]>
Commit | Line | Data |
---|---|---|
fee8b6da | 1 | /* Callgraph transformations to handle inlining |
a5544970 | 2 | Copyright (C) 2003-2019 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" | |
09fcc0c0 JH |
47 | #include "function.h" |
48 | #include "cfg.h" | |
49 | #include "basic-block.h" | |
fee8b6da JH |
50 | |
51 | int ncalls_inlined; | |
52 | int nfunctions_inlined; | |
53 | ||
1bad9c18 | 54 | /* Scale counts of NODE edges by NUM/DEN. */ |
fee8b6da JH |
55 | |
56 | static void | |
1bad9c18 JH |
57 | update_noncloned_counts (struct cgraph_node *node, |
58 | profile_count num, profile_count den) | |
fee8b6da JH |
59 | { |
60 | struct cgraph_edge *e; | |
8e7d1486 | 61 | |
e4373d41 | 62 | profile_count::adjust_for_ipa_scaling (&num, &den); |
fee8b6da | 63 | |
fee8b6da JH |
64 | for (e = node->callees; e; e = e->next_callee) |
65 | { | |
fee8b6da | 66 | if (!e->inline_failed) |
1bad9c18 | 67 | update_noncloned_counts (e->callee, num, den); |
8e7d1486 | 68 | e->count = e->count.apply_scale (num, den); |
898b8927 JH |
69 | } |
70 | for (e = node->indirect_calls; e; e = e->next_callee) | |
1bad9c18 | 71 | e->count = e->count.apply_scale (num, den); |
8e7d1486 | 72 | node->count = node->count.apply_scale (num, den); |
fee8b6da JH |
73 | } |
74 | ||
a5b1779f JH |
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 | |
d142079a | 81 | can_remove_node_now_p_1 (struct cgraph_node *node, struct cgraph_edge *e) |
a5b1779f | 82 | { |
d142079a JH |
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 | } | |
a5b1779f JH |
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 | |
a6a543bf JH |
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. */ | |
67348ccc | 98 | return (!node->address_taken |
a6a543bf | 99 | && node->can_remove_if_no_direct_calls_and_refs_p () |
a5b1779f JH |
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. */ | |
d142079a JH |
104 | && (!DECL_VIRTUAL_P (node->decl) |
105 | || !opt_for_fn (node->decl, flag_devirtualize)) | |
a5b1779f JH |
106 | /* During early inlining some unanalyzed cgraph nodes might be in the |
107 | callgraph and they might reffer the function in question. */ | |
31acf1bb | 108 | && !cgraph_new_nodes.exists ()); |
a5b1779f JH |
109 | } |
110 | ||
6c69a029 JH |
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; | |
d142079a | 119 | if (!can_remove_node_now_p_1 (node, e)) |
6c69a029 JH |
120 | return false; |
121 | ||
122 | /* When we see same comdat group, we need to be sure that all | |
123 | items can be removed. */ | |
8ccc8042 | 124 | if (!node->same_comdat_group || !node->externally_visible) |
6c69a029 | 125 | return true; |
d52f5295 ML |
126 | for (next = dyn_cast<cgraph_node *> (node->same_comdat_group); |
127 | next != node; next = dyn_cast<cgraph_node *> (next->same_comdat_group)) | |
d142079a JH |
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 | } | |
6c69a029 JH |
135 | return true; |
136 | } | |
137 | ||
d83fa499 EB |
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 | } | |
fee8b6da JH |
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. | |
bd936951 JH |
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 | |
1bad9c18 | 160 | transformation. */ |
fee8b6da JH |
161 | |
162 | void | |
163 | clone_inlined_nodes (struct cgraph_edge *e, bool duplicate, | |
1bad9c18 | 164 | bool update_original, int *overall_size) |
fee8b6da | 165 | { |
44a60244 | 166 | struct cgraph_node *inlining_into; |
09ce3660 | 167 | struct cgraph_edge *next; |
44a60244 MJ |
168 | |
169 | if (e->caller->global.inlined_to) | |
170 | inlining_into = e->caller->global.inlined_to; | |
171 | else | |
172 | inlining_into = e->caller; | |
173 | ||
fee8b6da JH |
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 | |
d83fa499 EB |
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)) | |
fee8b6da | 188 | { |
6c69a029 JH |
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. */ | |
fee8b6da | 196 | gcc_assert (!e->callee->global.inlined_to); |
b91b562c | 197 | e->callee->remove_from_same_comdat_group (); |
bb1e543c JH |
198 | if (e->callee->definition |
199 | && inline_account_function_p (e->callee)) | |
fee8b6da | 200 | { |
bb1e543c | 201 | gcc_assert (!e->callee->alias); |
fee8b6da | 202 | if (overall_size) |
f658ad30 | 203 | *overall_size -= ipa_size_summaries->get (e->callee)->size; |
fee8b6da JH |
204 | nfunctions_inlined++; |
205 | } | |
206 | duplicate = false; | |
67348ccc | 207 | e->callee->externally_visible = false; |
1bad9c18 | 208 | update_noncloned_counts (e->callee, e->count, e->callee->count); |
0bdad123 ML |
209 | |
210 | dump_callgraph_transformation (e->callee, inlining_into, | |
211 | "inlining to"); | |
fee8b6da JH |
212 | } |
213 | else | |
214 | { | |
215 | struct cgraph_node *n; | |
bd936951 | 216 | |
d52f5295 | 217 | n = e->callee->create_clone (e->callee->decl, |
1bad9c18 | 218 | e->count, |
d52f5295 ML |
219 | update_original, vNULL, true, |
220 | inlining_into, | |
221 | NULL); | |
4ad08ee8 | 222 | n->used_as_abstract_origin = e->callee->used_as_abstract_origin; |
3dafb85c | 223 | e->redirect_callee (n); |
fee8b6da JH |
224 | } |
225 | } | |
65d630d4 | 226 | else |
b91b562c | 227 | e->callee->remove_from_same_comdat_group (); |
fee8b6da | 228 | |
44a60244 | 229 | e->callee->global.inlined_to = inlining_into; |
fee8b6da JH |
230 | |
231 | /* Recursively clone all bodies. */ | |
09ce3660 JH |
232 | for (e = e->callee->callees; e; e = next) |
233 | { | |
234 | next = e->next_callee; | |
235 | if (!e->inline_failed) | |
1bad9c18 | 236 | clone_inlined_nodes (e, duplicate, update_original, overall_size); |
db66bf68 JH |
237 | } |
238 | } | |
239 | ||
4517b378 MJ |
240 | /* Check all speculations in N and if any seem useless, resolve them. When a |
241 | first edge is resolved, pop all edges from NEW_EDGES and insert them to | |
242 | EDGE_SET. Then remove each resolved edge from EDGE_SET, if it is there. */ | |
db66bf68 JH |
243 | |
244 | static bool | |
4517b378 MJ |
245 | check_speculations_1 (cgraph_node *n, vec<cgraph_edge *> *new_edges, |
246 | hash_set <cgraph_edge *> *edge_set) | |
db66bf68 JH |
247 | { |
248 | bool speculation_removed = false; | |
249 | cgraph_edge *next; | |
250 | ||
251 | for (cgraph_edge *e = n->callees; e; e = next) | |
252 | { | |
253 | next = e->next_callee; | |
09ce3660 JH |
254 | if (e->speculative && !speculation_useful_p (e, true)) |
255 | { | |
4517b378 MJ |
256 | while (new_edges && !new_edges->is_empty ()) |
257 | edge_set->add (new_edges->pop ()); | |
258 | edge_set->remove (e); | |
259 | ||
3dafb85c | 260 | e->resolve_speculation (NULL); |
09ce3660 JH |
261 | speculation_removed = true; |
262 | } | |
db66bf68 | 263 | else if (!e->inline_failed) |
4517b378 MJ |
264 | speculation_removed |= check_speculations_1 (e->callee, new_edges, |
265 | edge_set); | |
09ce3660 | 266 | } |
db66bf68 | 267 | return speculation_removed; |
fee8b6da JH |
268 | } |
269 | ||
4517b378 MJ |
270 | /* Push E to NEW_EDGES. Called from hash_set traverse method, which |
271 | unfortunately means this function has to have external linkage, otherwise | |
272 | the code will not compile with gcc 4.8. */ | |
273 | ||
274 | bool | |
275 | push_all_edges_in_set_to_vec (cgraph_edge * const &e, | |
276 | vec<cgraph_edge *> *new_edges) | |
277 | { | |
278 | new_edges->safe_push (e); | |
279 | return true; | |
280 | } | |
281 | ||
282 | /* Check all speculations in N and if any seem useless, resolve them and remove | |
283 | them from NEW_EDGES. */ | |
284 | ||
285 | static bool | |
286 | check_speculations (cgraph_node *n, vec<cgraph_edge *> *new_edges) | |
287 | { | |
288 | hash_set <cgraph_edge *> edge_set; | |
289 | bool res = check_speculations_1 (n, new_edges, &edge_set); | |
290 | if (!edge_set.is_empty ()) | |
291 | edge_set.traverse <vec<cgraph_edge *> *, | |
292 | push_all_edges_in_set_to_vec> (new_edges); | |
293 | return res; | |
294 | } | |
295 | ||
4fd94d1e MJ |
296 | /* Mark all call graph edges coming out of NODE and all nodes that have been |
297 | inlined to it as in_polymorphic_cdtor. */ | |
298 | ||
299 | static void | |
300 | mark_all_inlined_calls_cdtor (cgraph_node *node) | |
301 | { | |
302 | for (cgraph_edge *cs = node->callees; cs; cs = cs->next_callee) | |
303 | { | |
304 | cs->in_polymorphic_cdtor = true; | |
305 | if (!cs->inline_failed) | |
09fcc0c0 | 306 | mark_all_inlined_calls_cdtor (cs->callee); |
4fd94d1e MJ |
307 | } |
308 | for (cgraph_edge *cs = node->indirect_calls; cs; cs = cs->next_callee) | |
309 | cs->in_polymorphic_cdtor = true; | |
310 | } | |
311 | ||
fee8b6da JH |
312 | |
313 | /* Mark edge E as inlined and update callgraph accordingly. UPDATE_ORIGINAL | |
314 | specify whether profile of original function should be updated. If any new | |
315 | indirect edges are discovered in the process, add them to NEW_EDGES, unless | |
c170d40f JH |
316 | it is NULL. If UPDATE_OVERALL_SUMMARY is false, do not bother to recompute overall |
317 | size of caller after inlining. Caller is required to eventually do it via | |
0bceb671 | 318 | ipa_update_overall_fn_summary. |
1bbb87c4 | 319 | If callee_removed is non-NULL, set it to true if we removed callee node. |
c170d40f JH |
320 | |
321 | Return true iff any new callgraph edges were discovered as a | |
fee8b6da JH |
322 | result of inlining. */ |
323 | ||
324 | bool | |
325 | inline_call (struct cgraph_edge *e, bool update_original, | |
d52f5295 | 326 | vec<cgraph_edge *> *new_edges, |
1bbb87c4 JH |
327 | int *overall_size, bool update_overall_summary, |
328 | bool *callee_removed) | |
fee8b6da JH |
329 | { |
330 | int old_size = 0, new_size = 0; | |
331 | struct cgraph_node *to = NULL; | |
332 | struct cgraph_edge *curr = e; | |
d52f5295 | 333 | struct cgraph_node *callee = e->callee->ultimate_alias_target (); |
0f378cb5 JH |
334 | bool new_edges_found = false; |
335 | ||
bddead15 RB |
336 | int estimated_growth = 0; |
337 | if (! update_overall_summary) | |
338 | estimated_growth = estimate_edge_growth (e); | |
f107227b JH |
339 | /* This is used only for assert bellow. */ |
340 | #if 0 | |
0f378cb5 JH |
341 | bool predicated = inline_edge_summary (e)->predicate != NULL; |
342 | #endif | |
fee8b6da JH |
343 | |
344 | /* Don't inline inlined edges. */ | |
345 | gcc_assert (e->inline_failed); | |
346 | /* Don't even think of inlining inline clone. */ | |
a5b1779f | 347 | gcc_assert (!callee->global.inlined_to); |
fee8b6da | 348 | |
632b4f8e JH |
349 | to = e->caller; |
350 | if (to->global.inlined_to) | |
351 | to = to->global.inlined_to; | |
0b9004ed JH |
352 | if (to->thunk.thunk_p) |
353 | { | |
c7ed8938 | 354 | struct cgraph_node *target = to->callees->callee; |
a088d7b1 | 355 | symtab->call_cgraph_removal_hooks (to); |
0b9004ed JH |
356 | if (in_lto_p) |
357 | to->get_untransformed_body (); | |
358 | to->expand_thunk (false, true); | |
c7ed8938 IE |
359 | /* When thunk is instrumented we may have multiple callees. */ |
360 | for (e = to->callees; e && e->callee != target; e = e->next_callee) | |
361 | ; | |
a088d7b1 | 362 | symtab->call_cgraph_insertion_hooks (to); |
c7ed8938 | 363 | gcc_assert (e); |
0b9004ed JH |
364 | } |
365 | ||
366 | ||
367 | e->inline_failed = CIF_OK; | |
368 | DECL_POSSIBLY_INLINED (callee->decl) = true; | |
632b4f8e | 369 | |
5058c037 JH |
370 | if (DECL_FUNCTION_PERSONALITY (callee->decl)) |
371 | DECL_FUNCTION_PERSONALITY (to->decl) | |
372 | = DECL_FUNCTION_PERSONALITY (callee->decl); | |
77719b06 ML |
373 | |
374 | bool reload_optimization_node = false; | |
45285060 JH |
375 | if (!opt_for_fn (callee->decl, flag_strict_aliasing) |
376 | && opt_for_fn (to->decl, flag_strict_aliasing)) | |
377 | { | |
378 | struct gcc_options opts = global_options; | |
379 | ||
f7f32acd | 380 | cl_optimization_restore (&opts, opts_for_fn (to->decl)); |
45285060 JH |
381 | opts.x_flag_strict_aliasing = false; |
382 | if (dump_file) | |
464d0118 ML |
383 | fprintf (dump_file, "Dropping flag_strict_aliasing on %s\n", |
384 | to->dump_name ()); | |
45285060 JH |
385 | DECL_FUNCTION_SPECIFIC_OPTIMIZATION (to->decl) |
386 | = build_optimization_node (&opts); | |
77719b06 | 387 | reload_optimization_node = true; |
45285060 | 388 | } |
5c846a81 | 389 | |
56f62793 ML |
390 | ipa_fn_summary *caller_info = ipa_fn_summaries->get (to); |
391 | ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee); | |
818b88a7 JH |
392 | if (!caller_info->fp_expressions && callee_info->fp_expressions) |
393 | { | |
394 | caller_info->fp_expressions = true; | |
395 | if (opt_for_fn (callee->decl, flag_rounding_math) | |
396 | != opt_for_fn (to->decl, flag_rounding_math) | |
397 | || opt_for_fn (callee->decl, flag_trapping_math) | |
398 | != opt_for_fn (to->decl, flag_trapping_math) | |
399 | || opt_for_fn (callee->decl, flag_unsafe_math_optimizations) | |
400 | != opt_for_fn (to->decl, flag_unsafe_math_optimizations) | |
401 | || opt_for_fn (callee->decl, flag_finite_math_only) | |
402 | != opt_for_fn (to->decl, flag_finite_math_only) | |
403 | || opt_for_fn (callee->decl, flag_signaling_nans) | |
404 | != opt_for_fn (to->decl, flag_signaling_nans) | |
405 | || opt_for_fn (callee->decl, flag_cx_limited_range) | |
406 | != opt_for_fn (to->decl, flag_cx_limited_range) | |
407 | || opt_for_fn (callee->decl, flag_signed_zeros) | |
408 | != opt_for_fn (to->decl, flag_signed_zeros) | |
409 | || opt_for_fn (callee->decl, flag_associative_math) | |
410 | != opt_for_fn (to->decl, flag_associative_math) | |
411 | || opt_for_fn (callee->decl, flag_reciprocal_math) | |
412 | != opt_for_fn (to->decl, flag_reciprocal_math) | |
0d2f700f JM |
413 | || opt_for_fn (callee->decl, flag_fp_int_builtin_inexact) |
414 | != opt_for_fn (to->decl, flag_fp_int_builtin_inexact) | |
818b88a7 JH |
415 | || opt_for_fn (callee->decl, flag_errno_math) |
416 | != opt_for_fn (to->decl, flag_errno_math)) | |
417 | { | |
418 | struct gcc_options opts = global_options; | |
419 | ||
420 | cl_optimization_restore (&opts, opts_for_fn (to->decl)); | |
421 | opts.x_flag_rounding_math | |
422 | = opt_for_fn (callee->decl, flag_rounding_math); | |
423 | opts.x_flag_trapping_math | |
424 | = opt_for_fn (callee->decl, flag_trapping_math); | |
425 | opts.x_flag_unsafe_math_optimizations | |
426 | = opt_for_fn (callee->decl, flag_unsafe_math_optimizations); | |
427 | opts.x_flag_finite_math_only | |
428 | = opt_for_fn (callee->decl, flag_finite_math_only); | |
429 | opts.x_flag_signaling_nans | |
430 | = opt_for_fn (callee->decl, flag_signaling_nans); | |
431 | opts.x_flag_cx_limited_range | |
432 | = opt_for_fn (callee->decl, flag_cx_limited_range); | |
433 | opts.x_flag_signed_zeros | |
434 | = opt_for_fn (callee->decl, flag_signed_zeros); | |
435 | opts.x_flag_associative_math | |
436 | = opt_for_fn (callee->decl, flag_associative_math); | |
437 | opts.x_flag_reciprocal_math | |
438 | = opt_for_fn (callee->decl, flag_reciprocal_math); | |
0d2f700f JM |
439 | opts.x_flag_fp_int_builtin_inexact |
440 | = opt_for_fn (callee->decl, flag_fp_int_builtin_inexact); | |
818b88a7 JH |
441 | opts.x_flag_errno_math |
442 | = opt_for_fn (callee->decl, flag_errno_math); | |
443 | if (dump_file) | |
464d0118 ML |
444 | fprintf (dump_file, "Copying FP flags from %s to %s\n", |
445 | callee->dump_name (), to->dump_name ()); | |
818b88a7 JH |
446 | DECL_FUNCTION_SPECIFIC_OPTIMIZATION (to->decl) |
447 | = build_optimization_node (&opts); | |
77719b06 | 448 | reload_optimization_node = true; |
818b88a7 JH |
449 | } |
450 | } | |
5058c037 | 451 | |
77719b06 ML |
452 | /* Reload global optimization flags. */ |
453 | if (reload_optimization_node && DECL_STRUCT_FUNCTION (to->decl) == cfun) | |
454 | set_cfun (cfun, true); | |
455 | ||
a5b1779f JH |
456 | /* If aliases are involved, redirect edge to the actual destination and |
457 | possibly remove the aliases. */ | |
458 | if (e->callee != callee) | |
39e2db00 JH |
459 | { |
460 | struct cgraph_node *alias = e->callee, *next_alias; | |
3dafb85c | 461 | e->redirect_callee (callee); |
39e2db00 JH |
462 | while (alias && alias != callee) |
463 | { | |
464 | if (!alias->callers | |
8ccc8042 JH |
465 | && can_remove_node_now_p (alias, |
466 | !e->next_caller && !e->prev_caller ? e : NULL)) | |
39e2db00 | 467 | { |
d52f5295 ML |
468 | next_alias = alias->get_alias_target (); |
469 | alias->remove (); | |
1bbb87c4 JH |
470 | if (callee_removed) |
471 | *callee_removed = true; | |
39e2db00 JH |
472 | alias = next_alias; |
473 | } | |
474 | else | |
475 | break; | |
476 | } | |
477 | } | |
a5b1779f | 478 | |
1bad9c18 | 479 | clone_inlined_nodes (e, true, update_original, overall_size); |
fee8b6da | 480 | |
fee8b6da | 481 | gcc_assert (curr->callee->global.inlined_to == to); |
898b8927 | 482 | |
f658ad30 | 483 | old_size = ipa_size_summaries->get (to)->size; |
0bceb671 | 484 | ipa_merge_fn_summary_after_inlining (e); |
4fd94d1e MJ |
485 | if (e->in_polymorphic_cdtor) |
486 | mark_all_inlined_calls_cdtor (e->callee); | |
bb1e543c | 487 | if (opt_for_fn (e->caller->decl, optimize)) |
0f378cb5 | 488 | new_edges_found = ipa_propagate_indirect_call_infos (curr, new_edges); |
4517b378 | 489 | check_speculations (e->callee, new_edges); |
c170d40f | 490 | if (update_overall_summary) |
0bceb671 | 491 | ipa_update_overall_fn_summary (to); |
bddead15 RB |
492 | else |
493 | /* Update self size by the estimate so overall function growth limits | |
494 | work for further inlining into this function. Before inlining | |
495 | the function we inlined to again we expect the caller to update | |
496 | the overall summary. */ | |
f658ad30 JH |
497 | ipa_size_summaries->get (to)->size += estimated_growth; |
498 | new_size = ipa_size_summaries->get (to)->size; | |
d250540a | 499 | |
1f26ac87 JM |
500 | if (callee->calls_comdat_local) |
501 | to->calls_comdat_local = true; | |
d52f5295 | 502 | else if (to->calls_comdat_local && callee->comdat_local_p ()) |
1f26ac87 JM |
503 | { |
504 | struct cgraph_edge *se = to->callees; | |
505 | for (; se; se = se->next_callee) | |
d52f5295 | 506 | if (se->inline_failed && se->callee->comdat_local_p ()) |
1f26ac87 JM |
507 | break; |
508 | if (se == NULL) | |
509 | to->calls_comdat_local = false; | |
510 | } | |
511 | ||
f107227b JH |
512 | /* FIXME: This assert suffers from roundoff errors, disable it for GCC 5 |
513 | and revisit it after conversion to sreals in GCC 6. | |
514 | See PR 65654. */ | |
515 | #if 0 | |
0f378cb5 | 516 | /* Verify that estimated growth match real growth. Allow off-by-one |
0bceb671 | 517 | error due to ipa_fn_summary::size_scale roudoff errors. */ |
48b1474e | 518 | gcc_assert (!update_overall_summary || !overall_size || new_edges_found |
0f378cb5 | 519 | || abs (estimated_growth - (new_size - old_size)) <= 1 |
09ce3660 | 520 | || speculation_removed |
0f378cb5 JH |
521 | /* FIXME: a hack. Edges with false predicate are accounted |
522 | wrong, we should remove them from callgraph. */ | |
523 | || predicated); | |
524 | #endif | |
d250540a | 525 | |
8256d5ca JH |
526 | /* Account the change of overall unit size; external functions will be |
527 | removed and are thus not accounted. */ | |
bb1e543c | 528 | if (overall_size && inline_account_function_p (to)) |
fee8b6da JH |
529 | *overall_size += new_size - old_size; |
530 | ncalls_inlined++; | |
531 | ||
0bceb671 | 532 | /* This must happen after ipa_merge_fn_summary_after_inlining that rely on jump |
25837a2f | 533 | functions of callee to not be updated. */ |
0f378cb5 | 534 | return new_edges_found; |
fee8b6da JH |
535 | } |
536 | ||
537 | ||
538 | /* Copy function body of NODE and redirect all inline clones to it. | |
539 | This is done before inline plan is applied to NODE when there are | |
540 | still some inline clones if it. | |
541 | ||
073a8998 | 542 | This is necessary because inline decisions are not really transitive |
fee8b6da JH |
543 | and the other inline clones may have different bodies. */ |
544 | ||
545 | static struct cgraph_node * | |
546 | save_inline_function_body (struct cgraph_node *node) | |
547 | { | |
548 | struct cgraph_node *first_clone, *n; | |
549 | ||
550 | if (dump_file) | |
551 | fprintf (dump_file, "\nSaving body of %s for later reuse\n", | |
fec39fa6 | 552 | node->name ()); |
fee8b6da | 553 | |
d52f5295 | 554 | gcc_assert (node == cgraph_node::get (node->decl)); |
fee8b6da JH |
555 | |
556 | /* first_clone will be turned into real function. */ | |
557 | first_clone = node->clones; | |
ec6a1e35 JH |
558 | |
559 | /* Arrange first clone to not be thunk as those do not have bodies. */ | |
560 | if (first_clone->thunk.thunk_p) | |
561 | { | |
562 | while (first_clone->thunk.thunk_p) | |
563 | first_clone = first_clone->next_sibling_clone; | |
564 | first_clone->prev_sibling_clone->next_sibling_clone | |
565 | = first_clone->next_sibling_clone; | |
566 | if (first_clone->next_sibling_clone) | |
567 | first_clone->next_sibling_clone->prev_sibling_clone | |
568 | = first_clone->prev_sibling_clone; | |
569 | first_clone->next_sibling_clone = node->clones; | |
570 | first_clone->prev_sibling_clone = NULL; | |
571 | node->clones->prev_sibling_clone = first_clone; | |
572 | node->clones = first_clone; | |
573 | } | |
67348ccc | 574 | first_clone->decl = copy_node (node->decl); |
aede2c10 | 575 | first_clone->decl->decl_with_vis.symtab_node = first_clone; |
d52f5295 | 576 | gcc_assert (first_clone == cgraph_node::get (first_clone->decl)); |
fee8b6da JH |
577 | |
578 | /* Now reshape the clone tree, so all other clones descends from | |
579 | first_clone. */ | |
580 | if (first_clone->next_sibling_clone) | |
581 | { | |
ec6a1e35 JH |
582 | for (n = first_clone->next_sibling_clone; n->next_sibling_clone; |
583 | n = n->next_sibling_clone) | |
fee8b6da JH |
584 | n->clone_of = first_clone; |
585 | n->clone_of = first_clone; | |
586 | n->next_sibling_clone = first_clone->clones; | |
587 | if (first_clone->clones) | |
588 | first_clone->clones->prev_sibling_clone = n; | |
589 | first_clone->clones = first_clone->next_sibling_clone; | |
590 | first_clone->next_sibling_clone->prev_sibling_clone = NULL; | |
591 | first_clone->next_sibling_clone = NULL; | |
592 | gcc_assert (!first_clone->prev_sibling_clone); | |
593 | } | |
594 | first_clone->clone_of = NULL; | |
595 | ||
596 | /* Now node in question has no clones. */ | |
597 | node->clones = NULL; | |
598 | ||
1a3118e9 JH |
599 | /* Inline clones share decl with the function they are cloned |
600 | from. Walk the whole clone tree and redirect them all to the | |
601 | new decl. */ | |
fee8b6da JH |
602 | if (first_clone->clones) |
603 | for (n = first_clone->clones; n != first_clone;) | |
604 | { | |
67348ccc DM |
605 | gcc_assert (n->decl == node->decl); |
606 | n->decl = first_clone->decl; | |
fee8b6da JH |
607 | if (n->clones) |
608 | n = n->clones; | |
609 | else if (n->next_sibling_clone) | |
610 | n = n->next_sibling_clone; | |
611 | else | |
612 | { | |
613 | while (n != first_clone && !n->next_sibling_clone) | |
614 | n = n->clone_of; | |
615 | if (n != first_clone) | |
616 | n = n->next_sibling_clone; | |
617 | } | |
618 | } | |
619 | ||
620 | /* Copy the OLD_VERSION_NODE function tree to the new version. */ | |
67348ccc | 621 | tree_function_versioning (node->decl, first_clone->decl, |
ff6686d2 | 622 | NULL, NULL, true, NULL, NULL); |
fee8b6da | 623 | |
1a3118e9 JH |
624 | /* The function will be short lived and removed after we inline all the clones, |
625 | but make it internal so we won't confuse ourself. */ | |
67348ccc | 626 | DECL_EXTERNAL (first_clone->decl) = 0; |
67348ccc DM |
627 | TREE_PUBLIC (first_clone->decl) = 0; |
628 | DECL_COMDAT (first_clone->decl) = 0; | |
9771b263 | 629 | first_clone->ipa_transforms_to_apply.release (); |
fee8b6da | 630 | |
b4e93f45 JH |
631 | /* When doing recursive inlining, the clone may become unnecessary. |
632 | This is possible i.e. in the case when the recursive function is proved to be | |
633 | non-throwing and the recursion happens only in the EH landing pad. | |
67914693 | 634 | We cannot remove the clone until we are done with saving the body. |
b4e93f45 JH |
635 | Remove it now. */ |
636 | if (!first_clone->callers) | |
637 | { | |
d52f5295 | 638 | first_clone->remove_symbol_and_inline_clones (); |
b4e93f45 JH |
639 | first_clone = NULL; |
640 | } | |
b2b29377 | 641 | else if (flag_checking) |
d52f5295 | 642 | first_clone->verify (); |
b2b29377 | 643 | |
fee8b6da JH |
644 | return first_clone; |
645 | } | |
646 | ||
9c8305f8 JH |
647 | /* Return true when function body of DECL still needs to be kept around |
648 | for later re-use. */ | |
65d630d4 | 649 | static bool |
9c8305f8 JH |
650 | preserve_function_body_p (struct cgraph_node *node) |
651 | { | |
3dafb85c | 652 | gcc_assert (symtab->global_info_ready); |
67348ccc | 653 | gcc_assert (!node->alias && !node->thunk.thunk_p); |
9c8305f8 | 654 | |
ec6a1e35 JH |
655 | /* Look if there is any non-thunk clone around. */ |
656 | for (node = node->clones; node; node = node->next_sibling_clone) | |
657 | if (!node->thunk.thunk_p) | |
658 | return true; | |
9c8305f8 JH |
659 | return false; |
660 | } | |
fee8b6da JH |
661 | |
662 | /* Apply inline plan to function. */ | |
663 | ||
664 | unsigned int | |
665 | inline_transform (struct cgraph_node *node) | |
666 | { | |
667 | unsigned int todo = 0; | |
e8aec975 | 668 | struct cgraph_edge *e, *next; |
2bf86c84 | 669 | bool has_inline = false; |
c9fc06dc | 670 | |
fee8b6da JH |
671 | /* FIXME: Currently the pass manager is adding inline transform more than |
672 | once to some clones. This needs revisiting after WPA cleanups. */ | |
673 | if (cfun->after_inlining) | |
674 | return 0; | |
675 | ||
676 | /* We might need the body of this function so that we can expand | |
677 | it inline somewhere else. */ | |
9c8305f8 | 678 | if (preserve_function_body_p (node)) |
fee8b6da JH |
679 | save_inline_function_body (node); |
680 | ||
e8aec975 JH |
681 | for (e = node->callees; e; e = next) |
682 | { | |
2bf86c84 JH |
683 | if (!e->inline_failed) |
684 | has_inline = true; | |
e8aec975 | 685 | next = e->next_callee; |
3dafb85c | 686 | e->redirect_call_stmt_to_callee (); |
e8aec975 | 687 | } |
d122681a | 688 | node->remove_all_references (); |
c9fc06dc CB |
689 | |
690 | timevar_push (TV_INTEGRATION); | |
bb1e543c | 691 | if (node->callees && (opt_for_fn (node->decl, optimize) || has_inline)) |
09fcc0c0 JH |
692 | { |
693 | profile_count num = node->count; | |
694 | profile_count den = ENTRY_BLOCK_PTR_FOR_FN (cfun)->count; | |
517048ce | 695 | bool scale = num.initialized_p () && !(num == den); |
09fcc0c0 JH |
696 | if (scale) |
697 | { | |
517048ce | 698 | profile_count::adjust_for_ipa_scaling (&num, &den); |
09fcc0c0 JH |
699 | if (dump_file) |
700 | { | |
701 | fprintf (dump_file, "Applying count scale "); | |
702 | num.dump (dump_file); | |
703 | fprintf (dump_file, "/"); | |
704 | den.dump (dump_file); | |
705 | fprintf (dump_file, "\n"); | |
706 | } | |
707 | ||
708 | basic_block bb; | |
517048ce | 709 | cfun->cfg->count_max = profile_count::uninitialized (); |
09fcc0c0 | 710 | FOR_ALL_BB_FN (bb, cfun) |
517048ce | 711 | { |
8e7d1486 | 712 | bb->count = bb->count.apply_scale (num, den); |
517048ce JH |
713 | cfun->cfg->count_max = cfun->cfg->count_max.max (bb->count); |
714 | } | |
09fcc0c0 JH |
715 | ENTRY_BLOCK_PTR_FOR_FN (cfun)->count = node->count; |
716 | } | |
717 | todo = optimize_inline_calls (current_function_decl); | |
718 | } | |
c9fc06dc CB |
719 | timevar_pop (TV_INTEGRATION); |
720 | ||
f8698b37 RG |
721 | cfun->always_inline_functions_inlined = true; |
722 | cfun->after_inlining = true; | |
723 | todo |= execute_fixup_cfg (); | |
724 | ||
55f01229 RG |
725 | if (!(todo & TODO_update_ssa_any)) |
726 | /* Redirecting edges might lead to a need for vops to be recomputed. */ | |
727 | todo |= TODO_update_ssa_only_virtuals; | |
728 | ||
f8698b37 | 729 | return todo; |
fee8b6da | 730 | } |