1 //===-- ValueEnumerator.cpp - Number values and types for bitcode writer --===//
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 file implements the ValueEnumerator class.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "ValueEnumerator.h"
15 #include "llvm/ADT/STLExtras.h"
16 #include "llvm/ADT/SmallPtrSet.h"
17 #include "llvm/IR/Constants.h"
18 #include "llvm/IR/DebugInfoMetadata.h"
19 #include "llvm/IR/DerivedTypes.h"
20 #include "llvm/IR/Instructions.h"
21 #include "llvm/IR/Module.h"
22 #include "llvm/IR/UseListOrder.h"
23 #include "llvm/IR/ValueSymbolTable.h"
24 #include "llvm/Support/Debug.h"
25 #include "llvm/Support/raw_ostream.h"
26 #include <algorithm>
27 using namespace llvm;
28
29 namespace {
30 struct OrderMap {
31 DenseMap<const Value *, std::pair<unsigned, bool>> IDs;
32 unsigned LastGlobalConstantID;
33 unsigned LastGlobalValueID;
34
OrderMap__anon409929a60111::OrderMap35 OrderMap() : LastGlobalConstantID(0), LastGlobalValueID(0) {}
36
isGlobalConstant__anon409929a60111::OrderMap37 bool isGlobalConstant(unsigned ID) const {
38 return ID <= LastGlobalConstantID;
39 }
isGlobalValue__anon409929a60111::OrderMap40 bool isGlobalValue(unsigned ID) const {
41 return ID <= LastGlobalValueID && !isGlobalConstant(ID);
42 }
43
size__anon409929a60111::OrderMap44 unsigned size() const { return IDs.size(); }
operator []__anon409929a60111::OrderMap45 std::pair<unsigned, bool> &operator[](const Value *V) { return IDs[V]; }
lookup__anon409929a60111::OrderMap46 std::pair<unsigned, bool> lookup(const Value *V) const {
47 return IDs.lookup(V);
48 }
index__anon409929a60111::OrderMap49 void index(const Value *V) {
50 // Explicitly sequence get-size and insert-value operations to avoid UB.
51 unsigned ID = IDs.size() + 1;
52 IDs[V].first = ID;
53 }
54 };
55 }
56
orderValue(const Value * V,OrderMap & OM)57 static void orderValue(const Value *V, OrderMap &OM) {
58 if (OM.lookup(V).first)
59 return;
60
61 if (const Constant *C = dyn_cast<Constant>(V))
62 if (C->getNumOperands() && !isa<GlobalValue>(C))
63 for (const Value *Op : C->operands())
64 if (!isa<BasicBlock>(Op) && !isa<GlobalValue>(Op))
65 orderValue(Op, OM);
66
67 // Note: we cannot cache this lookup above, since inserting into the map
68 // changes the map's size, and thus affects the other IDs.
69 OM.index(V);
70 }
71
orderModule(const Module & M)72 static OrderMap orderModule(const Module &M) {
73 // This needs to match the order used by ValueEnumerator::ValueEnumerator()
74 // and ValueEnumerator::incorporateFunction().
75 OrderMap OM;
76
77 // In the reader, initializers of GlobalValues are set *after* all the
78 // globals have been read. Rather than awkwardly modeling this behaviour
79 // directly in predictValueUseListOrderImpl(), just assign IDs to
80 // initializers of GlobalValues before GlobalValues themselves to model this
81 // implicitly.
82 for (const GlobalVariable &G : M.globals())
83 if (G.hasInitializer())
84 if (!isa<GlobalValue>(G.getInitializer()))
85 orderValue(G.getInitializer(), OM);
86 for (const GlobalAlias &A : M.aliases())
87 if (!isa<GlobalValue>(A.getAliasee()))
88 orderValue(A.getAliasee(), OM);
89 for (const Function &F : M) {
90 if (F.hasPrefixData())
91 if (!isa<GlobalValue>(F.getPrefixData()))
92 orderValue(F.getPrefixData(), OM);
93 if (F.hasPrologueData())
94 if (!isa<GlobalValue>(F.getPrologueData()))
95 orderValue(F.getPrologueData(), OM);
96 }
97 OM.LastGlobalConstantID = OM.size();
98
99 // Initializers of GlobalValues are processed in
100 // BitcodeReader::ResolveGlobalAndAliasInits(). Match the order there rather
101 // than ValueEnumerator, and match the code in predictValueUseListOrderImpl()
102 // by giving IDs in reverse order.
103 //
104 // Since GlobalValues never reference each other directly (just through
105 // initializers), their relative IDs only matter for determining order of
106 // uses in their initializers.
107 for (const Function &F : M)
108 orderValue(&F, OM);
109 for (const GlobalAlias &A : M.aliases())
110 orderValue(&A, OM);
111 for (const GlobalVariable &G : M.globals())
112 orderValue(&G, OM);
113 OM.LastGlobalValueID = OM.size();
114
115 for (const Function &F : M) {
116 if (F.isDeclaration())
117 continue;
118 // Here we need to match the union of ValueEnumerator::incorporateFunction()
119 // and WriteFunction(). Basic blocks are implicitly declared before
120 // anything else (by declaring their size).
121 for (const BasicBlock &BB : F)
122 orderValue(&BB, OM);
123 for (const Argument &A : F.args())
124 orderValue(&A, OM);
125 for (const BasicBlock &BB : F)
126 for (const Instruction &I : BB)
127 for (const Value *Op : I.operands())
128 if ((isa<Constant>(*Op) && !isa<GlobalValue>(*Op)) ||
129 isa<InlineAsm>(*Op))
130 orderValue(Op, OM);
131 for (const BasicBlock &BB : F)
132 for (const Instruction &I : BB)
133 orderValue(&I, OM);
134 }
135 return OM;
136 }
137
predictValueUseListOrderImpl(const Value * V,const Function * F,unsigned ID,const OrderMap & OM,UseListOrderStack & Stack)138 static void predictValueUseListOrderImpl(const Value *V, const Function *F,
139 unsigned ID, const OrderMap &OM,
140 UseListOrderStack &Stack) {
141 // Predict use-list order for this one.
142 typedef std::pair<const Use *, unsigned> Entry;
143 SmallVector<Entry, 64> List;
144 for (const Use &U : V->uses())
145 // Check if this user will be serialized.
146 if (OM.lookup(U.getUser()).first)
147 List.push_back(std::make_pair(&U, List.size()));
148
149 if (List.size() < 2)
150 // We may have lost some users.
151 return;
152
153 bool IsGlobalValue = OM.isGlobalValue(ID);
154 std::sort(List.begin(), List.end(), [&](const Entry &L, const Entry &R) {
155 const Use *LU = L.first;
156 const Use *RU = R.first;
157 if (LU == RU)
158 return false;
159
160 auto LID = OM.lookup(LU->getUser()).first;
161 auto RID = OM.lookup(RU->getUser()).first;
162
163 // Global values are processed in reverse order.
164 //
165 // Moreover, initializers of GlobalValues are set *after* all the globals
166 // have been read (despite having earlier IDs). Rather than awkwardly
167 // modeling this behaviour here, orderModule() has assigned IDs to
168 // initializers of GlobalValues before GlobalValues themselves.
169 if (OM.isGlobalValue(LID) && OM.isGlobalValue(RID))
170 return LID < RID;
171
172 // If ID is 4, then expect: 7 6 5 1 2 3.
173 if (LID < RID) {
174 if (RID <= ID)
175 if (!IsGlobalValue) // GlobalValue uses don't get reversed.
176 return true;
177 return false;
178 }
179 if (RID < LID) {
180 if (LID <= ID)
181 if (!IsGlobalValue) // GlobalValue uses don't get reversed.
182 return false;
183 return true;
184 }
185
186 // LID and RID are equal, so we have different operands of the same user.
187 // Assume operands are added in order for all instructions.
188 if (LID <= ID)
189 if (!IsGlobalValue) // GlobalValue uses don't get reversed.
190 return LU->getOperandNo() < RU->getOperandNo();
191 return LU->getOperandNo() > RU->getOperandNo();
192 });
193
194 if (std::is_sorted(
195 List.begin(), List.end(),
196 [](const Entry &L, const Entry &R) { return L.second < R.second; }))
197 // Order is already correct.
198 return;
199
200 // Store the shuffle.
201 Stack.emplace_back(V, F, List.size());
202 assert(List.size() == Stack.back().Shuffle.size() && "Wrong size");
203 for (size_t I = 0, E = List.size(); I != E; ++I)
204 Stack.back().Shuffle[I] = List[I].second;
205 }
206
predictValueUseListOrder(const Value * V,const Function * F,OrderMap & OM,UseListOrderStack & Stack)207 static void predictValueUseListOrder(const Value *V, const Function *F,
208 OrderMap &OM, UseListOrderStack &Stack) {
209 auto &IDPair = OM[V];
210 assert(IDPair.first && "Unmapped value");
211 if (IDPair.second)
212 // Already predicted.
213 return;
214
215 // Do the actual prediction.
216 IDPair.second = true;
217 if (!V->use_empty() && std::next(V->use_begin()) != V->use_end())
218 predictValueUseListOrderImpl(V, F, IDPair.first, OM, Stack);
219
220 // Recursive descent into constants.
221 if (const Constant *C = dyn_cast<Constant>(V))
222 if (C->getNumOperands()) // Visit GlobalValues.
223 for (const Value *Op : C->operands())
224 if (isa<Constant>(Op)) // Visit GlobalValues.
225 predictValueUseListOrder(Op, F, OM, Stack);
226 }
227
predictUseListOrder(const Module & M)228 static UseListOrderStack predictUseListOrder(const Module &M) {
229 OrderMap OM = orderModule(M);
230
231 // Use-list orders need to be serialized after all the users have been added
232 // to a value, or else the shuffles will be incomplete. Store them per
233 // function in a stack.
234 //
235 // Aside from function order, the order of values doesn't matter much here.
236 UseListOrderStack Stack;
237
238 // We want to visit the functions backward now so we can list function-local
239 // constants in the last Function they're used in. Module-level constants
240 // have already been visited above.
241 for (auto I = M.rbegin(), E = M.rend(); I != E; ++I) {
242 const Function &F = *I;
243 if (F.isDeclaration())
244 continue;
245 for (const BasicBlock &BB : F)
246 predictValueUseListOrder(&BB, &F, OM, Stack);
247 for (const Argument &A : F.args())
248 predictValueUseListOrder(&A, &F, OM, Stack);
249 for (const BasicBlock &BB : F)
250 for (const Instruction &I : BB)
251 for (const Value *Op : I.operands())
252 if (isa<Constant>(*Op) || isa<InlineAsm>(*Op)) // Visit GlobalValues.
253 predictValueUseListOrder(Op, &F, OM, Stack);
254 for (const BasicBlock &BB : F)
255 for (const Instruction &I : BB)
256 predictValueUseListOrder(&I, &F, OM, Stack);
257 }
258
259 // Visit globals last, since the module-level use-list block will be seen
260 // before the function bodies are processed.
261 for (const GlobalVariable &G : M.globals())
262 predictValueUseListOrder(&G, nullptr, OM, Stack);
263 for (const Function &F : M)
264 predictValueUseListOrder(&F, nullptr, OM, Stack);
265 for (const GlobalAlias &A : M.aliases())
266 predictValueUseListOrder(&A, nullptr, OM, Stack);
267 for (const GlobalVariable &G : M.globals())
268 if (G.hasInitializer())
269 predictValueUseListOrder(G.getInitializer(), nullptr, OM, Stack);
270 for (const GlobalAlias &A : M.aliases())
271 predictValueUseListOrder(A.getAliasee(), nullptr, OM, Stack);
272 for (const Function &F : M) {
273 if (F.hasPrefixData())
274 predictValueUseListOrder(F.getPrefixData(), nullptr, OM, Stack);
275 if (F.hasPrologueData())
276 predictValueUseListOrder(F.getPrologueData(), nullptr, OM, Stack);
277 }
278
279 return Stack;
280 }
281
isIntOrIntVectorValue(const std::pair<const Value *,unsigned> & V)282 static bool isIntOrIntVectorValue(const std::pair<const Value*, unsigned> &V) {
283 return V.first->getType()->isIntOrIntVectorTy();
284 }
285
ValueEnumerator(const Module & M,bool ShouldPreserveUseListOrder)286 ValueEnumerator::ValueEnumerator(const Module &M,
287 bool ShouldPreserveUseListOrder)
288 : HasMDString(false), HasMDLocation(false), HasGenericDebugNode(false),
289 ShouldPreserveUseListOrder(ShouldPreserveUseListOrder) {
290 if (ShouldPreserveUseListOrder)
291 UseListOrders = predictUseListOrder(M);
292
293 // Enumerate the global variables.
294 for (Module::const_global_iterator I = M.global_begin(), E = M.global_end();
295 I != E; ++I)
296 EnumerateValue(I);
297
298 // Enumerate the functions.
299 for (Module::const_iterator I = M.begin(), E = M.end(); I != E; ++I) {
300 EnumerateValue(I);
301 EnumerateAttributes(cast<Function>(I)->getAttributes());
302 }
303
304 // Enumerate the aliases.
305 for (Module::const_alias_iterator I = M.alias_begin(), E = M.alias_end();
306 I != E; ++I)
307 EnumerateValue(I);
308
309 // Remember what is the cutoff between globalvalue's and other constants.
310 unsigned FirstConstant = Values.size();
311
312 // Enumerate the global variable initializers.
313 for (Module::const_global_iterator I = M.global_begin(), E = M.global_end();
314 I != E; ++I)
315 if (I->hasInitializer())
316 EnumerateValue(I->getInitializer());
317
318 // Enumerate the aliasees.
319 for (Module::const_alias_iterator I = M.alias_begin(), E = M.alias_end();
320 I != E; ++I)
321 EnumerateValue(I->getAliasee());
322
323 // Enumerate the prefix data constants.
324 for (Module::const_iterator I = M.begin(), E = M.end(); I != E; ++I)
325 if (I->hasPrefixData())
326 EnumerateValue(I->getPrefixData());
327
328 // Enumerate the prologue data constants.
329 for (Module::const_iterator I = M.begin(), E = M.end(); I != E; ++I)
330 if (I->hasPrologueData())
331 EnumerateValue(I->getPrologueData());
332
333 // Enumerate the metadata type.
334 //
335 // TODO: Move this to ValueEnumerator::EnumerateOperandType() once bitcode
336 // only encodes the metadata type when it's used as a value.
337 EnumerateType(Type::getMetadataTy(M.getContext()));
338
339 // Insert constants and metadata that are named at module level into the slot
340 // pool so that the module symbol table can refer to them...
341 EnumerateValueSymbolTable(M.getValueSymbolTable());
342 EnumerateNamedMetadata(M);
343
344 SmallVector<std::pair<unsigned, MDNode *>, 8> MDs;
345
346 // Enumerate types used by function bodies and argument lists.
347 for (const Function &F : M) {
348 for (const Argument &A : F.args())
349 EnumerateType(A.getType());
350
351 for (const BasicBlock &BB : F)
352 for (const Instruction &I : BB) {
353 for (const Use &Op : I.operands()) {
354 auto *MD = dyn_cast<MetadataAsValue>(&Op);
355 if (!MD) {
356 EnumerateOperandType(Op);
357 continue;
358 }
359
360 // Local metadata is enumerated during function-incorporation.
361 if (isa<LocalAsMetadata>(MD->getMetadata()))
362 continue;
363
364 EnumerateMetadata(MD->getMetadata());
365 }
366 EnumerateType(I.getType());
367 if (const CallInst *CI = dyn_cast<CallInst>(&I))
368 EnumerateAttributes(CI->getAttributes());
369 else if (const InvokeInst *II = dyn_cast<InvokeInst>(&I))
370 EnumerateAttributes(II->getAttributes());
371
372 // Enumerate metadata attached with this instruction.
373 MDs.clear();
374 I.getAllMetadataOtherThanDebugLoc(MDs);
375 for (unsigned i = 0, e = MDs.size(); i != e; ++i)
376 EnumerateMetadata(MDs[i].second);
377
378 // Don't enumerate the location directly -- it has a special record
379 // type -- but enumerate its operands.
380 if (MDLocation *L = I.getDebugLoc())
381 EnumerateMDNodeOperands(L);
382 }
383 }
384
385 // Optimize constant ordering.
386 OptimizeConstants(FirstConstant, Values.size());
387 }
388
getInstructionID(const Instruction * Inst) const389 unsigned ValueEnumerator::getInstructionID(const Instruction *Inst) const {
390 InstructionMapType::const_iterator I = InstructionMap.find(Inst);
391 assert(I != InstructionMap.end() && "Instruction is not mapped!");
392 return I->second;
393 }
394
getComdatID(const Comdat * C) const395 unsigned ValueEnumerator::getComdatID(const Comdat *C) const {
396 unsigned ComdatID = Comdats.idFor(C);
397 assert(ComdatID && "Comdat not found!");
398 return ComdatID;
399 }
400
setInstructionID(const Instruction * I)401 void ValueEnumerator::setInstructionID(const Instruction *I) {
402 InstructionMap[I] = InstructionCount++;
403 }
404
getValueID(const Value * V) const405 unsigned ValueEnumerator::getValueID(const Value *V) const {
406 if (auto *MD = dyn_cast<MetadataAsValue>(V))
407 return getMetadataID(MD->getMetadata());
408
409 ValueMapType::const_iterator I = ValueMap.find(V);
410 assert(I != ValueMap.end() && "Value not in slotcalculator!");
411 return I->second-1;
412 }
413
dump() const414 void ValueEnumerator::dump() const {
415 print(dbgs(), ValueMap, "Default");
416 dbgs() << '\n';
417 print(dbgs(), MDValueMap, "MetaData");
418 dbgs() << '\n';
419 }
420
print(raw_ostream & OS,const ValueMapType & Map,const char * Name) const421 void ValueEnumerator::print(raw_ostream &OS, const ValueMapType &Map,
422 const char *Name) const {
423
424 OS << "Map Name: " << Name << "\n";
425 OS << "Size: " << Map.size() << "\n";
426 for (ValueMapType::const_iterator I = Map.begin(),
427 E = Map.end(); I != E; ++I) {
428
429 const Value *V = I->first;
430 if (V->hasName())
431 OS << "Value: " << V->getName();
432 else
433 OS << "Value: [null]\n";
434 V->dump();
435
436 OS << " Uses(" << std::distance(V->use_begin(),V->use_end()) << "):";
437 for (const Use &U : V->uses()) {
438 if (&U != &*V->use_begin())
439 OS << ",";
440 if(U->hasName())
441 OS << " " << U->getName();
442 else
443 OS << " [null]";
444
445 }
446 OS << "\n\n";
447 }
448 }
449
print(raw_ostream & OS,const MetadataMapType & Map,const char * Name) const450 void ValueEnumerator::print(raw_ostream &OS, const MetadataMapType &Map,
451 const char *Name) const {
452
453 OS << "Map Name: " << Name << "\n";
454 OS << "Size: " << Map.size() << "\n";
455 for (auto I = Map.begin(), E = Map.end(); I != E; ++I) {
456 const Metadata *MD = I->first;
457 OS << "Metadata: slot = " << I->second << "\n";
458 MD->print(OS);
459 }
460 }
461
462 /// OptimizeConstants - Reorder constant pool for denser encoding.
OptimizeConstants(unsigned CstStart,unsigned CstEnd)463 void ValueEnumerator::OptimizeConstants(unsigned CstStart, unsigned CstEnd) {
464 if (CstStart == CstEnd || CstStart+1 == CstEnd) return;
465
466 if (ShouldPreserveUseListOrder)
467 // Optimizing constants makes the use-list order difficult to predict.
468 // Disable it for now when trying to preserve the order.
469 return;
470
471 std::stable_sort(Values.begin() + CstStart, Values.begin() + CstEnd,
472 [this](const std::pair<const Value *, unsigned> &LHS,
473 const std::pair<const Value *, unsigned> &RHS) {
474 // Sort by plane.
475 if (LHS.first->getType() != RHS.first->getType())
476 return getTypeID(LHS.first->getType()) < getTypeID(RHS.first->getType());
477 // Then by frequency.
478 return LHS.second > RHS.second;
479 });
480
481 // Ensure that integer and vector of integer constants are at the start of the
482 // constant pool. This is important so that GEP structure indices come before
483 // gep constant exprs.
484 std::partition(Values.begin()+CstStart, Values.begin()+CstEnd,
485 isIntOrIntVectorValue);
486
487 // Rebuild the modified portion of ValueMap.
488 for (; CstStart != CstEnd; ++CstStart)
489 ValueMap[Values[CstStart].first] = CstStart+1;
490 }
491
492
493 /// EnumerateValueSymbolTable - Insert all of the values in the specified symbol
494 /// table into the values table.
EnumerateValueSymbolTable(const ValueSymbolTable & VST)495 void ValueEnumerator::EnumerateValueSymbolTable(const ValueSymbolTable &VST) {
496 for (ValueSymbolTable::const_iterator VI = VST.begin(), VE = VST.end();
497 VI != VE; ++VI)
498 EnumerateValue(VI->getValue());
499 }
500
501 /// Insert all of the values referenced by named metadata in the specified
502 /// module.
EnumerateNamedMetadata(const Module & M)503 void ValueEnumerator::EnumerateNamedMetadata(const Module &M) {
504 for (Module::const_named_metadata_iterator I = M.named_metadata_begin(),
505 E = M.named_metadata_end();
506 I != E; ++I)
507 EnumerateNamedMDNode(I);
508 }
509
EnumerateNamedMDNode(const NamedMDNode * MD)510 void ValueEnumerator::EnumerateNamedMDNode(const NamedMDNode *MD) {
511 for (unsigned i = 0, e = MD->getNumOperands(); i != e; ++i)
512 EnumerateMetadata(MD->getOperand(i));
513 }
514
515 /// EnumerateMDNodeOperands - Enumerate all non-function-local values
516 /// and types referenced by the given MDNode.
EnumerateMDNodeOperands(const MDNode * N)517 void ValueEnumerator::EnumerateMDNodeOperands(const MDNode *N) {
518 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
519 Metadata *MD = N->getOperand(i);
520 if (!MD)
521 continue;
522 assert(!isa<LocalAsMetadata>(MD) && "MDNodes cannot be function-local");
523 EnumerateMetadata(MD);
524 }
525 }
526
EnumerateMetadata(const Metadata * MD)527 void ValueEnumerator::EnumerateMetadata(const Metadata *MD) {
528 assert(
529 (isa<MDNode>(MD) || isa<MDString>(MD) || isa<ConstantAsMetadata>(MD)) &&
530 "Invalid metadata kind");
531
532 // Insert a dummy ID to block the co-recursive call to
533 // EnumerateMDNodeOperands() from re-visiting MD in a cyclic graph.
534 //
535 // Return early if there's already an ID.
536 if (!MDValueMap.insert(std::make_pair(MD, 0)).second)
537 return;
538
539 // Visit operands first to minimize RAUW.
540 if (auto *N = dyn_cast<MDNode>(MD))
541 EnumerateMDNodeOperands(N);
542 else if (auto *C = dyn_cast<ConstantAsMetadata>(MD))
543 EnumerateValue(C->getValue());
544
545 HasMDString |= isa<MDString>(MD);
546 HasMDLocation |= isa<MDLocation>(MD);
547 HasGenericDebugNode |= isa<GenericDebugNode>(MD);
548
549 // Replace the dummy ID inserted above with the correct one. MDValueMap may
550 // have changed by inserting operands, so we need a fresh lookup here.
551 MDs.push_back(MD);
552 MDValueMap[MD] = MDs.size();
553 }
554
555 /// EnumerateFunctionLocalMetadataa - Incorporate function-local metadata
556 /// information reachable from the metadata.
EnumerateFunctionLocalMetadata(const LocalAsMetadata * Local)557 void ValueEnumerator::EnumerateFunctionLocalMetadata(
558 const LocalAsMetadata *Local) {
559 // Check to see if it's already in!
560 unsigned &MDValueID = MDValueMap[Local];
561 if (MDValueID)
562 return;
563
564 MDs.push_back(Local);
565 MDValueID = MDs.size();
566
567 EnumerateValue(Local->getValue());
568
569 // Also, collect all function-local metadata for easy access.
570 FunctionLocalMDs.push_back(Local);
571 }
572
EnumerateValue(const Value * V)573 void ValueEnumerator::EnumerateValue(const Value *V) {
574 assert(!V->getType()->isVoidTy() && "Can't insert void values!");
575 assert(!isa<MetadataAsValue>(V) && "EnumerateValue doesn't handle Metadata!");
576
577 // Check to see if it's already in!
578 unsigned &ValueID = ValueMap[V];
579 if (ValueID) {
580 // Increment use count.
581 Values[ValueID-1].second++;
582 return;
583 }
584
585 if (auto *GO = dyn_cast<GlobalObject>(V))
586 if (const Comdat *C = GO->getComdat())
587 Comdats.insert(C);
588
589 // Enumerate the type of this value.
590 EnumerateType(V->getType());
591
592 if (const Constant *C = dyn_cast<Constant>(V)) {
593 if (isa<GlobalValue>(C)) {
594 // Initializers for globals are handled explicitly elsewhere.
595 } else if (C->getNumOperands()) {
596 // If a constant has operands, enumerate them. This makes sure that if a
597 // constant has uses (for example an array of const ints), that they are
598 // inserted also.
599
600 // We prefer to enumerate them with values before we enumerate the user
601 // itself. This makes it more likely that we can avoid forward references
602 // in the reader. We know that there can be no cycles in the constants
603 // graph that don't go through a global variable.
604 for (User::const_op_iterator I = C->op_begin(), E = C->op_end();
605 I != E; ++I)
606 if (!isa<BasicBlock>(*I)) // Don't enumerate BB operand to BlockAddress.
607 EnumerateValue(*I);
608
609 // Finally, add the value. Doing this could make the ValueID reference be
610 // dangling, don't reuse it.
611 Values.push_back(std::make_pair(V, 1U));
612 ValueMap[V] = Values.size();
613 return;
614 }
615 }
616
617 // Add the value.
618 Values.push_back(std::make_pair(V, 1U));
619 ValueID = Values.size();
620 }
621
622
EnumerateType(Type * Ty)623 void ValueEnumerator::EnumerateType(Type *Ty) {
624 unsigned *TypeID = &TypeMap[Ty];
625
626 // We've already seen this type.
627 if (*TypeID)
628 return;
629
630 // If it is a non-anonymous struct, mark the type as being visited so that we
631 // don't recursively visit it. This is safe because we allow forward
632 // references of these in the bitcode reader.
633 if (StructType *STy = dyn_cast<StructType>(Ty))
634 if (!STy->isLiteral())
635 *TypeID = ~0U;
636
637 // Enumerate all of the subtypes before we enumerate this type. This ensures
638 // that the type will be enumerated in an order that can be directly built.
639 for (Type *SubTy : Ty->subtypes())
640 EnumerateType(SubTy);
641
642 // Refresh the TypeID pointer in case the table rehashed.
643 TypeID = &TypeMap[Ty];
644
645 // Check to see if we got the pointer another way. This can happen when
646 // enumerating recursive types that hit the base case deeper than they start.
647 //
648 // If this is actually a struct that we are treating as forward ref'able,
649 // then emit the definition now that all of its contents are available.
650 if (*TypeID && *TypeID != ~0U)
651 return;
652
653 // Add this type now that its contents are all happily enumerated.
654 Types.push_back(Ty);
655
656 *TypeID = Types.size();
657 }
658
659 // Enumerate the types for the specified value. If the value is a constant,
660 // walk through it, enumerating the types of the constant.
EnumerateOperandType(const Value * V)661 void ValueEnumerator::EnumerateOperandType(const Value *V) {
662 EnumerateType(V->getType());
663
664 if (auto *MD = dyn_cast<MetadataAsValue>(V)) {
665 assert(!isa<LocalAsMetadata>(MD->getMetadata()) &&
666 "Function-local metadata should be left for later");
667
668 EnumerateMetadata(MD->getMetadata());
669 return;
670 }
671
672 const Constant *C = dyn_cast<Constant>(V);
673 if (!C)
674 return;
675
676 // If this constant is already enumerated, ignore it, we know its type must
677 // be enumerated.
678 if (ValueMap.count(C))
679 return;
680
681 // This constant may have operands, make sure to enumerate the types in
682 // them.
683 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) {
684 const Value *Op = C->getOperand(i);
685
686 // Don't enumerate basic blocks here, this happens as operands to
687 // blockaddress.
688 if (isa<BasicBlock>(Op))
689 continue;
690
691 EnumerateOperandType(Op);
692 }
693 }
694
EnumerateAttributes(AttributeSet PAL)695 void ValueEnumerator::EnumerateAttributes(AttributeSet PAL) {
696 if (PAL.isEmpty()) return; // null is always 0.
697
698 // Do a lookup.
699 unsigned &Entry = AttributeMap[PAL];
700 if (Entry == 0) {
701 // Never saw this before, add it.
702 Attribute.push_back(PAL);
703 Entry = Attribute.size();
704 }
705
706 // Do lookups for all attribute groups.
707 for (unsigned i = 0, e = PAL.getNumSlots(); i != e; ++i) {
708 AttributeSet AS = PAL.getSlotAttributes(i);
709 unsigned &Entry = AttributeGroupMap[AS];
710 if (Entry == 0) {
711 AttributeGroups.push_back(AS);
712 Entry = AttributeGroups.size();
713 }
714 }
715 }
716
incorporateFunction(const Function & F)717 void ValueEnumerator::incorporateFunction(const Function &F) {
718 InstructionCount = 0;
719 NumModuleValues = Values.size();
720 NumModuleMDs = MDs.size();
721
722 // Adding function arguments to the value table.
723 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
724 I != E; ++I)
725 EnumerateValue(I);
726
727 FirstFuncConstantID = Values.size();
728
729 // Add all function-level constants to the value table.
730 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
731 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E; ++I)
732 for (User::const_op_iterator OI = I->op_begin(), E = I->op_end();
733 OI != E; ++OI) {
734 if ((isa<Constant>(*OI) && !isa<GlobalValue>(*OI)) ||
735 isa<InlineAsm>(*OI))
736 EnumerateValue(*OI);
737 }
738 BasicBlocks.push_back(BB);
739 ValueMap[BB] = BasicBlocks.size();
740 }
741
742 // Optimize the constant layout.
743 OptimizeConstants(FirstFuncConstantID, Values.size());
744
745 // Add the function's parameter attributes so they are available for use in
746 // the function's instruction.
747 EnumerateAttributes(F.getAttributes());
748
749 FirstInstID = Values.size();
750
751 SmallVector<LocalAsMetadata *, 8> FnLocalMDVector;
752 // Add all of the instructions.
753 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
754 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E; ++I) {
755 for (User::const_op_iterator OI = I->op_begin(), E = I->op_end();
756 OI != E; ++OI) {
757 if (auto *MD = dyn_cast<MetadataAsValue>(&*OI))
758 if (auto *Local = dyn_cast<LocalAsMetadata>(MD->getMetadata()))
759 // Enumerate metadata after the instructions they might refer to.
760 FnLocalMDVector.push_back(Local);
761 }
762
763 if (!I->getType()->isVoidTy())
764 EnumerateValue(I);
765 }
766 }
767
768 // Add all of the function-local metadata.
769 for (unsigned i = 0, e = FnLocalMDVector.size(); i != e; ++i)
770 EnumerateFunctionLocalMetadata(FnLocalMDVector[i]);
771 }
772
purgeFunction()773 void ValueEnumerator::purgeFunction() {
774 /// Remove purged values from the ValueMap.
775 for (unsigned i = NumModuleValues, e = Values.size(); i != e; ++i)
776 ValueMap.erase(Values[i].first);
777 for (unsigned i = NumModuleMDs, e = MDs.size(); i != e; ++i)
778 MDValueMap.erase(MDs[i]);
779 for (unsigned i = 0, e = BasicBlocks.size(); i != e; ++i)
780 ValueMap.erase(BasicBlocks[i]);
781
782 Values.resize(NumModuleValues);
783 MDs.resize(NumModuleMDs);
784 BasicBlocks.clear();
785 FunctionLocalMDs.clear();
786 }
787
IncorporateFunctionInfoGlobalBBIDs(const Function * F,DenseMap<const BasicBlock *,unsigned> & IDMap)788 static void IncorporateFunctionInfoGlobalBBIDs(const Function *F,
789 DenseMap<const BasicBlock*, unsigned> &IDMap) {
790 unsigned Counter = 0;
791 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
792 IDMap[BB] = ++Counter;
793 }
794
795 /// getGlobalBasicBlockID - This returns the function-specific ID for the
796 /// specified basic block. This is relatively expensive information, so it
797 /// should only be used by rare constructs such as address-of-label.
getGlobalBasicBlockID(const BasicBlock * BB) const798 unsigned ValueEnumerator::getGlobalBasicBlockID(const BasicBlock *BB) const {
799 unsigned &Idx = GlobalBasicBlockIDs[BB];
800 if (Idx != 0)
801 return Idx-1;
802
803 IncorporateFunctionInfoGlobalBBIDs(BB->getParent(), GlobalBasicBlockIDs);
804 return getGlobalBasicBlockID(BB);
805 }
806
computeBitsRequiredForTypeIndicies() const807 uint64_t ValueEnumerator::computeBitsRequiredForTypeIndicies() const {
808 return Log2_32_Ceil(getTypes().size() + 1);
809 }
810