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 file generated is <dumpbase>.bbg. The format is
43 described in full in gcov-io.h. */
45 /* ??? Register allocation should use basic block execution counts to
46 give preference to the most commonly executed blocks. */
48 /* ??? Should calculate branch probabilities before instrumenting code, since
49 then we can use arc counts to help decide which arcs to instrument. */
53 #include "coretypes.h"
63 #include "value-prof.h"
66 /* Additional information about the edges we need. */
68 unsigned int count_valid
: 1;
70 /* Is on the spanning tree. */
71 unsigned int on_tree
: 1;
73 /* Pretend this edge does not exist (it is abnormal and we've
74 inserted a fake to compensate). */
75 unsigned int ignore
: 1;
79 unsigned int count_valid
: 1;
81 /* Number of successor and predecessor edges. */
86 #define EDGE_INFO(e) ((struct edge_info *) (e)->aux)
87 #define BB_INFO(b) ((struct bb_info *) (b)->aux)
89 /* Counter summary from the last set of coverage counts read. */
91 const struct gcov_ctr_summary
*profile_info
;
93 /* Collect statistics on the performance of this pass for the entire source
96 static int total_num_blocks
;
97 static int total_num_edges
;
98 static int total_num_edges_ignored
;
99 static int total_num_edges_instrumented
;
100 static int total_num_blocks_created
;
101 static int total_num_passes
;
102 static int total_num_times_called
;
103 static int total_hist_br_prob
[20];
104 static int total_num_never_executed
;
105 static int total_num_branches
;
107 /* Forward declarations. */
108 static void find_spanning_tree (struct edge_list
*);
109 static rtx
gen_edge_profiler (int);
110 static rtx
gen_interval_profiler (struct histogram_value
*, unsigned,
112 static rtx
gen_pow2_profiler (struct histogram_value
*, unsigned, unsigned);
113 static rtx
gen_one_value_profiler (struct histogram_value
*, unsigned,
115 static rtx
gen_const_delta_profiler (struct histogram_value
*, unsigned,
117 static unsigned instrument_edges (struct edge_list
*);
118 static void instrument_values (unsigned, struct histogram_value
*);
119 static void compute_branch_probabilities (void);
120 static void compute_value_histograms (unsigned, struct histogram_value
*);
121 static gcov_type
* get_exec_counts (void);
122 static basic_block
find_group (basic_block
);
123 static void union_groups (basic_block
, basic_block
);
126 /* Add edge instrumentation code to the entire insn chain.
128 F is the first insn of the chain.
129 NUM_BLOCKS is the number of basic blocks found in F. */
132 instrument_edges (struct edge_list
*el
)
134 unsigned num_instr_edges
= 0;
135 int num_edges
= NUM_EDGES (el
);
138 remove_fake_edges ();
140 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
, NULL
, next_bb
)
144 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
146 struct edge_info
*inf
= EDGE_INFO (e
);
148 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 edge_profile
= gen_edge_profiler (num_instr_edges
++);
159 insert_insn_on_edge (edge_profile
, e
);
160 rebuild_jump_labels (e
->insns
);
165 total_num_blocks_created
+= num_edges
;
167 fprintf (dump_file
, "%d edges instrumented\n", num_instr_edges
);
168 return num_instr_edges
;
171 /* Add code to measure histograms list of VALUES of length N_VALUES. */
173 instrument_values (unsigned n_values
, struct histogram_value
*values
)
179 /* Emit code to generate the histograms before the insns. */
181 for (i
= 0; i
< n_values
; i
++)
183 e
= split_block (BLOCK_FOR_INSN (values
[i
].insn
),
184 PREV_INSN (values
[i
].insn
));
185 switch (values
[i
].type
)
187 case HIST_TYPE_INTERVAL
:
188 t
= GCOV_COUNTER_V_INTERVAL
;
192 t
= GCOV_COUNTER_V_POW2
;
195 case HIST_TYPE_SINGLE_VALUE
:
196 t
= GCOV_COUNTER_V_SINGLE
;
199 case HIST_TYPE_CONST_DELTA
:
200 t
= GCOV_COUNTER_V_DELTA
;
206 if (!coverage_counter_alloc (t
, values
[i
].n_counters
))
209 switch (values
[i
].type
)
211 case HIST_TYPE_INTERVAL
:
212 sequence
= gen_interval_profiler (values
+ i
, t
, 0);
216 sequence
= gen_pow2_profiler (values
+ i
, t
, 0);
219 case HIST_TYPE_SINGLE_VALUE
:
220 sequence
= gen_one_value_profiler (values
+ i
, t
, 0);
223 case HIST_TYPE_CONST_DELTA
:
224 sequence
= gen_const_delta_profiler (values
+ i
, t
, 0);
231 safe_insert_insn_on_edge (sequence
, e
);
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
)
249 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
250 if (!EDGE_INFO (e
)->ignore
&& !EDGE_INFO (e
)->on_tree
)
254 counts
= get_coverage_counts (GCOV_COUNTER_ARCS
, num_edges
, &profile_info
);
258 if (dump_file
&& profile_info
)
259 fprintf(dump_file
, "Merged %u profiles with maximal count %u.\n",
260 profile_info
->runs
, (unsigned) profile_info
->sum_max
);
266 /* Compute the branch probabilities for the various branches.
267 Annotate them accordingly. */
270 compute_branch_probabilities (void)
277 int hist_br_prob
[20];
278 int num_never_executed
;
280 gcov_type
*exec_counts
= get_exec_counts ();
281 int exec_counts_pos
= 0;
283 /* Very simple sanity checks so we catch bugs in our profiling code. */
286 if (profile_info
->run_max
* profile_info
->runs
< profile_info
->sum_max
)
288 error ("corrupted profile info: run_max * runs < sum_max");
292 if (profile_info
->sum_all
< profile_info
->sum_max
)
294 error ("corrupted profile info: sum_all is smaller than sum_max");
299 /* Attach extra info block to each bb. */
301 alloc_aux_for_blocks (sizeof (struct bb_info
));
302 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
, NULL
, next_bb
)
306 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
307 if (!EDGE_INFO (e
)->ignore
)
308 BB_INFO (bb
)->succ_count
++;
309 for (e
= bb
->pred
; e
; e
= e
->pred_next
)
310 if (!EDGE_INFO (e
)->ignore
)
311 BB_INFO (bb
)->pred_count
++;
314 /* Avoid predicting entry on exit nodes. */
315 BB_INFO (EXIT_BLOCK_PTR
)->succ_count
= 2;
316 BB_INFO (ENTRY_BLOCK_PTR
)->pred_count
= 2;
318 /* For each edge not on the spanning tree, set its execution count from
321 /* The first count in the .da file is the number of times that the function
322 was entered. This is the exec_count for block zero. */
324 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
, NULL
, next_bb
)
327 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
328 if (!EDGE_INFO (e
)->ignore
&& !EDGE_INFO (e
)->on_tree
)
333 e
->count
= exec_counts
[exec_counts_pos
++];
334 if (e
->count
> profile_info
->sum_max
)
336 error ("corrupted profile info: edge from %i to %i exceeds maximal count",
337 bb
->index
, e
->dest
->index
);
343 EDGE_INFO (e
)->count_valid
= 1;
344 BB_INFO (bb
)->succ_count
--;
345 BB_INFO (e
->dest
)->pred_count
--;
348 fprintf (dump_file
, "\nRead edge from %i to %i, count:",
349 bb
->index
, e
->dest
->index
);
350 fprintf (dump_file
, HOST_WIDEST_INT_PRINT_DEC
,
351 (HOST_WIDEST_INT
) e
->count
);
357 fprintf (dump_file
, "\n%d edge counts read\n", num_edges
);
359 /* For every block in the file,
360 - if every exit/entrance edge has a known count, then set the block count
361 - if the block count is known, and every exit/entrance edge but one has
362 a known execution count, then set the count of the remaining edge
364 As edge counts are set, decrement the succ/pred count, but don't delete
365 the edge, that way we can easily tell when all edges are known, or only
366 one edge is unknown. */
368 /* The order that the basic blocks are iterated through is important.
369 Since the code that finds spanning trees starts with block 0, low numbered
370 edges are put on the spanning tree in preference to high numbered edges.
371 Hence, most instrumented edges are at the end. Graph solving works much
372 faster if we propagate numbers from the end to the start.
374 This takes an average of slightly more than 3 passes. */
382 FOR_BB_BETWEEN (bb
, EXIT_BLOCK_PTR
, NULL
, prev_bb
)
384 struct bb_info
*bi
= BB_INFO (bb
);
385 if (! bi
->count_valid
)
387 if (bi
->succ_count
== 0)
392 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
398 else if (bi
->pred_count
== 0)
403 for (e
= bb
->pred
; e
; e
= e
->pred_next
)
412 if (bi
->succ_count
== 1)
417 /* One of the counts will be invalid, but it is zero,
418 so adding it in also doesn't hurt. */
419 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
422 /* Seedgeh for the invalid edge, and set its count. */
423 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
424 if (! EDGE_INFO (e
)->count_valid
&& ! EDGE_INFO (e
)->ignore
)
427 /* Calculate count for remaining edge by conservation. */
428 total
= bb
->count
- total
;
432 EDGE_INFO (e
)->count_valid
= 1;
436 BB_INFO (e
->dest
)->pred_count
--;
439 if (bi
->pred_count
== 1)
444 /* One of the counts will be invalid, but it is zero,
445 so adding it in also doesn't hurt. */
446 for (e
= bb
->pred
; e
; e
= e
->pred_next
)
449 /* Search for the invalid edge, and set its count. */
450 for (e
= bb
->pred
; e
; e
= e
->pred_next
)
451 if (!EDGE_INFO (e
)->count_valid
&& !EDGE_INFO (e
)->ignore
)
454 /* Calculate count for remaining edge by conservation. */
455 total
= bb
->count
- total
+ e
->count
;
459 EDGE_INFO (e
)->count_valid
= 1;
463 BB_INFO (e
->src
)->succ_count
--;
470 dump_flow_info (dump_file
);
472 total_num_passes
+= passes
;
474 fprintf (dump_file
, "Graph solving took %d passes.\n\n", passes
);
476 /* If the graph has been correctly solved, every block will have a
477 succ and pred count of zero. */
480 if (BB_INFO (bb
)->succ_count
|| BB_INFO (bb
)->pred_count
)
484 /* For every edge, calculate its branch probability and add a reg_note
485 to the branch insn to indicate this. */
487 for (i
= 0; i
< 20; i
++)
489 num_never_executed
= 0;
492 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
, NULL
, next_bb
)
499 error ("corrupted profile info: number of iterations for basic block %d thought to be %i",
500 bb
->index
, (int)bb
->count
);
503 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
505 /* Function may return twice in the cased the called function is
506 setjmp or calls fork, but we can't represent this by extra
507 edge from the entry, since extra edge from the exit is
508 already present. We get negative frequency from the entry
511 && e
->dest
== EXIT_BLOCK_PTR
)
512 || (e
->count
> bb
->count
513 && e
->dest
!= EXIT_BLOCK_PTR
))
515 rtx insn
= BB_END (bb
);
517 while (GET_CODE (insn
) != CALL_INSN
518 && insn
!= BB_HEAD (bb
)
519 && keep_with_call_p (insn
))
520 insn
= PREV_INSN (insn
);
521 if (GET_CODE (insn
) == CALL_INSN
)
522 e
->count
= e
->count
< 0 ? 0 : bb
->count
;
524 if (e
->count
< 0 || e
->count
> bb
->count
)
526 error ("corrupted profile info: number of executions for edge %d-%d thought to be %i",
527 e
->src
->index
, e
->dest
->index
,
529 e
->count
= bb
->count
/ 2;
534 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
535 e
->probability
= (e
->count
* REG_BR_PROB_BASE
+ bb
->count
/ 2) / bb
->count
;
537 && any_condjump_p (BB_END (bb
))
538 && bb
->succ
->succ_next
)
544 /* Find the branch edge. It is possible that we do have fake
546 for (e
= bb
->succ
; e
->flags
& (EDGE_FAKE
| EDGE_FALLTHRU
);
548 continue; /* Loop body has been intentionally left blank. */
550 prob
= e
->probability
;
551 index
= prob
* 20 / REG_BR_PROB_BASE
;
555 hist_br_prob
[index
]++;
557 note
= find_reg_note (BB_END (bb
), REG_BR_PROB
, 0);
558 /* There may be already note put by some other pass, such
559 as builtin_expect expander. */
561 XEXP (note
, 0) = GEN_INT (prob
);
563 REG_NOTES (BB_END (bb
))
564 = gen_rtx_EXPR_LIST (REG_BR_PROB
, GEN_INT (prob
),
565 REG_NOTES (BB_END (bb
)));
569 /* Otherwise distribute the probabilities evenly so we get sane
570 sum. Use simple heuristics that if there are normal edges,
571 give all abnormals frequency of 0, otherwise distribute the
572 frequency over abnormals (this is the case of noreturn
578 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
579 if (!(e
->flags
& (EDGE_COMPLEX
| EDGE_FAKE
)))
583 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
584 if (!(e
->flags
& (EDGE_COMPLEX
| EDGE_FAKE
)))
585 e
->probability
= REG_BR_PROB_BASE
/ total
;
591 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
593 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
594 e
->probability
= REG_BR_PROB_BASE
/ total
;
597 && any_condjump_p (BB_END (bb
))
598 && bb
->succ
->succ_next
)
599 num_branches
++, num_never_executed
;
605 fprintf (dump_file
, "%d branches\n", num_branches
);
606 fprintf (dump_file
, "%d branches never executed\n",
609 for (i
= 0; i
< 10; i
++)
610 fprintf (dump_file
, "%d%% branches in range %d-%d%%\n",
611 (hist_br_prob
[i
] + hist_br_prob
[19-i
]) * 100 / num_branches
,
614 total_num_branches
+= num_branches
;
615 total_num_never_executed
+= num_never_executed
;
616 for (i
= 0; i
< 20; i
++)
617 total_hist_br_prob
[i
] += hist_br_prob
[i
];
619 fputc ('\n', dump_file
);
620 fputc ('\n', dump_file
);
623 free_aux_for_blocks ();
626 /* Load value histograms for N_VALUES values whose description is stored
627 in VALUES array from .da file. */
629 compute_value_histograms (unsigned n_values
, struct histogram_value
*values
)
631 unsigned i
, j
, t
, any
;
632 unsigned n_histogram_counters
[GCOV_N_VALUE_COUNTERS
];
633 gcov_type
*histogram_counts
[GCOV_N_VALUE_COUNTERS
];
634 gcov_type
*act_count
[GCOV_N_VALUE_COUNTERS
];
635 gcov_type
*aact_count
;
637 for (t
= 0; t
< GCOV_N_VALUE_COUNTERS
; t
++)
638 n_histogram_counters
[t
] = 0;
640 for (i
= 0; i
< n_values
; i
++)
641 n_histogram_counters
[(int) (values
[i
].type
)] += values
[i
].n_counters
;
644 for (t
= 0; t
< GCOV_N_VALUE_COUNTERS
; t
++)
646 if (!n_histogram_counters
[t
])
648 histogram_counts
[t
] = NULL
;
652 histogram_counts
[t
] =
653 get_coverage_counts (COUNTER_FOR_HIST_TYPE (t
),
654 n_histogram_counters
[t
], NULL
);
655 if (histogram_counts
[t
])
657 act_count
[t
] = histogram_counts
[t
];
662 for (i
= 0; i
< n_values
; i
++)
664 rtx hist_list
= NULL_RTX
;
665 t
= (int) (values
[i
].type
);
667 aact_count
= act_count
[t
];
668 act_count
[t
] += values
[i
].n_counters
;
669 for (j
= values
[i
].n_counters
; j
> 0; j
--)
670 hist_list
= alloc_EXPR_LIST (0, GEN_INT (aact_count
[j
- 1]), hist_list
);
671 hist_list
= alloc_EXPR_LIST (0, copy_rtx (values
[i
].value
), hist_list
);
672 hist_list
= alloc_EXPR_LIST (0, GEN_INT (values
[i
].type
), hist_list
);
673 REG_NOTES (values
[i
].insn
) =
674 alloc_EXPR_LIST (REG_VALUE_PROFILE
, hist_list
,
675 REG_NOTES (values
[i
].insn
));
678 for (t
= 0; t
< GCOV_N_VALUE_COUNTERS
; t
++)
679 if (histogram_counts
[t
])
680 free (histogram_counts
[t
]);
683 /* Instrument and/or analyze program behavior based on program flow graph.
684 In either case, this function builds a flow graph for the function being
685 compiled. The flow graph is stored in BB_GRAPH.
687 When FLAG_PROFILE_ARCS is nonzero, this function instruments the edges in
688 the flow graph that are needed to reconstruct the dynamic behavior of the
691 When FLAG_BRANCH_PROBABILITIES is nonzero, this function reads auxiliary
692 information from a data file containing edge count information from previous
693 executions of the function being compiled. In this case, the flow graph is
694 annotated with actual execution counts, which are later propagated into the
695 rtl for optimization purposes.
697 Main entry point of this file. */
704 unsigned num_edges
, ignored_edges
;
705 unsigned num_instrumented
;
706 struct edge_list
*el
;
707 unsigned n_values
= 0;
708 struct histogram_value
*values
= NULL
;
710 total_num_times_called
++;
712 flow_call_edges_add (NULL
);
713 add_noreturn_fake_exit_edges ();
715 /* We can't handle cyclic regions constructed using abnormal edges.
716 To avoid these we replace every source of abnormal edge by a fake
717 edge from entry node and every destination by fake edge to exit.
718 This keeps graph acyclic and our calculation exact for all normal
719 edges except for exit and entrance ones.
721 We also add fake exit edges for each call and asm statement in the
722 basic, since it may not return. */
726 int need_exit_edge
= 0, need_entry_edge
= 0;
727 int have_exit_edge
= 0, have_entry_edge
= 0;
730 /* Functions returning multiple times are not handled by extra edges.
731 Instead we simply allow negative counts on edges from exit to the
732 block past call and corresponding probabilities. We can't go
733 with the extra edges because that would result in flowgraph that
734 needs to have fake edges outside the spanning tree. */
736 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
738 if ((e
->flags
& (EDGE_ABNORMAL
| EDGE_ABNORMAL_CALL
))
739 && e
->dest
!= EXIT_BLOCK_PTR
)
741 if (e
->dest
== EXIT_BLOCK_PTR
)
744 for (e
= bb
->pred
; e
; e
= e
->pred_next
)
746 if ((e
->flags
& (EDGE_ABNORMAL
| EDGE_ABNORMAL_CALL
))
747 && e
->src
!= ENTRY_BLOCK_PTR
)
749 if (e
->src
== ENTRY_BLOCK_PTR
)
753 if (need_exit_edge
&& !have_exit_edge
)
756 fprintf (dump_file
, "Adding fake exit edge to bb %i\n",
758 make_edge (bb
, EXIT_BLOCK_PTR
, EDGE_FAKE
);
760 if (need_entry_edge
&& !have_entry_edge
)
763 fprintf (dump_file
, "Adding fake entry edge to bb %i\n",
765 make_edge (ENTRY_BLOCK_PTR
, bb
, EDGE_FAKE
);
769 el
= create_edge_list ();
770 num_edges
= NUM_EDGES (el
);
771 alloc_aux_for_edges (sizeof (struct edge_info
));
773 /* The basic blocks are expected to be numbered sequentially. */
777 for (i
= 0 ; i
< num_edges
; i
++)
779 edge e
= INDEX_EDGE (el
, i
);
782 /* Mark edges we've replaced by fake edges above as ignored. */
783 if ((e
->flags
& (EDGE_ABNORMAL
| EDGE_ABNORMAL_CALL
))
784 && e
->src
!= ENTRY_BLOCK_PTR
&& e
->dest
!= EXIT_BLOCK_PTR
)
786 EDGE_INFO (e
)->ignore
= 1;
791 #ifdef ENABLE_CHECKING
795 /* Create spanning tree from basic block graph, mark each edge that is
796 on the spanning tree. We insert as many abnormal and critical edges
797 as possible to minimize number of edge splits necessary. */
799 find_spanning_tree (el
);
801 /* Fake edges that are not on the tree will not be instrumented, so
802 mark them ignored. */
803 for (num_instrumented
= i
= 0; i
< num_edges
; i
++)
805 edge e
= INDEX_EDGE (el
, i
);
806 struct edge_info
*inf
= EDGE_INFO (e
);
808 if (inf
->ignore
|| inf
->on_tree
)
810 else if (e
->flags
& EDGE_FAKE
)
819 total_num_blocks
+= n_basic_blocks
+ 2;
821 fprintf (dump_file
, "%d basic blocks\n", n_basic_blocks
);
823 total_num_edges
+= num_edges
;
825 fprintf (dump_file
, "%d edges\n", num_edges
);
827 total_num_edges_ignored
+= ignored_edges
;
829 fprintf (dump_file
, "%d ignored edges\n", ignored_edges
);
831 /* Write the data from which gcov can reconstruct the basic block
834 /* Basic block flags */
835 if (coverage_begin_output ())
837 gcov_position_t offset
;
839 offset
= gcov_write_tag (GCOV_TAG_BLOCKS
);
840 for (i
= 0; i
!= (unsigned) (n_basic_blocks
+ 2); i
++)
841 gcov_write_unsigned (0);
842 gcov_write_length (offset
);
845 /* Keep all basic block indexes nonnegative in the gcov output.
846 Index 0 is used for entry block, last index is for exit block.
848 ENTRY_BLOCK_PTR
->index
= -1;
849 EXIT_BLOCK_PTR
->index
= last_basic_block
;
850 #define BB_TO_GCOV_INDEX(bb) ((bb)->index + 1)
853 if (coverage_begin_output ())
855 gcov_position_t offset
;
857 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
, EXIT_BLOCK_PTR
, next_bb
)
861 offset
= gcov_write_tag (GCOV_TAG_ARCS
);
862 gcov_write_unsigned (BB_TO_GCOV_INDEX (bb
));
864 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
866 struct edge_info
*i
= EDGE_INFO (e
);
869 unsigned flag_bits
= 0;
872 flag_bits
|= GCOV_ARC_ON_TREE
;
873 if (e
->flags
& EDGE_FAKE
)
874 flag_bits
|= GCOV_ARC_FAKE
;
875 if (e
->flags
& EDGE_FALLTHRU
)
876 flag_bits
|= GCOV_ARC_FALLTHROUGH
;
878 gcov_write_unsigned (BB_TO_GCOV_INDEX (e
->dest
));
879 gcov_write_unsigned (flag_bits
);
883 gcov_write_length (offset
);
888 if (coverage_begin_output ())
890 char const *prev_file_name
= NULL
;
891 gcov_position_t offset
;
895 rtx insn
= BB_HEAD (bb
);
896 int ignore_next_note
= 0;
900 /* We are looking for line number notes. Search backward
901 before basic block to find correct ones. */
902 insn
= prev_nonnote_insn (insn
);
906 insn
= NEXT_INSN (insn
);
908 while (insn
!= BB_END (bb
))
910 if (GET_CODE (insn
) == NOTE
)
912 /* Must ignore the line number notes that
913 immediately follow the end of an inline function
914 to avoid counting it twice. There is a note
915 before the call, and one after the call. */
916 if (NOTE_LINE_NUMBER (insn
)
917 == NOTE_INSN_REPEATED_LINE_NUMBER
)
918 ignore_next_note
= 1;
919 else if (NOTE_LINE_NUMBER (insn
) <= 0)
921 else if (ignore_next_note
)
922 ignore_next_note
= 0;
927 offset
= gcov_write_tag (GCOV_TAG_LINES
);
928 gcov_write_unsigned (BB_TO_GCOV_INDEX (bb
));
931 /* If this is a new source file, then output the
932 file's name to the .bb file. */
934 || strcmp (NOTE_SOURCE_FILE (insn
),
937 prev_file_name
= NOTE_SOURCE_FILE (insn
);
938 gcov_write_unsigned (0);
939 gcov_write_string (prev_file_name
);
941 gcov_write_unsigned (NOTE_LINE_NUMBER (insn
));
944 insn
= NEXT_INSN (insn
);
949 /* A file of NULL indicates the end of run. */
950 gcov_write_unsigned (0);
951 gcov_write_string (NULL
);
952 gcov_write_length (offset
);
956 ENTRY_BLOCK_PTR
->index
= ENTRY_BLOCK
;
957 EXIT_BLOCK_PTR
->index
= EXIT_BLOCK
;
958 #undef BB_TO_GCOV_INDEX
960 if (flag_profile_values
)
962 life_analysis (get_insns (), NULL
, PROP_DEATH_NOTES
);
963 find_values_to_profile (&n_values
, &values
);
964 allocate_reg_info (max_reg_num (), FALSE
, FALSE
);
967 if (flag_branch_probabilities
)
969 compute_branch_probabilities ();
970 if (flag_profile_values
)
971 compute_value_histograms (n_values
, values
);
974 /* For each edge not on the spanning tree, add counting code as rtl. */
976 && coverage_counter_alloc (GCOV_COUNTER_ARCS
, num_instrumented
))
978 unsigned n_instrumented
= instrument_edges (el
);
980 if (n_instrumented
!= num_instrumented
)
983 if (flag_profile_values
)
984 instrument_values (n_values
, values
);
986 /* Commit changes done by instrumentation. */
987 commit_edge_insertions_watch_calls ();
988 allocate_reg_info (max_reg_num (), FALSE
, FALSE
);
991 remove_fake_edges ();
992 free_aux_for_edges ();
993 /* Re-merge split basic blocks and the mess introduced by
994 insert_insn_on_edge. */
995 cleanup_cfg (profile_arc_flag
? CLEANUP_EXPENSIVE
: 0);
997 dump_flow_info (dump_file
);
1002 /* Union find algorithm implementation for the basic blocks using
1006 find_group (basic_block bb
)
1008 basic_block group
= bb
, bb1
;
1010 while ((basic_block
) group
->aux
!= group
)
1011 group
= (basic_block
) group
->aux
;
1013 /* Compress path. */
1014 while ((basic_block
) bb
->aux
!= group
)
1016 bb1
= (basic_block
) bb
->aux
;
1017 bb
->aux
= (void *) group
;
1024 union_groups (basic_block bb1
, basic_block bb2
)
1026 basic_block bb1g
= find_group (bb1
);
1027 basic_block bb2g
= find_group (bb2
);
1029 /* ??? I don't have a place for the rank field. OK. Lets go w/o it,
1030 this code is unlikely going to be performance problem anyway. */
1037 /* This function searches all of the edges in the program flow graph, and puts
1038 as many bad edges as possible onto the spanning tree. Bad edges include
1039 abnormals edges, which can't be instrumented at the moment. Since it is
1040 possible for fake edges to form a cycle, we will have to develop some
1041 better way in the future. Also put critical edges to the tree, since they
1042 are more expensive to instrument. */
1045 find_spanning_tree (struct edge_list
*el
)
1048 int num_edges
= NUM_EDGES (el
);
1051 /* We use aux field for standard union-find algorithm. */
1052 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
, NULL
, next_bb
)
1055 /* Add fake edge exit to entry we can't instrument. */
1056 union_groups (EXIT_BLOCK_PTR
, ENTRY_BLOCK_PTR
);
1058 /* First add all abnormal edges to the tree unless they form a cycle. Also
1059 add all edges to EXIT_BLOCK_PTR to avoid inserting profiling code behind
1060 setting return value from function. */
1061 for (i
= 0; i
< num_edges
; i
++)
1063 edge e
= INDEX_EDGE (el
, i
);
1064 if (((e
->flags
& (EDGE_ABNORMAL
| EDGE_ABNORMAL_CALL
| EDGE_FAKE
))
1065 || e
->dest
== EXIT_BLOCK_PTR
)
1066 && !EDGE_INFO (e
)->ignore
1067 && (find_group (e
->src
) != find_group (e
->dest
)))
1070 fprintf (dump_file
, "Abnormal edge %d to %d put to tree\n",
1071 e
->src
->index
, e
->dest
->index
);
1072 EDGE_INFO (e
)->on_tree
= 1;
1073 union_groups (e
->src
, e
->dest
);
1077 /* Now insert all critical edges to the tree unless they form a cycle. */
1078 for (i
= 0; i
< num_edges
; i
++)
1080 edge e
= INDEX_EDGE (el
, i
);
1081 if (EDGE_CRITICAL_P (e
) && !EDGE_INFO (e
)->ignore
1082 && find_group (e
->src
) != find_group (e
->dest
))
1085 fprintf (dump_file
, "Critical edge %d to %d put to tree\n",
1086 e
->src
->index
, e
->dest
->index
);
1087 EDGE_INFO (e
)->on_tree
= 1;
1088 union_groups (e
->src
, e
->dest
);
1092 /* And now the rest. */
1093 for (i
= 0; i
< num_edges
; i
++)
1095 edge e
= INDEX_EDGE (el
, i
);
1096 if (!EDGE_INFO (e
)->ignore
1097 && find_group (e
->src
) != find_group (e
->dest
))
1100 fprintf (dump_file
, "Normal edge %d to %d put to tree\n",
1101 e
->src
->index
, e
->dest
->index
);
1102 EDGE_INFO (e
)->on_tree
= 1;
1103 union_groups (e
->src
, e
->dest
);
1107 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
, NULL
, next_bb
)
1111 /* Perform file-level initialization for branch-prob processing. */
1114 init_branch_prob (void)
1118 total_num_blocks
= 0;
1119 total_num_edges
= 0;
1120 total_num_edges_ignored
= 0;
1121 total_num_edges_instrumented
= 0;
1122 total_num_blocks_created
= 0;
1123 total_num_passes
= 0;
1124 total_num_times_called
= 0;
1125 total_num_branches
= 0;
1126 total_num_never_executed
= 0;
1127 for (i
= 0; i
< 20; i
++)
1128 total_hist_br_prob
[i
] = 0;
1131 /* Performs file-level cleanup after branch-prob processing
1135 end_branch_prob (void)
1139 fprintf (dump_file
, "\n");
1140 fprintf (dump_file
, "Total number of blocks: %d\n",
1142 fprintf (dump_file
, "Total number of edges: %d\n", total_num_edges
);
1143 fprintf (dump_file
, "Total number of ignored edges: %d\n",
1144 total_num_edges_ignored
);
1145 fprintf (dump_file
, "Total number of instrumented edges: %d\n",
1146 total_num_edges_instrumented
);
1147 fprintf (dump_file
, "Total number of blocks created: %d\n",
1148 total_num_blocks_created
);
1149 fprintf (dump_file
, "Total number of graph solution passes: %d\n",
1151 if (total_num_times_called
!= 0)
1152 fprintf (dump_file
, "Average number of graph solution passes: %d\n",
1153 (total_num_passes
+ (total_num_times_called
>> 1))
1154 / total_num_times_called
);
1155 fprintf (dump_file
, "Total number of branches: %d\n",
1156 total_num_branches
);
1157 fprintf (dump_file
, "Total number of branches never executed: %d\n",
1158 total_num_never_executed
);
1159 if (total_num_branches
)
1163 for (i
= 0; i
< 10; i
++)
1164 fprintf (dump_file
, "%d%% branches in range %d-%d%%\n",
1165 (total_hist_br_prob
[i
] + total_hist_br_prob
[19-i
]) * 100
1166 / total_num_branches
, 5*i
, 5*i
+5);
1172 /* Output instructions as RTL to increment the edge execution count. */
1175 gen_edge_profiler (int edgeno
)
1177 rtx ref
= coverage_counter_ref (GCOV_COUNTER_ARCS
, edgeno
);
1179 enum machine_mode mode
= GET_MODE (ref
);
1183 ref
= validize_mem (ref
);
1185 tmp
= expand_simple_binop (mode
, PLUS
, ref
, const1_rtx
,
1186 ref
, 0, OPTAB_WIDEN
);
1189 emit_move_insn (copy_rtx (ref
), tmp
);
1191 sequence
= get_insns ();
1196 /* Output instructions as RTL to increment the interval histogram counter.
1197 VALUE is the expression whose value is profiled. TAG is the tag of the
1198 section for counters, BASE is offset of the counter position. */
1201 gen_interval_profiler (struct histogram_value
*value
, unsigned tag
,
1204 unsigned gcov_size
= tree_low_cst (TYPE_SIZE (GCOV_TYPE_NODE
), 1);
1205 enum machine_mode mode
= mode_for_size (gcov_size
, MODE_INT
, 0);
1206 rtx mem_ref
, tmp
, tmp1
, mr
, val
;
1208 rtx more_label
= gen_label_rtx ();
1209 rtx less_label
= gen_label_rtx ();
1210 rtx end_of_code_label
= gen_label_rtx ();
1211 int per_counter
= gcov_size
/ BITS_PER_UNIT
;
1216 emit_insn (value
->seq
);
1218 mr
= gen_reg_rtx (Pmode
);
1220 tmp
= coverage_counter_ref (tag
, base
);
1221 tmp
= force_reg (Pmode
, XEXP (tmp
, 0));
1223 val
= expand_simple_binop (value
->mode
, MINUS
,
1224 copy_rtx (value
->value
),
1225 GEN_INT (value
->hdata
.intvl
.int_start
),
1226 NULL_RTX
, 0, OPTAB_WIDEN
);
1228 if (value
->hdata
.intvl
.may_be_more
)
1229 do_compare_rtx_and_jump (copy_rtx (val
), GEN_INT (value
->hdata
.intvl
.steps
),
1230 GE
, 0, value
->mode
, NULL_RTX
, NULL_RTX
, more_label
);
1231 if (value
->hdata
.intvl
.may_be_less
)
1232 do_compare_rtx_and_jump (copy_rtx (val
), const0_rtx
, LT
, 0, value
->mode
,
1233 NULL_RTX
, NULL_RTX
, less_label
);
1235 /* We are in range. */
1236 tmp1
= expand_simple_binop (value
->mode
, MULT
,
1237 copy_rtx (val
), GEN_INT (per_counter
),
1238 NULL_RTX
, 0, OPTAB_WIDEN
);
1239 tmp1
= expand_simple_binop (Pmode
, PLUS
, copy_rtx (tmp
), tmp1
, mr
,
1242 emit_move_insn (copy_rtx (mr
), tmp1
);
1244 if (value
->hdata
.intvl
.may_be_more
1245 || value
->hdata
.intvl
.may_be_less
)
1247 emit_jump_insn (gen_jump (end_of_code_label
));
1251 /* Above the interval. */
1252 if (value
->hdata
.intvl
.may_be_more
)
1254 emit_label (more_label
);
1255 tmp1
= expand_simple_binop (Pmode
, PLUS
, copy_rtx (tmp
),
1256 GEN_INT (per_counter
* value
->hdata
.intvl
.steps
),
1257 mr
, 0, OPTAB_WIDEN
);
1259 emit_move_insn (copy_rtx (mr
), tmp1
);
1260 if (value
->hdata
.intvl
.may_be_less
)
1262 emit_jump_insn (gen_jump (end_of_code_label
));
1267 /* Below the interval. */
1268 if (value
->hdata
.intvl
.may_be_less
)
1270 emit_label (less_label
);
1271 tmp1
= expand_simple_binop (Pmode
, PLUS
, copy_rtx (tmp
),
1272 GEN_INT (per_counter
* (value
->hdata
.intvl
.steps
1273 + (value
->hdata
.intvl
.may_be_more
? 1 : 0))),
1274 mr
, 0, OPTAB_WIDEN
);
1276 emit_move_insn (copy_rtx (mr
), tmp1
);
1279 if (value
->hdata
.intvl
.may_be_more
1280 || value
->hdata
.intvl
.may_be_less
)
1281 emit_label (end_of_code_label
);
1283 mem_ref
= validize_mem (gen_rtx_MEM (mode
, mr
));
1285 tmp
= expand_simple_binop (mode
, PLUS
, copy_rtx (mem_ref
), const1_rtx
,
1286 mem_ref
, 0, OPTAB_WIDEN
);
1289 emit_move_insn (copy_rtx (mem_ref
), tmp
);
1291 sequence
= get_insns ();
1293 rebuild_jump_labels (sequence
);
1297 /* Output instructions as RTL to increment the power of two histogram counter.
1298 VALUE is the expression whose value is profiled. TAG is the tag of the
1299 section for counters, BASE is offset of the counter position. */
1302 gen_pow2_profiler (struct histogram_value
*value
, unsigned tag
, unsigned base
)
1304 unsigned gcov_size
= tree_low_cst (TYPE_SIZE (GCOV_TYPE_NODE
), 1);
1305 enum machine_mode mode
= mode_for_size (gcov_size
, MODE_INT
, 0);
1306 rtx mem_ref
, tmp
, mr
, uval
;
1308 rtx end_of_code_label
= gen_label_rtx ();
1309 rtx loop_label
= gen_label_rtx ();
1310 int per_counter
= gcov_size
/ BITS_PER_UNIT
;
1315 emit_insn (value
->seq
);
1317 mr
= gen_reg_rtx (Pmode
);
1318 tmp
= coverage_counter_ref (tag
, base
);
1319 tmp
= force_reg (Pmode
, XEXP (tmp
, 0));
1320 emit_move_insn (mr
, tmp
);
1322 uval
= gen_reg_rtx (value
->mode
);
1323 emit_move_insn (uval
, copy_rtx (value
->value
));
1325 /* Check for non-power of 2. */
1326 if (value
->hdata
.pow2
.may_be_other
)
1328 do_compare_rtx_and_jump (copy_rtx (uval
), const0_rtx
, LE
, 0, value
->mode
,
1329 NULL_RTX
, NULL_RTX
, end_of_code_label
);
1330 tmp
= expand_simple_binop (value
->mode
, PLUS
, copy_rtx (uval
),
1331 constm1_rtx
, NULL_RTX
, 0, OPTAB_WIDEN
);
1332 tmp
= expand_simple_binop (value
->mode
, AND
, copy_rtx (uval
), tmp
,
1333 NULL_RTX
, 0, OPTAB_WIDEN
);
1334 do_compare_rtx_and_jump (tmp
, const0_rtx
, NE
, 0, value
->mode
, NULL_RTX
,
1335 NULL_RTX
, end_of_code_label
);
1338 /* Count log_2(value). */
1339 emit_label (loop_label
);
1341 tmp
= expand_simple_binop (Pmode
, PLUS
, copy_rtx (mr
), GEN_INT (per_counter
), mr
, 0, OPTAB_WIDEN
);
1343 emit_move_insn (copy_rtx (mr
), tmp
);
1345 tmp
= expand_simple_binop (value
->mode
, ASHIFTRT
, copy_rtx (uval
), const1_rtx
,
1346 uval
, 0, OPTAB_WIDEN
);
1348 emit_move_insn (copy_rtx (uval
), tmp
);
1350 do_compare_rtx_and_jump (copy_rtx (uval
), const0_rtx
, NE
, 0, value
->mode
,
1351 NULL_RTX
, NULL_RTX
, loop_label
);
1353 /* Increase the counter. */
1354 emit_label (end_of_code_label
);
1356 mem_ref
= validize_mem (gen_rtx_MEM (mode
, mr
));
1358 tmp
= expand_simple_binop (mode
, PLUS
, copy_rtx (mem_ref
), const1_rtx
,
1359 mem_ref
, 0, OPTAB_WIDEN
);
1362 emit_move_insn (copy_rtx (mem_ref
), tmp
);
1364 sequence
= get_insns ();
1366 rebuild_jump_labels (sequence
);
1370 /* Output instructions as RTL for code to find the most common value.
1371 VALUE is the expression whose value is profiled. TAG is the tag of the
1372 section for counters, BASE is offset of the counter position. */
1375 gen_one_value_profiler (struct histogram_value
*value
, unsigned tag
,
1378 unsigned gcov_size
= tree_low_cst (TYPE_SIZE (GCOV_TYPE_NODE
), 1);
1379 enum machine_mode mode
= mode_for_size (gcov_size
, MODE_INT
, 0);
1380 rtx stored_value_ref
, counter_ref
, all_ref
, stored_value
, counter
, all
;
1383 rtx same_label
= gen_label_rtx ();
1384 rtx zero_label
= gen_label_rtx ();
1385 rtx end_of_code_label
= gen_label_rtx ();
1390 emit_insn (value
->seq
);
1392 stored_value_ref
= coverage_counter_ref (tag
, base
);
1393 counter_ref
= coverage_counter_ref (tag
, base
+ 1);
1394 all_ref
= coverage_counter_ref (tag
, base
+ 2);
1395 stored_value
= validize_mem (stored_value_ref
);
1396 counter
= validize_mem (counter_ref
);
1397 all
= validize_mem (all_ref
);
1399 uval
= gen_reg_rtx (mode
);
1400 convert_move (uval
, copy_rtx (value
->value
), 0);
1402 /* Check if the stored value matches. */
1403 do_compare_rtx_and_jump (copy_rtx (uval
), copy_rtx (stored_value
), EQ
,
1404 0, mode
, NULL_RTX
, NULL_RTX
, same_label
);
1406 /* Does not match; check whether the counter is zero. */
1407 do_compare_rtx_and_jump (copy_rtx (counter
), const0_rtx
, EQ
, 0, mode
,
1408 NULL_RTX
, NULL_RTX
, zero_label
);
1410 /* The counter is not zero yet. */
1411 tmp
= expand_simple_binop (mode
, PLUS
, copy_rtx (counter
), constm1_rtx
,
1412 counter
, 0, OPTAB_WIDEN
);
1415 emit_move_insn (copy_rtx (counter
), tmp
);
1417 emit_jump_insn (gen_jump (end_of_code_label
));
1420 emit_label (zero_label
);
1421 /* Set new value. */
1422 emit_move_insn (copy_rtx (stored_value
), copy_rtx (uval
));
1424 emit_label (same_label
);
1425 /* Increase the counter. */
1426 tmp
= expand_simple_binop (mode
, PLUS
, copy_rtx (counter
), const1_rtx
,
1427 counter
, 0, OPTAB_WIDEN
);
1430 emit_move_insn (copy_rtx (counter
), tmp
);
1432 emit_label (end_of_code_label
);
1434 /* Increase the counter of all executions; this seems redundant given
1435 that ve have counts for edges in cfg, but it may happen that some
1436 optimization will change the counts for the block (either because
1437 it is unable to update them correctly, or because it will duplicate
1438 the block or its part). */
1439 tmp
= expand_simple_binop (mode
, PLUS
, copy_rtx (all
), const1_rtx
,
1440 all
, 0, OPTAB_WIDEN
);
1443 emit_move_insn (copy_rtx (all
), tmp
);
1444 sequence
= get_insns ();
1446 rebuild_jump_labels (sequence
);
1450 /* Output instructions as RTL for code to find the most common value of
1451 a difference between two evaluations of an expression.
1452 VALUE is the expression whose value is profiled. TAG is the tag of the
1453 section for counters, BASE is offset of the counter position. */
1456 gen_const_delta_profiler (struct histogram_value
*value
, unsigned tag
,
1459 struct histogram_value one_value_delta
;
1460 unsigned gcov_size
= tree_low_cst (TYPE_SIZE (GCOV_TYPE_NODE
), 1);
1461 enum machine_mode mode
= mode_for_size (gcov_size
, MODE_INT
, 0);
1462 rtx stored_value_ref
, stored_value
, tmp
, uval
;
1468 emit_insn (value
->seq
);
1470 stored_value_ref
= coverage_counter_ref (tag
, base
);
1471 stored_value
= validize_mem (stored_value_ref
);
1473 uval
= gen_reg_rtx (mode
);
1474 convert_move (uval
, copy_rtx (value
->value
), 0);
1475 tmp
= expand_simple_binop (mode
, MINUS
,
1476 copy_rtx (uval
), copy_rtx (stored_value
),
1477 NULL_RTX
, 0, OPTAB_WIDEN
);
1479 one_value_delta
.value
= tmp
;
1480 one_value_delta
.mode
= mode
;
1481 one_value_delta
.seq
= NULL_RTX
;
1482 one_value_delta
.insn
= value
->insn
;
1483 one_value_delta
.type
= HIST_TYPE_SINGLE_VALUE
;
1484 emit_insn (gen_one_value_profiler (&one_value_delta
, tag
, base
+ 1));
1486 emit_move_insn (copy_rtx (stored_value
), uval
);
1487 sequence
= get_insns ();
1489 rebuild_jump_labels (sequence
);