]>
Commit | Line | Data |
---|---|---|
fee8b6da | 1 | /* Callgraph transformations to handle inlining |
8d9254fc | 2 | Copyright (C) 2003-2020 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" | |
2bc2379b | 50 | #include "ipa-utils.h" |
fee8b6da JH |
51 | |
52 | int ncalls_inlined; | |
53 | int nfunctions_inlined; | |
54 | ||
1bad9c18 | 55 | /* Scale counts of NODE edges by NUM/DEN. */ |
fee8b6da JH |
56 | |
57 | static void | |
1bad9c18 JH |
58 | update_noncloned_counts (struct cgraph_node *node, |
59 | profile_count num, profile_count den) | |
fee8b6da JH |
60 | { |
61 | struct cgraph_edge *e; | |
8e7d1486 | 62 | |
e4373d41 | 63 | profile_count::adjust_for_ipa_scaling (&num, &den); |
fee8b6da | 64 | |
fee8b6da JH |
65 | for (e = node->callees; e; e = e->next_callee) |
66 | { | |
fee8b6da | 67 | if (!e->inline_failed) |
1bad9c18 | 68 | update_noncloned_counts (e->callee, num, den); |
8e7d1486 | 69 | e->count = e->count.apply_scale (num, den); |
898b8927 JH |
70 | } |
71 | for (e = node->indirect_calls; e; e = e->next_callee) | |
1bad9c18 | 72 | e->count = e->count.apply_scale (num, den); |
8e7d1486 | 73 | node->count = node->count.apply_scale (num, den); |
fee8b6da JH |
74 | } |
75 | ||
a5b1779f JH |
76 | /* We removed or are going to remove the last call to NODE. |
77 | Return true if we can and want proactively remove the NODE now. | |
78 | This is important to do, since we want inliner to know when offline | |
79 | copy of function was removed. */ | |
80 | ||
81 | static bool | |
d142079a | 82 | can_remove_node_now_p_1 (struct cgraph_node *node, struct cgraph_edge *e) |
a5b1779f | 83 | { |
d142079a JH |
84 | ipa_ref *ref; |
85 | ||
86 | FOR_EACH_ALIAS (node, ref) | |
87 | { | |
88 | cgraph_node *alias = dyn_cast <cgraph_node *> (ref->referring); | |
89 | if ((alias->callers && alias->callers != e) | |
90 | || !can_remove_node_now_p_1 (alias, e)) | |
91 | return false; | |
92 | } | |
a5b1779f JH |
93 | /* FIXME: When address is taken of DECL_EXTERNAL function we still |
94 | can remove its offline copy, but we would need to keep unanalyzed node in | |
a6a543bf JH |
95 | the callgraph so references can point to it. |
96 | ||
97 | Also for comdat group we can ignore references inside a group as we | |
98 | want to prove the group as a whole to be dead. */ | |
67348ccc | 99 | return (!node->address_taken |
a6a543bf | 100 | && node->can_remove_if_no_direct_calls_and_refs_p () |
a5b1779f JH |
101 | /* Inlining might enable more devirtualizing, so we want to remove |
102 | those only after all devirtualizable virtual calls are processed. | |
103 | Lacking may edges in callgraph we just preserve them post | |
104 | inlining. */ | |
d142079a JH |
105 | && (!DECL_VIRTUAL_P (node->decl) |
106 | || !opt_for_fn (node->decl, flag_devirtualize)) | |
a5b1779f | 107 | /* During early inlining some unanalyzed cgraph nodes might be in the |
956d615d | 108 | callgraph and they might refer the function in question. */ |
31acf1bb | 109 | && !cgraph_new_nodes.exists ()); |
a5b1779f JH |
110 | } |
111 | ||
6c69a029 JH |
112 | /* We are going to eliminate last direct call to NODE (or alias of it) via edge E. |
113 | Verify that the NODE can be removed from unit and if it is contained in comdat | |
114 | group that the whole comdat group is removable. */ | |
115 | ||
116 | static bool | |
117 | can_remove_node_now_p (struct cgraph_node *node, struct cgraph_edge *e) | |
118 | { | |
119 | struct cgraph_node *next; | |
d142079a | 120 | if (!can_remove_node_now_p_1 (node, e)) |
6c69a029 JH |
121 | return false; |
122 | ||
123 | /* When we see same comdat group, we need to be sure that all | |
124 | items can be removed. */ | |
8ccc8042 | 125 | if (!node->same_comdat_group || !node->externally_visible) |
6c69a029 | 126 | return true; |
d52f5295 ML |
127 | for (next = dyn_cast<cgraph_node *> (node->same_comdat_group); |
128 | next != node; next = dyn_cast<cgraph_node *> (next->same_comdat_group)) | |
d142079a JH |
129 | { |
130 | if (next->alias) | |
131 | continue; | |
132 | if ((next->callers && next->callers != e) | |
133 | || !can_remove_node_now_p_1 (next, e)) | |
134 | return false; | |
135 | } | |
6c69a029 JH |
136 | return true; |
137 | } | |
138 | ||
d83fa499 EB |
139 | /* Return true if NODE is a master clone with non-inline clones. */ |
140 | ||
141 | static bool | |
142 | master_clone_with_noninline_clones_p (struct cgraph_node *node) | |
143 | { | |
144 | if (node->clone_of) | |
145 | return false; | |
146 | ||
147 | for (struct cgraph_node *n = node->clones; n; n = n->next_sibling_clone) | |
148 | if (n->decl != node->decl) | |
149 | return true; | |
150 | ||
151 | return false; | |
152 | } | |
fee8b6da JH |
153 | |
154 | /* E is expected to be an edge being inlined. Clone destination node of | |
155 | the edge and redirect it to the new clone. | |
156 | DUPLICATE is used for bookkeeping on whether we are actually creating new | |
157 | clones or re-using node originally representing out-of-line function call. | |
bd936951 JH |
158 | By default the offline copy is removed, when it appears dead after inlining. |
159 | UPDATE_ORIGINAL prevents this transformation. | |
160 | If OVERALL_SIZE is non-NULL, the size is updated to reflect the | |
1bad9c18 | 161 | transformation. */ |
fee8b6da JH |
162 | |
163 | void | |
164 | clone_inlined_nodes (struct cgraph_edge *e, bool duplicate, | |
1bad9c18 | 165 | bool update_original, int *overall_size) |
fee8b6da | 166 | { |
44a60244 | 167 | struct cgraph_node *inlining_into; |
09ce3660 | 168 | struct cgraph_edge *next; |
44a60244 | 169 | |
a62bfab5 ML |
170 | if (e->caller->inlined_to) |
171 | inlining_into = e->caller->inlined_to; | |
44a60244 MJ |
172 | else |
173 | inlining_into = e->caller; | |
174 | ||
fee8b6da JH |
175 | if (duplicate) |
176 | { | |
177 | /* We may eliminate the need for out-of-line copy to be output. | |
178 | In that case just go ahead and re-use it. This is not just an | |
956d615d | 179 | memory optimization. Making offline copy of function disappear |
fee8b6da JH |
180 | from the program will improve future decisions on inlining. */ |
181 | if (!e->callee->callers->next_caller | |
182 | /* Recursive inlining never wants the master clone to | |
183 | be overwritten. */ | |
184 | && update_original | |
d83fa499 EB |
185 | && can_remove_node_now_p (e->callee, e) |
186 | /* We cannot overwrite a master clone with non-inline clones | |
187 | until after these clones are materialized. */ | |
188 | && !master_clone_with_noninline_clones_p (e->callee)) | |
fee8b6da | 189 | { |
6c69a029 JH |
190 | /* TODO: When callee is in a comdat group, we could remove all of it, |
191 | including all inline clones inlined into it. That would however | |
192 | need small function inlining to register edge removal hook to | |
193 | maintain the priority queue. | |
194 | ||
956d615d | 195 | For now we keep the other functions in the group in program until |
6c69a029 | 196 | cgraph_remove_unreachable_functions gets rid of them. */ |
a62bfab5 | 197 | gcc_assert (!e->callee->inlined_to); |
b91b562c | 198 | e->callee->remove_from_same_comdat_group (); |
bb1e543c JH |
199 | if (e->callee->definition |
200 | && inline_account_function_p (e->callee)) | |
fee8b6da | 201 | { |
bb1e543c | 202 | gcc_assert (!e->callee->alias); |
fee8b6da | 203 | if (overall_size) |
f658ad30 | 204 | *overall_size -= ipa_size_summaries->get (e->callee)->size; |
fee8b6da JH |
205 | nfunctions_inlined++; |
206 | } | |
207 | duplicate = false; | |
67348ccc | 208 | e->callee->externally_visible = false; |
1bad9c18 | 209 | update_noncloned_counts (e->callee, e->count, e->callee->count); |
0bdad123 ML |
210 | |
211 | dump_callgraph_transformation (e->callee, inlining_into, | |
212 | "inlining to"); | |
fee8b6da JH |
213 | } |
214 | else | |
215 | { | |
216 | struct cgraph_node *n; | |
bd936951 | 217 | |
d52f5295 | 218 | n = e->callee->create_clone (e->callee->decl, |
1bad9c18 | 219 | e->count, |
d52f5295 ML |
220 | update_original, vNULL, true, |
221 | inlining_into, | |
222 | NULL); | |
4ad08ee8 | 223 | n->used_as_abstract_origin = e->callee->used_as_abstract_origin; |
3dafb85c | 224 | e->redirect_callee (n); |
fee8b6da JH |
225 | } |
226 | } | |
65d630d4 | 227 | else |
b91b562c | 228 | e->callee->remove_from_same_comdat_group (); |
fee8b6da | 229 | |
a62bfab5 | 230 | e->callee->inlined_to = inlining_into; |
fee8b6da JH |
231 | |
232 | /* Recursively clone all bodies. */ | |
09ce3660 JH |
233 | for (e = e->callee->callees; e; e = next) |
234 | { | |
235 | next = e->next_callee; | |
236 | if (!e->inline_failed) | |
1bad9c18 | 237 | clone_inlined_nodes (e, duplicate, update_original, overall_size); |
db66bf68 JH |
238 | } |
239 | } | |
240 | ||
4517b378 MJ |
241 | /* Check all speculations in N and if any seem useless, resolve them. When a |
242 | first edge is resolved, pop all edges from NEW_EDGES and insert them to | |
243 | EDGE_SET. Then remove each resolved edge from EDGE_SET, if it is there. */ | |
db66bf68 JH |
244 | |
245 | static bool | |
4517b378 MJ |
246 | check_speculations_1 (cgraph_node *n, vec<cgraph_edge *> *new_edges, |
247 | hash_set <cgraph_edge *> *edge_set) | |
db66bf68 JH |
248 | { |
249 | bool speculation_removed = false; | |
250 | cgraph_edge *next; | |
251 | ||
252 | for (cgraph_edge *e = n->callees; e; e = next) | |
253 | { | |
254 | next = e->next_callee; | |
09ce3660 JH |
255 | if (e->speculative && !speculation_useful_p (e, true)) |
256 | { | |
4517b378 MJ |
257 | while (new_edges && !new_edges->is_empty ()) |
258 | edge_set->add (new_edges->pop ()); | |
259 | edge_set->remove (e); | |
260 | ||
27c5a177 | 261 | cgraph_edge::resolve_speculation (e, NULL); |
09ce3660 JH |
262 | speculation_removed = true; |
263 | } | |
db66bf68 | 264 | else if (!e->inline_failed) |
4517b378 MJ |
265 | speculation_removed |= check_speculations_1 (e->callee, new_edges, |
266 | edge_set); | |
09ce3660 | 267 | } |
db66bf68 | 268 | return speculation_removed; |
fee8b6da JH |
269 | } |
270 | ||
4517b378 MJ |
271 | /* Push E to NEW_EDGES. Called from hash_set traverse method, which |
272 | unfortunately means this function has to have external linkage, otherwise | |
273 | the code will not compile with gcc 4.8. */ | |
274 | ||
275 | bool | |
276 | push_all_edges_in_set_to_vec (cgraph_edge * const &e, | |
277 | vec<cgraph_edge *> *new_edges) | |
278 | { | |
279 | new_edges->safe_push (e); | |
280 | return true; | |
281 | } | |
282 | ||
283 | /* Check all speculations in N and if any seem useless, resolve them and remove | |
284 | them from NEW_EDGES. */ | |
285 | ||
286 | static bool | |
287 | check_speculations (cgraph_node *n, vec<cgraph_edge *> *new_edges) | |
288 | { | |
289 | hash_set <cgraph_edge *> edge_set; | |
290 | bool res = check_speculations_1 (n, new_edges, &edge_set); | |
291 | if (!edge_set.is_empty ()) | |
292 | edge_set.traverse <vec<cgraph_edge *> *, | |
293 | push_all_edges_in_set_to_vec> (new_edges); | |
294 | return res; | |
295 | } | |
296 | ||
4fd94d1e MJ |
297 | /* Mark all call graph edges coming out of NODE and all nodes that have been |
298 | inlined to it as in_polymorphic_cdtor. */ | |
299 | ||
300 | static void | |
301 | mark_all_inlined_calls_cdtor (cgraph_node *node) | |
302 | { | |
303 | for (cgraph_edge *cs = node->callees; cs; cs = cs->next_callee) | |
304 | { | |
305 | cs->in_polymorphic_cdtor = true; | |
306 | if (!cs->inline_failed) | |
09fcc0c0 | 307 | mark_all_inlined_calls_cdtor (cs->callee); |
4fd94d1e MJ |
308 | } |
309 | for (cgraph_edge *cs = node->indirect_calls; cs; cs = cs->next_callee) | |
310 | cs->in_polymorphic_cdtor = true; | |
311 | } | |
312 | ||
fee8b6da JH |
313 | |
314 | /* Mark edge E as inlined and update callgraph accordingly. UPDATE_ORIGINAL | |
315 | specify whether profile of original function should be updated. If any new | |
316 | indirect edges are discovered in the process, add them to NEW_EDGES, unless | |
c170d40f JH |
317 | it is NULL. If UPDATE_OVERALL_SUMMARY is false, do not bother to recompute overall |
318 | size of caller after inlining. Caller is required to eventually do it via | |
0bceb671 | 319 | ipa_update_overall_fn_summary. |
1bbb87c4 | 320 | If callee_removed is non-NULL, set it to true if we removed callee node. |
c170d40f JH |
321 | |
322 | Return true iff any new callgraph edges were discovered as a | |
fee8b6da JH |
323 | result of inlining. */ |
324 | ||
325 | bool | |
326 | inline_call (struct cgraph_edge *e, bool update_original, | |
d52f5295 | 327 | vec<cgraph_edge *> *new_edges, |
1bbb87c4 JH |
328 | int *overall_size, bool update_overall_summary, |
329 | bool *callee_removed) | |
fee8b6da JH |
330 | { |
331 | int old_size = 0, new_size = 0; | |
332 | struct cgraph_node *to = NULL; | |
333 | struct cgraph_edge *curr = e; | |
a414fd42 | 334 | bool comdat_local = e->callee->comdat_local_p (); |
d52f5295 | 335 | struct cgraph_node *callee = e->callee->ultimate_alias_target (); |
0f378cb5 JH |
336 | bool new_edges_found = false; |
337 | ||
bddead15 RB |
338 | int estimated_growth = 0; |
339 | if (! update_overall_summary) | |
340 | estimated_growth = estimate_edge_growth (e); | |
f107227b JH |
341 | /* This is used only for assert bellow. */ |
342 | #if 0 | |
0f378cb5 JH |
343 | bool predicated = inline_edge_summary (e)->predicate != NULL; |
344 | #endif | |
fee8b6da JH |
345 | |
346 | /* Don't inline inlined edges. */ | |
347 | gcc_assert (e->inline_failed); | |
348 | /* Don't even think of inlining inline clone. */ | |
a62bfab5 | 349 | gcc_assert (!callee->inlined_to); |
fee8b6da | 350 | |
632b4f8e | 351 | to = e->caller; |
a62bfab5 ML |
352 | if (to->inlined_to) |
353 | to = to->inlined_to; | |
0b9004ed JH |
354 | if (to->thunk.thunk_p) |
355 | { | |
c7ed8938 | 356 | struct cgraph_node *target = to->callees->callee; |
2bc2379b | 357 | thunk_expansion = true; |
a088d7b1 | 358 | symtab->call_cgraph_removal_hooks (to); |
0b9004ed JH |
359 | if (in_lto_p) |
360 | to->get_untransformed_body (); | |
361 | to->expand_thunk (false, true); | |
c7ed8938 IE |
362 | /* When thunk is instrumented we may have multiple callees. */ |
363 | for (e = to->callees; e && e->callee != target; e = e->next_callee) | |
364 | ; | |
a088d7b1 | 365 | symtab->call_cgraph_insertion_hooks (to); |
2bc2379b | 366 | thunk_expansion = false; |
c7ed8938 | 367 | gcc_assert (e); |
0b9004ed JH |
368 | } |
369 | ||
370 | ||
371 | e->inline_failed = CIF_OK; | |
372 | DECL_POSSIBLY_INLINED (callee->decl) = true; | |
632b4f8e | 373 | |
5058c037 JH |
374 | if (DECL_FUNCTION_PERSONALITY (callee->decl)) |
375 | DECL_FUNCTION_PERSONALITY (to->decl) | |
376 | = DECL_FUNCTION_PERSONALITY (callee->decl); | |
77719b06 ML |
377 | |
378 | bool reload_optimization_node = false; | |
45285060 JH |
379 | if (!opt_for_fn (callee->decl, flag_strict_aliasing) |
380 | && opt_for_fn (to->decl, flag_strict_aliasing)) | |
381 | { | |
382 | struct gcc_options opts = global_options; | |
383 | ||
f7f32acd | 384 | cl_optimization_restore (&opts, opts_for_fn (to->decl)); |
45285060 JH |
385 | opts.x_flag_strict_aliasing = false; |
386 | if (dump_file) | |
464d0118 ML |
387 | fprintf (dump_file, "Dropping flag_strict_aliasing on %s\n", |
388 | to->dump_name ()); | |
45285060 JH |
389 | DECL_FUNCTION_SPECIFIC_OPTIMIZATION (to->decl) |
390 | = build_optimization_node (&opts); | |
77719b06 | 391 | reload_optimization_node = true; |
45285060 | 392 | } |
5c846a81 | 393 | |
56f62793 ML |
394 | ipa_fn_summary *caller_info = ipa_fn_summaries->get (to); |
395 | ipa_fn_summary *callee_info = ipa_fn_summaries->get (callee); | |
818b88a7 JH |
396 | if (!caller_info->fp_expressions && callee_info->fp_expressions) |
397 | { | |
398 | caller_info->fp_expressions = true; | |
399 | if (opt_for_fn (callee->decl, flag_rounding_math) | |
400 | != opt_for_fn (to->decl, flag_rounding_math) | |
401 | || opt_for_fn (callee->decl, flag_trapping_math) | |
402 | != opt_for_fn (to->decl, flag_trapping_math) | |
403 | || opt_for_fn (callee->decl, flag_unsafe_math_optimizations) | |
404 | != opt_for_fn (to->decl, flag_unsafe_math_optimizations) | |
405 | || opt_for_fn (callee->decl, flag_finite_math_only) | |
406 | != opt_for_fn (to->decl, flag_finite_math_only) | |
407 | || opt_for_fn (callee->decl, flag_signaling_nans) | |
408 | != opt_for_fn (to->decl, flag_signaling_nans) | |
409 | || opt_for_fn (callee->decl, flag_cx_limited_range) | |
410 | != opt_for_fn (to->decl, flag_cx_limited_range) | |
411 | || opt_for_fn (callee->decl, flag_signed_zeros) | |
412 | != opt_for_fn (to->decl, flag_signed_zeros) | |
413 | || opt_for_fn (callee->decl, flag_associative_math) | |
414 | != opt_for_fn (to->decl, flag_associative_math) | |
415 | || opt_for_fn (callee->decl, flag_reciprocal_math) | |
416 | != opt_for_fn (to->decl, flag_reciprocal_math) | |
0d2f700f JM |
417 | || opt_for_fn (callee->decl, flag_fp_int_builtin_inexact) |
418 | != opt_for_fn (to->decl, flag_fp_int_builtin_inexact) | |
818b88a7 JH |
419 | || opt_for_fn (callee->decl, flag_errno_math) |
420 | != opt_for_fn (to->decl, flag_errno_math)) | |
421 | { | |
422 | struct gcc_options opts = global_options; | |
423 | ||
424 | cl_optimization_restore (&opts, opts_for_fn (to->decl)); | |
425 | opts.x_flag_rounding_math | |
426 | = opt_for_fn (callee->decl, flag_rounding_math); | |
427 | opts.x_flag_trapping_math | |
428 | = opt_for_fn (callee->decl, flag_trapping_math); | |
429 | opts.x_flag_unsafe_math_optimizations | |
430 | = opt_for_fn (callee->decl, flag_unsafe_math_optimizations); | |
431 | opts.x_flag_finite_math_only | |
432 | = opt_for_fn (callee->decl, flag_finite_math_only); | |
433 | opts.x_flag_signaling_nans | |
434 | = opt_for_fn (callee->decl, flag_signaling_nans); | |
435 | opts.x_flag_cx_limited_range | |
436 | = opt_for_fn (callee->decl, flag_cx_limited_range); | |
437 | opts.x_flag_signed_zeros | |
438 | = opt_for_fn (callee->decl, flag_signed_zeros); | |
439 | opts.x_flag_associative_math | |
440 | = opt_for_fn (callee->decl, flag_associative_math); | |
441 | opts.x_flag_reciprocal_math | |
442 | = opt_for_fn (callee->decl, flag_reciprocal_math); | |
0d2f700f JM |
443 | opts.x_flag_fp_int_builtin_inexact |
444 | = opt_for_fn (callee->decl, flag_fp_int_builtin_inexact); | |
818b88a7 JH |
445 | opts.x_flag_errno_math |
446 | = opt_for_fn (callee->decl, flag_errno_math); | |
447 | if (dump_file) | |
464d0118 ML |
448 | fprintf (dump_file, "Copying FP flags from %s to %s\n", |
449 | callee->dump_name (), to->dump_name ()); | |
818b88a7 JH |
450 | DECL_FUNCTION_SPECIFIC_OPTIMIZATION (to->decl) |
451 | = build_optimization_node (&opts); | |
77719b06 | 452 | reload_optimization_node = true; |
818b88a7 JH |
453 | } |
454 | } | |
5058c037 | 455 | |
77719b06 ML |
456 | /* Reload global optimization flags. */ |
457 | if (reload_optimization_node && DECL_STRUCT_FUNCTION (to->decl) == cfun) | |
458 | set_cfun (cfun, true); | |
459 | ||
a5b1779f JH |
460 | /* If aliases are involved, redirect edge to the actual destination and |
461 | possibly remove the aliases. */ | |
462 | if (e->callee != callee) | |
39e2db00 JH |
463 | { |
464 | struct cgraph_node *alias = e->callee, *next_alias; | |
3dafb85c | 465 | e->redirect_callee (callee); |
39e2db00 JH |
466 | while (alias && alias != callee) |
467 | { | |
468 | if (!alias->callers | |
8ccc8042 JH |
469 | && can_remove_node_now_p (alias, |
470 | !e->next_caller && !e->prev_caller ? e : NULL)) | |
39e2db00 | 471 | { |
d52f5295 ML |
472 | next_alias = alias->get_alias_target (); |
473 | alias->remove (); | |
1bbb87c4 JH |
474 | if (callee_removed) |
475 | *callee_removed = true; | |
39e2db00 JH |
476 | alias = next_alias; |
477 | } | |
478 | else | |
479 | break; | |
480 | } | |
481 | } | |
a5b1779f | 482 | |
1bad9c18 | 483 | clone_inlined_nodes (e, true, update_original, overall_size); |
fee8b6da | 484 | |
a62bfab5 | 485 | gcc_assert (curr->callee->inlined_to == to); |
898b8927 | 486 | |
f658ad30 | 487 | old_size = ipa_size_summaries->get (to)->size; |
0bceb671 | 488 | ipa_merge_fn_summary_after_inlining (e); |
4fd94d1e MJ |
489 | if (e->in_polymorphic_cdtor) |
490 | mark_all_inlined_calls_cdtor (e->callee); | |
bb1e543c | 491 | if (opt_for_fn (e->caller->decl, optimize)) |
0f378cb5 | 492 | new_edges_found = ipa_propagate_indirect_call_infos (curr, new_edges); |
d2bcf46c | 493 | bool removed_p = check_speculations (e->callee, new_edges); |
c170d40f | 494 | if (update_overall_summary) |
d2bcf46c | 495 | ipa_update_overall_fn_summary (to, new_edges_found || removed_p); |
bddead15 RB |
496 | else |
497 | /* Update self size by the estimate so overall function growth limits | |
498 | work for further inlining into this function. Before inlining | |
499 | the function we inlined to again we expect the caller to update | |
500 | the overall summary. */ | |
f658ad30 JH |
501 | ipa_size_summaries->get (to)->size += estimated_growth; |
502 | new_size = ipa_size_summaries->get (to)->size; | |
d250540a | 503 | |
1f26ac87 JM |
504 | if (callee->calls_comdat_local) |
505 | to->calls_comdat_local = true; | |
a414fd42 | 506 | else if (to->calls_comdat_local && comdat_local) |
72b3bc89 | 507 | to->calls_comdat_local = to->check_calls_comdat_local_p (); |
1f26ac87 | 508 | |
f107227b JH |
509 | /* FIXME: This assert suffers from roundoff errors, disable it for GCC 5 |
510 | and revisit it after conversion to sreals in GCC 6. | |
511 | See PR 65654. */ | |
512 | #if 0 | |
0f378cb5 | 513 | /* Verify that estimated growth match real growth. Allow off-by-one |
0bceb671 | 514 | error due to ipa_fn_summary::size_scale roudoff errors. */ |
48b1474e | 515 | gcc_assert (!update_overall_summary || !overall_size || new_edges_found |
0f378cb5 | 516 | || abs (estimated_growth - (new_size - old_size)) <= 1 |
09ce3660 | 517 | || speculation_removed |
0f378cb5 JH |
518 | /* FIXME: a hack. Edges with false predicate are accounted |
519 | wrong, we should remove them from callgraph. */ | |
520 | || predicated); | |
521 | #endif | |
d250540a | 522 | |
8256d5ca JH |
523 | /* Account the change of overall unit size; external functions will be |
524 | removed and are thus not accounted. */ | |
bb1e543c | 525 | if (overall_size && inline_account_function_p (to)) |
fee8b6da JH |
526 | *overall_size += new_size - old_size; |
527 | ncalls_inlined++; | |
528 | ||
0bceb671 | 529 | /* This must happen after ipa_merge_fn_summary_after_inlining that rely on jump |
25837a2f | 530 | functions of callee to not be updated. */ |
0f378cb5 | 531 | return new_edges_found; |
fee8b6da JH |
532 | } |
533 | ||
7123347c MJ |
534 | /* For each node that was made the holder of function body by |
535 | save_inline_function_body, this summary contains pointer to the previous | |
536 | holder of the body. */ | |
537 | ||
538 | function_summary <tree *> *ipa_saved_clone_sources; | |
fee8b6da JH |
539 | |
540 | /* Copy function body of NODE and redirect all inline clones to it. | |
541 | This is done before inline plan is applied to NODE when there are | |
542 | still some inline clones if it. | |
543 | ||
073a8998 | 544 | This is necessary because inline decisions are not really transitive |
fee8b6da JH |
545 | and the other inline clones may have different bodies. */ |
546 | ||
547 | static struct cgraph_node * | |
548 | save_inline_function_body (struct cgraph_node *node) | |
549 | { | |
550 | struct cgraph_node *first_clone, *n; | |
551 | ||
552 | if (dump_file) | |
553 | fprintf (dump_file, "\nSaving body of %s for later reuse\n", | |
3629ff8a | 554 | node->dump_name ()); |
fee8b6da | 555 | |
d52f5295 | 556 | gcc_assert (node == cgraph_node::get (node->decl)); |
fee8b6da JH |
557 | |
558 | /* first_clone will be turned into real function. */ | |
559 | first_clone = node->clones; | |
ec6a1e35 JH |
560 | |
561 | /* Arrange first clone to not be thunk as those do not have bodies. */ | |
562 | if (first_clone->thunk.thunk_p) | |
563 | { | |
564 | while (first_clone->thunk.thunk_p) | |
565 | first_clone = first_clone->next_sibling_clone; | |
566 | first_clone->prev_sibling_clone->next_sibling_clone | |
567 | = first_clone->next_sibling_clone; | |
568 | if (first_clone->next_sibling_clone) | |
569 | first_clone->next_sibling_clone->prev_sibling_clone | |
570 | = first_clone->prev_sibling_clone; | |
571 | first_clone->next_sibling_clone = node->clones; | |
572 | first_clone->prev_sibling_clone = NULL; | |
573 | node->clones->prev_sibling_clone = first_clone; | |
574 | node->clones = first_clone; | |
575 | } | |
67348ccc | 576 | first_clone->decl = copy_node (node->decl); |
aede2c10 | 577 | first_clone->decl->decl_with_vis.symtab_node = first_clone; |
d52f5295 | 578 | gcc_assert (first_clone == cgraph_node::get (first_clone->decl)); |
fee8b6da JH |
579 | |
580 | /* Now reshape the clone tree, so all other clones descends from | |
581 | first_clone. */ | |
582 | if (first_clone->next_sibling_clone) | |
583 | { | |
ec6a1e35 JH |
584 | for (n = first_clone->next_sibling_clone; n->next_sibling_clone; |
585 | n = n->next_sibling_clone) | |
fee8b6da JH |
586 | n->clone_of = first_clone; |
587 | n->clone_of = first_clone; | |
588 | n->next_sibling_clone = first_clone->clones; | |
589 | if (first_clone->clones) | |
590 | first_clone->clones->prev_sibling_clone = n; | |
591 | first_clone->clones = first_clone->next_sibling_clone; | |
592 | first_clone->next_sibling_clone->prev_sibling_clone = NULL; | |
593 | first_clone->next_sibling_clone = NULL; | |
594 | gcc_assert (!first_clone->prev_sibling_clone); | |
595 | } | |
7123347c MJ |
596 | |
597 | tree prev_body_holder = node->decl; | |
598 | if (!ipa_saved_clone_sources) | |
599 | ipa_saved_clone_sources = new function_summary <tree *> (symtab); | |
600 | else | |
601 | { | |
602 | tree *p = ipa_saved_clone_sources->get (node); | |
603 | if (p) | |
604 | { | |
605 | prev_body_holder = *p; | |
606 | gcc_assert (prev_body_holder); | |
607 | } | |
608 | } | |
609 | *ipa_saved_clone_sources->get_create (first_clone) = prev_body_holder; | |
b31ede6e MJ |
610 | first_clone->former_clone_of |
611 | = node->former_clone_of ? node->former_clone_of : node->decl; | |
fee8b6da JH |
612 | first_clone->clone_of = NULL; |
613 | ||
614 | /* Now node in question has no clones. */ | |
615 | node->clones = NULL; | |
616 | ||
1a3118e9 JH |
617 | /* Inline clones share decl with the function they are cloned |
618 | from. Walk the whole clone tree and redirect them all to the | |
619 | new decl. */ | |
fee8b6da JH |
620 | if (first_clone->clones) |
621 | for (n = first_clone->clones; n != first_clone;) | |
622 | { | |
67348ccc DM |
623 | gcc_assert (n->decl == node->decl); |
624 | n->decl = first_clone->decl; | |
fee8b6da JH |
625 | if (n->clones) |
626 | n = n->clones; | |
627 | else if (n->next_sibling_clone) | |
628 | n = n->next_sibling_clone; | |
629 | else | |
630 | { | |
631 | while (n != first_clone && !n->next_sibling_clone) | |
632 | n = n->clone_of; | |
633 | if (n != first_clone) | |
634 | n = n->next_sibling_clone; | |
635 | } | |
636 | } | |
637 | ||
638 | /* Copy the OLD_VERSION_NODE function tree to the new version. */ | |
67348ccc | 639 | tree_function_versioning (node->decl, first_clone->decl, |
ff6686d2 | 640 | NULL, NULL, true, NULL, NULL); |
fee8b6da | 641 | |
1a3118e9 JH |
642 | /* The function will be short lived and removed after we inline all the clones, |
643 | but make it internal so we won't confuse ourself. */ | |
67348ccc | 644 | DECL_EXTERNAL (first_clone->decl) = 0; |
67348ccc DM |
645 | TREE_PUBLIC (first_clone->decl) = 0; |
646 | DECL_COMDAT (first_clone->decl) = 0; | |
9771b263 | 647 | first_clone->ipa_transforms_to_apply.release (); |
fee8b6da | 648 | |
b4e93f45 JH |
649 | /* When doing recursive inlining, the clone may become unnecessary. |
650 | This is possible i.e. in the case when the recursive function is proved to be | |
651 | non-throwing and the recursion happens only in the EH landing pad. | |
67914693 | 652 | We cannot remove the clone until we are done with saving the body. |
b4e93f45 JH |
653 | Remove it now. */ |
654 | if (!first_clone->callers) | |
655 | { | |
d52f5295 | 656 | first_clone->remove_symbol_and_inline_clones (); |
b4e93f45 JH |
657 | first_clone = NULL; |
658 | } | |
b2b29377 | 659 | else if (flag_checking) |
d52f5295 | 660 | first_clone->verify (); |
b2b29377 | 661 | |
fee8b6da JH |
662 | return first_clone; |
663 | } | |
664 | ||
9c8305f8 JH |
665 | /* Return true when function body of DECL still needs to be kept around |
666 | for later re-use. */ | |
65d630d4 | 667 | static bool |
9c8305f8 JH |
668 | preserve_function_body_p (struct cgraph_node *node) |
669 | { | |
3dafb85c | 670 | gcc_assert (symtab->global_info_ready); |
67348ccc | 671 | gcc_assert (!node->alias && !node->thunk.thunk_p); |
9c8305f8 | 672 | |
ec6a1e35 JH |
673 | /* Look if there is any non-thunk clone around. */ |
674 | for (node = node->clones; node; node = node->next_sibling_clone) | |
675 | if (!node->thunk.thunk_p) | |
676 | return true; | |
9c8305f8 JH |
677 | return false; |
678 | } | |
fee8b6da JH |
679 | |
680 | /* Apply inline plan to function. */ | |
681 | ||
682 | unsigned int | |
683 | inline_transform (struct cgraph_node *node) | |
684 | { | |
685 | unsigned int todo = 0; | |
e8aec975 | 686 | struct cgraph_edge *e, *next; |
2bf86c84 | 687 | bool has_inline = false; |
c9fc06dc | 688 | |
fee8b6da JH |
689 | /* FIXME: Currently the pass manager is adding inline transform more than |
690 | once to some clones. This needs revisiting after WPA cleanups. */ | |
691 | if (cfun->after_inlining) | |
692 | return 0; | |
693 | ||
694 | /* We might need the body of this function so that we can expand | |
695 | it inline somewhere else. */ | |
9c8305f8 | 696 | if (preserve_function_body_p (node)) |
fee8b6da JH |
697 | save_inline_function_body (node); |
698 | ||
c3f1ae8a JH |
699 | profile_count num = node->count; |
700 | profile_count den = ENTRY_BLOCK_PTR_FOR_FN (cfun)->count; | |
701 | bool scale = num.initialized_p () && !(num == den); | |
702 | if (scale) | |
703 | { | |
704 | profile_count::adjust_for_ipa_scaling (&num, &den); | |
705 | if (dump_file) | |
706 | { | |
707 | fprintf (dump_file, "Applying count scale "); | |
708 | num.dump (dump_file); | |
709 | fprintf (dump_file, "/"); | |
710 | den.dump (dump_file); | |
711 | fprintf (dump_file, "\n"); | |
712 | } | |
713 | ||
714 | basic_block bb; | |
715 | cfun->cfg->count_max = profile_count::uninitialized (); | |
716 | FOR_ALL_BB_FN (bb, cfun) | |
717 | { | |
718 | bb->count = bb->count.apply_scale (num, den); | |
719 | cfun->cfg->count_max = cfun->cfg->count_max.max (bb->count); | |
720 | } | |
721 | ENTRY_BLOCK_PTR_FOR_FN (cfun)->count = node->count; | |
722 | } | |
723 | ||
e8aec975 JH |
724 | for (e = node->callees; e; e = next) |
725 | { | |
2bf86c84 JH |
726 | if (!e->inline_failed) |
727 | has_inline = true; | |
e8aec975 | 728 | next = e->next_callee; |
27c5a177 | 729 | cgraph_edge::redirect_call_stmt_to_callee (e); |
e8aec975 | 730 | } |
d122681a | 731 | node->remove_all_references (); |
c9fc06dc CB |
732 | |
733 | timevar_push (TV_INTEGRATION); | |
bb1e543c | 734 | if (node->callees && (opt_for_fn (node->decl, optimize) || has_inline)) |
09fcc0c0 | 735 | { |
09fcc0c0 | 736 | todo = optimize_inline_calls (current_function_decl); |
c3f1ae8a | 737 | } |
c9fc06dc CB |
738 | timevar_pop (TV_INTEGRATION); |
739 | ||
f8698b37 RG |
740 | cfun->always_inline_functions_inlined = true; |
741 | cfun->after_inlining = true; | |
742 | todo |= execute_fixup_cfg (); | |
743 | ||
55f01229 RG |
744 | if (!(todo & TODO_update_ssa_any)) |
745 | /* Redirecting edges might lead to a need for vops to be recomputed. */ | |
746 | todo |= TODO_update_ssa_only_virtuals; | |
747 | ||
f8698b37 | 748 | return todo; |
fee8b6da | 749 | } |