Use gather loads for strided accesses
[official-gcc.git] / gcc / ipa-icf.c
blobb9f2bf30744b6f350bae2861126142d5591a0c5c
1 /* Interprocedural Identical Code Folding pass
2 Copyright (C) 2014-2018 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 #define INCLUDE_LIST
56 #include "system.h"
57 #include "coretypes.h"
58 #include "backend.h"
59 #include "target.h"
60 #include "rtl.h"
61 #include "tree.h"
62 #include "gimple.h"
63 #include "alloc-pool.h"
64 #include "tree-pass.h"
65 #include "ssa.h"
66 #include "cgraph.h"
67 #include "coverage.h"
68 #include "gimple-pretty-print.h"
69 #include "data-streamer.h"
70 #include "fold-const.h"
71 #include "calls.h"
72 #include "varasm.h"
73 #include "gimple-iterator.h"
74 #include "tree-cfg.h"
75 #include "symbol-summary.h"
76 #include "ipa-prop.h"
77 #include "ipa-fnsummary.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"
86 #include "tree-vector-builder.h"
88 using namespace ipa_icf_gimple;
90 namespace ipa_icf {
92 /* Initialization and computation of symtab node hash, there data
93 are propagated later on. */
95 static sem_item_optimizer *optimizer = NULL;
97 /* Constructor. */
99 symbol_compare_collection::symbol_compare_collection (symtab_node *node)
101 m_references.create (0);
102 m_interposables.create (0);
104 ipa_ref *ref;
106 if (is_a <varpool_node *> (node) && DECL_VIRTUAL_P (node->decl))
107 return;
109 for (unsigned i = 0; node->iterate_reference (i, ref); i++)
111 if (ref->address_matters_p ())
112 m_references.safe_push (ref->referred);
114 if (ref->referred->get_availability () <= AVAIL_INTERPOSABLE)
116 if (ref->address_matters_p ())
117 m_references.safe_push (ref->referred);
118 else
119 m_interposables.safe_push (ref->referred);
123 if (is_a <cgraph_node *> (node))
125 cgraph_node *cnode = dyn_cast <cgraph_node *> (node);
127 for (cgraph_edge *e = cnode->callees; e; e = e->next_callee)
128 if (e->callee->get_availability () <= AVAIL_INTERPOSABLE)
129 m_interposables.safe_push (e->callee);
133 /* Constructor for key value pair, where _ITEM is key and _INDEX is a target. */
135 sem_usage_pair::sem_usage_pair (sem_item *_item, unsigned int _index)
136 : item (_item), index (_index)
140 sem_item::sem_item (sem_item_type _type, bitmap_obstack *stack)
141 : type (_type), m_hash (-1), m_hash_set (false)
143 setup (stack);
146 sem_item::sem_item (sem_item_type _type, symtab_node *_node,
147 bitmap_obstack *stack)
148 : type (_type), node (_node), m_hash (-1), m_hash_set (false)
150 decl = node->decl;
151 setup (stack);
154 /* Add reference to a semantic TARGET. */
156 void
157 sem_item::add_reference (sem_item *target)
159 refs.safe_push (target);
160 unsigned index = refs.length ();
161 target->usages.safe_push (new sem_usage_pair(this, index));
162 bitmap_set_bit (target->usage_index_bitmap, index);
163 refs_set.add (target->node);
166 /* Initialize internal data structures. Bitmap STACK is used for
167 bitmap memory allocation process. */
169 void
170 sem_item::setup (bitmap_obstack *stack)
172 gcc_checking_assert (node);
174 refs.create (0);
175 tree_refs.create (0);
176 usages.create (0);
177 usage_index_bitmap = BITMAP_ALLOC (stack);
180 sem_item::~sem_item ()
182 for (unsigned i = 0; i < usages.length (); i++)
183 delete usages[i];
185 refs.release ();
186 tree_refs.release ();
187 usages.release ();
189 BITMAP_FREE (usage_index_bitmap);
192 /* Dump function for debugging purpose. */
194 DEBUG_FUNCTION void
195 sem_item::dump (void)
197 if (dump_file)
199 fprintf (dump_file, "[%s] %s (tree:%p)\n", type == FUNC ? "func" : "var",
200 node->dump_name (), (void *) node->decl);
201 fprintf (dump_file, " hash: %u\n", get_hash ());
202 fprintf (dump_file, " references: ");
204 for (unsigned i = 0; i < refs.length (); i++)
205 fprintf (dump_file, "%s%s ", refs[i]->node->name (),
206 i < refs.length() - 1 ? "," : "");
208 fprintf (dump_file, "\n");
212 /* Return true if target supports alias symbols. */
214 bool
215 sem_item::target_supports_symbol_aliases_p (void)
217 #if !defined (ASM_OUTPUT_DEF) || (!defined(ASM_OUTPUT_WEAK_ALIAS) && !defined (ASM_WEAKEN_DECL))
218 return false;
219 #else
220 return true;
221 #endif
224 void sem_item::set_hash (hashval_t hash)
226 m_hash = hash;
227 m_hash_set = true;
230 /* Semantic function constructor that uses STACK as bitmap memory stack. */
232 sem_function::sem_function (bitmap_obstack *stack)
233 : sem_item (FUNC, stack), m_checker (NULL), m_compared_func (NULL)
235 bb_sizes.create (0);
236 bb_sorted.create (0);
239 sem_function::sem_function (cgraph_node *node, bitmap_obstack *stack)
240 : sem_item (FUNC, node, stack), m_checker (NULL), m_compared_func (NULL)
242 bb_sizes.create (0);
243 bb_sorted.create (0);
246 sem_function::~sem_function ()
248 for (unsigned i = 0; i < bb_sorted.length (); i++)
249 delete (bb_sorted[i]);
251 bb_sizes.release ();
252 bb_sorted.release ();
255 /* Calculates hash value based on a BASIC_BLOCK. */
257 hashval_t
258 sem_function::get_bb_hash (const sem_bb *basic_block)
260 inchash::hash hstate;
262 hstate.add_int (basic_block->nondbg_stmt_count);
263 hstate.add_int (basic_block->edge_count);
265 return hstate.end ();
268 /* References independent hash function. */
270 hashval_t
271 sem_function::get_hash (void)
273 if (!m_hash_set)
275 inchash::hash hstate;
276 hstate.add_int (177454); /* Random number for function type. */
278 hstate.add_int (arg_count);
279 hstate.add_int (cfg_checksum);
280 hstate.add_int (gcode_hash);
282 for (unsigned i = 0; i < bb_sorted.length (); i++)
283 hstate.merge_hash (get_bb_hash (bb_sorted[i]));
285 for (unsigned i = 0; i < bb_sizes.length (); i++)
286 hstate.add_int (bb_sizes[i]);
288 /* Add common features of declaration itself. */
289 if (DECL_FUNCTION_SPECIFIC_TARGET (decl))
290 hstate.add_hwi
291 (cl_target_option_hash
292 (TREE_TARGET_OPTION (DECL_FUNCTION_SPECIFIC_TARGET (decl))));
293 if (DECL_FUNCTION_SPECIFIC_OPTIMIZATION (decl))
294 hstate.add_hwi
295 (cl_optimization_hash
296 (TREE_OPTIMIZATION (DECL_FUNCTION_SPECIFIC_OPTIMIZATION (decl))));
297 hstate.add_flag (DECL_CXX_CONSTRUCTOR_P (decl));
298 hstate.add_flag (DECL_CXX_DESTRUCTOR_P (decl));
300 set_hash (hstate.end ());
303 return m_hash;
306 /* Return ture if A1 and A2 represent equivalent function attribute lists.
307 Based on comp_type_attributes. */
309 bool
310 sem_item::compare_attributes (const_tree a1, const_tree a2)
312 const_tree a;
313 if (a1 == a2)
314 return true;
315 for (a = a1; a != NULL_TREE; a = TREE_CHAIN (a))
317 const struct attribute_spec *as;
318 const_tree attr;
320 as = lookup_attribute_spec (get_attribute_name (a));
321 /* TODO: We can introduce as->affects_decl_identity
322 and as->affects_decl_reference_identity if attribute mismatch
323 gets a common reason to give up on merging. It may not be worth
324 the effort.
325 For example returns_nonnull affects only references, while
326 optimize attribute can be ignored because it is already lowered
327 into flags representation and compared separately. */
328 if (!as)
329 continue;
331 attr = lookup_attribute (as->name, CONST_CAST_TREE (a2));
332 if (!attr || !attribute_value_equal (a, attr))
333 break;
335 if (!a)
337 for (a = a2; a != NULL_TREE; a = TREE_CHAIN (a))
339 const struct attribute_spec *as;
341 as = lookup_attribute_spec (get_attribute_name (a));
342 if (!as)
343 continue;
345 if (!lookup_attribute (as->name, CONST_CAST_TREE (a1)))
346 break;
347 /* We don't need to compare trees again, as we did this
348 already in first loop. */
350 if (!a)
351 return true;
353 /* TODO: As in comp_type_attributes we may want to introduce target hook. */
354 return false;
357 /* Compare properties of symbols N1 and N2 that does not affect semantics of
358 symbol itself but affects semantics of its references from USED_BY (which
359 may be NULL if it is unknown). If comparsion is false, symbols
360 can still be merged but any symbols referring them can't.
362 If ADDRESS is true, do extra checking needed for IPA_REF_ADDR.
364 TODO: We can also split attributes to those that determine codegen of
365 a function body/variable constructor itself and those that are used when
366 referring to it. */
368 bool
369 sem_item::compare_referenced_symbol_properties (symtab_node *used_by,
370 symtab_node *n1,
371 symtab_node *n2,
372 bool address)
374 if (is_a <cgraph_node *> (n1))
376 /* Inline properties matters: we do now want to merge uses of inline
377 function to uses of normal function because inline hint would be lost.
378 We however can merge inline function to noinline because the alias
379 will keep its DECL_DECLARED_INLINE flag.
381 Also ignore inline flag when optimizing for size or when function
382 is known to not be inlinable.
384 TODO: the optimize_size checks can also be assumed to be true if
385 unit has no !optimize_size functions. */
387 if ((!used_by || address || !is_a <cgraph_node *> (used_by)
388 || !opt_for_fn (used_by->decl, optimize_size))
389 && !opt_for_fn (n1->decl, optimize_size)
390 && n1->get_availability () > AVAIL_INTERPOSABLE
391 && (!DECL_UNINLINABLE (n1->decl) || !DECL_UNINLINABLE (n2->decl)))
393 if (DECL_DISREGARD_INLINE_LIMITS (n1->decl)
394 != DECL_DISREGARD_INLINE_LIMITS (n2->decl))
395 return return_false_with_msg
396 ("DECL_DISREGARD_INLINE_LIMITS are different");
398 if (DECL_DECLARED_INLINE_P (n1->decl)
399 != DECL_DECLARED_INLINE_P (n2->decl))
400 return return_false_with_msg ("inline attributes are different");
403 if (DECL_IS_OPERATOR_NEW (n1->decl)
404 != DECL_IS_OPERATOR_NEW (n2->decl))
405 return return_false_with_msg ("operator new flags are different");
408 /* Merging two definitions with a reference to equivalent vtables, but
409 belonging to a different type may result in ipa-polymorphic-call analysis
410 giving a wrong answer about the dynamic type of instance. */
411 if (is_a <varpool_node *> (n1))
413 if ((DECL_VIRTUAL_P (n1->decl) || DECL_VIRTUAL_P (n2->decl))
414 && (DECL_VIRTUAL_P (n1->decl) != DECL_VIRTUAL_P (n2->decl)
415 || !types_must_be_same_for_odr (DECL_CONTEXT (n1->decl),
416 DECL_CONTEXT (n2->decl)))
417 && (!used_by || !is_a <cgraph_node *> (used_by) || address
418 || opt_for_fn (used_by->decl, flag_devirtualize)))
419 return return_false_with_msg
420 ("references to virtual tables can not be merged");
422 if (address && DECL_ALIGN (n1->decl) != DECL_ALIGN (n2->decl))
423 return return_false_with_msg ("alignment mismatch");
425 /* For functions we compare attributes in equals_wpa, because we do
426 not know what attributes may cause codegen differences, but for
427 variables just compare attributes for references - the codegen
428 for constructors is affected only by those attributes that we lower
429 to explicit representation (such as DECL_ALIGN or DECL_SECTION). */
430 if (!compare_attributes (DECL_ATTRIBUTES (n1->decl),
431 DECL_ATTRIBUTES (n2->decl)))
432 return return_false_with_msg ("different var decl attributes");
433 if (comp_type_attributes (TREE_TYPE (n1->decl),
434 TREE_TYPE (n2->decl)) != 1)
435 return return_false_with_msg ("different var type attributes");
438 /* When matching virtual tables, be sure to also match information
439 relevant for polymorphic call analysis. */
440 if (used_by && is_a <varpool_node *> (used_by)
441 && DECL_VIRTUAL_P (used_by->decl))
443 if (DECL_VIRTUAL_P (n1->decl) != DECL_VIRTUAL_P (n2->decl))
444 return return_false_with_msg ("virtual flag mismatch");
445 if (DECL_VIRTUAL_P (n1->decl) && is_a <cgraph_node *> (n1)
446 && (DECL_FINAL_P (n1->decl) != DECL_FINAL_P (n2->decl)))
447 return return_false_with_msg ("final flag mismatch");
449 return true;
452 /* Hash properties that are compared by compare_referenced_symbol_properties. */
454 void
455 sem_item::hash_referenced_symbol_properties (symtab_node *ref,
456 inchash::hash &hstate,
457 bool address)
459 if (is_a <cgraph_node *> (ref))
461 if ((type != FUNC || address || !opt_for_fn (decl, optimize_size))
462 && !opt_for_fn (ref->decl, optimize_size)
463 && !DECL_UNINLINABLE (ref->decl))
465 hstate.add_flag (DECL_DISREGARD_INLINE_LIMITS (ref->decl));
466 hstate.add_flag (DECL_DECLARED_INLINE_P (ref->decl));
468 hstate.add_flag (DECL_IS_OPERATOR_NEW (ref->decl));
470 else if (is_a <varpool_node *> (ref))
472 hstate.add_flag (DECL_VIRTUAL_P (ref->decl));
473 if (address)
474 hstate.add_int (DECL_ALIGN (ref->decl));
479 /* For a given symbol table nodes N1 and N2, we check that FUNCTION_DECLs
480 point to a same function. Comparison can be skipped if IGNORED_NODES
481 contains these nodes. ADDRESS indicate if address is taken. */
483 bool
484 sem_item::compare_symbol_references (
485 hash_map <symtab_node *, sem_item *> &ignored_nodes,
486 symtab_node *n1, symtab_node *n2, bool address)
488 enum availability avail1, avail2;
490 if (n1 == n2)
491 return true;
493 /* Never match variable and function. */
494 if (is_a <varpool_node *> (n1) != is_a <varpool_node *> (n2))
495 return false;
497 if (!compare_referenced_symbol_properties (node, n1, n2, address))
498 return false;
499 if (address && n1->equal_address_to (n2) == 1)
500 return true;
501 if (!address && n1->semantically_equivalent_p (n2))
502 return true;
504 n1 = n1->ultimate_alias_target (&avail1);
505 n2 = n2->ultimate_alias_target (&avail2);
507 if (avail1 > AVAIL_INTERPOSABLE && ignored_nodes.get (n1)
508 && avail2 > AVAIL_INTERPOSABLE && ignored_nodes.get (n2))
509 return true;
511 return return_false_with_msg ("different references");
514 /* If cgraph edges E1 and E2 are indirect calls, verify that
515 ECF flags are the same. */
517 bool sem_function::compare_edge_flags (cgraph_edge *e1, cgraph_edge *e2)
519 if (e1->indirect_info && e2->indirect_info)
521 int e1_flags = e1->indirect_info->ecf_flags;
522 int e2_flags = e2->indirect_info->ecf_flags;
524 if (e1_flags != e2_flags)
525 return return_false_with_msg ("ICF flags are different");
527 else if (e1->indirect_info || e2->indirect_info)
528 return false;
530 return true;
533 /* Return true if parameter I may be used. */
535 bool
536 sem_function::param_used_p (unsigned int i)
538 if (ipa_node_params_sum == NULL)
539 return true;
541 struct ipa_node_params *parms_info = IPA_NODE_REF (get_node ());
543 if (vec_safe_length (parms_info->descriptors) <= i)
544 return true;
546 return ipa_is_param_used (IPA_NODE_REF (get_node ()), i);
549 /* Perform additional check needed to match types function parameters that are
550 used. Unlike for normal decls it matters if type is TYPE_RESTRICT and we
551 make an assumption that REFERENCE_TYPE parameters are always non-NULL. */
553 bool
554 sem_function::compatible_parm_types_p (tree parm1, tree parm2)
556 /* Be sure that parameters are TBAA compatible. */
557 if (!func_checker::compatible_types_p (parm1, parm2))
558 return return_false_with_msg ("parameter type is not compatible");
560 if (POINTER_TYPE_P (parm1)
561 && (TYPE_RESTRICT (parm1) != TYPE_RESTRICT (parm2)))
562 return return_false_with_msg ("argument restrict flag mismatch");
564 /* nonnull_arg_p implies non-zero range to REFERENCE types. */
565 if (POINTER_TYPE_P (parm1)
566 && TREE_CODE (parm1) != TREE_CODE (parm2)
567 && opt_for_fn (decl, flag_delete_null_pointer_checks))
568 return return_false_with_msg ("pointer wrt reference mismatch");
570 return true;
573 /* Fast equality function based on knowledge known in WPA. */
575 bool
576 sem_function::equals_wpa (sem_item *item,
577 hash_map <symtab_node *, sem_item *> &ignored_nodes)
579 gcc_assert (item->type == FUNC);
580 cgraph_node *cnode = dyn_cast <cgraph_node *> (node);
581 cgraph_node *cnode2 = dyn_cast <cgraph_node *> (item->node);
583 m_compared_func = static_cast<sem_function *> (item);
585 if (cnode->thunk.thunk_p != cnode2->thunk.thunk_p)
586 return return_false_with_msg ("thunk_p mismatch");
588 if (cnode->thunk.thunk_p)
590 if (cnode->thunk.fixed_offset != cnode2->thunk.fixed_offset)
591 return return_false_with_msg ("thunk fixed_offset mismatch");
592 if (cnode->thunk.virtual_value != cnode2->thunk.virtual_value)
593 return return_false_with_msg ("thunk virtual_value mismatch");
594 if (cnode->thunk.this_adjusting != cnode2->thunk.this_adjusting)
595 return return_false_with_msg ("thunk this_adjusting mismatch");
596 if (cnode->thunk.virtual_offset_p != cnode2->thunk.virtual_offset_p)
597 return return_false_with_msg ("thunk virtual_offset_p mismatch");
598 if (cnode->thunk.add_pointer_bounds_args
599 != cnode2->thunk.add_pointer_bounds_args)
600 return return_false_with_msg ("thunk add_pointer_bounds_args mismatch");
603 /* Compare special function DECL attributes. */
604 if (DECL_FUNCTION_PERSONALITY (decl)
605 != DECL_FUNCTION_PERSONALITY (item->decl))
606 return return_false_with_msg ("function personalities are different");
608 if (DECL_NO_INSTRUMENT_FUNCTION_ENTRY_EXIT (decl)
609 != DECL_NO_INSTRUMENT_FUNCTION_ENTRY_EXIT (item->decl))
610 return return_false_with_msg ("intrument function entry exit "
611 "attributes are different");
613 if (DECL_NO_LIMIT_STACK (decl) != DECL_NO_LIMIT_STACK (item->decl))
614 return return_false_with_msg ("no stack limit attributes are different");
616 if (DECL_CXX_CONSTRUCTOR_P (decl) != DECL_CXX_CONSTRUCTOR_P (item->decl))
617 return return_false_with_msg ("DECL_CXX_CONSTRUCTOR mismatch");
619 if (DECL_CXX_DESTRUCTOR_P (decl) != DECL_CXX_DESTRUCTOR_P (item->decl))
620 return return_false_with_msg ("DECL_CXX_DESTRUCTOR mismatch");
622 /* TODO: pure/const flags mostly matters only for references, except for
623 the fact that codegen takes LOOPING flag as a hint that loops are
624 finite. We may arrange the code to always pick leader that has least
625 specified flags and then this can go into comparing symbol properties. */
626 if (flags_from_decl_or_type (decl) != flags_from_decl_or_type (item->decl))
627 return return_false_with_msg ("decl_or_type flags are different");
629 /* Do not match polymorphic constructors of different types. They calls
630 type memory location for ipa-polymorphic-call and we do not want
631 it to get confused by wrong type. */
632 if (DECL_CXX_CONSTRUCTOR_P (decl)
633 && TREE_CODE (TREE_TYPE (decl)) == METHOD_TYPE)
635 if (TREE_CODE (TREE_TYPE (item->decl)) != METHOD_TYPE)
636 return return_false_with_msg ("DECL_CXX_CONSTURCTOR type mismatch");
637 else if (!func_checker::compatible_polymorphic_types_p
638 (TYPE_METHOD_BASETYPE (TREE_TYPE (decl)),
639 TYPE_METHOD_BASETYPE (TREE_TYPE (item->decl)), false))
640 return return_false_with_msg ("ctor polymorphic type mismatch");
643 /* Checking function TARGET and OPTIMIZATION flags. */
644 cl_target_option *tar1 = target_opts_for_fn (decl);
645 cl_target_option *tar2 = target_opts_for_fn (item->decl);
647 if (tar1 != tar2 && !cl_target_option_eq (tar1, tar2))
649 if (dump_file && (dump_flags & TDF_DETAILS))
651 fprintf (dump_file, "target flags difference");
652 cl_target_option_print_diff (dump_file, 2, tar1, tar2);
655 return return_false_with_msg ("Target flags are different");
658 cl_optimization *opt1 = opts_for_fn (decl);
659 cl_optimization *opt2 = opts_for_fn (item->decl);
661 if (opt1 != opt2 && memcmp (opt1, opt2, sizeof(cl_optimization)))
663 if (dump_file && (dump_flags & TDF_DETAILS))
665 fprintf (dump_file, "optimization flags difference");
666 cl_optimization_print_diff (dump_file, 2, opt1, opt2);
669 return return_false_with_msg ("optimization flags are different");
672 /* Result type checking. */
673 if (!func_checker::compatible_types_p
674 (TREE_TYPE (TREE_TYPE (decl)),
675 TREE_TYPE (TREE_TYPE (m_compared_func->decl))))
676 return return_false_with_msg ("result types are different");
678 /* Checking types of arguments. */
679 tree list1 = TYPE_ARG_TYPES (TREE_TYPE (decl)),
680 list2 = TYPE_ARG_TYPES (TREE_TYPE (m_compared_func->decl));
681 for (unsigned i = 0; list1 && list2;
682 list1 = TREE_CHAIN (list1), list2 = TREE_CHAIN (list2), i++)
684 tree parm1 = TREE_VALUE (list1);
685 tree parm2 = TREE_VALUE (list2);
687 /* This guard is here for function pointer with attributes (pr59927.c). */
688 if (!parm1 || !parm2)
689 return return_false_with_msg ("NULL argument type");
691 /* Verify that types are compatible to ensure that both functions
692 have same calling conventions. */
693 if (!types_compatible_p (parm1, parm2))
694 return return_false_with_msg ("parameter types are not compatible");
696 if (!param_used_p (i))
697 continue;
699 /* Perform additional checks for used parameters. */
700 if (!compatible_parm_types_p (parm1, parm2))
701 return false;
704 if (list1 || list2)
705 return return_false_with_msg ("Mismatched number of parameters");
707 if (node->num_references () != item->node->num_references ())
708 return return_false_with_msg ("different number of references");
710 /* Checking function attributes.
711 This is quadratic in number of attributes */
712 if (comp_type_attributes (TREE_TYPE (decl),
713 TREE_TYPE (item->decl)) != 1)
714 return return_false_with_msg ("different type attributes");
715 if (!compare_attributes (DECL_ATTRIBUTES (decl),
716 DECL_ATTRIBUTES (item->decl)))
717 return return_false_with_msg ("different decl attributes");
719 /* The type of THIS pointer type memory location for
720 ipa-polymorphic-call-analysis. */
721 if (opt_for_fn (decl, flag_devirtualize)
722 && (TREE_CODE (TREE_TYPE (decl)) == METHOD_TYPE
723 || TREE_CODE (TREE_TYPE (item->decl)) == METHOD_TYPE)
724 && param_used_p (0)
725 && compare_polymorphic_p ())
727 if (TREE_CODE (TREE_TYPE (decl)) != TREE_CODE (TREE_TYPE (item->decl)))
728 return return_false_with_msg ("METHOD_TYPE and FUNCTION_TYPE mismatch");
729 if (!func_checker::compatible_polymorphic_types_p
730 (TYPE_METHOD_BASETYPE (TREE_TYPE (decl)),
731 TYPE_METHOD_BASETYPE (TREE_TYPE (item->decl)), false))
732 return return_false_with_msg ("THIS pointer ODR type mismatch");
735 ipa_ref *ref = NULL, *ref2 = NULL;
736 for (unsigned i = 0; node->iterate_reference (i, ref); i++)
738 item->node->iterate_reference (i, ref2);
740 if (ref->use != ref2->use)
741 return return_false_with_msg ("reference use mismatch");
743 if (!compare_symbol_references (ignored_nodes, ref->referred,
744 ref2->referred,
745 ref->address_matters_p ()))
746 return false;
749 cgraph_edge *e1 = dyn_cast <cgraph_node *> (node)->callees;
750 cgraph_edge *e2 = dyn_cast <cgraph_node *> (item->node)->callees;
752 while (e1 && e2)
754 if (!compare_symbol_references (ignored_nodes, e1->callee,
755 e2->callee, false))
756 return false;
757 if (!compare_edge_flags (e1, e2))
758 return false;
760 e1 = e1->next_callee;
761 e2 = e2->next_callee;
764 if (e1 || e2)
765 return return_false_with_msg ("different number of calls");
767 e1 = dyn_cast <cgraph_node *> (node)->indirect_calls;
768 e2 = dyn_cast <cgraph_node *> (item->node)->indirect_calls;
770 while (e1 && e2)
772 if (!compare_edge_flags (e1, e2))
773 return false;
775 e1 = e1->next_callee;
776 e2 = e2->next_callee;
779 if (e1 || e2)
780 return return_false_with_msg ("different number of indirect calls");
782 return true;
785 /* Update hash by address sensitive references. We iterate over all
786 sensitive references (address_matters_p) and we hash ultime alias
787 target of these nodes, which can improve a semantic item hash.
789 Also hash in referenced symbols properties. This can be done at any time
790 (as the properties should not change), but it is convenient to do it here
791 while we walk the references anyway. */
793 void
794 sem_item::update_hash_by_addr_refs (hash_map <symtab_node *,
795 sem_item *> &m_symtab_node_map)
797 ipa_ref* ref;
798 inchash::hash hstate (get_hash ());
800 for (unsigned i = 0; node->iterate_reference (i, ref); i++)
802 hstate.add_int (ref->use);
803 hash_referenced_symbol_properties (ref->referred, hstate,
804 ref->use == IPA_REF_ADDR);
805 if (ref->address_matters_p () || !m_symtab_node_map.get (ref->referred))
806 hstate.add_int (ref->referred->ultimate_alias_target ()->order);
809 if (is_a <cgraph_node *> (node))
811 for (cgraph_edge *e = dyn_cast <cgraph_node *> (node)->callers; e;
812 e = e->next_caller)
814 sem_item **result = m_symtab_node_map.get (e->callee);
815 hash_referenced_symbol_properties (e->callee, hstate, false);
816 if (!result)
817 hstate.add_int (e->callee->ultimate_alias_target ()->order);
821 set_hash (hstate.end ());
824 /* Update hash by computed local hash values taken from different
825 semantic items.
826 TODO: stronger SCC based hashing would be desirable here. */
828 void
829 sem_item::update_hash_by_local_refs (hash_map <symtab_node *,
830 sem_item *> &m_symtab_node_map)
832 ipa_ref* ref;
833 inchash::hash state (get_hash ());
835 for (unsigned j = 0; node->iterate_reference (j, ref); j++)
837 sem_item **result = m_symtab_node_map.get (ref->referring);
838 if (result)
839 state.merge_hash ((*result)->get_hash ());
842 if (type == FUNC)
844 for (cgraph_edge *e = dyn_cast <cgraph_node *> (node)->callees; e;
845 e = e->next_callee)
847 sem_item **result = m_symtab_node_map.get (e->caller);
848 if (result)
849 state.merge_hash ((*result)->get_hash ());
853 global_hash = state.end ();
856 /* Returns true if the item equals to ITEM given as argument. */
858 bool
859 sem_function::equals (sem_item *item,
860 hash_map <symtab_node *, sem_item *> &)
862 gcc_assert (item->type == FUNC);
863 bool eq = equals_private (item);
865 if (m_checker != NULL)
867 delete m_checker;
868 m_checker = NULL;
871 if (dump_file && (dump_flags & TDF_DETAILS))
872 fprintf (dump_file,
873 "Equals called for: %s:%s with result: %s\n\n",
874 node->dump_name (),
875 item->node->dump_name (),
876 eq ? "true" : "false");
878 return eq;
881 /* Processes function equality comparison. */
883 bool
884 sem_function::equals_private (sem_item *item)
886 if (item->type != FUNC)
887 return false;
889 basic_block bb1, bb2;
890 edge e1, e2;
891 edge_iterator ei1, ei2;
892 bool result = true;
893 tree arg1, arg2;
895 m_compared_func = static_cast<sem_function *> (item);
897 gcc_assert (decl != item->decl);
899 if (bb_sorted.length () != m_compared_func->bb_sorted.length ()
900 || edge_count != m_compared_func->edge_count
901 || cfg_checksum != m_compared_func->cfg_checksum)
902 return return_false ();
904 m_checker = new func_checker (decl, m_compared_func->decl,
905 compare_polymorphic_p (),
906 false,
907 &refs_set,
908 &m_compared_func->refs_set);
909 arg1 = DECL_ARGUMENTS (decl);
910 arg2 = DECL_ARGUMENTS (m_compared_func->decl);
911 for (unsigned i = 0;
912 arg1 && arg2; arg1 = DECL_CHAIN (arg1), arg2 = DECL_CHAIN (arg2), i++)
914 if (!types_compatible_p (TREE_TYPE (arg1), TREE_TYPE (arg2)))
915 return return_false_with_msg ("argument types are not compatible");
916 if (!param_used_p (i))
917 continue;
918 /* Perform additional checks for used parameters. */
919 if (!compatible_parm_types_p (TREE_TYPE (arg1), TREE_TYPE (arg2)))
920 return false;
921 if (!m_checker->compare_decl (arg1, arg2))
922 return return_false ();
924 if (arg1 || arg2)
925 return return_false_with_msg ("Mismatched number of arguments");
927 if (!dyn_cast <cgraph_node *> (node)->has_gimple_body_p ())
928 return true;
930 /* Fill-up label dictionary. */
931 for (unsigned i = 0; i < bb_sorted.length (); ++i)
933 m_checker->parse_labels (bb_sorted[i]);
934 m_checker->parse_labels (m_compared_func->bb_sorted[i]);
937 /* Checking all basic blocks. */
938 for (unsigned i = 0; i < bb_sorted.length (); ++i)
939 if(!m_checker->compare_bb (bb_sorted[i], m_compared_func->bb_sorted[i]))
940 return return_false();
942 dump_message ("All BBs are equal\n");
944 auto_vec <int> bb_dict;
946 /* Basic block edges check. */
947 for (unsigned i = 0; i < bb_sorted.length (); ++i)
949 bb1 = bb_sorted[i]->bb;
950 bb2 = m_compared_func->bb_sorted[i]->bb;
952 ei2 = ei_start (bb2->preds);
954 for (ei1 = ei_start (bb1->preds); ei_cond (ei1, &e1); ei_next (&ei1))
956 ei_cond (ei2, &e2);
958 if (e1->flags != e2->flags)
959 return return_false_with_msg ("flags comparison returns false");
961 if (!bb_dict_test (&bb_dict, e1->src->index, e2->src->index))
962 return return_false_with_msg ("edge comparison returns false");
964 if (!bb_dict_test (&bb_dict, e1->dest->index, e2->dest->index))
965 return return_false_with_msg ("BB comparison returns false");
967 if (!m_checker->compare_edge (e1, e2))
968 return return_false_with_msg ("edge comparison returns false");
970 ei_next (&ei2);
974 /* Basic block PHI nodes comparison. */
975 for (unsigned i = 0; i < bb_sorted.length (); i++)
976 if (!compare_phi_node (bb_sorted[i]->bb, m_compared_func->bb_sorted[i]->bb))
977 return return_false_with_msg ("PHI node comparison returns false");
979 return result;
982 /* Set LOCAL_P of NODE to true if DATA is non-NULL.
983 Helper for call_for_symbol_thunks_and_aliases. */
985 static bool
986 set_local (cgraph_node *node, void *data)
988 node->local.local = data != NULL;
989 return false;
992 /* TREE_ADDRESSABLE of NODE to true.
993 Helper for call_for_symbol_thunks_and_aliases. */
995 static bool
996 set_addressable (varpool_node *node, void *)
998 TREE_ADDRESSABLE (node->decl) = 1;
999 return false;
1002 /* Clear DECL_RTL of NODE.
1003 Helper for call_for_symbol_thunks_and_aliases. */
1005 static bool
1006 clear_decl_rtl (symtab_node *node, void *)
1008 SET_DECL_RTL (node->decl, NULL);
1009 return false;
1012 /* Redirect all callers of N and its aliases to TO. Remove aliases if
1013 possible. Return number of redirections made. */
1015 static int
1016 redirect_all_callers (cgraph_node *n, cgraph_node *to)
1018 int nredirected = 0;
1019 ipa_ref *ref;
1020 cgraph_edge *e = n->callers;
1022 while (e)
1024 /* Redirecting thunks to interposable symbols or symbols in other sections
1025 may not be supported by target output code. Play safe for now and
1026 punt on redirection. */
1027 if (!e->caller->thunk.thunk_p)
1029 struct cgraph_edge *nexte = e->next_caller;
1030 e->redirect_callee (to);
1031 e = nexte;
1032 nredirected++;
1034 else
1035 e = e->next_callee;
1037 for (unsigned i = 0; n->iterate_direct_aliases (i, ref);)
1039 bool removed = false;
1040 cgraph_node *n_alias = dyn_cast <cgraph_node *> (ref->referring);
1042 if ((DECL_COMDAT_GROUP (n->decl)
1043 && (DECL_COMDAT_GROUP (n->decl)
1044 == DECL_COMDAT_GROUP (n_alias->decl)))
1045 || (n_alias->get_availability () > AVAIL_INTERPOSABLE
1046 && n->get_availability () > AVAIL_INTERPOSABLE))
1048 nredirected += redirect_all_callers (n_alias, to);
1049 if (n_alias->can_remove_if_no_direct_calls_p ()
1050 && !n_alias->call_for_symbol_and_aliases (cgraph_node::has_thunk_p,
1051 NULL, true)
1052 && !n_alias->has_aliases_p ())
1053 n_alias->remove ();
1055 if (!removed)
1056 i++;
1058 return nredirected;
1061 /* Merges instance with an ALIAS_ITEM, where alias, thunk or redirection can
1062 be applied. */
1064 bool
1065 sem_function::merge (sem_item *alias_item)
1067 gcc_assert (alias_item->type == FUNC);
1069 sem_function *alias_func = static_cast<sem_function *> (alias_item);
1071 cgraph_node *original = get_node ();
1072 cgraph_node *local_original = NULL;
1073 cgraph_node *alias = alias_func->get_node ();
1075 bool create_wrapper = false;
1076 bool create_alias = false;
1077 bool redirect_callers = false;
1078 bool remove = false;
1080 bool original_discardable = false;
1081 bool original_discarded = false;
1083 bool original_address_matters = original->address_matters_p ();
1084 bool alias_address_matters = alias->address_matters_p ();
1086 if (DECL_EXTERNAL (alias->decl))
1088 if (dump_file)
1089 fprintf (dump_file, "Not unifying; alias is external.\n\n");
1090 return false;
1093 if (DECL_NO_INLINE_WARNING_P (original->decl)
1094 != DECL_NO_INLINE_WARNING_P (alias->decl))
1096 if (dump_file)
1097 fprintf (dump_file,
1098 "Not unifying; "
1099 "DECL_NO_INLINE_WARNING mismatch.\n\n");
1100 return false;
1103 /* Do not attempt to mix functions from different user sections;
1104 we do not know what user intends with those. */
1105 if (((DECL_SECTION_NAME (original->decl) && !original->implicit_section)
1106 || (DECL_SECTION_NAME (alias->decl) && !alias->implicit_section))
1107 && DECL_SECTION_NAME (original->decl) != DECL_SECTION_NAME (alias->decl))
1109 if (dump_file)
1110 fprintf (dump_file,
1111 "Not unifying; "
1112 "original and alias are in different sections.\n\n");
1113 return false;
1116 if (!original->in_same_comdat_group_p (alias)
1117 || original->comdat_local_p ())
1119 if (dump_file)
1120 fprintf (dump_file,
1121 "Not unifying; alias nor wrapper cannot be created; "
1122 "across comdat group boundary\n\n");
1124 return false;
1127 /* See if original is in a section that can be discarded if the main
1128 symbol is not used. */
1130 if (original->can_be_discarded_p ())
1131 original_discardable = true;
1132 /* Also consider case where we have resolution info and we know that
1133 original's definition is not going to be used. In this case we can not
1134 create alias to original. */
1135 if (node->resolution != LDPR_UNKNOWN
1136 && !decl_binds_to_current_def_p (node->decl))
1137 original_discardable = original_discarded = true;
1139 /* Creating a symtab alias is the optimal way to merge.
1140 It however can not be used in the following cases:
1142 1) if ORIGINAL and ALIAS may be possibly compared for address equality.
1143 2) if ORIGINAL is in a section that may be discarded by linker or if
1144 it is an external functions where we can not create an alias
1145 (ORIGINAL_DISCARDABLE)
1146 3) if target do not support symbol aliases.
1147 4) original and alias lie in different comdat groups.
1149 If we can not produce alias, we will turn ALIAS into WRAPPER of ORIGINAL
1150 and/or redirect all callers from ALIAS to ORIGINAL. */
1151 if ((original_address_matters && alias_address_matters)
1152 || (original_discardable
1153 && (!DECL_COMDAT_GROUP (alias->decl)
1154 || (DECL_COMDAT_GROUP (alias->decl)
1155 != DECL_COMDAT_GROUP (original->decl))))
1156 || original_discarded
1157 || !sem_item::target_supports_symbol_aliases_p ()
1158 || DECL_COMDAT_GROUP (alias->decl) != DECL_COMDAT_GROUP (original->decl))
1160 /* First see if we can produce wrapper. */
1162 /* Symbol properties that matter for references must be preserved.
1163 TODO: We can produce wrapper, but we need to produce alias of ORIGINAL
1164 with proper properties. */
1165 if (!sem_item::compare_referenced_symbol_properties (NULL, original, alias,
1166 alias->address_taken))
1168 if (dump_file)
1169 fprintf (dump_file,
1170 "Wrapper cannot be created because referenced symbol "
1171 "properties mismatch\n");
1173 /* Do not turn function in one comdat group into wrapper to another
1174 comdat group. Other compiler producing the body of the
1175 another comdat group may make opossite decision and with unfortunate
1176 linker choices this may close a loop. */
1177 else if (DECL_COMDAT_GROUP (original->decl)
1178 && DECL_COMDAT_GROUP (alias->decl)
1179 && (DECL_COMDAT_GROUP (alias->decl)
1180 != DECL_COMDAT_GROUP (original->decl)))
1182 if (dump_file)
1183 fprintf (dump_file,
1184 "Wrapper cannot be created because of COMDAT\n");
1186 else if (DECL_STATIC_CHAIN (alias->decl)
1187 || DECL_STATIC_CHAIN (original->decl))
1189 if (dump_file)
1190 fprintf (dump_file,
1191 "Cannot create wrapper of nested function.\n");
1193 /* TODO: We can also deal with variadic functions never calling
1194 VA_START. */
1195 else if (stdarg_p (TREE_TYPE (alias->decl)))
1197 if (dump_file)
1198 fprintf (dump_file,
1199 "can not create wrapper of stdarg function.\n");
1201 else if (ipa_fn_summaries
1202 && ipa_fn_summaries->get (alias)->self_size <= 2)
1204 if (dump_file)
1205 fprintf (dump_file, "Wrapper creation is not "
1206 "profitable (function is too small).\n");
1208 /* If user paid attention to mark function noinline, assume it is
1209 somewhat special and do not try to turn it into a wrapper that can
1210 not be undone by inliner. */
1211 else if (lookup_attribute ("noinline", DECL_ATTRIBUTES (alias->decl)))
1213 if (dump_file)
1214 fprintf (dump_file, "Wrappers are not created for noinline.\n");
1216 else
1217 create_wrapper = true;
1219 /* We can redirect local calls in the case both alias and orignal
1220 are not interposable. */
1221 redirect_callers
1222 = alias->get_availability () > AVAIL_INTERPOSABLE
1223 && original->get_availability () > AVAIL_INTERPOSABLE
1224 && !alias->instrumented_version;
1225 /* TODO: We can redirect, but we need to produce alias of ORIGINAL
1226 with proper properties. */
1227 if (!sem_item::compare_referenced_symbol_properties (NULL, original, alias,
1228 alias->address_taken))
1229 redirect_callers = false;
1231 if (!redirect_callers && !create_wrapper)
1233 if (dump_file)
1234 fprintf (dump_file, "Not unifying; can not redirect callers nor "
1235 "produce wrapper\n\n");
1236 return false;
1239 /* Work out the symbol the wrapper should call.
1240 If ORIGINAL is interposable, we need to call a local alias.
1241 Also produce local alias (if possible) as an optimization.
1243 Local aliases can not be created inside comdat groups because that
1244 prevents inlining. */
1245 if (!original_discardable && !original->get_comdat_group ())
1247 local_original
1248 = dyn_cast <cgraph_node *> (original->noninterposable_alias ());
1249 if (!local_original
1250 && original->get_availability () > AVAIL_INTERPOSABLE)
1251 local_original = original;
1253 /* If we can not use local alias, fallback to the original
1254 when possible. */
1255 else if (original->get_availability () > AVAIL_INTERPOSABLE)
1256 local_original = original;
1258 /* If original is COMDAT local, we can not really redirect calls outside
1259 of its comdat group to it. */
1260 if (original->comdat_local_p ())
1261 redirect_callers = false;
1262 if (!local_original)
1264 if (dump_file)
1265 fprintf (dump_file, "Not unifying; "
1266 "can not produce local alias.\n\n");
1267 return false;
1270 if (!redirect_callers && !create_wrapper)
1272 if (dump_file)
1273 fprintf (dump_file, "Not unifying; "
1274 "can not redirect callers nor produce a wrapper\n\n");
1275 return false;
1277 if (!create_wrapper
1278 && !alias->call_for_symbol_and_aliases (cgraph_node::has_thunk_p,
1279 NULL, true)
1280 && !alias->can_remove_if_no_direct_calls_p ())
1282 if (dump_file)
1283 fprintf (dump_file, "Not unifying; can not make wrapper and "
1284 "function has other uses than direct calls\n\n");
1285 return false;
1288 else
1289 create_alias = true;
1291 if (redirect_callers)
1293 int nredirected = redirect_all_callers (alias, local_original);
1295 if (nredirected)
1297 alias->icf_merged = true;
1298 local_original->icf_merged = true;
1300 if (dump_file && nredirected)
1301 fprintf (dump_file, "%i local calls have been "
1302 "redirected.\n", nredirected);
1305 /* If all callers was redirected, do not produce wrapper. */
1306 if (alias->can_remove_if_no_direct_calls_p ()
1307 && !DECL_VIRTUAL_P (alias->decl)
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 /* FIXME update local_original counts. */
1344 ipa_merge_profiles (original, alias, true);
1345 alias->create_wrapper (local_original);
1347 if (dump_file)
1348 fprintf (dump_file, "Unified; Wrapper has been created.\n\n");
1351 /* It's possible that redirection can hit thunks that block
1352 redirection opportunities. */
1353 gcc_assert (alias->icf_merged || remove || redirect_callers);
1354 original->icf_merged = true;
1356 /* We use merged flag to track cases where COMDAT function is known to be
1357 compatible its callers. If we merged in non-COMDAT, we need to give up
1358 on this optimization. */
1359 if (original->merged_comdat && !alias->merged_comdat)
1361 if (dump_file)
1362 fprintf (dump_file, "Dropping merged_comdat flag.\n\n");
1363 if (local_original)
1364 local_original->merged_comdat = false;
1365 original->merged_comdat = false;
1368 if (remove)
1370 ipa_merge_profiles (original, alias);
1371 alias->release_body ();
1372 alias->reset ();
1373 alias->body_removed = true;
1374 alias->icf_merged = true;
1375 if (dump_file)
1376 fprintf (dump_file, "Unified; Function body was removed.\n");
1379 return true;
1382 /* Semantic item initialization function. */
1384 void
1385 sem_function::init (void)
1387 if (in_lto_p)
1388 get_node ()->get_untransformed_body ();
1390 tree fndecl = node->decl;
1391 function *func = DECL_STRUCT_FUNCTION (fndecl);
1393 gcc_assert (func);
1394 gcc_assert (SSANAMES (func));
1396 ssa_names_size = SSANAMES (func)->length ();
1397 node = node;
1399 decl = fndecl;
1400 region_tree = func->eh->region_tree;
1402 /* iterating all function arguments. */
1403 arg_count = count_formal_params (fndecl);
1405 edge_count = n_edges_for_fn (func);
1406 cgraph_node *cnode = dyn_cast <cgraph_node *> (node);
1407 if (!cnode->thunk.thunk_p)
1409 cfg_checksum = coverage_compute_cfg_checksum (func);
1411 inchash::hash hstate;
1413 basic_block bb;
1414 FOR_EACH_BB_FN (bb, func)
1416 unsigned nondbg_stmt_count = 0;
1418 edge e;
1419 for (edge_iterator ei = ei_start (bb->preds); ei_cond (ei, &e);
1420 ei_next (&ei))
1421 cfg_checksum = iterative_hash_host_wide_int (e->flags,
1422 cfg_checksum);
1424 for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi);
1425 gsi_next (&gsi))
1427 gimple *stmt = gsi_stmt (gsi);
1429 if (gimple_code (stmt) != GIMPLE_DEBUG
1430 && gimple_code (stmt) != GIMPLE_PREDICT)
1432 hash_stmt (stmt, hstate);
1433 nondbg_stmt_count++;
1437 hstate.commit_flag ();
1438 gcode_hash = hstate.end ();
1439 bb_sizes.safe_push (nondbg_stmt_count);
1441 /* Inserting basic block to hash table. */
1442 sem_bb *semantic_bb = new sem_bb (bb, nondbg_stmt_count,
1443 EDGE_COUNT (bb->preds)
1444 + EDGE_COUNT (bb->succs));
1446 bb_sorted.safe_push (semantic_bb);
1449 else
1451 cfg_checksum = 0;
1452 inchash::hash hstate;
1453 hstate.add_hwi (cnode->thunk.fixed_offset);
1454 hstate.add_hwi (cnode->thunk.virtual_value);
1455 hstate.add_flag (cnode->thunk.this_adjusting);
1456 hstate.add_flag (cnode->thunk.virtual_offset_p);
1457 hstate.add_flag (cnode->thunk.add_pointer_bounds_args);
1458 gcode_hash = hstate.end ();
1462 /* Accumulate to HSTATE a hash of expression EXP.
1463 Identical to inchash::add_expr, but guaranteed to be stable across LTO
1464 and DECL equality classes. */
1466 void
1467 sem_item::add_expr (const_tree exp, inchash::hash &hstate)
1469 if (exp == NULL_TREE)
1471 hstate.merge_hash (0);
1472 return;
1475 /* Handled component can be matched in a cureful way proving equivalence
1476 even if they syntactically differ. Just skip them. */
1477 STRIP_NOPS (exp);
1478 while (handled_component_p (exp))
1479 exp = TREE_OPERAND (exp, 0);
1481 enum tree_code code = TREE_CODE (exp);
1482 hstate.add_int (code);
1484 switch (code)
1486 /* Use inchash::add_expr for everything that is LTO stable. */
1487 case VOID_CST:
1488 case INTEGER_CST:
1489 case REAL_CST:
1490 case FIXED_CST:
1491 case STRING_CST:
1492 case COMPLEX_CST:
1493 case VECTOR_CST:
1494 inchash::add_expr (exp, hstate);
1495 break;
1496 case CONSTRUCTOR:
1498 unsigned HOST_WIDE_INT idx;
1499 tree value;
1501 hstate.add_hwi (int_size_in_bytes (TREE_TYPE (exp)));
1503 FOR_EACH_CONSTRUCTOR_VALUE (CONSTRUCTOR_ELTS (exp), idx, value)
1504 if (value)
1505 add_expr (value, hstate);
1506 break;
1508 case ADDR_EXPR:
1509 case FDESC_EXPR:
1510 add_expr (get_base_address (TREE_OPERAND (exp, 0)), hstate);
1511 break;
1512 case SSA_NAME:
1513 case VAR_DECL:
1514 case CONST_DECL:
1515 case PARM_DECL:
1516 hstate.add_hwi (int_size_in_bytes (TREE_TYPE (exp)));
1517 break;
1518 case MEM_REF:
1519 case POINTER_PLUS_EXPR:
1520 case MINUS_EXPR:
1521 case RANGE_EXPR:
1522 add_expr (TREE_OPERAND (exp, 0), hstate);
1523 add_expr (TREE_OPERAND (exp, 1), hstate);
1524 break;
1525 case PLUS_EXPR:
1527 inchash::hash one, two;
1528 add_expr (TREE_OPERAND (exp, 0), one);
1529 add_expr (TREE_OPERAND (exp, 1), two);
1530 hstate.add_commutative (one, two);
1532 break;
1533 CASE_CONVERT:
1534 hstate.add_hwi (int_size_in_bytes (TREE_TYPE (exp)));
1535 return add_expr (TREE_OPERAND (exp, 0), hstate);
1536 default:
1537 break;
1541 /* Accumulate to HSTATE a hash of type t.
1542 TYpes that may end up being compatible after LTO type merging needs to have
1543 the same hash. */
1545 void
1546 sem_item::add_type (const_tree type, inchash::hash &hstate)
1548 if (type == NULL_TREE)
1550 hstate.merge_hash (0);
1551 return;
1554 type = TYPE_MAIN_VARIANT (type);
1556 hstate.add_int (TYPE_MODE (type));
1558 if (TREE_CODE (type) == COMPLEX_TYPE)
1560 hstate.add_int (COMPLEX_TYPE);
1561 sem_item::add_type (TREE_TYPE (type), hstate);
1563 else if (INTEGRAL_TYPE_P (type))
1565 hstate.add_int (INTEGER_TYPE);
1566 hstate.add_flag (TYPE_UNSIGNED (type));
1567 hstate.add_int (TYPE_PRECISION (type));
1569 else if (VECTOR_TYPE_P (type))
1571 hstate.add_int (VECTOR_TYPE);
1572 hstate.add_int (TYPE_PRECISION (type));
1573 sem_item::add_type (TREE_TYPE (type), hstate);
1575 else if (TREE_CODE (type) == ARRAY_TYPE)
1577 hstate.add_int (ARRAY_TYPE);
1578 /* Do not hash size, so complete and incomplete types can match. */
1579 sem_item::add_type (TREE_TYPE (type), hstate);
1581 else if (RECORD_OR_UNION_TYPE_P (type))
1583 gcc_checking_assert (COMPLETE_TYPE_P (type));
1584 hashval_t *val = optimizer->m_type_hash_cache.get (type);
1586 if (!val)
1588 inchash::hash hstate2;
1589 unsigned nf;
1590 tree f;
1591 hashval_t hash;
1593 hstate2.add_int (RECORD_TYPE);
1594 gcc_assert (COMPLETE_TYPE_P (type));
1596 for (f = TYPE_FIELDS (type), nf = 0; f; f = TREE_CHAIN (f))
1597 if (TREE_CODE (f) == FIELD_DECL)
1599 add_type (TREE_TYPE (f), hstate2);
1600 nf++;
1603 hstate2.add_int (nf);
1604 hash = hstate2.end ();
1605 hstate.add_hwi (hash);
1606 optimizer->m_type_hash_cache.put (type, hash);
1608 else
1609 hstate.add_hwi (*val);
1613 /* Improve accumulated hash for HSTATE based on a gimple statement STMT. */
1615 void
1616 sem_function::hash_stmt (gimple *stmt, inchash::hash &hstate)
1618 enum gimple_code code = gimple_code (stmt);
1620 hstate.add_int (code);
1622 switch (code)
1624 case GIMPLE_SWITCH:
1625 add_expr (gimple_switch_index (as_a <gswitch *> (stmt)), hstate);
1626 break;
1627 case GIMPLE_ASSIGN:
1628 hstate.add_int (gimple_assign_rhs_code (stmt));
1629 if (commutative_tree_code (gimple_assign_rhs_code (stmt))
1630 || commutative_ternary_tree_code (gimple_assign_rhs_code (stmt)))
1632 inchash::hash one, two;
1634 add_expr (gimple_assign_rhs1 (stmt), one);
1635 add_type (TREE_TYPE (gimple_assign_rhs1 (stmt)), one);
1636 add_expr (gimple_assign_rhs2 (stmt), two);
1637 hstate.add_commutative (one, two);
1638 if (commutative_ternary_tree_code (gimple_assign_rhs_code (stmt)))
1640 add_expr (gimple_assign_rhs3 (stmt), hstate);
1641 add_type (TREE_TYPE (gimple_assign_rhs3 (stmt)), hstate);
1643 add_expr (gimple_assign_lhs (stmt), hstate);
1644 add_type (TREE_TYPE (gimple_assign_lhs (stmt)), two);
1645 break;
1647 /* fall through */
1648 case GIMPLE_CALL:
1649 case GIMPLE_ASM:
1650 case GIMPLE_COND:
1651 case GIMPLE_GOTO:
1652 case GIMPLE_RETURN:
1653 /* All these statements are equivalent if their operands are. */
1654 for (unsigned i = 0; i < gimple_num_ops (stmt); ++i)
1656 add_expr (gimple_op (stmt, i), hstate);
1657 if (gimple_op (stmt, i))
1658 add_type (TREE_TYPE (gimple_op (stmt, i)), hstate);
1660 /* Consider nocf_check attribute in hash as it affects code
1661 generation. */
1662 if (code == GIMPLE_CALL
1663 && flag_cf_protection & CF_BRANCH)
1664 hstate.add_flag (gimple_call_nocf_check_p (as_a <gcall *> (stmt)));
1665 default:
1666 break;
1671 /* Return true if polymorphic comparison must be processed. */
1673 bool
1674 sem_function::compare_polymorphic_p (void)
1676 struct cgraph_edge *e;
1678 if (!opt_for_fn (get_node ()->decl, flag_devirtualize))
1679 return false;
1680 if (get_node ()->indirect_calls != NULL)
1681 return true;
1682 /* TODO: We can do simple propagation determining what calls may lead to
1683 a polymorphic call. */
1684 for (e = get_node ()->callees; e; e = e->next_callee)
1685 if (e->callee->definition
1686 && opt_for_fn (e->callee->decl, flag_devirtualize))
1687 return true;
1688 return false;
1691 /* For a given call graph NODE, the function constructs new
1692 semantic function item. */
1694 sem_function *
1695 sem_function::parse (cgraph_node *node, bitmap_obstack *stack)
1697 tree fndecl = node->decl;
1698 function *func = DECL_STRUCT_FUNCTION (fndecl);
1700 if (!func || (!node->has_gimple_body_p () && !node->thunk.thunk_p))
1701 return NULL;
1703 if (lookup_attribute_by_prefix ("omp ", DECL_ATTRIBUTES (node->decl)) != NULL)
1704 return NULL;
1706 if (lookup_attribute_by_prefix ("oacc ",
1707 DECL_ATTRIBUTES (node->decl)) != NULL)
1708 return NULL;
1710 /* PR ipa/70306. */
1711 if (DECL_STATIC_CONSTRUCTOR (node->decl)
1712 || DECL_STATIC_DESTRUCTOR (node->decl))
1713 return NULL;
1715 sem_function *f = new sem_function (node, stack);
1717 f->init ();
1719 return f;
1722 /* For given basic blocks BB1 and BB2 (from functions FUNC1 and FUNC),
1723 return true if phi nodes are semantically equivalent in these blocks . */
1725 bool
1726 sem_function::compare_phi_node (basic_block bb1, basic_block bb2)
1728 gphi_iterator si1, si2;
1729 gphi *phi1, *phi2;
1730 unsigned size1, size2, i;
1731 tree t1, t2;
1732 edge e1, e2;
1734 gcc_assert (bb1 != NULL);
1735 gcc_assert (bb2 != NULL);
1737 si2 = gsi_start_phis (bb2);
1738 for (si1 = gsi_start_phis (bb1); !gsi_end_p (si1);
1739 gsi_next (&si1))
1741 gsi_next_nonvirtual_phi (&si1);
1742 gsi_next_nonvirtual_phi (&si2);
1744 if (gsi_end_p (si1) && gsi_end_p (si2))
1745 break;
1747 if (gsi_end_p (si1) || gsi_end_p (si2))
1748 return return_false();
1750 phi1 = si1.phi ();
1751 phi2 = si2.phi ();
1753 tree phi_result1 = gimple_phi_result (phi1);
1754 tree phi_result2 = gimple_phi_result (phi2);
1756 if (!m_checker->compare_operand (phi_result1, phi_result2))
1757 return return_false_with_msg ("PHI results are different");
1759 size1 = gimple_phi_num_args (phi1);
1760 size2 = gimple_phi_num_args (phi2);
1762 if (size1 != size2)
1763 return return_false ();
1765 for (i = 0; i < size1; ++i)
1767 t1 = gimple_phi_arg (phi1, i)->def;
1768 t2 = gimple_phi_arg (phi2, i)->def;
1770 if (!m_checker->compare_operand (t1, t2))
1771 return return_false ();
1773 e1 = gimple_phi_arg_edge (phi1, i);
1774 e2 = gimple_phi_arg_edge (phi2, i);
1776 if (!m_checker->compare_edge (e1, e2))
1777 return return_false ();
1780 gsi_next (&si2);
1783 return true;
1786 /* Returns true if tree T can be compared as a handled component. */
1788 bool
1789 sem_function::icf_handled_component_p (tree t)
1791 tree_code tc = TREE_CODE (t);
1793 return (handled_component_p (t)
1794 || tc == ADDR_EXPR || tc == MEM_REF || tc == OBJ_TYPE_REF);
1797 /* Basic blocks dictionary BB_DICT returns true if SOURCE index BB
1798 corresponds to TARGET. */
1800 bool
1801 sem_function::bb_dict_test (vec<int> *bb_dict, int source, int target)
1803 source++;
1804 target++;
1806 if (bb_dict->length () <= (unsigned)source)
1807 bb_dict->safe_grow_cleared (source + 1);
1809 if ((*bb_dict)[source] == 0)
1811 (*bb_dict)[source] = target;
1812 return true;
1814 else
1815 return (*bb_dict)[source] == target;
1818 sem_variable::sem_variable (bitmap_obstack *stack): sem_item (VAR, stack)
1822 sem_variable::sem_variable (varpool_node *node, bitmap_obstack *stack)
1823 : sem_item (VAR, node, stack)
1825 gcc_checking_assert (node);
1826 gcc_checking_assert (get_node ());
1829 /* Fast equality function based on knowledge known in WPA. */
1831 bool
1832 sem_variable::equals_wpa (sem_item *item,
1833 hash_map <symtab_node *, sem_item *> &ignored_nodes)
1835 gcc_assert (item->type == VAR);
1837 if (node->num_references () != item->node->num_references ())
1838 return return_false_with_msg ("different number of references");
1840 if (DECL_TLS_MODEL (decl) || DECL_TLS_MODEL (item->decl))
1841 return return_false_with_msg ("TLS model");
1843 /* DECL_ALIGN is safe to merge, because we will always chose the largest
1844 alignment out of all aliases. */
1846 if (DECL_VIRTUAL_P (decl) != DECL_VIRTUAL_P (item->decl))
1847 return return_false_with_msg ("Virtual flag mismatch");
1849 if (DECL_SIZE (decl) != DECL_SIZE (item->decl)
1850 && ((!DECL_SIZE (decl) || !DECL_SIZE (item->decl))
1851 || !operand_equal_p (DECL_SIZE (decl),
1852 DECL_SIZE (item->decl), OEP_ONLY_CONST)))
1853 return return_false_with_msg ("size mismatch");
1855 /* Do not attempt to mix data from different user sections;
1856 we do not know what user intends with those. */
1857 if (((DECL_SECTION_NAME (decl) && !node->implicit_section)
1858 || (DECL_SECTION_NAME (item->decl) && !item->node->implicit_section))
1859 && DECL_SECTION_NAME (decl) != DECL_SECTION_NAME (item->decl))
1860 return return_false_with_msg ("user section mismatch");
1862 if (DECL_IN_TEXT_SECTION (decl) != DECL_IN_TEXT_SECTION (item->decl))
1863 return return_false_with_msg ("text section");
1865 ipa_ref *ref = NULL, *ref2 = NULL;
1866 for (unsigned i = 0; node->iterate_reference (i, ref); i++)
1868 item->node->iterate_reference (i, ref2);
1870 if (ref->use != ref2->use)
1871 return return_false_with_msg ("reference use mismatch");
1873 if (!compare_symbol_references (ignored_nodes,
1874 ref->referred, ref2->referred,
1875 ref->address_matters_p ()))
1876 return false;
1879 return true;
1882 /* Returns true if the item equals to ITEM given as argument. */
1884 bool
1885 sem_variable::equals (sem_item *item,
1886 hash_map <symtab_node *, sem_item *> &)
1888 gcc_assert (item->type == VAR);
1889 bool ret;
1891 if (DECL_INITIAL (decl) == error_mark_node && in_lto_p)
1892 dyn_cast <varpool_node *>(node)->get_constructor ();
1893 if (DECL_INITIAL (item->decl) == error_mark_node && in_lto_p)
1894 dyn_cast <varpool_node *>(item->node)->get_constructor ();
1896 /* As seen in PR ipa/65303 we have to compare variables types. */
1897 if (!func_checker::compatible_types_p (TREE_TYPE (decl),
1898 TREE_TYPE (item->decl)))
1899 return return_false_with_msg ("variables types are different");
1901 ret = sem_variable::equals (DECL_INITIAL (decl),
1902 DECL_INITIAL (item->node->decl));
1903 if (dump_file && (dump_flags & TDF_DETAILS))
1904 fprintf (dump_file,
1905 "Equals called for vars: %s:%s with result: %s\n\n",
1906 node->dump_name (), item->node->dump_name (),
1907 ret ? "true" : "false");
1909 return ret;
1912 /* Compares trees T1 and T2 for semantic equality. */
1914 bool
1915 sem_variable::equals (tree t1, tree t2)
1917 if (!t1 || !t2)
1918 return return_with_debug (t1 == t2);
1919 if (t1 == t2)
1920 return true;
1921 tree_code tc1 = TREE_CODE (t1);
1922 tree_code tc2 = TREE_CODE (t2);
1924 if (tc1 != tc2)
1925 return return_false_with_msg ("TREE_CODE mismatch");
1927 switch (tc1)
1929 case CONSTRUCTOR:
1931 vec<constructor_elt, va_gc> *v1, *v2;
1932 unsigned HOST_WIDE_INT idx;
1934 enum tree_code typecode = TREE_CODE (TREE_TYPE (t1));
1935 if (typecode != TREE_CODE (TREE_TYPE (t2)))
1936 return return_false_with_msg ("constructor type mismatch");
1938 if (typecode == ARRAY_TYPE)
1940 HOST_WIDE_INT size_1 = int_size_in_bytes (TREE_TYPE (t1));
1941 /* For arrays, check that the sizes all match. */
1942 if (TYPE_MODE (TREE_TYPE (t1)) != TYPE_MODE (TREE_TYPE (t2))
1943 || size_1 == -1
1944 || size_1 != int_size_in_bytes (TREE_TYPE (t2)))
1945 return return_false_with_msg ("constructor array size mismatch");
1947 else if (!func_checker::compatible_types_p (TREE_TYPE (t1),
1948 TREE_TYPE (t2)))
1949 return return_false_with_msg ("constructor type incompatible");
1951 v1 = CONSTRUCTOR_ELTS (t1);
1952 v2 = CONSTRUCTOR_ELTS (t2);
1953 if (vec_safe_length (v1) != vec_safe_length (v2))
1954 return return_false_with_msg ("constructor number of elts mismatch");
1956 for (idx = 0; idx < vec_safe_length (v1); ++idx)
1958 constructor_elt *c1 = &(*v1)[idx];
1959 constructor_elt *c2 = &(*v2)[idx];
1961 /* Check that each value is the same... */
1962 if (!sem_variable::equals (c1->value, c2->value))
1963 return false;
1964 /* ... and that they apply to the same fields! */
1965 if (!sem_variable::equals (c1->index, c2->index))
1966 return false;
1968 return true;
1970 case MEM_REF:
1972 tree x1 = TREE_OPERAND (t1, 0);
1973 tree x2 = TREE_OPERAND (t2, 0);
1974 tree y1 = TREE_OPERAND (t1, 1);
1975 tree y2 = TREE_OPERAND (t2, 1);
1977 if (!func_checker::compatible_types_p (TREE_TYPE (x1), TREE_TYPE (x2)))
1978 return return_false ();
1980 /* Type of the offset on MEM_REF does not matter. */
1981 return return_with_debug (sem_variable::equals (x1, x2)
1982 && wi::to_offset (y1)
1983 == wi::to_offset (y2));
1985 case ADDR_EXPR:
1986 case FDESC_EXPR:
1988 tree op1 = TREE_OPERAND (t1, 0);
1989 tree op2 = TREE_OPERAND (t2, 0);
1990 return sem_variable::equals (op1, op2);
1992 /* References to other vars/decls are compared using ipa-ref. */
1993 case FUNCTION_DECL:
1994 case VAR_DECL:
1995 if (decl_in_symtab_p (t1) && decl_in_symtab_p (t2))
1996 return true;
1997 return return_false_with_msg ("Declaration mismatch");
1998 case CONST_DECL:
1999 /* TODO: We can check CONST_DECL by its DECL_INITIAL, but for that we
2000 need to process its VAR/FUNCTION references without relying on ipa-ref
2001 compare. */
2002 case FIELD_DECL:
2003 case LABEL_DECL:
2004 return return_false_with_msg ("Declaration mismatch");
2005 case INTEGER_CST:
2006 /* Integer constants are the same only if the same width of type. */
2007 if (TYPE_PRECISION (TREE_TYPE (t1)) != TYPE_PRECISION (TREE_TYPE (t2)))
2008 return return_false_with_msg ("INTEGER_CST precision mismatch");
2009 if (TYPE_MODE (TREE_TYPE (t1)) != TYPE_MODE (TREE_TYPE (t2)))
2010 return return_false_with_msg ("INTEGER_CST mode mismatch");
2011 return return_with_debug (tree_int_cst_equal (t1, t2));
2012 case STRING_CST:
2013 if (TYPE_MODE (TREE_TYPE (t1)) != TYPE_MODE (TREE_TYPE (t2)))
2014 return return_false_with_msg ("STRING_CST mode mismatch");
2015 if (TREE_STRING_LENGTH (t1) != TREE_STRING_LENGTH (t2))
2016 return return_false_with_msg ("STRING_CST length mismatch");
2017 if (memcmp (TREE_STRING_POINTER (t1), TREE_STRING_POINTER (t2),
2018 TREE_STRING_LENGTH (t1)))
2019 return return_false_with_msg ("STRING_CST mismatch");
2020 return true;
2021 case FIXED_CST:
2022 /* Fixed constants are the same only if the same width of type. */
2023 if (TYPE_PRECISION (TREE_TYPE (t1)) != TYPE_PRECISION (TREE_TYPE (t2)))
2024 return return_false_with_msg ("FIXED_CST precision mismatch");
2026 return return_with_debug (FIXED_VALUES_IDENTICAL (TREE_FIXED_CST (t1),
2027 TREE_FIXED_CST (t2)));
2028 case COMPLEX_CST:
2029 return (sem_variable::equals (TREE_REALPART (t1), TREE_REALPART (t2))
2030 && sem_variable::equals (TREE_IMAGPART (t1), TREE_IMAGPART (t2)));
2031 case REAL_CST:
2032 /* Real constants are the same only if the same width of type. */
2033 if (TYPE_PRECISION (TREE_TYPE (t1)) != TYPE_PRECISION (TREE_TYPE (t2)))
2034 return return_false_with_msg ("REAL_CST precision mismatch");
2035 return return_with_debug (real_identical (&TREE_REAL_CST (t1),
2036 &TREE_REAL_CST (t2)));
2037 case VECTOR_CST:
2039 if (maybe_ne (VECTOR_CST_NELTS (t1), VECTOR_CST_NELTS (t2)))
2040 return return_false_with_msg ("VECTOR_CST nelts mismatch");
2042 unsigned int count
2043 = tree_vector_builder::binary_encoded_nelts (t1, t2);
2044 for (unsigned int i = 0; i < count; ++i)
2045 if (!sem_variable::equals (VECTOR_CST_ENCODED_ELT (t1, i),
2046 VECTOR_CST_ENCODED_ELT (t2, i)))
2047 return false;
2049 return true;
2051 case ARRAY_REF:
2052 case ARRAY_RANGE_REF:
2054 tree x1 = TREE_OPERAND (t1, 0);
2055 tree x2 = TREE_OPERAND (t2, 0);
2056 tree y1 = TREE_OPERAND (t1, 1);
2057 tree y2 = TREE_OPERAND (t2, 1);
2059 if (!sem_variable::equals (x1, x2) || !sem_variable::equals (y1, y2))
2060 return false;
2061 if (!sem_variable::equals (array_ref_low_bound (t1),
2062 array_ref_low_bound (t2)))
2063 return false;
2064 if (!sem_variable::equals (array_ref_element_size (t1),
2065 array_ref_element_size (t2)))
2066 return false;
2067 return true;
2070 case COMPONENT_REF:
2071 case POINTER_PLUS_EXPR:
2072 case PLUS_EXPR:
2073 case MINUS_EXPR:
2074 case RANGE_EXPR:
2076 tree x1 = TREE_OPERAND (t1, 0);
2077 tree x2 = TREE_OPERAND (t2, 0);
2078 tree y1 = TREE_OPERAND (t1, 1);
2079 tree y2 = TREE_OPERAND (t2, 1);
2081 return sem_variable::equals (x1, x2) && sem_variable::equals (y1, y2);
2084 CASE_CONVERT:
2085 case VIEW_CONVERT_EXPR:
2086 if (!func_checker::compatible_types_p (TREE_TYPE (t1), TREE_TYPE (t2)))
2087 return return_false ();
2088 return sem_variable::equals (TREE_OPERAND (t1, 0), TREE_OPERAND (t2, 0));
2089 case ERROR_MARK:
2090 return return_false_with_msg ("ERROR_MARK");
2091 default:
2092 return return_false_with_msg ("Unknown TREE code reached");
2096 /* Parser function that visits a varpool NODE. */
2098 sem_variable *
2099 sem_variable::parse (varpool_node *node, bitmap_obstack *stack)
2101 if (TREE_THIS_VOLATILE (node->decl) || DECL_HARD_REGISTER (node->decl)
2102 || node->alias)
2103 return NULL;
2105 sem_variable *v = new sem_variable (node, stack);
2107 v->init ();
2109 return v;
2112 /* References independent hash function. */
2114 hashval_t
2115 sem_variable::get_hash (void)
2117 if (m_hash_set)
2118 return m_hash;
2120 /* All WPA streamed in symbols should have their hashes computed at compile
2121 time. At this point, the constructor may not be in memory at all.
2122 DECL_INITIAL (decl) would be error_mark_node in that case. */
2123 gcc_assert (!node->lto_file_data);
2124 tree ctor = DECL_INITIAL (decl);
2125 inchash::hash hstate;
2127 hstate.add_int (456346417);
2128 if (DECL_SIZE (decl) && tree_fits_shwi_p (DECL_SIZE (decl)))
2129 hstate.add_hwi (tree_to_shwi (DECL_SIZE (decl)));
2130 add_expr (ctor, hstate);
2131 set_hash (hstate.end ());
2133 return m_hash;
2136 /* Set all points-to UIDs of aliases pointing to node N as UID. */
2138 static void
2139 set_alias_uids (symtab_node *n, int uid)
2141 ipa_ref *ref;
2142 FOR_EACH_ALIAS (n, ref)
2144 if (dump_file)
2145 fprintf (dump_file, " Setting points-to UID of [%s] as %d\n",
2146 xstrdup_for_dump (ref->referring->asm_name ()), uid);
2148 SET_DECL_PT_UID (ref->referring->decl, uid);
2149 set_alias_uids (ref->referring, uid);
2153 /* Merges instance with an ALIAS_ITEM, where alias, thunk or redirection can
2154 be applied. */
2156 bool
2157 sem_variable::merge (sem_item *alias_item)
2159 gcc_assert (alias_item->type == VAR);
2161 if (!sem_item::target_supports_symbol_aliases_p ())
2163 if (dump_file)
2164 fprintf (dump_file, "Not unifying; "
2165 "Symbol aliases are not supported by target\n\n");
2166 return false;
2169 if (DECL_EXTERNAL (alias_item->decl))
2171 if (dump_file)
2172 fprintf (dump_file, "Not unifying; alias is external.\n\n");
2173 return false;
2176 sem_variable *alias_var = static_cast<sem_variable *> (alias_item);
2178 varpool_node *original = get_node ();
2179 varpool_node *alias = alias_var->get_node ();
2180 bool original_discardable = false;
2182 bool alias_address_matters = alias->address_matters_p ();
2184 /* See if original is in a section that can be discarded if the main
2185 symbol is not used.
2186 Also consider case where we have resolution info and we know that
2187 original's definition is not going to be used. In this case we can not
2188 create alias to original. */
2189 if (original->can_be_discarded_p ()
2190 || (node->resolution != LDPR_UNKNOWN
2191 && !decl_binds_to_current_def_p (node->decl)))
2192 original_discardable = true;
2194 gcc_assert (!TREE_ASM_WRITTEN (alias->decl));
2196 /* Constant pool machinery is not quite ready for aliases.
2197 TODO: varasm code contains logic for merging DECL_IN_CONSTANT_POOL.
2198 For LTO merging does not happen that is an important missing feature.
2199 We can enable merging with LTO if the DECL_IN_CONSTANT_POOL
2200 flag is dropped and non-local symbol name is assigned. */
2201 if (DECL_IN_CONSTANT_POOL (alias->decl)
2202 || DECL_IN_CONSTANT_POOL (original->decl))
2204 if (dump_file)
2205 fprintf (dump_file,
2206 "Not unifying; constant pool variables.\n\n");
2207 return false;
2210 /* Do not attempt to mix functions from different user sections;
2211 we do not know what user intends with those. */
2212 if (((DECL_SECTION_NAME (original->decl) && !original->implicit_section)
2213 || (DECL_SECTION_NAME (alias->decl) && !alias->implicit_section))
2214 && DECL_SECTION_NAME (original->decl) != DECL_SECTION_NAME (alias->decl))
2216 if (dump_file)
2217 fprintf (dump_file,
2218 "Not unifying; "
2219 "original and alias are in different sections.\n\n");
2220 return false;
2223 /* We can not merge if address comparsion metters. */
2224 if (alias_address_matters && flag_merge_constants < 2)
2226 if (dump_file)
2227 fprintf (dump_file,
2228 "Not unifying; address of original may be compared.\n\n");
2229 return false;
2232 if (DECL_ALIGN (original->decl) < DECL_ALIGN (alias->decl))
2234 if (dump_file)
2235 fprintf (dump_file, "Not unifying; "
2236 "original and alias have incompatible alignments\n\n");
2238 return false;
2241 if (DECL_COMDAT_GROUP (original->decl) != DECL_COMDAT_GROUP (alias->decl))
2243 if (dump_file)
2244 fprintf (dump_file, "Not unifying; alias cannot be created; "
2245 "across comdat group boundary\n\n");
2247 return false;
2250 if (original_discardable)
2252 if (dump_file)
2253 fprintf (dump_file, "Not unifying; alias cannot be created; "
2254 "target is discardable\n\n");
2256 return false;
2258 else
2260 gcc_assert (!original->alias);
2261 gcc_assert (!alias->alias);
2263 alias->analyzed = false;
2265 DECL_INITIAL (alias->decl) = NULL;
2266 ((symtab_node *)alias)->call_for_symbol_and_aliases (clear_decl_rtl,
2267 NULL, true);
2268 alias->need_bounds_init = false;
2269 alias->remove_all_references ();
2270 if (TREE_ADDRESSABLE (alias->decl))
2271 original->call_for_symbol_and_aliases (set_addressable, NULL, true);
2273 varpool_node::create_alias (alias_var->decl, decl);
2274 alias->resolve_alias (original);
2276 if (dump_file)
2277 fprintf (dump_file, "Unified; Variable alias has been created.\n");
2279 set_alias_uids (original, DECL_UID (original->decl));
2280 return true;
2284 /* Dump symbol to FILE. */
2286 void
2287 sem_variable::dump_to_file (FILE *file)
2289 gcc_assert (file);
2291 print_node (file, "", decl, 0);
2292 fprintf (file, "\n\n");
2295 unsigned int sem_item_optimizer::class_id = 0;
2297 sem_item_optimizer::sem_item_optimizer ()
2298 : worklist (0), m_classes (0), m_classes_count (0), m_cgraph_node_hooks (NULL),
2299 m_varpool_node_hooks (NULL)
2301 m_items.create (0);
2302 bitmap_obstack_initialize (&m_bmstack);
2305 sem_item_optimizer::~sem_item_optimizer ()
2307 for (unsigned int i = 0; i < m_items.length (); i++)
2308 delete m_items[i];
2311 for (hash_table<congruence_class_hash>::iterator it = m_classes.begin ();
2312 it != m_classes.end (); ++it)
2314 for (unsigned int i = 0; i < (*it)->classes.length (); i++)
2315 delete (*it)->classes[i];
2317 (*it)->classes.release ();
2318 free (*it);
2321 m_items.release ();
2323 bitmap_obstack_release (&m_bmstack);
2326 /* Write IPA ICF summary for symbols. */
2328 void
2329 sem_item_optimizer::write_summary (void)
2331 unsigned int count = 0;
2333 output_block *ob = create_output_block (LTO_section_ipa_icf);
2334 lto_symtab_encoder_t encoder = ob->decl_state->symtab_node_encoder;
2335 ob->symbol = NULL;
2337 /* Calculate number of symbols to be serialized. */
2338 for (lto_symtab_encoder_iterator lsei = lsei_start_in_partition (encoder);
2339 !lsei_end_p (lsei);
2340 lsei_next_in_partition (&lsei))
2342 symtab_node *node = lsei_node (lsei);
2344 if (m_symtab_node_map.get (node))
2345 count++;
2348 streamer_write_uhwi (ob, count);
2350 /* Process all of the symbols. */
2351 for (lto_symtab_encoder_iterator lsei = lsei_start_in_partition (encoder);
2352 !lsei_end_p (lsei);
2353 lsei_next_in_partition (&lsei))
2355 symtab_node *node = lsei_node (lsei);
2357 sem_item **item = m_symtab_node_map.get (node);
2359 if (item && *item)
2361 int node_ref = lto_symtab_encoder_encode (encoder, node);
2362 streamer_write_uhwi_stream (ob->main_stream, node_ref);
2364 streamer_write_uhwi (ob, (*item)->get_hash ());
2368 streamer_write_char_stream (ob->main_stream, 0);
2369 produce_asm (ob, NULL);
2370 destroy_output_block (ob);
2373 /* Reads a section from LTO stream file FILE_DATA. Input block for DATA
2374 contains LEN bytes. */
2376 void
2377 sem_item_optimizer::read_section (lto_file_decl_data *file_data,
2378 const char *data, size_t len)
2380 const lto_function_header *header
2381 = (const lto_function_header *) data;
2382 const int cfg_offset = sizeof (lto_function_header);
2383 const int main_offset = cfg_offset + header->cfg_size;
2384 const int string_offset = main_offset + header->main_size;
2385 data_in *data_in;
2386 unsigned int i;
2387 unsigned int count;
2389 lto_input_block ib_main ((const char *) data + main_offset, 0,
2390 header->main_size, file_data->mode_table);
2392 data_in
2393 = lto_data_in_create (file_data, (const char *) data + string_offset,
2394 header->string_size, vNULL);
2396 count = streamer_read_uhwi (&ib_main);
2398 for (i = 0; i < count; i++)
2400 unsigned int index;
2401 symtab_node *node;
2402 lto_symtab_encoder_t encoder;
2404 index = streamer_read_uhwi (&ib_main);
2405 encoder = file_data->symtab_node_encoder;
2406 node = lto_symtab_encoder_deref (encoder, index);
2408 hashval_t hash = streamer_read_uhwi (&ib_main);
2410 gcc_assert (node->definition);
2412 if (dump_file)
2413 fprintf (dump_file, "Symbol added: %s (tree: %p)\n",
2414 node->dump_asm_name (), (void *) node->decl);
2416 if (is_a<cgraph_node *> (node))
2418 cgraph_node *cnode = dyn_cast <cgraph_node *> (node);
2420 sem_function *fn = new sem_function (cnode, &m_bmstack);
2421 fn->set_hash (hash);
2422 m_items.safe_push (fn);
2424 else
2426 varpool_node *vnode = dyn_cast <varpool_node *> (node);
2428 sem_variable *var = new sem_variable (vnode, &m_bmstack);
2429 var->set_hash (hash);
2430 m_items.safe_push (var);
2434 lto_free_section_data (file_data, LTO_section_ipa_icf, NULL, data,
2435 len);
2436 lto_data_in_delete (data_in);
2439 /* Read IPA ICF summary for symbols. */
2441 void
2442 sem_item_optimizer::read_summary (void)
2444 lto_file_decl_data **file_data_vec = lto_get_file_decl_data ();
2445 lto_file_decl_data *file_data;
2446 unsigned int j = 0;
2448 while ((file_data = file_data_vec[j++]))
2450 size_t len;
2451 const char *data = lto_get_section_data (file_data,
2452 LTO_section_ipa_icf, NULL, &len);
2454 if (data)
2455 read_section (file_data, data, len);
2459 /* Register callgraph and varpool hooks. */
2461 void
2462 sem_item_optimizer::register_hooks (void)
2464 if (!m_cgraph_node_hooks)
2465 m_cgraph_node_hooks = symtab->add_cgraph_removal_hook
2466 (&sem_item_optimizer::cgraph_removal_hook, this);
2468 if (!m_varpool_node_hooks)
2469 m_varpool_node_hooks = symtab->add_varpool_removal_hook
2470 (&sem_item_optimizer::varpool_removal_hook, this);
2473 /* Unregister callgraph and varpool hooks. */
2475 void
2476 sem_item_optimizer::unregister_hooks (void)
2478 if (m_cgraph_node_hooks)
2479 symtab->remove_cgraph_removal_hook (m_cgraph_node_hooks);
2481 if (m_varpool_node_hooks)
2482 symtab->remove_varpool_removal_hook (m_varpool_node_hooks);
2485 /* Adds a CLS to hashtable associated by hash value. */
2487 void
2488 sem_item_optimizer::add_class (congruence_class *cls)
2490 gcc_assert (cls->members.length ());
2492 congruence_class_group *group
2493 = get_group_by_hash (cls->members[0]->get_hash (),
2494 cls->members[0]->type);
2495 group->classes.safe_push (cls);
2498 /* Gets a congruence class group based on given HASH value and TYPE. */
2500 congruence_class_group *
2501 sem_item_optimizer::get_group_by_hash (hashval_t hash, sem_item_type type)
2503 congruence_class_group *item = XNEW (congruence_class_group);
2504 item->hash = hash;
2505 item->type = type;
2507 congruence_class_group **slot = m_classes.find_slot (item, INSERT);
2509 if (*slot)
2510 free (item);
2511 else
2513 item->classes.create (1);
2514 *slot = item;
2517 return *slot;
2520 /* Callgraph removal hook called for a NODE with a custom DATA. */
2522 void
2523 sem_item_optimizer::cgraph_removal_hook (cgraph_node *node, void *data)
2525 sem_item_optimizer *optimizer = (sem_item_optimizer *) data;
2526 optimizer->remove_symtab_node (node);
2529 /* Varpool removal hook called for a NODE with a custom DATA. */
2531 void
2532 sem_item_optimizer::varpool_removal_hook (varpool_node *node, void *data)
2534 sem_item_optimizer *optimizer = (sem_item_optimizer *) data;
2535 optimizer->remove_symtab_node (node);
2538 /* Remove symtab NODE triggered by symtab removal hooks. */
2540 void
2541 sem_item_optimizer::remove_symtab_node (symtab_node *node)
2543 gcc_assert (!m_classes.elements ());
2545 m_removed_items_set.add (node);
2548 void
2549 sem_item_optimizer::remove_item (sem_item *item)
2551 if (m_symtab_node_map.get (item->node))
2552 m_symtab_node_map.remove (item->node);
2553 delete item;
2556 /* Removes all callgraph and varpool nodes that are marked by symtab
2557 as deleted. */
2559 void
2560 sem_item_optimizer::filter_removed_items (void)
2562 auto_vec <sem_item *> filtered;
2564 for (unsigned int i = 0; i < m_items.length(); i++)
2566 sem_item *item = m_items[i];
2568 if (m_removed_items_set.contains (item->node))
2570 remove_item (item);
2571 continue;
2574 if (item->type == FUNC)
2576 cgraph_node *cnode = static_cast <sem_function *>(item)->get_node ();
2578 if (in_lto_p && (cnode->alias || cnode->body_removed))
2579 remove_item (item);
2580 else
2581 filtered.safe_push (item);
2583 else /* VAR. */
2585 if (!flag_ipa_icf_variables)
2586 remove_item (item);
2587 else
2589 /* Filter out non-readonly variables. */
2590 tree decl = item->decl;
2591 if (TREE_READONLY (decl))
2592 filtered.safe_push (item);
2593 else
2594 remove_item (item);
2599 /* Clean-up of released semantic items. */
2601 m_items.release ();
2602 for (unsigned int i = 0; i < filtered.length(); i++)
2603 m_items.safe_push (filtered[i]);
2606 /* Optimizer entry point which returns true in case it processes
2607 a merge operation. True is returned if there's a merge operation
2608 processed. */
2610 bool
2611 sem_item_optimizer::execute (void)
2613 filter_removed_items ();
2614 unregister_hooks ();
2616 build_graph ();
2617 update_hash_by_addr_refs ();
2618 build_hash_based_classes ();
2620 if (dump_file)
2621 fprintf (dump_file, "Dump after hash based groups\n");
2622 dump_cong_classes ();
2624 for (unsigned int i = 0; i < m_items.length(); i++)
2625 m_items[i]->init_wpa ();
2627 subdivide_classes_by_equality (true);
2629 if (dump_file)
2630 fprintf (dump_file, "Dump after WPA based types groups\n");
2632 dump_cong_classes ();
2634 process_cong_reduction ();
2635 checking_verify_classes ();
2637 if (dump_file)
2638 fprintf (dump_file, "Dump after callgraph-based congruence reduction\n");
2640 dump_cong_classes ();
2642 parse_nonsingleton_classes ();
2643 subdivide_classes_by_equality ();
2645 if (dump_file)
2646 fprintf (dump_file, "Dump after full equality comparison of groups\n");
2648 dump_cong_classes ();
2650 unsigned int prev_class_count = m_classes_count;
2652 process_cong_reduction ();
2653 dump_cong_classes ();
2654 checking_verify_classes ();
2655 bool merged_p = merge_classes (prev_class_count);
2657 if (dump_file && (dump_flags & TDF_DETAILS))
2658 symtab->dump (dump_file);
2660 return merged_p;
2663 /* Function responsible for visiting all potential functions and
2664 read-only variables that can be merged. */
2666 void
2667 sem_item_optimizer::parse_funcs_and_vars (void)
2669 cgraph_node *cnode;
2671 if (flag_ipa_icf_functions)
2672 FOR_EACH_DEFINED_FUNCTION (cnode)
2674 sem_function *f = sem_function::parse (cnode, &m_bmstack);
2675 if (f)
2677 m_items.safe_push (f);
2678 m_symtab_node_map.put (cnode, f);
2680 if (dump_file)
2681 fprintf (dump_file, "Parsed function:%s\n", f->node->asm_name ());
2683 if (dump_file && (dump_flags & TDF_DETAILS))
2684 f->dump_to_file (dump_file);
2686 else if (dump_file)
2687 fprintf (dump_file, "Not parsed function:%s\n", cnode->asm_name ());
2690 varpool_node *vnode;
2692 if (flag_ipa_icf_variables)
2693 FOR_EACH_DEFINED_VARIABLE (vnode)
2695 sem_variable *v = sem_variable::parse (vnode, &m_bmstack);
2697 if (v)
2699 m_items.safe_push (v);
2700 m_symtab_node_map.put (vnode, v);
2705 /* Makes pairing between a congruence class CLS and semantic ITEM. */
2707 void
2708 sem_item_optimizer::add_item_to_class (congruence_class *cls, sem_item *item)
2710 item->index_in_class = cls->members.length ();
2711 cls->members.safe_push (item);
2712 item->cls = cls;
2715 /* For each semantic item, append hash values of references. */
2717 void
2718 sem_item_optimizer::update_hash_by_addr_refs ()
2720 /* First, append to hash sensitive references and class type if it need to
2721 be matched for ODR. */
2722 for (unsigned i = 0; i < m_items.length (); i++)
2724 m_items[i]->update_hash_by_addr_refs (m_symtab_node_map);
2725 if (m_items[i]->type == FUNC)
2727 if (TREE_CODE (TREE_TYPE (m_items[i]->decl)) == METHOD_TYPE
2728 && contains_polymorphic_type_p
2729 (TYPE_METHOD_BASETYPE (TREE_TYPE (m_items[i]->decl)))
2730 && (DECL_CXX_CONSTRUCTOR_P (m_items[i]->decl)
2731 || (static_cast<sem_function *> (m_items[i])->param_used_p (0)
2732 && static_cast<sem_function *> (m_items[i])
2733 ->compare_polymorphic_p ())))
2735 tree class_type
2736 = TYPE_METHOD_BASETYPE (TREE_TYPE (m_items[i]->decl));
2737 inchash::hash hstate (m_items[i]->get_hash ());
2739 if (TYPE_NAME (class_type)
2740 && DECL_ASSEMBLER_NAME_SET_P (TYPE_NAME (class_type)))
2741 hstate.add_hwi
2742 (IDENTIFIER_HASH_VALUE
2743 (DECL_ASSEMBLER_NAME (TYPE_NAME (class_type))));
2745 m_items[i]->set_hash (hstate.end ());
2750 /* Once all symbols have enhanced hash value, we can append
2751 hash values of symbols that are seen by IPA ICF and are
2752 references by a semantic item. Newly computed values
2753 are saved to global_hash member variable. */
2754 for (unsigned i = 0; i < m_items.length (); i++)
2755 m_items[i]->update_hash_by_local_refs (m_symtab_node_map);
2757 /* Global hash value replace current hash values. */
2758 for (unsigned i = 0; i < m_items.length (); i++)
2759 m_items[i]->set_hash (m_items[i]->global_hash);
2762 /* Congruence classes are built by hash value. */
2764 void
2765 sem_item_optimizer::build_hash_based_classes (void)
2767 for (unsigned i = 0; i < m_items.length (); i++)
2769 sem_item *item = m_items[i];
2771 congruence_class_group *group
2772 = get_group_by_hash (item->get_hash (), item->type);
2774 if (!group->classes.length ())
2776 m_classes_count++;
2777 group->classes.safe_push (new congruence_class (class_id++));
2780 add_item_to_class (group->classes[0], item);
2784 /* Build references according to call graph. */
2786 void
2787 sem_item_optimizer::build_graph (void)
2789 for (unsigned i = 0; i < m_items.length (); i++)
2791 sem_item *item = m_items[i];
2792 m_symtab_node_map.put (item->node, item);
2794 /* Initialize hash values if we are not in LTO mode. */
2795 if (!in_lto_p)
2796 item->get_hash ();
2799 for (unsigned i = 0; i < m_items.length (); i++)
2801 sem_item *item = m_items[i];
2803 if (item->type == FUNC)
2805 cgraph_node *cnode = dyn_cast <cgraph_node *> (item->node);
2807 cgraph_edge *e = cnode->callees;
2808 while (e)
2810 sem_item **slot = m_symtab_node_map.get
2811 (e->callee->ultimate_alias_target ());
2812 if (slot)
2813 item->add_reference (*slot);
2815 e = e->next_callee;
2819 ipa_ref *ref = NULL;
2820 for (unsigned i = 0; item->node->iterate_reference (i, ref); i++)
2822 sem_item **slot = m_symtab_node_map.get
2823 (ref->referred->ultimate_alias_target ());
2824 if (slot)
2825 item->add_reference (*slot);
2830 /* Semantic items in classes having more than one element and initialized.
2831 In case of WPA, we load function body. */
2833 void
2834 sem_item_optimizer::parse_nonsingleton_classes (void)
2836 unsigned int init_called_count = 0;
2838 for (unsigned i = 0; i < m_items.length (); i++)
2839 if (m_items[i]->cls->members.length () > 1)
2841 m_items[i]->init ();
2842 init_called_count++;
2845 if (dump_file)
2846 fprintf (dump_file, "Init called for %u items (%.2f%%).\n",
2847 init_called_count,
2848 m_items.length () ? 100.0f * init_called_count / m_items.length ()
2849 : 0.0f);
2852 /* Equality function for semantic items is used to subdivide existing
2853 classes. If IN_WPA, fast equality function is invoked. */
2855 void
2856 sem_item_optimizer::subdivide_classes_by_equality (bool in_wpa)
2858 for (hash_table <congruence_class_hash>::iterator it = m_classes.begin ();
2859 it != m_classes.end (); ++it)
2861 unsigned int class_count = (*it)->classes.length ();
2863 for (unsigned i = 0; i < class_count; i++)
2865 congruence_class *c = (*it)->classes[i];
2867 if (c->members.length() > 1)
2869 auto_vec <sem_item *> new_vector;
2871 sem_item *first = c->members[0];
2872 new_vector.safe_push (first);
2874 unsigned class_split_first = (*it)->classes.length ();
2876 for (unsigned j = 1; j < c->members.length (); j++)
2878 sem_item *item = c->members[j];
2880 bool equals
2881 = in_wpa ? first->equals_wpa (item, m_symtab_node_map)
2882 : first->equals (item, m_symtab_node_map);
2884 if (equals)
2885 new_vector.safe_push (item);
2886 else
2888 bool integrated = false;
2890 for (unsigned k = class_split_first;
2891 k < (*it)->classes.length (); k++)
2893 sem_item *x = (*it)->classes[k]->members[0];
2894 bool equals
2895 = in_wpa ? x->equals_wpa (item, m_symtab_node_map)
2896 : x->equals (item, m_symtab_node_map);
2898 if (equals)
2900 integrated = true;
2901 add_item_to_class ((*it)->classes[k], item);
2903 break;
2907 if (!integrated)
2909 congruence_class *c
2910 = new congruence_class (class_id++);
2911 m_classes_count++;
2912 add_item_to_class (c, item);
2914 (*it)->classes.safe_push (c);
2919 // We replace newly created new_vector for the class we've just
2920 // splitted.
2921 c->members.release ();
2922 c->members.create (new_vector.length ());
2924 for (unsigned int j = 0; j < new_vector.length (); j++)
2925 add_item_to_class (c, new_vector[j]);
2930 checking_verify_classes ();
2933 /* Subdivide classes by address references that members of the class
2934 reference. Example can be a pair of functions that have an address
2935 taken from a function. If these addresses are different the class
2936 is split. */
2938 unsigned
2939 sem_item_optimizer::subdivide_classes_by_sensitive_refs ()
2941 typedef hash_map <symbol_compare_hash, vec <sem_item *> > subdivide_hash_map;
2943 unsigned newly_created_classes = 0;
2945 for (hash_table <congruence_class_hash>::iterator it = m_classes.begin ();
2946 it != m_classes.end (); ++it)
2948 unsigned int class_count = (*it)->classes.length ();
2949 auto_vec<congruence_class *> new_classes;
2951 for (unsigned i = 0; i < class_count; i++)
2953 congruence_class *c = (*it)->classes[i];
2955 if (c->members.length() > 1)
2957 subdivide_hash_map split_map;
2959 for (unsigned j = 0; j < c->members.length (); j++)
2961 sem_item *source_node = c->members[j];
2963 symbol_compare_collection *collection
2964 = new symbol_compare_collection (source_node->node);
2966 bool existed;
2967 vec <sem_item *> *slot
2968 = &split_map.get_or_insert (collection, &existed);
2969 gcc_checking_assert (slot);
2971 slot->safe_push (source_node);
2973 if (existed)
2974 delete collection;
2977 /* If the map contains more than one key, we have to split
2978 the map appropriately. */
2979 if (split_map.elements () != 1)
2981 bool first_class = true;
2983 for (subdivide_hash_map::iterator it2 = split_map.begin ();
2984 it2 != split_map.end (); ++it2)
2986 congruence_class *new_cls;
2987 new_cls = new congruence_class (class_id++);
2989 for (unsigned k = 0; k < (*it2).second.length (); k++)
2990 add_item_to_class (new_cls, (*it2).second[k]);
2992 worklist_push (new_cls);
2993 newly_created_classes++;
2995 if (first_class)
2997 (*it)->classes[i] = new_cls;
2998 first_class = false;
3000 else
3002 new_classes.safe_push (new_cls);
3003 m_classes_count++;
3008 /* Release memory. */
3009 for (subdivide_hash_map::iterator it2 = split_map.begin ();
3010 it2 != split_map.end (); ++it2)
3012 delete (*it2).first;
3013 (*it2).second.release ();
3018 for (unsigned i = 0; i < new_classes.length (); i++)
3019 (*it)->classes.safe_push (new_classes[i]);
3022 return newly_created_classes;
3025 /* Verify congruence classes, if checking is enabled. */
3027 void
3028 sem_item_optimizer::checking_verify_classes (void)
3030 if (flag_checking)
3031 verify_classes ();
3034 /* Verify congruence classes. */
3036 void
3037 sem_item_optimizer::verify_classes (void)
3039 for (hash_table<congruence_class_hash>::iterator it = m_classes.begin ();
3040 it != m_classes.end (); ++it)
3042 for (unsigned int i = 0; i < (*it)->classes.length (); i++)
3044 congruence_class *cls = (*it)->classes[i];
3046 gcc_assert (cls);
3047 gcc_assert (cls->members.length () > 0);
3049 for (unsigned int j = 0; j < cls->members.length (); j++)
3051 sem_item *item = cls->members[j];
3053 gcc_assert (item);
3054 gcc_assert (item->cls == cls);
3056 for (unsigned k = 0; k < item->usages.length (); k++)
3058 sem_usage_pair *usage = item->usages[k];
3059 gcc_assert (usage->item->index_in_class
3060 < usage->item->cls->members.length ());
3067 /* Disposes split map traverse function. CLS_PTR is pointer to congruence
3068 class, BSLOT is bitmap slot we want to release. DATA is mandatory,
3069 but unused argument. */
3071 bool
3072 sem_item_optimizer::release_split_map (congruence_class * const &,
3073 bitmap const &b, traverse_split_pair *)
3075 bitmap bmp = b;
3077 BITMAP_FREE (bmp);
3079 return true;
3082 /* Process split operation for a class given as pointer CLS_PTR,
3083 where bitmap B splits congruence class members. DATA is used
3084 as argument of split pair. */
3086 bool
3087 sem_item_optimizer::traverse_congruence_split (congruence_class * const &cls,
3088 bitmap const &b,
3089 traverse_split_pair *pair)
3091 sem_item_optimizer *optimizer = pair->optimizer;
3092 const congruence_class *splitter_cls = pair->cls;
3094 /* If counted bits are greater than zero and less than the number of members
3095 a group will be splitted. */
3096 unsigned popcount = bitmap_count_bits (b);
3098 if (popcount > 0 && popcount < cls->members.length ())
3100 auto_vec <congruence_class *, 2> newclasses;
3101 newclasses.quick_push (new congruence_class (class_id++));
3102 newclasses.quick_push (new congruence_class (class_id++));
3104 for (unsigned int i = 0; i < cls->members.length (); i++)
3106 int target = bitmap_bit_p (b, i);
3107 congruence_class *tc = newclasses[target];
3109 add_item_to_class (tc, cls->members[i]);
3112 if (flag_checking)
3114 for (unsigned int i = 0; i < 2; i++)
3115 gcc_assert (newclasses[i]->members.length ());
3118 if (splitter_cls == cls)
3119 optimizer->splitter_class_removed = true;
3121 /* Remove old class from worklist if presented. */
3122 bool in_worklist = cls->in_worklist;
3124 if (in_worklist)
3125 cls->in_worklist = false;
3127 congruence_class_group g;
3128 g.hash = cls->members[0]->get_hash ();
3129 g.type = cls->members[0]->type;
3131 congruence_class_group *slot = optimizer->m_classes.find (&g);
3133 for (unsigned int i = 0; i < slot->classes.length (); i++)
3134 if (slot->classes[i] == cls)
3136 slot->classes.ordered_remove (i);
3137 break;
3140 /* New class will be inserted and integrated to work list. */
3141 for (unsigned int i = 0; i < 2; i++)
3142 optimizer->add_class (newclasses[i]);
3144 /* Two classes replace one, so that increment just by one. */
3145 optimizer->m_classes_count++;
3147 /* If OLD class was presented in the worklist, we remove the class
3148 and replace it will both newly created classes. */
3149 if (in_worklist)
3150 for (unsigned int i = 0; i < 2; i++)
3151 optimizer->worklist_push (newclasses[i]);
3152 else /* Just smaller class is inserted. */
3154 unsigned int smaller_index
3155 = (newclasses[0]->members.length ()
3156 < newclasses[1]->members.length ()
3157 ? 0 : 1);
3158 optimizer->worklist_push (newclasses[smaller_index]);
3161 if (dump_file && (dump_flags & TDF_DETAILS))
3163 fprintf (dump_file, " congruence class splitted:\n");
3164 cls->dump (dump_file, 4);
3166 fprintf (dump_file, " newly created groups:\n");
3167 for (unsigned int i = 0; i < 2; i++)
3168 newclasses[i]->dump (dump_file, 4);
3171 /* Release class if not presented in work list. */
3172 if (!in_worklist)
3173 delete cls;
3177 return true;
3180 /* Tests if a class CLS used as INDEXth splits any congruence classes.
3181 Bitmap stack BMSTACK is used for bitmap allocation. */
3183 void
3184 sem_item_optimizer::do_congruence_step_for_index (congruence_class *cls,
3185 unsigned int index)
3187 hash_map <congruence_class *, bitmap> split_map;
3189 for (unsigned int i = 0; i < cls->members.length (); i++)
3191 sem_item *item = cls->members[i];
3193 /* Iterate all usages that have INDEX as usage of the item. */
3194 for (unsigned int j = 0; j < item->usages.length (); j++)
3196 sem_usage_pair *usage = item->usages[j];
3198 if (usage->index != index)
3199 continue;
3201 bitmap *slot = split_map.get (usage->item->cls);
3202 bitmap b;
3204 if(!slot)
3206 b = BITMAP_ALLOC (&m_bmstack);
3207 split_map.put (usage->item->cls, b);
3209 else
3210 b = *slot;
3212 gcc_checking_assert (usage->item->cls);
3213 gcc_checking_assert (usage->item->index_in_class
3214 < usage->item->cls->members.length ());
3216 bitmap_set_bit (b, usage->item->index_in_class);
3220 traverse_split_pair pair;
3221 pair.optimizer = this;
3222 pair.cls = cls;
3224 splitter_class_removed = false;
3225 split_map.traverse <traverse_split_pair *,
3226 sem_item_optimizer::traverse_congruence_split> (&pair);
3228 /* Bitmap clean-up. */
3229 split_map.traverse <traverse_split_pair *,
3230 sem_item_optimizer::release_split_map> (NULL);
3233 /* Every usage of a congruence class CLS is a candidate that can split the
3234 collection of classes. Bitmap stack BMSTACK is used for bitmap
3235 allocation. */
3237 void
3238 sem_item_optimizer::do_congruence_step (congruence_class *cls)
3240 bitmap_iterator bi;
3241 unsigned int i;
3243 bitmap usage = BITMAP_ALLOC (&m_bmstack);
3245 for (unsigned int i = 0; i < cls->members.length (); i++)
3246 bitmap_ior_into (usage, cls->members[i]->usage_index_bitmap);
3248 EXECUTE_IF_SET_IN_BITMAP (usage, 0, i, bi)
3250 if (dump_file && (dump_flags & TDF_DETAILS))
3251 fprintf (dump_file, " processing congruence step for class: %u, "
3252 "index: %u\n", cls->id, i);
3254 do_congruence_step_for_index (cls, i);
3256 if (splitter_class_removed)
3257 break;
3260 BITMAP_FREE (usage);
3263 /* Adds a newly created congruence class CLS to worklist. */
3265 void
3266 sem_item_optimizer::worklist_push (congruence_class *cls)
3268 /* Return if the class CLS is already presented in work list. */
3269 if (cls->in_worklist)
3270 return;
3272 cls->in_worklist = true;
3273 worklist.push_back (cls);
3276 /* Pops a class from worklist. */
3278 congruence_class *
3279 sem_item_optimizer::worklist_pop (void)
3281 congruence_class *cls;
3283 while (!worklist.empty ())
3285 cls = worklist.front ();
3286 worklist.pop_front ();
3287 if (cls->in_worklist)
3289 cls->in_worklist = false;
3291 return cls;
3293 else
3295 /* Work list item was already intended to be removed.
3296 The only reason for doing it is to split a class.
3297 Thus, the class CLS is deleted. */
3298 delete cls;
3302 return NULL;
3305 /* Iterative congruence reduction function. */
3307 void
3308 sem_item_optimizer::process_cong_reduction (void)
3310 for (hash_table<congruence_class_hash>::iterator it = m_classes.begin ();
3311 it != m_classes.end (); ++it)
3312 for (unsigned i = 0; i < (*it)->classes.length (); i++)
3313 if ((*it)->classes[i]->is_class_used ())
3314 worklist_push ((*it)->classes[i]);
3316 if (dump_file)
3317 fprintf (dump_file, "Worklist has been filled with: %lu\n",
3318 (unsigned long) worklist.size ());
3320 if (dump_file && (dump_flags & TDF_DETAILS))
3321 fprintf (dump_file, "Congruence class reduction\n");
3323 congruence_class *cls;
3325 /* Process complete congruence reduction. */
3326 while ((cls = worklist_pop ()) != NULL)
3327 do_congruence_step (cls);
3329 /* Subdivide newly created classes according to references. */
3330 unsigned new_classes = subdivide_classes_by_sensitive_refs ();
3332 if (dump_file)
3333 fprintf (dump_file, "Address reference subdivision created: %u "
3334 "new classes.\n", new_classes);
3337 /* Debug function prints all informations about congruence classes. */
3339 void
3340 sem_item_optimizer::dump_cong_classes (void)
3342 if (!dump_file)
3343 return;
3345 fprintf (dump_file,
3346 "Congruence classes: %u (unique hash values: %lu), with total: "
3347 "%u items\n", m_classes_count,
3348 (unsigned long) m_classes.elements (), m_items.length ());
3350 /* Histogram calculation. */
3351 unsigned int max_index = 0;
3352 unsigned int* histogram = XCNEWVEC (unsigned int, m_items.length () + 1);
3354 for (hash_table<congruence_class_hash>::iterator it = m_classes.begin ();
3355 it != m_classes.end (); ++it)
3356 for (unsigned i = 0; i < (*it)->classes.length (); i++)
3358 unsigned int c = (*it)->classes[i]->members.length ();
3359 histogram[c]++;
3361 if (c > max_index)
3362 max_index = c;
3365 fprintf (dump_file,
3366 "Class size histogram [num of members]: number of classe number "
3367 "of classess\n");
3369 for (unsigned int i = 0; i <= max_index; i++)
3370 if (histogram[i])
3371 fprintf (dump_file, "[%u]: %u classes\n", i, histogram[i]);
3373 fprintf (dump_file, "\n\n");
3375 if (dump_flags & TDF_DETAILS)
3376 for (hash_table<congruence_class_hash>::iterator it = m_classes.begin ();
3377 it != m_classes.end (); ++it)
3379 fprintf (dump_file, " group: with %u classes:\n",
3380 (*it)->classes.length ());
3382 for (unsigned i = 0; i < (*it)->classes.length (); i++)
3384 (*it)->classes[i]->dump (dump_file, 4);
3386 if (i < (*it)->classes.length () - 1)
3387 fprintf (dump_file, " ");
3391 free (histogram);
3394 /* Sort pair of sem_items A and B by DECL_UID. */
3396 static int
3397 sort_sem_items_by_decl_uid (const void *a, const void *b)
3399 const sem_item *i1 = *(const sem_item * const *)a;
3400 const sem_item *i2 = *(const sem_item * const *)b;
3402 int uid1 = DECL_UID (i1->decl);
3403 int uid2 = DECL_UID (i2->decl);
3405 if (uid1 < uid2)
3406 return -1;
3407 else if (uid1 > uid2)
3408 return 1;
3409 else
3410 return 0;
3413 /* Sort pair of congruence_classes A and B by DECL_UID of the first member. */
3415 static int
3416 sort_congruence_classes_by_decl_uid (const void *a, const void *b)
3418 const congruence_class *c1 = *(const congruence_class * const *)a;
3419 const congruence_class *c2 = *(const congruence_class * const *)b;
3421 int uid1 = DECL_UID (c1->members[0]->decl);
3422 int uid2 = DECL_UID (c2->members[0]->decl);
3424 if (uid1 < uid2)
3425 return -1;
3426 else if (uid1 > uid2)
3427 return 1;
3428 else
3429 return 0;
3432 /* Sort pair of congruence_class_groups A and B by
3433 DECL_UID of the first member of a first group. */
3435 static int
3436 sort_congruence_class_groups_by_decl_uid (const void *a, const void *b)
3438 const congruence_class_group *g1
3439 = *(const congruence_class_group * const *)a;
3440 const congruence_class_group *g2
3441 = *(const congruence_class_group * const *)b;
3443 int uid1 = DECL_UID (g1->classes[0]->members[0]->decl);
3444 int uid2 = DECL_UID (g2->classes[0]->members[0]->decl);
3446 if (uid1 < uid2)
3447 return -1;
3448 else if (uid1 > uid2)
3449 return 1;
3450 else
3451 return 0;
3454 /* After reduction is done, we can declare all items in a group
3455 to be equal. PREV_CLASS_COUNT is start number of classes
3456 before reduction. True is returned if there's a merge operation
3457 processed. */
3459 bool
3460 sem_item_optimizer::merge_classes (unsigned int prev_class_count)
3462 unsigned int item_count = m_items.length ();
3463 unsigned int class_count = m_classes_count;
3464 unsigned int equal_items = item_count - class_count;
3466 unsigned int non_singular_classes_count = 0;
3467 unsigned int non_singular_classes_sum = 0;
3469 bool merged_p = false;
3471 /* PR lto/78211
3472 Sort functions in congruence classes by DECL_UID and do the same
3473 for the classes to not to break -fcompare-debug. */
3475 for (hash_table<congruence_class_hash>::iterator it = m_classes.begin ();
3476 it != m_classes.end (); ++it)
3478 for (unsigned int i = 0; i < (*it)->classes.length (); i++)
3480 congruence_class *c = (*it)->classes[i];
3481 c->members.qsort (sort_sem_items_by_decl_uid);
3484 (*it)->classes.qsort (sort_congruence_classes_by_decl_uid);
3487 for (hash_table<congruence_class_hash>::iterator it = m_classes.begin ();
3488 it != m_classes.end (); ++it)
3489 for (unsigned int i = 0; i < (*it)->classes.length (); i++)
3491 congruence_class *c = (*it)->classes[i];
3492 if (c->members.length () > 1)
3494 non_singular_classes_count++;
3495 non_singular_classes_sum += c->members.length ();
3499 auto_vec <congruence_class_group *> classes (m_classes.elements ());
3500 for (hash_table<congruence_class_hash>::iterator it = m_classes.begin ();
3501 it != m_classes.end (); ++it)
3502 classes.quick_push (*it);
3504 classes.qsort (sort_congruence_class_groups_by_decl_uid);
3506 if (dump_file)
3508 fprintf (dump_file, "\nItem count: %u\n", item_count);
3509 fprintf (dump_file, "Congruent classes before: %u, after: %u\n",
3510 prev_class_count, class_count);
3511 fprintf (dump_file, "Average class size before: %.2f, after: %.2f\n",
3512 prev_class_count ? 1.0f * item_count / prev_class_count : 0.0f,
3513 class_count ? 1.0f * item_count / class_count : 0.0f);
3514 fprintf (dump_file, "Average non-singular class size: %.2f, count: %u\n",
3515 non_singular_classes_count ? 1.0f * non_singular_classes_sum /
3516 non_singular_classes_count : 0.0f,
3517 non_singular_classes_count);
3518 fprintf (dump_file, "Equal symbols: %u\n", equal_items);
3519 fprintf (dump_file, "Fraction of visited symbols: %.2f%%\n\n",
3520 item_count ? 100.0f * equal_items / item_count : 0.0f);
3523 unsigned int l;
3524 congruence_class_group *it;
3525 FOR_EACH_VEC_ELT (classes, l, it)
3526 for (unsigned int i = 0; i < it->classes.length (); i++)
3528 congruence_class *c = it->classes[i];
3530 if (c->members.length () == 1)
3531 continue;
3533 sem_item *source = c->members[0];
3535 if (DECL_NAME (source->decl)
3536 && MAIN_NAME_P (DECL_NAME (source->decl)))
3537 /* If merge via wrappers, picking main as the target can be
3538 problematic. */
3539 source = c->members[1];
3541 for (unsigned int j = 0; j < c->members.length (); j++)
3543 sem_item *alias = c->members[j];
3545 if (alias == source)
3546 continue;
3548 if (dump_file)
3550 fprintf (dump_file, "Semantic equality hit:%s->%s\n",
3551 xstrdup_for_dump (source->node->name ()),
3552 xstrdup_for_dump (alias->node->name ()));
3553 fprintf (dump_file, "Assembler symbol names:%s->%s\n",
3554 xstrdup_for_dump (source->node->asm_name ()),
3555 xstrdup_for_dump (alias->node->asm_name ()));
3558 if (lookup_attribute ("no_icf", DECL_ATTRIBUTES (alias->decl)))
3560 if (dump_file)
3561 fprintf (dump_file,
3562 "Merge operation is skipped due to no_icf "
3563 "attribute.\n\n");
3565 continue;
3568 if (dump_file && (dump_flags & TDF_DETAILS))
3570 source->dump_to_file (dump_file);
3571 alias->dump_to_file (dump_file);
3574 if (dbg_cnt (merged_ipa_icf))
3575 merged_p |= source->merge (alias);
3579 return merged_p;
3582 /* Dump function prints all class members to a FILE with an INDENT. */
3584 void
3585 congruence_class::dump (FILE *file, unsigned int indent) const
3587 FPRINTF_SPACES (file, indent, "class with id: %u, hash: %u, items: %u\n",
3588 id, members[0]->get_hash (), members.length ());
3590 FPUTS_SPACES (file, indent + 2, "");
3591 for (unsigned i = 0; i < members.length (); i++)
3592 fprintf (file, "%s ", members[i]->node->dump_asm_name ());
3594 fprintf (file, "\n");
3597 /* Returns true if there's a member that is used from another group. */
3599 bool
3600 congruence_class::is_class_used (void)
3602 for (unsigned int i = 0; i < members.length (); i++)
3603 if (members[i]->usages.length ())
3604 return true;
3606 return false;
3609 /* Generate pass summary for IPA ICF pass. */
3611 static void
3612 ipa_icf_generate_summary (void)
3614 if (!optimizer)
3615 optimizer = new sem_item_optimizer ();
3617 optimizer->register_hooks ();
3618 optimizer->parse_funcs_and_vars ();
3621 /* Write pass summary for IPA ICF pass. */
3623 static void
3624 ipa_icf_write_summary (void)
3626 gcc_assert (optimizer);
3628 optimizer->write_summary ();
3631 /* Read pass summary for IPA ICF pass. */
3633 static void
3634 ipa_icf_read_summary (void)
3636 if (!optimizer)
3637 optimizer = new sem_item_optimizer ();
3639 optimizer->read_summary ();
3640 optimizer->register_hooks ();
3643 /* Semantic equality exection function. */
3645 static unsigned int
3646 ipa_icf_driver (void)
3648 gcc_assert (optimizer);
3650 bool merged_p = optimizer->execute ();
3652 delete optimizer;
3653 optimizer = NULL;
3655 return merged_p ? TODO_remove_functions : 0;
3658 const pass_data pass_data_ipa_icf =
3660 IPA_PASS, /* type */
3661 "icf", /* name */
3662 OPTGROUP_IPA, /* optinfo_flags */
3663 TV_IPA_ICF, /* tv_id */
3664 0, /* properties_required */
3665 0, /* properties_provided */
3666 0, /* properties_destroyed */
3667 0, /* todo_flags_start */
3668 0, /* todo_flags_finish */
3671 class pass_ipa_icf : public ipa_opt_pass_d
3673 public:
3674 pass_ipa_icf (gcc::context *ctxt)
3675 : ipa_opt_pass_d (pass_data_ipa_icf, ctxt,
3676 ipa_icf_generate_summary, /* generate_summary */
3677 ipa_icf_write_summary, /* write_summary */
3678 ipa_icf_read_summary, /* read_summary */
3679 NULL, /*
3680 write_optimization_summary */
3681 NULL, /*
3682 read_optimization_summary */
3683 NULL, /* stmt_fixup */
3684 0, /* function_transform_todo_flags_start */
3685 NULL, /* function_transform */
3686 NULL) /* variable_transform */
3689 /* opt_pass methods: */
3690 virtual bool gate (function *)
3692 return in_lto_p || flag_ipa_icf_variables || flag_ipa_icf_functions;
3695 virtual unsigned int execute (function *)
3697 return ipa_icf_driver();
3699 }; // class pass_ipa_icf
3701 } // ipa_icf namespace
3703 ipa_opt_pass_d *
3704 make_pass_ipa_icf (gcc::context *ctxt)
3706 return new ipa_icf::pass_ipa_icf (ctxt);