1 /* Callgraph based intraprocedural optimizations.
2 Copyright (C) 2003, 2004 Free Software Foundation, Inc.
3 Contributed by Jan Hubicka
5 This file is part of GCC.
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 2, or (at your option) any later
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
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to the Free
19 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
22 /* This module implements main driver of compilation process as well as
23 few basic intraprocedural 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 function.)
37 - cgraph_varpool_finalize_variable
39 This function has same behaviour as the above but is used for static
42 - cgraph_finalize_compilation_unit
44 This function is called once compilation unit is finalized and it will
47 In the unit-at-a-time the call-graph construction and local function
48 analysis takes place here. Bodies of unreachable functions are released
49 to conserve memory usage.
51 ??? The compilation unit in this point of view should be compilation
52 unit as defined by the language - for instance C frontend allows multiple
53 compilation units to be parsed at once and it should call function each
54 time parsing is done so we save memory.
58 In this unit-at-a-time compilation the intra procedural analysis takes
59 place here. In particular the static functions whose address is never
60 taken are marked as local. Backend can then use this information to
61 modify calling conventions, do better inlining or similar optimizations.
63 - cgraph_assemble_pending_functions
64 - cgraph_varpool_assemble_pending_variables
66 In non-unit-at-a-time mode these functions can be used to force compilation
67 of functions or variables that are known to be needed at given stage
70 - cgraph_mark_needed_node
71 - cgraph_varpool_mark_needed_node
73 When function or variable is referenced by some hidden way (for instance
74 via assembly code and marked by attribute "used"), the call-graph data structure
75 must be updated accordingly by this function.
77 - analyze_expr callback
79 This function is responsible for lowering tree nodes not understood by
80 generic code into understandable ones or alternatively marking
81 callgraph and varpool nodes referenced by the as needed.
83 ??? On the tree-ssa genericizing should take place here and we will avoid
84 need for these hooks (replacing them by genericizing hook)
86 - expand_function callback
88 This function is used to expand function and pass it into RTL back-end.
89 Front-end should not make any assumptions about when this function can be
90 called. In particular cgraph_assemble_pending_functions,
91 cgraph_varpool_assemble_pending_variables, cgraph_finalize_function,
92 cgraph_varpool_finalize_function, cgraph_optimize can cause arbitrarily
93 previously finalized functions to be expanded.
95 We implement two compilation modes.
97 - unit-at-a-time: In this mode analyzing of all functions is deferred
98 to cgraph_finalize_compilation_unit and expansion into cgraph_optimize.
100 In cgraph_finalize_compilation_unit the reachable functions are
101 analyzed. During analysis the call-graph edges from reachable
102 functions are constructed and their destinations are marked as
103 reachable. References to functions and variables are discovered too
104 and variables found to be needed output to the assembly file. Via
105 mark_referenced call in assemble_variable functions referenced by
106 static variables are noticed too.
108 The intra-procedural information is produced and it's existence
109 indicated by global_info_ready. Once this flag is set it is impossible
110 to change function from !reachable to reachable and thus
111 assemble_variable no longer call mark_referenced.
113 Finally the call-graph is topologically sorted and all reachable functions
114 that has not been completely inlined or are not external are output.
116 ??? It is possible that reference to function or variable is optimized
117 out. We can not deal with this nicely because topological order is not
118 suitable for it. For tree-ssa we may consider another pass doing
119 optimization and re-discovering reachable functions.
121 ??? Reorganize code so variables are output very last and only if they
122 really has been referenced by produced code, so we catch more cases
123 where reference has been optimized out.
127 All functions are variables are output as early as possible to conserve
128 memory consumption. This may or may not result in less memory used but
129 it is still needed for some legacy code that rely on particular ordering
130 of things output from the compiler.
132 Varpool data structures are not used and variables are output directly.
134 Functions are output early using call of
135 cgraph_assemble_pending_function from cgraph_finalize_function. The
136 decision on whether function is needed is made more conservative so
137 uninlininable static functions are needed too. During the call-graph
138 construction the edge destinations are not marked as reachable and it
139 is completely relied upn assemble_variable to mark them.
141 Inlining decision heuristics
142 ??? Move this to separate file after tree-ssa merge.
144 We separate inlining decisions from the inliner itself and store it
145 inside callgraph as so called inline plan. Reffer to cgraph.c
146 documentation about particular representation of inline plans in the
149 The implementation of particular heuristics is separated from
150 the rest of code to make it easier to replace it with more complicated
151 implementation in the future. The rest of inlining code acts as a
152 library aimed to modify the callgraph and verify that the parameters
153 on code size growth fits.
155 To mark given call inline, use cgraph_mark_inline function, the
156 verification is performed by cgraph_default_inline_p and
157 cgraph_check_inline_limits.
159 The heuristics implements simple knapsack style algorithm ordering
160 all functions by their "profitability" (estimated by code size growth)
161 and inlining them in priority order.
163 cgraph_decide_inlining implements heuristics taking whole callgraph
164 into account, while cgraph_decide_inlining_incrementally considers
165 only one function at a time and is used in non-unit-at-a-time mode. */
168 #include "coretypes.h"
171 #include "tree-inline.h"
172 #include "langhooks.h"
180 #include "diagnostic.h"
184 #include "c-common.h"
186 #include "function.h"
188 #define INSNS_PER_CALL 10
190 static void cgraph_expand_all_functions (void);
191 static void cgraph_mark_functions_to_output (void);
192 static void cgraph_expand_function (struct cgraph_node *);
193 static tree record_call_1 (tree *, int *, void *);
194 static void cgraph_mark_local_functions (void);
195 static bool cgraph_default_inline_p (struct cgraph_node *n);
196 static void cgraph_analyze_function (struct cgraph_node *node);
197 static void cgraph_decide_inlining_incrementally (struct cgraph_node *);
199 /* Statistics we collect about inlining algorithm. */
200 static int ncalls_inlined;
201 static int nfunctions_inlined;
202 static int initial_insns;
203 static int overall_insns;
205 /* Records tree nodes seen in cgraph_create_edges. Simply using
206 walk_tree_without_duplicates doesn't guarantee each node is visited
207 once because it gets a new htab upon each recursive call from
209 static htab_t visited_nodes;
211 /* Determine if function DECL is needed. That is, visible to something
212 either outside this translation unit, something magic in the system
213 configury, or (if not doing unit-at-a-time) to something we havn't
217 decide_is_function_needed (struct cgraph_node *node, tree decl)
219 /* If we decided it was needed before, but at the time we didn't have
220 the body of the function available, then it's still needed. We have
221 to go back and re-check its dependencies now. */
225 /* Externally visible functions must be output. The exception is
226 COMDAT functions that must be output only when they are needed. */
227 if (TREE_PUBLIC (decl) && !DECL_COMDAT (decl) && !DECL_EXTERNAL (decl))
230 /* Constructors and destructors are reachable from the runtime by
232 if (DECL_STATIC_CONSTRUCTOR (decl) || DECL_STATIC_DESTRUCTOR (decl))
235 /* If the user told us it is used, then it must be so. */
236 if (lookup_attribute ("used", DECL_ATTRIBUTES (decl)))
239 /* ??? If the assembler name is set by hand, it is possible to assemble
240 the name later after finalizing the function and the fact is noticed
241 in assemble_name then. This is arguably a bug. */
242 if (DECL_ASSEMBLER_NAME_SET_P (decl)
243 && TREE_SYMBOL_REFERENCED (DECL_ASSEMBLER_NAME (decl)))
246 if (flag_unit_at_a_time)
249 /* If not doing unit at a time, then we'll only defer this function
250 if its marked for inlining. Otherwise we want to emit it now. */
252 /* "extern inline" functions are never output locally. */
253 if (DECL_EXTERNAL (decl))
255 /* We want to emit COMDAT functions only when absolutely necessary. */
256 if (DECL_COMDAT (decl))
258 if (!DECL_INLINE (decl)
259 || (!node->local.disregard_inline_limits
260 /* When declared inline, defer even the uninlinable functions.
261 This allows them to be eliminated when unused. */
262 && !DECL_DECLARED_INLINE_P (decl)
263 && (!node->local.inlinable || !cgraph_default_inline_p (node))))
269 /* When not doing unit-at-a-time, output all functions enqueued.
270 Return true when such a functions were found. */
273 cgraph_assemble_pending_functions (void)
277 if (flag_unit_at_a_time)
280 while (cgraph_nodes_queue)
282 struct cgraph_node *n = cgraph_nodes_queue;
284 cgraph_nodes_queue = cgraph_nodes_queue->next_needed;
285 n->next_needed = NULL;
286 if (!n->origin && !n->global.inlined_to && !DECL_EXTERNAL (n->decl))
288 cgraph_expand_function (n);
296 /* DECL has been parsed. Take it, queue it, compile it at the whim of the
297 logic in effect. If NESTED is true, then our caller cannot stand to have
298 the garbage collector run at the moment. We would need to either create
299 a new GC context, or just not compile right now. */
302 cgraph_finalize_function (tree decl, bool nested)
304 struct cgraph_node *node = cgraph_node (decl);
306 if (node->local.finalized)
308 /* As an GCC extension we allow redefinition of the function. The
309 semantics when both copies of bodies differ is not well defined.
310 We replace the old body with new body so in unit at a time mode
311 we always use new body, while in normal mode we may end up with
312 old body inlined into some functions and new body expanded and
315 ??? It may make more sense to use one body for inlining and other
316 body for expanding the function but this is difficult to do. */
318 /* If node->output is set, then this is a unit-at-a-time compilation
319 and we have already begun whole-unit analysis. This is *not*
320 testing for whether we've already emitted the function. That
321 case can be sort-of legitimately seen with real function
322 redefinition errors. I would argue that the front end should
323 never present us with such a case, but don't enforce that for now. */
327 /* Reset our datastructures so we can analyze the function again. */
328 memset (&node->local, 0, sizeof (node->local));
329 memset (&node->global, 0, sizeof (node->global));
330 memset (&node->rtl, 0, sizeof (node->rtl));
331 node->analyzed = false;
332 node->local.redefined_extern_inline = true;
333 while (node->callees)
334 cgraph_remove_edge (node->callees);
336 /* We may need to re-queue the node for assembling in case
337 we already proceeded it and ignored as not needed. */
338 if (node->reachable && !flag_unit_at_a_time)
340 struct cgraph_node *n;
342 for (n = cgraph_nodes_queue; n; n = n->next_needed)
350 notice_global_symbol (decl);
352 node->local.finalized = true;
354 /* If not unit at a time, then we need to create the call graph
355 now, so that called functions can be queued and emitted now. */
356 if (!flag_unit_at_a_time)
358 cgraph_analyze_function (node);
359 cgraph_decide_inlining_incrementally (node);
362 if (decide_is_function_needed (node, decl))
363 cgraph_mark_needed_node (node);
365 /* If not unit at a time, go ahead and emit everything we've found
366 to be reachable at this time. */
369 if (!cgraph_assemble_pending_functions ())
373 /* If we've not yet emitted decl, tell the debug info about it. */
374 if (!TREE_ASM_WRITTEN (decl))
375 (*debug_hooks->deferred_inline_function) (decl);
377 /* We will never really output the function body, clear the STRUCT_FUNCTION array
379 if (DECL_EXTERNAL (decl))
380 DECL_STRUCT_FUNCTION (decl) = NULL;
382 /* Possibly warn about unused parameters. */
383 if (warn_unused_parameter)
384 do_warn_unused_parameter (decl);
387 /* Walk tree and record all calls. Called via walk_tree. */
389 record_call_1 (tree *tp, int *walk_subtrees, void *data)
393 switch (TREE_CODE (t))
396 /* ??? Really, we should mark this decl as *potentially* referenced
397 by this function and re-examine whether the decl is actually used
398 after rtl has been generated. */
400 cgraph_varpool_mark_needed_node (cgraph_varpool_node (t));
404 if (flag_unit_at_a_time)
406 /* Record dereferences to the functions. This makes the
407 functions reachable unconditionally. */
408 tree decl = TREE_OPERAND (*tp, 0);
409 if (TREE_CODE (decl) == FUNCTION_DECL)
410 cgraph_mark_needed_node (cgraph_node (decl));
416 tree decl = get_callee_fndecl (*tp);
417 if (decl && TREE_CODE (decl) == FUNCTION_DECL)
419 cgraph_create_edge (data, cgraph_node (decl), *tp);
421 /* When we see a function call, we don't want to look at the
422 function reference in the ADDR_EXPR that is hanging from
423 the CALL_EXPR we're examining here, because we would
424 conclude incorrectly that the function's address could be
425 taken by something that is not a function call. So only
426 walk the function parameter list, skip the other subtrees. */
428 walk_tree (&TREE_OPERAND (*tp, 1), record_call_1, data,
436 /* Save some cycles by not walking types and declaration as we
437 won't find anything useful there anyway. */
438 if (DECL_P (*tp) || TYPE_P (*tp))
444 if ((unsigned int) TREE_CODE (t) >= LAST_AND_UNUSED_TREE_CODE)
445 return lang_hooks.callgraph.analyze_expr (tp, walk_subtrees, data);
452 /* Create cgraph edges for function calls inside BODY from NODE. */
455 cgraph_create_edges (struct cgraph_node *node, tree body)
457 /* The nodes we're interested in are never shared, so walk
458 the tree ignoring duplicates. */
459 visited_nodes = htab_create (37, htab_hash_pointer,
460 htab_eq_pointer, NULL);
461 walk_tree (&body, record_call_1, node, visited_nodes);
462 htab_delete (visited_nodes);
463 visited_nodes = NULL;
466 static bool error_found;
468 /* Callbrack of verify_cgraph_node. Check that all call_exprs have cgraph nodes. */
470 verify_cgraph_node_1 (tree *tp, int *walk_subtrees, void *data)
475 if (TREE_CODE (t) == CALL_EXPR && (decl = get_callee_fndecl (t)))
477 struct cgraph_edge *e = cgraph_edge (data, t);
482 error ("Shared call_expr:");
486 if (e->callee->decl != cgraph_node (decl)->decl)
488 error ("Edge points to wrong declaration:");
489 debug_tree (e->callee->decl);
490 fprintf (stderr," Instead of:");
497 error ("Missing callgraph edge for call expr:");
502 /* Save some cycles by not walking types and declaration as we
503 won't find anything useful there anyway. */
504 if (DECL_P (*tp) || TYPE_P (*tp))
511 /* Verify cgraph nodes of given cgraph node. */
513 verify_cgraph_node (struct cgraph_node *node)
515 struct cgraph_edge *e;
516 struct cgraph_node *main_clone;
518 timevar_push (TV_CGRAPH_VERIFY);
520 for (e = node->callees; e; e = e->next_callee)
523 error ("Aux field set for edge %s->%s",
524 cgraph_node_name (e->caller), cgraph_node_name (e->callee));
527 for (e = node->callers; e; e = e->next_caller)
529 if (!e->inline_failed)
531 if (node->global.inlined_to
532 != (e->caller->global.inlined_to
533 ? e->caller->global.inlined_to : e->caller))
535 error ("Inlined_to pointer is wrong");
538 if (node->callers->next_caller)
540 error ("Multiple inline callers");
545 if (node->global.inlined_to)
547 error ("Inlined_to pointer set for noninline callers");
551 if (!node->callers && node->global.inlined_to)
553 error ("Inlined_to pointer is set but no predecesors found");
556 if (node->global.inlined_to == node)
558 error ("Inlined_to pointer reffers to itself");
562 for (main_clone = cgraph_node (node->decl); main_clone;
563 main_clone = main_clone->next_clone)
564 if (main_clone == node)
568 error ("Node not found in DECL_ASSEMBLER_NAME hash");
573 && DECL_SAVED_TREE (node->decl) && !TREE_ASM_WRITTEN (node->decl)
574 && (!DECL_EXTERNAL (node->decl) || node->global.inlined_to))
576 walk_tree_without_duplicates (&DECL_SAVED_TREE (node->decl),
577 verify_cgraph_node_1, node);
578 for (e = node->callees; e; e = e->next_callee)
582 error ("Edge %s->%s has no corresponding call_expr",
583 cgraph_node_name (e->caller),
584 cgraph_node_name (e->callee));
592 dump_cgraph_node (stderr, node);
593 internal_error ("verify_cgraph_node failed.");
595 timevar_pop (TV_CGRAPH_VERIFY);
598 /* Verify whole cgraph structure. */
602 struct cgraph_node *node;
604 for (node = cgraph_nodes; node; node = node->next)
605 verify_cgraph_node (node);
608 /* Analyze the function scheduled to be output. */
610 cgraph_analyze_function (struct cgraph_node *node)
612 tree decl = node->decl;
613 struct cgraph_edge *e;
615 current_function_decl = decl;
617 /* First kill forward declaration so reverse inlining works properly. */
618 cgraph_create_edges (node, DECL_SAVED_TREE (decl));
620 node->local.inlinable = tree_inlinable_function_p (decl);
621 if (!node->local.self_insns)
622 node->local.self_insns
623 = lang_hooks.tree_inlining.estimate_num_insns (decl);
624 if (node->local.inlinable)
625 node->local.disregard_inline_limits
626 = lang_hooks.tree_inlining.disregard_inline_limits (decl);
627 for (e = node->callers; e; e = e->next_caller)
629 if (node->local.redefined_extern_inline)
630 e->inline_failed = N_("redefined extern inline functions are not "
631 "considered for inlining");
632 else if (!node->local.inlinable)
633 e->inline_failed = N_("function not inlinable");
635 e->inline_failed = N_("function not considered for inlining");
637 if (flag_really_no_inline && !node->local.disregard_inline_limits)
638 node->local.inlinable = 0;
639 /* Inlining characteristics are maintained by the cgraph_mark_inline. */
640 node->global.insns = node->local.self_insns;
642 node->analyzed = true;
643 current_function_decl = NULL;
646 /* Analyze the whole compilation unit once it is parsed completely. */
649 cgraph_finalize_compilation_unit (void)
651 struct cgraph_node *node;
653 if (!flag_unit_at_a_time)
655 cgraph_assemble_pending_functions ();
659 cgraph_varpool_assemble_pending_decls ();
661 fprintf (stderr, "\nAnalyzing compilation unit\n");
663 timevar_push (TV_CGRAPH);
664 if (cgraph_dump_file)
666 fprintf (cgraph_dump_file, "Initial entry points:");
667 for (node = cgraph_nodes; node; node = node->next)
668 if (node->needed && DECL_SAVED_TREE (node->decl))
669 fprintf (cgraph_dump_file, " %s", cgraph_node_name (node));
670 fprintf (cgraph_dump_file, "\n");
673 /* Propagate reachability flag and lower representation of all reachable
674 functions. In the future, lowering will introduce new functions and
675 new entry points on the way (by template instantiation and virtual
676 method table generation for instance). */
677 while (cgraph_nodes_queue)
679 struct cgraph_edge *edge;
680 tree decl = cgraph_nodes_queue->decl;
682 node = cgraph_nodes_queue;
683 cgraph_nodes_queue = cgraph_nodes_queue->next_needed;
684 node->next_needed = NULL;
686 /* ??? It is possible to create extern inline function and later using
687 weak alas attribute to kill its body. See
688 gcc.c-torture/compile/20011119-1.c */
689 if (!DECL_SAVED_TREE (decl))
692 if (node->analyzed || !node->reachable || !DECL_SAVED_TREE (decl))
695 cgraph_analyze_function (node);
697 for (edge = node->callees; edge; edge = edge->next_callee)
698 if (!edge->callee->reachable)
699 cgraph_mark_reachable_node (edge->callee);
701 cgraph_varpool_assemble_pending_decls ();
704 /* Collect entry points to the unit. */
706 if (cgraph_dump_file)
708 fprintf (cgraph_dump_file, "Unit entry points:");
709 for (node = cgraph_nodes; node; node = node->next)
710 if (node->needed && DECL_SAVED_TREE (node->decl))
711 fprintf (cgraph_dump_file, " %s", cgraph_node_name (node));
712 fprintf (cgraph_dump_file, "\n\nInitial ");
713 dump_cgraph (cgraph_dump_file);
716 if (cgraph_dump_file)
717 fprintf (cgraph_dump_file, "\nReclaiming functions:");
719 for (node = cgraph_nodes; node; node = node->next)
721 tree decl = node->decl;
723 if (!node->reachable && DECL_SAVED_TREE (decl))
725 if (cgraph_dump_file)
726 fprintf (cgraph_dump_file, " %s", cgraph_node_name (node));
727 cgraph_remove_node (node);
730 node->next_needed = NULL;
732 if (cgraph_dump_file)
734 fprintf (cgraph_dump_file, "\n\nReclaimed ");
735 dump_cgraph (cgraph_dump_file);
738 timevar_pop (TV_CGRAPH);
741 /* Figure out what functions we want to assemble. */
744 cgraph_mark_functions_to_output (void)
746 struct cgraph_node *node;
748 for (node = cgraph_nodes; node; node = node->next)
750 tree decl = node->decl;
751 struct cgraph_edge *e;
756 for (e = node->callers; e; e = e->next_caller)
757 if (e->inline_failed)
760 /* We need to output all local functions that are used and not
761 always inlined, as well as those that are reachable from
762 outside the current compilation unit. */
763 if (DECL_SAVED_TREE (decl)
764 && !node->global.inlined_to
766 || (e && node->reachable))
767 && !TREE_ASM_WRITTEN (decl) && !node->origin
768 && !DECL_EXTERNAL (decl))
770 /* We should've reclaimed all functions that are not needed. */
771 else if (!node->global.inlined_to && DECL_SAVED_TREE (decl)
772 && !node->origin && !DECL_EXTERNAL (decl))
774 dump_cgraph_node (stderr, node);
780 /* Expand function specified by NODE. */
783 cgraph_expand_function (struct cgraph_node *node)
785 tree decl = node->decl;
787 /* We ought to not compile any inline clones. */
788 if (node->global.inlined_to)
791 if (flag_unit_at_a_time)
792 announce_function (decl);
794 /* Generate RTL for the body of DECL. Nested functions are expanded
795 via lang_expand_decl_stmt. */
796 lang_hooks.callgraph.expand_function (decl);
797 if (DECL_DEFER_OUTPUT (decl))
800 /* Make sure that BE didn't gave up on compiling. */
801 if (!TREE_ASM_WRITTEN (node->decl)
802 && !(sorrycount || errorcount))
805 current_function_decl = NULL;
808 /* Fill array order with all nodes with output flag set in the reverse
809 topological order. */
812 cgraph_postorder (struct cgraph_node **order)
814 struct cgraph_node *node, *node2;
817 struct cgraph_edge *edge, last;
819 struct cgraph_node **stack =
820 xcalloc (cgraph_n_nodes, sizeof (struct cgraph_node *));
822 /* We have to deal with cycles nicely, so use a depth first traversal
823 output algorithm. Ignore the fact that some functions won't need
824 to be output and put them into order as well, so we get dependencies
825 right throughout inline functions. */
826 for (node = cgraph_nodes; node; node = node->next)
828 for (node = cgraph_nodes; node; node = node->next)
835 node->aux = node->callers;
838 while (node2->aux != &last)
841 if (edge->next_caller)
842 node2->aux = edge->next_caller;
845 if (!edge->caller->aux)
847 if (!edge->caller->callers)
848 edge->caller->aux = &last;
850 edge->caller->aux = edge->caller->callers;
851 stack[stack_size++] = node2;
852 node2 = edge->caller;
856 if (node2->aux == &last)
858 order[order_pos++] = node2;
860 node2 = stack[--stack_size];
870 /* Perform reachability analysis and reclaim all unreachable nodes.
871 This function also remove unneeded bodies of extern inline functions
872 and thus needs to be done only after inlining decisions has been made. */
874 cgraph_remove_unreachable_nodes (void)
876 struct cgraph_node *first = (void *) 1;
877 struct cgraph_node *node;
878 bool changed = false;
881 #ifdef ENABLE_CHECKING
884 if (cgraph_dump_file)
885 fprintf (cgraph_dump_file, "\nReclaiming functions:");
886 #ifdef ENABLE_CHECKING
887 for (node = cgraph_nodes; node; node = node->next)
891 for (node = cgraph_nodes; node; node = node->next)
892 if (node->needed && (!DECL_EXTERNAL (node->decl) || !node->analyzed))
900 /* Perform reachability analysis. As a special case do not consider
901 extern inline functions not inlined as live because we won't output
903 while (first != (void *) 1)
905 struct cgraph_edge *e;
909 for (e = node->callees; e; e = e->next_callee)
912 && (!e->inline_failed || !e->callee->analyzed
913 || !DECL_EXTERNAL (e->callee->decl)))
915 e->callee->aux = first;
920 /* Remove unreachable nodes. Extern inline functions need special care;
921 Unreachable extern inline functions shall be removed.
922 Reachable extern inline functions we never inlined shall get their bodies
924 Reachable extern inline functions we sometimes inlined will be turned into
925 unanalyzed nodes so they look like for true extern functions to the rest
926 of code. Body of such functions is released via remove_node once the
927 inline clones are eliminated. */
928 for (node = cgraph_nodes; node; node = node->next)
933 tree decl = node->decl;
935 if (DECL_STRUCT_FUNCTION (decl))
936 local_insns = node->local.self_insns;
939 if (cgraph_dump_file)
940 fprintf (cgraph_dump_file, " %s", cgraph_node_name (node));
941 if (!node->analyzed || !DECL_EXTERNAL (node->decl))
942 cgraph_remove_node (node);
945 struct cgraph_edge *e;
947 for (e = node->callers; e; e = e->next_caller)
950 if (e || node->needed)
952 struct cgraph_node *clone;
954 for (clone = node->next_clone; clone;
955 clone = clone->next_clone)
960 DECL_SAVED_TREE (node->decl) = NULL;
961 DECL_STRUCT_FUNCTION (node->decl) = NULL;
962 DECL_ARGUMENTS (node->decl) = NULL;
963 DECL_INITIAL (node->decl) = error_mark_node;
965 while (node->callees)
966 cgraph_remove_edge (node->callees);
967 node->analyzed = false;
970 cgraph_remove_node (node);
972 if (!DECL_SAVED_TREE (decl))
973 insns += local_insns;
977 for (node = cgraph_nodes; node; node = node->next)
979 if (cgraph_dump_file)
980 fprintf (cgraph_dump_file, "\nReclaimed %i insns", insns);
984 /* Estimate size of the function after inlining WHAT into TO. */
987 cgraph_estimate_size_after_inlining (int times, struct cgraph_node *to,
988 struct cgraph_node *what)
990 return (what->global.insns - INSNS_PER_CALL) * times + to->global.insns;
993 /* Estimate the growth caused by inlining NODE into all callees. */
996 cgraph_estimate_growth (struct cgraph_node *node)
999 struct cgraph_edge *e;
1001 for (e = node->callers; e; e = e->next_caller)
1002 if (e->inline_failed)
1003 growth += (cgraph_estimate_size_after_inlining (1, e->caller, node)
1004 - e->caller->global.insns);
1006 /* ??? Wrong for self recursive functions or cases where we decide to not
1007 inline for different reasons, but it is not big deal as in that case
1008 we will keep the body around, but we will also avoid some inlining. */
1009 if (!node->needed && !node->origin && !DECL_EXTERNAL (node->decl))
1010 growth -= node->global.insns;
1015 /* E is expected to be an edge being inlined. Clone destination node of
1016 the edge and redirect it to the new clone.
1017 DUPLICATE is used for bookeeping on whether we are actually creating new
1018 clones or re-using node originally representing out-of-line function call.
1021 cgraph_clone_inlined_nodes (struct cgraph_edge *e, bool duplicate)
1023 struct cgraph_node *n;
1025 /* We may elliminate the need for out-of-line copy to be output. In that
1026 case just go ahead and re-use it. */
1027 if (!e->callee->callers->next_caller
1028 && (!e->callee->needed || DECL_EXTERNAL (e->callee->decl))
1029 && !e->callee->origin
1031 && flag_unit_at_a_time)
1033 if (e->callee->global.inlined_to)
1035 if (!DECL_EXTERNAL (e->callee->decl))
1036 overall_insns -= e->callee->global.insns, nfunctions_inlined++;
1041 n = cgraph_clone_node (e->callee);
1042 cgraph_redirect_edge_callee (e, n);
1045 if (e->caller->global.inlined_to)
1046 e->callee->global.inlined_to = e->caller->global.inlined_to;
1048 e->callee->global.inlined_to = e->caller;
1050 /* Recursively clone all bodies. */
1051 for (e = e->callee->callees; e; e = e->next_callee)
1052 if (!e->inline_failed)
1053 cgraph_clone_inlined_nodes (e, duplicate);
1056 /* Mark edge E as inlined and update callgraph accordingly. */
1059 cgraph_mark_inline_edge (struct cgraph_edge *e)
1061 int old_insns = 0, new_insns = 0;
1062 struct cgraph_node *to = NULL, *what;
1064 if (!e->inline_failed)
1066 e->inline_failed = NULL;
1068 if (!e->callee->global.inlined && flag_unit_at_a_time)
1071 if (!cgraph_inline_hash)
1072 cgraph_inline_hash = htab_create_ggc (42, htab_hash_pointer,
1073 htab_eq_pointer, NULL);
1074 slot = htab_find_slot (cgraph_inline_hash,
1075 DECL_ASSEMBLER_NAME (e->callee->decl), INSERT);
1076 *slot = DECL_ASSEMBLER_NAME (e->callee->decl);
1078 e->callee->global.inlined = true;
1080 cgraph_clone_inlined_nodes (e, true);
1084 /* Now update size of caller and all functions caller is inlined into. */
1085 for (;e && !e->inline_failed; e = e->caller->callers)
1087 old_insns = e->caller->global.insns;
1088 new_insns = cgraph_estimate_size_after_inlining (1, e->caller,
1093 to->global.insns = new_insns;
1095 if (what->global.inlined_to != to)
1097 overall_insns += new_insns - old_insns;
1101 /* Mark all calls of EDGE->CALLEE inlined into EDGE->CALLER.
1102 Return following unredirected edge in the list of callers
1105 static struct cgraph_edge *
1106 cgraph_mark_inline (struct cgraph_edge *edge)
1108 struct cgraph_node *to = edge->caller;
1109 struct cgraph_node *what = edge->callee;
1110 struct cgraph_edge *e, *next;
1113 /* Look for all calls, mark them inline and clone recursively
1114 all inlined functions. */
1115 for (e = what->callers; e; e = next)
1117 next = e->next_caller;
1118 if (e->caller == to && e->inline_failed)
1120 cgraph_mark_inline_edge (e);
1131 /* Return false when inlining WHAT into TO is not good idea
1132 as it would cause too large growth of function bodies. */
1135 cgraph_check_inline_limits (struct cgraph_node *to, struct cgraph_node *what,
1136 const char **reason)
1139 struct cgraph_edge *e;
1143 if (to->global.inlined_to)
1144 to = to->global.inlined_to;
1146 for (e = to->callees; e; e = e->next_callee)
1147 if (e->callee == what)
1150 /* When inlining large function body called once into small function,
1151 take the inlined function as base for limiting the growth. */
1152 if (to->local.self_insns > what->local.self_insns)
1153 limit = to->local.self_insns;
1155 limit = what->local.self_insns;
1157 limit += limit * PARAM_VALUE (PARAM_LARGE_FUNCTION_GROWTH) / 100;
1159 newsize = cgraph_estimate_size_after_inlining (times, to, what);
1160 if (newsize > PARAM_VALUE (PARAM_LARGE_FUNCTION_INSNS)
1164 *reason = N_("--param large-function-growth limit reached");
1170 /* Return true when function N is small enough to be inlined. */
1173 cgraph_default_inline_p (struct cgraph_node *n)
1175 if (!DECL_INLINE (n->decl) || !DECL_SAVED_TREE (n->decl))
1177 if (DECL_DECLARED_INLINE_P (n->decl))
1178 return n->global.insns < MAX_INLINE_INSNS_SINGLE;
1180 return n->global.insns < MAX_INLINE_INSNS_AUTO;
1183 /* Return true when inlining WHAT would create recursive inlining.
1184 We call recursive inlining all cases where same function appears more than
1185 once in the single recursion nest path in the inline graph. */
1188 cgraph_recursive_inlining_p (struct cgraph_node *to,
1189 struct cgraph_node *what,
1190 const char **reason)
1192 struct cgraph_node *node;
1194 /* Walk TO and all functions TO is inlined in. */
1197 /* We create recursive inlining either by inlining WHAT into something
1198 already inlined in possibly different clone of WHAT. */
1199 if (what->decl == to->decl)
1201 /* Or by inlining WHAT into something that is already inlined in WHAT. */
1202 for (node = cgraph_node (to->decl); node; node = node->next_clone)
1203 if (node->global.inlined_to == what)
1205 if (!to->callers || to->callers->inline_failed)
1207 to = to->callers->caller;
1211 *reason = (what->local.disregard_inline_limits
1212 ? N_("recursive inlining") : "");
1216 /* Recompute heap nodes for each of callees. */
1218 update_callee_keys (fibheap_t heap, struct fibnode **heap_node,
1219 struct cgraph_node *node)
1221 struct cgraph_edge *e;
1223 for (e = node->callees; e; e = e->next_callee)
1224 if (e->inline_failed && heap_node[e->callee->uid])
1225 fibheap_replace_key (heap, heap_node[e->callee->uid],
1226 cgraph_estimate_growth (e->callee));
1227 else if (!e->inline_failed)
1228 update_callee_keys (heap, heap_node, e->callee);
1231 /* Set inline_failed for all callers of given function to REASON. */
1234 cgraph_set_inline_failed (struct cgraph_node *node, const char *reason)
1236 struct cgraph_edge *e;
1238 if (cgraph_dump_file)
1239 fprintf (cgraph_dump_file, "Inlining failed: %s\n", reason);
1240 for (e = node->callers; e; e = e->next_caller)
1241 if (e->inline_failed)
1242 e->inline_failed = reason;
1245 /* We use greedy algorithm for inlining of small functions:
1246 All inline candidates are put into prioritized heap based on estimated
1247 growth of the overall number of instructions and then update the estimates.
1249 INLINED and INLINED_CALEES are just pointers to arrays large enough
1250 to be passed to cgraph_inlined_into and cgraph_inlined_callees. */
1253 cgraph_decide_inlining_of_small_functions (void)
1255 struct cgraph_node *node;
1256 fibheap_t heap = fibheap_new ();
1257 struct fibnode **heap_node =
1258 xcalloc (cgraph_max_uid, sizeof (struct fibnode *));
1259 int max_insns = ((HOST_WIDEST_INT) initial_insns
1260 * (100 + PARAM_VALUE (PARAM_INLINE_UNIT_GROWTH)) / 100);
1262 /* Put all inline candidates into the heap. */
1264 for (node = cgraph_nodes; node; node = node->next)
1266 if (!node->local.inlinable || !node->callers
1267 || node->local.disregard_inline_limits)
1270 if (!cgraph_default_inline_p (node))
1272 cgraph_set_inline_failed (node,
1273 N_("--param max-inline-insns-single limit reached"));
1276 heap_node[node->uid] =
1277 fibheap_insert (heap, cgraph_estimate_growth (node), node);
1280 if (cgraph_dump_file)
1281 fprintf (cgraph_dump_file, "\nDeciding on smaller functions:\n");
1282 while (overall_insns <= max_insns && (node = fibheap_extract_min (heap)))
1284 struct cgraph_edge *e, *next;
1285 int old_insns = overall_insns;
1287 heap_node[node->uid] = NULL;
1288 if (cgraph_dump_file)
1289 fprintf (cgraph_dump_file,
1290 "\nConsidering %s with %i insns\n"
1291 " Estimated growth is %+i insns.\n",
1292 cgraph_node_name (node), node->global.insns,
1293 cgraph_estimate_growth (node));
1294 if (!cgraph_default_inline_p (node))
1296 cgraph_set_inline_failed (node,
1297 N_("--param max-inline-insns-single limit reached after inlining into the callee"));
1300 for (e = node->callers; e; e = next)
1302 next = e->next_caller;
1303 if (e->inline_failed)
1305 struct cgraph_node *where;
1307 if (cgraph_recursive_inlining_p (e->caller, e->callee,
1309 || !cgraph_check_inline_limits (e->caller, e->callee,
1312 if (cgraph_dump_file)
1313 fprintf (cgraph_dump_file, " Not inlining into %s:%s.\n",
1314 cgraph_node_name (e->caller), e->inline_failed);
1317 next = cgraph_mark_inline (e);
1319 if (where->global.inlined_to)
1320 where = where->global.inlined_to;
1322 if (heap_node[where->uid])
1323 fibheap_replace_key (heap, heap_node[where->uid],
1324 cgraph_estimate_growth (where));
1326 if (cgraph_dump_file)
1327 fprintf (cgraph_dump_file,
1328 " Inlined into %s which now has %i insns.\n",
1329 cgraph_node_name (e->caller),
1330 e->caller->global.insns);
1334 /* Similarly all functions called by the function we just inlined
1335 are now called more times; update keys. */
1336 update_callee_keys (heap, heap_node, node);
1338 if (cgraph_dump_file)
1339 fprintf (cgraph_dump_file,
1340 " Inlined for a net change of %+i insns.\n",
1341 overall_insns - old_insns);
1343 while ((node = fibheap_extract_min (heap)) != NULL)
1344 if (!node->local.disregard_inline_limits)
1345 cgraph_set_inline_failed (node, N_("--param inline-unit-growth limit reached"));
1346 fibheap_delete (heap);
1350 /* Decide on the inlining. We do so in the topological order to avoid
1351 expenses on updating datastructures. */
1354 cgraph_decide_inlining (void)
1356 struct cgraph_node *node;
1358 struct cgraph_node **order =
1359 xcalloc (cgraph_n_nodes, sizeof (struct cgraph_node *));
1363 for (node = cgraph_nodes; node; node = node->next)
1364 initial_insns += node->local.self_insns;
1365 overall_insns = initial_insns;
1367 nnodes = cgraph_postorder (order);
1369 if (cgraph_dump_file)
1370 fprintf (cgraph_dump_file,
1371 "\nDeciding on inlining. Starting with %i insns.\n",
1374 for (node = cgraph_nodes; node; node = node->next)
1377 if (cgraph_dump_file)
1378 fprintf (cgraph_dump_file, "\nInlining always_inline functions:\n");
1380 /* In the first pass mark all always_inline edges. Do this with a priority
1381 so none of our later choices will make this impossible. */
1382 for (i = nnodes - 1; i >= 0; i--)
1384 struct cgraph_edge *e;
1388 for (e = node->callees; e; e = e->next_callee)
1389 if (e->callee->local.disregard_inline_limits)
1393 if (cgraph_dump_file)
1394 fprintf (cgraph_dump_file,
1395 "\nConsidering %s %i insns (always inline)\n",
1396 cgraph_node_name (e->callee), e->callee->global.insns);
1397 for (; e; e = e->next_callee)
1399 old_insns = overall_insns;
1400 if (!e->inline_failed || !e->callee->local.disregard_inline_limits)
1402 if (cgraph_recursive_inlining_p (order[i], e->callee,
1405 cgraph_mark_inline (e);
1406 if (cgraph_dump_file)
1407 fprintf (cgraph_dump_file,
1408 " Inlined into %s which now has %i insns.\n",
1409 cgraph_node_name (node->callees->caller),
1410 node->callees->caller->global.insns);
1412 if (cgraph_dump_file)
1413 fprintf (cgraph_dump_file,
1414 " Inlined for a net change of %+i insns.\n",
1415 overall_insns - old_insns);
1418 if (!flag_really_no_inline)
1420 cgraph_decide_inlining_of_small_functions ();
1422 if (cgraph_dump_file)
1423 fprintf (cgraph_dump_file, "\nDeciding on functions called once:\n");
1425 /* And finally decide what functions are called once. */
1427 for (i = nnodes - 1; i >= 0; i--)
1431 if (node->callers && !node->callers->next_caller && !node->needed
1432 && node->local.inlinable && node->callers->inline_failed
1433 && !DECL_EXTERNAL (node->decl) && !DECL_COMDAT (node->decl))
1436 struct cgraph_node *node1;
1438 /* Verify that we won't duplicate the caller. */
1439 for (node1 = node->callers->caller;
1440 node1->callers && !node1->callers->inline_failed
1441 && ok; node1 = node1->callers->caller)
1442 if (node1->callers->next_caller || node1->needed)
1446 if (cgraph_dump_file)
1447 fprintf (cgraph_dump_file,
1448 "\nConsidering %s %i insns.\n"
1449 " Called once from %s %i insns.\n",
1450 cgraph_node_name (node), node->global.insns,
1451 cgraph_node_name (node->callers->caller),
1452 node->callers->caller->global.insns);
1454 old_insns = overall_insns;
1456 if (cgraph_check_inline_limits (node->callers->caller, node,
1459 cgraph_mark_inline (node->callers);
1460 if (cgraph_dump_file)
1461 fprintf (cgraph_dump_file,
1462 " Inlined into %s which now has %i insns"
1463 " for a net change of %+i insns.\n",
1464 cgraph_node_name (node->callers->caller),
1465 node->callers->caller->global.insns,
1466 overall_insns - old_insns);
1470 if (cgraph_dump_file)
1471 fprintf (cgraph_dump_file,
1472 " Inline limit reached, not inlined.\n");
1479 /* We will never output extern functions we didn't inline.
1480 ??? Perhaps we can prevent accounting of growth of external
1481 inline functions. */
1482 cgraph_remove_unreachable_nodes ();
1484 if (cgraph_dump_file)
1485 fprintf (cgraph_dump_file,
1486 "\nInlined %i calls, eliminated %i functions, "
1487 "%i insns turned to %i insns.\n\n",
1488 ncalls_inlined, nfunctions_inlined, initial_insns,
1493 /* Decide on the inlining. We do so in the topological order to avoid
1494 expenses on updating datastructures. */
1497 cgraph_decide_inlining_incrementally (struct cgraph_node *node)
1499 struct cgraph_edge *e;
1501 /* First of all look for always inline functions. */
1502 for (e = node->callees; e; e = e->next_callee)
1503 if (e->callee->local.disregard_inline_limits
1505 && !cgraph_recursive_inlining_p (node, e->callee, &e->inline_failed)
1506 /* ??? It is possible that renaming variable removed the function body
1507 in duplicate_decls. See gcc.c-torture/compile/20011119-2.c */
1508 && DECL_SAVED_TREE (e->callee->decl))
1509 cgraph_mark_inline (e);
1511 /* Now do the automatic inlining. */
1512 if (!flag_really_no_inline)
1513 for (e = node->callees; e; e = e->next_callee)
1514 if (e->callee->local.inlinable
1516 && !e->callee->local.disregard_inline_limits
1517 && !cgraph_recursive_inlining_p (node, e->callee, &e->inline_failed)
1518 && cgraph_check_inline_limits (node, e->callee, &e->inline_failed)
1519 && DECL_SAVED_TREE (e->callee->decl))
1521 if (cgraph_default_inline_p (e->callee))
1522 cgraph_mark_inline (e);
1525 = N_("--param max-inline-insns-single limit reached");
1530 /* Return true when CALLER_DECL should be inlined into CALLEE_DECL. */
1533 cgraph_inline_p (struct cgraph_edge *e, const char **reason)
1535 *reason = e->inline_failed;
1536 return !e->inline_failed;
1539 /* Expand all functions that must be output.
1541 Attempt to topologically sort the nodes so function is output when
1542 all called functions are already assembled to allow data to be
1543 propagated across the callgraph. Use a stack to get smaller distance
1544 between a function and its callees (later we may choose to use a more
1545 sophisticated algorithm for function reordering; we will likely want
1546 to use subsections to make the output functions appear in top-down
1550 cgraph_expand_all_functions (void)
1552 struct cgraph_node *node;
1553 struct cgraph_node **order =
1554 xcalloc (cgraph_n_nodes, sizeof (struct cgraph_node *));
1555 int order_pos = 0, new_order_pos = 0;
1558 cgraph_mark_functions_to_output ();
1560 order_pos = cgraph_postorder (order);
1561 if (order_pos != cgraph_n_nodes)
1564 /* Garbage collector may remove inline clones we elliminate during
1565 optimization. So we must be sure to not reference them. */
1566 for (i = 0; i < order_pos; i++)
1567 if (order[i]->output)
1568 order[new_order_pos++] = order[i];
1570 for (i = new_order_pos - 1; i >= 0; i--)
1575 if (!node->reachable)
1578 cgraph_expand_function (node);
1584 /* Mark all local functions.
1586 A local function is one whose calls can occur only in the
1587 current compilation unit and all its calls are explicit,
1588 so we can change its calling convention.
1589 We simply mark all static functions whose address is not taken
1593 cgraph_mark_local_functions (void)
1595 struct cgraph_node *node;
1597 if (cgraph_dump_file)
1598 fprintf (cgraph_dump_file, "\nMarking local functions:");
1600 /* Figure out functions we want to assemble. */
1601 for (node = cgraph_nodes; node; node = node->next)
1603 node->local.local = (!node->needed
1604 && DECL_SAVED_TREE (node->decl)
1605 && !TREE_PUBLIC (node->decl));
1606 if (cgraph_dump_file && node->local.local)
1607 fprintf (cgraph_dump_file, " %s", cgraph_node_name (node));
1609 if (cgraph_dump_file)
1610 fprintf (cgraph_dump_file, "\n\n");
1613 /* Return true when function body of DECL still needs to be kept around
1614 for later re-use. */
1616 cgraph_preserve_function_body_p (tree decl)
1618 struct cgraph_node *node;
1619 /* Keep the body; we're going to dump it. */
1620 if (dump_enabled_p (TDI_all))
1622 if (!cgraph_global_info_ready)
1623 return (DECL_INLINE (decl) && !flag_really_no_inline);
1624 /* Look if there is any clone around. */
1625 for (node = cgraph_node (decl); node; node = node->next_clone)
1626 if (node->global.inlined_to)
1631 /* Perform simple optimizations based on callgraph. */
1634 cgraph_optimize (void)
1636 #ifdef ENABLE_CHECKING
1639 if (!flag_unit_at_a_time)
1641 timevar_push (TV_CGRAPHOPT);
1643 fprintf (stderr, "Performing intraprocedural optimizations\n");
1645 cgraph_mark_local_functions ();
1646 if (cgraph_dump_file)
1648 fprintf (cgraph_dump_file, "Marked ");
1649 dump_cgraph (cgraph_dump_file);
1652 if (flag_inline_trees)
1653 cgraph_decide_inlining ();
1654 cgraph_global_info_ready = true;
1655 if (cgraph_dump_file)
1657 fprintf (cgraph_dump_file, "Optimized ");
1658 dump_cgraph (cgraph_dump_file);
1660 timevar_pop (TV_CGRAPHOPT);
1662 /* Output everything. */
1664 fprintf (stderr, "Assembling functions:\n");
1665 #ifdef ENABLE_CHECKING
1668 cgraph_expand_all_functions ();
1669 if (cgraph_dump_file)
1671 fprintf (cgraph_dump_file, "\nFinal ");
1672 dump_cgraph (cgraph_dump_file);
1674 #ifdef ENABLE_CHECKING