]>
Commit | Line | Data |
---|---|---|
8cbc43ff | 1 | /* Callgraph transformations to handle inlining |
fbd26352 | 2 | Copyright (C) 2003-2019 Free Software Foundation, Inc. |
8cbc43ff | 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" | |
7c29e30e | 35 | #include "function.h" |
8cbc43ff | 36 | #include "tree.h" |
7c29e30e | 37 | #include "alloc-pool.h" |
38 | #include "tree-pass.h" | |
39 | #include "cgraph.h" | |
073c1fd5 | 40 | #include "tree-cfg.h" |
2cc80ac3 | 41 | #include "symbol-summary.h" |
25a8e007 | 42 | #include "tree-vrp.h" |
8cbc43ff | 43 | #include "ipa-prop.h" |
b9a58fc5 | 44 | #include "ipa-fnsummary.h" |
8cbc43ff | 45 | #include "ipa-inline.h" |
46 | #include "tree-inline.h" | |
6e4303bf | 47 | #include "function.h" |
48 | #include "cfg.h" | |
49 | #include "basic-block.h" | |
8cbc43ff | 50 | |
51 | int ncalls_inlined; | |
52 | int nfunctions_inlined; | |
53 | ||
151b9ff5 | 54 | /* Scale counts of NODE edges by NUM/DEN. */ |
8cbc43ff | 55 | |
56 | static void | |
151b9ff5 | 57 | update_noncloned_counts (struct cgraph_node *node, |
58 | profile_count num, profile_count den) | |
8cbc43ff | 59 | { |
60 | struct cgraph_edge *e; | |
84895d28 | 61 | |
371858d4 | 62 | profile_count::adjust_for_ipa_scaling (&num, &den); |
8cbc43ff | 63 | |
8cbc43ff | 64 | for (e = node->callees; e; e = e->next_callee) |
65 | { | |
8cbc43ff | 66 | if (!e->inline_failed) |
151b9ff5 | 67 | update_noncloned_counts (e->callee, num, den); |
84895d28 | 68 | e->count = e->count.apply_scale (num, den); |
0835ad03 | 69 | } |
70 | for (e = node->indirect_calls; e; e = e->next_callee) | |
151b9ff5 | 71 | e->count = e->count.apply_scale (num, den); |
84895d28 | 72 | node->count = node->count.apply_scale (num, den); |
8cbc43ff | 73 | } |
74 | ||
82626cb0 | 75 | /* We removed or are going to remove the last call to NODE. |
76 | Return true if we can and want proactively remove the NODE now. | |
77 | This is important to do, since we want inliner to know when offline | |
78 | copy of function was removed. */ | |
79 | ||
80 | static bool | |
7fbf53b8 | 81 | can_remove_node_now_p_1 (struct cgraph_node *node, struct cgraph_edge *e) |
82626cb0 | 82 | { |
7fbf53b8 | 83 | ipa_ref *ref; |
84 | ||
85 | FOR_EACH_ALIAS (node, ref) | |
86 | { | |
87 | cgraph_node *alias = dyn_cast <cgraph_node *> (ref->referring); | |
88 | if ((alias->callers && alias->callers != e) | |
89 | || !can_remove_node_now_p_1 (alias, e)) | |
90 | return false; | |
91 | } | |
82626cb0 | 92 | /* FIXME: When address is taken of DECL_EXTERNAL function we still |
93 | can remove its offline copy, but we would need to keep unanalyzed node in | |
e88fecaf | 94 | the callgraph so references can point to it. |
95 | ||
96 | Also for comdat group we can ignore references inside a group as we | |
97 | want to prove the group as a whole to be dead. */ | |
02774f2d | 98 | return (!node->address_taken |
e88fecaf | 99 | && node->can_remove_if_no_direct_calls_and_refs_p () |
82626cb0 | 100 | /* Inlining might enable more devirtualizing, so we want to remove |
101 | those only after all devirtualizable virtual calls are processed. | |
102 | Lacking may edges in callgraph we just preserve them post | |
103 | inlining. */ | |
7fbf53b8 | 104 | && (!DECL_VIRTUAL_P (node->decl) |
105 | || !opt_for_fn (node->decl, flag_devirtualize)) | |
82626cb0 | 106 | /* During early inlining some unanalyzed cgraph nodes might be in the |
107 | callgraph and they might reffer the function in question. */ | |
347a47cb | 108 | && !cgraph_new_nodes.exists ()); |
82626cb0 | 109 | } |
110 | ||
7791b0eb | 111 | /* We are going to eliminate last direct call to NODE (or alias of it) via edge E. |
112 | Verify that the NODE can be removed from unit and if it is contained in comdat | |
113 | group that the whole comdat group is removable. */ | |
114 | ||
115 | static bool | |
116 | can_remove_node_now_p (struct cgraph_node *node, struct cgraph_edge *e) | |
117 | { | |
118 | struct cgraph_node *next; | |
7fbf53b8 | 119 | if (!can_remove_node_now_p_1 (node, e)) |
7791b0eb | 120 | return false; |
121 | ||
122 | /* When we see same comdat group, we need to be sure that all | |
123 | items can be removed. */ | |
6365c927 | 124 | if (!node->same_comdat_group || !node->externally_visible) |
7791b0eb | 125 | return true; |
415d1b9a | 126 | for (next = dyn_cast<cgraph_node *> (node->same_comdat_group); |
127 | next != node; next = dyn_cast<cgraph_node *> (next->same_comdat_group)) | |
7fbf53b8 | 128 | { |
129 | if (next->alias) | |
130 | continue; | |
131 | if ((next->callers && next->callers != e) | |
132 | || !can_remove_node_now_p_1 (next, e)) | |
133 | return false; | |
134 | } | |
7791b0eb | 135 | return true; |
136 | } | |
137 | ||
0ecf4b03 | 138 | /* Return true if NODE is a master clone with non-inline clones. */ |
139 | ||
140 | static bool | |
141 | master_clone_with_noninline_clones_p (struct cgraph_node *node) | |
142 | { | |
143 | if (node->clone_of) | |
144 | return false; | |
145 | ||
146 | for (struct cgraph_node *n = node->clones; n; n = n->next_sibling_clone) | |
147 | if (n->decl != node->decl) | |
148 | return true; | |
149 | ||
150 | return false; | |
151 | } | |
8cbc43ff | 152 | |
153 | /* E is expected to be an edge being inlined. Clone destination node of | |
154 | the edge and redirect it to the new clone. | |
155 | DUPLICATE is used for bookkeeping on whether we are actually creating new | |
156 | clones or re-using node originally representing out-of-line function call. | |
b8731470 | 157 | By default the offline copy is removed, when it appears dead after inlining. |
158 | UPDATE_ORIGINAL prevents this transformation. | |
159 | If OVERALL_SIZE is non-NULL, the size is updated to reflect the | |
151b9ff5 | 160 | transformation. */ |
8cbc43ff | 161 | |
162 | void | |
163 | clone_inlined_nodes (struct cgraph_edge *e, bool duplicate, | |
151b9ff5 | 164 | bool update_original, int *overall_size) |
8cbc43ff | 165 | { |
48f42a9a | 166 | struct cgraph_node *inlining_into; |
12d5ae9f | 167 | struct cgraph_edge *next; |
48f42a9a | 168 | |
169 | if (e->caller->global.inlined_to) | |
170 | inlining_into = e->caller->global.inlined_to; | |
171 | else | |
172 | inlining_into = e->caller; | |
173 | ||
8cbc43ff | 174 | if (duplicate) |
175 | { | |
176 | /* We may eliminate the need for out-of-line copy to be output. | |
177 | In that case just go ahead and re-use it. This is not just an | |
178 | memory optimization. Making offline copy of fuction disappear | |
179 | from the program will improve future decisions on inlining. */ | |
180 | if (!e->callee->callers->next_caller | |
181 | /* Recursive inlining never wants the master clone to | |
182 | be overwritten. */ | |
183 | && update_original | |
0ecf4b03 | 184 | && can_remove_node_now_p (e->callee, e) |
185 | /* We cannot overwrite a master clone with non-inline clones | |
186 | until after these clones are materialized. */ | |
187 | && !master_clone_with_noninline_clones_p (e->callee)) | |
8cbc43ff | 188 | { |
7791b0eb | 189 | /* TODO: When callee is in a comdat group, we could remove all of it, |
190 | including all inline clones inlined into it. That would however | |
191 | need small function inlining to register edge removal hook to | |
192 | maintain the priority queue. | |
193 | ||
194 | For now we keep the ohter functions in the group in program until | |
195 | cgraph_remove_unreachable_functions gets rid of them. */ | |
8cbc43ff | 196 | gcc_assert (!e->callee->global.inlined_to); |
ab89cd93 | 197 | e->callee->remove_from_same_comdat_group (); |
a6d60179 | 198 | if (e->callee->definition |
199 | && inline_account_function_p (e->callee)) | |
8cbc43ff | 200 | { |
a6d60179 | 201 | gcc_assert (!e->callee->alias); |
8cbc43ff | 202 | if (overall_size) |
d2c2513e | 203 | *overall_size -= ipa_fn_summaries->get (e->callee)->size; |
8cbc43ff | 204 | nfunctions_inlined++; |
205 | } | |
206 | duplicate = false; | |
02774f2d | 207 | e->callee->externally_visible = false; |
151b9ff5 | 208 | update_noncloned_counts (e->callee, e->count, e->callee->count); |
9dc70d59 | 209 | |
210 | dump_callgraph_transformation (e->callee, inlining_into, | |
211 | "inlining to"); | |
8cbc43ff | 212 | } |
213 | else | |
214 | { | |
215 | struct cgraph_node *n; | |
b8731470 | 216 | |
415d1b9a | 217 | n = e->callee->create_clone (e->callee->decl, |
151b9ff5 | 218 | e->count, |
415d1b9a | 219 | update_original, vNULL, true, |
220 | inlining_into, | |
221 | NULL); | |
ca92a251 | 222 | n->used_as_abstract_origin = e->callee->used_as_abstract_origin; |
35ee1c66 | 223 | e->redirect_callee (n); |
8cbc43ff | 224 | } |
225 | } | |
cf951b1a | 226 | else |
ab89cd93 | 227 | e->callee->remove_from_same_comdat_group (); |
8cbc43ff | 228 | |
48f42a9a | 229 | e->callee->global.inlined_to = inlining_into; |
8cbc43ff | 230 | |
231 | /* Recursively clone all bodies. */ | |
12d5ae9f | 232 | for (e = e->callee->callees; e; e = next) |
233 | { | |
234 | next = e->next_callee; | |
235 | if (!e->inline_failed) | |
151b9ff5 | 236 | clone_inlined_nodes (e, duplicate, update_original, overall_size); |
7ab096e0 | 237 | } |
238 | } | |
239 | ||
240 | /* Check all speculations in N and resolve them if they seems useless. */ | |
241 | ||
242 | static bool | |
243 | check_speculations (cgraph_node *n) | |
244 | { | |
245 | bool speculation_removed = false; | |
246 | cgraph_edge *next; | |
247 | ||
248 | for (cgraph_edge *e = n->callees; e; e = next) | |
249 | { | |
250 | next = e->next_callee; | |
12d5ae9f | 251 | if (e->speculative && !speculation_useful_p (e, true)) |
252 | { | |
35ee1c66 | 253 | e->resolve_speculation (NULL); |
12d5ae9f | 254 | speculation_removed = true; |
255 | } | |
7ab096e0 | 256 | else if (!e->inline_failed) |
257 | speculation_removed |= check_speculations (e->callee); | |
12d5ae9f | 258 | } |
7ab096e0 | 259 | return speculation_removed; |
8cbc43ff | 260 | } |
261 | ||
2bec7365 | 262 | /* Mark all call graph edges coming out of NODE and all nodes that have been |
263 | inlined to it as in_polymorphic_cdtor. */ | |
264 | ||
265 | static void | |
266 | mark_all_inlined_calls_cdtor (cgraph_node *node) | |
267 | { | |
268 | for (cgraph_edge *cs = node->callees; cs; cs = cs->next_callee) | |
269 | { | |
270 | cs->in_polymorphic_cdtor = true; | |
271 | if (!cs->inline_failed) | |
6e4303bf | 272 | mark_all_inlined_calls_cdtor (cs->callee); |
2bec7365 | 273 | } |
274 | for (cgraph_edge *cs = node->indirect_calls; cs; cs = cs->next_callee) | |
275 | cs->in_polymorphic_cdtor = true; | |
276 | } | |
277 | ||
8cbc43ff | 278 | |
279 | /* Mark edge E as inlined and update callgraph accordingly. UPDATE_ORIGINAL | |
280 | specify whether profile of original function should be updated. If any new | |
281 | indirect edges are discovered in the process, add them to NEW_EDGES, unless | |
6331b6fa | 282 | it is NULL. If UPDATE_OVERALL_SUMMARY is false, do not bother to recompute overall |
283 | size of caller after inlining. Caller is required to eventually do it via | |
1297cbcd | 284 | ipa_update_overall_fn_summary. |
7c5c01f1 | 285 | If callee_removed is non-NULL, set it to true if we removed callee node. |
6331b6fa | 286 | |
287 | Return true iff any new callgraph edges were discovered as a | |
8cbc43ff | 288 | result of inlining. */ |
289 | ||
290 | bool | |
291 | inline_call (struct cgraph_edge *e, bool update_original, | |
415d1b9a | 292 | vec<cgraph_edge *> *new_edges, |
7c5c01f1 | 293 | int *overall_size, bool update_overall_summary, |
294 | bool *callee_removed) | |
8cbc43ff | 295 | { |
296 | int old_size = 0, new_size = 0; | |
297 | struct cgraph_node *to = NULL; | |
298 | struct cgraph_edge *curr = e; | |
415d1b9a | 299 | struct cgraph_node *callee = e->callee->ultimate_alias_target (); |
18b64b34 | 300 | bool new_edges_found = false; |
301 | ||
6ec9c8c5 | 302 | int estimated_growth = 0; |
303 | if (! update_overall_summary) | |
304 | estimated_growth = estimate_edge_growth (e); | |
6737c56a | 305 | /* This is used only for assert bellow. */ |
306 | #if 0 | |
18b64b34 | 307 | bool predicated = inline_edge_summary (e)->predicate != NULL; |
308 | #endif | |
8cbc43ff | 309 | |
310 | /* Don't inline inlined edges. */ | |
311 | gcc_assert (e->inline_failed); | |
312 | /* Don't even think of inlining inline clone. */ | |
82626cb0 | 313 | gcc_assert (!callee->global.inlined_to); |
8cbc43ff | 314 | |
a41f2a28 | 315 | to = e->caller; |
316 | if (to->global.inlined_to) | |
317 | to = to->global.inlined_to; | |
46729204 | 318 | if (to->thunk.thunk_p) |
319 | { | |
839a6560 | 320 | struct cgraph_node *target = to->callees->callee; |
46729204 | 321 | if (in_lto_p) |
322 | to->get_untransformed_body (); | |
323 | to->expand_thunk (false, true); | |
839a6560 | 324 | /* When thunk is instrumented we may have multiple callees. */ |
325 | for (e = to->callees; e && e->callee != target; e = e->next_callee) | |
326 | ; | |
327 | gcc_assert (e); | |
46729204 | 328 | } |
329 | ||
330 | ||
331 | e->inline_failed = CIF_OK; | |
332 | DECL_POSSIBLY_INLINED (callee->decl) = true; | |
a41f2a28 | 333 | |
e806c56f | 334 | if (DECL_FUNCTION_PERSONALITY (callee->decl)) |
335 | DECL_FUNCTION_PERSONALITY (to->decl) | |
336 | = DECL_FUNCTION_PERSONALITY (callee->decl); | |
67b5f619 | 337 | |
338 | bool reload_optimization_node = false; | |
b1d47f0c | 339 | if (!opt_for_fn (callee->decl, flag_strict_aliasing) |
340 | && opt_for_fn (to->decl, flag_strict_aliasing)) | |
341 | { | |
342 | struct gcc_options opts = global_options; | |
343 | ||
42781674 | 344 | cl_optimization_restore (&opts, opts_for_fn (to->decl)); |
b1d47f0c | 345 | opts.x_flag_strict_aliasing = false; |
346 | if (dump_file) | |
0e388735 | 347 | fprintf (dump_file, "Dropping flag_strict_aliasing on %s\n", |
348 | to->dump_name ()); | |
b1d47f0c | 349 | DECL_FUNCTION_SPECIFIC_OPTIMIZATION (to->decl) |
350 | = build_optimization_node (&opts); | |
67b5f619 | 351 | reload_optimization_node = true; |
b1d47f0c | 352 | } |
9ed0eb00 | 353 | |
d2c2513e | 354 | ipa_fn_summary *caller_info = ipa_fn_summaries->get (to); |
355 | ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee); | |
46e5dccc | 356 | if (!caller_info->fp_expressions && callee_info->fp_expressions) |
357 | { | |
358 | caller_info->fp_expressions = true; | |
359 | if (opt_for_fn (callee->decl, flag_rounding_math) | |
360 | != opt_for_fn (to->decl, flag_rounding_math) | |
361 | || opt_for_fn (callee->decl, flag_trapping_math) | |
362 | != opt_for_fn (to->decl, flag_trapping_math) | |
363 | || opt_for_fn (callee->decl, flag_unsafe_math_optimizations) | |
364 | != opt_for_fn (to->decl, flag_unsafe_math_optimizations) | |
365 | || opt_for_fn (callee->decl, flag_finite_math_only) | |
366 | != opt_for_fn (to->decl, flag_finite_math_only) | |
367 | || opt_for_fn (callee->decl, flag_signaling_nans) | |
368 | != opt_for_fn (to->decl, flag_signaling_nans) | |
369 | || opt_for_fn (callee->decl, flag_cx_limited_range) | |
370 | != opt_for_fn (to->decl, flag_cx_limited_range) | |
371 | || opt_for_fn (callee->decl, flag_signed_zeros) | |
372 | != opt_for_fn (to->decl, flag_signed_zeros) | |
373 | || opt_for_fn (callee->decl, flag_associative_math) | |
374 | != opt_for_fn (to->decl, flag_associative_math) | |
375 | || opt_for_fn (callee->decl, flag_reciprocal_math) | |
376 | != opt_for_fn (to->decl, flag_reciprocal_math) | |
20702ea3 | 377 | || opt_for_fn (callee->decl, flag_fp_int_builtin_inexact) |
378 | != opt_for_fn (to->decl, flag_fp_int_builtin_inexact) | |
46e5dccc | 379 | || opt_for_fn (callee->decl, flag_errno_math) |
380 | != opt_for_fn (to->decl, flag_errno_math)) | |
381 | { | |
382 | struct gcc_options opts = global_options; | |
383 | ||
384 | cl_optimization_restore (&opts, opts_for_fn (to->decl)); | |
385 | opts.x_flag_rounding_math | |
386 | = opt_for_fn (callee->decl, flag_rounding_math); | |
387 | opts.x_flag_trapping_math | |
388 | = opt_for_fn (callee->decl, flag_trapping_math); | |
389 | opts.x_flag_unsafe_math_optimizations | |
390 | = opt_for_fn (callee->decl, flag_unsafe_math_optimizations); | |
391 | opts.x_flag_finite_math_only | |
392 | = opt_for_fn (callee->decl, flag_finite_math_only); | |
393 | opts.x_flag_signaling_nans | |
394 | = opt_for_fn (callee->decl, flag_signaling_nans); | |
395 | opts.x_flag_cx_limited_range | |
396 | = opt_for_fn (callee->decl, flag_cx_limited_range); | |
397 | opts.x_flag_signed_zeros | |
398 | = opt_for_fn (callee->decl, flag_signed_zeros); | |
399 | opts.x_flag_associative_math | |
400 | = opt_for_fn (callee->decl, flag_associative_math); | |
401 | opts.x_flag_reciprocal_math | |
402 | = opt_for_fn (callee->decl, flag_reciprocal_math); | |
20702ea3 | 403 | opts.x_flag_fp_int_builtin_inexact |
404 | = opt_for_fn (callee->decl, flag_fp_int_builtin_inexact); | |
46e5dccc | 405 | opts.x_flag_errno_math |
406 | = opt_for_fn (callee->decl, flag_errno_math); | |
407 | if (dump_file) | |
0e388735 | 408 | fprintf (dump_file, "Copying FP flags from %s to %s\n", |
409 | callee->dump_name (), to->dump_name ()); | |
46e5dccc | 410 | DECL_FUNCTION_SPECIFIC_OPTIMIZATION (to->decl) |
411 | = build_optimization_node (&opts); | |
67b5f619 | 412 | reload_optimization_node = true; |
46e5dccc | 413 | } |
414 | } | |
e806c56f | 415 | |
67b5f619 | 416 | /* Reload global optimization flags. */ |
417 | if (reload_optimization_node && DECL_STRUCT_FUNCTION (to->decl) == cfun) | |
418 | set_cfun (cfun, true); | |
419 | ||
82626cb0 | 420 | /* If aliases are involved, redirect edge to the actual destination and |
421 | possibly remove the aliases. */ | |
422 | if (e->callee != callee) | |
c70f46b0 | 423 | { |
424 | struct cgraph_node *alias = e->callee, *next_alias; | |
35ee1c66 | 425 | e->redirect_callee (callee); |
c70f46b0 | 426 | while (alias && alias != callee) |
427 | { | |
428 | if (!alias->callers | |
6365c927 | 429 | && can_remove_node_now_p (alias, |
430 | !e->next_caller && !e->prev_caller ? e : NULL)) | |
c70f46b0 | 431 | { |
415d1b9a | 432 | next_alias = alias->get_alias_target (); |
433 | alias->remove (); | |
7c5c01f1 | 434 | if (callee_removed) |
435 | *callee_removed = true; | |
c70f46b0 | 436 | alias = next_alias; |
437 | } | |
438 | else | |
439 | break; | |
440 | } | |
441 | } | |
82626cb0 | 442 | |
151b9ff5 | 443 | clone_inlined_nodes (e, true, update_original, overall_size); |
8cbc43ff | 444 | |
8cbc43ff | 445 | gcc_assert (curr->callee->global.inlined_to == to); |
0835ad03 | 446 | |
d2c2513e | 447 | old_size = ipa_fn_summaries->get (to)->size; |
1297cbcd | 448 | ipa_merge_fn_summary_after_inlining (e); |
2bec7365 | 449 | if (e->in_polymorphic_cdtor) |
450 | mark_all_inlined_calls_cdtor (e->callee); | |
a6d60179 | 451 | if (opt_for_fn (e->caller->decl, optimize)) |
18b64b34 | 452 | new_edges_found = ipa_propagate_indirect_call_infos (curr, new_edges); |
7ab096e0 | 453 | check_speculations (e->callee); |
6331b6fa | 454 | if (update_overall_summary) |
1297cbcd | 455 | ipa_update_overall_fn_summary (to); |
6ec9c8c5 | 456 | else |
457 | /* Update self size by the estimate so overall function growth limits | |
458 | work for further inlining into this function. Before inlining | |
459 | the function we inlined to again we expect the caller to update | |
460 | the overall summary. */ | |
d2c2513e | 461 | ipa_fn_summaries->get (to)->size += estimated_growth; |
462 | new_size = ipa_fn_summaries->get (to)->size; | |
12ecd4f9 | 463 | |
468088ac | 464 | if (callee->calls_comdat_local) |
465 | to->calls_comdat_local = true; | |
415d1b9a | 466 | else if (to->calls_comdat_local && callee->comdat_local_p ()) |
468088ac | 467 | { |
468 | struct cgraph_edge *se = to->callees; | |
469 | for (; se; se = se->next_callee) | |
415d1b9a | 470 | if (se->inline_failed && se->callee->comdat_local_p ()) |
468088ac | 471 | break; |
472 | if (se == NULL) | |
473 | to->calls_comdat_local = false; | |
474 | } | |
475 | ||
6737c56a | 476 | /* FIXME: This assert suffers from roundoff errors, disable it for GCC 5 |
477 | and revisit it after conversion to sreals in GCC 6. | |
478 | See PR 65654. */ | |
479 | #if 0 | |
18b64b34 | 480 | /* Verify that estimated growth match real growth. Allow off-by-one |
1297cbcd | 481 | error due to ipa_fn_summary::size_scale roudoff errors. */ |
95fb3203 | 482 | gcc_assert (!update_overall_summary || !overall_size || new_edges_found |
18b64b34 | 483 | || abs (estimated_growth - (new_size - old_size)) <= 1 |
12d5ae9f | 484 | || speculation_removed |
18b64b34 | 485 | /* FIXME: a hack. Edges with false predicate are accounted |
486 | wrong, we should remove them from callgraph. */ | |
487 | || predicated); | |
488 | #endif | |
12ecd4f9 | 489 | |
d94ff7ce | 490 | /* Account the change of overall unit size; external functions will be |
491 | removed and are thus not accounted. */ | |
a6d60179 | 492 | if (overall_size && inline_account_function_p (to)) |
8cbc43ff | 493 | *overall_size += new_size - old_size; |
494 | ncalls_inlined++; | |
495 | ||
1297cbcd | 496 | /* This must happen after ipa_merge_fn_summary_after_inlining that rely on jump |
eb4ae064 | 497 | functions of callee to not be updated. */ |
18b64b34 | 498 | return new_edges_found; |
8cbc43ff | 499 | } |
500 | ||
501 | ||
502 | /* Copy function body of NODE and redirect all inline clones to it. | |
503 | This is done before inline plan is applied to NODE when there are | |
504 | still some inline clones if it. | |
505 | ||
9d75589a | 506 | This is necessary because inline decisions are not really transitive |
8cbc43ff | 507 | and the other inline clones may have different bodies. */ |
508 | ||
509 | static struct cgraph_node * | |
510 | save_inline_function_body (struct cgraph_node *node) | |
511 | { | |
512 | struct cgraph_node *first_clone, *n; | |
513 | ||
514 | if (dump_file) | |
515 | fprintf (dump_file, "\nSaving body of %s for later reuse\n", | |
f1c8b4d7 | 516 | node->name ()); |
8cbc43ff | 517 | |
415d1b9a | 518 | gcc_assert (node == cgraph_node::get (node->decl)); |
8cbc43ff | 519 | |
520 | /* first_clone will be turned into real function. */ | |
521 | first_clone = node->clones; | |
6a353dda | 522 | |
523 | /* Arrange first clone to not be thunk as those do not have bodies. */ | |
524 | if (first_clone->thunk.thunk_p) | |
525 | { | |
526 | while (first_clone->thunk.thunk_p) | |
527 | first_clone = first_clone->next_sibling_clone; | |
528 | first_clone->prev_sibling_clone->next_sibling_clone | |
529 | = first_clone->next_sibling_clone; | |
530 | if (first_clone->next_sibling_clone) | |
531 | first_clone->next_sibling_clone->prev_sibling_clone | |
532 | = first_clone->prev_sibling_clone; | |
533 | first_clone->next_sibling_clone = node->clones; | |
534 | first_clone->prev_sibling_clone = NULL; | |
535 | node->clones->prev_sibling_clone = first_clone; | |
536 | node->clones = first_clone; | |
537 | } | |
02774f2d | 538 | first_clone->decl = copy_node (node->decl); |
8c016392 | 539 | first_clone->decl->decl_with_vis.symtab_node = first_clone; |
415d1b9a | 540 | gcc_assert (first_clone == cgraph_node::get (first_clone->decl)); |
8cbc43ff | 541 | |
542 | /* Now reshape the clone tree, so all other clones descends from | |
543 | first_clone. */ | |
544 | if (first_clone->next_sibling_clone) | |
545 | { | |
6a353dda | 546 | for (n = first_clone->next_sibling_clone; n->next_sibling_clone; |
547 | n = n->next_sibling_clone) | |
8cbc43ff | 548 | n->clone_of = first_clone; |
549 | n->clone_of = first_clone; | |
550 | n->next_sibling_clone = first_clone->clones; | |
551 | if (first_clone->clones) | |
552 | first_clone->clones->prev_sibling_clone = n; | |
553 | first_clone->clones = first_clone->next_sibling_clone; | |
554 | first_clone->next_sibling_clone->prev_sibling_clone = NULL; | |
555 | first_clone->next_sibling_clone = NULL; | |
556 | gcc_assert (!first_clone->prev_sibling_clone); | |
557 | } | |
558 | first_clone->clone_of = NULL; | |
559 | ||
560 | /* Now node in question has no clones. */ | |
561 | node->clones = NULL; | |
562 | ||
d826e131 | 563 | /* Inline clones share decl with the function they are cloned |
564 | from. Walk the whole clone tree and redirect them all to the | |
565 | new decl. */ | |
8cbc43ff | 566 | if (first_clone->clones) |
567 | for (n = first_clone->clones; n != first_clone;) | |
568 | { | |
02774f2d | 569 | gcc_assert (n->decl == node->decl); |
570 | n->decl = first_clone->decl; | |
8cbc43ff | 571 | if (n->clones) |
572 | n = n->clones; | |
573 | else if (n->next_sibling_clone) | |
574 | n = n->next_sibling_clone; | |
575 | else | |
576 | { | |
577 | while (n != first_clone && !n->next_sibling_clone) | |
578 | n = n->clone_of; | |
579 | if (n != first_clone) | |
580 | n = n->next_sibling_clone; | |
581 | } | |
582 | } | |
583 | ||
584 | /* Copy the OLD_VERSION_NODE function tree to the new version. */ | |
02774f2d | 585 | tree_function_versioning (node->decl, first_clone->decl, |
f1f41a6c | 586 | NULL, true, NULL, false, |
587 | NULL, NULL); | |
8cbc43ff | 588 | |
d826e131 | 589 | /* The function will be short lived and removed after we inline all the clones, |
590 | but make it internal so we won't confuse ourself. */ | |
02774f2d | 591 | DECL_EXTERNAL (first_clone->decl) = 0; |
02774f2d | 592 | TREE_PUBLIC (first_clone->decl) = 0; |
593 | DECL_COMDAT (first_clone->decl) = 0; | |
f1f41a6c | 594 | first_clone->ipa_transforms_to_apply.release (); |
8cbc43ff | 595 | |
41710b76 | 596 | /* When doing recursive inlining, the clone may become unnecessary. |
597 | This is possible i.e. in the case when the recursive function is proved to be | |
598 | non-throwing and the recursion happens only in the EH landing pad. | |
f4d3c071 | 599 | We cannot remove the clone until we are done with saving the body. |
41710b76 | 600 | Remove it now. */ |
601 | if (!first_clone->callers) | |
602 | { | |
415d1b9a | 603 | first_clone->remove_symbol_and_inline_clones (); |
41710b76 | 604 | first_clone = NULL; |
605 | } | |
382ecba7 | 606 | else if (flag_checking) |
415d1b9a | 607 | first_clone->verify (); |
382ecba7 | 608 | |
8cbc43ff | 609 | return first_clone; |
610 | } | |
611 | ||
da5e1e7c | 612 | /* Return true when function body of DECL still needs to be kept around |
613 | for later re-use. */ | |
cf951b1a | 614 | static bool |
da5e1e7c | 615 | preserve_function_body_p (struct cgraph_node *node) |
616 | { | |
35ee1c66 | 617 | gcc_assert (symtab->global_info_ready); |
02774f2d | 618 | gcc_assert (!node->alias && !node->thunk.thunk_p); |
da5e1e7c | 619 | |
6a353dda | 620 | /* Look if there is any non-thunk clone around. */ |
621 | for (node = node->clones; node; node = node->next_sibling_clone) | |
622 | if (!node->thunk.thunk_p) | |
623 | return true; | |
da5e1e7c | 624 | return false; |
625 | } | |
8cbc43ff | 626 | |
627 | /* Apply inline plan to function. */ | |
628 | ||
629 | unsigned int | |
630 | inline_transform (struct cgraph_node *node) | |
631 | { | |
632 | unsigned int todo = 0; | |
dc3d1030 | 633 | struct cgraph_edge *e, *next; |
d1f68cd8 | 634 | bool has_inline = false; |
a522e9eb | 635 | |
8cbc43ff | 636 | /* FIXME: Currently the pass manager is adding inline transform more than |
637 | once to some clones. This needs revisiting after WPA cleanups. */ | |
638 | if (cfun->after_inlining) | |
639 | return 0; | |
640 | ||
641 | /* We might need the body of this function so that we can expand | |
642 | it inline somewhere else. */ | |
da5e1e7c | 643 | if (preserve_function_body_p (node)) |
8cbc43ff | 644 | save_inline_function_body (node); |
645 | ||
dc3d1030 | 646 | for (e = node->callees; e; e = next) |
647 | { | |
d1f68cd8 | 648 | if (!e->inline_failed) |
649 | has_inline = true; | |
dc3d1030 | 650 | next = e->next_callee; |
35ee1c66 | 651 | e->redirect_call_stmt_to_callee (); |
dc3d1030 | 652 | } |
51ce5652 | 653 | node->remove_all_references (); |
a522e9eb | 654 | |
655 | timevar_push (TV_INTEGRATION); | |
a6d60179 | 656 | if (node->callees && (opt_for_fn (node->decl, optimize) || has_inline)) |
6e4303bf | 657 | { |
658 | profile_count num = node->count; | |
659 | profile_count den = ENTRY_BLOCK_PTR_FOR_FN (cfun)->count; | |
ed0831a9 | 660 | bool scale = num.initialized_p () && !(num == den); |
6e4303bf | 661 | if (scale) |
662 | { | |
ed0831a9 | 663 | profile_count::adjust_for_ipa_scaling (&num, &den); |
6e4303bf | 664 | if (dump_file) |
665 | { | |
666 | fprintf (dump_file, "Applying count scale "); | |
667 | num.dump (dump_file); | |
668 | fprintf (dump_file, "/"); | |
669 | den.dump (dump_file); | |
670 | fprintf (dump_file, "\n"); | |
671 | } | |
672 | ||
673 | basic_block bb; | |
ed0831a9 | 674 | cfun->cfg->count_max = profile_count::uninitialized (); |
6e4303bf | 675 | FOR_ALL_BB_FN (bb, cfun) |
ed0831a9 | 676 | { |
84895d28 | 677 | bb->count = bb->count.apply_scale (num, den); |
ed0831a9 | 678 | cfun->cfg->count_max = cfun->cfg->count_max.max (bb->count); |
679 | } | |
6e4303bf | 680 | ENTRY_BLOCK_PTR_FOR_FN (cfun)->count = node->count; |
681 | } | |
682 | todo = optimize_inline_calls (current_function_decl); | |
683 | } | |
a522e9eb | 684 | timevar_pop (TV_INTEGRATION); |
685 | ||
e723655c | 686 | cfun->always_inline_functions_inlined = true; |
687 | cfun->after_inlining = true; | |
688 | todo |= execute_fixup_cfg (); | |
689 | ||
4b0e0420 | 690 | if (!(todo & TODO_update_ssa_any)) |
691 | /* Redirecting edges might lead to a need for vops to be recomputed. */ | |
692 | todo |= TODO_update_ssa_only_virtuals; | |
693 | ||
e723655c | 694 | return todo; |
8cbc43ff | 695 | } |