1 /* CTF type deduplication.
2 Copyright (C) 2019-2021 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
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/>. */
26 /* (In the below, relevant functions are named in square brackets.) */
28 /* Type deduplication is a three-phase process:
30 [ctf_dedup, ctf_dedup_hash_type, ctf_dedup_rhash_type]
31 1) come up with unambiguous hash values for all types: no two types may have
32 the same hash value, and any given type should have only one hash value
33 (for optimal deduplication).
35 [ctf_dedup, ctf_dedup_detect_name_ambiguity,
36 ctf_dedup_conflictify_unshared, ctf_dedup_mark_conflicting_hash]
37 2) mark those distinct types with names that collide (and thus cannot be
38 declared simultaneously in the same translation unit) as conflicting, and
39 recursively mark all types that cite one of those types as conflicting as
40 well. Possibly mark all types cited in only one TU as conflicting, if
41 the CTF_LINK_SHARE_DUPLICATED link mode is active.
43 [ctf_dedup_emit, ctf_dedup_emit_struct_members, ctf_dedup_id_to_target]
44 3) emit all the types, one hash value at a time. Types not marked
45 conflicting are emitted once, into the shared dictionary: types marked
46 conflicting are emitted once per TU into a dictionary corresponding to
47 each TU in which they appear. Structs marked conflicting get at the very
48 least a forward emitted into the shared dict so that other dicts can cite
52 This all works over an array of inputs (usually in the same order as the
53 inputs on the link line). We don't use the ctf_link_inputs hash directly
54 because it is convenient to be able to address specific input types as a
55 *global type ID* or 'GID', a pair of an array offset and a ctf_id_t. Since
56 both are already 32 bits or less or can easily be constrained to that range,
57 we can pack them both into a single 64-bit hash word for easy lookups, which
58 would be much more annoying to do with a ctf_dict_t * and a ctf_id_t. (On
59 32-bit platforms, we must do that anyway, since pointers, and thus hash keys
60 and values, are only 32 bits wide). We track which inputs are parents of
61 which other inputs so that we can correctly recognize that types we have
62 traversed in children may cite types in parents, and so that we can process
65 Note that thanks to ld -r, the deduplicator can be fed its own output, so the
66 inputs may themselves have child dicts. Since we need to support this usage
67 anyway, we can use it in one other place. If the caller finds translation
68 units to be too small a unit ambiguous types, links can be 'cu-mapped', where
69 the caller provides a mapping of input TU names to output child dict names.
70 This mapping can fuse many child TUs into one potential child dict, so that
71 ambiguous types in any of those input TUs go into the same child dict.
72 When a many:1 cu-mapping is detected, the ctf_dedup machinery is called
73 repeatedly, once for every output name that has more than one input, to fuse
74 all the input TUs associated with a given output dict into one, and once again
75 as normal to deduplicate all those intermediate outputs (and any 1:1 inputs)
76 together. This has much higher memory usage than otherwise, because in the
77 intermediate state, all the output TUs are in memory at once and cannot be
78 lazily opened. It also has implications for the emission code: if types
79 appear ambiguously in multiple input TUs that are all mapped to the same
80 child dict, we cannot put them in children in the cu-mapping link phase
81 because this output is meant to *become* a child in the next link stage and
82 parent/child relationships are only one level deep: so instead, we just hide
83 all but one of the ambiguous types.
85 There are a few other subtleties here that make this more complex than it
86 seems. Let's go over the steps above in more detail.
90 [ctf_dedup_hash_type, ctf_dedup_rhash_type]
91 Hashing proceeds recursively, mixing in the properties of each input type
92 (including its name, if any), and then adding the hash values of every type
93 cited by that type. The result is stashed in the cd_type_hashes so other
94 phases can find the hash values of input types given their IDs, and so that
95 if we encounter this type again while hashing we can just return its hash
96 value: it is also stashed in the *output mapping*, a mapping from hash value
97 to the set of GIDs corresponding to that type in all inputs. We also keep
98 track of the GID of the first appearance of the type in any input (in
99 cd_output_first_gid), and the GID of structs, unions, and forwards that only
100 appear in one TU (in cd_struct_origin). See below for where these things are
103 Everything in this phase is time-critical, because it is operating over
104 non-deduplicated types and so may have hundreds or thousands of times the
105 data volume to deal with than later phases. Trace output is hidden behind
106 ENABLE_LIBCTF_HASH_DEBUGGING to prevent the sheer number of calls to
107 ctf_dprintf from slowing things down (tenfold slowdowns are observed purely
108 from the calls to ctf_dprintf(), even with debugging switched off), and keep
109 down the volume of output (hundreds of gigabytes of debug output are not
110 uncommon on larger links).
112 We have to do *something* about potential cycles in the type graph. We'd
113 like to avoid emitting forwards in the final output if possible, because
114 forwards aren't much use: they have no members. We are mostly saved from
115 needing to worry about this at emission time by ctf_add_struct*()
116 automatically replacing newly-created forwards when the real struct/union
117 comes along. So we only have to avoid getting stuck in cycles during the
118 hashing phase, while also not confusing types that cite members that are
119 structs with each other. It is easiest to solve this problem by noting two
122 - all cycles in C depend on the presence of tagged structs/unions
123 - all tagged structs/unions have a unique name they can be disambiguated by
126 This means that we can break all cycles by ceasing to hash in cited types at
127 every tagged struct/union and instead hashing in a stub consisting of the
128 struct/union's *decorated name*, which is the name preceded by "s " or "u "
129 depending on the namespace (cached in cd_decorated_names). Forwards are
130 decorated identically (so a forward to "struct foo" would be represented as
131 "s foo"): this means that a citation of a forward to a type and a citation of
132 a concrete definition of a type with the same name ends up getting the same
135 Of course, it is quite possible to have two TUs with structs with the same
136 name and different definitions, but that's OK because when we scan for types
137 with ambiguous names we will identify these and mark them conflicting.
139 We populate one thing to help conflictedness marking. No unconflicted type
140 may cite a conflicted one, but this means that conflictedness marking must
141 walk from types to the types that cite them, which is the opposite of the
142 usual order. We can make this easier to do by constructing a *citers* graph
143 in cd_citers, which points from types to the types that cite them: because we
144 emit forwards corresponding to every conflicted struct/union, we don't need
145 to do this for citations of structs/unions by other types. This is very
146 convenient for us, because that's the only type we don't traverse
147 recursively: so we can construct the citers graph at the same time as we
148 hash, rather than needing to add an extra pass. (This graph is a dynhash of
149 *type hash values*, so it's small: in effect it is automatically
152 2) COLLISIONAL MARKING.
154 [ctf_dedup_detect_name_ambiguity, ctf_dedup_mark_conflicting_hash]
155 We identify types whose names collide during the hashing process, and count
156 the rough number of uses of each name (caching may throw it off a bit: this
157 doesn't need to be accurate). We then mark the less-frequently-cited types
158 with each names conflicting: the most-frequently-cited one goes into the
159 shared type dictionary, while all others are duplicated into per-TU
160 dictionaries, named after the input TU, that have the shared dictionary as a
161 parent. For structures and unions this is not quite good enough: we'd like
162 to have citations of forwards to ambiguously named structures and unions
163 *stay* as citations of forwards, so that the user can tell that the caller
164 didn't actually know which structure definition was meant: but if we put one
165 of those structures into the shared dictionary, it would supplant and replace
166 the forward, leaving no sign. So structures and unions do not take part in
167 this popularity contest: if their names are ambiguous, they are just
168 duplicated, and only a forward appears in the shared dict.
170 [ctf_dedup_propagate_conflictedness]
171 The process of marking types conflicted is itself recursive: we recursively
172 traverse the cd_citers graph populated in the hashing pass above and mark
173 everything that we encounter conflicted (without wasting time re-marking
174 anything that is already marked). This naturally terminates just where we
175 want it to (at types that are cited by no other types, and at structures and
176 unions) and suffices to ensure that types that cite conflicted types are
177 always marked conflicted.
179 [ctf_dedup_conflictify_unshared, ctf_dedup_multiple_input_dicts]
180 When linking in CTF_LINK_SHARE_DUPLICATED mode, we would like all types that
181 are used in only one TU to end up in a per-CU dict. The easiest way to do
182 that is to mark them conflicted. ctf_dedup_conflictify_unshared does this,
183 traversing the output mapping and using ctf_dedup_multiple_input_dicts to
184 check the number of input dicts each distinct type hash value came from:
185 types that only came from one get marked conflicted. One caveat here is that
186 we need to consider both structs and forwards to them: a struct that appears
187 in one TU and has a dozen citations to an opaque forward in other TUs should
188 *not* be considered to be used in only one TU, because users would find it
189 useful to be able to traverse into opaque structures of that sort: so we use
190 cd_struct_origin to check both structs/unions and the forwards corresponding
195 [ctf_dedup_walk_output_mapping, ctf_dedup_rwalk_output_mapping,
196 ctf_dedup_rwalk_one_output_mapping]
197 Emission involves another walk of the entire output mapping, this time
198 traversing everything other than struct members, recursively. Types are
199 emitted from leaves to trunk, emitting all types a type cites before emitting
200 the type itself. We sort the output mapping before traversing it, for
201 reproducibility and also correctness: the input dicts may have parent/child
202 relationships, so we simply sort all types that first appear in parents
203 before all children, then sort types that first appear in dicts appearing
204 earlier on the linker command line before those that appear later, then sort
205 by input ctf_id_t. (This is where we use cd_output_first_gid, collected
208 The walking is done using a recursive traverser which arranges to not revisit
209 any type already visited and to call its callback once per input GID for
210 input GIDs corresponding to conflicted output types. The traverser only
211 finds input types and calls a callback for them as many times as the output
212 needs to appear: it doesn't try to figure out anything about where the output
213 might go. That's done by the callback based on whether the type is
214 marked conflicted or not.
216 [ctf_dedup_emit_type, ctf_dedup_id_to_target, ctf_dedup_synthesize_forward]
217 ctf_dedup_emit_type is the (sole) callback for ctf_dedup_walk_output_mapping.
218 Conflicted types have all necessary dictionaries created, and then we emit
219 the type into each dictionary in turn, working over each input CTF type
220 corresponding to each hash value and using ctf_dedup_id_to_target to map each
221 input ctf_id_t into the corresponding type in the output (dealing with input
222 ctf_id_t's with parents in the process by simply chasing to the parent dict
223 if the type we're looking up is in there). Emitting structures involves
224 simply noting that the members of this structure need emission later on:
225 because you cannot cite a single structure member from another type, we avoid
226 emitting the members at this stage to keep recursion depths down a bit.
228 At this point, if we have by some mischance decided that two different types
229 with child types that hash to different values have in fact got the same hash
230 value themselves and *not* marked it conflicting, the type walk will walk
231 only *one* of them and in all likelihood we'll find that we are trying to
232 emit a type into some child dictionary that references a type that was never
233 emitted into that dictionary and assertion-fail. This always indicates a bug
234 in the conflictedness marking machinery or the hashing code, or both.
236 ctf_dedup_id_to_target calls ctf_dedup_synthesize_forward to do one extra
237 thing, alluded to above: if this is a conflicted tagged structure or union,
238 and the target is the shared dict (i.e., the type we're being asked to emit
239 is not itself conflicted so can't just point straight at the conflicted
240 type), we instead synthesise a forward with the same name, emit it into the
241 shared dict, record it in cd_output_emission_conflicted_forwards so that we
242 don't re-emit it, and return it. This means that cycles that contain
243 conflicts do not cause the entire cycle to be replicated in every child: only
244 that piece of the cycle which takes you back as far as the closest tagged
245 struct/union needs to be replicated. This trick means that no part of the
246 deduplicator needs a cycle detector: every recursive walk can stop at tagged
249 [ctf_dedup_emit_struct_members]
250 The final stage of emission is to walk over all structures with members
251 that need emission and emit all of them. Every type has been emitted at
252 this stage, so emission cannot fail.
254 [ctf_dedup_populate_type_mappings, ctf_dedup_populate_type_mapping]
255 Finally, we update the input -> output type ID mappings used by the ctf-link
256 machinery to update all the other sections. This is surprisingly expensive
257 and may be replaced with a scheme which lets the ctf-link machinery extract
258 the needed info directly from the deduplicator. */
260 /* Possible future optimizations are flagged with 'optimization opportunity'
263 /* Global optimization opportunity: a GC pass, eliminating types with no direct
264 or indirect citations from the other sections in the dictionary. */
266 /* Internal flag values for ctf_dedup_hash_type. */
268 /* Child call: consider forwardable types equivalent to forwards or stubs below
270 #define CTF_DEDUP_HASH_INTERNAL_CHILD 0x01
272 /* Transform references to single ctf_id_ts in passed-in inputs into a number
273 that will fit in a uint64_t. Needs rethinking if CTF_MAX_TYPE is boosted.
275 On 32-bit platforms, we pack things together differently: see the note
278 #if UINTPTR_MAX < UINT64_MAX
279 # define IDS_NEED_ALLOCATION 1
280 # define CTF_DEDUP_GID(fp, input, type) id_to_packed_id (fp, input, type)
281 # define CTF_DEDUP_GID_TO_INPUT(id) packed_id_to_input (id)
282 # define CTF_DEDUP_GID_TO_TYPE(id) packed_id_to_type (id)
284 # define CTF_DEDUP_GID(fp, input, type) \
285 (void *) (((uint64_t) input) << 32 | (type))
286 # define CTF_DEDUP_GID_TO_INPUT(id) ((int) (((uint64_t) id) >> 32))
287 # define CTF_DEDUP_GID_TO_TYPE(id) (ctf_id_t) (((uint64_t) id) & ~(0xffffffff00000000ULL))
290 #ifdef IDS_NEED_ALLOCATION
292 /* This is the 32-bit path, which stores GIDs in a pool and returns a pointer
293 into the pool. It is notably less efficient than the 64-bit direct storage
294 approach, but with a smaller key, this is all we can do. */
297 id_to_packed_id (ctf_dict_t
*fp
, int input_num
, ctf_id_t type
)
300 ctf_type_id_key_t
*dynkey
= NULL
;
301 ctf_type_id_key_t key
= { input_num
, type
};
303 if (!ctf_dynhash_lookup_kv (fp
->ctf_dedup
.cd_id_to_dict_t
,
304 &key
, &lookup
, NULL
))
306 if ((dynkey
= malloc (sizeof (ctf_type_id_key_t
))) == NULL
)
308 memcpy (dynkey
, &key
, sizeof (ctf_type_id_key_t
));
310 if (ctf_dynhash_insert (fp
->ctf_dedup
.cd_id_to_dict_t
, dynkey
, NULL
) < 0)
313 ctf_dynhash_lookup_kv (fp
->ctf_dedup
.cd_id_to_dict_t
,
314 dynkey
, &lookup
, NULL
);
316 /* We use a raw assert() here because there isn't really a way to get any sort
317 of error back from this routine without vastly complicating things for the
318 much more common case of !IDS_NEED_ALLOCATION. */
320 return (void *) lookup
;
324 ctf_set_errno (fp
, ENOMEM
);
329 packed_id_to_input (const void *id
)
331 const ctf_type_id_key_t
*key
= (ctf_type_id_key_t
*) id
;
333 return key
->ctii_input_num
;
337 packed_id_to_type (const void *id
)
339 const ctf_type_id_key_t
*key
= (ctf_type_id_key_t
*) id
;
341 return key
->ctii_type
;
345 /* Make an element in a dynhash-of-dynsets, or return it if already present. */
347 static ctf_dynset_t
*
348 make_set_element (ctf_dynhash_t
*set
, const void *key
)
350 ctf_dynset_t
*element
;
352 if ((element
= ctf_dynhash_lookup (set
, key
)) == NULL
)
354 if ((element
= ctf_dynset_create (htab_hash_string
,
355 ctf_dynset_eq_string
,
359 if (ctf_dynhash_insert (set
, (void *) key
, element
) < 0)
361 ctf_dynset_destroy (element
);
369 /* Initialize the dedup atoms table. */
371 ctf_dedup_atoms_init (ctf_dict_t
*fp
)
373 if (fp
->ctf_dedup_atoms
)
376 if (!fp
->ctf_dedup_atoms_alloc
)
378 if ((fp
->ctf_dedup_atoms_alloc
379 = ctf_dynset_create (htab_hash_string
, ctf_dynset_eq_string
,
381 return ctf_set_errno (fp
, ENOMEM
);
383 fp
->ctf_dedup_atoms
= fp
->ctf_dedup_atoms_alloc
;
387 /* Intern things in the dedup atoms table. */
390 intern (ctf_dict_t
*fp
, char *atom
)
397 if (!ctf_dynset_exists (fp
->ctf_dedup_atoms
, atom
, &foo
))
399 if (ctf_dynset_insert (fp
->ctf_dedup_atoms
, atom
) < 0)
401 ctf_set_errno (fp
, ENOMEM
);
409 return (const char *) foo
;
412 /* Add an indication of the namespace to a type name in a way that is not valid
413 for C identifiers. Used to maintain hashes of type names to other things
414 while allowing for the four C namespaces (normal, struct, union, enum).
415 Return a new dynamically-allocated string. */
417 ctf_decorate_type_name (ctf_dict_t
*fp
, const char *name
, int kind
)
419 ctf_dedup_t
*d
= &fp
->ctf_dedup
;
444 if ((ret
= ctf_dynhash_lookup (d
->cd_decorated_names
[i
], name
)) == NULL
)
448 if ((str
= malloc (strlen (name
) + strlen (k
) + 1)) == NULL
)
453 ret
= intern (fp
, str
);
457 if (ctf_dynhash_cinsert (d
->cd_decorated_names
[i
], name
, ret
) < 0)
464 ctf_set_errno (fp
, ENOMEM
);
468 /* Hash a type, possibly debugging-dumping something about it as well. */
470 ctf_dedup_sha1_add (ctf_sha1_t
*sha1
, const void *buf
, size_t len
,
471 const char *description _libctf_unused_
,
472 unsigned long depth _libctf_unused_
)
474 ctf_sha1_add (sha1
, buf
, len
);
476 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
478 char tmp_hval
[CTF_SHA1_SIZE
];
480 ctf_sha1_fini (&tmp
, tmp_hval
);
481 ctf_dprintf ("%lu: after hash addition of %s: %s\n", depth
, description
,
487 ctf_dedup_hash_type (ctf_dict_t
*fp
, ctf_dict_t
*input
,
488 ctf_dict_t
**inputs
, uint32_t *parents
,
489 int input_num
, ctf_id_t type
, int flags
,
491 int (*populate_fun
) (ctf_dict_t
*fp
,
497 const char *decorated_name
,
500 /* Determine whether this type is being hashed as a stub (in which case it is
501 unsafe to cache it). */
503 ctf_dedup_is_stub (const char *name
, int kind
, int fwdkind
, int flags
)
505 /* We can cache all types unless we are recursing to children and are hashing
506 in a tagged struct, union or forward, all of which are replaced with their
507 decorated name as a stub and will have different hash values when hashed at
510 return ((flags
& CTF_DEDUP_HASH_INTERNAL_CHILD
) && name
511 && (kind
== CTF_K_STRUCT
|| kind
== CTF_K_UNION
512 || (kind
== CTF_K_FORWARD
&& (fwdkind
== CTF_K_STRUCT
513 || fwdkind
== CTF_K_UNION
))));
516 /* Populate struct_origin if need be (not already populated, or populated with
517 a different origin), in which case it must go to -1, "shared".)
519 Only called for forwards or forwardable types with names, when the link mode
520 is CTF_LINK_SHARE_DUPLICATED. */
522 ctf_dedup_record_origin (ctf_dict_t
*fp
, int input_num
, const char *decorated
,
525 ctf_dedup_t
*d
= &fp
->ctf_dedup
;
527 int populate_origin
= 0;
529 if (ctf_dynhash_lookup_kv (d
->cd_struct_origin
, decorated
, NULL
, &origin
))
531 if (CTF_DEDUP_GID_TO_INPUT (origin
) != input_num
532 && CTF_DEDUP_GID_TO_INPUT (origin
) != -1)
535 origin
= CTF_DEDUP_GID (fp
, -1, -1);
545 if (ctf_dynhash_cinsert (d
->cd_struct_origin
, decorated
, origin
) < 0)
546 return ctf_set_errno (fp
, errno
);
550 /* Do the underlying hashing and recursion for ctf_dedup_hash_type (which it
551 calls, recursively). */
554 ctf_dedup_rhash_type (ctf_dict_t
*fp
, ctf_dict_t
*input
, ctf_dict_t
**inputs
,
555 uint32_t *parents
, int input_num
, ctf_id_t type
,
556 void *type_id
, const ctf_type_t
*tp
, const char *name
,
557 const char *decorated
, int kind
, int flags
,
559 int (*populate_fun
) (ctf_dict_t
*fp
,
565 const char *decorated_name
,
568 ctf_dedup_t
*d
= &fp
->ctf_dedup
;
569 ctf_next_t
*i
= NULL
;
572 char hashbuf
[CTF_SHA1_SIZE
];
573 const char *hval
= NULL
;
577 const char *citer
= NULL
;
578 ctf_dynset_t
*citers
= NULL
;
580 /* Add a citer to the citers set. */
581 #define ADD_CITER(citers, hval) \
584 whaterr = N_("error updating citers"); \
586 if ((citers = ctf_dynset_create (htab_hash_string, \
587 ctf_dynset_eq_string, \
590 if (ctf_dynset_cinsert (citers, hval) < 0) \
594 /* If this is a named struct or union or a forward to one, and this is a child
595 traversal, treat this type as if it were a forward -- do not recurse to
596 children, ignore all content not already hashed in, and hash in the
597 decorated name of the type instead. */
599 if (ctf_dedup_is_stub (name
, kind
, tp
->ctt_type
, flags
))
601 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
602 ctf_dprintf ("Struct/union/forward citation: substituting forwarding "
603 "stub with decorated name %s\n", decorated
);
606 ctf_sha1_init (&hash
);
607 ctf_dedup_sha1_add (&hash
, decorated
, strlen (decorated
) + 1,
608 "decorated struct/union/forward name", depth
);
609 ctf_sha1_fini (&hash
, hashbuf
);
611 if ((hval
= intern (fp
, strdup (hashbuf
))) == NULL
)
613 ctf_err_warn (fp
, 0, 0, _("%s (%i): out of memory during forwarding-"
614 "stub hashing for type with GID %p"),
615 ctf_link_input_name (input
), input_num
, type_id
);
616 return NULL
; /* errno is set for us. */
619 /* In share-duplicated link mode, make sure the origin of this type is
620 recorded, even if this is a type in a parent dict which will not be
621 directly traversed. */
622 if (d
->cd_link_flags
& CTF_LINK_SHARE_DUPLICATED
623 && ctf_dedup_record_origin (fp
, input_num
, decorated
, type_id
) < 0)
624 return NULL
; /* errno is set for us. */
629 /* Now ensure that subsequent recursive calls (but *not* the top-level call)
630 get this treatment. */
631 flags
|= CTF_DEDUP_HASH_INTERNAL_CHILD
;
633 /* If this is a struct, union, or forward with a name, record the unique
634 originating input TU, if there is one. */
636 if (decorated
&& (ctf_forwardable_kind (kind
) || kind
!= CTF_K_FORWARD
))
637 if (d
->cd_link_flags
& CTF_LINK_SHARE_DUPLICATED
638 && ctf_dedup_record_origin (fp
, input_num
, decorated
, type_id
) < 0)
639 return NULL
; /* errno is set for us. */
641 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
642 ctf_dprintf ("%lu: hashing thing with ID %i/%lx (kind %i): %s.\n",
643 depth
, input_num
, type
, kind
, name
? name
: "");
646 /* Some type kinds don't have names: the API provides no way to set the name,
647 so the type the deduplicator outputs will be nameless even if the input
648 somehow has a name, and the name should not be mixed into the hash. */
662 /* Mix in invariant stuff, transforming the type kind if needed. Note that
663 the vlen is *not* hashed in: the actual variable-length info is hashed in
664 instead, piecewise. The vlen is not part of the type, only the
665 variable-length data is: identical types with distinct vlens are quite
666 possible. Equally, we do not want to hash in the isroot flag: both the
667 compiler and the deduplicator set the nonroot flag to indicate clashes with
668 *other types in the same TU* with the same name: so two types can easily
669 have distinct nonroot flags, yet be exactly the same type.*/
671 ctf_sha1_init (&hash
);
673 ctf_dedup_sha1_add (&hash
, name
, strlen (name
) + 1, "name", depth
);
674 ctf_dedup_sha1_add (&hash
, &kind
, sizeof (uint32_t), "kind", depth
);
676 /* Hash content of this type. */
680 /* No extra state. */
684 /* Add the forwarded kind, stored in the ctt_type. */
685 ctf_dedup_sha1_add (&hash
, &tp
->ctt_type
, sizeof (tp
->ctt_type
),
686 "forwarded kind", depth
);
692 memset (&ep
, 0, sizeof (ctf_encoding_t
));
694 ctf_dedup_sha1_add (&hash
, &tp
->ctt_size
, sizeof (uint32_t), "size",
696 if (ctf_type_encoding (input
, type
, &ep
) < 0)
698 whaterr
= N_("error getting encoding");
701 ctf_dedup_sha1_add (&hash
, &ep
, sizeof (ctf_encoding_t
), "encoding",
705 /* Types that reference other types. */
711 /* Hash the referenced type, if not already hashed, and mix it in. */
712 child_type
= ctf_type_reference (input
, type
);
713 if ((hval
= ctf_dedup_hash_type (fp
, input
, inputs
, parents
, input_num
,
714 child_type
, flags
, depth
,
715 populate_fun
)) == NULL
)
717 whaterr
= N_("error doing referenced type hashing");
720 ctf_dedup_sha1_add (&hash
, hval
, strlen (hval
) + 1, "referenced type",
726 /* The slices of two types hash identically only if the type they overlay
727 also has the same encoding. This is not ideal, but in practice will work
728 well enough. We work directly rather than using the CTF API because
729 we do not want the slice's normal automatically-shine-through
730 semantics to kick in here. */
733 const ctf_slice_t
*slice
;
734 const ctf_dtdef_t
*dtd
;
738 child_type
= ctf_type_reference (input
, type
);
739 ctf_get_ctt_size (input
, tp
, &size
, &increment
);
740 ctf_dedup_sha1_add (&hash
, &size
, sizeof (ssize_t
), "size", depth
);
742 if ((hval
= ctf_dedup_hash_type (fp
, input
, inputs
, parents
, input_num
,
743 child_type
, flags
, depth
,
744 populate_fun
)) == NULL
)
746 whaterr
= N_("error doing slice-referenced type hashing");
749 ctf_dedup_sha1_add (&hash
, hval
, strlen (hval
) + 1, "sliced type",
753 if ((dtd
= ctf_dynamic_type (input
, type
)) != NULL
)
754 slice
= &dtd
->dtd_u
.dtu_slice
;
756 slice
= (ctf_slice_t
*) ((uintptr_t) tp
+ increment
);
758 ctf_dedup_sha1_add (&hash
, &slice
->cts_offset
,
759 sizeof (slice
->cts_offset
), "slice offset", depth
);
760 ctf_dedup_sha1_add (&hash
, &slice
->cts_bits
,
761 sizeof (slice
->cts_bits
), "slice bits", depth
);
769 if (ctf_array_info (input
, type
, &ar
) < 0)
771 whaterr
= N_("error getting array info");
775 if ((hval
= ctf_dedup_hash_type (fp
, input
, inputs
, parents
, input_num
,
776 ar
.ctr_contents
, flags
, depth
,
777 populate_fun
)) == NULL
)
779 whaterr
= N_("error doing array contents type hashing");
782 ctf_dedup_sha1_add (&hash
, hval
, strlen (hval
) + 1, "array contents",
784 ADD_CITER (citers
, hval
);
786 if ((hval
= ctf_dedup_hash_type (fp
, input
, inputs
, parents
, input_num
,
787 ar
.ctr_index
, flags
, depth
,
788 populate_fun
)) == NULL
)
790 whaterr
= N_("error doing array index type hashing");
793 ctf_dedup_sha1_add (&hash
, hval
, strlen (hval
) + 1, "array index",
795 ctf_dedup_sha1_add (&hash
, &ar
.ctr_nelems
, sizeof (ar
.ctr_nelems
),
796 "element count", depth
);
797 ADD_CITER (citers
, hval
);
807 if (ctf_func_type_info (input
, type
, &fi
) < 0)
809 whaterr
= N_("error getting func type info");
813 if ((hval
= ctf_dedup_hash_type (fp
, input
, inputs
, parents
, input_num
,
814 fi
.ctc_return
, flags
, depth
,
815 populate_fun
)) == NULL
)
817 whaterr
= N_("error getting func return type");
820 ctf_dedup_sha1_add (&hash
, hval
, strlen (hval
) + 1, "func return",
822 ctf_dedup_sha1_add (&hash
, &fi
.ctc_argc
, sizeof (fi
.ctc_argc
),
824 ctf_dedup_sha1_add (&hash
, &fi
.ctc_flags
, sizeof (fi
.ctc_flags
),
825 "func flags", depth
);
826 ADD_CITER (citers
, hval
);
828 if ((args
= calloc (fi
.ctc_argc
, sizeof (ctf_id_t
))) == NULL
)
830 whaterr
= N_("error doing memory allocation");
834 if (ctf_func_type_args (input
, type
, fi
.ctc_argc
, args
) < 0)
837 whaterr
= N_("error getting func arg type");
840 for (j
= 0; j
< fi
.ctc_argc
; j
++)
842 if ((hval
= ctf_dedup_hash_type (fp
, input
, inputs
, parents
,
843 input_num
, args
[j
], flags
, depth
,
844 populate_fun
)) == NULL
)
847 whaterr
= N_("error doing func arg type hashing");
850 ctf_dedup_sha1_add (&hash
, hval
, strlen (hval
) + 1, "func arg type",
852 ADD_CITER (citers
, hval
);
862 ctf_dedup_sha1_add (&hash
, &tp
->ctt_size
, sizeof (uint32_t),
864 while ((ename
= ctf_enum_next (input
, type
, &i
, &val
)) != NULL
)
866 ctf_dedup_sha1_add (&hash
, ename
, strlen (ename
) + 1, "enumerator",
868 ctf_dedup_sha1_add (&hash
, &val
, sizeof (val
), "enumerand", depth
);
870 if (ctf_errno (input
) != ECTF_NEXT_END
)
872 whaterr
= N_("error doing enum member iteration");
877 /* Top-level only. */
886 ctf_get_ctt_size (input
, tp
, &size
, NULL
);
887 ctf_dedup_sha1_add (&hash
, &size
, sizeof (ssize_t
), "struct size",
890 while ((offset
= ctf_member_next (input
, type
, &i
, &mname
, &membtype
,
895 ctf_dedup_sha1_add (&hash
, mname
, strlen (mname
) + 1,
896 "member name", depth
);
898 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
899 ctf_dprintf ("%lu: Traversing to member %s\n", depth
, mname
);
901 if ((hval
= ctf_dedup_hash_type (fp
, input
, inputs
, parents
,
902 input_num
, membtype
, flags
, depth
,
903 populate_fun
)) == NULL
)
905 whaterr
= N_("error doing struct/union member type hashing");
909 ctf_dedup_sha1_add (&hash
, hval
, strlen (hval
) + 1, "member hash",
911 ctf_dedup_sha1_add (&hash
, &offset
, sizeof (offset
), "member offset",
913 ADD_CITER (citers
, hval
);
915 if (ctf_errno (input
) != ECTF_NEXT_END
)
917 whaterr
= N_("error doing struct/union member iteration");
923 whaterr
= N_("error: unknown type kind");
926 ctf_sha1_fini (&hash
, hashbuf
);
928 if ((hval
= intern (fp
, strdup (hashbuf
))) == NULL
)
930 whaterr
= N_("cannot intern hash");
934 /* Populate the citers for this type's subtypes, now the hash for the type
936 whaterr
= N_("error tracking citers");
940 ctf_dynset_t
*citer_hashes
;
942 if ((citer_hashes
= make_set_element (d
->cd_citers
, citer
)) == NULL
)
944 if (ctf_dynset_cinsert (citer_hashes
, hval
) < 0)
951 while ((err
= ctf_dynset_cnext (citers
, &i
, &k
)) == 0)
953 ctf_dynset_t
*citer_hashes
;
954 citer
= (const char *) k
;
956 if ((citer_hashes
= make_set_element (d
->cd_citers
, citer
)) == NULL
)
959 if (ctf_dynset_exists (citer_hashes
, hval
, NULL
))
961 if (ctf_dynset_cinsert (citer_hashes
, hval
) < 0)
964 if (err
!= ECTF_NEXT_END
)
966 ctf_dynset_destroy (citers
);
972 ctf_next_destroy (i
);
974 ctf_sha1_fini (&hash
, NULL
);
975 ctf_err_warn (fp
, 0, 0, _("%s (%i): %s: during type hashing for type %lx, "
976 "kind %i"), ctf_link_input_name (input
),
977 input_num
, gettext (whaterr
), type
, kind
);
980 ctf_set_errno (fp
, errno
);
981 ctf_err_warn (fp
, 0, 0, _("%s (%i): %s: during type hashing for type %lx, "
982 "kind %i"), ctf_link_input_name (input
),
983 input_num
, gettext (whaterr
), type
, kind
);
987 /* Hash a TYPE in the INPUT: FP is the eventual output, where the ctf_dedup
988 state is stored. INPUT_NUM is the number of this input in the set of inputs.
989 Record its hash in FP's cd_type_hashes once it is known. PARENTS is
990 described in the comment above ctf_dedup.
992 (The flags argument currently accepts only the flag
993 CTF_DEDUP_HASH_INTERNAL_CHILD, an implementation detail used to prevent
994 struct/union hashing in recursive traversals below the TYPE.)
996 We use the CTF API rather than direct access wherever possible, because types
997 that appear identical through the API should be considered identical, with
998 one exception: slices should only be considered identical to other slices,
999 not to the corresponding unsliced type.
1001 The POPULATE_FUN is a mandatory hook that populates other mappings with each
1002 type we see (excepting types that are recursively hashed as stubs). The
1003 caller should not rely on the order of calls to this hook, though it will be
1004 called at least once for every non-stub reference to every type.
1006 Returns a hash value (an atom), or NULL on error. */
1009 ctf_dedup_hash_type (ctf_dict_t
*fp
, ctf_dict_t
*input
,
1010 ctf_dict_t
**inputs
, uint32_t *parents
,
1011 int input_num
, ctf_id_t type
, int flags
,
1012 unsigned long depth
,
1013 int (*populate_fun
) (ctf_dict_t
*fp
,
1015 ctf_dict_t
**inputs
,
1019 const char *decorated_name
,
1022 ctf_dedup_t
*d
= &fp
->ctf_dedup
;
1023 const ctf_type_t
*tp
;
1025 const char *hval
= NULL
;
1027 const char *whaterr
;
1028 const char *decorated
= NULL
;
1029 uint32_t kind
, fwdkind
;
1033 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
1034 ctf_dprintf ("%lu: ctf_dedup_hash_type (%i, %lx, flags %x)\n", depth
, input_num
, type
, flags
);
1037 /* The unimplemented type doesn't really exist, but must be noted in parent
1038 hashes: so it gets a fixed, arbitrary hash. */
1040 return "00000000000000000000";
1042 /* Possible optimization: if the input type is in the parent type space, just
1043 copy recursively-cited hashes from the parent's types into the output
1044 mapping rather than rehashing them. */
1046 type_id
= CTF_DEDUP_GID (fp
, input_num
, type
);
1048 if ((tp
= ctf_lookup_by_id (&input
, type
)) == NULL
)
1050 ctf_set_errno (fp
, ctf_errno (input
));
1051 ctf_err_warn (fp
, 0, 0, _("%s (%i): lookup failure for type %lx: "
1052 "flags %x"), ctf_link_input_name (input
),
1053 input_num
, type
, flags
);
1054 return NULL
; /* errno is set for us. */
1057 kind
= LCTF_INFO_KIND (input
, tp
->ctt_info
);
1058 name
= ctf_strraw (input
, tp
->ctt_name
);
1060 if (tp
->ctt_name
== 0 || !name
|| name
[0] == '\0')
1063 /* Treat the unknown kind just like the unimplemented type. */
1064 if (kind
== CTF_K_UNKNOWN
)
1065 return "00000000000000000000";
1067 /* Decorate the name appropriately for the namespace it appears in: forwards
1068 appear in the namespace of their referent. */
1073 if (kind
== CTF_K_FORWARD
)
1074 fwdkind
= tp
->ctt_type
;
1076 if ((decorated
= ctf_decorate_type_name (fp
, name
, fwdkind
)) == NULL
)
1077 return NULL
; /* errno is set for us. */
1080 /* If not hashing a stub, we can rely on various sorts of caches.
1082 Optimization opportunity: we may be able to avoid calling the populate_fun
1085 if (!ctf_dedup_is_stub (name
, kind
, fwdkind
, flags
))
1087 if ((hval
= ctf_dynhash_lookup (d
->cd_type_hashes
, type_id
)) != NULL
)
1089 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
1090 ctf_dprintf ("%lu: Known hash for ID %i/%lx: %s\n", depth
, input_num
,
1093 populate_fun (fp
, input
, inputs
, input_num
, type
, type_id
,
1100 /* We have never seen this type before, and must figure out its hash and the
1101 hashes of the types it cites.
1103 Hash this type, and call ourselves recursively. (The hashing part is
1104 optional, and is disabled if overidden_hval is set.) */
1106 if ((hval
= ctf_dedup_rhash_type (fp
, input
, inputs
, parents
, input_num
,
1107 type
, type_id
, tp
, name
, decorated
,
1108 kind
, flags
, depth
, populate_fun
)) == NULL
)
1109 return NULL
; /* errno is set for us. */
1111 /* The hash of this type is now known: record it unless caching is unsafe
1112 because the hash value will change later. This will be the final storage
1113 of this type's hash, so we call the population function on it. */
1115 if (!ctf_dedup_is_stub (name
, kind
, fwdkind
, flags
))
1117 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
1118 ctf_dprintf ("Caching %lx, ID %p (%s), %s in final location\n", type
,
1119 type_id
, name
? name
: "", hval
);
1122 if (ctf_dynhash_cinsert (d
->cd_type_hashes
, type_id
, hval
) < 0)
1124 whaterr
= N_("error hash caching");
1128 if (populate_fun (fp
, input
, inputs
, input_num
, type
, type_id
,
1129 decorated
, hval
) < 0)
1131 whaterr
= N_("error calling population function");
1132 goto err
; /* errno is set for us. */
1136 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
1137 ctf_dprintf ("%lu: Returning final hash for ID %i/%lx: %s\n", depth
,
1138 input_num
, type
, hval
);
1143 ctf_set_errno (fp
, errno
);
1145 ctf_err_warn (fp
, 0, 0, _("%s (%i): %s: during type hashing, "
1146 "type %lx, kind %i"),
1147 ctf_link_input_name (input
), input_num
,
1148 gettext (whaterr
), type
, kind
);
1152 /* Populate a number of useful mappings not directly used by the hashing
1153 machinery: the output mapping, the cd_name_counts mapping from name -> hash
1154 -> count of hashval deduplication state for a given hashed type, and the
1155 cd_output_first_tu mapping. */
1158 ctf_dedup_populate_mappings (ctf_dict_t
*fp
, ctf_dict_t
*input _libctf_unused_
,
1159 ctf_dict_t
**inputs _libctf_unused_
,
1160 int input_num _libctf_unused_
,
1161 ctf_id_t type _libctf_unused_
, void *id
,
1162 const char *decorated_name
,
1165 ctf_dedup_t
*d
= &fp
->ctf_dedup
;
1166 ctf_dynset_t
*type_ids
;
1167 ctf_dynhash_t
*name_counts
;
1170 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
1171 ctf_dprintf ("Hash %s, %s, into output mapping for %i/%lx @ %s\n",
1172 hval
, decorated_name
? decorated_name
: "(unnamed)",
1173 input_num
, type
, ctf_link_input_name (input
));
1175 const char *orig_hval
;
1177 /* Make sure we never map a single GID to multiple hash values. */
1179 if ((orig_hval
= ctf_dynhash_lookup (d
->cd_output_mapping_guard
, id
)) != NULL
)
1181 /* We can rely on pointer identity here, since all hashes are
1183 if (!ctf_assert (fp
, orig_hval
== hval
))
1187 if (ctf_dynhash_cinsert (d
->cd_output_mapping_guard
, id
, hval
) < 0)
1188 return ctf_set_errno (fp
, errno
);
1191 /* Record the type in the output mapping: if this is the first time this type
1192 has been seen, also record it in the cd_output_first_gid. Because we
1193 traverse types in TU order and we do not merge types after the hashing
1194 phase, this will be the lowest TU this type ever appears in. */
1196 if ((type_ids
= ctf_dynhash_lookup (d
->cd_output_mapping
,
1199 if (ctf_dynhash_cinsert (d
->cd_output_first_gid
, hval
, id
) < 0)
1200 return ctf_set_errno (fp
, errno
);
1202 if ((type_ids
= ctf_dynset_create (htab_hash_pointer
,
1205 return ctf_set_errno (fp
, errno
);
1206 if (ctf_dynhash_insert (d
->cd_output_mapping
, (void *) hval
,
1209 ctf_dynset_destroy (type_ids
);
1210 return ctf_set_errno (fp
, errno
);
1213 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
1215 /* Verify that all types with this hash are of the same kind, and that the
1216 first TU a type was seen in never falls. */
1220 ctf_next_t
*i
= NULL
;
1221 int orig_kind
= ctf_type_kind_unsliced (input
, type
);
1224 orig_first_tu
= CTF_DEDUP_GID_TO_INPUT
1225 (ctf_dynhash_lookup (d
->cd_output_first_gid
, hval
));
1226 if (!ctf_assert (fp
, orig_first_tu
<= CTF_DEDUP_GID_TO_INPUT (id
)))
1229 while ((err
= ctf_dynset_cnext (type_ids
, &i
, &one_id
)) == 0)
1231 ctf_dict_t
*foo
= inputs
[CTF_DEDUP_GID_TO_INPUT (one_id
)];
1232 ctf_id_t bar
= CTF_DEDUP_GID_TO_TYPE (one_id
);
1233 if (ctf_type_kind_unsliced (foo
, bar
) != orig_kind
)
1235 ctf_err_warn (fp
, 1, 0, "added wrong kind to output mapping "
1236 "for hash %s named %s: %p/%lx from %s is "
1237 "kind %i, but newly-added %p/%lx from %s is "
1239 decorated_name
? decorated_name
: "(unnamed)",
1241 ctf_link_input_name (foo
),
1242 ctf_type_kind_unsliced (foo
, bar
),
1243 (void *) input
, type
,
1244 ctf_link_input_name (input
), orig_kind
);
1245 if (!ctf_assert (fp
, ctf_type_kind_unsliced (foo
, bar
)
1250 if (err
!= ECTF_NEXT_END
)
1251 return ctf_set_errno (fp
, err
);
1255 /* This function will be repeatedly called for the same types many times:
1256 don't waste time reinserting the same keys in that case. */
1257 if (!ctf_dynset_exists (type_ids
, id
, NULL
)
1258 && ctf_dynset_insert (type_ids
, id
) < 0)
1259 return ctf_set_errno (fp
, errno
);
1261 /* The rest only needs to happen for types with names. */
1262 if (!decorated_name
)
1265 /* Count the number of occurrences of the hash value for this GID. */
1267 hval
= ctf_dynhash_lookup (d
->cd_type_hashes
, id
);
1269 /* Mapping from name -> hash(hashval, count) not already present? */
1270 if ((name_counts
= ctf_dynhash_lookup (d
->cd_name_counts
,
1271 decorated_name
)) == NULL
)
1273 if ((name_counts
= ctf_dynhash_create (ctf_hash_string
,
1275 NULL
, NULL
)) == NULL
)
1276 return ctf_set_errno (fp
, errno
);
1277 if (ctf_dynhash_cinsert (d
->cd_name_counts
, decorated_name
,
1280 ctf_dynhash_destroy (name_counts
);
1281 return ctf_set_errno (fp
, errno
);
1285 /* This will, conveniently, return NULL (i.e. 0) for a new entry. */
1286 count
= (long int) (uintptr_t) ctf_dynhash_lookup (name_counts
, hval
);
1288 if (ctf_dynhash_cinsert (name_counts
, hval
,
1289 (const void *) (uintptr_t) (count
+ 1)) < 0)
1290 return ctf_set_errno (fp
, errno
);
1295 /* Mark a single hash as corresponding to a conflicting type. Mark all types
1296 that cite it as conflicting as well, terminating the recursive walk only when
1297 types that are already conflicted or types do not cite other types are seen.
1298 (Tagged structures and unions do not appear in the cd_citers graph, so the
1299 walk also terminates there, since any reference to a conflicting structure is
1300 just going to reference an unconflicting forward instead: see
1301 ctf_dedup_maybe_synthesize_forward.) */
1304 ctf_dedup_mark_conflicting_hash (ctf_dict_t
*fp
, const char *hval
)
1306 ctf_dedup_t
*d
= &fp
->ctf_dedup
;
1307 ctf_next_t
*i
= NULL
;
1310 ctf_dynset_t
*citers
;
1312 /* Mark conflicted if not already so marked. */
1313 if (ctf_dynset_exists (d
->cd_conflicting_types
, hval
, NULL
))
1316 ctf_dprintf ("Marking %s as conflicted\n", hval
);
1318 if (ctf_dynset_cinsert (d
->cd_conflicting_types
, hval
) < 0)
1320 ctf_dprintf ("Out of memory marking %s as conflicted\n", hval
);
1321 ctf_set_errno (fp
, errno
);
1325 /* If any types cite this type, mark them conflicted too. */
1326 if ((citers
= ctf_dynhash_lookup (d
->cd_citers
, hval
)) == NULL
)
1329 while ((err
= ctf_dynset_cnext (citers
, &i
, &k
)) == 0)
1331 const char *hv
= (const char *) k
;
1333 if (ctf_dynset_exists (d
->cd_conflicting_types
, hv
, NULL
))
1336 if (ctf_dedup_mark_conflicting_hash (fp
, hv
) < 0)
1338 ctf_next_destroy (i
);
1339 return -1; /* errno is set for us. */
1342 if (err
!= ECTF_NEXT_END
)
1343 return ctf_set_errno (fp
, err
);
1348 /* Look up a type kind from the output mapping, given a type hash value. */
1350 ctf_dedup_hash_kind (ctf_dict_t
*fp
, ctf_dict_t
**inputs
, const char *hash
)
1352 ctf_dedup_t
*d
= &fp
->ctf_dedup
;
1354 ctf_dynset_t
*type_ids
;
1356 /* Precondition: the output mapping is populated. */
1357 if (!ctf_assert (fp
, ctf_dynhash_elements (d
->cd_output_mapping
) > 0))
1360 /* Look up some GID from the output hash for this type. (They are all
1361 identical, so we can pick any). Don't assert if someone calls this
1362 function wrongly, but do assert if the output mapping knows about the hash,
1363 but has nothing associated with it. */
1365 type_ids
= ctf_dynhash_lookup (d
->cd_output_mapping
, hash
);
1368 ctf_dprintf ("Looked up type kind by nonexistent hash %s.\n", hash
);
1369 return ctf_set_errno (fp
, ECTF_INTERNAL
);
1371 id
= ctf_dynset_lookup_any (type_ids
);
1372 if (!ctf_assert (fp
, id
))
1375 return ctf_type_kind_unsliced (inputs
[CTF_DEDUP_GID_TO_INPUT (id
)],
1376 CTF_DEDUP_GID_TO_TYPE (id
));
1379 /* Used to keep a count of types: i.e. distinct type hash values. */
1380 typedef struct ctf_dedup_type_counter
1383 ctf_dict_t
**inputs
;
1384 int num_non_forwards
;
1385 } ctf_dedup_type_counter_t
;
1387 /* Add to the type counter for one name entry from the cd_name_counts. */
1389 ctf_dedup_count_types (void *key_
, void *value _libctf_unused_
, void *arg_
)
1391 const char *hval
= (const char *) key_
;
1393 ctf_dedup_type_counter_t
*arg
= (ctf_dedup_type_counter_t
*) arg_
;
1395 kind
= ctf_dedup_hash_kind (arg
->fp
, arg
->inputs
, hval
);
1397 /* We rely on ctf_dedup_hash_kind setting the fp to -ECTF_INTERNAL on error to
1398 smuggle errors out of here. */
1400 if (kind
!= CTF_K_FORWARD
)
1402 arg
->num_non_forwards
++;
1403 ctf_dprintf ("Counting hash %s: kind %i: num_non_forwards is %i\n",
1404 hval
, kind
, arg
->num_non_forwards
);
1407 /* We only need to know if there is more than one non-forward (an ambiguous
1408 type): don't waste time iterating any more than needed to figure that
1411 if (arg
->num_non_forwards
> 1)
1417 /* Detect name ambiguity and mark ambiguous names as conflicting, other than the
1420 ctf_dedup_detect_name_ambiguity (ctf_dict_t
*fp
, ctf_dict_t
**inputs
)
1422 ctf_dedup_t
*d
= &fp
->ctf_dedup
;
1423 ctf_next_t
*i
= NULL
;
1427 const char *whaterr
;
1429 /* Go through cd_name_counts for all CTF namespaces in turn. */
1431 while ((err
= ctf_dynhash_next (d
->cd_name_counts
, &i
, &k
, &v
)) == 0)
1433 const char *decorated
= (const char *) k
;
1434 ctf_dynhash_t
*name_counts
= (ctf_dynhash_t
*) v
;
1435 ctf_next_t
*j
= NULL
;
1437 /* If this is a forwardable kind or a forward (which we can tell without
1438 consulting the type because its decorated name has a space as its
1439 second character: see ctf_decorate_type_name), we are only interested
1440 in whether this name has many hashes associated with it: any such name
1441 is necessarily ambiguous, and types with that name are conflicting.
1442 Once we know whether this is true, we can skip to the next name: so use
1443 ctf_dynhash_iter_find for efficiency. */
1445 if (decorated
[0] != '\0' && decorated
[1] == ' ')
1447 ctf_dedup_type_counter_t counters
= { fp
, inputs
, 0 };
1448 ctf_dynhash_t
*counts
= (ctf_dynhash_t
*) v
;
1450 ctf_dynhash_iter_find (counts
, ctf_dedup_count_types
, &counters
);
1452 /* Check for assertion failure and pass it up. */
1453 if (ctf_errno (fp
) == ECTF_INTERNAL
)
1456 if (counters
.num_non_forwards
> 1)
1460 while ((err
= ctf_dynhash_cnext (counts
, &j
, &hval_
, NULL
)) == 0)
1462 const char *hval
= (const char *) hval_
;
1463 ctf_dynset_t
*type_ids
;
1467 /* Dig through the types in this hash to find the non-forwards
1468 and mark them ambiguous. */
1470 type_ids
= ctf_dynhash_lookup (d
->cd_output_mapping
, hval
);
1472 /* Nonexistent? Must be a forward with no referent. */
1476 id
= ctf_dynset_lookup_any (type_ids
);
1478 kind
= ctf_type_kind (inputs
[CTF_DEDUP_GID_TO_INPUT (id
)],
1479 CTF_DEDUP_GID_TO_TYPE (id
));
1481 if (kind
!= CTF_K_FORWARD
)
1483 ctf_dprintf ("Marking %p, with hash %s, conflicting: one "
1484 "of many non-forward GIDs for %s\n", id
,
1486 ctf_dedup_mark_conflicting_hash (fp
, hval
);
1489 if (err
!= ECTF_NEXT_END
)
1491 whaterr
= N_("error marking conflicting structs/unions");
1498 /* This is an ordinary type. Find the most common type with this
1499 name, and mark it unconflicting: all others are conflicting. (We
1500 cannot do this sort of popularity contest with forwardable types
1501 because any forwards to that type would be immediately unified with
1502 the most-popular type on insertion, and we want conflicting structs
1503 et al to have all forwards left intact, so the user is notified
1504 that this type is conflicting. TODO: improve this in future by
1505 setting such forwards non-root-visible.) */
1510 long max_hcount
= -1;
1511 const char *max_hval
= NULL
;
1513 if (ctf_dynhash_elements (name_counts
) <= 1)
1516 /* First find the most common. */
1517 while ((err
= ctf_dynhash_cnext (name_counts
, &j
, &key
, &count
)) == 0)
1519 hval
= (const char *) key
;
1520 if ((long int) (uintptr_t) count
> max_hcount
)
1522 max_hcount
= (long int) (uintptr_t) count
;
1526 if (err
!= ECTF_NEXT_END
)
1528 whaterr
= N_("error finding commonest conflicting type");
1532 /* Mark all the others as conflicting. */
1533 while ((err
= ctf_dynhash_cnext (name_counts
, &j
, &key
, NULL
)) == 0)
1535 hval
= (const char *) key
;
1536 if (strcmp (max_hval
, hval
) == 0)
1539 ctf_dprintf ("Marking %s, an uncommon hash for %s, conflicting\n",
1540 hval
, (const char *) k
);
1541 if (ctf_dedup_mark_conflicting_hash (fp
, hval
) < 0)
1543 whaterr
= N_("error marking hashes as conflicting");
1547 if (err
!= ECTF_NEXT_END
)
1549 whaterr
= N_("marking uncommon conflicting types");
1554 if (err
!= ECTF_NEXT_END
)
1556 whaterr
= N_("scanning for ambiguous names");
1563 ctf_next_destroy (i
);
1564 ctf_err_warn (fp
, 0, 0, "%s", gettext (whaterr
));
1565 return -1; /* errno is set for us. */
1568 ctf_err_warn (fp
, 0, err
, _("iteration failed: %s"), gettext (whaterr
));
1569 return ctf_set_errno (fp
, err
);
1572 ctf_next_destroy (i
);
1573 return -1; /* errno is set for us. */
1576 /* Initialize the deduplication machinery. */
1579 ctf_dedup_init (ctf_dict_t
*fp
)
1581 ctf_dedup_t
*d
= &fp
->ctf_dedup
;
1584 if (ctf_dedup_atoms_init (fp
) < 0)
1587 #if IDS_NEED_ALLOCATION
1588 if ((d
->cd_id_to_dict_t
= ctf_dynhash_create (ctf_hash_type_id_key
,
1589 ctf_hash_eq_type_id_key
,
1590 free
, NULL
)) == NULL
)
1594 for (i
= 0; i
< 4; i
++)
1596 if ((d
->cd_decorated_names
[i
] = ctf_dynhash_create (ctf_hash_string
,
1598 NULL
, NULL
)) == NULL
)
1602 if ((d
->cd_name_counts
1603 = ctf_dynhash_create (ctf_hash_string
,
1604 ctf_hash_eq_string
, NULL
,
1605 (ctf_hash_free_fun
) ctf_dynhash_destroy
)) == NULL
)
1608 if ((d
->cd_type_hashes
1609 = ctf_dynhash_create (ctf_hash_integer
,
1610 ctf_hash_eq_integer
,
1611 NULL
, NULL
)) == NULL
)
1614 if ((d
->cd_struct_origin
1615 = ctf_dynhash_create (ctf_hash_string
,
1617 NULL
, NULL
)) == NULL
)
1621 = ctf_dynhash_create (ctf_hash_string
,
1622 ctf_hash_eq_string
, NULL
,
1623 (ctf_hash_free_fun
) ctf_dynset_destroy
)) == NULL
)
1626 if ((d
->cd_output_mapping
1627 = ctf_dynhash_create (ctf_hash_string
,
1628 ctf_hash_eq_string
, NULL
,
1629 (ctf_hash_free_fun
) ctf_dynset_destroy
)) == NULL
)
1632 if ((d
->cd_output_first_gid
1633 = ctf_dynhash_create (ctf_hash_string
,
1635 NULL
, NULL
)) == NULL
)
1638 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
1639 if ((d
->cd_output_mapping_guard
1640 = ctf_dynhash_create (ctf_hash_integer
,
1641 ctf_hash_eq_integer
, NULL
, NULL
)) == NULL
)
1645 if ((d
->cd_input_nums
1646 = ctf_dynhash_create (ctf_hash_integer
,
1647 ctf_hash_eq_integer
,
1648 NULL
, NULL
)) == NULL
)
1651 if ((d
->cd_emission_struct_members
1652 = ctf_dynhash_create (ctf_hash_integer
,
1653 ctf_hash_eq_integer
,
1654 NULL
, NULL
)) == NULL
)
1657 if ((d
->cd_conflicting_types
1658 = ctf_dynset_create (htab_hash_string
,
1659 ctf_dynset_eq_string
, NULL
)) == NULL
)
1665 ctf_err_warn (fp
, 0, ENOMEM
, _("ctf_dedup_init: cannot initialize: "
1667 return ctf_set_errno (fp
, ENOMEM
);
1670 /* No ctf_dedup calls are allowed after this call other than starting a new
1671 deduplication via ctf_dedup (not even ctf_dedup_type_mapping lookups). */
1673 ctf_dedup_fini (ctf_dict_t
*fp
, ctf_dict_t
**outputs
, uint32_t noutputs
)
1675 ctf_dedup_t
*d
= &fp
->ctf_dedup
;
1678 /* ctf_dedup_atoms is kept across links. */
1679 #if IDS_NEED_ALLOCATION
1680 ctf_dynhash_destroy (d
->cd_id_to_dict_t
);
1682 for (i
= 0; i
< 4; i
++)
1683 ctf_dynhash_destroy (d
->cd_decorated_names
[i
]);
1684 ctf_dynhash_destroy (d
->cd_name_counts
);
1685 ctf_dynhash_destroy (d
->cd_type_hashes
);
1686 ctf_dynhash_destroy (d
->cd_struct_origin
);
1687 ctf_dynhash_destroy (d
->cd_citers
);
1688 ctf_dynhash_destroy (d
->cd_output_mapping
);
1689 ctf_dynhash_destroy (d
->cd_output_first_gid
);
1690 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
1691 ctf_dynhash_destroy (d
->cd_output_mapping_guard
);
1693 ctf_dynhash_destroy (d
->cd_input_nums
);
1694 ctf_dynhash_destroy (d
->cd_emission_struct_members
);
1695 ctf_dynset_destroy (d
->cd_conflicting_types
);
1697 /* Free the per-output state. */
1700 for (i
= 0; i
< noutputs
; i
++)
1702 ctf_dedup_t
*od
= &outputs
[i
]->ctf_dedup
;
1703 ctf_dynhash_destroy (od
->cd_output_emission_hashes
);
1704 ctf_dynhash_destroy (od
->cd_output_emission_conflicted_forwards
);
1705 ctf_dict_close (od
->cd_output
);
1708 memset (d
, 0, sizeof (ctf_dedup_t
));
1711 /* Return 1 if this type is cited by multiple input dictionaries. */
1714 ctf_dedup_multiple_input_dicts (ctf_dict_t
*output
, ctf_dict_t
**inputs
,
1717 ctf_dedup_t
*d
= &output
->ctf_dedup
;
1718 ctf_dynset_t
*type_ids
;
1719 ctf_next_t
*i
= NULL
;
1721 ctf_dict_t
*found
= NULL
, *relative_found
= NULL
;
1722 const char *type_id
;
1723 ctf_dict_t
*input_fp
;
1726 const char *decorated
;
1731 type_ids
= ctf_dynhash_lookup (d
->cd_output_mapping
, hval
);
1732 if (!ctf_assert (output
, type_ids
))
1735 /* Scan across the IDs until we find proof that two disjoint dictionaries
1736 are referenced. Exit as soon as possible. Optimization opportunity, but
1737 possibly not worth it, given that this is only executed in
1738 CTF_LINK_SHARE_DUPLICATED mode. */
1740 while ((err
= ctf_dynset_next (type_ids
, &i
, &id
)) == 0)
1742 ctf_dict_t
*fp
= inputs
[CTF_DEDUP_GID_TO_INPUT (id
)];
1744 if (fp
== found
|| fp
== relative_found
)
1754 && (fp
->ctf_parent
== found
|| found
->ctf_parent
== fp
))
1756 relative_found
= fp
;
1761 ctf_next_destroy (i
);
1764 if ((err
!= ECTF_NEXT_END
) && (err
!= 0))
1766 ctf_err_warn (output
, 0, err
, _("iteration error "
1767 "propagating conflictedness"));
1768 return ctf_set_errno (output
, err
);
1774 /* This type itself does not appear in multiple input dicts: how about another
1775 related type with the same name (e.g. a forward if this is a struct,
1778 type_id
= ctf_dynset_lookup_any (type_ids
);
1779 if (!ctf_assert (output
, type_id
))
1782 input_fp
= inputs
[CTF_DEDUP_GID_TO_INPUT (type_id
)];
1783 input_id
= CTF_DEDUP_GID_TO_TYPE (type_id
);
1784 fwdkind
= ctf_type_kind_forwarded (input_fp
, input_id
);
1785 name
= ctf_type_name_raw (input_fp
, input_id
);
1787 if ((fwdkind
== CTF_K_STRUCT
|| fwdkind
== CTF_K_UNION
)
1792 if ((decorated
= ctf_decorate_type_name (output
, name
,
1794 return -1; /* errno is set for us. */
1796 origin
= ctf_dynhash_lookup (d
->cd_struct_origin
, decorated
);
1797 if ((origin
!= NULL
) && (CTF_DEDUP_GID_TO_INPUT (origin
) < 0))
1804 /* Demote unconflicting types which reference only one input, or which reference
1805 two inputs where one input is the parent of the other, into conflicting
1806 types. Only used if the link mode is CTF_LINK_SHARE_DUPLICATED. */
1809 ctf_dedup_conflictify_unshared (ctf_dict_t
*output
, ctf_dict_t
**inputs
)
1811 ctf_dedup_t
*d
= &output
->ctf_dedup
;
1812 ctf_next_t
*i
= NULL
;
1815 ctf_dynset_t
*to_mark
= NULL
;
1817 if ((to_mark
= ctf_dynset_create (htab_hash_string
, ctf_dynset_eq_string
,
1821 while ((err
= ctf_dynhash_cnext (d
->cd_output_mapping
, &i
, &k
, NULL
)) == 0)
1823 const char *hval
= (const char *) k
;
1826 /* Types referenced by only one dict, with no type appearing under that
1827 name elsewhere, are marked conflicting. */
1829 conflicting
= !ctf_dedup_multiple_input_dicts (output
, inputs
, hval
);
1831 if (conflicting
< 0)
1832 goto err
; /* errno is set for us. */
1835 if (ctf_dynset_cinsert (to_mark
, hval
) < 0)
1838 if (err
!= ECTF_NEXT_END
)
1841 while ((err
= ctf_dynset_cnext (to_mark
, &i
, &k
)) == 0)
1843 const char *hval
= (const char *) k
;
1845 if (ctf_dedup_mark_conflicting_hash (output
, hval
) < 0)
1848 if (err
!= ECTF_NEXT_END
)
1851 ctf_dynset_destroy (to_mark
);
1856 ctf_set_errno (output
, errno
);
1858 err
= ctf_errno (output
);
1859 ctf_next_destroy (i
);
1861 ctf_dynset_destroy (to_mark
);
1862 ctf_err_warn (output
, 0, err
, _("conflictifying unshared types"));
1863 return ctf_set_errno (output
, err
);
1866 /* The core deduplicator. Populate cd_output_mapping in the output ctf_dedup
1867 with a mapping of all types that belong in this dictionary and where they
1868 come from, and cd_conflicting_types with an indication of whether each type
1869 is conflicted or not. OUTPUT is the top-level output: INPUTS is the array of
1870 input dicts; NINPUTS is the size of that array; PARENTS is an NINPUTS-element
1871 array with each element corresponding to a input which is a child dict set to
1872 the number in the INPUTS array of that input's parent.
1874 If CU_MAPPED is set, this is a first pass for a link with a non-empty CU
1875 mapping: only one output will result.
1877 Only deduplicates: does not emit the types into the output. Call
1878 ctf_dedup_emit afterwards to do that. */
1881 ctf_dedup (ctf_dict_t
*output
, ctf_dict_t
**inputs
, uint32_t ninputs
,
1882 uint32_t *parents
, int cu_mapped
)
1884 ctf_dedup_t
*d
= &output
->ctf_dedup
;
1886 ctf_next_t
*it
= NULL
;
1888 if (ctf_dedup_init (output
) < 0)
1889 return -1; /* errno is set for us. */
1891 for (i
= 0; i
< ninputs
; i
++)
1893 ctf_dprintf ("Input %i: %s\n", (int) i
, ctf_link_input_name (inputs
[i
]));
1894 if (ctf_dynhash_insert (d
->cd_input_nums
, inputs
[i
],
1895 (void *) (uintptr_t) i
) < 0)
1897 ctf_set_errno (output
, errno
);
1898 ctf_err_warn (output
, 0, errno
, _("ctf_dedup: cannot initialize: %s\n"),
1899 ctf_errmsg (errno
));
1904 /* Some flags do not apply when CU-mapping: this is not a duplicated link,
1905 because there is only one output and we really don't want to end up marking
1906 all nonconflicting but appears-only-once types as conflicting (which in the
1907 CU-mapped link means we'd mark them all as non-root-visible!). */
1908 d
->cd_link_flags
= output
->ctf_link_flags
;
1910 d
->cd_link_flags
&= ~(CTF_LINK_SHARE_DUPLICATED
);
1912 /* Compute hash values for all types, recursively, treating child structures
1913 and unions equivalent to forwards, and hashing in the name of the referent
1914 of each such type into structures, unions, and non-opaque forwards.
1915 Populate a mapping from decorated name (including an indication of
1916 struct/union/enum namespace) to count of type hash values in
1917 cd_name_counts, a mapping from and a mapping from hash values to input type
1918 IDs in cd_output_mapping. */
1920 ctf_dprintf ("Computing type hashes\n");
1921 for (i
= 0; i
< ninputs
; i
++)
1925 while ((id
= ctf_type_next (inputs
[i
], &it
, NULL
, 1)) != CTF_ERR
)
1927 ctf_dedup_hash_type (output
, inputs
[i
], inputs
, parents
,
1928 i
, id
, 0, 0, ctf_dedup_populate_mappings
);
1930 if (ctf_errno (inputs
[i
]) != ECTF_NEXT_END
)
1932 ctf_set_errno (output
, ctf_errno (inputs
[i
]));
1933 ctf_err_warn (output
, 0, 0, _("iteration failure "
1934 "computing type hashes"));
1939 /* Go through the cd_name_counts name->hash->count mapping for all CTF
1940 namespaces: any name with many hashes associated with it at this stage is
1941 necessarily ambiguous. Mark all the hashes except the most common as
1942 conflicting in the output. */
1944 ctf_dprintf ("Detecting type name ambiguity\n");
1945 if (ctf_dedup_detect_name_ambiguity (output
, inputs
) < 0)
1946 return -1; /* errno is set for us. */
1948 /* If the link mode is CTF_LINK_SHARE_DUPLICATED, we change any unconflicting
1949 types whose output mapping references only one input dict into a
1950 conflicting type, so that they end up in the per-CU dictionaries. */
1952 if (d
->cd_link_flags
& CTF_LINK_SHARE_DUPLICATED
)
1954 ctf_dprintf ("Conflictifying unshared types\n");
1955 if (ctf_dedup_conflictify_unshared (output
, inputs
) < 0)
1956 return -1; /* errno is set for us. */
1961 ctf_dedup_fini (output
, NULL
, 0);
1966 ctf_dedup_rwalk_output_mapping (ctf_dict_t
*output
, ctf_dict_t
**inputs
,
1967 uint32_t ninputs
, uint32_t *parents
,
1968 ctf_dynset_t
*already_visited
,
1970 int (*visit_fun
) (const char *hval
,
1972 ctf_dict_t
**inputs
,
1975 int already_visited
,
1981 void *arg
, unsigned long depth
);
1983 /* Like ctf_dedup_rwalk_output_mapping (which see), only takes a single target
1984 type and visits it. */
1986 ctf_dedup_rwalk_one_output_mapping (ctf_dict_t
*output
,
1987 ctf_dict_t
**inputs
, uint32_t ninputs
,
1989 ctf_dynset_t
*already_visited
,
1990 int visited
, void *type_id
,
1992 int (*visit_fun
) (const char *hval
,
1994 ctf_dict_t
**inputs
,
1997 int already_visited
,
2003 void *arg
, unsigned long depth
)
2005 ctf_dedup_t
*d
= &output
->ctf_dedup
;
2010 const char *whaterr
;
2012 input_num
= CTF_DEDUP_GID_TO_INPUT (type_id
);
2013 fp
= inputs
[input_num
];
2014 type
= CTF_DEDUP_GID_TO_TYPE (type_id
);
2016 ctf_dprintf ("%lu: Starting walk over type %s, %i/%lx (%p), from %s, "
2017 "kind %i\n", depth
, hval
, input_num
, type
, (void *) fp
,
2018 ctf_link_input_name (fp
), ctf_type_kind_unsliced (fp
, type
));
2020 /* Get the single call we do if this type has already been visited out of the
2023 return visit_fun (hval
, output
, inputs
, ninputs
, parents
, visited
, fp
,
2024 type
, type_id
, depth
, arg
);
2026 /* This macro is really ugly, but the alternative is repeating this code many
2027 times, which is worse. */
2029 #define CTF_TYPE_WALK(type, errlabel, errmsg) \
2032 const char *hashval; \
2033 int cited_type_input_num = input_num; \
2035 if ((fp->ctf_flags & LCTF_CHILD) && (LCTF_TYPE_ISPARENT (fp, type))) \
2036 cited_type_input_num = parents[input_num]; \
2038 type_id = CTF_DEDUP_GID (output, cited_type_input_num, type); \
2042 ctf_dprintf ("Walking: unimplemented type\n"); \
2046 ctf_dprintf ("Looking up ID %i/%lx in type hashes\n", \
2047 cited_type_input_num, type); \
2048 hashval = ctf_dynhash_lookup (d->cd_type_hashes, type_id); \
2049 if (!ctf_assert (output, hashval)) \
2051 whaterr = N_("error looking up ID in type hashes"); \
2054 ctf_dprintf ("ID %i/%lx has hash %s\n", cited_type_input_num, type, \
2057 ret = ctf_dedup_rwalk_output_mapping (output, inputs, ninputs, parents, \
2058 already_visited, hashval, \
2059 visit_fun, arg, depth); \
2067 switch (ctf_type_kind_unsliced (fp
, type
))
2070 /* Just skip things of unknown kind. */
2076 /* No types referenced. */
2080 case CTF_K_VOLATILE
:
2082 case CTF_K_RESTRICT
:
2085 CTF_TYPE_WALK (ctf_type_reference (fp
, type
), err
,
2086 N_("error during referenced type walk"));
2093 if (ctf_array_info (fp
, type
, &ar
) < 0)
2095 whaterr
= N_("error during array info lookup");
2099 CTF_TYPE_WALK (ar
.ctr_contents
, err
,
2100 N_("error during array contents type walk"));
2101 CTF_TYPE_WALK (ar
.ctr_index
, err
,
2102 N_("error during array index type walk"));
2106 case CTF_K_FUNCTION
:
2112 if (ctf_func_type_info (fp
, type
, &fi
) < 0)
2114 whaterr
= N_("error during func type info lookup");
2118 CTF_TYPE_WALK (fi
.ctc_return
, err
,
2119 N_("error during func return type walk"));
2121 if ((args
= calloc (fi
.ctc_argc
, sizeof (ctf_id_t
))) == NULL
)
2123 whaterr
= N_("error doing memory allocation");
2127 if (ctf_func_type_args (fp
, type
, fi
.ctc_argc
, args
) < 0)
2129 whaterr
= N_("error doing func arg type lookup");
2134 for (j
= 0; j
< fi
.ctc_argc
; j
++)
2135 CTF_TYPE_WALK (args
[j
], err_free_args
,
2136 N_("error during Func arg type walk"));
2146 /* We do not recursively traverse the members of structures: they are
2147 emitted later, in a separate pass. */
2150 whaterr
= N_("CTF dict corruption: unknown type kind");
2154 return visit_fun (hval
, output
, inputs
, ninputs
, parents
, visited
, fp
, type
,
2155 type_id
, depth
, arg
);
2158 ctf_set_errno (output
, ctf_errno (fp
));
2159 ctf_err_warn (output
, 0, 0, _("%s in input file %s at type ID %lx"),
2160 gettext (whaterr
), ctf_link_input_name (fp
), type
);
2164 /* Recursively traverse the output mapping, and do something with each type
2165 visited, from leaves to root. VISIT_FUN, called as recursion unwinds,
2166 returns a negative error code or zero. Type hashes may be visited more than
2167 once, but are not recursed through repeatedly: ALREADY_VISITED tracks whether
2168 types have already been visited. */
2170 ctf_dedup_rwalk_output_mapping (ctf_dict_t
*output
, ctf_dict_t
**inputs
,
2171 uint32_t ninputs
, uint32_t *parents
,
2172 ctf_dynset_t
*already_visited
,
2174 int (*visit_fun
) (const char *hval
,
2176 ctf_dict_t
**inputs
,
2179 int already_visited
,
2185 void *arg
, unsigned long depth
)
2187 ctf_dedup_t
*d
= &output
->ctf_dedup
;
2188 ctf_next_t
*i
= NULL
;
2191 ctf_dynset_t
*type_ids
;
2196 type_ids
= ctf_dynhash_lookup (d
->cd_output_mapping
, hval
);
2199 ctf_err_warn (output
, 0, ECTF_INTERNAL
,
2200 _("looked up type kind by nonexistent hash %s"), hval
);
2201 return ctf_set_errno (output
, ECTF_INTERNAL
);
2204 /* Have we seen this type before? */
2206 if (!ctf_dynset_exists (already_visited
, hval
, NULL
))
2208 /* Mark as already-visited immediately, to eliminate the possibility of
2209 cycles: but remember we have not actually visited it yet for the
2210 upcoming call to the visit_fun. (All our callers handle cycles
2211 properly themselves, so we can just abort them aggressively as soon as
2212 we find ourselves in one.) */
2215 if (ctf_dynset_cinsert (already_visited
, hval
) < 0)
2217 ctf_err_warn (output
, 0, ENOMEM
,
2218 _("out of memory tracking already-visited types"));
2219 return ctf_set_errno (output
, ENOMEM
);
2223 /* If this type is marked conflicted, traverse members and call
2224 ctf_dedup_rwalk_output_mapping_once on all the unique ones: otherwise, just
2225 pick a random one and use it. */
2227 if (!ctf_dynset_exists (d
->cd_conflicting_types
, hval
, NULL
))
2229 id
= ctf_dynset_lookup_any (type_ids
);
2230 if (!ctf_assert (output
, id
))
2233 return ctf_dedup_rwalk_one_output_mapping (output
, inputs
, ninputs
,
2234 parents
, already_visited
,
2235 visited
, id
, hval
, visit_fun
,
2239 while ((err
= ctf_dynset_next (type_ids
, &i
, &id
)) == 0)
2243 ret
= ctf_dedup_rwalk_one_output_mapping (output
, inputs
, ninputs
,
2244 parents
, already_visited
,
2246 visit_fun
, arg
, depth
);
2249 ctf_next_destroy (i
);
2250 return ret
; /* errno is set for us. */
2253 if (err
!= ECTF_NEXT_END
)
2255 ctf_err_warn (output
, 0, err
, _("cannot walk conflicted type"));
2256 return ctf_set_errno (output
, err
);
2262 typedef struct ctf_sort_om_cb_arg
2264 ctf_dict_t
**inputs
;
2267 } ctf_sort_om_cb_arg_t
;
2269 /* Sort the output mapping into order: types first appearing in earlier inputs
2270 first, parents preceding children: if types first appear in the same input,
2271 sort those with earlier ctf_id_t's first. */
2273 sort_output_mapping (const ctf_next_hkv_t
*one
, const ctf_next_hkv_t
*two
,
2276 ctf_sort_om_cb_arg_t
*arg
= (ctf_sort_om_cb_arg_t
*) arg_
;
2277 ctf_dedup_t
*d
= arg
->d
;
2278 const char *one_hval
= (const char *) one
->hkv_key
;
2279 const char *two_hval
= (const char *) two
->hkv_key
;
2280 void *one_gid
, *two_gid
;
2281 uint32_t one_ninput
;
2282 uint32_t two_ninput
;
2288 one_gid
= ctf_dynhash_lookup (d
->cd_output_first_gid
, one_hval
);
2289 two_gid
= ctf_dynhash_lookup (d
->cd_output_first_gid
, two_hval
);
2291 one_ninput
= CTF_DEDUP_GID_TO_INPUT (one_gid
);
2292 two_ninput
= CTF_DEDUP_GID_TO_INPUT (two_gid
);
2294 one_type
= CTF_DEDUP_GID_TO_TYPE (one_gid
);
2295 two_type
= CTF_DEDUP_GID_TO_TYPE (two_gid
);
2297 /* It's kind of hard to smuggle an assertion failure out of here. */
2298 assert (one_ninput
< arg
->ninputs
&& two_ninput
< arg
->ninputs
);
2300 one_fp
= arg
->inputs
[one_ninput
];
2301 two_fp
= arg
->inputs
[two_ninput
];
2303 /* Parents before children. */
2305 if (!(one_fp
->ctf_flags
& LCTF_CHILD
)
2306 && (two_fp
->ctf_flags
& LCTF_CHILD
))
2308 else if ((one_fp
->ctf_flags
& LCTF_CHILD
)
2309 && !(two_fp
->ctf_flags
& LCTF_CHILD
))
2312 /* ninput order, types appearing in earlier TUs first. */
2314 if (one_ninput
< two_ninput
)
2316 else if (two_ninput
< one_ninput
)
2319 /* Same TU. Earliest ctf_id_t first. They cannot be the same. */
2321 assert (one_type
!= two_type
);
2322 if (one_type
< two_type
)
2328 /* The public entry point to ctf_dedup_rwalk_output_mapping, above. */
2330 ctf_dedup_walk_output_mapping (ctf_dict_t
*output
, ctf_dict_t
**inputs
,
2331 uint32_t ninputs
, uint32_t *parents
,
2332 int (*visit_fun
) (const char *hval
,
2334 ctf_dict_t
**inputs
,
2337 int already_visited
,
2345 ctf_dynset_t
*already_visited
;
2346 ctf_next_t
*i
= NULL
;
2347 ctf_sort_om_cb_arg_t sort_arg
;
2351 if ((already_visited
= ctf_dynset_create (htab_hash_string
,
2352 ctf_dynset_eq_string
,
2354 return ctf_set_errno (output
, ENOMEM
);
2356 sort_arg
.inputs
= inputs
;
2357 sort_arg
.ninputs
= ninputs
;
2358 sort_arg
.d
= &output
->ctf_dedup
;
2360 while ((err
= ctf_dynhash_next_sorted (output
->ctf_dedup
.cd_output_mapping
,
2361 &i
, &k
, NULL
, sort_output_mapping
,
2364 const char *hval
= (const char *) k
;
2366 err
= ctf_dedup_rwalk_output_mapping (output
, inputs
, ninputs
, parents
,
2367 already_visited
, hval
, visit_fun
,
2371 ctf_next_destroy (i
);
2372 goto err
; /* errno is set for us. */
2375 if (err
!= ECTF_NEXT_END
)
2377 ctf_err_warn (output
, 0, err
, _("cannot recurse over output mapping"));
2378 ctf_set_errno (output
, err
);
2381 ctf_dynset_destroy (already_visited
);
2385 ctf_dynset_destroy (already_visited
);
2389 /* Possibly synthesise a synthetic forward in TARGET to subsitute for a
2390 conflicted per-TU type ID in INPUT with hash HVAL. Return its CTF ID, or 0
2391 if none was needed. */
2393 ctf_dedup_maybe_synthesize_forward (ctf_dict_t
*output
, ctf_dict_t
*target
,
2394 ctf_dict_t
*input
, ctf_id_t id
,
2397 ctf_dedup_t
*od
= &output
->ctf_dedup
;
2398 ctf_dedup_t
*td
= &target
->ctf_dedup
;
2401 const char *name
= ctf_type_name_raw (input
, id
);
2402 const char *decorated
;
2404 ctf_id_t emitted_forward
;
2406 if (!ctf_dynset_exists (od
->cd_conflicting_types
, hval
, NULL
)
2407 || target
->ctf_flags
& LCTF_CHILD
2409 || (((kind
= ctf_type_kind_unsliced (input
, id
)) != CTF_K_STRUCT
2410 && kind
!= CTF_K_UNION
&& kind
!= CTF_K_FORWARD
)))
2413 fwdkind
= ctf_type_kind_forwarded (input
, id
);
2415 ctf_dprintf ("Using synthetic forward for conflicted struct/union with "
2418 if (!ctf_assert (output
, name
))
2421 if ((decorated
= ctf_decorate_type_name (output
, name
, fwdkind
)) == NULL
)
2424 if (!ctf_dynhash_lookup_kv (td
->cd_output_emission_conflicted_forwards
,
2425 decorated
, NULL
, &v
))
2427 if ((emitted_forward
= ctf_add_forward (target
, CTF_ADD_ROOT
, name
,
2428 fwdkind
)) == CTF_ERR
)
2430 ctf_set_errno (output
, ctf_errno (target
));
2434 if (ctf_dynhash_cinsert (td
->cd_output_emission_conflicted_forwards
,
2435 decorated
, (void *) (uintptr_t)
2436 emitted_forward
) < 0)
2438 ctf_set_errno (output
, ENOMEM
);
2443 emitted_forward
= (ctf_id_t
) (uintptr_t) v
;
2445 ctf_dprintf ("Cross-TU conflicted struct: passing back forward, %lx\n",
2448 return emitted_forward
;
2451 /* Map a GID in some INPUT dict, in the form of an input number and a ctf_id_t,
2452 into a GID in a target output dict. If it returns 0, this is the
2453 unimplemented type, and the input type must have been 0. The OUTPUT dict is
2454 assumed to be the parent of the TARGET, if it is not the TARGET itself.
2456 Returns CTF_ERR on failure. Responds to an incoming CTF_ERR as an 'id' by
2457 returning CTF_ERR, to simplify callers. Errors are always propagated to the
2458 input, even if they relate to the target, for the same reason. (Target
2459 errors are expected to be very rare.)
2461 If the type in question is a citation of a conflicted type in a different TU,
2462 emit a forward of the right type in its place (if not already emitted), and
2463 record that forward in cd_output_emission_conflicted_forwards. This avoids
2464 the need to replicate the entire type graph below this point in the current
2465 TU (an appalling waste of space).
2467 TODO: maybe replace forwards in the same TU with their referents? Might
2468 make usability a bit better. */
2471 ctf_dedup_id_to_target (ctf_dict_t
*output
, ctf_dict_t
*target
,
2472 ctf_dict_t
**inputs
, uint32_t ninputs
,
2473 uint32_t *parents
, ctf_dict_t
*input
, int input_num
,
2476 ctf_dedup_t
*od
= &output
->ctf_dedup
;
2477 ctf_dedup_t
*td
= &target
->ctf_dedup
;
2478 ctf_dict_t
*err_fp
= input
;
2481 ctf_id_t emitted_forward
;
2483 /* The target type of an error is an error. */
2487 /* The unimplemented type's ID never changes. */
2490 ctf_dprintf ("%i/%lx: unimplemented type\n", input_num
, id
);
2494 ctf_dprintf ("Mapping %i/%lx to target %p (%s)\n", input_num
,
2495 id
, (void *) target
, ctf_link_input_name (target
));
2497 /* If the input type is in the parent type space, and this is a child, reset
2498 the input to the parent (which must already have been emitted, since
2499 emission of parent dicts happens before children). */
2500 if ((input
->ctf_flags
& LCTF_CHILD
) && (LCTF_TYPE_ISPARENT (input
, id
)))
2502 if (!ctf_assert (output
, parents
[input_num
] <= ninputs
))
2504 input
= inputs
[parents
[input_num
]];
2505 input_num
= parents
[input_num
];
2508 hval
= ctf_dynhash_lookup (od
->cd_type_hashes
,
2509 CTF_DEDUP_GID (output
, input_num
, id
));
2511 if (!ctf_assert (output
, hval
&& td
->cd_output_emission_hashes
))
2514 /* If this type is a conflicted tagged structure, union, or forward,
2515 substitute a synthetic forward instead, emitting it if need be. Only do
2516 this if the target is in the parent dict: if it's in the child dict, we can
2517 just point straight at the thing itself. Of course, we might be looking in
2518 the child dict right now and not find it and have to look in the parent, so
2519 we have to do this check twice. */
2521 emitted_forward
= ctf_dedup_maybe_synthesize_forward (output
, target
,
2523 switch (emitted_forward
)
2525 case 0: /* No forward needed. */
2528 ctf_set_errno (err_fp
, ctf_errno (output
));
2529 ctf_err_warn (err_fp
, 0, 0, _("cannot add synthetic forward for type "
2530 "%i/%lx"), input_num
, id
);
2533 return emitted_forward
;
2536 ctf_dprintf ("Looking up %i/%lx, hash %s, in target\n", input_num
, id
, hval
);
2538 target_id
= ctf_dynhash_lookup (td
->cd_output_emission_hashes
, hval
);
2541 /* Must be in the parent, so this must be a child, and they must not be
2543 ctf_dprintf ("Checking shared parent for target\n");
2544 if (!ctf_assert (output
, (target
!= output
)
2545 && (target
->ctf_flags
& LCTF_CHILD
)))
2548 target_id
= ctf_dynhash_lookup (od
->cd_output_emission_hashes
, hval
);
2550 emitted_forward
= ctf_dedup_maybe_synthesize_forward (output
, output
,
2552 switch (emitted_forward
)
2554 case 0: /* No forward needed. */
2557 ctf_err_warn (err_fp
, 0, ctf_errno (output
),
2558 _("cannot add synthetic forward for type %i/%lx"),
2560 return ctf_set_errno (err_fp
, ctf_errno (output
));
2562 return emitted_forward
;
2565 if (!ctf_assert (output
, target_id
))
2567 return (ctf_id_t
) (uintptr_t) target_id
;
2570 /* Emit a single deduplicated TYPE with the given HVAL, located in a given
2571 INPUT, with the given (G)ID, into the shared OUTPUT or a
2572 possibly-newly-created per-CU dict. All the types this type depends upon
2573 have already been emitted. (This type itself may also have been emitted.)
2575 If the ARG is 1, this is a CU-mapped deduplication round mapping many
2576 ctf_dict_t's into precisely one: conflicting types should be marked
2577 non-root-visible. If the ARG is 0, conflicting types go into per-CU
2578 dictionaries stored in the input's ctf_dedup.cd_output: otherwise, everything
2579 is emitted directly into the output. No struct/union members are emitted.
2581 Optimization opportunity: trace the ancestry of non-root-visible types and
2582 elide all that neither have a root-visible type somewhere towards their root,
2583 nor have the type visible via any other route (the function info section,
2584 data object section, backtrace section etc). */
2587 ctf_dedup_emit_type (const char *hval
, ctf_dict_t
*output
, ctf_dict_t
**inputs
,
2588 uint32_t ninputs
, uint32_t *parents
, int already_visited
,
2589 ctf_dict_t
*input
, ctf_id_t type
, void *id
, int depth
,
2592 ctf_dedup_t
*d
= &output
->ctf_dedup
;
2593 int kind
= ctf_type_kind_unsliced (input
, type
);
2595 ctf_dict_t
*target
= output
;
2596 ctf_dict_t
*real_input
;
2597 const ctf_type_t
*tp
;
2598 int input_num
= CTF_DEDUP_GID_TO_INPUT (id
);
2599 int output_num
= (uint32_t) -1; /* 'shared' */
2600 int cu_mapped
= *(int *)arg
;
2604 ctf_next_t
*i
= NULL
;
2607 ctf_id_t maybe_dup
= 0;
2609 const char *errtype
;
2610 int emission_hashed
= 0;
2612 /* We don't want to re-emit something we've already emitted. */
2614 if (already_visited
)
2617 ctf_dprintf ("%i: Emitting type with hash %s from %s: determining target\n",
2618 depth
, hval
, ctf_link_input_name (input
));
2620 /* Conflicting types go into a per-CU output dictionary, unless this is a
2621 CU-mapped run. The import is not refcounted, since it goes into the
2622 ctf_link_outputs dict of the output that is its parent. */
2623 is_conflicting
= ctf_dynset_exists (d
->cd_conflicting_types
, hval
, NULL
);
2625 if (is_conflicting
&& !cu_mapped
)
2627 ctf_dprintf ("%i: Type %s in %i/%lx is conflicted: "
2628 "inserting into per-CU target.\n",
2629 depth
, hval
, input_num
, type
);
2631 if (input
->ctf_dedup
.cd_output
)
2632 target
= input
->ctf_dedup
.cd_output
;
2637 if ((target
= ctf_create (&err
)) == NULL
)
2639 ctf_err_warn (output
, 0, err
,
2640 _("cannot create per-CU CTF archive for CU %s"),
2641 ctf_link_input_name (input
));
2642 return ctf_set_errno (output
, err
);
2645 ctf_import_unref (target
, output
);
2646 if (ctf_cuname (input
) != NULL
)
2647 ctf_cuname_set (target
, ctf_cuname (input
));
2649 ctf_cuname_set (target
, "unnamed-CU");
2650 ctf_parent_name_set (target
, _CTF_SECTION
);
2652 input
->ctf_dedup
.cd_output
= target
;
2654 output_num
= input_num
;
2658 if ((tp
= ctf_lookup_by_id (&real_input
, type
)) == NULL
)
2660 ctf_err_warn (output
, 0, ctf_errno (input
),
2661 _("%s: lookup failure for type %lx"),
2662 ctf_link_input_name (real_input
), type
);
2663 return ctf_set_errno (output
, ctf_errno (input
));
2666 name
= ctf_strraw (real_input
, tp
->ctt_name
);
2668 /* Hide conflicting types, if we were asked to: also hide if a type with this
2669 name already exists and is not a forward. */
2670 if (cu_mapped
&& is_conflicting
)
2673 && (maybe_dup
= ctf_lookup_by_rawname (target
, kind
, name
)) != 0)
2675 if (ctf_type_kind (target
, maybe_dup
) != CTF_K_FORWARD
)
2679 ctf_dprintf ("%i: Emitting type with hash %s (%s), into target %i/%p\n",
2680 depth
, hval
, name
? name
: "", input_num
, (void *) target
);
2682 if (!target
->ctf_dedup
.cd_output_emission_hashes
)
2683 if ((target
->ctf_dedup
.cd_output_emission_hashes
2684 = ctf_dynhash_create (ctf_hash_string
, ctf_hash_eq_string
,
2685 NULL
, NULL
)) == NULL
)
2688 if (!target
->ctf_dedup
.cd_output_emission_conflicted_forwards
)
2689 if ((target
->ctf_dedup
.cd_output_emission_conflicted_forwards
2690 = ctf_dynhash_create (ctf_hash_string
, ctf_hash_eq_string
,
2691 NULL
, NULL
)) == NULL
)
2697 /* These are types that CTF cannot encode, marked as such by the compile.
2698 We intentionally do not re-emit these. */
2702 /* This will do nothing if the type to which this forwards already exists,
2703 and will be replaced with such a type if it appears later. */
2705 errtype
= _("forward");
2706 if ((new_type
= ctf_add_forward (target
, isroot
, name
,
2707 ctf_type_kind_forwarded (input
, type
)))
2714 errtype
= _("float/int");
2715 if (ctf_type_encoding (input
, type
, &ep
) < 0)
2716 goto err_input
; /* errno is set for us. */
2717 if ((new_type
= ctf_add_encoded (target
, isroot
, name
, &ep
, kind
))
2725 errtype
= _("enum");
2726 if ((new_type
= ctf_add_enum (target
, isroot
, name
)) == CTF_ERR
)
2727 goto err_input
; /* errno is set for us. */
2729 while ((name
= ctf_enum_next (input
, type
, &i
, &val
)) != NULL
)
2731 if (ctf_add_enumerator (target
, new_type
, name
, val
) < 0)
2733 ctf_err_warn (target
, 0, ctf_errno (target
),
2734 _("%s (%i): cannot add enumeration value %s "
2735 "from input type %lx"),
2736 ctf_link_input_name (input
), input_num
, name
,
2738 ctf_next_destroy (i
);
2739 return ctf_set_errno (output
, ctf_errno (target
));
2742 if (ctf_errno (input
) != ECTF_NEXT_END
)
2748 errtype
= _("typedef");
2750 ref
= ctf_type_reference (input
, type
);
2751 if ((ref
= ctf_dedup_id_to_target (output
, target
, inputs
, ninputs
,
2752 parents
, input
, input_num
,
2754 goto err_input
; /* errno is set for us. */
2756 if ((new_type
= ctf_add_typedef (target
, isroot
, name
, ref
)) == CTF_ERR
)
2757 goto err_target
; /* errno is set for us. */
2760 case CTF_K_VOLATILE
:
2762 case CTF_K_RESTRICT
:
2764 errtype
= _("pointer or cvr-qual");
2766 ref
= ctf_type_reference (input
, type
);
2767 if ((ref
= ctf_dedup_id_to_target (output
, target
, inputs
, ninputs
,
2768 parents
, input
, input_num
,
2770 goto err_input
; /* errno is set for us. */
2772 if ((new_type
= ctf_add_reftype (target
, isroot
, ref
, kind
)) == CTF_ERR
)
2773 goto err_target
; /* errno is set for us. */
2777 errtype
= _("slice");
2779 if (ctf_type_encoding (input
, type
, &ep
) < 0)
2780 goto err_input
; /* errno is set for us. */
2782 ref
= ctf_type_reference (input
, type
);
2783 if ((ref
= ctf_dedup_id_to_target (output
, target
, inputs
, ninputs
,
2784 parents
, input
, input_num
,
2788 if ((new_type
= ctf_add_slice (target
, isroot
, ref
, &ep
)) == CTF_ERR
)
2796 errtype
= _("array info");
2797 if (ctf_array_info (input
, type
, &ar
) < 0)
2800 ar
.ctr_contents
= ctf_dedup_id_to_target (output
, target
, inputs
,
2801 ninputs
, parents
, input
,
2802 input_num
, ar
.ctr_contents
);
2803 ar
.ctr_index
= ctf_dedup_id_to_target (output
, target
, inputs
, ninputs
,
2804 parents
, input
, input_num
,
2807 if (ar
.ctr_contents
== CTF_ERR
|| ar
.ctr_index
== CTF_ERR
)
2810 if ((new_type
= ctf_add_array (target
, isroot
, &ar
)) == CTF_ERR
)
2816 case CTF_K_FUNCTION
:
2822 errtype
= _("function");
2823 if (ctf_func_type_info (input
, type
, &fi
) < 0)
2826 fi
.ctc_return
= ctf_dedup_id_to_target (output
, target
, inputs
, ninputs
,
2827 parents
, input
, input_num
,
2829 if (fi
.ctc_return
== CTF_ERR
)
2832 if ((args
= calloc (fi
.ctc_argc
, sizeof (ctf_id_t
))) == NULL
)
2834 ctf_set_errno (input
, ENOMEM
);
2838 errtype
= _("function args");
2839 if (ctf_func_type_args (input
, type
, fi
.ctc_argc
, args
) < 0)
2845 for (j
= 0; j
< fi
.ctc_argc
; j
++)
2847 args
[j
] = ctf_dedup_id_to_target (output
, target
, inputs
, ninputs
,
2848 parents
, input
, input_num
,
2850 if (args
[j
] == CTF_ERR
)
2854 if ((new_type
= ctf_add_function (target
, isroot
,
2855 &fi
, args
)) == CTF_ERR
)
2867 size_t size
= ctf_type_size (input
, type
);
2869 /* Insert the structure itself, so other types can refer to it. */
2871 errtype
= _("structure/union");
2872 if (kind
== CTF_K_STRUCT
)
2873 new_type
= ctf_add_struct_sized (target
, isroot
, name
, size
);
2875 new_type
= ctf_add_union_sized (target
, isroot
, name
, size
);
2877 if (new_type
== CTF_ERR
)
2880 out_id
= CTF_DEDUP_GID (output
, output_num
, new_type
);
2881 ctf_dprintf ("%i: Noting need to emit members of %p -> %p\n", depth
,
2883 /* Record the need to emit the members of this structure later. */
2884 if (ctf_dynhash_insert (d
->cd_emission_struct_members
, id
, out_id
) < 0)
2889 ctf_err_warn (output
, 0, ECTF_CORRUPT
, _("%s: unknown type kind for "
2891 ctf_link_input_name (input
), type
);
2892 return ctf_set_errno (output
, ECTF_CORRUPT
);
2895 if (!emission_hashed
2897 && ctf_dynhash_cinsert (target
->ctf_dedup
.cd_output_emission_hashes
,
2898 hval
, (void *) (uintptr_t) new_type
) < 0)
2900 ctf_err_warn (output
, 0, ENOMEM
, _("out of memory tracking deduplicated "
2901 "global type IDs"));
2902 return ctf_set_errno (output
, ENOMEM
);
2905 if (!emission_hashed
&& new_type
!= 0)
2906 ctf_dprintf ("%i: Inserted %s, %i/%lx -> %lx into emission hash for "
2907 "target %p (%s)\n", depth
, hval
, input_num
, type
, new_type
,
2908 (void *) target
, ctf_link_input_name (target
));
2913 ctf_err_warn (output
, 0, ENOMEM
, _("out of memory creating emission-tracking "
2915 return ctf_set_errno (output
, ENOMEM
);
2918 ctf_err_warn (output
, 0, ctf_errno (input
),
2919 _("%s (%i): while emitting deduplicated %s, error getting "
2920 "input type %lx"), ctf_link_input_name (input
),
2921 input_num
, errtype
, type
);
2922 return ctf_set_errno (output
, ctf_errno (input
));
2924 ctf_err_warn (output
, 0, ctf_errno (target
),
2925 _("%s (%i): while emitting deduplicated %s, error emitting "
2926 "target type from input type %lx"),
2927 ctf_link_input_name (input
), input_num
,
2929 return ctf_set_errno (output
, ctf_errno (target
));
2932 /* Traverse the cd_emission_struct_members and emit the members of all
2933 structures and unions. All other types are emitted and complete by this
2937 ctf_dedup_emit_struct_members (ctf_dict_t
*output
, ctf_dict_t
**inputs
,
2938 uint32_t ninputs
, uint32_t *parents
)
2940 ctf_dedup_t
*d
= &output
->ctf_dedup
;
2941 ctf_next_t
*i
= NULL
;
2942 void *input_id
, *target_id
;
2944 ctf_dict_t
*err_fp
, *input_fp
;
2948 while ((err
= ctf_dynhash_next (d
->cd_emission_struct_members
, &i
,
2949 &input_id
, &target_id
)) == 0)
2951 ctf_next_t
*j
= NULL
;
2953 uint32_t target_num
;
2954 ctf_id_t input_type
, target_type
;
2959 input_num
= CTF_DEDUP_GID_TO_INPUT (input_id
);
2960 input_fp
= inputs
[input_num
];
2961 input_type
= CTF_DEDUP_GID_TO_TYPE (input_id
);
2963 /* The output is either -1 (for the shared, parent output dict) or the
2964 number of the corresponding input. */
2965 target_num
= CTF_DEDUP_GID_TO_INPUT (target_id
);
2966 if (target_num
== (uint32_t) -1)
2970 target
= inputs
[target_num
]->ctf_dedup
.cd_output
;
2971 if (!ctf_assert (output
, target
))
2974 err_type
= input_type
;
2978 target_type
= CTF_DEDUP_GID_TO_TYPE (target_id
);
2980 while ((offset
= ctf_member_next (input_fp
, input_type
, &j
, &name
,
2981 &membtype
, 0)) >= 0)
2984 err_type
= target_type
;
2985 if ((membtype
= ctf_dedup_id_to_target (output
, target
, inputs
,
2986 ninputs
, parents
, input_fp
,
2988 membtype
)) == CTF_ERR
)
2990 ctf_next_destroy (j
);
2996 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
2997 ctf_dprintf ("Emitting %s, offset %zi\n", name
, offset
);
2999 if (ctf_add_member_offset (target
, target_type
, name
,
3000 membtype
, offset
) < 0)
3002 ctf_next_destroy (j
);
3006 if (ctf_errno (input_fp
) != ECTF_NEXT_END
)
3008 err
= ctf_errno (input_fp
);
3009 ctf_next_destroy (i
);
3013 if (err
!= ECTF_NEXT_END
)
3018 ctf_next_destroy (i
);
3019 ctf_err_warn (output
, 0, ctf_errno (err_fp
),
3020 _("%s (%i): error emitting members for structure type %lx"),
3021 ctf_link_input_name (input_fp
), input_num
, err_type
);
3022 return ctf_set_errno (output
, ctf_errno (err_fp
));
3024 ctf_err_warn (output
, 0, err
, _("iteration failure emitting "
3025 "structure members"));
3026 return ctf_set_errno (output
, err
);
3029 /* Emit deduplicated types into the outputs. The shared type repository is
3030 OUTPUT, on which the ctf_dedup function must have already been called. The
3031 PARENTS array contains the INPUTS index of the parent dict for every child
3032 dict at the corresponding index in the INPUTS (for non-child dicts, the value
3035 Return an array of fps with content emitted into them (starting with OUTPUT,
3036 which is the parent of all others, then all the newly-generated outputs).
3038 If CU_MAPPED is set, this is a first pass for a link with a non-empty CU
3039 mapping: only one output will result. */
3042 ctf_dedup_emit (ctf_dict_t
*output
, ctf_dict_t
**inputs
, uint32_t ninputs
,
3043 uint32_t *parents
, uint32_t *noutputs
, int cu_mapped
)
3045 size_t num_outputs
= 1; /* Always at least one output: us. */
3046 ctf_dict_t
**outputs
;
3050 ctf_dprintf ("Triggering emission.\n");
3051 if (ctf_dedup_walk_output_mapping (output
, inputs
, ninputs
, parents
,
3052 ctf_dedup_emit_type
, &cu_mapped
) < 0)
3053 return NULL
; /* errno is set for us. */
3055 ctf_dprintf ("Populating struct members.\n");
3056 if (ctf_dedup_emit_struct_members (output
, inputs
, ninputs
, parents
) < 0)
3057 return NULL
; /* errno is set for us. */
3059 for (i
= 0; i
< ninputs
; i
++)
3061 if (inputs
[i
]->ctf_dedup
.cd_output
)
3065 if (!ctf_assert (output
, !cu_mapped
|| (cu_mapped
&& num_outputs
== 1)))
3068 if ((outputs
= calloc (num_outputs
, sizeof (ctf_dict_t
*))) == NULL
)
3070 ctf_err_warn (output
, 0, ENOMEM
,
3071 _("out of memory allocating link outputs array"));
3072 ctf_set_errno (output
, ENOMEM
);
3075 *noutputs
= num_outputs
;
3079 output
->ctf_refcnt
++;
3082 for (i
= 0; i
< ninputs
; i
++)
3084 if (inputs
[i
]->ctf_dedup
.cd_output
)
3086 *walk
= inputs
[i
]->ctf_dedup
.cd_output
;
3087 inputs
[i
]->ctf_dedup
.cd_output
= NULL
;
3095 /* Determine what type SRC_FP / SRC_TYPE was emitted as in the FP, which
3096 must be the shared dict or have it as a parent: return 0 if none. The SRC_FP
3097 must be a past input to ctf_dedup. */
3100 ctf_dedup_type_mapping (ctf_dict_t
*fp
, ctf_dict_t
*src_fp
, ctf_id_t src_type
)
3102 ctf_dict_t
*output
= NULL
;
3110 /* It is an error (an internal error in the caller, in ctf-link.c) to call
3111 this with an FP that is not a per-CU output or shared output dict, or with
3112 a SRC_FP that was not passed to ctf_dedup as an input; it is an internal
3113 error in ctf-dedup for the type passed not to have been hashed, though if
3114 the src_fp is a child dict and the type is not a child type, it will have
3115 been hashed under the GID corresponding to the parent. */
3117 if (fp
->ctf_dedup
.cd_type_hashes
!= NULL
)
3119 else if (fp
->ctf_parent
&& fp
->ctf_parent
->ctf_dedup
.cd_type_hashes
!= NULL
)
3120 output
= fp
->ctf_parent
;
3123 ctf_set_errno (fp
, ECTF_INTERNAL
);
3124 ctf_err_warn (fp
, 0, ECTF_INTERNAL
,
3125 _("dict %p passed to ctf_dedup_type_mapping is not a "
3126 "deduplicated output"), (void *) fp
);
3130 if (src_fp
->ctf_parent
&& ctf_type_isparent (src_fp
, src_type
))
3131 src_fp
= src_fp
->ctf_parent
;
3133 d
= &output
->ctf_dedup
;
3135 found
= ctf_dynhash_lookup_kv (d
->cd_input_nums
, src_fp
, NULL
, &num_ptr
);
3136 if (!ctf_assert (output
, found
!= 0))
3137 return CTF_ERR
; /* errno is set for us. */
3138 input_num
= (uintptr_t) num_ptr
;
3140 hval
= ctf_dynhash_lookup (d
->cd_type_hashes
,
3141 CTF_DEDUP_GID (output
, input_num
, src_type
));
3143 if (!ctf_assert (output
, hval
!= NULL
))
3144 return CTF_ERR
; /* errno is set for us. */
3146 /* The emission hashes may be unset if this dict was created after
3147 deduplication to house variables or other things that would conflict if
3148 stored in the shared dict. */
3149 if (fp
->ctf_dedup
.cd_output_emission_hashes
)
3150 if (ctf_dynhash_lookup_kv (fp
->ctf_dedup
.cd_output_emission_hashes
, hval
,
3152 return (ctf_id_t
) (uintptr_t) type_ptr
;
3156 ctf_dict_t
*pfp
= fp
->ctf_parent
;
3157 if (pfp
->ctf_dedup
.cd_output_emission_hashes
)
3158 if (ctf_dynhash_lookup_kv (pfp
->ctf_dedup
.cd_output_emission_hashes
,
3159 hval
, NULL
, &type_ptr
))
3160 return (ctf_id_t
) (uintptr_t) type_ptr
;