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
;
164 count
-= new_edge
->count
;
168 symtab
->call_edge_duplication_hooks (this, new_edge
);
172 /* Build variant of function type ORIG_TYPE skipping ARGS_TO_SKIP and the
173 return value if SKIP_RETURN is true. */
176 build_function_type_skip_args (tree orig_type
, bitmap args_to_skip
,
179 tree new_type
= NULL
;
180 tree args
, new_args
= NULL
, t
;
184 for (args
= TYPE_ARG_TYPES (orig_type
); args
&& args
!= void_list_node
;
185 args
= TREE_CHAIN (args
), i
++)
186 if (!args_to_skip
|| !bitmap_bit_p (args_to_skip
, i
))
187 new_args
= tree_cons (NULL_TREE
, TREE_VALUE (args
), new_args
);
189 new_reversed
= nreverse (new_args
);
193 TREE_CHAIN (new_args
) = void_list_node
;
195 new_reversed
= void_list_node
;
198 /* Use copy_node to preserve as much as possible from original type
199 (debug info, attribute lists etc.)
200 Exception is METHOD_TYPEs must have THIS argument.
201 When we are asked to remove it, we need to build new FUNCTION_TYPE
203 if (TREE_CODE (orig_type
) != METHOD_TYPE
205 || !bitmap_bit_p (args_to_skip
, 0))
207 new_type
= build_distinct_type_copy (orig_type
);
208 TYPE_ARG_TYPES (new_type
) = new_reversed
;
213 = build_distinct_type_copy (build_function_type (TREE_TYPE (orig_type
),
215 TYPE_CONTEXT (new_type
) = TYPE_CONTEXT (orig_type
);
219 TREE_TYPE (new_type
) = void_type_node
;
221 /* This is a new type, not a copy of an old type. Need to reassociate
222 variants. We can handle everything except the main variant lazily. */
223 t
= TYPE_MAIN_VARIANT (orig_type
);
226 t
= build_function_type_skip_args (t
, args_to_skip
, skip_return
);
227 TYPE_MAIN_VARIANT (new_type
) = t
;
228 TYPE_NEXT_VARIANT (new_type
) = TYPE_NEXT_VARIANT (t
);
229 TYPE_NEXT_VARIANT (t
) = new_type
;
233 TYPE_MAIN_VARIANT (new_type
) = new_type
;
234 TYPE_NEXT_VARIANT (new_type
) = NULL
;
240 /* Build variant of function decl ORIG_DECL skipping ARGS_TO_SKIP and the
241 return value if SKIP_RETURN is true.
243 Arguments from DECL_ARGUMENTS list can't be removed now, since they are
244 linked by TREE_CHAIN directly. The caller is responsible for eliminating
245 them when they are being duplicated (i.e. copy_arguments_for_versioning). */
248 build_function_decl_skip_args (tree orig_decl
, bitmap args_to_skip
,
251 tree new_decl
= copy_node (orig_decl
);
254 new_type
= TREE_TYPE (orig_decl
);
255 if (prototype_p (new_type
)
256 || (skip_return
&& !VOID_TYPE_P (TREE_TYPE (new_type
))))
258 = build_function_type_skip_args (new_type
, args_to_skip
, skip_return
);
259 TREE_TYPE (new_decl
) = new_type
;
261 /* For declarations setting DECL_VINDEX (i.e. methods)
262 we expect first argument to be THIS pointer. */
263 if (args_to_skip
&& bitmap_bit_p (args_to_skip
, 0))
264 DECL_VINDEX (new_decl
) = NULL_TREE
;
266 /* When signature changes, we need to clear builtin info. */
267 if (DECL_BUILT_IN (new_decl
)
269 && !bitmap_empty_p (args_to_skip
))
271 DECL_BUILT_IN_CLASS (new_decl
) = NOT_BUILT_IN
;
272 DECL_FUNCTION_CODE (new_decl
) = (enum built_in_function
) 0;
274 /* The FE might have information and assumptions about the other
276 DECL_LANG_SPECIFIC (new_decl
) = NULL
;
280 /* Set flags of NEW_NODE and its decl. NEW_NODE is a newly created private
281 clone or its thunk. */
284 set_new_clone_decl_and_node_flags (cgraph_node
*new_node
)
286 DECL_EXTERNAL (new_node
->decl
) = 0;
287 TREE_PUBLIC (new_node
->decl
) = 0;
288 DECL_COMDAT (new_node
->decl
) = 0;
289 DECL_WEAK (new_node
->decl
) = 0;
290 DECL_VIRTUAL_P (new_node
->decl
) = 0;
291 DECL_STATIC_CONSTRUCTOR (new_node
->decl
) = 0;
292 DECL_STATIC_DESTRUCTOR (new_node
->decl
) = 0;
294 new_node
->externally_visible
= 0;
295 new_node
->local
.local
= 1;
296 new_node
->lowered
= true;
299 /* Duplicate thunk THUNK if necessary but make it to refer to NODE.
300 ARGS_TO_SKIP, if non-NULL, determines which parameters should be omitted.
301 Function can return NODE if no thunk is necessary, which can happen when
302 thunk is this_adjusting but we are removing this parameter. */
305 duplicate_thunk_for_node (cgraph_node
*thunk
, cgraph_node
*node
)
307 cgraph_node
*new_thunk
, *thunk_of
;
308 thunk_of
= thunk
->callees
->callee
->ultimate_alias_target ();
310 if (thunk_of
->thunk
.thunk_p
)
311 node
= duplicate_thunk_for_node (thunk_of
, node
);
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
)
323 if (!node
->clone
.args_to_skip
)
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 (node
->clone
.args_to_skip
, 0))
333 new_decl
= build_function_decl_skip_args (thunk
->decl
,
334 node
->clone
.args_to_skip
,
338 tree
*link
= &DECL_ARGUMENTS (new_decl
);
340 for (tree pd
= DECL_ARGUMENTS (thunk
->decl
); pd
; pd
= DECL_CHAIN (pd
), i
++)
342 if (!node
->clone
.args_to_skip
343 || !bitmap_bit_p (node
->clone
.args_to_skip
, i
))
345 tree nd
= copy_node (pd
);
346 DECL_CONTEXT (nd
) = new_decl
;
348 link
= &DECL_CHAIN (nd
);
353 gcc_checking_assert (!DECL_STRUCT_FUNCTION (new_decl
));
354 gcc_checking_assert (!DECL_INITIAL (new_decl
));
355 gcc_checking_assert (!DECL_RESULT (new_decl
));
356 gcc_checking_assert (!DECL_RTL_SET_P (new_decl
));
358 DECL_NAME (new_decl
) = clone_function_name (thunk
->decl
, "artificial_thunk");
359 SET_DECL_ASSEMBLER_NAME (new_decl
, DECL_NAME (new_decl
));
361 new_thunk
= cgraph_node::create (new_decl
);
362 set_new_clone_decl_and_node_flags (new_thunk
);
363 new_thunk
->definition
= true;
364 new_thunk
->thunk
= thunk
->thunk
;
365 new_thunk
->unique_name
= in_lto_p
;
366 new_thunk
->former_clone_of
= thunk
->decl
;
367 new_thunk
->clone
.args_to_skip
= node
->clone
.args_to_skip
;
368 new_thunk
->clone
.combined_args_to_skip
= node
->clone
.combined_args_to_skip
;
370 cgraph_edge
*e
= new_thunk
->create_edge (node
, NULL
, 0,
372 e
->call_stmt_cannot_inline_p
= true;
373 symtab
->call_edge_duplication_hooks (thunk
->callees
, e
);
374 if (new_thunk
->expand_thunk (false, false))
376 new_thunk
->thunk
.thunk_p
= false;
377 new_thunk
->analyze ();
380 symtab
->call_cgraph_duplication_hooks (thunk
, new_thunk
);
384 /* If E does not lead to a thunk, simply redirect it to N. Otherwise create
385 one or more equivalent thunks for N and redirect E to the first in the
389 redirect_edge_duplicating_thunks (cgraph_edge
*e
, cgraph_node
*n
)
391 cgraph_node
*orig_to
= e
->callee
->ultimate_alias_target ();
392 if (orig_to
->thunk
.thunk_p
)
393 n
= duplicate_thunk_for_node (orig_to
, n
);
395 e
->redirect_callee (n
);
398 /* Create node representing clone of N executed COUNT times. Decrease
399 the execution counts from original node too.
400 The new clone will have decl set to DECL that may or may not be the same
403 When UPDATE_ORIGINAL is true, the counts are subtracted from the original
404 function's profile to reflect the fact that part of execution is handled
406 When CALL_DUPLICATOIN_HOOK is true, the ipa passes are acknowledged about
407 the new clone. Otherwise the caller is responsible for doing so later.
409 If the new node is being inlined into another one, NEW_INLINED_TO should be
410 the outline function the new one is (even indirectly) inlined to. All hooks
411 will see this in node's global.inlined_to, when invoked. Can be NULL if the
412 node is not inlined. */
415 cgraph_node::create_clone (tree decl
, gcov_type gcov_count
, int freq
,
416 bool update_original
,
417 vec
<cgraph_edge
*> redirect_callers
,
418 bool call_duplication_hook
,
419 cgraph_node
*new_inlined_to
,
422 cgraph_node
*new_node
= symtab
->create_empty ();
424 gcov_type count_scale
;
427 new_node
->decl
= decl
;
428 new_node
->register_symbol ();
429 new_node
->origin
= origin
;
430 new_node
->lto_file_data
= lto_file_data
;
431 if (new_node
->origin
)
433 new_node
->next_nested
= new_node
->origin
->nested
;
434 new_node
->origin
->nested
= new_node
;
436 new_node
->analyzed
= analyzed
;
437 new_node
->definition
= definition
;
438 new_node
->local
= local
;
439 new_node
->externally_visible
= false;
440 new_node
->local
.local
= true;
441 new_node
->global
= global
;
442 new_node
->global
.inlined_to
= new_inlined_to
;
444 new_node
->count
= count
;
445 new_node
->frequency
= frequency
;
446 new_node
->tp_first_run
= tp_first_run
;
447 new_node
->tm_clone
= tm_clone
;
449 new_node
->clone
.tree_map
= NULL
;
450 new_node
->clone
.args_to_skip
= args_to_skip
;
452 new_node
->clone
.combined_args_to_skip
= clone
.combined_args_to_skip
;
453 else if (clone
.combined_args_to_skip
)
455 new_node
->clone
.combined_args_to_skip
= BITMAP_GGC_ALLOC ();
456 bitmap_ior (new_node
->clone
.combined_args_to_skip
,
457 clone
.combined_args_to_skip
, args_to_skip
);
460 new_node
->clone
.combined_args_to_skip
= args_to_skip
;
464 if (new_node
->count
> count
)
465 count_scale
= REG_BR_PROB_BASE
;
467 count_scale
= GCOV_COMPUTE_SCALE (new_node
->count
, count
);
478 FOR_EACH_VEC_ELT (redirect_callers
, i
, e
)
480 /* Redirect calls to the old version node to point to its new
481 version. The only exception is when the edge was proved to
482 be unreachable during the clonning procedure. */
484 || DECL_BUILT_IN_CLASS (e
->callee
->decl
) != BUILT_IN_NORMAL
485 || DECL_FUNCTION_CODE (e
->callee
->decl
) != BUILT_IN_UNREACHABLE
)
486 redirect_edge_duplicating_thunks (e
, new_node
);
489 for (e
= callees
;e
; e
=e
->next_callee
)
490 e
->clone (new_node
, e
->call_stmt
, e
->lto_stmt_uid
, count_scale
,
491 freq
, update_original
);
493 for (e
= indirect_calls
; e
; e
= e
->next_callee
)
494 e
->clone (new_node
, e
->call_stmt
, e
->lto_stmt_uid
,
495 count_scale
, freq
, update_original
);
496 new_node
->clone_references (this);
498 new_node
->next_sibling_clone
= clones
;
500 clones
->prev_sibling_clone
= new_node
;
502 new_node
->clone_of
= this;
504 if (call_duplication_hook
)
505 symtab
->call_cgraph_duplication_hooks (this, new_node
);
509 /* Return a new assembler name for a clone of DECL with SUFFIX. */
511 static GTY(()) unsigned int clone_fn_id_num
;
514 clone_function_name (tree decl
, const char *suffix
)
516 tree name
= DECL_ASSEMBLER_NAME (decl
);
517 size_t len
= IDENTIFIER_LENGTH (name
);
518 char *tmp_name
, *prefix
;
520 prefix
= XALLOCAVEC (char, len
+ strlen (suffix
) + 2);
521 memcpy (prefix
, IDENTIFIER_POINTER (name
), len
);
522 strcpy (prefix
+ len
+ 1, suffix
);
523 #ifndef NO_DOT_IN_LABEL
525 #elif !defined NO_DOLLAR_IN_LABEL
530 ASM_FORMAT_PRIVATE_NAME (tmp_name
, prefix
, clone_fn_id_num
++);
531 return get_identifier (tmp_name
);
534 /* Create callgraph node clone with new declaration. The actual body will
535 be copied later at compilation stage.
537 TODO: after merging in ipa-sra use function call notes instead of args_to_skip
541 cgraph_node::create_virtual_clone (vec
<cgraph_edge
*> redirect_callers
,
542 vec
<ipa_replace_map
*, va_gc
> *tree_map
,
543 bitmap args_to_skip
, const char * suffix
)
545 tree old_decl
= decl
;
546 cgraph_node
*new_node
= NULL
;
549 ipa_replace_map
*map
;
553 gcc_checking_assert (tree_versionable_function_p (old_decl
));
555 gcc_assert (local
.can_change_signature
|| !args_to_skip
);
557 /* Make a new FUNCTION_DECL tree node */
559 new_decl
= copy_node (old_decl
);
561 new_decl
= build_function_decl_skip_args (old_decl
, args_to_skip
, false);
563 /* These pointers represent function body and will be populated only when clone
565 gcc_assert (new_decl
!= old_decl
);
566 DECL_STRUCT_FUNCTION (new_decl
) = NULL
;
567 DECL_ARGUMENTS (new_decl
) = NULL
;
568 DECL_INITIAL (new_decl
) = NULL
;
569 DECL_RESULT (new_decl
) = NULL
;
570 /* We can not do DECL_RESULT (new_decl) = NULL; here because of LTO partitioning
571 sometimes storing only clone decl instead of original. */
573 /* Generate a new name for the new version. */
574 len
= IDENTIFIER_LENGTH (DECL_NAME (old_decl
));
575 name
= XALLOCAVEC (char, len
+ strlen (suffix
) + 2);
576 memcpy (name
, IDENTIFIER_POINTER (DECL_NAME (old_decl
)), len
);
577 strcpy (name
+ len
+ 1, suffix
);
579 DECL_NAME (new_decl
) = get_identifier (name
);
580 SET_DECL_ASSEMBLER_NAME (new_decl
, clone_function_name (old_decl
, suffix
));
581 SET_DECL_RTL (new_decl
, NULL
);
583 new_node
= create_clone (new_decl
, count
, CGRAPH_FREQ_BASE
, false,
584 redirect_callers
, false, NULL
, args_to_skip
);
586 /* Update the properties.
587 Make clone visible only within this translation unit. Make sure
588 that is not weak also.
589 ??? We cannot use COMDAT linkage because there is no
590 ABI support for this. */
591 set_new_clone_decl_and_node_flags (new_node
);
592 new_node
->clone
.tree_map
= tree_map
;
594 /* Clones of global symbols or symbols with unique names are unique. */
595 if ((TREE_PUBLIC (old_decl
)
596 && !DECL_EXTERNAL (old_decl
)
597 && !DECL_WEAK (old_decl
)
598 && !DECL_COMDAT (old_decl
))
600 new_node
->unique_name
= true;
601 FOR_EACH_VEC_SAFE_ELT (tree_map
, i
, map
)
602 new_node
->maybe_create_reference (map
->new_tree
, IPA_REF_ADDR
, NULL
);
604 if (ipa_transforms_to_apply
.exists ())
605 new_node
->ipa_transforms_to_apply
606 = ipa_transforms_to_apply
.copy ();
608 symtab
->call_cgraph_duplication_hooks (this, new_node
);
613 /* callgraph node being removed from symbol table; see if its entry can be
614 replaced by other inline clone. */
616 cgraph_node::find_replacement (void)
618 cgraph_node
*next_inline_clone
, *replacement
;
620 for (next_inline_clone
= clones
;
622 && next_inline_clone
->decl
!= decl
;
623 next_inline_clone
= next_inline_clone
->next_sibling_clone
)
626 /* If there is inline clone of the node being removed, we need
627 to put it into the position of removed node and reorganize all
628 other clones to be based on it. */
629 if (next_inline_clone
)
632 cgraph_node
*new_clones
;
634 replacement
= next_inline_clone
;
636 /* Unlink inline clone from the list of clones of removed node. */
637 if (next_inline_clone
->next_sibling_clone
)
638 next_inline_clone
->next_sibling_clone
->prev_sibling_clone
639 = next_inline_clone
->prev_sibling_clone
;
640 if (next_inline_clone
->prev_sibling_clone
)
642 gcc_assert (clones
!= next_inline_clone
);
643 next_inline_clone
->prev_sibling_clone
->next_sibling_clone
644 = next_inline_clone
->next_sibling_clone
;
648 gcc_assert (clones
== next_inline_clone
);
649 clones
= next_inline_clone
->next_sibling_clone
;
655 /* Copy clone info. */
656 next_inline_clone
->clone
= clone
;
658 /* Now place it into clone tree at same level at NODE. */
659 next_inline_clone
->clone_of
= clone_of
;
660 next_inline_clone
->prev_sibling_clone
= NULL
;
661 next_inline_clone
->next_sibling_clone
= NULL
;
664 if (clone_of
->clones
)
665 clone_of
->clones
->prev_sibling_clone
= next_inline_clone
;
666 next_inline_clone
->next_sibling_clone
= clone_of
->clones
;
667 clone_of
->clones
= next_inline_clone
;
670 /* Merge the clone list. */
673 if (!next_inline_clone
->clones
)
674 next_inline_clone
->clones
= new_clones
;
677 n
= next_inline_clone
->clones
;
678 while (n
->next_sibling_clone
)
679 n
= n
->next_sibling_clone
;
680 n
->next_sibling_clone
= new_clones
;
681 new_clones
->prev_sibling_clone
= n
;
685 /* Update clone_of pointers. */
689 n
->clone_of
= next_inline_clone
;
690 n
= n
->next_sibling_clone
;
698 /* Like cgraph_set_call_stmt but walk the clone tree and update all
699 clones sharing the same function body.
700 When WHOLE_SPECULATIVE_EDGES is true, all three components of
701 speculative edge gets updated. Otherwise we update only direct
705 cgraph_node::set_call_stmt_including_clones (gimple old_stmt
, gimple new_stmt
,
706 bool update_speculative
)
709 cgraph_edge
*edge
= get_edge (old_stmt
);
712 edge
->set_call_stmt (new_stmt
, update_speculative
);
718 cgraph_edge
*edge
= node
->get_edge (old_stmt
);
721 edge
->set_call_stmt (new_stmt
, update_speculative
);
722 /* If UPDATE_SPECULATIVE is false, it means that we are turning
723 speculative call into a real code sequence. Update the
725 if (edge
->speculative
&& !update_speculative
)
727 cgraph_edge
*direct
, *indirect
;
730 gcc_assert (!edge
->indirect_unknown_callee
);
731 edge
->speculative_call_info (direct
, indirect
, ref
);
732 direct
->speculative
= false;
733 indirect
->speculative
= false;
734 ref
->speculative
= false;
739 else if (node
->next_sibling_clone
)
740 node
= node
->next_sibling_clone
;
743 while (node
!= this && !node
->next_sibling_clone
)
744 node
= node
->clone_of
;
746 node
= node
->next_sibling_clone
;
751 /* Like cgraph_create_edge walk the clone tree and update all clones sharing
752 same function body. If clones already have edge for OLD_STMT; only
753 update the edge same way as cgraph_set_call_stmt_including_clones does.
755 TODO: COUNT and LOOP_DEPTH should be properly distributed based on relative
756 frequencies of the clones. */
759 cgraph_node::create_edge_including_clones (cgraph_node
*callee
,
760 gimple old_stmt
, gimple stmt
,
763 cgraph_inline_failed_t reason
)
768 if (!get_edge (stmt
))
770 edge
= create_edge (callee
, stmt
, count
, freq
);
771 edge
->inline_failed
= reason
;
778 cgraph_edge
*edge
= node
->get_edge (old_stmt
);
780 /* It is possible that clones already contain the edge while
781 master didn't. Either we promoted indirect call into direct
782 call in the clone or we are processing clones of unreachable
783 master where edges has been removed. */
785 edge
->set_call_stmt (stmt
);
786 else if (! node
->get_edge (stmt
))
788 edge
= node
->create_edge (callee
, stmt
, count
, freq
);
789 edge
->inline_failed
= reason
;
794 else if (node
->next_sibling_clone
)
795 node
= node
->next_sibling_clone
;
798 while (node
!= this && !node
->next_sibling_clone
)
799 node
= node
->clone_of
;
801 node
= node
->next_sibling_clone
;
806 /* Remove the node from cgraph and all inline clones inlined into it.
807 Skip however removal of FORBIDDEN_NODE and return true if it needs to be
808 removed. This allows to call the function from outer loop walking clone
812 cgraph_node::remove_symbol_and_inline_clones (cgraph_node
*forbidden_node
)
814 cgraph_edge
*e
, *next
;
817 if (this == forbidden_node
)
822 for (e
= callees
; e
; e
= next
)
824 next
= e
->next_callee
;
825 if (!e
->inline_failed
)
826 found
|= e
->callee
->remove_symbol_and_inline_clones (forbidden_node
);
832 /* The edges representing the callers of the NEW_VERSION node were
833 fixed by cgraph_function_versioning (), now the call_expr in their
834 respective tree code should be updated to call the NEW_VERSION. */
837 update_call_expr (cgraph_node
*new_version
)
841 gcc_assert (new_version
);
843 /* Update the call expr on the edges to call the new version. */
844 for (e
= new_version
->callers
; e
; e
= e
->next_caller
)
846 function
*inner_function
= DECL_STRUCT_FUNCTION (e
->caller
->decl
);
847 gimple_call_set_fndecl (e
->call_stmt
, new_version
->decl
);
848 maybe_clean_eh_stmt_fn (inner_function
, e
->call_stmt
);
853 /* Create a new cgraph node which is the new version of
854 callgraph node. REDIRECT_CALLERS holds the callers
855 edges which should be redirected to point to
856 NEW_VERSION. ALL the callees edges of the node
857 are cloned to the new version node. Return the new
860 If non-NULL BLOCK_TO_COPY determine what basic blocks
861 was copied to prevent duplications of calls that are dead
865 cgraph_node::create_version_clone (tree new_decl
,
866 vec
<cgraph_edge
*> redirect_callers
,
869 cgraph_node
*new_version
;
873 new_version
= cgraph_node::create (new_decl
);
875 new_version
->analyzed
= analyzed
;
876 new_version
->definition
= definition
;
877 new_version
->local
= local
;
878 new_version
->externally_visible
= false;
879 new_version
->local
.local
= new_version
->definition
;
880 new_version
->global
= global
;
881 new_version
->rtl
= rtl
;
882 new_version
->count
= count
;
884 for (e
= callees
; e
; e
=e
->next_callee
)
886 || bitmap_bit_p (bbs_to_copy
, gimple_bb (e
->call_stmt
)->index
))
887 e
->clone (new_version
, e
->call_stmt
,
888 e
->lto_stmt_uid
, REG_BR_PROB_BASE
,
891 for (e
= indirect_calls
; e
; e
=e
->next_callee
)
893 || bitmap_bit_p (bbs_to_copy
, gimple_bb (e
->call_stmt
)->index
))
894 e
->clone (new_version
, e
->call_stmt
,
895 e
->lto_stmt_uid
, REG_BR_PROB_BASE
,
898 FOR_EACH_VEC_ELT (redirect_callers
, i
, e
)
900 /* Redirect calls to the old version node to point to its new
902 e
->redirect_callee (new_version
);
905 symtab
->call_cgraph_duplication_hooks (this, new_version
);
910 /* Perform function versioning.
911 Function versioning includes copying of the tree and
912 a callgraph update (creating a new cgraph node and updating
913 its callees and callers).
915 REDIRECT_CALLERS varray includes the edges to be redirected
918 TREE_MAP is a mapping of tree nodes we want to replace with
919 new ones (according to results of prior analysis).
921 If non-NULL ARGS_TO_SKIP determine function parameters to remove
923 If SKIP_RETURN is true, the new version will return void.
924 If non-NULL BLOCK_TO_COPY determine what basic blocks to copy.
925 If non_NULL NEW_ENTRY determine new entry BB of the clone.
927 Return the new version's cgraph node. */
930 cgraph_node::create_version_clone_with_body
931 (vec
<cgraph_edge
*> redirect_callers
,
932 vec
<ipa_replace_map
*, va_gc
> *tree_map
, bitmap args_to_skip
,
933 bool skip_return
, bitmap bbs_to_copy
, basic_block new_entry_block
,
934 const char *clone_name
)
936 tree old_decl
= decl
;
937 cgraph_node
*new_version_node
= NULL
;
940 if (!tree_versionable_function_p (old_decl
))
943 gcc_assert (local
.can_change_signature
|| !args_to_skip
);
945 /* Make a new FUNCTION_DECL tree node for the new version. */
946 if (!args_to_skip
&& !skip_return
)
947 new_decl
= copy_node (old_decl
);
950 = build_function_decl_skip_args (old_decl
, args_to_skip
, skip_return
);
952 /* Generate a new name for the new version. */
953 DECL_NAME (new_decl
) = clone_function_name (old_decl
, clone_name
);
954 SET_DECL_ASSEMBLER_NAME (new_decl
, DECL_NAME (new_decl
));
955 SET_DECL_RTL (new_decl
, NULL
);
957 /* When the old decl was a con-/destructor make sure the clone isn't. */
958 DECL_STATIC_CONSTRUCTOR (new_decl
) = 0;
959 DECL_STATIC_DESTRUCTOR (new_decl
) = 0;
961 /* Create the new version's call-graph node.
962 and update the edges of the new node. */
963 new_version_node
= create_version_clone (new_decl
, redirect_callers
,
966 if (ipa_transforms_to_apply
.exists ())
967 new_version_node
->ipa_transforms_to_apply
968 = ipa_transforms_to_apply
.copy ();
969 /* Copy the OLD_VERSION_NODE function tree to the new version. */
970 tree_function_versioning (old_decl
, new_decl
, tree_map
, false, args_to_skip
,
971 skip_return
, bbs_to_copy
, new_entry_block
);
973 /* Update the new version's properties.
974 Make The new version visible only within this translation unit. Make sure
975 that is not weak also.
976 ??? We cannot use COMDAT linkage because there is no
977 ABI support for this. */
978 new_version_node
->make_decl_local ();
979 DECL_VIRTUAL_P (new_version_node
->decl
) = 0;
980 new_version_node
->externally_visible
= 0;
981 new_version_node
->local
.local
= 1;
982 new_version_node
->lowered
= true;
983 /* Clones of global symbols or symbols with unique names are unique. */
984 if ((TREE_PUBLIC (old_decl
)
985 && !DECL_EXTERNAL (old_decl
)
986 && !DECL_WEAK (old_decl
)
987 && !DECL_COMDAT (old_decl
))
989 new_version_node
->unique_name
= true;
991 /* Update the call_expr on the edges to call the new version node. */
992 update_call_expr (new_version_node
);
994 symtab
->call_cgraph_insertion_hooks (this);
995 return new_version_node
;
998 /* Given virtual clone, turn it into actual clone. */
1001 cgraph_materialize_clone (cgraph_node
*node
)
1003 bitmap_obstack_initialize (NULL
);
1004 node
->former_clone_of
= node
->clone_of
->decl
;
1005 if (node
->clone_of
->former_clone_of
)
1006 node
->former_clone_of
= node
->clone_of
->former_clone_of
;
1007 /* Copy the OLD_VERSION_NODE function tree to the new version. */
1008 tree_function_versioning (node
->clone_of
->decl
, node
->decl
,
1009 node
->clone
.tree_map
, true,
1010 node
->clone
.args_to_skip
, false,
1012 if (symtab
->dump_file
)
1014 dump_function_to_file (node
->clone_of
->decl
, symtab
->dump_file
,
1016 dump_function_to_file (node
->decl
, symtab
->dump_file
, dump_flags
);
1019 /* Function is no longer clone. */
1020 if (node
->next_sibling_clone
)
1021 node
->next_sibling_clone
->prev_sibling_clone
= node
->prev_sibling_clone
;
1022 if (node
->prev_sibling_clone
)
1023 node
->prev_sibling_clone
->next_sibling_clone
= node
->next_sibling_clone
;
1025 node
->clone_of
->clones
= node
->next_sibling_clone
;
1026 node
->next_sibling_clone
= NULL
;
1027 node
->prev_sibling_clone
= NULL
;
1028 if (!node
->clone_of
->analyzed
&& !node
->clone_of
->clones
)
1030 node
->clone_of
->release_body ();
1031 node
->clone_of
->remove_callees ();
1032 node
->clone_of
->remove_all_references ();
1034 node
->clone_of
= NULL
;
1035 bitmap_obstack_release (NULL
);
1038 /* Once all functions from compilation unit are in memory, produce all clones
1039 and update all calls. We might also do this on demand if we don't want to
1040 bring all functions to memory prior compilation, but current WHOPR
1041 implementation does that and it is is bit easier to keep everything right in
1045 symbol_table::materialize_all_clones (void)
1048 bool stabilized
= false;
1051 if (symtab
->dump_file
)
1052 fprintf (symtab
->dump_file
, "Materializing clones\n");
1053 #ifdef ENABLE_CHECKING
1054 cgraph_node::verify_cgraph_nodes ();
1057 /* We can also do topological order, but number of iterations should be
1058 bounded by number of IPA passes since single IPA pass is probably not
1059 going to create clones of clones it created itself. */
1063 FOR_EACH_FUNCTION (node
)
1065 if (node
->clone_of
&& node
->decl
!= node
->clone_of
->decl
1066 && !gimple_has_body_p (node
->decl
))
1068 if (!node
->clone_of
->clone_of
)
1069 node
->clone_of
->get_body ();
1070 if (gimple_has_body_p (node
->clone_of
->decl
))
1072 if (symtab
->dump_file
)
1074 fprintf (symtab
->dump_file
, "cloning %s to %s\n",
1075 xstrdup (node
->clone_of
->name ()),
1076 xstrdup (node
->name ()));
1077 if (node
->clone
.tree_map
)
1080 fprintf (symtab
->dump_file
, " replace map: ");
1082 i
< vec_safe_length (node
->clone
.tree_map
);
1085 ipa_replace_map
*replace_info
;
1086 replace_info
= (*node
->clone
.tree_map
)[i
];
1087 print_generic_expr (symtab
->dump_file
, replace_info
->old_tree
, 0);
1088 fprintf (symtab
->dump_file
, " -> ");
1089 print_generic_expr (symtab
->dump_file
, replace_info
->new_tree
, 0);
1090 fprintf (symtab
->dump_file
, "%s%s;",
1091 replace_info
->replace_p
? "(replace)":"",
1092 replace_info
->ref_p
? "(ref)":"");
1094 fprintf (symtab
->dump_file
, "\n");
1096 if (node
->clone
.args_to_skip
)
1098 fprintf (symtab
->dump_file
, " args_to_skip: ");
1099 dump_bitmap (symtab
->dump_file
,
1100 node
->clone
.args_to_skip
);
1102 if (node
->clone
.args_to_skip
)
1104 fprintf (symtab
->dump_file
, " combined_args_to_skip:");
1105 dump_bitmap (symtab
->dump_file
, node
->clone
.combined_args_to_skip
);
1108 cgraph_materialize_clone (node
);
1114 FOR_EACH_FUNCTION (node
)
1115 if (!node
->analyzed
&& node
->callees
)
1117 node
->remove_callees ();
1118 node
->remove_all_references ();
1121 node
->clear_stmts_in_references ();
1122 if (symtab
->dump_file
)
1123 fprintf (symtab
->dump_file
, "Materialization Call site updates done.\n");
1124 #ifdef ENABLE_CHECKING
1125 cgraph_node::verify_cgraph_nodes ();
1127 symtab
->remove_unreachable_nodes (false, symtab
->dump_file
);
1130 #include "gt-cgraphclones.h"