bisect: use "sha1_array" to store skipped revisions
[git/git.git] / bisect.c
CommitLineData
a2ad79ce
CC
1#include "cache.h"
2#include "commit.h"
3#include "diff.h"
4#include "revision.h"
1bf072e3
CC
5#include "refs.h"
6#include "list-objects.h"
3b437b0d 7#include "quote.h"
4eb5b646 8#include "sha1-lookup.h"
ef24c7ca 9#include "run-command.h"
a2ad79ce
CC
10#include "bisect.h"
11
6212b1aa
CC
12struct sha1_array {
13 unsigned char (*sha1)[20];
14 int sha1_nr;
15 int sha1_alloc;
16};
17
18static struct sha1_array skipped_revs;
1bf072e3
CC
19
20static const char **rev_argv;
21static int rev_argv_nr;
22static int rev_argv_alloc;
95188648 23
ef24c7ca
CC
24static const unsigned char *current_bad_sha1;
25
26static const char *argv_diff_tree[] = {"diff-tree", "--pretty", NULL, NULL};
27static const char *argv_checkout[] = {"checkout", "-q", NULL, "--", NULL};
28static const char *argv_show_branch[] = {"show-branch", NULL, NULL};
29
a2ad79ce
CC
30/* bits #0-15 in revision.h */
31
32#define COUNTED (1u<<16)
33
34/*
35 * This is a truly stupid algorithm, but it's only
36 * used for bisection, and we just don't care enough.
37 *
38 * We care just barely enough to avoid recursing for
39 * non-merge entries.
40 */
41static int count_distance(struct commit_list *entry)
42{
43 int nr = 0;
44
45 while (entry) {
46 struct commit *commit = entry->item;
47 struct commit_list *p;
48
49 if (commit->object.flags & (UNINTERESTING | COUNTED))
50 break;
51 if (!(commit->object.flags & TREESAME))
52 nr++;
53 commit->object.flags |= COUNTED;
54 p = commit->parents;
55 entry = p;
56 if (p) {
57 p = p->next;
58 while (p) {
59 nr += count_distance(p);
60 p = p->next;
61 }
62 }
63 }
64
65 return nr;
66}
67
68static void clear_distance(struct commit_list *list)
69{
70 while (list) {
71 struct commit *commit = list->item;
72 commit->object.flags &= ~COUNTED;
73 list = list->next;
74 }
75}
76
77#define DEBUG_BISECT 0
78
79static inline int weight(struct commit_list *elem)
80{
81 return *((int*)(elem->item->util));
82}
83
84static inline void weight_set(struct commit_list *elem, int weight)
85{
86 *((int*)(elem->item->util)) = weight;
87}
88
89static int count_interesting_parents(struct commit *commit)
90{
91 struct commit_list *p;
92 int count;
93
94 for (count = 0, p = commit->parents; p; p = p->next) {
95 if (p->item->object.flags & UNINTERESTING)
96 continue;
97 count++;
98 }
99 return count;
100}
101
102static inline int halfway(struct commit_list *p, int nr)
103{
104 /*
105 * Don't short-cut something we are not going to return!
106 */
107 if (p->item->object.flags & TREESAME)
108 return 0;
109 if (DEBUG_BISECT)
110 return 0;
111 /*
112 * 2 and 3 are halfway of 5.
113 * 3 is halfway of 6 but 2 and 4 are not.
114 */
115 switch (2 * weight(p) - nr) {
116 case -1: case 0: case 1:
117 return 1;
118 default:
119 return 0;
120 }
121}
122
123#if !DEBUG_BISECT
124#define show_list(a,b,c,d) do { ; } while (0)
125#else
126static void show_list(const char *debug, int counted, int nr,
127 struct commit_list *list)
128{
129 struct commit_list *p;
130
131 fprintf(stderr, "%s (%d/%d)\n", debug, counted, nr);
132
133 for (p = list; p; p = p->next) {
134 struct commit_list *pp;
135 struct commit *commit = p->item;
136 unsigned flags = commit->object.flags;
137 enum object_type type;
138 unsigned long size;
139 char *buf = read_sha1_file(commit->object.sha1, &type, &size);
140 char *ep, *sp;
141
142 fprintf(stderr, "%c%c%c ",
143 (flags & TREESAME) ? ' ' : 'T',
144 (flags & UNINTERESTING) ? 'U' : ' ',
145 (flags & COUNTED) ? 'C' : ' ');
146 if (commit->util)
147 fprintf(stderr, "%3d", weight(p));
148 else
149 fprintf(stderr, "---");
150 fprintf(stderr, " %.*s", 8, sha1_to_hex(commit->object.sha1));
151 for (pp = commit->parents; pp; pp = pp->next)
152 fprintf(stderr, " %.*s", 8,
153 sha1_to_hex(pp->item->object.sha1));
154
155 sp = strstr(buf, "\n\n");
156 if (sp) {
157 sp += 2;
158 for (ep = sp; *ep && *ep != '\n'; ep++)
159 ;
160 fprintf(stderr, " %.*s", (int)(ep - sp), sp);
161 }
162 fprintf(stderr, "\n");
163 }
164}
165#endif /* DEBUG_BISECT */
166
167static struct commit_list *best_bisection(struct commit_list *list, int nr)
168{
169 struct commit_list *p, *best;
170 int best_distance = -1;
171
172 best = list;
173 for (p = list; p; p = p->next) {
174 int distance;
175 unsigned flags = p->item->object.flags;
176
177 if (flags & TREESAME)
178 continue;
179 distance = weight(p);
180 if (nr - distance < distance)
181 distance = nr - distance;
182 if (distance > best_distance) {
183 best = p;
184 best_distance = distance;
185 }
186 }
187
188 return best;
189}
190
191struct commit_dist {
192 struct commit *commit;
193 int distance;
194};
195
196static int compare_commit_dist(const void *a_, const void *b_)
197{
198 struct commit_dist *a, *b;
199
200 a = (struct commit_dist *)a_;
201 b = (struct commit_dist *)b_;
202 if (a->distance != b->distance)
203 return b->distance - a->distance; /* desc sort */
204 return hashcmp(a->commit->object.sha1, b->commit->object.sha1);
205}
206
207static struct commit_list *best_bisection_sorted(struct commit_list *list, int nr)
208{
209 struct commit_list *p;
210 struct commit_dist *array = xcalloc(nr, sizeof(*array));
211 int cnt, i;
212
213 for (p = list, cnt = 0; p; p = p->next) {
214 int distance;
215 unsigned flags = p->item->object.flags;
216
217 if (flags & TREESAME)
218 continue;
219 distance = weight(p);
220 if (nr - distance < distance)
221 distance = nr - distance;
222 array[cnt].commit = p->item;
223 array[cnt].distance = distance;
224 cnt++;
225 }
226 qsort(array, cnt, sizeof(*array), compare_commit_dist);
227 for (p = list, i = 0; i < cnt; i++) {
228 struct name_decoration *r = xmalloc(sizeof(*r) + 100);
229 struct object *obj = &(array[i].commit->object);
230
231 sprintf(r->name, "dist=%d", array[i].distance);
232 r->next = add_decoration(&name_decoration, obj, r);
233 p->item = array[i].commit;
234 p = p->next;
235 }
236 if (p)
237 p->next = NULL;
238 free(array);
239 return list;
240}
241
242/*
243 * zero or positive weight is the number of interesting commits it can
244 * reach, including itself. Especially, weight = 0 means it does not
245 * reach any tree-changing commits (e.g. just above uninteresting one
246 * but traversal is with pathspec).
247 *
248 * weight = -1 means it has one parent and its distance is yet to
249 * be computed.
250 *
251 * weight = -2 means it has more than one parent and its distance is
252 * unknown. After running count_distance() first, they will get zero
253 * or positive distance.
254 */
255static struct commit_list *do_find_bisection(struct commit_list *list,
256 int nr, int *weights,
257 int find_all)
258{
259 int n, counted;
260 struct commit_list *p;
261
262 counted = 0;
263
264 for (n = 0, p = list; p; p = p->next) {
265 struct commit *commit = p->item;
266 unsigned flags = commit->object.flags;
267
268 p->item->util = &weights[n++];
269 switch (count_interesting_parents(commit)) {
270 case 0:
271 if (!(flags & TREESAME)) {
272 weight_set(p, 1);
273 counted++;
274 show_list("bisection 2 count one",
275 counted, nr, list);
276 }
277 /*
278 * otherwise, it is known not to reach any
279 * tree-changing commit and gets weight 0.
280 */
281 break;
282 case 1:
283 weight_set(p, -1);
284 break;
285 default:
286 weight_set(p, -2);
287 break;
288 }
289 }
290
291 show_list("bisection 2 initialize", counted, nr, list);
292
293 /*
294 * If you have only one parent in the resulting set
295 * then you can reach one commit more than that parent
296 * can reach. So we do not have to run the expensive
297 * count_distance() for single strand of pearls.
298 *
299 * However, if you have more than one parents, you cannot
300 * just add their distance and one for yourself, since
301 * they usually reach the same ancestor and you would
302 * end up counting them twice that way.
303 *
304 * So we will first count distance of merges the usual
305 * way, and then fill the blanks using cheaper algorithm.
306 */
307 for (p = list; p; p = p->next) {
308 if (p->item->object.flags & UNINTERESTING)
309 continue;
310 if (weight(p) != -2)
311 continue;
312 weight_set(p, count_distance(p));
313 clear_distance(list);
314
315 /* Does it happen to be at exactly half-way? */
316 if (!find_all && halfway(p, nr))
317 return p;
318 counted++;
319 }
320
321 show_list("bisection 2 count_distance", counted, nr, list);
322
323 while (counted < nr) {
324 for (p = list; p; p = p->next) {
325 struct commit_list *q;
326 unsigned flags = p->item->object.flags;
327
328 if (0 <= weight(p))
329 continue;
330 for (q = p->item->parents; q; q = q->next) {
331 if (q->item->object.flags & UNINTERESTING)
332 continue;
333 if (0 <= weight(q))
334 break;
335 }
336 if (!q)
337 continue;
338
339 /*
340 * weight for p is unknown but q is known.
341 * add one for p itself if p is to be counted,
342 * otherwise inherit it from q directly.
343 */
344 if (!(flags & TREESAME)) {
345 weight_set(p, weight(q)+1);
346 counted++;
347 show_list("bisection 2 count one",
348 counted, nr, list);
349 }
350 else
351 weight_set(p, weight(q));
352
353 /* Does it happen to be at exactly half-way? */
354 if (!find_all && halfway(p, nr))
355 return p;
356 }
357 }
358
359 show_list("bisection 2 counted all", counted, nr, list);
360
361 if (!find_all)
362 return best_bisection(list, nr);
363 else
364 return best_bisection_sorted(list, nr);
365}
366
367struct commit_list *find_bisection(struct commit_list *list,
368 int *reaches, int *all,
369 int find_all)
370{
371 int nr, on_list;
372 struct commit_list *p, *best, *next, *last;
373 int *weights;
374
375 show_list("bisection 2 entry", 0, 0, list);
376
377 /*
378 * Count the number of total and tree-changing items on the
379 * list, while reversing the list.
380 */
381 for (nr = on_list = 0, last = NULL, p = list;
382 p;
383 p = next) {
384 unsigned flags = p->item->object.flags;
385
386 next = p->next;
387 if (flags & UNINTERESTING)
388 continue;
389 p->next = last;
390 last = p;
391 if (!(flags & TREESAME))
392 nr++;
393 on_list++;
394 }
395 list = last;
396 show_list("bisection 2 sorted", 0, nr, list);
397
398 *all = nr;
399 weights = xcalloc(on_list, sizeof(*weights));
400
401 /* Do the real work of finding bisection commit. */
402 best = do_find_bisection(list, nr, weights, find_all);
403 if (best) {
404 if (!find_all)
405 best->next = NULL;
406 *reaches = weight(best);
407 }
408 free(weights);
409 return best;
410}
411
1bf072e3
CC
412static int register_ref(const char *refname, const unsigned char *sha1,
413 int flags, void *cb_data)
414{
415 if (!strcmp(refname, "bad")) {
416 ALLOC_GROW(rev_argv, rev_argv_nr + 1, rev_argv_alloc);
ef24c7ca 417 current_bad_sha1 = sha1;
1bf072e3
CC
418 rev_argv[rev_argv_nr++] = xstrdup(sha1_to_hex(sha1));
419 } else if (!prefixcmp(refname, "good-")) {
420 const char *hex = sha1_to_hex(sha1);
421 char *good = xmalloc(strlen(hex) + 2);
422 *good = '^';
423 memcpy(good + 1, hex, strlen(hex) + 1);
424 ALLOC_GROW(rev_argv, rev_argv_nr + 1, rev_argv_alloc);
425 rev_argv[rev_argv_nr++] = good;
426 } else if (!prefixcmp(refname, "skip-")) {
6212b1aa
CC
427 ALLOC_GROW(skipped_revs.sha1, skipped_revs.sha1_nr + 1,
428 skipped_revs.sha1_alloc);
429 hashcpy(skipped_revs.sha1[skipped_revs.sha1_nr++], sha1);
1bf072e3
CC
430 }
431
432 return 0;
433}
434
435static int read_bisect_refs(void)
436{
437 return for_each_ref_in("refs/bisect/", register_ref, NULL);
438}
439
3b437b0d
CC
440void read_bisect_paths(void)
441{
442 struct strbuf str = STRBUF_INIT;
443 const char *filename = git_path("BISECT_NAMES");
444 FILE *fp = fopen(filename, "r");
445
446 if (!fp)
447 die("Could not open file '%s': %s", filename, strerror(errno));
448
449 while (strbuf_getline(&str, fp, '\n') != EOF) {
450 char *quoted;
451 int res;
452
453 strbuf_trim(&str);
454 quoted = strbuf_detach(&str, NULL);
455 res = sq_dequote_to_argv(quoted, &rev_argv,
456 &rev_argv_nr, &rev_argv_alloc);
457 if (res)
458 die("Badly quoted content in file '%s': %s",
459 filename, quoted);
460 }
461
462 strbuf_release(&str);
463 fclose(fp);
464}
465
95188648
CC
466static int skipcmp(const void *a, const void *b)
467{
468 return hashcmp(a, b);
469}
470
471static void prepare_skipped(void)
472{
6212b1aa
CC
473 qsort(skipped_revs.sha1, skipped_revs.sha1_nr,
474 sizeof(*skipped_revs.sha1), skipcmp);
95188648
CC
475}
476
4eb5b646
CC
477static const unsigned char *skipped_sha1_access(size_t index, void *table)
478{
479 unsigned char (*skipped)[20] = table;
480 return skipped[index];
481}
482
95188648
CC
483static int lookup_skipped(unsigned char *sha1)
484{
6212b1aa 485 return sha1_pos(sha1, skipped_revs.sha1, skipped_revs.sha1_nr,
4eb5b646 486 skipped_sha1_access);
95188648
CC
487}
488
489struct commit_list *filter_skipped(struct commit_list *list,
490 struct commit_list **tried,
491 int show_all)
492{
493 struct commit_list *filtered = NULL, **f = &filtered;
494
495 *tried = NULL;
496
6212b1aa 497 if (!skipped_revs.sha1_nr)
95188648
CC
498 return list;
499
500 prepare_skipped();
501
502 while (list) {
503 struct commit_list *next = list->next;
504 list->next = NULL;
505 if (0 <= lookup_skipped(list->item->object.sha1)) {
506 /* Move current to tried list */
507 *tried = list;
508 tried = &list->next;
509 } else {
510 if (!show_all)
511 return list;
512 /* Move current to filtered list */
513 *f = list;
514 f = &list->next;
515 }
516 list = next;
517 }
518
519 return filtered;
520}
1bf072e3 521
3b437b0d 522static void bisect_rev_setup(struct rev_info *revs, const char *prefix)
1bf072e3 523{
3b437b0d
CC
524 init_revisions(revs, prefix);
525 revs->abbrev = 0;
526 revs->commit_format = CMIT_FMT_UNSPECIFIED;
1bf072e3
CC
527
528 /* argv[0] will be ignored by setup_revisions */
529 ALLOC_GROW(rev_argv, rev_argv_nr + 1, rev_argv_alloc);
530 rev_argv[rev_argv_nr++] = xstrdup("bisect_rev_setup");
531
532 if (read_bisect_refs())
533 die("reading bisect refs failed");
534
535 ALLOC_GROW(rev_argv, rev_argv_nr + 1, rev_argv_alloc);
536 rev_argv[rev_argv_nr++] = xstrdup("--");
537
3b437b0d
CC
538 read_bisect_paths();
539
540 ALLOC_GROW(rev_argv, rev_argv_nr + 1, rev_argv_alloc);
541 rev_argv[rev_argv_nr++] = NULL;
542
543 setup_revisions(rev_argv_nr, rev_argv, revs, NULL);
544
545 revs->limited = 1;
546}
547
2ace9727
CC
548static void bisect_common(struct rev_info *revs, const char *prefix,
549 int *reaches, int *all)
550{
551 bisect_rev_setup(revs, prefix);
552
553 if (prepare_revision_walk(revs))
554 die("revision walk setup failed");
555 if (revs->tree_objects)
556 mark_edges_uninteresting(revs->commits, revs, NULL);
557
558 revs->commits = find_bisection(revs->commits, reaches, all,
6212b1aa 559 !!skipped_revs.sha1_nr);
2ace9727
CC
560}
561
ef24c7ca
CC
562static void exit_if_skipped_commits(struct commit_list *tried,
563 const unsigned char *bad)
564{
565 if (!tried)
566 return;
567
568 printf("There are only 'skip'ped commits left to test.\n"
569 "The first bad commit could be any of:\n");
570 print_commit_list(tried, "%s\n", "%s\n");
571 if (bad)
572 printf("%s\n", sha1_to_hex(bad));
573 printf("We cannot bisect more!\n");
574 exit(2);
575}
576
577static void mark_expected_rev(char *bisect_rev_hex)
578{
579 int len = strlen(bisect_rev_hex);
580 const char *filename = git_path("BISECT_EXPECTED_REV");
581 int fd = open(filename, O_CREAT | O_TRUNC | O_WRONLY, 0600);
582
583 if (fd < 0)
584 die("could not create file '%s': %s",
585 filename, strerror(errno));
586
587 bisect_rev_hex[len] = '\n';
588 write_or_die(fd, bisect_rev_hex, len + 1);
589 bisect_rev_hex[len] = '\0';
590
591 if (close(fd) < 0)
592 die("closing file %s: %s", filename, strerror(errno));
593}
594
595static int bisect_checkout(char *bisect_rev_hex)
596{
597 int res;
598
599 mark_expected_rev(bisect_rev_hex);
600
601 argv_checkout[2] = bisect_rev_hex;
602 res = run_command_v_opt(argv_checkout, RUN_GIT_CMD);
603 if (res)
604 exit(res);
605
606 argv_show_branch[1] = bisect_rev_hex;
607 return run_command_v_opt(argv_show_branch, RUN_GIT_CMD);
608}
609
610/*
611 * We use the convention that exiting with an exit code 10 means that
612 * the bisection process finished successfully.
613 * In this case the calling shell script should exit 0.
614 */
615int bisect_next_exit(const char *prefix)
616{
617 struct rev_info revs;
618 struct commit_list *tried;
619 int reaches = 0, all = 0, nr;
620 const unsigned char *bisect_rev;
621 char bisect_rev_hex[41];
622
623 bisect_common(&revs, prefix, &reaches, &all);
624
625 revs.commits = filter_skipped(revs.commits, &tried, 0);
626
627 if (!revs.commits) {
628 /*
629 * We should exit here only if the "bad"
630 * commit is also a "skip" commit.
631 */
632 exit_if_skipped_commits(tried, NULL);
633
634 printf("%s was both good and bad\n",
635 sha1_to_hex(current_bad_sha1));
636 exit(1);
637 }
638
639 bisect_rev = revs.commits->item->object.sha1;
640 memcpy(bisect_rev_hex, sha1_to_hex(bisect_rev), 41);
641
642 if (!hashcmp(bisect_rev, current_bad_sha1)) {
643 exit_if_skipped_commits(tried, current_bad_sha1);
644 printf("%s is first bad commit\n", bisect_rev_hex);
645 argv_diff_tree[2] = bisect_rev_hex;
646 run_command_v_opt(argv_diff_tree, RUN_GIT_CMD);
647 /* This means the bisection process succeeded. */
648 exit(10);
649 }
650
651 nr = all - reaches - 1;
652 printf("Bisecting: %d revisions left to test after this "
653 "(roughly %d steps)\n", nr, estimate_bisect_steps(all));
654
655 return bisect_checkout(bisect_rev_hex);
656}
657