1 //===- GlobalOpt.cpp - Optimize Global Variables --------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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.
13 //
14 //===----------------------------------------------------------------------===//
15
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/Operator.h"
25 #include "llvm/Pass.h"
26 #include "llvm/Analysis/ConstantFolding.h"
27 #include "llvm/Analysis/MemoryBuiltins.h"
28 #include "llvm/Target/TargetData.h"
29 #include "llvm/Support/CallSite.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/GetElementPtrTypeIterator.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/ADT/DenseMap.h"
36 #include "llvm/ADT/SmallPtrSet.h"
37 #include "llvm/ADT/SmallVector.h"
38 #include "llvm/ADT/Statistic.h"
39 #include "llvm/ADT/STLExtras.h"
40 #include <algorithm>
41 using namespace llvm;
42
43 STATISTIC(NumMarked , "Number of globals marked constant");
44 STATISTIC(NumUnnamed , "Number of globals marked unnamed_addr");
45 STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
46 STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
47 STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
48 STATISTIC(NumDeleted , "Number of globals deleted");
49 STATISTIC(NumFnDeleted , "Number of functions deleted");
50 STATISTIC(NumGlobUses , "Number of global uses devirtualized");
51 STATISTIC(NumLocalized , "Number of globals localized");
52 STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans");
53 STATISTIC(NumFastCallFns , "Number of functions converted to fastcc");
54 STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
55 STATISTIC(NumNestRemoved , "Number of nest attributes removed");
56 STATISTIC(NumAliasesResolved, "Number of global aliases resolved");
57 STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated");
58 STATISTIC(NumCXXDtorsRemoved, "Number of global C++ destructors removed");
59
60 namespace {
61 struct GlobalStatus;
62 struct GlobalOpt : public ModulePass {
getAnalysisUsage__anon15e2b1800111::GlobalOpt63 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
64 }
65 static char ID; // Pass identification, replacement for typeid
GlobalOpt__anon15e2b1800111::GlobalOpt66 GlobalOpt() : ModulePass(ID) {
67 initializeGlobalOptPass(*PassRegistry::getPassRegistry());
68 }
69
70 bool runOnModule(Module &M);
71
72 private:
73 GlobalVariable *FindGlobalCtors(Module &M);
74 bool OptimizeFunctions(Module &M);
75 bool OptimizeGlobalVars(Module &M);
76 bool OptimizeGlobalAliases(Module &M);
77 bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
78 bool ProcessGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
79 bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI,
80 const SmallPtrSet<const PHINode*, 16> &PHIUsers,
81 const GlobalStatus &GS);
82 bool OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn);
83 };
84 }
85
86 char GlobalOpt::ID = 0;
87 INITIALIZE_PASS(GlobalOpt, "globalopt",
88 "Global Variable Optimizer", false, false)
89
createGlobalOptimizerPass()90 ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
91
92 namespace {
93
94 /// GlobalStatus - As we analyze each global, keep track of some information
95 /// about it. If we find out that the address of the global is taken, none of
96 /// this info will be accurate.
97 struct GlobalStatus {
98 /// isCompared - True if the global's address is used in a comparison.
99 bool isCompared;
100
101 /// isLoaded - True if the global is ever loaded. If the global isn't ever
102 /// loaded it can be deleted.
103 bool isLoaded;
104
105 /// StoredType - Keep track of what stores to the global look like.
106 ///
107 enum StoredType {
108 /// NotStored - There is no store to this global. It can thus be marked
109 /// constant.
110 NotStored,
111
112 /// isInitializerStored - This global is stored to, but the only thing
113 /// stored is the constant it was initialized with. This is only tracked
114 /// for scalar globals.
115 isInitializerStored,
116
117 /// isStoredOnce - This global is stored to, but only its initializer and
118 /// one other value is ever stored to it. If this global isStoredOnce, we
119 /// track the value stored to it in StoredOnceValue below. This is only
120 /// tracked for scalar globals.
121 isStoredOnce,
122
123 /// isStored - This global is stored to by multiple values or something else
124 /// that we cannot track.
125 isStored
126 } StoredType;
127
128 /// StoredOnceValue - If only one value (besides the initializer constant) is
129 /// ever stored to this global, keep track of what value it is.
130 Value *StoredOnceValue;
131
132 /// AccessingFunction/HasMultipleAccessingFunctions - These start out
133 /// null/false. When the first accessing function is noticed, it is recorded.
134 /// When a second different accessing function is noticed,
135 /// HasMultipleAccessingFunctions is set to true.
136 const Function *AccessingFunction;
137 bool HasMultipleAccessingFunctions;
138
139 /// HasNonInstructionUser - Set to true if this global has a user that is not
140 /// an instruction (e.g. a constant expr or GV initializer).
141 bool HasNonInstructionUser;
142
143 /// HasPHIUser - Set to true if this global has a user that is a PHI node.
144 bool HasPHIUser;
145
GlobalStatus__anon15e2b1800211::GlobalStatus146 GlobalStatus() : isCompared(false), isLoaded(false), StoredType(NotStored),
147 StoredOnceValue(0), AccessingFunction(0),
148 HasMultipleAccessingFunctions(false), HasNonInstructionUser(false),
149 HasPHIUser(false) {}
150 };
151
152 }
153
154 // SafeToDestroyConstant - It is safe to destroy a constant iff it is only used
155 // by constants itself. Note that constants cannot be cyclic, so this test is
156 // pretty easy to implement recursively.
157 //
SafeToDestroyConstant(const Constant * C)158 static bool SafeToDestroyConstant(const Constant *C) {
159 if (isa<GlobalValue>(C)) return false;
160
161 for (Value::const_use_iterator UI = C->use_begin(), E = C->use_end(); UI != E;
162 ++UI)
163 if (const Constant *CU = dyn_cast<Constant>(*UI)) {
164 if (!SafeToDestroyConstant(CU)) return false;
165 } else
166 return false;
167 return true;
168 }
169
170
171 /// AnalyzeGlobal - Look at all uses of the global and fill in the GlobalStatus
172 /// structure. If the global has its address taken, return true to indicate we
173 /// can't do anything with it.
174 ///
AnalyzeGlobal(const Value * V,GlobalStatus & GS,SmallPtrSet<const PHINode *,16> & PHIUsers)175 static bool AnalyzeGlobal(const Value *V, GlobalStatus &GS,
176 SmallPtrSet<const PHINode*, 16> &PHIUsers) {
177 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;
178 ++UI) {
179 const User *U = *UI;
180 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
181 GS.HasNonInstructionUser = true;
182
183 // If the result of the constantexpr isn't pointer type, then we won't
184 // know to expect it in various places. Just reject early.
185 if (!isa<PointerType>(CE->getType())) return true;
186
187 if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
188 } else if (const Instruction *I = dyn_cast<Instruction>(U)) {
189 if (!GS.HasMultipleAccessingFunctions) {
190 const Function *F = I->getParent()->getParent();
191 if (GS.AccessingFunction == 0)
192 GS.AccessingFunction = F;
193 else if (GS.AccessingFunction != F)
194 GS.HasMultipleAccessingFunctions = true;
195 }
196 if (const LoadInst *LI = dyn_cast<LoadInst>(I)) {
197 GS.isLoaded = true;
198 // Don't hack on volatile/atomic loads.
199 if (!LI->isSimple()) return true;
200 } else if (const StoreInst *SI = dyn_cast<StoreInst>(I)) {
201 // Don't allow a store OF the address, only stores TO the address.
202 if (SI->getOperand(0) == V) return true;
203
204 // Don't hack on volatile/atomic stores.
205 if (!SI->isSimple()) return true;
206
207 // If this is a direct store to the global (i.e., the global is a scalar
208 // value, not an aggregate), keep more specific information about
209 // stores.
210 if (GS.StoredType != GlobalStatus::isStored) {
211 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(
212 SI->getOperand(1))) {
213 Value *StoredVal = SI->getOperand(0);
214 if (StoredVal == GV->getInitializer()) {
215 if (GS.StoredType < GlobalStatus::isInitializerStored)
216 GS.StoredType = GlobalStatus::isInitializerStored;
217 } else if (isa<LoadInst>(StoredVal) &&
218 cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
219 if (GS.StoredType < GlobalStatus::isInitializerStored)
220 GS.StoredType = GlobalStatus::isInitializerStored;
221 } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
222 GS.StoredType = GlobalStatus::isStoredOnce;
223 GS.StoredOnceValue = StoredVal;
224 } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
225 GS.StoredOnceValue == StoredVal) {
226 // noop.
227 } else {
228 GS.StoredType = GlobalStatus::isStored;
229 }
230 } else {
231 GS.StoredType = GlobalStatus::isStored;
232 }
233 }
234 } else if (isa<GetElementPtrInst>(I)) {
235 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
236 } else if (isa<SelectInst>(I)) {
237 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
238 } else if (const PHINode *PN = dyn_cast<PHINode>(I)) {
239 // PHI nodes we can check just like select or GEP instructions, but we
240 // have to be careful about infinite recursion.
241 if (PHIUsers.insert(PN)) // Not already visited.
242 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
243 GS.HasPHIUser = true;
244 } else if (isa<CmpInst>(I)) {
245 GS.isCompared = true;
246 } else if (const MemTransferInst *MTI = dyn_cast<MemTransferInst>(I)) {
247 if (MTI->isVolatile()) return true;
248 if (MTI->getArgOperand(0) == V)
249 GS.StoredType = GlobalStatus::isStored;
250 if (MTI->getArgOperand(1) == V)
251 GS.isLoaded = true;
252 } else if (const MemSetInst *MSI = dyn_cast<MemSetInst>(I)) {
253 assert(MSI->getArgOperand(0) == V && "Memset only takes one pointer!");
254 if (MSI->isVolatile()) return true;
255 GS.StoredType = GlobalStatus::isStored;
256 } else {
257 return true; // Any other non-load instruction might take address!
258 }
259 } else if (const Constant *C = dyn_cast<Constant>(U)) {
260 GS.HasNonInstructionUser = true;
261 // We might have a dead and dangling constant hanging off of here.
262 if (!SafeToDestroyConstant(C))
263 return true;
264 } else {
265 GS.HasNonInstructionUser = true;
266 // Otherwise must be some other user.
267 return true;
268 }
269 }
270
271 return false;
272 }
273
getAggregateConstantElement(Constant * Agg,Constant * Idx)274 static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) {
275 ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
276 if (!CI) return 0;
277 unsigned IdxV = CI->getZExtValue();
278
279 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
280 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
281 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
282 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
283 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) {
284 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
285 } else if (isa<ConstantAggregateZero>(Agg)) {
286 if (StructType *STy = dyn_cast<StructType>(Agg->getType())) {
287 if (IdxV < STy->getNumElements())
288 return Constant::getNullValue(STy->getElementType(IdxV));
289 } else if (SequentialType *STy =
290 dyn_cast<SequentialType>(Agg->getType())) {
291 return Constant::getNullValue(STy->getElementType());
292 }
293 } else if (isa<UndefValue>(Agg)) {
294 if (StructType *STy = dyn_cast<StructType>(Agg->getType())) {
295 if (IdxV < STy->getNumElements())
296 return UndefValue::get(STy->getElementType(IdxV));
297 } else if (SequentialType *STy =
298 dyn_cast<SequentialType>(Agg->getType())) {
299 return UndefValue::get(STy->getElementType());
300 }
301 }
302 return 0;
303 }
304
305
306 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
307 /// users of the global, cleaning up the obvious ones. This is largely just a
308 /// quick scan over the use list to clean up the easy and obvious cruft. This
309 /// returns true if it made a change.
CleanupConstantGlobalUsers(Value * V,Constant * Init)310 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
311 bool Changed = false;
312 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
313 User *U = *UI++;
314
315 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
316 if (Init) {
317 // Replace the load with the initializer.
318 LI->replaceAllUsesWith(Init);
319 LI->eraseFromParent();
320 Changed = true;
321 }
322 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
323 // Store must be unreachable or storing Init into the global.
324 SI->eraseFromParent();
325 Changed = true;
326 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
327 if (CE->getOpcode() == Instruction::GetElementPtr) {
328 Constant *SubInit = 0;
329 if (Init)
330 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
331 Changed |= CleanupConstantGlobalUsers(CE, SubInit);
332 } else if (CE->getOpcode() == Instruction::BitCast &&
333 CE->getType()->isPointerTy()) {
334 // Pointer cast, delete any stores and memsets to the global.
335 Changed |= CleanupConstantGlobalUsers(CE, 0);
336 }
337
338 if (CE->use_empty()) {
339 CE->destroyConstant();
340 Changed = true;
341 }
342 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
343 // Do not transform "gepinst (gep constexpr (GV))" here, because forming
344 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
345 // and will invalidate our notion of what Init is.
346 Constant *SubInit = 0;
347 if (!isa<ConstantExpr>(GEP->getOperand(0))) {
348 ConstantExpr *CE =
349 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
350 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
351 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
352 }
353 Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
354
355 if (GEP->use_empty()) {
356 GEP->eraseFromParent();
357 Changed = true;
358 }
359 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
360 if (MI->getRawDest() == V) {
361 MI->eraseFromParent();
362 Changed = true;
363 }
364
365 } else if (Constant *C = dyn_cast<Constant>(U)) {
366 // If we have a chain of dead constantexprs or other things dangling from
367 // us, and if they are all dead, nuke them without remorse.
368 if (SafeToDestroyConstant(C)) {
369 C->destroyConstant();
370 // This could have invalidated UI, start over from scratch.
371 CleanupConstantGlobalUsers(V, Init);
372 return true;
373 }
374 }
375 }
376 return Changed;
377 }
378
379 /// isSafeSROAElementUse - Return true if the specified instruction is a safe
380 /// user of a derived expression from a global that we want to SROA.
isSafeSROAElementUse(Value * V)381 static bool isSafeSROAElementUse(Value *V) {
382 // We might have a dead and dangling constant hanging off of here.
383 if (Constant *C = dyn_cast<Constant>(V))
384 return SafeToDestroyConstant(C);
385
386 Instruction *I = dyn_cast<Instruction>(V);
387 if (!I) return false;
388
389 // Loads are ok.
390 if (isa<LoadInst>(I)) return true;
391
392 // Stores *to* the pointer are ok.
393 if (StoreInst *SI = dyn_cast<StoreInst>(I))
394 return SI->getOperand(0) != V;
395
396 // Otherwise, it must be a GEP.
397 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I);
398 if (GEPI == 0) return false;
399
400 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
401 !cast<Constant>(GEPI->getOperand(1))->isNullValue())
402 return false;
403
404 for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end();
405 I != E; ++I)
406 if (!isSafeSROAElementUse(*I))
407 return false;
408 return true;
409 }
410
411
412 /// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value.
413 /// Look at it and its uses and decide whether it is safe to SROA this global.
414 ///
IsUserOfGlobalSafeForSRA(User * U,GlobalValue * GV)415 static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
416 // The user of the global must be a GEP Inst or a ConstantExpr GEP.
417 if (!isa<GetElementPtrInst>(U) &&
418 (!isa<ConstantExpr>(U) ||
419 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
420 return false;
421
422 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
423 // don't like < 3 operand CE's, and we don't like non-constant integer
424 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some
425 // value of C.
426 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
427 !cast<Constant>(U->getOperand(1))->isNullValue() ||
428 !isa<ConstantInt>(U->getOperand(2)))
429 return false;
430
431 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
432 ++GEPI; // Skip over the pointer index.
433
434 // If this is a use of an array allocation, do a bit more checking for sanity.
435 if (ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
436 uint64_t NumElements = AT->getNumElements();
437 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
438
439 // Check to make sure that index falls within the array. If not,
440 // something funny is going on, so we won't do the optimization.
441 //
442 if (Idx->getZExtValue() >= NumElements)
443 return false;
444
445 // We cannot scalar repl this level of the array unless any array
446 // sub-indices are in-range constants. In particular, consider:
447 // A[0][i]. We cannot know that the user isn't doing invalid things like
448 // allowing i to index an out-of-range subscript that accesses A[1].
449 //
450 // Scalar replacing *just* the outer index of the array is probably not
451 // going to be a win anyway, so just give up.
452 for (++GEPI; // Skip array index.
453 GEPI != E;
454 ++GEPI) {
455 uint64_t NumElements;
456 if (ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
457 NumElements = SubArrayTy->getNumElements();
458 else if (VectorType *SubVectorTy = dyn_cast<VectorType>(*GEPI))
459 NumElements = SubVectorTy->getNumElements();
460 else {
461 assert((*GEPI)->isStructTy() &&
462 "Indexed GEP type is not array, vector, or struct!");
463 continue;
464 }
465
466 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
467 if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
468 return false;
469 }
470 }
471
472 for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I)
473 if (!isSafeSROAElementUse(*I))
474 return false;
475 return true;
476 }
477
478 /// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
479 /// is safe for us to perform this transformation.
480 ///
GlobalUsersSafeToSRA(GlobalValue * GV)481 static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
482 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
483 UI != E; ++UI) {
484 if (!IsUserOfGlobalSafeForSRA(*UI, GV))
485 return false;
486 }
487 return true;
488 }
489
490
491 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
492 /// variable. This opens the door for other optimizations by exposing the
493 /// behavior of the program in a more fine-grained way. We have determined that
494 /// this transformation is safe already. We return the first global variable we
495 /// insert so that the caller can reprocess it.
SRAGlobal(GlobalVariable * GV,const TargetData & TD)496 static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) {
497 // Make sure this global only has simple uses that we can SRA.
498 if (!GlobalUsersSafeToSRA(GV))
499 return 0;
500
501 assert(GV->hasLocalLinkage() && !GV->isConstant());
502 Constant *Init = GV->getInitializer();
503 Type *Ty = Init->getType();
504
505 std::vector<GlobalVariable*> NewGlobals;
506 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
507
508 // Get the alignment of the global, either explicit or target-specific.
509 unsigned StartAlignment = GV->getAlignment();
510 if (StartAlignment == 0)
511 StartAlignment = TD.getABITypeAlignment(GV->getType());
512
513 if (StructType *STy = dyn_cast<StructType>(Ty)) {
514 NewGlobals.reserve(STy->getNumElements());
515 const StructLayout &Layout = *TD.getStructLayout(STy);
516 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
517 Constant *In = getAggregateConstantElement(Init,
518 ConstantInt::get(Type::getInt32Ty(STy->getContext()), i));
519 assert(In && "Couldn't get element of initializer?");
520 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
521 GlobalVariable::InternalLinkage,
522 In, GV->getName()+"."+Twine(i),
523 GV->isThreadLocal(),
524 GV->getType()->getAddressSpace());
525 Globals.insert(GV, NGV);
526 NewGlobals.push_back(NGV);
527
528 // Calculate the known alignment of the field. If the original aggregate
529 // had 256 byte alignment for example, something might depend on that:
530 // propagate info to each field.
531 uint64_t FieldOffset = Layout.getElementOffset(i);
532 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset);
533 if (NewAlign > TD.getABITypeAlignment(STy->getElementType(i)))
534 NGV->setAlignment(NewAlign);
535 }
536 } else if (SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
537 unsigned NumElements = 0;
538 if (ArrayType *ATy = dyn_cast<ArrayType>(STy))
539 NumElements = ATy->getNumElements();
540 else
541 NumElements = cast<VectorType>(STy)->getNumElements();
542
543 if (NumElements > 16 && GV->hasNUsesOrMore(16))
544 return 0; // It's not worth it.
545 NewGlobals.reserve(NumElements);
546
547 uint64_t EltSize = TD.getTypeAllocSize(STy->getElementType());
548 unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType());
549 for (unsigned i = 0, e = NumElements; i != e; ++i) {
550 Constant *In = getAggregateConstantElement(Init,
551 ConstantInt::get(Type::getInt32Ty(Init->getContext()), i));
552 assert(In && "Couldn't get element of initializer?");
553
554 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
555 GlobalVariable::InternalLinkage,
556 In, GV->getName()+"."+Twine(i),
557 GV->isThreadLocal(),
558 GV->getType()->getAddressSpace());
559 Globals.insert(GV, NGV);
560 NewGlobals.push_back(NGV);
561
562 // Calculate the known alignment of the field. If the original aggregate
563 // had 256 byte alignment for example, something might depend on that:
564 // propagate info to each field.
565 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
566 if (NewAlign > EltAlign)
567 NGV->setAlignment(NewAlign);
568 }
569 }
570
571 if (NewGlobals.empty())
572 return 0;
573
574 DEBUG(dbgs() << "PERFORMING GLOBAL SRA ON: " << *GV);
575
576 Constant *NullInt =Constant::getNullValue(Type::getInt32Ty(GV->getContext()));
577
578 // Loop over all of the uses of the global, replacing the constantexpr geps,
579 // with smaller constantexpr geps or direct references.
580 while (!GV->use_empty()) {
581 User *GEP = GV->use_back();
582 assert(((isa<ConstantExpr>(GEP) &&
583 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
584 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
585
586 // Ignore the 1th operand, which has to be zero or else the program is quite
587 // broken (undefined). Get the 2nd operand, which is the structure or array
588 // index.
589 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
590 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
591
592 Value *NewPtr = NewGlobals[Val];
593
594 // Form a shorter GEP if needed.
595 if (GEP->getNumOperands() > 3) {
596 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
597 SmallVector<Constant*, 8> Idxs;
598 Idxs.push_back(NullInt);
599 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
600 Idxs.push_back(CE->getOperand(i));
601 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr), Idxs);
602 } else {
603 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
604 SmallVector<Value*, 8> Idxs;
605 Idxs.push_back(NullInt);
606 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
607 Idxs.push_back(GEPI->getOperand(i));
608 NewPtr = GetElementPtrInst::Create(NewPtr, Idxs,
609 GEPI->getName()+"."+Twine(Val),GEPI);
610 }
611 }
612 GEP->replaceAllUsesWith(NewPtr);
613
614 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
615 GEPI->eraseFromParent();
616 else
617 cast<ConstantExpr>(GEP)->destroyConstant();
618 }
619
620 // Delete the old global, now that it is dead.
621 Globals.erase(GV);
622 ++NumSRA;
623
624 // Loop over the new globals array deleting any globals that are obviously
625 // dead. This can arise due to scalarization of a structure or an array that
626 // has elements that are dead.
627 unsigned FirstGlobal = 0;
628 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
629 if (NewGlobals[i]->use_empty()) {
630 Globals.erase(NewGlobals[i]);
631 if (FirstGlobal == i) ++FirstGlobal;
632 }
633
634 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
635 }
636
637 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
638 /// value will trap if the value is dynamically null. PHIs keeps track of any
639 /// phi nodes we've seen to avoid reprocessing them.
AllUsesOfValueWillTrapIfNull(const Value * V,SmallPtrSet<const PHINode *,8> & PHIs)640 static bool AllUsesOfValueWillTrapIfNull(const Value *V,
641 SmallPtrSet<const PHINode*, 8> &PHIs) {
642 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;
643 ++UI) {
644 const User *U = *UI;
645
646 if (isa<LoadInst>(U)) {
647 // Will trap.
648 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
649 if (SI->getOperand(0) == V) {
650 //cerr << "NONTRAPPING USE: " << *U;
651 return false; // Storing the value.
652 }
653 } else if (const CallInst *CI = dyn_cast<CallInst>(U)) {
654 if (CI->getCalledValue() != V) {
655 //cerr << "NONTRAPPING USE: " << *U;
656 return false; // Not calling the ptr
657 }
658 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(U)) {
659 if (II->getCalledValue() != V) {
660 //cerr << "NONTRAPPING USE: " << *U;
661 return false; // Not calling the ptr
662 }
663 } else if (const BitCastInst *CI = dyn_cast<BitCastInst>(U)) {
664 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
665 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
666 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
667 } else if (const PHINode *PN = dyn_cast<PHINode>(U)) {
668 // If we've already seen this phi node, ignore it, it has already been
669 // checked.
670 if (PHIs.insert(PN) && !AllUsesOfValueWillTrapIfNull(PN, PHIs))
671 return false;
672 } else if (isa<ICmpInst>(U) &&
673 isa<ConstantPointerNull>(UI->getOperand(1))) {
674 // Ignore icmp X, null
675 } else {
676 //cerr << "NONTRAPPING USE: " << *U;
677 return false;
678 }
679 }
680 return true;
681 }
682
683 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
684 /// from GV will trap if the loaded value is null. Note that this also permits
685 /// comparisons of the loaded value against null, as a special case.
AllUsesOfLoadedValueWillTrapIfNull(const GlobalVariable * GV)686 static bool AllUsesOfLoadedValueWillTrapIfNull(const GlobalVariable *GV) {
687 for (Value::const_use_iterator UI = GV->use_begin(), E = GV->use_end();
688 UI != E; ++UI) {
689 const User *U = *UI;
690
691 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
692 SmallPtrSet<const PHINode*, 8> PHIs;
693 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
694 return false;
695 } else if (isa<StoreInst>(U)) {
696 // Ignore stores to the global.
697 } else {
698 // We don't know or understand this user, bail out.
699 //cerr << "UNKNOWN USER OF GLOBAL!: " << *U;
700 return false;
701 }
702 }
703 return true;
704 }
705
OptimizeAwayTrappingUsesOfValue(Value * V,Constant * NewV)706 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
707 bool Changed = false;
708 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
709 Instruction *I = cast<Instruction>(*UI++);
710 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
711 LI->setOperand(0, NewV);
712 Changed = true;
713 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
714 if (SI->getOperand(1) == V) {
715 SI->setOperand(1, NewV);
716 Changed = true;
717 }
718 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
719 CallSite CS(I);
720 if (CS.getCalledValue() == V) {
721 // Calling through the pointer! Turn into a direct call, but be careful
722 // that the pointer is not also being passed as an argument.
723 CS.setCalledFunction(NewV);
724 Changed = true;
725 bool PassedAsArg = false;
726 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
727 if (CS.getArgument(i) == V) {
728 PassedAsArg = true;
729 CS.setArgument(i, NewV);
730 }
731
732 if (PassedAsArg) {
733 // Being passed as an argument also. Be careful to not invalidate UI!
734 UI = V->use_begin();
735 }
736 }
737 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
738 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
739 ConstantExpr::getCast(CI->getOpcode(),
740 NewV, CI->getType()));
741 if (CI->use_empty()) {
742 Changed = true;
743 CI->eraseFromParent();
744 }
745 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
746 // Should handle GEP here.
747 SmallVector<Constant*, 8> Idxs;
748 Idxs.reserve(GEPI->getNumOperands()-1);
749 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
750 i != e; ++i)
751 if (Constant *C = dyn_cast<Constant>(*i))
752 Idxs.push_back(C);
753 else
754 break;
755 if (Idxs.size() == GEPI->getNumOperands()-1)
756 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
757 ConstantExpr::getGetElementPtr(NewV, Idxs));
758 if (GEPI->use_empty()) {
759 Changed = true;
760 GEPI->eraseFromParent();
761 }
762 }
763 }
764
765 return Changed;
766 }
767
768
769 /// OptimizeAwayTrappingUsesOfLoads - The specified global has only one non-null
770 /// value stored into it. If there are uses of the loaded value that would trap
771 /// if the loaded value is dynamically null, then we know that they cannot be
772 /// reachable with a null optimize away the load.
OptimizeAwayTrappingUsesOfLoads(GlobalVariable * GV,Constant * LV)773 static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) {
774 bool Changed = false;
775
776 // Keep track of whether we are able to remove all the uses of the global
777 // other than the store that defines it.
778 bool AllNonStoreUsesGone = true;
779
780 // Replace all uses of loads with uses of uses of the stored value.
781 for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end(); GUI != E;){
782 User *GlobalUser = *GUI++;
783 if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) {
784 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
785 // If we were able to delete all uses of the loads
786 if (LI->use_empty()) {
787 LI->eraseFromParent();
788 Changed = true;
789 } else {
790 AllNonStoreUsesGone = false;
791 }
792 } else if (isa<StoreInst>(GlobalUser)) {
793 // Ignore the store that stores "LV" to the global.
794 assert(GlobalUser->getOperand(1) == GV &&
795 "Must be storing *to* the global");
796 } else {
797 AllNonStoreUsesGone = false;
798
799 // If we get here we could have other crazy uses that are transitively
800 // loaded.
801 assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) ||
802 isa<ConstantExpr>(GlobalUser) || isa<CmpInst>(GlobalUser)) &&
803 "Only expect load and stores!");
804 }
805 }
806
807 if (Changed) {
808 DEBUG(dbgs() << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV);
809 ++NumGlobUses;
810 }
811
812 // If we nuked all of the loads, then none of the stores are needed either,
813 // nor is the global.
814 if (AllNonStoreUsesGone) {
815 DEBUG(dbgs() << " *** GLOBAL NOW DEAD!\n");
816 CleanupConstantGlobalUsers(GV, 0);
817 if (GV->use_empty()) {
818 GV->eraseFromParent();
819 ++NumDeleted;
820 }
821 Changed = true;
822 }
823 return Changed;
824 }
825
826 /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the
827 /// instructions that are foldable.
ConstantPropUsersOf(Value * V)828 static void ConstantPropUsersOf(Value *V) {
829 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; )
830 if (Instruction *I = dyn_cast<Instruction>(*UI++))
831 if (Constant *NewC = ConstantFoldInstruction(I)) {
832 I->replaceAllUsesWith(NewC);
833
834 // Advance UI to the next non-I use to avoid invalidating it!
835 // Instructions could multiply use V.
836 while (UI != E && *UI == I)
837 ++UI;
838 I->eraseFromParent();
839 }
840 }
841
842 /// OptimizeGlobalAddressOfMalloc - This function takes the specified global
843 /// variable, and transforms the program as if it always contained the result of
844 /// the specified malloc. Because it is always the result of the specified
845 /// malloc, there is no reason to actually DO the malloc. Instead, turn the
846 /// malloc into a global, and any loads of GV as uses of the new global.
OptimizeGlobalAddressOfMalloc(GlobalVariable * GV,CallInst * CI,Type * AllocTy,ConstantInt * NElements,TargetData * TD)847 static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV,
848 CallInst *CI,
849 Type *AllocTy,
850 ConstantInt *NElements,
851 TargetData* TD) {
852 DEBUG(errs() << "PROMOTING GLOBAL: " << *GV << " CALL = " << *CI << '\n');
853
854 Type *GlobalType;
855 if (NElements->getZExtValue() == 1)
856 GlobalType = AllocTy;
857 else
858 // If we have an array allocation, the global variable is of an array.
859 GlobalType = ArrayType::get(AllocTy, NElements->getZExtValue());
860
861 // Create the new global variable. The contents of the malloc'd memory is
862 // undefined, so initialize with an undef value.
863 GlobalVariable *NewGV = new GlobalVariable(*GV->getParent(),
864 GlobalType, false,
865 GlobalValue::InternalLinkage,
866 UndefValue::get(GlobalType),
867 GV->getName()+".body",
868 GV,
869 GV->isThreadLocal());
870
871 // If there are bitcast users of the malloc (which is typical, usually we have
872 // a malloc + bitcast) then replace them with uses of the new global. Update
873 // other users to use the global as well.
874 BitCastInst *TheBC = 0;
875 while (!CI->use_empty()) {
876 Instruction *User = cast<Instruction>(CI->use_back());
877 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
878 if (BCI->getType() == NewGV->getType()) {
879 BCI->replaceAllUsesWith(NewGV);
880 BCI->eraseFromParent();
881 } else {
882 BCI->setOperand(0, NewGV);
883 }
884 } else {
885 if (TheBC == 0)
886 TheBC = new BitCastInst(NewGV, CI->getType(), "newgv", CI);
887 User->replaceUsesOfWith(CI, TheBC);
888 }
889 }
890
891 Constant *RepValue = NewGV;
892 if (NewGV->getType() != GV->getType()->getElementType())
893 RepValue = ConstantExpr::getBitCast(RepValue,
894 GV->getType()->getElementType());
895
896 // If there is a comparison against null, we will insert a global bool to
897 // keep track of whether the global was initialized yet or not.
898 GlobalVariable *InitBool =
899 new GlobalVariable(Type::getInt1Ty(GV->getContext()), false,
900 GlobalValue::InternalLinkage,
901 ConstantInt::getFalse(GV->getContext()),
902 GV->getName()+".init", GV->isThreadLocal());
903 bool InitBoolUsed = false;
904
905 // Loop over all uses of GV, processing them in turn.
906 while (!GV->use_empty()) {
907 if (StoreInst *SI = dyn_cast<StoreInst>(GV->use_back())) {
908 // The global is initialized when the store to it occurs.
909 new StoreInst(ConstantInt::getTrue(GV->getContext()), InitBool, SI);
910 SI->eraseFromParent();
911 continue;
912 }
913
914 LoadInst *LI = cast<LoadInst>(GV->use_back());
915 while (!LI->use_empty()) {
916 Use &LoadUse = LI->use_begin().getUse();
917 if (!isa<ICmpInst>(LoadUse.getUser())) {
918 LoadUse = RepValue;
919 continue;
920 }
921
922 ICmpInst *ICI = cast<ICmpInst>(LoadUse.getUser());
923 // Replace the cmp X, 0 with a use of the bool value.
924 Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", ICI);
925 InitBoolUsed = true;
926 switch (ICI->getPredicate()) {
927 default: llvm_unreachable("Unknown ICmp Predicate!");
928 case ICmpInst::ICMP_ULT:
929 case ICmpInst::ICMP_SLT: // X < null -> always false
930 LV = ConstantInt::getFalse(GV->getContext());
931 break;
932 case ICmpInst::ICMP_ULE:
933 case ICmpInst::ICMP_SLE:
934 case ICmpInst::ICMP_EQ:
935 LV = BinaryOperator::CreateNot(LV, "notinit", ICI);
936 break;
937 case ICmpInst::ICMP_NE:
938 case ICmpInst::ICMP_UGE:
939 case ICmpInst::ICMP_SGE:
940 case ICmpInst::ICMP_UGT:
941 case ICmpInst::ICMP_SGT:
942 break; // no change.
943 }
944 ICI->replaceAllUsesWith(LV);
945 ICI->eraseFromParent();
946 }
947 LI->eraseFromParent();
948 }
949
950 // If the initialization boolean was used, insert it, otherwise delete it.
951 if (!InitBoolUsed) {
952 while (!InitBool->use_empty()) // Delete initializations
953 cast<StoreInst>(InitBool->use_back())->eraseFromParent();
954 delete InitBool;
955 } else
956 GV->getParent()->getGlobalList().insert(GV, InitBool);
957
958 // Now the GV is dead, nuke it and the malloc..
959 GV->eraseFromParent();
960 CI->eraseFromParent();
961
962 // To further other optimizations, loop over all users of NewGV and try to
963 // constant prop them. This will promote GEP instructions with constant
964 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
965 ConstantPropUsersOf(NewGV);
966 if (RepValue != NewGV)
967 ConstantPropUsersOf(RepValue);
968
969 return NewGV;
970 }
971
972 /// ValueIsOnlyUsedLocallyOrStoredToOneGlobal - Scan the use-list of V checking
973 /// to make sure that there are no complex uses of V. We permit simple things
974 /// like dereferencing the pointer, but not storing through the address, unless
975 /// it is to the specified global.
ValueIsOnlyUsedLocallyOrStoredToOneGlobal(const Instruction * V,const GlobalVariable * GV,SmallPtrSet<const PHINode *,8> & PHIs)976 static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(const Instruction *V,
977 const GlobalVariable *GV,
978 SmallPtrSet<const PHINode*, 8> &PHIs) {
979 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end();
980 UI != E; ++UI) {
981 const Instruction *Inst = cast<Instruction>(*UI);
982
983 if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) {
984 continue; // Fine, ignore.
985 }
986
987 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
988 if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
989 return false; // Storing the pointer itself... bad.
990 continue; // Otherwise, storing through it, or storing into GV... fine.
991 }
992
993 // Must index into the array and into the struct.
994 if (isa<GetElementPtrInst>(Inst) && Inst->getNumOperands() >= 3) {
995 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs))
996 return false;
997 continue;
998 }
999
1000 if (const PHINode *PN = dyn_cast<PHINode>(Inst)) {
1001 // PHIs are ok if all uses are ok. Don't infinitely recurse through PHI
1002 // cycles.
1003 if (PHIs.insert(PN))
1004 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
1005 return false;
1006 continue;
1007 }
1008
1009 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) {
1010 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs))
1011 return false;
1012 continue;
1013 }
1014
1015 return false;
1016 }
1017 return true;
1018 }
1019
1020 /// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
1021 /// somewhere. Transform all uses of the allocation into loads from the
1022 /// global and uses of the resultant pointer. Further, delete the store into
1023 /// GV. This assumes that these value pass the
1024 /// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
ReplaceUsesOfMallocWithGlobal(Instruction * Alloc,GlobalVariable * GV)1025 static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
1026 GlobalVariable *GV) {
1027 while (!Alloc->use_empty()) {
1028 Instruction *U = cast<Instruction>(*Alloc->use_begin());
1029 Instruction *InsertPt = U;
1030 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1031 // If this is the store of the allocation into the global, remove it.
1032 if (SI->getOperand(1) == GV) {
1033 SI->eraseFromParent();
1034 continue;
1035 }
1036 } else if (PHINode *PN = dyn_cast<PHINode>(U)) {
1037 // Insert the load in the corresponding predecessor, not right before the
1038 // PHI.
1039 InsertPt = PN->getIncomingBlock(Alloc->use_begin())->getTerminator();
1040 } else if (isa<BitCastInst>(U)) {
1041 // Must be bitcast between the malloc and store to initialize the global.
1042 ReplaceUsesOfMallocWithGlobal(U, GV);
1043 U->eraseFromParent();
1044 continue;
1045 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
1046 // If this is a "GEP bitcast" and the user is a store to the global, then
1047 // just process it as a bitcast.
1048 if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse())
1049 if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->use_back()))
1050 if (SI->getOperand(1) == GV) {
1051 // Must be bitcast GEP between the malloc and store to initialize
1052 // the global.
1053 ReplaceUsesOfMallocWithGlobal(GEPI, GV);
1054 GEPI->eraseFromParent();
1055 continue;
1056 }
1057 }
1058
1059 // Insert a load from the global, and use it instead of the malloc.
1060 Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
1061 U->replaceUsesOfWith(Alloc, NL);
1062 }
1063 }
1064
1065 /// LoadUsesSimpleEnoughForHeapSRA - Verify that all uses of V (a load, or a phi
1066 /// of a load) are simple enough to perform heap SRA on. This permits GEP's
1067 /// that index through the array and struct field, icmps of null, and PHIs.
LoadUsesSimpleEnoughForHeapSRA(const Value * V,SmallPtrSet<const PHINode *,32> & LoadUsingPHIs,SmallPtrSet<const PHINode *,32> & LoadUsingPHIsPerLoad)1068 static bool LoadUsesSimpleEnoughForHeapSRA(const Value *V,
1069 SmallPtrSet<const PHINode*, 32> &LoadUsingPHIs,
1070 SmallPtrSet<const PHINode*, 32> &LoadUsingPHIsPerLoad) {
1071 // We permit two users of the load: setcc comparing against the null
1072 // pointer, and a getelementptr of a specific form.
1073 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;
1074 ++UI) {
1075 const Instruction *User = cast<Instruction>(*UI);
1076
1077 // Comparison against null is ok.
1078 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(User)) {
1079 if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
1080 return false;
1081 continue;
1082 }
1083
1084 // getelementptr is also ok, but only a simple form.
1085 if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1086 // Must index into the array and into the struct.
1087 if (GEPI->getNumOperands() < 3)
1088 return false;
1089
1090 // Otherwise the GEP is ok.
1091 continue;
1092 }
1093
1094 if (const PHINode *PN = dyn_cast<PHINode>(User)) {
1095 if (!LoadUsingPHIsPerLoad.insert(PN))
1096 // This means some phi nodes are dependent on each other.
1097 // Avoid infinite looping!
1098 return false;
1099 if (!LoadUsingPHIs.insert(PN))
1100 // If we have already analyzed this PHI, then it is safe.
1101 continue;
1102
1103 // Make sure all uses of the PHI are simple enough to transform.
1104 if (!LoadUsesSimpleEnoughForHeapSRA(PN,
1105 LoadUsingPHIs, LoadUsingPHIsPerLoad))
1106 return false;
1107
1108 continue;
1109 }
1110
1111 // Otherwise we don't know what this is, not ok.
1112 return false;
1113 }
1114
1115 return true;
1116 }
1117
1118
1119 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA - If all users of values loaded from
1120 /// GV are simple enough to perform HeapSRA, return true.
AllGlobalLoadUsesSimpleEnoughForHeapSRA(const GlobalVariable * GV,Instruction * StoredVal)1121 static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(const GlobalVariable *GV,
1122 Instruction *StoredVal) {
1123 SmallPtrSet<const PHINode*, 32> LoadUsingPHIs;
1124 SmallPtrSet<const PHINode*, 32> LoadUsingPHIsPerLoad;
1125 for (Value::const_use_iterator UI = GV->use_begin(), E = GV->use_end();
1126 UI != E; ++UI)
1127 if (const LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
1128 if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs,
1129 LoadUsingPHIsPerLoad))
1130 return false;
1131 LoadUsingPHIsPerLoad.clear();
1132 }
1133
1134 // If we reach here, we know that all uses of the loads and transitive uses
1135 // (through PHI nodes) are simple enough to transform. However, we don't know
1136 // that all inputs the to the PHI nodes are in the same equivalence sets.
1137 // Check to verify that all operands of the PHIs are either PHIS that can be
1138 // transformed, loads from GV, or MI itself.
1139 for (SmallPtrSet<const PHINode*, 32>::const_iterator I = LoadUsingPHIs.begin()
1140 , E = LoadUsingPHIs.end(); I != E; ++I) {
1141 const PHINode *PN = *I;
1142 for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) {
1143 Value *InVal = PN->getIncomingValue(op);
1144
1145 // PHI of the stored value itself is ok.
1146 if (InVal == StoredVal) continue;
1147
1148 if (const PHINode *InPN = dyn_cast<PHINode>(InVal)) {
1149 // One of the PHIs in our set is (optimistically) ok.
1150 if (LoadUsingPHIs.count(InPN))
1151 continue;
1152 return false;
1153 }
1154
1155 // Load from GV is ok.
1156 if (const LoadInst *LI = dyn_cast<LoadInst>(InVal))
1157 if (LI->getOperand(0) == GV)
1158 continue;
1159
1160 // UNDEF? NULL?
1161
1162 // Anything else is rejected.
1163 return false;
1164 }
1165 }
1166
1167 return true;
1168 }
1169
GetHeapSROAValue(Value * V,unsigned FieldNo,DenseMap<Value *,std::vector<Value * >> & InsertedScalarizedValues,std::vector<std::pair<PHINode *,unsigned>> & PHIsToRewrite)1170 static Value *GetHeapSROAValue(Value *V, unsigned FieldNo,
1171 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1172 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1173 std::vector<Value*> &FieldVals = InsertedScalarizedValues[V];
1174
1175 if (FieldNo >= FieldVals.size())
1176 FieldVals.resize(FieldNo+1);
1177
1178 // If we already have this value, just reuse the previously scalarized
1179 // version.
1180 if (Value *FieldVal = FieldVals[FieldNo])
1181 return FieldVal;
1182
1183 // Depending on what instruction this is, we have several cases.
1184 Value *Result;
1185 if (LoadInst *LI = dyn_cast<LoadInst>(V)) {
1186 // This is a scalarized version of the load from the global. Just create
1187 // a new Load of the scalarized global.
1188 Result = new LoadInst(GetHeapSROAValue(LI->getOperand(0), FieldNo,
1189 InsertedScalarizedValues,
1190 PHIsToRewrite),
1191 LI->getName()+".f"+Twine(FieldNo), LI);
1192 } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
1193 // PN's type is pointer to struct. Make a new PHI of pointer to struct
1194 // field.
1195 StructType *ST =
1196 cast<StructType>(cast<PointerType>(PN->getType())->getElementType());
1197
1198 PHINode *NewPN =
1199 PHINode::Create(PointerType::getUnqual(ST->getElementType(FieldNo)),
1200 PN->getNumIncomingValues(),
1201 PN->getName()+".f"+Twine(FieldNo), PN);
1202 Result = NewPN;
1203 PHIsToRewrite.push_back(std::make_pair(PN, FieldNo));
1204 } else {
1205 llvm_unreachable("Unknown usable value");
1206 Result = 0;
1207 }
1208
1209 return FieldVals[FieldNo] = Result;
1210 }
1211
1212 /// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
1213 /// the load, rewrite the derived value to use the HeapSRoA'd load.
RewriteHeapSROALoadUser(Instruction * LoadUser,DenseMap<Value *,std::vector<Value * >> & InsertedScalarizedValues,std::vector<std::pair<PHINode *,unsigned>> & PHIsToRewrite)1214 static void RewriteHeapSROALoadUser(Instruction *LoadUser,
1215 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1216 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1217 // If this is a comparison against null, handle it.
1218 if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
1219 assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
1220 // If we have a setcc of the loaded pointer, we can use a setcc of any
1221 // field.
1222 Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0,
1223 InsertedScalarizedValues, PHIsToRewrite);
1224
1225 Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr,
1226 Constant::getNullValue(NPtr->getType()),
1227 SCI->getName());
1228 SCI->replaceAllUsesWith(New);
1229 SCI->eraseFromParent();
1230 return;
1231 }
1232
1233 // Handle 'getelementptr Ptr, Idx, i32 FieldNo ...'
1234 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
1235 assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
1236 && "Unexpected GEPI!");
1237
1238 // Load the pointer for this field.
1239 unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
1240 Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo,
1241 InsertedScalarizedValues, PHIsToRewrite);
1242
1243 // Create the new GEP idx vector.
1244 SmallVector<Value*, 8> GEPIdx;
1245 GEPIdx.push_back(GEPI->getOperand(1));
1246 GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
1247
1248 Value *NGEPI = GetElementPtrInst::Create(NewPtr, GEPIdx,
1249 GEPI->getName(), GEPI);
1250 GEPI->replaceAllUsesWith(NGEPI);
1251 GEPI->eraseFromParent();
1252 return;
1253 }
1254
1255 // Recursively transform the users of PHI nodes. This will lazily create the
1256 // PHIs that are needed for individual elements. Keep track of what PHIs we
1257 // see in InsertedScalarizedValues so that we don't get infinite loops (very
1258 // antisocial). If the PHI is already in InsertedScalarizedValues, it has
1259 // already been seen first by another load, so its uses have already been
1260 // processed.
1261 PHINode *PN = cast<PHINode>(LoadUser);
1262 if (!InsertedScalarizedValues.insert(std::make_pair(PN,
1263 std::vector<Value*>())).second)
1264 return;
1265
1266 // If this is the first time we've seen this PHI, recursively process all
1267 // users.
1268 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); UI != E; ) {
1269 Instruction *User = cast<Instruction>(*UI++);
1270 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1271 }
1272 }
1273
1274 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1275 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1276 /// use FieldGlobals instead. All uses of loaded values satisfy
1277 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA.
RewriteUsesOfLoadForHeapSRoA(LoadInst * Load,DenseMap<Value *,std::vector<Value * >> & InsertedScalarizedValues,std::vector<std::pair<PHINode *,unsigned>> & PHIsToRewrite)1278 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1279 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1280 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1281 for (Value::use_iterator UI = Load->use_begin(), E = Load->use_end();
1282 UI != E; ) {
1283 Instruction *User = cast<Instruction>(*UI++);
1284 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1285 }
1286
1287 if (Load->use_empty()) {
1288 Load->eraseFromParent();
1289 InsertedScalarizedValues.erase(Load);
1290 }
1291 }
1292
1293 /// PerformHeapAllocSRoA - CI is an allocation of an array of structures. Break
1294 /// it up into multiple allocations of arrays of the fields.
PerformHeapAllocSRoA(GlobalVariable * GV,CallInst * CI,Value * NElems,TargetData * TD)1295 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI,
1296 Value* NElems, TargetData *TD) {
1297 DEBUG(dbgs() << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *CI << '\n');
1298 Type* MAT = getMallocAllocatedType(CI);
1299 StructType *STy = cast<StructType>(MAT);
1300
1301 // There is guaranteed to be at least one use of the malloc (storing
1302 // it into GV). If there are other uses, change them to be uses of
1303 // the global to simplify later code. This also deletes the store
1304 // into GV.
1305 ReplaceUsesOfMallocWithGlobal(CI, GV);
1306
1307 // Okay, at this point, there are no users of the malloc. Insert N
1308 // new mallocs at the same place as CI, and N globals.
1309 std::vector<Value*> FieldGlobals;
1310 std::vector<Value*> FieldMallocs;
1311
1312 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1313 Type *FieldTy = STy->getElementType(FieldNo);
1314 PointerType *PFieldTy = PointerType::getUnqual(FieldTy);
1315
1316 GlobalVariable *NGV =
1317 new GlobalVariable(*GV->getParent(),
1318 PFieldTy, false, GlobalValue::InternalLinkage,
1319 Constant::getNullValue(PFieldTy),
1320 GV->getName() + ".f" + Twine(FieldNo), GV,
1321 GV->isThreadLocal());
1322 FieldGlobals.push_back(NGV);
1323
1324 unsigned TypeSize = TD->getTypeAllocSize(FieldTy);
1325 if (StructType *ST = dyn_cast<StructType>(FieldTy))
1326 TypeSize = TD->getStructLayout(ST)->getSizeInBytes();
1327 Type *IntPtrTy = TD->getIntPtrType(CI->getContext());
1328 Value *NMI = CallInst::CreateMalloc(CI, IntPtrTy, FieldTy,
1329 ConstantInt::get(IntPtrTy, TypeSize),
1330 NElems, 0,
1331 CI->getName() + ".f" + Twine(FieldNo));
1332 FieldMallocs.push_back(NMI);
1333 new StoreInst(NMI, NGV, CI);
1334 }
1335
1336 // The tricky aspect of this transformation is handling the case when malloc
1337 // fails. In the original code, malloc failing would set the result pointer
1338 // of malloc to null. In this case, some mallocs could succeed and others
1339 // could fail. As such, we emit code that looks like this:
1340 // F0 = malloc(field0)
1341 // F1 = malloc(field1)
1342 // F2 = malloc(field2)
1343 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1344 // if (F0) { free(F0); F0 = 0; }
1345 // if (F1) { free(F1); F1 = 0; }
1346 // if (F2) { free(F2); F2 = 0; }
1347 // }
1348 // The malloc can also fail if its argument is too large.
1349 Constant *ConstantZero = ConstantInt::get(CI->getArgOperand(0)->getType(), 0);
1350 Value *RunningOr = new ICmpInst(CI, ICmpInst::ICMP_SLT, CI->getArgOperand(0),
1351 ConstantZero, "isneg");
1352 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1353 Value *Cond = new ICmpInst(CI, ICmpInst::ICMP_EQ, FieldMallocs[i],
1354 Constant::getNullValue(FieldMallocs[i]->getType()),
1355 "isnull");
1356 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", CI);
1357 }
1358
1359 // Split the basic block at the old malloc.
1360 BasicBlock *OrigBB = CI->getParent();
1361 BasicBlock *ContBB = OrigBB->splitBasicBlock(CI, "malloc_cont");
1362
1363 // Create the block to check the first condition. Put all these blocks at the
1364 // end of the function as they are unlikely to be executed.
1365 BasicBlock *NullPtrBlock = BasicBlock::Create(OrigBB->getContext(),
1366 "malloc_ret_null",
1367 OrigBB->getParent());
1368
1369 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1370 // branch on RunningOr.
1371 OrigBB->getTerminator()->eraseFromParent();
1372 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1373
1374 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1375 // pointer, because some may be null while others are not.
1376 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1377 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1378 Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
1379 Constant::getNullValue(GVVal->getType()));
1380 BasicBlock *FreeBlock = BasicBlock::Create(Cmp->getContext(), "free_it",
1381 OrigBB->getParent());
1382 BasicBlock *NextBlock = BasicBlock::Create(Cmp->getContext(), "next",
1383 OrigBB->getParent());
1384 Instruction *BI = BranchInst::Create(FreeBlock, NextBlock,
1385 Cmp, NullPtrBlock);
1386
1387 // Fill in FreeBlock.
1388 CallInst::CreateFree(GVVal, BI);
1389 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1390 FreeBlock);
1391 BranchInst::Create(NextBlock, FreeBlock);
1392
1393 NullPtrBlock = NextBlock;
1394 }
1395
1396 BranchInst::Create(ContBB, NullPtrBlock);
1397
1398 // CI is no longer needed, remove it.
1399 CI->eraseFromParent();
1400
1401 /// InsertedScalarizedLoads - As we process loads, if we can't immediately
1402 /// update all uses of the load, keep track of what scalarized loads are
1403 /// inserted for a given load.
1404 DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues;
1405 InsertedScalarizedValues[GV] = FieldGlobals;
1406
1407 std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite;
1408
1409 // Okay, the malloc site is completely handled. All of the uses of GV are now
1410 // loads, and all uses of those loads are simple. Rewrite them to use loads
1411 // of the per-field globals instead.
1412 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;) {
1413 Instruction *User = cast<Instruction>(*UI++);
1414
1415 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1416 RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite);
1417 continue;
1418 }
1419
1420 // Must be a store of null.
1421 StoreInst *SI = cast<StoreInst>(User);
1422 assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
1423 "Unexpected heap-sra user!");
1424
1425 // Insert a store of null into each global.
1426 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1427 PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType());
1428 Constant *Null = Constant::getNullValue(PT->getElementType());
1429 new StoreInst(Null, FieldGlobals[i], SI);
1430 }
1431 // Erase the original store.
1432 SI->eraseFromParent();
1433 }
1434
1435 // While we have PHIs that are interesting to rewrite, do it.
1436 while (!PHIsToRewrite.empty()) {
1437 PHINode *PN = PHIsToRewrite.back().first;
1438 unsigned FieldNo = PHIsToRewrite.back().second;
1439 PHIsToRewrite.pop_back();
1440 PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]);
1441 assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi");
1442
1443 // Add all the incoming values. This can materialize more phis.
1444 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1445 Value *InVal = PN->getIncomingValue(i);
1446 InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues,
1447 PHIsToRewrite);
1448 FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
1449 }
1450 }
1451
1452 // Drop all inter-phi links and any loads that made it this far.
1453 for (DenseMap<Value*, std::vector<Value*> >::iterator
1454 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1455 I != E; ++I) {
1456 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1457 PN->dropAllReferences();
1458 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1459 LI->dropAllReferences();
1460 }
1461
1462 // Delete all the phis and loads now that inter-references are dead.
1463 for (DenseMap<Value*, std::vector<Value*> >::iterator
1464 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1465 I != E; ++I) {
1466 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1467 PN->eraseFromParent();
1468 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1469 LI->eraseFromParent();
1470 }
1471
1472 // The old global is now dead, remove it.
1473 GV->eraseFromParent();
1474
1475 ++NumHeapSRA;
1476 return cast<GlobalVariable>(FieldGlobals[0]);
1477 }
1478
1479 /// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1480 /// pointer global variable with a single value stored it that is a malloc or
1481 /// cast of malloc.
TryToOptimizeStoreOfMallocToGlobal(GlobalVariable * GV,CallInst * CI,Type * AllocTy,Module::global_iterator & GVI,TargetData * TD)1482 static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
1483 CallInst *CI,
1484 Type *AllocTy,
1485 Module::global_iterator &GVI,
1486 TargetData *TD) {
1487 if (!TD)
1488 return false;
1489
1490 // If this is a malloc of an abstract type, don't touch it.
1491 if (!AllocTy->isSized())
1492 return false;
1493
1494 // We can't optimize this global unless all uses of it are *known* to be
1495 // of the malloc value, not of the null initializer value (consider a use
1496 // that compares the global's value against zero to see if the malloc has
1497 // been reached). To do this, we check to see if all uses of the global
1498 // would trap if the global were null: this proves that they must all
1499 // happen after the malloc.
1500 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1501 return false;
1502
1503 // We can't optimize this if the malloc itself is used in a complex way,
1504 // for example, being stored into multiple globals. This allows the
1505 // malloc to be stored into the specified global, loaded setcc'd, and
1506 // GEP'd. These are all things we could transform to using the global
1507 // for.
1508 SmallPtrSet<const PHINode*, 8> PHIs;
1509 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV, PHIs))
1510 return false;
1511
1512 // If we have a global that is only initialized with a fixed size malloc,
1513 // transform the program to use global memory instead of malloc'd memory.
1514 // This eliminates dynamic allocation, avoids an indirection accessing the
1515 // data, and exposes the resultant global to further GlobalOpt.
1516 // We cannot optimize the malloc if we cannot determine malloc array size.
1517 Value *NElems = getMallocArraySize(CI, TD, true);
1518 if (!NElems)
1519 return false;
1520
1521 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
1522 // Restrict this transformation to only working on small allocations
1523 // (2048 bytes currently), as we don't want to introduce a 16M global or
1524 // something.
1525 if (NElements->getZExtValue() * TD->getTypeAllocSize(AllocTy) < 2048) {
1526 GVI = OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, TD);
1527 return true;
1528 }
1529
1530 // If the allocation is an array of structures, consider transforming this
1531 // into multiple malloc'd arrays, one for each field. This is basically
1532 // SRoA for malloc'd memory.
1533
1534 // If this is an allocation of a fixed size array of structs, analyze as a
1535 // variable size array. malloc [100 x struct],1 -> malloc struct, 100
1536 if (NElems == ConstantInt::get(CI->getArgOperand(0)->getType(), 1))
1537 if (ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
1538 AllocTy = AT->getElementType();
1539
1540 StructType *AllocSTy = dyn_cast<StructType>(AllocTy);
1541 if (!AllocSTy)
1542 return false;
1543
1544 // This the structure has an unreasonable number of fields, leave it
1545 // alone.
1546 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
1547 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, CI)) {
1548
1549 // If this is a fixed size array, transform the Malloc to be an alloc of
1550 // structs. malloc [100 x struct],1 -> malloc struct, 100
1551 if (ArrayType *AT = dyn_cast<ArrayType>(getMallocAllocatedType(CI))) {
1552 Type *IntPtrTy = TD->getIntPtrType(CI->getContext());
1553 unsigned TypeSize = TD->getStructLayout(AllocSTy)->getSizeInBytes();
1554 Value *AllocSize = ConstantInt::get(IntPtrTy, TypeSize);
1555 Value *NumElements = ConstantInt::get(IntPtrTy, AT->getNumElements());
1556 Instruction *Malloc = CallInst::CreateMalloc(CI, IntPtrTy, AllocSTy,
1557 AllocSize, NumElements,
1558 0, CI->getName());
1559 Instruction *Cast = new BitCastInst(Malloc, CI->getType(), "tmp", CI);
1560 CI->replaceAllUsesWith(Cast);
1561 CI->eraseFromParent();
1562 CI = dyn_cast<BitCastInst>(Malloc) ?
1563 extractMallocCallFromBitCast(Malloc) : cast<CallInst>(Malloc);
1564 }
1565
1566 GVI = PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, TD, true),TD);
1567 return true;
1568 }
1569
1570 return false;
1571 }
1572
1573 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1574 // that only one value (besides its initializer) is ever stored to the global.
OptimizeOnceStoredGlobal(GlobalVariable * GV,Value * StoredOnceVal,Module::global_iterator & GVI,TargetData * TD)1575 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1576 Module::global_iterator &GVI,
1577 TargetData *TD) {
1578 // Ignore no-op GEPs and bitcasts.
1579 StoredOnceVal = StoredOnceVal->stripPointerCasts();
1580
1581 // If we are dealing with a pointer global that is initialized to null and
1582 // only has one (non-null) value stored into it, then we can optimize any
1583 // users of the loaded value (often calls and loads) that would trap if the
1584 // value was null.
1585 if (GV->getInitializer()->getType()->isPointerTy() &&
1586 GV->getInitializer()->isNullValue()) {
1587 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1588 if (GV->getInitializer()->getType() != SOVC->getType())
1589 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1590
1591 // Optimize away any trapping uses of the loaded value.
1592 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1593 return true;
1594 } else if (CallInst *CI = extractMallocCall(StoredOnceVal)) {
1595 Type* MallocType = getMallocAllocatedType(CI);
1596 if (MallocType && TryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType,
1597 GVI, TD))
1598 return true;
1599 }
1600 }
1601
1602 return false;
1603 }
1604
1605 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1606 /// two values ever stored into GV are its initializer and OtherVal. See if we
1607 /// can shrink the global into a boolean and select between the two values
1608 /// whenever it is used. This exposes the values to other scalar optimizations.
TryToShrinkGlobalToBoolean(GlobalVariable * GV,Constant * OtherVal)1609 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1610 Type *GVElType = GV->getType()->getElementType();
1611
1612 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1613 // an FP value, pointer or vector, don't do this optimization because a select
1614 // between them is very expensive and unlikely to lead to later
1615 // simplification. In these cases, we typically end up with "cond ? v1 : v2"
1616 // where v1 and v2 both require constant pool loads, a big loss.
1617 if (GVElType == Type::getInt1Ty(GV->getContext()) ||
1618 GVElType->isFloatingPointTy() ||
1619 GVElType->isPointerTy() || GVElType->isVectorTy())
1620 return false;
1621
1622 // Walk the use list of the global seeing if all the uses are load or store.
1623 // If there is anything else, bail out.
1624 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I){
1625 User *U = *I;
1626 if (!isa<LoadInst>(U) && !isa<StoreInst>(U))
1627 return false;
1628 }
1629
1630 DEBUG(dbgs() << " *** SHRINKING TO BOOL: " << *GV);
1631
1632 // Create the new global, initializing it to false.
1633 GlobalVariable *NewGV = new GlobalVariable(Type::getInt1Ty(GV->getContext()),
1634 false,
1635 GlobalValue::InternalLinkage,
1636 ConstantInt::getFalse(GV->getContext()),
1637 GV->getName()+".b",
1638 GV->isThreadLocal());
1639 GV->getParent()->getGlobalList().insert(GV, NewGV);
1640
1641 Constant *InitVal = GV->getInitializer();
1642 assert(InitVal->getType() != Type::getInt1Ty(GV->getContext()) &&
1643 "No reason to shrink to bool!");
1644
1645 // If initialized to zero and storing one into the global, we can use a cast
1646 // instead of a select to synthesize the desired value.
1647 bool IsOneZero = false;
1648 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1649 IsOneZero = InitVal->isNullValue() && CI->isOne();
1650
1651 while (!GV->use_empty()) {
1652 Instruction *UI = cast<Instruction>(GV->use_back());
1653 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1654 // Change the store into a boolean store.
1655 bool StoringOther = SI->getOperand(0) == OtherVal;
1656 // Only do this if we weren't storing a loaded value.
1657 Value *StoreVal;
1658 if (StoringOther || SI->getOperand(0) == InitVal)
1659 StoreVal = ConstantInt::get(Type::getInt1Ty(GV->getContext()),
1660 StoringOther);
1661 else {
1662 // Otherwise, we are storing a previously loaded copy. To do this,
1663 // change the copy from copying the original value to just copying the
1664 // bool.
1665 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1666
1667 // If we've already replaced the input, StoredVal will be a cast or
1668 // select instruction. If not, it will be a load of the original
1669 // global.
1670 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1671 assert(LI->getOperand(0) == GV && "Not a copy!");
1672 // Insert a new load, to preserve the saved value.
1673 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1674 } else {
1675 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1676 "This is not a form that we understand!");
1677 StoreVal = StoredVal->getOperand(0);
1678 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1679 }
1680 }
1681 new StoreInst(StoreVal, NewGV, SI);
1682 } else {
1683 // Change the load into a load of bool then a select.
1684 LoadInst *LI = cast<LoadInst>(UI);
1685 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1686 Value *NSI;
1687 if (IsOneZero)
1688 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1689 else
1690 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1691 NSI->takeName(LI);
1692 LI->replaceAllUsesWith(NSI);
1693 }
1694 UI->eraseFromParent();
1695 }
1696
1697 GV->eraseFromParent();
1698 return true;
1699 }
1700
1701
1702 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1703 /// it if possible. If we make a change, return true.
ProcessGlobal(GlobalVariable * GV,Module::global_iterator & GVI)1704 bool GlobalOpt::ProcessGlobal(GlobalVariable *GV,
1705 Module::global_iterator &GVI) {
1706 if (!GV->hasLocalLinkage())
1707 return false;
1708
1709 // Do more involved optimizations if the global is internal.
1710 GV->removeDeadConstantUsers();
1711
1712 if (GV->use_empty()) {
1713 DEBUG(dbgs() << "GLOBAL DEAD: " << *GV);
1714 GV->eraseFromParent();
1715 ++NumDeleted;
1716 return true;
1717 }
1718
1719 SmallPtrSet<const PHINode*, 16> PHIUsers;
1720 GlobalStatus GS;
1721
1722 if (AnalyzeGlobal(GV, GS, PHIUsers))
1723 return false;
1724
1725 if (!GS.isCompared && !GV->hasUnnamedAddr()) {
1726 GV->setUnnamedAddr(true);
1727 NumUnnamed++;
1728 }
1729
1730 if (GV->isConstant() || !GV->hasInitializer())
1731 return false;
1732
1733 return ProcessInternalGlobal(GV, GVI, PHIUsers, GS);
1734 }
1735
1736 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1737 /// it if possible. If we make a change, return true.
ProcessInternalGlobal(GlobalVariable * GV,Module::global_iterator & GVI,const SmallPtrSet<const PHINode *,16> & PHIUsers,const GlobalStatus & GS)1738 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1739 Module::global_iterator &GVI,
1740 const SmallPtrSet<const PHINode*, 16> &PHIUsers,
1741 const GlobalStatus &GS) {
1742 // If this is a first class global and has only one accessing function
1743 // and this function is main (which we know is not recursive we can make
1744 // this global a local variable) we replace the global with a local alloca
1745 // in this function.
1746 //
1747 // NOTE: It doesn't make sense to promote non single-value types since we
1748 // are just replacing static memory to stack memory.
1749 //
1750 // If the global is in different address space, don't bring it to stack.
1751 if (!GS.HasMultipleAccessingFunctions &&
1752 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1753 GV->getType()->getElementType()->isSingleValueType() &&
1754 GS.AccessingFunction->getName() == "main" &&
1755 GS.AccessingFunction->hasExternalLinkage() &&
1756 GV->getType()->getAddressSpace() == 0) {
1757 DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV);
1758 Instruction& FirstI = const_cast<Instruction&>(*GS.AccessingFunction
1759 ->getEntryBlock().begin());
1760 Type* ElemTy = GV->getType()->getElementType();
1761 // FIXME: Pass Global's alignment when globals have alignment
1762 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), &FirstI);
1763 if (!isa<UndefValue>(GV->getInitializer()))
1764 new StoreInst(GV->getInitializer(), Alloca, &FirstI);
1765
1766 GV->replaceAllUsesWith(Alloca);
1767 GV->eraseFromParent();
1768 ++NumLocalized;
1769 return true;
1770 }
1771
1772 // If the global is never loaded (but may be stored to), it is dead.
1773 // Delete it now.
1774 if (!GS.isLoaded) {
1775 DEBUG(dbgs() << "GLOBAL NEVER LOADED: " << *GV);
1776
1777 // Delete any stores we can find to the global. We may not be able to
1778 // make it completely dead though.
1779 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1780
1781 // If the global is dead now, delete it.
1782 if (GV->use_empty()) {
1783 GV->eraseFromParent();
1784 ++NumDeleted;
1785 Changed = true;
1786 }
1787 return Changed;
1788
1789 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1790 DEBUG(dbgs() << "MARKING CONSTANT: " << *GV);
1791 GV->setConstant(true);
1792
1793 // Clean up any obviously simplifiable users now.
1794 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1795
1796 // If the global is dead now, just nuke it.
1797 if (GV->use_empty()) {
1798 DEBUG(dbgs() << " *** Marking constant allowed us to simplify "
1799 << "all users and delete global!\n");
1800 GV->eraseFromParent();
1801 ++NumDeleted;
1802 }
1803
1804 ++NumMarked;
1805 return true;
1806 } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
1807 if (TargetData *TD = getAnalysisIfAvailable<TargetData>())
1808 if (GlobalVariable *FirstNewGV = SRAGlobal(GV, *TD)) {
1809 GVI = FirstNewGV; // Don't skip the newly produced globals!
1810 return true;
1811 }
1812 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1813 // If the initial value for the global was an undef value, and if only
1814 // one other value was stored into it, we can just change the
1815 // initializer to be the stored value, then delete all stores to the
1816 // global. This allows us to mark it constant.
1817 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1818 if (isa<UndefValue>(GV->getInitializer())) {
1819 // Change the initial value here.
1820 GV->setInitializer(SOVConstant);
1821
1822 // Clean up any obviously simplifiable users now.
1823 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1824
1825 if (GV->use_empty()) {
1826 DEBUG(dbgs() << " *** Substituting initializer allowed us to "
1827 << "simplify all users and delete global!\n");
1828 GV->eraseFromParent();
1829 ++NumDeleted;
1830 } else {
1831 GVI = GV;
1832 }
1833 ++NumSubstitute;
1834 return true;
1835 }
1836
1837 // Try to optimize globals based on the knowledge that only one value
1838 // (besides its initializer) is ever stored to the global.
1839 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1840 getAnalysisIfAvailable<TargetData>()))
1841 return true;
1842
1843 // Otherwise, if the global was not a boolean, we can shrink it to be a
1844 // boolean.
1845 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1846 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1847 ++NumShrunkToBool;
1848 return true;
1849 }
1850 }
1851
1852 return false;
1853 }
1854
1855 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1856 /// function, changing them to FastCC.
ChangeCalleesToFastCall(Function * F)1857 static void ChangeCalleesToFastCall(Function *F) {
1858 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1859 CallSite User(cast<Instruction>(*UI));
1860 User.setCallingConv(CallingConv::Fast);
1861 }
1862 }
1863
StripNest(const AttrListPtr & Attrs)1864 static AttrListPtr StripNest(const AttrListPtr &Attrs) {
1865 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1866 if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0)
1867 continue;
1868
1869 // There can be only one.
1870 return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest);
1871 }
1872
1873 return Attrs;
1874 }
1875
RemoveNestAttribute(Function * F)1876 static void RemoveNestAttribute(Function *F) {
1877 F->setAttributes(StripNest(F->getAttributes()));
1878 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1879 CallSite User(cast<Instruction>(*UI));
1880 User.setAttributes(StripNest(User.getAttributes()));
1881 }
1882 }
1883
OptimizeFunctions(Module & M)1884 bool GlobalOpt::OptimizeFunctions(Module &M) {
1885 bool Changed = false;
1886 // Optimize functions.
1887 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1888 Function *F = FI++;
1889 // Functions without names cannot be referenced outside this module.
1890 if (!F->hasName() && !F->isDeclaration())
1891 F->setLinkage(GlobalValue::InternalLinkage);
1892 F->removeDeadConstantUsers();
1893 if (F->use_empty() && (F->hasLocalLinkage() || F->hasLinkOnceLinkage())) {
1894 F->eraseFromParent();
1895 Changed = true;
1896 ++NumFnDeleted;
1897 } else if (F->hasLocalLinkage()) {
1898 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1899 !F->hasAddressTaken()) {
1900 // If this function has C calling conventions, is not a varargs
1901 // function, and is only called directly, promote it to use the Fast
1902 // calling convention.
1903 F->setCallingConv(CallingConv::Fast);
1904 ChangeCalleesToFastCall(F);
1905 ++NumFastCallFns;
1906 Changed = true;
1907 }
1908
1909 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
1910 !F->hasAddressTaken()) {
1911 // The function is not used by a trampoline intrinsic, so it is safe
1912 // to remove the 'nest' attribute.
1913 RemoveNestAttribute(F);
1914 ++NumNestRemoved;
1915 Changed = true;
1916 }
1917 }
1918 }
1919 return Changed;
1920 }
1921
OptimizeGlobalVars(Module & M)1922 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1923 bool Changed = false;
1924 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1925 GVI != E; ) {
1926 GlobalVariable *GV = GVI++;
1927 // Global variables without names cannot be referenced outside this module.
1928 if (!GV->hasName() && !GV->isDeclaration())
1929 GV->setLinkage(GlobalValue::InternalLinkage);
1930 // Simplify the initializer.
1931 if (GV->hasInitializer())
1932 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GV->getInitializer())) {
1933 TargetData *TD = getAnalysisIfAvailable<TargetData>();
1934 Constant *New = ConstantFoldConstantExpression(CE, TD);
1935 if (New && New != CE)
1936 GV->setInitializer(New);
1937 }
1938
1939 Changed |= ProcessGlobal(GV, GVI);
1940 }
1941 return Changed;
1942 }
1943
1944 /// FindGlobalCtors - Find the llvm.global_ctors list, verifying that all
1945 /// initializers have an init priority of 65535.
FindGlobalCtors(Module & M)1946 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1947 GlobalVariable *GV = M.getGlobalVariable("llvm.global_ctors");
1948 if (GV == 0) return 0;
1949
1950 // Verify that the initializer is simple enough for us to handle. We are
1951 // only allowed to optimize the initializer if it is unique.
1952 if (!GV->hasUniqueInitializer()) return 0;
1953
1954 if (isa<ConstantAggregateZero>(GV->getInitializer()))
1955 return GV;
1956 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1957
1958 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
1959 if (isa<ConstantAggregateZero>(*i))
1960 continue;
1961 ConstantStruct *CS = cast<ConstantStruct>(*i);
1962 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1963 continue;
1964
1965 // Must have a function or null ptr.
1966 if (!isa<Function>(CS->getOperand(1)))
1967 return 0;
1968
1969 // Init priority must be standard.
1970 ConstantInt *CI = cast<ConstantInt>(CS->getOperand(0));
1971 if (CI->getZExtValue() != 65535)
1972 return 0;
1973 }
1974
1975 return GV;
1976 }
1977
1978 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1979 /// return a list of the functions and null terminator as a vector.
ParseGlobalCtors(GlobalVariable * GV)1980 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1981 if (GV->getInitializer()->isNullValue())
1982 return std::vector<Function*>();
1983 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1984 std::vector<Function*> Result;
1985 Result.reserve(CA->getNumOperands());
1986 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
1987 ConstantStruct *CS = cast<ConstantStruct>(*i);
1988 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1989 }
1990 return Result;
1991 }
1992
1993 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1994 /// specified array, returning the new global to use.
InstallGlobalCtors(GlobalVariable * GCL,const std::vector<Function * > & Ctors)1995 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
1996 const std::vector<Function*> &Ctors) {
1997 // If we made a change, reassemble the initializer list.
1998 Constant *CSVals[2];
1999 CSVals[0] = ConstantInt::get(Type::getInt32Ty(GCL->getContext()), 65535);
2000 CSVals[1] = 0;
2001
2002 StructType *StructTy =
2003 cast <StructType>(
2004 cast<ArrayType>(GCL->getType()->getElementType())->getElementType());
2005
2006 // Create the new init list.
2007 std::vector<Constant*> CAList;
2008 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
2009 if (Ctors[i]) {
2010 CSVals[1] = Ctors[i];
2011 } else {
2012 Type *FTy = FunctionType::get(Type::getVoidTy(GCL->getContext()),
2013 false);
2014 PointerType *PFTy = PointerType::getUnqual(FTy);
2015 CSVals[1] = Constant::getNullValue(PFTy);
2016 CSVals[0] = ConstantInt::get(Type::getInt32Ty(GCL->getContext()),
2017 0x7fffffff);
2018 }
2019 CAList.push_back(ConstantStruct::get(StructTy, CSVals));
2020 }
2021
2022 // Create the array initializer.
2023 Constant *CA = ConstantArray::get(ArrayType::get(StructTy,
2024 CAList.size()), CAList);
2025
2026 // If we didn't change the number of elements, don't create a new GV.
2027 if (CA->getType() == GCL->getInitializer()->getType()) {
2028 GCL->setInitializer(CA);
2029 return GCL;
2030 }
2031
2032 // Create the new global and insert it next to the existing list.
2033 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
2034 GCL->getLinkage(), CA, "",
2035 GCL->isThreadLocal());
2036 GCL->getParent()->getGlobalList().insert(GCL, NGV);
2037 NGV->takeName(GCL);
2038
2039 // Nuke the old list, replacing any uses with the new one.
2040 if (!GCL->use_empty()) {
2041 Constant *V = NGV;
2042 if (V->getType() != GCL->getType())
2043 V = ConstantExpr::getBitCast(V, GCL->getType());
2044 GCL->replaceAllUsesWith(V);
2045 }
2046 GCL->eraseFromParent();
2047
2048 if (Ctors.size())
2049 return NGV;
2050 else
2051 return 0;
2052 }
2053
2054
getVal(DenseMap<Value *,Constant * > & ComputedValues,Value * V)2055 static Constant *getVal(DenseMap<Value*, Constant*> &ComputedValues, Value *V) {
2056 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
2057 Constant *R = ComputedValues[V];
2058 assert(R && "Reference to an uncomputed value!");
2059 return R;
2060 }
2061
2062 static inline bool
2063 isSimpleEnoughValueToCommit(Constant *C,
2064 SmallPtrSet<Constant*, 8> &SimpleConstants);
2065
2066
2067 /// isSimpleEnoughValueToCommit - Return true if the specified constant can be
2068 /// handled by the code generator. We don't want to generate something like:
2069 /// void *X = &X/42;
2070 /// because the code generator doesn't have a relocation that can handle that.
2071 ///
2072 /// This function should be called if C was not found (but just got inserted)
2073 /// in SimpleConstants to avoid having to rescan the same constants all the
2074 /// time.
isSimpleEnoughValueToCommitHelper(Constant * C,SmallPtrSet<Constant *,8> & SimpleConstants)2075 static bool isSimpleEnoughValueToCommitHelper(Constant *C,
2076 SmallPtrSet<Constant*, 8> &SimpleConstants) {
2077 // Simple integer, undef, constant aggregate zero, global addresses, etc are
2078 // all supported.
2079 if (C->getNumOperands() == 0 || isa<BlockAddress>(C) ||
2080 isa<GlobalValue>(C))
2081 return true;
2082
2083 // Aggregate values are safe if all their elements are.
2084 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C) ||
2085 isa<ConstantVector>(C)) {
2086 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) {
2087 Constant *Op = cast<Constant>(C->getOperand(i));
2088 if (!isSimpleEnoughValueToCommit(Op, SimpleConstants))
2089 return false;
2090 }
2091 return true;
2092 }
2093
2094 // We don't know exactly what relocations are allowed in constant expressions,
2095 // so we allow &global+constantoffset, which is safe and uniformly supported
2096 // across targets.
2097 ConstantExpr *CE = cast<ConstantExpr>(C);
2098 switch (CE->getOpcode()) {
2099 case Instruction::BitCast:
2100 case Instruction::IntToPtr:
2101 case Instruction::PtrToInt:
2102 // These casts are always fine if the casted value is.
2103 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants);
2104
2105 // GEP is fine if it is simple + constant offset.
2106 case Instruction::GetElementPtr:
2107 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
2108 if (!isa<ConstantInt>(CE->getOperand(i)))
2109 return false;
2110 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants);
2111
2112 case Instruction::Add:
2113 // We allow simple+cst.
2114 if (!isa<ConstantInt>(CE->getOperand(1)))
2115 return false;
2116 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants);
2117 }
2118 return false;
2119 }
2120
2121 static inline bool
isSimpleEnoughValueToCommit(Constant * C,SmallPtrSet<Constant *,8> & SimpleConstants)2122 isSimpleEnoughValueToCommit(Constant *C,
2123 SmallPtrSet<Constant*, 8> &SimpleConstants) {
2124 // If we already checked this constant, we win.
2125 if (!SimpleConstants.insert(C)) return true;
2126 // Check the constant.
2127 return isSimpleEnoughValueToCommitHelper(C, SimpleConstants);
2128 }
2129
2130
2131 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
2132 /// enough for us to understand. In particular, if it is a cast to anything
2133 /// other than from one pointer type to another pointer type, we punt.
2134 /// We basically just support direct accesses to globals and GEP's of
2135 /// globals. This should be kept up to date with CommitValueTo.
isSimpleEnoughPointerToCommit(Constant * C)2136 static bool isSimpleEnoughPointerToCommit(Constant *C) {
2137 // Conservatively, avoid aggregate types. This is because we don't
2138 // want to worry about them partially overlapping other stores.
2139 if (!cast<PointerType>(C->getType())->getElementType()->isSingleValueType())
2140 return false;
2141
2142 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
2143 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2144 // external globals.
2145 return GV->hasUniqueInitializer();
2146
2147 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2148 // Handle a constantexpr gep.
2149 if (CE->getOpcode() == Instruction::GetElementPtr &&
2150 isa<GlobalVariable>(CE->getOperand(0)) &&
2151 cast<GEPOperator>(CE)->isInBounds()) {
2152 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2153 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2154 // external globals.
2155 if (!GV->hasUniqueInitializer())
2156 return false;
2157
2158 // The first index must be zero.
2159 ConstantInt *CI = dyn_cast<ConstantInt>(*llvm::next(CE->op_begin()));
2160 if (!CI || !CI->isZero()) return false;
2161
2162 // The remaining indices must be compile-time known integers within the
2163 // notional bounds of the corresponding static array types.
2164 if (!CE->isGEPWithNoNotionalOverIndexing())
2165 return false;
2166
2167 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2168
2169 // A constantexpr bitcast from a pointer to another pointer is a no-op,
2170 // and we know how to evaluate it by moving the bitcast from the pointer
2171 // operand to the value operand.
2172 } else if (CE->getOpcode() == Instruction::BitCast &&
2173 isa<GlobalVariable>(CE->getOperand(0))) {
2174 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2175 // external globals.
2176 return cast<GlobalVariable>(CE->getOperand(0))->hasUniqueInitializer();
2177 }
2178 }
2179
2180 return false;
2181 }
2182
2183 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
2184 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
2185 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
EvaluateStoreInto(Constant * Init,Constant * Val,ConstantExpr * Addr,unsigned OpNo)2186 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
2187 ConstantExpr *Addr, unsigned OpNo) {
2188 // Base case of the recursion.
2189 if (OpNo == Addr->getNumOperands()) {
2190 assert(Val->getType() == Init->getType() && "Type mismatch!");
2191 return Val;
2192 }
2193
2194 std::vector<Constant*> Elts;
2195 if (StructType *STy = dyn_cast<StructType>(Init->getType())) {
2196
2197 // Break up the constant into its elements.
2198 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
2199 for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i)
2200 Elts.push_back(cast<Constant>(*i));
2201 } else if (isa<ConstantAggregateZero>(Init)) {
2202 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2203 Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
2204 } else if (isa<UndefValue>(Init)) {
2205 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2206 Elts.push_back(UndefValue::get(STy->getElementType(i)));
2207 } else {
2208 llvm_unreachable("This code is out of sync with "
2209 " ConstantFoldLoadThroughGEPConstantExpr");
2210 }
2211
2212 // Replace the element that we are supposed to.
2213 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
2214 unsigned Idx = CU->getZExtValue();
2215 assert(Idx < STy->getNumElements() && "Struct index out of range!");
2216 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
2217
2218 // Return the modified struct.
2219 return ConstantStruct::get(STy, Elts);
2220 }
2221
2222 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
2223 SequentialType *InitTy = cast<SequentialType>(Init->getType());
2224
2225 uint64_t NumElts;
2226 if (ArrayType *ATy = dyn_cast<ArrayType>(InitTy))
2227 NumElts = ATy->getNumElements();
2228 else
2229 NumElts = cast<VectorType>(InitTy)->getNumElements();
2230
2231 // Break up the array into elements.
2232 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
2233 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
2234 Elts.push_back(cast<Constant>(*i));
2235 } else if (ConstantVector *CV = dyn_cast<ConstantVector>(Init)) {
2236 for (User::op_iterator i = CV->op_begin(), e = CV->op_end(); i != e; ++i)
2237 Elts.push_back(cast<Constant>(*i));
2238 } else if (isa<ConstantAggregateZero>(Init)) {
2239 Elts.assign(NumElts, Constant::getNullValue(InitTy->getElementType()));
2240 } else {
2241 assert(isa<UndefValue>(Init) && "This code is out of sync with "
2242 " ConstantFoldLoadThroughGEPConstantExpr");
2243 Elts.assign(NumElts, UndefValue::get(InitTy->getElementType()));
2244 }
2245
2246 assert(CI->getZExtValue() < NumElts);
2247 Elts[CI->getZExtValue()] =
2248 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
2249
2250 if (Init->getType()->isArrayTy())
2251 return ConstantArray::get(cast<ArrayType>(InitTy), Elts);
2252 return ConstantVector::get(Elts);
2253 }
2254
2255 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
2256 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
CommitValueTo(Constant * Val,Constant * Addr)2257 static void CommitValueTo(Constant *Val, Constant *Addr) {
2258 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
2259 assert(GV->hasInitializer());
2260 GV->setInitializer(Val);
2261 return;
2262 }
2263
2264 ConstantExpr *CE = cast<ConstantExpr>(Addr);
2265 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2266 GV->setInitializer(EvaluateStoreInto(GV->getInitializer(), Val, CE, 2));
2267 }
2268
2269 /// ComputeLoadResult - Return the value that would be computed by a load from
2270 /// P after the stores reflected by 'memory' have been performed. If we can't
2271 /// decide, return null.
ComputeLoadResult(Constant * P,const DenseMap<Constant *,Constant * > & Memory)2272 static Constant *ComputeLoadResult(Constant *P,
2273 const DenseMap<Constant*, Constant*> &Memory) {
2274 // If this memory location has been recently stored, use the stored value: it
2275 // is the most up-to-date.
2276 DenseMap<Constant*, Constant*>::const_iterator I = Memory.find(P);
2277 if (I != Memory.end()) return I->second;
2278
2279 // Access it.
2280 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
2281 if (GV->hasDefinitiveInitializer())
2282 return GV->getInitializer();
2283 return 0;
2284 }
2285
2286 // Handle a constantexpr getelementptr.
2287 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
2288 if (CE->getOpcode() == Instruction::GetElementPtr &&
2289 isa<GlobalVariable>(CE->getOperand(0))) {
2290 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2291 if (GV->hasDefinitiveInitializer())
2292 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2293 }
2294
2295 return 0; // don't know how to evaluate.
2296 }
2297
2298 /// EvaluateFunction - Evaluate a call to function F, returning true if
2299 /// successful, false if we can't evaluate it. ActualArgs contains the formal
2300 /// arguments for the function.
EvaluateFunction(Function * F,Constant * & RetVal,const SmallVectorImpl<Constant * > & ActualArgs,std::vector<Function * > & CallStack,DenseMap<Constant *,Constant * > & MutatedMemory,std::vector<GlobalVariable * > & AllocaTmps,SmallPtrSet<Constant *,8> & SimpleConstants,const TargetData * TD)2301 static bool EvaluateFunction(Function *F, Constant *&RetVal,
2302 const SmallVectorImpl<Constant*> &ActualArgs,
2303 std::vector<Function*> &CallStack,
2304 DenseMap<Constant*, Constant*> &MutatedMemory,
2305 std::vector<GlobalVariable*> &AllocaTmps,
2306 SmallPtrSet<Constant*, 8> &SimpleConstants,
2307 const TargetData *TD) {
2308 // Check to see if this function is already executing (recursion). If so,
2309 // bail out. TODO: we might want to accept limited recursion.
2310 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
2311 return false;
2312
2313 CallStack.push_back(F);
2314
2315 /// Values - As we compute SSA register values, we store their contents here.
2316 DenseMap<Value*, Constant*> Values;
2317
2318 // Initialize arguments to the incoming values specified.
2319 unsigned ArgNo = 0;
2320 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
2321 ++AI, ++ArgNo)
2322 Values[AI] = ActualArgs[ArgNo];
2323
2324 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
2325 /// we can only evaluate any one basic block at most once. This set keeps
2326 /// track of what we have executed so we can detect recursive cases etc.
2327 SmallPtrSet<BasicBlock*, 32> ExecutedBlocks;
2328
2329 // CurInst - The current instruction we're evaluating.
2330 BasicBlock::iterator CurInst = F->begin()->begin();
2331
2332 // This is the main evaluation loop.
2333 while (1) {
2334 Constant *InstResult = 0;
2335
2336 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
2337 if (!SI->isSimple()) return false; // no volatile/atomic accesses.
2338 Constant *Ptr = getVal(Values, SI->getOperand(1));
2339 if (!isSimpleEnoughPointerToCommit(Ptr))
2340 // If this is too complex for us to commit, reject it.
2341 return false;
2342
2343 Constant *Val = getVal(Values, SI->getOperand(0));
2344
2345 // If this might be too difficult for the backend to handle (e.g. the addr
2346 // of one global variable divided by another) then we can't commit it.
2347 if (!isSimpleEnoughValueToCommit(Val, SimpleConstants))
2348 return false;
2349
2350 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
2351 if (CE->getOpcode() == Instruction::BitCast) {
2352 // If we're evaluating a store through a bitcast, then we need
2353 // to pull the bitcast off the pointer type and push it onto the
2354 // stored value.
2355 Ptr = CE->getOperand(0);
2356
2357 Type *NewTy=cast<PointerType>(Ptr->getType())->getElementType();
2358
2359 // In order to push the bitcast onto the stored value, a bitcast
2360 // from NewTy to Val's type must be legal. If it's not, we can try
2361 // introspecting NewTy to find a legal conversion.
2362 while (!Val->getType()->canLosslesslyBitCastTo(NewTy)) {
2363 // If NewTy is a struct, we can convert the pointer to the struct
2364 // into a pointer to its first member.
2365 // FIXME: This could be extended to support arrays as well.
2366 if (StructType *STy = dyn_cast<StructType>(NewTy)) {
2367 NewTy = STy->getTypeAtIndex(0U);
2368
2369 IntegerType *IdxTy =IntegerType::get(NewTy->getContext(), 32);
2370 Constant *IdxZero = ConstantInt::get(IdxTy, 0, false);
2371 Constant * const IdxList[] = {IdxZero, IdxZero};
2372
2373 Ptr = ConstantExpr::getGetElementPtr(Ptr, IdxList);
2374
2375 // If we can't improve the situation by introspecting NewTy,
2376 // we have to give up.
2377 } else {
2378 return 0;
2379 }
2380 }
2381
2382 // If we found compatible types, go ahead and push the bitcast
2383 // onto the stored value.
2384 Val = ConstantExpr::getBitCast(Val, NewTy);
2385 }
2386
2387 MutatedMemory[Ptr] = Val;
2388 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2389 InstResult = ConstantExpr::get(BO->getOpcode(),
2390 getVal(Values, BO->getOperand(0)),
2391 getVal(Values, BO->getOperand(1)));
2392 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2393 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2394 getVal(Values, CI->getOperand(0)),
2395 getVal(Values, CI->getOperand(1)));
2396 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2397 InstResult = ConstantExpr::getCast(CI->getOpcode(),
2398 getVal(Values, CI->getOperand(0)),
2399 CI->getType());
2400 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2401 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
2402 getVal(Values, SI->getOperand(1)),
2403 getVal(Values, SI->getOperand(2)));
2404 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2405 Constant *P = getVal(Values, GEP->getOperand(0));
2406 SmallVector<Constant*, 8> GEPOps;
2407 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2408 i != e; ++i)
2409 GEPOps.push_back(getVal(Values, *i));
2410 InstResult =
2411 ConstantExpr::getGetElementPtr(P, GEPOps,
2412 cast<GEPOperator>(GEP)->isInBounds());
2413 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2414 if (!LI->isSimple()) return false; // no volatile/atomic accesses.
2415 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
2416 MutatedMemory);
2417 if (InstResult == 0) return false; // Could not evaluate load.
2418 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2419 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs.
2420 Type *Ty = AI->getType()->getElementType();
2421 AllocaTmps.push_back(new GlobalVariable(Ty, false,
2422 GlobalValue::InternalLinkage,
2423 UndefValue::get(Ty),
2424 AI->getName()));
2425 InstResult = AllocaTmps.back();
2426 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
2427
2428 // Debug info can safely be ignored here.
2429 if (isa<DbgInfoIntrinsic>(CI)) {
2430 ++CurInst;
2431 continue;
2432 }
2433
2434 // Cannot handle inline asm.
2435 if (isa<InlineAsm>(CI->getCalledValue())) return false;
2436
2437 if (MemSetInst *MSI = dyn_cast<MemSetInst>(CI)) {
2438 if (MSI->isVolatile()) return false;
2439 Constant *Ptr = getVal(Values, MSI->getDest());
2440 Constant *Val = getVal(Values, MSI->getValue());
2441 Constant *DestVal = ComputeLoadResult(getVal(Values, Ptr),
2442 MutatedMemory);
2443 if (Val->isNullValue() && DestVal && DestVal->isNullValue()) {
2444 // This memset is a no-op.
2445 ++CurInst;
2446 continue;
2447 }
2448 return false;
2449 }
2450
2451 // Resolve function pointers.
2452 Function *Callee = dyn_cast<Function>(getVal(Values,
2453 CI->getCalledValue()));
2454 if (!Callee) return false; // Cannot resolve.
2455
2456 SmallVector<Constant*, 8> Formals;
2457 CallSite CS(CI);
2458 for (User::op_iterator i = CS.arg_begin(), e = CS.arg_end();
2459 i != e; ++i)
2460 Formals.push_back(getVal(Values, *i));
2461
2462 if (Callee->isDeclaration()) {
2463 // If this is a function we can constant fold, do it.
2464 if (Constant *C = ConstantFoldCall(Callee, Formals)) {
2465 InstResult = C;
2466 } else {
2467 return false;
2468 }
2469 } else {
2470 if (Callee->getFunctionType()->isVarArg())
2471 return false;
2472
2473 Constant *RetVal;
2474 // Execute the call, if successful, use the return value.
2475 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
2476 MutatedMemory, AllocaTmps, SimpleConstants, TD))
2477 return false;
2478 InstResult = RetVal;
2479 }
2480 } else if (isa<TerminatorInst>(CurInst)) {
2481 BasicBlock *NewBB = 0;
2482 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2483 if (BI->isUnconditional()) {
2484 NewBB = BI->getSuccessor(0);
2485 } else {
2486 ConstantInt *Cond =
2487 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
2488 if (!Cond) return false; // Cannot determine.
2489
2490 NewBB = BI->getSuccessor(!Cond->getZExtValue());
2491 }
2492 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2493 ConstantInt *Val =
2494 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
2495 if (!Val) return false; // Cannot determine.
2496 NewBB = SI->getSuccessor(SI->findCaseValue(Val));
2497 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(CurInst)) {
2498 Value *Val = getVal(Values, IBI->getAddress())->stripPointerCasts();
2499 if (BlockAddress *BA = dyn_cast<BlockAddress>(Val))
2500 NewBB = BA->getBasicBlock();
2501 else
2502 return false; // Cannot determine.
2503 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
2504 if (RI->getNumOperands())
2505 RetVal = getVal(Values, RI->getOperand(0));
2506
2507 CallStack.pop_back(); // return from fn.
2508 return true; // We succeeded at evaluating this ctor!
2509 } else {
2510 // invoke, unwind, resume, unreachable.
2511 return false; // Cannot handle this terminator.
2512 }
2513
2514 // Okay, we succeeded in evaluating this control flow. See if we have
2515 // executed the new block before. If so, we have a looping function,
2516 // which we cannot evaluate in reasonable time.
2517 if (!ExecutedBlocks.insert(NewBB))
2518 return false; // looped!
2519
2520 // Okay, we have never been in this block before. Check to see if there
2521 // are any PHI nodes. If so, evaluate them with information about where
2522 // we came from.
2523 BasicBlock *OldBB = CurInst->getParent();
2524 CurInst = NewBB->begin();
2525 PHINode *PN;
2526 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2527 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
2528
2529 // Do NOT increment CurInst. We know that the terminator had no value.
2530 continue;
2531 } else {
2532 // Did not know how to evaluate this!
2533 return false;
2534 }
2535
2536 if (!CurInst->use_empty()) {
2537 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(InstResult))
2538 InstResult = ConstantFoldConstantExpression(CE, TD);
2539
2540 Values[CurInst] = InstResult;
2541 }
2542
2543 // Advance program counter.
2544 ++CurInst;
2545 }
2546 }
2547
2548 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2549 /// we can. Return true if we can, false otherwise.
EvaluateStaticConstructor(Function * F,const TargetData * TD)2550 static bool EvaluateStaticConstructor(Function *F, const TargetData *TD) {
2551 /// MutatedMemory - For each store we execute, we update this map. Loads
2552 /// check this to get the most up-to-date value. If evaluation is successful,
2553 /// this state is committed to the process.
2554 DenseMap<Constant*, Constant*> MutatedMemory;
2555
2556 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2557 /// to represent its body. This vector is needed so we can delete the
2558 /// temporary globals when we are done.
2559 std::vector<GlobalVariable*> AllocaTmps;
2560
2561 /// CallStack - This is used to detect recursion. In pathological situations
2562 /// we could hit exponential behavior, but at least there is nothing
2563 /// unbounded.
2564 std::vector<Function*> CallStack;
2565
2566 /// SimpleConstants - These are constants we have checked and know to be
2567 /// simple enough to live in a static initializer of a global.
2568 SmallPtrSet<Constant*, 8> SimpleConstants;
2569
2570 // Call the function.
2571 Constant *RetValDummy;
2572 bool EvalSuccess = EvaluateFunction(F, RetValDummy,
2573 SmallVector<Constant*, 0>(), CallStack,
2574 MutatedMemory, AllocaTmps,
2575 SimpleConstants, TD);
2576
2577 if (EvalSuccess) {
2578 // We succeeded at evaluation: commit the result.
2579 DEBUG(dbgs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2580 << F->getName() << "' to " << MutatedMemory.size()
2581 << " stores.\n");
2582 for (DenseMap<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2583 E = MutatedMemory.end(); I != E; ++I)
2584 CommitValueTo(I->second, I->first);
2585 }
2586
2587 // At this point, we are done interpreting. If we created any 'alloca'
2588 // temporaries, release them now.
2589 while (!AllocaTmps.empty()) {
2590 GlobalVariable *Tmp = AllocaTmps.back();
2591 AllocaTmps.pop_back();
2592
2593 // If there are still users of the alloca, the program is doing something
2594 // silly, e.g. storing the address of the alloca somewhere and using it
2595 // later. Since this is undefined, we'll just make it be null.
2596 if (!Tmp->use_empty())
2597 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2598 delete Tmp;
2599 }
2600
2601 return EvalSuccess;
2602 }
2603
2604
2605
2606 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2607 /// Return true if anything changed.
OptimizeGlobalCtorsList(GlobalVariable * & GCL)2608 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2609 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2610 bool MadeChange = false;
2611 if (Ctors.empty()) return false;
2612
2613 const TargetData *TD = getAnalysisIfAvailable<TargetData>();
2614 // Loop over global ctors, optimizing them when we can.
2615 for (unsigned i = 0; i != Ctors.size(); ++i) {
2616 Function *F = Ctors[i];
2617 // Found a null terminator in the middle of the list, prune off the rest of
2618 // the list.
2619 if (F == 0) {
2620 if (i != Ctors.size()-1) {
2621 Ctors.resize(i+1);
2622 MadeChange = true;
2623 }
2624 break;
2625 }
2626
2627 // We cannot simplify external ctor functions.
2628 if (F->empty()) continue;
2629
2630 // If we can evaluate the ctor at compile time, do.
2631 if (EvaluateStaticConstructor(F, TD)) {
2632 Ctors.erase(Ctors.begin()+i);
2633 MadeChange = true;
2634 --i;
2635 ++NumCtorsEvaluated;
2636 continue;
2637 }
2638 }
2639
2640 if (!MadeChange) return false;
2641
2642 GCL = InstallGlobalCtors(GCL, Ctors);
2643 return true;
2644 }
2645
OptimizeGlobalAliases(Module & M)2646 bool GlobalOpt::OptimizeGlobalAliases(Module &M) {
2647 bool Changed = false;
2648
2649 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end();
2650 I != E;) {
2651 Module::alias_iterator J = I++;
2652 // Aliases without names cannot be referenced outside this module.
2653 if (!J->hasName() && !J->isDeclaration())
2654 J->setLinkage(GlobalValue::InternalLinkage);
2655 // If the aliasee may change at link time, nothing can be done - bail out.
2656 if (J->mayBeOverridden())
2657 continue;
2658
2659 Constant *Aliasee = J->getAliasee();
2660 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2661 Target->removeDeadConstantUsers();
2662 bool hasOneUse = Target->hasOneUse() && Aliasee->hasOneUse();
2663
2664 // Make all users of the alias use the aliasee instead.
2665 if (!J->use_empty()) {
2666 J->replaceAllUsesWith(Aliasee);
2667 ++NumAliasesResolved;
2668 Changed = true;
2669 }
2670
2671 // If the alias is externally visible, we may still be able to simplify it.
2672 if (!J->hasLocalLinkage()) {
2673 // If the aliasee has internal linkage, give it the name and linkage
2674 // of the alias, and delete the alias. This turns:
2675 // define internal ... @f(...)
2676 // @a = alias ... @f
2677 // into:
2678 // define ... @a(...)
2679 if (!Target->hasLocalLinkage())
2680 continue;
2681
2682 // Do not perform the transform if multiple aliases potentially target the
2683 // aliasee. This check also ensures that it is safe to replace the section
2684 // and other attributes of the aliasee with those of the alias.
2685 if (!hasOneUse)
2686 continue;
2687
2688 // Give the aliasee the name, linkage and other attributes of the alias.
2689 Target->takeName(J);
2690 Target->setLinkage(J->getLinkage());
2691 Target->GlobalValue::copyAttributesFrom(J);
2692 }
2693
2694 // Delete the alias.
2695 M.getAliasList().erase(J);
2696 ++NumAliasesRemoved;
2697 Changed = true;
2698 }
2699
2700 return Changed;
2701 }
2702
FindCXAAtExit(Module & M)2703 static Function *FindCXAAtExit(Module &M) {
2704 Function *Fn = M.getFunction("__cxa_atexit");
2705
2706 if (!Fn)
2707 return 0;
2708
2709 FunctionType *FTy = Fn->getFunctionType();
2710
2711 // Checking that the function has the right return type, the right number of
2712 // parameters and that they all have pointer types should be enough.
2713 if (!FTy->getReturnType()->isIntegerTy() ||
2714 FTy->getNumParams() != 3 ||
2715 !FTy->getParamType(0)->isPointerTy() ||
2716 !FTy->getParamType(1)->isPointerTy() ||
2717 !FTy->getParamType(2)->isPointerTy())
2718 return 0;
2719
2720 return Fn;
2721 }
2722
2723 /// cxxDtorIsEmpty - Returns whether the given function is an empty C++
2724 /// destructor and can therefore be eliminated.
2725 /// Note that we assume that other optimization passes have already simplified
2726 /// the code so we only look for a function with a single basic block, where
2727 /// the only allowed instructions are 'ret' or 'call' to empty C++ dtor.
cxxDtorIsEmpty(const Function & Fn,SmallPtrSet<const Function *,8> & CalledFunctions)2728 static bool cxxDtorIsEmpty(const Function &Fn,
2729 SmallPtrSet<const Function *, 8> &CalledFunctions) {
2730 // FIXME: We could eliminate C++ destructors if they're readonly/readnone and
2731 // nounwind, but that doesn't seem worth doing.
2732 if (Fn.isDeclaration())
2733 return false;
2734
2735 if (++Fn.begin() != Fn.end())
2736 return false;
2737
2738 const BasicBlock &EntryBlock = Fn.getEntryBlock();
2739 for (BasicBlock::const_iterator I = EntryBlock.begin(), E = EntryBlock.end();
2740 I != E; ++I) {
2741 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
2742 // Ignore debug intrinsics.
2743 if (isa<DbgInfoIntrinsic>(CI))
2744 continue;
2745
2746 const Function *CalledFn = CI->getCalledFunction();
2747
2748 if (!CalledFn)
2749 return false;
2750
2751 SmallPtrSet<const Function *, 8> NewCalledFunctions(CalledFunctions);
2752
2753 // Don't treat recursive functions as empty.
2754 if (!NewCalledFunctions.insert(CalledFn))
2755 return false;
2756
2757 if (!cxxDtorIsEmpty(*CalledFn, NewCalledFunctions))
2758 return false;
2759 } else if (isa<ReturnInst>(*I))
2760 return true;
2761 else
2762 return false;
2763 }
2764
2765 return false;
2766 }
2767
OptimizeEmptyGlobalCXXDtors(Function * CXAAtExitFn)2768 bool GlobalOpt::OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) {
2769 /// Itanium C++ ABI p3.3.5:
2770 ///
2771 /// After constructing a global (or local static) object, that will require
2772 /// destruction on exit, a termination function is registered as follows:
2773 ///
2774 /// extern "C" int __cxa_atexit ( void (*f)(void *), void *p, void *d );
2775 ///
2776 /// This registration, e.g. __cxa_atexit(f,p,d), is intended to cause the
2777 /// call f(p) when DSO d is unloaded, before all such termination calls
2778 /// registered before this one. It returns zero if registration is
2779 /// successful, nonzero on failure.
2780
2781 // This pass will look for calls to __cxa_atexit where the function is trivial
2782 // and remove them.
2783 bool Changed = false;
2784
2785 for (Function::use_iterator I = CXAAtExitFn->use_begin(),
2786 E = CXAAtExitFn->use_end(); I != E;) {
2787 // We're only interested in calls. Theoretically, we could handle invoke
2788 // instructions as well, but neither llvm-gcc nor clang generate invokes
2789 // to __cxa_atexit.
2790 CallInst *CI = dyn_cast<CallInst>(*I++);
2791 if (!CI)
2792 continue;
2793
2794 Function *DtorFn =
2795 dyn_cast<Function>(CI->getArgOperand(0)->stripPointerCasts());
2796 if (!DtorFn)
2797 continue;
2798
2799 SmallPtrSet<const Function *, 8> CalledFunctions;
2800 if (!cxxDtorIsEmpty(*DtorFn, CalledFunctions))
2801 continue;
2802
2803 // Just remove the call.
2804 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
2805 CI->eraseFromParent();
2806
2807 ++NumCXXDtorsRemoved;
2808
2809 Changed |= true;
2810 }
2811
2812 return Changed;
2813 }
2814
runOnModule(Module & M)2815 bool GlobalOpt::runOnModule(Module &M) {
2816 bool Changed = false;
2817
2818 // Try to find the llvm.globalctors list.
2819 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2820
2821 Function *CXAAtExitFn = FindCXAAtExit(M);
2822
2823 bool LocalChange = true;
2824 while (LocalChange) {
2825 LocalChange = false;
2826
2827 // Delete functions that are trivially dead, ccc -> fastcc
2828 LocalChange |= OptimizeFunctions(M);
2829
2830 // Optimize global_ctors list.
2831 if (GlobalCtors)
2832 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2833
2834 // Optimize non-address-taken globals.
2835 LocalChange |= OptimizeGlobalVars(M);
2836
2837 // Resolve aliases, when possible.
2838 LocalChange |= OptimizeGlobalAliases(M);
2839
2840 // Try to remove trivial global destructors.
2841 if (CXAAtExitFn)
2842 LocalChange |= OptimizeEmptyGlobalCXXDtors(CXAAtExitFn);
2843
2844 Changed |= LocalChange;
2845 }
2846
2847 // TODO: Move all global ctors functions to the end of the module for code
2848 // layout.
2849
2850 return Changed;
2851 }
2852