1 /* Callgraph transformations to handle inlining
2 Copyright (C) 2003-2015 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 /* The inline decisions are stored in callgraph in "inline plan" and
24 To mark given call inline, use inline_call function.
25 The function marks the edge inlinable and, if necessary, produces
26 virtual clone in the callgraph representing the new copy of callee's
29 The inline plan is applied on given function body by inline_transform. */
33 #include "coretypes.h"
38 #include "double-int.h"
45 #include "langhooks.h"
52 #include "plugin-api.h"
53 #include "hard-reg-set.h"
58 #include "alloc-pool.h"
59 #include "symbol-summary.h"
61 #include "ipa-inline.h"
62 #include "tree-inline.h"
63 #include "tree-pass.h"
66 int nfunctions_inlined
;
67 bool speculation_removed
;
69 /* Scale frequency of NODE edges by FREQ_SCALE. */
72 update_noncloned_frequencies (struct cgraph_node
*node
,
75 struct cgraph_edge
*e
;
77 /* We do not want to ignore high loop nest after freq drops to 0. */
80 for (e
= node
->callees
; e
; e
= e
->next_callee
)
82 e
->frequency
= e
->frequency
* (gcov_type
) freq_scale
/ CGRAPH_FREQ_BASE
;
83 if (e
->frequency
> CGRAPH_FREQ_MAX
)
84 e
->frequency
= CGRAPH_FREQ_MAX
;
85 if (!e
->inline_failed
)
86 update_noncloned_frequencies (e
->callee
, freq_scale
);
88 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
90 e
->frequency
= e
->frequency
* (gcov_type
) freq_scale
/ CGRAPH_FREQ_BASE
;
91 if (e
->frequency
> CGRAPH_FREQ_MAX
)
92 e
->frequency
= CGRAPH_FREQ_MAX
;
96 /* We removed or are going to remove the last call to NODE.
97 Return true if we can and want proactively remove the NODE now.
98 This is important to do, since we want inliner to know when offline
99 copy of function was removed. */
102 can_remove_node_now_p_1 (struct cgraph_node
*node
, struct cgraph_edge
*e
)
106 FOR_EACH_ALIAS (node
, ref
)
108 cgraph_node
*alias
= dyn_cast
<cgraph_node
*> (ref
->referring
);
109 if ((alias
->callers
&& alias
->callers
!= e
)
110 || !can_remove_node_now_p_1 (alias
, e
))
113 /* FIXME: When address is taken of DECL_EXTERNAL function we still
114 can remove its offline copy, but we would need to keep unanalyzed node in
115 the callgraph so references can point to it.
117 Also for comdat group we can ignore references inside a group as we
118 want to prove the group as a whole to be dead. */
119 return (!node
->address_taken
120 && node
->can_remove_if_no_direct_calls_and_refs_p ()
121 /* Inlining might enable more devirtualizing, so we want to remove
122 those only after all devirtualizable virtual calls are processed.
123 Lacking may edges in callgraph we just preserve them post
125 && (!DECL_VIRTUAL_P (node
->decl
)
126 || !opt_for_fn (node
->decl
, flag_devirtualize
))
127 /* During early inlining some unanalyzed cgraph nodes might be in the
128 callgraph and they might reffer the function in question. */
129 && !cgraph_new_nodes
.exists ());
132 /* We are going to eliminate last direct call to NODE (or alias of it) via edge E.
133 Verify that the NODE can be removed from unit and if it is contained in comdat
134 group that the whole comdat group is removable. */
137 can_remove_node_now_p (struct cgraph_node
*node
, struct cgraph_edge
*e
)
139 struct cgraph_node
*next
;
140 if (!can_remove_node_now_p_1 (node
, e
))
143 /* When we see same comdat group, we need to be sure that all
144 items can be removed. */
145 if (!node
->same_comdat_group
|| !node
->externally_visible
)
147 for (next
= dyn_cast
<cgraph_node
*> (node
->same_comdat_group
);
148 next
!= node
; next
= dyn_cast
<cgraph_node
*> (next
->same_comdat_group
))
152 if ((next
->callers
&& next
->callers
!= e
)
153 || !can_remove_node_now_p_1 (next
, e
))
159 /* Return true if NODE is a master clone with non-inline clones. */
162 master_clone_with_noninline_clones_p (struct cgraph_node
*node
)
167 for (struct cgraph_node
*n
= node
->clones
; n
; n
= n
->next_sibling_clone
)
168 if (n
->decl
!= node
->decl
)
174 /* E is expected to be an edge being inlined. Clone destination node of
175 the edge and redirect it to the new clone.
176 DUPLICATE is used for bookkeeping on whether we are actually creating new
177 clones or re-using node originally representing out-of-line function call.
178 By default the offline copy is removed, when it appears dead after inlining.
179 UPDATE_ORIGINAL prevents this transformation.
180 If OVERALL_SIZE is non-NULL, the size is updated to reflect the
182 FREQ_SCALE specify the scaling of frequencies of call sites. */
185 clone_inlined_nodes (struct cgraph_edge
*e
, bool duplicate
,
186 bool update_original
, int *overall_size
, int freq_scale
)
188 struct cgraph_node
*inlining_into
;
189 struct cgraph_edge
*next
;
191 if (e
->caller
->global
.inlined_to
)
192 inlining_into
= e
->caller
->global
.inlined_to
;
194 inlining_into
= e
->caller
;
198 /* We may eliminate the need for out-of-line copy to be output.
199 In that case just go ahead and re-use it. This is not just an
200 memory optimization. Making offline copy of fuction disappear
201 from the program will improve future decisions on inlining. */
202 if (!e
->callee
->callers
->next_caller
203 /* Recursive inlining never wants the master clone to
206 && can_remove_node_now_p (e
->callee
, e
)
207 /* We cannot overwrite a master clone with non-inline clones
208 until after these clones are materialized. */
209 && !master_clone_with_noninline_clones_p (e
->callee
))
211 /* TODO: When callee is in a comdat group, we could remove all of it,
212 including all inline clones inlined into it. That would however
213 need small function inlining to register edge removal hook to
214 maintain the priority queue.
216 For now we keep the ohter functions in the group in program until
217 cgraph_remove_unreachable_functions gets rid of them. */
218 gcc_assert (!e
->callee
->global
.inlined_to
);
219 e
->callee
->remove_from_same_comdat_group ();
220 if (e
->callee
->definition
221 && inline_account_function_p (e
->callee
))
223 gcc_assert (!e
->callee
->alias
);
225 *overall_size
-= inline_summaries
->get (e
->callee
)->size
;
226 nfunctions_inlined
++;
229 e
->callee
->externally_visible
= false;
230 update_noncloned_frequencies (e
->callee
, e
->frequency
);
234 struct cgraph_node
*n
;
236 if (freq_scale
== -1)
237 freq_scale
= e
->frequency
;
238 n
= e
->callee
->create_clone (e
->callee
->decl
,
239 MIN (e
->count
, e
->callee
->count
),
241 update_original
, vNULL
, true,
244 n
->used_as_abstract_origin
= e
->callee
->used_as_abstract_origin
;
245 e
->redirect_callee (n
);
249 e
->callee
->remove_from_same_comdat_group ();
251 e
->callee
->global
.inlined_to
= inlining_into
;
253 /* Recursively clone all bodies. */
254 for (e
= e
->callee
->callees
; e
; e
= next
)
256 next
= e
->next_callee
;
257 if (!e
->inline_failed
)
258 clone_inlined_nodes (e
, duplicate
, update_original
, overall_size
, freq_scale
);
259 if (e
->speculative
&& !speculation_useful_p (e
, true))
261 e
->resolve_speculation (NULL
);
262 speculation_removed
= true;
267 /* Mark all call graph edges coming out of NODE and all nodes that have been
268 inlined to it as in_polymorphic_cdtor. */
271 mark_all_inlined_calls_cdtor (cgraph_node
*node
)
273 for (cgraph_edge
*cs
= node
->callees
; cs
; cs
= cs
->next_callee
)
275 cs
->in_polymorphic_cdtor
= true;
276 if (!cs
->inline_failed
)
277 mark_all_inlined_calls_cdtor (cs
->callee
);
279 for (cgraph_edge
*cs
= node
->indirect_calls
; cs
; cs
= cs
->next_callee
)
280 cs
->in_polymorphic_cdtor
= true;
284 /* Mark edge E as inlined and update callgraph accordingly. UPDATE_ORIGINAL
285 specify whether profile of original function should be updated. If any new
286 indirect edges are discovered in the process, add them to NEW_EDGES, unless
287 it is NULL. If UPDATE_OVERALL_SUMMARY is false, do not bother to recompute overall
288 size of caller after inlining. Caller is required to eventually do it via
289 inline_update_overall_summary.
290 If callee_removed is non-NULL, set it to true if we removed callee node.
292 Return true iff any new callgraph edges were discovered as a
293 result of inlining. */
296 inline_call (struct cgraph_edge
*e
, bool update_original
,
297 vec
<cgraph_edge
*> *new_edges
,
298 int *overall_size
, bool update_overall_summary
,
299 bool *callee_removed
)
301 int old_size
= 0, new_size
= 0;
302 struct cgraph_node
*to
= NULL
;
303 struct cgraph_edge
*curr
= e
;
304 struct cgraph_node
*callee
= e
->callee
->ultimate_alias_target ();
305 bool new_edges_found
= false;
307 /* This is used only for assert bellow. */
309 int estimated_growth
= estimate_edge_growth (e
);
310 bool predicated
= inline_edge_summary (e
)->predicate
!= NULL
;
313 speculation_removed
= false;
314 /* Don't inline inlined edges. */
315 gcc_assert (e
->inline_failed
);
316 /* Don't even think of inlining inline clone. */
317 gcc_assert (!callee
->global
.inlined_to
);
319 e
->inline_failed
= CIF_OK
;
320 DECL_POSSIBLY_INLINED (callee
->decl
) = true;
323 if (to
->global
.inlined_to
)
324 to
= to
->global
.inlined_to
;
326 if (DECL_FUNCTION_PERSONALITY (callee
->decl
))
327 DECL_FUNCTION_PERSONALITY (to
->decl
)
328 = DECL_FUNCTION_PERSONALITY (callee
->decl
);
330 /* If aliases are involved, redirect edge to the actual destination and
331 possibly remove the aliases. */
332 if (e
->callee
!= callee
)
334 struct cgraph_node
*alias
= e
->callee
, *next_alias
;
335 e
->redirect_callee (callee
);
336 while (alias
&& alias
!= callee
)
339 && can_remove_node_now_p (alias
,
340 !e
->next_caller
&& !e
->prev_caller
? e
: NULL
))
342 next_alias
= alias
->get_alias_target ();
345 *callee_removed
= true;
353 clone_inlined_nodes (e
, true, update_original
, overall_size
, e
->frequency
);
355 gcc_assert (curr
->callee
->global
.inlined_to
== to
);
357 old_size
= inline_summaries
->get (to
)->size
;
358 inline_merge_summary (e
);
359 if (e
->in_polymorphic_cdtor
)
360 mark_all_inlined_calls_cdtor (e
->callee
);
361 if (opt_for_fn (e
->caller
->decl
, optimize
))
362 new_edges_found
= ipa_propagate_indirect_call_infos (curr
, new_edges
);
363 if (update_overall_summary
)
364 inline_update_overall_summary (to
);
365 new_size
= inline_summaries
->get (to
)->size
;
367 if (callee
->calls_comdat_local
)
368 to
->calls_comdat_local
= true;
369 else if (to
->calls_comdat_local
&& callee
->comdat_local_p ())
371 struct cgraph_edge
*se
= to
->callees
;
372 for (; se
; se
= se
->next_callee
)
373 if (se
->inline_failed
&& se
->callee
->comdat_local_p ())
376 to
->calls_comdat_local
= false;
379 /* FIXME: This assert suffers from roundoff errors, disable it for GCC 5
380 and revisit it after conversion to sreals in GCC 6.
383 /* Verify that estimated growth match real growth. Allow off-by-one
384 error due to INLINE_SIZE_SCALE roudoff errors. */
385 gcc_assert (!update_overall_summary
|| !overall_size
|| new_edges_found
386 || abs (estimated_growth
- (new_size
- old_size
)) <= 1
387 || speculation_removed
388 /* FIXME: a hack. Edges with false predicate are accounted
389 wrong, we should remove them from callgraph. */
393 /* Account the change of overall unit size; external functions will be
394 removed and are thus not accounted. */
395 if (overall_size
&& inline_account_function_p (to
))
396 *overall_size
+= new_size
- old_size
;
399 /* This must happen after inline_merge_summary that rely on jump
400 functions of callee to not be updated. */
401 return new_edges_found
;
405 /* Copy function body of NODE and redirect all inline clones to it.
406 This is done before inline plan is applied to NODE when there are
407 still some inline clones if it.
409 This is necessary because inline decisions are not really transitive
410 and the other inline clones may have different bodies. */
412 static struct cgraph_node
*
413 save_inline_function_body (struct cgraph_node
*node
)
415 struct cgraph_node
*first_clone
, *n
;
418 fprintf (dump_file
, "\nSaving body of %s for later reuse\n",
421 gcc_assert (node
== cgraph_node::get (node
->decl
));
423 /* first_clone will be turned into real function. */
424 first_clone
= node
->clones
;
425 first_clone
->decl
= copy_node (node
->decl
);
426 first_clone
->decl
->decl_with_vis
.symtab_node
= first_clone
;
427 gcc_assert (first_clone
== cgraph_node::get (first_clone
->decl
));
429 /* Now reshape the clone tree, so all other clones descends from
431 if (first_clone
->next_sibling_clone
)
433 for (n
= first_clone
->next_sibling_clone
; n
->next_sibling_clone
; n
= n
->next_sibling_clone
)
434 n
->clone_of
= first_clone
;
435 n
->clone_of
= first_clone
;
436 n
->next_sibling_clone
= first_clone
->clones
;
437 if (first_clone
->clones
)
438 first_clone
->clones
->prev_sibling_clone
= n
;
439 first_clone
->clones
= first_clone
->next_sibling_clone
;
440 first_clone
->next_sibling_clone
->prev_sibling_clone
= NULL
;
441 first_clone
->next_sibling_clone
= NULL
;
442 gcc_assert (!first_clone
->prev_sibling_clone
);
444 first_clone
->clone_of
= NULL
;
446 /* Now node in question has no clones. */
449 /* Inline clones share decl with the function they are cloned
450 from. Walk the whole clone tree and redirect them all to the
452 if (first_clone
->clones
)
453 for (n
= first_clone
->clones
; n
!= first_clone
;)
455 gcc_assert (n
->decl
== node
->decl
);
456 n
->decl
= first_clone
->decl
;
459 else if (n
->next_sibling_clone
)
460 n
= n
->next_sibling_clone
;
463 while (n
!= first_clone
&& !n
->next_sibling_clone
)
465 if (n
!= first_clone
)
466 n
= n
->next_sibling_clone
;
470 /* Copy the OLD_VERSION_NODE function tree to the new version. */
471 tree_function_versioning (node
->decl
, first_clone
->decl
,
472 NULL
, true, NULL
, false,
475 /* The function will be short lived and removed after we inline all the clones,
476 but make it internal so we won't confuse ourself. */
477 DECL_EXTERNAL (first_clone
->decl
) = 0;
478 TREE_PUBLIC (first_clone
->decl
) = 0;
479 DECL_COMDAT (first_clone
->decl
) = 0;
480 first_clone
->ipa_transforms_to_apply
.release ();
482 /* When doing recursive inlining, the clone may become unnecessary.
483 This is possible i.e. in the case when the recursive function is proved to be
484 non-throwing and the recursion happens only in the EH landing pad.
485 We can not remove the clone until we are done with saving the body.
487 if (!first_clone
->callers
)
489 first_clone
->remove_symbol_and_inline_clones ();
492 #ifdef ENABLE_CHECKING
494 first_clone
->verify ();
499 /* Return true when function body of DECL still needs to be kept around
502 preserve_function_body_p (struct cgraph_node
*node
)
504 gcc_assert (symtab
->global_info_ready
);
505 gcc_assert (!node
->alias
&& !node
->thunk
.thunk_p
);
507 /* Look if there is any clone around. */
513 /* Apply inline plan to function. */
516 inline_transform (struct cgraph_node
*node
)
518 unsigned int todo
= 0;
519 struct cgraph_edge
*e
, *next
;
520 bool has_inline
= false;
522 /* FIXME: Currently the pass manager is adding inline transform more than
523 once to some clones. This needs revisiting after WPA cleanups. */
524 if (cfun
->after_inlining
)
527 /* We might need the body of this function so that we can expand
528 it inline somewhere else. */
529 if (preserve_function_body_p (node
))
530 save_inline_function_body (node
);
532 for (e
= node
->callees
; e
; e
= next
)
534 if (!e
->inline_failed
)
536 next
= e
->next_callee
;
537 e
->redirect_call_stmt_to_callee ();
539 node
->remove_all_references ();
541 timevar_push (TV_INTEGRATION
);
542 if (node
->callees
&& (opt_for_fn (node
->decl
, optimize
) || has_inline
))
543 todo
= optimize_inline_calls (current_function_decl
);
544 timevar_pop (TV_INTEGRATION
);
546 cfun
->always_inline_functions_inlined
= true;
547 cfun
->after_inlining
= true;
548 todo
|= execute_fixup_cfg ();
550 if (!(todo
& TODO_update_ssa_any
))
551 /* Redirecting edges might lead to a need for vops to be recomputed. */
552 todo
|= TODO_update_ssa_only_virtuals
;