2 Copyright (C) 2003-2014 Free Software Foundation, Inc.
3 Contributed by Jan Hubicka
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 /* This module provide facilities for clonning functions. I.e. creating
22 new functions based on existing functions with simple modifications,
23 such as replacement of parameters.
25 To allow whole program optimization without actual presence of function
26 bodies, an additional infrastructure is provided for so-called virtual
29 A virtual clone in the callgraph is a function that has no
30 associated body, just a description of how to create its body based
31 on a different function (which itself may be a virtual clone).
33 The description of function modifications includes adjustments to
34 the function's signature (which allows, for example, removing or
35 adding function arguments), substitutions to perform on the
36 function body, and, for inlined functions, a pointer to the
37 function that it will be inlined into.
39 It is also possible to redirect any edge of the callgraph from a
40 function to its virtual clone. This implies updating of the call
41 site to adjust for the new function signature.
43 Most of the transformations performed by inter-procedural
44 optimizations can be represented via virtual clones. For
45 instance, a constant propagation pass can produce a virtual clone
46 of the function which replaces one of its arguments by a
47 constant. The inliner can represent its decisions by producing a
48 clone of a function whose body will be later integrated into
51 Using virtual clones, the program can be easily updated
52 during the Execute stage, solving most of pass interactions
53 problems that would otherwise occur during Transform.
55 Virtual clones are later materialized in the LTRANS stage and
56 turned into real functions. Passes executed after the virtual
57 clone were introduced also perform their Transform stage
58 on new functions, so for a pass there is no significant
59 difference between operating on a real function or a virtual
60 clone introduced before its Execute stage.
62 Optimization passes then work on virtual clones introduced before
63 their Execute stage as if they were real functions. The
64 only difference is that clones are not visible during the
65 Generate Summary stage. */
69 #include "coretypes.h"
73 #include "stringpool.h"
76 #include "basic-block.h"
77 #include "tree-ssa-alias.h"
78 #include "internal-fn.h"
80 #include "gimple-expr.h"
85 #include "tree-inline.h"
86 #include "langhooks.h"
91 #include "diagnostic.h"
96 #include "tree-iterator.h"
97 #include "tree-dump.h"
98 #include "gimple-pretty-print.h"
100 #include "ipa-inline.h"
101 #include "ipa-utils.h"
102 #include "lto-streamer.h"
105 /* Create clone of edge in the node N represented by CALL_EXPR
109 cgraph_edge::clone (cgraph_node
*n
, gimple call_stmt
, unsigned stmt_uid
,
110 gcov_type count_scale
, int freq_scale
, bool update_original
)
112 cgraph_edge
*new_edge
;
113 gcov_type gcov_count
= apply_probability (count
, count_scale
);
116 /* We do not want to ignore loop nest after frequency drops to 0. */
119 freq
= frequency
* (gcov_type
) freq_scale
/ CGRAPH_FREQ_BASE
;
120 if (freq
> CGRAPH_FREQ_MAX
)
121 freq
= CGRAPH_FREQ_MAX
;
123 if (indirect_unknown_callee
)
127 if (call_stmt
&& (decl
= gimple_call_fndecl (call_stmt
))
128 /* When the call is speculative, we need to resolve it
129 via cgraph_resolve_speculation and not here. */
132 cgraph_node
*callee
= cgraph_node::get (decl
);
133 gcc_checking_assert (callee
);
134 new_edge
= n
->create_edge (callee
, call_stmt
, gcov_count
, freq
);
138 new_edge
= n
->create_indirect_edge (call_stmt
,
139 indirect_info
->ecf_flags
,
141 *new_edge
->indirect_info
= *indirect_info
;
146 new_edge
= n
->create_edge (callee
, call_stmt
, gcov_count
, freq
);
149 new_edge
->indirect_info
150 = ggc_cleared_alloc
<cgraph_indirect_call_info
> ();
151 *new_edge
->indirect_info
= *indirect_info
;
155 new_edge
->inline_failed
= inline_failed
;
156 new_edge
->indirect_inlining_edge
= indirect_inlining_edge
;
157 new_edge
->lto_stmt_uid
= stmt_uid
;
158 /* Clone flags that depend on call_stmt availability manually. */
159 new_edge
->can_throw_external
= can_throw_external
;
160 new_edge
->call_stmt_cannot_inline_p
= call_stmt_cannot_inline_p
;
161 new_edge
->speculative
= speculative
;
162 new_edge
->in_polymorphic_cdtor
= in_polymorphic_cdtor
;
165 count
-= new_edge
->count
;
169 symtab
->call_edge_duplication_hooks (this, new_edge
);
173 /* Build variant of function type ORIG_TYPE skipping ARGS_TO_SKIP and the
174 return value if SKIP_RETURN is true. */
177 build_function_type_skip_args (tree orig_type
, bitmap args_to_skip
,
180 tree new_type
= NULL
;
181 tree args
, new_args
= NULL
;
185 for (args
= TYPE_ARG_TYPES (orig_type
); args
&& args
!= void_list_node
;
186 args
= TREE_CHAIN (args
), i
++)
187 if (!args_to_skip
|| !bitmap_bit_p (args_to_skip
, i
))
188 new_args
= tree_cons (NULL_TREE
, TREE_VALUE (args
), new_args
);
190 new_reversed
= nreverse (new_args
);
194 TREE_CHAIN (new_args
) = void_list_node
;
196 new_reversed
= void_list_node
;
199 /* Use copy_node to preserve as much as possible from original type
200 (debug info, attribute lists etc.)
201 Exception is METHOD_TYPEs must have THIS argument.
202 When we are asked to remove it, we need to build new FUNCTION_TYPE
204 if (TREE_CODE (orig_type
) != METHOD_TYPE
206 || !bitmap_bit_p (args_to_skip
, 0))
208 new_type
= build_distinct_type_copy (orig_type
);
209 TYPE_ARG_TYPES (new_type
) = new_reversed
;
214 = build_distinct_type_copy (build_function_type (TREE_TYPE (orig_type
),
216 TYPE_CONTEXT (new_type
) = TYPE_CONTEXT (orig_type
);
220 TREE_TYPE (new_type
) = void_type_node
;
225 /* Build variant of function decl ORIG_DECL skipping ARGS_TO_SKIP and the
226 return value if SKIP_RETURN is true.
228 Arguments from DECL_ARGUMENTS list can't be removed now, since they are
229 linked by TREE_CHAIN directly. The caller is responsible for eliminating
230 them when they are being duplicated (i.e. copy_arguments_for_versioning). */
233 build_function_decl_skip_args (tree orig_decl
, bitmap args_to_skip
,
236 tree new_decl
= copy_node (orig_decl
);
239 new_type
= TREE_TYPE (orig_decl
);
240 if (prototype_p (new_type
)
241 || (skip_return
&& !VOID_TYPE_P (TREE_TYPE (new_type
))))
243 = build_function_type_skip_args (new_type
, args_to_skip
, skip_return
);
244 TREE_TYPE (new_decl
) = new_type
;
246 /* For declarations setting DECL_VINDEX (i.e. methods)
247 we expect first argument to be THIS pointer. */
248 if (args_to_skip
&& bitmap_bit_p (args_to_skip
, 0))
249 DECL_VINDEX (new_decl
) = NULL_TREE
;
251 /* When signature changes, we need to clear builtin info. */
252 if (DECL_BUILT_IN (new_decl
)
254 && !bitmap_empty_p (args_to_skip
))
256 DECL_BUILT_IN_CLASS (new_decl
) = NOT_BUILT_IN
;
257 DECL_FUNCTION_CODE (new_decl
) = (enum built_in_function
) 0;
259 /* The FE might have information and assumptions about the other
261 DECL_LANG_SPECIFIC (new_decl
) = NULL
;
265 /* Set flags of NEW_NODE and its decl. NEW_NODE is a newly created private
266 clone or its thunk. */
269 set_new_clone_decl_and_node_flags (cgraph_node
*new_node
)
271 DECL_EXTERNAL (new_node
->decl
) = 0;
272 TREE_PUBLIC (new_node
->decl
) = 0;
273 DECL_COMDAT (new_node
->decl
) = 0;
274 DECL_WEAK (new_node
->decl
) = 0;
275 DECL_VIRTUAL_P (new_node
->decl
) = 0;
276 DECL_STATIC_CONSTRUCTOR (new_node
->decl
) = 0;
277 DECL_STATIC_DESTRUCTOR (new_node
->decl
) = 0;
279 new_node
->externally_visible
= 0;
280 new_node
->local
.local
= 1;
281 new_node
->lowered
= true;
284 /* Duplicate thunk THUNK if necessary but make it to refer to NODE.
285 ARGS_TO_SKIP, if non-NULL, determines which parameters should be omitted.
286 Function can return NODE if no thunk is necessary, which can happen when
287 thunk is this_adjusting but we are removing this parameter. */
290 duplicate_thunk_for_node (cgraph_node
*thunk
, cgraph_node
*node
)
292 cgraph_node
*new_thunk
, *thunk_of
;
293 thunk_of
= thunk
->callees
->callee
->ultimate_alias_target ();
295 if (thunk_of
->thunk
.thunk_p
)
296 node
= duplicate_thunk_for_node (thunk_of
, node
);
298 if (!DECL_ARGUMENTS (thunk
->decl
))
302 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
303 if (cs
->caller
->thunk
.thunk_p
304 && cs
->caller
->thunk
.this_adjusting
== thunk
->thunk
.this_adjusting
305 && cs
->caller
->thunk
.fixed_offset
== thunk
->thunk
.fixed_offset
306 && cs
->caller
->thunk
.virtual_offset_p
== thunk
->thunk
.virtual_offset_p
307 && cs
->caller
->thunk
.virtual_value
== thunk
->thunk
.virtual_value
)
311 if (!node
->clone
.args_to_skip
)
312 new_decl
= copy_node (thunk
->decl
);
315 /* We do not need to duplicate this_adjusting thunks if we have removed
317 if (thunk
->thunk
.this_adjusting
318 && bitmap_bit_p (node
->clone
.args_to_skip
, 0))
321 new_decl
= build_function_decl_skip_args (thunk
->decl
,
322 node
->clone
.args_to_skip
,
326 tree
*link
= &DECL_ARGUMENTS (new_decl
);
328 for (tree pd
= DECL_ARGUMENTS (thunk
->decl
); pd
; pd
= DECL_CHAIN (pd
), i
++)
330 if (!node
->clone
.args_to_skip
331 || !bitmap_bit_p (node
->clone
.args_to_skip
, i
))
333 tree nd
= copy_node (pd
);
334 DECL_CONTEXT (nd
) = new_decl
;
336 link
= &DECL_CHAIN (nd
);
341 gcc_checking_assert (!DECL_STRUCT_FUNCTION (new_decl
));
342 gcc_checking_assert (!DECL_INITIAL (new_decl
));
343 gcc_checking_assert (!DECL_RESULT (new_decl
));
344 gcc_checking_assert (!DECL_RTL_SET_P (new_decl
));
346 DECL_NAME (new_decl
) = clone_function_name (thunk
->decl
, "artificial_thunk");
347 SET_DECL_ASSEMBLER_NAME (new_decl
, DECL_NAME (new_decl
));
349 new_thunk
= cgraph_node::create (new_decl
);
350 set_new_clone_decl_and_node_flags (new_thunk
);
351 new_thunk
->definition
= true;
352 new_thunk
->thunk
= thunk
->thunk
;
353 new_thunk
->unique_name
= in_lto_p
;
354 new_thunk
->former_clone_of
= thunk
->decl
;
355 new_thunk
->clone
.args_to_skip
= node
->clone
.args_to_skip
;
356 new_thunk
->clone
.combined_args_to_skip
= node
->clone
.combined_args_to_skip
;
358 cgraph_edge
*e
= new_thunk
->create_edge (node
, NULL
, 0,
360 e
->call_stmt_cannot_inline_p
= true;
361 symtab
->call_edge_duplication_hooks (thunk
->callees
, e
);
362 if (new_thunk
->expand_thunk (false, false))
364 new_thunk
->thunk
.thunk_p
= false;
365 new_thunk
->analyze ();
368 symtab
->call_cgraph_duplication_hooks (thunk
, new_thunk
);
372 /* If E does not lead to a thunk, simply redirect it to N. Otherwise create
373 one or more equivalent thunks for N and redirect E to the first in the
377 redirect_edge_duplicating_thunks (cgraph_edge
*e
, cgraph_node
*n
)
379 cgraph_node
*orig_to
= e
->callee
->ultimate_alias_target ();
380 if (orig_to
->thunk
.thunk_p
)
381 n
= duplicate_thunk_for_node (orig_to
, n
);
383 e
->redirect_callee (n
);
386 /* Create node representing clone of N executed COUNT times. Decrease
387 the execution counts from original node too.
388 The new clone will have decl set to DECL that may or may not be the same
391 When UPDATE_ORIGINAL is true, the counts are subtracted from the original
392 function's profile to reflect the fact that part of execution is handled
394 When CALL_DUPLICATOIN_HOOK is true, the ipa passes are acknowledged about
395 the new clone. Otherwise the caller is responsible for doing so later.
397 If the new node is being inlined into another one, NEW_INLINED_TO should be
398 the outline function the new one is (even indirectly) inlined to. All hooks
399 will see this in node's global.inlined_to, when invoked. Can be NULL if the
400 node is not inlined. */
403 cgraph_node::create_clone (tree decl
, gcov_type gcov_count
, int freq
,
404 bool update_original
,
405 vec
<cgraph_edge
*> redirect_callers
,
406 bool call_duplication_hook
,
407 cgraph_node
*new_inlined_to
,
410 cgraph_node
*new_node
= symtab
->create_empty ();
412 gcov_type count_scale
;
415 new_node
->decl
= decl
;
416 new_node
->register_symbol ();
417 new_node
->origin
= origin
;
418 new_node
->lto_file_data
= lto_file_data
;
419 if (new_node
->origin
)
421 new_node
->next_nested
= new_node
->origin
->nested
;
422 new_node
->origin
->nested
= new_node
;
424 new_node
->analyzed
= analyzed
;
425 new_node
->definition
= definition
;
426 new_node
->local
= local
;
427 new_node
->externally_visible
= false;
428 new_node
->no_reorder
= no_reorder
;
429 new_node
->local
.local
= true;
430 new_node
->global
= global
;
431 new_node
->global
.inlined_to
= new_inlined_to
;
433 new_node
->count
= count
;
434 new_node
->frequency
= frequency
;
435 new_node
->tp_first_run
= tp_first_run
;
436 new_node
->tm_clone
= tm_clone
;
438 new_node
->clone
.tree_map
= NULL
;
439 new_node
->clone
.args_to_skip
= args_to_skip
;
441 new_node
->clone
.combined_args_to_skip
= clone
.combined_args_to_skip
;
442 else if (clone
.combined_args_to_skip
)
444 new_node
->clone
.combined_args_to_skip
= BITMAP_GGC_ALLOC ();
445 bitmap_ior (new_node
->clone
.combined_args_to_skip
,
446 clone
.combined_args_to_skip
, args_to_skip
);
449 new_node
->clone
.combined_args_to_skip
= args_to_skip
;
453 if (new_node
->count
> count
)
454 count_scale
= REG_BR_PROB_BASE
;
456 count_scale
= GCOV_COMPUTE_SCALE (new_node
->count
, count
);
467 FOR_EACH_VEC_ELT (redirect_callers
, i
, e
)
469 /* Redirect calls to the old version node to point to its new
470 version. The only exception is when the edge was proved to
471 be unreachable during the clonning procedure. */
473 || DECL_BUILT_IN_CLASS (e
->callee
->decl
) != BUILT_IN_NORMAL
474 || DECL_FUNCTION_CODE (e
->callee
->decl
) != BUILT_IN_UNREACHABLE
)
475 redirect_edge_duplicating_thunks (e
, new_node
);
478 for (e
= callees
;e
; e
=e
->next_callee
)
479 e
->clone (new_node
, e
->call_stmt
, e
->lto_stmt_uid
, count_scale
,
480 freq
, update_original
);
482 for (e
= indirect_calls
; e
; e
= e
->next_callee
)
483 e
->clone (new_node
, e
->call_stmt
, e
->lto_stmt_uid
,
484 count_scale
, freq
, update_original
);
485 new_node
->clone_references (this);
487 new_node
->next_sibling_clone
= clones
;
489 clones
->prev_sibling_clone
= new_node
;
491 new_node
->clone_of
= this;
493 if (call_duplication_hook
)
494 symtab
->call_cgraph_duplication_hooks (this, new_node
);
498 /* Return a new assembler name for a clone of DECL with SUFFIX. */
500 static GTY(()) unsigned int clone_fn_id_num
;
503 clone_function_name (tree decl
, const char *suffix
)
505 tree name
= DECL_ASSEMBLER_NAME (decl
);
506 size_t len
= IDENTIFIER_LENGTH (name
);
507 char *tmp_name
, *prefix
;
509 prefix
= XALLOCAVEC (char, len
+ strlen (suffix
) + 2);
510 memcpy (prefix
, IDENTIFIER_POINTER (name
), len
);
511 strcpy (prefix
+ len
+ 1, suffix
);
512 #ifndef NO_DOT_IN_LABEL
514 #elif !defined NO_DOLLAR_IN_LABEL
519 ASM_FORMAT_PRIVATE_NAME (tmp_name
, prefix
, clone_fn_id_num
++);
520 return get_identifier (tmp_name
);
523 /* Create callgraph node clone with new declaration. The actual body will
524 be copied later at compilation stage.
526 TODO: after merging in ipa-sra use function call notes instead of args_to_skip
530 cgraph_node::create_virtual_clone (vec
<cgraph_edge
*> redirect_callers
,
531 vec
<ipa_replace_map
*, va_gc
> *tree_map
,
532 bitmap args_to_skip
, const char * suffix
)
534 tree old_decl
= decl
;
535 cgraph_node
*new_node
= NULL
;
538 ipa_replace_map
*map
;
542 gcc_checking_assert (tree_versionable_function_p (old_decl
));
544 gcc_assert (local
.can_change_signature
|| !args_to_skip
);
546 /* Make a new FUNCTION_DECL tree node */
548 new_decl
= copy_node (old_decl
);
550 new_decl
= build_function_decl_skip_args (old_decl
, args_to_skip
, false);
552 /* These pointers represent function body and will be populated only when clone
554 gcc_assert (new_decl
!= old_decl
);
555 DECL_STRUCT_FUNCTION (new_decl
) = NULL
;
556 DECL_ARGUMENTS (new_decl
) = NULL
;
557 DECL_INITIAL (new_decl
) = NULL
;
558 DECL_RESULT (new_decl
) = NULL
;
559 /* We can not do DECL_RESULT (new_decl) = NULL; here because of LTO partitioning
560 sometimes storing only clone decl instead of original. */
562 /* Generate a new name for the new version. */
563 len
= IDENTIFIER_LENGTH (DECL_NAME (old_decl
));
564 name
= XALLOCAVEC (char, len
+ strlen (suffix
) + 2);
565 memcpy (name
, IDENTIFIER_POINTER (DECL_NAME (old_decl
)), len
);
566 strcpy (name
+ len
+ 1, suffix
);
568 DECL_NAME (new_decl
) = get_identifier (name
);
569 SET_DECL_ASSEMBLER_NAME (new_decl
, clone_function_name (old_decl
, suffix
));
570 SET_DECL_RTL (new_decl
, NULL
);
572 new_node
= create_clone (new_decl
, count
, CGRAPH_FREQ_BASE
, false,
573 redirect_callers
, false, NULL
, args_to_skip
);
575 /* Update the properties.
576 Make clone visible only within this translation unit. Make sure
577 that is not weak also.
578 ??? We cannot use COMDAT linkage because there is no
579 ABI support for this. */
580 set_new_clone_decl_and_node_flags (new_node
);
581 new_node
->clone
.tree_map
= tree_map
;
583 /* Clones of global symbols or symbols with unique names are unique. */
584 if ((TREE_PUBLIC (old_decl
)
585 && !DECL_EXTERNAL (old_decl
)
586 && !DECL_WEAK (old_decl
)
587 && !DECL_COMDAT (old_decl
))
589 new_node
->unique_name
= true;
590 FOR_EACH_VEC_SAFE_ELT (tree_map
, i
, map
)
591 new_node
->maybe_create_reference (map
->new_tree
, IPA_REF_ADDR
, NULL
);
593 if (ipa_transforms_to_apply
.exists ())
594 new_node
->ipa_transforms_to_apply
595 = ipa_transforms_to_apply
.copy ();
597 symtab
->call_cgraph_duplication_hooks (this, new_node
);
602 /* callgraph node being removed from symbol table; see if its entry can be
603 replaced by other inline clone. */
605 cgraph_node::find_replacement (void)
607 cgraph_node
*next_inline_clone
, *replacement
;
609 for (next_inline_clone
= clones
;
611 && next_inline_clone
->decl
!= decl
;
612 next_inline_clone
= next_inline_clone
->next_sibling_clone
)
615 /* If there is inline clone of the node being removed, we need
616 to put it into the position of removed node and reorganize all
617 other clones to be based on it. */
618 if (next_inline_clone
)
621 cgraph_node
*new_clones
;
623 replacement
= next_inline_clone
;
625 /* Unlink inline clone from the list of clones of removed node. */
626 if (next_inline_clone
->next_sibling_clone
)
627 next_inline_clone
->next_sibling_clone
->prev_sibling_clone
628 = next_inline_clone
->prev_sibling_clone
;
629 if (next_inline_clone
->prev_sibling_clone
)
631 gcc_assert (clones
!= next_inline_clone
);
632 next_inline_clone
->prev_sibling_clone
->next_sibling_clone
633 = next_inline_clone
->next_sibling_clone
;
637 gcc_assert (clones
== next_inline_clone
);
638 clones
= next_inline_clone
->next_sibling_clone
;
644 /* Copy clone info. */
645 next_inline_clone
->clone
= clone
;
647 /* Now place it into clone tree at same level at NODE. */
648 next_inline_clone
->clone_of
= clone_of
;
649 next_inline_clone
->prev_sibling_clone
= NULL
;
650 next_inline_clone
->next_sibling_clone
= NULL
;
653 if (clone_of
->clones
)
654 clone_of
->clones
->prev_sibling_clone
= next_inline_clone
;
655 next_inline_clone
->next_sibling_clone
= clone_of
->clones
;
656 clone_of
->clones
= next_inline_clone
;
659 /* Merge the clone list. */
662 if (!next_inline_clone
->clones
)
663 next_inline_clone
->clones
= new_clones
;
666 n
= next_inline_clone
->clones
;
667 while (n
->next_sibling_clone
)
668 n
= n
->next_sibling_clone
;
669 n
->next_sibling_clone
= new_clones
;
670 new_clones
->prev_sibling_clone
= n
;
674 /* Update clone_of pointers. */
678 n
->clone_of
= next_inline_clone
;
679 n
= n
->next_sibling_clone
;
687 /* Like cgraph_set_call_stmt but walk the clone tree and update all
688 clones sharing the same function body.
689 When WHOLE_SPECULATIVE_EDGES is true, all three components of
690 speculative edge gets updated. Otherwise we update only direct
694 cgraph_node::set_call_stmt_including_clones (gimple old_stmt
, gimple new_stmt
,
695 bool update_speculative
)
698 cgraph_edge
*edge
= get_edge (old_stmt
);
701 edge
->set_call_stmt (new_stmt
, update_speculative
);
707 cgraph_edge
*edge
= node
->get_edge (old_stmt
);
710 edge
->set_call_stmt (new_stmt
, update_speculative
);
711 /* If UPDATE_SPECULATIVE is false, it means that we are turning
712 speculative call into a real code sequence. Update the
714 if (edge
->speculative
&& !update_speculative
)
716 cgraph_edge
*direct
, *indirect
;
719 gcc_assert (!edge
->indirect_unknown_callee
);
720 edge
->speculative_call_info (direct
, indirect
, ref
);
721 direct
->speculative
= false;
722 indirect
->speculative
= false;
723 ref
->speculative
= false;
728 else if (node
->next_sibling_clone
)
729 node
= node
->next_sibling_clone
;
732 while (node
!= this && !node
->next_sibling_clone
)
733 node
= node
->clone_of
;
735 node
= node
->next_sibling_clone
;
740 /* Like cgraph_create_edge walk the clone tree and update all clones sharing
741 same function body. If clones already have edge for OLD_STMT; only
742 update the edge same way as cgraph_set_call_stmt_including_clones does.
744 TODO: COUNT and LOOP_DEPTH should be properly distributed based on relative
745 frequencies of the clones. */
748 cgraph_node::create_edge_including_clones (cgraph_node
*callee
,
749 gimple old_stmt
, gimple stmt
,
752 cgraph_inline_failed_t reason
)
757 if (!get_edge (stmt
))
759 edge
= create_edge (callee
, stmt
, count
, freq
);
760 edge
->inline_failed
= reason
;
767 cgraph_edge
*edge
= node
->get_edge (old_stmt
);
769 /* It is possible that clones already contain the edge while
770 master didn't. Either we promoted indirect call into direct
771 call in the clone or we are processing clones of unreachable
772 master where edges has been removed. */
774 edge
->set_call_stmt (stmt
);
775 else if (! node
->get_edge (stmt
))
777 edge
= node
->create_edge (callee
, stmt
, count
, freq
);
778 edge
->inline_failed
= reason
;
783 else if (node
->next_sibling_clone
)
784 node
= node
->next_sibling_clone
;
787 while (node
!= this && !node
->next_sibling_clone
)
788 node
= node
->clone_of
;
790 node
= node
->next_sibling_clone
;
795 /* Remove the node from cgraph and all inline clones inlined into it.
796 Skip however removal of FORBIDDEN_NODE and return true if it needs to be
797 removed. This allows to call the function from outer loop walking clone
801 cgraph_node::remove_symbol_and_inline_clones (cgraph_node
*forbidden_node
)
803 cgraph_edge
*e
, *next
;
806 if (this == forbidden_node
)
811 for (e
= callees
; e
; e
= next
)
813 next
= e
->next_callee
;
814 if (!e
->inline_failed
)
815 found
|= e
->callee
->remove_symbol_and_inline_clones (forbidden_node
);
821 /* The edges representing the callers of the NEW_VERSION node were
822 fixed by cgraph_function_versioning (), now the call_expr in their
823 respective tree code should be updated to call the NEW_VERSION. */
826 update_call_expr (cgraph_node
*new_version
)
830 gcc_assert (new_version
);
832 /* Update the call expr on the edges to call the new version. */
833 for (e
= new_version
->callers
; e
; e
= e
->next_caller
)
835 function
*inner_function
= DECL_STRUCT_FUNCTION (e
->caller
->decl
);
836 gimple_call_set_fndecl (e
->call_stmt
, new_version
->decl
);
837 maybe_clean_eh_stmt_fn (inner_function
, e
->call_stmt
);
842 /* Create a new cgraph node which is the new version of
843 callgraph node. REDIRECT_CALLERS holds the callers
844 edges which should be redirected to point to
845 NEW_VERSION. ALL the callees edges of the node
846 are cloned to the new version node. Return the new
849 If non-NULL BLOCK_TO_COPY determine what basic blocks
850 was copied to prevent duplications of calls that are dead
854 cgraph_node::create_version_clone (tree new_decl
,
855 vec
<cgraph_edge
*> redirect_callers
,
858 cgraph_node
*new_version
;
862 new_version
= cgraph_node::create (new_decl
);
864 new_version
->analyzed
= analyzed
;
865 new_version
->definition
= definition
;
866 new_version
->local
= local
;
867 new_version
->externally_visible
= false;
868 new_version
->no_reorder
= no_reorder
;
869 new_version
->local
.local
= new_version
->definition
;
870 new_version
->global
= global
;
871 new_version
->rtl
= rtl
;
872 new_version
->count
= count
;
874 for (e
= callees
; e
; e
=e
->next_callee
)
876 || bitmap_bit_p (bbs_to_copy
, gimple_bb (e
->call_stmt
)->index
))
877 e
->clone (new_version
, e
->call_stmt
,
878 e
->lto_stmt_uid
, REG_BR_PROB_BASE
,
881 for (e
= indirect_calls
; e
; e
=e
->next_callee
)
883 || bitmap_bit_p (bbs_to_copy
, gimple_bb (e
->call_stmt
)->index
))
884 e
->clone (new_version
, e
->call_stmt
,
885 e
->lto_stmt_uid
, REG_BR_PROB_BASE
,
888 FOR_EACH_VEC_ELT (redirect_callers
, i
, e
)
890 /* Redirect calls to the old version node to point to its new
892 e
->redirect_callee (new_version
);
895 symtab
->call_cgraph_duplication_hooks (this, new_version
);
900 /* Perform function versioning.
901 Function versioning includes copying of the tree and
902 a callgraph update (creating a new cgraph node and updating
903 its callees and callers).
905 REDIRECT_CALLERS varray includes the edges to be redirected
908 TREE_MAP is a mapping of tree nodes we want to replace with
909 new ones (according to results of prior analysis).
911 If non-NULL ARGS_TO_SKIP determine function parameters to remove
913 If SKIP_RETURN is true, the new version will return void.
914 If non-NULL BLOCK_TO_COPY determine what basic blocks to copy.
915 If non_NULL NEW_ENTRY determine new entry BB of the clone.
917 Return the new version's cgraph node. */
920 cgraph_node::create_version_clone_with_body
921 (vec
<cgraph_edge
*> redirect_callers
,
922 vec
<ipa_replace_map
*, va_gc
> *tree_map
, bitmap args_to_skip
,
923 bool skip_return
, bitmap bbs_to_copy
, basic_block new_entry_block
,
924 const char *clone_name
)
926 tree old_decl
= decl
;
927 cgraph_node
*new_version_node
= NULL
;
930 if (!tree_versionable_function_p (old_decl
))
933 gcc_assert (local
.can_change_signature
|| !args_to_skip
);
935 /* Make a new FUNCTION_DECL tree node for the new version. */
936 if (!args_to_skip
&& !skip_return
)
937 new_decl
= copy_node (old_decl
);
940 = build_function_decl_skip_args (old_decl
, args_to_skip
, skip_return
);
942 /* Generate a new name for the new version. */
943 DECL_NAME (new_decl
) = clone_function_name (old_decl
, clone_name
);
944 SET_DECL_ASSEMBLER_NAME (new_decl
, DECL_NAME (new_decl
));
945 SET_DECL_RTL (new_decl
, NULL
);
947 /* When the old decl was a con-/destructor make sure the clone isn't. */
948 DECL_STATIC_CONSTRUCTOR (new_decl
) = 0;
949 DECL_STATIC_DESTRUCTOR (new_decl
) = 0;
951 /* Create the new version's call-graph node.
952 and update the edges of the new node. */
953 new_version_node
= create_version_clone (new_decl
, redirect_callers
,
956 if (ipa_transforms_to_apply
.exists ())
957 new_version_node
->ipa_transforms_to_apply
958 = ipa_transforms_to_apply
.copy ();
959 /* Copy the OLD_VERSION_NODE function tree to the new version. */
960 tree_function_versioning (old_decl
, new_decl
, tree_map
, false, args_to_skip
,
961 skip_return
, bbs_to_copy
, new_entry_block
);
963 /* Update the new version's properties.
964 Make The new version visible only within this translation unit. Make sure
965 that is not weak also.
966 ??? We cannot use COMDAT linkage because there is no
967 ABI support for this. */
968 new_version_node
->make_decl_local ();
969 DECL_VIRTUAL_P (new_version_node
->decl
) = 0;
970 new_version_node
->externally_visible
= 0;
971 new_version_node
->local
.local
= 1;
972 new_version_node
->lowered
= true;
973 /* Clones of global symbols or symbols with unique names are unique. */
974 if ((TREE_PUBLIC (old_decl
)
975 && !DECL_EXTERNAL (old_decl
)
976 && !DECL_WEAK (old_decl
)
977 && !DECL_COMDAT (old_decl
))
979 new_version_node
->unique_name
= true;
981 /* Update the call_expr on the edges to call the new version node. */
982 update_call_expr (new_version_node
);
984 symtab
->call_cgraph_insertion_hooks (this);
985 return new_version_node
;
988 /* Given virtual clone, turn it into actual clone. */
991 cgraph_materialize_clone (cgraph_node
*node
)
993 bitmap_obstack_initialize (NULL
);
994 node
->former_clone_of
= node
->clone_of
->decl
;
995 if (node
->clone_of
->former_clone_of
)
996 node
->former_clone_of
= node
->clone_of
->former_clone_of
;
997 /* Copy the OLD_VERSION_NODE function tree to the new version. */
998 tree_function_versioning (node
->clone_of
->decl
, node
->decl
,
999 node
->clone
.tree_map
, true,
1000 node
->clone
.args_to_skip
, false,
1002 if (symtab
->dump_file
)
1004 dump_function_to_file (node
->clone_of
->decl
, symtab
->dump_file
,
1006 dump_function_to_file (node
->decl
, symtab
->dump_file
, dump_flags
);
1009 /* Function is no longer clone. */
1010 if (node
->next_sibling_clone
)
1011 node
->next_sibling_clone
->prev_sibling_clone
= node
->prev_sibling_clone
;
1012 if (node
->prev_sibling_clone
)
1013 node
->prev_sibling_clone
->next_sibling_clone
= node
->next_sibling_clone
;
1015 node
->clone_of
->clones
= node
->next_sibling_clone
;
1016 node
->next_sibling_clone
= NULL
;
1017 node
->prev_sibling_clone
= NULL
;
1018 if (!node
->clone_of
->analyzed
&& !node
->clone_of
->clones
)
1020 node
->clone_of
->release_body ();
1021 node
->clone_of
->remove_callees ();
1022 node
->clone_of
->remove_all_references ();
1024 node
->clone_of
= NULL
;
1025 bitmap_obstack_release (NULL
);
1028 /* Once all functions from compilation unit are in memory, produce all clones
1029 and update all calls. We might also do this on demand if we don't want to
1030 bring all functions to memory prior compilation, but current WHOPR
1031 implementation does that and it is is bit easier to keep everything right in
1035 symbol_table::materialize_all_clones (void)
1038 bool stabilized
= false;
1041 if (symtab
->dump_file
)
1042 fprintf (symtab
->dump_file
, "Materializing clones\n");
1043 #ifdef ENABLE_CHECKING
1044 cgraph_node::verify_cgraph_nodes ();
1047 /* We can also do topological order, but number of iterations should be
1048 bounded by number of IPA passes since single IPA pass is probably not
1049 going to create clones of clones it created itself. */
1053 FOR_EACH_FUNCTION (node
)
1055 if (node
->clone_of
&& node
->decl
!= node
->clone_of
->decl
1056 && !gimple_has_body_p (node
->decl
))
1058 if (!node
->clone_of
->clone_of
)
1059 node
->clone_of
->get_body ();
1060 if (gimple_has_body_p (node
->clone_of
->decl
))
1062 if (symtab
->dump_file
)
1064 fprintf (symtab
->dump_file
, "cloning %s to %s\n",
1065 xstrdup (node
->clone_of
->name ()),
1066 xstrdup (node
->name ()));
1067 if (node
->clone
.tree_map
)
1070 fprintf (symtab
->dump_file
, " replace map: ");
1072 i
< vec_safe_length (node
->clone
.tree_map
);
1075 ipa_replace_map
*replace_info
;
1076 replace_info
= (*node
->clone
.tree_map
)[i
];
1077 print_generic_expr (symtab
->dump_file
, replace_info
->old_tree
, 0);
1078 fprintf (symtab
->dump_file
, " -> ");
1079 print_generic_expr (symtab
->dump_file
, replace_info
->new_tree
, 0);
1080 fprintf (symtab
->dump_file
, "%s%s;",
1081 replace_info
->replace_p
? "(replace)":"",
1082 replace_info
->ref_p
? "(ref)":"");
1084 fprintf (symtab
->dump_file
, "\n");
1086 if (node
->clone
.args_to_skip
)
1088 fprintf (symtab
->dump_file
, " args_to_skip: ");
1089 dump_bitmap (symtab
->dump_file
,
1090 node
->clone
.args_to_skip
);
1092 if (node
->clone
.args_to_skip
)
1094 fprintf (symtab
->dump_file
, " combined_args_to_skip:");
1095 dump_bitmap (symtab
->dump_file
, node
->clone
.combined_args_to_skip
);
1098 cgraph_materialize_clone (node
);
1104 FOR_EACH_FUNCTION (node
)
1105 if (!node
->analyzed
&& node
->callees
)
1107 node
->remove_callees ();
1108 node
->remove_all_references ();
1111 node
->clear_stmts_in_references ();
1112 if (symtab
->dump_file
)
1113 fprintf (symtab
->dump_file
, "Materialization Call site updates done.\n");
1114 #ifdef ENABLE_CHECKING
1115 cgraph_node::verify_cgraph_nodes ();
1117 symtab
->remove_unreachable_nodes (false, symtab
->dump_file
);
1120 #include "gt-cgraphclones.h"