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 V.first->getType()->isIntegerTy();
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) {
42 // Enumerate the global variables.
43 for (Module::const_global_iterator I = M->global_begin(),
44 E = M->global_end(); I != E; ++I)
47 // Enumerate the functions.
48 for (Module::const_iterator I = M->begin(), E = M->end(); I != E; ++I) {
50 EnumerateAttributes(cast<Function>(I)->getAttributes());
53 // Enumerate the aliases.
54 for (Module::const_alias_iterator I = M->alias_begin(), E = M->alias_end();
58 // Remember what is the cutoff between globalvalue's and other constants.
59 unsigned FirstConstant = Values.size();
61 // Enumerate the global variable initializers.
62 for (Module::const_global_iterator I = M->global_begin(),
63 E = M->global_end(); I != E; ++I)
64 if (I->hasInitializer())
65 EnumerateValue(I->getInitializer());
67 // Enumerate the aliasees.
68 for (Module::const_alias_iterator I = M->alias_begin(), E = M->alias_end();
70 EnumerateValue(I->getAliasee());
72 // Enumerate types used by the type symbol table.
73 EnumerateTypeSymbolTable(M->getTypeSymbolTable());
75 // Insert constants and metadata that are named at module level into the slot
76 // pool so that the module symbol table can refer to them...
77 EnumerateValueSymbolTable(M->getValueSymbolTable());
78 EnumerateMDSymbolTable(M->getMDSymbolTable());
80 SmallVector<std::pair<unsigned, MDNode*>, 8> MDs;
82 // Enumerate types used by function bodies and argument lists.
83 for (Module::const_iterator F = M->begin(), E = M->end(); F != E; ++F) {
85 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
87 EnumerateType(I->getType());
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 if (MDNode *MD = dyn_cast<MDNode>(*OI))
94 if (MD->isFunctionLocal() && MD->getFunction())
95 // These will get enumerated during function-incorporation.
97 EnumerateOperandType(*OI);
99 EnumerateType(I->getType());
100 if (const CallInst *CI = dyn_cast<CallInst>(I))
101 EnumerateAttributes(CI->getAttributes());
102 else if (const InvokeInst *II = dyn_cast<InvokeInst>(I))
103 EnumerateAttributes(II->getAttributes());
105 // Enumerate metadata attached with this instruction.
107 I->getAllMetadata(MDs);
108 for (unsigned i = 0, e = MDs.size(); i != e; ++i)
109 EnumerateMetadata(MDs[i].second);
113 // Optimize constant ordering.
114 OptimizeConstants(FirstConstant, Values.size());
116 // Sort the type table by frequency so that most commonly used types are early
117 // in the table (have low bit-width).
118 std::stable_sort(Types.begin(), Types.end(), CompareByFrequency);
120 // Partition the Type ID's so that the single-value types occur before the
121 // aggregate types. This allows the aggregate types to be dropped from the
122 // type table after parsing the global variable initializers.
123 std::partition(Types.begin(), Types.end(), isSingleValueType);
125 // Now that we rearranged the type table, rebuild TypeMap.
126 for (unsigned i = 0, e = Types.size(); i != e; ++i)
127 TypeMap[Types[i].first] = i+1;
130 unsigned ValueEnumerator::getInstructionID(const Instruction *Inst) const {
131 InstructionMapType::const_iterator I = InstructionMap.find(Inst);
132 assert (I != InstructionMap.end() && "Instruction is not mapped!");
136 void ValueEnumerator::setInstructionID(const Instruction *I) {
137 InstructionMap[I] = InstructionCount++;
140 unsigned ValueEnumerator::getValueID(const Value *V) const {
141 if (isa<MDNode>(V) || isa<MDString>(V)) {
142 ValueMapType::const_iterator I = MDValueMap.find(V);
143 assert(I != MDValueMap.end() && "Value not in slotcalculator!");
147 ValueMapType::const_iterator I = ValueMap.find(V);
148 assert(I != ValueMap.end() && "Value not in slotcalculator!");
152 // Optimize constant ordering.
154 struct CstSortPredicate {
156 explicit CstSortPredicate(ValueEnumerator &ve) : VE(ve) {}
157 bool operator()(const std::pair<const Value*, unsigned> &LHS,
158 const std::pair<const Value*, unsigned> &RHS) {
160 if (LHS.first->getType() != RHS.first->getType())
161 return VE.getTypeID(LHS.first->getType()) <
162 VE.getTypeID(RHS.first->getType());
163 // Then by frequency.
164 return LHS.second > RHS.second;
169 /// OptimizeConstants - Reorder constant pool for denser encoding.
170 void ValueEnumerator::OptimizeConstants(unsigned CstStart, unsigned CstEnd) {
171 if (CstStart == CstEnd || CstStart+1 == CstEnd) return;
173 CstSortPredicate P(*this);
174 std::stable_sort(Values.begin()+CstStart, Values.begin()+CstEnd, P);
176 // Ensure that integer constants are at the start of the constant pool. This
177 // is important so that GEP structure indices come before gep constant exprs.
178 std::partition(Values.begin()+CstStart, Values.begin()+CstEnd,
181 // Rebuild the modified portion of ValueMap.
182 for (; CstStart != CstEnd; ++CstStart)
183 ValueMap[Values[CstStart].first] = CstStart+1;
187 /// EnumerateTypeSymbolTable - Insert all of the types in the specified symbol
189 void ValueEnumerator::EnumerateTypeSymbolTable(const TypeSymbolTable &TST) {
190 for (TypeSymbolTable::const_iterator TI = TST.begin(), TE = TST.end();
192 EnumerateType(TI->second);
195 /// EnumerateValueSymbolTable - Insert all of the values in the specified symbol
196 /// table into the values table.
197 void ValueEnumerator::EnumerateValueSymbolTable(const ValueSymbolTable &VST) {
198 for (ValueSymbolTable::const_iterator VI = VST.begin(), VE = VST.end();
200 EnumerateValue(VI->getValue());
203 /// EnumerateMDSymbolTable - Insert all of the values in the specified metadata
205 void ValueEnumerator::EnumerateMDSymbolTable(const MDSymbolTable &MST) {
206 for (MDSymbolTable::const_iterator MI = MST.begin(), ME = MST.end();
208 EnumerateValue(MI->getValue());
211 void ValueEnumerator::EnumerateNamedMDNode(const NamedMDNode *MD) {
212 // Check to see if it's already in!
213 unsigned &MDValueID = MDValueMap[MD];
215 // Increment use count.
216 MDValues[MDValueID-1].second++;
220 // Enumerate the type of this value.
221 EnumerateType(MD->getType());
223 for (unsigned i = 0, e = MD->getNumOperands(); i != e; ++i)
224 if (MDNode *E = MD->getOperand(i))
226 MDValues.push_back(std::make_pair(MD, 1U));
227 MDValueMap[MD] = Values.size();
230 void ValueEnumerator::EnumerateMetadata(const Value *MD) {
231 assert((isa<MDNode>(MD) || isa<MDString>(MD)) && "Invalid metadata kind");
232 // Check to see if it's already in!
233 unsigned &MDValueID = MDValueMap[MD];
235 // Increment use count.
236 MDValues[MDValueID-1].second++;
240 // Enumerate the type of this value.
241 EnumerateType(MD->getType());
243 if (const MDNode *N = dyn_cast<MDNode>(MD)) {
244 MDValues.push_back(std::make_pair(MD, 1U));
245 MDValueMap[MD] = MDValues.size();
246 MDValueID = MDValues.size();
247 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
248 if (Value *V = N->getOperand(i))
251 EnumerateType(Type::getVoidTy(MD->getContext()));
257 assert(isa<MDString>(MD) && "Unknown metadata kind");
258 MDValues.push_back(std::make_pair(MD, 1U));
259 MDValueID = MDValues.size();
262 void ValueEnumerator::EnumerateValue(const Value *V) {
263 assert(!V->getType()->isVoidTy() && "Can't insert void values!");
264 if (isa<MDNode>(V) || isa<MDString>(V))
265 return EnumerateMetadata(V);
266 else if (const NamedMDNode *NMD = dyn_cast<NamedMDNode>(V))
267 return EnumerateNamedMDNode(NMD);
269 // Check to see if it's already in!
270 unsigned &ValueID = ValueMap[V];
272 // Increment use count.
273 Values[ValueID-1].second++;
277 // Enumerate the type of this value.
278 EnumerateType(V->getType());
280 if (const Constant *C = dyn_cast<Constant>(V)) {
281 if (isa<GlobalValue>(C)) {
282 // Initializers for globals are handled explicitly elsewhere.
283 } else if (isa<ConstantArray>(C) && cast<ConstantArray>(C)->isString()) {
284 // Do not enumerate the initializers for an array of simple characters.
285 // The initializers just polute the value table, and we emit the strings
287 } else if (C->getNumOperands()) {
288 // If a constant has operands, enumerate them. This makes sure that if a
289 // constant has uses (for example an array of const ints), that they are
292 // We prefer to enumerate them with values before we enumerate the user
293 // itself. This makes it more likely that we can avoid forward references
294 // in the reader. We know that there can be no cycles in the constants
295 // graph that don't go through a global variable.
296 for (User::const_op_iterator I = C->op_begin(), E = C->op_end();
298 if (!isa<BasicBlock>(*I)) // Don't enumerate BB operand to BlockAddress.
301 // Finally, add the value. Doing this could make the ValueID reference be
302 // dangling, don't reuse it.
303 Values.push_back(std::make_pair(V, 1U));
304 ValueMap[V] = Values.size();
310 Values.push_back(std::make_pair(V, 1U));
311 ValueID = Values.size();
315 void ValueEnumerator::EnumerateType(const Type *Ty) {
316 unsigned &TypeID = TypeMap[Ty];
319 // If we've already seen this type, just increase its occurrence count.
320 Types[TypeID-1].second++;
324 // First time we saw this type, add it.
325 Types.push_back(std::make_pair(Ty, 1U));
326 TypeID = Types.size();
328 // Enumerate subtypes.
329 for (Type::subtype_iterator I = Ty->subtype_begin(), E = Ty->subtype_end();
334 // Enumerate the types for the specified value. If the value is a constant,
335 // walk through it, enumerating the types of the constant.
336 void ValueEnumerator::EnumerateOperandType(const Value *V) {
337 EnumerateType(V->getType());
339 if (const Constant *C = dyn_cast<Constant>(V)) {
340 // If this constant is already enumerated, ignore it, we know its type must
342 if (ValueMap.count(V)) return;
344 // This constant may have operands, make sure to enumerate the types in
346 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) {
347 const User *Op = C->getOperand(i);
349 // Don't enumerate basic blocks here, this happens as operands to
351 if (isa<BasicBlock>(Op)) continue;
353 EnumerateOperandType(cast<Constant>(Op));
356 if (const MDNode *N = dyn_cast<MDNode>(V)) {
357 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
358 if (Value *Elem = N->getOperand(i))
359 EnumerateOperandType(Elem);
361 } else if (isa<MDString>(V) || isa<MDNode>(V))
365 void ValueEnumerator::EnumerateAttributes(const AttrListPtr &PAL) {
366 if (PAL.isEmpty()) return; // null is always 0.
368 unsigned &Entry = AttributeMap[PAL.getRawPointer()];
370 // Never saw this before, add it.
371 Attributes.push_back(PAL);
372 Entry = Attributes.size();
377 void ValueEnumerator::incorporateFunction(const Function &F) {
378 InstructionCount = 0;
379 NumModuleValues = Values.size();
381 // Adding function arguments to the value table.
382 for(Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
386 FirstFuncConstantID = Values.size();
388 // Add all function-level constants to the value table.
389 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
390 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E; ++I)
391 for (User::const_op_iterator OI = I->op_begin(), E = I->op_end();
393 if ((isa<Constant>(*OI) && !isa<GlobalValue>(*OI)) ||
397 BasicBlocks.push_back(BB);
398 ValueMap[BB] = BasicBlocks.size();
401 // Optimize the constant layout.
402 OptimizeConstants(FirstFuncConstantID, Values.size());
404 // Add the function's parameter attributes so they are available for use in
405 // the function's instruction.
406 EnumerateAttributes(F.getAttributes());
408 FirstInstID = Values.size();
410 SmallVector<MDNode *, 8> FunctionLocalMDs;
411 // Add all of the instructions.
412 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
413 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E; ++I) {
414 for (User::const_op_iterator OI = I->op_begin(), E = I->op_end();
416 if (MDNode *MD = dyn_cast<MDNode>(*OI))
417 if (MD->isFunctionLocal() && MD->getFunction())
418 // Enumerate metadata after the instructions they might refer to.
419 FunctionLocalMDs.push_back(MD);
421 if (!I->getType()->isVoidTy())
426 // Add all of the function-local metadata.
427 for (unsigned i = 0, e = FunctionLocalMDs.size(); i != e; ++i)
428 EnumerateOperandType(FunctionLocalMDs[i]);
431 void ValueEnumerator::purgeFunction() {
432 /// Remove purged values from the ValueMap.
433 for (unsigned i = NumModuleValues, e = Values.size(); i != e; ++i)
434 ValueMap.erase(Values[i].first);
435 for (unsigned i = 0, e = BasicBlocks.size(); i != e; ++i)
436 ValueMap.erase(BasicBlocks[i]);
438 Values.resize(NumModuleValues);
442 static void IncorporateFunctionInfoGlobalBBIDs(const Function *F,
443 DenseMap<const BasicBlock*, unsigned> &IDMap) {
444 unsigned Counter = 0;
445 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
446 IDMap[BB] = ++Counter;
449 /// getGlobalBasicBlockID - This returns the function-specific ID for the
450 /// specified basic block. This is relatively expensive information, so it
451 /// should only be used by rare constructs such as address-of-label.
452 unsigned ValueEnumerator::getGlobalBasicBlockID(const BasicBlock *BB) const {
453 unsigned &Idx = GlobalBasicBlockIDs[BB];
457 IncorporateFunctionInfoGlobalBBIDs(BB->getParent(), GlobalBasicBlockIDs);
458 return getGlobalBasicBlockID(BB);