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 E in the node N represented by CALL_EXPR the callgraph. */
107 cgraph_clone_edge (struct cgraph_edge
*e
, struct cgraph_node
*n
,
108 gimple call_stmt
, unsigned stmt_uid
, gcov_type count_scale
,
109 int freq_scale
, bool update_original
)
111 struct cgraph_edge
*new_edge
;
112 gcov_type count
= apply_probability (e
->count
, count_scale
);
115 /* We do not want to ignore loop nest after frequency drops to 0. */
118 freq
= e
->frequency
* (gcov_type
) freq_scale
/ CGRAPH_FREQ_BASE
;
119 if (freq
> CGRAPH_FREQ_MAX
)
120 freq
= CGRAPH_FREQ_MAX
;
122 if (e
->indirect_unknown_callee
)
126 if (call_stmt
&& (decl
= gimple_call_fndecl (call_stmt
))
127 /* When the call is speculative, we need to resolve it
128 via cgraph_resolve_speculation and not here. */
131 struct cgraph_node
*callee
= cgraph_get_node (decl
);
132 gcc_checking_assert (callee
);
133 new_edge
= cgraph_create_edge (n
, callee
, call_stmt
, count
, freq
);
137 new_edge
= cgraph_create_indirect_edge (n
, call_stmt
,
138 e
->indirect_info
->ecf_flags
,
140 *new_edge
->indirect_info
= *e
->indirect_info
;
145 new_edge
= cgraph_create_edge (n
, e
->callee
, call_stmt
, count
, freq
);
146 if (e
->indirect_info
)
148 new_edge
->indirect_info
149 = ggc_cleared_alloc
<cgraph_indirect_call_info
> ();
150 *new_edge
->indirect_info
= *e
->indirect_info
;
154 new_edge
->inline_failed
= e
->inline_failed
;
155 new_edge
->indirect_inlining_edge
= e
->indirect_inlining_edge
;
156 new_edge
->lto_stmt_uid
= stmt_uid
;
157 /* Clone flags that depend on call_stmt availability manually. */
158 new_edge
->can_throw_external
= e
->can_throw_external
;
159 new_edge
->call_stmt_cannot_inline_p
= e
->call_stmt_cannot_inline_p
;
160 new_edge
->speculative
= e
->speculative
;
163 e
->count
-= new_edge
->count
;
167 cgraph_call_edge_duplication_hooks (e
, new_edge
);
171 /* Build variant of function type ORIG_TYPE skipping ARGS_TO_SKIP and the
172 return value if SKIP_RETURN is true. */
175 build_function_type_skip_args (tree orig_type
, bitmap args_to_skip
,
178 tree new_type
= NULL
;
179 tree args
, new_args
= NULL
, t
;
183 for (args
= TYPE_ARG_TYPES (orig_type
); args
&& args
!= void_list_node
;
184 args
= TREE_CHAIN (args
), i
++)
185 if (!args_to_skip
|| !bitmap_bit_p (args_to_skip
, i
))
186 new_args
= tree_cons (NULL_TREE
, TREE_VALUE (args
), new_args
);
188 new_reversed
= nreverse (new_args
);
192 TREE_CHAIN (new_args
) = void_list_node
;
194 new_reversed
= void_list_node
;
197 /* Use copy_node to preserve as much as possible from original type
198 (debug info, attribute lists etc.)
199 Exception is METHOD_TYPEs must have THIS argument.
200 When we are asked to remove it, we need to build new FUNCTION_TYPE
202 if (TREE_CODE (orig_type
) != METHOD_TYPE
204 || !bitmap_bit_p (args_to_skip
, 0))
206 new_type
= build_distinct_type_copy (orig_type
);
207 TYPE_ARG_TYPES (new_type
) = new_reversed
;
212 = build_distinct_type_copy (build_function_type (TREE_TYPE (orig_type
),
214 TYPE_CONTEXT (new_type
) = TYPE_CONTEXT (orig_type
);
218 TREE_TYPE (new_type
) = void_type_node
;
220 /* This is a new type, not a copy of an old type. Need to reassociate
221 variants. We can handle everything except the main variant lazily. */
222 t
= TYPE_MAIN_VARIANT (orig_type
);
225 t
= build_function_type_skip_args (t
, args_to_skip
, skip_return
);
226 TYPE_MAIN_VARIANT (new_type
) = t
;
227 TYPE_NEXT_VARIANT (new_type
) = TYPE_NEXT_VARIANT (t
);
228 TYPE_NEXT_VARIANT (t
) = new_type
;
232 TYPE_MAIN_VARIANT (new_type
) = new_type
;
233 TYPE_NEXT_VARIANT (new_type
) = NULL
;
239 /* Build variant of function decl ORIG_DECL skipping ARGS_TO_SKIP and the
240 return value if SKIP_RETURN is true.
242 Arguments from DECL_ARGUMENTS list can't be removed now, since they are
243 linked by TREE_CHAIN directly. The caller is responsible for eliminating
244 them when they are being duplicated (i.e. copy_arguments_for_versioning). */
247 build_function_decl_skip_args (tree orig_decl
, bitmap args_to_skip
,
250 tree new_decl
= copy_node (orig_decl
);
253 new_type
= TREE_TYPE (orig_decl
);
254 if (prototype_p (new_type
)
255 || (skip_return
&& !VOID_TYPE_P (TREE_TYPE (new_type
))))
257 = build_function_type_skip_args (new_type
, args_to_skip
, skip_return
);
258 TREE_TYPE (new_decl
) = new_type
;
260 /* For declarations setting DECL_VINDEX (i.e. methods)
261 we expect first argument to be THIS pointer. */
262 if (args_to_skip
&& bitmap_bit_p (args_to_skip
, 0))
263 DECL_VINDEX (new_decl
) = NULL_TREE
;
265 /* When signature changes, we need to clear builtin info. */
266 if (DECL_BUILT_IN (new_decl
)
268 && !bitmap_empty_p (args_to_skip
))
270 DECL_BUILT_IN_CLASS (new_decl
) = NOT_BUILT_IN
;
271 DECL_FUNCTION_CODE (new_decl
) = (enum built_in_function
) 0;
273 /* The FE might have information and assumptions about the other
275 DECL_LANG_SPECIFIC (new_decl
) = NULL
;
279 /* Set flags of NEW_NODE and its decl. NEW_NODE is a newly created private
280 clone or its thunk. */
283 set_new_clone_decl_and_node_flags (cgraph_node
*new_node
)
285 DECL_EXTERNAL (new_node
->decl
) = 0;
286 TREE_PUBLIC (new_node
->decl
) = 0;
287 DECL_COMDAT (new_node
->decl
) = 0;
288 DECL_WEAK (new_node
->decl
) = 0;
289 DECL_VIRTUAL_P (new_node
->decl
) = 0;
290 DECL_STATIC_CONSTRUCTOR (new_node
->decl
) = 0;
291 DECL_STATIC_DESTRUCTOR (new_node
->decl
) = 0;
293 new_node
->externally_visible
= 0;
294 new_node
->local
.local
= 1;
295 new_node
->lowered
= true;
298 /* Duplicate thunk THUNK if necessary but make it to refer to NODE.
299 ARGS_TO_SKIP, if non-NULL, determines which parameters should be omitted.
300 Function can return NODE if no thunk is necessary, which can happen when
301 thunk is this_adjusting but we are removing this parameter. */
304 duplicate_thunk_for_node (cgraph_node
*thunk
, cgraph_node
*node
,
307 cgraph_node
*new_thunk
, *thunk_of
;
308 thunk_of
= cgraph_function_or_thunk_node (thunk
->callees
->callee
);
310 if (thunk_of
->thunk
.thunk_p
)
311 node
= duplicate_thunk_for_node (thunk_of
, node
, args_to_skip
);
313 struct cgraph_edge
*cs
;
314 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
315 if (cs
->caller
->thunk
.thunk_p
316 && cs
->caller
->thunk
.this_adjusting
== thunk
->thunk
.this_adjusting
317 && cs
->caller
->thunk
.fixed_offset
== thunk
->thunk
.fixed_offset
318 && cs
->caller
->thunk
.virtual_offset_p
== thunk
->thunk
.virtual_offset_p
319 && cs
->caller
->thunk
.virtual_value
== thunk
->thunk
.virtual_value
)
324 new_decl
= copy_node (thunk
->decl
);
327 /* We do not need to duplicate this_adjusting thunks if we have removed
329 if (thunk
->thunk
.this_adjusting
330 && bitmap_bit_p (args_to_skip
, 0))
333 new_decl
= build_function_decl_skip_args (thunk
->decl
, args_to_skip
,
336 gcc_checking_assert (!DECL_STRUCT_FUNCTION (new_decl
));
337 gcc_checking_assert (!DECL_INITIAL (new_decl
));
338 gcc_checking_assert (!DECL_RESULT (new_decl
));
339 gcc_checking_assert (!DECL_RTL_SET_P (new_decl
));
341 DECL_NAME (new_decl
) = clone_function_name (thunk
->decl
, "artificial_thunk");
342 SET_DECL_ASSEMBLER_NAME (new_decl
, DECL_NAME (new_decl
));
344 new_thunk
= cgraph_create_node (new_decl
);
345 set_new_clone_decl_and_node_flags (new_thunk
);
346 new_thunk
->definition
= true;
347 new_thunk
->thunk
= thunk
->thunk
;
348 new_thunk
->unique_name
= in_lto_p
;
349 new_thunk
->former_clone_of
= thunk
->decl
;
351 struct cgraph_edge
*e
= cgraph_create_edge (new_thunk
, node
, NULL
, 0,
353 e
->call_stmt_cannot_inline_p
= true;
354 cgraph_call_edge_duplication_hooks (thunk
->callees
, e
);
355 if (!expand_thunk (new_thunk
, false, false))
356 new_thunk
->analyzed
= true;
357 cgraph_call_node_duplication_hooks (thunk
, new_thunk
);
361 /* If E does not lead to a thunk, simply redirect it to N. Otherwise create
362 one or more equivalent thunks for N and redirect E to the first in the
366 redirect_edge_duplicating_thunks (struct cgraph_edge
*e
, struct cgraph_node
*n
,
369 cgraph_node
*orig_to
= cgraph_function_or_thunk_node (e
->callee
);
370 if (orig_to
->thunk
.thunk_p
)
371 n
= duplicate_thunk_for_node (orig_to
, n
, args_to_skip
);
373 cgraph_redirect_edge_callee (e
, n
);
376 /* Create node representing clone of N executed COUNT times. Decrease
377 the execution counts from original node too.
378 The new clone will have decl set to DECL that may or may not be the same
381 When UPDATE_ORIGINAL is true, the counts are subtracted from the original
382 function's profile to reflect the fact that part of execution is handled
384 When CALL_DUPLICATOIN_HOOK is true, the ipa passes are acknowledged about
385 the new clone. Otherwise the caller is responsible for doing so later.
387 If the new node is being inlined into another one, NEW_INLINED_TO should be
388 the outline function the new one is (even indirectly) inlined to. All hooks
389 will see this in node's global.inlined_to, when invoked. Can be NULL if the
390 node is not inlined. */
393 cgraph_clone_node (struct cgraph_node
*n
, tree decl
, gcov_type count
, int freq
,
394 bool update_original
,
395 vec
<cgraph_edge_p
> redirect_callers
,
396 bool call_duplication_hook
,
397 struct cgraph_node
*new_inlined_to
,
400 struct cgraph_node
*new_node
= cgraph_create_empty_node ();
401 struct cgraph_edge
*e
;
402 gcov_type count_scale
;
405 new_node
->decl
= decl
;
406 symtab_register_node (new_node
);
407 new_node
->origin
= n
->origin
;
408 new_node
->lto_file_data
= n
->lto_file_data
;
409 if (new_node
->origin
)
411 new_node
->next_nested
= new_node
->origin
->nested
;
412 new_node
->origin
->nested
= new_node
;
414 new_node
->analyzed
= n
->analyzed
;
415 new_node
->definition
= n
->definition
;
416 new_node
->local
= n
->local
;
417 new_node
->externally_visible
= false;
418 new_node
->local
.local
= true;
419 new_node
->global
= n
->global
;
420 new_node
->global
.inlined_to
= new_inlined_to
;
421 new_node
->rtl
= n
->rtl
;
422 new_node
->count
= count
;
423 new_node
->frequency
= n
->frequency
;
424 new_node
->clone
= n
->clone
;
425 new_node
->clone
.tree_map
= NULL
;
426 new_node
->tp_first_run
= n
->tp_first_run
;
429 if (new_node
->count
> n
->count
)
430 count_scale
= REG_BR_PROB_BASE
;
432 count_scale
= GCOV_COMPUTE_SCALE (new_node
->count
, n
->count
);
443 FOR_EACH_VEC_ELT (redirect_callers
, i
, e
)
445 /* Redirect calls to the old version node to point to its new
446 version. The only exception is when the edge was proved to
447 be unreachable during the clonning procedure. */
449 || DECL_BUILT_IN_CLASS (e
->callee
->decl
) != BUILT_IN_NORMAL
450 || DECL_FUNCTION_CODE (e
->callee
->decl
) != BUILT_IN_UNREACHABLE
)
451 redirect_edge_duplicating_thunks (e
, new_node
, args_to_skip
);
455 for (e
= n
->callees
;e
; e
=e
->next_callee
)
456 cgraph_clone_edge (e
, new_node
, e
->call_stmt
, e
->lto_stmt_uid
,
457 count_scale
, freq
, update_original
);
459 for (e
= n
->indirect_calls
; e
; e
= e
->next_callee
)
460 cgraph_clone_edge (e
, new_node
, e
->call_stmt
, e
->lto_stmt_uid
,
461 count_scale
, freq
, update_original
);
462 ipa_clone_references (new_node
, &n
->ref_list
);
464 new_node
->next_sibling_clone
= n
->clones
;
466 n
->clones
->prev_sibling_clone
= new_node
;
467 n
->clones
= new_node
;
468 new_node
->clone_of
= n
;
470 if (call_duplication_hook
)
471 cgraph_call_node_duplication_hooks (n
, new_node
);
475 /* Return a new assembler name for a clone of DECL with SUFFIX. */
477 static GTY(()) unsigned int clone_fn_id_num
;
480 clone_function_name (tree decl
, const char *suffix
)
482 tree name
= DECL_ASSEMBLER_NAME (decl
);
483 size_t len
= IDENTIFIER_LENGTH (name
);
484 char *tmp_name
, *prefix
;
486 prefix
= XALLOCAVEC (char, len
+ strlen (suffix
) + 2);
487 memcpy (prefix
, IDENTIFIER_POINTER (name
), len
);
488 strcpy (prefix
+ len
+ 1, suffix
);
489 #ifndef NO_DOT_IN_LABEL
491 #elif !defined NO_DOLLAR_IN_LABEL
496 ASM_FORMAT_PRIVATE_NAME (tmp_name
, prefix
, clone_fn_id_num
++);
497 return get_identifier (tmp_name
);
500 /* Create callgraph node clone with new declaration. The actual body will
501 be copied later at compilation stage.
503 TODO: after merging in ipa-sra use function call notes instead of args_to_skip
507 cgraph_create_virtual_clone (struct cgraph_node
*old_node
,
508 vec
<cgraph_edge_p
> redirect_callers
,
509 vec
<ipa_replace_map_p
, va_gc
> *tree_map
,
513 tree old_decl
= old_node
->decl
;
514 struct cgraph_node
*new_node
= NULL
;
517 struct ipa_replace_map
*map
;
521 gcc_checking_assert (tree_versionable_function_p (old_decl
));
523 gcc_assert (old_node
->local
.can_change_signature
|| !args_to_skip
);
525 /* Make a new FUNCTION_DECL tree node */
527 new_decl
= copy_node (old_decl
);
529 new_decl
= build_function_decl_skip_args (old_decl
, args_to_skip
, false);
531 /* These pointers represent function body and will be populated only when clone
533 gcc_assert (new_decl
!= old_decl
);
534 DECL_STRUCT_FUNCTION (new_decl
) = NULL
;
535 DECL_ARGUMENTS (new_decl
) = NULL
;
536 DECL_INITIAL (new_decl
) = NULL
;
537 DECL_RESULT (new_decl
) = NULL
;
538 /* We can not do DECL_RESULT (new_decl) = NULL; here because of LTO partitioning
539 sometimes storing only clone decl instead of original. */
541 /* Generate a new name for the new version. */
542 len
= IDENTIFIER_LENGTH (DECL_NAME (old_decl
));
543 name
= XALLOCAVEC (char, len
+ strlen (suffix
) + 2);
544 memcpy (name
, IDENTIFIER_POINTER (DECL_NAME (old_decl
)), len
);
545 strcpy (name
+ len
+ 1, suffix
);
547 DECL_NAME (new_decl
) = get_identifier (name
);
548 SET_DECL_ASSEMBLER_NAME (new_decl
, clone_function_name (old_decl
, suffix
));
549 SET_DECL_RTL (new_decl
, NULL
);
551 new_node
= cgraph_clone_node (old_node
, new_decl
, old_node
->count
,
552 CGRAPH_FREQ_BASE
, false,
553 redirect_callers
, false, NULL
, args_to_skip
);
554 /* Update the properties.
555 Make clone visible only within this translation unit. Make sure
556 that is not weak also.
557 ??? We cannot use COMDAT linkage because there is no
558 ABI support for this. */
559 set_new_clone_decl_and_node_flags (new_node
);
560 new_node
->clone
.tree_map
= tree_map
;
561 new_node
->clone
.args_to_skip
= args_to_skip
;
563 /* Clones of global symbols or symbols with unique names are unique. */
564 if ((TREE_PUBLIC (old_decl
)
565 && !DECL_EXTERNAL (old_decl
)
566 && !DECL_WEAK (old_decl
)
567 && !DECL_COMDAT (old_decl
))
569 new_node
->unique_name
= true;
570 FOR_EACH_VEC_SAFE_ELT (tree_map
, i
, map
)
571 ipa_maybe_record_reference (new_node
, map
->new_tree
,
574 new_node
->clone
.combined_args_to_skip
= old_node
->clone
.combined_args_to_skip
;
575 else if (old_node
->clone
.combined_args_to_skip
)
577 int newi
= 0, oldi
= 0;
579 bitmap new_args_to_skip
= BITMAP_GGC_ALLOC ();
580 struct cgraph_node
*orig_node
;
581 for (orig_node
= old_node
; orig_node
->clone_of
; orig_node
= orig_node
->clone_of
)
583 for (arg
= DECL_ARGUMENTS (orig_node
->decl
);
584 arg
; arg
= DECL_CHAIN (arg
), oldi
++)
586 if (bitmap_bit_p (old_node
->clone
.combined_args_to_skip
, oldi
))
588 bitmap_set_bit (new_args_to_skip
, oldi
);
591 if (bitmap_bit_p (args_to_skip
, newi
))
592 bitmap_set_bit (new_args_to_skip
, oldi
);
595 new_node
->clone
.combined_args_to_skip
= new_args_to_skip
;
598 new_node
->clone
.combined_args_to_skip
= args_to_skip
;
599 if (old_node
->ipa_transforms_to_apply
.exists ())
600 new_node
->ipa_transforms_to_apply
601 = old_node
->ipa_transforms_to_apply
.copy ();
603 cgraph_call_node_duplication_hooks (old_node
, new_node
);
609 /* NODE is being removed from symbol table; see if its entry can be replaced by
610 other inline clone. */
612 cgraph_find_replacement_node (struct cgraph_node
*node
)
614 struct cgraph_node
*next_inline_clone
, *replacement
;
616 for (next_inline_clone
= node
->clones
;
618 && next_inline_clone
->decl
!= node
->decl
;
619 next_inline_clone
= next_inline_clone
->next_sibling_clone
)
622 /* If there is inline clone of the node being removed, we need
623 to put it into the position of removed node and reorganize all
624 other clones to be based on it. */
625 if (next_inline_clone
)
627 struct cgraph_node
*n
;
628 struct cgraph_node
*new_clones
;
630 replacement
= next_inline_clone
;
632 /* Unlink inline clone from the list of clones of removed node. */
633 if (next_inline_clone
->next_sibling_clone
)
634 next_inline_clone
->next_sibling_clone
->prev_sibling_clone
635 = next_inline_clone
->prev_sibling_clone
;
636 if (next_inline_clone
->prev_sibling_clone
)
638 gcc_assert (node
->clones
!= next_inline_clone
);
639 next_inline_clone
->prev_sibling_clone
->next_sibling_clone
640 = next_inline_clone
->next_sibling_clone
;
644 gcc_assert (node
->clones
== next_inline_clone
);
645 node
->clones
= next_inline_clone
->next_sibling_clone
;
648 new_clones
= node
->clones
;
651 /* Copy clone info. */
652 next_inline_clone
->clone
= node
->clone
;
654 /* Now place it into clone tree at same level at NODE. */
655 next_inline_clone
->clone_of
= node
->clone_of
;
656 next_inline_clone
->prev_sibling_clone
= NULL
;
657 next_inline_clone
->next_sibling_clone
= NULL
;
660 if (node
->clone_of
->clones
)
661 node
->clone_of
->clones
->prev_sibling_clone
= next_inline_clone
;
662 next_inline_clone
->next_sibling_clone
= node
->clone_of
->clones
;
663 node
->clone_of
->clones
= next_inline_clone
;
666 /* Merge the clone list. */
669 if (!next_inline_clone
->clones
)
670 next_inline_clone
->clones
= new_clones
;
673 n
= next_inline_clone
->clones
;
674 while (n
->next_sibling_clone
)
675 n
= n
->next_sibling_clone
;
676 n
->next_sibling_clone
= new_clones
;
677 new_clones
->prev_sibling_clone
= n
;
681 /* Update clone_of pointers. */
685 n
->clone_of
= next_inline_clone
;
686 n
= n
->next_sibling_clone
;
694 /* Like cgraph_set_call_stmt but walk the clone tree and update all
695 clones sharing the same function body.
696 When WHOLE_SPECULATIVE_EDGES is true, all three components of
697 speculative edge gets updated. Otherwise we update only direct
701 cgraph_set_call_stmt_including_clones (struct cgraph_node
*orig
,
702 gimple old_stmt
, gimple new_stmt
,
703 bool update_speculative
)
705 struct cgraph_node
*node
;
706 struct cgraph_edge
*edge
= cgraph_edge (orig
, old_stmt
);
709 cgraph_set_call_stmt (edge
, new_stmt
, update_speculative
);
715 struct cgraph_edge
*edge
= cgraph_edge (node
, old_stmt
);
718 cgraph_set_call_stmt (edge
, new_stmt
, update_speculative
);
719 /* If UPDATE_SPECULATIVE is false, it means that we are turning
720 speculative call into a real code sequence. Update the
722 if (edge
->speculative
&& !update_speculative
)
724 struct cgraph_edge
*direct
, *indirect
;
727 gcc_assert (!edge
->indirect_unknown_callee
);
728 cgraph_speculative_call_info (edge
, direct
, indirect
, ref
);
729 direct
->speculative
= false;
730 indirect
->speculative
= false;
731 ref
->speculative
= false;
736 else if (node
->next_sibling_clone
)
737 node
= node
->next_sibling_clone
;
740 while (node
!= orig
&& !node
->next_sibling_clone
)
741 node
= node
->clone_of
;
743 node
= node
->next_sibling_clone
;
748 /* Like cgraph_create_edge walk the clone tree and update all clones sharing
749 same function body. If clones already have edge for OLD_STMT; only
750 update the edge same way as cgraph_set_call_stmt_including_clones does.
752 TODO: COUNT and LOOP_DEPTH should be properly distributed based on relative
753 frequencies of the clones. */
756 cgraph_create_edge_including_clones (struct cgraph_node
*orig
,
757 struct cgraph_node
*callee
,
759 gimple stmt
, gcov_type count
,
761 cgraph_inline_failed_t reason
)
763 struct cgraph_node
*node
;
764 struct cgraph_edge
*edge
;
766 if (!cgraph_edge (orig
, stmt
))
768 edge
= cgraph_create_edge (orig
, callee
, stmt
, count
, freq
);
769 edge
->inline_failed
= reason
;
776 struct cgraph_edge
*edge
= cgraph_edge (node
, old_stmt
);
778 /* It is possible that clones already contain the edge while
779 master didn't. Either we promoted indirect call into direct
780 call in the clone or we are processing clones of unreachable
781 master where edges has been removed. */
783 cgraph_set_call_stmt (edge
, stmt
);
784 else if (!cgraph_edge (node
, stmt
))
786 edge
= cgraph_create_edge (node
, callee
, stmt
, count
,
788 edge
->inline_failed
= reason
;
793 else if (node
->next_sibling_clone
)
794 node
= node
->next_sibling_clone
;
797 while (node
!= orig
&& !node
->next_sibling_clone
)
798 node
= node
->clone_of
;
800 node
= node
->next_sibling_clone
;
805 /* Remove the node from cgraph and all inline clones inlined into it.
806 Skip however removal of FORBIDDEN_NODE and return true if it needs to be
807 removed. This allows to call the function from outer loop walking clone
811 cgraph_remove_node_and_inline_clones (struct cgraph_node
*node
, struct cgraph_node
*forbidden_node
)
813 struct cgraph_edge
*e
, *next
;
816 if (node
== forbidden_node
)
818 cgraph_remove_edge (node
->callers
);
821 for (e
= node
->callees
; e
; e
= next
)
823 next
= e
->next_callee
;
824 if (!e
->inline_failed
)
825 found
|= cgraph_remove_node_and_inline_clones (e
->callee
, forbidden_node
);
827 cgraph_remove_node (node
);
831 /* The edges representing the callers of the NEW_VERSION node were
832 fixed by cgraph_function_versioning (), now the call_expr in their
833 respective tree code should be updated to call the NEW_VERSION. */
836 update_call_expr (struct cgraph_node
*new_version
)
838 struct cgraph_edge
*e
;
840 gcc_assert (new_version
);
842 /* Update the call expr on the edges to call the new version. */
843 for (e
= new_version
->callers
; e
; e
= e
->next_caller
)
845 struct function
*inner_function
= DECL_STRUCT_FUNCTION (e
->caller
->decl
);
846 gimple_call_set_fndecl (e
->call_stmt
, new_version
->decl
);
847 maybe_clean_eh_stmt_fn (inner_function
, e
->call_stmt
);
852 /* Create a new cgraph node which is the new version of
853 OLD_VERSION node. REDIRECT_CALLERS holds the callers
854 edges which should be redirected to point to
855 NEW_VERSION. ALL the callees edges of OLD_VERSION
856 are cloned to the new version node. Return the new
859 If non-NULL BLOCK_TO_COPY determine what basic blocks
860 was copied to prevent duplications of calls that are dead
864 cgraph_copy_node_for_versioning (struct cgraph_node
*old_version
,
866 vec
<cgraph_edge_p
> redirect_callers
,
869 struct cgraph_node
*new_version
;
870 struct cgraph_edge
*e
;
873 gcc_assert (old_version
);
875 new_version
= cgraph_create_node (new_decl
);
877 new_version
->analyzed
= old_version
->analyzed
;
878 new_version
->definition
= old_version
->definition
;
879 new_version
->local
= old_version
->local
;
880 new_version
->externally_visible
= false;
881 new_version
->local
.local
= new_version
->definition
;
882 new_version
->global
= old_version
->global
;
883 new_version
->rtl
= old_version
->rtl
;
884 new_version
->count
= old_version
->count
;
886 for (e
= old_version
->callees
; e
; e
=e
->next_callee
)
888 || bitmap_bit_p (bbs_to_copy
, gimple_bb (e
->call_stmt
)->index
))
889 cgraph_clone_edge (e
, new_version
, e
->call_stmt
,
890 e
->lto_stmt_uid
, REG_BR_PROB_BASE
,
893 for (e
= old_version
->indirect_calls
; e
; e
=e
->next_callee
)
895 || bitmap_bit_p (bbs_to_copy
, gimple_bb (e
->call_stmt
)->index
))
896 cgraph_clone_edge (e
, new_version
, e
->call_stmt
,
897 e
->lto_stmt_uid
, REG_BR_PROB_BASE
,
900 FOR_EACH_VEC_ELT (redirect_callers
, i
, e
)
902 /* Redirect calls to the old version node to point to its new
904 cgraph_redirect_edge_callee (e
, new_version
);
907 cgraph_call_node_duplication_hooks (old_version
, new_version
);
912 /* Perform function versioning.
913 Function versioning includes copying of the tree and
914 a callgraph update (creating a new cgraph node and updating
915 its callees and callers).
917 REDIRECT_CALLERS varray includes the edges to be redirected
920 TREE_MAP is a mapping of tree nodes we want to replace with
921 new ones (according to results of prior analysis).
922 OLD_VERSION_NODE is the node that is versioned.
924 If non-NULL ARGS_TO_SKIP determine function parameters to remove
926 If SKIP_RETURN is true, the new version will return void.
927 If non-NULL BLOCK_TO_COPY determine what basic blocks to copy.
928 If non_NULL NEW_ENTRY determine new entry BB of the clone.
930 Return the new version's cgraph node. */
933 cgraph_function_versioning (struct cgraph_node
*old_version_node
,
934 vec
<cgraph_edge_p
> redirect_callers
,
935 vec
<ipa_replace_map_p
, va_gc
> *tree_map
,
939 basic_block new_entry_block
,
940 const char *clone_name
)
942 tree old_decl
= old_version_node
->decl
;
943 struct cgraph_node
*new_version_node
= NULL
;
946 if (!tree_versionable_function_p (old_decl
))
949 gcc_assert (old_version_node
->local
.can_change_signature
|| !args_to_skip
);
951 /* Make a new FUNCTION_DECL tree node for the new version. */
952 if (!args_to_skip
&& !skip_return
)
953 new_decl
= copy_node (old_decl
);
956 = build_function_decl_skip_args (old_decl
, args_to_skip
, skip_return
);
958 /* Generate a new name for the new version. */
959 DECL_NAME (new_decl
) = clone_function_name (old_decl
, clone_name
);
960 SET_DECL_ASSEMBLER_NAME (new_decl
, DECL_NAME (new_decl
));
961 SET_DECL_RTL (new_decl
, NULL
);
963 /* When the old decl was a con-/destructor make sure the clone isn't. */
964 DECL_STATIC_CONSTRUCTOR (new_decl
) = 0;
965 DECL_STATIC_DESTRUCTOR (new_decl
) = 0;
967 /* Create the new version's call-graph node.
968 and update the edges of the new node. */
970 cgraph_copy_node_for_versioning (old_version_node
, new_decl
,
971 redirect_callers
, bbs_to_copy
);
973 if (old_version_node
->ipa_transforms_to_apply
.exists ())
974 new_version_node
->ipa_transforms_to_apply
975 = old_version_node
->ipa_transforms_to_apply
.copy ();
976 /* Copy the OLD_VERSION_NODE function tree to the new version. */
977 tree_function_versioning (old_decl
, new_decl
, tree_map
, false, args_to_skip
,
978 skip_return
, bbs_to_copy
, new_entry_block
);
980 /* Update the new version's properties.
981 Make The new version visible only within this translation unit. Make sure
982 that is not weak also.
983 ??? We cannot use COMDAT linkage because there is no
984 ABI support for this. */
985 symtab_make_decl_local (new_version_node
->decl
);
986 DECL_VIRTUAL_P (new_version_node
->decl
) = 0;
987 new_version_node
->externally_visible
= 0;
988 new_version_node
->local
.local
= 1;
989 new_version_node
->lowered
= true;
990 /* Clones of global symbols or symbols with unique names are unique. */
991 if ((TREE_PUBLIC (old_decl
)
992 && !DECL_EXTERNAL (old_decl
)
993 && !DECL_WEAK (old_decl
)
994 && !DECL_COMDAT (old_decl
))
996 new_version_node
->unique_name
= true;
998 /* Update the call_expr on the edges to call the new version node. */
999 update_call_expr (new_version_node
);
1001 cgraph_call_function_insertion_hooks (new_version_node
);
1002 return new_version_node
;
1005 /* Given virtual clone, turn it into actual clone. */
1008 cgraph_materialize_clone (struct cgraph_node
*node
)
1010 bitmap_obstack_initialize (NULL
);
1011 node
->former_clone_of
= node
->clone_of
->decl
;
1012 if (node
->clone_of
->former_clone_of
)
1013 node
->former_clone_of
= node
->clone_of
->former_clone_of
;
1014 /* Copy the OLD_VERSION_NODE function tree to the new version. */
1015 tree_function_versioning (node
->clone_of
->decl
, node
->decl
,
1016 node
->clone
.tree_map
, true,
1017 node
->clone
.args_to_skip
, false,
1019 if (cgraph_dump_file
)
1021 dump_function_to_file (node
->clone_of
->decl
, cgraph_dump_file
, dump_flags
);
1022 dump_function_to_file (node
->decl
, cgraph_dump_file
, dump_flags
);
1025 /* Function is no longer clone. */
1026 if (node
->next_sibling_clone
)
1027 node
->next_sibling_clone
->prev_sibling_clone
= node
->prev_sibling_clone
;
1028 if (node
->prev_sibling_clone
)
1029 node
->prev_sibling_clone
->next_sibling_clone
= node
->next_sibling_clone
;
1031 node
->clone_of
->clones
= node
->next_sibling_clone
;
1032 node
->next_sibling_clone
= NULL
;
1033 node
->prev_sibling_clone
= NULL
;
1034 if (!node
->clone_of
->analyzed
&& !node
->clone_of
->clones
)
1036 cgraph_release_function_body (node
->clone_of
);
1037 cgraph_node_remove_callees (node
->clone_of
);
1038 ipa_remove_all_references (&node
->clone_of
->ref_list
);
1040 node
->clone_of
= NULL
;
1041 bitmap_obstack_release (NULL
);
1044 /* Once all functions from compilation unit are in memory, produce all clones
1045 and update all calls. We might also do this on demand if we don't want to
1046 bring all functions to memory prior compilation, but current WHOPR
1047 implementation does that and it is is bit easier to keep everything right in
1051 cgraph_materialize_all_clones (void)
1053 struct cgraph_node
*node
;
1054 bool stabilized
= false;
1057 if (cgraph_dump_file
)
1058 fprintf (cgraph_dump_file
, "Materializing clones\n");
1059 #ifdef ENABLE_CHECKING
1063 /* We can also do topological order, but number of iterations should be
1064 bounded by number of IPA passes since single IPA pass is probably not
1065 going to create clones of clones it created itself. */
1069 FOR_EACH_FUNCTION (node
)
1071 if (node
->clone_of
&& node
->decl
!= node
->clone_of
->decl
1072 && !gimple_has_body_p (node
->decl
))
1074 if (!node
->clone_of
->clone_of
)
1075 cgraph_get_body (node
->clone_of
);
1076 if (gimple_has_body_p (node
->clone_of
->decl
))
1078 if (cgraph_dump_file
)
1080 fprintf (cgraph_dump_file
, "cloning %s to %s\n",
1081 xstrdup (node
->clone_of
->name ()),
1082 xstrdup (node
->name ()));
1083 if (node
->clone
.tree_map
)
1086 fprintf (cgraph_dump_file
, " replace map: ");
1088 i
< vec_safe_length (node
->clone
.tree_map
);
1091 struct ipa_replace_map
*replace_info
;
1092 replace_info
= (*node
->clone
.tree_map
)[i
];
1093 print_generic_expr (cgraph_dump_file
, replace_info
->old_tree
, 0);
1094 fprintf (cgraph_dump_file
, " -> ");
1095 print_generic_expr (cgraph_dump_file
, replace_info
->new_tree
, 0);
1096 fprintf (cgraph_dump_file
, "%s%s;",
1097 replace_info
->replace_p
? "(replace)":"",
1098 replace_info
->ref_p
? "(ref)":"");
1100 fprintf (cgraph_dump_file
, "\n");
1102 if (node
->clone
.args_to_skip
)
1104 fprintf (cgraph_dump_file
, " args_to_skip: ");
1105 dump_bitmap (cgraph_dump_file
, node
->clone
.args_to_skip
);
1107 if (node
->clone
.args_to_skip
)
1109 fprintf (cgraph_dump_file
, " combined_args_to_skip:");
1110 dump_bitmap (cgraph_dump_file
, node
->clone
.combined_args_to_skip
);
1113 cgraph_materialize_clone (node
);
1119 FOR_EACH_FUNCTION (node
)
1120 if (!node
->analyzed
&& node
->callees
)
1122 cgraph_node_remove_callees (node
);
1123 ipa_remove_all_references (&node
->ref_list
);
1126 ipa_clear_stmts_in_references (node
);
1127 if (cgraph_dump_file
)
1128 fprintf (cgraph_dump_file
, "Materialization Call site updates done.\n");
1129 #ifdef ENABLE_CHECKING
1132 symtab_remove_unreachable_nodes (false, cgraph_dump_file
);
1135 #include "gt-cgraphclones.h"