[thirdparty/binutils-gdb.git] / gold / symtab.cc

// symtab.cc -- the gold symbol table

#include "gold.h"

#include <cassert>
#include <stdint.h>
#include <string>
#include <utility>

#include "object.h"
#include "symtab.h"

namespace gold
{

// Class Symbol.

// Initialize the fields in the base class Symbol.

template<int size, bool big_endian>
void
Symbol::init_base(const char* name, const char* version, Object* object,
		  const elfcpp::Sym<size, big_endian>& sym)
{
  this->name_ = name;
  this->version_ = version;
  this->object_ = object;
  this->shnum_ = sym.get_st_shndx(); // FIXME: Handle SHN_XINDEX.
  this->type_ = sym.get_st_type();
  this->binding_ = sym.get_st_bind();
  this->visibility_ = sym.get_st_visibility();
  this->other_ = sym.get_st_nonvis();
  this->is_special_ = false;
  this->is_def_ = false;
  this->is_forwarder_ = false;
  this->in_dyn_ = object->is_dynamic();
}

// Initialize the fields in Sized_symbol.

template<int size>
template<bool big_endian>
void
Sized_symbol<size>::init(const char* name, const char* version, Object* object,
			 const elfcpp::Sym<size, big_endian>& sym)
{
  this->init_base(name, version, object, sym);
  this->value_ = sym.get_st_value();
  this->size_ = sym.get_st_size();
}

// Class Symbol_table.

Symbol_table::Symbol_table()
  : size_(0), table_(), namepool_(), forwarders_()
{
}

Symbol_table::~Symbol_table()
{
}

// The hash function.  The key is always canonicalized, so we use a
// simple combination of the pointers.

size_t
Symbol_table::Symbol_table_hash::operator()(const Symbol_table_key& key) const
{
  return (reinterpret_cast<size_t>(key.first)
	  ^ reinterpret_cast<size_t>(key.second));
}

// The symbol table key equality function.  This is only called with
// canonicalized name and version strings, so we can use pointer
// comparison.

bool
Symbol_table::Symbol_table_eq::operator()(const Symbol_table_key& k1,
					  const Symbol_table_key& k2) const
{
  return k1.first == k2.first && k1.second == k2.second;
}

// Make TO a symbol which forwards to FROM.  

void
Symbol_table::make_forwarder(Symbol* from, Symbol* to)
{
  assert(!from->is_forwarder() && !to->is_forwarder());
  this->forwarders_[from] = to;
  from->set_forwarder();
}

Symbol*
Symbol_table::resolve_forwards(Symbol* from) const
{
  assert(from->is_forwarder());
  Unordered_map<Symbol*, Symbol*>::const_iterator p =
    this->forwarders_.find(from);
  assert(p != this->forwarders_.end());
  return p->second;
}

// Resolve a Symbol with another Symbol.  This is only used in the
// unusual case where there are references to both an unversioned
// symbol and a symbol with a version, and we then discover that that
// version is the default version.  Because this is unusual, we do
// this the slow way, by converting back to an ELF symbol.

template<int size, bool big_endian>
void
Symbol_table::resolve(Sized_symbol<size>* to, const Sized_symbol<size>* from)
{
  unsigned char buf[elfcpp::Elf_sizes<size>::sym_size];
  elfcpp::Sym_write<size, big_endian> esym(buf);
  // We don't bother to set the st_name field.
  esym.put_st_value(from->value());
  esym.put_st_size(from->symsize());
  esym.put_st_info(from->binding(), from->type());
  esym.put_st_other(from->visibility(), from->other());
  esym.put_st_shndx(from->shnum());
  Symbol_table::resolve(to, esym.sym(), from->object());
}

// Add one symbol from OBJECT to the symbol table.  NAME is symbol
// name and VERSION is the version; both are canonicalized.  DEF is
// whether this is the default version.

// If DEF is true, then this is the definition of a default version of
// a symbol.  That means that any lookup of NAME/NULL and any lookup
// of NAME/VERSION should always return the same symbol.  This is
// obvious for references, but in particular we want to do this for
// definitions: overriding NAME/NULL should also override
// NAME/VERSION.  If we don't do that, it would be very hard to
// override functions in a shared library which uses versioning.

// We implement this by simply making both entries in the hash table
// point to the same Symbol structure.  That is easy enough if this is
// the first time we see NAME/NULL or NAME/VERSION, but it is possible
// that we have seen both already, in which case they will both have
// independent entries in the symbol table.  We can't simply change
// the symbol table entry, because we have pointers to the entries
// attached to the object files.  So we mark the entry attached to the
// object file as a forwarder, and record it in the forwarders_ map.
// Note that entries in the hash table will never be marked as
// forwarders.

template<int size, bool big_endian>
Symbol*
Symbol_table::add_from_object(Sized_object<size, big_endian>* object,
			      const char *name,
			      const char *version, bool def,
			      const elfcpp::Sym<size, big_endian>& sym)
{
  Symbol* const snull = NULL;
  std::pair<typename Symbol_table_type::iterator, bool> ins =
    this->table_.insert(std::make_pair(std::make_pair(name, version), snull));

  std::pair<typename Symbol_table_type::iterator, bool> insdef =
    std::make_pair(this->table_.end(), false);
  if (def)
    {
      const char* const vnull = NULL;
      insdef = this->table_.insert(std::make_pair(std::make_pair(name, vnull),
						  snull));
    }

  // ins.first: an iterator, which is a pointer to a pair.
  // ins.first->first: the key (a pair of name and version).
  // ins.first->second: the value (Symbol*).
  // ins.second: true if new entry was inserted, false if not.

  Sized_symbol<size>* ret;
  if (!ins.second)
    {
      // We already have an entry for NAME/VERSION.
      ret = this->get_sized_symbol<size>(ins.first->second);
      assert(ret != NULL);
      Symbol_table::resolve(ret, sym, object);

      if (def)
	{
	  if (insdef.second)
	    {
	      // This is the first time we have seen NAME/NULL.  Make
	      // NAME/NULL point to NAME/VERSION.
	      insdef.first->second = ret;
	    }
	  else
	    {
	      // This is the unfortunate case where we already have
	      // entries for both NAME/VERSION and NAME/NULL.
	      const Sized_symbol<size>* sym2 =
		this->get_sized_symbol<size>(insdef.first->second);
	      Symbol_table::resolve<size, big_endian>(ret, sym2);
	      this->make_forwarder(insdef.first->second, ret);
	      insdef.first->second = ret;
	    }
	}
    }
  else
    {
      // This is the first time we have seen NAME/VERSION.
      assert(ins.first->second == NULL);
      if (def && !insdef.second)
	{
	  // We already have an entry for NAME/NULL.  Make
	  // NAME/VERSION point to it.
	  ret = this->get_sized_symbol<size>(insdef.first->second);
	  Symbol_table::resolve(ret, sym, object);
	  ins.first->second = ret;
	}
      else
	{
	  Sized_target<size, big_endian>* target = object->sized_target();
	  if (!target->has_make_symbol())
	    ret = new Sized_symbol<size>();
	  else
	    {
	      ret = target->make_symbol();
	      if (ret == NULL)
		{
		  // This means that we don't want a symbol table
		  // entry after all.
		  if (!def)
		    this->table_.erase(ins.first);
		  else
		    {
		      this->table_.erase(insdef.first);
		      // Inserting insdef invalidated ins.
		      this->table_.erase(std::make_pair(name, version));
		    }
		  return NULL;
		}
	    }

	  ret->init(name, version, object, sym);

	  ins.first->second = ret;
	  if (def)
	    {
	      // This is the first time we have seen NAME/NULL.  Point
	      // it at the new entry for NAME/VERSION.
	      assert(insdef.second);
	      insdef.first->second = ret;
	    }
	}
    }

  return ret;
}

// Add all the symbols in an object to the hash table.

template<int size, bool big_endian>
void
Symbol_table::add_from_object(
    Sized_object<size, big_endian>* object,
    const elfcpp::Sym<size, big_endian>* syms,
    size_t count,
    const char* sym_names,
    size_t sym_name_size,
    Symbol** sympointers)
{
  // We take the size from the first object we see.
  if (this->get_size() == 0)
    this->set_size(size);

  if (size != this->get_size() || size != object->target()->get_size())
    {
      fprintf(stderr, _("%s: %s: mixing 32-bit and 64-bit ELF objects\n"),
	      program_name, object->name().c_str());
      gold_exit(false);
    }

  const unsigned char* p = reinterpret_cast<const unsigned char*>(syms);
  for (size_t i = 0; i < count; ++i)
    {
      elfcpp::Sym<size, big_endian> sym(p);

      unsigned int st_name = sym.get_st_name();
      if (st_name >= sym_name_size)
	{
	  fprintf(stderr, _("%s: %s: bad symbol name offset %u at %lu\n"),
		  program_name, object->name().c_str(), st_name,
		  static_cast<unsigned long>(i));
	  gold_exit(false);
	}

      const char* name = sym_names + st_name;

      // In an object file, an '@' in the name separates the symbol
      // name from the version name.  If there are two '@' characters,
      // this is the default version.
      const char* ver = strchr(name, '@');

      Symbol* res;
      if (ver == NULL)
	{
	  name = this->namepool_.add(name);
	  res = this->add_from_object(object, name, NULL, false, sym);
	}
      else
	{
	  name = this->namepool_.add(name, ver - name);
	  bool def = false;
	  ++ver;
	  if (*ver == '@')
	    {
	      def = true;
	      ++ver;
	    }
	  ver = this->namepool_.add(ver);
	  res = this->add_from_object(object, name, ver, def, sym);
	}

      *sympointers++ = res;

      p += elfcpp::Elf_sizes<size>::sym_size;
    }
}

// Instantiate the templates we need.  We could use the configure
// script to restrict this to only the ones needed for implemented
// targets.

template
void
Symbol_table::add_from_object<32, true>(
    Sized_object<32, true>* object,
    const elfcpp::Sym<32, true>* syms,
    size_t count,
    const char* sym_names,
    size_t sym_name_size,
    Symbol** sympointers);

template
void
Symbol_table::add_from_object<32, false>(
    Sized_object<32, false>* object,
    const elfcpp::Sym<32, false>* syms,
    size_t count,
    const char* sym_names,
    size_t sym_name_size,
    Symbol** sympointers);

template
void
Symbol_table::add_from_object<64, true>(
    Sized_object<64, true>* object,
    const elfcpp::Sym<64, true>* syms,
    size_t count,
    const char* sym_names,
    size_t sym_name_size,
    Symbol** sympointers);

template
void
Symbol_table::add_from_object<64, false>(
    Sized_object<64, false>* object,
    const elfcpp::Sym<64, false>* syms,
    size_t count,
    const char* sym_names,
    size_t sym_name_size,
    Symbol** sympointers);

} // End namespace gold.
Commit	Line	Data
14bfc3f5 ILT	1	// symtab.cc -- the gold symbol table
	2
	3	#include "gold.h"
	4
	5	#include <cassert>
	6	#include <stdint.h>
	7	#include <string>
	8	#include <utility>
	9
	10	#include "object.h"
	11	#include "symtab.h"
	12
	13	namespace gold
	14	{
	15
	16	// Class Symbol.
	17
14bfc3f5 ILT	18	// Initialize the fields in the base class Symbol.
	19
	20	template<int size, bool big_endian>
	21	void
	22	Symbol::init_base(const char* name, const char* version, Object* object,
	23	const elfcpp::Sym<size, big_endian>& sym)
	24	{
	25	this->name_ = name;
	26	this->version_ = version;
	27	this->object_ = object;
	28	this->shnum_ = sym.get_st_shndx(); // FIXME: Handle SHN_XINDEX.
	29	this->type_ = sym.get_st_type();
	30	this->binding_ = sym.get_st_bind();
	31	this->visibility_ = sym.get_st_visibility();
	32	this->other_ = sym.get_st_nonvis();
1564db8d ILT	33	this->is_special_ = false;
	34	this->is_def_ = false;
	35	this->is_forwarder_ = false;
	36	this->in_dyn_ = object->is_dynamic();
14bfc3f5 ILT	37	}
	38
	39	// Initialize the fields in Sized_symbol.
	40
	41	template<int size>
	42	template<bool big_endian>
	43	void
	44	Sized_symbol<size>::init(const char* name, const char* version, Object* object,
	45	const elfcpp::Sym<size, big_endian>& sym)
	46	{
	47	this->init_base(name, version, object, sym);
	48	this->value_ = sym.get_st_value();
	49	this->size_ = sym.get_st_size();
	50	}
	51
	52	// Class Symbol_table.
	53
	54	Symbol_table::Symbol_table()
	55	: size_(0), table_(), namepool_(), forwarders_()
	56	{
	57	}
	58
	59	Symbol_table::~Symbol_table()
	60	{
	61	}
	62
	63	// The hash function. The key is always canonicalized, so we use a
	64	// simple combination of the pointers.
	65
	66	size_t
	67	Symbol_table::Symbol_table_hash::operator()(const Symbol_table_key& key) const
	68	{
	69	return (reinterpret_cast<size_t>(key.first)
	70	^ reinterpret_cast<size_t>(key.second));
	71	}
	72
	73	// The symbol table key equality function. This is only called with
	74	// canonicalized name and version strings, so we can use pointer
	75	// comparison.
	76
	77	bool
	78	Symbol_table::Symbol_table_eq::operator()(const Symbol_table_key& k1,
	79	const Symbol_table_key& k2) const
	80	{
	81	return k1.first == k2.first && k1.second == k2.second;
	82	}
	83
	84	// Make TO a symbol which forwards to FROM.
	85
	86	void
	87	Symbol_table::make_forwarder(Symbol* from, Symbol* to)
	88	{
	89	assert(!from->is_forwarder() && !to->is_forwarder());
	90	this->forwarders_[from] = to;
	91	from->set_forwarder();
	92	}
	93
	94	Symbol*
	95	Symbol_table::resolve_forwards(Symbol* from) const
	96	{
	97	assert(from->is_forwarder());
	98	Unordered_map<Symbol, Symbol>::const_iterator p =
	99	this->forwarders_.find(from);
	100	assert(p != this->forwarders_.end());
101	return p->second;
102	}
103
104	// Resolve a Symbol with another Symbol. This is only used in the
105	// unusual case where there are references to both an unversioned
106	// symbol and a symbol with a version, and we then discover that that
1564db8d ILT	107	// version is the default version. Because this is unusual, we do
1564db8d ILT	108	// this the slow way, by converting back to an ELF symbol.
14bfc3f5	109
1564db8d	110	template<int size, bool big_endian>
14bfc3f5	111	void
1564db8d	112	Symbol_table::resolve(Sized_symbol<size>* to, const Sized_symbol<size>* from)
14bfc3f5	113	{
1564db8d ILT	114	unsigned char buf[elfcpp::Elf_sizes<size>::sym_size];
	115	elfcpp::Sym_write<size, big_endian> esym(buf);
	116	// We don't bother to set the st_name field.
	117	esym.put_st_value(from->value());
	118	esym.put_st_size(from->symsize());
	119	esym.put_st_info(from->binding(), from->type());
	120	esym.put_st_other(from->visibility(), from->other());
	121	esym.put_st_shndx(from->shnum());
	122	Symbol_table::resolve(to, esym.sym(), from->object());
14bfc3f5 ILT	123	}
	124
	125	// Add one symbol from OBJECT to the symbol table. NAME is symbol
	126	// name and VERSION is the version; both are canonicalized. DEF is
	127	// whether this is the default version.
	128
	129	// If DEF is true, then this is the definition of a default version of
	130	// a symbol. That means that any lookup of NAME/NULL and any lookup
	131	// of NAME/VERSION should always return the same symbol. This is
	132	// obvious for references, but in particular we want to do this for
	133	// definitions: overriding NAME/NULL should also override
	134	// NAME/VERSION. If we don't do that, it would be very hard to
	135	// override functions in a shared library which uses versioning.
	136
	137	// We implement this by simply making both entries in the hash table
	138	// point to the same Symbol structure. That is easy enough if this is
	139	// the first time we see NAME/NULL or NAME/VERSION, but it is possible
	140	// that we have seen both already, in which case they will both have
	141	// independent entries in the symbol table. We can't simply change
	142	// the symbol table entry, because we have pointers to the entries
	143	// attached to the object files. So we mark the entry attached to the
	144	// object file as a forwarder, and record it in the forwarders_ map.
	145	// Note that entries in the hash table will never be marked as
	146	// forwarders.
	147
	148	template<int size, bool big_endian>
	149	Symbol*
	150	Symbol_table::add_from_object(Sized_object<size, big_endian>* object,
	151	const char *name,
	152	const char *version, bool def,
	153	const elfcpp::Sym<size, big_endian>& sym)
	154	{
	155	Symbol* const snull = NULL;
	156	std::pair<typename Symbol_table_type::iterator, bool> ins =
	157	this->table_.insert(std::make_pair(std::make_pair(name, version), snull));
	158
	159	std::pair<typename Symbol_table_type::iterator, bool> insdef =
	160	std::make_pair(this->table_.end(), false);
	161	if (def)
	162	{
	163	const char* const vnull = NULL;
	164	insdef = this->table_.insert(std::make_pair(std::make_pair(name, vnull),
	165	snull));
	166	}
	167
	168	// ins.first: an iterator, which is a pointer to a pair.
	169	// ins.first->first: the key (a pair of name and version).
	170	// ins.first->second: the value (Symbol*).
	171	// ins.second: true if new entry was inserted, false if not.
	172
1564db8d	173	Sized_symbol<size>* ret;
14bfc3f5 ILT	174	if (!ins.second)
	175	{
	176	// We already have an entry for NAME/VERSION.
1564db8d	177	ret = this->get_sized_symbol<size>(ins.first->second);
14bfc3f5 ILT	178	assert(ret != NULL);
	179	Symbol_table::resolve(ret, sym, object);
	180
	181	if (def)
	182	{
	183	if (insdef.second)
	184	{
	185	// This is the first time we have seen NAME/NULL. Make
	186	// NAME/NULL point to NAME/VERSION.
	187	insdef.first->second = ret;
	188	}
	189	else
	190	{
	191	// This is the unfortunate case where we already have
	192	// entries for both NAME/VERSION and NAME/NULL.
1564db8d ILT	193	const Sized_symbol<size>* sym2 =
	194	this->get_sized_symbol<size>(insdef.first->second);
	195	Symbol_table::resolve<size, big_endian>(ret, sym2);
14bfc3f5 ILT	196	this->make_forwarder(insdef.first->second, ret);
	197	insdef.first->second = ret;
	198	}
	199	}
	200	}
	201	else
	202	{
	203	// This is the first time we have seen NAME/VERSION.
	204	assert(ins.first->second == NULL);
	205	if (def && !insdef.second)
	206	{
	207	// We already have an entry for NAME/NULL. Make
	208	// NAME/VERSION point to it.
1564db8d	209	ret = this->get_sized_symbol<size>(insdef.first->second);
14bfc3f5 ILT	210	Symbol_table::resolve(ret, sym, object);
	211	ins.first->second = ret;
	212	}
	213	else
	214	{
14bfc3f5	215	Sized_target<size, big_endian>* target = object->sized_target();
1564db8d ILT	216	if (!target->has_make_symbol())
	217	ret = new Sized_symbol<size>();
	218	else
14bfc3f5	219	{
1564db8d ILT	220	ret = target->make_symbol();
1564db8d ILT	221	if (ret == NULL)
14bfc3f5 ILT	222	{
	223	// This means that we don't want a symbol table
	224	// entry after all.
	225	if (!def)
	226	this->table_.erase(ins.first);
	227	else
	228	{
	229	this->table_.erase(insdef.first);
	230	// Inserting insdef invalidated ins.
	231	this->table_.erase(std::make_pair(name, version));
	232	}
	233	return NULL;
	234	}
	235	}
14bfc3f5	236
1564db8d ILT	237	ret->init(name, version, object, sym);
1564db8d ILT	238
14bfc3f5 ILT	239	ins.first->second = ret;
	240	if (def)
	241	{
	242	// This is the first time we have seen NAME/NULL. Point
	243	// it at the new entry for NAME/VERSION.
	244	assert(insdef.second);
	245	insdef.first->second = ret;
	246	}
	247	}
	248	}
	249
	250	return ret;
	251	}
	252
	253	// Add all the symbols in an object to the hash table.
	254
	255	template<int size, bool big_endian>
	256	void
	257	Symbol_table::add_from_object(
	258	Sized_object<size, big_endian>* object,
	259	const elfcpp::Sym<size, big_endian>* syms,
	260	size_t count,
	261	const char* sym_names,
	262	size_t sym_name_size,
	263	Symbol** sympointers)
	264	{
	265	// We take the size from the first object we see.
	266	if (this->get_size() == 0)
	267	this->set_size(size);
	268
	269	if (size != this->get_size() \|\| size != object->target()->get_size())
	270	{
	271	fprintf(stderr, _("%s: %s: mixing 32-bit and 64-bit ELF objects\n"),
	272	program_name, object->name().c_str());
	273	gold_exit(false);
	274	}
	275
	276	const unsigned char* p = reinterpret_cast<const unsigned char*>(syms);
	277	for (size_t i = 0; i < count; ++i)
	278	{
	279	elfcpp::Sym<size, big_endian> sym(p);
	280
	281	unsigned int st_name = sym.get_st_name();
	282	if (st_name >= sym_name_size)
	283	{
	284	fprintf(stderr, _("%s: %s: bad symbol name offset %u at %lu\n"),
	285	program_name, object->name().c_str(), st_name,
	286	static_cast<unsigned long>(i));
	287	gold_exit(false);
	288	}
	289
	290	const char* name = sym_names + st_name;
	291
	292	// In an object file, an '@' in the name separates the symbol
	293	// name from the version name. If there are two '@' characters,
	294	// this is the default version.
	295	const char* ver = strchr(name, '@');
	296
	297	Symbol* res;
	298	if (ver == NULL)
	299	{
	300	name = this->namepool_.add(name);
	301	res = this->add_from_object(object, name, NULL, false, sym);
	302	}
303	else
304	{
305	name = this->namepool_.add(name, ver - name);
306	bool def = false;
307	++ver;
308	if (*ver == '@')
309	{
310	def = true;
311	++ver;
312	}
313	ver = this->namepool_.add(ver);
314	res = this->add_from_object(object, name, ver, def, sym);
315	}
316
317	*sympointers++ = res;
318
319	p += elfcpp::Elf_sizes<size>::sym_size;
320	}
321	}
322
323	// Instantiate the templates we need. We could use the configure
324	// script to restrict this to only the ones needed for implemented
325	// targets.
326
327	template
328	void
329	Symbol_table::add_from_object<32, true>(
330	Sized_object<32, true>* object,
331	const elfcpp::Sym<32, true>* syms,
332	size_t count,
333	const char* sym_names,
334	size_t sym_name_size,
335	Symbol** sympointers);
336
337	template
338	void
339	Symbol_table::add_from_object<32, false>(
340	Sized_object<32, false>* object,
341	const elfcpp::Sym<32, false>* syms,
342	size_t count,
343	const char* sym_names,
344	size_t sym_name_size,
345	Symbol** sympointers);
346
347	template
348	void
349	Symbol_table::add_from_object<64, true>(
350	Sized_object<64, true>* object,
351	const elfcpp::Sym<64, true>* syms,
352	size_t count,
353	const char* sym_names,
354	size_t sym_name_size,
355	Symbol** sympointers);
356
357	template
358	void
359	Symbol_table::add_from_object<64, false>(
360	Sized_object<64, false>* object,
361	const elfcpp::Sym<64, false>* syms,
362	size_t count,
363	const char* sym_names,
364	size_t sym_name_size,
365	Symbol** sympointers);
366
367	} // End namespace gold.