1 /* Calculate branch probabilities, and basic block execution counts.
2 Copyright (C) 1990, 1991, 1992, 1993, 1994, 1996, 1997, 1998, 1999,
3 2000, 2001, 2002, 2003, 2004 Free Software Foundation, Inc.
4 Contributed by James E. Wilson, UC Berkeley/Cygnus Support;
5 based on some ideas from Dain Samples of UC Berkeley.
6 Further mangling by Bob Manson, Cygnus Support.
8 This file is part of GCC.
10 GCC is free software; you can redistribute it and/or modify it under
11 the terms of the GNU General Public License as published by the Free
12 Software Foundation; either version 2, or (at your option) any later
15 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
16 WARRANTY; without even the implied warranty of MERCHANTABILITY or
17 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
20 You should have received a copy of the GNU General Public License
21 along with GCC; see the file COPYING. If not, write to the Free
22 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
25 /* Generate basic block profile instrumentation and auxiliary files.
26 Profile generation is optimized, so that not all arcs in the basic
27 block graph need instrumenting. First, the BB graph is closed with
28 one entry (function start), and one exit (function exit). Any
29 ABNORMAL_EDGE cannot be instrumented (because there is no control
30 path to place the code). We close the graph by inserting fake
31 EDGE_FAKE edges to the EXIT_BLOCK, from the sources of abnormal
32 edges that do not go to the exit_block. We ignore such abnormal
33 edges. Naturally these fake edges are never directly traversed,
34 and so *cannot* be directly instrumented. Some other graph
35 massaging is done. To optimize the instrumentation we generate the
36 BB minimal span tree, only edges that are not on the span tree
37 (plus the entry point) need instrumenting. From that information
38 all other edge counts can be deduced. By construction all fake
39 edges must be on the spanning tree. We also attempt to place
40 EDGE_CRITICAL edges on the spanning tree.
42 The auxiliary files generated are <dumpbase>.gcno (at compile time)
43 and <dumpbase>.gcda (at run time). The format is
44 described in full in gcov-io.h. */
46 /* ??? Register allocation should use basic block execution counts to
47 give preference to the most commonly executed blocks. */
49 /* ??? Should calculate branch probabilities before instrumenting code, since
50 then we can use arc counts to help decide which arcs to instrument. */
54 #include "coretypes.h"
64 #include "value-prof.h"
67 #include "tree-flow.h"
69 /* Hooks for profiling. */
70 static struct profile_hooks* profile_hooks;
72 /* File for profiling debug output. */
74 profile_dump_file (void) {
75 return profile_hooks->profile_dump_file ();
78 /* Additional information about the edges we need. */
80 unsigned int count_valid : 1;
82 /* Is on the spanning tree. */
83 unsigned int on_tree : 1;
85 /* Pretend this edge does not exist (it is abnormal and we've
86 inserted a fake to compensate). */
87 unsigned int ignore : 1;
91 unsigned int count_valid : 1;
93 /* Number of successor and predecessor edges. */
98 #define EDGE_INFO(e) ((struct edge_info *) (e)->aux)
99 #define BB_INFO(b) ((struct bb_info *) (b)->aux)
101 /* Counter summary from the last set of coverage counts read. */
103 const struct gcov_ctr_summary *profile_info;
105 /* Collect statistics on the performance of this pass for the entire source
108 static int total_num_blocks;
109 static int total_num_edges;
110 static int total_num_edges_ignored;
111 static int total_num_edges_instrumented;
112 static int total_num_blocks_created;
113 static int total_num_passes;
114 static int total_num_times_called;
115 static int total_hist_br_prob[20];
116 static int total_num_never_executed;
117 static int total_num_branches;
119 /* Forward declarations. */
120 static void find_spanning_tree (struct edge_list *);
121 static unsigned instrument_edges (struct edge_list *);
122 static void instrument_values (histogram_values);
123 static void compute_branch_probabilities (void);
124 static void compute_value_histograms (histogram_values);
125 static gcov_type * get_exec_counts (void);
126 static basic_block find_group (basic_block);
127 static void union_groups (basic_block, basic_block);
130 /* Add edge instrumentation code to the entire insn chain.
132 F is the first insn of the chain.
133 NUM_BLOCKS is the number of basic blocks found in F. */
136 instrument_edges (struct edge_list *el)
138 unsigned num_instr_edges = 0;
139 int num_edges = NUM_EDGES (el);
142 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
146 for (e = bb->succ; e; e = e->succ_next)
148 struct edge_info *inf = EDGE_INFO (e);
150 if (!inf->ignore && !inf->on_tree)
152 if (e->flags & EDGE_ABNORMAL)
155 fprintf (dump_file, "Edge %d to %d instrumented%s\n",
156 e->src->index, e->dest->index,
157 EDGE_CRITICAL_P (e) ? " (and split)" : "");
158 (profile_hooks->gen_edge_profiler) (num_instr_edges++, e);
163 total_num_blocks_created += num_edges;
165 fprintf (dump_file, "%d edges instrumented\n", num_instr_edges);
166 return num_instr_edges;
169 /* Add code to measure histograms for values in list VALUES. */
171 instrument_values (histogram_values values)
175 /* Emit code to generate the histograms before the insns. */
177 for (i = 0; i < VEC_length (histogram_value, values); i++)
179 histogram_value hist = VEC_index (histogram_value, values, i);
182 case HIST_TYPE_INTERVAL:
183 t = GCOV_COUNTER_V_INTERVAL;
187 t = GCOV_COUNTER_V_POW2;
190 case HIST_TYPE_SINGLE_VALUE:
191 t = GCOV_COUNTER_V_SINGLE;
194 case HIST_TYPE_CONST_DELTA:
195 t = GCOV_COUNTER_V_DELTA;
201 if (!coverage_counter_alloc (t, hist->n_counters))
206 case HIST_TYPE_INTERVAL:
207 (profile_hooks->gen_interval_profiler) (hist, t, 0);
211 (profile_hooks->gen_pow2_profiler) (hist, t, 0);
214 case HIST_TYPE_SINGLE_VALUE:
215 (profile_hooks->gen_one_value_profiler) (hist, t, 0);
218 case HIST_TYPE_CONST_DELTA:
219 (profile_hooks->gen_const_delta_profiler) (hist, t, 0);
229 /* Computes hybrid profile for all matching entries in da_file. */
232 get_exec_counts (void)
234 unsigned num_edges = 0;
238 /* Count the edges to be (possibly) instrumented. */
239 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
242 for (e = bb->succ; e; e = e->succ_next)
243 if (!EDGE_INFO (e)->ignore && !EDGE_INFO (e)->on_tree)
247 counts = get_coverage_counts (GCOV_COUNTER_ARCS, num_edges, &profile_info);
251 if (dump_file && profile_info)
252 fprintf(dump_file, "Merged %u profiles with maximal count %u.\n",
253 profile_info->runs, (unsigned) profile_info->sum_max);
259 /* Compute the branch probabilities for the various branches.
260 Annotate them accordingly. */
263 compute_branch_probabilities (void)
270 int hist_br_prob[20];
271 int num_never_executed;
273 gcov_type *exec_counts = get_exec_counts ();
274 int exec_counts_pos = 0;
276 /* Very simple sanity checks so we catch bugs in our profiling code. */
279 if (profile_info->run_max * profile_info->runs < profile_info->sum_max)
281 error ("corrupted profile info: run_max * runs < sum_max");
285 if (profile_info->sum_all < profile_info->sum_max)
287 error ("corrupted profile info: sum_all is smaller than sum_max");
292 /* Attach extra info block to each bb. */
294 alloc_aux_for_blocks (sizeof (struct bb_info));
295 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
299 for (e = bb->succ; e; e = e->succ_next)
300 if (!EDGE_INFO (e)->ignore)
301 BB_INFO (bb)->succ_count++;
302 for (e = bb->pred; e; e = e->pred_next)
303 if (!EDGE_INFO (e)->ignore)
304 BB_INFO (bb)->pred_count++;
307 /* Avoid predicting entry on exit nodes. */
308 BB_INFO (EXIT_BLOCK_PTR)->succ_count = 2;
309 BB_INFO (ENTRY_BLOCK_PTR)->pred_count = 2;
311 /* For each edge not on the spanning tree, set its execution count from
314 /* The first count in the .da file is the number of times that the function
315 was entered. This is the exec_count for block zero. */
317 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
320 for (e = bb->succ; e; e = e->succ_next)
321 if (!EDGE_INFO (e)->ignore && !EDGE_INFO (e)->on_tree)
326 e->count = exec_counts[exec_counts_pos++];
327 if (e->count > profile_info->sum_max)
329 error ("corrupted profile info: edge from %i to %i exceeds maximal count",
330 bb->index, e->dest->index);
336 EDGE_INFO (e)->count_valid = 1;
337 BB_INFO (bb)->succ_count--;
338 BB_INFO (e->dest)->pred_count--;
341 fprintf (dump_file, "\nRead edge from %i to %i, count:",
342 bb->index, e->dest->index);
343 fprintf (dump_file, HOST_WIDEST_INT_PRINT_DEC,
344 (HOST_WIDEST_INT) e->count);
350 fprintf (dump_file, "\n%d edge counts read\n", num_edges);
352 /* For every block in the file,
353 - if every exit/entrance edge has a known count, then set the block count
354 - if the block count is known, and every exit/entrance edge but one has
355 a known execution count, then set the count of the remaining edge
357 As edge counts are set, decrement the succ/pred count, but don't delete
358 the edge, that way we can easily tell when all edges are known, or only
359 one edge is unknown. */
361 /* The order that the basic blocks are iterated through is important.
362 Since the code that finds spanning trees starts with block 0, low numbered
363 edges are put on the spanning tree in preference to high numbered edges.
364 Hence, most instrumented edges are at the end. Graph solving works much
365 faster if we propagate numbers from the end to the start.
367 This takes an average of slightly more than 3 passes. */
375 FOR_BB_BETWEEN (bb, EXIT_BLOCK_PTR, NULL, prev_bb)
377 struct bb_info *bi = BB_INFO (bb);
378 if (! bi->count_valid)
380 if (bi->succ_count == 0)
385 for (e = bb->succ; e; e = e->succ_next)
391 else if (bi->pred_count == 0)
396 for (e = bb->pred; e; e = e->pred_next)
405 if (bi->succ_count == 1)
410 /* One of the counts will be invalid, but it is zero,
411 so adding it in also doesn't hurt. */
412 for (e = bb->succ; e; e = e->succ_next)
415 /* Seedgeh for the invalid edge, and set its count. */
416 for (e = bb->succ; e; e = e->succ_next)
417 if (! EDGE_INFO (e)->count_valid && ! EDGE_INFO (e)->ignore)
420 /* Calculate count for remaining edge by conservation. */
421 total = bb->count - total;
425 EDGE_INFO (e)->count_valid = 1;
429 BB_INFO (e->dest)->pred_count--;
432 if (bi->pred_count == 1)
437 /* One of the counts will be invalid, but it is zero,
438 so adding it in also doesn't hurt. */
439 for (e = bb->pred; e; e = e->pred_next)
442 /* Search for the invalid edge, and set its count. */
443 for (e = bb->pred; e; e = e->pred_next)
444 if (!EDGE_INFO (e)->count_valid && !EDGE_INFO (e)->ignore)
447 /* Calculate count for remaining edge by conservation. */
448 total = bb->count - total + e->count;
452 EDGE_INFO (e)->count_valid = 1;
456 BB_INFO (e->src)->succ_count--;
463 dump_flow_info (dump_file);
465 total_num_passes += passes;
467 fprintf (dump_file, "Graph solving took %d passes.\n\n", passes);
469 /* If the graph has been correctly solved, every block will have a
470 succ and pred count of zero. */
473 if (BB_INFO (bb)->succ_count || BB_INFO (bb)->pred_count)
477 /* For every edge, calculate its branch probability and add a reg_note
478 to the branch insn to indicate this. */
480 for (i = 0; i < 20; i++)
482 num_never_executed = 0;
485 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
492 error ("corrupted profile info: number of iterations for basic block %d thought to be %i",
493 bb->index, (int)bb->count);
496 for (e = bb->succ; e; e = e->succ_next)
498 /* Function may return twice in the cased the called function is
499 setjmp or calls fork, but we can't represent this by extra
500 edge from the entry, since extra edge from the exit is
501 already present. We get negative frequency from the entry
504 && e->dest == EXIT_BLOCK_PTR)
505 || (e->count > bb->count
506 && e->dest != EXIT_BLOCK_PTR))
508 if (block_ends_with_call_p (bb))
509 e->count = e->count < 0 ? 0 : bb->count;
511 if (e->count < 0 || e->count > bb->count)
513 error ("corrupted profile info: number of executions for edge %d-%d thought to be %i",
514 e->src->index, e->dest->index,
516 e->count = bb->count / 2;
521 for (e = bb->succ; e; e = e->succ_next)
522 e->probability = (e->count * REG_BR_PROB_BASE + bb->count / 2) / bb->count;
524 && block_ends_with_condjump_p (bb)
525 && bb->succ->succ_next)
531 /* Find the branch edge. It is possible that we do have fake
533 for (e = bb->succ; e->flags & (EDGE_FAKE | EDGE_FALLTHRU);
535 continue; /* Loop body has been intentionally left blank. */
537 prob = e->probability;
538 index = prob * 20 / REG_BR_PROB_BASE;
542 hist_br_prob[index]++;
544 /* Do this for RTL only. */
547 note = find_reg_note (BB_END (bb), REG_BR_PROB, 0);
548 /* There may be already note put by some other pass, such
549 as builtin_expect expander. */
551 XEXP (note, 0) = GEN_INT (prob);
553 REG_NOTES (BB_END (bb))
554 = gen_rtx_EXPR_LIST (REG_BR_PROB, GEN_INT (prob),
555 REG_NOTES (BB_END (bb)));
560 /* Otherwise distribute the probabilities evenly so we get sane
561 sum. Use simple heuristics that if there are normal edges,
562 give all abnormals frequency of 0, otherwise distribute the
563 frequency over abnormals (this is the case of noreturn
569 for (e = bb->succ; e; e = e->succ_next)
570 if (!(e->flags & (EDGE_COMPLEX | EDGE_FAKE)))
574 for (e = bb->succ; e; e = e->succ_next)
575 if (!(e->flags & (EDGE_COMPLEX | EDGE_FAKE)))
576 e->probability = REG_BR_PROB_BASE / total;
582 for (e = bb->succ; e; e = e->succ_next)
584 for (e = bb->succ; e; e = e->succ_next)
585 e->probability = REG_BR_PROB_BASE / total;
588 && block_ends_with_condjump_p (bb)
589 && bb->succ->succ_next)
590 num_branches++, num_never_executed;
596 fprintf (dump_file, "%d branches\n", num_branches);
597 fprintf (dump_file, "%d branches never executed\n",
600 for (i = 0; i < 10; i++)
601 fprintf (dump_file, "%d%% branches in range %d-%d%%\n",
602 (hist_br_prob[i] + hist_br_prob[19-i]) * 100 / num_branches,
605 total_num_branches += num_branches;
606 total_num_never_executed += num_never_executed;
607 for (i = 0; i < 20; i++)
608 total_hist_br_prob[i] += hist_br_prob[i];
610 fputc ('\n', dump_file);
611 fputc ('\n', dump_file);
614 free_aux_for_blocks ();
617 /* Load value histograms values whose description is stored in VALUES array
621 compute_value_histograms (histogram_values values)
623 unsigned i, j, t, any;
624 unsigned n_histogram_counters[GCOV_N_VALUE_COUNTERS];
625 gcov_type *histogram_counts[GCOV_N_VALUE_COUNTERS];
626 gcov_type *act_count[GCOV_N_VALUE_COUNTERS];
627 gcov_type *aact_count;
628 histogram_value hist;
630 for (t = 0; t < GCOV_N_VALUE_COUNTERS; t++)
631 n_histogram_counters[t] = 0;
633 for (i = 0; i < VEC_length (histogram_value, values); i++)
635 hist = VEC_index (histogram_value, values, i);
636 n_histogram_counters[(int) hist->type] += hist->n_counters;
640 for (t = 0; t < GCOV_N_VALUE_COUNTERS; t++)
642 if (!n_histogram_counters[t])
644 histogram_counts[t] = NULL;
648 histogram_counts[t] =
649 get_coverage_counts (COUNTER_FOR_HIST_TYPE (t),
650 n_histogram_counters[t], NULL);
651 if (histogram_counts[t])
653 act_count[t] = histogram_counts[t];
658 for (i = 0; i < VEC_length (histogram_value, values); i++)
660 rtx hist_list = NULL_RTX;
662 hist = VEC_index (histogram_value, values, i);
663 t = (int) hist->type;
665 /* FIXME: make this work for trees. */
668 aact_count = act_count[t];
669 act_count[t] += hist->n_counters;
670 for (j = hist->n_counters; j > 0; j--)
671 hist_list = alloc_EXPR_LIST (0, GEN_INT (aact_count[j - 1]),
673 hist_list = alloc_EXPR_LIST (0,
674 copy_rtx ((rtx) hist->value), hist_list);
675 hist_list = alloc_EXPR_LIST (0, GEN_INT (hist->type), hist_list);
676 REG_NOTES ((rtx) hist->insn) =
677 alloc_EXPR_LIST (REG_VALUE_PROFILE, hist_list,
678 REG_NOTES ((rtx) hist->insn));
682 for (t = 0; t < GCOV_N_VALUE_COUNTERS; t++)
683 if (histogram_counts[t])
684 free (histogram_counts[t]);
687 /* Instrument and/or analyze program behavior based on program flow graph.
688 In either case, this function builds a flow graph for the function being
689 compiled. The flow graph is stored in BB_GRAPH.
691 When FLAG_PROFILE_ARCS is nonzero, this function instruments the edges in
692 the flow graph that are needed to reconstruct the dynamic behavior of the
695 When FLAG_BRANCH_PROBABILITIES is nonzero, this function reads auxiliary
696 information from a data file containing edge count information from previous
697 executions of the function being compiled. In this case, the flow graph is
698 annotated with actual execution counts, which are later propagated into the
699 rtl for optimization purposes.
701 Main entry point of this file. */
708 unsigned num_edges, ignored_edges;
709 unsigned num_instrumented;
710 struct edge_list *el;
711 histogram_values values = NULL;
713 total_num_times_called++;
715 flow_call_edges_add (NULL);
716 add_noreturn_fake_exit_edges ();
718 /* We can't handle cyclic regions constructed using abnormal edges.
719 To avoid these we replace every source of abnormal edge by a fake
720 edge from entry node and every destination by fake edge to exit.
721 This keeps graph acyclic and our calculation exact for all normal
722 edges except for exit and entrance ones.
724 We also add fake exit edges for each call and asm statement in the
725 basic, since it may not return. */
729 int need_exit_edge = 0, need_entry_edge = 0;
730 int have_exit_edge = 0, have_entry_edge = 0;
733 /* Functions returning multiple times are not handled by extra edges.
734 Instead we simply allow negative counts on edges from exit to the
735 block past call and corresponding probabilities. We can't go
736 with the extra edges because that would result in flowgraph that
737 needs to have fake edges outside the spanning tree. */
739 for (e = bb->succ; e; e = e->succ_next)
741 if ((e->flags & (EDGE_ABNORMAL | EDGE_ABNORMAL_CALL))
742 && e->dest != EXIT_BLOCK_PTR)
744 if (e->dest == EXIT_BLOCK_PTR)
747 for (e = bb->pred; e; e = e->pred_next)
749 if ((e->flags & (EDGE_ABNORMAL | EDGE_ABNORMAL_CALL))
750 && e->src != ENTRY_BLOCK_PTR)
752 if (e->src == ENTRY_BLOCK_PTR)
756 if (need_exit_edge && !have_exit_edge)
759 fprintf (dump_file, "Adding fake exit edge to bb %i\n",
761 make_edge (bb, EXIT_BLOCK_PTR, EDGE_FAKE);
763 if (need_entry_edge && !have_entry_edge)
766 fprintf (dump_file, "Adding fake entry edge to bb %i\n",
768 make_edge (ENTRY_BLOCK_PTR, bb, EDGE_FAKE);
772 el = create_edge_list ();
773 num_edges = NUM_EDGES (el);
774 alloc_aux_for_edges (sizeof (struct edge_info));
776 /* The basic blocks are expected to be numbered sequentially. */
780 for (i = 0 ; i < num_edges ; i++)
782 edge e = INDEX_EDGE (el, i);
785 /* Mark edges we've replaced by fake edges above as ignored. */
786 if ((e->flags & (EDGE_ABNORMAL | EDGE_ABNORMAL_CALL))
787 && e->src != ENTRY_BLOCK_PTR && e->dest != EXIT_BLOCK_PTR)
789 EDGE_INFO (e)->ignore = 1;
794 #ifdef ENABLE_CHECKING
798 /* Create spanning tree from basic block graph, mark each edge that is
799 on the spanning tree. We insert as many abnormal and critical edges
800 as possible to minimize number of edge splits necessary. */
802 find_spanning_tree (el);
804 /* Fake edges that are not on the tree will not be instrumented, so
805 mark them ignored. */
806 for (num_instrumented = i = 0; i < num_edges; i++)
808 edge e = INDEX_EDGE (el, i);
809 struct edge_info *inf = EDGE_INFO (e);
811 if (inf->ignore || inf->on_tree)
813 else if (e->flags & EDGE_FAKE)
822 total_num_blocks += n_basic_blocks + 2;
824 fprintf (dump_file, "%d basic blocks\n", n_basic_blocks);
826 total_num_edges += num_edges;
828 fprintf (dump_file, "%d edges\n", num_edges);
830 total_num_edges_ignored += ignored_edges;
832 fprintf (dump_file, "%d ignored edges\n", ignored_edges);
834 /* Write the data from which gcov can reconstruct the basic block
837 /* Basic block flags */
838 if (coverage_begin_output ())
840 gcov_position_t offset;
842 offset = gcov_write_tag (GCOV_TAG_BLOCKS);
843 for (i = 0; i != (unsigned) (n_basic_blocks + 2); i++)
844 gcov_write_unsigned (0);
845 gcov_write_length (offset);
848 /* Keep all basic block indexes nonnegative in the gcov output.
849 Index 0 is used for entry block, last index is for exit block.
851 ENTRY_BLOCK_PTR->index = -1;
852 EXIT_BLOCK_PTR->index = last_basic_block;
853 #define BB_TO_GCOV_INDEX(bb) ((bb)->index + 1)
856 if (coverage_begin_output ())
858 gcov_position_t offset;
860 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR, next_bb)
864 offset = gcov_write_tag (GCOV_TAG_ARCS);
865 gcov_write_unsigned (BB_TO_GCOV_INDEX (bb));
867 for (e = bb->succ; e; e = e->succ_next)
869 struct edge_info *i = EDGE_INFO (e);
872 unsigned flag_bits = 0;
875 flag_bits |= GCOV_ARC_ON_TREE;
876 if (e->flags & EDGE_FAKE)
877 flag_bits |= GCOV_ARC_FAKE;
878 if (e->flags & EDGE_FALLTHRU)
879 flag_bits |= GCOV_ARC_FALLTHROUGH;
881 gcov_write_unsigned (BB_TO_GCOV_INDEX (e->dest));
882 gcov_write_unsigned (flag_bits);
886 gcov_write_length (offset);
891 /* FIXME: make this work for trees. (Line numbers are in location_t
892 objects, but aren't always attached to the obvious tree...) */
893 if (coverage_begin_output () && !ir_type ())
895 char const *prev_file_name = NULL;
896 gcov_position_t offset;
900 rtx insn = BB_HEAD (bb);
901 int ignore_next_note = 0;
905 /* We are looking for line number notes. Search backward
906 before basic block to find correct ones. */
907 insn = prev_nonnote_insn (insn);
911 insn = NEXT_INSN (insn);
913 while (insn != BB_END (bb))
917 /* Must ignore the line number notes that
918 immediately follow the end of an inline function
919 to avoid counting it twice. There is a note
920 before the call, and one after the call. */
921 if (NOTE_LINE_NUMBER (insn)
922 == NOTE_INSN_REPEATED_LINE_NUMBER)
923 ignore_next_note = 1;
924 else if (NOTE_LINE_NUMBER (insn) <= 0)
926 else if (ignore_next_note)
927 ignore_next_note = 0;
934 offset = gcov_write_tag (GCOV_TAG_LINES);
935 gcov_write_unsigned (BB_TO_GCOV_INDEX (bb));
938 NOTE_EXPANDED_LOCATION (s, insn);
940 /* If this is a new source file, then output the
941 file's name to the .bb file. */
943 || strcmp (s.file, prev_file_name))
945 prev_file_name = s.file;
946 gcov_write_unsigned (0);
947 gcov_write_string (prev_file_name);
949 gcov_write_unsigned (s.line);
952 insn = NEXT_INSN (insn);
957 /* A file of NULL indicates the end of run. */
958 gcov_write_unsigned (0);
959 gcov_write_string (NULL);
960 gcov_write_length (offset);
965 ENTRY_BLOCK_PTR->index = ENTRY_BLOCK;
966 EXIT_BLOCK_PTR->index = EXIT_BLOCK;
967 #undef BB_TO_GCOV_INDEX
969 if (flag_profile_values)
970 find_values_to_profile (&values);
972 if (flag_branch_probabilities)
974 compute_branch_probabilities ();
975 if (flag_profile_values)
976 compute_value_histograms (values);
979 remove_fake_edges ();
981 /* For each edge not on the spanning tree, add counting code. */
983 && coverage_counter_alloc (GCOV_COUNTER_ARCS, num_instrumented))
985 unsigned n_instrumented = instrument_edges (el);
987 if (n_instrumented != num_instrumented)
990 if (flag_profile_values)
991 instrument_values (values);
993 /* Commit changes done by instrumentation. */
995 bsi_commit_edge_inserts ((int *)NULL);
998 commit_edge_insertions_watch_calls ();
999 allocate_reg_info (max_reg_num (), FALSE, FALSE);
1003 free_aux_for_edges ();
1007 /* Re-merge split basic blocks and the mess introduced by
1008 insert_insn_on_edge. */
1009 cleanup_cfg (profile_arc_flag ? CLEANUP_EXPENSIVE : 0);
1010 if (profile_dump_file())
1011 dump_flow_info (profile_dump_file());
1014 free_edge_list (el);
1015 profile_status = PROFILE_READ;
1018 /* Union find algorithm implementation for the basic blocks using
1022 find_group (basic_block bb)
1024 basic_block group = bb, bb1;
1026 while ((basic_block) group->aux != group)
1027 group = (basic_block) group->aux;
1029 /* Compress path. */
1030 while ((basic_block) bb->aux != group)
1032 bb1 = (basic_block) bb->aux;
1033 bb->aux = (void *) group;
1040 union_groups (basic_block bb1, basic_block bb2)
1042 basic_block bb1g = find_group (bb1);
1043 basic_block bb2g = find_group (bb2);
1045 /* ??? I don't have a place for the rank field. OK. Lets go w/o it,
1046 this code is unlikely going to be performance problem anyway. */
1053 /* This function searches all of the edges in the program flow graph, and puts
1054 as many bad edges as possible onto the spanning tree. Bad edges include
1055 abnormals edges, which can't be instrumented at the moment. Since it is
1056 possible for fake edges to form a cycle, we will have to develop some
1057 better way in the future. Also put critical edges to the tree, since they
1058 are more expensive to instrument. */
1061 find_spanning_tree (struct edge_list *el)
1064 int num_edges = NUM_EDGES (el);
1067 /* We use aux field for standard union-find algorithm. */
1068 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
1071 /* Add fake edge exit to entry we can't instrument. */
1072 union_groups (EXIT_BLOCK_PTR, ENTRY_BLOCK_PTR);
1074 /* First add all abnormal edges to the tree unless they form a cycle. Also
1075 add all edges to EXIT_BLOCK_PTR to avoid inserting profiling code behind
1076 setting return value from function. */
1077 for (i = 0; i < num_edges; i++)
1079 edge e = INDEX_EDGE (el, i);
1080 if (((e->flags & (EDGE_ABNORMAL | EDGE_ABNORMAL_CALL | EDGE_FAKE))
1081 || e->dest == EXIT_BLOCK_PTR)
1082 && !EDGE_INFO (e)->ignore
1083 && (find_group (e->src) != find_group (e->dest)))
1086 fprintf (dump_file, "Abnormal edge %d to %d put to tree\n",
1087 e->src->index, e->dest->index);
1088 EDGE_INFO (e)->on_tree = 1;
1089 union_groups (e->src, e->dest);
1093 /* Now insert all critical edges to the tree unless they form a cycle. */
1094 for (i = 0; i < num_edges; i++)
1096 edge e = INDEX_EDGE (el, i);
1097 if (EDGE_CRITICAL_P (e) && !EDGE_INFO (e)->ignore
1098 && find_group (e->src) != find_group (e->dest))
1101 fprintf (dump_file, "Critical edge %d to %d put to tree\n",
1102 e->src->index, e->dest->index);
1103 EDGE_INFO (e)->on_tree = 1;
1104 union_groups (e->src, e->dest);
1108 /* And now the rest. */
1109 for (i = 0; i < num_edges; i++)
1111 edge e = INDEX_EDGE (el, i);
1112 if (!EDGE_INFO (e)->ignore
1113 && find_group (e->src) != find_group (e->dest))
1116 fprintf (dump_file, "Normal edge %d to %d put to tree\n",
1117 e->src->index, e->dest->index);
1118 EDGE_INFO (e)->on_tree = 1;
1119 union_groups (e->src, e->dest);
1123 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
1127 /* Perform file-level initialization for branch-prob processing. */
1130 init_branch_prob (void)
1134 total_num_blocks = 0;
1135 total_num_edges = 0;
1136 total_num_edges_ignored = 0;
1137 total_num_edges_instrumented = 0;
1138 total_num_blocks_created = 0;
1139 total_num_passes = 0;
1140 total_num_times_called = 0;
1141 total_num_branches = 0;
1142 total_num_never_executed = 0;
1143 for (i = 0; i < 20; i++)
1144 total_hist_br_prob[i] = 0;
1147 /* Performs file-level cleanup after branch-prob processing
1151 end_branch_prob (void)
1155 fprintf (dump_file, "\n");
1156 fprintf (dump_file, "Total number of blocks: %d\n",
1158 fprintf (dump_file, "Total number of edges: %d\n", total_num_edges);
1159 fprintf (dump_file, "Total number of ignored edges: %d\n",
1160 total_num_edges_ignored);
1161 fprintf (dump_file, "Total number of instrumented edges: %d\n",
1162 total_num_edges_instrumented);
1163 fprintf (dump_file, "Total number of blocks created: %d\n",
1164 total_num_blocks_created);
1165 fprintf (dump_file, "Total number of graph solution passes: %d\n",
1167 if (total_num_times_called != 0)
1168 fprintf (dump_file, "Average number of graph solution passes: %d\n",
1169 (total_num_passes + (total_num_times_called >> 1))
1170 / total_num_times_called);
1171 fprintf (dump_file, "Total number of branches: %d\n",
1172 total_num_branches);
1173 fprintf (dump_file, "Total number of branches never executed: %d\n",
1174 total_num_never_executed);
1175 if (total_num_branches)
1179 for (i = 0; i < 10; i++)
1180 fprintf (dump_file, "%d%% branches in range %d-%d%%\n",
1181 (total_hist_br_prob[i] + total_hist_br_prob[19-i]) * 100
1182 / total_num_branches, 5*i, 5*i+5);
1187 /* Set up hooks to enable tree-based profiling. */
1190 tree_register_profile_hooks (void)
1192 profile_hooks = &tree_profile_hooks;
1197 /* Set up hooks to enable RTL-based profiling. */
1200 rtl_register_profile_hooks (void)
1202 profile_hooks = &rtl_profile_hooks;