]>
Commit | Line | Data |
---|---|---|
08f835dc | 1 | /* Basic IPA optimizations based on profile. |
cbe34bb5 | 2 | Copyright (C) 2003-2017 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 | ||
28 | Similar histogram is also estimated by coverage runtime. This histogram | |
29 | is not dependent on LTO, but it suffers from various defects; first | |
30 | gcov runtime is not weighting individual basic block by estimated execution | |
31 | time and second the merging of multiple runs makes assumption that the | |
32 | histogram distribution did not change. Consequentely histogram constructed | |
33 | here may be more precise. | |
34 | ||
35 | The information is used to set hot/cold thresholds. | |
36 | - Next speculative indirect call resolution is performed: the local | |
37 | profile pass assigns profile-id to each function and provide us with a | |
38 | histogram specifying the most common target. We look up the callgraph | |
39 | node corresponding to the target and produce a speculative call. | |
40 | ||
41 | This call may or may not survive through IPA optimization based on decision | |
42 | of inliner. | |
43 | - Finally we propagate the following flags: unlikely executed, executed | |
44 | once, executed at startup and executed at exit. These flags are used to | |
026c3cfd | 45 | control code size/performance threshold and code placement (by producing |
daf5c770 | 46 | .text.unlikely/.text.hot/.text.startup/.text.exit subsections). */ |
08f835dc JH |
47 | #include "config.h" |
48 | #include "system.h" | |
49 | #include "coretypes.h" | |
c7131fb2 | 50 | #include "backend.h" |
4d648807 | 51 | #include "tree.h" |
c7131fb2 | 52 | #include "gimple.h" |
957060b5 AM |
53 | #include "predict.h" |
54 | #include "alloc-pool.h" | |
55 | #include "tree-pass.h" | |
56 | #include "cgraph.h" | |
57 | #include "data-streamer.h" | |
5be5c238 | 58 | #include "gimple-iterator.h" |
08f835dc | 59 | #include "ipa-utils.h" |
08f835dc JH |
60 | #include "profile.h" |
61 | #include "params.h" | |
62 | #include "value-prof.h" | |
08f835dc | 63 | #include "tree-inline.h" |
dd912cb8 | 64 | #include "symbol-summary.h" |
8bc5448f | 65 | #include "tree-vrp.h" |
c582198b | 66 | #include "ipa-prop.h" |
27d020cf | 67 | #include "ipa-fnsummary.h" |
08f835dc JH |
68 | |
69 | /* Entry in the histogram. */ | |
70 | ||
71 | struct histogram_entry | |
72 | { | |
73 | gcov_type count; | |
74 | int time; | |
75 | int size; | |
76 | }; | |
77 | ||
78 | /* Histogram of profile values. | |
79 | The histogram is represented as an ordered vector of entries allocated via | |
80 | histogram_pool. During construction a separate hashtable is kept to lookup | |
81 | duplicate entries. */ | |
82 | ||
83 | vec<histogram_entry *> histogram; | |
fcb87c50 | 84 | static object_allocator<histogram_entry> histogram_pool ("IPA histogram"); |
08f835dc JH |
85 | |
86 | /* Hashtable support for storing SSA names hashed by their SSA_NAME_VAR. */ | |
87 | ||
8d67ee55 | 88 | struct histogram_hash : nofree_ptr_hash <histogram_entry> |
08f835dc | 89 | { |
67f58944 TS |
90 | static inline hashval_t hash (const histogram_entry *); |
91 | static inline int equal (const histogram_entry *, const histogram_entry *); | |
08f835dc JH |
92 | }; |
93 | ||
94 | inline hashval_t | |
95 | histogram_hash::hash (const histogram_entry *val) | |
96 | { | |
97 | return val->count; | |
98 | } | |
99 | ||
100 | inline int | |
101 | histogram_hash::equal (const histogram_entry *val, const histogram_entry *val2) | |
102 | { | |
103 | return val->count == val2->count; | |
104 | } | |
105 | ||
106 | /* Account TIME and SIZE executed COUNT times into HISTOGRAM. | |
107 | HASHTABLE is the on-side hash kept to avoid duplicates. */ | |
108 | ||
109 | static void | |
c203e8a7 | 110 | account_time_size (hash_table<histogram_hash> *hashtable, |
08f835dc JH |
111 | vec<histogram_entry *> &histogram, |
112 | gcov_type count, int time, int size) | |
113 | { | |
114 | histogram_entry key = {count, 0, 0}; | |
c203e8a7 | 115 | histogram_entry **val = hashtable->find_slot (&key, INSERT); |
08f835dc JH |
116 | |
117 | if (!*val) | |
118 | { | |
d7809518 | 119 | *val = histogram_pool.allocate (); |
08f835dc JH |
120 | **val = key; |
121 | histogram.safe_push (*val); | |
122 | } | |
123 | (*val)->time += time; | |
124 | (*val)->size += size; | |
125 | } | |
126 | ||
127 | int | |
128 | cmp_counts (const void *v1, const void *v2) | |
129 | { | |
130 | const histogram_entry *h1 = *(const histogram_entry * const *)v1; | |
131 | const histogram_entry *h2 = *(const histogram_entry * const *)v2; | |
132 | if (h1->count < h2->count) | |
133 | return 1; | |
134 | if (h1->count > h2->count) | |
135 | return -1; | |
136 | return 0; | |
137 | } | |
138 | ||
139 | /* Dump HISTOGRAM to FILE. */ | |
140 | ||
141 | static void | |
142 | dump_histogram (FILE *file, vec<histogram_entry *> histogram) | |
143 | { | |
144 | unsigned int i; | |
145 | gcov_type overall_time = 0, cumulated_time = 0, cumulated_size = 0, overall_size = 0; | |
146 | ||
147 | fprintf (dump_file, "Histogram:\n"); | |
148 | for (i = 0; i < histogram.length (); i++) | |
149 | { | |
150 | overall_time += histogram[i]->count * histogram[i]->time; | |
151 | overall_size += histogram[i]->size; | |
152 | } | |
153 | if (!overall_time) | |
154 | overall_time = 1; | |
155 | if (!overall_size) | |
156 | overall_size = 1; | |
157 | for (i = 0; i < histogram.length (); i++) | |
158 | { | |
159 | cumulated_time += histogram[i]->count * histogram[i]->time; | |
160 | cumulated_size += histogram[i]->size; | |
16998094 | 161 | fprintf (file, " %" PRId64": time:%i (%2.2f) size:%i (%2.2f)\n", |
a9243bfc | 162 | (int64_t) histogram[i]->count, |
08f835dc JH |
163 | histogram[i]->time, |
164 | cumulated_time * 100.0 / overall_time, | |
165 | histogram[i]->size, | |
166 | cumulated_size * 100.0 / overall_size); | |
167 | } | |
168 | } | |
169 | ||
170 | /* Collect histogram from CFG profiles. */ | |
171 | ||
172 | static void | |
173 | ipa_profile_generate_summary (void) | |
174 | { | |
175 | struct cgraph_node *node; | |
176 | gimple_stmt_iterator gsi; | |
08f835dc JH |
177 | basic_block bb; |
178 | ||
c203e8a7 | 179 | hash_table<histogram_hash> hashtable (10); |
08f835dc JH |
180 | |
181 | FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node) | |
67348ccc | 182 | FOR_EACH_BB_FN (bb, DECL_STRUCT_FUNCTION (node->decl)) |
08f835dc JH |
183 | { |
184 | int time = 0; | |
185 | int size = 0; | |
186 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
187 | { | |
355fe088 | 188 | gimple *stmt = gsi_stmt (gsi); |
08f835dc JH |
189 | if (gimple_code (stmt) == GIMPLE_CALL |
190 | && !gimple_call_fndecl (stmt)) | |
191 | { | |
192 | histogram_value h; | |
193 | h = gimple_histogram_value_of_type | |
67348ccc | 194 | (DECL_STRUCT_FUNCTION (node->decl), |
08f835dc JH |
195 | stmt, HIST_TYPE_INDIR_CALL); |
196 | /* No need to do sanity check: gimple_ic_transform already | |
197 | takes away bad histograms. */ | |
198 | if (h) | |
199 | { | |
200 | /* counter 0 is target, counter 1 is number of execution we called target, | |
201 | counter 2 is total number of executions. */ | |
202 | if (h->hvalue.counters[2]) | |
203 | { | |
d52f5295 | 204 | struct cgraph_edge * e = node->get_edge (stmt); |
fd3c9a7e JH |
205 | if (e && !e->indirect_unknown_callee) |
206 | continue; | |
08f835dc JH |
207 | e->indirect_info->common_target_id |
208 | = h->hvalue.counters [0]; | |
209 | e->indirect_info->common_target_probability | |
210 | = GCOV_COMPUTE_SCALE (h->hvalue.counters [1], h->hvalue.counters [2]); | |
211 | if (e->indirect_info->common_target_probability > REG_BR_PROB_BASE) | |
212 | { | |
213 | if (dump_file) | |
214 | fprintf (dump_file, "Probability capped to 1\n"); | |
215 | e->indirect_info->common_target_probability = REG_BR_PROB_BASE; | |
216 | } | |
217 | } | |
67348ccc | 218 | gimple_remove_histogram_value (DECL_STRUCT_FUNCTION (node->decl), |
08f835dc JH |
219 | stmt, h); |
220 | } | |
221 | } | |
222 | time += estimate_num_insns (stmt, &eni_time_weights); | |
223 | size += estimate_num_insns (stmt, &eni_size_weights); | |
224 | } | |
c203e8a7 | 225 | account_time_size (&hashtable, histogram, bb->count, time, size); |
08f835dc | 226 | } |
08f835dc JH |
227 | histogram.qsort (cmp_counts); |
228 | } | |
229 | ||
230 | /* Serialize the ipa info for lto. */ | |
231 | ||
232 | static void | |
233 | ipa_profile_write_summary (void) | |
234 | { | |
235 | struct lto_simple_output_block *ob | |
236 | = lto_create_simple_output_block (LTO_section_ipa_profile); | |
237 | unsigned int i; | |
238 | ||
c3284718 | 239 | streamer_write_uhwi_stream (ob->main_stream, histogram.length ()); |
08f835dc JH |
240 | for (i = 0; i < histogram.length (); i++) |
241 | { | |
242 | streamer_write_gcov_count_stream (ob->main_stream, histogram[i]->count); | |
243 | streamer_write_uhwi_stream (ob->main_stream, histogram[i]->time); | |
244 | streamer_write_uhwi_stream (ob->main_stream, histogram[i]->size); | |
245 | } | |
246 | lto_destroy_simple_output_block (ob); | |
247 | } | |
248 | ||
249 | /* Deserialize the ipa info for lto. */ | |
250 | ||
251 | static void | |
252 | ipa_profile_read_summary (void) | |
253 | { | |
254 | struct lto_file_decl_data ** file_data_vec | |
255 | = lto_get_file_decl_data (); | |
256 | struct lto_file_decl_data * file_data; | |
08f835dc JH |
257 | int j = 0; |
258 | ||
c203e8a7 | 259 | hash_table<histogram_hash> hashtable (10); |
08f835dc JH |
260 | |
261 | while ((file_data = file_data_vec[j++])) | |
262 | { | |
263 | const char *data; | |
264 | size_t len; | |
265 | struct lto_input_block *ib | |
266 | = lto_create_simple_input_block (file_data, | |
267 | LTO_section_ipa_profile, | |
268 | &data, &len); | |
269 | if (ib) | |
270 | { | |
271 | unsigned int num = streamer_read_uhwi (ib); | |
272 | unsigned int n; | |
273 | for (n = 0; n < num; n++) | |
274 | { | |
275 | gcov_type count = streamer_read_gcov_count (ib); | |
276 | int time = streamer_read_uhwi (ib); | |
277 | int size = streamer_read_uhwi (ib); | |
c203e8a7 | 278 | account_time_size (&hashtable, histogram, |
08f835dc JH |
279 | count, time, size); |
280 | } | |
281 | lto_destroy_simple_input_block (file_data, | |
282 | LTO_section_ipa_profile, | |
283 | ib, data, len); | |
284 | } | |
285 | } | |
08f835dc JH |
286 | histogram.qsort (cmp_counts); |
287 | } | |
288 | ||
289 | /* Data used by ipa_propagate_frequency. */ | |
290 | ||
291 | struct ipa_propagate_frequency_data | |
292 | { | |
1ede94c5 | 293 | cgraph_node *function_symbol; |
08f835dc JH |
294 | bool maybe_unlikely_executed; |
295 | bool maybe_executed_once; | |
296 | bool only_called_at_startup; | |
297 | bool only_called_at_exit; | |
298 | }; | |
299 | ||
300 | /* Worker for ipa_propagate_frequency_1. */ | |
301 | ||
302 | static bool | |
303 | ipa_propagate_frequency_1 (struct cgraph_node *node, void *data) | |
304 | { | |
305 | struct ipa_propagate_frequency_data *d; | |
306 | struct cgraph_edge *edge; | |
307 | ||
308 | d = (struct ipa_propagate_frequency_data *)data; | |
309 | for (edge = node->callers; | |
310 | edge && (d->maybe_unlikely_executed || d->maybe_executed_once | |
311 | || d->only_called_at_startup || d->only_called_at_exit); | |
312 | edge = edge->next_caller) | |
313 | { | |
1ede94c5 | 314 | if (edge->caller != d->function_symbol) |
08f835dc JH |
315 | { |
316 | d->only_called_at_startup &= edge->caller->only_called_at_startup; | |
317 | /* It makes sense to put main() together with the static constructors. | |
318 | It will be executed for sure, but rest of functions called from | |
319 | main are definitely not at startup only. */ | |
67348ccc | 320 | if (MAIN_NAME_P (DECL_NAME (edge->caller->decl))) |
08f835dc JH |
321 | d->only_called_at_startup = 0; |
322 | d->only_called_at_exit &= edge->caller->only_called_at_exit; | |
323 | } | |
daf5c770 JH |
324 | |
325 | /* When profile feedback is available, do not try to propagate too hard; | |
326 | counts are already good guide on function frequencies and roundoff | |
327 | errors can make us to push function into unlikely section even when | |
328 | it is executed by the train run. Transfer the function only if all | |
329 | callers are unlikely executed. */ | |
1ede94c5 JH |
330 | if (profile_info |
331 | && opt_for_fn (d->function_symbol->decl, flag_branch_probabilities) | |
332 | /* Thunks are not profiled. This is more or less implementation | |
333 | bug. */ | |
334 | && !d->function_symbol->thunk.thunk_p | |
daf5c770 JH |
335 | && (edge->caller->frequency != NODE_FREQUENCY_UNLIKELY_EXECUTED |
336 | || (edge->caller->global.inlined_to | |
337 | && edge->caller->global.inlined_to->frequency | |
338 | != NODE_FREQUENCY_UNLIKELY_EXECUTED))) | |
339 | d->maybe_unlikely_executed = false; | |
08f835dc JH |
340 | if (!edge->frequency) |
341 | continue; | |
342 | switch (edge->caller->frequency) | |
343 | { | |
344 | case NODE_FREQUENCY_UNLIKELY_EXECUTED: | |
345 | break; | |
346 | case NODE_FREQUENCY_EXECUTED_ONCE: | |
347 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
348 | fprintf (dump_file, " Called by %s that is executed once\n", | |
fec39fa6 | 349 | edge->caller->name ()); |
08f835dc | 350 | d->maybe_unlikely_executed = false; |
263e19c7 | 351 | if (ipa_call_summaries->get (edge)->loop_depth) |
08f835dc JH |
352 | { |
353 | d->maybe_executed_once = false; | |
354 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
355 | fprintf (dump_file, " Called in loop\n"); | |
356 | } | |
357 | break; | |
358 | case NODE_FREQUENCY_HOT: | |
359 | case NODE_FREQUENCY_NORMAL: | |
360 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
361 | fprintf (dump_file, " Called by %s that is normal or hot\n", | |
fec39fa6 | 362 | edge->caller->name ()); |
08f835dc JH |
363 | d->maybe_unlikely_executed = false; |
364 | d->maybe_executed_once = false; | |
365 | break; | |
366 | } | |
367 | } | |
368 | return edge != NULL; | |
369 | } | |
370 | ||
daf5c770 JH |
371 | /* Return ture if NODE contains hot calls. */ |
372 | ||
373 | bool | |
374 | contains_hot_call_p (struct cgraph_node *node) | |
375 | { | |
376 | struct cgraph_edge *e; | |
377 | for (e = node->callees; e; e = e->next_callee) | |
3dafb85c | 378 | if (e->maybe_hot_p ()) |
daf5c770 JH |
379 | return true; |
380 | else if (!e->inline_failed | |
381 | && contains_hot_call_p (e->callee)) | |
382 | return true; | |
383 | for (e = node->indirect_calls; e; e = e->next_callee) | |
3dafb85c | 384 | if (e->maybe_hot_p ()) |
daf5c770 JH |
385 | return true; |
386 | return false; | |
387 | } | |
388 | ||
08f835dc JH |
389 | /* See if the frequency of NODE can be updated based on frequencies of its |
390 | callers. */ | |
391 | bool | |
392 | ipa_propagate_frequency (struct cgraph_node *node) | |
393 | { | |
1ede94c5 | 394 | struct ipa_propagate_frequency_data d = {node, true, true, true, true}; |
08f835dc JH |
395 | bool changed = false; |
396 | ||
397 | /* We can not propagate anything useful about externally visible functions | |
398 | nor about virtuals. */ | |
399 | if (!node->local.local | |
67348ccc | 400 | || node->alias |
2bf86c84 JH |
401 | || (opt_for_fn (node->decl, flag_devirtualize) |
402 | && DECL_VIRTUAL_P (node->decl))) | |
08f835dc | 403 | return false; |
67348ccc | 404 | gcc_assert (node->analyzed); |
08f835dc | 405 | if (dump_file && (dump_flags & TDF_DETAILS)) |
fec39fa6 | 406 | fprintf (dump_file, "Processing frequency %s\n", node->name ()); |
08f835dc | 407 | |
1ede94c5 JH |
408 | node->call_for_symbol_and_aliases (ipa_propagate_frequency_1, &d, |
409 | true); | |
08f835dc JH |
410 | |
411 | if ((d.only_called_at_startup && !d.only_called_at_exit) | |
412 | && !node->only_called_at_startup) | |
413 | { | |
414 | node->only_called_at_startup = true; | |
415 | if (dump_file) | |
416 | fprintf (dump_file, "Node %s promoted to only called at startup.\n", | |
fec39fa6 | 417 | node->name ()); |
08f835dc JH |
418 | changed = true; |
419 | } | |
420 | if ((d.only_called_at_exit && !d.only_called_at_startup) | |
421 | && !node->only_called_at_exit) | |
422 | { | |
423 | node->only_called_at_exit = true; | |
424 | if (dump_file) | |
425 | fprintf (dump_file, "Node %s promoted to only called at exit.\n", | |
fec39fa6 | 426 | node->name ()); |
08f835dc JH |
427 | changed = true; |
428 | } | |
daf5c770 JH |
429 | |
430 | /* With profile we can decide on hot/normal based on count. */ | |
431 | if (node->count) | |
432 | { | |
433 | bool hot = false; | |
434 | if (node->count >= get_hot_bb_threshold ()) | |
435 | hot = true; | |
436 | if (!hot) | |
437 | hot |= contains_hot_call_p (node); | |
438 | if (hot) | |
439 | { | |
440 | if (node->frequency != NODE_FREQUENCY_HOT) | |
441 | { | |
442 | if (dump_file) | |
443 | fprintf (dump_file, "Node %s promoted to hot.\n", | |
fec39fa6 | 444 | node->name ()); |
daf5c770 JH |
445 | node->frequency = NODE_FREQUENCY_HOT; |
446 | return true; | |
447 | } | |
448 | return false; | |
449 | } | |
450 | else if (node->frequency == NODE_FREQUENCY_HOT) | |
451 | { | |
452 | if (dump_file) | |
453 | fprintf (dump_file, "Node %s reduced to normal.\n", | |
fec39fa6 | 454 | node->name ()); |
daf5c770 JH |
455 | node->frequency = NODE_FREQUENCY_NORMAL; |
456 | changed = true; | |
457 | } | |
458 | } | |
08f835dc JH |
459 | /* These come either from profile or user hints; never update them. */ |
460 | if (node->frequency == NODE_FREQUENCY_HOT | |
461 | || node->frequency == NODE_FREQUENCY_UNLIKELY_EXECUTED) | |
462 | return changed; | |
463 | if (d.maybe_unlikely_executed) | |
464 | { | |
465 | node->frequency = NODE_FREQUENCY_UNLIKELY_EXECUTED; | |
466 | if (dump_file) | |
467 | fprintf (dump_file, "Node %s promoted to unlikely executed.\n", | |
fec39fa6 | 468 | node->name ()); |
08f835dc JH |
469 | changed = true; |
470 | } | |
471 | else if (d.maybe_executed_once && node->frequency != NODE_FREQUENCY_EXECUTED_ONCE) | |
472 | { | |
473 | node->frequency = NODE_FREQUENCY_EXECUTED_ONCE; | |
474 | if (dump_file) | |
475 | fprintf (dump_file, "Node %s promoted to executed once.\n", | |
fec39fa6 | 476 | node->name ()); |
08f835dc JH |
477 | changed = true; |
478 | } | |
479 | return changed; | |
480 | } | |
481 | ||
482 | /* Simple ipa profile pass propagating frequencies across the callgraph. */ | |
483 | ||
484 | static unsigned int | |
485 | ipa_profile (void) | |
486 | { | |
487 | struct cgraph_node **order; | |
488 | struct cgraph_edge *e; | |
489 | int order_pos; | |
490 | bool something_changed = false; | |
491 | int i; | |
492 | gcov_type overall_time = 0, cutoff = 0, cumulated = 0, overall_size = 0; | |
493 | struct cgraph_node *n,*n2; | |
494 | int nindirect = 0, ncommon = 0, nunknown = 0, nuseless = 0, nconverted = 0; | |
95d81ba5 | 495 | int nmismatch = 0, nimpossible = 0; |
08f835dc JH |
496 | bool node_map_initialized = false; |
497 | ||
498 | if (dump_file) | |
499 | dump_histogram (dump_file, histogram); | |
500 | for (i = 0; i < (int)histogram.length (); i++) | |
501 | { | |
502 | overall_time += histogram[i]->count * histogram[i]->time; | |
503 | overall_size += histogram[i]->size; | |
504 | } | |
505 | if (overall_time) | |
506 | { | |
507 | gcov_type threshold; | |
508 | ||
509 | gcc_assert (overall_size); | |
510 | if (dump_file) | |
511 | { | |
512 | gcov_type min, cumulated_time = 0, cumulated_size = 0; | |
513 | ||
16998094 | 514 | fprintf (dump_file, "Overall time: %" PRId64"\n", |
a9243bfc | 515 | (int64_t)overall_time); |
08f835dc JH |
516 | min = get_hot_bb_threshold (); |
517 | for (i = 0; i < (int)histogram.length () && histogram[i]->count >= min; | |
518 | i++) | |
519 | { | |
520 | cumulated_time += histogram[i]->count * histogram[i]->time; | |
521 | cumulated_size += histogram[i]->size; | |
522 | } | |
16998094 | 523 | fprintf (dump_file, "GCOV min count: %" PRId64 |
08f835dc | 524 | " Time:%3.2f%% Size:%3.2f%%\n", |
a9243bfc | 525 | (int64_t)min, |
08f835dc JH |
526 | cumulated_time * 100.0 / overall_time, |
527 | cumulated_size * 100.0 / overall_size); | |
528 | } | |
529 | cutoff = (overall_time * PARAM_VALUE (HOT_BB_COUNT_WS_PERMILLE) + 500) / 1000; | |
530 | threshold = 0; | |
531 | for (i = 0; cumulated < cutoff; i++) | |
532 | { | |
533 | cumulated += histogram[i]->count * histogram[i]->time; | |
534 | threshold = histogram[i]->count; | |
535 | } | |
536 | if (!threshold) | |
537 | threshold = 1; | |
538 | if (dump_file) | |
539 | { | |
540 | gcov_type cumulated_time = 0, cumulated_size = 0; | |
541 | ||
542 | for (i = 0; | |
543 | i < (int)histogram.length () && histogram[i]->count >= threshold; | |
544 | i++) | |
545 | { | |
546 | cumulated_time += histogram[i]->count * histogram[i]->time; | |
547 | cumulated_size += histogram[i]->size; | |
548 | } | |
16998094 | 549 | fprintf (dump_file, "Determined min count: %" PRId64 |
08f835dc | 550 | " Time:%3.2f%% Size:%3.2f%%\n", |
a9243bfc | 551 | (int64_t)threshold, |
08f835dc JH |
552 | cumulated_time * 100.0 / overall_time, |
553 | cumulated_size * 100.0 / overall_size); | |
554 | } | |
555 | if (threshold > get_hot_bb_threshold () | |
556 | || in_lto_p) | |
557 | { | |
558 | if (dump_file) | |
559 | fprintf (dump_file, "Threshold updated.\n"); | |
560 | set_hot_bb_threshold (threshold); | |
561 | } | |
562 | } | |
c3284718 | 563 | histogram.release (); |
d7809518 | 564 | histogram_pool.release (); |
08f835dc JH |
565 | |
566 | /* Produce speculative calls: we saved common traget from porfiling into | |
567 | e->common_target_id. Now, at link time, we can look up corresponding | |
568 | function node and produce speculative call. */ | |
569 | ||
570 | FOR_EACH_DEFINED_FUNCTION (n) | |
571 | { | |
572 | bool update = false; | |
573 | ||
1ede94c5 JH |
574 | if (!opt_for_fn (n->decl, flag_ipa_profile)) |
575 | continue; | |
576 | ||
08f835dc JH |
577 | for (e = n->indirect_calls; e; e = e->next_callee) |
578 | { | |
579 | if (n->count) | |
580 | nindirect++; | |
581 | if (e->indirect_info->common_target_id) | |
582 | { | |
583 | if (!node_map_initialized) | |
584 | init_node_map (false); | |
585 | node_map_initialized = true; | |
586 | ncommon++; | |
587 | n2 = find_func_by_profile_id (e->indirect_info->common_target_id); | |
588 | if (n2) | |
589 | { | |
590 | if (dump_file) | |
591 | { | |
592 | fprintf (dump_file, "Indirect call -> direct call from" | |
593 | " other module %s/%i => %s/%i, prob %3.2f\n", | |
2a72a953 DM |
594 | xstrdup_for_dump (n->name ()), n->order, |
595 | xstrdup_for_dump (n2->name ()), n2->order, | |
08f835dc JH |
596 | e->indirect_info->common_target_probability |
597 | / (float)REG_BR_PROB_BASE); | |
598 | } | |
599 | if (e->indirect_info->common_target_probability | |
600 | < REG_BR_PROB_BASE / 2) | |
601 | { | |
602 | nuseless++; | |
603 | if (dump_file) | |
604 | fprintf (dump_file, | |
605 | "Not speculating: probability is too low.\n"); | |
606 | } | |
3dafb85c | 607 | else if (!e->maybe_hot_p ()) |
08f835dc JH |
608 | { |
609 | nuseless++; | |
610 | if (dump_file) | |
611 | fprintf (dump_file, | |
612 | "Not speculating: call is cold.\n"); | |
613 | } | |
d52f5295 ML |
614 | else if (n2->get_availability () <= AVAIL_INTERPOSABLE |
615 | && n2->can_be_discarded_p ()) | |
08f835dc JH |
616 | { |
617 | nuseless++; | |
618 | if (dump_file) | |
619 | fprintf (dump_file, | |
620 | "Not speculating: target is overwritable " | |
621 | "and can be discarded.\n"); | |
622 | } | |
6fe906a3 | 623 | else if (ipa_node_params_sum && ipa_edge_args_sum |
f65f1ae3 MJ |
624 | && (!vec_safe_is_empty |
625 | (IPA_NODE_REF (n2)->descriptors)) | |
95d81ba5 JH |
626 | && ipa_get_param_count (IPA_NODE_REF (n2)) |
627 | != ipa_get_cs_argument_count (IPA_EDGE_REF (e)) | |
628 | && (ipa_get_param_count (IPA_NODE_REF (n2)) | |
629 | >= ipa_get_cs_argument_count (IPA_EDGE_REF (e)) | |
630 | || !stdarg_p (TREE_TYPE (n2->decl)))) | |
631 | { | |
632 | nmismatch++; | |
633 | if (dump_file) | |
634 | fprintf (dump_file, | |
635 | "Not speculating: " | |
636 | "parameter count mistmatch\n"); | |
637 | } | |
638 | else if (e->indirect_info->polymorphic | |
639 | && !opt_for_fn (n->decl, flag_devirtualize) | |
640 | && !possible_polymorphic_call_target_p (e, n2)) | |
641 | { | |
642 | nimpossible++; | |
643 | if (dump_file) | |
644 | fprintf (dump_file, | |
645 | "Not speculating: " | |
646 | "function is not in the polymorphic " | |
647 | "call target list\n"); | |
648 | } | |
08f835dc JH |
649 | else |
650 | { | |
651 | /* Target may be overwritable, but profile says that | |
652 | control flow goes to this particular implementation | |
653 | of N2. Speculate on the local alias to allow inlining. | |
654 | */ | |
d52f5295 | 655 | if (!n2->can_be_discarded_p ()) |
5b79657a JH |
656 | { |
657 | cgraph_node *alias; | |
d52f5295 | 658 | alias = dyn_cast<cgraph_node *> (n2->noninterposable_alias ()); |
5b79657a JH |
659 | if (alias) |
660 | n2 = alias; | |
661 | } | |
08f835dc | 662 | nconverted++; |
3dafb85c ML |
663 | e->make_speculative |
664 | (n2, | |
08f835dc JH |
665 | apply_scale (e->count, |
666 | e->indirect_info->common_target_probability), | |
667 | apply_scale (e->frequency, | |
668 | e->indirect_info->common_target_probability)); | |
669 | update = true; | |
670 | } | |
671 | } | |
672 | else | |
673 | { | |
674 | if (dump_file) | |
675 | fprintf (dump_file, "Function with profile-id %i not found.\n", | |
676 | e->indirect_info->common_target_id); | |
677 | nunknown++; | |
678 | } | |
679 | } | |
680 | } | |
681 | if (update) | |
682 | inline_update_overall_summary (n); | |
683 | } | |
684 | if (node_map_initialized) | |
685 | del_node_map (); | |
686 | if (dump_file && nindirect) | |
687 | fprintf (dump_file, | |
688 | "%i indirect calls trained.\n" | |
689 | "%i (%3.2f%%) have common target.\n" | |
690 | "%i (%3.2f%%) targets was not found.\n" | |
95d81ba5 JH |
691 | "%i (%3.2f%%) targets had parameter count mismatch.\n" |
692 | "%i (%3.2f%%) targets was not in polymorphic call target list.\n" | |
08f835dc JH |
693 | "%i (%3.2f%%) speculations seems useless.\n" |
694 | "%i (%3.2f%%) speculations produced.\n", | |
695 | nindirect, | |
696 | ncommon, ncommon * 100.0 / nindirect, | |
697 | nunknown, nunknown * 100.0 / nindirect, | |
95d81ba5 JH |
698 | nmismatch, nmismatch * 100.0 / nindirect, |
699 | nimpossible, nimpossible * 100.0 / nindirect, | |
08f835dc JH |
700 | nuseless, nuseless * 100.0 / nindirect, |
701 | nconverted, nconverted * 100.0 / nindirect); | |
702 | ||
3dafb85c | 703 | order = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count); |
08f835dc JH |
704 | order_pos = ipa_reverse_postorder (order); |
705 | for (i = order_pos - 1; i >= 0; i--) | |
706 | { | |
1ede94c5 JH |
707 | if (order[i]->local.local |
708 | && opt_for_fn (order[i]->decl, flag_ipa_profile) | |
709 | && ipa_propagate_frequency (order[i])) | |
08f835dc JH |
710 | { |
711 | for (e = order[i]->callees; e; e = e->next_callee) | |
67348ccc | 712 | if (e->callee->local.local && !e->callee->aux) |
08f835dc JH |
713 | { |
714 | something_changed = true; | |
67348ccc | 715 | e->callee->aux = (void *)1; |
08f835dc JH |
716 | } |
717 | } | |
67348ccc | 718 | order[i]->aux = NULL; |
08f835dc JH |
719 | } |
720 | ||
721 | while (something_changed) | |
722 | { | |
723 | something_changed = false; | |
724 | for (i = order_pos - 1; i >= 0; i--) | |
725 | { | |
1ede94c5 JH |
726 | if (order[i]->aux |
727 | && opt_for_fn (order[i]->decl, flag_ipa_profile) | |
728 | && ipa_propagate_frequency (order[i])) | |
08f835dc JH |
729 | { |
730 | for (e = order[i]->callees; e; e = e->next_callee) | |
67348ccc | 731 | if (e->callee->local.local && !e->callee->aux) |
08f835dc JH |
732 | { |
733 | something_changed = true; | |
67348ccc | 734 | e->callee->aux = (void *)1; |
08f835dc JH |
735 | } |
736 | } | |
67348ccc | 737 | order[i]->aux = NULL; |
08f835dc JH |
738 | } |
739 | } | |
740 | free (order); | |
741 | return 0; | |
742 | } | |
743 | ||
08f835dc JH |
744 | namespace { |
745 | ||
746 | const pass_data pass_data_ipa_profile = | |
747 | { | |
748 | IPA_PASS, /* type */ | |
749 | "profile_estimate", /* name */ | |
750 | OPTGROUP_NONE, /* optinfo_flags */ | |
08f835dc JH |
751 | TV_IPA_PROFILE, /* tv_id */ |
752 | 0, /* properties_required */ | |
753 | 0, /* properties_provided */ | |
754 | 0, /* properties_destroyed */ | |
755 | 0, /* todo_flags_start */ | |
756 | 0, /* todo_flags_finish */ | |
757 | }; | |
758 | ||
759 | class pass_ipa_profile : public ipa_opt_pass_d | |
760 | { | |
761 | public: | |
c3284718 RS |
762 | pass_ipa_profile (gcc::context *ctxt) |
763 | : ipa_opt_pass_d (pass_data_ipa_profile, ctxt, | |
764 | ipa_profile_generate_summary, /* generate_summary */ | |
765 | ipa_profile_write_summary, /* write_summary */ | |
766 | ipa_profile_read_summary, /* read_summary */ | |
767 | NULL, /* write_optimization_summary */ | |
768 | NULL, /* read_optimization_summary */ | |
769 | NULL, /* stmt_fixup */ | |
770 | 0, /* function_transform_todo_flags_start */ | |
771 | NULL, /* function_transform */ | |
772 | NULL) /* variable_transform */ | |
08f835dc JH |
773 | {} |
774 | ||
775 | /* opt_pass methods: */ | |
2bf86c84 | 776 | virtual bool gate (function *) { return flag_ipa_profile || in_lto_p; } |
be55bfe6 | 777 | virtual unsigned int execute (function *) { return ipa_profile (); } |
08f835dc JH |
778 | |
779 | }; // class pass_ipa_profile | |
780 | ||
781 | } // anon namespace | |
782 | ||
783 | ipa_opt_pass_d * | |
784 | make_pass_ipa_profile (gcc::context *ctxt) | |
785 | { | |
786 | return new pass_ipa_profile (ctxt); | |
787 | } |