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