Fix PR ipa/68790
[official-gcc.git] / gcc / ipa-icf.c
blob889a07de697a47fb61823b5d760ba0ddb72a4cef
1 /* Interprocedural Identical Code Folding pass
2 Copyright (C) 2014-2015 Free Software Foundation, Inc.
4 Contributed by Jan Hubicka <hubicka@ucw.cz> and Martin Liska <mliska@suse.cz>
6 This file is part of GCC.
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.
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.
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/>. */
22 /* Interprocedural Identical Code Folding for functions and
23 read-only variables.
25 The goal of this transformation is to discover functions and read-only
26 variables which do have exactly the same semantics.
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.
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.
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 <list>
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 "ipa-prop.h"
77 #include "ipa-inline.h"
78 #include "except.h"
79 #include "attribs.h"
80 #include "print-tree.h"
81 #include "ipa-utils.h"
82 #include "ipa-icf-gimple.h"
83 #include "ipa-icf.h"
84 #include "stor-layout.h"
85 #include "dbgcnt.h"
87 using namespace ipa_icf_gimple;
89 namespace ipa_icf {
91 /* Initialization and computation of symtab node hash, there data
92 are propagated later on. */
94 static sem_item_optimizer *optimizer = NULL;
96 /* Constructor. */
98 symbol_compare_collection::symbol_compare_collection (symtab_node *node)
100 m_references.create (0);
101 m_interposables.create (0);
103 ipa_ref *ref;
105 if (is_a <varpool_node *> (node) && DECL_VIRTUAL_P (node->decl))
106 return;
108 for (unsigned i = 0; node->iterate_reference (i, ref); i++)
110 if (ref->address_matters_p ())
111 m_references.safe_push (ref->referred);
113 if (ref->referred->get_availability () <= AVAIL_INTERPOSABLE)
115 if (ref->address_matters_p ())
116 m_references.safe_push (ref->referred);
117 else
118 m_interposables.safe_push (ref->referred);
122 if (is_a <cgraph_node *> (node))
124 cgraph_node *cnode = dyn_cast <cgraph_node *> (node);
126 for (cgraph_edge *e = cnode->callees; e; e = e->next_callee)
127 if (e->callee->get_availability () <= AVAIL_INTERPOSABLE)
128 m_interposables.safe_push (e->callee);
132 /* Constructor for key value pair, where _ITEM is key and _INDEX is a target. */
134 sem_usage_pair::sem_usage_pair (sem_item *_item, unsigned int _index):
135 item (_item), index (_index)
139 /* Semantic item constructor for a node of _TYPE, where STACK is used
140 for bitmap memory allocation. */
142 sem_item::sem_item (sem_item_type _type,
143 bitmap_obstack *stack): type (_type), m_hash (0)
145 setup (stack);
148 /* Semantic item constructor for a node of _TYPE, where STACK is used
149 for bitmap memory allocation. The item is based on symtab node _NODE
150 with computed _HASH. */
152 sem_item::sem_item (sem_item_type _type, symtab_node *_node,
153 hashval_t _hash, bitmap_obstack *stack): type(_type),
154 node (_node), m_hash (_hash)
156 decl = node->decl;
157 setup (stack);
160 /* Add reference to a semantic TARGET. */
162 void
163 sem_item::add_reference (sem_item *target)
165 refs.safe_push (target);
166 unsigned index = refs.length ();
167 target->usages.safe_push (new sem_usage_pair(this, index));
168 bitmap_set_bit (target->usage_index_bitmap, index);
169 refs_set.add (target->node);
172 /* Initialize internal data structures. Bitmap STACK is used for
173 bitmap memory allocation process. */
175 void
176 sem_item::setup (bitmap_obstack *stack)
178 gcc_checking_assert (node);
180 refs.create (0);
181 tree_refs.create (0);
182 usages.create (0);
183 usage_index_bitmap = BITMAP_ALLOC (stack);
186 sem_item::~sem_item ()
188 for (unsigned i = 0; i < usages.length (); i++)
189 delete usages[i];
191 refs.release ();
192 tree_refs.release ();
193 usages.release ();
195 BITMAP_FREE (usage_index_bitmap);
198 /* Dump function for debugging purpose. */
200 DEBUG_FUNCTION void
201 sem_item::dump (void)
203 if (dump_file)
205 fprintf (dump_file, "[%s] %s (%u) (tree:%p)\n", type == FUNC ? "func" : "var",
206 node->name(), node->order, (void *) node->decl);
207 fprintf (dump_file, " hash: %u\n", get_hash ());
208 fprintf (dump_file, " references: ");
210 for (unsigned i = 0; i < refs.length (); i++)
211 fprintf (dump_file, "%s%s ", refs[i]->node->name (),
212 i < refs.length() - 1 ? "," : "");
214 fprintf (dump_file, "\n");
218 /* Return true if target supports alias symbols. */
220 bool
221 sem_item::target_supports_symbol_aliases_p (void)
223 #if !defined (ASM_OUTPUT_DEF) || (!defined(ASM_OUTPUT_WEAK_ALIAS) && !defined (ASM_WEAKEN_DECL))
224 return false;
225 #else
226 return true;
227 #endif
230 void sem_item::set_hash (hashval_t hash)
232 m_hash = hash;
235 /* Semantic function constructor that uses STACK as bitmap memory stack. */
237 sem_function::sem_function (bitmap_obstack *stack): sem_item (FUNC, stack),
238 m_checker (NULL), m_compared_func (NULL)
240 bb_sizes.create (0);
241 bb_sorted.create (0);
244 /* Constructor based on callgraph node _NODE with computed hash _HASH.
245 Bitmap STACK is used for memory allocation. */
246 sem_function::sem_function (cgraph_node *node, hashval_t hash,
247 bitmap_obstack *stack):
248 sem_item (FUNC, node, hash, stack),
249 m_checker (NULL), m_compared_func (NULL)
251 bb_sizes.create (0);
252 bb_sorted.create (0);
255 sem_function::~sem_function ()
257 for (unsigned i = 0; i < bb_sorted.length (); i++)
258 delete (bb_sorted[i]);
260 bb_sizes.release ();
261 bb_sorted.release ();
264 /* Calculates hash value based on a BASIC_BLOCK. */
266 hashval_t
267 sem_function::get_bb_hash (const sem_bb *basic_block)
269 inchash::hash hstate;
271 hstate.add_int (basic_block->nondbg_stmt_count);
272 hstate.add_int (basic_block->edge_count);
274 return hstate.end ();
277 /* References independent hash function. */
279 hashval_t
280 sem_function::get_hash (void)
282 if (!m_hash)
284 inchash::hash hstate;
285 hstate.add_int (177454); /* Random number for function type. */
287 hstate.add_int (arg_count);
288 hstate.add_int (cfg_checksum);
289 hstate.add_int (gcode_hash);
291 for (unsigned i = 0; i < bb_sorted.length (); i++)
292 hstate.merge_hash (get_bb_hash (bb_sorted[i]));
294 for (unsigned i = 0; i < bb_sizes.length (); i++)
295 hstate.add_int (bb_sizes[i]);
297 /* Add common features of declaration itself. */
298 if (DECL_FUNCTION_SPECIFIC_TARGET (decl))
299 hstate.add_wide_int
300 (cl_target_option_hash
301 (TREE_TARGET_OPTION (DECL_FUNCTION_SPECIFIC_TARGET (decl))));
302 if (DECL_FUNCTION_SPECIFIC_OPTIMIZATION (decl))
303 (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));
308 set_hash (hstate.end ());
311 return m_hash;
314 /* Return ture if A1 and A2 represent equivalent function attribute lists.
315 Based on comp_type_attributes. */
317 bool
318 sem_item::compare_attributes (const_tree a1, const_tree a2)
320 const_tree a;
321 if (a1 == a2)
322 return true;
323 for (a = a1; a != NULL_TREE; a = TREE_CHAIN (a))
325 const struct attribute_spec *as;
326 const_tree attr;
328 as = lookup_attribute_spec (get_attribute_name (a));
329 /* TODO: We can introduce as->affects_decl_identity
330 and as->affects_decl_reference_identity if attribute mismatch
331 gets a common reason to give up on merging. It may not be worth
332 the effort.
333 For example returns_nonnull affects only references, while
334 optimize attribute can be ignored because it is already lowered
335 into flags representation and compared separately. */
336 if (!as)
337 continue;
339 attr = lookup_attribute (as->name, CONST_CAST_TREE (a2));
340 if (!attr || !attribute_value_equal (a, attr))
341 break;
343 if (!a)
345 for (a = a2; a != NULL_TREE; a = TREE_CHAIN (a))
347 const struct attribute_spec *as;
349 as = lookup_attribute_spec (get_attribute_name (a));
350 if (!as)
351 continue;
353 if (!lookup_attribute (as->name, CONST_CAST_TREE (a1)))
354 break;
355 /* We don't need to compare trees again, as we did this
356 already in first loop. */
358 if (!a)
359 return true;
361 /* TODO: As in comp_type_attributes we may want to introduce target hook. */
362 return false;
365 /* Compare properties of symbols N1 and N2 that does not affect semantics of
366 symbol itself but affects semantics of its references from USED_BY (which
367 may be NULL if it is unknown). If comparsion is false, symbols
368 can still be merged but any symbols referring them can't.
370 If ADDRESS is true, do extra checking needed for IPA_REF_ADDR.
372 TODO: We can also split attributes to those that determine codegen of
373 a function body/variable constructor itself and those that are used when
374 referring to it. */
376 bool
377 sem_item::compare_referenced_symbol_properties (symtab_node *used_by,
378 symtab_node *n1,
379 symtab_node *n2,
380 bool address)
382 if (is_a <cgraph_node *> (n1))
384 /* Inline properties matters: we do now want to merge uses of inline
385 function to uses of normal function because inline hint would be lost.
386 We however can merge inline function to noinline because the alias
387 will keep its DECL_DECLARED_INLINE flag.
389 Also ignore inline flag when optimizing for size or when function
390 is known to not be inlinable.
392 TODO: the optimize_size checks can also be assumed to be true if
393 unit has no !optimize_size functions. */
395 if ((!used_by || address || !is_a <cgraph_node *> (used_by)
396 || !opt_for_fn (used_by->decl, optimize_size))
397 && !opt_for_fn (n1->decl, optimize_size)
398 && n1->get_availability () > AVAIL_INTERPOSABLE
399 && (!DECL_UNINLINABLE (n1->decl) || !DECL_UNINLINABLE (n2->decl)))
401 if (DECL_DISREGARD_INLINE_LIMITS (n1->decl)
402 != DECL_DISREGARD_INLINE_LIMITS (n2->decl))
403 return return_false_with_msg
404 ("DECL_DISREGARD_INLINE_LIMITS are different");
406 if (DECL_DECLARED_INLINE_P (n1->decl)
407 != DECL_DECLARED_INLINE_P (n2->decl))
408 return return_false_with_msg ("inline attributes are different");
411 if (DECL_IS_OPERATOR_NEW (n1->decl)
412 != DECL_IS_OPERATOR_NEW (n2->decl))
413 return return_false_with_msg ("operator new flags are different");
416 /* Merging two definitions with a reference to equivalent vtables, but
417 belonging to a different type may result in ipa-polymorphic-call analysis
418 giving a wrong answer about the dynamic type of instance. */
419 if (is_a <varpool_node *> (n1))
421 if ((DECL_VIRTUAL_P (n1->decl) || DECL_VIRTUAL_P (n2->decl))
422 && (DECL_VIRTUAL_P (n1->decl) != DECL_VIRTUAL_P (n2->decl)
423 || !types_must_be_same_for_odr (DECL_CONTEXT (n1->decl),
424 DECL_CONTEXT (n2->decl)))
425 && (!used_by || !is_a <cgraph_node *> (used_by) || address
426 || opt_for_fn (used_by->decl, flag_devirtualize)))
427 return return_false_with_msg
428 ("references to virtual tables can not be merged");
430 if (address && DECL_ALIGN (n1->decl) != DECL_ALIGN (n2->decl))
431 return return_false_with_msg ("alignment mismatch");
433 /* For functions we compare attributes in equals_wpa, because we do
434 not know what attributes may cause codegen differences, but for
435 variables just compare attributes for references - the codegen
436 for constructors is affected only by those attributes that we lower
437 to explicit representation (such as DECL_ALIGN or DECL_SECTION). */
438 if (!compare_attributes (DECL_ATTRIBUTES (n1->decl),
439 DECL_ATTRIBUTES (n2->decl)))
440 return return_false_with_msg ("different var decl attributes");
441 if (comp_type_attributes (TREE_TYPE (n1->decl),
442 TREE_TYPE (n2->decl)) != 1)
443 return return_false_with_msg ("different var type attributes");
446 /* When matching virtual tables, be sure to also match information
447 relevant for polymorphic call analysis. */
448 if (used_by && is_a <varpool_node *> (used_by)
449 && DECL_VIRTUAL_P (used_by->decl))
451 if (DECL_VIRTUAL_P (n1->decl) != DECL_VIRTUAL_P (n2->decl))
452 return return_false_with_msg ("virtual flag mismatch");
453 if (DECL_VIRTUAL_P (n1->decl) && is_a <cgraph_node *> (n1)
454 && (DECL_FINAL_P (n1->decl) != DECL_FINAL_P (n2->decl)))
455 return return_false_with_msg ("final flag mismatch");
457 return true;
460 /* Hash properties that are compared by compare_referenced_symbol_properties. */
462 void
463 sem_item::hash_referenced_symbol_properties (symtab_node *ref,
464 inchash::hash &hstate,
465 bool address)
467 if (is_a <cgraph_node *> (ref))
469 if ((type != FUNC || address || !opt_for_fn (decl, optimize_size))
470 && !opt_for_fn (ref->decl, optimize_size)
471 && !DECL_UNINLINABLE (ref->decl))
473 hstate.add_flag (DECL_DISREGARD_INLINE_LIMITS (ref->decl));
474 hstate.add_flag (DECL_DECLARED_INLINE_P (ref->decl));
476 hstate.add_flag (DECL_IS_OPERATOR_NEW (ref->decl));
478 else if (is_a <varpool_node *> (ref))
480 hstate.add_flag (DECL_VIRTUAL_P (ref->decl));
481 if (address)
482 hstate.add_int (DECL_ALIGN (ref->decl));
487 /* For a given symbol table nodes N1 and N2, we check that FUNCTION_DECLs
488 point to a same function. Comparison can be skipped if IGNORED_NODES
489 contains these nodes. ADDRESS indicate if address is taken. */
491 bool
492 sem_item::compare_symbol_references (
493 hash_map <symtab_node *, sem_item *> &ignored_nodes,
494 symtab_node *n1, symtab_node *n2, bool address)
496 enum availability avail1, avail2;
498 if (n1 == n2)
499 return true;
501 /* Never match variable and function. */
502 if (is_a <varpool_node *> (n1) != is_a <varpool_node *> (n2))
503 return false;
505 if (!compare_referenced_symbol_properties (node, n1, n2, address))
506 return false;
507 if (address && n1->equal_address_to (n2) == 1)
508 return true;
509 if (!address && n1->semantically_equivalent_p (n2))
510 return true;
512 n1 = n1->ultimate_alias_target (&avail1);
513 n2 = n2->ultimate_alias_target (&avail2);
515 if (avail1 > AVAIL_INTERPOSABLE && ignored_nodes.get (n1)
516 && avail2 > AVAIL_INTERPOSABLE && ignored_nodes.get (n2))
517 return true;
519 return return_false_with_msg ("different references");
522 /* If cgraph edges E1 and E2 are indirect calls, verify that
523 ECF flags are the same. */
525 bool sem_function::compare_edge_flags (cgraph_edge *e1, cgraph_edge *e2)
527 if (e1->indirect_info && e2->indirect_info)
529 int e1_flags = e1->indirect_info->ecf_flags;
530 int e2_flags = e2->indirect_info->ecf_flags;
532 if (e1_flags != e2_flags)
533 return return_false_with_msg ("ICF flags are different");
535 else if (e1->indirect_info || e2->indirect_info)
536 return false;
538 return true;
541 /* Return true if parameter I may be used. */
543 bool
544 sem_function::param_used_p (unsigned int i)
546 if (ipa_node_params_sum == NULL)
547 return true;
549 struct ipa_node_params *parms_info = IPA_NODE_REF (get_node ());
551 if (parms_info->descriptors.is_empty ()
552 || parms_info->descriptors.length () <= i)
553 return true;
555 return ipa_is_param_used (IPA_NODE_REF (get_node ()), i);
558 /* Perform additional check needed to match types function parameters that are
559 used. Unlike for normal decls it matters if type is TYPE_RESTRICT and we
560 make an assumption that REFERENCE_TYPE parameters are always non-NULL. */
562 bool
563 sem_function::compatible_parm_types_p (tree parm1, tree parm2)
565 /* Be sure that parameters are TBAA compatible. */
566 if (!func_checker::compatible_types_p (parm1, parm2))
567 return return_false_with_msg ("parameter type is not compatible");
569 if (POINTER_TYPE_P (parm1)
570 && (TYPE_RESTRICT (parm1) != TYPE_RESTRICT (parm2)))
571 return return_false_with_msg ("argument restrict flag mismatch");
573 /* nonnull_arg_p implies non-zero range to REFERENCE types. */
574 if (POINTER_TYPE_P (parm1)
575 && TREE_CODE (parm1) != TREE_CODE (parm2)
576 && opt_for_fn (decl, flag_delete_null_pointer_checks))
577 return return_false_with_msg ("pointer wrt reference mismatch");
579 return true;
582 /* Fast equality function based on knowledge known in WPA. */
584 bool
585 sem_function::equals_wpa (sem_item *item,
586 hash_map <symtab_node *, sem_item *> &ignored_nodes)
588 gcc_assert (item->type == FUNC);
589 cgraph_node *cnode = dyn_cast <cgraph_node *> (node);
590 cgraph_node *cnode2 = dyn_cast <cgraph_node *> (item->node);
592 m_compared_func = static_cast<sem_function *> (item);
594 if (cnode->thunk.thunk_p != cnode2->thunk.thunk_p)
595 return return_false_with_msg ("thunk_p mismatch");
597 if (cnode->thunk.thunk_p)
599 if (cnode->thunk.fixed_offset != cnode2->thunk.fixed_offset)
600 return return_false_with_msg ("thunk fixed_offset mismatch");
601 if (cnode->thunk.virtual_value != cnode2->thunk.virtual_value)
602 return return_false_with_msg ("thunk virtual_value mismatch");
603 if (cnode->thunk.this_adjusting != cnode2->thunk.this_adjusting)
604 return return_false_with_msg ("thunk this_adjusting mismatch");
605 if (cnode->thunk.virtual_offset_p != cnode2->thunk.virtual_offset_p)
606 return return_false_with_msg ("thunk virtual_offset_p mismatch");
607 if (cnode->thunk.add_pointer_bounds_args
608 != cnode2->thunk.add_pointer_bounds_args)
609 return return_false_with_msg ("thunk add_pointer_bounds_args mismatch");
612 /* Compare special function DECL attributes. */
613 if (DECL_FUNCTION_PERSONALITY (decl)
614 != DECL_FUNCTION_PERSONALITY (item->decl))
615 return return_false_with_msg ("function personalities are different");
617 if (DECL_NO_INSTRUMENT_FUNCTION_ENTRY_EXIT (decl)
618 != DECL_NO_INSTRUMENT_FUNCTION_ENTRY_EXIT (item->decl))
619 return return_false_with_msg ("intrument function entry exit "
620 "attributes are different");
622 if (DECL_NO_LIMIT_STACK (decl) != DECL_NO_LIMIT_STACK (item->decl))
623 return return_false_with_msg ("no stack limit attributes are different");
625 if (DECL_CXX_CONSTRUCTOR_P (decl) != DECL_CXX_CONSTRUCTOR_P (item->decl))
626 return return_false_with_msg ("DECL_CXX_CONSTRUCTOR mismatch");
628 if (DECL_CXX_DESTRUCTOR_P (decl) != DECL_CXX_DESTRUCTOR_P (item->decl))
629 return return_false_with_msg ("DECL_CXX_DESTRUCTOR mismatch");
631 /* TODO: pure/const flags mostly matters only for references, except for
632 the fact that codegen takes LOOPING flag as a hint that loops are
633 finite. We may arrange the code to always pick leader that has least
634 specified flags and then this can go into comparing symbol properties. */
635 if (flags_from_decl_or_type (decl) != flags_from_decl_or_type (item->decl))
636 return return_false_with_msg ("decl_or_type flags are different");
638 /* Do not match polymorphic constructors of different types. They calls
639 type memory location for ipa-polymorphic-call and we do not want
640 it to get confused by wrong type. */
641 if (DECL_CXX_CONSTRUCTOR_P (decl)
642 && TREE_CODE (TREE_TYPE (decl)) == METHOD_TYPE)
644 if (TREE_CODE (TREE_TYPE (item->decl)) != METHOD_TYPE)
645 return return_false_with_msg ("DECL_CXX_CONSTURCTOR type mismatch");
646 else if (!func_checker::compatible_polymorphic_types_p
647 (TYPE_METHOD_BASETYPE (TREE_TYPE (decl)),
648 TYPE_METHOD_BASETYPE (TREE_TYPE (item->decl)), false))
649 return return_false_with_msg ("ctor polymorphic type mismatch");
652 /* Checking function TARGET and OPTIMIZATION flags. */
653 cl_target_option *tar1 = target_opts_for_fn (decl);
654 cl_target_option *tar2 = target_opts_for_fn (item->decl);
656 if (tar1 != tar2 && !cl_target_option_eq (tar1, tar2))
658 if (dump_file && (dump_flags & TDF_DETAILS))
660 fprintf (dump_file, "target flags difference");
661 cl_target_option_print_diff (dump_file, 2, tar1, tar2);
664 return return_false_with_msg ("Target flags are different");
667 cl_optimization *opt1 = opts_for_fn (decl);
668 cl_optimization *opt2 = opts_for_fn (item->decl);
670 if (opt1 != opt2 && memcmp (opt1, opt2, sizeof(cl_optimization)))
672 if (dump_file && (dump_flags & TDF_DETAILS))
674 fprintf (dump_file, "optimization flags difference");
675 cl_optimization_print_diff (dump_file, 2, opt1, opt2);
678 return return_false_with_msg ("optimization flags are different");
681 /* Result type checking. */
682 if (!func_checker::compatible_types_p
683 (TREE_TYPE (TREE_TYPE (decl)),
684 TREE_TYPE (TREE_TYPE (m_compared_func->decl))))
685 return return_false_with_msg ("result types are different");
687 /* Checking types of arguments. */
688 tree list1 = TYPE_ARG_TYPES (TREE_TYPE (decl)),
689 list2 = TYPE_ARG_TYPES (TREE_TYPE (m_compared_func->decl));
690 for (unsigned i = 0; list1 && list2;
691 list1 = TREE_CHAIN (list1), list2 = TREE_CHAIN (list2), i++)
693 tree parm1 = TREE_VALUE (list1);
694 tree parm2 = TREE_VALUE (list2);
696 /* This guard is here for function pointer with attributes (pr59927.c). */
697 if (!parm1 || !parm2)
698 return return_false_with_msg ("NULL argument type");
700 /* Verify that types are compatible to ensure that both functions
701 have same calling conventions. */
702 if (!types_compatible_p (parm1, parm2))
703 return return_false_with_msg ("parameter types are not compatible");
705 if (!param_used_p (i))
706 continue;
708 /* Perform additional checks for used parameters. */
709 if (!compatible_parm_types_p (parm1, parm2))
710 return false;
713 if (list1 || list2)
714 return return_false_with_msg ("Mismatched number of parameters");
716 if (node->num_references () != item->node->num_references ())
717 return return_false_with_msg ("different number of references");
719 /* Checking function attributes.
720 This is quadratic in number of attributes */
721 if (comp_type_attributes (TREE_TYPE (decl),
722 TREE_TYPE (item->decl)) != 1)
723 return return_false_with_msg ("different type attributes");
724 if (!compare_attributes (DECL_ATTRIBUTES (decl),
725 DECL_ATTRIBUTES (item->decl)))
726 return return_false_with_msg ("different decl attributes");
728 /* The type of THIS pointer type memory location for
729 ipa-polymorphic-call-analysis. */
730 if (opt_for_fn (decl, flag_devirtualize)
731 && (TREE_CODE (TREE_TYPE (decl)) == METHOD_TYPE
732 || TREE_CODE (TREE_TYPE (item->decl)) == METHOD_TYPE)
733 && param_used_p (0)
734 && compare_polymorphic_p ())
736 if (TREE_CODE (TREE_TYPE (decl)) != TREE_CODE (TREE_TYPE (item->decl)))
737 return return_false_with_msg ("METHOD_TYPE and FUNCTION_TYPE mismatch");
738 if (!func_checker::compatible_polymorphic_types_p
739 (TYPE_METHOD_BASETYPE (TREE_TYPE (decl)),
740 TYPE_METHOD_BASETYPE (TREE_TYPE (item->decl)), false))
741 return return_false_with_msg ("THIS pointer ODR type mismatch");
744 ipa_ref *ref = NULL, *ref2 = NULL;
745 for (unsigned i = 0; node->iterate_reference (i, ref); i++)
747 item->node->iterate_reference (i, ref2);
749 if (ref->use != ref2->use)
750 return return_false_with_msg ("reference use mismatch");
752 if (!compare_symbol_references (ignored_nodes, ref->referred,
753 ref2->referred,
754 ref->address_matters_p ()))
755 return false;
758 cgraph_edge *e1 = dyn_cast <cgraph_node *> (node)->callees;
759 cgraph_edge *e2 = dyn_cast <cgraph_node *> (item->node)->callees;
761 while (e1 && e2)
763 if (!compare_symbol_references (ignored_nodes, e1->callee,
764 e2->callee, false))
765 return false;
766 if (!compare_edge_flags (e1, e2))
767 return false;
769 e1 = e1->next_callee;
770 e2 = e2->next_callee;
773 if (e1 || e2)
774 return return_false_with_msg ("different number of calls");
776 e1 = dyn_cast <cgraph_node *> (node)->indirect_calls;
777 e2 = dyn_cast <cgraph_node *> (item->node)->indirect_calls;
779 while (e1 && e2)
781 if (!compare_edge_flags (e1, e2))
782 return false;
784 e1 = e1->next_callee;
785 e2 = e2->next_callee;
788 if (e1 || e2)
789 return return_false_with_msg ("different number of indirect calls");
791 return true;
794 /* Update hash by address sensitive references. We iterate over all
795 sensitive references (address_matters_p) and we hash ultime alias
796 target of these nodes, which can improve a semantic item hash.
798 Also hash in referenced symbols properties. This can be done at any time
799 (as the properties should not change), but it is convenient to do it here
800 while we walk the references anyway. */
802 void
803 sem_item::update_hash_by_addr_refs (hash_map <symtab_node *,
804 sem_item *> &m_symtab_node_map)
806 ipa_ref* ref;
807 inchash::hash hstate (get_hash ());
809 for (unsigned i = 0; node->iterate_reference (i, ref); i++)
811 hstate.add_int (ref->use);
812 hash_referenced_symbol_properties (ref->referred, hstate,
813 ref->use == IPA_REF_ADDR);
814 if (ref->address_matters_p () || !m_symtab_node_map.get (ref->referred))
815 hstate.add_int (ref->referred->ultimate_alias_target ()->order);
818 if (is_a <cgraph_node *> (node))
820 for (cgraph_edge *e = dyn_cast <cgraph_node *> (node)->callers; e;
821 e = e->next_caller)
823 sem_item **result = m_symtab_node_map.get (e->callee);
824 hash_referenced_symbol_properties (e->callee, hstate, false);
825 if (!result)
826 hstate.add_int (e->callee->ultimate_alias_target ()->order);
830 set_hash (hstate.end ());
833 /* Update hash by computed local hash values taken from different
834 semantic items.
835 TODO: stronger SCC based hashing would be desirable here. */
837 void
838 sem_item::update_hash_by_local_refs (hash_map <symtab_node *,
839 sem_item *> &m_symtab_node_map)
841 ipa_ref* ref;
842 inchash::hash state (get_hash ());
844 for (unsigned j = 0; node->iterate_reference (j, ref); j++)
846 sem_item **result = m_symtab_node_map.get (ref->referring);
847 if (result)
848 state.merge_hash ((*result)->get_hash ());
851 if (type == FUNC)
853 for (cgraph_edge *e = dyn_cast <cgraph_node *> (node)->callees; e;
854 e = e->next_callee)
856 sem_item **result = m_symtab_node_map.get (e->caller);
857 if (result)
858 state.merge_hash ((*result)->get_hash ());
862 global_hash = state.end ();
865 /* Returns true if the item equals to ITEM given as argument. */
867 bool
868 sem_function::equals (sem_item *item,
869 hash_map <symtab_node *, sem_item *> &)
871 gcc_assert (item->type == FUNC);
872 bool eq = equals_private (item);
874 if (m_checker != NULL)
876 delete m_checker;
877 m_checker = NULL;
880 if (dump_file && (dump_flags & TDF_DETAILS))
881 fprintf (dump_file,
882 "Equals called for:%s:%s (%u:%u) (%s:%s) with result: %s\n\n",
883 xstrdup_for_dump (node->name()),
884 xstrdup_for_dump (item->node->name ()),
885 node->order,
886 item->node->order,
887 xstrdup_for_dump (node->asm_name ()),
888 xstrdup_for_dump (item->node->asm_name ()),
889 eq ? "true" : "false");
891 return eq;
894 /* Processes function equality comparison. */
896 bool
897 sem_function::equals_private (sem_item *item)
899 if (item->type != FUNC)
900 return false;
902 basic_block bb1, bb2;
903 edge e1, e2;
904 edge_iterator ei1, ei2;
905 bool result = true;
906 tree arg1, arg2;
908 m_compared_func = static_cast<sem_function *> (item);
910 gcc_assert (decl != item->decl);
912 if (bb_sorted.length () != m_compared_func->bb_sorted.length ()
913 || edge_count != m_compared_func->edge_count
914 || cfg_checksum != m_compared_func->cfg_checksum)
915 return return_false ();
917 m_checker = new func_checker (decl, m_compared_func->decl,
918 compare_polymorphic_p (),
919 false,
920 &refs_set,
921 &m_compared_func->refs_set);
922 arg1 = DECL_ARGUMENTS (decl);
923 arg2 = DECL_ARGUMENTS (m_compared_func->decl);
924 for (unsigned i = 0;
925 arg1 && arg2; arg1 = DECL_CHAIN (arg1), arg2 = DECL_CHAIN (arg2), i++)
927 if (!types_compatible_p (TREE_TYPE (arg1), TREE_TYPE (arg2)))
928 return return_false_with_msg ("argument types are not compatible");
929 if (!param_used_p (i))
930 continue;
931 /* Perform additional checks for used parameters. */
932 if (!compatible_parm_types_p (TREE_TYPE (arg1), TREE_TYPE (arg2)))
933 return false;
934 if (!m_checker->compare_decl (arg1, arg2))
935 return return_false ();
937 if (arg1 || arg2)
938 return return_false_with_msg ("Mismatched number of arguments");
940 if (!dyn_cast <cgraph_node *> (node)->has_gimple_body_p ())
941 return true;
943 /* Fill-up label dictionary. */
944 for (unsigned i = 0; i < bb_sorted.length (); ++i)
946 m_checker->parse_labels (bb_sorted[i]);
947 m_checker->parse_labels (m_compared_func->bb_sorted[i]);
950 /* Checking all basic blocks. */
951 for (unsigned i = 0; i < bb_sorted.length (); ++i)
952 if(!m_checker->compare_bb (bb_sorted[i], m_compared_func->bb_sorted[i]))
953 return return_false();
955 dump_message ("All BBs are equal\n");
957 auto_vec <int> bb_dict;
959 /* Basic block edges check. */
960 for (unsigned i = 0; i < bb_sorted.length (); ++i)
962 bb1 = bb_sorted[i]->bb;
963 bb2 = m_compared_func->bb_sorted[i]->bb;
965 ei2 = ei_start (bb2->preds);
967 for (ei1 = ei_start (bb1->preds); ei_cond (ei1, &e1); ei_next (&ei1))
969 ei_cond (ei2, &e2);
971 if (e1->flags != e2->flags)
972 return return_false_with_msg ("flags comparison returns false");
974 if (!bb_dict_test (&bb_dict, e1->src->index, e2->src->index))
975 return return_false_with_msg ("edge comparison returns false");
977 if (!bb_dict_test (&bb_dict, e1->dest->index, e2->dest->index))
978 return return_false_with_msg ("BB comparison returns false");
980 if (!m_checker->compare_edge (e1, e2))
981 return return_false_with_msg ("edge comparison returns false");
983 ei_next (&ei2);
987 /* Basic block PHI nodes comparison. */
988 for (unsigned i = 0; i < bb_sorted.length (); i++)
989 if (!compare_phi_node (bb_sorted[i]->bb, m_compared_func->bb_sorted[i]->bb))
990 return return_false_with_msg ("PHI node comparison returns false");
992 return result;
995 /* Set LOCAL_P of NODE to true if DATA is non-NULL.
996 Helper for call_for_symbol_thunks_and_aliases. */
998 static bool
999 set_local (cgraph_node *node, void *data)
1001 node->local.local = data != NULL;
1002 return false;
1005 /* TREE_ADDRESSABLE of NODE to true.
1006 Helper for call_for_symbol_thunks_and_aliases. */
1008 static bool
1009 set_addressable (varpool_node *node, void *)
1011 TREE_ADDRESSABLE (node->decl) = 1;
1012 return false;
1015 /* Clear DECL_RTL of NODE.
1016 Helper for call_for_symbol_thunks_and_aliases. */
1018 static bool
1019 clear_decl_rtl (symtab_node *node, void *)
1021 SET_DECL_RTL (node->decl, NULL);
1022 return false;
1025 /* Redirect all callers of N and its aliases to TO. Remove aliases if
1026 possible. Return number of redirections made. */
1028 static int
1029 redirect_all_callers (cgraph_node *n, cgraph_node *to)
1031 int nredirected = 0;
1032 ipa_ref *ref;
1033 cgraph_edge *e = n->callers;
1035 while (e)
1037 /* Redirecting thunks to interposable symbols or symbols in other sections
1038 may not be supported by target output code. Play safe for now and
1039 punt on redirection. */
1040 if (!e->caller->thunk.thunk_p)
1042 struct cgraph_edge *nexte = e->next_caller;
1043 e->redirect_callee (to);
1044 e = nexte;
1045 nredirected++;
1047 else
1048 e = e->next_callee;
1050 for (unsigned i = 0; n->iterate_direct_aliases (i, ref);)
1052 bool removed = false;
1053 cgraph_node *n_alias = dyn_cast <cgraph_node *> (ref->referring);
1055 if ((DECL_COMDAT_GROUP (n->decl)
1056 && (DECL_COMDAT_GROUP (n->decl)
1057 == DECL_COMDAT_GROUP (n_alias->decl)))
1058 || (n_alias->get_availability () > AVAIL_INTERPOSABLE
1059 && n->get_availability () > AVAIL_INTERPOSABLE))
1061 nredirected += redirect_all_callers (n_alias, to);
1062 if (n_alias->can_remove_if_no_direct_calls_p ()
1063 && !n_alias->call_for_symbol_and_aliases (cgraph_node::has_thunk_p,
1064 NULL, true)
1065 && !n_alias->has_aliases_p ())
1066 n_alias->remove ();
1068 if (!removed)
1069 i++;
1071 return nredirected;
1074 /* Merges instance with an ALIAS_ITEM, where alias, thunk or redirection can
1075 be applied. */
1077 bool
1078 sem_function::merge (sem_item *alias_item)
1080 gcc_assert (alias_item->type == FUNC);
1082 sem_function *alias_func = static_cast<sem_function *> (alias_item);
1084 cgraph_node *original = get_node ();
1085 cgraph_node *local_original = NULL;
1086 cgraph_node *alias = alias_func->get_node ();
1088 bool create_wrapper = false;
1089 bool create_alias = false;
1090 bool redirect_callers = false;
1091 bool remove = false;
1093 bool original_discardable = false;
1094 bool original_discarded = false;
1096 bool original_address_matters = original->address_matters_p ();
1097 bool alias_address_matters = alias->address_matters_p ();
1099 if (DECL_EXTERNAL (alias->decl))
1101 if (dump_file)
1102 fprintf (dump_file, "Not unifying; alias is external.\n\n");
1103 return false;
1106 if (DECL_NO_INLINE_WARNING_P (original->decl)
1107 != DECL_NO_INLINE_WARNING_P (alias->decl))
1109 if (dump_file)
1110 fprintf (dump_file,
1111 "Not unifying; "
1112 "DECL_NO_INLINE_WARNING mismatch.\n\n");
1113 return false;
1116 /* Do not attempt to mix functions from different user sections;
1117 we do not know what user intends with those. */
1118 if (((DECL_SECTION_NAME (original->decl) && !original->implicit_section)
1119 || (DECL_SECTION_NAME (alias->decl) && !alias->implicit_section))
1120 && DECL_SECTION_NAME (original->decl) != DECL_SECTION_NAME (alias->decl))
1122 if (dump_file)
1123 fprintf (dump_file,
1124 "Not unifying; "
1125 "original and alias are in different sections.\n\n");
1126 return false;
1129 /* See if original is in a section that can be discarded if the main
1130 symbol is not used. */
1132 if (original->can_be_discarded_p ())
1133 original_discardable = true;
1134 /* Also consider case where we have resolution info and we know that
1135 original's definition is not going to be used. In this case we can not
1136 create alias to original. */
1137 if (node->resolution != LDPR_UNKNOWN
1138 && !decl_binds_to_current_def_p (node->decl))
1139 original_discardable = original_discarded = true;
1141 /* Creating a symtab alias is the optimal way to merge.
1142 It however can not be used in the following cases:
1144 1) if ORIGINAL and ALIAS may be possibly compared for address equality.
1145 2) if ORIGINAL is in a section that may be discarded by linker or if
1146 it is an external functions where we can not create an alias
1147 (ORIGINAL_DISCARDABLE)
1148 3) if target do not support symbol aliases.
1149 4) original and alias lie in different comdat groups.
1151 If we can not produce alias, we will turn ALIAS into WRAPPER of ORIGINAL
1152 and/or redirect all callers from ALIAS to ORIGINAL. */
1153 if ((original_address_matters && alias_address_matters)
1154 || (original_discardable
1155 && (!DECL_COMDAT_GROUP (alias->decl)
1156 || (DECL_COMDAT_GROUP (alias->decl)
1157 != DECL_COMDAT_GROUP (original->decl))))
1158 || original_discarded
1159 || !sem_item::target_supports_symbol_aliases_p ()
1160 || DECL_COMDAT_GROUP (alias->decl) != DECL_COMDAT_GROUP (original->decl))
1162 /* First see if we can produce wrapper. */
1164 /* Symbol properties that matter for references must be preserved.
1165 TODO: We can produce wrapper, but we need to produce alias of ORIGINAL
1166 with proper properties. */
1167 if (!sem_item::compare_referenced_symbol_properties (NULL, original, alias,
1168 alias->address_taken))
1170 if (dump_file)
1171 fprintf (dump_file,
1172 "Wrapper cannot be created because referenced symbol "
1173 "properties mismatch\n");
1175 /* Do not turn function in one comdat group into wrapper to another
1176 comdat group. Other compiler producing the body of the
1177 another comdat group may make opossite decision and with unfortunate
1178 linker choices this may close a loop. */
1179 else if (DECL_COMDAT_GROUP (original->decl)
1180 && DECL_COMDAT_GROUP (alias->decl)
1181 && (DECL_COMDAT_GROUP (alias->decl)
1182 != DECL_COMDAT_GROUP (original->decl)))
1184 if (dump_file)
1185 fprintf (dump_file,
1186 "Wrapper cannot be created because of COMDAT\n");
1188 else if (DECL_STATIC_CHAIN (alias->decl))
1190 if (dump_file)
1191 fprintf (dump_file,
1192 "Can not create wrapper of nested functions.\n");
1194 /* TODO: We can also deal with variadic functions never calling
1195 VA_START. */
1196 else if (stdarg_p (TREE_TYPE (alias->decl)))
1198 if (dump_file)
1199 fprintf (dump_file,
1200 "can not create wrapper of stdarg function.\n");
1202 else if (inline_summaries
1203 && inline_summaries->get (alias)->self_size <= 2)
1205 if (dump_file)
1206 fprintf (dump_file, "Wrapper creation is not "
1207 "profitable (function is too small).\n");
1209 /* If user paid attention to mark function noinline, assume it is
1210 somewhat special and do not try to turn it into a wrapper that can
1211 not be undone by inliner. */
1212 else if (lookup_attribute ("noinline", DECL_ATTRIBUTES (alias->decl)))
1214 if (dump_file)
1215 fprintf (dump_file, "Wrappers are not created for noinline.\n");
1217 else
1218 create_wrapper = true;
1220 /* We can redirect local calls in the case both alias and orignal
1221 are not interposable. */
1222 redirect_callers
1223 = alias->get_availability () > AVAIL_INTERPOSABLE
1224 && original->get_availability () > AVAIL_INTERPOSABLE
1225 && !alias->instrumented_version;
1226 /* TODO: We can redirect, but we need to produce alias of ORIGINAL
1227 with proper properties. */
1228 if (!sem_item::compare_referenced_symbol_properties (NULL, original, alias,
1229 alias->address_taken))
1230 redirect_callers = false;
1232 if (!redirect_callers && !create_wrapper)
1234 if (dump_file)
1235 fprintf (dump_file, "Not unifying; can not redirect callers nor "
1236 "produce wrapper\n\n");
1237 return false;
1240 /* Work out the symbol the wrapper should call.
1241 If ORIGINAL is interposable, we need to call a local alias.
1242 Also produce local alias (if possible) as an optimization.
1244 Local aliases can not be created inside comdat groups because that
1245 prevents inlining. */
1246 if (!original_discardable && !original->get_comdat_group ())
1248 local_original
1249 = dyn_cast <cgraph_node *> (original->noninterposable_alias ());
1250 if (!local_original
1251 && original->get_availability () > AVAIL_INTERPOSABLE)
1252 local_original = original;
1254 /* If we can not use local alias, fallback to the original
1255 when possible. */
1256 else if (original->get_availability () > AVAIL_INTERPOSABLE)
1257 local_original = original;
1259 /* If original is COMDAT local, we can not really redirect calls outside
1260 of its comdat group to it. */
1261 if (original->comdat_local_p ())
1262 redirect_callers = false;
1263 if (!local_original)
1265 if (dump_file)
1266 fprintf (dump_file, "Not unifying; "
1267 "can not produce local alias.\n\n");
1268 return false;
1271 if (!redirect_callers && !create_wrapper)
1273 if (dump_file)
1274 fprintf (dump_file, "Not unifying; "
1275 "can not redirect callers nor produce a wrapper\n\n");
1276 return false;
1278 if (!create_wrapper
1279 && !alias->call_for_symbol_and_aliases (cgraph_node::has_thunk_p,
1280 NULL, true)
1281 && !alias->can_remove_if_no_direct_calls_p ())
1283 if (dump_file)
1284 fprintf (dump_file, "Not unifying; can not make wrapper and "
1285 "function has other uses than direct calls\n\n");
1286 return false;
1289 else
1290 create_alias = true;
1292 if (redirect_callers)
1294 int nredirected = redirect_all_callers (alias, local_original);
1296 if (nredirected)
1298 alias->icf_merged = true;
1299 local_original->icf_merged = true;
1301 if (dump_file && nredirected)
1302 fprintf (dump_file, "%i local calls have been "
1303 "redirected.\n", nredirected);
1306 /* If all callers was redirected, do not produce wrapper. */
1307 if (alias->can_remove_if_no_direct_calls_p ()
1308 && !alias->has_aliases_p ())
1310 create_wrapper = false;
1311 remove = true;
1313 gcc_assert (!create_alias);
1315 else if (create_alias)
1317 alias->icf_merged = true;
1319 /* Remove the function's body. */
1320 ipa_merge_profiles (original, alias);
1321 alias->release_body (true);
1322 alias->reset ();
1323 /* Notice global symbol possibly produced RTL. */
1324 ((symtab_node *)alias)->call_for_symbol_and_aliases (clear_decl_rtl,
1325 NULL, true);
1327 /* Create the alias. */
1328 cgraph_node::create_alias (alias_func->decl, decl);
1329 alias->resolve_alias (original);
1331 original->call_for_symbol_thunks_and_aliases
1332 (set_local, (void *)(size_t) original->local_p (), true);
1334 if (dump_file)
1335 fprintf (dump_file, "Unified; Function alias has been created.\n\n");
1337 if (create_wrapper)
1339 gcc_assert (!create_alias);
1340 alias->icf_merged = true;
1341 local_original->icf_merged = true;
1343 ipa_merge_profiles (local_original, alias, true);
1344 alias->create_wrapper (local_original);
1346 if (dump_file)
1347 fprintf (dump_file, "Unified; Wrapper has been created.\n\n");
1350 /* It's possible that redirection can hit thunks that block
1351 redirection opportunities. */
1352 gcc_assert (alias->icf_merged || remove || redirect_callers);
1353 original->icf_merged = true;
1355 /* We use merged flag to track cases where COMDAT function is known to be
1356 compatible its callers. If we merged in non-COMDAT, we need to give up
1357 on this optimization. */
1358 if (original->merged_comdat && !alias->merged_comdat)
1360 if (dump_file)
1361 fprintf (dump_file, "Dropping merged_comdat flag.\n\n");
1362 if (local_original)
1363 local_original->merged_comdat = false;
1364 original->merged_comdat = false;
1367 if (remove)
1369 ipa_merge_profiles (original, alias);
1370 alias->release_body ();
1371 alias->reset ();
1372 alias->body_removed = true;
1373 alias->icf_merged = true;
1374 if (dump_file)
1375 fprintf (dump_file, "Unified; Function body was removed.\n");
1378 return true;
1381 /* Semantic item initialization function. */
1383 void
1384 sem_function::init (void)
1386 if (in_lto_p)
1387 get_node ()->get_untransformed_body ();
1389 tree fndecl = node->decl;
1390 function *func = DECL_STRUCT_FUNCTION (fndecl);
1392 gcc_assert (func);
1393 gcc_assert (SSANAMES (func));
1395 ssa_names_size = SSANAMES (func)->length ();
1396 node = node;
1398 decl = fndecl;
1399 region_tree = func->eh->region_tree;
1401 /* iterating all function arguments. */
1402 arg_count = count_formal_params (fndecl);
1404 edge_count = n_edges_for_fn (func);
1405 cgraph_node *cnode = dyn_cast <cgraph_node *> (node);
1406 if (!cnode->thunk.thunk_p)
1408 cfg_checksum = coverage_compute_cfg_checksum (func);
1410 inchash::hash hstate;
1412 basic_block bb;
1413 FOR_EACH_BB_FN (bb, func)
1415 unsigned nondbg_stmt_count = 0;
1417 edge e;
1418 for (edge_iterator ei = ei_start (bb->preds); ei_cond (ei, &e);
1419 ei_next (&ei))
1420 cfg_checksum = iterative_hash_host_wide_int (e->flags,
1421 cfg_checksum);
1423 for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi);
1424 gsi_next (&gsi))
1426 gimple *stmt = gsi_stmt (gsi);
1428 if (gimple_code (stmt) != GIMPLE_DEBUG
1429 && gimple_code (stmt) != GIMPLE_PREDICT)
1431 hash_stmt (stmt, hstate);
1432 nondbg_stmt_count++;
1436 gcode_hash = hstate.end ();
1437 bb_sizes.safe_push (nondbg_stmt_count);
1439 /* Inserting basic block to hash table. */
1440 sem_bb *semantic_bb = new sem_bb (bb, nondbg_stmt_count,
1441 EDGE_COUNT (bb->preds)
1442 + EDGE_COUNT (bb->succs));
1444 bb_sorted.safe_push (semantic_bb);
1447 else
1449 cfg_checksum = 0;
1450 inchash::hash hstate;
1451 hstate.add_wide_int (cnode->thunk.fixed_offset);
1452 hstate.add_wide_int (cnode->thunk.virtual_value);
1453 hstate.add_flag (cnode->thunk.this_adjusting);
1454 hstate.add_flag (cnode->thunk.virtual_offset_p);
1455 hstate.add_flag (cnode->thunk.add_pointer_bounds_args);
1456 gcode_hash = hstate.end ();
1460 /* Accumulate to HSTATE a hash of expression EXP.
1461 Identical to inchash::add_expr, but guaranteed to be stable across LTO
1462 and DECL equality classes. */
1464 void
1465 sem_item::add_expr (const_tree exp, inchash::hash &hstate)
1467 if (exp == NULL_TREE)
1469 hstate.merge_hash (0);
1470 return;
1473 /* Handled component can be matched in a cureful way proving equivalence
1474 even if they syntactically differ. Just skip them. */
1475 STRIP_NOPS (exp);
1476 while (handled_component_p (exp))
1477 exp = TREE_OPERAND (exp, 0);
1479 enum tree_code code = TREE_CODE (exp);
1480 hstate.add_int (code);
1482 switch (code)
1484 /* Use inchash::add_expr for everything that is LTO stable. */
1485 case VOID_CST:
1486 case INTEGER_CST:
1487 case REAL_CST:
1488 case FIXED_CST:
1489 case STRING_CST:
1490 case COMPLEX_CST:
1491 case VECTOR_CST:
1492 inchash::add_expr (exp, hstate);
1493 break;
1494 case CONSTRUCTOR:
1496 unsigned HOST_WIDE_INT idx;
1497 tree value;
1499 hstate.add_wide_int (int_size_in_bytes (TREE_TYPE (exp)));
1501 FOR_EACH_CONSTRUCTOR_VALUE (CONSTRUCTOR_ELTS (exp), idx, value)
1502 if (value)
1503 add_expr (value, hstate);
1504 break;
1506 case ADDR_EXPR:
1507 case FDESC_EXPR:
1508 add_expr (get_base_address (TREE_OPERAND (exp, 0)), hstate);
1509 break;
1510 case SSA_NAME:
1511 case VAR_DECL:
1512 case CONST_DECL:
1513 case PARM_DECL:
1514 hstate.add_wide_int (int_size_in_bytes (TREE_TYPE (exp)));
1515 break;
1516 case MEM_REF:
1517 case POINTER_PLUS_EXPR:
1518 case MINUS_EXPR:
1519 case RANGE_EXPR:
1520 add_expr (TREE_OPERAND (exp, 0), hstate);
1521 add_expr (TREE_OPERAND (exp, 1), hstate);
1522 break;
1523 case PLUS_EXPR:
1525 inchash::hash one, two;
1526 add_expr (TREE_OPERAND (exp, 0), one);
1527 add_expr (TREE_OPERAND (exp, 1), two);
1528 hstate.add_commutative (one, two);
1530 break;
1531 CASE_CONVERT:
1532 hstate.add_wide_int (int_size_in_bytes (TREE_TYPE (exp)));
1533 return add_expr (TREE_OPERAND (exp, 0), hstate);
1534 default:
1535 break;
1539 /* Accumulate to HSTATE a hash of type t.
1540 TYpes that may end up being compatible after LTO type merging needs to have
1541 the same hash. */
1543 void
1544 sem_item::add_type (const_tree type, inchash::hash &hstate)
1546 if (type == NULL_TREE)
1548 hstate.merge_hash (0);
1549 return;
1552 type = TYPE_MAIN_VARIANT (type);
1554 hstate.add_int (TYPE_MODE (type));
1556 if (TREE_CODE (type) == COMPLEX_TYPE)
1558 hstate.add_int (COMPLEX_TYPE);
1559 sem_item::add_type (TREE_TYPE (type), hstate);
1561 else if (INTEGRAL_TYPE_P (type))
1563 hstate.add_int (INTEGER_TYPE);
1564 hstate.add_flag (TYPE_UNSIGNED (type));
1565 hstate.add_int (TYPE_PRECISION (type));
1567 else if (VECTOR_TYPE_P (type))
1569 hstate.add_int (VECTOR_TYPE);
1570 hstate.add_int (TYPE_PRECISION (type));
1571 sem_item::add_type (TREE_TYPE (type), hstate);
1573 else if (TREE_CODE (type) == ARRAY_TYPE)
1575 hstate.add_int (ARRAY_TYPE);
1576 /* Do not hash size, so complete and incomplete types can match. */
1577 sem_item::add_type (TREE_TYPE (type), hstate);
1579 else if (RECORD_OR_UNION_TYPE_P (type))
1581 gcc_checking_assert (COMPLETE_TYPE_P (type));
1582 hashval_t *val = optimizer->m_type_hash_cache.get (type);
1584 if (!val)
1586 inchash::hash hstate2;
1587 unsigned nf;
1588 tree f;
1589 hashval_t hash;
1591 hstate2.add_int (RECORD_TYPE);
1592 gcc_assert (COMPLETE_TYPE_P (type));
1594 for (f = TYPE_FIELDS (type), nf = 0; f; f = TREE_CHAIN (f))
1595 if (TREE_CODE (f) == FIELD_DECL)
1597 add_type (TREE_TYPE (f), hstate2);
1598 nf++;
1601 hstate2.add_int (nf);
1602 hash = hstate2.end ();
1603 hstate.add_wide_int (hash);
1604 optimizer->m_type_hash_cache.put (type, hash);
1606 else
1607 hstate.add_wide_int (*val);
1611 /* Improve accumulated hash for HSTATE based on a gimple statement STMT. */
1613 void
1614 sem_function::hash_stmt (gimple *stmt, inchash::hash &hstate)
1616 enum gimple_code code = gimple_code (stmt);
1618 hstate.add_int (code);
1620 switch (code)
1622 case GIMPLE_SWITCH:
1623 add_expr (gimple_switch_index (as_a <gswitch *> (stmt)), hstate);
1624 break;
1625 case GIMPLE_ASSIGN:
1626 hstate.add_int (gimple_assign_rhs_code (stmt));
1627 if (commutative_tree_code (gimple_assign_rhs_code (stmt))
1628 || commutative_ternary_tree_code (gimple_assign_rhs_code (stmt)))
1630 inchash::hash one, two;
1632 add_expr (gimple_assign_rhs1 (stmt), one);
1633 add_type (TREE_TYPE (gimple_assign_rhs1 (stmt)), one);
1634 add_expr (gimple_assign_rhs2 (stmt), two);
1635 hstate.add_commutative (one, two);
1636 if (commutative_ternary_tree_code (gimple_assign_rhs_code (stmt)))
1638 add_expr (gimple_assign_rhs3 (stmt), hstate);
1639 add_type (TREE_TYPE (gimple_assign_rhs3 (stmt)), hstate);
1641 add_expr (gimple_assign_lhs (stmt), hstate);
1642 add_type (TREE_TYPE (gimple_assign_lhs (stmt)), two);
1643 break;
1645 /* ... fall through ... */
1646 case GIMPLE_CALL:
1647 case GIMPLE_ASM:
1648 case GIMPLE_COND:
1649 case GIMPLE_GOTO:
1650 case GIMPLE_RETURN:
1651 /* All these statements are equivalent if their operands are. */
1652 for (unsigned i = 0; i < gimple_num_ops (stmt); ++i)
1654 add_expr (gimple_op (stmt, i), hstate);
1655 if (gimple_op (stmt, i))
1656 add_type (TREE_TYPE (gimple_op (stmt, i)), hstate);
1658 default:
1659 break;
1664 /* Return true if polymorphic comparison must be processed. */
1666 bool
1667 sem_function::compare_polymorphic_p (void)
1669 struct cgraph_edge *e;
1671 if (!opt_for_fn (get_node ()->decl, flag_devirtualize))
1672 return false;
1673 if (get_node ()->indirect_calls != NULL)
1674 return true;
1675 /* TODO: We can do simple propagation determining what calls may lead to
1676 a polymorphic call. */
1677 for (e = get_node ()->callees; e; e = e->next_callee)
1678 if (e->callee->definition
1679 && opt_for_fn (e->callee->decl, flag_devirtualize))
1680 return true;
1681 return false;
1684 /* For a given call graph NODE, the function constructs new
1685 semantic function item. */
1687 sem_function *
1688 sem_function::parse (cgraph_node *node, bitmap_obstack *stack)
1690 tree fndecl = node->decl;
1691 function *func = DECL_STRUCT_FUNCTION (fndecl);
1693 if (!func || (!node->has_gimple_body_p () && !node->thunk.thunk_p))
1694 return NULL;
1696 if (lookup_attribute_by_prefix ("omp ", DECL_ATTRIBUTES (node->decl)) != NULL)
1697 return NULL;
1699 sem_function *f = new sem_function (node, 0, stack);
1701 f->init ();
1703 return f;
1706 /* For given basic blocks BB1 and BB2 (from functions FUNC1 and FUNC),
1707 return true if phi nodes are semantically equivalent in these blocks . */
1709 bool
1710 sem_function::compare_phi_node (basic_block bb1, basic_block bb2)
1712 gphi_iterator si1, si2;
1713 gphi *phi1, *phi2;
1714 unsigned size1, size2, i;
1715 tree t1, t2;
1716 edge e1, e2;
1718 gcc_assert (bb1 != NULL);
1719 gcc_assert (bb2 != NULL);
1721 si2 = gsi_start_phis (bb2);
1722 for (si1 = gsi_start_phis (bb1); !gsi_end_p (si1);
1723 gsi_next (&si1))
1725 gsi_next_nonvirtual_phi (&si1);
1726 gsi_next_nonvirtual_phi (&si2);
1728 if (gsi_end_p (si1) && gsi_end_p (si2))
1729 break;
1731 if (gsi_end_p (si1) || gsi_end_p (si2))
1732 return return_false();
1734 phi1 = si1.phi ();
1735 phi2 = si2.phi ();
1737 tree phi_result1 = gimple_phi_result (phi1);
1738 tree phi_result2 = gimple_phi_result (phi2);
1740 if (!m_checker->compare_operand (phi_result1, phi_result2))
1741 return return_false_with_msg ("PHI results are different");
1743 size1 = gimple_phi_num_args (phi1);
1744 size2 = gimple_phi_num_args (phi2);
1746 if (size1 != size2)
1747 return return_false ();
1749 for (i = 0; i < size1; ++i)
1751 t1 = gimple_phi_arg (phi1, i)->def;
1752 t2 = gimple_phi_arg (phi2, i)->def;
1754 if (!m_checker->compare_operand (t1, t2))
1755 return return_false ();
1757 e1 = gimple_phi_arg_edge (phi1, i);
1758 e2 = gimple_phi_arg_edge (phi2, i);
1760 if (!m_checker->compare_edge (e1, e2))
1761 return return_false ();
1764 gsi_next (&si2);
1767 return true;
1770 /* Returns true if tree T can be compared as a handled component. */
1772 bool
1773 sem_function::icf_handled_component_p (tree t)
1775 tree_code tc = TREE_CODE (t);
1777 return (handled_component_p (t)
1778 || tc == ADDR_EXPR || tc == MEM_REF || tc == OBJ_TYPE_REF);
1781 /* Basic blocks dictionary BB_DICT returns true if SOURCE index BB
1782 corresponds to TARGET. */
1784 bool
1785 sem_function::bb_dict_test (vec<int> *bb_dict, int source, int target)
1787 source++;
1788 target++;
1790 if (bb_dict->length () <= (unsigned)source)
1791 bb_dict->safe_grow_cleared (source + 1);
1793 if ((*bb_dict)[source] == 0)
1795 (*bb_dict)[source] = target;
1796 return true;
1798 else
1799 return (*bb_dict)[source] == target;
1803 /* Semantic variable constructor that uses STACK as bitmap memory stack. */
1805 sem_variable::sem_variable (bitmap_obstack *stack): sem_item (VAR, stack)
1809 /* Constructor based on varpool node _NODE with computed hash _HASH.
1810 Bitmap STACK is used for memory allocation. */
1812 sem_variable::sem_variable (varpool_node *node, hashval_t _hash,
1813 bitmap_obstack *stack): sem_item(VAR,
1814 node, _hash, stack)
1816 gcc_checking_assert (node);
1817 gcc_checking_assert (get_node ());
1820 /* Fast equality function based on knowledge known in WPA. */
1822 bool
1823 sem_variable::equals_wpa (sem_item *item,
1824 hash_map <symtab_node *, sem_item *> &ignored_nodes)
1826 gcc_assert (item->type == VAR);
1828 if (node->num_references () != item->node->num_references ())
1829 return return_false_with_msg ("different number of references");
1831 if (DECL_TLS_MODEL (decl) || DECL_TLS_MODEL (item->decl))
1832 return return_false_with_msg ("TLS model");
1834 /* DECL_ALIGN is safe to merge, because we will always chose the largest
1835 alignment out of all aliases. */
1837 if (DECL_VIRTUAL_P (decl) != DECL_VIRTUAL_P (item->decl))
1838 return return_false_with_msg ("Virtual flag mismatch");
1840 if (DECL_SIZE (decl) != DECL_SIZE (item->decl)
1841 && ((!DECL_SIZE (decl) || !DECL_SIZE (item->decl))
1842 || !operand_equal_p (DECL_SIZE (decl),
1843 DECL_SIZE (item->decl), OEP_ONLY_CONST)))
1844 return return_false_with_msg ("size mismatch");
1846 /* Do not attempt to mix data from different user sections;
1847 we do not know what user intends with those. */
1848 if (((DECL_SECTION_NAME (decl) && !node->implicit_section)
1849 || (DECL_SECTION_NAME (item->decl) && !item->node->implicit_section))
1850 && DECL_SECTION_NAME (decl) != DECL_SECTION_NAME (item->decl))
1851 return return_false_with_msg ("user section mismatch");
1853 if (DECL_IN_TEXT_SECTION (decl) != DECL_IN_TEXT_SECTION (item->decl))
1854 return return_false_with_msg ("text section");
1856 ipa_ref *ref = NULL, *ref2 = NULL;
1857 for (unsigned i = 0; node->iterate_reference (i, ref); i++)
1859 item->node->iterate_reference (i, ref2);
1861 if (ref->use != ref2->use)
1862 return return_false_with_msg ("reference use mismatch");
1864 if (!compare_symbol_references (ignored_nodes,
1865 ref->referred, ref2->referred,
1866 ref->address_matters_p ()))
1867 return false;
1870 return true;
1873 /* Returns true if the item equals to ITEM given as argument. */
1875 bool
1876 sem_variable::equals (sem_item *item,
1877 hash_map <symtab_node *, sem_item *> &)
1879 gcc_assert (item->type == VAR);
1880 bool ret;
1882 if (DECL_INITIAL (decl) == error_mark_node && in_lto_p)
1883 dyn_cast <varpool_node *>(node)->get_constructor ();
1884 if (DECL_INITIAL (item->decl) == error_mark_node && in_lto_p)
1885 dyn_cast <varpool_node *>(item->node)->get_constructor ();
1887 /* As seen in PR ipa/65303 we have to compare variables types. */
1888 if (!func_checker::compatible_types_p (TREE_TYPE (decl),
1889 TREE_TYPE (item->decl)))
1890 return return_false_with_msg ("variables types are different");
1892 ret = sem_variable::equals (DECL_INITIAL (decl),
1893 DECL_INITIAL (item->node->decl));
1894 if (dump_file && (dump_flags & TDF_DETAILS))
1895 fprintf (dump_file,
1896 "Equals called for vars:%s:%s (%u:%u) (%s:%s) with result: %s\n\n",
1897 xstrdup_for_dump (node->name()),
1898 xstrdup_for_dump (item->node->name ()),
1899 node->order, item->node->order,
1900 xstrdup_for_dump (node->asm_name ()),
1901 xstrdup_for_dump (item->node->asm_name ()), ret ? "true" : "false");
1903 return ret;
1906 /* Compares trees T1 and T2 for semantic equality. */
1908 bool
1909 sem_variable::equals (tree t1, tree t2)
1911 if (!t1 || !t2)
1912 return return_with_debug (t1 == t2);
1913 if (t1 == t2)
1914 return true;
1915 tree_code tc1 = TREE_CODE (t1);
1916 tree_code tc2 = TREE_CODE (t2);
1918 if (tc1 != tc2)
1919 return return_false_with_msg ("TREE_CODE mismatch");
1921 switch (tc1)
1923 case CONSTRUCTOR:
1925 vec<constructor_elt, va_gc> *v1, *v2;
1926 unsigned HOST_WIDE_INT idx;
1928 enum tree_code typecode = TREE_CODE (TREE_TYPE (t1));
1929 if (typecode != TREE_CODE (TREE_TYPE (t2)))
1930 return return_false_with_msg ("constructor type mismatch");
1932 if (typecode == ARRAY_TYPE)
1934 HOST_WIDE_INT size_1 = int_size_in_bytes (TREE_TYPE (t1));
1935 /* For arrays, check that the sizes all match. */
1936 if (TYPE_MODE (TREE_TYPE (t1)) != TYPE_MODE (TREE_TYPE (t2))
1937 || size_1 == -1
1938 || size_1 != int_size_in_bytes (TREE_TYPE (t2)))
1939 return return_false_with_msg ("constructor array size mismatch");
1941 else if (!func_checker::compatible_types_p (TREE_TYPE (t1),
1942 TREE_TYPE (t2)))
1943 return return_false_with_msg ("constructor type incompatible");
1945 v1 = CONSTRUCTOR_ELTS (t1);
1946 v2 = CONSTRUCTOR_ELTS (t2);
1947 if (vec_safe_length (v1) != vec_safe_length (v2))
1948 return return_false_with_msg ("constructor number of elts mismatch");
1950 for (idx = 0; idx < vec_safe_length (v1); ++idx)
1952 constructor_elt *c1 = &(*v1)[idx];
1953 constructor_elt *c2 = &(*v2)[idx];
1955 /* Check that each value is the same... */
1956 if (!sem_variable::equals (c1->value, c2->value))
1957 return false;
1958 /* ... and that they apply to the same fields! */
1959 if (!sem_variable::equals (c1->index, c2->index))
1960 return false;
1962 return true;
1964 case MEM_REF:
1966 tree x1 = TREE_OPERAND (t1, 0);
1967 tree x2 = TREE_OPERAND (t2, 0);
1968 tree y1 = TREE_OPERAND (t1, 1);
1969 tree y2 = TREE_OPERAND (t2, 1);
1971 if (!func_checker::compatible_types_p (TREE_TYPE (x1), TREE_TYPE (x2)))
1972 return return_false ();
1974 /* Type of the offset on MEM_REF does not matter. */
1975 return return_with_debug (sem_variable::equals (x1, x2)
1976 && wi::to_offset (y1)
1977 == wi::to_offset (y2));
1979 case ADDR_EXPR:
1980 case FDESC_EXPR:
1982 tree op1 = TREE_OPERAND (t1, 0);
1983 tree op2 = TREE_OPERAND (t2, 0);
1984 return sem_variable::equals (op1, op2);
1986 /* References to other vars/decls are compared using ipa-ref. */
1987 case FUNCTION_DECL:
1988 case VAR_DECL:
1989 if (decl_in_symtab_p (t1) && decl_in_symtab_p (t2))
1990 return true;
1991 return return_false_with_msg ("Declaration mismatch");
1992 case CONST_DECL:
1993 /* TODO: We can check CONST_DECL by its DECL_INITIAL, but for that we
1994 need to process its VAR/FUNCTION references without relying on ipa-ref
1995 compare. */
1996 case FIELD_DECL:
1997 case LABEL_DECL:
1998 return return_false_with_msg ("Declaration mismatch");
1999 case INTEGER_CST:
2000 /* Integer constants are the same only if the same width of type. */
2001 if (TYPE_PRECISION (TREE_TYPE (t1)) != TYPE_PRECISION (TREE_TYPE (t2)))
2002 return return_false_with_msg ("INTEGER_CST precision mismatch");
2003 if (TYPE_MODE (TREE_TYPE (t1)) != TYPE_MODE (TREE_TYPE (t2)))
2004 return return_false_with_msg ("INTEGER_CST mode mismatch");
2005 return return_with_debug (tree_int_cst_equal (t1, t2));
2006 case STRING_CST:
2007 if (TYPE_MODE (TREE_TYPE (t1)) != TYPE_MODE (TREE_TYPE (t2)))
2008 return return_false_with_msg ("STRING_CST mode mismatch");
2009 if (TREE_STRING_LENGTH (t1) != TREE_STRING_LENGTH (t2))
2010 return return_false_with_msg ("STRING_CST length mismatch");
2011 if (memcmp (TREE_STRING_POINTER (t1), TREE_STRING_POINTER (t2),
2012 TREE_STRING_LENGTH (t1)))
2013 return return_false_with_msg ("STRING_CST mismatch");
2014 return true;
2015 case FIXED_CST:
2016 /* Fixed constants are the same only if the same width of type. */
2017 if (TYPE_PRECISION (TREE_TYPE (t1)) != TYPE_PRECISION (TREE_TYPE (t2)))
2018 return return_false_with_msg ("FIXED_CST precision mismatch");
2020 return return_with_debug (FIXED_VALUES_IDENTICAL (TREE_FIXED_CST (t1),
2021 TREE_FIXED_CST (t2)));
2022 case COMPLEX_CST:
2023 return (sem_variable::equals (TREE_REALPART (t1), TREE_REALPART (t2))
2024 && sem_variable::equals (TREE_IMAGPART (t1), TREE_IMAGPART (t2)));
2025 case REAL_CST:
2026 /* Real constants are the same only if the same width of type. */
2027 if (TYPE_PRECISION (TREE_TYPE (t1)) != TYPE_PRECISION (TREE_TYPE (t2)))
2028 return return_false_with_msg ("REAL_CST precision mismatch");
2029 return return_with_debug (real_identical (&TREE_REAL_CST (t1),
2030 &TREE_REAL_CST (t2)));
2031 case VECTOR_CST:
2033 unsigned i;
2035 if (VECTOR_CST_NELTS (t1) != VECTOR_CST_NELTS (t2))
2036 return return_false_with_msg ("VECTOR_CST nelts mismatch");
2038 for (i = 0; i < VECTOR_CST_NELTS (t1); ++i)
2039 if (!sem_variable::equals (VECTOR_CST_ELT (t1, i),
2040 VECTOR_CST_ELT (t2, i)))
2041 return 0;
2043 return 1;
2045 case ARRAY_REF:
2046 case ARRAY_RANGE_REF:
2048 tree x1 = TREE_OPERAND (t1, 0);
2049 tree x2 = TREE_OPERAND (t2, 0);
2050 tree y1 = TREE_OPERAND (t1, 1);
2051 tree y2 = TREE_OPERAND (t2, 1);
2053 if (!sem_variable::equals (x1, x2) || !sem_variable::equals (y1, y2))
2054 return false;
2055 if (!sem_variable::equals (array_ref_low_bound (t1),
2056 array_ref_low_bound (t2)))
2057 return false;
2058 if (!sem_variable::equals (array_ref_element_size (t1),
2059 array_ref_element_size (t2)))
2060 return false;
2061 return true;
2064 case COMPONENT_REF:
2065 case POINTER_PLUS_EXPR:
2066 case PLUS_EXPR:
2067 case MINUS_EXPR:
2068 case RANGE_EXPR:
2070 tree x1 = TREE_OPERAND (t1, 0);
2071 tree x2 = TREE_OPERAND (t2, 0);
2072 tree y1 = TREE_OPERAND (t1, 1);
2073 tree y2 = TREE_OPERAND (t2, 1);
2075 return sem_variable::equals (x1, x2) && sem_variable::equals (y1, y2);
2078 CASE_CONVERT:
2079 case VIEW_CONVERT_EXPR:
2080 if (!func_checker::compatible_types_p (TREE_TYPE (t1), TREE_TYPE (t2)))
2081 return return_false ();
2082 return sem_variable::equals (TREE_OPERAND (t1, 0), TREE_OPERAND (t2, 0));
2083 case ERROR_MARK:
2084 return return_false_with_msg ("ERROR_MARK");
2085 default:
2086 return return_false_with_msg ("Unknown TREE code reached");
2090 /* Parser function that visits a varpool NODE. */
2092 sem_variable *
2093 sem_variable::parse (varpool_node *node, bitmap_obstack *stack)
2095 if (TREE_THIS_VOLATILE (node->decl) || DECL_HARD_REGISTER (node->decl)
2096 || node->alias)
2097 return NULL;
2099 sem_variable *v = new sem_variable (node, 0, stack);
2101 v->init ();
2103 return v;
2106 /* References independent hash function. */
2108 hashval_t
2109 sem_variable::get_hash (void)
2111 if (m_hash)
2112 return m_hash;
2114 /* All WPA streamed in symbols should have their hashes computed at compile
2115 time. At this point, the constructor may not be in memory at all.
2116 DECL_INITIAL (decl) would be error_mark_node in that case. */
2117 gcc_assert (!node->lto_file_data);
2118 tree ctor = DECL_INITIAL (decl);
2119 inchash::hash hstate;
2121 hstate.add_int (456346417);
2122 if (DECL_SIZE (decl) && tree_fits_shwi_p (DECL_SIZE (decl)))
2123 hstate.add_wide_int (tree_to_shwi (DECL_SIZE (decl)));
2124 add_expr (ctor, hstate);
2125 set_hash (hstate.end ());
2127 return m_hash;
2130 /* Merges instance with an ALIAS_ITEM, where alias, thunk or redirection can
2131 be applied. */
2133 bool
2134 sem_variable::merge (sem_item *alias_item)
2136 gcc_assert (alias_item->type == VAR);
2138 if (!sem_item::target_supports_symbol_aliases_p ())
2140 if (dump_file)
2141 fprintf (dump_file, "Not unifying; "
2142 "Symbol aliases are not supported by target\n\n");
2143 return false;
2146 if (DECL_EXTERNAL (alias_item->decl))
2148 if (dump_file)
2149 fprintf (dump_file, "Not unifying; alias is external.\n\n");
2150 return false;
2153 sem_variable *alias_var = static_cast<sem_variable *> (alias_item);
2155 varpool_node *original = get_node ();
2156 varpool_node *alias = alias_var->get_node ();
2157 bool original_discardable = false;
2159 bool original_address_matters = original->address_matters_p ();
2160 bool alias_address_matters = alias->address_matters_p ();
2162 /* See if original is in a section that can be discarded if the main
2163 symbol is not used.
2164 Also consider case where we have resolution info and we know that
2165 original's definition is not going to be used. In this case we can not
2166 create alias to original. */
2167 if (original->can_be_discarded_p ()
2168 || (node->resolution != LDPR_UNKNOWN
2169 && !decl_binds_to_current_def_p (node->decl)))
2170 original_discardable = true;
2172 gcc_assert (!TREE_ASM_WRITTEN (alias->decl));
2174 /* Constant pool machinery is not quite ready for aliases.
2175 TODO: varasm code contains logic for merging DECL_IN_CONSTANT_POOL.
2176 For LTO merging does not happen that is an important missing feature.
2177 We can enable merging with LTO if the DECL_IN_CONSTANT_POOL
2178 flag is dropped and non-local symbol name is assigned. */
2179 if (DECL_IN_CONSTANT_POOL (alias->decl)
2180 || DECL_IN_CONSTANT_POOL (original->decl))
2182 if (dump_file)
2183 fprintf (dump_file,
2184 "Not unifying; constant pool variables.\n\n");
2185 return false;
2188 /* Do not attempt to mix functions from different user sections;
2189 we do not know what user intends with those. */
2190 if (((DECL_SECTION_NAME (original->decl) && !original->implicit_section)
2191 || (DECL_SECTION_NAME (alias->decl) && !alias->implicit_section))
2192 && DECL_SECTION_NAME (original->decl) != DECL_SECTION_NAME (alias->decl))
2194 if (dump_file)
2195 fprintf (dump_file,
2196 "Not unifying; "
2197 "original and alias are in different sections.\n\n");
2198 return false;
2201 /* We can not merge if address comparsion metters. */
2202 if (original_address_matters && alias_address_matters
2203 && flag_merge_constants < 2)
2205 if (dump_file)
2206 fprintf (dump_file,
2207 "Not unifying; "
2208 "adress of original and alias may be compared.\n\n");
2209 return false;
2211 if (DECL_COMDAT_GROUP (original->decl) != DECL_COMDAT_GROUP (alias->decl))
2213 if (dump_file)
2214 fprintf (dump_file, "Not unifying; alias cannot be created; "
2215 "across comdat group boundary\n\n");
2217 return false;
2220 if (original_discardable)
2222 if (dump_file)
2223 fprintf (dump_file, "Not unifying; alias cannot be created; "
2224 "target is discardable\n\n");
2226 return false;
2228 else
2230 gcc_assert (!original->alias);
2231 gcc_assert (!alias->alias);
2233 alias->analyzed = false;
2235 DECL_INITIAL (alias->decl) = NULL;
2236 ((symtab_node *)alias)->call_for_symbol_and_aliases (clear_decl_rtl,
2237 NULL, true);
2238 alias->need_bounds_init = false;
2239 alias->remove_all_references ();
2240 if (TREE_ADDRESSABLE (alias->decl))
2241 original->call_for_symbol_and_aliases (set_addressable, NULL, true);
2243 varpool_node::create_alias (alias_var->decl, decl);
2244 alias->resolve_alias (original);
2246 if (dump_file)
2247 fprintf (dump_file, "Unified; Variable alias has been created.\n\n");
2249 return true;
2253 /* Dump symbol to FILE. */
2255 void
2256 sem_variable::dump_to_file (FILE *file)
2258 gcc_assert (file);
2260 print_node (file, "", decl, 0);
2261 fprintf (file, "\n\n");
2264 unsigned int sem_item_optimizer::class_id = 0;
2266 sem_item_optimizer::sem_item_optimizer (): worklist (0), m_classes (0),
2267 m_classes_count (0), m_cgraph_node_hooks (NULL), m_varpool_node_hooks (NULL)
2269 m_items.create (0);
2270 bitmap_obstack_initialize (&m_bmstack);
2273 sem_item_optimizer::~sem_item_optimizer ()
2275 for (unsigned int i = 0; i < m_items.length (); i++)
2276 delete m_items[i];
2278 for (hash_table<congruence_class_group_hash>::iterator it = m_classes.begin ();
2279 it != m_classes.end (); ++it)
2281 for (unsigned int i = 0; i < (*it)->classes.length (); i++)
2282 delete (*it)->classes[i];
2284 (*it)->classes.release ();
2285 free (*it);
2288 m_items.release ();
2290 bitmap_obstack_release (&m_bmstack);
2293 /* Write IPA ICF summary for symbols. */
2295 void
2296 sem_item_optimizer::write_summary (void)
2298 unsigned int count = 0;
2300 output_block *ob = create_output_block (LTO_section_ipa_icf);
2301 lto_symtab_encoder_t encoder = ob->decl_state->symtab_node_encoder;
2302 ob->symbol = NULL;
2304 /* Calculate number of symbols to be serialized. */
2305 for (lto_symtab_encoder_iterator lsei = lsei_start_in_partition (encoder);
2306 !lsei_end_p (lsei);
2307 lsei_next_in_partition (&lsei))
2309 symtab_node *node = lsei_node (lsei);
2311 if (m_symtab_node_map.get (node))
2312 count++;
2315 streamer_write_uhwi (ob, count);
2317 /* Process all of the symbols. */
2318 for (lto_symtab_encoder_iterator lsei = lsei_start_in_partition (encoder);
2319 !lsei_end_p (lsei);
2320 lsei_next_in_partition (&lsei))
2322 symtab_node *node = lsei_node (lsei);
2324 sem_item **item = m_symtab_node_map.get (node);
2326 if (item && *item)
2328 int node_ref = lto_symtab_encoder_encode (encoder, node);
2329 streamer_write_uhwi_stream (ob->main_stream, node_ref);
2331 streamer_write_uhwi (ob, (*item)->get_hash ());
2335 streamer_write_char_stream (ob->main_stream, 0);
2336 produce_asm (ob, NULL);
2337 destroy_output_block (ob);
2340 /* Reads a section from LTO stream file FILE_DATA. Input block for DATA
2341 contains LEN bytes. */
2343 void
2344 sem_item_optimizer::read_section (lto_file_decl_data *file_data,
2345 const char *data, size_t len)
2347 const lto_function_header *header =
2348 (const lto_function_header *) data;
2349 const int cfg_offset = sizeof (lto_function_header);
2350 const int main_offset = cfg_offset + header->cfg_size;
2351 const int string_offset = main_offset + header->main_size;
2352 data_in *data_in;
2353 unsigned int i;
2354 unsigned int count;
2356 lto_input_block ib_main ((const char *) data + main_offset, 0,
2357 header->main_size, file_data->mode_table);
2359 data_in =
2360 lto_data_in_create (file_data, (const char *) data + string_offset,
2361 header->string_size, vNULL);
2363 count = streamer_read_uhwi (&ib_main);
2365 for (i = 0; i < count; i++)
2367 unsigned int index;
2368 symtab_node *node;
2369 lto_symtab_encoder_t encoder;
2371 index = streamer_read_uhwi (&ib_main);
2372 encoder = file_data->symtab_node_encoder;
2373 node = lto_symtab_encoder_deref (encoder, index);
2375 hashval_t hash = streamer_read_uhwi (&ib_main);
2377 gcc_assert (node->definition);
2379 if (dump_file)
2380 fprintf (dump_file, "Symbol added:%s (tree: %p, uid:%u)\n",
2381 node->asm_name (), (void *) node->decl, node->order);
2383 if (is_a<cgraph_node *> (node))
2385 cgraph_node *cnode = dyn_cast <cgraph_node *> (node);
2387 m_items.safe_push (new sem_function (cnode, hash, &m_bmstack));
2389 else
2391 varpool_node *vnode = dyn_cast <varpool_node *> (node);
2393 m_items.safe_push (new sem_variable (vnode, hash, &m_bmstack));
2397 lto_free_section_data (file_data, LTO_section_ipa_icf, NULL, data,
2398 len);
2399 lto_data_in_delete (data_in);
2402 /* Read IPA ICF summary for symbols. */
2404 void
2405 sem_item_optimizer::read_summary (void)
2407 lto_file_decl_data **file_data_vec = lto_get_file_decl_data ();
2408 lto_file_decl_data *file_data;
2409 unsigned int j = 0;
2411 while ((file_data = file_data_vec[j++]))
2413 size_t len;
2414 const char *data = lto_get_section_data (file_data,
2415 LTO_section_ipa_icf, NULL, &len);
2417 if (data)
2418 read_section (file_data, data, len);
2422 /* Register callgraph and varpool hooks. */
2424 void
2425 sem_item_optimizer::register_hooks (void)
2427 if (!m_cgraph_node_hooks)
2428 m_cgraph_node_hooks = symtab->add_cgraph_removal_hook
2429 (&sem_item_optimizer::cgraph_removal_hook, this);
2431 if (!m_varpool_node_hooks)
2432 m_varpool_node_hooks = symtab->add_varpool_removal_hook
2433 (&sem_item_optimizer::varpool_removal_hook, this);
2436 /* Unregister callgraph and varpool hooks. */
2438 void
2439 sem_item_optimizer::unregister_hooks (void)
2441 if (m_cgraph_node_hooks)
2442 symtab->remove_cgraph_removal_hook (m_cgraph_node_hooks);
2444 if (m_varpool_node_hooks)
2445 symtab->remove_varpool_removal_hook (m_varpool_node_hooks);
2448 /* Adds a CLS to hashtable associated by hash value. */
2450 void
2451 sem_item_optimizer::add_class (congruence_class *cls)
2453 gcc_assert (cls->members.length ());
2455 congruence_class_group *group = get_group_by_hash (
2456 cls->members[0]->get_hash (),
2457 cls->members[0]->type);
2458 group->classes.safe_push (cls);
2461 /* Gets a congruence class group based on given HASH value and TYPE. */
2463 congruence_class_group *
2464 sem_item_optimizer::get_group_by_hash (hashval_t hash, sem_item_type type)
2466 congruence_class_group *item = XNEW (congruence_class_group);
2467 item->hash = hash;
2468 item->type = type;
2470 congruence_class_group **slot = m_classes.find_slot (item, INSERT);
2472 if (*slot)
2473 free (item);
2474 else
2476 item->classes.create (1);
2477 *slot = item;
2480 return *slot;
2483 /* Callgraph removal hook called for a NODE with a custom DATA. */
2485 void
2486 sem_item_optimizer::cgraph_removal_hook (cgraph_node *node, void *data)
2488 sem_item_optimizer *optimizer = (sem_item_optimizer *) data;
2489 optimizer->remove_symtab_node (node);
2492 /* Varpool removal hook called for a NODE with a custom DATA. */
2494 void
2495 sem_item_optimizer::varpool_removal_hook (varpool_node *node, void *data)
2497 sem_item_optimizer *optimizer = (sem_item_optimizer *) data;
2498 optimizer->remove_symtab_node (node);
2501 /* Remove symtab NODE triggered by symtab removal hooks. */
2503 void
2504 sem_item_optimizer::remove_symtab_node (symtab_node *node)
2506 gcc_assert (!m_classes.elements());
2508 m_removed_items_set.add (node);
2511 void
2512 sem_item_optimizer::remove_item (sem_item *item)
2514 if (m_symtab_node_map.get (item->node))
2515 m_symtab_node_map.remove (item->node);
2516 delete item;
2519 /* Removes all callgraph and varpool nodes that are marked by symtab
2520 as deleted. */
2522 void
2523 sem_item_optimizer::filter_removed_items (void)
2525 auto_vec <sem_item *> filtered;
2527 for (unsigned int i = 0; i < m_items.length(); i++)
2529 sem_item *item = m_items[i];
2531 if (m_removed_items_set.contains (item->node))
2533 remove_item (item);
2534 continue;
2537 if (item->type == FUNC)
2539 cgraph_node *cnode = static_cast <sem_function *>(item)->get_node ();
2541 if (in_lto_p && (cnode->alias || cnode->body_removed))
2542 remove_item (item);
2543 else
2544 filtered.safe_push (item);
2546 else /* VAR. */
2548 if (!flag_ipa_icf_variables)
2549 remove_item (item);
2550 else
2552 /* Filter out non-readonly variables. */
2553 tree decl = item->decl;
2554 if (TREE_READONLY (decl))
2555 filtered.safe_push (item);
2556 else
2557 remove_item (item);
2562 /* Clean-up of released semantic items. */
2564 m_items.release ();
2565 for (unsigned int i = 0; i < filtered.length(); i++)
2566 m_items.safe_push (filtered[i]);
2569 /* Optimizer entry point which returns true in case it processes
2570 a merge operation. True is returned if there's a merge operation
2571 processed. */
2573 bool
2574 sem_item_optimizer::execute (void)
2576 filter_removed_items ();
2577 unregister_hooks ();
2579 build_graph ();
2580 update_hash_by_addr_refs ();
2581 build_hash_based_classes ();
2583 if (dump_file)
2584 fprintf (dump_file, "Dump after hash based groups\n");
2585 dump_cong_classes ();
2587 for (unsigned int i = 0; i < m_items.length(); i++)
2588 m_items[i]->init_wpa ();
2590 subdivide_classes_by_equality (true);
2592 if (dump_file)
2593 fprintf (dump_file, "Dump after WPA based types groups\n");
2595 dump_cong_classes ();
2597 process_cong_reduction ();
2598 checking_verify_classes ();
2600 if (dump_file)
2601 fprintf (dump_file, "Dump after callgraph-based congruence reduction\n");
2603 dump_cong_classes ();
2605 parse_nonsingleton_classes ();
2606 subdivide_classes_by_equality ();
2608 if (dump_file)
2609 fprintf (dump_file, "Dump after full equality comparison of groups\n");
2611 dump_cong_classes ();
2613 unsigned int prev_class_count = m_classes_count;
2615 process_cong_reduction ();
2616 dump_cong_classes ();
2617 checking_verify_classes ();
2618 bool merged_p = merge_classes (prev_class_count);
2620 if (dump_file && (dump_flags & TDF_DETAILS))
2621 symtab_node::dump_table (dump_file);
2623 return merged_p;
2626 /* Function responsible for visiting all potential functions and
2627 read-only variables that can be merged. */
2629 void
2630 sem_item_optimizer::parse_funcs_and_vars (void)
2632 cgraph_node *cnode;
2634 if (flag_ipa_icf_functions)
2635 FOR_EACH_DEFINED_FUNCTION (cnode)
2637 sem_function *f = sem_function::parse (cnode, &m_bmstack);
2638 if (f)
2640 m_items.safe_push (f);
2641 m_symtab_node_map.put (cnode, f);
2643 if (dump_file)
2644 fprintf (dump_file, "Parsed function:%s\n", f->node->asm_name ());
2646 if (dump_file && (dump_flags & TDF_DETAILS))
2647 f->dump_to_file (dump_file);
2649 else if (dump_file)
2650 fprintf (dump_file, "Not parsed function:%s\n", cnode->asm_name ());
2653 varpool_node *vnode;
2655 if (flag_ipa_icf_variables)
2656 FOR_EACH_DEFINED_VARIABLE (vnode)
2658 sem_variable *v = sem_variable::parse (vnode, &m_bmstack);
2660 if (v)
2662 m_items.safe_push (v);
2663 m_symtab_node_map.put (vnode, v);
2668 /* Makes pairing between a congruence class CLS and semantic ITEM. */
2670 void
2671 sem_item_optimizer::add_item_to_class (congruence_class *cls, sem_item *item)
2673 item->index_in_class = cls->members.length ();
2674 cls->members.safe_push (item);
2675 item->cls = cls;
2678 /* For each semantic item, append hash values of references. */
2680 void
2681 sem_item_optimizer::update_hash_by_addr_refs ()
2683 /* First, append to hash sensitive references and class type if it need to
2684 be matched for ODR. */
2685 for (unsigned i = 0; i < m_items.length (); i++)
2687 m_items[i]->update_hash_by_addr_refs (m_symtab_node_map);
2688 if (m_items[i]->type == FUNC)
2690 if (TREE_CODE (TREE_TYPE (m_items[i]->decl)) == METHOD_TYPE
2691 && contains_polymorphic_type_p
2692 (TYPE_METHOD_BASETYPE (TREE_TYPE (m_items[i]->decl)))
2693 && (DECL_CXX_CONSTRUCTOR_P (m_items[i]->decl)
2694 || (static_cast<sem_function *> (m_items[i])->param_used_p (0)
2695 && static_cast<sem_function *> (m_items[i])
2696 ->compare_polymorphic_p ())))
2698 tree class_type
2699 = TYPE_METHOD_BASETYPE (TREE_TYPE (m_items[i]->decl));
2700 inchash::hash hstate (m_items[i]->get_hash ());
2702 if (TYPE_NAME (class_type)
2703 && DECL_ASSEMBLER_NAME_SET_P (TYPE_NAME (class_type)))
2704 hstate.add_wide_int
2705 (IDENTIFIER_HASH_VALUE
2706 (DECL_ASSEMBLER_NAME (TYPE_NAME (class_type))));
2708 m_items[i]->set_hash (hstate.end ());
2713 /* Once all symbols have enhanced hash value, we can append
2714 hash values of symbols that are seen by IPA ICF and are
2715 references by a semantic item. Newly computed values
2716 are saved to global_hash member variable. */
2717 for (unsigned i = 0; i < m_items.length (); i++)
2718 m_items[i]->update_hash_by_local_refs (m_symtab_node_map);
2720 /* Global hash value replace current hash values. */
2721 for (unsigned i = 0; i < m_items.length (); i++)
2722 m_items[i]->set_hash (m_items[i]->global_hash);
2725 /* Congruence classes are built by hash value. */
2727 void
2728 sem_item_optimizer::build_hash_based_classes (void)
2730 for (unsigned i = 0; i < m_items.length (); i++)
2732 sem_item *item = m_items[i];
2734 congruence_class_group *group = get_group_by_hash (item->get_hash (),
2735 item->type);
2737 if (!group->classes.length ())
2739 m_classes_count++;
2740 group->classes.safe_push (new congruence_class (class_id++));
2743 add_item_to_class (group->classes[0], item);
2747 /* Build references according to call graph. */
2749 void
2750 sem_item_optimizer::build_graph (void)
2752 for (unsigned i = 0; i < m_items.length (); i++)
2754 sem_item *item = m_items[i];
2755 m_symtab_node_map.put (item->node, item);
2757 /* Initialize hash values if we are not in LTO mode. */
2758 if (!in_lto_p)
2759 item->get_hash ();
2762 for (unsigned i = 0; i < m_items.length (); i++)
2764 sem_item *item = m_items[i];
2766 if (item->type == FUNC)
2768 cgraph_node *cnode = dyn_cast <cgraph_node *> (item->node);
2770 cgraph_edge *e = cnode->callees;
2771 while (e)
2773 sem_item **slot = m_symtab_node_map.get
2774 (e->callee->ultimate_alias_target ());
2775 if (slot)
2776 item->add_reference (*slot);
2778 e = e->next_callee;
2782 ipa_ref *ref = NULL;
2783 for (unsigned i = 0; item->node->iterate_reference (i, ref); i++)
2785 sem_item **slot = m_symtab_node_map.get
2786 (ref->referred->ultimate_alias_target ());
2787 if (slot)
2788 item->add_reference (*slot);
2793 /* Semantic items in classes having more than one element and initialized.
2794 In case of WPA, we load function body. */
2796 void
2797 sem_item_optimizer::parse_nonsingleton_classes (void)
2799 unsigned int init_called_count = 0;
2801 for (unsigned i = 0; i < m_items.length (); i++)
2802 if (m_items[i]->cls->members.length () > 1)
2804 m_items[i]->init ();
2805 init_called_count++;
2808 if (dump_file)
2809 fprintf (dump_file, "Init called for %u items (%.2f%%).\n", init_called_count,
2810 m_items.length () ? 100.0f * init_called_count / m_items.length (): 0.0f);
2813 /* Equality function for semantic items is used to subdivide existing
2814 classes. If IN_WPA, fast equality function is invoked. */
2816 void
2817 sem_item_optimizer::subdivide_classes_by_equality (bool in_wpa)
2819 for (hash_table <congruence_class_group_hash>::iterator it = m_classes.begin ();
2820 it != m_classes.end (); ++it)
2822 unsigned int class_count = (*it)->classes.length ();
2824 for (unsigned i = 0; i < class_count; i++)
2826 congruence_class *c = (*it)->classes [i];
2828 if (c->members.length() > 1)
2830 auto_vec <sem_item *> new_vector;
2832 sem_item *first = c->members[0];
2833 new_vector.safe_push (first);
2835 unsigned class_split_first = (*it)->classes.length ();
2837 for (unsigned j = 1; j < c->members.length (); j++)
2839 sem_item *item = c->members[j];
2841 bool equals = in_wpa ? first->equals_wpa (item,
2842 m_symtab_node_map) : first->equals (item, m_symtab_node_map);
2844 if (equals)
2845 new_vector.safe_push (item);
2846 else
2848 bool integrated = false;
2850 for (unsigned k = class_split_first; k < (*it)->classes.length (); k++)
2852 sem_item *x = (*it)->classes[k]->members[0];
2853 bool equals = in_wpa ? x->equals_wpa (item,
2854 m_symtab_node_map) : x->equals (item, m_symtab_node_map);
2856 if (equals)
2858 integrated = true;
2859 add_item_to_class ((*it)->classes[k], item);
2861 break;
2865 if (!integrated)
2867 congruence_class *c = new congruence_class (class_id++);
2868 m_classes_count++;
2869 add_item_to_class (c, item);
2871 (*it)->classes.safe_push (c);
2876 // we replace newly created new_vector for the class we've just splitted
2877 c->members.release ();
2878 c->members.create (new_vector.length ());
2880 for (unsigned int j = 0; j < new_vector.length (); j++)
2881 add_item_to_class (c, new_vector[j]);
2886 checking_verify_classes ();
2889 /* Subdivide classes by address references that members of the class
2890 reference. Example can be a pair of functions that have an address
2891 taken from a function. If these addresses are different the class
2892 is split. */
2894 unsigned
2895 sem_item_optimizer::subdivide_classes_by_sensitive_refs ()
2897 typedef hash_map <symbol_compare_hash, vec <sem_item *> > subdivide_hash_map;
2899 unsigned newly_created_classes = 0;
2901 for (hash_table <congruence_class_group_hash>::iterator it = m_classes.begin ();
2902 it != m_classes.end (); ++it)
2904 unsigned int class_count = (*it)->classes.length ();
2905 auto_vec<congruence_class *> new_classes;
2907 for (unsigned i = 0; i < class_count; i++)
2909 congruence_class *c = (*it)->classes [i];
2911 if (c->members.length() > 1)
2913 subdivide_hash_map split_map;
2915 for (unsigned j = 0; j < c->members.length (); j++)
2917 sem_item *source_node = c->members[j];
2919 symbol_compare_collection *collection = new symbol_compare_collection (source_node->node);
2921 bool existed;
2922 vec <sem_item *> *slot = &split_map.get_or_insert (collection,
2923 &existed);
2924 gcc_checking_assert (slot);
2926 slot->safe_push (source_node);
2928 if (existed)
2929 delete collection;
2932 /* If the map contains more than one key, we have to split the map
2933 appropriately. */
2934 if (split_map.elements () != 1)
2936 bool first_class = true;
2938 for (subdivide_hash_map::iterator it2 = split_map.begin ();
2939 it2 != split_map.end (); ++it2)
2941 congruence_class *new_cls;
2942 new_cls = new congruence_class (class_id++);
2944 for (unsigned k = 0; k < (*it2).second.length (); k++)
2945 add_item_to_class (new_cls, (*it2).second[k]);
2947 worklist_push (new_cls);
2948 newly_created_classes++;
2950 if (first_class)
2952 (*it)->classes[i] = new_cls;
2953 first_class = false;
2955 else
2957 new_classes.safe_push (new_cls);
2958 m_classes_count++;
2963 /* Release memory. */
2964 for (subdivide_hash_map::iterator it2 = split_map.begin ();
2965 it2 != split_map.end (); ++it2)
2967 delete (*it2).first;
2968 (*it2).second.release ();
2973 for (unsigned i = 0; i < new_classes.length (); i++)
2974 (*it)->classes.safe_push (new_classes[i]);
2977 return newly_created_classes;
2980 /* Verify congruence classes, if checking is enabled. */
2982 void
2983 sem_item_optimizer::checking_verify_classes (void)
2985 if (flag_checking)
2986 verify_classes ();
2989 /* Verify congruence classes. */
2991 void
2992 sem_item_optimizer::verify_classes (void)
2994 for (hash_table <congruence_class_group_hash>::iterator it = m_classes.begin ();
2995 it != m_classes.end (); ++it)
2997 for (unsigned int i = 0; i < (*it)->classes.length (); i++)
2999 congruence_class *cls = (*it)->classes[i];
3001 gcc_assert (cls);
3002 gcc_assert (cls->members.length () > 0);
3004 for (unsigned int j = 0; j < cls->members.length (); j++)
3006 sem_item *item = cls->members[j];
3008 gcc_assert (item);
3009 gcc_assert (item->cls == cls);
3011 for (unsigned k = 0; k < item->usages.length (); k++)
3013 sem_usage_pair *usage = item->usages[k];
3014 gcc_assert (usage->item->index_in_class <
3015 usage->item->cls->members.length ());
3022 /* Disposes split map traverse function. CLS_PTR is pointer to congruence
3023 class, BSLOT is bitmap slot we want to release. DATA is mandatory,
3024 but unused argument. */
3026 bool
3027 sem_item_optimizer::release_split_map (congruence_class * const &,
3028 bitmap const &b, traverse_split_pair *)
3030 bitmap bmp = b;
3032 BITMAP_FREE (bmp);
3034 return true;
3037 /* Process split operation for a class given as pointer CLS_PTR,
3038 where bitmap B splits congruence class members. DATA is used
3039 as argument of split pair. */
3041 bool
3042 sem_item_optimizer::traverse_congruence_split (congruence_class * const &cls,
3043 bitmap const &b, traverse_split_pair *pair)
3045 sem_item_optimizer *optimizer = pair->optimizer;
3046 const congruence_class *splitter_cls = pair->cls;
3048 /* If counted bits are greater than zero and less than the number of members
3049 a group will be splitted. */
3050 unsigned popcount = bitmap_count_bits (b);
3052 if (popcount > 0 && popcount < cls->members.length ())
3054 auto_vec <congruence_class *, 2> newclasses;
3055 newclasses.quick_push (new congruence_class (class_id++));
3056 newclasses.quick_push (new congruence_class (class_id++));
3058 for (unsigned int i = 0; i < cls->members.length (); i++)
3060 int target = bitmap_bit_p (b, i);
3061 congruence_class *tc = newclasses[target];
3063 add_item_to_class (tc, cls->members[i]);
3066 if (flag_checking)
3068 for (unsigned int i = 0; i < 2; i++)
3069 gcc_assert (newclasses[i]->members.length ());
3072 if (splitter_cls == cls)
3073 optimizer->splitter_class_removed = true;
3075 /* Remove old class from worklist if presented. */
3076 bool in_worklist = cls->in_worklist;
3078 if (in_worklist)
3079 cls->in_worklist = false;
3081 congruence_class_group g;
3082 g.hash = cls->members[0]->get_hash ();
3083 g.type = cls->members[0]->type;
3085 congruence_class_group *slot = optimizer->m_classes.find(&g);
3087 for (unsigned int i = 0; i < slot->classes.length (); i++)
3088 if (slot->classes[i] == cls)
3090 slot->classes.ordered_remove (i);
3091 break;
3094 /* New class will be inserted and integrated to work list. */
3095 for (unsigned int i = 0; i < 2; i++)
3096 optimizer->add_class (newclasses[i]);
3098 /* Two classes replace one, so that increment just by one. */
3099 optimizer->m_classes_count++;
3101 /* If OLD class was presented in the worklist, we remove the class
3102 and replace it will both newly created classes. */
3103 if (in_worklist)
3104 for (unsigned int i = 0; i < 2; i++)
3105 optimizer->worklist_push (newclasses[i]);
3106 else /* Just smaller class is inserted. */
3108 unsigned int smaller_index = newclasses[0]->members.length () <
3109 newclasses[1]->members.length () ?
3110 0 : 1;
3111 optimizer->worklist_push (newclasses[smaller_index]);
3114 if (dump_file && (dump_flags & TDF_DETAILS))
3116 fprintf (dump_file, " congruence class splitted:\n");
3117 cls->dump (dump_file, 4);
3119 fprintf (dump_file, " newly created groups:\n");
3120 for (unsigned int i = 0; i < 2; i++)
3121 newclasses[i]->dump (dump_file, 4);
3124 /* Release class if not presented in work list. */
3125 if (!in_worklist)
3126 delete cls;
3130 return true;
3133 /* Tests if a class CLS used as INDEXth splits any congruence classes.
3134 Bitmap stack BMSTACK is used for bitmap allocation. */
3136 void
3137 sem_item_optimizer::do_congruence_step_for_index (congruence_class *cls,
3138 unsigned int index)
3140 hash_map <congruence_class *, bitmap> split_map;
3142 for (unsigned int i = 0; i < cls->members.length (); i++)
3144 sem_item *item = cls->members[i];
3146 /* Iterate all usages that have INDEX as usage of the item. */
3147 for (unsigned int j = 0; j < item->usages.length (); j++)
3149 sem_usage_pair *usage = item->usages[j];
3151 if (usage->index != index)
3152 continue;
3154 bitmap *slot = split_map.get (usage->item->cls);
3155 bitmap b;
3157 if(!slot)
3159 b = BITMAP_ALLOC (&m_bmstack);
3160 split_map.put (usage->item->cls, b);
3162 else
3163 b = *slot;
3165 gcc_checking_assert (usage->item->cls);
3166 gcc_checking_assert (usage->item->index_in_class <
3167 usage->item->cls->members.length ());
3169 bitmap_set_bit (b, usage->item->index_in_class);
3173 traverse_split_pair pair;
3174 pair.optimizer = this;
3175 pair.cls = cls;
3177 splitter_class_removed = false;
3178 split_map.traverse
3179 <traverse_split_pair *, sem_item_optimizer::traverse_congruence_split> (&pair);
3181 /* Bitmap clean-up. */
3182 split_map.traverse
3183 <traverse_split_pair *, sem_item_optimizer::release_split_map> (NULL);
3186 /* Every usage of a congruence class CLS is a candidate that can split the
3187 collection of classes. Bitmap stack BMSTACK is used for bitmap
3188 allocation. */
3190 void
3191 sem_item_optimizer::do_congruence_step (congruence_class *cls)
3193 bitmap_iterator bi;
3194 unsigned int i;
3196 bitmap usage = BITMAP_ALLOC (&m_bmstack);
3198 for (unsigned int i = 0; i < cls->members.length (); i++)
3199 bitmap_ior_into (usage, cls->members[i]->usage_index_bitmap);
3201 EXECUTE_IF_SET_IN_BITMAP (usage, 0, i, bi)
3203 if (dump_file && (dump_flags & TDF_DETAILS))
3204 fprintf (dump_file, " processing congruence step for class: %u, index: %u\n",
3205 cls->id, i);
3207 do_congruence_step_for_index (cls, i);
3209 if (splitter_class_removed)
3210 break;
3213 BITMAP_FREE (usage);
3216 /* Adds a newly created congruence class CLS to worklist. */
3218 void
3219 sem_item_optimizer::worklist_push (congruence_class *cls)
3221 /* Return if the class CLS is already presented in work list. */
3222 if (cls->in_worklist)
3223 return;
3225 cls->in_worklist = true;
3226 worklist.push_back (cls);
3229 /* Pops a class from worklist. */
3231 congruence_class *
3232 sem_item_optimizer::worklist_pop (void)
3234 congruence_class *cls;
3236 while (!worklist.empty ())
3238 cls = worklist.front ();
3239 worklist.pop_front ();
3240 if (cls->in_worklist)
3242 cls->in_worklist = false;
3244 return cls;
3246 else
3248 /* Work list item was already intended to be removed.
3249 The only reason for doing it is to split a class.
3250 Thus, the class CLS is deleted. */
3251 delete cls;
3255 return NULL;
3258 /* Iterative congruence reduction function. */
3260 void
3261 sem_item_optimizer::process_cong_reduction (void)
3263 for (hash_table<congruence_class_group_hash>::iterator it = m_classes.begin ();
3264 it != m_classes.end (); ++it)
3265 for (unsigned i = 0; i < (*it)->classes.length (); i++)
3266 if ((*it)->classes[i]->is_class_used ())
3267 worklist_push ((*it)->classes[i]);
3269 if (dump_file)
3270 fprintf (dump_file, "Worklist has been filled with: %lu\n",
3271 (unsigned long) worklist.size ());
3273 if (dump_file && (dump_flags & TDF_DETAILS))
3274 fprintf (dump_file, "Congruence class reduction\n");
3276 congruence_class *cls;
3278 /* Process complete congruence reduction. */
3279 while ((cls = worklist_pop ()) != NULL)
3280 do_congruence_step (cls);
3282 /* Subdivide newly created classes according to references. */
3283 unsigned new_classes = subdivide_classes_by_sensitive_refs ();
3285 if (dump_file)
3286 fprintf (dump_file, "Address reference subdivision created: %u "
3287 "new classes.\n", new_classes);
3290 /* Debug function prints all informations about congruence classes. */
3292 void
3293 sem_item_optimizer::dump_cong_classes (void)
3295 if (!dump_file)
3296 return;
3298 fprintf (dump_file,
3299 "Congruence classes: %u (unique hash values: %lu), with total: %u items\n",
3300 m_classes_count, (unsigned long) m_classes.elements(), m_items.length ());
3302 /* Histogram calculation. */
3303 unsigned int max_index = 0;
3304 unsigned int* histogram = XCNEWVEC (unsigned int, m_items.length () + 1);
3306 for (hash_table<congruence_class_group_hash>::iterator it = m_classes.begin ();
3307 it != m_classes.end (); ++it)
3309 for (unsigned i = 0; i < (*it)->classes.length (); i++)
3311 unsigned int c = (*it)->classes[i]->members.length ();
3312 histogram[c]++;
3314 if (c > max_index)
3315 max_index = c;
3318 fprintf (dump_file,
3319 "Class size histogram [num of members]: number of classe number of classess\n");
3321 for (unsigned int i = 0; i <= max_index; i++)
3322 if (histogram[i])
3323 fprintf (dump_file, "[%u]: %u classes\n", i, histogram[i]);
3325 fprintf (dump_file, "\n\n");
3328 if (dump_flags & TDF_DETAILS)
3329 for (hash_table<congruence_class_group_hash>::iterator it = m_classes.begin ();
3330 it != m_classes.end (); ++it)
3332 fprintf (dump_file, " group: with %u classes:\n", (*it)->classes.length ());
3334 for (unsigned i = 0; i < (*it)->classes.length (); i++)
3336 (*it)->classes[i]->dump (dump_file, 4);
3338 if(i < (*it)->classes.length () - 1)
3339 fprintf (dump_file, " ");
3343 free (histogram);
3346 /* After reduction is done, we can declare all items in a group
3347 to be equal. PREV_CLASS_COUNT is start number of classes
3348 before reduction. True is returned if there's a merge operation
3349 processed. */
3351 bool
3352 sem_item_optimizer::merge_classes (unsigned int prev_class_count)
3354 unsigned int item_count = m_items.length ();
3355 unsigned int class_count = m_classes_count;
3356 unsigned int equal_items = item_count - class_count;
3358 unsigned int non_singular_classes_count = 0;
3359 unsigned int non_singular_classes_sum = 0;
3361 bool merged_p = false;
3363 for (hash_table<congruence_class_group_hash>::iterator it = m_classes.begin ();
3364 it != m_classes.end (); ++it)
3365 for (unsigned int i = 0; i < (*it)->classes.length (); i++)
3367 congruence_class *c = (*it)->classes[i];
3368 if (c->members.length () > 1)
3370 non_singular_classes_count++;
3371 non_singular_classes_sum += c->members.length ();
3375 if (dump_file)
3377 fprintf (dump_file, "\nItem count: %u\n", item_count);
3378 fprintf (dump_file, "Congruent classes before: %u, after: %u\n",
3379 prev_class_count, class_count);
3380 fprintf (dump_file, "Average class size before: %.2f, after: %.2f\n",
3381 prev_class_count ? 1.0f * item_count / prev_class_count : 0.0f,
3382 class_count ? 1.0f * item_count / class_count : 0.0f);
3383 fprintf (dump_file, "Average non-singular class size: %.2f, count: %u\n",
3384 non_singular_classes_count ? 1.0f * non_singular_classes_sum /
3385 non_singular_classes_count : 0.0f,
3386 non_singular_classes_count);
3387 fprintf (dump_file, "Equal symbols: %u\n", equal_items);
3388 fprintf (dump_file, "Fraction of visited symbols: %.2f%%\n\n",
3389 item_count ? 100.0f * equal_items / item_count : 0.0f);
3392 for (hash_table<congruence_class_group_hash>::iterator it = m_classes.begin ();
3393 it != m_classes.end (); ++it)
3394 for (unsigned int i = 0; i < (*it)->classes.length (); i++)
3396 congruence_class *c = (*it)->classes[i];
3398 if (c->members.length () == 1)
3399 continue;
3401 gcc_assert (c->members.length ());
3403 sem_item *source = c->members[0];
3405 for (unsigned int j = 1; j < c->members.length (); j++)
3407 sem_item *alias = c->members[j];
3409 if (dump_file)
3411 fprintf (dump_file, "Semantic equality hit:%s->%s\n",
3412 xstrdup_for_dump (source->node->name ()),
3413 xstrdup_for_dump (alias->node->name ()));
3414 fprintf (dump_file, "Assembler symbol names:%s->%s\n",
3415 xstrdup_for_dump (source->node->asm_name ()),
3416 xstrdup_for_dump (alias->node->asm_name ()));
3419 if (lookup_attribute ("no_icf", DECL_ATTRIBUTES (alias->decl)))
3421 if (dump_file)
3422 fprintf (dump_file,
3423 "Merge operation is skipped due to no_icf "
3424 "attribute.\n\n");
3426 continue;
3429 if (dump_file && (dump_flags & TDF_DETAILS))
3431 source->dump_to_file (dump_file);
3432 alias->dump_to_file (dump_file);
3435 if (dbg_cnt (merged_ipa_icf))
3436 merged_p |= source->merge (alias);
3440 return merged_p;
3443 /* Dump function prints all class members to a FILE with an INDENT. */
3445 void
3446 congruence_class::dump (FILE *file, unsigned int indent) const
3448 FPRINTF_SPACES (file, indent, "class with id: %u, hash: %u, items: %u\n",
3449 id, members[0]->get_hash (), members.length ());
3451 FPUTS_SPACES (file, indent + 2, "");
3452 for (unsigned i = 0; i < members.length (); i++)
3453 fprintf (file, "%s(%p/%u) ", members[i]->node->asm_name (),
3454 (void *) members[i]->decl,
3455 members[i]->node->order);
3457 fprintf (file, "\n");
3460 /* Returns true if there's a member that is used from another group. */
3462 bool
3463 congruence_class::is_class_used (void)
3465 for (unsigned int i = 0; i < members.length (); i++)
3466 if (members[i]->usages.length ())
3467 return true;
3469 return false;
3472 /* Generate pass summary for IPA ICF pass. */
3474 static void
3475 ipa_icf_generate_summary (void)
3477 if (!optimizer)
3478 optimizer = new sem_item_optimizer ();
3480 optimizer->register_hooks ();
3481 optimizer->parse_funcs_and_vars ();
3484 /* Write pass summary for IPA ICF pass. */
3486 static void
3487 ipa_icf_write_summary (void)
3489 gcc_assert (optimizer);
3491 optimizer->write_summary ();
3494 /* Read pass summary for IPA ICF pass. */
3496 static void
3497 ipa_icf_read_summary (void)
3499 if (!optimizer)
3500 optimizer = new sem_item_optimizer ();
3502 optimizer->read_summary ();
3503 optimizer->register_hooks ();
3506 /* Semantic equality exection function. */
3508 static unsigned int
3509 ipa_icf_driver (void)
3511 gcc_assert (optimizer);
3513 bool merged_p = optimizer->execute ();
3515 delete optimizer;
3516 optimizer = NULL;
3518 return merged_p ? TODO_remove_functions : 0;
3521 const pass_data pass_data_ipa_icf =
3523 IPA_PASS, /* type */
3524 "icf", /* name */
3525 OPTGROUP_IPA, /* optinfo_flags */
3526 TV_IPA_ICF, /* tv_id */
3527 0, /* properties_required */
3528 0, /* properties_provided */
3529 0, /* properties_destroyed */
3530 0, /* todo_flags_start */
3531 0, /* todo_flags_finish */
3534 class pass_ipa_icf : public ipa_opt_pass_d
3536 public:
3537 pass_ipa_icf (gcc::context *ctxt)
3538 : ipa_opt_pass_d (pass_data_ipa_icf, ctxt,
3539 ipa_icf_generate_summary, /* generate_summary */
3540 ipa_icf_write_summary, /* write_summary */
3541 ipa_icf_read_summary, /* read_summary */
3542 NULL, /*
3543 write_optimization_summary */
3544 NULL, /*
3545 read_optimization_summary */
3546 NULL, /* stmt_fixup */
3547 0, /* function_transform_todo_flags_start */
3548 NULL, /* function_transform */
3549 NULL) /* variable_transform */
3552 /* opt_pass methods: */
3553 virtual bool gate (function *)
3555 return in_lto_p || flag_ipa_icf_variables || flag_ipa_icf_functions;
3558 virtual unsigned int execute (function *)
3560 return ipa_icf_driver();
3562 }; // class pass_ipa_icf
3564 } // ipa_icf namespace
3566 ipa_opt_pass_d *
3567 make_pass_ipa_icf (gcc::context *ctxt)
3569 return new ipa_icf::pass_ipa_icf (ctxt);