blame.c

   1 #include "cache.h"
   2 #include "refs.h"
   3 #include "object-store.h"
   4 #include "cache-tree.h"
   5 #include "mergesort.h"
   6 #include "diff.h"
   7 #include "diffcore.h"
   8 #include "tag.h"
   9 #include "blame.h"
  10 #include "alloc.h"
  11 #include "commit-slab.h"
  12
  13 define_commit_slab(blame_suspects, struct blame_origin *);
  14 static struct blame_suspects blame_suspects;
  15
  16 struct blame_origin *get_blame_suspects(struct commit *commit)
  17 {
  18         struct blame_origin **result;
  19
  20         result = blame_suspects_peek(&blame_suspects, commit);
  21
  22         return result ? *result : NULL;
  23 }
  24
  25 static void set_blame_suspects(struct commit *commit, struct blame_origin *origin)
  26 {
  27         *blame_suspects_at(&blame_suspects, commit) = origin;
  28 }
  29
  30 void blame_origin_decref(struct blame_origin *o)
  31 {
  32         if (o && --o->refcnt <= 0) {
  33                 struct blame_origin *p, *l = NULL;
  34                 if (o->previous)
  35                         blame_origin_decref(o->previous);
  36                 free(o->file.ptr);
  37                 /* Should be present exactly once in commit chain */
  38                 for (p = get_blame_suspects(o->commit); p; l = p, p = p->next) {
  39                         if (p == o) {
  40                                 if (l)
  41                                         l->next = p->next;
  42                                 else
  43                                         set_blame_suspects(o->commit, p->next);
  44                                 free(o);
  45                                 return;
  46                         }
  47                 }
  48                 die("internal error in blame_origin_decref");
  49         }
  50 }
  51
  52 /*
  53  * Given a commit and a path in it, create a new origin structure.
  54  * The callers that add blame to the scoreboard should use
  55  * get_origin() to obtain shared, refcounted copy instead of calling
  56  * this function directly.
  57  */
  58 static struct blame_origin *make_origin(struct commit *commit, const char *path)
  59 {
  60         struct blame_origin *o;
  61         FLEX_ALLOC_STR(o, path, path);
  62         o->commit = commit;
  63         o->refcnt = 1;
  64         o->next = get_blame_suspects(commit);
  65         set_blame_suspects(commit, o);
  66         return o;
  67 }
  68
  69 /*
  70  * Locate an existing origin or create a new one.
  71  * This moves the origin to front position in the commit util list.
  72  */
  73 static struct blame_origin *get_origin(struct commit *commit, const char *path)
  74 {
  75         struct blame_origin *o, *l;
  76
  77         for (o = get_blame_suspects(commit), l = NULL; o; l = o, o = o->next) {
  78                 if (!strcmp(o->path, path)) {
  79                         /* bump to front */
  80                         if (l) {
  81                                 l->next = o->next;
  82                                 o->next = get_blame_suspects(commit);
  83                                 set_blame_suspects(commit, o);
  84                         }
  85                         return blame_origin_incref(o);
  86                 }
  87         }
  88         return make_origin(commit, path);
  89 }
  90
  91
  92
  93 static void verify_working_tree_path(struct repository *repo,
  94                                      struct commit *work_tree, const char *path)
  95 {
  96         struct commit_list *parents;
  97         int pos;
  98
  99         for (parents = work_tree->parents; parents; parents = parents->next) {
 100                 const struct object_id *commit_oid = &parents->item->object.oid;
 101                 struct object_id blob_oid;
 102                 unsigned mode;
 103
 104                 if (!get_tree_entry(commit_oid, path, &blob_oid, &mode) &&
 105                     oid_object_info(repo, &blob_oid, NULL) == OBJ_BLOB)
 106                         return;
 107         }
 108
 109         pos = index_name_pos(repo->index, path, strlen(path));
 110         if (pos >= 0)
 111                 ; /* path is in the index */
 112         else if (-1 - pos < repo->index->cache_nr &&
 113                  !strcmp(repo->index->cache[-1 - pos]->name, path))
 114                 ; /* path is in the index, unmerged */
 115         else
 116                 die("no such path '%s' in HEAD", path);
 117 }
 118
 119 static struct commit_list **append_parent(struct commit_list **tail, const struct object_id *oid)
 120 {
 121         struct commit *parent;
 122
 123         parent = lookup_commit_reference(the_repository, oid);
 124         if (!parent)
 125                 die("no such commit %s", oid_to_hex(oid));
 126         return &commit_list_insert(parent, tail)->next;
 127 }
 128
 129 static void append_merge_parents(struct commit_list **tail)
 130 {
 131         int merge_head;
 132         struct strbuf line = STRBUF_INIT;
 133
 134         merge_head = open(git_path_merge_head(the_repository), O_RDONLY);
 135         if (merge_head < 0) {
 136                 if (errno == ENOENT)
 137                         return;
 138                 die("cannot open '%s' for reading",
 139                     git_path_merge_head(the_repository));
 140         }
 141
 142         while (!strbuf_getwholeline_fd(&line, merge_head, '\n')) {
 143                 struct object_id oid;
 144                 if (line.len < GIT_SHA1_HEXSZ || get_oid_hex(line.buf, &oid))
 145                         die("unknown line in '%s': %s",
 146                             git_path_merge_head(the_repository), line.buf);
 147                 tail = append_parent(tail, &oid);
 148         }
 149         close(merge_head);
 150         strbuf_release(&line);
 151 }
 152
 153 /*
 154  * This isn't as simple as passing sb->buf and sb->len, because we
 155  * want to transfer ownership of the buffer to the commit (so we
 156  * must use detach).
 157  */
 158 static void set_commit_buffer_from_strbuf(struct commit *c, struct strbuf *sb)
 159 {
 160         size_t len;
 161         void *buf = strbuf_detach(sb, &len);
 162         set_commit_buffer(the_repository, c, buf, len);
 163 }
 164
 165 /*
 166  * Prepare a dummy commit that represents the work tree (or staged) item.
 167  * Note that annotating work tree item never works in the reverse.
 168  */
 169 static struct commit *fake_working_tree_commit(struct repository *repo,
 170                                                struct diff_options *opt,
 171                                                const char *path,
 172                                                const char *contents_from)
 173 {
 174         struct commit *commit;
 175         struct blame_origin *origin;
 176         struct commit_list **parent_tail, *parent;
 177         struct object_id head_oid;
 178         struct strbuf buf = STRBUF_INIT;
 179         const char *ident;
 180         time_t now;
 181         int len;
 182         struct cache_entry *ce;
 183         unsigned mode;
 184         struct strbuf msg = STRBUF_INIT;
 185
 186         read_index(repo->index);
 187         time(&now);
 188         commit = alloc_commit_node(the_repository);
 189         commit->object.parsed = 1;
 190         commit->date = now;
 191         parent_tail = &commit->parents;
 192
 193         if (!resolve_ref_unsafe("HEAD", RESOLVE_REF_READING, &head_oid, NULL))
 194                 die("no such ref: HEAD");
 195
 196         parent_tail = append_parent(parent_tail, &head_oid);
 197         append_merge_parents(parent_tail);
 198         verify_working_tree_path(repo, commit, path);
 199
 200         origin = make_origin(commit, path);
 201
 202         ident = fmt_ident("Not Committed Yet", "not.committed.yet", NULL, 0);
 203         strbuf_addstr(&msg, "tree 0000000000000000000000000000000000000000\n");
 204         for (parent = commit->parents; parent; parent = parent->next)
 205                 strbuf_addf(&msg, "parent %s\n",
 206                             oid_to_hex(&parent->item->object.oid));
 207         strbuf_addf(&msg,
 208                     "author %s\n"
 209                     "committer %s\n\n"
 210                     "Version of %s from %s\n",
 211                     ident, ident, path,
 212                     (!contents_from ? path :
 213                      (!strcmp(contents_from, "-") ? "standard input" : contents_from)));
 214         set_commit_buffer_from_strbuf(commit, &msg);
 215
 216         if (!contents_from || strcmp("-", contents_from)) {
 217                 struct stat st;
 218                 const char *read_from;
 219                 char *buf_ptr;
 220                 unsigned long buf_len;
 221
 222                 if (contents_from) {
 223                         if (stat(contents_from, &st) < 0)
 224                                 die_errno("Cannot stat '%s'", contents_from);
 225                         read_from = contents_from;
 226                 }
 227                 else {
 228                         if (lstat(path, &st) < 0)
 229                                 die_errno("Cannot lstat '%s'", path);
 230                         read_from = path;
 231                 }
 232                 mode = canon_mode(st.st_mode);
 233
 234                 switch (st.st_mode & S_IFMT) {
 235                 case S_IFREG:
 236                         if (opt->flags.allow_textconv &&
 237                             textconv_object(read_from, mode, &null_oid, 0, &buf_ptr, &buf_len))
 238                                 strbuf_attach(&buf, buf_ptr, buf_len, buf_len + 1);
 239                         else if (strbuf_read_file(&buf, read_from, st.st_size) != st.st_size)
 240                                 die_errno("cannot open or read '%s'", read_from);
 241                         break;
 242                 case S_IFLNK:
 243                         if (strbuf_readlink(&buf, read_from, st.st_size) < 0)
 244                                 die_errno("cannot readlink '%s'", read_from);
 245                         break;
 246                 default:
 247                         die("unsupported file type %s", read_from);
 248                 }
 249         }
 250         else {
 251                 /* Reading from stdin */
 252                 mode = 0;
 253                 if (strbuf_read(&buf, 0, 0) < 0)
 254                         die_errno("failed to read from stdin");
 255         }
 256         convert_to_git(repo->index, path, buf.buf, buf.len, &buf, 0);
 257         origin->file.ptr = buf.buf;
 258         origin->file.size = buf.len;
 259         pretend_object_file(buf.buf, buf.len, OBJ_BLOB, &origin->blob_oid);
 260
 261         /*
 262          * Read the current index, replace the path entry with
 263          * origin->blob_sha1 without mucking with its mode or type
 264          * bits; we are not going to write this index out -- we just
 265          * want to run "diff-index --cached".
 266          */
 267         discard_index(repo->index);
 268         read_index(repo->index);
 269
 270         len = strlen(path);
 271         if (!mode) {
 272                 int pos = index_name_pos(repo->index, path, len);
 273                 if (0 <= pos)
 274                         mode = repo->index->cache[pos]->ce_mode;
 275                 else
 276                         /* Let's not bother reading from HEAD tree */
 277                         mode = S_IFREG | 0644;
 278         }
 279         ce = make_empty_cache_entry(repo->index, len);
 280         oidcpy(&ce->oid, &origin->blob_oid);
 281         memcpy(ce->name, path, len);
 282         ce->ce_flags = create_ce_flags(0);
 283         ce->ce_namelen = len;
 284         ce->ce_mode = create_ce_mode(mode);
 285         add_index_entry(repo->index, ce,
 286                         ADD_CACHE_OK_TO_ADD | ADD_CACHE_OK_TO_REPLACE);
 287
 288         cache_tree_invalidate_path(repo->index, path);
 289
 290         return commit;
 291 }
 292
 293
 294
 295 static int diff_hunks(mmfile_t *file_a, mmfile_t *file_b,
 296                       xdl_emit_hunk_consume_func_t hunk_func, void *cb_data, int xdl_opts)
 297 {
 298         xpparam_t xpp = {0};
 299         xdemitconf_t xecfg = {0};
 300         xdemitcb_t ecb = {NULL};
 301
 302         xpp.flags = xdl_opts;
 303         xecfg.hunk_func = hunk_func;
 304         ecb.priv = cb_data;
 305         return xdi_diff(file_a, file_b, &xpp, &xecfg, &ecb);
 306 }
 307
 308 /*
 309  * Given an origin, prepare mmfile_t structure to be used by the
 310  * diff machinery
 311  */
 312 static void fill_origin_blob(struct diff_options *opt,
 313                              struct blame_origin *o, mmfile_t *file, int *num_read_blob)
 314 {
 315         if (!o->file.ptr) {
 316                 enum object_type type;
 317                 unsigned long file_size;
 318
 319                 (*num_read_blob)++;
 320                 if (opt->flags.allow_textconv &&
 321                     textconv_object(o->path, o->mode, &o->blob_oid, 1, &file->ptr, &file_size))
 322                         ;
 323                 else
 324                         file->ptr = read_object_file(&o->blob_oid, &type,
 325                                                      &file_size);
 326                 file->size = file_size;
 327
 328                 if (!file->ptr)
 329                         die("Cannot read blob %s for path %s",
 330                             oid_to_hex(&o->blob_oid),
 331                             o->path);
 332                 o->file = *file;
 333         }
 334         else
 335                 *file = o->file;
 336 }
 337
 338 static void drop_origin_blob(struct blame_origin *o)
 339 {
 340         FREE_AND_NULL(o->file.ptr);
 341 }
 342
 343 /*
 344  * Any merge of blames happens on lists of blames that arrived via
 345  * different parents in a single suspect.  In this case, we want to
 346  * sort according to the suspect line numbers as opposed to the final
 347  * image line numbers.  The function body is somewhat longish because
 348  * it avoids unnecessary writes.
 349  */
 350
 351 static struct blame_entry *blame_merge(struct blame_entry *list1,
 352                                        struct blame_entry *list2)
 353 {
 354         struct blame_entry *p1 = list1, *p2 = list2,
 355                 **tail = &list1;
 356
 357         if (!p1)
 358                 return p2;
 359         if (!p2)
 360                 return p1;
 361
 362         if (p1->s_lno <= p2->s_lno) {
 363                 do {
 364                         tail = &p1->next;
 365                         if ((p1 = *tail) == NULL) {
 366                                 *tail = p2;
 367                                 return list1;
 368                         }
 369                 } while (p1->s_lno <= p2->s_lno);
 370         }
 371         for (;;) {
 372                 *tail = p2;
 373                 do {
 374                         tail = &p2->next;
 375                         if ((p2 = *tail) == NULL)  {
 376                                 *tail = p1;
 377                                 return list1;
 378                         }
 379                 } while (p1->s_lno > p2->s_lno);
 380                 *tail = p1;
 381                 do {
 382                         tail = &p1->next;
 383                         if ((p1 = *tail) == NULL) {
 384                                 *tail = p2;
 385                                 return list1;
 386                         }
 387                 } while (p1->s_lno <= p2->s_lno);
 388         }
 389 }
 390
 391 static void *get_next_blame(const void *p)
 392 {
 393         return ((struct blame_entry *)p)->next;
 394 }
 395
 396 static void set_next_blame(void *p1, void *p2)
 397 {
 398         ((struct blame_entry *)p1)->next = p2;
 399 }
 400
 401 /*
 402  * Final image line numbers are all different, so we don't need a
 403  * three-way comparison here.
 404  */
 405
 406 static int compare_blame_final(const void *p1, const void *p2)
 407 {
 408         return ((struct blame_entry *)p1)->lno > ((struct blame_entry *)p2)->lno
 409                 ? 1 : -1;
 410 }
 411
 412 static int compare_blame_suspect(const void *p1, const void *p2)
 413 {
 414         const struct blame_entry *s1 = p1, *s2 = p2;
 415         /*
 416          * to allow for collating suspects, we sort according to the
 417          * respective pointer value as the primary sorting criterion.
 418          * The actual relation is pretty unimportant as long as it
 419          * establishes a total order.  Comparing as integers gives us
 420          * that.
 421          */
 422         if (s1->suspect != s2->suspect)
 423                 return (intptr_t)s1->suspect > (intptr_t)s2->suspect ? 1 : -1;
 424         if (s1->s_lno == s2->s_lno)
 425                 return 0;
 426         return s1->s_lno > s2->s_lno ? 1 : -1;
 427 }
 428
 429 void blame_sort_final(struct blame_scoreboard *sb)
 430 {
 431         sb->ent = llist_mergesort(sb->ent, get_next_blame, set_next_blame,
 432                                   compare_blame_final);
 433 }
 434
 435 static int compare_commits_by_reverse_commit_date(const void *a,
 436                                                   const void *b,
 437                                                   void *c)
 438 {
 439         return -compare_commits_by_commit_date(a, b, c);
 440 }
 441
 442 /*
 443  * For debugging -- origin is refcounted, and this asserts that
 444  * we do not underflow.
 445  */
 446 static void sanity_check_refcnt(struct blame_scoreboard *sb)
 447 {
 448         int baa = 0;
 449         struct blame_entry *ent;
 450
 451         for (ent = sb->ent; ent; ent = ent->next) {
 452                 /* Nobody should have zero or negative refcnt */
 453                 if (ent->suspect->refcnt <= 0) {
 454                         fprintf(stderr, "%s in %s has negative refcnt %d\n",
 455                                 ent->suspect->path,
 456                                 oid_to_hex(&ent->suspect->commit->object.oid),
 457                                 ent->suspect->refcnt);
 458                         baa = 1;
 459                 }
 460         }
 461         if (baa)
 462                 sb->on_sanity_fail(sb, baa);
 463 }
 464
 465 /*
 466  * If two blame entries that are next to each other came from
 467  * contiguous lines in the same origin (i.e. <commit, path> pair),
 468  * merge them together.
 469  */
 470 void blame_coalesce(struct blame_scoreboard *sb)
 471 {
 472         struct blame_entry *ent, *next;
 473
 474         for (ent = sb->ent; ent && (next = ent->next); ent = next) {
 475                 if (ent->suspect == next->suspect &&
 476                     ent->s_lno + ent->num_lines == next->s_lno) {
 477                         ent->num_lines += next->num_lines;
 478                         ent->next = next->next;
 479                         blame_origin_decref(next->suspect);
 480                         free(next);
 481                         ent->score = 0;
 482                         next = ent; /* again */
 483                 }
 484         }
 485
 486         if (sb->debug) /* sanity */
 487                 sanity_check_refcnt(sb);
 488 }
 489
 490 /*
 491  * Merge the given sorted list of blames into a preexisting origin.
 492  * If there were no previous blames to that commit, it is entered into
 493  * the commit priority queue of the score board.
 494  */
 495
 496 static void queue_blames(struct blame_scoreboard *sb, struct blame_origin *porigin,
 497                          struct blame_entry *sorted)
 498 {
 499         if (porigin->suspects)
 500                 porigin->suspects = blame_merge(porigin->suspects, sorted);
 501         else {
 502                 struct blame_origin *o;
 503                 for (o = get_blame_suspects(porigin->commit); o; o = o->next) {
 504                         if (o->suspects) {
 505                                 porigin->suspects = sorted;
 506                                 return;
 507                         }
 508                 }
 509                 porigin->suspects = sorted;
 510                 prio_queue_put(&sb->commits, porigin->commit);
 511         }
 512 }
 513
 514 /*
 515  * Fill the blob_sha1 field of an origin if it hasn't, so that later
 516  * call to fill_origin_blob() can use it to locate the data.  blob_sha1
 517  * for an origin is also used to pass the blame for the entire file to
 518  * the parent to detect the case where a child's blob is identical to
 519  * that of its parent's.
 520  *
 521  * This also fills origin->mode for corresponding tree path.
 522  */
 523 static int fill_blob_sha1_and_mode(struct repository *repo,
 524                                    struct blame_origin *origin)
 525 {
 526         if (!is_null_oid(&origin->blob_oid))
 527                 return 0;
 528         if (get_tree_entry(&origin->commit->object.oid, origin->path, &origin->blob_oid, &origin->mode))
 529                 goto error_out;
 530         if (oid_object_info(repo, &origin->blob_oid, NULL) != OBJ_BLOB)
 531                 goto error_out;
 532         return 0;
 533  error_out:
 534         oidclr(&origin->blob_oid);
 535         origin->mode = S_IFINVALID;
 536         return -1;
 537 }
 538
 539 /*
 540  * We have an origin -- check if the same path exists in the
 541  * parent and return an origin structure to represent it.
 542  */
 543 static struct blame_origin *find_origin(struct commit *parent,
 544                                   struct blame_origin *origin)
 545 {
 546         struct blame_origin *porigin;
 547         struct diff_options diff_opts;
 548         const char *paths[2];
 549
 550         /* First check any existing origins */
 551         for (porigin = get_blame_suspects(parent); porigin; porigin = porigin->next)
 552                 if (!strcmp(porigin->path, origin->path)) {
 553                         /*
 554                          * The same path between origin and its parent
 555                          * without renaming -- the most common case.
 556                          */
 557                         return blame_origin_incref (porigin);
 558                 }
 559
 560         /* See if the origin->path is different between parent
 561          * and origin first.  Most of the time they are the
 562          * same and diff-tree is fairly efficient about this.
 563          */
 564         diff_setup(&diff_opts);
 565         diff_opts.flags.recursive = 1;
 566         diff_opts.detect_rename = 0;
 567         diff_opts.output_format = DIFF_FORMAT_NO_OUTPUT;
 568         paths[0] = origin->path;
 569         paths[1] = NULL;
 570
 571         parse_pathspec(&diff_opts.pathspec,
 572                        PATHSPEC_ALL_MAGIC & ~PATHSPEC_LITERAL,
 573                        PATHSPEC_LITERAL_PATH, "", paths);
 574         diff_setup_done(&diff_opts);
 575
 576         if (is_null_oid(&origin->commit->object.oid))
 577                 do_diff_cache(get_commit_tree_oid(parent), &diff_opts);
 578         else
 579                 diff_tree_oid(get_commit_tree_oid(parent),
 580                               get_commit_tree_oid(origin->commit),
 581                               "", &diff_opts);
 582         diffcore_std(&diff_opts);
 583
 584         if (!diff_queued_diff.nr) {
 585                 /* The path is the same as parent */
 586                 porigin = get_origin(parent, origin->path);
 587                 oidcpy(&porigin->blob_oid, &origin->blob_oid);
 588                 porigin->mode = origin->mode;
 589         } else {
 590                 /*
 591                  * Since origin->path is a pathspec, if the parent
 592                  * commit had it as a directory, we will see a whole
 593                  * bunch of deletion of files in the directory that we
 594                  * do not care about.
 595                  */
 596                 int i;
 597                 struct diff_filepair *p = NULL;
 598                 for (i = 0; i < diff_queued_diff.nr; i++) {
 599                         const char *name;
 600                         p = diff_queued_diff.queue[i];
 601                         name = p->one->path ? p->one->path : p->two->path;
 602                         if (!strcmp(name, origin->path))
 603                                 break;
 604                 }
 605                 if (!p)
 606                         die("internal error in blame::find_origin");
 607                 switch (p->status) {
 608                 default:
 609                         die("internal error in blame::find_origin (%c)",
 610                             p->status);
 611                 case 'M':
 612                         porigin = get_origin(parent, origin->path);
 613                         oidcpy(&porigin->blob_oid, &p->one->oid);
 614                         porigin->mode = p->one->mode;
 615                         break;
 616                 case 'A':
 617                 case 'T':
 618                         /* Did not exist in parent, or type changed */
 619                         break;
 620                 }
 621         }
 622         diff_flush(&diff_opts);
 623         clear_pathspec(&diff_opts.pathspec);
 624         return porigin;
 625 }
 626
 627 /*
 628  * We have an origin -- find the path that corresponds to it in its
 629  * parent and return an origin structure to represent it.
 630  */
 631 static struct blame_origin *find_rename(struct commit *parent,
 632                                   struct blame_origin *origin)
 633 {
 634         struct blame_origin *porigin = NULL;
 635         struct diff_options diff_opts;
 636         int i;
 637
 638         diff_setup(&diff_opts);
 639         diff_opts.flags.recursive = 1;
 640         diff_opts.detect_rename = DIFF_DETECT_RENAME;
 641         diff_opts.output_format = DIFF_FORMAT_NO_OUTPUT;
 642         diff_opts.single_follow = origin->path;
 643         diff_setup_done(&diff_opts);
 644
 645         if (is_null_oid(&origin->commit->object.oid))
 646                 do_diff_cache(get_commit_tree_oid(parent), &diff_opts);
 647         else
 648                 diff_tree_oid(get_commit_tree_oid(parent),
 649                               get_commit_tree_oid(origin->commit),
 650                               "", &diff_opts);
 651         diffcore_std(&diff_opts);
 652
 653         for (i = 0; i < diff_queued_diff.nr; i++) {
 654                 struct diff_filepair *p = diff_queued_diff.queue[i];
 655                 if ((p->status == 'R' || p->status == 'C') &&
 656                     !strcmp(p->two->path, origin->path)) {
 657                         porigin = get_origin(parent, p->one->path);
 658                         oidcpy(&porigin->blob_oid, &p->one->oid);
 659                         porigin->mode = p->one->mode;
 660                         break;
 661                 }
 662         }
 663         diff_flush(&diff_opts);
 664         clear_pathspec(&diff_opts.pathspec);
 665         return porigin;
 666 }
 667
 668 /*
 669  * Append a new blame entry to a given output queue.
 670  */
 671 static void add_blame_entry(struct blame_entry ***queue,
 672                             const struct blame_entry *src)
 673 {
 674         struct blame_entry *e = xmalloc(sizeof(*e));
 675         memcpy(e, src, sizeof(*e));
 676         blame_origin_incref(e->suspect);
 677
 678         e->next = **queue;
 679         **queue = e;
 680         *queue = &e->next;
 681 }
 682
 683 /*
 684  * src typically is on-stack; we want to copy the information in it to
 685  * a malloced blame_entry that gets added to the given queue.  The
 686  * origin of dst loses a refcnt.
 687  */
 688 static void dup_entry(struct blame_entry ***queue,
 689                       struct blame_entry *dst, struct blame_entry *src)
 690 {
 691         blame_origin_incref(src->suspect);
 692         blame_origin_decref(dst->suspect);
 693         memcpy(dst, src, sizeof(*src));
 694         dst->next = **queue;
 695         **queue = dst;
 696         *queue = &dst->next;
 697 }
 698
 699 const char *blame_nth_line(struct blame_scoreboard *sb, long lno)
 700 {
 701         return sb->final_buf + sb->lineno[lno];
 702 }
 703
 704 /*
 705  * It is known that lines between tlno to same came from parent, and e
 706  * has an overlap with that range.  it also is known that parent's
 707  * line plno corresponds to e's line tlno.
 708  *
 709  *                <---- e ----->
 710  *                   <------>
 711  *                   <------------>
 712  *             <------------>
 713  *             <------------------>
 714  *
 715  * Split e into potentially three parts; before this chunk, the chunk
 716  * to be blamed for the parent, and after that portion.
 717  */
 718 static void split_overlap(struct blame_entry *split,
 719                           struct blame_entry *e,
 720                           int tlno, int plno, int same,
 721                           struct blame_origin *parent)
 722 {
 723         int chunk_end_lno;
 724         memset(split, 0, sizeof(struct blame_entry [3]));
 725
 726         if (e->s_lno < tlno) {
 727                 /* there is a pre-chunk part not blamed on parent */
 728                 split[0].suspect = blame_origin_incref(e->suspect);
 729                 split[0].lno = e->lno;
 730                 split[0].s_lno = e->s_lno;
 731                 split[0].num_lines = tlno - e->s_lno;
 732                 split[1].lno = e->lno + tlno - e->s_lno;
 733                 split[1].s_lno = plno;
 734         }
 735         else {
 736                 split[1].lno = e->lno;
 737                 split[1].s_lno = plno + (e->s_lno - tlno);
 738         }
 739
 740         if (same < e->s_lno + e->num_lines) {
 741                 /* there is a post-chunk part not blamed on parent */
 742                 split[2].suspect = blame_origin_incref(e->suspect);
 743                 split[2].lno = e->lno + (same - e->s_lno);
 744                 split[2].s_lno = e->s_lno + (same - e->s_lno);
 745                 split[2].num_lines = e->s_lno + e->num_lines - same;
 746                 chunk_end_lno = split[2].lno;
 747         }
 748         else
 749                 chunk_end_lno = e->lno + e->num_lines;
 750         split[1].num_lines = chunk_end_lno - split[1].lno;
 751
 752         /*
 753          * if it turns out there is nothing to blame the parent for,
 754          * forget about the splitting.  !split[1].suspect signals this.
 755          */
 756         if (split[1].num_lines < 1)
 757                 return;
 758         split[1].suspect = blame_origin_incref(parent);
 759 }
 760
 761 /*
 762  * split_overlap() divided an existing blame e into up to three parts
 763  * in split.  Any assigned blame is moved to queue to
 764  * reflect the split.
 765  */
 766 static void split_blame(struct blame_entry ***blamed,
 767                         struct blame_entry ***unblamed,
 768                         struct blame_entry *split,
 769                         struct blame_entry *e)
 770 {
 771         if (split[0].suspect && split[2].suspect) {
 772                 /* The first part (reuse storage for the existing entry e) */
 773                 dup_entry(unblamed, e, &split[0]);
 774
 775                 /* The last part -- me */
 776                 add_blame_entry(unblamed, &split[2]);
 777
 778                 /* ... and the middle part -- parent */
 779                 add_blame_entry(blamed, &split[1]);
 780         }
 781         else if (!split[0].suspect && !split[2].suspect)
 782                 /*
 783                  * The parent covers the entire area; reuse storage for
 784                  * e and replace it with the parent.
 785                  */
 786                 dup_entry(blamed, e, &split[1]);
 787         else if (split[0].suspect) {
 788                 /* me and then parent */
 789                 dup_entry(unblamed, e, &split[0]);
 790                 add_blame_entry(blamed, &split[1]);
 791         }
 792         else {
 793                 /* parent and then me */
 794                 dup_entry(blamed, e, &split[1]);
 795                 add_blame_entry(unblamed, &split[2]);
 796         }
 797 }
 798
 799 /*
 800  * After splitting the blame, the origins used by the
 801  * on-stack blame_entry should lose one refcnt each.
 802  */
 803 static void decref_split(struct blame_entry *split)
 804 {
 805         int i;
 806
 807         for (i = 0; i < 3; i++)
 808                 blame_origin_decref(split[i].suspect);
 809 }
 810
 811 /*
 812  * reverse_blame reverses the list given in head, appending tail.
 813  * That allows us to build lists in reverse order, then reverse them
 814  * afterwards.  This can be faster than building the list in proper
 815  * order right away.  The reason is that building in proper order
 816  * requires writing a link in the _previous_ element, while building
 817  * in reverse order just requires placing the list head into the
 818  * _current_ element.
 819  */
 820
 821 static struct blame_entry *reverse_blame(struct blame_entry *head,
 822                                          struct blame_entry *tail)
 823 {
 824         while (head) {
 825                 struct blame_entry *next = head->next;
 826                 head->next = tail;
 827                 tail = head;
 828                 head = next;
 829         }
 830         return tail;
 831 }
 832
 833 /*
 834  * Process one hunk from the patch between the current suspect for
 835  * blame_entry e and its parent.  This first blames any unfinished
 836  * entries before the chunk (which is where target and parent start
 837  * differing) on the parent, and then splits blame entries at the
 838  * start and at the end of the difference region.  Since use of -M and
 839  * -C options may lead to overlapping/duplicate source line number
 840  * ranges, all we can rely on from sorting/merging is the order of the
 841  * first suspect line number.
 842  */
 843 static void blame_chunk(struct blame_entry ***dstq, struct blame_entry ***srcq,
 844                         int tlno, int offset, int same,
 845                         struct blame_origin *parent)
 846 {
 847         struct blame_entry *e = **srcq;
 848         struct blame_entry *samep = NULL, *diffp = NULL;
 849
 850         while (e && e->s_lno < tlno) {
 851                 struct blame_entry *next = e->next;
 852                 /*
 853                  * current record starts before differing portion.  If
 854                  * it reaches into it, we need to split it up and
 855                  * examine the second part separately.
 856                  */
 857                 if (e->s_lno + e->num_lines > tlno) {
 858                         /* Move second half to a new record */
 859                         int len = tlno - e->s_lno;
 860                         struct blame_entry *n = xcalloc(1, sizeof (struct blame_entry));
 861                         n->suspect = e->suspect;
 862                         n->lno = e->lno + len;
 863                         n->s_lno = e->s_lno + len;
 864                         n->num_lines = e->num_lines - len;
 865                         e->num_lines = len;
 866                         e->score = 0;
 867                         /* Push new record to diffp */
 868                         n->next = diffp;
 869                         diffp = n;
 870                 } else
 871                         blame_origin_decref(e->suspect);
 872                 /* Pass blame for everything before the differing
 873                  * chunk to the parent */
 874                 e->suspect = blame_origin_incref(parent);
 875                 e->s_lno += offset;
 876                 e->next = samep;
 877                 samep = e;
 878                 e = next;
 879         }
 880         /*
 881          * As we don't know how much of a common stretch after this
 882          * diff will occur, the currently blamed parts are all that we
 883          * can assign to the parent for now.
 884          */
 885
 886         if (samep) {
 887                 **dstq = reverse_blame(samep, **dstq);
 888                 *dstq = &samep->next;
 889         }
 890         /*
 891          * Prepend the split off portions: everything after e starts
 892          * after the blameable portion.
 893          */
 894         e = reverse_blame(diffp, e);
 895
 896         /*
 897          * Now retain records on the target while parts are different
 898          * from the parent.
 899          */
 900         samep = NULL;
 901         diffp = NULL;
 902         while (e && e->s_lno < same) {
 903                 struct blame_entry *next = e->next;
 904
 905                 /*
 906                  * If current record extends into sameness, need to split.
 907                  */
 908                 if (e->s_lno + e->num_lines > same) {
 909                         /*
 910                          * Move second half to a new record to be
 911                          * processed by later chunks
 912                          */
 913                         int len = same - e->s_lno;
 914                         struct blame_entry *n = xcalloc(1, sizeof (struct blame_entry));
 915                         n->suspect = blame_origin_incref(e->suspect);
 916                         n->lno = e->lno + len;
 917                         n->s_lno = e->s_lno + len;
 918                         n->num_lines = e->num_lines - len;
 919                         e->num_lines = len;
 920                         e->score = 0;
 921                         /* Push new record to samep */
 922                         n->next = samep;
 923                         samep = n;
 924                 }
 925                 e->next = diffp;
 926                 diffp = e;
 927                 e = next;
 928         }
 929         **srcq = reverse_blame(diffp, reverse_blame(samep, e));
 930         /* Move across elements that are in the unblamable portion */
 931         if (diffp)
 932                 *srcq = &diffp->next;
 933 }
 934
 935 struct blame_chunk_cb_data {
 936         struct blame_origin *parent;
 937         long offset;
 938         struct blame_entry **dstq;
 939         struct blame_entry **srcq;
 940 };
 941
 942 /* diff chunks are from parent to target */
 943 static int blame_chunk_cb(long start_a, long count_a,
 944                           long start_b, long count_b, void *data)
 945 {
 946         struct blame_chunk_cb_data *d = data;
 947         if (start_a - start_b != d->offset)
 948                 die("internal error in blame::blame_chunk_cb");
 949         blame_chunk(&d->dstq, &d->srcq, start_b, start_a - start_b,
 950                     start_b + count_b, d->parent);
 951         d->offset = start_a + count_a - (start_b + count_b);
 952         return 0;
 953 }
 954
 955 /*
 956  * We are looking at the origin 'target' and aiming to pass blame
 957  * for the lines it is suspected to its parent.  Run diff to find
 958  * which lines came from parent and pass blame for them.
 959  */
 960 static void pass_blame_to_parent(struct blame_scoreboard *sb,
 961                                  struct blame_origin *target,
 962                                  struct blame_origin *parent)
 963 {
 964         mmfile_t file_p, file_o;
 965         struct blame_chunk_cb_data d;
 966         struct blame_entry *newdest = NULL;
 967
 968         if (!target->suspects)
 969                 return; /* nothing remains for this target */
 970
 971         d.parent = parent;
 972         d.offset = 0;
 973         d.dstq = &newdest; d.srcq = &target->suspects;
 974
 975         fill_origin_blob(&sb->revs->diffopt, parent, &file_p, &sb->num_read_blob);
 976         fill_origin_blob(&sb->revs->diffopt, target, &file_o, &sb->num_read_blob);
 977         sb->num_get_patch++;
 978
 979         if (diff_hunks(&file_p, &file_o, blame_chunk_cb, &d, sb->xdl_opts))
 980                 die("unable to generate diff (%s -> %s)",
 981                     oid_to_hex(&parent->commit->object.oid),
 982                     oid_to_hex(&target->commit->object.oid));
 983         /* The rest are the same as the parent */
 984         blame_chunk(&d.dstq, &d.srcq, INT_MAX, d.offset, INT_MAX, parent);
 985         *d.dstq = NULL;
 986         queue_blames(sb, parent, newdest);
 987
 988         return;
 989 }
 990
 991 /*
 992  * The lines in blame_entry after splitting blames many times can become
 993  * very small and trivial, and at some point it becomes pointless to
 994  * blame the parents.  E.g. "\t\t}\n\t}\n\n" appears everywhere in any
 995  * ordinary C program, and it is not worth to say it was copied from
 996  * totally unrelated file in the parent.
 997  *
 998  * Compute how trivial the lines in the blame_entry are.
 999  */
1000 unsigned blame_entry_score(struct blame_scoreboard *sb, struct blame_entry *e)
1001 {
1002         unsigned score;
1003         const char *cp, *ep;
1004
1005         if (e->score)
1006                 return e->score;
1007
1008         score = 1;
1009         cp = blame_nth_line(sb, e->lno);
1010         ep = blame_nth_line(sb, e->lno + e->num_lines);
1011         while (cp < ep) {
1012                 unsigned ch = *((unsigned char *)cp);
1013                 if (isalnum(ch))
1014                         score++;
1015                 cp++;
1016         }
1017         e->score = score;
1018         return score;
1019 }
1020
1021 /*
1022  * best_so_far[] and potential[] are both a split of an existing blame_entry
1023  * that passes blame to the parent.  Maintain best_so_far the best split so
1024  * far, by comparing potential and best_so_far and copying potential into
1025  * bst_so_far as needed.
1026  */
1027 static void copy_split_if_better(struct blame_scoreboard *sb,
1028                                  struct blame_entry *best_so_far,
1029                                  struct blame_entry *potential)
1030 {
1031         int i;
1032
1033         if (!potential[1].suspect)
1034                 return;
1035         if (best_so_far[1].suspect) {
1036                 if (blame_entry_score(sb, &potential[1]) <
1037                     blame_entry_score(sb, &best_so_far[1]))
1038                         return;
1039         }
1040
1041         for (i = 0; i < 3; i++)
1042                 blame_origin_incref(potential[i].suspect);
1043         decref_split(best_so_far);
1044         memcpy(best_so_far, potential, sizeof(struct blame_entry[3]));
1045 }
1046
1047 /*
1048  * We are looking at a part of the final image represented by
1049  * ent (tlno and same are offset by ent->s_lno).
1050  * tlno is where we are looking at in the final image.
1051  * up to (but not including) same match preimage.
1052  * plno is where we are looking at in the preimage.
1053  *
1054  * <-------------- final image ---------------------->
1055  *       <------ent------>
1056  *         ^tlno ^same
1057  *    <---------preimage----->
1058  *         ^plno
1059  *
1060  * All line numbers are 0-based.
1061  */
1062 static void handle_split(struct blame_scoreboard *sb,
1063                          struct blame_entry *ent,
1064                          int tlno, int plno, int same,
1065                          struct blame_origin *parent,
1066                          struct blame_entry *split)
1067 {
1068         if (ent->num_lines <= tlno)
1069                 return;
1070         if (tlno < same) {
1071                 struct blame_entry potential[3];
1072                 tlno += ent->s_lno;
1073                 same += ent->s_lno;
1074                 split_overlap(potential, ent, tlno, plno, same, parent);
1075                 copy_split_if_better(sb, split, potential);
1076                 decref_split(potential);
1077         }
1078 }
1079
1080 struct handle_split_cb_data {
1081         struct blame_scoreboard *sb;
1082         struct blame_entry *ent;
1083         struct blame_origin *parent;
1084         struct blame_entry *split;
1085         long plno;
1086         long tlno;
1087 };
1088
1089 static int handle_split_cb(long start_a, long count_a,
1090                            long start_b, long count_b, void *data)
1091 {
1092         struct handle_split_cb_data *d = data;
1093         handle_split(d->sb, d->ent, d->tlno, d->plno, start_b, d->parent,
1094                      d->split);
1095         d->plno = start_a + count_a;
1096         d->tlno = start_b + count_b;
1097         return 0;
1098 }
1099
1100 /*
1101  * Find the lines from parent that are the same as ent so that
1102  * we can pass blames to it.  file_p has the blob contents for
1103  * the parent.
1104  */
1105 static void find_copy_in_blob(struct blame_scoreboard *sb,
1106                               struct blame_entry *ent,
1107                               struct blame_origin *parent,
1108                               struct blame_entry *split,
1109                               mmfile_t *file_p)
1110 {
1111         const char *cp;
1112         mmfile_t file_o;
1113         struct handle_split_cb_data d;
1114
1115         memset(&d, 0, sizeof(d));
1116         d.sb = sb; d.ent = ent; d.parent = parent; d.split = split;
1117         /*
1118          * Prepare mmfile that contains only the lines in ent.
1119          */
1120         cp = blame_nth_line(sb, ent->lno);
1121         file_o.ptr = (char *) cp;
1122         file_o.size = blame_nth_line(sb, ent->lno + ent->num_lines) - cp;
1123
1124         /*
1125          * file_o is a part of final image we are annotating.
1126          * file_p partially may match that image.
1127          */
1128         memset(split, 0, sizeof(struct blame_entry [3]));
1129         if (diff_hunks(file_p, &file_o, handle_split_cb, &d, sb->xdl_opts))
1130                 die("unable to generate diff (%s)",
1131                     oid_to_hex(&parent->commit->object.oid));
1132         /* remainder, if any, all match the preimage */
1133         handle_split(sb, ent, d.tlno, d.plno, ent->num_lines, parent, split);
1134 }
1135
1136 /* Move all blame entries from list *source that have a score smaller
1137  * than score_min to the front of list *small.
1138  * Returns a pointer to the link pointing to the old head of the small list.
1139  */
1140
1141 static struct blame_entry **filter_small(struct blame_scoreboard *sb,
1142                                          struct blame_entry **small,
1143                                          struct blame_entry **source,
1144                                          unsigned score_min)
1145 {
1146         struct blame_entry *p = *source;
1147         struct blame_entry *oldsmall = *small;
1148         while (p) {
1149                 if (blame_entry_score(sb, p) <= score_min) {
1150                         *small = p;
1151                         small = &p->next;
1152                         p = *small;
1153                 } else {
1154                         *source = p;
1155                         source = &p->next;
1156                         p = *source;
1157                 }
1158         }
1159         *small = oldsmall;
1160         *source = NULL;
1161         return small;
1162 }
1163
1164 /*
1165  * See if lines currently target is suspected for can be attributed to
1166  * parent.
1167  */
1168 static void find_move_in_parent(struct blame_scoreboard *sb,
1169                                 struct blame_entry ***blamed,
1170                                 struct blame_entry **toosmall,
1171                                 struct blame_origin *target,
1172                                 struct blame_origin *parent)
1173 {
1174         struct blame_entry *e, split[3];
1175         struct blame_entry *unblamed = target->suspects;
1176         struct blame_entry *leftover = NULL;
1177         mmfile_t file_p;
1178
1179         if (!unblamed)
1180                 return; /* nothing remains for this target */
1181
1182         fill_origin_blob(&sb->revs->diffopt, parent, &file_p, &sb->num_read_blob);
1183         if (!file_p.ptr)
1184                 return;
1185
1186         /* At each iteration, unblamed has a NULL-terminated list of
1187          * entries that have not yet been tested for blame.  leftover
1188          * contains the reversed list of entries that have been tested
1189          * without being assignable to the parent.
1190          */
1191         do {
1192                 struct blame_entry **unblamedtail = &unblamed;
1193                 struct blame_entry *next;
1194                 for (e = unblamed; e; e = next) {
1195                         next = e->next;
1196                         find_copy_in_blob(sb, e, parent, split, &file_p);
1197                         if (split[1].suspect &&
1198                             sb->move_score < blame_entry_score(sb, &split[1])) {
1199                                 split_blame(blamed, &unblamedtail, split, e);
1200                         } else {
1201                                 e->next = leftover;
1202                                 leftover = e;
1203                         }
1204                         decref_split(split);
1205                 }
1206                 *unblamedtail = NULL;
1207                 toosmall = filter_small(sb, toosmall, &unblamed, sb->move_score);
1208         } while (unblamed);
1209         target->suspects = reverse_blame(leftover, NULL);
1210 }
1211
1212 struct blame_list {
1213         struct blame_entry *ent;
1214         struct blame_entry split[3];
1215 };
1216
1217 /*
1218  * Count the number of entries the target is suspected for,
1219  * and prepare a list of entry and the best split.
1220  */
1221 static struct blame_list *setup_blame_list(struct blame_entry *unblamed,
1222                                            int *num_ents_p)
1223 {
1224         struct blame_entry *e;
1225         int num_ents, i;
1226         struct blame_list *blame_list = NULL;
1227
1228         for (e = unblamed, num_ents = 0; e; e = e->next)
1229                 num_ents++;
1230         if (num_ents) {
1231                 blame_list = xcalloc(num_ents, sizeof(struct blame_list));
1232                 for (e = unblamed, i = 0; e; e = e->next)
1233                         blame_list[i++].ent = e;
1234         }
1235         *num_ents_p = num_ents;
1236         return blame_list;
1237 }
1238
1239 /*
1240  * For lines target is suspected for, see if we can find code movement
1241  * across file boundary from the parent commit.  porigin is the path
1242  * in the parent we already tried.
1243  */
1244 static void find_copy_in_parent(struct blame_scoreboard *sb,
1245                                 struct blame_entry ***blamed,
1246                                 struct blame_entry **toosmall,
1247                                 struct blame_origin *target,
1248                                 struct commit *parent,
1249                                 struct blame_origin *porigin,
1250                                 int opt)
1251 {
1252         struct diff_options diff_opts;
1253         int i, j;
1254         struct blame_list *blame_list;
1255         int num_ents;
1256         struct blame_entry *unblamed = target->suspects;
1257         struct blame_entry *leftover = NULL;
1258
1259         if (!unblamed)
1260                 return; /* nothing remains for this target */
1261
1262         diff_setup(&diff_opts);
1263         diff_opts.flags.recursive = 1;
1264         diff_opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1265
1266         diff_setup_done(&diff_opts);
1267
1268         /* Try "find copies harder" on new path if requested;
1269          * we do not want to use diffcore_rename() actually to
1270          * match things up; find_copies_harder is set only to
1271          * force diff_tree_oid() to feed all filepairs to diff_queue,
1272          * and this code needs to be after diff_setup_done(), which
1273          * usually makes find-copies-harder imply copy detection.
1274          */
1275         if ((opt & PICKAXE_BLAME_COPY_HARDEST)
1276             || ((opt & PICKAXE_BLAME_COPY_HARDER)
1277                 && (!porigin || strcmp(target->path, porigin->path))))
1278                 diff_opts.flags.find_copies_harder = 1;
1279
1280         if (is_null_oid(&target->commit->object.oid))
1281                 do_diff_cache(get_commit_tree_oid(parent), &diff_opts);
1282         else
1283                 diff_tree_oid(get_commit_tree_oid(parent),
1284                               get_commit_tree_oid(target->commit),
1285                               "", &diff_opts);
1286
1287         if (!diff_opts.flags.find_copies_harder)
1288                 diffcore_std(&diff_opts);
1289
1290         do {
1291                 struct blame_entry **unblamedtail = &unblamed;
1292                 blame_list = setup_blame_list(unblamed, &num_ents);
1293
1294                 for (i = 0; i < diff_queued_diff.nr; i++) {
1295                         struct diff_filepair *p = diff_queued_diff.queue[i];
1296                         struct blame_origin *norigin;
1297                         mmfile_t file_p;
1298                         struct blame_entry potential[3];
1299
1300                         if (!DIFF_FILE_VALID(p->one))
1301                                 continue; /* does not exist in parent */
1302                         if (S_ISGITLINK(p->one->mode))
1303                                 continue; /* ignore git links */
1304                         if (porigin && !strcmp(p->one->path, porigin->path))
1305                                 /* find_move already dealt with this path */
1306                                 continue;
1307
1308                         norigin = get_origin(parent, p->one->path);
1309                         oidcpy(&norigin->blob_oid, &p->one->oid);
1310                         norigin->mode = p->one->mode;
1311                         fill_origin_blob(&sb->revs->diffopt, norigin, &file_p, &sb->num_read_blob);
1312                         if (!file_p.ptr)
1313                                 continue;
1314
1315                         for (j = 0; j < num_ents; j++) {
1316                                 find_copy_in_blob(sb, blame_list[j].ent,
1317                                                   norigin, potential, &file_p);
1318                                 copy_split_if_better(sb, blame_list[j].split,
1319                                                      potential);
1320                                 decref_split(potential);
1321                         }
1322                         blame_origin_decref(norigin);
1323                 }
1324
1325                 for (j = 0; j < num_ents; j++) {
1326                         struct blame_entry *split = blame_list[j].split;
1327                         if (split[1].suspect &&
1328                             sb->copy_score < blame_entry_score(sb, &split[1])) {
1329                                 split_blame(blamed, &unblamedtail, split,
1330                                             blame_list[j].ent);
1331                         } else {
1332                                 blame_list[j].ent->next = leftover;
1333                                 leftover = blame_list[j].ent;
1334                         }
1335                         decref_split(split);
1336                 }
1337                 free(blame_list);
1338                 *unblamedtail = NULL;
1339                 toosmall = filter_small(sb, toosmall, &unblamed, sb->copy_score);
1340         } while (unblamed);
1341         target->suspects = reverse_blame(leftover, NULL);
1342         diff_flush(&diff_opts);
1343         clear_pathspec(&diff_opts.pathspec);
1344 }
1345
1346 /*
1347  * The blobs of origin and porigin exactly match, so everything
1348  * origin is suspected for can be blamed on the parent.
1349  */
1350 static void pass_whole_blame(struct blame_scoreboard *sb,
1351                              struct blame_origin *origin, struct blame_origin *porigin)
1352 {
1353         struct blame_entry *e, *suspects;
1354
1355         if (!porigin->file.ptr && origin->file.ptr) {
1356                 /* Steal its file */
1357                 porigin->file = origin->file;
1358                 origin->file.ptr = NULL;
1359         }
1360         suspects = origin->suspects;
1361         origin->suspects = NULL;
1362         for (e = suspects; e; e = e->next) {
1363                 blame_origin_incref(porigin);
1364                 blame_origin_decref(e->suspect);
1365                 e->suspect = porigin;
1366         }
1367         queue_blames(sb, porigin, suspects);
1368 }
1369
1370 /*
1371  * We pass blame from the current commit to its parents.  We keep saying
1372  * "parent" (and "porigin"), but what we mean is to find scapegoat to
1373  * exonerate ourselves.
1374  */
1375 static struct commit_list *first_scapegoat(struct rev_info *revs, struct commit *commit,
1376                                         int reverse)
1377 {
1378         if (!reverse) {
1379                 if (revs->first_parent_only &&
1380                     commit->parents &&
1381                     commit->parents->next) {
1382                         free_commit_list(commit->parents->next);
1383                         commit->parents->next = NULL;
1384                 }
1385                 return commit->parents;
1386         }
1387         return lookup_decoration(&revs->children, &commit->object);
1388 }
1389
1390 static int num_scapegoats(struct rev_info *revs, struct commit *commit, int reverse)
1391 {
1392         struct commit_list *l = first_scapegoat(revs, commit, reverse);
1393         return commit_list_count(l);
1394 }
1395
1396 /* Distribute collected unsorted blames to the respected sorted lists
1397  * in the various origins.
1398  */
1399 static void distribute_blame(struct blame_scoreboard *sb, struct blame_entry *blamed)
1400 {
1401         blamed = llist_mergesort(blamed, get_next_blame, set_next_blame,
1402                                  compare_blame_suspect);
1403         while (blamed)
1404         {
1405                 struct blame_origin *porigin = blamed->suspect;
1406                 struct blame_entry *suspects = NULL;
1407                 do {
1408                         struct blame_entry *next = blamed->next;
1409                         blamed->next = suspects;
1410                         suspects = blamed;
1411                         blamed = next;
1412                 } while (blamed && blamed->suspect == porigin);
1413                 suspects = reverse_blame(suspects, NULL);
1414                 queue_blames(sb, porigin, suspects);
1415         }
1416 }
1417
1418 #define MAXSG 16
1419
1420 static void pass_blame(struct blame_scoreboard *sb, struct blame_origin *origin, int opt)
1421 {
1422         struct rev_info *revs = sb->revs;
1423         int i, pass, num_sg;
1424         struct commit *commit = origin->commit;
1425         struct commit_list *sg;
1426         struct blame_origin *sg_buf[MAXSG];
1427         struct blame_origin *porigin, **sg_origin = sg_buf;
1428         struct blame_entry *toosmall = NULL;
1429         struct blame_entry *blames, **blametail = &blames;
1430
1431         num_sg = num_scapegoats(revs, commit, sb->reverse);
1432         if (!num_sg)
1433                 goto finish;
1434         else if (num_sg < ARRAY_SIZE(sg_buf))
1435                 memset(sg_buf, 0, sizeof(sg_buf));
1436         else
1437                 sg_origin = xcalloc(num_sg, sizeof(*sg_origin));
1438
1439         /*
1440          * The first pass looks for unrenamed path to optimize for
1441          * common cases, then we look for renames in the second pass.
1442          */
1443         for (pass = 0; pass < 2 - sb->no_whole_file_rename; pass++) {
1444                 struct blame_origin *(*find)(struct commit *, struct blame_origin *);
1445                 find = pass ? find_rename : find_origin;
1446
1447                 for (i = 0, sg = first_scapegoat(revs, commit, sb->reverse);
1448                      i < num_sg && sg;
1449                      sg = sg->next, i++) {
1450                         struct commit *p = sg->item;
1451                         int j, same;
1452
1453                         if (sg_origin[i])
1454                                 continue;
1455                         if (parse_commit(p))
1456                                 continue;
1457                         porigin = find(p, origin);
1458                         if (!porigin)
1459                                 continue;
1460                         if (!oidcmp(&porigin->blob_oid, &origin->blob_oid)) {
1461                                 pass_whole_blame(sb, origin, porigin);
1462                                 blame_origin_decref(porigin);
1463                                 goto finish;
1464                         }
1465                         for (j = same = 0; j < i; j++)
1466                                 if (sg_origin[j] &&
1467                                     !oidcmp(&sg_origin[j]->blob_oid, &porigin->blob_oid)) {
1468                                         same = 1;
1469                                         break;
1470                                 }
1471                         if (!same)
1472                                 sg_origin[i] = porigin;
1473                         else
1474                                 blame_origin_decref(porigin);
1475                 }
1476         }
1477
1478         sb->num_commits++;
1479         for (i = 0, sg = first_scapegoat(revs, commit, sb->reverse);
1480              i < num_sg && sg;
1481              sg = sg->next, i++) {
1482                 struct blame_origin *porigin = sg_origin[i];
1483                 if (!porigin)
1484                         continue;
1485                 if (!origin->previous) {
1486                         blame_origin_incref(porigin);
1487                         origin->previous = porigin;
1488                 }
1489                 pass_blame_to_parent(sb, origin, porigin);
1490                 if (!origin->suspects)
1491                         goto finish;
1492         }
1493
1494         /*
1495          * Optionally find moves in parents' files.
1496          */
1497         if (opt & PICKAXE_BLAME_MOVE) {
1498                 filter_small(sb, &toosmall, &origin->suspects, sb->move_score);
1499                 if (origin->suspects) {
1500                         for (i = 0, sg = first_scapegoat(revs, commit, sb->reverse);
1501                              i < num_sg && sg;
1502                              sg = sg->next, i++) {
1503                                 struct blame_origin *porigin = sg_origin[i];
1504                                 if (!porigin)
1505                                         continue;
1506                                 find_move_in_parent(sb, &blametail, &toosmall, origin, porigin);
1507                                 if (!origin->suspects)
1508                                         break;
1509                         }
1510                 }
1511         }
1512
1513         /*
1514          * Optionally find copies from parents' files.
1515          */
1516         if (opt & PICKAXE_BLAME_COPY) {
1517                 if (sb->copy_score > sb->move_score)
1518                         filter_small(sb, &toosmall, &origin->suspects, sb->copy_score);
1519                 else if (sb->copy_score < sb->move_score) {
1520                         origin->suspects = blame_merge(origin->suspects, toosmall);
1521                         toosmall = NULL;
1522                         filter_small(sb, &toosmall, &origin->suspects, sb->copy_score);
1523                 }
1524                 if (!origin->suspects)
1525                         goto finish;
1526
1527                 for (i = 0, sg = first_scapegoat(revs, commit, sb->reverse);
1528                      i < num_sg && sg;
1529                      sg = sg->next, i++) {
1530                         struct blame_origin *porigin = sg_origin[i];
1531                         find_copy_in_parent(sb, &blametail, &toosmall,
1532                                             origin, sg->item, porigin, opt);
1533                         if (!origin->suspects)
1534                                 goto finish;
1535                 }
1536         }
1537
1538 finish:
1539         *blametail = NULL;
1540         distribute_blame(sb, blames);
1541         /*
1542          * prepend toosmall to origin->suspects
1543          *
1544          * There is no point in sorting: this ends up on a big
1545          * unsorted list in the caller anyway.
1546          */
1547         if (toosmall) {
1548                 struct blame_entry **tail = &toosmall;
1549                 while (*tail)
1550                         tail = &(*tail)->next;
1551                 *tail = origin->suspects;
1552                 origin->suspects = toosmall;
1553         }
1554         for (i = 0; i < num_sg; i++) {
1555                 if (sg_origin[i]) {
1556                         drop_origin_blob(sg_origin[i]);
1557                         blame_origin_decref(sg_origin[i]);
1558                 }
1559         }
1560         drop_origin_blob(origin);
1561         if (sg_buf != sg_origin)
1562                 free(sg_origin);
1563 }
1564
1565 /*
1566  * The main loop -- while we have blobs with lines whose true origin
1567  * is still unknown, pick one blob, and allow its lines to pass blames
1568  * to its parents. */
1569 void assign_blame(struct blame_scoreboard *sb, int opt)
1570 {
1571         struct rev_info *revs = sb->revs;
1572         struct commit *commit = prio_queue_get(&sb->commits);
1573
1574         while (commit) {
1575                 struct blame_entry *ent;
1576                 struct blame_origin *suspect = get_blame_suspects(commit);
1577
1578                 /* find one suspect to break down */
1579                 while (suspect && !suspect->suspects)
1580                         suspect = suspect->next;
1581
1582                 if (!suspect) {
1583                         commit = prio_queue_get(&sb->commits);
1584                         continue;
1585                 }
1586
1587                 assert(commit == suspect->commit);
1588
1589                 /*
1590                  * We will use this suspect later in the loop,
1591                  * so hold onto it in the meantime.
1592                  */
1593                 blame_origin_incref(suspect);
1594                 parse_commit(commit);
1595                 if (sb->reverse ||
1596                     (!(commit->object.flags & UNINTERESTING) &&
1597                      !(revs->max_age != -1 && commit->date < revs->max_age)))
1598                         pass_blame(sb, suspect, opt);
1599                 else {
1600                         commit->object.flags |= UNINTERESTING;
1601                         if (commit->object.parsed)
1602                                 mark_parents_uninteresting(commit);
1603                 }
1604                 /* treat root commit as boundary */
1605                 if (!commit->parents && !sb->show_root)
1606                         commit->object.flags |= UNINTERESTING;
1607
1608                 /* Take responsibility for the remaining entries */
1609                 ent = suspect->suspects;
1610                 if (ent) {
1611                         suspect->guilty = 1;
1612                         for (;;) {
1613                                 struct blame_entry *next = ent->next;
1614                                 if (sb->found_guilty_entry)
1615                                         sb->found_guilty_entry(ent, sb->found_guilty_entry_data);
1616                                 if (next) {
1617                                         ent = next;
1618                                         continue;
1619                                 }
1620                                 ent->next = sb->ent;
1621                                 sb->ent = suspect->suspects;
1622                                 suspect->suspects = NULL;
1623                                 break;
1624                         }
1625                 }
1626                 blame_origin_decref(suspect);
1627
1628                 if (sb->debug) /* sanity */
1629                         sanity_check_refcnt(sb);
1630         }
1631 }
1632
1633 static const char *get_next_line(const char *start, const char *end)
1634 {
1635         const char *nl = memchr(start, '\n', end - start);
1636         return nl ? nl + 1 : end;
1637 }
1638
1639 /*
1640  * To allow quick access to the contents of nth line in the
1641  * final image, prepare an index in the scoreboard.
1642  */
1643 static int prepare_lines(struct blame_scoreboard *sb)
1644 {
1645         const char *buf = sb->final_buf;
1646         unsigned long len = sb->final_buf_size;
1647         const char *end = buf + len;
1648         const char *p;
1649         int *lineno;
1650         int num = 0;
1651
1652         for (p = buf; p < end; p = get_next_line(p, end))
1653                 num++;
1654
1655         ALLOC_ARRAY(sb->lineno, num + 1);
1656         lineno = sb->lineno;
1657
1658         for (p = buf; p < end; p = get_next_line(p, end))
1659                 *lineno++ = p - buf;
1660
1661         *lineno = len;
1662
1663         sb->num_lines = num;
1664         return sb->num_lines;
1665 }
1666
1667 static struct commit *find_single_final(struct rev_info *revs,
1668                                         const char **name_p)
1669 {
1670         int i;
1671         struct commit *found = NULL;
1672         const char *name = NULL;
1673
1674         for (i = 0; i < revs->pending.nr; i++) {
1675                 struct object *obj = revs->pending.objects[i].item;
1676                 if (obj->flags & UNINTERESTING)
1677                         continue;
1678                 obj = deref_tag(the_repository, obj, NULL, 0);
1679                 if (obj->type != OBJ_COMMIT)
1680                         die("Non commit %s?", revs->pending.objects[i].name);
1681                 if (found)
1682                         die("More than one commit to dig from %s and %s?",
1683                             revs->pending.objects[i].name, name);
1684                 found = (struct commit *)obj;
1685                 name = revs->pending.objects[i].name;
1686         }
1687         if (name_p)
1688                 *name_p = xstrdup_or_null(name);
1689         return found;
1690 }
1691
1692 static struct commit *dwim_reverse_initial(struct rev_info *revs,
1693                                            const char **name_p)
1694 {
1695         /*
1696          * DWIM "git blame --reverse ONE -- PATH" as
1697          * "git blame --reverse ONE..HEAD -- PATH" but only do so
1698          * when it makes sense.
1699          */
1700         struct object *obj;
1701         struct commit *head_commit;
1702         struct object_id head_oid;
1703
1704         if (revs->pending.nr != 1)
1705                 return NULL;
1706
1707         /* Is that sole rev a committish? */
1708         obj = revs->pending.objects[0].item;
1709         obj = deref_tag(the_repository, obj, NULL, 0);
1710         if (obj->type != OBJ_COMMIT)
1711                 return NULL;
1712
1713         /* Do we have HEAD? */
1714         if (!resolve_ref_unsafe("HEAD", RESOLVE_REF_READING, &head_oid, NULL))
1715                 return NULL;
1716         head_commit = lookup_commit_reference_gently(the_repository,
1717                                                      &head_oid, 1);
1718         if (!head_commit)
1719                 return NULL;
1720
1721         /* Turn "ONE" into "ONE..HEAD" then */
1722         obj->flags |= UNINTERESTING;
1723         add_pending_object(revs, &head_commit->object, "HEAD");
1724
1725         if (name_p)
1726                 *name_p = revs->pending.objects[0].name;
1727         return (struct commit *)obj;
1728 }
1729
1730 static struct commit *find_single_initial(struct rev_info *revs,
1731                                           const char **name_p)
1732 {
1733         int i;
1734         struct commit *found = NULL;
1735         const char *name = NULL;
1736
1737         /*
1738          * There must be one and only one negative commit, and it must be
1739          * the boundary.
1740          */
1741         for (i = 0; i < revs->pending.nr; i++) {
1742                 struct object *obj = revs->pending.objects[i].item;
1743                 if (!(obj->flags & UNINTERESTING))
1744                         continue;
1745                 obj = deref_tag(the_repository, obj, NULL, 0);
1746                 if (obj->type != OBJ_COMMIT)
1747                         die("Non commit %s?", revs->pending.objects[i].name);
1748                 if (found)
1749                         die("More than one commit to dig up from, %s and %s?",
1750                             revs->pending.objects[i].name, name);
1751                 found = (struct commit *) obj;
1752                 name = revs->pending.objects[i].name;
1753         }
1754
1755         if (!name)
1756                 found = dwim_reverse_initial(revs, &name);
1757         if (!name)
1758                 die("No commit to dig up from?");
1759
1760         if (name_p)
1761                 *name_p = xstrdup(name);
1762         return found;
1763 }
1764
1765 void init_scoreboard(struct blame_scoreboard *sb)
1766 {
1767         memset(sb, 0, sizeof(struct blame_scoreboard));
1768         sb->move_score = BLAME_DEFAULT_MOVE_SCORE;
1769         sb->copy_score = BLAME_DEFAULT_COPY_SCORE;
1770 }
1771
1772 void setup_scoreboard(struct blame_scoreboard *sb,
1773                       const char *path,
1774                       struct blame_origin **orig)
1775 {
1776         const char *final_commit_name = NULL;
1777         struct blame_origin *o;
1778         struct commit *final_commit = NULL;
1779         enum object_type type;
1780
1781         init_blame_suspects(&blame_suspects);
1782
1783         if (sb->reverse && sb->contents_from)
1784                 die(_("--contents and --reverse do not blend well."));
1785
1786         if (!sb->repo)
1787                 BUG("repo is NULL");
1788
1789         if (!sb->reverse) {
1790                 sb->final = find_single_final(sb->revs, &final_commit_name);
1791                 sb->commits.compare = compare_commits_by_commit_date;
1792         } else {
1793                 sb->final = find_single_initial(sb->revs, &final_commit_name);
1794                 sb->commits.compare = compare_commits_by_reverse_commit_date;
1795         }
1796
1797         if (sb->final && sb->contents_from)
1798                 die(_("cannot use --contents with final commit object name"));
1799
1800         if (sb->reverse && sb->revs->first_parent_only)
1801                 sb->revs->children.name = NULL;
1802
1803         if (!sb->final) {
1804                 /*
1805                  * "--not A B -- path" without anything positive;
1806                  * do not default to HEAD, but use the working tree
1807                  * or "--contents".
1808                  */
1809                 setup_work_tree();
1810                 sb->final = fake_working_tree_commit(sb->repo,
1811                                                      &sb->revs->diffopt,
1812                                                      path, sb->contents_from);
1813                 add_pending_object(sb->revs, &(sb->final->object), ":");
1814         }
1815
1816         if (sb->reverse && sb->revs->first_parent_only) {
1817                 final_commit = find_single_final(sb->revs, NULL);
1818                 if (!final_commit)
1819                         die(_("--reverse and --first-parent together require specified latest commit"));
1820         }
1821
1822         /*
1823          * If we have bottom, this will mark the ancestors of the
1824          * bottom commits we would reach while traversing as
1825          * uninteresting.
1826          */
1827         if (prepare_revision_walk(sb->revs))
1828                 die(_("revision walk setup failed"));
1829
1830         if (sb->reverse && sb->revs->first_parent_only) {
1831                 struct commit *c = final_commit;
1832
1833                 sb->revs->children.name = "children";
1834                 while (c->parents &&
1835                        oidcmp(&c->object.oid, &sb->final->object.oid)) {
1836                         struct commit_list *l = xcalloc(1, sizeof(*l));
1837
1838                         l->item = c;
1839                         if (add_decoration(&sb->revs->children,
1840                                            &c->parents->item->object, l))
1841                                 BUG("not unique item in first-parent chain");
1842                         c = c->parents->item;
1843                 }
1844
1845                 if (oidcmp(&c->object.oid, &sb->final->object.oid))
1846                         die(_("--reverse --first-parent together require range along first-parent chain"));
1847         }
1848
1849         if (is_null_oid(&sb->final->object.oid)) {
1850                 o = get_blame_suspects(sb->final);
1851                 sb->final_buf = xmemdupz(o->file.ptr, o->file.size);
1852                 sb->final_buf_size = o->file.size;
1853         }
1854         else {
1855                 o = get_origin(sb->final, path);
1856                 if (fill_blob_sha1_and_mode(sb->repo, o))
1857                         die(_("no such path %s in %s"), path, final_commit_name);
1858
1859                 if (sb->revs->diffopt.flags.allow_textconv &&
1860                     textconv_object(path, o->mode, &o->blob_oid, 1, (char **) &sb->final_buf,
1861                                     &sb->final_buf_size))
1862                         ;
1863                 else
1864                         sb->final_buf = read_object_file(&o->blob_oid, &type,
1865                                                          &sb->final_buf_size);
1866
1867                 if (!sb->final_buf)
1868                         die(_("cannot read blob %s for path %s"),
1869                             oid_to_hex(&o->blob_oid),
1870                             path);
1871         }
1872         sb->num_read_blob++;
1873         prepare_lines(sb);
1874
1875         if (orig)
1876                 *orig = o;
1877
1878         free((char *)final_commit_name);
1879 }
1880
1881
1882
1883 struct blame_entry *blame_entry_prepend(struct blame_entry *head,
1884                                         long start, long end,
1885                                         struct blame_origin *o)
1886 {
1887         struct blame_entry *new_head = xcalloc(1, sizeof(struct blame_entry));
1888         new_head->lno = start;
1889         new_head->num_lines = end - start;
1890         new_head->suspect = o;
1891         new_head->s_lno = start;
1892         new_head->next = head;
1893         blame_origin_incref(o);
1894         return new_head;
1895 }