]> git.ipfire.org Git - thirdparty/gcc.git/blame - gcc/ipa-profile.c
New symbol_summary class introduced.
[thirdparty/gcc.git] / gcc / ipa-profile.c
CommitLineData
08f835dc 1/* Basic IPA optimizations based on profile.
23a5b65a 2 Copyright (C) 2003-2014 Free Software Foundation, Inc.
08f835dc
JH
3
4This file is part of GCC.
5
6GCC is free software; you can redistribute it and/or modify it under
7the terms of the GNU General Public License as published by the Free
8Software Foundation; either version 3, or (at your option) any later
9version.
10
11GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12WARRANTY; without even the implied warranty of MERCHANTABILITY or
13FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14for more details.
15
16You should have received a copy of the GNU General Public License
17along 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
45 control code size/performance threshold and and code placement (by producing
46 .text.unlikely/.text.hot/.text.startup/.text.exit subsections). */
08f835dc
JH
47#include "config.h"
48#include "system.h"
49#include "coretypes.h"
50#include "tm.h"
4d648807 51#include "tree.h"
60393bbc
AM
52#include "predict.h"
53#include "dominance.h"
54#include "cfg.h"
55#include "basic-block.h"
c582198b
AM
56#include "hash-map.h"
57#include "is-a.h"
58#include "plugin-api.h"
59#include "vec.h"
60#include "hashtab.h"
61#include "hash-set.h"
62#include "machmode.h"
63#include "hard-reg-set.h"
64#include "input.h"
65#include "function.h"
66#include "ipa-ref.h"
08f835dc
JH
67#include "cgraph.h"
68#include "tree-pass.h"
2fb9a547
AM
69#include "tree-ssa-alias.h"
70#include "internal-fn.h"
71#include "gimple-expr.h"
08f835dc 72#include "gimple.h"
5be5c238 73#include "gimple-iterator.h"
08f835dc
JH
74#include "flags.h"
75#include "target.h"
76#include "tree-iterator.h"
77#include "ipa-utils.h"
08f835dc
JH
78#include "profile.h"
79#include "params.h"
80#include "value-prof.h"
81#include "alloc-pool.h"
82#include "tree-inline.h"
83#include "lto-streamer.h"
84#include "data-streamer.h"
c582198b 85#include "ipa-prop.h"
08f835dc
JH
86#include "ipa-inline.h"
87
88/* Entry in the histogram. */
89
90struct histogram_entry
91{
92 gcov_type count;
93 int time;
94 int size;
95};
96
97/* Histogram of profile values.
98 The histogram is represented as an ordered vector of entries allocated via
99 histogram_pool. During construction a separate hashtable is kept to lookup
100 duplicate entries. */
101
102vec<histogram_entry *> histogram;
103static alloc_pool histogram_pool;
104
105/* Hashtable support for storing SSA names hashed by their SSA_NAME_VAR. */
106
107struct histogram_hash : typed_noop_remove <histogram_entry>
108{
109 typedef histogram_entry value_type;
110 typedef histogram_entry compare_type;
111 static inline hashval_t hash (const value_type *);
112 static inline int equal (const value_type *, const compare_type *);
113};
114
115inline hashval_t
116histogram_hash::hash (const histogram_entry *val)
117{
118 return val->count;
119}
120
121inline int
122histogram_hash::equal (const histogram_entry *val, const histogram_entry *val2)
123{
124 return val->count == val2->count;
125}
126
127/* Account TIME and SIZE executed COUNT times into HISTOGRAM.
128 HASHTABLE is the on-side hash kept to avoid duplicates. */
129
130static void
c203e8a7 131account_time_size (hash_table<histogram_hash> *hashtable,
08f835dc
JH
132 vec<histogram_entry *> &histogram,
133 gcov_type count, int time, int size)
134{
135 histogram_entry key = {count, 0, 0};
c203e8a7 136 histogram_entry **val = hashtable->find_slot (&key, INSERT);
08f835dc
JH
137
138 if (!*val)
139 {
140 *val = (histogram_entry *) pool_alloc (histogram_pool);
141 **val = key;
142 histogram.safe_push (*val);
143 }
144 (*val)->time += time;
145 (*val)->size += size;
146}
147
148int
149cmp_counts (const void *v1, const void *v2)
150{
151 const histogram_entry *h1 = *(const histogram_entry * const *)v1;
152 const histogram_entry *h2 = *(const histogram_entry * const *)v2;
153 if (h1->count < h2->count)
154 return 1;
155 if (h1->count > h2->count)
156 return -1;
157 return 0;
158}
159
160/* Dump HISTOGRAM to FILE. */
161
162static void
163dump_histogram (FILE *file, vec<histogram_entry *> histogram)
164{
165 unsigned int i;
166 gcov_type overall_time = 0, cumulated_time = 0, cumulated_size = 0, overall_size = 0;
167
168 fprintf (dump_file, "Histogram:\n");
169 for (i = 0; i < histogram.length (); i++)
170 {
171 overall_time += histogram[i]->count * histogram[i]->time;
172 overall_size += histogram[i]->size;
173 }
174 if (!overall_time)
175 overall_time = 1;
176 if (!overall_size)
177 overall_size = 1;
178 for (i = 0; i < histogram.length (); i++)
179 {
180 cumulated_time += histogram[i]->count * histogram[i]->time;
181 cumulated_size += histogram[i]->size;
a9243bfc
RB
182 fprintf (file, " %"PRId64": time:%i (%2.2f) size:%i (%2.2f)\n",
183 (int64_t) histogram[i]->count,
08f835dc
JH
184 histogram[i]->time,
185 cumulated_time * 100.0 / overall_time,
186 histogram[i]->size,
187 cumulated_size * 100.0 / overall_size);
188 }
189}
190
191/* Collect histogram from CFG profiles. */
192
193static void
194ipa_profile_generate_summary (void)
195{
196 struct cgraph_node *node;
197 gimple_stmt_iterator gsi;
08f835dc
JH
198 basic_block bb;
199
c203e8a7 200 hash_table<histogram_hash> hashtable (10);
08f835dc
JH
201 histogram_pool = create_alloc_pool ("IPA histogram", sizeof (struct histogram_entry),
202 10);
203
204 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
67348ccc 205 FOR_EACH_BB_FN (bb, DECL_STRUCT_FUNCTION (node->decl))
08f835dc
JH
206 {
207 int time = 0;
208 int size = 0;
209 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
210 {
211 gimple stmt = gsi_stmt (gsi);
212 if (gimple_code (stmt) == GIMPLE_CALL
213 && !gimple_call_fndecl (stmt))
214 {
215 histogram_value h;
216 h = gimple_histogram_value_of_type
67348ccc 217 (DECL_STRUCT_FUNCTION (node->decl),
08f835dc
JH
218 stmt, HIST_TYPE_INDIR_CALL);
219 /* No need to do sanity check: gimple_ic_transform already
220 takes away bad histograms. */
221 if (h)
222 {
223 /* counter 0 is target, counter 1 is number of execution we called target,
224 counter 2 is total number of executions. */
225 if (h->hvalue.counters[2])
226 {
d52f5295 227 struct cgraph_edge * e = node->get_edge (stmt);
fd3c9a7e
JH
228 if (e && !e->indirect_unknown_callee)
229 continue;
08f835dc
JH
230 e->indirect_info->common_target_id
231 = h->hvalue.counters [0];
232 e->indirect_info->common_target_probability
233 = GCOV_COMPUTE_SCALE (h->hvalue.counters [1], h->hvalue.counters [2]);
234 if (e->indirect_info->common_target_probability > REG_BR_PROB_BASE)
235 {
236 if (dump_file)
237 fprintf (dump_file, "Probability capped to 1\n");
238 e->indirect_info->common_target_probability = REG_BR_PROB_BASE;
239 }
240 }
67348ccc 241 gimple_remove_histogram_value (DECL_STRUCT_FUNCTION (node->decl),
08f835dc
JH
242 stmt, h);
243 }
244 }
245 time += estimate_num_insns (stmt, &eni_time_weights);
246 size += estimate_num_insns (stmt, &eni_size_weights);
247 }
c203e8a7 248 account_time_size (&hashtable, histogram, bb->count, time, size);
08f835dc 249 }
08f835dc
JH
250 histogram.qsort (cmp_counts);
251}
252
253/* Serialize the ipa info for lto. */
254
255static void
256ipa_profile_write_summary (void)
257{
258 struct lto_simple_output_block *ob
259 = lto_create_simple_output_block (LTO_section_ipa_profile);
260 unsigned int i;
261
c3284718 262 streamer_write_uhwi_stream (ob->main_stream, histogram.length ());
08f835dc
JH
263 for (i = 0; i < histogram.length (); i++)
264 {
265 streamer_write_gcov_count_stream (ob->main_stream, histogram[i]->count);
266 streamer_write_uhwi_stream (ob->main_stream, histogram[i]->time);
267 streamer_write_uhwi_stream (ob->main_stream, histogram[i]->size);
268 }
269 lto_destroy_simple_output_block (ob);
270}
271
272/* Deserialize the ipa info for lto. */
273
274static void
275ipa_profile_read_summary (void)
276{
277 struct lto_file_decl_data ** file_data_vec
278 = lto_get_file_decl_data ();
279 struct lto_file_decl_data * file_data;
08f835dc
JH
280 int j = 0;
281
c203e8a7 282 hash_table<histogram_hash> hashtable (10);
08f835dc
JH
283 histogram_pool = create_alloc_pool ("IPA histogram", sizeof (struct histogram_entry),
284 10);
285
286 while ((file_data = file_data_vec[j++]))
287 {
288 const char *data;
289 size_t len;
290 struct lto_input_block *ib
291 = lto_create_simple_input_block (file_data,
292 LTO_section_ipa_profile,
293 &data, &len);
294 if (ib)
295 {
296 unsigned int num = streamer_read_uhwi (ib);
297 unsigned int n;
298 for (n = 0; n < num; n++)
299 {
300 gcov_type count = streamer_read_gcov_count (ib);
301 int time = streamer_read_uhwi (ib);
302 int size = streamer_read_uhwi (ib);
c203e8a7 303 account_time_size (&hashtable, histogram,
08f835dc
JH
304 count, time, size);
305 }
306 lto_destroy_simple_input_block (file_data,
307 LTO_section_ipa_profile,
308 ib, data, len);
309 }
310 }
08f835dc
JH
311 histogram.qsort (cmp_counts);
312}
313
314/* Data used by ipa_propagate_frequency. */
315
316struct ipa_propagate_frequency_data
317{
318 bool maybe_unlikely_executed;
319 bool maybe_executed_once;
320 bool only_called_at_startup;
321 bool only_called_at_exit;
322};
323
324/* Worker for ipa_propagate_frequency_1. */
325
326static bool
327ipa_propagate_frequency_1 (struct cgraph_node *node, void *data)
328{
329 struct ipa_propagate_frequency_data *d;
330 struct cgraph_edge *edge;
331
332 d = (struct ipa_propagate_frequency_data *)data;
333 for (edge = node->callers;
334 edge && (d->maybe_unlikely_executed || d->maybe_executed_once
335 || d->only_called_at_startup || d->only_called_at_exit);
336 edge = edge->next_caller)
337 {
338 if (edge->caller != node)
339 {
340 d->only_called_at_startup &= edge->caller->only_called_at_startup;
341 /* It makes sense to put main() together with the static constructors.
342 It will be executed for sure, but rest of functions called from
343 main are definitely not at startup only. */
67348ccc 344 if (MAIN_NAME_P (DECL_NAME (edge->caller->decl)))
08f835dc
JH
345 d->only_called_at_startup = 0;
346 d->only_called_at_exit &= edge->caller->only_called_at_exit;
347 }
daf5c770
JH
348
349 /* When profile feedback is available, do not try to propagate too hard;
350 counts are already good guide on function frequencies and roundoff
351 errors can make us to push function into unlikely section even when
352 it is executed by the train run. Transfer the function only if all
353 callers are unlikely executed. */
354 if (profile_info && flag_branch_probabilities
355 && (edge->caller->frequency != NODE_FREQUENCY_UNLIKELY_EXECUTED
356 || (edge->caller->global.inlined_to
357 && edge->caller->global.inlined_to->frequency
358 != NODE_FREQUENCY_UNLIKELY_EXECUTED)))
359 d->maybe_unlikely_executed = false;
08f835dc
JH
360 if (!edge->frequency)
361 continue;
362 switch (edge->caller->frequency)
363 {
364 case NODE_FREQUENCY_UNLIKELY_EXECUTED:
365 break;
366 case NODE_FREQUENCY_EXECUTED_ONCE:
367 if (dump_file && (dump_flags & TDF_DETAILS))
368 fprintf (dump_file, " Called by %s that is executed once\n",
fec39fa6 369 edge->caller->name ());
08f835dc
JH
370 d->maybe_unlikely_executed = false;
371 if (inline_edge_summary (edge)->loop_depth)
372 {
373 d->maybe_executed_once = false;
374 if (dump_file && (dump_flags & TDF_DETAILS))
375 fprintf (dump_file, " Called in loop\n");
376 }
377 break;
378 case NODE_FREQUENCY_HOT:
379 case NODE_FREQUENCY_NORMAL:
380 if (dump_file && (dump_flags & TDF_DETAILS))
381 fprintf (dump_file, " Called by %s that is normal or hot\n",
fec39fa6 382 edge->caller->name ());
08f835dc
JH
383 d->maybe_unlikely_executed = false;
384 d->maybe_executed_once = false;
385 break;
386 }
387 }
388 return edge != NULL;
389}
390
daf5c770
JH
391/* Return ture if NODE contains hot calls. */
392
393bool
394contains_hot_call_p (struct cgraph_node *node)
395{
396 struct cgraph_edge *e;
397 for (e = node->callees; e; e = e->next_callee)
3dafb85c 398 if (e->maybe_hot_p ())
daf5c770
JH
399 return true;
400 else if (!e->inline_failed
401 && contains_hot_call_p (e->callee))
402 return true;
403 for (e = node->indirect_calls; e; e = e->next_callee)
3dafb85c 404 if (e->maybe_hot_p ())
daf5c770
JH
405 return true;
406 return false;
407}
408
08f835dc
JH
409/* See if the frequency of NODE can be updated based on frequencies of its
410 callers. */
411bool
412ipa_propagate_frequency (struct cgraph_node *node)
413{
414 struct ipa_propagate_frequency_data d = {true, true, true, true};
415 bool changed = false;
416
417 /* We can not propagate anything useful about externally visible functions
418 nor about virtuals. */
419 if (!node->local.local
67348ccc 420 || node->alias
2bf86c84
JH
421 || (opt_for_fn (node->decl, flag_devirtualize)
422 && DECL_VIRTUAL_P (node->decl)))
08f835dc 423 return false;
67348ccc 424 gcc_assert (node->analyzed);
08f835dc 425 if (dump_file && (dump_flags & TDF_DETAILS))
fec39fa6 426 fprintf (dump_file, "Processing frequency %s\n", node->name ());
08f835dc 427
d52f5295
ML
428 node->call_for_symbol_thunks_and_aliases (ipa_propagate_frequency_1, &d,
429 true);
08f835dc
JH
430
431 if ((d.only_called_at_startup && !d.only_called_at_exit)
432 && !node->only_called_at_startup)
433 {
434 node->only_called_at_startup = true;
435 if (dump_file)
436 fprintf (dump_file, "Node %s promoted to only called at startup.\n",
fec39fa6 437 node->name ());
08f835dc
JH
438 changed = true;
439 }
440 if ((d.only_called_at_exit && !d.only_called_at_startup)
441 && !node->only_called_at_exit)
442 {
443 node->only_called_at_exit = true;
444 if (dump_file)
445 fprintf (dump_file, "Node %s promoted to only called at exit.\n",
fec39fa6 446 node->name ());
08f835dc
JH
447 changed = true;
448 }
daf5c770
JH
449
450 /* With profile we can decide on hot/normal based on count. */
451 if (node->count)
452 {
453 bool hot = false;
454 if (node->count >= get_hot_bb_threshold ())
455 hot = true;
456 if (!hot)
457 hot |= contains_hot_call_p (node);
458 if (hot)
459 {
460 if (node->frequency != NODE_FREQUENCY_HOT)
461 {
462 if (dump_file)
463 fprintf (dump_file, "Node %s promoted to hot.\n",
fec39fa6 464 node->name ());
daf5c770
JH
465 node->frequency = NODE_FREQUENCY_HOT;
466 return true;
467 }
468 return false;
469 }
470 else if (node->frequency == NODE_FREQUENCY_HOT)
471 {
472 if (dump_file)
473 fprintf (dump_file, "Node %s reduced to normal.\n",
fec39fa6 474 node->name ());
daf5c770
JH
475 node->frequency = NODE_FREQUENCY_NORMAL;
476 changed = true;
477 }
478 }
08f835dc
JH
479 /* These come either from profile or user hints; never update them. */
480 if (node->frequency == NODE_FREQUENCY_HOT
481 || node->frequency == NODE_FREQUENCY_UNLIKELY_EXECUTED)
482 return changed;
483 if (d.maybe_unlikely_executed)
484 {
485 node->frequency = NODE_FREQUENCY_UNLIKELY_EXECUTED;
486 if (dump_file)
487 fprintf (dump_file, "Node %s promoted to unlikely executed.\n",
fec39fa6 488 node->name ());
08f835dc
JH
489 changed = true;
490 }
491 else if (d.maybe_executed_once && node->frequency != NODE_FREQUENCY_EXECUTED_ONCE)
492 {
493 node->frequency = NODE_FREQUENCY_EXECUTED_ONCE;
494 if (dump_file)
495 fprintf (dump_file, "Node %s promoted to executed once.\n",
fec39fa6 496 node->name ());
08f835dc
JH
497 changed = true;
498 }
499 return changed;
500}
501
502/* Simple ipa profile pass propagating frequencies across the callgraph. */
503
504static unsigned int
505ipa_profile (void)
506{
507 struct cgraph_node **order;
508 struct cgraph_edge *e;
509 int order_pos;
510 bool something_changed = false;
511 int i;
512 gcov_type overall_time = 0, cutoff = 0, cumulated = 0, overall_size = 0;
513 struct cgraph_node *n,*n2;
514 int nindirect = 0, ncommon = 0, nunknown = 0, nuseless = 0, nconverted = 0;
515 bool node_map_initialized = false;
516
517 if (dump_file)
518 dump_histogram (dump_file, histogram);
519 for (i = 0; i < (int)histogram.length (); i++)
520 {
521 overall_time += histogram[i]->count * histogram[i]->time;
522 overall_size += histogram[i]->size;
523 }
524 if (overall_time)
525 {
526 gcov_type threshold;
527
528 gcc_assert (overall_size);
529 if (dump_file)
530 {
531 gcov_type min, cumulated_time = 0, cumulated_size = 0;
532
a9243bfc
RB
533 fprintf (dump_file, "Overall time: %"PRId64"\n",
534 (int64_t)overall_time);
08f835dc
JH
535 min = get_hot_bb_threshold ();
536 for (i = 0; i < (int)histogram.length () && histogram[i]->count >= min;
537 i++)
538 {
539 cumulated_time += histogram[i]->count * histogram[i]->time;
540 cumulated_size += histogram[i]->size;
541 }
a9243bfc 542 fprintf (dump_file, "GCOV min count: %"PRId64
08f835dc 543 " Time:%3.2f%% Size:%3.2f%%\n",
a9243bfc 544 (int64_t)min,
08f835dc
JH
545 cumulated_time * 100.0 / overall_time,
546 cumulated_size * 100.0 / overall_size);
547 }
548 cutoff = (overall_time * PARAM_VALUE (HOT_BB_COUNT_WS_PERMILLE) + 500) / 1000;
549 threshold = 0;
550 for (i = 0; cumulated < cutoff; i++)
551 {
552 cumulated += histogram[i]->count * histogram[i]->time;
553 threshold = histogram[i]->count;
554 }
555 if (!threshold)
556 threshold = 1;
557 if (dump_file)
558 {
559 gcov_type cumulated_time = 0, cumulated_size = 0;
560
561 for (i = 0;
562 i < (int)histogram.length () && histogram[i]->count >= threshold;
563 i++)
564 {
565 cumulated_time += histogram[i]->count * histogram[i]->time;
566 cumulated_size += histogram[i]->size;
567 }
a9243bfc 568 fprintf (dump_file, "Determined min count: %"PRId64
08f835dc 569 " Time:%3.2f%% Size:%3.2f%%\n",
a9243bfc 570 (int64_t)threshold,
08f835dc
JH
571 cumulated_time * 100.0 / overall_time,
572 cumulated_size * 100.0 / overall_size);
573 }
574 if (threshold > get_hot_bb_threshold ()
575 || in_lto_p)
576 {
577 if (dump_file)
578 fprintf (dump_file, "Threshold updated.\n");
579 set_hot_bb_threshold (threshold);
580 }
581 }
c3284718 582 histogram.release ();
08f835dc
JH
583 free_alloc_pool (histogram_pool);
584
585 /* Produce speculative calls: we saved common traget from porfiling into
586 e->common_target_id. Now, at link time, we can look up corresponding
587 function node and produce speculative call. */
588
589 FOR_EACH_DEFINED_FUNCTION (n)
590 {
591 bool update = false;
592
593 for (e = n->indirect_calls; e; e = e->next_callee)
594 {
595 if (n->count)
596 nindirect++;
597 if (e->indirect_info->common_target_id)
598 {
599 if (!node_map_initialized)
600 init_node_map (false);
601 node_map_initialized = true;
602 ncommon++;
603 n2 = find_func_by_profile_id (e->indirect_info->common_target_id);
604 if (n2)
605 {
606 if (dump_file)
607 {
608 fprintf (dump_file, "Indirect call -> direct call from"
609 " other module %s/%i => %s/%i, prob %3.2f\n",
2a72a953
DM
610 xstrdup_for_dump (n->name ()), n->order,
611 xstrdup_for_dump (n2->name ()), n2->order,
08f835dc
JH
612 e->indirect_info->common_target_probability
613 / (float)REG_BR_PROB_BASE);
614 }
615 if (e->indirect_info->common_target_probability
616 < REG_BR_PROB_BASE / 2)
617 {
618 nuseless++;
619 if (dump_file)
620 fprintf (dump_file,
621 "Not speculating: probability is too low.\n");
622 }
3dafb85c 623 else if (!e->maybe_hot_p ())
08f835dc
JH
624 {
625 nuseless++;
626 if (dump_file)
627 fprintf (dump_file,
628 "Not speculating: call is cold.\n");
629 }
d52f5295
ML
630 else if (n2->get_availability () <= AVAIL_INTERPOSABLE
631 && n2->can_be_discarded_p ())
08f835dc
JH
632 {
633 nuseless++;
634 if (dump_file)
635 fprintf (dump_file,
636 "Not speculating: target is overwritable "
637 "and can be discarded.\n");
638 }
639 else
640 {
641 /* Target may be overwritable, but profile says that
642 control flow goes to this particular implementation
643 of N2. Speculate on the local alias to allow inlining.
644 */
d52f5295 645 if (!n2->can_be_discarded_p ())
5b79657a
JH
646 {
647 cgraph_node *alias;
d52f5295 648 alias = dyn_cast<cgraph_node *> (n2->noninterposable_alias ());
5b79657a
JH
649 if (alias)
650 n2 = alias;
651 }
08f835dc 652 nconverted++;
3dafb85c
ML
653 e->make_speculative
654 (n2,
08f835dc
JH
655 apply_scale (e->count,
656 e->indirect_info->common_target_probability),
657 apply_scale (e->frequency,
658 e->indirect_info->common_target_probability));
659 update = true;
660 }
661 }
662 else
663 {
664 if (dump_file)
665 fprintf (dump_file, "Function with profile-id %i not found.\n",
666 e->indirect_info->common_target_id);
667 nunknown++;
668 }
669 }
670 }
671 if (update)
672 inline_update_overall_summary (n);
673 }
674 if (node_map_initialized)
675 del_node_map ();
676 if (dump_file && nindirect)
677 fprintf (dump_file,
678 "%i indirect calls trained.\n"
679 "%i (%3.2f%%) have common target.\n"
680 "%i (%3.2f%%) targets was not found.\n"
681 "%i (%3.2f%%) speculations seems useless.\n"
682 "%i (%3.2f%%) speculations produced.\n",
683 nindirect,
684 ncommon, ncommon * 100.0 / nindirect,
685 nunknown, nunknown * 100.0 / nindirect,
686 nuseless, nuseless * 100.0 / nindirect,
687 nconverted, nconverted * 100.0 / nindirect);
688
3dafb85c 689 order = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count);
08f835dc
JH
690 order_pos = ipa_reverse_postorder (order);
691 for (i = order_pos - 1; i >= 0; i--)
692 {
693 if (order[i]->local.local && ipa_propagate_frequency (order[i]))
694 {
695 for (e = order[i]->callees; e; e = e->next_callee)
67348ccc 696 if (e->callee->local.local && !e->callee->aux)
08f835dc
JH
697 {
698 something_changed = true;
67348ccc 699 e->callee->aux = (void *)1;
08f835dc
JH
700 }
701 }
67348ccc 702 order[i]->aux = NULL;
08f835dc
JH
703 }
704
705 while (something_changed)
706 {
707 something_changed = false;
708 for (i = order_pos - 1; i >= 0; i--)
709 {
67348ccc 710 if (order[i]->aux && ipa_propagate_frequency (order[i]))
08f835dc
JH
711 {
712 for (e = order[i]->callees; e; e = e->next_callee)
67348ccc 713 if (e->callee->local.local && !e->callee->aux)
08f835dc
JH
714 {
715 something_changed = true;
67348ccc 716 e->callee->aux = (void *)1;
08f835dc
JH
717 }
718 }
67348ccc 719 order[i]->aux = NULL;
08f835dc
JH
720 }
721 }
722 free (order);
723 return 0;
724}
725
08f835dc
JH
726namespace {
727
728const pass_data pass_data_ipa_profile =
729{
730 IPA_PASS, /* type */
731 "profile_estimate", /* name */
732 OPTGROUP_NONE, /* optinfo_flags */
08f835dc
JH
733 TV_IPA_PROFILE, /* tv_id */
734 0, /* properties_required */
735 0, /* properties_provided */
736 0, /* properties_destroyed */
737 0, /* todo_flags_start */
738 0, /* todo_flags_finish */
739};
740
741class pass_ipa_profile : public ipa_opt_pass_d
742{
743public:
c3284718
RS
744 pass_ipa_profile (gcc::context *ctxt)
745 : ipa_opt_pass_d (pass_data_ipa_profile, ctxt,
746 ipa_profile_generate_summary, /* generate_summary */
747 ipa_profile_write_summary, /* write_summary */
748 ipa_profile_read_summary, /* read_summary */
749 NULL, /* write_optimization_summary */
750 NULL, /* read_optimization_summary */
751 NULL, /* stmt_fixup */
752 0, /* function_transform_todo_flags_start */
753 NULL, /* function_transform */
754 NULL) /* variable_transform */
08f835dc
JH
755 {}
756
757 /* opt_pass methods: */
2bf86c84 758 virtual bool gate (function *) { return flag_ipa_profile || in_lto_p; }
be55bfe6 759 virtual unsigned int execute (function *) { return ipa_profile (); }
08f835dc
JH
760
761}; // class pass_ipa_profile
762
763} // anon namespace
764
765ipa_opt_pass_d *
766make_pass_ipa_profile (gcc::context *ctxt)
767{
768 return new pass_ipa_profile (ctxt);
769}