1 /* Basic IPA optimizations and utilities.
2 Copyright (C) 2003-2018 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
22 #include "coretypes.h"
27 #include "alloc-pool.h"
28 #include "tree-pass.h"
29 #include "stringpool.h"
32 #include "tree-iterator.h"
33 #include "ipa-utils.h"
34 #include "symbol-summary.h"
37 #include "ipa-fnsummary.h"
40 #include "stringpool.h"
43 /* Return true when NODE has ADDR reference. */
46 has_addr_references_p (struct cgraph_node
*node
,
50 struct ipa_ref
*ref
= NULL
;
52 for (i
= 0; node
->iterate_referring (i
, ref
); i
++)
53 if (ref
->use
== IPA_REF_ADDR
)
58 /* Return true when NODE can be target of an indirect call. */
61 is_indirect_call_target_p (struct cgraph_node
*node
, void *)
63 return node
->indirect_call_target
;
66 /* Look for all functions inlined to NODE and update their inlined_to pointers
70 update_inlined_to_pointer (struct cgraph_node
*node
, struct cgraph_node
*inlined_to
)
72 struct cgraph_edge
*e
;
73 for (e
= node
->callees
; e
; e
= e
->next_callee
)
74 if (e
->callee
->global
.inlined_to
)
76 e
->callee
->global
.inlined_to
= inlined_to
;
77 update_inlined_to_pointer (e
->callee
, inlined_to
);
81 /* Add symtab NODE to queue starting at FIRST.
83 The queue is linked via AUX pointers and terminated by pointer to 1.
84 We enqueue nodes at two occasions: when we find them reachable or when we find
85 their bodies needed for further clonning. In the second case we mark them
86 by pointer to 2 after processing so they are re-queue when they become
90 enqueue_node (symtab_node
*node
, symtab_node
**first
,
91 hash_set
<symtab_node
*> *reachable
)
93 /* Node is still in queue; do nothing. */
94 if (node
->aux
&& node
->aux
!= (void *) 2)
96 /* Node was already processed as unreachable, re-enqueue
97 only if it became reachable now. */
98 if (node
->aux
== (void *)2 && !reachable
->contains (node
))
104 /* Process references. */
107 process_references (symtab_node
*snode
,
109 bool before_inlining_p
,
110 hash_set
<symtab_node
*> *reachable
)
113 struct ipa_ref
*ref
= NULL
;
114 for (i
= 0; snode
->iterate_reference (i
, ref
); i
++)
116 symtab_node
*node
= ref
->referred
;
117 symtab_node
*body
= node
->ultimate_alias_target ();
119 if (node
->definition
&& !node
->in_other_partition
120 && ((!DECL_EXTERNAL (node
->decl
) || node
->alias
)
121 || (((before_inlining_p
122 && (TREE_CODE (node
->decl
) != FUNCTION_DECL
123 || (TREE_CODE (node
->decl
) == FUNCTION_DECL
124 && opt_for_fn (body
->decl
, optimize
))
125 || (symtab
->state
< IPA_SSA
128 DECL_ATTRIBUTES (body
->decl
))))))
129 /* We use variable constructors during late compilation for
130 constant folding. Keep references alive so partitioning
131 knows about potential references. */
132 || (VAR_P (node
->decl
)
134 || flag_incremental_link
135 == INCREMENTAL_LINK_LTO
)
136 && dyn_cast
<varpool_node
*> (node
)
137 ->ctor_useable_for_folding_p ()))))
139 /* Be sure that we will not optimize out alias target
141 if (DECL_EXTERNAL (node
->decl
)
143 && before_inlining_p
)
144 reachable
->add (body
);
145 reachable
->add (node
);
147 enqueue_node (node
, first
, reachable
);
151 /* EDGE is an polymorphic call. If BEFORE_INLINING_P is set, mark
152 all its potential targets as reachable to permit later inlining if
153 devirtualization happens. After inlining still keep their declarations
154 around, so we can devirtualize to a direct call.
156 Also try to make trivial devirutalization when no or only one target is
160 walk_polymorphic_call_targets (hash_set
<void *> *reachable_call_targets
,
161 struct cgraph_edge
*edge
,
163 hash_set
<symtab_node
*> *reachable
,
164 bool before_inlining_p
)
169 vec
<cgraph_node
*>targets
170 = possible_polymorphic_call_targets
171 (edge
, &final
, &cache_token
);
173 if (!reachable_call_targets
->add (cache_token
))
175 for (i
= 0; i
< targets
.length (); i
++)
177 struct cgraph_node
*n
= targets
[i
];
179 /* Do not bother to mark virtual methods in anonymous namespace;
180 either we will find use of virtual table defining it, or it is
182 if (TREE_CODE (TREE_TYPE (n
->decl
)) == METHOD_TYPE
183 && type_in_anonymous_namespace_p
184 (TYPE_METHOD_BASETYPE (TREE_TYPE (n
->decl
))))
187 n
->indirect_call_target
= true;
188 symtab_node
*body
= n
->function_symbol ();
190 /* Prior inlining, keep alive bodies of possible targets for
193 && (before_inlining_p
194 && opt_for_fn (body
->decl
, optimize
)
195 && opt_for_fn (body
->decl
, flag_devirtualize
)))
197 /* Be sure that we will not optimize out alias target
199 if (DECL_EXTERNAL (n
->decl
)
201 && before_inlining_p
)
202 reachable
->add (body
);
205 /* Even after inlining we want to keep the possible targets in the
206 boundary, so late passes can still produce direct call even if
207 the chance for inlining is lost. */
208 enqueue_node (n
, first
, reachable
);
212 /* Very trivial devirtualization; when the type is
213 final or anonymous (so we know all its derivation)
214 and there is only one possible virtual call target,
215 make the edge direct. */
218 if (targets
.length () <= 1 && dbg_cnt (devirt
))
220 cgraph_node
*target
, *node
= edge
->caller
;
221 if (targets
.length () == 1)
224 target
= cgraph_node::get_create
225 (builtin_decl_implicit (BUILT_IN_UNREACHABLE
));
227 if (dump_enabled_p ())
229 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, edge
->call_stmt
,
230 "devirtualizing call in %s to %s\n",
231 edge
->caller
->dump_name (),
232 target
->dump_name ());
234 edge
= edge
->make_direct (target
);
235 if (ipa_fn_summaries
)
236 ipa_update_overall_fn_summary (node
);
237 else if (edge
->call_stmt
)
238 edge
->redirect_call_stmt_to_callee ();
243 /* Perform reachability analysis and reclaim all unreachable nodes.
245 The algorithm is basically mark&sweep but with some extra refinements:
247 - reachable extern inline functions needs special handling; the bodies needs
248 to stay in memory until inlining in hope that they will be inlined.
249 After inlining we release their bodies and turn them into unanalyzed
250 nodes even when they are reachable.
252 - virtual functions are kept in callgraph even if they seem unreachable in
253 hope calls to them will be devirtualized.
255 Again we remove them after inlining. In late optimization some
256 devirtualization may happen, but it is not important since we won't inline
257 the call. In theory early opts and IPA should work out all important cases.
259 - virtual clones needs bodies of their origins for later materialization;
260 this means that we want to keep the body even if the origin is unreachable
261 otherwise. To avoid origin from sitting in the callgraph and being
262 walked by IPA passes, we turn them into unanalyzed nodes with body
265 We maintain set of function declaration where body needs to stay in
266 body_needed_for_clonning
268 Inline clones represent special case: their declaration match the
269 declaration of origin and cgraph_remove_node already knows how to
270 reshape callgraph and preserve body when offline copy of function or
271 inline clone is being removed.
273 - C++ virtual tables keyed to other unit are represented as DECL_EXTERNAL
274 variables with DECL_INITIAL set. We finalize these and keep reachable
275 ones around for constant folding purposes. After inlining we however
276 stop walking their references to let everything static referneced by them
277 to be removed when it is otherwise unreachable.
279 We maintain queue of both reachable symbols (i.e. defined symbols that needs
280 to stay) and symbols that are in boundary (i.e. external symbols referenced
281 by reachable symbols or origins of clones). The queue is represented
282 as linked list by AUX pointer terminated by 1.
284 At the end we keep all reachable symbols. For symbols in boundary we always
285 turn definition into a declaration, but we may keep function body around
286 based on body_needed_for_clonning
288 All symbols that enter the queue have AUX pointer non-zero and are in the
289 boundary. Pointer set REACHABLE is used to track reachable symbols.
291 Every symbol can be visited twice - once as part of boundary and once
292 as real reachable symbol. enqueue_node needs to decide whether the
293 node needs to be re-queued for second processing. For this purpose
294 we set AUX pointer of processed symbols in the boundary to constant 2. */
297 symbol_table::remove_unreachable_nodes (FILE *file
)
299 symtab_node
*first
= (symtab_node
*) (void *) 1;
300 struct cgraph_node
*node
, *next
;
301 varpool_node
*vnode
, *vnext
;
302 bool changed
= false;
303 hash_set
<symtab_node
*> reachable
;
304 hash_set
<tree
> body_needed_for_clonning
;
305 hash_set
<void *> reachable_call_targets
;
306 bool before_inlining_p
= symtab
->state
< (!optimize
&& !in_lto_p
? IPA_SSA
307 : IPA_SSA_AFTER_INLINING
);
309 timevar_push (TV_IPA_UNREACHABLE
);
310 build_type_inheritance_graph ();
312 fprintf (file
, "\nReclaiming functions:");
315 FOR_EACH_FUNCTION (node
)
316 gcc_assert (!node
->aux
);
317 FOR_EACH_VARIABLE (vnode
)
318 gcc_assert (!vnode
->aux
);
320 /* Mark functions whose bodies are obviously needed.
321 This is mostly when they can be referenced externally. Inline clones
322 are special since their declarations are shared with master clone and thus
323 cgraph_can_remove_if_no_direct_calls_and_refs_p should not be called on them. */
324 FOR_EACH_FUNCTION (node
)
326 node
->used_as_abstract_origin
= false;
327 node
->indirect_call_target
= false;
329 && !node
->global
.inlined_to
330 && !node
->in_other_partition
331 && !node
->can_remove_if_no_direct_calls_and_refs_p ())
333 gcc_assert (!node
->global
.inlined_to
);
334 reachable
.add (node
);
335 enqueue_node (node
, &first
, &reachable
);
338 gcc_assert (!node
->aux
);
341 /* Mark variables that are obviously needed. */
342 FOR_EACH_DEFINED_VARIABLE (vnode
)
343 if (!vnode
->can_remove_if_no_refs_p()
344 && !vnode
->in_other_partition
)
346 reachable
.add (vnode
);
347 enqueue_node (vnode
, &first
, &reachable
);
350 /* Perform reachability analysis. */
351 while (first
!= (symtab_node
*) (void *) 1)
353 bool in_boundary_p
= !reachable
.contains (first
);
354 symtab_node
*node
= first
;
356 first
= (symtab_node
*)first
->aux
;
358 /* If we are processing symbol in boundary, mark its AUX pointer for
359 possible later re-processing in enqueue_node. */
362 node
->aux
= (void *)2;
363 if (node
->alias
&& node
->analyzed
)
364 enqueue_node (node
->get_alias_target (), &first
, &reachable
);
368 if (TREE_CODE (node
->decl
) == FUNCTION_DECL
369 && DECL_ABSTRACT_ORIGIN (node
->decl
))
371 struct cgraph_node
*origin_node
372 = cgraph_node::get (DECL_ABSTRACT_ORIGIN (node
->decl
));
373 if (origin_node
&& !origin_node
->used_as_abstract_origin
)
375 origin_node
->used_as_abstract_origin
= true;
376 gcc_assert (!origin_node
->prev_sibling_clone
);
377 gcc_assert (!origin_node
->next_sibling_clone
);
378 for (cgraph_node
*n
= origin_node
->clones
; n
;
379 n
= n
->next_sibling_clone
)
380 if (n
->decl
== DECL_ABSTRACT_ORIGIN (node
->decl
))
381 n
->used_as_abstract_origin
= true;
384 /* If any symbol in a comdat group is reachable, force
385 all externally visible symbols in the same comdat
386 group to be reachable as well. Comdat-local symbols
387 can be discarded if all uses were inlined. */
388 if (node
->same_comdat_group
)
391 for (next
= node
->same_comdat_group
;
393 next
= next
->same_comdat_group
)
394 if (!next
->comdat_local_p ()
395 && !reachable
.add (next
))
396 enqueue_node (next
, &first
, &reachable
);
398 /* Mark references as reachable. */
399 process_references (node
, &first
, before_inlining_p
, &reachable
);
402 if (cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (node
))
404 /* Mark the callees reachable unless they are direct calls to extern
405 inline functions we decided to not inline. */
408 struct cgraph_edge
*e
;
409 /* Keep alive possible targets for devirtualization. */
410 if (opt_for_fn (cnode
->decl
, optimize
)
411 && opt_for_fn (cnode
->decl
, flag_devirtualize
))
413 struct cgraph_edge
*next
;
414 for (e
= cnode
->indirect_calls
; e
; e
= next
)
416 next
= e
->next_callee
;
417 if (e
->indirect_info
->polymorphic
)
418 walk_polymorphic_call_targets (&reachable_call_targets
,
419 e
, &first
, &reachable
,
423 for (e
= cnode
->callees
; e
; e
= e
->next_callee
)
425 symtab_node
*body
= e
->callee
->function_symbol ();
426 if (e
->callee
->definition
427 && !e
->callee
->in_other_partition
428 && (!e
->inline_failed
429 || !DECL_EXTERNAL (e
->callee
->decl
)
431 || (before_inlining_p
432 && (opt_for_fn (body
->decl
, optimize
)
433 || (symtab
->state
< IPA_SSA
436 DECL_ATTRIBUTES (body
->decl
)))))))
438 /* Be sure that we will not optimize out alias target
440 if (DECL_EXTERNAL (e
->callee
->decl
)
442 && before_inlining_p
)
443 reachable
.add (body
);
444 reachable
.add (e
->callee
);
446 enqueue_node (e
->callee
, &first
, &reachable
);
449 /* When inline clone exists, mark body to be preserved so when removing
450 offline copy of the function we don't kill it. */
451 if (cnode
->global
.inlined_to
)
452 body_needed_for_clonning
.add (cnode
->decl
);
454 /* For non-inline clones, force their origins to the boundary and ensure
455 that body is not removed. */
456 while (cnode
->clone_of
)
458 bool noninline
= cnode
->clone_of
->decl
!= cnode
->decl
;
459 cnode
= cnode
->clone_of
;
462 body_needed_for_clonning
.add (cnode
->decl
);
463 enqueue_node (cnode
, &first
, &reachable
);
468 else if (cnode
->thunk
.thunk_p
)
469 enqueue_node (cnode
->callees
->callee
, &first
, &reachable
);
471 /* If any reachable function has simd clones, mark them as
472 reachable as well. */
473 if (cnode
->simd_clones
)
476 for (next
= cnode
->simd_clones
;
478 next
= next
->simdclone
->next_clone
)
480 || !reachable
.add (next
))
481 enqueue_node (next
, &first
, &reachable
);
484 /* When we see constructor of external variable, keep referred nodes in the
485 boundary. This will also hold initializers of the external vars NODE
487 varpool_node
*vnode
= dyn_cast
<varpool_node
*> (node
);
489 && DECL_EXTERNAL (node
->decl
)
493 struct ipa_ref
*ref
= NULL
;
494 for (int i
= 0; node
->iterate_reference (i
, ref
); i
++)
495 enqueue_node (ref
->referred
, &first
, &reachable
);
499 /* Remove unreachable functions. */
500 for (node
= first_function (); node
; node
= next
)
502 next
= next_function (node
);
504 /* If node is not needed at all, remove it. */
508 fprintf (file
, " %s", node
->dump_name ());
512 /* If node is unreachable, remove its body. */
513 else if (!reachable
.contains (node
))
515 /* We keep definitions of thunks and aliases in the boundary so
516 we can walk to the ultimate alias targets and function symbols
518 if (node
->alias
|| node
->thunk
.thunk_p
)
520 else if (!body_needed_for_clonning
.contains (node
->decl
)
521 && !node
->alias
&& !node
->thunk
.thunk_p
)
522 node
->release_body ();
523 else if (!node
->clone_of
)
524 gcc_assert (in_lto_p
|| DECL_RESULT (node
->decl
));
525 if (node
->definition
&& !node
->alias
&& !node
->thunk
.thunk_p
)
528 fprintf (file
, " %s", node
->dump_name ());
529 node
->body_removed
= true;
530 node
->analyzed
= false;
531 node
->definition
= false;
532 node
->cpp_implicit_alias
= false;
534 node
->transparent_alias
= false;
535 node
->thunk
.thunk_p
= false;
536 node
->weakref
= false;
537 /* After early inlining we drop always_inline attributes on
538 bodies of functions that are still referenced (have their
540 DECL_ATTRIBUTES (node
->decl
)
541 = remove_attribute ("always_inline",
542 DECL_ATTRIBUTES (node
->decl
));
543 if (!node
->in_other_partition
)
544 node
->local
.local
= false;
545 node
->remove_callees ();
546 node
->remove_all_references ();
548 if (node
->thunk
.thunk_p
549 && node
->thunk
.add_pointer_bounds_args
)
551 node
->thunk
.thunk_p
= false;
552 node
->thunk
.add_pointer_bounds_args
= false;
557 gcc_assert (node
->clone_of
|| !node
->has_gimple_body_p ()
558 || in_lto_p
|| DECL_RESULT (node
->decl
));
561 /* Inline clones might be kept around so their materializing allows further
562 cloning. If the function the clone is inlined into is removed, we need
563 to turn it into normal cone. */
564 FOR_EACH_FUNCTION (node
)
566 if (node
->global
.inlined_to
569 gcc_assert (node
->clones
);
570 node
->global
.inlined_to
= NULL
;
571 update_inlined_to_pointer (node
, node
);
576 /* Remove unreachable variables. */
578 fprintf (file
, "\nReclaiming variables:");
579 for (vnode
= first_variable (); vnode
; vnode
= vnext
)
581 vnext
= next_variable (vnode
);
583 /* For can_refer_decl_in_current_unit_p we want to track for
584 all external variables if they are defined in other partition
586 && (!flag_ltrans
|| !DECL_EXTERNAL (vnode
->decl
)))
588 struct ipa_ref
*ref
= NULL
;
590 /* First remove the aliases, so varpool::remove can possibly lookup
591 the constructor and save it for future use. */
592 while (vnode
->iterate_direct_aliases (0, ref
))
595 fprintf (file
, " %s", ref
->referred
->dump_name ());
596 ref
->referring
->remove ();
599 fprintf (file
, " %s", vnode
->dump_name ());
600 vnext
= next_variable (vnode
);
601 /* Signal removal to the debug machinery. */
602 if (! flag_wpa
|| flag_incremental_link
== INCREMENTAL_LINK_LTO
)
604 vnode
->definition
= false;
605 (*debug_hooks
->late_global_decl
) (vnode
->decl
);
610 else if (!reachable
.contains (vnode
) && !vnode
->alias
)
613 if (vnode
->definition
)
616 fprintf (file
, " %s", vnode
->name ());
619 /* Keep body if it may be useful for constant folding. */
620 if ((flag_wpa
|| flag_incremental_link
== INCREMENTAL_LINK_LTO
)
621 || ((init
= ctor_for_folding (vnode
->decl
)) == error_mark_node
))
622 vnode
->remove_initializer ();
624 DECL_INITIAL (vnode
->decl
) = init
;
625 vnode
->body_removed
= true;
626 vnode
->definition
= false;
627 vnode
->analyzed
= false;
630 vnode
->remove_from_same_comdat_group ();
632 vnode
->remove_all_references ();
638 /* Now update address_taken flags and try to promote functions to be local. */
640 fprintf (file
, "\nClearing address taken flags:");
641 FOR_EACH_DEFINED_FUNCTION (node
)
642 if (node
->address_taken
643 && !node
->used_from_other_partition
)
645 if (!node
->call_for_symbol_and_aliases
646 (has_addr_references_p
, NULL
, true))
649 fprintf (file
, " %s", node
->name ());
650 node
->address_taken
= false;
653 /* Virtual functions may be kept in cgraph just because
654 of possible later devirtualization. Do not mark them as
655 local too early so we won't optimize them out before
656 we are done with polymorphic call analysis. */
657 && (!before_inlining_p
658 || !node
->call_for_symbol_and_aliases
659 (is_indirect_call_target_p
, NULL
, true)))
661 node
->local
.local
= true;
663 fprintf (file
, " (local)");
668 fprintf (file
, "\n");
670 symtab_node::checking_verify_symtab_nodes ();
672 /* If we removed something, perhaps profile could be improved. */
673 if (changed
&& (optimize
|| in_lto_p
) && ipa_call_summaries
)
674 FOR_EACH_DEFINED_FUNCTION (node
)
675 ipa_propagate_frequency (node
);
677 timevar_pop (TV_IPA_UNREACHABLE
);
681 /* Process references to VNODE and set flags WRITTEN, ADDRESS_TAKEN, READ
682 as needed, also clear EXPLICIT_REFS if the references to given variable
683 do not need to be explicit. */
686 process_references (varpool_node
*vnode
,
687 bool *written
, bool *address_taken
,
688 bool *read
, bool *explicit_refs
)
693 if (!vnode
->all_refs_explicit_p ()
694 || TREE_THIS_VOLATILE (vnode
->decl
))
695 *explicit_refs
= false;
697 for (i
= 0; vnode
->iterate_referring (i
, ref
)
698 && *explicit_refs
&& (!*written
|| !*address_taken
|| !*read
); i
++)
702 *address_taken
= true;
711 process_references (dyn_cast
<varpool_node
*> (ref
->referring
), written
,
712 address_taken
, read
, explicit_refs
);
717 /* Set TREE_READONLY bit. */
720 set_readonly_bit (varpool_node
*vnode
, void *data ATTRIBUTE_UNUSED
)
722 TREE_READONLY (vnode
->decl
) = true;
726 /* Set writeonly bit and clear the initalizer, since it will not be needed. */
729 set_writeonly_bit (varpool_node
*vnode
, void *data
)
731 vnode
->writeonly
= true;
732 if (optimize
|| in_lto_p
)
734 DECL_INITIAL (vnode
->decl
) = NULL
;
737 if (vnode
->num_references ())
738 *(bool *)data
= true;
739 vnode
->remove_all_references ();
745 /* Clear addressale bit of VNODE. */
748 clear_addressable_bit (varpool_node
*vnode
, void *data ATTRIBUTE_UNUSED
)
750 vnode
->address_taken
= false;
751 TREE_ADDRESSABLE (vnode
->decl
) = 0;
755 /* Discover variables that have no longer address taken or that are read only
756 and update their flags.
758 Return true when unreachable symbol removan should be done.
760 FIXME: This can not be done in between gimplify and omp_expand since
761 readonly flag plays role on what is shared and what is not. Currently we do
762 this transformation as part of whole program visibility and re-do at
763 ipa-reference pass (to take into account clonning), but it would
764 make sense to do it before early optimizations. */
767 ipa_discover_readonly_nonaddressable_vars (void)
769 bool remove_p
= false;
772 fprintf (dump_file
, "Clearing variable flags:");
773 FOR_EACH_VARIABLE (vnode
)
775 && (TREE_ADDRESSABLE (vnode
->decl
)
777 || !TREE_READONLY (vnode
->decl
)))
779 bool written
= false;
780 bool address_taken
= false;
782 bool explicit_refs
= true;
784 process_references (vnode
, &written
, &address_taken
, &read
,
790 if (TREE_ADDRESSABLE (vnode
->decl
) && dump_file
)
791 fprintf (dump_file
, " %s (non-addressable)", vnode
->name ());
792 vnode
->call_for_symbol_and_aliases (clear_addressable_bit
, NULL
,
795 if (!address_taken
&& !written
796 /* Making variable in explicit section readonly can cause section
798 See e.g. gcc.c-torture/compile/pr23237.c */
799 && vnode
->get_section () == NULL
)
801 if (!TREE_READONLY (vnode
->decl
) && dump_file
)
802 fprintf (dump_file
, " %s (read-only)", vnode
->name ());
803 vnode
->call_for_symbol_and_aliases (set_readonly_bit
, NULL
, true);
805 if (!vnode
->writeonly
&& !read
&& !address_taken
&& written
)
808 fprintf (dump_file
, " %s (write-only)", vnode
->name ());
809 vnode
->call_for_symbol_and_aliases (set_writeonly_bit
, &remove_p
,
814 fprintf (dump_file
, "\n");
818 /* Generate and emit a static constructor or destructor. WHICH must
819 be one of 'I' (for a constructor), 'D' (for a destructor).
820 BODY is a STATEMENT_LIST containing GENERIC
821 statements. PRIORITY is the initialization priority for this
822 constructor or destructor.
824 FINAL specify whether the externally visible name for collect2 should
828 cgraph_build_static_cdtor_1 (char which
, tree body
, int priority
, bool final
)
830 static int counter
= 0;
832 tree decl
, name
, resdecl
;
834 /* The priority is encoded in the constructor or destructor name.
835 collect2 will sort the names and arrange that they are called at
838 sprintf (which_buf
, "%c_%.5d_%d", which
, priority
, counter
++);
840 /* Proudce sane name but one not recognizable by collect2, just for the
841 case we fail to inline the function. */
842 sprintf (which_buf
, "sub_%c_%.5d_%d", which
, priority
, counter
++);
843 name
= get_file_function_name (which_buf
);
845 decl
= build_decl (input_location
, FUNCTION_DECL
, name
,
846 build_function_type_list (void_type_node
, NULL_TREE
));
847 current_function_decl
= decl
;
849 resdecl
= build_decl (input_location
,
850 RESULT_DECL
, NULL_TREE
, void_type_node
);
851 DECL_ARTIFICIAL (resdecl
) = 1;
852 DECL_RESULT (decl
) = resdecl
;
853 DECL_CONTEXT (resdecl
) = decl
;
855 allocate_struct_function (decl
, false);
857 TREE_STATIC (decl
) = 1;
858 TREE_USED (decl
) = 1;
859 DECL_ARTIFICIAL (decl
) = 1;
860 DECL_IGNORED_P (decl
) = 1;
861 DECL_NO_INSTRUMENT_FUNCTION_ENTRY_EXIT (decl
) = 1;
862 DECL_SAVED_TREE (decl
) = body
;
863 if (!targetm
.have_ctors_dtors
&& final
)
865 TREE_PUBLIC (decl
) = 1;
866 DECL_PRESERVE_P (decl
) = 1;
868 DECL_UNINLINABLE (decl
) = 1;
870 DECL_INITIAL (decl
) = make_node (BLOCK
);
871 BLOCK_SUPERCONTEXT (DECL_INITIAL (decl
)) = decl
;
872 TREE_USED (DECL_INITIAL (decl
)) = 1;
874 DECL_SOURCE_LOCATION (decl
) = input_location
;
875 cfun
->function_end_locus
= input_location
;
880 DECL_STATIC_CONSTRUCTOR (decl
) = 1;
881 decl_init_priority_insert (decl
, priority
);
884 DECL_STATIC_DESTRUCTOR (decl
) = 1;
885 decl_fini_priority_insert (decl
, priority
);
891 gimplify_function_tree (decl
);
893 cgraph_node::add_new_function (decl
, false);
896 current_function_decl
= NULL
;
899 /* Generate and emit a static constructor or destructor. WHICH must
900 be one of 'I' (for a constructor) or 'D' (for a destructor).
901 BODY is a STATEMENT_LIST containing GENERIC
902 statements. PRIORITY is the initialization priority for this
903 constructor or destructor. */
906 cgraph_build_static_cdtor (char which
, tree body
, int priority
)
908 cgraph_build_static_cdtor_1 (which
, body
, priority
, false);
911 /* When target does not have ctors and dtors, we call all constructor
912 and destructor by special initialization/destruction function
913 recognized by collect2.
915 When we are going to build this function, collect all constructors and
916 destructors and turn them into normal functions. */
919 record_cdtor_fn (struct cgraph_node
*node
, vec
<tree
> *ctors
, vec
<tree
> *dtors
)
921 if (DECL_STATIC_CONSTRUCTOR (node
->decl
))
922 ctors
->safe_push (node
->decl
);
923 if (DECL_STATIC_DESTRUCTOR (node
->decl
))
924 dtors
->safe_push (node
->decl
);
925 node
= cgraph_node::get (node
->decl
);
926 DECL_DISREGARD_INLINE_LIMITS (node
->decl
) = 1;
929 /* Define global constructors/destructor functions for the CDTORS, of
930 which they are LEN. The CDTORS are sorted by initialization
931 priority. If CTOR_P is true, these are constructors; otherwise,
932 they are destructors. */
935 build_cdtor (bool ctor_p
, const vec
<tree
> &cdtors
)
938 size_t len
= cdtors
.length ();
945 priority_type priority
;
954 p
= ctor_p
? DECL_INIT_PRIORITY (fn
) : DECL_FINI_PRIORITY (fn
);
957 else if (p
!= priority
)
963 /* When there is only one cdtor and target supports them, do nothing. */
965 && targetm
.have_ctors_dtors
)
970 /* Find the next batch of constructors/destructors with the same
971 initialization priority. */
976 call
= build_call_expr (fn
, 0);
978 DECL_STATIC_CONSTRUCTOR (fn
) = 0;
980 DECL_STATIC_DESTRUCTOR (fn
) = 0;
981 /* We do not want to optimize away pure/const calls here.
982 When optimizing, these should be already removed, when not
983 optimizing, we want user to be able to breakpoint in them. */
984 TREE_SIDE_EFFECTS (call
) = 1;
985 append_to_statement_list (call
, &body
);
987 gcc_assert (body
!= NULL_TREE
);
988 /* Generate a function to call all the function of like
990 cgraph_build_static_cdtor_1 (ctor_p
? 'I' : 'D', body
, priority
, true);
994 /* Comparison function for qsort. P1 and P2 are actually of type
995 "tree *" and point to static constructors. DECL_INIT_PRIORITY is
996 used to determine the sort order. */
999 compare_ctor (const void *p1
, const void *p2
)
1006 f1
= *(const tree
*)p1
;
1007 f2
= *(const tree
*)p2
;
1008 priority1
= DECL_INIT_PRIORITY (f1
);
1009 priority2
= DECL_INIT_PRIORITY (f2
);
1011 if (priority1
< priority2
)
1013 else if (priority1
> priority2
)
1016 /* Ensure a stable sort. Constructors are executed in backwarding
1017 order to make LTO initialize braries first. */
1018 return DECL_UID (f2
) - DECL_UID (f1
);
1021 /* Comparison function for qsort. P1 and P2 are actually of type
1022 "tree *" and point to static destructors. DECL_FINI_PRIORITY is
1023 used to determine the sort order. */
1026 compare_dtor (const void *p1
, const void *p2
)
1033 f1
= *(const tree
*)p1
;
1034 f2
= *(const tree
*)p2
;
1035 priority1
= DECL_FINI_PRIORITY (f1
);
1036 priority2
= DECL_FINI_PRIORITY (f2
);
1038 if (priority1
< priority2
)
1040 else if (priority1
> priority2
)
1043 /* Ensure a stable sort. */
1044 return DECL_UID (f1
) - DECL_UID (f2
);
1047 /* Generate functions to call static constructors and destructors
1048 for targets that do not support .ctors/.dtors sections. These
1049 functions have magic names which are detected by collect2. */
1052 build_cdtor_fns (vec
<tree
> *ctors
, vec
<tree
> *dtors
)
1054 if (!ctors
->is_empty ())
1056 gcc_assert (!targetm
.have_ctors_dtors
|| in_lto_p
);
1057 ctors
->qsort (compare_ctor
);
1058 build_cdtor (/*ctor_p=*/true, *ctors
);
1061 if (!dtors
->is_empty ())
1063 gcc_assert (!targetm
.have_ctors_dtors
|| in_lto_p
);
1064 dtors
->qsort (compare_dtor
);
1065 build_cdtor (/*ctor_p=*/false, *dtors
);
1069 /* Look for constructors and destructors and produce function calling them.
1070 This is needed for targets not supporting ctors or dtors, but we perform the
1071 transformation also at linktime to merge possibly numerous
1072 constructors/destructors into single function to improve code locality and
1076 ipa_cdtor_merge (void)
1078 /* A vector of FUNCTION_DECLs declared as static constructors. */
1079 auto_vec
<tree
, 20> ctors
;
1080 /* A vector of FUNCTION_DECLs declared as static destructors. */
1081 auto_vec
<tree
, 20> dtors
;
1082 struct cgraph_node
*node
;
1083 FOR_EACH_DEFINED_FUNCTION (node
)
1084 if (DECL_STATIC_CONSTRUCTOR (node
->decl
)
1085 || DECL_STATIC_DESTRUCTOR (node
->decl
))
1086 record_cdtor_fn (node
, &ctors
, &dtors
);
1087 build_cdtor_fns (&ctors
, &dtors
);
1093 const pass_data pass_data_ipa_cdtor_merge
=
1095 IPA_PASS
, /* type */
1097 OPTGROUP_NONE
, /* optinfo_flags */
1098 TV_CGRAPHOPT
, /* tv_id */
1099 0, /* properties_required */
1100 0, /* properties_provided */
1101 0, /* properties_destroyed */
1102 0, /* todo_flags_start */
1103 0, /* todo_flags_finish */
1106 class pass_ipa_cdtor_merge
: public ipa_opt_pass_d
1109 pass_ipa_cdtor_merge (gcc::context
*ctxt
)
1110 : ipa_opt_pass_d (pass_data_ipa_cdtor_merge
, ctxt
,
1111 NULL
, /* generate_summary */
1112 NULL
, /* write_summary */
1113 NULL
, /* read_summary */
1114 NULL
, /* write_optimization_summary */
1115 NULL
, /* read_optimization_summary */
1116 NULL
, /* stmt_fixup */
1117 0, /* function_transform_todo_flags_start */
1118 NULL
, /* function_transform */
1119 NULL
) /* variable_transform */
1122 /* opt_pass methods: */
1123 virtual bool gate (function
*);
1124 virtual unsigned int execute (function
*) { return ipa_cdtor_merge (); }
1126 }; // class pass_ipa_cdtor_merge
1129 pass_ipa_cdtor_merge::gate (function
*)
1131 /* Perform the pass when we have no ctors/dtors support
1132 or at LTO time to merge multiple constructors into single
1134 return !targetm
.have_ctors_dtors
|| in_lto_p
;
1140 make_pass_ipa_cdtor_merge (gcc::context
*ctxt
)
1142 return new pass_ipa_cdtor_merge (ctxt
);
1145 /* Invalid pointer representing BOTTOM for single user dataflow. */
1146 #define BOTTOM ((cgraph_node *)(size_t) 2)
1148 /* Meet operation for single user dataflow.
1149 Here we want to associate variables with sigle function that may access it.
1151 FUNCTION is current single user of a variable, VAR is variable that uses it.
1152 Latttice is stored in SINGLE_USER_MAP.
1155 - TOP by no entry in SIGNLE_USER_MAP
1156 - BOTTOM by BOTTOM in AUX pointer (to save lookups)
1157 - known single user by cgraph pointer in SINGLE_USER_MAP. */
1160 meet (cgraph_node
*function
, varpool_node
*var
,
1161 hash_map
<varpool_node
*, cgraph_node
*> &single_user_map
)
1163 struct cgraph_node
*user
, **f
;
1165 if (var
->aux
== BOTTOM
)
1168 f
= single_user_map
.get (var
);
1174 else if (function
!= user
)
1180 /* Propagation step of single-use dataflow.
1182 Check all uses of VNODE and see if they are used by single function FUNCTION.
1183 SINGLE_USER_MAP represents the dataflow lattice. */
1186 propagate_single_user (varpool_node
*vnode
, cgraph_node
*function
,
1187 hash_map
<varpool_node
*, cgraph_node
*> &single_user_map
)
1190 struct ipa_ref
*ref
;
1192 gcc_assert (!vnode
->externally_visible
);
1194 /* If node is an alias, first meet with its target. */
1196 function
= meet (function
, vnode
->get_alias_target (), single_user_map
);
1198 /* Check all users and see if they correspond to a single function. */
1199 for (i
= 0; vnode
->iterate_referring (i
, ref
) && function
!= BOTTOM
; i
++)
1201 struct cgraph_node
*cnode
= dyn_cast
<cgraph_node
*> (ref
->referring
);
1204 if (cnode
->global
.inlined_to
)
1205 cnode
= cnode
->global
.inlined_to
;
1208 else if (function
!= cnode
)
1212 function
= meet (function
, dyn_cast
<varpool_node
*> (ref
->referring
),
1218 /* Pass setting used_by_single_function flag.
1219 This flag is set on variable when there is only one function that may
1220 possibly referr to it. */
1223 ipa_single_use (void)
1225 varpool_node
*first
= (varpool_node
*) (void *) 1;
1227 hash_map
<varpool_node
*, cgraph_node
*> single_user_map
;
1229 FOR_EACH_DEFINED_VARIABLE (var
)
1230 if (!var
->all_refs_explicit_p ())
1234 /* Enqueue symbol for dataflow. */
1239 /* The actual dataflow. */
1241 while (first
!= (void *) 1)
1243 cgraph_node
*user
, *orig_user
, **f
;
1246 first
= (varpool_node
*)first
->aux
;
1248 f
= single_user_map
.get (var
);
1253 user
= propagate_single_user (var
, orig_user
, single_user_map
);
1255 gcc_checking_assert (var
->aux
!= BOTTOM
);
1257 /* If user differs, enqueue all references. */
1258 if (user
!= orig_user
)
1263 single_user_map
.put (var
, user
);
1265 /* Enqueue all aliases for re-processing. */
1266 for (i
= 0; var
->iterate_direct_aliases (i
, ref
); i
++)
1267 if (!ref
->referring
->aux
)
1269 ref
->referring
->aux
= first
;
1270 first
= dyn_cast
<varpool_node
*> (ref
->referring
);
1272 /* Enqueue all users for re-processing. */
1273 for (i
= 0; var
->iterate_reference (i
, ref
); i
++)
1274 if (!ref
->referred
->aux
1275 && ref
->referred
->definition
1276 && is_a
<varpool_node
*> (ref
->referred
))
1278 ref
->referred
->aux
= first
;
1279 first
= dyn_cast
<varpool_node
*> (ref
->referred
);
1282 /* If user is BOTTOM, just punt on this var. */
1292 FOR_EACH_DEFINED_VARIABLE (var
)
1294 if (var
->aux
!= BOTTOM
)
1296 /* Not having the single user known means that the VAR is
1297 unreachable. Either someone forgot to remove unreachable
1298 variables or the reachability here is wrong. */
1300 gcc_checking_assert (single_user_map
.get (var
));
1304 fprintf (dump_file
, "Variable %s is used by single function\n",
1307 var
->used_by_single_function
= true;
1316 const pass_data pass_data_ipa_single_use
=
1318 IPA_PASS
, /* type */
1319 "single-use", /* name */
1320 OPTGROUP_NONE
, /* optinfo_flags */
1321 TV_CGRAPHOPT
, /* tv_id */
1322 0, /* properties_required */
1323 0, /* properties_provided */
1324 0, /* properties_destroyed */
1325 0, /* todo_flags_start */
1326 0, /* todo_flags_finish */
1329 class pass_ipa_single_use
: public ipa_opt_pass_d
1332 pass_ipa_single_use (gcc::context
*ctxt
)
1333 : ipa_opt_pass_d (pass_data_ipa_single_use
, ctxt
,
1334 NULL
, /* generate_summary */
1335 NULL
, /* write_summary */
1336 NULL
, /* read_summary */
1337 NULL
, /* write_optimization_summary */
1338 NULL
, /* read_optimization_summary */
1339 NULL
, /* stmt_fixup */
1340 0, /* function_transform_todo_flags_start */
1341 NULL
, /* function_transform */
1342 NULL
) /* variable_transform */
1345 /* opt_pass methods: */
1346 virtual unsigned int execute (function
*) { return ipa_single_use (); }
1348 }; // class pass_ipa_single_use
1353 make_pass_ipa_single_use (gcc::context
*ctxt
)
1355 return new pass_ipa_single_use (ctxt
);
1358 /* Materialize all clones. */
1362 const pass_data pass_data_materialize_all_clones
=
1364 SIMPLE_IPA_PASS
, /* type */
1365 "materialize-all-clones", /* name */
1366 OPTGROUP_NONE
, /* optinfo_flags */
1367 TV_IPA_OPT
, /* tv_id */
1368 0, /* properties_required */
1369 0, /* properties_provided */
1370 0, /* properties_destroyed */
1371 0, /* todo_flags_start */
1372 0, /* todo_flags_finish */
1375 class pass_materialize_all_clones
: public simple_ipa_opt_pass
1378 pass_materialize_all_clones (gcc::context
*ctxt
)
1379 : simple_ipa_opt_pass (pass_data_materialize_all_clones
, ctxt
)
1382 /* opt_pass methods: */
1383 virtual unsigned int execute (function
*)
1385 symtab
->materialize_all_clones ();
1389 }; // class pass_materialize_all_clones
1393 simple_ipa_opt_pass
*
1394 make_pass_materialize_all_clones (gcc::context
*ctxt
)
1396 return new pass_materialize_all_clones (ctxt
);