Merge branch 'jk/pack-idx-corruption-safety' into maint
[git] / builtin / blame.c
1 /*
2  * Blame
3  *
4  * Copyright (c) 2006, 2014 by its authors
5  * See COPYING for licensing conditions
6  */
7
8 #include "cache.h"
9 #include "refs.h"
10 #include "builtin.h"
11 #include "blob.h"
12 #include "commit.h"
13 #include "tag.h"
14 #include "tree-walk.h"
15 #include "diff.h"
16 #include "diffcore.h"
17 #include "revision.h"
18 #include "quote.h"
19 #include "xdiff-interface.h"
20 #include "cache-tree.h"
21 #include "string-list.h"
22 #include "mailmap.h"
23 #include "mergesort.h"
24 #include "parse-options.h"
25 #include "prio-queue.h"
26 #include "utf8.h"
27 #include "userdiff.h"
28 #include "line-range.h"
29 #include "line-log.h"
30 #include "dir.h"
31
32 static char blame_usage[] = N_("git blame [<options>] [<rev-opts>] [<rev>] [--] <file>");
33
34 static const char *blame_opt_usage[] = {
35         blame_usage,
36         "",
37         N_("<rev-opts> are documented in git-rev-list(1)"),
38         NULL
39 };
40
41 static int longest_file;
42 static int longest_author;
43 static int max_orig_digits;
44 static int max_digits;
45 static int max_score_digits;
46 static int show_root;
47 static int reverse;
48 static int blank_boundary;
49 static int incremental;
50 static int xdl_opts;
51 static int abbrev = -1;
52 static int no_whole_file_rename;
53
54 static struct date_mode blame_date_mode = { DATE_ISO8601 };
55 static size_t blame_date_width;
56
57 static struct string_list mailmap;
58
59 #ifndef DEBUG
60 #define DEBUG 0
61 #endif
62
63 /* stats */
64 static int num_read_blob;
65 static int num_get_patch;
66 static int num_commits;
67
68 #define PICKAXE_BLAME_MOVE              01
69 #define PICKAXE_BLAME_COPY              02
70 #define PICKAXE_BLAME_COPY_HARDER       04
71 #define PICKAXE_BLAME_COPY_HARDEST      010
72
73 /*
74  * blame for a blame_entry with score lower than these thresholds
75  * is not passed to the parent using move/copy logic.
76  */
77 static unsigned blame_move_score;
78 static unsigned blame_copy_score;
79 #define BLAME_DEFAULT_MOVE_SCORE        20
80 #define BLAME_DEFAULT_COPY_SCORE        40
81
82 /* Remember to update object flag allocation in object.h */
83 #define METAINFO_SHOWN          (1u<<12)
84 #define MORE_THAN_ONE_PATH      (1u<<13)
85
86 /*
87  * One blob in a commit that is being suspected
88  */
89 struct origin {
90         int refcnt;
91         /* Record preceding blame record for this blob */
92         struct origin *previous;
93         /* origins are put in a list linked via `next' hanging off the
94          * corresponding commit's util field in order to make finding
95          * them fast.  The presence in this chain does not count
96          * towards the origin's reference count.  It is tempting to
97          * let it count as long as the commit is pending examination,
98          * but even under circumstances where the commit will be
99          * present multiple times in the priority queue of unexamined
100          * commits, processing the first instance will not leave any
101          * work requiring the origin data for the second instance.  An
102          * interspersed commit changing that would have to be
103          * preexisting with a different ancestry and with the same
104          * commit date in order to wedge itself between two instances
105          * of the same commit in the priority queue _and_ produce
106          * blame entries relevant for it.  While we don't want to let
107          * us get tripped up by this case, it certainly does not seem
108          * worth optimizing for.
109          */
110         struct origin *next;
111         struct commit *commit;
112         /* `suspects' contains blame entries that may be attributed to
113          * this origin's commit or to parent commits.  When a commit
114          * is being processed, all suspects will be moved, either by
115          * assigning them to an origin in a different commit, or by
116          * shipping them to the scoreboard's ent list because they
117          * cannot be attributed to a different commit.
118          */
119         struct blame_entry *suspects;
120         mmfile_t file;
121         unsigned char blob_sha1[20];
122         unsigned mode;
123         /* guilty gets set when shipping any suspects to the final
124          * blame list instead of other commits
125          */
126         char guilty;
127         char path[FLEX_ARRAY];
128 };
129
130 static int diff_hunks(mmfile_t *file_a, mmfile_t *file_b, long ctxlen,
131                       xdl_emit_hunk_consume_func_t hunk_func, void *cb_data)
132 {
133         xpparam_t xpp = {0};
134         xdemitconf_t xecfg = {0};
135         xdemitcb_t ecb = {NULL};
136
137         xpp.flags = xdl_opts;
138         xecfg.ctxlen = ctxlen;
139         xecfg.hunk_func = hunk_func;
140         ecb.priv = cb_data;
141         return xdi_diff(file_a, file_b, &xpp, &xecfg, &ecb);
142 }
143
144 /*
145  * Prepare diff_filespec and convert it using diff textconv API
146  * if the textconv driver exists.
147  * Return 1 if the conversion succeeds, 0 otherwise.
148  */
149 int textconv_object(const char *path,
150                     unsigned mode,
151                     const unsigned char *sha1,
152                     int sha1_valid,
153                     char **buf,
154                     unsigned long *buf_size)
155 {
156         struct diff_filespec *df;
157         struct userdiff_driver *textconv;
158
159         df = alloc_filespec(path);
160         fill_filespec(df, sha1, sha1_valid, mode);
161         textconv = get_textconv(df);
162         if (!textconv) {
163                 free_filespec(df);
164                 return 0;
165         }
166
167         *buf_size = fill_textconv(textconv, df, buf);
168         free_filespec(df);
169         return 1;
170 }
171
172 /*
173  * Given an origin, prepare mmfile_t structure to be used by the
174  * diff machinery
175  */
176 static void fill_origin_blob(struct diff_options *opt,
177                              struct origin *o, mmfile_t *file)
178 {
179         if (!o->file.ptr) {
180                 enum object_type type;
181                 unsigned long file_size;
182
183                 num_read_blob++;
184                 if (DIFF_OPT_TST(opt, ALLOW_TEXTCONV) &&
185                     textconv_object(o->path, o->mode, o->blob_sha1, 1, &file->ptr, &file_size))
186                         ;
187                 else
188                         file->ptr = read_sha1_file(o->blob_sha1, &type, &file_size);
189                 file->size = file_size;
190
191                 if (!file->ptr)
192                         die("Cannot read blob %s for path %s",
193                             sha1_to_hex(o->blob_sha1),
194                             o->path);
195                 o->file = *file;
196         }
197         else
198                 *file = o->file;
199 }
200
201 /*
202  * Origin is refcounted and usually we keep the blob contents to be
203  * reused.
204  */
205 static inline struct origin *origin_incref(struct origin *o)
206 {
207         if (o)
208                 o->refcnt++;
209         return o;
210 }
211
212 static void origin_decref(struct origin *o)
213 {
214         if (o && --o->refcnt <= 0) {
215                 struct origin *p, *l = NULL;
216                 if (o->previous)
217                         origin_decref(o->previous);
218                 free(o->file.ptr);
219                 /* Should be present exactly once in commit chain */
220                 for (p = o->commit->util; p; l = p, p = p->next) {
221                         if (p == o) {
222                                 if (l)
223                                         l->next = p->next;
224                                 else
225                                         o->commit->util = p->next;
226                                 free(o);
227                                 return;
228                         }
229                 }
230                 die("internal error in blame::origin_decref");
231         }
232 }
233
234 static void drop_origin_blob(struct origin *o)
235 {
236         if (o->file.ptr) {
237                 free(o->file.ptr);
238                 o->file.ptr = NULL;
239         }
240 }
241
242 /*
243  * Each group of lines is described by a blame_entry; it can be split
244  * as we pass blame to the parents.  They are arranged in linked lists
245  * kept as `suspects' of some unprocessed origin, or entered (when the
246  * blame origin has been finalized) into the scoreboard structure.
247  * While the scoreboard structure is only sorted at the end of
248  * processing (according to final image line number), the lists
249  * attached to an origin are sorted by the target line number.
250  */
251 struct blame_entry {
252         struct blame_entry *next;
253
254         /* the first line of this group in the final image;
255          * internally all line numbers are 0 based.
256          */
257         int lno;
258
259         /* how many lines this group has */
260         int num_lines;
261
262         /* the commit that introduced this group into the final image */
263         struct origin *suspect;
264
265         /* the line number of the first line of this group in the
266          * suspect's file; internally all line numbers are 0 based.
267          */
268         int s_lno;
269
270         /* how significant this entry is -- cached to avoid
271          * scanning the lines over and over.
272          */
273         unsigned score;
274 };
275
276 /*
277  * Any merge of blames happens on lists of blames that arrived via
278  * different parents in a single suspect.  In this case, we want to
279  * sort according to the suspect line numbers as opposed to the final
280  * image line numbers.  The function body is somewhat longish because
281  * it avoids unnecessary writes.
282  */
283
284 static struct blame_entry *blame_merge(struct blame_entry *list1,
285                                        struct blame_entry *list2)
286 {
287         struct blame_entry *p1 = list1, *p2 = list2,
288                 **tail = &list1;
289
290         if (!p1)
291                 return p2;
292         if (!p2)
293                 return p1;
294
295         if (p1->s_lno <= p2->s_lno) {
296                 do {
297                         tail = &p1->next;
298                         if ((p1 = *tail) == NULL) {
299                                 *tail = p2;
300                                 return list1;
301                         }
302                 } while (p1->s_lno <= p2->s_lno);
303         }
304         for (;;) {
305                 *tail = p2;
306                 do {
307                         tail = &p2->next;
308                         if ((p2 = *tail) == NULL)  {
309                                 *tail = p1;
310                                 return list1;
311                         }
312                 } while (p1->s_lno > p2->s_lno);
313                 *tail = p1;
314                 do {
315                         tail = &p1->next;
316                         if ((p1 = *tail) == NULL) {
317                                 *tail = p2;
318                                 return list1;
319                         }
320                 } while (p1->s_lno <= p2->s_lno);
321         }
322 }
323
324 static void *get_next_blame(const void *p)
325 {
326         return ((struct blame_entry *)p)->next;
327 }
328
329 static void set_next_blame(void *p1, void *p2)
330 {
331         ((struct blame_entry *)p1)->next = p2;
332 }
333
334 /*
335  * Final image line numbers are all different, so we don't need a
336  * three-way comparison here.
337  */
338
339 static int compare_blame_final(const void *p1, const void *p2)
340 {
341         return ((struct blame_entry *)p1)->lno > ((struct blame_entry *)p2)->lno
342                 ? 1 : -1;
343 }
344
345 static int compare_blame_suspect(const void *p1, const void *p2)
346 {
347         const struct blame_entry *s1 = p1, *s2 = p2;
348         /*
349          * to allow for collating suspects, we sort according to the
350          * respective pointer value as the primary sorting criterion.
351          * The actual relation is pretty unimportant as long as it
352          * establishes a total order.  Comparing as integers gives us
353          * that.
354          */
355         if (s1->suspect != s2->suspect)
356                 return (intptr_t)s1->suspect > (intptr_t)s2->suspect ? 1 : -1;
357         if (s1->s_lno == s2->s_lno)
358                 return 0;
359         return s1->s_lno > s2->s_lno ? 1 : -1;
360 }
361
362 static struct blame_entry *blame_sort(struct blame_entry *head,
363                                       int (*compare_fn)(const void *, const void *))
364 {
365         return llist_mergesort (head, get_next_blame, set_next_blame, compare_fn);
366 }
367
368 static int compare_commits_by_reverse_commit_date(const void *a,
369                                                   const void *b,
370                                                   void *c)
371 {
372         return -compare_commits_by_commit_date(a, b, c);
373 }
374
375 /*
376  * The current state of the blame assignment.
377  */
378 struct scoreboard {
379         /* the final commit (i.e. where we started digging from) */
380         struct commit *final;
381         /* Priority queue for commits with unassigned blame records */
382         struct prio_queue commits;
383         struct rev_info *revs;
384         const char *path;
385
386         /*
387          * The contents in the final image.
388          * Used by many functions to obtain contents of the nth line,
389          * indexed with scoreboard.lineno[blame_entry.lno].
390          */
391         const char *final_buf;
392         unsigned long final_buf_size;
393
394         /* linked list of blames */
395         struct blame_entry *ent;
396
397         /* look-up a line in the final buffer */
398         int num_lines;
399         int *lineno;
400 };
401
402 static void sanity_check_refcnt(struct scoreboard *);
403
404 /*
405  * If two blame entries that are next to each other came from
406  * contiguous lines in the same origin (i.e. <commit, path> pair),
407  * merge them together.
408  */
409 static void coalesce(struct scoreboard *sb)
410 {
411         struct blame_entry *ent, *next;
412
413         for (ent = sb->ent; ent && (next = ent->next); ent = next) {
414                 if (ent->suspect == next->suspect &&
415                     ent->s_lno + ent->num_lines == next->s_lno) {
416                         ent->num_lines += next->num_lines;
417                         ent->next = next->next;
418                         origin_decref(next->suspect);
419                         free(next);
420                         ent->score = 0;
421                         next = ent; /* again */
422                 }
423         }
424
425         if (DEBUG) /* sanity */
426                 sanity_check_refcnt(sb);
427 }
428
429 /*
430  * Merge the given sorted list of blames into a preexisting origin.
431  * If there were no previous blames to that commit, it is entered into
432  * the commit priority queue of the score board.
433  */
434
435 static void queue_blames(struct scoreboard *sb, struct origin *porigin,
436                          struct blame_entry *sorted)
437 {
438         if (porigin->suspects)
439                 porigin->suspects = blame_merge(porigin->suspects, sorted);
440         else {
441                 struct origin *o;
442                 for (o = porigin->commit->util; o; o = o->next) {
443                         if (o->suspects) {
444                                 porigin->suspects = sorted;
445                                 return;
446                         }
447                 }
448                 porigin->suspects = sorted;
449                 prio_queue_put(&sb->commits, porigin->commit);
450         }
451 }
452
453 /*
454  * Given a commit and a path in it, create a new origin structure.
455  * The callers that add blame to the scoreboard should use
456  * get_origin() to obtain shared, refcounted copy instead of calling
457  * this function directly.
458  */
459 static struct origin *make_origin(struct commit *commit, const char *path)
460 {
461         struct origin *o;
462         FLEX_ALLOC_STR(o, path, path);
463         o->commit = commit;
464         o->refcnt = 1;
465         o->next = commit->util;
466         commit->util = o;
467         return o;
468 }
469
470 /*
471  * Locate an existing origin or create a new one.
472  * This moves the origin to front position in the commit util list.
473  */
474 static struct origin *get_origin(struct scoreboard *sb,
475                                  struct commit *commit,
476                                  const char *path)
477 {
478         struct origin *o, *l;
479
480         for (o = commit->util, l = NULL; o; l = o, o = o->next) {
481                 if (!strcmp(o->path, path)) {
482                         /* bump to front */
483                         if (l) {
484                                 l->next = o->next;
485                                 o->next = commit->util;
486                                 commit->util = o;
487                         }
488                         return origin_incref(o);
489                 }
490         }
491         return make_origin(commit, path);
492 }
493
494 /*
495  * Fill the blob_sha1 field of an origin if it hasn't, so that later
496  * call to fill_origin_blob() can use it to locate the data.  blob_sha1
497  * for an origin is also used to pass the blame for the entire file to
498  * the parent to detect the case where a child's blob is identical to
499  * that of its parent's.
500  *
501  * This also fills origin->mode for corresponding tree path.
502  */
503 static int fill_blob_sha1_and_mode(struct origin *origin)
504 {
505         if (!is_null_sha1(origin->blob_sha1))
506                 return 0;
507         if (get_tree_entry(origin->commit->object.oid.hash,
508                            origin->path,
509                            origin->blob_sha1, &origin->mode))
510                 goto error_out;
511         if (sha1_object_info(origin->blob_sha1, NULL) != OBJ_BLOB)
512                 goto error_out;
513         return 0;
514  error_out:
515         hashclr(origin->blob_sha1);
516         origin->mode = S_IFINVALID;
517         return -1;
518 }
519
520 /*
521  * We have an origin -- check if the same path exists in the
522  * parent and return an origin structure to represent it.
523  */
524 static struct origin *find_origin(struct scoreboard *sb,
525                                   struct commit *parent,
526                                   struct origin *origin)
527 {
528         struct origin *porigin;
529         struct diff_options diff_opts;
530         const char *paths[2];
531
532         /* First check any existing origins */
533         for (porigin = parent->util; porigin; porigin = porigin->next)
534                 if (!strcmp(porigin->path, origin->path)) {
535                         /*
536                          * The same path between origin and its parent
537                          * without renaming -- the most common case.
538                          */
539                         return origin_incref (porigin);
540                 }
541
542         /* See if the origin->path is different between parent
543          * and origin first.  Most of the time they are the
544          * same and diff-tree is fairly efficient about this.
545          */
546         diff_setup(&diff_opts);
547         DIFF_OPT_SET(&diff_opts, RECURSIVE);
548         diff_opts.detect_rename = 0;
549         diff_opts.output_format = DIFF_FORMAT_NO_OUTPUT;
550         paths[0] = origin->path;
551         paths[1] = NULL;
552
553         parse_pathspec(&diff_opts.pathspec,
554                        PATHSPEC_ALL_MAGIC & ~PATHSPEC_LITERAL,
555                        PATHSPEC_LITERAL_PATH, "", paths);
556         diff_setup_done(&diff_opts);
557
558         if (is_null_oid(&origin->commit->object.oid))
559                 do_diff_cache(parent->tree->object.oid.hash, &diff_opts);
560         else
561                 diff_tree_sha1(parent->tree->object.oid.hash,
562                                origin->commit->tree->object.oid.hash,
563                                "", &diff_opts);
564         diffcore_std(&diff_opts);
565
566         if (!diff_queued_diff.nr) {
567                 /* The path is the same as parent */
568                 porigin = get_origin(sb, parent, origin->path);
569                 hashcpy(porigin->blob_sha1, origin->blob_sha1);
570                 porigin->mode = origin->mode;
571         } else {
572                 /*
573                  * Since origin->path is a pathspec, if the parent
574                  * commit had it as a directory, we will see a whole
575                  * bunch of deletion of files in the directory that we
576                  * do not care about.
577                  */
578                 int i;
579                 struct diff_filepair *p = NULL;
580                 for (i = 0; i < diff_queued_diff.nr; i++) {
581                         const char *name;
582                         p = diff_queued_diff.queue[i];
583                         name = p->one->path ? p->one->path : p->two->path;
584                         if (!strcmp(name, origin->path))
585                                 break;
586                 }
587                 if (!p)
588                         die("internal error in blame::find_origin");
589                 switch (p->status) {
590                 default:
591                         die("internal error in blame::find_origin (%c)",
592                             p->status);
593                 case 'M':
594                         porigin = get_origin(sb, parent, origin->path);
595                         hashcpy(porigin->blob_sha1, p->one->sha1);
596                         porigin->mode = p->one->mode;
597                         break;
598                 case 'A':
599                 case 'T':
600                         /* Did not exist in parent, or type changed */
601                         break;
602                 }
603         }
604         diff_flush(&diff_opts);
605         free_pathspec(&diff_opts.pathspec);
606         return porigin;
607 }
608
609 /*
610  * We have an origin -- find the path that corresponds to it in its
611  * parent and return an origin structure to represent it.
612  */
613 static struct origin *find_rename(struct scoreboard *sb,
614                                   struct commit *parent,
615                                   struct origin *origin)
616 {
617         struct origin *porigin = NULL;
618         struct diff_options diff_opts;
619         int i;
620
621         diff_setup(&diff_opts);
622         DIFF_OPT_SET(&diff_opts, RECURSIVE);
623         diff_opts.detect_rename = DIFF_DETECT_RENAME;
624         diff_opts.output_format = DIFF_FORMAT_NO_OUTPUT;
625         diff_opts.single_follow = origin->path;
626         diff_setup_done(&diff_opts);
627
628         if (is_null_oid(&origin->commit->object.oid))
629                 do_diff_cache(parent->tree->object.oid.hash, &diff_opts);
630         else
631                 diff_tree_sha1(parent->tree->object.oid.hash,
632                                origin->commit->tree->object.oid.hash,
633                                "", &diff_opts);
634         diffcore_std(&diff_opts);
635
636         for (i = 0; i < diff_queued_diff.nr; i++) {
637                 struct diff_filepair *p = diff_queued_diff.queue[i];
638                 if ((p->status == 'R' || p->status == 'C') &&
639                     !strcmp(p->two->path, origin->path)) {
640                         porigin = get_origin(sb, parent, p->one->path);
641                         hashcpy(porigin->blob_sha1, p->one->sha1);
642                         porigin->mode = p->one->mode;
643                         break;
644                 }
645         }
646         diff_flush(&diff_opts);
647         free_pathspec(&diff_opts.pathspec);
648         return porigin;
649 }
650
651 /*
652  * Append a new blame entry to a given output queue.
653  */
654 static void add_blame_entry(struct blame_entry ***queue, struct blame_entry *e)
655 {
656         origin_incref(e->suspect);
657
658         e->next = **queue;
659         **queue = e;
660         *queue = &e->next;
661 }
662
663 /*
664  * src typically is on-stack; we want to copy the information in it to
665  * a malloced blame_entry that gets added to the given queue.  The
666  * origin of dst loses a refcnt.
667  */
668 static void dup_entry(struct blame_entry ***queue,
669                       struct blame_entry *dst, struct blame_entry *src)
670 {
671         origin_incref(src->suspect);
672         origin_decref(dst->suspect);
673         memcpy(dst, src, sizeof(*src));
674         dst->next = **queue;
675         **queue = dst;
676         *queue = &dst->next;
677 }
678
679 static const char *nth_line(struct scoreboard *sb, long lno)
680 {
681         return sb->final_buf + sb->lineno[lno];
682 }
683
684 static const char *nth_line_cb(void *data, long lno)
685 {
686         return nth_line((struct scoreboard *)data, lno);
687 }
688
689 /*
690  * It is known that lines between tlno to same came from parent, and e
691  * has an overlap with that range.  it also is known that parent's
692  * line plno corresponds to e's line tlno.
693  *
694  *                <---- e ----->
695  *                   <------>
696  *                   <------------>
697  *             <------------>
698  *             <------------------>
699  *
700  * Split e into potentially three parts; before this chunk, the chunk
701  * to be blamed for the parent, and after that portion.
702  */
703 static void split_overlap(struct blame_entry *split,
704                           struct blame_entry *e,
705                           int tlno, int plno, int same,
706                           struct origin *parent)
707 {
708         int chunk_end_lno;
709         memset(split, 0, sizeof(struct blame_entry [3]));
710
711         if (e->s_lno < tlno) {
712                 /* there is a pre-chunk part not blamed on parent */
713                 split[0].suspect = origin_incref(e->suspect);
714                 split[0].lno = e->lno;
715                 split[0].s_lno = e->s_lno;
716                 split[0].num_lines = tlno - e->s_lno;
717                 split[1].lno = e->lno + tlno - e->s_lno;
718                 split[1].s_lno = plno;
719         }
720         else {
721                 split[1].lno = e->lno;
722                 split[1].s_lno = plno + (e->s_lno - tlno);
723         }
724
725         if (same < e->s_lno + e->num_lines) {
726                 /* there is a post-chunk part not blamed on parent */
727                 split[2].suspect = origin_incref(e->suspect);
728                 split[2].lno = e->lno + (same - e->s_lno);
729                 split[2].s_lno = e->s_lno + (same - e->s_lno);
730                 split[2].num_lines = e->s_lno + e->num_lines - same;
731                 chunk_end_lno = split[2].lno;
732         }
733         else
734                 chunk_end_lno = e->lno + e->num_lines;
735         split[1].num_lines = chunk_end_lno - split[1].lno;
736
737         /*
738          * if it turns out there is nothing to blame the parent for,
739          * forget about the splitting.  !split[1].suspect signals this.
740          */
741         if (split[1].num_lines < 1)
742                 return;
743         split[1].suspect = origin_incref(parent);
744 }
745
746 /*
747  * split_overlap() divided an existing blame e into up to three parts
748  * in split.  Any assigned blame is moved to queue to
749  * reflect the split.
750  */
751 static void split_blame(struct blame_entry ***blamed,
752                         struct blame_entry ***unblamed,
753                         struct blame_entry *split,
754                         struct blame_entry *e)
755 {
756         struct blame_entry *new_entry;
757
758         if (split[0].suspect && split[2].suspect) {
759                 /* The first part (reuse storage for the existing entry e) */
760                 dup_entry(unblamed, e, &split[0]);
761
762                 /* The last part -- me */
763                 new_entry = xmalloc(sizeof(*new_entry));
764                 memcpy(new_entry, &(split[2]), sizeof(struct blame_entry));
765                 add_blame_entry(unblamed, new_entry);
766
767                 /* ... and the middle part -- parent */
768                 new_entry = xmalloc(sizeof(*new_entry));
769                 memcpy(new_entry, &(split[1]), sizeof(struct blame_entry));
770                 add_blame_entry(blamed, new_entry);
771         }
772         else if (!split[0].suspect && !split[2].suspect)
773                 /*
774                  * The parent covers the entire area; reuse storage for
775                  * e and replace it with the parent.
776                  */
777                 dup_entry(blamed, e, &split[1]);
778         else if (split[0].suspect) {
779                 /* me and then parent */
780                 dup_entry(unblamed, e, &split[0]);
781
782                 new_entry = xmalloc(sizeof(*new_entry));
783                 memcpy(new_entry, &(split[1]), sizeof(struct blame_entry));
784                 add_blame_entry(blamed, new_entry);
785         }
786         else {
787                 /* parent and then me */
788                 dup_entry(blamed, e, &split[1]);
789
790                 new_entry = xmalloc(sizeof(*new_entry));
791                 memcpy(new_entry, &(split[2]), sizeof(struct blame_entry));
792                 add_blame_entry(unblamed, new_entry);
793         }
794 }
795
796 /*
797  * After splitting the blame, the origins used by the
798  * on-stack blame_entry should lose one refcnt each.
799  */
800 static void decref_split(struct blame_entry *split)
801 {
802         int i;
803
804         for (i = 0; i < 3; i++)
805                 origin_decref(split[i].suspect);
806 }
807
808 /*
809  * reverse_blame reverses the list given in head, appending tail.
810  * That allows us to build lists in reverse order, then reverse them
811  * afterwards.  This can be faster than building the list in proper
812  * order right away.  The reason is that building in proper order
813  * requires writing a link in the _previous_ element, while building
814  * in reverse order just requires placing the list head into the
815  * _current_ element.
816  */
817
818 static struct blame_entry *reverse_blame(struct blame_entry *head,
819                                          struct blame_entry *tail)
820 {
821         while (head) {
822                 struct blame_entry *next = head->next;
823                 head->next = tail;
824                 tail = head;
825                 head = next;
826         }
827         return tail;
828 }
829
830 /*
831  * Process one hunk from the patch between the current suspect for
832  * blame_entry e and its parent.  This first blames any unfinished
833  * entries before the chunk (which is where target and parent start
834  * differing) on the parent, and then splits blame entries at the
835  * start and at the end of the difference region.  Since use of -M and
836  * -C options may lead to overlapping/duplicate source line number
837  * ranges, all we can rely on from sorting/merging is the order of the
838  * first suspect line number.
839  */
840 static void blame_chunk(struct blame_entry ***dstq, struct blame_entry ***srcq,
841                         int tlno, int offset, int same,
842                         struct origin *parent)
843 {
844         struct blame_entry *e = **srcq;
845         struct blame_entry *samep = NULL, *diffp = NULL;
846
847         while (e && e->s_lno < tlno) {
848                 struct blame_entry *next = e->next;
849                 /*
850                  * current record starts before differing portion.  If
851                  * it reaches into it, we need to split it up and
852                  * examine the second part separately.
853                  */
854                 if (e->s_lno + e->num_lines > tlno) {
855                         /* Move second half to a new record */
856                         int len = tlno - e->s_lno;
857                         struct blame_entry *n = xcalloc(1, sizeof (struct blame_entry));
858                         n->suspect = e->suspect;
859                         n->lno = e->lno + len;
860                         n->s_lno = e->s_lno + len;
861                         n->num_lines = e->num_lines - len;
862                         e->num_lines = len;
863                         e->score = 0;
864                         /* Push new record to diffp */
865                         n->next = diffp;
866                         diffp = n;
867                 } else
868                         origin_decref(e->suspect);
869                 /* Pass blame for everything before the differing
870                  * chunk to the parent */
871                 e->suspect = origin_incref(parent);
872                 e->s_lno += offset;
873                 e->next = samep;
874                 samep = e;
875                 e = next;
876         }
877         /*
878          * As we don't know how much of a common stretch after this
879          * diff will occur, the currently blamed parts are all that we
880          * can assign to the parent for now.
881          */
882
883         if (samep) {
884                 **dstq = reverse_blame(samep, **dstq);
885                 *dstq = &samep->next;
886         }
887         /*
888          * Prepend the split off portions: everything after e starts
889          * after the blameable portion.
890          */
891         e = reverse_blame(diffp, e);
892
893         /*
894          * Now retain records on the target while parts are different
895          * from the parent.
896          */
897         samep = NULL;
898         diffp = NULL;
899         while (e && e->s_lno < same) {
900                 struct blame_entry *next = e->next;
901
902                 /*
903                  * If current record extends into sameness, need to split.
904                  */
905                 if (e->s_lno + e->num_lines > same) {
906                         /*
907                          * Move second half to a new record to be
908                          * processed by later chunks
909                          */
910                         int len = same - e->s_lno;
911                         struct blame_entry *n = xcalloc(1, sizeof (struct blame_entry));
912                         n->suspect = origin_incref(e->suspect);
913                         n->lno = e->lno + len;
914                         n->s_lno = e->s_lno + len;
915                         n->num_lines = e->num_lines - len;
916                         e->num_lines = len;
917                         e->score = 0;
918                         /* Push new record to samep */
919                         n->next = samep;
920                         samep = n;
921                 }
922                 e->next = diffp;
923                 diffp = e;
924                 e = next;
925         }
926         **srcq = reverse_blame(diffp, reverse_blame(samep, e));
927         /* Move across elements that are in the unblamable portion */
928         if (diffp)
929                 *srcq = &diffp->next;
930 }
931
932 struct blame_chunk_cb_data {
933         struct origin *parent;
934         long offset;
935         struct blame_entry **dstq;
936         struct blame_entry **srcq;
937 };
938
939 /* diff chunks are from parent to target */
940 static int blame_chunk_cb(long start_a, long count_a,
941                           long start_b, long count_b, void *data)
942 {
943         struct blame_chunk_cb_data *d = data;
944         if (start_a - start_b != d->offset)
945                 die("internal error in blame::blame_chunk_cb");
946         blame_chunk(&d->dstq, &d->srcq, start_b, start_a - start_b,
947                     start_b + count_b, d->parent);
948         d->offset = start_a + count_a - (start_b + count_b);
949         return 0;
950 }
951
952 /*
953  * We are looking at the origin 'target' and aiming to pass blame
954  * for the lines it is suspected to its parent.  Run diff to find
955  * which lines came from parent and pass blame for them.
956  */
957 static void pass_blame_to_parent(struct scoreboard *sb,
958                                  struct origin *target,
959                                  struct origin *parent)
960 {
961         mmfile_t file_p, file_o;
962         struct blame_chunk_cb_data d;
963         struct blame_entry *newdest = NULL;
964
965         if (!target->suspects)
966                 return; /* nothing remains for this target */
967
968         d.parent = parent;
969         d.offset = 0;
970         d.dstq = &newdest; d.srcq = &target->suspects;
971
972         fill_origin_blob(&sb->revs->diffopt, parent, &file_p);
973         fill_origin_blob(&sb->revs->diffopt, target, &file_o);
974         num_get_patch++;
975
976         if (diff_hunks(&file_p, &file_o, 0, blame_chunk_cb, &d))
977                 die("unable to generate diff (%s -> %s)",
978                     oid_to_hex(&parent->commit->object.oid),
979                     oid_to_hex(&target->commit->object.oid));
980         /* The rest are the same as the parent */
981         blame_chunk(&d.dstq, &d.srcq, INT_MAX, d.offset, INT_MAX, parent);
982         *d.dstq = NULL;
983         queue_blames(sb, parent, newdest);
984
985         return;
986 }
987
988 /*
989  * The lines in blame_entry after splitting blames many times can become
990  * very small and trivial, and at some point it becomes pointless to
991  * blame the parents.  E.g. "\t\t}\n\t}\n\n" appears everywhere in any
992  * ordinary C program, and it is not worth to say it was copied from
993  * totally unrelated file in the parent.
994  *
995  * Compute how trivial the lines in the blame_entry are.
996  */
997 static unsigned ent_score(struct scoreboard *sb, struct blame_entry *e)
998 {
999         unsigned score;
1000         const char *cp, *ep;
1001
1002         if (e->score)
1003                 return e->score;
1004
1005         score = 1;
1006         cp = nth_line(sb, e->lno);
1007         ep = nth_line(sb, e->lno + e->num_lines);
1008         while (cp < ep) {
1009                 unsigned ch = *((unsigned char *)cp);
1010                 if (isalnum(ch))
1011                         score++;
1012                 cp++;
1013         }
1014         e->score = score;
1015         return score;
1016 }
1017
1018 /*
1019  * best_so_far[] and this[] are both a split of an existing blame_entry
1020  * that passes blame to the parent.  Maintain best_so_far the best split
1021  * so far, by comparing this and best_so_far and copying this into
1022  * bst_so_far as needed.
1023  */
1024 static void copy_split_if_better(struct scoreboard *sb,
1025                                  struct blame_entry *best_so_far,
1026                                  struct blame_entry *this)
1027 {
1028         int i;
1029
1030         if (!this[1].suspect)
1031                 return;
1032         if (best_so_far[1].suspect) {
1033                 if (ent_score(sb, &this[1]) < ent_score(sb, &best_so_far[1]))
1034                         return;
1035         }
1036
1037         for (i = 0; i < 3; i++)
1038                 origin_incref(this[i].suspect);
1039         decref_split(best_so_far);
1040         memcpy(best_so_far, this, 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 scoreboard *sb,
1059                          struct blame_entry *ent,
1060                          int tlno, int plno, int same,
1061                          struct origin *parent,
1062                          struct blame_entry *split)
1063 {
1064         if (ent->num_lines <= tlno)
1065                 return;
1066         if (tlno < same) {
1067                 struct blame_entry this[3];
1068                 tlno += ent->s_lno;
1069                 same += ent->s_lno;
1070                 split_overlap(this, ent, tlno, plno, same, parent);
1071                 copy_split_if_better(sb, split, this);
1072                 decref_split(this);
1073         }
1074 }
1075
1076 struct handle_split_cb_data {
1077         struct scoreboard *sb;
1078         struct blame_entry *ent;
1079         struct 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 scoreboard *sb,
1102                               struct blame_entry *ent,
1103                               struct 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 = nth_line(sb, ent->lno);
1117         file_o.ptr = (char *) cp;
1118         file_o.size = 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, 1, handle_split_cb, &d))
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 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 (ent_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 scoreboard *sb,
1165                                 struct blame_entry ***blamed,
1166                                 struct blame_entry **toosmall,
1167                                 struct origin *target,
1168                                 struct 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);
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                             blame_move_score < ent_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, blame_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 scoreboard *sb,
1241                                 struct blame_entry ***blamed,
1242                                 struct blame_entry **toosmall,
1243                                 struct origin *target,
1244                                 struct commit *parent,
1245                                 struct 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_OPT_SET(&diff_opts, RECURSIVE);
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_sha1() 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_OPT_SET(&diff_opts, FIND_COPIES_HARDER);
1275
1276         if (is_null_oid(&target->commit->object.oid))
1277                 do_diff_cache(parent->tree->object.oid.hash, &diff_opts);
1278         else
1279                 diff_tree_sha1(parent->tree->object.oid.hash,
1280                                target->commit->tree->object.oid.hash,
1281                                "", &diff_opts);
1282
1283         if (!DIFF_OPT_TST(&diff_opts, 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 origin *norigin;
1293                         mmfile_t file_p;
1294                         struct blame_entry this[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(sb, parent, p->one->path);
1305                         hashcpy(norigin->blob_sha1, p->one->sha1);
1306                         norigin->mode = p->one->mode;
1307                         fill_origin_blob(&sb->revs->diffopt, norigin, &file_p);
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, this, &file_p);
1314                                 copy_split_if_better(sb, blame_list[j].split,
1315                                                      this);
1316                                 decref_split(this);
1317                         }
1318                         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                             blame_copy_score < ent_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, blame_copy_score);
1336         } while (unblamed);
1337         target->suspects = reverse_blame(leftover, NULL);
1338         diff_flush(&diff_opts);
1339         free_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 scoreboard *sb,
1347                              struct origin *origin, struct 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                 origin_incref(porigin);
1360                 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 {
1373         if (!reverse) {
1374                 if (revs->first_parent_only &&
1375                     commit->parents &&
1376                     commit->parents->next) {
1377                         free_commit_list(commit->parents->next);
1378                         commit->parents->next = NULL;
1379                 }
1380                 return commit->parents;
1381         }
1382         return lookup_decoration(&revs->children, &commit->object);
1383 }
1384
1385 static int num_scapegoats(struct rev_info *revs, struct commit *commit)
1386 {
1387         struct commit_list *l = first_scapegoat(revs, commit);
1388         return commit_list_count(l);
1389 }
1390
1391 /* Distribute collected unsorted blames to the respected sorted lists
1392  * in the various origins.
1393  */
1394 static void distribute_blame(struct scoreboard *sb, struct blame_entry *blamed)
1395 {
1396         blamed = blame_sort(blamed, compare_blame_suspect);
1397         while (blamed)
1398         {
1399                 struct origin *porigin = blamed->suspect;
1400                 struct blame_entry *suspects = NULL;
1401                 do {
1402                         struct blame_entry *next = blamed->next;
1403                         blamed->next = suspects;
1404                         suspects = blamed;
1405                         blamed = next;
1406                 } while (blamed && blamed->suspect == porigin);
1407                 suspects = reverse_blame(suspects, NULL);
1408                 queue_blames(sb, porigin, suspects);
1409         }
1410 }
1411
1412 #define MAXSG 16
1413
1414 static void pass_blame(struct scoreboard *sb, struct origin *origin, int opt)
1415 {
1416         struct rev_info *revs = sb->revs;
1417         int i, pass, num_sg;
1418         struct commit *commit = origin->commit;
1419         struct commit_list *sg;
1420         struct origin *sg_buf[MAXSG];
1421         struct origin *porigin, **sg_origin = sg_buf;
1422         struct blame_entry *toosmall = NULL;
1423         struct blame_entry *blames, **blametail = &blames;
1424
1425         num_sg = num_scapegoats(revs, commit);
1426         if (!num_sg)
1427                 goto finish;
1428         else if (num_sg < ARRAY_SIZE(sg_buf))
1429                 memset(sg_buf, 0, sizeof(sg_buf));
1430         else
1431                 sg_origin = xcalloc(num_sg, sizeof(*sg_origin));
1432
1433         /*
1434          * The first pass looks for unrenamed path to optimize for
1435          * common cases, then we look for renames in the second pass.
1436          */
1437         for (pass = 0; pass < 2 - no_whole_file_rename; pass++) {
1438                 struct origin *(*find)(struct scoreboard *,
1439                                        struct commit *, struct origin *);
1440                 find = pass ? find_rename : find_origin;
1441
1442                 for (i = 0, sg = first_scapegoat(revs, commit);
1443                      i < num_sg && sg;
1444                      sg = sg->next, i++) {
1445                         struct commit *p = sg->item;
1446                         int j, same;
1447
1448                         if (sg_origin[i])
1449                                 continue;
1450                         if (parse_commit(p))
1451                                 continue;
1452                         porigin = find(sb, p, origin);
1453                         if (!porigin)
1454                                 continue;
1455                         if (!hashcmp(porigin->blob_sha1, origin->blob_sha1)) {
1456                                 pass_whole_blame(sb, origin, porigin);
1457                                 origin_decref(porigin);
1458                                 goto finish;
1459                         }
1460                         for (j = same = 0; j < i; j++)
1461                                 if (sg_origin[j] &&
1462                                     !hashcmp(sg_origin[j]->blob_sha1,
1463                                              porigin->blob_sha1)) {
1464                                         same = 1;
1465                                         break;
1466                                 }
1467                         if (!same)
1468                                 sg_origin[i] = porigin;
1469                         else
1470                                 origin_decref(porigin);
1471                 }
1472         }
1473
1474         num_commits++;
1475         for (i = 0, sg = first_scapegoat(revs, commit);
1476              i < num_sg && sg;
1477              sg = sg->next, i++) {
1478                 struct origin *porigin = sg_origin[i];
1479                 if (!porigin)
1480                         continue;
1481                 if (!origin->previous) {
1482                         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, blame_move_score);
1495                 if (origin->suspects) {
1496                         for (i = 0, sg = first_scapegoat(revs, commit);
1497                              i < num_sg && sg;
1498                              sg = sg->next, i++) {
1499                                 struct 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 (blame_copy_score > blame_move_score)
1514                         filter_small(sb, &toosmall, &origin->suspects, blame_copy_score);
1515                 else if (blame_copy_score < blame_move_score) {
1516                         origin->suspects = blame_merge(origin->suspects, toosmall);
1517                         toosmall = NULL;
1518                         filter_small(sb, &toosmall, &origin->suspects, blame_copy_score);
1519                 }
1520                 if (!origin->suspects)
1521                         goto finish;
1522
1523                 for (i = 0, sg = first_scapegoat(revs, commit);
1524                      i < num_sg && sg;
1525                      sg = sg->next, i++) {
1526                         struct 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                         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  * Information on commits, used for output.
1563  */
1564 struct commit_info {
1565         struct strbuf author;
1566         struct strbuf author_mail;
1567         unsigned long author_time;
1568         struct strbuf author_tz;
1569
1570         /* filled only when asked for details */
1571         struct strbuf committer;
1572         struct strbuf committer_mail;
1573         unsigned long committer_time;
1574         struct strbuf committer_tz;
1575
1576         struct strbuf summary;
1577 };
1578
1579 /*
1580  * Parse author/committer line in the commit object buffer
1581  */
1582 static void get_ac_line(const char *inbuf, const char *what,
1583         struct strbuf *name, struct strbuf *mail,
1584         unsigned long *time, struct strbuf *tz)
1585 {
1586         struct ident_split ident;
1587         size_t len, maillen, namelen;
1588         char *tmp, *endp;
1589         const char *namebuf, *mailbuf;
1590
1591         tmp = strstr(inbuf, what);
1592         if (!tmp)
1593                 goto error_out;
1594         tmp += strlen(what);
1595         endp = strchr(tmp, '\n');
1596         if (!endp)
1597                 len = strlen(tmp);
1598         else
1599                 len = endp - tmp;
1600
1601         if (split_ident_line(&ident, tmp, len)) {
1602         error_out:
1603                 /* Ugh */
1604                 tmp = "(unknown)";
1605                 strbuf_addstr(name, tmp);
1606                 strbuf_addstr(mail, tmp);
1607                 strbuf_addstr(tz, tmp);
1608                 *time = 0;
1609                 return;
1610         }
1611
1612         namelen = ident.name_end - ident.name_begin;
1613         namebuf = ident.name_begin;
1614
1615         maillen = ident.mail_end - ident.mail_begin;
1616         mailbuf = ident.mail_begin;
1617
1618         if (ident.date_begin && ident.date_end)
1619                 *time = strtoul(ident.date_begin, NULL, 10);
1620         else
1621                 *time = 0;
1622
1623         if (ident.tz_begin && ident.tz_end)
1624                 strbuf_add(tz, ident.tz_begin, ident.tz_end - ident.tz_begin);
1625         else
1626                 strbuf_addstr(tz, "(unknown)");
1627
1628         /*
1629          * Now, convert both name and e-mail using mailmap
1630          */
1631         map_user(&mailmap, &mailbuf, &maillen,
1632                  &namebuf, &namelen);
1633
1634         strbuf_addf(mail, "<%.*s>", (int)maillen, mailbuf);
1635         strbuf_add(name, namebuf, namelen);
1636 }
1637
1638 static void commit_info_init(struct commit_info *ci)
1639 {
1640
1641         strbuf_init(&ci->author, 0);
1642         strbuf_init(&ci->author_mail, 0);
1643         strbuf_init(&ci->author_tz, 0);
1644         strbuf_init(&ci->committer, 0);
1645         strbuf_init(&ci->committer_mail, 0);
1646         strbuf_init(&ci->committer_tz, 0);
1647         strbuf_init(&ci->summary, 0);
1648 }
1649
1650 static void commit_info_destroy(struct commit_info *ci)
1651 {
1652
1653         strbuf_release(&ci->author);
1654         strbuf_release(&ci->author_mail);
1655         strbuf_release(&ci->author_tz);
1656         strbuf_release(&ci->committer);
1657         strbuf_release(&ci->committer_mail);
1658         strbuf_release(&ci->committer_tz);
1659         strbuf_release(&ci->summary);
1660 }
1661
1662 static void get_commit_info(struct commit *commit,
1663                             struct commit_info *ret,
1664                             int detailed)
1665 {
1666         int len;
1667         const char *subject, *encoding;
1668         const char *message;
1669
1670         commit_info_init(ret);
1671
1672         encoding = get_log_output_encoding();
1673         message = logmsg_reencode(commit, NULL, encoding);
1674         get_ac_line(message, "\nauthor ",
1675                     &ret->author, &ret->author_mail,
1676                     &ret->author_time, &ret->author_tz);
1677
1678         if (!detailed) {
1679                 unuse_commit_buffer(commit, message);
1680                 return;
1681         }
1682
1683         get_ac_line(message, "\ncommitter ",
1684                     &ret->committer, &ret->committer_mail,
1685                     &ret->committer_time, &ret->committer_tz);
1686
1687         len = find_commit_subject(message, &subject);
1688         if (len)
1689                 strbuf_add(&ret->summary, subject, len);
1690         else
1691                 strbuf_addf(&ret->summary, "(%s)", oid_to_hex(&commit->object.oid));
1692
1693         unuse_commit_buffer(commit, message);
1694 }
1695
1696 /*
1697  * To allow LF and other nonportable characters in pathnames,
1698  * they are c-style quoted as needed.
1699  */
1700 static void write_filename_info(const char *path)
1701 {
1702         printf("filename ");
1703         write_name_quoted(path, stdout, '\n');
1704 }
1705
1706 /*
1707  * Porcelain/Incremental format wants to show a lot of details per
1708  * commit.  Instead of repeating this every line, emit it only once,
1709  * the first time each commit appears in the output (unless the
1710  * user has specifically asked for us to repeat).
1711  */
1712 static int emit_one_suspect_detail(struct origin *suspect, int repeat)
1713 {
1714         struct commit_info ci;
1715
1716         if (!repeat && (suspect->commit->object.flags & METAINFO_SHOWN))
1717                 return 0;
1718
1719         suspect->commit->object.flags |= METAINFO_SHOWN;
1720         get_commit_info(suspect->commit, &ci, 1);
1721         printf("author %s\n", ci.author.buf);
1722         printf("author-mail %s\n", ci.author_mail.buf);
1723         printf("author-time %lu\n", ci.author_time);
1724         printf("author-tz %s\n", ci.author_tz.buf);
1725         printf("committer %s\n", ci.committer.buf);
1726         printf("committer-mail %s\n", ci.committer_mail.buf);
1727         printf("committer-time %lu\n", ci.committer_time);
1728         printf("committer-tz %s\n", ci.committer_tz.buf);
1729         printf("summary %s\n", ci.summary.buf);
1730         if (suspect->commit->object.flags & UNINTERESTING)
1731                 printf("boundary\n");
1732         if (suspect->previous) {
1733                 struct origin *prev = suspect->previous;
1734                 printf("previous %s ", oid_to_hex(&prev->commit->object.oid));
1735                 write_name_quoted(prev->path, stdout, '\n');
1736         }
1737
1738         commit_info_destroy(&ci);
1739
1740         return 1;
1741 }
1742
1743 /*
1744  * The blame_entry is found to be guilty for the range.
1745  * Show it in incremental output.
1746  */
1747 static void found_guilty_entry(struct blame_entry *ent)
1748 {
1749         if (incremental) {
1750                 struct origin *suspect = ent->suspect;
1751
1752                 printf("%s %d %d %d\n",
1753                        oid_to_hex(&suspect->commit->object.oid),
1754                        ent->s_lno + 1, ent->lno + 1, ent->num_lines);
1755                 emit_one_suspect_detail(suspect, 0);
1756                 write_filename_info(suspect->path);
1757                 maybe_flush_or_die(stdout, "stdout");
1758         }
1759 }
1760
1761 /*
1762  * The main loop -- while we have blobs with lines whose true origin
1763  * is still unknown, pick one blob, and allow its lines to pass blames
1764  * to its parents. */
1765 static void assign_blame(struct scoreboard *sb, int opt)
1766 {
1767         struct rev_info *revs = sb->revs;
1768         struct commit *commit = prio_queue_get(&sb->commits);
1769
1770         while (commit) {
1771                 struct blame_entry *ent;
1772                 struct origin *suspect = commit->util;
1773
1774                 /* find one suspect to break down */
1775                 while (suspect && !suspect->suspects)
1776                         suspect = suspect->next;
1777
1778                 if (!suspect) {
1779                         commit = prio_queue_get(&sb->commits);
1780                         continue;
1781                 }
1782
1783                 assert(commit == suspect->commit);
1784
1785                 /*
1786                  * We will use this suspect later in the loop,
1787                  * so hold onto it in the meantime.
1788                  */
1789                 origin_incref(suspect);
1790                 parse_commit(commit);
1791                 if (reverse ||
1792                     (!(commit->object.flags & UNINTERESTING) &&
1793                      !(revs->max_age != -1 && commit->date < revs->max_age)))
1794                         pass_blame(sb, suspect, opt);
1795                 else {
1796                         commit->object.flags |= UNINTERESTING;
1797                         if (commit->object.parsed)
1798                                 mark_parents_uninteresting(commit);
1799                 }
1800                 /* treat root commit as boundary */
1801                 if (!commit->parents && !show_root)
1802                         commit->object.flags |= UNINTERESTING;
1803
1804                 /* Take responsibility for the remaining entries */
1805                 ent = suspect->suspects;
1806                 if (ent) {
1807                         suspect->guilty = 1;
1808                         for (;;) {
1809                                 struct blame_entry *next = ent->next;
1810                                 found_guilty_entry(ent);
1811                                 if (next) {
1812                                         ent = next;
1813                                         continue;
1814                                 }
1815                                 ent->next = sb->ent;
1816                                 sb->ent = suspect->suspects;
1817                                 suspect->suspects = NULL;
1818                                 break;
1819                         }
1820                 }
1821                 origin_decref(suspect);
1822
1823                 if (DEBUG) /* sanity */
1824                         sanity_check_refcnt(sb);
1825         }
1826 }
1827
1828 static const char *format_time(unsigned long time, const char *tz_str,
1829                                int show_raw_time)
1830 {
1831         static struct strbuf time_buf = STRBUF_INIT;
1832
1833         strbuf_reset(&time_buf);
1834         if (show_raw_time) {
1835                 strbuf_addf(&time_buf, "%lu %s", time, tz_str);
1836         }
1837         else {
1838                 const char *time_str;
1839                 size_t time_width;
1840                 int tz;
1841                 tz = atoi(tz_str);
1842                 time_str = show_date(time, tz, &blame_date_mode);
1843                 strbuf_addstr(&time_buf, time_str);
1844                 /*
1845                  * Add space paddings to time_buf to display a fixed width
1846                  * string, and use time_width for display width calibration.
1847                  */
1848                 for (time_width = utf8_strwidth(time_str);
1849                      time_width < blame_date_width;
1850                      time_width++)
1851                         strbuf_addch(&time_buf, ' ');
1852         }
1853         return time_buf.buf;
1854 }
1855
1856 #define OUTPUT_ANNOTATE_COMPAT  001
1857 #define OUTPUT_LONG_OBJECT_NAME 002
1858 #define OUTPUT_RAW_TIMESTAMP    004
1859 #define OUTPUT_PORCELAIN        010
1860 #define OUTPUT_SHOW_NAME        020
1861 #define OUTPUT_SHOW_NUMBER      040
1862 #define OUTPUT_SHOW_SCORE      0100
1863 #define OUTPUT_NO_AUTHOR       0200
1864 #define OUTPUT_SHOW_EMAIL       0400
1865 #define OUTPUT_LINE_PORCELAIN 01000
1866
1867 static void emit_porcelain_details(struct origin *suspect, int repeat)
1868 {
1869         if (emit_one_suspect_detail(suspect, repeat) ||
1870             (suspect->commit->object.flags & MORE_THAN_ONE_PATH))
1871                 write_filename_info(suspect->path);
1872 }
1873
1874 static void emit_porcelain(struct scoreboard *sb, struct blame_entry *ent,
1875                            int opt)
1876 {
1877         int repeat = opt & OUTPUT_LINE_PORCELAIN;
1878         int cnt;
1879         const char *cp;
1880         struct origin *suspect = ent->suspect;
1881         char hex[GIT_SHA1_HEXSZ + 1];
1882
1883         sha1_to_hex_r(hex, suspect->commit->object.oid.hash);
1884         printf("%s %d %d %d\n",
1885                hex,
1886                ent->s_lno + 1,
1887                ent->lno + 1,
1888                ent->num_lines);
1889         emit_porcelain_details(suspect, repeat);
1890
1891         cp = nth_line(sb, ent->lno);
1892         for (cnt = 0; cnt < ent->num_lines; cnt++) {
1893                 char ch;
1894                 if (cnt) {
1895                         printf("%s %d %d\n", hex,
1896                                ent->s_lno + 1 + cnt,
1897                                ent->lno + 1 + cnt);
1898                         if (repeat)
1899                                 emit_porcelain_details(suspect, 1);
1900                 }
1901                 putchar('\t');
1902                 do {
1903                         ch = *cp++;
1904                         putchar(ch);
1905                 } while (ch != '\n' &&
1906                          cp < sb->final_buf + sb->final_buf_size);
1907         }
1908
1909         if (sb->final_buf_size && cp[-1] != '\n')
1910                 putchar('\n');
1911 }
1912
1913 static void emit_other(struct scoreboard *sb, struct blame_entry *ent, int opt)
1914 {
1915         int cnt;
1916         const char *cp;
1917         struct origin *suspect = ent->suspect;
1918         struct commit_info ci;
1919         char hex[GIT_SHA1_HEXSZ + 1];
1920         int show_raw_time = !!(opt & OUTPUT_RAW_TIMESTAMP);
1921
1922         get_commit_info(suspect->commit, &ci, 1);
1923         sha1_to_hex_r(hex, suspect->commit->object.oid.hash);
1924
1925         cp = nth_line(sb, ent->lno);
1926         for (cnt = 0; cnt < ent->num_lines; cnt++) {
1927                 char ch;
1928                 int length = (opt & OUTPUT_LONG_OBJECT_NAME) ? 40 : abbrev;
1929
1930                 if (suspect->commit->object.flags & UNINTERESTING) {
1931                         if (blank_boundary)
1932                                 memset(hex, ' ', length);
1933                         else if (!(opt & OUTPUT_ANNOTATE_COMPAT)) {
1934                                 length--;
1935                                 putchar('^');
1936                         }
1937                 }
1938
1939                 printf("%.*s", length, hex);
1940                 if (opt & OUTPUT_ANNOTATE_COMPAT) {
1941                         const char *name;
1942                         if (opt & OUTPUT_SHOW_EMAIL)
1943                                 name = ci.author_mail.buf;
1944                         else
1945                                 name = ci.author.buf;
1946                         printf("\t(%10s\t%10s\t%d)", name,
1947                                format_time(ci.author_time, ci.author_tz.buf,
1948                                            show_raw_time),
1949                                ent->lno + 1 + cnt);
1950                 } else {
1951                         if (opt & OUTPUT_SHOW_SCORE)
1952                                 printf(" %*d %02d",
1953                                        max_score_digits, ent->score,
1954                                        ent->suspect->refcnt);
1955                         if (opt & OUTPUT_SHOW_NAME)
1956                                 printf(" %-*.*s", longest_file, longest_file,
1957                                        suspect->path);
1958                         if (opt & OUTPUT_SHOW_NUMBER)
1959                                 printf(" %*d", max_orig_digits,
1960                                        ent->s_lno + 1 + cnt);
1961
1962                         if (!(opt & OUTPUT_NO_AUTHOR)) {
1963                                 const char *name;
1964                                 int pad;
1965                                 if (opt & OUTPUT_SHOW_EMAIL)
1966                                         name = ci.author_mail.buf;
1967                                 else
1968                                         name = ci.author.buf;
1969                                 pad = longest_author - utf8_strwidth(name);
1970                                 printf(" (%s%*s %10s",
1971                                        name, pad, "",
1972                                        format_time(ci.author_time,
1973                                                    ci.author_tz.buf,
1974                                                    show_raw_time));
1975                         }
1976                         printf(" %*d) ",
1977                                max_digits, ent->lno + 1 + cnt);
1978                 }
1979                 do {
1980                         ch = *cp++;
1981                         putchar(ch);
1982                 } while (ch != '\n' &&
1983                          cp < sb->final_buf + sb->final_buf_size);
1984         }
1985
1986         if (sb->final_buf_size && cp[-1] != '\n')
1987                 putchar('\n');
1988
1989         commit_info_destroy(&ci);
1990 }
1991
1992 static void output(struct scoreboard *sb, int option)
1993 {
1994         struct blame_entry *ent;
1995
1996         if (option & OUTPUT_PORCELAIN) {
1997                 for (ent = sb->ent; ent; ent = ent->next) {
1998                         int count = 0;
1999                         struct origin *suspect;
2000                         struct commit *commit = ent->suspect->commit;
2001                         if (commit->object.flags & MORE_THAN_ONE_PATH)
2002                                 continue;
2003                         for (suspect = commit->util; suspect; suspect = suspect->next) {
2004                                 if (suspect->guilty && count++) {
2005                                         commit->object.flags |= MORE_THAN_ONE_PATH;
2006                                         break;
2007                                 }
2008                         }
2009                 }
2010         }
2011
2012         for (ent = sb->ent; ent; ent = ent->next) {
2013                 if (option & OUTPUT_PORCELAIN)
2014                         emit_porcelain(sb, ent, option);
2015                 else {
2016                         emit_other(sb, ent, option);
2017                 }
2018         }
2019 }
2020
2021 static const char *get_next_line(const char *start, const char *end)
2022 {
2023         const char *nl = memchr(start, '\n', end - start);
2024         return nl ? nl + 1 : end;
2025 }
2026
2027 /*
2028  * To allow quick access to the contents of nth line in the
2029  * final image, prepare an index in the scoreboard.
2030  */
2031 static int prepare_lines(struct scoreboard *sb)
2032 {
2033         const char *buf = sb->final_buf;
2034         unsigned long len = sb->final_buf_size;
2035         const char *end = buf + len;
2036         const char *p;
2037         int *lineno;
2038         int num = 0;
2039
2040         for (p = buf; p < end; p = get_next_line(p, end))
2041                 num++;
2042
2043         ALLOC_ARRAY(sb->lineno, num + 1);
2044         lineno = sb->lineno;
2045
2046         for (p = buf; p < end; p = get_next_line(p, end))
2047                 *lineno++ = p - buf;
2048
2049         *lineno = len;
2050
2051         sb->num_lines = num;
2052         return sb->num_lines;
2053 }
2054
2055 /*
2056  * Add phony grafts for use with -S; this is primarily to
2057  * support git's cvsserver that wants to give a linear history
2058  * to its clients.
2059  */
2060 static int read_ancestry(const char *graft_file)
2061 {
2062         FILE *fp = fopen(graft_file, "r");
2063         struct strbuf buf = STRBUF_INIT;
2064         if (!fp)
2065                 return -1;
2066         while (!strbuf_getwholeline(&buf, fp, '\n')) {
2067                 /* The format is just "Commit Parent1 Parent2 ...\n" */
2068                 struct commit_graft *graft = read_graft_line(buf.buf, buf.len);
2069                 if (graft)
2070                         register_commit_graft(graft, 0);
2071         }
2072         fclose(fp);
2073         strbuf_release(&buf);
2074         return 0;
2075 }
2076
2077 static int update_auto_abbrev(int auto_abbrev, struct origin *suspect)
2078 {
2079         const char *uniq = find_unique_abbrev(suspect->commit->object.oid.hash,
2080                                               auto_abbrev);
2081         int len = strlen(uniq);
2082         if (auto_abbrev < len)
2083                 return len;
2084         return auto_abbrev;
2085 }
2086
2087 /*
2088  * How many columns do we need to show line numbers, authors,
2089  * and filenames?
2090  */
2091 static void find_alignment(struct scoreboard *sb, int *option)
2092 {
2093         int longest_src_lines = 0;
2094         int longest_dst_lines = 0;
2095         unsigned largest_score = 0;
2096         struct blame_entry *e;
2097         int compute_auto_abbrev = (abbrev < 0);
2098         int auto_abbrev = default_abbrev;
2099
2100         for (e = sb->ent; e; e = e->next) {
2101                 struct origin *suspect = e->suspect;
2102                 int num;
2103
2104                 if (compute_auto_abbrev)
2105                         auto_abbrev = update_auto_abbrev(auto_abbrev, suspect);
2106                 if (strcmp(suspect->path, sb->path))
2107                         *option |= OUTPUT_SHOW_NAME;
2108                 num = strlen(suspect->path);
2109                 if (longest_file < num)
2110                         longest_file = num;
2111                 if (!(suspect->commit->object.flags & METAINFO_SHOWN)) {
2112                         struct commit_info ci;
2113                         suspect->commit->object.flags |= METAINFO_SHOWN;
2114                         get_commit_info(suspect->commit, &ci, 1);
2115                         if (*option & OUTPUT_SHOW_EMAIL)
2116                                 num = utf8_strwidth(ci.author_mail.buf);
2117                         else
2118                                 num = utf8_strwidth(ci.author.buf);
2119                         if (longest_author < num)
2120                                 longest_author = num;
2121                         commit_info_destroy(&ci);
2122                 }
2123                 num = e->s_lno + e->num_lines;
2124                 if (longest_src_lines < num)
2125                         longest_src_lines = num;
2126                 num = e->lno + e->num_lines;
2127                 if (longest_dst_lines < num)
2128                         longest_dst_lines = num;
2129                 if (largest_score < ent_score(sb, e))
2130                         largest_score = ent_score(sb, e);
2131         }
2132         max_orig_digits = decimal_width(longest_src_lines);
2133         max_digits = decimal_width(longest_dst_lines);
2134         max_score_digits = decimal_width(largest_score);
2135
2136         if (compute_auto_abbrev)
2137                 /* one more abbrev length is needed for the boundary commit */
2138                 abbrev = auto_abbrev + 1;
2139 }
2140
2141 /*
2142  * For debugging -- origin is refcounted, and this asserts that
2143  * we do not underflow.
2144  */
2145 static void sanity_check_refcnt(struct scoreboard *sb)
2146 {
2147         int baa = 0;
2148         struct blame_entry *ent;
2149
2150         for (ent = sb->ent; ent; ent = ent->next) {
2151                 /* Nobody should have zero or negative refcnt */
2152                 if (ent->suspect->refcnt <= 0) {
2153                         fprintf(stderr, "%s in %s has negative refcnt %d\n",
2154                                 ent->suspect->path,
2155                                 oid_to_hex(&ent->suspect->commit->object.oid),
2156                                 ent->suspect->refcnt);
2157                         baa = 1;
2158                 }
2159         }
2160         if (baa) {
2161                 int opt = 0160;
2162                 find_alignment(sb, &opt);
2163                 output(sb, opt);
2164                 die("Baa %d!", baa);
2165         }
2166 }
2167
2168 static unsigned parse_score(const char *arg)
2169 {
2170         char *end;
2171         unsigned long score = strtoul(arg, &end, 10);
2172         if (*end)
2173                 return 0;
2174         return score;
2175 }
2176
2177 static const char *add_prefix(const char *prefix, const char *path)
2178 {
2179         return prefix_path(prefix, prefix ? strlen(prefix) : 0, path);
2180 }
2181
2182 static int git_blame_config(const char *var, const char *value, void *cb)
2183 {
2184         if (!strcmp(var, "blame.showroot")) {
2185                 show_root = git_config_bool(var, value);
2186                 return 0;
2187         }
2188         if (!strcmp(var, "blame.blankboundary")) {
2189                 blank_boundary = git_config_bool(var, value);
2190                 return 0;
2191         }
2192         if (!strcmp(var, "blame.showemail")) {
2193                 int *output_option = cb;
2194                 if (git_config_bool(var, value))
2195                         *output_option |= OUTPUT_SHOW_EMAIL;
2196                 else
2197                         *output_option &= ~OUTPUT_SHOW_EMAIL;
2198                 return 0;
2199         }
2200         if (!strcmp(var, "blame.date")) {
2201                 if (!value)
2202                         return config_error_nonbool(var);
2203                 parse_date_format(value, &blame_date_mode);
2204                 return 0;
2205         }
2206
2207         if (userdiff_config(var, value) < 0)
2208                 return -1;
2209
2210         return git_default_config(var, value, cb);
2211 }
2212
2213 static void verify_working_tree_path(struct commit *work_tree, const char *path)
2214 {
2215         struct commit_list *parents;
2216
2217         for (parents = work_tree->parents; parents; parents = parents->next) {
2218                 const unsigned char *commit_sha1 = parents->item->object.oid.hash;
2219                 unsigned char blob_sha1[20];
2220                 unsigned mode;
2221
2222                 if (!get_tree_entry(commit_sha1, path, blob_sha1, &mode) &&
2223                     sha1_object_info(blob_sha1, NULL) == OBJ_BLOB)
2224                         return;
2225         }
2226         die("no such path '%s' in HEAD", path);
2227 }
2228
2229 static struct commit_list **append_parent(struct commit_list **tail, const unsigned char *sha1)
2230 {
2231         struct commit *parent;
2232
2233         parent = lookup_commit_reference(sha1);
2234         if (!parent)
2235                 die("no such commit %s", sha1_to_hex(sha1));
2236         return &commit_list_insert(parent, tail)->next;
2237 }
2238
2239 static void append_merge_parents(struct commit_list **tail)
2240 {
2241         int merge_head;
2242         struct strbuf line = STRBUF_INIT;
2243
2244         merge_head = open(git_path_merge_head(), O_RDONLY);
2245         if (merge_head < 0) {
2246                 if (errno == ENOENT)
2247                         return;
2248                 die("cannot open '%s' for reading", git_path_merge_head());
2249         }
2250
2251         while (!strbuf_getwholeline_fd(&line, merge_head, '\n')) {
2252                 unsigned char sha1[20];
2253                 if (line.len < 40 || get_sha1_hex(line.buf, sha1))
2254                         die("unknown line in '%s': %s", git_path_merge_head(), line.buf);
2255                 tail = append_parent(tail, sha1);
2256         }
2257         close(merge_head);
2258         strbuf_release(&line);
2259 }
2260
2261 /*
2262  * This isn't as simple as passing sb->buf and sb->len, because we
2263  * want to transfer ownership of the buffer to the commit (so we
2264  * must use detach).
2265  */
2266 static void set_commit_buffer_from_strbuf(struct commit *c, struct strbuf *sb)
2267 {
2268         size_t len;
2269         void *buf = strbuf_detach(sb, &len);
2270         set_commit_buffer(c, buf, len);
2271 }
2272
2273 /*
2274  * Prepare a dummy commit that represents the work tree (or staged) item.
2275  * Note that annotating work tree item never works in the reverse.
2276  */
2277 static struct commit *fake_working_tree_commit(struct diff_options *opt,
2278                                                const char *path,
2279                                                const char *contents_from)
2280 {
2281         struct commit *commit;
2282         struct origin *origin;
2283         struct commit_list **parent_tail, *parent;
2284         unsigned char head_sha1[20];
2285         struct strbuf buf = STRBUF_INIT;
2286         const char *ident;
2287         time_t now;
2288         int size, len;
2289         struct cache_entry *ce;
2290         unsigned mode;
2291         struct strbuf msg = STRBUF_INIT;
2292
2293         time(&now);
2294         commit = alloc_commit_node();
2295         commit->object.parsed = 1;
2296         commit->date = now;
2297         parent_tail = &commit->parents;
2298
2299         if (!resolve_ref_unsafe("HEAD", RESOLVE_REF_READING, head_sha1, NULL))
2300                 die("no such ref: HEAD");
2301
2302         parent_tail = append_parent(parent_tail, head_sha1);
2303         append_merge_parents(parent_tail);
2304         verify_working_tree_path(commit, path);
2305
2306         origin = make_origin(commit, path);
2307
2308         ident = fmt_ident("Not Committed Yet", "not.committed.yet", NULL, 0);
2309         strbuf_addstr(&msg, "tree 0000000000000000000000000000000000000000\n");
2310         for (parent = commit->parents; parent; parent = parent->next)
2311                 strbuf_addf(&msg, "parent %s\n",
2312                             oid_to_hex(&parent->item->object.oid));
2313         strbuf_addf(&msg,
2314                     "author %s\n"
2315                     "committer %s\n\n"
2316                     "Version of %s from %s\n",
2317                     ident, ident, path,
2318                     (!contents_from ? path :
2319                      (!strcmp(contents_from, "-") ? "standard input" : contents_from)));
2320         set_commit_buffer_from_strbuf(commit, &msg);
2321
2322         if (!contents_from || strcmp("-", contents_from)) {
2323                 struct stat st;
2324                 const char *read_from;
2325                 char *buf_ptr;
2326                 unsigned long buf_len;
2327
2328                 if (contents_from) {
2329                         if (stat(contents_from, &st) < 0)
2330                                 die_errno("Cannot stat '%s'", contents_from);
2331                         read_from = contents_from;
2332                 }
2333                 else {
2334                         if (lstat(path, &st) < 0)
2335                                 die_errno("Cannot lstat '%s'", path);
2336                         read_from = path;
2337                 }
2338                 mode = canon_mode(st.st_mode);
2339
2340                 switch (st.st_mode & S_IFMT) {
2341                 case S_IFREG:
2342                         if (DIFF_OPT_TST(opt, ALLOW_TEXTCONV) &&
2343                             textconv_object(read_from, mode, null_sha1, 0, &buf_ptr, &buf_len))
2344                                 strbuf_attach(&buf, buf_ptr, buf_len, buf_len + 1);
2345                         else if (strbuf_read_file(&buf, read_from, st.st_size) != st.st_size)
2346                                 die_errno("cannot open or read '%s'", read_from);
2347                         break;
2348                 case S_IFLNK:
2349                         if (strbuf_readlink(&buf, read_from, st.st_size) < 0)
2350                                 die_errno("cannot readlink '%s'", read_from);
2351                         break;
2352                 default:
2353                         die("unsupported file type %s", read_from);
2354                 }
2355         }
2356         else {
2357                 /* Reading from stdin */
2358                 mode = 0;
2359                 if (strbuf_read(&buf, 0, 0) < 0)
2360                         die_errno("failed to read from stdin");
2361         }
2362         convert_to_git(path, buf.buf, buf.len, &buf, 0);
2363         origin->file.ptr = buf.buf;
2364         origin->file.size = buf.len;
2365         pretend_sha1_file(buf.buf, buf.len, OBJ_BLOB, origin->blob_sha1);
2366
2367         /*
2368          * Read the current index, replace the path entry with
2369          * origin->blob_sha1 without mucking with its mode or type
2370          * bits; we are not going to write this index out -- we just
2371          * want to run "diff-index --cached".
2372          */
2373         discard_cache();
2374         read_cache();
2375
2376         len = strlen(path);
2377         if (!mode) {
2378                 int pos = cache_name_pos(path, len);
2379                 if (0 <= pos)
2380                         mode = active_cache[pos]->ce_mode;
2381                 else
2382                         /* Let's not bother reading from HEAD tree */
2383                         mode = S_IFREG | 0644;
2384         }
2385         size = cache_entry_size(len);
2386         ce = xcalloc(1, size);
2387         hashcpy(ce->sha1, origin->blob_sha1);
2388         memcpy(ce->name, path, len);
2389         ce->ce_flags = create_ce_flags(0);
2390         ce->ce_namelen = len;
2391         ce->ce_mode = create_ce_mode(mode);
2392         add_cache_entry(ce, ADD_CACHE_OK_TO_ADD|ADD_CACHE_OK_TO_REPLACE);
2393
2394         cache_tree_invalidate_path(&the_index, path);
2395
2396         return commit;
2397 }
2398
2399 static struct commit *find_single_final(struct rev_info *revs,
2400                                         const char **name_p)
2401 {
2402         int i;
2403         struct commit *found = NULL;
2404         const char *name = NULL;
2405
2406         for (i = 0; i < revs->pending.nr; i++) {
2407                 struct object *obj = revs->pending.objects[i].item;
2408                 if (obj->flags & UNINTERESTING)
2409                         continue;
2410                 while (obj->type == OBJ_TAG)
2411                         obj = deref_tag(obj, NULL, 0);
2412                 if (obj->type != OBJ_COMMIT)
2413                         die("Non commit %s?", revs->pending.objects[i].name);
2414                 if (found)
2415                         die("More than one commit to dig from %s and %s?",
2416                             revs->pending.objects[i].name, name);
2417                 found = (struct commit *)obj;
2418                 name = revs->pending.objects[i].name;
2419         }
2420         if (name_p)
2421                 *name_p = name;
2422         return found;
2423 }
2424
2425 static char *prepare_final(struct scoreboard *sb)
2426 {
2427         const char *name;
2428         sb->final = find_single_final(sb->revs, &name);
2429         return xstrdup_or_null(name);
2430 }
2431
2432 static char *prepare_initial(struct scoreboard *sb)
2433 {
2434         int i;
2435         const char *final_commit_name = NULL;
2436         struct rev_info *revs = sb->revs;
2437
2438         /*
2439          * There must be one and only one negative commit, and it must be
2440          * the boundary.
2441          */
2442         for (i = 0; i < revs->pending.nr; i++) {
2443                 struct object *obj = revs->pending.objects[i].item;
2444                 if (!(obj->flags & UNINTERESTING))
2445                         continue;
2446                 while (obj->type == OBJ_TAG)
2447                         obj = deref_tag(obj, NULL, 0);
2448                 if (obj->type != OBJ_COMMIT)
2449                         die("Non commit %s?", revs->pending.objects[i].name);
2450                 if (sb->final)
2451                         die("More than one commit to dig down to %s and %s?",
2452                             revs->pending.objects[i].name,
2453                             final_commit_name);
2454                 sb->final = (struct commit *) obj;
2455                 final_commit_name = revs->pending.objects[i].name;
2456         }
2457         if (!final_commit_name)
2458                 die("No commit to dig down to?");
2459         return xstrdup(final_commit_name);
2460 }
2461
2462 static int blame_copy_callback(const struct option *option, const char *arg, int unset)
2463 {
2464         int *opt = option->value;
2465
2466         /*
2467          * -C enables copy from removed files;
2468          * -C -C enables copy from existing files, but only
2469          *       when blaming a new file;
2470          * -C -C -C enables copy from existing files for
2471          *          everybody
2472          */
2473         if (*opt & PICKAXE_BLAME_COPY_HARDER)
2474                 *opt |= PICKAXE_BLAME_COPY_HARDEST;
2475         if (*opt & PICKAXE_BLAME_COPY)
2476                 *opt |= PICKAXE_BLAME_COPY_HARDER;
2477         *opt |= PICKAXE_BLAME_COPY | PICKAXE_BLAME_MOVE;
2478
2479         if (arg)
2480                 blame_copy_score = parse_score(arg);
2481         return 0;
2482 }
2483
2484 static int blame_move_callback(const struct option *option, const char *arg, int unset)
2485 {
2486         int *opt = option->value;
2487
2488         *opt |= PICKAXE_BLAME_MOVE;
2489
2490         if (arg)
2491                 blame_move_score = parse_score(arg);
2492         return 0;
2493 }
2494
2495 int cmd_blame(int argc, const char **argv, const char *prefix)
2496 {
2497         struct rev_info revs;
2498         const char *path;
2499         struct scoreboard sb;
2500         struct origin *o;
2501         struct blame_entry *ent = NULL;
2502         long dashdash_pos, lno;
2503         char *final_commit_name = NULL;
2504         enum object_type type;
2505         struct commit *final_commit = NULL;
2506
2507         static struct string_list range_list;
2508         static int output_option = 0, opt = 0;
2509         static int show_stats = 0;
2510         static const char *revs_file = NULL;
2511         static const char *contents_from = NULL;
2512         static const struct option options[] = {
2513                 OPT_BOOL(0, "incremental", &incremental, N_("Show blame entries as we find them, incrementally")),
2514                 OPT_BOOL('b', NULL, &blank_boundary, N_("Show blank SHA-1 for boundary commits (Default: off)")),
2515                 OPT_BOOL(0, "root", &show_root, N_("Do not treat root commits as boundaries (Default: off)")),
2516                 OPT_BOOL(0, "show-stats", &show_stats, N_("Show work cost statistics")),
2517                 OPT_BIT(0, "score-debug", &output_option, N_("Show output score for blame entries"), OUTPUT_SHOW_SCORE),
2518                 OPT_BIT('f', "show-name", &output_option, N_("Show original filename (Default: auto)"), OUTPUT_SHOW_NAME),
2519                 OPT_BIT('n', "show-number", &output_option, N_("Show original linenumber (Default: off)"), OUTPUT_SHOW_NUMBER),
2520                 OPT_BIT('p', "porcelain", &output_option, N_("Show in a format designed for machine consumption"), OUTPUT_PORCELAIN),
2521                 OPT_BIT(0, "line-porcelain", &output_option, N_("Show porcelain format with per-line commit information"), OUTPUT_PORCELAIN|OUTPUT_LINE_PORCELAIN),
2522                 OPT_BIT('c', NULL, &output_option, N_("Use the same output mode as git-annotate (Default: off)"), OUTPUT_ANNOTATE_COMPAT),
2523                 OPT_BIT('t', NULL, &output_option, N_("Show raw timestamp (Default: off)"), OUTPUT_RAW_TIMESTAMP),
2524                 OPT_BIT('l', NULL, &output_option, N_("Show long commit SHA1 (Default: off)"), OUTPUT_LONG_OBJECT_NAME),
2525                 OPT_BIT('s', NULL, &output_option, N_("Suppress author name and timestamp (Default: off)"), OUTPUT_NO_AUTHOR),
2526                 OPT_BIT('e', "show-email", &output_option, N_("Show author email instead of name (Default: off)"), OUTPUT_SHOW_EMAIL),
2527                 OPT_BIT('w', NULL, &xdl_opts, N_("Ignore whitespace differences"), XDF_IGNORE_WHITESPACE),
2528                 OPT_BIT(0, "minimal", &xdl_opts, N_("Spend extra cycles to find better match"), XDF_NEED_MINIMAL),
2529                 OPT_STRING('S', NULL, &revs_file, N_("file"), N_("Use revisions from <file> instead of calling git-rev-list")),
2530                 OPT_STRING(0, "contents", &contents_from, N_("file"), N_("Use <file>'s contents as the final image")),
2531                 { OPTION_CALLBACK, 'C', NULL, &opt, N_("score"), N_("Find line copies within and across files"), PARSE_OPT_OPTARG, blame_copy_callback },
2532                 { OPTION_CALLBACK, 'M', NULL, &opt, N_("score"), N_("Find line movements within and across files"), PARSE_OPT_OPTARG, blame_move_callback },
2533                 OPT_STRING_LIST('L', NULL, &range_list, N_("n,m"), N_("Process only line range n,m, counting from 1")),
2534                 OPT__ABBREV(&abbrev),
2535                 OPT_END()
2536         };
2537
2538         struct parse_opt_ctx_t ctx;
2539         int cmd_is_annotate = !strcmp(argv[0], "annotate");
2540         struct range_set ranges;
2541         unsigned int range_i;
2542         long anchor;
2543
2544         git_config(git_blame_config, &output_option);
2545         init_revisions(&revs, NULL);
2546         revs.date_mode = blame_date_mode;
2547         DIFF_OPT_SET(&revs.diffopt, ALLOW_TEXTCONV);
2548         DIFF_OPT_SET(&revs.diffopt, FOLLOW_RENAMES);
2549
2550         save_commit_buffer = 0;
2551         dashdash_pos = 0;
2552
2553         parse_options_start(&ctx, argc, argv, prefix, options,
2554                             PARSE_OPT_KEEP_DASHDASH | PARSE_OPT_KEEP_ARGV0);
2555         for (;;) {
2556                 switch (parse_options_step(&ctx, options, blame_opt_usage)) {
2557                 case PARSE_OPT_HELP:
2558                         exit(129);
2559                 case PARSE_OPT_DONE:
2560                         if (ctx.argv[0])
2561                                 dashdash_pos = ctx.cpidx;
2562                         goto parse_done;
2563                 }
2564
2565                 if (!strcmp(ctx.argv[0], "--reverse")) {
2566                         ctx.argv[0] = "--children";
2567                         reverse = 1;
2568                 }
2569                 parse_revision_opt(&revs, &ctx, options, blame_opt_usage);
2570         }
2571 parse_done:
2572         no_whole_file_rename = !DIFF_OPT_TST(&revs.diffopt, FOLLOW_RENAMES);
2573         DIFF_OPT_CLR(&revs.diffopt, FOLLOW_RENAMES);
2574         argc = parse_options_end(&ctx);
2575
2576         if (0 < abbrev)
2577                 /* one more abbrev length is needed for the boundary commit */
2578                 abbrev++;
2579
2580         if (revs_file && read_ancestry(revs_file))
2581                 die_errno("reading graft file '%s' failed", revs_file);
2582
2583         if (cmd_is_annotate) {
2584                 output_option |= OUTPUT_ANNOTATE_COMPAT;
2585                 blame_date_mode.type = DATE_ISO8601;
2586         } else {
2587                 blame_date_mode = revs.date_mode;
2588         }
2589
2590         /* The maximum width used to show the dates */
2591         switch (blame_date_mode.type) {
2592         case DATE_RFC2822:
2593                 blame_date_width = sizeof("Thu, 19 Oct 2006 16:00:04 -0700");
2594                 break;
2595         case DATE_ISO8601_STRICT:
2596                 blame_date_width = sizeof("2006-10-19T16:00:04-07:00");
2597                 break;
2598         case DATE_ISO8601:
2599                 blame_date_width = sizeof("2006-10-19 16:00:04 -0700");
2600                 break;
2601         case DATE_RAW:
2602                 blame_date_width = sizeof("1161298804 -0700");
2603                 break;
2604         case DATE_SHORT:
2605                 blame_date_width = sizeof("2006-10-19");
2606                 break;
2607         case DATE_RELATIVE:
2608                 /* TRANSLATORS: This string is used to tell us the maximum
2609                    display width for a relative timestamp in "git blame"
2610                    output.  For C locale, "4 years, 11 months ago", which
2611                    takes 22 places, is the longest among various forms of
2612                    relative timestamps, but your language may need more or
2613                    fewer display columns. */
2614                 blame_date_width = utf8_strwidth(_("4 years, 11 months ago")) + 1; /* add the null */
2615                 break;
2616         case DATE_NORMAL:
2617                 blame_date_width = sizeof("Thu Oct 19 16:00:04 2006 -0700");
2618                 break;
2619         case DATE_STRFTIME:
2620                 blame_date_width = strlen(show_date(0, 0, &blame_date_mode)) + 1; /* add the null */
2621                 break;
2622         }
2623         blame_date_width -= 1; /* strip the null */
2624
2625         if (DIFF_OPT_TST(&revs.diffopt, FIND_COPIES_HARDER))
2626                 opt |= (PICKAXE_BLAME_COPY | PICKAXE_BLAME_MOVE |
2627                         PICKAXE_BLAME_COPY_HARDER);
2628
2629         if (!blame_move_score)
2630                 blame_move_score = BLAME_DEFAULT_MOVE_SCORE;
2631         if (!blame_copy_score)
2632                 blame_copy_score = BLAME_DEFAULT_COPY_SCORE;
2633
2634         /*
2635          * We have collected options unknown to us in argv[1..unk]
2636          * which are to be passed to revision machinery if we are
2637          * going to do the "bottom" processing.
2638          *
2639          * The remaining are:
2640          *
2641          * (1) if dashdash_pos != 0, it is either
2642          *     "blame [revisions] -- <path>" or
2643          *     "blame -- <path> <rev>"
2644          *
2645          * (2) otherwise, it is one of the two:
2646          *     "blame [revisions] <path>"
2647          *     "blame <path> <rev>"
2648          *
2649          * Note that we must strip out <path> from the arguments: we do not
2650          * want the path pruning but we may want "bottom" processing.
2651          */
2652         if (dashdash_pos) {
2653                 switch (argc - dashdash_pos - 1) {
2654                 case 2: /* (1b) */
2655                         if (argc != 4)
2656                                 usage_with_options(blame_opt_usage, options);
2657                         /* reorder for the new way: <rev> -- <path> */
2658                         argv[1] = argv[3];
2659                         argv[3] = argv[2];
2660                         argv[2] = "--";
2661                         /* FALLTHROUGH */
2662                 case 1: /* (1a) */
2663                         path = add_prefix(prefix, argv[--argc]);
2664                         argv[argc] = NULL;
2665                         break;
2666                 default:
2667                         usage_with_options(blame_opt_usage, options);
2668                 }
2669         } else {
2670                 if (argc < 2)
2671                         usage_with_options(blame_opt_usage, options);
2672                 path = add_prefix(prefix, argv[argc - 1]);
2673                 if (argc == 3 && !file_exists(path)) { /* (2b) */
2674                         path = add_prefix(prefix, argv[1]);
2675                         argv[1] = argv[2];
2676                 }
2677                 argv[argc - 1] = "--";
2678
2679                 setup_work_tree();
2680                 if (!file_exists(path))
2681                         die_errno("cannot stat path '%s'", path);
2682         }
2683
2684         revs.disable_stdin = 1;
2685         setup_revisions(argc, argv, &revs, NULL);
2686         memset(&sb, 0, sizeof(sb));
2687
2688         sb.revs = &revs;
2689         if (!reverse) {
2690                 final_commit_name = prepare_final(&sb);
2691                 sb.commits.compare = compare_commits_by_commit_date;
2692         }
2693         else if (contents_from)
2694                 die("--contents and --reverse do not blend well.");
2695         else {
2696                 final_commit_name = prepare_initial(&sb);
2697                 sb.commits.compare = compare_commits_by_reverse_commit_date;
2698                 if (revs.first_parent_only)
2699                         revs.children.name = NULL;
2700         }
2701
2702         if (!sb.final) {
2703                 /*
2704                  * "--not A B -- path" without anything positive;
2705                  * do not default to HEAD, but use the working tree
2706                  * or "--contents".
2707                  */
2708                 setup_work_tree();
2709                 sb.final = fake_working_tree_commit(&sb.revs->diffopt,
2710                                                     path, contents_from);
2711                 add_pending_object(&revs, &(sb.final->object), ":");
2712         }
2713         else if (contents_from)
2714                 die("Cannot use --contents with final commit object name");
2715
2716         if (reverse && revs.first_parent_only) {
2717                 final_commit = find_single_final(sb.revs, NULL);
2718                 if (!final_commit)
2719                         die("--reverse and --first-parent together require specified latest commit");
2720         }
2721
2722         /*
2723          * If we have bottom, this will mark the ancestors of the
2724          * bottom commits we would reach while traversing as
2725          * uninteresting.
2726          */
2727         if (prepare_revision_walk(&revs))
2728                 die(_("revision walk setup failed"));
2729
2730         if (reverse && revs.first_parent_only) {
2731                 struct commit *c = final_commit;
2732
2733                 sb.revs->children.name = "children";
2734                 while (c->parents &&
2735                        oidcmp(&c->object.oid, &sb.final->object.oid)) {
2736                         struct commit_list *l = xcalloc(1, sizeof(*l));
2737
2738                         l->item = c;
2739                         if (add_decoration(&sb.revs->children,
2740                                            &c->parents->item->object, l))
2741                                 die("BUG: not unique item in first-parent chain");
2742                         c = c->parents->item;
2743                 }
2744
2745                 if (oidcmp(&c->object.oid, &sb.final->object.oid))
2746                         die("--reverse --first-parent together require range along first-parent chain");
2747         }
2748
2749         if (is_null_oid(&sb.final->object.oid)) {
2750                 o = sb.final->util;
2751                 sb.final_buf = xmemdupz(o->file.ptr, o->file.size);
2752                 sb.final_buf_size = o->file.size;
2753         }
2754         else {
2755                 o = get_origin(&sb, sb.final, path);
2756                 if (fill_blob_sha1_and_mode(o))
2757                         die("no such path %s in %s", path, final_commit_name);
2758
2759                 if (DIFF_OPT_TST(&sb.revs->diffopt, ALLOW_TEXTCONV) &&
2760                     textconv_object(path, o->mode, o->blob_sha1, 1, (char **) &sb.final_buf,
2761                                     &sb.final_buf_size))
2762                         ;
2763                 else
2764                         sb.final_buf = read_sha1_file(o->blob_sha1, &type,
2765                                                       &sb.final_buf_size);
2766
2767                 if (!sb.final_buf)
2768                         die("Cannot read blob %s for path %s",
2769                             sha1_to_hex(o->blob_sha1),
2770                             path);
2771         }
2772         num_read_blob++;
2773         lno = prepare_lines(&sb);
2774
2775         if (lno && !range_list.nr)
2776                 string_list_append(&range_list, xstrdup("1"));
2777
2778         anchor = 1;
2779         range_set_init(&ranges, range_list.nr);
2780         for (range_i = 0; range_i < range_list.nr; ++range_i) {
2781                 long bottom, top;
2782                 if (parse_range_arg(range_list.items[range_i].string,
2783                                     nth_line_cb, &sb, lno, anchor,
2784                                     &bottom, &top, sb.path))
2785                         usage(blame_usage);
2786                 if (lno < top || ((lno || bottom) && lno < bottom))
2787                         die("file %s has only %lu lines", path, lno);
2788                 if (bottom < 1)
2789                         bottom = 1;
2790                 if (top < 1)
2791                         top = lno;
2792                 bottom--;
2793                 range_set_append_unsafe(&ranges, bottom, top);
2794                 anchor = top + 1;
2795         }
2796         sort_and_merge_range_set(&ranges);
2797
2798         for (range_i = ranges.nr; range_i > 0; --range_i) {
2799                 const struct range *r = &ranges.ranges[range_i - 1];
2800                 long bottom = r->start;
2801                 long top = r->end;
2802                 struct blame_entry *next = ent;
2803                 ent = xcalloc(1, sizeof(*ent));
2804                 ent->lno = bottom;
2805                 ent->num_lines = top - bottom;
2806                 ent->suspect = o;
2807                 ent->s_lno = bottom;
2808                 ent->next = next;
2809                 origin_incref(o);
2810         }
2811
2812         o->suspects = ent;
2813         prio_queue_put(&sb.commits, o->commit);
2814
2815         origin_decref(o);
2816
2817         range_set_release(&ranges);
2818         string_list_clear(&range_list, 0);
2819
2820         sb.ent = NULL;
2821         sb.path = path;
2822
2823         read_mailmap(&mailmap, NULL);
2824
2825         if (!incremental)
2826                 setup_pager();
2827
2828         assign_blame(&sb, opt);
2829
2830         free(final_commit_name);
2831
2832         if (incremental)
2833                 return 0;
2834
2835         sb.ent = blame_sort(sb.ent, compare_blame_final);
2836
2837         coalesce(&sb);
2838
2839         if (!(output_option & OUTPUT_PORCELAIN))
2840                 find_alignment(&sb, &output_option);
2841
2842         output(&sb, output_option);
2843         free((void *)sb.final_buf);
2844         for (ent = sb.ent; ent; ) {
2845                 struct blame_entry *e = ent->next;
2846                 free(ent);
2847                 ent = e;
2848         }
2849
2850         if (show_stats) {
2851                 printf("num read blob: %d\n", num_read_blob);
2852                 printf("num get patch: %d\n", num_get_patch);
2853                 printf("num commits: %d\n", num_commits);
2854         }
2855         return 0;
2856 }