111
|
1 /* Pass for parsing functions with multiple target attributes.
|
|
2
|
|
3 Contributed by Evgeny Stupachenko <evstupac@gmail.com>
|
|
4
|
145
|
5 Copyright (C) 2015-2020 Free Software Foundation, Inc.
|
111
|
6
|
|
7 This file is part of GCC.
|
|
8
|
|
9 GCC is free software; you can redistribute it and/or modify it under
|
|
10 the terms of the GNU General Public License as published by the Free
|
|
11 Software Foundation; either version 3, or (at your option) any later
|
|
12 version.
|
|
13
|
|
14 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
|
|
15 WARRANTY; without even the implied warranty of MERCHANTABILITY or
|
|
16 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
|
17 for more details.
|
|
18
|
|
19 You should have received a copy of the GNU General Public License
|
|
20 along with GCC; see the file COPYING3. If not see
|
|
21 <http://www.gnu.org/licenses/>. */
|
|
22
|
|
23 #include "config.h"
|
|
24 #include "system.h"
|
|
25 #include "coretypes.h"
|
|
26 #include "backend.h"
|
|
27 #include "tree.h"
|
|
28 #include "stringpool.h"
|
|
29 #include "gimple.h"
|
|
30 #include "diagnostic-core.h"
|
|
31 #include "gimple-ssa.h"
|
|
32 #include "cgraph.h"
|
|
33 #include "tree-pass.h"
|
|
34 #include "target.h"
|
|
35 #include "attribs.h"
|
|
36 #include "pretty-print.h"
|
|
37 #include "gimple-iterator.h"
|
|
38 #include "gimple-walk.h"
|
131
|
39 #include "tree-inline.h"
|
|
40 #include "intl.h"
|
111
|
41
|
|
42 /* Walker callback that replaces all FUNCTION_DECL of a function that's
|
|
43 going to be versioned. */
|
|
44
|
|
45 static tree
|
|
46 replace_function_decl (tree *op, int *walk_subtrees, void *data)
|
|
47 {
|
|
48 struct walk_stmt_info *wi = (struct walk_stmt_info *) data;
|
|
49 cgraph_function_version_info *info = (cgraph_function_version_info *)wi->info;
|
|
50
|
|
51 if (TREE_CODE (*op) == FUNCTION_DECL
|
|
52 && info->this_node->decl == *op)
|
|
53 {
|
|
54 *op = info->dispatcher_resolver;
|
|
55 *walk_subtrees = 0;
|
|
56 }
|
|
57
|
|
58 return NULL;
|
|
59 }
|
|
60
|
|
61 /* If the call in NODE has multiple target attribute with multiple fields,
|
|
62 replace it with dispatcher call and create dispatcher (once). */
|
|
63
|
|
64 static void
|
|
65 create_dispatcher_calls (struct cgraph_node *node)
|
|
66 {
|
|
67 ipa_ref *ref;
|
|
68
|
|
69 if (!DECL_FUNCTION_VERSIONED (node->decl)
|
|
70 || !is_function_default_version (node->decl))
|
|
71 return;
|
|
72
|
|
73 if (!targetm.has_ifunc_p ())
|
|
74 {
|
|
75 error_at (DECL_SOURCE_LOCATION (node->decl),
|
145
|
76 "the call requires %<ifunc%>, which is not"
|
111
|
77 " supported by this target");
|
|
78 return;
|
|
79 }
|
|
80 else if (!targetm.get_function_versions_dispatcher)
|
|
81 {
|
|
82 error_at (DECL_SOURCE_LOCATION (node->decl),
|
|
83 "target does not support function version dispatcher");
|
|
84 return;
|
|
85 }
|
|
86
|
|
87 tree idecl = targetm.get_function_versions_dispatcher (node->decl);
|
|
88 if (!idecl)
|
|
89 {
|
|
90 error_at (DECL_SOURCE_LOCATION (node->decl),
|
131
|
91 "default %<target_clones%> attribute was not set");
|
111
|
92 return;
|
|
93 }
|
|
94
|
|
95 cgraph_node *inode = cgraph_node::get (idecl);
|
|
96 gcc_assert (inode);
|
|
97 tree resolver_decl = targetm.generate_version_dispatcher_body (inode);
|
|
98
|
|
99 /* Update aliases. */
|
|
100 inode->alias = true;
|
|
101 inode->alias_target = resolver_decl;
|
|
102 if (!inode->analyzed)
|
|
103 inode->resolve_alias (cgraph_node::get (resolver_decl));
|
|
104
|
|
105 auto_vec<cgraph_edge *> edges_to_redirect;
|
145
|
106 /* We need to capture the references by value rather than just pointers to them
|
|
107 and remove them right away, as removing them later would invalidate what
|
|
108 some other reference pointers point to. */
|
|
109 auto_vec<ipa_ref> references_to_redirect;
|
111
|
110
|
145
|
111 while (node->iterate_referring (0, ref))
|
|
112 {
|
|
113 references_to_redirect.safe_push (*ref);
|
|
114 ref->remove_reference ();
|
|
115 }
|
111
|
116
|
|
117 /* We need to remember NEXT_CALLER as it could be modified in the loop. */
|
|
118 for (cgraph_edge *e = node->callers; e ; e = e->next_caller)
|
|
119 edges_to_redirect.safe_push (e);
|
|
120
|
|
121 if (!edges_to_redirect.is_empty () || !references_to_redirect.is_empty ())
|
|
122 {
|
|
123 /* Redirect edges. */
|
|
124 unsigned i;
|
|
125 cgraph_edge *e;
|
|
126 FOR_EACH_VEC_ELT (edges_to_redirect, i, e)
|
|
127 {
|
|
128 e->redirect_callee (inode);
|
145
|
129 cgraph_edge::redirect_call_stmt_to_callee (e);
|
111
|
130 }
|
|
131
|
|
132 /* Redirect references. */
|
|
133 FOR_EACH_VEC_ELT (references_to_redirect, i, ref)
|
|
134 {
|
|
135 if (ref->use == IPA_REF_ADDR)
|
|
136 {
|
|
137 struct walk_stmt_info wi;
|
|
138 memset (&wi, 0, sizeof (wi));
|
|
139 wi.info = (void *)node->function_version ();
|
|
140
|
|
141 if (dyn_cast<varpool_node *> (ref->referring))
|
|
142 {
|
|
143 hash_set<tree> visited_nodes;
|
|
144 walk_tree (&DECL_INITIAL (ref->referring->decl),
|
|
145 replace_function_decl, &wi, &visited_nodes);
|
|
146 }
|
|
147 else
|
|
148 {
|
|
149 gimple_stmt_iterator it = gsi_for_stmt (ref->stmt);
|
|
150 if (ref->referring->decl != resolver_decl)
|
|
151 walk_gimple_stmt (&it, NULL, replace_function_decl, &wi);
|
|
152 }
|
131
|
153
|
|
154 symtab_node *source = ref->referring;
|
|
155 source->create_reference (inode, IPA_REF_ADDR);
|
|
156 }
|
|
157 else if (ref->use == IPA_REF_ALIAS)
|
|
158 {
|
|
159 symtab_node *source = ref->referring;
|
|
160 source->create_reference (inode, IPA_REF_ALIAS);
|
145
|
161 if (inode->get_comdat_group ())
|
|
162 source->add_to_same_comdat_group (inode);
|
111
|
163 }
|
|
164 else
|
|
165 gcc_unreachable ();
|
|
166 }
|
|
167 }
|
|
168
|
|
169 symtab->change_decl_assembler_name (node->decl,
|
145
|
170 clone_function_name_numbered (
|
|
171 node->decl, "default"));
|
131
|
172
|
|
173 /* FIXME: copy of cgraph_node::make_local that should be cleaned up
|
|
174 in next stage1. */
|
|
175 node->make_decl_local ();
|
|
176 node->set_section (NULL);
|
|
177 node->set_comdat_group (NULL);
|
|
178 node->externally_visible = false;
|
|
179 node->forced_by_abi = false;
|
|
180 node->set_section (NULL);
|
|
181 node->unique_name = ((node->resolution == LDPR_PREVAILING_DEF_IRONLY
|
|
182 || node->resolution == LDPR_PREVAILING_DEF_IRONLY_EXP)
|
|
183 && !flag_incremental_link);
|
|
184 node->resolution = LDPR_PREVAILING_DEF_IRONLY;
|
|
185
|
|
186 DECL_ARTIFICIAL (node->decl) = 1;
|
|
187 node->force_output = true;
|
111
|
188 }
|
|
189
|
|
190 /* Return length of attribute names string,
|
|
191 if arglist chain > 1, -1 otherwise. */
|
|
192
|
|
193 static int
|
|
194 get_attr_len (tree arglist)
|
|
195 {
|
|
196 tree arg;
|
|
197 int str_len_sum = 0;
|
|
198 int argnum = 0;
|
|
199
|
|
200 for (arg = arglist; arg; arg = TREE_CHAIN (arg))
|
|
201 {
|
|
202 const char *str = TREE_STRING_POINTER (TREE_VALUE (arg));
|
|
203 size_t len = strlen (str);
|
|
204 str_len_sum += len + 1;
|
|
205 for (const char *p = strchr (str, ','); p; p = strchr (p + 1, ','))
|
|
206 argnum++;
|
|
207 argnum++;
|
|
208 }
|
|
209 if (argnum <= 1)
|
|
210 return -1;
|
|
211 return str_len_sum;
|
|
212 }
|
|
213
|
|
214 /* Create string with attributes separated by comma.
|
|
215 Return number of attributes. */
|
|
216
|
|
217 static int
|
|
218 get_attr_str (tree arglist, char *attr_str)
|
|
219 {
|
|
220 tree arg;
|
|
221 size_t str_len_sum = 0;
|
|
222 int argnum = 0;
|
|
223
|
|
224 for (arg = arglist; arg; arg = TREE_CHAIN (arg))
|
|
225 {
|
|
226 const char *str = TREE_STRING_POINTER (TREE_VALUE (arg));
|
|
227 size_t len = strlen (str);
|
|
228 for (const char *p = strchr (str, ','); p; p = strchr (p + 1, ','))
|
|
229 argnum++;
|
|
230 memcpy (attr_str + str_len_sum, str, len);
|
|
231 attr_str[str_len_sum + len] = TREE_CHAIN (arg) ? ',' : '\0';
|
|
232 str_len_sum += len + 1;
|
|
233 argnum++;
|
|
234 }
|
|
235 return argnum;
|
|
236 }
|
|
237
|
|
238 /* Return number of attributes separated by comma and put them into ARGS.
|
145
|
239 If there is no DEFAULT attribute return -1.
|
|
240 If there is an empty string in attribute return -2.
|
|
241 If there are multiple DEFAULT attributes return -3.
|
|
242 */
|
111
|
243
|
|
244 static int
|
131
|
245 separate_attrs (char *attr_str, char **attrs, int attrnum)
|
111
|
246 {
|
|
247 int i = 0;
|
131
|
248 int default_count = 0;
|
111
|
249
|
|
250 for (char *attr = strtok (attr_str, ",");
|
|
251 attr != NULL; attr = strtok (NULL, ","))
|
|
252 {
|
|
253 if (strcmp (attr, "default") == 0)
|
|
254 {
|
131
|
255 default_count++;
|
111
|
256 continue;
|
|
257 }
|
|
258 attrs[i++] = attr;
|
|
259 }
|
131
|
260 if (default_count == 0)
|
111
|
261 return -1;
|
145
|
262 else if (default_count > 1)
|
|
263 return -3;
|
131
|
264 else if (i + default_count < attrnum)
|
|
265 return -2;
|
|
266
|
111
|
267 return i;
|
|
268 }
|
|
269
|
|
270 /* Return true if symbol is valid in assembler name. */
|
|
271
|
|
272 static bool
|
|
273 is_valid_asm_symbol (char c)
|
|
274 {
|
|
275 if ('a' <= c && c <= 'z')
|
|
276 return true;
|
|
277 if ('A' <= c && c <= 'Z')
|
|
278 return true;
|
|
279 if ('0' <= c && c <= '9')
|
|
280 return true;
|
|
281 if (c == '_')
|
|
282 return true;
|
|
283 return false;
|
|
284 }
|
|
285
|
|
286 /* Replace all not valid assembler symbols with '_'. */
|
|
287
|
|
288 static void
|
|
289 create_new_asm_name (char *old_asm_name, char *new_asm_name)
|
|
290 {
|
|
291 int i;
|
|
292 int old_name_len = strlen (old_asm_name);
|
|
293
|
|
294 /* Replace all not valid assembler symbols with '_'. */
|
|
295 for (i = 0; i < old_name_len; i++)
|
|
296 if (!is_valid_asm_symbol (old_asm_name[i]))
|
|
297 new_asm_name[i] = '_';
|
|
298 else
|
|
299 new_asm_name[i] = old_asm_name[i];
|
|
300 new_asm_name[old_name_len] = '\0';
|
|
301 }
|
|
302
|
|
303 /* Creates target clone of NODE. */
|
|
304
|
|
305 static cgraph_node *
|
145
|
306 create_target_clone (cgraph_node *node, bool definition, char *name,
|
|
307 tree attributes)
|
111
|
308 {
|
|
309 cgraph_node *new_node;
|
|
310
|
|
311 if (definition)
|
|
312 {
|
|
313 new_node = node->create_version_clone_with_body (vNULL, NULL,
|
145
|
314 NULL, NULL,
|
|
315 NULL, name, attributes);
|
|
316 if (new_node == NULL)
|
|
317 return NULL;
|
111
|
318 new_node->force_output = true;
|
|
319 }
|
|
320 else
|
|
321 {
|
|
322 tree new_decl = copy_node (node->decl);
|
|
323 new_node = cgraph_node::get_create (new_decl);
|
145
|
324 DECL_ATTRIBUTES (new_decl) = attributes;
|
111
|
325 /* Generate a new name for the new version. */
|
|
326 symtab->change_decl_assembler_name (new_node->decl,
|
145
|
327 clone_function_name_numbered (
|
|
328 node->decl, name));
|
111
|
329 }
|
|
330 return new_node;
|
|
331 }
|
|
332
|
|
333 /* If the function in NODE has multiple target attributes
|
|
334 create the appropriate clone for each valid target attribute. */
|
|
335
|
|
336 static bool
|
|
337 expand_target_clones (struct cgraph_node *node, bool definition)
|
|
338 {
|
|
339 int i;
|
|
340 /* Parsing target attributes separated by comma. */
|
|
341 tree attr_target = lookup_attribute ("target_clones",
|
|
342 DECL_ATTRIBUTES (node->decl));
|
|
343 /* No targets specified. */
|
|
344 if (!attr_target)
|
|
345 return false;
|
|
346
|
|
347 tree arglist = TREE_VALUE (attr_target);
|
|
348 int attr_len = get_attr_len (arglist);
|
|
349
|
|
350 /* No need to clone for 1 target attribute. */
|
|
351 if (attr_len == -1)
|
|
352 {
|
|
353 warning_at (DECL_SOURCE_LOCATION (node->decl),
|
145
|
354 0, "single %<target_clones%> attribute is ignored");
|
131
|
355 return false;
|
|
356 }
|
|
357
|
|
358 if (node->definition
|
145
|
359 && (node->alias || !tree_versionable_function_p (node->decl)))
|
131
|
360 {
|
|
361 auto_diagnostic_group d;
|
|
362 error_at (DECL_SOURCE_LOCATION (node->decl),
|
|
363 "clones for %<target_clones%> attribute cannot be created");
|
|
364 const char *reason = NULL;
|
|
365 if (lookup_attribute ("noclone", DECL_ATTRIBUTES (node->decl)))
|
|
366 reason = G_("function %q+F can never be copied "
|
|
367 "because it has %<noclone%> attribute");
|
145
|
368 else if (node->alias)
|
|
369 reason
|
|
370 = "%<target_clones%> cannot be combined with %<alias%> attribute";
|
131
|
371 else
|
|
372 reason = copy_forbidden (DECL_STRUCT_FUNCTION (node->decl));
|
|
373 if (reason)
|
|
374 inform (DECL_SOURCE_LOCATION (node->decl), reason, node->decl);
|
111
|
375 return false;
|
|
376 }
|
|
377
|
|
378 char *attr_str = XNEWVEC (char, attr_len);
|
|
379 int attrnum = get_attr_str (arglist, attr_str);
|
|
380 char **attrs = XNEWVEC (char *, attrnum);
|
|
381
|
131
|
382 attrnum = separate_attrs (attr_str, attrs, attrnum);
|
145
|
383 switch (attrnum)
|
111
|
384 {
|
145
|
385 case -1:
|
111
|
386 error_at (DECL_SOURCE_LOCATION (node->decl),
|
145
|
387 "%<default%> target was not set");
|
|
388 break;
|
|
389 case -2:
|
131
|
390 error_at (DECL_SOURCE_LOCATION (node->decl),
|
|
391 "an empty string cannot be in %<target_clones%> attribute");
|
145
|
392 break;
|
|
393 case -3:
|
|
394 error_at (DECL_SOURCE_LOCATION (node->decl),
|
|
395 "multiple %<default%> targets were set");
|
|
396 break;
|
|
397 default:
|
|
398 break;
|
|
399 }
|
|
400
|
|
401 if (attrnum < 0)
|
|
402 {
|
131
|
403 XDELETEVEC (attrs);
|
|
404 XDELETEVEC (attr_str);
|
|
405 return false;
|
|
406 }
|
111
|
407
|
|
408 cgraph_function_version_info *decl1_v = NULL;
|
|
409 cgraph_function_version_info *decl2_v = NULL;
|
|
410 cgraph_function_version_info *before = NULL;
|
|
411 cgraph_function_version_info *after = NULL;
|
|
412 decl1_v = node->function_version ();
|
|
413 if (decl1_v == NULL)
|
|
414 decl1_v = node->insert_new_function_version ();
|
|
415 before = decl1_v;
|
|
416 DECL_FUNCTION_VERSIONED (node->decl) = 1;
|
|
417
|
|
418 for (i = 0; i < attrnum; i++)
|
|
419 {
|
|
420 char *attr = attrs[i];
|
|
421 char *suffix = XNEWVEC (char, strlen (attr) + 1);
|
|
422
|
|
423 create_new_asm_name (attr, suffix);
|
|
424 /* Create new target clone. */
|
145
|
425 tree attributes = make_attribute ("target", attr,
|
|
426 DECL_ATTRIBUTES (node->decl));
|
|
427
|
|
428 cgraph_node *new_node = create_target_clone (node, definition, suffix,
|
|
429 attributes);
|
|
430 if (new_node == NULL)
|
|
431 return false;
|
|
432 new_node->local = false;
|
111
|
433 XDELETEVEC (suffix);
|
|
434
|
|
435 decl2_v = new_node->function_version ();
|
|
436 if (decl2_v != NULL)
|
|
437 continue;
|
|
438 decl2_v = new_node->insert_new_function_version ();
|
|
439
|
|
440 /* Chain decl2_v and decl1_v. All semantically identical versions
|
|
441 will be chained together. */
|
|
442 after = decl2_v;
|
|
443 while (before->next != NULL)
|
|
444 before = before->next;
|
|
445 while (after->prev != NULL)
|
|
446 after = after->prev;
|
|
447
|
|
448 before->next = after;
|
|
449 after->prev = before;
|
|
450 DECL_FUNCTION_VERSIONED (new_node->decl) = 1;
|
|
451 }
|
|
452
|
|
453 XDELETEVEC (attrs);
|
|
454 XDELETEVEC (attr_str);
|
|
455
|
|
456 /* Setting new attribute to initial function. */
|
|
457 tree attributes = make_attribute ("target", "default",
|
|
458 DECL_ATTRIBUTES (node->decl));
|
|
459 DECL_ATTRIBUTES (node->decl) = attributes;
|
145
|
460 node->local = false;
|
|
461 return true;
|
111
|
462 }
|
|
463
|
131
|
464 /* When NODE is a target clone, consider all callees and redirect
|
|
465 to a clone with equal target attributes. That prevents multiple
|
|
466 multi-versioning dispatches and a call-chain can be optimized. */
|
|
467
|
|
468 static void
|
|
469 redirect_to_specific_clone (cgraph_node *node)
|
|
470 {
|
|
471 cgraph_function_version_info *fv = node->function_version ();
|
|
472 if (fv == NULL)
|
|
473 return;
|
|
474
|
|
475 tree attr_target = lookup_attribute ("target", DECL_ATTRIBUTES (node->decl));
|
|
476 if (attr_target == NULL_TREE)
|
|
477 return;
|
|
478
|
|
479 /* We need to remember NEXT_CALLER as it could be modified in the loop. */
|
|
480 for (cgraph_edge *e = node->callees; e ; e = e->next_callee)
|
|
481 {
|
|
482 cgraph_function_version_info *fv2 = e->callee->function_version ();
|
|
483 if (!fv2)
|
|
484 continue;
|
|
485
|
|
486 tree attr_target2 = lookup_attribute ("target",
|
|
487 DECL_ATTRIBUTES (e->callee->decl));
|
|
488
|
|
489 /* Function is not calling proper target clone. */
|
|
490 if (!attribute_list_equal (attr_target, attr_target2))
|
|
491 {
|
|
492 while (fv2->prev != NULL)
|
|
493 fv2 = fv2->prev;
|
|
494
|
|
495 /* Try to find a clone with equal target attribute. */
|
|
496 for (; fv2 != NULL; fv2 = fv2->next)
|
|
497 {
|
|
498 cgraph_node *callee = fv2->this_node;
|
|
499 attr_target2 = lookup_attribute ("target",
|
|
500 DECL_ATTRIBUTES (callee->decl));
|
|
501 if (attribute_list_equal (attr_target, attr_target2))
|
|
502 {
|
|
503 e->redirect_callee (callee);
|
145
|
504 cgraph_edge::redirect_call_stmt_to_callee (e);
|
131
|
505 break;
|
|
506 }
|
|
507 }
|
|
508 }
|
|
509 }
|
|
510 }
|
|
511
|
111
|
512 static unsigned int
|
|
513 ipa_target_clone (void)
|
|
514 {
|
|
515 struct cgraph_node *node;
|
131
|
516 auto_vec<cgraph_node *> to_dispatch;
|
111
|
517
|
|
518 FOR_EACH_FUNCTION (node)
|
131
|
519 if (expand_target_clones (node, node->definition))
|
|
520 to_dispatch.safe_push (node);
|
111
|
521
|
131
|
522 for (unsigned i = 0; i < to_dispatch.length (); i++)
|
|
523 create_dispatcher_calls (to_dispatch[i]);
|
|
524
|
|
525 FOR_EACH_FUNCTION (node)
|
|
526 redirect_to_specific_clone (node);
|
111
|
527
|
|
528 return 0;
|
|
529 }
|
|
530
|
|
531 namespace {
|
|
532
|
|
533 const pass_data pass_data_target_clone =
|
|
534 {
|
|
535 SIMPLE_IPA_PASS, /* type */
|
|
536 "targetclone", /* name */
|
|
537 OPTGROUP_NONE, /* optinfo_flags */
|
|
538 TV_NONE, /* tv_id */
|
|
539 ( PROP_ssa | PROP_cfg ), /* properties_required */
|
|
540 0, /* properties_provided */
|
|
541 0, /* properties_destroyed */
|
|
542 0, /* todo_flags_start */
|
|
543 TODO_update_ssa /* todo_flags_finish */
|
|
544 };
|
|
545
|
|
546 class pass_target_clone : public simple_ipa_opt_pass
|
|
547 {
|
|
548 public:
|
|
549 pass_target_clone (gcc::context *ctxt)
|
|
550 : simple_ipa_opt_pass (pass_data_target_clone, ctxt)
|
|
551 {}
|
|
552
|
|
553 /* opt_pass methods: */
|
|
554 virtual bool gate (function *);
|
|
555 virtual unsigned int execute (function *) { return ipa_target_clone (); }
|
|
556 };
|
|
557
|
|
558 bool
|
|
559 pass_target_clone::gate (function *)
|
|
560 {
|
|
561 return true;
|
|
562 }
|
|
563
|
|
564 } // anon namespace
|
|
565
|
|
566 simple_ipa_opt_pass *
|
|
567 make_pass_target_clone (gcc::context *ctxt)
|
|
568 {
|
|
569 return new pass_target_clone (ctxt);
|
|
570 }
|