1 /* Callgraph based interprocedural optimizations.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009
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"
130 #include "function.h"
131 #include "ipa-prop.h"
133 #include "tree-iterator.h"
134 #include "tree-pass.h"
135 #include "tree-dump.h"
137 #include "coverage.h"
139 static void cgraph_expand_all_functions (void);
140 static void cgraph_mark_functions_to_output (void);
141 static void cgraph_expand_function (struct cgraph_node
*);
142 static void cgraph_output_pending_asms (void);
143 static void cgraph_analyze_function (struct cgraph_node
*);
145 static FILE *cgraph_dump_file
;
147 /* A vector of FUNCTION_DECLs declared as static constructors. */
148 static GTY (()) VEC(tree
, gc
) *static_ctors
;
149 /* A vector of FUNCTION_DECLs declared as static destructors. */
150 static GTY (()) VEC(tree
, gc
) *static_dtors
;
152 /* When target does not have ctors and dtors, we call all constructor
153 and destructor by special initialization/destruction function
154 recognized by collect2.
156 When we are going to build this function, collect all constructors and
157 destructors and turn them into normal functions. */
160 record_cdtor_fn (tree fndecl
)
162 struct cgraph_node
*node
;
163 if (targetm
.have_ctors_dtors
164 || (!DECL_STATIC_CONSTRUCTOR (fndecl
)
165 && !DECL_STATIC_DESTRUCTOR (fndecl
)))
168 if (DECL_STATIC_CONSTRUCTOR (fndecl
))
170 VEC_safe_push (tree
, gc
, static_ctors
, fndecl
);
171 DECL_STATIC_CONSTRUCTOR (fndecl
) = 0;
173 if (DECL_STATIC_DESTRUCTOR (fndecl
))
175 VEC_safe_push (tree
, gc
, static_dtors
, fndecl
);
176 DECL_STATIC_DESTRUCTOR (fndecl
) = 0;
178 node
= cgraph_node (fndecl
);
179 node
->local
.disregard_inline_limits
= 1;
180 cgraph_mark_reachable_node (node
);
183 /* Define global constructors/destructor functions for the CDTORS, of
184 which they are LEN. The CDTORS are sorted by initialization
185 priority. If CTOR_P is true, these are constructors; otherwise,
186 they are destructors. */
189 build_cdtor (bool ctor_p
, tree
*cdtors
, size_t len
)
198 priority_type priority
;
202 /* Find the next batch of constructors/destructors with the same
203 initialization priority. */
208 p
= ctor_p
? DECL_INIT_PRIORITY (fn
) : DECL_FINI_PRIORITY (fn
);
211 else if (p
!= priority
)
213 append_to_statement_list (build_function_call_expr (UNKNOWN_LOCATION
,
219 gcc_assert (body
!= NULL_TREE
);
220 /* Generate a function to call all the function of like
222 cgraph_build_static_cdtor (ctor_p
? 'I' : 'D', body
, priority
);
226 /* Comparison function for qsort. P1 and P2 are actually of type
227 "tree *" and point to static constructors. DECL_INIT_PRIORITY is
228 used to determine the sort order. */
231 compare_ctor (const void *p1
, const void *p2
)
238 f1
= *(const tree
*)p1
;
239 f2
= *(const tree
*)p2
;
240 priority1
= DECL_INIT_PRIORITY (f1
);
241 priority2
= DECL_INIT_PRIORITY (f2
);
243 if (priority1
< priority2
)
245 else if (priority1
> priority2
)
248 /* Ensure a stable sort. */
249 return (const tree
*)p1
- (const tree
*)p2
;
252 /* Comparison function for qsort. P1 and P2 are actually of type
253 "tree *" and point to static destructors. DECL_FINI_PRIORITY is
254 used to determine the sort order. */
257 compare_dtor (const void *p1
, const void *p2
)
264 f1
= *(const tree
*)p1
;
265 f2
= *(const tree
*)p2
;
266 priority1
= DECL_FINI_PRIORITY (f1
);
267 priority2
= DECL_FINI_PRIORITY (f2
);
269 if (priority1
< priority2
)
271 else if (priority1
> priority2
)
274 /* Ensure a stable sort. */
275 return (const tree
*)p1
- (const tree
*)p2
;
278 /* Generate functions to call static constructors and destructors
279 for targets that do not support .ctors/.dtors sections. These
280 functions have magic names which are detected by collect2. */
283 cgraph_build_cdtor_fns (void)
285 if (!VEC_empty (tree
, static_ctors
))
287 gcc_assert (!targetm
.have_ctors_dtors
);
288 qsort (VEC_address (tree
, static_ctors
),
289 VEC_length (tree
, static_ctors
),
292 build_cdtor (/*ctor_p=*/true,
293 VEC_address (tree
, static_ctors
),
294 VEC_length (tree
, static_ctors
));
295 VEC_truncate (tree
, static_ctors
, 0);
298 if (!VEC_empty (tree
, static_dtors
))
300 gcc_assert (!targetm
.have_ctors_dtors
);
301 qsort (VEC_address (tree
, static_dtors
),
302 VEC_length (tree
, static_dtors
),
305 build_cdtor (/*ctor_p=*/false,
306 VEC_address (tree
, static_dtors
),
307 VEC_length (tree
, static_dtors
));
308 VEC_truncate (tree
, static_dtors
, 0);
312 /* Determine if function DECL is needed. That is, visible to something
313 either outside this translation unit, something magic in the system
317 cgraph_decide_is_function_needed (struct cgraph_node
*node
, tree decl
)
319 /* If the user told us it is used, then it must be so. */
320 if (node
->local
.externally_visible
)
323 /* ??? If the assembler name is set by hand, it is possible to assemble
324 the name later after finalizing the function and the fact is noticed
325 in assemble_name then. This is arguably a bug. */
326 if (DECL_ASSEMBLER_NAME_SET_P (decl
)
327 && TREE_SYMBOL_REFERENCED (DECL_ASSEMBLER_NAME (decl
)))
330 /* With -fkeep-inline-functions we are keeping all inline functions except
331 for extern inline ones. */
332 if (flag_keep_inline_functions
333 && DECL_DECLARED_INLINE_P (decl
)
334 && !DECL_EXTERNAL (decl
)
335 && !lookup_attribute ("always_inline", DECL_ATTRIBUTES (decl
)))
338 /* If we decided it was needed before, but at the time we didn't have
339 the body of the function available, then it's still needed. We have
340 to go back and re-check its dependencies now. */
344 /* Externally visible functions must be output. The exception is
345 COMDAT functions that must be output only when they are needed.
347 When not optimizing, also output the static functions. (see
348 PR24561), but don't do so for always_inline functions, functions
349 declared inline and nested functions. These was optimized out
350 in the original implementation and it is unclear whether we want
351 to change the behavior here. */
352 if (((TREE_PUBLIC (decl
)
353 || (!optimize
&& !node
->local
.disregard_inline_limits
354 && !DECL_DECLARED_INLINE_P (decl
)
356 && !flag_whole_program
359 && !DECL_COMDAT (decl
) && !DECL_EXTERNAL (decl
))
362 /* Constructors and destructors are reachable from the runtime by
364 if (DECL_STATIC_CONSTRUCTOR (decl
) || DECL_STATIC_DESTRUCTOR (decl
))
370 /* Process CGRAPH_NEW_FUNCTIONS and perform actions necessary to add these
371 functions into callgraph in a way so they look like ordinary reachable
372 functions inserted into callgraph already at construction time. */
375 cgraph_process_new_functions (void)
379 struct cgraph_node
*node
;
381 /* Note that this queue may grow as its being processed, as the new
382 functions may generate new ones. */
383 while (cgraph_new_nodes
)
385 node
= cgraph_new_nodes
;
387 cgraph_new_nodes
= cgraph_new_nodes
->next_needed
;
388 switch (cgraph_state
)
390 case CGRAPH_STATE_CONSTRUCTION
:
391 /* At construction time we just need to finalize function and move
392 it into reachable functions list. */
394 node
->next_needed
= NULL
;
395 cgraph_finalize_function (fndecl
, false);
396 cgraph_mark_reachable_node (node
);
400 case CGRAPH_STATE_IPA
:
401 case CGRAPH_STATE_IPA_SSA
:
402 /* When IPA optimization already started, do all essential
403 transformations that has been already performed on the whole
404 cgraph but not on this function. */
406 gimple_register_cfg_hooks ();
408 cgraph_analyze_function (node
);
409 push_cfun (DECL_STRUCT_FUNCTION (fndecl
));
410 current_function_decl
= fndecl
;
411 compute_inline_parameters (node
);
412 if ((cgraph_state
== CGRAPH_STATE_IPA_SSA
413 && !gimple_in_ssa_p (DECL_STRUCT_FUNCTION (fndecl
)))
414 /* When not optimizing, be sure we run early local passes anyway
417 execute_pass_list (pass_early_local_passes
.pass
.sub
);
418 free_dominance_info (CDI_POST_DOMINATORS
);
419 free_dominance_info (CDI_DOMINATORS
);
421 current_function_decl
= NULL
;
424 case CGRAPH_STATE_EXPANSION
:
425 /* Functions created during expansion shall be compiled
428 cgraph_expand_function (node
);
435 cgraph_call_function_insertion_hooks (node
);
440 /* As an GCC extension we allow redefinition of the function. The
441 semantics when both copies of bodies differ is not well defined.
442 We replace the old body with new body so in unit at a time mode
443 we always use new body, while in normal mode we may end up with
444 old body inlined into some functions and new body expanded and
447 ??? It may make more sense to use one body for inlining and other
448 body for expanding the function but this is difficult to do. */
451 cgraph_reset_node (struct cgraph_node
*node
)
453 /* If node->process is set, then we have already begun whole-unit analysis.
454 This is *not* testing for whether we've already emitted the function.
455 That case can be sort-of legitimately seen with real function redefinition
456 errors. I would argue that the front end should never present us with
457 such a case, but don't enforce that for now. */
458 gcc_assert (!node
->process
);
460 /* Reset our data structures so we can analyze the function again. */
461 memset (&node
->local
, 0, sizeof (node
->local
));
462 memset (&node
->global
, 0, sizeof (node
->global
));
463 memset (&node
->rtl
, 0, sizeof (node
->rtl
));
464 node
->analyzed
= false;
465 node
->local
.redefined_extern_inline
= true;
466 node
->local
.finalized
= false;
468 cgraph_node_remove_callees (node
);
470 /* We may need to re-queue the node for assembling in case
471 we already proceeded it and ignored as not needed or got
472 a re-declaration in IMA mode. */
475 struct cgraph_node
*n
;
477 for (n
= cgraph_nodes_queue
; n
; n
= n
->next_needed
)
486 cgraph_lower_function (struct cgraph_node
*node
)
492 lower_nested_functions (node
->decl
);
493 gcc_assert (!node
->nested
);
495 tree_lowering_passes (node
->decl
);
496 node
->lowered
= true;
499 /* DECL has been parsed. Take it, queue it, compile it at the whim of the
500 logic in effect. If NESTED is true, then our caller cannot stand to have
501 the garbage collector run at the moment. We would need to either create
502 a new GC context, or just not compile right now. */
505 cgraph_finalize_function (tree decl
, bool nested
)
507 struct cgraph_node
*node
= cgraph_node (decl
);
509 if (node
->local
.finalized
)
510 cgraph_reset_node (node
);
512 node
->pid
= cgraph_max_pid
++;
513 notice_global_symbol (decl
);
514 node
->local
.finalized
= true;
515 node
->lowered
= DECL_STRUCT_FUNCTION (decl
)->cfg
!= NULL
;
516 node
->finalized_by_frontend
= true;
517 record_cdtor_fn (node
->decl
);
519 if (cgraph_decide_is_function_needed (node
, decl
))
520 cgraph_mark_needed_node (node
);
522 /* Since we reclaim unreachable nodes at the end of every language
523 level unit, we need to be conservative about possible entry points
525 if ((TREE_PUBLIC (decl
) && !DECL_COMDAT (decl
) && !DECL_EXTERNAL (decl
)))
526 cgraph_mark_reachable_node (node
);
528 /* If we've not yet emitted decl, tell the debug info about it. */
529 if (!TREE_ASM_WRITTEN (decl
))
530 (*debug_hooks
->deferred_inline_function
) (decl
);
532 /* Possibly warn about unused parameters. */
533 if (warn_unused_parameter
)
534 do_warn_unused_parameter (decl
);
540 /* C99 extern inline keywords allow changing of declaration after function
541 has been finalized. We need to re-decide if we want to mark the function as
545 cgraph_mark_if_needed (tree decl
)
547 struct cgraph_node
*node
= cgraph_node (decl
);
548 if (node
->local
.finalized
&& cgraph_decide_is_function_needed (node
, decl
))
549 cgraph_mark_needed_node (node
);
552 /* Return TRUE if NODE2 is equivalent to NODE or its clone. */
554 clone_of_p (struct cgraph_node
*node
, struct cgraph_node
*node2
)
556 while (node
!= node2
&& node2
)
557 node2
= node2
->clone_of
;
558 return node2
!= NULL
;
561 /* Verify cgraph nodes of given cgraph node. */
563 verify_cgraph_node (struct cgraph_node
*node
)
565 struct cgraph_edge
*e
;
566 struct function
*this_cfun
= DECL_STRUCT_FUNCTION (node
->decl
);
567 struct function
*saved_cfun
= cfun
;
568 basic_block this_block
;
569 gimple_stmt_iterator gsi
;
570 bool error_found
= false;
572 if (errorcount
|| sorrycount
)
575 timevar_push (TV_CGRAPH_VERIFY
);
576 /* debug_generic_stmt needs correct cfun */
577 set_cfun (this_cfun
);
578 for (e
= node
->callees
; e
; e
= e
->next_callee
)
581 error ("aux field set for edge %s->%s",
582 identifier_to_locale (cgraph_node_name (e
->caller
)),
583 identifier_to_locale (cgraph_node_name (e
->callee
)));
588 error ("Execution count is negative");
591 if (node
->global
.inlined_to
&& node
->local
.externally_visible
)
593 error ("Externally visible inline clone");
596 if (node
->global
.inlined_to
&& node
->address_taken
)
598 error ("Inline clone with address taken");
601 if (node
->global
.inlined_to
&& node
->needed
)
603 error ("Inline clone is needed");
606 for (e
= node
->callers
; e
; e
= e
->next_caller
)
610 error ("caller edge count is negative");
613 if (e
->frequency
< 0)
615 error ("caller edge frequency is negative");
618 if (e
->frequency
> CGRAPH_FREQ_MAX
)
620 error ("caller edge frequency is too large");
623 if (gimple_has_body_p (e
->caller
->decl
)
624 && !e
->caller
->global
.inlined_to
626 != compute_call_stmt_bb_frequency (e
->caller
->decl
,
627 gimple_bb (e
->call_stmt
))))
629 error ("caller edge frequency %i does not match BB freqency %i",
631 compute_call_stmt_bb_frequency (e
->caller
->decl
,
632 gimple_bb (e
->call_stmt
)));
635 if (!e
->inline_failed
)
637 if (node
->global
.inlined_to
638 != (e
->caller
->global
.inlined_to
639 ? e
->caller
->global
.inlined_to
: e
->caller
))
641 error ("inlined_to pointer is wrong");
644 if (node
->callers
->next_caller
)
646 error ("multiple inline callers");
651 if (node
->global
.inlined_to
)
653 error ("inlined_to pointer set for noninline callers");
657 if (!node
->callers
&& node
->global
.inlined_to
)
659 error ("inlined_to pointer is set but no predecessors found");
662 if (node
->global
.inlined_to
== node
)
664 error ("inlined_to pointer refers to itself");
668 if (!cgraph_node (node
->decl
))
670 error ("node not found in cgraph_hash");
676 struct cgraph_node
*n
;
677 for (n
= node
->clone_of
->clones
; n
; n
= n
->next_sibling_clone
)
682 error ("node has wrong clone_of");
688 struct cgraph_node
*n
;
689 for (n
= node
->clones
; n
; n
= n
->next_sibling_clone
)
690 if (n
->clone_of
!= node
)
694 error ("node has wrong clone list");
698 if ((node
->prev_sibling_clone
|| node
->next_sibling_clone
) && !node
->clone_of
)
700 error ("node is in clone list but it is not clone");
703 if (!node
->prev_sibling_clone
&& node
->clone_of
&& node
->clone_of
->clones
!= node
)
705 error ("node has wrong prev_clone pointer");
708 if (node
->prev_sibling_clone
&& node
->prev_sibling_clone
->next_sibling_clone
!= node
)
710 error ("double linked list of clones corrupted");
714 if (node
->analyzed
&& gimple_has_body_p (node
->decl
)
715 && !TREE_ASM_WRITTEN (node
->decl
)
716 && (!DECL_EXTERNAL (node
->decl
) || node
->global
.inlined_to
)
721 /* The nodes we're interested in are never shared, so walk
722 the tree ignoring duplicates. */
723 struct pointer_set_t
*visited_nodes
= pointer_set_create ();
724 /* Reach the trees by walking over the CFG, and note the
725 enclosing basic-blocks in the call edges. */
726 FOR_EACH_BB_FN (this_block
, this_cfun
)
727 for (gsi
= gsi_start_bb (this_block
);
731 gimple stmt
= gsi_stmt (gsi
);
733 if (is_gimple_call (stmt
) && (decl
= gimple_call_fndecl (stmt
)))
735 struct cgraph_edge
*e
= cgraph_edge (node
, stmt
);
740 error ("shared call_stmt:");
741 debug_gimple_stmt (stmt
);
744 if (!clone_of_p (cgraph_node (decl
), e
->callee
)
745 && !e
->callee
->global
.inlined_to
)
747 error ("edge points to wrong declaration:");
748 debug_tree (e
->callee
->decl
);
749 fprintf (stderr
," Instead of:");
756 error ("missing callgraph edge for call stmt:");
757 debug_gimple_stmt (stmt
);
762 pointer_set_destroy (visited_nodes
);
765 /* No CFG available?! */
768 for (e
= node
->callees
; e
; e
= e
->next_callee
)
770 if (!e
->aux
&& !e
->indirect_call
)
772 error ("edge %s->%s has no corresponding call_stmt",
773 identifier_to_locale (cgraph_node_name (e
->caller
)),
774 identifier_to_locale (cgraph_node_name (e
->callee
)));
775 debug_gimple_stmt (e
->call_stmt
);
783 dump_cgraph_node (stderr
, node
);
784 internal_error ("verify_cgraph_node failed");
786 set_cfun (saved_cfun
);
787 timevar_pop (TV_CGRAPH_VERIFY
);
790 /* Verify whole cgraph structure. */
794 struct cgraph_node
*node
;
796 if (sorrycount
|| errorcount
)
799 for (node
= cgraph_nodes
; node
; node
= node
->next
)
800 verify_cgraph_node (node
);
803 /* Output all asm statements we have stored up to be output. */
806 cgraph_output_pending_asms (void)
808 struct cgraph_asm_node
*can
;
810 if (errorcount
|| sorrycount
)
813 for (can
= cgraph_asm_nodes
; can
; can
= can
->next
)
814 assemble_asm (can
->asm_str
);
815 cgraph_asm_nodes
= NULL
;
818 /* Analyze the function scheduled to be output. */
820 cgraph_analyze_function (struct cgraph_node
*node
)
822 tree save
= current_function_decl
;
823 tree decl
= node
->decl
;
825 current_function_decl
= decl
;
826 push_cfun (DECL_STRUCT_FUNCTION (decl
));
828 /* Make sure to gimplify bodies only once. During analyzing a
829 function we lower it, which will require gimplified nested
830 functions, so we can end up here with an already gimplified
832 if (!gimple_body (decl
))
833 gimplify_function_tree (decl
);
834 dump_function (TDI_generic
, decl
);
836 cgraph_lower_function (node
);
837 node
->analyzed
= true;
840 current_function_decl
= save
;
843 /* Look for externally_visible and used attributes and mark cgraph nodes
846 We cannot mark the nodes at the point the attributes are processed (in
847 handle_*_attribute) because the copy of the declarations available at that
848 point may not be canonical. For example, in:
851 void f() __attribute__((used));
853 the declaration we see in handle_used_attribute will be the second
854 declaration -- but the front end will subsequently merge that declaration
855 with the original declaration and discard the second declaration.
857 Furthermore, we can't mark these nodes in cgraph_finalize_function because:
860 void f() __attribute__((externally_visible));
864 So, we walk the nodes at the end of the translation unit, applying the
865 attributes at that point. */
868 process_function_and_variable_attributes (struct cgraph_node
*first
,
869 struct varpool_node
*first_var
)
871 struct cgraph_node
*node
;
872 struct varpool_node
*vnode
;
874 for (node
= cgraph_nodes
; node
!= first
; node
= node
->next
)
876 tree decl
= node
->decl
;
877 if (lookup_attribute ("used", DECL_ATTRIBUTES (decl
)))
879 mark_decl_referenced (decl
);
880 if (node
->local
.finalized
)
881 cgraph_mark_needed_node (node
);
883 if (lookup_attribute ("externally_visible", DECL_ATTRIBUTES (decl
)))
885 if (! TREE_PUBLIC (node
->decl
))
886 warning_at (DECL_SOURCE_LOCATION (node
->decl
), OPT_Wattributes
,
887 "%<externally_visible%>"
888 " attribute have effect only on public objects");
889 else if (node
->local
.finalized
)
890 cgraph_mark_needed_node (node
);
893 for (vnode
= varpool_nodes
; vnode
!= first_var
; vnode
= vnode
->next
)
895 tree decl
= vnode
->decl
;
896 if (lookup_attribute ("used", DECL_ATTRIBUTES (decl
)))
898 mark_decl_referenced (decl
);
899 vnode
->force_output
= true;
900 if (vnode
->finalized
)
901 varpool_mark_needed_node (vnode
);
903 if (lookup_attribute ("externally_visible", DECL_ATTRIBUTES (decl
)))
905 if (! TREE_PUBLIC (vnode
->decl
))
906 warning_at (DECL_SOURCE_LOCATION (vnode
->decl
), OPT_Wattributes
,
907 "%<externally_visible%>"
908 " attribute have effect only on public objects");
909 else if (vnode
->finalized
)
910 varpool_mark_needed_node (vnode
);
915 /* Process CGRAPH_NODES_NEEDED queue, analyze each function (and transitively
916 each reachable functions) and build cgraph.
917 The function can be called multiple times after inserting new nodes
918 into beginning of queue. Just the new part of queue is re-scanned then. */
921 cgraph_analyze_functions (void)
923 /* Keep track of already processed nodes when called multiple times for
924 intermodule optimization. */
925 static struct cgraph_node
*first_analyzed
;
926 struct cgraph_node
*first_processed
= first_analyzed
;
927 static struct varpool_node
*first_analyzed_var
;
928 struct cgraph_node
*node
, *next
;
930 process_function_and_variable_attributes (first_processed
,
932 first_processed
= cgraph_nodes
;
933 first_analyzed_var
= varpool_nodes
;
934 varpool_analyze_pending_decls ();
935 if (cgraph_dump_file
)
937 fprintf (cgraph_dump_file
, "Initial entry points:");
938 for (node
= cgraph_nodes
; node
!= first_analyzed
; node
= node
->next
)
940 fprintf (cgraph_dump_file
, " %s", cgraph_node_name (node
));
941 fprintf (cgraph_dump_file
, "\n");
943 cgraph_process_new_functions ();
945 /* Propagate reachability flag and lower representation of all reachable
946 functions. In the future, lowering will introduce new functions and
947 new entry points on the way (by template instantiation and virtual
948 method table generation for instance). */
949 while (cgraph_nodes_queue
)
951 struct cgraph_edge
*edge
;
952 tree decl
= cgraph_nodes_queue
->decl
;
954 node
= cgraph_nodes_queue
;
955 cgraph_nodes_queue
= cgraph_nodes_queue
->next_needed
;
956 node
->next_needed
= NULL
;
958 /* ??? It is possible to create extern inline function and later using
959 weak alias attribute to kill its body. See
960 gcc.c-torture/compile/20011119-1.c */
961 if (!DECL_STRUCT_FUNCTION (decl
))
963 cgraph_reset_node (node
);
968 cgraph_analyze_function (node
);
970 for (edge
= node
->callees
; edge
; edge
= edge
->next_callee
)
971 if (!edge
->callee
->reachable
)
972 cgraph_mark_reachable_node (edge
->callee
);
974 /* If decl is a clone of an abstract function, mark that abstract
975 function so that we don't release its body. The DECL_INITIAL() of that
976 abstract function declaration will be later needed to output debug info. */
977 if (DECL_ABSTRACT_ORIGIN (decl
))
979 struct cgraph_node
*origin_node
= cgraph_node (DECL_ABSTRACT_ORIGIN (decl
));
980 origin_node
->abstract_and_needed
= true;
983 /* We finalize local static variables during constructing callgraph
984 edges. Process their attributes too. */
985 process_function_and_variable_attributes (first_processed
,
987 first_processed
= cgraph_nodes
;
988 first_analyzed_var
= varpool_nodes
;
989 varpool_analyze_pending_decls ();
990 cgraph_process_new_functions ();
993 /* Collect entry points to the unit. */
994 if (cgraph_dump_file
)
996 fprintf (cgraph_dump_file
, "Unit entry points:");
997 for (node
= cgraph_nodes
; node
!= first_analyzed
; node
= node
->next
)
999 fprintf (cgraph_dump_file
, " %s", cgraph_node_name (node
));
1000 fprintf (cgraph_dump_file
, "\n\nInitial ");
1001 dump_cgraph (cgraph_dump_file
);
1004 if (cgraph_dump_file
)
1005 fprintf (cgraph_dump_file
, "\nReclaiming functions:");
1007 for (node
= cgraph_nodes
; node
!= first_analyzed
; node
= next
)
1009 tree decl
= node
->decl
;
1012 if (node
->local
.finalized
&& !gimple_has_body_p (decl
))
1013 cgraph_reset_node (node
);
1015 if (!node
->reachable
&& gimple_has_body_p (decl
))
1017 if (cgraph_dump_file
)
1018 fprintf (cgraph_dump_file
, " %s", cgraph_node_name (node
));
1019 cgraph_remove_node (node
);
1023 node
->next_needed
= NULL
;
1024 gcc_assert (!node
->local
.finalized
|| gimple_has_body_p (decl
));
1025 gcc_assert (node
->analyzed
== node
->local
.finalized
);
1027 if (cgraph_dump_file
)
1029 fprintf (cgraph_dump_file
, "\n\nReclaimed ");
1030 dump_cgraph (cgraph_dump_file
);
1032 first_analyzed
= cgraph_nodes
;
1037 /* Emit thunks for every node in the cgraph.
1038 FIXME: We really ought to emit thunks only for functions that are needed. */
1041 cgraph_emit_thunks (void)
1043 struct cgraph_node
*n
;
1045 for (n
= cgraph_nodes
; n
; n
= n
->next
)
1047 /* Only emit thunks on functions defined in this TU.
1048 Note that this may emit more thunks than strictly necessary.
1049 During optimization some nodes may disappear. It would be
1050 nice to only emit thunks only for the functions that will be
1051 emitted, but we cannot know that until the inliner and other
1052 IPA passes have run (see the sequencing of the call to
1053 cgraph_mark_functions_to_output in cgraph_optimize). */
1055 && !DECL_EXTERNAL (n
->decl
))
1056 lang_hooks
.callgraph
.emit_associated_thunks (n
->decl
);
1061 /* Analyze the whole compilation unit once it is parsed completely. */
1064 cgraph_finalize_compilation_unit (void)
1066 timevar_push (TV_CGRAPH
);
1068 /* Do not skip analyzing the functions if there were errors, we
1069 miss diagnostics for following functions otherwise. */
1071 /* Emit size functions we didn't inline. */
1072 finalize_size_functions ();
1074 /* Call functions declared with the "constructor" or "destructor"
1076 cgraph_build_cdtor_fns ();
1078 /* Mark alias targets necessary and emit diagnostics. */
1079 finish_aliases_1 ();
1083 fprintf (stderr
, "\nAnalyzing compilation unit\n");
1087 /* Gimplify and lower all functions, compute reachability and
1088 remove unreachable nodes. */
1089 cgraph_analyze_functions ();
1091 /* Emit thunks for reachable nodes, if needed. */
1092 if (lang_hooks
.callgraph
.emit_associated_thunks
)
1093 cgraph_emit_thunks ();
1095 /* Mark alias targets necessary and emit diagnostics. */
1096 finish_aliases_1 ();
1098 /* Gimplify and lower thunks. */
1099 cgraph_analyze_functions ();
1101 /* Finally drive the pass manager. */
1104 timevar_pop (TV_CGRAPH
);
1108 /* Figure out what functions we want to assemble. */
1111 cgraph_mark_functions_to_output (void)
1113 struct cgraph_node
*node
;
1115 for (node
= cgraph_nodes
; node
; node
= node
->next
)
1117 tree decl
= node
->decl
;
1118 struct cgraph_edge
*e
;
1120 gcc_assert (!node
->process
);
1122 for (e
= node
->callers
; e
; e
= e
->next_caller
)
1123 if (e
->inline_failed
)
1126 /* We need to output all local functions that are used and not
1127 always inlined, as well as those that are reachable from
1128 outside the current compilation unit. */
1130 && !node
->global
.inlined_to
1132 || (e
&& node
->reachable
))
1133 && !TREE_ASM_WRITTEN (decl
)
1134 && !DECL_EXTERNAL (decl
))
1138 /* We should've reclaimed all functions that are not needed. */
1139 #ifdef ENABLE_CHECKING
1140 if (!node
->global
.inlined_to
1141 && gimple_has_body_p (decl
)
1142 && !DECL_EXTERNAL (decl
))
1144 dump_cgraph_node (stderr
, node
);
1145 internal_error ("failed to reclaim unneeded function");
1148 gcc_assert (node
->global
.inlined_to
1149 || !gimple_has_body_p (decl
)
1150 || DECL_EXTERNAL (decl
));
1157 /* Expand function specified by NODE. */
1160 cgraph_expand_function (struct cgraph_node
*node
)
1162 tree decl
= node
->decl
;
1164 /* We ought to not compile any inline clones. */
1165 gcc_assert (!node
->global
.inlined_to
);
1167 announce_function (decl
);
1170 gcc_assert (node
->lowered
);
1172 /* Generate RTL for the body of DECL. */
1173 tree_rest_of_compilation (decl
);
1175 /* Make sure that BE didn't give up on compiling. */
1176 gcc_assert (TREE_ASM_WRITTEN (decl
));
1177 current_function_decl
= NULL
;
1178 gcc_assert (!cgraph_preserve_function_body_p (decl
));
1179 cgraph_release_function_body (node
);
1180 /* Eliminate all call edges. This is important so the GIMPLE_CALL no longer
1181 points to the dead function body. */
1182 cgraph_node_remove_callees (node
);
1184 cgraph_function_flags_ready
= true;
1187 /* Return true when CALLER_DECL should be inlined into CALLEE_DECL. */
1190 cgraph_inline_p (struct cgraph_edge
*e
, cgraph_inline_failed_t
*reason
)
1192 *reason
= e
->inline_failed
;
1193 return !e
->inline_failed
;
1198 /* Expand all functions that must be output.
1200 Attempt to topologically sort the nodes so function is output when
1201 all called functions are already assembled to allow data to be
1202 propagated across the callgraph. Use a stack to get smaller distance
1203 between a function and its callees (later we may choose to use a more
1204 sophisticated algorithm for function reordering; we will likely want
1205 to use subsections to make the output functions appear in top-down
1209 cgraph_expand_all_functions (void)
1211 struct cgraph_node
*node
;
1212 struct cgraph_node
**order
= XCNEWVEC (struct cgraph_node
*, cgraph_n_nodes
);
1213 int order_pos
, new_order_pos
= 0;
1216 order_pos
= cgraph_postorder (order
);
1217 gcc_assert (order_pos
== cgraph_n_nodes
);
1219 /* Garbage collector may remove inline clones we eliminate during
1220 optimization. So we must be sure to not reference them. */
1221 for (i
= 0; i
< order_pos
; i
++)
1222 if (order
[i
]->process
)
1223 order
[new_order_pos
++] = order
[i
];
1225 for (i
= new_order_pos
- 1; i
>= 0; i
--)
1230 gcc_assert (node
->reachable
);
1232 cgraph_expand_function (node
);
1235 cgraph_process_new_functions ();
1241 /* This is used to sort the node types by the cgraph order number. */
1243 enum cgraph_order_sort_kind
1245 ORDER_UNDEFINED
= 0,
1251 struct cgraph_order_sort
1253 enum cgraph_order_sort_kind kind
;
1256 struct cgraph_node
*f
;
1257 struct varpool_node
*v
;
1258 struct cgraph_asm_node
*a
;
1262 /* Output all functions, variables, and asm statements in the order
1263 according to their order fields, which is the order in which they
1264 appeared in the file. This implements -fno-toplevel-reorder. In
1265 this mode we may output functions and variables which don't really
1266 need to be output. */
1269 cgraph_output_in_order (void)
1273 struct cgraph_order_sort
*nodes
;
1275 struct cgraph_node
*pf
;
1276 struct varpool_node
*pv
;
1277 struct cgraph_asm_node
*pa
;
1280 size
= max
* sizeof (struct cgraph_order_sort
);
1281 nodes
= (struct cgraph_order_sort
*) alloca (size
);
1282 memset (nodes
, 0, size
);
1284 varpool_analyze_pending_decls ();
1286 for (pf
= cgraph_nodes
; pf
; pf
= pf
->next
)
1291 gcc_assert (nodes
[i
].kind
== ORDER_UNDEFINED
);
1292 nodes
[i
].kind
= ORDER_FUNCTION
;
1297 for (pv
= varpool_nodes_queue
; pv
; pv
= pv
->next_needed
)
1300 gcc_assert (nodes
[i
].kind
== ORDER_UNDEFINED
);
1301 nodes
[i
].kind
= ORDER_VAR
;
1305 for (pa
= cgraph_asm_nodes
; pa
; pa
= pa
->next
)
1308 gcc_assert (nodes
[i
].kind
== ORDER_UNDEFINED
);
1309 nodes
[i
].kind
= ORDER_ASM
;
1313 /* In toplevel reorder mode we output all statics; mark them as needed. */
1314 for (i
= 0; i
< max
; ++i
)
1316 if (nodes
[i
].kind
== ORDER_VAR
)
1318 varpool_mark_needed_node (nodes
[i
].u
.v
);
1321 varpool_empty_needed_queue ();
1323 for (i
= 0; i
< max
; ++i
)
1325 switch (nodes
[i
].kind
)
1327 case ORDER_FUNCTION
:
1328 nodes
[i
].u
.f
->process
= 0;
1329 cgraph_expand_function (nodes
[i
].u
.f
);
1333 varpool_assemble_decl (nodes
[i
].u
.v
);
1337 assemble_asm (nodes
[i
].u
.a
->asm_str
);
1340 case ORDER_UNDEFINED
:
1348 cgraph_asm_nodes
= NULL
;
1351 /* Return true when function body of DECL still needs to be kept around
1352 for later re-use. */
1354 cgraph_preserve_function_body_p (tree decl
)
1356 struct cgraph_node
*node
;
1358 gcc_assert (cgraph_global_info_ready
);
1359 /* Look if there is any clone around. */
1360 node
= cgraph_node (decl
);
1370 current_function_decl
= NULL
;
1371 gimple_register_cfg_hooks ();
1372 bitmap_obstack_initialize (NULL
);
1375 execute_ipa_pass_list (all_small_ipa_passes
);
1377 /* If pass_all_early_optimizations was not scheduled, the state of
1378 the cgraph will not be properly updated. Update it now. */
1379 if (cgraph_state
< CGRAPH_STATE_IPA_SSA
)
1380 cgraph_state
= CGRAPH_STATE_IPA_SSA
;
1384 /* Generate coverage variables and constructors. */
1387 /* Process new functions added. */
1389 current_function_decl
= NULL
;
1390 cgraph_process_new_functions ();
1392 execute_ipa_summary_passes ((struct ipa_opt_pass_d
*) all_regular_ipa_passes
);
1394 execute_ipa_summary_passes ((struct ipa_opt_pass_d
*) all_lto_gen_passes
);
1397 ipa_write_summaries ();
1400 execute_ipa_pass_list (all_regular_ipa_passes
);
1402 bitmap_obstack_release (NULL
);
1406 /* Perform simple optimizations based on callgraph. */
1409 cgraph_optimize (void)
1411 if (errorcount
|| sorrycount
)
1414 #ifdef ENABLE_CHECKING
1418 /* Frontend may output common variables after the unit has been finalized.
1419 It is safe to deal with them here as they are always zero initialized. */
1420 varpool_analyze_pending_decls ();
1422 timevar_push (TV_CGRAPHOPT
);
1423 if (pre_ipa_mem_report
)
1425 fprintf (stderr
, "Memory consumption before IPA\n");
1426 dump_memory_report (false);
1429 fprintf (stderr
, "Performing interprocedural optimizations\n");
1430 cgraph_state
= CGRAPH_STATE_IPA
;
1432 /* Don't run the IPA passes if there was any error or sorry messages. */
1433 if (errorcount
== 0 && sorrycount
== 0)
1436 /* Do nothing else if any IPA pass found errors. */
1437 if (errorcount
|| sorrycount
)
1439 timevar_pop (TV_CGRAPHOPT
);
1443 /* This pass remove bodies of extern inline functions we never inlined.
1444 Do this later so other IPA passes see what is really going on. */
1445 cgraph_remove_unreachable_nodes (false, dump_file
);
1446 cgraph_global_info_ready
= true;
1447 if (cgraph_dump_file
)
1449 fprintf (cgraph_dump_file
, "Optimized ");
1450 dump_cgraph (cgraph_dump_file
);
1451 dump_varpool (cgraph_dump_file
);
1453 if (post_ipa_mem_report
)
1455 fprintf (stderr
, "Memory consumption after IPA\n");
1456 dump_memory_report (false);
1458 timevar_pop (TV_CGRAPHOPT
);
1460 /* Output everything. */
1461 (*debug_hooks
->assembly_start
) ();
1463 fprintf (stderr
, "Assembling functions:\n");
1464 #ifdef ENABLE_CHECKING
1468 cgraph_materialize_all_clones ();
1469 cgraph_mark_functions_to_output ();
1471 cgraph_state
= CGRAPH_STATE_EXPANSION
;
1472 if (!flag_toplevel_reorder
)
1473 cgraph_output_in_order ();
1476 cgraph_output_pending_asms ();
1478 cgraph_expand_all_functions ();
1479 varpool_remove_unreferenced_decls ();
1481 varpool_assemble_pending_decls ();
1483 cgraph_process_new_functions ();
1484 cgraph_state
= CGRAPH_STATE_FINISHED
;
1486 if (cgraph_dump_file
)
1488 fprintf (cgraph_dump_file
, "\nFinal ");
1489 dump_cgraph (cgraph_dump_file
);
1491 #ifdef ENABLE_CHECKING
1493 /* Double check that all inline clones are gone and that all
1494 function bodies have been released from memory. */
1495 if (!(sorrycount
|| errorcount
))
1497 struct cgraph_node
*node
;
1498 bool error_found
= false;
1500 for (node
= cgraph_nodes
; node
; node
= node
->next
)
1502 && (node
->global
.inlined_to
1503 || gimple_has_body_p (node
->decl
)))
1506 dump_cgraph_node (stderr
, node
);
1509 internal_error ("nodes with unreleased memory found");
1515 /* Generate and emit a static constructor or destructor. WHICH must
1516 be one of 'I' (for a constructor) or 'D' (for a destructor). BODY
1517 is a STATEMENT_LIST containing GENERIC statements. PRIORITY is the
1518 initialization priority for this constructor or destructor. */
1521 cgraph_build_static_cdtor (char which
, tree body
, int priority
)
1523 static int counter
= 0;
1525 tree decl
, name
, resdecl
;
1527 /* The priority is encoded in the constructor or destructor name.
1528 collect2 will sort the names and arrange that they are called at
1530 sprintf (which_buf
, "%c_%.5d_%d", which
, priority
, counter
++);
1531 name
= get_file_function_name (which_buf
);
1533 decl
= build_decl (input_location
, FUNCTION_DECL
, name
,
1534 build_function_type (void_type_node
, void_list_node
));
1535 current_function_decl
= decl
;
1537 resdecl
= build_decl (input_location
,
1538 RESULT_DECL
, NULL_TREE
, void_type_node
);
1539 DECL_ARTIFICIAL (resdecl
) = 1;
1540 DECL_RESULT (decl
) = resdecl
;
1541 DECL_CONTEXT (resdecl
) = decl
;
1543 allocate_struct_function (decl
, false);
1545 TREE_STATIC (decl
) = 1;
1546 TREE_USED (decl
) = 1;
1547 DECL_ARTIFICIAL (decl
) = 1;
1548 DECL_NO_INSTRUMENT_FUNCTION_ENTRY_EXIT (decl
) = 1;
1549 DECL_SAVED_TREE (decl
) = body
;
1550 TREE_PUBLIC (decl
) = ! targetm
.have_ctors_dtors
;
1551 DECL_UNINLINABLE (decl
) = 1;
1553 DECL_INITIAL (decl
) = make_node (BLOCK
);
1554 TREE_USED (DECL_INITIAL (decl
)) = 1;
1556 DECL_SOURCE_LOCATION (decl
) = input_location
;
1557 cfun
->function_end_locus
= input_location
;
1562 DECL_STATIC_CONSTRUCTOR (decl
) = 1;
1563 decl_init_priority_insert (decl
, priority
);
1566 DECL_STATIC_DESTRUCTOR (decl
) = 1;
1567 decl_fini_priority_insert (decl
, priority
);
1573 gimplify_function_tree (decl
);
1575 cgraph_add_new_function (decl
, false);
1576 cgraph_mark_needed_node (cgraph_node (decl
));
1583 cgraph_dump_file
= dump_begin (TDI_cgraph
, NULL
);
1586 /* The edges representing the callers of the NEW_VERSION node were
1587 fixed by cgraph_function_versioning (), now the call_expr in their
1588 respective tree code should be updated to call the NEW_VERSION. */
1591 update_call_expr (struct cgraph_node
*new_version
)
1593 struct cgraph_edge
*e
;
1595 gcc_assert (new_version
);
1597 /* Update the call expr on the edges to call the new version. */
1598 for (e
= new_version
->callers
; e
; e
= e
->next_caller
)
1600 struct function
*inner_function
= DECL_STRUCT_FUNCTION (e
->caller
->decl
);
1601 gimple_call_set_fndecl (e
->call_stmt
, new_version
->decl
);
1602 maybe_clean_eh_stmt_fn (inner_function
, e
->call_stmt
);
1607 /* Create a new cgraph node which is the new version of
1608 OLD_VERSION node. REDIRECT_CALLERS holds the callers
1609 edges which should be redirected to point to
1610 NEW_VERSION. ALL the callees edges of OLD_VERSION
1611 are cloned to the new version node. Return the new
1614 static struct cgraph_node
*
1615 cgraph_copy_node_for_versioning (struct cgraph_node
*old_version
,
1617 VEC(cgraph_edge_p
,heap
) *redirect_callers
)
1619 struct cgraph_node
*new_version
;
1620 struct cgraph_edge
*e
, *new_e
;
1621 struct cgraph_edge
*next_callee
;
1624 gcc_assert (old_version
);
1626 new_version
= cgraph_node (new_decl
);
1628 new_version
->analyzed
= true;
1629 new_version
->local
= old_version
->local
;
1630 new_version
->global
= old_version
->global
;
1631 new_version
->rtl
= new_version
->rtl
;
1632 new_version
->reachable
= true;
1633 new_version
->count
= old_version
->count
;
1635 /* Clone the old node callees. Recursive calls are
1637 for (e
= old_version
->callees
;e
; e
=e
->next_callee
)
1639 new_e
= cgraph_clone_edge (e
, new_version
, e
->call_stmt
,
1640 e
->lto_stmt_uid
, 0, e
->frequency
,
1641 e
->loop_nest
, true);
1642 new_e
->count
= e
->count
;
1644 /* Fix recursive calls.
1645 If OLD_VERSION has a recursive call after the
1646 previous edge cloning, the new version will have an edge
1647 pointing to the old version, which is wrong;
1648 Redirect it to point to the new version. */
1649 for (e
= new_version
->callees
; e
; e
= next_callee
)
1651 next_callee
= e
->next_callee
;
1652 if (e
->callee
== old_version
)
1653 cgraph_redirect_edge_callee (e
, new_version
);
1658 for (i
= 0; VEC_iterate (cgraph_edge_p
, redirect_callers
, i
, e
); i
++)
1660 /* Redirect calls to the old version node to point to its new
1662 cgraph_redirect_edge_callee (e
, new_version
);
1668 /* Perform function versioning.
1669 Function versioning includes copying of the tree and
1670 a callgraph update (creating a new cgraph node and updating
1671 its callees and callers).
1673 REDIRECT_CALLERS varray includes the edges to be redirected
1676 TREE_MAP is a mapping of tree nodes we want to replace with
1677 new ones (according to results of prior analysis).
1678 OLD_VERSION_NODE is the node that is versioned.
1679 It returns the new version's cgraph node.
1680 ARGS_TO_SKIP lists arguments to be omitted from functions
1683 struct cgraph_node
*
1684 cgraph_function_versioning (struct cgraph_node
*old_version_node
,
1685 VEC(cgraph_edge_p
,heap
) *redirect_callers
,
1686 VEC (ipa_replace_map_p
,gc
)* tree_map
,
1687 bitmap args_to_skip
)
1689 tree old_decl
= old_version_node
->decl
;
1690 struct cgraph_node
*new_version_node
= NULL
;
1693 if (!tree_versionable_function_p (old_decl
))
1696 /* Make a new FUNCTION_DECL tree node for the
1699 new_decl
= copy_node (old_decl
);
1701 new_decl
= build_function_decl_skip_args (old_decl
, args_to_skip
);
1703 /* Create the new version's call-graph node.
1704 and update the edges of the new node. */
1706 cgraph_copy_node_for_versioning (old_version_node
, new_decl
,
1709 /* Copy the OLD_VERSION_NODE function tree to the new version. */
1710 tree_function_versioning (old_decl
, new_decl
, tree_map
, false, args_to_skip
);
1712 /* Update the new version's properties.
1713 Make The new version visible only within this translation unit. Make sure
1714 that is not weak also.
1715 ??? We cannot use COMDAT linkage because there is no
1716 ABI support for this. */
1717 DECL_EXTERNAL (new_version_node
->decl
) = 0;
1718 DECL_COMDAT_GROUP (new_version_node
->decl
) = NULL_TREE
;
1719 TREE_PUBLIC (new_version_node
->decl
) = 0;
1720 DECL_COMDAT (new_version_node
->decl
) = 0;
1721 DECL_WEAK (new_version_node
->decl
) = 0;
1722 DECL_VIRTUAL_P (new_version_node
->decl
) = 0;
1723 new_version_node
->local
.externally_visible
= 0;
1724 new_version_node
->local
.local
= 1;
1725 new_version_node
->lowered
= true;
1727 /* Update the call_expr on the edges to call the new version node. */
1728 update_call_expr (new_version_node
);
1730 cgraph_call_function_insertion_hooks (new_version_node
);
1731 return new_version_node
;
1734 /* Produce separate function body for inline clones so the offline copy can be
1735 modified without affecting them. */
1736 struct cgraph_node
*
1737 save_inline_function_body (struct cgraph_node
*node
)
1739 struct cgraph_node
*first_clone
, *n
;
1741 gcc_assert (node
== cgraph_node (node
->decl
));
1743 cgraph_lower_function (node
);
1745 first_clone
= node
->clones
;
1747 first_clone
->decl
= copy_node (node
->decl
);
1748 cgraph_insert_node_to_hashtable (first_clone
);
1749 gcc_assert (first_clone
== cgraph_node (first_clone
->decl
));
1750 if (first_clone
->next_sibling_clone
)
1752 for (n
= first_clone
->next_sibling_clone
; n
->next_sibling_clone
; n
= n
->next_sibling_clone
)
1753 n
->clone_of
= first_clone
;
1754 n
->clone_of
= first_clone
;
1755 n
->next_sibling_clone
= first_clone
->clones
;
1756 if (first_clone
->clones
)
1757 first_clone
->clones
->prev_sibling_clone
= n
;
1758 first_clone
->clones
= first_clone
->next_sibling_clone
;
1759 first_clone
->next_sibling_clone
->prev_sibling_clone
= NULL
;
1760 first_clone
->next_sibling_clone
= NULL
;
1761 gcc_assert (!first_clone
->prev_sibling_clone
);
1763 first_clone
->clone_of
= NULL
;
1764 node
->clones
= NULL
;
1766 if (first_clone
->clones
)
1767 for (n
= first_clone
->clones
; n
!= first_clone
;)
1769 gcc_assert (n
->decl
== node
->decl
);
1770 n
->decl
= first_clone
->decl
;
1773 else if (n
->next_sibling_clone
)
1774 n
= n
->next_sibling_clone
;
1777 while (n
!= first_clone
&& !n
->next_sibling_clone
)
1779 if (n
!= first_clone
)
1780 n
= n
->next_sibling_clone
;
1784 /* Copy the OLD_VERSION_NODE function tree to the new version. */
1785 tree_function_versioning (node
->decl
, first_clone
->decl
, NULL
, true, NULL
);
1787 DECL_EXTERNAL (first_clone
->decl
) = 0;
1788 DECL_COMDAT_GROUP (first_clone
->decl
) = NULL_TREE
;
1789 TREE_PUBLIC (first_clone
->decl
) = 0;
1790 DECL_COMDAT (first_clone
->decl
) = 0;
1791 VEC_free (ipa_opt_pass
, heap
,
1792 first_clone
->ipa_transforms_to_apply
);
1793 first_clone
->ipa_transforms_to_apply
= NULL
;
1795 #ifdef ENABLE_CHECKING
1796 verify_cgraph_node (first_clone
);
1801 /* Given virtual clone, turn it into actual clone. */
1803 cgraph_materialize_clone (struct cgraph_node
*node
)
1805 bitmap_obstack_initialize (NULL
);
1806 /* Copy the OLD_VERSION_NODE function tree to the new version. */
1807 tree_function_versioning (node
->clone_of
->decl
, node
->decl
,
1808 node
->clone
.tree_map
, true,
1809 node
->clone
.args_to_skip
);
1810 if (cgraph_dump_file
)
1812 dump_function_to_file (node
->clone_of
->decl
, cgraph_dump_file
, dump_flags
);
1813 dump_function_to_file (node
->decl
, cgraph_dump_file
, dump_flags
);
1816 /* Function is no longer clone. */
1817 if (node
->next_sibling_clone
)
1818 node
->next_sibling_clone
->prev_sibling_clone
= node
->prev_sibling_clone
;
1819 if (node
->prev_sibling_clone
)
1820 node
->prev_sibling_clone
->next_sibling_clone
= node
->next_sibling_clone
;
1822 node
->clone_of
->clones
= node
->next_sibling_clone
;
1823 node
->next_sibling_clone
= NULL
;
1824 node
->prev_sibling_clone
= NULL
;
1825 if (!node
->clone_of
->analyzed
&& !node
->clone_of
->clones
)
1826 cgraph_remove_node (node
->clone_of
);
1827 node
->clone_of
= NULL
;
1828 bitmap_obstack_release (NULL
);
1831 /* Once all functions from compilation unit are in memory, produce all clones
1832 and update all calls.
1833 We might also do this on demand if we don't want to bring all functions to
1834 memory prior compilation, but current WHOPR implementation does that and it is
1835 is bit easier to keep everything right in this order. */
1837 cgraph_materialize_all_clones (void)
1839 struct cgraph_node
*node
;
1840 bool stabilized
= false;
1842 if (cgraph_dump_file
)
1843 fprintf (cgraph_dump_file
, "Materializing clones\n");
1844 #ifdef ENABLE_CHECKING
1848 /* We can also do topological order, but number of iterations should be
1849 bounded by number of IPA passes since single IPA pass is probably not
1850 going to create clones of clones it created itself. */
1854 for (node
= cgraph_nodes
; node
; node
= node
->next
)
1856 if (node
->clone_of
&& node
->decl
!= node
->clone_of
->decl
1857 && !gimple_has_body_p (node
->decl
))
1859 if (gimple_has_body_p (node
->clone_of
->decl
))
1861 if (cgraph_dump_file
)
1863 fprintf (cgraph_dump_file
, "clonning %s to %s\n",
1864 cgraph_node_name (node
->clone_of
),
1865 cgraph_node_name (node
));
1866 if (node
->clone
.tree_map
)
1869 fprintf (cgraph_dump_file
, " replace map: ");
1870 for (i
= 0; i
< VEC_length (ipa_replace_map_p
,
1871 node
->clone
.tree_map
);
1874 struct ipa_replace_map
*replace_info
;
1875 replace_info
= VEC_index (ipa_replace_map_p
,
1876 node
->clone
.tree_map
,
1878 print_generic_expr (cgraph_dump_file
, replace_info
->old_tree
, 0);
1879 fprintf (cgraph_dump_file
, " -> ");
1880 print_generic_expr (cgraph_dump_file
, replace_info
->new_tree
, 0);
1881 fprintf (cgraph_dump_file
, "%s%s;",
1882 replace_info
->replace_p
? "(replace)":"",
1883 replace_info
->ref_p
? "(ref)":"");
1885 fprintf (cgraph_dump_file
, "\n");
1887 if (node
->clone
.args_to_skip
)
1889 fprintf (cgraph_dump_file
, " args_to_skip: ");
1890 dump_bitmap (cgraph_dump_file
, node
->clone
.args_to_skip
);
1892 if (node
->clone
.args_to_skip
)
1894 fprintf (cgraph_dump_file
, " combined_args_to_skip:");
1895 dump_bitmap (cgraph_dump_file
, node
->clone
.combined_args_to_skip
);
1898 cgraph_materialize_clone (node
);
1905 if (cgraph_dump_file
)
1906 fprintf (cgraph_dump_file
, "Updating call sites\n");
1907 for (node
= cgraph_nodes
; node
; node
= node
->next
)
1908 if (node
->analyzed
&& gimple_has_body_p (node
->decl
)
1909 && (!node
->clone_of
|| node
->clone_of
->decl
!= node
->decl
))
1911 struct cgraph_edge
*e
;
1913 current_function_decl
= node
->decl
;
1914 push_cfun (DECL_STRUCT_FUNCTION (node
->decl
));
1915 for (e
= node
->callees
; e
; e
= e
->next_callee
)
1917 tree decl
= gimple_call_fndecl (e
->call_stmt
);
1918 /* When function gets inlined, indirect inlining might've invented
1919 new edge for orginally indirect stmt. Since we are not
1920 preserving clones in the original form, we must not update here
1921 since other inline clones don't need to contain call to the same
1922 call. Inliner will do the substitution for us later. */
1923 if (decl
&& decl
!= e
->callee
->decl
)
1926 gimple_stmt_iterator gsi
;
1928 if (cgraph_dump_file
)
1930 fprintf (cgraph_dump_file
, "updating call of %s in %s:",
1931 cgraph_node_name (node
),
1932 cgraph_node_name (e
->callee
));
1933 print_gimple_stmt (cgraph_dump_file
, e
->call_stmt
, 0, dump_flags
);
1936 if (e
->callee
->clone
.combined_args_to_skip
)
1937 new_stmt
= gimple_call_copy_skip_args (e
->call_stmt
,
1938 e
->callee
->clone
.combined_args_to_skip
);
1940 new_stmt
= e
->call_stmt
;
1941 if (gimple_vdef (new_stmt
)
1942 && TREE_CODE (gimple_vdef (new_stmt
)) == SSA_NAME
)
1943 SSA_NAME_DEF_STMT (gimple_vdef (new_stmt
)) = new_stmt
;
1944 gimple_call_set_fndecl (new_stmt
, e
->callee
->decl
);
1946 gsi
= gsi_for_stmt (e
->call_stmt
);
1947 gsi_replace (&gsi
, new_stmt
, true);
1949 /* Update EH information too, just in case. */
1950 maybe_clean_or_replace_eh_stmt (e
->call_stmt
, new_stmt
);
1952 cgraph_set_call_stmt_including_clones (node
, e
->call_stmt
, new_stmt
);
1954 if (cgraph_dump_file
)
1956 fprintf (cgraph_dump_file
, " updated to:");
1957 print_gimple_stmt (cgraph_dump_file
, e
->call_stmt
, 0, dump_flags
);
1962 current_function_decl
= NULL
;
1963 #ifdef ENABLE_CHECKING
1964 verify_cgraph_node (node
);
1967 #ifdef ENABLE_CHECKING
1970 cgraph_remove_unreachable_nodes (false, cgraph_dump_file
);
1973 #include "gt-cgraphunit.h"