]>
Commit | Line | Data |
---|---|---|
fee8b6da | 1 | /* Callgraph transformations to handle inlining |
cbe34bb5 | 2 | Copyright (C) 2003-2017 Free Software Foundation, Inc. |
fee8b6da JH |
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" | |
957060b5 | 35 | #include "function.h" |
fee8b6da | 36 | #include "tree.h" |
957060b5 AM |
37 | #include "alloc-pool.h" |
38 | #include "tree-pass.h" | |
39 | #include "cgraph.h" | |
442b4905 | 40 | #include "tree-cfg.h" |
dd912cb8 | 41 | #include "symbol-summary.h" |
8bc5448f | 42 | #include "tree-vrp.h" |
fee8b6da | 43 | #include "ipa-prop.h" |
27d020cf | 44 | #include "ipa-fnsummary.h" |
fee8b6da JH |
45 | #include "ipa-inline.h" |
46 | #include "tree-inline.h" | |
47 | ||
48 | int ncalls_inlined; | |
49 | int nfunctions_inlined; | |
50 | ||
898b8927 | 51 | /* Scale frequency of NODE edges by FREQ_SCALE. */ |
fee8b6da JH |
52 | |
53 | static void | |
54 | update_noncloned_frequencies (struct cgraph_node *node, | |
898b8927 | 55 | int freq_scale) |
fee8b6da JH |
56 | { |
57 | struct cgraph_edge *e; | |
58 | ||
59 | /* We do not want to ignore high loop nest after freq drops to 0. */ | |
60 | if (!freq_scale) | |
61 | freq_scale = 1; | |
62 | for (e = node->callees; e; e = e->next_callee) | |
63 | { | |
fee8b6da JH |
64 | e->frequency = e->frequency * (gcov_type) freq_scale / CGRAPH_FREQ_BASE; |
65 | if (e->frequency > CGRAPH_FREQ_MAX) | |
66 | e->frequency = CGRAPH_FREQ_MAX; | |
67 | if (!e->inline_failed) | |
898b8927 JH |
68 | update_noncloned_frequencies (e->callee, freq_scale); |
69 | } | |
70 | for (e = node->indirect_calls; e; e = e->next_callee) | |
71 | { | |
72 | e->frequency = e->frequency * (gcov_type) freq_scale / CGRAPH_FREQ_BASE; | |
73 | if (e->frequency > CGRAPH_FREQ_MAX) | |
74 | e->frequency = CGRAPH_FREQ_MAX; | |
fee8b6da JH |
75 | } |
76 | } | |
77 | ||
a5b1779f JH |
78 | /* We removed or are going to remove the last call to NODE. |
79 | Return true if we can and want proactively remove the NODE now. | |
80 | This is important to do, since we want inliner to know when offline | |
81 | copy of function was removed. */ | |
82 | ||
83 | static bool | |
d142079a | 84 | can_remove_node_now_p_1 (struct cgraph_node *node, struct cgraph_edge *e) |
a5b1779f | 85 | { |
d142079a JH |
86 | ipa_ref *ref; |
87 | ||
88 | FOR_EACH_ALIAS (node, ref) | |
89 | { | |
90 | cgraph_node *alias = dyn_cast <cgraph_node *> (ref->referring); | |
91 | if ((alias->callers && alias->callers != e) | |
92 | || !can_remove_node_now_p_1 (alias, e)) | |
93 | return false; | |
94 | } | |
a5b1779f JH |
95 | /* FIXME: When address is taken of DECL_EXTERNAL function we still |
96 | can remove its offline copy, but we would need to keep unanalyzed node in | |
a6a543bf JH |
97 | the callgraph so references can point to it. |
98 | ||
99 | Also for comdat group we can ignore references inside a group as we | |
100 | want to prove the group as a whole to be dead. */ | |
67348ccc | 101 | return (!node->address_taken |
a6a543bf | 102 | && node->can_remove_if_no_direct_calls_and_refs_p () |
a5b1779f JH |
103 | /* Inlining might enable more devirtualizing, so we want to remove |
104 | those only after all devirtualizable virtual calls are processed. | |
105 | Lacking may edges in callgraph we just preserve them post | |
106 | inlining. */ | |
d142079a JH |
107 | && (!DECL_VIRTUAL_P (node->decl) |
108 | || !opt_for_fn (node->decl, flag_devirtualize)) | |
a5b1779f JH |
109 | /* During early inlining some unanalyzed cgraph nodes might be in the |
110 | callgraph and they might reffer the function in question. */ | |
31acf1bb | 111 | && !cgraph_new_nodes.exists ()); |
a5b1779f JH |
112 | } |
113 | ||
6c69a029 JH |
114 | /* We are going to eliminate last direct call to NODE (or alias of it) via edge E. |
115 | Verify that the NODE can be removed from unit and if it is contained in comdat | |
116 | group that the whole comdat group is removable. */ | |
117 | ||
118 | static bool | |
119 | can_remove_node_now_p (struct cgraph_node *node, struct cgraph_edge *e) | |
120 | { | |
121 | struct cgraph_node *next; | |
d142079a | 122 | if (!can_remove_node_now_p_1 (node, e)) |
6c69a029 JH |
123 | return false; |
124 | ||
125 | /* When we see same comdat group, we need to be sure that all | |
126 | items can be removed. */ | |
8ccc8042 | 127 | if (!node->same_comdat_group || !node->externally_visible) |
6c69a029 | 128 | return true; |
d52f5295 ML |
129 | for (next = dyn_cast<cgraph_node *> (node->same_comdat_group); |
130 | next != node; next = dyn_cast<cgraph_node *> (next->same_comdat_group)) | |
d142079a JH |
131 | { |
132 | if (next->alias) | |
133 | continue; | |
134 | if ((next->callers && next->callers != e) | |
135 | || !can_remove_node_now_p_1 (next, e)) | |
136 | return false; | |
137 | } | |
6c69a029 JH |
138 | return true; |
139 | } | |
140 | ||
d83fa499 EB |
141 | /* Return true if NODE is a master clone with non-inline clones. */ |
142 | ||
143 | static bool | |
144 | master_clone_with_noninline_clones_p (struct cgraph_node *node) | |
145 | { | |
146 | if (node->clone_of) | |
147 | return false; | |
148 | ||
149 | for (struct cgraph_node *n = node->clones; n; n = n->next_sibling_clone) | |
150 | if (n->decl != node->decl) | |
151 | return true; | |
152 | ||
153 | return false; | |
154 | } | |
fee8b6da JH |
155 | |
156 | /* E is expected to be an edge being inlined. Clone destination node of | |
157 | the edge and redirect it to the new clone. | |
158 | DUPLICATE is used for bookkeeping on whether we are actually creating new | |
159 | clones or re-using node originally representing out-of-line function call. | |
bd936951 JH |
160 | By default the offline copy is removed, when it appears dead after inlining. |
161 | UPDATE_ORIGINAL prevents this transformation. | |
162 | If OVERALL_SIZE is non-NULL, the size is updated to reflect the | |
163 | transformation. | |
164 | FREQ_SCALE specify the scaling of frequencies of call sites. */ | |
fee8b6da JH |
165 | |
166 | void | |
167 | clone_inlined_nodes (struct cgraph_edge *e, bool duplicate, | |
bd936951 | 168 | bool update_original, int *overall_size, int freq_scale) |
fee8b6da | 169 | { |
44a60244 | 170 | struct cgraph_node *inlining_into; |
09ce3660 | 171 | struct cgraph_edge *next; |
44a60244 MJ |
172 | |
173 | if (e->caller->global.inlined_to) | |
174 | inlining_into = e->caller->global.inlined_to; | |
175 | else | |
176 | inlining_into = e->caller; | |
177 | ||
fee8b6da JH |
178 | if (duplicate) |
179 | { | |
180 | /* We may eliminate the need for out-of-line copy to be output. | |
181 | In that case just go ahead and re-use it. This is not just an | |
182 | memory optimization. Making offline copy of fuction disappear | |
183 | from the program will improve future decisions on inlining. */ | |
184 | if (!e->callee->callers->next_caller | |
185 | /* Recursive inlining never wants the master clone to | |
186 | be overwritten. */ | |
187 | && update_original | |
d83fa499 EB |
188 | && can_remove_node_now_p (e->callee, e) |
189 | /* We cannot overwrite a master clone with non-inline clones | |
190 | until after these clones are materialized. */ | |
191 | && !master_clone_with_noninline_clones_p (e->callee)) | |
fee8b6da | 192 | { |
6c69a029 JH |
193 | /* TODO: When callee is in a comdat group, we could remove all of it, |
194 | including all inline clones inlined into it. That would however | |
195 | need small function inlining to register edge removal hook to | |
196 | maintain the priority queue. | |
197 | ||
198 | For now we keep the ohter functions in the group in program until | |
199 | cgraph_remove_unreachable_functions gets rid of them. */ | |
fee8b6da | 200 | gcc_assert (!e->callee->global.inlined_to); |
b91b562c | 201 | e->callee->remove_from_same_comdat_group (); |
bb1e543c JH |
202 | if (e->callee->definition |
203 | && inline_account_function_p (e->callee)) | |
fee8b6da | 204 | { |
bb1e543c | 205 | gcc_assert (!e->callee->alias); |
fee8b6da | 206 | if (overall_size) |
0bceb671 | 207 | *overall_size -= ipa_fn_summaries->get (e->callee)->size; |
fee8b6da JH |
208 | nfunctions_inlined++; |
209 | } | |
210 | duplicate = false; | |
67348ccc | 211 | e->callee->externally_visible = false; |
898b8927 | 212 | update_noncloned_frequencies (e->callee, e->frequency); |
0bdad123 ML |
213 | |
214 | dump_callgraph_transformation (e->callee, inlining_into, | |
215 | "inlining to"); | |
fee8b6da JH |
216 | } |
217 | else | |
218 | { | |
219 | struct cgraph_node *n; | |
bd936951 JH |
220 | |
221 | if (freq_scale == -1) | |
222 | freq_scale = e->frequency; | |
d52f5295 ML |
223 | n = e->callee->create_clone (e->callee->decl, |
224 | MIN (e->count, e->callee->count), | |
225 | freq_scale, | |
226 | update_original, vNULL, true, | |
227 | inlining_into, | |
228 | NULL); | |
4ad08ee8 | 229 | n->used_as_abstract_origin = e->callee->used_as_abstract_origin; |
3dafb85c | 230 | e->redirect_callee (n); |
fee8b6da JH |
231 | } |
232 | } | |
65d630d4 | 233 | else |
b91b562c | 234 | e->callee->remove_from_same_comdat_group (); |
fee8b6da | 235 | |
44a60244 | 236 | e->callee->global.inlined_to = inlining_into; |
fee8b6da JH |
237 | |
238 | /* Recursively clone all bodies. */ | |
09ce3660 JH |
239 | for (e = e->callee->callees; e; e = next) |
240 | { | |
241 | next = e->next_callee; | |
242 | if (!e->inline_failed) | |
bd936951 | 243 | clone_inlined_nodes (e, duplicate, update_original, overall_size, freq_scale); |
db66bf68 JH |
244 | } |
245 | } | |
246 | ||
247 | /* Check all speculations in N and resolve them if they seems useless. */ | |
248 | ||
249 | static bool | |
250 | check_speculations (cgraph_node *n) | |
251 | { | |
252 | bool speculation_removed = false; | |
253 | cgraph_edge *next; | |
254 | ||
255 | for (cgraph_edge *e = n->callees; e; e = next) | |
256 | { | |
257 | next = e->next_callee; | |
09ce3660 JH |
258 | if (e->speculative && !speculation_useful_p (e, true)) |
259 | { | |
3dafb85c | 260 | e->resolve_speculation (NULL); |
09ce3660 JH |
261 | speculation_removed = true; |
262 | } | |
db66bf68 JH |
263 | else if (!e->inline_failed) |
264 | speculation_removed |= check_speculations (e->callee); | |
09ce3660 | 265 | } |
db66bf68 | 266 | return speculation_removed; |
fee8b6da JH |
267 | } |
268 | ||
4fd94d1e MJ |
269 | /* Mark all call graph edges coming out of NODE and all nodes that have been |
270 | inlined to it as in_polymorphic_cdtor. */ | |
271 | ||
272 | static void | |
273 | mark_all_inlined_calls_cdtor (cgraph_node *node) | |
274 | { | |
275 | for (cgraph_edge *cs = node->callees; cs; cs = cs->next_callee) | |
276 | { | |
277 | cs->in_polymorphic_cdtor = true; | |
278 | if (!cs->inline_failed) | |
279 | mark_all_inlined_calls_cdtor (cs->callee); | |
280 | } | |
281 | for (cgraph_edge *cs = node->indirect_calls; cs; cs = cs->next_callee) | |
282 | cs->in_polymorphic_cdtor = true; | |
283 | } | |
284 | ||
fee8b6da JH |
285 | |
286 | /* Mark edge E as inlined and update callgraph accordingly. UPDATE_ORIGINAL | |
287 | specify whether profile of original function should be updated. If any new | |
288 | indirect edges are discovered in the process, add them to NEW_EDGES, unless | |
c170d40f JH |
289 | it is NULL. If UPDATE_OVERALL_SUMMARY is false, do not bother to recompute overall |
290 | size of caller after inlining. Caller is required to eventually do it via | |
0bceb671 | 291 | ipa_update_overall_fn_summary. |
1bbb87c4 | 292 | If callee_removed is non-NULL, set it to true if we removed callee node. |
c170d40f JH |
293 | |
294 | Return true iff any new callgraph edges were discovered as a | |
fee8b6da JH |
295 | result of inlining. */ |
296 | ||
297 | bool | |
298 | inline_call (struct cgraph_edge *e, bool update_original, | |
d52f5295 | 299 | vec<cgraph_edge *> *new_edges, |
1bbb87c4 JH |
300 | int *overall_size, bool update_overall_summary, |
301 | bool *callee_removed) | |
fee8b6da JH |
302 | { |
303 | int old_size = 0, new_size = 0; | |
304 | struct cgraph_node *to = NULL; | |
305 | struct cgraph_edge *curr = e; | |
d52f5295 | 306 | struct cgraph_node *callee = e->callee->ultimate_alias_target (); |
0f378cb5 JH |
307 | bool new_edges_found = false; |
308 | ||
bddead15 RB |
309 | int estimated_growth = 0; |
310 | if (! update_overall_summary) | |
311 | estimated_growth = estimate_edge_growth (e); | |
f107227b JH |
312 | /* This is used only for assert bellow. */ |
313 | #if 0 | |
0f378cb5 JH |
314 | bool predicated = inline_edge_summary (e)->predicate != NULL; |
315 | #endif | |
fee8b6da JH |
316 | |
317 | /* Don't inline inlined edges. */ | |
318 | gcc_assert (e->inline_failed); | |
319 | /* Don't even think of inlining inline clone. */ | |
a5b1779f | 320 | gcc_assert (!callee->global.inlined_to); |
fee8b6da | 321 | |
632b4f8e JH |
322 | to = e->caller; |
323 | if (to->global.inlined_to) | |
324 | to = to->global.inlined_to; | |
0b9004ed JH |
325 | if (to->thunk.thunk_p) |
326 | { | |
c7ed8938 | 327 | struct cgraph_node *target = to->callees->callee; |
0b9004ed JH |
328 | if (in_lto_p) |
329 | to->get_untransformed_body (); | |
330 | to->expand_thunk (false, true); | |
c7ed8938 IE |
331 | /* When thunk is instrumented we may have multiple callees. */ |
332 | for (e = to->callees; e && e->callee != target; e = e->next_callee) | |
333 | ; | |
334 | gcc_assert (e); | |
0b9004ed JH |
335 | } |
336 | ||
337 | ||
338 | e->inline_failed = CIF_OK; | |
339 | DECL_POSSIBLY_INLINED (callee->decl) = true; | |
632b4f8e | 340 | |
5058c037 JH |
341 | if (DECL_FUNCTION_PERSONALITY (callee->decl)) |
342 | DECL_FUNCTION_PERSONALITY (to->decl) | |
343 | = DECL_FUNCTION_PERSONALITY (callee->decl); | |
77719b06 ML |
344 | |
345 | bool reload_optimization_node = false; | |
45285060 JH |
346 | if (!opt_for_fn (callee->decl, flag_strict_aliasing) |
347 | && opt_for_fn (to->decl, flag_strict_aliasing)) | |
348 | { | |
349 | struct gcc_options opts = global_options; | |
350 | ||
f7f32acd | 351 | cl_optimization_restore (&opts, opts_for_fn (to->decl)); |
45285060 JH |
352 | opts.x_flag_strict_aliasing = false; |
353 | if (dump_file) | |
354 | fprintf (dump_file, "Dropping flag_strict_aliasing on %s:%i\n", | |
355 | to->name (), to->order); | |
45285060 JH |
356 | DECL_FUNCTION_SPECIFIC_OPTIMIZATION (to->decl) |
357 | = build_optimization_node (&opts); | |
77719b06 | 358 | reload_optimization_node = true; |
45285060 | 359 | } |
5c846a81 | 360 | |
0bceb671 JH |
361 | ipa_fn_summary *caller_info = ipa_fn_summaries->get (to); |
362 | ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee); | |
818b88a7 JH |
363 | if (!caller_info->fp_expressions && callee_info->fp_expressions) |
364 | { | |
365 | caller_info->fp_expressions = true; | |
366 | if (opt_for_fn (callee->decl, flag_rounding_math) | |
367 | != opt_for_fn (to->decl, flag_rounding_math) | |
368 | || opt_for_fn (callee->decl, flag_trapping_math) | |
369 | != opt_for_fn (to->decl, flag_trapping_math) | |
370 | || opt_for_fn (callee->decl, flag_unsafe_math_optimizations) | |
371 | != opt_for_fn (to->decl, flag_unsafe_math_optimizations) | |
372 | || opt_for_fn (callee->decl, flag_finite_math_only) | |
373 | != opt_for_fn (to->decl, flag_finite_math_only) | |
374 | || opt_for_fn (callee->decl, flag_signaling_nans) | |
375 | != opt_for_fn (to->decl, flag_signaling_nans) | |
376 | || opt_for_fn (callee->decl, flag_cx_limited_range) | |
377 | != opt_for_fn (to->decl, flag_cx_limited_range) | |
378 | || opt_for_fn (callee->decl, flag_signed_zeros) | |
379 | != opt_for_fn (to->decl, flag_signed_zeros) | |
380 | || opt_for_fn (callee->decl, flag_associative_math) | |
381 | != opt_for_fn (to->decl, flag_associative_math) | |
382 | || opt_for_fn (callee->decl, flag_reciprocal_math) | |
383 | != opt_for_fn (to->decl, flag_reciprocal_math) | |
0d2f700f JM |
384 | || opt_for_fn (callee->decl, flag_fp_int_builtin_inexact) |
385 | != opt_for_fn (to->decl, flag_fp_int_builtin_inexact) | |
818b88a7 JH |
386 | || opt_for_fn (callee->decl, flag_errno_math) |
387 | != opt_for_fn (to->decl, flag_errno_math)) | |
388 | { | |
389 | struct gcc_options opts = global_options; | |
390 | ||
391 | cl_optimization_restore (&opts, opts_for_fn (to->decl)); | |
392 | opts.x_flag_rounding_math | |
393 | = opt_for_fn (callee->decl, flag_rounding_math); | |
394 | opts.x_flag_trapping_math | |
395 | = opt_for_fn (callee->decl, flag_trapping_math); | |
396 | opts.x_flag_unsafe_math_optimizations | |
397 | = opt_for_fn (callee->decl, flag_unsafe_math_optimizations); | |
398 | opts.x_flag_finite_math_only | |
399 | = opt_for_fn (callee->decl, flag_finite_math_only); | |
400 | opts.x_flag_signaling_nans | |
401 | = opt_for_fn (callee->decl, flag_signaling_nans); | |
402 | opts.x_flag_cx_limited_range | |
403 | = opt_for_fn (callee->decl, flag_cx_limited_range); | |
404 | opts.x_flag_signed_zeros | |
405 | = opt_for_fn (callee->decl, flag_signed_zeros); | |
406 | opts.x_flag_associative_math | |
407 | = opt_for_fn (callee->decl, flag_associative_math); | |
408 | opts.x_flag_reciprocal_math | |
409 | = opt_for_fn (callee->decl, flag_reciprocal_math); | |
0d2f700f JM |
410 | opts.x_flag_fp_int_builtin_inexact |
411 | = opt_for_fn (callee->decl, flag_fp_int_builtin_inexact); | |
818b88a7 JH |
412 | opts.x_flag_errno_math |
413 | = opt_for_fn (callee->decl, flag_errno_math); | |
414 | if (dump_file) | |
415 | fprintf (dump_file, "Copying FP flags from %s:%i to %s:%i\n", | |
416 | callee->name (), callee->order, to->name (), to->order); | |
818b88a7 JH |
417 | DECL_FUNCTION_SPECIFIC_OPTIMIZATION (to->decl) |
418 | = build_optimization_node (&opts); | |
77719b06 | 419 | reload_optimization_node = true; |
818b88a7 JH |
420 | } |
421 | } | |
5058c037 | 422 | |
77719b06 ML |
423 | /* Reload global optimization flags. */ |
424 | if (reload_optimization_node && DECL_STRUCT_FUNCTION (to->decl) == cfun) | |
425 | set_cfun (cfun, true); | |
426 | ||
a5b1779f JH |
427 | /* If aliases are involved, redirect edge to the actual destination and |
428 | possibly remove the aliases. */ | |
429 | if (e->callee != callee) | |
39e2db00 JH |
430 | { |
431 | struct cgraph_node *alias = e->callee, *next_alias; | |
3dafb85c | 432 | e->redirect_callee (callee); |
39e2db00 JH |
433 | while (alias && alias != callee) |
434 | { | |
435 | if (!alias->callers | |
8ccc8042 JH |
436 | && can_remove_node_now_p (alias, |
437 | !e->next_caller && !e->prev_caller ? e : NULL)) | |
39e2db00 | 438 | { |
d52f5295 ML |
439 | next_alias = alias->get_alias_target (); |
440 | alias->remove (); | |
1bbb87c4 JH |
441 | if (callee_removed) |
442 | *callee_removed = true; | |
39e2db00 JH |
443 | alias = next_alias; |
444 | } | |
445 | else | |
446 | break; | |
447 | } | |
448 | } | |
a5b1779f | 449 | |
bd936951 | 450 | clone_inlined_nodes (e, true, update_original, overall_size, e->frequency); |
fee8b6da | 451 | |
fee8b6da | 452 | gcc_assert (curr->callee->global.inlined_to == to); |
898b8927 | 453 | |
0bceb671 JH |
454 | old_size = ipa_fn_summaries->get (to)->size; |
455 | ipa_merge_fn_summary_after_inlining (e); | |
4fd94d1e MJ |
456 | if (e->in_polymorphic_cdtor) |
457 | mark_all_inlined_calls_cdtor (e->callee); | |
bb1e543c | 458 | if (opt_for_fn (e->caller->decl, optimize)) |
0f378cb5 | 459 | new_edges_found = ipa_propagate_indirect_call_infos (curr, new_edges); |
db66bf68 | 460 | check_speculations (e->callee); |
c170d40f | 461 | if (update_overall_summary) |
0bceb671 | 462 | ipa_update_overall_fn_summary (to); |
bddead15 RB |
463 | else |
464 | /* Update self size by the estimate so overall function growth limits | |
465 | work for further inlining into this function. Before inlining | |
466 | the function we inlined to again we expect the caller to update | |
467 | the overall summary. */ | |
0bceb671 JH |
468 | ipa_fn_summaries->get (to)->size += estimated_growth; |
469 | new_size = ipa_fn_summaries->get (to)->size; | |
d250540a | 470 | |
1f26ac87 JM |
471 | if (callee->calls_comdat_local) |
472 | to->calls_comdat_local = true; | |
d52f5295 | 473 | else if (to->calls_comdat_local && callee->comdat_local_p ()) |
1f26ac87 JM |
474 | { |
475 | struct cgraph_edge *se = to->callees; | |
476 | for (; se; se = se->next_callee) | |
d52f5295 | 477 | if (se->inline_failed && se->callee->comdat_local_p ()) |
1f26ac87 JM |
478 | break; |
479 | if (se == NULL) | |
480 | to->calls_comdat_local = false; | |
481 | } | |
482 | ||
f107227b JH |
483 | /* FIXME: This assert suffers from roundoff errors, disable it for GCC 5 |
484 | and revisit it after conversion to sreals in GCC 6. | |
485 | See PR 65654. */ | |
486 | #if 0 | |
0f378cb5 | 487 | /* Verify that estimated growth match real growth. Allow off-by-one |
0bceb671 | 488 | error due to ipa_fn_summary::size_scale roudoff errors. */ |
48b1474e | 489 | gcc_assert (!update_overall_summary || !overall_size || new_edges_found |
0f378cb5 | 490 | || abs (estimated_growth - (new_size - old_size)) <= 1 |
09ce3660 | 491 | || speculation_removed |
0f378cb5 JH |
492 | /* FIXME: a hack. Edges with false predicate are accounted |
493 | wrong, we should remove them from callgraph. */ | |
494 | || predicated); | |
495 | #endif | |
d250540a | 496 | |
8256d5ca JH |
497 | /* Account the change of overall unit size; external functions will be |
498 | removed and are thus not accounted. */ | |
bb1e543c | 499 | if (overall_size && inline_account_function_p (to)) |
fee8b6da JH |
500 | *overall_size += new_size - old_size; |
501 | ncalls_inlined++; | |
502 | ||
0bceb671 | 503 | /* This must happen after ipa_merge_fn_summary_after_inlining that rely on jump |
25837a2f | 504 | functions of callee to not be updated. */ |
0f378cb5 | 505 | return new_edges_found; |
fee8b6da JH |
506 | } |
507 | ||
508 | ||
509 | /* Copy function body of NODE and redirect all inline clones to it. | |
510 | This is done before inline plan is applied to NODE when there are | |
511 | still some inline clones if it. | |
512 | ||
073a8998 | 513 | This is necessary because inline decisions are not really transitive |
fee8b6da JH |
514 | and the other inline clones may have different bodies. */ |
515 | ||
516 | static struct cgraph_node * | |
517 | save_inline_function_body (struct cgraph_node *node) | |
518 | { | |
519 | struct cgraph_node *first_clone, *n; | |
520 | ||
521 | if (dump_file) | |
522 | fprintf (dump_file, "\nSaving body of %s for later reuse\n", | |
fec39fa6 | 523 | node->name ()); |
fee8b6da | 524 | |
d52f5295 | 525 | gcc_assert (node == cgraph_node::get (node->decl)); |
fee8b6da JH |
526 | |
527 | /* first_clone will be turned into real function. */ | |
528 | first_clone = node->clones; | |
ec6a1e35 JH |
529 | |
530 | /* Arrange first clone to not be thunk as those do not have bodies. */ | |
531 | if (first_clone->thunk.thunk_p) | |
532 | { | |
533 | while (first_clone->thunk.thunk_p) | |
534 | first_clone = first_clone->next_sibling_clone; | |
535 | first_clone->prev_sibling_clone->next_sibling_clone | |
536 | = first_clone->next_sibling_clone; | |
537 | if (first_clone->next_sibling_clone) | |
538 | first_clone->next_sibling_clone->prev_sibling_clone | |
539 | = first_clone->prev_sibling_clone; | |
540 | first_clone->next_sibling_clone = node->clones; | |
541 | first_clone->prev_sibling_clone = NULL; | |
542 | node->clones->prev_sibling_clone = first_clone; | |
543 | node->clones = first_clone; | |
544 | } | |
67348ccc | 545 | first_clone->decl = copy_node (node->decl); |
aede2c10 | 546 | first_clone->decl->decl_with_vis.symtab_node = first_clone; |
d52f5295 | 547 | gcc_assert (first_clone == cgraph_node::get (first_clone->decl)); |
fee8b6da JH |
548 | |
549 | /* Now reshape the clone tree, so all other clones descends from | |
550 | first_clone. */ | |
551 | if (first_clone->next_sibling_clone) | |
552 | { | |
ec6a1e35 JH |
553 | for (n = first_clone->next_sibling_clone; n->next_sibling_clone; |
554 | n = n->next_sibling_clone) | |
fee8b6da JH |
555 | n->clone_of = first_clone; |
556 | n->clone_of = first_clone; | |
557 | n->next_sibling_clone = first_clone->clones; | |
558 | if (first_clone->clones) | |
559 | first_clone->clones->prev_sibling_clone = n; | |
560 | first_clone->clones = first_clone->next_sibling_clone; | |
561 | first_clone->next_sibling_clone->prev_sibling_clone = NULL; | |
562 | first_clone->next_sibling_clone = NULL; | |
563 | gcc_assert (!first_clone->prev_sibling_clone); | |
564 | } | |
565 | first_clone->clone_of = NULL; | |
566 | ||
567 | /* Now node in question has no clones. */ | |
568 | node->clones = NULL; | |
569 | ||
1a3118e9 JH |
570 | /* Inline clones share decl with the function they are cloned |
571 | from. Walk the whole clone tree and redirect them all to the | |
572 | new decl. */ | |
fee8b6da JH |
573 | if (first_clone->clones) |
574 | for (n = first_clone->clones; n != first_clone;) | |
575 | { | |
67348ccc DM |
576 | gcc_assert (n->decl == node->decl); |
577 | n->decl = first_clone->decl; | |
fee8b6da JH |
578 | if (n->clones) |
579 | n = n->clones; | |
580 | else if (n->next_sibling_clone) | |
581 | n = n->next_sibling_clone; | |
582 | else | |
583 | { | |
584 | while (n != first_clone && !n->next_sibling_clone) | |
585 | n = n->clone_of; | |
586 | if (n != first_clone) | |
587 | n = n->next_sibling_clone; | |
588 | } | |
589 | } | |
590 | ||
591 | /* Copy the OLD_VERSION_NODE function tree to the new version. */ | |
67348ccc | 592 | tree_function_versioning (node->decl, first_clone->decl, |
9771b263 DN |
593 | NULL, true, NULL, false, |
594 | NULL, NULL); | |
fee8b6da | 595 | |
1a3118e9 JH |
596 | /* The function will be short lived and removed after we inline all the clones, |
597 | but make it internal so we won't confuse ourself. */ | |
67348ccc | 598 | DECL_EXTERNAL (first_clone->decl) = 0; |
67348ccc DM |
599 | TREE_PUBLIC (first_clone->decl) = 0; |
600 | DECL_COMDAT (first_clone->decl) = 0; | |
9771b263 | 601 | first_clone->ipa_transforms_to_apply.release (); |
fee8b6da | 602 | |
b4e93f45 JH |
603 | /* When doing recursive inlining, the clone may become unnecessary. |
604 | This is possible i.e. in the case when the recursive function is proved to be | |
605 | non-throwing and the recursion happens only in the EH landing pad. | |
606 | We can not remove the clone until we are done with saving the body. | |
607 | Remove it now. */ | |
608 | if (!first_clone->callers) | |
609 | { | |
d52f5295 | 610 | first_clone->remove_symbol_and_inline_clones (); |
b4e93f45 JH |
611 | first_clone = NULL; |
612 | } | |
b2b29377 | 613 | else if (flag_checking) |
d52f5295 | 614 | first_clone->verify (); |
b2b29377 | 615 | |
fee8b6da JH |
616 | return first_clone; |
617 | } | |
618 | ||
9c8305f8 JH |
619 | /* Return true when function body of DECL still needs to be kept around |
620 | for later re-use. */ | |
65d630d4 | 621 | static bool |
9c8305f8 JH |
622 | preserve_function_body_p (struct cgraph_node *node) |
623 | { | |
3dafb85c | 624 | gcc_assert (symtab->global_info_ready); |
67348ccc | 625 | gcc_assert (!node->alias && !node->thunk.thunk_p); |
9c8305f8 | 626 | |
ec6a1e35 JH |
627 | /* Look if there is any non-thunk clone around. */ |
628 | for (node = node->clones; node; node = node->next_sibling_clone) | |
629 | if (!node->thunk.thunk_p) | |
630 | return true; | |
9c8305f8 JH |
631 | return false; |
632 | } | |
fee8b6da JH |
633 | |
634 | /* Apply inline plan to function. */ | |
635 | ||
636 | unsigned int | |
637 | inline_transform (struct cgraph_node *node) | |
638 | { | |
639 | unsigned int todo = 0; | |
e8aec975 | 640 | struct cgraph_edge *e, *next; |
2bf86c84 | 641 | bool has_inline = false; |
c9fc06dc | 642 | |
fee8b6da JH |
643 | /* FIXME: Currently the pass manager is adding inline transform more than |
644 | once to some clones. This needs revisiting after WPA cleanups. */ | |
645 | if (cfun->after_inlining) | |
646 | return 0; | |
647 | ||
648 | /* We might need the body of this function so that we can expand | |
649 | it inline somewhere else. */ | |
9c8305f8 | 650 | if (preserve_function_body_p (node)) |
fee8b6da JH |
651 | save_inline_function_body (node); |
652 | ||
e8aec975 JH |
653 | for (e = node->callees; e; e = next) |
654 | { | |
2bf86c84 JH |
655 | if (!e->inline_failed) |
656 | has_inline = true; | |
e8aec975 | 657 | next = e->next_callee; |
3dafb85c | 658 | e->redirect_call_stmt_to_callee (); |
e8aec975 | 659 | } |
d122681a | 660 | node->remove_all_references (); |
c9fc06dc CB |
661 | |
662 | timevar_push (TV_INTEGRATION); | |
bb1e543c | 663 | if (node->callees && (opt_for_fn (node->decl, optimize) || has_inline)) |
55f01229 | 664 | todo = optimize_inline_calls (current_function_decl); |
c9fc06dc CB |
665 | timevar_pop (TV_INTEGRATION); |
666 | ||
f8698b37 RG |
667 | cfun->always_inline_functions_inlined = true; |
668 | cfun->after_inlining = true; | |
669 | todo |= execute_fixup_cfg (); | |
670 | ||
55f01229 RG |
671 | if (!(todo & TODO_update_ssa_any)) |
672 | /* Redirecting edges might lead to a need for vops to be recomputed. */ | |
673 | todo |= TODO_update_ssa_only_virtuals; | |
674 | ||
f8698b37 | 675 | return todo; |
fee8b6da | 676 | } |