1 /* Calculate branch probabilities, and basic block execution counts.
2 Copyright (C) 1990-2017 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
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
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. */
52 #include "coretypes.h"
60 #include "diagnostic-core.h"
62 #include "value-prof.h"
63 #include "gimple-iterator.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. */
82 #define BB_INFO(b) ((struct bb_profile_info *) (b)->aux)
85 /* Counter summary from the last set of coverage counts read. */
87 const struct gcov_ctr_summary
*profile_info
;
89 /* Counter working set information computed from the current counter
90 summary. Not initialized unless profile_info summary is non-NULL. */
91 static gcov_working_set_t gcov_working_sets
[NUM_GCOV_WORKING_SETS
];
93 /* Collect statistics on the performance of this pass for the entire source
96 static int total_num_blocks
;
97 static int total_num_edges
;
98 static int total_num_edges_ignored
;
99 static int total_num_edges_instrumented
;
100 static int total_num_blocks_created
;
101 static int total_num_passes
;
102 static int total_num_times_called
;
103 static int total_hist_br_prob
[20];
104 static int total_num_branches
;
106 /* Helper function to update gcov_working_sets. */
108 void add_working_set (gcov_working_set_t
*set
) {
110 for (; i
< NUM_GCOV_WORKING_SETS
; i
++)
111 gcov_working_sets
[i
] = set
[i
];
114 /* Forward declarations. */
115 static void find_spanning_tree (struct edge_list
*);
117 /* Add edge instrumentation code to the entire insn chain.
119 F is the first insn of the chain.
120 NUM_BLOCKS is the number of basic blocks found in F. */
123 instrument_edges (struct edge_list
*el
)
125 unsigned num_instr_edges
= 0;
126 int num_edges
= NUM_EDGES (el
);
129 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
134 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
136 struct edge_profile_info
*inf
= EDGE_INFO (e
);
138 if (!inf
->ignore
&& !inf
->on_tree
)
140 gcc_assert (!(e
->flags
& EDGE_ABNORMAL
));
142 fprintf (dump_file
, "Edge %d to %d instrumented%s\n",
143 e
->src
->index
, e
->dest
->index
,
144 EDGE_CRITICAL_P (e
) ? " (and split)" : "");
145 gimple_gen_edge_profiler (num_instr_edges
++, e
);
150 total_num_blocks_created
+= num_edges
;
152 fprintf (dump_file
, "%d edges instrumented\n", num_instr_edges
);
153 return num_instr_edges
;
156 /* Add code to measure histograms for values in list VALUES. */
158 instrument_values (histogram_values values
)
162 /* Emit code to generate the histograms before the insns. */
164 for (i
= 0; i
< values
.length (); i
++)
166 histogram_value hist
= values
[i
];
167 unsigned t
= COUNTER_FOR_HIST_TYPE (hist
->type
);
169 if (!coverage_counter_alloc (t
, hist
->n_counters
))
174 case HIST_TYPE_INTERVAL
:
175 gimple_gen_interval_profiler (hist
, t
, 0);
179 gimple_gen_pow2_profiler (hist
, t
, 0);
182 case HIST_TYPE_SINGLE_VALUE
:
183 gimple_gen_one_value_profiler (hist
, t
, 0);
186 case HIST_TYPE_INDIR_CALL
:
187 case HIST_TYPE_INDIR_CALL_TOPN
:
188 gimple_gen_ic_profiler (hist
, t
, 0);
191 case HIST_TYPE_AVERAGE
:
192 gimple_gen_average_profiler (hist
, t
, 0);
196 gimple_gen_ior_profiler (hist
, t
, 0);
199 case HIST_TYPE_TIME_PROFILE
:
200 gimple_gen_time_profiler (t
, 0);
210 /* Fill the working set information into the profile_info structure. */
213 get_working_sets (void)
215 unsigned ws_ix
, pctinc
, pct
;
216 gcov_working_set_t
*ws_info
;
221 compute_working_sets (profile_info
, gcov_working_sets
);
225 fprintf (dump_file
, "Counter working sets:\n");
226 /* Multiply the percentage by 100 to avoid float. */
227 pctinc
= 100 * 100 / NUM_GCOV_WORKING_SETS
;
228 for (ws_ix
= 0, pct
= pctinc
; ws_ix
< NUM_GCOV_WORKING_SETS
;
229 ws_ix
++, pct
+= pctinc
)
231 if (ws_ix
== NUM_GCOV_WORKING_SETS
- 1)
233 ws_info
= &gcov_working_sets
[ws_ix
];
234 /* Print out the percentage using int arithmatic to avoid float. */
235 fprintf (dump_file
, "\t\t%u.%02u%%: num counts=%u, min counter="
237 pct
/ 100, pct
- (pct
/ 100 * 100),
238 ws_info
->num_counters
,
239 (int64_t)ws_info
->min_counter
);
244 /* Given a the desired percentage of the full profile (sum_all from the
245 summary), multiplied by 10 to avoid float in PCT_TIMES_10, returns
246 the corresponding working set information. If an exact match for
247 the percentage isn't found, the closest value is used. */
250 find_working_set (unsigned pct_times_10
)
255 gcc_assert (pct_times_10
<= 1000);
256 if (pct_times_10
>= 999)
257 return &gcov_working_sets
[NUM_GCOV_WORKING_SETS
- 1];
258 i
= pct_times_10
* NUM_GCOV_WORKING_SETS
/ 1000;
260 return &gcov_working_sets
[0];
261 return &gcov_working_sets
[i
- 1];
264 /* Computes hybrid profile for all matching entries in da_file.
266 CFG_CHECKSUM is the precomputed checksum for the CFG. */
269 get_exec_counts (unsigned cfg_checksum
, unsigned lineno_checksum
)
271 unsigned num_edges
= 0;
275 /* Count the edges to be (possibly) instrumented. */
276 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
281 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
282 if (!EDGE_INFO (e
)->ignore
&& !EDGE_INFO (e
)->on_tree
)
286 counts
= get_coverage_counts (GCOV_COUNTER_ARCS
, num_edges
, cfg_checksum
,
287 lineno_checksum
, &profile_info
);
293 if (dump_file
&& profile_info
)
294 fprintf (dump_file
, "Merged %u profiles with maximal count %u.\n",
295 profile_info
->runs
, (unsigned) profile_info
->sum_max
);
302 is_edge_inconsistent (vec
<edge
, va_gc
> *edges
)
306 FOR_EACH_EDGE (e
, ei
, edges
)
308 if (!EDGE_INFO (e
)->ignore
)
310 if (edge_gcov_count (e
) < 0
311 && (!(e
->flags
& EDGE_FAKE
)
312 || !block_ends_with_call_p (e
->src
)))
317 "Edge %i->%i is inconsistent, count%" PRId64
,
318 e
->src
->index
, e
->dest
->index
, edge_gcov_count (e
));
319 dump_bb (dump_file
, e
->src
, 0, TDF_DETAILS
);
320 dump_bb (dump_file
, e
->dest
, 0, TDF_DETAILS
);
330 correct_negative_edge_counts (void)
336 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
338 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
340 if (edge_gcov_count (e
) < 0)
341 edge_gcov_count (e
) = 0;
346 /* Check consistency.
347 Return true if inconsistency is found. */
349 is_inconsistent (void)
352 bool inconsistent
= false;
353 FOR_EACH_BB_FN (bb
, cfun
)
355 inconsistent
|= is_edge_inconsistent (bb
->preds
);
356 if (!dump_file
&& inconsistent
)
358 inconsistent
|= is_edge_inconsistent (bb
->succs
);
359 if (!dump_file
&& inconsistent
)
361 if (bb_gcov_count (bb
) < 0)
365 fprintf (dump_file
, "BB %i count is negative "
369 dump_bb (dump_file
, bb
, 0, TDF_DETAILS
);
373 if (bb_gcov_count (bb
) != sum_edge_counts (bb
->preds
))
377 fprintf (dump_file
, "BB %i count does not match sum of incoming edges "
378 "%" PRId64
" should be %" PRId64
,
381 sum_edge_counts (bb
->preds
));
382 dump_bb (dump_file
, bb
, 0, TDF_DETAILS
);
386 if (bb_gcov_count (bb
) != sum_edge_counts (bb
->succs
) &&
387 ! (find_edge (bb
, EXIT_BLOCK_PTR_FOR_FN (cfun
)) != NULL
388 && block_ends_with_call_p (bb
)))
392 fprintf (dump_file
, "BB %i count does not match sum of outgoing edges "
393 "%" PRId64
" should be %" PRId64
,
396 sum_edge_counts (bb
->succs
));
397 dump_bb (dump_file
, bb
, 0, TDF_DETAILS
);
401 if (!dump_file
&& inconsistent
)
408 /* Set each basic block count to the sum of its outgoing edge counts */
413 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
415 bb_gcov_count (bb
) = sum_edge_counts (bb
->succs
);
416 gcc_assert (bb_gcov_count (bb
) >= 0);
420 /* Reads profile data and returns total number of edge counts read */
422 read_profile_edge_counts (gcov_type
*exec_counts
)
426 int exec_counts_pos
= 0;
427 /* For each edge not on the spanning tree, set its execution count from
429 /* The first count in the .da file is the number of times that the function
430 was entered. This is the exec_count for block zero. */
432 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
437 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
438 if (!EDGE_INFO (e
)->ignore
&& !EDGE_INFO (e
)->on_tree
)
443 edge_gcov_count (e
) = exec_counts
[exec_counts_pos
++];
444 if (edge_gcov_count (e
) > profile_info
->sum_max
)
446 if (flag_profile_correction
)
448 static bool informed
= 0;
449 if (dump_enabled_p () && !informed
)
450 dump_printf_loc (MSG_NOTE
, input_location
,
451 "corrupted profile info: edge count"
452 " exceeds maximal count\n");
456 error ("corrupted profile info: edge from %i to %i exceeds maximal count",
457 bb
->index
, e
->dest
->index
);
461 edge_gcov_count (e
) = 0;
463 EDGE_INFO (e
)->count_valid
= 1;
464 BB_INFO (bb
)->succ_count
--;
465 BB_INFO (e
->dest
)->pred_count
--;
468 fprintf (dump_file
, "\nRead edge from %i to %i, count:",
469 bb
->index
, e
->dest
->index
);
470 fprintf (dump_file
, "%" PRId64
,
471 (int64_t) edge_gcov_count (e
));
479 #define OVERLAP_BASE 10000
481 /* Compare the static estimated profile to the actual profile, and
482 return the "degree of overlap" measure between them.
484 Degree of overlap is a number between 0 and OVERLAP_BASE. It is
485 the sum of each basic block's minimum relative weights between
486 two profiles. And overlap of OVERLAP_BASE means two profiles are
490 compute_frequency_overlap (void)
492 gcov_type count_total
= 0, freq_total
= 0;
496 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
498 count_total
+= bb_gcov_count (bb
);
499 freq_total
+= bb
->frequency
;
502 if (count_total
== 0 || freq_total
== 0)
505 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
506 overlap
+= MIN (bb_gcov_count (bb
) * OVERLAP_BASE
/ count_total
,
507 bb
->frequency
* OVERLAP_BASE
/ freq_total
);
512 /* Compute the branch probabilities for the various branches.
513 Annotate them accordingly.
515 CFG_CHECKSUM is the precomputed checksum for the CFG. */
518 compute_branch_probabilities (unsigned cfg_checksum
, unsigned lineno_checksum
)
525 int hist_br_prob
[20];
527 gcov_type
*exec_counts
= get_exec_counts (cfg_checksum
, lineno_checksum
);
528 int inconsistent
= 0;
530 /* Very simple sanity checks so we catch bugs in our profiling code. */
534 bb_gcov_counts
.safe_grow_cleared (last_basic_block_for_fn (cfun
));
535 edge_gcov_counts
= new hash_map
<edge
,gcov_type
>;
537 if (profile_info
->sum_all
< profile_info
->sum_max
)
539 error ("corrupted profile info: sum_all is smaller than sum_max");
543 /* Attach extra info block to each bb. */
544 alloc_aux_for_blocks (sizeof (struct bb_profile_info
));
545 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
550 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
551 if (!EDGE_INFO (e
)->ignore
)
552 BB_INFO (bb
)->succ_count
++;
553 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
554 if (!EDGE_INFO (e
)->ignore
)
555 BB_INFO (bb
)->pred_count
++;
558 /* Avoid predicting entry on exit nodes. */
559 BB_INFO (EXIT_BLOCK_PTR_FOR_FN (cfun
))->succ_count
= 2;
560 BB_INFO (ENTRY_BLOCK_PTR_FOR_FN (cfun
))->pred_count
= 2;
562 num_edges
= read_profile_edge_counts (exec_counts
);
565 fprintf (dump_file
, "\n%d edge counts read\n", num_edges
);
567 /* For every block in the file,
568 - if every exit/entrance edge has a known count, then set the block count
569 - if the block count is known, and every exit/entrance edge but one has
570 a known execution count, then set the count of the remaining edge
572 As edge counts are set, decrement the succ/pred count, but don't delete
573 the edge, that way we can easily tell when all edges are known, or only
574 one edge is unknown. */
576 /* The order that the basic blocks are iterated through is important.
577 Since the code that finds spanning trees starts with block 0, low numbered
578 edges are put on the spanning tree in preference to high numbered edges.
579 Hence, most instrumented edges are at the end. Graph solving works much
580 faster if we propagate numbers from the end to the start.
582 This takes an average of slightly more than 3 passes. */
590 FOR_BB_BETWEEN (bb
, EXIT_BLOCK_PTR_FOR_FN (cfun
), NULL
, prev_bb
)
592 struct bb_profile_info
*bi
= BB_INFO (bb
);
593 if (! bi
->count_valid
)
595 if (bi
->succ_count
== 0)
601 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
602 total
+= edge_gcov_count (e
);
603 bb_gcov_count (bb
) = total
;
607 else if (bi
->pred_count
== 0)
613 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
614 total
+= edge_gcov_count (e
);
615 bb_gcov_count (bb
) = total
;
622 if (bi
->succ_count
== 1)
628 /* One of the counts will be invalid, but it is zero,
629 so adding it in also doesn't hurt. */
630 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
631 total
+= edge_gcov_count (e
);
633 /* Search for the invalid edge, and set its count. */
634 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
635 if (! EDGE_INFO (e
)->count_valid
&& ! EDGE_INFO (e
)->ignore
)
638 /* Calculate count for remaining edge by conservation. */
639 total
= bb_gcov_count (bb
) - total
;
642 EDGE_INFO (e
)->count_valid
= 1;
643 edge_gcov_count (e
) = total
;
646 BB_INFO (e
->dest
)->pred_count
--;
649 if (bi
->pred_count
== 1)
655 /* One of the counts will be invalid, but it is zero,
656 so adding it in also doesn't hurt. */
657 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
658 total
+= edge_gcov_count (e
);
660 /* Search for the invalid edge, and set its count. */
661 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
662 if (!EDGE_INFO (e
)->count_valid
&& !EDGE_INFO (e
)->ignore
)
665 /* Calculate count for remaining edge by conservation. */
666 total
= bb_gcov_count (bb
) - total
+ edge_gcov_count (e
);
669 EDGE_INFO (e
)->count_valid
= 1;
670 edge_gcov_count (e
) = total
;
673 BB_INFO (e
->src
)->succ_count
--;
681 int overlap
= compute_frequency_overlap ();
682 gimple_dump_cfg (dump_file
, dump_flags
);
683 fprintf (dump_file
, "Static profile overlap: %d.%d%%\n",
684 overlap
/ (OVERLAP_BASE
/ 100),
685 overlap
% (OVERLAP_BASE
/ 100));
688 total_num_passes
+= passes
;
690 fprintf (dump_file
, "Graph solving took %d passes.\n\n", passes
);
692 /* If the graph has been correctly solved, every block will have a
693 succ and pred count of zero. */
694 FOR_EACH_BB_FN (bb
, cfun
)
696 gcc_assert (!BB_INFO (bb
)->succ_count
&& !BB_INFO (bb
)->pred_count
);
699 /* Check for inconsistent basic block counts */
700 inconsistent
= is_inconsistent ();
704 if (flag_profile_correction
)
706 /* Inconsistency detected. Make it flow-consistent. */
707 static int informed
= 0;
708 if (dump_enabled_p () && informed
== 0)
711 dump_printf_loc (MSG_NOTE
, input_location
,
712 "correcting inconsistent profile data\n");
714 correct_negative_edge_counts ();
715 /* Set bb counts to the sum of the outgoing edge counts */
718 fprintf (dump_file
, "\nCalling mcf_smooth_cfg\n");
722 error ("corrupted profile info: profile data is not flow-consistent");
725 /* For every edge, calculate its branch probability and add a reg_note
726 to the branch insn to indicate this. */
728 for (i
= 0; i
< 20; i
++)
732 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
737 if (bb_gcov_count (bb
) < 0)
739 error ("corrupted profile info: number of iterations for basic block %d thought to be %i",
740 bb
->index
, (int)bb_gcov_count (bb
));
741 bb_gcov_count (bb
) = 0;
743 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
745 /* Function may return twice in the cased the called function is
746 setjmp or calls fork, but we can't represent this by extra
747 edge from the entry, since extra edge from the exit is
748 already present. We get negative frequency from the entry
750 if ((edge_gcov_count (e
) < 0
751 && e
->dest
== EXIT_BLOCK_PTR_FOR_FN (cfun
))
752 || (edge_gcov_count (e
) > bb_gcov_count (bb
)
753 && e
->dest
!= EXIT_BLOCK_PTR_FOR_FN (cfun
)))
755 if (block_ends_with_call_p (bb
))
756 edge_gcov_count (e
) = edge_gcov_count (e
) < 0
757 ? 0 : bb_gcov_count (bb
);
759 if (edge_gcov_count (e
) < 0
760 || edge_gcov_count (e
) > bb_gcov_count (bb
))
762 error ("corrupted profile info: number of executions for edge %d-%d thought to be %i",
763 e
->src
->index
, e
->dest
->index
,
764 (int)edge_gcov_count (e
));
765 edge_gcov_count (e
) = bb_gcov_count (bb
) / 2;
768 if (bb_gcov_count (bb
))
770 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
771 e
->probability
= profile_probability::probability_in_gcov_type
772 (edge_gcov_count (e
), bb_gcov_count (bb
));
773 if (bb
->index
>= NUM_FIXED_BLOCKS
774 && block_ends_with_condjump_p (bb
)
775 && EDGE_COUNT (bb
->succs
) >= 2)
781 /* Find the branch edge. It is possible that we do have fake
783 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
784 if (!(e
->flags
& (EDGE_FAKE
| EDGE_FALLTHRU
)))
787 prob
= e
->probability
.to_reg_br_prob_base ();
788 index
= prob
* 20 / REG_BR_PROB_BASE
;
792 hist_br_prob
[index
]++;
797 /* As a last resort, distribute the probabilities evenly.
798 Use simple heuristics that if there are normal edges,
799 give all abnormals frequency of 0, otherwise distribute the
800 frequency over abnormals (this is the case of noreturn
802 else if (profile_status_for_fn (cfun
) == PROFILE_ABSENT
)
806 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
807 if (!(e
->flags
& (EDGE_COMPLEX
| EDGE_FAKE
)))
811 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
812 if (!(e
->flags
& (EDGE_COMPLEX
| EDGE_FAKE
)))
814 = profile_probability::guessed_always ().apply_scale (1, total
);
816 e
->probability
= profile_probability::never ();
820 total
+= EDGE_COUNT (bb
->succs
);
821 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
823 = profile_probability::guessed_always ().apply_scale (1, total
);
825 if (bb
->index
>= NUM_FIXED_BLOCKS
826 && block_ends_with_condjump_p (bb
)
827 && EDGE_COUNT (bb
->succs
) >= 2)
832 FOR_ALL_BB_FN (bb
, cfun
)
837 bb
->count
= profile_count::from_gcov_type (bb_gcov_count (bb
));
838 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
839 e
->count
= profile_count::from_gcov_type (edge_gcov_count (e
));
841 bb_gcov_counts
.release ();
842 delete edge_gcov_counts
;
843 edge_gcov_counts
= NULL
;
849 fprintf (dump_file
, "%d branches\n", num_branches
);
851 for (i
= 0; i
< 10; i
++)
852 fprintf (dump_file
, "%d%% branches in range %d-%d%%\n",
853 (hist_br_prob
[i
] + hist_br_prob
[19-i
]) * 100 / num_branches
,
856 total_num_branches
+= num_branches
;
857 for (i
= 0; i
< 20; i
++)
858 total_hist_br_prob
[i
] += hist_br_prob
[i
];
860 fputc ('\n', dump_file
);
861 fputc ('\n', dump_file
);
864 free_aux_for_blocks ();
867 /* Load value histograms values whose description is stored in VALUES array
870 CFG_CHECKSUM is the precomputed checksum for the CFG. */
873 compute_value_histograms (histogram_values values
, unsigned cfg_checksum
,
874 unsigned lineno_checksum
)
876 unsigned i
, j
, t
, any
;
877 unsigned n_histogram_counters
[GCOV_N_VALUE_COUNTERS
];
878 gcov_type
*histogram_counts
[GCOV_N_VALUE_COUNTERS
];
879 gcov_type
*act_count
[GCOV_N_VALUE_COUNTERS
];
880 gcov_type
*aact_count
;
881 struct cgraph_node
*node
;
883 for (t
= 0; t
< GCOV_N_VALUE_COUNTERS
; t
++)
884 n_histogram_counters
[t
] = 0;
886 for (i
= 0; i
< values
.length (); i
++)
888 histogram_value hist
= values
[i
];
889 n_histogram_counters
[(int) hist
->type
] += hist
->n_counters
;
893 for (t
= 0; t
< GCOV_N_VALUE_COUNTERS
; t
++)
895 if (!n_histogram_counters
[t
])
897 histogram_counts
[t
] = NULL
;
901 histogram_counts
[t
] =
902 get_coverage_counts (COUNTER_FOR_HIST_TYPE (t
),
903 n_histogram_counters
[t
], cfg_checksum
,
904 lineno_checksum
, NULL
);
905 if (histogram_counts
[t
])
907 act_count
[t
] = histogram_counts
[t
];
912 for (i
= 0; i
< values
.length (); i
++)
914 histogram_value hist
= values
[i
];
915 gimple
*stmt
= hist
->hvalue
.stmt
;
917 t
= (int) hist
->type
;
919 aact_count
= act_count
[t
];
922 act_count
[t
] += hist
->n_counters
;
924 gimple_add_histogram_value (cfun
, stmt
, hist
);
925 hist
->hvalue
.counters
= XNEWVEC (gcov_type
, hist
->n_counters
);
926 for (j
= 0; j
< hist
->n_counters
; j
++)
928 hist
->hvalue
.counters
[j
] = aact_count
[j
];
930 hist
->hvalue
.counters
[j
] = 0;
932 /* Time profiler counter is not related to any statement,
933 so that we have to read the counter and set the value to
934 the corresponding call graph node. */
935 if (hist
->type
== HIST_TYPE_TIME_PROFILE
)
937 node
= cgraph_node::get (hist
->fun
->decl
);
938 node
->tp_first_run
= hist
->hvalue
.counters
[0];
941 fprintf (dump_file
, "Read tp_first_run: %d\n", node
->tp_first_run
);
945 for (t
= 0; t
< GCOV_N_VALUE_COUNTERS
; t
++)
946 free (histogram_counts
[t
]);
949 /* When passed NULL as file_name, initialize.
950 When passed something else, output the necessary commands to change
951 line to LINE and offset to FILE_NAME. */
953 output_location (char const *file_name
, int line
,
954 gcov_position_t
*offset
, basic_block bb
)
956 static char const *prev_file_name
;
957 static int prev_line
;
958 bool name_differs
, line_differs
;
962 prev_file_name
= NULL
;
967 name_differs
= !prev_file_name
|| filename_cmp (file_name
, prev_file_name
);
968 line_differs
= prev_line
!= line
;
972 *offset
= gcov_write_tag (GCOV_TAG_LINES
);
973 gcov_write_unsigned (bb
->index
);
974 name_differs
= line_differs
= true;
977 /* If this is a new source file, then output the
978 file's name to the .bb file. */
981 prev_file_name
= file_name
;
982 gcov_write_unsigned (0);
983 gcov_write_filename (prev_file_name
);
987 gcov_write_unsigned (line
);
992 /* Helper for qsort so edges get sorted from highest frequency to smallest.
993 This controls the weight for minimal spanning tree algorithm */
995 compare_freqs (const void *p1
, const void *p2
)
997 const_edge e1
= *(const const_edge
*)p1
;
998 const_edge e2
= *(const const_edge
*)p2
;
1000 /* Critical edges needs to be split which introduce extra control flow.
1001 Make them more heavy. */
1002 int m1
= EDGE_CRITICAL_P (e1
) ? 2 : 1;
1003 int m2
= EDGE_CRITICAL_P (e2
) ? 2 : 1;
1005 if (EDGE_FREQUENCY (e1
) * m1
+ m1
!= EDGE_FREQUENCY (e2
) * m2
+ m2
)
1006 return EDGE_FREQUENCY (e2
) * m2
+ m2
- EDGE_FREQUENCY (e1
) * m1
- m1
;
1007 /* Stabilize sort. */
1008 if (e1
->src
->index
!= e2
->src
->index
)
1009 return e2
->src
->index
- e1
->src
->index
;
1010 return e2
->dest
->index
- e1
->dest
->index
;
1013 /* Instrument and/or analyze program behavior based on program the CFG.
1015 This function creates a representation of the control flow graph (of
1016 the function being compiled) that is suitable for the instrumentation
1017 of edges and/or converting measured edge counts to counts on the
1020 When FLAG_PROFILE_ARCS is nonzero, this function instruments the edges in
1021 the flow graph that are needed to reconstruct the dynamic behavior of the
1022 flow graph. This data is written to the gcno file for gcov.
1024 When FLAG_BRANCH_PROBABILITIES is nonzero, this function reads auxiliary
1025 information from the gcda file containing edge count information from
1026 previous executions of the function being compiled. In this case, the
1027 control flow graph is annotated with actual execution counts by
1028 compute_branch_probabilities().
1030 Main entry point of this file. */
1037 unsigned num_edges
, ignored_edges
;
1038 unsigned num_instrumented
;
1039 struct edge_list
*el
;
1040 histogram_values values
= histogram_values ();
1041 unsigned cfg_checksum
, lineno_checksum
;
1043 total_num_times_called
++;
1045 flow_call_edges_add (NULL
);
1046 add_noreturn_fake_exit_edges ();
1048 /* We can't handle cyclic regions constructed using abnormal edges.
1049 To avoid these we replace every source of abnormal edge by a fake
1050 edge from entry node and every destination by fake edge to exit.
1051 This keeps graph acyclic and our calculation exact for all normal
1052 edges except for exit and entrance ones.
1054 We also add fake exit edges for each call and asm statement in the
1055 basic, since it may not return. */
1057 FOR_EACH_BB_FN (bb
, cfun
)
1059 int need_exit_edge
= 0, need_entry_edge
= 0;
1060 int have_exit_edge
= 0, have_entry_edge
= 0;
1064 /* Functions returning multiple times are not handled by extra edges.
1065 Instead we simply allow negative counts on edges from exit to the
1066 block past call and corresponding probabilities. We can't go
1067 with the extra edges because that would result in flowgraph that
1068 needs to have fake edges outside the spanning tree. */
1070 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1072 gimple_stmt_iterator gsi
;
1073 gimple
*last
= NULL
;
1075 /* It may happen that there are compiler generated statements
1076 without a locus at all. Go through the basic block from the
1077 last to the first statement looking for a locus. */
1078 for (gsi
= gsi_last_nondebug_bb (bb
);
1080 gsi_prev_nondebug (&gsi
))
1082 last
= gsi_stmt (gsi
);
1083 if (!RESERVED_LOCATION_P (gimple_location (last
)))
1087 /* Edge with goto locus might get wrong coverage info unless
1088 it is the only edge out of BB.
1089 Don't do that when the locuses match, so
1090 if (blah) goto something;
1091 is not computed twice. */
1093 && gimple_has_location (last
)
1094 && !RESERVED_LOCATION_P (e
->goto_locus
)
1095 && !single_succ_p (bb
)
1096 && (LOCATION_FILE (e
->goto_locus
)
1097 != LOCATION_FILE (gimple_location (last
))
1098 || (LOCATION_LINE (e
->goto_locus
)
1099 != LOCATION_LINE (gimple_location (last
)))))
1101 basic_block new_bb
= split_edge (e
);
1102 edge ne
= single_succ_edge (new_bb
);
1103 ne
->goto_locus
= e
->goto_locus
;
1105 if ((e
->flags
& (EDGE_ABNORMAL
| EDGE_ABNORMAL_CALL
))
1106 && e
->dest
!= EXIT_BLOCK_PTR_FOR_FN (cfun
))
1108 if (e
->dest
== EXIT_BLOCK_PTR_FOR_FN (cfun
))
1111 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1113 if ((e
->flags
& (EDGE_ABNORMAL
| EDGE_ABNORMAL_CALL
))
1114 && e
->src
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
))
1115 need_entry_edge
= 1;
1116 if (e
->src
== ENTRY_BLOCK_PTR_FOR_FN (cfun
))
1117 have_entry_edge
= 1;
1120 if (need_exit_edge
&& !have_exit_edge
)
1123 fprintf (dump_file
, "Adding fake exit edge to bb %i\n",
1125 make_edge (bb
, EXIT_BLOCK_PTR_FOR_FN (cfun
), EDGE_FAKE
);
1127 if (need_entry_edge
&& !have_entry_edge
)
1130 fprintf (dump_file
, "Adding fake entry edge to bb %i\n",
1132 make_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun
), bb
, EDGE_FAKE
);
1133 /* Avoid bbs that have both fake entry edge and also some
1134 exit edge. One of those edges wouldn't be added to the
1135 spanning tree, but we can't instrument any of them. */
1136 if (have_exit_edge
|| need_exit_edge
)
1138 gimple_stmt_iterator gsi
;
1141 gsi
= gsi_start_nondebug_after_labels_bb (bb
);
1142 gcc_checking_assert (!gsi_end_p (gsi
));
1143 first
= gsi_stmt (gsi
);
1144 /* Don't split the bbs containing __builtin_setjmp_receiver
1145 or ABNORMAL_DISPATCHER calls. These are very
1146 special and don't expect anything to be inserted before
1148 if (is_gimple_call (first
)
1149 && (gimple_call_builtin_p (first
, BUILT_IN_SETJMP_RECEIVER
)
1150 || (gimple_call_flags (first
) & ECF_RETURNS_TWICE
)
1151 || (gimple_call_internal_p (first
)
1152 && (gimple_call_internal_fn (first
)
1153 == IFN_ABNORMAL_DISPATCHER
))))
1157 fprintf (dump_file
, "Splitting bb %i after labels\n",
1159 split_block_after_labels (bb
);
1164 el
= create_edge_list ();
1165 num_edges
= NUM_EDGES (el
);
1166 qsort (el
->index_to_edge
, num_edges
, sizeof (edge
), compare_freqs
);
1167 alloc_aux_for_edges (sizeof (struct edge_profile_info
));
1169 /* The basic blocks are expected to be numbered sequentially. */
1173 for (i
= 0 ; i
< num_edges
; i
++)
1175 edge e
= INDEX_EDGE (el
, i
);
1177 /* Mark edges we've replaced by fake edges above as ignored. */
1178 if ((e
->flags
& (EDGE_ABNORMAL
| EDGE_ABNORMAL_CALL
))
1179 && e
->src
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
)
1180 && e
->dest
!= EXIT_BLOCK_PTR_FOR_FN (cfun
))
1182 EDGE_INFO (e
)->ignore
= 1;
1187 /* Create spanning tree from basic block graph, mark each edge that is
1188 on the spanning tree. We insert as many abnormal and critical edges
1189 as possible to minimize number of edge splits necessary. */
1191 find_spanning_tree (el
);
1193 /* Fake edges that are not on the tree will not be instrumented, so
1194 mark them ignored. */
1195 for (num_instrumented
= i
= 0; i
< num_edges
; i
++)
1197 edge e
= INDEX_EDGE (el
, i
);
1198 struct edge_profile_info
*inf
= EDGE_INFO (e
);
1200 if (inf
->ignore
|| inf
->on_tree
)
1202 else if (e
->flags
& EDGE_FAKE
)
1211 total_num_blocks
+= n_basic_blocks_for_fn (cfun
);
1213 fprintf (dump_file
, "%d basic blocks\n", n_basic_blocks_for_fn (cfun
));
1215 total_num_edges
+= num_edges
;
1217 fprintf (dump_file
, "%d edges\n", num_edges
);
1219 total_num_edges_ignored
+= ignored_edges
;
1221 fprintf (dump_file
, "%d ignored edges\n", ignored_edges
);
1223 total_num_edges_instrumented
+= num_instrumented
;
1225 fprintf (dump_file
, "%d instrumentation edges\n", num_instrumented
);
1227 /* Compute two different checksums. Note that we want to compute
1228 the checksum in only once place, since it depends on the shape
1229 of the control flow which can change during
1230 various transformations. */
1231 cfg_checksum
= coverage_compute_cfg_checksum (cfun
);
1232 lineno_checksum
= coverage_compute_lineno_checksum ();
1234 /* Write the data from which gcov can reconstruct the basic block
1235 graph and function line numbers (the gcno file). */
1236 if (coverage_begin_function (lineno_checksum
, cfg_checksum
))
1238 gcov_position_t offset
;
1240 /* Basic block flags */
1241 offset
= gcov_write_tag (GCOV_TAG_BLOCKS
);
1242 gcov_write_unsigned (n_basic_blocks_for_fn (cfun
));
1243 gcov_write_length (offset
);
1246 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
),
1247 EXIT_BLOCK_PTR_FOR_FN (cfun
), next_bb
)
1252 offset
= gcov_write_tag (GCOV_TAG_ARCS
);
1253 gcov_write_unsigned (bb
->index
);
1255 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1257 struct edge_profile_info
*i
= EDGE_INFO (e
);
1260 unsigned flag_bits
= 0;
1263 flag_bits
|= GCOV_ARC_ON_TREE
;
1264 if (e
->flags
& EDGE_FAKE
)
1265 flag_bits
|= GCOV_ARC_FAKE
;
1266 if (e
->flags
& EDGE_FALLTHRU
)
1267 flag_bits
|= GCOV_ARC_FALLTHROUGH
;
1268 /* On trees we don't have fallthru flags, but we can
1269 recompute them from CFG shape. */
1270 if (e
->flags
& (EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
)
1271 && e
->src
->next_bb
== e
->dest
)
1272 flag_bits
|= GCOV_ARC_FALLTHROUGH
;
1274 gcov_write_unsigned (e
->dest
->index
);
1275 gcov_write_unsigned (flag_bits
);
1279 gcov_write_length (offset
);
1283 /* Initialize the output. */
1284 output_location (NULL
, 0, NULL
, NULL
);
1286 FOR_EACH_BB_FN (bb
, cfun
)
1288 gimple_stmt_iterator gsi
;
1289 gcov_position_t offset
= 0;
1291 if (bb
== ENTRY_BLOCK_PTR_FOR_FN (cfun
)->next_bb
)
1293 expanded_location curr_location
=
1294 expand_location (DECL_SOURCE_LOCATION (current_function_decl
));
1295 output_location (curr_location
.file
, curr_location
.line
,
1299 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1301 gimple
*stmt
= gsi_stmt (gsi
);
1302 if (!RESERVED_LOCATION_P (gimple_location (stmt
)))
1303 output_location (gimple_filename (stmt
), gimple_lineno (stmt
),
1307 /* Notice GOTO expressions eliminated while constructing the CFG. */
1308 if (single_succ_p (bb
)
1309 && !RESERVED_LOCATION_P (single_succ_edge (bb
)->goto_locus
))
1311 expanded_location curr_location
1312 = expand_location (single_succ_edge (bb
)->goto_locus
);
1313 output_location (curr_location
.file
, curr_location
.line
,
1319 /* A file of NULL indicates the end of run. */
1320 gcov_write_unsigned (0);
1321 gcov_write_string (NULL
);
1322 gcov_write_length (offset
);
1327 if (flag_profile_values
)
1328 gimple_find_values_to_profile (&values
);
1330 if (flag_branch_probabilities
)
1332 compute_branch_probabilities (cfg_checksum
, lineno_checksum
);
1333 if (flag_profile_values
)
1334 compute_value_histograms (values
, cfg_checksum
, lineno_checksum
);
1337 remove_fake_edges ();
1339 /* For each edge not on the spanning tree, add counting code. */
1340 if (profile_arc_flag
1341 && coverage_counter_alloc (GCOV_COUNTER_ARCS
, num_instrumented
))
1343 unsigned n_instrumented
;
1345 gimple_init_gcov_profiler ();
1347 n_instrumented
= instrument_edges (el
);
1349 gcc_assert (n_instrumented
== num_instrumented
);
1351 if (flag_profile_values
)
1352 instrument_values (values
);
1354 /* Commit changes done by instrumentation. */
1355 gsi_commit_edge_inserts ();
1358 free_aux_for_edges ();
1361 free_edge_list (el
);
1362 coverage_end_function (lineno_checksum
, cfg_checksum
);
1363 if (flag_branch_probabilities
&& profile_info
)
1366 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1367 report_predictor_hitrates ();
1368 profile_status_for_fn (cfun
) = PROFILE_READ
;
1370 /* At this moment we have precise loop iteration count estimates.
1371 Record them to loop structure before the profile gets out of date. */
1372 FOR_EACH_LOOP (loop
, 0)
1373 if (loop
->header
->count
> 0)
1375 gcov_type nit
= expected_loop_iterations_unbounded (loop
);
1376 widest_int bound
= gcov_type_to_wide_int (nit
);
1377 loop
->any_estimate
= false;
1378 record_niter_bound (loop
, bound
, true, false);
1380 compute_function_frequency ();
1384 /* Union find algorithm implementation for the basic blocks using
1388 find_group (basic_block bb
)
1390 basic_block group
= bb
, bb1
;
1392 while ((basic_block
) group
->aux
!= group
)
1393 group
= (basic_block
) group
->aux
;
1395 /* Compress path. */
1396 while ((basic_block
) bb
->aux
!= group
)
1398 bb1
= (basic_block
) bb
->aux
;
1399 bb
->aux
= (void *) group
;
1406 union_groups (basic_block bb1
, basic_block bb2
)
1408 basic_block bb1g
= find_group (bb1
);
1409 basic_block bb2g
= find_group (bb2
);
1411 /* ??? I don't have a place for the rank field. OK. Lets go w/o it,
1412 this code is unlikely going to be performance problem anyway. */
1413 gcc_assert (bb1g
!= bb2g
);
1418 /* This function searches all of the edges in the program flow graph, and puts
1419 as many bad edges as possible onto the spanning tree. Bad edges include
1420 abnormals edges, which can't be instrumented at the moment. Since it is
1421 possible for fake edges to form a cycle, we will have to develop some
1422 better way in the future. Also put critical edges to the tree, since they
1423 are more expensive to instrument. */
1426 find_spanning_tree (struct edge_list
*el
)
1429 int num_edges
= NUM_EDGES (el
);
1432 /* We use aux field for standard union-find algorithm. */
1433 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
1436 /* Add fake edge exit to entry we can't instrument. */
1437 union_groups (EXIT_BLOCK_PTR_FOR_FN (cfun
), ENTRY_BLOCK_PTR_FOR_FN (cfun
));
1439 /* First add all abnormal edges to the tree unless they form a cycle. Also
1440 add all edges to the exit block to avoid inserting profiling code behind
1441 setting return value from function. */
1442 for (i
= 0; i
< num_edges
; i
++)
1444 edge e
= INDEX_EDGE (el
, i
);
1445 if (((e
->flags
& (EDGE_ABNORMAL
| EDGE_ABNORMAL_CALL
| EDGE_FAKE
))
1446 || e
->dest
== EXIT_BLOCK_PTR_FOR_FN (cfun
))
1447 && !EDGE_INFO (e
)->ignore
1448 && (find_group (e
->src
) != find_group (e
->dest
)))
1451 fprintf (dump_file
, "Abnormal edge %d to %d put to tree\n",
1452 e
->src
->index
, e
->dest
->index
);
1453 EDGE_INFO (e
)->on_tree
= 1;
1454 union_groups (e
->src
, e
->dest
);
1458 /* And now the rest. Edge list is sorted according to frequencies and
1459 thus we will produce minimal spanning tree. */
1460 for (i
= 0; i
< num_edges
; i
++)
1462 edge e
= INDEX_EDGE (el
, i
);
1463 if (!EDGE_INFO (e
)->ignore
1464 && find_group (e
->src
) != find_group (e
->dest
))
1467 fprintf (dump_file
, "Normal edge %d to %d put to tree\n",
1468 e
->src
->index
, e
->dest
->index
);
1469 EDGE_INFO (e
)->on_tree
= 1;
1470 union_groups (e
->src
, e
->dest
);
1474 clear_aux_for_blocks ();
1477 /* Perform file-level initialization for branch-prob processing. */
1480 init_branch_prob (void)
1484 total_num_blocks
= 0;
1485 total_num_edges
= 0;
1486 total_num_edges_ignored
= 0;
1487 total_num_edges_instrumented
= 0;
1488 total_num_blocks_created
= 0;
1489 total_num_passes
= 0;
1490 total_num_times_called
= 0;
1491 total_num_branches
= 0;
1492 for (i
= 0; i
< 20; i
++)
1493 total_hist_br_prob
[i
] = 0;
1496 /* Performs file-level cleanup after branch-prob processing
1500 end_branch_prob (void)
1504 fprintf (dump_file
, "\n");
1505 fprintf (dump_file
, "Total number of blocks: %d\n",
1507 fprintf (dump_file
, "Total number of edges: %d\n", total_num_edges
);
1508 fprintf (dump_file
, "Total number of ignored edges: %d\n",
1509 total_num_edges_ignored
);
1510 fprintf (dump_file
, "Total number of instrumented edges: %d\n",
1511 total_num_edges_instrumented
);
1512 fprintf (dump_file
, "Total number of blocks created: %d\n",
1513 total_num_blocks_created
);
1514 fprintf (dump_file
, "Total number of graph solution passes: %d\n",
1516 if (total_num_times_called
!= 0)
1517 fprintf (dump_file
, "Average number of graph solution passes: %d\n",
1518 (total_num_passes
+ (total_num_times_called
>> 1))
1519 / total_num_times_called
);
1520 fprintf (dump_file
, "Total number of branches: %d\n",
1521 total_num_branches
);
1522 if (total_num_branches
)
1526 for (i
= 0; i
< 10; i
++)
1527 fprintf (dump_file
, "%d%% branches in range %d-%d%%\n",
1528 (total_hist_br_prob
[i
] + total_hist_br_prob
[19-i
]) * 100
1529 / total_num_branches
, 5*i
, 5*i
+5);