blame, line-log: do not loop around deref_tag()
[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 "builtin.h"
10 #include "blob.h"
11 #include "commit.h"
12 #include "tag.h"
13 #include "tree-walk.h"
14 #include "diff.h"
15 #include "diffcore.h"
16 #include "revision.h"
17 #include "quote.h"
18 #include "xdiff-interface.h"
19 #include "cache-tree.h"
20 #include "string-list.h"
21 #include "mailmap.h"
22 #include "mergesort.h"
23 #include "parse-options.h"
24 #include "prio-queue.h"
25 #include "utf8.h"
26 #include "userdiff.h"
27 #include "line-range.h"
28 #include "line-log.h"
29 #include "dir.h"
30
31 static char blame_usage[] = N_("git blame [<options>] [<rev-opts>] [<rev>] [--] <file>");
32
33 static const char *blame_opt_usage[] = {
34         blame_usage,
35         "",
36         N_("<rev-opts> are documented in git-rev-list(1)"),
37         NULL
38 };
39
40 static int longest_file;
41 static int longest_author;
42 static int max_orig_digits;
43 static int max_digits;
44 static int max_score_digits;
45 static int show_root;
46 static int reverse;
47 static int blank_boundary;
48 static int incremental;
49 static int xdl_opts;
50 static int abbrev = -1;
51 static int no_whole_file_rename;
52
53 static enum date_mode blame_date_mode = DATE_ISO8601;
54 static size_t blame_date_width;
55
56 static struct string_list mailmap;
57
58 #ifndef DEBUG
59 #define DEBUG 0
60 #endif
61
62 /* stats */
63 static int num_read_blob;
64 static int num_get_patch;
65 static int num_commits;
66
67 #define PICKAXE_BLAME_MOVE              01
68 #define PICKAXE_BLAME_COPY              02
69 #define PICKAXE_BLAME_COPY_HARDER       04
70 #define PICKAXE_BLAME_COPY_HARDEST      010
71
72 /*
73  * blame for a blame_entry with score lower than these thresholds
74  * is not passed to the parent using move/copy logic.
75  */
76 static unsigned blame_move_score;
77 static unsigned blame_copy_score;
78 #define BLAME_DEFAULT_MOVE_SCORE        20
79 #define BLAME_DEFAULT_COPY_SCORE        40
80
81 /* Remember to update object flag allocation in object.h */
82 #define METAINFO_SHOWN          (1u<<12)
83 #define MORE_THAN_ONE_PATH      (1u<<13)
84
85 /*
86  * One blob in a commit that is being suspected
87  */
88 struct origin {
89         int refcnt;
90         /* Record preceding blame record for this blob */
91         struct origin *previous;
92         /* origins are put in a list linked via `next' hanging off the
93          * corresponding commit's util field in order to make finding
94          * them fast.  The presence in this chain does not count
95          * towards the origin's reference count.  It is tempting to
96          * let it count as long as the commit is pending examination,
97          * but even under circumstances where the commit will be
98          * present multiple times in the priority queue of unexamined
99          * commits, processing the first instance will not leave any
100          * work requiring the origin data for the second instance.  An
101          * interspersed commit changing that would have to be
102          * preexisting with a different ancestry and with the same
103          * commit date in order to wedge itself between two instances
104          * of the same commit in the priority queue _and_ produce
105          * blame entries relevant for it.  While we don't want to let
106          * us get tripped up by this case, it certainly does not seem
107          * worth optimizing for.
108          */
109         struct origin *next;
110         struct commit *commit;
111         /* `suspects' contains blame entries that may be attributed to
112          * this origin's commit or to parent commits.  When a commit
113          * is being processed, all suspects will be moved, either by
114          * assigning them to an origin in a different commit, or by
115          * shipping them to the scoreboard's ent list because they
116          * cannot be attributed to a different commit.
117          */
118         struct blame_entry *suspects;
119         mmfile_t file;
120         unsigned char blob_sha1[20];
121         unsigned mode;
122         /* guilty gets set when shipping any suspects to the final
123          * blame list instead of other commits
124          */
125         char guilty;
126         char path[FLEX_ARRAY];
127 };
128
129 static int diff_hunks(mmfile_t *file_a, mmfile_t *file_b, long ctxlen,
130                       xdl_emit_hunk_consume_func_t hunk_func, void *cb_data)
131 {
132         xpparam_t xpp = {0};
133         xdemitconf_t xecfg = {0};
134         xdemitcb_t ecb = {NULL};
135
136         xpp.flags = xdl_opts;
137         xecfg.ctxlen = ctxlen;
138         xecfg.hunk_func = hunk_func;
139         ecb.priv = cb_data;
140         return xdi_diff(file_a, file_b, &xpp, &xecfg, &ecb);
141 }
142
143 /*
144  * Prepare diff_filespec and convert it using diff textconv API
145  * if the textconv driver exists.
146  * Return 1 if the conversion succeeds, 0 otherwise.
147  */
148 int textconv_object(const char *path,
149                     unsigned mode,
150                     const unsigned char *sha1,
151                     int sha1_valid,
152                     char **buf,
153                     unsigned long *buf_size)
154 {
155         struct diff_filespec *df;
156         struct userdiff_driver *textconv;
157
158         df = alloc_filespec(path);
159         fill_filespec(df, sha1, sha1_valid, mode);
160         textconv = get_textconv(df);
161         if (!textconv) {
162                 free_filespec(df);
163                 return 0;
164         }
165
166         *buf_size = fill_textconv(textconv, df, buf);
167         free_filespec(df);
168         return 1;
169 }
170
171 /*
172  * Given an origin, prepare mmfile_t structure to be used by the
173  * diff machinery
174  */
175 static void fill_origin_blob(struct diff_options *opt,
176                              struct origin *o, mmfile_t *file)
177 {
178         if (!o->file.ptr) {
179                 enum object_type type;
180                 unsigned long file_size;
181
182                 num_read_blob++;
183                 if (DIFF_OPT_TST(opt, ALLOW_TEXTCONV) &&
184                     textconv_object(o->path, o->mode, o->blob_sha1, 1, &file->ptr, &file_size))
185                         ;
186                 else
187                         file->ptr = read_sha1_file(o->blob_sha1, &type, &file_size);
188                 file->size = file_size;
189
190                 if (!file->ptr)
191                         die("Cannot read blob %s for path %s",
192                             sha1_to_hex(o->blob_sha1),
193                             o->path);
194                 o->file = *file;
195         }
196         else
197                 *file = o->file;
198 }
199
200 /*
201  * Origin is refcounted and usually we keep the blob contents to be
202  * reused.
203  */
204 static inline struct origin *origin_incref(struct origin *o)
205 {
206         if (o)
207                 o->refcnt++;
208         return o;
209 }
210
211 static void origin_decref(struct origin *o)
212 {
213         if (o && --o->refcnt <= 0) {
214                 struct origin *p, *l = NULL;
215                 if (o->previous)
216                         origin_decref(o->previous);
217                 free(o->file.ptr);
218                 /* Should be present exactly once in commit chain */
219                 for (p = o->commit->util; p; l = p, p = p->next) {
220                         if (p == o) {
221                                 if (l)
222                                         l->next = p->next;
223                                 else
224                                         o->commit->util = p->next;
225                                 free(o);
226                                 return;
227                         }
228                 }
229                 die("internal error in blame::origin_decref");
230         }
231 }
232
233 static void drop_origin_blob(struct origin *o)
234 {
235         if (o->file.ptr) {
236                 free(o->file.ptr);
237                 o->file.ptr = NULL;
238         }
239 }
240
241 /*
242  * Each group of lines is described by a blame_entry; it can be split
243  * as we pass blame to the parents.  They are arranged in linked lists
244  * kept as `suspects' of some unprocessed origin, or entered (when the
245  * blame origin has been finalized) into the scoreboard structure.
246  * While the scoreboard structure is only sorted at the end of
247  * processing (according to final image line number), the lists
248  * attached to an origin are sorted by the target line number.
249  */
250 struct blame_entry {
251         struct blame_entry *next;
252
253         /* the first line of this group in the final image;
254          * internally all line numbers are 0 based.
255          */
256         int lno;
257
258         /* how many lines this group has */
259         int num_lines;
260
261         /* the commit that introduced this group into the final image */
262         struct origin *suspect;
263
264         /* the line number of the first line of this group in the
265          * suspect's file; internally all line numbers are 0 based.
266          */
267         int s_lno;
268
269         /* how significant this entry is -- cached to avoid
270          * scanning the lines over and over.
271          */
272         unsigned score;
273 };
274
275 /*
276  * Any merge of blames happens on lists of blames that arrived via
277  * different parents in a single suspect.  In this case, we want to
278  * sort according to the suspect line numbers as opposed to the final
279  * image line numbers.  The function body is somewhat longish because
280  * it avoids unnecessary writes.
281  */
282
283 static struct blame_entry *blame_merge(struct blame_entry *list1,
284                                        struct blame_entry *list2)
285 {
286         struct blame_entry *p1 = list1, *p2 = list2,
287                 **tail = &list1;
288
289         if (!p1)
290                 return p2;
291         if (!p2)
292                 return p1;
293
294         if (p1->s_lno <= p2->s_lno) {
295                 do {
296                         tail = &p1->next;
297                         if ((p1 = *tail) == NULL) {
298                                 *tail = p2;
299                                 return list1;
300                         }
301                 } while (p1->s_lno <= p2->s_lno);
302         }
303         for (;;) {
304                 *tail = p2;
305                 do {
306                         tail = &p2->next;
307                         if ((p2 = *tail) == NULL)  {
308                                 *tail = p1;
309                                 return list1;
310                         }
311                 } while (p1->s_lno > p2->s_lno);
312                 *tail = p1;
313                 do {
314                         tail = &p1->next;
315                         if ((p1 = *tail) == NULL) {
316                                 *tail = p2;
317                                 return list1;
318                         }
319                 } while (p1->s_lno <= p2->s_lno);
320         }
321 }
322
323 static void *get_next_blame(const void *p)
324 {
325         return ((struct blame_entry *)p)->next;
326 }
327
328 static void set_next_blame(void *p1, void *p2)
329 {
330         ((struct blame_entry *)p1)->next = p2;
331 }
332
333 /*
334  * Final image line numbers are all different, so we don't need a
335  * three-way comparison here.
336  */
337
338 static int compare_blame_final(const void *p1, const void *p2)
339 {
340         return ((struct blame_entry *)p1)->lno > ((struct blame_entry *)p2)->lno
341                 ? 1 : -1;
342 }
343
344 static int compare_blame_suspect(const void *p1, const void *p2)
345 {
346         const struct blame_entry *s1 = p1, *s2 = p2;
347         /*
348          * to allow for collating suspects, we sort according to the
349          * respective pointer value as the primary sorting criterion.
350          * The actual relation is pretty unimportant as long as it
351          * establishes a total order.  Comparing as integers gives us
352          * that.
353          */
354         if (s1->suspect != s2->suspect)
355                 return (intptr_t)s1->suspect > (intptr_t)s2->suspect ? 1 : -1;
356         if (s1->s_lno == s2->s_lno)
357                 return 0;
358         return s1->s_lno > s2->s_lno ? 1 : -1;
359 }
360
361 static struct blame_entry *blame_sort(struct blame_entry *head,
362                                       int (*compare_fn)(const void *, const void *))
363 {
364         return llist_mergesort (head, get_next_blame, set_next_blame, compare_fn);
365 }
366
367 static int compare_commits_by_reverse_commit_date(const void *a,
368                                                   const void *b,
369                                                   void *c)
370 {
371         return -compare_commits_by_commit_date(a, b, c);
372 }
373
374 /*
375  * The current state of the blame assignment.
376  */
377 struct scoreboard {
378         /* the final commit (i.e. where we started digging from) */
379         struct commit *final;
380         /* Priority queue for commits with unassigned blame records */
381         struct prio_queue commits;
382         struct rev_info *revs;
383         const char *path;
384
385         /*
386          * The contents in the final image.
387          * Used by many functions to obtain contents of the nth line,
388          * indexed with scoreboard.lineno[blame_entry.lno].
389          */
390         const char *final_buf;
391         unsigned long final_buf_size;
392
393         /* linked list of blames */
394         struct blame_entry *ent;
395
396         /* look-up a line in the final buffer */
397         int num_lines;
398         int *lineno;
399 };
400
401 static void sanity_check_refcnt(struct scoreboard *);
402
403 /*
404  * If two blame entries that are next to each other came from
405  * contiguous lines in the same origin (i.e. <commit, path> pair),
406  * merge them together.
407  */
408 static void coalesce(struct scoreboard *sb)
409 {
410         struct blame_entry *ent, *next;
411
412         for (ent = sb->ent; ent && (next = ent->next); ent = next) {
413                 if (ent->suspect == next->suspect &&
414                     ent->s_lno + ent->num_lines == next->s_lno) {
415                         ent->num_lines += next->num_lines;
416                         ent->next = next->next;
417                         origin_decref(next->suspect);
418                         free(next);
419                         ent->score = 0;
420                         next = ent; /* again */
421                 }
422         }
423
424         if (DEBUG) /* sanity */
425                 sanity_check_refcnt(sb);
426 }
427
428 /*
429  * Merge the given sorted list of blames into a preexisting origin.
430  * If there were no previous blames to that commit, it is entered into
431  * the commit priority queue of the score board.
432  */
433
434 static void queue_blames(struct scoreboard *sb, struct origin *porigin,
435                          struct blame_entry *sorted)
436 {
437         if (porigin->suspects)
438                 porigin->suspects = blame_merge(porigin->suspects, sorted);
439         else {
440                 struct origin *o;
441                 for (o = porigin->commit->util; o; o = o->next) {
442                         if (o->suspects) {
443                                 porigin->suspects = sorted;
444                                 return;
445                         }
446                 }
447                 porigin->suspects = sorted;
448                 prio_queue_put(&sb->commits, porigin->commit);
449         }
450 }
451
452 /*
453  * Given a commit and a path in it, create a new origin structure.
454  * The callers that add blame to the scoreboard should use
455  * get_origin() to obtain shared, refcounted copy instead of calling
456  * this function directly.
457  */
458 static struct origin *make_origin(struct commit *commit, const char *path)
459 {
460         struct origin *o;
461         o = xcalloc(1, sizeof(*o) + strlen(path) + 1);
462         o->commit = commit;
463         o->refcnt = 1;
464         o->next = commit->util;
465         commit->util = o;
466         strcpy(o->path, path);
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.sha1,
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_sha1(origin->commit->object.sha1))
559                 do_diff_cache(parent->tree->object.sha1, &diff_opts);
560         else
561                 diff_tree_sha1(parent->tree->object.sha1,
562                                origin->commit->tree->object.sha1,
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_sha1(origin->commit->object.sha1))
629                 do_diff_cache(parent->tree->object.sha1, &diff_opts);
630         else
631                 diff_tree_sha1(parent->tree->object.sha1,
632                                origin->commit->tree->object.sha1,
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                     sha1_to_hex(parent->commit->object.sha1),
979                     sha1_to_hex(target->commit->object.sha1));
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                     sha1_to_hex(parent->commit->object.sha1));
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_sha1(target->commit->object.sha1))
1277                 do_diff_cache(parent->tree->object.sha1, &diff_opts);
1278         else
1279                 diff_tree_sha1(parent->tree->object.sha1,
1280                                target->commit->tree->object.sha1,
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                 return commit->parents;
1375         return lookup_decoration(&revs->children, &commit->object);
1376 }
1377
1378 static int num_scapegoats(struct rev_info *revs, struct commit *commit)
1379 {
1380         struct commit_list *l = first_scapegoat(revs, commit);
1381         return commit_list_count(l);
1382 }
1383
1384 /* Distribute collected unsorted blames to the respected sorted lists
1385  * in the various origins.
1386  */
1387 static void distribute_blame(struct scoreboard *sb, struct blame_entry *blamed)
1388 {
1389         blamed = blame_sort(blamed, compare_blame_suspect);
1390         while (blamed)
1391         {
1392                 struct origin *porigin = blamed->suspect;
1393                 struct blame_entry *suspects = NULL;
1394                 do {
1395                         struct blame_entry *next = blamed->next;
1396                         blamed->next = suspects;
1397                         suspects = blamed;
1398                         blamed = next;
1399                 } while (blamed && blamed->suspect == porigin);
1400                 suspects = reverse_blame(suspects, NULL);
1401                 queue_blames(sb, porigin, suspects);
1402         }
1403 }
1404
1405 #define MAXSG 16
1406
1407 static void pass_blame(struct scoreboard *sb, struct origin *origin, int opt)
1408 {
1409         struct rev_info *revs = sb->revs;
1410         int i, pass, num_sg;
1411         struct commit *commit = origin->commit;
1412         struct commit_list *sg;
1413         struct origin *sg_buf[MAXSG];
1414         struct origin *porigin, **sg_origin = sg_buf;
1415         struct blame_entry *toosmall = NULL;
1416         struct blame_entry *blames, **blametail = &blames;
1417
1418         num_sg = num_scapegoats(revs, commit);
1419         if (!num_sg)
1420                 goto finish;
1421         else if (num_sg < ARRAY_SIZE(sg_buf))
1422                 memset(sg_buf, 0, sizeof(sg_buf));
1423         else
1424                 sg_origin = xcalloc(num_sg, sizeof(*sg_origin));
1425
1426         /*
1427          * The first pass looks for unrenamed path to optimize for
1428          * common cases, then we look for renames in the second pass.
1429          */
1430         for (pass = 0; pass < 2 - no_whole_file_rename; pass++) {
1431                 struct origin *(*find)(struct scoreboard *,
1432                                        struct commit *, struct origin *);
1433                 find = pass ? find_rename : find_origin;
1434
1435                 for (i = 0, sg = first_scapegoat(revs, commit);
1436                      i < num_sg && sg;
1437                      sg = sg->next, i++) {
1438                         struct commit *p = sg->item;
1439                         int j, same;
1440
1441                         if (sg_origin[i])
1442                                 continue;
1443                         if (parse_commit(p))
1444                                 continue;
1445                         porigin = find(sb, p, origin);
1446                         if (!porigin)
1447                                 continue;
1448                         if (!hashcmp(porigin->blob_sha1, origin->blob_sha1)) {
1449                                 pass_whole_blame(sb, origin, porigin);
1450                                 origin_decref(porigin);
1451                                 goto finish;
1452                         }
1453                         for (j = same = 0; j < i; j++)
1454                                 if (sg_origin[j] &&
1455                                     !hashcmp(sg_origin[j]->blob_sha1,
1456                                              porigin->blob_sha1)) {
1457                                         same = 1;
1458                                         break;
1459                                 }
1460                         if (!same)
1461                                 sg_origin[i] = porigin;
1462                         else
1463                                 origin_decref(porigin);
1464                 }
1465         }
1466
1467         num_commits++;
1468         for (i = 0, sg = first_scapegoat(revs, commit);
1469              i < num_sg && sg;
1470              sg = sg->next, i++) {
1471                 struct origin *porigin = sg_origin[i];
1472                 if (!porigin)
1473                         continue;
1474                 if (!origin->previous) {
1475                         origin_incref(porigin);
1476                         origin->previous = porigin;
1477                 }
1478                 pass_blame_to_parent(sb, origin, porigin);
1479                 if (!origin->suspects)
1480                         goto finish;
1481         }
1482
1483         /*
1484          * Optionally find moves in parents' files.
1485          */
1486         if (opt & PICKAXE_BLAME_MOVE) {
1487                 filter_small(sb, &toosmall, &origin->suspects, blame_move_score);
1488                 if (origin->suspects) {
1489                         for (i = 0, sg = first_scapegoat(revs, commit);
1490                              i < num_sg && sg;
1491                              sg = sg->next, i++) {
1492                                 struct origin *porigin = sg_origin[i];
1493                                 if (!porigin)
1494                                         continue;
1495                                 find_move_in_parent(sb, &blametail, &toosmall, origin, porigin);
1496                                 if (!origin->suspects)
1497                                         break;
1498                         }
1499                 }
1500         }
1501
1502         /*
1503          * Optionally find copies from parents' files.
1504          */
1505         if (opt & PICKAXE_BLAME_COPY) {
1506                 if (blame_copy_score > blame_move_score)
1507                         filter_small(sb, &toosmall, &origin->suspects, blame_copy_score);
1508                 else if (blame_copy_score < blame_move_score) {
1509                         origin->suspects = blame_merge(origin->suspects, toosmall);
1510                         toosmall = NULL;
1511                         filter_small(sb, &toosmall, &origin->suspects, blame_copy_score);
1512                 }
1513                 if (!origin->suspects)
1514                         goto finish;
1515
1516                 for (i = 0, sg = first_scapegoat(revs, commit);
1517                      i < num_sg && sg;
1518                      sg = sg->next, i++) {
1519                         struct origin *porigin = sg_origin[i];
1520                         find_copy_in_parent(sb, &blametail, &toosmall,
1521                                             origin, sg->item, porigin, opt);
1522                         if (!origin->suspects)
1523                                 goto finish;
1524                 }
1525         }
1526
1527 finish:
1528         *blametail = NULL;
1529         distribute_blame(sb, blames);
1530         /*
1531          * prepend toosmall to origin->suspects
1532          *
1533          * There is no point in sorting: this ends up on a big
1534          * unsorted list in the caller anyway.
1535          */
1536         if (toosmall) {
1537                 struct blame_entry **tail = &toosmall;
1538                 while (*tail)
1539                         tail = &(*tail)->next;
1540                 *tail = origin->suspects;
1541                 origin->suspects = toosmall;
1542         }
1543         for (i = 0; i < num_sg; i++) {
1544                 if (sg_origin[i]) {
1545                         drop_origin_blob(sg_origin[i]);
1546                         origin_decref(sg_origin[i]);
1547                 }
1548         }
1549         drop_origin_blob(origin);
1550         if (sg_buf != sg_origin)
1551                 free(sg_origin);
1552 }
1553
1554 /*
1555  * Information on commits, used for output.
1556  */
1557 struct commit_info {
1558         struct strbuf author;
1559         struct strbuf author_mail;
1560         unsigned long author_time;
1561         struct strbuf author_tz;
1562
1563         /* filled only when asked for details */
1564         struct strbuf committer;
1565         struct strbuf committer_mail;
1566         unsigned long committer_time;
1567         struct strbuf committer_tz;
1568
1569         struct strbuf summary;
1570 };
1571
1572 /*
1573  * Parse author/committer line in the commit object buffer
1574  */
1575 static void get_ac_line(const char *inbuf, const char *what,
1576         struct strbuf *name, struct strbuf *mail,
1577         unsigned long *time, struct strbuf *tz)
1578 {
1579         struct ident_split ident;
1580         size_t len, maillen, namelen;
1581         char *tmp, *endp;
1582         const char *namebuf, *mailbuf;
1583
1584         tmp = strstr(inbuf, what);
1585         if (!tmp)
1586                 goto error_out;
1587         tmp += strlen(what);
1588         endp = strchr(tmp, '\n');
1589         if (!endp)
1590                 len = strlen(tmp);
1591         else
1592                 len = endp - tmp;
1593
1594         if (split_ident_line(&ident, tmp, len)) {
1595         error_out:
1596                 /* Ugh */
1597                 tmp = "(unknown)";
1598                 strbuf_addstr(name, tmp);
1599                 strbuf_addstr(mail, tmp);
1600                 strbuf_addstr(tz, tmp);
1601                 *time = 0;
1602                 return;
1603         }
1604
1605         namelen = ident.name_end - ident.name_begin;
1606         namebuf = ident.name_begin;
1607
1608         maillen = ident.mail_end - ident.mail_begin;
1609         mailbuf = ident.mail_begin;
1610
1611         if (ident.date_begin && ident.date_end)
1612                 *time = strtoul(ident.date_begin, NULL, 10);
1613         else
1614                 *time = 0;
1615
1616         if (ident.tz_begin && ident.tz_end)
1617                 strbuf_add(tz, ident.tz_begin, ident.tz_end - ident.tz_begin);
1618         else
1619                 strbuf_addstr(tz, "(unknown)");
1620
1621         /*
1622          * Now, convert both name and e-mail using mailmap
1623          */
1624         map_user(&mailmap, &mailbuf, &maillen,
1625                  &namebuf, &namelen);
1626
1627         strbuf_addf(mail, "<%.*s>", (int)maillen, mailbuf);
1628         strbuf_add(name, namebuf, namelen);
1629 }
1630
1631 static void commit_info_init(struct commit_info *ci)
1632 {
1633
1634         strbuf_init(&ci->author, 0);
1635         strbuf_init(&ci->author_mail, 0);
1636         strbuf_init(&ci->author_tz, 0);
1637         strbuf_init(&ci->committer, 0);
1638         strbuf_init(&ci->committer_mail, 0);
1639         strbuf_init(&ci->committer_tz, 0);
1640         strbuf_init(&ci->summary, 0);
1641 }
1642
1643 static void commit_info_destroy(struct commit_info *ci)
1644 {
1645
1646         strbuf_release(&ci->author);
1647         strbuf_release(&ci->author_mail);
1648         strbuf_release(&ci->author_tz);
1649         strbuf_release(&ci->committer);
1650         strbuf_release(&ci->committer_mail);
1651         strbuf_release(&ci->committer_tz);
1652         strbuf_release(&ci->summary);
1653 }
1654
1655 static void get_commit_info(struct commit *commit,
1656                             struct commit_info *ret,
1657                             int detailed)
1658 {
1659         int len;
1660         const char *subject, *encoding;
1661         const char *message;
1662
1663         commit_info_init(ret);
1664
1665         encoding = get_log_output_encoding();
1666         message = logmsg_reencode(commit, NULL, encoding);
1667         get_ac_line(message, "\nauthor ",
1668                     &ret->author, &ret->author_mail,
1669                     &ret->author_time, &ret->author_tz);
1670
1671         if (!detailed) {
1672                 unuse_commit_buffer(commit, message);
1673                 return;
1674         }
1675
1676         get_ac_line(message, "\ncommitter ",
1677                     &ret->committer, &ret->committer_mail,
1678                     &ret->committer_time, &ret->committer_tz);
1679
1680         len = find_commit_subject(message, &subject);
1681         if (len)
1682                 strbuf_add(&ret->summary, subject, len);
1683         else
1684                 strbuf_addf(&ret->summary, "(%s)", sha1_to_hex(commit->object.sha1));
1685
1686         unuse_commit_buffer(commit, message);
1687 }
1688
1689 /*
1690  * To allow LF and other nonportable characters in pathnames,
1691  * they are c-style quoted as needed.
1692  */
1693 static void write_filename_info(const char *path)
1694 {
1695         printf("filename ");
1696         write_name_quoted(path, stdout, '\n');
1697 }
1698
1699 /*
1700  * Porcelain/Incremental format wants to show a lot of details per
1701  * commit.  Instead of repeating this every line, emit it only once,
1702  * the first time each commit appears in the output (unless the
1703  * user has specifically asked for us to repeat).
1704  */
1705 static int emit_one_suspect_detail(struct origin *suspect, int repeat)
1706 {
1707         struct commit_info ci;
1708
1709         if (!repeat && (suspect->commit->object.flags & METAINFO_SHOWN))
1710                 return 0;
1711
1712         suspect->commit->object.flags |= METAINFO_SHOWN;
1713         get_commit_info(suspect->commit, &ci, 1);
1714         printf("author %s\n", ci.author.buf);
1715         printf("author-mail %s\n", ci.author_mail.buf);
1716         printf("author-time %lu\n", ci.author_time);
1717         printf("author-tz %s\n", ci.author_tz.buf);
1718         printf("committer %s\n", ci.committer.buf);
1719         printf("committer-mail %s\n", ci.committer_mail.buf);
1720         printf("committer-time %lu\n", ci.committer_time);
1721         printf("committer-tz %s\n", ci.committer_tz.buf);
1722         printf("summary %s\n", ci.summary.buf);
1723         if (suspect->commit->object.flags & UNINTERESTING)
1724                 printf("boundary\n");
1725         if (suspect->previous) {
1726                 struct origin *prev = suspect->previous;
1727                 printf("previous %s ", sha1_to_hex(prev->commit->object.sha1));
1728                 write_name_quoted(prev->path, stdout, '\n');
1729         }
1730
1731         commit_info_destroy(&ci);
1732
1733         return 1;
1734 }
1735
1736 /*
1737  * The blame_entry is found to be guilty for the range.
1738  * Show it in incremental output.
1739  */
1740 static void found_guilty_entry(struct blame_entry *ent)
1741 {
1742         if (incremental) {
1743                 struct origin *suspect = ent->suspect;
1744
1745                 printf("%s %d %d %d\n",
1746                        sha1_to_hex(suspect->commit->object.sha1),
1747                        ent->s_lno + 1, ent->lno + 1, ent->num_lines);
1748                 emit_one_suspect_detail(suspect, 0);
1749                 write_filename_info(suspect->path);
1750                 maybe_flush_or_die(stdout, "stdout");
1751         }
1752 }
1753
1754 /*
1755  * The main loop -- while we have blobs with lines whose true origin
1756  * is still unknown, pick one blob, and allow its lines to pass blames
1757  * to its parents. */
1758 static void assign_blame(struct scoreboard *sb, int opt)
1759 {
1760         struct rev_info *revs = sb->revs;
1761         struct commit *commit = prio_queue_get(&sb->commits);
1762
1763         while (commit) {
1764                 struct blame_entry *ent;
1765                 struct origin *suspect = commit->util;
1766
1767                 /* find one suspect to break down */
1768                 while (suspect && !suspect->suspects)
1769                         suspect = suspect->next;
1770
1771                 if (!suspect) {
1772                         commit = prio_queue_get(&sb->commits);
1773                         continue;
1774                 }
1775
1776                 assert(commit == suspect->commit);
1777
1778                 /*
1779                  * We will use this suspect later in the loop,
1780                  * so hold onto it in the meantime.
1781                  */
1782                 origin_incref(suspect);
1783                 parse_commit(commit);
1784                 if (reverse ||
1785                     (!(commit->object.flags & UNINTERESTING) &&
1786                      !(revs->max_age != -1 && commit->date < revs->max_age)))
1787                         pass_blame(sb, suspect, opt);
1788                 else {
1789                         commit->object.flags |= UNINTERESTING;
1790                         if (commit->object.parsed)
1791                                 mark_parents_uninteresting(commit);
1792                 }
1793                 /* treat root commit as boundary */
1794                 if (!commit->parents && !show_root)
1795                         commit->object.flags |= UNINTERESTING;
1796
1797                 /* Take responsibility for the remaining entries */
1798                 ent = suspect->suspects;
1799                 if (ent) {
1800                         suspect->guilty = 1;
1801                         for (;;) {
1802                                 struct blame_entry *next = ent->next;
1803                                 found_guilty_entry(ent);
1804                                 if (next) {
1805                                         ent = next;
1806                                         continue;
1807                                 }
1808                                 ent->next = sb->ent;
1809                                 sb->ent = suspect->suspects;
1810                                 suspect->suspects = NULL;
1811                                 break;
1812                         }
1813                 }
1814                 origin_decref(suspect);
1815
1816                 if (DEBUG) /* sanity */
1817                         sanity_check_refcnt(sb);
1818         }
1819 }
1820
1821 static const char *format_time(unsigned long time, const char *tz_str,
1822                                int show_raw_time)
1823 {
1824         static struct strbuf time_buf = STRBUF_INIT;
1825
1826         strbuf_reset(&time_buf);
1827         if (show_raw_time) {
1828                 strbuf_addf(&time_buf, "%lu %s", time, tz_str);
1829         }
1830         else {
1831                 const char *time_str;
1832                 size_t time_width;
1833                 int tz;
1834                 tz = atoi(tz_str);
1835                 time_str = show_date(time, tz, blame_date_mode);
1836                 strbuf_addstr(&time_buf, time_str);
1837                 /*
1838                  * Add space paddings to time_buf to display a fixed width
1839                  * string, and use time_width for display width calibration.
1840                  */
1841                 for (time_width = utf8_strwidth(time_str);
1842                      time_width < blame_date_width;
1843                      time_width++)
1844                         strbuf_addch(&time_buf, ' ');
1845         }
1846         return time_buf.buf;
1847 }
1848
1849 #define OUTPUT_ANNOTATE_COMPAT  001
1850 #define OUTPUT_LONG_OBJECT_NAME 002
1851 #define OUTPUT_RAW_TIMESTAMP    004
1852 #define OUTPUT_PORCELAIN        010
1853 #define OUTPUT_SHOW_NAME        020
1854 #define OUTPUT_SHOW_NUMBER      040
1855 #define OUTPUT_SHOW_SCORE      0100
1856 #define OUTPUT_NO_AUTHOR       0200
1857 #define OUTPUT_SHOW_EMAIL       0400
1858 #define OUTPUT_LINE_PORCELAIN 01000
1859
1860 static void emit_porcelain_details(struct origin *suspect, int repeat)
1861 {
1862         if (emit_one_suspect_detail(suspect, repeat) ||
1863             (suspect->commit->object.flags & MORE_THAN_ONE_PATH))
1864                 write_filename_info(suspect->path);
1865 }
1866
1867 static void emit_porcelain(struct scoreboard *sb, struct blame_entry *ent,
1868                            int opt)
1869 {
1870         int repeat = opt & OUTPUT_LINE_PORCELAIN;
1871         int cnt;
1872         const char *cp;
1873         struct origin *suspect = ent->suspect;
1874         char hex[41];
1875
1876         strcpy(hex, sha1_to_hex(suspect->commit->object.sha1));
1877         printf("%s %d %d %d\n",
1878                hex,
1879                ent->s_lno + 1,
1880                ent->lno + 1,
1881                ent->num_lines);
1882         emit_porcelain_details(suspect, repeat);
1883
1884         cp = nth_line(sb, ent->lno);
1885         for (cnt = 0; cnt < ent->num_lines; cnt++) {
1886                 char ch;
1887                 if (cnt) {
1888                         printf("%s %d %d\n", hex,
1889                                ent->s_lno + 1 + cnt,
1890                                ent->lno + 1 + cnt);
1891                         if (repeat)
1892                                 emit_porcelain_details(suspect, 1);
1893                 }
1894                 putchar('\t');
1895                 do {
1896                         ch = *cp++;
1897                         putchar(ch);
1898                 } while (ch != '\n' &&
1899                          cp < sb->final_buf + sb->final_buf_size);
1900         }
1901
1902         if (sb->final_buf_size && cp[-1] != '\n')
1903                 putchar('\n');
1904 }
1905
1906 static void emit_other(struct scoreboard *sb, struct blame_entry *ent, int opt)
1907 {
1908         int cnt;
1909         const char *cp;
1910         struct origin *suspect = ent->suspect;
1911         struct commit_info ci;
1912         char hex[41];
1913         int show_raw_time = !!(opt & OUTPUT_RAW_TIMESTAMP);
1914
1915         get_commit_info(suspect->commit, &ci, 1);
1916         strcpy(hex, sha1_to_hex(suspect->commit->object.sha1));
1917
1918         cp = nth_line(sb, ent->lno);
1919         for (cnt = 0; cnt < ent->num_lines; cnt++) {
1920                 char ch;
1921                 int length = (opt & OUTPUT_LONG_OBJECT_NAME) ? 40 : abbrev;
1922
1923                 if (suspect->commit->object.flags & UNINTERESTING) {
1924                         if (blank_boundary)
1925                                 memset(hex, ' ', length);
1926                         else if (!(opt & OUTPUT_ANNOTATE_COMPAT)) {
1927                                 length--;
1928                                 putchar('^');
1929                         }
1930                 }
1931
1932                 printf("%.*s", length, hex);
1933                 if (opt & OUTPUT_ANNOTATE_COMPAT) {
1934                         const char *name;
1935                         if (opt & OUTPUT_SHOW_EMAIL)
1936                                 name = ci.author_mail.buf;
1937                         else
1938                                 name = ci.author.buf;
1939                         printf("\t(%10s\t%10s\t%d)", name,
1940                                format_time(ci.author_time, ci.author_tz.buf,
1941                                            show_raw_time),
1942                                ent->lno + 1 + cnt);
1943                 } else {
1944                         if (opt & OUTPUT_SHOW_SCORE)
1945                                 printf(" %*d %02d",
1946                                        max_score_digits, ent->score,
1947                                        ent->suspect->refcnt);
1948                         if (opt & OUTPUT_SHOW_NAME)
1949                                 printf(" %-*.*s", longest_file, longest_file,
1950                                        suspect->path);
1951                         if (opt & OUTPUT_SHOW_NUMBER)
1952                                 printf(" %*d", max_orig_digits,
1953                                        ent->s_lno + 1 + cnt);
1954
1955                         if (!(opt & OUTPUT_NO_AUTHOR)) {
1956                                 const char *name;
1957                                 int pad;
1958                                 if (opt & OUTPUT_SHOW_EMAIL)
1959                                         name = ci.author_mail.buf;
1960                                 else
1961                                         name = ci.author.buf;
1962                                 pad = longest_author - utf8_strwidth(name);
1963                                 printf(" (%s%*s %10s",
1964                                        name, pad, "",
1965                                        format_time(ci.author_time,
1966                                                    ci.author_tz.buf,
1967                                                    show_raw_time));
1968                         }
1969                         printf(" %*d) ",
1970                                max_digits, ent->lno + 1 + cnt);
1971                 }
1972                 do {
1973                         ch = *cp++;
1974                         putchar(ch);
1975                 } while (ch != '\n' &&
1976                          cp < sb->final_buf + sb->final_buf_size);
1977         }
1978
1979         if (sb->final_buf_size && cp[-1] != '\n')
1980                 putchar('\n');
1981
1982         commit_info_destroy(&ci);
1983 }
1984
1985 static void output(struct scoreboard *sb, int option)
1986 {
1987         struct blame_entry *ent;
1988
1989         if (option & OUTPUT_PORCELAIN) {
1990                 for (ent = sb->ent; ent; ent = ent->next) {
1991                         int count = 0;
1992                         struct origin *suspect;
1993                         struct commit *commit = ent->suspect->commit;
1994                         if (commit->object.flags & MORE_THAN_ONE_PATH)
1995                                 continue;
1996                         for (suspect = commit->util; suspect; suspect = suspect->next) {
1997                                 if (suspect->guilty && count++) {
1998                                         commit->object.flags |= MORE_THAN_ONE_PATH;
1999                                         break;
2000                                 }
2001                         }
2002                 }
2003         }
2004
2005         for (ent = sb->ent; ent; ent = ent->next) {
2006                 if (option & OUTPUT_PORCELAIN)
2007                         emit_porcelain(sb, ent, option);
2008                 else {
2009                         emit_other(sb, ent, option);
2010                 }
2011         }
2012 }
2013
2014 static const char *get_next_line(const char *start, const char *end)
2015 {
2016         const char *nl = memchr(start, '\n', end - start);
2017         return nl ? nl + 1 : end;
2018 }
2019
2020 /*
2021  * To allow quick access to the contents of nth line in the
2022  * final image, prepare an index in the scoreboard.
2023  */
2024 static int prepare_lines(struct scoreboard *sb)
2025 {
2026         const char *buf = sb->final_buf;
2027         unsigned long len = sb->final_buf_size;
2028         const char *end = buf + len;
2029         const char *p;
2030         int *lineno;
2031         int num = 0;
2032
2033         for (p = buf; p < end; p = get_next_line(p, end))
2034                 num++;
2035
2036         sb->lineno = lineno = xmalloc(sizeof(*sb->lineno) * (num + 1));
2037
2038         for (p = buf; p < end; p = get_next_line(p, end))
2039                 *lineno++ = p - buf;
2040
2041         *lineno = len;
2042
2043         sb->num_lines = num;
2044         return sb->num_lines;
2045 }
2046
2047 /*
2048  * Add phony grafts for use with -S; this is primarily to
2049  * support git's cvsserver that wants to give a linear history
2050  * to its clients.
2051  */
2052 static int read_ancestry(const char *graft_file)
2053 {
2054         FILE *fp = fopen(graft_file, "r");
2055         struct strbuf buf = STRBUF_INIT;
2056         if (!fp)
2057                 return -1;
2058         while (!strbuf_getwholeline(&buf, fp, '\n')) {
2059                 /* The format is just "Commit Parent1 Parent2 ...\n" */
2060                 struct commit_graft *graft = read_graft_line(buf.buf, buf.len);
2061                 if (graft)
2062                         register_commit_graft(graft, 0);
2063         }
2064         fclose(fp);
2065         strbuf_release(&buf);
2066         return 0;
2067 }
2068
2069 static int update_auto_abbrev(int auto_abbrev, struct origin *suspect)
2070 {
2071         const char *uniq = find_unique_abbrev(suspect->commit->object.sha1,
2072                                               auto_abbrev);
2073         int len = strlen(uniq);
2074         if (auto_abbrev < len)
2075                 return len;
2076         return auto_abbrev;
2077 }
2078
2079 /*
2080  * How many columns do we need to show line numbers, authors,
2081  * and filenames?
2082  */
2083 static void find_alignment(struct scoreboard *sb, int *option)
2084 {
2085         int longest_src_lines = 0;
2086         int longest_dst_lines = 0;
2087         unsigned largest_score = 0;
2088         struct blame_entry *e;
2089         int compute_auto_abbrev = (abbrev < 0);
2090         int auto_abbrev = default_abbrev;
2091
2092         for (e = sb->ent; e; e = e->next) {
2093                 struct origin *suspect = e->suspect;
2094                 int num;
2095
2096                 if (compute_auto_abbrev)
2097                         auto_abbrev = update_auto_abbrev(auto_abbrev, suspect);
2098                 if (strcmp(suspect->path, sb->path))
2099                         *option |= OUTPUT_SHOW_NAME;
2100                 num = strlen(suspect->path);
2101                 if (longest_file < num)
2102                         longest_file = num;
2103                 if (!(suspect->commit->object.flags & METAINFO_SHOWN)) {
2104                         struct commit_info ci;
2105                         suspect->commit->object.flags |= METAINFO_SHOWN;
2106                         get_commit_info(suspect->commit, &ci, 1);
2107                         if (*option & OUTPUT_SHOW_EMAIL)
2108                                 num = utf8_strwidth(ci.author_mail.buf);
2109                         else
2110                                 num = utf8_strwidth(ci.author.buf);
2111                         if (longest_author < num)
2112                                 longest_author = num;
2113                         commit_info_destroy(&ci);
2114                 }
2115                 num = e->s_lno + e->num_lines;
2116                 if (longest_src_lines < num)
2117                         longest_src_lines = num;
2118                 num = e->lno + e->num_lines;
2119                 if (longest_dst_lines < num)
2120                         longest_dst_lines = num;
2121                 if (largest_score < ent_score(sb, e))
2122                         largest_score = ent_score(sb, e);
2123         }
2124         max_orig_digits = decimal_width(longest_src_lines);
2125         max_digits = decimal_width(longest_dst_lines);
2126         max_score_digits = decimal_width(largest_score);
2127
2128         if (compute_auto_abbrev)
2129                 /* one more abbrev length is needed for the boundary commit */
2130                 abbrev = auto_abbrev + 1;
2131 }
2132
2133 /*
2134  * For debugging -- origin is refcounted, and this asserts that
2135  * we do not underflow.
2136  */
2137 static void sanity_check_refcnt(struct scoreboard *sb)
2138 {
2139         int baa = 0;
2140         struct blame_entry *ent;
2141
2142         for (ent = sb->ent; ent; ent = ent->next) {
2143                 /* Nobody should have zero or negative refcnt */
2144                 if (ent->suspect->refcnt <= 0) {
2145                         fprintf(stderr, "%s in %s has negative refcnt %d\n",
2146                                 ent->suspect->path,
2147                                 sha1_to_hex(ent->suspect->commit->object.sha1),
2148                                 ent->suspect->refcnt);
2149                         baa = 1;
2150                 }
2151         }
2152         if (baa) {
2153                 int opt = 0160;
2154                 find_alignment(sb, &opt);
2155                 output(sb, opt);
2156                 die("Baa %d!", baa);
2157         }
2158 }
2159
2160 static unsigned parse_score(const char *arg)
2161 {
2162         char *end;
2163         unsigned long score = strtoul(arg, &end, 10);
2164         if (*end)
2165                 return 0;
2166         return score;
2167 }
2168
2169 static const char *add_prefix(const char *prefix, const char *path)
2170 {
2171         return prefix_path(prefix, prefix ? strlen(prefix) : 0, path);
2172 }
2173
2174 static int git_blame_config(const char *var, const char *value, void *cb)
2175 {
2176         if (!strcmp(var, "blame.showroot")) {
2177                 show_root = git_config_bool(var, value);
2178                 return 0;
2179         }
2180         if (!strcmp(var, "blame.blankboundary")) {
2181                 blank_boundary = git_config_bool(var, value);
2182                 return 0;
2183         }
2184         if (!strcmp(var, "blame.date")) {
2185                 if (!value)
2186                         return config_error_nonbool(var);
2187                 blame_date_mode = parse_date_format(value);
2188                 return 0;
2189         }
2190
2191         if (userdiff_config(var, value) < 0)
2192                 return -1;
2193
2194         return git_default_config(var, value, cb);
2195 }
2196
2197 static void verify_working_tree_path(struct commit *work_tree, const char *path)
2198 {
2199         struct commit_list *parents;
2200
2201         for (parents = work_tree->parents; parents; parents = parents->next) {
2202                 const unsigned char *commit_sha1 = parents->item->object.sha1;
2203                 unsigned char blob_sha1[20];
2204                 unsigned mode;
2205
2206                 if (!get_tree_entry(commit_sha1, path, blob_sha1, &mode) &&
2207                     sha1_object_info(blob_sha1, NULL) == OBJ_BLOB)
2208                         return;
2209         }
2210         die("no such path '%s' in HEAD", path);
2211 }
2212
2213 static struct commit_list **append_parent(struct commit_list **tail, const unsigned char *sha1)
2214 {
2215         struct commit *parent;
2216
2217         parent = lookup_commit_reference(sha1);
2218         if (!parent)
2219                 die("no such commit %s", sha1_to_hex(sha1));
2220         return &commit_list_insert(parent, tail)->next;
2221 }
2222
2223 static void append_merge_parents(struct commit_list **tail)
2224 {
2225         int merge_head;
2226         const char *merge_head_file = git_path("MERGE_HEAD");
2227         struct strbuf line = STRBUF_INIT;
2228
2229         merge_head = open(merge_head_file, O_RDONLY);
2230         if (merge_head < 0) {
2231                 if (errno == ENOENT)
2232                         return;
2233                 die("cannot open '%s' for reading", merge_head_file);
2234         }
2235
2236         while (!strbuf_getwholeline_fd(&line, merge_head, '\n')) {
2237                 unsigned char sha1[20];
2238                 if (line.len < 40 || get_sha1_hex(line.buf, sha1))
2239                         die("unknown line in '%s': %s", merge_head_file, line.buf);
2240                 tail = append_parent(tail, sha1);
2241         }
2242         close(merge_head);
2243         strbuf_release(&line);
2244 }
2245
2246 /*
2247  * This isn't as simple as passing sb->buf and sb->len, because we
2248  * want to transfer ownership of the buffer to the commit (so we
2249  * must use detach).
2250  */
2251 static void set_commit_buffer_from_strbuf(struct commit *c, struct strbuf *sb)
2252 {
2253         size_t len;
2254         void *buf = strbuf_detach(sb, &len);
2255         set_commit_buffer(c, buf, len);
2256 }
2257
2258 /*
2259  * Prepare a dummy commit that represents the work tree (or staged) item.
2260  * Note that annotating work tree item never works in the reverse.
2261  */
2262 static struct commit *fake_working_tree_commit(struct diff_options *opt,
2263                                                const char *path,
2264                                                const char *contents_from)
2265 {
2266         struct commit *commit;
2267         struct origin *origin;
2268         struct commit_list **parent_tail, *parent;
2269         unsigned char head_sha1[20];
2270         struct strbuf buf = STRBUF_INIT;
2271         const char *ident;
2272         time_t now;
2273         int size, len;
2274         struct cache_entry *ce;
2275         unsigned mode;
2276         struct strbuf msg = STRBUF_INIT;
2277
2278         time(&now);
2279         commit = alloc_commit_node();
2280         commit->object.parsed = 1;
2281         commit->date = now;
2282         parent_tail = &commit->parents;
2283
2284         if (!resolve_ref_unsafe("HEAD", RESOLVE_REF_READING, head_sha1, NULL))
2285                 die("no such ref: HEAD");
2286
2287         parent_tail = append_parent(parent_tail, head_sha1);
2288         append_merge_parents(parent_tail);
2289         verify_working_tree_path(commit, path);
2290
2291         origin = make_origin(commit, path);
2292
2293         ident = fmt_ident("Not Committed Yet", "not.committed.yet", NULL, 0);
2294         strbuf_addstr(&msg, "tree 0000000000000000000000000000000000000000\n");
2295         for (parent = commit->parents; parent; parent = parent->next)
2296                 strbuf_addf(&msg, "parent %s\n",
2297                             sha1_to_hex(parent->item->object.sha1));
2298         strbuf_addf(&msg,
2299                     "author %s\n"
2300                     "committer %s\n\n"
2301                     "Version of %s from %s\n",
2302                     ident, ident, path,
2303                     (!contents_from ? path :
2304                      (!strcmp(contents_from, "-") ? "standard input" : contents_from)));
2305         set_commit_buffer_from_strbuf(commit, &msg);
2306
2307         if (!contents_from || strcmp("-", contents_from)) {
2308                 struct stat st;
2309                 const char *read_from;
2310                 char *buf_ptr;
2311                 unsigned long buf_len;
2312
2313                 if (contents_from) {
2314                         if (stat(contents_from, &st) < 0)
2315                                 die_errno("Cannot stat '%s'", contents_from);
2316                         read_from = contents_from;
2317                 }
2318                 else {
2319                         if (lstat(path, &st) < 0)
2320                                 die_errno("Cannot lstat '%s'", path);
2321                         read_from = path;
2322                 }
2323                 mode = canon_mode(st.st_mode);
2324
2325                 switch (st.st_mode & S_IFMT) {
2326                 case S_IFREG:
2327                         if (DIFF_OPT_TST(opt, ALLOW_TEXTCONV) &&
2328                             textconv_object(read_from, mode, null_sha1, 0, &buf_ptr, &buf_len))
2329                                 strbuf_attach(&buf, buf_ptr, buf_len, buf_len + 1);
2330                         else if (strbuf_read_file(&buf, read_from, st.st_size) != st.st_size)
2331                                 die_errno("cannot open or read '%s'", read_from);
2332                         break;
2333                 case S_IFLNK:
2334                         if (strbuf_readlink(&buf, read_from, st.st_size) < 0)
2335                                 die_errno("cannot readlink '%s'", read_from);
2336                         break;
2337                 default:
2338                         die("unsupported file type %s", read_from);
2339                 }
2340         }
2341         else {
2342                 /* Reading from stdin */
2343                 mode = 0;
2344                 if (strbuf_read(&buf, 0, 0) < 0)
2345                         die_errno("failed to read from stdin");
2346         }
2347         convert_to_git(path, buf.buf, buf.len, &buf, 0);
2348         origin->file.ptr = buf.buf;
2349         origin->file.size = buf.len;
2350         pretend_sha1_file(buf.buf, buf.len, OBJ_BLOB, origin->blob_sha1);
2351
2352         /*
2353          * Read the current index, replace the path entry with
2354          * origin->blob_sha1 without mucking with its mode or type
2355          * bits; we are not going to write this index out -- we just
2356          * want to run "diff-index --cached".
2357          */
2358         discard_cache();
2359         read_cache();
2360
2361         len = strlen(path);
2362         if (!mode) {
2363                 int pos = cache_name_pos(path, len);
2364                 if (0 <= pos)
2365                         mode = active_cache[pos]->ce_mode;
2366                 else
2367                         /* Let's not bother reading from HEAD tree */
2368                         mode = S_IFREG | 0644;
2369         }
2370         size = cache_entry_size(len);
2371         ce = xcalloc(1, size);
2372         hashcpy(ce->sha1, origin->blob_sha1);
2373         memcpy(ce->name, path, len);
2374         ce->ce_flags = create_ce_flags(0);
2375         ce->ce_namelen = len;
2376         ce->ce_mode = create_ce_mode(mode);
2377         add_cache_entry(ce, ADD_CACHE_OK_TO_ADD|ADD_CACHE_OK_TO_REPLACE);
2378
2379         /*
2380          * We are not going to write this out, so this does not matter
2381          * right now, but someday we might optimize diff-index --cached
2382          * with cache-tree information.
2383          */
2384         cache_tree_invalidate_path(&the_index, path);
2385
2386         return commit;
2387 }
2388
2389 static char *prepare_final(struct scoreboard *sb)
2390 {
2391         int i;
2392         const char *final_commit_name = NULL;
2393         struct rev_info *revs = sb->revs;
2394
2395         /*
2396          * There must be one and only one positive commit in the
2397          * revs->pending array.
2398          */
2399         for (i = 0; i < revs->pending.nr; i++) {
2400                 struct object *obj = revs->pending.objects[i].item;
2401                 if (obj->flags & UNINTERESTING)
2402                         continue;
2403                 obj = deref_tag(obj, NULL, 0);
2404                 if (obj->type != OBJ_COMMIT)
2405                         die("Non commit %s?", revs->pending.objects[i].name);
2406                 if (sb->final)
2407                         die("More than one commit to dig from %s and %s?",
2408                             revs->pending.objects[i].name,
2409                             final_commit_name);
2410                 sb->final = (struct commit *) obj;
2411                 final_commit_name = revs->pending.objects[i].name;
2412         }
2413         return xstrdup_or_null(final_commit_name);
2414 }
2415
2416 static char *prepare_initial(struct scoreboard *sb)
2417 {
2418         int i;
2419         const char *final_commit_name = NULL;
2420         struct rev_info *revs = sb->revs;
2421
2422         /*
2423          * There must be one and only one negative commit, and it must be
2424          * the boundary.
2425          */
2426         for (i = 0; i < revs->pending.nr; i++) {
2427                 struct object *obj = revs->pending.objects[i].item;
2428                 if (!(obj->flags & UNINTERESTING))
2429                         continue;
2430                 obj = deref_tag(obj, NULL, 0);
2431                 if (obj->type != OBJ_COMMIT)
2432                         die("Non commit %s?", revs->pending.objects[i].name);
2433                 if (sb->final)
2434                         die("More than one commit to dig down to %s and %s?",
2435                             revs->pending.objects[i].name,
2436                             final_commit_name);
2437                 sb->final = (struct commit *) obj;
2438                 final_commit_name = revs->pending.objects[i].name;
2439         }
2440         if (!final_commit_name)
2441                 die("No commit to dig down to?");
2442         return xstrdup(final_commit_name);
2443 }
2444
2445 static int blame_copy_callback(const struct option *option, const char *arg, int unset)
2446 {
2447         int *opt = option->value;
2448
2449         /*
2450          * -C enables copy from removed files;
2451          * -C -C enables copy from existing files, but only
2452          *       when blaming a new file;
2453          * -C -C -C enables copy from existing files for
2454          *          everybody
2455          */
2456         if (*opt & PICKAXE_BLAME_COPY_HARDER)
2457                 *opt |= PICKAXE_BLAME_COPY_HARDEST;
2458         if (*opt & PICKAXE_BLAME_COPY)
2459                 *opt |= PICKAXE_BLAME_COPY_HARDER;
2460         *opt |= PICKAXE_BLAME_COPY | PICKAXE_BLAME_MOVE;
2461
2462         if (arg)
2463                 blame_copy_score = parse_score(arg);
2464         return 0;
2465 }
2466
2467 static int blame_move_callback(const struct option *option, const char *arg, int unset)
2468 {
2469         int *opt = option->value;
2470
2471         *opt |= PICKAXE_BLAME_MOVE;
2472
2473         if (arg)
2474                 blame_move_score = parse_score(arg);
2475         return 0;
2476 }
2477
2478 int cmd_blame(int argc, const char **argv, const char *prefix)
2479 {
2480         struct rev_info revs;
2481         const char *path;
2482         struct scoreboard sb;
2483         struct origin *o;
2484         struct blame_entry *ent = NULL;
2485         long dashdash_pos, lno;
2486         char *final_commit_name = NULL;
2487         enum object_type type;
2488
2489         static struct string_list range_list;
2490         static int output_option = 0, opt = 0;
2491         static int show_stats = 0;
2492         static const char *revs_file = NULL;
2493         static const char *contents_from = NULL;
2494         static const struct option options[] = {
2495                 OPT_BOOL(0, "incremental", &incremental, N_("Show blame entries as we find them, incrementally")),
2496                 OPT_BOOL('b', NULL, &blank_boundary, N_("Show blank SHA-1 for boundary commits (Default: off)")),
2497                 OPT_BOOL(0, "root", &show_root, N_("Do not treat root commits as boundaries (Default: off)")),
2498                 OPT_BOOL(0, "show-stats", &show_stats, N_("Show work cost statistics")),
2499                 OPT_BIT(0, "score-debug", &output_option, N_("Show output score for blame entries"), OUTPUT_SHOW_SCORE),
2500                 OPT_BIT('f', "show-name", &output_option, N_("Show original filename (Default: auto)"), OUTPUT_SHOW_NAME),
2501                 OPT_BIT('n', "show-number", &output_option, N_("Show original linenumber (Default: off)"), OUTPUT_SHOW_NUMBER),
2502                 OPT_BIT('p', "porcelain", &output_option, N_("Show in a format designed for machine consumption"), OUTPUT_PORCELAIN),
2503                 OPT_BIT(0, "line-porcelain", &output_option, N_("Show porcelain format with per-line commit information"), OUTPUT_PORCELAIN|OUTPUT_LINE_PORCELAIN),
2504                 OPT_BIT('c', NULL, &output_option, N_("Use the same output mode as git-annotate (Default: off)"), OUTPUT_ANNOTATE_COMPAT),
2505                 OPT_BIT('t', NULL, &output_option, N_("Show raw timestamp (Default: off)"), OUTPUT_RAW_TIMESTAMP),
2506                 OPT_BIT('l', NULL, &output_option, N_("Show long commit SHA1 (Default: off)"), OUTPUT_LONG_OBJECT_NAME),
2507                 OPT_BIT('s', NULL, &output_option, N_("Suppress author name and timestamp (Default: off)"), OUTPUT_NO_AUTHOR),
2508                 OPT_BIT('e', "show-email", &output_option, N_("Show author email instead of name (Default: off)"), OUTPUT_SHOW_EMAIL),
2509                 OPT_BIT('w', NULL, &xdl_opts, N_("Ignore whitespace differences"), XDF_IGNORE_WHITESPACE),
2510                 OPT_BIT(0, "minimal", &xdl_opts, N_("Spend extra cycles to find better match"), XDF_NEED_MINIMAL),
2511                 OPT_STRING('S', NULL, &revs_file, N_("file"), N_("Use revisions from <file> instead of calling git-rev-list")),
2512                 OPT_STRING(0, "contents", &contents_from, N_("file"), N_("Use <file>'s contents as the final image")),
2513                 { OPTION_CALLBACK, 'C', NULL, &opt, N_("score"), N_("Find line copies within and across files"), PARSE_OPT_OPTARG, blame_copy_callback },
2514                 { OPTION_CALLBACK, 'M', NULL, &opt, N_("score"), N_("Find line movements within and across files"), PARSE_OPT_OPTARG, blame_move_callback },
2515                 OPT_STRING_LIST('L', NULL, &range_list, N_("n,m"), N_("Process only line range n,m, counting from 1")),
2516                 OPT__ABBREV(&abbrev),
2517                 OPT_END()
2518         };
2519
2520         struct parse_opt_ctx_t ctx;
2521         int cmd_is_annotate = !strcmp(argv[0], "annotate");
2522         struct range_set ranges;
2523         unsigned int range_i;
2524         long anchor;
2525
2526         git_config(git_blame_config, NULL);
2527         init_revisions(&revs, NULL);
2528         revs.date_mode = blame_date_mode;
2529         DIFF_OPT_SET(&revs.diffopt, ALLOW_TEXTCONV);
2530         DIFF_OPT_SET(&revs.diffopt, FOLLOW_RENAMES);
2531
2532         save_commit_buffer = 0;
2533         dashdash_pos = 0;
2534
2535         parse_options_start(&ctx, argc, argv, prefix, options,
2536                             PARSE_OPT_KEEP_DASHDASH | PARSE_OPT_KEEP_ARGV0);
2537         for (;;) {
2538                 switch (parse_options_step(&ctx, options, blame_opt_usage)) {
2539                 case PARSE_OPT_HELP:
2540                         exit(129);
2541                 case PARSE_OPT_DONE:
2542                         if (ctx.argv[0])
2543                                 dashdash_pos = ctx.cpidx;
2544                         goto parse_done;
2545                 }
2546
2547                 if (!strcmp(ctx.argv[0], "--reverse")) {
2548                         ctx.argv[0] = "--children";
2549                         reverse = 1;
2550                 }
2551                 parse_revision_opt(&revs, &ctx, options, blame_opt_usage);
2552         }
2553 parse_done:
2554         no_whole_file_rename = !DIFF_OPT_TST(&revs.diffopt, FOLLOW_RENAMES);
2555         DIFF_OPT_CLR(&revs.diffopt, FOLLOW_RENAMES);
2556         argc = parse_options_end(&ctx);
2557
2558         if (0 < abbrev)
2559                 /* one more abbrev length is needed for the boundary commit */
2560                 abbrev++;
2561
2562         if (revs_file && read_ancestry(revs_file))
2563                 die_errno("reading graft file '%s' failed", revs_file);
2564
2565         if (cmd_is_annotate) {
2566                 output_option |= OUTPUT_ANNOTATE_COMPAT;
2567                 blame_date_mode = DATE_ISO8601;
2568         } else {
2569                 blame_date_mode = revs.date_mode;
2570         }
2571
2572         /* The maximum width used to show the dates */
2573         switch (blame_date_mode) {
2574         case DATE_RFC2822:
2575                 blame_date_width = sizeof("Thu, 19 Oct 2006 16:00:04 -0700");
2576                 break;
2577         case DATE_ISO8601_STRICT:
2578                 blame_date_width = sizeof("2006-10-19T16:00:04-07:00");
2579                 break;
2580         case DATE_ISO8601:
2581                 blame_date_width = sizeof("2006-10-19 16:00:04 -0700");
2582                 break;
2583         case DATE_RAW:
2584                 blame_date_width = sizeof("1161298804 -0700");
2585                 break;
2586         case DATE_SHORT:
2587                 blame_date_width = sizeof("2006-10-19");
2588                 break;
2589         case DATE_RELATIVE:
2590                 /* TRANSLATORS: This string is used to tell us the maximum
2591                    display width for a relative timestamp in "git blame"
2592                    output.  For C locale, "4 years, 11 months ago", which
2593                    takes 22 places, is the longest among various forms of
2594                    relative timestamps, but your language may need more or
2595                    fewer display columns. */
2596                 blame_date_width = utf8_strwidth(_("4 years, 11 months ago")) + 1; /* add the null */
2597                 break;
2598         case DATE_LOCAL:
2599         case DATE_NORMAL:
2600                 blame_date_width = sizeof("Thu Oct 19 16:00:04 2006 -0700");
2601                 break;
2602         }
2603         blame_date_width -= 1; /* strip the null */
2604
2605         if (DIFF_OPT_TST(&revs.diffopt, FIND_COPIES_HARDER))
2606                 opt |= (PICKAXE_BLAME_COPY | PICKAXE_BLAME_MOVE |
2607                         PICKAXE_BLAME_COPY_HARDER);
2608
2609         if (!blame_move_score)
2610                 blame_move_score = BLAME_DEFAULT_MOVE_SCORE;
2611         if (!blame_copy_score)
2612                 blame_copy_score = BLAME_DEFAULT_COPY_SCORE;
2613
2614         /*
2615          * We have collected options unknown to us in argv[1..unk]
2616          * which are to be passed to revision machinery if we are
2617          * going to do the "bottom" processing.
2618          *
2619          * The remaining are:
2620          *
2621          * (1) if dashdash_pos != 0, it is either
2622          *     "blame [revisions] -- <path>" or
2623          *     "blame -- <path> <rev>"
2624          *
2625          * (2) otherwise, it is one of the two:
2626          *     "blame [revisions] <path>"
2627          *     "blame <path> <rev>"
2628          *
2629          * Note that we must strip out <path> from the arguments: we do not
2630          * want the path pruning but we may want "bottom" processing.
2631          */
2632         if (dashdash_pos) {
2633                 switch (argc - dashdash_pos - 1) {
2634                 case 2: /* (1b) */
2635                         if (argc != 4)
2636                                 usage_with_options(blame_opt_usage, options);
2637                         /* reorder for the new way: <rev> -- <path> */
2638                         argv[1] = argv[3];
2639                         argv[3] = argv[2];
2640                         argv[2] = "--";
2641                         /* FALLTHROUGH */
2642                 case 1: /* (1a) */
2643                         path = add_prefix(prefix, argv[--argc]);
2644                         argv[argc] = NULL;
2645                         break;
2646                 default:
2647                         usage_with_options(blame_opt_usage, options);
2648                 }
2649         } else {
2650                 if (argc < 2)
2651                         usage_with_options(blame_opt_usage, options);
2652                 path = add_prefix(prefix, argv[argc - 1]);
2653                 if (argc == 3 && !file_exists(path)) { /* (2b) */
2654                         path = add_prefix(prefix, argv[1]);
2655                         argv[1] = argv[2];
2656                 }
2657                 argv[argc - 1] = "--";
2658
2659                 setup_work_tree();
2660                 if (!file_exists(path))
2661                         die_errno("cannot stat path '%s'", path);
2662         }
2663
2664         revs.disable_stdin = 1;
2665         setup_revisions(argc, argv, &revs, NULL);
2666         memset(&sb, 0, sizeof(sb));
2667
2668         sb.revs = &revs;
2669         if (!reverse) {
2670                 final_commit_name = prepare_final(&sb);
2671                 sb.commits.compare = compare_commits_by_commit_date;
2672         }
2673         else if (contents_from)
2674                 die("--contents and --children do not blend well.");
2675         else {
2676                 final_commit_name = prepare_initial(&sb);
2677                 sb.commits.compare = compare_commits_by_reverse_commit_date;
2678         }
2679
2680         if (!sb.final) {
2681                 /*
2682                  * "--not A B -- path" without anything positive;
2683                  * do not default to HEAD, but use the working tree
2684                  * or "--contents".
2685                  */
2686                 setup_work_tree();
2687                 sb.final = fake_working_tree_commit(&sb.revs->diffopt,
2688                                                     path, contents_from);
2689                 add_pending_object(&revs, &(sb.final->object), ":");
2690         }
2691         else if (contents_from)
2692                 die("Cannot use --contents with final commit object name");
2693
2694         /*
2695          * If we have bottom, this will mark the ancestors of the
2696          * bottom commits we would reach while traversing as
2697          * uninteresting.
2698          */
2699         if (prepare_revision_walk(&revs))
2700                 die(_("revision walk setup failed"));
2701
2702         if (is_null_sha1(sb.final->object.sha1)) {
2703                 o = sb.final->util;
2704                 sb.final_buf = xmemdupz(o->file.ptr, o->file.size);
2705                 sb.final_buf_size = o->file.size;
2706         }
2707         else {
2708                 o = get_origin(&sb, sb.final, path);
2709                 if (fill_blob_sha1_and_mode(o))
2710                         die("no such path %s in %s", path, final_commit_name);
2711
2712                 if (DIFF_OPT_TST(&sb.revs->diffopt, ALLOW_TEXTCONV) &&
2713                     textconv_object(path, o->mode, o->blob_sha1, 1, (char **) &sb.final_buf,
2714                                     &sb.final_buf_size))
2715                         ;
2716                 else
2717                         sb.final_buf = read_sha1_file(o->blob_sha1, &type,
2718                                                       &sb.final_buf_size);
2719
2720                 if (!sb.final_buf)
2721                         die("Cannot read blob %s for path %s",
2722                             sha1_to_hex(o->blob_sha1),
2723                             path);
2724         }
2725         num_read_blob++;
2726         lno = prepare_lines(&sb);
2727
2728         if (lno && !range_list.nr)
2729                 string_list_append(&range_list, xstrdup("1"));
2730
2731         anchor = 1;
2732         range_set_init(&ranges, range_list.nr);
2733         for (range_i = 0; range_i < range_list.nr; ++range_i) {
2734                 long bottom, top;
2735                 if (parse_range_arg(range_list.items[range_i].string,
2736                                     nth_line_cb, &sb, lno, anchor,
2737                                     &bottom, &top, sb.path))
2738                         usage(blame_usage);
2739                 if (lno < top || ((lno || bottom) && lno < bottom))
2740                         die("file %s has only %lu lines", path, lno);
2741                 if (bottom < 1)
2742                         bottom = 1;
2743                 if (top < 1)
2744                         top = lno;
2745                 bottom--;
2746                 range_set_append_unsafe(&ranges, bottom, top);
2747                 anchor = top + 1;
2748         }
2749         sort_and_merge_range_set(&ranges);
2750
2751         for (range_i = ranges.nr; range_i > 0; --range_i) {
2752                 const struct range *r = &ranges.ranges[range_i - 1];
2753                 long bottom = r->start;
2754                 long top = r->end;
2755                 struct blame_entry *next = ent;
2756                 ent = xcalloc(1, sizeof(*ent));
2757                 ent->lno = bottom;
2758                 ent->num_lines = top - bottom;
2759                 ent->suspect = o;
2760                 ent->s_lno = bottom;
2761                 ent->next = next;
2762                 origin_incref(o);
2763         }
2764
2765         o->suspects = ent;
2766         prio_queue_put(&sb.commits, o->commit);
2767
2768         origin_decref(o);
2769
2770         range_set_release(&ranges);
2771         string_list_clear(&range_list, 0);
2772
2773         sb.ent = NULL;
2774         sb.path = path;
2775
2776         read_mailmap(&mailmap, NULL);
2777
2778         if (!incremental)
2779                 setup_pager();
2780
2781         assign_blame(&sb, opt);
2782
2783         free(final_commit_name);
2784
2785         if (incremental)
2786                 return 0;
2787
2788         sb.ent = blame_sort(sb.ent, compare_blame_final);
2789
2790         coalesce(&sb);
2791
2792         if (!(output_option & OUTPUT_PORCELAIN))
2793                 find_alignment(&sb, &output_option);
2794
2795         output(&sb, output_option);
2796         free((void *)sb.final_buf);
2797         for (ent = sb.ent; ent; ) {
2798                 struct blame_entry *e = ent->next;
2799                 free(ent);
2800                 ent = e;
2801         }
2802
2803         if (show_stats) {
2804                 printf("num read blob: %d\n", num_read_blob);
2805                 printf("num get patch: %d\n", num_get_patch);
2806                 printf("num commits: %d\n", num_commits);
2807         }
2808         return 0;
2809 }