still users of the alloca, the program is doing something
2156 // silly, e.g. storing the address of the alloca somewhere and using it
2157 // later. Since this is undefined, we'll just make it be null.
2158 if (!Tmp->use_empty())
2159 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2168 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2169 /// Return true if anything changed.
2170 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2171 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2172 bool MadeChange = false;
2173 if (Ctors.empty()) return false;
2175 // Loop over global ctors, optimizing them when we can.
2176 for (unsigned i = 0; i != Ctors.size(); ++i) {
2177 Function *F = Ctors[i];
2178 // Found a null terminator in the middle of the list, prune off the rest of
2181 if (i != Ctors.size()-1) {
2188 // We cannot simplify external ctor functions.
2189 if (F->empty()) continue;
2191 // If we can evaluate the ctor at compile time, do.
2192 if (EvaluateStaticConstructor(F)) {
2193 Ctors.erase(Ctors.begin()+i);
2196 ++NumCtorsEvaluated;
2201 if (!MadeChange) return false;
2203 GCL = InstallGlobalCtors(GCL, Ctors);
2208 bool GlobalOpt::runOnModule(Module &M) {
2209 bool Changed = false;
2211 // Try to find the llvm.globalctors list.
2212 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2214 bool LocalChange = true;
2215 while (LocalChange) {
2216 LocalChange = false;
2218 // Delete functions that are trivially dead, ccc -> fastcc
2219 LocalChange |= OptimizeFunctions(M);
2221 // Optimize global_ctors list.
2223 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2225 // Optimize non-address-taken globals.
2226 LocalChange |= OptimizeGlobalVars(M);
2227 Changed |= LocalChange;
2230 // TODO: Move all global ctors functions to the end of the module for code