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, 2005, 2007, 2008, 2009, 2010
4 Free Software Foundation, Inc.
5 Contributed by James E. Wilson, UC Berkeley/Cygnus Support;
6 based on some ideas from Dain Samples of UC Berkeley.
7 Further mangling by Bob Manson, Cygnus Support.
9 This file is part of GCC.
11 GCC is free software; you can redistribute it and/or modify it under
12 the terms of the GNU General Public License as published by the Free
13 Software Foundation; either version 3, or (at your option) any later
16 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
17 WARRANTY; without even the implied warranty of MERCHANTABILITY or
18 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
21 You should have received a copy of the GNU General Public License
22 along with GCC; see the file COPYING3. If not see
23 <http://www.gnu.org/licenses/>. */
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"
62 #include "basic-block.h"
63 #include "diagnostic-core.h"
65 #include "value-prof.h"
68 #include "tree-flow.h"
71 #include "tree-pass.h"
76 unsigned int count_valid : 1;
78 /* Number of successor and predecessor edges. */
83 #define BB_INFO(b) ((struct bb_info *) (b)->aux)
86 /* Counter summary from the last set of coverage counts read. */
88 const struct gcov_ctr_summary *profile_info;
90 /* Collect statistics on the performance of this pass for the entire source
93 static int total_num_blocks;
94 static int total_num_edges;
95 static int total_num_edges_ignored;
96 static int total_num_edges_instrumented;
97 static int total_num_blocks_created;
98 static int total_num_passes;
99 static int total_num_times_called;
100 static int total_hist_br_prob[20];
101 static int total_num_branches;
103 /* Forward declarations. */
104 static void find_spanning_tree (struct edge_list *);
105 static unsigned instrument_edges (struct edge_list *);
106 static void instrument_values (histogram_values);
107 static void compute_branch_probabilities (void);
108 static void compute_value_histograms (histogram_values);
109 static gcov_type * get_exec_counts (void);
110 static basic_block find_group (basic_block);
111 static void union_groups (basic_block, basic_block);
113 /* Add edge instrumentation code to the entire insn chain.
115 F is the first insn of the chain.
116 NUM_BLOCKS is the number of basic blocks found in F. */
119 instrument_edges (struct edge_list *el)
121 unsigned num_instr_edges = 0;
122 int num_edges = NUM_EDGES (el);
125 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
130 FOR_EACH_EDGE (e, ei, bb->succs)
132 struct edge_info *inf = EDGE_INFO (e);
134 if (!inf->ignore && !inf->on_tree)
136 gcc_assert (!(e->flags & EDGE_ABNORMAL));
138 fprintf (dump_file, "Edge %d to %d instrumented%s\n",
139 e->src->index, e->dest->index,
140 EDGE_CRITICAL_P (e) ? " (and split)" : "");
141 gimple_gen_edge_profiler (num_instr_edges++, e);
146 total_num_blocks_created += num_edges;
148 fprintf (dump_file, "%d edges instrumented\n", num_instr_edges);
149 return num_instr_edges;
152 /* Add code to measure histograms for values in list VALUES. */
154 instrument_values (histogram_values values)
158 /* Emit code to generate the histograms before the insns. */
160 for (i = 0; i < VEC_length (histogram_value, values); i++)
162 histogram_value hist = VEC_index (histogram_value, values, i);
165 case HIST_TYPE_INTERVAL:
166 t = GCOV_COUNTER_V_INTERVAL;
170 t = GCOV_COUNTER_V_POW2;
173 case HIST_TYPE_SINGLE_VALUE:
174 t = GCOV_COUNTER_V_SINGLE;
177 case HIST_TYPE_CONST_DELTA:
178 t = GCOV_COUNTER_V_DELTA;
181 case HIST_TYPE_INDIR_CALL:
182 t = GCOV_COUNTER_V_INDIR;
185 case HIST_TYPE_AVERAGE:
186 t = GCOV_COUNTER_AVERAGE;
190 t = GCOV_COUNTER_IOR;
196 if (!coverage_counter_alloc (t, hist->n_counters))
201 case HIST_TYPE_INTERVAL:
202 gimple_gen_interval_profiler (hist, t, 0);
206 gimple_gen_pow2_profiler (hist, t, 0);
209 case HIST_TYPE_SINGLE_VALUE:
210 gimple_gen_one_value_profiler (hist, t, 0);
213 case HIST_TYPE_CONST_DELTA:
214 gimple_gen_const_delta_profiler (hist, t, 0);
217 case HIST_TYPE_INDIR_CALL:
218 gimple_gen_ic_profiler (hist, t, 0);
221 case HIST_TYPE_AVERAGE:
222 gimple_gen_average_profiler (hist, t, 0);
226 gimple_gen_ior_profiler (hist, t, 0);
236 /* Computes hybrid profile for all matching entries in da_file. */
239 get_exec_counts (void)
241 unsigned num_edges = 0;
245 /* Count the edges to be (possibly) instrumented. */
246 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
251 FOR_EACH_EDGE (e, ei, bb->succs)
252 if (!EDGE_INFO (e)->ignore && !EDGE_INFO (e)->on_tree)
256 counts = get_coverage_counts (GCOV_COUNTER_ARCS, num_edges, &profile_info);
260 if (dump_file && profile_info)
261 fprintf(dump_file, "Merged %u profiles with maximal count %u.\n",
262 profile_info->runs, (unsigned) profile_info->sum_max);
269 is_edge_inconsistent (VEC(edge,gc) *edges)
273 FOR_EACH_EDGE (e, ei, edges)
275 if (!EDGE_INFO (e)->ignore)
278 && (!(e->flags & EDGE_FAKE)
279 || !block_ends_with_call_p (e->src)))
284 "Edge %i->%i is inconsistent, count"HOST_WIDEST_INT_PRINT_DEC,
285 e->src->index, e->dest->index, e->count);
286 dump_bb (e->src, dump_file, 0);
287 dump_bb (e->dest, dump_file, 0);
297 correct_negative_edge_counts (void)
303 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
305 FOR_EACH_EDGE (e, ei, bb->succs)
313 /* Check consistency.
314 Return true if inconsistency is found. */
316 is_inconsistent (void)
319 bool inconsistent = false;
322 inconsistent |= is_edge_inconsistent (bb->preds);
323 if (!dump_file && inconsistent)
325 inconsistent |= is_edge_inconsistent (bb->succs);
326 if (!dump_file && inconsistent)
332 fprintf (dump_file, "BB %i count is negative "
333 HOST_WIDEST_INT_PRINT_DEC,
336 dump_bb (bb, dump_file, 0);
340 if (bb->count != sum_edge_counts (bb->preds))
344 fprintf (dump_file, "BB %i count does not match sum of incoming edges "
345 HOST_WIDEST_INT_PRINT_DEC" should be " HOST_WIDEST_INT_PRINT_DEC,
348 sum_edge_counts (bb->preds));
349 dump_bb (bb, dump_file, 0);
353 if (bb->count != sum_edge_counts (bb->succs) &&
354 ! (find_edge (bb, EXIT_BLOCK_PTR) != NULL && block_ends_with_call_p (bb)))
358 fprintf (dump_file, "BB %i count does not match sum of outgoing edges "
359 HOST_WIDEST_INT_PRINT_DEC" should be " HOST_WIDEST_INT_PRINT_DEC,
362 sum_edge_counts (bb->succs));
363 dump_bb (bb, dump_file, 0);
367 if (!dump_file && inconsistent)
374 /* Set each basic block count to the sum of its outgoing edge counts */
379 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
381 bb->count = sum_edge_counts (bb->succs);
382 gcc_assert (bb->count >= 0);
386 /* Reads profile data and returns total number of edge counts read */
388 read_profile_edge_counts (gcov_type *exec_counts)
392 int exec_counts_pos = 0;
393 /* For each edge not on the spanning tree, set its execution count from
395 /* The first count in the .da file is the number of times that the function
396 was entered. This is the exec_count for block zero. */
398 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
403 FOR_EACH_EDGE (e, ei, bb->succs)
404 if (!EDGE_INFO (e)->ignore && !EDGE_INFO (e)->on_tree)
409 e->count = exec_counts[exec_counts_pos++];
410 if (e->count > profile_info->sum_max)
412 error ("corrupted profile info: edge from %i to %i exceeds maximal count",
413 bb->index, e->dest->index);
419 EDGE_INFO (e)->count_valid = 1;
420 BB_INFO (bb)->succ_count--;
421 BB_INFO (e->dest)->pred_count--;
424 fprintf (dump_file, "\nRead edge from %i to %i, count:",
425 bb->index, e->dest->index);
426 fprintf (dump_file, HOST_WIDEST_INT_PRINT_DEC,
427 (HOST_WIDEST_INT) e->count);
435 /* Compute the branch probabilities for the various branches.
436 Annotate them accordingly. */
439 compute_branch_probabilities (void)
446 int hist_br_prob[20];
448 gcov_type *exec_counts = get_exec_counts ();
449 int inconsistent = 0;
451 /* Very simple sanity checks so we catch bugs in our profiling code. */
454 if (profile_info->run_max * profile_info->runs < profile_info->sum_max)
456 error ("corrupted profile info: run_max * runs < sum_max");
460 if (profile_info->sum_all < profile_info->sum_max)
462 error ("corrupted profile info: sum_all is smaller than sum_max");
466 /* Attach extra info block to each bb. */
467 alloc_aux_for_blocks (sizeof (struct bb_info));
468 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
473 FOR_EACH_EDGE (e, ei, bb->succs)
474 if (!EDGE_INFO (e)->ignore)
475 BB_INFO (bb)->succ_count++;
476 FOR_EACH_EDGE (e, ei, bb->preds)
477 if (!EDGE_INFO (e)->ignore)
478 BB_INFO (bb)->pred_count++;
481 /* Avoid predicting entry on exit nodes. */
482 BB_INFO (EXIT_BLOCK_PTR)->succ_count = 2;
483 BB_INFO (ENTRY_BLOCK_PTR)->pred_count = 2;
485 num_edges = read_profile_edge_counts (exec_counts);
488 fprintf (dump_file, "\n%d edge counts read\n", num_edges);
490 /* For every block in the file,
491 - if every exit/entrance edge has a known count, then set the block count
492 - if the block count is known, and every exit/entrance edge but one has
493 a known execution count, then set the count of the remaining edge
495 As edge counts are set, decrement the succ/pred count, but don't delete
496 the edge, that way we can easily tell when all edges are known, or only
497 one edge is unknown. */
499 /* The order that the basic blocks are iterated through is important.
500 Since the code that finds spanning trees starts with block 0, low numbered
501 edges are put on the spanning tree in preference to high numbered edges.
502 Hence, most instrumented edges are at the end. Graph solving works much
503 faster if we propagate numbers from the end to the start.
505 This takes an average of slightly more than 3 passes. */
513 FOR_BB_BETWEEN (bb, EXIT_BLOCK_PTR, NULL, prev_bb)
515 struct bb_info *bi = BB_INFO (bb);
516 if (! bi->count_valid)
518 if (bi->succ_count == 0)
524 FOR_EACH_EDGE (e, ei, bb->succs)
530 else if (bi->pred_count == 0)
536 FOR_EACH_EDGE (e, ei, bb->preds)
545 if (bi->succ_count == 1)
551 /* One of the counts will be invalid, but it is zero,
552 so adding it in also doesn't hurt. */
553 FOR_EACH_EDGE (e, ei, bb->succs)
556 /* Search for the invalid edge, and set its count. */
557 FOR_EACH_EDGE (e, ei, bb->succs)
558 if (! EDGE_INFO (e)->count_valid && ! EDGE_INFO (e)->ignore)
561 /* Calculate count for remaining edge by conservation. */
562 total = bb->count - total;
565 EDGE_INFO (e)->count_valid = 1;
569 BB_INFO (e->dest)->pred_count--;
572 if (bi->pred_count == 1)
578 /* One of the counts will be invalid, but it is zero,
579 so adding it in also doesn't hurt. */
580 FOR_EACH_EDGE (e, ei, bb->preds)
583 /* Search for the invalid edge, and set its count. */
584 FOR_EACH_EDGE (e, ei, bb->preds)
585 if (!EDGE_INFO (e)->count_valid && !EDGE_INFO (e)->ignore)
588 /* Calculate count for remaining edge by conservation. */
589 total = bb->count - total + e->count;
592 EDGE_INFO (e)->count_valid = 1;
596 BB_INFO (e->src)->succ_count--;
603 dump_flow_info (dump_file, dump_flags);
605 total_num_passes += passes;
607 fprintf (dump_file, "Graph solving took %d passes.\n\n", passes);
609 /* If the graph has been correctly solved, every block will have a
610 succ and pred count of zero. */
613 gcc_assert (!BB_INFO (bb)->succ_count && !BB_INFO (bb)->pred_count);
616 /* Check for inconsistent basic block counts */
617 inconsistent = is_inconsistent ();
621 if (flag_profile_correction)
623 /* Inconsistency detected. Make it flow-consistent. */
624 static int informed = 0;
628 inform (input_location, "correcting inconsistent profile data");
630 correct_negative_edge_counts ();
631 /* Set bb counts to the sum of the outgoing edge counts */
634 fprintf (dump_file, "\nCalling mcf_smooth_cfg\n");
638 error ("corrupted profile info: profile data is not flow-consistent");
641 /* For every edge, calculate its branch probability and add a reg_note
642 to the branch insn to indicate this. */
644 for (i = 0; i < 20; i++)
648 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
655 error ("corrupted profile info: number of iterations for basic block %d thought to be %i",
656 bb->index, (int)bb->count);
659 FOR_EACH_EDGE (e, ei, bb->succs)
661 /* Function may return twice in the cased the called function is
662 setjmp or calls fork, but we can't represent this by extra
663 edge from the entry, since extra edge from the exit is
664 already present. We get negative frequency from the entry
667 && e->dest == EXIT_BLOCK_PTR)
668 || (e->count > bb->count
669 && e->dest != EXIT_BLOCK_PTR))
671 if (block_ends_with_call_p (bb))
672 e->count = e->count < 0 ? 0 : bb->count;
674 if (e->count < 0 || e->count > bb->count)
676 error ("corrupted profile info: number of executions for edge %d-%d thought to be %i",
677 e->src->index, e->dest->index,
679 e->count = bb->count / 2;
684 FOR_EACH_EDGE (e, ei, bb->succs)
685 e->probability = (e->count * REG_BR_PROB_BASE + bb->count / 2) / bb->count;
686 if (bb->index >= NUM_FIXED_BLOCKS
687 && block_ends_with_condjump_p (bb)
688 && EDGE_COUNT (bb->succs) >= 2)
694 /* Find the branch edge. It is possible that we do have fake
696 FOR_EACH_EDGE (e, ei, bb->succs)
697 if (!(e->flags & (EDGE_FAKE | EDGE_FALLTHRU)))
700 prob = e->probability;
701 index = prob * 20 / REG_BR_PROB_BASE;
705 hist_br_prob[index]++;
710 /* As a last resort, distribute the probabilities evenly.
711 Use simple heuristics that if there are normal edges,
712 give all abnormals frequency of 0, otherwise distribute the
713 frequency over abnormals (this is the case of noreturn
715 else if (profile_status == PROFILE_ABSENT)
719 FOR_EACH_EDGE (e, ei, bb->succs)
720 if (!(e->flags & (EDGE_COMPLEX | EDGE_FAKE)))
724 FOR_EACH_EDGE (e, ei, bb->succs)
725 if (!(e->flags & (EDGE_COMPLEX | EDGE_FAKE)))
726 e->probability = REG_BR_PROB_BASE / total;
732 total += EDGE_COUNT (bb->succs);
733 FOR_EACH_EDGE (e, ei, bb->succs)
734 e->probability = REG_BR_PROB_BASE / total;
736 if (bb->index >= NUM_FIXED_BLOCKS
737 && block_ends_with_condjump_p (bb)
738 && EDGE_COUNT (bb->succs) >= 2)
743 profile_status = PROFILE_READ;
747 fprintf (dump_file, "%d branches\n", num_branches);
749 for (i = 0; i < 10; i++)
750 fprintf (dump_file, "%d%% branches in range %d-%d%%\n",
751 (hist_br_prob[i] + hist_br_prob[19-i]) * 100 / num_branches,
754 total_num_branches += num_branches;
755 for (i = 0; i < 20; i++)
756 total_hist_br_prob[i] += hist_br_prob[i];
758 fputc ('\n', dump_file);
759 fputc ('\n', dump_file);
762 free_aux_for_blocks ();
765 /* Load value histograms values whose description is stored in VALUES array
769 compute_value_histograms (histogram_values values)
771 unsigned i, j, t, any;
772 unsigned n_histogram_counters[GCOV_N_VALUE_COUNTERS];
773 gcov_type *histogram_counts[GCOV_N_VALUE_COUNTERS];
774 gcov_type *act_count[GCOV_N_VALUE_COUNTERS];
775 gcov_type *aact_count;
777 for (t = 0; t < GCOV_N_VALUE_COUNTERS; t++)
778 n_histogram_counters[t] = 0;
780 for (i = 0; i < VEC_length (histogram_value, values); i++)
782 histogram_value hist = VEC_index (histogram_value, values, i);
783 n_histogram_counters[(int) hist->type] += hist->n_counters;
787 for (t = 0; t < GCOV_N_VALUE_COUNTERS; t++)
789 if (!n_histogram_counters[t])
791 histogram_counts[t] = NULL;
795 histogram_counts[t] =
796 get_coverage_counts (COUNTER_FOR_HIST_TYPE (t),
797 n_histogram_counters[t], NULL);
798 if (histogram_counts[t])
800 act_count[t] = histogram_counts[t];
805 for (i = 0; i < VEC_length (histogram_value, values); i++)
807 histogram_value hist = VEC_index (histogram_value, values, i);
808 gimple stmt = hist->hvalue.stmt;
810 t = (int) hist->type;
812 aact_count = act_count[t];
813 act_count[t] += hist->n_counters;
815 gimple_add_histogram_value (cfun, stmt, hist);
816 hist->hvalue.counters = XNEWVEC (gcov_type, hist->n_counters);
817 for (j = 0; j < hist->n_counters; j++)
818 hist->hvalue.counters[j] = aact_count[j];
821 for (t = 0; t < GCOV_N_VALUE_COUNTERS; t++)
822 if (histogram_counts[t])
823 free (histogram_counts[t]);
826 /* The entry basic block will be moved around so that it has index=1,
827 there is nothing at index 0 and the exit is at n_basic_block. */
828 #define BB_TO_GCOV_INDEX(bb) ((bb)->index - 1)
829 /* When passed NULL as file_name, initialize.
830 When passed something else, output the necessary commands to change
831 line to LINE and offset to FILE_NAME. */
833 output_location (char const *file_name, int line,
834 gcov_position_t *offset, basic_block bb)
836 static char const *prev_file_name;
837 static int prev_line;
838 bool name_differs, line_differs;
842 prev_file_name = NULL;
847 name_differs = !prev_file_name || strcmp (file_name, prev_file_name);
848 line_differs = prev_line != line;
850 if (name_differs || line_differs)
854 *offset = gcov_write_tag (GCOV_TAG_LINES);
855 gcov_write_unsigned (BB_TO_GCOV_INDEX (bb));
856 name_differs = line_differs=true;
859 /* If this is a new source file, then output the
860 file's name to the .bb file. */
863 prev_file_name = file_name;
864 gcov_write_unsigned (0);
865 gcov_write_string (prev_file_name);
869 gcov_write_unsigned (line);
875 /* Instrument and/or analyze program behavior based on program flow graph.
876 In either case, this function builds a flow graph for the function being
877 compiled. The flow graph is stored in BB_GRAPH.
879 When FLAG_PROFILE_ARCS is nonzero, this function instruments the edges in
880 the flow graph that are needed to reconstruct the dynamic behavior of the
883 When FLAG_BRANCH_PROBABILITIES is nonzero, this function reads auxiliary
884 information from a data file containing edge count information from previous
885 executions of the function being compiled. In this case, the flow graph is
886 annotated with actual execution counts, which are later propagated into the
887 rtl for optimization purposes.
889 Main entry point of this file. */
896 unsigned num_edges, ignored_edges;
897 unsigned num_instrumented;
898 struct edge_list *el;
899 histogram_values values = NULL;
901 total_num_times_called++;
903 flow_call_edges_add (NULL);
904 add_noreturn_fake_exit_edges ();
906 /* We can't handle cyclic regions constructed using abnormal edges.
907 To avoid these we replace every source of abnormal edge by a fake
908 edge from entry node and every destination by fake edge to exit.
909 This keeps graph acyclic and our calculation exact for all normal
910 edges except for exit and entrance ones.
912 We also add fake exit edges for each call and asm statement in the
913 basic, since it may not return. */
917 int need_exit_edge = 0, need_entry_edge = 0;
918 int have_exit_edge = 0, have_entry_edge = 0;
922 /* Functions returning multiple times are not handled by extra edges.
923 Instead we simply allow negative counts on edges from exit to the
924 block past call and corresponding probabilities. We can't go
925 with the extra edges because that would result in flowgraph that
926 needs to have fake edges outside the spanning tree. */
928 FOR_EACH_EDGE (e, ei, bb->succs)
930 gimple_stmt_iterator gsi;
933 /* It may happen that there are compiler generated statements
934 without a locus at all. Go through the basic block from the
935 last to the first statement looking for a locus. */
936 for (gsi = gsi_last_nondebug_bb (bb);
938 gsi_prev_nondebug (&gsi))
940 last = gsi_stmt (gsi);
941 if (gimple_has_location (last))
945 /* Edge with goto locus might get wrong coverage info unless
946 it is the only edge out of BB.
947 Don't do that when the locuses match, so
948 if (blah) goto something;
949 is not computed twice. */
951 && gimple_has_location (last)
952 && e->goto_locus != UNKNOWN_LOCATION
953 && !single_succ_p (bb)
954 && (LOCATION_FILE (e->goto_locus)
955 != LOCATION_FILE (gimple_location (last))
956 || (LOCATION_LINE (e->goto_locus)
957 != LOCATION_LINE (gimple_location (last)))))
959 basic_block new_bb = split_edge (e);
960 edge ne = single_succ_edge (new_bb);
961 ne->goto_locus = e->goto_locus;
962 ne->goto_block = e->goto_block;
964 if ((e->flags & (EDGE_ABNORMAL | EDGE_ABNORMAL_CALL))
965 && e->dest != EXIT_BLOCK_PTR)
967 if (e->dest == EXIT_BLOCK_PTR)
970 FOR_EACH_EDGE (e, ei, bb->preds)
972 if ((e->flags & (EDGE_ABNORMAL | EDGE_ABNORMAL_CALL))
973 && e->src != ENTRY_BLOCK_PTR)
975 if (e->src == ENTRY_BLOCK_PTR)
979 if (need_exit_edge && !have_exit_edge)
982 fprintf (dump_file, "Adding fake exit edge to bb %i\n",
984 make_edge (bb, EXIT_BLOCK_PTR, EDGE_FAKE);
986 if (need_entry_edge && !have_entry_edge)
989 fprintf (dump_file, "Adding fake entry edge to bb %i\n",
991 make_edge (ENTRY_BLOCK_PTR, bb, EDGE_FAKE);
995 el = create_edge_list ();
996 num_edges = NUM_EDGES (el);
997 alloc_aux_for_edges (sizeof (struct edge_info));
999 /* The basic blocks are expected to be numbered sequentially. */
1003 for (i = 0 ; i < num_edges ; i++)
1005 edge e = INDEX_EDGE (el, i);
1008 /* Mark edges we've replaced by fake edges above as ignored. */
1009 if ((e->flags & (EDGE_ABNORMAL | EDGE_ABNORMAL_CALL))
1010 && e->src != ENTRY_BLOCK_PTR && e->dest != EXIT_BLOCK_PTR)
1012 EDGE_INFO (e)->ignore = 1;
1017 /* Create spanning tree from basic block graph, mark each edge that is
1018 on the spanning tree. We insert as many abnormal and critical edges
1019 as possible to minimize number of edge splits necessary. */
1021 find_spanning_tree (el);
1023 /* Fake edges that are not on the tree will not be instrumented, so
1024 mark them ignored. */
1025 for (num_instrumented = i = 0; i < num_edges; i++)
1027 edge e = INDEX_EDGE (el, i);
1028 struct edge_info *inf = EDGE_INFO (e);
1030 if (inf->ignore || inf->on_tree)
1032 else if (e->flags & EDGE_FAKE)
1041 total_num_blocks += n_basic_blocks;
1043 fprintf (dump_file, "%d basic blocks\n", n_basic_blocks);
1045 total_num_edges += num_edges;
1047 fprintf (dump_file, "%d edges\n", num_edges);
1049 total_num_edges_ignored += ignored_edges;
1051 fprintf (dump_file, "%d ignored edges\n", ignored_edges);
1053 /* Write the data from which gcov can reconstruct the basic block
1056 /* Basic block flags */
1057 if (coverage_begin_output ())
1059 gcov_position_t offset;
1061 offset = gcov_write_tag (GCOV_TAG_BLOCKS);
1062 for (i = 0; i != (unsigned) (n_basic_blocks); i++)
1063 gcov_write_unsigned (0);
1064 gcov_write_length (offset);
1067 /* Keep all basic block indexes nonnegative in the gcov output.
1068 Index 0 is used for entry block, last index is for exit block.
1070 ENTRY_BLOCK_PTR->index = 1;
1071 EXIT_BLOCK_PTR->index = last_basic_block;
1074 if (coverage_begin_output ())
1076 gcov_position_t offset;
1078 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR, next_bb)
1083 offset = gcov_write_tag (GCOV_TAG_ARCS);
1084 gcov_write_unsigned (BB_TO_GCOV_INDEX (bb));
1086 FOR_EACH_EDGE (e, ei, bb->succs)
1088 struct edge_info *i = EDGE_INFO (e);
1091 unsigned flag_bits = 0;
1094 flag_bits |= GCOV_ARC_ON_TREE;
1095 if (e->flags & EDGE_FAKE)
1096 flag_bits |= GCOV_ARC_FAKE;
1097 if (e->flags & EDGE_FALLTHRU)
1098 flag_bits |= GCOV_ARC_FALLTHROUGH;
1099 /* On trees we don't have fallthru flags, but we can
1100 recompute them from CFG shape. */
1101 if (e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE)
1102 && e->src->next_bb == e->dest)
1103 flag_bits |= GCOV_ARC_FALLTHROUGH;
1105 gcov_write_unsigned (BB_TO_GCOV_INDEX (e->dest));
1106 gcov_write_unsigned (flag_bits);
1110 gcov_write_length (offset);
1115 if (coverage_begin_output ())
1117 gcov_position_t offset;
1119 /* Initialize the output. */
1120 output_location (NULL, 0, NULL, NULL);
1124 gimple_stmt_iterator gsi;
1128 if (bb == ENTRY_BLOCK_PTR->next_bb)
1130 expanded_location curr_location =
1131 expand_location (DECL_SOURCE_LOCATION (current_function_decl));
1132 output_location (curr_location.file, curr_location.line,
1136 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1138 gimple stmt = gsi_stmt (gsi);
1139 if (gimple_has_location (stmt))
1140 output_location (gimple_filename (stmt), gimple_lineno (stmt),
1144 /* Notice GOTO expressions we eliminated while constructing the
1146 if (single_succ_p (bb)
1147 && single_succ_edge (bb)->goto_locus != UNKNOWN_LOCATION)
1149 location_t curr_location = single_succ_edge (bb)->goto_locus;
1150 /* ??? The FILE/LINE API is inconsistent for these cases. */
1151 output_location (LOCATION_FILE (curr_location),
1152 LOCATION_LINE (curr_location), &offset, bb);
1157 /* A file of NULL indicates the end of run. */
1158 gcov_write_unsigned (0);
1159 gcov_write_string (NULL);
1160 gcov_write_length (offset);
1165 ENTRY_BLOCK_PTR->index = ENTRY_BLOCK;
1166 EXIT_BLOCK_PTR->index = EXIT_BLOCK;
1167 #undef BB_TO_GCOV_INDEX
1169 if (flag_profile_values)
1170 gimple_find_values_to_profile (&values);
1172 if (flag_branch_probabilities)
1174 compute_branch_probabilities ();
1175 if (flag_profile_values)
1176 compute_value_histograms (values);
1179 remove_fake_edges ();
1181 /* For each edge not on the spanning tree, add counting code. */
1182 if (profile_arc_flag
1183 && coverage_counter_alloc (GCOV_COUNTER_ARCS, num_instrumented))
1185 unsigned n_instrumented;
1187 gimple_init_edge_profiler ();
1189 n_instrumented = instrument_edges (el);
1191 gcc_assert (n_instrumented == num_instrumented);
1193 if (flag_profile_values)
1194 instrument_values (values);
1196 /* Commit changes done by instrumentation. */
1197 gsi_commit_edge_inserts ();
1200 free_aux_for_edges ();
1202 VEC_free (histogram_value, heap, values);
1203 free_edge_list (el);
1204 coverage_end_function ();
1207 /* Union find algorithm implementation for the basic blocks using
1211 find_group (basic_block bb)
1213 basic_block group = bb, bb1;
1215 while ((basic_block) group->aux != group)
1216 group = (basic_block) group->aux;
1218 /* Compress path. */
1219 while ((basic_block) bb->aux != group)
1221 bb1 = (basic_block) bb->aux;
1222 bb->aux = (void *) group;
1229 union_groups (basic_block bb1, basic_block bb2)
1231 basic_block bb1g = find_group (bb1);
1232 basic_block bb2g = find_group (bb2);
1234 /* ??? I don't have a place for the rank field. OK. Lets go w/o it,
1235 this code is unlikely going to be performance problem anyway. */
1236 gcc_assert (bb1g != bb2g);
1241 /* This function searches all of the edges in the program flow graph, and puts
1242 as many bad edges as possible onto the spanning tree. Bad edges include
1243 abnormals edges, which can't be instrumented at the moment. Since it is
1244 possible for fake edges to form a cycle, we will have to develop some
1245 better way in the future. Also put critical edges to the tree, since they
1246 are more expensive to instrument. */
1249 find_spanning_tree (struct edge_list *el)
1252 int num_edges = NUM_EDGES (el);
1255 /* We use aux field for standard union-find algorithm. */
1256 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
1259 /* Add fake edge exit to entry we can't instrument. */
1260 union_groups (EXIT_BLOCK_PTR, ENTRY_BLOCK_PTR);
1262 /* First add all abnormal edges to the tree unless they form a cycle. Also
1263 add all edges to EXIT_BLOCK_PTR to avoid inserting profiling code behind
1264 setting return value from function. */
1265 for (i = 0; i < num_edges; i++)
1267 edge e = INDEX_EDGE (el, i);
1268 if (((e->flags & (EDGE_ABNORMAL | EDGE_ABNORMAL_CALL | EDGE_FAKE))
1269 || e->dest == EXIT_BLOCK_PTR)
1270 && !EDGE_INFO (e)->ignore
1271 && (find_group (e->src) != find_group (e->dest)))
1274 fprintf (dump_file, "Abnormal edge %d to %d put to tree\n",
1275 e->src->index, e->dest->index);
1276 EDGE_INFO (e)->on_tree = 1;
1277 union_groups (e->src, e->dest);
1281 /* Now insert all critical edges to the tree unless they form a cycle. */
1282 for (i = 0; i < num_edges; i++)
1284 edge e = INDEX_EDGE (el, i);
1285 if (EDGE_CRITICAL_P (e) && !EDGE_INFO (e)->ignore
1286 && find_group (e->src) != find_group (e->dest))
1289 fprintf (dump_file, "Critical edge %d to %d put to tree\n",
1290 e->src->index, e->dest->index);
1291 EDGE_INFO (e)->on_tree = 1;
1292 union_groups (e->src, e->dest);
1296 /* And now the rest. */
1297 for (i = 0; i < num_edges; i++)
1299 edge e = INDEX_EDGE (el, i);
1300 if (!EDGE_INFO (e)->ignore
1301 && find_group (e->src) != find_group (e->dest))
1304 fprintf (dump_file, "Normal edge %d to %d put to tree\n",
1305 e->src->index, e->dest->index);
1306 EDGE_INFO (e)->on_tree = 1;
1307 union_groups (e->src, e->dest);
1311 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
1315 /* Perform file-level initialization for branch-prob processing. */
1318 init_branch_prob (void)
1322 total_num_blocks = 0;
1323 total_num_edges = 0;
1324 total_num_edges_ignored = 0;
1325 total_num_edges_instrumented = 0;
1326 total_num_blocks_created = 0;
1327 total_num_passes = 0;
1328 total_num_times_called = 0;
1329 total_num_branches = 0;
1330 for (i = 0; i < 20; i++)
1331 total_hist_br_prob[i] = 0;
1334 /* Performs file-level cleanup after branch-prob processing
1338 end_branch_prob (void)
1342 fprintf (dump_file, "\n");
1343 fprintf (dump_file, "Total number of blocks: %d\n",
1345 fprintf (dump_file, "Total number of edges: %d\n", total_num_edges);
1346 fprintf (dump_file, "Total number of ignored edges: %d\n",
1347 total_num_edges_ignored);
1348 fprintf (dump_file, "Total number of instrumented edges: %d\n",
1349 total_num_edges_instrumented);
1350 fprintf (dump_file, "Total number of blocks created: %d\n",
1351 total_num_blocks_created);
1352 fprintf (dump_file, "Total number of graph solution passes: %d\n",
1354 if (total_num_times_called != 0)
1355 fprintf (dump_file, "Average number of graph solution passes: %d\n",
1356 (total_num_passes + (total_num_times_called >> 1))
1357 / total_num_times_called);
1358 fprintf (dump_file, "Total number of branches: %d\n",
1359 total_num_branches);
1360 if (total_num_branches)
1364 for (i = 0; i < 10; i++)
1365 fprintf (dump_file, "%d%% branches in range %d-%d%%\n",
1366 (total_hist_br_prob[i] + total_hist_br_prob[19-i]) * 100
1367 / total_num_branches, 5*i, 5*i+5);