Merge branch 'jk/sha1dc'
[git] / blame.c
1 #include "cache.h"
2 #include "refs.h"
3 #include "object-store.h"
4 #include "cache-tree.h"
5 #include "mergesort.h"
6 #include "diff.h"
7 #include "diffcore.h"
8 #include "tag.h"
9 #include "blame.h"
10 #include "alloc.h"
11 #include "commit-slab.h"
12
13 define_commit_slab(blame_suspects, struct blame_origin *);
14 static struct blame_suspects blame_suspects;
15
16 struct blame_origin *get_blame_suspects(struct commit *commit)
17 {
18         struct blame_origin **result;
19
20         result = blame_suspects_peek(&blame_suspects, commit);
21
22         return result ? *result : NULL;
23 }
24
25 static void set_blame_suspects(struct commit *commit, struct blame_origin *origin)
26 {
27         *blame_suspects_at(&blame_suspects, commit) = origin;
28 }
29
30 void blame_origin_decref(struct blame_origin *o)
31 {
32         if (o && --o->refcnt <= 0) {
33                 struct blame_origin *p, *l = NULL;
34                 if (o->previous)
35                         blame_origin_decref(o->previous);
36                 free(o->file.ptr);
37                 /* Should be present exactly once in commit chain */
38                 for (p = get_blame_suspects(o->commit); p; l = p, p = p->next) {
39                         if (p == o) {
40                                 if (l)
41                                         l->next = p->next;
42                                 else
43                                         set_blame_suspects(o->commit, p->next);
44                                 free(o);
45                                 return;
46                         }
47                 }
48                 die("internal error in blame_origin_decref");
49         }
50 }
51
52 /*
53  * Given a commit and a path in it, create a new origin structure.
54  * The callers that add blame to the scoreboard should use
55  * get_origin() to obtain shared, refcounted copy instead of calling
56  * this function directly.
57  */
58 static struct blame_origin *make_origin(struct commit *commit, const char *path)
59 {
60         struct blame_origin *o;
61         FLEX_ALLOC_STR(o, path, path);
62         o->commit = commit;
63         o->refcnt = 1;
64         o->next = get_blame_suspects(commit);
65         set_blame_suspects(commit, o);
66         return o;
67 }
68
69 /*
70  * Locate an existing origin or create a new one.
71  * This moves the origin to front position in the commit util list.
72  */
73 static struct blame_origin *get_origin(struct commit *commit, const char *path)
74 {
75         struct blame_origin *o, *l;
76
77         for (o = get_blame_suspects(commit), l = NULL; o; l = o, o = o->next) {
78                 if (!strcmp(o->path, path)) {
79                         /* bump to front */
80                         if (l) {
81                                 l->next = o->next;
82                                 o->next = get_blame_suspects(commit);
83                                 set_blame_suspects(commit, o);
84                         }
85                         return blame_origin_incref(o);
86                 }
87         }
88         return make_origin(commit, path);
89 }
90
91
92
93 static void verify_working_tree_path(struct repository *r,
94                                      struct commit *work_tree, const char *path)
95 {
96         struct commit_list *parents;
97         int pos;
98
99         for (parents = work_tree->parents; parents; parents = parents->next) {
100                 const struct object_id *commit_oid = &parents->item->object.oid;
101                 struct object_id blob_oid;
102                 unsigned mode;
103
104                 if (!get_tree_entry(commit_oid, path, &blob_oid, &mode) &&
105                     oid_object_info(r, &blob_oid, NULL) == OBJ_BLOB)
106                         return;
107         }
108
109         pos = index_name_pos(r->index, path, strlen(path));
110         if (pos >= 0)
111                 ; /* path is in the index */
112         else if (-1 - pos < r->index->cache_nr &&
113                  !strcmp(r->index->cache[-1 - pos]->name, path))
114                 ; /* path is in the index, unmerged */
115         else
116                 die("no such path '%s' in HEAD", path);
117 }
118
119 static struct commit_list **append_parent(struct repository *r,
120                                           struct commit_list **tail,
121                                           const struct object_id *oid)
122 {
123         struct commit *parent;
124
125         parent = lookup_commit_reference(r, oid);
126         if (!parent)
127                 die("no such commit %s", oid_to_hex(oid));
128         return &commit_list_insert(parent, tail)->next;
129 }
130
131 static void append_merge_parents(struct repository *r,
132                                  struct commit_list **tail)
133 {
134         int merge_head;
135         struct strbuf line = STRBUF_INIT;
136
137         merge_head = open(git_path_merge_head(r), O_RDONLY);
138         if (merge_head < 0) {
139                 if (errno == ENOENT)
140                         return;
141                 die("cannot open '%s' for reading",
142                     git_path_merge_head(r));
143         }
144
145         while (!strbuf_getwholeline_fd(&line, merge_head, '\n')) {
146                 struct object_id oid;
147                 if (line.len < GIT_SHA1_HEXSZ || get_oid_hex(line.buf, &oid))
148                         die("unknown line in '%s': %s",
149                             git_path_merge_head(r), line.buf);
150                 tail = append_parent(r, tail, &oid);
151         }
152         close(merge_head);
153         strbuf_release(&line);
154 }
155
156 /*
157  * This isn't as simple as passing sb->buf and sb->len, because we
158  * want to transfer ownership of the buffer to the commit (so we
159  * must use detach).
160  */
161 static void set_commit_buffer_from_strbuf(struct repository *r,
162                                           struct commit *c,
163                                           struct strbuf *sb)
164 {
165         size_t len;
166         void *buf = strbuf_detach(sb, &len);
167         set_commit_buffer(r, c, buf, len);
168 }
169
170 /*
171  * Prepare a dummy commit that represents the work tree (or staged) item.
172  * Note that annotating work tree item never works in the reverse.
173  */
174 static struct commit *fake_working_tree_commit(struct repository *r,
175                                                struct diff_options *opt,
176                                                const char *path,
177                                                const char *contents_from)
178 {
179         struct commit *commit;
180         struct blame_origin *origin;
181         struct commit_list **parent_tail, *parent;
182         struct object_id head_oid;
183         struct strbuf buf = STRBUF_INIT;
184         const char *ident;
185         time_t now;
186         int len;
187         struct cache_entry *ce;
188         unsigned mode;
189         struct strbuf msg = STRBUF_INIT;
190
191         repo_read_index(r);
192         time(&now);
193         commit = alloc_commit_node(r);
194         commit->object.parsed = 1;
195         commit->date = now;
196         parent_tail = &commit->parents;
197
198         if (!resolve_ref_unsafe("HEAD", RESOLVE_REF_READING, &head_oid, NULL))
199                 die("no such ref: HEAD");
200
201         parent_tail = append_parent(r, parent_tail, &head_oid);
202         append_merge_parents(r, parent_tail);
203         verify_working_tree_path(r, commit, path);
204
205         origin = make_origin(commit, path);
206
207         ident = fmt_ident("Not Committed Yet", "not.committed.yet",
208                         WANT_BLANK_IDENT, NULL, 0);
209         strbuf_addstr(&msg, "tree 0000000000000000000000000000000000000000\n");
210         for (parent = commit->parents; parent; parent = parent->next)
211                 strbuf_addf(&msg, "parent %s\n",
212                             oid_to_hex(&parent->item->object.oid));
213         strbuf_addf(&msg,
214                     "author %s\n"
215                     "committer %s\n\n"
216                     "Version of %s from %s\n",
217                     ident, ident, path,
218                     (!contents_from ? path :
219                      (!strcmp(contents_from, "-") ? "standard input" : contents_from)));
220         set_commit_buffer_from_strbuf(r, commit, &msg);
221
222         if (!contents_from || strcmp("-", contents_from)) {
223                 struct stat st;
224                 const char *read_from;
225                 char *buf_ptr;
226                 unsigned long buf_len;
227
228                 if (contents_from) {
229                         if (stat(contents_from, &st) < 0)
230                                 die_errno("Cannot stat '%s'", contents_from);
231                         read_from = contents_from;
232                 }
233                 else {
234                         if (lstat(path, &st) < 0)
235                                 die_errno("Cannot lstat '%s'", path);
236                         read_from = path;
237                 }
238                 mode = canon_mode(st.st_mode);
239
240                 switch (st.st_mode & S_IFMT) {
241                 case S_IFREG:
242                         if (opt->flags.allow_textconv &&
243                             textconv_object(r, read_from, mode, &null_oid, 0, &buf_ptr, &buf_len))
244                                 strbuf_attach(&buf, buf_ptr, buf_len, buf_len + 1);
245                         else if (strbuf_read_file(&buf, read_from, st.st_size) != st.st_size)
246                                 die_errno("cannot open or read '%s'", read_from);
247                         break;
248                 case S_IFLNK:
249                         if (strbuf_readlink(&buf, read_from, st.st_size) < 0)
250                                 die_errno("cannot readlink '%s'", read_from);
251                         break;
252                 default:
253                         die("unsupported file type %s", read_from);
254                 }
255         }
256         else {
257                 /* Reading from stdin */
258                 mode = 0;
259                 if (strbuf_read(&buf, 0, 0) < 0)
260                         die_errno("failed to read from stdin");
261         }
262         convert_to_git(r->index, path, buf.buf, buf.len, &buf, 0);
263         origin->file.ptr = buf.buf;
264         origin->file.size = buf.len;
265         pretend_object_file(buf.buf, buf.len, OBJ_BLOB, &origin->blob_oid);
266
267         /*
268          * Read the current index, replace the path entry with
269          * origin->blob_sha1 without mucking with its mode or type
270          * bits; we are not going to write this index out -- we just
271          * want to run "diff-index --cached".
272          */
273         discard_index(r->index);
274         repo_read_index(r);
275
276         len = strlen(path);
277         if (!mode) {
278                 int pos = index_name_pos(r->index, path, len);
279                 if (0 <= pos)
280                         mode = r->index->cache[pos]->ce_mode;
281                 else
282                         /* Let's not bother reading from HEAD tree */
283                         mode = S_IFREG | 0644;
284         }
285         ce = make_empty_cache_entry(r->index, len);
286         oidcpy(&ce->oid, &origin->blob_oid);
287         memcpy(ce->name, path, len);
288         ce->ce_flags = create_ce_flags(0);
289         ce->ce_namelen = len;
290         ce->ce_mode = create_ce_mode(mode);
291         add_index_entry(r->index, ce,
292                         ADD_CACHE_OK_TO_ADD | ADD_CACHE_OK_TO_REPLACE);
293
294         cache_tree_invalidate_path(r->index, path);
295
296         return commit;
297 }
298
299
300
301 static int diff_hunks(mmfile_t *file_a, mmfile_t *file_b,
302                       xdl_emit_hunk_consume_func_t hunk_func, void *cb_data, int xdl_opts)
303 {
304         xpparam_t xpp = {0};
305         xdemitconf_t xecfg = {0};
306         xdemitcb_t ecb = {NULL};
307
308         xpp.flags = xdl_opts;
309         xecfg.hunk_func = hunk_func;
310         ecb.priv = cb_data;
311         return xdi_diff(file_a, file_b, &xpp, &xecfg, &ecb);
312 }
313
314 /*
315  * Given an origin, prepare mmfile_t structure to be used by the
316  * diff machinery
317  */
318 static void fill_origin_blob(struct diff_options *opt,
319                              struct blame_origin *o, mmfile_t *file, int *num_read_blob)
320 {
321         if (!o->file.ptr) {
322                 enum object_type type;
323                 unsigned long file_size;
324
325                 (*num_read_blob)++;
326                 if (opt->flags.allow_textconv &&
327                     textconv_object(opt->repo, o->path, o->mode,
328                                     &o->blob_oid, 1, &file->ptr, &file_size))
329                         ;
330                 else
331                         file->ptr = read_object_file(&o->blob_oid, &type,
332                                                      &file_size);
333                 file->size = file_size;
334
335                 if (!file->ptr)
336                         die("Cannot read blob %s for path %s",
337                             oid_to_hex(&o->blob_oid),
338                             o->path);
339                 o->file = *file;
340         }
341         else
342                 *file = o->file;
343 }
344
345 static void drop_origin_blob(struct blame_origin *o)
346 {
347         FREE_AND_NULL(o->file.ptr);
348 }
349
350 /*
351  * Any merge of blames happens on lists of blames that arrived via
352  * different parents in a single suspect.  In this case, we want to
353  * sort according to the suspect line numbers as opposed to the final
354  * image line numbers.  The function body is somewhat longish because
355  * it avoids unnecessary writes.
356  */
357
358 static struct blame_entry *blame_merge(struct blame_entry *list1,
359                                        struct blame_entry *list2)
360 {
361         struct blame_entry *p1 = list1, *p2 = list2,
362                 **tail = &list1;
363
364         if (!p1)
365                 return p2;
366         if (!p2)
367                 return p1;
368
369         if (p1->s_lno <= p2->s_lno) {
370                 do {
371                         tail = &p1->next;
372                         if ((p1 = *tail) == NULL) {
373                                 *tail = p2;
374                                 return list1;
375                         }
376                 } while (p1->s_lno <= p2->s_lno);
377         }
378         for (;;) {
379                 *tail = p2;
380                 do {
381                         tail = &p2->next;
382                         if ((p2 = *tail) == NULL)  {
383                                 *tail = p1;
384                                 return list1;
385                         }
386                 } while (p1->s_lno > p2->s_lno);
387                 *tail = p1;
388                 do {
389                         tail = &p1->next;
390                         if ((p1 = *tail) == NULL) {
391                                 *tail = p2;
392                                 return list1;
393                         }
394                 } while (p1->s_lno <= p2->s_lno);
395         }
396 }
397
398 static void *get_next_blame(const void *p)
399 {
400         return ((struct blame_entry *)p)->next;
401 }
402
403 static void set_next_blame(void *p1, void *p2)
404 {
405         ((struct blame_entry *)p1)->next = p2;
406 }
407
408 /*
409  * Final image line numbers are all different, so we don't need a
410  * three-way comparison here.
411  */
412
413 static int compare_blame_final(const void *p1, const void *p2)
414 {
415         return ((struct blame_entry *)p1)->lno > ((struct blame_entry *)p2)->lno
416                 ? 1 : -1;
417 }
418
419 static int compare_blame_suspect(const void *p1, const void *p2)
420 {
421         const struct blame_entry *s1 = p1, *s2 = p2;
422         /*
423          * to allow for collating suspects, we sort according to the
424          * respective pointer value as the primary sorting criterion.
425          * The actual relation is pretty unimportant as long as it
426          * establishes a total order.  Comparing as integers gives us
427          * that.
428          */
429         if (s1->suspect != s2->suspect)
430                 return (intptr_t)s1->suspect > (intptr_t)s2->suspect ? 1 : -1;
431         if (s1->s_lno == s2->s_lno)
432                 return 0;
433         return s1->s_lno > s2->s_lno ? 1 : -1;
434 }
435
436 void blame_sort_final(struct blame_scoreboard *sb)
437 {
438         sb->ent = llist_mergesort(sb->ent, get_next_blame, set_next_blame,
439                                   compare_blame_final);
440 }
441
442 static int compare_commits_by_reverse_commit_date(const void *a,
443                                                   const void *b,
444                                                   void *c)
445 {
446         return -compare_commits_by_commit_date(a, b, c);
447 }
448
449 /*
450  * For debugging -- origin is refcounted, and this asserts that
451  * we do not underflow.
452  */
453 static void sanity_check_refcnt(struct blame_scoreboard *sb)
454 {
455         int baa = 0;
456         struct blame_entry *ent;
457
458         for (ent = sb->ent; ent; ent = ent->next) {
459                 /* Nobody should have zero or negative refcnt */
460                 if (ent->suspect->refcnt <= 0) {
461                         fprintf(stderr, "%s in %s has negative refcnt %d\n",
462                                 ent->suspect->path,
463                                 oid_to_hex(&ent->suspect->commit->object.oid),
464                                 ent->suspect->refcnt);
465                         baa = 1;
466                 }
467         }
468         if (baa)
469                 sb->on_sanity_fail(sb, baa);
470 }
471
472 /*
473  * If two blame entries that are next to each other came from
474  * contiguous lines in the same origin (i.e. <commit, path> pair),
475  * merge them together.
476  */
477 void blame_coalesce(struct blame_scoreboard *sb)
478 {
479         struct blame_entry *ent, *next;
480
481         for (ent = sb->ent; ent && (next = ent->next); ent = next) {
482                 if (ent->suspect == next->suspect &&
483                     ent->s_lno + ent->num_lines == next->s_lno) {
484                         ent->num_lines += next->num_lines;
485                         ent->next = next->next;
486                         blame_origin_decref(next->suspect);
487                         free(next);
488                         ent->score = 0;
489                         next = ent; /* again */
490                 }
491         }
492
493         if (sb->debug) /* sanity */
494                 sanity_check_refcnt(sb);
495 }
496
497 /*
498  * Merge the given sorted list of blames into a preexisting origin.
499  * If there were no previous blames to that commit, it is entered into
500  * the commit priority queue of the score board.
501  */
502
503 static void queue_blames(struct blame_scoreboard *sb, struct blame_origin *porigin,
504                          struct blame_entry *sorted)
505 {
506         if (porigin->suspects)
507                 porigin->suspects = blame_merge(porigin->suspects, sorted);
508         else {
509                 struct blame_origin *o;
510                 for (o = get_blame_suspects(porigin->commit); o; o = o->next) {
511                         if (o->suspects) {
512                                 porigin->suspects = sorted;
513                                 return;
514                         }
515                 }
516                 porigin->suspects = sorted;
517                 prio_queue_put(&sb->commits, porigin->commit);
518         }
519 }
520
521 /*
522  * Fill the blob_sha1 field of an origin if it hasn't, so that later
523  * call to fill_origin_blob() can use it to locate the data.  blob_sha1
524  * for an origin is also used to pass the blame for the entire file to
525  * the parent to detect the case where a child's blob is identical to
526  * that of its parent's.
527  *
528  * This also fills origin->mode for corresponding tree path.
529  */
530 static int fill_blob_sha1_and_mode(struct repository *r,
531                                    struct blame_origin *origin)
532 {
533         if (!is_null_oid(&origin->blob_oid))
534                 return 0;
535         if (get_tree_entry(&origin->commit->object.oid, origin->path, &origin->blob_oid, &origin->mode))
536                 goto error_out;
537         if (oid_object_info(r, &origin->blob_oid, NULL) != OBJ_BLOB)
538                 goto error_out;
539         return 0;
540  error_out:
541         oidclr(&origin->blob_oid);
542         origin->mode = S_IFINVALID;
543         return -1;
544 }
545
546 /*
547  * We have an origin -- check if the same path exists in the
548  * parent and return an origin structure to represent it.
549  */
550 static struct blame_origin *find_origin(struct repository *r,
551                                         struct commit *parent,
552                                         struct blame_origin *origin)
553 {
554         struct blame_origin *porigin;
555         struct diff_options diff_opts;
556         const char *paths[2];
557
558         /* First check any existing origins */
559         for (porigin = get_blame_suspects(parent); porigin; porigin = porigin->next)
560                 if (!strcmp(porigin->path, origin->path)) {
561                         /*
562                          * The same path between origin and its parent
563                          * without renaming -- the most common case.
564                          */
565                         return blame_origin_incref (porigin);
566                 }
567
568         /* See if the origin->path is different between parent
569          * and origin first.  Most of the time they are the
570          * same and diff-tree is fairly efficient about this.
571          */
572         repo_diff_setup(r, &diff_opts);
573         diff_opts.flags.recursive = 1;
574         diff_opts.detect_rename = 0;
575         diff_opts.output_format = DIFF_FORMAT_NO_OUTPUT;
576         paths[0] = origin->path;
577         paths[1] = NULL;
578
579         parse_pathspec(&diff_opts.pathspec,
580                        PATHSPEC_ALL_MAGIC & ~PATHSPEC_LITERAL,
581                        PATHSPEC_LITERAL_PATH, "", paths);
582         diff_setup_done(&diff_opts);
583
584         if (is_null_oid(&origin->commit->object.oid))
585                 do_diff_cache(get_commit_tree_oid(parent), &diff_opts);
586         else
587                 diff_tree_oid(get_commit_tree_oid(parent),
588                               get_commit_tree_oid(origin->commit),
589                               "", &diff_opts);
590         diffcore_std(&diff_opts);
591
592         if (!diff_queued_diff.nr) {
593                 /* The path is the same as parent */
594                 porigin = get_origin(parent, origin->path);
595                 oidcpy(&porigin->blob_oid, &origin->blob_oid);
596                 porigin->mode = origin->mode;
597         } else {
598                 /*
599                  * Since origin->path is a pathspec, if the parent
600                  * commit had it as a directory, we will see a whole
601                  * bunch of deletion of files in the directory that we
602                  * do not care about.
603                  */
604                 int i;
605                 struct diff_filepair *p = NULL;
606                 for (i = 0; i < diff_queued_diff.nr; i++) {
607                         const char *name;
608                         p = diff_queued_diff.queue[i];
609                         name = p->one->path ? p->one->path : p->two->path;
610                         if (!strcmp(name, origin->path))
611                                 break;
612                 }
613                 if (!p)
614                         die("internal error in blame::find_origin");
615                 switch (p->status) {
616                 default:
617                         die("internal error in blame::find_origin (%c)",
618                             p->status);
619                 case 'M':
620                         porigin = get_origin(parent, origin->path);
621                         oidcpy(&porigin->blob_oid, &p->one->oid);
622                         porigin->mode = p->one->mode;
623                         break;
624                 case 'A':
625                 case 'T':
626                         /* Did not exist in parent, or type changed */
627                         break;
628                 }
629         }
630         diff_flush(&diff_opts);
631         clear_pathspec(&diff_opts.pathspec);
632         return porigin;
633 }
634
635 /*
636  * We have an origin -- find the path that corresponds to it in its
637  * parent and return an origin structure to represent it.
638  */
639 static struct blame_origin *find_rename(struct repository *r,
640                                         struct commit *parent,
641                                         struct blame_origin *origin)
642 {
643         struct blame_origin *porigin = NULL;
644         struct diff_options diff_opts;
645         int i;
646
647         repo_diff_setup(r, &diff_opts);
648         diff_opts.flags.recursive = 1;
649         diff_opts.detect_rename = DIFF_DETECT_RENAME;
650         diff_opts.output_format = DIFF_FORMAT_NO_OUTPUT;
651         diff_opts.single_follow = origin->path;
652         diff_setup_done(&diff_opts);
653
654         if (is_null_oid(&origin->commit->object.oid))
655                 do_diff_cache(get_commit_tree_oid(parent), &diff_opts);
656         else
657                 diff_tree_oid(get_commit_tree_oid(parent),
658                               get_commit_tree_oid(origin->commit),
659                               "", &diff_opts);
660         diffcore_std(&diff_opts);
661
662         for (i = 0; i < diff_queued_diff.nr; i++) {
663                 struct diff_filepair *p = diff_queued_diff.queue[i];
664                 if ((p->status == 'R' || p->status == 'C') &&
665                     !strcmp(p->two->path, origin->path)) {
666                         porigin = get_origin(parent, p->one->path);
667                         oidcpy(&porigin->blob_oid, &p->one->oid);
668                         porigin->mode = p->one->mode;
669                         break;
670                 }
671         }
672         diff_flush(&diff_opts);
673         clear_pathspec(&diff_opts.pathspec);
674         return porigin;
675 }
676
677 /*
678  * Append a new blame entry to a given output queue.
679  */
680 static void add_blame_entry(struct blame_entry ***queue,
681                             const struct blame_entry *src)
682 {
683         struct blame_entry *e = xmalloc(sizeof(*e));
684         memcpy(e, src, sizeof(*e));
685         blame_origin_incref(e->suspect);
686
687         e->next = **queue;
688         **queue = e;
689         *queue = &e->next;
690 }
691
692 /*
693  * src typically is on-stack; we want to copy the information in it to
694  * a malloced blame_entry that gets added to the given queue.  The
695  * origin of dst loses a refcnt.
696  */
697 static void dup_entry(struct blame_entry ***queue,
698                       struct blame_entry *dst, struct blame_entry *src)
699 {
700         blame_origin_incref(src->suspect);
701         blame_origin_decref(dst->suspect);
702         memcpy(dst, src, sizeof(*src));
703         dst->next = **queue;
704         **queue = dst;
705         *queue = &dst->next;
706 }
707
708 const char *blame_nth_line(struct blame_scoreboard *sb, long lno)
709 {
710         return sb->final_buf + sb->lineno[lno];
711 }
712
713 /*
714  * It is known that lines between tlno to same came from parent, and e
715  * has an overlap with that range.  it also is known that parent's
716  * line plno corresponds to e's line tlno.
717  *
718  *                <---- e ----->
719  *                   <------>
720  *                   <------------>
721  *             <------------>
722  *             <------------------>
723  *
724  * Split e into potentially three parts; before this chunk, the chunk
725  * to be blamed for the parent, and after that portion.
726  */
727 static void split_overlap(struct blame_entry *split,
728                           struct blame_entry *e,
729                           int tlno, int plno, int same,
730                           struct blame_origin *parent)
731 {
732         int chunk_end_lno;
733         memset(split, 0, sizeof(struct blame_entry [3]));
734
735         if (e->s_lno < tlno) {
736                 /* there is a pre-chunk part not blamed on parent */
737                 split[0].suspect = blame_origin_incref(e->suspect);
738                 split[0].lno = e->lno;
739                 split[0].s_lno = e->s_lno;
740                 split[0].num_lines = tlno - e->s_lno;
741                 split[1].lno = e->lno + tlno - e->s_lno;
742                 split[1].s_lno = plno;
743         }
744         else {
745                 split[1].lno = e->lno;
746                 split[1].s_lno = plno + (e->s_lno - tlno);
747         }
748
749         if (same < e->s_lno + e->num_lines) {
750                 /* there is a post-chunk part not blamed on parent */
751                 split[2].suspect = blame_origin_incref(e->suspect);
752                 split[2].lno = e->lno + (same - e->s_lno);
753                 split[2].s_lno = e->s_lno + (same - e->s_lno);
754                 split[2].num_lines = e->s_lno + e->num_lines - same;
755                 chunk_end_lno = split[2].lno;
756         }
757         else
758                 chunk_end_lno = e->lno + e->num_lines;
759         split[1].num_lines = chunk_end_lno - split[1].lno;
760
761         /*
762          * if it turns out there is nothing to blame the parent for,
763          * forget about the splitting.  !split[1].suspect signals this.
764          */
765         if (split[1].num_lines < 1)
766                 return;
767         split[1].suspect = blame_origin_incref(parent);
768 }
769
770 /*
771  * split_overlap() divided an existing blame e into up to three parts
772  * in split.  Any assigned blame is moved to queue to
773  * reflect the split.
774  */
775 static void split_blame(struct blame_entry ***blamed,
776                         struct blame_entry ***unblamed,
777                         struct blame_entry *split,
778                         struct blame_entry *e)
779 {
780         if (split[0].suspect && split[2].suspect) {
781                 /* The first part (reuse storage for the existing entry e) */
782                 dup_entry(unblamed, e, &split[0]);
783
784                 /* The last part -- me */
785                 add_blame_entry(unblamed, &split[2]);
786
787                 /* ... and the middle part -- parent */
788                 add_blame_entry(blamed, &split[1]);
789         }
790         else if (!split[0].suspect && !split[2].suspect)
791                 /*
792                  * The parent covers the entire area; reuse storage for
793                  * e and replace it with the parent.
794                  */
795                 dup_entry(blamed, e, &split[1]);
796         else if (split[0].suspect) {
797                 /* me and then parent */
798                 dup_entry(unblamed, e, &split[0]);
799                 add_blame_entry(blamed, &split[1]);
800         }
801         else {
802                 /* parent and then me */
803                 dup_entry(blamed, e, &split[1]);
804                 add_blame_entry(unblamed, &split[2]);
805         }
806 }
807
808 /*
809  * After splitting the blame, the origins used by the
810  * on-stack blame_entry should lose one refcnt each.
811  */
812 static void decref_split(struct blame_entry *split)
813 {
814         int i;
815
816         for (i = 0; i < 3; i++)
817                 blame_origin_decref(split[i].suspect);
818 }
819
820 /*
821  * reverse_blame reverses the list given in head, appending tail.
822  * That allows us to build lists in reverse order, then reverse them
823  * afterwards.  This can be faster than building the list in proper
824  * order right away.  The reason is that building in proper order
825  * requires writing a link in the _previous_ element, while building
826  * in reverse order just requires placing the list head into the
827  * _current_ element.
828  */
829
830 static struct blame_entry *reverse_blame(struct blame_entry *head,
831                                          struct blame_entry *tail)
832 {
833         while (head) {
834                 struct blame_entry *next = head->next;
835                 head->next = tail;
836                 tail = head;
837                 head = next;
838         }
839         return tail;
840 }
841
842 /*
843  * Process one hunk from the patch between the current suspect for
844  * blame_entry e and its parent.  This first blames any unfinished
845  * entries before the chunk (which is where target and parent start
846  * differing) on the parent, and then splits blame entries at the
847  * start and at the end of the difference region.  Since use of -M and
848  * -C options may lead to overlapping/duplicate source line number
849  * ranges, all we can rely on from sorting/merging is the order of the
850  * first suspect line number.
851  */
852 static void blame_chunk(struct blame_entry ***dstq, struct blame_entry ***srcq,
853                         int tlno, int offset, int same,
854                         struct blame_origin *parent)
855 {
856         struct blame_entry *e = **srcq;
857         struct blame_entry *samep = NULL, *diffp = NULL;
858
859         while (e && e->s_lno < tlno) {
860                 struct blame_entry *next = e->next;
861                 /*
862                  * current record starts before differing portion.  If
863                  * it reaches into it, we need to split it up and
864                  * examine the second part separately.
865                  */
866                 if (e->s_lno + e->num_lines > tlno) {
867                         /* Move second half to a new record */
868                         int len = tlno - e->s_lno;
869                         struct blame_entry *n = xcalloc(1, sizeof (struct blame_entry));
870                         n->suspect = e->suspect;
871                         n->lno = e->lno + len;
872                         n->s_lno = e->s_lno + len;
873                         n->num_lines = e->num_lines - len;
874                         e->num_lines = len;
875                         e->score = 0;
876                         /* Push new record to diffp */
877                         n->next = diffp;
878                         diffp = n;
879                 } else
880                         blame_origin_decref(e->suspect);
881                 /* Pass blame for everything before the differing
882                  * chunk to the parent */
883                 e->suspect = blame_origin_incref(parent);
884                 e->s_lno += offset;
885                 e->next = samep;
886                 samep = e;
887                 e = next;
888         }
889         /*
890          * As we don't know how much of a common stretch after this
891          * diff will occur, the currently blamed parts are all that we
892          * can assign to the parent for now.
893          */
894
895         if (samep) {
896                 **dstq = reverse_blame(samep, **dstq);
897                 *dstq = &samep->next;
898         }
899         /*
900          * Prepend the split off portions: everything after e starts
901          * after the blameable portion.
902          */
903         e = reverse_blame(diffp, e);
904
905         /*
906          * Now retain records on the target while parts are different
907          * from the parent.
908          */
909         samep = NULL;
910         diffp = NULL;
911         while (e && e->s_lno < same) {
912                 struct blame_entry *next = e->next;
913
914                 /*
915                  * If current record extends into sameness, need to split.
916                  */
917                 if (e->s_lno + e->num_lines > same) {
918                         /*
919                          * Move second half to a new record to be
920                          * processed by later chunks
921                          */
922                         int len = same - e->s_lno;
923                         struct blame_entry *n = xcalloc(1, sizeof (struct blame_entry));
924                         n->suspect = blame_origin_incref(e->suspect);
925                         n->lno = e->lno + len;
926                         n->s_lno = e->s_lno + len;
927                         n->num_lines = e->num_lines - len;
928                         e->num_lines = len;
929                         e->score = 0;
930                         /* Push new record to samep */
931                         n->next = samep;
932                         samep = n;
933                 }
934                 e->next = diffp;
935                 diffp = e;
936                 e = next;
937         }
938         **srcq = reverse_blame(diffp, reverse_blame(samep, e));
939         /* Move across elements that are in the unblamable portion */
940         if (diffp)
941                 *srcq = &diffp->next;
942 }
943
944 struct blame_chunk_cb_data {
945         struct blame_origin *parent;
946         long offset;
947         struct blame_entry **dstq;
948         struct blame_entry **srcq;
949 };
950
951 /* diff chunks are from parent to target */
952 static int blame_chunk_cb(long start_a, long count_a,
953                           long start_b, long count_b, void *data)
954 {
955         struct blame_chunk_cb_data *d = data;
956         if (start_a - start_b != d->offset)
957                 die("internal error in blame::blame_chunk_cb");
958         blame_chunk(&d->dstq, &d->srcq, start_b, start_a - start_b,
959                     start_b + count_b, d->parent);
960         d->offset = start_a + count_a - (start_b + count_b);
961         return 0;
962 }
963
964 /*
965  * We are looking at the origin 'target' and aiming to pass blame
966  * for the lines it is suspected to its parent.  Run diff to find
967  * which lines came from parent and pass blame for them.
968  */
969 static void pass_blame_to_parent(struct blame_scoreboard *sb,
970                                  struct blame_origin *target,
971                                  struct blame_origin *parent)
972 {
973         mmfile_t file_p, file_o;
974         struct blame_chunk_cb_data d;
975         struct blame_entry *newdest = NULL;
976
977         if (!target->suspects)
978                 return; /* nothing remains for this target */
979
980         d.parent = parent;
981         d.offset = 0;
982         d.dstq = &newdest; d.srcq = &target->suspects;
983
984         fill_origin_blob(&sb->revs->diffopt, parent, &file_p, &sb->num_read_blob);
985         fill_origin_blob(&sb->revs->diffopt, target, &file_o, &sb->num_read_blob);
986         sb->num_get_patch++;
987
988         if (diff_hunks(&file_p, &file_o, blame_chunk_cb, &d, sb->xdl_opts))
989                 die("unable to generate diff (%s -> %s)",
990                     oid_to_hex(&parent->commit->object.oid),
991                     oid_to_hex(&target->commit->object.oid));
992         /* The rest are the same as the parent */
993         blame_chunk(&d.dstq, &d.srcq, INT_MAX, d.offset, INT_MAX, parent);
994         *d.dstq = NULL;
995         queue_blames(sb, parent, newdest);
996
997         return;
998 }
999
1000 /*
1001  * The lines in blame_entry after splitting blames many times can become
1002  * very small and trivial, and at some point it becomes pointless to
1003  * blame the parents.  E.g. "\t\t}\n\t}\n\n" appears everywhere in any
1004  * ordinary C program, and it is not worth to say it was copied from
1005  * totally unrelated file in the parent.
1006  *
1007  * Compute how trivial the lines in the blame_entry are.
1008  */
1009 unsigned blame_entry_score(struct blame_scoreboard *sb, struct blame_entry *e)
1010 {
1011         unsigned score;
1012         const char *cp, *ep;
1013
1014         if (e->score)
1015                 return e->score;
1016
1017         score = 1;
1018         cp = blame_nth_line(sb, e->lno);
1019         ep = blame_nth_line(sb, e->lno + e->num_lines);
1020         while (cp < ep) {
1021                 unsigned ch = *((unsigned char *)cp);
1022                 if (isalnum(ch))
1023                         score++;
1024                 cp++;
1025         }
1026         e->score = score;
1027         return score;
1028 }
1029
1030 /*
1031  * best_so_far[] and potential[] are both a split of an existing blame_entry
1032  * that passes blame to the parent.  Maintain best_so_far the best split so
1033  * far, by comparing potential and best_so_far and copying potential into
1034  * bst_so_far as needed.
1035  */
1036 static void copy_split_if_better(struct blame_scoreboard *sb,
1037                                  struct blame_entry *best_so_far,
1038                                  struct blame_entry *potential)
1039 {
1040         int i;
1041
1042         if (!potential[1].suspect)
1043                 return;
1044         if (best_so_far[1].suspect) {
1045                 if (blame_entry_score(sb, &potential[1]) <
1046                     blame_entry_score(sb, &best_so_far[1]))
1047                         return;
1048         }
1049
1050         for (i = 0; i < 3; i++)
1051                 blame_origin_incref(potential[i].suspect);
1052         decref_split(best_so_far);
1053         memcpy(best_so_far, potential, sizeof(struct blame_entry[3]));
1054 }
1055
1056 /*
1057  * We are looking at a part of the final image represented by
1058  * ent (tlno and same are offset by ent->s_lno).
1059  * tlno is where we are looking at in the final image.
1060  * up to (but not including) same match preimage.
1061  * plno is where we are looking at in the preimage.
1062  *
1063  * <-------------- final image ---------------------->
1064  *       <------ent------>
1065  *         ^tlno ^same
1066  *    <---------preimage----->
1067  *         ^plno
1068  *
1069  * All line numbers are 0-based.
1070  */
1071 static void handle_split(struct blame_scoreboard *sb,
1072                          struct blame_entry *ent,
1073                          int tlno, int plno, int same,
1074                          struct blame_origin *parent,
1075                          struct blame_entry *split)
1076 {
1077         if (ent->num_lines <= tlno)
1078                 return;
1079         if (tlno < same) {
1080                 struct blame_entry potential[3];
1081                 tlno += ent->s_lno;
1082                 same += ent->s_lno;
1083                 split_overlap(potential, ent, tlno, plno, same, parent);
1084                 copy_split_if_better(sb, split, potential);
1085                 decref_split(potential);
1086         }
1087 }
1088
1089 struct handle_split_cb_data {
1090         struct blame_scoreboard *sb;
1091         struct blame_entry *ent;
1092         struct blame_origin *parent;
1093         struct blame_entry *split;
1094         long plno;
1095         long tlno;
1096 };
1097
1098 static int handle_split_cb(long start_a, long count_a,
1099                            long start_b, long count_b, void *data)
1100 {
1101         struct handle_split_cb_data *d = data;
1102         handle_split(d->sb, d->ent, d->tlno, d->plno, start_b, d->parent,
1103                      d->split);
1104         d->plno = start_a + count_a;
1105         d->tlno = start_b + count_b;
1106         return 0;
1107 }
1108
1109 /*
1110  * Find the lines from parent that are the same as ent so that
1111  * we can pass blames to it.  file_p has the blob contents for
1112  * the parent.
1113  */
1114 static void find_copy_in_blob(struct blame_scoreboard *sb,
1115                               struct blame_entry *ent,
1116                               struct blame_origin *parent,
1117                               struct blame_entry *split,
1118                               mmfile_t *file_p)
1119 {
1120         const char *cp;
1121         mmfile_t file_o;
1122         struct handle_split_cb_data d;
1123
1124         memset(&d, 0, sizeof(d));
1125         d.sb = sb; d.ent = ent; d.parent = parent; d.split = split;
1126         /*
1127          * Prepare mmfile that contains only the lines in ent.
1128          */
1129         cp = blame_nth_line(sb, ent->lno);
1130         file_o.ptr = (char *) cp;
1131         file_o.size = blame_nth_line(sb, ent->lno + ent->num_lines) - cp;
1132
1133         /*
1134          * file_o is a part of final image we are annotating.
1135          * file_p partially may match that image.
1136          */
1137         memset(split, 0, sizeof(struct blame_entry [3]));
1138         if (diff_hunks(file_p, &file_o, handle_split_cb, &d, sb->xdl_opts))
1139                 die("unable to generate diff (%s)",
1140                     oid_to_hex(&parent->commit->object.oid));
1141         /* remainder, if any, all match the preimage */
1142         handle_split(sb, ent, d.tlno, d.plno, ent->num_lines, parent, split);
1143 }
1144
1145 /* Move all blame entries from list *source that have a score smaller
1146  * than score_min to the front of list *small.
1147  * Returns a pointer to the link pointing to the old head of the small list.
1148  */
1149
1150 static struct blame_entry **filter_small(struct blame_scoreboard *sb,
1151                                          struct blame_entry **small,
1152                                          struct blame_entry **source,
1153                                          unsigned score_min)
1154 {
1155         struct blame_entry *p = *source;
1156         struct blame_entry *oldsmall = *small;
1157         while (p) {
1158                 if (blame_entry_score(sb, p) <= score_min) {
1159                         *small = p;
1160                         small = &p->next;
1161                         p = *small;
1162                 } else {
1163                         *source = p;
1164                         source = &p->next;
1165                         p = *source;
1166                 }
1167         }
1168         *small = oldsmall;
1169         *source = NULL;
1170         return small;
1171 }
1172
1173 /*
1174  * See if lines currently target is suspected for can be attributed to
1175  * parent.
1176  */
1177 static void find_move_in_parent(struct blame_scoreboard *sb,
1178                                 struct blame_entry ***blamed,
1179                                 struct blame_entry **toosmall,
1180                                 struct blame_origin *target,
1181                                 struct blame_origin *parent)
1182 {
1183         struct blame_entry *e, split[3];
1184         struct blame_entry *unblamed = target->suspects;
1185         struct blame_entry *leftover = NULL;
1186         mmfile_t file_p;
1187
1188         if (!unblamed)
1189                 return; /* nothing remains for this target */
1190
1191         fill_origin_blob(&sb->revs->diffopt, parent, &file_p, &sb->num_read_blob);
1192         if (!file_p.ptr)
1193                 return;
1194
1195         /* At each iteration, unblamed has a NULL-terminated list of
1196          * entries that have not yet been tested for blame.  leftover
1197          * contains the reversed list of entries that have been tested
1198          * without being assignable to the parent.
1199          */
1200         do {
1201                 struct blame_entry **unblamedtail = &unblamed;
1202                 struct blame_entry *next;
1203                 for (e = unblamed; e; e = next) {
1204                         next = e->next;
1205                         find_copy_in_blob(sb, e, parent, split, &file_p);
1206                         if (split[1].suspect &&
1207                             sb->move_score < blame_entry_score(sb, &split[1])) {
1208                                 split_blame(blamed, &unblamedtail, split, e);
1209                         } else {
1210                                 e->next = leftover;
1211                                 leftover = e;
1212                         }
1213                         decref_split(split);
1214                 }
1215                 *unblamedtail = NULL;
1216                 toosmall = filter_small(sb, toosmall, &unblamed, sb->move_score);
1217         } while (unblamed);
1218         target->suspects = reverse_blame(leftover, NULL);
1219 }
1220
1221 struct blame_list {
1222         struct blame_entry *ent;
1223         struct blame_entry split[3];
1224 };
1225
1226 /*
1227  * Count the number of entries the target is suspected for,
1228  * and prepare a list of entry and the best split.
1229  */
1230 static struct blame_list *setup_blame_list(struct blame_entry *unblamed,
1231                                            int *num_ents_p)
1232 {
1233         struct blame_entry *e;
1234         int num_ents, i;
1235         struct blame_list *blame_list = NULL;
1236
1237         for (e = unblamed, num_ents = 0; e; e = e->next)
1238                 num_ents++;
1239         if (num_ents) {
1240                 blame_list = xcalloc(num_ents, sizeof(struct blame_list));
1241                 for (e = unblamed, i = 0; e; e = e->next)
1242                         blame_list[i++].ent = e;
1243         }
1244         *num_ents_p = num_ents;
1245         return blame_list;
1246 }
1247
1248 /*
1249  * For lines target is suspected for, see if we can find code movement
1250  * across file boundary from the parent commit.  porigin is the path
1251  * in the parent we already tried.
1252  */
1253 static void find_copy_in_parent(struct blame_scoreboard *sb,
1254                                 struct blame_entry ***blamed,
1255                                 struct blame_entry **toosmall,
1256                                 struct blame_origin *target,
1257                                 struct commit *parent,
1258                                 struct blame_origin *porigin,
1259                                 int opt)
1260 {
1261         struct diff_options diff_opts;
1262         int i, j;
1263         struct blame_list *blame_list;
1264         int num_ents;
1265         struct blame_entry *unblamed = target->suspects;
1266         struct blame_entry *leftover = NULL;
1267
1268         if (!unblamed)
1269                 return; /* nothing remains for this target */
1270
1271         repo_diff_setup(sb->repo, &diff_opts);
1272         diff_opts.flags.recursive = 1;
1273         diff_opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1274
1275         diff_setup_done(&diff_opts);
1276
1277         /* Try "find copies harder" on new path if requested;
1278          * we do not want to use diffcore_rename() actually to
1279          * match things up; find_copies_harder is set only to
1280          * force diff_tree_oid() to feed all filepairs to diff_queue,
1281          * and this code needs to be after diff_setup_done(), which
1282          * usually makes find-copies-harder imply copy detection.
1283          */
1284         if ((opt & PICKAXE_BLAME_COPY_HARDEST)
1285             || ((opt & PICKAXE_BLAME_COPY_HARDER)
1286                 && (!porigin || strcmp(target->path, porigin->path))))
1287                 diff_opts.flags.find_copies_harder = 1;
1288
1289         if (is_null_oid(&target->commit->object.oid))
1290                 do_diff_cache(get_commit_tree_oid(parent), &diff_opts);
1291         else
1292                 diff_tree_oid(get_commit_tree_oid(parent),
1293                               get_commit_tree_oid(target->commit),
1294                               "", &diff_opts);
1295
1296         if (!diff_opts.flags.find_copies_harder)
1297                 diffcore_std(&diff_opts);
1298
1299         do {
1300                 struct blame_entry **unblamedtail = &unblamed;
1301                 blame_list = setup_blame_list(unblamed, &num_ents);
1302
1303                 for (i = 0; i < diff_queued_diff.nr; i++) {
1304                         struct diff_filepair *p = diff_queued_diff.queue[i];
1305                         struct blame_origin *norigin;
1306                         mmfile_t file_p;
1307                         struct blame_entry potential[3];
1308
1309                         if (!DIFF_FILE_VALID(p->one))
1310                                 continue; /* does not exist in parent */
1311                         if (S_ISGITLINK(p->one->mode))
1312                                 continue; /* ignore git links */
1313                         if (porigin && !strcmp(p->one->path, porigin->path))
1314                                 /* find_move already dealt with this path */
1315                                 continue;
1316
1317                         norigin = get_origin(parent, p->one->path);
1318                         oidcpy(&norigin->blob_oid, &p->one->oid);
1319                         norigin->mode = p->one->mode;
1320                         fill_origin_blob(&sb->revs->diffopt, norigin, &file_p, &sb->num_read_blob);
1321                         if (!file_p.ptr)
1322                                 continue;
1323
1324                         for (j = 0; j < num_ents; j++) {
1325                                 find_copy_in_blob(sb, blame_list[j].ent,
1326                                                   norigin, potential, &file_p);
1327                                 copy_split_if_better(sb, blame_list[j].split,
1328                                                      potential);
1329                                 decref_split(potential);
1330                         }
1331                         blame_origin_decref(norigin);
1332                 }
1333
1334                 for (j = 0; j < num_ents; j++) {
1335                         struct blame_entry *split = blame_list[j].split;
1336                         if (split[1].suspect &&
1337                             sb->copy_score < blame_entry_score(sb, &split[1])) {
1338                                 split_blame(blamed, &unblamedtail, split,
1339                                             blame_list[j].ent);
1340                         } else {
1341                                 blame_list[j].ent->next = leftover;
1342                                 leftover = blame_list[j].ent;
1343                         }
1344                         decref_split(split);
1345                 }
1346                 free(blame_list);
1347                 *unblamedtail = NULL;
1348                 toosmall = filter_small(sb, toosmall, &unblamed, sb->copy_score);
1349         } while (unblamed);
1350         target->suspects = reverse_blame(leftover, NULL);
1351         diff_flush(&diff_opts);
1352         clear_pathspec(&diff_opts.pathspec);
1353 }
1354
1355 /*
1356  * The blobs of origin and porigin exactly match, so everything
1357  * origin is suspected for can be blamed on the parent.
1358  */
1359 static void pass_whole_blame(struct blame_scoreboard *sb,
1360                              struct blame_origin *origin, struct blame_origin *porigin)
1361 {
1362         struct blame_entry *e, *suspects;
1363
1364         if (!porigin->file.ptr && origin->file.ptr) {
1365                 /* Steal its file */
1366                 porigin->file = origin->file;
1367                 origin->file.ptr = NULL;
1368         }
1369         suspects = origin->suspects;
1370         origin->suspects = NULL;
1371         for (e = suspects; e; e = e->next) {
1372                 blame_origin_incref(porigin);
1373                 blame_origin_decref(e->suspect);
1374                 e->suspect = porigin;
1375         }
1376         queue_blames(sb, porigin, suspects);
1377 }
1378
1379 /*
1380  * We pass blame from the current commit to its parents.  We keep saying
1381  * "parent" (and "porigin"), but what we mean is to find scapegoat to
1382  * exonerate ourselves.
1383  */
1384 static struct commit_list *first_scapegoat(struct rev_info *revs, struct commit *commit,
1385                                         int reverse)
1386 {
1387         if (!reverse) {
1388                 if (revs->first_parent_only &&
1389                     commit->parents &&
1390                     commit->parents->next) {
1391                         free_commit_list(commit->parents->next);
1392                         commit->parents->next = NULL;
1393                 }
1394                 return commit->parents;
1395         }
1396         return lookup_decoration(&revs->children, &commit->object);
1397 }
1398
1399 static int num_scapegoats(struct rev_info *revs, struct commit *commit, int reverse)
1400 {
1401         struct commit_list *l = first_scapegoat(revs, commit, reverse);
1402         return commit_list_count(l);
1403 }
1404
1405 /* Distribute collected unsorted blames to the respected sorted lists
1406  * in the various origins.
1407  */
1408 static void distribute_blame(struct blame_scoreboard *sb, struct blame_entry *blamed)
1409 {
1410         blamed = llist_mergesort(blamed, get_next_blame, set_next_blame,
1411                                  compare_blame_suspect);
1412         while (blamed)
1413         {
1414                 struct blame_origin *porigin = blamed->suspect;
1415                 struct blame_entry *suspects = NULL;
1416                 do {
1417                         struct blame_entry *next = blamed->next;
1418                         blamed->next = suspects;
1419                         suspects = blamed;
1420                         blamed = next;
1421                 } while (blamed && blamed->suspect == porigin);
1422                 suspects = reverse_blame(suspects, NULL);
1423                 queue_blames(sb, porigin, suspects);
1424         }
1425 }
1426
1427 #define MAXSG 16
1428
1429 static void pass_blame(struct blame_scoreboard *sb, struct blame_origin *origin, int opt)
1430 {
1431         struct rev_info *revs = sb->revs;
1432         int i, pass, num_sg;
1433         struct commit *commit = origin->commit;
1434         struct commit_list *sg;
1435         struct blame_origin *sg_buf[MAXSG];
1436         struct blame_origin *porigin, **sg_origin = sg_buf;
1437         struct blame_entry *toosmall = NULL;
1438         struct blame_entry *blames, **blametail = &blames;
1439
1440         num_sg = num_scapegoats(revs, commit, sb->reverse);
1441         if (!num_sg)
1442                 goto finish;
1443         else if (num_sg < ARRAY_SIZE(sg_buf))
1444                 memset(sg_buf, 0, sizeof(sg_buf));
1445         else
1446                 sg_origin = xcalloc(num_sg, sizeof(*sg_origin));
1447
1448         /*
1449          * The first pass looks for unrenamed path to optimize for
1450          * common cases, then we look for renames in the second pass.
1451          */
1452         for (pass = 0; pass < 2 - sb->no_whole_file_rename; pass++) {
1453                 struct blame_origin *(*find)(struct repository *, struct commit *, struct blame_origin *);
1454                 find = pass ? find_rename : find_origin;
1455
1456                 for (i = 0, sg = first_scapegoat(revs, commit, sb->reverse);
1457                      i < num_sg && sg;
1458                      sg = sg->next, i++) {
1459                         struct commit *p = sg->item;
1460                         int j, same;
1461
1462                         if (sg_origin[i])
1463                                 continue;
1464                         if (parse_commit(p))
1465                                 continue;
1466                         porigin = find(sb->repo, p, origin);
1467                         if (!porigin)
1468                                 continue;
1469                         if (oideq(&porigin->blob_oid, &origin->blob_oid)) {
1470                                 pass_whole_blame(sb, origin, porigin);
1471                                 blame_origin_decref(porigin);
1472                                 goto finish;
1473                         }
1474                         for (j = same = 0; j < i; j++)
1475                                 if (sg_origin[j] &&
1476                                     oideq(&sg_origin[j]->blob_oid, &porigin->blob_oid)) {
1477                                         same = 1;
1478                                         break;
1479                                 }
1480                         if (!same)
1481                                 sg_origin[i] = porigin;
1482                         else
1483                                 blame_origin_decref(porigin);
1484                 }
1485         }
1486
1487         sb->num_commits++;
1488         for (i = 0, sg = first_scapegoat(revs, commit, sb->reverse);
1489              i < num_sg && sg;
1490              sg = sg->next, i++) {
1491                 struct blame_origin *porigin = sg_origin[i];
1492                 if (!porigin)
1493                         continue;
1494                 if (!origin->previous) {
1495                         blame_origin_incref(porigin);
1496                         origin->previous = porigin;
1497                 }
1498                 pass_blame_to_parent(sb, origin, porigin);
1499                 if (!origin->suspects)
1500                         goto finish;
1501         }
1502
1503         /*
1504          * Optionally find moves in parents' files.
1505          */
1506         if (opt & PICKAXE_BLAME_MOVE) {
1507                 filter_small(sb, &toosmall, &origin->suspects, sb->move_score);
1508                 if (origin->suspects) {
1509                         for (i = 0, sg = first_scapegoat(revs, commit, sb->reverse);
1510                              i < num_sg && sg;
1511                              sg = sg->next, i++) {
1512                                 struct blame_origin *porigin = sg_origin[i];
1513                                 if (!porigin)
1514                                         continue;
1515                                 find_move_in_parent(sb, &blametail, &toosmall, origin, porigin);
1516                                 if (!origin->suspects)
1517                                         break;
1518                         }
1519                 }
1520         }
1521
1522         /*
1523          * Optionally find copies from parents' files.
1524          */
1525         if (opt & PICKAXE_BLAME_COPY) {
1526                 if (sb->copy_score > sb->move_score)
1527                         filter_small(sb, &toosmall, &origin->suspects, sb->copy_score);
1528                 else if (sb->copy_score < sb->move_score) {
1529                         origin->suspects = blame_merge(origin->suspects, toosmall);
1530                         toosmall = NULL;
1531                         filter_small(sb, &toosmall, &origin->suspects, sb->copy_score);
1532                 }
1533                 if (!origin->suspects)
1534                         goto finish;
1535
1536                 for (i = 0, sg = first_scapegoat(revs, commit, sb->reverse);
1537                      i < num_sg && sg;
1538                      sg = sg->next, i++) {
1539                         struct blame_origin *porigin = sg_origin[i];
1540                         find_copy_in_parent(sb, &blametail, &toosmall,
1541                                             origin, sg->item, porigin, opt);
1542                         if (!origin->suspects)
1543                                 goto finish;
1544                 }
1545         }
1546
1547 finish:
1548         *blametail = NULL;
1549         distribute_blame(sb, blames);
1550         /*
1551          * prepend toosmall to origin->suspects
1552          *
1553          * There is no point in sorting: this ends up on a big
1554          * unsorted list in the caller anyway.
1555          */
1556         if (toosmall) {
1557                 struct blame_entry **tail = &toosmall;
1558                 while (*tail)
1559                         tail = &(*tail)->next;
1560                 *tail = origin->suspects;
1561                 origin->suspects = toosmall;
1562         }
1563         for (i = 0; i < num_sg; i++) {
1564                 if (sg_origin[i]) {
1565                         drop_origin_blob(sg_origin[i]);
1566                         blame_origin_decref(sg_origin[i]);
1567                 }
1568         }
1569         drop_origin_blob(origin);
1570         if (sg_buf != sg_origin)
1571                 free(sg_origin);
1572 }
1573
1574 /*
1575  * The main loop -- while we have blobs with lines whose true origin
1576  * is still unknown, pick one blob, and allow its lines to pass blames
1577  * to its parents. */
1578 void assign_blame(struct blame_scoreboard *sb, int opt)
1579 {
1580         struct rev_info *revs = sb->revs;
1581         struct commit *commit = prio_queue_get(&sb->commits);
1582
1583         while (commit) {
1584                 struct blame_entry *ent;
1585                 struct blame_origin *suspect = get_blame_suspects(commit);
1586
1587                 /* find one suspect to break down */
1588                 while (suspect && !suspect->suspects)
1589                         suspect = suspect->next;
1590
1591                 if (!suspect) {
1592                         commit = prio_queue_get(&sb->commits);
1593                         continue;
1594                 }
1595
1596                 assert(commit == suspect->commit);
1597
1598                 /*
1599                  * We will use this suspect later in the loop,
1600                  * so hold onto it in the meantime.
1601                  */
1602                 blame_origin_incref(suspect);
1603                 parse_commit(commit);
1604                 if (sb->reverse ||
1605                     (!(commit->object.flags & UNINTERESTING) &&
1606                      !(revs->max_age != -1 && commit->date < revs->max_age)))
1607                         pass_blame(sb, suspect, opt);
1608                 else {
1609                         commit->object.flags |= UNINTERESTING;
1610                         if (commit->object.parsed)
1611                                 mark_parents_uninteresting(commit);
1612                 }
1613                 /* treat root commit as boundary */
1614                 if (!commit->parents && !sb->show_root)
1615                         commit->object.flags |= UNINTERESTING;
1616
1617                 /* Take responsibility for the remaining entries */
1618                 ent = suspect->suspects;
1619                 if (ent) {
1620                         suspect->guilty = 1;
1621                         for (;;) {
1622                                 struct blame_entry *next = ent->next;
1623                                 if (sb->found_guilty_entry)
1624                                         sb->found_guilty_entry(ent, sb->found_guilty_entry_data);
1625                                 if (next) {
1626                                         ent = next;
1627                                         continue;
1628                                 }
1629                                 ent->next = sb->ent;
1630                                 sb->ent = suspect->suspects;
1631                                 suspect->suspects = NULL;
1632                                 break;
1633                         }
1634                 }
1635                 blame_origin_decref(suspect);
1636
1637                 if (sb->debug) /* sanity */
1638                         sanity_check_refcnt(sb);
1639         }
1640 }
1641
1642 static const char *get_next_line(const char *start, const char *end)
1643 {
1644         const char *nl = memchr(start, '\n', end - start);
1645         return nl ? nl + 1 : end;
1646 }
1647
1648 /*
1649  * To allow quick access to the contents of nth line in the
1650  * final image, prepare an index in the scoreboard.
1651  */
1652 static int prepare_lines(struct blame_scoreboard *sb)
1653 {
1654         const char *buf = sb->final_buf;
1655         unsigned long len = sb->final_buf_size;
1656         const char *end = buf + len;
1657         const char *p;
1658         int *lineno;
1659         int num = 0;
1660
1661         for (p = buf; p < end; p = get_next_line(p, end))
1662                 num++;
1663
1664         ALLOC_ARRAY(sb->lineno, num + 1);
1665         lineno = sb->lineno;
1666
1667         for (p = buf; p < end; p = get_next_line(p, end))
1668                 *lineno++ = p - buf;
1669
1670         *lineno = len;
1671
1672         sb->num_lines = num;
1673         return sb->num_lines;
1674 }
1675
1676 static struct commit *find_single_final(struct rev_info *revs,
1677                                         const char **name_p)
1678 {
1679         int i;
1680         struct commit *found = NULL;
1681         const char *name = NULL;
1682
1683         for (i = 0; i < revs->pending.nr; i++) {
1684                 struct object *obj = revs->pending.objects[i].item;
1685                 if (obj->flags & UNINTERESTING)
1686                         continue;
1687                 obj = deref_tag(revs->repo, obj, NULL, 0);
1688                 if (obj->type != OBJ_COMMIT)
1689                         die("Non commit %s?", revs->pending.objects[i].name);
1690                 if (found)
1691                         die("More than one commit to dig from %s and %s?",
1692                             revs->pending.objects[i].name, name);
1693                 found = (struct commit *)obj;
1694                 name = revs->pending.objects[i].name;
1695         }
1696         if (name_p)
1697                 *name_p = xstrdup_or_null(name);
1698         return found;
1699 }
1700
1701 static struct commit *dwim_reverse_initial(struct rev_info *revs,
1702                                            const char **name_p)
1703 {
1704         /*
1705          * DWIM "git blame --reverse ONE -- PATH" as
1706          * "git blame --reverse ONE..HEAD -- PATH" but only do so
1707          * when it makes sense.
1708          */
1709         struct object *obj;
1710         struct commit *head_commit;
1711         struct object_id head_oid;
1712
1713         if (revs->pending.nr != 1)
1714                 return NULL;
1715
1716         /* Is that sole rev a committish? */
1717         obj = revs->pending.objects[0].item;
1718         obj = deref_tag(revs->repo, obj, NULL, 0);
1719         if (obj->type != OBJ_COMMIT)
1720                 return NULL;
1721
1722         /* Do we have HEAD? */
1723         if (!resolve_ref_unsafe("HEAD", RESOLVE_REF_READING, &head_oid, NULL))
1724                 return NULL;
1725         head_commit = lookup_commit_reference_gently(revs->repo,
1726                                                      &head_oid, 1);
1727         if (!head_commit)
1728                 return NULL;
1729
1730         /* Turn "ONE" into "ONE..HEAD" then */
1731         obj->flags |= UNINTERESTING;
1732         add_pending_object(revs, &head_commit->object, "HEAD");
1733
1734         if (name_p)
1735                 *name_p = revs->pending.objects[0].name;
1736         return (struct commit *)obj;
1737 }
1738
1739 static struct commit *find_single_initial(struct rev_info *revs,
1740                                           const char **name_p)
1741 {
1742         int i;
1743         struct commit *found = NULL;
1744         const char *name = NULL;
1745
1746         /*
1747          * There must be one and only one negative commit, and it must be
1748          * the boundary.
1749          */
1750         for (i = 0; i < revs->pending.nr; i++) {
1751                 struct object *obj = revs->pending.objects[i].item;
1752                 if (!(obj->flags & UNINTERESTING))
1753                         continue;
1754                 obj = deref_tag(revs->repo, obj, NULL, 0);
1755                 if (obj->type != OBJ_COMMIT)
1756                         die("Non commit %s?", revs->pending.objects[i].name);
1757                 if (found)
1758                         die("More than one commit to dig up from, %s and %s?",
1759                             revs->pending.objects[i].name, name);
1760                 found = (struct commit *) obj;
1761                 name = revs->pending.objects[i].name;
1762         }
1763
1764         if (!name)
1765                 found = dwim_reverse_initial(revs, &name);
1766         if (!name)
1767                 die("No commit to dig up from?");
1768
1769         if (name_p)
1770                 *name_p = xstrdup(name);
1771         return found;
1772 }
1773
1774 void init_scoreboard(struct blame_scoreboard *sb)
1775 {
1776         memset(sb, 0, sizeof(struct blame_scoreboard));
1777         sb->move_score = BLAME_DEFAULT_MOVE_SCORE;
1778         sb->copy_score = BLAME_DEFAULT_COPY_SCORE;
1779 }
1780
1781 void setup_scoreboard(struct blame_scoreboard *sb,
1782                       const char *path,
1783                       struct blame_origin **orig)
1784 {
1785         const char *final_commit_name = NULL;
1786         struct blame_origin *o;
1787         struct commit *final_commit = NULL;
1788         enum object_type type;
1789
1790         init_blame_suspects(&blame_suspects);
1791
1792         if (sb->reverse && sb->contents_from)
1793                 die(_("--contents and --reverse do not blend well."));
1794
1795         if (!sb->repo)
1796                 BUG("repo is NULL");
1797
1798         if (!sb->reverse) {
1799                 sb->final = find_single_final(sb->revs, &final_commit_name);
1800                 sb->commits.compare = compare_commits_by_commit_date;
1801         } else {
1802                 sb->final = find_single_initial(sb->revs, &final_commit_name);
1803                 sb->commits.compare = compare_commits_by_reverse_commit_date;
1804         }
1805
1806         if (sb->final && sb->contents_from)
1807                 die(_("cannot use --contents with final commit object name"));
1808
1809         if (sb->reverse && sb->revs->first_parent_only)
1810                 sb->revs->children.name = NULL;
1811
1812         if (!sb->final) {
1813                 /*
1814                  * "--not A B -- path" without anything positive;
1815                  * do not default to HEAD, but use the working tree
1816                  * or "--contents".
1817                  */
1818                 setup_work_tree();
1819                 sb->final = fake_working_tree_commit(sb->repo,
1820                                                      &sb->revs->diffopt,
1821                                                      path, sb->contents_from);
1822                 add_pending_object(sb->revs, &(sb->final->object), ":");
1823         }
1824
1825         if (sb->reverse && sb->revs->first_parent_only) {
1826                 final_commit = find_single_final(sb->revs, NULL);
1827                 if (!final_commit)
1828                         die(_("--reverse and --first-parent together require specified latest commit"));
1829         }
1830
1831         /*
1832          * If we have bottom, this will mark the ancestors of the
1833          * bottom commits we would reach while traversing as
1834          * uninteresting.
1835          */
1836         if (prepare_revision_walk(sb->revs))
1837                 die(_("revision walk setup failed"));
1838
1839         if (sb->reverse && sb->revs->first_parent_only) {
1840                 struct commit *c = final_commit;
1841
1842                 sb->revs->children.name = "children";
1843                 while (c->parents &&
1844                        !oideq(&c->object.oid, &sb->final->object.oid)) {
1845                         struct commit_list *l = xcalloc(1, sizeof(*l));
1846
1847                         l->item = c;
1848                         if (add_decoration(&sb->revs->children,
1849                                            &c->parents->item->object, l))
1850                                 BUG("not unique item in first-parent chain");
1851                         c = c->parents->item;
1852                 }
1853
1854                 if (!oideq(&c->object.oid, &sb->final->object.oid))
1855                         die(_("--reverse --first-parent together require range along first-parent chain"));
1856         }
1857
1858         if (is_null_oid(&sb->final->object.oid)) {
1859                 o = get_blame_suspects(sb->final);
1860                 sb->final_buf = xmemdupz(o->file.ptr, o->file.size);
1861                 sb->final_buf_size = o->file.size;
1862         }
1863         else {
1864                 o = get_origin(sb->final, path);
1865                 if (fill_blob_sha1_and_mode(sb->repo, o))
1866                         die(_("no such path %s in %s"), path, final_commit_name);
1867
1868                 if (sb->revs->diffopt.flags.allow_textconv &&
1869                     textconv_object(sb->repo, path, o->mode, &o->blob_oid, 1, (char **) &sb->final_buf,
1870                                     &sb->final_buf_size))
1871                         ;
1872                 else
1873                         sb->final_buf = read_object_file(&o->blob_oid, &type,
1874                                                          &sb->final_buf_size);
1875
1876                 if (!sb->final_buf)
1877                         die(_("cannot read blob %s for path %s"),
1878                             oid_to_hex(&o->blob_oid),
1879                             path);
1880         }
1881         sb->num_read_blob++;
1882         prepare_lines(sb);
1883
1884         if (orig)
1885                 *orig = o;
1886
1887         free((char *)final_commit_name);
1888 }
1889
1890
1891
1892 struct blame_entry *blame_entry_prepend(struct blame_entry *head,
1893                                         long start, long end,
1894                                         struct blame_origin *o)
1895 {
1896         struct blame_entry *new_head = xcalloc(1, sizeof(struct blame_entry));
1897         new_head->lno = start;
1898         new_head->num_lines = end - start;
1899         new_head->suspect = o;
1900         new_head->s_lno = start;
1901         new_head->next = head;
1902         blame_origin_incref(o);
1903         return new_head;
1904 }