1 /* Interprocedural Identical Code Folding pass
2 Copyright (C) 2014-2017 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-inline.h"
80 #include "print-tree.h"
81 #include "ipa-utils.h"
82 #include "ipa-icf-gimple.h"
84 #include "stor-layout.h"
87 using namespace ipa_icf_gimple
;
91 /* Initialization and computation of symtab node hash, there data
92 are propagated later on. */
94 static sem_item_optimizer
*optimizer
= NULL
;
98 symbol_compare_collection::symbol_compare_collection (symtab_node
*node
)
100 m_references
.create (0);
101 m_interposables
.create (0);
105 if (is_a
<varpool_node
*> (node
) && DECL_VIRTUAL_P (node
->decl
))
108 for (unsigned i
= 0; node
->iterate_reference (i
, ref
); i
++)
110 if (ref
->address_matters_p ())
111 m_references
.safe_push (ref
->referred
);
113 if (ref
->referred
->get_availability () <= AVAIL_INTERPOSABLE
)
115 if (ref
->address_matters_p ())
116 m_references
.safe_push (ref
->referred
);
118 m_interposables
.safe_push (ref
->referred
);
122 if (is_a
<cgraph_node
*> (node
))
124 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (node
);
126 for (cgraph_edge
*e
= cnode
->callees
; e
; e
= e
->next_callee
)
127 if (e
->callee
->get_availability () <= AVAIL_INTERPOSABLE
)
128 m_interposables
.safe_push (e
->callee
);
132 /* Constructor for key value pair, where _ITEM is key and _INDEX is a target. */
134 sem_usage_pair::sem_usage_pair (sem_item
*_item
, unsigned int _index
)
135 : item (_item
), index (_index
)
139 sem_item::sem_item (sem_item_type _type
, bitmap_obstack
*stack
)
140 : type (_type
), m_hash (-1), m_hash_set (false)
145 sem_item::sem_item (sem_item_type _type
, symtab_node
*_node
,
146 bitmap_obstack
*stack
)
147 : type (_type
), node (_node
), m_hash (-1), m_hash_set (false)
153 /* Add reference to a semantic TARGET. */
156 sem_item::add_reference (sem_item
*target
)
158 refs
.safe_push (target
);
159 unsigned index
= refs
.length ();
160 target
->usages
.safe_push (new sem_usage_pair(this, index
));
161 bitmap_set_bit (target
->usage_index_bitmap
, index
);
162 refs_set
.add (target
->node
);
165 /* Initialize internal data structures. Bitmap STACK is used for
166 bitmap memory allocation process. */
169 sem_item::setup (bitmap_obstack
*stack
)
171 gcc_checking_assert (node
);
174 tree_refs
.create (0);
176 usage_index_bitmap
= BITMAP_ALLOC (stack
);
179 sem_item::~sem_item ()
181 for (unsigned i
= 0; i
< usages
.length (); i
++)
185 tree_refs
.release ();
188 BITMAP_FREE (usage_index_bitmap
);
191 /* Dump function for debugging purpose. */
194 sem_item::dump (void)
198 fprintf (dump_file
, "[%s] %s (%u) (tree:%p)\n", type
== FUNC
? "func" : "var",
199 node
->name(), node
->order
, (void *) node
->decl
);
200 fprintf (dump_file
, " hash: %u\n", get_hash ());
201 fprintf (dump_file
, " references: ");
203 for (unsigned i
= 0; i
< refs
.length (); i
++)
204 fprintf (dump_file
, "%s%s ", refs
[i
]->node
->name (),
205 i
< refs
.length() - 1 ? "," : "");
207 fprintf (dump_file
, "\n");
211 /* Return true if target supports alias symbols. */
214 sem_item::target_supports_symbol_aliases_p (void)
216 #if !defined (ASM_OUTPUT_DEF) || (!defined(ASM_OUTPUT_WEAK_ALIAS) && !defined (ASM_WEAKEN_DECL))
223 void sem_item::set_hash (hashval_t hash
)
229 /* Semantic function constructor that uses STACK as bitmap memory stack. */
231 sem_function::sem_function (bitmap_obstack
*stack
)
232 : sem_item (FUNC
, stack
), m_checker (NULL
), m_compared_func (NULL
)
235 bb_sorted
.create (0);
238 sem_function::sem_function (cgraph_node
*node
, bitmap_obstack
*stack
)
239 : sem_item (FUNC
, node
, stack
), m_checker (NULL
), m_compared_func (NULL
)
242 bb_sorted
.create (0);
245 sem_function::~sem_function ()
247 for (unsigned i
= 0; i
< bb_sorted
.length (); i
++)
248 delete (bb_sorted
[i
]);
251 bb_sorted
.release ();
254 /* Calculates hash value based on a BASIC_BLOCK. */
257 sem_function::get_bb_hash (const sem_bb
*basic_block
)
259 inchash::hash hstate
;
261 hstate
.add_int (basic_block
->nondbg_stmt_count
);
262 hstate
.add_int (basic_block
->edge_count
);
264 return hstate
.end ();
267 /* References independent hash function. */
270 sem_function::get_hash (void)
274 inchash::hash hstate
;
275 hstate
.add_int (177454); /* Random number for function type. */
277 hstate
.add_int (arg_count
);
278 hstate
.add_int (cfg_checksum
);
279 hstate
.add_int (gcode_hash
);
281 for (unsigned i
= 0; i
< bb_sorted
.length (); i
++)
282 hstate
.merge_hash (get_bb_hash (bb_sorted
[i
]));
284 for (unsigned i
= 0; i
< bb_sizes
.length (); i
++)
285 hstate
.add_int (bb_sizes
[i
]);
287 /* Add common features of declaration itself. */
288 if (DECL_FUNCTION_SPECIFIC_TARGET (decl
))
290 (cl_target_option_hash
291 (TREE_TARGET_OPTION (DECL_FUNCTION_SPECIFIC_TARGET (decl
))));
292 if (DECL_FUNCTION_SPECIFIC_OPTIMIZATION (decl
))
294 (cl_optimization_hash
295 (TREE_OPTIMIZATION (DECL_FUNCTION_SPECIFIC_OPTIMIZATION (decl
))));
296 hstate
.add_flag (DECL_CXX_CONSTRUCTOR_P (decl
));
297 hstate
.add_flag (DECL_CXX_DESTRUCTOR_P (decl
));
299 set_hash (hstate
.end ());
305 /* Return ture if A1 and A2 represent equivalent function attribute lists.
306 Based on comp_type_attributes. */
309 sem_item::compare_attributes (const_tree a1
, const_tree a2
)
314 for (a
= a1
; a
!= NULL_TREE
; a
= TREE_CHAIN (a
))
316 const struct attribute_spec
*as
;
319 as
= lookup_attribute_spec (get_attribute_name (a
));
320 /* TODO: We can introduce as->affects_decl_identity
321 and as->affects_decl_reference_identity if attribute mismatch
322 gets a common reason to give up on merging. It may not be worth
324 For example returns_nonnull affects only references, while
325 optimize attribute can be ignored because it is already lowered
326 into flags representation and compared separately. */
330 attr
= lookup_attribute (as
->name
, CONST_CAST_TREE (a2
));
331 if (!attr
|| !attribute_value_equal (a
, attr
))
336 for (a
= a2
; a
!= NULL_TREE
; a
= TREE_CHAIN (a
))
338 const struct attribute_spec
*as
;
340 as
= lookup_attribute_spec (get_attribute_name (a
));
344 if (!lookup_attribute (as
->name
, CONST_CAST_TREE (a1
)))
346 /* We don't need to compare trees again, as we did this
347 already in first loop. */
352 /* TODO: As in comp_type_attributes we may want to introduce target hook. */
356 /* Compare properties of symbols N1 and N2 that does not affect semantics of
357 symbol itself but affects semantics of its references from USED_BY (which
358 may be NULL if it is unknown). If comparsion is false, symbols
359 can still be merged but any symbols referring them can't.
361 If ADDRESS is true, do extra checking needed for IPA_REF_ADDR.
363 TODO: We can also split attributes to those that determine codegen of
364 a function body/variable constructor itself and those that are used when
368 sem_item::compare_referenced_symbol_properties (symtab_node
*used_by
,
373 if (is_a
<cgraph_node
*> (n1
))
375 /* Inline properties matters: we do now want to merge uses of inline
376 function to uses of normal function because inline hint would be lost.
377 We however can merge inline function to noinline because the alias
378 will keep its DECL_DECLARED_INLINE flag.
380 Also ignore inline flag when optimizing for size or when function
381 is known to not be inlinable.
383 TODO: the optimize_size checks can also be assumed to be true if
384 unit has no !optimize_size functions. */
386 if ((!used_by
|| address
|| !is_a
<cgraph_node
*> (used_by
)
387 || !opt_for_fn (used_by
->decl
, optimize_size
))
388 && !opt_for_fn (n1
->decl
, optimize_size
)
389 && n1
->get_availability () > AVAIL_INTERPOSABLE
390 && (!DECL_UNINLINABLE (n1
->decl
) || !DECL_UNINLINABLE (n2
->decl
)))
392 if (DECL_DISREGARD_INLINE_LIMITS (n1
->decl
)
393 != DECL_DISREGARD_INLINE_LIMITS (n2
->decl
))
394 return return_false_with_msg
395 ("DECL_DISREGARD_INLINE_LIMITS are different");
397 if (DECL_DECLARED_INLINE_P (n1
->decl
)
398 != DECL_DECLARED_INLINE_P (n2
->decl
))
399 return return_false_with_msg ("inline attributes are different");
402 if (DECL_IS_OPERATOR_NEW (n1
->decl
)
403 != DECL_IS_OPERATOR_NEW (n2
->decl
))
404 return return_false_with_msg ("operator new flags are different");
407 /* Merging two definitions with a reference to equivalent vtables, but
408 belonging to a different type may result in ipa-polymorphic-call analysis
409 giving a wrong answer about the dynamic type of instance. */
410 if (is_a
<varpool_node
*> (n1
))
412 if ((DECL_VIRTUAL_P (n1
->decl
) || DECL_VIRTUAL_P (n2
->decl
))
413 && (DECL_VIRTUAL_P (n1
->decl
) != DECL_VIRTUAL_P (n2
->decl
)
414 || !types_must_be_same_for_odr (DECL_CONTEXT (n1
->decl
),
415 DECL_CONTEXT (n2
->decl
)))
416 && (!used_by
|| !is_a
<cgraph_node
*> (used_by
) || address
417 || opt_for_fn (used_by
->decl
, flag_devirtualize
)))
418 return return_false_with_msg
419 ("references to virtual tables can not be merged");
421 if (address
&& DECL_ALIGN (n1
->decl
) != DECL_ALIGN (n2
->decl
))
422 return return_false_with_msg ("alignment mismatch");
424 /* For functions we compare attributes in equals_wpa, because we do
425 not know what attributes may cause codegen differences, but for
426 variables just compare attributes for references - the codegen
427 for constructors is affected only by those attributes that we lower
428 to explicit representation (such as DECL_ALIGN or DECL_SECTION). */
429 if (!compare_attributes (DECL_ATTRIBUTES (n1
->decl
),
430 DECL_ATTRIBUTES (n2
->decl
)))
431 return return_false_with_msg ("different var decl attributes");
432 if (comp_type_attributes (TREE_TYPE (n1
->decl
),
433 TREE_TYPE (n2
->decl
)) != 1)
434 return return_false_with_msg ("different var type attributes");
437 /* When matching virtual tables, be sure to also match information
438 relevant for polymorphic call analysis. */
439 if (used_by
&& is_a
<varpool_node
*> (used_by
)
440 && DECL_VIRTUAL_P (used_by
->decl
))
442 if (DECL_VIRTUAL_P (n1
->decl
) != DECL_VIRTUAL_P (n2
->decl
))
443 return return_false_with_msg ("virtual flag mismatch");
444 if (DECL_VIRTUAL_P (n1
->decl
) && is_a
<cgraph_node
*> (n1
)
445 && (DECL_FINAL_P (n1
->decl
) != DECL_FINAL_P (n2
->decl
)))
446 return return_false_with_msg ("final flag mismatch");
451 /* Hash properties that are compared by compare_referenced_symbol_properties. */
454 sem_item::hash_referenced_symbol_properties (symtab_node
*ref
,
455 inchash::hash
&hstate
,
458 if (is_a
<cgraph_node
*> (ref
))
460 if ((type
!= FUNC
|| address
|| !opt_for_fn (decl
, optimize_size
))
461 && !opt_for_fn (ref
->decl
, optimize_size
)
462 && !DECL_UNINLINABLE (ref
->decl
))
464 hstate
.add_flag (DECL_DISREGARD_INLINE_LIMITS (ref
->decl
));
465 hstate
.add_flag (DECL_DECLARED_INLINE_P (ref
->decl
));
467 hstate
.add_flag (DECL_IS_OPERATOR_NEW (ref
->decl
));
469 else if (is_a
<varpool_node
*> (ref
))
471 hstate
.add_flag (DECL_VIRTUAL_P (ref
->decl
));
473 hstate
.add_int (DECL_ALIGN (ref
->decl
));
478 /* For a given symbol table nodes N1 and N2, we check that FUNCTION_DECLs
479 point to a same function. Comparison can be skipped if IGNORED_NODES
480 contains these nodes. ADDRESS indicate if address is taken. */
483 sem_item::compare_symbol_references (
484 hash_map
<symtab_node
*, sem_item
*> &ignored_nodes
,
485 symtab_node
*n1
, symtab_node
*n2
, bool address
)
487 enum availability avail1
, avail2
;
492 /* Never match variable and function. */
493 if (is_a
<varpool_node
*> (n1
) != is_a
<varpool_node
*> (n2
))
496 if (!compare_referenced_symbol_properties (node
, n1
, n2
, address
))
498 if (address
&& n1
->equal_address_to (n2
) == 1)
500 if (!address
&& n1
->semantically_equivalent_p (n2
))
503 n1
= n1
->ultimate_alias_target (&avail1
);
504 n2
= n2
->ultimate_alias_target (&avail2
);
506 if (avail1
> AVAIL_INTERPOSABLE
&& ignored_nodes
.get (n1
)
507 && avail2
> AVAIL_INTERPOSABLE
&& ignored_nodes
.get (n2
))
510 return return_false_with_msg ("different references");
513 /* If cgraph edges E1 and E2 are indirect calls, verify that
514 ECF flags are the same. */
516 bool sem_function::compare_edge_flags (cgraph_edge
*e1
, cgraph_edge
*e2
)
518 if (e1
->indirect_info
&& e2
->indirect_info
)
520 int e1_flags
= e1
->indirect_info
->ecf_flags
;
521 int e2_flags
= e2
->indirect_info
->ecf_flags
;
523 if (e1_flags
!= e2_flags
)
524 return return_false_with_msg ("ICF flags are different");
526 else if (e1
->indirect_info
|| e2
->indirect_info
)
532 /* Return true if parameter I may be used. */
535 sem_function::param_used_p (unsigned int i
)
537 if (ipa_node_params_sum
== NULL
)
540 struct ipa_node_params
*parms_info
= IPA_NODE_REF (get_node ());
542 if (vec_safe_length (parms_info
->descriptors
) <= i
)
545 return ipa_is_param_used (IPA_NODE_REF (get_node ()), i
);
548 /* Perform additional check needed to match types function parameters that are
549 used. Unlike for normal decls it matters if type is TYPE_RESTRICT and we
550 make an assumption that REFERENCE_TYPE parameters are always non-NULL. */
553 sem_function::compatible_parm_types_p (tree parm1
, tree parm2
)
555 /* Be sure that parameters are TBAA compatible. */
556 if (!func_checker::compatible_types_p (parm1
, parm2
))
557 return return_false_with_msg ("parameter type is not compatible");
559 if (POINTER_TYPE_P (parm1
)
560 && (TYPE_RESTRICT (parm1
) != TYPE_RESTRICT (parm2
)))
561 return return_false_with_msg ("argument restrict flag mismatch");
563 /* nonnull_arg_p implies non-zero range to REFERENCE types. */
564 if (POINTER_TYPE_P (parm1
)
565 && TREE_CODE (parm1
) != TREE_CODE (parm2
)
566 && opt_for_fn (decl
, flag_delete_null_pointer_checks
))
567 return return_false_with_msg ("pointer wrt reference mismatch");
572 /* Fast equality function based on knowledge known in WPA. */
575 sem_function::equals_wpa (sem_item
*item
,
576 hash_map
<symtab_node
*, sem_item
*> &ignored_nodes
)
578 gcc_assert (item
->type
== FUNC
);
579 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (node
);
580 cgraph_node
*cnode2
= dyn_cast
<cgraph_node
*> (item
->node
);
582 m_compared_func
= static_cast<sem_function
*> (item
);
584 if (cnode
->thunk
.thunk_p
!= cnode2
->thunk
.thunk_p
)
585 return return_false_with_msg ("thunk_p mismatch");
587 if (cnode
->thunk
.thunk_p
)
589 if (cnode
->thunk
.fixed_offset
!= cnode2
->thunk
.fixed_offset
)
590 return return_false_with_msg ("thunk fixed_offset mismatch");
591 if (cnode
->thunk
.virtual_value
!= cnode2
->thunk
.virtual_value
)
592 return return_false_with_msg ("thunk virtual_value mismatch");
593 if (cnode
->thunk
.this_adjusting
!= cnode2
->thunk
.this_adjusting
)
594 return return_false_with_msg ("thunk this_adjusting mismatch");
595 if (cnode
->thunk
.virtual_offset_p
!= cnode2
->thunk
.virtual_offset_p
)
596 return return_false_with_msg ("thunk virtual_offset_p mismatch");
597 if (cnode
->thunk
.add_pointer_bounds_args
598 != cnode2
->thunk
.add_pointer_bounds_args
)
599 return return_false_with_msg ("thunk add_pointer_bounds_args mismatch");
602 /* Compare special function DECL attributes. */
603 if (DECL_FUNCTION_PERSONALITY (decl
)
604 != DECL_FUNCTION_PERSONALITY (item
->decl
))
605 return return_false_with_msg ("function personalities are different");
607 if (DECL_NO_INSTRUMENT_FUNCTION_ENTRY_EXIT (decl
)
608 != DECL_NO_INSTRUMENT_FUNCTION_ENTRY_EXIT (item
->decl
))
609 return return_false_with_msg ("intrument function entry exit "
610 "attributes are different");
612 if (DECL_NO_LIMIT_STACK (decl
) != DECL_NO_LIMIT_STACK (item
->decl
))
613 return return_false_with_msg ("no stack limit attributes are different");
615 if (DECL_CXX_CONSTRUCTOR_P (decl
) != DECL_CXX_CONSTRUCTOR_P (item
->decl
))
616 return return_false_with_msg ("DECL_CXX_CONSTRUCTOR mismatch");
618 if (DECL_CXX_DESTRUCTOR_P (decl
) != DECL_CXX_DESTRUCTOR_P (item
->decl
))
619 return return_false_with_msg ("DECL_CXX_DESTRUCTOR mismatch");
621 /* TODO: pure/const flags mostly matters only for references, except for
622 the fact that codegen takes LOOPING flag as a hint that loops are
623 finite. We may arrange the code to always pick leader that has least
624 specified flags and then this can go into comparing symbol properties. */
625 if (flags_from_decl_or_type (decl
) != flags_from_decl_or_type (item
->decl
))
626 return return_false_with_msg ("decl_or_type flags are different");
628 /* Do not match polymorphic constructors of different types. They calls
629 type memory location for ipa-polymorphic-call and we do not want
630 it to get confused by wrong type. */
631 if (DECL_CXX_CONSTRUCTOR_P (decl
)
632 && TREE_CODE (TREE_TYPE (decl
)) == METHOD_TYPE
)
634 if (TREE_CODE (TREE_TYPE (item
->decl
)) != METHOD_TYPE
)
635 return return_false_with_msg ("DECL_CXX_CONSTURCTOR type mismatch");
636 else if (!func_checker::compatible_polymorphic_types_p
637 (TYPE_METHOD_BASETYPE (TREE_TYPE (decl
)),
638 TYPE_METHOD_BASETYPE (TREE_TYPE (item
->decl
)), false))
639 return return_false_with_msg ("ctor polymorphic type mismatch");
642 /* Checking function TARGET and OPTIMIZATION flags. */
643 cl_target_option
*tar1
= target_opts_for_fn (decl
);
644 cl_target_option
*tar2
= target_opts_for_fn (item
->decl
);
646 if (tar1
!= tar2
&& !cl_target_option_eq (tar1
, tar2
))
648 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
650 fprintf (dump_file
, "target flags difference");
651 cl_target_option_print_diff (dump_file
, 2, tar1
, tar2
);
654 return return_false_with_msg ("Target flags are different");
657 cl_optimization
*opt1
= opts_for_fn (decl
);
658 cl_optimization
*opt2
= opts_for_fn (item
->decl
);
660 if (opt1
!= opt2
&& memcmp (opt1
, opt2
, sizeof(cl_optimization
)))
662 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
664 fprintf (dump_file
, "optimization flags difference");
665 cl_optimization_print_diff (dump_file
, 2, opt1
, opt2
);
668 return return_false_with_msg ("optimization flags are different");
671 /* Result type checking. */
672 if (!func_checker::compatible_types_p
673 (TREE_TYPE (TREE_TYPE (decl
)),
674 TREE_TYPE (TREE_TYPE (m_compared_func
->decl
))))
675 return return_false_with_msg ("result types are different");
677 /* Checking types of arguments. */
678 tree list1
= TYPE_ARG_TYPES (TREE_TYPE (decl
)),
679 list2
= TYPE_ARG_TYPES (TREE_TYPE (m_compared_func
->decl
));
680 for (unsigned i
= 0; list1
&& list2
;
681 list1
= TREE_CHAIN (list1
), list2
= TREE_CHAIN (list2
), i
++)
683 tree parm1
= TREE_VALUE (list1
);
684 tree parm2
= TREE_VALUE (list2
);
686 /* This guard is here for function pointer with attributes (pr59927.c). */
687 if (!parm1
|| !parm2
)
688 return return_false_with_msg ("NULL argument type");
690 /* Verify that types are compatible to ensure that both functions
691 have same calling conventions. */
692 if (!types_compatible_p (parm1
, parm2
))
693 return return_false_with_msg ("parameter types are not compatible");
695 if (!param_used_p (i
))
698 /* Perform additional checks for used parameters. */
699 if (!compatible_parm_types_p (parm1
, parm2
))
704 return return_false_with_msg ("Mismatched number of parameters");
706 if (node
->num_references () != item
->node
->num_references ())
707 return return_false_with_msg ("different number of references");
709 /* Checking function attributes.
710 This is quadratic in number of attributes */
711 if (comp_type_attributes (TREE_TYPE (decl
),
712 TREE_TYPE (item
->decl
)) != 1)
713 return return_false_with_msg ("different type attributes");
714 if (!compare_attributes (DECL_ATTRIBUTES (decl
),
715 DECL_ATTRIBUTES (item
->decl
)))
716 return return_false_with_msg ("different decl attributes");
718 /* The type of THIS pointer type memory location for
719 ipa-polymorphic-call-analysis. */
720 if (opt_for_fn (decl
, flag_devirtualize
)
721 && (TREE_CODE (TREE_TYPE (decl
)) == METHOD_TYPE
722 || TREE_CODE (TREE_TYPE (item
->decl
)) == METHOD_TYPE
)
724 && compare_polymorphic_p ())
726 if (TREE_CODE (TREE_TYPE (decl
)) != TREE_CODE (TREE_TYPE (item
->decl
)))
727 return return_false_with_msg ("METHOD_TYPE and FUNCTION_TYPE mismatch");
728 if (!func_checker::compatible_polymorphic_types_p
729 (TYPE_METHOD_BASETYPE (TREE_TYPE (decl
)),
730 TYPE_METHOD_BASETYPE (TREE_TYPE (item
->decl
)), false))
731 return return_false_with_msg ("THIS pointer ODR type mismatch");
734 ipa_ref
*ref
= NULL
, *ref2
= NULL
;
735 for (unsigned i
= 0; node
->iterate_reference (i
, ref
); i
++)
737 item
->node
->iterate_reference (i
, ref2
);
739 if (ref
->use
!= ref2
->use
)
740 return return_false_with_msg ("reference use mismatch");
742 if (!compare_symbol_references (ignored_nodes
, ref
->referred
,
744 ref
->address_matters_p ()))
748 cgraph_edge
*e1
= dyn_cast
<cgraph_node
*> (node
)->callees
;
749 cgraph_edge
*e2
= dyn_cast
<cgraph_node
*> (item
->node
)->callees
;
753 if (!compare_symbol_references (ignored_nodes
, e1
->callee
,
756 if (!compare_edge_flags (e1
, e2
))
759 e1
= e1
->next_callee
;
760 e2
= e2
->next_callee
;
764 return return_false_with_msg ("different number of calls");
766 e1
= dyn_cast
<cgraph_node
*> (node
)->indirect_calls
;
767 e2
= dyn_cast
<cgraph_node
*> (item
->node
)->indirect_calls
;
771 if (!compare_edge_flags (e1
, e2
))
774 e1
= e1
->next_callee
;
775 e2
= e2
->next_callee
;
779 return return_false_with_msg ("different number of indirect calls");
784 /* Update hash by address sensitive references. We iterate over all
785 sensitive references (address_matters_p) and we hash ultime alias
786 target of these nodes, which can improve a semantic item hash.
788 Also hash in referenced symbols properties. This can be done at any time
789 (as the properties should not change), but it is convenient to do it here
790 while we walk the references anyway. */
793 sem_item::update_hash_by_addr_refs (hash_map
<symtab_node
*,
794 sem_item
*> &m_symtab_node_map
)
797 inchash::hash
hstate (get_hash ());
799 for (unsigned i
= 0; node
->iterate_reference (i
, ref
); i
++)
801 hstate
.add_int (ref
->use
);
802 hash_referenced_symbol_properties (ref
->referred
, hstate
,
803 ref
->use
== IPA_REF_ADDR
);
804 if (ref
->address_matters_p () || !m_symtab_node_map
.get (ref
->referred
))
805 hstate
.add_int (ref
->referred
->ultimate_alias_target ()->order
);
808 if (is_a
<cgraph_node
*> (node
))
810 for (cgraph_edge
*e
= dyn_cast
<cgraph_node
*> (node
)->callers
; e
;
813 sem_item
**result
= m_symtab_node_map
.get (e
->callee
);
814 hash_referenced_symbol_properties (e
->callee
, hstate
, false);
816 hstate
.add_int (e
->callee
->ultimate_alias_target ()->order
);
820 set_hash (hstate
.end ());
823 /* Update hash by computed local hash values taken from different
825 TODO: stronger SCC based hashing would be desirable here. */
828 sem_item::update_hash_by_local_refs (hash_map
<symtab_node
*,
829 sem_item
*> &m_symtab_node_map
)
832 inchash::hash
state (get_hash ());
834 for (unsigned j
= 0; node
->iterate_reference (j
, ref
); j
++)
836 sem_item
**result
= m_symtab_node_map
.get (ref
->referring
);
838 state
.merge_hash ((*result
)->get_hash ());
843 for (cgraph_edge
*e
= dyn_cast
<cgraph_node
*> (node
)->callees
; e
;
846 sem_item
**result
= m_symtab_node_map
.get (e
->caller
);
848 state
.merge_hash ((*result
)->get_hash ());
852 global_hash
= state
.end ();
855 /* Returns true if the item equals to ITEM given as argument. */
858 sem_function::equals (sem_item
*item
,
859 hash_map
<symtab_node
*, sem_item
*> &)
861 gcc_assert (item
->type
== FUNC
);
862 bool eq
= equals_private (item
);
864 if (m_checker
!= NULL
)
870 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
872 "Equals called for:%s:%s (%u:%u) (%s:%s) with result: %s\n\n",
873 xstrdup_for_dump (node
->name()),
874 xstrdup_for_dump (item
->node
->name ()),
877 xstrdup_for_dump (node
->asm_name ()),
878 xstrdup_for_dump (item
->node
->asm_name ()),
879 eq
? "true" : "false");
884 /* Processes function equality comparison. */
887 sem_function::equals_private (sem_item
*item
)
889 if (item
->type
!= FUNC
)
892 basic_block bb1
, bb2
;
894 edge_iterator ei1
, ei2
;
898 m_compared_func
= static_cast<sem_function
*> (item
);
900 gcc_assert (decl
!= item
->decl
);
902 if (bb_sorted
.length () != m_compared_func
->bb_sorted
.length ()
903 || edge_count
!= m_compared_func
->edge_count
904 || cfg_checksum
!= m_compared_func
->cfg_checksum
)
905 return return_false ();
907 m_checker
= new func_checker (decl
, m_compared_func
->decl
,
908 compare_polymorphic_p (),
911 &m_compared_func
->refs_set
);
912 arg1
= DECL_ARGUMENTS (decl
);
913 arg2
= DECL_ARGUMENTS (m_compared_func
->decl
);
915 arg1
&& arg2
; arg1
= DECL_CHAIN (arg1
), arg2
= DECL_CHAIN (arg2
), i
++)
917 if (!types_compatible_p (TREE_TYPE (arg1
), TREE_TYPE (arg2
)))
918 return return_false_with_msg ("argument types are not compatible");
919 if (!param_used_p (i
))
921 /* Perform additional checks for used parameters. */
922 if (!compatible_parm_types_p (TREE_TYPE (arg1
), TREE_TYPE (arg2
)))
924 if (!m_checker
->compare_decl (arg1
, arg2
))
925 return return_false ();
928 return return_false_with_msg ("Mismatched number of arguments");
930 if (!dyn_cast
<cgraph_node
*> (node
)->has_gimple_body_p ())
933 /* Fill-up label dictionary. */
934 for (unsigned i
= 0; i
< bb_sorted
.length (); ++i
)
936 m_checker
->parse_labels (bb_sorted
[i
]);
937 m_checker
->parse_labels (m_compared_func
->bb_sorted
[i
]);
940 /* Checking all basic blocks. */
941 for (unsigned i
= 0; i
< bb_sorted
.length (); ++i
)
942 if(!m_checker
->compare_bb (bb_sorted
[i
], m_compared_func
->bb_sorted
[i
]))
943 return return_false();
945 dump_message ("All BBs are equal\n");
947 auto_vec
<int> bb_dict
;
949 /* Basic block edges check. */
950 for (unsigned i
= 0; i
< bb_sorted
.length (); ++i
)
952 bb1
= bb_sorted
[i
]->bb
;
953 bb2
= m_compared_func
->bb_sorted
[i
]->bb
;
955 ei2
= ei_start (bb2
->preds
);
957 for (ei1
= ei_start (bb1
->preds
); ei_cond (ei1
, &e1
); ei_next (&ei1
))
961 if (e1
->flags
!= e2
->flags
)
962 return return_false_with_msg ("flags comparison returns false");
964 if (!bb_dict_test (&bb_dict
, e1
->src
->index
, e2
->src
->index
))
965 return return_false_with_msg ("edge comparison returns false");
967 if (!bb_dict_test (&bb_dict
, e1
->dest
->index
, e2
->dest
->index
))
968 return return_false_with_msg ("BB comparison returns false");
970 if (!m_checker
->compare_edge (e1
, e2
))
971 return return_false_with_msg ("edge comparison returns false");
977 /* Basic block PHI nodes comparison. */
978 for (unsigned i
= 0; i
< bb_sorted
.length (); i
++)
979 if (!compare_phi_node (bb_sorted
[i
]->bb
, m_compared_func
->bb_sorted
[i
]->bb
))
980 return return_false_with_msg ("PHI node comparison returns false");
985 /* Set LOCAL_P of NODE to true if DATA is non-NULL.
986 Helper for call_for_symbol_thunks_and_aliases. */
989 set_local (cgraph_node
*node
, void *data
)
991 node
->local
.local
= data
!= NULL
;
995 /* TREE_ADDRESSABLE of NODE to true.
996 Helper for call_for_symbol_thunks_and_aliases. */
999 set_addressable (varpool_node
*node
, void *)
1001 TREE_ADDRESSABLE (node
->decl
) = 1;
1005 /* Clear DECL_RTL of NODE.
1006 Helper for call_for_symbol_thunks_and_aliases. */
1009 clear_decl_rtl (symtab_node
*node
, void *)
1011 SET_DECL_RTL (node
->decl
, NULL
);
1015 /* Redirect all callers of N and its aliases to TO. Remove aliases if
1016 possible. Return number of redirections made. */
1019 redirect_all_callers (cgraph_node
*n
, cgraph_node
*to
)
1021 int nredirected
= 0;
1023 cgraph_edge
*e
= n
->callers
;
1027 /* Redirecting thunks to interposable symbols or symbols in other sections
1028 may not be supported by target output code. Play safe for now and
1029 punt on redirection. */
1030 if (!e
->caller
->thunk
.thunk_p
)
1032 struct cgraph_edge
*nexte
= e
->next_caller
;
1033 e
->redirect_callee (to
);
1040 for (unsigned i
= 0; n
->iterate_direct_aliases (i
, ref
);)
1042 bool removed
= false;
1043 cgraph_node
*n_alias
= dyn_cast
<cgraph_node
*> (ref
->referring
);
1045 if ((DECL_COMDAT_GROUP (n
->decl
)
1046 && (DECL_COMDAT_GROUP (n
->decl
)
1047 == DECL_COMDAT_GROUP (n_alias
->decl
)))
1048 || (n_alias
->get_availability () > AVAIL_INTERPOSABLE
1049 && n
->get_availability () > AVAIL_INTERPOSABLE
))
1051 nredirected
+= redirect_all_callers (n_alias
, to
);
1052 if (n_alias
->can_remove_if_no_direct_calls_p ()
1053 && !n_alias
->call_for_symbol_and_aliases (cgraph_node::has_thunk_p
,
1055 && !n_alias
->has_aliases_p ())
1064 /* Merges instance with an ALIAS_ITEM, where alias, thunk or redirection can
1068 sem_function::merge (sem_item
*alias_item
)
1070 gcc_assert (alias_item
->type
== FUNC
);
1072 sem_function
*alias_func
= static_cast<sem_function
*> (alias_item
);
1074 cgraph_node
*original
= get_node ();
1075 cgraph_node
*local_original
= NULL
;
1076 cgraph_node
*alias
= alias_func
->get_node ();
1078 bool create_wrapper
= false;
1079 bool create_alias
= false;
1080 bool redirect_callers
= false;
1081 bool remove
= false;
1083 bool original_discardable
= false;
1084 bool original_discarded
= false;
1086 bool original_address_matters
= original
->address_matters_p ();
1087 bool alias_address_matters
= alias
->address_matters_p ();
1089 if (DECL_EXTERNAL (alias
->decl
))
1092 fprintf (dump_file
, "Not unifying; alias is external.\n\n");
1096 if (DECL_NO_INLINE_WARNING_P (original
->decl
)
1097 != DECL_NO_INLINE_WARNING_P (alias
->decl
))
1102 "DECL_NO_INLINE_WARNING mismatch.\n\n");
1106 /* Do not attempt to mix functions from different user sections;
1107 we do not know what user intends with those. */
1108 if (((DECL_SECTION_NAME (original
->decl
) && !original
->implicit_section
)
1109 || (DECL_SECTION_NAME (alias
->decl
) && !alias
->implicit_section
))
1110 && DECL_SECTION_NAME (original
->decl
) != DECL_SECTION_NAME (alias
->decl
))
1115 "original and alias are in different sections.\n\n");
1119 /* See if original is in a section that can be discarded if the main
1120 symbol is not used. */
1122 if (original
->can_be_discarded_p ())
1123 original_discardable
= true;
1124 /* Also consider case where we have resolution info and we know that
1125 original's definition is not going to be used. In this case we can not
1126 create alias to original. */
1127 if (node
->resolution
!= LDPR_UNKNOWN
1128 && !decl_binds_to_current_def_p (node
->decl
))
1129 original_discardable
= original_discarded
= true;
1131 /* Creating a symtab alias is the optimal way to merge.
1132 It however can not be used in the following cases:
1134 1) if ORIGINAL and ALIAS may be possibly compared for address equality.
1135 2) if ORIGINAL is in a section that may be discarded by linker or if
1136 it is an external functions where we can not create an alias
1137 (ORIGINAL_DISCARDABLE)
1138 3) if target do not support symbol aliases.
1139 4) original and alias lie in different comdat groups.
1141 If we can not produce alias, we will turn ALIAS into WRAPPER of ORIGINAL
1142 and/or redirect all callers from ALIAS to ORIGINAL. */
1143 if ((original_address_matters
&& alias_address_matters
)
1144 || (original_discardable
1145 && (!DECL_COMDAT_GROUP (alias
->decl
)
1146 || (DECL_COMDAT_GROUP (alias
->decl
)
1147 != DECL_COMDAT_GROUP (original
->decl
))))
1148 || original_discarded
1149 || !sem_item::target_supports_symbol_aliases_p ()
1150 || DECL_COMDAT_GROUP (alias
->decl
) != DECL_COMDAT_GROUP (original
->decl
))
1152 /* First see if we can produce wrapper. */
1154 /* Symbol properties that matter for references must be preserved.
1155 TODO: We can produce wrapper, but we need to produce alias of ORIGINAL
1156 with proper properties. */
1157 if (!sem_item::compare_referenced_symbol_properties (NULL
, original
, alias
,
1158 alias
->address_taken
))
1162 "Wrapper cannot be created because referenced symbol "
1163 "properties mismatch\n");
1165 /* Do not turn function in one comdat group into wrapper to another
1166 comdat group. Other compiler producing the body of the
1167 another comdat group may make opossite decision and with unfortunate
1168 linker choices this may close a loop. */
1169 else if (DECL_COMDAT_GROUP (original
->decl
)
1170 && DECL_COMDAT_GROUP (alias
->decl
)
1171 && (DECL_COMDAT_GROUP (alias
->decl
)
1172 != DECL_COMDAT_GROUP (original
->decl
)))
1176 "Wrapper cannot be created because of COMDAT\n");
1178 else if (DECL_STATIC_CHAIN (alias
->decl
)
1179 || DECL_STATIC_CHAIN (original
->decl
))
1183 "Cannot create wrapper of nested function.\n");
1185 /* TODO: We can also deal with variadic functions never calling
1187 else if (stdarg_p (TREE_TYPE (alias
->decl
)))
1191 "can not create wrapper of stdarg function.\n");
1193 else if (inline_summaries
1194 && inline_summaries
->get (alias
)->self_size
<= 2)
1197 fprintf (dump_file
, "Wrapper creation is not "
1198 "profitable (function is too small).\n");
1200 /* If user paid attention to mark function noinline, assume it is
1201 somewhat special and do not try to turn it into a wrapper that can
1202 not be undone by inliner. */
1203 else if (lookup_attribute ("noinline", DECL_ATTRIBUTES (alias
->decl
)))
1206 fprintf (dump_file
, "Wrappers are not created for noinline.\n");
1209 create_wrapper
= true;
1211 /* We can redirect local calls in the case both alias and orignal
1212 are not interposable. */
1214 = alias
->get_availability () > AVAIL_INTERPOSABLE
1215 && original
->get_availability () > AVAIL_INTERPOSABLE
1216 && !alias
->instrumented_version
;
1217 /* TODO: We can redirect, but we need to produce alias of ORIGINAL
1218 with proper properties. */
1219 if (!sem_item::compare_referenced_symbol_properties (NULL
, original
, alias
,
1220 alias
->address_taken
))
1221 redirect_callers
= false;
1223 if (!redirect_callers
&& !create_wrapper
)
1226 fprintf (dump_file
, "Not unifying; can not redirect callers nor "
1227 "produce wrapper\n\n");
1231 /* Work out the symbol the wrapper should call.
1232 If ORIGINAL is interposable, we need to call a local alias.
1233 Also produce local alias (if possible) as an optimization.
1235 Local aliases can not be created inside comdat groups because that
1236 prevents inlining. */
1237 if (!original_discardable
&& !original
->get_comdat_group ())
1240 = dyn_cast
<cgraph_node
*> (original
->noninterposable_alias ());
1242 && original
->get_availability () > AVAIL_INTERPOSABLE
)
1243 local_original
= original
;
1245 /* If we can not use local alias, fallback to the original
1247 else if (original
->get_availability () > AVAIL_INTERPOSABLE
)
1248 local_original
= original
;
1250 /* If original is COMDAT local, we can not really redirect calls outside
1251 of its comdat group to it. */
1252 if (original
->comdat_local_p ())
1253 redirect_callers
= false;
1254 if (!local_original
)
1257 fprintf (dump_file
, "Not unifying; "
1258 "can not produce local alias.\n\n");
1262 if (!redirect_callers
&& !create_wrapper
)
1265 fprintf (dump_file
, "Not unifying; "
1266 "can not redirect callers nor produce a wrapper\n\n");
1270 && !alias
->call_for_symbol_and_aliases (cgraph_node::has_thunk_p
,
1272 && !alias
->can_remove_if_no_direct_calls_p ())
1275 fprintf (dump_file
, "Not unifying; can not make wrapper and "
1276 "function has other uses than direct calls\n\n");
1281 create_alias
= true;
1283 if (redirect_callers
)
1285 int nredirected
= redirect_all_callers (alias
, local_original
);
1289 alias
->icf_merged
= true;
1290 local_original
->icf_merged
= true;
1292 if (dump_file
&& nredirected
)
1293 fprintf (dump_file
, "%i local calls have been "
1294 "redirected.\n", nredirected
);
1297 /* If all callers was redirected, do not produce wrapper. */
1298 if (alias
->can_remove_if_no_direct_calls_p ()
1299 && !DECL_VIRTUAL_P (alias
->decl
)
1300 && !alias
->has_aliases_p ())
1302 create_wrapper
= false;
1305 gcc_assert (!create_alias
);
1307 else if (create_alias
)
1309 alias
->icf_merged
= true;
1311 /* Remove the function's body. */
1312 ipa_merge_profiles (original
, alias
);
1313 alias
->release_body (true);
1315 /* Notice global symbol possibly produced RTL. */
1316 ((symtab_node
*)alias
)->call_for_symbol_and_aliases (clear_decl_rtl
,
1319 /* Create the alias. */
1320 cgraph_node::create_alias (alias_func
->decl
, decl
);
1321 alias
->resolve_alias (original
);
1323 original
->call_for_symbol_thunks_and_aliases
1324 (set_local
, (void *)(size_t) original
->local_p (), true);
1327 fprintf (dump_file
, "Unified; Function alias has been created.\n\n");
1331 gcc_assert (!create_alias
);
1332 alias
->icf_merged
= true;
1333 local_original
->icf_merged
= true;
1335 ipa_merge_profiles (local_original
, alias
, true);
1336 alias
->create_wrapper (local_original
);
1339 fprintf (dump_file
, "Unified; Wrapper has been created.\n\n");
1342 /* It's possible that redirection can hit thunks that block
1343 redirection opportunities. */
1344 gcc_assert (alias
->icf_merged
|| remove
|| redirect_callers
);
1345 original
->icf_merged
= true;
1347 /* We use merged flag to track cases where COMDAT function is known to be
1348 compatible its callers. If we merged in non-COMDAT, we need to give up
1349 on this optimization. */
1350 if (original
->merged_comdat
&& !alias
->merged_comdat
)
1353 fprintf (dump_file
, "Dropping merged_comdat flag.\n\n");
1355 local_original
->merged_comdat
= false;
1356 original
->merged_comdat
= false;
1361 ipa_merge_profiles (original
, alias
);
1362 alias
->release_body ();
1364 alias
->body_removed
= true;
1365 alias
->icf_merged
= true;
1367 fprintf (dump_file
, "Unified; Function body was removed.\n");
1373 /* Semantic item initialization function. */
1376 sem_function::init (void)
1379 get_node ()->get_untransformed_body ();
1381 tree fndecl
= node
->decl
;
1382 function
*func
= DECL_STRUCT_FUNCTION (fndecl
);
1385 gcc_assert (SSANAMES (func
));
1387 ssa_names_size
= SSANAMES (func
)->length ();
1391 region_tree
= func
->eh
->region_tree
;
1393 /* iterating all function arguments. */
1394 arg_count
= count_formal_params (fndecl
);
1396 edge_count
= n_edges_for_fn (func
);
1397 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (node
);
1398 if (!cnode
->thunk
.thunk_p
)
1400 cfg_checksum
= coverage_compute_cfg_checksum (func
);
1402 inchash::hash hstate
;
1405 FOR_EACH_BB_FN (bb
, func
)
1407 unsigned nondbg_stmt_count
= 0;
1410 for (edge_iterator ei
= ei_start (bb
->preds
); ei_cond (ei
, &e
);
1412 cfg_checksum
= iterative_hash_host_wide_int (e
->flags
,
1415 for (gimple_stmt_iterator gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
);
1418 gimple
*stmt
= gsi_stmt (gsi
);
1420 if (gimple_code (stmt
) != GIMPLE_DEBUG
1421 && gimple_code (stmt
) != GIMPLE_PREDICT
)
1423 hash_stmt (stmt
, hstate
);
1424 nondbg_stmt_count
++;
1428 gcode_hash
= hstate
.end ();
1429 bb_sizes
.safe_push (nondbg_stmt_count
);
1431 /* Inserting basic block to hash table. */
1432 sem_bb
*semantic_bb
= new sem_bb (bb
, nondbg_stmt_count
,
1433 EDGE_COUNT (bb
->preds
)
1434 + EDGE_COUNT (bb
->succs
));
1436 bb_sorted
.safe_push (semantic_bb
);
1442 inchash::hash hstate
;
1443 hstate
.add_wide_int (cnode
->thunk
.fixed_offset
);
1444 hstate
.add_wide_int (cnode
->thunk
.virtual_value
);
1445 hstate
.add_flag (cnode
->thunk
.this_adjusting
);
1446 hstate
.add_flag (cnode
->thunk
.virtual_offset_p
);
1447 hstate
.add_flag (cnode
->thunk
.add_pointer_bounds_args
);
1448 gcode_hash
= hstate
.end ();
1452 /* Accumulate to HSTATE a hash of expression EXP.
1453 Identical to inchash::add_expr, but guaranteed to be stable across LTO
1454 and DECL equality classes. */
1457 sem_item::add_expr (const_tree exp
, inchash::hash
&hstate
)
1459 if (exp
== NULL_TREE
)
1461 hstate
.merge_hash (0);
1465 /* Handled component can be matched in a cureful way proving equivalence
1466 even if they syntactically differ. Just skip them. */
1468 while (handled_component_p (exp
))
1469 exp
= TREE_OPERAND (exp
, 0);
1471 enum tree_code code
= TREE_CODE (exp
);
1472 hstate
.add_int (code
);
1476 /* Use inchash::add_expr for everything that is LTO stable. */
1484 inchash::add_expr (exp
, hstate
);
1488 unsigned HOST_WIDE_INT idx
;
1491 hstate
.add_wide_int (int_size_in_bytes (TREE_TYPE (exp
)));
1493 FOR_EACH_CONSTRUCTOR_VALUE (CONSTRUCTOR_ELTS (exp
), idx
, value
)
1495 add_expr (value
, hstate
);
1500 add_expr (get_base_address (TREE_OPERAND (exp
, 0)), hstate
);
1506 hstate
.add_wide_int (int_size_in_bytes (TREE_TYPE (exp
)));
1509 case POINTER_PLUS_EXPR
:
1512 add_expr (TREE_OPERAND (exp
, 0), hstate
);
1513 add_expr (TREE_OPERAND (exp
, 1), hstate
);
1517 inchash::hash one
, two
;
1518 add_expr (TREE_OPERAND (exp
, 0), one
);
1519 add_expr (TREE_OPERAND (exp
, 1), two
);
1520 hstate
.add_commutative (one
, two
);
1524 hstate
.add_wide_int (int_size_in_bytes (TREE_TYPE (exp
)));
1525 return add_expr (TREE_OPERAND (exp
, 0), hstate
);
1531 /* Accumulate to HSTATE a hash of type t.
1532 TYpes that may end up being compatible after LTO type merging needs to have
1536 sem_item::add_type (const_tree type
, inchash::hash
&hstate
)
1538 if (type
== NULL_TREE
)
1540 hstate
.merge_hash (0);
1544 type
= TYPE_MAIN_VARIANT (type
);
1546 hstate
.add_int (TYPE_MODE (type
));
1548 if (TREE_CODE (type
) == COMPLEX_TYPE
)
1550 hstate
.add_int (COMPLEX_TYPE
);
1551 sem_item::add_type (TREE_TYPE (type
), hstate
);
1553 else if (INTEGRAL_TYPE_P (type
))
1555 hstate
.add_int (INTEGER_TYPE
);
1556 hstate
.add_flag (TYPE_UNSIGNED (type
));
1557 hstate
.add_int (TYPE_PRECISION (type
));
1559 else if (VECTOR_TYPE_P (type
))
1561 hstate
.add_int (VECTOR_TYPE
);
1562 hstate
.add_int (TYPE_PRECISION (type
));
1563 sem_item::add_type (TREE_TYPE (type
), hstate
);
1565 else if (TREE_CODE (type
) == ARRAY_TYPE
)
1567 hstate
.add_int (ARRAY_TYPE
);
1568 /* Do not hash size, so complete and incomplete types can match. */
1569 sem_item::add_type (TREE_TYPE (type
), hstate
);
1571 else if (RECORD_OR_UNION_TYPE_P (type
))
1573 gcc_checking_assert (COMPLETE_TYPE_P (type
));
1574 hashval_t
*val
= optimizer
->m_type_hash_cache
.get (type
);
1578 inchash::hash hstate2
;
1583 hstate2
.add_int (RECORD_TYPE
);
1584 gcc_assert (COMPLETE_TYPE_P (type
));
1586 for (f
= TYPE_FIELDS (type
), nf
= 0; f
; f
= TREE_CHAIN (f
))
1587 if (TREE_CODE (f
) == FIELD_DECL
)
1589 add_type (TREE_TYPE (f
), hstate2
);
1593 hstate2
.add_int (nf
);
1594 hash
= hstate2
.end ();
1595 hstate
.add_wide_int (hash
);
1596 optimizer
->m_type_hash_cache
.put (type
, hash
);
1599 hstate
.add_wide_int (*val
);
1603 /* Improve accumulated hash for HSTATE based on a gimple statement STMT. */
1606 sem_function::hash_stmt (gimple
*stmt
, inchash::hash
&hstate
)
1608 enum gimple_code code
= gimple_code (stmt
);
1610 hstate
.add_int (code
);
1615 add_expr (gimple_switch_index (as_a
<gswitch
*> (stmt
)), hstate
);
1618 hstate
.add_int (gimple_assign_rhs_code (stmt
));
1619 if (commutative_tree_code (gimple_assign_rhs_code (stmt
))
1620 || commutative_ternary_tree_code (gimple_assign_rhs_code (stmt
)))
1622 inchash::hash one
, two
;
1624 add_expr (gimple_assign_rhs1 (stmt
), one
);
1625 add_type (TREE_TYPE (gimple_assign_rhs1 (stmt
)), one
);
1626 add_expr (gimple_assign_rhs2 (stmt
), two
);
1627 hstate
.add_commutative (one
, two
);
1628 if (commutative_ternary_tree_code (gimple_assign_rhs_code (stmt
)))
1630 add_expr (gimple_assign_rhs3 (stmt
), hstate
);
1631 add_type (TREE_TYPE (gimple_assign_rhs3 (stmt
)), hstate
);
1633 add_expr (gimple_assign_lhs (stmt
), hstate
);
1634 add_type (TREE_TYPE (gimple_assign_lhs (stmt
)), two
);
1643 /* All these statements are equivalent if their operands are. */
1644 for (unsigned i
= 0; i
< gimple_num_ops (stmt
); ++i
)
1646 add_expr (gimple_op (stmt
, i
), hstate
);
1647 if (gimple_op (stmt
, i
))
1648 add_type (TREE_TYPE (gimple_op (stmt
, i
)), hstate
);
1656 /* Return true if polymorphic comparison must be processed. */
1659 sem_function::compare_polymorphic_p (void)
1661 struct cgraph_edge
*e
;
1663 if (!opt_for_fn (get_node ()->decl
, flag_devirtualize
))
1665 if (get_node ()->indirect_calls
!= NULL
)
1667 /* TODO: We can do simple propagation determining what calls may lead to
1668 a polymorphic call. */
1669 for (e
= get_node ()->callees
; e
; e
= e
->next_callee
)
1670 if (e
->callee
->definition
1671 && opt_for_fn (e
->callee
->decl
, flag_devirtualize
))
1676 /* For a given call graph NODE, the function constructs new
1677 semantic function item. */
1680 sem_function::parse (cgraph_node
*node
, bitmap_obstack
*stack
)
1682 tree fndecl
= node
->decl
;
1683 function
*func
= DECL_STRUCT_FUNCTION (fndecl
);
1685 if (!func
|| (!node
->has_gimple_body_p () && !node
->thunk
.thunk_p
))
1688 if (lookup_attribute_by_prefix ("omp ", DECL_ATTRIBUTES (node
->decl
)) != NULL
)
1691 if (lookup_attribute_by_prefix ("oacc ",
1692 DECL_ATTRIBUTES (node
->decl
)) != NULL
)
1696 if (DECL_STATIC_CONSTRUCTOR (node
->decl
)
1697 || DECL_STATIC_DESTRUCTOR (node
->decl
))
1700 sem_function
*f
= new sem_function (node
, stack
);
1707 /* For given basic blocks BB1 and BB2 (from functions FUNC1 and FUNC),
1708 return true if phi nodes are semantically equivalent in these blocks . */
1711 sem_function::compare_phi_node (basic_block bb1
, basic_block bb2
)
1713 gphi_iterator si1
, si2
;
1715 unsigned size1
, size2
, i
;
1719 gcc_assert (bb1
!= NULL
);
1720 gcc_assert (bb2
!= NULL
);
1722 si2
= gsi_start_phis (bb2
);
1723 for (si1
= gsi_start_phis (bb1
); !gsi_end_p (si1
);
1726 gsi_next_nonvirtual_phi (&si1
);
1727 gsi_next_nonvirtual_phi (&si2
);
1729 if (gsi_end_p (si1
) && gsi_end_p (si2
))
1732 if (gsi_end_p (si1
) || gsi_end_p (si2
))
1733 return return_false();
1738 tree phi_result1
= gimple_phi_result (phi1
);
1739 tree phi_result2
= gimple_phi_result (phi2
);
1741 if (!m_checker
->compare_operand (phi_result1
, phi_result2
))
1742 return return_false_with_msg ("PHI results are different");
1744 size1
= gimple_phi_num_args (phi1
);
1745 size2
= gimple_phi_num_args (phi2
);
1748 return return_false ();
1750 for (i
= 0; i
< size1
; ++i
)
1752 t1
= gimple_phi_arg (phi1
, i
)->def
;
1753 t2
= gimple_phi_arg (phi2
, i
)->def
;
1755 if (!m_checker
->compare_operand (t1
, t2
))
1756 return return_false ();
1758 e1
= gimple_phi_arg_edge (phi1
, i
);
1759 e2
= gimple_phi_arg_edge (phi2
, i
);
1761 if (!m_checker
->compare_edge (e1
, e2
))
1762 return return_false ();
1771 /* Returns true if tree T can be compared as a handled component. */
1774 sem_function::icf_handled_component_p (tree t
)
1776 tree_code tc
= TREE_CODE (t
);
1778 return (handled_component_p (t
)
1779 || tc
== ADDR_EXPR
|| tc
== MEM_REF
|| tc
== OBJ_TYPE_REF
);
1782 /* Basic blocks dictionary BB_DICT returns true if SOURCE index BB
1783 corresponds to TARGET. */
1786 sem_function::bb_dict_test (vec
<int> *bb_dict
, int source
, int target
)
1791 if (bb_dict
->length () <= (unsigned)source
)
1792 bb_dict
->safe_grow_cleared (source
+ 1);
1794 if ((*bb_dict
)[source
] == 0)
1796 (*bb_dict
)[source
] = target
;
1800 return (*bb_dict
)[source
] == target
;
1803 sem_variable::sem_variable (bitmap_obstack
*stack
): sem_item (VAR
, stack
)
1807 sem_variable::sem_variable (varpool_node
*node
, bitmap_obstack
*stack
)
1808 : sem_item (VAR
, node
, stack
)
1810 gcc_checking_assert (node
);
1811 gcc_checking_assert (get_node ());
1814 /* Fast equality function based on knowledge known in WPA. */
1817 sem_variable::equals_wpa (sem_item
*item
,
1818 hash_map
<symtab_node
*, sem_item
*> &ignored_nodes
)
1820 gcc_assert (item
->type
== VAR
);
1822 if (node
->num_references () != item
->node
->num_references ())
1823 return return_false_with_msg ("different number of references");
1825 if (DECL_TLS_MODEL (decl
) || DECL_TLS_MODEL (item
->decl
))
1826 return return_false_with_msg ("TLS model");
1828 /* DECL_ALIGN is safe to merge, because we will always chose the largest
1829 alignment out of all aliases. */
1831 if (DECL_VIRTUAL_P (decl
) != DECL_VIRTUAL_P (item
->decl
))
1832 return return_false_with_msg ("Virtual flag mismatch");
1834 if (DECL_SIZE (decl
) != DECL_SIZE (item
->decl
)
1835 && ((!DECL_SIZE (decl
) || !DECL_SIZE (item
->decl
))
1836 || !operand_equal_p (DECL_SIZE (decl
),
1837 DECL_SIZE (item
->decl
), OEP_ONLY_CONST
)))
1838 return return_false_with_msg ("size mismatch");
1840 /* Do not attempt to mix data from different user sections;
1841 we do not know what user intends with those. */
1842 if (((DECL_SECTION_NAME (decl
) && !node
->implicit_section
)
1843 || (DECL_SECTION_NAME (item
->decl
) && !item
->node
->implicit_section
))
1844 && DECL_SECTION_NAME (decl
) != DECL_SECTION_NAME (item
->decl
))
1845 return return_false_with_msg ("user section mismatch");
1847 if (DECL_IN_TEXT_SECTION (decl
) != DECL_IN_TEXT_SECTION (item
->decl
))
1848 return return_false_with_msg ("text section");
1850 ipa_ref
*ref
= NULL
, *ref2
= NULL
;
1851 for (unsigned i
= 0; node
->iterate_reference (i
, ref
); i
++)
1853 item
->node
->iterate_reference (i
, ref2
);
1855 if (ref
->use
!= ref2
->use
)
1856 return return_false_with_msg ("reference use mismatch");
1858 if (!compare_symbol_references (ignored_nodes
,
1859 ref
->referred
, ref2
->referred
,
1860 ref
->address_matters_p ()))
1867 /* Returns true if the item equals to ITEM given as argument. */
1870 sem_variable::equals (sem_item
*item
,
1871 hash_map
<symtab_node
*, sem_item
*> &)
1873 gcc_assert (item
->type
== VAR
);
1876 if (DECL_INITIAL (decl
) == error_mark_node
&& in_lto_p
)
1877 dyn_cast
<varpool_node
*>(node
)->get_constructor ();
1878 if (DECL_INITIAL (item
->decl
) == error_mark_node
&& in_lto_p
)
1879 dyn_cast
<varpool_node
*>(item
->node
)->get_constructor ();
1881 /* As seen in PR ipa/65303 we have to compare variables types. */
1882 if (!func_checker::compatible_types_p (TREE_TYPE (decl
),
1883 TREE_TYPE (item
->decl
)))
1884 return return_false_with_msg ("variables types are different");
1886 ret
= sem_variable::equals (DECL_INITIAL (decl
),
1887 DECL_INITIAL (item
->node
->decl
));
1888 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1890 "Equals called for vars:%s:%s (%u:%u) (%s:%s) with result: %s\n\n",
1891 xstrdup_for_dump (node
->name()),
1892 xstrdup_for_dump (item
->node
->name ()),
1893 node
->order
, item
->node
->order
,
1894 xstrdup_for_dump (node
->asm_name ()),
1895 xstrdup_for_dump (item
->node
->asm_name ()), ret
? "true" : "false");
1900 /* Compares trees T1 and T2 for semantic equality. */
1903 sem_variable::equals (tree t1
, tree t2
)
1906 return return_with_debug (t1
== t2
);
1909 tree_code tc1
= TREE_CODE (t1
);
1910 tree_code tc2
= TREE_CODE (t2
);
1913 return return_false_with_msg ("TREE_CODE mismatch");
1919 vec
<constructor_elt
, va_gc
> *v1
, *v2
;
1920 unsigned HOST_WIDE_INT idx
;
1922 enum tree_code typecode
= TREE_CODE (TREE_TYPE (t1
));
1923 if (typecode
!= TREE_CODE (TREE_TYPE (t2
)))
1924 return return_false_with_msg ("constructor type mismatch");
1926 if (typecode
== ARRAY_TYPE
)
1928 HOST_WIDE_INT size_1
= int_size_in_bytes (TREE_TYPE (t1
));
1929 /* For arrays, check that the sizes all match. */
1930 if (TYPE_MODE (TREE_TYPE (t1
)) != TYPE_MODE (TREE_TYPE (t2
))
1932 || size_1
!= int_size_in_bytes (TREE_TYPE (t2
)))
1933 return return_false_with_msg ("constructor array size mismatch");
1935 else if (!func_checker::compatible_types_p (TREE_TYPE (t1
),
1937 return return_false_with_msg ("constructor type incompatible");
1939 v1
= CONSTRUCTOR_ELTS (t1
);
1940 v2
= CONSTRUCTOR_ELTS (t2
);
1941 if (vec_safe_length (v1
) != vec_safe_length (v2
))
1942 return return_false_with_msg ("constructor number of elts mismatch");
1944 for (idx
= 0; idx
< vec_safe_length (v1
); ++idx
)
1946 constructor_elt
*c1
= &(*v1
)[idx
];
1947 constructor_elt
*c2
= &(*v2
)[idx
];
1949 /* Check that each value is the same... */
1950 if (!sem_variable::equals (c1
->value
, c2
->value
))
1952 /* ... and that they apply to the same fields! */
1953 if (!sem_variable::equals (c1
->index
, c2
->index
))
1960 tree x1
= TREE_OPERAND (t1
, 0);
1961 tree x2
= TREE_OPERAND (t2
, 0);
1962 tree y1
= TREE_OPERAND (t1
, 1);
1963 tree y2
= TREE_OPERAND (t2
, 1);
1965 if (!func_checker::compatible_types_p (TREE_TYPE (x1
), TREE_TYPE (x2
)))
1966 return return_false ();
1968 /* Type of the offset on MEM_REF does not matter. */
1969 return return_with_debug (sem_variable::equals (x1
, x2
)
1970 && wi::to_offset (y1
)
1971 == wi::to_offset (y2
));
1976 tree op1
= TREE_OPERAND (t1
, 0);
1977 tree op2
= TREE_OPERAND (t2
, 0);
1978 return sem_variable::equals (op1
, op2
);
1980 /* References to other vars/decls are compared using ipa-ref. */
1983 if (decl_in_symtab_p (t1
) && decl_in_symtab_p (t2
))
1985 return return_false_with_msg ("Declaration mismatch");
1987 /* TODO: We can check CONST_DECL by its DECL_INITIAL, but for that we
1988 need to process its VAR/FUNCTION references without relying on ipa-ref
1992 return return_false_with_msg ("Declaration mismatch");
1994 /* Integer constants are the same only if the same width of type. */
1995 if (TYPE_PRECISION (TREE_TYPE (t1
)) != TYPE_PRECISION (TREE_TYPE (t2
)))
1996 return return_false_with_msg ("INTEGER_CST precision mismatch");
1997 if (TYPE_MODE (TREE_TYPE (t1
)) != TYPE_MODE (TREE_TYPE (t2
)))
1998 return return_false_with_msg ("INTEGER_CST mode mismatch");
1999 return return_with_debug (tree_int_cst_equal (t1
, t2
));
2001 if (TYPE_MODE (TREE_TYPE (t1
)) != TYPE_MODE (TREE_TYPE (t2
)))
2002 return return_false_with_msg ("STRING_CST mode mismatch");
2003 if (TREE_STRING_LENGTH (t1
) != TREE_STRING_LENGTH (t2
))
2004 return return_false_with_msg ("STRING_CST length mismatch");
2005 if (memcmp (TREE_STRING_POINTER (t1
), TREE_STRING_POINTER (t2
),
2006 TREE_STRING_LENGTH (t1
)))
2007 return return_false_with_msg ("STRING_CST mismatch");
2010 /* Fixed constants are the same only if the same width of type. */
2011 if (TYPE_PRECISION (TREE_TYPE (t1
)) != TYPE_PRECISION (TREE_TYPE (t2
)))
2012 return return_false_with_msg ("FIXED_CST precision mismatch");
2014 return return_with_debug (FIXED_VALUES_IDENTICAL (TREE_FIXED_CST (t1
),
2015 TREE_FIXED_CST (t2
)));
2017 return (sem_variable::equals (TREE_REALPART (t1
), TREE_REALPART (t2
))
2018 && sem_variable::equals (TREE_IMAGPART (t1
), TREE_IMAGPART (t2
)));
2020 /* Real constants are the same only if the same width of type. */
2021 if (TYPE_PRECISION (TREE_TYPE (t1
)) != TYPE_PRECISION (TREE_TYPE (t2
)))
2022 return return_false_with_msg ("REAL_CST precision mismatch");
2023 return return_with_debug (real_identical (&TREE_REAL_CST (t1
),
2024 &TREE_REAL_CST (t2
)));
2029 if (VECTOR_CST_NELTS (t1
) != VECTOR_CST_NELTS (t2
))
2030 return return_false_with_msg ("VECTOR_CST nelts mismatch");
2032 for (i
= 0; i
< VECTOR_CST_NELTS (t1
); ++i
)
2033 if (!sem_variable::equals (VECTOR_CST_ELT (t1
, i
),
2034 VECTOR_CST_ELT (t2
, i
)))
2040 case ARRAY_RANGE_REF
:
2042 tree x1
= TREE_OPERAND (t1
, 0);
2043 tree x2
= TREE_OPERAND (t2
, 0);
2044 tree y1
= TREE_OPERAND (t1
, 1);
2045 tree y2
= TREE_OPERAND (t2
, 1);
2047 if (!sem_variable::equals (x1
, x2
) || !sem_variable::equals (y1
, y2
))
2049 if (!sem_variable::equals (array_ref_low_bound (t1
),
2050 array_ref_low_bound (t2
)))
2052 if (!sem_variable::equals (array_ref_element_size (t1
),
2053 array_ref_element_size (t2
)))
2059 case POINTER_PLUS_EXPR
:
2064 tree x1
= TREE_OPERAND (t1
, 0);
2065 tree x2
= TREE_OPERAND (t2
, 0);
2066 tree y1
= TREE_OPERAND (t1
, 1);
2067 tree y2
= TREE_OPERAND (t2
, 1);
2069 return sem_variable::equals (x1
, x2
) && sem_variable::equals (y1
, y2
);
2073 case VIEW_CONVERT_EXPR
:
2074 if (!func_checker::compatible_types_p (TREE_TYPE (t1
), TREE_TYPE (t2
)))
2075 return return_false ();
2076 return sem_variable::equals (TREE_OPERAND (t1
, 0), TREE_OPERAND (t2
, 0));
2078 return return_false_with_msg ("ERROR_MARK");
2080 return return_false_with_msg ("Unknown TREE code reached");
2084 /* Parser function that visits a varpool NODE. */
2087 sem_variable::parse (varpool_node
*node
, bitmap_obstack
*stack
)
2089 if (TREE_THIS_VOLATILE (node
->decl
) || DECL_HARD_REGISTER (node
->decl
)
2093 sem_variable
*v
= new sem_variable (node
, stack
);
2100 /* References independent hash function. */
2103 sem_variable::get_hash (void)
2108 /* All WPA streamed in symbols should have their hashes computed at compile
2109 time. At this point, the constructor may not be in memory at all.
2110 DECL_INITIAL (decl) would be error_mark_node in that case. */
2111 gcc_assert (!node
->lto_file_data
);
2112 tree ctor
= DECL_INITIAL (decl
);
2113 inchash::hash hstate
;
2115 hstate
.add_int (456346417);
2116 if (DECL_SIZE (decl
) && tree_fits_shwi_p (DECL_SIZE (decl
)))
2117 hstate
.add_wide_int (tree_to_shwi (DECL_SIZE (decl
)));
2118 add_expr (ctor
, hstate
);
2119 set_hash (hstate
.end ());
2124 /* Set all points-to UIDs of aliases pointing to node N as UID. */
2127 set_alias_uids (symtab_node
*n
, int uid
)
2130 FOR_EACH_ALIAS (n
, ref
)
2133 fprintf (dump_file
, " Setting points-to UID of [%s] as %d\n",
2134 xstrdup_for_dump (ref
->referring
->asm_name ()), uid
);
2136 SET_DECL_PT_UID (ref
->referring
->decl
, uid
);
2137 set_alias_uids (ref
->referring
, uid
);
2141 /* Merges instance with an ALIAS_ITEM, where alias, thunk or redirection can
2145 sem_variable::merge (sem_item
*alias_item
)
2147 gcc_assert (alias_item
->type
== VAR
);
2149 if (!sem_item::target_supports_symbol_aliases_p ())
2152 fprintf (dump_file
, "Not unifying; "
2153 "Symbol aliases are not supported by target\n\n");
2157 if (DECL_EXTERNAL (alias_item
->decl
))
2160 fprintf (dump_file
, "Not unifying; alias is external.\n\n");
2164 sem_variable
*alias_var
= static_cast<sem_variable
*> (alias_item
);
2166 varpool_node
*original
= get_node ();
2167 varpool_node
*alias
= alias_var
->get_node ();
2168 bool original_discardable
= false;
2170 bool alias_address_matters
= alias
->address_matters_p ();
2172 /* See if original is in a section that can be discarded if the main
2174 Also consider case where we have resolution info and we know that
2175 original's definition is not going to be used. In this case we can not
2176 create alias to original. */
2177 if (original
->can_be_discarded_p ()
2178 || (node
->resolution
!= LDPR_UNKNOWN
2179 && !decl_binds_to_current_def_p (node
->decl
)))
2180 original_discardable
= true;
2182 gcc_assert (!TREE_ASM_WRITTEN (alias
->decl
));
2184 /* Constant pool machinery is not quite ready for aliases.
2185 TODO: varasm code contains logic for merging DECL_IN_CONSTANT_POOL.
2186 For LTO merging does not happen that is an important missing feature.
2187 We can enable merging with LTO if the DECL_IN_CONSTANT_POOL
2188 flag is dropped and non-local symbol name is assigned. */
2189 if (DECL_IN_CONSTANT_POOL (alias
->decl
)
2190 || DECL_IN_CONSTANT_POOL (original
->decl
))
2194 "Not unifying; constant pool variables.\n\n");
2198 /* Do not attempt to mix functions from different user sections;
2199 we do not know what user intends with those. */
2200 if (((DECL_SECTION_NAME (original
->decl
) && !original
->implicit_section
)
2201 || (DECL_SECTION_NAME (alias
->decl
) && !alias
->implicit_section
))
2202 && DECL_SECTION_NAME (original
->decl
) != DECL_SECTION_NAME (alias
->decl
))
2207 "original and alias are in different sections.\n\n");
2211 /* We can not merge if address comparsion metters. */
2212 if (alias_address_matters
&& flag_merge_constants
< 2)
2216 "Not unifying; address of original may be compared.\n\n");
2220 if (DECL_ALIGN (original
->decl
) < DECL_ALIGN (alias
->decl
))
2223 fprintf (dump_file
, "Not unifying; "
2224 "original and alias have incompatible alignments\n\n");
2229 if (DECL_COMDAT_GROUP (original
->decl
) != DECL_COMDAT_GROUP (alias
->decl
))
2232 fprintf (dump_file
, "Not unifying; alias cannot be created; "
2233 "across comdat group boundary\n\n");
2238 if (original_discardable
)
2241 fprintf (dump_file
, "Not unifying; alias cannot be created; "
2242 "target is discardable\n\n");
2248 gcc_assert (!original
->alias
);
2249 gcc_assert (!alias
->alias
);
2251 alias
->analyzed
= false;
2253 DECL_INITIAL (alias
->decl
) = NULL
;
2254 ((symtab_node
*)alias
)->call_for_symbol_and_aliases (clear_decl_rtl
,
2256 alias
->need_bounds_init
= false;
2257 alias
->remove_all_references ();
2258 if (TREE_ADDRESSABLE (alias
->decl
))
2259 original
->call_for_symbol_and_aliases (set_addressable
, NULL
, true);
2261 varpool_node::create_alias (alias_var
->decl
, decl
);
2262 alias
->resolve_alias (original
);
2265 fprintf (dump_file
, "Unified; Variable alias has been created.\n");
2267 set_alias_uids (original
, DECL_UID (original
->decl
));
2272 /* Dump symbol to FILE. */
2275 sem_variable::dump_to_file (FILE *file
)
2279 print_node (file
, "", decl
, 0);
2280 fprintf (file
, "\n\n");
2283 unsigned int sem_item_optimizer::class_id
= 0;
2285 sem_item_optimizer::sem_item_optimizer ()
2286 : worklist (0), m_classes (0), m_classes_count (0), m_cgraph_node_hooks (NULL
),
2287 m_varpool_node_hooks (NULL
)
2290 bitmap_obstack_initialize (&m_bmstack
);
2293 sem_item_optimizer::~sem_item_optimizer ()
2295 for (unsigned int i
= 0; i
< m_items
.length (); i
++)
2299 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
2300 it
!= m_classes
.end (); ++it
)
2302 for (unsigned int i
= 0; i
< (*it
)->classes
.length (); i
++)
2303 delete (*it
)->classes
[i
];
2305 (*it
)->classes
.release ();
2311 bitmap_obstack_release (&m_bmstack
);
2314 /* Write IPA ICF summary for symbols. */
2317 sem_item_optimizer::write_summary (void)
2319 unsigned int count
= 0;
2321 output_block
*ob
= create_output_block (LTO_section_ipa_icf
);
2322 lto_symtab_encoder_t encoder
= ob
->decl_state
->symtab_node_encoder
;
2325 /* Calculate number of symbols to be serialized. */
2326 for (lto_symtab_encoder_iterator lsei
= lsei_start_in_partition (encoder
);
2328 lsei_next_in_partition (&lsei
))
2330 symtab_node
*node
= lsei_node (lsei
);
2332 if (m_symtab_node_map
.get (node
))
2336 streamer_write_uhwi (ob
, count
);
2338 /* Process all of the symbols. */
2339 for (lto_symtab_encoder_iterator lsei
= lsei_start_in_partition (encoder
);
2341 lsei_next_in_partition (&lsei
))
2343 symtab_node
*node
= lsei_node (lsei
);
2345 sem_item
**item
= m_symtab_node_map
.get (node
);
2349 int node_ref
= lto_symtab_encoder_encode (encoder
, node
);
2350 streamer_write_uhwi_stream (ob
->main_stream
, node_ref
);
2352 streamer_write_uhwi (ob
, (*item
)->get_hash ());
2356 streamer_write_char_stream (ob
->main_stream
, 0);
2357 produce_asm (ob
, NULL
);
2358 destroy_output_block (ob
);
2361 /* Reads a section from LTO stream file FILE_DATA. Input block for DATA
2362 contains LEN bytes. */
2365 sem_item_optimizer::read_section (lto_file_decl_data
*file_data
,
2366 const char *data
, size_t len
)
2368 const lto_function_header
*header
2369 = (const lto_function_header
*) data
;
2370 const int cfg_offset
= sizeof (lto_function_header
);
2371 const int main_offset
= cfg_offset
+ header
->cfg_size
;
2372 const int string_offset
= main_offset
+ header
->main_size
;
2377 lto_input_block
ib_main ((const char *) data
+ main_offset
, 0,
2378 header
->main_size
, file_data
->mode_table
);
2381 = lto_data_in_create (file_data
, (const char *) data
+ string_offset
,
2382 header
->string_size
, vNULL
);
2384 count
= streamer_read_uhwi (&ib_main
);
2386 for (i
= 0; i
< count
; i
++)
2390 lto_symtab_encoder_t encoder
;
2392 index
= streamer_read_uhwi (&ib_main
);
2393 encoder
= file_data
->symtab_node_encoder
;
2394 node
= lto_symtab_encoder_deref (encoder
, index
);
2396 hashval_t hash
= streamer_read_uhwi (&ib_main
);
2398 gcc_assert (node
->definition
);
2401 fprintf (dump_file
, "Symbol added:%s (tree: %p, uid:%u)\n",
2402 node
->asm_name (), (void *) node
->decl
, node
->order
);
2404 if (is_a
<cgraph_node
*> (node
))
2406 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (node
);
2408 sem_function
*fn
= new sem_function (cnode
, &m_bmstack
);
2409 fn
->set_hash (hash
);
2410 m_items
.safe_push (fn
);
2414 varpool_node
*vnode
= dyn_cast
<varpool_node
*> (node
);
2416 sem_variable
*var
= new sem_variable (vnode
, &m_bmstack
);
2417 var
->set_hash (hash
);
2418 m_items
.safe_push (var
);
2422 lto_free_section_data (file_data
, LTO_section_ipa_icf
, NULL
, data
,
2424 lto_data_in_delete (data_in
);
2427 /* Read IPA ICF summary for symbols. */
2430 sem_item_optimizer::read_summary (void)
2432 lto_file_decl_data
**file_data_vec
= lto_get_file_decl_data ();
2433 lto_file_decl_data
*file_data
;
2436 while ((file_data
= file_data_vec
[j
++]))
2439 const char *data
= lto_get_section_data (file_data
,
2440 LTO_section_ipa_icf
, NULL
, &len
);
2443 read_section (file_data
, data
, len
);
2447 /* Register callgraph and varpool hooks. */
2450 sem_item_optimizer::register_hooks (void)
2452 if (!m_cgraph_node_hooks
)
2453 m_cgraph_node_hooks
= symtab
->add_cgraph_removal_hook
2454 (&sem_item_optimizer::cgraph_removal_hook
, this);
2456 if (!m_varpool_node_hooks
)
2457 m_varpool_node_hooks
= symtab
->add_varpool_removal_hook
2458 (&sem_item_optimizer::varpool_removal_hook
, this);
2461 /* Unregister callgraph and varpool hooks. */
2464 sem_item_optimizer::unregister_hooks (void)
2466 if (m_cgraph_node_hooks
)
2467 symtab
->remove_cgraph_removal_hook (m_cgraph_node_hooks
);
2469 if (m_varpool_node_hooks
)
2470 symtab
->remove_varpool_removal_hook (m_varpool_node_hooks
);
2473 /* Adds a CLS to hashtable associated by hash value. */
2476 sem_item_optimizer::add_class (congruence_class
*cls
)
2478 gcc_assert (cls
->members
.length ());
2480 congruence_class_group
*group
2481 = get_group_by_hash (cls
->members
[0]->get_hash (),
2482 cls
->members
[0]->type
);
2483 group
->classes
.safe_push (cls
);
2486 /* Gets a congruence class group based on given HASH value and TYPE. */
2488 congruence_class_group
*
2489 sem_item_optimizer::get_group_by_hash (hashval_t hash
, sem_item_type type
)
2491 congruence_class_group
*item
= XNEW (congruence_class_group
);
2495 congruence_class_group
**slot
= m_classes
.find_slot (item
, INSERT
);
2501 item
->classes
.create (1);
2508 /* Callgraph removal hook called for a NODE with a custom DATA. */
2511 sem_item_optimizer::cgraph_removal_hook (cgraph_node
*node
, void *data
)
2513 sem_item_optimizer
*optimizer
= (sem_item_optimizer
*) data
;
2514 optimizer
->remove_symtab_node (node
);
2517 /* Varpool removal hook called for a NODE with a custom DATA. */
2520 sem_item_optimizer::varpool_removal_hook (varpool_node
*node
, void *data
)
2522 sem_item_optimizer
*optimizer
= (sem_item_optimizer
*) data
;
2523 optimizer
->remove_symtab_node (node
);
2526 /* Remove symtab NODE triggered by symtab removal hooks. */
2529 sem_item_optimizer::remove_symtab_node (symtab_node
*node
)
2531 gcc_assert (!m_classes
.elements ());
2533 m_removed_items_set
.add (node
);
2537 sem_item_optimizer::remove_item (sem_item
*item
)
2539 if (m_symtab_node_map
.get (item
->node
))
2540 m_symtab_node_map
.remove (item
->node
);
2544 /* Removes all callgraph and varpool nodes that are marked by symtab
2548 sem_item_optimizer::filter_removed_items (void)
2550 auto_vec
<sem_item
*> filtered
;
2552 for (unsigned int i
= 0; i
< m_items
.length(); i
++)
2554 sem_item
*item
= m_items
[i
];
2556 if (m_removed_items_set
.contains (item
->node
))
2562 if (item
->type
== FUNC
)
2564 cgraph_node
*cnode
= static_cast <sem_function
*>(item
)->get_node ();
2566 if (in_lto_p
&& (cnode
->alias
|| cnode
->body_removed
))
2569 filtered
.safe_push (item
);
2573 if (!flag_ipa_icf_variables
)
2577 /* Filter out non-readonly variables. */
2578 tree decl
= item
->decl
;
2579 if (TREE_READONLY (decl
))
2580 filtered
.safe_push (item
);
2587 /* Clean-up of released semantic items. */
2590 for (unsigned int i
= 0; i
< filtered
.length(); i
++)
2591 m_items
.safe_push (filtered
[i
]);
2594 /* Optimizer entry point which returns true in case it processes
2595 a merge operation. True is returned if there's a merge operation
2599 sem_item_optimizer::execute (void)
2601 filter_removed_items ();
2602 unregister_hooks ();
2605 update_hash_by_addr_refs ();
2606 build_hash_based_classes ();
2609 fprintf (dump_file
, "Dump after hash based groups\n");
2610 dump_cong_classes ();
2612 for (unsigned int i
= 0; i
< m_items
.length(); i
++)
2613 m_items
[i
]->init_wpa ();
2615 subdivide_classes_by_equality (true);
2618 fprintf (dump_file
, "Dump after WPA based types groups\n");
2620 dump_cong_classes ();
2622 process_cong_reduction ();
2623 checking_verify_classes ();
2626 fprintf (dump_file
, "Dump after callgraph-based congruence reduction\n");
2628 dump_cong_classes ();
2630 parse_nonsingleton_classes ();
2631 subdivide_classes_by_equality ();
2634 fprintf (dump_file
, "Dump after full equality comparison of groups\n");
2636 dump_cong_classes ();
2638 unsigned int prev_class_count
= m_classes_count
;
2640 process_cong_reduction ();
2641 dump_cong_classes ();
2642 checking_verify_classes ();
2643 bool merged_p
= merge_classes (prev_class_count
);
2645 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2646 symtab_node::dump_table (dump_file
);
2651 /* Function responsible for visiting all potential functions and
2652 read-only variables that can be merged. */
2655 sem_item_optimizer::parse_funcs_and_vars (void)
2659 if (flag_ipa_icf_functions
)
2660 FOR_EACH_DEFINED_FUNCTION (cnode
)
2662 sem_function
*f
= sem_function::parse (cnode
, &m_bmstack
);
2665 m_items
.safe_push (f
);
2666 m_symtab_node_map
.put (cnode
, f
);
2669 fprintf (dump_file
, "Parsed function:%s\n", f
->node
->asm_name ());
2671 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2672 f
->dump_to_file (dump_file
);
2675 fprintf (dump_file
, "Not parsed function:%s\n", cnode
->asm_name ());
2678 varpool_node
*vnode
;
2680 if (flag_ipa_icf_variables
)
2681 FOR_EACH_DEFINED_VARIABLE (vnode
)
2683 sem_variable
*v
= sem_variable::parse (vnode
, &m_bmstack
);
2687 m_items
.safe_push (v
);
2688 m_symtab_node_map
.put (vnode
, v
);
2693 /* Makes pairing between a congruence class CLS and semantic ITEM. */
2696 sem_item_optimizer::add_item_to_class (congruence_class
*cls
, sem_item
*item
)
2698 item
->index_in_class
= cls
->members
.length ();
2699 cls
->members
.safe_push (item
);
2703 /* For each semantic item, append hash values of references. */
2706 sem_item_optimizer::update_hash_by_addr_refs ()
2708 /* First, append to hash sensitive references and class type if it need to
2709 be matched for ODR. */
2710 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2712 m_items
[i
]->update_hash_by_addr_refs (m_symtab_node_map
);
2713 if (m_items
[i
]->type
== FUNC
)
2715 if (TREE_CODE (TREE_TYPE (m_items
[i
]->decl
)) == METHOD_TYPE
2716 && contains_polymorphic_type_p
2717 (TYPE_METHOD_BASETYPE (TREE_TYPE (m_items
[i
]->decl
)))
2718 && (DECL_CXX_CONSTRUCTOR_P (m_items
[i
]->decl
)
2719 || (static_cast<sem_function
*> (m_items
[i
])->param_used_p (0)
2720 && static_cast<sem_function
*> (m_items
[i
])
2721 ->compare_polymorphic_p ())))
2724 = TYPE_METHOD_BASETYPE (TREE_TYPE (m_items
[i
]->decl
));
2725 inchash::hash
hstate (m_items
[i
]->get_hash ());
2727 if (TYPE_NAME (class_type
)
2728 && DECL_ASSEMBLER_NAME_SET_P (TYPE_NAME (class_type
)))
2730 (IDENTIFIER_HASH_VALUE
2731 (DECL_ASSEMBLER_NAME (TYPE_NAME (class_type
))));
2733 m_items
[i
]->set_hash (hstate
.end ());
2738 /* Once all symbols have enhanced hash value, we can append
2739 hash values of symbols that are seen by IPA ICF and are
2740 references by a semantic item. Newly computed values
2741 are saved to global_hash member variable. */
2742 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2743 m_items
[i
]->update_hash_by_local_refs (m_symtab_node_map
);
2745 /* Global hash value replace current hash values. */
2746 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2747 m_items
[i
]->set_hash (m_items
[i
]->global_hash
);
2750 /* Congruence classes are built by hash value. */
2753 sem_item_optimizer::build_hash_based_classes (void)
2755 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2757 sem_item
*item
= m_items
[i
];
2759 congruence_class_group
*group
2760 = get_group_by_hash (item
->get_hash (), item
->type
);
2762 if (!group
->classes
.length ())
2765 group
->classes
.safe_push (new congruence_class (class_id
++));
2768 add_item_to_class (group
->classes
[0], item
);
2772 /* Build references according to call graph. */
2775 sem_item_optimizer::build_graph (void)
2777 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2779 sem_item
*item
= m_items
[i
];
2780 m_symtab_node_map
.put (item
->node
, item
);
2782 /* Initialize hash values if we are not in LTO mode. */
2787 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2789 sem_item
*item
= m_items
[i
];
2791 if (item
->type
== FUNC
)
2793 cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (item
->node
);
2795 cgraph_edge
*e
= cnode
->callees
;
2798 sem_item
**slot
= m_symtab_node_map
.get
2799 (e
->callee
->ultimate_alias_target ());
2801 item
->add_reference (*slot
);
2807 ipa_ref
*ref
= NULL
;
2808 for (unsigned i
= 0; item
->node
->iterate_reference (i
, ref
); i
++)
2810 sem_item
**slot
= m_symtab_node_map
.get
2811 (ref
->referred
->ultimate_alias_target ());
2813 item
->add_reference (*slot
);
2818 /* Semantic items in classes having more than one element and initialized.
2819 In case of WPA, we load function body. */
2822 sem_item_optimizer::parse_nonsingleton_classes (void)
2824 unsigned int init_called_count
= 0;
2826 for (unsigned i
= 0; i
< m_items
.length (); i
++)
2827 if (m_items
[i
]->cls
->members
.length () > 1)
2829 m_items
[i
]->init ();
2830 init_called_count
++;
2834 fprintf (dump_file
, "Init called for %u items (%.2f%%).\n",
2836 m_items
.length () ? 100.0f
* init_called_count
/ m_items
.length ()
2840 /* Equality function for semantic items is used to subdivide existing
2841 classes. If IN_WPA, fast equality function is invoked. */
2844 sem_item_optimizer::subdivide_classes_by_equality (bool in_wpa
)
2846 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
2847 it
!= m_classes
.end (); ++it
)
2849 unsigned int class_count
= (*it
)->classes
.length ();
2851 for (unsigned i
= 0; i
< class_count
; i
++)
2853 congruence_class
*c
= (*it
)->classes
[i
];
2855 if (c
->members
.length() > 1)
2857 auto_vec
<sem_item
*> new_vector
;
2859 sem_item
*first
= c
->members
[0];
2860 new_vector
.safe_push (first
);
2862 unsigned class_split_first
= (*it
)->classes
.length ();
2864 for (unsigned j
= 1; j
< c
->members
.length (); j
++)
2866 sem_item
*item
= c
->members
[j
];
2869 = in_wpa
? first
->equals_wpa (item
, m_symtab_node_map
)
2870 : first
->equals (item
, m_symtab_node_map
);
2873 new_vector
.safe_push (item
);
2876 bool integrated
= false;
2878 for (unsigned k
= class_split_first
;
2879 k
< (*it
)->classes
.length (); k
++)
2881 sem_item
*x
= (*it
)->classes
[k
]->members
[0];
2883 = in_wpa
? x
->equals_wpa (item
, m_symtab_node_map
)
2884 : x
->equals (item
, m_symtab_node_map
);
2889 add_item_to_class ((*it
)->classes
[k
], item
);
2898 = new congruence_class (class_id
++);
2900 add_item_to_class (c
, item
);
2902 (*it
)->classes
.safe_push (c
);
2907 // We replace newly created new_vector for the class we've just
2909 c
->members
.release ();
2910 c
->members
.create (new_vector
.length ());
2912 for (unsigned int j
= 0; j
< new_vector
.length (); j
++)
2913 add_item_to_class (c
, new_vector
[j
]);
2918 checking_verify_classes ();
2921 /* Subdivide classes by address references that members of the class
2922 reference. Example can be a pair of functions that have an address
2923 taken from a function. If these addresses are different the class
2927 sem_item_optimizer::subdivide_classes_by_sensitive_refs ()
2929 typedef hash_map
<symbol_compare_hash
, vec
<sem_item
*> > subdivide_hash_map
;
2931 unsigned newly_created_classes
= 0;
2933 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
2934 it
!= m_classes
.end (); ++it
)
2936 unsigned int class_count
= (*it
)->classes
.length ();
2937 auto_vec
<congruence_class
*> new_classes
;
2939 for (unsigned i
= 0; i
< class_count
; i
++)
2941 congruence_class
*c
= (*it
)->classes
[i
];
2943 if (c
->members
.length() > 1)
2945 subdivide_hash_map split_map
;
2947 for (unsigned j
= 0; j
< c
->members
.length (); j
++)
2949 sem_item
*source_node
= c
->members
[j
];
2951 symbol_compare_collection
*collection
2952 = new symbol_compare_collection (source_node
->node
);
2955 vec
<sem_item
*> *slot
2956 = &split_map
.get_or_insert (collection
, &existed
);
2957 gcc_checking_assert (slot
);
2959 slot
->safe_push (source_node
);
2965 /* If the map contains more than one key, we have to split
2966 the map appropriately. */
2967 if (split_map
.elements () != 1)
2969 bool first_class
= true;
2971 for (subdivide_hash_map::iterator it2
= split_map
.begin ();
2972 it2
!= split_map
.end (); ++it2
)
2974 congruence_class
*new_cls
;
2975 new_cls
= new congruence_class (class_id
++);
2977 for (unsigned k
= 0; k
< (*it2
).second
.length (); k
++)
2978 add_item_to_class (new_cls
, (*it2
).second
[k
]);
2980 worklist_push (new_cls
);
2981 newly_created_classes
++;
2985 (*it
)->classes
[i
] = new_cls
;
2986 first_class
= false;
2990 new_classes
.safe_push (new_cls
);
2996 /* Release memory. */
2997 for (subdivide_hash_map::iterator it2
= split_map
.begin ();
2998 it2
!= split_map
.end (); ++it2
)
3000 delete (*it2
).first
;
3001 (*it2
).second
.release ();
3006 for (unsigned i
= 0; i
< new_classes
.length (); i
++)
3007 (*it
)->classes
.safe_push (new_classes
[i
]);
3010 return newly_created_classes
;
3013 /* Verify congruence classes, if checking is enabled. */
3016 sem_item_optimizer::checking_verify_classes (void)
3022 /* Verify congruence classes. */
3025 sem_item_optimizer::verify_classes (void)
3027 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3028 it
!= m_classes
.end (); ++it
)
3030 for (unsigned int i
= 0; i
< (*it
)->classes
.length (); i
++)
3032 congruence_class
*cls
= (*it
)->classes
[i
];
3035 gcc_assert (cls
->members
.length () > 0);
3037 for (unsigned int j
= 0; j
< cls
->members
.length (); j
++)
3039 sem_item
*item
= cls
->members
[j
];
3042 gcc_assert (item
->cls
== cls
);
3044 for (unsigned k
= 0; k
< item
->usages
.length (); k
++)
3046 sem_usage_pair
*usage
= item
->usages
[k
];
3047 gcc_assert (usage
->item
->index_in_class
3048 < usage
->item
->cls
->members
.length ());
3055 /* Disposes split map traverse function. CLS_PTR is pointer to congruence
3056 class, BSLOT is bitmap slot we want to release. DATA is mandatory,
3057 but unused argument. */
3060 sem_item_optimizer::release_split_map (congruence_class
* const &,
3061 bitmap
const &b
, traverse_split_pair
*)
3070 /* Process split operation for a class given as pointer CLS_PTR,
3071 where bitmap B splits congruence class members. DATA is used
3072 as argument of split pair. */
3075 sem_item_optimizer::traverse_congruence_split (congruence_class
* const &cls
,
3077 traverse_split_pair
*pair
)
3079 sem_item_optimizer
*optimizer
= pair
->optimizer
;
3080 const congruence_class
*splitter_cls
= pair
->cls
;
3082 /* If counted bits are greater than zero and less than the number of members
3083 a group will be splitted. */
3084 unsigned popcount
= bitmap_count_bits (b
);
3086 if (popcount
> 0 && popcount
< cls
->members
.length ())
3088 auto_vec
<congruence_class
*, 2> newclasses
;
3089 newclasses
.quick_push (new congruence_class (class_id
++));
3090 newclasses
.quick_push (new congruence_class (class_id
++));
3092 for (unsigned int i
= 0; i
< cls
->members
.length (); i
++)
3094 int target
= bitmap_bit_p (b
, i
);
3095 congruence_class
*tc
= newclasses
[target
];
3097 add_item_to_class (tc
, cls
->members
[i
]);
3102 for (unsigned int i
= 0; i
< 2; i
++)
3103 gcc_assert (newclasses
[i
]->members
.length ());
3106 if (splitter_cls
== cls
)
3107 optimizer
->splitter_class_removed
= true;
3109 /* Remove old class from worklist if presented. */
3110 bool in_worklist
= cls
->in_worklist
;
3113 cls
->in_worklist
= false;
3115 congruence_class_group g
;
3116 g
.hash
= cls
->members
[0]->get_hash ();
3117 g
.type
= cls
->members
[0]->type
;
3119 congruence_class_group
*slot
= optimizer
->m_classes
.find (&g
);
3121 for (unsigned int i
= 0; i
< slot
->classes
.length (); i
++)
3122 if (slot
->classes
[i
] == cls
)
3124 slot
->classes
.ordered_remove (i
);
3128 /* New class will be inserted and integrated to work list. */
3129 for (unsigned int i
= 0; i
< 2; i
++)
3130 optimizer
->add_class (newclasses
[i
]);
3132 /* Two classes replace one, so that increment just by one. */
3133 optimizer
->m_classes_count
++;
3135 /* If OLD class was presented in the worklist, we remove the class
3136 and replace it will both newly created classes. */
3138 for (unsigned int i
= 0; i
< 2; i
++)
3139 optimizer
->worklist_push (newclasses
[i
]);
3140 else /* Just smaller class is inserted. */
3142 unsigned int smaller_index
3143 = (newclasses
[0]->members
.length ()
3144 < newclasses
[1]->members
.length ()
3146 optimizer
->worklist_push (newclasses
[smaller_index
]);
3149 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3151 fprintf (dump_file
, " congruence class splitted:\n");
3152 cls
->dump (dump_file
, 4);
3154 fprintf (dump_file
, " newly created groups:\n");
3155 for (unsigned int i
= 0; i
< 2; i
++)
3156 newclasses
[i
]->dump (dump_file
, 4);
3159 /* Release class if not presented in work list. */
3168 /* Tests if a class CLS used as INDEXth splits any congruence classes.
3169 Bitmap stack BMSTACK is used for bitmap allocation. */
3172 sem_item_optimizer::do_congruence_step_for_index (congruence_class
*cls
,
3175 hash_map
<congruence_class
*, bitmap
> split_map
;
3177 for (unsigned int i
= 0; i
< cls
->members
.length (); i
++)
3179 sem_item
*item
= cls
->members
[i
];
3181 /* Iterate all usages that have INDEX as usage of the item. */
3182 for (unsigned int j
= 0; j
< item
->usages
.length (); j
++)
3184 sem_usage_pair
*usage
= item
->usages
[j
];
3186 if (usage
->index
!= index
)
3189 bitmap
*slot
= split_map
.get (usage
->item
->cls
);
3194 b
= BITMAP_ALLOC (&m_bmstack
);
3195 split_map
.put (usage
->item
->cls
, b
);
3200 gcc_checking_assert (usage
->item
->cls
);
3201 gcc_checking_assert (usage
->item
->index_in_class
3202 < usage
->item
->cls
->members
.length ());
3204 bitmap_set_bit (b
, usage
->item
->index_in_class
);
3208 traverse_split_pair pair
;
3209 pair
.optimizer
= this;
3212 splitter_class_removed
= false;
3213 split_map
.traverse
<traverse_split_pair
*,
3214 sem_item_optimizer::traverse_congruence_split
> (&pair
);
3216 /* Bitmap clean-up. */
3217 split_map
.traverse
<traverse_split_pair
*,
3218 sem_item_optimizer::release_split_map
> (NULL
);
3221 /* Every usage of a congruence class CLS is a candidate that can split the
3222 collection of classes. Bitmap stack BMSTACK is used for bitmap
3226 sem_item_optimizer::do_congruence_step (congruence_class
*cls
)
3231 bitmap usage
= BITMAP_ALLOC (&m_bmstack
);
3233 for (unsigned int i
= 0; i
< cls
->members
.length (); i
++)
3234 bitmap_ior_into (usage
, cls
->members
[i
]->usage_index_bitmap
);
3236 EXECUTE_IF_SET_IN_BITMAP (usage
, 0, i
, bi
)
3238 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3239 fprintf (dump_file
, " processing congruence step for class: %u, "
3240 "index: %u\n", cls
->id
, i
);
3242 do_congruence_step_for_index (cls
, i
);
3244 if (splitter_class_removed
)
3248 BITMAP_FREE (usage
);
3251 /* Adds a newly created congruence class CLS to worklist. */
3254 sem_item_optimizer::worklist_push (congruence_class
*cls
)
3256 /* Return if the class CLS is already presented in work list. */
3257 if (cls
->in_worklist
)
3260 cls
->in_worklist
= true;
3261 worklist
.push_back (cls
);
3264 /* Pops a class from worklist. */
3267 sem_item_optimizer::worklist_pop (void)
3269 congruence_class
*cls
;
3271 while (!worklist
.empty ())
3273 cls
= worklist
.front ();
3274 worklist
.pop_front ();
3275 if (cls
->in_worklist
)
3277 cls
->in_worklist
= false;
3283 /* Work list item was already intended to be removed.
3284 The only reason for doing it is to split a class.
3285 Thus, the class CLS is deleted. */
3293 /* Iterative congruence reduction function. */
3296 sem_item_optimizer::process_cong_reduction (void)
3298 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3299 it
!= m_classes
.end (); ++it
)
3300 for (unsigned i
= 0; i
< (*it
)->classes
.length (); i
++)
3301 if ((*it
)->classes
[i
]->is_class_used ())
3302 worklist_push ((*it
)->classes
[i
]);
3305 fprintf (dump_file
, "Worklist has been filled with: %lu\n",
3306 (unsigned long) worklist
.size ());
3308 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3309 fprintf (dump_file
, "Congruence class reduction\n");
3311 congruence_class
*cls
;
3313 /* Process complete congruence reduction. */
3314 while ((cls
= worklist_pop ()) != NULL
)
3315 do_congruence_step (cls
);
3317 /* Subdivide newly created classes according to references. */
3318 unsigned new_classes
= subdivide_classes_by_sensitive_refs ();
3321 fprintf (dump_file
, "Address reference subdivision created: %u "
3322 "new classes.\n", new_classes
);
3325 /* Debug function prints all informations about congruence classes. */
3328 sem_item_optimizer::dump_cong_classes (void)
3334 "Congruence classes: %u (unique hash values: %lu), with total: "
3335 "%u items\n", m_classes_count
,
3336 (unsigned long) m_classes
.elements (), m_items
.length ());
3338 /* Histogram calculation. */
3339 unsigned int max_index
= 0;
3340 unsigned int* histogram
= XCNEWVEC (unsigned int, m_items
.length () + 1);
3342 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3343 it
!= m_classes
.end (); ++it
)
3344 for (unsigned i
= 0; i
< (*it
)->classes
.length (); i
++)
3346 unsigned int c
= (*it
)->classes
[i
]->members
.length ();
3354 "Class size histogram [num of members]: number of classe number "
3357 for (unsigned int i
= 0; i
<= max_index
; i
++)
3359 fprintf (dump_file
, "[%u]: %u classes\n", i
, histogram
[i
]);
3361 fprintf (dump_file
, "\n\n");
3363 if (dump_flags
& TDF_DETAILS
)
3364 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3365 it
!= m_classes
.end (); ++it
)
3367 fprintf (dump_file
, " group: with %u classes:\n",
3368 (*it
)->classes
.length ());
3370 for (unsigned i
= 0; i
< (*it
)->classes
.length (); i
++)
3372 (*it
)->classes
[i
]->dump (dump_file
, 4);
3374 if (i
< (*it
)->classes
.length () - 1)
3375 fprintf (dump_file
, " ");
3382 /* Sort pair of sem_items A and B by DECL_UID. */
3385 sort_sem_items_by_decl_uid (const void *a
, const void *b
)
3387 const sem_item
*i1
= *(const sem_item
* const *)a
;
3388 const sem_item
*i2
= *(const sem_item
* const *)b
;
3390 int uid1
= DECL_UID (i1
->decl
);
3391 int uid2
= DECL_UID (i2
->decl
);
3395 else if (uid1
> uid2
)
3401 /* Sort pair of congruence_classes A and B by DECL_UID of the first member. */
3404 sort_congruence_classes_by_decl_uid (const void *a
, const void *b
)
3406 const congruence_class
*c1
= *(const congruence_class
* const *)a
;
3407 const congruence_class
*c2
= *(const congruence_class
* const *)b
;
3409 int uid1
= DECL_UID (c1
->members
[0]->decl
);
3410 int uid2
= DECL_UID (c2
->members
[0]->decl
);
3414 else if (uid1
> uid2
)
3420 /* Sort pair of congruence_class_groups A and B by
3421 DECL_UID of the first member of a first group. */
3424 sort_congruence_class_groups_by_decl_uid (const void *a
, const void *b
)
3426 const congruence_class_group
*g1
3427 = *(const congruence_class_group
* const *)a
;
3428 const congruence_class_group
*g2
3429 = *(const congruence_class_group
* const *)b
;
3431 int uid1
= DECL_UID (g1
->classes
[0]->members
[0]->decl
);
3432 int uid2
= DECL_UID (g2
->classes
[0]->members
[0]->decl
);
3436 else if (uid1
> uid2
)
3442 /* After reduction is done, we can declare all items in a group
3443 to be equal. PREV_CLASS_COUNT is start number of classes
3444 before reduction. True is returned if there's a merge operation
3448 sem_item_optimizer::merge_classes (unsigned int prev_class_count
)
3450 unsigned int item_count
= m_items
.length ();
3451 unsigned int class_count
= m_classes_count
;
3452 unsigned int equal_items
= item_count
- class_count
;
3454 unsigned int non_singular_classes_count
= 0;
3455 unsigned int non_singular_classes_sum
= 0;
3457 bool merged_p
= false;
3460 Sort functions in congruence classes by DECL_UID and do the same
3461 for the classes to not to break -fcompare-debug. */
3463 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3464 it
!= m_classes
.end (); ++it
)
3466 for (unsigned int i
= 0; i
< (*it
)->classes
.length (); i
++)
3468 congruence_class
*c
= (*it
)->classes
[i
];
3469 c
->members
.qsort (sort_sem_items_by_decl_uid
);
3472 (*it
)->classes
.qsort (sort_congruence_classes_by_decl_uid
);
3475 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3476 it
!= m_classes
.end (); ++it
)
3477 for (unsigned int i
= 0; i
< (*it
)->classes
.length (); i
++)
3479 congruence_class
*c
= (*it
)->classes
[i
];
3480 if (c
->members
.length () > 1)
3482 non_singular_classes_count
++;
3483 non_singular_classes_sum
+= c
->members
.length ();
3487 auto_vec
<congruence_class_group
*> classes (m_classes
.elements ());
3488 for (hash_table
<congruence_class_hash
>::iterator it
= m_classes
.begin ();
3489 it
!= m_classes
.end (); ++it
)
3490 classes
.quick_push (*it
);
3492 classes
.qsort (sort_congruence_class_groups_by_decl_uid
);
3496 fprintf (dump_file
, "\nItem count: %u\n", item_count
);
3497 fprintf (dump_file
, "Congruent classes before: %u, after: %u\n",
3498 prev_class_count
, class_count
);
3499 fprintf (dump_file
, "Average class size before: %.2f, after: %.2f\n",
3500 prev_class_count
? 1.0f
* item_count
/ prev_class_count
: 0.0f
,
3501 class_count
? 1.0f
* item_count
/ class_count
: 0.0f
);
3502 fprintf (dump_file
, "Average non-singular class size: %.2f, count: %u\n",
3503 non_singular_classes_count
? 1.0f
* non_singular_classes_sum
/
3504 non_singular_classes_count
: 0.0f
,
3505 non_singular_classes_count
);
3506 fprintf (dump_file
, "Equal symbols: %u\n", equal_items
);
3507 fprintf (dump_file
, "Fraction of visited symbols: %.2f%%\n\n",
3508 item_count
? 100.0f
* equal_items
/ item_count
: 0.0f
);
3512 congruence_class_group
*it
;
3513 FOR_EACH_VEC_ELT (classes
, l
, it
)
3514 for (unsigned int i
= 0; i
< it
->classes
.length (); i
++)
3516 congruence_class
*c
= it
->classes
[i
];
3518 if (c
->members
.length () == 1)
3521 sem_item
*source
= c
->members
[0];
3523 if (DECL_NAME (source
->decl
)
3524 && MAIN_NAME_P (DECL_NAME (source
->decl
)))
3525 /* If merge via wrappers, picking main as the target can be
3527 source
= c
->members
[1];
3529 for (unsigned int j
= 0; j
< c
->members
.length (); j
++)
3531 sem_item
*alias
= c
->members
[j
];
3533 if (alias
== source
)
3538 fprintf (dump_file
, "Semantic equality hit:%s->%s\n",
3539 xstrdup_for_dump (source
->node
->name ()),
3540 xstrdup_for_dump (alias
->node
->name ()));
3541 fprintf (dump_file
, "Assembler symbol names:%s->%s\n",
3542 xstrdup_for_dump (source
->node
->asm_name ()),
3543 xstrdup_for_dump (alias
->node
->asm_name ()));
3546 if (lookup_attribute ("no_icf", DECL_ATTRIBUTES (alias
->decl
)))
3550 "Merge operation is skipped due to no_icf "
3556 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3558 source
->dump_to_file (dump_file
);
3559 alias
->dump_to_file (dump_file
);
3562 if (dbg_cnt (merged_ipa_icf
))
3563 merged_p
|= source
->merge (alias
);
3570 /* Dump function prints all class members to a FILE with an INDENT. */
3573 congruence_class::dump (FILE *file
, unsigned int indent
) const
3575 FPRINTF_SPACES (file
, indent
, "class with id: %u, hash: %u, items: %u\n",
3576 id
, members
[0]->get_hash (), members
.length ());
3578 FPUTS_SPACES (file
, indent
+ 2, "");
3579 for (unsigned i
= 0; i
< members
.length (); i
++)
3580 fprintf (file
, "%s(%p/%u) ", members
[i
]->node
->asm_name (),
3581 (void *) members
[i
]->decl
,
3582 members
[i
]->node
->order
);
3584 fprintf (file
, "\n");
3587 /* Returns true if there's a member that is used from another group. */
3590 congruence_class::is_class_used (void)
3592 for (unsigned int i
= 0; i
< members
.length (); i
++)
3593 if (members
[i
]->usages
.length ())
3599 /* Generate pass summary for IPA ICF pass. */
3602 ipa_icf_generate_summary (void)
3605 optimizer
= new sem_item_optimizer ();
3607 optimizer
->register_hooks ();
3608 optimizer
->parse_funcs_and_vars ();
3611 /* Write pass summary for IPA ICF pass. */
3614 ipa_icf_write_summary (void)
3616 gcc_assert (optimizer
);
3618 optimizer
->write_summary ();
3621 /* Read pass summary for IPA ICF pass. */
3624 ipa_icf_read_summary (void)
3627 optimizer
= new sem_item_optimizer ();
3629 optimizer
->read_summary ();
3630 optimizer
->register_hooks ();
3633 /* Semantic equality exection function. */
3636 ipa_icf_driver (void)
3638 gcc_assert (optimizer
);
3640 bool merged_p
= optimizer
->execute ();
3645 return merged_p
? TODO_remove_functions
: 0;
3648 const pass_data pass_data_ipa_icf
=
3650 IPA_PASS
, /* type */
3652 OPTGROUP_IPA
, /* optinfo_flags */
3653 TV_IPA_ICF
, /* tv_id */
3654 0, /* properties_required */
3655 0, /* properties_provided */
3656 0, /* properties_destroyed */
3657 0, /* todo_flags_start */
3658 0, /* todo_flags_finish */
3661 class pass_ipa_icf
: public ipa_opt_pass_d
3664 pass_ipa_icf (gcc::context
*ctxt
)
3665 : ipa_opt_pass_d (pass_data_ipa_icf
, ctxt
,
3666 ipa_icf_generate_summary
, /* generate_summary */
3667 ipa_icf_write_summary
, /* write_summary */
3668 ipa_icf_read_summary
, /* read_summary */
3670 write_optimization_summary */
3672 read_optimization_summary */
3673 NULL
, /* stmt_fixup */
3674 0, /* function_transform_todo_flags_start */
3675 NULL
, /* function_transform */
3676 NULL
) /* variable_transform */
3679 /* opt_pass methods: */
3680 virtual bool gate (function
*)
3682 return in_lto_p
|| flag_ipa_icf_variables
|| flag_ipa_icf_functions
;
3685 virtual unsigned int execute (function
*)
3687 return ipa_icf_driver();
3689 }; // class pass_ipa_icf
3691 } // ipa_icf namespace
3694 make_pass_ipa_icf (gcc::context
*ctxt
)
3696 return new ipa_icf::pass_ipa_icf (ctxt
);