// icf.cc -- Identical Code Folding.
//
-// Copyright (C) 2009-2016 Free Software Foundation, Inc.
+// Copyright (C) 2009-2020 Free Software Foundation, Inc.
// Written by Sriraman Tallam <tmsriram@google.com>.
// This file is part of gold.
// folded causing unpredictable run-time behaviour if the pointers were used
// in comparisons.
//
+// Notes regarding C++ exception handling :
+// --------------------------------------
+//
+// It is possible for two sections to have identical text, identical
+// relocations, but different exception handling metadata (unwind
+// information in the .eh_frame section, and/or handler information in
+// a .gcc_except_table section). Thus, if a foldable section is
+// referenced from a .eh_frame FDE, we must include in its checksum
+// the contents of that FDE as well as of the CIE that the FDE refers
+// to. The CIE and FDE in turn probably contain relocations to the
+// personality routine and LSDA, which are handled like any other
+// relocation for ICF purposes. This logic is helped by the fact that
+// gcc with -ffunction-sections puts each function's LSDA in its own
+// .gcc_except_table.<functionname> section. Given sections for two
+// functions with nontrivial exception handling logic, we will
+// determine on the first iteration that their .gcc_except_table
+// sections are identical and can be folded, and on the second
+// iteration that their .text and .eh_frame contents (including the
+// now-merged .gcc_except_table relocations for the LSDA) are
+// identical and can be folded.
//
//
// How to run : --icf=[safe|all|none]
#include "elfcpp.h"
#include "int_encoding.h"
+#include <limits>
+
namespace gold
{
// subsequent invocations of this function.
// Parameters :
// FIRST_ITERATION : true if it is the first invocation.
+// FIXED_CACHE : String that stores the portion of the result that
+// does not change from iteration to iteration;
+// written if first_iteration is true, read if it's false.
// SECN : Section for which contents are desired.
-// SECTION_NUM : Unique section number of this section.
+// SELF_SECN : Relocations that target this section will be
+// considered "relocations to self" so that recursive
+// functions can be folded. Should normally be the
+// same as `secn` except when processing extra identity
+// regions.
// 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.
+// START_OFFSET : Only consider the part of the section at and after
+// this offset.
+// END_OFFSET : Only consider the part of the section before this
+// offset.
static std::string
get_section_contents(bool first_iteration,
+ std::string* fixed_cache,
const Section_id& secn,
- unsigned int section_num,
+ const Section_id& self_secn,
unsigned int* num_tracked_relocs,
Symbol_table* symtab,
const std::vector<unsigned int>& kept_section_id,
- std::vector<std::string>* section_contents)
+ section_offset_type start_offset = 0,
+ section_offset_type end_offset =
+ std::numeric_limits<section_offset_type>::max())
{
- // 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);
-
section_size_type plen;
const unsigned char* contents = NULL;
if (first_iteration)
std::string buffer;
std::string icf_reloc_buffer;
- if (num_tracked_relocs)
- *num_tracked_relocs = 0;
-
Icf::Reloc_info_list& reloc_info_list =
symtab->icf()->reloc_info_list();
for (; it_v != v.end(); ++it_v, ++it_s, ++it_a, ++it_o, ++it_addend_size)
{
+ Symbol* gsym = *it_s;
+ bool is_section_symbol = false;
+
+ // Ignore relocations outside the region we were told to look at
+ if (static_cast<section_offset_type>(*it_o) < start_offset
+ || static_cast<section_offset_type>(*it_o) >= end_offset)
+ continue;
+
+ // A -1 value in the symbol vector indicates a local section symbol.
+ if (gsym == reinterpret_cast<Symbol*>(-1))
+ {
+ is_section_symbol = true;
+ gsym = NULL;
+ }
+
if (first_iteration
&& it_v->first != NULL)
{
// 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",
+ snprintf(addend_str, sizeof(addend_str), "%llx %llx %llx",
static_cast<long long>((*it_a).first),
static_cast<long long>((*it_a).second),
- static_cast<unsigned long long>(*it_o));
+ static_cast<unsigned long long>(*it_o - start_offset));
// 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
if (first_iteration)
{
// If the symbol name is available, use it.
- if ((*it_s) != NULL)
- buffer.append((*it_s)->name());
+ if (gsym != NULL)
+ buffer.append(gsym->name());
// Append the addend.
buffer.append(addend_str);
buffer.append("@");
// 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 (reloc_secn.first == self_secn.first
+ && reloc_secn.second == self_secn.second)
{
if (first_iteration)
{
symtab->icf()->section_to_int_map();
Icf::Uniq_secn_id_map::iterator section_id_map_it =
section_id_map.find(reloc_secn);
- bool is_sym_preemptible = (*it_s != NULL
- && !(*it_s)->is_from_dynobj()
- && !(*it_s)->is_undefined()
- && (*it_s)->is_preemptible());
+ bool is_sym_preemptible = (gsym != NULL
+ && !gsym->is_from_dynobj()
+ && !gsym->is_undefined()
+ && gsym->is_preemptible());
if (!is_sym_preemptible
&& section_id_map_it != section_id_map.end())
{
uint64_t entsize =
(it_v->first)->section_entsize(it_v->second);
long long offset = it_a->first;
- // Handle SHT_RELA and SHT_REL addends, only one of these
- // addends exists.
- // Get the SHT_RELA addend. For RELA relocations, we have
- // the addend from the relocation.
- uint64_t reloc_addend_value = it_a->second;
-
- // Handle SHT_REL addends.
- // For REL relocations, we need to fetch the addend from the
- // section contents.
- const unsigned char* reloc_addend_ptr =
- contents + static_cast<unsigned long long>(*it_o);
-
- // Update the addend value with the SHT_REL addend if
- // available.
- get_rel_addend(reloc_addend_ptr, *it_addend_size,
- &reloc_addend_value);
-
- // Ignore the addend when it is a negative value. See the
- // comments in Merged_symbol_value::value in object.h.
- if (reloc_addend_value < 0xffffff00)
- offset = offset + reloc_addend_value;
+
+ // Handle SHT_RELA and SHT_REL addends. Only one of these
+ // addends exists. When pointing to a merge section, the
+ // addend only matters if it's relative to a section
+ // symbol. In order to unambiguously identify the target
+ // of the relocation, the compiler (and assembler) must use
+ // a local non-section symbol unless Symbol+Addend does in
+ // fact point directly to the target. (In other words,
+ // a bias for a pc-relative reference or a non-zero based
+ // access forces the use of a local symbol, and the addend
+ // is used only to provide that bias.)
+ uint64_t reloc_addend_value = 0;
+ if (is_section_symbol)
+ {
+ // Get the SHT_RELA addend. For RELA relocations,
+ // we have the addend from the relocation.
+ reloc_addend_value = it_a->second;
+
+ // Handle SHT_REL addends.
+ // For REL relocations, we need to fetch the addend
+ // from the section contents.
+ const unsigned char* reloc_addend_ptr =
+ contents + static_cast<unsigned long long>(*it_o);
+
+ // Update the addend value with the SHT_REL addend if
+ // available.
+ get_rel_addend(reloc_addend_ptr, *it_addend_size,
+ &reloc_addend_value);
+
+ // Ignore the addend when it is a negative value.
+ // See the comments in Merged_symbol_value::value
+ // in object.h.
+ if (reloc_addend_value < 0xffffff00)
+ offset = offset + reloc_addend_value;
+ }
section_size_type secn_len;
}
buffer.append("@");
}
- else if ((*it_s) != NULL)
+ else if (gsym != NULL)
{
// If symbol name is available use that.
- buffer.append((*it_s)->name());
+ buffer.append(gsym->name());
// Append the addend.
buffer.append(addend_str);
buffer.append("@");
if (first_iteration)
{
buffer.append("Contents = ");
- buffer.append(reinterpret_cast<const char*>(contents), plen);
+
+ const unsigned char* slice_end =
+ contents + std::min<section_offset_type>(plen, end_offset);
+
+ if (contents + start_offset < slice_end)
+ {
+ buffer.append(reinterpret_cast<const char*>(contents + start_offset),
+ slice_end - (contents + start_offset));
+ }
+ }
+
+ // Add any extra identity regions.
+ std::pair<Icf::Extra_identity_list::const_iterator,
+ Icf::Extra_identity_list::const_iterator>
+ extra_range = symtab->icf()->extra_identity_list().equal_range(secn);
+ for (Icf::Extra_identity_list::const_iterator it_ext = extra_range.first;
+ it_ext != extra_range.second; ++it_ext)
+ {
+ std::string external_fixed;
+ std::string external_all =
+ get_section_contents(first_iteration, &external_fixed,
+ it_ext->second.section, self_secn,
+ num_tracked_relocs, symtab,
+ kept_section_id, it_ext->second.offset,
+ it_ext->second.offset + it_ext->second.length);
+ buffer.append(external_fixed);
+ icf_reloc_buffer.append(external_all, external_fixed.length(),
+ std::string::npos);
+ }
+
+ if (first_iteration)
+ {
// Store the section contents that don't change to avoid recomputing
// during the next call to this function.
- (*section_contents)[section_num] = buffer;
+ *fixed_cache = buffer;
}
else
{
gold_assert(buffer.empty());
+
// Reuse the contents computed in the previous iteration.
- buffer.append((*section_contents)[section_num]);
+ buffer.append(*fixed_cache);
}
buffer.append(icf_reloc_buffer);
continue;
Section_id secn = id_section[i];
+
+ // 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);
+
std::string this_secn_contents;
uint32_t cksum;
+ std::string* this_secn_cache = &((*section_contents)[i]);
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);
+ this_secn_contents = get_section_contents(true, this_secn_cache,
+ secn, secn, &num_relocs,
+ symtab, (*kept_section_id));
(*num_tracked_relocs)[i] = num_relocs;
}
else
// This section is already folded into something.
continue;
}
- this_secn_contents = get_section_contents(false, secn, i, NULL,
- symtab, (*kept_section_id),
- section_contents);
+ this_secn_contents = get_section_contents(false, this_secn_cache,
+ secn, secn, NULL,
+ symtab, (*kept_section_id));
}
const unsigned char* this_secn_contents_array =
return false;
}
+// Iterate through the .eh_frame section that has index
+// `ehframe_shndx` in `object`, adding entries to extra_identity_list_
+// that will cause the contents of each FDE and its CIE to be included
+// in the logical ICF identity of the function that the FDE refers to.
+
+bool
+Icf::add_ehframe_links(Relobj* object, unsigned int ehframe_shndx,
+ Reloc_info& relocs)
+{
+ section_size_type contents_len;
+ const unsigned char* pcontents = object->section_contents(ehframe_shndx,
+ &contents_len,
+ false);
+ const unsigned char* p = pcontents;
+ const unsigned char* pend = pcontents + contents_len;
+
+ Sections_reachable_info::iterator it_target = relocs.section_info.begin();
+ Sections_reachable_info::iterator it_target_end = relocs.section_info.end();
+ Offset_info::iterator it_offset = relocs.offset_info.begin();
+ Offset_info::iterator it_offset_end = relocs.offset_info.end();
+
+ // Maps section offset to the length of the CIE defined at that offset.
+ typedef Unordered_map<section_offset_type, section_size_type> Cie_map;
+ Cie_map cies;
+
+ uint32_t (*read_swap_32)(const unsigned char*);
+ if (object->is_big_endian())
+ read_swap_32 = &elfcpp::Swap<32, true>::readval;
+ else
+ read_swap_32 = &elfcpp::Swap<32, false>::readval;
+
+ // TODO: The logic for parsing the CIE/FDE framing is copied from
+ // Eh_frame::do_add_ehframe_input_section() and might want to be
+ // factored into a shared helper function.
+ while (p < pend)
+ {
+ if (pend - p < 4)
+ return false;
+
+ unsigned int len = read_swap_32(p);
+ p += 4;
+ if (len == 0)
+ {
+ // We should only find a zero-length entry at the end of the
+ // section.
+ if (p < pend)
+ return false;
+ break;
+ }
+ // We don't support a 64-bit .eh_frame.
+ if (len == 0xffffffff)
+ return false;
+ if (static_cast<unsigned int>(pend - p) < len)
+ return false;
+
+ const unsigned char* const pentend = p + len;
+
+ if (pend - p < 4)
+ return false;
+
+ unsigned int id = read_swap_32(p);
+ p += 4;
+
+ if (id == 0)
+ {
+ // CIE.
+ cies.insert(std::make_pair(p - pcontents, len - 4));
+ }
+ else
+ {
+ // FDE.
+ Cie_map::const_iterator it;
+ it = cies.find((p - pcontents) - (id - 4));
+ if (it == cies.end())
+ return false;
+
+ // Figure out which section this FDE refers into. The word at `p`
+ // is an address, and we expect to see a relocation there. If not,
+ // this FDE isn't ICF-relevant.
+ while (it_offset != it_offset_end
+ && it_target != it_target_end
+ && static_cast<ptrdiff_t>(*it_offset) < (p - pcontents))
+ {
+ ++it_offset;
+ ++it_target;
+ }
+ if (it_offset != it_offset_end
+ && it_target != it_target_end
+ && static_cast<ptrdiff_t>(*it_offset) == (p - pcontents))
+ {
+ // Found a reloc. Add this FDE and its CIE as extra identity
+ // info for the section it refers to.
+ Extra_identity_info rec_fde = {Section_id(object, ehframe_shndx),
+ p - pcontents, len - 4};
+ Extra_identity_info rec_cie = {Section_id(object, ehframe_shndx),
+ it->first, it->second};
+ extra_identity_list_.insert(std::make_pair(*it_target, rec_fde));
+ extra_identity_list_.insert(std::make_pair(*it_target, rec_cie));
+ }
+ }
+
+ p = pentend;
+ }
+
+ return true;
+}
+
// This is the main ICF function called in gold.cc. This does the
-// initialization and calls match_sections repeatedly (twice by default)
+// initialization and calls match_sections repeatedly (thrice by default)
// which computes the crc checksums and detects identical functions.
void
// 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);
+ std::vector<unsigned int> eh_frame_ind;
- for (unsigned int i = 0;i < (*p)->shnum(); ++i)
+ for (unsigned int i = 0; i < (*p)->shnum(); ++i)
{
const std::string section_name = (*p)->section_name(i);
if (!is_section_foldable_candidate(section_name))
- continue;
+ {
+ if (is_prefix_of(".eh_frame", section_name.c_str()))
+ eh_frame_ind.push_back(i);
+ continue;
+ }
+
if (!(*p)->is_section_included(i))
continue;
if (parameters->options().gc_sections()
section_contents.push_back("");
section_num++;
}
+
+ for (std::vector<unsigned int>::iterator it_eh_ind = eh_frame_ind.begin();
+ it_eh_ind != eh_frame_ind.end(); ++it_eh_ind)
+ {
+ // gc_process_relocs() recorded relocations for this
+ // section even though we can't fold it. We need to
+ // use those relocations to associate other foldable
+ // sections with the FDEs and CIEs that are relevant
+ // to them, so we can avoid merging sections that
+ // don't have identical exception-handling behavior.
+
+ Section_id sect(*p, *it_eh_ind);
+ Reloc_info_list::iterator it_rel = this->reloc_info_list().find(sect);
+ if (it_rel != this->reloc_info_list().end())
+ {
+ if (!add_ehframe_links(*p, *it_eh_ind, it_rel->second))
+ {
+ gold_warning(_("could not parse eh_frame section %s(%s); ICF "
+ "might not preserve exception handling "
+ "behavior"),
+ (*p)->name().c_str(),
+ (*p)->section_name(*it_eh_ind).c_str());
+ }
+ }
+ }
}
unsigned int num_iterations = 0;
- // Default number of iterations to run ICF is 2.
+ // Default number of iterations to run ICF is 3.
unsigned int max_iterations = (parameters->options().icf_iterations() > 0)
? parameters->options().icf_iterations()
- : 2;
+ : 3;
bool converged = false;
projects / deliverable / binutils-gdb.git / blobdiff