1 /* Callgraph based interprocedural optimizations.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008
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"
129 #include "c-common.h"
131 #include "function.h"
132 #include "ipa-prop.h"
134 #include "tree-iterator.h"
135 #include "tree-pass.h"
138 static void cgraph_expand_all_functions (void);
139 static void cgraph_mark_functions_to_output (void);
140 static void cgraph_expand_function (struct cgraph_node
*);
141 static void cgraph_output_pending_asms (void);
143 static FILE *cgraph_dump_file
;
145 /* A vector of FUNCTION_DECLs declared as static constructors. */
146 static GTY (()) VEC(tree
, gc
) *static_ctors
;
147 /* A vector of FUNCTION_DECLs declared as static destructors. */
148 static GTY (()) VEC(tree
, gc
) *static_dtors
;
150 /* When target does not have ctors and dtors, we call all constructor
151 and destructor by special initialization/destruction function
152 recognized by collect2.
154 When we are going to build this function, collect all constructors and
155 destructors and turn them into normal functions. */
158 record_cdtor_fn (tree fndecl
)
160 struct cgraph_node
*node
;
161 if (targetm
.have_ctors_dtors
162 || (!DECL_STATIC_CONSTRUCTOR (fndecl
)
163 && !DECL_STATIC_DESTRUCTOR (fndecl
)))
166 if (DECL_STATIC_CONSTRUCTOR (fndecl
))
168 VEC_safe_push (tree
, gc
, static_ctors
, fndecl
);
169 DECL_STATIC_CONSTRUCTOR (fndecl
) = 0;
171 if (DECL_STATIC_DESTRUCTOR (fndecl
))
173 VEC_safe_push (tree
, gc
, static_dtors
, fndecl
);
174 DECL_STATIC_DESTRUCTOR (fndecl
) = 0;
176 node
= cgraph_node (fndecl
);
177 node
->local
.disregard_inline_limits
= 1;
178 cgraph_mark_reachable_node (node
);
181 /* Define global constructors/destructor functions for the CDTORS, of
182 which they are LEN. The CDTORS are sorted by initialization
183 priority. If CTOR_P is true, these are constructors; otherwise,
184 they are destructors. */
187 build_cdtor (bool ctor_p
, tree
*cdtors
, size_t len
)
196 priority_type priority
;
200 /* Find the next batch of constructors/destructors with the same
201 initialization priority. */
206 p
= ctor_p
? DECL_INIT_PRIORITY (fn
) : DECL_FINI_PRIORITY (fn
);
209 else if (p
!= priority
)
211 append_to_statement_list (build_function_call_expr (fn
, 0),
216 gcc_assert (body
!= NULL_TREE
);
217 /* Generate a function to call all the function of like
219 cgraph_build_static_cdtor (ctor_p
? 'I' : 'D', body
, priority
);
223 /* Comparison function for qsort. P1 and P2 are actually of type
224 "tree *" and point to static constructors. DECL_INIT_PRIORITY is
225 used to determine the sort order. */
228 compare_ctor (const void *p1
, const void *p2
)
235 f1
= *(const tree
*)p1
;
236 f2
= *(const tree
*)p2
;
237 priority1
= DECL_INIT_PRIORITY (f1
);
238 priority2
= DECL_INIT_PRIORITY (f2
);
240 if (priority1
< priority2
)
242 else if (priority1
> priority2
)
245 /* Ensure a stable sort. */
246 return (const tree
*)p1
- (const tree
*)p2
;
249 /* Comparison function for qsort. P1 and P2 are actually of type
250 "tree *" and point to static destructors. DECL_FINI_PRIORITY is
251 used to determine the sort order. */
254 compare_dtor (const void *p1
, const void *p2
)
261 f1
= *(const tree
*)p1
;
262 f2
= *(const tree
*)p2
;
263 priority1
= DECL_FINI_PRIORITY (f1
);
264 priority2
= DECL_FINI_PRIORITY (f2
);
266 if (priority1
< priority2
)
268 else if (priority1
> priority2
)
271 /* Ensure a stable sort. */
272 return (const tree
*)p1
- (const tree
*)p2
;
275 /* Generate functions to call static constructors and destructors
276 for targets that do not support .ctors/.dtors sections. These
277 functions have magic names which are detected by collect2. */
280 cgraph_build_cdtor_fns (void)
282 if (!VEC_empty (tree
, static_ctors
))
284 gcc_assert (!targetm
.have_ctors_dtors
);
285 qsort (VEC_address (tree
, static_ctors
),
286 VEC_length (tree
, static_ctors
),
289 build_cdtor (/*ctor_p=*/true,
290 VEC_address (tree
, static_ctors
),
291 VEC_length (tree
, static_ctors
));
292 VEC_truncate (tree
, static_ctors
, 0);
295 if (!VEC_empty (tree
, static_dtors
))
297 gcc_assert (!targetm
.have_ctors_dtors
);
298 qsort (VEC_address (tree
, static_dtors
),
299 VEC_length (tree
, static_dtors
),
302 build_cdtor (/*ctor_p=*/false,
303 VEC_address (tree
, static_dtors
),
304 VEC_length (tree
, static_dtors
));
305 VEC_truncate (tree
, static_dtors
, 0);
309 /* Determine if function DECL is needed. That is, visible to something
310 either outside this translation unit, something magic in the system
314 decide_is_function_needed (struct cgraph_node
*node
, tree decl
)
316 if (MAIN_NAME_P (DECL_NAME (decl
))
317 && TREE_PUBLIC (decl
))
319 node
->local
.externally_visible
= true;
323 /* If the user told us it is used, then it must be so. */
324 if (node
->local
.externally_visible
)
327 /* ??? If the assembler name is set by hand, it is possible to assemble
328 the name later after finalizing the function and the fact is noticed
329 in assemble_name then. This is arguably a bug. */
330 if (DECL_ASSEMBLER_NAME_SET_P (decl
)
331 && TREE_SYMBOL_REFERENCED (DECL_ASSEMBLER_NAME (decl
)))
334 /* With -fkeep-inline-functions we are keeping all inline functions except
335 for extern inline ones. */
336 if (flag_keep_inline_functions
337 && DECL_DECLARED_INLINE_P (decl
)
338 && !DECL_EXTERNAL (decl
)
339 && !lookup_attribute ("always_inline", DECL_ATTRIBUTES (decl
)))
342 /* If we decided it was needed before, but at the time we didn't have
343 the body of the function available, then it's still needed. We have
344 to go back and re-check its dependencies now. */
348 /* Externally visible functions must be output. The exception is
349 COMDAT functions that must be output only when they are needed.
351 When not optimizing, also output the static functions. (see
352 PR24561), but don't do so for always_inline functions, functions
353 declared inline and nested functions. These was optimized out
354 in the original implementation and it is unclear whether we want
355 to change the behavior here. */
356 if (((TREE_PUBLIC (decl
)
357 || (!optimize
&& !node
->local
.disregard_inline_limits
358 && !DECL_DECLARED_INLINE_P (decl
)
360 && !flag_whole_program
)
361 && !DECL_COMDAT (decl
) && !DECL_EXTERNAL (decl
))
364 /* Constructors and destructors are reachable from the runtime by
366 if (DECL_STATIC_CONSTRUCTOR (decl
) || DECL_STATIC_DESTRUCTOR (decl
))
372 /* Process CGRAPH_NEW_FUNCTIONS and perform actions necessary to add these
373 functions into callgraph in a way so they look like ordinary reachable
374 functions inserted into callgraph already at construction time. */
377 cgraph_process_new_functions (void)
381 struct cgraph_node
*node
;
383 /* Note that this queue may grow as its being processed, as the new
384 functions may generate new ones. */
385 while (cgraph_new_nodes
)
387 node
= cgraph_new_nodes
;
389 cgraph_new_nodes
= cgraph_new_nodes
->next_needed
;
390 switch (cgraph_state
)
392 case CGRAPH_STATE_CONSTRUCTION
:
393 /* At construction time we just need to finalize function and move
394 it into reachable functions list. */
396 node
->next_needed
= NULL
;
397 cgraph_finalize_function (fndecl
, false);
398 cgraph_mark_reachable_node (node
);
402 case CGRAPH_STATE_IPA
:
403 case CGRAPH_STATE_IPA_SSA
:
404 /* When IPA optimization already started, do all essential
405 transformations that has been already performed on the whole
406 cgraph but not on this function. */
408 gimple_register_cfg_hooks ();
410 cgraph_analyze_function (node
);
411 push_cfun (DECL_STRUCT_FUNCTION (fndecl
));
412 current_function_decl
= fndecl
;
413 compute_inline_parameters (node
);
414 if ((cgraph_state
== CGRAPH_STATE_IPA_SSA
415 && !gimple_in_ssa_p (DECL_STRUCT_FUNCTION (fndecl
)))
416 /* When not optimizing, be sure we run early local passes anyway
419 execute_pass_list (pass_early_local_passes
.pass
.sub
);
420 free_dominance_info (CDI_POST_DOMINATORS
);
421 free_dominance_info (CDI_DOMINATORS
);
423 current_function_decl
= NULL
;
426 case CGRAPH_STATE_EXPANSION
:
427 /* Functions created during expansion shall be compiled
430 cgraph_expand_function (node
);
437 cgraph_call_function_insertion_hooks (node
);
442 /* As an GCC extension we allow redefinition of the function. The
443 semantics when both copies of bodies differ is not well defined.
444 We replace the old body with new body so in unit at a time mode
445 we always use new body, while in normal mode we may end up with
446 old body inlined into some functions and new body expanded and
449 ??? It may make more sense to use one body for inlining and other
450 body for expanding the function but this is difficult to do. */
453 cgraph_reset_node (struct cgraph_node
*node
)
455 /* If node->output is set, then we have already begun whole-unit analysis.
456 This is *not* testing for whether we've already emitted the function.
457 That case can be sort-of legitimately seen with real function redefinition
458 errors. I would argue that the front end should never present us with
459 such a case, but don't enforce that for now. */
460 gcc_assert (!node
->output
);
462 /* Reset our data structures so we can analyze the function again. */
463 memset (&node
->local
, 0, sizeof (node
->local
));
464 memset (&node
->global
, 0, sizeof (node
->global
));
465 memset (&node
->rtl
, 0, sizeof (node
->rtl
));
466 node
->analyzed
= false;
467 node
->local
.redefined_extern_inline
= true;
468 node
->local
.finalized
= false;
470 cgraph_node_remove_callees (node
);
472 /* We may need to re-queue the node for assembling in case
473 we already proceeded it and ignored as not needed or got
474 a re-declaration in IMA mode. */
477 struct cgraph_node
*n
;
479 for (n
= cgraph_nodes_queue
; n
; n
= n
->next_needed
)
488 cgraph_lower_function (struct cgraph_node
*node
)
492 tree_lowering_passes (node
->decl
);
493 node
->lowered
= true;
496 /* DECL has been parsed. Take it, queue it, compile it at the whim of the
497 logic in effect. If NESTED is true, then our caller cannot stand to have
498 the garbage collector run at the moment. We would need to either create
499 a new GC context, or just not compile right now. */
502 cgraph_finalize_function (tree decl
, bool nested
)
504 struct cgraph_node
*node
= cgraph_node (decl
);
506 if (node
->local
.finalized
)
507 cgraph_reset_node (node
);
509 node
->pid
= cgraph_max_pid
++;
510 notice_global_symbol (decl
);
511 node
->local
.finalized
= true;
512 node
->lowered
= DECL_STRUCT_FUNCTION (decl
)->cfg
!= NULL
;
513 record_cdtor_fn (node
->decl
);
515 lower_nested_functions (decl
);
516 gcc_assert (!node
->nested
);
518 if (decide_is_function_needed (node
, decl
))
519 cgraph_mark_needed_node (node
);
521 /* Since we reclaim unreachable nodes at the end of every language
522 level unit, we need to be conservative about possible entry points
524 if ((TREE_PUBLIC (decl
) && !DECL_COMDAT (decl
) && !DECL_EXTERNAL (decl
)))
525 cgraph_mark_reachable_node (node
);
527 /* If we've not yet emitted decl, tell the debug info about it. */
528 if (!TREE_ASM_WRITTEN (decl
))
529 (*debug_hooks
->deferred_inline_function
) (decl
);
531 /* Possibly warn about unused parameters. */
532 if (warn_unused_parameter
)
533 do_warn_unused_parameter (decl
);
539 /* C99 extern inline keywords allow changing of declaration after function
540 has been finalized. We need to re-decide if we want to mark the function as
544 cgraph_mark_if_needed (tree decl
)
546 struct cgraph_node
*node
= cgraph_node (decl
);
547 if (node
->local
.finalized
&& decide_is_function_needed (node
, decl
))
548 cgraph_mark_needed_node (node
);
551 /* Verify cgraph nodes of given cgraph node. */
553 verify_cgraph_node (struct cgraph_node
*node
)
555 struct cgraph_edge
*e
;
556 struct cgraph_node
*main_clone
;
557 struct function
*this_cfun
= DECL_STRUCT_FUNCTION (node
->decl
);
558 struct function
*saved_cfun
= cfun
;
559 basic_block this_block
;
560 gimple_stmt_iterator gsi
;
561 bool error_found
= false;
563 if (errorcount
|| sorrycount
)
566 timevar_push (TV_CGRAPH_VERIFY
);
567 /* debug_generic_stmt needs correct cfun */
568 set_cfun (this_cfun
);
569 for (e
= node
->callees
; e
; e
= e
->next_callee
)
572 error ("aux field set for edge %s->%s",
573 cgraph_node_name (e
->caller
), cgraph_node_name (e
->callee
));
578 error ("Execution count is negative");
581 for (e
= node
->callers
; e
; e
= e
->next_caller
)
585 error ("caller edge count is negative");
588 if (e
->frequency
< 0)
590 error ("caller edge frequency is negative");
593 if (e
->frequency
> CGRAPH_FREQ_MAX
)
595 error ("caller edge frequency is too large");
598 if (!e
->inline_failed
)
600 if (node
->global
.inlined_to
601 != (e
->caller
->global
.inlined_to
602 ? e
->caller
->global
.inlined_to
: e
->caller
))
604 error ("inlined_to pointer is wrong");
607 if (node
->callers
->next_caller
)
609 error ("multiple inline callers");
614 if (node
->global
.inlined_to
)
616 error ("inlined_to pointer set for noninline callers");
620 if (!node
->callers
&& node
->global
.inlined_to
)
622 error ("inlined_to pointer is set but no predecessors found");
625 if (node
->global
.inlined_to
== node
)
627 error ("inlined_to pointer refers to itself");
631 for (main_clone
= cgraph_node (node
->decl
); main_clone
;
632 main_clone
= main_clone
->next_clone
)
633 if (main_clone
== node
)
635 if (!cgraph_node (node
->decl
))
637 error ("node not found in cgraph_hash");
642 && gimple_body (node
->decl
)
643 && !TREE_ASM_WRITTEN (node
->decl
)
644 && (!DECL_EXTERNAL (node
->decl
) || node
->global
.inlined_to
))
648 /* The nodes we're interested in are never shared, so walk
649 the tree ignoring duplicates. */
650 struct pointer_set_t
*visited_nodes
= pointer_set_create ();
651 /* Reach the trees by walking over the CFG, and note the
652 enclosing basic-blocks in the call edges. */
653 FOR_EACH_BB_FN (this_block
, this_cfun
)
654 for (gsi
= gsi_start_bb (this_block
);
658 gimple stmt
= gsi_stmt (gsi
);
660 if (is_gimple_call (stmt
) && (decl
= gimple_call_fndecl (stmt
)))
662 struct cgraph_edge
*e
= cgraph_edge (node
, stmt
);
667 error ("shared call_stmt:");
668 debug_gimple_stmt (stmt
);
671 if (e
->callee
->decl
!= cgraph_node (decl
)->decl
674 error ("edge points to wrong declaration:");
675 debug_tree (e
->callee
->decl
);
676 fprintf (stderr
," Instead of:");
683 error ("missing callgraph edge for call stmt:");
684 debug_gimple_stmt (stmt
);
689 pointer_set_destroy (visited_nodes
);
692 /* No CFG available?! */
695 for (e
= node
->callees
; e
; e
= e
->next_callee
)
697 if (!e
->aux
&& !e
->indirect_call
)
699 error ("edge %s->%s has no corresponding call_stmt",
700 cgraph_node_name (e
->caller
),
701 cgraph_node_name (e
->callee
));
702 debug_gimple_stmt (e
->call_stmt
);
710 dump_cgraph_node (stderr
, node
);
711 internal_error ("verify_cgraph_node failed");
713 set_cfun (saved_cfun
);
714 timevar_pop (TV_CGRAPH_VERIFY
);
717 /* Verify whole cgraph structure. */
721 struct cgraph_node
*node
;
723 if (sorrycount
|| errorcount
)
726 for (node
= cgraph_nodes
; node
; node
= node
->next
)
727 verify_cgraph_node (node
);
730 /* Output all asm statements we have stored up to be output. */
733 cgraph_output_pending_asms (void)
735 struct cgraph_asm_node
*can
;
737 if (errorcount
|| sorrycount
)
740 for (can
= cgraph_asm_nodes
; can
; can
= can
->next
)
741 assemble_asm (can
->asm_str
);
742 cgraph_asm_nodes
= NULL
;
745 /* Analyze the function scheduled to be output. */
747 cgraph_analyze_function (struct cgraph_node
*node
)
749 tree decl
= node
->decl
;
751 current_function_decl
= decl
;
752 push_cfun (DECL_STRUCT_FUNCTION (decl
));
753 cgraph_lower_function (node
);
754 node
->analyzed
= true;
757 current_function_decl
= NULL
;
760 /* Look for externally_visible and used attributes and mark cgraph nodes
763 We cannot mark the nodes at the point the attributes are processed (in
764 handle_*_attribute) because the copy of the declarations available at that
765 point may not be canonical. For example, in:
768 void f() __attribute__((used));
770 the declaration we see in handle_used_attribute will be the second
771 declaration -- but the front end will subsequently merge that declaration
772 with the original declaration and discard the second declaration.
774 Furthermore, we can't mark these nodes in cgraph_finalize_function because:
777 void f() __attribute__((externally_visible));
781 So, we walk the nodes at the end of the translation unit, applying the
782 attributes at that point. */
785 process_function_and_variable_attributes (struct cgraph_node
*first
,
786 struct varpool_node
*first_var
)
788 struct cgraph_node
*node
;
789 struct varpool_node
*vnode
;
791 for (node
= cgraph_nodes
; node
!= first
; node
= node
->next
)
793 tree decl
= node
->decl
;
794 if (lookup_attribute ("used", DECL_ATTRIBUTES (decl
)))
796 mark_decl_referenced (decl
);
797 if (node
->local
.finalized
)
798 cgraph_mark_needed_node (node
);
800 if (lookup_attribute ("externally_visible", DECL_ATTRIBUTES (decl
)))
802 if (! TREE_PUBLIC (node
->decl
))
803 warning (OPT_Wattributes
,
804 "%J%<externally_visible%> attribute have effect only on public objects",
808 if (node
->local
.finalized
)
809 cgraph_mark_needed_node (node
);
810 node
->local
.externally_visible
= true;
814 for (vnode
= varpool_nodes
; vnode
!= first_var
; vnode
= vnode
->next
)
816 tree decl
= vnode
->decl
;
817 if (lookup_attribute ("used", DECL_ATTRIBUTES (decl
)))
819 mark_decl_referenced (decl
);
820 if (vnode
->finalized
)
821 varpool_mark_needed_node (vnode
);
823 if (lookup_attribute ("externally_visible", DECL_ATTRIBUTES (decl
)))
825 if (! TREE_PUBLIC (vnode
->decl
))
826 warning (OPT_Wattributes
,
827 "%J%<externally_visible%> attribute have effect only on public objects",
831 if (vnode
->finalized
)
832 varpool_mark_needed_node (vnode
);
833 vnode
->externally_visible
= true;
839 /* Process CGRAPH_NODES_NEEDED queue, analyze each function (and transitively
840 each reachable functions) and build cgraph.
841 The function can be called multiple times after inserting new nodes
842 into beginning of queue. Just the new part of queue is re-scanned then. */
845 cgraph_analyze_functions (void)
847 /* Keep track of already processed nodes when called multiple times for
848 intermodule optimization. */
849 static struct cgraph_node
*first_analyzed
;
850 struct cgraph_node
*first_processed
= first_analyzed
;
851 static struct varpool_node
*first_analyzed_var
;
852 struct cgraph_node
*node
, *next
;
854 process_function_and_variable_attributes (first_processed
,
856 first_processed
= cgraph_nodes
;
857 first_analyzed_var
= varpool_nodes
;
858 varpool_analyze_pending_decls ();
859 if (cgraph_dump_file
)
861 fprintf (cgraph_dump_file
, "Initial entry points:");
862 for (node
= cgraph_nodes
; node
!= first_analyzed
; node
= node
->next
)
863 if (node
->needed
&& gimple_body (node
->decl
))
864 fprintf (cgraph_dump_file
, " %s", cgraph_node_name (node
));
865 fprintf (cgraph_dump_file
, "\n");
867 cgraph_process_new_functions ();
869 /* Propagate reachability flag and lower representation of all reachable
870 functions. In the future, lowering will introduce new functions and
871 new entry points on the way (by template instantiation and virtual
872 method table generation for instance). */
873 while (cgraph_nodes_queue
)
875 struct cgraph_edge
*edge
;
876 tree decl
= cgraph_nodes_queue
->decl
;
878 node
= cgraph_nodes_queue
;
879 cgraph_nodes_queue
= cgraph_nodes_queue
->next_needed
;
880 node
->next_needed
= NULL
;
882 /* ??? It is possible to create extern inline function and later using
883 weak alias attribute to kill its body. See
884 gcc.c-torture/compile/20011119-1.c */
885 if (!DECL_STRUCT_FUNCTION (decl
))
887 cgraph_reset_node (node
);
891 gcc_assert (!node
->analyzed
&& node
->reachable
);
892 gcc_assert (gimple_body (decl
));
894 cgraph_analyze_function (node
);
896 for (edge
= node
->callees
; edge
; edge
= edge
->next_callee
)
897 if (!edge
->callee
->reachable
)
898 cgraph_mark_reachable_node (edge
->callee
);
900 /* We finalize local static variables during constructing callgraph
901 edges. Process their attributes too. */
902 process_function_and_variable_attributes (first_processed
,
904 first_processed
= cgraph_nodes
;
905 first_analyzed_var
= varpool_nodes
;
906 varpool_analyze_pending_decls ();
907 cgraph_process_new_functions ();
910 /* Collect entry points to the unit. */
911 if (cgraph_dump_file
)
913 fprintf (cgraph_dump_file
, "Unit entry points:");
914 for (node
= cgraph_nodes
; node
!= first_analyzed
; node
= node
->next
)
915 if (node
->needed
&& gimple_body (node
->decl
))
916 fprintf (cgraph_dump_file
, " %s", cgraph_node_name (node
));
917 fprintf (cgraph_dump_file
, "\n\nInitial ");
918 dump_cgraph (cgraph_dump_file
);
921 if (cgraph_dump_file
)
922 fprintf (cgraph_dump_file
, "\nReclaiming functions:");
924 for (node
= cgraph_nodes
; node
!= first_analyzed
; node
= next
)
926 tree decl
= node
->decl
;
929 if (node
->local
.finalized
&& !gimple_body (decl
))
930 cgraph_reset_node (node
);
932 if (!node
->reachable
&& gimple_body (decl
))
934 if (cgraph_dump_file
)
935 fprintf (cgraph_dump_file
, " %s", cgraph_node_name (node
));
936 cgraph_remove_node (node
);
940 node
->next_needed
= NULL
;
941 gcc_assert (!node
->local
.finalized
|| gimple_body (decl
));
942 gcc_assert (node
->analyzed
== node
->local
.finalized
);
944 if (cgraph_dump_file
)
946 fprintf (cgraph_dump_file
, "\n\nReclaimed ");
947 dump_cgraph (cgraph_dump_file
);
949 first_analyzed
= cgraph_nodes
;
953 /* Analyze the whole compilation unit once it is parsed completely. */
956 cgraph_finalize_compilation_unit (void)
958 if (errorcount
|| sorrycount
)
965 fprintf (stderr
, "\nAnalyzing compilation unit\n");
969 timevar_push (TV_CGRAPH
);
970 cgraph_analyze_functions ();
971 timevar_pop (TV_CGRAPH
);
973 /* Figure out what functions we want to assemble. */
976 cgraph_mark_functions_to_output (void)
978 struct cgraph_node
*node
;
980 for (node
= cgraph_nodes
; node
; node
= node
->next
)
982 tree decl
= node
->decl
;
983 struct cgraph_edge
*e
;
985 gcc_assert (!node
->output
);
987 for (e
= node
->callers
; e
; e
= e
->next_caller
)
988 if (e
->inline_failed
)
991 /* We need to output all local functions that are used and not
992 always inlined, as well as those that are reachable from
993 outside the current compilation unit. */
994 if (gimple_body (decl
)
995 && !node
->global
.inlined_to
997 || (e
&& node
->reachable
))
998 && !TREE_ASM_WRITTEN (decl
)
999 && !DECL_EXTERNAL (decl
))
1003 /* We should've reclaimed all functions that are not needed. */
1004 #ifdef ENABLE_CHECKING
1005 if (!node
->global
.inlined_to
1006 && gimple_body (decl
)
1007 && !DECL_EXTERNAL (decl
))
1009 dump_cgraph_node (stderr
, node
);
1010 internal_error ("failed to reclaim unneeded function");
1013 gcc_assert (node
->global
.inlined_to
1014 || !gimple_body (decl
)
1015 || DECL_EXTERNAL (decl
));
1022 /* Expand function specified by NODE. */
1025 cgraph_expand_function (struct cgraph_node
*node
)
1027 tree decl
= node
->decl
;
1029 /* We ought to not compile any inline clones. */
1030 gcc_assert (!node
->global
.inlined_to
);
1032 announce_function (decl
);
1034 gcc_assert (node
->lowered
);
1036 /* Generate RTL for the body of DECL. */
1037 if (lang_hooks
.callgraph
.emit_associated_thunks
)
1038 lang_hooks
.callgraph
.emit_associated_thunks (decl
);
1039 tree_rest_of_compilation (decl
);
1041 /* Make sure that BE didn't give up on compiling. */
1042 /* ??? Can happen with nested function of extern inline. */
1043 gcc_assert (TREE_ASM_WRITTEN (decl
));
1044 current_function_decl
= NULL
;
1045 if (!cgraph_preserve_function_body_p (decl
))
1047 cgraph_release_function_body (node
);
1048 /* Eliminate all call edges. This is important so the call_expr no longer
1049 points to the dead function body. */
1050 cgraph_node_remove_callees (node
);
1053 cgraph_function_flags_ready
= true;
1056 /* Return true when CALLER_DECL should be inlined into CALLEE_DECL. */
1059 cgraph_inline_p (struct cgraph_edge
*e
, const char **reason
)
1061 *reason
= e
->inline_failed
;
1062 return !e
->inline_failed
;
1067 /* Expand all functions that must be output.
1069 Attempt to topologically sort the nodes so function is output when
1070 all called functions are already assembled to allow data to be
1071 propagated across the callgraph. Use a stack to get smaller distance
1072 between a function and its callees (later we may choose to use a more
1073 sophisticated algorithm for function reordering; we will likely want
1074 to use subsections to make the output functions appear in top-down
1078 cgraph_expand_all_functions (void)
1080 struct cgraph_node
*node
;
1081 struct cgraph_node
**order
= XCNEWVEC (struct cgraph_node
*, cgraph_n_nodes
);
1082 int order_pos
, new_order_pos
= 0;
1085 order_pos
= cgraph_postorder (order
);
1086 gcc_assert (order_pos
== cgraph_n_nodes
);
1088 /* Garbage collector may remove inline clones we eliminate during
1089 optimization. So we must be sure to not reference them. */
1090 for (i
= 0; i
< order_pos
; i
++)
1091 if (order
[i
]->output
)
1092 order
[new_order_pos
++] = order
[i
];
1094 for (i
= new_order_pos
- 1; i
>= 0; i
--)
1099 gcc_assert (node
->reachable
);
1101 cgraph_expand_function (node
);
1104 cgraph_process_new_functions ();
1110 /* This is used to sort the node types by the cgraph order number. */
1112 struct cgraph_order_sort
1114 enum { ORDER_UNDEFINED
= 0, ORDER_FUNCTION
, ORDER_VAR
, ORDER_ASM
} kind
;
1117 struct cgraph_node
*f
;
1118 struct varpool_node
*v
;
1119 struct cgraph_asm_node
*a
;
1123 /* Output all functions, variables, and asm statements in the order
1124 according to their order fields, which is the order in which they
1125 appeared in the file. This implements -fno-toplevel-reorder. In
1126 this mode we may output functions and variables which don't really
1127 need to be output. */
1130 cgraph_output_in_order (void)
1134 struct cgraph_order_sort
*nodes
;
1136 struct cgraph_node
*pf
;
1137 struct varpool_node
*pv
;
1138 struct cgraph_asm_node
*pa
;
1141 size
= max
* sizeof (struct cgraph_order_sort
);
1142 nodes
= (struct cgraph_order_sort
*) alloca (size
);
1143 memset (nodes
, 0, size
);
1145 varpool_analyze_pending_decls ();
1147 for (pf
= cgraph_nodes
; pf
; pf
= pf
->next
)
1152 gcc_assert (nodes
[i
].kind
== ORDER_UNDEFINED
);
1153 nodes
[i
].kind
= ORDER_FUNCTION
;
1158 for (pv
= varpool_nodes_queue
; pv
; pv
= pv
->next_needed
)
1161 gcc_assert (nodes
[i
].kind
== ORDER_UNDEFINED
);
1162 nodes
[i
].kind
= ORDER_VAR
;
1166 for (pa
= cgraph_asm_nodes
; pa
; pa
= pa
->next
)
1169 gcc_assert (nodes
[i
].kind
== ORDER_UNDEFINED
);
1170 nodes
[i
].kind
= ORDER_ASM
;
1174 /* In toplevel reorder mode we output all statics; mark them as needed. */
1175 for (i
= 0; i
< max
; ++i
)
1177 if (nodes
[i
].kind
== ORDER_VAR
)
1179 varpool_mark_needed_node (nodes
[i
].u
.v
);
1182 varpool_empty_needed_queue ();
1184 for (i
= 0; i
< max
; ++i
)
1186 switch (nodes
[i
].kind
)
1188 case ORDER_FUNCTION
:
1189 nodes
[i
].u
.f
->output
= 0;
1190 cgraph_expand_function (nodes
[i
].u
.f
);
1194 varpool_assemble_decl (nodes
[i
].u
.v
);
1198 assemble_asm (nodes
[i
].u
.a
->asm_str
);
1201 case ORDER_UNDEFINED
:
1209 cgraph_asm_nodes
= NULL
;
1212 /* Return true when function body of DECL still needs to be kept around
1213 for later re-use. */
1215 cgraph_preserve_function_body_p (tree decl
)
1217 struct cgraph_node
*node
;
1219 gcc_assert (cgraph_global_info_ready
);
1220 /* Look if there is any clone around. */
1221 for (node
= cgraph_node (decl
); node
; node
= node
->next_clone
)
1222 if (node
->global
.inlined_to
)
1231 current_function_decl
= NULL
;
1232 gimple_register_cfg_hooks ();
1233 bitmap_obstack_initialize (NULL
);
1234 execute_ipa_pass_list (all_ipa_passes
);
1235 bitmap_obstack_release (NULL
);
1238 /* Perform simple optimizations based on callgraph. */
1241 cgraph_optimize (void)
1243 if (errorcount
|| sorrycount
)
1246 #ifdef ENABLE_CHECKING
1250 /* Call functions declared with the "constructor" or "destructor"
1252 cgraph_build_cdtor_fns ();
1254 /* Frontend may output common variables after the unit has been finalized.
1255 It is safe to deal with them here as they are always zero initialized. */
1256 varpool_analyze_pending_decls ();
1257 cgraph_analyze_functions ();
1259 timevar_push (TV_CGRAPHOPT
);
1260 if (pre_ipa_mem_report
)
1262 fprintf (stderr
, "Memory consumption before IPA\n");
1263 dump_memory_report (false);
1266 fprintf (stderr
, "Performing interprocedural optimizations\n");
1267 cgraph_state
= CGRAPH_STATE_IPA
;
1269 /* Don't run the IPA passes if there was any error or sorry messages. */
1270 if (errorcount
== 0 && sorrycount
== 0)
1273 /* This pass remove bodies of extern inline functions we never inlined.
1274 Do this later so other IPA passes see what is really going on. */
1275 cgraph_remove_unreachable_nodes (false, dump_file
);
1276 cgraph_global_info_ready
= true;
1277 if (cgraph_dump_file
)
1279 fprintf (cgraph_dump_file
, "Optimized ");
1280 dump_cgraph (cgraph_dump_file
);
1281 dump_varpool (cgraph_dump_file
);
1283 if (post_ipa_mem_report
)
1285 fprintf (stderr
, "Memory consumption after IPA\n");
1286 dump_memory_report (false);
1288 timevar_pop (TV_CGRAPHOPT
);
1290 /* Output everything. */
1292 fprintf (stderr
, "Assembling functions:\n");
1293 #ifdef ENABLE_CHECKING
1297 cgraph_mark_functions_to_output ();
1299 cgraph_state
= CGRAPH_STATE_EXPANSION
;
1300 if (!flag_toplevel_reorder
)
1301 cgraph_output_in_order ();
1304 cgraph_output_pending_asms ();
1306 cgraph_expand_all_functions ();
1307 varpool_remove_unreferenced_decls ();
1309 varpool_assemble_pending_decls ();
1311 varpool_output_debug_info ();
1312 cgraph_process_new_functions ();
1313 cgraph_state
= CGRAPH_STATE_FINISHED
;
1315 if (cgraph_dump_file
)
1317 fprintf (cgraph_dump_file
, "\nFinal ");
1318 dump_cgraph (cgraph_dump_file
);
1320 #ifdef ENABLE_CHECKING
1322 /* Double check that all inline clones are gone and that all
1323 function bodies have been released from memory. */
1324 if (!(sorrycount
|| errorcount
))
1326 struct cgraph_node
*node
;
1327 bool error_found
= false;
1329 for (node
= cgraph_nodes
; node
; node
= node
->next
)
1331 && (node
->global
.inlined_to
1332 || gimple_body (node
->decl
)))
1335 dump_cgraph_node (stderr
, node
);
1338 internal_error ("nodes with unreleased memory found");
1342 /* Generate and emit a static constructor or destructor. WHICH must
1343 be one of 'I' (for a constructor) or 'D' (for a destructor). BODY
1344 is a STATEMENT_LIST containing GENERIC statements. PRIORITY is the
1345 initialization priority for this constructor or destructor. */
1348 cgraph_build_static_cdtor (char which
, tree body
, int priority
)
1350 static int counter
= 0;
1352 tree decl
, name
, resdecl
;
1354 /* The priority is encoded in the constructor or destructor name.
1355 collect2 will sort the names and arrange that they are called at
1357 sprintf (which_buf
, "%c_%.5d_%d", which
, priority
, counter
++);
1358 name
= get_file_function_name (which_buf
);
1360 decl
= build_decl (FUNCTION_DECL
, name
,
1361 build_function_type (void_type_node
, void_list_node
));
1362 current_function_decl
= decl
;
1364 resdecl
= build_decl (RESULT_DECL
, NULL_TREE
, void_type_node
);
1365 DECL_ARTIFICIAL (resdecl
) = 1;
1366 DECL_RESULT (decl
) = resdecl
;
1368 allocate_struct_function (decl
, false);
1370 TREE_STATIC (decl
) = 1;
1371 TREE_USED (decl
) = 1;
1372 DECL_ARTIFICIAL (decl
) = 1;
1373 DECL_NO_INSTRUMENT_FUNCTION_ENTRY_EXIT (decl
) = 1;
1374 DECL_SAVED_TREE (decl
) = body
;
1375 TREE_PUBLIC (decl
) = ! targetm
.have_ctors_dtors
;
1376 DECL_UNINLINABLE (decl
) = 1;
1378 DECL_INITIAL (decl
) = make_node (BLOCK
);
1379 TREE_USED (DECL_INITIAL (decl
)) = 1;
1381 DECL_SOURCE_LOCATION (decl
) = input_location
;
1382 cfun
->function_end_locus
= input_location
;
1387 DECL_STATIC_CONSTRUCTOR (decl
) = 1;
1388 decl_init_priority_insert (decl
, priority
);
1391 DECL_STATIC_DESTRUCTOR (decl
) = 1;
1392 decl_fini_priority_insert (decl
, priority
);
1398 gimplify_function_tree (decl
);
1400 cgraph_add_new_function (decl
, false);
1401 cgraph_mark_needed_node (cgraph_node (decl
));
1408 cgraph_dump_file
= dump_begin (TDI_cgraph
, NULL
);
1411 /* The edges representing the callers of the NEW_VERSION node were
1412 fixed by cgraph_function_versioning (), now the call_expr in their
1413 respective tree code should be updated to call the NEW_VERSION. */
1416 update_call_expr (struct cgraph_node
*new_version
)
1418 struct cgraph_edge
*e
;
1420 gcc_assert (new_version
);
1422 /* Update the call expr on the edges to call the new version. */
1423 for (e
= new_version
->callers
; e
; e
= e
->next_caller
)
1424 gimple_call_set_fndecl (e
->call_stmt
, new_version
->decl
);
1428 /* Create a new cgraph node which is the new version of
1429 OLD_VERSION node. REDIRECT_CALLERS holds the callers
1430 edges which should be redirected to point to
1431 NEW_VERSION. ALL the callees edges of OLD_VERSION
1432 are cloned to the new version node. Return the new
1435 static struct cgraph_node
*
1436 cgraph_copy_node_for_versioning (struct cgraph_node
*old_version
,
1438 VEC(cgraph_edge_p
,heap
) *redirect_callers
)
1440 struct cgraph_node
*new_version
;
1441 struct cgraph_edge
*e
, *new_e
;
1442 struct cgraph_edge
*next_callee
;
1445 gcc_assert (old_version
);
1447 new_version
= cgraph_node (new_decl
);
1449 new_version
->analyzed
= true;
1450 new_version
->local
= old_version
->local
;
1451 new_version
->global
= old_version
->global
;
1452 new_version
->rtl
= new_version
->rtl
;
1453 new_version
->reachable
= true;
1454 new_version
->count
= old_version
->count
;
1456 /* Clone the old node callees. Recursive calls are
1458 for (e
= old_version
->callees
;e
; e
=e
->next_callee
)
1460 new_e
= cgraph_clone_edge (e
, new_version
, e
->call_stmt
, 0, e
->frequency
,
1461 e
->loop_nest
, true);
1462 new_e
->count
= e
->count
;
1464 /* Fix recursive calls.
1465 If OLD_VERSION has a recursive call after the
1466 previous edge cloning, the new version will have an edge
1467 pointing to the old version, which is wrong;
1468 Redirect it to point to the new version. */
1469 for (e
= new_version
->callees
; e
; e
= next_callee
)
1471 next_callee
= e
->next_callee
;
1472 if (e
->callee
== old_version
)
1473 cgraph_redirect_edge_callee (e
, new_version
);
1478 for (i
= 0; VEC_iterate (cgraph_edge_p
, redirect_callers
, i
, e
); i
++)
1480 /* Redirect calls to the old version node to point to its new
1482 cgraph_redirect_edge_callee (e
, new_version
);
1488 /* Perform function versioning.
1489 Function versioning includes copying of the tree and
1490 a callgraph update (creating a new cgraph node and updating
1491 its callees and callers).
1493 REDIRECT_CALLERS varray includes the edges to be redirected
1496 TREE_MAP is a mapping of tree nodes we want to replace with
1497 new ones (according to results of prior analysis).
1498 OLD_VERSION_NODE is the node that is versioned.
1499 It returns the new version's cgraph node.
1500 ARGS_TO_SKIP lists arguments to be omitted from functions
1503 struct cgraph_node
*
1504 cgraph_function_versioning (struct cgraph_node
*old_version_node
,
1505 VEC(cgraph_edge_p
,heap
) *redirect_callers
,
1506 varray_type tree_map
,
1507 bitmap args_to_skip
)
1509 tree old_decl
= old_version_node
->decl
;
1510 struct cgraph_node
*new_version_node
= NULL
;
1513 if (!tree_versionable_function_p (old_decl
))
1516 /* Make a new FUNCTION_DECL tree node for the
1519 new_decl
= copy_node (old_decl
);
1521 new_decl
= build_function_decl_skip_args (old_decl
, args_to_skip
);
1523 /* Create the new version's call-graph node.
1524 and update the edges of the new node. */
1526 cgraph_copy_node_for_versioning (old_version_node
, new_decl
,
1529 /* Copy the OLD_VERSION_NODE function tree to the new version. */
1530 tree_function_versioning (old_decl
, new_decl
, tree_map
, false, args_to_skip
);
1531 /* Update the call_expr on the edges to call the new version node. */
1532 update_call_expr (new_version_node
);
1534 /* Update the new version's properties.
1535 Make The new version visible only within this translation unit.
1536 ??? We cannot use COMDAT linkage because there is no
1537 ABI support for this. */
1538 DECL_EXTERNAL (new_version_node
->decl
) = 0;
1539 DECL_ONE_ONLY (new_version_node
->decl
) = 0;
1540 TREE_PUBLIC (new_version_node
->decl
) = 0;
1541 DECL_COMDAT (new_version_node
->decl
) = 0;
1542 new_version_node
->local
.externally_visible
= 0;
1543 new_version_node
->local
.local
= 1;
1544 new_version_node
->lowered
= true;
1545 cgraph_call_function_insertion_hooks (new_version_node
);
1546 return new_version_node
;
1549 /* Produce separate function body for inline clones so the offline copy can be
1550 modified without affecting them. */
1551 struct cgraph_node
*
1552 save_inline_function_body (struct cgraph_node
*node
)
1554 struct cgraph_node
*first_clone
;
1556 gcc_assert (node
== cgraph_node (node
->decl
));
1558 cgraph_lower_function (node
);
1560 first_clone
= node
->next_clone
;
1562 first_clone
->decl
= copy_node (node
->decl
);
1563 node
->next_clone
= NULL
;
1564 first_clone
->prev_clone
= NULL
;
1565 cgraph_insert_node_to_hashtable (first_clone
);
1566 gcc_assert (first_clone
== cgraph_node (first_clone
->decl
));
1568 /* Copy the OLD_VERSION_NODE function tree to the new version. */
1569 tree_function_versioning (node
->decl
, first_clone
->decl
, NULL
, true, NULL
);
1571 DECL_EXTERNAL (first_clone
->decl
) = 0;
1572 DECL_ONE_ONLY (first_clone
->decl
) = 0;
1573 TREE_PUBLIC (first_clone
->decl
) = 0;
1574 DECL_COMDAT (first_clone
->decl
) = 0;
1576 for (node
= first_clone
->next_clone
; node
; node
= node
->next_clone
)
1577 node
->decl
= first_clone
->decl
;
1578 #ifdef ENABLE_CHECKING
1579 verify_cgraph_node (first_clone
);
1584 #include "gt-cgraphunit.h"