1 /* Callgraph transformations to handle inlining
2 Copyright (C) 2003-2014 Free Software Foundation, Inc.
3 Contributed by Jan Hubicka
4
5 This file is part of GCC.
6
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
10 version.
11
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
15 for more details.
16
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/>. */
20
21 /* The inline decisions are stored in callgraph in "inline plan" and
22 applied later.
23
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
27 function body.
28
29 The inline plan is applied on given function body by inline_transform. */
30
31 #include "config.h"
32 #include "system.h"
33 #include "coretypes.h"
34 #include "tm.h"
35 #include "tree.h"
36 #include "langhooks.h"
37 #include "intl.h"
38 #include "coverage.h"
39 #include "ggc.h"
40 #include "tree-cfg.h"
41 #include "ipa-prop.h"
42 #include "ipa-inline.h"
43 #include "tree-inline.h"
44 #include "tree-pass.h"
45
46 int ncalls_inlined;
47 int nfunctions_inlined;
48 bool speculation_removed;
49
50 /* Scale frequency of NODE edges by FREQ_SCALE. */
51
52 static void
update_noncloned_frequencies(struct cgraph_node * node,int freq_scale)53 update_noncloned_frequencies (struct cgraph_node *node,
54 int freq_scale)
55 {
56 struct cgraph_edge *e;
57
58 /* We do not want to ignore high loop nest after freq drops to 0. */
59 if (!freq_scale)
60 freq_scale = 1;
61 for (e = node->callees; e; e = e->next_callee)
62 {
63 e->frequency = e->frequency * (gcov_type) freq_scale / CGRAPH_FREQ_BASE;
64 if (e->frequency > CGRAPH_FREQ_MAX)
65 e->frequency = CGRAPH_FREQ_MAX;
66 if (!e->inline_failed)
67 update_noncloned_frequencies (e->callee, freq_scale);
68 }
69 for (e = node->indirect_calls; e; e = e->next_callee)
70 {
71 e->frequency = e->frequency * (gcov_type) freq_scale / CGRAPH_FREQ_BASE;
72 if (e->frequency > CGRAPH_FREQ_MAX)
73 e->frequency = CGRAPH_FREQ_MAX;
74 }
75 }
76
77 /* We removed or are going to remove the last call to NODE.
78 Return true if we can and want proactively remove the NODE now.
79 This is important to do, since we want inliner to know when offline
80 copy of function was removed. */
81
82 static bool
can_remove_node_now_p_1(struct cgraph_node * node)83 can_remove_node_now_p_1 (struct cgraph_node *node)
84 {
85 /* FIXME: When address is taken of DECL_EXTERNAL function we still
86 can remove its offline copy, but we would need to keep unanalyzed node in
87 the callgraph so references can point to it. */
88 return (!node->address_taken
89 && !ipa_ref_has_aliases_p (&node->ref_list)
90 && !node->used_as_abstract_origin
91 && cgraph_can_remove_if_no_direct_calls_p (node)
92 /* Inlining might enable more devirtualizing, so we want to remove
93 those only after all devirtualizable virtual calls are processed.
94 Lacking may edges in callgraph we just preserve them post
95 inlining. */
96 && !DECL_VIRTUAL_P (node->decl)
97 /* During early inlining some unanalyzed cgraph nodes might be in the
98 callgraph and they might reffer the function in question. */
99 && !cgraph_new_nodes);
100 }
101
102 /* We are going to eliminate last direct call to NODE (or alias of it) via edge E.
103 Verify that the NODE can be removed from unit and if it is contained in comdat
104 group that the whole comdat group is removable. */
105
106 static bool
can_remove_node_now_p(struct cgraph_node * node,struct cgraph_edge * e)107 can_remove_node_now_p (struct cgraph_node *node, struct cgraph_edge *e)
108 {
109 struct cgraph_node *next;
110 if (!can_remove_node_now_p_1 (node))
111 return false;
112
113 /* When we see same comdat group, we need to be sure that all
114 items can be removed. */
115 if (!node->same_comdat_group)
116 return true;
117 for (next = cgraph (node->same_comdat_group);
118 next != node; next = cgraph (next->same_comdat_group))
119 if ((next->callers && next->callers != e)
120 || !can_remove_node_now_p_1 (next))
121 return false;
122 return true;
123 }
124
125
126 /* E is expected to be an edge being inlined. Clone destination node of
127 the edge and redirect it to the new clone.
128 DUPLICATE is used for bookkeeping on whether we are actually creating new
129 clones or re-using node originally representing out-of-line function call.
130 By default the offline copy is removed, when it appears dead after inlining.
131 UPDATE_ORIGINAL prevents this transformation.
132 If OVERALL_SIZE is non-NULL, the size is updated to reflect the
133 transformation.
134 FREQ_SCALE specify the scaling of frequencies of call sites. */
135
136 void
clone_inlined_nodes(struct cgraph_edge * e,bool duplicate,bool update_original,int * overall_size,int freq_scale)137 clone_inlined_nodes (struct cgraph_edge *e, bool duplicate,
138 bool update_original, int *overall_size, int freq_scale)
139 {
140 struct cgraph_node *inlining_into;
141 struct cgraph_edge *next;
142
143 if (e->caller->global.inlined_to)
144 inlining_into = e->caller->global.inlined_to;
145 else
146 inlining_into = e->caller;
147
148 if (duplicate)
149 {
150 /* We may eliminate the need for out-of-line copy to be output.
151 In that case just go ahead and re-use it. This is not just an
152 memory optimization. Making offline copy of fuction disappear
153 from the program will improve future decisions on inlining. */
154 if (!e->callee->callers->next_caller
155 /* Recursive inlining never wants the master clone to
156 be overwritten. */
157 && update_original
158 && can_remove_node_now_p (e->callee, e))
159 {
160 /* TODO: When callee is in a comdat group, we could remove all of it,
161 including all inline clones inlined into it. That would however
162 need small function inlining to register edge removal hook to
163 maintain the priority queue.
164
165 For now we keep the ohter functions in the group in program until
166 cgraph_remove_unreachable_functions gets rid of them. */
167 gcc_assert (!e->callee->global.inlined_to);
168 symtab_dissolve_same_comdat_group_list (e->callee);
169 if (e->callee->definition && !DECL_EXTERNAL (e->callee->decl))
170 {
171 if (overall_size)
172 *overall_size -= inline_summary (e->callee)->size;
173 nfunctions_inlined++;
174 }
175 duplicate = false;
176 e->callee->externally_visible = false;
177 update_noncloned_frequencies (e->callee, e->frequency);
178 }
179 else
180 {
181 struct cgraph_node *n;
182
183 if (freq_scale == -1)
184 freq_scale = e->frequency;
185 n = cgraph_clone_node (e->callee, e->callee->decl,
186 e->count, freq_scale, update_original,
187 vNULL, true, inlining_into, NULL);
188 cgraph_redirect_edge_callee (e, n);
189 }
190 }
191 else
192 symtab_dissolve_same_comdat_group_list (e->callee);
193
194 e->callee->global.inlined_to = inlining_into;
195
196 /* Recursively clone all bodies. */
197 for (e = e->callee->callees; e; e = next)
198 {
199 next = e->next_callee;
200 if (!e->inline_failed)
201 clone_inlined_nodes (e, duplicate, update_original, overall_size, freq_scale);
202 if (e->speculative && !speculation_useful_p (e, true))
203 {
204 cgraph_resolve_speculation (e, NULL);
205 speculation_removed = true;
206 }
207 }
208 }
209
210
211 /* Mark edge E as inlined and update callgraph accordingly. UPDATE_ORIGINAL
212 specify whether profile of original function should be updated. If any new
213 indirect edges are discovered in the process, add them to NEW_EDGES, unless
214 it is NULL. If UPDATE_OVERALL_SUMMARY is false, do not bother to recompute overall
215 size of caller after inlining. Caller is required to eventually do it via
216 inline_update_overall_summary.
217 If callee_removed is non-NULL, set it to true if we removed callee node.
218
219 Return true iff any new callgraph edges were discovered as a
220 result of inlining. */
221
222 bool
inline_call(struct cgraph_edge * e,bool update_original,vec<cgraph_edge_p> * new_edges,int * overall_size,bool update_overall_summary,bool * callee_removed)223 inline_call (struct cgraph_edge *e, bool update_original,
224 vec<cgraph_edge_p> *new_edges,
225 int *overall_size, bool update_overall_summary,
226 bool *callee_removed)
227 {
228 int old_size = 0, new_size = 0;
229 struct cgraph_node *to = NULL;
230 struct cgraph_edge *curr = e;
231 struct cgraph_node *callee = cgraph_function_or_thunk_node (e->callee, NULL);
232 bool new_edges_found = false;
233
234 #ifdef ENABLE_CHECKING
235 int estimated_growth = estimate_edge_growth (e);
236 bool predicated = inline_edge_summary (e)->predicate != NULL;
237 #endif
238
239 speculation_removed = false;
240 /* Don't inline inlined edges. */
241 gcc_assert (e->inline_failed);
242 /* Don't even think of inlining inline clone. */
243 gcc_assert (!callee->global.inlined_to);
244
245 e->inline_failed = CIF_OK;
246 DECL_POSSIBLY_INLINED (callee->decl) = true;
247
248 to = e->caller;
249 if (to->global.inlined_to)
250 to = to->global.inlined_to;
251
252 /* If aliases are involved, redirect edge to the actual destination and
253 possibly remove the aliases. */
254 if (e->callee != callee)
255 {
256 struct cgraph_node *alias = e->callee, *next_alias;
257 cgraph_redirect_edge_callee (e, callee);
258 while (alias && alias != callee)
259 {
260 if (!alias->callers
261 && can_remove_node_now_p (alias, e))
262 {
263 next_alias = cgraph_alias_target (alias);
264 cgraph_remove_node (alias);
265 if (callee_removed)
266 *callee_removed = true;
267 alias = next_alias;
268 }
269 else
270 break;
271 }
272 }
273
274 clone_inlined_nodes (e, true, update_original, overall_size, e->frequency);
275
276 gcc_assert (curr->callee->global.inlined_to == to);
277
278 old_size = inline_summary (to)->size;
279 inline_merge_summary (e);
280 if (optimize)
281 new_edges_found = ipa_propagate_indirect_call_infos (curr, new_edges);
282 if (update_overall_summary)
283 inline_update_overall_summary (to);
284 new_size = inline_summary (to)->size;
285
286 if (callee->calls_comdat_local)
287 to->calls_comdat_local = true;
288 else if (to->calls_comdat_local && symtab_comdat_local_p (callee))
289 {
290 struct cgraph_edge *se = to->callees;
291 for (; se; se = se->next_callee)
292 if (se->inline_failed && symtab_comdat_local_p (se->callee))
293 break;
294 if (se == NULL)
295 to->calls_comdat_local = false;
296 }
297
298 #ifdef ENABLE_CHECKING
299 /* Verify that estimated growth match real growth. Allow off-by-one
300 error due to INLINE_SIZE_SCALE roudoff errors. */
301 gcc_assert (!update_overall_summary || !overall_size || new_edges_found
302 || abs (estimated_growth - (new_size - old_size)) <= 1
303 || speculation_removed
304 /* FIXME: a hack. Edges with false predicate are accounted
305 wrong, we should remove them from callgraph. */
306 || predicated);
307 #endif
308
309 /* Account the change of overall unit size; external functions will be
310 removed and are thus not accounted. */
311 if (overall_size
312 && !DECL_EXTERNAL (to->decl))
313 *overall_size += new_size - old_size;
314 ncalls_inlined++;
315
316 /* This must happen after inline_merge_summary that rely on jump
317 functions of callee to not be updated. */
318 return new_edges_found;
319 }
320
321
322 /* Copy function body of NODE and redirect all inline clones to it.
323 This is done before inline plan is applied to NODE when there are
324 still some inline clones if it.
325
326 This is necessary because inline decisions are not really transitive
327 and the other inline clones may have different bodies. */
328
329 static struct cgraph_node *
save_inline_function_body(struct cgraph_node * node)330 save_inline_function_body (struct cgraph_node *node)
331 {
332 struct cgraph_node *first_clone, *n;
333
334 if (dump_file)
335 fprintf (dump_file, "\nSaving body of %s for later reuse\n",
336 node->name ());
337
338 gcc_assert (node == cgraph_get_node (node->decl));
339
340 /* first_clone will be turned into real function. */
341 first_clone = node->clones;
342 first_clone->decl = copy_node (node->decl);
343 symtab_insert_node_to_hashtable (first_clone);
344 gcc_assert (first_clone == cgraph_get_node (first_clone->decl));
345
346 /* Now reshape the clone tree, so all other clones descends from
347 first_clone. */
348 if (first_clone->next_sibling_clone)
349 {
350 for (n = first_clone->next_sibling_clone; n->next_sibling_clone; n = n->next_sibling_clone)
351 n->clone_of = first_clone;
352 n->clone_of = first_clone;
353 n->next_sibling_clone = first_clone->clones;
354 if (first_clone->clones)
355 first_clone->clones->prev_sibling_clone = n;
356 first_clone->clones = first_clone->next_sibling_clone;
357 first_clone->next_sibling_clone->prev_sibling_clone = NULL;
358 first_clone->next_sibling_clone = NULL;
359 gcc_assert (!first_clone->prev_sibling_clone);
360 }
361 first_clone->clone_of = NULL;
362
363 /* Now node in question has no clones. */
364 node->clones = NULL;
365
366 /* Inline clones share decl with the function they are cloned
367 from. Walk the whole clone tree and redirect them all to the
368 new decl. */
369 if (first_clone->clones)
370 for (n = first_clone->clones; n != first_clone;)
371 {
372 gcc_assert (n->decl == node->decl);
373 n->decl = first_clone->decl;
374 if (n->clones)
375 n = n->clones;
376 else if (n->next_sibling_clone)
377 n = n->next_sibling_clone;
378 else
379 {
380 while (n != first_clone && !n->next_sibling_clone)
381 n = n->clone_of;
382 if (n != first_clone)
383 n = n->next_sibling_clone;
384 }
385 }
386
387 /* Copy the OLD_VERSION_NODE function tree to the new version. */
388 tree_function_versioning (node->decl, first_clone->decl,
389 NULL, true, NULL, false,
390 NULL, NULL);
391
392 /* The function will be short lived and removed after we inline all the clones,
393 but make it internal so we won't confuse ourself. */
394 DECL_EXTERNAL (first_clone->decl) = 0;
395 TREE_PUBLIC (first_clone->decl) = 0;
396 DECL_COMDAT (first_clone->decl) = 0;
397 first_clone->ipa_transforms_to_apply.release ();
398
399 /* When doing recursive inlining, the clone may become unnecessary.
400 This is possible i.e. in the case when the recursive function is proved to be
401 non-throwing and the recursion happens only in the EH landing pad.
402 We can not remove the clone until we are done with saving the body.
403 Remove it now. */
404 if (!first_clone->callers)
405 {
406 cgraph_remove_node_and_inline_clones (first_clone, NULL);
407 first_clone = NULL;
408 }
409 #ifdef ENABLE_CHECKING
410 else
411 verify_cgraph_node (first_clone);
412 #endif
413 return first_clone;
414 }
415
416 /* Return true when function body of DECL still needs to be kept around
417 for later re-use. */
418 static bool
preserve_function_body_p(struct cgraph_node * node)419 preserve_function_body_p (struct cgraph_node *node)
420 {
421 gcc_assert (cgraph_global_info_ready);
422 gcc_assert (!node->alias && !node->thunk.thunk_p);
423
424 /* Look if there is any clone around. */
425 if (node->clones)
426 return true;
427 return false;
428 }
429
430 /* Apply inline plan to function. */
431
432 unsigned int
inline_transform(struct cgraph_node * node)433 inline_transform (struct cgraph_node *node)
434 {
435 unsigned int todo = 0;
436 struct cgraph_edge *e, *next;
437
438 /* FIXME: Currently the pass manager is adding inline transform more than
439 once to some clones. This needs revisiting after WPA cleanups. */
440 if (cfun->after_inlining)
441 return 0;
442
443 /* We might need the body of this function so that we can expand
444 it inline somewhere else. */
445 if (preserve_function_body_p (node))
446 save_inline_function_body (node);
447
448 for (e = node->callees; e; e = next)
449 {
450 next = e->next_callee;
451 cgraph_redirect_edge_call_stmt_to_callee (e);
452 }
453 ipa_remove_all_references (&node->ref_list);
454
455 timevar_push (TV_INTEGRATION);
456 if (node->callees && optimize)
457 todo = optimize_inline_calls (current_function_decl);
458 timevar_pop (TV_INTEGRATION);
459
460 cfun->always_inline_functions_inlined = true;
461 cfun->after_inlining = true;
462 todo |= execute_fixup_cfg ();
463
464 if (!(todo & TODO_update_ssa_any))
465 /* Redirecting edges might lead to a need for vops to be recomputed. */
466 todo |= TODO_update_ssa_only_virtuals;
467
468 return todo;
469 }
470