1 /* Callgraph based interprocedural optimizations.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
3 Free Software Foundation, Inc.
4 Contributed by Jan Hubicka
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 /* This module implements main driver of compilation process as well as
23 few basic interprocedural optimizers.
25 The main scope of this file is to act as an interface in between
26 tree based frontends and the backend (and middle end)
28 The front-end is supposed to use following functionality:
30 - cgraph_finalize_function
32 This function is called once front-end has parsed whole body of function
33 and it is certain that the function body nor the declaration will change.
35 (There is one exception needed for implementing GCC extern inline
38 - varpool_finalize_variable
40 This function has same behavior as the above but is used for static
43 - cgraph_finalize_compilation_unit
45 This function is called once (source level) compilation unit is finalized
46 and it will no longer change.
48 In the the call-graph construction and local function
49 analysis takes place here. Bodies of unreachable functions are released
50 to conserve memory usage.
52 The function can be called multiple times when multiple source level
53 compilation units are combined (such as in C frontend)
57 In this unit-at-a-time compilation the intra procedural analysis takes
58 place here. In particular the static functions whose address is never
59 taken are marked as local. Backend can then use this information to
60 modify calling conventions, do better inlining or similar optimizations.
62 - cgraph_mark_needed_node
63 - varpool_mark_needed_node
65 When function or variable is referenced by some hidden way the call-graph
66 data structure must be updated accordingly by this function.
67 There should be little need to call this function and all the references
68 should be made explicit to cgraph code. At present these functions are
69 used by C++ frontend to explicitly mark the keyed methods.
71 - analyze_expr callback
73 This function is responsible for lowering tree nodes not understood by
74 generic code into understandable ones or alternatively marking
75 callgraph and varpool nodes referenced by the as needed.
77 ??? On the tree-ssa genericizing should take place here and we will avoid
78 need for these hooks (replacing them by genericizing hook)
80 Analyzing of all functions is deferred
81 to cgraph_finalize_compilation_unit and expansion into cgraph_optimize.
83 In cgraph_finalize_compilation_unit the reachable functions are
84 analyzed. During analysis the call-graph edges from reachable
85 functions are constructed and their destinations are marked as
86 reachable. References to functions and variables are discovered too
87 and variables found to be needed output to the assembly file. Via
88 mark_referenced call in assemble_variable functions referenced by
89 static variables are noticed too.
91 The intra-procedural information is produced and its existence
92 indicated by global_info_ready. Once this flag is set it is impossible
93 to change function from !reachable to reachable and thus
94 assemble_variable no longer call mark_referenced.
96 Finally the call-graph is topologically sorted and all reachable functions
97 that has not been completely inlined or are not external are output.
99 ??? It is possible that reference to function or variable is optimized
100 out. We can not deal with this nicely because topological order is not
101 suitable for it. For tree-ssa we may consider another pass doing
102 optimization and re-discovering reachable functions.
104 ??? Reorganize code so variables are output very last and only if they
105 really has been referenced by produced code, so we catch more cases
106 where reference has been optimized out. */
111 #include "coretypes.h"
115 #include "tree-flow.h"
116 #include "tree-inline.h"
117 #include "langhooks.h"
118 #include "pointer-set.h"
125 #include "diagnostic.h"
126 #include "tree-pretty-print.h"
127 #include "gimple-pretty-print.h"
132 #include "function.h"
133 #include "ipa-prop.h"
135 #include "tree-iterator.h"
136 #include "tree-pass.h"
137 #include "tree-dump.h"
139 #include "coverage.h"
142 static void cgraph_expand_all_functions (void);
143 static void cgraph_mark_functions_to_output (void);
144 static void cgraph_expand_function (struct cgraph_node *);
145 static void cgraph_output_pending_asms (void);
146 static void cgraph_analyze_function (struct cgraph_node *);
148 FILE *cgraph_dump_file;
150 /* Used for vtable lookup in thunk adjusting. */
151 static GTY (()) tree vtable_entry_type;
153 /* Determine if function DECL is needed. That is, visible to something
154 either outside this translation unit, something magic in the system
158 cgraph_decide_is_function_needed (struct cgraph_node *node, tree decl)
160 /* If the user told us it is used, then it must be so. */
161 if (node->local.externally_visible)
164 /* ??? If the assembler name is set by hand, it is possible to assemble
165 the name later after finalizing the function and the fact is noticed
166 in assemble_name then. This is arguably a bug. */
167 if (DECL_ASSEMBLER_NAME_SET_P (decl)
168 && TREE_SYMBOL_REFERENCED (DECL_ASSEMBLER_NAME (decl)))
171 /* With -fkeep-inline-functions we are keeping all inline functions except
172 for extern inline ones. */
173 if (flag_keep_inline_functions
174 && DECL_DECLARED_INLINE_P (decl)
175 && !DECL_EXTERNAL (decl)
176 && !lookup_attribute ("always_inline", DECL_ATTRIBUTES (decl)))
179 /* If we decided it was needed before, but at the time we didn't have
180 the body of the function available, then it's still needed. We have
181 to go back and re-check its dependencies now. */
185 /* Externally visible functions must be output. The exception is
186 COMDAT functions that must be output only when they are needed.
188 When not optimizing, also output the static functions. (see
189 PR24561), but don't do so for always_inline functions, functions
190 declared inline and nested functions. These were optimized out
191 in the original implementation and it is unclear whether we want
192 to change the behavior here. */
193 if (((TREE_PUBLIC (decl)
195 && !node->local.disregard_inline_limits
196 && !DECL_DECLARED_INLINE_P (decl)
197 && !(DECL_CONTEXT (decl)
198 && TREE_CODE (DECL_CONTEXT (decl)) == FUNCTION_DECL)))
199 && !flag_whole_program
202 && !DECL_COMDAT (decl) && !DECL_EXTERNAL (decl))
208 /* Process CGRAPH_NEW_FUNCTIONS and perform actions necessary to add these
209 functions into callgraph in a way so they look like ordinary reachable
210 functions inserted into callgraph already at construction time. */
213 cgraph_process_new_functions (void)
217 struct cgraph_node *node;
219 varpool_analyze_pending_decls ();
220 /* Note that this queue may grow as its being processed, as the new
221 functions may generate new ones. */
222 while (cgraph_new_nodes)
224 node = cgraph_new_nodes;
226 cgraph_new_nodes = cgraph_new_nodes->next_needed;
227 switch (cgraph_state)
229 case CGRAPH_STATE_CONSTRUCTION:
230 /* At construction time we just need to finalize function and move
231 it into reachable functions list. */
233 node->next_needed = NULL;
234 cgraph_finalize_function (fndecl, false);
235 cgraph_mark_reachable_node (node);
239 case CGRAPH_STATE_IPA:
240 case CGRAPH_STATE_IPA_SSA:
241 /* When IPA optimization already started, do all essential
242 transformations that has been already performed on the whole
243 cgraph but not on this function. */
245 gimple_register_cfg_hooks ();
247 cgraph_analyze_function (node);
248 push_cfun (DECL_STRUCT_FUNCTION (fndecl));
249 current_function_decl = fndecl;
250 compute_inline_parameters (node);
251 if ((cgraph_state == CGRAPH_STATE_IPA_SSA
252 && !gimple_in_ssa_p (DECL_STRUCT_FUNCTION (fndecl)))
253 /* When not optimizing, be sure we run early local passes anyway
256 execute_pass_list (pass_early_local_passes.pass.sub);
257 free_dominance_info (CDI_POST_DOMINATORS);
258 free_dominance_info (CDI_DOMINATORS);
260 current_function_decl = NULL;
263 case CGRAPH_STATE_EXPANSION:
264 /* Functions created during expansion shall be compiled
267 cgraph_expand_function (node);
274 cgraph_call_function_insertion_hooks (node);
275 varpool_analyze_pending_decls ();
280 /* As an GCC extension we allow redefinition of the function. The
281 semantics when both copies of bodies differ is not well defined.
282 We replace the old body with new body so in unit at a time mode
283 we always use new body, while in normal mode we may end up with
284 old body inlined into some functions and new body expanded and
287 ??? It may make more sense to use one body for inlining and other
288 body for expanding the function but this is difficult to do. */
291 cgraph_reset_node (struct cgraph_node *node)
293 /* If node->process is set, then we have already begun whole-unit analysis.
294 This is *not* testing for whether we've already emitted the function.
295 That case can be sort-of legitimately seen with real function redefinition
296 errors. I would argue that the front end should never present us with
297 such a case, but don't enforce that for now. */
298 gcc_assert (!node->process);
300 /* Reset our data structures so we can analyze the function again. */
301 memset (&node->local, 0, sizeof (node->local));
302 memset (&node->global, 0, sizeof (node->global));
303 memset (&node->rtl, 0, sizeof (node->rtl));
304 node->analyzed = false;
305 node->local.redefined_extern_inline = true;
306 node->local.finalized = false;
308 cgraph_node_remove_callees (node);
310 /* We may need to re-queue the node for assembling in case
311 we already proceeded it and ignored as not needed or got
312 a re-declaration in IMA mode. */
315 struct cgraph_node *n;
317 for (n = cgraph_nodes_queue; n; n = n->next_needed)
326 cgraph_lower_function (struct cgraph_node *node)
332 lower_nested_functions (node->decl);
333 gcc_assert (!node->nested);
335 tree_lowering_passes (node->decl);
336 node->lowered = true;
339 /* DECL has been parsed. Take it, queue it, compile it at the whim of the
340 logic in effect. If NESTED is true, then our caller cannot stand to have
341 the garbage collector run at the moment. We would need to either create
342 a new GC context, or just not compile right now. */
345 cgraph_finalize_function (tree decl, bool nested)
347 struct cgraph_node *node = cgraph_node (decl);
349 if (node->local.finalized)
350 cgraph_reset_node (node);
352 node->pid = cgraph_max_pid ++;
353 notice_global_symbol (decl);
354 node->local.finalized = true;
355 node->lowered = DECL_STRUCT_FUNCTION (decl)->cfg != NULL;
356 node->finalized_by_frontend = true;
358 if (cgraph_decide_is_function_needed (node, decl))
359 cgraph_mark_needed_node (node);
361 /* Since we reclaim unreachable nodes at the end of every language
362 level unit, we need to be conservative about possible entry points
364 if ((TREE_PUBLIC (decl) && !DECL_COMDAT (decl) && !DECL_EXTERNAL (decl))
365 || DECL_STATIC_CONSTRUCTOR (decl)
366 || DECL_STATIC_DESTRUCTOR (decl)
367 /* COMDAT virtual functions may be referenced by vtable from
368 other compilatoin unit. Still we want to devirtualize calls
369 to those so we need to analyze them.
370 FIXME: We should introduce may edges for this purpose and update
371 their handling in unreachable function removal and inliner too. */
372 || (DECL_VIRTUAL_P (decl) && (DECL_COMDAT (decl) || DECL_EXTERNAL (decl))))
373 cgraph_mark_reachable_node (node);
375 /* If we've not yet emitted decl, tell the debug info about it. */
376 if (!TREE_ASM_WRITTEN (decl))
377 (*debug_hooks->deferred_inline_function) (decl);
379 /* Possibly warn about unused parameters. */
380 if (warn_unused_parameter)
381 do_warn_unused_parameter (decl);
387 /* C99 extern inline keywords allow changing of declaration after function
388 has been finalized. We need to re-decide if we want to mark the function as
392 cgraph_mark_if_needed (tree decl)
394 struct cgraph_node *node = cgraph_node (decl);
395 if (node->local.finalized && cgraph_decide_is_function_needed (node, decl))
396 cgraph_mark_needed_node (node);
399 #ifdef ENABLE_CHECKING
400 /* Return TRUE if NODE2 is equivalent to NODE or its clone. */
402 clone_of_p (struct cgraph_node *node, struct cgraph_node *node2)
404 while (node != node2 && node2)
405 node2 = node2->clone_of;
406 return node2 != NULL;
410 /* Verify edge E count and frequency. */
413 verify_edge_count_and_frequency (struct cgraph_edge *e)
415 bool error_found = false;
418 error ("caller edge count is negative");
421 if (e->frequency < 0)
423 error ("caller edge frequency is negative");
426 if (e->frequency > CGRAPH_FREQ_MAX)
428 error ("caller edge frequency is too large");
431 if (gimple_has_body_p (e->caller->decl)
432 && !e->caller->global.inlined_to
434 != compute_call_stmt_bb_frequency (e->caller->decl,
435 gimple_bb (e->call_stmt))))
437 error ("caller edge frequency %i does not match BB freqency %i",
439 compute_call_stmt_bb_frequency (e->caller->decl,
440 gimple_bb (e->call_stmt)));
446 /* Verify cgraph nodes of given cgraph node. */
448 verify_cgraph_node (struct cgraph_node *node)
450 struct cgraph_edge *e;
451 struct function *this_cfun = DECL_STRUCT_FUNCTION (node->decl);
452 struct function *saved_cfun = cfun;
453 basic_block this_block;
454 gimple_stmt_iterator gsi;
455 bool error_found = false;
460 timevar_push (TV_CGRAPH_VERIFY);
461 /* debug_generic_stmt needs correct cfun */
462 set_cfun (this_cfun);
463 for (e = node->callees; e; e = e->next_callee)
466 error ("aux field set for edge %s->%s",
467 identifier_to_locale (cgraph_node_name (e->caller)),
468 identifier_to_locale (cgraph_node_name (e->callee)));
473 error ("Execution count is negative");
476 if (node->global.inlined_to && node->local.externally_visible)
478 error ("Externally visible inline clone");
481 if (node->global.inlined_to && node->address_taken)
483 error ("Inline clone with address taken");
486 if (node->global.inlined_to && node->needed)
488 error ("Inline clone is needed");
491 for (e = node->indirect_calls; e; e = e->next_callee)
495 error ("aux field set for indirect edge from %s",
496 identifier_to_locale (cgraph_node_name (e->caller)));
499 if (!e->indirect_unknown_callee
500 || !e->indirect_info)
502 error ("An indirect edge from %s is not marked as indirect or has "
503 "associated indirect_info, the corresponding statement is: ",
504 identifier_to_locale (cgraph_node_name (e->caller)));
505 debug_gimple_stmt (e->call_stmt);
509 for (e = node->callers; e; e = e->next_caller)
511 if (verify_edge_count_and_frequency (e))
513 if (!e->inline_failed)
515 if (node->global.inlined_to
516 != (e->caller->global.inlined_to
517 ? e->caller->global.inlined_to : e->caller))
519 error ("inlined_to pointer is wrong");
522 if (node->callers->next_caller)
524 error ("multiple inline callers");
529 if (node->global.inlined_to)
531 error ("inlined_to pointer set for noninline callers");
535 for (e = node->indirect_calls; e; e = e->next_callee)
536 if (verify_edge_count_and_frequency (e))
538 if (!node->callers && node->global.inlined_to)
540 error ("inlined_to pointer is set but no predecessors found");
543 if (node->global.inlined_to == node)
545 error ("inlined_to pointer refers to itself");
549 if (!cgraph_node (node->decl))
551 error ("node not found in cgraph_hash");
557 struct cgraph_node *n;
558 for (n = node->clone_of->clones; n; n = n->next_sibling_clone)
563 error ("node has wrong clone_of");
569 struct cgraph_node *n;
570 for (n = node->clones; n; n = n->next_sibling_clone)
571 if (n->clone_of != node)
575 error ("node has wrong clone list");
579 if ((node->prev_sibling_clone || node->next_sibling_clone) && !node->clone_of)
581 error ("node is in clone list but it is not clone");
584 if (!node->prev_sibling_clone && node->clone_of && node->clone_of->clones != node)
586 error ("node has wrong prev_clone pointer");
589 if (node->prev_sibling_clone && node->prev_sibling_clone->next_sibling_clone != node)
591 error ("double linked list of clones corrupted");
594 if (node->same_comdat_group)
596 struct cgraph_node *n = node->same_comdat_group;
598 if (!DECL_ONE_ONLY (node->decl))
600 error ("non-DECL_ONE_ONLY node in a same_comdat_group list");
605 error ("node is alone in a comdat group");
610 if (!n->same_comdat_group)
612 error ("same_comdat_group is not a circular list");
616 n = n->same_comdat_group;
621 if (node->analyzed && gimple_has_body_p (node->decl)
622 && !TREE_ASM_WRITTEN (node->decl)
623 && (!DECL_EXTERNAL (node->decl) || node->global.inlined_to)
628 /* The nodes we're interested in are never shared, so walk
629 the tree ignoring duplicates. */
630 struct pointer_set_t *visited_nodes = pointer_set_create ();
631 /* Reach the trees by walking over the CFG, and note the
632 enclosing basic-blocks in the call edges. */
633 FOR_EACH_BB_FN (this_block, this_cfun)
634 for (gsi = gsi_start_bb (this_block);
638 gimple stmt = gsi_stmt (gsi);
639 if (is_gimple_call (stmt))
641 struct cgraph_edge *e = cgraph_edge (node, stmt);
642 tree decl = gimple_call_fndecl (stmt);
647 error ("shared call_stmt:");
648 debug_gimple_stmt (stmt);
651 if (!e->indirect_unknown_callee)
653 if (e->callee->same_body_alias)
655 error ("edge points to same body alias:");
656 debug_tree (e->callee->decl);
659 #ifdef ENABLE_CHECKING
660 else if (!e->callee->global.inlined_to
662 && cgraph_get_node (decl)
663 && (e->callee->former_clone_of
664 != cgraph_get_node (decl)->decl)
665 && !clone_of_p (cgraph_node (decl),
668 error ("edge points to wrong declaration:");
669 debug_tree (e->callee->decl);
670 fprintf (stderr," Instead of:");
678 error ("an indirect edge with unknown callee "
679 "corresponding to a call_stmt with "
680 "a known declaration:");
682 debug_gimple_stmt (e->call_stmt);
688 error ("missing callgraph edge for call stmt:");
689 debug_gimple_stmt (stmt);
694 pointer_set_destroy (visited_nodes);
697 /* No CFG available?! */
700 for (e = node->callees; e; e = e->next_callee)
704 error ("edge %s->%s has no corresponding call_stmt",
705 identifier_to_locale (cgraph_node_name (e->caller)),
706 identifier_to_locale (cgraph_node_name (e->callee)));
707 debug_gimple_stmt (e->call_stmt);
712 for (e = node->indirect_calls; e; e = e->next_callee)
716 error ("an indirect edge from %s has no corresponding call_stmt",
717 identifier_to_locale (cgraph_node_name (e->caller)));
718 debug_gimple_stmt (e->call_stmt);
726 dump_cgraph_node (stderr, node);
727 internal_error ("verify_cgraph_node failed");
729 set_cfun (saved_cfun);
730 timevar_pop (TV_CGRAPH_VERIFY);
733 /* Verify whole cgraph structure. */
737 struct cgraph_node *node;
742 for (node = cgraph_nodes; node; node = node->next)
743 verify_cgraph_node (node);
746 /* Output all asm statements we have stored up to be output. */
749 cgraph_output_pending_asms (void)
751 struct cgraph_asm_node *can;
756 for (can = cgraph_asm_nodes; can; can = can->next)
757 assemble_asm (can->asm_str);
758 cgraph_asm_nodes = NULL;
761 /* Analyze the function scheduled to be output. */
763 cgraph_analyze_function (struct cgraph_node *node)
765 tree save = current_function_decl;
766 tree decl = node->decl;
768 current_function_decl = decl;
769 push_cfun (DECL_STRUCT_FUNCTION (decl));
771 assign_assembler_name_if_neeeded (node->decl);
773 /* Make sure to gimplify bodies only once. During analyzing a
774 function we lower it, which will require gimplified nested
775 functions, so we can end up here with an already gimplified
777 if (!gimple_body (decl))
778 gimplify_function_tree (decl);
779 dump_function (TDI_generic, decl);
781 cgraph_lower_function (node);
782 node->analyzed = true;
785 current_function_decl = save;
788 /* Look for externally_visible and used attributes and mark cgraph nodes
791 We cannot mark the nodes at the point the attributes are processed (in
792 handle_*_attribute) because the copy of the declarations available at that
793 point may not be canonical. For example, in:
796 void f() __attribute__((used));
798 the declaration we see in handle_used_attribute will be the second
799 declaration -- but the front end will subsequently merge that declaration
800 with the original declaration and discard the second declaration.
802 Furthermore, we can't mark these nodes in cgraph_finalize_function because:
805 void f() __attribute__((externally_visible));
809 So, we walk the nodes at the end of the translation unit, applying the
810 attributes at that point. */
813 process_function_and_variable_attributes (struct cgraph_node *first,
814 struct varpool_node *first_var)
816 struct cgraph_node *node;
817 struct varpool_node *vnode;
819 for (node = cgraph_nodes; node != first; node = node->next)
821 tree decl = node->decl;
822 if (DECL_PRESERVE_P (decl))
823 cgraph_mark_needed_node (node);
824 if (lookup_attribute ("externally_visible", DECL_ATTRIBUTES (decl)))
826 if (! TREE_PUBLIC (node->decl))
827 warning_at (DECL_SOURCE_LOCATION (node->decl), OPT_Wattributes,
828 "%<externally_visible%>"
829 " attribute have effect only on public objects");
830 else if (node->local.finalized)
831 cgraph_mark_needed_node (node);
834 for (vnode = varpool_nodes; vnode != first_var; vnode = vnode->next)
836 tree decl = vnode->decl;
837 if (DECL_PRESERVE_P (decl))
839 vnode->force_output = true;
840 if (vnode->finalized)
841 varpool_mark_needed_node (vnode);
843 if (lookup_attribute ("externally_visible", DECL_ATTRIBUTES (decl)))
845 if (! TREE_PUBLIC (vnode->decl))
846 warning_at (DECL_SOURCE_LOCATION (vnode->decl), OPT_Wattributes,
847 "%<externally_visible%>"
848 " attribute have effect only on public objects");
849 else if (vnode->finalized)
850 varpool_mark_needed_node (vnode);
855 /* Process CGRAPH_NODES_NEEDED queue, analyze each function (and transitively
856 each reachable functions) and build cgraph.
857 The function can be called multiple times after inserting new nodes
858 into beginning of queue. Just the new part of queue is re-scanned then. */
861 cgraph_analyze_functions (void)
863 /* Keep track of already processed nodes when called multiple times for
864 intermodule optimization. */
865 static struct cgraph_node *first_analyzed;
866 struct cgraph_node *first_processed = first_analyzed;
867 static struct varpool_node *first_analyzed_var;
868 struct cgraph_node *node, *next;
870 bitmap_obstack_initialize (NULL);
871 process_function_and_variable_attributes (first_processed,
873 first_processed = cgraph_nodes;
874 first_analyzed_var = varpool_nodes;
875 varpool_analyze_pending_decls ();
876 if (cgraph_dump_file)
878 fprintf (cgraph_dump_file, "Initial entry points:");
879 for (node = cgraph_nodes; node != first_analyzed; node = node->next)
881 fprintf (cgraph_dump_file, " %s", cgraph_node_name (node));
882 fprintf (cgraph_dump_file, "\n");
884 cgraph_process_new_functions ();
886 /* Propagate reachability flag and lower representation of all reachable
887 functions. In the future, lowering will introduce new functions and
888 new entry points on the way (by template instantiation and virtual
889 method table generation for instance). */
890 while (cgraph_nodes_queue)
892 struct cgraph_edge *edge;
893 tree decl = cgraph_nodes_queue->decl;
895 node = cgraph_nodes_queue;
896 cgraph_nodes_queue = cgraph_nodes_queue->next_needed;
897 node->next_needed = NULL;
899 /* ??? It is possible to create extern inline function and later using
900 weak alias attribute to kill its body. See
901 gcc.c-torture/compile/20011119-1.c */
902 if (!DECL_STRUCT_FUNCTION (decl))
904 cgraph_reset_node (node);
909 cgraph_analyze_function (node);
911 for (edge = node->callees; edge; edge = edge->next_callee)
912 if (!edge->callee->reachable)
913 cgraph_mark_reachable_node (edge->callee);
915 if (node->same_comdat_group)
917 for (next = node->same_comdat_group;
919 next = next->same_comdat_group)
920 cgraph_mark_reachable_node (next);
923 /* If decl is a clone of an abstract function, mark that abstract
924 function so that we don't release its body. The DECL_INITIAL() of that
925 abstract function declaration will be later needed to output debug info. */
926 if (DECL_ABSTRACT_ORIGIN (decl))
928 struct cgraph_node *origin_node = cgraph_node (DECL_ABSTRACT_ORIGIN (decl));
929 origin_node->abstract_and_needed = true;
932 /* We finalize local static variables during constructing callgraph
933 edges. Process their attributes too. */
934 process_function_and_variable_attributes (first_processed,
936 first_processed = cgraph_nodes;
937 first_analyzed_var = varpool_nodes;
938 varpool_analyze_pending_decls ();
939 cgraph_process_new_functions ();
942 /* Collect entry points to the unit. */
943 if (cgraph_dump_file)
945 fprintf (cgraph_dump_file, "Unit entry points:");
946 for (node = cgraph_nodes; node != first_analyzed; node = node->next)
948 fprintf (cgraph_dump_file, " %s", cgraph_node_name (node));
949 fprintf (cgraph_dump_file, "\n\nInitial ");
950 dump_cgraph (cgraph_dump_file);
953 if (cgraph_dump_file)
954 fprintf (cgraph_dump_file, "\nReclaiming functions:");
956 for (node = cgraph_nodes; node != first_analyzed; node = next)
958 tree decl = node->decl;
961 if (node->local.finalized && !gimple_has_body_p (decl))
962 cgraph_reset_node (node);
964 if (!node->reachable && gimple_has_body_p (decl))
966 if (cgraph_dump_file)
967 fprintf (cgraph_dump_file, " %s", cgraph_node_name (node));
968 cgraph_remove_node (node);
972 node->next_needed = NULL;
973 gcc_assert (!node->local.finalized || gimple_has_body_p (decl));
974 gcc_assert (node->analyzed == node->local.finalized);
976 if (cgraph_dump_file)
978 fprintf (cgraph_dump_file, "\n\nReclaimed ");
979 dump_cgraph (cgraph_dump_file);
981 bitmap_obstack_release (NULL);
982 first_analyzed = cgraph_nodes;
987 /* Analyze the whole compilation unit once it is parsed completely. */
990 cgraph_finalize_compilation_unit (void)
992 timevar_push (TV_CGRAPH);
994 /* Do not skip analyzing the functions if there were errors, we
995 miss diagnostics for following functions otherwise. */
997 /* Emit size functions we didn't inline. */
998 finalize_size_functions ();
1000 /* Mark alias targets necessary and emit diagnostics. */
1001 finish_aliases_1 ();
1005 fprintf (stderr, "\nAnalyzing compilation unit\n");
1009 /* Gimplify and lower all functions, compute reachability and
1010 remove unreachable nodes. */
1011 cgraph_analyze_functions ();
1013 /* Mark alias targets necessary and emit diagnostics. */
1014 finish_aliases_1 ();
1016 /* Gimplify and lower thunks. */
1017 cgraph_analyze_functions ();
1019 /* Finally drive the pass manager. */
1022 timevar_pop (TV_CGRAPH);
1026 /* Figure out what functions we want to assemble. */
1029 cgraph_mark_functions_to_output (void)
1031 struct cgraph_node *node;
1032 #ifdef ENABLE_CHECKING
1033 bool check_same_comdat_groups = false;
1035 for (node = cgraph_nodes; node; node = node->next)
1036 gcc_assert (!node->process);
1039 for (node = cgraph_nodes; node; node = node->next)
1041 tree decl = node->decl;
1042 struct cgraph_edge *e;
1044 gcc_assert (!node->process || node->same_comdat_group);
1048 for (e = node->callers; e; e = e->next_caller)
1049 if (e->inline_failed)
1052 /* We need to output all local functions that are used and not
1053 always inlined, as well as those that are reachable from
1054 outside the current compilation unit. */
1056 && !node->global.inlined_to
1057 && (!cgraph_only_called_directly_p (node)
1058 || (e && node->reachable))
1059 && !TREE_ASM_WRITTEN (decl)
1060 && !DECL_EXTERNAL (decl))
1063 if (node->same_comdat_group)
1065 struct cgraph_node *next;
1066 for (next = node->same_comdat_group;
1068 next = next->same_comdat_group)
1072 else if (node->same_comdat_group)
1074 #ifdef ENABLE_CHECKING
1075 check_same_comdat_groups = true;
1080 /* We should've reclaimed all functions that are not needed. */
1081 #ifdef ENABLE_CHECKING
1082 if (!node->global.inlined_to
1083 && gimple_has_body_p (decl)
1084 /* FIXME: in ltrans unit when offline copy is outside partition but inline copies
1085 are inside partition, we can end up not removing the body since we no longer
1086 have analyzed node pointing to it. */
1087 && !node->in_other_partition
1088 && !DECL_EXTERNAL (decl))
1090 dump_cgraph_node (stderr, node);
1091 internal_error ("failed to reclaim unneeded function");
1094 gcc_assert (node->global.inlined_to
1095 || !gimple_has_body_p (decl)
1096 || node->in_other_partition
1097 || DECL_EXTERNAL (decl));
1102 #ifdef ENABLE_CHECKING
1103 if (check_same_comdat_groups)
1104 for (node = cgraph_nodes; node; node = node->next)
1105 if (node->same_comdat_group && !node->process)
1107 tree decl = node->decl;
1108 if (!node->global.inlined_to
1109 && gimple_has_body_p (decl)
1110 /* FIXME: in ltrans unit when offline copy is outside partition but inline copies
1111 are inside partition, we can end up not removing the body since we no longer
1112 have analyzed node pointing to it. */
1113 && !node->in_other_partition
1114 && !DECL_EXTERNAL (decl))
1116 dump_cgraph_node (stderr, node);
1117 internal_error ("failed to reclaim unneeded function");
1123 /* DECL is FUNCTION_DECL. Initialize datastructures so DECL is a function
1124 in lowered gimple form.
1126 Set current_function_decl and cfun to newly constructed empty function body.
1127 return basic block in the function body. */
1130 init_lowered_empty_function (tree decl)
1134 current_function_decl = decl;
1135 allocate_struct_function (decl, false);
1136 gimple_register_cfg_hooks ();
1137 init_empty_tree_cfg ();
1138 init_tree_ssa (cfun);
1139 init_ssa_operands ();
1140 cfun->gimple_df->in_ssa_p = true;
1141 DECL_INITIAL (decl) = make_node (BLOCK);
1143 DECL_SAVED_TREE (decl) = error_mark_node;
1144 cfun->curr_properties |=
1145 (PROP_gimple_lcf | PROP_gimple_leh | PROP_cfg | PROP_referenced_vars |
1148 /* Create BB for body of the function and connect it properly. */
1149 bb = create_basic_block (NULL, (void *) 0, ENTRY_BLOCK_PTR);
1150 make_edge (ENTRY_BLOCK_PTR, bb, 0);
1151 make_edge (bb, EXIT_BLOCK_PTR, 0);
1156 /* Adjust PTR by the constant FIXED_OFFSET, and by the vtable
1157 offset indicated by VIRTUAL_OFFSET, if that is
1158 non-null. THIS_ADJUSTING is nonzero for a this adjusting thunk and
1159 zero for a result adjusting thunk. */
1162 thunk_adjust (gimple_stmt_iterator * bsi,
1163 tree ptr, bool this_adjusting,
1164 HOST_WIDE_INT fixed_offset, tree virtual_offset)
1170 && fixed_offset != 0)
1172 stmt = gimple_build_assign (ptr,
1173 fold_build2_loc (input_location,
1175 TREE_TYPE (ptr), ptr,
1176 size_int (fixed_offset)));
1177 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1180 /* If there's a virtual offset, look up that value in the vtable and
1181 adjust the pointer again. */
1189 if (!vtable_entry_type)
1191 tree vfunc_type = make_node (FUNCTION_TYPE);
1192 TREE_TYPE (vfunc_type) = integer_type_node;
1193 TYPE_ARG_TYPES (vfunc_type) = NULL_TREE;
1194 layout_type (vfunc_type);
1196 vtable_entry_type = build_pointer_type (vfunc_type);
1200 create_tmp_var (build_pointer_type
1201 (build_pointer_type (vtable_entry_type)), "vptr");
1203 /* The vptr is always at offset zero in the object. */
1204 stmt = gimple_build_assign (vtabletmp,
1205 build1 (NOP_EXPR, TREE_TYPE (vtabletmp),
1207 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1208 mark_symbols_for_renaming (stmt);
1209 find_referenced_vars_in (stmt);
1211 /* Form the vtable address. */
1212 vtabletmp2 = create_tmp_var (TREE_TYPE (TREE_TYPE (vtabletmp)),
1214 stmt = gimple_build_assign (vtabletmp2,
1215 build_simple_mem_ref (vtabletmp));
1216 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1217 mark_symbols_for_renaming (stmt);
1218 find_referenced_vars_in (stmt);
1220 /* Find the entry with the vcall offset. */
1221 stmt = gimple_build_assign (vtabletmp2,
1222 fold_build2_loc (input_location,
1224 TREE_TYPE (vtabletmp2),
1226 fold_convert (sizetype,
1228 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1230 /* Get the offset itself. */
1231 vtabletmp3 = create_tmp_var (TREE_TYPE (TREE_TYPE (vtabletmp2)),
1233 stmt = gimple_build_assign (vtabletmp3,
1234 build_simple_mem_ref (vtabletmp2));
1235 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1236 mark_symbols_for_renaming (stmt);
1237 find_referenced_vars_in (stmt);
1239 /* Cast to sizetype. */
1240 offsettmp = create_tmp_var (sizetype, "offset");
1241 stmt = gimple_build_assign (offsettmp, fold_convert (sizetype, vtabletmp3));
1242 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1243 mark_symbols_for_renaming (stmt);
1244 find_referenced_vars_in (stmt);
1246 /* Adjust the `this' pointer. */
1247 ptr = fold_build2_loc (input_location,
1248 POINTER_PLUS_EXPR, TREE_TYPE (ptr), ptr,
1253 && fixed_offset != 0)
1254 /* Adjust the pointer by the constant. */
1258 if (TREE_CODE (ptr) == VAR_DECL)
1262 ptrtmp = create_tmp_var (TREE_TYPE (ptr), "ptr");
1263 stmt = gimple_build_assign (ptrtmp, ptr);
1264 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1265 mark_symbols_for_renaming (stmt);
1266 find_referenced_vars_in (stmt);
1268 ptr = fold_build2_loc (input_location,
1269 POINTER_PLUS_EXPR, TREE_TYPE (ptrtmp), ptrtmp,
1270 size_int (fixed_offset));
1273 /* Emit the statement and gimplify the adjustment expression. */
1274 ret = create_tmp_var (TREE_TYPE (ptr), "adjusted_this");
1275 stmt = gimple_build_assign (ret, ptr);
1276 mark_symbols_for_renaming (stmt);
1277 find_referenced_vars_in (stmt);
1278 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1283 /* Produce assembler for thunk NODE. */
1286 assemble_thunk (struct cgraph_node *node)
1288 bool this_adjusting = node->thunk.this_adjusting;
1289 HOST_WIDE_INT fixed_offset = node->thunk.fixed_offset;
1290 HOST_WIDE_INT virtual_value = node->thunk.virtual_value;
1291 tree virtual_offset = NULL;
1292 tree alias = node->thunk.alias;
1293 tree thunk_fndecl = node->decl;
1294 tree a = DECL_ARGUMENTS (thunk_fndecl);
1296 current_function_decl = thunk_fndecl;
1299 && targetm.asm_out.can_output_mi_thunk (thunk_fndecl, fixed_offset,
1300 virtual_value, alias))
1305 DECL_RESULT (thunk_fndecl)
1306 = build_decl (DECL_SOURCE_LOCATION (thunk_fndecl),
1307 RESULT_DECL, 0, integer_type_node);
1308 fnname = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (thunk_fndecl));
1310 /* The back end expects DECL_INITIAL to contain a BLOCK, so we
1312 fn_block = make_node (BLOCK);
1313 BLOCK_VARS (fn_block) = a;
1314 DECL_INITIAL (thunk_fndecl) = fn_block;
1315 init_function_start (thunk_fndecl);
1317 assemble_start_function (thunk_fndecl, fnname);
1319 targetm.asm_out.output_mi_thunk (asm_out_file, thunk_fndecl,
1320 fixed_offset, virtual_value, alias);
1322 assemble_end_function (thunk_fndecl, fnname);
1323 init_insn_lengths ();
1324 free_after_compilation (cfun);
1326 TREE_ASM_WRITTEN (thunk_fndecl) = 1;
1331 basic_block bb, then_bb, else_bb, return_bb;
1332 gimple_stmt_iterator bsi;
1338 VEC(tree, heap) *vargs;
1343 DECL_IGNORED_P (thunk_fndecl) = 1;
1344 bitmap_obstack_initialize (NULL);
1346 if (node->thunk.virtual_offset_p)
1347 virtual_offset = size_int (virtual_value);
1349 /* Build the return declaration for the function. */
1350 restype = TREE_TYPE (TREE_TYPE (thunk_fndecl));
1351 if (DECL_RESULT (thunk_fndecl) == NULL_TREE)
1353 resdecl = build_decl (input_location, RESULT_DECL, 0, restype);
1354 DECL_ARTIFICIAL (resdecl) = 1;
1355 DECL_IGNORED_P (resdecl) = 1;
1356 DECL_RESULT (thunk_fndecl) = resdecl;
1359 resdecl = DECL_RESULT (thunk_fndecl);
1361 bb = then_bb = else_bb = return_bb = init_lowered_empty_function (thunk_fndecl);
1363 bsi = gsi_start_bb (bb);
1365 /* Build call to the function being thunked. */
1366 if (!VOID_TYPE_P (restype))
1368 if (!is_gimple_reg_type (restype))
1371 add_local_decl (cfun, restmp);
1372 BLOCK_VARS (DECL_INITIAL (current_function_decl)) = restmp;
1375 restmp = create_tmp_var_raw (restype, "retval");
1378 for (arg = a; arg; arg = DECL_CHAIN (arg))
1380 vargs = VEC_alloc (tree, heap, nargs);
1382 VEC_quick_push (tree, vargs,
1387 VEC_quick_push (tree, vargs, a);
1388 for (i = 1, arg = DECL_CHAIN (a); i < nargs; i++, arg = DECL_CHAIN (arg))
1389 VEC_quick_push (tree, vargs, arg);
1390 call = gimple_build_call_vec (build_fold_addr_expr_loc (0, alias), vargs);
1391 VEC_free (tree, heap, vargs);
1392 gimple_call_set_cannot_inline (call, true);
1393 gimple_call_set_from_thunk (call, true);
1395 gimple_call_set_lhs (call, restmp);
1396 gsi_insert_after (&bsi, call, GSI_NEW_STMT);
1397 mark_symbols_for_renaming (call);
1398 find_referenced_vars_in (call);
1401 if (restmp && !this_adjusting)
1403 tree true_label = NULL_TREE;
1405 if (TREE_CODE (TREE_TYPE (restmp)) == POINTER_TYPE)
1408 /* If the return type is a pointer, we need to
1409 protect against NULL. We know there will be an
1410 adjustment, because that's why we're emitting a
1412 then_bb = create_basic_block (NULL, (void *) 0, bb);
1413 return_bb = create_basic_block (NULL, (void *) 0, then_bb);
1414 else_bb = create_basic_block (NULL, (void *) 0, else_bb);
1415 remove_edge (single_succ_edge (bb));
1416 true_label = gimple_block_label (then_bb);
1417 stmt = gimple_build_cond (NE_EXPR, restmp,
1418 fold_convert (TREE_TYPE (restmp),
1420 NULL_TREE, NULL_TREE);
1421 gsi_insert_after (&bsi, stmt, GSI_NEW_STMT);
1422 make_edge (bb, then_bb, EDGE_TRUE_VALUE);
1423 make_edge (bb, else_bb, EDGE_FALSE_VALUE);
1424 make_edge (return_bb, EXIT_BLOCK_PTR, 0);
1425 make_edge (then_bb, return_bb, EDGE_FALLTHRU);
1426 make_edge (else_bb, return_bb, EDGE_FALLTHRU);
1427 bsi = gsi_last_bb (then_bb);
1430 restmp = thunk_adjust (&bsi, restmp, /*this_adjusting=*/0,
1431 fixed_offset, virtual_offset);
1435 bsi = gsi_last_bb (else_bb);
1436 stmt = gimple_build_assign (restmp, fold_convert (TREE_TYPE (restmp),
1437 integer_zero_node));
1438 gsi_insert_after (&bsi, stmt, GSI_NEW_STMT);
1439 bsi = gsi_last_bb (return_bb);
1443 gimple_call_set_tail (call, true);
1445 /* Build return value. */
1446 ret = gimple_build_return (restmp);
1447 gsi_insert_after (&bsi, ret, GSI_NEW_STMT);
1449 delete_unreachable_blocks ();
1450 update_ssa (TODO_update_ssa);
1452 cgraph_remove_same_body_alias (node);
1453 /* Since we want to emit the thunk, we explicitly mark its name as
1455 cgraph_add_new_function (thunk_fndecl, true);
1456 bitmap_obstack_release (NULL);
1458 current_function_decl = NULL;
1461 /* Expand function specified by NODE. */
1464 cgraph_expand_function (struct cgraph_node *node)
1466 tree decl = node->decl;
1468 /* We ought to not compile any inline clones. */
1469 gcc_assert (!node->global.inlined_to);
1471 announce_function (decl);
1473 if (node->same_body)
1475 struct cgraph_node *alias, *next;
1476 bool saved_alias = node->alias;
1477 for (alias = node->same_body;
1478 alias && alias->next; alias = alias->next)
1480 /* Walk aliases in the order they were created; it is possible that
1481 thunks reffers to the aliases made earlier. */
1482 for (; alias; alias = next)
1484 next = alias->previous;
1485 if (!alias->thunk.thunk_p)
1486 assemble_alias (alias->decl,
1487 DECL_ASSEMBLER_NAME (alias->thunk.alias));
1489 assemble_thunk (alias);
1491 node->alias = saved_alias;
1492 cgraph_process_new_functions ();
1495 gcc_assert (node->lowered);
1497 /* Generate RTL for the body of DECL. */
1498 tree_rest_of_compilation (decl);
1500 /* Make sure that BE didn't give up on compiling. */
1501 gcc_assert (TREE_ASM_WRITTEN (decl));
1502 current_function_decl = NULL;
1503 gcc_assert (!cgraph_preserve_function_body_p (decl));
1504 cgraph_release_function_body (node);
1505 /* Eliminate all call edges. This is important so the GIMPLE_CALL no longer
1506 points to the dead function body. */
1507 cgraph_node_remove_callees (node);
1509 cgraph_function_flags_ready = true;
1512 /* Return true when CALLER_DECL should be inlined into CALLEE_DECL. */
1515 cgraph_inline_p (struct cgraph_edge *e, cgraph_inline_failed_t *reason)
1517 *reason = e->inline_failed;
1518 return !e->inline_failed;
1523 /* Expand all functions that must be output.
1525 Attempt to topologically sort the nodes so function is output when
1526 all called functions are already assembled to allow data to be
1527 propagated across the callgraph. Use a stack to get smaller distance
1528 between a function and its callees (later we may choose to use a more
1529 sophisticated algorithm for function reordering; we will likely want
1530 to use subsections to make the output functions appear in top-down
1534 cgraph_expand_all_functions (void)
1536 struct cgraph_node *node;
1537 struct cgraph_node **order = XCNEWVEC (struct cgraph_node *, cgraph_n_nodes);
1538 int order_pos, new_order_pos = 0;
1541 order_pos = cgraph_postorder (order);
1542 gcc_assert (order_pos == cgraph_n_nodes);
1544 /* Garbage collector may remove inline clones we eliminate during
1545 optimization. So we must be sure to not reference them. */
1546 for (i = 0; i < order_pos; i++)
1547 if (order[i]->process)
1548 order[new_order_pos++] = order[i];
1550 for (i = new_order_pos - 1; i >= 0; i--)
1555 gcc_assert (node->reachable);
1557 cgraph_expand_function (node);
1560 cgraph_process_new_functions ();
1566 /* This is used to sort the node types by the cgraph order number. */
1568 enum cgraph_order_sort_kind
1570 ORDER_UNDEFINED = 0,
1576 struct cgraph_order_sort
1578 enum cgraph_order_sort_kind kind;
1581 struct cgraph_node *f;
1582 struct varpool_node *v;
1583 struct cgraph_asm_node *a;
1587 /* Output all functions, variables, and asm statements in the order
1588 according to their order fields, which is the order in which they
1589 appeared in the file. This implements -fno-toplevel-reorder. In
1590 this mode we may output functions and variables which don't really
1591 need to be output. */
1594 cgraph_output_in_order (void)
1597 struct cgraph_order_sort *nodes;
1599 struct cgraph_node *pf;
1600 struct varpool_node *pv;
1601 struct cgraph_asm_node *pa;
1604 nodes = XCNEWVEC (struct cgraph_order_sort, max);
1606 varpool_analyze_pending_decls ();
1608 for (pf = cgraph_nodes; pf; pf = pf->next)
1613 gcc_assert (nodes[i].kind == ORDER_UNDEFINED);
1614 nodes[i].kind = ORDER_FUNCTION;
1619 for (pv = varpool_nodes_queue; pv; pv = pv->next_needed)
1622 gcc_assert (nodes[i].kind == ORDER_UNDEFINED);
1623 nodes[i].kind = ORDER_VAR;
1627 for (pa = cgraph_asm_nodes; pa; pa = pa->next)
1630 gcc_assert (nodes[i].kind == ORDER_UNDEFINED);
1631 nodes[i].kind = ORDER_ASM;
1635 /* In toplevel reorder mode we output all statics; mark them as needed. */
1636 for (i = 0; i < max; ++i)
1638 if (nodes[i].kind == ORDER_VAR)
1640 varpool_mark_needed_node (nodes[i].u.v);
1643 varpool_empty_needed_queue ();
1645 for (i = 0; i < max; ++i)
1647 switch (nodes[i].kind)
1649 case ORDER_FUNCTION:
1650 nodes[i].u.f->process = 0;
1651 cgraph_expand_function (nodes[i].u.f);
1655 varpool_assemble_decl (nodes[i].u.v);
1659 assemble_asm (nodes[i].u.a->asm_str);
1662 case ORDER_UNDEFINED:
1670 cgraph_asm_nodes = NULL;
1674 /* Return true when function body of DECL still needs to be kept around
1675 for later re-use. */
1677 cgraph_preserve_function_body_p (tree decl)
1679 struct cgraph_node *node;
1681 gcc_assert (cgraph_global_info_ready);
1682 /* Look if there is any clone around. */
1683 node = cgraph_node (decl);
1693 current_function_decl = NULL;
1694 gimple_register_cfg_hooks ();
1695 bitmap_obstack_initialize (NULL);
1697 invoke_plugin_callbacks (PLUGIN_ALL_IPA_PASSES_START, NULL);
1700 execute_ipa_pass_list (all_small_ipa_passes);
1702 /* If pass_all_early_optimizations was not scheduled, the state of
1703 the cgraph will not be properly updated. Update it now. */
1704 if (cgraph_state < CGRAPH_STATE_IPA_SSA)
1705 cgraph_state = CGRAPH_STATE_IPA_SSA;
1709 /* Generate coverage variables and constructors. */
1712 /* Process new functions added. */
1714 current_function_decl = NULL;
1715 cgraph_process_new_functions ();
1717 execute_ipa_summary_passes
1718 ((struct ipa_opt_pass_d *) all_regular_ipa_passes);
1721 /* Some targets need to handle LTO assembler output specially. */
1722 if (flag_generate_lto)
1723 targetm.asm_out.lto_start ();
1725 execute_ipa_summary_passes ((struct ipa_opt_pass_d *) all_lto_gen_passes);
1728 ipa_write_summaries ();
1730 if (flag_generate_lto)
1731 targetm.asm_out.lto_end ();
1734 execute_ipa_pass_list (all_regular_ipa_passes);
1735 invoke_plugin_callbacks (PLUGIN_ALL_IPA_PASSES_END, NULL);
1737 bitmap_obstack_release (NULL);
1741 /* Perform simple optimizations based on callgraph. */
1744 cgraph_optimize (void)
1749 #ifdef ENABLE_CHECKING
1753 /* Frontend may output common variables after the unit has been finalized.
1754 It is safe to deal with them here as they are always zero initialized. */
1755 varpool_analyze_pending_decls ();
1757 timevar_push (TV_CGRAPHOPT);
1758 if (pre_ipa_mem_report)
1760 fprintf (stderr, "Memory consumption before IPA\n");
1761 dump_memory_report (false);
1764 fprintf (stderr, "Performing interprocedural optimizations\n");
1765 cgraph_state = CGRAPH_STATE_IPA;
1767 /* Don't run the IPA passes if there was any error or sorry messages. */
1771 /* Do nothing else if any IPA pass found errors. */
1774 timevar_pop (TV_CGRAPHOPT);
1778 /* This pass remove bodies of extern inline functions we never inlined.
1779 Do this later so other IPA passes see what is really going on. */
1780 cgraph_remove_unreachable_nodes (false, dump_file);
1781 cgraph_global_info_ready = true;
1782 if (cgraph_dump_file)
1784 fprintf (cgraph_dump_file, "Optimized ");
1785 dump_cgraph (cgraph_dump_file);
1786 dump_varpool (cgraph_dump_file);
1788 if (post_ipa_mem_report)
1790 fprintf (stderr, "Memory consumption after IPA\n");
1791 dump_memory_report (false);
1793 timevar_pop (TV_CGRAPHOPT);
1795 /* Output everything. */
1796 (*debug_hooks->assembly_start) ();
1798 fprintf (stderr, "Assembling functions:\n");
1799 #ifdef ENABLE_CHECKING
1803 cgraph_materialize_all_clones ();
1804 cgraph_mark_functions_to_output ();
1806 cgraph_state = CGRAPH_STATE_EXPANSION;
1807 if (!flag_toplevel_reorder)
1808 cgraph_output_in_order ();
1811 cgraph_output_pending_asms ();
1813 cgraph_expand_all_functions ();
1814 varpool_remove_unreferenced_decls ();
1816 varpool_assemble_pending_decls ();
1818 cgraph_process_new_functions ();
1819 cgraph_state = CGRAPH_STATE_FINISHED;
1821 if (cgraph_dump_file)
1823 fprintf (cgraph_dump_file, "\nFinal ");
1824 dump_cgraph (cgraph_dump_file);
1826 #ifdef ENABLE_CHECKING
1828 /* Double check that all inline clones are gone and that all
1829 function bodies have been released from memory. */
1832 struct cgraph_node *node;
1833 bool error_found = false;
1835 for (node = cgraph_nodes; node; node = node->next)
1837 && (node->global.inlined_to
1838 || gimple_has_body_p (node->decl)))
1841 dump_cgraph_node (stderr, node);
1844 internal_error ("nodes with unreleased memory found");
1852 if (!cgraph_dump_file)
1853 cgraph_dump_file = dump_begin (TDI_cgraph, NULL);
1856 /* The edges representing the callers of the NEW_VERSION node were
1857 fixed by cgraph_function_versioning (), now the call_expr in their
1858 respective tree code should be updated to call the NEW_VERSION. */
1861 update_call_expr (struct cgraph_node *new_version)
1863 struct cgraph_edge *e;
1865 gcc_assert (new_version);
1867 /* Update the call expr on the edges to call the new version. */
1868 for (e = new_version->callers; e; e = e->next_caller)
1870 struct function *inner_function = DECL_STRUCT_FUNCTION (e->caller->decl);
1871 gimple_call_set_fndecl (e->call_stmt, new_version->decl);
1872 maybe_clean_eh_stmt_fn (inner_function, e->call_stmt);
1877 /* Create a new cgraph node which is the new version of
1878 OLD_VERSION node. REDIRECT_CALLERS holds the callers
1879 edges which should be redirected to point to
1880 NEW_VERSION. ALL the callees edges of OLD_VERSION
1881 are cloned to the new version node. Return the new
1884 If non-NULL BLOCK_TO_COPY determine what basic blocks
1885 was copied to prevent duplications of calls that are dead
1888 static struct cgraph_node *
1889 cgraph_copy_node_for_versioning (struct cgraph_node *old_version,
1891 VEC(cgraph_edge_p,heap) *redirect_callers,
1894 struct cgraph_node *new_version;
1895 struct cgraph_edge *e;
1898 gcc_assert (old_version);
1900 new_version = cgraph_node (new_decl);
1902 new_version->analyzed = true;
1903 new_version->local = old_version->local;
1904 new_version->local.externally_visible = false;
1905 new_version->local.local = true;
1906 new_version->local.vtable_method = false;
1907 new_version->global = old_version->global;
1908 new_version->rtl = old_version->rtl;
1909 new_version->reachable = true;
1910 new_version->count = old_version->count;
1912 for (e = old_version->callees; e; e=e->next_callee)
1914 || bitmap_bit_p (bbs_to_copy, gimple_bb (e->call_stmt)->index))
1915 cgraph_clone_edge (e, new_version, e->call_stmt,
1916 e->lto_stmt_uid, REG_BR_PROB_BASE,
1918 e->loop_nest, true);
1919 for (e = old_version->indirect_calls; e; e=e->next_callee)
1921 || bitmap_bit_p (bbs_to_copy, gimple_bb (e->call_stmt)->index))
1922 cgraph_clone_edge (e, new_version, e->call_stmt,
1923 e->lto_stmt_uid, REG_BR_PROB_BASE,
1925 e->loop_nest, true);
1926 FOR_EACH_VEC_ELT (cgraph_edge_p, redirect_callers, i, e)
1928 /* Redirect calls to the old version node to point to its new
1930 cgraph_redirect_edge_callee (e, new_version);
1936 /* Perform function versioning.
1937 Function versioning includes copying of the tree and
1938 a callgraph update (creating a new cgraph node and updating
1939 its callees and callers).
1941 REDIRECT_CALLERS varray includes the edges to be redirected
1944 TREE_MAP is a mapping of tree nodes we want to replace with
1945 new ones (according to results of prior analysis).
1946 OLD_VERSION_NODE is the node that is versioned.
1947 It returns the new version's cgraph node.
1948 If non-NULL ARGS_TO_SKIP determine function parameters to remove
1950 If non-NULL BLOCK_TO_COPY determine what basic blocks to copy.
1951 If non_NULL NEW_ENTRY determine new entry BB of the clone. */
1953 struct cgraph_node *
1954 cgraph_function_versioning (struct cgraph_node *old_version_node,
1955 VEC(cgraph_edge_p,heap) *redirect_callers,
1956 VEC (ipa_replace_map_p,gc)* tree_map,
1957 bitmap args_to_skip,
1959 basic_block new_entry_block,
1960 const char *clone_name)
1962 tree old_decl = old_version_node->decl;
1963 struct cgraph_node *new_version_node = NULL;
1966 if (!tree_versionable_function_p (old_decl))
1969 /* Make a new FUNCTION_DECL tree node for the
1972 new_decl = copy_node (old_decl);
1974 new_decl = build_function_decl_skip_args (old_decl, args_to_skip);
1976 /* Generate a new name for the new version. */
1977 DECL_NAME (new_decl) = clone_function_name (old_decl, clone_name);
1978 SET_DECL_ASSEMBLER_NAME (new_decl, DECL_NAME (new_decl));
1979 SET_DECL_RTL (new_decl, NULL);
1981 /* Create the new version's call-graph node.
1982 and update the edges of the new node. */
1984 cgraph_copy_node_for_versioning (old_version_node, new_decl,
1985 redirect_callers, bbs_to_copy);
1987 /* Copy the OLD_VERSION_NODE function tree to the new version. */
1988 tree_function_versioning (old_decl, new_decl, tree_map, false, args_to_skip,
1989 bbs_to_copy, new_entry_block);
1991 /* Update the new version's properties.
1992 Make The new version visible only within this translation unit. Make sure
1993 that is not weak also.
1994 ??? We cannot use COMDAT linkage because there is no
1995 ABI support for this. */
1996 cgraph_make_decl_local (new_version_node->decl);
1997 DECL_VIRTUAL_P (new_version_node->decl) = 0;
1998 new_version_node->local.externally_visible = 0;
1999 new_version_node->local.local = 1;
2000 new_version_node->lowered = true;
2002 /* Update the call_expr on the edges to call the new version node. */
2003 update_call_expr (new_version_node);
2005 cgraph_call_function_insertion_hooks (new_version_node);
2006 return new_version_node;
2009 /* Produce separate function body for inline clones so the offline copy can be
2010 modified without affecting them. */
2011 struct cgraph_node *
2012 save_inline_function_body (struct cgraph_node *node)
2014 struct cgraph_node *first_clone, *n;
2016 gcc_assert (node == cgraph_node (node->decl));
2018 cgraph_lower_function (node);
2020 first_clone = node->clones;
2022 first_clone->decl = copy_node (node->decl);
2023 cgraph_insert_node_to_hashtable (first_clone);
2024 gcc_assert (first_clone == cgraph_node (first_clone->decl));
2025 if (first_clone->next_sibling_clone)
2027 for (n = first_clone->next_sibling_clone; n->next_sibling_clone; n = n->next_sibling_clone)
2028 n->clone_of = first_clone;
2029 n->clone_of = first_clone;
2030 n->next_sibling_clone = first_clone->clones;
2031 if (first_clone->clones)
2032 first_clone->clones->prev_sibling_clone = n;
2033 first_clone->clones = first_clone->next_sibling_clone;
2034 first_clone->next_sibling_clone->prev_sibling_clone = NULL;
2035 first_clone->next_sibling_clone = NULL;
2036 gcc_assert (!first_clone->prev_sibling_clone);
2038 first_clone->clone_of = NULL;
2039 node->clones = NULL;
2041 if (first_clone->clones)
2042 for (n = first_clone->clones; n != first_clone;)
2044 gcc_assert (n->decl == node->decl);
2045 n->decl = first_clone->decl;
2048 else if (n->next_sibling_clone)
2049 n = n->next_sibling_clone;
2052 while (n != first_clone && !n->next_sibling_clone)
2054 if (n != first_clone)
2055 n = n->next_sibling_clone;
2059 /* Copy the OLD_VERSION_NODE function tree to the new version. */
2060 tree_function_versioning (node->decl, first_clone->decl, NULL, true, NULL,
2063 DECL_EXTERNAL (first_clone->decl) = 0;
2064 DECL_COMDAT_GROUP (first_clone->decl) = NULL_TREE;
2065 TREE_PUBLIC (first_clone->decl) = 0;
2066 DECL_COMDAT (first_clone->decl) = 0;
2067 VEC_free (ipa_opt_pass, heap,
2068 first_clone->ipa_transforms_to_apply);
2069 first_clone->ipa_transforms_to_apply = NULL;
2071 #ifdef ENABLE_CHECKING
2072 verify_cgraph_node (first_clone);
2077 /* Given virtual clone, turn it into actual clone. */
2079 cgraph_materialize_clone (struct cgraph_node *node)
2081 bitmap_obstack_initialize (NULL);
2082 #ifdef ENABLE_CHECKING
2083 node->former_clone_of = node->clone_of->decl;
2084 if (node->clone_of->former_clone_of)
2085 node->former_clone_of = node->clone_of->former_clone_of;
2087 /* Copy the OLD_VERSION_NODE function tree to the new version. */
2088 tree_function_versioning (node->clone_of->decl, node->decl,
2089 node->clone.tree_map, true,
2090 node->clone.args_to_skip, NULL, NULL);
2091 if (cgraph_dump_file)
2093 dump_function_to_file (node->clone_of->decl, cgraph_dump_file, dump_flags);
2094 dump_function_to_file (node->decl, cgraph_dump_file, dump_flags);
2097 /* Function is no longer clone. */
2098 if (node->next_sibling_clone)
2099 node->next_sibling_clone->prev_sibling_clone = node->prev_sibling_clone;
2100 if (node->prev_sibling_clone)
2101 node->prev_sibling_clone->next_sibling_clone = node->next_sibling_clone;
2103 node->clone_of->clones = node->next_sibling_clone;
2104 node->next_sibling_clone = NULL;
2105 node->prev_sibling_clone = NULL;
2106 if (!node->clone_of->analyzed && !node->clone_of->clones)
2108 cgraph_release_function_body (node->clone_of);
2109 cgraph_node_remove_callees (node->clone_of);
2110 ipa_remove_all_references (&node->clone_of->ref_list);
2112 node->clone_of = NULL;
2113 bitmap_obstack_release (NULL);
2116 /* If necessary, change the function declaration in the call statement
2117 associated with E so that it corresponds to the edge callee. */
2120 cgraph_redirect_edge_call_stmt_to_callee (struct cgraph_edge *e)
2122 tree decl = gimple_call_fndecl (e->call_stmt);
2124 #ifdef ENABLE_CHECKING
2125 struct cgraph_node *node;
2128 if (e->indirect_unknown_callee
2129 || decl == e->callee->decl
2130 /* Don't update call from same body alias to the real function. */
2131 || (decl && cgraph_get_node (decl) == cgraph_get_node (e->callee->decl)))
2132 return e->call_stmt;
2134 #ifdef ENABLE_CHECKING
2137 node = cgraph_get_node (decl);
2138 gcc_assert (!node || !node->clone.combined_args_to_skip);
2142 if (cgraph_dump_file)
2144 fprintf (cgraph_dump_file, "updating call of %s/%i -> %s/%i: ",
2145 cgraph_node_name (e->caller), e->caller->uid,
2146 cgraph_node_name (e->callee), e->callee->uid);
2147 print_gimple_stmt (cgraph_dump_file, e->call_stmt, 0, dump_flags);
2148 if (e->callee->clone.combined_args_to_skip)
2150 fprintf (cgraph_dump_file, " combined args to skip: ");
2151 dump_bitmap (cgraph_dump_file,
2152 e->callee->clone.combined_args_to_skip);
2156 if (e->callee->clone.combined_args_to_skip)
2158 gimple_stmt_iterator gsi;
2161 = gimple_call_copy_skip_args (e->call_stmt,
2162 e->callee->clone.combined_args_to_skip);
2163 gimple_call_set_fndecl (new_stmt, e->callee->decl);
2165 if (gimple_vdef (new_stmt)
2166 && TREE_CODE (gimple_vdef (new_stmt)) == SSA_NAME)
2167 SSA_NAME_DEF_STMT (gimple_vdef (new_stmt)) = new_stmt;
2169 gsi = gsi_for_stmt (e->call_stmt);
2170 gsi_replace (&gsi, new_stmt, false);
2171 if (maybe_clean_or_replace_eh_stmt (e->call_stmt, new_stmt))
2172 gimple_purge_dead_eh_edges (gimple_bb (new_stmt));
2176 new_stmt = e->call_stmt;
2177 gimple_call_set_fndecl (new_stmt, e->callee->decl);
2178 update_stmt (new_stmt);
2179 if (maybe_clean_eh_stmt (new_stmt))
2180 gimple_purge_dead_eh_edges (gimple_bb (new_stmt));
2183 cgraph_set_call_stmt_including_clones (e->caller, e->call_stmt, new_stmt);
2185 if (cgraph_dump_file)
2187 fprintf (cgraph_dump_file, " updated to:");
2188 print_gimple_stmt (cgraph_dump_file, e->call_stmt, 0, dump_flags);
2193 /* Once all functions from compilation unit are in memory, produce all clones
2194 and update all calls. We might also do this on demand if we don't want to
2195 bring all functions to memory prior compilation, but current WHOPR
2196 implementation does that and it is is bit easier to keep everything right in
2199 cgraph_materialize_all_clones (void)
2201 struct cgraph_node *node;
2202 bool stabilized = false;
2204 if (cgraph_dump_file)
2205 fprintf (cgraph_dump_file, "Materializing clones\n");
2206 #ifdef ENABLE_CHECKING
2210 /* We can also do topological order, but number of iterations should be
2211 bounded by number of IPA passes since single IPA pass is probably not
2212 going to create clones of clones it created itself. */
2216 for (node = cgraph_nodes; node; node = node->next)
2218 if (node->clone_of && node->decl != node->clone_of->decl
2219 && !gimple_has_body_p (node->decl))
2221 if (gimple_has_body_p (node->clone_of->decl))
2223 if (cgraph_dump_file)
2225 fprintf (cgraph_dump_file, "clonning %s to %s\n",
2226 cgraph_node_name (node->clone_of),
2227 cgraph_node_name (node));
2228 if (node->clone.tree_map)
2231 fprintf (cgraph_dump_file, " replace map: ");
2232 for (i = 0; i < VEC_length (ipa_replace_map_p,
2233 node->clone.tree_map);
2236 struct ipa_replace_map *replace_info;
2237 replace_info = VEC_index (ipa_replace_map_p,
2238 node->clone.tree_map,
2240 print_generic_expr (cgraph_dump_file, replace_info->old_tree, 0);
2241 fprintf (cgraph_dump_file, " -> ");
2242 print_generic_expr (cgraph_dump_file, replace_info->new_tree, 0);
2243 fprintf (cgraph_dump_file, "%s%s;",
2244 replace_info->replace_p ? "(replace)":"",
2245 replace_info->ref_p ? "(ref)":"");
2247 fprintf (cgraph_dump_file, "\n");
2249 if (node->clone.args_to_skip)
2251 fprintf (cgraph_dump_file, " args_to_skip: ");
2252 dump_bitmap (cgraph_dump_file, node->clone.args_to_skip);
2254 if (node->clone.args_to_skip)
2256 fprintf (cgraph_dump_file, " combined_args_to_skip:");
2257 dump_bitmap (cgraph_dump_file, node->clone.combined_args_to_skip);
2260 cgraph_materialize_clone (node);
2266 for (node = cgraph_nodes; node; node = node->next)
2267 if (!node->analyzed && node->callees)
2268 cgraph_node_remove_callees (node);
2269 if (cgraph_dump_file)
2270 fprintf (cgraph_dump_file, "Materialization Call site updates done.\n");
2271 #ifdef ENABLE_CHECKING
2274 cgraph_remove_unreachable_nodes (false, cgraph_dump_file);
2277 #include "gt-cgraphunit.h"