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 cgraph_mark_reachable_node (node);
369 /* If we've not yet emitted decl, tell the debug info about it. */
370 if (!TREE_ASM_WRITTEN (decl))
371 (*debug_hooks->deferred_inline_function) (decl);
373 /* Possibly warn about unused parameters. */
374 if (warn_unused_parameter)
375 do_warn_unused_parameter (decl);
381 /* C99 extern inline keywords allow changing of declaration after function
382 has been finalized. We need to re-decide if we want to mark the function as
386 cgraph_mark_if_needed (tree decl)
388 struct cgraph_node *node = cgraph_node (decl);
389 if (node->local.finalized && cgraph_decide_is_function_needed (node, decl))
390 cgraph_mark_needed_node (node);
393 #ifdef ENABLE_CHECKING
394 /* Return TRUE if NODE2 is equivalent to NODE or its clone. */
396 clone_of_p (struct cgraph_node *node, struct cgraph_node *node2)
398 while (node != node2 && node2)
399 node2 = node2->clone_of;
400 return node2 != NULL;
404 /* Verify edge E count and frequency. */
407 verify_edge_count_and_frequency (struct cgraph_edge *e)
409 bool error_found = false;
412 error ("caller edge count is negative");
415 if (e->frequency < 0)
417 error ("caller edge frequency is negative");
420 if (e->frequency > CGRAPH_FREQ_MAX)
422 error ("caller edge frequency is too large");
425 if (gimple_has_body_p (e->caller->decl)
426 && !e->caller->global.inlined_to
428 != compute_call_stmt_bb_frequency (e->caller->decl,
429 gimple_bb (e->call_stmt))))
431 error ("caller edge frequency %i does not match BB freqency %i",
433 compute_call_stmt_bb_frequency (e->caller->decl,
434 gimple_bb (e->call_stmt)));
440 /* Verify cgraph nodes of given cgraph node. */
442 verify_cgraph_node (struct cgraph_node *node)
444 struct cgraph_edge *e;
445 struct function *this_cfun = DECL_STRUCT_FUNCTION (node->decl);
446 struct function *saved_cfun = cfun;
447 basic_block this_block;
448 gimple_stmt_iterator gsi;
449 bool error_found = false;
454 timevar_push (TV_CGRAPH_VERIFY);
455 /* debug_generic_stmt needs correct cfun */
456 set_cfun (this_cfun);
457 for (e = node->callees; e; e = e->next_callee)
460 error ("aux field set for edge %s->%s",
461 identifier_to_locale (cgraph_node_name (e->caller)),
462 identifier_to_locale (cgraph_node_name (e->callee)));
467 error ("Execution count is negative");
470 if (node->global.inlined_to && node->local.externally_visible)
472 error ("Externally visible inline clone");
475 if (node->global.inlined_to && node->address_taken)
477 error ("Inline clone with address taken");
480 if (node->global.inlined_to && node->needed)
482 error ("Inline clone is needed");
485 for (e = node->indirect_calls; e; e = e->next_callee)
489 error ("aux field set for indirect edge from %s",
490 identifier_to_locale (cgraph_node_name (e->caller)));
493 if (!e->indirect_unknown_callee
494 || !e->indirect_info)
496 error ("An indirect edge from %s is not marked as indirect or has "
497 "associated indirect_info, the corresponding statement is: ",
498 identifier_to_locale (cgraph_node_name (e->caller)));
499 debug_gimple_stmt (e->call_stmt);
503 for (e = node->callers; e; e = e->next_caller)
505 if (verify_edge_count_and_frequency (e))
507 if (!e->inline_failed)
509 if (node->global.inlined_to
510 != (e->caller->global.inlined_to
511 ? e->caller->global.inlined_to : e->caller))
513 error ("inlined_to pointer is wrong");
516 if (node->callers->next_caller)
518 error ("multiple inline callers");
523 if (node->global.inlined_to)
525 error ("inlined_to pointer set for noninline callers");
529 for (e = node->indirect_calls; e; e = e->next_callee)
530 if (verify_edge_count_and_frequency (e))
532 if (!node->callers && node->global.inlined_to)
534 error ("inlined_to pointer is set but no predecessors found");
537 if (node->global.inlined_to == node)
539 error ("inlined_to pointer refers to itself");
543 if (!cgraph_node (node->decl))
545 error ("node not found in cgraph_hash");
551 struct cgraph_node *n;
552 for (n = node->clone_of->clones; n; n = n->next_sibling_clone)
557 error ("node has wrong clone_of");
563 struct cgraph_node *n;
564 for (n = node->clones; n; n = n->next_sibling_clone)
565 if (n->clone_of != node)
569 error ("node has wrong clone list");
573 if ((node->prev_sibling_clone || node->next_sibling_clone) && !node->clone_of)
575 error ("node is in clone list but it is not clone");
578 if (!node->prev_sibling_clone && node->clone_of && node->clone_of->clones != node)
580 error ("node has wrong prev_clone pointer");
583 if (node->prev_sibling_clone && node->prev_sibling_clone->next_sibling_clone != node)
585 error ("double linked list of clones corrupted");
588 if (node->same_comdat_group)
590 struct cgraph_node *n = node->same_comdat_group;
592 if (!DECL_ONE_ONLY (node->decl))
594 error ("non-DECL_ONE_ONLY node in a same_comdat_group list");
599 error ("node is alone in a comdat group");
604 if (!n->same_comdat_group)
606 error ("same_comdat_group is not a circular list");
610 n = n->same_comdat_group;
615 if (node->analyzed && gimple_has_body_p (node->decl)
616 && !TREE_ASM_WRITTEN (node->decl)
617 && (!DECL_EXTERNAL (node->decl) || node->global.inlined_to)
622 /* The nodes we're interested in are never shared, so walk
623 the tree ignoring duplicates. */
624 struct pointer_set_t *visited_nodes = pointer_set_create ();
625 /* Reach the trees by walking over the CFG, and note the
626 enclosing basic-blocks in the call edges. */
627 FOR_EACH_BB_FN (this_block, this_cfun)
628 for (gsi = gsi_start_bb (this_block);
632 gimple stmt = gsi_stmt (gsi);
633 if (is_gimple_call (stmt))
635 struct cgraph_edge *e = cgraph_edge (node, stmt);
636 tree decl = gimple_call_fndecl (stmt);
641 error ("shared call_stmt:");
642 debug_gimple_stmt (stmt);
645 if (!e->indirect_unknown_callee)
647 if (e->callee->same_body_alias)
649 error ("edge points to same body alias:");
650 debug_tree (e->callee->decl);
653 #ifdef ENABLE_CHECKING
654 else if (!e->callee->global.inlined_to
656 && cgraph_get_node (decl)
657 && (e->callee->former_clone_of
658 != cgraph_get_node (decl)->decl)
659 && !clone_of_p (cgraph_node (decl),
662 error ("edge points to wrong declaration:");
663 debug_tree (e->callee->decl);
664 fprintf (stderr," Instead of:");
672 error ("an indirect edge with unknown callee "
673 "corresponding to a call_stmt with "
674 "a known declaration:");
676 debug_gimple_stmt (e->call_stmt);
682 error ("missing callgraph edge for call stmt:");
683 debug_gimple_stmt (stmt);
688 pointer_set_destroy (visited_nodes);
691 /* No CFG available?! */
694 for (e = node->callees; e; e = e->next_callee)
698 error ("edge %s->%s has no corresponding call_stmt",
699 identifier_to_locale (cgraph_node_name (e->caller)),
700 identifier_to_locale (cgraph_node_name (e->callee)));
701 debug_gimple_stmt (e->call_stmt);
706 for (e = node->indirect_calls; e; e = e->next_callee)
710 error ("an indirect edge from %s has no corresponding call_stmt",
711 identifier_to_locale (cgraph_node_name (e->caller)));
712 debug_gimple_stmt (e->call_stmt);
720 dump_cgraph_node (stderr, node);
721 internal_error ("verify_cgraph_node failed");
723 set_cfun (saved_cfun);
724 timevar_pop (TV_CGRAPH_VERIFY);
727 /* Verify whole cgraph structure. */
731 struct cgraph_node *node;
736 for (node = cgraph_nodes; node; node = node->next)
737 verify_cgraph_node (node);
740 /* Output all asm statements we have stored up to be output. */
743 cgraph_output_pending_asms (void)
745 struct cgraph_asm_node *can;
750 for (can = cgraph_asm_nodes; can; can = can->next)
751 assemble_asm (can->asm_str);
752 cgraph_asm_nodes = NULL;
755 /* Analyze the function scheduled to be output. */
757 cgraph_analyze_function (struct cgraph_node *node)
759 tree save = current_function_decl;
760 tree decl = node->decl;
762 current_function_decl = decl;
763 push_cfun (DECL_STRUCT_FUNCTION (decl));
765 assign_assembler_name_if_neeeded (node->decl);
767 /* Make sure to gimplify bodies only once. During analyzing a
768 function we lower it, which will require gimplified nested
769 functions, so we can end up here with an already gimplified
771 if (!gimple_body (decl))
772 gimplify_function_tree (decl);
773 dump_function (TDI_generic, decl);
775 cgraph_lower_function (node);
776 node->analyzed = true;
779 current_function_decl = save;
782 /* Look for externally_visible and used attributes and mark cgraph nodes
785 We cannot mark the nodes at the point the attributes are processed (in
786 handle_*_attribute) because the copy of the declarations available at that
787 point may not be canonical. For example, in:
790 void f() __attribute__((used));
792 the declaration we see in handle_used_attribute will be the second
793 declaration -- but the front end will subsequently merge that declaration
794 with the original declaration and discard the second declaration.
796 Furthermore, we can't mark these nodes in cgraph_finalize_function because:
799 void f() __attribute__((externally_visible));
803 So, we walk the nodes at the end of the translation unit, applying the
804 attributes at that point. */
807 process_function_and_variable_attributes (struct cgraph_node *first,
808 struct varpool_node *first_var)
810 struct cgraph_node *node;
811 struct varpool_node *vnode;
813 for (node = cgraph_nodes; node != first; node = node->next)
815 tree decl = node->decl;
816 if (DECL_PRESERVE_P (decl))
817 cgraph_mark_needed_node (node);
818 if (lookup_attribute ("externally_visible", DECL_ATTRIBUTES (decl)))
820 if (! TREE_PUBLIC (node->decl))
821 warning_at (DECL_SOURCE_LOCATION (node->decl), OPT_Wattributes,
822 "%<externally_visible%>"
823 " attribute have effect only on public objects");
824 else if (node->local.finalized)
825 cgraph_mark_needed_node (node);
828 for (vnode = varpool_nodes; vnode != first_var; vnode = vnode->next)
830 tree decl = vnode->decl;
831 if (DECL_PRESERVE_P (decl))
833 vnode->force_output = true;
834 if (vnode->finalized)
835 varpool_mark_needed_node (vnode);
837 if (lookup_attribute ("externally_visible", DECL_ATTRIBUTES (decl)))
839 if (! TREE_PUBLIC (vnode->decl))
840 warning_at (DECL_SOURCE_LOCATION (vnode->decl), OPT_Wattributes,
841 "%<externally_visible%>"
842 " attribute have effect only on public objects");
843 else if (vnode->finalized)
844 varpool_mark_needed_node (vnode);
849 /* Process CGRAPH_NODES_NEEDED queue, analyze each function (and transitively
850 each reachable functions) and build cgraph.
851 The function can be called multiple times after inserting new nodes
852 into beginning of queue. Just the new part of queue is re-scanned then. */
855 cgraph_analyze_functions (void)
857 /* Keep track of already processed nodes when called multiple times for
858 intermodule optimization. */
859 static struct cgraph_node *first_analyzed;
860 struct cgraph_node *first_processed = first_analyzed;
861 static struct varpool_node *first_analyzed_var;
862 struct cgraph_node *node, *next;
864 process_function_and_variable_attributes (first_processed,
866 first_processed = cgraph_nodes;
867 first_analyzed_var = varpool_nodes;
868 varpool_analyze_pending_decls ();
869 if (cgraph_dump_file)
871 fprintf (cgraph_dump_file, "Initial entry points:");
872 for (node = cgraph_nodes; node != first_analyzed; node = node->next)
874 fprintf (cgraph_dump_file, " %s", cgraph_node_name (node));
875 fprintf (cgraph_dump_file, "\n");
877 cgraph_process_new_functions ();
879 /* Propagate reachability flag and lower representation of all reachable
880 functions. In the future, lowering will introduce new functions and
881 new entry points on the way (by template instantiation and virtual
882 method table generation for instance). */
883 while (cgraph_nodes_queue)
885 struct cgraph_edge *edge;
886 tree decl = cgraph_nodes_queue->decl;
888 node = cgraph_nodes_queue;
889 cgraph_nodes_queue = cgraph_nodes_queue->next_needed;
890 node->next_needed = NULL;
892 /* ??? It is possible to create extern inline function and later using
893 weak alias attribute to kill its body. See
894 gcc.c-torture/compile/20011119-1.c */
895 if (!DECL_STRUCT_FUNCTION (decl))
897 cgraph_reset_node (node);
902 cgraph_analyze_function (node);
904 for (edge = node->callees; edge; edge = edge->next_callee)
905 if (!edge->callee->reachable)
906 cgraph_mark_reachable_node (edge->callee);
908 if (node->same_comdat_group)
910 for (next = node->same_comdat_group;
912 next = next->same_comdat_group)
913 cgraph_mark_reachable_node (next);
916 /* If decl is a clone of an abstract function, mark that abstract
917 function so that we don't release its body. The DECL_INITIAL() of that
918 abstract function declaration will be later needed to output debug info. */
919 if (DECL_ABSTRACT_ORIGIN (decl))
921 struct cgraph_node *origin_node = cgraph_node (DECL_ABSTRACT_ORIGIN (decl));
922 origin_node->abstract_and_needed = true;
925 /* We finalize local static variables during constructing callgraph
926 edges. Process their attributes too. */
927 process_function_and_variable_attributes (first_processed,
929 first_processed = cgraph_nodes;
930 first_analyzed_var = varpool_nodes;
931 varpool_analyze_pending_decls ();
932 cgraph_process_new_functions ();
935 /* Collect entry points to the unit. */
936 if (cgraph_dump_file)
938 fprintf (cgraph_dump_file, "Unit entry points:");
939 for (node = cgraph_nodes; node != first_analyzed; node = node->next)
941 fprintf (cgraph_dump_file, " %s", cgraph_node_name (node));
942 fprintf (cgraph_dump_file, "\n\nInitial ");
943 dump_cgraph (cgraph_dump_file);
946 if (cgraph_dump_file)
947 fprintf (cgraph_dump_file, "\nReclaiming functions:");
949 for (node = cgraph_nodes; node != first_analyzed; node = next)
951 tree decl = node->decl;
954 if (node->local.finalized && !gimple_has_body_p (decl))
955 cgraph_reset_node (node);
957 if (!node->reachable && gimple_has_body_p (decl))
959 if (cgraph_dump_file)
960 fprintf (cgraph_dump_file, " %s", cgraph_node_name (node));
961 cgraph_remove_node (node);
965 node->next_needed = NULL;
966 gcc_assert (!node->local.finalized || gimple_has_body_p (decl));
967 gcc_assert (node->analyzed == node->local.finalized);
969 if (cgraph_dump_file)
971 fprintf (cgraph_dump_file, "\n\nReclaimed ");
972 dump_cgraph (cgraph_dump_file);
974 first_analyzed = cgraph_nodes;
979 /* Analyze the whole compilation unit once it is parsed completely. */
982 cgraph_finalize_compilation_unit (void)
984 timevar_push (TV_CGRAPH);
986 /* Do not skip analyzing the functions if there were errors, we
987 miss diagnostics for following functions otherwise. */
989 /* Emit size functions we didn't inline. */
990 finalize_size_functions ();
992 /* Mark alias targets necessary and emit diagnostics. */
997 fprintf (stderr, "\nAnalyzing compilation unit\n");
1001 /* Gimplify and lower all functions, compute reachability and
1002 remove unreachable nodes. */
1003 cgraph_analyze_functions ();
1005 /* Mark alias targets necessary and emit diagnostics. */
1006 finish_aliases_1 ();
1008 /* Gimplify and lower thunks. */
1009 cgraph_analyze_functions ();
1011 /* Finally drive the pass manager. */
1014 timevar_pop (TV_CGRAPH);
1018 /* Figure out what functions we want to assemble. */
1021 cgraph_mark_functions_to_output (void)
1023 struct cgraph_node *node;
1024 #ifdef ENABLE_CHECKING
1025 bool check_same_comdat_groups = false;
1027 for (node = cgraph_nodes; node; node = node->next)
1028 gcc_assert (!node->process);
1031 for (node = cgraph_nodes; node; node = node->next)
1033 tree decl = node->decl;
1034 struct cgraph_edge *e;
1036 gcc_assert (!node->process || node->same_comdat_group);
1040 for (e = node->callers; e; e = e->next_caller)
1041 if (e->inline_failed)
1044 /* We need to output all local functions that are used and not
1045 always inlined, as well as those that are reachable from
1046 outside the current compilation unit. */
1048 && !node->global.inlined_to
1049 && (!cgraph_only_called_directly_p (node)
1050 || (e && node->reachable))
1051 && !TREE_ASM_WRITTEN (decl)
1052 && !DECL_EXTERNAL (decl))
1055 if (node->same_comdat_group)
1057 struct cgraph_node *next;
1058 for (next = node->same_comdat_group;
1060 next = next->same_comdat_group)
1064 else if (node->same_comdat_group)
1066 #ifdef ENABLE_CHECKING
1067 check_same_comdat_groups = true;
1072 /* We should've reclaimed all functions that are not needed. */
1073 #ifdef ENABLE_CHECKING
1074 if (!node->global.inlined_to
1075 && gimple_has_body_p (decl)
1076 /* FIXME: in ltrans unit when offline copy is outside partition but inline copies
1077 are inside partition, we can end up not removing the body since we no longer
1078 have analyzed node pointing to it. */
1079 && !node->in_other_partition
1080 && !DECL_EXTERNAL (decl))
1082 dump_cgraph_node (stderr, node);
1083 internal_error ("failed to reclaim unneeded function");
1086 gcc_assert (node->global.inlined_to
1087 || !gimple_has_body_p (decl)
1088 || node->in_other_partition
1089 || DECL_EXTERNAL (decl));
1094 #ifdef ENABLE_CHECKING
1095 if (check_same_comdat_groups)
1096 for (node = cgraph_nodes; node; node = node->next)
1097 if (node->same_comdat_group && !node->process)
1099 tree decl = node->decl;
1100 if (!node->global.inlined_to
1101 && gimple_has_body_p (decl)
1102 /* FIXME: in ltrans unit when offline copy is outside partition but inline copies
1103 are inside partition, we can end up not removing the body since we no longer
1104 have analyzed node pointing to it. */
1105 && !node->in_other_partition
1106 && !DECL_EXTERNAL (decl))
1108 dump_cgraph_node (stderr, node);
1109 internal_error ("failed to reclaim unneeded function");
1115 /* DECL is FUNCTION_DECL. Initialize datastructures so DECL is a function
1116 in lowered gimple form.
1118 Set current_function_decl and cfun to newly constructed empty function body.
1119 return basic block in the function body. */
1122 init_lowered_empty_function (tree decl)
1126 current_function_decl = decl;
1127 allocate_struct_function (decl, false);
1128 gimple_register_cfg_hooks ();
1129 init_empty_tree_cfg ();
1130 init_tree_ssa (cfun);
1131 init_ssa_operands ();
1132 cfun->gimple_df->in_ssa_p = true;
1133 DECL_INITIAL (decl) = make_node (BLOCK);
1135 DECL_SAVED_TREE (decl) = error_mark_node;
1136 cfun->curr_properties |=
1137 (PROP_gimple_lcf | PROP_gimple_leh | PROP_cfg | PROP_referenced_vars |
1140 /* Create BB for body of the function and connect it properly. */
1141 bb = create_basic_block (NULL, (void *) 0, ENTRY_BLOCK_PTR);
1142 make_edge (ENTRY_BLOCK_PTR, bb, 0);
1143 make_edge (bb, EXIT_BLOCK_PTR, 0);
1148 /* Adjust PTR by the constant FIXED_OFFSET, and by the vtable
1149 offset indicated by VIRTUAL_OFFSET, if that is
1150 non-null. THIS_ADJUSTING is nonzero for a this adjusting thunk and
1151 zero for a result adjusting thunk. */
1154 thunk_adjust (gimple_stmt_iterator * bsi,
1155 tree ptr, bool this_adjusting,
1156 HOST_WIDE_INT fixed_offset, tree virtual_offset)
1162 && fixed_offset != 0)
1164 stmt = gimple_build_assign (ptr,
1165 fold_build2_loc (input_location,
1167 TREE_TYPE (ptr), ptr,
1168 size_int (fixed_offset)));
1169 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1172 /* If there's a virtual offset, look up that value in the vtable and
1173 adjust the pointer again. */
1181 if (!vtable_entry_type)
1183 tree vfunc_type = make_node (FUNCTION_TYPE);
1184 TREE_TYPE (vfunc_type) = integer_type_node;
1185 TYPE_ARG_TYPES (vfunc_type) = NULL_TREE;
1186 layout_type (vfunc_type);
1188 vtable_entry_type = build_pointer_type (vfunc_type);
1192 create_tmp_var (build_pointer_type
1193 (build_pointer_type (vtable_entry_type)), "vptr");
1195 /* The vptr is always at offset zero in the object. */
1196 stmt = gimple_build_assign (vtabletmp,
1197 build1 (NOP_EXPR, TREE_TYPE (vtabletmp),
1199 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1200 mark_symbols_for_renaming (stmt);
1201 find_referenced_vars_in (stmt);
1203 /* Form the vtable address. */
1204 vtabletmp2 = create_tmp_var (TREE_TYPE (TREE_TYPE (vtabletmp)),
1206 stmt = gimple_build_assign (vtabletmp2,
1207 build_simple_mem_ref (vtabletmp));
1208 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1209 mark_symbols_for_renaming (stmt);
1210 find_referenced_vars_in (stmt);
1212 /* Find the entry with the vcall offset. */
1213 stmt = gimple_build_assign (vtabletmp2,
1214 fold_build2_loc (input_location,
1216 TREE_TYPE (vtabletmp2),
1218 fold_convert (sizetype,
1220 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1222 /* Get the offset itself. */
1223 vtabletmp3 = create_tmp_var (TREE_TYPE (TREE_TYPE (vtabletmp2)),
1225 stmt = gimple_build_assign (vtabletmp3,
1226 build_simple_mem_ref (vtabletmp2));
1227 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1228 mark_symbols_for_renaming (stmt);
1229 find_referenced_vars_in (stmt);
1231 /* Cast to sizetype. */
1232 offsettmp = create_tmp_var (sizetype, "offset");
1233 stmt = gimple_build_assign (offsettmp, fold_convert (sizetype, vtabletmp3));
1234 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1235 mark_symbols_for_renaming (stmt);
1236 find_referenced_vars_in (stmt);
1238 /* Adjust the `this' pointer. */
1239 ptr = fold_build2_loc (input_location,
1240 POINTER_PLUS_EXPR, TREE_TYPE (ptr), ptr,
1245 && fixed_offset != 0)
1246 /* Adjust the pointer by the constant. */
1250 if (TREE_CODE (ptr) == VAR_DECL)
1254 ptrtmp = create_tmp_var (TREE_TYPE (ptr), "ptr");
1255 stmt = gimple_build_assign (ptrtmp, ptr);
1256 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1257 mark_symbols_for_renaming (stmt);
1258 find_referenced_vars_in (stmt);
1260 ptr = fold_build2_loc (input_location,
1261 POINTER_PLUS_EXPR, TREE_TYPE (ptrtmp), ptrtmp,
1262 size_int (fixed_offset));
1265 /* Emit the statement and gimplify the adjustment expression. */
1266 ret = create_tmp_var (TREE_TYPE (ptr), "adjusted_this");
1267 stmt = gimple_build_assign (ret, ptr);
1268 mark_symbols_for_renaming (stmt);
1269 find_referenced_vars_in (stmt);
1270 gsi_insert_after (bsi, stmt, GSI_NEW_STMT);
1275 /* Produce assembler for thunk NODE. */
1278 assemble_thunk (struct cgraph_node *node)
1280 bool this_adjusting = node->thunk.this_adjusting;
1281 HOST_WIDE_INT fixed_offset = node->thunk.fixed_offset;
1282 HOST_WIDE_INT virtual_value = node->thunk.virtual_value;
1283 tree virtual_offset = NULL;
1284 tree alias = node->thunk.alias;
1285 tree thunk_fndecl = node->decl;
1286 tree a = DECL_ARGUMENTS (thunk_fndecl);
1288 current_function_decl = thunk_fndecl;
1291 && targetm.asm_out.can_output_mi_thunk (thunk_fndecl, fixed_offset,
1292 virtual_value, alias))
1297 DECL_RESULT (thunk_fndecl)
1298 = build_decl (DECL_SOURCE_LOCATION (thunk_fndecl),
1299 RESULT_DECL, 0, integer_type_node);
1300 fnname = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (thunk_fndecl));
1302 /* The back end expects DECL_INITIAL to contain a BLOCK, so we
1304 fn_block = make_node (BLOCK);
1305 BLOCK_VARS (fn_block) = a;
1306 DECL_INITIAL (thunk_fndecl) = fn_block;
1307 init_function_start (thunk_fndecl);
1309 assemble_start_function (thunk_fndecl, fnname);
1311 targetm.asm_out.output_mi_thunk (asm_out_file, thunk_fndecl,
1312 fixed_offset, virtual_value, alias);
1314 assemble_end_function (thunk_fndecl, fnname);
1315 init_insn_lengths ();
1316 free_after_compilation (cfun);
1318 TREE_ASM_WRITTEN (thunk_fndecl) = 1;
1323 basic_block bb, then_bb, else_bb, return_bb;
1324 gimple_stmt_iterator bsi;
1330 VEC(tree, heap) *vargs;
1335 DECL_IGNORED_P (thunk_fndecl) = 1;
1336 bitmap_obstack_initialize (NULL);
1338 if (node->thunk.virtual_offset_p)
1339 virtual_offset = size_int (virtual_value);
1341 /* Build the return declaration for the function. */
1342 restype = TREE_TYPE (TREE_TYPE (thunk_fndecl));
1343 if (DECL_RESULT (thunk_fndecl) == NULL_TREE)
1345 resdecl = build_decl (input_location, RESULT_DECL, 0, restype);
1346 DECL_ARTIFICIAL (resdecl) = 1;
1347 DECL_IGNORED_P (resdecl) = 1;
1348 DECL_RESULT (thunk_fndecl) = resdecl;
1351 resdecl = DECL_RESULT (thunk_fndecl);
1353 bb = then_bb = else_bb = return_bb = init_lowered_empty_function (thunk_fndecl);
1355 bsi = gsi_start_bb (bb);
1357 /* Build call to the function being thunked. */
1358 if (!VOID_TYPE_P (restype))
1360 if (!is_gimple_reg_type (restype))
1363 add_local_decl (cfun, restmp);
1364 BLOCK_VARS (DECL_INITIAL (current_function_decl)) = restmp;
1367 restmp = create_tmp_var_raw (restype, "retval");
1370 for (arg = a; arg; arg = DECL_CHAIN (arg))
1372 vargs = VEC_alloc (tree, heap, nargs);
1374 VEC_quick_push (tree, vargs,
1379 VEC_quick_push (tree, vargs, a);
1380 for (i = 1, arg = DECL_CHAIN (a); i < nargs; i++, arg = DECL_CHAIN (arg))
1381 VEC_quick_push (tree, vargs, arg);
1382 call = gimple_build_call_vec (build_fold_addr_expr_loc (0, alias), vargs);
1383 VEC_free (tree, heap, vargs);
1384 gimple_call_set_cannot_inline (call, true);
1385 gimple_call_set_from_thunk (call, true);
1387 gimple_call_set_lhs (call, restmp);
1388 gsi_insert_after (&bsi, call, GSI_NEW_STMT);
1389 mark_symbols_for_renaming (call);
1390 find_referenced_vars_in (call);
1393 if (restmp && !this_adjusting)
1395 tree true_label = NULL_TREE;
1397 if (TREE_CODE (TREE_TYPE (restmp)) == POINTER_TYPE)
1400 /* If the return type is a pointer, we need to
1401 protect against NULL. We know there will be an
1402 adjustment, because that's why we're emitting a
1404 then_bb = create_basic_block (NULL, (void *) 0, bb);
1405 return_bb = create_basic_block (NULL, (void *) 0, then_bb);
1406 else_bb = create_basic_block (NULL, (void *) 0, else_bb);
1407 remove_edge (single_succ_edge (bb));
1408 true_label = gimple_block_label (then_bb);
1409 stmt = gimple_build_cond (NE_EXPR, restmp,
1410 fold_convert (TREE_TYPE (restmp),
1412 NULL_TREE, NULL_TREE);
1413 gsi_insert_after (&bsi, stmt, GSI_NEW_STMT);
1414 make_edge (bb, then_bb, EDGE_TRUE_VALUE);
1415 make_edge (bb, else_bb, EDGE_FALSE_VALUE);
1416 make_edge (return_bb, EXIT_BLOCK_PTR, 0);
1417 make_edge (then_bb, return_bb, EDGE_FALLTHRU);
1418 make_edge (else_bb, return_bb, EDGE_FALLTHRU);
1419 bsi = gsi_last_bb (then_bb);
1422 restmp = thunk_adjust (&bsi, restmp, /*this_adjusting=*/0,
1423 fixed_offset, virtual_offset);
1427 bsi = gsi_last_bb (else_bb);
1428 stmt = gimple_build_assign (restmp, fold_convert (TREE_TYPE (restmp),
1429 integer_zero_node));
1430 gsi_insert_after (&bsi, stmt, GSI_NEW_STMT);
1431 bsi = gsi_last_bb (return_bb);
1435 gimple_call_set_tail (call, true);
1437 /* Build return value. */
1438 ret = gimple_build_return (restmp);
1439 gsi_insert_after (&bsi, ret, GSI_NEW_STMT);
1441 delete_unreachable_blocks ();
1442 update_ssa (TODO_update_ssa);
1444 cgraph_remove_same_body_alias (node);
1445 /* Since we want to emit the thunk, we explicitly mark its name as
1447 cgraph_add_new_function (thunk_fndecl, true);
1448 bitmap_obstack_release (NULL);
1450 current_function_decl = NULL;
1453 /* Expand function specified by NODE. */
1456 cgraph_expand_function (struct cgraph_node *node)
1458 tree decl = node->decl;
1460 /* We ought to not compile any inline clones. */
1461 gcc_assert (!node->global.inlined_to);
1463 announce_function (decl);
1466 gcc_assert (node->lowered);
1468 /* Generate RTL for the body of DECL. */
1469 tree_rest_of_compilation (decl);
1471 /* Make sure that BE didn't give up on compiling. */
1472 gcc_assert (TREE_ASM_WRITTEN (decl));
1473 current_function_decl = NULL;
1474 if (node->same_body)
1476 struct cgraph_node *alias, *next;
1477 bool saved_alias = node->alias;
1478 for (alias = node->same_body;
1479 alias && alias->next; alias = alias->next)
1481 /* Walk aliases in the order they were created; it is possible that
1482 thunks reffers to the aliases made earlier. */
1483 for (; alias; alias = next)
1485 next = alias->previous;
1486 if (!alias->thunk.thunk_p)
1487 assemble_alias (alias->decl,
1488 DECL_ASSEMBLER_NAME (alias->thunk.alias));
1490 assemble_thunk (alias);
1492 node->alias = saved_alias;
1494 gcc_assert (!cgraph_preserve_function_body_p (decl));
1495 cgraph_release_function_body (node);
1496 /* Eliminate all call edges. This is important so the GIMPLE_CALL no longer
1497 points to the dead function body. */
1498 cgraph_node_remove_callees (node);
1500 cgraph_function_flags_ready = true;
1503 /* Return true when CALLER_DECL should be inlined into CALLEE_DECL. */
1506 cgraph_inline_p (struct cgraph_edge *e, cgraph_inline_failed_t *reason)
1508 *reason = e->inline_failed;
1509 return !e->inline_failed;
1514 /* Expand all functions that must be output.
1516 Attempt to topologically sort the nodes so function is output when
1517 all called functions are already assembled to allow data to be
1518 propagated across the callgraph. Use a stack to get smaller distance
1519 between a function and its callees (later we may choose to use a more
1520 sophisticated algorithm for function reordering; we will likely want
1521 to use subsections to make the output functions appear in top-down
1525 cgraph_expand_all_functions (void)
1527 struct cgraph_node *node;
1528 struct cgraph_node **order = XCNEWVEC (struct cgraph_node *, cgraph_n_nodes);
1529 int order_pos, new_order_pos = 0;
1532 order_pos = cgraph_postorder (order);
1533 gcc_assert (order_pos == cgraph_n_nodes);
1535 /* Garbage collector may remove inline clones we eliminate during
1536 optimization. So we must be sure to not reference them. */
1537 for (i = 0; i < order_pos; i++)
1538 if (order[i]->process)
1539 order[new_order_pos++] = order[i];
1541 for (i = new_order_pos - 1; i >= 0; i--)
1546 gcc_assert (node->reachable);
1548 cgraph_expand_function (node);
1551 cgraph_process_new_functions ();
1557 /* This is used to sort the node types by the cgraph order number. */
1559 enum cgraph_order_sort_kind
1561 ORDER_UNDEFINED = 0,
1567 struct cgraph_order_sort
1569 enum cgraph_order_sort_kind kind;
1572 struct cgraph_node *f;
1573 struct varpool_node *v;
1574 struct cgraph_asm_node *a;
1578 /* Output all functions, variables, and asm statements in the order
1579 according to their order fields, which is the order in which they
1580 appeared in the file. This implements -fno-toplevel-reorder. In
1581 this mode we may output functions and variables which don't really
1582 need to be output. */
1585 cgraph_output_in_order (void)
1588 struct cgraph_order_sort *nodes;
1590 struct cgraph_node *pf;
1591 struct varpool_node *pv;
1592 struct cgraph_asm_node *pa;
1595 nodes = XCNEWVEC (struct cgraph_order_sort, max);
1597 varpool_analyze_pending_decls ();
1599 for (pf = cgraph_nodes; pf; pf = pf->next)
1604 gcc_assert (nodes[i].kind == ORDER_UNDEFINED);
1605 nodes[i].kind = ORDER_FUNCTION;
1610 for (pv = varpool_nodes_queue; pv; pv = pv->next_needed)
1613 gcc_assert (nodes[i].kind == ORDER_UNDEFINED);
1614 nodes[i].kind = ORDER_VAR;
1618 for (pa = cgraph_asm_nodes; pa; pa = pa->next)
1621 gcc_assert (nodes[i].kind == ORDER_UNDEFINED);
1622 nodes[i].kind = ORDER_ASM;
1626 /* In toplevel reorder mode we output all statics; mark them as needed. */
1627 for (i = 0; i < max; ++i)
1629 if (nodes[i].kind == ORDER_VAR)
1631 varpool_mark_needed_node (nodes[i].u.v);
1634 varpool_empty_needed_queue ();
1636 for (i = 0; i < max; ++i)
1638 switch (nodes[i].kind)
1640 case ORDER_FUNCTION:
1641 nodes[i].u.f->process = 0;
1642 cgraph_expand_function (nodes[i].u.f);
1646 varpool_assemble_decl (nodes[i].u.v);
1650 assemble_asm (nodes[i].u.a->asm_str);
1653 case ORDER_UNDEFINED:
1661 cgraph_asm_nodes = NULL;
1665 /* Return true when function body of DECL still needs to be kept around
1666 for later re-use. */
1668 cgraph_preserve_function_body_p (tree decl)
1670 struct cgraph_node *node;
1672 gcc_assert (cgraph_global_info_ready);
1673 /* Look if there is any clone around. */
1674 node = cgraph_node (decl);
1684 current_function_decl = NULL;
1685 gimple_register_cfg_hooks ();
1686 bitmap_obstack_initialize (NULL);
1688 invoke_plugin_callbacks (PLUGIN_ALL_IPA_PASSES_START, NULL);
1691 execute_ipa_pass_list (all_small_ipa_passes);
1693 /* If pass_all_early_optimizations was not scheduled, the state of
1694 the cgraph will not be properly updated. Update it now. */
1695 if (cgraph_state < CGRAPH_STATE_IPA_SSA)
1696 cgraph_state = CGRAPH_STATE_IPA_SSA;
1700 /* Generate coverage variables and constructors. */
1703 /* Process new functions added. */
1705 current_function_decl = NULL;
1706 cgraph_process_new_functions ();
1708 execute_ipa_summary_passes
1709 ((struct ipa_opt_pass_d *) all_regular_ipa_passes);
1712 /* Some targets need to handle LTO assembler output specially. */
1713 if (flag_generate_lto)
1714 targetm.asm_out.lto_start ();
1716 execute_ipa_summary_passes ((struct ipa_opt_pass_d *) all_lto_gen_passes);
1719 ipa_write_summaries ();
1721 if (flag_generate_lto)
1722 targetm.asm_out.lto_end ();
1725 execute_ipa_pass_list (all_regular_ipa_passes);
1726 invoke_plugin_callbacks (PLUGIN_ALL_IPA_PASSES_END, NULL);
1728 bitmap_obstack_release (NULL);
1732 /* Perform simple optimizations based on callgraph. */
1735 cgraph_optimize (void)
1740 #ifdef ENABLE_CHECKING
1744 /* Frontend may output common variables after the unit has been finalized.
1745 It is safe to deal with them here as they are always zero initialized. */
1746 varpool_analyze_pending_decls ();
1748 timevar_push (TV_CGRAPHOPT);
1749 if (pre_ipa_mem_report)
1751 fprintf (stderr, "Memory consumption before IPA\n");
1752 dump_memory_report (false);
1755 fprintf (stderr, "Performing interprocedural optimizations\n");
1756 cgraph_state = CGRAPH_STATE_IPA;
1758 /* Don't run the IPA passes if there was any error or sorry messages. */
1762 /* Do nothing else if any IPA pass found errors. */
1765 timevar_pop (TV_CGRAPHOPT);
1769 /* This pass remove bodies of extern inline functions we never inlined.
1770 Do this later so other IPA passes see what is really going on. */
1771 cgraph_remove_unreachable_nodes (false, dump_file);
1772 cgraph_global_info_ready = true;
1773 if (cgraph_dump_file)
1775 fprintf (cgraph_dump_file, "Optimized ");
1776 dump_cgraph (cgraph_dump_file);
1777 dump_varpool (cgraph_dump_file);
1779 if (post_ipa_mem_report)
1781 fprintf (stderr, "Memory consumption after IPA\n");
1782 dump_memory_report (false);
1784 timevar_pop (TV_CGRAPHOPT);
1786 /* Output everything. */
1787 (*debug_hooks->assembly_start) ();
1789 fprintf (stderr, "Assembling functions:\n");
1790 #ifdef ENABLE_CHECKING
1794 cgraph_materialize_all_clones ();
1795 cgraph_mark_functions_to_output ();
1797 cgraph_state = CGRAPH_STATE_EXPANSION;
1798 if (!flag_toplevel_reorder)
1799 cgraph_output_in_order ();
1802 cgraph_output_pending_asms ();
1804 cgraph_expand_all_functions ();
1805 varpool_remove_unreferenced_decls ();
1807 varpool_assemble_pending_decls ();
1809 cgraph_process_new_functions ();
1810 cgraph_state = CGRAPH_STATE_FINISHED;
1812 if (cgraph_dump_file)
1814 fprintf (cgraph_dump_file, "\nFinal ");
1815 dump_cgraph (cgraph_dump_file);
1817 #ifdef ENABLE_CHECKING
1819 /* Double check that all inline clones are gone and that all
1820 function bodies have been released from memory. */
1823 struct cgraph_node *node;
1824 bool error_found = false;
1826 for (node = cgraph_nodes; node; node = node->next)
1828 && (node->global.inlined_to
1829 || gimple_has_body_p (node->decl)))
1832 dump_cgraph_node (stderr, node);
1835 internal_error ("nodes with unreleased memory found");
1843 if (!cgraph_dump_file)
1844 cgraph_dump_file = dump_begin (TDI_cgraph, NULL);
1847 /* The edges representing the callers of the NEW_VERSION node were
1848 fixed by cgraph_function_versioning (), now the call_expr in their
1849 respective tree code should be updated to call the NEW_VERSION. */
1852 update_call_expr (struct cgraph_node *new_version)
1854 struct cgraph_edge *e;
1856 gcc_assert (new_version);
1858 /* Update the call expr on the edges to call the new version. */
1859 for (e = new_version->callers; e; e = e->next_caller)
1861 struct function *inner_function = DECL_STRUCT_FUNCTION (e->caller->decl);
1862 gimple_call_set_fndecl (e->call_stmt, new_version->decl);
1863 maybe_clean_eh_stmt_fn (inner_function, e->call_stmt);
1868 /* Create a new cgraph node which is the new version of
1869 OLD_VERSION node. REDIRECT_CALLERS holds the callers
1870 edges which should be redirected to point to
1871 NEW_VERSION. ALL the callees edges of OLD_VERSION
1872 are cloned to the new version node. Return the new
1875 If non-NULL BLOCK_TO_COPY determine what basic blocks
1876 was copied to prevent duplications of calls that are dead
1879 static struct cgraph_node *
1880 cgraph_copy_node_for_versioning (struct cgraph_node *old_version,
1882 VEC(cgraph_edge_p,heap) *redirect_callers,
1885 struct cgraph_node *new_version;
1886 struct cgraph_edge *e;
1889 gcc_assert (old_version);
1891 new_version = cgraph_node (new_decl);
1893 new_version->analyzed = true;
1894 new_version->local = old_version->local;
1895 new_version->local.externally_visible = false;
1896 new_version->local.local = true;
1897 new_version->local.vtable_method = false;
1898 new_version->global = old_version->global;
1899 new_version->rtl = old_version->rtl;
1900 new_version->reachable = true;
1901 new_version->count = old_version->count;
1903 for (e = old_version->callees; e; e=e->next_callee)
1905 || bitmap_bit_p (bbs_to_copy, gimple_bb (e->call_stmt)->index))
1906 cgraph_clone_edge (e, new_version, e->call_stmt,
1907 e->lto_stmt_uid, REG_BR_PROB_BASE,
1909 e->loop_nest, true);
1910 for (e = old_version->indirect_calls; e; e=e->next_callee)
1912 || bitmap_bit_p (bbs_to_copy, gimple_bb (e->call_stmt)->index))
1913 cgraph_clone_edge (e, new_version, e->call_stmt,
1914 e->lto_stmt_uid, REG_BR_PROB_BASE,
1916 e->loop_nest, true);
1917 FOR_EACH_VEC_ELT (cgraph_edge_p, redirect_callers, i, e)
1919 /* Redirect calls to the old version node to point to its new
1921 cgraph_redirect_edge_callee (e, new_version);
1927 /* Perform function versioning.
1928 Function versioning includes copying of the tree and
1929 a callgraph update (creating a new cgraph node and updating
1930 its callees and callers).
1932 REDIRECT_CALLERS varray includes the edges to be redirected
1935 TREE_MAP is a mapping of tree nodes we want to replace with
1936 new ones (according to results of prior analysis).
1937 OLD_VERSION_NODE is the node that is versioned.
1938 It returns the new version's cgraph node.
1939 If non-NULL ARGS_TO_SKIP determine function parameters to remove
1941 If non-NULL BLOCK_TO_COPY determine what basic blocks to copy.
1942 If non_NULL NEW_ENTRY determine new entry BB of the clone. */
1944 struct cgraph_node *
1945 cgraph_function_versioning (struct cgraph_node *old_version_node,
1946 VEC(cgraph_edge_p,heap) *redirect_callers,
1947 VEC (ipa_replace_map_p,gc)* tree_map,
1948 bitmap args_to_skip,
1950 basic_block new_entry_block,
1951 const char *clone_name)
1953 tree old_decl = old_version_node->decl;
1954 struct cgraph_node *new_version_node = NULL;
1957 if (!tree_versionable_function_p (old_decl))
1960 /* Make a new FUNCTION_DECL tree node for the
1963 new_decl = copy_node (old_decl);
1965 new_decl = build_function_decl_skip_args (old_decl, args_to_skip);
1967 /* Generate a new name for the new version. */
1968 DECL_NAME (new_decl) = clone_function_name (old_decl, clone_name);
1969 SET_DECL_ASSEMBLER_NAME (new_decl, DECL_NAME (new_decl));
1970 SET_DECL_RTL (new_decl, NULL);
1972 /* Create the new version's call-graph node.
1973 and update the edges of the new node. */
1975 cgraph_copy_node_for_versioning (old_version_node, new_decl,
1976 redirect_callers, bbs_to_copy);
1978 /* Copy the OLD_VERSION_NODE function tree to the new version. */
1979 tree_function_versioning (old_decl, new_decl, tree_map, false, args_to_skip,
1980 bbs_to_copy, new_entry_block);
1982 /* Update the new version's properties.
1983 Make The new version visible only within this translation unit. Make sure
1984 that is not weak also.
1985 ??? We cannot use COMDAT linkage because there is no
1986 ABI support for this. */
1987 cgraph_make_decl_local (new_version_node->decl);
1988 DECL_VIRTUAL_P (new_version_node->decl) = 0;
1989 new_version_node->local.externally_visible = 0;
1990 new_version_node->local.local = 1;
1991 new_version_node->lowered = true;
1993 /* Update the call_expr on the edges to call the new version node. */
1994 update_call_expr (new_version_node);
1996 cgraph_call_function_insertion_hooks (new_version_node);
1997 return new_version_node;
2000 /* Produce separate function body for inline clones so the offline copy can be
2001 modified without affecting them. */
2002 struct cgraph_node *
2003 save_inline_function_body (struct cgraph_node *node)
2005 struct cgraph_node *first_clone, *n;
2007 gcc_assert (node == cgraph_node (node->decl));
2009 cgraph_lower_function (node);
2011 first_clone = node->clones;
2013 first_clone->decl = copy_node (node->decl);
2014 cgraph_insert_node_to_hashtable (first_clone);
2015 gcc_assert (first_clone == cgraph_node (first_clone->decl));
2016 if (first_clone->next_sibling_clone)
2018 for (n = first_clone->next_sibling_clone; n->next_sibling_clone; n = n->next_sibling_clone)
2019 n->clone_of = first_clone;
2020 n->clone_of = first_clone;
2021 n->next_sibling_clone = first_clone->clones;
2022 if (first_clone->clones)
2023 first_clone->clones->prev_sibling_clone = n;
2024 first_clone->clones = first_clone->next_sibling_clone;
2025 first_clone->next_sibling_clone->prev_sibling_clone = NULL;
2026 first_clone->next_sibling_clone = NULL;
2027 gcc_assert (!first_clone->prev_sibling_clone);
2029 first_clone->clone_of = NULL;
2030 node->clones = NULL;
2032 if (first_clone->clones)
2033 for (n = first_clone->clones; n != first_clone;)
2035 gcc_assert (n->decl == node->decl);
2036 n->decl = first_clone->decl;
2039 else if (n->next_sibling_clone)
2040 n = n->next_sibling_clone;
2043 while (n != first_clone && !n->next_sibling_clone)
2045 if (n != first_clone)
2046 n = n->next_sibling_clone;
2050 /* Copy the OLD_VERSION_NODE function tree to the new version. */
2051 tree_function_versioning (node->decl, first_clone->decl, NULL, true, NULL,
2054 DECL_EXTERNAL (first_clone->decl) = 0;
2055 DECL_COMDAT_GROUP (first_clone->decl) = NULL_TREE;
2056 TREE_PUBLIC (first_clone->decl) = 0;
2057 DECL_COMDAT (first_clone->decl) = 0;
2058 VEC_free (ipa_opt_pass, heap,
2059 first_clone->ipa_transforms_to_apply);
2060 first_clone->ipa_transforms_to_apply = NULL;
2062 #ifdef ENABLE_CHECKING
2063 verify_cgraph_node (first_clone);
2068 /* Given virtual clone, turn it into actual clone. */
2070 cgraph_materialize_clone (struct cgraph_node *node)
2072 bitmap_obstack_initialize (NULL);
2073 #ifdef ENABLE_CHECKING
2074 node->former_clone_of = node->clone_of->decl;
2075 if (node->clone_of->former_clone_of)
2076 node->former_clone_of = node->clone_of->former_clone_of;
2078 /* Copy the OLD_VERSION_NODE function tree to the new version. */
2079 tree_function_versioning (node->clone_of->decl, node->decl,
2080 node->clone.tree_map, true,
2081 node->clone.args_to_skip, NULL, NULL);
2082 if (cgraph_dump_file)
2084 dump_function_to_file (node->clone_of->decl, cgraph_dump_file, dump_flags);
2085 dump_function_to_file (node->decl, cgraph_dump_file, dump_flags);
2088 /* Function is no longer clone. */
2089 if (node->next_sibling_clone)
2090 node->next_sibling_clone->prev_sibling_clone = node->prev_sibling_clone;
2091 if (node->prev_sibling_clone)
2092 node->prev_sibling_clone->next_sibling_clone = node->next_sibling_clone;
2094 node->clone_of->clones = node->next_sibling_clone;
2095 node->next_sibling_clone = NULL;
2096 node->prev_sibling_clone = NULL;
2097 if (!node->clone_of->analyzed && !node->clone_of->clones)
2099 cgraph_release_function_body (node->clone_of);
2100 cgraph_node_remove_callees (node->clone_of);
2101 ipa_remove_all_references (&node->clone_of->ref_list);
2103 node->clone_of = NULL;
2104 bitmap_obstack_release (NULL);
2107 /* If necessary, change the function declaration in the call statement
2108 associated with E so that it corresponds to the edge callee. */
2111 cgraph_redirect_edge_call_stmt_to_callee (struct cgraph_edge *e)
2113 tree decl = gimple_call_fndecl (e->call_stmt);
2115 #ifdef ENABLE_CHECKING
2116 struct cgraph_node *node;
2119 if (e->indirect_unknown_callee
2120 || decl == e->callee->decl
2121 /* Don't update call from same body alias to the real function. */
2122 || (decl && cgraph_get_node (decl) == cgraph_get_node (e->callee->decl)))
2123 return e->call_stmt;
2125 #ifdef ENABLE_CHECKING
2128 node = cgraph_get_node (decl);
2129 gcc_assert (!node || !node->clone.combined_args_to_skip);
2133 if (cgraph_dump_file)
2135 fprintf (cgraph_dump_file, "updating call of %s/%i -> %s/%i: ",
2136 cgraph_node_name (e->caller), e->caller->uid,
2137 cgraph_node_name (e->callee), e->callee->uid);
2138 print_gimple_stmt (cgraph_dump_file, e->call_stmt, 0, dump_flags);
2139 if (e->callee->clone.combined_args_to_skip)
2141 fprintf (cgraph_dump_file, " combined args to skip: ");
2142 dump_bitmap (cgraph_dump_file,
2143 e->callee->clone.combined_args_to_skip);
2147 if (e->callee->clone.combined_args_to_skip)
2149 gimple_stmt_iterator gsi;
2152 = gimple_call_copy_skip_args (e->call_stmt,
2153 e->callee->clone.combined_args_to_skip);
2155 if (gimple_vdef (new_stmt)
2156 && TREE_CODE (gimple_vdef (new_stmt)) == SSA_NAME)
2157 SSA_NAME_DEF_STMT (gimple_vdef (new_stmt)) = new_stmt;
2159 gsi = gsi_for_stmt (e->call_stmt);
2160 gsi_replace (&gsi, new_stmt, true);
2163 new_stmt = e->call_stmt;
2165 gimple_call_set_fndecl (new_stmt, e->callee->decl);
2166 update_stmt (new_stmt);
2168 cgraph_set_call_stmt_including_clones (e->caller, e->call_stmt, new_stmt);
2170 if (cgraph_dump_file)
2172 fprintf (cgraph_dump_file, " updated to:");
2173 print_gimple_stmt (cgraph_dump_file, e->call_stmt, 0, dump_flags);
2178 /* Once all functions from compilation unit are in memory, produce all clones
2179 and update all calls. We might also do this on demand if we don't want to
2180 bring all functions to memory prior compilation, but current WHOPR
2181 implementation does that and it is is bit easier to keep everything right in
2184 cgraph_materialize_all_clones (void)
2186 struct cgraph_node *node;
2187 bool stabilized = false;
2189 if (cgraph_dump_file)
2190 fprintf (cgraph_dump_file, "Materializing clones\n");
2191 #ifdef ENABLE_CHECKING
2195 /* We can also do topological order, but number of iterations should be
2196 bounded by number of IPA passes since single IPA pass is probably not
2197 going to create clones of clones it created itself. */
2201 for (node = cgraph_nodes; node; node = node->next)
2203 if (node->clone_of && node->decl != node->clone_of->decl
2204 && !gimple_has_body_p (node->decl))
2206 if (gimple_has_body_p (node->clone_of->decl))
2208 if (cgraph_dump_file)
2210 fprintf (cgraph_dump_file, "clonning %s to %s\n",
2211 cgraph_node_name (node->clone_of),
2212 cgraph_node_name (node));
2213 if (node->clone.tree_map)
2216 fprintf (cgraph_dump_file, " replace map: ");
2217 for (i = 0; i < VEC_length (ipa_replace_map_p,
2218 node->clone.tree_map);
2221 struct ipa_replace_map *replace_info;
2222 replace_info = VEC_index (ipa_replace_map_p,
2223 node->clone.tree_map,
2225 print_generic_expr (cgraph_dump_file, replace_info->old_tree, 0);
2226 fprintf (cgraph_dump_file, " -> ");
2227 print_generic_expr (cgraph_dump_file, replace_info->new_tree, 0);
2228 fprintf (cgraph_dump_file, "%s%s;",
2229 replace_info->replace_p ? "(replace)":"",
2230 replace_info->ref_p ? "(ref)":"");
2232 fprintf (cgraph_dump_file, "\n");
2234 if (node->clone.args_to_skip)
2236 fprintf (cgraph_dump_file, " args_to_skip: ");
2237 dump_bitmap (cgraph_dump_file, node->clone.args_to_skip);
2239 if (node->clone.args_to_skip)
2241 fprintf (cgraph_dump_file, " combined_args_to_skip:");
2242 dump_bitmap (cgraph_dump_file, node->clone.combined_args_to_skip);
2245 cgraph_materialize_clone (node);
2251 for (node = cgraph_nodes; node; node = node->next)
2252 if (!node->analyzed && node->callees)
2253 cgraph_node_remove_callees (node);
2254 if (cgraph_dump_file)
2255 fprintf (cgraph_dump_file, "Materialization Call site updates done.\n");
2256 #ifdef ENABLE_CHECKING
2259 cgraph_remove_unreachable_nodes (false, cgraph_dump_file);
2262 #include "gt-cgraphunit.h"