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
&& memcmp (opt1
, opt2
, sizeof(cl_optimization
)))
663 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
665 fprintf (dump_file
, "optimization flags difference");
666 cl_optimization_print_diff (dump_file
, 2, opt1
, opt2
);
669 return return_false_with_msg ("optimization flags are different");
672 /* Result type checking. */
673 if (!func_checker::compatible_types_p
674 (TREE_TYPE (TREE_TYPE (decl
)),
675 TREE_TYPE (TREE_TYPE (m_compared_func
->decl
))))
676 return return_false_with_msg ("result types are different");
678 /* Checking types of arguments. */
679 tree list1
= TYPE_ARG_TYPES (TREE_TYPE (decl
)),
680 list2
= TYPE_ARG_TYPES (TREE_TYPE (m_compared_func
->decl
));
681 for (unsigned i
= 0; list1
&& list2
;
682 list1
= TREE_CHAIN (list1
), list2
= TREE_CHAIN (list2
), i
++)
684 tree parm1
= TREE_VALUE (list1
);
685 tree parm2
= TREE_VALUE (list2
);
687 /* This guard is here for function pointer with attributes (pr59927.c). */
688 if (!parm1
|| !parm2
)
689 return return_false_with_msg ("NULL argument type");
691 /* Verify that types are compatible to ensure that both functions
692 have same calling conventions. */
693 if (!types_compatible_p (parm1
, parm2
))
694 return return_false_with_msg ("parameter types are not compatible");
696 if (!param_used_p (i
))
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
)->self_size
<= 2)
1205 fprintf (dump_file
, "Wrapper creation is not "
1206 "profitable (function is too small).\n");
1208 /* If user paid attention to mark function noinline, assume it is
1209 somewhat special and do not try to turn it into a wrapper that can
1210 not be undone by inliner. */
1211 else if (lookup_attribute ("noinline", DECL_ATTRIBUTES (alias
->decl
)))
1214 fprintf (dump_file
, "Wrappers are not created for noinline.\n");
1217 create_wrapper
= true;
1219 /* We can redirect local calls in the case both alias and orignal
1220 are not interposable. */
1222 = alias
->get_availability () > AVAIL_INTERPOSABLE
1223 && original
->get_availability () > AVAIL_INTERPOSABLE
1224 && !alias
->instrumented_version
;
1225 /* TODO: We can redirect, but we need to produce alias of ORIGINAL
1226 with proper properties. */
1227 if (!sem_item::compare_referenced_symbol_properties (NULL
, original
, alias
,
1228 alias
->address_taken
))
1229 redirect_callers
= false;
1231 if (!redirect_callers
&& !create_wrapper
)
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 gcc_checking_assert (COMPLETE_TYPE_P (type
));
1584 hashval_t
*val
= optimizer
->m_type_hash_cache
.get (type
);
1588 inchash::hash hstate2
;
1593 hstate2
.add_int (RECORD_TYPE
);
1594 gcc_assert (COMPLETE_TYPE_P (type
));
1596 for (f
= TYPE_FIELDS (type
), nf
= 0; f
; f
= TREE_CHAIN (f
))
1597 if (TREE_CODE (f
) == FIELD_DECL
)
1599 add_type (TREE_TYPE (f
), hstate2
);
1603 hstate2
.add_int (nf
);
1604 hash
= hstate2
.end ();
1605 hstate
.add_hwi (hash
);
1606 optimizer
->m_type_hash_cache
.put (type
, hash
);
1609 hstate
.add_hwi (*val
);
1613 /* Improve accumulated hash for HSTATE based on a gimple statement STMT. */
1616 sem_function::hash_stmt (gimple
*stmt
, inchash::hash
&hstate
)
1618 enum gimple_code code
= gimple_code (stmt
);
1620 hstate
.add_int (code
);
1625 add_expr (gimple_switch_index (as_a
<gswitch
*> (stmt
)), hstate
);
1628 hstate
.add_int (gimple_assign_rhs_code (stmt
));
1629 if (commutative_tree_code (gimple_assign_rhs_code (stmt
))
1630 || commutative_ternary_tree_code (gimple_assign_rhs_code (stmt
)))
1632 inchash::hash one
, two
;
1634 add_expr (gimple_assign_rhs1 (stmt
), one
);
1635 add_type (TREE_TYPE (gimple_assign_rhs1 (stmt
)), one
);
1636 add_expr (gimple_assign_rhs2 (stmt
), two
);
1637 hstate
.add_commutative (one
, two
);
1638 if (commutative_ternary_tree_code (gimple_assign_rhs_code (stmt
)))
1640 add_expr (gimple_assign_rhs3 (stmt
), hstate
);
1641 add_type (TREE_TYPE (gimple_assign_rhs3 (stmt
)), hstate
);
1643 add_expr (gimple_assign_lhs (stmt
), hstate
);
1644 add_type (TREE_TYPE (gimple_assign_lhs (stmt
)), two
);
1653 /* All these statements are equivalent if their operands are. */
1654 for (unsigned i
= 0; i
< gimple_num_ops (stmt
); ++i
)
1656 add_expr (gimple_op (stmt
, i
), hstate
);
1657 if (gimple_op (stmt
, i
))
1658 add_type (TREE_TYPE (gimple_op (stmt
, i
)), hstate
);
1660 /* Consider nocf_check attribute in hash as it affects code
1662 if (code
== GIMPLE_CALL
1663 && flag_cf_protection
& CF_BRANCH
)
1664 hstate
.add_flag (gimple_call_nocf_check_p (as_a
<gcall
*> (stmt
)));
1671 /* Return true if polymorphic comparison must be processed. */
1674 sem_function::compare_polymorphic_p (void)
1676 struct cgraph_edge
*e
;
1678 if (!opt_for_fn (get_node ()->decl
, flag_devirtualize
))
1680 if (get_node ()->indirect_calls
!= NULL
)
1682 /* TODO: We can do simple propagation determining what calls may lead to
1683 a polymorphic call. */
1684 for (e
= get_node ()->callees
; e
; e
= e
->next_callee
)
1685 if (e
->callee
->definition
1686 && opt_for_fn (e
->callee
->decl
, flag_devirtualize
))
1691 /* For a given call graph NODE, the function constructs new
1692 semantic function item. */
1695 sem_function::parse (cgraph_node
*node
, bitmap_obstack
*stack
)
1697 tree fndecl
= node
->decl
;
1698 function
*func
= DECL_STRUCT_FUNCTION (fndecl
);
1700 if (!func
|| (!node
->has_gimple_body_p () && !node
->thunk
.thunk_p
))
1703 if (lookup_attribute_by_prefix ("omp ", DECL_ATTRIBUTES (node
->decl
)) != NULL
)
1706 if (lookup_attribute_by_prefix ("oacc ",
1707 DECL_ATTRIBUTES (node
->decl
)) != NULL
)
1711 if (DECL_STATIC_CONSTRUCTOR (node
->decl
)
1712 || DECL_STATIC_DESTRUCTOR (node
->decl
))
1715 sem_function
*f
= new sem_function (node
, stack
);
1722 /* For given basic blocks BB1 and BB2 (from functions FUNC1 and FUNC),
1723 return true if phi nodes are semantically equivalent in these blocks . */
1726 sem_function::compare_phi_node (basic_block bb1
, basic_block bb2
)
1728 gphi_iterator si1
, si2
;
1730 unsigned size1
, size2
, i
;
1734 gcc_assert (bb1
!= NULL
);
1735 gcc_assert (bb2
!= NULL
);
1737 si2
= gsi_start_phis (bb2
);
1738 for (si1
= gsi_start_phis (bb1
); !gsi_end_p (si1
);
1741 gsi_next_nonvirtual_phi (&si1
);
1742 gsi_next_nonvirtual_phi (&si2
);
1744 if (gsi_end_p (si1
) && gsi_end_p (si2
))
1747 if (gsi_end_p (si1
) || gsi_end_p (si2
))
1748 return return_false();
1753 tree phi_result1
= gimple_phi_result (phi1
);
1754 tree phi_result2
= gimple_phi_result (phi2
);
1756 if (!m_checker
->compare_operand (phi_result1
, phi_result2
))
1757 return return_false_with_msg ("PHI results are different");
1759 size1
= gimple_phi_num_args (phi1
);
1760 size2
= gimple_phi_num_args (phi2
);
1763 return return_false ();
1765 for (i
= 0; i
< size1
; ++i
)
1767 t1
= gimple_phi_arg (phi1
, i
)->def
;
1768 t2
= gimple_phi_arg (phi2
, i
)->def
;
1770 if (!m_checker
->compare_operand (t1
, t2
))
1771 return return_false ();
1773 e1
= gimple_phi_arg_edge (phi1
, i
);
1774 e2
= gimple_phi_arg_edge (phi2
, i
);
1776 if (!m_checker
->compare_edge (e1
, e2
))
1777 return return_false ();
1786 /* Returns true if tree T can be compared as a handled component. */
1789 sem_function::icf_handled_component_p (tree t
)
1791 tree_code tc
= TREE_CODE (t
);
1793 return (handled_component_p (t
)
1794 || tc
== ADDR_EXPR
|| tc
== MEM_REF
|| tc
== OBJ_TYPE_REF
);
1797 /* Basic blocks dictionary BB_DICT returns true if SOURCE index BB
1798 corresponds to TARGET. */
1801 sem_function::bb_dict_test (vec
<int> *bb_dict
, int source
, int target
)
1806 if (bb_dict
->length () <= (unsigned)source
)
1807 bb_dict
->safe_grow_cleared (source
+ 1);
1809 if ((*bb_dict
)[source
] == 0)
1811 (*bb_dict
)[source
] = target
;
1815 return (*bb_dict
)[source
] == target
;
1818 sem_variable::sem_variable (bitmap_obstack
*stack
): sem_item (VAR
, stack
)
1822 sem_variable::sem_variable (varpool_node
*node
, bitmap_obstack
*stack
)
1823 : sem_item (VAR
, node
, stack
)
1825 gcc_checking_assert (node
);
1826 gcc_checking_assert (get_node ());
1829 /* Fast equality function based on knowledge known in WPA. */
1832 sem_variable::equals_wpa (sem_item
*item
,
1833 hash_map
<symtab_node
*, sem_item
*> &ignored_nodes
)
1835 gcc_assert (item
->type
== VAR
);
1837 if (node
->num_references () != item
->node
->num_references ())
1838 return return_false_with_msg ("different number of references");
1840 if (DECL_TLS_MODEL (decl
) || DECL_TLS_MODEL (item
->decl
))
1841 return return_false_with_msg ("TLS model");
1843 /* DECL_ALIGN is safe to merge, because we will always chose the largest
1844 alignment out of all aliases. */
1846 if (DECL_VIRTUAL_P (decl
) != DECL_VIRTUAL_P (item
->decl
))
1847 return return_false_with_msg ("Virtual flag mismatch");
1849 if (DECL_SIZE (decl
) != DECL_SIZE (item
->decl
)
1850 && ((!DECL_SIZE (decl
) || !DECL_SIZE (item
->decl
))
1851 || !operand_equal_p (DECL_SIZE (decl
),
1852 DECL_SIZE (item
->decl
), OEP_ONLY_CONST
)))
1853 return return_false_with_msg ("size mismatch");
1855 /* Do not attempt to mix data from different user sections;
1856 we do not know what user intends with those. */
1857 if (((DECL_SECTION_NAME (decl
) && !node
->implicit_section
)
1858 || (DECL_SECTION_NAME (item
->decl
) && !item
->node
->implicit_section
))
1859 && DECL_SECTION_NAME (decl
) != DECL_SECTION_NAME (item
->decl
))
1860 return return_false_with_msg ("user section mismatch");
1862 if (DECL_IN_TEXT_SECTION (decl
) != DECL_IN_TEXT_SECTION (item
->decl
))
1863 return return_false_with_msg ("text section");
1865 ipa_ref
*ref
= NULL
, *ref2
= NULL
;
1866 for (unsigned i
= 0; node
->iterate_reference (i
, ref
); i
++)
1868 item
->node
->iterate_reference (i
, ref2
);
1870 if (ref
->use
!= ref2
->use
)
1871 return return_false_with_msg ("reference use mismatch");
1873 if (!compare_symbol_references (ignored_nodes
,
1874 ref
->referred
, ref2
->referred
,
1875 ref
->address_matters_p ()))
1882 /* Returns true if the item equals to ITEM given as argument. */
1885 sem_variable::equals (sem_item
*item
,
1886 hash_map
<symtab_node
*, sem_item
*> &)
1888 gcc_assert (item
->type
== VAR
);
1891 if (DECL_INITIAL (decl
) == error_mark_node
&& in_lto_p
)
1892 dyn_cast
<varpool_node
*>(node
)->get_constructor ();
1893 if (DECL_INITIAL (item
->decl
) == error_mark_node
&& in_lto_p
)
1894 dyn_cast
<varpool_node
*>(item
->node
)->get_constructor ();
1896 /* As seen in PR ipa/65303 we have to compare variables types. */
1897 if (!func_checker::compatible_types_p (TREE_TYPE (decl
),
1898 TREE_TYPE (item
->decl
)))
1899 return return_false_with_msg ("variables types are different");
1901 ret
= sem_variable::equals (DECL_INITIAL (decl
),
1902 DECL_INITIAL (item
->node
->decl
));
1903 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1905 "Equals called for vars: %s:%s with result: %s\n\n",
1906 node
->dump_name (), item
->node
->dump_name (),
1907 ret
? "true" : "false");
1912 /* Compares trees T1 and T2 for semantic equality. */
1915 sem_variable::equals (tree t1
, tree t2
)
1918 return return_with_debug (t1
== t2
);
1921 tree_code tc1
= TREE_CODE (t1
);
1922 tree_code tc2
= TREE_CODE (t2
);
1925 return return_false_with_msg ("TREE_CODE mismatch");
1931 vec
<constructor_elt
, va_gc
> *v1
, *v2
;
1932 unsigned HOST_WIDE_INT idx
;
1934 enum tree_code typecode
= TREE_CODE (TREE_TYPE (t1
));
1935 if (typecode
!= TREE_CODE (TREE_TYPE (t2
)))
1936 return return_false_with_msg ("constructor type mismatch");
1938 if (typecode
== ARRAY_TYPE
)
1940 HOST_WIDE_INT size_1
= int_size_in_bytes (TREE_TYPE (t1
));
1941 /* For arrays, check that the sizes all match. */
1942 if (TYPE_MODE (TREE_TYPE (t1
)) != TYPE_MODE (TREE_TYPE (t2
))
1944 || size_1
!= int_size_in_bytes (TREE_TYPE (t2
)))
1945 return return_false_with_msg ("constructor array size mismatch");
1947 else if (!func_checker::compatible_types_p (TREE_TYPE (t1
),
1949 return return_false_with_msg ("constructor type incompatible");
1951 v1
= CONSTRUCTOR_ELTS (t1
);
1952 v2
= CONSTRUCTOR_ELTS (t2
);
1953 if (vec_safe_length (v1
) != vec_safe_length (v2
))
1954 return return_false_with_msg ("constructor number of elts mismatch");
1956 for (idx
= 0; idx
< vec_safe_length (v1
); ++idx
)
1958 constructor_elt
*c1
= &(*v1
)[idx
];
1959 constructor_elt
*c2
= &(*v2
)[idx
];
1961 /* Check that each value is the same... */
1962 if (!sem_variable::equals (c1
->value
, c2
->value
))
1964 /* ... and that they apply to the same fields! */
1965 if (!sem_variable::equals (c1
->index
, c2
->index
))
1972 tree x1
= TREE_OPERAND (t1
, 0);
1973 tree x2
= TREE_OPERAND (t2
, 0);
1974 tree y1
= TREE_OPERAND (t1
, 1);
1975 tree y2
= TREE_OPERAND (t2
, 1);
1977 if (!func_checker::compatible_types_p (TREE_TYPE (x1
), TREE_TYPE (x2
)))
1978 return return_false ();
1980 /* Type of the offset on MEM_REF does not matter. */
1981 return return_with_debug (sem_variable::equals (x1
, x2
)
1982 && wi::to_offset (y1
)
1983 == wi::to_offset (y2
));
1988 tree op1
= TREE_OPERAND (t1
, 0);
1989 tree op2
= TREE_OPERAND (t2
, 0);
1990 return sem_variable::equals (op1
, op2
);
1992 /* References to other vars/decls are compared using ipa-ref. */
1995 if (decl_in_symtab_p (t1
) && decl_in_symtab_p (t2
))
1997 return return_false_with_msg ("Declaration mismatch");
1999 /* TODO: We can check CONST_DECL by its DECL_INITIAL, but for that we
2000 need to process its VAR/FUNCTION references without relying on ipa-ref
2004 return return_false_with_msg ("Declaration mismatch");
2006 /* Integer constants are the same only if the same width of type. */
2007 if (TYPE_PRECISION (TREE_TYPE (t1
)) != TYPE_PRECISION (TREE_TYPE (t2
)))
2008 return return_false_with_msg ("INTEGER_CST precision mismatch");
2009 if (TYPE_MODE (TREE_TYPE (t1
)) != TYPE_MODE (TREE_TYPE (t2
)))
2010 return return_false_with_msg ("INTEGER_CST mode mismatch");
2011 return return_with_debug (tree_int_cst_equal (t1
, t2
));
2013 if (TYPE_MODE (TREE_TYPE (t1
)) != TYPE_MODE (TREE_TYPE (t2
)))
2014 return return_false_with_msg ("STRING_CST mode mismatch");
2015 if (TREE_STRING_LENGTH (t1
) != TREE_STRING_LENGTH (t2
))
2016 return return_false_with_msg ("STRING_CST length mismatch");
2017 if (memcmp (TREE_STRING_POINTER (t1
), TREE_STRING_POINTER (t2
),
2018 TREE_STRING_LENGTH (t1
)))
2019 return return_false_with_msg ("STRING_CST mismatch");
2022 /* Fixed constants are the same only if the same width of type. */
2023 if (TYPE_PRECISION (TREE_TYPE (t1
)) != TYPE_PRECISION (TREE_TYPE (t2
)))
2024 return return_false_with_msg ("FIXED_CST precision mismatch");
2026 return return_with_debug (FIXED_VALUES_IDENTICAL (TREE_FIXED_CST (t1
),
2027 TREE_FIXED_CST (t2
)));
2029 return (sem_variable::equals (TREE_REALPART (t1
), TREE_REALPART (t2
))
2030 && sem_variable::equals (TREE_IMAGPART (t1
), TREE_IMAGPART (t2
)));
2032 /* Real constants are the same only if the same width of type. */
2033 if (TYPE_PRECISION (TREE_TYPE (t1
)) != TYPE_PRECISION (TREE_TYPE (t2
)))
2034 return return_false_with_msg ("REAL_CST precision mismatch");
2035 return return_with_debug (real_identical (&TREE_REAL_CST (t1
),
2036 &TREE_REAL_CST (t2
)));
2039 if (maybe_ne (VECTOR_CST_NELTS (t1
), VECTOR_CST_NELTS (t2
)))
2040 return return_false_with_msg ("VECTOR_CST nelts mismatch");
2043 = tree_vector_builder::binary_encoded_nelts (t1
, t2
);
2044 for (unsigned int i
= 0; i
< count
; ++i
)
2045 if (!sem_variable::equals (VECTOR_CST_ENCODED_ELT (t1
, i
),
2046 VECTOR_CST_ENCODED_ELT (t2
, i
)))
2052 case ARRAY_RANGE_REF
:
2054 tree x1
= TREE_OPERAND (t1
, 0);
2055 tree x2
= TREE_OPERAND (t2
, 0);
2056 tree y1
= TREE_OPERAND (t1
, 1);
2057 tree y2
= TREE_OPERAND (t2
, 1);
2059 if (!sem_variable::equals (x1
, x2
) || !sem_variable::equals (y1
, y2
))
2061 if (!sem_variable::equals (array_ref_low_bound (t1
),
2062 array_ref_low_bound (t2
)))
2064 if (!sem_variable::equals (array_ref_element_size (t1
),
2065 array_ref_element_size (t2
)))
2071 case POINTER_PLUS_EXPR
:
2076 tree x1
= TREE_OPERAND (t1
, 0);
2077 tree x2
= TREE_OPERAND (t2
, 0);
2078 tree y1
= TREE_OPERAND (t1
, 1);
2079 tree y2
= TREE_OPERAND (t2
, 1);
2081 return sem_variable::equals (x1
, x2
) && sem_variable::equals (y1
, y2
);
2085 case VIEW_CONVERT_EXPR
:
2086 if (!func_checker::compatible_types_p (TREE_TYPE (t1
), TREE_TYPE (t2
)))
2087 return return_false ();
2088 return sem_variable::equals (TREE_OPERAND (t1
, 0), TREE_OPERAND (t2
, 0));
2090 return return_false_with_msg ("ERROR_MARK");
2092 return return_false_with_msg ("Unknown TREE code reached");
2096 /* Parser function that visits a varpool NODE. */
2099 sem_variable::parse (varpool_node
*node
, bitmap_obstack
*stack
)
2101 if (TREE_THIS_VOLATILE (node
->decl
) || DECL_HARD_REGISTER (node
->decl
)
2105 sem_variable
*v
= new sem_variable (node
, stack
);
2112 /* References independent hash function. */
2115 sem_variable::get_hash (void)
2120 /* All WPA streamed in symbols should have their hashes computed at compile
2121 time. At this point, the constructor may not be in memory at all.
2122 DECL_INITIAL (decl) would be error_mark_node in that case. */
2123 gcc_assert (!node
->lto_file_data
);
2124 tree ctor
= DECL_INITIAL (decl
);
2125 inchash::hash hstate
;
2127 hstate
.add_int (456346417);
2128 if (DECL_SIZE (decl
) && tree_fits_shwi_p (DECL_SIZE (decl
)))
2129 hstate
.add_hwi (tree_to_shwi (DECL_SIZE (decl
)));
2130 add_expr (ctor
, hstate
);
2131 set_hash (hstate
.end ());
2136 /* Merges instance with an ALIAS_ITEM, where alias, thunk or redirection can
2140 sem_variable::merge (sem_item
*alias_item
)
2142 gcc_assert (alias_item
->type
== VAR
);
2144 if (!sem_item::target_supports_symbol_aliases_p ())
2147 fprintf (dump_file
, "Not unifying; "
2148 "Symbol aliases are not supported by target\n\n");
2152 if (DECL_EXTERNAL (alias_item
->decl
))
2155 fprintf (dump_file
, "Not unifying; alias is external.\n\n");
2159 sem_variable
*alias_var
= static_cast<sem_variable
*> (alias_item
);
2161 varpool_node
*original
= get_node ();
2162 varpool_node
*alias
= alias_var
->get_node ();
2163 bool original_discardable
= false;
2165 bool alias_address_matters
= alias
->address_matters_p ();
2167 /* See if original is in a section that can be discarded if the main
2169 Also consider case where we have resolution info and we know that
2170 original's definition is not going to be used. In this case we can not
2171 create alias to original. */
2172 if (original
->can_be_discarded_p ()
2173 || (node
->resolution
!= LDPR_UNKNOWN
2174 && !decl_binds_to_current_def_p (node
->decl
)))
2175 original_discardable
= true;
2177 gcc_assert (!TREE_ASM_WRITTEN (alias
->decl
));
2179 /* Constant pool machinery is not quite ready for aliases.
2180 TODO: varasm code contains logic for merging DECL_IN_CONSTANT_POOL.
2181 For LTO merging does not happen that is an important missing feature.
2182 We can enable merging with LTO if the DECL_IN_CONSTANT_POOL
2183 flag is dropped and non-local symbol name is assigned. */
2184 if (DECL_IN_CONSTANT_POOL (alias
->decl
)
2185 || DECL_IN_CONSTANT_POOL (original
->decl
))
2189 "Not unifying; constant pool variables.\n\n");
2193 /* Do not attempt to mix functions from different user sections;
2194 we do not know what user intends with those. */
2195 if (((DECL_SECTION_NAME (original
->decl
) && !original
->implicit_section
)
2196 || (DECL_SECTION_NAME (alias
->decl
) && !alias
->implicit_section
))
2197 && DECL_SECTION_NAME (original
->decl
) != DECL_SECTION_NAME (alias
->decl
))
2202 "original and alias are in different sections.\n\n");
2206 /* We can not merge if address comparsion metters. */
2207 if (alias_address_matters
&& flag_merge_constants
< 2)
2211 "Not unifying; address of original may be compared.\n\n");
2215 if (DECL_ALIGN (original
->decl
) < DECL_ALIGN (alias
->decl
))
2218 fprintf (dump_file
, "Not unifying; "
2219 "original and alias have incompatible alignments\n\n");
2224 if (DECL_COMDAT_GROUP (original
->decl
) != DECL_COMDAT_GROUP (alias
->decl
))
2227 fprintf (dump_file
, "Not unifying; alias cannot be created; "
2228 "across comdat group boundary\n\n");
2233 if (original_discardable
)
2236 fprintf (dump_file
, "Not unifying; alias cannot be created; "
2237 "target is discardable\n\n");
2243 gcc_assert (!original
->alias
);
2244 gcc_assert (!alias
->alias
);
2246 alias
->analyzed
= false;
2248 DECL_INITIAL (alias
->decl
) = NULL
;
2249 ((symtab_node
*)alias
)->call_for_symbol_and_aliases (clear_decl_rtl
,
2251 alias
->need_bounds_init
= false;
2252 alias
->remove_all_references ();
2253 if (TREE_ADDRESSABLE (alias
->decl
))
2254 original
->call_for_symbol_and_aliases (set_addressable
, NULL
, true);
2256 varpool_node::create_alias (alias_var
->decl
, decl
);
2257 alias
->resolve_alias (original
);
2260 fprintf (dump_file
, "Unified; Variable alias has been created.\n");
2266 /* Dump symbol to FILE. */
2269 sem_variable::dump_to_file (FILE *file
)
2273 print_node (file
, "", decl
, 0);
2274 fprintf (file
, "\n\n");
2277 unsigned int sem_item_optimizer::class_id
= 0;
2279 sem_item_optimizer::sem_item_optimizer ()
2280 : worklist (0), m_classes (0), m_classes_count (0), m_cgraph_node_hooks (NULL
),
2281 m_varpool_node_hooks (NULL
), m_merged_variables ()
2284 bitmap_obstack_initialize (&m_bmstack
);
2287 sem_item_optimizer::~sem_item_optimizer ()
2289 for (unsigned int i
= 0; i
< m_items
.length (); i
++)
2293 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
2294 it
!= m_classes
.end (); ++it
)
2296 for (unsigned int i
= 0; i
< (*it
)->classes
.length (); i
++)
2297 delete (*it
)->classes
[i
];
2299 (*it
)->classes
.release ();
2305 bitmap_obstack_release (&m_bmstack
);
2306 m_merged_variables
.release ();
2309 /* Write IPA ICF summary for symbols. */
2312 sem_item_optimizer::write_summary (void)
2314 unsigned int count
= 0;
2316 output_block
*ob
= create_output_block (LTO_section_ipa_icf
);
2317 lto_symtab_encoder_t encoder
= ob
->decl_state
->symtab_node_encoder
;
2320 /* Calculate number of symbols to be serialized. */
2321 for (lto_symtab_encoder_iterator lsei
= lsei_start_in_partition (encoder
);
2323 lsei_next_in_partition (&lsei
))
2325 symtab_node
*node
= lsei_node (lsei
);
2327 if (m_symtab_node_map
.get (node
))
2331 streamer_write_uhwi (ob
, count
);
2333 /* Process all of the symbols. */
2334 for (lto_symtab_encoder_iterator lsei
= lsei_start_in_partition (encoder
);
2336 lsei_next_in_partition (&lsei
))
2338 symtab_node
*node
= lsei_node (lsei
);
2340 sem_item
**item
= m_symtab_node_map
.get (node
);
2344 int node_ref
= lto_symtab_encoder_encode (encoder
, node
);
2345 streamer_write_uhwi_stream (ob
->main_stream
, node_ref
);
2347 streamer_write_uhwi (ob
, (*item
)->get_hash ());
2351 streamer_write_char_stream (ob
->main_stream
, 0);
2352 produce_asm (ob
, NULL
);
2353 destroy_output_block (ob
);
2356 /* Reads a section from LTO stream file FILE_DATA. Input block for DATA
2357 contains LEN bytes. */
2360 sem_item_optimizer::read_section (lto_file_decl_data
*file_data
,
2361 const char *data
, size_t len
)
2363 const lto_function_header
*header
2364 = (const lto_function_header
*) data
;
2365 const int cfg_offset
= sizeof (lto_function_header
);
2366 const int main_offset
= cfg_offset
+ header
->cfg_size
;
2367 const int string_offset
= main_offset
+ header
->main_size
;
2372 lto_input_block
ib_main ((const char *) data
+ main_offset
, 0,
2373 header
->main_size
, file_data
->mode_table
);
2376 = lto_data_in_create (file_data
, (const char *) data
+ string_offset
,
2377 header
->string_size
, vNULL
);
2379 count
= streamer_read_uhwi (&ib_main
);
2381 for (i
= 0; i
< count
; i
++)
2385 lto_symtab_encoder_t encoder
;
2387 index
= streamer_read_uhwi (&ib_main
);
2388 encoder
= file_data
->symtab_node_encoder
;
2389 node
= lto_symtab_encoder_deref (encoder
, index
);
2391 hashval_t hash
= streamer_read_uhwi (&ib_main
);
2393 gcc_assert (node
->definition
);
2396 fprintf (dump_file
, "Symbol added: %s (tree: %p)\n",
2397 node
->dump_asm_name (), (void *) node
->decl
);
2399 if (is_a
<cgraph_node
*> (node
))
2401 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (node
);
2403 sem_function
*fn
= new sem_function (cnode
, &m_bmstack
);
2404 fn
->set_hash (hash
);
2405 m_items
.safe_push (fn
);
2409 varpool_node
*vnode
= dyn_cast
<varpool_node
*> (node
);
2411 sem_variable
*var
= new sem_variable (vnode
, &m_bmstack
);
2412 var
->set_hash (hash
);
2413 m_items
.safe_push (var
);
2417 lto_free_section_data (file_data
, LTO_section_ipa_icf
, NULL
, data
,
2419 lto_data_in_delete (data_in
);
2422 /* Read IPA ICF summary for symbols. */
2425 sem_item_optimizer::read_summary (void)
2427 lto_file_decl_data
**file_data_vec
= lto_get_file_decl_data ();
2428 lto_file_decl_data
*file_data
;
2431 while ((file_data
= file_data_vec
[j
++]))
2434 const char *data
= lto_get_section_data (file_data
,
2435 LTO_section_ipa_icf
, NULL
, &len
);
2438 read_section (file_data
, data
, len
);
2442 /* Register callgraph and varpool hooks. */
2445 sem_item_optimizer::register_hooks (void)
2447 if (!m_cgraph_node_hooks
)
2448 m_cgraph_node_hooks
= symtab
->add_cgraph_removal_hook
2449 (&sem_item_optimizer::cgraph_removal_hook
, this);
2451 if (!m_varpool_node_hooks
)
2452 m_varpool_node_hooks
= symtab
->add_varpool_removal_hook
2453 (&sem_item_optimizer::varpool_removal_hook
, this);
2456 /* Unregister callgraph and varpool hooks. */
2459 sem_item_optimizer::unregister_hooks (void)
2461 if (m_cgraph_node_hooks
)
2462 symtab
->remove_cgraph_removal_hook (m_cgraph_node_hooks
);
2464 if (m_varpool_node_hooks
)
2465 symtab
->remove_varpool_removal_hook (m_varpool_node_hooks
);
2468 /* Adds a CLS to hashtable associated by hash value. */
2471 sem_item_optimizer::add_class (congruence_class
*cls
)
2473 gcc_assert (cls
->members
.length ());
2475 congruence_class_group
*group
2476 = get_group_by_hash (cls
->members
[0]->get_hash (),
2477 cls
->members
[0]->type
);
2478 group
->classes
.safe_push (cls
);
2481 /* Gets a congruence class group based on given HASH value and TYPE. */
2483 congruence_class_group
*
2484 sem_item_optimizer::get_group_by_hash (hashval_t hash
, sem_item_type type
)
2486 congruence_class_group
*item
= XNEW (congruence_class_group
);
2490 congruence_class_group
**slot
= m_classes
.find_slot (item
, INSERT
);
2496 item
->classes
.create (1);
2503 /* Callgraph removal hook called for a NODE with a custom DATA. */
2506 sem_item_optimizer::cgraph_removal_hook (cgraph_node
*node
, void *data
)
2508 sem_item_optimizer
*optimizer
= (sem_item_optimizer
*) data
;
2509 optimizer
->remove_symtab_node (node
);
2512 /* Varpool removal hook called for a NODE with a custom DATA. */
2515 sem_item_optimizer::varpool_removal_hook (varpool_node
*node
, void *data
)
2517 sem_item_optimizer
*optimizer
= (sem_item_optimizer
*) data
;
2518 optimizer
->remove_symtab_node (node
);
2521 /* Remove symtab NODE triggered by symtab removal hooks. */
2524 sem_item_optimizer::remove_symtab_node (symtab_node
*node
)
2526 gcc_assert (!m_classes
.elements ());
2528 m_removed_items_set
.add (node
);
2532 sem_item_optimizer::remove_item (sem_item
*item
)
2534 if (m_symtab_node_map
.get (item
->node
))
2535 m_symtab_node_map
.remove (item
->node
);
2539 /* Removes all callgraph and varpool nodes that are marked by symtab
2543 sem_item_optimizer::filter_removed_items (void)
2545 auto_vec
<sem_item
*> filtered
;
2547 for (unsigned int i
= 0; i
< m_items
.length(); i
++)
2549 sem_item
*item
= m_items
[i
];
2551 if (m_removed_items_set
.contains (item
->node
))
2557 if (item
->type
== FUNC
)
2559 cgraph_node
*cnode
= static_cast <sem_function
*>(item
)->get_node ();
2561 if (in_lto_p
&& (cnode
->alias
|| cnode
->body_removed
))
2564 filtered
.safe_push (item
);
2568 if (!flag_ipa_icf_variables
)
2572 /* Filter out non-readonly variables. */
2573 tree decl
= item
->decl
;
2574 if (TREE_READONLY (decl
))
2575 filtered
.safe_push (item
);
2582 /* Clean-up of released semantic items. */
2585 for (unsigned int i
= 0; i
< filtered
.length(); i
++)
2586 m_items
.safe_push (filtered
[i
]);
2589 /* Optimizer entry point which returns true in case it processes
2590 a merge operation. True is returned if there's a merge operation
2594 sem_item_optimizer::execute (void)
2596 filter_removed_items ();
2597 unregister_hooks ();
2600 update_hash_by_addr_refs ();
2601 build_hash_based_classes ();
2604 fprintf (dump_file
, "Dump after hash based groups\n");
2605 dump_cong_classes ();
2607 for (unsigned int i
= 0; i
< m_items
.length(); i
++)
2608 m_items
[i
]->init_wpa ();
2610 subdivide_classes_by_equality (true);
2613 fprintf (dump_file
, "Dump after WPA based types groups\n");
2615 dump_cong_classes ();
2617 process_cong_reduction ();
2618 checking_verify_classes ();
2621 fprintf (dump_file
, "Dump after callgraph-based congruence reduction\n");
2623 dump_cong_classes ();
2625 parse_nonsingleton_classes ();
2626 subdivide_classes_by_equality ();
2629 fprintf (dump_file
, "Dump after full equality comparison of groups\n");
2631 dump_cong_classes ();
2633 unsigned int prev_class_count
= m_classes_count
;
2635 process_cong_reduction ();
2636 dump_cong_classes ();
2637 checking_verify_classes ();
2638 bool merged_p
= merge_classes (prev_class_count
);
2640 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2641 symtab
->dump (dump_file
);
2646 /* Function responsible for visiting all potential functions and
2647 read-only variables that can be merged. */
2650 sem_item_optimizer::parse_funcs_and_vars (void)
2654 if (flag_ipa_icf_functions
)
2655 FOR_EACH_DEFINED_FUNCTION (cnode
)
2657 sem_function
*f
= sem_function::parse (cnode
, &m_bmstack
);
2660 m_items
.safe_push (f
);
2661 m_symtab_node_map
.put (cnode
, f
);
2664 fprintf (dump_file
, "Parsed function:%s\n", f
->node
->asm_name ());
2666 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2667 f
->dump_to_file (dump_file
);
2670 fprintf (dump_file
, "Not parsed function:%s\n", cnode
->asm_name ());
2673 varpool_node
*vnode
;
2675 if (flag_ipa_icf_variables
)
2676 FOR_EACH_DEFINED_VARIABLE (vnode
)
2678 sem_variable
*v
= sem_variable::parse (vnode
, &m_bmstack
);
2682 m_items
.safe_push (v
);
2683 m_symtab_node_map
.put (vnode
, v
);
2688 /* Makes pairing between a congruence class CLS and semantic ITEM. */
2691 sem_item_optimizer::add_item_to_class (congruence_class
*cls
, sem_item
*item
)
2693 item
->index_in_class
= cls
->members
.length ();
2694 cls
->members
.safe_push (item
);
2698 /* For each semantic item, append hash values of references. */
2701 sem_item_optimizer::update_hash_by_addr_refs ()
2703 /* First, append to hash sensitive references and class type if it need to
2704 be matched for ODR. */
2705 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2707 m_items
[i
]->update_hash_by_addr_refs (m_symtab_node_map
);
2708 if (m_items
[i
]->type
== FUNC
)
2710 if (TREE_CODE (TREE_TYPE (m_items
[i
]->decl
)) == METHOD_TYPE
2711 && contains_polymorphic_type_p
2712 (TYPE_METHOD_BASETYPE (TREE_TYPE (m_items
[i
]->decl
)))
2713 && (DECL_CXX_CONSTRUCTOR_P (m_items
[i
]->decl
)
2714 || (static_cast<sem_function
*> (m_items
[i
])->param_used_p (0)
2715 && static_cast<sem_function
*> (m_items
[i
])
2716 ->compare_polymorphic_p ())))
2719 = TYPE_METHOD_BASETYPE (TREE_TYPE (m_items
[i
]->decl
));
2720 inchash::hash
hstate (m_items
[i
]->get_hash ());
2722 if (TYPE_NAME (class_type
)
2723 && DECL_ASSEMBLER_NAME_SET_P (TYPE_NAME (class_type
)))
2725 (IDENTIFIER_HASH_VALUE
2726 (DECL_ASSEMBLER_NAME (TYPE_NAME (class_type
))));
2728 m_items
[i
]->set_hash (hstate
.end ());
2733 /* Once all symbols have enhanced hash value, we can append
2734 hash values of symbols that are seen by IPA ICF and are
2735 references by a semantic item. Newly computed values
2736 are saved to global_hash member variable. */
2737 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2738 m_items
[i
]->update_hash_by_local_refs (m_symtab_node_map
);
2740 /* Global hash value replace current hash values. */
2741 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2742 m_items
[i
]->set_hash (m_items
[i
]->global_hash
);
2745 /* Congruence classes are built by hash value. */
2748 sem_item_optimizer::build_hash_based_classes (void)
2750 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2752 sem_item
*item
= m_items
[i
];
2754 congruence_class_group
*group
2755 = get_group_by_hash (item
->get_hash (), item
->type
);
2757 if (!group
->classes
.length ())
2760 group
->classes
.safe_push (new congruence_class (class_id
++));
2763 add_item_to_class (group
->classes
[0], item
);
2767 /* Build references according to call graph. */
2770 sem_item_optimizer::build_graph (void)
2772 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2774 sem_item
*item
= m_items
[i
];
2775 m_symtab_node_map
.put (item
->node
, item
);
2777 /* Initialize hash values if we are not in LTO mode. */
2782 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2784 sem_item
*item
= m_items
[i
];
2786 if (item
->type
== FUNC
)
2788 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (item
->node
);
2790 cgraph_edge
*e
= cnode
->callees
;
2793 sem_item
**slot
= m_symtab_node_map
.get
2794 (e
->callee
->ultimate_alias_target ());
2796 item
->add_reference (*slot
);
2802 ipa_ref
*ref
= NULL
;
2803 for (unsigned i
= 0; item
->node
->iterate_reference (i
, ref
); i
++)
2805 sem_item
**slot
= m_symtab_node_map
.get
2806 (ref
->referred
->ultimate_alias_target ());
2808 item
->add_reference (*slot
);
2813 /* Semantic items in classes having more than one element and initialized.
2814 In case of WPA, we load function body. */
2817 sem_item_optimizer::parse_nonsingleton_classes (void)
2819 unsigned int init_called_count
= 0;
2821 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2822 if (m_items
[i
]->cls
->members
.length () > 1)
2824 m_items
[i
]->init ();
2825 init_called_count
++;
2829 fprintf (dump_file
, "Init called for %u items (%.2f%%).\n",
2831 m_items
.length () ? 100.0f
* init_called_count
/ m_items
.length ()
2835 /* Equality function for semantic items is used to subdivide existing
2836 classes. If IN_WPA, fast equality function is invoked. */
2839 sem_item_optimizer::subdivide_classes_by_equality (bool in_wpa
)
2841 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
2842 it
!= m_classes
.end (); ++it
)
2844 unsigned int class_count
= (*it
)->classes
.length ();
2846 for (unsigned i
= 0; i
< class_count
; i
++)
2848 congruence_class
*c
= (*it
)->classes
[i
];
2850 if (c
->members
.length() > 1)
2852 auto_vec
<sem_item
*> new_vector
;
2854 sem_item
*first
= c
->members
[0];
2855 new_vector
.safe_push (first
);
2857 unsigned class_split_first
= (*it
)->classes
.length ();
2859 for (unsigned j
= 1; j
< c
->members
.length (); j
++)
2861 sem_item
*item
= c
->members
[j
];
2864 = in_wpa
? first
->equals_wpa (item
, m_symtab_node_map
)
2865 : first
->equals (item
, m_symtab_node_map
);
2868 new_vector
.safe_push (item
);
2871 bool integrated
= false;
2873 for (unsigned k
= class_split_first
;
2874 k
< (*it
)->classes
.length (); k
++)
2876 sem_item
*x
= (*it
)->classes
[k
]->members
[0];
2878 = in_wpa
? x
->equals_wpa (item
, m_symtab_node_map
)
2879 : x
->equals (item
, m_symtab_node_map
);
2884 add_item_to_class ((*it
)->classes
[k
], item
);
2893 = new congruence_class (class_id
++);
2895 add_item_to_class (c
, item
);
2897 (*it
)->classes
.safe_push (c
);
2902 // We replace newly created new_vector for the class we've just
2904 c
->members
.release ();
2905 c
->members
.create (new_vector
.length ());
2907 for (unsigned int j
= 0; j
< new_vector
.length (); j
++)
2908 add_item_to_class (c
, new_vector
[j
]);
2913 checking_verify_classes ();
2916 /* Subdivide classes by address references that members of the class
2917 reference. Example can be a pair of functions that have an address
2918 taken from a function. If these addresses are different the class
2922 sem_item_optimizer::subdivide_classes_by_sensitive_refs ()
2924 typedef hash_map
<symbol_compare_hash
, vec
<sem_item
*> > subdivide_hash_map
;
2926 unsigned newly_created_classes
= 0;
2928 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
2929 it
!= m_classes
.end (); ++it
)
2931 unsigned int class_count
= (*it
)->classes
.length ();
2932 auto_vec
<congruence_class
*> new_classes
;
2934 for (unsigned i
= 0; i
< class_count
; i
++)
2936 congruence_class
*c
= (*it
)->classes
[i
];
2938 if (c
->members
.length() > 1)
2940 subdivide_hash_map split_map
;
2942 for (unsigned j
= 0; j
< c
->members
.length (); j
++)
2944 sem_item
*source_node
= c
->members
[j
];
2946 symbol_compare_collection
*collection
2947 = new symbol_compare_collection (source_node
->node
);
2950 vec
<sem_item
*> *slot
2951 = &split_map
.get_or_insert (collection
, &existed
);
2952 gcc_checking_assert (slot
);
2954 slot
->safe_push (source_node
);
2960 /* If the map contains more than one key, we have to split
2961 the map appropriately. */
2962 if (split_map
.elements () != 1)
2964 bool first_class
= true;
2966 for (subdivide_hash_map::iterator it2
= split_map
.begin ();
2967 it2
!= split_map
.end (); ++it2
)
2969 congruence_class
*new_cls
;
2970 new_cls
= new congruence_class (class_id
++);
2972 for (unsigned k
= 0; k
< (*it2
).second
.length (); k
++)
2973 add_item_to_class (new_cls
, (*it2
).second
[k
]);
2975 worklist_push (new_cls
);
2976 newly_created_classes
++;
2980 (*it
)->classes
[i
] = new_cls
;
2981 first_class
= false;
2985 new_classes
.safe_push (new_cls
);
2991 /* Release memory. */
2992 for (subdivide_hash_map::iterator it2
= split_map
.begin ();
2993 it2
!= split_map
.end (); ++it2
)
2995 delete (*it2
).first
;
2996 (*it2
).second
.release ();
3001 for (unsigned i
= 0; i
< new_classes
.length (); i
++)
3002 (*it
)->classes
.safe_push (new_classes
[i
]);
3005 return newly_created_classes
;
3008 /* Verify congruence classes, if checking is enabled. */
3011 sem_item_optimizer::checking_verify_classes (void)
3017 /* Verify congruence classes. */
3020 sem_item_optimizer::verify_classes (void)
3022 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3023 it
!= m_classes
.end (); ++it
)
3025 for (unsigned int i
= 0; i
< (*it
)->classes
.length (); i
++)
3027 congruence_class
*cls
= (*it
)->classes
[i
];
3030 gcc_assert (cls
->members
.length () > 0);
3032 for (unsigned int j
= 0; j
< cls
->members
.length (); j
++)
3034 sem_item
*item
= cls
->members
[j
];
3037 gcc_assert (item
->cls
== cls
);
3039 for (unsigned k
= 0; k
< item
->usages
.length (); k
++)
3041 sem_usage_pair
*usage
= item
->usages
[k
];
3042 gcc_assert (usage
->item
->index_in_class
3043 < usage
->item
->cls
->members
.length ());
3050 /* Disposes split map traverse function. CLS_PTR is pointer to congruence
3051 class, BSLOT is bitmap slot we want to release. DATA is mandatory,
3052 but unused argument. */
3055 sem_item_optimizer::release_split_map (congruence_class
* const &,
3056 bitmap
const &b
, traverse_split_pair
*)
3065 /* Process split operation for a class given as pointer CLS_PTR,
3066 where bitmap B splits congruence class members. DATA is used
3067 as argument of split pair. */
3070 sem_item_optimizer::traverse_congruence_split (congruence_class
* const &cls
,
3072 traverse_split_pair
*pair
)
3074 sem_item_optimizer
*optimizer
= pair
->optimizer
;
3075 const congruence_class
*splitter_cls
= pair
->cls
;
3077 /* If counted bits are greater than zero and less than the number of members
3078 a group will be splitted. */
3079 unsigned popcount
= bitmap_count_bits (b
);
3081 if (popcount
> 0 && popcount
< cls
->members
.length ())
3083 auto_vec
<congruence_class
*, 2> newclasses
;
3084 newclasses
.quick_push (new congruence_class (class_id
++));
3085 newclasses
.quick_push (new congruence_class (class_id
++));
3087 for (unsigned int i
= 0; i
< cls
->members
.length (); i
++)
3089 int target
= bitmap_bit_p (b
, i
);
3090 congruence_class
*tc
= newclasses
[target
];
3092 add_item_to_class (tc
, cls
->members
[i
]);
3097 for (unsigned int i
= 0; i
< 2; i
++)
3098 gcc_assert (newclasses
[i
]->members
.length ());
3101 if (splitter_cls
== cls
)
3102 optimizer
->splitter_class_removed
= true;
3104 /* Remove old class from worklist if presented. */
3105 bool in_worklist
= cls
->in_worklist
;
3108 cls
->in_worklist
= false;
3110 congruence_class_group g
;
3111 g
.hash
= cls
->members
[0]->get_hash ();
3112 g
.type
= cls
->members
[0]->type
;
3114 congruence_class_group
*slot
= optimizer
->m_classes
.find (&g
);
3116 for (unsigned int i
= 0; i
< slot
->classes
.length (); i
++)
3117 if (slot
->classes
[i
] == cls
)
3119 slot
->classes
.ordered_remove (i
);
3123 /* New class will be inserted and integrated to work list. */
3124 for (unsigned int i
= 0; i
< 2; i
++)
3125 optimizer
->add_class (newclasses
[i
]);
3127 /* Two classes replace one, so that increment just by one. */
3128 optimizer
->m_classes_count
++;
3130 /* If OLD class was presented in the worklist, we remove the class
3131 and replace it will both newly created classes. */
3133 for (unsigned int i
= 0; i
< 2; i
++)
3134 optimizer
->worklist_push (newclasses
[i
]);
3135 else /* Just smaller class is inserted. */
3137 unsigned int smaller_index
3138 = (newclasses
[0]->members
.length ()
3139 < newclasses
[1]->members
.length ()
3141 optimizer
->worklist_push (newclasses
[smaller_index
]);
3144 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3146 fprintf (dump_file
, " congruence class splitted:\n");
3147 cls
->dump (dump_file
, 4);
3149 fprintf (dump_file
, " newly created groups:\n");
3150 for (unsigned int i
= 0; i
< 2; i
++)
3151 newclasses
[i
]->dump (dump_file
, 4);
3154 /* Release class if not presented in work list. */
3163 /* Tests if a class CLS used as INDEXth splits any congruence classes.
3164 Bitmap stack BMSTACK is used for bitmap allocation. */
3167 sem_item_optimizer::do_congruence_step_for_index (congruence_class
*cls
,
3170 hash_map
<congruence_class
*, bitmap
> split_map
;
3172 for (unsigned int i
= 0; i
< cls
->members
.length (); i
++)
3174 sem_item
*item
= cls
->members
[i
];
3176 /* Iterate all usages that have INDEX as usage of the item. */
3177 for (unsigned int j
= 0; j
< item
->usages
.length (); j
++)
3179 sem_usage_pair
*usage
= item
->usages
[j
];
3181 if (usage
->index
!= index
)
3184 bitmap
*slot
= split_map
.get (usage
->item
->cls
);
3189 b
= BITMAP_ALLOC (&m_bmstack
);
3190 split_map
.put (usage
->item
->cls
, b
);
3195 gcc_checking_assert (usage
->item
->cls
);
3196 gcc_checking_assert (usage
->item
->index_in_class
3197 < usage
->item
->cls
->members
.length ());
3199 bitmap_set_bit (b
, usage
->item
->index_in_class
);
3203 traverse_split_pair pair
;
3204 pair
.optimizer
= this;
3207 splitter_class_removed
= false;
3208 split_map
.traverse
<traverse_split_pair
*,
3209 sem_item_optimizer::traverse_congruence_split
> (&pair
);
3211 /* Bitmap clean-up. */
3212 split_map
.traverse
<traverse_split_pair
*,
3213 sem_item_optimizer::release_split_map
> (NULL
);
3216 /* Every usage of a congruence class CLS is a candidate that can split the
3217 collection of classes. Bitmap stack BMSTACK is used for bitmap
3221 sem_item_optimizer::do_congruence_step (congruence_class
*cls
)
3226 bitmap usage
= BITMAP_ALLOC (&m_bmstack
);
3228 for (unsigned int i
= 0; i
< cls
->members
.length (); i
++)
3229 bitmap_ior_into (usage
, cls
->members
[i
]->usage_index_bitmap
);
3231 EXECUTE_IF_SET_IN_BITMAP (usage
, 0, i
, bi
)
3233 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3234 fprintf (dump_file
, " processing congruence step for class: %u, "
3235 "index: %u\n", cls
->id
, i
);
3237 do_congruence_step_for_index (cls
, i
);
3239 if (splitter_class_removed
)
3243 BITMAP_FREE (usage
);
3246 /* Adds a newly created congruence class CLS to worklist. */
3249 sem_item_optimizer::worklist_push (congruence_class
*cls
)
3251 /* Return if the class CLS is already presented in work list. */
3252 if (cls
->in_worklist
)
3255 cls
->in_worklist
= true;
3256 worklist
.push_back (cls
);
3259 /* Pops a class from worklist. */
3262 sem_item_optimizer::worklist_pop (void)
3264 congruence_class
*cls
;
3266 while (!worklist
.empty ())
3268 cls
= worklist
.front ();
3269 worklist
.pop_front ();
3270 if (cls
->in_worklist
)
3272 cls
->in_worklist
= false;
3278 /* Work list item was already intended to be removed.
3279 The only reason for doing it is to split a class.
3280 Thus, the class CLS is deleted. */
3288 /* Iterative congruence reduction function. */
3291 sem_item_optimizer::process_cong_reduction (void)
3293 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3294 it
!= m_classes
.end (); ++it
)
3295 for (unsigned i
= 0; i
< (*it
)->classes
.length (); i
++)
3296 if ((*it
)->classes
[i
]->is_class_used ())
3297 worklist_push ((*it
)->classes
[i
]);
3300 fprintf (dump_file
, "Worklist has been filled with: %lu\n",
3301 (unsigned long) worklist
.size ());
3303 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3304 fprintf (dump_file
, "Congruence class reduction\n");
3306 congruence_class
*cls
;
3308 /* Process complete congruence reduction. */
3309 while ((cls
= worklist_pop ()) != NULL
)
3310 do_congruence_step (cls
);
3312 /* Subdivide newly created classes according to references. */
3313 unsigned new_classes
= subdivide_classes_by_sensitive_refs ();
3316 fprintf (dump_file
, "Address reference subdivision created: %u "
3317 "new classes.\n", new_classes
);
3320 /* Debug function prints all informations about congruence classes. */
3323 sem_item_optimizer::dump_cong_classes (void)
3329 "Congruence classes: %u (unique hash values: %lu), with total: "
3330 "%u items\n", m_classes_count
,
3331 (unsigned long) m_classes
.elements (), m_items
.length ());
3333 /* Histogram calculation. */
3334 unsigned int max_index
= 0;
3335 unsigned int* histogram
= XCNEWVEC (unsigned int, m_items
.length () + 1);
3337 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3338 it
!= m_classes
.end (); ++it
)
3339 for (unsigned i
= 0; i
< (*it
)->classes
.length (); i
++)
3341 unsigned int c
= (*it
)->classes
[i
]->members
.length ();
3349 "Class size histogram [num of members]: number of classe number "
3352 for (unsigned int i
= 0; i
<= max_index
; i
++)
3354 fprintf (dump_file
, "[%u]: %u classes\n", i
, histogram
[i
]);
3356 fprintf (dump_file
, "\n\n");
3358 if (dump_flags
& TDF_DETAILS
)
3359 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3360 it
!= m_classes
.end (); ++it
)
3362 fprintf (dump_file
, " group: with %u classes:\n",
3363 (*it
)->classes
.length ());
3365 for (unsigned i
= 0; i
< (*it
)->classes
.length (); i
++)
3367 (*it
)->classes
[i
]->dump (dump_file
, 4);
3369 if (i
< (*it
)->classes
.length () - 1)
3370 fprintf (dump_file
, " ");
3377 /* Sort pair of sem_items A and B by DECL_UID. */
3380 sort_sem_items_by_decl_uid (const void *a
, const void *b
)
3382 const sem_item
*i1
= *(const sem_item
* const *)a
;
3383 const sem_item
*i2
= *(const sem_item
* const *)b
;
3385 int uid1
= DECL_UID (i1
->decl
);
3386 int uid2
= DECL_UID (i2
->decl
);
3390 else if (uid1
> uid2
)
3396 /* Sort pair of congruence_classes A and B by DECL_UID of the first member. */
3399 sort_congruence_classes_by_decl_uid (const void *a
, const void *b
)
3401 const congruence_class
*c1
= *(const congruence_class
* const *)a
;
3402 const congruence_class
*c2
= *(const congruence_class
* const *)b
;
3404 int uid1
= DECL_UID (c1
->members
[0]->decl
);
3405 int uid2
= DECL_UID (c2
->members
[0]->decl
);
3409 else if (uid1
> uid2
)
3415 /* Sort pair of congruence_class_groups A and B by
3416 DECL_UID of the first member of a first group. */
3419 sort_congruence_class_groups_by_decl_uid (const void *a
, const void *b
)
3421 const congruence_class_group
*g1
3422 = *(const congruence_class_group
* const *)a
;
3423 const congruence_class_group
*g2
3424 = *(const congruence_class_group
* const *)b
;
3426 int uid1
= DECL_UID (g1
->classes
[0]->members
[0]->decl
);
3427 int uid2
= DECL_UID (g2
->classes
[0]->members
[0]->decl
);
3431 else if (uid1
> uid2
)
3437 /* After reduction is done, we can declare all items in a group
3438 to be equal. PREV_CLASS_COUNT is start number of classes
3439 before reduction. True is returned if there's a merge operation
3443 sem_item_optimizer::merge_classes (unsigned int prev_class_count
)
3445 unsigned int item_count
= m_items
.length ();
3446 unsigned int class_count
= m_classes_count
;
3447 unsigned int equal_items
= item_count
- class_count
;
3449 unsigned int non_singular_classes_count
= 0;
3450 unsigned int non_singular_classes_sum
= 0;
3452 bool merged_p
= false;
3455 Sort functions in congruence classes by DECL_UID and do the same
3456 for the classes to not to break -fcompare-debug. */
3458 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3459 it
!= m_classes
.end (); ++it
)
3461 for (unsigned int i
= 0; i
< (*it
)->classes
.length (); i
++)
3463 congruence_class
*c
= (*it
)->classes
[i
];
3464 c
->members
.qsort (sort_sem_items_by_decl_uid
);
3467 (*it
)->classes
.qsort (sort_congruence_classes_by_decl_uid
);
3470 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3471 it
!= m_classes
.end (); ++it
)
3472 for (unsigned int i
= 0; i
< (*it
)->classes
.length (); i
++)
3474 congruence_class
*c
= (*it
)->classes
[i
];
3475 if (c
->members
.length () > 1)
3477 non_singular_classes_count
++;
3478 non_singular_classes_sum
+= c
->members
.length ();
3482 auto_vec
<congruence_class_group
*> classes (m_classes
.elements ());
3483 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3484 it
!= m_classes
.end (); ++it
)
3485 classes
.quick_push (*it
);
3487 classes
.qsort (sort_congruence_class_groups_by_decl_uid
);
3491 fprintf (dump_file
, "\nItem count: %u\n", item_count
);
3492 fprintf (dump_file
, "Congruent classes before: %u, after: %u\n",
3493 prev_class_count
, class_count
);
3494 fprintf (dump_file
, "Average class size before: %.2f, after: %.2f\n",
3495 prev_class_count
? 1.0f
* item_count
/ prev_class_count
: 0.0f
,
3496 class_count
? 1.0f
* item_count
/ class_count
: 0.0f
);
3497 fprintf (dump_file
, "Average non-singular class size: %.2f, count: %u\n",
3498 non_singular_classes_count
? 1.0f
* non_singular_classes_sum
/
3499 non_singular_classes_count
: 0.0f
,
3500 non_singular_classes_count
);
3501 fprintf (dump_file
, "Equal symbols: %u\n", equal_items
);
3502 fprintf (dump_file
, "Fraction of visited symbols: %.2f%%\n\n",
3503 item_count
? 100.0f
* equal_items
/ item_count
: 0.0f
);
3507 congruence_class_group
*it
;
3508 FOR_EACH_VEC_ELT (classes
, l
, it
)
3509 for (unsigned int i
= 0; i
< it
->classes
.length (); i
++)
3511 congruence_class
*c
= it
->classes
[i
];
3513 if (c
->members
.length () == 1)
3516 sem_item
*source
= c
->members
[0];
3518 if (DECL_NAME (source
->decl
)
3519 && MAIN_NAME_P (DECL_NAME (source
->decl
)))
3520 /* If merge via wrappers, picking main as the target can be
3522 source
= c
->members
[1];
3524 for (unsigned int j
= 0; j
< c
->members
.length (); j
++)
3526 sem_item
*alias
= c
->members
[j
];
3528 if (alias
== source
)
3533 fprintf (dump_file
, "Semantic equality hit:%s->%s\n",
3534 xstrdup_for_dump (source
->node
->name ()),
3535 xstrdup_for_dump (alias
->node
->name ()));
3536 fprintf (dump_file
, "Assembler symbol names:%s->%s\n",
3537 xstrdup_for_dump (source
->node
->asm_name ()),
3538 xstrdup_for_dump (alias
->node
->asm_name ()));
3541 if (lookup_attribute ("no_icf", DECL_ATTRIBUTES (alias
->decl
)))
3545 "Merge operation is skipped due to no_icf "
3551 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3553 source
->dump_to_file (dump_file
);
3554 alias
->dump_to_file (dump_file
);
3557 if (dbg_cnt (merged_ipa_icf
))
3559 bool merged
= source
->merge (alias
);
3562 if (merged
&& alias
->type
== VAR
)
3564 symtab_pair p
= symtab_pair (source
->node
, alias
->node
);
3565 m_merged_variables
.safe_push (p
);
3571 if (!m_merged_variables
.is_empty ())
3572 fixup_points_to_sets ();
3577 /* Fixup points to set PT. */
3580 sem_item_optimizer::fixup_pt_set (struct pt_solution
*pt
)
3582 if (pt
->vars
== NULL
)
3587 FOR_EACH_VEC_ELT (m_merged_variables
, i
, item
)
3588 if (bitmap_bit_p (pt
->vars
, DECL_UID (item
->second
->decl
)))
3589 bitmap_set_bit (pt
->vars
, DECL_UID (item
->first
->decl
));
3592 /* Set all points-to UIDs of aliases pointing to node N as UID. */
3595 set_alias_uids (symtab_node
*n
, int uid
)
3598 FOR_EACH_ALIAS (n
, ref
)
3601 fprintf (dump_file
, " Setting points-to UID of [%s] as %d\n",
3602 xstrdup_for_dump (ref
->referring
->asm_name ()), uid
);
3604 SET_DECL_PT_UID (ref
->referring
->decl
, uid
);
3605 set_alias_uids (ref
->referring
, uid
);
3609 /* Fixup points to analysis info. */
3612 sem_item_optimizer::fixup_points_to_sets (void)
3614 /* TODO: remove in GCC 9 and trigger PTA re-creation after IPA passes. */
3617 FOR_EACH_DEFINED_FUNCTION (cnode
)
3621 function
*fn
= DECL_STRUCT_FUNCTION (cnode
->decl
);
3622 if (!gimple_in_ssa_p (fn
))
3625 FOR_EACH_SSA_NAME (i
, name
, fn
)
3626 if (POINTER_TYPE_P (TREE_TYPE (name
))
3627 && SSA_NAME_PTR_INFO (name
))
3628 fixup_pt_set (&SSA_NAME_PTR_INFO (name
)->pt
);
3629 fixup_pt_set (&fn
->gimple_df
->escaped
);
3631 /* The above get's us to 99% I guess, at least catching the
3632 address compares. Below also gets us aliasing correct
3633 but as said we're giving leeway to the situation with
3634 readonly vars anyway, so ... */
3636 FOR_EACH_BB_FN (bb
, fn
)
3637 for (gimple_stmt_iterator gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
);
3640 gcall
*call
= dyn_cast
<gcall
*> (gsi_stmt (gsi
));
3643 fixup_pt_set (gimple_call_use_set (call
));
3644 fixup_pt_set (gimple_call_clobber_set (call
));
3651 FOR_EACH_VEC_ELT (m_merged_variables
, i
, item
)
3652 set_alias_uids (item
->first
, DECL_UID (item
->first
->decl
));
3655 /* Dump function prints all class members to a FILE with an INDENT. */
3658 congruence_class::dump (FILE *file
, unsigned int indent
) const
3660 FPRINTF_SPACES (file
, indent
, "class with id: %u, hash: %u, items: %u\n",
3661 id
, members
[0]->get_hash (), members
.length ());
3663 FPUTS_SPACES (file
, indent
+ 2, "");
3664 for (unsigned i
= 0; i
< members
.length (); i
++)
3665 fprintf (file
, "%s ", members
[i
]->node
->dump_asm_name ());
3667 fprintf (file
, "\n");
3670 /* Returns true if there's a member that is used from another group. */
3673 congruence_class::is_class_used (void)
3675 for (unsigned int i
= 0; i
< members
.length (); i
++)
3676 if (members
[i
]->usages
.length ())
3682 /* Generate pass summary for IPA ICF pass. */
3685 ipa_icf_generate_summary (void)
3688 optimizer
= new sem_item_optimizer ();
3690 optimizer
->register_hooks ();
3691 optimizer
->parse_funcs_and_vars ();
3694 /* Write pass summary for IPA ICF pass. */
3697 ipa_icf_write_summary (void)
3699 gcc_assert (optimizer
);
3701 optimizer
->write_summary ();
3704 /* Read pass summary for IPA ICF pass. */
3707 ipa_icf_read_summary (void)
3710 optimizer
= new sem_item_optimizer ();
3712 optimizer
->read_summary ();
3713 optimizer
->register_hooks ();
3716 /* Semantic equality exection function. */
3719 ipa_icf_driver (void)
3721 gcc_assert (optimizer
);
3723 bool merged_p
= optimizer
->execute ();
3728 return merged_p
? TODO_remove_functions
: 0;
3731 const pass_data pass_data_ipa_icf
=
3733 IPA_PASS
, /* type */
3735 OPTGROUP_IPA
, /* optinfo_flags */
3736 TV_IPA_ICF
, /* tv_id */
3737 0, /* properties_required */
3738 0, /* properties_provided */
3739 0, /* properties_destroyed */
3740 0, /* todo_flags_start */
3741 0, /* todo_flags_finish */
3744 class pass_ipa_icf
: public ipa_opt_pass_d
3747 pass_ipa_icf (gcc::context
*ctxt
)
3748 : ipa_opt_pass_d (pass_data_ipa_icf
, ctxt
,
3749 ipa_icf_generate_summary
, /* generate_summary */
3750 ipa_icf_write_summary
, /* write_summary */
3751 ipa_icf_read_summary
, /* read_summary */
3753 write_optimization_summary */
3755 read_optimization_summary */
3756 NULL
, /* stmt_fixup */
3757 0, /* function_transform_todo_flags_start */
3758 NULL
, /* function_transform */
3759 NULL
) /* variable_transform */
3762 /* opt_pass methods: */
3763 virtual bool gate (function
*)
3765 return in_lto_p
|| flag_ipa_icf_variables
|| flag_ipa_icf_functions
;
3768 virtual unsigned int execute (function
*)
3770 return ipa_icf_driver();
3772 }; // class pass_ipa_icf
3774 } // ipa_icf namespace
3777 make_pass_ipa_icf (gcc::context
*ctxt
)
3779 return new ipa_icf::pass_ipa_icf (ctxt
);