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__anon4a207e4e0111::GlobalOpt63 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
64 }
65 static char ID; // Pass identification, replacement for typeid
GlobalOpt__anon4a207e4e0111::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__anon4a207e4e0211::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 if (LI->isVolatile()) return true; // Don't hack on volatile loads.
199 } else if (const StoreInst *SI = dyn_cast<StoreInst>(I)) {
200 // Don't allow a store OF the address, only stores TO the address.
201 if (SI->getOperand(0) == V) return true;
202
203 if (SI->isVolatile()) return true; // Don't hack on volatile stores.
204
205 // If this is a direct store to the global (i.e., the global is a scalar
206 // value, not an aggregate), keep more specific information about
207 // stores.
208 if (GS.StoredType != GlobalStatus::isStored) {
209 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(
210 SI->getOperand(1))) {
211 Value *StoredVal = SI->getOperand(0);
212 if (StoredVal == GV->getInitializer()) {
213 if (GS.StoredType < GlobalStatus::isInitializerStored)
214 GS.StoredType = GlobalStatus::isInitializerStored;
215 } else if (isa<LoadInst>(StoredVal) &&
216 cast<LoadInst>(StoredVal)->getOperand(0) == GV) {
217 if (GS.StoredType < GlobalStatus::isInitializerStored)
218 GS.StoredType = GlobalStatus::isInitializerStored;
219 } else if (GS.StoredType < GlobalStatus::isStoredOnce) {
220 GS.StoredType = GlobalStatus::isStoredOnce;
221 GS.StoredOnceValue = StoredVal;
222 } else if (GS.StoredType == GlobalStatus::isStoredOnce &&
223 GS.StoredOnceValue == StoredVal) {
224 // noop.
225 } else {
226 GS.StoredType = GlobalStatus::isStored;
227 }
228 } else {
229 GS.StoredType = GlobalStatus::isStored;
230 }
231 }
232 } else if (isa<GetElementPtrInst>(I)) {
233 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
234 } else if (isa<SelectInst>(I)) {
235 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
236 } else if (const PHINode *PN = dyn_cast<PHINode>(I)) {
237 // PHI nodes we can check just like select or GEP instructions, but we
238 // have to be careful about infinite recursion.
239 if (PHIUsers.insert(PN)) // Not already visited.
240 if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
241 GS.HasPHIUser = true;
242 } else if (isa<CmpInst>(I)) {
243 GS.isCompared = true;
244 } else if (const MemTransferInst *MTI = dyn_cast<MemTransferInst>(I)) {
245 if (MTI->isVolatile()) return true;
246 if (MTI->getArgOperand(0) == V)
247 GS.StoredType = GlobalStatus::isStored;
248 if (MTI->getArgOperand(1) == V)
249 GS.isLoaded = true;
250 } else if (const MemSetInst *MSI = dyn_cast<MemSetInst>(I)) {
251 assert(MSI->getArgOperand(0) == V && "Memset only takes one pointer!");
252 if (MSI->isVolatile()) return true;
253 GS.StoredType = GlobalStatus::isStored;
254 } else {
255 return true; // Any other non-load instruction might take address!
256 }
257 } else if (const Constant *C = dyn_cast<Constant>(U)) {
258 GS.HasNonInstructionUser = true;
259 // We might have a dead and dangling constant hanging off of here.
260 if (!SafeToDestroyConstant(C))
261 return true;
262 } else {
263 GS.HasNonInstructionUser = true;
264 // Otherwise must be some other user.
265 return true;
266 }
267 }
268
269 return false;
270 }
271
getAggregateConstantElement(Constant * Agg,Constant * Idx)272 static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) {
273 ConstantInt *CI = dyn_cast<ConstantInt>(Idx);
274 if (!CI) return 0;
275 unsigned IdxV = CI->getZExtValue();
276
277 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Agg)) {
278 if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV);
279 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Agg)) {
280 if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV);
281 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Agg)) {
282 if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV);
283 } else if (isa<ConstantAggregateZero>(Agg)) {
284 if (StructType *STy = dyn_cast<StructType>(Agg->getType())) {
285 if (IdxV < STy->getNumElements())
286 return Constant::getNullValue(STy->getElementType(IdxV));
287 } else if (SequentialType *STy =
288 dyn_cast<SequentialType>(Agg->getType())) {
289 return Constant::getNullValue(STy->getElementType());
290 }
291 } else if (isa<UndefValue>(Agg)) {
292 if (StructType *STy = dyn_cast<StructType>(Agg->getType())) {
293 if (IdxV < STy->getNumElements())
294 return UndefValue::get(STy->getElementType(IdxV));
295 } else if (SequentialType *STy =
296 dyn_cast<SequentialType>(Agg->getType())) {
297 return UndefValue::get(STy->getElementType());
298 }
299 }
300 return 0;
301 }
302
303
304 /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all
305 /// users of the global, cleaning up the obvious ones. This is largely just a
306 /// quick scan over the use list to clean up the easy and obvious cruft. This
307 /// returns true if it made a change.
CleanupConstantGlobalUsers(Value * V,Constant * Init)308 static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) {
309 bool Changed = false;
310 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) {
311 User *U = *UI++;
312
313 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
314 if (Init) {
315 // Replace the load with the initializer.
316 LI->replaceAllUsesWith(Init);
317 LI->eraseFromParent();
318 Changed = true;
319 }
320 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
321 // Store must be unreachable or storing Init into the global.
322 SI->eraseFromParent();
323 Changed = true;
324 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
325 if (CE->getOpcode() == Instruction::GetElementPtr) {
326 Constant *SubInit = 0;
327 if (Init)
328 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
329 Changed |= CleanupConstantGlobalUsers(CE, SubInit);
330 } else if (CE->getOpcode() == Instruction::BitCast &&
331 CE->getType()->isPointerTy()) {
332 // Pointer cast, delete any stores and memsets to the global.
333 Changed |= CleanupConstantGlobalUsers(CE, 0);
334 }
335
336 if (CE->use_empty()) {
337 CE->destroyConstant();
338 Changed = true;
339 }
340 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
341 // Do not transform "gepinst (gep constexpr (GV))" here, because forming
342 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
343 // and will invalidate our notion of what Init is.
344 Constant *SubInit = 0;
345 if (!isa<ConstantExpr>(GEP->getOperand(0))) {
346 ConstantExpr *CE =
347 dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
348 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
349 SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
350 }
351 Changed |= CleanupConstantGlobalUsers(GEP, SubInit);
352
353 if (GEP->use_empty()) {
354 GEP->eraseFromParent();
355 Changed = true;
356 }
357 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
358 if (MI->getRawDest() == V) {
359 MI->eraseFromParent();
360 Changed = true;
361 }
362
363 } else if (Constant *C = dyn_cast<Constant>(U)) {
364 // If we have a chain of dead constantexprs or other things dangling from
365 // us, and if they are all dead, nuke them without remorse.
366 if (SafeToDestroyConstant(C)) {
367 C->destroyConstant();
368 // This could have invalidated UI, start over from scratch.
369 CleanupConstantGlobalUsers(V, Init);
370 return true;
371 }
372 }
373 }
374 return Changed;
375 }
376
377 /// isSafeSROAElementUse - Return true if the specified instruction is a safe
378 /// user of a derived expression from a global that we want to SROA.
isSafeSROAElementUse(Value * V)379 static bool isSafeSROAElementUse(Value *V) {
380 // We might have a dead and dangling constant hanging off of here.
381 if (Constant *C = dyn_cast<Constant>(V))
382 return SafeToDestroyConstant(C);
383
384 Instruction *I = dyn_cast<Instruction>(V);
385 if (!I) return false;
386
387 // Loads are ok.
388 if (isa<LoadInst>(I)) return true;
389
390 // Stores *to* the pointer are ok.
391 if (StoreInst *SI = dyn_cast<StoreInst>(I))
392 return SI->getOperand(0) != V;
393
394 // Otherwise, it must be a GEP.
395 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I);
396 if (GEPI == 0) return false;
397
398 if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
399 !cast<Constant>(GEPI->getOperand(1))->isNullValue())
400 return false;
401
402 for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end();
403 I != E; ++I)
404 if (!isSafeSROAElementUse(*I))
405 return false;
406 return true;
407 }
408
409
410 /// IsUserOfGlobalSafeForSRA - U is a direct user of the specified global value.
411 /// Look at it and its uses and decide whether it is safe to SROA this global.
412 ///
IsUserOfGlobalSafeForSRA(User * U,GlobalValue * GV)413 static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
414 // The user of the global must be a GEP Inst or a ConstantExpr GEP.
415 if (!isa<GetElementPtrInst>(U) &&
416 (!isa<ConstantExpr>(U) ||
417 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
418 return false;
419
420 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
421 // don't like < 3 operand CE's, and we don't like non-constant integer
422 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some
423 // value of C.
424 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
425 !cast<Constant>(U->getOperand(1))->isNullValue() ||
426 !isa<ConstantInt>(U->getOperand(2)))
427 return false;
428
429 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
430 ++GEPI; // Skip over the pointer index.
431
432 // If this is a use of an array allocation, do a bit more checking for sanity.
433 if (ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
434 uint64_t NumElements = AT->getNumElements();
435 ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
436
437 // Check to make sure that index falls within the array. If not,
438 // something funny is going on, so we won't do the optimization.
439 //
440 if (Idx->getZExtValue() >= NumElements)
441 return false;
442
443 // We cannot scalar repl this level of the array unless any array
444 // sub-indices are in-range constants. In particular, consider:
445 // A[0][i]. We cannot know that the user isn't doing invalid things like
446 // allowing i to index an out-of-range subscript that accesses A[1].
447 //
448 // Scalar replacing *just* the outer index of the array is probably not
449 // going to be a win anyway, so just give up.
450 for (++GEPI; // Skip array index.
451 GEPI != E;
452 ++GEPI) {
453 uint64_t NumElements;
454 if (ArrayType *SubArrayTy = dyn_cast<ArrayType>(*GEPI))
455 NumElements = SubArrayTy->getNumElements();
456 else if (VectorType *SubVectorTy = dyn_cast<VectorType>(*GEPI))
457 NumElements = SubVectorTy->getNumElements();
458 else {
459 assert((*GEPI)->isStructTy() &&
460 "Indexed GEP type is not array, vector, or struct!");
461 continue;
462 }
463
464 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
465 if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
466 return false;
467 }
468 }
469
470 for (Value::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I)
471 if (!isSafeSROAElementUse(*I))
472 return false;
473 return true;
474 }
475
476 /// GlobalUsersSafeToSRA - Look at all uses of the global and decide whether it
477 /// is safe for us to perform this transformation.
478 ///
GlobalUsersSafeToSRA(GlobalValue * GV)479 static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
480 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end();
481 UI != E; ++UI) {
482 if (!IsUserOfGlobalSafeForSRA(*UI, GV))
483 return false;
484 }
485 return true;
486 }
487
488
489 /// SRAGlobal - Perform scalar replacement of aggregates on the specified global
490 /// variable. This opens the door for other optimizations by exposing the
491 /// behavior of the program in a more fine-grained way. We have determined that
492 /// this transformation is safe already. We return the first global variable we
493 /// insert so that the caller can reprocess it.
SRAGlobal(GlobalVariable * GV,const TargetData & TD)494 static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) {
495 // Make sure this global only has simple uses that we can SRA.
496 if (!GlobalUsersSafeToSRA(GV))
497 return 0;
498
499 assert(GV->hasLocalLinkage() && !GV->isConstant());
500 Constant *Init = GV->getInitializer();
501 Type *Ty = Init->getType();
502
503 std::vector<GlobalVariable*> NewGlobals;
504 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
505
506 // Get the alignment of the global, either explicit or target-specific.
507 unsigned StartAlignment = GV->getAlignment();
508 if (StartAlignment == 0)
509 StartAlignment = TD.getABITypeAlignment(GV->getType());
510
511 if (StructType *STy = dyn_cast<StructType>(Ty)) {
512 NewGlobals.reserve(STy->getNumElements());
513 const StructLayout &Layout = *TD.getStructLayout(STy);
514 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
515 Constant *In = getAggregateConstantElement(Init,
516 ConstantInt::get(Type::getInt32Ty(STy->getContext()), i));
517 assert(In && "Couldn't get element of initializer?");
518 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false,
519 GlobalVariable::InternalLinkage,
520 In, GV->getName()+"."+Twine(i),
521 GV->isThreadLocal(),
522 GV->getType()->getAddressSpace());
523 Globals.insert(GV, NGV);
524 NewGlobals.push_back(NGV);
525
526 // Calculate the known alignment of the field. If the original aggregate
527 // had 256 byte alignment for example, something might depend on that:
528 // propagate info to each field.
529 uint64_t FieldOffset = Layout.getElementOffset(i);
530 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, FieldOffset);
531 if (NewAlign > TD.getABITypeAlignment(STy->getElementType(i)))
532 NGV->setAlignment(NewAlign);
533 }
534 } else if (SequentialType *STy = dyn_cast<SequentialType>(Ty)) {
535 unsigned NumElements = 0;
536 if (ArrayType *ATy = dyn_cast<ArrayType>(STy))
537 NumElements = ATy->getNumElements();
538 else
539 NumElements = cast<VectorType>(STy)->getNumElements();
540
541 if (NumElements > 16 && GV->hasNUsesOrMore(16))
542 return 0; // It's not worth it.
543 NewGlobals.reserve(NumElements);
544
545 uint64_t EltSize = TD.getTypeAllocSize(STy->getElementType());
546 unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType());
547 for (unsigned i = 0, e = NumElements; i != e; ++i) {
548 Constant *In = getAggregateConstantElement(Init,
549 ConstantInt::get(Type::getInt32Ty(Init->getContext()), i));
550 assert(In && "Couldn't get element of initializer?");
551
552 GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false,
553 GlobalVariable::InternalLinkage,
554 In, GV->getName()+"."+Twine(i),
555 GV->isThreadLocal(),
556 GV->getType()->getAddressSpace());
557 Globals.insert(GV, NGV);
558 NewGlobals.push_back(NGV);
559
560 // Calculate the known alignment of the field. If the original aggregate
561 // had 256 byte alignment for example, something might depend on that:
562 // propagate info to each field.
563 unsigned NewAlign = (unsigned)MinAlign(StartAlignment, EltSize*i);
564 if (NewAlign > EltAlign)
565 NGV->setAlignment(NewAlign);
566 }
567 }
568
569 if (NewGlobals.empty())
570 return 0;
571
572 DEBUG(dbgs() << "PERFORMING GLOBAL SRA ON: " << *GV);
573
574 Constant *NullInt =Constant::getNullValue(Type::getInt32Ty(GV->getContext()));
575
576 // Loop over all of the uses of the global, replacing the constantexpr geps,
577 // with smaller constantexpr geps or direct references.
578 while (!GV->use_empty()) {
579 User *GEP = GV->use_back();
580 assert(((isa<ConstantExpr>(GEP) &&
581 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||
582 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!");
583
584 // Ignore the 1th operand, which has to be zero or else the program is quite
585 // broken (undefined). Get the 2nd operand, which is the structure or array
586 // index.
587 unsigned Val = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
588 if (Val >= NewGlobals.size()) Val = 0; // Out of bound array access.
589
590 Value *NewPtr = NewGlobals[Val];
591
592 // Form a shorter GEP if needed.
593 if (GEP->getNumOperands() > 3) {
594 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
595 SmallVector<Constant*, 8> Idxs;
596 Idxs.push_back(NullInt);
597 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
598 Idxs.push_back(CE->getOperand(i));
599 NewPtr = ConstantExpr::getGetElementPtr(cast<Constant>(NewPtr),
600 &Idxs[0], Idxs.size());
601 } else {
602 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
603 SmallVector<Value*, 8> Idxs;
604 Idxs.push_back(NullInt);
605 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
606 Idxs.push_back(GEPI->getOperand(i));
607 NewPtr = GetElementPtrInst::Create(NewPtr, Idxs.begin(), Idxs.end(),
608 GEPI->getName()+"."+Twine(Val),GEPI);
609 }
610 }
611 GEP->replaceAllUsesWith(NewPtr);
612
613 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
614 GEPI->eraseFromParent();
615 else
616 cast<ConstantExpr>(GEP)->destroyConstant();
617 }
618
619 // Delete the old global, now that it is dead.
620 Globals.erase(GV);
621 ++NumSRA;
622
623 // Loop over the new globals array deleting any globals that are obviously
624 // dead. This can arise due to scalarization of a structure or an array that
625 // has elements that are dead.
626 unsigned FirstGlobal = 0;
627 for (unsigned i = 0, e = NewGlobals.size(); i != e; ++i)
628 if (NewGlobals[i]->use_empty()) {
629 Globals.erase(NewGlobals[i]);
630 if (FirstGlobal == i) ++FirstGlobal;
631 }
632
633 return FirstGlobal != NewGlobals.size() ? NewGlobals[FirstGlobal] : 0;
634 }
635
636 /// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
637 /// value will trap if the value is dynamically null. PHIs keeps track of any
638 /// phi nodes we've seen to avoid reprocessing them.
AllUsesOfValueWillTrapIfNull(const Value * V,SmallPtrSet<const PHINode *,8> & PHIs)639 static bool AllUsesOfValueWillTrapIfNull(const Value *V,
640 SmallPtrSet<const PHINode*, 8> &PHIs) {
641 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;
642 ++UI) {
643 const User *U = *UI;
644
645 if (isa<LoadInst>(U)) {
646 // Will trap.
647 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
648 if (SI->getOperand(0) == V) {
649 //cerr << "NONTRAPPING USE: " << *U;
650 return false; // Storing the value.
651 }
652 } else if (const CallInst *CI = dyn_cast<CallInst>(U)) {
653 if (CI->getCalledValue() != V) {
654 //cerr << "NONTRAPPING USE: " << *U;
655 return false; // Not calling the ptr
656 }
657 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(U)) {
658 if (II->getCalledValue() != V) {
659 //cerr << "NONTRAPPING USE: " << *U;
660 return false; // Not calling the ptr
661 }
662 } else if (const BitCastInst *CI = dyn_cast<BitCastInst>(U)) {
663 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
664 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
665 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
666 } else if (const PHINode *PN = dyn_cast<PHINode>(U)) {
667 // If we've already seen this phi node, ignore it, it has already been
668 // checked.
669 if (PHIs.insert(PN) && !AllUsesOfValueWillTrapIfNull(PN, PHIs))
670 return false;
671 } else if (isa<ICmpInst>(U) &&
672 isa<ConstantPointerNull>(UI->getOperand(1))) {
673 // Ignore icmp X, null
674 } else {
675 //cerr << "NONTRAPPING USE: " << *U;
676 return false;
677 }
678 }
679 return true;
680 }
681
682 /// AllUsesOfLoadedValueWillTrapIfNull - Return true if all uses of any loads
683 /// from GV will trap if the loaded value is null. Note that this also permits
684 /// comparisons of the loaded value against null, as a special case.
AllUsesOfLoadedValueWillTrapIfNull(const GlobalVariable * GV)685 static bool AllUsesOfLoadedValueWillTrapIfNull(const GlobalVariable *GV) {
686 for (Value::const_use_iterator UI = GV->use_begin(), E = GV->use_end();
687 UI != E; ++UI) {
688 const User *U = *UI;
689
690 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
691 SmallPtrSet<const PHINode*, 8> PHIs;
692 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
693 return false;
694 } else if (isa<StoreInst>(U)) {
695 // Ignore stores to the global.
696 } else {
697 // We don't know or understand this user, bail out.
698 //cerr << "UNKNOWN USER OF GLOBAL!: " << *U;
699 return false;
700 }
701 }
702 return true;
703 }
704
OptimizeAwayTrappingUsesOfValue(Value * V,Constant * NewV)705 static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
706 bool Changed = false;
707 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) {
708 Instruction *I = cast<Instruction>(*UI++);
709 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
710 LI->setOperand(0, NewV);
711 Changed = true;
712 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
713 if (SI->getOperand(1) == V) {
714 SI->setOperand(1, NewV);
715 Changed = true;
716 }
717 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
718 CallSite CS(I);
719 if (CS.getCalledValue() == V) {
720 // Calling through the pointer! Turn into a direct call, but be careful
721 // that the pointer is not also being passed as an argument.
722 CS.setCalledFunction(NewV);
723 Changed = true;
724 bool PassedAsArg = false;
725 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
726 if (CS.getArgument(i) == V) {
727 PassedAsArg = true;
728 CS.setArgument(i, NewV);
729 }
730
731 if (PassedAsArg) {
732 // Being passed as an argument also. Be careful to not invalidate UI!
733 UI = V->use_begin();
734 }
735 }
736 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
737 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
738 ConstantExpr::getCast(CI->getOpcode(),
739 NewV, CI->getType()));
740 if (CI->use_empty()) {
741 Changed = true;
742 CI->eraseFromParent();
743 }
744 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
745 // Should handle GEP here.
746 SmallVector<Constant*, 8> Idxs;
747 Idxs.reserve(GEPI->getNumOperands()-1);
748 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
749 i != e; ++i)
750 if (Constant *C = dyn_cast<Constant>(*i))
751 Idxs.push_back(C);
752 else
753 break;
754 if (Idxs.size() == GEPI->getNumOperands()-1)
755 Changed |= OptimizeAwayTrappingUsesOfValue(GEPI,
756 ConstantExpr::getGetElementPtr(NewV, &Idxs[0],
757 Idxs.size()));
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,
1249 GEPIdx.begin(), GEPIdx.end(),
1250 GEPI->getName(), GEPI);
1251 GEPI->replaceAllUsesWith(NGEPI);
1252 GEPI->eraseFromParent();
1253 return;
1254 }
1255
1256 // Recursively transform the users of PHI nodes. This will lazily create the
1257 // PHIs that are needed for individual elements. Keep track of what PHIs we
1258 // see in InsertedScalarizedValues so that we don't get infinite loops (very
1259 // antisocial). If the PHI is already in InsertedScalarizedValues, it has
1260 // already been seen first by another load, so its uses have already been
1261 // processed.
1262 PHINode *PN = cast<PHINode>(LoadUser);
1263 bool Inserted;
1264 DenseMap<Value*, std::vector<Value*> >::iterator InsertPos;
1265 tie(InsertPos, Inserted) =
1266 InsertedScalarizedValues.insert(std::make_pair(PN, std::vector<Value*>()));
1267 if (!Inserted) return;
1268
1269 // If this is the first time we've seen this PHI, recursively process all
1270 // users.
1271 for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); UI != E; ) {
1272 Instruction *User = cast<Instruction>(*UI++);
1273 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1274 }
1275 }
1276
1277 /// RewriteUsesOfLoadForHeapSRoA - We are performing Heap SRoA on a global. Ptr
1278 /// is a value loaded from the global. Eliminate all uses of Ptr, making them
1279 /// use FieldGlobals instead. All uses of loaded values satisfy
1280 /// AllGlobalLoadUsesSimpleEnoughForHeapSRA.
RewriteUsesOfLoadForHeapSRoA(LoadInst * Load,DenseMap<Value *,std::vector<Value * >> & InsertedScalarizedValues,std::vector<std::pair<PHINode *,unsigned>> & PHIsToRewrite)1281 static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
1282 DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
1283 std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
1284 for (Value::use_iterator UI = Load->use_begin(), E = Load->use_end();
1285 UI != E; ) {
1286 Instruction *User = cast<Instruction>(*UI++);
1287 RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
1288 }
1289
1290 if (Load->use_empty()) {
1291 Load->eraseFromParent();
1292 InsertedScalarizedValues.erase(Load);
1293 }
1294 }
1295
1296 /// PerformHeapAllocSRoA - CI is an allocation of an array of structures. Break
1297 /// it up into multiple allocations of arrays of the fields.
PerformHeapAllocSRoA(GlobalVariable * GV,CallInst * CI,Value * NElems,TargetData * TD)1298 static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI,
1299 Value* NElems, TargetData *TD) {
1300 DEBUG(dbgs() << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *CI << '\n');
1301 Type* MAT = getMallocAllocatedType(CI);
1302 StructType *STy = cast<StructType>(MAT);
1303
1304 // There is guaranteed to be at least one use of the malloc (storing
1305 // it into GV). If there are other uses, change them to be uses of
1306 // the global to simplify later code. This also deletes the store
1307 // into GV.
1308 ReplaceUsesOfMallocWithGlobal(CI, GV);
1309
1310 // Okay, at this point, there are no users of the malloc. Insert N
1311 // new mallocs at the same place as CI, and N globals.
1312 std::vector<Value*> FieldGlobals;
1313 std::vector<Value*> FieldMallocs;
1314
1315 for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
1316 Type *FieldTy = STy->getElementType(FieldNo);
1317 PointerType *PFieldTy = PointerType::getUnqual(FieldTy);
1318
1319 GlobalVariable *NGV =
1320 new GlobalVariable(*GV->getParent(),
1321 PFieldTy, false, GlobalValue::InternalLinkage,
1322 Constant::getNullValue(PFieldTy),
1323 GV->getName() + ".f" + Twine(FieldNo), GV,
1324 GV->isThreadLocal());
1325 FieldGlobals.push_back(NGV);
1326
1327 unsigned TypeSize = TD->getTypeAllocSize(FieldTy);
1328 if (StructType *ST = dyn_cast<StructType>(FieldTy))
1329 TypeSize = TD->getStructLayout(ST)->getSizeInBytes();
1330 Type *IntPtrTy = TD->getIntPtrType(CI->getContext());
1331 Value *NMI = CallInst::CreateMalloc(CI, IntPtrTy, FieldTy,
1332 ConstantInt::get(IntPtrTy, TypeSize),
1333 NElems, 0,
1334 CI->getName() + ".f" + Twine(FieldNo));
1335 FieldMallocs.push_back(NMI);
1336 new StoreInst(NMI, NGV, CI);
1337 }
1338
1339 // The tricky aspect of this transformation is handling the case when malloc
1340 // fails. In the original code, malloc failing would set the result pointer
1341 // of malloc to null. In this case, some mallocs could succeed and others
1342 // could fail. As such, we emit code that looks like this:
1343 // F0 = malloc(field0)
1344 // F1 = malloc(field1)
1345 // F2 = malloc(field2)
1346 // if (F0 == 0 || F1 == 0 || F2 == 0) {
1347 // if (F0) { free(F0); F0 = 0; }
1348 // if (F1) { free(F1); F1 = 0; }
1349 // if (F2) { free(F2); F2 = 0; }
1350 // }
1351 // The malloc can also fail if its argument is too large.
1352 Constant *ConstantZero = ConstantInt::get(CI->getArgOperand(0)->getType(), 0);
1353 Value *RunningOr = new ICmpInst(CI, ICmpInst::ICMP_SLT, CI->getArgOperand(0),
1354 ConstantZero, "isneg");
1355 for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
1356 Value *Cond = new ICmpInst(CI, ICmpInst::ICMP_EQ, FieldMallocs[i],
1357 Constant::getNullValue(FieldMallocs[i]->getType()),
1358 "isnull");
1359 RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", CI);
1360 }
1361
1362 // Split the basic block at the old malloc.
1363 BasicBlock *OrigBB = CI->getParent();
1364 BasicBlock *ContBB = OrigBB->splitBasicBlock(CI, "malloc_cont");
1365
1366 // Create the block to check the first condition. Put all these blocks at the
1367 // end of the function as they are unlikely to be executed.
1368 BasicBlock *NullPtrBlock = BasicBlock::Create(OrigBB->getContext(),
1369 "malloc_ret_null",
1370 OrigBB->getParent());
1371
1372 // Remove the uncond branch from OrigBB to ContBB, turning it into a cond
1373 // branch on RunningOr.
1374 OrigBB->getTerminator()->eraseFromParent();
1375 BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
1376
1377 // Within the NullPtrBlock, we need to emit a comparison and branch for each
1378 // pointer, because some may be null while others are not.
1379 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1380 Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
1381 Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
1382 Constant::getNullValue(GVVal->getType()),
1383 "tmp");
1384 BasicBlock *FreeBlock = BasicBlock::Create(Cmp->getContext(), "free_it",
1385 OrigBB->getParent());
1386 BasicBlock *NextBlock = BasicBlock::Create(Cmp->getContext(), "next",
1387 OrigBB->getParent());
1388 Instruction *BI = BranchInst::Create(FreeBlock, NextBlock,
1389 Cmp, NullPtrBlock);
1390
1391 // Fill in FreeBlock.
1392 CallInst::CreateFree(GVVal, BI);
1393 new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
1394 FreeBlock);
1395 BranchInst::Create(NextBlock, FreeBlock);
1396
1397 NullPtrBlock = NextBlock;
1398 }
1399
1400 BranchInst::Create(ContBB, NullPtrBlock);
1401
1402 // CI is no longer needed, remove it.
1403 CI->eraseFromParent();
1404
1405 /// InsertedScalarizedLoads - As we process loads, if we can't immediately
1406 /// update all uses of the load, keep track of what scalarized loads are
1407 /// inserted for a given load.
1408 DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues;
1409 InsertedScalarizedValues[GV] = FieldGlobals;
1410
1411 std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite;
1412
1413 // Okay, the malloc site is completely handled. All of the uses of GV are now
1414 // loads, and all uses of those loads are simple. Rewrite them to use loads
1415 // of the per-field globals instead.
1416 for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;) {
1417 Instruction *User = cast<Instruction>(*UI++);
1418
1419 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1420 RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite);
1421 continue;
1422 }
1423
1424 // Must be a store of null.
1425 StoreInst *SI = cast<StoreInst>(User);
1426 assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
1427 "Unexpected heap-sra user!");
1428
1429 // Insert a store of null into each global.
1430 for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
1431 PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType());
1432 Constant *Null = Constant::getNullValue(PT->getElementType());
1433 new StoreInst(Null, FieldGlobals[i], SI);
1434 }
1435 // Erase the original store.
1436 SI->eraseFromParent();
1437 }
1438
1439 // While we have PHIs that are interesting to rewrite, do it.
1440 while (!PHIsToRewrite.empty()) {
1441 PHINode *PN = PHIsToRewrite.back().first;
1442 unsigned FieldNo = PHIsToRewrite.back().second;
1443 PHIsToRewrite.pop_back();
1444 PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]);
1445 assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi");
1446
1447 // Add all the incoming values. This can materialize more phis.
1448 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1449 Value *InVal = PN->getIncomingValue(i);
1450 InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues,
1451 PHIsToRewrite);
1452 FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
1453 }
1454 }
1455
1456 // Drop all inter-phi links and any loads that made it this far.
1457 for (DenseMap<Value*, std::vector<Value*> >::iterator
1458 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1459 I != E; ++I) {
1460 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1461 PN->dropAllReferences();
1462 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1463 LI->dropAllReferences();
1464 }
1465
1466 // Delete all the phis and loads now that inter-references are dead.
1467 for (DenseMap<Value*, std::vector<Value*> >::iterator
1468 I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
1469 I != E; ++I) {
1470 if (PHINode *PN = dyn_cast<PHINode>(I->first))
1471 PN->eraseFromParent();
1472 else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
1473 LI->eraseFromParent();
1474 }
1475
1476 // The old global is now dead, remove it.
1477 GV->eraseFromParent();
1478
1479 ++NumHeapSRA;
1480 return cast<GlobalVariable>(FieldGlobals[0]);
1481 }
1482
1483 /// TryToOptimizeStoreOfMallocToGlobal - This function is called when we see a
1484 /// pointer global variable with a single value stored it that is a malloc or
1485 /// cast of malloc.
TryToOptimizeStoreOfMallocToGlobal(GlobalVariable * GV,CallInst * CI,Type * AllocTy,Module::global_iterator & GVI,TargetData * TD)1486 static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
1487 CallInst *CI,
1488 Type *AllocTy,
1489 Module::global_iterator &GVI,
1490 TargetData *TD) {
1491 if (!TD)
1492 return false;
1493
1494 // If this is a malloc of an abstract type, don't touch it.
1495 if (!AllocTy->isSized())
1496 return false;
1497
1498 // We can't optimize this global unless all uses of it are *known* to be
1499 // of the malloc value, not of the null initializer value (consider a use
1500 // that compares the global's value against zero to see if the malloc has
1501 // been reached). To do this, we check to see if all uses of the global
1502 // would trap if the global were null: this proves that they must all
1503 // happen after the malloc.
1504 if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
1505 return false;
1506
1507 // We can't optimize this if the malloc itself is used in a complex way,
1508 // for example, being stored into multiple globals. This allows the
1509 // malloc to be stored into the specified global, loaded setcc'd, and
1510 // GEP'd. These are all things we could transform to using the global
1511 // for.
1512 SmallPtrSet<const PHINode*, 8> PHIs;
1513 if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV, PHIs))
1514 return false;
1515
1516 // If we have a global that is only initialized with a fixed size malloc,
1517 // transform the program to use global memory instead of malloc'd memory.
1518 // This eliminates dynamic allocation, avoids an indirection accessing the
1519 // data, and exposes the resultant global to further GlobalOpt.
1520 // We cannot optimize the malloc if we cannot determine malloc array size.
1521 Value *NElems = getMallocArraySize(CI, TD, true);
1522 if (!NElems)
1523 return false;
1524
1525 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
1526 // Restrict this transformation to only working on small allocations
1527 // (2048 bytes currently), as we don't want to introduce a 16M global or
1528 // something.
1529 if (NElements->getZExtValue() * TD->getTypeAllocSize(AllocTy) < 2048) {
1530 GVI = OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, TD);
1531 return true;
1532 }
1533
1534 // If the allocation is an array of structures, consider transforming this
1535 // into multiple malloc'd arrays, one for each field. This is basically
1536 // SRoA for malloc'd memory.
1537
1538 // If this is an allocation of a fixed size array of structs, analyze as a
1539 // variable size array. malloc [100 x struct],1 -> malloc struct, 100
1540 if (NElems == ConstantInt::get(CI->getArgOperand(0)->getType(), 1))
1541 if (ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
1542 AllocTy = AT->getElementType();
1543
1544 StructType *AllocSTy = dyn_cast<StructType>(AllocTy);
1545 if (!AllocSTy)
1546 return false;
1547
1548 // This the structure has an unreasonable number of fields, leave it
1549 // alone.
1550 if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
1551 AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, CI)) {
1552
1553 // If this is a fixed size array, transform the Malloc to be an alloc of
1554 // structs. malloc [100 x struct],1 -> malloc struct, 100
1555 if (ArrayType *AT = dyn_cast<ArrayType>(getMallocAllocatedType(CI))) {
1556 Type *IntPtrTy = TD->getIntPtrType(CI->getContext());
1557 unsigned TypeSize = TD->getStructLayout(AllocSTy)->getSizeInBytes();
1558 Value *AllocSize = ConstantInt::get(IntPtrTy, TypeSize);
1559 Value *NumElements = ConstantInt::get(IntPtrTy, AT->getNumElements());
1560 Instruction *Malloc = CallInst::CreateMalloc(CI, IntPtrTy, AllocSTy,
1561 AllocSize, NumElements,
1562 0, CI->getName());
1563 Instruction *Cast = new BitCastInst(Malloc, CI->getType(), "tmp", CI);
1564 CI->replaceAllUsesWith(Cast);
1565 CI->eraseFromParent();
1566 CI = dyn_cast<BitCastInst>(Malloc) ?
1567 extractMallocCallFromBitCast(Malloc) : cast<CallInst>(Malloc);
1568 }
1569
1570 GVI = PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, TD, true),TD);
1571 return true;
1572 }
1573
1574 return false;
1575 }
1576
1577 // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
1578 // that only one value (besides its initializer) is ever stored to the global.
OptimizeOnceStoredGlobal(GlobalVariable * GV,Value * StoredOnceVal,Module::global_iterator & GVI,TargetData * TD)1579 static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1580 Module::global_iterator &GVI,
1581 TargetData *TD) {
1582 // Ignore no-op GEPs and bitcasts.
1583 StoredOnceVal = StoredOnceVal->stripPointerCasts();
1584
1585 // If we are dealing with a pointer global that is initialized to null and
1586 // only has one (non-null) value stored into it, then we can optimize any
1587 // users of the loaded value (often calls and loads) that would trap if the
1588 // value was null.
1589 if (GV->getInitializer()->getType()->isPointerTy() &&
1590 GV->getInitializer()->isNullValue()) {
1591 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1592 if (GV->getInitializer()->getType() != SOVC->getType())
1593 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1594
1595 // Optimize away any trapping uses of the loaded value.
1596 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC))
1597 return true;
1598 } else if (CallInst *CI = extractMallocCall(StoredOnceVal)) {
1599 Type* MallocType = getMallocAllocatedType(CI);
1600 if (MallocType && TryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType,
1601 GVI, TD))
1602 return true;
1603 }
1604 }
1605
1606 return false;
1607 }
1608
1609 /// TryToShrinkGlobalToBoolean - At this point, we have learned that the only
1610 /// two values ever stored into GV are its initializer and OtherVal. See if we
1611 /// can shrink the global into a boolean and select between the two values
1612 /// whenever it is used. This exposes the values to other scalar optimizations.
TryToShrinkGlobalToBoolean(GlobalVariable * GV,Constant * OtherVal)1613 static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1614 Type *GVElType = GV->getType()->getElementType();
1615
1616 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1617 // an FP value, pointer or vector, don't do this optimization because a select
1618 // between them is very expensive and unlikely to lead to later
1619 // simplification. In these cases, we typically end up with "cond ? v1 : v2"
1620 // where v1 and v2 both require constant pool loads, a big loss.
1621 if (GVElType == Type::getInt1Ty(GV->getContext()) ||
1622 GVElType->isFloatingPointTy() ||
1623 GVElType->isPointerTy() || GVElType->isVectorTy())
1624 return false;
1625
1626 // Walk the use list of the global seeing if all the uses are load or store.
1627 // If there is anything else, bail out.
1628 for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I){
1629 User *U = *I;
1630 if (!isa<LoadInst>(U) && !isa<StoreInst>(U))
1631 return false;
1632 }
1633
1634 DEBUG(dbgs() << " *** SHRINKING TO BOOL: " << *GV);
1635
1636 // Create the new global, initializing it to false.
1637 GlobalVariable *NewGV = new GlobalVariable(Type::getInt1Ty(GV->getContext()),
1638 false,
1639 GlobalValue::InternalLinkage,
1640 ConstantInt::getFalse(GV->getContext()),
1641 GV->getName()+".b",
1642 GV->isThreadLocal());
1643 GV->getParent()->getGlobalList().insert(GV, NewGV);
1644
1645 Constant *InitVal = GV->getInitializer();
1646 assert(InitVal->getType() != Type::getInt1Ty(GV->getContext()) &&
1647 "No reason to shrink to bool!");
1648
1649 // If initialized to zero and storing one into the global, we can use a cast
1650 // instead of a select to synthesize the desired value.
1651 bool IsOneZero = false;
1652 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal))
1653 IsOneZero = InitVal->isNullValue() && CI->isOne();
1654
1655 while (!GV->use_empty()) {
1656 Instruction *UI = cast<Instruction>(GV->use_back());
1657 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1658 // Change the store into a boolean store.
1659 bool StoringOther = SI->getOperand(0) == OtherVal;
1660 // Only do this if we weren't storing a loaded value.
1661 Value *StoreVal;
1662 if (StoringOther || SI->getOperand(0) == InitVal)
1663 StoreVal = ConstantInt::get(Type::getInt1Ty(GV->getContext()),
1664 StoringOther);
1665 else {
1666 // Otherwise, we are storing a previously loaded copy. To do this,
1667 // change the copy from copying the original value to just copying the
1668 // bool.
1669 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1670
1671 // If we've already replaced the input, StoredVal will be a cast or
1672 // select instruction. If not, it will be a load of the original
1673 // global.
1674 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1675 assert(LI->getOperand(0) == GV && "Not a copy!");
1676 // Insert a new load, to preserve the saved value.
1677 StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI);
1678 } else {
1679 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&
1680 "This is not a form that we understand!");
1681 StoreVal = StoredVal->getOperand(0);
1682 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!");
1683 }
1684 }
1685 new StoreInst(StoreVal, NewGV, SI);
1686 } else {
1687 // Change the load into a load of bool then a select.
1688 LoadInst *LI = cast<LoadInst>(UI);
1689 LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI);
1690 Value *NSI;
1691 if (IsOneZero)
1692 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1693 else
1694 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1695 NSI->takeName(LI);
1696 LI->replaceAllUsesWith(NSI);
1697 }
1698 UI->eraseFromParent();
1699 }
1700
1701 GV->eraseFromParent();
1702 return true;
1703 }
1704
1705
1706 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1707 /// it if possible. If we make a change, return true.
ProcessGlobal(GlobalVariable * GV,Module::global_iterator & GVI)1708 bool GlobalOpt::ProcessGlobal(GlobalVariable *GV,
1709 Module::global_iterator &GVI) {
1710 if (!GV->hasLocalLinkage())
1711 return false;
1712
1713 // Do more involved optimizations if the global is internal.
1714 GV->removeDeadConstantUsers();
1715
1716 if (GV->use_empty()) {
1717 DEBUG(dbgs() << "GLOBAL DEAD: " << *GV);
1718 GV->eraseFromParent();
1719 ++NumDeleted;
1720 return true;
1721 }
1722
1723 SmallPtrSet<const PHINode*, 16> PHIUsers;
1724 GlobalStatus GS;
1725
1726 if (AnalyzeGlobal(GV, GS, PHIUsers))
1727 return false;
1728
1729 if (!GS.isCompared && !GV->hasUnnamedAddr()) {
1730 GV->setUnnamedAddr(true);
1731 NumUnnamed++;
1732 }
1733
1734 if (GV->isConstant() || !GV->hasInitializer())
1735 return false;
1736
1737 return ProcessInternalGlobal(GV, GVI, PHIUsers, GS);
1738 }
1739
1740 /// ProcessInternalGlobal - Analyze the specified global variable and optimize
1741 /// 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)1742 bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
1743 Module::global_iterator &GVI,
1744 const SmallPtrSet<const PHINode*, 16> &PHIUsers,
1745 const GlobalStatus &GS) {
1746 // If this is a first class global and has only one accessing function
1747 // and this function is main (which we know is not recursive we can make
1748 // this global a local variable) we replace the global with a local alloca
1749 // in this function.
1750 //
1751 // NOTE: It doesn't make sense to promote non single-value types since we
1752 // are just replacing static memory to stack memory.
1753 //
1754 // If the global is in different address space, don't bring it to stack.
1755 if (!GS.HasMultipleAccessingFunctions &&
1756 GS.AccessingFunction && !GS.HasNonInstructionUser &&
1757 GV->getType()->getElementType()->isSingleValueType() &&
1758 GS.AccessingFunction->getName() == "main" &&
1759 GS.AccessingFunction->hasExternalLinkage() &&
1760 GV->getType()->getAddressSpace() == 0) {
1761 DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV);
1762 Instruction& FirstI = const_cast<Instruction&>(*GS.AccessingFunction
1763 ->getEntryBlock().begin());
1764 Type* ElemTy = GV->getType()->getElementType();
1765 // FIXME: Pass Global's alignment when globals have alignment
1766 AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), &FirstI);
1767 if (!isa<UndefValue>(GV->getInitializer()))
1768 new StoreInst(GV->getInitializer(), Alloca, &FirstI);
1769
1770 GV->replaceAllUsesWith(Alloca);
1771 GV->eraseFromParent();
1772 ++NumLocalized;
1773 return true;
1774 }
1775
1776 // If the global is never loaded (but may be stored to), it is dead.
1777 // Delete it now.
1778 if (!GS.isLoaded) {
1779 DEBUG(dbgs() << "GLOBAL NEVER LOADED: " << *GV);
1780
1781 // Delete any stores we can find to the global. We may not be able to
1782 // make it completely dead though.
1783 bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
1784
1785 // If the global is dead now, delete it.
1786 if (GV->use_empty()) {
1787 GV->eraseFromParent();
1788 ++NumDeleted;
1789 Changed = true;
1790 }
1791 return Changed;
1792
1793 } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
1794 DEBUG(dbgs() << "MARKING CONSTANT: " << *GV);
1795 GV->setConstant(true);
1796
1797 // Clean up any obviously simplifiable users now.
1798 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1799
1800 // If the global is dead now, just nuke it.
1801 if (GV->use_empty()) {
1802 DEBUG(dbgs() << " *** Marking constant allowed us to simplify "
1803 << "all users and delete global!\n");
1804 GV->eraseFromParent();
1805 ++NumDeleted;
1806 }
1807
1808 ++NumMarked;
1809 return true;
1810 } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
1811 if (TargetData *TD = getAnalysisIfAvailable<TargetData>())
1812 if (GlobalVariable *FirstNewGV = SRAGlobal(GV, *TD)) {
1813 GVI = FirstNewGV; // Don't skip the newly produced globals!
1814 return true;
1815 }
1816 } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
1817 // If the initial value for the global was an undef value, and if only
1818 // one other value was stored into it, we can just change the
1819 // initializer to be the stored value, then delete all stores to the
1820 // global. This allows us to mark it constant.
1821 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1822 if (isa<UndefValue>(GV->getInitializer())) {
1823 // Change the initial value here.
1824 GV->setInitializer(SOVConstant);
1825
1826 // Clean up any obviously simplifiable users now.
1827 CleanupConstantGlobalUsers(GV, GV->getInitializer());
1828
1829 if (GV->use_empty()) {
1830 DEBUG(dbgs() << " *** Substituting initializer allowed us to "
1831 << "simplify all users and delete global!\n");
1832 GV->eraseFromParent();
1833 ++NumDeleted;
1834 } else {
1835 GVI = GV;
1836 }
1837 ++NumSubstitute;
1838 return true;
1839 }
1840
1841 // Try to optimize globals based on the knowledge that only one value
1842 // (besides its initializer) is ever stored to the global.
1843 if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
1844 getAnalysisIfAvailable<TargetData>()))
1845 return true;
1846
1847 // Otherwise, if the global was not a boolean, we can shrink it to be a
1848 // boolean.
1849 if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
1850 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1851 ++NumShrunkToBool;
1852 return true;
1853 }
1854 }
1855
1856 return false;
1857 }
1858
1859 /// ChangeCalleesToFastCall - Walk all of the direct calls of the specified
1860 /// function, changing them to FastCC.
ChangeCalleesToFastCall(Function * F)1861 static void ChangeCalleesToFastCall(Function *F) {
1862 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1863 CallSite User(cast<Instruction>(*UI));
1864 User.setCallingConv(CallingConv::Fast);
1865 }
1866 }
1867
StripNest(const AttrListPtr & Attrs)1868 static AttrListPtr StripNest(const AttrListPtr &Attrs) {
1869 for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) {
1870 if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0)
1871 continue;
1872
1873 // There can be only one.
1874 return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest);
1875 }
1876
1877 return Attrs;
1878 }
1879
RemoveNestAttribute(Function * F)1880 static void RemoveNestAttribute(Function *F) {
1881 F->setAttributes(StripNest(F->getAttributes()));
1882 for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){
1883 CallSite User(cast<Instruction>(*UI));
1884 User.setAttributes(StripNest(User.getAttributes()));
1885 }
1886 }
1887
OptimizeFunctions(Module & M)1888 bool GlobalOpt::OptimizeFunctions(Module &M) {
1889 bool Changed = false;
1890 // Optimize functions.
1891 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
1892 Function *F = FI++;
1893 // Functions without names cannot be referenced outside this module.
1894 if (!F->hasName() && !F->isDeclaration())
1895 F->setLinkage(GlobalValue::InternalLinkage);
1896 F->removeDeadConstantUsers();
1897 if (F->use_empty() && (F->hasLocalLinkage() || F->hasLinkOnceLinkage())) {
1898 F->eraseFromParent();
1899 Changed = true;
1900 ++NumFnDeleted;
1901 } else if (F->hasLocalLinkage()) {
1902 if (F->getCallingConv() == CallingConv::C && !F->isVarArg() &&
1903 !F->hasAddressTaken()) {
1904 // If this function has C calling conventions, is not a varargs
1905 // function, and is only called directly, promote it to use the Fast
1906 // calling convention.
1907 F->setCallingConv(CallingConv::Fast);
1908 ChangeCalleesToFastCall(F);
1909 ++NumFastCallFns;
1910 Changed = true;
1911 }
1912
1913 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
1914 !F->hasAddressTaken()) {
1915 // The function is not used by a trampoline intrinsic, so it is safe
1916 // to remove the 'nest' attribute.
1917 RemoveNestAttribute(F);
1918 ++NumNestRemoved;
1919 Changed = true;
1920 }
1921 }
1922 }
1923 return Changed;
1924 }
1925
OptimizeGlobalVars(Module & M)1926 bool GlobalOpt::OptimizeGlobalVars(Module &M) {
1927 bool Changed = false;
1928 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
1929 GVI != E; ) {
1930 GlobalVariable *GV = GVI++;
1931 // Global variables without names cannot be referenced outside this module.
1932 if (!GV->hasName() && !GV->isDeclaration())
1933 GV->setLinkage(GlobalValue::InternalLinkage);
1934 // Simplify the initializer.
1935 if (GV->hasInitializer())
1936 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GV->getInitializer())) {
1937 TargetData *TD = getAnalysisIfAvailable<TargetData>();
1938 Constant *New = ConstantFoldConstantExpression(CE, TD);
1939 if (New && New != CE)
1940 GV->setInitializer(New);
1941 }
1942
1943 Changed |= ProcessGlobal(GV, GVI);
1944 }
1945 return Changed;
1946 }
1947
1948 /// FindGlobalCtors - Find the llvm.global_ctors list, verifying that all
1949 /// initializers have an init priority of 65535.
FindGlobalCtors(Module & M)1950 GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
1951 GlobalVariable *GV = M.getGlobalVariable("llvm.global_ctors");
1952 if (GV == 0) return 0;
1953
1954 // Verify that the initializer is simple enough for us to handle. We are
1955 // only allowed to optimize the initializer if it is unique.
1956 if (!GV->hasUniqueInitializer()) return 0;
1957
1958 if (isa<ConstantAggregateZero>(GV->getInitializer()))
1959 return GV;
1960 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1961
1962 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
1963 if (isa<ConstantAggregateZero>(*i))
1964 continue;
1965 ConstantStruct *CS = cast<ConstantStruct>(*i);
1966 if (isa<ConstantPointerNull>(CS->getOperand(1)))
1967 continue;
1968
1969 // Must have a function or null ptr.
1970 if (!isa<Function>(CS->getOperand(1)))
1971 return 0;
1972
1973 // Init priority must be standard.
1974 ConstantInt *CI = cast<ConstantInt>(CS->getOperand(0));
1975 if (CI->getZExtValue() != 65535)
1976 return 0;
1977 }
1978
1979 return GV;
1980 }
1981
1982 /// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
1983 /// return a list of the functions and null terminator as a vector.
ParseGlobalCtors(GlobalVariable * GV)1984 static std::vector<Function*> ParseGlobalCtors(GlobalVariable *GV) {
1985 if (GV->getInitializer()->isNullValue())
1986 return std::vector<Function*>();
1987 ConstantArray *CA = cast<ConstantArray>(GV->getInitializer());
1988 std::vector<Function*> Result;
1989 Result.reserve(CA->getNumOperands());
1990 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
1991 ConstantStruct *CS = cast<ConstantStruct>(*i);
1992 Result.push_back(dyn_cast<Function>(CS->getOperand(1)));
1993 }
1994 return Result;
1995 }
1996
1997 /// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
1998 /// specified array, returning the new global to use.
InstallGlobalCtors(GlobalVariable * GCL,const std::vector<Function * > & Ctors)1999 static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
2000 const std::vector<Function*> &Ctors) {
2001 // If we made a change, reassemble the initializer list.
2002 Constant *CSVals[2];
2003 CSVals[0] = ConstantInt::get(Type::getInt32Ty(GCL->getContext()), 65535);
2004 CSVals[1] = 0;
2005
2006 StructType *StructTy =
2007 cast <StructType>(
2008 cast<ArrayType>(GCL->getType()->getElementType())->getElementType());
2009
2010 // Create the new init list.
2011 std::vector<Constant*> CAList;
2012 for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
2013 if (Ctors[i]) {
2014 CSVals[1] = Ctors[i];
2015 } else {
2016 Type *FTy = FunctionType::get(Type::getVoidTy(GCL->getContext()),
2017 false);
2018 PointerType *PFTy = PointerType::getUnqual(FTy);
2019 CSVals[1] = Constant::getNullValue(PFTy);
2020 CSVals[0] = ConstantInt::get(Type::getInt32Ty(GCL->getContext()),
2021 0x7fffffff);
2022 }
2023 CAList.push_back(ConstantStruct::get(StructTy, CSVals));
2024 }
2025
2026 // Create the array initializer.
2027 Constant *CA = ConstantArray::get(ArrayType::get(StructTy,
2028 CAList.size()), CAList);
2029
2030 // If we didn't change the number of elements, don't create a new GV.
2031 if (CA->getType() == GCL->getInitializer()->getType()) {
2032 GCL->setInitializer(CA);
2033 return GCL;
2034 }
2035
2036 // Create the new global and insert it next to the existing list.
2037 GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
2038 GCL->getLinkage(), CA, "",
2039 GCL->isThreadLocal());
2040 GCL->getParent()->getGlobalList().insert(GCL, NGV);
2041 NGV->takeName(GCL);
2042
2043 // Nuke the old list, replacing any uses with the new one.
2044 if (!GCL->use_empty()) {
2045 Constant *V = NGV;
2046 if (V->getType() != GCL->getType())
2047 V = ConstantExpr::getBitCast(V, GCL->getType());
2048 GCL->replaceAllUsesWith(V);
2049 }
2050 GCL->eraseFromParent();
2051
2052 if (Ctors.size())
2053 return NGV;
2054 else
2055 return 0;
2056 }
2057
2058
getVal(DenseMap<Value *,Constant * > & ComputedValues,Value * V)2059 static Constant *getVal(DenseMap<Value*, Constant*> &ComputedValues, Value *V) {
2060 if (Constant *CV = dyn_cast<Constant>(V)) return CV;
2061 Constant *R = ComputedValues[V];
2062 assert(R && "Reference to an uncomputed value!");
2063 return R;
2064 }
2065
2066 static inline bool
2067 isSimpleEnoughValueToCommit(Constant *C,
2068 SmallPtrSet<Constant*, 8> &SimpleConstants);
2069
2070
2071 /// isSimpleEnoughValueToCommit - Return true if the specified constant can be
2072 /// handled by the code generator. We don't want to generate something like:
2073 /// void *X = &X/42;
2074 /// because the code generator doesn't have a relocation that can handle that.
2075 ///
2076 /// This function should be called if C was not found (but just got inserted)
2077 /// in SimpleConstants to avoid having to rescan the same constants all the
2078 /// time.
isSimpleEnoughValueToCommitHelper(Constant * C,SmallPtrSet<Constant *,8> & SimpleConstants)2079 static bool isSimpleEnoughValueToCommitHelper(Constant *C,
2080 SmallPtrSet<Constant*, 8> &SimpleConstants) {
2081 // Simple integer, undef, constant aggregate zero, global addresses, etc are
2082 // all supported.
2083 if (C->getNumOperands() == 0 || isa<BlockAddress>(C) ||
2084 isa<GlobalValue>(C))
2085 return true;
2086
2087 // Aggregate values are safe if all their elements are.
2088 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C) ||
2089 isa<ConstantVector>(C)) {
2090 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) {
2091 Constant *Op = cast<Constant>(C->getOperand(i));
2092 if (!isSimpleEnoughValueToCommit(Op, SimpleConstants))
2093 return false;
2094 }
2095 return true;
2096 }
2097
2098 // We don't know exactly what relocations are allowed in constant expressions,
2099 // so we allow &global+constantoffset, which is safe and uniformly supported
2100 // across targets.
2101 ConstantExpr *CE = cast<ConstantExpr>(C);
2102 switch (CE->getOpcode()) {
2103 case Instruction::BitCast:
2104 case Instruction::IntToPtr:
2105 case Instruction::PtrToInt:
2106 // These casts are always fine if the casted value is.
2107 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants);
2108
2109 // GEP is fine if it is simple + constant offset.
2110 case Instruction::GetElementPtr:
2111 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
2112 if (!isa<ConstantInt>(CE->getOperand(i)))
2113 return false;
2114 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants);
2115
2116 case Instruction::Add:
2117 // We allow simple+cst.
2118 if (!isa<ConstantInt>(CE->getOperand(1)))
2119 return false;
2120 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants);
2121 }
2122 return false;
2123 }
2124
2125 static inline bool
isSimpleEnoughValueToCommit(Constant * C,SmallPtrSet<Constant *,8> & SimpleConstants)2126 isSimpleEnoughValueToCommit(Constant *C,
2127 SmallPtrSet<Constant*, 8> &SimpleConstants) {
2128 // If we already checked this constant, we win.
2129 if (!SimpleConstants.insert(C)) return true;
2130 // Check the constant.
2131 return isSimpleEnoughValueToCommitHelper(C, SimpleConstants);
2132 }
2133
2134
2135 /// isSimpleEnoughPointerToCommit - Return true if this constant is simple
2136 /// enough for us to understand. In particular, if it is a cast to anything
2137 /// other than from one pointer type to another pointer type, we punt.
2138 /// We basically just support direct accesses to globals and GEP's of
2139 /// globals. This should be kept up to date with CommitValueTo.
isSimpleEnoughPointerToCommit(Constant * C)2140 static bool isSimpleEnoughPointerToCommit(Constant *C) {
2141 // Conservatively, avoid aggregate types. This is because we don't
2142 // want to worry about them partially overlapping other stores.
2143 if (!cast<PointerType>(C->getType())->getElementType()->isSingleValueType())
2144 return false;
2145
2146 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
2147 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2148 // external globals.
2149 return GV->hasUniqueInitializer();
2150
2151 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2152 // Handle a constantexpr gep.
2153 if (CE->getOpcode() == Instruction::GetElementPtr &&
2154 isa<GlobalVariable>(CE->getOperand(0)) &&
2155 cast<GEPOperator>(CE)->isInBounds()) {
2156 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2157 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2158 // external globals.
2159 if (!GV->hasUniqueInitializer())
2160 return false;
2161
2162 // The first index must be zero.
2163 ConstantInt *CI = dyn_cast<ConstantInt>(*llvm::next(CE->op_begin()));
2164 if (!CI || !CI->isZero()) return false;
2165
2166 // The remaining indices must be compile-time known integers within the
2167 // notional bounds of the corresponding static array types.
2168 if (!CE->isGEPWithNoNotionalOverIndexing())
2169 return false;
2170
2171 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2172
2173 // A constantexpr bitcast from a pointer to another pointer is a no-op,
2174 // and we know how to evaluate it by moving the bitcast from the pointer
2175 // operand to the value operand.
2176 } else if (CE->getOpcode() == Instruction::BitCast &&
2177 isa<GlobalVariable>(CE->getOperand(0))) {
2178 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
2179 // external globals.
2180 return cast<GlobalVariable>(CE->getOperand(0))->hasUniqueInitializer();
2181 }
2182 }
2183
2184 return false;
2185 }
2186
2187 /// EvaluateStoreInto - Evaluate a piece of a constantexpr store into a global
2188 /// initializer. This returns 'Init' modified to reflect 'Val' stored into it.
2189 /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
EvaluateStoreInto(Constant * Init,Constant * Val,ConstantExpr * Addr,unsigned OpNo)2190 static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
2191 ConstantExpr *Addr, unsigned OpNo) {
2192 // Base case of the recursion.
2193 if (OpNo == Addr->getNumOperands()) {
2194 assert(Val->getType() == Init->getType() && "Type mismatch!");
2195 return Val;
2196 }
2197
2198 std::vector<Constant*> Elts;
2199 if (StructType *STy = dyn_cast<StructType>(Init->getType())) {
2200
2201 // Break up the constant into its elements.
2202 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) {
2203 for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i)
2204 Elts.push_back(cast<Constant>(*i));
2205 } else if (isa<ConstantAggregateZero>(Init)) {
2206 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2207 Elts.push_back(Constant::getNullValue(STy->getElementType(i)));
2208 } else if (isa<UndefValue>(Init)) {
2209 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2210 Elts.push_back(UndefValue::get(STy->getElementType(i)));
2211 } else {
2212 llvm_unreachable("This code is out of sync with "
2213 " ConstantFoldLoadThroughGEPConstantExpr");
2214 }
2215
2216 // Replace the element that we are supposed to.
2217 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
2218 unsigned Idx = CU->getZExtValue();
2219 assert(Idx < STy->getNumElements() && "Struct index out of range!");
2220 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
2221
2222 // Return the modified struct.
2223 return ConstantStruct::get(STy, Elts);
2224 }
2225
2226 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
2227 SequentialType *InitTy = cast<SequentialType>(Init->getType());
2228
2229 uint64_t NumElts;
2230 if (ArrayType *ATy = dyn_cast<ArrayType>(InitTy))
2231 NumElts = ATy->getNumElements();
2232 else
2233 NumElts = cast<VectorType>(InitTy)->getNumElements();
2234
2235 // Break up the array into elements.
2236 if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
2237 for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
2238 Elts.push_back(cast<Constant>(*i));
2239 } else if (ConstantVector *CV = dyn_cast<ConstantVector>(Init)) {
2240 for (User::op_iterator i = CV->op_begin(), e = CV->op_end(); i != e; ++i)
2241 Elts.push_back(cast<Constant>(*i));
2242 } else if (isa<ConstantAggregateZero>(Init)) {
2243 Elts.assign(NumElts, Constant::getNullValue(InitTy->getElementType()));
2244 } else {
2245 assert(isa<UndefValue>(Init) && "This code is out of sync with "
2246 " ConstantFoldLoadThroughGEPConstantExpr");
2247 Elts.assign(NumElts, UndefValue::get(InitTy->getElementType()));
2248 }
2249
2250 assert(CI->getZExtValue() < NumElts);
2251 Elts[CI->getZExtValue()] =
2252 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
2253
2254 if (Init->getType()->isArrayTy())
2255 return ConstantArray::get(cast<ArrayType>(InitTy), Elts);
2256 return ConstantVector::get(Elts);
2257 }
2258
2259 /// CommitValueTo - We have decided that Addr (which satisfies the predicate
2260 /// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
CommitValueTo(Constant * Val,Constant * Addr)2261 static void CommitValueTo(Constant *Val, Constant *Addr) {
2262 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
2263 assert(GV->hasInitializer());
2264 GV->setInitializer(Val);
2265 return;
2266 }
2267
2268 ConstantExpr *CE = cast<ConstantExpr>(Addr);
2269 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2270 GV->setInitializer(EvaluateStoreInto(GV->getInitializer(), Val, CE, 2));
2271 }
2272
2273 /// ComputeLoadResult - Return the value that would be computed by a load from
2274 /// P after the stores reflected by 'memory' have been performed. If we can't
2275 /// decide, return null.
ComputeLoadResult(Constant * P,const DenseMap<Constant *,Constant * > & Memory)2276 static Constant *ComputeLoadResult(Constant *P,
2277 const DenseMap<Constant*, Constant*> &Memory) {
2278 // If this memory location has been recently stored, use the stored value: it
2279 // is the most up-to-date.
2280 DenseMap<Constant*, Constant*>::const_iterator I = Memory.find(P);
2281 if (I != Memory.end()) return I->second;
2282
2283 // Access it.
2284 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
2285 if (GV->hasDefinitiveInitializer())
2286 return GV->getInitializer();
2287 return 0;
2288 }
2289
2290 // Handle a constantexpr getelementptr.
2291 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
2292 if (CE->getOpcode() == Instruction::GetElementPtr &&
2293 isa<GlobalVariable>(CE->getOperand(0))) {
2294 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2295 if (GV->hasDefinitiveInitializer())
2296 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
2297 }
2298
2299 return 0; // don't know how to evaluate.
2300 }
2301
2302 /// EvaluateFunction - Evaluate a call to function F, returning true if
2303 /// successful, false if we can't evaluate it. ActualArgs contains the formal
2304 /// 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)2305 static bool EvaluateFunction(Function *F, Constant *&RetVal,
2306 const SmallVectorImpl<Constant*> &ActualArgs,
2307 std::vector<Function*> &CallStack,
2308 DenseMap<Constant*, Constant*> &MutatedMemory,
2309 std::vector<GlobalVariable*> &AllocaTmps,
2310 SmallPtrSet<Constant*, 8> &SimpleConstants,
2311 const TargetData *TD) {
2312 // Check to see if this function is already executing (recursion). If so,
2313 // bail out. TODO: we might want to accept limited recursion.
2314 if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
2315 return false;
2316
2317 CallStack.push_back(F);
2318
2319 /// Values - As we compute SSA register values, we store their contents here.
2320 DenseMap<Value*, Constant*> Values;
2321
2322 // Initialize arguments to the incoming values specified.
2323 unsigned ArgNo = 0;
2324 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
2325 ++AI, ++ArgNo)
2326 Values[AI] = ActualArgs[ArgNo];
2327
2328 /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such,
2329 /// we can only evaluate any one basic block at most once. This set keeps
2330 /// track of what we have executed so we can detect recursive cases etc.
2331 SmallPtrSet<BasicBlock*, 32> ExecutedBlocks;
2332
2333 // CurInst - The current instruction we're evaluating.
2334 BasicBlock::iterator CurInst = F->begin()->begin();
2335
2336 // This is the main evaluation loop.
2337 while (1) {
2338 Constant *InstResult = 0;
2339
2340 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
2341 if (SI->isVolatile()) return false; // no volatile accesses.
2342 Constant *Ptr = getVal(Values, SI->getOperand(1));
2343 if (!isSimpleEnoughPointerToCommit(Ptr))
2344 // If this is too complex for us to commit, reject it.
2345 return false;
2346
2347 Constant *Val = getVal(Values, SI->getOperand(0));
2348
2349 // If this might be too difficult for the backend to handle (e.g. the addr
2350 // of one global variable divided by another) then we can't commit it.
2351 if (!isSimpleEnoughValueToCommit(Val, SimpleConstants))
2352 return false;
2353
2354 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
2355 if (CE->getOpcode() == Instruction::BitCast) {
2356 // If we're evaluating a store through a bitcast, then we need
2357 // to pull the bitcast off the pointer type and push it onto the
2358 // stored value.
2359 Ptr = CE->getOperand(0);
2360
2361 Type *NewTy=cast<PointerType>(Ptr->getType())->getElementType();
2362
2363 // In order to push the bitcast onto the stored value, a bitcast
2364 // from NewTy to Val's type must be legal. If it's not, we can try
2365 // introspecting NewTy to find a legal conversion.
2366 while (!Val->getType()->canLosslesslyBitCastTo(NewTy)) {
2367 // If NewTy is a struct, we can convert the pointer to the struct
2368 // into a pointer to its first member.
2369 // FIXME: This could be extended to support arrays as well.
2370 if (StructType *STy = dyn_cast<StructType>(NewTy)) {
2371 NewTy = STy->getTypeAtIndex(0U);
2372
2373 IntegerType *IdxTy =IntegerType::get(NewTy->getContext(), 32);
2374 Constant *IdxZero = ConstantInt::get(IdxTy, 0, false);
2375 Constant * const IdxList[] = {IdxZero, IdxZero};
2376
2377 Ptr = ConstantExpr::getGetElementPtr(Ptr, IdxList, 2);
2378
2379 // If we can't improve the situation by introspecting NewTy,
2380 // we have to give up.
2381 } else {
2382 return 0;
2383 }
2384 }
2385
2386 // If we found compatible types, go ahead and push the bitcast
2387 // onto the stored value.
2388 Val = ConstantExpr::getBitCast(Val, NewTy);
2389 }
2390
2391 MutatedMemory[Ptr] = Val;
2392 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
2393 InstResult = ConstantExpr::get(BO->getOpcode(),
2394 getVal(Values, BO->getOperand(0)),
2395 getVal(Values, BO->getOperand(1)));
2396 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) {
2397 InstResult = ConstantExpr::getCompare(CI->getPredicate(),
2398 getVal(Values, CI->getOperand(0)),
2399 getVal(Values, CI->getOperand(1)));
2400 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) {
2401 InstResult = ConstantExpr::getCast(CI->getOpcode(),
2402 getVal(Values, CI->getOperand(0)),
2403 CI->getType());
2404 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) {
2405 InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)),
2406 getVal(Values, SI->getOperand(1)),
2407 getVal(Values, SI->getOperand(2)));
2408 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) {
2409 Constant *P = getVal(Values, GEP->getOperand(0));
2410 SmallVector<Constant*, 8> GEPOps;
2411 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
2412 i != e; ++i)
2413 GEPOps.push_back(getVal(Values, *i));
2414 InstResult = cast<GEPOperator>(GEP)->isInBounds() ?
2415 ConstantExpr::getInBoundsGetElementPtr(P, &GEPOps[0], GEPOps.size()) :
2416 ConstantExpr::getGetElementPtr(P, &GEPOps[0], GEPOps.size());
2417 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) {
2418 if (LI->isVolatile()) return false; // no volatile accesses.
2419 InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)),
2420 MutatedMemory);
2421 if (InstResult == 0) return false; // Could not evaluate load.
2422 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) {
2423 if (AI->isArrayAllocation()) return false; // Cannot handle array allocs.
2424 Type *Ty = AI->getType()->getElementType();
2425 AllocaTmps.push_back(new GlobalVariable(Ty, false,
2426 GlobalValue::InternalLinkage,
2427 UndefValue::get(Ty),
2428 AI->getName()));
2429 InstResult = AllocaTmps.back();
2430 } else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
2431
2432 // Debug info can safely be ignored here.
2433 if (isa<DbgInfoIntrinsic>(CI)) {
2434 ++CurInst;
2435 continue;
2436 }
2437
2438 // Cannot handle inline asm.
2439 if (isa<InlineAsm>(CI->getCalledValue())) return false;
2440
2441 if (MemSetInst *MSI = dyn_cast<MemSetInst>(CI)) {
2442 if (MSI->isVolatile()) return false;
2443 Constant *Ptr = getVal(Values, MSI->getDest());
2444 Constant *Val = getVal(Values, MSI->getValue());
2445 Constant *DestVal = ComputeLoadResult(getVal(Values, Ptr),
2446 MutatedMemory);
2447 if (Val->isNullValue() && DestVal && DestVal->isNullValue()) {
2448 // This memset is a no-op.
2449 ++CurInst;
2450 continue;
2451 }
2452 return false;
2453 }
2454
2455 // Resolve function pointers.
2456 Function *Callee = dyn_cast<Function>(getVal(Values,
2457 CI->getCalledValue()));
2458 if (!Callee) return false; // Cannot resolve.
2459
2460 SmallVector<Constant*, 8> Formals;
2461 CallSite CS(CI);
2462 for (User::op_iterator i = CS.arg_begin(), e = CS.arg_end();
2463 i != e; ++i)
2464 Formals.push_back(getVal(Values, *i));
2465
2466 if (Callee->isDeclaration()) {
2467 // If this is a function we can constant fold, do it.
2468 if (Constant *C = ConstantFoldCall(Callee, Formals)) {
2469 InstResult = C;
2470 } else {
2471 return false;
2472 }
2473 } else {
2474 if (Callee->getFunctionType()->isVarArg())
2475 return false;
2476
2477 Constant *RetVal;
2478 // Execute the call, if successful, use the return value.
2479 if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
2480 MutatedMemory, AllocaTmps, SimpleConstants, TD))
2481 return false;
2482 InstResult = RetVal;
2483 }
2484 } else if (isa<TerminatorInst>(CurInst)) {
2485 BasicBlock *NewBB = 0;
2486 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) {
2487 if (BI->isUnconditional()) {
2488 NewBB = BI->getSuccessor(0);
2489 } else {
2490 ConstantInt *Cond =
2491 dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
2492 if (!Cond) return false; // Cannot determine.
2493
2494 NewBB = BI->getSuccessor(!Cond->getZExtValue());
2495 }
2496 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
2497 ConstantInt *Val =
2498 dyn_cast<ConstantInt>(getVal(Values, SI->getCondition()));
2499 if (!Val) return false; // Cannot determine.
2500 NewBB = SI->getSuccessor(SI->findCaseValue(Val));
2501 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(CurInst)) {
2502 Value *Val = getVal(Values, IBI->getAddress())->stripPointerCasts();
2503 if (BlockAddress *BA = dyn_cast<BlockAddress>(Val))
2504 NewBB = BA->getBasicBlock();
2505 else
2506 return false; // Cannot determine.
2507 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
2508 if (RI->getNumOperands())
2509 RetVal = getVal(Values, RI->getOperand(0));
2510
2511 CallStack.pop_back(); // return from fn.
2512 return true; // We succeeded at evaluating this ctor!
2513 } else {
2514 // invoke, unwind, unreachable.
2515 return false; // Cannot handle this terminator.
2516 }
2517
2518 // Okay, we succeeded in evaluating this control flow. See if we have
2519 // executed the new block before. If so, we have a looping function,
2520 // which we cannot evaluate in reasonable time.
2521 if (!ExecutedBlocks.insert(NewBB))
2522 return false; // looped!
2523
2524 // Okay, we have never been in this block before. Check to see if there
2525 // are any PHI nodes. If so, evaluate them with information about where
2526 // we came from.
2527 BasicBlock *OldBB = CurInst->getParent();
2528 CurInst = NewBB->begin();
2529 PHINode *PN;
2530 for (; (PN = dyn_cast<PHINode>(CurInst)); ++CurInst)
2531 Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB));
2532
2533 // Do NOT increment CurInst. We know that the terminator had no value.
2534 continue;
2535 } else {
2536 // Did not know how to evaluate this!
2537 return false;
2538 }
2539
2540 if (!CurInst->use_empty()) {
2541 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(InstResult))
2542 InstResult = ConstantFoldConstantExpression(CE, TD);
2543
2544 Values[CurInst] = InstResult;
2545 }
2546
2547 // Advance program counter.
2548 ++CurInst;
2549 }
2550 }
2551
2552 /// EvaluateStaticConstructor - Evaluate static constructors in the function, if
2553 /// we can. Return true if we can, false otherwise.
EvaluateStaticConstructor(Function * F,const TargetData * TD)2554 static bool EvaluateStaticConstructor(Function *F, const TargetData *TD) {
2555 /// MutatedMemory - For each store we execute, we update this map. Loads
2556 /// check this to get the most up-to-date value. If evaluation is successful,
2557 /// this state is committed to the process.
2558 DenseMap<Constant*, Constant*> MutatedMemory;
2559
2560 /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable
2561 /// to represent its body. This vector is needed so we can delete the
2562 /// temporary globals when we are done.
2563 std::vector<GlobalVariable*> AllocaTmps;
2564
2565 /// CallStack - This is used to detect recursion. In pathological situations
2566 /// we could hit exponential behavior, but at least there is nothing
2567 /// unbounded.
2568 std::vector<Function*> CallStack;
2569
2570 /// SimpleConstants - These are constants we have checked and know to be
2571 /// simple enough to live in a static initializer of a global.
2572 SmallPtrSet<Constant*, 8> SimpleConstants;
2573
2574 // Call the function.
2575 Constant *RetValDummy;
2576 bool EvalSuccess = EvaluateFunction(F, RetValDummy,
2577 SmallVector<Constant*, 0>(), CallStack,
2578 MutatedMemory, AllocaTmps,
2579 SimpleConstants, TD);
2580
2581 if (EvalSuccess) {
2582 // We succeeded at evaluation: commit the result.
2583 DEBUG(dbgs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
2584 << F->getName() << "' to " << MutatedMemory.size()
2585 << " stores.\n");
2586 for (DenseMap<Constant*, Constant*>::iterator I = MutatedMemory.begin(),
2587 E = MutatedMemory.end(); I != E; ++I)
2588 CommitValueTo(I->second, I->first);
2589 }
2590
2591 // At this point, we are done interpreting. If we created any 'alloca'
2592 // temporaries, release them now.
2593 while (!AllocaTmps.empty()) {
2594 GlobalVariable *Tmp = AllocaTmps.back();
2595 AllocaTmps.pop_back();
2596
2597 // If there are still users of the alloca, the program is doing something
2598 // silly, e.g. storing the address of the alloca somewhere and using it
2599 // later. Since this is undefined, we'll just make it be null.
2600 if (!Tmp->use_empty())
2601 Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
2602 delete Tmp;
2603 }
2604
2605 return EvalSuccess;
2606 }
2607
2608
2609
2610 /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible.
2611 /// Return true if anything changed.
OptimizeGlobalCtorsList(GlobalVariable * & GCL)2612 bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) {
2613 std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
2614 bool MadeChange = false;
2615 if (Ctors.empty()) return false;
2616
2617 const TargetData *TD = getAnalysisIfAvailable<TargetData>();
2618 // Loop over global ctors, optimizing them when we can.
2619 for (unsigned i = 0; i != Ctors.size(); ++i) {
2620 Function *F = Ctors[i];
2621 // Found a null terminator in the middle of the list, prune off the rest of
2622 // the list.
2623 if (F == 0) {
2624 if (i != Ctors.size()-1) {
2625 Ctors.resize(i+1);
2626 MadeChange = true;
2627 }
2628 break;
2629 }
2630
2631 // We cannot simplify external ctor functions.
2632 if (F->empty()) continue;
2633
2634 // If we can evaluate the ctor at compile time, do.
2635 if (EvaluateStaticConstructor(F, TD)) {
2636 Ctors.erase(Ctors.begin()+i);
2637 MadeChange = true;
2638 --i;
2639 ++NumCtorsEvaluated;
2640 continue;
2641 }
2642 }
2643
2644 if (!MadeChange) return false;
2645
2646 GCL = InstallGlobalCtors(GCL, Ctors);
2647 return true;
2648 }
2649
OptimizeGlobalAliases(Module & M)2650 bool GlobalOpt::OptimizeGlobalAliases(Module &M) {
2651 bool Changed = false;
2652
2653 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end();
2654 I != E;) {
2655 Module::alias_iterator J = I++;
2656 // Aliases without names cannot be referenced outside this module.
2657 if (!J->hasName() && !J->isDeclaration())
2658 J->setLinkage(GlobalValue::InternalLinkage);
2659 // If the aliasee may change at link time, nothing can be done - bail out.
2660 if (J->mayBeOverridden())
2661 continue;
2662
2663 Constant *Aliasee = J->getAliasee();
2664 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2665 Target->removeDeadConstantUsers();
2666 bool hasOneUse = Target->hasOneUse() && Aliasee->hasOneUse();
2667
2668 // Make all users of the alias use the aliasee instead.
2669 if (!J->use_empty()) {
2670 J->replaceAllUsesWith(Aliasee);
2671 ++NumAliasesResolved;
2672 Changed = true;
2673 }
2674
2675 // If the alias is externally visible, we may still be able to simplify it.
2676 if (!J->hasLocalLinkage()) {
2677 // If the aliasee has internal linkage, give it the name and linkage
2678 // of the alias, and delete the alias. This turns:
2679 // define internal ... @f(...)
2680 // @a = alias ... @f
2681 // into:
2682 // define ... @a(...)
2683 if (!Target->hasLocalLinkage())
2684 continue;
2685
2686 // Do not perform the transform if multiple aliases potentially target the
2687 // aliasee. This check also ensures that it is safe to replace the section
2688 // and other attributes of the aliasee with those of the alias.
2689 if (!hasOneUse)
2690 continue;
2691
2692 // Give the aliasee the name, linkage and other attributes of the alias.
2693 Target->takeName(J);
2694 Target->setLinkage(J->getLinkage());
2695 Target->GlobalValue::copyAttributesFrom(J);
2696 }
2697
2698 // Delete the alias.
2699 M.getAliasList().erase(J);
2700 ++NumAliasesRemoved;
2701 Changed = true;
2702 }
2703
2704 return Changed;
2705 }
2706
FindCXAAtExit(Module & M)2707 static Function *FindCXAAtExit(Module &M) {
2708 Function *Fn = M.getFunction("__cxa_atexit");
2709
2710 if (!Fn)
2711 return 0;
2712
2713 FunctionType *FTy = Fn->getFunctionType();
2714
2715 // Checking that the function has the right return type, the right number of
2716 // parameters and that they all have pointer types should be enough.
2717 if (!FTy->getReturnType()->isIntegerTy() ||
2718 FTy->getNumParams() != 3 ||
2719 !FTy->getParamType(0)->isPointerTy() ||
2720 !FTy->getParamType(1)->isPointerTy() ||
2721 !FTy->getParamType(2)->isPointerTy())
2722 return 0;
2723
2724 return Fn;
2725 }
2726
2727 /// cxxDtorIsEmpty - Returns whether the given function is an empty C++
2728 /// destructor and can therefore be eliminated.
2729 /// Note that we assume that other optimization passes have already simplified
2730 /// the code so we only look for a function with a single basic block, where
2731 /// the only allowed instructions are 'ret' or 'call' to empty C++ dtor.
cxxDtorIsEmpty(const Function & Fn,SmallPtrSet<const Function *,8> & CalledFunctions)2732 static bool cxxDtorIsEmpty(const Function &Fn,
2733 SmallPtrSet<const Function *, 8> &CalledFunctions) {
2734 // FIXME: We could eliminate C++ destructors if they're readonly/readnone and
2735 // nounwind, but that doesn't seem worth doing.
2736 if (Fn.isDeclaration())
2737 return false;
2738
2739 if (++Fn.begin() != Fn.end())
2740 return false;
2741
2742 const BasicBlock &EntryBlock = Fn.getEntryBlock();
2743 for (BasicBlock::const_iterator I = EntryBlock.begin(), E = EntryBlock.end();
2744 I != E; ++I) {
2745 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
2746 // Ignore debug intrinsics.
2747 if (isa<DbgInfoIntrinsic>(CI))
2748 continue;
2749
2750 const Function *CalledFn = CI->getCalledFunction();
2751
2752 if (!CalledFn)
2753 return false;
2754
2755 SmallPtrSet<const Function *, 8> NewCalledFunctions(CalledFunctions);
2756
2757 // Don't treat recursive functions as empty.
2758 if (!NewCalledFunctions.insert(CalledFn))
2759 return false;
2760
2761 if (!cxxDtorIsEmpty(*CalledFn, NewCalledFunctions))
2762 return false;
2763 } else if (isa<ReturnInst>(*I))
2764 return true;
2765 else
2766 return false;
2767 }
2768
2769 return false;
2770 }
2771
OptimizeEmptyGlobalCXXDtors(Function * CXAAtExitFn)2772 bool GlobalOpt::OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) {
2773 /// Itanium C++ ABI p3.3.5:
2774 ///
2775 /// After constructing a global (or local static) object, that will require
2776 /// destruction on exit, a termination function is registered as follows:
2777 ///
2778 /// extern "C" int __cxa_atexit ( void (*f)(void *), void *p, void *d );
2779 ///
2780 /// This registration, e.g. __cxa_atexit(f,p,d), is intended to cause the
2781 /// call f(p) when DSO d is unloaded, before all such termination calls
2782 /// registered before this one. It returns zero if registration is
2783 /// successful, nonzero on failure.
2784
2785 // This pass will look for calls to __cxa_atexit where the function is trivial
2786 // and remove them.
2787 bool Changed = false;
2788
2789 for (Function::use_iterator I = CXAAtExitFn->use_begin(),
2790 E = CXAAtExitFn->use_end(); I != E;) {
2791 // We're only interested in calls. Theoretically, we could handle invoke
2792 // instructions as well, but neither llvm-gcc nor clang generate invokes
2793 // to __cxa_atexit.
2794 CallInst *CI = dyn_cast<CallInst>(*I++);
2795 if (!CI)
2796 continue;
2797
2798 Function *DtorFn =
2799 dyn_cast<Function>(CI->getArgOperand(0)->stripPointerCasts());
2800 if (!DtorFn)
2801 continue;
2802
2803 SmallPtrSet<const Function *, 8> CalledFunctions;
2804 if (!cxxDtorIsEmpty(*DtorFn, CalledFunctions))
2805 continue;
2806
2807 // Just remove the call.
2808 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
2809 CI->eraseFromParent();
2810
2811 ++NumCXXDtorsRemoved;
2812
2813 Changed |= true;
2814 }
2815
2816 return Changed;
2817 }
2818
runOnModule(Module & M)2819 bool GlobalOpt::runOnModule(Module &M) {
2820 bool Changed = false;
2821
2822 // Try to find the llvm.globalctors list.
2823 GlobalVariable *GlobalCtors = FindGlobalCtors(M);
2824
2825 Function *CXAAtExitFn = FindCXAAtExit(M);
2826
2827 bool LocalChange = true;
2828 while (LocalChange) {
2829 LocalChange = false;
2830
2831 // Delete functions that are trivially dead, ccc -> fastcc
2832 LocalChange |= OptimizeFunctions(M);
2833
2834 // Optimize global_ctors list.
2835 if (GlobalCtors)
2836 LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
2837
2838 // Optimize non-address-taken globals.
2839 LocalChange |= OptimizeGlobalVars(M);
2840
2841 // Resolve aliases, when possible.
2842 LocalChange |= OptimizeGlobalAliases(M);
2843
2844 // Try to remove trivial global destructors.
2845 if (CXAAtExitFn)
2846 LocalChange |= OptimizeEmptyGlobalCXXDtors(CXAAtExitFn);
2847
2848 Changed |= LocalChange;
2849 }
2850
2851 // TODO: Move all global ctors functions to the end of the module for code
2852 // layout.
2853
2854 return Changed;
2855 }
2856