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
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
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
25 The goal of this transformation is to discover functions and read-only
26 variables which do have exactly the same semantics.
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
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
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.
57 #include "coretypes.h"
63 #include "alloc-pool.h"
64 #include "tree-pass.h"
68 #include "gimple-pretty-print.h"
69 #include "data-streamer.h"
70 #include "fold-const.h"
73 #include "gimple-iterator.h"
75 #include "symbol-summary.h"
77 #include "ipa-fnsummary.h"
80 #include "print-tree.h"
81 #include "ipa-utils.h"
82 #include "ipa-icf-gimple.h"
84 #include "stor-layout.h"
86 #include "tree-vector-builder.h"
88 using namespace ipa_icf_gimple
;
92 /* Initialization and computation of symtab node hash, there data
93 are propagated later on. */
95 static sem_item_optimizer
*optimizer
= NULL
;
99 symbol_compare_collection::symbol_compare_collection (symtab_node
*node
)
101 m_references
.create (0);
102 m_interposables
.create (0);
106 if (is_a
<varpool_node
*> (node
) && DECL_VIRTUAL_P (node
->decl
))
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
);
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)
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)
154 /* Add reference to a semantic TARGET. */
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. */
170 sem_item::setup (bitmap_obstack
*stack
)
172 gcc_checking_assert (node
);
175 tree_refs
.create (0);
177 usage_index_bitmap
= BITMAP_ALLOC (stack
);
180 sem_item::~sem_item ()
182 for (unsigned i
= 0; i
< usages
.length (); i
++)
186 tree_refs
.release ();
189 BITMAP_FREE (usage_index_bitmap
);
192 /* Dump function for debugging purpose. */
195 sem_item::dump (void)
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. */
215 sem_item::target_supports_symbol_aliases_p (void)
217 #if !defined (ASM_OUTPUT_DEF) || (!defined(ASM_OUTPUT_WEAK_ALIAS) && !defined (ASM_WEAKEN_DECL))
224 void sem_item::set_hash (hashval_t hash
)
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
)
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
)
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
]);
252 bb_sorted
.release ();
255 /* Calculates hash value based on a BASIC_BLOCK. */
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. */
271 sem_function::get_hash (void)
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
))
291 (cl_target_option_hash
292 (TREE_TARGET_OPTION (DECL_FUNCTION_SPECIFIC_TARGET (decl
))));
293 if (DECL_FUNCTION_SPECIFIC_OPTIMIZATION (decl
))
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 ());
306 /* Return ture if A1 and A2 represent equivalent function attribute lists.
307 Based on comp_type_attributes. */
310 sem_item::compare_attributes (const_tree a1
, const_tree a2
)
315 for (a
= a1
; a
!= NULL_TREE
; a
= TREE_CHAIN (a
))
317 const struct attribute_spec
*as
;
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
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. */
331 attr
= lookup_attribute (as
->name
, CONST_CAST_TREE (a2
));
332 if (!attr
|| !attribute_value_equal (a
, attr
))
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
));
345 if (!lookup_attribute (as
->name
, CONST_CAST_TREE (a1
)))
347 /* We don't need to compare trees again, as we did this
348 already in first loop. */
353 /* TODO: As in comp_type_attributes we may want to introduce target hook. */
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
369 sem_item::compare_referenced_symbol_properties (symtab_node
*used_by
,
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");
452 /* Hash properties that are compared by compare_referenced_symbol_properties. */
455 sem_item::hash_referenced_symbol_properties (symtab_node
*ref
,
456 inchash::hash
&hstate
,
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
));
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. */
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
;
493 /* Never match variable and function. */
494 if (is_a
<varpool_node
*> (n1
) != is_a
<varpool_node
*> (n2
))
497 if (!compare_referenced_symbol_properties (node
, n1
, n2
, address
))
499 if (address
&& n1
->equal_address_to (n2
) == 1)
501 if (!address
&& n1
->semantically_equivalent_p (n2
))
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
))
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
)
533 /* Return true if parameter I may be used. */
536 sem_function::param_used_p (unsigned int i
)
538 if (ipa_node_params_sum
== NULL
)
541 struct ipa_node_params
*parms_info
= IPA_NODE_REF (get_node ());
543 if (vec_safe_length (parms_info
->descriptors
) <= i
)
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. */
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");
573 /* Fast equality function based on knowledge known in WPA. */
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
&& !cl_optimization_option_eq (opt1
, opt2
))
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
))
699 /* Perform additional checks for used parameters. */
700 if (!compatible_parm_types_p (parm1
, parm2
))
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
)
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
,
745 ref
->address_matters_p ()))
749 cgraph_edge
*e1
= dyn_cast
<cgraph_node
*> (node
)->callees
;
750 cgraph_edge
*e2
= dyn_cast
<cgraph_node
*> (item
->node
)->callees
;
754 if (!compare_symbol_references (ignored_nodes
, e1
->callee
,
757 if (!compare_edge_flags (e1
, e2
))
760 e1
= e1
->next_callee
;
761 e2
= e2
->next_callee
;
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
;
772 if (!compare_edge_flags (e1
, e2
))
775 e1
= e1
->next_callee
;
776 e2
= e2
->next_callee
;
780 return return_false_with_msg ("different number of indirect calls");
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. */
794 sem_item::update_hash_by_addr_refs (hash_map
<symtab_node
*,
795 sem_item
*> &m_symtab_node_map
)
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
;
814 sem_item
**result
= m_symtab_node_map
.get (e
->callee
);
815 hash_referenced_symbol_properties (e
->callee
, hstate
, false);
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
826 TODO: stronger SCC based hashing would be desirable here. */
829 sem_item::update_hash_by_local_refs (hash_map
<symtab_node
*,
830 sem_item
*> &m_symtab_node_map
)
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
);
839 state
.merge_hash ((*result
)->get_hash ());
844 for (cgraph_edge
*e
= dyn_cast
<cgraph_node
*> (node
)->callees
; e
;
847 sem_item
**result
= m_symtab_node_map
.get (e
->caller
);
849 state
.merge_hash ((*result
)->get_hash ());
853 global_hash
= state
.end ();
856 /* Returns true if the item equals to ITEM given as argument. */
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
)
871 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
873 "Equals called for: %s:%s with result: %s\n\n",
875 item
->node
->dump_name (),
876 eq
? "true" : "false");
881 /* Processes function equality comparison. */
884 sem_function::equals_private (sem_item
*item
)
886 if (item
->type
!= FUNC
)
889 basic_block bb1
, bb2
;
891 edge_iterator ei1
, ei2
;
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 (),
908 &m_compared_func
->refs_set
);
909 arg1
= DECL_ARGUMENTS (decl
);
910 arg2
= DECL_ARGUMENTS (m_compared_func
->decl
);
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
))
918 /* Perform additional checks for used parameters. */
919 if (!compatible_parm_types_p (TREE_TYPE (arg1
), TREE_TYPE (arg2
)))
921 if (!m_checker
->compare_decl (arg1
, arg2
))
922 return return_false ();
925 return return_false_with_msg ("Mismatched number of arguments");
927 if (!dyn_cast
<cgraph_node
*> (node
)->has_gimple_body_p ())
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
))
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");
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");
982 /* Set LOCAL_P of NODE to true if DATA is non-NULL.
983 Helper for call_for_symbol_thunks_and_aliases. */
986 set_local (cgraph_node
*node
, void *data
)
988 node
->local
.local
= data
!= NULL
;
992 /* TREE_ADDRESSABLE of NODE to true.
993 Helper for call_for_symbol_thunks_and_aliases. */
996 set_addressable (varpool_node
*node
, void *)
998 TREE_ADDRESSABLE (node
->decl
) = 1;
1002 /* Clear DECL_RTL of NODE.
1003 Helper for call_for_symbol_thunks_and_aliases. */
1006 clear_decl_rtl (symtab_node
*node
, void *)
1008 SET_DECL_RTL (node
->decl
, NULL
);
1012 /* Redirect all callers of N and its aliases to TO. Remove aliases if
1013 possible. Return number of redirections made. */
1016 redirect_all_callers (cgraph_node
*n
, cgraph_node
*to
)
1018 int nredirected
= 0;
1020 cgraph_edge
*e
= n
->callers
;
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
);
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
,
1052 && !n_alias
->has_aliases_p ())
1061 /* Merges instance with an ALIAS_ITEM, where alias, thunk or redirection can
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
))
1089 fprintf (dump_file
, "Not unifying; alias is external.\n\n");
1093 if (DECL_NO_INLINE_WARNING_P (original
->decl
)
1094 != DECL_NO_INLINE_WARNING_P (alias
->decl
))
1099 "DECL_NO_INLINE_WARNING mismatch.\n\n");
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
))
1112 "original and alias are in different sections.\n\n");
1116 if (!original
->in_same_comdat_group_p (alias
)
1117 || original
->comdat_local_p ())
1121 "Not unifying; alias nor wrapper cannot be created; "
1122 "across comdat group boundary\n\n");
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
))
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
)))
1184 "Wrapper cannot be created because of COMDAT\n");
1186 else if (DECL_STATIC_CHAIN (alias
->decl
)
1187 || DECL_STATIC_CHAIN (original
->decl
))
1191 "Cannot create wrapper of nested function.\n");
1193 /* TODO: We can also deal with variadic functions never calling
1195 else if (stdarg_p (TREE_TYPE (alias
->decl
)))
1199 "can not create wrapper of stdarg function.\n");
1201 else if (ipa_fn_summaries
1202 && ipa_fn_summaries
->get (alias
) != NULL
1203 && ipa_fn_summaries
->get (alias
)->self_size
<= 2)
1206 fprintf (dump_file
, "Wrapper creation is not "
1207 "profitable (function is too small).\n");
1209 /* If user paid attention to mark function noinline, assume it is
1210 somewhat special and do not try to turn it into a wrapper that can
1211 not be undone by inliner. */
1212 else if (lookup_attribute ("noinline", DECL_ATTRIBUTES (alias
->decl
)))
1215 fprintf (dump_file
, "Wrappers are not created for noinline.\n");
1218 create_wrapper
= true;
1220 /* We can redirect local calls in the case both alias and orignal
1221 are not interposable. */
1223 = alias
->get_availability () > AVAIL_INTERPOSABLE
1224 && original
->get_availability () > AVAIL_INTERPOSABLE
;
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
)
1234 fprintf (dump_file
, "Not unifying; can not redirect callers nor "
1235 "produce wrapper\n\n");
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 ())
1248 = dyn_cast
<cgraph_node
*> (original
->noninterposable_alias ());
1250 && original
->get_availability () > AVAIL_INTERPOSABLE
)
1251 local_original
= original
;
1253 /* If we can not use local alias, fallback to the original
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
)
1265 fprintf (dump_file
, "Not unifying; "
1266 "can not produce local alias.\n\n");
1270 if (!redirect_callers
&& !create_wrapper
)
1273 fprintf (dump_file
, "Not unifying; "
1274 "can not redirect callers nor produce a wrapper\n\n");
1278 && !alias
->call_for_symbol_and_aliases (cgraph_node::has_thunk_p
,
1280 && !alias
->can_remove_if_no_direct_calls_p ())
1283 fprintf (dump_file
, "Not unifying; can not make wrapper and "
1284 "function has other uses than direct calls\n\n");
1289 create_alias
= true;
1291 if (redirect_callers
)
1293 int nredirected
= redirect_all_callers (alias
, local_original
);
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;
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);
1323 /* Notice global symbol possibly produced RTL. */
1324 ((symtab_node
*)alias
)->call_for_symbol_and_aliases (clear_decl_rtl
,
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);
1335 fprintf (dump_file
, "Unified; Function alias has been created.\n\n");
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
);
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
)
1362 fprintf (dump_file
, "Dropping merged_comdat flag.\n\n");
1364 local_original
->merged_comdat
= false;
1365 original
->merged_comdat
= false;
1370 ipa_merge_profiles (original
, alias
);
1371 alias
->release_body ();
1373 alias
->body_removed
= true;
1374 alias
->icf_merged
= true;
1376 fprintf (dump_file
, "Unified; Function body was removed.\n");
1382 /* Semantic item initialization function. */
1385 sem_function::init (void)
1388 get_node ()->get_untransformed_body ();
1390 tree fndecl
= node
->decl
;
1391 function
*func
= DECL_STRUCT_FUNCTION (fndecl
);
1394 gcc_assert (SSANAMES (func
));
1396 ssa_names_size
= SSANAMES (func
)->length ();
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
;
1414 FOR_EACH_BB_FN (bb
, func
)
1416 unsigned nondbg_stmt_count
= 0;
1419 for (edge_iterator ei
= ei_start (bb
->preds
); ei_cond (ei
, &e
);
1421 cfg_checksum
= iterative_hash_host_wide_int (e
->flags
,
1424 for (gimple_stmt_iterator gsi
= gsi_start_bb (bb
); !gsi_end_p (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
);
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. */
1467 sem_item::add_expr (const_tree exp
, inchash::hash
&hstate
)
1469 if (exp
== NULL_TREE
)
1471 hstate
.merge_hash (0);
1475 /* Handled component can be matched in a cureful way proving equivalence
1476 even if they syntactically differ. Just skip them. */
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
);
1486 /* Use inchash::add_expr for everything that is LTO stable. */
1494 inchash::add_expr (exp
, hstate
);
1498 unsigned HOST_WIDE_INT idx
;
1501 hstate
.add_hwi (int_size_in_bytes (TREE_TYPE (exp
)));
1503 FOR_EACH_CONSTRUCTOR_VALUE (CONSTRUCTOR_ELTS (exp
), idx
, value
)
1505 add_expr (value
, hstate
);
1510 add_expr (get_base_address (TREE_OPERAND (exp
, 0)), hstate
);
1516 hstate
.add_hwi (int_size_in_bytes (TREE_TYPE (exp
)));
1519 case POINTER_PLUS_EXPR
:
1522 add_expr (TREE_OPERAND (exp
, 0), hstate
);
1523 add_expr (TREE_OPERAND (exp
, 1), hstate
);
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
);
1534 hstate
.add_hwi (int_size_in_bytes (TREE_TYPE (exp
)));
1535 return add_expr (TREE_OPERAND (exp
, 0), hstate
);
1541 /* Accumulate to HSTATE a hash of type t.
1542 TYpes that may end up being compatible after LTO type merging needs to have
1546 sem_item::add_type (const_tree type
, inchash::hash
&hstate
)
1548 if (type
== NULL_TREE
)
1550 hstate
.merge_hash (0);
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 /* Incomplete types must be skipped here. */
1584 if (!COMPLETE_TYPE_P (type
))
1586 hstate
.add_int (RECORD_TYPE
);
1590 hashval_t
*val
= optimizer
->m_type_hash_cache
.get (type
);
1594 inchash::hash hstate2
;
1599 hstate2
.add_int (RECORD_TYPE
);
1600 for (f
= TYPE_FIELDS (type
), nf
= 0; f
; f
= TREE_CHAIN (f
))
1601 if (TREE_CODE (f
) == FIELD_DECL
)
1603 add_type (TREE_TYPE (f
), hstate2
);
1607 hstate2
.add_int (nf
);
1608 hash
= hstate2
.end ();
1609 hstate
.add_hwi (hash
);
1610 optimizer
->m_type_hash_cache
.put (type
, hash
);
1613 hstate
.add_hwi (*val
);
1617 /* Improve accumulated hash for HSTATE based on a gimple statement STMT. */
1620 sem_function::hash_stmt (gimple
*stmt
, inchash::hash
&hstate
)
1622 enum gimple_code code
= gimple_code (stmt
);
1624 hstate
.add_int (code
);
1629 add_expr (gimple_switch_index (as_a
<gswitch
*> (stmt
)), hstate
);
1632 hstate
.add_int (gimple_assign_rhs_code (stmt
));
1633 if (commutative_tree_code (gimple_assign_rhs_code (stmt
))
1634 || commutative_ternary_tree_code (gimple_assign_rhs_code (stmt
)))
1636 inchash::hash one
, two
;
1638 add_expr (gimple_assign_rhs1 (stmt
), one
);
1639 add_type (TREE_TYPE (gimple_assign_rhs1 (stmt
)), one
);
1640 add_expr (gimple_assign_rhs2 (stmt
), two
);
1641 hstate
.add_commutative (one
, two
);
1642 if (commutative_ternary_tree_code (gimple_assign_rhs_code (stmt
)))
1644 add_expr (gimple_assign_rhs3 (stmt
), hstate
);
1645 add_type (TREE_TYPE (gimple_assign_rhs3 (stmt
)), hstate
);
1647 add_expr (gimple_assign_lhs (stmt
), hstate
);
1648 add_type (TREE_TYPE (gimple_assign_lhs (stmt
)), two
);
1657 /* All these statements are equivalent if their operands are. */
1658 for (unsigned i
= 0; i
< gimple_num_ops (stmt
); ++i
)
1660 add_expr (gimple_op (stmt
, i
), hstate
);
1661 if (gimple_op (stmt
, i
))
1662 add_type (TREE_TYPE (gimple_op (stmt
, i
)), hstate
);
1664 /* Consider nocf_check attribute in hash as it affects code
1666 if (code
== GIMPLE_CALL
1667 && flag_cf_protection
& CF_BRANCH
)
1668 hstate
.add_flag (gimple_call_nocf_check_p (as_a
<gcall
*> (stmt
)));
1675 /* Return true if polymorphic comparison must be processed. */
1678 sem_function::compare_polymorphic_p (void)
1680 struct cgraph_edge
*e
;
1682 if (!opt_for_fn (get_node ()->decl
, flag_devirtualize
))
1684 if (get_node ()->indirect_calls
!= NULL
)
1686 /* TODO: We can do simple propagation determining what calls may lead to
1687 a polymorphic call. */
1688 for (e
= get_node ()->callees
; e
; e
= e
->next_callee
)
1689 if (e
->callee
->definition
1690 && opt_for_fn (e
->callee
->decl
, flag_devirtualize
))
1695 /* For a given call graph NODE, the function constructs new
1696 semantic function item. */
1699 sem_function::parse (cgraph_node
*node
, bitmap_obstack
*stack
)
1701 tree fndecl
= node
->decl
;
1702 function
*func
= DECL_STRUCT_FUNCTION (fndecl
);
1704 if (!func
|| (!node
->has_gimple_body_p () && !node
->thunk
.thunk_p
))
1707 if (lookup_attribute_by_prefix ("omp ", DECL_ATTRIBUTES (node
->decl
)) != NULL
)
1710 if (lookup_attribute_by_prefix ("oacc ",
1711 DECL_ATTRIBUTES (node
->decl
)) != NULL
)
1715 if (DECL_STATIC_CONSTRUCTOR (node
->decl
)
1716 || DECL_STATIC_DESTRUCTOR (node
->decl
))
1719 sem_function
*f
= new sem_function (node
, stack
);
1726 /* For given basic blocks BB1 and BB2 (from functions FUNC1 and FUNC),
1727 return true if phi nodes are semantically equivalent in these blocks . */
1730 sem_function::compare_phi_node (basic_block bb1
, basic_block bb2
)
1732 gphi_iterator si1
, si2
;
1734 unsigned size1
, size2
, i
;
1738 gcc_assert (bb1
!= NULL
);
1739 gcc_assert (bb2
!= NULL
);
1741 si2
= gsi_start_phis (bb2
);
1742 for (si1
= gsi_start_phis (bb1
); !gsi_end_p (si1
);
1745 gsi_next_nonvirtual_phi (&si1
);
1746 gsi_next_nonvirtual_phi (&si2
);
1748 if (gsi_end_p (si1
) && gsi_end_p (si2
))
1751 if (gsi_end_p (si1
) || gsi_end_p (si2
))
1752 return return_false();
1757 tree phi_result1
= gimple_phi_result (phi1
);
1758 tree phi_result2
= gimple_phi_result (phi2
);
1760 if (!m_checker
->compare_operand (phi_result1
, phi_result2
))
1761 return return_false_with_msg ("PHI results are different");
1763 size1
= gimple_phi_num_args (phi1
);
1764 size2
= gimple_phi_num_args (phi2
);
1767 return return_false ();
1769 for (i
= 0; i
< size1
; ++i
)
1771 t1
= gimple_phi_arg (phi1
, i
)->def
;
1772 t2
= gimple_phi_arg (phi2
, i
)->def
;
1774 if (!m_checker
->compare_operand (t1
, t2
))
1775 return return_false ();
1777 e1
= gimple_phi_arg_edge (phi1
, i
);
1778 e2
= gimple_phi_arg_edge (phi2
, i
);
1780 if (!m_checker
->compare_edge (e1
, e2
))
1781 return return_false ();
1790 /* Returns true if tree T can be compared as a handled component. */
1793 sem_function::icf_handled_component_p (tree t
)
1795 tree_code tc
= TREE_CODE (t
);
1797 return (handled_component_p (t
)
1798 || tc
== ADDR_EXPR
|| tc
== MEM_REF
|| tc
== OBJ_TYPE_REF
);
1801 /* Basic blocks dictionary BB_DICT returns true if SOURCE index BB
1802 corresponds to TARGET. */
1805 sem_function::bb_dict_test (vec
<int> *bb_dict
, int source
, int target
)
1810 if (bb_dict
->length () <= (unsigned)source
)
1811 bb_dict
->safe_grow_cleared (source
+ 1);
1813 if ((*bb_dict
)[source
] == 0)
1815 (*bb_dict
)[source
] = target
;
1819 return (*bb_dict
)[source
] == target
;
1822 sem_variable::sem_variable (bitmap_obstack
*stack
): sem_item (VAR
, stack
)
1826 sem_variable::sem_variable (varpool_node
*node
, bitmap_obstack
*stack
)
1827 : sem_item (VAR
, node
, stack
)
1829 gcc_checking_assert (node
);
1830 gcc_checking_assert (get_node ());
1833 /* Fast equality function based on knowledge known in WPA. */
1836 sem_variable::equals_wpa (sem_item
*item
,
1837 hash_map
<symtab_node
*, sem_item
*> &ignored_nodes
)
1839 gcc_assert (item
->type
== VAR
);
1841 if (node
->num_references () != item
->node
->num_references ())
1842 return return_false_with_msg ("different number of references");
1844 if (DECL_TLS_MODEL (decl
) || DECL_TLS_MODEL (item
->decl
))
1845 return return_false_with_msg ("TLS model");
1847 /* DECL_ALIGN is safe to merge, because we will always chose the largest
1848 alignment out of all aliases. */
1850 if (DECL_VIRTUAL_P (decl
) != DECL_VIRTUAL_P (item
->decl
))
1851 return return_false_with_msg ("Virtual flag mismatch");
1853 if (DECL_SIZE (decl
) != DECL_SIZE (item
->decl
)
1854 && ((!DECL_SIZE (decl
) || !DECL_SIZE (item
->decl
))
1855 || !operand_equal_p (DECL_SIZE (decl
),
1856 DECL_SIZE (item
->decl
), OEP_ONLY_CONST
)))
1857 return return_false_with_msg ("size mismatch");
1859 /* Do not attempt to mix data from different user sections;
1860 we do not know what user intends with those. */
1861 if (((DECL_SECTION_NAME (decl
) && !node
->implicit_section
)
1862 || (DECL_SECTION_NAME (item
->decl
) && !item
->node
->implicit_section
))
1863 && DECL_SECTION_NAME (decl
) != DECL_SECTION_NAME (item
->decl
))
1864 return return_false_with_msg ("user section mismatch");
1866 if (DECL_IN_TEXT_SECTION (decl
) != DECL_IN_TEXT_SECTION (item
->decl
))
1867 return return_false_with_msg ("text section");
1869 ipa_ref
*ref
= NULL
, *ref2
= NULL
;
1870 for (unsigned i
= 0; node
->iterate_reference (i
, ref
); i
++)
1872 item
->node
->iterate_reference (i
, ref2
);
1874 if (ref
->use
!= ref2
->use
)
1875 return return_false_with_msg ("reference use mismatch");
1877 if (!compare_symbol_references (ignored_nodes
,
1878 ref
->referred
, ref2
->referred
,
1879 ref
->address_matters_p ()))
1886 /* Returns true if the item equals to ITEM given as argument. */
1889 sem_variable::equals (sem_item
*item
,
1890 hash_map
<symtab_node
*, sem_item
*> &)
1892 gcc_assert (item
->type
== VAR
);
1895 if (DECL_INITIAL (decl
) == error_mark_node
&& in_lto_p
)
1896 dyn_cast
<varpool_node
*>(node
)->get_constructor ();
1897 if (DECL_INITIAL (item
->decl
) == error_mark_node
&& in_lto_p
)
1898 dyn_cast
<varpool_node
*>(item
->node
)->get_constructor ();
1900 /* As seen in PR ipa/65303 we have to compare variables types. */
1901 if (!func_checker::compatible_types_p (TREE_TYPE (decl
),
1902 TREE_TYPE (item
->decl
)))
1903 return return_false_with_msg ("variables types are different");
1905 ret
= sem_variable::equals (DECL_INITIAL (decl
),
1906 DECL_INITIAL (item
->node
->decl
));
1907 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1909 "Equals called for vars: %s:%s with result: %s\n\n",
1910 node
->dump_name (), item
->node
->dump_name (),
1911 ret
? "true" : "false");
1916 /* Compares trees T1 and T2 for semantic equality. */
1919 sem_variable::equals (tree t1
, tree t2
)
1922 return return_with_debug (t1
== t2
);
1925 tree_code tc1
= TREE_CODE (t1
);
1926 tree_code tc2
= TREE_CODE (t2
);
1929 return return_false_with_msg ("TREE_CODE mismatch");
1935 vec
<constructor_elt
, va_gc
> *v1
, *v2
;
1936 unsigned HOST_WIDE_INT idx
;
1938 enum tree_code typecode
= TREE_CODE (TREE_TYPE (t1
));
1939 if (typecode
!= TREE_CODE (TREE_TYPE (t2
)))
1940 return return_false_with_msg ("constructor type mismatch");
1942 if (typecode
== ARRAY_TYPE
)
1944 HOST_WIDE_INT size_1
= int_size_in_bytes (TREE_TYPE (t1
));
1945 /* For arrays, check that the sizes all match. */
1946 if (TYPE_MODE (TREE_TYPE (t1
)) != TYPE_MODE (TREE_TYPE (t2
))
1948 || size_1
!= int_size_in_bytes (TREE_TYPE (t2
)))
1949 return return_false_with_msg ("constructor array size mismatch");
1951 else if (!func_checker::compatible_types_p (TREE_TYPE (t1
),
1953 return return_false_with_msg ("constructor type incompatible");
1955 v1
= CONSTRUCTOR_ELTS (t1
);
1956 v2
= CONSTRUCTOR_ELTS (t2
);
1957 if (vec_safe_length (v1
) != vec_safe_length (v2
))
1958 return return_false_with_msg ("constructor number of elts mismatch");
1960 for (idx
= 0; idx
< vec_safe_length (v1
); ++idx
)
1962 constructor_elt
*c1
= &(*v1
)[idx
];
1963 constructor_elt
*c2
= &(*v2
)[idx
];
1965 /* Check that each value is the same... */
1966 if (!sem_variable::equals (c1
->value
, c2
->value
))
1968 /* ... and that they apply to the same fields! */
1969 if (!sem_variable::equals (c1
->index
, c2
->index
))
1976 tree x1
= TREE_OPERAND (t1
, 0);
1977 tree x2
= TREE_OPERAND (t2
, 0);
1978 tree y1
= TREE_OPERAND (t1
, 1);
1979 tree y2
= TREE_OPERAND (t2
, 1);
1981 if (!func_checker::compatible_types_p (TREE_TYPE (x1
), TREE_TYPE (x2
)))
1982 return return_false ();
1984 /* Type of the offset on MEM_REF does not matter. */
1985 return return_with_debug (sem_variable::equals (x1
, x2
)
1986 && known_eq (wi::to_poly_offset (y1
),
1987 wi::to_poly_offset (y2
)));
1992 tree op1
= TREE_OPERAND (t1
, 0);
1993 tree op2
= TREE_OPERAND (t2
, 0);
1994 return sem_variable::equals (op1
, op2
);
1996 /* References to other vars/decls are compared using ipa-ref. */
1999 if (decl_in_symtab_p (t1
) && decl_in_symtab_p (t2
))
2001 return return_false_with_msg ("Declaration mismatch");
2003 /* TODO: We can check CONST_DECL by its DECL_INITIAL, but for that we
2004 need to process its VAR/FUNCTION references without relying on ipa-ref
2008 return return_false_with_msg ("Declaration mismatch");
2010 /* Integer constants are the same only if the same width of type. */
2011 if (TYPE_PRECISION (TREE_TYPE (t1
)) != TYPE_PRECISION (TREE_TYPE (t2
)))
2012 return return_false_with_msg ("INTEGER_CST precision mismatch");
2013 if (TYPE_MODE (TREE_TYPE (t1
)) != TYPE_MODE (TREE_TYPE (t2
)))
2014 return return_false_with_msg ("INTEGER_CST mode mismatch");
2015 return return_with_debug (tree_int_cst_equal (t1
, t2
));
2017 if (TYPE_MODE (TREE_TYPE (t1
)) != TYPE_MODE (TREE_TYPE (t2
)))
2018 return return_false_with_msg ("STRING_CST mode mismatch");
2019 if (TREE_STRING_LENGTH (t1
) != TREE_STRING_LENGTH (t2
))
2020 return return_false_with_msg ("STRING_CST length mismatch");
2021 if (memcmp (TREE_STRING_POINTER (t1
), TREE_STRING_POINTER (t2
),
2022 TREE_STRING_LENGTH (t1
)))
2023 return return_false_with_msg ("STRING_CST mismatch");
2026 /* Fixed constants are the same only if the same width of type. */
2027 if (TYPE_PRECISION (TREE_TYPE (t1
)) != TYPE_PRECISION (TREE_TYPE (t2
)))
2028 return return_false_with_msg ("FIXED_CST precision mismatch");
2030 return return_with_debug (FIXED_VALUES_IDENTICAL (TREE_FIXED_CST (t1
),
2031 TREE_FIXED_CST (t2
)));
2033 return (sem_variable::equals (TREE_REALPART (t1
), TREE_REALPART (t2
))
2034 && sem_variable::equals (TREE_IMAGPART (t1
), TREE_IMAGPART (t2
)));
2036 /* Real constants are the same only if the same width of type. */
2037 if (TYPE_PRECISION (TREE_TYPE (t1
)) != TYPE_PRECISION (TREE_TYPE (t2
)))
2038 return return_false_with_msg ("REAL_CST precision mismatch");
2039 return return_with_debug (real_identical (&TREE_REAL_CST (t1
),
2040 &TREE_REAL_CST (t2
)));
2043 if (maybe_ne (VECTOR_CST_NELTS (t1
), VECTOR_CST_NELTS (t2
)))
2044 return return_false_with_msg ("VECTOR_CST nelts mismatch");
2047 = tree_vector_builder::binary_encoded_nelts (t1
, t2
);
2048 for (unsigned int i
= 0; i
< count
; ++i
)
2049 if (!sem_variable::equals (VECTOR_CST_ENCODED_ELT (t1
, i
),
2050 VECTOR_CST_ENCODED_ELT (t2
, i
)))
2056 case ARRAY_RANGE_REF
:
2058 tree x1
= TREE_OPERAND (t1
, 0);
2059 tree x2
= TREE_OPERAND (t2
, 0);
2060 tree y1
= TREE_OPERAND (t1
, 1);
2061 tree y2
= TREE_OPERAND (t2
, 1);
2063 if (!sem_variable::equals (x1
, x2
) || !sem_variable::equals (y1
, y2
))
2065 if (!sem_variable::equals (array_ref_low_bound (t1
),
2066 array_ref_low_bound (t2
)))
2068 if (!sem_variable::equals (array_ref_element_size (t1
),
2069 array_ref_element_size (t2
)))
2075 case POINTER_PLUS_EXPR
:
2080 tree x1
= TREE_OPERAND (t1
, 0);
2081 tree x2
= TREE_OPERAND (t2
, 0);
2082 tree y1
= TREE_OPERAND (t1
, 1);
2083 tree y2
= TREE_OPERAND (t2
, 1);
2085 return sem_variable::equals (x1
, x2
) && sem_variable::equals (y1
, y2
);
2089 case VIEW_CONVERT_EXPR
:
2090 if (!func_checker::compatible_types_p (TREE_TYPE (t1
), TREE_TYPE (t2
)))
2091 return return_false ();
2092 return sem_variable::equals (TREE_OPERAND (t1
, 0), TREE_OPERAND (t2
, 0));
2094 return return_false_with_msg ("ERROR_MARK");
2096 return return_false_with_msg ("Unknown TREE code reached");
2100 /* Parser function that visits a varpool NODE. */
2103 sem_variable::parse (varpool_node
*node
, bitmap_obstack
*stack
)
2105 if (TREE_THIS_VOLATILE (node
->decl
) || DECL_HARD_REGISTER (node
->decl
)
2109 sem_variable
*v
= new sem_variable (node
, stack
);
2116 /* References independent hash function. */
2119 sem_variable::get_hash (void)
2124 /* All WPA streamed in symbols should have their hashes computed at compile
2125 time. At this point, the constructor may not be in memory at all.
2126 DECL_INITIAL (decl) would be error_mark_node in that case. */
2127 gcc_assert (!node
->lto_file_data
);
2128 tree ctor
= DECL_INITIAL (decl
);
2129 inchash::hash hstate
;
2131 hstate
.add_int (456346417);
2132 if (DECL_SIZE (decl
) && tree_fits_shwi_p (DECL_SIZE (decl
)))
2133 hstate
.add_hwi (tree_to_shwi (DECL_SIZE (decl
)));
2134 add_expr (ctor
, hstate
);
2135 set_hash (hstate
.end ());
2140 /* Merges instance with an ALIAS_ITEM, where alias, thunk or redirection can
2144 sem_variable::merge (sem_item
*alias_item
)
2146 gcc_assert (alias_item
->type
== VAR
);
2148 if (!sem_item::target_supports_symbol_aliases_p ())
2151 fprintf (dump_file
, "Not unifying; "
2152 "Symbol aliases are not supported by target\n\n");
2156 if (DECL_EXTERNAL (alias_item
->decl
))
2159 fprintf (dump_file
, "Not unifying; alias is external.\n\n");
2163 sem_variable
*alias_var
= static_cast<sem_variable
*> (alias_item
);
2165 varpool_node
*original
= get_node ();
2166 varpool_node
*alias
= alias_var
->get_node ();
2167 bool original_discardable
= false;
2169 bool alias_address_matters
= alias
->address_matters_p ();
2171 /* See if original is in a section that can be discarded if the main
2173 Also consider case where we have resolution info and we know that
2174 original's definition is not going to be used. In this case we can not
2175 create alias to original. */
2176 if (original
->can_be_discarded_p ()
2177 || (node
->resolution
!= LDPR_UNKNOWN
2178 && !decl_binds_to_current_def_p (node
->decl
)))
2179 original_discardable
= true;
2181 gcc_assert (!TREE_ASM_WRITTEN (alias
->decl
));
2183 /* Constant pool machinery is not quite ready for aliases.
2184 TODO: varasm code contains logic for merging DECL_IN_CONSTANT_POOL.
2185 For LTO merging does not happen that is an important missing feature.
2186 We can enable merging with LTO if the DECL_IN_CONSTANT_POOL
2187 flag is dropped and non-local symbol name is assigned. */
2188 if (DECL_IN_CONSTANT_POOL (alias
->decl
)
2189 || DECL_IN_CONSTANT_POOL (original
->decl
))
2193 "Not unifying; constant pool variables.\n\n");
2197 /* Do not attempt to mix functions from different user sections;
2198 we do not know what user intends with those. */
2199 if (((DECL_SECTION_NAME (original
->decl
) && !original
->implicit_section
)
2200 || (DECL_SECTION_NAME (alias
->decl
) && !alias
->implicit_section
))
2201 && DECL_SECTION_NAME (original
->decl
) != DECL_SECTION_NAME (alias
->decl
))
2206 "original and alias are in different sections.\n\n");
2210 /* We can not merge if address comparsion metters. */
2211 if (alias_address_matters
&& flag_merge_constants
< 2)
2215 "Not unifying; address of original may be compared.\n\n");
2219 if (DECL_ALIGN (original
->decl
) < DECL_ALIGN (alias
->decl
))
2222 fprintf (dump_file
, "Not unifying; "
2223 "original and alias have incompatible alignments\n\n");
2228 if (DECL_COMDAT_GROUP (original
->decl
) != DECL_COMDAT_GROUP (alias
->decl
))
2231 fprintf (dump_file
, "Not unifying; alias cannot be created; "
2232 "across comdat group boundary\n\n");
2237 if (original_discardable
)
2240 fprintf (dump_file
, "Not unifying; alias cannot be created; "
2241 "target is discardable\n\n");
2247 gcc_assert (!original
->alias
);
2248 gcc_assert (!alias
->alias
);
2250 alias
->analyzed
= false;
2252 DECL_INITIAL (alias
->decl
) = NULL
;
2253 ((symtab_node
*)alias
)->call_for_symbol_and_aliases (clear_decl_rtl
,
2255 alias
->need_bounds_init
= false;
2256 alias
->remove_all_references ();
2257 if (TREE_ADDRESSABLE (alias
->decl
))
2258 original
->call_for_symbol_and_aliases (set_addressable
, NULL
, true);
2260 varpool_node::create_alias (alias_var
->decl
, decl
);
2261 alias
->resolve_alias (original
);
2264 fprintf (dump_file
, "Unified; Variable alias has been created.\n");
2270 /* Dump symbol to FILE. */
2273 sem_variable::dump_to_file (FILE *file
)
2277 print_node (file
, "", decl
, 0);
2278 fprintf (file
, "\n\n");
2281 unsigned int sem_item_optimizer::class_id
= 0;
2283 sem_item_optimizer::sem_item_optimizer ()
2284 : worklist (0), m_classes (0), m_classes_count (0), m_cgraph_node_hooks (NULL
),
2285 m_varpool_node_hooks (NULL
), m_merged_variables ()
2288 bitmap_obstack_initialize (&m_bmstack
);
2291 sem_item_optimizer::~sem_item_optimizer ()
2293 for (unsigned int i
= 0; i
< m_items
.length (); i
++)
2297 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
2298 it
!= m_classes
.end (); ++it
)
2300 for (unsigned int i
= 0; i
< (*it
)->classes
.length (); i
++)
2301 delete (*it
)->classes
[i
];
2303 (*it
)->classes
.release ();
2309 bitmap_obstack_release (&m_bmstack
);
2310 m_merged_variables
.release ();
2313 /* Write IPA ICF summary for symbols. */
2316 sem_item_optimizer::write_summary (void)
2318 unsigned int count
= 0;
2320 output_block
*ob
= create_output_block (LTO_section_ipa_icf
);
2321 lto_symtab_encoder_t encoder
= ob
->decl_state
->symtab_node_encoder
;
2324 /* Calculate number of symbols to be serialized. */
2325 for (lto_symtab_encoder_iterator lsei
= lsei_start_in_partition (encoder
);
2327 lsei_next_in_partition (&lsei
))
2329 symtab_node
*node
= lsei_node (lsei
);
2331 if (m_symtab_node_map
.get (node
))
2335 streamer_write_uhwi (ob
, count
);
2337 /* Process all of the symbols. */
2338 for (lto_symtab_encoder_iterator lsei
= lsei_start_in_partition (encoder
);
2340 lsei_next_in_partition (&lsei
))
2342 symtab_node
*node
= lsei_node (lsei
);
2344 sem_item
**item
= m_symtab_node_map
.get (node
);
2348 int node_ref
= lto_symtab_encoder_encode (encoder
, node
);
2349 streamer_write_uhwi_stream (ob
->main_stream
, node_ref
);
2351 streamer_write_uhwi (ob
, (*item
)->get_hash ());
2355 streamer_write_char_stream (ob
->main_stream
, 0);
2356 produce_asm (ob
, NULL
);
2357 destroy_output_block (ob
);
2360 /* Reads a section from LTO stream file FILE_DATA. Input block for DATA
2361 contains LEN bytes. */
2364 sem_item_optimizer::read_section (lto_file_decl_data
*file_data
,
2365 const char *data
, size_t len
)
2367 const lto_function_header
*header
2368 = (const lto_function_header
*) data
;
2369 const int cfg_offset
= sizeof (lto_function_header
);
2370 const int main_offset
= cfg_offset
+ header
->cfg_size
;
2371 const int string_offset
= main_offset
+ header
->main_size
;
2376 lto_input_block
ib_main ((const char *) data
+ main_offset
, 0,
2377 header
->main_size
, file_data
->mode_table
);
2380 = lto_data_in_create (file_data
, (const char *) data
+ string_offset
,
2381 header
->string_size
, vNULL
);
2383 count
= streamer_read_uhwi (&ib_main
);
2385 for (i
= 0; i
< count
; i
++)
2389 lto_symtab_encoder_t encoder
;
2391 index
= streamer_read_uhwi (&ib_main
);
2392 encoder
= file_data
->symtab_node_encoder
;
2393 node
= lto_symtab_encoder_deref (encoder
, index
);
2395 hashval_t hash
= streamer_read_uhwi (&ib_main
);
2397 gcc_assert (node
->definition
);
2400 fprintf (dump_file
, "Symbol added: %s (tree: %p)\n",
2401 node
->dump_asm_name (), (void *) node
->decl
);
2403 if (is_a
<cgraph_node
*> (node
))
2405 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (node
);
2407 sem_function
*fn
= new sem_function (cnode
, &m_bmstack
);
2408 fn
->set_hash (hash
);
2409 m_items
.safe_push (fn
);
2413 varpool_node
*vnode
= dyn_cast
<varpool_node
*> (node
);
2415 sem_variable
*var
= new sem_variable (vnode
, &m_bmstack
);
2416 var
->set_hash (hash
);
2417 m_items
.safe_push (var
);
2421 lto_free_section_data (file_data
, LTO_section_ipa_icf
, NULL
, data
,
2423 lto_data_in_delete (data_in
);
2426 /* Read IPA ICF summary for symbols. */
2429 sem_item_optimizer::read_summary (void)
2431 lto_file_decl_data
**file_data_vec
= lto_get_file_decl_data ();
2432 lto_file_decl_data
*file_data
;
2435 while ((file_data
= file_data_vec
[j
++]))
2438 const char *data
= lto_get_section_data (file_data
,
2439 LTO_section_ipa_icf
, NULL
, &len
);
2442 read_section (file_data
, data
, len
);
2446 /* Register callgraph and varpool hooks. */
2449 sem_item_optimizer::register_hooks (void)
2451 if (!m_cgraph_node_hooks
)
2452 m_cgraph_node_hooks
= symtab
->add_cgraph_removal_hook
2453 (&sem_item_optimizer::cgraph_removal_hook
, this);
2455 if (!m_varpool_node_hooks
)
2456 m_varpool_node_hooks
= symtab
->add_varpool_removal_hook
2457 (&sem_item_optimizer::varpool_removal_hook
, this);
2460 /* Unregister callgraph and varpool hooks. */
2463 sem_item_optimizer::unregister_hooks (void)
2465 if (m_cgraph_node_hooks
)
2466 symtab
->remove_cgraph_removal_hook (m_cgraph_node_hooks
);
2468 if (m_varpool_node_hooks
)
2469 symtab
->remove_varpool_removal_hook (m_varpool_node_hooks
);
2472 /* Adds a CLS to hashtable associated by hash value. */
2475 sem_item_optimizer::add_class (congruence_class
*cls
)
2477 gcc_assert (cls
->members
.length ());
2479 congruence_class_group
*group
2480 = get_group_by_hash (cls
->members
[0]->get_hash (),
2481 cls
->members
[0]->type
);
2482 group
->classes
.safe_push (cls
);
2485 /* Gets a congruence class group based on given HASH value and TYPE. */
2487 congruence_class_group
*
2488 sem_item_optimizer::get_group_by_hash (hashval_t hash
, sem_item_type type
)
2490 congruence_class_group
*item
= XNEW (congruence_class_group
);
2494 congruence_class_group
**slot
= m_classes
.find_slot (item
, INSERT
);
2500 item
->classes
.create (1);
2507 /* Callgraph removal hook called for a NODE with a custom DATA. */
2510 sem_item_optimizer::cgraph_removal_hook (cgraph_node
*node
, void *data
)
2512 sem_item_optimizer
*optimizer
= (sem_item_optimizer
*) data
;
2513 optimizer
->remove_symtab_node (node
);
2516 /* Varpool removal hook called for a NODE with a custom DATA. */
2519 sem_item_optimizer::varpool_removal_hook (varpool_node
*node
, void *data
)
2521 sem_item_optimizer
*optimizer
= (sem_item_optimizer
*) data
;
2522 optimizer
->remove_symtab_node (node
);
2525 /* Remove symtab NODE triggered by symtab removal hooks. */
2528 sem_item_optimizer::remove_symtab_node (symtab_node
*node
)
2530 gcc_assert (!m_classes
.elements ());
2532 m_removed_items_set
.add (node
);
2536 sem_item_optimizer::remove_item (sem_item
*item
)
2538 if (m_symtab_node_map
.get (item
->node
))
2539 m_symtab_node_map
.remove (item
->node
);
2543 /* Removes all callgraph and varpool nodes that are marked by symtab
2547 sem_item_optimizer::filter_removed_items (void)
2549 auto_vec
<sem_item
*> filtered
;
2551 for (unsigned int i
= 0; i
< m_items
.length(); i
++)
2553 sem_item
*item
= m_items
[i
];
2555 if (m_removed_items_set
.contains (item
->node
))
2561 if (item
->type
== FUNC
)
2563 cgraph_node
*cnode
= static_cast <sem_function
*>(item
)->get_node ();
2565 if (in_lto_p
&& (cnode
->alias
|| cnode
->body_removed
))
2568 filtered
.safe_push (item
);
2572 if (!flag_ipa_icf_variables
)
2576 /* Filter out non-readonly variables. */
2577 tree decl
= item
->decl
;
2578 if (TREE_READONLY (decl
))
2579 filtered
.safe_push (item
);
2586 /* Clean-up of released semantic items. */
2589 for (unsigned int i
= 0; i
< filtered
.length(); i
++)
2590 m_items
.safe_push (filtered
[i
]);
2593 /* Optimizer entry point which returns true in case it processes
2594 a merge operation. True is returned if there's a merge operation
2598 sem_item_optimizer::execute (void)
2600 filter_removed_items ();
2601 unregister_hooks ();
2604 update_hash_by_addr_refs ();
2605 build_hash_based_classes ();
2608 fprintf (dump_file
, "Dump after hash based groups\n");
2609 dump_cong_classes ();
2611 for (unsigned int i
= 0; i
< m_items
.length(); i
++)
2612 m_items
[i
]->init_wpa ();
2614 subdivide_classes_by_equality (true);
2617 fprintf (dump_file
, "Dump after WPA based types groups\n");
2619 dump_cong_classes ();
2621 process_cong_reduction ();
2622 checking_verify_classes ();
2625 fprintf (dump_file
, "Dump after callgraph-based congruence reduction\n");
2627 dump_cong_classes ();
2629 parse_nonsingleton_classes ();
2630 subdivide_classes_by_equality ();
2633 fprintf (dump_file
, "Dump after full equality comparison of groups\n");
2635 dump_cong_classes ();
2637 unsigned int prev_class_count
= m_classes_count
;
2639 process_cong_reduction ();
2640 dump_cong_classes ();
2641 checking_verify_classes ();
2642 bool merged_p
= merge_classes (prev_class_count
);
2644 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2645 symtab
->dump (dump_file
);
2650 /* Function responsible for visiting all potential functions and
2651 read-only variables that can be merged. */
2654 sem_item_optimizer::parse_funcs_and_vars (void)
2658 if (flag_ipa_icf_functions
)
2659 FOR_EACH_DEFINED_FUNCTION (cnode
)
2661 sem_function
*f
= sem_function::parse (cnode
, &m_bmstack
);
2664 m_items
.safe_push (f
);
2665 m_symtab_node_map
.put (cnode
, f
);
2668 fprintf (dump_file
, "Parsed function:%s\n", f
->node
->asm_name ());
2670 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2671 f
->dump_to_file (dump_file
);
2674 fprintf (dump_file
, "Not parsed function:%s\n", cnode
->asm_name ());
2677 varpool_node
*vnode
;
2679 if (flag_ipa_icf_variables
)
2680 FOR_EACH_DEFINED_VARIABLE (vnode
)
2682 sem_variable
*v
= sem_variable::parse (vnode
, &m_bmstack
);
2686 m_items
.safe_push (v
);
2687 m_symtab_node_map
.put (vnode
, v
);
2692 /* Makes pairing between a congruence class CLS and semantic ITEM. */
2695 sem_item_optimizer::add_item_to_class (congruence_class
*cls
, sem_item
*item
)
2697 item
->index_in_class
= cls
->members
.length ();
2698 cls
->members
.safe_push (item
);
2702 /* For each semantic item, append hash values of references. */
2705 sem_item_optimizer::update_hash_by_addr_refs ()
2707 /* First, append to hash sensitive references and class type if it need to
2708 be matched for ODR. */
2709 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2711 m_items
[i
]->update_hash_by_addr_refs (m_symtab_node_map
);
2712 if (m_items
[i
]->type
== FUNC
)
2714 if (TREE_CODE (TREE_TYPE (m_items
[i
]->decl
)) == METHOD_TYPE
2715 && contains_polymorphic_type_p
2716 (TYPE_METHOD_BASETYPE (TREE_TYPE (m_items
[i
]->decl
)))
2717 && (DECL_CXX_CONSTRUCTOR_P (m_items
[i
]->decl
)
2718 || (static_cast<sem_function
*> (m_items
[i
])->param_used_p (0)
2719 && static_cast<sem_function
*> (m_items
[i
])
2720 ->compare_polymorphic_p ())))
2723 = TYPE_METHOD_BASETYPE (TREE_TYPE (m_items
[i
]->decl
));
2724 inchash::hash
hstate (m_items
[i
]->get_hash ());
2726 if (TYPE_NAME (class_type
)
2727 && DECL_ASSEMBLER_NAME_SET_P (TYPE_NAME (class_type
)))
2729 (IDENTIFIER_HASH_VALUE
2730 (DECL_ASSEMBLER_NAME (TYPE_NAME (class_type
))));
2732 m_items
[i
]->set_hash (hstate
.end ());
2737 /* Once all symbols have enhanced hash value, we can append
2738 hash values of symbols that are seen by IPA ICF and are
2739 references by a semantic item. Newly computed values
2740 are saved to global_hash member variable. */
2741 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2742 m_items
[i
]->update_hash_by_local_refs (m_symtab_node_map
);
2744 /* Global hash value replace current hash values. */
2745 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2746 m_items
[i
]->set_hash (m_items
[i
]->global_hash
);
2749 /* Congruence classes are built by hash value. */
2752 sem_item_optimizer::build_hash_based_classes (void)
2754 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2756 sem_item
*item
= m_items
[i
];
2758 congruence_class_group
*group
2759 = get_group_by_hash (item
->get_hash (), item
->type
);
2761 if (!group
->classes
.length ())
2764 group
->classes
.safe_push (new congruence_class (class_id
++));
2767 add_item_to_class (group
->classes
[0], item
);
2771 /* Build references according to call graph. */
2774 sem_item_optimizer::build_graph (void)
2776 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2778 sem_item
*item
= m_items
[i
];
2779 m_symtab_node_map
.put (item
->node
, item
);
2781 /* Initialize hash values if we are not in LTO mode. */
2786 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2788 sem_item
*item
= m_items
[i
];
2790 if (item
->type
== FUNC
)
2792 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (item
->node
);
2794 cgraph_edge
*e
= cnode
->callees
;
2797 sem_item
**slot
= m_symtab_node_map
.get
2798 (e
->callee
->ultimate_alias_target ());
2800 item
->add_reference (*slot
);
2806 ipa_ref
*ref
= NULL
;
2807 for (unsigned i
= 0; item
->node
->iterate_reference (i
, ref
); i
++)
2809 sem_item
**slot
= m_symtab_node_map
.get
2810 (ref
->referred
->ultimate_alias_target ());
2812 item
->add_reference (*slot
);
2817 /* Semantic items in classes having more than one element and initialized.
2818 In case of WPA, we load function body. */
2821 sem_item_optimizer::parse_nonsingleton_classes (void)
2823 unsigned int init_called_count
= 0;
2825 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2826 if (m_items
[i
]->cls
->members
.length () > 1)
2828 m_items
[i
]->init ();
2829 init_called_count
++;
2833 fprintf (dump_file
, "Init called for %u items (%.2f%%).\n",
2835 m_items
.length () ? 100.0f
* init_called_count
/ m_items
.length ()
2839 /* Equality function for semantic items is used to subdivide existing
2840 classes. If IN_WPA, fast equality function is invoked. */
2843 sem_item_optimizer::subdivide_classes_by_equality (bool in_wpa
)
2845 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
2846 it
!= m_classes
.end (); ++it
)
2848 unsigned int class_count
= (*it
)->classes
.length ();
2850 for (unsigned i
= 0; i
< class_count
; i
++)
2852 congruence_class
*c
= (*it
)->classes
[i
];
2854 if (c
->members
.length() > 1)
2856 auto_vec
<sem_item
*> new_vector
;
2858 sem_item
*first
= c
->members
[0];
2859 new_vector
.safe_push (first
);
2861 unsigned class_split_first
= (*it
)->classes
.length ();
2863 for (unsigned j
= 1; j
< c
->members
.length (); j
++)
2865 sem_item
*item
= c
->members
[j
];
2868 = in_wpa
? first
->equals_wpa (item
, m_symtab_node_map
)
2869 : first
->equals (item
, m_symtab_node_map
);
2872 new_vector
.safe_push (item
);
2875 bool integrated
= false;
2877 for (unsigned k
= class_split_first
;
2878 k
< (*it
)->classes
.length (); k
++)
2880 sem_item
*x
= (*it
)->classes
[k
]->members
[0];
2882 = in_wpa
? x
->equals_wpa (item
, m_symtab_node_map
)
2883 : x
->equals (item
, m_symtab_node_map
);
2888 add_item_to_class ((*it
)->classes
[k
], item
);
2897 = new congruence_class (class_id
++);
2899 add_item_to_class (c
, item
);
2901 (*it
)->classes
.safe_push (c
);
2906 // We replace newly created new_vector for the class we've just
2908 c
->members
.release ();
2909 c
->members
.create (new_vector
.length ());
2911 for (unsigned int j
= 0; j
< new_vector
.length (); j
++)
2912 add_item_to_class (c
, new_vector
[j
]);
2917 checking_verify_classes ();
2920 /* Subdivide classes by address references that members of the class
2921 reference. Example can be a pair of functions that have an address
2922 taken from a function. If these addresses are different the class
2926 sem_item_optimizer::subdivide_classes_by_sensitive_refs ()
2928 typedef hash_map
<symbol_compare_hash
, vec
<sem_item
*> > subdivide_hash_map
;
2930 unsigned newly_created_classes
= 0;
2932 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
2933 it
!= m_classes
.end (); ++it
)
2935 unsigned int class_count
= (*it
)->classes
.length ();
2936 auto_vec
<congruence_class
*> new_classes
;
2938 for (unsigned i
= 0; i
< class_count
; i
++)
2940 congruence_class
*c
= (*it
)->classes
[i
];
2942 if (c
->members
.length() > 1)
2944 subdivide_hash_map split_map
;
2946 for (unsigned j
= 0; j
< c
->members
.length (); j
++)
2948 sem_item
*source_node
= c
->members
[j
];
2950 symbol_compare_collection
*collection
2951 = new symbol_compare_collection (source_node
->node
);
2954 vec
<sem_item
*> *slot
2955 = &split_map
.get_or_insert (collection
, &existed
);
2956 gcc_checking_assert (slot
);
2958 slot
->safe_push (source_node
);
2964 /* If the map contains more than one key, we have to split
2965 the map appropriately. */
2966 if (split_map
.elements () != 1)
2968 bool first_class
= true;
2970 for (subdivide_hash_map::iterator it2
= split_map
.begin ();
2971 it2
!= split_map
.end (); ++it2
)
2973 congruence_class
*new_cls
;
2974 new_cls
= new congruence_class (class_id
++);
2976 for (unsigned k
= 0; k
< (*it2
).second
.length (); k
++)
2977 add_item_to_class (new_cls
, (*it2
).second
[k
]);
2979 worklist_push (new_cls
);
2980 newly_created_classes
++;
2984 (*it
)->classes
[i
] = new_cls
;
2985 first_class
= false;
2989 new_classes
.safe_push (new_cls
);
2995 /* Release memory. */
2996 for (subdivide_hash_map::iterator it2
= split_map
.begin ();
2997 it2
!= split_map
.end (); ++it2
)
2999 delete (*it2
).first
;
3000 (*it2
).second
.release ();
3005 for (unsigned i
= 0; i
< new_classes
.length (); i
++)
3006 (*it
)->classes
.safe_push (new_classes
[i
]);
3009 return newly_created_classes
;
3012 /* Verify congruence classes, if checking is enabled. */
3015 sem_item_optimizer::checking_verify_classes (void)
3021 /* Verify congruence classes. */
3024 sem_item_optimizer::verify_classes (void)
3026 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3027 it
!= m_classes
.end (); ++it
)
3029 for (unsigned int i
= 0; i
< (*it
)->classes
.length (); i
++)
3031 congruence_class
*cls
= (*it
)->classes
[i
];
3034 gcc_assert (cls
->members
.length () > 0);
3036 for (unsigned int j
= 0; j
< cls
->members
.length (); j
++)
3038 sem_item
*item
= cls
->members
[j
];
3041 gcc_assert (item
->cls
== cls
);
3043 for (unsigned k
= 0; k
< item
->usages
.length (); k
++)
3045 sem_usage_pair
*usage
= item
->usages
[k
];
3046 gcc_assert (usage
->item
->index_in_class
3047 < usage
->item
->cls
->members
.length ());
3054 /* Disposes split map traverse function. CLS_PTR is pointer to congruence
3055 class, BSLOT is bitmap slot we want to release. DATA is mandatory,
3056 but unused argument. */
3059 sem_item_optimizer::release_split_map (congruence_class
* const &,
3060 bitmap
const &b
, traverse_split_pair
*)
3069 /* Process split operation for a class given as pointer CLS_PTR,
3070 where bitmap B splits congruence class members. DATA is used
3071 as argument of split pair. */
3074 sem_item_optimizer::traverse_congruence_split (congruence_class
* const &cls
,
3076 traverse_split_pair
*pair
)
3078 sem_item_optimizer
*optimizer
= pair
->optimizer
;
3079 const congruence_class
*splitter_cls
= pair
->cls
;
3081 /* If counted bits are greater than zero and less than the number of members
3082 a group will be splitted. */
3083 unsigned popcount
= bitmap_count_bits (b
);
3085 if (popcount
> 0 && popcount
< cls
->members
.length ())
3087 auto_vec
<congruence_class
*, 2> newclasses
;
3088 newclasses
.quick_push (new congruence_class (class_id
++));
3089 newclasses
.quick_push (new congruence_class (class_id
++));
3091 for (unsigned int i
= 0; i
< cls
->members
.length (); i
++)
3093 int target
= bitmap_bit_p (b
, i
);
3094 congruence_class
*tc
= newclasses
[target
];
3096 add_item_to_class (tc
, cls
->members
[i
]);
3101 for (unsigned int i
= 0; i
< 2; i
++)
3102 gcc_assert (newclasses
[i
]->members
.length ());
3105 if (splitter_cls
== cls
)
3106 optimizer
->splitter_class_removed
= true;
3108 /* Remove old class from worklist if presented. */
3109 bool in_worklist
= cls
->in_worklist
;
3112 cls
->in_worklist
= false;
3114 congruence_class_group g
;
3115 g
.hash
= cls
->members
[0]->get_hash ();
3116 g
.type
= cls
->members
[0]->type
;
3118 congruence_class_group
*slot
= optimizer
->m_classes
.find (&g
);
3120 for (unsigned int i
= 0; i
< slot
->classes
.length (); i
++)
3121 if (slot
->classes
[i
] == cls
)
3123 slot
->classes
.ordered_remove (i
);
3127 /* New class will be inserted and integrated to work list. */
3128 for (unsigned int i
= 0; i
< 2; i
++)
3129 optimizer
->add_class (newclasses
[i
]);
3131 /* Two classes replace one, so that increment just by one. */
3132 optimizer
->m_classes_count
++;
3134 /* If OLD class was presented in the worklist, we remove the class
3135 and replace it will both newly created classes. */
3137 for (unsigned int i
= 0; i
< 2; i
++)
3138 optimizer
->worklist_push (newclasses
[i
]);
3139 else /* Just smaller class is inserted. */
3141 unsigned int smaller_index
3142 = (newclasses
[0]->members
.length ()
3143 < newclasses
[1]->members
.length ()
3145 optimizer
->worklist_push (newclasses
[smaller_index
]);
3148 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3150 fprintf (dump_file
, " congruence class splitted:\n");
3151 cls
->dump (dump_file
, 4);
3153 fprintf (dump_file
, " newly created groups:\n");
3154 for (unsigned int i
= 0; i
< 2; i
++)
3155 newclasses
[i
]->dump (dump_file
, 4);
3158 /* Release class if not presented in work list. */
3167 /* Tests if a class CLS used as INDEXth splits any congruence classes.
3168 Bitmap stack BMSTACK is used for bitmap allocation. */
3171 sem_item_optimizer::do_congruence_step_for_index (congruence_class
*cls
,
3174 hash_map
<congruence_class
*, bitmap
> split_map
;
3176 for (unsigned int i
= 0; i
< cls
->members
.length (); i
++)
3178 sem_item
*item
= cls
->members
[i
];
3180 /* Iterate all usages that have INDEX as usage of the item. */
3181 for (unsigned int j
= 0; j
< item
->usages
.length (); j
++)
3183 sem_usage_pair
*usage
= item
->usages
[j
];
3185 if (usage
->index
!= index
)
3188 bitmap
*slot
= split_map
.get (usage
->item
->cls
);
3193 b
= BITMAP_ALLOC (&m_bmstack
);
3194 split_map
.put (usage
->item
->cls
, b
);
3199 gcc_checking_assert (usage
->item
->cls
);
3200 gcc_checking_assert (usage
->item
->index_in_class
3201 < usage
->item
->cls
->members
.length ());
3203 bitmap_set_bit (b
, usage
->item
->index_in_class
);
3207 traverse_split_pair pair
;
3208 pair
.optimizer
= this;
3211 splitter_class_removed
= false;
3212 split_map
.traverse
<traverse_split_pair
*,
3213 sem_item_optimizer::traverse_congruence_split
> (&pair
);
3215 /* Bitmap clean-up. */
3216 split_map
.traverse
<traverse_split_pair
*,
3217 sem_item_optimizer::release_split_map
> (NULL
);
3220 /* Every usage of a congruence class CLS is a candidate that can split the
3221 collection of classes. Bitmap stack BMSTACK is used for bitmap
3225 sem_item_optimizer::do_congruence_step (congruence_class
*cls
)
3230 bitmap usage
= BITMAP_ALLOC (&m_bmstack
);
3232 for (unsigned int i
= 0; i
< cls
->members
.length (); i
++)
3233 bitmap_ior_into (usage
, cls
->members
[i
]->usage_index_bitmap
);
3235 EXECUTE_IF_SET_IN_BITMAP (usage
, 0, i
, bi
)
3237 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3238 fprintf (dump_file
, " processing congruence step for class: %u, "
3239 "index: %u\n", cls
->id
, i
);
3241 do_congruence_step_for_index (cls
, i
);
3243 if (splitter_class_removed
)
3247 BITMAP_FREE (usage
);
3250 /* Adds a newly created congruence class CLS to worklist. */
3253 sem_item_optimizer::worklist_push (congruence_class
*cls
)
3255 /* Return if the class CLS is already presented in work list. */
3256 if (cls
->in_worklist
)
3259 cls
->in_worklist
= true;
3260 worklist
.push_back (cls
);
3263 /* Pops a class from worklist. */
3266 sem_item_optimizer::worklist_pop (void)
3268 congruence_class
*cls
;
3270 while (!worklist
.empty ())
3272 cls
= worklist
.front ();
3273 worklist
.pop_front ();
3274 if (cls
->in_worklist
)
3276 cls
->in_worklist
= false;
3282 /* Work list item was already intended to be removed.
3283 The only reason for doing it is to split a class.
3284 Thus, the class CLS is deleted. */
3292 /* Iterative congruence reduction function. */
3295 sem_item_optimizer::process_cong_reduction (void)
3297 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3298 it
!= m_classes
.end (); ++it
)
3299 for (unsigned i
= 0; i
< (*it
)->classes
.length (); i
++)
3300 if ((*it
)->classes
[i
]->is_class_used ())
3301 worklist_push ((*it
)->classes
[i
]);
3304 fprintf (dump_file
, "Worklist has been filled with: %lu\n",
3305 (unsigned long) worklist
.size ());
3307 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3308 fprintf (dump_file
, "Congruence class reduction\n");
3310 congruence_class
*cls
;
3312 /* Process complete congruence reduction. */
3313 while ((cls
= worklist_pop ()) != NULL
)
3314 do_congruence_step (cls
);
3316 /* Subdivide newly created classes according to references. */
3317 unsigned new_classes
= subdivide_classes_by_sensitive_refs ();
3320 fprintf (dump_file
, "Address reference subdivision created: %u "
3321 "new classes.\n", new_classes
);
3324 /* Debug function prints all informations about congruence classes. */
3327 sem_item_optimizer::dump_cong_classes (void)
3333 "Congruence classes: %u (unique hash values: %lu), with total: "
3334 "%u items\n", m_classes_count
,
3335 (unsigned long) m_classes
.elements (), m_items
.length ());
3337 /* Histogram calculation. */
3338 unsigned int max_index
= 0;
3339 unsigned int* histogram
= XCNEWVEC (unsigned int, m_items
.length () + 1);
3341 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3342 it
!= m_classes
.end (); ++it
)
3343 for (unsigned i
= 0; i
< (*it
)->classes
.length (); i
++)
3345 unsigned int c
= (*it
)->classes
[i
]->members
.length ();
3353 "Class size histogram [num of members]: number of classe number "
3356 for (unsigned int i
= 0; i
<= max_index
; i
++)
3358 fprintf (dump_file
, "[%u]: %u classes\n", i
, histogram
[i
]);
3360 fprintf (dump_file
, "\n\n");
3362 if (dump_flags
& TDF_DETAILS
)
3363 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3364 it
!= m_classes
.end (); ++it
)
3366 fprintf (dump_file
, " group: with %u classes:\n",
3367 (*it
)->classes
.length ());
3369 for (unsigned i
= 0; i
< (*it
)->classes
.length (); i
++)
3371 (*it
)->classes
[i
]->dump (dump_file
, 4);
3373 if (i
< (*it
)->classes
.length () - 1)
3374 fprintf (dump_file
, " ");
3381 /* Sort pair of sem_items A and B by DECL_UID. */
3384 sort_sem_items_by_decl_uid (const void *a
, const void *b
)
3386 const sem_item
*i1
= *(const sem_item
* const *)a
;
3387 const sem_item
*i2
= *(const sem_item
* const *)b
;
3389 int uid1
= DECL_UID (i1
->decl
);
3390 int uid2
= DECL_UID (i2
->decl
);
3394 else if (uid1
> uid2
)
3400 /* Sort pair of congruence_classes A and B by DECL_UID of the first member. */
3403 sort_congruence_classes_by_decl_uid (const void *a
, const void *b
)
3405 const congruence_class
*c1
= *(const congruence_class
* const *)a
;
3406 const congruence_class
*c2
= *(const congruence_class
* const *)b
;
3408 int uid1
= DECL_UID (c1
->members
[0]->decl
);
3409 int uid2
= DECL_UID (c2
->members
[0]->decl
);
3413 else if (uid1
> uid2
)
3419 /* Sort pair of congruence_class_groups A and B by
3420 DECL_UID of the first member of a first group. */
3423 sort_congruence_class_groups_by_decl_uid (const void *a
, const void *b
)
3425 const congruence_class_group
*g1
3426 = *(const congruence_class_group
* const *)a
;
3427 const congruence_class_group
*g2
3428 = *(const congruence_class_group
* const *)b
;
3430 int uid1
= DECL_UID (g1
->classes
[0]->members
[0]->decl
);
3431 int uid2
= DECL_UID (g2
->classes
[0]->members
[0]->decl
);
3435 else if (uid1
> uid2
)
3441 /* After reduction is done, we can declare all items in a group
3442 to be equal. PREV_CLASS_COUNT is start number of classes
3443 before reduction. True is returned if there's a merge operation
3447 sem_item_optimizer::merge_classes (unsigned int prev_class_count
)
3449 unsigned int item_count
= m_items
.length ();
3450 unsigned int class_count
= m_classes_count
;
3451 unsigned int equal_items
= item_count
- class_count
;
3453 unsigned int non_singular_classes_count
= 0;
3454 unsigned int non_singular_classes_sum
= 0;
3456 bool merged_p
= false;
3459 Sort functions in congruence classes by DECL_UID and do the same
3460 for the classes to not to break -fcompare-debug. */
3462 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3463 it
!= m_classes
.end (); ++it
)
3465 for (unsigned int i
= 0; i
< (*it
)->classes
.length (); i
++)
3467 congruence_class
*c
= (*it
)->classes
[i
];
3468 c
->members
.qsort (sort_sem_items_by_decl_uid
);
3471 (*it
)->classes
.qsort (sort_congruence_classes_by_decl_uid
);
3474 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3475 it
!= m_classes
.end (); ++it
)
3476 for (unsigned int i
= 0; i
< (*it
)->classes
.length (); i
++)
3478 congruence_class
*c
= (*it
)->classes
[i
];
3479 if (c
->members
.length () > 1)
3481 non_singular_classes_count
++;
3482 non_singular_classes_sum
+= c
->members
.length ();
3486 auto_vec
<congruence_class_group
*> classes (m_classes
.elements ());
3487 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3488 it
!= m_classes
.end (); ++it
)
3489 classes
.quick_push (*it
);
3491 classes
.qsort (sort_congruence_class_groups_by_decl_uid
);
3495 fprintf (dump_file
, "\nItem count: %u\n", item_count
);
3496 fprintf (dump_file
, "Congruent classes before: %u, after: %u\n",
3497 prev_class_count
, class_count
);
3498 fprintf (dump_file
, "Average class size before: %.2f, after: %.2f\n",
3499 prev_class_count
? 1.0f
* item_count
/ prev_class_count
: 0.0f
,
3500 class_count
? 1.0f
* item_count
/ class_count
: 0.0f
);
3501 fprintf (dump_file
, "Average non-singular class size: %.2f, count: %u\n",
3502 non_singular_classes_count
? 1.0f
* non_singular_classes_sum
/
3503 non_singular_classes_count
: 0.0f
,
3504 non_singular_classes_count
);
3505 fprintf (dump_file
, "Equal symbols: %u\n", equal_items
);
3506 fprintf (dump_file
, "Fraction of visited symbols: %.2f%%\n\n",
3507 item_count
? 100.0f
* equal_items
/ item_count
: 0.0f
);
3511 congruence_class_group
*it
;
3512 FOR_EACH_VEC_ELT (classes
, l
, it
)
3513 for (unsigned int i
= 0; i
< it
->classes
.length (); i
++)
3515 congruence_class
*c
= it
->classes
[i
];
3517 if (c
->members
.length () == 1)
3520 sem_item
*source
= c
->members
[0];
3522 if (DECL_NAME (source
->decl
)
3523 && MAIN_NAME_P (DECL_NAME (source
->decl
)))
3524 /* If merge via wrappers, picking main as the target can be
3526 source
= c
->members
[1];
3528 for (unsigned int j
= 0; j
< c
->members
.length (); j
++)
3530 sem_item
*alias
= c
->members
[j
];
3532 if (alias
== source
)
3537 fprintf (dump_file
, "Semantic equality hit:%s->%s\n",
3538 xstrdup_for_dump (source
->node
->name ()),
3539 xstrdup_for_dump (alias
->node
->name ()));
3540 fprintf (dump_file
, "Assembler symbol names:%s->%s\n",
3541 xstrdup_for_dump (source
->node
->asm_name ()),
3542 xstrdup_for_dump (alias
->node
->asm_name ()));
3545 if (lookup_attribute ("no_icf", DECL_ATTRIBUTES (alias
->decl
)))
3549 "Merge operation is skipped due to no_icf "
3555 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3557 source
->dump_to_file (dump_file
);
3558 alias
->dump_to_file (dump_file
);
3561 if (dbg_cnt (merged_ipa_icf
))
3563 bool merged
= source
->merge (alias
);
3566 if (merged
&& alias
->type
== VAR
)
3568 symtab_pair p
= symtab_pair (source
->node
, alias
->node
);
3569 m_merged_variables
.safe_push (p
);
3575 if (!m_merged_variables
.is_empty ())
3576 fixup_points_to_sets ();
3581 /* Fixup points to set PT. */
3584 sem_item_optimizer::fixup_pt_set (struct pt_solution
*pt
)
3586 if (pt
->vars
== NULL
)
3591 FOR_EACH_VEC_ELT (m_merged_variables
, i
, item
)
3592 if (bitmap_bit_p (pt
->vars
, DECL_UID (item
->second
->decl
)))
3593 bitmap_set_bit (pt
->vars
, DECL_UID (item
->first
->decl
));
3596 /* Set all points-to UIDs of aliases pointing to node N as UID. */
3599 set_alias_uids (symtab_node
*n
, int uid
)
3602 FOR_EACH_ALIAS (n
, ref
)
3605 fprintf (dump_file
, " Setting points-to UID of [%s] as %d\n",
3606 xstrdup_for_dump (ref
->referring
->asm_name ()), uid
);
3608 SET_DECL_PT_UID (ref
->referring
->decl
, uid
);
3609 set_alias_uids (ref
->referring
, uid
);
3613 /* Fixup points to analysis info. */
3616 sem_item_optimizer::fixup_points_to_sets (void)
3618 /* TODO: remove in GCC 9 and trigger PTA re-creation after IPA passes. */
3621 FOR_EACH_DEFINED_FUNCTION (cnode
)
3625 function
*fn
= DECL_STRUCT_FUNCTION (cnode
->decl
);
3626 if (!gimple_in_ssa_p (fn
))
3629 FOR_EACH_SSA_NAME (i
, name
, fn
)
3630 if (POINTER_TYPE_P (TREE_TYPE (name
))
3631 && SSA_NAME_PTR_INFO (name
))
3632 fixup_pt_set (&SSA_NAME_PTR_INFO (name
)->pt
);
3633 fixup_pt_set (&fn
->gimple_df
->escaped
);
3635 /* The above get's us to 99% I guess, at least catching the
3636 address compares. Below also gets us aliasing correct
3637 but as said we're giving leeway to the situation with
3638 readonly vars anyway, so ... */
3640 FOR_EACH_BB_FN (bb
, fn
)
3641 for (gimple_stmt_iterator gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
);
3644 gcall
*call
= dyn_cast
<gcall
*> (gsi_stmt (gsi
));
3647 fixup_pt_set (gimple_call_use_set (call
));
3648 fixup_pt_set (gimple_call_clobber_set (call
));
3655 FOR_EACH_VEC_ELT (m_merged_variables
, i
, item
)
3656 set_alias_uids (item
->first
, DECL_UID (item
->first
->decl
));
3659 /* Dump function prints all class members to a FILE with an INDENT. */
3662 congruence_class::dump (FILE *file
, unsigned int indent
) const
3664 FPRINTF_SPACES (file
, indent
, "class with id: %u, hash: %u, items: %u\n",
3665 id
, members
[0]->get_hash (), members
.length ());
3667 FPUTS_SPACES (file
, indent
+ 2, "");
3668 for (unsigned i
= 0; i
< members
.length (); i
++)
3669 fprintf (file
, "%s ", members
[i
]->node
->dump_asm_name ());
3671 fprintf (file
, "\n");
3674 /* Returns true if there's a member that is used from another group. */
3677 congruence_class::is_class_used (void)
3679 for (unsigned int i
= 0; i
< members
.length (); i
++)
3680 if (members
[i
]->usages
.length ())
3686 /* Generate pass summary for IPA ICF pass. */
3689 ipa_icf_generate_summary (void)
3692 optimizer
= new sem_item_optimizer ();
3694 optimizer
->register_hooks ();
3695 optimizer
->parse_funcs_and_vars ();
3698 /* Write pass summary for IPA ICF pass. */
3701 ipa_icf_write_summary (void)
3703 gcc_assert (optimizer
);
3705 optimizer
->write_summary ();
3708 /* Read pass summary for IPA ICF pass. */
3711 ipa_icf_read_summary (void)
3714 optimizer
= new sem_item_optimizer ();
3716 optimizer
->read_summary ();
3717 optimizer
->register_hooks ();
3720 /* Semantic equality exection function. */
3723 ipa_icf_driver (void)
3725 gcc_assert (optimizer
);
3727 bool merged_p
= optimizer
->execute ();
3732 return merged_p
? TODO_remove_functions
: 0;
3735 const pass_data pass_data_ipa_icf
=
3737 IPA_PASS
, /* type */
3739 OPTGROUP_IPA
, /* optinfo_flags */
3740 TV_IPA_ICF
, /* tv_id */
3741 0, /* properties_required */
3742 0, /* properties_provided */
3743 0, /* properties_destroyed */
3744 0, /* todo_flags_start */
3745 0, /* todo_flags_finish */
3748 class pass_ipa_icf
: public ipa_opt_pass_d
3751 pass_ipa_icf (gcc::context
*ctxt
)
3752 : ipa_opt_pass_d (pass_data_ipa_icf
, ctxt
,
3753 ipa_icf_generate_summary
, /* generate_summary */
3754 ipa_icf_write_summary
, /* write_summary */
3755 ipa_icf_read_summary
, /* read_summary */
3757 write_optimization_summary */
3759 read_optimization_summary */
3760 NULL
, /* stmt_fixup */
3761 0, /* function_transform_todo_flags_start */
3762 NULL
, /* function_transform */
3763 NULL
) /* variable_transform */
3766 /* opt_pass methods: */
3767 virtual bool gate (function
*)
3769 return in_lto_p
|| flag_ipa_icf_variables
|| flag_ipa_icf_functions
;
3772 virtual unsigned int execute (function
*)
3774 return ipa_icf_driver();
3776 }; // class pass_ipa_icf
3778 } // ipa_icf namespace
3781 make_pass_ipa_icf (gcc::context
*ctxt
)
3783 return new ipa_icf::pass_ipa_icf (ctxt
);