git-rev-list --bisect: optimization
[git] / builtin-rev-list.c
1 #include "cache.h"
2 #include "refs.h"
3 #include "tag.h"
4 #include "commit.h"
5 #include "tree.h"
6 #include "blob.h"
7 #include "tree-walk.h"
8 #include "diff.h"
9 #include "revision.h"
10 #include "list-objects.h"
11 #include "builtin.h"
12
13 /* bits #0-15 in revision.h */
14
15 #define COUNTED         (1u<<16)
16
17 static const char rev_list_usage[] =
18 "git-rev-list [OPTION] <commit-id>... [ -- paths... ]\n"
19 "  limiting output:\n"
20 "    --max-count=nr\n"
21 "    --max-age=epoch\n"
22 "    --min-age=epoch\n"
23 "    --sparse\n"
24 "    --no-merges\n"
25 "    --remove-empty\n"
26 "    --all\n"
27 "    --stdin\n"
28 "  ordering output:\n"
29 "    --topo-order\n"
30 "    --date-order\n"
31 "  formatting output:\n"
32 "    --parents\n"
33 "    --objects | --objects-edge\n"
34 "    --unpacked\n"
35 "    --header | --pretty\n"
36 "    --abbrev=nr | --no-abbrev\n"
37 "    --abbrev-commit\n"
38 "  special purpose:\n"
39 "    --bisect\n"
40 "    --bisect-vars"
41 ;
42
43 static struct rev_info revs;
44
45 static int bisect_list;
46 static int show_timestamp;
47 static int hdr_termination;
48 static const char *header_prefix;
49
50 static void show_commit(struct commit *commit)
51 {
52         if (show_timestamp)
53                 printf("%lu ", commit->date);
54         if (header_prefix)
55                 fputs(header_prefix, stdout);
56         if (commit->object.flags & BOUNDARY)
57                 putchar('-');
58         else if (revs.left_right) {
59                 if (commit->object.flags & SYMMETRIC_LEFT)
60                         putchar('<');
61                 else
62                         putchar('>');
63         }
64         if (revs.abbrev_commit && revs.abbrev)
65                 fputs(find_unique_abbrev(commit->object.sha1, revs.abbrev),
66                       stdout);
67         else
68                 fputs(sha1_to_hex(commit->object.sha1), stdout);
69         if (revs.parents) {
70                 struct commit_list *parents = commit->parents;
71                 while (parents) {
72                         struct object *o = &(parents->item->object);
73                         parents = parents->next;
74                         if (o->flags & TMP_MARK)
75                                 continue;
76                         printf(" %s", sha1_to_hex(o->sha1));
77                         o->flags |= TMP_MARK;
78                 }
79                 /* TMP_MARK is a general purpose flag that can
80                  * be used locally, but the user should clean
81                  * things up after it is done with them.
82                  */
83                 for (parents = commit->parents;
84                      parents;
85                      parents = parents->next)
86                         parents->item->object.flags &= ~TMP_MARK;
87         }
88         if (revs.commit_format == CMIT_FMT_ONELINE)
89                 putchar(' ');
90         else
91                 putchar('\n');
92
93         if (revs.verbose_header) {
94                 static char pretty_header[16384];
95                 pretty_print_commit(revs.commit_format, commit, ~0,
96                                     pretty_header, sizeof(pretty_header),
97                                     revs.abbrev, NULL, NULL, revs.relative_date);
98                 printf("%s%c", pretty_header, hdr_termination);
99         }
100         fflush(stdout);
101         if (commit->parents) {
102                 free_commit_list(commit->parents);
103                 commit->parents = NULL;
104         }
105         free(commit->buffer);
106         commit->buffer = NULL;
107 }
108
109 static void show_object(struct object_array_entry *p)
110 {
111         /* An object with name "foo\n0000000..." can be used to
112          * confuse downstream git-pack-objects very badly.
113          */
114         const char *ep = strchr(p->name, '\n');
115         if (ep) {
116                 printf("%s %.*s\n", sha1_to_hex(p->item->sha1),
117                        (int) (ep - p->name),
118                        p->name);
119         }
120         else
121                 printf("%s %s\n", sha1_to_hex(p->item->sha1), p->name);
122 }
123
124 static void show_edge(struct commit *commit)
125 {
126         printf("-%s\n", sha1_to_hex(commit->object.sha1));
127 }
128
129 /*
130  * This is a truly stupid algorithm, but it's only
131  * used for bisection, and we just don't care enough.
132  *
133  * We care just barely enough to avoid recursing for
134  * non-merge entries.
135  */
136 static int count_distance(struct commit_list *entry)
137 {
138         int nr = 0;
139
140         while (entry) {
141                 struct commit *commit = entry->item;
142                 struct commit_list *p;
143
144                 if (commit->object.flags & (UNINTERESTING | COUNTED))
145                         break;
146                 if (!revs.prune_fn || (commit->object.flags & TREECHANGE))
147                         nr++;
148                 commit->object.flags |= COUNTED;
149                 p = commit->parents;
150                 entry = p;
151                 if (p) {
152                         p = p->next;
153                         while (p) {
154                                 nr += count_distance(p);
155                                 p = p->next;
156                         }
157                 }
158         }
159
160         return nr;
161 }
162
163 static void clear_distance(struct commit_list *list)
164 {
165         while (list) {
166                 struct commit *commit = list->item;
167                 commit->object.flags &= ~COUNTED;
168                 list = list->next;
169         }
170 }
171
172 static struct commit_list *find_bisection(struct commit_list *list,
173                                           int *reaches, int *all)
174 {
175         int nr, closest;
176         struct commit_list *p, *best;
177
178         nr = 0;
179         p = list;
180         while (p) {
181                 if (!revs.prune_fn || (p->item->object.flags & TREECHANGE))
182                         nr++;
183                 p = p->next;
184         }
185         *all = nr;
186         closest = 0;
187         best = list;
188
189         for (p = list; p; p = p->next) {
190                 int distance, reach;
191
192                 if (revs.prune_fn && !(p->item->object.flags & TREECHANGE))
193                         continue;
194
195                 distance = reach = count_distance(p);
196                 clear_distance(list);
197                 if (nr - distance < distance)
198                         distance = nr - distance;
199                 if (distance > closest) {
200                         best = p;
201                         *reaches = reach;
202                         closest = distance;
203                 }
204         }
205         if (best)
206                 best->next = NULL;
207         return best;
208 }
209
210 static inline int commit_interesting(struct commit_list *elem)
211 {
212         unsigned flags = elem->item->object.flags;
213         if (flags & UNINTERESTING)
214                 return 0;
215         return (!revs.prune_fn || (flags & TREECHANGE));
216 }
217
218 static inline int weight(struct commit_list *elem)
219 {
220         return *((int*)(elem->item->util));
221 }
222
223 static inline void weight_set(struct commit_list *elem, int weight)
224 {
225         *((int*)(elem->item->util)) = weight;
226 }
227
228 static int count_interesting_parents(struct commit_list *elem)
229 {
230         int cnt = 0;
231         if (!elem->item->parents)
232                 return cnt;
233         for (elem = elem->item->parents; elem; elem = elem->next) {
234                 if (commit_interesting(elem))
235                         cnt++;
236         }
237         return cnt;
238 }
239
240 static struct commit_list *find_bisection_2(struct commit_list *list,
241                                             int *reaches, int *all)
242 {
243         int n, nr, counted, distance;
244         struct commit_list *p, *best;
245         int *weights;
246
247         for (nr = 0, p = list; p; p = p->next) {
248                 if (commit_interesting(p))
249                         nr++;
250         }
251         *all = nr;
252         weights = xcalloc(nr, sizeof(int*));
253         counted = 0;
254
255         for (n = 0, p = list; p; p = p->next) {
256                 if (!commit_interesting(p))
257                         continue;
258                 if (commit_interesting(p)) {
259                         /*
260                          * positive weight is the number of interesting
261                          * commits it can reach, including itself.
262                          * weight = 0 means it has one parent and
263                          * its distance is unknown.
264                          * weight < 0 means it has more than one
265                          * parent and its distance is unknown.
266                          */
267                         p->item->util = &weights[n++];
268                         switch (count_interesting_parents(p)) {
269                         case 0:
270                                 weight_set(p, 1);
271                                 counted++;
272                                 break;
273                         case 1:
274                                 weight_set(p, 0);
275                                 break;
276                         default:
277                                 weight_set(p, -1);
278                                 break;
279                         }
280                 }
281         }
282
283         /*
284          * If you have only one parent in the resulting set
285          * then you can reach one commit more than that parent
286          * can reach.  So we do not have to run the expensive
287          * count_distance() for single strand of pearls.
288          *
289          * However, if you have more than one parents, you cannot
290          * just add their distance and one for yourself, since
291          * they usually reach the same ancestor and you would
292          * end up counting them twice that way.
293          *
294          * So we will first count distance of merges the usual
295          * way, and then fill the blanks using cheaper algorithm.
296          */
297         for (p = list; p; p = p->next) {
298                 if (!commit_interesting(p))
299                         continue;
300                 n = weight(p);
301                 if (0 <= n)
302                         continue;
303                 distance = count_distance(p);
304                 clear_distance(p);
305                 weight_set(p, distance);
306
307                 /* Does it happen to be at exactly half-way? */
308                 distance *= 2;
309                 if (nr == distance || (nr+1) == distance) {
310                         p->next = NULL;
311                         *reaches = weight(p);
312                         free(weights);
313                         return p;
314                 }
315                 counted++;
316         }
317
318         while (counted < nr) {
319                 for (p = list; p; p = p->next) {
320                         struct commit_list *q;
321
322                         if (!commit_interesting(p) || 0 < weight(p))
323                                 continue;
324                         for (q = p->item->parents; q; q = q->next)
325                                 if (commit_interesting(q) && 0 < weight(q))
326                                         break;
327                         if (!q)
328                                 continue;
329                         weight_set(p, weight(q)+1);
330                         counted++;
331
332                         /* Does it happen to be at exactly half-way? */
333                         distance = weight(p) * 2;
334                         if (nr == distance || (nr+1) == distance) {
335                                 p->next = NULL;
336                                 *reaches = weight(p);
337                                 free(weights);
338                                 return p;
339                         }
340                 }
341         }
342
343         /* Then find the best one */
344         counted = 0;
345         best = list;
346         for (p = list; p; p = p->next) {
347                 if (!commit_interesting(p))
348                         continue;
349                 distance = weight(p);
350                 if (nr - distance < distance)
351                         distance = nr - distance;
352                 if (distance > counted) {
353                         best = p;
354                         counted = distance;
355                         *reaches = weight(p);
356                 }
357         }
358         if (best)
359                 best->next = NULL;
360         free(weights);
361         return best;
362 }
363
364 static void read_revisions_from_stdin(struct rev_info *revs)
365 {
366         char line[1000];
367
368         while (fgets(line, sizeof(line), stdin) != NULL) {
369                 int len = strlen(line);
370                 if (line[len - 1] == '\n')
371                         line[--len] = 0;
372                 if (!len)
373                         break;
374                 if (line[0] == '-')
375                         die("options not supported in --stdin mode");
376                 if (handle_revision_arg(line, revs, 0, 1))
377                         die("bad revision '%s'", line);
378         }
379 }
380
381 int cmd_rev_list(int argc, const char **argv, const char *prefix)
382 {
383         struct commit_list *list;
384         int i;
385         int read_from_stdin = 0;
386         int bisect_show_vars = 0;
387
388         git_config(git_default_config);
389         init_revisions(&revs, prefix);
390         revs.abbrev = 0;
391         revs.commit_format = CMIT_FMT_UNSPECIFIED;
392         argc = setup_revisions(argc, argv, &revs, NULL);
393
394         for (i = 1 ; i < argc; i++) {
395                 const char *arg = argv[i];
396
397                 if (!strcmp(arg, "--header")) {
398                         revs.verbose_header = 1;
399                         continue;
400                 }
401                 if (!strcmp(arg, "--timestamp")) {
402                         show_timestamp = 1;
403                         continue;
404                 }
405                 if (!strcmp(arg, "--bisect")) {
406                         bisect_list = 1;
407                         continue;
408                 }
409                 if (!strcmp(arg, "--bisect-vars")) {
410                         bisect_list = 1;
411                         bisect_show_vars = 1;
412                         continue;
413                 }
414                 if (!strcmp(arg, "--stdin")) {
415                         if (read_from_stdin++)
416                                 die("--stdin given twice?");
417                         read_revisions_from_stdin(&revs);
418                         continue;
419                 }
420                 usage(rev_list_usage);
421
422         }
423         if (revs.commit_format != CMIT_FMT_UNSPECIFIED) {
424                 /* The command line has a --pretty  */
425                 hdr_termination = '\n';
426                 if (revs.commit_format == CMIT_FMT_ONELINE)
427                         header_prefix = "";
428                 else
429                         header_prefix = "commit ";
430         }
431         else if (revs.verbose_header)
432                 /* Only --header was specified */
433                 revs.commit_format = CMIT_FMT_RAW;
434
435         list = revs.commits;
436
437         if ((!list &&
438              (!(revs.tag_objects||revs.tree_objects||revs.blob_objects) &&
439               !revs.pending.nr)) ||
440             revs.diff)
441                 usage(rev_list_usage);
442
443         save_commit_buffer = revs.verbose_header || revs.grep_filter;
444         track_object_refs = 0;
445         if (bisect_list)
446                 revs.limited = 1;
447
448         prepare_revision_walk(&revs);
449         if (revs.tree_objects)
450                 mark_edges_uninteresting(revs.commits, &revs, show_edge);
451
452         if (bisect_list) {
453                 int reaches = reaches, all = all;
454
455                 if (!revs.prune_fn)
456                         revs.commits = find_bisection_2(revs.commits,
457                                                         &reaches, &all);
458                 else
459                         revs.commits = find_bisection(revs.commits,
460                                                       &reaches, &all);
461                 if (bisect_show_vars) {
462                         int cnt;
463                         if (!revs.commits)
464                                 return 1;
465                         /*
466                          * revs.commits can reach "reaches" commits among
467                          * "all" commits.  If it is good, then there are
468                          * (all-reaches) commits left to be bisected.
469                          * On the other hand, if it is bad, then the set
470                          * to bisect is "reaches".
471                          * A bisect set of size N has (N-1) commits further
472                          * to test, as we already know one bad one.
473                          */
474                         cnt = all-reaches;
475                         if (cnt < reaches)
476                                 cnt = reaches;
477                         printf("bisect_rev=%s\n"
478                                "bisect_nr=%d\n"
479                                "bisect_good=%d\n"
480                                "bisect_bad=%d\n"
481                                "bisect_all=%d\n",
482                                sha1_to_hex(revs.commits->item->object.sha1),
483                                cnt - 1,
484                                all - reaches - 1,
485                                reaches - 1,
486                                all);
487                         return 0;
488                 }
489         }
490
491         traverse_commit_list(&revs, show_commit, show_object);
492
493         return 0;
494 }