1 // icf.cc -- Identical Code Folding.
3 // Copyright 2009, 2010 Free Software Foundation, Inc.
4 // Written by Sriraman Tallam <tmsriram@google.com>.
6 // This file is part of gold.
8 // This program is free software; you can redistribute it and/or modify
9 // it under the terms of the GNU General Public License as published by
10 // the Free Software Foundation; either version 3 of the License, or
11 // (at your option) any later version.
13 // This program is distributed in the hope that it will be useful,
14 // but WITHOUT ANY WARRANTY; without even the implied warranty of
15 // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 // GNU General Public License for more details.
18 // You should have received a copy of the GNU General Public License
19 // along with this program; if not, write to the Free Software
20 // Foundation, Inc., 51 Franklin Street - Fifth Floor, Boston,
21 // MA 02110-1301, USA.
23 // Identical Code Folding Algorithm
24 // ----------------------------------
25 // Detecting identical functions is done here and the basic algorithm
26 // is as follows. A checksum is computed on each foldable section using
27 // its contents and relocations. If the symbol name corresponding to
28 // a relocation is known it is used to compute the checksum. If the
29 // symbol name is not known the stringified name of the object and the
30 // section number pointed to by the relocation is used. The checksums
31 // are stored as keys in a hash map and a section is identical to some
32 // other section if its checksum is already present in the hash map.
33 // Checksum collisions are handled by using a multimap and explicitly
34 // checking the contents when two sections have the same checksum.
36 // However, two functions A and B with identical text but with
37 // relocations pointing to different foldable sections can be identical if
38 // the corresponding foldable sections to which their relocations point to
39 // turn out to be identical. Hence, this checksumming process must be
40 // done repeatedly until convergence is obtained. Here is an example for
41 // the following case :
43 // int funcA () int funcB ()
45 // return foo(); return goo();
48 // The functions funcA and funcB are identical if functions foo() and
49 // goo() are identical.
51 // Hence, as described above, we repeatedly do the checksumming,
52 // assigning identical functions to the same group, until convergence is
53 // obtained. Now, we have two different ways to do this depending on how
58 // We can start with marking all functions as different and repeatedly do
59 // the checksumming. This has the advantage that we do not need to wait
60 // for convergence. We can stop at any point and correctness will be
61 // guaranteed although not all cases would have been found. However, this
62 // has a problem that some cases can never be found even if it is run until
63 // convergence. Here is an example with mutually recursive functions :
65 // int funcA (int a) int funcB (int a)
67 // if (a == 1) if (a == 1)
68 // return 1; return 1;
69 // return 1 + funcB(a - 1); return 1 + funcA(a - 1);
72 // In this example funcA and funcB are identical and one of them could be
73 // folded into the other. However, if we start with assuming that funcA
74 // and funcB are not identical, the algorithm, even after it is run to
75 // convergence, cannot detect that they are identical. It should be noted
76 // that even if the functions were self-recursive, Algorithm I cannot catch
77 // that they are identical, at least as is.
81 // Here we start with marking all functions as identical and then repeat
82 // the checksumming until convergence. This can detect the above case
83 // mentioned above. It can detect all cases that Algorithm I can and more.
84 // However, the caveat is that it has to be run to convergence. It cannot
85 // be stopped arbitrarily like Algorithm I as correctness cannot be
86 // guaranteed. Algorithm II is not implemented.
88 // Algorithm I is used because experiments show that about three
89 // iterations are more than enough to achieve convergence. Algorithm I can
90 // handle recursive calls if it is changed to use a special common symbol
91 // for recursive relocs. This seems to be the most common case that
92 // Algorithm I could not catch as is. Mutually recursive calls are not
93 // frequent and Algorithm I wins because of its ability to be stopped
96 // Caveat with using function pointers :
97 // ------------------------------------
99 // Programs using function pointer comparisons/checks should use function
100 // folding with caution as the result of such comparisons could be different
101 // when folding takes place. This could lead to unexpected run-time
107 // ICF in safe mode folds only ctors and dtors if their function pointers can
108 // never be taken. Also, for X86-64, safe folding uses the relocation
109 // type to determine if a function's pointer is taken or not and only folds
110 // functions whose pointers are definitely not taken.
112 // Caveat with safe folding :
113 // ------------------------
115 // This applies only to x86_64.
117 // Position independent executables are created from PIC objects (compiled
118 // with -fPIC) and/or PIE objects (compiled with -fPIE). For PIE objects, the
119 // relocation types for function pointer taken and a call are the same.
120 // Now, it is not always possible to tell if an object used in the link of
121 // a pie executable is a PIC object or a PIE object. Hence, for pie
122 // executables, using relocation types to disambiguate function pointers is
123 // currently disabled.
125 // Further, it is not correct to use safe folding to build non-pie
126 // executables using PIC/PIE objects. PIC/PIE objects have different
127 // relocation types for function pointers than non-PIC objects, and the
128 // current implementation of safe folding does not handle those relocation
129 // types. Hence, if used, functions whose pointers are taken could still be
130 // folded causing unpredictable run-time behaviour if the pointers were used
135 // How to run : --icf=[safe|all|none]
136 // Optional parameters : --icf-iterations <num> --print-icf-sections
138 // Performance : Less than 20 % link-time overhead on industry strength
139 // applications. Up to 6 % text size reductions.
146 #include "libiberty.h"
147 #include "demangle.h"
152 // This function determines if a section or a group of identical
153 // sections has unique contents. Such unique sections or groups can be
154 // declared final and need not be processed any further.
156 // ID_SECTION : Vector mapping a section index to a Section_id pair.
157 // IS_SECN_OR_GROUP_UNIQUE : To check if a section or a group of identical
158 // sections is already known to be unique.
159 // SECTION_CONTENTS : Contains the section's text and relocs to sections
160 // that cannot be folded. SECTION_CONTENTS are NULL
161 // implies that this function is being called for the
162 // first time before the first iteration of icf.
165 preprocess_for_unique_sections(const std::vector
<Section_id
>& id_section
,
166 std::vector
<bool>* is_secn_or_group_unique
,
167 std::vector
<std::string
>* section_contents
)
169 Unordered_map
<uint32_t, unsigned int> uniq_map
;
170 std::pair
<Unordered_map
<uint32_t, unsigned int>::iterator
, bool>
173 for (unsigned int i
= 0; i
< id_section
.size(); i
++)
175 if ((*is_secn_or_group_unique
)[i
])
179 Section_id secn
= id_section
[i
];
180 section_size_type plen
;
181 if (section_contents
== NULL
)
183 const unsigned char* contents
;
184 contents
= secn
.first
->section_contents(secn
.second
,
187 cksum
= xcrc32(contents
, plen
, 0xffffffff);
191 const unsigned char* contents_array
= reinterpret_cast
192 <const unsigned char*>((*section_contents
)[i
].c_str());
193 cksum
= xcrc32(contents_array
, (*section_contents
)[i
].length(),
196 uniq_map_insert
= uniq_map
.insert(std::make_pair(cksum
, i
));
197 if (uniq_map_insert
.second
)
199 (*is_secn_or_group_unique
)[i
] = true;
203 (*is_secn_or_group_unique
)[i
] = false;
204 (*is_secn_or_group_unique
)[uniq_map_insert
.first
->second
] = false;
209 // This returns the buffer containing the section's contents, both
210 // text and relocs. Relocs are differentiated as those pointing to
211 // sections that could be folded and those that cannot. Only relocs
212 // pointing to sections that could be folded are recomputed on
213 // subsequent invocations of this function.
215 // FIRST_ITERATION : true if it is the first invocation.
216 // SECN : Section for which contents are desired.
217 // SECTION_NUM : Unique section number of this section.
218 // NUM_TRACKED_RELOCS : Vector reference to store the number of relocs
220 // KEPT_SECTION_ID : Vector which maps folded sections to kept sections.
221 // SECTION_CONTENTS : Store the section's text and relocs to non-ICF
225 get_section_contents(bool first_iteration
,
226 const Section_id
& secn
,
227 unsigned int section_num
,
228 unsigned int* num_tracked_relocs
,
229 Symbol_table
* symtab
,
230 const std::vector
<unsigned int>& kept_section_id
,
231 std::vector
<std::string
>* section_contents
)
233 section_size_type plen
;
234 const unsigned char* contents
= NULL
;
238 contents
= secn
.first
->section_contents(secn
.second
,
243 // The buffer to hold all the contents including relocs. A checksum
244 // is then computed on this buffer.
246 std::string icf_reloc_buffer
;
248 if (num_tracked_relocs
)
249 *num_tracked_relocs
= 0;
251 Icf::Reloc_info_list
& reloc_info_list
=
252 symtab
->icf()->reloc_info_list();
254 Icf::Reloc_info_list::iterator it_reloc_info_list
=
255 reloc_info_list
.find(secn
);
258 icf_reloc_buffer
.clear();
260 // Process relocs and put them into the buffer.
262 if (it_reloc_info_list
!= reloc_info_list
.end())
264 Icf::Sections_reachable_info v
=
265 (it_reloc_info_list
->second
).section_info
;
266 // Stores the information of the symbol pointed to by the reloc.
267 Icf::Symbol_info s
= (it_reloc_info_list
->second
).symbol_info
;
268 // Stores the addend and the symbol value.
269 Icf::Addend_info a
= (it_reloc_info_list
->second
).addend_info
;
270 // Stores the offset of the reloc.
271 Icf::Offset_info o
= (it_reloc_info_list
->second
).offset_info
;
272 Icf::Sections_reachable_info::iterator it_v
= v
.begin();
273 Icf::Symbol_info::iterator it_s
= s
.begin();
274 Icf::Addend_info::iterator it_a
= a
.begin();
275 Icf::Offset_info::iterator it_o
= o
.begin();
277 for (; it_v
!= v
.end(); ++it_v
, ++it_s
, ++it_a
, ++it_o
)
279 // ADDEND_STR stores the symbol value and addend and offset,
280 // each atmost 16 hex digits long. it_a points to a pair
281 // where first is the symbol value and second is the
285 // It would be nice if we could use format macros in inttypes.h
286 // here but there are not in ISO/IEC C++ 1998.
287 snprintf(addend_str
, sizeof(addend_str
), "%llx %llx %llux",
288 static_cast<long long>((*it_a
).first
),
289 static_cast<long long>((*it_a
).second
),
290 static_cast<unsigned long long>(*it_o
));
292 // If the symbol pointed to by the reloc is not in an ordinary
293 // section or if the symbol type is not FROM_OBJECT, then the
295 if (it_v
->first
== NULL
)
299 // If the symbol name is available, use it.
301 buffer
.append((*it_s
)->name());
302 // Append the addend.
303 buffer
.append(addend_str
);
309 Section_id
reloc_secn(it_v
->first
, it_v
->second
);
311 // If this reloc turns back and points to the same section,
312 // like a recursive call, use a special symbol to mark this.
313 if (reloc_secn
.first
== secn
.first
314 && reloc_secn
.second
== secn
.second
)
319 buffer
.append(addend_str
);
324 Icf::Uniq_secn_id_map
& section_id_map
=
325 symtab
->icf()->section_to_int_map();
326 Icf::Uniq_secn_id_map::iterator section_id_map_it
=
327 section_id_map
.find(reloc_secn
);
328 bool is_sym_preemptible
= (*it_s
!= NULL
329 && !(*it_s
)->is_from_dynobj()
330 && !(*it_s
)->is_undefined()
331 && (*it_s
)->is_preemptible());
332 if (!is_sym_preemptible
333 && section_id_map_it
!= section_id_map
.end())
335 // This is a reloc to a section that might be folded.
336 if (num_tracked_relocs
)
337 (*num_tracked_relocs
)++;
339 char kept_section_str
[10];
340 unsigned int secn_id
= section_id_map_it
->second
;
341 snprintf(kept_section_str
, sizeof(kept_section_str
), "%u",
342 kept_section_id
[secn_id
]);
345 buffer
.append("ICF_R");
346 buffer
.append(addend_str
);
348 icf_reloc_buffer
.append(kept_section_str
);
349 // Append the addend.
350 icf_reloc_buffer
.append(addend_str
);
351 icf_reloc_buffer
.append("@");
355 // This is a reloc to a section that cannot be folded.
356 // Process it only in the first iteration.
357 if (!first_iteration
)
360 uint64_t secn_flags
= (it_v
->first
)->section_flags(it_v
->second
);
361 // This reloc points to a merge section. Hash the
362 // contents of this section.
363 if ((secn_flags
& elfcpp::SHF_MERGE
) != 0)
366 (it_v
->first
)->section_entsize(it_v
->second
);
367 long long offset
= it_a
->first
;
369 unsigned long long addend
= it_a
->second
;
370 // Ignoring the addend when it is a negative value. See the
371 // comments in Merged_symbol_value::Value in object.h.
372 if (addend
< 0xffffff00)
373 offset
= offset
+ addend
;
375 section_size_type secn_len
;
376 const unsigned char* str_contents
=
377 (it_v
->first
)->section_contents(it_v
->second
,
380 if ((secn_flags
& elfcpp::SHF_STRINGS
) != 0)
382 // String merge section.
383 const char* str_char
=
384 reinterpret_cast<const char*>(str_contents
);
389 buffer
.append(str_char
);
394 const uint16_t* ptr_16
=
395 reinterpret_cast<const uint16_t*>(str_char
);
396 unsigned int strlen_16
= 0;
397 // Find the NULL character.
398 while(*(ptr_16
+ strlen_16
) != 0)
400 buffer
.append(str_char
, strlen_16
* 2);
405 const uint32_t* ptr_32
=
406 reinterpret_cast<const uint32_t*>(str_char
);
407 unsigned int strlen_32
= 0;
408 // Find the NULL character.
409 while(*(ptr_32
+ strlen_32
) != 0)
411 buffer
.append(str_char
, strlen_32
* 4);
420 // Use the entsize to determine the length.
421 buffer
.append(reinterpret_cast<const
422 char*>(str_contents
),
427 else if ((*it_s
) != NULL
)
429 // If symbol name is available use that.
430 buffer
.append((*it_s
)->name());
431 // Append the addend.
432 buffer
.append(addend_str
);
437 // Symbol name is not available, like for a local symbol,
438 // use object and section id.
439 buffer
.append(it_v
->first
->name());
441 snprintf(secn_id
, sizeof(secn_id
), "%u",it_v
->second
);
442 buffer
.append(secn_id
);
443 // Append the addend.
444 buffer
.append(addend_str
);
453 buffer
.append("Contents = ");
454 buffer
.append(reinterpret_cast<const char*>(contents
), plen
);
455 // Store the section contents that dont change to avoid recomputing
456 // during the next call to this function.
457 (*section_contents
)[section_num
] = buffer
;
461 gold_assert(buffer
.empty());
462 // Reuse the contents computed in the previous iteration.
463 buffer
.append((*section_contents
)[section_num
]);
466 buffer
.append(icf_reloc_buffer
);
470 // This function computes a checksum on each section to detect and form
471 // groups of identical sections. The first iteration does this for all
473 // Further iterations do this only for the kept sections from each group to
474 // determine if larger groups of identical sections could be formed. The
475 // first section in each group is the kept section for that group.
477 // CRC32 is the checksumming algorithm and can have collisions. That is,
478 // two sections with different contents can have the same checksum. Hence,
479 // a multimap is used to maintain more than one group of checksum
480 // identical sections. A section is added to a group only after its
481 // contents are explicitly compared with the kept section of the group.
484 // ITERATION_NUM : Invocation instance of this function.
485 // NUM_TRACKED_RELOCS : Vector reference to store the number of relocs
487 // KEPT_SECTION_ID : Vector which maps folded sections to kept sections.
488 // ID_SECTION : Vector mapping a section to an unique integer.
489 // IS_SECN_OR_GROUP_UNIQUE : To check if a section or a group of identical
490 // sectionsis already known to be unique.
491 // SECTION_CONTENTS : Store the section's text and relocs to non-ICF
495 match_sections(unsigned int iteration_num
,
496 Symbol_table
* symtab
,
497 std::vector
<unsigned int>* num_tracked_relocs
,
498 std::vector
<unsigned int>* kept_section_id
,
499 const std::vector
<Section_id
>& id_section
,
500 std::vector
<bool>* is_secn_or_group_unique
,
501 std::vector
<std::string
>* section_contents
)
503 Unordered_multimap
<uint32_t, unsigned int> section_cksum
;
504 std::pair
<Unordered_multimap
<uint32_t, unsigned int>::iterator
,
505 Unordered_multimap
<uint32_t, unsigned int>::iterator
> key_range
;
506 bool converged
= true;
508 if (iteration_num
== 1)
509 preprocess_for_unique_sections(id_section
,
510 is_secn_or_group_unique
,
513 preprocess_for_unique_sections(id_section
,
514 is_secn_or_group_unique
,
517 std::vector
<std::string
> full_section_contents
;
519 for (unsigned int i
= 0; i
< id_section
.size(); i
++)
521 full_section_contents
.push_back("");
522 if ((*is_secn_or_group_unique
)[i
])
525 Section_id secn
= id_section
[i
];
526 std::string this_secn_contents
;
528 if (iteration_num
== 1)
530 unsigned int num_relocs
= 0;
531 this_secn_contents
= get_section_contents(true, secn
, i
, &num_relocs
,
532 symtab
, (*kept_section_id
),
534 (*num_tracked_relocs
)[i
] = num_relocs
;
538 if ((*kept_section_id
)[i
] != i
)
540 // This section is already folded into something. See
541 // if it should point to a different kept section.
542 unsigned int kept_section
= (*kept_section_id
)[i
];
543 if (kept_section
!= (*kept_section_id
)[kept_section
])
545 (*kept_section_id
)[i
] = (*kept_section_id
)[kept_section
];
549 this_secn_contents
= get_section_contents(false, secn
, i
, NULL
,
550 symtab
, (*kept_section_id
),
554 const unsigned char* this_secn_contents_array
=
555 reinterpret_cast<const unsigned char*>(this_secn_contents
.c_str());
556 cksum
= xcrc32(this_secn_contents_array
, this_secn_contents
.length(),
558 size_t count
= section_cksum
.count(cksum
);
562 // Start a group with this cksum.
563 section_cksum
.insert(std::make_pair(cksum
, i
));
564 full_section_contents
[i
] = this_secn_contents
;
568 key_range
= section_cksum
.equal_range(cksum
);
569 Unordered_multimap
<uint32_t, unsigned int>::iterator it
;
570 // Search all the groups with this cksum for a match.
571 for (it
= key_range
.first
; it
!= key_range
.second
; ++it
)
573 unsigned int kept_section
= it
->second
;
574 if (full_section_contents
[kept_section
].length()
575 != this_secn_contents
.length())
577 if (memcmp(full_section_contents
[kept_section
].c_str(),
578 this_secn_contents
.c_str(),
579 this_secn_contents
.length()) != 0)
581 (*kept_section_id
)[i
] = kept_section
;
585 if (it
== key_range
.second
)
587 // Create a new group for this cksum.
588 section_cksum
.insert(std::make_pair(cksum
, i
));
589 full_section_contents
[i
] = this_secn_contents
;
592 // If there are no relocs to foldable sections do not process
593 // this section any further.
594 if (iteration_num
== 1 && (*num_tracked_relocs
)[i
] == 0)
595 (*is_secn_or_group_unique
)[i
] = true;
601 // During safe icf (--icf=safe), only fold functions that are ctors or dtors.
602 // This function returns true if the mangled function name is a ctor or a
606 is_function_ctor_or_dtor(const char* mangled_func_name
)
608 if ((is_prefix_of("_ZN", mangled_func_name
)
609 || is_prefix_of("_ZZ", mangled_func_name
))
610 && (is_gnu_v3_mangled_ctor(mangled_func_name
)
611 || is_gnu_v3_mangled_dtor(mangled_func_name
)))
618 // This is the main ICF function called in gold.cc. This does the
619 // initialization and calls match_sections repeatedly (twice by default)
620 // which computes the crc checksums and detects identical functions.
623 Icf::find_identical_sections(const Input_objects
* input_objects
,
624 Symbol_table
* symtab
)
626 unsigned int section_num
= 0;
627 std::vector
<unsigned int> num_tracked_relocs
;
628 std::vector
<bool> is_secn_or_group_unique
;
629 std::vector
<std::string
> section_contents
;
630 const Target
& target
= parameters
->target();
632 // Decide which sections are possible candidates first.
634 for (Input_objects::Relobj_iterator p
= input_objects
->relobj_begin();
635 p
!= input_objects
->relobj_end();
638 for (unsigned int i
= 0;i
< (*p
)->shnum(); ++i
)
640 const char* section_name
= (*p
)->section_name(i
).c_str();
641 if (!is_section_foldable_candidate(section_name
))
643 if (!(*p
)->is_section_included(i
))
645 if (parameters
->options().gc_sections()
646 && symtab
->gc()->is_section_garbage(*p
, i
))
648 const char* mangled_func_name
= strrchr(section_name
, '.');
649 gold_assert(mangled_func_name
!= NULL
);
650 // With --icf=safe, check if the mangled function name is a ctor
651 // or a dtor. The mangled function name can be obtained from the
652 // section name by stripping the section prefix.
653 if (parameters
->options().icf_safe_folding()
654 && !is_function_ctor_or_dtor(mangled_func_name
+ 1)
655 && (!target
.can_check_for_function_pointers()
656 || section_has_function_pointers(*p
, i
)))
660 this->id_section_
.push_back(Section_id(*p
, i
));
661 this->section_id_
[Section_id(*p
, i
)] = section_num
;
662 this->kept_section_id_
.push_back(section_num
);
663 num_tracked_relocs
.push_back(0);
664 is_secn_or_group_unique
.push_back(false);
665 section_contents
.push_back("");
670 unsigned int num_iterations
= 0;
672 // Default number of iterations to run ICF is 2.
673 unsigned int max_iterations
= (parameters
->options().icf_iterations() > 0)
674 ? parameters
->options().icf_iterations()
677 bool converged
= false;
679 while (!converged
&& (num_iterations
< max_iterations
))
682 converged
= match_sections(num_iterations
, symtab
,
683 &num_tracked_relocs
, &this->kept_section_id_
,
684 this->id_section_
, &is_secn_or_group_unique
,
688 if (parameters
->options().print_icf_sections())
691 gold_info(_("%s: ICF Converged after %u iteration(s)"),
692 program_name
, num_iterations
);
694 gold_info(_("%s: ICF stopped after %u iteration(s)"),
695 program_name
, num_iterations
);
698 // Unfold --keep-unique symbols.
699 for (options::String_set::const_iterator p
=
700 parameters
->options().keep_unique_begin();
701 p
!= parameters
->options().keep_unique_end();
704 const char* name
= p
->c_str();
705 Symbol
* sym
= symtab
->lookup(name
);
708 gold_warning(_("Could not find symbol %s to unfold\n"), name
);
710 else if (sym
->source() == Symbol::FROM_OBJECT
711 && !sym
->object()->is_dynamic())
713 Object
* obj
= sym
->object();
715 unsigned int shndx
= sym
->shndx(&is_ordinary
);
718 this->unfold_section(obj
, shndx
);
727 // Unfolds the section denoted by OBJ and SHNDX if folded.
730 Icf::unfold_section(Object
* obj
, unsigned int shndx
)
732 Section_id
secn(obj
, shndx
);
733 Uniq_secn_id_map::iterator it
= this->section_id_
.find(secn
);
734 if (it
== this->section_id_
.end())
736 unsigned int section_num
= it
->second
;
737 unsigned int kept_section_id
= this->kept_section_id_
[section_num
];
738 if (kept_section_id
!= section_num
)
739 this->kept_section_id_
[section_num
] = section_num
;
742 // This function determines if the section corresponding to the
743 // given object and index is folded based on if the kept section
744 // is different from this section.
747 Icf::is_section_folded(Object
* obj
, unsigned int shndx
)
749 Section_id
secn(obj
, shndx
);
750 Uniq_secn_id_map::iterator it
= this->section_id_
.find(secn
);
751 if (it
== this->section_id_
.end())
753 unsigned int section_num
= it
->second
;
754 unsigned int kept_section_id
= this->kept_section_id_
[section_num
];
755 return kept_section_id
!= section_num
;
758 // This function returns the folded section for the given section.
761 Icf::get_folded_section(Object
* dup_obj
, unsigned int dup_shndx
)
763 Section_id
dup_secn(dup_obj
, dup_shndx
);
764 Uniq_secn_id_map::iterator it
= this->section_id_
.find(dup_secn
);
765 gold_assert(it
!= this->section_id_
.end());
766 unsigned int section_num
= it
->second
;
767 unsigned int kept_section_id
= this->kept_section_id_
[section_num
];
768 Section_id folded_section
= this->id_section_
[kept_section_id
];
769 return folded_section
;
772 } // End of namespace gold.