ValueEnumerator.cpp revision a279bc3da55691784064cb47200a1c584408b8ab
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/Constants.h" 16#include "llvm/DerivedTypes.h" 17#include "llvm/Metadata.h" 18#include "llvm/Module.h" 19#include "llvm/TypeSymbolTable.h" 20#include "llvm/ValueSymbolTable.h" 21#include "llvm/Instructions.h" 22#include <algorithm> 23using namespace llvm; 24 25static bool isSingleValueType(const std::pair<const llvm::Type*, 26 unsigned int> &P) { 27 return P.first->isSingleValueType(); 28} 29 30static bool isIntegerValue(const std::pair<const Value*, unsigned> &V) { 31 return isa<IntegerType>(V.first->getType()); 32} 33 34static bool CompareByFrequency(const std::pair<const llvm::Type*, 35 unsigned int> &P1, 36 const std::pair<const llvm::Type*, 37 unsigned int> &P2) { 38 return P1.second > P2.second; 39} 40 41/// ValueEnumerator - Enumerate module-level information. 42ValueEnumerator::ValueEnumerator(const Module *M) { 43 InstructionCount = 0; 44 45 // Enumerate the global variables. 46 for (Module::const_global_iterator I = M->global_begin(), 47 E = M->global_end(); I != E; ++I) 48 EnumerateValue(I); 49 50 // Enumerate the functions. 51 for (Module::const_iterator I = M->begin(), E = M->end(); I != E; ++I) { 52 EnumerateValue(I); 53 EnumerateAttributes(cast<Function>(I)->getAttributes()); 54 } 55 56 // Enumerate the aliases. 57 for (Module::const_alias_iterator I = M->alias_begin(), E = M->alias_end(); 58 I != E; ++I) 59 EnumerateValue(I); 60 61 // Remember what is the cutoff between globalvalue's and other constants. 62 unsigned FirstConstant = Values.size(); 63 64 // Enumerate the global variable initializers. 65 for (Module::const_global_iterator I = M->global_begin(), 66 E = M->global_end(); I != E; ++I) 67 if (I->hasInitializer()) 68 EnumerateValue(I->getInitializer()); 69 70 // Enumerate the aliasees. 71 for (Module::const_alias_iterator I = M->alias_begin(), E = M->alias_end(); 72 I != E; ++I) 73 EnumerateValue(I->getAliasee()); 74 75 // Enumerate types used by the type symbol table. 76 EnumerateTypeSymbolTable(M->getTypeSymbolTable()); 77 78 // Insert constants that are named at module level into the slot pool so that 79 // the module symbol table can refer to them... 80 EnumerateValueSymbolTable(M->getValueSymbolTable()); 81 82 // Enumerate types used by function bodies and argument lists. 83 for (Module::const_iterator F = M->begin(), E = M->end(); F != E; ++F) { 84 85 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end(); 86 I != E; ++I) 87 EnumerateType(I->getType()); 88 89 Metadata &TheMetadata = F->getContext().getMetadata(); 90 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) 91 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E;++I){ 92 for (User::const_op_iterator OI = I->op_begin(), E = I->op_end(); 93 OI != E; ++OI) 94 EnumerateOperandType(*OI); 95 EnumerateType(I->getType()); 96 if (const CallInst *CI = dyn_cast<CallInst>(I)) 97 EnumerateAttributes(CI->getAttributes()); 98 else if (const InvokeInst *II = dyn_cast<InvokeInst>(I)) 99 EnumerateAttributes(II->getAttributes()); 100 101 // Enumerate metadata attached with this instruction. 102 const Metadata::MDMapTy *MDs = TheMetadata.getMDs(I); 103 if (MDs) 104 for (Metadata::MDMapTy::const_iterator MI = MDs->begin(), 105 ME = MDs->end(); MI != ME; ++MI) 106 if (MDNode *MDN = dyn_cast_or_null<MDNode>(MI->second)) 107 EnumerateMetadata(MDN); 108 } 109 } 110 111 // Optimize constant ordering. 112 OptimizeConstants(FirstConstant, Values.size()); 113 114 // Sort the type table by frequency so that most commonly used types are early 115 // in the table (have low bit-width). 116 std::stable_sort(Types.begin(), Types.end(), CompareByFrequency); 117 118 // Partition the Type ID's so that the single-value types occur before the 119 // aggregate types. This allows the aggregate types to be dropped from the 120 // type table after parsing the global variable initializers. 121 std::partition(Types.begin(), Types.end(), isSingleValueType); 122 123 // Now that we rearranged the type table, rebuild TypeMap. 124 for (unsigned i = 0, e = Types.size(); i != e; ++i) 125 TypeMap[Types[i].first] = i+1; 126} 127 128unsigned ValueEnumerator::getInstructionID(const Instruction *Inst) const { 129 InstructionMapType::const_iterator I = InstructionMap.find(Inst); 130 assert (I != InstructionMap.end() && "Instruction is not mapped!"); 131 return I->second; 132} 133 134void ValueEnumerator::setInstructionID(const Instruction *I) { 135 InstructionMap[I] = InstructionCount++; 136} 137 138unsigned ValueEnumerator::getValueID(const Value *V) const { 139 if (isa<MetadataBase>(V)) { 140 ValueMapType::const_iterator I = MDValueMap.find(V); 141 assert(I != MDValueMap.end() && "Value not in slotcalculator!"); 142 return I->second-1; 143 } 144 145 ValueMapType::const_iterator I = ValueMap.find(V); 146 assert(I != ValueMap.end() && "Value not in slotcalculator!"); 147 return I->second-1; 148} 149 150// Optimize constant ordering. 151namespace { 152 struct CstSortPredicate { 153 ValueEnumerator &VE; 154 explicit CstSortPredicate(ValueEnumerator &ve) : VE(ve) {} 155 bool operator()(const std::pair<const Value*, unsigned> &LHS, 156 const std::pair<const Value*, unsigned> &RHS) { 157 // Sort by plane. 158 if (LHS.first->getType() != RHS.first->getType()) 159 return VE.getTypeID(LHS.first->getType()) < 160 VE.getTypeID(RHS.first->getType()); 161 // Then by frequency. 162 return LHS.second > RHS.second; 163 } 164 }; 165} 166 167/// OptimizeConstants - Reorder constant pool for denser encoding. 168void ValueEnumerator::OptimizeConstants(unsigned CstStart, unsigned CstEnd) { 169 if (CstStart == CstEnd || CstStart+1 == CstEnd) return; 170 171 CstSortPredicate P(*this); 172 std::stable_sort(Values.begin()+CstStart, Values.begin()+CstEnd, P); 173 174 // Ensure that integer constants are at the start of the constant pool. This 175 // is important so that GEP structure indices come before gep constant exprs. 176 std::partition(Values.begin()+CstStart, Values.begin()+CstEnd, 177 isIntegerValue); 178 179 // Rebuild the modified portion of ValueMap. 180 for (; CstStart != CstEnd; ++CstStart) 181 ValueMap[Values[CstStart].first] = CstStart+1; 182} 183 184 185/// EnumerateTypeSymbolTable - Insert all of the types in the specified symbol 186/// table. 187void ValueEnumerator::EnumerateTypeSymbolTable(const TypeSymbolTable &TST) { 188 for (TypeSymbolTable::const_iterator TI = TST.begin(), TE = TST.end(); 189 TI != TE; ++TI) 190 EnumerateType(TI->second); 191} 192 193/// EnumerateValueSymbolTable - Insert all of the values in the specified symbol 194/// table into the values table. 195void ValueEnumerator::EnumerateValueSymbolTable(const ValueSymbolTable &VST) { 196 for (ValueSymbolTable::const_iterator VI = VST.begin(), VE = VST.end(); 197 VI != VE; ++VI) 198 EnumerateValue(VI->getValue()); 199} 200 201void ValueEnumerator::EnumerateMetadata(const MetadataBase *MD) { 202 // Check to see if it's already in! 203 unsigned &MDValueID = MDValueMap[MD]; 204 if (MDValueID) { 205 // Increment use count. 206 MDValues[MDValueID-1].second++; 207 return; 208 } 209 210 // Enumerate the type of this value. 211 EnumerateType(MD->getType()); 212 213 if (const MDNode *N = dyn_cast<MDNode>(MD)) { 214 MDValues.push_back(std::make_pair(MD, 1U)); 215 MDValueMap[MD] = MDValues.size(); 216 MDValueID = MDValues.size(); 217 for (MDNode::const_elem_iterator I = N->elem_begin(), E = N->elem_end(); 218 I != E; ++I) { 219 if (*I) 220 EnumerateValue(*I); 221 else 222 EnumerateType(Type::getVoidTy(MD->getContext())); 223 } 224 return; 225 } else if (const NamedMDNode *N = dyn_cast<NamedMDNode>(MD)) { 226 for(NamedMDNode::const_elem_iterator I = N->elem_begin(), 227 E = N->elem_end(); I != E; ++I) { 228 MetadataBase *M = *I; 229 EnumerateValue(M); 230 } 231 MDValues.push_back(std::make_pair(MD, 1U)); 232 MDValueMap[MD] = Values.size(); 233 return; 234 } 235 236 // Add the value. 237 MDValues.push_back(std::make_pair(MD, 1U)); 238 MDValueID = MDValues.size(); 239} 240 241void ValueEnumerator::EnumerateValue(const Value *V) { 242 assert(V->getType() != Type::getVoidTy(V->getContext()) && 243 "Can't insert void values!"); 244 if (const MetadataBase *MB = dyn_cast<MetadataBase>(V)) 245 return EnumerateMetadata(MB); 246 247 // Check to see if it's already in! 248 unsigned &ValueID = ValueMap[V]; 249 if (ValueID) { 250 // Increment use count. 251 Values[ValueID-1].second++; 252 return; 253 } 254 255 // Enumerate the type of this value. 256 EnumerateType(V->getType()); 257 258 if (const Constant *C = dyn_cast<Constant>(V)) { 259 if (isa<GlobalValue>(C)) { 260 // Initializers for globals are handled explicitly elsewhere. 261 } else if (isa<ConstantArray>(C) && cast<ConstantArray>(C)->isString()) { 262 // Do not enumerate the initializers for an array of simple characters. 263 // The initializers just polute the value table, and we emit the strings 264 // specially. 265 } else if (C->getNumOperands()) { 266 // If a constant has operands, enumerate them. This makes sure that if a 267 // constant has uses (for example an array of const ints), that they are 268 // inserted also. 269 270 // We prefer to enumerate them with values before we enumerate the user 271 // itself. This makes it more likely that we can avoid forward references 272 // in the reader. We know that there can be no cycles in the constants 273 // graph that don't go through a global variable. 274 for (User::const_op_iterator I = C->op_begin(), E = C->op_end(); 275 I != E; ++I) 276 EnumerateValue(*I); 277 278 // Finally, add the value. Doing this could make the ValueID reference be 279 // dangling, don't reuse it. 280 Values.push_back(std::make_pair(V, 1U)); 281 ValueMap[V] = Values.size(); 282 return; 283 } 284 } 285 286 // Add the value. 287 Values.push_back(std::make_pair(V, 1U)); 288 ValueID = Values.size(); 289} 290 291 292void ValueEnumerator::EnumerateType(const Type *Ty) { 293 unsigned &TypeID = TypeMap[Ty]; 294 295 if (TypeID) { 296 // If we've already seen this type, just increase its occurrence count. 297 Types[TypeID-1].second++; 298 return; 299 } 300 301 // First time we saw this type, add it. 302 Types.push_back(std::make_pair(Ty, 1U)); 303 TypeID = Types.size(); 304 305 // Enumerate subtypes. 306 for (Type::subtype_iterator I = Ty->subtype_begin(), E = Ty->subtype_end(); 307 I != E; ++I) 308 EnumerateType(*I); 309} 310 311// Enumerate the types for the specified value. If the value is a constant, 312// walk through it, enumerating the types of the constant. 313void ValueEnumerator::EnumerateOperandType(const Value *V) { 314 EnumerateType(V->getType()); 315 if (const Constant *C = dyn_cast<Constant>(V)) { 316 // If this constant is already enumerated, ignore it, we know its type must 317 // be enumerated. 318 if (ValueMap.count(V)) return; 319 320 // This constant may have operands, make sure to enumerate the types in 321 // them. 322 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) 323 EnumerateOperandType(C->getOperand(i)); 324 325 if (const MDNode *N = dyn_cast<MDNode>(V)) { 326 for (unsigned i = 0, e = N->getNumElements(); i != e; ++i) { 327 Value *Elem = N->getElement(i); 328 if (Elem) 329 EnumerateOperandType(Elem); 330 } 331 } 332 } else if (isa<MDString>(V) || isa<MDNode>(V)) 333 EnumerateValue(V); 334} 335 336void ValueEnumerator::EnumerateAttributes(const AttrListPtr &PAL) { 337 if (PAL.isEmpty()) return; // null is always 0. 338 // Do a lookup. 339 unsigned &Entry = AttributeMap[PAL.getRawPointer()]; 340 if (Entry == 0) { 341 // Never saw this before, add it. 342 Attributes.push_back(PAL); 343 Entry = Attributes.size(); 344 } 345} 346 347 348void ValueEnumerator::incorporateFunction(const Function &F) { 349 NumModuleValues = Values.size(); 350 351 // Adding function arguments to the value table. 352 for(Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); 353 I != E; ++I) 354 EnumerateValue(I); 355 356 FirstFuncConstantID = Values.size(); 357 358 // Add all function-level constants to the value table. 359 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 360 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E; ++I) 361 for (User::const_op_iterator OI = I->op_begin(), E = I->op_end(); 362 OI != E; ++OI) { 363 if ((isa<Constant>(*OI) && !isa<GlobalValue>(*OI)) || 364 isa<InlineAsm>(*OI)) 365 EnumerateValue(*OI); 366 } 367 BasicBlocks.push_back(BB); 368 ValueMap[BB] = BasicBlocks.size(); 369 } 370 371 // Optimize the constant layout. 372 OptimizeConstants(FirstFuncConstantID, Values.size()); 373 374 // Add the function's parameter attributes so they are available for use in 375 // the function's instruction. 376 EnumerateAttributes(F.getAttributes()); 377 378 FirstInstID = Values.size(); 379 380 // Add all of the instructions. 381 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 382 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E; ++I) { 383 if (I->getType() != Type::getVoidTy(F.getContext())) 384 EnumerateValue(I); 385 } 386 } 387} 388 389void ValueEnumerator::purgeFunction() { 390 /// Remove purged values from the ValueMap. 391 for (unsigned i = NumModuleValues, e = Values.size(); i != e; ++i) 392 ValueMap.erase(Values[i].first); 393 for (unsigned i = 0, e = BasicBlocks.size(); i != e; ++i) 394 ValueMap.erase(BasicBlocks[i]); 395 396 Values.resize(NumModuleValues); 397 BasicBlocks.clear(); 398} 399