Merge branch 'ah/show-ref-usage-string'
[git] / builtin / blame.c
1 /*
2  * Blame
3  *
4  * Copyright (c) 2006, 2014 by its authors
5  * See COPYING for licensing conditions
6  */
7
8 #include "cache.h"
9 #include "refs.h"
10 #include "builtin.h"
11 #include "blob.h"
12 #include "commit.h"
13 #include "tag.h"
14 #include "tree-walk.h"
15 #include "diff.h"
16 #include "diffcore.h"
17 #include "revision.h"
18 #include "quote.h"
19 #include "xdiff-interface.h"
20 #include "cache-tree.h"
21 #include "string-list.h"
22 #include "mailmap.h"
23 #include "mergesort.h"
24 #include "parse-options.h"
25 #include "prio-queue.h"
26 #include "utf8.h"
27 #include "userdiff.h"
28 #include "line-range.h"
29 #include "line-log.h"
30 #include "dir.h"
31
32 static char blame_usage[] = N_("git blame [<options>] [<rev-opts>] [<rev>] [--] <file>");
33
34 static const char *blame_opt_usage[] = {
35         blame_usage,
36         "",
37         N_("<rev-opts> are documented in git-rev-list(1)"),
38         NULL
39 };
40
41 static int longest_file;
42 static int longest_author;
43 static int max_orig_digits;
44 static int max_digits;
45 static int max_score_digits;
46 static int show_root;
47 static int reverse;
48 static int blank_boundary;
49 static int incremental;
50 static int xdl_opts;
51 static int abbrev = -1;
52 static int no_whole_file_rename;
53
54 static struct date_mode blame_date_mode = { DATE_ISO8601 };
55 static size_t blame_date_width;
56
57 static struct string_list mailmap;
58
59 #ifndef DEBUG
60 #define DEBUG 0
61 #endif
62
63 /* stats */
64 static int num_read_blob;
65 static int num_get_patch;
66 static int num_commits;
67
68 #define PICKAXE_BLAME_MOVE              01
69 #define PICKAXE_BLAME_COPY              02
70 #define PICKAXE_BLAME_COPY_HARDER       04
71 #define PICKAXE_BLAME_COPY_HARDEST      010
72
73 /*
74  * blame for a blame_entry with score lower than these thresholds
75  * is not passed to the parent using move/copy logic.
76  */
77 static unsigned blame_move_score;
78 static unsigned blame_copy_score;
79 #define BLAME_DEFAULT_MOVE_SCORE        20
80 #define BLAME_DEFAULT_COPY_SCORE        40
81
82 /* Remember to update object flag allocation in object.h */
83 #define METAINFO_SHOWN          (1u<<12)
84 #define MORE_THAN_ONE_PATH      (1u<<13)
85
86 /*
87  * One blob in a commit that is being suspected
88  */
89 struct origin {
90         int refcnt;
91         /* Record preceding blame record for this blob */
92         struct origin *previous;
93         /* origins are put in a list linked via `next' hanging off the
94          * corresponding commit's util field in order to make finding
95          * them fast.  The presence in this chain does not count
96          * towards the origin's reference count.  It is tempting to
97          * let it count as long as the commit is pending examination,
98          * but even under circumstances where the commit will be
99          * present multiple times in the priority queue of unexamined
100          * commits, processing the first instance will not leave any
101          * work requiring the origin data for the second instance.  An
102          * interspersed commit changing that would have to be
103          * preexisting with a different ancestry and with the same
104          * commit date in order to wedge itself between two instances
105          * of the same commit in the priority queue _and_ produce
106          * blame entries relevant for it.  While we don't want to let
107          * us get tripped up by this case, it certainly does not seem
108          * worth optimizing for.
109          */
110         struct origin *next;
111         struct commit *commit;
112         /* `suspects' contains blame entries that may be attributed to
113          * this origin's commit or to parent commits.  When a commit
114          * is being processed, all suspects will be moved, either by
115          * assigning them to an origin in a different commit, or by
116          * shipping them to the scoreboard's ent list because they
117          * cannot be attributed to a different commit.
118          */
119         struct blame_entry *suspects;
120         mmfile_t file;
121         unsigned char blob_sha1[20];
122         unsigned mode;
123         /* guilty gets set when shipping any suspects to the final
124          * blame list instead of other commits
125          */
126         char guilty;
127         char path[FLEX_ARRAY];
128 };
129
130 static int diff_hunks(mmfile_t *file_a, mmfile_t *file_b, long ctxlen,
131                       xdl_emit_hunk_consume_func_t hunk_func, void *cb_data)
132 {
133         xpparam_t xpp = {0};
134         xdemitconf_t xecfg = {0};
135         xdemitcb_t ecb = {NULL};
136
137         xpp.flags = xdl_opts;
138         xecfg.ctxlen = ctxlen;
139         xecfg.hunk_func = hunk_func;
140         ecb.priv = cb_data;
141         return xdi_diff(file_a, file_b, &xpp, &xecfg, &ecb);
142 }
143
144 /*
145  * Prepare diff_filespec and convert it using diff textconv API
146  * if the textconv driver exists.
147  * Return 1 if the conversion succeeds, 0 otherwise.
148  */
149 int textconv_object(const char *path,
150                     unsigned mode,
151                     const unsigned char *sha1,
152                     int sha1_valid,
153                     char **buf,
154                     unsigned long *buf_size)
155 {
156         struct diff_filespec *df;
157         struct userdiff_driver *textconv;
158
159         df = alloc_filespec(path);
160         fill_filespec(df, sha1, sha1_valid, mode);
161         textconv = get_textconv(df);
162         if (!textconv) {
163                 free_filespec(df);
164                 return 0;
165         }
166
167         *buf_size = fill_textconv(textconv, df, buf);
168         free_filespec(df);
169         return 1;
170 }
171
172 /*
173  * Given an origin, prepare mmfile_t structure to be used by the
174  * diff machinery
175  */
176 static void fill_origin_blob(struct diff_options *opt,
177                              struct origin *o, mmfile_t *file)
178 {
179         if (!o->file.ptr) {
180                 enum object_type type;
181                 unsigned long file_size;
182
183                 num_read_blob++;
184                 if (DIFF_OPT_TST(opt, ALLOW_TEXTCONV) &&
185                     textconv_object(o->path, o->mode, o->blob_sha1, 1, &file->ptr, &file_size))
186                         ;
187                 else
188                         file->ptr = read_sha1_file(o->blob_sha1, &type, &file_size);
189                 file->size = file_size;
190
191                 if (!file->ptr)
192                         die("Cannot read blob %s for path %s",
193                             sha1_to_hex(o->blob_sha1),
194                             o->path);
195                 o->file = *file;
196         }
197         else
198                 *file = o->file;
199 }
200
201 /*
202  * Origin is refcounted and usually we keep the blob contents to be
203  * reused.
204  */
205 static inline struct origin *origin_incref(struct origin *o)
206 {
207         if (o)
208                 o->refcnt++;
209         return o;
210 }
211
212 static void origin_decref(struct origin *o)
213 {
214         if (o && --o->refcnt <= 0) {
215                 struct origin *p, *l = NULL;
216                 if (o->previous)
217                         origin_decref(o->previous);
218                 free(o->file.ptr);
219                 /* Should be present exactly once in commit chain */
220                 for (p = o->commit->util; p; l = p, p = p->next) {
221                         if (p == o) {
222                                 if (l)
223                                         l->next = p->next;
224                                 else
225                                         o->commit->util = p->next;
226                                 free(o);
227                                 return;
228                         }
229                 }
230                 die("internal error in blame::origin_decref");
231         }
232 }
233
234 static void drop_origin_blob(struct origin *o)
235 {
236         if (o->file.ptr) {
237                 free(o->file.ptr);
238                 o->file.ptr = NULL;
239         }
240 }
241
242 /*
243  * Each group of lines is described by a blame_entry; it can be split
244  * as we pass blame to the parents.  They are arranged in linked lists
245  * kept as `suspects' of some unprocessed origin, or entered (when the
246  * blame origin has been finalized) into the scoreboard structure.
247  * While the scoreboard structure is only sorted at the end of
248  * processing (according to final image line number), the lists
249  * attached to an origin are sorted by the target line number.
250  */
251 struct blame_entry {
252         struct blame_entry *next;
253
254         /* the first line of this group in the final image;
255          * internally all line numbers are 0 based.
256          */
257         int lno;
258
259         /* how many lines this group has */
260         int num_lines;
261
262         /* the commit that introduced this group into the final image */
263         struct origin *suspect;
264
265         /* the line number of the first line of this group in the
266          * suspect's file; internally all line numbers are 0 based.
267          */
268         int s_lno;
269
270         /* how significant this entry is -- cached to avoid
271          * scanning the lines over and over.
272          */
273         unsigned score;
274 };
275
276 /*
277  * Any merge of blames happens on lists of blames that arrived via
278  * different parents in a single suspect.  In this case, we want to
279  * sort according to the suspect line numbers as opposed to the final
280  * image line numbers.  The function body is somewhat longish because
281  * it avoids unnecessary writes.
282  */
283
284 static struct blame_entry *blame_merge(struct blame_entry *list1,
285                                        struct blame_entry *list2)
286 {
287         struct blame_entry *p1 = list1, *p2 = list2,
288                 **tail = &list1;
289
290         if (!p1)
291                 return p2;
292         if (!p2)
293                 return p1;
294
295         if (p1->s_lno <= p2->s_lno) {
296                 do {
297                         tail = &p1->next;
298                         if ((p1 = *tail) == NULL) {
299                                 *tail = p2;
300                                 return list1;
301                         }
302                 } while (p1->s_lno <= p2->s_lno);
303         }
304         for (;;) {
305                 *tail = p2;
306                 do {
307                         tail = &p2->next;
308                         if ((p2 = *tail) == NULL)  {
309                                 *tail = p1;
310                                 return list1;
311                         }
312                 } while (p1->s_lno > p2->s_lno);
313                 *tail = p1;
314                 do {
315                         tail = &p1->next;
316                         if ((p1 = *tail) == NULL) {
317                                 *tail = p2;
318                                 return list1;
319                         }
320                 } while (p1->s_lno <= p2->s_lno);
321         }
322 }
323
324 static void *get_next_blame(const void *p)
325 {
326         return ((struct blame_entry *)p)->next;
327 }
328
329 static void set_next_blame(void *p1, void *p2)
330 {
331         ((struct blame_entry *)p1)->next = p2;
332 }
333
334 /*
335  * Final image line numbers are all different, so we don't need a
336  * three-way comparison here.
337  */
338
339 static int compare_blame_final(const void *p1, const void *p2)
340 {
341         return ((struct blame_entry *)p1)->lno > ((struct blame_entry *)p2)->lno
342                 ? 1 : -1;
343 }
344
345 static int compare_blame_suspect(const void *p1, const void *p2)
346 {
347         const struct blame_entry *s1 = p1, *s2 = p2;
348         /*
349          * to allow for collating suspects, we sort according to the
350          * respective pointer value as the primary sorting criterion.
351          * The actual relation is pretty unimportant as long as it
352          * establishes a total order.  Comparing as integers gives us
353          * that.
354          */
355         if (s1->suspect != s2->suspect)
356                 return (intptr_t)s1->suspect > (intptr_t)s2->suspect ? 1 : -1;
357         if (s1->s_lno == s2->s_lno)
358                 return 0;
359         return s1->s_lno > s2->s_lno ? 1 : -1;
360 }
361
362 static struct blame_entry *blame_sort(struct blame_entry *head,
363                                       int (*compare_fn)(const void *, const void *))
364 {
365         return llist_mergesort (head, get_next_blame, set_next_blame, compare_fn);
366 }
367
368 static int compare_commits_by_reverse_commit_date(const void *a,
369                                                   const void *b,
370                                                   void *c)
371 {
372         return -compare_commits_by_commit_date(a, b, c);
373 }
374
375 /*
376  * The current state of the blame assignment.
377  */
378 struct scoreboard {
379         /* the final commit (i.e. where we started digging from) */
380         struct commit *final;
381         /* Priority queue for commits with unassigned blame records */
382         struct prio_queue commits;
383         struct rev_info *revs;
384         const char *path;
385
386         /*
387          * The contents in the final image.
388          * Used by many functions to obtain contents of the nth line,
389          * indexed with scoreboard.lineno[blame_entry.lno].
390          */
391         const char *final_buf;
392         unsigned long final_buf_size;
393
394         /* linked list of blames */
395         struct blame_entry *ent;
396
397         /* look-up a line in the final buffer */
398         int num_lines;
399         int *lineno;
400 };
401
402 static void sanity_check_refcnt(struct scoreboard *);
403
404 /*
405  * If two blame entries that are next to each other came from
406  * contiguous lines in the same origin (i.e. <commit, path> pair),
407  * merge them together.
408  */
409 static void coalesce(struct scoreboard *sb)
410 {
411         struct blame_entry *ent, *next;
412
413         for (ent = sb->ent; ent && (next = ent->next); ent = next) {
414                 if (ent->suspect == next->suspect &&
415                     ent->s_lno + ent->num_lines == next->s_lno) {
416                         ent->num_lines += next->num_lines;
417                         ent->next = next->next;
418                         origin_decref(next->suspect);
419                         free(next);
420                         ent->score = 0;
421                         next = ent; /* again */
422                 }
423         }
424
425         if (DEBUG) /* sanity */
426                 sanity_check_refcnt(sb);
427 }
428
429 /*
430  * Merge the given sorted list of blames into a preexisting origin.
431  * If there were no previous blames to that commit, it is entered into
432  * the commit priority queue of the score board.
433  */
434
435 static void queue_blames(struct scoreboard *sb, struct origin *porigin,
436                          struct blame_entry *sorted)
437 {
438         if (porigin->suspects)
439                 porigin->suspects = blame_merge(porigin->suspects, sorted);
440         else {
441                 struct origin *o;
442                 for (o = porigin->commit->util; o; o = o->next) {
443                         if (o->suspects) {
444                                 porigin->suspects = sorted;
445                                 return;
446                         }
447                 }
448                 porigin->suspects = sorted;
449                 prio_queue_put(&sb->commits, porigin->commit);
450         }
451 }
452
453 /*
454  * Given a commit and a path in it, create a new origin structure.
455  * The callers that add blame to the scoreboard should use
456  * get_origin() to obtain shared, refcounted copy instead of calling
457  * this function directly.
458  */
459 static struct origin *make_origin(struct commit *commit, const char *path)
460 {
461         struct origin *o;
462         o = xcalloc(1, sizeof(*o) + strlen(path) + 1);
463         o->commit = commit;
464         o->refcnt = 1;
465         o->next = commit->util;
466         commit->util = o;
467         strcpy(o->path, path);
468         return o;
469 }
470
471 /*
472  * Locate an existing origin or create a new one.
473  * This moves the origin to front position in the commit util list.
474  */
475 static struct origin *get_origin(struct scoreboard *sb,
476                                  struct commit *commit,
477                                  const char *path)
478 {
479         struct origin *o, *l;
480
481         for (o = commit->util, l = NULL; o; l = o, o = o->next) {
482                 if (!strcmp(o->path, path)) {
483                         /* bump to front */
484                         if (l) {
485                                 l->next = o->next;
486                                 o->next = commit->util;
487                                 commit->util = o;
488                         }
489                         return origin_incref(o);
490                 }
491         }
492         return make_origin(commit, path);
493 }
494
495 /*
496  * Fill the blob_sha1 field of an origin if it hasn't, so that later
497  * call to fill_origin_blob() can use it to locate the data.  blob_sha1
498  * for an origin is also used to pass the blame for the entire file to
499  * the parent to detect the case where a child's blob is identical to
500  * that of its parent's.
501  *
502  * This also fills origin->mode for corresponding tree path.
503  */
504 static int fill_blob_sha1_and_mode(struct origin *origin)
505 {
506         if (!is_null_sha1(origin->blob_sha1))
507                 return 0;
508         if (get_tree_entry(origin->commit->object.sha1,
509                            origin->path,
510                            origin->blob_sha1, &origin->mode))
511                 goto error_out;
512         if (sha1_object_info(origin->blob_sha1, NULL) != OBJ_BLOB)
513                 goto error_out;
514         return 0;
515  error_out:
516         hashclr(origin->blob_sha1);
517         origin->mode = S_IFINVALID;
518         return -1;
519 }
520
521 /*
522  * We have an origin -- check if the same path exists in the
523  * parent and return an origin structure to represent it.
524  */
525 static struct origin *find_origin(struct scoreboard *sb,
526                                   struct commit *parent,
527                                   struct origin *origin)
528 {
529         struct origin *porigin;
530         struct diff_options diff_opts;
531         const char *paths[2];
532
533         /* First check any existing origins */
534         for (porigin = parent->util; porigin; porigin = porigin->next)
535                 if (!strcmp(porigin->path, origin->path)) {
536                         /*
537                          * The same path between origin and its parent
538                          * without renaming -- the most common case.
539                          */
540                         return origin_incref (porigin);
541                 }
542
543         /* See if the origin->path is different between parent
544          * and origin first.  Most of the time they are the
545          * same and diff-tree is fairly efficient about this.
546          */
547         diff_setup(&diff_opts);
548         DIFF_OPT_SET(&diff_opts, RECURSIVE);
549         diff_opts.detect_rename = 0;
550         diff_opts.output_format = DIFF_FORMAT_NO_OUTPUT;
551         paths[0] = origin->path;
552         paths[1] = NULL;
553
554         parse_pathspec(&diff_opts.pathspec,
555                        PATHSPEC_ALL_MAGIC & ~PATHSPEC_LITERAL,
556                        PATHSPEC_LITERAL_PATH, "", paths);
557         diff_setup_done(&diff_opts);
558
559         if (is_null_sha1(origin->commit->object.sha1))
560                 do_diff_cache(parent->tree->object.sha1, &diff_opts);
561         else
562                 diff_tree_sha1(parent->tree->object.sha1,
563                                origin->commit->tree->object.sha1,
564                                "", &diff_opts);
565         diffcore_std(&diff_opts);
566
567         if (!diff_queued_diff.nr) {
568                 /* The path is the same as parent */
569                 porigin = get_origin(sb, parent, origin->path);
570                 hashcpy(porigin->blob_sha1, origin->blob_sha1);
571                 porigin->mode = origin->mode;
572         } else {
573                 /*
574                  * Since origin->path is a pathspec, if the parent
575                  * commit had it as a directory, we will see a whole
576                  * bunch of deletion of files in the directory that we
577                  * do not care about.
578                  */
579                 int i;
580                 struct diff_filepair *p = NULL;
581                 for (i = 0; i < diff_queued_diff.nr; i++) {
582                         const char *name;
583                         p = diff_queued_diff.queue[i];
584                         name = p->one->path ? p->one->path : p->two->path;
585                         if (!strcmp(name, origin->path))
586                                 break;
587                 }
588                 if (!p)
589                         die("internal error in blame::find_origin");
590                 switch (p->status) {
591                 default:
592                         die("internal error in blame::find_origin (%c)",
593                             p->status);
594                 case 'M':
595                         porigin = get_origin(sb, parent, origin->path);
596                         hashcpy(porigin->blob_sha1, p->one->sha1);
597                         porigin->mode = p->one->mode;
598                         break;
599                 case 'A':
600                 case 'T':
601                         /* Did not exist in parent, or type changed */
602                         break;
603                 }
604         }
605         diff_flush(&diff_opts);
606         free_pathspec(&diff_opts.pathspec);
607         return porigin;
608 }
609
610 /*
611  * We have an origin -- find the path that corresponds to it in its
612  * parent and return an origin structure to represent it.
613  */
614 static struct origin *find_rename(struct scoreboard *sb,
615                                   struct commit *parent,
616                                   struct origin *origin)
617 {
618         struct origin *porigin = NULL;
619         struct diff_options diff_opts;
620         int i;
621
622         diff_setup(&diff_opts);
623         DIFF_OPT_SET(&diff_opts, RECURSIVE);
624         diff_opts.detect_rename = DIFF_DETECT_RENAME;
625         diff_opts.output_format = DIFF_FORMAT_NO_OUTPUT;
626         diff_opts.single_follow = origin->path;
627         diff_setup_done(&diff_opts);
628
629         if (is_null_sha1(origin->commit->object.sha1))
630                 do_diff_cache(parent->tree->object.sha1, &diff_opts);
631         else
632                 diff_tree_sha1(parent->tree->object.sha1,
633                                origin->commit->tree->object.sha1,
634                                "", &diff_opts);
635         diffcore_std(&diff_opts);
636
637         for (i = 0; i < diff_queued_diff.nr; i++) {
638                 struct diff_filepair *p = diff_queued_diff.queue[i];
639                 if ((p->status == 'R' || p->status == 'C') &&
640                     !strcmp(p->two->path, origin->path)) {
641                         porigin = get_origin(sb, parent, p->one->path);
642                         hashcpy(porigin->blob_sha1, p->one->sha1);
643                         porigin->mode = p->one->mode;
644                         break;
645                 }
646         }
647         diff_flush(&diff_opts);
648         free_pathspec(&diff_opts.pathspec);
649         return porigin;
650 }
651
652 /*
653  * Append a new blame entry to a given output queue.
654  */
655 static void add_blame_entry(struct blame_entry ***queue, struct blame_entry *e)
656 {
657         origin_incref(e->suspect);
658
659         e->next = **queue;
660         **queue = e;
661         *queue = &e->next;
662 }
663
664 /*
665  * src typically is on-stack; we want to copy the information in it to
666  * a malloced blame_entry that gets added to the given queue.  The
667  * origin of dst loses a refcnt.
668  */
669 static void dup_entry(struct blame_entry ***queue,
670                       struct blame_entry *dst, struct blame_entry *src)
671 {
672         origin_incref(src->suspect);
673         origin_decref(dst->suspect);
674         memcpy(dst, src, sizeof(*src));
675         dst->next = **queue;
676         **queue = dst;
677         *queue = &dst->next;
678 }
679
680 static const char *nth_line(struct scoreboard *sb, long lno)
681 {
682         return sb->final_buf + sb->lineno[lno];
683 }
684
685 static const char *nth_line_cb(void *data, long lno)
686 {
687         return nth_line((struct scoreboard *)data, lno);
688 }
689
690 /*
691  * It is known that lines between tlno to same came from parent, and e
692  * has an overlap with that range.  it also is known that parent's
693  * line plno corresponds to e's line tlno.
694  *
695  *                <---- e ----->
696  *                   <------>
697  *                   <------------>
698  *             <------------>
699  *             <------------------>
700  *
701  * Split e into potentially three parts; before this chunk, the chunk
702  * to be blamed for the parent, and after that portion.
703  */
704 static void split_overlap(struct blame_entry *split,
705                           struct blame_entry *e,
706                           int tlno, int plno, int same,
707                           struct origin *parent)
708 {
709         int chunk_end_lno;
710         memset(split, 0, sizeof(struct blame_entry [3]));
711
712         if (e->s_lno < tlno) {
713                 /* there is a pre-chunk part not blamed on parent */
714                 split[0].suspect = origin_incref(e->suspect);
715                 split[0].lno = e->lno;
716                 split[0].s_lno = e->s_lno;
717                 split[0].num_lines = tlno - e->s_lno;
718                 split[1].lno = e->lno + tlno - e->s_lno;
719                 split[1].s_lno = plno;
720         }
721         else {
722                 split[1].lno = e->lno;
723                 split[1].s_lno = plno + (e->s_lno - tlno);
724         }
725
726         if (same < e->s_lno + e->num_lines) {
727                 /* there is a post-chunk part not blamed on parent */
728                 split[2].suspect = origin_incref(e->suspect);
729                 split[2].lno = e->lno + (same - e->s_lno);
730                 split[2].s_lno = e->s_lno + (same - e->s_lno);
731                 split[2].num_lines = e->s_lno + e->num_lines - same;
732                 chunk_end_lno = split[2].lno;
733         }
734         else
735                 chunk_end_lno = e->lno + e->num_lines;
736         split[1].num_lines = chunk_end_lno - split[1].lno;
737
738         /*
739          * if it turns out there is nothing to blame the parent for,
740          * forget about the splitting.  !split[1].suspect signals this.
741          */
742         if (split[1].num_lines < 1)
743                 return;
744         split[1].suspect = origin_incref(parent);
745 }
746
747 /*
748  * split_overlap() divided an existing blame e into up to three parts
749  * in split.  Any assigned blame is moved to queue to
750  * reflect the split.
751  */
752 static void split_blame(struct blame_entry ***blamed,
753                         struct blame_entry ***unblamed,
754                         struct blame_entry *split,
755                         struct blame_entry *e)
756 {
757         struct blame_entry *new_entry;
758
759         if (split[0].suspect && split[2].suspect) {
760                 /* The first part (reuse storage for the existing entry e) */
761                 dup_entry(unblamed, e, &split[0]);
762
763                 /* The last part -- me */
764                 new_entry = xmalloc(sizeof(*new_entry));
765                 memcpy(new_entry, &(split[2]), sizeof(struct blame_entry));
766                 add_blame_entry(unblamed, new_entry);
767
768                 /* ... and the middle part -- parent */
769                 new_entry = xmalloc(sizeof(*new_entry));
770                 memcpy(new_entry, &(split[1]), sizeof(struct blame_entry));
771                 add_blame_entry(blamed, new_entry);
772         }
773         else if (!split[0].suspect && !split[2].suspect)
774                 /*
775                  * The parent covers the entire area; reuse storage for
776                  * e and replace it with the parent.
777                  */
778                 dup_entry(blamed, e, &split[1]);
779         else if (split[0].suspect) {
780                 /* me and then parent */
781                 dup_entry(unblamed, e, &split[0]);
782
783                 new_entry = xmalloc(sizeof(*new_entry));
784                 memcpy(new_entry, &(split[1]), sizeof(struct blame_entry));
785                 add_blame_entry(blamed, new_entry);
786         }
787         else {
788                 /* parent and then me */
789                 dup_entry(blamed, e, &split[1]);
790
791                 new_entry = xmalloc(sizeof(*new_entry));
792                 memcpy(new_entry, &(split[2]), sizeof(struct blame_entry));
793                 add_blame_entry(unblamed, new_entry);
794         }
795 }
796
797 /*
798  * After splitting the blame, the origins used by the
799  * on-stack blame_entry should lose one refcnt each.
800  */
801 static void decref_split(struct blame_entry *split)
802 {
803         int i;
804
805         for (i = 0; i < 3; i++)
806                 origin_decref(split[i].suspect);
807 }
808
809 /*
810  * reverse_blame reverses the list given in head, appending tail.
811  * That allows us to build lists in reverse order, then reverse them
812  * afterwards.  This can be faster than building the list in proper
813  * order right away.  The reason is that building in proper order
814  * requires writing a link in the _previous_ element, while building
815  * in reverse order just requires placing the list head into the
816  * _current_ element.
817  */
818
819 static struct blame_entry *reverse_blame(struct blame_entry *head,
820                                          struct blame_entry *tail)
821 {
822         while (head) {
823                 struct blame_entry *next = head->next;
824                 head->next = tail;
825                 tail = head;
826                 head = next;
827         }
828         return tail;
829 }
830
831 /*
832  * Process one hunk from the patch between the current suspect for
833  * blame_entry e and its parent.  This first blames any unfinished
834  * entries before the chunk (which is where target and parent start
835  * differing) on the parent, and then splits blame entries at the
836  * start and at the end of the difference region.  Since use of -M and
837  * -C options may lead to overlapping/duplicate source line number
838  * ranges, all we can rely on from sorting/merging is the order of the
839  * first suspect line number.
840  */
841 static void blame_chunk(struct blame_entry ***dstq, struct blame_entry ***srcq,
842                         int tlno, int offset, int same,
843                         struct origin *parent)
844 {
845         struct blame_entry *e = **srcq;
846         struct blame_entry *samep = NULL, *diffp = NULL;
847
848         while (e && e->s_lno < tlno) {
849                 struct blame_entry *next = e->next;
850                 /*
851                  * current record starts before differing portion.  If
852                  * it reaches into it, we need to split it up and
853                  * examine the second part separately.
854                  */
855                 if (e->s_lno + e->num_lines > tlno) {
856                         /* Move second half to a new record */
857                         int len = tlno - e->s_lno;
858                         struct blame_entry *n = xcalloc(1, sizeof (struct blame_entry));
859                         n->suspect = e->suspect;
860                         n->lno = e->lno + len;
861                         n->s_lno = e->s_lno + len;
862                         n->num_lines = e->num_lines - len;
863                         e->num_lines = len;
864                         e->score = 0;
865                         /* Push new record to diffp */
866                         n->next = diffp;
867                         diffp = n;
868                 } else
869                         origin_decref(e->suspect);
870                 /* Pass blame for everything before the differing
871                  * chunk to the parent */
872                 e->suspect = origin_incref(parent);
873                 e->s_lno += offset;
874                 e->next = samep;
875                 samep = e;
876                 e = next;
877         }
878         /*
879          * As we don't know how much of a common stretch after this
880          * diff will occur, the currently blamed parts are all that we
881          * can assign to the parent for now.
882          */
883
884         if (samep) {
885                 **dstq = reverse_blame(samep, **dstq);
886                 *dstq = &samep->next;
887         }
888         /*
889          * Prepend the split off portions: everything after e starts
890          * after the blameable portion.
891          */
892         e = reverse_blame(diffp, e);
893
894         /*
895          * Now retain records on the target while parts are different
896          * from the parent.
897          */
898         samep = NULL;
899         diffp = NULL;
900         while (e && e->s_lno < same) {
901                 struct blame_entry *next = e->next;
902
903                 /*
904                  * If current record extends into sameness, need to split.
905                  */
906                 if (e->s_lno + e->num_lines > same) {
907                         /*
908                          * Move second half to a new record to be
909                          * processed by later chunks
910                          */
911                         int len = same - e->s_lno;
912                         struct blame_entry *n = xcalloc(1, sizeof (struct blame_entry));
913                         n->suspect = origin_incref(e->suspect);
914                         n->lno = e->lno + len;
915                         n->s_lno = e->s_lno + len;
916                         n->num_lines = e->num_lines - len;
917                         e->num_lines = len;
918                         e->score = 0;
919                         /* Push new record to samep */
920                         n->next = samep;
921                         samep = n;
922                 }
923                 e->next = diffp;
924                 diffp = e;
925                 e = next;
926         }
927         **srcq = reverse_blame(diffp, reverse_blame(samep, e));
928         /* Move across elements that are in the unblamable portion */
929         if (diffp)
930                 *srcq = &diffp->next;
931 }
932
933 struct blame_chunk_cb_data {
934         struct origin *parent;
935         long offset;
936         struct blame_entry **dstq;
937         struct blame_entry **srcq;
938 };
939
940 /* diff chunks are from parent to target */
941 static int blame_chunk_cb(long start_a, long count_a,
942                           long start_b, long count_b, void *data)
943 {
944         struct blame_chunk_cb_data *d = data;
945         if (start_a - start_b != d->offset)
946                 die("internal error in blame::blame_chunk_cb");
947         blame_chunk(&d->dstq, &d->srcq, start_b, start_a - start_b,
948                     start_b + count_b, d->parent);
949         d->offset = start_a + count_a - (start_b + count_b);
950         return 0;
951 }
952
953 /*
954  * We are looking at the origin 'target' and aiming to pass blame
955  * for the lines it is suspected to its parent.  Run diff to find
956  * which lines came from parent and pass blame for them.
957  */
958 static void pass_blame_to_parent(struct scoreboard *sb,
959                                  struct origin *target,
960                                  struct origin *parent)
961 {
962         mmfile_t file_p, file_o;
963         struct blame_chunk_cb_data d;
964         struct blame_entry *newdest = NULL;
965
966         if (!target->suspects)
967                 return; /* nothing remains for this target */
968
969         d.parent = parent;
970         d.offset = 0;
971         d.dstq = &newdest; d.srcq = &target->suspects;
972
973         fill_origin_blob(&sb->revs->diffopt, parent, &file_p);
974         fill_origin_blob(&sb->revs->diffopt, target, &file_o);
975         num_get_patch++;
976
977         diff_hunks(&file_p, &file_o, 0, blame_chunk_cb, &d);
978         /* The rest are the same as the parent */
979         blame_chunk(&d.dstq, &d.srcq, INT_MAX, d.offset, INT_MAX, parent);
980         *d.dstq = NULL;
981         queue_blames(sb, parent, newdest);
982
983         return;
984 }
985
986 /*
987  * The lines in blame_entry after splitting blames many times can become
988  * very small and trivial, and at some point it becomes pointless to
989  * blame the parents.  E.g. "\t\t}\n\t}\n\n" appears everywhere in any
990  * ordinary C program, and it is not worth to say it was copied from
991  * totally unrelated file in the parent.
992  *
993  * Compute how trivial the lines in the blame_entry are.
994  */
995 static unsigned ent_score(struct scoreboard *sb, struct blame_entry *e)
996 {
997         unsigned score;
998         const char *cp, *ep;
999
1000         if (e->score)
1001                 return e->score;
1002
1003         score = 1;
1004         cp = nth_line(sb, e->lno);
1005         ep = nth_line(sb, e->lno + e->num_lines);
1006         while (cp < ep) {
1007                 unsigned ch = *((unsigned char *)cp);
1008                 if (isalnum(ch))
1009                         score++;
1010                 cp++;
1011         }
1012         e->score = score;
1013         return score;
1014 }
1015
1016 /*
1017  * best_so_far[] and this[] are both a split of an existing blame_entry
1018  * that passes blame to the parent.  Maintain best_so_far the best split
1019  * so far, by comparing this and best_so_far and copying this into
1020  * bst_so_far as needed.
1021  */
1022 static void copy_split_if_better(struct scoreboard *sb,
1023                                  struct blame_entry *best_so_far,
1024                                  struct blame_entry *this)
1025 {
1026         int i;
1027
1028         if (!this[1].suspect)
1029                 return;
1030         if (best_so_far[1].suspect) {
1031                 if (ent_score(sb, &this[1]) < ent_score(sb, &best_so_far[1]))
1032                         return;
1033         }
1034
1035         for (i = 0; i < 3; i++)
1036                 origin_incref(this[i].suspect);
1037         decref_split(best_so_far);
1038         memcpy(best_so_far, this, sizeof(struct blame_entry [3]));
1039 }
1040
1041 /*
1042  * We are looking at a part of the final image represented by
1043  * ent (tlno and same are offset by ent->s_lno).
1044  * tlno is where we are looking at in the final image.
1045  * up to (but not including) same match preimage.
1046  * plno is where we are looking at in the preimage.
1047  *
1048  * <-------------- final image ---------------------->
1049  *       <------ent------>
1050  *         ^tlno ^same
1051  *    <---------preimage----->
1052  *         ^plno
1053  *
1054  * All line numbers are 0-based.
1055  */
1056 static void handle_split(struct scoreboard *sb,
1057                          struct blame_entry *ent,
1058                          int tlno, int plno, int same,
1059                          struct origin *parent,
1060                          struct blame_entry *split)
1061 {
1062         if (ent->num_lines <= tlno)
1063                 return;
1064         if (tlno < same) {
1065                 struct blame_entry this[3];
1066                 tlno += ent->s_lno;
1067                 same += ent->s_lno;
1068                 split_overlap(this, ent, tlno, plno, same, parent);
1069                 copy_split_if_better(sb, split, this);
1070                 decref_split(this);
1071         }
1072 }
1073
1074 struct handle_split_cb_data {
1075         struct scoreboard *sb;
1076         struct blame_entry *ent;
1077         struct origin *parent;
1078         struct blame_entry *split;
1079         long plno;
1080         long tlno;
1081 };
1082
1083 static int handle_split_cb(long start_a, long count_a,
1084                            long start_b, long count_b, void *data)
1085 {
1086         struct handle_split_cb_data *d = data;
1087         handle_split(d->sb, d->ent, d->tlno, d->plno, start_b, d->parent,
1088                      d->split);
1089         d->plno = start_a + count_a;
1090         d->tlno = start_b + count_b;
1091         return 0;
1092 }
1093
1094 /*
1095  * Find the lines from parent that are the same as ent so that
1096  * we can pass blames to it.  file_p has the blob contents for
1097  * the parent.
1098  */
1099 static void find_copy_in_blob(struct scoreboard *sb,
1100                               struct blame_entry *ent,
1101                               struct origin *parent,
1102                               struct blame_entry *split,
1103                               mmfile_t *file_p)
1104 {
1105         const char *cp;
1106         mmfile_t file_o;
1107         struct handle_split_cb_data d;
1108
1109         memset(&d, 0, sizeof(d));
1110         d.sb = sb; d.ent = ent; d.parent = parent; d.split = split;
1111         /*
1112          * Prepare mmfile that contains only the lines in ent.
1113          */
1114         cp = nth_line(sb, ent->lno);
1115         file_o.ptr = (char *) cp;
1116         file_o.size = nth_line(sb, ent->lno + ent->num_lines) - cp;
1117
1118         /*
1119          * file_o is a part of final image we are annotating.
1120          * file_p partially may match that image.
1121          */
1122         memset(split, 0, sizeof(struct blame_entry [3]));
1123         diff_hunks(file_p, &file_o, 1, handle_split_cb, &d);
1124         /* remainder, if any, all match the preimage */
1125         handle_split(sb, ent, d.tlno, d.plno, ent->num_lines, parent, split);
1126 }
1127
1128 /* Move all blame entries from list *source that have a score smaller
1129  * than score_min to the front of list *small.
1130  * Returns a pointer to the link pointing to the old head of the small list.
1131  */
1132
1133 static struct blame_entry **filter_small(struct scoreboard *sb,
1134                                          struct blame_entry **small,
1135                                          struct blame_entry **source,
1136                                          unsigned score_min)
1137 {
1138         struct blame_entry *p = *source;
1139         struct blame_entry *oldsmall = *small;
1140         while (p) {
1141                 if (ent_score(sb, p) <= score_min) {
1142                         *small = p;
1143                         small = &p->next;
1144                         p = *small;
1145                 } else {
1146                         *source = p;
1147                         source = &p->next;
1148                         p = *source;
1149                 }
1150         }
1151         *small = oldsmall;
1152         *source = NULL;
1153         return small;
1154 }
1155
1156 /*
1157  * See if lines currently target is suspected for can be attributed to
1158  * parent.
1159  */
1160 static void find_move_in_parent(struct scoreboard *sb,
1161                                 struct blame_entry ***blamed,
1162                                 struct blame_entry **toosmall,
1163                                 struct origin *target,
1164                                 struct origin *parent)
1165 {
1166         struct blame_entry *e, split[3];
1167         struct blame_entry *unblamed = target->suspects;
1168         struct blame_entry *leftover = NULL;
1169         mmfile_t file_p;
1170
1171         if (!unblamed)
1172                 return; /* nothing remains for this target */
1173
1174         fill_origin_blob(&sb->revs->diffopt, parent, &file_p);
1175         if (!file_p.ptr)
1176                 return;
1177
1178         /* At each iteration, unblamed has a NULL-terminated list of
1179          * entries that have not yet been tested for blame.  leftover
1180          * contains the reversed list of entries that have been tested
1181          * without being assignable to the parent.
1182          */
1183         do {
1184                 struct blame_entry **unblamedtail = &unblamed;
1185                 struct blame_entry *next;
1186                 for (e = unblamed; e; e = next) {
1187                         next = e->next;
1188                         find_copy_in_blob(sb, e, parent, split, &file_p);
1189                         if (split[1].suspect &&
1190                             blame_move_score < ent_score(sb, &split[1])) {
1191                                 split_blame(blamed, &unblamedtail, split, e);
1192                         } else {
1193                                 e->next = leftover;
1194                                 leftover = e;
1195                         }
1196                         decref_split(split);
1197                 }
1198                 *unblamedtail = NULL;
1199                 toosmall = filter_small(sb, toosmall, &unblamed, blame_move_score);
1200         } while (unblamed);
1201         target->suspects = reverse_blame(leftover, NULL);
1202 }
1203
1204 struct blame_list {
1205         struct blame_entry *ent;
1206         struct blame_entry split[3];
1207 };
1208
1209 /*
1210  * Count the number of entries the target is suspected for,
1211  * and prepare a list of entry and the best split.
1212  */
1213 static struct blame_list *setup_blame_list(struct blame_entry *unblamed,
1214                                            int *num_ents_p)
1215 {
1216         struct blame_entry *e;
1217         int num_ents, i;
1218         struct blame_list *blame_list = NULL;
1219
1220         for (e = unblamed, num_ents = 0; e; e = e->next)
1221                 num_ents++;
1222         if (num_ents) {
1223                 blame_list = xcalloc(num_ents, sizeof(struct blame_list));
1224                 for (e = unblamed, i = 0; e; e = e->next)
1225                         blame_list[i++].ent = e;
1226         }
1227         *num_ents_p = num_ents;
1228         return blame_list;
1229 }
1230
1231 /*
1232  * For lines target is suspected for, see if we can find code movement
1233  * across file boundary from the parent commit.  porigin is the path
1234  * in the parent we already tried.
1235  */
1236 static void find_copy_in_parent(struct scoreboard *sb,
1237                                 struct blame_entry ***blamed,
1238                                 struct blame_entry **toosmall,
1239                                 struct origin *target,
1240                                 struct commit *parent,
1241                                 struct origin *porigin,
1242                                 int opt)
1243 {
1244         struct diff_options diff_opts;
1245         int i, j;
1246         struct blame_list *blame_list;
1247         int num_ents;
1248         struct blame_entry *unblamed = target->suspects;
1249         struct blame_entry *leftover = NULL;
1250
1251         if (!unblamed)
1252                 return; /* nothing remains for this target */
1253
1254         diff_setup(&diff_opts);
1255         DIFF_OPT_SET(&diff_opts, RECURSIVE);
1256         diff_opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1257
1258         diff_setup_done(&diff_opts);
1259
1260         /* Try "find copies harder" on new path if requested;
1261          * we do not want to use diffcore_rename() actually to
1262          * match things up; find_copies_harder is set only to
1263          * force diff_tree_sha1() to feed all filepairs to diff_queue,
1264          * and this code needs to be after diff_setup_done(), which
1265          * usually makes find-copies-harder imply copy detection.
1266          */
1267         if ((opt & PICKAXE_BLAME_COPY_HARDEST)
1268             || ((opt & PICKAXE_BLAME_COPY_HARDER)
1269                 && (!porigin || strcmp(target->path, porigin->path))))
1270                 DIFF_OPT_SET(&diff_opts, FIND_COPIES_HARDER);
1271
1272         if (is_null_sha1(target->commit->object.sha1))
1273                 do_diff_cache(parent->tree->object.sha1, &diff_opts);
1274         else
1275                 diff_tree_sha1(parent->tree->object.sha1,
1276                                target->commit->tree->object.sha1,
1277                                "", &diff_opts);
1278
1279         if (!DIFF_OPT_TST(&diff_opts, FIND_COPIES_HARDER))
1280                 diffcore_std(&diff_opts);
1281
1282         do {
1283                 struct blame_entry **unblamedtail = &unblamed;
1284                 blame_list = setup_blame_list(unblamed, &num_ents);
1285
1286                 for (i = 0; i < diff_queued_diff.nr; i++) {
1287                         struct diff_filepair *p = diff_queued_diff.queue[i];
1288                         struct origin *norigin;
1289                         mmfile_t file_p;
1290                         struct blame_entry this[3];
1291
1292                         if (!DIFF_FILE_VALID(p->one))
1293                                 continue; /* does not exist in parent */
1294                         if (S_ISGITLINK(p->one->mode))
1295                                 continue; /* ignore git links */
1296                         if (porigin && !strcmp(p->one->path, porigin->path))
1297                                 /* find_move already dealt with this path */
1298                                 continue;
1299
1300                         norigin = get_origin(sb, parent, p->one->path);
1301                         hashcpy(norigin->blob_sha1, p->one->sha1);
1302                         norigin->mode = p->one->mode;
1303                         fill_origin_blob(&sb->revs->diffopt, norigin, &file_p);
1304                         if (!file_p.ptr)
1305                                 continue;
1306
1307                         for (j = 0; j < num_ents; j++) {
1308                                 find_copy_in_blob(sb, blame_list[j].ent,
1309                                                   norigin, this, &file_p);
1310                                 copy_split_if_better(sb, blame_list[j].split,
1311                                                      this);
1312                                 decref_split(this);
1313                         }
1314                         origin_decref(norigin);
1315                 }
1316
1317                 for (j = 0; j < num_ents; j++) {
1318                         struct blame_entry *split = blame_list[j].split;
1319                         if (split[1].suspect &&
1320                             blame_copy_score < ent_score(sb, &split[1])) {
1321                                 split_blame(blamed, &unblamedtail, split,
1322                                             blame_list[j].ent);
1323                         } else {
1324                                 blame_list[j].ent->next = leftover;
1325                                 leftover = blame_list[j].ent;
1326                         }
1327                         decref_split(split);
1328                 }
1329                 free(blame_list);
1330                 *unblamedtail = NULL;
1331                 toosmall = filter_small(sb, toosmall, &unblamed, blame_copy_score);
1332         } while (unblamed);
1333         target->suspects = reverse_blame(leftover, NULL);
1334         diff_flush(&diff_opts);
1335         free_pathspec(&diff_opts.pathspec);
1336 }
1337
1338 /*
1339  * The blobs of origin and porigin exactly match, so everything
1340  * origin is suspected for can be blamed on the parent.
1341  */
1342 static void pass_whole_blame(struct scoreboard *sb,
1343                              struct origin *origin, struct origin *porigin)
1344 {
1345         struct blame_entry *e, *suspects;
1346
1347         if (!porigin->file.ptr && origin->file.ptr) {
1348                 /* Steal its file */
1349                 porigin->file = origin->file;
1350                 origin->file.ptr = NULL;
1351         }
1352         suspects = origin->suspects;
1353         origin->suspects = NULL;
1354         for (e = suspects; e; e = e->next) {
1355                 origin_incref(porigin);
1356                 origin_decref(e->suspect);
1357                 e->suspect = porigin;
1358         }
1359         queue_blames(sb, porigin, suspects);
1360 }
1361
1362 /*
1363  * We pass blame from the current commit to its parents.  We keep saying
1364  * "parent" (and "porigin"), but what we mean is to find scapegoat to
1365  * exonerate ourselves.
1366  */
1367 static struct commit_list *first_scapegoat(struct rev_info *revs, struct commit *commit)
1368 {
1369         if (!reverse)
1370                 return commit->parents;
1371         return lookup_decoration(&revs->children, &commit->object);
1372 }
1373
1374 static int num_scapegoats(struct rev_info *revs, struct commit *commit)
1375 {
1376         struct commit_list *l = first_scapegoat(revs, commit);
1377         return commit_list_count(l);
1378 }
1379
1380 /* Distribute collected unsorted blames to the respected sorted lists
1381  * in the various origins.
1382  */
1383 static void distribute_blame(struct scoreboard *sb, struct blame_entry *blamed)
1384 {
1385         blamed = blame_sort(blamed, compare_blame_suspect);
1386         while (blamed)
1387         {
1388                 struct origin *porigin = blamed->suspect;
1389                 struct blame_entry *suspects = NULL;
1390                 do {
1391                         struct blame_entry *next = blamed->next;
1392                         blamed->next = suspects;
1393                         suspects = blamed;
1394                         blamed = next;
1395                 } while (blamed && blamed->suspect == porigin);
1396                 suspects = reverse_blame(suspects, NULL);
1397                 queue_blames(sb, porigin, suspects);
1398         }
1399 }
1400
1401 #define MAXSG 16
1402
1403 static void pass_blame(struct scoreboard *sb, struct origin *origin, int opt)
1404 {
1405         struct rev_info *revs = sb->revs;
1406         int i, pass, num_sg;
1407         struct commit *commit = origin->commit;
1408         struct commit_list *sg;
1409         struct origin *sg_buf[MAXSG];
1410         struct origin *porigin, **sg_origin = sg_buf;
1411         struct blame_entry *toosmall = NULL;
1412         struct blame_entry *blames, **blametail = &blames;
1413
1414         num_sg = num_scapegoats(revs, commit);
1415         if (!num_sg)
1416                 goto finish;
1417         else if (num_sg < ARRAY_SIZE(sg_buf))
1418                 memset(sg_buf, 0, sizeof(sg_buf));
1419         else
1420                 sg_origin = xcalloc(num_sg, sizeof(*sg_origin));
1421
1422         /*
1423          * The first pass looks for unrenamed path to optimize for
1424          * common cases, then we look for renames in the second pass.
1425          */
1426         for (pass = 0; pass < 2 - no_whole_file_rename; pass++) {
1427                 struct origin *(*find)(struct scoreboard *,
1428                                        struct commit *, struct origin *);
1429                 find = pass ? find_rename : find_origin;
1430
1431                 for (i = 0, sg = first_scapegoat(revs, commit);
1432                      i < num_sg && sg;
1433                      sg = sg->next, i++) {
1434                         struct commit *p = sg->item;
1435                         int j, same;
1436
1437                         if (sg_origin[i])
1438                                 continue;
1439                         if (parse_commit(p))
1440                                 continue;
1441                         porigin = find(sb, p, origin);
1442                         if (!porigin)
1443                                 continue;
1444                         if (!hashcmp(porigin->blob_sha1, origin->blob_sha1)) {
1445                                 pass_whole_blame(sb, origin, porigin);
1446                                 origin_decref(porigin);
1447                                 goto finish;
1448                         }
1449                         for (j = same = 0; j < i; j++)
1450                                 if (sg_origin[j] &&
1451                                     !hashcmp(sg_origin[j]->blob_sha1,
1452                                              porigin->blob_sha1)) {
1453                                         same = 1;
1454                                         break;
1455                                 }
1456                         if (!same)
1457                                 sg_origin[i] = porigin;
1458                         else
1459                                 origin_decref(porigin);
1460                 }
1461         }
1462
1463         num_commits++;
1464         for (i = 0, sg = first_scapegoat(revs, commit);
1465              i < num_sg && sg;
1466              sg = sg->next, i++) {
1467                 struct origin *porigin = sg_origin[i];
1468                 if (!porigin)
1469                         continue;
1470                 if (!origin->previous) {
1471                         origin_incref(porigin);
1472                         origin->previous = porigin;
1473                 }
1474                 pass_blame_to_parent(sb, origin, porigin);
1475                 if (!origin->suspects)
1476                         goto finish;
1477         }
1478
1479         /*
1480          * Optionally find moves in parents' files.
1481          */
1482         if (opt & PICKAXE_BLAME_MOVE) {
1483                 filter_small(sb, &toosmall, &origin->suspects, blame_move_score);
1484                 if (origin->suspects) {
1485                         for (i = 0, sg = first_scapegoat(revs, commit);
1486                              i < num_sg && sg;
1487                              sg = sg->next, i++) {
1488                                 struct origin *porigin = sg_origin[i];
1489                                 if (!porigin)
1490                                         continue;
1491                                 find_move_in_parent(sb, &blametail, &toosmall, origin, porigin);
1492                                 if (!origin->suspects)
1493                                         break;
1494                         }
1495                 }
1496         }
1497
1498         /*
1499          * Optionally find copies from parents' files.
1500          */
1501         if (opt & PICKAXE_BLAME_COPY) {
1502                 if (blame_copy_score > blame_move_score)
1503                         filter_small(sb, &toosmall, &origin->suspects, blame_copy_score);
1504                 else if (blame_copy_score < blame_move_score) {
1505                         origin->suspects = blame_merge(origin->suspects, toosmall);
1506                         toosmall = NULL;
1507                         filter_small(sb, &toosmall, &origin->suspects, blame_copy_score);
1508                 }
1509                 if (!origin->suspects)
1510                         goto finish;
1511
1512                 for (i = 0, sg = first_scapegoat(revs, commit);
1513                      i < num_sg && sg;
1514                      sg = sg->next, i++) {
1515                         struct origin *porigin = sg_origin[i];
1516                         find_copy_in_parent(sb, &blametail, &toosmall,
1517                                             origin, sg->item, porigin, opt);
1518                         if (!origin->suspects)
1519                                 goto finish;
1520                 }
1521         }
1522
1523 finish:
1524         *blametail = NULL;
1525         distribute_blame(sb, blames);
1526         /*
1527          * prepend toosmall to origin->suspects
1528          *
1529          * There is no point in sorting: this ends up on a big
1530          * unsorted list in the caller anyway.
1531          */
1532         if (toosmall) {
1533                 struct blame_entry **tail = &toosmall;
1534                 while (*tail)
1535                         tail = &(*tail)->next;
1536                 *tail = origin->suspects;
1537                 origin->suspects = toosmall;
1538         }
1539         for (i = 0; i < num_sg; i++) {
1540                 if (sg_origin[i]) {
1541                         drop_origin_blob(sg_origin[i]);
1542                         origin_decref(sg_origin[i]);
1543                 }
1544         }
1545         drop_origin_blob(origin);
1546         if (sg_buf != sg_origin)
1547                 free(sg_origin);
1548 }
1549
1550 /*
1551  * Information on commits, used for output.
1552  */
1553 struct commit_info {
1554         struct strbuf author;
1555         struct strbuf author_mail;
1556         unsigned long author_time;
1557         struct strbuf author_tz;
1558
1559         /* filled only when asked for details */
1560         struct strbuf committer;
1561         struct strbuf committer_mail;
1562         unsigned long committer_time;
1563         struct strbuf committer_tz;
1564
1565         struct strbuf summary;
1566 };
1567
1568 /*
1569  * Parse author/committer line in the commit object buffer
1570  */
1571 static void get_ac_line(const char *inbuf, const char *what,
1572         struct strbuf *name, struct strbuf *mail,
1573         unsigned long *time, struct strbuf *tz)
1574 {
1575         struct ident_split ident;
1576         size_t len, maillen, namelen;
1577         char *tmp, *endp;
1578         const char *namebuf, *mailbuf;
1579
1580         tmp = strstr(inbuf, what);
1581         if (!tmp)
1582                 goto error_out;
1583         tmp += strlen(what);
1584         endp = strchr(tmp, '\n');
1585         if (!endp)
1586                 len = strlen(tmp);
1587         else
1588                 len = endp - tmp;
1589
1590         if (split_ident_line(&ident, tmp, len)) {
1591         error_out:
1592                 /* Ugh */
1593                 tmp = "(unknown)";
1594                 strbuf_addstr(name, tmp);
1595                 strbuf_addstr(mail, tmp);
1596                 strbuf_addstr(tz, tmp);
1597                 *time = 0;
1598                 return;
1599         }
1600
1601         namelen = ident.name_end - ident.name_begin;
1602         namebuf = ident.name_begin;
1603
1604         maillen = ident.mail_end - ident.mail_begin;
1605         mailbuf = ident.mail_begin;
1606
1607         if (ident.date_begin && ident.date_end)
1608                 *time = strtoul(ident.date_begin, NULL, 10);
1609         else
1610                 *time = 0;
1611
1612         if (ident.tz_begin && ident.tz_end)
1613                 strbuf_add(tz, ident.tz_begin, ident.tz_end - ident.tz_begin);
1614         else
1615                 strbuf_addstr(tz, "(unknown)");
1616
1617         /*
1618          * Now, convert both name and e-mail using mailmap
1619          */
1620         map_user(&mailmap, &mailbuf, &maillen,
1621                  &namebuf, &namelen);
1622
1623         strbuf_addf(mail, "<%.*s>", (int)maillen, mailbuf);
1624         strbuf_add(name, namebuf, namelen);
1625 }
1626
1627 static void commit_info_init(struct commit_info *ci)
1628 {
1629
1630         strbuf_init(&ci->author, 0);
1631         strbuf_init(&ci->author_mail, 0);
1632         strbuf_init(&ci->author_tz, 0);
1633         strbuf_init(&ci->committer, 0);
1634         strbuf_init(&ci->committer_mail, 0);
1635         strbuf_init(&ci->committer_tz, 0);
1636         strbuf_init(&ci->summary, 0);
1637 }
1638
1639 static void commit_info_destroy(struct commit_info *ci)
1640 {
1641
1642         strbuf_release(&ci->author);
1643         strbuf_release(&ci->author_mail);
1644         strbuf_release(&ci->author_tz);
1645         strbuf_release(&ci->committer);
1646         strbuf_release(&ci->committer_mail);
1647         strbuf_release(&ci->committer_tz);
1648         strbuf_release(&ci->summary);
1649 }
1650
1651 static void get_commit_info(struct commit *commit,
1652                             struct commit_info *ret,
1653                             int detailed)
1654 {
1655         int len;
1656         const char *subject, *encoding;
1657         const char *message;
1658
1659         commit_info_init(ret);
1660
1661         encoding = get_log_output_encoding();
1662         message = logmsg_reencode(commit, NULL, encoding);
1663         get_ac_line(message, "\nauthor ",
1664                     &ret->author, &ret->author_mail,
1665                     &ret->author_time, &ret->author_tz);
1666
1667         if (!detailed) {
1668                 unuse_commit_buffer(commit, message);
1669                 return;
1670         }
1671
1672         get_ac_line(message, "\ncommitter ",
1673                     &ret->committer, &ret->committer_mail,
1674                     &ret->committer_time, &ret->committer_tz);
1675
1676         len = find_commit_subject(message, &subject);
1677         if (len)
1678                 strbuf_add(&ret->summary, subject, len);
1679         else
1680                 strbuf_addf(&ret->summary, "(%s)", sha1_to_hex(commit->object.sha1));
1681
1682         unuse_commit_buffer(commit, message);
1683 }
1684
1685 /*
1686  * To allow LF and other nonportable characters in pathnames,
1687  * they are c-style quoted as needed.
1688  */
1689 static void write_filename_info(const char *path)
1690 {
1691         printf("filename ");
1692         write_name_quoted(path, stdout, '\n');
1693 }
1694
1695 /*
1696  * Porcelain/Incremental format wants to show a lot of details per
1697  * commit.  Instead of repeating this every line, emit it only once,
1698  * the first time each commit appears in the output (unless the
1699  * user has specifically asked for us to repeat).
1700  */
1701 static int emit_one_suspect_detail(struct origin *suspect, int repeat)
1702 {
1703         struct commit_info ci;
1704
1705         if (!repeat && (suspect->commit->object.flags & METAINFO_SHOWN))
1706                 return 0;
1707
1708         suspect->commit->object.flags |= METAINFO_SHOWN;
1709         get_commit_info(suspect->commit, &ci, 1);
1710         printf("author %s\n", ci.author.buf);
1711         printf("author-mail %s\n", ci.author_mail.buf);
1712         printf("author-time %lu\n", ci.author_time);
1713         printf("author-tz %s\n", ci.author_tz.buf);
1714         printf("committer %s\n", ci.committer.buf);
1715         printf("committer-mail %s\n", ci.committer_mail.buf);
1716         printf("committer-time %lu\n", ci.committer_time);
1717         printf("committer-tz %s\n", ci.committer_tz.buf);
1718         printf("summary %s\n", ci.summary.buf);
1719         if (suspect->commit->object.flags & UNINTERESTING)
1720                 printf("boundary\n");
1721         if (suspect->previous) {
1722                 struct origin *prev = suspect->previous;
1723                 printf("previous %s ", sha1_to_hex(prev->commit->object.sha1));
1724                 write_name_quoted(prev->path, stdout, '\n');
1725         }
1726
1727         commit_info_destroy(&ci);
1728
1729         return 1;
1730 }
1731
1732 /*
1733  * The blame_entry is found to be guilty for the range.
1734  * Show it in incremental output.
1735  */
1736 static void found_guilty_entry(struct blame_entry *ent)
1737 {
1738         if (incremental) {
1739                 struct origin *suspect = ent->suspect;
1740
1741                 printf("%s %d %d %d\n",
1742                        sha1_to_hex(suspect->commit->object.sha1),
1743                        ent->s_lno + 1, ent->lno + 1, ent->num_lines);
1744                 emit_one_suspect_detail(suspect, 0);
1745                 write_filename_info(suspect->path);
1746                 maybe_flush_or_die(stdout, "stdout");
1747         }
1748 }
1749
1750 /*
1751  * The main loop -- while we have blobs with lines whose true origin
1752  * is still unknown, pick one blob, and allow its lines to pass blames
1753  * to its parents. */
1754 static void assign_blame(struct scoreboard *sb, int opt)
1755 {
1756         struct rev_info *revs = sb->revs;
1757         struct commit *commit = prio_queue_get(&sb->commits);
1758
1759         while (commit) {
1760                 struct blame_entry *ent;
1761                 struct origin *suspect = commit->util;
1762
1763                 /* find one suspect to break down */
1764                 while (suspect && !suspect->suspects)
1765                         suspect = suspect->next;
1766
1767                 if (!suspect) {
1768                         commit = prio_queue_get(&sb->commits);
1769                         continue;
1770                 }
1771
1772                 assert(commit == suspect->commit);
1773
1774                 /*
1775                  * We will use this suspect later in the loop,
1776                  * so hold onto it in the meantime.
1777                  */
1778                 origin_incref(suspect);
1779                 parse_commit(commit);
1780                 if (reverse ||
1781                     (!(commit->object.flags & UNINTERESTING) &&
1782                      !(revs->max_age != -1 && commit->date < revs->max_age)))
1783                         pass_blame(sb, suspect, opt);
1784                 else {
1785                         commit->object.flags |= UNINTERESTING;
1786                         if (commit->object.parsed)
1787                                 mark_parents_uninteresting(commit);
1788                 }
1789                 /* treat root commit as boundary */
1790                 if (!commit->parents && !show_root)
1791                         commit->object.flags |= UNINTERESTING;
1792
1793                 /* Take responsibility for the remaining entries */
1794                 ent = suspect->suspects;
1795                 if (ent) {
1796                         suspect->guilty = 1;
1797                         for (;;) {
1798                                 struct blame_entry *next = ent->next;
1799                                 found_guilty_entry(ent);
1800                                 if (next) {
1801                                         ent = next;
1802                                         continue;
1803                                 }
1804                                 ent->next = sb->ent;
1805                                 sb->ent = suspect->suspects;
1806                                 suspect->suspects = NULL;
1807                                 break;
1808                         }
1809                 }
1810                 origin_decref(suspect);
1811
1812                 if (DEBUG) /* sanity */
1813                         sanity_check_refcnt(sb);
1814         }
1815 }
1816
1817 static const char *format_time(unsigned long time, const char *tz_str,
1818                                int show_raw_time)
1819 {
1820         static struct strbuf time_buf = STRBUF_INIT;
1821
1822         strbuf_reset(&time_buf);
1823         if (show_raw_time) {
1824                 strbuf_addf(&time_buf, "%lu %s", time, tz_str);
1825         }
1826         else {
1827                 const char *time_str;
1828                 size_t time_width;
1829                 int tz;
1830                 tz = atoi(tz_str);
1831                 time_str = show_date(time, tz, &blame_date_mode);
1832                 strbuf_addstr(&time_buf, time_str);
1833                 /*
1834                  * Add space paddings to time_buf to display a fixed width
1835                  * string, and use time_width for display width calibration.
1836                  */
1837                 for (time_width = utf8_strwidth(time_str);
1838                      time_width < blame_date_width;
1839                      time_width++)
1840                         strbuf_addch(&time_buf, ' ');
1841         }
1842         return time_buf.buf;
1843 }
1844
1845 #define OUTPUT_ANNOTATE_COMPAT  001
1846 #define OUTPUT_LONG_OBJECT_NAME 002
1847 #define OUTPUT_RAW_TIMESTAMP    004
1848 #define OUTPUT_PORCELAIN        010
1849 #define OUTPUT_SHOW_NAME        020
1850 #define OUTPUT_SHOW_NUMBER      040
1851 #define OUTPUT_SHOW_SCORE      0100
1852 #define OUTPUT_NO_AUTHOR       0200
1853 #define OUTPUT_SHOW_EMAIL       0400
1854 #define OUTPUT_LINE_PORCELAIN 01000
1855
1856 static void emit_porcelain_details(struct origin *suspect, int repeat)
1857 {
1858         if (emit_one_suspect_detail(suspect, repeat) ||
1859             (suspect->commit->object.flags & MORE_THAN_ONE_PATH))
1860                 write_filename_info(suspect->path);
1861 }
1862
1863 static void emit_porcelain(struct scoreboard *sb, struct blame_entry *ent,
1864                            int opt)
1865 {
1866         int repeat = opt & OUTPUT_LINE_PORCELAIN;
1867         int cnt;
1868         const char *cp;
1869         struct origin *suspect = ent->suspect;
1870         char hex[41];
1871
1872         strcpy(hex, sha1_to_hex(suspect->commit->object.sha1));
1873         printf("%s %d %d %d\n",
1874                hex,
1875                ent->s_lno + 1,
1876                ent->lno + 1,
1877                ent->num_lines);
1878         emit_porcelain_details(suspect, repeat);
1879
1880         cp = nth_line(sb, ent->lno);
1881         for (cnt = 0; cnt < ent->num_lines; cnt++) {
1882                 char ch;
1883                 if (cnt) {
1884                         printf("%s %d %d\n", hex,
1885                                ent->s_lno + 1 + cnt,
1886                                ent->lno + 1 + cnt);
1887                         if (repeat)
1888                                 emit_porcelain_details(suspect, 1);
1889                 }
1890                 putchar('\t');
1891                 do {
1892                         ch = *cp++;
1893                         putchar(ch);
1894                 } while (ch != '\n' &&
1895                          cp < sb->final_buf + sb->final_buf_size);
1896         }
1897
1898         if (sb->final_buf_size && cp[-1] != '\n')
1899                 putchar('\n');
1900 }
1901
1902 static void emit_other(struct scoreboard *sb, struct blame_entry *ent, int opt)
1903 {
1904         int cnt;
1905         const char *cp;
1906         struct origin *suspect = ent->suspect;
1907         struct commit_info ci;
1908         char hex[41];
1909         int show_raw_time = !!(opt & OUTPUT_RAW_TIMESTAMP);
1910
1911         get_commit_info(suspect->commit, &ci, 1);
1912         strcpy(hex, sha1_to_hex(suspect->commit->object.sha1));
1913
1914         cp = nth_line(sb, ent->lno);
1915         for (cnt = 0; cnt < ent->num_lines; cnt++) {
1916                 char ch;
1917                 int length = (opt & OUTPUT_LONG_OBJECT_NAME) ? 40 : abbrev;
1918
1919                 if (suspect->commit->object.flags & UNINTERESTING) {
1920                         if (blank_boundary)
1921                                 memset(hex, ' ', length);
1922                         else if (!(opt & OUTPUT_ANNOTATE_COMPAT)) {
1923                                 length--;
1924                                 putchar('^');
1925                         }
1926                 }
1927
1928                 printf("%.*s", length, hex);
1929                 if (opt & OUTPUT_ANNOTATE_COMPAT) {
1930                         const char *name;
1931                         if (opt & OUTPUT_SHOW_EMAIL)
1932                                 name = ci.author_mail.buf;
1933                         else
1934                                 name = ci.author.buf;
1935                         printf("\t(%10s\t%10s\t%d)", name,
1936                                format_time(ci.author_time, ci.author_tz.buf,
1937                                            show_raw_time),
1938                                ent->lno + 1 + cnt);
1939                 } else {
1940                         if (opt & OUTPUT_SHOW_SCORE)
1941                                 printf(" %*d %02d",
1942                                        max_score_digits, ent->score,
1943                                        ent->suspect->refcnt);
1944                         if (opt & OUTPUT_SHOW_NAME)
1945                                 printf(" %-*.*s", longest_file, longest_file,
1946                                        suspect->path);
1947                         if (opt & OUTPUT_SHOW_NUMBER)
1948                                 printf(" %*d", max_orig_digits,
1949                                        ent->s_lno + 1 + cnt);
1950
1951                         if (!(opt & OUTPUT_NO_AUTHOR)) {
1952                                 const char *name;
1953                                 int pad;
1954                                 if (opt & OUTPUT_SHOW_EMAIL)
1955                                         name = ci.author_mail.buf;
1956                                 else
1957                                         name = ci.author.buf;
1958                                 pad = longest_author - utf8_strwidth(name);
1959                                 printf(" (%s%*s %10s",
1960                                        name, pad, "",
1961                                        format_time(ci.author_time,
1962                                                    ci.author_tz.buf,
1963                                                    show_raw_time));
1964                         }
1965                         printf(" %*d) ",
1966                                max_digits, ent->lno + 1 + cnt);
1967                 }
1968                 do {
1969                         ch = *cp++;
1970                         putchar(ch);
1971                 } while (ch != '\n' &&
1972                          cp < sb->final_buf + sb->final_buf_size);
1973         }
1974
1975         if (sb->final_buf_size && cp[-1] != '\n')
1976                 putchar('\n');
1977
1978         commit_info_destroy(&ci);
1979 }
1980
1981 static void output(struct scoreboard *sb, int option)
1982 {
1983         struct blame_entry *ent;
1984
1985         if (option & OUTPUT_PORCELAIN) {
1986                 for (ent = sb->ent; ent; ent = ent->next) {
1987                         int count = 0;
1988                         struct origin *suspect;
1989                         struct commit *commit = ent->suspect->commit;
1990                         if (commit->object.flags & MORE_THAN_ONE_PATH)
1991                                 continue;
1992                         for (suspect = commit->util; suspect; suspect = suspect->next) {
1993                                 if (suspect->guilty && count++) {
1994                                         commit->object.flags |= MORE_THAN_ONE_PATH;
1995                                         break;
1996                                 }
1997                         }
1998                 }
1999         }
2000
2001         for (ent = sb->ent; ent; ent = ent->next) {
2002                 if (option & OUTPUT_PORCELAIN)
2003                         emit_porcelain(sb, ent, option);
2004                 else {
2005                         emit_other(sb, ent, option);
2006                 }
2007         }
2008 }
2009
2010 static const char *get_next_line(const char *start, const char *end)
2011 {
2012         const char *nl = memchr(start, '\n', end - start);
2013         return nl ? nl + 1 : end;
2014 }
2015
2016 /*
2017  * To allow quick access to the contents of nth line in the
2018  * final image, prepare an index in the scoreboard.
2019  */
2020 static int prepare_lines(struct scoreboard *sb)
2021 {
2022         const char *buf = sb->final_buf;
2023         unsigned long len = sb->final_buf_size;
2024         const char *end = buf + len;
2025         const char *p;
2026         int *lineno;
2027         int num = 0;
2028
2029         for (p = buf; p < end; p = get_next_line(p, end))
2030                 num++;
2031
2032         sb->lineno = lineno = xmalloc(sizeof(*sb->lineno) * (num + 1));
2033
2034         for (p = buf; p < end; p = get_next_line(p, end))
2035                 *lineno++ = p - buf;
2036
2037         *lineno = len;
2038
2039         sb->num_lines = num;
2040         return sb->num_lines;
2041 }
2042
2043 /*
2044  * Add phony grafts for use with -S; this is primarily to
2045  * support git's cvsserver that wants to give a linear history
2046  * to its clients.
2047  */
2048 static int read_ancestry(const char *graft_file)
2049 {
2050         FILE *fp = fopen(graft_file, "r");
2051         struct strbuf buf = STRBUF_INIT;
2052         if (!fp)
2053                 return -1;
2054         while (!strbuf_getwholeline(&buf, fp, '\n')) {
2055                 /* The format is just "Commit Parent1 Parent2 ...\n" */
2056                 struct commit_graft *graft = read_graft_line(buf.buf, buf.len);
2057                 if (graft)
2058                         register_commit_graft(graft, 0);
2059         }
2060         fclose(fp);
2061         strbuf_release(&buf);
2062         return 0;
2063 }
2064
2065 static int update_auto_abbrev(int auto_abbrev, struct origin *suspect)
2066 {
2067         const char *uniq = find_unique_abbrev(suspect->commit->object.sha1,
2068                                               auto_abbrev);
2069         int len = strlen(uniq);
2070         if (auto_abbrev < len)
2071                 return len;
2072         return auto_abbrev;
2073 }
2074
2075 /*
2076  * How many columns do we need to show line numbers, authors,
2077  * and filenames?
2078  */
2079 static void find_alignment(struct scoreboard *sb, int *option)
2080 {
2081         int longest_src_lines = 0;
2082         int longest_dst_lines = 0;
2083         unsigned largest_score = 0;
2084         struct blame_entry *e;
2085         int compute_auto_abbrev = (abbrev < 0);
2086         int auto_abbrev = default_abbrev;
2087
2088         for (e = sb->ent; e; e = e->next) {
2089                 struct origin *suspect = e->suspect;
2090                 int num;
2091
2092                 if (compute_auto_abbrev)
2093                         auto_abbrev = update_auto_abbrev(auto_abbrev, suspect);
2094                 if (strcmp(suspect->path, sb->path))
2095                         *option |= OUTPUT_SHOW_NAME;
2096                 num = strlen(suspect->path);
2097                 if (longest_file < num)
2098                         longest_file = num;
2099                 if (!(suspect->commit->object.flags & METAINFO_SHOWN)) {
2100                         struct commit_info ci;
2101                         suspect->commit->object.flags |= METAINFO_SHOWN;
2102                         get_commit_info(suspect->commit, &ci, 1);
2103                         if (*option & OUTPUT_SHOW_EMAIL)
2104                                 num = utf8_strwidth(ci.author_mail.buf);
2105                         else
2106                                 num = utf8_strwidth(ci.author.buf);
2107                         if (longest_author < num)
2108                                 longest_author = num;
2109                         commit_info_destroy(&ci);
2110                 }
2111                 num = e->s_lno + e->num_lines;
2112                 if (longest_src_lines < num)
2113                         longest_src_lines = num;
2114                 num = e->lno + e->num_lines;
2115                 if (longest_dst_lines < num)
2116                         longest_dst_lines = num;
2117                 if (largest_score < ent_score(sb, e))
2118                         largest_score = ent_score(sb, e);
2119         }
2120         max_orig_digits = decimal_width(longest_src_lines);
2121         max_digits = decimal_width(longest_dst_lines);
2122         max_score_digits = decimal_width(largest_score);
2123
2124         if (compute_auto_abbrev)
2125                 /* one more abbrev length is needed for the boundary commit */
2126                 abbrev = auto_abbrev + 1;
2127 }
2128
2129 /*
2130  * For debugging -- origin is refcounted, and this asserts that
2131  * we do not underflow.
2132  */
2133 static void sanity_check_refcnt(struct scoreboard *sb)
2134 {
2135         int baa = 0;
2136         struct blame_entry *ent;
2137
2138         for (ent = sb->ent; ent; ent = ent->next) {
2139                 /* Nobody should have zero or negative refcnt */
2140                 if (ent->suspect->refcnt <= 0) {
2141                         fprintf(stderr, "%s in %s has negative refcnt %d\n",
2142                                 ent->suspect->path,
2143                                 sha1_to_hex(ent->suspect->commit->object.sha1),
2144                                 ent->suspect->refcnt);
2145                         baa = 1;
2146                 }
2147         }
2148         if (baa) {
2149                 int opt = 0160;
2150                 find_alignment(sb, &opt);
2151                 output(sb, opt);
2152                 die("Baa %d!", baa);
2153         }
2154 }
2155
2156 static unsigned parse_score(const char *arg)
2157 {
2158         char *end;
2159         unsigned long score = strtoul(arg, &end, 10);
2160         if (*end)
2161                 return 0;
2162         return score;
2163 }
2164
2165 static const char *add_prefix(const char *prefix, const char *path)
2166 {
2167         return prefix_path(prefix, prefix ? strlen(prefix) : 0, path);
2168 }
2169
2170 static int git_blame_config(const char *var, const char *value, void *cb)
2171 {
2172         if (!strcmp(var, "blame.showroot")) {
2173                 show_root = git_config_bool(var, value);
2174                 return 0;
2175         }
2176         if (!strcmp(var, "blame.blankboundary")) {
2177                 blank_boundary = git_config_bool(var, value);
2178                 return 0;
2179         }
2180         if (!strcmp(var, "blame.showemail")) {
2181                 int *output_option = cb;
2182                 if (git_config_bool(var, value))
2183                         *output_option |= OUTPUT_SHOW_EMAIL;
2184                 else
2185                         *output_option &= ~OUTPUT_SHOW_EMAIL;
2186                 return 0;
2187         }
2188         if (!strcmp(var, "blame.date")) {
2189                 if (!value)
2190                         return config_error_nonbool(var);
2191                 parse_date_format(value, &blame_date_mode);
2192                 return 0;
2193         }
2194
2195         if (userdiff_config(var, value) < 0)
2196                 return -1;
2197
2198         return git_default_config(var, value, cb);
2199 }
2200
2201 static void verify_working_tree_path(struct commit *work_tree, const char *path)
2202 {
2203         struct commit_list *parents;
2204
2205         for (parents = work_tree->parents; parents; parents = parents->next) {
2206                 const unsigned char *commit_sha1 = parents->item->object.sha1;
2207                 unsigned char blob_sha1[20];
2208                 unsigned mode;
2209
2210                 if (!get_tree_entry(commit_sha1, path, blob_sha1, &mode) &&
2211                     sha1_object_info(blob_sha1, NULL) == OBJ_BLOB)
2212                         return;
2213         }
2214         die("no such path '%s' in HEAD", path);
2215 }
2216
2217 static struct commit_list **append_parent(struct commit_list **tail, const unsigned char *sha1)
2218 {
2219         struct commit *parent;
2220
2221         parent = lookup_commit_reference(sha1);
2222         if (!parent)
2223                 die("no such commit %s", sha1_to_hex(sha1));
2224         return &commit_list_insert(parent, tail)->next;
2225 }
2226
2227 static void append_merge_parents(struct commit_list **tail)
2228 {
2229         int merge_head;
2230         struct strbuf line = STRBUF_INIT;
2231
2232         merge_head = open(git_path_merge_head(), O_RDONLY);
2233         if (merge_head < 0) {
2234                 if (errno == ENOENT)
2235                         return;
2236                 die("cannot open '%s' for reading", git_path_merge_head());
2237         }
2238
2239         while (!strbuf_getwholeline_fd(&line, merge_head, '\n')) {
2240                 unsigned char sha1[20];
2241                 if (line.len < 40 || get_sha1_hex(line.buf, sha1))
2242                         die("unknown line in '%s': %s", git_path_merge_head(), line.buf);
2243                 tail = append_parent(tail, sha1);
2244         }
2245         close(merge_head);
2246         strbuf_release(&line);
2247 }
2248
2249 /*
2250  * This isn't as simple as passing sb->buf and sb->len, because we
2251  * want to transfer ownership of the buffer to the commit (so we
2252  * must use detach).
2253  */
2254 static void set_commit_buffer_from_strbuf(struct commit *c, struct strbuf *sb)
2255 {
2256         size_t len;
2257         void *buf = strbuf_detach(sb, &len);
2258         set_commit_buffer(c, buf, len);
2259 }
2260
2261 /*
2262  * Prepare a dummy commit that represents the work tree (or staged) item.
2263  * Note that annotating work tree item never works in the reverse.
2264  */
2265 static struct commit *fake_working_tree_commit(struct diff_options *opt,
2266                                                const char *path,
2267                                                const char *contents_from)
2268 {
2269         struct commit *commit;
2270         struct origin *origin;
2271         struct commit_list **parent_tail, *parent;
2272         unsigned char head_sha1[20];
2273         struct strbuf buf = STRBUF_INIT;
2274         const char *ident;
2275         time_t now;
2276         int size, len;
2277         struct cache_entry *ce;
2278         unsigned mode;
2279         struct strbuf msg = STRBUF_INIT;
2280
2281         time(&now);
2282         commit = alloc_commit_node();
2283         commit->object.parsed = 1;
2284         commit->date = now;
2285         parent_tail = &commit->parents;
2286
2287         if (!resolve_ref_unsafe("HEAD", RESOLVE_REF_READING, head_sha1, NULL))
2288                 die("no such ref: HEAD");
2289
2290         parent_tail = append_parent(parent_tail, head_sha1);
2291         append_merge_parents(parent_tail);
2292         verify_working_tree_path(commit, path);
2293
2294         origin = make_origin(commit, path);
2295
2296         ident = fmt_ident("Not Committed Yet", "not.committed.yet", NULL, 0);
2297         strbuf_addstr(&msg, "tree 0000000000000000000000000000000000000000\n");
2298         for (parent = commit->parents; parent; parent = parent->next)
2299                 strbuf_addf(&msg, "parent %s\n",
2300                             sha1_to_hex(parent->item->object.sha1));
2301         strbuf_addf(&msg,
2302                     "author %s\n"
2303                     "committer %s\n\n"
2304                     "Version of %s from %s\n",
2305                     ident, ident, path,
2306                     (!contents_from ? path :
2307                      (!strcmp(contents_from, "-") ? "standard input" : contents_from)));
2308         set_commit_buffer_from_strbuf(commit, &msg);
2309
2310         if (!contents_from || strcmp("-", contents_from)) {
2311                 struct stat st;
2312                 const char *read_from;
2313                 char *buf_ptr;
2314                 unsigned long buf_len;
2315
2316                 if (contents_from) {
2317                         if (stat(contents_from, &st) < 0)
2318                                 die_errno("Cannot stat '%s'", contents_from);
2319                         read_from = contents_from;
2320                 }
2321                 else {
2322                         if (lstat(path, &st) < 0)
2323                                 die_errno("Cannot lstat '%s'", path);
2324                         read_from = path;
2325                 }
2326                 mode = canon_mode(st.st_mode);
2327
2328                 switch (st.st_mode & S_IFMT) {
2329                 case S_IFREG:
2330                         if (DIFF_OPT_TST(opt, ALLOW_TEXTCONV) &&
2331                             textconv_object(read_from, mode, null_sha1, 0, &buf_ptr, &buf_len))
2332                                 strbuf_attach(&buf, buf_ptr, buf_len, buf_len + 1);
2333                         else if (strbuf_read_file(&buf, read_from, st.st_size) != st.st_size)
2334                                 die_errno("cannot open or read '%s'", read_from);
2335                         break;
2336                 case S_IFLNK:
2337                         if (strbuf_readlink(&buf, read_from, st.st_size) < 0)
2338                                 die_errno("cannot readlink '%s'", read_from);
2339                         break;
2340                 default:
2341                         die("unsupported file type %s", read_from);
2342                 }
2343         }
2344         else {
2345                 /* Reading from stdin */
2346                 mode = 0;
2347                 if (strbuf_read(&buf, 0, 0) < 0)
2348                         die_errno("failed to read from stdin");
2349         }
2350         convert_to_git(path, buf.buf, buf.len, &buf, 0);
2351         origin->file.ptr = buf.buf;
2352         origin->file.size = buf.len;
2353         pretend_sha1_file(buf.buf, buf.len, OBJ_BLOB, origin->blob_sha1);
2354
2355         /*
2356          * Read the current index, replace the path entry with
2357          * origin->blob_sha1 without mucking with its mode or type
2358          * bits; we are not going to write this index out -- we just
2359          * want to run "diff-index --cached".
2360          */
2361         discard_cache();
2362         read_cache();
2363
2364         len = strlen(path);
2365         if (!mode) {
2366                 int pos = cache_name_pos(path, len);
2367                 if (0 <= pos)
2368                         mode = active_cache[pos]->ce_mode;
2369                 else
2370                         /* Let's not bother reading from HEAD tree */
2371                         mode = S_IFREG | 0644;
2372         }
2373         size = cache_entry_size(len);
2374         ce = xcalloc(1, size);
2375         hashcpy(ce->sha1, origin->blob_sha1);
2376         memcpy(ce->name, path, len);
2377         ce->ce_flags = create_ce_flags(0);
2378         ce->ce_namelen = len;
2379         ce->ce_mode = create_ce_mode(mode);
2380         add_cache_entry(ce, ADD_CACHE_OK_TO_ADD|ADD_CACHE_OK_TO_REPLACE);
2381
2382         /*
2383          * We are not going to write this out, so this does not matter
2384          * right now, but someday we might optimize diff-index --cached
2385          * with cache-tree information.
2386          */
2387         cache_tree_invalidate_path(&the_index, path);
2388
2389         return commit;
2390 }
2391
2392 static char *prepare_final(struct scoreboard *sb)
2393 {
2394         int i;
2395         const char *final_commit_name = NULL;
2396         struct rev_info *revs = sb->revs;
2397
2398         /*
2399          * There must be one and only one positive commit in the
2400          * revs->pending array.
2401          */
2402         for (i = 0; i < revs->pending.nr; i++) {
2403                 struct object *obj = revs->pending.objects[i].item;
2404                 if (obj->flags & UNINTERESTING)
2405                         continue;
2406                 while (obj->type == OBJ_TAG)
2407                         obj = deref_tag(obj, NULL, 0);
2408                 if (obj->type != OBJ_COMMIT)
2409                         die("Non commit %s?", revs->pending.objects[i].name);
2410                 if (sb->final)
2411                         die("More than one commit to dig from %s and %s?",
2412                             revs->pending.objects[i].name,
2413                             final_commit_name);
2414                 sb->final = (struct commit *) obj;
2415                 final_commit_name = revs->pending.objects[i].name;
2416         }
2417         return xstrdup_or_null(final_commit_name);
2418 }
2419
2420 static char *prepare_initial(struct scoreboard *sb)
2421 {
2422         int i;
2423         const char *final_commit_name = NULL;
2424         struct rev_info *revs = sb->revs;
2425
2426         /*
2427          * There must be one and only one negative commit, and it must be
2428          * the boundary.
2429          */
2430         for (i = 0; i < revs->pending.nr; i++) {
2431                 struct object *obj = revs->pending.objects[i].item;
2432                 if (!(obj->flags & UNINTERESTING))
2433                         continue;
2434                 while (obj->type == OBJ_TAG)
2435                         obj = deref_tag(obj, NULL, 0);
2436                 if (obj->type != OBJ_COMMIT)
2437                         die("Non commit %s?", revs->pending.objects[i].name);
2438                 if (sb->final)
2439                         die("More than one commit to dig down to %s and %s?",
2440                             revs->pending.objects[i].name,
2441                             final_commit_name);
2442                 sb->final = (struct commit *) obj;
2443                 final_commit_name = revs->pending.objects[i].name;
2444         }
2445         if (!final_commit_name)
2446                 die("No commit to dig down to?");
2447         return xstrdup(final_commit_name);
2448 }
2449
2450 static int blame_copy_callback(const struct option *option, const char *arg, int unset)
2451 {
2452         int *opt = option->value;
2453
2454         /*
2455          * -C enables copy from removed files;
2456          * -C -C enables copy from existing files, but only
2457          *       when blaming a new file;
2458          * -C -C -C enables copy from existing files for
2459          *          everybody
2460          */
2461         if (*opt & PICKAXE_BLAME_COPY_HARDER)
2462                 *opt |= PICKAXE_BLAME_COPY_HARDEST;
2463         if (*opt & PICKAXE_BLAME_COPY)
2464                 *opt |= PICKAXE_BLAME_COPY_HARDER;
2465         *opt |= PICKAXE_BLAME_COPY | PICKAXE_BLAME_MOVE;
2466
2467         if (arg)
2468                 blame_copy_score = parse_score(arg);
2469         return 0;
2470 }
2471
2472 static int blame_move_callback(const struct option *option, const char *arg, int unset)
2473 {
2474         int *opt = option->value;
2475
2476         *opt |= PICKAXE_BLAME_MOVE;
2477
2478         if (arg)
2479                 blame_move_score = parse_score(arg);
2480         return 0;
2481 }
2482
2483 int cmd_blame(int argc, const char **argv, const char *prefix)
2484 {
2485         struct rev_info revs;
2486         const char *path;
2487         struct scoreboard sb;
2488         struct origin *o;
2489         struct blame_entry *ent = NULL;
2490         long dashdash_pos, lno;
2491         char *final_commit_name = NULL;
2492         enum object_type type;
2493
2494         static struct string_list range_list;
2495         static int output_option = 0, opt = 0;
2496         static int show_stats = 0;
2497         static const char *revs_file = NULL;
2498         static const char *contents_from = NULL;
2499         static const struct option options[] = {
2500                 OPT_BOOL(0, "incremental", &incremental, N_("Show blame entries as we find them, incrementally")),
2501                 OPT_BOOL('b', NULL, &blank_boundary, N_("Show blank SHA-1 for boundary commits (Default: off)")),
2502                 OPT_BOOL(0, "root", &show_root, N_("Do not treat root commits as boundaries (Default: off)")),
2503                 OPT_BOOL(0, "show-stats", &show_stats, N_("Show work cost statistics")),
2504                 OPT_BIT(0, "score-debug", &output_option, N_("Show output score for blame entries"), OUTPUT_SHOW_SCORE),
2505                 OPT_BIT('f', "show-name", &output_option, N_("Show original filename (Default: auto)"), OUTPUT_SHOW_NAME),
2506                 OPT_BIT('n', "show-number", &output_option, N_("Show original linenumber (Default: off)"), OUTPUT_SHOW_NUMBER),
2507                 OPT_BIT('p', "porcelain", &output_option, N_("Show in a format designed for machine consumption"), OUTPUT_PORCELAIN),
2508                 OPT_BIT(0, "line-porcelain", &output_option, N_("Show porcelain format with per-line commit information"), OUTPUT_PORCELAIN|OUTPUT_LINE_PORCELAIN),
2509                 OPT_BIT('c', NULL, &output_option, N_("Use the same output mode as git-annotate (Default: off)"), OUTPUT_ANNOTATE_COMPAT),
2510                 OPT_BIT('t', NULL, &output_option, N_("Show raw timestamp (Default: off)"), OUTPUT_RAW_TIMESTAMP),
2511                 OPT_BIT('l', NULL, &output_option, N_("Show long commit SHA1 (Default: off)"), OUTPUT_LONG_OBJECT_NAME),
2512                 OPT_BIT('s', NULL, &output_option, N_("Suppress author name and timestamp (Default: off)"), OUTPUT_NO_AUTHOR),
2513                 OPT_BIT('e', "show-email", &output_option, N_("Show author email instead of name (Default: off)"), OUTPUT_SHOW_EMAIL),
2514                 OPT_BIT('w', NULL, &xdl_opts, N_("Ignore whitespace differences"), XDF_IGNORE_WHITESPACE),
2515                 OPT_BIT(0, "minimal", &xdl_opts, N_("Spend extra cycles to find better match"), XDF_NEED_MINIMAL),
2516                 OPT_STRING('S', NULL, &revs_file, N_("file"), N_("Use revisions from <file> instead of calling git-rev-list")),
2517                 OPT_STRING(0, "contents", &contents_from, N_("file"), N_("Use <file>'s contents as the final image")),
2518                 { OPTION_CALLBACK, 'C', NULL, &opt, N_("score"), N_("Find line copies within and across files"), PARSE_OPT_OPTARG, blame_copy_callback },
2519                 { OPTION_CALLBACK, 'M', NULL, &opt, N_("score"), N_("Find line movements within and across files"), PARSE_OPT_OPTARG, blame_move_callback },
2520                 OPT_STRING_LIST('L', NULL, &range_list, N_("n,m"), N_("Process only line range n,m, counting from 1")),
2521                 OPT__ABBREV(&abbrev),
2522                 OPT_END()
2523         };
2524
2525         struct parse_opt_ctx_t ctx;
2526         int cmd_is_annotate = !strcmp(argv[0], "annotate");
2527         struct range_set ranges;
2528         unsigned int range_i;
2529         long anchor;
2530
2531         git_config(git_blame_config, &output_option);
2532         init_revisions(&revs, NULL);
2533         revs.date_mode = blame_date_mode;
2534         DIFF_OPT_SET(&revs.diffopt, ALLOW_TEXTCONV);
2535         DIFF_OPT_SET(&revs.diffopt, FOLLOW_RENAMES);
2536
2537         save_commit_buffer = 0;
2538         dashdash_pos = 0;
2539
2540         parse_options_start(&ctx, argc, argv, prefix, options,
2541                             PARSE_OPT_KEEP_DASHDASH | PARSE_OPT_KEEP_ARGV0);
2542         for (;;) {
2543                 switch (parse_options_step(&ctx, options, blame_opt_usage)) {
2544                 case PARSE_OPT_HELP:
2545                         exit(129);
2546                 case PARSE_OPT_DONE:
2547                         if (ctx.argv[0])
2548                                 dashdash_pos = ctx.cpidx;
2549                         goto parse_done;
2550                 }
2551
2552                 if (!strcmp(ctx.argv[0], "--reverse")) {
2553                         ctx.argv[0] = "--children";
2554                         reverse = 1;
2555                 }
2556                 parse_revision_opt(&revs, &ctx, options, blame_opt_usage);
2557         }
2558 parse_done:
2559         no_whole_file_rename = !DIFF_OPT_TST(&revs.diffopt, FOLLOW_RENAMES);
2560         DIFF_OPT_CLR(&revs.diffopt, FOLLOW_RENAMES);
2561         argc = parse_options_end(&ctx);
2562
2563         if (0 < abbrev)
2564                 /* one more abbrev length is needed for the boundary commit */
2565                 abbrev++;
2566
2567         if (revs_file && read_ancestry(revs_file))
2568                 die_errno("reading graft file '%s' failed", revs_file);
2569
2570         if (cmd_is_annotate) {
2571                 output_option |= OUTPUT_ANNOTATE_COMPAT;
2572                 blame_date_mode.type = DATE_ISO8601;
2573         } else {
2574                 blame_date_mode = revs.date_mode;
2575         }
2576
2577         /* The maximum width used to show the dates */
2578         switch (blame_date_mode.type) {
2579         case DATE_RFC2822:
2580                 blame_date_width = sizeof("Thu, 19 Oct 2006 16:00:04 -0700");
2581                 break;
2582         case DATE_ISO8601_STRICT:
2583                 blame_date_width = sizeof("2006-10-19T16:00:04-07:00");
2584                 break;
2585         case DATE_ISO8601:
2586                 blame_date_width = sizeof("2006-10-19 16:00:04 -0700");
2587                 break;
2588         case DATE_RAW:
2589                 blame_date_width = sizeof("1161298804 -0700");
2590                 break;
2591         case DATE_SHORT:
2592                 blame_date_width = sizeof("2006-10-19");
2593                 break;
2594         case DATE_RELATIVE:
2595                 /* TRANSLATORS: This string is used to tell us the maximum
2596                    display width for a relative timestamp in "git blame"
2597                    output.  For C locale, "4 years, 11 months ago", which
2598                    takes 22 places, is the longest among various forms of
2599                    relative timestamps, but your language may need more or
2600                    fewer display columns. */
2601                 blame_date_width = utf8_strwidth(_("4 years, 11 months ago")) + 1; /* add the null */
2602                 break;
2603         case DATE_LOCAL:
2604         case DATE_NORMAL:
2605                 blame_date_width = sizeof("Thu Oct 19 16:00:04 2006 -0700");
2606                 break;
2607         case DATE_STRFTIME:
2608                 blame_date_width = strlen(show_date(0, 0, &blame_date_mode)) + 1; /* add the null */
2609                 break;
2610         }
2611         blame_date_width -= 1; /* strip the null */
2612
2613         if (DIFF_OPT_TST(&revs.diffopt, FIND_COPIES_HARDER))
2614                 opt |= (PICKAXE_BLAME_COPY | PICKAXE_BLAME_MOVE |
2615                         PICKAXE_BLAME_COPY_HARDER);
2616
2617         if (!blame_move_score)
2618                 blame_move_score = BLAME_DEFAULT_MOVE_SCORE;
2619         if (!blame_copy_score)
2620                 blame_copy_score = BLAME_DEFAULT_COPY_SCORE;
2621
2622         /*
2623          * We have collected options unknown to us in argv[1..unk]
2624          * which are to be passed to revision machinery if we are
2625          * going to do the "bottom" processing.
2626          *
2627          * The remaining are:
2628          *
2629          * (1) if dashdash_pos != 0, it is either
2630          *     "blame [revisions] -- <path>" or
2631          *     "blame -- <path> <rev>"
2632          *
2633          * (2) otherwise, it is one of the two:
2634          *     "blame [revisions] <path>"
2635          *     "blame <path> <rev>"
2636          *
2637          * Note that we must strip out <path> from the arguments: we do not
2638          * want the path pruning but we may want "bottom" processing.
2639          */
2640         if (dashdash_pos) {
2641                 switch (argc - dashdash_pos - 1) {
2642                 case 2: /* (1b) */
2643                         if (argc != 4)
2644                                 usage_with_options(blame_opt_usage, options);
2645                         /* reorder for the new way: <rev> -- <path> */
2646                         argv[1] = argv[3];
2647                         argv[3] = argv[2];
2648                         argv[2] = "--";
2649                         /* FALLTHROUGH */
2650                 case 1: /* (1a) */
2651                         path = add_prefix(prefix, argv[--argc]);
2652                         argv[argc] = NULL;
2653                         break;
2654                 default:
2655                         usage_with_options(blame_opt_usage, options);
2656                 }
2657         } else {
2658                 if (argc < 2)
2659                         usage_with_options(blame_opt_usage, options);
2660                 path = add_prefix(prefix, argv[argc - 1]);
2661                 if (argc == 3 && !file_exists(path)) { /* (2b) */
2662                         path = add_prefix(prefix, argv[1]);
2663                         argv[1] = argv[2];
2664                 }
2665                 argv[argc - 1] = "--";
2666
2667                 setup_work_tree();
2668                 if (!file_exists(path))
2669                         die_errno("cannot stat path '%s'", path);
2670         }
2671
2672         revs.disable_stdin = 1;
2673         setup_revisions(argc, argv, &revs, NULL);
2674         memset(&sb, 0, sizeof(sb));
2675
2676         sb.revs = &revs;
2677         if (!reverse) {
2678                 final_commit_name = prepare_final(&sb);
2679                 sb.commits.compare = compare_commits_by_commit_date;
2680         }
2681         else if (contents_from)
2682                 die("--contents and --children do not blend well.");
2683         else {
2684                 final_commit_name = prepare_initial(&sb);
2685                 sb.commits.compare = compare_commits_by_reverse_commit_date;
2686         }
2687
2688         if (!sb.final) {
2689                 /*
2690                  * "--not A B -- path" without anything positive;
2691                  * do not default to HEAD, but use the working tree
2692                  * or "--contents".
2693                  */
2694                 setup_work_tree();
2695                 sb.final = fake_working_tree_commit(&sb.revs->diffopt,
2696                                                     path, contents_from);
2697                 add_pending_object(&revs, &(sb.final->object), ":");
2698         }
2699         else if (contents_from)
2700                 die("Cannot use --contents with final commit object name");
2701
2702         /*
2703          * If we have bottom, this will mark the ancestors of the
2704          * bottom commits we would reach while traversing as
2705          * uninteresting.
2706          */
2707         if (prepare_revision_walk(&revs))
2708                 die(_("revision walk setup failed"));
2709
2710         if (is_null_sha1(sb.final->object.sha1)) {
2711                 o = sb.final->util;
2712                 sb.final_buf = xmemdupz(o->file.ptr, o->file.size);
2713                 sb.final_buf_size = o->file.size;
2714         }
2715         else {
2716                 o = get_origin(&sb, sb.final, path);
2717                 if (fill_blob_sha1_and_mode(o))
2718                         die("no such path %s in %s", path, final_commit_name);
2719
2720                 if (DIFF_OPT_TST(&sb.revs->diffopt, ALLOW_TEXTCONV) &&
2721                     textconv_object(path, o->mode, o->blob_sha1, 1, (char **) &sb.final_buf,
2722                                     &sb.final_buf_size))
2723                         ;
2724                 else
2725                         sb.final_buf = read_sha1_file(o->blob_sha1, &type,
2726                                                       &sb.final_buf_size);
2727
2728                 if (!sb.final_buf)
2729                         die("Cannot read blob %s for path %s",
2730                             sha1_to_hex(o->blob_sha1),
2731                             path);
2732         }
2733         num_read_blob++;
2734         lno = prepare_lines(&sb);
2735
2736         if (lno && !range_list.nr)
2737                 string_list_append(&range_list, xstrdup("1"));
2738
2739         anchor = 1;
2740         range_set_init(&ranges, range_list.nr);
2741         for (range_i = 0; range_i < range_list.nr; ++range_i) {
2742                 long bottom, top;
2743                 if (parse_range_arg(range_list.items[range_i].string,
2744                                     nth_line_cb, &sb, lno, anchor,
2745                                     &bottom, &top, sb.path))
2746                         usage(blame_usage);
2747                 if (lno < top || ((lno || bottom) && lno < bottom))
2748                         die("file %s has only %lu lines", path, lno);
2749                 if (bottom < 1)
2750                         bottom = 1;
2751                 if (top < 1)
2752                         top = lno;
2753                 bottom--;
2754                 range_set_append_unsafe(&ranges, bottom, top);
2755                 anchor = top + 1;
2756         }
2757         sort_and_merge_range_set(&ranges);
2758
2759         for (range_i = ranges.nr; range_i > 0; --range_i) {
2760                 const struct range *r = &ranges.ranges[range_i - 1];
2761                 long bottom = r->start;
2762                 long top = r->end;
2763                 struct blame_entry *next = ent;
2764                 ent = xcalloc(1, sizeof(*ent));
2765                 ent->lno = bottom;
2766                 ent->num_lines = top - bottom;
2767                 ent->suspect = o;
2768                 ent->s_lno = bottom;
2769                 ent->next = next;
2770                 origin_incref(o);
2771         }
2772
2773         o->suspects = ent;
2774         prio_queue_put(&sb.commits, o->commit);
2775
2776         origin_decref(o);
2777
2778         range_set_release(&ranges);
2779         string_list_clear(&range_list, 0);
2780
2781         sb.ent = NULL;
2782         sb.path = path;
2783
2784         read_mailmap(&mailmap, NULL);
2785
2786         if (!incremental)
2787                 setup_pager();
2788
2789         assign_blame(&sb, opt);
2790
2791         free(final_commit_name);
2792
2793         if (incremental)
2794                 return 0;
2795
2796         sb.ent = blame_sort(sb.ent, compare_blame_final);
2797
2798         coalesce(&sb);
2799
2800         if (!(output_option & OUTPUT_PORCELAIN))
2801                 find_alignment(&sb, &output_option);
2802
2803         output(&sb, output_option);
2804         free((void *)sb.final_buf);
2805         for (ent = sb.ent; ent; ) {
2806                 struct blame_entry *e = ent->next;
2807                 free(ent);
2808                 ent = e;
2809         }
2810
2811         if (show_stats) {
2812                 printf("num read blob: %d\n", num_read_blob);
2813                 printf("num get patch: %d\n", num_get_patch);
2814                 printf("num commits: %d\n", num_commits);
2815         }
2816         return 0;
2817 }