1 /* Calculate branch probabilities, and basic block execution counts.
2 Copyright (C) 1990-2019 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 gcov_summary
*profile_info
;
89 /* Collect statistics on the performance of this pass for the entire source
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. */
111 instrument_edges (struct edge_list
*el
)
113 unsigned num_instr_edges
= 0;
114 int num_edges
= NUM_EDGES (el
);
117 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
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
));
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
;
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. */
146 instrument_values (histogram_values values
)
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
))
162 case HIST_TYPE_INTERVAL
:
163 gimple_gen_interval_profiler (hist
, t
, 0);
167 gimple_gen_pow2_profiler (hist
, t
, 0);
170 case HIST_TYPE_TOPN_VALUES
:
171 gimple_gen_topn_values_profiler (hist
, t
, 0);
174 case HIST_TYPE_INDIR_CALL
:
175 gimple_gen_ic_profiler (hist
, t
, 0);
178 case HIST_TYPE_AVERAGE
:
179 gimple_gen_average_profiler (hist
, t
, 0);
183 gimple_gen_ior_profiler (hist
, t
, 0);
186 case HIST_TYPE_TIME_PROFILE
:
187 gimple_gen_time_profiler (t
, 0);
197 /* Computes hybrid profile for all matching entries in da_file.
199 CFG_CHECKSUM is the precomputed checksum for the CFG. */
202 get_exec_counts (unsigned cfg_checksum
, unsigned lineno_checksum
)
204 unsigned num_edges
= 0;
208 /* Count the edges to be (possibly) instrumented. */
209 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
214 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
215 if (!EDGE_INFO (e
)->ignore
&& !EDGE_INFO (e
)->on_tree
)
219 counts
= get_coverage_counts (GCOV_COUNTER_ARCS
, cfg_checksum
,
220 lineno_checksum
, num_edges
);
228 is_edge_inconsistent (vec
<edge
, va_gc
> *edges
)
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
)))
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
);
256 correct_negative_edge_counts (void)
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. */
275 is_inconsistent (void)
278 bool inconsistent
= false;
279 FOR_EACH_BB_FN (bb
, cfun
)
281 inconsistent
|= is_edge_inconsistent (bb
->preds
);
282 if (!dump_file
&& inconsistent
)
284 inconsistent
|= is_edge_inconsistent (bb
->succs
);
285 if (!dump_file
&& inconsistent
)
287 if (bb_gcov_count (bb
) < 0)
291 fprintf (dump_file
, "BB %i count is negative "
295 dump_bb (dump_file
, bb
, 0, TDF_DETAILS
);
299 if (bb_gcov_count (bb
) != sum_edge_counts (bb
->preds
))
303 fprintf (dump_file
, "BB %i count does not match sum of incoming edges "
304 "%" PRId64
" should be %" PRId64
,
307 sum_edge_counts (bb
->preds
));
308 dump_bb (dump_file
, bb
, 0, TDF_DETAILS
);
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
)))
318 fprintf (dump_file
, "BB %i count does not match sum of outgoing edges "
319 "%" PRId64
" should be %" PRId64
,
322 sum_edge_counts (bb
->succs
));
323 dump_bb (dump_file
, bb
, 0, TDF_DETAILS
);
327 if (!dump_file
&& inconsistent
)
334 /* Set each basic block count to the sum of its outgoing edge counts */
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 */
348 read_profile_edge_counts (gcov_type
*exec_counts
)
352 int exec_counts_pos
= 0;
353 /* For each edge not on the spanning tree, set its execution count from
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
)
363 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
364 if (!EDGE_INFO (e
)->ignore
&& !EDGE_INFO (e
)->on_tree
)
368 edge_gcov_count (e
) = exec_counts
[exec_counts_pos
++];
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
--;
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
));
389 /* Compute the branch probabilities for the various branches.
390 Annotate them accordingly.
392 CFG_CHECKSUM is the precomputed checksum for the CFG. */
395 compute_branch_probabilities (unsigned cfg_checksum
, unsigned lineno_checksum
)
402 int hist_br_prob
[20];
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. */
411 fprintf (dump_file
, "Profile info is missing; giving up\n");
415 bb_gcov_counts
.safe_grow_cleared (last_basic_block_for_fn (cfun
));
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
)
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
);
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. */
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)
476 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
477 total
+= edge_gcov_count (e
);
478 bb_gcov_count (bb
) = total
;
482 else if (bi
->pred_count
== 0)
488 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
489 total
+= edge_gcov_count (e
);
490 bb_gcov_count (bb
) = total
;
497 if (bi
->succ_count
== 1)
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
)
513 /* Calculate count for remaining edge by conservation. */
514 total
= bb_gcov_count (bb
) - total
;
517 EDGE_INFO (e
)->count_valid
= 1;
518 edge_gcov_count (e
) = total
;
521 BB_INFO (e
->dest
)->pred_count
--;
524 if (bi
->pred_count
== 1)
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
)
540 /* Calculate count for remaining edge by conservation. */
541 total
= bb_gcov_count (bb
) - total
+ edge_gcov_count (e
);
544 EDGE_INFO (e
)->count_valid
= 1;
545 edge_gcov_count (e
) = total
;
548 BB_INFO (e
->src
)->succ_count
--;
555 total_num_passes
+= passes
;
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 ();
571 if (flag_profile_correction
)
573 /* Inconsistency detected. Make it flow-consistent. */
574 static int informed
= 0;
575 if (dump_enabled_p () && informed
== 0)
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 */
586 fprintf (dump_file
, "\nCalling mcf_smooth_cfg\n");
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
++)
600 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
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
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 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
639 e
->probability
= profile_probability::probability_in_gcov_type
640 (edge_gcov_count (e
), bb_gcov_count (bb
));
641 if (bb
->index
>= NUM_FIXED_BLOCKS
642 && block_ends_with_condjump_p (bb
)
643 && EDGE_COUNT (bb
->succs
) >= 2)
649 /* Find the branch edge. It is possible that we do have fake
651 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
652 if (!(e
->flags
& (EDGE_FAKE
| EDGE_FALLTHRU
)))
655 prob
= e
->probability
.to_reg_br_prob_base ();
656 index
= prob
* 20 / REG_BR_PROB_BASE
;
660 hist_br_prob
[index
]++;
665 /* As a last resort, distribute the probabilities evenly.
666 Use simple heuristics that if there are normal edges,
667 give all abnormals frequency of 0, otherwise distribute the
668 frequency over abnormals (this is the case of noreturn
670 else if (profile_status_for_fn (cfun
) == PROFILE_ABSENT
)
674 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
675 if (!(e
->flags
& (EDGE_COMPLEX
| EDGE_FAKE
)))
679 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
680 if (!(e
->flags
& (EDGE_COMPLEX
| EDGE_FAKE
)))
682 = profile_probability::guessed_always ().apply_scale (1, total
);
684 e
->probability
= profile_probability::never ();
688 total
+= EDGE_COUNT (bb
->succs
);
689 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
691 = profile_probability::guessed_always ().apply_scale (1, total
);
693 if (bb
->index
>= NUM_FIXED_BLOCKS
694 && block_ends_with_condjump_p (bb
)
695 && EDGE_COUNT (bb
->succs
) >= 2)
701 profile_status_for_fn (cfun
) = PROFILE_READ
;
703 /* If we have real data, use them! */
704 if (bb_gcov_count (ENTRY_BLOCK_PTR_FOR_FN (cfun
))
705 || !flag_guess_branch_prob
)
706 FOR_ALL_BB_FN (bb
, cfun
)
707 bb
->count
= profile_count::from_gcov_type (bb_gcov_count (bb
));
708 /* If function was not trained, preserve local estimates including statically
709 determined zero counts. */
710 else if (profile_status_for_fn (cfun
) == PROFILE_READ
)
711 FOR_ALL_BB_FN (bb
, cfun
)
712 if (!(bb
->count
== profile_count::zero ()))
713 bb
->count
= bb
->count
.global0 ();
715 bb_gcov_counts
.release ();
716 delete edge_gcov_counts
;
717 edge_gcov_counts
= NULL
;
719 update_max_bb_count ();
723 fprintf (dump_file
, " Profile feedback for function");
724 fprintf (dump_file
, ((profile_status_for_fn (cfun
) == PROFILE_READ
)
726 : " is not available \n"));
728 fprintf (dump_file
, "%d branches\n", num_branches
);
730 for (i
= 0; i
< 10; i
++)
731 fprintf (dump_file
, "%d%% branches in range %d-%d%%\n",
732 (hist_br_prob
[i
] + hist_br_prob
[19-i
]) * 100 / num_branches
,
735 total_num_branches
+= num_branches
;
736 for (i
= 0; i
< 20; i
++)
737 total_hist_br_prob
[i
] += hist_br_prob
[i
];
739 fputc ('\n', dump_file
);
740 fputc ('\n', dump_file
);
743 free_aux_for_blocks ();
746 /* Load value histograms values whose description is stored in VALUES array
749 CFG_CHECKSUM is the precomputed checksum for the CFG. */
752 compute_value_histograms (histogram_values values
, unsigned cfg_checksum
,
753 unsigned lineno_checksum
)
755 unsigned i
, j
, t
, any
;
756 unsigned n_histogram_counters
[GCOV_N_VALUE_COUNTERS
];
757 gcov_type
*histogram_counts
[GCOV_N_VALUE_COUNTERS
];
758 gcov_type
*act_count
[GCOV_N_VALUE_COUNTERS
];
759 gcov_type
*aact_count
;
760 struct cgraph_node
*node
;
762 for (t
= 0; t
< GCOV_N_VALUE_COUNTERS
; t
++)
763 n_histogram_counters
[t
] = 0;
765 for (i
= 0; i
< values
.length (); i
++)
767 histogram_value hist
= values
[i
];
768 n_histogram_counters
[(int) hist
->type
] += hist
->n_counters
;
772 for (t
= 0; t
< GCOV_N_VALUE_COUNTERS
; t
++)
774 if (!n_histogram_counters
[t
])
776 histogram_counts
[t
] = NULL
;
780 histogram_counts
[t
] = get_coverage_counts (COUNTER_FOR_HIST_TYPE (t
),
783 n_histogram_counters
[t
]);
784 if (histogram_counts
[t
])
786 act_count
[t
] = histogram_counts
[t
];
791 for (i
= 0; i
< values
.length (); i
++)
793 histogram_value hist
= values
[i
];
794 gimple
*stmt
= hist
->hvalue
.stmt
;
796 t
= (int) hist
->type
;
798 aact_count
= act_count
[t
];
801 act_count
[t
] += hist
->n_counters
;
803 gimple_add_histogram_value (cfun
, stmt
, hist
);
804 hist
->hvalue
.counters
= XNEWVEC (gcov_type
, hist
->n_counters
);
805 for (j
= 0; j
< hist
->n_counters
; j
++)
807 hist
->hvalue
.counters
[j
] = aact_count
[j
];
809 hist
->hvalue
.counters
[j
] = 0;
811 /* Time profiler counter is not related to any statement,
812 so that we have to read the counter and set the value to
813 the corresponding call graph node. */
814 if (hist
->type
== HIST_TYPE_TIME_PROFILE
)
816 node
= cgraph_node::get (hist
->fun
->decl
);
817 node
->tp_first_run
= hist
->hvalue
.counters
[0];
820 fprintf (dump_file
, "Read tp_first_run: %d\n", node
->tp_first_run
);
824 for (t
= 0; t
< GCOV_N_VALUE_COUNTERS
; t
++)
825 free (histogram_counts
[t
]);
828 /* Location triplet which records a location. */
829 struct location_triplet
831 const char *filename
;
836 /* Traits class for streamed_locations hash set below. */
838 struct location_triplet_hash
: typed_noop_remove
<location_triplet
>
840 typedef location_triplet value_type
;
841 typedef location_triplet compare_type
;
844 hash (const location_triplet
&ref
)
846 inchash::hash
hstate (0);
848 hstate
.add_int (strlen (ref
.filename
));
849 hstate
.add_int (ref
.lineno
);
850 hstate
.add_int (ref
.bb_index
);
851 return hstate
.end ();
855 equal (const location_triplet
&ref1
, const location_triplet
&ref2
)
857 return ref1
.lineno
== ref2
.lineno
858 && ref1
.bb_index
== ref2
.bb_index
859 && ref1
.filename
!= NULL
860 && ref2
.filename
!= NULL
861 && strcmp (ref1
.filename
, ref2
.filename
) == 0;
865 mark_deleted (location_triplet
&ref
)
871 mark_empty (location_triplet
&ref
)
877 is_deleted (const location_triplet
&ref
)
879 return ref
.lineno
== -1;
883 is_empty (const location_triplet
&ref
)
885 return ref
.lineno
== -2;
892 /* When passed NULL as file_name, initialize.
893 When passed something else, output the necessary commands to change
894 line to LINE and offset to FILE_NAME. */
896 output_location (hash_set
<location_triplet_hash
> *streamed_locations
,
897 char const *file_name
, int line
,
898 gcov_position_t
*offset
, basic_block bb
)
900 static char const *prev_file_name
;
901 static int prev_line
;
902 bool name_differs
, line_differs
;
904 location_triplet triplet
;
905 triplet
.filename
= file_name
;
906 triplet
.lineno
= line
;
907 triplet
.bb_index
= bb
? bb
->index
: 0;
909 if (streamed_locations
->add (triplet
))
914 prev_file_name
= NULL
;
919 name_differs
= !prev_file_name
|| filename_cmp (file_name
, prev_file_name
);
920 line_differs
= prev_line
!= line
;
924 *offset
= gcov_write_tag (GCOV_TAG_LINES
);
925 gcov_write_unsigned (bb
->index
);
926 name_differs
= line_differs
= true;
929 /* If this is a new source file, then output the
930 file's name to the .bb file. */
933 prev_file_name
= file_name
;
934 gcov_write_unsigned (0);
935 gcov_write_filename (prev_file_name
);
939 gcov_write_unsigned (line
);
944 /* Helper for qsort so edges get sorted from highest frequency to smallest.
945 This controls the weight for minimal spanning tree algorithm */
947 compare_freqs (const void *p1
, const void *p2
)
949 const_edge e1
= *(const const_edge
*)p1
;
950 const_edge e2
= *(const const_edge
*)p2
;
952 /* Critical edges needs to be split which introduce extra control flow.
953 Make them more heavy. */
954 int m1
= EDGE_CRITICAL_P (e1
) ? 2 : 1;
955 int m2
= EDGE_CRITICAL_P (e2
) ? 2 : 1;
957 if (EDGE_FREQUENCY (e1
) * m1
+ m1
!= EDGE_FREQUENCY (e2
) * m2
+ m2
)
958 return EDGE_FREQUENCY (e2
) * m2
+ m2
- EDGE_FREQUENCY (e1
) * m1
- m1
;
959 /* Stabilize sort. */
960 if (e1
->src
->index
!= e2
->src
->index
)
961 return e2
->src
->index
- e1
->src
->index
;
962 return e2
->dest
->index
- e1
->dest
->index
;
965 /* Only read execution count for thunks. */
968 read_thunk_profile (struct cgraph_node
*node
)
970 tree old
= current_function_decl
;
971 current_function_decl
= node
->decl
;
972 gcov_type
*counts
= get_coverage_counts (GCOV_COUNTER_ARCS
, 0, 0, 1);
975 node
->callees
->count
= node
->count
976 = profile_count::from_gcov_type (counts
[0]);
979 current_function_decl
= old
;
984 /* Instrument and/or analyze program behavior based on program the CFG.
986 This function creates a representation of the control flow graph (of
987 the function being compiled) that is suitable for the instrumentation
988 of edges and/or converting measured edge counts to counts on the
991 When FLAG_PROFILE_ARCS is nonzero, this function instruments the edges in
992 the flow graph that are needed to reconstruct the dynamic behavior of the
993 flow graph. This data is written to the gcno file for gcov.
995 When FLAG_BRANCH_PROBABILITIES is nonzero, this function reads auxiliary
996 information from the gcda file containing edge count information from
997 previous executions of the function being compiled. In this case, the
998 control flow graph is annotated with actual execution counts by
999 compute_branch_probabilities().
1001 Main entry point of this file. */
1004 branch_prob (bool thunk
)
1008 unsigned num_edges
, ignored_edges
;
1009 unsigned num_instrumented
;
1010 struct edge_list
*el
;
1011 histogram_values values
= histogram_values ();
1012 unsigned cfg_checksum
, lineno_checksum
;
1014 total_num_times_called
++;
1016 flow_call_edges_add (NULL
);
1017 add_noreturn_fake_exit_edges ();
1019 hash_set
<location_triplet_hash
> streamed_locations
;
1023 /* We can't handle cyclic regions constructed using abnormal edges.
1024 To avoid these we replace every source of abnormal edge by a fake
1025 edge from entry node and every destination by fake edge to exit.
1026 This keeps graph acyclic and our calculation exact for all normal
1027 edges except for exit and entrance ones.
1029 We also add fake exit edges for each call and asm statement in the
1030 basic, since it may not return. */
1032 FOR_EACH_BB_FN (bb
, cfun
)
1034 int need_exit_edge
= 0, need_entry_edge
= 0;
1035 int have_exit_edge
= 0, have_entry_edge
= 0;
1039 /* Functions returning multiple times are not handled by extra edges.
1040 Instead we simply allow negative counts on edges from exit to the
1041 block past call and corresponding probabilities. We can't go
1042 with the extra edges because that would result in flowgraph that
1043 needs to have fake edges outside the spanning tree. */
1045 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1047 gimple_stmt_iterator gsi
;
1048 gimple
*last
= NULL
;
1050 /* It may happen that there are compiler generated statements
1051 without a locus at all. Go through the basic block from the
1052 last to the first statement looking for a locus. */
1053 for (gsi
= gsi_last_nondebug_bb (bb
);
1055 gsi_prev_nondebug (&gsi
))
1057 last
= gsi_stmt (gsi
);
1058 if (!RESERVED_LOCATION_P (gimple_location (last
)))
1062 /* Edge with goto locus might get wrong coverage info unless
1063 it is the only edge out of BB.
1064 Don't do that when the locuses match, so
1065 if (blah) goto something;
1066 is not computed twice. */
1068 && gimple_has_location (last
)
1069 && !RESERVED_LOCATION_P (e
->goto_locus
)
1070 && !single_succ_p (bb
)
1071 && (LOCATION_FILE (e
->goto_locus
)
1072 != LOCATION_FILE (gimple_location (last
))
1073 || (LOCATION_LINE (e
->goto_locus
)
1074 != LOCATION_LINE (gimple_location (last
)))))
1076 basic_block new_bb
= split_edge (e
);
1077 edge ne
= single_succ_edge (new_bb
);
1078 ne
->goto_locus
= e
->goto_locus
;
1080 if ((e
->flags
& (EDGE_ABNORMAL
| EDGE_ABNORMAL_CALL
))
1081 && e
->dest
!= EXIT_BLOCK_PTR_FOR_FN (cfun
))
1083 if (e
->dest
== EXIT_BLOCK_PTR_FOR_FN (cfun
))
1086 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1088 if ((e
->flags
& (EDGE_ABNORMAL
| EDGE_ABNORMAL_CALL
))
1089 && e
->src
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
))
1090 need_entry_edge
= 1;
1091 if (e
->src
== ENTRY_BLOCK_PTR_FOR_FN (cfun
))
1092 have_entry_edge
= 1;
1095 if (need_exit_edge
&& !have_exit_edge
)
1098 fprintf (dump_file
, "Adding fake exit edge to bb %i\n",
1100 make_edge (bb
, EXIT_BLOCK_PTR_FOR_FN (cfun
), EDGE_FAKE
);
1102 if (need_entry_edge
&& !have_entry_edge
)
1105 fprintf (dump_file
, "Adding fake entry edge to bb %i\n",
1107 make_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun
), bb
, EDGE_FAKE
);
1108 /* Avoid bbs that have both fake entry edge and also some
1109 exit edge. One of those edges wouldn't be added to the
1110 spanning tree, but we can't instrument any of them. */
1111 if (have_exit_edge
|| need_exit_edge
)
1113 gimple_stmt_iterator gsi
;
1116 gsi
= gsi_start_nondebug_after_labels_bb (bb
);
1117 gcc_checking_assert (!gsi_end_p (gsi
));
1118 first
= gsi_stmt (gsi
);
1119 /* Don't split the bbs containing __builtin_setjmp_receiver
1120 or ABNORMAL_DISPATCHER calls. These are very
1121 special and don't expect anything to be inserted before
1123 if (is_gimple_call (first
)
1124 && (gimple_call_builtin_p (first
, BUILT_IN_SETJMP_RECEIVER
)
1125 || (gimple_call_flags (first
) & ECF_RETURNS_TWICE
)
1126 || (gimple_call_internal_p (first
)
1127 && (gimple_call_internal_fn (first
)
1128 == IFN_ABNORMAL_DISPATCHER
))))
1132 fprintf (dump_file
, "Splitting bb %i after labels\n",
1134 split_block_after_labels (bb
);
1140 el
= create_edge_list ();
1141 num_edges
= NUM_EDGES (el
);
1142 qsort (el
->index_to_edge
, num_edges
, sizeof (edge
), compare_freqs
);
1143 alloc_aux_for_edges (sizeof (struct edge_profile_info
));
1145 /* The basic blocks are expected to be numbered sequentially. */
1149 for (i
= 0 ; i
< num_edges
; i
++)
1151 edge e
= INDEX_EDGE (el
, i
);
1153 /* Mark edges we've replaced by fake edges above as ignored. */
1154 if ((e
->flags
& (EDGE_ABNORMAL
| EDGE_ABNORMAL_CALL
))
1155 && e
->src
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
)
1156 && e
->dest
!= EXIT_BLOCK_PTR_FOR_FN (cfun
))
1158 EDGE_INFO (e
)->ignore
= 1;
1163 /* Create spanning tree from basic block graph, mark each edge that is
1164 on the spanning tree. We insert as many abnormal and critical edges
1165 as possible to minimize number of edge splits necessary. */
1168 find_spanning_tree (el
);
1173 /* Keep only edge from entry block to be instrumented. */
1174 FOR_EACH_BB_FN (bb
, cfun
)
1175 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1176 EDGE_INFO (e
)->ignore
= true;
1180 /* Fake edges that are not on the tree will not be instrumented, so
1181 mark them ignored. */
1182 for (num_instrumented
= i
= 0; i
< num_edges
; i
++)
1184 edge e
= INDEX_EDGE (el
, i
);
1185 struct edge_profile_info
*inf
= EDGE_INFO (e
);
1187 if (inf
->ignore
|| inf
->on_tree
)
1189 else if (e
->flags
& EDGE_FAKE
)
1198 total_num_blocks
+= n_basic_blocks_for_fn (cfun
);
1200 fprintf (dump_file
, "%d basic blocks\n", n_basic_blocks_for_fn (cfun
));
1202 total_num_edges
+= num_edges
;
1204 fprintf (dump_file
, "%d edges\n", num_edges
);
1206 total_num_edges_ignored
+= ignored_edges
;
1208 fprintf (dump_file
, "%d ignored edges\n", ignored_edges
);
1210 total_num_edges_instrumented
+= num_instrumented
;
1212 fprintf (dump_file
, "%d instrumentation edges\n", num_instrumented
);
1214 /* Compute two different checksums. Note that we want to compute
1215 the checksum in only once place, since it depends on the shape
1216 of the control flow which can change during
1217 various transformations. */
1220 /* At stream in time we do not have CFG, so we cannot do checksums. */
1222 lineno_checksum
= 0;
1226 cfg_checksum
= coverage_compute_cfg_checksum (cfun
);
1227 lineno_checksum
= coverage_compute_lineno_checksum ();
1230 /* Write the data from which gcov can reconstruct the basic block
1231 graph and function line numbers (the gcno file). */
1232 if (coverage_begin_function (lineno_checksum
, cfg_checksum
))
1234 gcov_position_t offset
;
1236 /* Basic block flags */
1237 offset
= gcov_write_tag (GCOV_TAG_BLOCKS
);
1238 gcov_write_unsigned (n_basic_blocks_for_fn (cfun
));
1239 gcov_write_length (offset
);
1242 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
),
1243 EXIT_BLOCK_PTR_FOR_FN (cfun
), next_bb
)
1248 offset
= gcov_write_tag (GCOV_TAG_ARCS
);
1249 gcov_write_unsigned (bb
->index
);
1251 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1253 struct edge_profile_info
*i
= EDGE_INFO (e
);
1256 unsigned flag_bits
= 0;
1259 flag_bits
|= GCOV_ARC_ON_TREE
;
1260 if (e
->flags
& EDGE_FAKE
)
1261 flag_bits
|= GCOV_ARC_FAKE
;
1262 if (e
->flags
& EDGE_FALLTHRU
)
1263 flag_bits
|= GCOV_ARC_FALLTHROUGH
;
1264 /* On trees we don't have fallthru flags, but we can
1265 recompute them from CFG shape. */
1266 if (e
->flags
& (EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
)
1267 && e
->src
->next_bb
== e
->dest
)
1268 flag_bits
|= GCOV_ARC_FALLTHROUGH
;
1270 gcov_write_unsigned (e
->dest
->index
);
1271 gcov_write_unsigned (flag_bits
);
1275 gcov_write_length (offset
);
1279 /* Initialize the output. */
1280 output_location (&streamed_locations
, NULL
, 0, NULL
, NULL
);
1282 hash_set
<int_hash
<location_t
, 0, 2> > seen_locations
;
1284 FOR_EACH_BB_FN (bb
, cfun
)
1286 gimple_stmt_iterator gsi
;
1287 gcov_position_t offset
= 0;
1289 if (bb
== ENTRY_BLOCK_PTR_FOR_FN (cfun
)->next_bb
)
1291 location_t loc
= DECL_SOURCE_LOCATION (current_function_decl
);
1292 seen_locations
.add (loc
);
1293 expanded_location curr_location
= expand_location (loc
);
1294 output_location (&streamed_locations
, curr_location
.file
,
1295 curr_location
.line
, &offset
, bb
);
1298 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1300 gimple
*stmt
= gsi_stmt (gsi
);
1301 location_t loc
= gimple_location (stmt
);
1302 if (!RESERVED_LOCATION_P (loc
))
1304 seen_locations
.add (loc
);
1305 output_location (&streamed_locations
, gimple_filename (stmt
),
1306 gimple_lineno (stmt
), &offset
, bb
);
1310 /* Notice GOTO expressions eliminated while constructing the CFG.
1311 It's hard to distinguish such expression, but goto_locus should
1312 not be any of already seen location. */
1314 if (single_succ_p (bb
)
1315 && (loc
= single_succ_edge (bb
)->goto_locus
)
1316 && !RESERVED_LOCATION_P (loc
)
1317 && !seen_locations
.contains (loc
))
1319 expanded_location curr_location
= expand_location (loc
);
1320 output_location (&streamed_locations
, curr_location
.file
,
1321 curr_location
.line
, &offset
, bb
);
1326 /* A file of NULL indicates the end of run. */
1327 gcov_write_unsigned (0);
1328 gcov_write_string (NULL
);
1329 gcov_write_length (offset
);
1334 if (flag_profile_values
)
1335 gimple_find_values_to_profile (&values
);
1337 if (flag_branch_probabilities
)
1339 compute_branch_probabilities (cfg_checksum
, lineno_checksum
);
1340 if (flag_profile_values
)
1341 compute_value_histograms (values
, cfg_checksum
, lineno_checksum
);
1344 remove_fake_edges ();
1346 /* For each edge not on the spanning tree, add counting code. */
1347 if (profile_arc_flag
1348 && coverage_counter_alloc (GCOV_COUNTER_ARCS
, num_instrumented
))
1350 unsigned n_instrumented
;
1352 gimple_init_gcov_profiler ();
1354 n_instrumented
= instrument_edges (el
);
1356 gcc_assert (n_instrumented
== num_instrumented
);
1358 if (flag_profile_values
)
1359 instrument_values (values
);
1361 /* Commit changes done by instrumentation. */
1362 gsi_commit_edge_inserts ();
1365 free_aux_for_edges ();
1368 free_edge_list (el
);
1369 coverage_end_function (lineno_checksum
, cfg_checksum
);
1370 if (flag_branch_probabilities
1371 && (profile_status_for_fn (cfun
) == PROFILE_READ
))
1374 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1375 report_predictor_hitrates ();
1377 /* At this moment we have precise loop iteration count estimates.
1378 Record them to loop structure before the profile gets out of date. */
1379 FOR_EACH_LOOP (loop
, 0)
1380 if (loop
->header
->count
> 0)
1382 gcov_type nit
= expected_loop_iterations_unbounded (loop
);
1383 widest_int bound
= gcov_type_to_wide_int (nit
);
1384 loop
->any_estimate
= false;
1385 record_niter_bound (loop
, bound
, true, false);
1387 compute_function_frequency ();
1391 /* Union find algorithm implementation for the basic blocks using
1395 find_group (basic_block bb
)
1397 basic_block group
= bb
, bb1
;
1399 while ((basic_block
) group
->aux
!= group
)
1400 group
= (basic_block
) group
->aux
;
1402 /* Compress path. */
1403 while ((basic_block
) bb
->aux
!= group
)
1405 bb1
= (basic_block
) bb
->aux
;
1406 bb
->aux
= (void *) group
;
1413 union_groups (basic_block bb1
, basic_block bb2
)
1415 basic_block bb1g
= find_group (bb1
);
1416 basic_block bb2g
= find_group (bb2
);
1418 /* ??? I don't have a place for the rank field. OK. Lets go w/o it,
1419 this code is unlikely going to be performance problem anyway. */
1420 gcc_assert (bb1g
!= bb2g
);
1425 /* This function searches all of the edges in the program flow graph, and puts
1426 as many bad edges as possible onto the spanning tree. Bad edges include
1427 abnormals edges, which can't be instrumented at the moment. Since it is
1428 possible for fake edges to form a cycle, we will have to develop some
1429 better way in the future. Also put critical edges to the tree, since they
1430 are more expensive to instrument. */
1433 find_spanning_tree (struct edge_list
*el
)
1436 int num_edges
= NUM_EDGES (el
);
1439 /* We use aux field for standard union-find algorithm. */
1440 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
1443 /* Add fake edge exit to entry we can't instrument. */
1444 union_groups (EXIT_BLOCK_PTR_FOR_FN (cfun
), ENTRY_BLOCK_PTR_FOR_FN (cfun
));
1446 /* First add all abnormal edges to the tree unless they form a cycle. Also
1447 add all edges to the exit block to avoid inserting profiling code behind
1448 setting return value from function. */
1449 for (i
= 0; i
< num_edges
; i
++)
1451 edge e
= INDEX_EDGE (el
, i
);
1452 if (((e
->flags
& (EDGE_ABNORMAL
| EDGE_ABNORMAL_CALL
| EDGE_FAKE
))
1453 || e
->dest
== EXIT_BLOCK_PTR_FOR_FN (cfun
))
1454 && !EDGE_INFO (e
)->ignore
1455 && (find_group (e
->src
) != find_group (e
->dest
)))
1458 fprintf (dump_file
, "Abnormal edge %d to %d put to tree\n",
1459 e
->src
->index
, e
->dest
->index
);
1460 EDGE_INFO (e
)->on_tree
= 1;
1461 union_groups (e
->src
, e
->dest
);
1465 /* And now the rest. Edge list is sorted according to frequencies and
1466 thus we will produce minimal spanning tree. */
1467 for (i
= 0; i
< num_edges
; i
++)
1469 edge e
= INDEX_EDGE (el
, i
);
1470 if (!EDGE_INFO (e
)->ignore
1471 && find_group (e
->src
) != find_group (e
->dest
))
1474 fprintf (dump_file
, "Normal edge %d to %d put to tree\n",
1475 e
->src
->index
, e
->dest
->index
);
1476 EDGE_INFO (e
)->on_tree
= 1;
1477 union_groups (e
->src
, e
->dest
);
1481 clear_aux_for_blocks ();
1484 /* Perform file-level initialization for branch-prob processing. */
1487 init_branch_prob (void)
1491 total_num_blocks
= 0;
1492 total_num_edges
= 0;
1493 total_num_edges_ignored
= 0;
1494 total_num_edges_instrumented
= 0;
1495 total_num_blocks_created
= 0;
1496 total_num_passes
= 0;
1497 total_num_times_called
= 0;
1498 total_num_branches
= 0;
1499 for (i
= 0; i
< 20; i
++)
1500 total_hist_br_prob
[i
] = 0;
1503 /* Performs file-level cleanup after branch-prob processing
1507 end_branch_prob (void)
1511 fprintf (dump_file
, "\n");
1512 fprintf (dump_file
, "Total number of blocks: %d\n",
1514 fprintf (dump_file
, "Total number of edges: %d\n", total_num_edges
);
1515 fprintf (dump_file
, "Total number of ignored edges: %d\n",
1516 total_num_edges_ignored
);
1517 fprintf (dump_file
, "Total number of instrumented edges: %d\n",
1518 total_num_edges_instrumented
);
1519 fprintf (dump_file
, "Total number of blocks created: %d\n",
1520 total_num_blocks_created
);
1521 fprintf (dump_file
, "Total number of graph solution passes: %d\n",
1523 if (total_num_times_called
!= 0)
1524 fprintf (dump_file
, "Average number of graph solution passes: %d\n",
1525 (total_num_passes
+ (total_num_times_called
>> 1))
1526 / total_num_times_called
);
1527 fprintf (dump_file
, "Total number of branches: %d\n",
1528 total_num_branches
);
1529 if (total_num_branches
)
1533 for (i
= 0; i
< 10; i
++)
1534 fprintf (dump_file
, "%d%% branches in range %d-%d%%\n",
1535 (total_hist_br_prob
[i
] + total_hist_br_prob
[19-i
]) * 100
1536 / total_num_branches
, 5*i
, 5*i
+5);