1 //===-- SlotCalculator.cpp - Calculate what slots values land in ----------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements a useful analysis step to figure out what numbered
11 // slots values in a program will land in (keeping track of per plane
12 // information as required.
14 // This is used primarily for when writing a file to disk, either in bytecode
17 //===----------------------------------------------------------------------===//
19 #include "llvm/SlotCalculator.h"
20 #include "llvm/Analysis/ConstantsScanner.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/iOther.h"
24 #include "llvm/Module.h"
25 #include "llvm/SymbolTable.h"
26 #include "Support/PostOrderIterator.h"
27 #include "Support/STLExtras.h"
32 #define SC_DEBUG(X) std::cerr << X
37 SlotCalculator::SlotCalculator(const Module *M, bool buildBytecodeInfo) {
38 BuildBytecodeInfo = buildBytecodeInfo;
41 // Preload table... Make sure that all of the primitive types are in the table
42 // and that their Primitive ID is equal to their slot #
44 SC_DEBUG("Inserting primitive types:\n");
45 for (unsigned i = 0; i < Type::FirstDerivedTyID; ++i) {
46 assert(Type::getPrimitiveType((Type::PrimitiveID)i));
47 insertValue(Type::getPrimitiveType((Type::PrimitiveID)i), true);
50 if (M == 0) return; // Empty table...
54 SlotCalculator::SlotCalculator(const Function *M, bool buildBytecodeInfo) {
55 BuildBytecodeInfo = buildBytecodeInfo;
56 TheModule = M ? M->getParent() : 0;
58 // Preload table... Make sure that all of the primitive types are in the table
59 // and that their Primitive ID is equal to their slot #
61 SC_DEBUG("Inserting primitive types:\n");
62 for (unsigned i = 0; i < Type::FirstDerivedTyID; ++i) {
63 assert(Type::getPrimitiveType((Type::PrimitiveID)i));
64 insertValue(Type::getPrimitiveType((Type::PrimitiveID)i), true);
67 if (TheModule == 0) return; // Empty table...
69 processModule(); // Process module level stuff
70 incorporateFunction(M); // Start out in incorporated state
74 // processModule - Process all of the module level function declarations and
75 // types that are available.
77 void SlotCalculator::processModule() {
78 SC_DEBUG("begin processModule!\n");
80 // Add all of the global variables to the value table...
82 for (Module::const_giterator I = TheModule->gbegin(), E = TheModule->gend();
86 // Scavenge the types out of the functions, then add the functions themselves
87 // to the value table...
89 for (Module::const_iterator I = TheModule->begin(), E = TheModule->end();
93 // Add all of the module level constants used as initializers
95 for (Module::const_giterator I = TheModule->gbegin(), E = TheModule->gend();
97 if (I->hasInitializer())
98 getOrCreateSlot(I->getInitializer());
100 // Now that all global constants have been added, rearrange constant planes
101 // that contain constant strings so that the strings occur at the start of the
102 // plane, not somewhere in the middle.
104 if (BuildBytecodeInfo) {
105 TypePlane &Types = Table[Type::TypeTyID];
106 for (unsigned plane = 0, e = Table.size(); plane != e; ++plane) {
107 if (const ArrayType *AT = dyn_cast<ArrayType>(Types[plane]))
108 if (AT->getElementType() == Type::SByteTy ||
109 AT->getElementType() == Type::UByteTy) {
110 TypePlane &Plane = Table[plane];
111 unsigned FirstNonStringID = 0;
112 for (unsigned i = 0, e = Plane.size(); i != e; ++i)
113 if (cast<ConstantArray>(Plane[i])->isString()) {
114 // Check to see if we have to shuffle this string around. If not,
115 // don't do anything.
116 if (i != FirstNonStringID) {
117 // Swap the plane entries....
118 std::swap(Plane[i], Plane[FirstNonStringID]);
120 // Keep the NodeMap up to date.
121 NodeMap[Plane[i]] = i;
122 NodeMap[Plane[FirstNonStringID]] = FirstNonStringID;
131 // FIXME: Empirically, this causes the bytecode files to get BIGGER, because
132 // it explodes the operand size numbers to be bigger than can be handled
133 // compactly, which offsets the ~40% savings in constant sizes. Whoops.
135 // If we are emitting a bytecode file, scan all of the functions for their
136 // constants, which allows us to emit more compact modules. This is optional,
137 // and is just used to compactify the constants used by different functions
139 if (BuildBytecodeInfo) {
140 SC_DEBUG("Inserting function constants:\n");
141 for (Module::const_iterator F = TheModule->begin(), E = TheModule->end();
143 for_each(constant_begin(F), constant_end(F),
144 bind_obj(this, &SlotCalculator::getOrCreateSlot));
148 // Insert constants that are named at module level into the slot pool so that
149 // the module symbol table can refer to them...
151 if (BuildBytecodeInfo) {
152 SC_DEBUG("Inserting SymbolTable values:\n");
153 processSymbolTable(&TheModule->getSymbolTable());
156 // Now that we have collected together all of the information relevant to the
157 // module, compactify the type table if it is particularly big and outputting
158 // a bytecode file. The basic problem we run into is that some programs have
159 // a large number of types, which causes the type field to overflow its size,
160 // which causes instructions to explode in size (particularly call
161 // instructions). To avoid this behavior, we "sort" the type table so that
162 // all non-value types are pushed to the end of the type table, giving nice
163 // low numbers to the types that can be used by instructions, thus reducing
164 // the amount of explodage we suffer.
165 if (BuildBytecodeInfo && Table[Type::TypeTyID].size() >= 64) {
166 // Scan through the type table moving value types to the start of the table.
167 TypePlane *Types = &Table[Type::TypeTyID];
168 unsigned FirstNonValueTypeID = 0;
169 for (unsigned i = 0, e = Types->size(); i != e; ++i)
170 if (cast<Type>((*Types)[i])->isFirstClassType() ||
171 cast<Type>((*Types)[i])->isPrimitiveType()) {
172 // Check to see if we have to shuffle this type around. If not, don't
174 if (i != FirstNonValueTypeID) {
175 assert(i != Type::TypeTyID && FirstNonValueTypeID != Type::TypeTyID &&
176 "Cannot move around the type plane!");
178 // Swap the type ID's.
179 std::swap((*Types)[i], (*Types)[FirstNonValueTypeID]);
181 // Keep the NodeMap up to date.
182 NodeMap[(*Types)[i]] = i;
183 NodeMap[(*Types)[FirstNonValueTypeID]] = FirstNonValueTypeID;
185 // When we move a type, make sure to move its value plane as needed.
186 if (Table.size() > FirstNonValueTypeID) {
187 if (Table.size() <= i) Table.resize(i+1);
188 std::swap(Table[i], Table[FirstNonValueTypeID]);
189 Types = &Table[Type::TypeTyID];
192 ++FirstNonValueTypeID;
196 SC_DEBUG("end processModule!\n");
199 // processSymbolTable - Insert all of the values in the specified symbol table
200 // into the values table...
202 void SlotCalculator::processSymbolTable(const SymbolTable *ST) {
203 for (SymbolTable::const_iterator I = ST->begin(), E = ST->end(); I != E; ++I)
204 for (SymbolTable::type_const_iterator TI = I->second.begin(),
205 TE = I->second.end(); TI != TE; ++TI)
206 getOrCreateSlot(TI->second);
209 void SlotCalculator::processSymbolTableConstants(const SymbolTable *ST) {
210 for (SymbolTable::const_iterator I = ST->begin(), E = ST->end(); I != E; ++I)
211 for (SymbolTable::type_const_iterator TI = I->second.begin(),
212 TE = I->second.end(); TI != TE; ++TI)
213 if (isa<Constant>(TI->second) || isa<Type>(TI->second))
214 getOrCreateSlot(TI->second);
218 void SlotCalculator::incorporateFunction(const Function *F) {
219 assert(ModuleLevel.size() == 0 && "Module already incorporated!");
221 SC_DEBUG("begin processFunction!\n");
223 // Save the Table state before we process the function...
224 for (unsigned i = 0; i < Table.size(); ++i)
225 ModuleLevel.push_back(Table[i].size());
227 SC_DEBUG("Inserting function arguments\n");
229 // Iterate over function arguments, adding them to the value table...
230 for(Function::const_aiterator I = F->abegin(), E = F->aend(); I != E; ++I)
233 // Iterate over all of the instructions in the function, looking for constant
234 // values that are referenced. Add these to the value pools before any
235 // nonconstant values. This will be turned into the constant pool for the
238 if (BuildBytecodeInfo) { // Assembly writer does not need this!
239 // Emit all of the constants that are being used by the instructions in the
241 for_each(constant_begin(F), constant_end(F),
242 bind_obj(this, &SlotCalculator::getOrCreateSlot));
244 // If there is a symbol table, it is possible that the user has names for
245 // constants that are not being used. In this case, we will have problems
246 // if we don't emit the constants now, because otherwise we will get
247 // symbol table references to constants not in the output. Scan for these
250 processSymbolTableConstants(&F->getSymbolTable());
253 SC_DEBUG("Inserting Instructions:\n");
255 // Add all of the instructions to the type planes...
256 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
258 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E; ++I) {
260 if (const VANextInst *VAN = dyn_cast<VANextInst>(I))
261 getOrCreateSlot(VAN->getArgType());
265 SC_DEBUG("end processFunction!\n");
268 void SlotCalculator::purgeFunction() {
269 assert(ModuleLevel.size() != 0 && "Module not incorporated!");
270 unsigned NumModuleTypes = ModuleLevel.size();
272 SC_DEBUG("begin purgeFunction!\n");
274 // First, remove values from existing type planes
275 for (unsigned i = 0; i < NumModuleTypes; ++i) {
276 unsigned ModuleSize = ModuleLevel[i]; // Size of plane before function came
277 TypePlane &CurPlane = Table[i];
278 //SC_DEBUG("Processing Plane " <<i<< " of size " << CurPlane.size() <<"\n");
280 while (CurPlane.size() != ModuleSize) {
281 //SC_DEBUG(" Removing [" << i << "] Value=" << CurPlane.back() << "\n");
282 std::map<const Value *, unsigned>::iterator NI =
283 NodeMap.find(CurPlane.back());
284 assert(NI != NodeMap.end() && "Node not in nodemap?");
285 NodeMap.erase(NI); // Erase from nodemap
286 CurPlane.pop_back(); // Shrink plane
290 // We don't need this state anymore, free it up.
293 // Next, remove any type planes defined by the function...
294 while (NumModuleTypes != Table.size()) {
295 TypePlane &Plane = Table.back();
296 SC_DEBUG("Removing Plane " << (Table.size()-1) << " of size "
297 << Plane.size() << "\n");
298 while (Plane.size()) {
299 NodeMap.erase(NodeMap.find(Plane.back())); // Erase from nodemap
300 Plane.pop_back(); // Shrink plane
303 Table.pop_back(); // Nuke the plane, we don't like it.
306 SC_DEBUG("end purgeFunction!\n");
309 int SlotCalculator::getSlot(const Value *V) const {
310 std::map<const Value*, unsigned>::const_iterator I = NodeMap.find(V);
311 if (I != NodeMap.end())
312 return (int)I->second;
314 // Do not number ConstantPointerRef's at all. They are an abomination.
315 if (const ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(V))
316 return getSlot(CPR->getValue());
322 int SlotCalculator::getOrCreateSlot(const Value *V) {
323 int SlotNo = getSlot(V); // Check to see if it's already in!
324 if (SlotNo != -1) return SlotNo;
326 // Do not number ConstantPointerRef's at all. They are an abomination.
327 if (const ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(V))
328 return getOrCreateSlot(CPR->getValue());
330 if (!isa<GlobalValue>(V))
331 if (const Constant *C = dyn_cast<Constant>(V)) {
332 // If we are emitting a bytecode file, do not index the characters that
333 // make up constant strings. We emit constant strings as special
334 // entities that don't require their individual characters to be emitted.
335 if (!BuildBytecodeInfo || !isa<ConstantArray>(C) ||
336 !cast<ConstantArray>(C)->isString()) {
337 // This makes sure that if a constant has uses (for example an array of
338 // const ints), that they are inserted also.
340 for (User::const_op_iterator I = C->op_begin(), E = C->op_end();
344 assert(ModuleLevel.empty() &&
345 "How can a constant string be directly accessed in a function?");
346 // Otherwise, if we are emitting a bytecode file and this IS a string,
348 if (!C->isNullValue())
349 ConstantStrings.push_back(cast<ConstantArray>(C));
353 return insertValue(V);
357 int SlotCalculator::insertValue(const Value *D, bool dontIgnore) {
358 assert(D && "Can't insert a null value!");
359 assert(getSlot(D) == -1 && "Value is already in the table!");
361 // If this node does not contribute to a plane, or if the node has a
362 // name and we don't want names, then ignore the silly node... Note that types
363 // do need slot numbers so that we can keep track of where other values land.
365 if (!dontIgnore) // Don't ignore nonignorables!
366 if (D->getType() == Type::VoidTy || // Ignore void type nodes
367 (!BuildBytecodeInfo && // Ignore named and constants
368 (D->hasName() || isa<Constant>(D)) && !isa<Type>(D))) {
369 SC_DEBUG("ignored value " << *D << "\n");
370 return -1; // We do need types unconditionally though
373 // If it's a type, make sure that all subtypes of the type are included...
374 if (const Type *TheTy = dyn_cast<Type>(D)) {
376 // Insert the current type before any subtypes. This is important because
377 // recursive types elements are inserted in a bottom up order. Changing
378 // this here can break things. For example:
380 // global { \2 * } { { \2 }* null }
382 int ResultSlot = doInsertValue(TheTy);
383 SC_DEBUG(" Inserted type: " << TheTy->getDescription() << " slot=" <<
386 // Loop over any contained types in the definition... in post
389 for (po_iterator<const Type*> I = po_begin(TheTy), E = po_end(TheTy);
392 const Type *SubTy = *I;
393 // If we haven't seen this sub type before, add it to our type table!
394 if (getSlot(SubTy) == -1) {
395 SC_DEBUG(" Inserting subtype: " << SubTy->getDescription() << "\n");
396 int Slot = doInsertValue(SubTy);
397 SC_DEBUG(" Inserted subtype: " << SubTy->getDescription() <<
398 " slot=" << Slot << "\n");
405 // Okay, everything is happy, actually insert the silly value now...
406 return doInsertValue(D);
409 static inline bool hasNullValue(unsigned TyID) {
410 return TyID != Type::LabelTyID && TyID != Type::TypeTyID &&
411 TyID != Type::VoidTyID;
414 // doInsertValue - This is a small helper function to be called only
417 int SlotCalculator::doInsertValue(const Value *D) {
418 const Type *Typ = D->getType();
421 // Used for debugging DefSlot=-1 assertion...
422 //if (Typ == Type::TypeTy)
423 // cerr << "Inserting type '" << cast<Type>(D)->getDescription() << "'!\n";
425 if (Typ->isDerivedType()) {
426 int ValSlot = getSlot(Typ);
427 if (ValSlot == -1) { // Have we already entered this type?
428 // Nope, this is the first we have seen the type, process it.
429 ValSlot = insertValue(Typ, true);
430 assert(ValSlot != -1 && "ProcessType returned -1 for a type?");
432 Ty = (unsigned)ValSlot;
434 Ty = Typ->getPrimitiveID();
437 if (Table.size() <= Ty) // Make sure we have the type plane allocated...
438 Table.resize(Ty+1, TypePlane());
440 // If this is the first value to get inserted into the type plane, make sure
441 // to insert the implicit null value...
442 if (Table[Ty].empty() && BuildBytecodeInfo && hasNullValue(Ty)) {
443 Value *ZeroInitializer = Constant::getNullValue(Typ);
445 // If we are pushing zeroinit, it will be handled below.
446 if (D != ZeroInitializer) {
447 Table[Ty].push_back(ZeroInitializer);
448 NodeMap[ZeroInitializer] = 0;
452 // Insert node into table and NodeMap...
453 unsigned DestSlot = NodeMap[D] = Table[Ty].size();
454 Table[Ty].push_back(D);
456 SC_DEBUG(" Inserting value [" << Ty << "] = " << D << " slot=" <<
458 // G = Global, C = Constant, T = Type, F = Function, o = other
459 SC_DEBUG((isa<GlobalVariable>(D) ? "G" : (isa<Constant>(D) ? "C" :
460 (isa<Type>(D) ? "T" : (isa<Function>(D) ? "F" : "o")))));
462 return (int)DestSlot;