1 | /* Interprocedural Identical Code Folding pass |
2 | Copyright (C) 2014-2024 Free Software Foundation, Inc. |
3 | |
4 | Contributed by Jan Hubicka <hubicka@ucw.cz> and Martin Liska <mliska@suse.cz> |
5 | |
6 | This file is part of GCC. |
7 | |
8 | GCC is free software; you can redistribute it and/or modify it under |
9 | the terms of the GNU General Public License as published by the Free |
10 | Software Foundation; either version 3, or (at your option) any later |
11 | version. |
12 | |
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
16 | for more details. |
17 | |
18 | You should have received a copy of the GNU General Public License |
19 | along with GCC; see the file COPYING3. If not see |
20 | <http://www.gnu.org/licenses/>. */ |
21 | |
22 | /* Interprocedural Identical Code Folding for functions and |
23 | read-only variables. |
24 | |
25 | The goal of this transformation is to discover functions and read-only |
26 | variables which do have exactly the same semantics. |
27 | |
28 | In case of functions, |
29 | we could either create a virtual clone or do a simple function wrapper |
30 | that will call equivalent function. If the function is just locally visible, |
31 | all function calls can be redirected. For read-only variables, we create |
32 | aliases if possible. |
33 | |
34 | Optimization pass arranges as follows: |
35 | 1) All functions and read-only variables are visited and internal |
36 | data structure, either sem_function or sem_variables is created. |
37 | 2) For every symbol from the previous step, VAR_DECL and FUNCTION_DECL are |
38 | saved and matched to corresponding sem_items. |
39 | 3) These declaration are ignored for equality check and are solved |
40 | by Value Numbering algorithm published by Alpert, Zadeck in 1992. |
41 | 4) We compute hash value for each symbol. |
42 | 5) Congruence classes are created based on hash value. If hash value are |
43 | equal, equals function is called and symbols are deeply compared. |
44 | We must prove that all SSA names, declarations and other items |
45 | correspond. |
46 | 6) Value Numbering is executed for these classes. At the end of the process |
47 | all symbol members in remaining classes can be merged. |
48 | 7) Merge operation creates alias in case of read-only variables. For |
49 | callgraph node, we must decide if we can redirect local calls, |
50 | create an alias or a thunk. |
51 | |
52 | */ |
53 | |
54 | #include "config.h" |
55 | #include "system.h" |
56 | #include "coretypes.h" |
57 | #include "backend.h" |
58 | #include "target.h" |
59 | #include "rtl.h" |
60 | #include "tree.h" |
61 | #include "gimple.h" |
62 | #include "alloc-pool.h" |
63 | #include "tree-pass.h" |
64 | #include "ssa.h" |
65 | #include "cgraph.h" |
66 | #include "coverage.h" |
67 | #include "gimple-pretty-print.h" |
68 | #include "data-streamer.h" |
69 | #include "tree-streamer.h" |
70 | #include "fold-const.h" |
71 | #include "calls.h" |
72 | #include "varasm.h" |
73 | #include "gimple-iterator.h" |
74 | #include "tree-cfg.h" |
75 | #include "symbol-summary.h" |
76 | #include "sreal.h" |
77 | #include "ipa-cp.h" |
78 | #include "ipa-prop.h" |
79 | #include "ipa-fnsummary.h" |
80 | #include "except.h" |
81 | #include "attribs.h" |
82 | #include "print-tree.h" |
83 | #include "ipa-utils.h" |
84 | #include "tree-ssa-alias-compare.h" |
85 | #include "ipa-icf-gimple.h" |
86 | #include "fibonacci_heap.h" |
87 | #include "ipa-icf.h" |
88 | #include "stor-layout.h" |
89 | #include "dbgcnt.h" |
90 | #include "tree-vector-builder.h" |
91 | #include "symtab-thunks.h" |
92 | #include "alias.h" |
93 | #include "asan.h" |
94 | |
95 | using namespace ipa_icf_gimple; |
96 | |
97 | namespace ipa_icf { |
98 | |
99 | /* Initialization and computation of symtab node hash, there data |
100 | are propagated later on. */ |
101 | |
102 | static sem_item_optimizer *optimizer = NULL; |
103 | |
104 | /* Constructor. */ |
105 | |
106 | symbol_compare_collection::symbol_compare_collection (symtab_node *node) |
107 | { |
108 | m_references.create (nelems: 0); |
109 | m_interposables.create (nelems: 0); |
110 | |
111 | ipa_ref *ref; |
112 | |
113 | if (is_a <varpool_node *> (p: node) && DECL_VIRTUAL_P (node->decl)) |
114 | return; |
115 | |
116 | for (unsigned i = 0; node->iterate_reference (i, ref); i++) |
117 | { |
118 | if (ref->address_matters_p ()) |
119 | m_references.safe_push (obj: ref->referred); |
120 | |
121 | if (ref->referred->get_availability () <= AVAIL_INTERPOSABLE) |
122 | { |
123 | if (ref->address_matters_p ()) |
124 | m_references.safe_push (obj: ref->referred); |
125 | else |
126 | m_interposables.safe_push (obj: ref->referred); |
127 | } |
128 | } |
129 | |
130 | if (is_a <cgraph_node *> (p: node)) |
131 | { |
132 | cgraph_node *cnode = dyn_cast <cgraph_node *> (p: node); |
133 | |
134 | for (cgraph_edge *e = cnode->callees; e; e = e->next_callee) |
135 | if (e->callee->get_availability () <= AVAIL_INTERPOSABLE) |
136 | m_interposables.safe_push (obj: e->callee); |
137 | } |
138 | } |
139 | |
140 | /* Constructor for key value pair, where _ITEM is key and _INDEX is a target. */ |
141 | |
142 | sem_usage_pair::sem_usage_pair (sem_item *_item, unsigned int _index) |
143 | : item (_item), index (_index) |
144 | { |
145 | } |
146 | |
147 | sem_item::sem_item (sem_item_type _type, bitmap_obstack *stack) |
148 | : type (_type), referenced_by_count (0), m_hash (-1), m_hash_set (false) |
149 | { |
150 | setup (stack); |
151 | } |
152 | |
153 | sem_item::sem_item (sem_item_type _type, symtab_node *_node, |
154 | bitmap_obstack *stack) |
155 | : type (_type), node (_node), referenced_by_count (0), m_hash (-1), |
156 | m_hash_set (false) |
157 | { |
158 | decl = node->decl; |
159 | setup (stack); |
160 | } |
161 | |
162 | /* Add reference to a semantic TARGET. */ |
163 | |
164 | void |
165 | sem_item::add_reference (ref_map *refs, |
166 | sem_item *target) |
167 | { |
168 | unsigned index = reference_count++; |
169 | bool existed; |
170 | |
171 | sem_usage_pair *pair = new sem_usage_pair (target, index); |
172 | vec<sem_item *> &v = refs->get_or_insert (k: pair, existed: &existed); |
173 | if (existed) |
174 | delete pair; |
175 | |
176 | v.safe_push (obj: this); |
177 | bitmap_set_bit (target->usage_index_bitmap, index); |
178 | refs_set.add (k: target->node); |
179 | ++target->referenced_by_count; |
180 | } |
181 | |
182 | /* Initialize internal data structures. Bitmap STACK is used for |
183 | bitmap memory allocation process. */ |
184 | |
185 | void |
186 | sem_item::setup (bitmap_obstack *stack) |
187 | { |
188 | gcc_checking_assert (node); |
189 | |
190 | reference_count = 0; |
191 | tree_refs.create (nelems: 0); |
192 | usage_index_bitmap = BITMAP_ALLOC (obstack: stack); |
193 | } |
194 | |
195 | sem_item::~sem_item () |
196 | { |
197 | tree_refs.release (); |
198 | |
199 | BITMAP_FREE (usage_index_bitmap); |
200 | } |
201 | |
202 | /* Dump function for debugging purpose. */ |
203 | |
204 | DEBUG_FUNCTION void |
205 | sem_item::dump (void) |
206 | { |
207 | if (dump_file) |
208 | { |
209 | fprintf (stream: dump_file, format: "[%s] %s (tree:%p)\n" , type == FUNC ? "func" : "var" , |
210 | node->dump_name (), (void *) node->decl); |
211 | fprintf (stream: dump_file, format: " hash: %u\n" , get_hash ()); |
212 | } |
213 | } |
214 | |
215 | /* Return true if target supports alias symbols. */ |
216 | |
217 | bool |
218 | sem_item::target_supports_symbol_aliases_p (void) |
219 | { |
220 | #if !defined (ASM_OUTPUT_DEF) || (!defined(ASM_OUTPUT_WEAK_ALIAS) && !defined (ASM_WEAKEN_DECL)) |
221 | return false; |
222 | #else |
223 | gcc_checking_assert (TARGET_SUPPORTS_ALIASES); |
224 | return true; |
225 | #endif |
226 | } |
227 | |
228 | void sem_item::set_hash (hashval_t hash) |
229 | { |
230 | m_hash = hash; |
231 | m_hash_set = true; |
232 | } |
233 | |
234 | hash_map<const_tree, hashval_t> sem_item::m_type_hash_cache; |
235 | |
236 | /* Semantic function constructor that uses STACK as bitmap memory stack. */ |
237 | |
238 | sem_function::sem_function (bitmap_obstack *stack) |
239 | : sem_item (FUNC, stack), memory_access_types (), m_alias_sets_hash (0), |
240 | m_checker (NULL), m_compared_func (NULL) |
241 | { |
242 | bb_sizes.create (nelems: 0); |
243 | bb_sorted.create (nelems: 0); |
244 | } |
245 | |
246 | sem_function::sem_function (cgraph_node *node, bitmap_obstack *stack) |
247 | : sem_item (FUNC, node, stack), memory_access_types (), |
248 | m_alias_sets_hash (0), m_checker (NULL), m_compared_func (NULL) |
249 | { |
250 | bb_sizes.create (nelems: 0); |
251 | bb_sorted.create (nelems: 0); |
252 | } |
253 | |
254 | sem_function::~sem_function () |
255 | { |
256 | for (unsigned i = 0; i < bb_sorted.length (); i++) |
257 | delete (bb_sorted[i]); |
258 | |
259 | bb_sizes.release (); |
260 | bb_sorted.release (); |
261 | } |
262 | |
263 | /* Calculates hash value based on a BASIC_BLOCK. */ |
264 | |
265 | hashval_t |
266 | sem_function::get_bb_hash (const sem_bb *basic_block) |
267 | { |
268 | inchash::hash hstate; |
269 | |
270 | hstate.add_int (v: basic_block->nondbg_stmt_count); |
271 | hstate.add_int (v: basic_block->edge_count); |
272 | |
273 | return hstate.end (); |
274 | } |
275 | |
276 | /* References independent hash function. */ |
277 | |
278 | hashval_t |
279 | sem_function::get_hash (void) |
280 | { |
281 | if (!m_hash_set) |
282 | { |
283 | inchash::hash hstate; |
284 | hstate.add_int (v: 177454); /* Random number for function type. */ |
285 | |
286 | hstate.add_int (v: arg_count); |
287 | hstate.add_int (v: cfg_checksum); |
288 | hstate.add_int (v: gcode_hash); |
289 | |
290 | for (unsigned i = 0; i < bb_sorted.length (); i++) |
291 | hstate.merge_hash (other: get_bb_hash (basic_block: bb_sorted[i])); |
292 | |
293 | for (unsigned i = 0; i < bb_sizes.length (); i++) |
294 | hstate.add_int (v: bb_sizes[i]); |
295 | |
296 | /* Add common features of declaration itself. */ |
297 | if (DECL_FUNCTION_SPECIFIC_TARGET (decl)) |
298 | hstate.add_hwi |
299 | (v: cl_target_option_hash |
300 | (TREE_TARGET_OPTION (DECL_FUNCTION_SPECIFIC_TARGET (decl)))); |
301 | if (DECL_FUNCTION_SPECIFIC_OPTIMIZATION (decl)) |
302 | hstate.add_hwi |
303 | (v: cl_optimization_hash |
304 | (TREE_OPTIMIZATION (DECL_FUNCTION_SPECIFIC_OPTIMIZATION (decl)))); |
305 | hstate.add_flag (DECL_CXX_CONSTRUCTOR_P (decl)); |
306 | hstate.add_flag (DECL_CXX_DESTRUCTOR_P (decl)); |
307 | |
308 | set_hash (hstate.end ()); |
309 | } |
310 | |
311 | return m_hash; |
312 | } |
313 | |
314 | /* Compare properties of symbols N1 and N2 that does not affect semantics of |
315 | symbol itself but affects semantics of its references from USED_BY (which |
316 | may be NULL if it is unknown). If comparison is false, symbols |
317 | can still be merged but any symbols referring them can't. |
318 | |
319 | If ADDRESS is true, do extra checking needed for IPA_REF_ADDR. |
320 | |
321 | TODO: We can also split attributes to those that determine codegen of |
322 | a function body/variable constructor itself and those that are used when |
323 | referring to it. */ |
324 | |
325 | bool |
326 | sem_item::compare_referenced_symbol_properties (symtab_node *used_by, |
327 | symtab_node *n1, |
328 | symtab_node *n2, |
329 | bool address) |
330 | { |
331 | if (is_a <cgraph_node *> (p: n1)) |
332 | { |
333 | /* Inline properties matters: we do now want to merge uses of inline |
334 | function to uses of normal function because inline hint would be lost. |
335 | We however can merge inline function to noinline because the alias |
336 | will keep its DECL_DECLARED_INLINE flag. |
337 | |
338 | Also ignore inline flag when optimizing for size or when function |
339 | is known to not be inlinable. |
340 | |
341 | TODO: the optimize_size checks can also be assumed to be true if |
342 | unit has no !optimize_size functions. */ |
343 | |
344 | if ((!used_by || address || !is_a <cgraph_node *> (p: used_by) |
345 | || !opt_for_fn (used_by->decl, optimize_size)) |
346 | && !opt_for_fn (n1->decl, optimize_size) |
347 | && n1->get_availability () > AVAIL_INTERPOSABLE |
348 | && (!DECL_UNINLINABLE (n1->decl) || !DECL_UNINLINABLE (n2->decl))) |
349 | { |
350 | if (DECL_DISREGARD_INLINE_LIMITS (n1->decl) |
351 | != DECL_DISREGARD_INLINE_LIMITS (n2->decl)) |
352 | return return_false_with_msg |
353 | ("DECL_DISREGARD_INLINE_LIMITS are different" ); |
354 | |
355 | if (DECL_DECLARED_INLINE_P (n1->decl) |
356 | != DECL_DECLARED_INLINE_P (n2->decl)) |
357 | return return_false_with_msg ("inline attributes are different" ); |
358 | } |
359 | |
360 | if (DECL_IS_OPERATOR_NEW_P (n1->decl) |
361 | != DECL_IS_OPERATOR_NEW_P (n2->decl)) |
362 | return return_false_with_msg ("operator new flags are different" ); |
363 | |
364 | if (DECL_IS_REPLACEABLE_OPERATOR (n1->decl) |
365 | != DECL_IS_REPLACEABLE_OPERATOR (n2->decl)) |
366 | return return_false_with_msg ("replaceable operator flags are different" ); |
367 | } |
368 | |
369 | /* Merging two definitions with a reference to equivalent vtables, but |
370 | belonging to a different type may result in ipa-polymorphic-call analysis |
371 | giving a wrong answer about the dynamic type of instance. */ |
372 | if (is_a <varpool_node *> (p: n1)) |
373 | { |
374 | if ((DECL_VIRTUAL_P (n1->decl) || DECL_VIRTUAL_P (n2->decl)) |
375 | && (DECL_VIRTUAL_P (n1->decl) != DECL_VIRTUAL_P (n2->decl) |
376 | || !types_must_be_same_for_odr (DECL_CONTEXT (n1->decl), |
377 | DECL_CONTEXT (n2->decl))) |
378 | && (!used_by || !is_a <cgraph_node *> (p: used_by) || address |
379 | || opt_for_fn (used_by->decl, flag_devirtualize))) |
380 | return return_false_with_msg |
381 | ("references to virtual tables cannot be merged" ); |
382 | |
383 | if (address && DECL_ALIGN (n1->decl) != DECL_ALIGN (n2->decl)) |
384 | return return_false_with_msg ("alignment mismatch" ); |
385 | |
386 | /* For functions we compare attributes in equals_wpa, because we do |
387 | not know what attributes may cause codegen differences, but for |
388 | variables just compare attributes for references - the codegen |
389 | for constructors is affected only by those attributes that we lower |
390 | to explicit representation (such as DECL_ALIGN or DECL_SECTION). */ |
391 | if (!attribute_list_equal (DECL_ATTRIBUTES (n1->decl), |
392 | DECL_ATTRIBUTES (n2->decl))) |
393 | return return_false_with_msg ("different var decl attributes" ); |
394 | if (comp_type_attributes (TREE_TYPE (n1->decl), |
395 | TREE_TYPE (n2->decl)) != 1) |
396 | return return_false_with_msg ("different var type attributes" ); |
397 | } |
398 | |
399 | /* When matching virtual tables, be sure to also match information |
400 | relevant for polymorphic call analysis. */ |
401 | if (used_by && is_a <varpool_node *> (p: used_by) |
402 | && DECL_VIRTUAL_P (used_by->decl)) |
403 | { |
404 | if (DECL_VIRTUAL_P (n1->decl) != DECL_VIRTUAL_P (n2->decl)) |
405 | return return_false_with_msg ("virtual flag mismatch" ); |
406 | if (DECL_VIRTUAL_P (n1->decl) && is_a <cgraph_node *> (p: n1) |
407 | && (DECL_FINAL_P (n1->decl) != DECL_FINAL_P (n2->decl))) |
408 | return return_false_with_msg ("final flag mismatch" ); |
409 | } |
410 | return true; |
411 | } |
412 | |
413 | /* Hash properties that are compared by compare_referenced_symbol_properties. */ |
414 | |
415 | void |
416 | sem_item::hash_referenced_symbol_properties (symtab_node *ref, |
417 | inchash::hash &hstate, |
418 | bool address) |
419 | { |
420 | if (is_a <cgraph_node *> (p: ref)) |
421 | { |
422 | if ((type != FUNC || address || !opt_for_fn (decl, optimize_size)) |
423 | && !opt_for_fn (ref->decl, optimize_size) |
424 | && !DECL_UNINLINABLE (ref->decl)) |
425 | { |
426 | hstate.add_flag (DECL_DISREGARD_INLINE_LIMITS (ref->decl)); |
427 | hstate.add_flag (DECL_DECLARED_INLINE_P (ref->decl)); |
428 | } |
429 | hstate.add_flag (DECL_IS_OPERATOR_NEW_P (ref->decl)); |
430 | } |
431 | else if (is_a <varpool_node *> (p: ref)) |
432 | { |
433 | hstate.add_flag (DECL_VIRTUAL_P (ref->decl)); |
434 | if (address) |
435 | hstate.add_int (DECL_ALIGN (ref->decl)); |
436 | } |
437 | } |
438 | |
439 | |
440 | /* For a given symbol table nodes N1 and N2, we check that FUNCTION_DECLs |
441 | point to a same function. Comparison can be skipped if IGNORED_NODES |
442 | contains these nodes. ADDRESS indicate if address is taken. */ |
443 | |
444 | bool |
445 | sem_item::compare_symbol_references ( |
446 | hash_map <symtab_node *, sem_item *> &ignored_nodes, |
447 | symtab_node *n1, symtab_node *n2, bool address) |
448 | { |
449 | enum availability avail1, avail2; |
450 | |
451 | if (n1 == n2) |
452 | return true; |
453 | |
454 | /* Never match variable and function. */ |
455 | if (is_a <varpool_node *> (p: n1) != is_a <varpool_node *> (p: n2)) |
456 | return false; |
457 | |
458 | if (!compare_referenced_symbol_properties (used_by: node, n1, n2, address)) |
459 | return false; |
460 | if (address && n1->equal_address_to (s2: n2) == 1) |
461 | return true; |
462 | if (!address && n1->semantically_equivalent_p (target: n2)) |
463 | return true; |
464 | |
465 | n1 = n1->ultimate_alias_target (availability: &avail1); |
466 | n2 = n2->ultimate_alias_target (availability: &avail2); |
467 | |
468 | if (avail1 > AVAIL_INTERPOSABLE && ignored_nodes.get (k: n1) |
469 | && avail2 > AVAIL_INTERPOSABLE && ignored_nodes.get (k: n2)) |
470 | return true; |
471 | |
472 | return return_false_with_msg ("different references" ); |
473 | } |
474 | |
475 | /* If cgraph edges E1 and E2 are indirect calls, verify that |
476 | ECF flags are the same. */ |
477 | |
478 | bool sem_function::compare_edge_flags (cgraph_edge *e1, cgraph_edge *e2) |
479 | { |
480 | if (e1->indirect_info && e2->indirect_info) |
481 | { |
482 | int e1_flags = e1->indirect_info->ecf_flags; |
483 | int e2_flags = e2->indirect_info->ecf_flags; |
484 | |
485 | if (e1_flags != e2_flags) |
486 | return return_false_with_msg ("ICF flags are different" ); |
487 | } |
488 | else if (e1->indirect_info || e2->indirect_info) |
489 | return false; |
490 | |
491 | return true; |
492 | } |
493 | |
494 | /* Return true if parameter I may be used. */ |
495 | |
496 | bool |
497 | sem_function::param_used_p (unsigned int i) |
498 | { |
499 | if (ipa_node_params_sum == NULL) |
500 | return true; |
501 | |
502 | ipa_node_params *parms_info = ipa_node_params_sum->get (node: get_node ()); |
503 | |
504 | if (!parms_info || vec_safe_length (v: parms_info->descriptors) <= i) |
505 | return true; |
506 | |
507 | return ipa_is_param_used (info: parms_info, i); |
508 | } |
509 | |
510 | /* Perform additional check needed to match types function parameters that are |
511 | used. Unlike for normal decls it matters if type is TYPE_RESTRICT and we |
512 | make an assumption that REFERENCE_TYPE parameters are always non-NULL. */ |
513 | |
514 | bool |
515 | sem_function::compatible_parm_types_p (tree parm1, tree parm2) |
516 | { |
517 | /* Be sure that parameters are TBAA compatible. */ |
518 | if (!func_checker::compatible_types_p (t1: parm1, t2: parm2)) |
519 | return return_false_with_msg ("parameter type is not compatible" ); |
520 | |
521 | if (POINTER_TYPE_P (parm1) |
522 | && (TYPE_RESTRICT (parm1) != TYPE_RESTRICT (parm2))) |
523 | return return_false_with_msg ("argument restrict flag mismatch" ); |
524 | |
525 | /* nonnull_arg_p implies non-zero range to REFERENCE types. */ |
526 | if (POINTER_TYPE_P (parm1) |
527 | && TREE_CODE (parm1) != TREE_CODE (parm2) |
528 | && opt_for_fn (decl, flag_delete_null_pointer_checks)) |
529 | return return_false_with_msg ("pointer wrt reference mismatch" ); |
530 | |
531 | return true; |
532 | } |
533 | |
534 | /* Fast equality function based on knowledge known in WPA. */ |
535 | |
536 | bool |
537 | sem_function::equals_wpa (sem_item *item, |
538 | hash_map <symtab_node *, sem_item *> &ignored_nodes) |
539 | { |
540 | gcc_assert (item->type == FUNC); |
541 | cgraph_node *cnode = dyn_cast <cgraph_node *> (p: node); |
542 | cgraph_node *cnode2 = dyn_cast <cgraph_node *> (p: item->node); |
543 | |
544 | m_compared_func = static_cast<sem_function *> (item); |
545 | |
546 | if (cnode->thunk != cnode2->thunk) |
547 | return return_false_with_msg ("thunk mismatch" ); |
548 | if (cnode->former_thunk_p () != cnode2->former_thunk_p ()) |
549 | return return_false_with_msg ("former_thunk_p mismatch" ); |
550 | |
551 | if ((cnode->thunk || cnode->former_thunk_p ()) |
552 | && thunk_info::get (node: cnode) != thunk_info::get (node: cnode2)) |
553 | return return_false_with_msg ("thunk_info mismatch" ); |
554 | |
555 | /* Compare special function DECL attributes. */ |
556 | if (DECL_FUNCTION_PERSONALITY (decl) |
557 | != DECL_FUNCTION_PERSONALITY (item->decl)) |
558 | return return_false_with_msg ("function personalities are different" ); |
559 | |
560 | if (DECL_NO_INSTRUMENT_FUNCTION_ENTRY_EXIT (decl) |
561 | != DECL_NO_INSTRUMENT_FUNCTION_ENTRY_EXIT (item->decl)) |
562 | return return_false_with_msg ("instrument function entry exit " |
563 | "attributes are different" ); |
564 | |
565 | if (DECL_NO_LIMIT_STACK (decl) != DECL_NO_LIMIT_STACK (item->decl)) |
566 | return return_false_with_msg ("no stack limit attributes are different" ); |
567 | |
568 | if (DECL_CXX_CONSTRUCTOR_P (decl) != DECL_CXX_CONSTRUCTOR_P (item->decl)) |
569 | return return_false_with_msg ("DECL_CXX_CONSTRUCTOR mismatch" ); |
570 | |
571 | if (DECL_CXX_DESTRUCTOR_P (decl) != DECL_CXX_DESTRUCTOR_P (item->decl)) |
572 | return return_false_with_msg ("DECL_CXX_DESTRUCTOR mismatch" ); |
573 | |
574 | /* TODO: pure/const flags mostly matters only for references, except for |
575 | the fact that codegen takes LOOPING flag as a hint that loops are |
576 | finite. We may arrange the code to always pick leader that has least |
577 | specified flags and then this can go into comparing symbol properties. */ |
578 | if (flags_from_decl_or_type (decl) != flags_from_decl_or_type (item->decl)) |
579 | return return_false_with_msg ("decl_or_type flags are different" ); |
580 | |
581 | /* Do not match polymorphic constructors of different types. They calls |
582 | type memory location for ipa-polymorphic-call and we do not want |
583 | it to get confused by wrong type. */ |
584 | if (DECL_CXX_CONSTRUCTOR_P (decl) |
585 | && opt_for_fn (decl, flag_devirtualize) |
586 | && TREE_CODE (TREE_TYPE (decl)) == METHOD_TYPE) |
587 | { |
588 | if (TREE_CODE (TREE_TYPE (item->decl)) != METHOD_TYPE) |
589 | return return_false_with_msg ("DECL_CXX_CONSTRUCTOR type mismatch" ); |
590 | else if (!func_checker::compatible_polymorphic_types_p |
591 | (TYPE_METHOD_BASETYPE (TREE_TYPE (decl)), |
592 | TYPE_METHOD_BASETYPE (TREE_TYPE (item->decl)), compare_ptr: false)) |
593 | return return_false_with_msg ("ctor polymorphic type mismatch" ); |
594 | } |
595 | |
596 | /* Checking function TARGET and OPTIMIZATION flags. */ |
597 | cl_target_option *tar1 = target_opts_for_fn (fndecl: decl); |
598 | cl_target_option *tar2 = target_opts_for_fn (fndecl: item->decl); |
599 | |
600 | if (tar1 != tar2 && !cl_target_option_eq (tar1, tar2)) |
601 | { |
602 | if (dump_file && (dump_flags & TDF_DETAILS)) |
603 | { |
604 | fprintf (stream: dump_file, format: "target flags difference" ); |
605 | cl_target_option_print_diff (dump_file, 2, ptr1: tar1, ptr2: tar2); |
606 | } |
607 | |
608 | return return_false_with_msg ("Target flags are different" ); |
609 | } |
610 | |
611 | cl_optimization *opt1 = opts_for_fn (fndecl: decl); |
612 | cl_optimization *opt2 = opts_for_fn (fndecl: item->decl); |
613 | |
614 | if (opt1 != opt2 && !cl_optimization_option_eq (ptr1: opt1, ptr2: opt2)) |
615 | { |
616 | if (dump_file && (dump_flags & TDF_DETAILS)) |
617 | { |
618 | fprintf (stream: dump_file, format: "optimization flags difference" ); |
619 | cl_optimization_print_diff (dump_file, 2, ptr1: opt1, ptr2: opt2); |
620 | } |
621 | |
622 | return return_false_with_msg ("optimization flags are different" ); |
623 | } |
624 | |
625 | /* Result type checking. */ |
626 | if (!func_checker::compatible_types_p |
627 | (TREE_TYPE (TREE_TYPE (decl)), |
628 | TREE_TYPE (TREE_TYPE (m_compared_func->decl)))) |
629 | return return_false_with_msg ("result types are different" ); |
630 | |
631 | /* Checking types of arguments. */ |
632 | tree list1 = TYPE_ARG_TYPES (TREE_TYPE (decl)), |
633 | list2 = TYPE_ARG_TYPES (TREE_TYPE (m_compared_func->decl)); |
634 | for (unsigned i = 0; list1 && list2; |
635 | list1 = TREE_CHAIN (list1), list2 = TREE_CHAIN (list2), i++) |
636 | { |
637 | tree parm1 = TREE_VALUE (list1); |
638 | tree parm2 = TREE_VALUE (list2); |
639 | |
640 | /* This guard is here for function pointer with attributes (pr59927.c). */ |
641 | if (!parm1 || !parm2) |
642 | return return_false_with_msg ("NULL argument type" ); |
643 | |
644 | /* Verify that types are compatible to ensure that both functions |
645 | have same calling conventions. */ |
646 | if (!types_compatible_p (type1: parm1, type2: parm2)) |
647 | return return_false_with_msg ("parameter types are not compatible" ); |
648 | |
649 | if (!param_used_p (i)) |
650 | continue; |
651 | |
652 | /* Perform additional checks for used parameters. */ |
653 | if (!compatible_parm_types_p (parm1, parm2)) |
654 | return false; |
655 | } |
656 | |
657 | if (list1 || list2) |
658 | return return_false_with_msg ("Mismatched number of parameters" ); |
659 | |
660 | if (node->num_references () != item->node->num_references ()) |
661 | return return_false_with_msg ("different number of references" ); |
662 | |
663 | /* Checking function attributes. |
664 | This is quadratic in number of attributes */ |
665 | if (comp_type_attributes (TREE_TYPE (decl), |
666 | TREE_TYPE (item->decl)) != 1) |
667 | return return_false_with_msg ("different type attributes" ); |
668 | if (!attribute_list_equal (DECL_ATTRIBUTES (decl), |
669 | DECL_ATTRIBUTES (item->decl))) |
670 | return return_false_with_msg ("different decl attributes" ); |
671 | |
672 | /* The type of THIS pointer type memory location for |
673 | ipa-polymorphic-call-analysis. */ |
674 | if (opt_for_fn (decl, flag_devirtualize) |
675 | && (TREE_CODE (TREE_TYPE (decl)) == METHOD_TYPE |
676 | || TREE_CODE (TREE_TYPE (item->decl)) == METHOD_TYPE) |
677 | && param_used_p (i: 0) |
678 | && compare_polymorphic_p ()) |
679 | { |
680 | if (TREE_CODE (TREE_TYPE (decl)) != TREE_CODE (TREE_TYPE (item->decl))) |
681 | return return_false_with_msg ("METHOD_TYPE and FUNCTION_TYPE mismatch" ); |
682 | if (!func_checker::compatible_polymorphic_types_p |
683 | (TYPE_METHOD_BASETYPE (TREE_TYPE (decl)), |
684 | TYPE_METHOD_BASETYPE (TREE_TYPE (item->decl)), compare_ptr: false)) |
685 | return return_false_with_msg ("THIS pointer ODR type mismatch" ); |
686 | } |
687 | |
688 | ipa_ref *ref = NULL, *ref2 = NULL; |
689 | for (unsigned i = 0; node->iterate_reference (i, ref); i++) |
690 | { |
691 | item->node->iterate_reference (i, ref&: ref2); |
692 | |
693 | if (ref->use != ref2->use) |
694 | return return_false_with_msg ("reference use mismatch" ); |
695 | |
696 | if (!compare_symbol_references (ignored_nodes, n1: ref->referred, |
697 | n2: ref2->referred, |
698 | address: ref->address_matters_p ())) |
699 | return false; |
700 | } |
701 | |
702 | cgraph_edge *e1 = dyn_cast <cgraph_node *> (p: node)->callees; |
703 | cgraph_edge *e2 = dyn_cast <cgraph_node *> (p: item->node)->callees; |
704 | |
705 | while (e1 && e2) |
706 | { |
707 | if (!compare_symbol_references (ignored_nodes, n1: e1->callee, |
708 | n2: e2->callee, address: false)) |
709 | return false; |
710 | if (!compare_edge_flags (e1, e2)) |
711 | return false; |
712 | |
713 | e1 = e1->next_callee; |
714 | e2 = e2->next_callee; |
715 | } |
716 | |
717 | if (e1 || e2) |
718 | return return_false_with_msg ("different number of calls" ); |
719 | |
720 | e1 = dyn_cast <cgraph_node *> (p: node)->indirect_calls; |
721 | e2 = dyn_cast <cgraph_node *> (p: item->node)->indirect_calls; |
722 | |
723 | while (e1 && e2) |
724 | { |
725 | if (!compare_edge_flags (e1, e2)) |
726 | return false; |
727 | |
728 | e1 = e1->next_callee; |
729 | e2 = e2->next_callee; |
730 | } |
731 | |
732 | if (e1 || e2) |
733 | return return_false_with_msg ("different number of indirect calls" ); |
734 | |
735 | return true; |
736 | } |
737 | |
738 | /* Update hash by address sensitive references. We iterate over all |
739 | sensitive references (address_matters_p) and we hash ultimate alias |
740 | target of these nodes, which can improve a semantic item hash. |
741 | |
742 | Also hash in referenced symbols properties. This can be done at any time |
743 | (as the properties should not change), but it is convenient to do it here |
744 | while we walk the references anyway. */ |
745 | |
746 | void |
747 | sem_item::update_hash_by_addr_refs (hash_map <symtab_node *, |
748 | sem_item *> &m_symtab_node_map) |
749 | { |
750 | ipa_ref* ref; |
751 | inchash::hash hstate (get_hash ()); |
752 | |
753 | for (unsigned i = 0; node->iterate_reference (i, ref); i++) |
754 | { |
755 | hstate.add_int (v: ref->use); |
756 | hash_referenced_symbol_properties (ref: ref->referred, hstate, |
757 | address: ref->use == IPA_REF_ADDR); |
758 | if (ref->address_matters_p () || !m_symtab_node_map.get (k: ref->referred)) |
759 | hstate.add_int (v: ref->referred->ultimate_alias_target ()->order); |
760 | } |
761 | |
762 | if (is_a <cgraph_node *> (p: node)) |
763 | { |
764 | for (cgraph_edge *e = dyn_cast <cgraph_node *> (p: node)->callers; e; |
765 | e = e->next_caller) |
766 | { |
767 | sem_item **result = m_symtab_node_map.get (k: e->callee); |
768 | hash_referenced_symbol_properties (ref: e->callee, hstate, address: false); |
769 | if (!result) |
770 | hstate.add_int (v: e->callee->ultimate_alias_target ()->order); |
771 | } |
772 | } |
773 | |
774 | set_hash (hstate.end ()); |
775 | } |
776 | |
777 | /* Update hash by computed local hash values taken from different |
778 | semantic items. |
779 | TODO: stronger SCC based hashing would be desirable here. */ |
780 | |
781 | void |
782 | sem_item::update_hash_by_local_refs (hash_map <symtab_node *, |
783 | sem_item *> &m_symtab_node_map) |
784 | { |
785 | ipa_ref* ref; |
786 | inchash::hash state (get_hash ()); |
787 | |
788 | for (unsigned j = 0; node->iterate_reference (i: j, ref); j++) |
789 | { |
790 | sem_item **result = m_symtab_node_map.get (k: ref->referring); |
791 | if (result) |
792 | state.merge_hash (other: (*result)->get_hash ()); |
793 | } |
794 | |
795 | if (type == FUNC) |
796 | { |
797 | for (cgraph_edge *e = dyn_cast <cgraph_node *> (p: node)->callees; e; |
798 | e = e->next_callee) |
799 | { |
800 | sem_item **result = m_symtab_node_map.get (k: e->caller); |
801 | if (result) |
802 | state.merge_hash (other: (*result)->get_hash ()); |
803 | } |
804 | } |
805 | |
806 | global_hash = state.end (); |
807 | } |
808 | |
809 | /* Returns true if the item equals to ITEM given as argument. */ |
810 | |
811 | bool |
812 | sem_function::equals (sem_item *item, |
813 | hash_map <symtab_node *, sem_item *> &) |
814 | { |
815 | gcc_assert (item->type == FUNC); |
816 | bool eq = equals_private (item); |
817 | |
818 | if (m_checker != NULL) |
819 | { |
820 | delete m_checker; |
821 | m_checker = NULL; |
822 | } |
823 | |
824 | if (dump_file && (dump_flags & TDF_DETAILS)) |
825 | fprintf (stream: dump_file, |
826 | format: "Equals called for: %s:%s with result: %s\n\n" , |
827 | node->dump_name (), |
828 | item->node->dump_name (), |
829 | eq ? "true" : "false" ); |
830 | |
831 | return eq; |
832 | } |
833 | |
834 | /* Processes function equality comparison. */ |
835 | |
836 | bool |
837 | sem_function::equals_private (sem_item *item) |
838 | { |
839 | if (item->type != FUNC) |
840 | return false; |
841 | |
842 | basic_block bb1, bb2; |
843 | edge e1, e2; |
844 | edge_iterator ei1, ei2; |
845 | bool result = true; |
846 | tree arg1, arg2; |
847 | |
848 | m_compared_func = static_cast<sem_function *> (item); |
849 | |
850 | gcc_assert (decl != item->decl); |
851 | |
852 | if (bb_sorted.length () != m_compared_func->bb_sorted.length () |
853 | || edge_count != m_compared_func->edge_count |
854 | || cfg_checksum != m_compared_func->cfg_checksum) |
855 | return return_false (); |
856 | |
857 | m_checker = new func_checker (decl, m_compared_func->decl, |
858 | false, |
859 | opt_for_fn (m_compared_func->decl, |
860 | flag_strict_aliasing), |
861 | &refs_set, |
862 | &m_compared_func->refs_set); |
863 | arg1 = DECL_ARGUMENTS (decl); |
864 | arg2 = DECL_ARGUMENTS (m_compared_func->decl); |
865 | for (unsigned i = 0; |
866 | arg1 && arg2; arg1 = DECL_CHAIN (arg1), arg2 = DECL_CHAIN (arg2), i++) |
867 | { |
868 | if (!types_compatible_p (TREE_TYPE (arg1), TREE_TYPE (arg2))) |
869 | return return_false_with_msg ("argument types are not compatible" ); |
870 | if (!param_used_p (i)) |
871 | continue; |
872 | /* Perform additional checks for used parameters. */ |
873 | if (!compatible_parm_types_p (TREE_TYPE (arg1), TREE_TYPE (arg2))) |
874 | return false; |
875 | if (!m_checker->compare_decl (t1: arg1, t2: arg2)) |
876 | return return_false (); |
877 | } |
878 | if (arg1 || arg2) |
879 | return return_false_with_msg ("Mismatched number of arguments" ); |
880 | |
881 | if (!dyn_cast <cgraph_node *> (p: node)->has_gimple_body_p ()) |
882 | return true; |
883 | |
884 | /* Fill-up label dictionary. */ |
885 | for (unsigned i = 0; i < bb_sorted.length (); ++i) |
886 | { |
887 | m_checker->parse_labels (bb: bb_sorted[i]); |
888 | m_checker->parse_labels (bb: m_compared_func->bb_sorted[i]); |
889 | } |
890 | |
891 | /* Checking all basic blocks. */ |
892 | for (unsigned i = 0; i < bb_sorted.length (); ++i) |
893 | if(!m_checker->compare_bb (bb1: bb_sorted[i], bb2: m_compared_func->bb_sorted[i])) |
894 | return return_false (); |
895 | |
896 | auto_vec <int> bb_dict; |
897 | |
898 | /* Basic block edges check. */ |
899 | for (unsigned i = 0; i < bb_sorted.length (); ++i) |
900 | { |
901 | bb1 = bb_sorted[i]->bb; |
902 | bb2 = m_compared_func->bb_sorted[i]->bb; |
903 | |
904 | ei2 = ei_start (bb2->preds); |
905 | |
906 | for (ei1 = ei_start (bb1->preds); ei_cond (ei: ei1, p: &e1); ei_next (i: &ei1)) |
907 | { |
908 | ei_cond (ei: ei2, p: &e2); |
909 | |
910 | if (e1->flags != e2->flags) |
911 | return return_false_with_msg ("flags comparison returns false" ); |
912 | |
913 | if (!bb_dict_test (bb_dict: &bb_dict, source: e1->src->index, target: e2->src->index)) |
914 | return return_false_with_msg ("edge comparison returns false" ); |
915 | |
916 | if (!bb_dict_test (bb_dict: &bb_dict, source: e1->dest->index, target: e2->dest->index)) |
917 | return return_false_with_msg ("BB comparison returns false" ); |
918 | |
919 | if (!m_checker->compare_edge (e1, e2)) |
920 | return return_false_with_msg ("edge comparison returns false" ); |
921 | |
922 | ei_next (i: &ei2); |
923 | } |
924 | } |
925 | |
926 | /* Basic block PHI nodes comparison. */ |
927 | for (unsigned i = 0; i < bb_sorted.length (); i++) |
928 | if (!compare_phi_node (bb1: bb_sorted[i]->bb, bb2: m_compared_func->bb_sorted[i]->bb)) |
929 | return return_false_with_msg ("PHI node comparison returns false" ); |
930 | |
931 | return result; |
932 | } |
933 | |
934 | /* Set LOCAL_P of NODE to true if DATA is non-NULL. |
935 | Helper for call_for_symbol_thunks_and_aliases. */ |
936 | |
937 | static bool |
938 | set_local (cgraph_node *node, void *data) |
939 | { |
940 | node->local = data != NULL; |
941 | return false; |
942 | } |
943 | |
944 | /* TREE_ADDRESSABLE of NODE to true. |
945 | Helper for call_for_symbol_thunks_and_aliases. */ |
946 | |
947 | static bool |
948 | set_addressable (varpool_node *node, void *) |
949 | { |
950 | TREE_ADDRESSABLE (node->decl) = 1; |
951 | return false; |
952 | } |
953 | |
954 | /* Clear DECL_RTL of NODE. |
955 | Helper for call_for_symbol_thunks_and_aliases. */ |
956 | |
957 | static bool |
958 | clear_decl_rtl (symtab_node *node, void *) |
959 | { |
960 | SET_DECL_RTL (node->decl, NULL); |
961 | return false; |
962 | } |
963 | |
964 | /* Redirect all callers of N and its aliases to TO. Remove aliases if |
965 | possible. Return number of redirections made. */ |
966 | |
967 | static int |
968 | redirect_all_callers (cgraph_node *n, cgraph_node *to) |
969 | { |
970 | int nredirected = 0; |
971 | ipa_ref *ref; |
972 | cgraph_edge *e = n->callers; |
973 | |
974 | while (e) |
975 | { |
976 | /* Redirecting thunks to interposable symbols or symbols in other sections |
977 | may not be supported by target output code. Play safe for now and |
978 | punt on redirection. */ |
979 | if (!e->caller->thunk) |
980 | { |
981 | struct cgraph_edge *nexte = e->next_caller; |
982 | e->redirect_callee (n: to); |
983 | e = nexte; |
984 | nredirected++; |
985 | } |
986 | else |
987 | e = e->next_callee; |
988 | } |
989 | for (unsigned i = 0; n->iterate_direct_aliases (i, ref);) |
990 | { |
991 | bool removed = false; |
992 | cgraph_node *n_alias = dyn_cast <cgraph_node *> (p: ref->referring); |
993 | |
994 | if ((DECL_COMDAT_GROUP (n->decl) |
995 | && (DECL_COMDAT_GROUP (n->decl) |
996 | == DECL_COMDAT_GROUP (n_alias->decl))) |
997 | || (n_alias->get_availability () > AVAIL_INTERPOSABLE |
998 | && n->get_availability () > AVAIL_INTERPOSABLE)) |
999 | { |
1000 | nredirected += redirect_all_callers (n: n_alias, to); |
1001 | if (n_alias->can_remove_if_no_direct_calls_p () |
1002 | && !n_alias->call_for_symbol_and_aliases (callback: cgraph_node::has_thunk_p, |
1003 | NULL, include_overwritable: true) |
1004 | && !n_alias->has_aliases_p ()) |
1005 | n_alias->remove (); |
1006 | } |
1007 | if (!removed) |
1008 | i++; |
1009 | } |
1010 | return nredirected; |
1011 | } |
1012 | |
1013 | /* Merges instance with an ALIAS_ITEM, where alias, thunk or redirection can |
1014 | be applied. */ |
1015 | |
1016 | bool |
1017 | sem_function::merge (sem_item *alias_item) |
1018 | { |
1019 | gcc_assert (alias_item->type == FUNC); |
1020 | |
1021 | sem_function *alias_func = static_cast<sem_function *> (alias_item); |
1022 | |
1023 | cgraph_node *original = get_node (); |
1024 | cgraph_node *local_original = NULL; |
1025 | cgraph_node *alias = alias_func->get_node (); |
1026 | |
1027 | bool create_wrapper = false; |
1028 | bool create_alias = false; |
1029 | bool redirect_callers = false; |
1030 | bool remove = false; |
1031 | |
1032 | bool original_discardable = false; |
1033 | bool original_discarded = false; |
1034 | |
1035 | bool original_address_matters = original->address_matters_p (); |
1036 | bool alias_address_matters = alias->address_matters_p (); |
1037 | |
1038 | AUTO_DUMP_SCOPE ("merge" , |
1039 | dump_user_location_t::from_function_decl (decl)); |
1040 | |
1041 | if (DECL_EXTERNAL (alias->decl)) |
1042 | { |
1043 | if (dump_enabled_p ()) |
1044 | dump_printf (MSG_MISSED_OPTIMIZATION, |
1045 | "Not unifying; alias is external.\n" ); |
1046 | return false; |
1047 | } |
1048 | |
1049 | if (DECL_NO_INLINE_WARNING_P (original->decl) |
1050 | != DECL_NO_INLINE_WARNING_P (alias->decl)) |
1051 | { |
1052 | if (dump_enabled_p ()) |
1053 | dump_printf (MSG_MISSED_OPTIMIZATION, |
1054 | "Not unifying; DECL_NO_INLINE_WARNING mismatch.\n" ); |
1055 | return false; |
1056 | } |
1057 | |
1058 | /* Do not attempt to mix functions from different user sections; |
1059 | we do not know what user intends with those. */ |
1060 | if (((DECL_SECTION_NAME (original->decl) && !original->implicit_section) |
1061 | || (DECL_SECTION_NAME (alias->decl) && !alias->implicit_section)) |
1062 | && DECL_SECTION_NAME (original->decl) != DECL_SECTION_NAME (alias->decl)) |
1063 | { |
1064 | if (dump_enabled_p ()) |
1065 | dump_printf (MSG_MISSED_OPTIMIZATION, |
1066 | "Not unifying; " |
1067 | "original and alias are in different sections.\n" ); |
1068 | return false; |
1069 | } |
1070 | |
1071 | if (!original->in_same_comdat_group_p (target: alias) |
1072 | || original->comdat_local_p ()) |
1073 | { |
1074 | if (dump_enabled_p ()) |
1075 | dump_printf (MSG_MISSED_OPTIMIZATION, |
1076 | "Not unifying; alias nor wrapper cannot be created; " |
1077 | "across comdat group boundary\n" ); |
1078 | return false; |
1079 | } |
1080 | |
1081 | /* See if original is in a section that can be discarded if the main |
1082 | symbol is not used. */ |
1083 | |
1084 | if (original->can_be_discarded_p ()) |
1085 | original_discardable = true; |
1086 | /* Also consider case where we have resolution info and we know that |
1087 | original's definition is not going to be used. In this case we cannot |
1088 | create alias to original. */ |
1089 | if (node->resolution != LDPR_UNKNOWN |
1090 | && !decl_binds_to_current_def_p (node->decl)) |
1091 | original_discardable = original_discarded = true; |
1092 | |
1093 | /* Creating a symtab alias is the optimal way to merge. |
1094 | It however cannot be used in the following cases: |
1095 | |
1096 | 1) if ORIGINAL and ALIAS may be possibly compared for address equality. |
1097 | 2) if ORIGINAL is in a section that may be discarded by linker or if |
1098 | it is an external functions where we cannot create an alias |
1099 | (ORIGINAL_DISCARDABLE) |
1100 | 3) if target do not support symbol aliases. |
1101 | 4) original and alias lie in different comdat groups. |
1102 | |
1103 | If we cannot produce alias, we will turn ALIAS into WRAPPER of ORIGINAL |
1104 | and/or redirect all callers from ALIAS to ORIGINAL. */ |
1105 | if ((original_address_matters && alias_address_matters) |
1106 | || (original_discardable |
1107 | && (!DECL_COMDAT_GROUP (alias->decl) |
1108 | || (DECL_COMDAT_GROUP (alias->decl) |
1109 | != DECL_COMDAT_GROUP (original->decl)))) |
1110 | || original_discarded |
1111 | || !sem_item::target_supports_symbol_aliases_p () |
1112 | || DECL_COMDAT_GROUP (alias->decl) != DECL_COMDAT_GROUP (original->decl)) |
1113 | { |
1114 | /* First see if we can produce wrapper. */ |
1115 | |
1116 | /* Symbol properties that matter for references must be preserved. |
1117 | TODO: We can produce wrapper, but we need to produce alias of ORIGINAL |
1118 | with proper properties. */ |
1119 | if (!sem_item::compare_referenced_symbol_properties (NULL, n1: original, n2: alias, |
1120 | address: alias->address_taken)) |
1121 | { |
1122 | if (dump_enabled_p ()) |
1123 | dump_printf (MSG_MISSED_OPTIMIZATION, |
1124 | "Wrapper cannot be created because referenced symbol " |
1125 | "properties mismatch\n" ); |
1126 | } |
1127 | /* Do not turn function in one comdat group into wrapper to another |
1128 | comdat group. Other compiler producing the body of the |
1129 | another comdat group may make opposite decision and with unfortunate |
1130 | linker choices this may close a loop. */ |
1131 | else if (DECL_COMDAT_GROUP (original->decl) |
1132 | && DECL_COMDAT_GROUP (alias->decl) |
1133 | && (DECL_COMDAT_GROUP (alias->decl) |
1134 | != DECL_COMDAT_GROUP (original->decl))) |
1135 | { |
1136 | if (dump_enabled_p ()) |
1137 | dump_printf (MSG_MISSED_OPTIMIZATION, |
1138 | "Wrapper cannot be created because of COMDAT\n" ); |
1139 | } |
1140 | else if (DECL_STATIC_CHAIN (alias->decl) |
1141 | || DECL_STATIC_CHAIN (original->decl)) |
1142 | { |
1143 | if (dump_enabled_p ()) |
1144 | dump_printf (MSG_MISSED_OPTIMIZATION, |
1145 | "Cannot create wrapper of nested function.\n" ); |
1146 | } |
1147 | /* TODO: We can also deal with variadic functions never calling |
1148 | VA_START. */ |
1149 | else if (stdarg_p (TREE_TYPE (alias->decl))) |
1150 | { |
1151 | if (dump_enabled_p ()) |
1152 | dump_printf (MSG_MISSED_OPTIMIZATION, |
1153 | "cannot create wrapper of stdarg function.\n" ); |
1154 | } |
1155 | else if (ipa_fn_summaries |
1156 | && ipa_size_summaries->get (node: alias) != NULL |
1157 | && ipa_size_summaries->get (node: alias)->self_size <= 2) |
1158 | { |
1159 | if (dump_enabled_p ()) |
1160 | dump_printf (MSG_MISSED_OPTIMIZATION, "Wrapper creation is not " |
1161 | "profitable (function is too small).\n" ); |
1162 | } |
1163 | /* If user paid attention to mark function noinline, assume it is |
1164 | somewhat special and do not try to turn it into a wrapper that |
1165 | cannot be undone by inliner. */ |
1166 | else if (lookup_attribute (attr_name: "noinline" , DECL_ATTRIBUTES (alias->decl))) |
1167 | { |
1168 | if (dump_enabled_p ()) |
1169 | dump_printf (MSG_MISSED_OPTIMIZATION, |
1170 | "Wrappers are not created for noinline.\n" ); |
1171 | } |
1172 | else |
1173 | create_wrapper = true; |
1174 | |
1175 | /* We can redirect local calls in the case both alias and original |
1176 | are not interposable. */ |
1177 | redirect_callers |
1178 | = alias->get_availability () > AVAIL_INTERPOSABLE |
1179 | && original->get_availability () > AVAIL_INTERPOSABLE; |
1180 | /* TODO: We can redirect, but we need to produce alias of ORIGINAL |
1181 | with proper properties. */ |
1182 | if (!sem_item::compare_referenced_symbol_properties (NULL, n1: original, n2: alias, |
1183 | address: alias->address_taken)) |
1184 | redirect_callers = false; |
1185 | |
1186 | if (!redirect_callers && !create_wrapper) |
1187 | { |
1188 | if (dump_enabled_p ()) |
1189 | dump_printf (MSG_MISSED_OPTIMIZATION, |
1190 | "Not unifying; cannot redirect callers nor " |
1191 | "produce wrapper\n" ); |
1192 | return false; |
1193 | } |
1194 | |
1195 | /* Work out the symbol the wrapper should call. |
1196 | If ORIGINAL is interposable, we need to call a local alias. |
1197 | Also produce local alias (if possible) as an optimization. |
1198 | |
1199 | Local aliases cannot be created inside comdat groups because that |
1200 | prevents inlining. */ |
1201 | if (!original_discardable && !original->get_comdat_group ()) |
1202 | { |
1203 | local_original |
1204 | = dyn_cast <cgraph_node *> (p: original->noninterposable_alias ()); |
1205 | if (!local_original |
1206 | && original->get_availability () > AVAIL_INTERPOSABLE) |
1207 | local_original = original; |
1208 | } |
1209 | /* If we cannot use local alias, fallback to the original |
1210 | when possible. */ |
1211 | else if (original->get_availability () > AVAIL_INTERPOSABLE) |
1212 | local_original = original; |
1213 | |
1214 | /* If original is COMDAT local, we cannot really redirect calls outside |
1215 | of its comdat group to it. */ |
1216 | if (original->comdat_local_p ()) |
1217 | redirect_callers = false; |
1218 | if (!local_original) |
1219 | { |
1220 | if (dump_enabled_p ()) |
1221 | dump_printf (MSG_MISSED_OPTIMIZATION, |
1222 | "Not unifying; cannot produce local alias.\n" ); |
1223 | return false; |
1224 | } |
1225 | |
1226 | if (!redirect_callers && !create_wrapper) |
1227 | { |
1228 | if (dump_enabled_p ()) |
1229 | dump_printf (MSG_MISSED_OPTIMIZATION, |
1230 | "Not unifying; " |
1231 | "cannot redirect callers nor produce a wrapper\n" ); |
1232 | return false; |
1233 | } |
1234 | if (!create_wrapper |
1235 | && !alias->call_for_symbol_and_aliases (callback: cgraph_node::has_thunk_p, |
1236 | NULL, include_overwritable: true) |
1237 | && !alias->can_remove_if_no_direct_calls_p ()) |
1238 | { |
1239 | if (dump_enabled_p ()) |
1240 | dump_printf (MSG_MISSED_OPTIMIZATION, |
1241 | "Not unifying; cannot make wrapper and " |
1242 | "function has other uses than direct calls\n" ); |
1243 | return false; |
1244 | } |
1245 | } |
1246 | else |
1247 | create_alias = true; |
1248 | |
1249 | if (redirect_callers) |
1250 | { |
1251 | int nredirected = redirect_all_callers (n: alias, to: local_original); |
1252 | |
1253 | if (nredirected) |
1254 | { |
1255 | alias->icf_merged = true; |
1256 | local_original->icf_merged = true; |
1257 | |
1258 | if (dump_enabled_p ()) |
1259 | dump_printf (MSG_NOTE, |
1260 | "%i local calls have been " |
1261 | "redirected.\n" , nredirected); |
1262 | } |
1263 | |
1264 | /* If all callers was redirected, do not produce wrapper. */ |
1265 | if (alias->can_remove_if_no_direct_calls_p () |
1266 | && !DECL_VIRTUAL_P (alias->decl) |
1267 | && !alias->has_aliases_p ()) |
1268 | { |
1269 | create_wrapper = false; |
1270 | remove = true; |
1271 | } |
1272 | gcc_assert (!create_alias); |
1273 | } |
1274 | else if (create_alias) |
1275 | { |
1276 | alias->icf_merged = true; |
1277 | |
1278 | /* Remove the function's body. */ |
1279 | ipa_merge_profiles (dst: original, src: alias); |
1280 | symtab->call_cgraph_removal_hooks (node: alias); |
1281 | alias->release_body (keep_arguments: true); |
1282 | alias->reset (); |
1283 | /* Notice global symbol possibly produced RTL. */ |
1284 | ((symtab_node *)alias)->call_for_symbol_and_aliases (callback: clear_decl_rtl, |
1285 | NULL, include_overwritable: true); |
1286 | |
1287 | /* Create the alias. */ |
1288 | cgraph_node::create_alias (alias: alias_func->decl, target: decl); |
1289 | alias->resolve_alias (target: original); |
1290 | |
1291 | original->call_for_symbol_thunks_and_aliases |
1292 | (callback: set_local, data: (void *)(size_t) original->local_p (), include_overwritable: true); |
1293 | |
1294 | if (dump_enabled_p ()) |
1295 | dump_printf (MSG_OPTIMIZED_LOCATIONS, |
1296 | "Unified; Function alias has been created.\n" ); |
1297 | } |
1298 | if (create_wrapper) |
1299 | { |
1300 | gcc_assert (!create_alias); |
1301 | alias->icf_merged = true; |
1302 | symtab->call_cgraph_removal_hooks (node: alias); |
1303 | local_original->icf_merged = true; |
1304 | |
1305 | /* FIXME update local_original counts. */ |
1306 | ipa_merge_profiles (dst: original, src: alias, preserve_body: true); |
1307 | alias->create_wrapper (target: local_original); |
1308 | symtab->call_cgraph_insertion_hooks (node: alias); |
1309 | |
1310 | if (dump_enabled_p ()) |
1311 | dump_printf (MSG_OPTIMIZED_LOCATIONS, |
1312 | "Unified; Wrapper has been created.\n" ); |
1313 | } |
1314 | |
1315 | /* It's possible that redirection can hit thunks that block |
1316 | redirection opportunities. */ |
1317 | gcc_assert (alias->icf_merged || remove || redirect_callers); |
1318 | original->icf_merged = true; |
1319 | |
1320 | /* We use merged flag to track cases where COMDAT function is known to be |
1321 | compatible its callers. If we merged in non-COMDAT, we need to give up |
1322 | on this optimization. */ |
1323 | if (original->merged_comdat && !alias->merged_comdat) |
1324 | { |
1325 | if (dump_enabled_p ()) |
1326 | dump_printf (MSG_NOTE, "Dropping merged_comdat flag.\n" ); |
1327 | if (local_original) |
1328 | local_original->merged_comdat = false; |
1329 | original->merged_comdat = false; |
1330 | } |
1331 | |
1332 | if (remove) |
1333 | { |
1334 | ipa_merge_profiles (dst: original, src: alias); |
1335 | alias->release_body (); |
1336 | alias->reset (); |
1337 | alias->body_removed = true; |
1338 | alias->icf_merged = true; |
1339 | if (dump_enabled_p ()) |
1340 | dump_printf (MSG_OPTIMIZED_LOCATIONS, |
1341 | "Unified; Function body was removed.\n" ); |
1342 | } |
1343 | |
1344 | return true; |
1345 | } |
1346 | |
1347 | /* Semantic item initialization function. */ |
1348 | |
1349 | void |
1350 | sem_function::init (ipa_icf_gimple::func_checker *checker) |
1351 | { |
1352 | m_checker = checker; |
1353 | if (in_lto_p) |
1354 | get_node ()->get_untransformed_body (); |
1355 | |
1356 | tree fndecl = node->decl; |
1357 | function *func = DECL_STRUCT_FUNCTION (fndecl); |
1358 | |
1359 | gcc_assert (func); |
1360 | gcc_assert (SSANAMES (func)); |
1361 | |
1362 | ssa_names_size = SSANAMES (func)->length (); |
1363 | node = node; |
1364 | |
1365 | decl = fndecl; |
1366 | region_tree = func->eh->region_tree; |
1367 | |
1368 | /* iterating all function arguments. */ |
1369 | arg_count = count_formal_params (fndecl); |
1370 | |
1371 | edge_count = n_edges_for_fn (func); |
1372 | cgraph_node *cnode = dyn_cast <cgraph_node *> (p: node); |
1373 | if (!cnode->thunk) |
1374 | { |
1375 | cfg_checksum = coverage_compute_cfg_checksum (fn: func); |
1376 | |
1377 | inchash::hash hstate; |
1378 | |
1379 | basic_block bb; |
1380 | FOR_EACH_BB_FN (bb, func) |
1381 | { |
1382 | unsigned nondbg_stmt_count = 0; |
1383 | |
1384 | edge e; |
1385 | for (edge_iterator ei = ei_start (bb->preds); ei_cond (ei, p: &e); |
1386 | ei_next (i: &ei)) |
1387 | cfg_checksum = iterative_hash_host_wide_int (val: e->flags, |
1388 | val2: cfg_checksum); |
1389 | |
1390 | /* TODO: We should be able to match PHIs with different order of |
1391 | parameters. This needs to be also updated in |
1392 | sem_function::compare_phi_node. */ |
1393 | gphi_iterator si; |
1394 | for (si = gsi_start_nonvirtual_phis (bb); !gsi_end_p (i: si); |
1395 | gsi_next_nonvirtual_phi (i: &si)) |
1396 | { |
1397 | hstate.add_int (v: GIMPLE_PHI); |
1398 | gphi *phi = si.phi (); |
1399 | m_checker->hash_operand (gimple_phi_result (gs: phi), hstate, flags: 0, |
1400 | access: func_checker::OP_NORMAL); |
1401 | hstate.add_int (v: gimple_phi_num_args (gs: phi)); |
1402 | for (unsigned int i = 0; i < gimple_phi_num_args (gs: phi); i++) |
1403 | m_checker->hash_operand (gimple_phi_arg_def (gs: phi, index: i), |
1404 | hstate, flags: 0, access: func_checker::OP_NORMAL); |
1405 | } |
1406 | |
1407 | for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); |
1408 | gsi_next (i: &gsi)) |
1409 | { |
1410 | gimple *stmt = gsi_stmt (i: gsi); |
1411 | |
1412 | if (gimple_code (g: stmt) != GIMPLE_DEBUG |
1413 | && gimple_code (g: stmt) != GIMPLE_PREDICT) |
1414 | { |
1415 | hash_stmt (stmt, inchash&: hstate); |
1416 | nondbg_stmt_count++; |
1417 | } |
1418 | } |
1419 | |
1420 | hstate.commit_flag (); |
1421 | gcode_hash = hstate.end (); |
1422 | bb_sizes.safe_push (obj: nondbg_stmt_count); |
1423 | |
1424 | /* Inserting basic block to hash table. */ |
1425 | sem_bb *semantic_bb = new sem_bb (bb, nondbg_stmt_count, |
1426 | EDGE_COUNT (bb->preds) |
1427 | + EDGE_COUNT (bb->succs)); |
1428 | |
1429 | bb_sorted.safe_push (obj: semantic_bb); |
1430 | } |
1431 | } |
1432 | else |
1433 | { |
1434 | cfg_checksum = 0; |
1435 | gcode_hash = thunk_info::get (node: cnode)->hash (); |
1436 | } |
1437 | |
1438 | m_checker = NULL; |
1439 | } |
1440 | |
1441 | /* Improve accumulated hash for HSTATE based on a gimple statement STMT. */ |
1442 | |
1443 | void |
1444 | sem_function::hash_stmt (gimple *stmt, inchash::hash &hstate) |
1445 | { |
1446 | enum gimple_code code = gimple_code (g: stmt); |
1447 | |
1448 | hstate.add_int (v: code); |
1449 | |
1450 | switch (code) |
1451 | { |
1452 | case GIMPLE_SWITCH: |
1453 | m_checker->hash_operand (gimple_switch_index (gs: as_a <gswitch *> (p: stmt)), |
1454 | hstate, flags: 0, access: func_checker::OP_NORMAL); |
1455 | break; |
1456 | case GIMPLE_ASSIGN: |
1457 | hstate.add_int (v: gimple_assign_rhs_code (gs: stmt)); |
1458 | /* fall through */ |
1459 | case GIMPLE_CALL: |
1460 | case GIMPLE_ASM: |
1461 | case GIMPLE_COND: |
1462 | case GIMPLE_GOTO: |
1463 | case GIMPLE_RETURN: |
1464 | { |
1465 | func_checker::operand_access_type_map map (5); |
1466 | func_checker::classify_operands (stmt, map: &map); |
1467 | |
1468 | /* All these statements are equivalent if their operands are. */ |
1469 | for (unsigned i = 0; i < gimple_num_ops (gs: stmt); ++i) |
1470 | { |
1471 | func_checker::operand_access_type |
1472 | access_type = func_checker::get_operand_access_type |
1473 | (map: &map, gimple_op (gs: stmt, i)); |
1474 | m_checker->hash_operand (gimple_op (gs: stmt, i), hstate, flags: 0, |
1475 | access: access_type); |
1476 | /* For memory accesses when hasing for LTO stremaing record |
1477 | base and ref alias ptr types so we can compare them at WPA |
1478 | time without having to read actual function body. */ |
1479 | if (access_type == func_checker::OP_MEMORY |
1480 | && lto_streaming_expected_p () |
1481 | && flag_strict_aliasing) |
1482 | { |
1483 | ao_ref ref; |
1484 | |
1485 | ao_ref_init (&ref, gimple_op (gs: stmt, i)); |
1486 | tree t = ao_ref_alias_ptr_type (&ref); |
1487 | if (!variably_modified_type_p (t, NULL_TREE)) |
1488 | memory_access_types.safe_push (obj: t); |
1489 | t = ao_ref_base_alias_ptr_type (&ref); |
1490 | if (!variably_modified_type_p (t, NULL_TREE)) |
1491 | memory_access_types.safe_push (obj: t); |
1492 | } |
1493 | } |
1494 | /* Consider nocf_check attribute in hash as it affects code |
1495 | generation. */ |
1496 | if (code == GIMPLE_CALL |
1497 | && flag_cf_protection & CF_BRANCH) |
1498 | hstate.add_flag (flag: gimple_call_nocf_check_p (gs: as_a <gcall *> (p: stmt))); |
1499 | } |
1500 | break; |
1501 | default: |
1502 | break; |
1503 | } |
1504 | } |
1505 | |
1506 | |
1507 | /* Return true if polymorphic comparison must be processed. */ |
1508 | |
1509 | bool |
1510 | sem_function::compare_polymorphic_p (void) |
1511 | { |
1512 | struct cgraph_edge *e; |
1513 | |
1514 | if (!opt_for_fn (get_node ()->decl, flag_devirtualize)) |
1515 | return false; |
1516 | if (get_node ()->indirect_calls != NULL) |
1517 | return true; |
1518 | /* TODO: We can do simple propagation determining what calls may lead to |
1519 | a polymorphic call. */ |
1520 | for (e = get_node ()->callees; e; e = e->next_callee) |
1521 | if (e->callee->definition |
1522 | && opt_for_fn (e->callee->decl, flag_devirtualize)) |
1523 | return true; |
1524 | return false; |
1525 | } |
1526 | |
1527 | /* For a given call graph NODE, the function constructs new |
1528 | semantic function item. */ |
1529 | |
1530 | sem_function * |
1531 | sem_function::parse (cgraph_node *node, bitmap_obstack *stack, |
1532 | func_checker *checker) |
1533 | { |
1534 | tree fndecl = node->decl; |
1535 | function *func = DECL_STRUCT_FUNCTION (fndecl); |
1536 | |
1537 | if (!func || (!node->has_gimple_body_p () && !node->thunk)) |
1538 | return NULL; |
1539 | |
1540 | if (lookup_attribute_by_prefix (attr_name: "omp " , DECL_ATTRIBUTES (node->decl)) != NULL) |
1541 | return NULL; |
1542 | |
1543 | if (lookup_attribute_by_prefix (attr_name: "oacc " , |
1544 | DECL_ATTRIBUTES (node->decl)) != NULL) |
1545 | return NULL; |
1546 | |
1547 | /* PR ipa/70306. */ |
1548 | if (DECL_STATIC_CONSTRUCTOR (node->decl) |
1549 | || DECL_STATIC_DESTRUCTOR (node->decl)) |
1550 | return NULL; |
1551 | |
1552 | sem_function *f = new sem_function (node, stack); |
1553 | f->init (checker); |
1554 | |
1555 | return f; |
1556 | } |
1557 | |
1558 | /* For given basic blocks BB1 and BB2 (from functions FUNC1 and FUNC), |
1559 | return true if phi nodes are semantically equivalent in these blocks . */ |
1560 | |
1561 | bool |
1562 | sem_function::compare_phi_node (basic_block bb1, basic_block bb2) |
1563 | { |
1564 | gphi_iterator si1, si2; |
1565 | gphi *phi1, *phi2; |
1566 | unsigned size1, size2, i; |
1567 | tree t1, t2; |
1568 | edge e1, e2; |
1569 | |
1570 | gcc_assert (bb1 != NULL); |
1571 | gcc_assert (bb2 != NULL); |
1572 | |
1573 | si2 = gsi_start_nonvirtual_phis (bb: bb2); |
1574 | for (si1 = gsi_start_nonvirtual_phis (bb: bb1); !gsi_end_p (i: si1); |
1575 | gsi_next_nonvirtual_phi (i: &si1)) |
1576 | { |
1577 | if (gsi_end_p (i: si1) && gsi_end_p (i: si2)) |
1578 | break; |
1579 | |
1580 | if (gsi_end_p (i: si1) || gsi_end_p (i: si2)) |
1581 | return return_false(); |
1582 | |
1583 | phi1 = si1.phi (); |
1584 | phi2 = si2.phi (); |
1585 | |
1586 | tree phi_result1 = gimple_phi_result (gs: phi1); |
1587 | tree phi_result2 = gimple_phi_result (gs: phi2); |
1588 | |
1589 | if (!m_checker->compare_operand (t1: phi_result1, t2: phi_result2, |
1590 | type: func_checker::OP_NORMAL)) |
1591 | return return_false_with_msg ("PHI results are different" ); |
1592 | |
1593 | size1 = gimple_phi_num_args (gs: phi1); |
1594 | size2 = gimple_phi_num_args (gs: phi2); |
1595 | |
1596 | if (size1 != size2) |
1597 | return return_false (); |
1598 | |
1599 | /* TODO: We should be able to match PHIs with different order of |
1600 | parameters. This needs to be also updated in sem_function::init. */ |
1601 | for (i = 0; i < size1; ++i) |
1602 | { |
1603 | t1 = gimple_phi_arg (gs: phi1, index: i)->def; |
1604 | t2 = gimple_phi_arg (gs: phi2, index: i)->def; |
1605 | |
1606 | if (!m_checker->compare_operand (t1, t2, type: func_checker::OP_NORMAL)) |
1607 | return return_false (); |
1608 | |
1609 | e1 = gimple_phi_arg_edge (phi: phi1, i); |
1610 | e2 = gimple_phi_arg_edge (phi: phi2, i); |
1611 | |
1612 | if (!m_checker->compare_edge (e1, e2)) |
1613 | return return_false (); |
1614 | } |
1615 | |
1616 | gsi_next_nonvirtual_phi (i: &si2); |
1617 | } |
1618 | |
1619 | return true; |
1620 | } |
1621 | |
1622 | /* Basic blocks dictionary BB_DICT returns true if SOURCE index BB |
1623 | corresponds to TARGET. */ |
1624 | |
1625 | bool |
1626 | sem_function::bb_dict_test (vec<int> *bb_dict, int source, int target) |
1627 | { |
1628 | source++; |
1629 | target++; |
1630 | |
1631 | if (bb_dict->length () <= (unsigned)source) |
1632 | bb_dict->safe_grow_cleared (len: source + 1, exact: true); |
1633 | |
1634 | if ((*bb_dict)[source] == 0) |
1635 | { |
1636 | (*bb_dict)[source] = target; |
1637 | return true; |
1638 | } |
1639 | else |
1640 | return (*bb_dict)[source] == target; |
1641 | } |
1642 | |
1643 | sem_variable::sem_variable (bitmap_obstack *stack): sem_item (VAR, stack) |
1644 | { |
1645 | } |
1646 | |
1647 | sem_variable::sem_variable (varpool_node *node, bitmap_obstack *stack) |
1648 | : sem_item (VAR, node, stack) |
1649 | { |
1650 | gcc_checking_assert (node); |
1651 | gcc_checking_assert (get_node ()); |
1652 | } |
1653 | |
1654 | /* Fast equality function based on knowledge known in WPA. */ |
1655 | |
1656 | bool |
1657 | sem_variable::equals_wpa (sem_item *item, |
1658 | hash_map <symtab_node *, sem_item *> &ignored_nodes) |
1659 | { |
1660 | gcc_assert (item->type == VAR); |
1661 | |
1662 | if (node->num_references () != item->node->num_references ()) |
1663 | return return_false_with_msg ("different number of references" ); |
1664 | |
1665 | if (DECL_TLS_MODEL (decl) || DECL_TLS_MODEL (item->decl)) |
1666 | return return_false_with_msg ("TLS model" ); |
1667 | |
1668 | /* DECL_ALIGN is safe to merge, because we will always chose the largest |
1669 | alignment out of all aliases. */ |
1670 | |
1671 | if (DECL_VIRTUAL_P (decl) != DECL_VIRTUAL_P (item->decl)) |
1672 | return return_false_with_msg ("Virtual flag mismatch" ); |
1673 | |
1674 | if (DECL_SIZE (decl) != DECL_SIZE (item->decl) |
1675 | && ((!DECL_SIZE (decl) || !DECL_SIZE (item->decl)) |
1676 | || !operand_equal_p (DECL_SIZE (decl), |
1677 | DECL_SIZE (item->decl), flags: OEP_ONLY_CONST))) |
1678 | return return_false_with_msg ("size mismatch" ); |
1679 | |
1680 | /* Do not attempt to mix data from different user sections; |
1681 | we do not know what user intends with those. */ |
1682 | if (((DECL_SECTION_NAME (decl) && !node->implicit_section) |
1683 | || (DECL_SECTION_NAME (item->decl) && !item->node->implicit_section)) |
1684 | && DECL_SECTION_NAME (decl) != DECL_SECTION_NAME (item->decl)) |
1685 | return return_false_with_msg ("user section mismatch" ); |
1686 | |
1687 | if (DECL_IN_TEXT_SECTION (decl) != DECL_IN_TEXT_SECTION (item->decl)) |
1688 | return return_false_with_msg ("text section" ); |
1689 | |
1690 | if (TYPE_ADDR_SPACE (TREE_TYPE (decl)) |
1691 | != TYPE_ADDR_SPACE (TREE_TYPE (item->decl))) |
1692 | return return_false_with_msg ("address-space" ); |
1693 | |
1694 | ipa_ref *ref = NULL, *ref2 = NULL; |
1695 | for (unsigned i = 0; node->iterate_reference (i, ref); i++) |
1696 | { |
1697 | item->node->iterate_reference (i, ref&: ref2); |
1698 | |
1699 | if (ref->use != ref2->use) |
1700 | return return_false_with_msg ("reference use mismatch" ); |
1701 | |
1702 | if (!compare_symbol_references (ignored_nodes, |
1703 | n1: ref->referred, n2: ref2->referred, |
1704 | address: ref->address_matters_p ())) |
1705 | return false; |
1706 | } |
1707 | |
1708 | return true; |
1709 | } |
1710 | |
1711 | /* Returns true if the item equals to ITEM given as argument. */ |
1712 | |
1713 | bool |
1714 | sem_variable::equals (sem_item *item, |
1715 | hash_map <symtab_node *, sem_item *> &) |
1716 | { |
1717 | gcc_assert (item->type == VAR); |
1718 | bool ret; |
1719 | |
1720 | if (DECL_INITIAL (decl) == error_mark_node && in_lto_p) |
1721 | dyn_cast <varpool_node *>(p: node)->get_constructor (); |
1722 | if (DECL_INITIAL (item->decl) == error_mark_node && in_lto_p) |
1723 | dyn_cast <varpool_node *>(p: item->node)->get_constructor (); |
1724 | |
1725 | /* As seen in PR ipa/65303 we have to compare variables types. */ |
1726 | if (!func_checker::compatible_types_p (TREE_TYPE (decl), |
1727 | TREE_TYPE (item->decl))) |
1728 | return return_false_with_msg ("variables types are different" ); |
1729 | |
1730 | ret = sem_variable::equals (DECL_INITIAL (decl), |
1731 | DECL_INITIAL (item->node->decl)); |
1732 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1733 | fprintf (stream: dump_file, |
1734 | format: "Equals called for vars: %s:%s with result: %s\n\n" , |
1735 | node->dump_name (), item->node->dump_name (), |
1736 | ret ? "true" : "false" ); |
1737 | |
1738 | return ret; |
1739 | } |
1740 | |
1741 | /* Compares trees T1 and T2 for semantic equality. */ |
1742 | |
1743 | bool |
1744 | sem_variable::equals (tree t1, tree t2) |
1745 | { |
1746 | if (!t1 || !t2) |
1747 | return return_with_debug (t1 == t2); |
1748 | if (t1 == t2) |
1749 | return true; |
1750 | tree_code tc1 = TREE_CODE (t1); |
1751 | tree_code tc2 = TREE_CODE (t2); |
1752 | |
1753 | if (tc1 != tc2) |
1754 | return return_false_with_msg ("TREE_CODE mismatch" ); |
1755 | |
1756 | switch (tc1) |
1757 | { |
1758 | case CONSTRUCTOR: |
1759 | { |
1760 | vec<constructor_elt, va_gc> *v1, *v2; |
1761 | unsigned HOST_WIDE_INT idx; |
1762 | |
1763 | enum tree_code typecode = TREE_CODE (TREE_TYPE (t1)); |
1764 | if (typecode != TREE_CODE (TREE_TYPE (t2))) |
1765 | return return_false_with_msg ("constructor type mismatch" ); |
1766 | |
1767 | if (typecode == ARRAY_TYPE) |
1768 | { |
1769 | HOST_WIDE_INT size_1 = int_size_in_bytes (TREE_TYPE (t1)); |
1770 | /* For arrays, check that the sizes all match. */ |
1771 | if (TYPE_MODE (TREE_TYPE (t1)) != TYPE_MODE (TREE_TYPE (t2)) |
1772 | || size_1 == -1 |
1773 | || size_1 != int_size_in_bytes (TREE_TYPE (t2))) |
1774 | return return_false_with_msg ("constructor array size mismatch" ); |
1775 | } |
1776 | else if (!func_checker::compatible_types_p (TREE_TYPE (t1), |
1777 | TREE_TYPE (t2))) |
1778 | return return_false_with_msg ("constructor type incompatible" ); |
1779 | |
1780 | v1 = CONSTRUCTOR_ELTS (t1); |
1781 | v2 = CONSTRUCTOR_ELTS (t2); |
1782 | if (vec_safe_length (v: v1) != vec_safe_length (v: v2)) |
1783 | return return_false_with_msg ("constructor number of elts mismatch" ); |
1784 | |
1785 | for (idx = 0; idx < vec_safe_length (v: v1); ++idx) |
1786 | { |
1787 | constructor_elt *c1 = &(*v1)[idx]; |
1788 | constructor_elt *c2 = &(*v2)[idx]; |
1789 | |
1790 | /* Check that each value is the same... */ |
1791 | if (!sem_variable::equals (t1: c1->value, t2: c2->value)) |
1792 | return false; |
1793 | /* ... and that they apply to the same fields! */ |
1794 | if (!sem_variable::equals (t1: c1->index, t2: c2->index)) |
1795 | return false; |
1796 | } |
1797 | return true; |
1798 | } |
1799 | case MEM_REF: |
1800 | { |
1801 | tree x1 = TREE_OPERAND (t1, 0); |
1802 | tree x2 = TREE_OPERAND (t2, 0); |
1803 | tree y1 = TREE_OPERAND (t1, 1); |
1804 | tree y2 = TREE_OPERAND (t2, 1); |
1805 | |
1806 | if (!func_checker::compatible_types_p (TREE_TYPE (x1), TREE_TYPE (x2))) |
1807 | return return_false (); |
1808 | |
1809 | /* Type of the offset on MEM_REF does not matter. */ |
1810 | return return_with_debug (sem_variable::equals (x1, x2) |
1811 | && known_eq (wi::to_poly_offset (y1), |
1812 | wi::to_poly_offset (y2))); |
1813 | } |
1814 | case ADDR_EXPR: |
1815 | case FDESC_EXPR: |
1816 | { |
1817 | tree op1 = TREE_OPERAND (t1, 0); |
1818 | tree op2 = TREE_OPERAND (t2, 0); |
1819 | return sem_variable::equals (t1: op1, t2: op2); |
1820 | } |
1821 | /* References to other vars/decls are compared using ipa-ref. */ |
1822 | case FUNCTION_DECL: |
1823 | case VAR_DECL: |
1824 | if (decl_in_symtab_p (decl: t1) && decl_in_symtab_p (decl: t2)) |
1825 | return true; |
1826 | return return_false_with_msg ("Declaration mismatch" ); |
1827 | case CONST_DECL: |
1828 | /* TODO: We can check CONST_DECL by its DECL_INITIAL, but for that we |
1829 | need to process its VAR/FUNCTION references without relying on ipa-ref |
1830 | compare. */ |
1831 | case FIELD_DECL: |
1832 | case LABEL_DECL: |
1833 | return return_false_with_msg ("Declaration mismatch" ); |
1834 | case INTEGER_CST: |
1835 | /* Integer constants are the same only if the same width of type. */ |
1836 | if (TYPE_PRECISION (TREE_TYPE (t1)) != TYPE_PRECISION (TREE_TYPE (t2))) |
1837 | return return_false_with_msg ("INTEGER_CST precision mismatch" ); |
1838 | if (TYPE_MODE (TREE_TYPE (t1)) != TYPE_MODE (TREE_TYPE (t2))) |
1839 | return return_false_with_msg ("INTEGER_CST mode mismatch" ); |
1840 | return return_with_debug (tree_int_cst_equal (t1, t2)); |
1841 | case STRING_CST: |
1842 | if (TYPE_MODE (TREE_TYPE (t1)) != TYPE_MODE (TREE_TYPE (t2))) |
1843 | return return_false_with_msg ("STRING_CST mode mismatch" ); |
1844 | if (TREE_STRING_LENGTH (t1) != TREE_STRING_LENGTH (t2)) |
1845 | return return_false_with_msg ("STRING_CST length mismatch" ); |
1846 | if (memcmp (TREE_STRING_POINTER (t1), TREE_STRING_POINTER (t2), |
1847 | TREE_STRING_LENGTH (t1))) |
1848 | return return_false_with_msg ("STRING_CST mismatch" ); |
1849 | return true; |
1850 | case FIXED_CST: |
1851 | /* Fixed constants are the same only if the same width of type. */ |
1852 | if (TYPE_PRECISION (TREE_TYPE (t1)) != TYPE_PRECISION (TREE_TYPE (t2))) |
1853 | return return_false_with_msg ("FIXED_CST precision mismatch" ); |
1854 | |
1855 | return return_with_debug (FIXED_VALUES_IDENTICAL (TREE_FIXED_CST (t1), |
1856 | TREE_FIXED_CST (t2))); |
1857 | case COMPLEX_CST: |
1858 | return (sem_variable::equals (TREE_REALPART (t1), TREE_REALPART (t2)) |
1859 | && sem_variable::equals (TREE_IMAGPART (t1), TREE_IMAGPART (t2))); |
1860 | case REAL_CST: |
1861 | /* Real constants are the same only if the same width of type. */ |
1862 | if (TYPE_PRECISION (TREE_TYPE (t1)) != TYPE_PRECISION (TREE_TYPE (t2))) |
1863 | return return_false_with_msg ("REAL_CST precision mismatch" ); |
1864 | return return_with_debug (real_identical (&TREE_REAL_CST (t1), |
1865 | &TREE_REAL_CST (t2))); |
1866 | case VECTOR_CST: |
1867 | { |
1868 | if (maybe_ne (VECTOR_CST_NELTS (t1), VECTOR_CST_NELTS (t2))) |
1869 | return return_false_with_msg ("VECTOR_CST nelts mismatch" ); |
1870 | |
1871 | unsigned int count |
1872 | = tree_vector_builder::binary_encoded_nelts (vec1: t1, vec2: t2); |
1873 | for (unsigned int i = 0; i < count; ++i) |
1874 | if (!sem_variable::equals (VECTOR_CST_ENCODED_ELT (t1, i), |
1875 | VECTOR_CST_ENCODED_ELT (t2, i))) |
1876 | return false; |
1877 | |
1878 | return true; |
1879 | } |
1880 | case ARRAY_REF: |
1881 | case ARRAY_RANGE_REF: |
1882 | { |
1883 | tree x1 = TREE_OPERAND (t1, 0); |
1884 | tree x2 = TREE_OPERAND (t2, 0); |
1885 | tree y1 = TREE_OPERAND (t1, 1); |
1886 | tree y2 = TREE_OPERAND (t2, 1); |
1887 | |
1888 | if (!sem_variable::equals (t1: x1, t2: x2) || !sem_variable::equals (t1: y1, t2: y2)) |
1889 | return false; |
1890 | if (!sem_variable::equals (t1: array_ref_low_bound (t1), |
1891 | t2: array_ref_low_bound (t2))) |
1892 | return false; |
1893 | if (!sem_variable::equals (t1: array_ref_element_size (t1), |
1894 | t2: array_ref_element_size (t2))) |
1895 | return false; |
1896 | return true; |
1897 | } |
1898 | |
1899 | case COMPONENT_REF: |
1900 | case POINTER_PLUS_EXPR: |
1901 | case PLUS_EXPR: |
1902 | case MINUS_EXPR: |
1903 | case RANGE_EXPR: |
1904 | { |
1905 | tree x1 = TREE_OPERAND (t1, 0); |
1906 | tree x2 = TREE_OPERAND (t2, 0); |
1907 | tree y1 = TREE_OPERAND (t1, 1); |
1908 | tree y2 = TREE_OPERAND (t2, 1); |
1909 | |
1910 | return sem_variable::equals (t1: x1, t2: x2) && sem_variable::equals (t1: y1, t2: y2); |
1911 | } |
1912 | |
1913 | CASE_CONVERT: |
1914 | case VIEW_CONVERT_EXPR: |
1915 | if (!func_checker::compatible_types_p (TREE_TYPE (t1), TREE_TYPE (t2))) |
1916 | return return_false (); |
1917 | return sem_variable::equals (TREE_OPERAND (t1, 0), TREE_OPERAND (t2, 0)); |
1918 | case ERROR_MARK: |
1919 | return return_false_with_msg ("ERROR_MARK" ); |
1920 | default: |
1921 | return return_false_with_msg ("Unknown TREE code reached" ); |
1922 | } |
1923 | } |
1924 | |
1925 | /* Parser function that visits a varpool NODE. */ |
1926 | |
1927 | sem_variable * |
1928 | sem_variable::parse (varpool_node *node, bitmap_obstack *stack, |
1929 | func_checker *checker) |
1930 | { |
1931 | if (TREE_THIS_VOLATILE (node->decl) || DECL_HARD_REGISTER (node->decl) |
1932 | || node->alias) |
1933 | return NULL; |
1934 | |
1935 | sem_variable *v = new sem_variable (node, stack); |
1936 | v->init (checker); |
1937 | |
1938 | return v; |
1939 | } |
1940 | |
1941 | /* Semantic variable initialization function. */ |
1942 | |
1943 | void |
1944 | sem_variable::init (ipa_icf_gimple::func_checker *checker) |
1945 | { |
1946 | decl = get_node ()->decl; |
1947 | |
1948 | /* All WPA streamed in symbols should have their hashes computed at compile |
1949 | time. At this point, the constructor may not be in memory at all. |
1950 | DECL_INITIAL (decl) would be error_mark_node in that case. */ |
1951 | if (!m_hash_set) |
1952 | { |
1953 | gcc_assert (!node->lto_file_data); |
1954 | inchash::hash hstate; |
1955 | hstate.add_int (v: 456346417); |
1956 | checker->hash_operand (DECL_INITIAL (decl), hstate, flags: 0); |
1957 | set_hash (hstate.end ()); |
1958 | } |
1959 | } |
1960 | |
1961 | /* References independent hash function. */ |
1962 | |
1963 | hashval_t |
1964 | sem_variable::get_hash (void) |
1965 | { |
1966 | gcc_checking_assert (m_hash_set); |
1967 | return m_hash; |
1968 | } |
1969 | |
1970 | /* Merges instance with an ALIAS_ITEM, where alias, thunk or redirection can |
1971 | be applied. */ |
1972 | |
1973 | bool |
1974 | sem_variable::merge (sem_item *alias_item) |
1975 | { |
1976 | gcc_assert (alias_item->type == VAR); |
1977 | |
1978 | AUTO_DUMP_SCOPE ("merge" , |
1979 | dump_user_location_t::from_function_decl (decl)); |
1980 | if (!sem_item::target_supports_symbol_aliases_p ()) |
1981 | { |
1982 | if (dump_enabled_p ()) |
1983 | dump_printf (MSG_MISSED_OPTIMIZATION, "Not unifying; " |
1984 | "Symbol aliases are not supported by target\n" ); |
1985 | return false; |
1986 | } |
1987 | |
1988 | if (DECL_EXTERNAL (alias_item->decl)) |
1989 | { |
1990 | if (dump_enabled_p ()) |
1991 | dump_printf (MSG_MISSED_OPTIMIZATION, |
1992 | "Not unifying; alias is external.\n" ); |
1993 | return false; |
1994 | } |
1995 | |
1996 | sem_variable *alias_var = static_cast<sem_variable *> (alias_item); |
1997 | |
1998 | varpool_node *original = get_node (); |
1999 | varpool_node *alias = alias_var->get_node (); |
2000 | bool original_discardable = false; |
2001 | |
2002 | bool alias_address_matters = alias->address_matters_p (); |
2003 | |
2004 | /* See if original is in a section that can be discarded if the main |
2005 | symbol is not used. |
2006 | Also consider case where we have resolution info and we know that |
2007 | original's definition is not going to be used. In this case we cannot |
2008 | create alias to original. */ |
2009 | if (original->can_be_discarded_p () |
2010 | || (node->resolution != LDPR_UNKNOWN |
2011 | && !decl_binds_to_current_def_p (node->decl))) |
2012 | original_discardable = true; |
2013 | |
2014 | gcc_assert (!TREE_ASM_WRITTEN (alias->decl)); |
2015 | |
2016 | /* Constant pool machinery is not quite ready for aliases. |
2017 | TODO: varasm code contains logic for merging DECL_IN_CONSTANT_POOL. |
2018 | For LTO merging does not happen that is an important missing feature. |
2019 | We can enable merging with LTO if the DECL_IN_CONSTANT_POOL |
2020 | flag is dropped and non-local symbol name is assigned. */ |
2021 | if (DECL_IN_CONSTANT_POOL (alias->decl) |
2022 | || DECL_IN_CONSTANT_POOL (original->decl)) |
2023 | { |
2024 | if (dump_enabled_p ()) |
2025 | dump_printf (MSG_MISSED_OPTIMIZATION, |
2026 | "Not unifying; constant pool variables.\n" ); |
2027 | return false; |
2028 | } |
2029 | |
2030 | /* Do not attempt to mix functions from different user sections; |
2031 | we do not know what user intends with those. */ |
2032 | if (((DECL_SECTION_NAME (original->decl) && !original->implicit_section) |
2033 | || (DECL_SECTION_NAME (alias->decl) && !alias->implicit_section)) |
2034 | && DECL_SECTION_NAME (original->decl) != DECL_SECTION_NAME (alias->decl)) |
2035 | { |
2036 | if (dump_enabled_p ()) |
2037 | dump_printf (MSG_MISSED_OPTIMIZATION, |
2038 | "Not unifying; " |
2039 | "original and alias are in different sections.\n" ); |
2040 | return false; |
2041 | } |
2042 | |
2043 | /* We cannot merge if address comparison matters. */ |
2044 | if (alias_address_matters && flag_merge_constants < 2) |
2045 | { |
2046 | if (dump_enabled_p ()) |
2047 | dump_printf (MSG_MISSED_OPTIMIZATION, |
2048 | "Not unifying; address of original may be compared.\n" ); |
2049 | return false; |
2050 | } |
2051 | |
2052 | if (DECL_ALIGN (original->decl) != DECL_ALIGN (alias->decl) |
2053 | && (sanitize_flags_p (flag: SANITIZE_ADDRESS, fn: original->decl) |
2054 | || sanitize_flags_p (flag: SANITIZE_ADDRESS, fn: alias->decl))) |
2055 | { |
2056 | if (dump_enabled_p ()) |
2057 | dump_printf (MSG_MISSED_OPTIMIZATION, |
2058 | "Not unifying; " |
2059 | "ASAN requires equal alignments for original and alias\n" ); |
2060 | |
2061 | return false; |
2062 | } |
2063 | |
2064 | if (DECL_ALIGN (original->decl) < DECL_ALIGN (alias->decl)) |
2065 | { |
2066 | if (dump_enabled_p ()) |
2067 | dump_printf (MSG_MISSED_OPTIMIZATION, |
2068 | "Not unifying; " |
2069 | "original and alias have incompatible alignments\n" ); |
2070 | |
2071 | return false; |
2072 | } |
2073 | |
2074 | if (DECL_COMDAT_GROUP (original->decl) != DECL_COMDAT_GROUP (alias->decl)) |
2075 | { |
2076 | if (dump_enabled_p ()) |
2077 | dump_printf (MSG_MISSED_OPTIMIZATION, |
2078 | "Not unifying; alias cannot be created; " |
2079 | "across comdat group boundary\n" ); |
2080 | |
2081 | return false; |
2082 | } |
2083 | |
2084 | if (original_discardable) |
2085 | { |
2086 | if (dump_enabled_p ()) |
2087 | dump_printf (MSG_MISSED_OPTIMIZATION, |
2088 | "Not unifying; alias cannot be created; " |
2089 | "target is discardable\n" ); |
2090 | |
2091 | return false; |
2092 | } |
2093 | else |
2094 | { |
2095 | gcc_assert (!original->alias); |
2096 | gcc_assert (!alias->alias); |
2097 | |
2098 | alias->analyzed = false; |
2099 | |
2100 | DECL_INITIAL (alias->decl) = NULL; |
2101 | ((symtab_node *)alias)->call_for_symbol_and_aliases (callback: clear_decl_rtl, |
2102 | NULL, include_overwritable: true); |
2103 | alias->remove_all_references (); |
2104 | if (TREE_ADDRESSABLE (alias->decl)) |
2105 | original->call_for_symbol_and_aliases (callback: set_addressable, NULL, include_overwritable: true); |
2106 | |
2107 | varpool_node::create_alias (alias_var->decl, decl); |
2108 | alias->resolve_alias (target: original); |
2109 | |
2110 | if (dump_enabled_p ()) |
2111 | dump_printf (MSG_OPTIMIZED_LOCATIONS, |
2112 | "Unified; Variable alias has been created.\n" ); |
2113 | |
2114 | return true; |
2115 | } |
2116 | } |
2117 | |
2118 | /* Dump symbol to FILE. */ |
2119 | |
2120 | void |
2121 | sem_variable::dump_to_file (FILE *file) |
2122 | { |
2123 | gcc_assert (file); |
2124 | |
2125 | print_node (file, "" , decl, 0); |
2126 | fprintf (stream: file, format: "\n\n" ); |
2127 | } |
2128 | |
2129 | unsigned int sem_item_optimizer::class_id = 0; |
2130 | |
2131 | sem_item_optimizer::sem_item_optimizer () |
2132 | : worklist (0), m_classes (0), m_classes_count (0), m_cgraph_node_hooks (NULL), |
2133 | m_varpool_node_hooks (NULL), m_merged_variables (), m_references () |
2134 | { |
2135 | m_items.create (nelems: 0); |
2136 | bitmap_obstack_initialize (&m_bmstack); |
2137 | } |
2138 | |
2139 | sem_item_optimizer::~sem_item_optimizer () |
2140 | { |
2141 | for (unsigned int i = 0; i < m_items.length (); i++) |
2142 | delete m_items[i]; |
2143 | |
2144 | |
2145 | for (hash_table<congruence_class_hash>::iterator it = m_classes.begin (); |
2146 | it != m_classes.end (); ++it) |
2147 | { |
2148 | for (unsigned int i = 0; i < (*it)->classes.length (); i++) |
2149 | delete (*it)->classes[i]; |
2150 | |
2151 | (*it)->classes.release (); |
2152 | free (ptr: *it); |
2153 | } |
2154 | |
2155 | m_items.release (); |
2156 | |
2157 | bitmap_obstack_release (&m_bmstack); |
2158 | m_merged_variables.release (); |
2159 | } |
2160 | |
2161 | /* Write IPA ICF summary for symbols. */ |
2162 | |
2163 | void |
2164 | sem_item_optimizer::write_summary (void) |
2165 | { |
2166 | unsigned int count = 0; |
2167 | |
2168 | output_block *ob = create_output_block (LTO_section_ipa_icf); |
2169 | lto_symtab_encoder_t encoder = ob->decl_state->symtab_node_encoder; |
2170 | ob->symbol = NULL; |
2171 | |
2172 | /* Calculate number of symbols to be serialized. */ |
2173 | for (lto_symtab_encoder_iterator lsei = lsei_start_in_partition (encoder); |
2174 | !lsei_end_p (lsei); |
2175 | lsei_next_in_partition (lsei: &lsei)) |
2176 | { |
2177 | symtab_node *node = lsei_node (lsei); |
2178 | |
2179 | if (m_symtab_node_map.get (k: node)) |
2180 | count++; |
2181 | } |
2182 | |
2183 | streamer_write_uhwi (ob, count); |
2184 | |
2185 | /* Process all of the symbols. */ |
2186 | for (lto_symtab_encoder_iterator lsei = lsei_start_in_partition (encoder); |
2187 | !lsei_end_p (lsei); |
2188 | lsei_next_in_partition (lsei: &lsei)) |
2189 | { |
2190 | symtab_node *node = lsei_node (lsei); |
2191 | |
2192 | sem_item **item = m_symtab_node_map.get (k: node); |
2193 | |
2194 | if (item && *item) |
2195 | { |
2196 | int node_ref = lto_symtab_encoder_encode (encoder, node); |
2197 | streamer_write_uhwi_stream (ob->main_stream, node_ref); |
2198 | |
2199 | streamer_write_uhwi (ob, (*item)->get_hash ()); |
2200 | |
2201 | if ((*item)->type == FUNC) |
2202 | { |
2203 | sem_function *fn = static_cast<sem_function *> (*item); |
2204 | streamer_write_uhwi (ob, fn->memory_access_types.length ()); |
2205 | for (unsigned i = 0; i < fn->memory_access_types.length (); i++) |
2206 | stream_write_tree (ob, fn->memory_access_types[i], true); |
2207 | } |
2208 | } |
2209 | } |
2210 | |
2211 | streamer_write_char_stream (obs: ob->main_stream, c: 0); |
2212 | produce_asm (ob, NULL); |
2213 | destroy_output_block (ob); |
2214 | } |
2215 | |
2216 | /* Reads a section from LTO stream file FILE_DATA. Input block for DATA |
2217 | contains LEN bytes. */ |
2218 | |
2219 | void |
2220 | sem_item_optimizer::read_section (lto_file_decl_data *file_data, |
2221 | const char *data, size_t len) |
2222 | { |
2223 | const lto_function_header * |
2224 | = (const lto_function_header *) data; |
2225 | const int cfg_offset = sizeof (lto_function_header); |
2226 | const int main_offset = cfg_offset + header->cfg_size; |
2227 | const int string_offset = main_offset + header->main_size; |
2228 | data_in *data_in; |
2229 | unsigned int i; |
2230 | unsigned int count; |
2231 | |
2232 | lto_input_block ib_main ((const char *) data + main_offset, 0, |
2233 | header->main_size, file_data); |
2234 | |
2235 | data_in |
2236 | = lto_data_in_create (file_data, (const char *) data + string_offset, |
2237 | header->string_size, vNULL); |
2238 | |
2239 | count = streamer_read_uhwi (&ib_main); |
2240 | |
2241 | for (i = 0; i < count; i++) |
2242 | { |
2243 | unsigned int index; |
2244 | symtab_node *node; |
2245 | lto_symtab_encoder_t encoder; |
2246 | |
2247 | index = streamer_read_uhwi (&ib_main); |
2248 | encoder = file_data->symtab_node_encoder; |
2249 | node = lto_symtab_encoder_deref (encoder, ref: index); |
2250 | |
2251 | hashval_t hash = streamer_read_uhwi (&ib_main); |
2252 | gcc_assert (node->definition); |
2253 | |
2254 | if (is_a<cgraph_node *> (p: node)) |
2255 | { |
2256 | cgraph_node *cnode = dyn_cast <cgraph_node *> (p: node); |
2257 | |
2258 | sem_function *fn = new sem_function (cnode, &m_bmstack); |
2259 | unsigned count = streamer_read_uhwi (&ib_main); |
2260 | inchash::hash hstate (0); |
2261 | if (flag_incremental_link == INCREMENTAL_LINK_LTO) |
2262 | fn->memory_access_types.reserve_exact (nelems: count); |
2263 | for (unsigned i = 0; i < count; i++) |
2264 | { |
2265 | tree type = stream_read_tree (&ib_main, data_in); |
2266 | hstate.add_int (v: get_deref_alias_set (type)); |
2267 | if (flag_incremental_link == INCREMENTAL_LINK_LTO) |
2268 | fn->memory_access_types.quick_push (obj: type); |
2269 | } |
2270 | fn->m_alias_sets_hash = hstate.end (); |
2271 | fn->set_hash (hash); |
2272 | m_items.safe_push (obj: fn); |
2273 | } |
2274 | else |
2275 | { |
2276 | varpool_node *vnode = dyn_cast <varpool_node *> (p: node); |
2277 | |
2278 | sem_variable *var = new sem_variable (vnode, &m_bmstack); |
2279 | var->set_hash (hash); |
2280 | m_items.safe_push (obj: var); |
2281 | } |
2282 | } |
2283 | |
2284 | lto_free_section_data (file_data, LTO_section_ipa_icf, NULL, data, |
2285 | len); |
2286 | lto_data_in_delete (data_in); |
2287 | } |
2288 | |
2289 | /* Read IPA ICF summary for symbols. */ |
2290 | |
2291 | void |
2292 | sem_item_optimizer::read_summary (void) |
2293 | { |
2294 | lto_file_decl_data **file_data_vec = lto_get_file_decl_data (); |
2295 | lto_file_decl_data *file_data; |
2296 | unsigned int j = 0; |
2297 | |
2298 | while ((file_data = file_data_vec[j++])) |
2299 | { |
2300 | size_t len; |
2301 | const char *data |
2302 | = lto_get_summary_section_data (file_data, LTO_section_ipa_icf, &len); |
2303 | if (data) |
2304 | read_section (file_data, data, len); |
2305 | } |
2306 | } |
2307 | |
2308 | /* Register callgraph and varpool hooks. */ |
2309 | |
2310 | void |
2311 | sem_item_optimizer::register_hooks (void) |
2312 | { |
2313 | if (!m_cgraph_node_hooks) |
2314 | m_cgraph_node_hooks = symtab->add_cgraph_removal_hook |
2315 | (hook: &sem_item_optimizer::cgraph_removal_hook, data: this); |
2316 | |
2317 | if (!m_varpool_node_hooks) |
2318 | m_varpool_node_hooks = symtab->add_varpool_removal_hook |
2319 | (hook: &sem_item_optimizer::varpool_removal_hook, data: this); |
2320 | } |
2321 | |
2322 | /* Unregister callgraph and varpool hooks. */ |
2323 | |
2324 | void |
2325 | sem_item_optimizer::unregister_hooks (void) |
2326 | { |
2327 | if (m_cgraph_node_hooks) |
2328 | symtab->remove_cgraph_removal_hook (entry: m_cgraph_node_hooks); |
2329 | |
2330 | if (m_varpool_node_hooks) |
2331 | symtab->remove_varpool_removal_hook (entry: m_varpool_node_hooks); |
2332 | } |
2333 | |
2334 | /* Adds a CLS to hashtable associated by hash value. */ |
2335 | |
2336 | void |
2337 | sem_item_optimizer::add_class (congruence_class *cls) |
2338 | { |
2339 | gcc_assert (cls->members.length ()); |
2340 | |
2341 | congruence_class_group *group |
2342 | = get_group_by_hash (hash: cls->members[0]->get_hash (), |
2343 | type: cls->members[0]->type); |
2344 | group->classes.safe_push (obj: cls); |
2345 | } |
2346 | |
2347 | /* Gets a congruence class group based on given HASH value and TYPE. */ |
2348 | |
2349 | congruence_class_group * |
2350 | sem_item_optimizer::get_group_by_hash (hashval_t hash, sem_item_type type) |
2351 | { |
2352 | congruence_class_group *item = XNEW (congruence_class_group); |
2353 | item->hash = hash; |
2354 | item->type = type; |
2355 | |
2356 | congruence_class_group **slot = m_classes.find_slot (value: item, insert: INSERT); |
2357 | |
2358 | if (*slot) |
2359 | free (ptr: item); |
2360 | else |
2361 | { |
2362 | item->classes.create (nelems: 1); |
2363 | *slot = item; |
2364 | } |
2365 | |
2366 | return *slot; |
2367 | } |
2368 | |
2369 | /* Callgraph removal hook called for a NODE with a custom DATA. */ |
2370 | |
2371 | void |
2372 | sem_item_optimizer::cgraph_removal_hook (cgraph_node *node, void *data) |
2373 | { |
2374 | sem_item_optimizer *optimizer = (sem_item_optimizer *) data; |
2375 | optimizer->remove_symtab_node (node); |
2376 | } |
2377 | |
2378 | /* Varpool removal hook called for a NODE with a custom DATA. */ |
2379 | |
2380 | void |
2381 | sem_item_optimizer::varpool_removal_hook (varpool_node *node, void *data) |
2382 | { |
2383 | sem_item_optimizer *optimizer = (sem_item_optimizer *) data; |
2384 | optimizer->remove_symtab_node (node); |
2385 | } |
2386 | |
2387 | /* Remove symtab NODE triggered by symtab removal hooks. */ |
2388 | |
2389 | void |
2390 | sem_item_optimizer::remove_symtab_node (symtab_node *node) |
2391 | { |
2392 | gcc_assert (m_classes.is_empty ()); |
2393 | |
2394 | m_removed_items_set.add (k: node); |
2395 | } |
2396 | |
2397 | void |
2398 | sem_item_optimizer::remove_item (sem_item *item) |
2399 | { |
2400 | if (m_symtab_node_map.get (k: item->node)) |
2401 | m_symtab_node_map.remove (k: item->node); |
2402 | delete item; |
2403 | } |
2404 | |
2405 | /* Removes all callgraph and varpool nodes that are marked by symtab |
2406 | as deleted. */ |
2407 | |
2408 | void |
2409 | sem_item_optimizer::filter_removed_items (void) |
2410 | { |
2411 | auto_vec <sem_item *> filtered; |
2412 | |
2413 | for (unsigned int i = 0; i < m_items.length(); i++) |
2414 | { |
2415 | sem_item *item = m_items[i]; |
2416 | |
2417 | if (m_removed_items_set.contains (k: item->node)) |
2418 | { |
2419 | remove_item (item); |
2420 | continue; |
2421 | } |
2422 | |
2423 | if (item->type == FUNC) |
2424 | { |
2425 | cgraph_node *cnode = static_cast <sem_function *>(item)->get_node (); |
2426 | |
2427 | if (in_lto_p && (cnode->alias || cnode->body_removed)) |
2428 | remove_item (item); |
2429 | else |
2430 | filtered.safe_push (obj: item); |
2431 | } |
2432 | else /* VAR. */ |
2433 | { |
2434 | if (!flag_ipa_icf_variables) |
2435 | remove_item (item); |
2436 | else |
2437 | { |
2438 | /* Filter out non-readonly variables. */ |
2439 | tree decl = item->decl; |
2440 | varpool_node *vnode = static_cast <sem_variable *>(item)->get_node (); |
2441 | if (!TREE_READONLY (decl) || vnode->body_removed) |
2442 | remove_item (item); |
2443 | else |
2444 | filtered.safe_push (obj: item); |
2445 | } |
2446 | } |
2447 | } |
2448 | |
2449 | /* Clean-up of released semantic items. */ |
2450 | |
2451 | m_items.release (); |
2452 | for (unsigned int i = 0; i < filtered.length(); i++) |
2453 | m_items.safe_push (obj: filtered[i]); |
2454 | } |
2455 | |
2456 | /* Optimizer entry point which returns true in case it processes |
2457 | a merge operation. True is returned if there's a merge operation |
2458 | processed. */ |
2459 | |
2460 | bool |
2461 | sem_item_optimizer::execute (void) |
2462 | { |
2463 | filter_removed_items (); |
2464 | unregister_hooks (); |
2465 | |
2466 | build_graph (); |
2467 | update_hash_by_addr_refs (); |
2468 | update_hash_by_memory_access_type (); |
2469 | build_hash_based_classes (); |
2470 | |
2471 | if (dump_file) |
2472 | fprintf (stream: dump_file, format: "Dump after hash based groups\n" ); |
2473 | dump_cong_classes (); |
2474 | |
2475 | subdivide_classes_by_equality (in_wpa: true); |
2476 | |
2477 | if (dump_file) |
2478 | fprintf (stream: dump_file, format: "Dump after WPA based types groups\n" ); |
2479 | |
2480 | dump_cong_classes (); |
2481 | |
2482 | process_cong_reduction (); |
2483 | checking_verify_classes (); |
2484 | |
2485 | if (dump_file) |
2486 | fprintf (stream: dump_file, format: "Dump after callgraph-based congruence reduction\n" ); |
2487 | |
2488 | dump_cong_classes (); |
2489 | |
2490 | unsigned int loaded_symbols = parse_nonsingleton_classes (); |
2491 | subdivide_classes_by_equality (); |
2492 | |
2493 | if (dump_file) |
2494 | fprintf (stream: dump_file, format: "Dump after full equality comparison of groups\n" ); |
2495 | |
2496 | dump_cong_classes (); |
2497 | |
2498 | unsigned int prev_class_count = m_classes_count; |
2499 | |
2500 | process_cong_reduction (); |
2501 | dump_cong_classes (); |
2502 | checking_verify_classes (); |
2503 | bool merged_p = merge_classes (prev_class_count, loaded_symbols); |
2504 | |
2505 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2506 | symtab->dump (f: dump_file); |
2507 | |
2508 | return merged_p; |
2509 | } |
2510 | |
2511 | /* Function responsible for visiting all potential functions and |
2512 | read-only variables that can be merged. */ |
2513 | |
2514 | void |
2515 | sem_item_optimizer::parse_funcs_and_vars (void) |
2516 | { |
2517 | cgraph_node *cnode; |
2518 | |
2519 | /* Create dummy func_checker for hashing purpose. */ |
2520 | func_checker checker; |
2521 | |
2522 | if (flag_ipa_icf_functions) |
2523 | FOR_EACH_DEFINED_FUNCTION (cnode) |
2524 | { |
2525 | sem_function *f = sem_function::parse (node: cnode, stack: &m_bmstack, checker: &checker); |
2526 | if (f) |
2527 | { |
2528 | m_items.safe_push (obj: f); |
2529 | m_symtab_node_map.put (k: cnode, v: f); |
2530 | } |
2531 | } |
2532 | |
2533 | varpool_node *vnode; |
2534 | |
2535 | if (flag_ipa_icf_variables) |
2536 | FOR_EACH_DEFINED_VARIABLE (vnode) |
2537 | { |
2538 | sem_variable *v = sem_variable::parse (node: vnode, stack: &m_bmstack, checker: &checker); |
2539 | |
2540 | if (v) |
2541 | { |
2542 | m_items.safe_push (obj: v); |
2543 | m_symtab_node_map.put (k: vnode, v); |
2544 | } |
2545 | } |
2546 | } |
2547 | |
2548 | /* Makes pairing between a congruence class CLS and semantic ITEM. */ |
2549 | |
2550 | void |
2551 | sem_item_optimizer::add_item_to_class (congruence_class *cls, sem_item *item) |
2552 | { |
2553 | item->index_in_class = cls->members.length (); |
2554 | cls->members.safe_push (obj: item); |
2555 | cls->referenced_by_count += item->referenced_by_count; |
2556 | item->cls = cls; |
2557 | } |
2558 | |
2559 | /* For each semantic item, append hash values of references. */ |
2560 | |
2561 | void |
2562 | sem_item_optimizer::update_hash_by_addr_refs () |
2563 | { |
2564 | /* First, append to hash sensitive references and class type if it need to |
2565 | be matched for ODR. */ |
2566 | for (unsigned i = 0; i < m_items.length (); i++) |
2567 | { |
2568 | m_items[i]->update_hash_by_addr_refs (m_symtab_node_map); |
2569 | if (m_items[i]->type == FUNC) |
2570 | { |
2571 | if (TREE_CODE (TREE_TYPE (m_items[i]->decl)) == METHOD_TYPE |
2572 | && contains_polymorphic_type_p |
2573 | (TYPE_METHOD_BASETYPE (TREE_TYPE (m_items[i]->decl))) |
2574 | && (DECL_CXX_CONSTRUCTOR_P (m_items[i]->decl) |
2575 | || (static_cast<sem_function *> (m_items[i])->param_used_p (i: 0) |
2576 | && static_cast<sem_function *> (m_items[i]) |
2577 | ->compare_polymorphic_p ()))) |
2578 | { |
2579 | tree class_type |
2580 | = TYPE_METHOD_BASETYPE (TREE_TYPE (m_items[i]->decl)); |
2581 | inchash::hash hstate (m_items[i]->get_hash ()); |
2582 | |
2583 | /* Hash ODR types by mangled name if it is defined. |
2584 | If not we know that type is anonymous of free_lang_data |
2585 | was not run and in that case type main variants are |
2586 | unique. */ |
2587 | if (TYPE_NAME (class_type) |
2588 | && DECL_ASSEMBLER_NAME_SET_P (TYPE_NAME (class_type)) |
2589 | && !type_in_anonymous_namespace_p |
2590 | (t: class_type)) |
2591 | hstate.add_hwi |
2592 | (IDENTIFIER_HASH_VALUE |
2593 | (DECL_ASSEMBLER_NAME (TYPE_NAME (class_type)))); |
2594 | else |
2595 | { |
2596 | gcc_checking_assert |
2597 | (!in_lto_p |
2598 | || type_in_anonymous_namespace_p (class_type)); |
2599 | hstate.add_hwi (TYPE_UID (TYPE_MAIN_VARIANT (class_type))); |
2600 | } |
2601 | |
2602 | m_items[i]->set_hash (hstate.end ()); |
2603 | } |
2604 | } |
2605 | } |
2606 | |
2607 | /* Once all symbols have enhanced hash value, we can append |
2608 | hash values of symbols that are seen by IPA ICF and are |
2609 | references by a semantic item. Newly computed values |
2610 | are saved to global_hash member variable. */ |
2611 | for (unsigned i = 0; i < m_items.length (); i++) |
2612 | m_items[i]->update_hash_by_local_refs (m_symtab_node_map); |
2613 | |
2614 | /* Global hash value replace current hash values. */ |
2615 | for (unsigned i = 0; i < m_items.length (); i++) |
2616 | m_items[i]->set_hash (m_items[i]->global_hash); |
2617 | } |
2618 | |
2619 | void |
2620 | sem_item_optimizer::update_hash_by_memory_access_type () |
2621 | { |
2622 | for (unsigned i = 0; i < m_items.length (); i++) |
2623 | { |
2624 | if (m_items[i]->type == FUNC) |
2625 | { |
2626 | sem_function *fn = static_cast<sem_function *> (m_items[i]); |
2627 | inchash::hash hstate (fn->get_hash ()); |
2628 | hstate.add_int (v: fn->m_alias_sets_hash); |
2629 | fn->set_hash (hstate.end ()); |
2630 | } |
2631 | } |
2632 | } |
2633 | |
2634 | /* Congruence classes are built by hash value. */ |
2635 | |
2636 | void |
2637 | sem_item_optimizer::build_hash_based_classes (void) |
2638 | { |
2639 | for (unsigned i = 0; i < m_items.length (); i++) |
2640 | { |
2641 | sem_item *item = m_items[i]; |
2642 | |
2643 | congruence_class_group *group |
2644 | = get_group_by_hash (hash: item->get_hash (), type: item->type); |
2645 | |
2646 | if (!group->classes.length ()) |
2647 | { |
2648 | m_classes_count++; |
2649 | group->classes.safe_push (obj: new congruence_class (class_id++)); |
2650 | } |
2651 | |
2652 | add_item_to_class (cls: group->classes[0], item); |
2653 | } |
2654 | } |
2655 | |
2656 | /* Build references according to call graph. */ |
2657 | |
2658 | void |
2659 | sem_item_optimizer::build_graph (void) |
2660 | { |
2661 | for (unsigned i = 0; i < m_items.length (); i++) |
2662 | { |
2663 | sem_item *item = m_items[i]; |
2664 | m_symtab_node_map.put (k: item->node, v: item); |
2665 | |
2666 | /* Initialize hash values if we are not in LTO mode. */ |
2667 | if (!in_lto_p) |
2668 | item->get_hash (); |
2669 | } |
2670 | |
2671 | for (unsigned i = 0; i < m_items.length (); i++) |
2672 | { |
2673 | sem_item *item = m_items[i]; |
2674 | |
2675 | if (item->type == FUNC) |
2676 | { |
2677 | cgraph_node *cnode = dyn_cast <cgraph_node *> (p: item->node); |
2678 | |
2679 | cgraph_edge *e = cnode->callees; |
2680 | while (e) |
2681 | { |
2682 | sem_item **slot = m_symtab_node_map.get |
2683 | (k: e->callee->ultimate_alias_target ()); |
2684 | if (slot) |
2685 | item->add_reference (refs: &m_references, target: *slot); |
2686 | |
2687 | e = e->next_callee; |
2688 | } |
2689 | } |
2690 | |
2691 | ipa_ref *ref = NULL; |
2692 | for (unsigned i = 0; item->node->iterate_reference (i, ref); i++) |
2693 | { |
2694 | sem_item **slot = m_symtab_node_map.get |
2695 | (k: ref->referred->ultimate_alias_target ()); |
2696 | if (slot) |
2697 | item->add_reference (refs: &m_references, target: *slot); |
2698 | } |
2699 | } |
2700 | } |
2701 | |
2702 | /* Semantic items in classes having more than one element and initialized. |
2703 | In case of WPA, we load function body. */ |
2704 | |
2705 | unsigned int |
2706 | sem_item_optimizer::parse_nonsingleton_classes (void) |
2707 | { |
2708 | unsigned int counter = 0; |
2709 | |
2710 | /* Create dummy func_checker for hashing purpose. */ |
2711 | func_checker checker; |
2712 | |
2713 | for (unsigned i = 0; i < m_items.length (); i++) |
2714 | if (m_items[i]->cls->members.length () > 1) |
2715 | { |
2716 | m_items[i]->init (&checker); |
2717 | ++counter; |
2718 | } |
2719 | |
2720 | if (dump_file) |
2721 | { |
2722 | float f = m_items.length () ? 100.0f * counter / m_items.length () : 0.0f; |
2723 | fprintf (stream: dump_file, format: "Init called for %u items (%.2f%%).\n" , counter, f); |
2724 | } |
2725 | |
2726 | return counter; |
2727 | } |
2728 | |
2729 | /* Equality function for semantic items is used to subdivide existing |
2730 | classes. If IN_WPA, fast equality function is invoked. */ |
2731 | |
2732 | void |
2733 | sem_item_optimizer::subdivide_classes_by_equality (bool in_wpa) |
2734 | { |
2735 | for (hash_table <congruence_class_hash>::iterator it = m_classes.begin (); |
2736 | it != m_classes.end (); ++it) |
2737 | { |
2738 | unsigned int class_count = (*it)->classes.length (); |
2739 | |
2740 | for (unsigned i = 0; i < class_count; i++) |
2741 | { |
2742 | congruence_class *c = (*it)->classes[i]; |
2743 | |
2744 | if (c->members.length() > 1) |
2745 | { |
2746 | auto_vec <sem_item *> new_vector; |
2747 | |
2748 | sem_item *first = c->members[0]; |
2749 | new_vector.safe_push (obj: first); |
2750 | |
2751 | unsigned class_split_first = (*it)->classes.length (); |
2752 | |
2753 | for (unsigned j = 1; j < c->members.length (); j++) |
2754 | { |
2755 | sem_item *item = c->members[j]; |
2756 | |
2757 | bool equals |
2758 | = in_wpa ? first->equals_wpa (item, ignored_nodes&: m_symtab_node_map) |
2759 | : first->equals (item, ignored_nodes&: m_symtab_node_map); |
2760 | |
2761 | if (equals) |
2762 | new_vector.safe_push (obj: item); |
2763 | else |
2764 | { |
2765 | bool integrated = false; |
2766 | |
2767 | for (unsigned k = class_split_first; |
2768 | k < (*it)->classes.length (); k++) |
2769 | { |
2770 | sem_item *x = (*it)->classes[k]->members[0]; |
2771 | bool equals |
2772 | = in_wpa ? x->equals_wpa (item, ignored_nodes&: m_symtab_node_map) |
2773 | : x->equals (item, ignored_nodes&: m_symtab_node_map); |
2774 | |
2775 | if (equals) |
2776 | { |
2777 | integrated = true; |
2778 | add_item_to_class (cls: (*it)->classes[k], item); |
2779 | |
2780 | break; |
2781 | } |
2782 | } |
2783 | |
2784 | if (!integrated) |
2785 | { |
2786 | congruence_class *c |
2787 | = new congruence_class (class_id++); |
2788 | m_classes_count++; |
2789 | add_item_to_class (cls: c, item); |
2790 | |
2791 | (*it)->classes.safe_push (obj: c); |
2792 | } |
2793 | } |
2794 | } |
2795 | |
2796 | // We replace newly created new_vector for the class we've just |
2797 | // splitted. |
2798 | c->members.release (); |
2799 | c->members.create (nelems: new_vector.length ()); |
2800 | |
2801 | for (unsigned int j = 0; j < new_vector.length (); j++) |
2802 | add_item_to_class (cls: c, item: new_vector[j]); |
2803 | } |
2804 | } |
2805 | } |
2806 | |
2807 | checking_verify_classes (); |
2808 | } |
2809 | |
2810 | /* Subdivide classes by address references that members of the class |
2811 | reference. Example can be a pair of functions that have an address |
2812 | taken from a function. If these addresses are different the class |
2813 | is split. */ |
2814 | |
2815 | unsigned |
2816 | sem_item_optimizer::subdivide_classes_by_sensitive_refs () |
2817 | { |
2818 | typedef hash_map <symbol_compare_hash, vec <sem_item *> > subdivide_hash_map; |
2819 | |
2820 | unsigned newly_created_classes = 0; |
2821 | |
2822 | for (hash_table <congruence_class_hash>::iterator it = m_classes.begin (); |
2823 | it != m_classes.end (); ++it) |
2824 | { |
2825 | unsigned int class_count = (*it)->classes.length (); |
2826 | auto_vec<congruence_class *> new_classes; |
2827 | |
2828 | for (unsigned i = 0; i < class_count; i++) |
2829 | { |
2830 | congruence_class *c = (*it)->classes[i]; |
2831 | |
2832 | if (c->members.length() > 1) |
2833 | { |
2834 | subdivide_hash_map split_map; |
2835 | |
2836 | for (unsigned j = 0; j < c->members.length (); j++) |
2837 | { |
2838 | sem_item *source_node = c->members[j]; |
2839 | |
2840 | symbol_compare_collection *collection |
2841 | = new symbol_compare_collection (source_node->node); |
2842 | |
2843 | bool existed; |
2844 | vec <sem_item *> *slot |
2845 | = &split_map.get_or_insert (k: collection, existed: &existed); |
2846 | gcc_checking_assert (slot); |
2847 | |
2848 | slot->safe_push (obj: source_node); |
2849 | |
2850 | if (existed) |
2851 | delete collection; |
2852 | } |
2853 | |
2854 | /* If the map contains more than one key, we have to split |
2855 | the map appropriately. */ |
2856 | if (split_map.elements () != 1) |
2857 | { |
2858 | bool first_class = true; |
2859 | |
2860 | for (subdivide_hash_map::iterator it2 = split_map.begin (); |
2861 | it2 != split_map.end (); ++it2) |
2862 | { |
2863 | congruence_class *new_cls; |
2864 | new_cls = new congruence_class (class_id++); |
2865 | |
2866 | for (unsigned k = 0; k < (*it2).second.length (); k++) |
2867 | add_item_to_class (cls: new_cls, item: (*it2).second[k]); |
2868 | |
2869 | worklist_push (cls: new_cls); |
2870 | newly_created_classes++; |
2871 | |
2872 | if (first_class) |
2873 | { |
2874 | (*it)->classes[i] = new_cls; |
2875 | first_class = false; |
2876 | } |
2877 | else |
2878 | { |
2879 | new_classes.safe_push (obj: new_cls); |
2880 | m_classes_count++; |
2881 | } |
2882 | } |
2883 | } |
2884 | |
2885 | /* Release memory. */ |
2886 | for (subdivide_hash_map::iterator it2 = split_map.begin (); |
2887 | it2 != split_map.end (); ++it2) |
2888 | { |
2889 | delete (*it2).first; |
2890 | (*it2).second.release (); |
2891 | } |
2892 | } |
2893 | } |
2894 | |
2895 | for (unsigned i = 0; i < new_classes.length (); i++) |
2896 | (*it)->classes.safe_push (obj: new_classes[i]); |
2897 | } |
2898 | |
2899 | return newly_created_classes; |
2900 | } |
2901 | |
2902 | /* Verify congruence classes, if checking is enabled. */ |
2903 | |
2904 | void |
2905 | sem_item_optimizer::checking_verify_classes (void) |
2906 | { |
2907 | if (flag_checking) |
2908 | verify_classes (); |
2909 | } |
2910 | |
2911 | /* Verify congruence classes. */ |
2912 | |
2913 | void |
2914 | sem_item_optimizer::verify_classes (void) |
2915 | { |
2916 | for (hash_table<congruence_class_hash>::iterator it = m_classes.begin (); |
2917 | it != m_classes.end (); ++it) |
2918 | { |
2919 | for (unsigned int i = 0; i < (*it)->classes.length (); i++) |
2920 | { |
2921 | congruence_class *cls = (*it)->classes[i]; |
2922 | |
2923 | gcc_assert (cls); |
2924 | gcc_assert (cls->members.length () > 0); |
2925 | |
2926 | for (unsigned int j = 0; j < cls->members.length (); j++) |
2927 | { |
2928 | sem_item *item = cls->members[j]; |
2929 | |
2930 | gcc_assert (item); |
2931 | gcc_assert (item->cls == cls); |
2932 | } |
2933 | } |
2934 | } |
2935 | } |
2936 | |
2937 | /* Disposes split map traverse function. CLS_PTR is pointer to congruence |
2938 | class, BSLOT is bitmap slot we want to release. DATA is mandatory, |
2939 | but unused argument. */ |
2940 | |
2941 | bool |
2942 | sem_item_optimizer::release_split_map (congruence_class * const &, |
2943 | bitmap const &b, traverse_split_pair *) |
2944 | { |
2945 | bitmap bmp = b; |
2946 | |
2947 | BITMAP_FREE (bmp); |
2948 | |
2949 | return true; |
2950 | } |
2951 | |
2952 | /* Process split operation for a class given as pointer CLS_PTR, |
2953 | where bitmap B splits congruence class members. DATA is used |
2954 | as argument of split pair. */ |
2955 | |
2956 | bool |
2957 | sem_item_optimizer::traverse_congruence_split (congruence_class * const &cls, |
2958 | bitmap const &b, |
2959 | traverse_split_pair *pair) |
2960 | { |
2961 | sem_item_optimizer *optimizer = pair->optimizer; |
2962 | const congruence_class *splitter_cls = pair->cls; |
2963 | |
2964 | /* If counted bits are greater than zero and less than the number of members |
2965 | a group will be splitted. */ |
2966 | unsigned popcount = bitmap_count_bits (b); |
2967 | |
2968 | if (popcount > 0 && popcount < cls->members.length ()) |
2969 | { |
2970 | auto_vec <congruence_class *, 2> newclasses; |
2971 | newclasses.quick_push (obj: new congruence_class (class_id++)); |
2972 | newclasses.quick_push (obj: new congruence_class (class_id++)); |
2973 | |
2974 | for (unsigned int i = 0; i < cls->members.length (); i++) |
2975 | { |
2976 | int target = bitmap_bit_p (b, i); |
2977 | congruence_class *tc = newclasses[target]; |
2978 | |
2979 | add_item_to_class (cls: tc, item: cls->members[i]); |
2980 | } |
2981 | |
2982 | if (flag_checking) |
2983 | { |
2984 | for (unsigned int i = 0; i < 2; i++) |
2985 | gcc_assert (newclasses[i]->members.length ()); |
2986 | } |
2987 | |
2988 | if (splitter_cls == cls) |
2989 | optimizer->splitter_class_removed = true; |
2990 | |
2991 | /* Remove old class from worklist if presented. */ |
2992 | bool in_worklist = cls->in_worklist; |
2993 | |
2994 | if (in_worklist) |
2995 | cls->in_worklist = false; |
2996 | |
2997 | congruence_class_group g; |
2998 | g.hash = cls->members[0]->get_hash (); |
2999 | g.type = cls->members[0]->type; |
3000 | |
3001 | congruence_class_group *slot = optimizer->m_classes.find (value: &g); |
3002 | |
3003 | for (unsigned int i = 0; i < slot->classes.length (); i++) |
3004 | if (slot->classes[i] == cls) |
3005 | { |
3006 | slot->classes.ordered_remove (ix: i); |
3007 | break; |
3008 | } |
3009 | |
3010 | /* New class will be inserted and integrated to work list. */ |
3011 | for (unsigned int i = 0; i < 2; i++) |
3012 | optimizer->add_class (cls: newclasses[i]); |
3013 | |
3014 | /* Two classes replace one, so that increment just by one. */ |
3015 | optimizer->m_classes_count++; |
3016 | |
3017 | /* If OLD class was presented in the worklist, we remove the class |
3018 | and replace it will both newly created classes. */ |
3019 | if (in_worklist) |
3020 | for (unsigned int i = 0; i < 2; i++) |
3021 | optimizer->worklist_push (cls: newclasses[i]); |
3022 | else /* Just smaller class is inserted. */ |
3023 | { |
3024 | unsigned int smaller_index |
3025 | = (newclasses[0]->members.length () |
3026 | < newclasses[1]->members.length () |
3027 | ? 0 : 1); |
3028 | optimizer->worklist_push (cls: newclasses[smaller_index]); |
3029 | } |
3030 | |
3031 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3032 | { |
3033 | fprintf (stream: dump_file, format: " congruence class splitted:\n" ); |
3034 | cls->dump (file: dump_file, indent: 4); |
3035 | |
3036 | fprintf (stream: dump_file, format: " newly created groups:\n" ); |
3037 | for (unsigned int i = 0; i < 2; i++) |
3038 | newclasses[i]->dump (file: dump_file, indent: 4); |
3039 | } |
3040 | |
3041 | /* Release class if not presented in work list. */ |
3042 | if (!in_worklist) |
3043 | delete cls; |
3044 | |
3045 | return true; |
3046 | } |
3047 | |
3048 | return false; |
3049 | } |
3050 | |
3051 | /* Compare function for sorting pairs in do_congruence_step_f. */ |
3052 | |
3053 | int |
3054 | sem_item_optimizer::sort_congruence_split (const void *a_, const void *b_) |
3055 | { |
3056 | const std::pair<congruence_class *, bitmap> *a |
3057 | = (const std::pair<congruence_class *, bitmap> *)a_; |
3058 | const std::pair<congruence_class *, bitmap> *b |
3059 | = (const std::pair<congruence_class *, bitmap> *)b_; |
3060 | if (a->first->id < b->first->id) |
3061 | return -1; |
3062 | else if (a->first->id > b->first->id) |
3063 | return 1; |
3064 | return 0; |
3065 | } |
3066 | |
3067 | /* Tests if a class CLS used as INDEXth splits any congruence classes. |
3068 | Bitmap stack BMSTACK is used for bitmap allocation. */ |
3069 | |
3070 | bool |
3071 | sem_item_optimizer::do_congruence_step_for_index (congruence_class *cls, |
3072 | unsigned int index) |
3073 | { |
3074 | hash_map <congruence_class *, bitmap> split_map; |
3075 | |
3076 | for (unsigned int i = 0; i < cls->members.length (); i++) |
3077 | { |
3078 | sem_item *item = cls->members[i]; |
3079 | sem_usage_pair needle (item, index); |
3080 | vec<sem_item *> *callers = m_references.get (k: &needle); |
3081 | if (callers == NULL) |
3082 | continue; |
3083 | |
3084 | for (unsigned int j = 0; j < callers->length (); j++) |
3085 | { |
3086 | sem_item *caller = (*callers)[j]; |
3087 | if (caller->cls->members.length () < 2) |
3088 | continue; |
3089 | bitmap *slot = split_map.get (k: caller->cls); |
3090 | bitmap b; |
3091 | |
3092 | if(!slot) |
3093 | { |
3094 | b = BITMAP_ALLOC (obstack: &m_bmstack); |
3095 | split_map.put (k: caller->cls, v: b); |
3096 | } |
3097 | else |
3098 | b = *slot; |
3099 | |
3100 | gcc_checking_assert (caller->cls); |
3101 | gcc_checking_assert (caller->index_in_class |
3102 | < caller->cls->members.length ()); |
3103 | |
3104 | bitmap_set_bit (b, caller->index_in_class); |
3105 | } |
3106 | } |
3107 | |
3108 | auto_vec<std::pair<congruence_class *, bitmap> > to_split; |
3109 | to_split.reserve_exact (nelems: split_map.elements ()); |
3110 | for (hash_map <congruence_class *, bitmap>::iterator i = split_map.begin (); |
3111 | i != split_map.end (); ++i) |
3112 | to_split.safe_push (obj: *i); |
3113 | to_split.qsort (sort_congruence_split); |
3114 | |
3115 | traverse_split_pair pair; |
3116 | pair.optimizer = this; |
3117 | pair.cls = cls; |
3118 | |
3119 | splitter_class_removed = false; |
3120 | bool r = false; |
3121 | for (unsigned i = 0; i < to_split.length (); ++i) |
3122 | r |= traverse_congruence_split (cls: to_split[i].first, b: to_split[i].second, |
3123 | pair: &pair); |
3124 | |
3125 | /* Bitmap clean-up. */ |
3126 | split_map.traverse <traverse_split_pair *, |
3127 | sem_item_optimizer::release_split_map> (NULL); |
3128 | |
3129 | return r; |
3130 | } |
3131 | |
3132 | /* Every usage of a congruence class CLS is a candidate that can split the |
3133 | collection of classes. Bitmap stack BMSTACK is used for bitmap |
3134 | allocation. */ |
3135 | |
3136 | void |
3137 | sem_item_optimizer::do_congruence_step (congruence_class *cls) |
3138 | { |
3139 | bitmap_iterator bi; |
3140 | unsigned int i; |
3141 | |
3142 | bitmap usage = BITMAP_ALLOC (obstack: &m_bmstack); |
3143 | |
3144 | for (unsigned int i = 0; i < cls->members.length (); i++) |
3145 | bitmap_ior_into (usage, cls->members[i]->usage_index_bitmap); |
3146 | |
3147 | EXECUTE_IF_SET_IN_BITMAP (usage, 0, i, bi) |
3148 | { |
3149 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3150 | fprintf (stream: dump_file, format: " processing congruence step for class: %u " |
3151 | "(%u items, %u references), index: %u\n" , cls->id, |
3152 | cls->referenced_by_count, cls->members.length (), i); |
3153 | do_congruence_step_for_index (cls, index: i); |
3154 | |
3155 | if (splitter_class_removed) |
3156 | break; |
3157 | } |
3158 | |
3159 | BITMAP_FREE (usage); |
3160 | } |
3161 | |
3162 | /* Adds a newly created congruence class CLS to worklist. */ |
3163 | |
3164 | void |
3165 | sem_item_optimizer::worklist_push (congruence_class *cls) |
3166 | { |
3167 | /* Return if the class CLS is already presented in work list. */ |
3168 | if (cls->in_worklist) |
3169 | return; |
3170 | |
3171 | cls->in_worklist = true; |
3172 | worklist.insert (key: cls->referenced_by_count, data: cls); |
3173 | } |
3174 | |
3175 | /* Pops a class from worklist. */ |
3176 | |
3177 | congruence_class * |
3178 | sem_item_optimizer::worklist_pop (void) |
3179 | { |
3180 | congruence_class *cls; |
3181 | |
3182 | while (!worklist.empty ()) |
3183 | { |
3184 | cls = worklist.extract_min (); |
3185 | if (cls->in_worklist) |
3186 | { |
3187 | cls->in_worklist = false; |
3188 | |
3189 | return cls; |
3190 | } |
3191 | else |
3192 | { |
3193 | /* Work list item was already intended to be removed. |
3194 | The only reason for doing it is to split a class. |
3195 | Thus, the class CLS is deleted. */ |
3196 | delete cls; |
3197 | } |
3198 | } |
3199 | |
3200 | return NULL; |
3201 | } |
3202 | |
3203 | /* Iterative congruence reduction function. */ |
3204 | |
3205 | void |
3206 | sem_item_optimizer::process_cong_reduction (void) |
3207 | { |
3208 | for (hash_table<congruence_class_hash>::iterator it = m_classes.begin (); |
3209 | it != m_classes.end (); ++it) |
3210 | for (unsigned i = 0; i < (*it)->classes.length (); i++) |
3211 | if ((*it)->classes[i]->is_class_used ()) |
3212 | worklist_push (cls: (*it)->classes[i]); |
3213 | |
3214 | if (dump_file) |
3215 | fprintf (stream: dump_file, format: "Worklist has been filled with: " |
3216 | HOST_SIZE_T_PRINT_UNSIGNED "\n" , |
3217 | (fmt_size_t) worklist.nodes ()); |
3218 | |
3219 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3220 | fprintf (stream: dump_file, format: "Congruence class reduction\n" ); |
3221 | |
3222 | congruence_class *cls; |
3223 | |
3224 | /* Process complete congruence reduction. */ |
3225 | while ((cls = worklist_pop ()) != NULL) |
3226 | do_congruence_step (cls); |
3227 | |
3228 | /* Subdivide newly created classes according to references. */ |
3229 | unsigned new_classes = subdivide_classes_by_sensitive_refs (); |
3230 | |
3231 | if (dump_file) |
3232 | fprintf (stream: dump_file, format: "Address reference subdivision created: %u " |
3233 | "new classes.\n" , new_classes); |
3234 | } |
3235 | |
3236 | /* Debug function prints all informations about congruence classes. */ |
3237 | |
3238 | void |
3239 | sem_item_optimizer::dump_cong_classes (void) |
3240 | { |
3241 | if (!dump_file) |
3242 | return; |
3243 | |
3244 | /* Histogram calculation. */ |
3245 | unsigned int max_index = 0; |
3246 | unsigned int single_element_classes = 0; |
3247 | unsigned int* histogram = XCNEWVEC (unsigned int, m_items.length () + 1); |
3248 | |
3249 | for (hash_table<congruence_class_hash>::iterator it = m_classes.begin (); |
3250 | it != m_classes.end (); ++it) |
3251 | for (unsigned i = 0; i < (*it)->classes.length (); i++) |
3252 | { |
3253 | unsigned int c = (*it)->classes[i]->members.length (); |
3254 | histogram[c]++; |
3255 | |
3256 | if (c > max_index) |
3257 | max_index = c; |
3258 | |
3259 | if (c == 1) |
3260 | ++single_element_classes; |
3261 | } |
3262 | |
3263 | fprintf (stream: dump_file, |
3264 | format: "Congruence classes: " HOST_SIZE_T_PRINT_UNSIGNED " with total: " |
3265 | "%u items (in a non-singular class: %u)\n" , |
3266 | (fmt_size_t) m_classes.elements (), |
3267 | m_items.length (), m_items.length () - single_element_classes); |
3268 | fprintf (stream: dump_file, |
3269 | format: "Class size histogram [number of members]: number of classes\n" ); |
3270 | for (unsigned int i = 0; i <= max_index; i++) |
3271 | if (histogram[i]) |
3272 | fprintf (stream: dump_file, format: "%6u: %6u\n" , i, histogram[i]); |
3273 | |
3274 | if (dump_flags & TDF_DETAILS) |
3275 | for (hash_table<congruence_class_hash>::iterator it = m_classes.begin (); |
3276 | it != m_classes.end (); ++it) |
3277 | { |
3278 | fprintf (stream: dump_file, format: " group: with %u classes:\n" , |
3279 | (*it)->classes.length ()); |
3280 | |
3281 | for (unsigned i = 0; i < (*it)->classes.length (); i++) |
3282 | { |
3283 | (*it)->classes[i]->dump (file: dump_file, indent: 4); |
3284 | |
3285 | if (i < (*it)->classes.length () - 1) |
3286 | fprintf (stream: dump_file, format: " " ); |
3287 | } |
3288 | } |
3289 | |
3290 | free (ptr: histogram); |
3291 | } |
3292 | |
3293 | /* Sort pair of sem_items A and B by DECL_UID. */ |
3294 | |
3295 | static int |
3296 | sort_sem_items_by_decl_uid (const void *a, const void *b) |
3297 | { |
3298 | const sem_item *i1 = *(const sem_item * const *)a; |
3299 | const sem_item *i2 = *(const sem_item * const *)b; |
3300 | |
3301 | int uid1 = DECL_UID (i1->decl); |
3302 | int uid2 = DECL_UID (i2->decl); |
3303 | return uid1 - uid2; |
3304 | } |
3305 | |
3306 | /* Sort pair of congruence_classes A and B by DECL_UID of the first member. */ |
3307 | |
3308 | static int |
3309 | sort_congruence_classes_by_decl_uid (const void *a, const void *b) |
3310 | { |
3311 | const congruence_class *c1 = *(const congruence_class * const *)a; |
3312 | const congruence_class *c2 = *(const congruence_class * const *)b; |
3313 | |
3314 | int uid1 = DECL_UID (c1->members[0]->decl); |
3315 | int uid2 = DECL_UID (c2->members[0]->decl); |
3316 | return uid1 - uid2; |
3317 | } |
3318 | |
3319 | /* Sort pair of congruence_class_groups A and B by |
3320 | DECL_UID of the first member of a first group. */ |
3321 | |
3322 | static int |
3323 | sort_congruence_class_groups_by_decl_uid (const void *a, const void *b) |
3324 | { |
3325 | const std::pair<congruence_class_group *, int> *g1 |
3326 | = (const std::pair<congruence_class_group *, int> *) a; |
3327 | const std::pair<congruence_class_group *, int> *g2 |
3328 | = (const std::pair<congruence_class_group *, int> *) b; |
3329 | return g1->second - g2->second; |
3330 | } |
3331 | |
3332 | /* After reduction is done, we can declare all items in a group |
3333 | to be equal. PREV_CLASS_COUNT is start number of classes |
3334 | before reduction. True is returned if there's a merge operation |
3335 | processed. LOADED_SYMBOLS is number of symbols that were loaded |
3336 | in WPA. */ |
3337 | |
3338 | bool |
3339 | sem_item_optimizer::merge_classes (unsigned int prev_class_count, |
3340 | unsigned int loaded_symbols) |
3341 | { |
3342 | unsigned int item_count = m_items.length (); |
3343 | unsigned int class_count = m_classes_count; |
3344 | unsigned int equal_items = item_count - class_count; |
3345 | |
3346 | unsigned int non_singular_classes_count = 0; |
3347 | unsigned int non_singular_classes_sum = 0; |
3348 | |
3349 | bool merged_p = false; |
3350 | |
3351 | /* PR lto/78211 |
3352 | Sort functions in congruence classes by DECL_UID and do the same |
3353 | for the classes to not to break -fcompare-debug. */ |
3354 | |
3355 | for (hash_table<congruence_class_hash>::iterator it = m_classes.begin (); |
3356 | it != m_classes.end (); ++it) |
3357 | { |
3358 | for (unsigned int i = 0; i < (*it)->classes.length (); i++) |
3359 | { |
3360 | congruence_class *c = (*it)->classes[i]; |
3361 | c->members.qsort (sort_sem_items_by_decl_uid); |
3362 | } |
3363 | |
3364 | (*it)->classes.qsort (sort_congruence_classes_by_decl_uid); |
3365 | } |
3366 | |
3367 | for (hash_table<congruence_class_hash>::iterator it = m_classes.begin (); |
3368 | it != m_classes.end (); ++it) |
3369 | for (unsigned int i = 0; i < (*it)->classes.length (); i++) |
3370 | { |
3371 | congruence_class *c = (*it)->classes[i]; |
3372 | if (c->members.length () > 1) |
3373 | { |
3374 | non_singular_classes_count++; |
3375 | non_singular_classes_sum += c->members.length (); |
3376 | } |
3377 | } |
3378 | |
3379 | auto_vec<std::pair<congruence_class_group *, int> > classes ( |
3380 | m_classes.elements ()); |
3381 | for (hash_table<congruence_class_hash>::iterator it = m_classes.begin (); |
3382 | it != m_classes.end (); ++it) |
3383 | { |
3384 | int uid = DECL_UID ((*it)->classes[0]->members[0]->decl); |
3385 | classes.quick_push (obj: std::pair<congruence_class_group *, int> (*it, uid)); |
3386 | } |
3387 | |
3388 | classes.qsort (sort_congruence_class_groups_by_decl_uid); |
3389 | |
3390 | if (dump_file) |
3391 | { |
3392 | fprintf (stream: dump_file, format: "\nItem count: %u\n" , item_count); |
3393 | fprintf (stream: dump_file, format: "Congruent classes before: %u, after: %u\n" , |
3394 | prev_class_count, class_count); |
3395 | fprintf (stream: dump_file, format: "Average class size before: %.2f, after: %.2f\n" , |
3396 | prev_class_count ? 1.0f * item_count / prev_class_count : 0.0f, |
3397 | class_count ? 1.0f * item_count / class_count : 0.0f); |
3398 | fprintf (stream: dump_file, format: "Average non-singular class size: %.2f, count: %u\n" , |
3399 | non_singular_classes_count ? 1.0f * non_singular_classes_sum / |
3400 | non_singular_classes_count : 0.0f, |
3401 | non_singular_classes_count); |
3402 | fprintf (stream: dump_file, format: "Equal symbols: %u\n" , equal_items); |
3403 | unsigned total = equal_items + non_singular_classes_count; |
3404 | fprintf (stream: dump_file, format: "Totally needed symbols: %u" |
3405 | ", fraction of loaded symbols: %.2f%%\n\n" , total, |
3406 | loaded_symbols ? 100.0f * total / loaded_symbols: 0.0f); |
3407 | } |
3408 | |
3409 | unsigned int l; |
3410 | std::pair<congruence_class_group *, int> *it; |
3411 | FOR_EACH_VEC_ELT (classes, l, it) |
3412 | for (unsigned int i = 0; i < it->first->classes.length (); i++) |
3413 | { |
3414 | congruence_class *c = it->first->classes[i]; |
3415 | |
3416 | if (c->members.length () == 1) |
3417 | continue; |
3418 | |
3419 | sem_item *source = c->members[0]; |
3420 | bool this_merged_p = false; |
3421 | |
3422 | if (DECL_NAME (source->decl) |
3423 | && MAIN_NAME_P (DECL_NAME (source->decl))) |
3424 | /* If merge via wrappers, picking main as the target can be |
3425 | problematic. */ |
3426 | source = c->members[1]; |
3427 | |
3428 | for (unsigned int j = 0; j < c->members.length (); j++) |
3429 | { |
3430 | sem_item *alias = c->members[j]; |
3431 | |
3432 | if (alias == source) |
3433 | continue; |
3434 | |
3435 | dump_user_location_t loc |
3436 | = dump_user_location_t::from_function_decl (fndecl: source->decl); |
3437 | if (dump_enabled_p ()) |
3438 | { |
3439 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, loc, |
3440 | "Semantic equality hit:%s->%s\n" , |
3441 | source->node->dump_name (), |
3442 | alias->node->dump_name ()); |
3443 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, loc, |
3444 | "Assembler symbol names:%s->%s\n" , |
3445 | source->node->dump_asm_name (), |
3446 | alias->node->dump_asm_name ()); |
3447 | } |
3448 | |
3449 | if (lookup_attribute (attr_name: "no_icf" , DECL_ATTRIBUTES (alias->decl)) |
3450 | || lookup_attribute (attr_name: "no_icf" , DECL_ATTRIBUTES (source->decl))) |
3451 | { |
3452 | if (dump_enabled_p ()) |
3453 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, loc, |
3454 | "Merge operation is skipped due to no_icf " |
3455 | "attribute.\n" ); |
3456 | continue; |
3457 | } |
3458 | |
3459 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3460 | { |
3461 | source->dump_to_file (file: dump_file); |
3462 | alias->dump_to_file (file: dump_file); |
3463 | } |
3464 | |
3465 | if (dbg_cnt (index: merged_ipa_icf)) |
3466 | { |
3467 | bool merged = source->merge (alias_item: alias); |
3468 | this_merged_p |= merged; |
3469 | |
3470 | if (merged && alias->type == VAR) |
3471 | { |
3472 | symtab_pair p = symtab_pair (source->node, alias->node); |
3473 | m_merged_variables.safe_push (obj: p); |
3474 | } |
3475 | } |
3476 | } |
3477 | |
3478 | merged_p |= this_merged_p; |
3479 | if (this_merged_p |
3480 | && source->type == FUNC |
3481 | && (!flag_wpa || flag_checking)) |
3482 | { |
3483 | unsigned i; |
3484 | tree name; |
3485 | FOR_EACH_SSA_NAME (i, name, DECL_STRUCT_FUNCTION (source->decl)) |
3486 | { |
3487 | /* We need to either merge or reset SSA_NAME_*_INFO. |
3488 | For merging we don't preserve the mapping between |
3489 | original and alias SSA_NAMEs from successful equals |
3490 | calls. */ |
3491 | if (POINTER_TYPE_P (TREE_TYPE (name))) |
3492 | { |
3493 | if (SSA_NAME_PTR_INFO (name)) |
3494 | { |
3495 | gcc_checking_assert (!flag_wpa); |
3496 | SSA_NAME_PTR_INFO (name) = NULL; |
3497 | } |
3498 | } |
3499 | else if (SSA_NAME_RANGE_INFO (name)) |
3500 | { |
3501 | gcc_checking_assert (!flag_wpa); |
3502 | SSA_NAME_RANGE_INFO (name) = NULL; |
3503 | } |
3504 | } |
3505 | } |
3506 | } |
3507 | |
3508 | if (!m_merged_variables.is_empty ()) |
3509 | fixup_points_to_sets (); |
3510 | |
3511 | return merged_p; |
3512 | } |
3513 | |
3514 | /* Fixup points to set PT. */ |
3515 | |
3516 | void |
3517 | sem_item_optimizer::fixup_pt_set (struct pt_solution *pt) |
3518 | { |
3519 | if (pt->vars == NULL) |
3520 | return; |
3521 | |
3522 | unsigned i; |
3523 | symtab_pair *item; |
3524 | FOR_EACH_VEC_ELT (m_merged_variables, i, item) |
3525 | if (bitmap_bit_p (pt->vars, DECL_UID (item->second->decl))) |
3526 | bitmap_set_bit (pt->vars, DECL_UID (item->first->decl)); |
3527 | } |
3528 | |
3529 | /* Set all points-to UIDs of aliases pointing to node N as UID. */ |
3530 | |
3531 | static void |
3532 | set_alias_uids (symtab_node *n, int uid) |
3533 | { |
3534 | ipa_ref *ref; |
3535 | FOR_EACH_ALIAS (n, ref) |
3536 | { |
3537 | if (dump_file) |
3538 | fprintf (stream: dump_file, format: " Setting points-to UID of [%s] as %d\n" , |
3539 | ref->referring->dump_asm_name (), uid); |
3540 | |
3541 | SET_DECL_PT_UID (ref->referring->decl, uid); |
3542 | set_alias_uids (n: ref->referring, uid); |
3543 | } |
3544 | } |
3545 | |
3546 | /* Fixup points to analysis info. */ |
3547 | |
3548 | void |
3549 | sem_item_optimizer::fixup_points_to_sets (void) |
3550 | { |
3551 | /* TODO: remove in GCC 9 and trigger PTA re-creation after IPA passes. */ |
3552 | cgraph_node *cnode; |
3553 | |
3554 | FOR_EACH_DEFINED_FUNCTION (cnode) |
3555 | { |
3556 | tree name; |
3557 | unsigned i; |
3558 | function *fn = DECL_STRUCT_FUNCTION (cnode->decl); |
3559 | if (!gimple_in_ssa_p (fun: fn)) |
3560 | continue; |
3561 | |
3562 | FOR_EACH_SSA_NAME (i, name, fn) |
3563 | if (POINTER_TYPE_P (TREE_TYPE (name)) |
3564 | && SSA_NAME_PTR_INFO (name)) |
3565 | fixup_pt_set (pt: &SSA_NAME_PTR_INFO (name)->pt); |
3566 | fixup_pt_set (pt: &fn->gimple_df->escaped); |
3567 | fixup_pt_set (pt: &fn->gimple_df->escaped_return); |
3568 | |
3569 | /* The above gets us to 99% I guess, at least catching the |
3570 | address compares. Below also gets us aliasing correct |
3571 | but as said we're giving leeway to the situation with |
3572 | readonly vars anyway, so ... */ |
3573 | basic_block bb; |
3574 | FOR_EACH_BB_FN (bb, fn) |
3575 | for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); |
3576 | gsi_next (i: &gsi)) |
3577 | { |
3578 | gcall *call = dyn_cast<gcall *> (p: gsi_stmt (i: gsi)); |
3579 | if (call) |
3580 | { |
3581 | fixup_pt_set (pt: gimple_call_use_set (call_stmt: call)); |
3582 | fixup_pt_set (pt: gimple_call_clobber_set (call_stmt: call)); |
3583 | } |
3584 | } |
3585 | } |
3586 | |
3587 | unsigned i; |
3588 | symtab_pair *item; |
3589 | FOR_EACH_VEC_ELT (m_merged_variables, i, item) |
3590 | set_alias_uids (n: item->first, DECL_UID (item->first->decl)); |
3591 | } |
3592 | |
3593 | /* Dump function prints all class members to a FILE with an INDENT. */ |
3594 | |
3595 | void |
3596 | congruence_class::dump (FILE *file, unsigned int indent) const |
3597 | { |
3598 | FPRINTF_SPACES (file, indent, "class with id: %u, hash: %u, items: %u\n" , |
3599 | id, members[0]->get_hash (), members.length ()); |
3600 | |
3601 | FPUTS_SPACES (file, indent + 2, "" ); |
3602 | for (unsigned i = 0; i < members.length (); i++) |
3603 | fprintf (stream: file, format: "%s " , members[i]->node->dump_asm_name ()); |
3604 | |
3605 | fprintf (stream: file, format: "\n" ); |
3606 | } |
3607 | |
3608 | /* Returns true if there's a member that is used from another group. */ |
3609 | |
3610 | bool |
3611 | congruence_class::is_class_used (void) |
3612 | { |
3613 | for (unsigned int i = 0; i < members.length (); i++) |
3614 | if (members[i]->referenced_by_count) |
3615 | return true; |
3616 | |
3617 | return false; |
3618 | } |
3619 | |
3620 | /* Generate pass summary for IPA ICF pass. */ |
3621 | |
3622 | static void |
3623 | ipa_icf_generate_summary (void) |
3624 | { |
3625 | if (!optimizer) |
3626 | optimizer = new sem_item_optimizer (); |
3627 | |
3628 | optimizer->register_hooks (); |
3629 | optimizer->parse_funcs_and_vars (); |
3630 | } |
3631 | |
3632 | /* Write pass summary for IPA ICF pass. */ |
3633 | |
3634 | static void |
3635 | ipa_icf_write_summary (void) |
3636 | { |
3637 | gcc_assert (optimizer); |
3638 | |
3639 | optimizer->write_summary (); |
3640 | } |
3641 | |
3642 | /* Read pass summary for IPA ICF pass. */ |
3643 | |
3644 | static void |
3645 | ipa_icf_read_summary (void) |
3646 | { |
3647 | if (!optimizer) |
3648 | optimizer = new sem_item_optimizer (); |
3649 | |
3650 | optimizer->read_summary (); |
3651 | optimizer->register_hooks (); |
3652 | } |
3653 | |
3654 | /* Semantic equality execution function. */ |
3655 | |
3656 | static unsigned int |
3657 | ipa_icf_driver (void) |
3658 | { |
3659 | gcc_assert (optimizer); |
3660 | |
3661 | bool merged_p = optimizer->execute (); |
3662 | |
3663 | delete optimizer; |
3664 | optimizer = NULL; |
3665 | |
3666 | return merged_p ? TODO_remove_functions : 0; |
3667 | } |
3668 | |
3669 | const pass_data pass_data_ipa_icf = |
3670 | { |
3671 | .type: IPA_PASS, /* type */ |
3672 | .name: "icf" , /* name */ |
3673 | .optinfo_flags: OPTGROUP_IPA, /* optinfo_flags */ |
3674 | .tv_id: TV_IPA_ICF, /* tv_id */ |
3675 | .properties_required: 0, /* properties_required */ |
3676 | .properties_provided: 0, /* properties_provided */ |
3677 | .properties_destroyed: 0, /* properties_destroyed */ |
3678 | .todo_flags_start: 0, /* todo_flags_start */ |
3679 | .todo_flags_finish: 0, /* todo_flags_finish */ |
3680 | }; |
3681 | |
3682 | class pass_ipa_icf : public ipa_opt_pass_d |
3683 | { |
3684 | public: |
3685 | pass_ipa_icf (gcc::context *ctxt) |
3686 | : ipa_opt_pass_d (pass_data_ipa_icf, ctxt, |
3687 | ipa_icf_generate_summary, /* generate_summary */ |
3688 | ipa_icf_write_summary, /* write_summary */ |
3689 | ipa_icf_read_summary, /* read_summary */ |
3690 | NULL, /* |
3691 | write_optimization_summary */ |
3692 | NULL, /* |
3693 | read_optimization_summary */ |
3694 | NULL, /* stmt_fixup */ |
3695 | 0, /* function_transform_todo_flags_start */ |
3696 | NULL, /* function_transform */ |
3697 | NULL) /* variable_transform */ |
3698 | {} |
3699 | |
3700 | /* opt_pass methods: */ |
3701 | bool gate (function *) final override |
3702 | { |
3703 | return in_lto_p || flag_ipa_icf_variables || flag_ipa_icf_functions; |
3704 | } |
3705 | |
3706 | unsigned int execute (function *) final override |
3707 | { |
3708 | return ipa_icf_driver(); |
3709 | } |
3710 | }; // class pass_ipa_icf |
3711 | |
3712 | } // ipa_icf namespace |
3713 | |
3714 | ipa_opt_pass_d * |
3715 | make_pass_ipa_icf (gcc::context *ctxt) |
3716 | { |
3717 | return new ipa_icf::pass_ipa_icf (ctxt); |
3718 | } |
3719 | |
3720 | /* Reset all state within ipa-icf.cc so that we can rerun the compiler |
3721 | within the same process. For use by toplev::finalize. */ |
3722 | |
3723 | void |
3724 | ipa_icf_cc_finalize (void) |
3725 | { |
3726 | ipa_icf::optimizer = NULL; |
3727 | } |
3728 | |