LoongArch: implement count_{leading,trailing}_zeros
[binutils-gdb.git] / libctf / ctf-link.c
blob9babec2aa37bf99ff36239802fd0c92422fdc2f1
1 /* CTF linking.
2 Copyright (C) 2019-2023 Free Software Foundation, Inc.
4 This file is part of libctf.
6 libctf is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
9 version.
11 This program is distributed in the hope that it will be useful, but
12 WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
14 See the GNU General Public License for more details.
16 You should have received a copy of the GNU General Public License
17 along with this program; see the file COPYING. If not see
18 <http://www.gnu.org/licenses/>. */
20 #include <ctf-impl.h>
21 #include <string.h>
23 #if defined (PIC)
24 #pragma weak ctf_open
25 #endif
27 /* CTF linking consists of adding CTF archives full of content to be merged into
28 this one to the current file (which must be writable) by calling
29 ctf_link_add_ctf. Once this is done, a call to ctf_link will merge the type
30 tables together, generating new CTF files as needed, with this one as a
31 parent, to contain types from the inputs which conflict. ctf_link_add_strtab
32 takes a callback which provides string/offset pairs to be added to the
33 external symbol table and deduplicated from all CTF string tables in the
34 output link; ctf_link_shuffle_syms takes a callback which provides symtab
35 entries in ascending order, and shuffles the function and data sections to
36 match; and ctf_link_write emits a CTF file (if there are no conflicts
37 requiring per-compilation-unit sub-CTF files) or CTF archives (otherwise) and
38 returns it, suitable for addition in the .ctf section of the output. */
40 /* Return the name of the compilation unit this CTF dict or its parent applies
41 to, or a non-null string otherwise: prefer the parent. Used in debugging
42 output. Sometimes used for outputs too. */
43 const char *
44 ctf_link_input_name (ctf_dict_t *fp)
46 if (fp->ctf_parent && fp->ctf_parent->ctf_cuname)
47 return fp->ctf_parent->ctf_cuname;
48 else if (fp->ctf_cuname)
49 return fp->ctf_cuname;
50 else
51 return "(unnamed)";
54 /* Return the cuname of a dict, or the string "unnamed-CU" if none. */
56 static const char *
57 ctf_unnamed_cuname (ctf_dict_t *fp)
59 const char *cuname = ctf_cuname (fp);
61 if (!cuname)
62 cuname = "unnamed-CU";
64 return cuname;
67 /* The linker inputs look like this. clin_fp is used for short-circuited
68 CU-mapped links that can entirely avoid the first link phase in some
69 situations in favour of just passing on the contained ctf_dict_t: it is
70 always the sole ctf_dict_t inside the corresponding clin_arc. If set, it
71 gets assigned directly to the final link inputs and freed from there, so it
72 never gets explicitly freed in the ctf_link_input. */
73 typedef struct ctf_link_input
75 char *clin_filename;
76 ctf_archive_t *clin_arc;
77 ctf_dict_t *clin_fp;
78 int n;
79 } ctf_link_input_t;
81 static void
82 ctf_link_input_close (void *input)
84 ctf_link_input_t *i = (ctf_link_input_t *) input;
85 if (i->clin_arc)
86 ctf_arc_close (i->clin_arc);
87 free (i->clin_filename);
88 free (i);
91 /* Like ctf_link_add_ctf, below, but with no error-checking, so it can be called
92 in the middle of an ongoing link. */
93 static int
94 ctf_link_add_ctf_internal (ctf_dict_t *fp, ctf_archive_t *ctf,
95 ctf_dict_t *fp_input, const char *name)
97 int existing = 0;
98 ctf_link_input_t *input;
99 char *filename, *keyname;
101 /* Existing: return it, or (if a different dict with the same name
102 is already there) make up a new unique name. Always use the actual name
103 for the filename, because that needs to be ctf_open()ed. */
105 if ((input = ctf_dynhash_lookup (fp->ctf_link_inputs, name)) != NULL)
107 if ((fp_input != NULL && (input->clin_fp == fp_input))
108 || (ctf != NULL && (input->clin_arc == ctf)))
109 return 0;
110 existing = 1;
113 if ((filename = strdup (name)) == NULL)
114 goto oom;
116 if ((input = calloc (1, sizeof (ctf_link_input_t))) == NULL)
117 goto oom1;
119 input->clin_arc = ctf;
120 input->clin_fp = fp_input;
121 input->clin_filename = filename;
122 input->n = ctf_dynhash_elements (fp->ctf_link_inputs);
124 if (existing)
126 if (asprintf (&keyname, "%s#%li", name, (long int)
127 ctf_dynhash_elements (fp->ctf_link_inputs)) < 0)
128 goto oom2;
130 else if ((keyname = strdup (name)) == NULL)
131 goto oom2;
133 if (ctf_dynhash_insert (fp->ctf_link_inputs, keyname, input) < 0)
134 goto oom3;
136 return 0;
138 oom3:
139 free (keyname);
140 oom2:
141 free (input);
142 oom1:
143 free (filename);
144 oom:
145 return ctf_set_errno (fp, ENOMEM);
148 /* Add a file, memory buffer, or unopened file (by name) to a link.
150 You can call this with:
152 CTF and NAME: link the passed ctf_archive_t, with the given NAME.
153 NAME alone: open NAME as a CTF file when needed.
154 BUF and NAME: open the BUF (of length N) as CTF, with the given NAME. (Not
155 yet implemented.)
157 Passed in CTF args are owned by the dictionary and will be freed by it.
158 The BUF arg is *not* owned by the dictionary, and the user should not free
159 its referent until the link is done.
161 The order of calls to this function influences the order of types in the
162 final link output, but otherwise is not important.
164 Repeated additions of the same NAME have no effect; repeated additions of
165 different dicts with the same NAME add all the dicts with unique NAMEs
166 derived from NAME.
168 Private for now, but may in time become public once support for BUF is
169 implemented. */
171 static int
172 ctf_link_add (ctf_dict_t *fp, ctf_archive_t *ctf, const char *name,
173 void *buf _libctf_unused_, size_t n _libctf_unused_)
175 if (buf)
176 return (ctf_set_errno (fp, ECTF_NOTYET));
178 if (!((ctf && name && !buf)
179 || (name && !buf && !ctf)
180 || (buf && name && !ctf)))
181 return (ctf_set_errno (fp, EINVAL));
183 /* We can only lazily open files if libctf.so is in use rather than
184 libctf-nobfd.so. This is a little tricky: in shared libraries, we can use
185 a weak symbol so that -lctf -lctf-nobfd works, but in static libraries we
186 must distinguish between the two libraries explicitly. */
188 #if defined (PIC)
189 if (!buf && !ctf && name && !ctf_open)
190 return (ctf_set_errno (fp, ECTF_NEEDSBFD));
191 #elif NOBFD
192 if (!buf && !ctf && name)
193 return (ctf_set_errno (fp, ECTF_NEEDSBFD));
194 #endif
196 if (fp->ctf_link_outputs)
197 return (ctf_set_errno (fp, ECTF_LINKADDEDLATE));
198 if (fp->ctf_link_inputs == NULL)
199 fp->ctf_link_inputs = ctf_dynhash_create (ctf_hash_string,
200 ctf_hash_eq_string, free,
201 ctf_link_input_close);
203 if (fp->ctf_link_inputs == NULL)
204 return (ctf_set_errno (fp, ENOMEM));
206 return ctf_link_add_ctf_internal (fp, ctf, NULL, name);
209 /* Add an opened CTF archive or unopened file (by name) to a link.
210 If CTF is NULL and NAME is non-null, an unopened file is meant:
211 otherwise, the specified archive is assumed to have the given NAME.
213 Passed in CTF args are owned by the dictionary and will be freed by it.
215 The order of calls to this function influences the order of types in the
216 final link output, but otherwise is not important. */
219 ctf_link_add_ctf (ctf_dict_t *fp, ctf_archive_t *ctf, const char *name)
221 return ctf_link_add (fp, ctf, name, NULL, 0);
224 /* Lazily open a CTF archive for linking, if not already open.
226 Returns the number of files contained within the opened archive (0 for none),
227 or -1 on error, as usual. */
228 static ssize_t
229 ctf_link_lazy_open (ctf_dict_t *fp, ctf_link_input_t *input)
231 size_t count;
232 int err;
234 if (input->clin_arc)
235 return ctf_archive_count (input->clin_arc);
237 if (input->clin_fp)
238 return 1;
240 /* See ctf_link_add_ctf. */
241 #if defined (PIC) || !NOBFD
242 input->clin_arc = ctf_open (input->clin_filename, NULL, &err);
243 #else
244 ctf_err_warn (fp, 0, ECTF_NEEDSBFD, _("cannot open %s lazily"),
245 input->clin_filename);
246 ctf_set_errno (fp, ECTF_NEEDSBFD);
247 return -1;
248 #endif
250 /* Having no CTF sections is not an error. We just don't need to do
251 anything. */
253 if (!input->clin_arc)
255 if (err == ECTF_NOCTFDATA)
256 return 0;
258 ctf_err_warn (fp, 0, err, _("opening CTF %s failed"),
259 input->clin_filename);
260 ctf_set_errno (fp, err);
261 return -1;
264 if ((count = ctf_archive_count (input->clin_arc)) == 0)
265 ctf_arc_close (input->clin_arc);
267 return (ssize_t) count;
270 /* Find a non-clashing unique name for a per-CU output dict, to prevent distinct
271 members corresponding to inputs with identical cunames from overwriting each
272 other. The name should be something like NAME. */
274 static char *
275 ctf_new_per_cu_name (ctf_dict_t *fp, const char *name)
277 char *dynname;
278 long int i = 0;
280 if ((dynname = strdup (name)) == NULL)
281 return NULL;
283 while ((ctf_dynhash_lookup (fp->ctf_link_outputs, dynname)) != NULL)
285 free (dynname);
286 if (asprintf (&dynname, "%s#%li", name, i++) < 0)
287 return NULL;
290 return dynname;
293 /* Return a per-CU output CTF dictionary suitable for the given INPUT or CU,
294 creating and interning it if need be. */
296 static ctf_dict_t *
297 ctf_create_per_cu (ctf_dict_t *fp, ctf_dict_t *input, const char *cu_name)
299 ctf_dict_t *cu_fp;
300 const char *ctf_name = NULL;
301 char *dynname = NULL;
303 /* Already has a per-CU mapping? Just return it. */
305 if (input && input->ctf_link_in_out)
306 return input->ctf_link_in_out;
308 /* Check the mapping table and translate the per-CU name we use
309 accordingly. */
311 if (cu_name == NULL)
312 cu_name = ctf_unnamed_cuname (input);
314 if (fp->ctf_link_in_cu_mapping)
316 if ((ctf_name = ctf_dynhash_lookup (fp->ctf_link_in_cu_mapping,
317 cu_name)) == NULL)
318 ctf_name = cu_name;
321 if (ctf_name == NULL)
322 ctf_name = cu_name;
324 /* Look up the per-CU dict. If we don't know of one, or it is for a different input
325 CU which just happens to have the same name, create a new one. If we are creating
326 a dict with no input specified, anything will do. */
328 if ((cu_fp = ctf_dynhash_lookup (fp->ctf_link_outputs, ctf_name)) == NULL
329 || (input && cu_fp->ctf_link_in_out != fp))
331 int err;
333 if ((cu_fp = ctf_create (&err)) == NULL)
335 ctf_err_warn (fp, 0, err, _("cannot create per-CU CTF archive for "
336 "input CU %s"), cu_name);
337 ctf_set_errno (fp, err);
338 return NULL;
341 ctf_import_unref (cu_fp, fp);
343 if ((dynname = ctf_new_per_cu_name (fp, ctf_name)) == NULL)
344 goto oom;
346 ctf_cuname_set (cu_fp, cu_name);
348 ctf_parent_name_set (cu_fp, _CTF_SECTION);
349 cu_fp->ctf_link_in_out = fp;
350 fp->ctf_link_in_out = cu_fp;
352 if (ctf_dynhash_insert (fp->ctf_link_outputs, dynname, cu_fp) < 0)
353 goto oom;
355 return cu_fp;
357 oom:
358 free (dynname);
359 ctf_dict_close (cu_fp);
360 ctf_set_errno (fp, ENOMEM);
361 return NULL;
364 /* Add a mapping directing that the CU named FROM should have its
365 conflicting/non-duplicate types (depending on link mode) go into a dict
366 named TO. Many FROMs can share a TO.
368 We forcibly add a dict named TO in every case, even though it may well
369 wind up empty, because clients that use this facility usually expect to find
370 every TO dict present, even if empty, and malfunction otherwise. */
373 ctf_link_add_cu_mapping (ctf_dict_t *fp, const char *from, const char *to)
375 int err;
376 char *f = NULL, *t = NULL;
377 ctf_dynhash_t *one_out;
379 /* Mappings cannot be set up if per-CU output dicts already exist. */
380 if (fp->ctf_link_outputs && ctf_dynhash_elements (fp->ctf_link_outputs) != 0)
381 return (ctf_set_errno (fp, ECTF_LINKADDEDLATE));
383 if (fp->ctf_link_in_cu_mapping == NULL)
384 fp->ctf_link_in_cu_mapping = ctf_dynhash_create (ctf_hash_string,
385 ctf_hash_eq_string, free,
386 free);
387 if (fp->ctf_link_in_cu_mapping == NULL)
388 goto oom;
390 if (fp->ctf_link_out_cu_mapping == NULL)
391 fp->ctf_link_out_cu_mapping = ctf_dynhash_create (ctf_hash_string,
392 ctf_hash_eq_string, free,
393 (ctf_hash_free_fun)
394 ctf_dynhash_destroy);
395 if (fp->ctf_link_out_cu_mapping == NULL)
396 goto oom;
398 f = strdup (from);
399 t = strdup (to);
400 if (!f || !t)
401 goto oom;
403 /* Track both in a list from FROM to TO and in a list from TO to a list of
404 FROM. The former is used to create TUs with the mapped-to name at need:
405 the latter is used in deduplicating links to pull in all input CUs
406 corresponding to a single output CU. */
408 if ((err = ctf_dynhash_insert (fp->ctf_link_in_cu_mapping, f, t)) < 0)
410 ctf_set_errno (fp, err);
411 goto oom_noerrno;
414 /* f and t are now owned by the in_cu_mapping: reallocate them. */
415 f = strdup (from);
416 t = strdup (to);
417 if (!f || !t)
418 goto oom;
420 if ((one_out = ctf_dynhash_lookup (fp->ctf_link_out_cu_mapping, t)) == NULL)
422 if ((one_out = ctf_dynhash_create (ctf_hash_string, ctf_hash_eq_string,
423 free, NULL)) == NULL)
424 goto oom;
425 if ((err = ctf_dynhash_insert (fp->ctf_link_out_cu_mapping,
426 t, one_out)) < 0)
428 ctf_dynhash_destroy (one_out);
429 ctf_set_errno (fp, err);
430 goto oom_noerrno;
433 else
435 free (t);
436 t = NULL;
439 if (ctf_dynhash_insert (one_out, f, NULL) < 0)
441 ctf_set_errno (fp, err);
442 goto oom_noerrno;
445 return 0;
447 oom:
448 ctf_set_errno (fp, errno);
449 oom_noerrno:
450 free (f);
451 free (t);
452 return -1;
455 /* Set a function which is called to transform the names of archive members.
456 This is useful for applying regular transformations to many names, where
457 ctf_link_add_cu_mapping applies arbitrarily irregular changes to single
458 names. The member name changer is applied at ctf_link_write time, so it
459 cannot conflate multiple CUs into one the way ctf_link_add_cu_mapping can.
460 The changer function accepts a name and should return a new
461 dynamically-allocated name, or NULL if the name should be left unchanged. */
462 void
463 ctf_link_set_memb_name_changer (ctf_dict_t *fp,
464 ctf_link_memb_name_changer_f *changer,
465 void *arg)
467 fp->ctf_link_memb_name_changer = changer;
468 fp->ctf_link_memb_name_changer_arg = arg;
471 /* Set a function which is used to filter out unwanted variables from the link. */
473 ctf_link_set_variable_filter (ctf_dict_t *fp, ctf_link_variable_filter_f *filter,
474 void *arg)
476 fp->ctf_link_variable_filter = filter;
477 fp->ctf_link_variable_filter_arg = arg;
478 return 0;
481 /* Check if we can safely add a variable with the given type to this dict. */
483 static int
484 check_variable (const char *name, ctf_dict_t *fp, ctf_id_t type,
485 ctf_dvdef_t **out_dvd)
487 ctf_dvdef_t *dvd;
489 dvd = ctf_dynhash_lookup (fp->ctf_dvhash, name);
490 *out_dvd = dvd;
491 if (!dvd)
492 return 1;
494 if (dvd->dvd_type != type)
496 /* Variable here. Wrong type: cannot add. Just skip it, because there is
497 no way to express this in CTF. Don't even warn: this case is too
498 common. (This might be the parent, in which case we'll try adding in
499 the child first, and only then give up.) */
500 ctf_dprintf ("Inexpressible duplicate variable %s skipped.\n", name);
503 return 0; /* Already exists. */
506 /* Link one variable named NAME of type TYPE found in IN_FP into FP. */
508 static int
509 ctf_link_one_variable (ctf_dict_t *fp, ctf_dict_t *in_fp, const char *name,
510 ctf_id_t type, int cu_mapped)
512 ctf_dict_t *per_cu_out_fp;
513 ctf_id_t dst_type = 0;
514 ctf_dvdef_t *dvd;
516 /* See if this variable is filtered out. */
518 if (fp->ctf_link_variable_filter)
520 void *farg = fp->ctf_link_variable_filter_arg;
521 if (fp->ctf_link_variable_filter (in_fp, name, type, farg))
522 return 0;
525 /* If this type is mapped to a type in the parent dict, we want to try to add
526 to that first: if it reports a duplicate, or if the type is in a child
527 already, add straight to the child. */
529 if ((dst_type = ctf_dedup_type_mapping (fp, in_fp, type)) == CTF_ERR)
530 return -1; /* errno is set for us. */
532 if (dst_type != 0)
534 if (!ctf_assert (fp, ctf_type_isparent (fp, dst_type)))
535 return -1; /* errno is set for us. */
537 if (check_variable (name, fp, dst_type, &dvd))
539 /* No variable here: we can add it. */
540 if (ctf_add_variable (fp, name, dst_type) < 0)
541 return -1; /* errno is set for us. */
542 return 0;
545 /* Already present? Nothing to do. */
546 if (dvd && dvd->dvd_type == dst_type)
547 return 0;
550 /* Can't add to the parent due to a name clash, or because it references a
551 type only present in the child. Try adding to the child, creating if need
552 be. If we can't do that, skip it. Don't add to a child if we're doing a
553 CU-mapped link, since that has only one output. */
555 if (cu_mapped)
557 ctf_dprintf ("Variable %s in input file %s depends on a type %lx hidden "
558 "due to conflicts: skipped.\n", name,
559 ctf_unnamed_cuname (in_fp), type);
560 return 0;
563 if ((per_cu_out_fp = ctf_create_per_cu (fp, in_fp, NULL)) == NULL)
564 return -1; /* errno is set for us. */
566 /* If the type was not found, check for it in the child too. */
567 if (dst_type == 0)
569 if ((dst_type = ctf_dedup_type_mapping (per_cu_out_fp,
570 in_fp, type)) == CTF_ERR)
571 return -1; /* errno is set for us. */
573 if (dst_type == 0)
575 ctf_err_warn (fp, 1, 0, _("type %lx for variable %s in input file %s "
576 "not found: skipped"), type, name,
577 ctf_unnamed_cuname (in_fp));
578 /* Do not terminate the link: just skip the variable. */
579 return 0;
583 if (check_variable (name, per_cu_out_fp, dst_type, &dvd))
584 if (ctf_add_variable (per_cu_out_fp, name, dst_type) < 0)
585 return (ctf_set_errno (fp, ctf_errno (per_cu_out_fp)));
586 return 0;
589 typedef struct link_sort_inputs_cb_arg
591 int is_cu_mapped;
592 ctf_dict_t *fp;
593 } link_sort_inputs_cb_arg_t;
595 /* Sort the inputs by N (the link order). For CU-mapped links, this is a
596 mapping of input to output name, not a mapping of input name to input
597 ctf_link_input_t: compensate accordingly. */
598 static int
599 ctf_link_sort_inputs (const ctf_next_hkv_t *one, const ctf_next_hkv_t *two,
600 void *arg)
602 ctf_link_input_t *input_1;
603 ctf_link_input_t *input_2;
604 link_sort_inputs_cb_arg_t *cu_mapped = (link_sort_inputs_cb_arg_t *) arg;
606 if (!cu_mapped || !cu_mapped->is_cu_mapped)
608 input_1 = (ctf_link_input_t *) one->hkv_value;
609 input_2 = (ctf_link_input_t *) two->hkv_value;
611 else
613 const char *name_1 = (const char *) one->hkv_key;
614 const char *name_2 = (const char *) two->hkv_key;
616 input_1 = ctf_dynhash_lookup (cu_mapped->fp->ctf_link_inputs, name_1);
617 input_2 = ctf_dynhash_lookup (cu_mapped->fp->ctf_link_inputs, name_2);
619 /* There is no guarantee that CU-mappings actually have corresponding
620 inputs: the relative ordering in that case is unimportant. */
621 if (!input_1)
622 return -1;
623 if (!input_2)
624 return 1;
627 if (input_1->n < input_2->n)
628 return -1;
629 else if (input_1->n > input_2->n)
630 return 1;
631 else
632 return 0;
635 /* Count the number of input dicts in the ctf_link_inputs, or that subset of the
636 ctf_link_inputs given by CU_NAMES if set. Return the number of input dicts,
637 and optionally the name and ctf_link_input_t of the single input archive if
638 only one exists (no matter how many dicts it contains). */
639 static ssize_t
640 ctf_link_deduplicating_count_inputs (ctf_dict_t *fp, ctf_dynhash_t *cu_names,
641 ctf_link_input_t **only_one_input)
643 ctf_dynhash_t *inputs = fp->ctf_link_inputs;
644 ctf_next_t *i = NULL;
645 void *name, *input;
646 ctf_link_input_t *one_input = NULL;
647 const char *one_name = NULL;
648 ssize_t count = 0, narcs = 0;
649 int err;
651 if (cu_names)
652 inputs = cu_names;
654 while ((err = ctf_dynhash_next (inputs, &i, &name, &input)) == 0)
656 ssize_t one_count;
658 one_name = (const char *) name;
659 /* If we are processing CU names, get the real input. */
660 if (cu_names)
661 one_input = ctf_dynhash_lookup (fp->ctf_link_inputs, one_name);
662 else
663 one_input = (ctf_link_input_t *) input;
665 if (!one_input)
666 continue;
668 one_count = ctf_link_lazy_open (fp, one_input);
670 if (one_count < 0)
672 ctf_next_destroy (i);
673 return -1; /* errno is set for us. */
676 count += one_count;
677 narcs++;
679 if (err != ECTF_NEXT_END)
681 ctf_err_warn (fp, 0, err, _("iteration error counting deduplicating "
682 "CTF link inputs"));
683 ctf_set_errno (fp, err);
684 return -1;
687 if (!count)
688 return 0;
690 if (narcs == 1)
692 if (only_one_input)
693 *only_one_input = one_input;
695 else if (only_one_input)
696 *only_one_input = NULL;
698 return count;
701 /* Allocate and populate an inputs array big enough for a given set of inputs:
702 either a specific set of CU names (those from that set found in the
703 ctf_link_inputs), or the entire ctf_link_inputs (if cu_names is not set).
704 The number of inputs (from ctf_link_deduplicating_count_inputs, above) is
705 passed in NINPUTS: an array of uint32_t containing parent pointers
706 (corresponding to those members of the inputs that have parents) is allocated
707 and returned in PARENTS.
709 The inputs are *archives*, not files: the archive can have multiple members
710 if it is the result of a previous incremental link. We want to add every one
711 in turn, including the shared parent. (The dedup machinery knows that a type
712 used by a single dictionary and its parent should not be shared in
713 CTF_LINK_SHARE_DUPLICATED mode.)
715 If no inputs exist that correspond to these CUs, return NULL with the errno
716 set to ECTF_NOCTFDATA. */
717 static ctf_dict_t **
718 ctf_link_deduplicating_open_inputs (ctf_dict_t *fp, ctf_dynhash_t *cu_names,
719 ssize_t ninputs, uint32_t **parents)
721 ctf_dynhash_t *inputs = fp->ctf_link_inputs;
722 ctf_next_t *i = NULL;
723 void *name, *input;
724 link_sort_inputs_cb_arg_t sort_arg;
725 ctf_dict_t **dedup_inputs = NULL;
726 ctf_dict_t **walk;
727 uint32_t *parents_ = NULL;
728 int err;
730 if (cu_names)
731 inputs = cu_names;
733 if ((dedup_inputs = calloc (ninputs, sizeof (ctf_dict_t *))) == NULL)
734 goto oom;
736 if ((parents_ = calloc (ninputs, sizeof (uint32_t))) == NULL)
737 goto oom;
739 walk = dedup_inputs;
741 /* Counting done: push every input into the array, in the order they were
742 passed to ctf_link_add_ctf (and ultimately ld). */
744 sort_arg.is_cu_mapped = (cu_names != NULL);
745 sort_arg.fp = fp;
747 while ((err = ctf_dynhash_next_sorted (inputs, &i, &name, &input,
748 ctf_link_sort_inputs, &sort_arg)) == 0)
750 const char *one_name = (const char *) name;
751 ctf_link_input_t *one_input;
752 ctf_dict_t *one_fp;
753 ctf_dict_t *parent_fp = NULL;
754 uint32_t parent_i;
755 ctf_next_t *j = NULL;
757 /* If we are processing CU names, get the real input. All the inputs
758 will have been opened, if they contained any CTF at all. */
759 if (cu_names)
760 one_input = ctf_dynhash_lookup (fp->ctf_link_inputs, one_name);
761 else
762 one_input = (ctf_link_input_t *) input;
764 if (!one_input || (!one_input->clin_arc && !one_input->clin_fp))
765 continue;
767 /* Short-circuit: if clin_fp is set, just use it. */
768 if (one_input->clin_fp)
770 parents_[walk - dedup_inputs] = walk - dedup_inputs;
771 *walk = one_input->clin_fp;
772 walk++;
773 continue;
776 /* Get and insert the parent archive (if any), if this archive has
777 multiple members. We assume, as elsewhere, that the parent is named
778 _CTF_SECTION. */
780 if ((parent_fp = ctf_dict_open (one_input->clin_arc, _CTF_SECTION,
781 &err)) == NULL)
783 if (err != ECTF_NOMEMBNAM)
785 ctf_next_destroy (i);
786 ctf_set_errno (fp, err);
787 goto err;
790 else
792 *walk = parent_fp;
793 parent_i = walk - dedup_inputs;
794 walk++;
797 /* We disregard the input archive name: either it is the parent (which we
798 already have), or we want to put everything into one TU sharing the
799 cuname anyway (if this is a CU-mapped link), or this is the final phase
800 of a relink with CU-mapping off (i.e. ld -r) in which case the cuname
801 is correctly set regardless. */
802 while ((one_fp = ctf_archive_next (one_input->clin_arc, &j, NULL,
803 1, &err)) != NULL)
805 if (one_fp->ctf_flags & LCTF_CHILD)
807 /* The contents of the parents array for elements not
808 corresponding to children is undefined. If there is no parent
809 (itself a sign of a likely linker bug or corrupt input), we set
810 it to itself. */
812 ctf_import (one_fp, parent_fp);
813 if (parent_fp)
814 parents_[walk - dedup_inputs] = parent_i;
815 else
816 parents_[walk - dedup_inputs] = walk - dedup_inputs;
818 *walk = one_fp;
819 walk++;
821 if (err != ECTF_NEXT_END)
823 ctf_next_destroy (i);
824 goto iterr;
827 if (err != ECTF_NEXT_END)
828 goto iterr;
830 *parents = parents_;
832 return dedup_inputs;
834 oom:
835 err = ENOMEM;
837 iterr:
838 ctf_set_errno (fp, err);
840 err:
841 free (dedup_inputs);
842 free (parents_);
843 ctf_err_warn (fp, 0, 0, _("error in deduplicating CTF link "
844 "input allocation"));
845 return NULL;
848 /* Close INPUTS that have already been linked, first the passed array, and then
849 that subset of the ctf_link_inputs archives they came from cited by the
850 CU_NAMES. If CU_NAMES is not specified, close all the ctf_link_inputs in one
851 go, leaving it empty. */
852 static int
853 ctf_link_deduplicating_close_inputs (ctf_dict_t *fp, ctf_dynhash_t *cu_names,
854 ctf_dict_t **inputs, ssize_t ninputs)
856 ctf_next_t *it = NULL;
857 void *name;
858 int err;
859 ssize_t i;
861 /* This is the inverse of ctf_link_deduplicating_open_inputs: so first, close
862 all the individual input dicts, opened by the archive iterator. */
863 for (i = 0; i < ninputs; i++)
864 ctf_dict_close (inputs[i]);
866 /* Now close the archives they are part of. */
867 if (cu_names)
869 while ((err = ctf_dynhash_next (cu_names, &it, &name, NULL)) == 0)
871 /* Remove the input from the linker inputs, if it exists, which also
872 closes it. */
874 ctf_dynhash_remove (fp->ctf_link_inputs, (const char *) name);
876 if (err != ECTF_NEXT_END)
878 ctf_err_warn (fp, 0, err, _("iteration error in deduplicating link "
879 "input freeing"));
880 ctf_set_errno (fp, err);
883 else
884 ctf_dynhash_empty (fp->ctf_link_inputs);
886 return 0;
889 /* Do a deduplicating link of all variables in the inputs.
891 Also, if we are not omitting the variable section, integrate all symbols from
892 the symtypetabs into the variable section too. (Duplication with the
893 symtypetab section in the output will be eliminated at serialization time.) */
895 static int
896 ctf_link_deduplicating_variables (ctf_dict_t *fp, ctf_dict_t **inputs,
897 size_t ninputs, int cu_mapped)
899 size_t i;
901 for (i = 0; i < ninputs; i++)
903 ctf_next_t *it = NULL;
904 ctf_id_t type;
905 const char *name;
907 /* First the variables on the inputs. */
909 while ((type = ctf_variable_next (inputs[i], &it, &name)) != CTF_ERR)
911 if (ctf_link_one_variable (fp, inputs[i], name, type, cu_mapped) < 0)
913 ctf_next_destroy (it);
914 return -1; /* errno is set for us. */
917 if (ctf_errno (inputs[i]) != ECTF_NEXT_END)
918 return ctf_set_errno (fp, ctf_errno (inputs[i]));
920 /* Next the symbols. We integrate data symbols even though the compiler
921 is currently doing the same, to allow the compiler to stop in
922 future. */
924 while ((type = ctf_symbol_next (inputs[i], &it, &name, 0)) != CTF_ERR)
926 if (ctf_link_one_variable (fp, inputs[i], name, type, 1) < 0)
928 ctf_next_destroy (it);
929 return -1; /* errno is set for us. */
932 if (ctf_errno (inputs[i]) != ECTF_NEXT_END)
933 return ctf_set_errno (fp, ctf_errno (inputs[i]));
935 /* Finally the function symbols. */
937 while ((type = ctf_symbol_next (inputs[i], &it, &name, 1)) != CTF_ERR)
939 if (ctf_link_one_variable (fp, inputs[i], name, type, 1) < 0)
941 ctf_next_destroy (it);
942 return -1; /* errno is set for us. */
945 if (ctf_errno (inputs[i]) != ECTF_NEXT_END)
946 return ctf_set_errno (fp, ctf_errno (inputs[i]));
948 return 0;
951 /* Check for symbol conflicts during linking. Three possibilities: already
952 exists, conflicting, or nonexistent. We don't have a dvd structure we can
953 use as a flag like check_variable does, so we use a tristate return
954 value instead: -1: conflicting; 1: nonexistent: 0: already exists. */
956 static int
957 check_sym (ctf_dict_t *fp, const char *name, ctf_id_t type, int functions)
959 ctf_dynhash_t *thishash = functions ? fp->ctf_funchash : fp->ctf_objthash;
960 ctf_dynhash_t *thathash = functions ? fp->ctf_objthash : fp->ctf_funchash;
961 void *value;
963 /* Wrong type (function when object is wanted, etc). */
964 if (ctf_dynhash_lookup_kv (thathash, name, NULL, NULL))
965 return -1;
967 /* Not present at all yet. */
968 if (!ctf_dynhash_lookup_kv (thishash, name, NULL, &value))
969 return 1;
971 /* Already present. */
972 if ((ctf_id_t) (uintptr_t) value == type)
973 return 0;
975 /* Wrong type. */
976 return -1;
979 /* Do a deduplicating link of one symtypetab (function info or data object) in
980 one input dict. */
982 static int
983 ctf_link_deduplicating_one_symtypetab (ctf_dict_t *fp, ctf_dict_t *input,
984 int cu_mapped, int functions)
986 ctf_next_t *it = NULL;
987 const char *name;
988 ctf_id_t type;
990 while ((type = ctf_symbol_next (input, &it, &name, functions)) != CTF_ERR)
992 ctf_id_t dst_type;
993 ctf_dict_t *per_cu_out_fp;
994 int sym;
996 /* Look in the parent first. */
998 if ((dst_type = ctf_dedup_type_mapping (fp, input, type)) == CTF_ERR)
999 return -1; /* errno is set for us. */
1001 if (dst_type != 0)
1003 if (!ctf_assert (fp, ctf_type_isparent (fp, dst_type)))
1004 return -1; /* errno is set for us. */
1006 sym = check_sym (fp, name, dst_type, functions);
1008 /* Already present: next symbol. */
1009 if (sym == 0)
1010 continue;
1011 /* Not present: add it. */
1012 else if (sym > 0)
1014 if (ctf_add_funcobjt_sym (fp, functions,
1015 name, dst_type) < 0)
1016 return -1; /* errno is set for us. */
1017 continue;
1021 /* Can't add to the parent due to a name clash (most unlikely), or because
1022 it references a type only present in the child. Try adding to the
1023 child, creating if need be. If we can't do that, skip it. Don't add
1024 to a child if we're doing a CU-mapped link, since that has only one
1025 output. */
1026 if (cu_mapped)
1028 ctf_dprintf ("Symbol %s in input file %s depends on a type %lx "
1029 "hidden due to conflicts: skipped.\n", name,
1030 ctf_unnamed_cuname (input), type);
1031 continue;
1034 if ((per_cu_out_fp = ctf_create_per_cu (fp, input, NULL)) == NULL)
1035 return -1; /* errno is set for us. */
1037 /* If the type was not found, check for it in the child too. */
1038 if (dst_type == 0)
1040 if ((dst_type = ctf_dedup_type_mapping (per_cu_out_fp,
1041 input, type)) == CTF_ERR)
1042 return -1; /* errno is set for us. */
1044 if (dst_type == 0)
1046 ctf_err_warn (fp, 1, 0,
1047 _("type %lx for symbol %s in input file %s "
1048 "not found: skipped"), type, name,
1049 ctf_unnamed_cuname (input));
1050 continue;
1054 sym = check_sym (per_cu_out_fp, name, dst_type, functions);
1056 /* Already present: next symbol. */
1057 if (sym == 0)
1058 continue;
1059 /* Not present: add it. */
1060 else if (sym > 0)
1062 if (ctf_add_funcobjt_sym (per_cu_out_fp, functions,
1063 name, dst_type) < 0)
1064 return -1; /* errno is set for us. */
1066 else
1068 /* Perhaps this should be an assertion failure. */
1069 ctf_err_warn (fp, 0, ECTF_DUPLICATE,
1070 _("symbol %s in input file %s found conflicting "
1071 "even when trying in per-CU dict."), name,
1072 ctf_unnamed_cuname (input));
1073 return (ctf_set_errno (fp, ECTF_DUPLICATE));
1076 if (ctf_errno (input) != ECTF_NEXT_END)
1078 ctf_set_errno (fp, ctf_errno (input));
1079 ctf_err_warn (fp, 0, ctf_errno (input),
1080 functions ? _("iterating over function symbols") :
1081 _("iterating over data symbols"));
1082 return -1;
1085 return 0;
1088 /* Do a deduplicating link of the function info and data objects
1089 in the inputs. */
1090 static int
1091 ctf_link_deduplicating_syms (ctf_dict_t *fp, ctf_dict_t **inputs,
1092 size_t ninputs, int cu_mapped)
1094 size_t i;
1096 for (i = 0; i < ninputs; i++)
1098 if (ctf_link_deduplicating_one_symtypetab (fp, inputs[i],
1099 cu_mapped, 0) < 0)
1100 return -1; /* errno is set for us. */
1102 if (ctf_link_deduplicating_one_symtypetab (fp, inputs[i],
1103 cu_mapped, 1) < 0)
1104 return -1; /* errno is set for us. */
1107 return 0;
1110 /* Do the per-CU part of a deduplicating link. */
1111 static int
1112 ctf_link_deduplicating_per_cu (ctf_dict_t *fp)
1114 ctf_next_t *i = NULL;
1115 int err;
1116 void *out_cu;
1117 void *in_cus;
1119 /* Links with a per-CU mapping in force get a first pass of deduplication,
1120 dedupping the inputs for a given CU mapping into the output for that
1121 mapping. The outputs from this process get fed back into the final pass
1122 that is carried out even for non-CU links. */
1124 while ((err = ctf_dynhash_next (fp->ctf_link_out_cu_mapping, &i, &out_cu,
1125 &in_cus)) == 0)
1127 const char *out_name = (const char *) out_cu;
1128 ctf_dynhash_t *in = (ctf_dynhash_t *) in_cus;
1129 ctf_dict_t *out = NULL;
1130 ctf_dict_t **inputs;
1131 ctf_dict_t **outputs;
1132 ctf_archive_t *in_arc;
1133 ssize_t ninputs;
1134 ctf_link_input_t *only_input;
1135 uint32_t noutputs;
1136 uint32_t *parents;
1138 if ((ninputs = ctf_link_deduplicating_count_inputs (fp, in,
1139 &only_input)) == -1)
1140 goto err_open_inputs;
1142 /* CU mapping with no inputs? Skip. */
1143 if (ninputs == 0)
1144 continue;
1146 if (labs ((long int) ninputs) > 0xfffffffe)
1148 ctf_err_warn (fp, 0, EFBIG, _("too many inputs in deduplicating "
1149 "link: %li"), (long int) ninputs);
1150 ctf_set_errno (fp, EFBIG);
1151 goto err_open_inputs;
1154 /* Short-circuit: a cu-mapped link with only one input archive with
1155 unconflicting contents is a do-nothing, and we can just leave the input
1156 in place: we do have to change the cuname, though, so we unwrap it,
1157 change the cuname, then stuff it back in the linker input again, via
1158 the clin_fp short-circuit member. ctf_link_deduplicating_open_inputs
1159 will spot this member and jam it straight into the next link phase,
1160 ignoring the corresponding archive. */
1161 if (only_input && ninputs == 1)
1163 ctf_next_t *ai = NULL;
1164 int err;
1166 /* We can abuse an archive iterator to get the only member cheaply, no
1167 matter what its name. */
1168 only_input->clin_fp = ctf_archive_next (only_input->clin_arc,
1169 &ai, NULL, 0, &err);
1170 if (!only_input->clin_fp)
1172 ctf_err_warn (fp, 0, err, _("cannot open archive %s in "
1173 "CU-mapped CTF link"),
1174 only_input->clin_filename);
1175 ctf_set_errno (fp, err);
1176 goto err_open_inputs;
1178 ctf_next_destroy (ai);
1180 if (strcmp (only_input->clin_filename, out_name) != 0)
1182 /* Renaming. We need to add a new input, then null out the
1183 clin_arc and clin_fp of the old one to stop it being
1184 auto-closed on removal. The new input needs its cuname changed
1185 to out_name, which is doable only because the cuname is a
1186 dynamic property which can be changed even in readonly
1187 dicts. */
1189 ctf_cuname_set (only_input->clin_fp, out_name);
1190 if (ctf_link_add_ctf_internal (fp, only_input->clin_arc,
1191 only_input->clin_fp,
1192 out_name) < 0)
1194 ctf_err_warn (fp, 0, 0, _("cannot add intermediate files "
1195 "to link"));
1196 goto err_open_inputs;
1198 only_input->clin_arc = NULL;
1199 only_input->clin_fp = NULL;
1200 ctf_dynhash_remove (fp->ctf_link_inputs,
1201 only_input->clin_filename);
1203 continue;
1206 /* This is a real CU many-to-one mapping: we must dedup the inputs into
1207 a new output to be used in the final link phase. */
1209 if ((inputs = ctf_link_deduplicating_open_inputs (fp, in, ninputs,
1210 &parents)) == NULL)
1212 ctf_next_destroy (i);
1213 goto err_inputs;
1216 if ((out = ctf_create (&err)) == NULL)
1218 ctf_err_warn (fp, 0, err, _("cannot create per-CU CTF archive "
1219 "for %s"),
1220 out_name);
1221 ctf_set_errno (fp, err);
1222 goto err_inputs;
1225 /* Share the atoms table to reduce memory usage. */
1226 out->ctf_dedup_atoms = fp->ctf_dedup_atoms_alloc;
1228 /* No ctf_imports at this stage: this per-CU dictionary has no parents.
1229 Parent/child deduplication happens in the link's final pass. However,
1230 the cuname *is* important, as it is propagated into the final
1231 dictionary. */
1232 ctf_cuname_set (out, out_name);
1234 if (ctf_dedup (out, inputs, ninputs, parents, 1) < 0)
1236 ctf_set_errno (fp, ctf_errno (out));
1237 ctf_err_warn (fp, 0, 0, _("CU-mapped deduplication failed for %s"),
1238 out_name);
1239 goto err_inputs;
1242 if ((outputs = ctf_dedup_emit (out, inputs, ninputs, parents,
1243 &noutputs, 1)) == NULL)
1245 ctf_set_errno (fp, ctf_errno (out));
1246 ctf_err_warn (fp, 0, 0, _("CU-mapped deduplicating link type emission "
1247 "failed for %s"), out_name);
1248 goto err_inputs;
1250 if (!ctf_assert (fp, noutputs == 1))
1252 size_t j;
1253 for (j = 1; j < noutputs; j++)
1254 ctf_dict_close (outputs[j]);
1255 goto err_inputs_outputs;
1258 if (!(fp->ctf_link_flags & CTF_LINK_OMIT_VARIABLES_SECTION)
1259 && ctf_link_deduplicating_variables (out, inputs, ninputs, 1) < 0)
1261 ctf_set_errno (fp, ctf_errno (out));
1262 ctf_err_warn (fp, 0, 0, _("CU-mapped deduplicating link variable "
1263 "emission failed for %s"), out_name);
1264 goto err_inputs_outputs;
1267 ctf_dedup_fini (out, outputs, noutputs);
1269 /* For now, we omit symbol section linking for CU-mapped links, until it
1270 is clear how to unify the symbol table across such links. (Perhaps we
1271 should emit an unconditionally indexed symtab, like the compiler
1272 does.) */
1274 if (ctf_link_deduplicating_close_inputs (fp, in, inputs, ninputs) < 0)
1276 free (inputs);
1277 free (parents);
1278 goto err_outputs;
1280 free (inputs);
1281 free (parents);
1283 /* Splice any errors or warnings created during this link back into the
1284 dict that the caller knows about. */
1285 ctf_list_splice (&fp->ctf_errs_warnings, &outputs[0]->ctf_errs_warnings);
1287 /* This output now becomes an input to the next link phase, with a name
1288 equal to the CU name. We have to wrap it in an archive wrapper
1289 first. */
1291 if ((in_arc = ctf_new_archive_internal (0, 0, NULL, outputs[0], NULL,
1292 NULL, &err)) == NULL)
1294 ctf_set_errno (fp, err);
1295 goto err_outputs;
1298 if (ctf_link_add_ctf_internal (fp, in_arc, NULL,
1299 ctf_cuname (outputs[0])) < 0)
1301 ctf_err_warn (fp, 0, 0, _("cannot add intermediate files to link"));
1302 goto err_outputs;
1305 ctf_dict_close (out);
1306 free (outputs);
1307 continue;
1309 err_inputs_outputs:
1310 ctf_list_splice (&fp->ctf_errs_warnings, &outputs[0]->ctf_errs_warnings);
1311 ctf_dict_close (outputs[0]);
1312 free (outputs);
1313 err_inputs:
1314 ctf_link_deduplicating_close_inputs (fp, in, inputs, ninputs);
1315 ctf_dict_close (out);
1316 free (inputs);
1317 free (parents);
1318 err_open_inputs:
1319 ctf_next_destroy (i);
1320 return -1;
1322 err_outputs:
1323 ctf_list_splice (&fp->ctf_errs_warnings, &outputs[0]->ctf_errs_warnings);
1324 ctf_dict_close (outputs[0]);
1325 free (outputs);
1326 ctf_next_destroy (i);
1327 return -1; /* Errno is set for us. */
1329 if (err != ECTF_NEXT_END)
1331 ctf_err_warn (fp, 0, err, _("iteration error in CU-mapped deduplicating "
1332 "link"));
1333 return ctf_set_errno (fp, err);
1336 return 0;
1339 /* Empty all the ctf_link_outputs. */
1340 static int
1341 ctf_link_empty_outputs (ctf_dict_t *fp)
1343 ctf_next_t *i = NULL;
1344 void *v;
1345 int err;
1347 ctf_dynhash_empty (fp->ctf_link_outputs);
1349 while ((err = ctf_dynhash_next (fp->ctf_link_inputs, &i, NULL, &v)) == 0)
1351 ctf_dict_t *in = (ctf_dict_t *) v;
1352 in->ctf_link_in_out = NULL;
1354 if (err != ECTF_NEXT_END)
1356 fp->ctf_flags &= ~LCTF_LINKING;
1357 ctf_err_warn (fp, 1, err, _("iteration error removing old outputs"));
1358 ctf_set_errno (fp, err);
1359 return -1;
1361 return 0;
1364 /* Do a deduplicating link using the ctf-dedup machinery. */
1365 static void
1366 ctf_link_deduplicating (ctf_dict_t *fp)
1368 size_t i;
1369 ctf_dict_t **inputs, **outputs = NULL;
1370 ssize_t ninputs;
1371 uint32_t noutputs;
1372 uint32_t *parents;
1374 if (ctf_dedup_atoms_init (fp) < 0)
1376 ctf_err_warn (fp, 0, 0, _("allocating CTF dedup atoms table"));
1377 return; /* Errno is set for us. */
1380 if (fp->ctf_link_out_cu_mapping
1381 && (ctf_link_deduplicating_per_cu (fp) < 0))
1382 return; /* Errno is set for us. */
1384 if ((ninputs = ctf_link_deduplicating_count_inputs (fp, NULL, NULL)) < 0)
1385 return; /* Errno is set for us. */
1387 if ((inputs = ctf_link_deduplicating_open_inputs (fp, NULL, ninputs,
1388 &parents)) == NULL)
1389 return; /* Errno is set for us. */
1391 if (ninputs == 1 && ctf_cuname (inputs[0]) != NULL)
1392 ctf_cuname_set (fp, ctf_cuname (inputs[0]));
1394 if (ctf_dedup (fp, inputs, ninputs, parents, 0) < 0)
1396 ctf_err_warn (fp, 0, 0, _("deduplication failed for %s"),
1397 ctf_link_input_name (fp));
1398 goto err;
1401 if ((outputs = ctf_dedup_emit (fp, inputs, ninputs, parents, &noutputs,
1402 0)) == NULL)
1404 ctf_err_warn (fp, 0, 0, _("deduplicating link type emission failed "
1405 "for %s"), ctf_link_input_name (fp));
1406 goto err;
1409 if (!ctf_assert (fp, outputs[0] == fp))
1411 for (i = 1; i < noutputs; i++)
1412 ctf_dict_close (outputs[i]);
1413 goto err;
1416 for (i = 0; i < noutputs; i++)
1418 char *dynname;
1420 /* We already have access to this one. Close the duplicate. */
1421 if (i == 0)
1423 ctf_dict_close (outputs[0]);
1424 continue;
1427 if ((dynname = ctf_new_per_cu_name (fp, ctf_cuname (outputs[i]))) == NULL)
1428 goto oom_one_output;
1430 if (ctf_dynhash_insert (fp->ctf_link_outputs, dynname, outputs[i]) < 0)
1431 goto oom_one_output;
1433 continue;
1435 oom_one_output:
1436 ctf_set_errno (fp, ENOMEM);
1437 ctf_err_warn (fp, 0, 0, _("out of memory allocating link outputs"));
1438 free (dynname);
1440 for (; i < noutputs; i++)
1441 ctf_dict_close (outputs[i]);
1442 goto err;
1445 if (!(fp->ctf_link_flags & CTF_LINK_OMIT_VARIABLES_SECTION)
1446 && ctf_link_deduplicating_variables (fp, inputs, ninputs, 0) < 0)
1448 ctf_err_warn (fp, 0, 0, _("deduplicating link variable emission failed for "
1449 "%s"), ctf_link_input_name (fp));
1450 goto err_clean_outputs;
1453 if (ctf_link_deduplicating_syms (fp, inputs, ninputs, 0) < 0)
1455 ctf_err_warn (fp, 0, 0, _("deduplicating link symbol emission failed for "
1456 "%s"), ctf_link_input_name (fp));
1457 goto err_clean_outputs;
1460 ctf_dedup_fini (fp, outputs, noutputs);
1462 /* Now close all the inputs, including per-CU intermediates. */
1464 if (ctf_link_deduplicating_close_inputs (fp, NULL, inputs, ninputs) < 0)
1465 return; /* errno is set for us. */
1467 ninputs = 0; /* Prevent double-close. */
1468 ctf_set_errno (fp, 0);
1470 /* Fall through. */
1472 err:
1473 for (i = 0; i < (size_t) ninputs; i++)
1474 ctf_dict_close (inputs[i]);
1475 free (inputs);
1476 free (parents);
1477 free (outputs);
1478 return;
1480 err_clean_outputs:
1481 ctf_link_empty_outputs (fp);
1482 goto err;
1485 /* Merge types and variable sections in all dicts added to the link together.
1486 The result of any previous link is discarded. */
1488 ctf_link (ctf_dict_t *fp, int flags)
1490 int err;
1492 fp->ctf_link_flags = flags;
1494 if (fp->ctf_link_inputs == NULL)
1495 return 0; /* Nothing to do. */
1497 if (fp->ctf_link_outputs != NULL)
1498 ctf_link_empty_outputs (fp);
1499 else
1500 fp->ctf_link_outputs = ctf_dynhash_create (ctf_hash_string,
1501 ctf_hash_eq_string, free,
1502 (ctf_hash_free_fun)
1503 ctf_dict_close);
1505 if (fp->ctf_link_outputs == NULL)
1506 return ctf_set_errno (fp, ENOMEM);
1508 fp->ctf_flags |= LCTF_LINKING;
1509 ctf_link_deduplicating (fp);
1510 fp->ctf_flags &= ~LCTF_LINKING;
1512 if ((ctf_errno (fp) != 0) && (ctf_errno (fp) != ECTF_NOCTFDATA))
1513 return -1;
1515 /* Create empty CUs if requested. We do not currently claim that multiple
1516 links in succession with CTF_LINK_EMPTY_CU_MAPPINGS set in some calls and
1517 not set in others will do anything especially sensible. */
1519 if (fp->ctf_link_out_cu_mapping && (flags & CTF_LINK_EMPTY_CU_MAPPINGS))
1521 ctf_next_t *i = NULL;
1522 void *k;
1524 while ((err = ctf_dynhash_next (fp->ctf_link_out_cu_mapping, &i, &k,
1525 NULL)) == 0)
1527 const char *to = (const char *) k;
1528 if (ctf_create_per_cu (fp, NULL, to) == NULL)
1530 fp->ctf_flags &= ~LCTF_LINKING;
1531 ctf_next_destroy (i);
1532 return -1; /* Errno is set for us. */
1535 if (err != ECTF_NEXT_END)
1537 fp->ctf_flags &= ~LCTF_LINKING;
1538 ctf_err_warn (fp, 1, err, _("iteration error creating empty CUs"));
1539 ctf_set_errno (fp, err);
1540 return -1;
1544 return 0;
1547 typedef struct ctf_link_out_string_cb_arg
1549 const char *str;
1550 uint32_t offset;
1551 int err;
1552 } ctf_link_out_string_cb_arg_t;
1554 /* Intern a string in the string table of an output per-CU CTF file. */
1555 static void
1556 ctf_link_intern_extern_string (void *key _libctf_unused_, void *value,
1557 void *arg_)
1559 ctf_dict_t *fp = (ctf_dict_t *) value;
1560 ctf_link_out_string_cb_arg_t *arg = (ctf_link_out_string_cb_arg_t *) arg_;
1562 fp->ctf_flags |= LCTF_DIRTY;
1563 if (!ctf_str_add_external (fp, arg->str, arg->offset))
1564 arg->err = ENOMEM;
1567 /* Repeatedly call ADD_STRING to acquire strings from the external string table,
1568 adding them to the atoms table for this CU and all subsidiary CUs.
1570 If ctf_link is also called, it must be called first if you want the new CTF
1571 files ctf_link can create to get their strings dedupped against the ELF
1572 strtab properly. */
1574 ctf_link_add_strtab (ctf_dict_t *fp, ctf_link_strtab_string_f *add_string,
1575 void *arg)
1577 const char *str;
1578 uint32_t offset;
1579 int err = 0;
1581 while ((str = add_string (&offset, arg)) != NULL)
1583 ctf_link_out_string_cb_arg_t iter_arg = { str, offset, 0 };
1585 fp->ctf_flags |= LCTF_DIRTY;
1586 if (!ctf_str_add_external (fp, str, offset))
1587 err = ENOMEM;
1589 ctf_dynhash_iter (fp->ctf_link_outputs, ctf_link_intern_extern_string,
1590 &iter_arg);
1591 if (iter_arg.err)
1592 err = iter_arg.err;
1595 if (err)
1596 ctf_set_errno (fp, err);
1598 return -err;
1601 /* Inform the ctf-link machinery of a new symbol in the target symbol table
1602 (which must be some symtab that is not usually stripped, and which
1603 is in agreement with ctf_bfdopen_ctfsect). May be called either before or
1604 after ctf_link_add_strtab. */
1606 ctf_link_add_linker_symbol (ctf_dict_t *fp, ctf_link_sym_t *sym)
1608 ctf_in_flight_dynsym_t *cid;
1610 /* Cheat a little: if there is already an ENOMEM error code recorded against
1611 this dict, we shouldn't even try to add symbols because there will be no
1612 memory to do so: probably we failed to add some previous symbol. This
1613 makes out-of-memory exits 'sticky' across calls to this function, so the
1614 caller doesn't need to worry about error conditions. */
1616 if (ctf_errno (fp) == ENOMEM)
1617 return -ENOMEM; /* errno is set for us. */
1619 if (ctf_symtab_skippable (sym))
1620 return 0;
1622 if (sym->st_type != STT_OBJECT && sym->st_type != STT_FUNC)
1623 return 0;
1625 /* Add the symbol to the in-flight list. */
1627 if ((cid = malloc (sizeof (ctf_in_flight_dynsym_t))) == NULL)
1628 goto oom;
1630 cid->cid_sym = *sym;
1631 ctf_list_append (&fp->ctf_in_flight_dynsyms, cid);
1633 return 0;
1635 oom:
1636 ctf_dynhash_destroy (fp->ctf_dynsyms);
1637 fp->ctf_dynsyms = NULL;
1638 ctf_set_errno (fp, ENOMEM);
1639 return -ENOMEM;
1642 /* Impose an ordering on symbols. The ordering takes effect immediately, but
1643 since the ordering info does not include type IDs, lookups may return nothing
1644 until such IDs are added by calls to ctf_add_*_sym. Must be called after
1645 ctf_link_add_strtab and ctf_link_add_linker_symbol. */
1647 ctf_link_shuffle_syms (ctf_dict_t *fp)
1649 ctf_in_flight_dynsym_t *did, *nid;
1650 ctf_next_t *i = NULL;
1651 int err = ENOMEM;
1652 void *name_, *sym_;
1654 if (!fp->ctf_dynsyms)
1656 fp->ctf_dynsyms = ctf_dynhash_create (ctf_hash_string,
1657 ctf_hash_eq_string,
1658 NULL, free);
1659 if (!fp->ctf_dynsyms)
1661 ctf_set_errno (fp, ENOMEM);
1662 return -ENOMEM;
1666 /* Add all the symbols, excluding only those we already know are prohibited
1667 from appearing in symtypetabs. */
1669 for (did = ctf_list_next (&fp->ctf_in_flight_dynsyms); did != NULL; did = nid)
1671 ctf_link_sym_t *new_sym;
1673 nid = ctf_list_next (did);
1674 ctf_list_delete (&fp->ctf_in_flight_dynsyms, did);
1676 /* We might get a name or an external strtab offset. The strtab offset is
1677 guaranteed resolvable at this point, so turn it into a string. */
1679 if (did->cid_sym.st_name == NULL)
1681 uint32_t off = CTF_SET_STID (did->cid_sym.st_nameidx, CTF_STRTAB_1);
1683 did->cid_sym.st_name = ctf_strraw (fp, off);
1684 did->cid_sym.st_nameidx_set = 0;
1685 if (!ctf_assert (fp, did->cid_sym.st_name != NULL))
1686 return -ECTF_INTERNAL; /* errno is set for us. */
1689 /* The symbol might have turned out to be nameless, so we have to recheck
1690 for skippability here. */
1691 if (!ctf_symtab_skippable (&did->cid_sym))
1693 ctf_dprintf ("symbol from linker: %s (%x)\n", did->cid_sym.st_name,
1694 did->cid_sym.st_symidx);
1696 if ((new_sym = malloc (sizeof (ctf_link_sym_t))) == NULL)
1697 goto local_oom;
1699 memcpy (new_sym, &did->cid_sym, sizeof (ctf_link_sym_t));
1700 if (ctf_dynhash_cinsert (fp->ctf_dynsyms, new_sym->st_name, new_sym) < 0)
1701 goto local_oom;
1703 if (fp->ctf_dynsymmax < new_sym->st_symidx)
1704 fp->ctf_dynsymmax = new_sym->st_symidx;
1707 free (did);
1708 continue;
1710 local_oom:
1711 free (did);
1712 free (new_sym);
1713 goto err;
1716 /* If no symbols are reported, unwind what we have done and return. This
1717 makes it a bit easier for the serializer to tell that no symbols have been
1718 reported and that it should look elsewhere for reported symbols. */
1719 if (!ctf_dynhash_elements (fp->ctf_dynsyms))
1721 ctf_dprintf ("No symbols: not a final link.\n");
1722 ctf_dynhash_destroy (fp->ctf_dynsyms);
1723 fp->ctf_dynsyms = NULL;
1724 return 0;
1727 /* Construct a mapping from shndx to the symbol info. */
1728 free (fp->ctf_dynsymidx);
1729 if ((fp->ctf_dynsymidx = calloc (fp->ctf_dynsymmax + 1,
1730 sizeof (ctf_link_sym_t *))) == NULL)
1731 goto err;
1733 while ((err = ctf_dynhash_next (fp->ctf_dynsyms, &i, &name_, &sym_)) == 0)
1735 const char *name = (const char *) name;
1736 ctf_link_sym_t *symp = (ctf_link_sym_t *) sym_;
1738 if (!ctf_assert (fp, symp->st_symidx <= fp->ctf_dynsymmax))
1740 ctf_next_destroy (i);
1741 err = ctf_errno (fp);
1742 goto err;
1744 fp->ctf_dynsymidx[symp->st_symidx] = symp;
1746 if (err != ECTF_NEXT_END)
1748 ctf_err_warn (fp, 0, err, _("error iterating over shuffled symbols"));
1749 goto err;
1751 return 0;
1753 err:
1754 /* Leave the in-flight symbols around: they'll be freed at
1755 dict close time regardless. */
1756 ctf_dynhash_destroy (fp->ctf_dynsyms);
1757 fp->ctf_dynsyms = NULL;
1758 free (fp->ctf_dynsymidx);
1759 fp->ctf_dynsymidx = NULL;
1760 fp->ctf_dynsymmax = 0;
1761 ctf_set_errno (fp, err);
1762 return -err;
1765 typedef struct ctf_name_list_accum_cb_arg
1767 char **names;
1768 ctf_dict_t *fp;
1769 ctf_dict_t **files;
1770 size_t i;
1771 char **dynames;
1772 size_t ndynames;
1773 } ctf_name_list_accum_cb_arg_t;
1775 /* Accumulate the names and a count of the names in the link output hash. */
1776 static void
1777 ctf_accumulate_archive_names (void *key, void *value, void *arg_)
1779 const char *name = (const char *) key;
1780 ctf_dict_t *fp = (ctf_dict_t *) value;
1781 char **names;
1782 ctf_dict_t **files;
1783 ctf_name_list_accum_cb_arg_t *arg = (ctf_name_list_accum_cb_arg_t *) arg_;
1785 if ((names = realloc (arg->names, sizeof (char *) * ++(arg->i))) == NULL)
1787 (arg->i)--;
1788 ctf_set_errno (arg->fp, ENOMEM);
1789 return;
1792 if ((files = realloc (arg->files, sizeof (ctf_dict_t *) * arg->i)) == NULL)
1794 (arg->i)--;
1795 ctf_set_errno (arg->fp, ENOMEM);
1796 return;
1799 /* Allow the caller to get in and modify the name at the last minute. If the
1800 caller *does* modify the name, we have to stash away the new name the
1801 caller returned so we can free it later on. (The original name is the key
1802 of the ctf_link_outputs hash and is freed by the dynhash machinery.) */
1804 if (fp->ctf_link_memb_name_changer)
1806 char **dynames;
1807 char *dyname;
1808 void *nc_arg = fp->ctf_link_memb_name_changer_arg;
1810 dyname = fp->ctf_link_memb_name_changer (fp, name, nc_arg);
1812 if (dyname != NULL)
1814 if ((dynames = realloc (arg->dynames,
1815 sizeof (char *) * ++(arg->ndynames))) == NULL)
1817 (arg->ndynames)--;
1818 ctf_set_errno (arg->fp, ENOMEM);
1819 return;
1821 arg->dynames = dynames;
1822 name = (const char *) dyname;
1826 arg->names = names;
1827 arg->names[(arg->i) - 1] = (char *) name;
1828 arg->files = files;
1829 arg->files[(arg->i) - 1] = fp;
1832 /* Change the name of the parent CTF section, if the name transformer has got to
1833 it. */
1834 static void
1835 ctf_change_parent_name (void *key _libctf_unused_, void *value, void *arg)
1837 ctf_dict_t *fp = (ctf_dict_t *) value;
1838 const char *name = (const char *) arg;
1840 ctf_parent_name_set (fp, name);
1843 /* Warn if we may suffer information loss because the CTF input files are too
1844 old. Usually we provide complete backward compatibility, but compiler
1845 changes etc which never hit a release may have a flag in the header that
1846 simply prevents those changes from being used. */
1847 static void
1848 ctf_link_warn_outdated_inputs (ctf_dict_t *fp)
1850 ctf_next_t *i = NULL;
1851 void *name_;
1852 void *input_;
1853 int err;
1855 while ((err = ctf_dynhash_next (fp->ctf_link_inputs, &i, &name_, &input_)) == 0)
1857 const char *name = (const char *) name_;
1858 ctf_link_input_t *input = (ctf_link_input_t *) input_;
1859 ctf_next_t *j = NULL;
1860 ctf_dict_t *ifp;
1861 int err;
1863 /* We only care about CTF archives by this point: lazy-opened archives
1864 have always been opened by this point, and short-circuited entries have
1865 a matching corresponding archive member. Entries with NULL clin_arc can
1866 exist, and constitute old entries renamed via a name changer: the
1867 renamed entries exist elsewhere in the list, so we can just skip
1868 those. */
1870 if (!input->clin_arc)
1871 continue;
1873 /* All entries in the archive will necessarily contain the same
1874 CTF_F_NEWFUNCINFO flag, so we only need to check the first. We don't
1875 even need to do that if we can't open it for any reason at all: the
1876 link will fail later on regardless, since an input can't be opened. */
1878 ifp = ctf_archive_next (input->clin_arc, &j, NULL, 0, &err);
1879 if (!ifp)
1880 continue;
1881 ctf_next_destroy (j);
1883 if (!(ifp->ctf_header->cth_flags & CTF_F_NEWFUNCINFO)
1884 && (ifp->ctf_header->cth_varoff - ifp->ctf_header->cth_funcoff) > 0)
1885 ctf_err_warn (fp, 1, 0, _("linker input %s has CTF func info but uses "
1886 "an old, unreleased func info format: "
1887 "this func info section will be dropped."),
1888 name);
1890 if (err != ECTF_NEXT_END)
1891 ctf_err_warn (fp, 0, err, _("error checking for outdated inputs"));
1894 /* Write out a CTF archive (if there are per-CU CTF files) or a CTF file
1895 (otherwise) into a new dynamically-allocated string, and return it.
1896 Members with sizes above THRESHOLD are compressed. */
1897 unsigned char *
1898 ctf_link_write (ctf_dict_t *fp, size_t *size, size_t threshold)
1900 ctf_name_list_accum_cb_arg_t arg;
1901 char **names;
1902 char *transformed_name = NULL;
1903 ctf_dict_t **files;
1904 FILE *f = NULL;
1905 size_t i;
1906 int err;
1907 long fsize;
1908 const char *errloc;
1909 unsigned char *buf = NULL;
1911 memset (&arg, 0, sizeof (ctf_name_list_accum_cb_arg_t));
1912 arg.fp = fp;
1913 fp->ctf_flags |= LCTF_LINKING;
1915 ctf_link_warn_outdated_inputs (fp);
1917 if (fp->ctf_link_outputs)
1919 ctf_dynhash_iter (fp->ctf_link_outputs, ctf_accumulate_archive_names, &arg);
1920 if (ctf_errno (fp) < 0)
1922 errloc = "hash creation";
1923 goto err;
1927 /* No extra outputs? Just write a simple ctf_dict_t. */
1928 if (arg.i == 0)
1930 unsigned char *ret = ctf_write_mem (fp, size, threshold);
1931 fp->ctf_flags &= ~LCTF_LINKING;
1932 return ret;
1935 /* Writing an archive. Stick ourselves (the shared repository, parent of all
1936 other archives) on the front of it with the default name. */
1937 if ((names = realloc (arg.names, sizeof (char *) * (arg.i + 1))) == NULL)
1939 errloc = "name reallocation";
1940 goto err_no;
1942 arg.names = names;
1943 memmove (&(arg.names[1]), arg.names, sizeof (char *) * (arg.i));
1945 arg.names[0] = (char *) _CTF_SECTION;
1946 if (fp->ctf_link_memb_name_changer)
1948 void *nc_arg = fp->ctf_link_memb_name_changer_arg;
1950 transformed_name = fp->ctf_link_memb_name_changer (fp, _CTF_SECTION,
1951 nc_arg);
1953 if (transformed_name != NULL)
1955 arg.names[0] = transformed_name;
1956 ctf_dynhash_iter (fp->ctf_link_outputs, ctf_change_parent_name,
1957 transformed_name);
1961 /* Propagate the link flags to all the dicts in this link. */
1962 for (i = 0; i < arg.i; i++)
1964 arg.files[i]->ctf_link_flags = fp->ctf_link_flags;
1965 arg.files[i]->ctf_flags |= LCTF_LINKING;
1968 if ((files = realloc (arg.files,
1969 sizeof (struct ctf_dict *) * (arg.i + 1))) == NULL)
1971 errloc = "ctf_dict reallocation";
1972 goto err_no;
1974 arg.files = files;
1975 memmove (&(arg.files[1]), arg.files, sizeof (ctf_dict_t *) * (arg.i));
1976 arg.files[0] = fp;
1978 if ((f = tmpfile ()) == NULL)
1980 errloc = "tempfile creation";
1981 goto err_no;
1984 if ((err = ctf_arc_write_fd (fileno (f), arg.files, arg.i + 1,
1985 (const char **) arg.names,
1986 threshold)) < 0)
1988 errloc = "archive writing";
1989 ctf_set_errno (fp, err);
1990 goto err;
1993 if (fseek (f, 0, SEEK_END) < 0)
1995 errloc = "seeking to end";
1996 goto err_no;
1999 if ((fsize = ftell (f)) < 0)
2001 errloc = "filesize determination";
2002 goto err_no;
2005 if (fseek (f, 0, SEEK_SET) < 0)
2007 errloc = "filepos resetting";
2008 goto err_no;
2011 if ((buf = malloc (fsize)) == NULL)
2013 errloc = "CTF archive buffer allocation";
2014 goto err_no;
2017 while (!feof (f) && fread (buf, fsize, 1, f) == 0)
2018 if (ferror (f))
2020 errloc = "reading archive from temporary file";
2021 goto err_no;
2024 *size = fsize;
2025 free (arg.names);
2026 free (arg.files);
2027 free (transformed_name);
2028 if (arg.ndynames)
2030 size_t i;
2031 for (i = 0; i < arg.ndynames; i++)
2032 free (arg.dynames[i]);
2033 free (arg.dynames);
2035 fclose (f);
2036 return buf;
2038 err_no:
2039 ctf_set_errno (fp, errno);
2041 /* Turn off the is-linking flag on all the dicts in this link. */
2042 for (i = 0; i < arg.i; i++)
2043 arg.files[i]->ctf_flags &= ~LCTF_LINKING;
2044 err:
2045 free (buf);
2046 if (f)
2047 fclose (f);
2048 free (arg.names);
2049 free (arg.files);
2050 free (transformed_name);
2051 if (arg.ndynames)
2053 size_t i;
2054 for (i = 0; i < arg.ndynames; i++)
2055 free (arg.dynames[i]);
2056 free (arg.dynames);
2058 ctf_err_warn (fp, 0, 0, _("cannot write archive in link: %s failure"),
2059 errloc);
2060 return NULL;