2009-08-06 04:51:56 +08:00
|
|
|
// icf.cc -- Identical Code Folding.
|
|
|
|
//
|
2014-03-05 19:46:15 +08:00
|
|
|
// Copyright (C) 2009-2014 Free Software Foundation, Inc.
|
2009-08-06 04:51:56 +08:00
|
|
|
// Written by Sriraman Tallam <tmsriram@google.com>.
|
|
|
|
|
|
|
|
// This file is part of gold.
|
|
|
|
|
|
|
|
// This program is free software; you can redistribute it and/or modify
|
|
|
|
// it under the terms of the GNU General Public License as published by
|
|
|
|
// the Free Software Foundation; either version 3 of the License, or
|
|
|
|
// (at your option) any later version.
|
|
|
|
|
|
|
|
// This program is distributed in the hope that it will be useful,
|
|
|
|
// but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
|
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
|
|
// GNU General Public License for more details.
|
|
|
|
|
|
|
|
// You should have received a copy of the GNU General Public License
|
|
|
|
// along with this program; if not, write to the Free Software
|
|
|
|
// Foundation, Inc., 51 Franklin Street - Fifth Floor, Boston,
|
|
|
|
// MA 02110-1301, USA.
|
|
|
|
|
|
|
|
// Identical Code Folding Algorithm
|
|
|
|
// ----------------------------------
|
|
|
|
// Detecting identical functions is done here and the basic algorithm
|
2010-01-05 03:08:39 +08:00
|
|
|
// is as follows. A checksum is computed on each foldable section using
|
2009-08-06 04:51:56 +08:00
|
|
|
// its contents and relocations. If the symbol name corresponding to
|
|
|
|
// a relocation is known it is used to compute the checksum. If the
|
|
|
|
// symbol name is not known the stringified name of the object and the
|
|
|
|
// section number pointed to by the relocation is used. The checksums
|
|
|
|
// are stored as keys in a hash map and a section is identical to some
|
|
|
|
// other section if its checksum is already present in the hash map.
|
|
|
|
// Checksum collisions are handled by using a multimap and explicitly
|
|
|
|
// checking the contents when two sections have the same checksum.
|
|
|
|
//
|
|
|
|
// However, two functions A and B with identical text but with
|
2010-01-05 03:08:39 +08:00
|
|
|
// relocations pointing to different foldable sections can be identical if
|
|
|
|
// the corresponding foldable sections to which their relocations point to
|
2009-08-06 04:51:56 +08:00
|
|
|
// turn out to be identical. Hence, this checksumming process must be
|
|
|
|
// done repeatedly until convergence is obtained. Here is an example for
|
|
|
|
// the following case :
|
|
|
|
//
|
|
|
|
// int funcA () int funcB ()
|
|
|
|
// { {
|
|
|
|
// return foo(); return goo();
|
|
|
|
// } }
|
|
|
|
//
|
|
|
|
// The functions funcA and funcB are identical if functions foo() and
|
|
|
|
// goo() are identical.
|
|
|
|
//
|
|
|
|
// Hence, as described above, we repeatedly do the checksumming,
|
|
|
|
// assigning identical functions to the same group, until convergence is
|
|
|
|
// obtained. Now, we have two different ways to do this depending on how
|
|
|
|
// we initialize.
|
|
|
|
//
|
|
|
|
// Algorithm I :
|
|
|
|
// -----------
|
|
|
|
// We can start with marking all functions as different and repeatedly do
|
|
|
|
// the checksumming. This has the advantage that we do not need to wait
|
|
|
|
// for convergence. We can stop at any point and correctness will be
|
|
|
|
// guaranteed although not all cases would have been found. However, this
|
|
|
|
// has a problem that some cases can never be found even if it is run until
|
|
|
|
// convergence. Here is an example with mutually recursive functions :
|
|
|
|
//
|
|
|
|
// int funcA (int a) int funcB (int a)
|
|
|
|
// { {
|
|
|
|
// if (a == 1) if (a == 1)
|
|
|
|
// return 1; return 1;
|
|
|
|
// return 1 + funcB(a - 1); return 1 + funcA(a - 1);
|
|
|
|
// } }
|
|
|
|
//
|
|
|
|
// In this example funcA and funcB are identical and one of them could be
|
|
|
|
// folded into the other. However, if we start with assuming that funcA
|
|
|
|
// and funcB are not identical, the algorithm, even after it is run to
|
|
|
|
// convergence, cannot detect that they are identical. It should be noted
|
|
|
|
// that even if the functions were self-recursive, Algorithm I cannot catch
|
|
|
|
// that they are identical, at least as is.
|
|
|
|
//
|
|
|
|
// Algorithm II :
|
|
|
|
// ------------
|
|
|
|
// Here we start with marking all functions as identical and then repeat
|
|
|
|
// the checksumming until convergence. This can detect the above case
|
|
|
|
// mentioned above. It can detect all cases that Algorithm I can and more.
|
|
|
|
// However, the caveat is that it has to be run to convergence. It cannot
|
|
|
|
// be stopped arbitrarily like Algorithm I as correctness cannot be
|
|
|
|
// guaranteed. Algorithm II is not implemented.
|
|
|
|
//
|
|
|
|
// Algorithm I is used because experiments show that about three
|
|
|
|
// iterations are more than enough to achieve convergence. Algorithm I can
|
|
|
|
// handle recursive calls if it is changed to use a special common symbol
|
|
|
|
// for recursive relocs. This seems to be the most common case that
|
|
|
|
// Algorithm I could not catch as is. Mutually recursive calls are not
|
|
|
|
// frequent and Algorithm I wins because of its ability to be stopped
|
|
|
|
// arbitrarily.
|
|
|
|
//
|
|
|
|
// Caveat with using function pointers :
|
|
|
|
// ------------------------------------
|
|
|
|
//
|
|
|
|
// Programs using function pointer comparisons/checks should use function
|
|
|
|
// folding with caution as the result of such comparisons could be different
|
|
|
|
// when folding takes place. This could lead to unexpected run-time
|
|
|
|
// behaviour.
|
|
|
|
//
|
2010-02-13 10:04:21 +08:00
|
|
|
// Safe Folding :
|
|
|
|
// ------------
|
|
|
|
//
|
|
|
|
// ICF in safe mode folds only ctors and dtors if their function pointers can
|
|
|
|
// never be taken. Also, for X86-64, safe folding uses the relocation
|
|
|
|
// type to determine if a function's pointer is taken or not and only folds
|
|
|
|
// functions whose pointers are definitely not taken.
|
|
|
|
//
|
|
|
|
// Caveat with safe folding :
|
|
|
|
// ------------------------
|
|
|
|
//
|
|
|
|
// This applies only to x86_64.
|
|
|
|
//
|
|
|
|
// Position independent executables are created from PIC objects (compiled
|
|
|
|
// with -fPIC) and/or PIE objects (compiled with -fPIE). For PIE objects, the
|
|
|
|
// relocation types for function pointer taken and a call are the same.
|
|
|
|
// Now, it is not always possible to tell if an object used in the link of
|
|
|
|
// a pie executable is a PIC object or a PIE object. Hence, for pie
|
|
|
|
// executables, using relocation types to disambiguate function pointers is
|
|
|
|
// currently disabled.
|
|
|
|
//
|
|
|
|
// Further, it is not correct to use safe folding to build non-pie
|
|
|
|
// executables using PIC/PIE objects. PIC/PIE objects have different
|
|
|
|
// relocation types for function pointers than non-PIC objects, and the
|
|
|
|
// current implementation of safe folding does not handle those relocation
|
|
|
|
// types. Hence, if used, functions whose pointers are taken could still be
|
|
|
|
// folded causing unpredictable run-time behaviour if the pointers were used
|
|
|
|
// in comparisons.
|
|
|
|
//
|
|
|
|
//
|
2009-08-06 04:51:56 +08:00
|
|
|
//
|
2010-01-05 03:08:39 +08:00
|
|
|
// How to run : --icf=[safe|all|none]
|
2009-08-06 04:51:56 +08:00
|
|
|
// Optional parameters : --icf-iterations <num> --print-icf-sections
|
|
|
|
//
|
|
|
|
// Performance : Less than 20 % link-time overhead on industry strength
|
|
|
|
// applications. Up to 6 % text size reductions.
|
|
|
|
|
|
|
|
#include "gold.h"
|
|
|
|
#include "object.h"
|
|
|
|
#include "gc.h"
|
|
|
|
#include "icf.h"
|
|
|
|
#include "symtab.h"
|
|
|
|
#include "libiberty.h"
|
2009-10-14 05:17:43 +08:00
|
|
|
#include "demangle.h"
|
2010-07-30 02:57:28 +08:00
|
|
|
#include "elfcpp.h"
|
|
|
|
#include "int_encoding.h"
|
2009-08-06 04:51:56 +08:00
|
|
|
|
|
|
|
namespace gold
|
|
|
|
{
|
|
|
|
|
|
|
|
// This function determines if a section or a group of identical
|
|
|
|
// sections has unique contents. Such unique sections or groups can be
|
|
|
|
// declared final and need not be processed any further.
|
|
|
|
// Parameters :
|
|
|
|
// ID_SECTION : Vector mapping a section index to a Section_id pair.
|
|
|
|
// IS_SECN_OR_GROUP_UNIQUE : To check if a section or a group of identical
|
|
|
|
// sections is already known to be unique.
|
|
|
|
// SECTION_CONTENTS : Contains the section's text and relocs to sections
|
|
|
|
// that cannot be folded. SECTION_CONTENTS are NULL
|
|
|
|
// implies that this function is being called for the
|
|
|
|
// first time before the first iteration of icf.
|
|
|
|
|
|
|
|
static void
|
|
|
|
preprocess_for_unique_sections(const std::vector<Section_id>& id_section,
|
|
|
|
std::vector<bool>* is_secn_or_group_unique,
|
|
|
|
std::vector<std::string>* section_contents)
|
|
|
|
{
|
|
|
|
Unordered_map<uint32_t, unsigned int> uniq_map;
|
|
|
|
std::pair<Unordered_map<uint32_t, unsigned int>::iterator, bool>
|
|
|
|
uniq_map_insert;
|
|
|
|
|
|
|
|
for (unsigned int i = 0; i < id_section.size(); i++)
|
|
|
|
{
|
|
|
|
if ((*is_secn_or_group_unique)[i])
|
|
|
|
continue;
|
|
|
|
|
|
|
|
uint32_t cksum;
|
|
|
|
Section_id secn = id_section[i];
|
|
|
|
section_size_type plen;
|
|
|
|
if (section_contents == NULL)
|
|
|
|
{
|
2010-11-06 05:14:33 +08:00
|
|
|
// Lock the object so we can read from it. This is only called
|
|
|
|
// single-threaded from queue_middle_tasks, so it is OK to lock.
|
|
|
|
// Unfortunately we have no way to pass in a Task token.
|
|
|
|
const Task* dummy_task = reinterpret_cast<const Task*>(-1);
|
|
|
|
Task_lock_obj<Object> tl(dummy_task, secn.first);
|
2009-08-06 04:51:56 +08:00
|
|
|
const unsigned char* contents;
|
|
|
|
contents = secn.first->section_contents(secn.second,
|
|
|
|
&plen,
|
|
|
|
false);
|
|
|
|
cksum = xcrc32(contents, plen, 0xffffffff);
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
const unsigned char* contents_array = reinterpret_cast
|
|
|
|
<const unsigned char*>((*section_contents)[i].c_str());
|
|
|
|
cksum = xcrc32(contents_array, (*section_contents)[i].length(),
|
|
|
|
0xffffffff);
|
|
|
|
}
|
|
|
|
uniq_map_insert = uniq_map.insert(std::make_pair(cksum, i));
|
|
|
|
if (uniq_map_insert.second)
|
|
|
|
{
|
|
|
|
(*is_secn_or_group_unique)[i] = true;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
(*is_secn_or_group_unique)[i] = false;
|
|
|
|
(*is_secn_or_group_unique)[uniq_map_insert.first->second] = false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// This returns the buffer containing the section's contents, both
|
|
|
|
// text and relocs. Relocs are differentiated as those pointing to
|
|
|
|
// sections that could be folded and those that cannot. Only relocs
|
|
|
|
// pointing to sections that could be folded are recomputed on
|
|
|
|
// subsequent invocations of this function.
|
|
|
|
// Parameters :
|
|
|
|
// FIRST_ITERATION : true if it is the first invocation.
|
|
|
|
// SECN : Section for which contents are desired.
|
|
|
|
// SECTION_NUM : Unique section number of this section.
|
|
|
|
// NUM_TRACKED_RELOCS : Vector reference to store the number of relocs
|
|
|
|
// to ICF sections.
|
|
|
|
// KEPT_SECTION_ID : Vector which maps folded sections to kept sections.
|
|
|
|
// SECTION_CONTENTS : Store the section's text and relocs to non-ICF
|
|
|
|
// sections.
|
|
|
|
|
|
|
|
static std::string
|
|
|
|
get_section_contents(bool first_iteration,
|
|
|
|
const Section_id& secn,
|
|
|
|
unsigned int section_num,
|
|
|
|
unsigned int* num_tracked_relocs,
|
|
|
|
Symbol_table* symtab,
|
|
|
|
const std::vector<unsigned int>& kept_section_id,
|
|
|
|
std::vector<std::string>* section_contents)
|
|
|
|
{
|
2011-01-26 01:14:59 +08:00
|
|
|
// Lock the object so we can read from it. This is only called
|
|
|
|
// single-threaded from queue_middle_tasks, so it is OK to lock.
|
|
|
|
// Unfortunately we have no way to pass in a Task token.
|
|
|
|
const Task* dummy_task = reinterpret_cast<const Task*>(-1);
|
|
|
|
Task_lock_obj<Object> tl(dummy_task, secn.first);
|
|
|
|
|
2009-08-06 04:51:56 +08:00
|
|
|
section_size_type plen;
|
|
|
|
const unsigned char* contents = NULL;
|
|
|
|
if (first_iteration)
|
2011-01-26 01:14:59 +08:00
|
|
|
contents = secn.first->section_contents(secn.second, &plen, false);
|
2009-08-06 04:51:56 +08:00
|
|
|
|
|
|
|
// The buffer to hold all the contents including relocs. A checksum
|
|
|
|
// is then computed on this buffer.
|
|
|
|
std::string buffer;
|
|
|
|
std::string icf_reloc_buffer;
|
|
|
|
|
|
|
|
if (num_tracked_relocs)
|
|
|
|
*num_tracked_relocs = 0;
|
|
|
|
|
2010-02-21 08:57:59 +08:00
|
|
|
Icf::Reloc_info_list& reloc_info_list =
|
|
|
|
symtab->icf()->reloc_info_list();
|
2009-08-06 04:51:56 +08:00
|
|
|
|
2010-02-21 08:57:59 +08:00
|
|
|
Icf::Reloc_info_list::iterator it_reloc_info_list =
|
|
|
|
reloc_info_list.find(secn);
|
2009-08-06 04:51:56 +08:00
|
|
|
|
|
|
|
buffer.clear();
|
|
|
|
icf_reloc_buffer.clear();
|
|
|
|
|
|
|
|
// Process relocs and put them into the buffer.
|
|
|
|
|
2010-02-21 08:57:59 +08:00
|
|
|
if (it_reloc_info_list != reloc_info_list.end())
|
2009-08-06 04:51:56 +08:00
|
|
|
{
|
2014-04-03 07:46:02 +08:00
|
|
|
Icf::Sections_reachable_info &v =
|
2010-02-21 08:57:59 +08:00
|
|
|
(it_reloc_info_list->second).section_info;
|
2010-04-24 02:49:23 +08:00
|
|
|
// Stores the information of the symbol pointed to by the reloc.
|
2014-04-03 07:46:02 +08:00
|
|
|
const Icf::Symbol_info &s = (it_reloc_info_list->second).symbol_info;
|
2010-04-24 02:49:23 +08:00
|
|
|
// Stores the addend and the symbol value.
|
2014-04-03 07:46:02 +08:00
|
|
|
Icf::Addend_info &a = (it_reloc_info_list->second).addend_info;
|
2010-04-24 02:49:23 +08:00
|
|
|
// Stores the offset of the reloc.
|
2014-04-03 07:46:02 +08:00
|
|
|
const Icf::Offset_info &o = (it_reloc_info_list->second).offset_info;
|
|
|
|
const Icf::Reloc_addend_size_info &reloc_addend_size_info =
|
2010-07-30 02:57:28 +08:00
|
|
|
(it_reloc_info_list->second).reloc_addend_size_info;
|
2010-02-21 08:57:59 +08:00
|
|
|
Icf::Sections_reachable_info::iterator it_v = v.begin();
|
2014-04-03 07:46:02 +08:00
|
|
|
Icf::Symbol_info::const_iterator it_s = s.begin();
|
2009-08-06 04:51:56 +08:00
|
|
|
Icf::Addend_info::iterator it_a = a.begin();
|
2014-04-03 07:46:02 +08:00
|
|
|
Icf::Offset_info::const_iterator it_o = o.begin();
|
|
|
|
Icf::Reloc_addend_size_info::const_iterator it_addend_size =
|
2010-07-30 02:57:28 +08:00
|
|
|
reloc_addend_size_info.begin();
|
2009-08-06 04:51:56 +08:00
|
|
|
|
2010-07-30 02:57:28 +08:00
|
|
|
for (; it_v != v.end(); ++it_v, ++it_s, ++it_a, ++it_o, ++it_addend_size)
|
2009-08-06 04:51:56 +08:00
|
|
|
{
|
2013-03-15 15:51:32 +08:00
|
|
|
if (first_iteration
|
|
|
|
&& it_v->first != NULL)
|
|
|
|
{
|
|
|
|
Symbol_location loc;
|
|
|
|
loc.object = it_v->first;
|
|
|
|
loc.shndx = it_v->second;
|
|
|
|
loc.offset = convert_types<off_t, long long>(it_a->first
|
|
|
|
+ it_a->second);
|
|
|
|
// Look through function descriptors
|
|
|
|
parameters->target().function_location(&loc);
|
|
|
|
if (loc.shndx != it_v->second)
|
|
|
|
{
|
|
|
|
it_v->second = loc.shndx;
|
|
|
|
// Modify symvalue/addend to the code entry.
|
|
|
|
it_a->first = loc.offset;
|
|
|
|
it_a->second = 0;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2010-02-21 08:57:59 +08:00
|
|
|
// ADDEND_STR stores the symbol value and addend and offset,
|
2010-12-15 03:03:30 +08:00
|
|
|
// each at most 16 hex digits long. it_a points to a pair
|
2009-08-06 04:51:56 +08:00
|
|
|
// where first is the symbol value and second is the
|
|
|
|
// addend.
|
2010-02-21 08:57:59 +08:00
|
|
|
char addend_str[50];
|
2010-02-23 02:26:29 +08:00
|
|
|
|
|
|
|
// It would be nice if we could use format macros in inttypes.h
|
|
|
|
// here but there are not in ISO/IEC C++ 1998.
|
|
|
|
snprintf(addend_str, sizeof(addend_str), "%llx %llx %llux",
|
|
|
|
static_cast<long long>((*it_a).first),
|
|
|
|
static_cast<long long>((*it_a).second),
|
|
|
|
static_cast<unsigned long long>(*it_o));
|
2010-04-24 02:49:23 +08:00
|
|
|
|
|
|
|
// If the symbol pointed to by the reloc is not in an ordinary
|
|
|
|
// section or if the symbol type is not FROM_OBJECT, then the
|
|
|
|
// object is NULL.
|
|
|
|
if (it_v->first == NULL)
|
|
|
|
{
|
|
|
|
if (first_iteration)
|
|
|
|
{
|
|
|
|
// If the symbol name is available, use it.
|
|
|
|
if ((*it_s) != NULL)
|
|
|
|
buffer.append((*it_s)->name());
|
|
|
|
// Append the addend.
|
|
|
|
buffer.append(addend_str);
|
|
|
|
buffer.append("@");
|
|
|
|
}
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2009-08-06 04:51:56 +08:00
|
|
|
Section_id reloc_secn(it_v->first, it_v->second);
|
|
|
|
|
|
|
|
// If this reloc turns back and points to the same section,
|
|
|
|
// like a recursive call, use a special symbol to mark this.
|
|
|
|
if (reloc_secn.first == secn.first
|
|
|
|
&& reloc_secn.second == secn.second)
|
|
|
|
{
|
|
|
|
if (first_iteration)
|
|
|
|
{
|
|
|
|
buffer.append("R");
|
|
|
|
buffer.append(addend_str);
|
|
|
|
buffer.append("@");
|
|
|
|
}
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
Icf::Uniq_secn_id_map& section_id_map =
|
|
|
|
symtab->icf()->section_to_int_map();
|
|
|
|
Icf::Uniq_secn_id_map::iterator section_id_map_it =
|
|
|
|
section_id_map.find(reloc_secn);
|
2010-04-21 05:13:30 +08:00
|
|
|
bool is_sym_preemptible = (*it_s != NULL
|
|
|
|
&& !(*it_s)->is_from_dynobj()
|
|
|
|
&& !(*it_s)->is_undefined()
|
|
|
|
&& (*it_s)->is_preemptible());
|
|
|
|
if (!is_sym_preemptible
|
|
|
|
&& section_id_map_it != section_id_map.end())
|
2009-08-06 04:51:56 +08:00
|
|
|
{
|
|
|
|
// This is a reloc to a section that might be folded.
|
|
|
|
if (num_tracked_relocs)
|
|
|
|
(*num_tracked_relocs)++;
|
|
|
|
|
|
|
|
char kept_section_str[10];
|
|
|
|
unsigned int secn_id = section_id_map_it->second;
|
|
|
|
snprintf(kept_section_str, sizeof(kept_section_str), "%u",
|
|
|
|
kept_section_id[secn_id]);
|
|
|
|
if (first_iteration)
|
|
|
|
{
|
|
|
|
buffer.append("ICF_R");
|
|
|
|
buffer.append(addend_str);
|
|
|
|
}
|
|
|
|
icf_reloc_buffer.append(kept_section_str);
|
|
|
|
// Append the addend.
|
|
|
|
icf_reloc_buffer.append(addend_str);
|
|
|
|
icf_reloc_buffer.append("@");
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
// This is a reloc to a section that cannot be folded.
|
|
|
|
// Process it only in the first iteration.
|
|
|
|
if (!first_iteration)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
uint64_t secn_flags = (it_v->first)->section_flags(it_v->second);
|
|
|
|
// This reloc points to a merge section. Hash the
|
|
|
|
// contents of this section.
|
2010-09-29 01:14:15 +08:00
|
|
|
if ((secn_flags & elfcpp::SHF_MERGE) != 0
|
2011-06-29 05:15:42 +08:00
|
|
|
&& parameters->target().can_icf_inline_merge_sections())
|
2009-08-06 04:51:56 +08:00
|
|
|
{
|
|
|
|
uint64_t entsize =
|
|
|
|
(it_v->first)->section_entsize(it_v->second);
|
2010-04-21 05:13:30 +08:00
|
|
|
long long offset = it_a->first;
|
|
|
|
|
|
|
|
unsigned long long addend = it_a->second;
|
|
|
|
// Ignoring the addend when it is a negative value. See the
|
|
|
|
// comments in Merged_symbol_value::Value in object.h.
|
|
|
|
if (addend < 0xffffff00)
|
|
|
|
offset = offset + addend;
|
|
|
|
|
2010-07-30 02:57:28 +08:00
|
|
|
// For SHT_REL relocation sections, the addend is stored in the
|
|
|
|
// text section at the relocation offset.
|
|
|
|
uint64_t reloc_addend_value = 0;
|
|
|
|
const unsigned char* reloc_addend_ptr =
|
|
|
|
contents + static_cast<unsigned long long>(*it_o);
|
|
|
|
switch(*it_addend_size)
|
|
|
|
{
|
|
|
|
case 0:
|
|
|
|
{
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case 1:
|
|
|
|
{
|
|
|
|
reloc_addend_value =
|
|
|
|
read_from_pointer<8>(reloc_addend_ptr);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case 2:
|
|
|
|
{
|
|
|
|
reloc_addend_value =
|
|
|
|
read_from_pointer<16>(reloc_addend_ptr);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case 4:
|
|
|
|
{
|
|
|
|
reloc_addend_value =
|
|
|
|
read_from_pointer<32>(reloc_addend_ptr);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case 8:
|
|
|
|
{
|
|
|
|
reloc_addend_value =
|
|
|
|
read_from_pointer<64>(reloc_addend_ptr);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
default:
|
|
|
|
gold_unreachable();
|
|
|
|
}
|
|
|
|
offset = offset + reloc_addend_value;
|
|
|
|
|
2009-08-06 04:51:56 +08:00
|
|
|
section_size_type secn_len;
|
|
|
|
const unsigned char* str_contents =
|
|
|
|
(it_v->first)->section_contents(it_v->second,
|
|
|
|
&secn_len,
|
|
|
|
false) + offset;
|
|
|
|
if ((secn_flags & elfcpp::SHF_STRINGS) != 0)
|
|
|
|
{
|
|
|
|
// String merge section.
|
|
|
|
const char* str_char =
|
|
|
|
reinterpret_cast<const char*>(str_contents);
|
|
|
|
switch(entsize)
|
|
|
|
{
|
|
|
|
case 1:
|
|
|
|
{
|
|
|
|
buffer.append(str_char);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case 2:
|
|
|
|
{
|
|
|
|
const uint16_t* ptr_16 =
|
|
|
|
reinterpret_cast<const uint16_t*>(str_char);
|
|
|
|
unsigned int strlen_16 = 0;
|
|
|
|
// Find the NULL character.
|
|
|
|
while(*(ptr_16 + strlen_16) != 0)
|
|
|
|
strlen_16++;
|
|
|
|
buffer.append(str_char, strlen_16 * 2);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case 4:
|
|
|
|
{
|
|
|
|
const uint32_t* ptr_32 =
|
|
|
|
reinterpret_cast<const uint32_t*>(str_char);
|
|
|
|
unsigned int strlen_32 = 0;
|
|
|
|
// Find the NULL character.
|
|
|
|
while(*(ptr_32 + strlen_32) != 0)
|
|
|
|
strlen_32++;
|
|
|
|
buffer.append(str_char, strlen_32 * 4);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
gold_unreachable();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
// Use the entsize to determine the length.
|
|
|
|
buffer.append(reinterpret_cast<const
|
|
|
|
char*>(str_contents),
|
|
|
|
entsize);
|
|
|
|
}
|
2010-03-10 09:26:57 +08:00
|
|
|
buffer.append("@");
|
2009-08-06 04:51:56 +08:00
|
|
|
}
|
|
|
|
else if ((*it_s) != NULL)
|
|
|
|
{
|
|
|
|
// If symbol name is available use that.
|
2010-04-24 02:49:23 +08:00
|
|
|
buffer.append((*it_s)->name());
|
2009-08-06 04:51:56 +08:00
|
|
|
// Append the addend.
|
|
|
|
buffer.append(addend_str);
|
|
|
|
buffer.append("@");
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
// Symbol name is not available, like for a local symbol,
|
|
|
|
// use object and section id.
|
|
|
|
buffer.append(it_v->first->name());
|
|
|
|
char secn_id[10];
|
|
|
|
snprintf(secn_id, sizeof(secn_id), "%u",it_v->second);
|
|
|
|
buffer.append(secn_id);
|
|
|
|
// Append the addend.
|
|
|
|
buffer.append(addend_str);
|
|
|
|
buffer.append("@");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (first_iteration)
|
|
|
|
{
|
|
|
|
buffer.append("Contents = ");
|
|
|
|
buffer.append(reinterpret_cast<const char*>(contents), plen);
|
|
|
|
// Store the section contents that dont change to avoid recomputing
|
|
|
|
// during the next call to this function.
|
|
|
|
(*section_contents)[section_num] = buffer;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
gold_assert(buffer.empty());
|
|
|
|
// Reuse the contents computed in the previous iteration.
|
|
|
|
buffer.append((*section_contents)[section_num]);
|
|
|
|
}
|
|
|
|
|
|
|
|
buffer.append(icf_reloc_buffer);
|
|
|
|
return buffer;
|
|
|
|
}
|
|
|
|
|
|
|
|
// This function computes a checksum on each section to detect and form
|
|
|
|
// groups of identical sections. The first iteration does this for all
|
|
|
|
// sections.
|
|
|
|
// Further iterations do this only for the kept sections from each group to
|
|
|
|
// determine if larger groups of identical sections could be formed. The
|
|
|
|
// first section in each group is the kept section for that group.
|
|
|
|
//
|
|
|
|
// CRC32 is the checksumming algorithm and can have collisions. That is,
|
|
|
|
// two sections with different contents can have the same checksum. Hence,
|
|
|
|
// a multimap is used to maintain more than one group of checksum
|
|
|
|
// identical sections. A section is added to a group only after its
|
|
|
|
// contents are explicitly compared with the kept section of the group.
|
|
|
|
//
|
|
|
|
// Parameters :
|
|
|
|
// ITERATION_NUM : Invocation instance of this function.
|
|
|
|
// NUM_TRACKED_RELOCS : Vector reference to store the number of relocs
|
|
|
|
// to ICF sections.
|
|
|
|
// KEPT_SECTION_ID : Vector which maps folded sections to kept sections.
|
|
|
|
// ID_SECTION : Vector mapping a section to an unique integer.
|
|
|
|
// IS_SECN_OR_GROUP_UNIQUE : To check if a section or a group of identical
|
2010-12-15 03:03:30 +08:00
|
|
|
// sections is already known to be unique.
|
2009-08-06 04:51:56 +08:00
|
|
|
// SECTION_CONTENTS : Store the section's text and relocs to non-ICF
|
|
|
|
// sections.
|
|
|
|
|
|
|
|
static bool
|
|
|
|
match_sections(unsigned int iteration_num,
|
|
|
|
Symbol_table* symtab,
|
|
|
|
std::vector<unsigned int>* num_tracked_relocs,
|
|
|
|
std::vector<unsigned int>* kept_section_id,
|
|
|
|
const std::vector<Section_id>& id_section,
|
|
|
|
std::vector<bool>* is_secn_or_group_unique,
|
|
|
|
std::vector<std::string>* section_contents)
|
|
|
|
{
|
|
|
|
Unordered_multimap<uint32_t, unsigned int> section_cksum;
|
|
|
|
std::pair<Unordered_multimap<uint32_t, unsigned int>::iterator,
|
|
|
|
Unordered_multimap<uint32_t, unsigned int>::iterator> key_range;
|
|
|
|
bool converged = true;
|
|
|
|
|
|
|
|
if (iteration_num == 1)
|
|
|
|
preprocess_for_unique_sections(id_section,
|
|
|
|
is_secn_or_group_unique,
|
|
|
|
NULL);
|
|
|
|
else
|
|
|
|
preprocess_for_unique_sections(id_section,
|
|
|
|
is_secn_or_group_unique,
|
|
|
|
section_contents);
|
|
|
|
|
|
|
|
std::vector<std::string> full_section_contents;
|
|
|
|
|
|
|
|
for (unsigned int i = 0; i < id_section.size(); i++)
|
|
|
|
{
|
|
|
|
full_section_contents.push_back("");
|
|
|
|
if ((*is_secn_or_group_unique)[i])
|
|
|
|
continue;
|
|
|
|
|
|
|
|
Section_id secn = id_section[i];
|
|
|
|
std::string this_secn_contents;
|
|
|
|
uint32_t cksum;
|
|
|
|
if (iteration_num == 1)
|
|
|
|
{
|
|
|
|
unsigned int num_relocs = 0;
|
|
|
|
this_secn_contents = get_section_contents(true, secn, i, &num_relocs,
|
|
|
|
symtab, (*kept_section_id),
|
|
|
|
section_contents);
|
|
|
|
(*num_tracked_relocs)[i] = num_relocs;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
if ((*kept_section_id)[i] != i)
|
|
|
|
{
|
|
|
|
// This section is already folded into something. See
|
|
|
|
// if it should point to a different kept section.
|
|
|
|
unsigned int kept_section = (*kept_section_id)[i];
|
|
|
|
if (kept_section != (*kept_section_id)[kept_section])
|
|
|
|
{
|
|
|
|
(*kept_section_id)[i] = (*kept_section_id)[kept_section];
|
|
|
|
}
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
this_secn_contents = get_section_contents(false, secn, i, NULL,
|
|
|
|
symtab, (*kept_section_id),
|
|
|
|
section_contents);
|
|
|
|
}
|
|
|
|
|
|
|
|
const unsigned char* this_secn_contents_array =
|
|
|
|
reinterpret_cast<const unsigned char*>(this_secn_contents.c_str());
|
|
|
|
cksum = xcrc32(this_secn_contents_array, this_secn_contents.length(),
|
|
|
|
0xffffffff);
|
|
|
|
size_t count = section_cksum.count(cksum);
|
|
|
|
|
|
|
|
if (count == 0)
|
|
|
|
{
|
|
|
|
// Start a group with this cksum.
|
|
|
|
section_cksum.insert(std::make_pair(cksum, i));
|
|
|
|
full_section_contents[i] = this_secn_contents;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
key_range = section_cksum.equal_range(cksum);
|
|
|
|
Unordered_multimap<uint32_t, unsigned int>::iterator it;
|
|
|
|
// Search all the groups with this cksum for a match.
|
|
|
|
for (it = key_range.first; it != key_range.second; ++it)
|
|
|
|
{
|
|
|
|
unsigned int kept_section = it->second;
|
|
|
|
if (full_section_contents[kept_section].length()
|
|
|
|
!= this_secn_contents.length())
|
|
|
|
continue;
|
|
|
|
if (memcmp(full_section_contents[kept_section].c_str(),
|
|
|
|
this_secn_contents.c_str(),
|
|
|
|
this_secn_contents.length()) != 0)
|
|
|
|
continue;
|
|
|
|
(*kept_section_id)[i] = kept_section;
|
|
|
|
converged = false;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
if (it == key_range.second)
|
|
|
|
{
|
|
|
|
// Create a new group for this cksum.
|
|
|
|
section_cksum.insert(std::make_pair(cksum, i));
|
|
|
|
full_section_contents[i] = this_secn_contents;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
// If there are no relocs to foldable sections do not process
|
|
|
|
// this section any further.
|
|
|
|
if (iteration_num == 1 && (*num_tracked_relocs)[i] == 0)
|
|
|
|
(*is_secn_or_group_unique)[i] = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
return converged;
|
|
|
|
}
|
|
|
|
|
2009-10-14 05:17:43 +08:00
|
|
|
// During safe icf (--icf=safe), only fold functions that are ctors or dtors.
|
2011-02-03 05:55:29 +08:00
|
|
|
// This function returns true if the section name is that of a ctor or a dtor.
|
2009-10-14 05:17:43 +08:00
|
|
|
|
|
|
|
static bool
|
2011-02-03 05:55:29 +08:00
|
|
|
is_function_ctor_or_dtor(const std::string& section_name)
|
2009-10-14 05:17:43 +08:00
|
|
|
{
|
2011-02-03 05:55:29 +08:00
|
|
|
const char* mangled_func_name = strrchr(section_name.c_str(), '.');
|
|
|
|
gold_assert(mangled_func_name != NULL);
|
|
|
|
if ((is_prefix_of("._ZN", mangled_func_name)
|
|
|
|
|| is_prefix_of("._ZZ", mangled_func_name))
|
|
|
|
&& (is_gnu_v3_mangled_ctor(mangled_func_name + 1)
|
|
|
|
|| is_gnu_v3_mangled_dtor(mangled_func_name + 1)))
|
2009-10-14 05:17:43 +08:00
|
|
|
{
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
2009-08-06 04:51:56 +08:00
|
|
|
|
|
|
|
// This is the main ICF function called in gold.cc. This does the
|
|
|
|
// initialization and calls match_sections repeatedly (twice by default)
|
|
|
|
// which computes the crc checksums and detects identical functions.
|
|
|
|
|
|
|
|
void
|
|
|
|
Icf::find_identical_sections(const Input_objects* input_objects,
|
|
|
|
Symbol_table* symtab)
|
|
|
|
{
|
|
|
|
unsigned int section_num = 0;
|
2009-12-15 03:53:05 +08:00
|
|
|
std::vector<unsigned int> num_tracked_relocs;
|
2009-08-06 04:51:56 +08:00
|
|
|
std::vector<bool> is_secn_or_group_unique;
|
|
|
|
std::vector<std::string> section_contents;
|
2010-02-13 10:04:21 +08:00
|
|
|
const Target& target = parameters->target();
|
2009-08-06 04:51:56 +08:00
|
|
|
|
|
|
|
// Decide which sections are possible candidates first.
|
|
|
|
|
|
|
|
for (Input_objects::Relobj_iterator p = input_objects->relobj_begin();
|
|
|
|
p != input_objects->relobj_end();
|
|
|
|
++p)
|
|
|
|
{
|
2010-11-06 05:14:33 +08:00
|
|
|
// Lock the object so we can read from it. This is only called
|
|
|
|
// single-threaded from queue_middle_tasks, so it is OK to lock.
|
|
|
|
// Unfortunately we have no way to pass in a Task token.
|
|
|
|
const Task* dummy_task = reinterpret_cast<const Task*>(-1);
|
|
|
|
Task_lock_obj<Object> tl(dummy_task, *p);
|
|
|
|
|
2009-08-06 04:51:56 +08:00
|
|
|
for (unsigned int i = 0;i < (*p)->shnum(); ++i)
|
|
|
|
{
|
2011-02-03 05:55:29 +08:00
|
|
|
const std::string section_name = (*p)->section_name(i);
|
2010-01-05 03:08:39 +08:00
|
|
|
if (!is_section_foldable_candidate(section_name))
|
2009-08-06 04:51:56 +08:00
|
|
|
continue;
|
|
|
|
if (!(*p)->is_section_included(i))
|
|
|
|
continue;
|
|
|
|
if (parameters->options().gc_sections()
|
|
|
|
&& symtab->gc()->is_section_garbage(*p, i))
|
|
|
|
continue;
|
2010-01-05 03:08:39 +08:00
|
|
|
// With --icf=safe, check if the mangled function name is a ctor
|
|
|
|
// or a dtor. The mangled function name can be obtained from the
|
|
|
|
// section name by stripping the section prefix.
|
2009-10-14 05:17:43 +08:00
|
|
|
if (parameters->options().icf_safe_folding()
|
2011-02-03 05:55:29 +08:00
|
|
|
&& !is_function_ctor_or_dtor(section_name)
|
2010-02-13 10:04:21 +08:00
|
|
|
&& (!target.can_check_for_function_pointers()
|
|
|
|
|| section_has_function_pointers(*p, i)))
|
|
|
|
{
|
|
|
|
continue;
|
|
|
|
}
|
2009-08-06 04:51:56 +08:00
|
|
|
this->id_section_.push_back(Section_id(*p, i));
|
|
|
|
this->section_id_[Section_id(*p, i)] = section_num;
|
|
|
|
this->kept_section_id_.push_back(section_num);
|
2009-12-15 03:53:05 +08:00
|
|
|
num_tracked_relocs.push_back(0);
|
2009-08-06 04:51:56 +08:00
|
|
|
is_secn_or_group_unique.push_back(false);
|
|
|
|
section_contents.push_back("");
|
|
|
|
section_num++;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
unsigned int num_iterations = 0;
|
|
|
|
|
|
|
|
// Default number of iterations to run ICF is 2.
|
|
|
|
unsigned int max_iterations = (parameters->options().icf_iterations() > 0)
|
|
|
|
? parameters->options().icf_iterations()
|
|
|
|
: 2;
|
|
|
|
|
|
|
|
bool converged = false;
|
|
|
|
|
|
|
|
while (!converged && (num_iterations < max_iterations))
|
|
|
|
{
|
|
|
|
num_iterations++;
|
|
|
|
converged = match_sections(num_iterations, symtab,
|
2009-12-15 03:53:05 +08:00
|
|
|
&num_tracked_relocs, &this->kept_section_id_,
|
2009-08-06 04:51:56 +08:00
|
|
|
this->id_section_, &is_secn_or_group_unique,
|
|
|
|
§ion_contents);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (parameters->options().print_icf_sections())
|
|
|
|
{
|
|
|
|
if (converged)
|
|
|
|
gold_info(_("%s: ICF Converged after %u iteration(s)"),
|
|
|
|
program_name, num_iterations);
|
|
|
|
else
|
|
|
|
gold_info(_("%s: ICF stopped after %u iteration(s)"),
|
|
|
|
program_name, num_iterations);
|
|
|
|
}
|
|
|
|
|
2009-08-13 03:03:16 +08:00
|
|
|
// Unfold --keep-unique symbols.
|
|
|
|
for (options::String_set::const_iterator p =
|
|
|
|
parameters->options().keep_unique_begin();
|
|
|
|
p != parameters->options().keep_unique_end();
|
|
|
|
++p)
|
|
|
|
{
|
|
|
|
const char* name = p->c_str();
|
|
|
|
Symbol* sym = symtab->lookup(name);
|
2009-08-13 08:03:43 +08:00
|
|
|
if (sym == NULL)
|
|
|
|
{
|
|
|
|
gold_warning(_("Could not find symbol %s to unfold\n"), name);
|
|
|
|
}
|
|
|
|
else if (sym->source() == Symbol::FROM_OBJECT
|
|
|
|
&& !sym->object()->is_dynamic())
|
2009-08-13 03:03:16 +08:00
|
|
|
{
|
|
|
|
Object* obj = sym->object();
|
|
|
|
bool is_ordinary;
|
|
|
|
unsigned int shndx = sym->shndx(&is_ordinary);
|
|
|
|
if (is_ordinary)
|
|
|
|
{
|
|
|
|
this->unfold_section(obj, shndx);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
}
|
|
|
|
|
2009-08-06 04:51:56 +08:00
|
|
|
this->icf_ready();
|
|
|
|
}
|
|
|
|
|
2009-08-13 03:03:16 +08:00
|
|
|
// Unfolds the section denoted by OBJ and SHNDX if folded.
|
|
|
|
|
|
|
|
void
|
|
|
|
Icf::unfold_section(Object* obj, unsigned int shndx)
|
|
|
|
{
|
|
|
|
Section_id secn(obj, shndx);
|
|
|
|
Uniq_secn_id_map::iterator it = this->section_id_.find(secn);
|
|
|
|
if (it == this->section_id_.end())
|
|
|
|
return;
|
|
|
|
unsigned int section_num = it->second;
|
|
|
|
unsigned int kept_section_id = this->kept_section_id_[section_num];
|
|
|
|
if (kept_section_id != section_num)
|
|
|
|
this->kept_section_id_[section_num] = section_num;
|
|
|
|
}
|
|
|
|
|
2009-08-06 04:51:56 +08:00
|
|
|
// This function determines if the section corresponding to the
|
|
|
|
// given object and index is folded based on if the kept section
|
|
|
|
// is different from this section.
|
|
|
|
|
|
|
|
bool
|
|
|
|
Icf::is_section_folded(Object* obj, unsigned int shndx)
|
|
|
|
{
|
|
|
|
Section_id secn(obj, shndx);
|
|
|
|
Uniq_secn_id_map::iterator it = this->section_id_.find(secn);
|
|
|
|
if (it == this->section_id_.end())
|
|
|
|
return false;
|
|
|
|
unsigned int section_num = it->second;
|
|
|
|
unsigned int kept_section_id = this->kept_section_id_[section_num];
|
|
|
|
return kept_section_id != section_num;
|
|
|
|
}
|
|
|
|
|
|
|
|
// This function returns the folded section for the given section.
|
|
|
|
|
|
|
|
Section_id
|
|
|
|
Icf::get_folded_section(Object* dup_obj, unsigned int dup_shndx)
|
|
|
|
{
|
|
|
|
Section_id dup_secn(dup_obj, dup_shndx);
|
|
|
|
Uniq_secn_id_map::iterator it = this->section_id_.find(dup_secn);
|
|
|
|
gold_assert(it != this->section_id_.end());
|
|
|
|
unsigned int section_num = it->second;
|
|
|
|
unsigned int kept_section_id = this->kept_section_id_[section_num];
|
|
|
|
Section_id folded_section = this->id_section_[kept_section_id];
|
|
|
|
return folded_section;
|
|
|
|
}
|
|
|
|
|
|
|
|
} // End of namespace gold.
|