Re: Refactor copying decl section names
[official-gcc.git] / gcc / profile.c
blob454095916294194052b98499ae6143e256613417
1 /* Calculate branch probabilities, and basic block execution counts.
2 Copyright (C) 1990-2020 Free Software Foundation, Inc.
3 Contributed by James E. Wilson, UC Berkeley/Cygnus Support;
4 based on some ideas from Dain Samples of UC Berkeley.
5 Further mangling by Bob Manson, Cygnus Support.
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify it under
10 the terms of the GNU General Public License as published by the Free
11 Software Foundation; either version 3, or (at your option) any later
12 version.
14 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
15 WARRANTY; without even the implied warranty of MERCHANTABILITY or
16 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 for more details.
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
23 /* Generate basic block profile instrumentation and auxiliary files.
24 Profile generation is optimized, so that not all arcs in the basic
25 block graph need instrumenting. First, the BB graph is closed with
26 one entry (function start), and one exit (function exit). Any
27 ABNORMAL_EDGE cannot be instrumented (because there is no control
28 path to place the code). We close the graph by inserting fake
29 EDGE_FAKE edges to the EXIT_BLOCK, from the sources of abnormal
30 edges that do not go to the exit_block. We ignore such abnormal
31 edges. Naturally these fake edges are never directly traversed,
32 and so *cannot* be directly instrumented. Some other graph
33 massaging is done. To optimize the instrumentation we generate the
34 BB minimal span tree, only edges that are not on the span tree
35 (plus the entry point) need instrumenting. From that information
36 all other edge counts can be deduced. By construction all fake
37 edges must be on the spanning tree. We also attempt to place
38 EDGE_CRITICAL edges on the spanning tree.
40 The auxiliary files generated are <dumpbase>.gcno (at compile time)
41 and <dumpbase>.gcda (at run time). The format is
42 described in full in gcov-io.h. */
44 /* ??? Register allocation should use basic block execution counts to
45 give preference to the most commonly executed blocks. */
47 /* ??? Should calculate branch probabilities before instrumenting code, since
48 then we can use arc counts to help decide which arcs to instrument. */
50 #include "config.h"
51 #include "system.h"
52 #include "coretypes.h"
53 #include "backend.h"
54 #include "rtl.h"
55 #include "tree.h"
56 #include "gimple.h"
57 #include "cfghooks.h"
58 #include "cgraph.h"
59 #include "coverage.h"
60 #include "diagnostic-core.h"
61 #include "cfganal.h"
62 #include "value-prof.h"
63 #include "gimple-iterator.h"
64 #include "tree-cfg.h"
65 #include "dumpfile.h"
66 #include "cfgloop.h"
68 #include "profile.h"
70 /* Map from BBs/edges to gcov counters. */
71 vec<gcov_type> bb_gcov_counts;
72 hash_map<edge,gcov_type> *edge_gcov_counts;
74 struct bb_profile_info {
75 unsigned int count_valid : 1;
77 /* Number of successor and predecessor edges. */
78 gcov_type succ_count;
79 gcov_type pred_count;
82 #define BB_INFO(b) ((struct bb_profile_info *) (b)->aux)
85 /* Counter summary from the last set of coverage counts read. */
87 gcov_summary *profile_info;
89 /* Collect statistics on the performance of this pass for the entire source
90 file. */
92 static int total_num_blocks;
93 static int total_num_edges;
94 static int total_num_edges_ignored;
95 static int total_num_edges_instrumented;
96 static int total_num_blocks_created;
97 static int total_num_passes;
98 static int total_num_times_called;
99 static int total_hist_br_prob[20];
100 static int total_num_branches;
102 /* Forward declarations. */
103 static void find_spanning_tree (struct edge_list *);
105 /* Add edge instrumentation code to the entire insn chain.
107 F is the first insn of the chain.
108 NUM_BLOCKS is the number of basic blocks found in F. */
110 static unsigned
111 instrument_edges (struct edge_list *el)
113 unsigned num_instr_edges = 0;
114 int num_edges = NUM_EDGES (el);
115 basic_block bb;
117 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR_FOR_FN (cfun), NULL, next_bb)
119 edge e;
120 edge_iterator ei;
122 FOR_EACH_EDGE (e, ei, bb->succs)
124 struct edge_profile_info *inf = EDGE_INFO (e);
126 if (!inf->ignore && !inf->on_tree)
128 gcc_assert (!(e->flags & EDGE_ABNORMAL));
129 if (dump_file)
130 fprintf (dump_file, "Edge %d to %d instrumented%s\n",
131 e->src->index, e->dest->index,
132 EDGE_CRITICAL_P (e) ? " (and split)" : "");
133 gimple_gen_edge_profiler (num_instr_edges++, e);
138 total_num_blocks_created += num_edges;
139 if (dump_file)
140 fprintf (dump_file, "%d edges instrumented\n", num_instr_edges);
141 return num_instr_edges;
144 /* Add code to measure histograms for values in list VALUES. */
145 static void
146 instrument_values (histogram_values values)
148 unsigned i;
150 /* Emit code to generate the histograms before the insns. */
152 for (i = 0; i < values.length (); i++)
154 histogram_value hist = values[i];
155 unsigned t = COUNTER_FOR_HIST_TYPE (hist->type);
157 if (!coverage_counter_alloc (t, hist->n_counters))
158 continue;
160 switch (hist->type)
162 case HIST_TYPE_INTERVAL:
163 gimple_gen_interval_profiler (hist, t);
164 break;
166 case HIST_TYPE_POW2:
167 gimple_gen_pow2_profiler (hist, t);
168 break;
170 case HIST_TYPE_TOPN_VALUES:
171 gimple_gen_topn_values_profiler (hist, t);
172 break;
174 case HIST_TYPE_INDIR_CALL:
175 gimple_gen_ic_profiler (hist, t);
176 break;
178 case HIST_TYPE_AVERAGE:
179 gimple_gen_average_profiler (hist, t);
180 break;
182 case HIST_TYPE_IOR:
183 gimple_gen_ior_profiler (hist, t);
184 break;
186 case HIST_TYPE_TIME_PROFILE:
187 gimple_gen_time_profiler (t);
188 break;
190 default:
191 gcc_unreachable ();
197 /* Computes hybrid profile for all matching entries in da_file.
199 CFG_CHECKSUM is the precomputed checksum for the CFG. */
201 static gcov_type *
202 get_exec_counts (unsigned cfg_checksum, unsigned lineno_checksum)
204 unsigned num_edges = 0;
205 basic_block bb;
206 gcov_type *counts;
208 /* Count the edges to be (possibly) instrumented. */
209 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR_FOR_FN (cfun), NULL, next_bb)
211 edge e;
212 edge_iterator ei;
214 FOR_EACH_EDGE (e, ei, bb->succs)
215 if (!EDGE_INFO (e)->ignore && !EDGE_INFO (e)->on_tree)
216 num_edges++;
219 counts = get_coverage_counts (GCOV_COUNTER_ARCS, cfg_checksum,
220 lineno_checksum, num_edges);
221 if (!counts)
222 return NULL;
224 return counts;
227 static bool
228 is_edge_inconsistent (vec<edge, va_gc> *edges)
230 edge e;
231 edge_iterator ei;
232 FOR_EACH_EDGE (e, ei, edges)
234 if (!EDGE_INFO (e)->ignore)
236 if (edge_gcov_count (e) < 0
237 && (!(e->flags & EDGE_FAKE)
238 || !block_ends_with_call_p (e->src)))
240 if (dump_file)
242 fprintf (dump_file,
243 "Edge %i->%i is inconsistent, count%" PRId64,
244 e->src->index, e->dest->index, edge_gcov_count (e));
245 dump_bb (dump_file, e->src, 0, TDF_DETAILS);
246 dump_bb (dump_file, e->dest, 0, TDF_DETAILS);
248 return true;
252 return false;
255 static void
256 correct_negative_edge_counts (void)
258 basic_block bb;
259 edge e;
260 edge_iterator ei;
262 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR_FOR_FN (cfun), NULL, next_bb)
264 FOR_EACH_EDGE (e, ei, bb->succs)
266 if (edge_gcov_count (e) < 0)
267 edge_gcov_count (e) = 0;
272 /* Check consistency.
273 Return true if inconsistency is found. */
274 static bool
275 is_inconsistent (void)
277 basic_block bb;
278 bool inconsistent = false;
279 FOR_EACH_BB_FN (bb, cfun)
281 inconsistent |= is_edge_inconsistent (bb->preds);
282 if (!dump_file && inconsistent)
283 return true;
284 inconsistent |= is_edge_inconsistent (bb->succs);
285 if (!dump_file && inconsistent)
286 return true;
287 if (bb_gcov_count (bb) < 0)
289 if (dump_file)
291 fprintf (dump_file, "BB %i count is negative "
292 "%" PRId64,
293 bb->index,
294 bb_gcov_count (bb));
295 dump_bb (dump_file, bb, 0, TDF_DETAILS);
297 inconsistent = true;
299 if (bb_gcov_count (bb) != sum_edge_counts (bb->preds))
301 if (dump_file)
303 fprintf (dump_file, "BB %i count does not match sum of incoming edges "
304 "%" PRId64" should be %" PRId64,
305 bb->index,
306 bb_gcov_count (bb),
307 sum_edge_counts (bb->preds));
308 dump_bb (dump_file, bb, 0, TDF_DETAILS);
310 inconsistent = true;
312 if (bb_gcov_count (bb) != sum_edge_counts (bb->succs) &&
313 ! (find_edge (bb, EXIT_BLOCK_PTR_FOR_FN (cfun)) != NULL
314 && block_ends_with_call_p (bb)))
316 if (dump_file)
318 fprintf (dump_file, "BB %i count does not match sum of outgoing edges "
319 "%" PRId64" should be %" PRId64,
320 bb->index,
321 bb_gcov_count (bb),
322 sum_edge_counts (bb->succs));
323 dump_bb (dump_file, bb, 0, TDF_DETAILS);
325 inconsistent = true;
327 if (!dump_file && inconsistent)
328 return true;
331 return inconsistent;
334 /* Set each basic block count to the sum of its outgoing edge counts */
335 static void
336 set_bb_counts (void)
338 basic_block bb;
339 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR_FOR_FN (cfun), NULL, next_bb)
341 bb_gcov_count (bb) = sum_edge_counts (bb->succs);
342 gcc_assert (bb_gcov_count (bb) >= 0);
346 /* Reads profile data and returns total number of edge counts read */
347 static int
348 read_profile_edge_counts (gcov_type *exec_counts)
350 basic_block bb;
351 int num_edges = 0;
352 int exec_counts_pos = 0;
353 /* For each edge not on the spanning tree, set its execution count from
354 the .da file. */
355 /* The first count in the .da file is the number of times that the function
356 was entered. This is the exec_count for block zero. */
358 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR_FOR_FN (cfun), NULL, next_bb)
360 edge e;
361 edge_iterator ei;
363 FOR_EACH_EDGE (e, ei, bb->succs)
364 if (!EDGE_INFO (e)->ignore && !EDGE_INFO (e)->on_tree)
366 num_edges++;
367 if (exec_counts)
368 edge_gcov_count (e) = exec_counts[exec_counts_pos++];
369 else
370 edge_gcov_count (e) = 0;
372 EDGE_INFO (e)->count_valid = 1;
373 BB_INFO (bb)->succ_count--;
374 BB_INFO (e->dest)->pred_count--;
375 if (dump_file)
377 fprintf (dump_file, "\nRead edge from %i to %i, count:",
378 bb->index, e->dest->index);
379 fprintf (dump_file, "%" PRId64,
380 (int64_t) edge_gcov_count (e));
385 return num_edges;
389 /* Compute the branch probabilities for the various branches.
390 Annotate them accordingly.
392 CFG_CHECKSUM is the precomputed checksum for the CFG. */
394 static void
395 compute_branch_probabilities (unsigned cfg_checksum, unsigned lineno_checksum)
397 basic_block bb;
398 int i;
399 int num_edges = 0;
400 int changes;
401 int passes;
402 int hist_br_prob[20];
403 int num_branches;
404 gcov_type *exec_counts = get_exec_counts (cfg_checksum, lineno_checksum);
405 int inconsistent = 0;
407 /* Very simple sanity checks so we catch bugs in our profiling code. */
408 if (!profile_info)
410 if (dump_file)
411 fprintf (dump_file, "Profile info is missing; giving up\n");
412 return;
415 bb_gcov_counts.safe_grow_cleared (last_basic_block_for_fn (cfun), true);
416 edge_gcov_counts = new hash_map<edge,gcov_type>;
418 /* Attach extra info block to each bb. */
419 alloc_aux_for_blocks (sizeof (struct bb_profile_info));
420 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR_FOR_FN (cfun), NULL, next_bb)
422 edge e;
423 edge_iterator ei;
425 FOR_EACH_EDGE (e, ei, bb->succs)
426 if (!EDGE_INFO (e)->ignore)
427 BB_INFO (bb)->succ_count++;
428 FOR_EACH_EDGE (e, ei, bb->preds)
429 if (!EDGE_INFO (e)->ignore)
430 BB_INFO (bb)->pred_count++;
433 /* Avoid predicting entry on exit nodes. */
434 BB_INFO (EXIT_BLOCK_PTR_FOR_FN (cfun))->succ_count = 2;
435 BB_INFO (ENTRY_BLOCK_PTR_FOR_FN (cfun))->pred_count = 2;
437 num_edges = read_profile_edge_counts (exec_counts);
439 if (dump_file)
440 fprintf (dump_file, "\n%d edge counts read\n", num_edges);
442 /* For every block in the file,
443 - if every exit/entrance edge has a known count, then set the block count
444 - if the block count is known, and every exit/entrance edge but one has
445 a known execution count, then set the count of the remaining edge
447 As edge counts are set, decrement the succ/pred count, but don't delete
448 the edge, that way we can easily tell when all edges are known, or only
449 one edge is unknown. */
451 /* The order that the basic blocks are iterated through is important.
452 Since the code that finds spanning trees starts with block 0, low numbered
453 edges are put on the spanning tree in preference to high numbered edges.
454 Hence, most instrumented edges are at the end. Graph solving works much
455 faster if we propagate numbers from the end to the start.
457 This takes an average of slightly more than 3 passes. */
459 changes = 1;
460 passes = 0;
461 while (changes)
463 passes++;
464 changes = 0;
465 FOR_BB_BETWEEN (bb, EXIT_BLOCK_PTR_FOR_FN (cfun), NULL, prev_bb)
467 struct bb_profile_info *bi = BB_INFO (bb);
468 if (! bi->count_valid)
470 if (bi->succ_count == 0)
472 edge e;
473 edge_iterator ei;
474 gcov_type total = 0;
476 FOR_EACH_EDGE (e, ei, bb->succs)
477 total += edge_gcov_count (e);
478 bb_gcov_count (bb) = total;
479 bi->count_valid = 1;
480 changes = 1;
482 else if (bi->pred_count == 0)
484 edge e;
485 edge_iterator ei;
486 gcov_type total = 0;
488 FOR_EACH_EDGE (e, ei, bb->preds)
489 total += edge_gcov_count (e);
490 bb_gcov_count (bb) = total;
491 bi->count_valid = 1;
492 changes = 1;
495 if (bi->count_valid)
497 if (bi->succ_count == 1)
499 edge e;
500 edge_iterator ei;
501 gcov_type total = 0;
503 /* One of the counts will be invalid, but it is zero,
504 so adding it in also doesn't hurt. */
505 FOR_EACH_EDGE (e, ei, bb->succs)
506 total += edge_gcov_count (e);
508 /* Search for the invalid edge, and set its count. */
509 FOR_EACH_EDGE (e, ei, bb->succs)
510 if (! EDGE_INFO (e)->count_valid && ! EDGE_INFO (e)->ignore)
511 break;
513 /* Calculate count for remaining edge by conservation. */
514 total = bb_gcov_count (bb) - total;
516 gcc_assert (e);
517 EDGE_INFO (e)->count_valid = 1;
518 edge_gcov_count (e) = total;
519 bi->succ_count--;
521 BB_INFO (e->dest)->pred_count--;
522 changes = 1;
524 if (bi->pred_count == 1)
526 edge e;
527 edge_iterator ei;
528 gcov_type total = 0;
530 /* One of the counts will be invalid, but it is zero,
531 so adding it in also doesn't hurt. */
532 FOR_EACH_EDGE (e, ei, bb->preds)
533 total += edge_gcov_count (e);
535 /* Search for the invalid edge, and set its count. */
536 FOR_EACH_EDGE (e, ei, bb->preds)
537 if (!EDGE_INFO (e)->count_valid && !EDGE_INFO (e)->ignore)
538 break;
540 /* Calculate count for remaining edge by conservation. */
541 total = bb_gcov_count (bb) - total + edge_gcov_count (e);
543 gcc_assert (e);
544 EDGE_INFO (e)->count_valid = 1;
545 edge_gcov_count (e) = total;
546 bi->pred_count--;
548 BB_INFO (e->src)->succ_count--;
549 changes = 1;
555 total_num_passes += passes;
556 if (dump_file)
557 fprintf (dump_file, "Graph solving took %d passes.\n\n", passes);
559 /* If the graph has been correctly solved, every block will have a
560 succ and pred count of zero. */
561 FOR_EACH_BB_FN (bb, cfun)
563 gcc_assert (!BB_INFO (bb)->succ_count && !BB_INFO (bb)->pred_count);
566 /* Check for inconsistent basic block counts */
567 inconsistent = is_inconsistent ();
569 if (inconsistent)
571 if (flag_profile_correction)
573 /* Inconsistency detected. Make it flow-consistent. */
574 static int informed = 0;
575 if (dump_enabled_p () && informed == 0)
577 informed = 1;
578 dump_printf_loc (MSG_NOTE,
579 dump_user_location_t::from_location_t (input_location),
580 "correcting inconsistent profile data\n");
582 correct_negative_edge_counts ();
583 /* Set bb counts to the sum of the outgoing edge counts */
584 set_bb_counts ();
585 if (dump_file)
586 fprintf (dump_file, "\nCalling mcf_smooth_cfg\n");
587 mcf_smooth_cfg ();
589 else
590 error ("corrupted profile info: profile data is not flow-consistent");
593 /* For every edge, calculate its branch probability and add a reg_note
594 to the branch insn to indicate this. */
596 for (i = 0; i < 20; i++)
597 hist_br_prob[i] = 0;
598 num_branches = 0;
600 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR_FOR_FN (cfun), NULL, next_bb)
602 edge e;
603 edge_iterator ei;
605 if (bb_gcov_count (bb) < 0)
607 error ("corrupted profile info: number of iterations for basic block %d thought to be %i",
608 bb->index, (int)bb_gcov_count (bb));
609 bb_gcov_count (bb) = 0;
611 FOR_EACH_EDGE (e, ei, bb->succs)
613 /* Function may return twice in the cased the called function is
614 setjmp or calls fork, but we can't represent this by extra
615 edge from the entry, since extra edge from the exit is
616 already present. We get negative frequency from the entry
617 point. */
618 if ((edge_gcov_count (e) < 0
619 && e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
620 || (edge_gcov_count (e) > bb_gcov_count (bb)
621 && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)))
623 if (block_ends_with_call_p (bb))
624 edge_gcov_count (e) = edge_gcov_count (e) < 0
625 ? 0 : bb_gcov_count (bb);
627 if (edge_gcov_count (e) < 0
628 || edge_gcov_count (e) > bb_gcov_count (bb))
630 error ("corrupted profile info: number of executions for edge %d-%d thought to be %i",
631 e->src->index, e->dest->index,
632 (int)edge_gcov_count (e));
633 edge_gcov_count (e) = bb_gcov_count (bb) / 2;
636 if (bb_gcov_count (bb))
638 bool set_to_guessed = false;
639 FOR_EACH_EDGE (e, ei, bb->succs)
641 bool prev_never = e->probability == profile_probability::never ();
642 e->probability = profile_probability::probability_in_gcov_type
643 (edge_gcov_count (e), bb_gcov_count (bb));
644 if (e->probability == profile_probability::never ()
645 && !prev_never
646 && flag_profile_partial_training)
647 set_to_guessed = true;
649 if (set_to_guessed)
650 FOR_EACH_EDGE (e, ei, bb->succs)
651 e->probability = e->probability.guessed ();
652 if (bb->index >= NUM_FIXED_BLOCKS
653 && block_ends_with_condjump_p (bb)
654 && EDGE_COUNT (bb->succs) >= 2)
656 int prob;
657 edge e;
658 int index;
660 /* Find the branch edge. It is possible that we do have fake
661 edges here. */
662 FOR_EACH_EDGE (e, ei, bb->succs)
663 if (!(e->flags & (EDGE_FAKE | EDGE_FALLTHRU)))
664 break;
666 prob = e->probability.to_reg_br_prob_base ();
667 index = prob * 20 / REG_BR_PROB_BASE;
669 if (index == 20)
670 index = 19;
671 hist_br_prob[index]++;
673 num_branches++;
676 /* As a last resort, distribute the probabilities evenly.
677 Use simple heuristics that if there are normal edges,
678 give all abnormals frequency of 0, otherwise distribute the
679 frequency over abnormals (this is the case of noreturn
680 calls). */
681 else if (profile_status_for_fn (cfun) == PROFILE_ABSENT)
683 int total = 0;
685 FOR_EACH_EDGE (e, ei, bb->succs)
686 if (!(e->flags & (EDGE_COMPLEX | EDGE_FAKE)))
687 total ++;
688 if (total)
690 FOR_EACH_EDGE (e, ei, bb->succs)
691 if (!(e->flags & (EDGE_COMPLEX | EDGE_FAKE)))
692 e->probability
693 = profile_probability::guessed_always ().apply_scale (1, total);
694 else
695 e->probability = profile_probability::never ();
697 else
699 total += EDGE_COUNT (bb->succs);
700 FOR_EACH_EDGE (e, ei, bb->succs)
701 e->probability
702 = profile_probability::guessed_always ().apply_scale (1, total);
704 if (bb->index >= NUM_FIXED_BLOCKS
705 && block_ends_with_condjump_p (bb)
706 && EDGE_COUNT (bb->succs) >= 2)
707 num_branches++;
711 if (exec_counts
712 && (bb_gcov_count (ENTRY_BLOCK_PTR_FOR_FN (cfun))
713 || !flag_profile_partial_training))
714 profile_status_for_fn (cfun) = PROFILE_READ;
716 /* If we have real data, use them! */
717 if (bb_gcov_count (ENTRY_BLOCK_PTR_FOR_FN (cfun))
718 || !flag_guess_branch_prob)
719 FOR_ALL_BB_FN (bb, cfun)
720 if (bb_gcov_count (bb) || !flag_profile_partial_training)
721 bb->count = profile_count::from_gcov_type (bb_gcov_count (bb));
722 else
723 bb->count = profile_count::guessed_zero ();
724 /* If function was not trained, preserve local estimates including statically
725 determined zero counts. */
726 else if (profile_status_for_fn (cfun) == PROFILE_READ
727 && !flag_profile_partial_training)
728 FOR_ALL_BB_FN (bb, cfun)
729 if (!(bb->count == profile_count::zero ()))
730 bb->count = bb->count.global0 ();
732 bb_gcov_counts.release ();
733 delete edge_gcov_counts;
734 edge_gcov_counts = NULL;
736 update_max_bb_count ();
738 if (dump_file)
740 fprintf (dump_file, " Profile feedback for function");
741 fprintf (dump_file, ((profile_status_for_fn (cfun) == PROFILE_READ)
742 ? " is available \n"
743 : " is not available \n"));
745 fprintf (dump_file, "%d branches\n", num_branches);
746 if (num_branches)
747 for (i = 0; i < 10; i++)
748 fprintf (dump_file, "%d%% branches in range %d-%d%%\n",
749 (hist_br_prob[i] + hist_br_prob[19-i]) * 100 / num_branches,
750 5 * i, 5 * i + 5);
752 total_num_branches += num_branches;
753 for (i = 0; i < 20; i++)
754 total_hist_br_prob[i] += hist_br_prob[i];
756 fputc ('\n', dump_file);
757 fputc ('\n', dump_file);
760 free_aux_for_blocks ();
763 /* Sort the histogram value and count for TOPN and INDIR_CALL type. */
765 static void
766 sort_hist_values (histogram_value hist)
768 gcc_assert (hist->type == HIST_TYPE_TOPN_VALUES
769 || hist->type == HIST_TYPE_INDIR_CALL);
771 int counters = hist->hvalue.counters[1];
772 for (int i = 0; i < counters - 1; i++)
773 /* Hist value is organized as:
774 [total_executions, N, value1, counter1, ..., valueN, counterN]
775 Use decrease bubble sort to rearrange it. The sort starts from <value1,
776 counter1> and compares counter first. If counter is same, compares the
777 value, exchange it if small to keep stable. */
780 bool swapped = false;
781 for (int j = 0; j < counters - 1 - i; j++)
783 gcov_type *p = &hist->hvalue.counters[2 * j + 2];
784 if (p[1] < p[3] || (p[1] == p[3] && p[0] < p[2]))
786 std::swap (p[0], p[2]);
787 std::swap (p[1], p[3]);
788 swapped = true;
791 if (!swapped)
792 break;
795 /* Load value histograms values whose description is stored in VALUES array
796 from .gcda file.
798 CFG_CHECKSUM is the precomputed checksum for the CFG. */
800 static void
801 compute_value_histograms (histogram_values values, unsigned cfg_checksum,
802 unsigned lineno_checksum)
804 unsigned i, j, t, any;
805 unsigned n_histogram_counters[GCOV_N_VALUE_COUNTERS];
806 gcov_type *histogram_counts[GCOV_N_VALUE_COUNTERS];
807 gcov_type *act_count[GCOV_N_VALUE_COUNTERS];
808 gcov_type *aact_count;
809 struct cgraph_node *node;
811 for (t = 0; t < GCOV_N_VALUE_COUNTERS; t++)
812 n_histogram_counters[t] = 0;
814 for (i = 0; i < values.length (); i++)
816 histogram_value hist = values[i];
817 n_histogram_counters[(int) hist->type] += hist->n_counters;
820 any = 0;
821 for (t = 0; t < GCOV_N_VALUE_COUNTERS; t++)
823 if (!n_histogram_counters[t])
825 histogram_counts[t] = NULL;
826 continue;
829 histogram_counts[t] = get_coverage_counts (COUNTER_FOR_HIST_TYPE (t),
830 cfg_checksum,
831 lineno_checksum,
832 n_histogram_counters[t]);
833 if (histogram_counts[t])
834 any = 1;
835 act_count[t] = histogram_counts[t];
837 if (!any)
838 return;
840 for (i = 0; i < values.length (); i++)
842 histogram_value hist = values[i];
843 gimple *stmt = hist->hvalue.stmt;
845 t = (int) hist->type;
846 bool topn_p = (hist->type == HIST_TYPE_TOPN_VALUES
847 || hist->type == HIST_TYPE_INDIR_CALL);
849 /* TOP N counter uses variable number of counters. */
850 if (topn_p)
852 unsigned total_size;
853 if (act_count[t])
854 total_size = 2 + 2 * act_count[t][1];
855 else
856 total_size = 2;
857 gimple_add_histogram_value (cfun, stmt, hist);
858 hist->n_counters = total_size;
859 hist->hvalue.counters = XNEWVEC (gcov_type, hist->n_counters);
860 for (j = 0; j < hist->n_counters; j++)
861 if (act_count[t])
862 hist->hvalue.counters[j] = act_count[t][j];
863 else
864 hist->hvalue.counters[j] = 0;
865 act_count[t] += hist->n_counters;
866 sort_hist_values (hist);
868 else
870 aact_count = act_count[t];
872 if (act_count[t])
873 act_count[t] += hist->n_counters;
875 gimple_add_histogram_value (cfun, stmt, hist);
876 hist->hvalue.counters = XNEWVEC (gcov_type, hist->n_counters);
877 for (j = 0; j < hist->n_counters; j++)
878 if (aact_count)
879 hist->hvalue.counters[j] = aact_count[j];
880 else
881 hist->hvalue.counters[j] = 0;
884 /* Time profiler counter is not related to any statement,
885 so that we have to read the counter and set the value to
886 the corresponding call graph node. */
887 if (hist->type == HIST_TYPE_TIME_PROFILE)
889 node = cgraph_node::get (hist->fun->decl);
890 if (hist->hvalue.counters[0] >= 0
891 && hist->hvalue.counters[0] < INT_MAX / 2)
892 node->tp_first_run = hist->hvalue.counters[0];
893 else
895 if (flag_profile_correction)
896 error ("corrupted profile info: invalid time profile");
897 node->tp_first_run = 0;
900 if (dump_file)
901 fprintf (dump_file, "Read tp_first_run: %d\n", node->tp_first_run);
905 for (t = 0; t < GCOV_N_VALUE_COUNTERS; t++)
906 free (histogram_counts[t]);
909 /* Location triplet which records a location. */
910 struct location_triplet
912 const char *filename;
913 int lineno;
914 int bb_index;
917 /* Traits class for streamed_locations hash set below. */
919 struct location_triplet_hash : typed_noop_remove <location_triplet>
921 typedef location_triplet value_type;
922 typedef location_triplet compare_type;
924 static hashval_t
925 hash (const location_triplet &ref)
927 inchash::hash hstate (0);
928 if (ref.filename)
929 hstate.add_int (strlen (ref.filename));
930 hstate.add_int (ref.lineno);
931 hstate.add_int (ref.bb_index);
932 return hstate.end ();
935 static bool
936 equal (const location_triplet &ref1, const location_triplet &ref2)
938 return ref1.lineno == ref2.lineno
939 && ref1.bb_index == ref2.bb_index
940 && ref1.filename != NULL
941 && ref2.filename != NULL
942 && strcmp (ref1.filename, ref2.filename) == 0;
945 static void
946 mark_deleted (location_triplet &ref)
948 ref.lineno = -1;
951 static const bool empty_zero_p = false;
953 static void
954 mark_empty (location_triplet &ref)
956 ref.lineno = -2;
959 static bool
960 is_deleted (const location_triplet &ref)
962 return ref.lineno == -1;
965 static bool
966 is_empty (const location_triplet &ref)
968 return ref.lineno == -2;
975 /* When passed NULL as file_name, initialize.
976 When passed something else, output the necessary commands to change
977 line to LINE and offset to FILE_NAME. */
978 static void
979 output_location (hash_set<location_triplet_hash> *streamed_locations,
980 char const *file_name, int line,
981 gcov_position_t *offset, basic_block bb)
983 static char const *prev_file_name;
984 static int prev_line;
985 bool name_differs, line_differs;
987 location_triplet triplet;
988 triplet.filename = file_name;
989 triplet.lineno = line;
990 triplet.bb_index = bb ? bb->index : 0;
992 if (streamed_locations->add (triplet))
993 return;
995 if (!file_name)
997 prev_file_name = NULL;
998 prev_line = -1;
999 return;
1002 name_differs = !prev_file_name || filename_cmp (file_name, prev_file_name);
1003 line_differs = prev_line != line;
1005 if (!*offset)
1007 *offset = gcov_write_tag (GCOV_TAG_LINES);
1008 gcov_write_unsigned (bb->index);
1009 name_differs = line_differs = true;
1012 /* If this is a new source file, then output the
1013 file's name to the .bb file. */
1014 if (name_differs)
1016 prev_file_name = file_name;
1017 gcov_write_unsigned (0);
1018 gcov_write_filename (prev_file_name);
1020 if (line_differs)
1022 gcov_write_unsigned (line);
1023 prev_line = line;
1027 /* Helper for qsort so edges get sorted from highest frequency to smallest.
1028 This controls the weight for minimal spanning tree algorithm */
1029 static int
1030 compare_freqs (const void *p1, const void *p2)
1032 const_edge e1 = *(const const_edge *)p1;
1033 const_edge e2 = *(const const_edge *)p2;
1035 /* Critical edges needs to be split which introduce extra control flow.
1036 Make them more heavy. */
1037 int m1 = EDGE_CRITICAL_P (e1) ? 2 : 1;
1038 int m2 = EDGE_CRITICAL_P (e2) ? 2 : 1;
1040 if (EDGE_FREQUENCY (e1) * m1 + m1 != EDGE_FREQUENCY (e2) * m2 + m2)
1041 return EDGE_FREQUENCY (e2) * m2 + m2 - EDGE_FREQUENCY (e1) * m1 - m1;
1042 /* Stabilize sort. */
1043 if (e1->src->index != e2->src->index)
1044 return e2->src->index - e1->src->index;
1045 return e2->dest->index - e1->dest->index;
1048 /* Only read execution count for thunks. */
1050 void
1051 read_thunk_profile (struct cgraph_node *node)
1053 tree old = current_function_decl;
1054 current_function_decl = node->decl;
1055 gcov_type *counts = get_coverage_counts (GCOV_COUNTER_ARCS, 0, 0, 1);
1056 if (counts)
1058 node->callees->count = node->count
1059 = profile_count::from_gcov_type (counts[0]);
1060 free (counts);
1062 current_function_decl = old;
1063 return;
1067 /* Instrument and/or analyze program behavior based on program the CFG.
1069 This function creates a representation of the control flow graph (of
1070 the function being compiled) that is suitable for the instrumentation
1071 of edges and/or converting measured edge counts to counts on the
1072 complete CFG.
1074 When FLAG_PROFILE_ARCS is nonzero, this function instruments the edges in
1075 the flow graph that are needed to reconstruct the dynamic behavior of the
1076 flow graph. This data is written to the gcno file for gcov.
1078 When FLAG_BRANCH_PROBABILITIES is nonzero, this function reads auxiliary
1079 information from the gcda file containing edge count information from
1080 previous executions of the function being compiled. In this case, the
1081 control flow graph is annotated with actual execution counts by
1082 compute_branch_probabilities().
1084 Main entry point of this file. */
1086 void
1087 branch_prob (bool thunk)
1089 basic_block bb;
1090 unsigned i;
1091 unsigned num_edges, ignored_edges;
1092 unsigned num_instrumented;
1093 struct edge_list *el;
1094 histogram_values values = histogram_values ();
1095 unsigned cfg_checksum, lineno_checksum;
1097 total_num_times_called++;
1099 flow_call_edges_add (NULL);
1100 add_noreturn_fake_exit_edges ();
1102 hash_set <location_triplet_hash> streamed_locations;
1104 if (!thunk)
1106 /* We can't handle cyclic regions constructed using abnormal edges.
1107 To avoid these we replace every source of abnormal edge by a fake
1108 edge from entry node and every destination by fake edge to exit.
1109 This keeps graph acyclic and our calculation exact for all normal
1110 edges except for exit and entrance ones.
1112 We also add fake exit edges for each call and asm statement in the
1113 basic, since it may not return. */
1115 FOR_EACH_BB_FN (bb, cfun)
1117 int need_exit_edge = 0, need_entry_edge = 0;
1118 int have_exit_edge = 0, have_entry_edge = 0;
1119 edge e;
1120 edge_iterator ei;
1122 /* Functions returning multiple times are not handled by extra edges.
1123 Instead we simply allow negative counts on edges from exit to the
1124 block past call and corresponding probabilities. We can't go
1125 with the extra edges because that would result in flowgraph that
1126 needs to have fake edges outside the spanning tree. */
1128 FOR_EACH_EDGE (e, ei, bb->succs)
1130 gimple_stmt_iterator gsi;
1131 gimple *last = NULL;
1133 /* It may happen that there are compiler generated statements
1134 without a locus at all. Go through the basic block from the
1135 last to the first statement looking for a locus. */
1136 for (gsi = gsi_last_nondebug_bb (bb);
1137 !gsi_end_p (gsi);
1138 gsi_prev_nondebug (&gsi))
1140 last = gsi_stmt (gsi);
1141 if (!RESERVED_LOCATION_P (gimple_location (last)))
1142 break;
1145 /* Edge with goto locus might get wrong coverage info unless
1146 it is the only edge out of BB.
1147 Don't do that when the locuses match, so
1148 if (blah) goto something;
1149 is not computed twice. */
1150 if (last
1151 && gimple_has_location (last)
1152 && !RESERVED_LOCATION_P (e->goto_locus)
1153 && !single_succ_p (bb)
1154 && (LOCATION_FILE (e->goto_locus)
1155 != LOCATION_FILE (gimple_location (last))
1156 || (LOCATION_LINE (e->goto_locus)
1157 != LOCATION_LINE (gimple_location (last)))))
1159 basic_block new_bb = split_edge (e);
1160 edge ne = single_succ_edge (new_bb);
1161 ne->goto_locus = e->goto_locus;
1163 if ((e->flags & (EDGE_ABNORMAL | EDGE_ABNORMAL_CALL))
1164 && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun))
1165 need_exit_edge = 1;
1166 if (e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
1167 have_exit_edge = 1;
1169 FOR_EACH_EDGE (e, ei, bb->preds)
1171 if ((e->flags & (EDGE_ABNORMAL | EDGE_ABNORMAL_CALL))
1172 && e->src != ENTRY_BLOCK_PTR_FOR_FN (cfun))
1173 need_entry_edge = 1;
1174 if (e->src == ENTRY_BLOCK_PTR_FOR_FN (cfun))
1175 have_entry_edge = 1;
1178 if (need_exit_edge && !have_exit_edge)
1180 if (dump_file)
1181 fprintf (dump_file, "Adding fake exit edge to bb %i\n",
1182 bb->index);
1183 make_edge (bb, EXIT_BLOCK_PTR_FOR_FN (cfun), EDGE_FAKE);
1185 if (need_entry_edge && !have_entry_edge)
1187 if (dump_file)
1188 fprintf (dump_file, "Adding fake entry edge to bb %i\n",
1189 bb->index);
1190 make_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun), bb, EDGE_FAKE);
1191 /* Avoid bbs that have both fake entry edge and also some
1192 exit edge. One of those edges wouldn't be added to the
1193 spanning tree, but we can't instrument any of them. */
1194 if (have_exit_edge || need_exit_edge)
1196 gimple_stmt_iterator gsi;
1197 gimple *first;
1199 gsi = gsi_start_nondebug_after_labels_bb (bb);
1200 gcc_checking_assert (!gsi_end_p (gsi));
1201 first = gsi_stmt (gsi);
1202 /* Don't split the bbs containing __builtin_setjmp_receiver
1203 or ABNORMAL_DISPATCHER calls. These are very
1204 special and don't expect anything to be inserted before
1205 them. */
1206 if (is_gimple_call (first)
1207 && (gimple_call_builtin_p (first, BUILT_IN_SETJMP_RECEIVER)
1208 || (gimple_call_flags (first) & ECF_RETURNS_TWICE)
1209 || (gimple_call_internal_p (first)
1210 && (gimple_call_internal_fn (first)
1211 == IFN_ABNORMAL_DISPATCHER))))
1212 continue;
1214 if (dump_file)
1215 fprintf (dump_file, "Splitting bb %i after labels\n",
1216 bb->index);
1217 split_block_after_labels (bb);
1223 el = create_edge_list ();
1224 num_edges = NUM_EDGES (el);
1225 qsort (el->index_to_edge, num_edges, sizeof (edge), compare_freqs);
1226 alloc_aux_for_edges (sizeof (struct edge_profile_info));
1228 /* The basic blocks are expected to be numbered sequentially. */
1229 compact_blocks ();
1231 ignored_edges = 0;
1232 for (i = 0 ; i < num_edges ; i++)
1234 edge e = INDEX_EDGE (el, i);
1236 /* Mark edges we've replaced by fake edges above as ignored. */
1237 if ((e->flags & (EDGE_ABNORMAL | EDGE_ABNORMAL_CALL))
1238 && e->src != ENTRY_BLOCK_PTR_FOR_FN (cfun)
1239 && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun))
1241 EDGE_INFO (e)->ignore = 1;
1242 ignored_edges++;
1246 /* Create spanning tree from basic block graph, mark each edge that is
1247 on the spanning tree. We insert as many abnormal and critical edges
1248 as possible to minimize number of edge splits necessary. */
1250 if (!thunk)
1251 find_spanning_tree (el);
1252 else
1254 edge e;
1255 edge_iterator ei;
1256 /* Keep only edge from entry block to be instrumented. */
1257 FOR_EACH_BB_FN (bb, cfun)
1258 FOR_EACH_EDGE (e, ei, bb->succs)
1259 EDGE_INFO (e)->ignore = true;
1263 /* Fake edges that are not on the tree will not be instrumented, so
1264 mark them ignored. */
1265 for (num_instrumented = i = 0; i < num_edges; i++)
1267 edge e = INDEX_EDGE (el, i);
1268 struct edge_profile_info *inf = EDGE_INFO (e);
1270 if (inf->ignore || inf->on_tree)
1271 /*NOP*/;
1272 else if (e->flags & EDGE_FAKE)
1274 inf->ignore = 1;
1275 ignored_edges++;
1277 else
1278 num_instrumented++;
1281 total_num_blocks += n_basic_blocks_for_fn (cfun);
1282 if (dump_file)
1283 fprintf (dump_file, "%d basic blocks\n", n_basic_blocks_for_fn (cfun));
1285 total_num_edges += num_edges;
1286 if (dump_file)
1287 fprintf (dump_file, "%d edges\n", num_edges);
1289 total_num_edges_ignored += ignored_edges;
1290 if (dump_file)
1291 fprintf (dump_file, "%d ignored edges\n", ignored_edges);
1293 total_num_edges_instrumented += num_instrumented;
1294 if (dump_file)
1295 fprintf (dump_file, "%d instrumentation edges\n", num_instrumented);
1297 /* Compute two different checksums. Note that we want to compute
1298 the checksum in only once place, since it depends on the shape
1299 of the control flow which can change during
1300 various transformations. */
1301 if (thunk)
1303 /* At stream in time we do not have CFG, so we cannot do checksums. */
1304 cfg_checksum = 0;
1305 lineno_checksum = 0;
1307 else
1309 cfg_checksum = coverage_compute_cfg_checksum (cfun);
1310 lineno_checksum = coverage_compute_lineno_checksum ();
1313 /* Write the data from which gcov can reconstruct the basic block
1314 graph and function line numbers (the gcno file). */
1315 if (coverage_begin_function (lineno_checksum, cfg_checksum))
1317 gcov_position_t offset;
1319 /* Basic block flags */
1320 offset = gcov_write_tag (GCOV_TAG_BLOCKS);
1321 gcov_write_unsigned (n_basic_blocks_for_fn (cfun));
1322 gcov_write_length (offset);
1324 /* Arcs */
1325 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR_FOR_FN (cfun),
1326 EXIT_BLOCK_PTR_FOR_FN (cfun), next_bb)
1328 edge e;
1329 edge_iterator ei;
1331 offset = gcov_write_tag (GCOV_TAG_ARCS);
1332 gcov_write_unsigned (bb->index);
1334 FOR_EACH_EDGE (e, ei, bb->succs)
1336 struct edge_profile_info *i = EDGE_INFO (e);
1337 if (!i->ignore)
1339 unsigned flag_bits = 0;
1341 if (i->on_tree)
1342 flag_bits |= GCOV_ARC_ON_TREE;
1343 if (e->flags & EDGE_FAKE)
1344 flag_bits |= GCOV_ARC_FAKE;
1345 if (e->flags & EDGE_FALLTHRU)
1346 flag_bits |= GCOV_ARC_FALLTHROUGH;
1347 /* On trees we don't have fallthru flags, but we can
1348 recompute them from CFG shape. */
1349 if (e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE)
1350 && e->src->next_bb == e->dest)
1351 flag_bits |= GCOV_ARC_FALLTHROUGH;
1353 gcov_write_unsigned (e->dest->index);
1354 gcov_write_unsigned (flag_bits);
1358 gcov_write_length (offset);
1361 /* Line numbers. */
1362 /* Initialize the output. */
1363 output_location (&streamed_locations, NULL, 0, NULL, NULL);
1365 hash_set<int_hash <location_t, 0, 2> > seen_locations;
1367 FOR_EACH_BB_FN (bb, cfun)
1369 gimple_stmt_iterator gsi;
1370 gcov_position_t offset = 0;
1372 if (bb == ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb)
1374 location_t loc = DECL_SOURCE_LOCATION (current_function_decl);
1375 seen_locations.add (loc);
1376 expanded_location curr_location = expand_location (loc);
1377 output_location (&streamed_locations, curr_location.file,
1378 MAX (1, curr_location.line), &offset, bb);
1381 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1383 gimple *stmt = gsi_stmt (gsi);
1384 location_t loc = gimple_location (stmt);
1385 if (!RESERVED_LOCATION_P (loc))
1387 seen_locations.add (loc);
1388 output_location (&streamed_locations, gimple_filename (stmt),
1389 MAX (1, gimple_lineno (stmt)), &offset, bb);
1393 /* Notice GOTO expressions eliminated while constructing the CFG.
1394 It's hard to distinguish such expression, but goto_locus should
1395 not be any of already seen location. */
1396 location_t loc;
1397 if (single_succ_p (bb)
1398 && (loc = single_succ_edge (bb)->goto_locus)
1399 && !RESERVED_LOCATION_P (loc)
1400 && !seen_locations.contains (loc))
1402 expanded_location curr_location = expand_location (loc);
1403 output_location (&streamed_locations, curr_location.file,
1404 MAX (1, curr_location.line), &offset, bb);
1407 if (offset)
1409 /* A file of NULL indicates the end of run. */
1410 gcov_write_unsigned (0);
1411 gcov_write_string (NULL);
1412 gcov_write_length (offset);
1417 if (flag_profile_values)
1418 gimple_find_values_to_profile (&values);
1420 if (flag_branch_probabilities)
1422 compute_branch_probabilities (cfg_checksum, lineno_checksum);
1423 if (flag_profile_values)
1424 compute_value_histograms (values, cfg_checksum, lineno_checksum);
1427 remove_fake_edges ();
1429 /* For each edge not on the spanning tree, add counting code. */
1430 if (profile_arc_flag
1431 && coverage_counter_alloc (GCOV_COUNTER_ARCS, num_instrumented))
1433 unsigned n_instrumented;
1435 gimple_init_gcov_profiler ();
1437 n_instrumented = instrument_edges (el);
1439 gcc_assert (n_instrumented == num_instrumented);
1441 if (flag_profile_values)
1442 instrument_values (values);
1444 /* Commit changes done by instrumentation. */
1445 gsi_commit_edge_inserts ();
1448 free_aux_for_edges ();
1450 values.release ();
1451 free_edge_list (el);
1452 coverage_end_function (lineno_checksum, cfg_checksum);
1453 if (flag_branch_probabilities
1454 && (profile_status_for_fn (cfun) == PROFILE_READ))
1456 class loop *loop;
1457 if (dump_file && (dump_flags & TDF_DETAILS))
1458 report_predictor_hitrates ();
1460 /* At this moment we have precise loop iteration count estimates.
1461 Record them to loop structure before the profile gets out of date. */
1462 FOR_EACH_LOOP (loop, 0)
1463 if (loop->header->count > 0 && loop->header->count.reliable_p ())
1465 gcov_type nit = expected_loop_iterations_unbounded (loop);
1466 widest_int bound = gcov_type_to_wide_int (nit);
1467 loop->any_estimate = false;
1468 record_niter_bound (loop, bound, true, false);
1470 compute_function_frequency ();
1474 /* Union find algorithm implementation for the basic blocks using
1475 aux fields. */
1477 static basic_block
1478 find_group (basic_block bb)
1480 basic_block group = bb, bb1;
1482 while ((basic_block) group->aux != group)
1483 group = (basic_block) group->aux;
1485 /* Compress path. */
1486 while ((basic_block) bb->aux != group)
1488 bb1 = (basic_block) bb->aux;
1489 bb->aux = (void *) group;
1490 bb = bb1;
1492 return group;
1495 static void
1496 union_groups (basic_block bb1, basic_block bb2)
1498 basic_block bb1g = find_group (bb1);
1499 basic_block bb2g = find_group (bb2);
1501 /* ??? I don't have a place for the rank field. OK. Lets go w/o it,
1502 this code is unlikely going to be performance problem anyway. */
1503 gcc_assert (bb1g != bb2g);
1505 bb1g->aux = bb2g;
1508 /* This function searches all of the edges in the program flow graph, and puts
1509 as many bad edges as possible onto the spanning tree. Bad edges include
1510 abnormals edges, which can't be instrumented at the moment. Since it is
1511 possible for fake edges to form a cycle, we will have to develop some
1512 better way in the future. Also put critical edges to the tree, since they
1513 are more expensive to instrument. */
1515 static void
1516 find_spanning_tree (struct edge_list *el)
1518 int i;
1519 int num_edges = NUM_EDGES (el);
1520 basic_block bb;
1522 /* We use aux field for standard union-find algorithm. */
1523 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR_FOR_FN (cfun), NULL, next_bb)
1524 bb->aux = bb;
1526 /* Add fake edge exit to entry we can't instrument. */
1527 union_groups (EXIT_BLOCK_PTR_FOR_FN (cfun), ENTRY_BLOCK_PTR_FOR_FN (cfun));
1529 /* First add all abnormal edges to the tree unless they form a cycle. Also
1530 add all edges to the exit block to avoid inserting profiling code behind
1531 setting return value from function. */
1532 for (i = 0; i < num_edges; i++)
1534 edge e = INDEX_EDGE (el, i);
1535 if (((e->flags & (EDGE_ABNORMAL | EDGE_ABNORMAL_CALL | EDGE_FAKE))
1536 || e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
1537 && !EDGE_INFO (e)->ignore
1538 && (find_group (e->src) != find_group (e->dest)))
1540 if (dump_file)
1541 fprintf (dump_file, "Abnormal edge %d to %d put to tree\n",
1542 e->src->index, e->dest->index);
1543 EDGE_INFO (e)->on_tree = 1;
1544 union_groups (e->src, e->dest);
1548 /* And now the rest. Edge list is sorted according to frequencies and
1549 thus we will produce minimal spanning tree. */
1550 for (i = 0; i < num_edges; i++)
1552 edge e = INDEX_EDGE (el, i);
1553 if (!EDGE_INFO (e)->ignore
1554 && find_group (e->src) != find_group (e->dest))
1556 if (dump_file)
1557 fprintf (dump_file, "Normal edge %d to %d put to tree\n",
1558 e->src->index, e->dest->index);
1559 EDGE_INFO (e)->on_tree = 1;
1560 union_groups (e->src, e->dest);
1564 clear_aux_for_blocks ();
1567 /* Perform file-level initialization for branch-prob processing. */
1569 void
1570 init_branch_prob (void)
1572 int i;
1574 total_num_blocks = 0;
1575 total_num_edges = 0;
1576 total_num_edges_ignored = 0;
1577 total_num_edges_instrumented = 0;
1578 total_num_blocks_created = 0;
1579 total_num_passes = 0;
1580 total_num_times_called = 0;
1581 total_num_branches = 0;
1582 for (i = 0; i < 20; i++)
1583 total_hist_br_prob[i] = 0;
1586 /* Performs file-level cleanup after branch-prob processing
1587 is completed. */
1589 void
1590 end_branch_prob (void)
1592 if (dump_file)
1594 fprintf (dump_file, "\n");
1595 fprintf (dump_file, "Total number of blocks: %d\n",
1596 total_num_blocks);
1597 fprintf (dump_file, "Total number of edges: %d\n", total_num_edges);
1598 fprintf (dump_file, "Total number of ignored edges: %d\n",
1599 total_num_edges_ignored);
1600 fprintf (dump_file, "Total number of instrumented edges: %d\n",
1601 total_num_edges_instrumented);
1602 fprintf (dump_file, "Total number of blocks created: %d\n",
1603 total_num_blocks_created);
1604 fprintf (dump_file, "Total number of graph solution passes: %d\n",
1605 total_num_passes);
1606 if (total_num_times_called != 0)
1607 fprintf (dump_file, "Average number of graph solution passes: %d\n",
1608 (total_num_passes + (total_num_times_called >> 1))
1609 / total_num_times_called);
1610 fprintf (dump_file, "Total number of branches: %d\n",
1611 total_num_branches);
1612 if (total_num_branches)
1614 int i;
1616 for (i = 0; i < 10; i++)
1617 fprintf (dump_file, "%d%% branches in range %d-%d%%\n",
1618 (total_hist_br_prob[i] + total_hist_br_prob[19-i]) * 100
1619 / total_num_branches, 5*i, 5*i+5);