1 //===-- ValueEnumerator.cpp - Number values and types for bitcode writer --===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the ValueEnumerator class.
12 //===----------------------------------------------------------------------===//
14 #include "ValueEnumerator.h"
15 #include "llvm/Constants.h"
16 #include "llvm/DerivedTypes.h"
17 #include "llvm/Module.h"
18 #include "llvm/TypeSymbolTable.h"
19 #include "llvm/ValueSymbolTable.h"
20 #include "llvm/Instructions.h"
24 static bool isSingleValueType(const std::pair<const llvm::Type*,
26 return P.first->isSingleValueType();
29 static bool isIntegerValue(const std::pair<const Value*, unsigned> &V) {
30 return isa<IntegerType>(V.first->getType());
33 static bool CompareByFrequency(const std::pair<const llvm::Type*,
35 const std::pair<const llvm::Type*,
37 return P1.second > P2.second;
40 /// ValueEnumerator - Enumerate module-level information.
41 ValueEnumerator::ValueEnumerator(const Module *M) {
44 // Enumerate the global variables.
45 for (Module::const_global_iterator I = M->global_begin(),
46 E = M->global_end(); I != E; ++I)
49 // Enumerate the functions.
50 for (Module::const_iterator I = M->begin(), E = M->end(); I != E; ++I) {
52 EnumerateAttributes(cast<Function>(I)->getAttributes());
55 // Enumerate the aliases.
56 for (Module::const_alias_iterator I = M->alias_begin(), E = M->alias_end();
60 // Remember what is the cutoff between globalvalue's and other constants.
61 unsigned FirstConstant = Values.size();
63 // Enumerate the global variable initializers.
64 for (Module::const_global_iterator I = M->global_begin(),
65 E = M->global_end(); I != E; ++I)
66 if (I->hasInitializer())
67 EnumerateValue(I->getInitializer());
69 // Enumerate the aliasees.
70 for (Module::const_alias_iterator I = M->alias_begin(), E = M->alias_end();
72 EnumerateValue(I->getAliasee());
74 // Enumerate types used by the type symbol table.
75 EnumerateTypeSymbolTable(M->getTypeSymbolTable());
77 // Insert constants that are named at module level into the slot pool so that
78 // the module symbol table can refer to them...
79 EnumerateValueSymbolTable(M->getValueSymbolTable());
81 // Enumerate types used by function bodies and argument lists.
82 for (Module::const_iterator F = M->begin(), E = M->end(); F != E; ++F) {
84 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
86 EnumerateType(I->getType());
88 SmallVector<std::pair<unsigned, MDNode*>, 2> MDs;
89 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
90 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E;++I){
91 for (User::const_op_iterator OI = I->op_begin(), E = I->op_end();
93 EnumerateOperandType(*OI);
94 EnumerateType(I->getType());
95 if (const CallInst *CI = dyn_cast<CallInst>(I))
96 EnumerateAttributes(CI->getAttributes());
97 else if (const InvokeInst *II = dyn_cast<InvokeInst>(I))
98 EnumerateAttributes(II->getAttributes());
100 // Enumerate metadata attached with this instruction.
102 I->getAllMetadata(MDs);
103 for (unsigned i = 0, e = MDs.size(); i != e; ++i)
104 EnumerateMetadata(MDs[i].second);
108 // Optimize constant ordering.
109 OptimizeConstants(FirstConstant, Values.size());
111 // Sort the type table by frequency so that most commonly used types are early
112 // in the table (have low bit-width).
113 std::stable_sort(Types.begin(), Types.end(), CompareByFrequency);
115 // Partition the Type ID's so that the single-value types occur before the
116 // aggregate types. This allows the aggregate types to be dropped from the
117 // type table after parsing the global variable initializers.
118 std::partition(Types.begin(), Types.end(), isSingleValueType);
120 // Now that we rearranged the type table, rebuild TypeMap.
121 for (unsigned i = 0, e = Types.size(); i != e; ++i)
122 TypeMap[Types[i].first] = i+1;
125 unsigned ValueEnumerator::getInstructionID(const Instruction *Inst) const {
126 InstructionMapType::const_iterator I = InstructionMap.find(Inst);
127 assert (I != InstructionMap.end() && "Instruction is not mapped!");
131 void ValueEnumerator::setInstructionID(const Instruction *I) {
132 InstructionMap[I] = InstructionCount++;
135 unsigned ValueEnumerator::getValueID(const Value *V) const {
136 if (isa<MetadataBase>(V)) {
137 ValueMapType::const_iterator I = MDValueMap.find(V);
138 assert(I != MDValueMap.end() && "Value not in slotcalculator!");
142 ValueMapType::const_iterator I = ValueMap.find(V);
143 assert(I != ValueMap.end() && "Value not in slotcalculator!");
147 // Optimize constant ordering.
149 struct CstSortPredicate {
151 explicit CstSortPredicate(ValueEnumerator &ve) : VE(ve) {}
152 bool operator()(const std::pair<const Value*, unsigned> &LHS,
153 const std::pair<const Value*, unsigned> &RHS) {
155 if (LHS.first->getType() != RHS.first->getType())
156 return VE.getTypeID(LHS.first->getType()) <
157 VE.getTypeID(RHS.first->getType());
158 // Then by frequency.
159 return LHS.second > RHS.second;
164 /// OptimizeConstants - Reorder constant pool for denser encoding.
165 void ValueEnumerator::OptimizeConstants(unsigned CstStart, unsigned CstEnd) {
166 if (CstStart == CstEnd || CstStart+1 == CstEnd) return;
168 CstSortPredicate P(*this);
169 std::stable_sort(Values.begin()+CstStart, Values.begin()+CstEnd, P);
171 // Ensure that integer constants are at the start of the constant pool. This
172 // is important so that GEP structure indices come before gep constant exprs.
173 std::partition(Values.begin()+CstStart, Values.begin()+CstEnd,
176 // Rebuild the modified portion of ValueMap.
177 for (; CstStart != CstEnd; ++CstStart)
178 ValueMap[Values[CstStart].first] = CstStart+1;
182 /// EnumerateTypeSymbolTable - Insert all of the types in the specified symbol
184 void ValueEnumerator::EnumerateTypeSymbolTable(const TypeSymbolTable &TST) {
185 for (TypeSymbolTable::const_iterator TI = TST.begin(), TE = TST.end();
187 EnumerateType(TI->second);
190 /// EnumerateValueSymbolTable - Insert all of the values in the specified symbol
191 /// table into the values table.
192 void ValueEnumerator::EnumerateValueSymbolTable(const ValueSymbolTable &VST) {
193 for (ValueSymbolTable::const_iterator VI = VST.begin(), VE = VST.end();
195 EnumerateValue(VI->getValue());
198 void ValueEnumerator::EnumerateMetadata(const MetadataBase *MD) {
199 // Check to see if it's already in!
200 unsigned &MDValueID = MDValueMap[MD];
202 // Increment use count.
203 MDValues[MDValueID-1].second++;
207 // Enumerate the type of this value.
208 EnumerateType(MD->getType());
210 if (const MDNode *N = dyn_cast<MDNode>(MD)) {
211 MDValues.push_back(std::make_pair(MD, 1U));
212 MDValueMap[MD] = MDValues.size();
213 MDValueID = MDValues.size();
214 for (unsigned i = 0, e = N->getNumElements(); i != e; ++i) {
215 if (Value *V = N->getElement(i))
218 EnumerateType(Type::getVoidTy(MD->getContext()));
223 if (const NamedMDNode *N = dyn_cast<NamedMDNode>(MD)) {
224 for (unsigned i = 0, e = N->getNumElements(); i != e; ++i)
225 EnumerateValue(N->getElement(i));
226 MDValues.push_back(std::make_pair(MD, 1U));
227 MDValueMap[MD] = Values.size();
232 MDValues.push_back(std::make_pair(MD, 1U));
233 MDValueID = MDValues.size();
236 void ValueEnumerator::EnumerateValue(const Value *V) {
237 assert(V->getType() != Type::getVoidTy(V->getContext()) &&
238 "Can't insert void values!");
239 if (const MetadataBase *MB = dyn_cast<MetadataBase>(V))
240 return EnumerateMetadata(MB);
242 // Check to see if it's already in!
243 unsigned &ValueID = ValueMap[V];
245 // Increment use count.
246 Values[ValueID-1].second++;
250 // Enumerate the type of this value.
251 EnumerateType(V->getType());
253 if (const Constant *C = dyn_cast<Constant>(V)) {
254 if (isa<GlobalValue>(C)) {
255 // Initializers for globals are handled explicitly elsewhere.
256 } else if (isa<ConstantArray>(C) && cast<ConstantArray>(C)->isString()) {
257 // Do not enumerate the initializers for an array of simple characters.
258 // The initializers just polute the value table, and we emit the strings
260 } else if (C->getNumOperands()) {
261 // If a constant has operands, enumerate them. This makes sure that if a
262 // constant has uses (for example an array of const ints), that they are
265 // We prefer to enumerate them with values before we enumerate the user
266 // itself. This makes it more likely that we can avoid forward references
267 // in the reader. We know that there can be no cycles in the constants
268 // graph that don't go through a global variable.
269 for (User::const_op_iterator I = C->op_begin(), E = C->op_end();
271 if (!isa<BasicBlock>(*I)) // Don't enumerate BB operand to BlockAddress.
274 // Finally, add the value. Doing this could make the ValueID reference be
275 // dangling, don't reuse it.
276 Values.push_back(std::make_pair(V, 1U));
277 ValueMap[V] = Values.size();
283 Values.push_back(std::make_pair(V, 1U));
284 ValueID = Values.size();
288 void ValueEnumerator::EnumerateType(const Type *Ty) {
289 unsigned &TypeID = TypeMap[Ty];
292 // If we've already seen this type, just increase its occurrence count.
293 Types[TypeID-1].second++;
297 // First time we saw this type, add it.
298 Types.push_back(std::make_pair(Ty, 1U));
299 TypeID = Types.size();
301 // Enumerate subtypes.
302 for (Type::subtype_iterator I = Ty->subtype_begin(), E = Ty->subtype_end();
307 // Enumerate the types for the specified value. If the value is a constant,
308 // walk through it, enumerating the types of the constant.
309 void ValueEnumerator::EnumerateOperandType(const Value *V) {
310 EnumerateType(V->getType());
311 if (const Constant *C = dyn_cast<Constant>(V)) {
312 // If this constant is already enumerated, ignore it, we know its type must
314 if (ValueMap.count(V)) return;
316 // This constant may have operands, make sure to enumerate the types in
318 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) {
319 const User *Op = C->getOperand(i);
321 // Don't enumerate basic blocks here, this happens as operands to
323 if (isa<BasicBlock>(Op)) continue;
325 EnumerateOperandType(cast<Constant>(Op));
328 if (const MDNode *N = dyn_cast<MDNode>(V)) {
329 for (unsigned i = 0, e = N->getNumElements(); i != e; ++i)
330 if (Value *Elem = N->getElement(i))
331 EnumerateOperandType(Elem);
333 } else if (isa<MDString>(V) || isa<MDNode>(V))
337 void ValueEnumerator::EnumerateAttributes(const AttrListPtr &PAL) {
338 if (PAL.isEmpty()) return; // null is always 0.
340 unsigned &Entry = AttributeMap[PAL.getRawPointer()];
342 // Never saw this before, add it.
343 Attributes.push_back(PAL);
344 Entry = Attributes.size();
349 void ValueEnumerator::incorporateFunction(const Function &F) {
350 NumModuleValues = Values.size();
352 // Adding function arguments to the value table.
353 for(Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
357 FirstFuncConstantID = Values.size();
359 // Add all function-level constants to the value table.
360 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
361 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E; ++I)
362 for (User::const_op_iterator OI = I->op_begin(), E = I->op_end();
364 if ((isa<Constant>(*OI) && !isa<GlobalValue>(*OI)) ||
368 BasicBlocks.push_back(BB);
369 ValueMap[BB] = BasicBlocks.size();
372 // Optimize the constant layout.
373 OptimizeConstants(FirstFuncConstantID, Values.size());
375 // Add the function's parameter attributes so they are available for use in
376 // the function's instruction.
377 EnumerateAttributes(F.getAttributes());
379 FirstInstID = Values.size();
381 // Add all of the instructions.
382 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
383 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E; ++I) {
384 if (I->getType() != Type::getVoidTy(F.getContext()))
390 void ValueEnumerator::purgeFunction() {
391 /// Remove purged values from the ValueMap.
392 for (unsigned i = NumModuleValues, e = Values.size(); i != e; ++i)
393 ValueMap.erase(Values[i].first);
394 for (unsigned i = 0, e = BasicBlocks.size(); i != e; ++i)
395 ValueMap.erase(BasicBlocks[i]);
397 Values.resize(NumModuleValues);
401 static void IncorporateFunctionInfoGlobalBBIDs(const Function *F,
402 DenseMap<const BasicBlock*, unsigned> &IDMap) {
403 unsigned Counter = 0;
404 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
405 IDMap[BB] = ++Counter;
408 /// getGlobalBasicBlockID - This returns the function-specific ID for the
409 /// specified basic block. This is relatively expensive information, so it
410 /// should only be used by rare constructs such as address-of-label.
411 unsigned ValueEnumerator::getGlobalBasicBlockID(const BasicBlock *BB) const {
412 unsigned &Idx = GlobalBasicBlockIDs[BB];
416 IncorporateFunctionInfoGlobalBBIDs(BB->getParent(), GlobalBasicBlockIDs);
417 return getGlobalBasicBlockID(BB);