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) \
595 /* If this is a named struct or union or a forward to one, and this is a child
596 traversal, treat this type as if it were a forward -- do not recurse to
597 children, ignore all content not already hashed in, and hash in the
598 decorated name of the type instead. */
600 if (ctf_dedup_is_stub (name
, kind
, tp
->ctt_type
, flags
))
602 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
603 ctf_dprintf ("Struct/union/forward citation: substituting forwarding "
604 "stub with decorated name %s\n", decorated
);
607 ctf_sha1_init (&hash
);
608 ctf_dedup_sha1_add (&hash
, decorated
, strlen (decorated
) + 1,
609 "decorated struct/union/forward name", depth
);
610 ctf_sha1_fini (&hash
, hashbuf
);
612 if ((hval
= intern (fp
, strdup (hashbuf
))) == NULL
)
614 ctf_err_warn (fp
, 0, 0, _("%s (%i): out of memory during forwarding-"
615 "stub hashing for type with GID %p"),
616 ctf_link_input_name (input
), input_num
, type_id
);
617 return NULL
; /* errno is set for us. */
620 /* In share-duplicated link mode, make sure the origin of this type is
621 recorded, even if this is a type in a parent dict which will not be
622 directly traversed. */
623 if (d
->cd_link_flags
& CTF_LINK_SHARE_DUPLICATED
624 && ctf_dedup_record_origin (fp
, input_num
, decorated
, type_id
) < 0)
625 return NULL
; /* errno is set for us. */
630 /* Now ensure that subsequent recursive calls (but *not* the top-level call)
631 get this treatment. */
632 flags
|= CTF_DEDUP_HASH_INTERNAL_CHILD
;
634 /* If this is a struct, union, or forward with a name, record the unique
635 originating input TU, if there is one. */
637 if (decorated
&& (ctf_forwardable_kind (kind
) || kind
!= CTF_K_FORWARD
))
638 if (d
->cd_link_flags
& CTF_LINK_SHARE_DUPLICATED
639 && ctf_dedup_record_origin (fp
, input_num
, decorated
, type_id
) < 0)
640 return NULL
; /* errno is set for us. */
642 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
643 ctf_dprintf ("%lu: hashing thing with ID %i/%lx (kind %i): %s.\n",
644 depth
, input_num
, type
, kind
, name
? name
: "");
647 /* Some type kinds don't have names: the API provides no way to set the name,
648 so the type the deduplicator outputs will be nameless even if the input
649 somehow has a name, and the name should not be mixed into the hash. */
663 /* Mix in invariant stuff, transforming the type kind if needed. Note that
664 the vlen is *not* hashed in: the actual variable-length info is hashed in
665 instead, piecewise. The vlen is not part of the type, only the
666 variable-length data is: identical types with distinct vlens are quite
667 possible. Equally, we do not want to hash in the isroot flag: both the
668 compiler and the deduplicator set the nonroot flag to indicate clashes with
669 *other types in the same TU* with the same name: so two types can easily
670 have distinct nonroot flags, yet be exactly the same type.*/
672 ctf_sha1_init (&hash
);
674 ctf_dedup_sha1_add (&hash
, name
, strlen (name
) + 1, "name", depth
);
675 ctf_dedup_sha1_add (&hash
, &kind
, sizeof (uint32_t), "kind", depth
);
677 /* Hash content of this type. */
681 /* No extra state. */
685 /* Add the forwarded kind, stored in the ctt_type. */
686 ctf_dedup_sha1_add (&hash
, &tp
->ctt_type
, sizeof (tp
->ctt_type
),
687 "forwarded kind", depth
);
693 memset (&ep
, 0, sizeof (ctf_encoding_t
));
695 ctf_dedup_sha1_add (&hash
, &tp
->ctt_size
, sizeof (uint32_t), "size",
697 if (ctf_type_encoding (input
, type
, &ep
) < 0)
699 whaterr
= N_("error getting encoding");
702 ctf_dedup_sha1_add (&hash
, &ep
, sizeof (ctf_encoding_t
), "encoding",
706 /* Types that reference other types. */
712 /* Hash the referenced type, if not already hashed, and mix it in. */
713 child_type
= ctf_type_reference (input
, type
);
714 if ((hval
= ctf_dedup_hash_type (fp
, input
, inputs
, parents
, input_num
,
715 child_type
, flags
, depth
,
716 populate_fun
)) == NULL
)
718 whaterr
= N_("error doing referenced type hashing");
721 ctf_dedup_sha1_add (&hash
, hval
, strlen (hval
) + 1, "referenced type",
727 /* The slices of two types hash identically only if the type they overlay
728 also has the same encoding. This is not ideal, but in practice will work
729 well enough. We work directly rather than using the CTF API because
730 we do not want the slice's normal automatically-shine-through
731 semantics to kick in here. */
734 const ctf_slice_t
*slice
;
735 const ctf_dtdef_t
*dtd
;
739 child_type
= ctf_type_reference (input
, type
);
740 ctf_get_ctt_size (input
, tp
, &size
, &increment
);
741 ctf_dedup_sha1_add (&hash
, &size
, sizeof (ssize_t
), "size", depth
);
743 if ((hval
= ctf_dedup_hash_type (fp
, input
, inputs
, parents
, input_num
,
744 child_type
, flags
, depth
,
745 populate_fun
)) == NULL
)
747 whaterr
= N_("error doing slice-referenced type hashing");
750 ctf_dedup_sha1_add (&hash
, hval
, strlen (hval
) + 1, "sliced type",
754 if ((dtd
= ctf_dynamic_type (input
, type
)) != NULL
)
755 slice
= (ctf_slice_t
*) dtd
->dtd_vlen
;
757 slice
= (ctf_slice_t
*) ((uintptr_t) tp
+ increment
);
759 ctf_dedup_sha1_add (&hash
, &slice
->cts_offset
,
760 sizeof (slice
->cts_offset
), "slice offset", depth
);
761 ctf_dedup_sha1_add (&hash
, &slice
->cts_bits
,
762 sizeof (slice
->cts_bits
), "slice bits", depth
);
770 if (ctf_array_info (input
, type
, &ar
) < 0)
772 whaterr
= N_("error getting array info");
776 if ((hval
= ctf_dedup_hash_type (fp
, input
, inputs
, parents
, input_num
,
777 ar
.ctr_contents
, flags
, depth
,
778 populate_fun
)) == NULL
)
780 whaterr
= N_("error doing array contents type hashing");
783 ctf_dedup_sha1_add (&hash
, hval
, strlen (hval
) + 1, "array contents",
785 ADD_CITER (citers
, hval
);
787 if ((hval
= ctf_dedup_hash_type (fp
, input
, inputs
, parents
, input_num
,
788 ar
.ctr_index
, flags
, depth
,
789 populate_fun
)) == NULL
)
791 whaterr
= N_("error doing array index type hashing");
794 ctf_dedup_sha1_add (&hash
, hval
, strlen (hval
) + 1, "array index",
796 ctf_dedup_sha1_add (&hash
, &ar
.ctr_nelems
, sizeof (ar
.ctr_nelems
),
797 "element count", depth
);
798 ADD_CITER (citers
, hval
);
808 if (ctf_func_type_info (input
, type
, &fi
) < 0)
810 whaterr
= N_("error getting func type info");
814 if ((hval
= ctf_dedup_hash_type (fp
, input
, inputs
, parents
, input_num
,
815 fi
.ctc_return
, flags
, depth
,
816 populate_fun
)) == NULL
)
818 whaterr
= N_("error getting func return type");
821 ctf_dedup_sha1_add (&hash
, hval
, strlen (hval
) + 1, "func return",
823 ctf_dedup_sha1_add (&hash
, &fi
.ctc_argc
, sizeof (fi
.ctc_argc
),
825 ctf_dedup_sha1_add (&hash
, &fi
.ctc_flags
, sizeof (fi
.ctc_flags
),
826 "func flags", depth
);
827 ADD_CITER (citers
, hval
);
829 if ((args
= calloc (fi
.ctc_argc
, sizeof (ctf_id_t
))) == NULL
)
832 whaterr
= N_("error doing memory allocation");
836 if (ctf_func_type_args (input
, type
, fi
.ctc_argc
, args
) < 0)
839 whaterr
= N_("error getting func arg type");
842 for (j
= 0; j
< fi
.ctc_argc
; j
++)
844 if ((hval
= ctf_dedup_hash_type (fp
, input
, inputs
, parents
,
845 input_num
, args
[j
], flags
, depth
,
846 populate_fun
)) == NULL
)
849 whaterr
= N_("error doing func arg type hashing");
852 ctf_dedup_sha1_add (&hash
, hval
, strlen (hval
) + 1, "func arg type",
854 ADD_CITER (citers
, hval
);
864 ctf_dedup_sha1_add (&hash
, &tp
->ctt_size
, sizeof (uint32_t),
866 while ((ename
= ctf_enum_next (input
, type
, &i
, &val
)) != NULL
)
868 ctf_dedup_sha1_add (&hash
, ename
, strlen (ename
) + 1, "enumerator",
870 ctf_dedup_sha1_add (&hash
, &val
, sizeof (val
), "enumerand", depth
);
872 if (ctf_errno (input
) != ECTF_NEXT_END
)
874 whaterr
= N_("error doing enum member iteration");
879 /* Top-level only. */
888 ctf_get_ctt_size (input
, tp
, &size
, NULL
);
889 ctf_dedup_sha1_add (&hash
, &size
, sizeof (ssize_t
), "struct size",
892 while ((offset
= ctf_member_next (input
, type
, &i
, &mname
, &membtype
,
897 ctf_dedup_sha1_add (&hash
, mname
, strlen (mname
) + 1,
898 "member name", depth
);
900 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
901 ctf_dprintf ("%lu: Traversing to member %s\n", depth
, mname
);
903 if ((hval
= ctf_dedup_hash_type (fp
, input
, inputs
, parents
,
904 input_num
, membtype
, flags
, depth
,
905 populate_fun
)) == NULL
)
907 whaterr
= N_("error doing struct/union member type hashing");
911 ctf_dedup_sha1_add (&hash
, hval
, strlen (hval
) + 1, "member hash",
913 ctf_dedup_sha1_add (&hash
, &offset
, sizeof (offset
), "member offset",
915 ADD_CITER (citers
, hval
);
917 if (ctf_errno (input
) != ECTF_NEXT_END
)
919 whaterr
= N_("error doing struct/union member iteration");
925 whaterr
= N_("error: unknown type kind");
928 ctf_sha1_fini (&hash
, hashbuf
);
930 if ((hval
= intern (fp
, strdup (hashbuf
))) == NULL
)
932 whaterr
= N_("cannot intern hash");
936 /* Populate the citers for this type's subtypes, now the hash for the type
938 whaterr
= N_("error tracking citers");
942 ctf_dynset_t
*citer_hashes
;
944 if ((citer_hashes
= make_set_element (d
->cd_citers
, citer
)) == NULL
)
946 if (ctf_dynset_cinsert (citer_hashes
, hval
) < 0)
953 while ((err
= ctf_dynset_cnext (citers
, &i
, &k
)) == 0)
955 ctf_dynset_t
*citer_hashes
;
956 citer
= (const char *) k
;
958 if ((citer_hashes
= make_set_element (d
->cd_citers
, citer
)) == NULL
)
961 if (ctf_dynset_exists (citer_hashes
, hval
, NULL
))
963 if (ctf_dynset_cinsert (citer_hashes
, hval
) < 0)
966 if (err
!= ECTF_NEXT_END
)
968 ctf_dynset_destroy (citers
);
974 ctf_next_destroy (i
);
976 err
= ctf_errno (input
);
978 ctf_sha1_fini (&hash
, NULL
);
979 ctf_err_warn (fp
, 0, err
, _("%s (%i): %s: during type hashing for type %lx, "
980 "kind %i"), ctf_link_input_name (input
),
981 input_num
, gettext (whaterr
), type
, kind
);
984 ctf_set_errno (fp
, errno
);
985 ctf_err_warn (fp
, 0, 0, _("%s (%i): %s: during type hashing for type %lx, "
986 "kind %i"), ctf_link_input_name (input
),
987 input_num
, gettext (whaterr
), type
, kind
);
991 /* Hash a TYPE in the INPUT: FP is the eventual output, where the ctf_dedup
992 state is stored. INPUT_NUM is the number of this input in the set of inputs.
993 Record its hash in FP's cd_type_hashes once it is known. PARENTS is
994 described in the comment above ctf_dedup.
996 (The flags argument currently accepts only the flag
997 CTF_DEDUP_HASH_INTERNAL_CHILD, an implementation detail used to prevent
998 struct/union hashing in recursive traversals below the TYPE.)
1000 We use the CTF API rather than direct access wherever possible, because types
1001 that appear identical through the API should be considered identical, with
1002 one exception: slices should only be considered identical to other slices,
1003 not to the corresponding unsliced type.
1005 The POPULATE_FUN is a mandatory hook that populates other mappings with each
1006 type we see (excepting types that are recursively hashed as stubs). The
1007 caller should not rely on the order of calls to this hook, though it will be
1008 called at least once for every non-stub reference to every type.
1010 Returns a hash value (an atom), or NULL on error. */
1013 ctf_dedup_hash_type (ctf_dict_t
*fp
, ctf_dict_t
*input
,
1014 ctf_dict_t
**inputs
, uint32_t *parents
,
1015 int input_num
, ctf_id_t type
, int flags
,
1016 unsigned long depth
,
1017 int (*populate_fun
) (ctf_dict_t
*fp
,
1019 ctf_dict_t
**inputs
,
1023 const char *decorated_name
,
1026 ctf_dedup_t
*d
= &fp
->ctf_dedup
;
1027 const ctf_type_t
*tp
;
1029 const char *hval
= NULL
;
1031 const char *whaterr
;
1032 const char *decorated
= NULL
;
1033 uint32_t kind
, fwdkind
;
1037 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
1038 ctf_dprintf ("%lu: ctf_dedup_hash_type (%i, %lx, flags %x)\n", depth
, input_num
, type
, flags
);
1041 /* The unimplemented type doesn't really exist, but must be noted in parent
1042 hashes: so it gets a fixed, arbitrary hash. */
1044 return "00000000000000000000";
1046 /* Possible optimization: if the input type is in the parent type space, just
1047 copy recursively-cited hashes from the parent's types into the output
1048 mapping rather than rehashing them. */
1050 type_id
= CTF_DEDUP_GID (fp
, input_num
, type
);
1052 if ((tp
= ctf_lookup_by_id (&input
, type
)) == NULL
)
1054 ctf_set_errno (fp
, ctf_errno (input
));
1055 ctf_err_warn (fp
, 0, 0, _("%s (%i): lookup failure for type %lx: "
1056 "flags %x"), ctf_link_input_name (input
),
1057 input_num
, type
, flags
);
1058 return NULL
; /* errno is set for us. */
1061 kind
= LCTF_INFO_KIND (input
, tp
->ctt_info
);
1062 name
= ctf_strraw (input
, tp
->ctt_name
);
1064 if (tp
->ctt_name
== 0 || !name
|| name
[0] == '\0')
1067 /* Treat the unknown kind just like the unimplemented type. */
1068 if (kind
== CTF_K_UNKNOWN
)
1069 return "00000000000000000000";
1071 /* Decorate the name appropriately for the namespace it appears in: forwards
1072 appear in the namespace of their referent. */
1077 if (kind
== CTF_K_FORWARD
)
1078 fwdkind
= tp
->ctt_type
;
1080 if ((decorated
= ctf_decorate_type_name (fp
, name
, fwdkind
)) == NULL
)
1081 return NULL
; /* errno is set for us. */
1084 /* If not hashing a stub, we can rely on various sorts of caches.
1086 Optimization opportunity: we may be able to avoid calling the populate_fun
1089 if (!ctf_dedup_is_stub (name
, kind
, fwdkind
, flags
))
1091 if ((hval
= ctf_dynhash_lookup (d
->cd_type_hashes
, type_id
)) != NULL
)
1093 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
1094 ctf_dprintf ("%lu: Known hash for ID %i/%lx: %s\n", depth
, input_num
,
1097 populate_fun (fp
, input
, inputs
, input_num
, type
, type_id
,
1104 /* We have never seen this type before, and must figure out its hash and the
1105 hashes of the types it cites.
1107 Hash this type, and call ourselves recursively. (The hashing part is
1108 optional, and is disabled if overidden_hval is set.) */
1110 if ((hval
= ctf_dedup_rhash_type (fp
, input
, inputs
, parents
, input_num
,
1111 type
, type_id
, tp
, name
, decorated
,
1112 kind
, flags
, depth
, populate_fun
)) == NULL
)
1113 return NULL
; /* errno is set for us. */
1115 /* The hash of this type is now known: record it unless caching is unsafe
1116 because the hash value will change later. This will be the final storage
1117 of this type's hash, so we call the population function on it. */
1119 if (!ctf_dedup_is_stub (name
, kind
, fwdkind
, flags
))
1121 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
1122 ctf_dprintf ("Caching %lx, ID %p (%s), %s in final location\n", type
,
1123 type_id
, name
? name
: "", hval
);
1126 if (ctf_dynhash_cinsert (d
->cd_type_hashes
, type_id
, hval
) < 0)
1128 whaterr
= N_("error hash caching");
1132 if (populate_fun (fp
, input
, inputs
, input_num
, type
, type_id
,
1133 decorated
, hval
) < 0)
1135 whaterr
= N_("error calling population function");
1136 goto err
; /* errno is set for us. */
1140 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
1141 ctf_dprintf ("%lu: Returning final hash for ID %i/%lx: %s\n", depth
,
1142 input_num
, type
, hval
);
1147 ctf_set_errno (fp
, errno
);
1149 ctf_err_warn (fp
, 0, 0, _("%s (%i): %s: during type hashing, "
1150 "type %lx, kind %i"),
1151 ctf_link_input_name (input
), input_num
,
1152 gettext (whaterr
), type
, kind
);
1156 /* Populate a number of useful mappings not directly used by the hashing
1157 machinery: the output mapping, the cd_name_counts mapping from name -> hash
1158 -> count of hashval deduplication state for a given hashed type, and the
1159 cd_output_first_tu mapping. */
1162 ctf_dedup_populate_mappings (ctf_dict_t
*fp
, ctf_dict_t
*input _libctf_unused_
,
1163 ctf_dict_t
**inputs _libctf_unused_
,
1164 int input_num _libctf_unused_
,
1165 ctf_id_t type _libctf_unused_
, void *id
,
1166 const char *decorated_name
,
1169 ctf_dedup_t
*d
= &fp
->ctf_dedup
;
1170 ctf_dynset_t
*type_ids
;
1171 ctf_dynhash_t
*name_counts
;
1174 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
1175 ctf_dprintf ("Hash %s, %s, into output mapping for %i/%lx @ %s\n",
1176 hval
, decorated_name
? decorated_name
: "(unnamed)",
1177 input_num
, type
, ctf_link_input_name (input
));
1179 const char *orig_hval
;
1181 /* Make sure we never map a single GID to multiple hash values. */
1183 if ((orig_hval
= ctf_dynhash_lookup (d
->cd_output_mapping_guard
, id
)) != NULL
)
1185 /* We can rely on pointer identity here, since all hashes are
1187 if (!ctf_assert (fp
, orig_hval
== hval
))
1191 if (ctf_dynhash_cinsert (d
->cd_output_mapping_guard
, id
, hval
) < 0)
1192 return ctf_set_errno (fp
, errno
);
1195 /* Record the type in the output mapping: if this is the first time this type
1196 has been seen, also record it in the cd_output_first_gid. Because we
1197 traverse types in TU order and we do not merge types after the hashing
1198 phase, this will be the lowest TU this type ever appears in. */
1200 if ((type_ids
= ctf_dynhash_lookup (d
->cd_output_mapping
,
1203 if (ctf_dynhash_cinsert (d
->cd_output_first_gid
, hval
, id
) < 0)
1204 return ctf_set_errno (fp
, errno
);
1206 if ((type_ids
= ctf_dynset_create (htab_hash_pointer
,
1209 return ctf_set_errno (fp
, errno
);
1210 if (ctf_dynhash_insert (d
->cd_output_mapping
, (void *) hval
,
1213 ctf_dynset_destroy (type_ids
);
1214 return ctf_set_errno (fp
, errno
);
1217 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
1219 /* Verify that all types with this hash are of the same kind, and that the
1220 first TU a type was seen in never falls. */
1224 ctf_next_t
*i
= NULL
;
1225 int orig_kind
= ctf_type_kind_unsliced (input
, type
);
1228 orig_first_tu
= CTF_DEDUP_GID_TO_INPUT
1229 (ctf_dynhash_lookup (d
->cd_output_first_gid
, hval
));
1230 if (!ctf_assert (fp
, orig_first_tu
<= CTF_DEDUP_GID_TO_INPUT (id
)))
1233 while ((err
= ctf_dynset_cnext (type_ids
, &i
, &one_id
)) == 0)
1235 ctf_dict_t
*foo
= inputs
[CTF_DEDUP_GID_TO_INPUT (one_id
)];
1236 ctf_id_t bar
= CTF_DEDUP_GID_TO_TYPE (one_id
);
1237 if (ctf_type_kind_unsliced (foo
, bar
) != orig_kind
)
1239 ctf_err_warn (fp
, 1, 0, "added wrong kind to output mapping "
1240 "for hash %s named %s: %p/%lx from %s is "
1241 "kind %i, but newly-added %p/%lx from %s is "
1243 decorated_name
? decorated_name
: "(unnamed)",
1245 ctf_link_input_name (foo
),
1246 ctf_type_kind_unsliced (foo
, bar
),
1247 (void *) input
, type
,
1248 ctf_link_input_name (input
), orig_kind
);
1249 if (!ctf_assert (fp
, ctf_type_kind_unsliced (foo
, bar
)
1254 if (err
!= ECTF_NEXT_END
)
1255 return ctf_set_errno (fp
, err
);
1259 /* This function will be repeatedly called for the same types many times:
1260 don't waste time reinserting the same keys in that case. */
1261 if (!ctf_dynset_exists (type_ids
, id
, NULL
)
1262 && ctf_dynset_insert (type_ids
, id
) < 0)
1263 return ctf_set_errno (fp
, errno
);
1265 /* The rest only needs to happen for types with names. */
1266 if (!decorated_name
)
1269 /* Count the number of occurrences of the hash value for this GID. */
1271 hval
= ctf_dynhash_lookup (d
->cd_type_hashes
, id
);
1273 /* Mapping from name -> hash(hashval, count) not already present? */
1274 if ((name_counts
= ctf_dynhash_lookup (d
->cd_name_counts
,
1275 decorated_name
)) == NULL
)
1277 if ((name_counts
= ctf_dynhash_create (ctf_hash_string
,
1279 NULL
, NULL
)) == NULL
)
1280 return ctf_set_errno (fp
, errno
);
1281 if (ctf_dynhash_cinsert (d
->cd_name_counts
, decorated_name
,
1284 ctf_dynhash_destroy (name_counts
);
1285 return ctf_set_errno (fp
, errno
);
1289 /* This will, conveniently, return NULL (i.e. 0) for a new entry. */
1290 count
= (long int) (uintptr_t) ctf_dynhash_lookup (name_counts
, hval
);
1292 if (ctf_dynhash_cinsert (name_counts
, hval
,
1293 (const void *) (uintptr_t) (count
+ 1)) < 0)
1294 return ctf_set_errno (fp
, errno
);
1299 /* Mark a single hash as corresponding to a conflicting type. Mark all types
1300 that cite it as conflicting as well, terminating the recursive walk only when
1301 types that are already conflicted or types do not cite other types are seen.
1302 (Tagged structures and unions do not appear in the cd_citers graph, so the
1303 walk also terminates there, since any reference to a conflicting structure is
1304 just going to reference an unconflicting forward instead: see
1305 ctf_dedup_maybe_synthesize_forward.) */
1308 ctf_dedup_mark_conflicting_hash (ctf_dict_t
*fp
, const char *hval
)
1310 ctf_dedup_t
*d
= &fp
->ctf_dedup
;
1311 ctf_next_t
*i
= NULL
;
1314 ctf_dynset_t
*citers
;
1316 /* Mark conflicted if not already so marked. */
1317 if (ctf_dynset_exists (d
->cd_conflicting_types
, hval
, NULL
))
1320 ctf_dprintf ("Marking %s as conflicted\n", hval
);
1322 if (ctf_dynset_cinsert (d
->cd_conflicting_types
, hval
) < 0)
1324 ctf_dprintf ("Out of memory marking %s as conflicted\n", hval
);
1325 ctf_set_errno (fp
, errno
);
1329 /* If any types cite this type, mark them conflicted too. */
1330 if ((citers
= ctf_dynhash_lookup (d
->cd_citers
, hval
)) == NULL
)
1333 while ((err
= ctf_dynset_cnext (citers
, &i
, &k
)) == 0)
1335 const char *hv
= (const char *) k
;
1337 if (ctf_dynset_exists (d
->cd_conflicting_types
, hv
, NULL
))
1340 if (ctf_dedup_mark_conflicting_hash (fp
, hv
) < 0)
1342 ctf_next_destroy (i
);
1343 return -1; /* errno is set for us. */
1346 if (err
!= ECTF_NEXT_END
)
1347 return ctf_set_errno (fp
, err
);
1352 /* Look up a type kind from the output mapping, given a type hash value. */
1354 ctf_dedup_hash_kind (ctf_dict_t
*fp
, ctf_dict_t
**inputs
, const char *hash
)
1356 ctf_dedup_t
*d
= &fp
->ctf_dedup
;
1358 ctf_dynset_t
*type_ids
;
1360 /* Precondition: the output mapping is populated. */
1361 if (!ctf_assert (fp
, ctf_dynhash_elements (d
->cd_output_mapping
) > 0))
1364 /* Look up some GID from the output hash for this type. (They are all
1365 identical, so we can pick any). Don't assert if someone calls this
1366 function wrongly, but do assert if the output mapping knows about the hash,
1367 but has nothing associated with it. */
1369 type_ids
= ctf_dynhash_lookup (d
->cd_output_mapping
, hash
);
1372 ctf_dprintf ("Looked up type kind by nonexistent hash %s.\n", hash
);
1373 return ctf_set_errno (fp
, ECTF_INTERNAL
);
1375 id
= ctf_dynset_lookup_any (type_ids
);
1376 if (!ctf_assert (fp
, id
))
1379 return ctf_type_kind_unsliced (inputs
[CTF_DEDUP_GID_TO_INPUT (id
)],
1380 CTF_DEDUP_GID_TO_TYPE (id
));
1383 /* Used to keep a count of types: i.e. distinct type hash values. */
1384 typedef struct ctf_dedup_type_counter
1387 ctf_dict_t
**inputs
;
1388 int num_non_forwards
;
1389 } ctf_dedup_type_counter_t
;
1391 /* Add to the type counter for one name entry from the cd_name_counts. */
1393 ctf_dedup_count_types (void *key_
, void *value _libctf_unused_
, void *arg_
)
1395 const char *hval
= (const char *) key_
;
1397 ctf_dedup_type_counter_t
*arg
= (ctf_dedup_type_counter_t
*) arg_
;
1399 kind
= ctf_dedup_hash_kind (arg
->fp
, arg
->inputs
, hval
);
1401 /* We rely on ctf_dedup_hash_kind setting the fp to -ECTF_INTERNAL on error to
1402 smuggle errors out of here. */
1404 if (kind
!= CTF_K_FORWARD
)
1406 arg
->num_non_forwards
++;
1407 ctf_dprintf ("Counting hash %s: kind %i: num_non_forwards is %i\n",
1408 hval
, kind
, arg
->num_non_forwards
);
1411 /* We only need to know if there is more than one non-forward (an ambiguous
1412 type): don't waste time iterating any more than needed to figure that
1415 if (arg
->num_non_forwards
> 1)
1421 /* Detect name ambiguity and mark ambiguous names as conflicting, other than the
1424 ctf_dedup_detect_name_ambiguity (ctf_dict_t
*fp
, ctf_dict_t
**inputs
)
1426 ctf_dedup_t
*d
= &fp
->ctf_dedup
;
1427 ctf_next_t
*i
= NULL
;
1431 const char *whaterr
;
1433 /* Go through cd_name_counts for all CTF namespaces in turn. */
1435 while ((err
= ctf_dynhash_next (d
->cd_name_counts
, &i
, &k
, &v
)) == 0)
1437 const char *decorated
= (const char *) k
;
1438 ctf_dynhash_t
*name_counts
= (ctf_dynhash_t
*) v
;
1439 ctf_next_t
*j
= NULL
;
1441 /* If this is a forwardable kind or a forward (which we can tell without
1442 consulting the type because its decorated name has a space as its
1443 second character: see ctf_decorate_type_name), we are only interested
1444 in whether this name has many hashes associated with it: any such name
1445 is necessarily ambiguous, and types with that name are conflicting.
1446 Once we know whether this is true, we can skip to the next name: so use
1447 ctf_dynhash_iter_find for efficiency. */
1449 if (decorated
[0] != '\0' && decorated
[1] == ' ')
1451 ctf_dedup_type_counter_t counters
= { fp
, inputs
, 0 };
1452 ctf_dynhash_t
*counts
= (ctf_dynhash_t
*) v
;
1454 ctf_dynhash_iter_find (counts
, ctf_dedup_count_types
, &counters
);
1456 /* Check for assertion failure and pass it up. */
1457 if (ctf_errno (fp
) == ECTF_INTERNAL
)
1460 if (counters
.num_non_forwards
> 1)
1464 while ((err
= ctf_dynhash_cnext (counts
, &j
, &hval_
, NULL
)) == 0)
1466 const char *hval
= (const char *) hval_
;
1467 ctf_dynset_t
*type_ids
;
1471 /* Dig through the types in this hash to find the non-forwards
1472 and mark them ambiguous. */
1474 type_ids
= ctf_dynhash_lookup (d
->cd_output_mapping
, hval
);
1476 /* Nonexistent? Must be a forward with no referent. */
1480 id
= ctf_dynset_lookup_any (type_ids
);
1482 kind
= ctf_type_kind (inputs
[CTF_DEDUP_GID_TO_INPUT (id
)],
1483 CTF_DEDUP_GID_TO_TYPE (id
));
1485 if (kind
!= CTF_K_FORWARD
)
1487 ctf_dprintf ("Marking %p, with hash %s, conflicting: one "
1488 "of many non-forward GIDs for %s\n", id
,
1490 ctf_dedup_mark_conflicting_hash (fp
, hval
);
1493 if (err
!= ECTF_NEXT_END
)
1495 whaterr
= N_("error marking conflicting structs/unions");
1502 /* This is an ordinary type. Find the most common type with this
1503 name, and mark it unconflicting: all others are conflicting. (We
1504 cannot do this sort of popularity contest with forwardable types
1505 because any forwards to that type would be immediately unified with
1506 the most-popular type on insertion, and we want conflicting structs
1507 et al to have all forwards left intact, so the user is notified
1508 that this type is conflicting. TODO: improve this in future by
1509 setting such forwards non-root-visible.) */
1514 long max_hcount
= -1;
1515 const char *max_hval
= NULL
;
1517 if (ctf_dynhash_elements (name_counts
) <= 1)
1520 /* First find the most common. */
1521 while ((err
= ctf_dynhash_cnext (name_counts
, &j
, &key
, &count
)) == 0)
1523 hval
= (const char *) key
;
1524 if ((long int) (uintptr_t) count
> max_hcount
)
1526 max_hcount
= (long int) (uintptr_t) count
;
1530 if (err
!= ECTF_NEXT_END
)
1532 whaterr
= N_("error finding commonest conflicting type");
1536 /* Mark all the others as conflicting. */
1537 while ((err
= ctf_dynhash_cnext (name_counts
, &j
, &key
, NULL
)) == 0)
1539 hval
= (const char *) key
;
1540 if (strcmp (max_hval
, hval
) == 0)
1543 ctf_dprintf ("Marking %s, an uncommon hash for %s, conflicting\n",
1544 hval
, (const char *) k
);
1545 if (ctf_dedup_mark_conflicting_hash (fp
, hval
) < 0)
1547 whaterr
= N_("error marking hashes as conflicting");
1551 if (err
!= ECTF_NEXT_END
)
1553 whaterr
= N_("marking uncommon conflicting types");
1558 if (err
!= ECTF_NEXT_END
)
1560 whaterr
= N_("scanning for ambiguous names");
1567 ctf_next_destroy (i
);
1568 ctf_err_warn (fp
, 0, 0, "%s", gettext (whaterr
));
1569 return -1; /* errno is set for us. */
1572 ctf_err_warn (fp
, 0, err
, _("iteration failed: %s"), gettext (whaterr
));
1573 return ctf_set_errno (fp
, err
);
1576 ctf_next_destroy (i
);
1577 return -1; /* errno is set for us. */
1580 /* Initialize the deduplication machinery. */
1583 ctf_dedup_init (ctf_dict_t
*fp
)
1585 ctf_dedup_t
*d
= &fp
->ctf_dedup
;
1588 if (ctf_dedup_atoms_init (fp
) < 0)
1591 #if IDS_NEED_ALLOCATION
1592 if ((d
->cd_id_to_dict_t
= ctf_dynhash_create (ctf_hash_type_id_key
,
1593 ctf_hash_eq_type_id_key
,
1594 free
, NULL
)) == NULL
)
1598 for (i
= 0; i
< 4; i
++)
1600 if ((d
->cd_decorated_names
[i
] = ctf_dynhash_create (ctf_hash_string
,
1602 NULL
, NULL
)) == NULL
)
1606 if ((d
->cd_name_counts
1607 = ctf_dynhash_create (ctf_hash_string
,
1608 ctf_hash_eq_string
, NULL
,
1609 (ctf_hash_free_fun
) ctf_dynhash_destroy
)) == NULL
)
1612 if ((d
->cd_type_hashes
1613 = ctf_dynhash_create (ctf_hash_integer
,
1614 ctf_hash_eq_integer
,
1615 NULL
, NULL
)) == NULL
)
1618 if ((d
->cd_struct_origin
1619 = ctf_dynhash_create (ctf_hash_string
,
1621 NULL
, NULL
)) == NULL
)
1625 = ctf_dynhash_create (ctf_hash_string
,
1626 ctf_hash_eq_string
, NULL
,
1627 (ctf_hash_free_fun
) ctf_dynset_destroy
)) == NULL
)
1630 if ((d
->cd_output_mapping
1631 = ctf_dynhash_create (ctf_hash_string
,
1632 ctf_hash_eq_string
, NULL
,
1633 (ctf_hash_free_fun
) ctf_dynset_destroy
)) == NULL
)
1636 if ((d
->cd_output_first_gid
1637 = ctf_dynhash_create (ctf_hash_string
,
1639 NULL
, NULL
)) == NULL
)
1642 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
1643 if ((d
->cd_output_mapping_guard
1644 = ctf_dynhash_create (ctf_hash_integer
,
1645 ctf_hash_eq_integer
, NULL
, NULL
)) == NULL
)
1649 if ((d
->cd_input_nums
1650 = ctf_dynhash_create (ctf_hash_integer
,
1651 ctf_hash_eq_integer
,
1652 NULL
, NULL
)) == NULL
)
1655 if ((d
->cd_emission_struct_members
1656 = ctf_dynhash_create (ctf_hash_integer
,
1657 ctf_hash_eq_integer
,
1658 NULL
, NULL
)) == NULL
)
1661 if ((d
->cd_conflicting_types
1662 = ctf_dynset_create (htab_hash_string
,
1663 ctf_dynset_eq_string
, NULL
)) == NULL
)
1669 ctf_err_warn (fp
, 0, ENOMEM
, _("ctf_dedup_init: cannot initialize: "
1671 return ctf_set_errno (fp
, ENOMEM
);
1674 /* No ctf_dedup calls are allowed after this call other than starting a new
1675 deduplication via ctf_dedup (not even ctf_dedup_type_mapping lookups). */
1677 ctf_dedup_fini (ctf_dict_t
*fp
, ctf_dict_t
**outputs
, uint32_t noutputs
)
1679 ctf_dedup_t
*d
= &fp
->ctf_dedup
;
1682 /* ctf_dedup_atoms is kept across links. */
1683 #if IDS_NEED_ALLOCATION
1684 ctf_dynhash_destroy (d
->cd_id_to_dict_t
);
1686 for (i
= 0; i
< 4; i
++)
1687 ctf_dynhash_destroy (d
->cd_decorated_names
[i
]);
1688 ctf_dynhash_destroy (d
->cd_name_counts
);
1689 ctf_dynhash_destroy (d
->cd_type_hashes
);
1690 ctf_dynhash_destroy (d
->cd_struct_origin
);
1691 ctf_dynhash_destroy (d
->cd_citers
);
1692 ctf_dynhash_destroy (d
->cd_output_mapping
);
1693 ctf_dynhash_destroy (d
->cd_output_first_gid
);
1694 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
1695 ctf_dynhash_destroy (d
->cd_output_mapping_guard
);
1697 ctf_dynhash_destroy (d
->cd_input_nums
);
1698 ctf_dynhash_destroy (d
->cd_emission_struct_members
);
1699 ctf_dynset_destroy (d
->cd_conflicting_types
);
1701 /* Free the per-output state. */
1704 for (i
= 0; i
< noutputs
; i
++)
1706 ctf_dedup_t
*od
= &outputs
[i
]->ctf_dedup
;
1707 ctf_dynhash_destroy (od
->cd_output_emission_hashes
);
1708 ctf_dynhash_destroy (od
->cd_output_emission_conflicted_forwards
);
1709 ctf_dict_close (od
->cd_output
);
1712 memset (d
, 0, sizeof (ctf_dedup_t
));
1715 /* Return 1 if this type is cited by multiple input dictionaries. */
1718 ctf_dedup_multiple_input_dicts (ctf_dict_t
*output
, ctf_dict_t
**inputs
,
1721 ctf_dedup_t
*d
= &output
->ctf_dedup
;
1722 ctf_dynset_t
*type_ids
;
1723 ctf_next_t
*i
= NULL
;
1725 ctf_dict_t
*found
= NULL
, *relative_found
= NULL
;
1726 const char *type_id
;
1727 ctf_dict_t
*input_fp
;
1730 const char *decorated
;
1735 type_ids
= ctf_dynhash_lookup (d
->cd_output_mapping
, hval
);
1736 if (!ctf_assert (output
, type_ids
))
1739 /* Scan across the IDs until we find proof that two disjoint dictionaries
1740 are referenced. Exit as soon as possible. Optimization opportunity, but
1741 possibly not worth it, given that this is only executed in
1742 CTF_LINK_SHARE_DUPLICATED mode. */
1744 while ((err
= ctf_dynset_next (type_ids
, &i
, &id
)) == 0)
1746 ctf_dict_t
*fp
= inputs
[CTF_DEDUP_GID_TO_INPUT (id
)];
1748 if (fp
== found
|| fp
== relative_found
)
1758 && (fp
->ctf_parent
== found
|| found
->ctf_parent
== fp
))
1760 relative_found
= fp
;
1765 ctf_next_destroy (i
);
1768 if ((err
!= ECTF_NEXT_END
) && (err
!= 0))
1770 ctf_err_warn (output
, 0, err
, _("iteration error "
1771 "propagating conflictedness"));
1772 return ctf_set_errno (output
, err
);
1778 /* This type itself does not appear in multiple input dicts: how about another
1779 related type with the same name (e.g. a forward if this is a struct,
1782 type_id
= ctf_dynset_lookup_any (type_ids
);
1783 if (!ctf_assert (output
, type_id
))
1786 input_fp
= inputs
[CTF_DEDUP_GID_TO_INPUT (type_id
)];
1787 input_id
= CTF_DEDUP_GID_TO_TYPE (type_id
);
1788 fwdkind
= ctf_type_kind_forwarded (input_fp
, input_id
);
1789 name
= ctf_type_name_raw (input_fp
, input_id
);
1791 if ((fwdkind
== CTF_K_STRUCT
|| fwdkind
== CTF_K_UNION
)
1796 if ((decorated
= ctf_decorate_type_name (output
, name
,
1798 return -1; /* errno is set for us. */
1800 origin
= ctf_dynhash_lookup (d
->cd_struct_origin
, decorated
);
1801 if ((origin
!= NULL
) && (CTF_DEDUP_GID_TO_INPUT (origin
) < 0))
1808 /* Demote unconflicting types which reference only one input, or which reference
1809 two inputs where one input is the parent of the other, into conflicting
1810 types. Only used if the link mode is CTF_LINK_SHARE_DUPLICATED. */
1813 ctf_dedup_conflictify_unshared (ctf_dict_t
*output
, ctf_dict_t
**inputs
)
1815 ctf_dedup_t
*d
= &output
->ctf_dedup
;
1816 ctf_next_t
*i
= NULL
;
1819 ctf_dynset_t
*to_mark
= NULL
;
1821 if ((to_mark
= ctf_dynset_create (htab_hash_string
, ctf_dynset_eq_string
,
1825 while ((err
= ctf_dynhash_cnext (d
->cd_output_mapping
, &i
, &k
, NULL
)) == 0)
1827 const char *hval
= (const char *) k
;
1830 /* Types referenced by only one dict, with no type appearing under that
1831 name elsewhere, are marked conflicting. */
1833 conflicting
= !ctf_dedup_multiple_input_dicts (output
, inputs
, hval
);
1835 if (conflicting
< 0)
1836 goto err
; /* errno is set for us. */
1839 if (ctf_dynset_cinsert (to_mark
, hval
) < 0)
1842 if (err
!= ECTF_NEXT_END
)
1845 while ((err
= ctf_dynset_cnext (to_mark
, &i
, &k
)) == 0)
1847 const char *hval
= (const char *) k
;
1849 if (ctf_dedup_mark_conflicting_hash (output
, hval
) < 0)
1852 if (err
!= ECTF_NEXT_END
)
1855 ctf_dynset_destroy (to_mark
);
1860 ctf_set_errno (output
, errno
);
1862 err
= ctf_errno (output
);
1863 ctf_next_destroy (i
);
1865 ctf_dynset_destroy (to_mark
);
1866 ctf_err_warn (output
, 0, err
, _("conflictifying unshared types"));
1867 return ctf_set_errno (output
, err
);
1870 /* The core deduplicator. Populate cd_output_mapping in the output ctf_dedup
1871 with a mapping of all types that belong in this dictionary and where they
1872 come from, and cd_conflicting_types with an indication of whether each type
1873 is conflicted or not. OUTPUT is the top-level output: INPUTS is the array of
1874 input dicts; NINPUTS is the size of that array; PARENTS is an NINPUTS-element
1875 array with each element corresponding to a input which is a child dict set to
1876 the number in the INPUTS array of that input's parent.
1878 If CU_MAPPED is set, this is a first pass for a link with a non-empty CU
1879 mapping: only one output will result.
1881 Only deduplicates: does not emit the types into the output. Call
1882 ctf_dedup_emit afterwards to do that. */
1885 ctf_dedup (ctf_dict_t
*output
, ctf_dict_t
**inputs
, uint32_t ninputs
,
1886 uint32_t *parents
, int cu_mapped
)
1888 ctf_dedup_t
*d
= &output
->ctf_dedup
;
1890 ctf_next_t
*it
= NULL
;
1892 if (ctf_dedup_init (output
) < 0)
1893 return -1; /* errno is set for us. */
1895 for (i
= 0; i
< ninputs
; i
++)
1897 ctf_dprintf ("Input %i: %s\n", (int) i
, ctf_link_input_name (inputs
[i
]));
1898 if (ctf_dynhash_insert (d
->cd_input_nums
, inputs
[i
],
1899 (void *) (uintptr_t) i
) < 0)
1901 ctf_set_errno (output
, errno
);
1902 ctf_err_warn (output
, 0, errno
, _("ctf_dedup: cannot initialize: %s\n"),
1903 ctf_errmsg (errno
));
1908 /* Some flags do not apply when CU-mapping: this is not a duplicated link,
1909 because there is only one output and we really don't want to end up marking
1910 all nonconflicting but appears-only-once types as conflicting (which in the
1911 CU-mapped link means we'd mark them all as non-root-visible!). */
1912 d
->cd_link_flags
= output
->ctf_link_flags
;
1914 d
->cd_link_flags
&= ~(CTF_LINK_SHARE_DUPLICATED
);
1916 /* Compute hash values for all types, recursively, treating child structures
1917 and unions equivalent to forwards, and hashing in the name of the referent
1918 of each such type into structures, unions, and non-opaque forwards.
1919 Populate a mapping from decorated name (including an indication of
1920 struct/union/enum namespace) to count of type hash values in
1921 cd_name_counts, a mapping from and a mapping from hash values to input type
1922 IDs in cd_output_mapping. */
1924 ctf_dprintf ("Computing type hashes\n");
1925 for (i
= 0; i
< ninputs
; i
++)
1929 while ((id
= ctf_type_next (inputs
[i
], &it
, NULL
, 1)) != CTF_ERR
)
1931 if (ctf_dedup_hash_type (output
, inputs
[i
], inputs
,
1932 parents
, i
, id
, 0, 0,
1933 ctf_dedup_populate_mappings
) == NULL
)
1934 goto err
; /* errno is set for us. */
1936 if (ctf_errno (inputs
[i
]) != ECTF_NEXT_END
)
1938 ctf_set_errno (output
, ctf_errno (inputs
[i
]));
1939 ctf_err_warn (output
, 0, 0, _("iteration failure "
1940 "computing type hashes"));
1945 /* Go through the cd_name_counts name->hash->count mapping for all CTF
1946 namespaces: any name with many hashes associated with it at this stage is
1947 necessarily ambiguous. Mark all the hashes except the most common as
1948 conflicting in the output. */
1950 ctf_dprintf ("Detecting type name ambiguity\n");
1951 if (ctf_dedup_detect_name_ambiguity (output
, inputs
) < 0)
1952 goto err
; /* errno is set for us. */
1954 /* If the link mode is CTF_LINK_SHARE_DUPLICATED, we change any unconflicting
1955 types whose output mapping references only one input dict into a
1956 conflicting type, so that they end up in the per-CU dictionaries. */
1958 if (d
->cd_link_flags
& CTF_LINK_SHARE_DUPLICATED
)
1960 ctf_dprintf ("Conflictifying unshared types\n");
1961 if (ctf_dedup_conflictify_unshared (output
, inputs
) < 0)
1962 goto err
; /* errno is set for us. */
1967 ctf_dedup_fini (output
, NULL
, 0);
1972 ctf_dedup_rwalk_output_mapping (ctf_dict_t
*output
, ctf_dict_t
**inputs
,
1973 uint32_t ninputs
, uint32_t *parents
,
1974 ctf_dynset_t
*already_visited
,
1976 int (*visit_fun
) (const char *hval
,
1978 ctf_dict_t
**inputs
,
1981 int already_visited
,
1987 void *arg
, unsigned long depth
);
1989 /* Like ctf_dedup_rwalk_output_mapping (which see), only takes a single target
1990 type and visits it. */
1992 ctf_dedup_rwalk_one_output_mapping (ctf_dict_t
*output
,
1993 ctf_dict_t
**inputs
, uint32_t ninputs
,
1995 ctf_dynset_t
*already_visited
,
1996 int visited
, void *type_id
,
1998 int (*visit_fun
) (const char *hval
,
2000 ctf_dict_t
**inputs
,
2003 int already_visited
,
2009 void *arg
, unsigned long depth
)
2011 ctf_dedup_t
*d
= &output
->ctf_dedup
;
2016 const char *whaterr
;
2018 input_num
= CTF_DEDUP_GID_TO_INPUT (type_id
);
2019 fp
= inputs
[input_num
];
2020 type
= CTF_DEDUP_GID_TO_TYPE (type_id
);
2022 ctf_dprintf ("%lu: Starting walk over type %s, %i/%lx (%p), from %s, "
2023 "kind %i\n", depth
, hval
, input_num
, type
, (void *) fp
,
2024 ctf_link_input_name (fp
), ctf_type_kind_unsliced (fp
, type
));
2026 /* Get the single call we do if this type has already been visited out of the
2029 return visit_fun (hval
, output
, inputs
, ninputs
, parents
, visited
, fp
,
2030 type
, type_id
, depth
, arg
);
2032 /* This macro is really ugly, but the alternative is repeating this code many
2033 times, which is worse. */
2035 #define CTF_TYPE_WALK(type, errlabel, errmsg) \
2039 const char *hashval; \
2040 int cited_type_input_num = input_num; \
2042 if ((fp->ctf_flags & LCTF_CHILD) && (LCTF_TYPE_ISPARENT (fp, type))) \
2043 cited_type_input_num = parents[input_num]; \
2045 type_id = CTF_DEDUP_GID (output, cited_type_input_num, type); \
2049 ctf_dprintf ("Walking: unimplemented type\n"); \
2053 ctf_dprintf ("Looking up ID %i/%lx in type hashes\n", \
2054 cited_type_input_num, type); \
2055 hashval = ctf_dynhash_lookup (d->cd_type_hashes, type_id); \
2056 if (!ctf_assert (output, hashval)) \
2058 whaterr = N_("error looking up ID in type hashes"); \
2061 ctf_dprintf ("ID %i/%lx has hash %s\n", cited_type_input_num, type, \
2064 ret = ctf_dedup_rwalk_output_mapping (output, inputs, ninputs, parents, \
2065 already_visited, hashval, \
2066 visit_fun, arg, depth); \
2075 switch (ctf_type_kind_unsliced (fp
, type
))
2078 /* Just skip things of unknown kind. */
2084 /* No types referenced. */
2088 case CTF_K_VOLATILE
:
2090 case CTF_K_RESTRICT
:
2093 CTF_TYPE_WALK (ctf_type_reference (fp
, type
), err
,
2094 N_("error during referenced type walk"));
2101 if (ctf_array_info (fp
, type
, &ar
) < 0)
2103 whaterr
= N_("error during array info lookup");
2107 CTF_TYPE_WALK (ar
.ctr_contents
, err
,
2108 N_("error during array contents type walk"));
2109 CTF_TYPE_WALK (ar
.ctr_index
, err
,
2110 N_("error during array index type walk"));
2114 case CTF_K_FUNCTION
:
2120 if (ctf_func_type_info (fp
, type
, &fi
) < 0)
2122 whaterr
= N_("error during func type info lookup");
2126 CTF_TYPE_WALK (fi
.ctc_return
, err
,
2127 N_("error during func return type walk"));
2129 if ((args
= calloc (fi
.ctc_argc
, sizeof (ctf_id_t
))) == NULL
)
2131 whaterr
= N_("error doing memory allocation");
2135 if (ctf_func_type_args (fp
, type
, fi
.ctc_argc
, args
) < 0)
2137 whaterr
= N_("error doing func arg type lookup");
2142 for (j
= 0; j
< fi
.ctc_argc
; j
++)
2143 CTF_TYPE_WALK (args
[j
], err_free_args
,
2144 N_("error during Func arg type walk"));
2154 /* We do not recursively traverse the members of structures: they are
2155 emitted later, in a separate pass. */
2158 whaterr
= N_("CTF dict corruption: unknown type kind");
2162 return visit_fun (hval
, output
, inputs
, ninputs
, parents
, visited
, fp
, type
,
2163 type_id
, depth
, arg
);
2166 ctf_set_errno (output
, ctf_errno (fp
));
2167 ctf_err_warn (output
, 0, 0, _("%s in input file %s at type ID %lx"),
2168 gettext (whaterr
), ctf_link_input_name (fp
), type
);
2172 /* Recursively traverse the output mapping, and do something with each type
2173 visited, from leaves to root. VISIT_FUN, called as recursion unwinds,
2174 returns a negative error code or zero. Type hashes may be visited more than
2175 once, but are not recursed through repeatedly: ALREADY_VISITED tracks whether
2176 types have already been visited. */
2178 ctf_dedup_rwalk_output_mapping (ctf_dict_t
*output
, ctf_dict_t
**inputs
,
2179 uint32_t ninputs
, uint32_t *parents
,
2180 ctf_dynset_t
*already_visited
,
2182 int (*visit_fun
) (const char *hval
,
2184 ctf_dict_t
**inputs
,
2187 int already_visited
,
2193 void *arg
, unsigned long depth
)
2195 ctf_dedup_t
*d
= &output
->ctf_dedup
;
2196 ctf_next_t
*i
= NULL
;
2199 ctf_dynset_t
*type_ids
;
2204 type_ids
= ctf_dynhash_lookup (d
->cd_output_mapping
, hval
);
2207 ctf_err_warn (output
, 0, ECTF_INTERNAL
,
2208 _("looked up type kind by nonexistent hash %s"), hval
);
2209 return ctf_set_errno (output
, ECTF_INTERNAL
);
2212 /* Have we seen this type before? */
2214 if (!ctf_dynset_exists (already_visited
, hval
, NULL
))
2216 /* Mark as already-visited immediately, to eliminate the possibility of
2217 cycles: but remember we have not actually visited it yet for the
2218 upcoming call to the visit_fun. (All our callers handle cycles
2219 properly themselves, so we can just abort them aggressively as soon as
2220 we find ourselves in one.) */
2223 if (ctf_dynset_cinsert (already_visited
, hval
) < 0)
2225 ctf_err_warn (output
, 0, ENOMEM
,
2226 _("out of memory tracking already-visited types"));
2227 return ctf_set_errno (output
, ENOMEM
);
2231 /* If this type is marked conflicted, traverse members and call
2232 ctf_dedup_rwalk_output_mapping_once on all the unique ones: otherwise, just
2233 pick a random one and use it. */
2235 if (!ctf_dynset_exists (d
->cd_conflicting_types
, hval
, NULL
))
2237 id
= ctf_dynset_lookup_any (type_ids
);
2238 if (!ctf_assert (output
, id
))
2241 return ctf_dedup_rwalk_one_output_mapping (output
, inputs
, ninputs
,
2242 parents
, already_visited
,
2243 visited
, id
, hval
, visit_fun
,
2247 while ((err
= ctf_dynset_next (type_ids
, &i
, &id
)) == 0)
2251 ret
= ctf_dedup_rwalk_one_output_mapping (output
, inputs
, ninputs
,
2252 parents
, already_visited
,
2254 visit_fun
, arg
, depth
);
2257 ctf_next_destroy (i
);
2258 return ret
; /* errno is set for us. */
2261 if (err
!= ECTF_NEXT_END
)
2263 ctf_err_warn (output
, 0, err
, _("cannot walk conflicted type"));
2264 return ctf_set_errno (output
, err
);
2270 typedef struct ctf_sort_om_cb_arg
2272 ctf_dict_t
**inputs
;
2275 } ctf_sort_om_cb_arg_t
;
2277 /* Sort the output mapping into order: types first appearing in earlier inputs
2278 first, parents preceding children: if types first appear in the same input,
2279 sort those with earlier ctf_id_t's first. */
2281 sort_output_mapping (const ctf_next_hkv_t
*one
, const ctf_next_hkv_t
*two
,
2284 ctf_sort_om_cb_arg_t
*arg
= (ctf_sort_om_cb_arg_t
*) arg_
;
2285 ctf_dedup_t
*d
= arg
->d
;
2286 const char *one_hval
= (const char *) one
->hkv_key
;
2287 const char *two_hval
= (const char *) two
->hkv_key
;
2288 void *one_gid
, *two_gid
;
2289 uint32_t one_ninput
;
2290 uint32_t two_ninput
;
2296 one_gid
= ctf_dynhash_lookup (d
->cd_output_first_gid
, one_hval
);
2297 two_gid
= ctf_dynhash_lookup (d
->cd_output_first_gid
, two_hval
);
2299 one_ninput
= CTF_DEDUP_GID_TO_INPUT (one_gid
);
2300 two_ninput
= CTF_DEDUP_GID_TO_INPUT (two_gid
);
2302 one_type
= CTF_DEDUP_GID_TO_TYPE (one_gid
);
2303 two_type
= CTF_DEDUP_GID_TO_TYPE (two_gid
);
2305 /* It's kind of hard to smuggle an assertion failure out of here. */
2306 assert (one_ninput
< arg
->ninputs
&& two_ninput
< arg
->ninputs
);
2308 one_fp
= arg
->inputs
[one_ninput
];
2309 two_fp
= arg
->inputs
[two_ninput
];
2311 /* Parents before children. */
2313 if (!(one_fp
->ctf_flags
& LCTF_CHILD
)
2314 && (two_fp
->ctf_flags
& LCTF_CHILD
))
2316 else if ((one_fp
->ctf_flags
& LCTF_CHILD
)
2317 && !(two_fp
->ctf_flags
& LCTF_CHILD
))
2320 /* ninput order, types appearing in earlier TUs first. */
2322 if (one_ninput
< two_ninput
)
2324 else if (two_ninput
< one_ninput
)
2327 /* Same TU. Earliest ctf_id_t first. They cannot be the same. */
2329 assert (one_type
!= two_type
);
2330 if (one_type
< two_type
)
2336 /* The public entry point to ctf_dedup_rwalk_output_mapping, above. */
2338 ctf_dedup_walk_output_mapping (ctf_dict_t
*output
, ctf_dict_t
**inputs
,
2339 uint32_t ninputs
, uint32_t *parents
,
2340 int (*visit_fun
) (const char *hval
,
2342 ctf_dict_t
**inputs
,
2345 int already_visited
,
2353 ctf_dynset_t
*already_visited
;
2354 ctf_next_t
*i
= NULL
;
2355 ctf_sort_om_cb_arg_t sort_arg
;
2359 if ((already_visited
= ctf_dynset_create (htab_hash_string
,
2360 ctf_dynset_eq_string
,
2362 return ctf_set_errno (output
, ENOMEM
);
2364 sort_arg
.inputs
= inputs
;
2365 sort_arg
.ninputs
= ninputs
;
2366 sort_arg
.d
= &output
->ctf_dedup
;
2368 while ((err
= ctf_dynhash_next_sorted (output
->ctf_dedup
.cd_output_mapping
,
2369 &i
, &k
, NULL
, sort_output_mapping
,
2372 const char *hval
= (const char *) k
;
2374 err
= ctf_dedup_rwalk_output_mapping (output
, inputs
, ninputs
, parents
,
2375 already_visited
, hval
, visit_fun
,
2379 ctf_next_destroy (i
);
2380 goto err
; /* errno is set for us. */
2383 if (err
!= ECTF_NEXT_END
)
2385 ctf_err_warn (output
, 0, err
, _("cannot recurse over output mapping"));
2386 ctf_set_errno (output
, err
);
2389 ctf_dynset_destroy (already_visited
);
2393 ctf_dynset_destroy (already_visited
);
2397 /* Possibly synthesise a synthetic forward in TARGET to subsitute for a
2398 conflicted per-TU type ID in INPUT with hash HVAL. Return its CTF ID, or 0
2399 if none was needed. */
2401 ctf_dedup_maybe_synthesize_forward (ctf_dict_t
*output
, ctf_dict_t
*target
,
2402 ctf_dict_t
*input
, ctf_id_t id
,
2405 ctf_dedup_t
*od
= &output
->ctf_dedup
;
2406 ctf_dedup_t
*td
= &target
->ctf_dedup
;
2409 const char *name
= ctf_type_name_raw (input
, id
);
2410 const char *decorated
;
2412 ctf_id_t emitted_forward
;
2414 if (!ctf_dynset_exists (od
->cd_conflicting_types
, hval
, NULL
)
2415 || target
->ctf_flags
& LCTF_CHILD
2417 || (((kind
= ctf_type_kind_unsliced (input
, id
)) != CTF_K_STRUCT
2418 && kind
!= CTF_K_UNION
&& kind
!= CTF_K_FORWARD
)))
2421 fwdkind
= ctf_type_kind_forwarded (input
, id
);
2423 ctf_dprintf ("Using synthetic forward for conflicted struct/union with "
2426 if (!ctf_assert (output
, name
))
2429 if ((decorated
= ctf_decorate_type_name (output
, name
, fwdkind
)) == NULL
)
2432 if (!ctf_dynhash_lookup_kv (td
->cd_output_emission_conflicted_forwards
,
2433 decorated
, NULL
, &v
))
2435 if ((emitted_forward
= ctf_add_forward (target
, CTF_ADD_ROOT
, name
,
2436 fwdkind
)) == CTF_ERR
)
2438 ctf_set_errno (output
, ctf_errno (target
));
2442 if (ctf_dynhash_cinsert (td
->cd_output_emission_conflicted_forwards
,
2443 decorated
, (void *) (uintptr_t)
2444 emitted_forward
) < 0)
2446 ctf_set_errno (output
, ENOMEM
);
2451 emitted_forward
= (ctf_id_t
) (uintptr_t) v
;
2453 ctf_dprintf ("Cross-TU conflicted struct: passing back forward, %lx\n",
2456 return emitted_forward
;
2459 /* Map a GID in some INPUT dict, in the form of an input number and a ctf_id_t,
2460 into a GID in a target output dict. If it returns 0, this is the
2461 unimplemented type, and the input type must have been 0. The OUTPUT dict is
2462 assumed to be the parent of the TARGET, if it is not the TARGET itself.
2464 Returns CTF_ERR on failure. Responds to an incoming CTF_ERR as an 'id' by
2465 returning CTF_ERR, to simplify callers. Errors are always propagated to the
2466 input, even if they relate to the target, for the same reason. (Target
2467 errors are expected to be very rare.)
2469 If the type in question is a citation of a conflicted type in a different TU,
2470 emit a forward of the right type in its place (if not already emitted), and
2471 record that forward in cd_output_emission_conflicted_forwards. This avoids
2472 the need to replicate the entire type graph below this point in the current
2473 TU (an appalling waste of space).
2475 TODO: maybe replace forwards in the same TU with their referents? Might
2476 make usability a bit better. */
2479 ctf_dedup_id_to_target (ctf_dict_t
*output
, ctf_dict_t
*target
,
2480 ctf_dict_t
**inputs
, uint32_t ninputs
,
2481 uint32_t *parents
, ctf_dict_t
*input
, int input_num
,
2484 ctf_dedup_t
*od
= &output
->ctf_dedup
;
2485 ctf_dedup_t
*td
= &target
->ctf_dedup
;
2486 ctf_dict_t
*err_fp
= input
;
2489 ctf_id_t emitted_forward
;
2491 /* The target type of an error is an error. */
2495 /* The unimplemented type's ID never changes. */
2498 ctf_dprintf ("%i/%lx: unimplemented type\n", input_num
, id
);
2502 ctf_dprintf ("Mapping %i/%lx to target %p (%s)\n", input_num
,
2503 id
, (void *) target
, ctf_link_input_name (target
));
2505 /* If the input type is in the parent type space, and this is a child, reset
2506 the input to the parent (which must already have been emitted, since
2507 emission of parent dicts happens before children). */
2508 if ((input
->ctf_flags
& LCTF_CHILD
) && (LCTF_TYPE_ISPARENT (input
, id
)))
2510 if (!ctf_assert (output
, parents
[input_num
] <= ninputs
))
2512 input
= inputs
[parents
[input_num
]];
2513 input_num
= parents
[input_num
];
2516 hval
= ctf_dynhash_lookup (od
->cd_type_hashes
,
2517 CTF_DEDUP_GID (output
, input_num
, id
));
2519 if (!ctf_assert (output
, hval
&& td
->cd_output_emission_hashes
))
2522 /* If this type is a conflicted tagged structure, union, or forward,
2523 substitute a synthetic forward instead, emitting it if need be. Only do
2524 this if the target is in the parent dict: if it's in the child dict, we can
2525 just point straight at the thing itself. Of course, we might be looking in
2526 the child dict right now and not find it and have to look in the parent, so
2527 we have to do this check twice. */
2529 emitted_forward
= ctf_dedup_maybe_synthesize_forward (output
, target
,
2531 switch (emitted_forward
)
2533 case 0: /* No forward needed. */
2536 ctf_set_errno (err_fp
, ctf_errno (output
));
2537 ctf_err_warn (err_fp
, 0, 0, _("cannot add synthetic forward for type "
2538 "%i/%lx"), input_num
, id
);
2541 return emitted_forward
;
2544 ctf_dprintf ("Looking up %i/%lx, hash %s, in target\n", input_num
, id
, hval
);
2546 target_id
= ctf_dynhash_lookup (td
->cd_output_emission_hashes
, hval
);
2549 /* Must be in the parent, so this must be a child, and they must not be
2551 ctf_dprintf ("Checking shared parent for target\n");
2552 if (!ctf_assert (output
, (target
!= output
)
2553 && (target
->ctf_flags
& LCTF_CHILD
)))
2556 target_id
= ctf_dynhash_lookup (od
->cd_output_emission_hashes
, hval
);
2558 emitted_forward
= ctf_dedup_maybe_synthesize_forward (output
, output
,
2560 switch (emitted_forward
)
2562 case 0: /* No forward needed. */
2565 ctf_err_warn (err_fp
, 0, ctf_errno (output
),
2566 _("cannot add synthetic forward for type %i/%lx"),
2568 return ctf_set_errno (err_fp
, ctf_errno (output
));
2570 return emitted_forward
;
2573 if (!ctf_assert (output
, target_id
))
2575 return (ctf_id_t
) (uintptr_t) target_id
;
2578 /* Emit a single deduplicated TYPE with the given HVAL, located in a given
2579 INPUT, with the given (G)ID, into the shared OUTPUT or a
2580 possibly-newly-created per-CU dict. All the types this type depends upon
2581 have already been emitted. (This type itself may also have been emitted.)
2583 If the ARG is 1, this is a CU-mapped deduplication round mapping many
2584 ctf_dict_t's into precisely one: conflicting types should be marked
2585 non-root-visible. If the ARG is 0, conflicting types go into per-CU
2586 dictionaries stored in the input's ctf_dedup.cd_output: otherwise, everything
2587 is emitted directly into the output. No struct/union members are emitted.
2589 Optimization opportunity: trace the ancestry of non-root-visible types and
2590 elide all that neither have a root-visible type somewhere towards their root,
2591 nor have the type visible via any other route (the function info section,
2592 data object section, backtrace section etc). */
2595 ctf_dedup_emit_type (const char *hval
, ctf_dict_t
*output
, ctf_dict_t
**inputs
,
2596 uint32_t ninputs
, uint32_t *parents
, int already_visited
,
2597 ctf_dict_t
*input
, ctf_id_t type
, void *id
, int depth
,
2600 ctf_dedup_t
*d
= &output
->ctf_dedup
;
2601 int kind
= ctf_type_kind_unsliced (input
, type
);
2603 ctf_dict_t
*target
= output
;
2604 ctf_dict_t
*real_input
;
2605 const ctf_type_t
*tp
;
2606 int input_num
= CTF_DEDUP_GID_TO_INPUT (id
);
2607 int output_num
= (uint32_t) -1; /* 'shared' */
2608 int cu_mapped
= *(int *)arg
;
2612 ctf_next_t
*i
= NULL
;
2615 ctf_id_t maybe_dup
= 0;
2617 const char *errtype
;
2618 int emission_hashed
= 0;
2620 /* We don't want to re-emit something we've already emitted. */
2622 if (already_visited
)
2625 ctf_dprintf ("%i: Emitting type with hash %s from %s: determining target\n",
2626 depth
, hval
, ctf_link_input_name (input
));
2628 /* Conflicting types go into a per-CU output dictionary, unless this is a
2629 CU-mapped run. The import is not refcounted, since it goes into the
2630 ctf_link_outputs dict of the output that is its parent. */
2631 is_conflicting
= ctf_dynset_exists (d
->cd_conflicting_types
, hval
, NULL
);
2633 if (is_conflicting
&& !cu_mapped
)
2635 ctf_dprintf ("%i: Type %s in %i/%lx is conflicted: "
2636 "inserting into per-CU target.\n",
2637 depth
, hval
, input_num
, type
);
2639 if (input
->ctf_dedup
.cd_output
)
2640 target
= input
->ctf_dedup
.cd_output
;
2645 if ((target
= ctf_create (&err
)) == NULL
)
2647 ctf_err_warn (output
, 0, err
,
2648 _("cannot create per-CU CTF archive for CU %s"),
2649 ctf_link_input_name (input
));
2650 return ctf_set_errno (output
, err
);
2653 ctf_import_unref (target
, output
);
2654 if (ctf_cuname (input
) != NULL
)
2655 ctf_cuname_set (target
, ctf_cuname (input
));
2657 ctf_cuname_set (target
, "unnamed-CU");
2658 ctf_parent_name_set (target
, _CTF_SECTION
);
2660 input
->ctf_dedup
.cd_output
= target
;
2662 output_num
= input_num
;
2666 if ((tp
= ctf_lookup_by_id (&real_input
, type
)) == NULL
)
2668 ctf_err_warn (output
, 0, ctf_errno (input
),
2669 _("%s: lookup failure for type %lx"),
2670 ctf_link_input_name (real_input
), type
);
2671 return ctf_set_errno (output
, ctf_errno (input
));
2674 name
= ctf_strraw (real_input
, tp
->ctt_name
);
2676 /* Hide conflicting types, if we were asked to: also hide if a type with this
2677 name already exists and is not a forward. */
2678 if (cu_mapped
&& is_conflicting
)
2681 && (maybe_dup
= ctf_lookup_by_rawname (target
, kind
, name
)) != 0)
2683 if (ctf_type_kind (target
, maybe_dup
) != CTF_K_FORWARD
)
2687 ctf_dprintf ("%i: Emitting type with hash %s (%s), into target %i/%p\n",
2688 depth
, hval
, name
? name
: "", input_num
, (void *) target
);
2690 if (!target
->ctf_dedup
.cd_output_emission_hashes
)
2691 if ((target
->ctf_dedup
.cd_output_emission_hashes
2692 = ctf_dynhash_create (ctf_hash_string
, ctf_hash_eq_string
,
2693 NULL
, NULL
)) == NULL
)
2696 if (!target
->ctf_dedup
.cd_output_emission_conflicted_forwards
)
2697 if ((target
->ctf_dedup
.cd_output_emission_conflicted_forwards
2698 = ctf_dynhash_create (ctf_hash_string
, ctf_hash_eq_string
,
2699 NULL
, NULL
)) == NULL
)
2705 /* These are types that CTF cannot encode, marked as such by the compile.
2706 We intentionally do not re-emit these. */
2710 /* This will do nothing if the type to which this forwards already exists,
2711 and will be replaced with such a type if it appears later. */
2713 errtype
= _("forward");
2714 if ((new_type
= ctf_add_forward (target
, isroot
, name
,
2715 ctf_type_kind_forwarded (input
, type
)))
2722 errtype
= _("float/int");
2723 if (ctf_type_encoding (input
, type
, &ep
) < 0)
2724 goto err_input
; /* errno is set for us. */
2725 if ((new_type
= ctf_add_encoded (target
, isroot
, name
, &ep
, kind
))
2733 errtype
= _("enum");
2734 if ((new_type
= ctf_add_enum (target
, isroot
, name
)) == CTF_ERR
)
2735 goto err_input
; /* errno is set for us. */
2737 while ((name
= ctf_enum_next (input
, type
, &i
, &val
)) != NULL
)
2739 if (ctf_add_enumerator (target
, new_type
, name
, val
) < 0)
2741 ctf_err_warn (target
, 0, ctf_errno (target
),
2742 _("%s (%i): cannot add enumeration value %s "
2743 "from input type %lx"),
2744 ctf_link_input_name (input
), input_num
, name
,
2746 ctf_next_destroy (i
);
2747 return ctf_set_errno (output
, ctf_errno (target
));
2750 if (ctf_errno (input
) != ECTF_NEXT_END
)
2756 errtype
= _("typedef");
2758 ref
= ctf_type_reference (input
, type
);
2759 if ((ref
= ctf_dedup_id_to_target (output
, target
, inputs
, ninputs
,
2760 parents
, input
, input_num
,
2762 goto err_input
; /* errno is set for us. */
2764 if ((new_type
= ctf_add_typedef (target
, isroot
, name
, ref
)) == CTF_ERR
)
2765 goto err_target
; /* errno is set for us. */
2768 case CTF_K_VOLATILE
:
2770 case CTF_K_RESTRICT
:
2772 errtype
= _("pointer or cvr-qual");
2774 ref
= ctf_type_reference (input
, type
);
2775 if ((ref
= ctf_dedup_id_to_target (output
, target
, inputs
, ninputs
,
2776 parents
, input
, input_num
,
2778 goto err_input
; /* errno is set for us. */
2780 if ((new_type
= ctf_add_reftype (target
, isroot
, ref
, kind
)) == CTF_ERR
)
2781 goto err_target
; /* errno is set for us. */
2785 errtype
= _("slice");
2787 if (ctf_type_encoding (input
, type
, &ep
) < 0)
2788 goto err_input
; /* errno is set for us. */
2790 ref
= ctf_type_reference (input
, type
);
2791 if ((ref
= ctf_dedup_id_to_target (output
, target
, inputs
, ninputs
,
2792 parents
, input
, input_num
,
2796 if ((new_type
= ctf_add_slice (target
, isroot
, ref
, &ep
)) == CTF_ERR
)
2804 errtype
= _("array info");
2805 if (ctf_array_info (input
, type
, &ar
) < 0)
2808 ar
.ctr_contents
= ctf_dedup_id_to_target (output
, target
, inputs
,
2809 ninputs
, parents
, input
,
2810 input_num
, ar
.ctr_contents
);
2811 ar
.ctr_index
= ctf_dedup_id_to_target (output
, target
, inputs
, ninputs
,
2812 parents
, input
, input_num
,
2815 if (ar
.ctr_contents
== CTF_ERR
|| ar
.ctr_index
== CTF_ERR
)
2818 if ((new_type
= ctf_add_array (target
, isroot
, &ar
)) == CTF_ERR
)
2824 case CTF_K_FUNCTION
:
2830 errtype
= _("function");
2831 if (ctf_func_type_info (input
, type
, &fi
) < 0)
2834 fi
.ctc_return
= ctf_dedup_id_to_target (output
, target
, inputs
, ninputs
,
2835 parents
, input
, input_num
,
2837 if (fi
.ctc_return
== CTF_ERR
)
2840 if ((args
= calloc (fi
.ctc_argc
, sizeof (ctf_id_t
))) == NULL
)
2842 ctf_set_errno (input
, ENOMEM
);
2846 errtype
= _("function args");
2847 if (ctf_func_type_args (input
, type
, fi
.ctc_argc
, args
) < 0)
2853 for (j
= 0; j
< fi
.ctc_argc
; j
++)
2855 args
[j
] = ctf_dedup_id_to_target (output
, target
, inputs
, ninputs
,
2856 parents
, input
, input_num
,
2858 if (args
[j
] == CTF_ERR
)
2862 if ((new_type
= ctf_add_function (target
, isroot
,
2863 &fi
, args
)) == CTF_ERR
)
2875 size_t size
= ctf_type_size (input
, type
);
2877 /* Insert the structure itself, so other types can refer to it. */
2879 errtype
= _("structure/union");
2880 if (kind
== CTF_K_STRUCT
)
2881 new_type
= ctf_add_struct_sized (target
, isroot
, name
, size
);
2883 new_type
= ctf_add_union_sized (target
, isroot
, name
, size
);
2885 if (new_type
== CTF_ERR
)
2888 out_id
= CTF_DEDUP_GID (output
, output_num
, new_type
);
2889 ctf_dprintf ("%i: Noting need to emit members of %p -> %p\n", depth
,
2891 /* Record the need to emit the members of this structure later. */
2892 if (ctf_dynhash_insert (d
->cd_emission_struct_members
, id
, out_id
) < 0)
2894 ctf_set_errno (target
, errno
);
2900 ctf_err_warn (output
, 0, ECTF_CORRUPT
, _("%s: unknown type kind for "
2902 ctf_link_input_name (input
), type
);
2903 return ctf_set_errno (output
, ECTF_CORRUPT
);
2906 if (!emission_hashed
2908 && ctf_dynhash_cinsert (target
->ctf_dedup
.cd_output_emission_hashes
,
2909 hval
, (void *) (uintptr_t) new_type
) < 0)
2911 ctf_err_warn (output
, 0, ENOMEM
, _("out of memory tracking deduplicated "
2912 "global type IDs"));
2913 return ctf_set_errno (output
, ENOMEM
);
2916 if (!emission_hashed
&& new_type
!= 0)
2917 ctf_dprintf ("%i: Inserted %s, %i/%lx -> %lx into emission hash for "
2918 "target %p (%s)\n", depth
, hval
, input_num
, type
, new_type
,
2919 (void *) target
, ctf_link_input_name (target
));
2924 ctf_err_warn (output
, 0, ENOMEM
, _("out of memory creating emission-tracking "
2926 return ctf_set_errno (output
, ENOMEM
);
2929 ctf_err_warn (output
, 0, ctf_errno (input
),
2930 _("%s (%i): while emitting deduplicated %s, error getting "
2931 "input type %lx"), ctf_link_input_name (input
),
2932 input_num
, errtype
, type
);
2933 return ctf_set_errno (output
, ctf_errno (input
));
2935 ctf_err_warn (output
, 0, ctf_errno (target
),
2936 _("%s (%i): while emitting deduplicated %s, error emitting "
2937 "target type from input type %lx"),
2938 ctf_link_input_name (input
), input_num
,
2940 return ctf_set_errno (output
, ctf_errno (target
));
2943 /* Traverse the cd_emission_struct_members and emit the members of all
2944 structures and unions. All other types are emitted and complete by this
2948 ctf_dedup_emit_struct_members (ctf_dict_t
*output
, ctf_dict_t
**inputs
,
2949 uint32_t ninputs
, uint32_t *parents
)
2951 ctf_dedup_t
*d
= &output
->ctf_dedup
;
2952 ctf_next_t
*i
= NULL
;
2953 void *input_id
, *target_id
;
2955 ctf_dict_t
*err_fp
, *input_fp
;
2959 while ((err
= ctf_dynhash_next (d
->cd_emission_struct_members
, &i
,
2960 &input_id
, &target_id
)) == 0)
2962 ctf_next_t
*j
= NULL
;
2964 uint32_t target_num
;
2965 ctf_id_t input_type
, target_type
;
2970 input_num
= CTF_DEDUP_GID_TO_INPUT (input_id
);
2971 input_fp
= inputs
[input_num
];
2972 input_type
= CTF_DEDUP_GID_TO_TYPE (input_id
);
2974 /* The output is either -1 (for the shared, parent output dict) or the
2975 number of the corresponding input. */
2976 target_num
= CTF_DEDUP_GID_TO_INPUT (target_id
);
2977 if (target_num
== (uint32_t) -1)
2981 target
= inputs
[target_num
]->ctf_dedup
.cd_output
;
2982 if (!ctf_assert (output
, target
))
2985 err_type
= input_type
;
2989 target_type
= CTF_DEDUP_GID_TO_TYPE (target_id
);
2991 while ((offset
= ctf_member_next (input_fp
, input_type
, &j
, &name
,
2992 &membtype
, 0)) >= 0)
2995 err_type
= target_type
;
2996 if ((membtype
= ctf_dedup_id_to_target (output
, target
, inputs
,
2997 ninputs
, parents
, input_fp
,
2999 membtype
)) == CTF_ERR
)
3001 ctf_next_destroy (j
);
3007 #ifdef ENABLE_LIBCTF_HASH_DEBUGGING
3008 ctf_dprintf ("Emitting %s, offset %zi\n", name
, offset
);
3010 if (ctf_add_member_offset (target
, target_type
, name
,
3011 membtype
, offset
) < 0)
3013 ctf_next_destroy (j
);
3017 if (ctf_errno (input_fp
) != ECTF_NEXT_END
)
3019 err
= ctf_errno (input_fp
);
3020 ctf_next_destroy (i
);
3024 if (err
!= ECTF_NEXT_END
)
3029 ctf_next_destroy (i
);
3030 ctf_err_warn (output
, 0, ctf_errno (err_fp
),
3031 _("%s (%i): error emitting members for structure type %lx"),
3032 ctf_link_input_name (input_fp
), input_num
, err_type
);
3033 return ctf_set_errno (output
, ctf_errno (err_fp
));
3035 ctf_err_warn (output
, 0, err
, _("iteration failure emitting "
3036 "structure members"));
3037 return ctf_set_errno (output
, err
);
3040 /* Emit deduplicated types into the outputs. The shared type repository is
3041 OUTPUT, on which the ctf_dedup function must have already been called. The
3042 PARENTS array contains the INPUTS index of the parent dict for every child
3043 dict at the corresponding index in the INPUTS (for non-child dicts, the value
3046 Return an array of fps with content emitted into them (starting with OUTPUT,
3047 which is the parent of all others, then all the newly-generated outputs).
3049 If CU_MAPPED is set, this is a first pass for a link with a non-empty CU
3050 mapping: only one output will result. */
3053 ctf_dedup_emit (ctf_dict_t
*output
, ctf_dict_t
**inputs
, uint32_t ninputs
,
3054 uint32_t *parents
, uint32_t *noutputs
, int cu_mapped
)
3056 size_t num_outputs
= 1; /* Always at least one output: us. */
3057 ctf_dict_t
**outputs
;
3061 ctf_dprintf ("Triggering emission.\n");
3062 if (ctf_dedup_walk_output_mapping (output
, inputs
, ninputs
, parents
,
3063 ctf_dedup_emit_type
, &cu_mapped
) < 0)
3064 return NULL
; /* errno is set for us. */
3066 ctf_dprintf ("Populating struct members.\n");
3067 if (ctf_dedup_emit_struct_members (output
, inputs
, ninputs
, parents
) < 0)
3068 return NULL
; /* errno is set for us. */
3070 for (i
= 0; i
< ninputs
; i
++)
3072 if (inputs
[i
]->ctf_dedup
.cd_output
)
3076 if (!ctf_assert (output
, !cu_mapped
|| (cu_mapped
&& num_outputs
== 1)))
3079 if ((outputs
= calloc (num_outputs
, sizeof (ctf_dict_t
*))) == NULL
)
3081 ctf_err_warn (output
, 0, ENOMEM
,
3082 _("out of memory allocating link outputs array"));
3083 ctf_set_errno (output
, ENOMEM
);
3086 *noutputs
= num_outputs
;
3090 output
->ctf_refcnt
++;
3093 for (i
= 0; i
< ninputs
; i
++)
3095 if (inputs
[i
]->ctf_dedup
.cd_output
)
3097 *walk
= inputs
[i
]->ctf_dedup
.cd_output
;
3098 inputs
[i
]->ctf_dedup
.cd_output
= NULL
;
3106 /* Determine what type SRC_FP / SRC_TYPE was emitted as in the FP, which
3107 must be the shared dict or have it as a parent: return 0 if none. The SRC_FP
3108 must be a past input to ctf_dedup. */
3111 ctf_dedup_type_mapping (ctf_dict_t
*fp
, ctf_dict_t
*src_fp
, ctf_id_t src_type
)
3113 ctf_dict_t
*output
= NULL
;
3121 /* It is an error (an internal error in the caller, in ctf-link.c) to call
3122 this with an FP that is not a per-CU output or shared output dict, or with
3123 a SRC_FP that was not passed to ctf_dedup as an input; it is an internal
3124 error in ctf-dedup for the type passed not to have been hashed, though if
3125 the src_fp is a child dict and the type is not a child type, it will have
3126 been hashed under the GID corresponding to the parent. */
3128 if (fp
->ctf_dedup
.cd_type_hashes
!= NULL
)
3130 else if (fp
->ctf_parent
&& fp
->ctf_parent
->ctf_dedup
.cd_type_hashes
!= NULL
)
3131 output
= fp
->ctf_parent
;
3134 ctf_set_errno (fp
, ECTF_INTERNAL
);
3135 ctf_err_warn (fp
, 0, ECTF_INTERNAL
,
3136 _("dict %p passed to ctf_dedup_type_mapping is not a "
3137 "deduplicated output"), (void *) fp
);
3141 if (src_fp
->ctf_parent
&& ctf_type_isparent (src_fp
, src_type
))
3142 src_fp
= src_fp
->ctf_parent
;
3144 d
= &output
->ctf_dedup
;
3146 found
= ctf_dynhash_lookup_kv (d
->cd_input_nums
, src_fp
, NULL
, &num_ptr
);
3147 if (!ctf_assert (output
, found
!= 0))
3148 return CTF_ERR
; /* errno is set for us. */
3149 input_num
= (uintptr_t) num_ptr
;
3151 hval
= ctf_dynhash_lookup (d
->cd_type_hashes
,
3152 CTF_DEDUP_GID (output
, input_num
, src_type
));
3154 if (!ctf_assert (output
, hval
!= NULL
))
3155 return CTF_ERR
; /* errno is set for us. */
3157 /* The emission hashes may be unset if this dict was created after
3158 deduplication to house variables or other things that would conflict if
3159 stored in the shared dict. */
3160 if (fp
->ctf_dedup
.cd_output_emission_hashes
)
3161 if (ctf_dynhash_lookup_kv (fp
->ctf_dedup
.cd_output_emission_hashes
, hval
,
3163 return (ctf_id_t
) (uintptr_t) type_ptr
;
3167 ctf_dict_t
*pfp
= fp
->ctf_parent
;
3168 if (pfp
->ctf_dedup
.cd_output_emission_hashes
)
3169 if (ctf_dynhash_lookup_kv (pfp
->ctf_dedup
.cd_output_emission_hashes
,
3170 hval
, NULL
, &type_ptr
))
3171 return (ctf_id_t
) (uintptr_t) type_ptr
;