2011-06-28 Tristan Gingold <gingold@adacore.com>
[deliverable/binutils-gdb.git] / gold / icf.cc
CommitLineData
ef15dade
ST
1// icf.cc -- Identical Code Folding.
2//
55a2bb35 3// Copyright 2009, 2010 Free Software Foundation, Inc.
ef15dade
ST
4// Written by Sriraman Tallam <tmsriram@google.com>.
5
6// This file is part of gold.
7
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.
12
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.
17
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.
22
23// Identical Code Folding Algorithm
24// ----------------------------------
25// Detecting identical functions is done here and the basic algorithm
55a2bb35 26// is as follows. A checksum is computed on each foldable section using
ef15dade
ST
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.
35//
36// However, two functions A and B with identical text but with
55a2bb35
ST
37// relocations pointing to different foldable sections can be identical if
38// the corresponding foldable sections to which their relocations point to
ef15dade
ST
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 :
42//
43// int funcA () int funcB ()
44// { {
45// return foo(); return goo();
46// } }
47//
48// The functions funcA and funcB are identical if functions foo() and
49// goo() are identical.
50//
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
54// we initialize.
55//
56// Algorithm I :
57// -----------
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 :
64//
65// int funcA (int a) int funcB (int a)
66// { {
67// if (a == 1) if (a == 1)
68// return 1; return 1;
69// return 1 + funcB(a - 1); return 1 + funcA(a - 1);
70// } }
71//
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.
78//
79// Algorithm II :
80// ------------
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.
87//
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
94// arbitrarily.
95//
96// Caveat with using function pointers :
97// ------------------------------------
98//
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
102// behaviour.
103//
21bb3914
ST
104// Safe Folding :
105// ------------
106//
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.
111//
112// Caveat with safe folding :
113// ------------------------
114//
115// This applies only to x86_64.
116//
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.
124//
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
131// in comparisons.
132//
133//
ef15dade 134//
55a2bb35 135// How to run : --icf=[safe|all|none]
ef15dade
ST
136// Optional parameters : --icf-iterations <num> --print-icf-sections
137//
138// Performance : Less than 20 % link-time overhead on industry strength
139// applications. Up to 6 % text size reductions.
140
141#include "gold.h"
142#include "object.h"
143#include "gc.h"
144#include "icf.h"
145#include "symtab.h"
146#include "libiberty.h"
032ce4e9 147#include "demangle.h"
41cbeecc
ST
148#include "elfcpp.h"
149#include "int_encoding.h"
ef15dade
ST
150
151namespace gold
152{
153
154// This function determines if a section or a group of identical
155// sections has unique contents. Such unique sections or groups can be
156// declared final and need not be processed any further.
157// Parameters :
158// ID_SECTION : Vector mapping a section index to a Section_id pair.
159// IS_SECN_OR_GROUP_UNIQUE : To check if a section or a group of identical
160// sections is already known to be unique.
161// SECTION_CONTENTS : Contains the section's text and relocs to sections
162// that cannot be folded. SECTION_CONTENTS are NULL
163// implies that this function is being called for the
164// first time before the first iteration of icf.
165
166static void
167preprocess_for_unique_sections(const std::vector<Section_id>& id_section,
168 std::vector<bool>* is_secn_or_group_unique,
169 std::vector<std::string>* section_contents)
170{
171 Unordered_map<uint32_t, unsigned int> uniq_map;
172 std::pair<Unordered_map<uint32_t, unsigned int>::iterator, bool>
173 uniq_map_insert;
174
175 for (unsigned int i = 0; i < id_section.size(); i++)
176 {
177 if ((*is_secn_or_group_unique)[i])
178 continue;
179
180 uint32_t cksum;
181 Section_id secn = id_section[i];
182 section_size_type plen;
183 if (section_contents == NULL)
184 {
5f9bcf58
CC
185 // Lock the object so we can read from it. This is only called
186 // single-threaded from queue_middle_tasks, so it is OK to lock.
187 // Unfortunately we have no way to pass in a Task token.
188 const Task* dummy_task = reinterpret_cast<const Task*>(-1);
189 Task_lock_obj<Object> tl(dummy_task, secn.first);
ef15dade
ST
190 const unsigned char* contents;
191 contents = secn.first->section_contents(secn.second,
192 &plen,
193 false);
194 cksum = xcrc32(contents, plen, 0xffffffff);
195 }
196 else
197 {
198 const unsigned char* contents_array = reinterpret_cast
199 <const unsigned char*>((*section_contents)[i].c_str());
200 cksum = xcrc32(contents_array, (*section_contents)[i].length(),
201 0xffffffff);
202 }
203 uniq_map_insert = uniq_map.insert(std::make_pair(cksum, i));
204 if (uniq_map_insert.second)
205 {
206 (*is_secn_or_group_unique)[i] = true;
207 }
208 else
209 {
210 (*is_secn_or_group_unique)[i] = false;
211 (*is_secn_or_group_unique)[uniq_map_insert.first->second] = false;
212 }
213 }
214}
215
216// This returns the buffer containing the section's contents, both
217// text and relocs. Relocs are differentiated as those pointing to
218// sections that could be folded and those that cannot. Only relocs
219// pointing to sections that could be folded are recomputed on
220// subsequent invocations of this function.
221// Parameters :
222// FIRST_ITERATION : true if it is the first invocation.
223// SECN : Section for which contents are desired.
224// SECTION_NUM : Unique section number of this section.
225// NUM_TRACKED_RELOCS : Vector reference to store the number of relocs
226// to ICF sections.
227// KEPT_SECTION_ID : Vector which maps folded sections to kept sections.
228// SECTION_CONTENTS : Store the section's text and relocs to non-ICF
229// sections.
230
231static std::string
232get_section_contents(bool first_iteration,
233 const Section_id& secn,
234 unsigned int section_num,
235 unsigned int* num_tracked_relocs,
236 Symbol_table* symtab,
237 const std::vector<unsigned int>& kept_section_id,
238 std::vector<std::string>* section_contents)
239{
880473a6
DK
240 // Lock the object so we can read from it. This is only called
241 // single-threaded from queue_middle_tasks, so it is OK to lock.
242 // Unfortunately we have no way to pass in a Task token.
243 const Task* dummy_task = reinterpret_cast<const Task*>(-1);
244 Task_lock_obj<Object> tl(dummy_task, secn.first);
245
ef15dade
ST
246 section_size_type plen;
247 const unsigned char* contents = NULL;
ef15dade 248 if (first_iteration)
880473a6 249 contents = secn.first->section_contents(secn.second, &plen, false);
ef15dade
ST
250
251 // The buffer to hold all the contents including relocs. A checksum
252 // is then computed on this buffer.
253 std::string buffer;
254 std::string icf_reloc_buffer;
255
256 if (num_tracked_relocs)
257 *num_tracked_relocs = 0;
258
b487ad64
ST
259 Icf::Reloc_info_list& reloc_info_list =
260 symtab->icf()->reloc_info_list();
ef15dade 261
b487ad64
ST
262 Icf::Reloc_info_list::iterator it_reloc_info_list =
263 reloc_info_list.find(secn);
ef15dade
ST
264
265 buffer.clear();
266 icf_reloc_buffer.clear();
267
268 // Process relocs and put them into the buffer.
269
b487ad64 270 if (it_reloc_info_list != reloc_info_list.end())
ef15dade 271 {
b487ad64
ST
272 Icf::Sections_reachable_info v =
273 (it_reloc_info_list->second).section_info;
ef38fd8a 274 // Stores the information of the symbol pointed to by the reloc.
b487ad64 275 Icf::Symbol_info s = (it_reloc_info_list->second).symbol_info;
ef38fd8a 276 // Stores the addend and the symbol value.
b487ad64 277 Icf::Addend_info a = (it_reloc_info_list->second).addend_info;
ef38fd8a 278 // Stores the offset of the reloc.
b487ad64 279 Icf::Offset_info o = (it_reloc_info_list->second).offset_info;
41cbeecc
ST
280 Icf::Reloc_addend_size_info reloc_addend_size_info =
281 (it_reloc_info_list->second).reloc_addend_size_info;
b487ad64 282 Icf::Sections_reachable_info::iterator it_v = v.begin();
ef15dade
ST
283 Icf::Symbol_info::iterator it_s = s.begin();
284 Icf::Addend_info::iterator it_a = a.begin();
b487ad64 285 Icf::Offset_info::iterator it_o = o.begin();
41cbeecc
ST
286 Icf::Reloc_addend_size_info::iterator it_addend_size =
287 reloc_addend_size_info.begin();
ef15dade 288
41cbeecc 289 for (; it_v != v.end(); ++it_v, ++it_s, ++it_a, ++it_o, ++it_addend_size)
ef15dade 290 {
b487ad64 291 // ADDEND_STR stores the symbol value and addend and offset,
9b547ce6 292 // each at most 16 hex digits long. it_a points to a pair
ef15dade
ST
293 // where first is the symbol value and second is the
294 // addend.
b487ad64 295 char addend_str[50];
bb0bfe4f
DK
296
297 // It would be nice if we could use format macros in inttypes.h
298 // here but there are not in ISO/IEC C++ 1998.
299 snprintf(addend_str, sizeof(addend_str), "%llx %llx %llux",
300 static_cast<long long>((*it_a).first),
301 static_cast<long long>((*it_a).second),
302 static_cast<unsigned long long>(*it_o));
ef38fd8a
ST
303
304 // If the symbol pointed to by the reloc is not in an ordinary
305 // section or if the symbol type is not FROM_OBJECT, then the
306 // object is NULL.
307 if (it_v->first == NULL)
308 {
309 if (first_iteration)
310 {
311 // If the symbol name is available, use it.
312 if ((*it_s) != NULL)
313 buffer.append((*it_s)->name());
314 // Append the addend.
315 buffer.append(addend_str);
316 buffer.append("@");
317 }
318 continue;
319 }
320
ef15dade
ST
321 Section_id reloc_secn(it_v->first, it_v->second);
322
323 // If this reloc turns back and points to the same section,
324 // like a recursive call, use a special symbol to mark this.
325 if (reloc_secn.first == secn.first
326 && reloc_secn.second == secn.second)
327 {
328 if (first_iteration)
329 {
330 buffer.append("R");
331 buffer.append(addend_str);
332 buffer.append("@");
333 }
334 continue;
335 }
336 Icf::Uniq_secn_id_map& section_id_map =
337 symtab->icf()->section_to_int_map();
338 Icf::Uniq_secn_id_map::iterator section_id_map_it =
339 section_id_map.find(reloc_secn);
ce97fa81
ST
340 bool is_sym_preemptible = (*it_s != NULL
341 && !(*it_s)->is_from_dynobj()
342 && !(*it_s)->is_undefined()
343 && (*it_s)->is_preemptible());
344 if (!is_sym_preemptible
345 && section_id_map_it != section_id_map.end())
ef15dade
ST
346 {
347 // This is a reloc to a section that might be folded.
348 if (num_tracked_relocs)
349 (*num_tracked_relocs)++;
350
351 char kept_section_str[10];
352 unsigned int secn_id = section_id_map_it->second;
353 snprintf(kept_section_str, sizeof(kept_section_str), "%u",
354 kept_section_id[secn_id]);
355 if (first_iteration)
356 {
357 buffer.append("ICF_R");
358 buffer.append(addend_str);
359 }
360 icf_reloc_buffer.append(kept_section_str);
361 // Append the addend.
362 icf_reloc_buffer.append(addend_str);
363 icf_reloc_buffer.append("@");
364 }
365 else
366 {
367 // This is a reloc to a section that cannot be folded.
368 // Process it only in the first iteration.
369 if (!first_iteration)
370 continue;
371
372 uint64_t secn_flags = (it_v->first)->section_flags(it_v->second);
373 // This reloc points to a merge section. Hash the
374 // contents of this section.
c95e9f27
ST
375 if ((secn_flags & elfcpp::SHF_MERGE) != 0
376 && parameters->target().can_icf_inline_merge_sections ())
ef15dade
ST
377 {
378 uint64_t entsize =
379 (it_v->first)->section_entsize(it_v->second);
ce97fa81
ST
380 long long offset = it_a->first;
381
382 unsigned long long addend = it_a->second;
383 // Ignoring the addend when it is a negative value. See the
384 // comments in Merged_symbol_value::Value in object.h.
385 if (addend < 0xffffff00)
386 offset = offset + addend;
387
41cbeecc
ST
388 // For SHT_REL relocation sections, the addend is stored in the
389 // text section at the relocation offset.
390 uint64_t reloc_addend_value = 0;
391 const unsigned char* reloc_addend_ptr =
392 contents + static_cast<unsigned long long>(*it_o);
393 switch(*it_addend_size)
394 {
395 case 0:
396 {
397 break;
398 }
399 case 1:
400 {
401 reloc_addend_value =
402 read_from_pointer<8>(reloc_addend_ptr);
403 break;
404 }
405 case 2:
406 {
407 reloc_addend_value =
408 read_from_pointer<16>(reloc_addend_ptr);
409 break;
410 }
411 case 4:
412 {
413 reloc_addend_value =
414 read_from_pointer<32>(reloc_addend_ptr);
415 break;
416 }
417 case 8:
418 {
419 reloc_addend_value =
420 read_from_pointer<64>(reloc_addend_ptr);
421 break;
422 }
423 default:
424 gold_unreachable();
425 }
426 offset = offset + reloc_addend_value;
427
ef15dade
ST
428 section_size_type secn_len;
429 const unsigned char* str_contents =
430 (it_v->first)->section_contents(it_v->second,
431 &secn_len,
432 false) + offset;
433 if ((secn_flags & elfcpp::SHF_STRINGS) != 0)
434 {
435 // String merge section.
436 const char* str_char =
437 reinterpret_cast<const char*>(str_contents);
438 switch(entsize)
439 {
440 case 1:
441 {
442 buffer.append(str_char);
443 break;
444 }
445 case 2:
446 {
447 const uint16_t* ptr_16 =
448 reinterpret_cast<const uint16_t*>(str_char);
449 unsigned int strlen_16 = 0;
450 // Find the NULL character.
451 while(*(ptr_16 + strlen_16) != 0)
452 strlen_16++;
453 buffer.append(str_char, strlen_16 * 2);
454 }
455 break;
456 case 4:
457 {
458 const uint32_t* ptr_32 =
459 reinterpret_cast<const uint32_t*>(str_char);
460 unsigned int strlen_32 = 0;
461 // Find the NULL character.
462 while(*(ptr_32 + strlen_32) != 0)
463 strlen_32++;
464 buffer.append(str_char, strlen_32 * 4);
465 }
466 break;
467 default:
468 gold_unreachable();
469 }
470 }
471 else
472 {
473 // Use the entsize to determine the length.
474 buffer.append(reinterpret_cast<const
475 char*>(str_contents),
476 entsize);
477 }
d62d0f5f 478 buffer.append("@");
ef15dade
ST
479 }
480 else if ((*it_s) != NULL)
481 {
482 // If symbol name is available use that.
ef38fd8a 483 buffer.append((*it_s)->name());
ef15dade
ST
484 // Append the addend.
485 buffer.append(addend_str);
486 buffer.append("@");
487 }
488 else
489 {
490 // Symbol name is not available, like for a local symbol,
491 // use object and section id.
492 buffer.append(it_v->first->name());
493 char secn_id[10];
494 snprintf(secn_id, sizeof(secn_id), "%u",it_v->second);
495 buffer.append(secn_id);
496 // Append the addend.
497 buffer.append(addend_str);
498 buffer.append("@");
499 }
500 }
501 }
502 }
503
504 if (first_iteration)
505 {
506 buffer.append("Contents = ");
507 buffer.append(reinterpret_cast<const char*>(contents), plen);
508 // Store the section contents that dont change to avoid recomputing
509 // during the next call to this function.
510 (*section_contents)[section_num] = buffer;
511 }
512 else
513 {
514 gold_assert(buffer.empty());
515 // Reuse the contents computed in the previous iteration.
516 buffer.append((*section_contents)[section_num]);
517 }
518
519 buffer.append(icf_reloc_buffer);
520 return buffer;
521}
522
523// This function computes a checksum on each section to detect and form
524// groups of identical sections. The first iteration does this for all
525// sections.
526// Further iterations do this only for the kept sections from each group to
527// determine if larger groups of identical sections could be formed. The
528// first section in each group is the kept section for that group.
529//
530// CRC32 is the checksumming algorithm and can have collisions. That is,
531// two sections with different contents can have the same checksum. Hence,
532// a multimap is used to maintain more than one group of checksum
533// identical sections. A section is added to a group only after its
534// contents are explicitly compared with the kept section of the group.
535//
536// Parameters :
537// ITERATION_NUM : Invocation instance of this function.
538// NUM_TRACKED_RELOCS : Vector reference to store the number of relocs
539// to ICF sections.
540// KEPT_SECTION_ID : Vector which maps folded sections to kept sections.
541// ID_SECTION : Vector mapping a section to an unique integer.
542// IS_SECN_OR_GROUP_UNIQUE : To check if a section or a group of identical
9b547ce6 543// sections is already known to be unique.
ef15dade
ST
544// SECTION_CONTENTS : Store the section's text and relocs to non-ICF
545// sections.
546
547static bool
548match_sections(unsigned int iteration_num,
549 Symbol_table* symtab,
550 std::vector<unsigned int>* num_tracked_relocs,
551 std::vector<unsigned int>* kept_section_id,
552 const std::vector<Section_id>& id_section,
553 std::vector<bool>* is_secn_or_group_unique,
554 std::vector<std::string>* section_contents)
555{
556 Unordered_multimap<uint32_t, unsigned int> section_cksum;
557 std::pair<Unordered_multimap<uint32_t, unsigned int>::iterator,
558 Unordered_multimap<uint32_t, unsigned int>::iterator> key_range;
559 bool converged = true;
560
561 if (iteration_num == 1)
562 preprocess_for_unique_sections(id_section,
563 is_secn_or_group_unique,
564 NULL);
565 else
566 preprocess_for_unique_sections(id_section,
567 is_secn_or_group_unique,
568 section_contents);
569
570 std::vector<std::string> full_section_contents;
571
572 for (unsigned int i = 0; i < id_section.size(); i++)
573 {
574 full_section_contents.push_back("");
575 if ((*is_secn_or_group_unique)[i])
576 continue;
577
578 Section_id secn = id_section[i];
579 std::string this_secn_contents;
580 uint32_t cksum;
581 if (iteration_num == 1)
582 {
583 unsigned int num_relocs = 0;
584 this_secn_contents = get_section_contents(true, secn, i, &num_relocs,
585 symtab, (*kept_section_id),
586 section_contents);
587 (*num_tracked_relocs)[i] = num_relocs;
588 }
589 else
590 {
591 if ((*kept_section_id)[i] != i)
592 {
593 // This section is already folded into something. See
594 // if it should point to a different kept section.
595 unsigned int kept_section = (*kept_section_id)[i];
596 if (kept_section != (*kept_section_id)[kept_section])
597 {
598 (*kept_section_id)[i] = (*kept_section_id)[kept_section];
599 }
600 continue;
601 }
602 this_secn_contents = get_section_contents(false, secn, i, NULL,
603 symtab, (*kept_section_id),
604 section_contents);
605 }
606
607 const unsigned char* this_secn_contents_array =
608 reinterpret_cast<const unsigned char*>(this_secn_contents.c_str());
609 cksum = xcrc32(this_secn_contents_array, this_secn_contents.length(),
610 0xffffffff);
611 size_t count = section_cksum.count(cksum);
612
613 if (count == 0)
614 {
615 // Start a group with this cksum.
616 section_cksum.insert(std::make_pair(cksum, i));
617 full_section_contents[i] = this_secn_contents;
618 }
619 else
620 {
621 key_range = section_cksum.equal_range(cksum);
622 Unordered_multimap<uint32_t, unsigned int>::iterator it;
623 // Search all the groups with this cksum for a match.
624 for (it = key_range.first; it != key_range.second; ++it)
625 {
626 unsigned int kept_section = it->second;
627 if (full_section_contents[kept_section].length()
628 != this_secn_contents.length())
629 continue;
630 if (memcmp(full_section_contents[kept_section].c_str(),
631 this_secn_contents.c_str(),
632 this_secn_contents.length()) != 0)
633 continue;
634 (*kept_section_id)[i] = kept_section;
635 converged = false;
636 break;
637 }
638 if (it == key_range.second)
639 {
640 // Create a new group for this cksum.
641 section_cksum.insert(std::make_pair(cksum, i));
642 full_section_contents[i] = this_secn_contents;
643 }
644 }
645 // If there are no relocs to foldable sections do not process
646 // this section any further.
647 if (iteration_num == 1 && (*num_tracked_relocs)[i] == 0)
648 (*is_secn_or_group_unique)[i] = true;
649 }
650
651 return converged;
652}
653
032ce4e9 654// During safe icf (--icf=safe), only fold functions that are ctors or dtors.
4e271fff 655// This function returns true if the section name is that of a ctor or a dtor.
032ce4e9
ST
656
657static bool
4e271fff 658is_function_ctor_or_dtor(const std::string& section_name)
032ce4e9 659{
4e271fff
ST
660 const char* mangled_func_name = strrchr(section_name.c_str(), '.');
661 gold_assert(mangled_func_name != NULL);
662 if ((is_prefix_of("._ZN", mangled_func_name)
663 || is_prefix_of("._ZZ", mangled_func_name))
664 && (is_gnu_v3_mangled_ctor(mangled_func_name + 1)
665 || is_gnu_v3_mangled_dtor(mangled_func_name + 1)))
032ce4e9
ST
666 {
667 return true;
668 }
669 return false;
670}
ef15dade
ST
671
672// This is the main ICF function called in gold.cc. This does the
673// initialization and calls match_sections repeatedly (twice by default)
674// which computes the crc checksums and detects identical functions.
675
676void
677Icf::find_identical_sections(const Input_objects* input_objects,
678 Symbol_table* symtab)
679{
680 unsigned int section_num = 0;
2ea97941 681 std::vector<unsigned int> num_tracked_relocs;
ef15dade
ST
682 std::vector<bool> is_secn_or_group_unique;
683 std::vector<std::string> section_contents;
21bb3914 684 const Target& target = parameters->target();
ef15dade
ST
685
686 // Decide which sections are possible candidates first.
687
688 for (Input_objects::Relobj_iterator p = input_objects->relobj_begin();
689 p != input_objects->relobj_end();
690 ++p)
691 {
5f9bcf58
CC
692 // Lock the object so we can read from it. This is only called
693 // single-threaded from queue_middle_tasks, so it is OK to lock.
694 // Unfortunately we have no way to pass in a Task token.
695 const Task* dummy_task = reinterpret_cast<const Task*>(-1);
696 Task_lock_obj<Object> tl(dummy_task, *p);
697
ef15dade
ST
698 for (unsigned int i = 0;i < (*p)->shnum(); ++i)
699 {
4e271fff 700 const std::string section_name = (*p)->section_name(i);
55a2bb35 701 if (!is_section_foldable_candidate(section_name))
ef15dade
ST
702 continue;
703 if (!(*p)->is_section_included(i))
704 continue;
705 if (parameters->options().gc_sections()
706 && symtab->gc()->is_section_garbage(*p, i))
707 continue;
55a2bb35
ST
708 // With --icf=safe, check if the mangled function name is a ctor
709 // or a dtor. The mangled function name can be obtained from the
710 // section name by stripping the section prefix.
032ce4e9 711 if (parameters->options().icf_safe_folding()
4e271fff 712 && !is_function_ctor_or_dtor(section_name)
21bb3914
ST
713 && (!target.can_check_for_function_pointers()
714 || section_has_function_pointers(*p, i)))
715 {
716 continue;
717 }
ef15dade
ST
718 this->id_section_.push_back(Section_id(*p, i));
719 this->section_id_[Section_id(*p, i)] = section_num;
720 this->kept_section_id_.push_back(section_num);
2ea97941 721 num_tracked_relocs.push_back(0);
ef15dade
ST
722 is_secn_or_group_unique.push_back(false);
723 section_contents.push_back("");
724 section_num++;
725 }
726 }
727
728 unsigned int num_iterations = 0;
729
730 // Default number of iterations to run ICF is 2.
731 unsigned int max_iterations = (parameters->options().icf_iterations() > 0)
732 ? parameters->options().icf_iterations()
733 : 2;
734
735 bool converged = false;
736
737 while (!converged && (num_iterations < max_iterations))
738 {
739 num_iterations++;
740 converged = match_sections(num_iterations, symtab,
2ea97941 741 &num_tracked_relocs, &this->kept_section_id_,
ef15dade
ST
742 this->id_section_, &is_secn_or_group_unique,
743 &section_contents);
744 }
745
746 if (parameters->options().print_icf_sections())
747 {
748 if (converged)
749 gold_info(_("%s: ICF Converged after %u iteration(s)"),
750 program_name, num_iterations);
751 else
752 gold_info(_("%s: ICF stopped after %u iteration(s)"),
753 program_name, num_iterations);
754 }
755
48c187ce
ST
756 // Unfold --keep-unique symbols.
757 for (options::String_set::const_iterator p =
758 parameters->options().keep_unique_begin();
759 p != parameters->options().keep_unique_end();
760 ++p)
761 {
762 const char* name = p->c_str();
763 Symbol* sym = symtab->lookup(name);
ef5e0cb1
ST
764 if (sym == NULL)
765 {
766 gold_warning(_("Could not find symbol %s to unfold\n"), name);
767 }
768 else if (sym->source() == Symbol::FROM_OBJECT
769 && !sym->object()->is_dynamic())
48c187ce
ST
770 {
771 Object* obj = sym->object();
772 bool is_ordinary;
773 unsigned int shndx = sym->shndx(&is_ordinary);
774 if (is_ordinary)
775 {
776 this->unfold_section(obj, shndx);
777 }
778 }
779
780 }
781
ef15dade
ST
782 this->icf_ready();
783}
784
48c187ce
ST
785// Unfolds the section denoted by OBJ and SHNDX if folded.
786
787void
788Icf::unfold_section(Object* obj, unsigned int shndx)
789{
790 Section_id secn(obj, shndx);
791 Uniq_secn_id_map::iterator it = this->section_id_.find(secn);
792 if (it == this->section_id_.end())
793 return;
794 unsigned int section_num = it->second;
795 unsigned int kept_section_id = this->kept_section_id_[section_num];
796 if (kept_section_id != section_num)
797 this->kept_section_id_[section_num] = section_num;
798}
799
ef15dade
ST
800// This function determines if the section corresponding to the
801// given object and index is folded based on if the kept section
802// is different from this section.
803
804bool
805Icf::is_section_folded(Object* obj, unsigned int shndx)
806{
807 Section_id secn(obj, shndx);
808 Uniq_secn_id_map::iterator it = this->section_id_.find(secn);
809 if (it == this->section_id_.end())
810 return false;
811 unsigned int section_num = it->second;
812 unsigned int kept_section_id = this->kept_section_id_[section_num];
813 return kept_section_id != section_num;
814}
815
816// This function returns the folded section for the given section.
817
818Section_id
819Icf::get_folded_section(Object* dup_obj, unsigned int dup_shndx)
820{
821 Section_id dup_secn(dup_obj, dup_shndx);
822 Uniq_secn_id_map::iterator it = this->section_id_.find(dup_secn);
823 gold_assert(it != this->section_id_.end());
824 unsigned int section_num = it->second;
825 unsigned int kept_section_id = this->kept_section_id_[section_num];
826 Section_id folded_section = this->id_section_[kept_section_id];
827 return folded_section;
828}
829
830} // End of namespace gold.
This page took 0.151517 seconds and 4 git commands to generate.