blame.c

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