1 /* Driver of optimization process
2 Copyright (C) 2003-2016 Free Software Foundation, Inc.
3 Contributed by Jan Hubicka
4
5 This file is part of GCC.
6
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
11
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
16
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
20
21 /* This module implements main driver of compilation process.
22
23 The main scope of this file is to act as an interface in between
24 tree based frontends and the backend.
25
26 The front-end is supposed to use following functionality:
27
28 - finalize_function
29
30 This function is called once front-end has parsed whole body of function
31 and it is certain that the function body nor the declaration will change.
32
33 (There is one exception needed for implementing GCC extern inline
34 function.)
35
36 - varpool_finalize_decl
37
38 This function has same behavior as the above but is used for static
39 variables.
40
41 - add_asm_node
42
43 Insert new toplevel ASM statement
44
45 - finalize_compilation_unit
46
47 This function is called once (source level) compilation unit is finalized
48 and it will no longer change.
49
50 The symbol table is constructed starting from the trivially needed
51 symbols finalized by the frontend. Functions are lowered into
52 GIMPLE representation and callgraph/reference lists are constructed.
53 Those are used to discover other necessary functions and variables.
54
55 At the end the bodies of unreachable functions are removed.
56
57 The function can be called multiple times when multiple source level
58 compilation units are combined.
59
60 - compile
61
62 This passes control to the back-end. Optimizations are performed and
63 final assembler is generated. This is done in the following way. Note
64 that with link time optimization the process is split into three
65 stages (compile time, linktime analysis and parallel linktime as
66 indicated bellow).
67
68 Compile time:
69
70 1) Inter-procedural optimization.
71 (ipa_passes)
72
73 This part is further split into:
74
75 a) early optimizations. These are local passes executed in
76 the topological order on the callgraph.
77
78 The purpose of early optimiations is to optimize away simple
79 things that may otherwise confuse IP analysis. Very simple
80 propagation across the callgraph is done i.e. to discover
81 functions without side effects and simple inlining is performed.
82
83 b) early small interprocedural passes.
84
85 Those are interprocedural passes executed only at compilation
86 time. These include, for example, transational memory lowering,
87 unreachable code removal and other simple transformations.
88
89 c) IP analysis stage. All interprocedural passes do their
90 analysis.
91
92 Interprocedural passes differ from small interprocedural
93 passes by their ability to operate across whole program
94 at linktime. Their analysis stage is performed early to
95 both reduce linking times and linktime memory usage by
96 not having to represent whole program in memory.
97
98 d) LTO sreaming. When doing LTO, everything important gets
99 streamed into the object file.
100
101 Compile time and or linktime analysis stage (WPA):
102
103 At linktime units gets streamed back and symbol table is
104 merged. Function bodies are not streamed in and not
105 available.
106 e) IP propagation stage. All IP passes execute their
107 IP propagation. This is done based on the earlier analysis
108 without having function bodies at hand.
109 f) Ltrans streaming. When doing WHOPR LTO, the program
110 is partitioned and streamed into multple object files.
111
112 Compile time and/or parallel linktime stage (ltrans)
113
114 Each of the object files is streamed back and compiled
115 separately. Now the function bodies becomes available
116 again.
117
118 2) Virtual clone materialization
119 (cgraph_materialize_clone)
120
121 IP passes can produce copies of existing functoins (such
122 as versioned clones or inline clones) without actually
123 manipulating their bodies by creating virtual clones in
124 the callgraph. At this time the virtual clones are
125 turned into real functions
126 3) IP transformation
127
128 All IP passes transform function bodies based on earlier
129 decision of the IP propagation.
130
131 4) late small IP passes
132
133 Simple IP passes working within single program partition.
134
135 5) Expansion
136 (expand_all_functions)
137
138 At this stage functions that needs to be output into
139 assembler are identified and compiled in topological order
140 6) Output of variables and aliases
141 Now it is known what variable references was not optimized
142 out and thus all variables are output to the file.
143
144 Note that with -fno-toplevel-reorder passes 5 and 6
145 are combined together in cgraph_output_in_order.
146
147 Finally there are functions to manipulate the callgraph from
148 backend.
149 - cgraph_add_new_function is used to add backend produced
150 functions introduced after the unit is finalized.
151 The functions are enqueue for later processing and inserted
152 into callgraph with cgraph_process_new_functions.
153
154 - cgraph_function_versioning
155
156 produces a copy of function into new one (a version)
157 and apply simple transformations
158 */
159
160 #include "config.h"
161 #include "system.h"
162 #include "coretypes.h"
163 #include "backend.h"
164 #include "target.h"
165 #include "rtl.h"
166 #include "tree.h"
167 #include "gimple.h"
168 #include "cfghooks.h"
169 #include "regset.h" /* FIXME: For reg_obstack. */
170 #include "alloc-pool.h"
171 #include "tree-pass.h"
172 #include "stringpool.h"
173 #include "gimple-ssa.h"
174 #include "cgraph.h"
175 #include "coverage.h"
176 #include "lto-streamer.h"
177 #include "fold-const.h"
178 #include "varasm.h"
179 #include "stor-layout.h"
180 #include "output.h"
181 #include "cfgcleanup.h"
182 #include "gimple-fold.h"
183 #include "gimplify.h"
184 #include "gimple-iterator.h"
185 #include "gimplify-me.h"
186 #include "tree-cfg.h"
187 #include "tree-into-ssa.h"
188 #include "tree-ssa.h"
189 #include "langhooks.h"
190 #include "toplev.h"
191 #include "debug.h"
192 #include "symbol-summary.h"
193 #include "ipa-prop.h"
194 #include "gimple-pretty-print.h"
195 #include "plugin.h"
196 #include "ipa-inline.h"
197 #include "ipa-utils.h"
198 #include "except.h"
199 #include "cfgloop.h"
200 #include "context.h"
201 #include "pass_manager.h"
202 #include "tree-nested.h"
203 #include "dbgcnt.h"
204 #include "tree-chkp.h"
205 #include "lto-section-names.h"
206
207 /* Queue of cgraph nodes scheduled to be added into cgraph. This is a
208 secondary queue used during optimization to accommodate passes that
209 may generate new functions that need to be optimized and expanded. */
210 vec<cgraph_node *> cgraph_new_nodes;
211
212 static void expand_all_functions (void);
213 static void mark_functions_to_output (void);
214 static void handle_alias_pairs (void);
215
216 /* Used for vtable lookup in thunk adjusting. */
217 static GTY (()) tree vtable_entry_type;
218
219 /* Determine if symbol declaration is needed. That is, visible to something
220 either outside this translation unit, something magic in the system
221 configury */
222 bool
needed_p(void)223 symtab_node::needed_p (void)
224 {
225 /* Double check that no one output the function into assembly file
226 early. */
227 gcc_checking_assert (!DECL_ASSEMBLER_NAME_SET_P (decl)
228 || !TREE_SYMBOL_REFERENCED (DECL_ASSEMBLER_NAME (decl)));
229
230 if (!definition)
231 return false;
232
233 if (DECL_EXTERNAL (decl))
234 return false;
235
236 /* If the user told us it is used, then it must be so. */
237 if (force_output)
238 return true;
239
240 /* ABI forced symbols are needed when they are external. */
241 if (forced_by_abi && TREE_PUBLIC (decl))
242 return true;
243
244 /* Keep constructors, destructors and virtual functions. */
245 if (TREE_CODE (decl) == FUNCTION_DECL
246 && (DECL_STATIC_CONSTRUCTOR (decl) || DECL_STATIC_DESTRUCTOR (decl)))
247 return true;
248
249 /* Externally visible variables must be output. The exception is
250 COMDAT variables that must be output only when they are needed. */
251 if (TREE_PUBLIC (decl) && !DECL_COMDAT (decl))
252 return true;
253
254 return false;
255 }
256
257 /* Head and terminator of the queue of nodes to be processed while building
258 callgraph. */
259
260 static symtab_node symtab_terminator;
261 static symtab_node *queued_nodes = &symtab_terminator;
262
263 /* Add NODE to queue starting at QUEUED_NODES.
264 The queue is linked via AUX pointers and terminated by pointer to 1. */
265
266 static void
enqueue_node(symtab_node * node)267 enqueue_node (symtab_node *node)
268 {
269 if (node->aux)
270 return;
271 gcc_checking_assert (queued_nodes);
272 node->aux = queued_nodes;
273 queued_nodes = node;
274 }
275
276 /* Process CGRAPH_NEW_FUNCTIONS and perform actions necessary to add these
277 functions into callgraph in a way so they look like ordinary reachable
278 functions inserted into callgraph already at construction time. */
279
280 void
process_new_functions(void)281 symbol_table::process_new_functions (void)
282 {
283 tree fndecl;
284
285 if (!cgraph_new_nodes.exists ())
286 return;
287
288 handle_alias_pairs ();
289 /* Note that this queue may grow as its being processed, as the new
290 functions may generate new ones. */
291 for (unsigned i = 0; i < cgraph_new_nodes.length (); i++)
292 {
293 cgraph_node *node = cgraph_new_nodes[i];
294 fndecl = node->decl;
295 switch (state)
296 {
297 case CONSTRUCTION:
298 /* At construction time we just need to finalize function and move
299 it into reachable functions list. */
300
301 cgraph_node::finalize_function (fndecl, false);
302 call_cgraph_insertion_hooks (node);
303 enqueue_node (node);
304 break;
305
306 case IPA:
307 case IPA_SSA:
308 case IPA_SSA_AFTER_INLINING:
309 /* When IPA optimization already started, do all essential
310 transformations that has been already performed on the whole
311 cgraph but not on this function. */
312
313 gimple_register_cfg_hooks ();
314 if (!node->analyzed)
315 node->analyze ();
316 push_cfun (DECL_STRUCT_FUNCTION (fndecl));
317 if ((state == IPA_SSA || state == IPA_SSA_AFTER_INLINING)
318 && !gimple_in_ssa_p (DECL_STRUCT_FUNCTION (fndecl)))
319 g->get_passes ()->execute_early_local_passes ();
320 else if (inline_summaries != NULL)
321 compute_inline_parameters (node, true);
322 free_dominance_info (CDI_POST_DOMINATORS);
323 free_dominance_info (CDI_DOMINATORS);
324 pop_cfun ();
325 call_cgraph_insertion_hooks (node);
326 break;
327
328 case EXPANSION:
329 /* Functions created during expansion shall be compiled
330 directly. */
331 node->process = 0;
332 call_cgraph_insertion_hooks (node);
333 node->expand ();
334 break;
335
336 default:
337 gcc_unreachable ();
338 break;
339 }
340 }
341
342 cgraph_new_nodes.release ();
343 }
344
345 /* As an GCC extension we allow redefinition of the function. The
346 semantics when both copies of bodies differ is not well defined.
347 We replace the old body with new body so in unit at a time mode
348 we always use new body, while in normal mode we may end up with
349 old body inlined into some functions and new body expanded and
350 inlined in others.
351
352 ??? It may make more sense to use one body for inlining and other
353 body for expanding the function but this is difficult to do. */
354
355 void
reset(void)356 cgraph_node::reset (void)
357 {
358 /* If process is set, then we have already begun whole-unit analysis.
359 This is *not* testing for whether we've already emitted the function.
360 That case can be sort-of legitimately seen with real function redefinition
361 errors. I would argue that the front end should never present us with
362 such a case, but don't enforce that for now. */
363 gcc_assert (!process);
364
365 /* Reset our data structures so we can analyze the function again. */
366 memset (&local, 0, sizeof (local));
367 memset (&global, 0, sizeof (global));
368 memset (&rtl, 0, sizeof (rtl));
369 analyzed = false;
370 definition = false;
371 alias = false;
372 transparent_alias = false;
373 weakref = false;
374 cpp_implicit_alias = false;
375
376 remove_callees ();
377 remove_all_references ();
378 }
379
380 /* Return true when there are references to the node. INCLUDE_SELF is
381 true if a self reference counts as a reference. */
382
383 bool
referred_to_p(bool include_self)384 symtab_node::referred_to_p (bool include_self)
385 {
386 ipa_ref *ref = NULL;
387
388 /* See if there are any references at all. */
389 if (iterate_referring (0, ref))
390 return true;
391 /* For functions check also calls. */
392 cgraph_node *cn = dyn_cast <cgraph_node *> (this);
393 if (cn && cn->callers)
394 {
395 if (include_self)
396 return true;
397 for (cgraph_edge *e = cn->callers; e; e = e->next_caller)
398 if (e->caller != this)
399 return true;
400 }
401 return false;
402 }
403
404 /* DECL has been parsed. Take it, queue it, compile it at the whim of the
405 logic in effect. If NO_COLLECT is true, then our caller cannot stand to have
406 the garbage collector run at the moment. We would need to either create
407 a new GC context, or just not compile right now. */
408
409 void
finalize_function(tree decl,bool no_collect)410 cgraph_node::finalize_function (tree decl, bool no_collect)
411 {
412 cgraph_node *node = cgraph_node::get_create (decl);
413
414 if (node->definition)
415 {
416 /* Nested functions should only be defined once. */
417 gcc_assert (!DECL_CONTEXT (decl)
418 || TREE_CODE (DECL_CONTEXT (decl)) != FUNCTION_DECL);
419 node->reset ();
420 node->local.redefined_extern_inline = true;
421 }
422
423 /* Set definition first before calling notice_global_symbol so that
424 it is available to notice_global_symbol. */
425 node->definition = true;
426 notice_global_symbol (decl);
427 node->lowered = DECL_STRUCT_FUNCTION (decl)->cfg != NULL;
428
429 /* With -fkeep-inline-functions we are keeping all inline functions except
430 for extern inline ones. */
431 if (flag_keep_inline_functions
432 && DECL_DECLARED_INLINE_P (decl)
433 && !DECL_EXTERNAL (decl)
434 && !DECL_DISREGARD_INLINE_LIMITS (decl))
435 node->force_output = 1;
436
437 /* When not optimizing, also output the static functions. (see
438 PR24561), but don't do so for always_inline functions, functions
439 declared inline and nested functions. These were optimized out
440 in the original implementation and it is unclear whether we want
441 to change the behavior here. */
442 if (((!opt_for_fn (decl, optimize) || flag_keep_static_functions)
443 && !node->cpp_implicit_alias
444 && !DECL_DISREGARD_INLINE_LIMITS (decl)
445 && !DECL_DECLARED_INLINE_P (decl)
446 && !(DECL_CONTEXT (decl)
447 && TREE_CODE (DECL_CONTEXT (decl)) == FUNCTION_DECL))
448 && !DECL_COMDAT (decl) && !DECL_EXTERNAL (decl))
449 node->force_output = 1;
450
451 /* If we've not yet emitted decl, tell the debug info about it. */
452 if (!TREE_ASM_WRITTEN (decl))
453 (*debug_hooks->deferred_inline_function) (decl);
454
455 if (!no_collect)
456 ggc_collect ();
457
458 if (symtab->state == CONSTRUCTION
459 && (node->needed_p () || node->referred_to_p ()))
460 enqueue_node (node);
461 }
462
463 /* Add the function FNDECL to the call graph.
464 Unlike finalize_function, this function is intended to be used
465 by middle end and allows insertion of new function at arbitrary point
466 of compilation. The function can be either in high, low or SSA form
467 GIMPLE.
468
469 The function is assumed to be reachable and have address taken (so no
470 API breaking optimizations are performed on it).
471
472 Main work done by this function is to enqueue the function for later
473 processing to avoid need the passes to be re-entrant. */
474
475 void
add_new_function(tree fndecl,bool lowered)476 cgraph_node::add_new_function (tree fndecl, bool lowered)
477 {
478 gcc::pass_manager *passes = g->get_passes ();
479 cgraph_node *node;
480
481 if (dump_file)
482 {
483 struct function *fn = DECL_STRUCT_FUNCTION (fndecl);
484 const char *function_type = ((gimple_has_body_p (fndecl))
485 ? (lowered
486 ? (gimple_in_ssa_p (fn)
487 ? "ssa gimple"
488 : "low gimple")
489 : "high gimple")
490 : "to-be-gimplified");
491 fprintf (dump_file,
492 "Added new %s function %s to callgraph\n",
493 function_type,
494 fndecl_name (fndecl));
495 }
496
497 switch (symtab->state)
498 {
499 case PARSING:
500 cgraph_node::finalize_function (fndecl, false);
501 break;
502 case CONSTRUCTION:
503 /* Just enqueue function to be processed at nearest occurrence. */
504 node = cgraph_node::get_create (fndecl);
505 if (lowered)
506 node->lowered = true;
507 cgraph_new_nodes.safe_push (node);
508 break;
509
510 case IPA:
511 case IPA_SSA:
512 case IPA_SSA_AFTER_INLINING:
513 case EXPANSION:
514 /* Bring the function into finalized state and enqueue for later
515 analyzing and compilation. */
516 node = cgraph_node::get_create (fndecl);
517 node->local.local = false;
518 node->definition = true;
519 node->force_output = true;
520 if (!lowered && symtab->state == EXPANSION)
521 {
522 push_cfun (DECL_STRUCT_FUNCTION (fndecl));
523 gimple_register_cfg_hooks ();
524 bitmap_obstack_initialize (NULL);
525 execute_pass_list (cfun, passes->all_lowering_passes);
526 passes->execute_early_local_passes ();
527 bitmap_obstack_release (NULL);
528 pop_cfun ();
529
530 lowered = true;
531 }
532 if (lowered)
533 node->lowered = true;
534 cgraph_new_nodes.safe_push (node);
535 break;
536
537 case FINISHED:
538 /* At the very end of compilation we have to do all the work up
539 to expansion. */
540 node = cgraph_node::create (fndecl);
541 if (lowered)
542 node->lowered = true;
543 node->definition = true;
544 node->analyze ();
545 push_cfun (DECL_STRUCT_FUNCTION (fndecl));
546 gimple_register_cfg_hooks ();
547 bitmap_obstack_initialize (NULL);
548 if (!gimple_in_ssa_p (DECL_STRUCT_FUNCTION (fndecl)))
549 g->get_passes ()->execute_early_local_passes ();
550 bitmap_obstack_release (NULL);
551 pop_cfun ();
552 node->expand ();
553 break;
554
555 default:
556 gcc_unreachable ();
557 }
558
559 /* Set a personality if required and we already passed EH lowering. */
560 if (lowered
561 && (function_needs_eh_personality (DECL_STRUCT_FUNCTION (fndecl))
562 == eh_personality_lang))
563 DECL_FUNCTION_PERSONALITY (fndecl) = lang_hooks.eh_personality ();
564 }
565
566 /* Analyze the function scheduled to be output. */
567 void
analyze(void)568 cgraph_node::analyze (void)
569 {
570 tree decl = this->decl;
571 location_t saved_loc = input_location;
572 input_location = DECL_SOURCE_LOCATION (decl);
573
574 if (thunk.thunk_p)
575 {
576 cgraph_node *t = cgraph_node::get (thunk.alias);
577
578 create_edge (t, NULL, 0, CGRAPH_FREQ_BASE);
579 callees->can_throw_external = !TREE_NOTHROW (t->decl);
580 /* Target code in expand_thunk may need the thunk's target
581 to be analyzed, so recurse here. */
582 if (!t->analyzed)
583 t->analyze ();
584 if (t->alias)
585 {
586 t = t->get_alias_target ();
587 if (!t->analyzed)
588 t->analyze ();
589 }
590 if (!expand_thunk (false, false))
591 {
592 thunk.alias = NULL;
593 return;
594 }
595 thunk.alias = NULL;
596 }
597 if (alias)
598 resolve_alias (cgraph_node::get (alias_target), transparent_alias);
599 else if (dispatcher_function)
600 {
601 /* Generate the dispatcher body of multi-versioned functions. */
602 cgraph_function_version_info *dispatcher_version_info
603 = function_version ();
604 if (dispatcher_version_info != NULL
605 && (dispatcher_version_info->dispatcher_resolver
606 == NULL_TREE))
607 {
608 tree resolver = NULL_TREE;
609 gcc_assert (targetm.generate_version_dispatcher_body);
610 resolver = targetm.generate_version_dispatcher_body (this);
611 gcc_assert (resolver != NULL_TREE);
612 }
613 }
614 else
615 {
616 push_cfun (DECL_STRUCT_FUNCTION (decl));
617
618 assign_assembler_name_if_neeeded (decl);
619
620 /* Make sure to gimplify bodies only once. During analyzing a
621 function we lower it, which will require gimplified nested
622 functions, so we can end up here with an already gimplified
623 body. */
624 if (!gimple_has_body_p (decl))
625 gimplify_function_tree (decl);
626
627 /* Lower the function. */
628 if (!lowered)
629 {
630 if (nested)
631 lower_nested_functions (decl);
632 gcc_assert (!nested);
633
634 gimple_register_cfg_hooks ();
635 bitmap_obstack_initialize (NULL);
636 execute_pass_list (cfun, g->get_passes ()->all_lowering_passes);
637 free_dominance_info (CDI_POST_DOMINATORS);
638 free_dominance_info (CDI_DOMINATORS);
639 compact_blocks ();
640 bitmap_obstack_release (NULL);
641 lowered = true;
642 }
643
644 pop_cfun ();
645 }
646 analyzed = true;
647
648 input_location = saved_loc;
649 }
650
651 /* C++ frontend produce same body aliases all over the place, even before PCH
652 gets streamed out. It relies on us linking the aliases with their function
653 in order to do the fixups, but ipa-ref is not PCH safe. Consequentely we
654 first produce aliases without links, but once C++ FE is sure he won't sream
655 PCH we build the links via this function. */
656
657 void
process_same_body_aliases(void)658 symbol_table::process_same_body_aliases (void)
659 {
660 symtab_node *node;
661 FOR_EACH_SYMBOL (node)
662 if (node->cpp_implicit_alias && !node->analyzed)
663 node->resolve_alias
664 (TREE_CODE (node->alias_target) == VAR_DECL
665 ? (symtab_node *)varpool_node::get_create (node->alias_target)
666 : (symtab_node *)cgraph_node::get_create (node->alias_target));
667 cpp_implicit_aliases_done = true;
668 }
669
670 /* Process attributes common for vars and functions. */
671
672 static void
process_common_attributes(symtab_node * node,tree decl)673 process_common_attributes (symtab_node *node, tree decl)
674 {
675 tree weakref = lookup_attribute ("weakref", DECL_ATTRIBUTES (decl));
676
677 if (weakref && !lookup_attribute ("alias", DECL_ATTRIBUTES (decl)))
678 {
679 warning_at (DECL_SOURCE_LOCATION (decl), OPT_Wattributes,
680 "%<weakref%> attribute should be accompanied with"
681 " an %<alias%> attribute");
682 DECL_WEAK (decl) = 0;
683 DECL_ATTRIBUTES (decl) = remove_attribute ("weakref",
684 DECL_ATTRIBUTES (decl));
685 }
686
687 if (lookup_attribute ("no_reorder", DECL_ATTRIBUTES (decl)))
688 node->no_reorder = 1;
689 }
690
691 /* Look for externally_visible and used attributes and mark cgraph nodes
692 accordingly.
693
694 We cannot mark the nodes at the point the attributes are processed (in
695 handle_*_attribute) because the copy of the declarations available at that
696 point may not be canonical. For example, in:
697
698 void f();
699 void f() __attribute__((used));
700
701 the declaration we see in handle_used_attribute will be the second
702 declaration -- but the front end will subsequently merge that declaration
703 with the original declaration and discard the second declaration.
704
705 Furthermore, we can't mark these nodes in finalize_function because:
706
707 void f() {}
708 void f() __attribute__((externally_visible));
709
710 is valid.
711
712 So, we walk the nodes at the end of the translation unit, applying the
713 attributes at that point. */
714
715 static void
process_function_and_variable_attributes(cgraph_node * first,varpool_node * first_var)716 process_function_and_variable_attributes (cgraph_node *first,
717 varpool_node *first_var)
718 {
719 cgraph_node *node;
720 varpool_node *vnode;
721
722 for (node = symtab->first_function (); node != first;
723 node = symtab->next_function (node))
724 {
725 tree decl = node->decl;
726 if (DECL_PRESERVE_P (decl))
727 node->mark_force_output ();
728 else if (lookup_attribute ("externally_visible", DECL_ATTRIBUTES (decl)))
729 {
730 if (! TREE_PUBLIC (node->decl))
731 warning_at (DECL_SOURCE_LOCATION (node->decl), OPT_Wattributes,
732 "%<externally_visible%>"
733 " attribute have effect only on public objects");
734 }
735 if (lookup_attribute ("weakref", DECL_ATTRIBUTES (decl))
736 && (node->definition && !node->alias))
737 {
738 warning_at (DECL_SOURCE_LOCATION (node->decl), OPT_Wattributes,
739 "%<weakref%> attribute ignored"
740 " because function is defined");
741 DECL_WEAK (decl) = 0;
742 DECL_ATTRIBUTES (decl) = remove_attribute ("weakref",
743 DECL_ATTRIBUTES (decl));
744 }
745
746 if (lookup_attribute ("always_inline", DECL_ATTRIBUTES (decl))
747 && !DECL_DECLARED_INLINE_P (decl)
748 /* redefining extern inline function makes it DECL_UNINLINABLE. */
749 && !DECL_UNINLINABLE (decl))
750 warning_at (DECL_SOURCE_LOCATION (decl), OPT_Wattributes,
751 "always_inline function might not be inlinable");
752
753 process_common_attributes (node, decl);
754 }
755 for (vnode = symtab->first_variable (); vnode != first_var;
756 vnode = symtab->next_variable (vnode))
757 {
758 tree decl = vnode->decl;
759 if (DECL_EXTERNAL (decl)
760 && DECL_INITIAL (decl))
761 varpool_node::finalize_decl (decl);
762 if (DECL_PRESERVE_P (decl))
763 vnode->force_output = true;
764 else if (lookup_attribute ("externally_visible", DECL_ATTRIBUTES (decl)))
765 {
766 if (! TREE_PUBLIC (vnode->decl))
767 warning_at (DECL_SOURCE_LOCATION (vnode->decl), OPT_Wattributes,
768 "%<externally_visible%>"
769 " attribute have effect only on public objects");
770 }
771 if (lookup_attribute ("weakref", DECL_ATTRIBUTES (decl))
772 && vnode->definition
773 && DECL_INITIAL (decl))
774 {
775 warning_at (DECL_SOURCE_LOCATION (vnode->decl), OPT_Wattributes,
776 "%<weakref%> attribute ignored"
777 " because variable is initialized");
778 DECL_WEAK (decl) = 0;
779 DECL_ATTRIBUTES (decl) = remove_attribute ("weakref",
780 DECL_ATTRIBUTES (decl));
781 }
782 process_common_attributes (vnode, decl);
783 }
784 }
785
786 /* Mark DECL as finalized. By finalizing the declaration, frontend instruct the
787 middle end to output the variable to asm file, if needed or externally
788 visible. */
789
790 void
finalize_decl(tree decl)791 varpool_node::finalize_decl (tree decl)
792 {
793 varpool_node *node = varpool_node::get_create (decl);
794
795 gcc_assert (TREE_STATIC (decl) || DECL_EXTERNAL (decl));
796
797 if (node->definition)
798 return;
799 /* Set definition first before calling notice_global_symbol so that
800 it is available to notice_global_symbol. */
801 node->definition = true;
802 notice_global_symbol (decl);
803 if (TREE_THIS_VOLATILE (decl) || DECL_PRESERVE_P (decl)
804 /* Traditionally we do not eliminate static variables when not
805 optimizing and when not doing toplevel reoder. */
806 || node->no_reorder
807 || ((!flag_toplevel_reorder
808 && !DECL_COMDAT (node->decl)
809 && !DECL_ARTIFICIAL (node->decl))))
810 node->force_output = true;
811
812 if (symtab->state == CONSTRUCTION
813 && (node->needed_p () || node->referred_to_p ()))
814 enqueue_node (node);
815 if (symtab->state >= IPA_SSA)
816 node->analyze ();
817 /* Some frontends produce various interface variables after compilation
818 finished. */
819 if (symtab->state == FINISHED
820 || (!flag_toplevel_reorder
821 && symtab->state == EXPANSION))
822 node->assemble_decl ();
823
824 if (DECL_INITIAL (decl))
825 chkp_register_var_initializer (decl);
826 }
827
828 /* EDGE is an polymorphic call. Mark all possible targets as reachable
829 and if there is only one target, perform trivial devirtualization.
830 REACHABLE_CALL_TARGETS collects target lists we already walked to
831 avoid udplicate work. */
832
833 static void
walk_polymorphic_call_targets(hash_set<void * > * reachable_call_targets,cgraph_edge * edge)834 walk_polymorphic_call_targets (hash_set<void *> *reachable_call_targets,
835 cgraph_edge *edge)
836 {
837 unsigned int i;
838 void *cache_token;
839 bool final;
840 vec <cgraph_node *>targets
841 = possible_polymorphic_call_targets
842 (edge, &final, &cache_token);
843
844 if (!reachable_call_targets->add (cache_token))
845 {
846 if (symtab->dump_file)
847 dump_possible_polymorphic_call_targets
848 (symtab->dump_file, edge);
849
850 for (i = 0; i < targets.length (); i++)
851 {
852 /* Do not bother to mark virtual methods in anonymous namespace;
853 either we will find use of virtual table defining it, or it is
854 unused. */
855 if (targets[i]->definition
856 && TREE_CODE
857 (TREE_TYPE (targets[i]->decl))
858 == METHOD_TYPE
859 && !type_in_anonymous_namespace_p
860 (TYPE_METHOD_BASETYPE (TREE_TYPE (targets[i]->decl))))
861 enqueue_node (targets[i]);
862 }
863 }
864
865 /* Very trivial devirtualization; when the type is
866 final or anonymous (so we know all its derivation)
867 and there is only one possible virtual call target,
868 make the edge direct. */
869 if (final)
870 {
871 if (targets.length () <= 1 && dbg_cnt (devirt))
872 {
873 cgraph_node *target;
874 if (targets.length () == 1)
875 target = targets[0];
876 else
877 target = cgraph_node::create
878 (builtin_decl_implicit (BUILT_IN_UNREACHABLE));
879
880 if (symtab->dump_file)
881 {
882 fprintf (symtab->dump_file,
883 "Devirtualizing call: ");
884 print_gimple_stmt (symtab->dump_file,
885 edge->call_stmt, 0,
886 TDF_SLIM);
887 }
888 if (dump_enabled_p ())
889 {
890 location_t locus = gimple_location_safe (edge->call_stmt);
891 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, locus,
892 "devirtualizing call in %s to %s\n",
893 edge->caller->name (), target->name ());
894 }
895
896 edge->make_direct (target);
897 edge->redirect_call_stmt_to_callee ();
898
899 /* Call to __builtin_unreachable shouldn't be instrumented. */
900 if (!targets.length ())
901 gimple_call_set_with_bounds (edge->call_stmt, false);
902
903 if (symtab->dump_file)
904 {
905 fprintf (symtab->dump_file,
906 "Devirtualized as: ");
907 print_gimple_stmt (symtab->dump_file,
908 edge->call_stmt, 0,
909 TDF_SLIM);
910 }
911 }
912 }
913 }
914
915 /* Issue appropriate warnings for the global declaration DECL. */
916
917 static void
check_global_declaration(symtab_node * snode)918 check_global_declaration (symtab_node *snode)
919 {
920 const char *decl_file;
921 tree decl = snode->decl;
922
923 /* Warn about any function declared static but not defined. We don't
924 warn about variables, because many programs have static variables
925 that exist only to get some text into the object file. */
926 if (TREE_CODE (decl) == FUNCTION_DECL
927 && DECL_INITIAL (decl) == 0
928 && DECL_EXTERNAL (decl)
929 && ! DECL_ARTIFICIAL (decl)
930 && ! TREE_NO_WARNING (decl)
931 && ! TREE_PUBLIC (decl)
932 && (warn_unused_function
933 || snode->referred_to_p (/*include_self=*/false)))
934 {
935 if (snode->referred_to_p (/*include_self=*/false))
936 pedwarn (input_location, 0, "%q+F used but never defined", decl);
937 else
938 warning (OPT_Wunused_function, "%q+F declared %<static%> but never defined", decl);
939 /* This symbol is effectively an "extern" declaration now. */
940 TREE_PUBLIC (decl) = 1;
941 }
942
943 /* Warn about static fns or vars defined but not used. */
944 if (((warn_unused_function && TREE_CODE (decl) == FUNCTION_DECL)
945 || (((warn_unused_variable && ! TREE_READONLY (decl))
946 || (warn_unused_const_variable > 0 && TREE_READONLY (decl)
947 && (warn_unused_const_variable == 2
948 || (main_input_filename != NULL
949 && (decl_file = DECL_SOURCE_FILE (decl)) != NULL
950 && filename_cmp (main_input_filename,
951 decl_file) == 0))))
952 && TREE_CODE (decl) == VAR_DECL))
953 && ! DECL_IN_SYSTEM_HEADER (decl)
954 && ! snode->referred_to_p (/*include_self=*/false)
955 /* This TREE_USED check is needed in addition to referred_to_p
956 above, because the `__unused__' attribute is not being
957 considered for referred_to_p. */
958 && ! TREE_USED (decl)
959 /* The TREE_USED bit for file-scope decls is kept in the identifier,
960 to handle multiple external decls in different scopes. */
961 && ! (DECL_NAME (decl) && TREE_USED (DECL_NAME (decl)))
962 && ! DECL_EXTERNAL (decl)
963 && ! DECL_ARTIFICIAL (decl)
964 && ! DECL_ABSTRACT_ORIGIN (decl)
965 && ! TREE_PUBLIC (decl)
966 /* A volatile variable might be used in some non-obvious way. */
967 && (! VAR_P (decl) || ! TREE_THIS_VOLATILE (decl))
968 /* Global register variables must be declared to reserve them. */
969 && ! (TREE_CODE (decl) == VAR_DECL && DECL_REGISTER (decl))
970 /* Global ctors and dtors are called by the runtime. */
971 && (TREE_CODE (decl) != FUNCTION_DECL
972 || (!DECL_STATIC_CONSTRUCTOR (decl)
973 && !DECL_STATIC_DESTRUCTOR (decl)))
974 /* Otherwise, ask the language. */
975 && lang_hooks.decls.warn_unused_global (decl))
976 warning_at (DECL_SOURCE_LOCATION (decl),
977 (TREE_CODE (decl) == FUNCTION_DECL)
978 ? OPT_Wunused_function
979 : (TREE_READONLY (decl)
980 ? OPT_Wunused_const_variable_
981 : OPT_Wunused_variable),
982 "%qD defined but not used", decl);
983 }
984
985 /* Discover all functions and variables that are trivially needed, analyze
986 them as well as all functions and variables referred by them */
987 static cgraph_node *first_analyzed;
988 static varpool_node *first_analyzed_var;
989
990 /* FIRST_TIME is set to TRUE for the first time we are called for a
991 translation unit from finalize_compilation_unit() or false
992 otherwise. */
993
994 static void
analyze_functions(bool first_time)995 analyze_functions (bool first_time)
996 {
997 /* Keep track of already processed nodes when called multiple times for
998 intermodule optimization. */
999 cgraph_node *first_handled = first_analyzed;
1000 varpool_node *first_handled_var = first_analyzed_var;
1001 hash_set<void *> reachable_call_targets;
1002
1003 symtab_node *node;
1004 symtab_node *next;
1005 int i;
1006 ipa_ref *ref;
1007 bool changed = true;
1008 location_t saved_loc = input_location;
1009
1010 bitmap_obstack_initialize (NULL);
1011 symtab->state = CONSTRUCTION;
1012 input_location = UNKNOWN_LOCATION;
1013
1014 /* Ugly, but the fixup can not happen at a time same body alias is created;
1015 C++ FE is confused about the COMDAT groups being right. */
1016 if (symtab->cpp_implicit_aliases_done)
1017 FOR_EACH_SYMBOL (node)
1018 if (node->cpp_implicit_alias)
1019 node->fixup_same_cpp_alias_visibility (node->get_alias_target ());
1020 build_type_inheritance_graph ();
1021
1022 /* Analysis adds static variables that in turn adds references to new functions.
1023 So we need to iterate the process until it stabilize. */
1024 while (changed)
1025 {
1026 changed = false;
1027 process_function_and_variable_attributes (first_analyzed,
1028 first_analyzed_var);
1029
1030 /* First identify the trivially needed symbols. */
1031 for (node = symtab->first_symbol ();
1032 node != first_analyzed
1033 && node != first_analyzed_var; node = node->next)
1034 {
1035 /* Convert COMDAT group designators to IDENTIFIER_NODEs. */
1036 node->get_comdat_group_id ();
1037 if (node->needed_p ())
1038 {
1039 enqueue_node (node);
1040 if (!changed && symtab->dump_file)
1041 fprintf (symtab->dump_file, "Trivially needed symbols:");
1042 changed = true;
1043 if (symtab->dump_file)
1044 fprintf (symtab->dump_file, " %s", node->asm_name ());
1045 if (!changed && symtab->dump_file)
1046 fprintf (symtab->dump_file, "\n");
1047 }
1048 if (node == first_analyzed
1049 || node == first_analyzed_var)
1050 break;
1051 }
1052 symtab->process_new_functions ();
1053 first_analyzed_var = symtab->first_variable ();
1054 first_analyzed = symtab->first_function ();
1055
1056 if (changed && symtab->dump_file)
1057 fprintf (symtab->dump_file, "\n");
1058
1059 /* Lower representation, build callgraph edges and references for all trivially
1060 needed symbols and all symbols referred by them. */
1061 while (queued_nodes != &symtab_terminator)
1062 {
1063 changed = true;
1064 node = queued_nodes;
1065 queued_nodes = (symtab_node *)queued_nodes->aux;
1066 cgraph_node *cnode = dyn_cast <cgraph_node *> (node);
1067 if (cnode && cnode->definition)
1068 {
1069 cgraph_edge *edge;
1070 tree decl = cnode->decl;
1071
1072 /* ??? It is possible to create extern inline function
1073 and later using weak alias attribute to kill its body.
1074 See gcc.c-torture/compile/20011119-1.c */
1075 if (!DECL_STRUCT_FUNCTION (decl)
1076 && !cnode->alias
1077 && !cnode->thunk.thunk_p
1078 && !cnode->dispatcher_function)
1079 {
1080 cnode->reset ();
1081 cnode->local.redefined_extern_inline = true;
1082 continue;
1083 }
1084
1085 if (!cnode->analyzed)
1086 cnode->analyze ();
1087
1088 for (edge = cnode->callees; edge; edge = edge->next_callee)
1089 if (edge->callee->definition
1090 && (!DECL_EXTERNAL (edge->callee->decl)
1091 /* When not optimizing, do not try to analyze extern
1092 inline functions. Doing so is pointless. */
1093 || opt_for_fn (edge->callee->decl, optimize)
1094 /* Weakrefs needs to be preserved. */
1095 || edge->callee->alias
1096 /* always_inline functions are inlined aven at -O0. */
1097 || lookup_attribute
1098 ("always_inline",
1099 DECL_ATTRIBUTES (edge->callee->decl))
1100 /* Multiversioned functions needs the dispatcher to
1101 be produced locally even for extern functions. */
1102 || edge->callee->function_version ()))
1103 enqueue_node (edge->callee);
1104 if (opt_for_fn (cnode->decl, optimize)
1105 && opt_for_fn (cnode->decl, flag_devirtualize))
1106 {
1107 cgraph_edge *next;
1108
1109 for (edge = cnode->indirect_calls; edge; edge = next)
1110 {
1111 next = edge->next_callee;
1112 if (edge->indirect_info->polymorphic)
1113 walk_polymorphic_call_targets (&reachable_call_targets,
1114 edge);
1115 }
1116 }
1117
1118 /* If decl is a clone of an abstract function,
1119 mark that abstract function so that we don't release its body.
1120 The DECL_INITIAL() of that abstract function declaration
1121 will be later needed to output debug info. */
1122 if (DECL_ABSTRACT_ORIGIN (decl))
1123 {
1124 cgraph_node *origin_node
1125 = cgraph_node::get_create (DECL_ABSTRACT_ORIGIN (decl));
1126 origin_node->used_as_abstract_origin = true;
1127 }
1128 }
1129 else
1130 {
1131 varpool_node *vnode = dyn_cast <varpool_node *> (node);
1132 if (vnode && vnode->definition && !vnode->analyzed)
1133 vnode->analyze ();
1134 }
1135
1136 if (node->same_comdat_group)
1137 {
1138 symtab_node *next;
1139 for (next = node->same_comdat_group;
1140 next != node;
1141 next = next->same_comdat_group)
1142 if (!next->comdat_local_p ())
1143 enqueue_node (next);
1144 }
1145 for (i = 0; node->iterate_reference (i, ref); i++)
1146 if (ref->referred->definition
1147 && (!DECL_EXTERNAL (ref->referred->decl)
1148 || ((TREE_CODE (ref->referred->decl) != FUNCTION_DECL
1149 && optimize)
1150 || (TREE_CODE (ref->referred->decl) == FUNCTION_DECL
1151 && opt_for_fn (ref->referred->decl, optimize))
1152 || node->alias
1153 || ref->referred->alias)))
1154 enqueue_node (ref->referred);
1155 symtab->process_new_functions ();
1156 }
1157 }
1158 update_type_inheritance_graph ();
1159
1160 /* Collect entry points to the unit. */
1161 if (symtab->dump_file)
1162 {
1163 fprintf (symtab->dump_file, "\n\nInitial ");
1164 symtab_node::dump_table (symtab->dump_file);
1165 }
1166
1167 if (first_time)
1168 {
1169 symtab_node *snode;
1170 FOR_EACH_SYMBOL (snode)
1171 check_global_declaration (snode);
1172 }
1173
1174 if (symtab->dump_file)
1175 fprintf (symtab->dump_file, "\nRemoving unused symbols:");
1176
1177 for (node = symtab->first_symbol ();
1178 node != first_handled
1179 && node != first_handled_var; node = next)
1180 {
1181 next = node->next;
1182 if (!node->aux && !node->referred_to_p ())
1183 {
1184 if (symtab->dump_file)
1185 fprintf (symtab->dump_file, " %s", node->name ());
1186
1187 /* See if the debugger can use anything before the DECL
1188 passes away. Perhaps it can notice a DECL that is now a
1189 constant and can tag the early DIE with an appropriate
1190 attribute.
1191
1192 Otherwise, this is the last chance the debug_hooks have
1193 at looking at optimized away DECLs, since
1194 late_global_decl will subsequently be called from the
1195 contents of the now pruned symbol table. */
1196 if (VAR_P (node->decl)
1197 && !decl_function_context (node->decl))
1198 {
1199 /* We are reclaiming totally unreachable code and variables
1200 so they effectively appear as readonly. Show that to
1201 the debug machinery. */
1202 TREE_READONLY (node->decl) = 1;
1203 node->definition = false;
1204 (*debug_hooks->late_global_decl) (node->decl);
1205 }
1206
1207 node->remove ();
1208 continue;
1209 }
1210 if (cgraph_node *cnode = dyn_cast <cgraph_node *> (node))
1211 {
1212 tree decl = node->decl;
1213
1214 if (cnode->definition && !gimple_has_body_p (decl)
1215 && !cnode->alias
1216 && !cnode->thunk.thunk_p)
1217 cnode->reset ();
1218
1219 gcc_assert (!cnode->definition || cnode->thunk.thunk_p
1220 || cnode->alias
1221 || gimple_has_body_p (decl));
1222 gcc_assert (cnode->analyzed == cnode->definition);
1223 }
1224 node->aux = NULL;
1225 }
1226 for (;node; node = node->next)
1227 node->aux = NULL;
1228 first_analyzed = symtab->first_function ();
1229 first_analyzed_var = symtab->first_variable ();
1230 if (symtab->dump_file)
1231 {
1232 fprintf (symtab->dump_file, "\n\nReclaimed ");
1233 symtab_node::dump_table (symtab->dump_file);
1234 }
1235 bitmap_obstack_release (NULL);
1236 ggc_collect ();
1237 /* Initialize assembler name hash, in particular we want to trigger C++
1238 mangling and same body alias creation before we free DECL_ARGUMENTS
1239 used by it. */
1240 if (!seen_error ())
1241 symtab->symtab_initialize_asm_name_hash ();
1242
1243 input_location = saved_loc;
1244 }
1245
1246 /* Translate the ugly representation of aliases as alias pairs into nice
1247 representation in callgraph. We don't handle all cases yet,
1248 unfortunately. */
1249
1250 static void
handle_alias_pairs(void)1251 handle_alias_pairs (void)
1252 {
1253 alias_pair *p;
1254 unsigned i;
1255
1256 for (i = 0; alias_pairs && alias_pairs->iterate (i, &p);)
1257 {
1258 symtab_node *target_node = symtab_node::get_for_asmname (p->target);
1259
1260 /* Weakrefs with target not defined in current unit are easy to handle:
1261 they behave just as external variables except we need to note the
1262 alias flag to later output the weakref pseudo op into asm file. */
1263 if (!target_node
1264 && lookup_attribute ("weakref", DECL_ATTRIBUTES (p->decl)) != NULL)
1265 {
1266 symtab_node *node = symtab_node::get (p->decl);
1267 if (node)
1268 {
1269 node->alias_target = p->target;
1270 node->weakref = true;
1271 node->alias = true;
1272 node->transparent_alias = true;
1273 }
1274 alias_pairs->unordered_remove (i);
1275 continue;
1276 }
1277 else if (!target_node)
1278 {
1279 error ("%q+D aliased to undefined symbol %qE", p->decl, p->target);
1280 symtab_node *node = symtab_node::get (p->decl);
1281 if (node)
1282 node->alias = false;
1283 alias_pairs->unordered_remove (i);
1284 continue;
1285 }
1286
1287 if (DECL_EXTERNAL (target_node->decl)
1288 /* We use local aliases for C++ thunks to force the tailcall
1289 to bind locally. This is a hack - to keep it working do
1290 the following (which is not strictly correct). */
1291 && (TREE_CODE (target_node->decl) != FUNCTION_DECL
1292 || ! DECL_VIRTUAL_P (target_node->decl))
1293 && ! lookup_attribute ("weakref", DECL_ATTRIBUTES (p->decl)))
1294 {
1295 error ("%q+D aliased to external symbol %qE",
1296 p->decl, p->target);
1297 }
1298
1299 if (TREE_CODE (p->decl) == FUNCTION_DECL
1300 && target_node && is_a <cgraph_node *> (target_node))
1301 {
1302 cgraph_node *src_node = cgraph_node::get (p->decl);
1303 if (src_node && src_node->definition)
1304 src_node->reset ();
1305 cgraph_node::create_alias (p->decl, target_node->decl);
1306 alias_pairs->unordered_remove (i);
1307 }
1308 else if (TREE_CODE (p->decl) == VAR_DECL
1309 && target_node && is_a <varpool_node *> (target_node))
1310 {
1311 varpool_node::create_alias (p->decl, target_node->decl);
1312 alias_pairs->unordered_remove (i);
1313 }
1314 else
1315 {
1316 error ("%q+D alias in between function and variable is not supported",
1317 p->decl);
1318 warning (0, "%q+D aliased declaration",
1319 target_node->decl);
1320 alias_pairs->unordered_remove (i);
1321 }
1322 }
1323 vec_free (alias_pairs);
1324 }
1325
1326
1327 /* Figure out what functions we want to assemble. */
1328
1329 static void
mark_functions_to_output(void)1330 mark_functions_to_output (void)
1331 {
1332 bool check_same_comdat_groups = false;
1333 cgraph_node *node;
1334
1335 if (flag_checking)
1336 FOR_EACH_FUNCTION (node)
1337 gcc_assert (!node->process);
1338
1339 FOR_EACH_FUNCTION (node)
1340 {
1341 tree decl = node->decl;
1342
1343 gcc_assert (!node->process || node->same_comdat_group);
1344 if (node->process)
1345 continue;
1346
1347 /* We need to output all local functions that are used and not
1348 always inlined, as well as those that are reachable from
1349 outside the current compilation unit. */
1350 if (node->analyzed
1351 && !node->thunk.thunk_p
1352 && !node->alias
1353 && !node->global.inlined_to
1354 && !TREE_ASM_WRITTEN (decl)
1355 && !DECL_EXTERNAL (decl))
1356 {
1357 node->process = 1;
1358 if (node->same_comdat_group)
1359 {
1360 cgraph_node *next;
1361 for (next = dyn_cast<cgraph_node *> (node->same_comdat_group);
1362 next != node;
1363 next = dyn_cast<cgraph_node *> (next->same_comdat_group))
1364 if (!next->thunk.thunk_p && !next->alias
1365 && !next->comdat_local_p ())
1366 next->process = 1;
1367 }
1368 }
1369 else if (node->same_comdat_group)
1370 {
1371 if (flag_checking)
1372 check_same_comdat_groups = true;
1373 }
1374 else
1375 {
1376 /* We should've reclaimed all functions that are not needed. */
1377 if (flag_checking
1378 && !node->global.inlined_to
1379 && gimple_has_body_p (decl)
1380 /* FIXME: in ltrans unit when offline copy is outside partition but inline copies
1381 are inside partition, we can end up not removing the body since we no longer
1382 have analyzed node pointing to it. */
1383 && !node->in_other_partition
1384 && !node->alias
1385 && !node->clones
1386 && !DECL_EXTERNAL (decl))
1387 {
1388 node->debug ();
1389 internal_error ("failed to reclaim unneeded function");
1390 }
1391 gcc_assert (node->global.inlined_to
1392 || !gimple_has_body_p (decl)
1393 || node->in_other_partition
1394 || node->clones
1395 || DECL_ARTIFICIAL (decl)
1396 || DECL_EXTERNAL (decl));
1397
1398 }
1399
1400 }
1401 if (flag_checking && check_same_comdat_groups)
1402 FOR_EACH_FUNCTION (node)
1403 if (node->same_comdat_group && !node->process)
1404 {
1405 tree decl = node->decl;
1406 if (!node->global.inlined_to
1407 && gimple_has_body_p (decl)
1408 /* FIXME: in an ltrans unit when the offline copy is outside a
1409 partition but inline copies are inside a partition, we can
1410 end up not removing the body since we no longer have an
1411 analyzed node pointing to it. */
1412 && !node->in_other_partition
1413 && !node->clones
1414 && !DECL_EXTERNAL (decl))
1415 {
1416 node->debug ();
1417 internal_error ("failed to reclaim unneeded function in same "
1418 "comdat group");
1419 }
1420 }
1421 }
1422
1423 /* DECL is FUNCTION_DECL. Initialize datastructures so DECL is a function
1424 in lowered gimple form. IN_SSA is true if the gimple is in SSA.
1425
1426 Set current_function_decl and cfun to newly constructed empty function body.
1427 return basic block in the function body. */
1428
1429 basic_block
init_lowered_empty_function(tree decl,bool in_ssa,gcov_type count)1430 init_lowered_empty_function (tree decl, bool in_ssa, gcov_type count)
1431 {
1432 basic_block bb;
1433 edge e;
1434
1435 current_function_decl = decl;
1436 allocate_struct_function (decl, false);
1437 gimple_register_cfg_hooks ();
1438 init_empty_tree_cfg ();
1439
1440 if (in_ssa)
1441 {
1442 init_tree_ssa (cfun);
1443 init_ssa_operands (cfun);
1444 cfun->gimple_df->in_ssa_p = true;
1445 cfun->curr_properties |= PROP_ssa;
1446 }
1447
1448 DECL_INITIAL (decl) = make_node (BLOCK);
1449
1450 DECL_SAVED_TREE (decl) = error_mark_node;
1451 cfun->curr_properties |= (PROP_gimple_lcf | PROP_gimple_leh | PROP_gimple_any
1452 | PROP_cfg | PROP_loops);
1453
1454 set_loops_for_fn (cfun, ggc_cleared_alloc<loops> ());
1455 init_loops_structure (cfun, loops_for_fn (cfun), 1);
1456 loops_for_fn (cfun)->state |= LOOPS_MAY_HAVE_MULTIPLE_LATCHES;
1457
1458 /* Create BB for body of the function and connect it properly. */
1459 ENTRY_BLOCK_PTR_FOR_FN (cfun)->count = count;
1460 ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency = REG_BR_PROB_BASE;
1461 EXIT_BLOCK_PTR_FOR_FN (cfun)->count = count;
1462 EXIT_BLOCK_PTR_FOR_FN (cfun)->frequency = REG_BR_PROB_BASE;
1463 bb = create_basic_block (NULL, ENTRY_BLOCK_PTR_FOR_FN (cfun));
1464 bb->count = count;
1465 bb->frequency = BB_FREQ_MAX;
1466 e = make_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun), bb, EDGE_FALLTHRU);
1467 e->count = count;
1468 e->probability = REG_BR_PROB_BASE;
1469 e = make_edge (bb, EXIT_BLOCK_PTR_FOR_FN (cfun), 0);
1470 e->count = count;
1471 e->probability = REG_BR_PROB_BASE;
1472 add_bb_to_loop (bb, ENTRY_BLOCK_PTR_FOR_FN (cfun)->loop_father);
1473
1474 return bb;
1475 }
1476
1477 /* Adjust PTR by the constant FIXED_OFFSET, and by the vtable
1478 offset indicated by VIRTUAL_OFFSET, if that is
1479 non-null. THIS_ADJUSTING is nonzero for a this adjusting thunk and
1480 zero for a result adjusting thunk. */
1481
1482 static tree
thunk_adjust(gimple_stmt_iterator * bsi,tree ptr,bool this_adjusting,HOST_WIDE_INT fixed_offset,tree virtual_offset)1483 thunk_adjust (gimple_stmt_iterator * bsi,
1484 tree ptr, bool this_adjusting,
1485 HOST_WIDE_INT fixed_offset, tree virtual_offset)
1486 {
1487 gassign *stmt;
1488 tree ret;
1489
1490 if (this_adjusting
1491 && fixed_offset != 0)
1492 {
1493 stmt = gimple_build_assign
1494 (ptr, fold_build_pointer_plus_hwi_loc (input_location,
1495 ptr,
1496 fixed_offset));
1497 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1498 }
1499
1500 /* If there's a virtual offset, look up that value in the vtable and
1501 adjust the pointer again. */
1502 if (virtual_offset)
1503 {
1504 tree vtabletmp;
1505 tree vtabletmp2;
1506 tree vtabletmp3;
1507
1508 if (!vtable_entry_type)
1509 {
1510 tree vfunc_type = make_node (FUNCTION_TYPE);
1511 TREE_TYPE (vfunc_type) = integer_type_node;
1512 TYPE_ARG_TYPES (vfunc_type) = NULL_TREE;
1513 layout_type (vfunc_type);
1514
1515 vtable_entry_type = build_pointer_type (vfunc_type);
1516 }
1517
1518 vtabletmp =
1519 create_tmp_reg (build_pointer_type
1520 (build_pointer_type (vtable_entry_type)), "vptr");
1521
1522 /* The vptr is always at offset zero in the object. */
1523 stmt = gimple_build_assign (vtabletmp,
1524 build1 (NOP_EXPR, TREE_TYPE (vtabletmp),
1525 ptr));
1526 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1527
1528 /* Form the vtable address. */
1529 vtabletmp2 = create_tmp_reg (TREE_TYPE (TREE_TYPE (vtabletmp)),
1530 "vtableaddr");
1531 stmt = gimple_build_assign (vtabletmp2,
1532 build_simple_mem_ref (vtabletmp));
1533 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1534
1535 /* Find the entry with the vcall offset. */
1536 stmt = gimple_build_assign (vtabletmp2,
1537 fold_build_pointer_plus_loc (input_location,
1538 vtabletmp2,
1539 virtual_offset));
1540 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1541
1542 /* Get the offset itself. */
1543 vtabletmp3 = create_tmp_reg (TREE_TYPE (TREE_TYPE (vtabletmp2)),
1544 "vcalloffset");
1545 stmt = gimple_build_assign (vtabletmp3,
1546 build_simple_mem_ref (vtabletmp2));
1547 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1548
1549 /* Adjust the `this' pointer. */
1550 ptr = fold_build_pointer_plus_loc (input_location, ptr, vtabletmp3);
1551 ptr = force_gimple_operand_gsi (bsi, ptr, true, NULL_TREE, false,
1552 GSI_CONTINUE_LINKING);
1553 }
1554
1555 if (!this_adjusting
1556 && fixed_offset != 0)
1557 /* Adjust the pointer by the constant. */
1558 {
1559 tree ptrtmp;
1560
1561 if (TREE_CODE (ptr) == VAR_DECL)
1562 ptrtmp = ptr;
1563 else
1564 {
1565 ptrtmp = create_tmp_reg (TREE_TYPE (ptr), "ptr");
1566 stmt = gimple_build_assign (ptrtmp, ptr);
1567 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1568 }
1569 ptr = fold_build_pointer_plus_hwi_loc (input_location,
1570 ptrtmp, fixed_offset);
1571 }
1572
1573 /* Emit the statement and gimplify the adjustment expression. */
1574 ret = create_tmp_reg (TREE_TYPE (ptr), "adjusted_this");
1575 stmt = gimple_build_assign (ret, ptr);
1576 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1577
1578 return ret;
1579 }
1580
1581 /* Expand thunk NODE to gimple if possible.
1582 When FORCE_GIMPLE_THUNK is true, gimple thunk is created and
1583 no assembler is produced.
1584 When OUTPUT_ASM_THUNK is true, also produce assembler for
1585 thunks that are not lowered. */
1586
1587 bool
expand_thunk(bool output_asm_thunks,bool force_gimple_thunk)1588 cgraph_node::expand_thunk (bool output_asm_thunks, bool force_gimple_thunk)
1589 {
1590 bool this_adjusting = thunk.this_adjusting;
1591 HOST_WIDE_INT fixed_offset = thunk.fixed_offset;
1592 HOST_WIDE_INT virtual_value = thunk.virtual_value;
1593 tree virtual_offset = NULL;
1594 tree alias = callees->callee->decl;
1595 tree thunk_fndecl = decl;
1596 tree a;
1597
1598 /* Instrumentation thunk is the same function with
1599 a different signature. Never need to expand it. */
1600 if (thunk.add_pointer_bounds_args)
1601 return false;
1602
1603 if (!force_gimple_thunk && this_adjusting
1604 && targetm.asm_out.can_output_mi_thunk (thunk_fndecl, fixed_offset,
1605 virtual_value, alias))
1606 {
1607 const char *fnname;
1608 tree fn_block;
1609 tree restype = TREE_TYPE (TREE_TYPE (thunk_fndecl));
1610
1611 if (!output_asm_thunks)
1612 {
1613 analyzed = true;
1614 return false;
1615 }
1616
1617 if (in_lto_p)
1618 get_untransformed_body ();
1619 a = DECL_ARGUMENTS (thunk_fndecl);
1620
1621 current_function_decl = thunk_fndecl;
1622
1623 /* Ensure thunks are emitted in their correct sections. */
1624 resolve_unique_section (thunk_fndecl, 0,
1625 flag_function_sections);
1626
1627 DECL_RESULT (thunk_fndecl)
1628 = build_decl (DECL_SOURCE_LOCATION (thunk_fndecl),
1629 RESULT_DECL, 0, restype);
1630 DECL_CONTEXT (DECL_RESULT (thunk_fndecl)) = thunk_fndecl;
1631 fnname = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (thunk_fndecl));
1632
1633 /* The back end expects DECL_INITIAL to contain a BLOCK, so we
1634 create one. */
1635 fn_block = make_node (BLOCK);
1636 BLOCK_VARS (fn_block) = a;
1637 DECL_INITIAL (thunk_fndecl) = fn_block;
1638 allocate_struct_function (thunk_fndecl, false);
1639 init_function_start (thunk_fndecl);
1640 cfun->is_thunk = 1;
1641 insn_locations_init ();
1642 set_curr_insn_location (DECL_SOURCE_LOCATION (thunk_fndecl));
1643 prologue_location = curr_insn_location ();
1644 assemble_start_function (thunk_fndecl, fnname);
1645
1646 targetm.asm_out.output_mi_thunk (asm_out_file, thunk_fndecl,
1647 fixed_offset, virtual_value, alias);
1648
1649 assemble_end_function (thunk_fndecl, fnname);
1650 insn_locations_finalize ();
1651 init_insn_lengths ();
1652 free_after_compilation (cfun);
1653 TREE_ASM_WRITTEN (thunk_fndecl) = 1;
1654 thunk.thunk_p = false;
1655 analyzed = false;
1656 }
1657 else if (stdarg_p (TREE_TYPE (thunk_fndecl)))
1658 {
1659 error ("generic thunk code fails for method %qD which uses %<...%>",
1660 thunk_fndecl);
1661 TREE_ASM_WRITTEN (thunk_fndecl) = 1;
1662 analyzed = true;
1663 return false;
1664 }
1665 else
1666 {
1667 tree restype;
1668 basic_block bb, then_bb, else_bb, return_bb;
1669 gimple_stmt_iterator bsi;
1670 int nargs = 0;
1671 tree arg;
1672 int i;
1673 tree resdecl;
1674 tree restmp = NULL;
1675 tree resbnd = NULL;
1676
1677 gcall *call;
1678 greturn *ret;
1679 bool alias_is_noreturn = TREE_THIS_VOLATILE (alias);
1680
1681 /* We may be called from expand_thunk that releses body except for
1682 DECL_ARGUMENTS. In this case force_gimple_thunk is true. */
1683 if (in_lto_p && !force_gimple_thunk)
1684 get_untransformed_body ();
1685 a = DECL_ARGUMENTS (thunk_fndecl);
1686
1687 current_function_decl = thunk_fndecl;
1688
1689 /* Ensure thunks are emitted in their correct sections. */
1690 resolve_unique_section (thunk_fndecl, 0,
1691 flag_function_sections);
1692
1693 DECL_IGNORED_P (thunk_fndecl) = 1;
1694 bitmap_obstack_initialize (NULL);
1695
1696 if (thunk.virtual_offset_p)
1697 virtual_offset = size_int (virtual_value);
1698
1699 /* Build the return declaration for the function. */
1700 restype = TREE_TYPE (TREE_TYPE (thunk_fndecl));
1701 if (DECL_RESULT (thunk_fndecl) == NULL_TREE)
1702 {
1703 resdecl = build_decl (input_location, RESULT_DECL, 0, restype);
1704 DECL_ARTIFICIAL (resdecl) = 1;
1705 DECL_IGNORED_P (resdecl) = 1;
1706 DECL_RESULT (thunk_fndecl) = resdecl;
1707 DECL_CONTEXT (DECL_RESULT (thunk_fndecl)) = thunk_fndecl;
1708 }
1709 else
1710 resdecl = DECL_RESULT (thunk_fndecl);
1711
1712 bb = then_bb = else_bb = return_bb
1713 = init_lowered_empty_function (thunk_fndecl, true, count);
1714
1715 bsi = gsi_start_bb (bb);
1716
1717 /* Build call to the function being thunked. */
1718 if (!VOID_TYPE_P (restype)
1719 && (!alias_is_noreturn
1720 || TREE_ADDRESSABLE (restype)
1721 || TREE_CODE (TYPE_SIZE_UNIT (restype)) != INTEGER_CST))
1722 {
1723 if (DECL_BY_REFERENCE (resdecl))
1724 {
1725 restmp = gimple_fold_indirect_ref (resdecl);
1726 if (!restmp)
1727 restmp = build2 (MEM_REF,
1728 TREE_TYPE (TREE_TYPE (DECL_RESULT (alias))),
1729 resdecl,
1730 build_int_cst (TREE_TYPE
1731 (DECL_RESULT (alias)), 0));
1732 }
1733 else if (!is_gimple_reg_type (restype))
1734 {
1735 if (aggregate_value_p (resdecl, TREE_TYPE (thunk_fndecl)))
1736 {
1737 restmp = resdecl;
1738
1739 if (TREE_CODE (restmp) == VAR_DECL)
1740 add_local_decl (cfun, restmp);
1741 BLOCK_VARS (DECL_INITIAL (current_function_decl)) = restmp;
1742 }
1743 else
1744 restmp = create_tmp_var (restype, "retval");
1745 }
1746 else
1747 restmp = create_tmp_reg (restype, "retval");
1748 }
1749
1750 for (arg = a; arg; arg = DECL_CHAIN (arg))
1751 nargs++;
1752 auto_vec<tree> vargs (nargs);
1753 i = 0;
1754 arg = a;
1755 if (this_adjusting)
1756 {
1757 vargs.quick_push (thunk_adjust (&bsi, a, 1, fixed_offset,
1758 virtual_offset));
1759 arg = DECL_CHAIN (a);
1760 i = 1;
1761 }
1762
1763 if (nargs)
1764 for (; i < nargs; i++, arg = DECL_CHAIN (arg))
1765 {
1766 tree tmp = arg;
1767 if (VECTOR_TYPE_P (TREE_TYPE (arg))
1768 || TREE_CODE (TREE_TYPE (arg)) == COMPLEX_TYPE)
1769 DECL_GIMPLE_REG_P (arg) = 1;
1770
1771 if (!is_gimple_val (arg))
1772 {
1773 tmp = create_tmp_reg (TYPE_MAIN_VARIANT
1774 (TREE_TYPE (arg)), "arg");
1775 gimple *stmt = gimple_build_assign (tmp, arg);
1776 gsi_insert_after (&bsi, stmt, GSI_NEW_STMT);
1777 }
1778 vargs.quick_push (tmp);
1779 }
1780 call = gimple_build_call_vec (build_fold_addr_expr_loc (0, alias), vargs);
1781 callees->call_stmt = call;
1782 gimple_call_set_from_thunk (call, true);
1783 gimple_call_set_with_bounds (call, instrumentation_clone);
1784
1785 /* Return slot optimization is always possible and in fact requred to
1786 return values with DECL_BY_REFERENCE. */
1787 if (aggregate_value_p (resdecl, TREE_TYPE (thunk_fndecl))
1788 && (!is_gimple_reg_type (TREE_TYPE (resdecl))
1789 || DECL_BY_REFERENCE (resdecl)))
1790 gimple_call_set_return_slot_opt (call, true);
1791
1792 if (restmp)
1793 {
1794 gimple_call_set_lhs (call, restmp);
1795 gcc_assert (useless_type_conversion_p (TREE_TYPE (restmp),
1796 TREE_TYPE (TREE_TYPE (alias))));
1797 }
1798 gsi_insert_after (&bsi, call, GSI_NEW_STMT);
1799 if (!alias_is_noreturn)
1800 {
1801 if (instrumentation_clone
1802 && !DECL_BY_REFERENCE (resdecl)
1803 && restmp
1804 && BOUNDED_P (restmp))
1805 {
1806 resbnd = chkp_insert_retbnd_call (NULL, restmp, &bsi);
1807 create_edge (get_create (gimple_call_fndecl (gsi_stmt (bsi))),
1808 as_a <gcall *> (gsi_stmt (bsi)),
1809 callees->count, callees->frequency);
1810 }
1811
1812 if (restmp && !this_adjusting
1813 && (fixed_offset || virtual_offset))
1814 {
1815 tree true_label = NULL_TREE;
1816
1817 if (TREE_CODE (TREE_TYPE (restmp)) == POINTER_TYPE)
1818 {
1819 gimple *stmt;
1820 edge e;
1821 /* If the return type is a pointer, we need to
1822 protect against NULL. We know there will be an
1823 adjustment, because that's why we're emitting a
1824 thunk. */
1825 then_bb = create_basic_block (NULL, bb);
1826 then_bb->count = count - count / 16;
1827 then_bb->frequency = BB_FREQ_MAX - BB_FREQ_MAX / 16;
1828 return_bb = create_basic_block (NULL, then_bb);
1829 return_bb->count = count;
1830 return_bb->frequency = BB_FREQ_MAX;
1831 else_bb = create_basic_block (NULL, else_bb);
1832 then_bb->count = count / 16;
1833 then_bb->frequency = BB_FREQ_MAX / 16;
1834 add_bb_to_loop (then_bb, bb->loop_father);
1835 add_bb_to_loop (return_bb, bb->loop_father);
1836 add_bb_to_loop (else_bb, bb->loop_father);
1837 remove_edge (single_succ_edge (bb));
1838 true_label = gimple_block_label (then_bb);
1839 stmt = gimple_build_cond (NE_EXPR, restmp,
1840 build_zero_cst (TREE_TYPE (restmp)),
1841 NULL_TREE, NULL_TREE);
1842 gsi_insert_after (&bsi, stmt, GSI_NEW_STMT);
1843 e = make_edge (bb, then_bb, EDGE_TRUE_VALUE);
1844 e->probability = REG_BR_PROB_BASE - REG_BR_PROB_BASE / 16;
1845 e->count = count - count / 16;
1846 e = make_edge (bb, else_bb, EDGE_FALSE_VALUE);
1847 e->probability = REG_BR_PROB_BASE / 16;
1848 e->count = count / 16;
1849 e = make_edge (return_bb, EXIT_BLOCK_PTR_FOR_FN (cfun), 0);
1850 e->probability = REG_BR_PROB_BASE;
1851 e->count = count;
1852 e = make_edge (then_bb, return_bb, EDGE_FALLTHRU);
1853 e->probability = REG_BR_PROB_BASE;
1854 e->count = count - count / 16;
1855 e = make_edge (else_bb, return_bb, EDGE_FALLTHRU);
1856 e->probability = REG_BR_PROB_BASE;
1857 e->count = count / 16;
1858 bsi = gsi_last_bb (then_bb);
1859 }
1860
1861 restmp = thunk_adjust (&bsi, restmp, /*this_adjusting=*/0,
1862 fixed_offset, virtual_offset);
1863 if (true_label)
1864 {
1865 gimple *stmt;
1866 bsi = gsi_last_bb (else_bb);
1867 stmt = gimple_build_assign (restmp,
1868 build_zero_cst (TREE_TYPE (restmp)));
1869 gsi_insert_after (&bsi, stmt, GSI_NEW_STMT);
1870 bsi = gsi_last_bb (return_bb);
1871 }
1872 }
1873 else
1874 gimple_call_set_tail (call, true);
1875
1876 /* Build return value. */
1877 if (!DECL_BY_REFERENCE (resdecl))
1878 ret = gimple_build_return (restmp);
1879 else
1880 ret = gimple_build_return (resdecl);
1881 gimple_return_set_retbnd (ret, resbnd);
1882
1883 gsi_insert_after (&bsi, ret, GSI_NEW_STMT);
1884 }
1885 else
1886 {
1887 gimple_call_set_tail (call, true);
1888 remove_edge (single_succ_edge (bb));
1889 }
1890
1891 cfun->gimple_df->in_ssa_p = true;
1892 profile_status_for_fn (cfun)
1893 = count ? PROFILE_READ : PROFILE_GUESSED;
1894 /* FIXME: C++ FE should stop setting TREE_ASM_WRITTEN on thunks. */
1895 TREE_ASM_WRITTEN (thunk_fndecl) = false;
1896 delete_unreachable_blocks ();
1897 update_ssa (TODO_update_ssa);
1898 checking_verify_flow_info ();
1899 free_dominance_info (CDI_DOMINATORS);
1900
1901 /* Since we want to emit the thunk, we explicitly mark its name as
1902 referenced. */
1903 thunk.thunk_p = false;
1904 lowered = true;
1905 bitmap_obstack_release (NULL);
1906 }
1907 current_function_decl = NULL;
1908 set_cfun (NULL);
1909 return true;
1910 }
1911
1912 /* Assemble thunks and aliases associated to node. */
1913
1914 void
assemble_thunks_and_aliases(void)1915 cgraph_node::assemble_thunks_and_aliases (void)
1916 {
1917 cgraph_edge *e;
1918 ipa_ref *ref;
1919
1920 for (e = callers; e;)
1921 if (e->caller->thunk.thunk_p
1922 && !e->caller->thunk.add_pointer_bounds_args)
1923 {
1924 cgraph_node *thunk = e->caller;
1925
1926 e = e->next_caller;
1927 thunk->expand_thunk (true, false);
1928 thunk->assemble_thunks_and_aliases ();
1929 }
1930 else
1931 e = e->next_caller;
1932
1933 FOR_EACH_ALIAS (this, ref)
1934 {
1935 cgraph_node *alias = dyn_cast <cgraph_node *> (ref->referring);
1936 if (!alias->transparent_alias)
1937 {
1938 bool saved_written = TREE_ASM_WRITTEN (decl);
1939
1940 /* Force assemble_alias to really output the alias this time instead
1941 of buffering it in same alias pairs. */
1942 TREE_ASM_WRITTEN (decl) = 1;
1943 do_assemble_alias (alias->decl,
1944 DECL_ASSEMBLER_NAME (decl));
1945 alias->assemble_thunks_and_aliases ();
1946 TREE_ASM_WRITTEN (decl) = saved_written;
1947 }
1948 }
1949 }
1950
1951 /* Expand function specified by node. */
1952
1953 void
expand(void)1954 cgraph_node::expand (void)
1955 {
1956 location_t saved_loc;
1957
1958 /* We ought to not compile any inline clones. */
1959 gcc_assert (!global.inlined_to);
1960
1961 announce_function (decl);
1962 process = 0;
1963 gcc_assert (lowered);
1964 get_untransformed_body ();
1965
1966 /* Generate RTL for the body of DECL. */
1967
1968 timevar_push (TV_REST_OF_COMPILATION);
1969
1970 gcc_assert (symtab->global_info_ready);
1971
1972 /* Initialize the default bitmap obstack. */
1973 bitmap_obstack_initialize (NULL);
1974
1975 /* Initialize the RTL code for the function. */
1976 saved_loc = input_location;
1977 input_location = DECL_SOURCE_LOCATION (decl);
1978
1979 gcc_assert (DECL_STRUCT_FUNCTION (decl));
1980 push_cfun (DECL_STRUCT_FUNCTION (decl));
1981 init_function_start (decl);
1982
1983 gimple_register_cfg_hooks ();
1984
1985 bitmap_obstack_initialize (®_obstack); /* FIXME, only at RTL generation*/
1986
1987 execute_all_ipa_transforms ();
1988
1989 /* Perform all tree transforms and optimizations. */
1990
1991 /* Signal the start of passes. */
1992 invoke_plugin_callbacks (PLUGIN_ALL_PASSES_START, NULL);
1993
1994 execute_pass_list (cfun, g->get_passes ()->all_passes);
1995
1996 /* Signal the end of passes. */
1997 invoke_plugin_callbacks (PLUGIN_ALL_PASSES_END, NULL);
1998
1999 bitmap_obstack_release (®_obstack);
2000
2001 /* Release the default bitmap obstack. */
2002 bitmap_obstack_release (NULL);
2003
2004 /* If requested, warn about function definitions where the function will
2005 return a value (usually of some struct or union type) which itself will
2006 take up a lot of stack space. */
2007 if (warn_larger_than && !DECL_EXTERNAL (decl) && TREE_TYPE (decl))
2008 {
2009 tree ret_type = TREE_TYPE (TREE_TYPE (decl));
2010
2011 if (ret_type && TYPE_SIZE_UNIT (ret_type)
2012 && TREE_CODE (TYPE_SIZE_UNIT (ret_type)) == INTEGER_CST
2013 && 0 < compare_tree_int (TYPE_SIZE_UNIT (ret_type),
2014 larger_than_size))
2015 {
2016 unsigned int size_as_int
2017 = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (ret_type));
2018
2019 if (compare_tree_int (TYPE_SIZE_UNIT (ret_type), size_as_int) == 0)
2020 warning (OPT_Wlarger_than_, "size of return value of %q+D is %u bytes",
2021 decl, size_as_int);
2022 else
2023 warning (OPT_Wlarger_than_, "size of return value of %q+D is larger than %wd bytes",
2024 decl, larger_than_size);
2025 }
2026 }
2027
2028 gimple_set_body (decl, NULL);
2029 if (DECL_STRUCT_FUNCTION (decl) == 0
2030 && !cgraph_node::get (decl)->origin)
2031 {
2032 /* Stop pointing to the local nodes about to be freed.
2033 But DECL_INITIAL must remain nonzero so we know this
2034 was an actual function definition.
2035 For a nested function, this is done in c_pop_function_context.
2036 If rest_of_compilation set this to 0, leave it 0. */
2037 if (DECL_INITIAL (decl) != 0)
2038 DECL_INITIAL (decl) = error_mark_node;
2039 }
2040
2041 input_location = saved_loc;
2042
2043 ggc_collect ();
2044 timevar_pop (TV_REST_OF_COMPILATION);
2045
2046 /* Make sure that BE didn't give up on compiling. */
2047 gcc_assert (TREE_ASM_WRITTEN (decl));
2048 if (cfun)
2049 pop_cfun ();
2050
2051 /* It would make a lot more sense to output thunks before function body to get more
2052 forward and lest backwarding jumps. This however would need solving problem
2053 with comdats. See PR48668. Also aliases must come after function itself to
2054 make one pass assemblers, like one on AIX, happy. See PR 50689.
2055 FIXME: Perhaps thunks should be move before function IFF they are not in comdat
2056 groups. */
2057 assemble_thunks_and_aliases ();
2058 release_body ();
2059 /* Eliminate all call edges. This is important so the GIMPLE_CALL no longer
2060 points to the dead function body. */
2061 remove_callees ();
2062 remove_all_references ();
2063 }
2064
2065 /* Node comparer that is responsible for the order that corresponds
2066 to time when a function was launched for the first time. */
2067
2068 static int
node_cmp(const void * pa,const void * pb)2069 node_cmp (const void *pa, const void *pb)
2070 {
2071 const cgraph_node *a = *(const cgraph_node * const *) pa;
2072 const cgraph_node *b = *(const cgraph_node * const *) pb;
2073
2074 /* Functions with time profile must be before these without profile. */
2075 if (!a->tp_first_run || !b->tp_first_run)
2076 return a->tp_first_run - b->tp_first_run;
2077
2078 return a->tp_first_run != b->tp_first_run
2079 ? b->tp_first_run - a->tp_first_run
2080 : b->order - a->order;
2081 }
2082
2083 /* Expand all functions that must be output.
2084
2085 Attempt to topologically sort the nodes so function is output when
2086 all called functions are already assembled to allow data to be
2087 propagated across the callgraph. Use a stack to get smaller distance
2088 between a function and its callees (later we may choose to use a more
2089 sophisticated algorithm for function reordering; we will likely want
2090 to use subsections to make the output functions appear in top-down
2091 order). */
2092
2093 static void
expand_all_functions(void)2094 expand_all_functions (void)
2095 {
2096 cgraph_node *node;
2097 cgraph_node **order = XCNEWVEC (cgraph_node *,
2098 symtab->cgraph_count);
2099 unsigned int expanded_func_count = 0, profiled_func_count = 0;
2100 int order_pos, new_order_pos = 0;
2101 int i;
2102
2103 order_pos = ipa_reverse_postorder (order);
2104 gcc_assert (order_pos == symtab->cgraph_count);
2105
2106 /* Garbage collector may remove inline clones we eliminate during
2107 optimization. So we must be sure to not reference them. */
2108 for (i = 0; i < order_pos; i++)
2109 if (order[i]->process)
2110 order[new_order_pos++] = order[i];
2111
2112 if (flag_profile_reorder_functions)
2113 qsort (order, new_order_pos, sizeof (cgraph_node *), node_cmp);
2114
2115 for (i = new_order_pos - 1; i >= 0; i--)
2116 {
2117 node = order[i];
2118
2119 if (node->process)
2120 {
2121 expanded_func_count++;
2122 if(node->tp_first_run)
2123 profiled_func_count++;
2124
2125 if (symtab->dump_file)
2126 fprintf (symtab->dump_file,
2127 "Time profile order in expand_all_functions:%s:%d\n",
2128 node->asm_name (), node->tp_first_run);
2129 node->process = 0;
2130 node->expand ();
2131 }
2132 }
2133
2134 if (dump_file)
2135 fprintf (dump_file, "Expanded functions with time profile (%s):%u/%u\n",
2136 main_input_filename, profiled_func_count, expanded_func_count);
2137
2138 if (symtab->dump_file && flag_profile_reorder_functions)
2139 fprintf (symtab->dump_file, "Expanded functions with time profile:%u/%u\n",
2140 profiled_func_count, expanded_func_count);
2141
2142 symtab->process_new_functions ();
2143 free_gimplify_stack ();
2144
2145 free (order);
2146 }
2147
2148 /* This is used to sort the node types by the cgraph order number. */
2149
2150 enum cgraph_order_sort_kind
2151 {
2152 ORDER_UNDEFINED = 0,
2153 ORDER_FUNCTION,
2154 ORDER_VAR,
2155 ORDER_ASM
2156 };
2157
2158 struct cgraph_order_sort
2159 {
2160 enum cgraph_order_sort_kind kind;
2161 union
2162 {
2163 cgraph_node *f;
2164 varpool_node *v;
2165 asm_node *a;
2166 } u;
2167 };
2168
2169 /* Output all functions, variables, and asm statements in the order
2170 according to their order fields, which is the order in which they
2171 appeared in the file. This implements -fno-toplevel-reorder. In
2172 this mode we may output functions and variables which don't really
2173 need to be output.
2174 When NO_REORDER is true only do this for symbols marked no reorder. */
2175
2176 static void
output_in_order(bool no_reorder)2177 output_in_order (bool no_reorder)
2178 {
2179 int max;
2180 cgraph_order_sort *nodes;
2181 int i;
2182 cgraph_node *pf;
2183 varpool_node *pv;
2184 asm_node *pa;
2185 max = symtab->order;
2186 nodes = XCNEWVEC (cgraph_order_sort, max);
2187
2188 FOR_EACH_DEFINED_FUNCTION (pf)
2189 {
2190 if (pf->process && !pf->thunk.thunk_p && !pf->alias)
2191 {
2192 if (no_reorder && !pf->no_reorder)
2193 continue;
2194 i = pf->order;
2195 gcc_assert (nodes[i].kind == ORDER_UNDEFINED);
2196 nodes[i].kind = ORDER_FUNCTION;
2197 nodes[i].u.f = pf;
2198 }
2199 }
2200
2201 FOR_EACH_DEFINED_VARIABLE (pv)
2202 if (!DECL_EXTERNAL (pv->decl))
2203 {
2204 if (no_reorder && !pv->no_reorder)
2205 continue;
2206 i = pv->order;
2207 gcc_assert (nodes[i].kind == ORDER_UNDEFINED);
2208 nodes[i].kind = ORDER_VAR;
2209 nodes[i].u.v = pv;
2210 }
2211
2212 for (pa = symtab->first_asm_symbol (); pa; pa = pa->next)
2213 {
2214 i = pa->order;
2215 gcc_assert (nodes[i].kind == ORDER_UNDEFINED);
2216 nodes[i].kind = ORDER_ASM;
2217 nodes[i].u.a = pa;
2218 }
2219
2220 /* In toplevel reorder mode we output all statics; mark them as needed. */
2221
2222 for (i = 0; i < max; ++i)
2223 if (nodes[i].kind == ORDER_VAR)
2224 nodes[i].u.v->finalize_named_section_flags ();
2225
2226 for (i = 0; i < max; ++i)
2227 {
2228 switch (nodes[i].kind)
2229 {
2230 case ORDER_FUNCTION:
2231 nodes[i].u.f->process = 0;
2232 nodes[i].u.f->expand ();
2233 break;
2234
2235 case ORDER_VAR:
2236 #ifdef ACCEL_COMPILER
2237 /* Do not assemble "omp declare target link" vars. */
2238 if (DECL_HAS_VALUE_EXPR_P (nodes[i].u.v->decl)
2239 && lookup_attribute ("omp declare target link",
2240 DECL_ATTRIBUTES (nodes[i].u.v->decl)))
2241 break;
2242 #endif
2243 nodes[i].u.v->assemble_decl ();
2244 break;
2245
2246 case ORDER_ASM:
2247 assemble_asm (nodes[i].u.a->asm_str);
2248 break;
2249
2250 case ORDER_UNDEFINED:
2251 break;
2252
2253 default:
2254 gcc_unreachable ();
2255 }
2256 }
2257
2258 symtab->clear_asm_symbols ();
2259
2260 free (nodes);
2261 }
2262
2263 static void
ipa_passes(void)2264 ipa_passes (void)
2265 {
2266 gcc::pass_manager *passes = g->get_passes ();
2267
2268 set_cfun (NULL);
2269 current_function_decl = NULL;
2270 gimple_register_cfg_hooks ();
2271 bitmap_obstack_initialize (NULL);
2272
2273 invoke_plugin_callbacks (PLUGIN_ALL_IPA_PASSES_START, NULL);
2274
2275 if (!in_lto_p)
2276 {
2277 execute_ipa_pass_list (passes->all_small_ipa_passes);
2278 if (seen_error ())
2279 return;
2280 }
2281
2282 /* This extra symtab_remove_unreachable_nodes pass tends to catch some
2283 devirtualization and other changes where removal iterate. */
2284 symtab->remove_unreachable_nodes (symtab->dump_file);
2285
2286 /* If pass_all_early_optimizations was not scheduled, the state of
2287 the cgraph will not be properly updated. Update it now. */
2288 if (symtab->state < IPA_SSA)
2289 symtab->state = IPA_SSA;
2290
2291 if (!in_lto_p)
2292 {
2293 /* Generate coverage variables and constructors. */
2294 coverage_finish ();
2295
2296 /* Process new functions added. */
2297 set_cfun (NULL);
2298 current_function_decl = NULL;
2299 symtab->process_new_functions ();
2300
2301 execute_ipa_summary_passes
2302 ((ipa_opt_pass_d *) passes->all_regular_ipa_passes);
2303 }
2304
2305 /* Some targets need to handle LTO assembler output specially. */
2306 if (flag_generate_lto || flag_generate_offload)
2307 targetm.asm_out.lto_start ();
2308
2309 if (!in_lto_p)
2310 {
2311 if (g->have_offload)
2312 {
2313 section_name_prefix = OFFLOAD_SECTION_NAME_PREFIX;
2314 lto_stream_offload_p = true;
2315 ipa_write_summaries ();
2316 lto_stream_offload_p = false;
2317 }
2318 if (flag_lto)
2319 {
2320 section_name_prefix = LTO_SECTION_NAME_PREFIX;
2321 lto_stream_offload_p = false;
2322 ipa_write_summaries ();
2323 }
2324 }
2325
2326 if (flag_generate_lto || flag_generate_offload)
2327 targetm.asm_out.lto_end ();
2328
2329 if (!flag_ltrans && (in_lto_p || !flag_lto || flag_fat_lto_objects))
2330 execute_ipa_pass_list (passes->all_regular_ipa_passes);
2331 invoke_plugin_callbacks (PLUGIN_ALL_IPA_PASSES_END, NULL);
2332
2333 bitmap_obstack_release (NULL);
2334 }
2335
2336
2337 /* Return string alias is alias of. */
2338
2339 static tree
get_alias_symbol(tree decl)2340 get_alias_symbol (tree decl)
2341 {
2342 tree alias = lookup_attribute ("alias", DECL_ATTRIBUTES (decl));
2343 return get_identifier (TREE_STRING_POINTER
2344 (TREE_VALUE (TREE_VALUE (alias))));
2345 }
2346
2347
2348 /* Weakrefs may be associated to external decls and thus not output
2349 at expansion time. Emit all necessary aliases. */
2350
2351 void
output_weakrefs(void)2352 symbol_table::output_weakrefs (void)
2353 {
2354 symtab_node *node;
2355 cgraph_node *cnode;
2356 FOR_EACH_SYMBOL (node)
2357 if (node->alias
2358 && !TREE_ASM_WRITTEN (node->decl)
2359 && (!(cnode = dyn_cast <cgraph_node *> (node))
2360 || !cnode->instrumented_version
2361 || !TREE_ASM_WRITTEN (cnode->instrumented_version->decl))
2362 && node->weakref)
2363 {
2364 tree target;
2365
2366 /* Weakrefs are special by not requiring target definition in current
2367 compilation unit. It is thus bit hard to work out what we want to
2368 alias.
2369 When alias target is defined, we need to fetch it from symtab reference,
2370 otherwise it is pointed to by alias_target. */
2371 if (node->alias_target)
2372 target = (DECL_P (node->alias_target)
2373 ? DECL_ASSEMBLER_NAME (node->alias_target)
2374 : node->alias_target);
2375 else if (node->analyzed)
2376 target = DECL_ASSEMBLER_NAME (node->get_alias_target ()->decl);
2377 else
2378 {
2379 gcc_unreachable ();
2380 target = get_alias_symbol (node->decl);
2381 }
2382 do_assemble_alias (node->decl, target);
2383 }
2384 }
2385
2386 /* Perform simple optimizations based on callgraph. */
2387
2388 void
compile(void)2389 symbol_table::compile (void)
2390 {
2391 if (seen_error ())
2392 return;
2393
2394 symtab_node::checking_verify_symtab_nodes ();
2395
2396 timevar_push (TV_CGRAPHOPT);
2397 if (pre_ipa_mem_report)
2398 {
2399 fprintf (stderr, "Memory consumption before IPA\n");
2400 dump_memory_report (false);
2401 }
2402 if (!quiet_flag)
2403 fprintf (stderr, "Performing interprocedural optimizations\n");
2404 state = IPA;
2405
2406 /* Offloading requires LTO infrastructure. */
2407 if (!in_lto_p && g->have_offload)
2408 flag_generate_offload = 1;
2409
2410 /* If LTO is enabled, initialize the streamer hooks needed by GIMPLE. */
2411 if (flag_generate_lto || flag_generate_offload)
2412 lto_streamer_hooks_init ();
2413
2414 /* Don't run the IPA passes if there was any error or sorry messages. */
2415 if (!seen_error ())
2416 ipa_passes ();
2417
2418 /* Do nothing else if any IPA pass found errors or if we are just streaming LTO. */
2419 if (seen_error ()
2420 || (!in_lto_p && flag_lto && !flag_fat_lto_objects))
2421 {
2422 timevar_pop (TV_CGRAPHOPT);
2423 return;
2424 }
2425
2426 global_info_ready = true;
2427 if (dump_file)
2428 {
2429 fprintf (dump_file, "Optimized ");
2430 symtab_node:: dump_table (dump_file);
2431 }
2432 if (post_ipa_mem_report)
2433 {
2434 fprintf (stderr, "Memory consumption after IPA\n");
2435 dump_memory_report (false);
2436 }
2437 timevar_pop (TV_CGRAPHOPT);
2438
2439 /* Output everything. */
2440 (*debug_hooks->assembly_start) ();
2441 if (!quiet_flag)
2442 fprintf (stderr, "Assembling functions:\n");
2443 symtab_node::checking_verify_symtab_nodes ();
2444
2445 materialize_all_clones ();
2446 bitmap_obstack_initialize (NULL);
2447 execute_ipa_pass_list (g->get_passes ()->all_late_ipa_passes);
2448 bitmap_obstack_release (NULL);
2449 mark_functions_to_output ();
2450
2451 /* When weakref support is missing, we autmatically translate all
2452 references to NODE to references to its ultimate alias target.
2453 The renaming mechanizm uses flag IDENTIFIER_TRANSPARENT_ALIAS and
2454 TREE_CHAIN.
2455
2456 Set up this mapping before we output any assembler but once we are sure
2457 that all symbol renaming is done.
2458
2459 FIXME: All this uglyness can go away if we just do renaming at gimple
2460 level by physically rewritting the IL. At the moment we can only redirect
2461 calls, so we need infrastructure for renaming references as well. */
2462 #ifndef ASM_OUTPUT_WEAKREF
2463 symtab_node *node;
2464
2465 FOR_EACH_SYMBOL (node)
2466 if (node->alias
2467 && lookup_attribute ("weakref", DECL_ATTRIBUTES (node->decl)))
2468 {
2469 IDENTIFIER_TRANSPARENT_ALIAS
2470 (DECL_ASSEMBLER_NAME (node->decl)) = 1;
2471 TREE_CHAIN (DECL_ASSEMBLER_NAME (node->decl))
2472 = (node->alias_target ? node->alias_target
2473 : DECL_ASSEMBLER_NAME (node->get_alias_target ()->decl));
2474 }
2475 #endif
2476
2477 state = EXPANSION;
2478
2479 if (!flag_toplevel_reorder)
2480 output_in_order (false);
2481 else
2482 {
2483 /* Output first asm statements and anything ordered. The process
2484 flag is cleared for these nodes, so we skip them later. */
2485 output_in_order (true);
2486 expand_all_functions ();
2487 output_variables ();
2488 }
2489
2490 process_new_functions ();
2491 state = FINISHED;
2492 output_weakrefs ();
2493
2494 if (dump_file)
2495 {
2496 fprintf (dump_file, "\nFinal ");
2497 symtab_node::dump_table (dump_file);
2498 }
2499 if (!flag_checking)
2500 return;
2501 symtab_node::verify_symtab_nodes ();
2502 /* Double check that all inline clones are gone and that all
2503 function bodies have been released from memory. */
2504 if (!seen_error ())
2505 {
2506 cgraph_node *node;
2507 bool error_found = false;
2508
2509 FOR_EACH_DEFINED_FUNCTION (node)
2510 if (node->global.inlined_to
2511 || gimple_has_body_p (node->decl))
2512 {
2513 error_found = true;
2514 node->debug ();
2515 }
2516 if (error_found)
2517 internal_error ("nodes with unreleased memory found");
2518 }
2519 }
2520
2521
2522 /* Analyze the whole compilation unit once it is parsed completely. */
2523
2524 void
finalize_compilation_unit(void)2525 symbol_table::finalize_compilation_unit (void)
2526 {
2527 timevar_push (TV_CGRAPH);
2528
2529 /* If we're here there's no current function anymore. Some frontends
2530 are lazy in clearing these. */
2531 current_function_decl = NULL;
2532 set_cfun (NULL);
2533
2534 /* Do not skip analyzing the functions if there were errors, we
2535 miss diagnostics for following functions otherwise. */
2536
2537 /* Emit size functions we didn't inline. */
2538 finalize_size_functions ();
2539
2540 /* Mark alias targets necessary and emit diagnostics. */
2541 handle_alias_pairs ();
2542
2543 if (!quiet_flag)
2544 {
2545 fprintf (stderr, "\nAnalyzing compilation unit\n");
2546 fflush (stderr);
2547 }
2548
2549 if (flag_dump_passes)
2550 dump_passes ();
2551
2552 /* Gimplify and lower all functions, compute reachability and
2553 remove unreachable nodes. */
2554 analyze_functions (/*first_time=*/true);
2555
2556 /* Mark alias targets necessary and emit diagnostics. */
2557 handle_alias_pairs ();
2558
2559 /* Gimplify and lower thunks. */
2560 analyze_functions (/*first_time=*/false);
2561
2562 if (!seen_error ())
2563 {
2564 /* Emit early debug for reachable functions, and by consequence,
2565 locally scoped symbols. */
2566 struct cgraph_node *cnode;
2567 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (cnode)
2568 (*debug_hooks->early_global_decl) (cnode->decl);
2569
2570 /* Clean up anything that needs cleaning up after initial debug
2571 generation. */
2572 (*debug_hooks->early_finish) ();
2573 }
2574
2575 /* Finally drive the pass manager. */
2576 compile ();
2577
2578 timevar_pop (TV_CGRAPH);
2579 }
2580
2581 /* Reset all state within cgraphunit.c so that we can rerun the compiler
2582 within the same process. For use by toplev::finalize. */
2583
2584 void
cgraphunit_c_finalize(void)2585 cgraphunit_c_finalize (void)
2586 {
2587 gcc_assert (cgraph_new_nodes.length () == 0);
2588 cgraph_new_nodes.truncate (0);
2589
2590 vtable_entry_type = NULL;
2591 queued_nodes = &symtab_terminator;
2592
2593 first_analyzed = NULL;
2594 first_analyzed_var = NULL;
2595 }
2596
2597 /* Creates a wrapper from cgraph_node to TARGET node. Thunk is used for this
2598 kind of wrapper method. */
2599
2600 void
create_wrapper(cgraph_node * target)2601 cgraph_node::create_wrapper (cgraph_node *target)
2602 {
2603 /* Preserve DECL_RESULT so we get right by reference flag. */
2604 tree decl_result = DECL_RESULT (decl);
2605
2606 /* Remove the function's body but keep arguments to be reused
2607 for thunk. */
2608 release_body (true);
2609 reset ();
2610
2611 DECL_UNINLINABLE (decl) = false;
2612 DECL_RESULT (decl) = decl_result;
2613 DECL_INITIAL (decl) = NULL;
2614 allocate_struct_function (decl, false);
2615 set_cfun (NULL);
2616
2617 /* Turn alias into thunk and expand it into GIMPLE representation. */
2618 definition = true;
2619
2620 memset (&thunk, 0, sizeof (cgraph_thunk_info));
2621 thunk.thunk_p = true;
2622 create_edge (target, NULL, count, CGRAPH_FREQ_BASE);
2623 callees->can_throw_external = !TREE_NOTHROW (target->decl);
2624
2625 tree arguments = DECL_ARGUMENTS (decl);
2626
2627 while (arguments)
2628 {
2629 TREE_ADDRESSABLE (arguments) = false;
2630 arguments = TREE_CHAIN (arguments);
2631 }
2632
2633 expand_thunk (false, true);
2634
2635 /* Inline summary set-up. */
2636 analyze ();
2637 inline_analyze_function (this);
2638 }
2639
2640 #include "gt-cgraphunit.h"
2641