]>
Commit | Line | Data |
---|---|---|
08f835dc | 1 | /* Basic IPA optimizations based on profile. |
8d9254fc | 2 | Copyright (C) 2003-2020 Free Software Foundation, Inc. |
08f835dc JH |
3 | |
4 | This file is part of GCC. | |
5 | ||
6 | GCC is free software; you can redistribute it and/or modify it under | |
7 | the terms of the GNU General Public License as published by the Free | |
8 | Software Foundation; either version 3, or (at your option) any later | |
9 | version. | |
10 | ||
11 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
12 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
14 | for more details. | |
15 | ||
16 | You should have received a copy of the GNU General Public License | |
17 | along with GCC; see the file COPYING3. If not see | |
18 | <http://www.gnu.org/licenses/>. */ | |
19 | ||
daf5c770 JH |
20 | /* ipa-profile pass implements the following analysis propagating profille |
21 | inter-procedurally. | |
22 | ||
23 | - Count histogram construction. This is a histogram analyzing how much | |
24 | time is spent executing statements with a given execution count read | |
1c5fd343 | 25 | from profile feedback. This histogram is complete only with LTO, |
daf5c770 JH |
26 | otherwise it contains information only about the current unit. |
27 | ||
daf5c770 JH |
28 | The information is used to set hot/cold thresholds. |
29 | - Next speculative indirect call resolution is performed: the local | |
30 | profile pass assigns profile-id to each function and provide us with a | |
31 | histogram specifying the most common target. We look up the callgraph | |
32 | node corresponding to the target and produce a speculative call. | |
33 | ||
34 | This call may or may not survive through IPA optimization based on decision | |
35 | of inliner. | |
36 | - Finally we propagate the following flags: unlikely executed, executed | |
37 | once, executed at startup and executed at exit. These flags are used to | |
026c3cfd | 38 | control code size/performance threshold and code placement (by producing |
daf5c770 | 39 | .text.unlikely/.text.hot/.text.startup/.text.exit subsections). */ |
08f835dc JH |
40 | #include "config.h" |
41 | #include "system.h" | |
42 | #include "coretypes.h" | |
c7131fb2 | 43 | #include "backend.h" |
4d648807 | 44 | #include "tree.h" |
c7131fb2 | 45 | #include "gimple.h" |
957060b5 AM |
46 | #include "predict.h" |
47 | #include "alloc-pool.h" | |
48 | #include "tree-pass.h" | |
49 | #include "cgraph.h" | |
50 | #include "data-streamer.h" | |
5be5c238 | 51 | #include "gimple-iterator.h" |
08f835dc | 52 | #include "ipa-utils.h" |
08f835dc | 53 | #include "profile.h" |
08f835dc | 54 | #include "value-prof.h" |
08f835dc | 55 | #include "tree-inline.h" |
dd912cb8 | 56 | #include "symbol-summary.h" |
8bc5448f | 57 | #include "tree-vrp.h" |
c582198b | 58 | #include "ipa-prop.h" |
27d020cf | 59 | #include "ipa-fnsummary.h" |
08f835dc JH |
60 | |
61 | /* Entry in the histogram. */ | |
62 | ||
63 | struct histogram_entry | |
64 | { | |
65 | gcov_type count; | |
66 | int time; | |
67 | int size; | |
68 | }; | |
69 | ||
70 | /* Histogram of profile values. | |
71 | The histogram is represented as an ordered vector of entries allocated via | |
72 | histogram_pool. During construction a separate hashtable is kept to lookup | |
73 | duplicate entries. */ | |
74 | ||
75 | vec<histogram_entry *> histogram; | |
fcb87c50 | 76 | static object_allocator<histogram_entry> histogram_pool ("IPA histogram"); |
08f835dc JH |
77 | |
78 | /* Hashtable support for storing SSA names hashed by their SSA_NAME_VAR. */ | |
79 | ||
8d67ee55 | 80 | struct histogram_hash : nofree_ptr_hash <histogram_entry> |
08f835dc | 81 | { |
67f58944 TS |
82 | static inline hashval_t hash (const histogram_entry *); |
83 | static inline int equal (const histogram_entry *, const histogram_entry *); | |
08f835dc JH |
84 | }; |
85 | ||
86 | inline hashval_t | |
87 | histogram_hash::hash (const histogram_entry *val) | |
88 | { | |
89 | return val->count; | |
90 | } | |
91 | ||
92 | inline int | |
93 | histogram_hash::equal (const histogram_entry *val, const histogram_entry *val2) | |
94 | { | |
95 | return val->count == val2->count; | |
96 | } | |
97 | ||
98 | /* Account TIME and SIZE executed COUNT times into HISTOGRAM. | |
99 | HASHTABLE is the on-side hash kept to avoid duplicates. */ | |
100 | ||
101 | static void | |
c203e8a7 | 102 | account_time_size (hash_table<histogram_hash> *hashtable, |
08f835dc JH |
103 | vec<histogram_entry *> &histogram, |
104 | gcov_type count, int time, int size) | |
105 | { | |
106 | histogram_entry key = {count, 0, 0}; | |
c203e8a7 | 107 | histogram_entry **val = hashtable->find_slot (&key, INSERT); |
08f835dc JH |
108 | |
109 | if (!*val) | |
110 | { | |
d7809518 | 111 | *val = histogram_pool.allocate (); |
08f835dc JH |
112 | **val = key; |
113 | histogram.safe_push (*val); | |
114 | } | |
115 | (*val)->time += time; | |
116 | (*val)->size += size; | |
117 | } | |
118 | ||
119 | int | |
120 | cmp_counts (const void *v1, const void *v2) | |
121 | { | |
122 | const histogram_entry *h1 = *(const histogram_entry * const *)v1; | |
123 | const histogram_entry *h2 = *(const histogram_entry * const *)v2; | |
124 | if (h1->count < h2->count) | |
125 | return 1; | |
126 | if (h1->count > h2->count) | |
127 | return -1; | |
128 | return 0; | |
129 | } | |
130 | ||
131 | /* Dump HISTOGRAM to FILE. */ | |
132 | ||
133 | static void | |
134 | dump_histogram (FILE *file, vec<histogram_entry *> histogram) | |
135 | { | |
136 | unsigned int i; | |
137 | gcov_type overall_time = 0, cumulated_time = 0, cumulated_size = 0, overall_size = 0; | |
138 | ||
139 | fprintf (dump_file, "Histogram:\n"); | |
140 | for (i = 0; i < histogram.length (); i++) | |
141 | { | |
142 | overall_time += histogram[i]->count * histogram[i]->time; | |
143 | overall_size += histogram[i]->size; | |
144 | } | |
145 | if (!overall_time) | |
146 | overall_time = 1; | |
147 | if (!overall_size) | |
148 | overall_size = 1; | |
149 | for (i = 0; i < histogram.length (); i++) | |
150 | { | |
151 | cumulated_time += histogram[i]->count * histogram[i]->time; | |
152 | cumulated_size += histogram[i]->size; | |
16998094 | 153 | fprintf (file, " %" PRId64": time:%i (%2.2f) size:%i (%2.2f)\n", |
a9243bfc | 154 | (int64_t) histogram[i]->count, |
08f835dc JH |
155 | histogram[i]->time, |
156 | cumulated_time * 100.0 / overall_time, | |
157 | histogram[i]->size, | |
158 | cumulated_size * 100.0 / overall_size); | |
159 | } | |
160 | } | |
161 | ||
f1ba88b1 XHL |
162 | /* Structure containing speculative target information from profile. */ |
163 | ||
164 | struct speculative_call_target | |
165 | { | |
166 | speculative_call_target (unsigned int id = 0, int prob = 0) | |
167 | : target_id (id), target_probability (prob) | |
168 | { | |
169 | } | |
170 | ||
171 | /* Profile_id of target obtained from profile. */ | |
172 | unsigned int target_id; | |
173 | /* Probability that call will land in function with target_id. */ | |
174 | unsigned int target_probability; | |
175 | }; | |
176 | ||
177 | class speculative_call_summary | |
178 | { | |
179 | public: | |
180 | speculative_call_summary () : speculative_call_targets () | |
181 | {} | |
182 | ||
183 | auto_vec<speculative_call_target> speculative_call_targets; | |
184 | ||
185 | void dump (FILE *f); | |
186 | ||
187 | }; | |
188 | ||
189 | /* Class to manage call summaries. */ | |
190 | ||
191 | class ipa_profile_call_summaries | |
192 | : public call_summary<speculative_call_summary *> | |
193 | { | |
194 | public: | |
195 | ipa_profile_call_summaries (symbol_table *table) | |
196 | : call_summary<speculative_call_summary *> (table) | |
197 | {} | |
198 | ||
199 | /* Duplicate info when an edge is cloned. */ | |
200 | virtual void duplicate (cgraph_edge *, cgraph_edge *, | |
201 | speculative_call_summary *old_sum, | |
202 | speculative_call_summary *new_sum); | |
203 | }; | |
204 | ||
205 | static ipa_profile_call_summaries *call_sums = NULL; | |
206 | ||
207 | /* Dump all information in speculative call summary to F. */ | |
208 | ||
209 | void | |
210 | speculative_call_summary::dump (FILE *f) | |
211 | { | |
212 | cgraph_node *n2; | |
213 | ||
214 | unsigned spec_count = speculative_call_targets.length (); | |
215 | for (unsigned i = 0; i < spec_count; i++) | |
216 | { | |
217 | speculative_call_target item = speculative_call_targets[i]; | |
218 | n2 = find_func_by_profile_id (item.target_id); | |
219 | if (n2) | |
220 | fprintf (f, " The %i speculative target is %s with prob %3.2f\n", i, | |
221 | n2->dump_name (), | |
222 | item.target_probability / (float) REG_BR_PROB_BASE); | |
223 | else | |
224 | fprintf (f, " The %i speculative target is %u with prob %3.2f\n", i, | |
225 | item.target_id, | |
226 | item.target_probability / (float) REG_BR_PROB_BASE); | |
227 | } | |
228 | } | |
229 | ||
230 | /* Duplicate info when an edge is cloned. */ | |
231 | ||
232 | void | |
233 | ipa_profile_call_summaries::duplicate (cgraph_edge *, cgraph_edge *, | |
234 | speculative_call_summary *old_sum, | |
235 | speculative_call_summary *new_sum) | |
236 | { | |
237 | if (!old_sum) | |
238 | return; | |
239 | ||
240 | unsigned old_count = old_sum->speculative_call_targets.length (); | |
241 | if (!old_count) | |
242 | return; | |
243 | ||
244 | new_sum->speculative_call_targets.reserve_exact (old_count); | |
245 | new_sum->speculative_call_targets.quick_grow_cleared (old_count); | |
246 | ||
247 | for (unsigned i = 0; i < old_count; i++) | |
248 | { | |
249 | new_sum->speculative_call_targets[i] | |
250 | = old_sum->speculative_call_targets[i]; | |
251 | } | |
252 | } | |
253 | ||
254 | /* Collect histogram and speculative target summaries from CFG profiles. */ | |
08f835dc JH |
255 | |
256 | static void | |
257 | ipa_profile_generate_summary (void) | |
258 | { | |
259 | struct cgraph_node *node; | |
260 | gimple_stmt_iterator gsi; | |
08f835dc JH |
261 | basic_block bb; |
262 | ||
c203e8a7 | 263 | hash_table<histogram_hash> hashtable (10); |
f1ba88b1 XHL |
264 | |
265 | gcc_checking_assert (!call_sums); | |
266 | call_sums = new ipa_profile_call_summaries (symtab); | |
267 | ||
08f835dc | 268 | FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node) |
e7a74006 JH |
269 | if (ENTRY_BLOCK_PTR_FOR_FN (DECL_STRUCT_FUNCTION (node->decl))->count.ipa_p ()) |
270 | FOR_EACH_BB_FN (bb, DECL_STRUCT_FUNCTION (node->decl)) | |
271 | { | |
272 | int time = 0; | |
273 | int size = 0; | |
274 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
275 | { | |
276 | gimple *stmt = gsi_stmt (gsi); | |
277 | if (gimple_code (stmt) == GIMPLE_CALL | |
278 | && !gimple_call_fndecl (stmt)) | |
279 | { | |
280 | histogram_value h; | |
281 | h = gimple_histogram_value_of_type | |
282 | (DECL_STRUCT_FUNCTION (node->decl), | |
283 | stmt, HIST_TYPE_INDIR_CALL); | |
284 | /* No need to do sanity check: gimple_ic_transform already | |
285 | takes away bad histograms. */ | |
286 | if (h) | |
287 | { | |
92d41717 | 288 | gcov_type val, count, all; |
f1ba88b1 XHL |
289 | struct cgraph_edge *e = node->get_edge (stmt); |
290 | if (e && !e->indirect_unknown_callee) | |
291 | continue; | |
292 | ||
293 | speculative_call_summary *csum | |
294 | = call_sums->get_create (e); | |
295 | ||
296 | for (unsigned j = 0; j < GCOV_TOPN_VALUES; j++) | |
e7a74006 | 297 | { |
f1ba88b1 XHL |
298 | if (!get_nth_most_common_value (NULL, "indirect call", |
299 | h, &val, &count, &all, | |
300 | j)) | |
e7a74006 | 301 | continue; |
92d41717 | 302 | |
f1ba88b1 XHL |
303 | if (val == 0) |
304 | continue; | |
305 | ||
306 | speculative_call_target item ( | |
307 | val, GCOV_COMPUTE_SCALE (count, all)); | |
308 | if (item.target_probability > REG_BR_PROB_BASE) | |
e7a74006 JH |
309 | { |
310 | if (dump_file) | |
f1ba88b1 XHL |
311 | fprintf (dump_file, |
312 | "Probability capped to 1\n"); | |
313 | item.target_probability = REG_BR_PROB_BASE; | |
e7a74006 | 314 | } |
f1ba88b1 | 315 | csum->speculative_call_targets.safe_push (item); |
e7a74006 | 316 | } |
f1ba88b1 | 317 | |
e7a74006 JH |
318 | gimple_remove_histogram_value (DECL_STRUCT_FUNCTION (node->decl), |
319 | stmt, h); | |
320 | } | |
321 | } | |
322 | time += estimate_num_insns (stmt, &eni_time_weights); | |
323 | size += estimate_num_insns (stmt, &eni_size_weights); | |
324 | } | |
325 | if (bb->count.ipa_p () && bb->count.initialized_p ()) | |
326 | account_time_size (&hashtable, histogram, bb->count.ipa ().to_gcov_type (), | |
327 | time, size); | |
328 | } | |
08f835dc JH |
329 | histogram.qsort (cmp_counts); |
330 | } | |
331 | ||
f1ba88b1 XHL |
332 | /* Serialize the speculative summary info for LTO. */ |
333 | ||
334 | static void | |
335 | ipa_profile_write_edge_summary (lto_simple_output_block *ob, | |
336 | speculative_call_summary *csum) | |
337 | { | |
338 | unsigned len = 0; | |
339 | ||
340 | len = csum->speculative_call_targets.length (); | |
341 | ||
342 | gcc_assert (len <= GCOV_TOPN_VALUES); | |
343 | ||
344 | streamer_write_hwi_stream (ob->main_stream, len); | |
345 | ||
346 | if (len) | |
347 | { | |
348 | unsigned spec_count = csum->speculative_call_targets.length (); | |
349 | for (unsigned i = 0; i < spec_count; i++) | |
350 | { | |
351 | speculative_call_target item = csum->speculative_call_targets[i]; | |
352 | gcc_assert (item.target_id); | |
353 | streamer_write_hwi_stream (ob->main_stream, item.target_id); | |
354 | streamer_write_hwi_stream (ob->main_stream, item.target_probability); | |
355 | } | |
356 | } | |
357 | } | |
358 | ||
08f835dc JH |
359 | /* Serialize the ipa info for lto. */ |
360 | ||
361 | static void | |
362 | ipa_profile_write_summary (void) | |
363 | { | |
364 | struct lto_simple_output_block *ob | |
365 | = lto_create_simple_output_block (LTO_section_ipa_profile); | |
366 | unsigned int i; | |
367 | ||
c3284718 | 368 | streamer_write_uhwi_stream (ob->main_stream, histogram.length ()); |
08f835dc JH |
369 | for (i = 0; i < histogram.length (); i++) |
370 | { | |
371 | streamer_write_gcov_count_stream (ob->main_stream, histogram[i]->count); | |
372 | streamer_write_uhwi_stream (ob->main_stream, histogram[i]->time); | |
373 | streamer_write_uhwi_stream (ob->main_stream, histogram[i]->size); | |
374 | } | |
f1ba88b1 XHL |
375 | |
376 | if (!call_sums) | |
377 | return; | |
378 | ||
379 | /* Serialize speculative targets information. */ | |
380 | unsigned int count = 0; | |
381 | lto_symtab_encoder_t encoder = ob->decl_state->symtab_node_encoder; | |
382 | lto_symtab_encoder_iterator lsei; | |
383 | cgraph_node *node; | |
384 | ||
385 | for (lsei = lsei_start_function_in_partition (encoder); !lsei_end_p (lsei); | |
386 | lsei_next_function_in_partition (&lsei)) | |
387 | { | |
388 | node = lsei_cgraph_node (lsei); | |
389 | if (node->definition && node->has_gimple_body_p () | |
390 | && node->indirect_calls) | |
391 | count++; | |
392 | } | |
393 | ||
394 | streamer_write_uhwi_stream (ob->main_stream, count); | |
395 | ||
396 | /* Process all of the functions. */ | |
397 | for (lsei = lsei_start_function_in_partition (encoder); | |
398 | !lsei_end_p (lsei) && count; lsei_next_function_in_partition (&lsei)) | |
399 | { | |
400 | cgraph_node *node = lsei_cgraph_node (lsei); | |
401 | if (node->definition && node->has_gimple_body_p () | |
402 | && node->indirect_calls) | |
403 | { | |
404 | int node_ref = lto_symtab_encoder_encode (encoder, node); | |
405 | streamer_write_uhwi_stream (ob->main_stream, node_ref); | |
406 | ||
407 | for (cgraph_edge *e = node->indirect_calls; e; e = e->next_callee) | |
408 | { | |
409 | speculative_call_summary *csum = call_sums->get_create (e); | |
410 | ipa_profile_write_edge_summary (ob, csum); | |
411 | } | |
412 | } | |
413 | } | |
414 | ||
08f835dc JH |
415 | lto_destroy_simple_output_block (ob); |
416 | } | |
417 | ||
f1ba88b1 XHL |
418 | /* Dump all profile summary data for all cgraph nodes and edges to file F. */ |
419 | ||
420 | static void | |
421 | ipa_profile_dump_all_summaries (FILE *f) | |
422 | { | |
423 | fprintf (dump_file, | |
424 | "\n========== IPA-profile speculative targets: ==========\n"); | |
425 | cgraph_node *node; | |
426 | FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node) | |
427 | { | |
428 | fprintf (f, "\nSummary for node %s:\n", node->dump_name ()); | |
429 | for (cgraph_edge *e = node->indirect_calls; e; e = e->next_callee) | |
430 | { | |
431 | fprintf (f, " Summary for %s of indirect edge %d:\n", | |
432 | e->caller->dump_name (), e->lto_stmt_uid); | |
433 | speculative_call_summary *csum = call_sums->get_create (e); | |
434 | csum->dump (f); | |
435 | } | |
436 | } | |
437 | fprintf (f, "\n\n"); | |
438 | } | |
439 | ||
440 | /* Read speculative targets information about edge for LTO WPA. */ | |
441 | ||
442 | static void | |
443 | ipa_profile_read_edge_summary (class lto_input_block *ib, cgraph_edge *edge) | |
444 | { | |
445 | unsigned i, len; | |
446 | ||
447 | len = streamer_read_hwi (ib); | |
448 | gcc_assert (len <= GCOV_TOPN_VALUES); | |
449 | ||
450 | speculative_call_summary *csum = call_sums->get_create (edge); | |
451 | ||
452 | for (i = 0; i < len; i++) | |
453 | { | |
84a3effa ML |
454 | unsigned int target_id = streamer_read_hwi (ib); |
455 | int target_probability = streamer_read_hwi (ib); | |
456 | speculative_call_target item (target_id, target_probability); | |
f1ba88b1 XHL |
457 | csum->speculative_call_targets.safe_push (item); |
458 | } | |
459 | } | |
460 | ||
461 | /* Read profile speculative targets section information for LTO WPA. */ | |
462 | ||
463 | static void | |
464 | ipa_profile_read_summary_section (struct lto_file_decl_data *file_data, | |
465 | class lto_input_block *ib) | |
466 | { | |
467 | if (!ib) | |
468 | return; | |
469 | ||
470 | lto_symtab_encoder_t encoder = file_data->symtab_node_encoder; | |
471 | ||
472 | unsigned int count = streamer_read_uhwi (ib); | |
473 | ||
474 | unsigned int i; | |
475 | unsigned int index; | |
476 | cgraph_node * node; | |
477 | ||
478 | for (i = 0; i < count; i++) | |
479 | { | |
480 | index = streamer_read_uhwi (ib); | |
481 | encoder = file_data->symtab_node_encoder; | |
482 | node | |
483 | = dyn_cast<cgraph_node *> (lto_symtab_encoder_deref (encoder, index)); | |
484 | ||
485 | for (cgraph_edge *e = node->indirect_calls; e; e = e->next_callee) | |
486 | ipa_profile_read_edge_summary (ib, e); | |
487 | } | |
488 | } | |
489 | ||
490 | /* Deserialize the IPA histogram and speculative targets summary info for LTO. | |
491 | */ | |
08f835dc JH |
492 | |
493 | static void | |
494 | ipa_profile_read_summary (void) | |
495 | { | |
496 | struct lto_file_decl_data ** file_data_vec | |
497 | = lto_get_file_decl_data (); | |
498 | struct lto_file_decl_data * file_data; | |
08f835dc JH |
499 | int j = 0; |
500 | ||
c203e8a7 | 501 | hash_table<histogram_hash> hashtable (10); |
08f835dc | 502 | |
f1ba88b1 XHL |
503 | gcc_checking_assert (!call_sums); |
504 | call_sums = new ipa_profile_call_summaries (symtab); | |
505 | ||
08f835dc JH |
506 | while ((file_data = file_data_vec[j++])) |
507 | { | |
508 | const char *data; | |
509 | size_t len; | |
99b1c316 | 510 | class lto_input_block *ib |
08f835dc JH |
511 | = lto_create_simple_input_block (file_data, |
512 | LTO_section_ipa_profile, | |
513 | &data, &len); | |
514 | if (ib) | |
515 | { | |
516 | unsigned int num = streamer_read_uhwi (ib); | |
517 | unsigned int n; | |
518 | for (n = 0; n < num; n++) | |
519 | { | |
520 | gcov_type count = streamer_read_gcov_count (ib); | |
521 | int time = streamer_read_uhwi (ib); | |
522 | int size = streamer_read_uhwi (ib); | |
c203e8a7 | 523 | account_time_size (&hashtable, histogram, |
08f835dc JH |
524 | count, time, size); |
525 | } | |
f1ba88b1 XHL |
526 | |
527 | ipa_profile_read_summary_section (file_data, ib); | |
528 | ||
08f835dc JH |
529 | lto_destroy_simple_input_block (file_data, |
530 | LTO_section_ipa_profile, | |
531 | ib, data, len); | |
532 | } | |
533 | } | |
08f835dc JH |
534 | histogram.qsort (cmp_counts); |
535 | } | |
536 | ||
537 | /* Data used by ipa_propagate_frequency. */ | |
538 | ||
539 | struct ipa_propagate_frequency_data | |
540 | { | |
1ede94c5 | 541 | cgraph_node *function_symbol; |
08f835dc JH |
542 | bool maybe_unlikely_executed; |
543 | bool maybe_executed_once; | |
544 | bool only_called_at_startup; | |
545 | bool only_called_at_exit; | |
546 | }; | |
547 | ||
548 | /* Worker for ipa_propagate_frequency_1. */ | |
549 | ||
550 | static bool | |
551 | ipa_propagate_frequency_1 (struct cgraph_node *node, void *data) | |
552 | { | |
553 | struct ipa_propagate_frequency_data *d; | |
554 | struct cgraph_edge *edge; | |
555 | ||
556 | d = (struct ipa_propagate_frequency_data *)data; | |
557 | for (edge = node->callers; | |
558 | edge && (d->maybe_unlikely_executed || d->maybe_executed_once | |
559 | || d->only_called_at_startup || d->only_called_at_exit); | |
560 | edge = edge->next_caller) | |
561 | { | |
1ede94c5 | 562 | if (edge->caller != d->function_symbol) |
08f835dc JH |
563 | { |
564 | d->only_called_at_startup &= edge->caller->only_called_at_startup; | |
565 | /* It makes sense to put main() together with the static constructors. | |
566 | It will be executed for sure, but rest of functions called from | |
567 | main are definitely not at startup only. */ | |
67348ccc | 568 | if (MAIN_NAME_P (DECL_NAME (edge->caller->decl))) |
08f835dc JH |
569 | d->only_called_at_startup = 0; |
570 | d->only_called_at_exit &= edge->caller->only_called_at_exit; | |
571 | } | |
daf5c770 JH |
572 | |
573 | /* When profile feedback is available, do not try to propagate too hard; | |
574 | counts are already good guide on function frequencies and roundoff | |
575 | errors can make us to push function into unlikely section even when | |
576 | it is executed by the train run. Transfer the function only if all | |
577 | callers are unlikely executed. */ | |
1ede94c5 | 578 | if (profile_info |
35d93d1d | 579 | && !(edge->callee->count.ipa () == profile_count::zero ()) |
daf5c770 | 580 | && (edge->caller->frequency != NODE_FREQUENCY_UNLIKELY_EXECUTED |
a62bfab5 ML |
581 | || (edge->caller->inlined_to |
582 | && edge->caller->inlined_to->frequency | |
daf5c770 JH |
583 | != NODE_FREQUENCY_UNLIKELY_EXECUTED))) |
584 | d->maybe_unlikely_executed = false; | |
35d93d1d JH |
585 | if (edge->count.ipa ().initialized_p () |
586 | && !edge->count.ipa ().nonzero_p ()) | |
08f835dc JH |
587 | continue; |
588 | switch (edge->caller->frequency) | |
589 | { | |
590 | case NODE_FREQUENCY_UNLIKELY_EXECUTED: | |
591 | break; | |
592 | case NODE_FREQUENCY_EXECUTED_ONCE: | |
56f62793 ML |
593 | { |
594 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
595 | fprintf (dump_file, " Called by %s that is executed once\n", | |
3629ff8a | 596 | edge->caller->dump_name ()); |
56f62793 ML |
597 | d->maybe_unlikely_executed = false; |
598 | ipa_call_summary *s = ipa_call_summaries->get (edge); | |
599 | if (s != NULL && s->loop_depth) | |
600 | { | |
601 | d->maybe_executed_once = false; | |
602 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
603 | fprintf (dump_file, " Called in loop\n"); | |
604 | } | |
605 | break; | |
606 | } | |
08f835dc JH |
607 | case NODE_FREQUENCY_HOT: |
608 | case NODE_FREQUENCY_NORMAL: | |
609 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
610 | fprintf (dump_file, " Called by %s that is normal or hot\n", | |
3629ff8a | 611 | edge->caller->dump_name ()); |
08f835dc JH |
612 | d->maybe_unlikely_executed = false; |
613 | d->maybe_executed_once = false; | |
614 | break; | |
615 | } | |
616 | } | |
617 | return edge != NULL; | |
618 | } | |
619 | ||
daf5c770 JH |
620 | /* Return ture if NODE contains hot calls. */ |
621 | ||
622 | bool | |
623 | contains_hot_call_p (struct cgraph_node *node) | |
624 | { | |
625 | struct cgraph_edge *e; | |
626 | for (e = node->callees; e; e = e->next_callee) | |
3dafb85c | 627 | if (e->maybe_hot_p ()) |
daf5c770 JH |
628 | return true; |
629 | else if (!e->inline_failed | |
630 | && contains_hot_call_p (e->callee)) | |
631 | return true; | |
632 | for (e = node->indirect_calls; e; e = e->next_callee) | |
3dafb85c | 633 | if (e->maybe_hot_p ()) |
daf5c770 JH |
634 | return true; |
635 | return false; | |
636 | } | |
637 | ||
08f835dc JH |
638 | /* See if the frequency of NODE can be updated based on frequencies of its |
639 | callers. */ | |
640 | bool | |
641 | ipa_propagate_frequency (struct cgraph_node *node) | |
642 | { | |
1ede94c5 | 643 | struct ipa_propagate_frequency_data d = {node, true, true, true, true}; |
08f835dc JH |
644 | bool changed = false; |
645 | ||
67914693 | 646 | /* We cannot propagate anything useful about externally visible functions |
08f835dc | 647 | nor about virtuals. */ |
87f94429 | 648 | if (!node->local |
67348ccc | 649 | || node->alias |
2bf86c84 JH |
650 | || (opt_for_fn (node->decl, flag_devirtualize) |
651 | && DECL_VIRTUAL_P (node->decl))) | |
08f835dc | 652 | return false; |
67348ccc | 653 | gcc_assert (node->analyzed); |
08f835dc | 654 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3629ff8a | 655 | fprintf (dump_file, "Processing frequency %s\n", node->dump_name ()); |
08f835dc | 656 | |
1ede94c5 JH |
657 | node->call_for_symbol_and_aliases (ipa_propagate_frequency_1, &d, |
658 | true); | |
08f835dc JH |
659 | |
660 | if ((d.only_called_at_startup && !d.only_called_at_exit) | |
661 | && !node->only_called_at_startup) | |
662 | { | |
663 | node->only_called_at_startup = true; | |
664 | if (dump_file) | |
665 | fprintf (dump_file, "Node %s promoted to only called at startup.\n", | |
3629ff8a | 666 | node->dump_name ()); |
08f835dc JH |
667 | changed = true; |
668 | } | |
669 | if ((d.only_called_at_exit && !d.only_called_at_startup) | |
670 | && !node->only_called_at_exit) | |
671 | { | |
672 | node->only_called_at_exit = true; | |
673 | if (dump_file) | |
674 | fprintf (dump_file, "Node %s promoted to only called at exit.\n", | |
3629ff8a | 675 | node->dump_name ()); |
08f835dc JH |
676 | changed = true; |
677 | } | |
daf5c770 JH |
678 | |
679 | /* With profile we can decide on hot/normal based on count. */ | |
1bad9c18 | 680 | if (node->count. ipa().initialized_p ()) |
daf5c770 JH |
681 | { |
682 | bool hot = false; | |
1bad9c18 JH |
683 | if (!(node->count. ipa() == profile_count::zero ()) |
684 | && node->count. ipa() >= get_hot_bb_threshold ()) | |
daf5c770 JH |
685 | hot = true; |
686 | if (!hot) | |
687 | hot |= contains_hot_call_p (node); | |
688 | if (hot) | |
689 | { | |
690 | if (node->frequency != NODE_FREQUENCY_HOT) | |
691 | { | |
692 | if (dump_file) | |
693 | fprintf (dump_file, "Node %s promoted to hot.\n", | |
3629ff8a | 694 | node->dump_name ()); |
daf5c770 JH |
695 | node->frequency = NODE_FREQUENCY_HOT; |
696 | return true; | |
697 | } | |
698 | return false; | |
699 | } | |
700 | else if (node->frequency == NODE_FREQUENCY_HOT) | |
701 | { | |
702 | if (dump_file) | |
703 | fprintf (dump_file, "Node %s reduced to normal.\n", | |
3629ff8a | 704 | node->dump_name ()); |
daf5c770 JH |
705 | node->frequency = NODE_FREQUENCY_NORMAL; |
706 | changed = true; | |
707 | } | |
708 | } | |
08f835dc JH |
709 | /* These come either from profile or user hints; never update them. */ |
710 | if (node->frequency == NODE_FREQUENCY_HOT | |
711 | || node->frequency == NODE_FREQUENCY_UNLIKELY_EXECUTED) | |
712 | return changed; | |
713 | if (d.maybe_unlikely_executed) | |
714 | { | |
715 | node->frequency = NODE_FREQUENCY_UNLIKELY_EXECUTED; | |
716 | if (dump_file) | |
717 | fprintf (dump_file, "Node %s promoted to unlikely executed.\n", | |
3629ff8a | 718 | node->dump_name ()); |
08f835dc JH |
719 | changed = true; |
720 | } | |
721 | else if (d.maybe_executed_once && node->frequency != NODE_FREQUENCY_EXECUTED_ONCE) | |
722 | { | |
723 | node->frequency = NODE_FREQUENCY_EXECUTED_ONCE; | |
724 | if (dump_file) | |
725 | fprintf (dump_file, "Node %s promoted to executed once.\n", | |
3629ff8a | 726 | node->dump_name ()); |
08f835dc JH |
727 | changed = true; |
728 | } | |
729 | return changed; | |
730 | } | |
731 | ||
d200a49f JH |
732 | /* Check that number of arguments of N agrees with E. |
733 | Be conservative when summaries are not present. */ | |
734 | ||
735 | static bool | |
736 | check_argument_count (struct cgraph_node *n, struct cgraph_edge *e) | |
737 | { | |
738 | if (!ipa_node_params_sum || !ipa_edge_args_sum) | |
739 | return true; | |
740 | class ipa_node_params *info = IPA_NODE_REF (n->function_symbol ()); | |
741 | if (!info) | |
742 | return true; | |
743 | ipa_edge_args *e_info = IPA_EDGE_REF (e); | |
d772e360 | 744 | if (!e_info) |
d200a49f JH |
745 | return true; |
746 | if (ipa_get_param_count (info) != ipa_get_cs_argument_count (e_info) | |
747 | && (ipa_get_param_count (info) >= ipa_get_cs_argument_count (e_info) | |
748 | || !stdarg_p (TREE_TYPE (n->decl)))) | |
749 | return false; | |
750 | return true; | |
751 | } | |
752 | ||
08f835dc JH |
753 | /* Simple ipa profile pass propagating frequencies across the callgraph. */ |
754 | ||
755 | static unsigned int | |
756 | ipa_profile (void) | |
757 | { | |
758 | struct cgraph_node **order; | |
759 | struct cgraph_edge *e; | |
760 | int order_pos; | |
761 | bool something_changed = false; | |
762 | int i; | |
763 | gcov_type overall_time = 0, cutoff = 0, cumulated = 0, overall_size = 0; | |
764 | struct cgraph_node *n,*n2; | |
765 | int nindirect = 0, ncommon = 0, nunknown = 0, nuseless = 0, nconverted = 0; | |
95d81ba5 | 766 | int nmismatch = 0, nimpossible = 0; |
08f835dc | 767 | bool node_map_initialized = false; |
f1ba88b1 | 768 | gcov_type threshold; |
08f835dc JH |
769 | |
770 | if (dump_file) | |
771 | dump_histogram (dump_file, histogram); | |
772 | for (i = 0; i < (int)histogram.length (); i++) | |
773 | { | |
774 | overall_time += histogram[i]->count * histogram[i]->time; | |
775 | overall_size += histogram[i]->size; | |
776 | } | |
f1ba88b1 | 777 | threshold = 0; |
08f835dc JH |
778 | if (overall_time) |
779 | { | |
08f835dc | 780 | gcc_assert (overall_size); |
08f835dc | 781 | |
028d4092 | 782 | cutoff = (overall_time * param_hot_bb_count_ws_permille + 500) / 1000; |
08f835dc JH |
783 | for (i = 0; cumulated < cutoff; i++) |
784 | { | |
785 | cumulated += histogram[i]->count * histogram[i]->time; | |
786 | threshold = histogram[i]->count; | |
787 | } | |
788 | if (!threshold) | |
789 | threshold = 1; | |
790 | if (dump_file) | |
791 | { | |
792 | gcov_type cumulated_time = 0, cumulated_size = 0; | |
793 | ||
794 | for (i = 0; | |
795 | i < (int)histogram.length () && histogram[i]->count >= threshold; | |
796 | i++) | |
797 | { | |
798 | cumulated_time += histogram[i]->count * histogram[i]->time; | |
799 | cumulated_size += histogram[i]->size; | |
800 | } | |
16998094 | 801 | fprintf (dump_file, "Determined min count: %" PRId64 |
08f835dc | 802 | " Time:%3.2f%% Size:%3.2f%%\n", |
a9243bfc | 803 | (int64_t)threshold, |
08f835dc JH |
804 | cumulated_time * 100.0 / overall_time, |
805 | cumulated_size * 100.0 / overall_size); | |
806 | } | |
512cc015 | 807 | |
e53f77c6 | 808 | if (in_lto_p) |
08f835dc JH |
809 | { |
810 | if (dump_file) | |
e53f77c6 | 811 | fprintf (dump_file, "Setting hotness threshold in LTO mode.\n"); |
08f835dc JH |
812 | set_hot_bb_threshold (threshold); |
813 | } | |
814 | } | |
c3284718 | 815 | histogram.release (); |
d7809518 | 816 | histogram_pool.release (); |
08f835dc | 817 | |
f1ba88b1 XHL |
818 | /* Produce speculative calls: we saved common target from profiling into |
819 | e->target_id. Now, at link time, we can look up corresponding | |
08f835dc JH |
820 | function node and produce speculative call. */ |
821 | ||
f1ba88b1 XHL |
822 | gcc_checking_assert (call_sums); |
823 | ||
824 | if (dump_file) | |
825 | { | |
826 | if (!node_map_initialized) | |
827 | init_node_map (false); | |
828 | node_map_initialized = true; | |
829 | ||
830 | ipa_profile_dump_all_summaries (dump_file); | |
831 | } | |
832 | ||
08f835dc JH |
833 | FOR_EACH_DEFINED_FUNCTION (n) |
834 | { | |
835 | bool update = false; | |
836 | ||
1ede94c5 JH |
837 | if (!opt_for_fn (n->decl, flag_ipa_profile)) |
838 | continue; | |
839 | ||
08f835dc JH |
840 | for (e = n->indirect_calls; e; e = e->next_callee) |
841 | { | |
3995f3a2 | 842 | if (n->count.initialized_p ()) |
08f835dc | 843 | nindirect++; |
f1ba88b1 XHL |
844 | |
845 | speculative_call_summary *csum = call_sums->get_create (e); | |
846 | unsigned spec_count = csum->speculative_call_targets.length (); | |
847 | if (spec_count) | |
08f835dc JH |
848 | { |
849 | if (!node_map_initialized) | |
f1ba88b1 | 850 | init_node_map (false); |
08f835dc JH |
851 | node_map_initialized = true; |
852 | ncommon++; | |
f1ba88b1 XHL |
853 | |
854 | if (in_lto_p) | |
855 | { | |
856 | if (dump_file) | |
857 | { | |
858 | fprintf (dump_file, | |
859 | "Updating hotness threshold in LTO mode.\n"); | |
860 | fprintf (dump_file, "Updated min count: %" PRId64 "\n", | |
861 | (int64_t) threshold / spec_count); | |
862 | } | |
863 | set_hot_bb_threshold (threshold / spec_count); | |
864 | } | |
865 | ||
866 | unsigned speculative_id = 0; | |
867 | bool speculative_found = false; | |
868 | for (unsigned i = 0; i < spec_count; i++) | |
869 | { | |
870 | speculative_call_target item | |
871 | = csum->speculative_call_targets[i]; | |
872 | n2 = find_func_by_profile_id (item.target_id); | |
08f835dc JH |
873 | if (n2) |
874 | { | |
875 | if (dump_file) | |
876 | { | |
877 | fprintf (dump_file, "Indirect call -> direct call from" | |
464d0118 ML |
878 | " other module %s => %s, prob %3.2f\n", |
879 | n->dump_name (), | |
880 | n2->dump_name (), | |
f1ba88b1 XHL |
881 | item.target_probability |
882 | / (float) REG_BR_PROB_BASE); | |
08f835dc | 883 | } |
f1ba88b1 | 884 | if (item.target_probability < REG_BR_PROB_BASE / 2) |
08f835dc JH |
885 | { |
886 | nuseless++; | |
887 | if (dump_file) | |
888 | fprintf (dump_file, | |
889 | "Not speculating: probability is too low.\n"); | |
890 | } | |
3dafb85c | 891 | else if (!e->maybe_hot_p ()) |
08f835dc JH |
892 | { |
893 | nuseless++; | |
894 | if (dump_file) | |
895 | fprintf (dump_file, | |
896 | "Not speculating: call is cold.\n"); | |
897 | } | |
d52f5295 ML |
898 | else if (n2->get_availability () <= AVAIL_INTERPOSABLE |
899 | && n2->can_be_discarded_p ()) | |
08f835dc JH |
900 | { |
901 | nuseless++; | |
902 | if (dump_file) | |
903 | fprintf (dump_file, | |
904 | "Not speculating: target is overwritable " | |
905 | "and can be discarded.\n"); | |
906 | } | |
7b34a284 | 907 | else if (!check_argument_count (n2, e)) |
95d81ba5 JH |
908 | { |
909 | nmismatch++; | |
910 | if (dump_file) | |
911 | fprintf (dump_file, | |
912 | "Not speculating: " | |
4e8e460b | 913 | "parameter count mismatch\n"); |
95d81ba5 JH |
914 | } |
915 | else if (e->indirect_info->polymorphic | |
916 | && !opt_for_fn (n->decl, flag_devirtualize) | |
917 | && !possible_polymorphic_call_target_p (e, n2)) | |
918 | { | |
919 | nimpossible++; | |
920 | if (dump_file) | |
921 | fprintf (dump_file, | |
922 | "Not speculating: " | |
923 | "function is not in the polymorphic " | |
924 | "call target list\n"); | |
925 | } | |
08f835dc JH |
926 | else |
927 | { | |
928 | /* Target may be overwritable, but profile says that | |
929 | control flow goes to this particular implementation | |
930 | of N2. Speculate on the local alias to allow inlining. | |
931 | */ | |
d52f5295 | 932 | if (!n2->can_be_discarded_p ()) |
5b79657a JH |
933 | { |
934 | cgraph_node *alias; | |
d52f5295 | 935 | alias = dyn_cast<cgraph_node *> (n2->noninterposable_alias ()); |
5b79657a JH |
936 | if (alias) |
937 | n2 = alias; | |
938 | } | |
08f835dc | 939 | nconverted++; |
f1ba88b1 XHL |
940 | e->make_speculative (n2, |
941 | e->count.apply_probability ( | |
942 | item.target_probability), | |
943 | speculative_id, | |
944 | item.target_probability); | |
08f835dc | 945 | update = true; |
f1ba88b1 XHL |
946 | speculative_id++; |
947 | speculative_found = true; | |
08f835dc JH |
948 | } |
949 | } | |
950 | else | |
951 | { | |
952 | if (dump_file) | |
953 | fprintf (dump_file, "Function with profile-id %i not found.\n", | |
f1ba88b1 | 954 | item.target_id); |
08f835dc JH |
955 | nunknown++; |
956 | } | |
f1ba88b1 XHL |
957 | } |
958 | if (speculative_found) | |
959 | e->indirect_info->num_speculative_call_targets = speculative_id; | |
08f835dc JH |
960 | } |
961 | } | |
962 | if (update) | |
0bceb671 | 963 | ipa_update_overall_fn_summary (n); |
08f835dc JH |
964 | } |
965 | if (node_map_initialized) | |
966 | del_node_map (); | |
967 | if (dump_file && nindirect) | |
968 | fprintf (dump_file, | |
969 | "%i indirect calls trained.\n" | |
970 | "%i (%3.2f%%) have common target.\n" | |
971 | "%i (%3.2f%%) targets was not found.\n" | |
95d81ba5 JH |
972 | "%i (%3.2f%%) targets had parameter count mismatch.\n" |
973 | "%i (%3.2f%%) targets was not in polymorphic call target list.\n" | |
08f835dc JH |
974 | "%i (%3.2f%%) speculations seems useless.\n" |
975 | "%i (%3.2f%%) speculations produced.\n", | |
976 | nindirect, | |
977 | ncommon, ncommon * 100.0 / nindirect, | |
978 | nunknown, nunknown * 100.0 / nindirect, | |
95d81ba5 JH |
979 | nmismatch, nmismatch * 100.0 / nindirect, |
980 | nimpossible, nimpossible * 100.0 / nindirect, | |
08f835dc JH |
981 | nuseless, nuseless * 100.0 / nindirect, |
982 | nconverted, nconverted * 100.0 / nindirect); | |
983 | ||
3dafb85c | 984 | order = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count); |
08f835dc JH |
985 | order_pos = ipa_reverse_postorder (order); |
986 | for (i = order_pos - 1; i >= 0; i--) | |
987 | { | |
87f94429 | 988 | if (order[i]->local |
1ede94c5 JH |
989 | && opt_for_fn (order[i]->decl, flag_ipa_profile) |
990 | && ipa_propagate_frequency (order[i])) | |
08f835dc JH |
991 | { |
992 | for (e = order[i]->callees; e; e = e->next_callee) | |
87f94429 | 993 | if (e->callee->local && !e->callee->aux) |
08f835dc JH |
994 | { |
995 | something_changed = true; | |
67348ccc | 996 | e->callee->aux = (void *)1; |
08f835dc JH |
997 | } |
998 | } | |
67348ccc | 999 | order[i]->aux = NULL; |
08f835dc JH |
1000 | } |
1001 | ||
1002 | while (something_changed) | |
1003 | { | |
1004 | something_changed = false; | |
1005 | for (i = order_pos - 1; i >= 0; i--) | |
1006 | { | |
1ede94c5 JH |
1007 | if (order[i]->aux |
1008 | && opt_for_fn (order[i]->decl, flag_ipa_profile) | |
1009 | && ipa_propagate_frequency (order[i])) | |
08f835dc JH |
1010 | { |
1011 | for (e = order[i]->callees; e; e = e->next_callee) | |
87f94429 | 1012 | if (e->callee->local && !e->callee->aux) |
08f835dc JH |
1013 | { |
1014 | something_changed = true; | |
67348ccc | 1015 | e->callee->aux = (void *)1; |
08f835dc JH |
1016 | } |
1017 | } | |
67348ccc | 1018 | order[i]->aux = NULL; |
08f835dc JH |
1019 | } |
1020 | } | |
1021 | free (order); | |
f1ba88b1 XHL |
1022 | |
1023 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1024 | symtab->dump (dump_file); | |
1025 | ||
08f835dc JH |
1026 | return 0; |
1027 | } | |
1028 | ||
08f835dc JH |
1029 | namespace { |
1030 | ||
1031 | const pass_data pass_data_ipa_profile = | |
1032 | { | |
1033 | IPA_PASS, /* type */ | |
1034 | "profile_estimate", /* name */ | |
1035 | OPTGROUP_NONE, /* optinfo_flags */ | |
08f835dc JH |
1036 | TV_IPA_PROFILE, /* tv_id */ |
1037 | 0, /* properties_required */ | |
1038 | 0, /* properties_provided */ | |
1039 | 0, /* properties_destroyed */ | |
1040 | 0, /* todo_flags_start */ | |
1041 | 0, /* todo_flags_finish */ | |
1042 | }; | |
1043 | ||
1044 | class pass_ipa_profile : public ipa_opt_pass_d | |
1045 | { | |
1046 | public: | |
c3284718 RS |
1047 | pass_ipa_profile (gcc::context *ctxt) |
1048 | : ipa_opt_pass_d (pass_data_ipa_profile, ctxt, | |
1049 | ipa_profile_generate_summary, /* generate_summary */ | |
1050 | ipa_profile_write_summary, /* write_summary */ | |
1051 | ipa_profile_read_summary, /* read_summary */ | |
1052 | NULL, /* write_optimization_summary */ | |
1053 | NULL, /* read_optimization_summary */ | |
1054 | NULL, /* stmt_fixup */ | |
1055 | 0, /* function_transform_todo_flags_start */ | |
1056 | NULL, /* function_transform */ | |
1057 | NULL) /* variable_transform */ | |
08f835dc JH |
1058 | {} |
1059 | ||
1060 | /* opt_pass methods: */ | |
2bf86c84 | 1061 | virtual bool gate (function *) { return flag_ipa_profile || in_lto_p; } |
be55bfe6 | 1062 | virtual unsigned int execute (function *) { return ipa_profile (); } |
08f835dc JH |
1063 | |
1064 | }; // class pass_ipa_profile | |
1065 | ||
1066 | } // anon namespace | |
1067 | ||
1068 | ipa_opt_pass_d * | |
1069 | make_pass_ipa_profile (gcc::context *ctxt) | |
1070 | { | |
1071 | return new pass_ipa_profile (ctxt); | |
1072 | } |