Merge branch 'sb/submodule-merge-in-merge-recursive'
authorJunio C Hamano <gitster@pobox.com>
Wed, 13 Jun 2018 19:50:44 +0000 (12:50 -0700)
committerJunio C Hamano <gitster@pobox.com>
Wed, 13 Jun 2018 19:50:45 +0000 (12:50 -0700)
Finishing touches to a topic that already is in 'master'.

* sb/submodule-merge-in-merge-recursive:
merge-submodule: reduce output verbosity

1  2 
merge-recursive.c
diff --combined merge-recursive.c
index ac27abbd4cd708fdc2526c6508680b7d246773b6,11c858d422f61d936c061a6d9bca03eaf422a409..5eb907f46e04bc560833ed42074a60d217ef19bd
@@@ -50,67 -50,6 +50,67 @@@ static unsigned int path_hash(const cha
        return ignore_case ? strihash(path) : strhash(path);
  }
  
 +static struct dir_rename_entry *dir_rename_find_entry(struct hashmap *hashmap,
 +                                                    char *dir)
 +{
 +      struct dir_rename_entry key;
 +
 +      if (dir == NULL)
 +              return NULL;
 +      hashmap_entry_init(&key, strhash(dir));
 +      key.dir = dir;
 +      return hashmap_get(hashmap, &key, NULL);
 +}
 +
 +static int dir_rename_cmp(const void *unused_cmp_data,
 +                        const void *entry,
 +                        const void *entry_or_key,
 +                        const void *unused_keydata)
 +{
 +      const struct dir_rename_entry *e1 = entry;
 +      const struct dir_rename_entry *e2 = entry_or_key;
 +
 +      return strcmp(e1->dir, e2->dir);
 +}
 +
 +static void dir_rename_init(struct hashmap *map)
 +{
 +      hashmap_init(map, dir_rename_cmp, NULL, 0);
 +}
 +
 +static void dir_rename_entry_init(struct dir_rename_entry *entry,
 +                                char *directory)
 +{
 +      hashmap_entry_init(entry, strhash(directory));
 +      entry->dir = directory;
 +      entry->non_unique_new_dir = 0;
 +      strbuf_init(&entry->new_dir, 0);
 +      string_list_init(&entry->possible_new_dirs, 0);
 +}
 +
 +static struct collision_entry *collision_find_entry(struct hashmap *hashmap,
 +                                                  char *target_file)
 +{
 +      struct collision_entry key;
 +
 +      hashmap_entry_init(&key, strhash(target_file));
 +      key.target_file = target_file;
 +      return hashmap_get(hashmap, &key, NULL);
 +}
 +
 +static int collision_cmp(void *unused_cmp_data,
 +                       const struct collision_entry *e1,
 +                       const struct collision_entry *e2,
 +                       const void *unused_keydata)
 +{
 +      return strcmp(e1->target_file, e2->target_file);
 +}
 +
 +static void collision_init(struct hashmap *map)
 +{
 +      hashmap_init(map, (hashmap_cmp_fn) collision_cmp, NULL, 0);
 +}
 +
  static void flush_output(struct merge_options *o)
  {
        if (o->buffer_output < 2 && o->obuf.len) {
@@@ -163,7 -102,7 +163,7 @@@ static struct commit *make_virtual_comm
        struct commit *commit = alloc_commit_node();
  
        set_merge_remote_desc(commit, comment, (struct object *)commit);
 -      commit->tree = tree;
 +      commit->maybe_tree = tree;
        commit->object.parsed = 1;
        return commit;
  }
@@@ -181,7 -120,6 +181,7 @@@ static int oid_eq(const struct object_i
  
  enum rename_type {
        RENAME_NORMAL = 0,
 +      RENAME_DIR,
        RENAME_DELETE,
        RENAME_ONE_FILE_TO_ONE,
        RENAME_ONE_FILE_TO_TWO,
@@@ -317,7 -255,7 +317,7 @@@ static int add_cacheinfo(struct merge_o
  
        ce = make_cache_entry(mode, oid ? oid->hash : null_sha1, path, stage, 0);
        if (!ce)
 -              return err(o, _("addinfo_cache failed for path '%s'"), path);
 +              return err(o, _("add_cacheinfo failed for path '%s'; merge aborting."), path);
  
        ret = add_cache_entry(ce, options);
        if (refresh) {
  
                nce = refresh_cache_entry(ce, CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
                if (!nce)
 -                      return err(o, _("addinfo_cache failed for path '%s'"), path);
 +                      return err(o, _("add_cacheinfo failed to refresh for path '%s'; merge aborting."), path);
                if (nce != ce)
                        ret = add_cache_entry(nce, options);
        }
@@@ -338,54 -276,36 +338,54 @@@ static void init_tree_desc_from_tree(st
        init_tree_desc(desc, tree->buffer, tree->size);
  }
  
 -static int git_merge_trees(int index_only,
 -                         struct tree *common,
 -                         struct tree *head,
 -                         struct tree *merge)
 +static int unpack_trees_start(struct merge_options *o,
 +                            struct tree *common,
 +                            struct tree *head,
 +                            struct tree *merge)
  {
        int rc;
        struct tree_desc t[3];
 -      struct unpack_trees_options opts;
 +      struct index_state tmp_index = { NULL };
  
 -      memset(&opts, 0, sizeof(opts));
 -      if (index_only)
 -              opts.index_only = 1;
 +      memset(&o->unpack_opts, 0, sizeof(o->unpack_opts));
 +      if (o->call_depth)
 +              o->unpack_opts.index_only = 1;
        else
 -              opts.update = 1;
 -      opts.merge = 1;
 -      opts.head_idx = 2;
 -      opts.fn = threeway_merge;
 -      opts.src_index = &the_index;
 -      opts.dst_index = &the_index;
 -      setup_unpack_trees_porcelain(&opts, "merge");
 +              o->unpack_opts.update = 1;
 +      o->unpack_opts.merge = 1;
 +      o->unpack_opts.head_idx = 2;
 +      o->unpack_opts.fn = threeway_merge;
 +      o->unpack_opts.src_index = &the_index;
 +      o->unpack_opts.dst_index = &tmp_index;
 +      o->unpack_opts.aggressive = !merge_detect_rename(o);
 +      setup_unpack_trees_porcelain(&o->unpack_opts, "merge");
  
        init_tree_desc_from_tree(t+0, common);
        init_tree_desc_from_tree(t+1, head);
        init_tree_desc_from_tree(t+2, merge);
  
 -      rc = unpack_trees(3, t, &opts);
 +      rc = unpack_trees(3, t, &o->unpack_opts);
        cache_tree_free(&active_cache_tree);
 +
 +      /*
 +       * Update the_index to match the new results, AFTER saving a copy
 +       * in o->orig_index.  Update src_index to point to the saved copy.
 +       * (verify_uptodate() checks src_index, and the original index is
 +       * the one that had the necessary modification timestamps.)
 +       */
 +      o->orig_index = the_index;
 +      the_index = tmp_index;
 +      o->unpack_opts.src_index = &o->orig_index;
 +
        return rc;
  }
  
 +static void unpack_trees_finish(struct merge_options *o)
 +{
 +      discard_index(&o->orig_index);
 +      clear_unpack_trees_porcelain(&o->unpack_opts);
 +}
 +
  struct tree *write_tree_from_memory(struct merge_options *o)
  {
        struct tree *result = NULL;
                                fprintf(stderr, "BUG: %d %.*s\n", ce_stage(ce),
                                        (int)ce_namelen(ce), ce->name);
                }
 -              die("BUG: unmerged index entries in merge-recursive.c");
 +              BUG("unmerged index entries in merge-recursive.c");
        }
  
        if (!active_cache_tree)
@@@ -441,21 -361,6 +441,21 @@@ static void get_files_dirs(struct merge
        read_tree_recursive(tree, "", 0, 0, &match_all, save_files_dirs, o);
  }
  
 +static int get_tree_entry_if_blob(const struct object_id *tree,
 +                                const char *path,
 +                                struct object_id *hashy,
 +                                unsigned int *mode_o)
 +{
 +      int ret;
 +
 +      ret = get_tree_entry(tree, path, hashy, mode_o);
 +      if (S_ISDIR(*mode_o)) {
 +              oidcpy(hashy, &null_oid);
 +              *mode_o = 0;
 +      }
 +      return ret;
 +}
 +
  /*
   * Returns an index_entry instance which doesn't have to correspond to
   * a real cache entry in Git's index.
@@@ -466,12 -371,12 +466,12 @@@ static struct stage_data *insert_stage_
  {
        struct string_list_item *item;
        struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
 -      get_tree_entry(&o->object.oid, path,
 -                      &e->stages[1].oid, &e->stages[1].mode);
 -      get_tree_entry(&a->object.oid, path,
 -                      &e->stages[2].oid, &e->stages[2].mode);
 -      get_tree_entry(&b->object.oid, path,
 -                      &e->stages[3].oid, &e->stages[3].mode);
 +      get_tree_entry_if_blob(&o->object.oid, path,
 +                             &e->stages[1].oid, &e->stages[1].mode);
 +      get_tree_entry_if_blob(&a->object.oid, path,
 +                             &e->stages[2].oid, &e->stages[2].mode);
 +      get_tree_entry_if_blob(&b->object.oid, path,
 +                             &e->stages[3].oid, &e->stages[3].mode);
        item = string_list_insert(entries, path);
        item->util = e;
        return e;
@@@ -630,10 -535,78 +630,10 @@@ struct rename 
         */
        struct stage_data *src_entry;
        struct stage_data *dst_entry;
 +      unsigned add_turned_into_rename:1;
        unsigned processed:1;
  };
  
 -/*
 - * Get information of all renames which occurred between 'o_tree' and
 - * 'tree'. We need the three trees in the merge ('o_tree', 'a_tree' and
 - * 'b_tree') to be able to associate the correct cache entries with
 - * the rename information. 'tree' is always equal to either a_tree or b_tree.
 - */
 -static struct string_list *get_renames(struct merge_options *o,
 -                                     struct tree *tree,
 -                                     struct tree *o_tree,
 -                                     struct tree *a_tree,
 -                                     struct tree *b_tree,
 -                                     struct string_list *entries)
 -{
 -      int i;
 -      struct string_list *renames;
 -      struct diff_options opts;
 -
 -      renames = xcalloc(1, sizeof(struct string_list));
 -      if (!o->detect_rename)
 -              return renames;
 -
 -      diff_setup(&opts);
 -      opts.flags.recursive = 1;
 -      opts.flags.rename_empty = 0;
 -      opts.detect_rename = DIFF_DETECT_RENAME;
 -      opts.rename_limit = o->merge_rename_limit >= 0 ? o->merge_rename_limit :
 -                          o->diff_rename_limit >= 0 ? o->diff_rename_limit :
 -                          1000;
 -      opts.rename_score = o->rename_score;
 -      opts.show_rename_progress = o->show_rename_progress;
 -      opts.output_format = DIFF_FORMAT_NO_OUTPUT;
 -      diff_setup_done(&opts);
 -      diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
 -      diffcore_std(&opts);
 -      if (opts.needed_rename_limit > o->needed_rename_limit)
 -              o->needed_rename_limit = opts.needed_rename_limit;
 -      for (i = 0; i < diff_queued_diff.nr; ++i) {
 -              struct string_list_item *item;
 -              struct rename *re;
 -              struct diff_filepair *pair = diff_queued_diff.queue[i];
 -              if (pair->status != 'R') {
 -                      diff_free_filepair(pair);
 -                      continue;
 -              }
 -              re = xmalloc(sizeof(*re));
 -              re->processed = 0;
 -              re->pair = pair;
 -              item = string_list_lookup(entries, re->pair->one->path);
 -              if (!item)
 -                      re->src_entry = insert_stage_data(re->pair->one->path,
 -                                      o_tree, a_tree, b_tree, entries);
 -              else
 -                      re->src_entry = item->util;
 -
 -              item = string_list_lookup(entries, re->pair->two->path);
 -              if (!item)
 -                      re->dst_entry = insert_stage_data(re->pair->two->path,
 -                                      o_tree, a_tree, b_tree, entries);
 -              else
 -                      re->dst_entry = item->util;
 -              item = string_list_insert(renames, pair->one->path);
 -              item->util = re;
 -      }
 -      opts.output_format = DIFF_FORMAT_NO_OUTPUT;
 -      diff_queued_diff.nr = 0;
 -      diff_flush(&opts);
 -      return renames;
 -}
 -
  static int update_stages(struct merge_options *opt, const char *path,
                         const struct diff_filespec *o,
                         const struct diff_filespec *a,
        return 0;
  }
  
 +static int update_stages_for_stage_data(struct merge_options *opt,
 +                                      const char *path,
 +                                      const struct stage_data *stage_data)
 +{
 +      struct diff_filespec o, a, b;
 +
 +      o.mode = stage_data->stages[1].mode;
 +      oidcpy(&o.oid, &stage_data->stages[1].oid);
 +
 +      a.mode = stage_data->stages[2].mode;
 +      oidcpy(&a.oid, &stage_data->stages[2].oid);
 +
 +      b.mode = stage_data->stages[3].mode;
 +      oidcpy(&b.oid, &stage_data->stages[3].oid);
 +
 +      return update_stages(opt, path,
 +                           is_null_oid(&o.oid) ? NULL : &o,
 +                           is_null_oid(&a.oid) ? NULL : &a,
 +                           is_null_oid(&b.oid) ? NULL : &b);
 +}
 +
  static void update_entry(struct stage_data *entry,
                         struct diff_filespec *o,
                         struct diff_filespec *a,
@@@ -787,92 -739,31 +787,92 @@@ static int dir_in_way(const char *path
                !(empty_ok && is_empty_dir(path));
  }
  
 -static int was_tracked(const char *path)
 +/*
 + * Returns whether path was tracked in the index before the merge started,
 + * and its oid and mode match the specified values
 + */
 +static int was_tracked_and_matches(struct merge_options *o, const char *path,
 +                                 const struct object_id *oid, unsigned mode)
  {
 -      int pos = cache_name_pos(path, strlen(path));
 +      int pos = index_name_pos(&o->orig_index, path, strlen(path));
 +      struct cache_entry *ce;
 +
 +      if (0 > pos)
 +              /* we were not tracking this path before the merge */
 +              return 0;
 +
 +      /* See if the file we were tracking before matches */
 +      ce = o->orig_index.cache[pos];
 +      return (oid_eq(&ce->oid, oid) && ce->ce_mode == mode);
 +}
 +
 +/*
 + * Returns whether path was tracked in the index before the merge started
 + */
 +static int was_tracked(struct merge_options *o, const char *path)
 +{
 +      int pos = index_name_pos(&o->orig_index, path, strlen(path));
  
        if (0 <= pos)
 -              /* we have been tracking this path */
 +              /* we were tracking this path before the merge */
                return 1;
  
 -      /*
 -       * Look for an unmerged entry for the path,
 -       * specifically stage #2, which would indicate
 -       * that "our" side before the merge started
 -       * had the path tracked (and resulted in a conflict).
 -       */
 -      for (pos = -1 - pos;
 -           pos < active_nr && !strcmp(path, active_cache[pos]->name);
 -           pos++)
 -              if (ce_stage(active_cache[pos]) == 2)
 -                      return 1;
        return 0;
  }
  
  static int would_lose_untracked(const char *path)
  {
 -      return !was_tracked(path) && file_exists(path);
 +      /*
 +       * This may look like it can be simplified to:
 +       *   return !was_tracked(o, path) && file_exists(path)
 +       * but it can't.  This function needs to know whether path was in
 +       * the working tree due to EITHER having been tracked in the index
 +       * before the merge OR having been put into the working copy and
 +       * index by unpack_trees().  Due to that either-or requirement, we
 +       * check the current index instead of the original one.
 +       *
 +       * Note that we do not need to worry about merge-recursive itself
 +       * updating the index after unpack_trees() and before calling this
 +       * function, because we strictly require all code paths in
 +       * merge-recursive to update the working tree first and the index
 +       * second.  Doing otherwise would break
 +       * update_file()/would_lose_untracked(); see every comment in this
 +       * file which mentions "update_stages".
 +       */
 +      int pos = cache_name_pos(path, strlen(path));
 +
 +      if (pos < 0)
 +              pos = -1 - pos;
 +      while (pos < active_nr &&
 +             !strcmp(path, active_cache[pos]->name)) {
 +              /*
 +               * If stage #0, it is definitely tracked.
 +               * If it has stage #2 then it was tracked
 +               * before this merge started.  All other
 +               * cases the path was not tracked.
 +               */
 +              switch (ce_stage(active_cache[pos])) {
 +              case 0:
 +              case 2:
 +                      return 0;
 +              }
 +              pos++;
 +      }
 +      return file_exists(path);
 +}
 +
 +static int was_dirty(struct merge_options *o, const char *path)
 +{
 +      struct cache_entry *ce;
 +      int dirty = 1;
 +
 +      if (o->call_depth || !was_tracked(o, path))
 +              return !dirty;
 +
 +      ce = index_file_exists(o->unpack_opts.src_index,
 +                             path, strlen(path), ignore_case);
 +      dirty = verify_uptodate(ce, &o->unpack_opts) != 0;
 +      return dirty;
  }
  
  static int make_room_for_path(struct merge_options *o, const char *path)
@@@ -1003,9 -894,7 +1003,9 @@@ static int update_file_flags(struct mer
        }
   update_index:
        if (!ret && update_cache)
 -              add_cacheinfo(o, mode, oid, path, 0, update_wd, ADD_CACHE_OK_TO_ADD);
 +              if (add_cacheinfo(o, mode, oid, path, 0, update_wd,
 +                                ADD_CACHE_OK_TO_ADD))
 +                      return -1;
        return ret;
  }
  
@@@ -1208,7 -1097,7 +1208,7 @@@ static int merge_submodule(struct merge
                        output(o, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
                        output_commit_title(o, commit_b);
                } else if (show(o, 2))
-                       output(o, 2, _("Fast-forwarding submodule %s to %s"), path, oid_to_hex(b));
+                       output(o, 2, _("Fast-forwarding submodule %s"), path);
                else
                        ; /* no output */
  
                        output(o, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
                        output_commit_title(o, commit_a);
                } else if (show(o, 2))
-                       output(o, 2, _("Fast-forwarding submodule %s to %s"), path, oid_to_hex(a));
+                       output(o, 2, _("Fast-forwarding submodule %s"), path);
                else
                        ; /* no output */
  
  }
  
  static int merge_file_1(struct merge_options *o,
 -                                         const struct diff_filespec *one,
 -                                         const struct diff_filespec *a,
 -                                         const struct diff_filespec *b,
 -                                         const char *branch1,
 -                                         const char *branch2,
 -                                         struct merge_file_info *result)
 +                      const struct diff_filespec *one,
 +                      const struct diff_filespec *a,
 +                      const struct diff_filespec *b,
 +                      const char *filename,
 +                      const char *branch1,
 +                      const char *branch2,
 +                      struct merge_file_info *result)
  {
        result->merge = 0;
        result->clean = 1;
                                break;
                        }
                } else
 -                      die("BUG: unsupported object type in the tree");
 +                      BUG("unsupported object type in the tree");
        }
  
 +      if (result->merge)
 +              output(o, 2, _("Auto-merging %s"), filename);
 +
        return 0;
  }
  
  static int merge_file_special_markers(struct merge_options *o,
 -                         const struct diff_filespec *one,
 -                         const struct diff_filespec *a,
 -                         const struct diff_filespec *b,
 -                         const char *branch1,
 -                         const char *filename1,
 -                         const char *branch2,
 -                         const char *filename2,
 -                         struct merge_file_info *mfi)
 +                                    const struct diff_filespec *one,
 +                                    const struct diff_filespec *a,
 +                                    const struct diff_filespec *b,
 +                                    const char *target_filename,
 +                                    const char *branch1,
 +                                    const char *filename1,
 +                                    const char *branch2,
 +                                    const char *filename2,
 +                                    struct merge_file_info *mfi)
  {
        char *side1 = NULL;
        char *side2 = NULL;
        if (filename2)
                side2 = xstrfmt("%s:%s", branch2, filename2);
  
 -      ret = merge_file_1(o, one, a, b,
 +      ret = merge_file_1(o, one, a, b, target_filename,
                           side1 ? side1 : branch1,
                           side2 ? side2 : branch2, mfi);
 +
        free(side1);
        free(side2);
        return ret;
  }
  
  static int merge_file_one(struct merge_options *o,
 -                                       const char *path,
 -                                       const struct object_id *o_oid, int o_mode,
 -                                       const struct object_id *a_oid, int a_mode,
 -                                       const struct object_id *b_oid, int b_mode,
 -                                       const char *branch1,
 -                                       const char *branch2,
 -                                       struct merge_file_info *mfi)
 +                        const char *path,
 +                        const struct object_id *o_oid, int o_mode,
 +                        const struct object_id *a_oid, int a_mode,
 +                        const struct object_id *b_oid, int b_mode,
 +                        const char *branch1,
 +                        const char *branch2,
 +                        struct merge_file_info *mfi)
  {
        struct diff_filespec one, a, b;
  
        a.mode = a_mode;
        oidcpy(&b.oid, b_oid);
        b.mode = b_mode;
 -      return merge_file_1(o, &one, &a, &b, branch1, branch2, mfi);
 +      return merge_file_1(o, &one, &a, &b, path, branch1, branch2, mfi);
 +}
 +
 +static int conflict_rename_dir(struct merge_options *o,
 +                             struct diff_filepair *pair,
 +                             const char *rename_branch,
 +                             const char *other_branch)
 +{
 +      const struct diff_filespec *dest = pair->two;
 +
 +      if (!o->call_depth && would_lose_untracked(dest->path)) {
 +              char *alt_path = unique_path(o, dest->path, rename_branch);
 +
 +              output(o, 1, _("Error: Refusing to lose untracked file at %s; "
 +                             "writing to %s instead."),
 +                     dest->path, alt_path);
 +              /*
 +               * Write the file in worktree at alt_path, but not in the
 +               * index.  Instead, write to dest->path for the index but
 +               * only at the higher appropriate stage.
 +               */
 +              if (update_file(o, 0, &dest->oid, dest->mode, alt_path))
 +                      return -1;
 +              free(alt_path);
 +              return update_stages(o, dest->path, NULL,
 +                                   rename_branch == o->branch1 ? dest : NULL,
 +                                   rename_branch == o->branch1 ? NULL : dest);
 +      }
 +
 +      /* Update dest->path both in index and in worktree */
 +      if (update_file(o, 1, &dest->oid, dest->mode, dest->path))
 +              return -1;
 +      return 0;
  }
  
  static int handle_change_delete(struct merge_options *o,
        const char *update_path = path;
        int ret = 0;
  
 -      if (dir_in_way(path, !o->call_depth, 0)) {
 +      if (dir_in_way(path, !o->call_depth, 0) ||
 +          (!o->call_depth && would_lose_untracked(path))) {
                update_path = alt_path = unique_path(o, path, change_branch);
        }
  
@@@ -1571,34 -1421,17 +1571,34 @@@ static int handle_file(struct merge_opt
  
        add = filespec_from_entry(&other, dst_entry, stage ^ 1);
        if (add) {
 +              int ren_src_was_dirty = was_dirty(o, rename->path);
                char *add_name = unique_path(o, rename->path, other_branch);
                if (update_file(o, 0, &add->oid, add->mode, add_name))
                        return -1;
  
 -              remove_file(o, 0, rename->path, 0);
 +              if (ren_src_was_dirty) {
 +                      output(o, 1, _("Refusing to lose dirty file at %s"),
 +                             rename->path);
 +              }
 +              /*
 +               * Because the double negatives somehow keep confusing me...
 +               *    1) update_wd iff !ren_src_was_dirty.
 +               *    2) no_wd iff !update_wd
 +               *    3) so, no_wd == !!ren_src_was_dirty == ren_src_was_dirty
 +               */
 +              remove_file(o, 0, rename->path, ren_src_was_dirty);
                dst_name = unique_path(o, rename->path, cur_branch);
        } else {
                if (dir_in_way(rename->path, !o->call_depth, 0)) {
                        dst_name = unique_path(o, rename->path, cur_branch);
                        output(o, 1, _("%s is a directory in %s adding as %s instead"),
                               rename->path, other_branch, dst_name);
 +              } else if (!o->call_depth &&
 +                         would_lose_untracked(rename->path)) {
 +                      dst_name = unique_path(o, rename->path, cur_branch);
 +                      output(o, 1, _("Refusing to lose untracked file at %s; "
 +                                     "adding as %s instead"),
 +                             rename->path, dst_name);
                }
        }
        if ((ret = update_file(o, 0, &rename->oid, rename->mode, dst_name)))
@@@ -1685,8 -1518,6 +1685,8 @@@ static int conflict_rename_rename_2to1(
        struct diff_filespec *c1 = ci->pair1->two;
        struct diff_filespec *c2 = ci->pair2->two;
        char *path = c1->path; /* == c2->path */
 +      char *path_side_1_desc;
 +      char *path_side_2_desc;
        struct merge_file_info mfi_c1;
        struct merge_file_info mfi_c2;
        int ret;
        remove_file(o, 1, a->path, o->call_depth || would_lose_untracked(a->path));
        remove_file(o, 1, b->path, o->call_depth || would_lose_untracked(b->path));
  
 +      path_side_1_desc = xstrfmt("%s (was %s)", path, a->path);
 +      path_side_2_desc = xstrfmt("%s (was %s)", path, b->path);
        if (merge_file_special_markers(o, a, c1, &ci->ren1_other,
 +                                     path_side_1_desc,
                                       o->branch1, c1->path,
                                       o->branch2, ci->ren1_other.path, &mfi_c1) ||
            merge_file_special_markers(o, b, &ci->ren2_other, c2,
 +                                     path_side_2_desc,
                                       o->branch1, ci->ren2_other.path,
                                       o->branch2, c2->path, &mfi_c2))
                return -1;
 +      free(path_side_1_desc);
 +      free(path_side_2_desc);
  
        if (o->call_depth) {
                /*
                char *new_path2 = unique_path(o, path, ci->branch2);
                output(o, 1, _("Renaming %s to %s and %s to %s instead"),
                       a->path, new_path1, b->path, new_path2);
 -              remove_file(o, 0, path, 0);
 +              if (was_dirty(o, path))
 +                      output(o, 1, _("Refusing to lose dirty file at %s"),
 +                             path);
 +              else if (would_lose_untracked(path))
 +                      /*
 +                       * Only way we get here is if both renames were from
 +                       * a directory rename AND user had an untracked file
 +                       * at the location where both files end up after the
 +                       * two directory renames.  See testcase 10d of t6043.
 +                       */
 +                      output(o, 1, _("Refusing to lose untracked file at "
 +                                     "%s, even though it's in the way."),
 +                             path);
 +              else
 +                      remove_file(o, 0, path, 0);
                ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, new_path1);
                if (!ret)
                        ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
                                          new_path2);
 +              /*
 +               * unpack_trees() actually populates the index for us for
 +               * "normal" rename/rename(2to1) situtations so that the
 +               * correct entries are at the higher stages, which would
 +               * make the call below to update_stages_for_stage_data
 +               * unnecessary.  However, if either of the renames came
 +               * from a directory rename, then unpack_trees() will not
 +               * have gotten the right data loaded into the index, so we
 +               * need to do so now.  (While it'd be tempting to move this
 +               * call to update_stages_for_stage_data() to
 +               * apply_directory_rename_modifications(), that would break
 +               * our intermediate calls to would_lose_untracked() since
 +               * those rely on the current in-memory index.  See also the
 +               * big "NOTE" in update_stages()).
 +               */
 +              if (update_stages_for_stage_data(o, path, ci->dst_entry1))
 +                      ret = -1;
 +
                free(new_path2);
                free(new_path1);
        }
        return ret;
  }
  
 +/*
 + * Get the diff_filepairs changed between o_tree and tree.
 + */
 +static struct diff_queue_struct *get_diffpairs(struct merge_options *o,
 +                                             struct tree *o_tree,
 +                                             struct tree *tree)
 +{
 +      struct diff_queue_struct *ret;
 +      struct diff_options opts;
 +
 +      diff_setup(&opts);
 +      opts.flags.recursive = 1;
 +      opts.flags.rename_empty = 0;
 +      opts.detect_rename = merge_detect_rename(o);
 +      /*
 +       * We do not have logic to handle the detection of copies.  In
 +       * fact, it may not even make sense to add such logic: would we
 +       * really want a change to a base file to be propagated through
 +       * multiple other files by a merge?
 +       */
 +      if (opts.detect_rename > DIFF_DETECT_RENAME)
 +              opts.detect_rename = DIFF_DETECT_RENAME;
 +      opts.rename_limit = o->merge_rename_limit >= 0 ? o->merge_rename_limit :
 +                          o->diff_rename_limit >= 0 ? o->diff_rename_limit :
 +                          1000;
 +      opts.rename_score = o->rename_score;
 +      opts.show_rename_progress = o->show_rename_progress;
 +      opts.output_format = DIFF_FORMAT_NO_OUTPUT;
 +      diff_setup_done(&opts);
 +      diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
 +      diffcore_std(&opts);
 +      if (opts.needed_rename_limit > o->needed_rename_limit)
 +              o->needed_rename_limit = opts.needed_rename_limit;
 +
 +      ret = xmalloc(sizeof(*ret));
 +      *ret = diff_queued_diff;
 +
 +      opts.output_format = DIFF_FORMAT_NO_OUTPUT;
 +      diff_queued_diff.nr = 0;
 +      diff_queued_diff.queue = NULL;
 +      diff_flush(&opts);
 +      return ret;
 +}
 +
 +static int tree_has_path(struct tree *tree, const char *path)
 +{
 +      struct object_id hashy;
 +      unsigned int mode_o;
 +
 +      return !get_tree_entry(&tree->object.oid, path,
 +                             &hashy, &mode_o);
 +}
 +
 +/*
 + * Return a new string that replaces the beginning portion (which matches
 + * entry->dir), with entry->new_dir.  In perl-speak:
 + *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
 + * NOTE:
 + *   Caller must ensure that old_path starts with entry->dir + '/'.
 + */
 +static char *apply_dir_rename(struct dir_rename_entry *entry,
 +                            const char *old_path)
 +{
 +      struct strbuf new_path = STRBUF_INIT;
 +      int oldlen, newlen;
 +
 +      if (entry->non_unique_new_dir)
 +              return NULL;
 +
 +      oldlen = strlen(entry->dir);
 +      newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
 +      strbuf_grow(&new_path, newlen);
 +      strbuf_addbuf(&new_path, &entry->new_dir);
 +      strbuf_addstr(&new_path, &old_path[oldlen]);
 +
 +      return strbuf_detach(&new_path, NULL);
 +}
 +
 +static void get_renamed_dir_portion(const char *old_path, const char *new_path,
 +                                  char **old_dir, char **new_dir)
 +{
 +      char *end_of_old, *end_of_new;
 +      int old_len, new_len;
 +
 +      *old_dir = NULL;
 +      *new_dir = NULL;
 +
 +      /*
 +       * For
 +       *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
 +       * the "e/foo.c" part is the same, we just want to know that
 +       *    "a/b/c/d" was renamed to "a/b/some/thing/else"
 +       * so, for this example, this function returns "a/b/c/d" in
 +       * *old_dir and "a/b/some/thing/else" in *new_dir.
 +       *
 +       * Also, if the basename of the file changed, we don't care.  We
 +       * want to know which portion of the directory, if any, changed.
 +       */
 +      end_of_old = strrchr(old_path, '/');
 +      end_of_new = strrchr(new_path, '/');
 +
 +      if (end_of_old == NULL || end_of_new == NULL)
 +              return;
 +      while (*--end_of_new == *--end_of_old &&
 +             end_of_old != old_path &&
 +             end_of_new != new_path)
 +              ; /* Do nothing; all in the while loop */
 +      /*
 +       * We've found the first non-matching character in the directory
 +       * paths.  That means the current directory we were comparing
 +       * represents the rename.  Move end_of_old and end_of_new back
 +       * to the full directory name.
 +       */
 +      if (*end_of_old == '/')
 +              end_of_old++;
 +      if (*end_of_old != '/')
 +              end_of_new++;
 +      end_of_old = strchr(end_of_old, '/');
 +      end_of_new = strchr(end_of_new, '/');
 +
 +      /*
 +       * It may have been the case that old_path and new_path were the same
 +       * directory all along.  Don't claim a rename if they're the same.
 +       */
 +      old_len = end_of_old - old_path;
 +      new_len = end_of_new - new_path;
 +
 +      if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
 +              *old_dir = xstrndup(old_path, old_len);
 +              *new_dir = xstrndup(new_path, new_len);
 +      }
 +}
 +
 +static void remove_hashmap_entries(struct hashmap *dir_renames,
 +                                 struct string_list *items_to_remove)
 +{
 +      int i;
 +      struct dir_rename_entry *entry;
 +
 +      for (i = 0; i < items_to_remove->nr; i++) {
 +              entry = items_to_remove->items[i].util;
 +              hashmap_remove(dir_renames, entry, NULL);
 +      }
 +      string_list_clear(items_to_remove, 0);
 +}
 +
 +/*
 + * See if there is a directory rename for path, and if there are any file
 + * level conflicts for the renamed location.  If there is a rename and
 + * there are no conflicts, return the new name.  Otherwise, return NULL.
 + */
 +static char *handle_path_level_conflicts(struct merge_options *o,
 +                                       const char *path,
 +                                       struct dir_rename_entry *entry,
 +                                       struct hashmap *collisions,
 +                                       struct tree *tree)
 +{
 +      char *new_path = NULL;
 +      struct collision_entry *collision_ent;
 +      int clean = 1;
 +      struct strbuf collision_paths = STRBUF_INIT;
 +
 +      /*
 +       * entry has the mapping of old directory name to new directory name
 +       * that we want to apply to path.
 +       */
 +      new_path = apply_dir_rename(entry, path);
 +
 +      if (!new_path) {
 +              /* This should only happen when entry->non_unique_new_dir set */
 +              if (!entry->non_unique_new_dir)
 +                      BUG("entry->non_unqiue_dir not set and !new_path");
 +              output(o, 1, _("CONFLICT (directory rename split): "
 +                             "Unclear where to place %s because directory "
 +                             "%s was renamed to multiple other directories, "
 +                             "with no destination getting a majority of the "
 +                             "files."),
 +                     path, entry->dir);
 +              clean = 0;
 +              return NULL;
 +      }
 +
 +      /*
 +       * The caller needs to have ensured that it has pre-populated
 +       * collisions with all paths that map to new_path.  Do a quick check
 +       * to ensure that's the case.
 +       */
 +      collision_ent = collision_find_entry(collisions, new_path);
 +      if (collision_ent == NULL)
 +              BUG("collision_ent is NULL");
 +
 +      /*
 +       * Check for one-sided add/add/.../add conflicts, i.e.
 +       * where implicit renames from the other side doing
 +       * directory rename(s) can affect this side of history
 +       * to put multiple paths into the same location.  Warn
 +       * and bail on directory renames for such paths.
 +       */
 +      if (collision_ent->reported_already) {
 +              clean = 0;
 +      } else if (tree_has_path(tree, new_path)) {
 +              collision_ent->reported_already = 1;
 +              strbuf_add_separated_string_list(&collision_paths, ", ",
 +                                               &collision_ent->source_files);
 +              output(o, 1, _("CONFLICT (implicit dir rename): Existing "
 +                             "file/dir at %s in the way of implicit "
 +                             "directory rename(s) putting the following "
 +                             "path(s) there: %s."),
 +                     new_path, collision_paths.buf);
 +              clean = 0;
 +      } else if (collision_ent->source_files.nr > 1) {
 +              collision_ent->reported_already = 1;
 +              strbuf_add_separated_string_list(&collision_paths, ", ",
 +                                               &collision_ent->source_files);
 +              output(o, 1, _("CONFLICT (implicit dir rename): Cannot map "
 +                             "more than one path to %s; implicit directory "
 +                             "renames tried to put these paths there: %s"),
 +                     new_path, collision_paths.buf);
 +              clean = 0;
 +      }
 +
 +      /* Free memory we no longer need */
 +      strbuf_release(&collision_paths);
 +      if (!clean && new_path) {
 +              free(new_path);
 +              return NULL;
 +      }
 +
 +      return new_path;
 +}
 +
 +/*
 + * There are a couple things we want to do at the directory level:
 + *   1. Check for both sides renaming to the same thing, in order to avoid
 + *      implicit renaming of files that should be left in place.  (See
 + *      testcase 6b in t6043 for details.)
 + *   2. Prune directory renames if there are still files left in the
 + *      the original directory.  These represent a partial directory rename,
 + *      i.e. a rename where only some of the files within the directory
 + *      were renamed elsewhere.  (Technically, this could be done earlier
 + *      in get_directory_renames(), except that would prevent us from
 + *      doing the previous check and thus failing testcase 6b.)
 + *   3. Check for rename/rename(1to2) conflicts (at the directory level).
 + *      In the future, we could potentially record this info as well and
 + *      omit reporting rename/rename(1to2) conflicts for each path within
 + *      the affected directories, thus cleaning up the merge output.
 + *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
 + *         directory level, because merging directories is fine.  If it
 + *         causes conflicts for files within those merged directories, then
 + *         that should be detected at the individual path level.
 + */
 +static void handle_directory_level_conflicts(struct merge_options *o,
 +                                           struct hashmap *dir_re_head,
 +                                           struct tree *head,
 +                                           struct hashmap *dir_re_merge,
 +                                           struct tree *merge)
 +{
 +      struct hashmap_iter iter;
 +      struct dir_rename_entry *head_ent;
 +      struct dir_rename_entry *merge_ent;
 +
 +      struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
 +      struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
 +
 +      hashmap_iter_init(dir_re_head, &iter);
 +      while ((head_ent = hashmap_iter_next(&iter))) {
 +              merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
 +              if (merge_ent &&
 +                  !head_ent->non_unique_new_dir &&
 +                  !merge_ent->non_unique_new_dir &&
 +                  !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
 +                      /* 1. Renamed identically; remove it from both sides */
 +                      string_list_append(&remove_from_head,
 +                                         head_ent->dir)->util = head_ent;
 +                      strbuf_release(&head_ent->new_dir);
 +                      string_list_append(&remove_from_merge,
 +                                         merge_ent->dir)->util = merge_ent;
 +                      strbuf_release(&merge_ent->new_dir);
 +              } else if (tree_has_path(head, head_ent->dir)) {
 +                      /* 2. This wasn't a directory rename after all */
 +                      string_list_append(&remove_from_head,
 +                                         head_ent->dir)->util = head_ent;
 +                      strbuf_release(&head_ent->new_dir);
 +              }
 +      }
 +
 +      remove_hashmap_entries(dir_re_head, &remove_from_head);
 +      remove_hashmap_entries(dir_re_merge, &remove_from_merge);
 +
 +      hashmap_iter_init(dir_re_merge, &iter);
 +      while ((merge_ent = hashmap_iter_next(&iter))) {
 +              head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
 +              if (tree_has_path(merge, merge_ent->dir)) {
 +                      /* 2. This wasn't a directory rename after all */
 +                      string_list_append(&remove_from_merge,
 +                                         merge_ent->dir)->util = merge_ent;
 +              } else if (head_ent &&
 +                         !head_ent->non_unique_new_dir &&
 +                         !merge_ent->non_unique_new_dir) {
 +                      /* 3. rename/rename(1to2) */
 +                      /*
 +                       * We can assume it's not rename/rename(1to1) because
 +                       * that was case (1), already checked above.  So we
 +                       * know that head_ent->new_dir and merge_ent->new_dir
 +                       * are different strings.
 +                       */
 +                      output(o, 1, _("CONFLICT (rename/rename): "
 +                                     "Rename directory %s->%s in %s. "
 +                                     "Rename directory %s->%s in %s"),
 +                             head_ent->dir, head_ent->new_dir.buf, o->branch1,
 +                             head_ent->dir, merge_ent->new_dir.buf, o->branch2);
 +                      string_list_append(&remove_from_head,
 +                                         head_ent->dir)->util = head_ent;
 +                      strbuf_release(&head_ent->new_dir);
 +                      string_list_append(&remove_from_merge,
 +                                         merge_ent->dir)->util = merge_ent;
 +                      strbuf_release(&merge_ent->new_dir);
 +              }
 +      }
 +
 +      remove_hashmap_entries(dir_re_head, &remove_from_head);
 +      remove_hashmap_entries(dir_re_merge, &remove_from_merge);
 +}
 +
 +static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs,
 +                                           struct tree *tree)
 +{
 +      struct hashmap *dir_renames;
 +      struct hashmap_iter iter;
 +      struct dir_rename_entry *entry;
 +      int i;
 +
 +      /*
 +       * Typically, we think of a directory rename as all files from a
 +       * certain directory being moved to a target directory.  However,
 +       * what if someone first moved two files from the original
 +       * directory in one commit, and then renamed the directory
 +       * somewhere else in a later commit?  At merge time, we just know
 +       * that files from the original directory went to two different
 +       * places, and that the bulk of them ended up in the same place.
 +       * We want each directory rename to represent where the bulk of the
 +       * files from that directory end up; this function exists to find
 +       * where the bulk of the files went.
 +       *
 +       * The first loop below simply iterates through the list of file
 +       * renames, finding out how often each directory rename pair
 +       * possibility occurs.
 +       */
 +      dir_renames = xmalloc(sizeof(*dir_renames));
 +      dir_rename_init(dir_renames);
 +      for (i = 0; i < pairs->nr; ++i) {
 +              struct string_list_item *item;
 +              int *count;
 +              struct diff_filepair *pair = pairs->queue[i];
 +              char *old_dir, *new_dir;
 +
 +              /* File not part of directory rename if it wasn't renamed */
 +              if (pair->status != 'R')
 +                      continue;
 +
 +              get_renamed_dir_portion(pair->one->path, pair->two->path,
 +                                      &old_dir,        &new_dir);
 +              if (!old_dir)
 +                      /* Directory didn't change at all; ignore this one. */
 +                      continue;
 +
 +              entry = dir_rename_find_entry(dir_renames, old_dir);
 +              if (!entry) {
 +                      entry = xmalloc(sizeof(*entry));
 +                      dir_rename_entry_init(entry, old_dir);
 +                      hashmap_put(dir_renames, entry);
 +              } else {
 +                      free(old_dir);
 +              }
 +              item = string_list_lookup(&entry->possible_new_dirs, new_dir);
 +              if (!item) {
 +                      item = string_list_insert(&entry->possible_new_dirs,
 +                                                new_dir);
 +                      item->util = xcalloc(1, sizeof(int));
 +              } else {
 +                      free(new_dir);
 +              }
 +              count = item->util;
 +              *count += 1;
 +      }
 +
 +      /*
 +       * For each directory with files moved out of it, we find out which
 +       * target directory received the most files so we can declare it to
 +       * be the "winning" target location for the directory rename.  This
 +       * winner gets recorded in new_dir.  If there is no winner
 +       * (multiple target directories received the same number of files),
 +       * we set non_unique_new_dir.  Once we've determined the winner (or
 +       * that there is no winner), we no longer need possible_new_dirs.
 +       */
 +      hashmap_iter_init(dir_renames, &iter);
 +      while ((entry = hashmap_iter_next(&iter))) {
 +              int max = 0;
 +              int bad_max = 0;
 +              char *best = NULL;
 +
 +              for (i = 0; i < entry->possible_new_dirs.nr; i++) {
 +                      int *count = entry->possible_new_dirs.items[i].util;
 +
 +                      if (*count == max)
 +                              bad_max = max;
 +                      else if (*count > max) {
 +                              max = *count;
 +                              best = entry->possible_new_dirs.items[i].string;
 +                      }
 +              }
 +              if (bad_max == max)
 +                      entry->non_unique_new_dir = 1;
 +              else {
 +                      assert(entry->new_dir.len == 0);
 +                      strbuf_addstr(&entry->new_dir, best);
 +              }
 +              /*
 +               * The relevant directory sub-portion of the original full
 +               * filepaths were xstrndup'ed before inserting into
 +               * possible_new_dirs, and instead of manually iterating the
 +               * list and free'ing each, just lie and tell
 +               * possible_new_dirs that it did the strdup'ing so that it
 +               * will free them for us.
 +               */
 +              entry->possible_new_dirs.strdup_strings = 1;
 +              string_list_clear(&entry->possible_new_dirs, 1);
 +      }
 +
 +      return dir_renames;
 +}
 +
 +static struct dir_rename_entry *check_dir_renamed(const char *path,
 +                                                struct hashmap *dir_renames)
 +{
 +      char temp[PATH_MAX];
 +      char *end;
 +      struct dir_rename_entry *entry;
 +
 +      strcpy(temp, path);
 +      while ((end = strrchr(temp, '/'))) {
 +              *end = '\0';
 +              entry = dir_rename_find_entry(dir_renames, temp);
 +              if (entry)
 +                      return entry;
 +      }
 +      return NULL;
 +}
 +
 +static void compute_collisions(struct hashmap *collisions,
 +                             struct hashmap *dir_renames,
 +                             struct diff_queue_struct *pairs)
 +{
 +      int i;
 +
 +      /*
 +       * Multiple files can be mapped to the same path due to directory
 +       * renames done by the other side of history.  Since that other
 +       * side of history could have merged multiple directories into one,
 +       * if our side of history added the same file basename to each of
 +       * those directories, then all N of them would get implicitly
 +       * renamed by the directory rename detection into the same path,
 +       * and we'd get an add/add/.../add conflict, and all those adds
 +       * from *this* side of history.  This is not representable in the
 +       * index, and users aren't going to easily be able to make sense of
 +       * it.  So we need to provide a good warning about what's
 +       * happening, and fall back to no-directory-rename detection
 +       * behavior for those paths.
 +       *
 +       * See testcases 9e and all of section 5 from t6043 for examples.
 +       */
 +      collision_init(collisions);
 +
 +      for (i = 0; i < pairs->nr; ++i) {
 +              struct dir_rename_entry *dir_rename_ent;
 +              struct collision_entry *collision_ent;
 +              char *new_path;
 +              struct diff_filepair *pair = pairs->queue[i];
 +
 +              if (pair->status != 'A' && pair->status != 'R')
 +                      continue;
 +              dir_rename_ent = check_dir_renamed(pair->two->path,
 +                                                 dir_renames);
 +              if (!dir_rename_ent)
 +                      continue;
 +
 +              new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
 +              if (!new_path)
 +                      /*
 +                       * dir_rename_ent->non_unique_new_path is true, which
 +                       * means there is no directory rename for us to use,
 +                       * which means it won't cause us any additional
 +                       * collisions.
 +                       */
 +                      continue;
 +              collision_ent = collision_find_entry(collisions, new_path);
 +              if (!collision_ent) {
 +                      collision_ent = xcalloc(1,
 +                                              sizeof(struct collision_entry));
 +                      hashmap_entry_init(collision_ent, strhash(new_path));
 +                      hashmap_put(collisions, collision_ent);
 +                      collision_ent->target_file = new_path;
 +              } else {
 +                      free(new_path);
 +              }
 +              string_list_insert(&collision_ent->source_files,
 +                                 pair->two->path);
 +      }
 +}
 +
 +static char *check_for_directory_rename(struct merge_options *o,
 +                                      const char *path,
 +                                      struct tree *tree,
 +                                      struct hashmap *dir_renames,
 +                                      struct hashmap *dir_rename_exclusions,
 +                                      struct hashmap *collisions,
 +                                      int *clean_merge)
 +{
 +      char *new_path = NULL;
 +      struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
 +      struct dir_rename_entry *oentry = NULL;
 +
 +      if (!entry)
 +              return new_path;
 +
 +      /*
 +       * This next part is a little weird.  We do not want to do an
 +       * implicit rename into a directory we renamed on our side, because
 +       * that will result in a spurious rename/rename(1to2) conflict.  An
 +       * example:
 +       *   Base commit: dumbdir/afile, otherdir/bfile
 +       *   Side 1:      smrtdir/afile, otherdir/bfile
 +       *   Side 2:      dumbdir/afile, dumbdir/bfile
 +       * Here, while working on Side 1, we could notice that otherdir was
 +       * renamed/merged to dumbdir, and change the diff_filepair for
 +       * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
 +       * 2 will notice the rename from dumbdir to smrtdir, and do the
 +       * transitive rename to move it from dumbdir/bfile to
 +       * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
 +       * smrtdir, a rename/rename(1to2) conflict.  We really just want
 +       * the file to end up in smrtdir.  And the way to achieve that is
 +       * to not let Side1 do the rename to dumbdir, since we know that is
 +       * the source of one of our directory renames.
 +       *
 +       * That's why oentry and dir_rename_exclusions is here.
 +       *
 +       * As it turns out, this also prevents N-way transient rename
 +       * confusion; See testcases 9c and 9d of t6043.
 +       */
 +      oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
 +      if (oentry) {
 +              output(o, 1, _("WARNING: Avoiding applying %s -> %s rename "
 +                             "to %s, because %s itself was renamed."),
 +                     entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
 +      } else {
 +              new_path = handle_path_level_conflicts(o, path, entry,
 +                                                     collisions, tree);
 +              *clean_merge &= (new_path != NULL);
 +      }
 +
 +      return new_path;
 +}
 +
 +static void apply_directory_rename_modifications(struct merge_options *o,
 +                                               struct diff_filepair *pair,
 +                                               char *new_path,
 +                                               struct rename *re,
 +                                               struct tree *tree,
 +                                               struct tree *o_tree,
 +                                               struct tree *a_tree,
 +                                               struct tree *b_tree,
 +                                               struct string_list *entries,
 +                                               int *clean)
 +{
 +      struct string_list_item *item;
 +      int stage = (tree == a_tree ? 2 : 3);
 +      int update_wd;
 +
 +      /*
 +       * In all cases where we can do directory rename detection,
 +       * unpack_trees() will have read pair->two->path into the
 +       * index and the working copy.  We need to remove it so that
 +       * we can instead place it at new_path.  It is guaranteed to
 +       * not be untracked (unpack_trees() would have errored out
 +       * saying the file would have been overwritten), but it might
 +       * be dirty, though.
 +       */
 +      update_wd = !was_dirty(o, pair->two->path);
 +      if (!update_wd)
 +              output(o, 1, _("Refusing to lose dirty file at %s"),
 +                     pair->two->path);
 +      remove_file(o, 1, pair->two->path, !update_wd);
 +
 +      /* Find or create a new re->dst_entry */
 +      item = string_list_lookup(entries, new_path);
 +      if (item) {
 +              /*
 +               * Since we're renaming on this side of history, and it's
 +               * due to a directory rename on the other side of history
 +               * (which we only allow when the directory in question no
 +               * longer exists on the other side of history), the
 +               * original entry for re->dst_entry is no longer
 +               * necessary...
 +               */
 +              re->dst_entry->processed = 1;
 +
 +              /*
 +               * ...because we'll be using this new one.
 +               */
 +              re->dst_entry = item->util;
 +      } else {
 +              /*
 +               * re->dst_entry is for the before-dir-rename path, and we
 +               * need it to hold information for the after-dir-rename
 +               * path.  Before creating a new entry, we need to mark the
 +               * old one as unnecessary (...unless it is shared by
 +               * src_entry, i.e. this didn't use to be a rename, in which
 +               * case we can just allow the normal processing to happen
 +               * for it).
 +               */
 +              if (pair->status == 'R')
 +                      re->dst_entry->processed = 1;
 +
 +              re->dst_entry = insert_stage_data(new_path,
 +                                                o_tree, a_tree, b_tree,
 +                                                entries);
 +              item = string_list_insert(entries, new_path);
 +              item->util = re->dst_entry;
 +      }
 +
 +      /*
 +       * Update the stage_data with the information about the path we are
 +       * moving into place.  That slot will be empty and available for us
 +       * to write to because of the collision checks in
 +       * handle_path_level_conflicts().  In other words,
 +       * re->dst_entry->stages[stage].oid will be the null_oid, so it's
 +       * open for us to write to.
 +       *
 +       * It may be tempting to actually update the index at this point as
 +       * well, using update_stages_for_stage_data(), but as per the big
 +       * "NOTE" in update_stages(), doing so will modify the current
 +       * in-memory index which will break calls to would_lose_untracked()
 +       * that we need to make.  Instead, we need to just make sure that
 +       * the various conflict_rename_*() functions update the index
 +       * explicitly rather than relying on unpack_trees() to have done it.
 +       */
 +      get_tree_entry(&tree->object.oid,
 +                     pair->two->path,
 +                     &re->dst_entry->stages[stage].oid,
 +                     &re->dst_entry->stages[stage].mode);
 +
 +      /* Update pair status */
 +      if (pair->status == 'A') {
 +              /*
 +               * Recording rename information for this add makes it look
 +               * like a rename/delete conflict.  Make sure we can
 +               * correctly handle this as an add that was moved to a new
 +               * directory instead of reporting a rename/delete conflict.
 +               */
 +              re->add_turned_into_rename = 1;
 +      }
 +      /*
 +       * We don't actually look at pair->status again, but it seems
 +       * pedagogically correct to adjust it.
 +       */
 +      pair->status = 'R';
 +
 +      /*
 +       * Finally, record the new location.
 +       */
 +      pair->two->path = new_path;
 +}
 +
 +/*
 + * Get information of all renames which occurred in 'pairs', making use of
 + * any implicit directory renames inferred from the other side of history.
 + * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
 + * to be able to associate the correct cache entries with the rename
 + * information; tree is always equal to either a_tree or b_tree.
 + */
 +static struct string_list *get_renames(struct merge_options *o,
 +                                     struct diff_queue_struct *pairs,
 +                                     struct hashmap *dir_renames,
 +                                     struct hashmap *dir_rename_exclusions,
 +                                     struct tree *tree,
 +                                     struct tree *o_tree,
 +                                     struct tree *a_tree,
 +                                     struct tree *b_tree,
 +                                     struct string_list *entries,
 +                                     int *clean_merge)
 +{
 +      int i;
 +      struct hashmap collisions;
 +      struct hashmap_iter iter;
 +      struct collision_entry *e;
 +      struct string_list *renames;
 +
 +      compute_collisions(&collisions, dir_renames, pairs);
 +      renames = xcalloc(1, sizeof(struct string_list));
 +
 +      for (i = 0; i < pairs->nr; ++i) {
 +              struct string_list_item *item;
 +              struct rename *re;
 +              struct diff_filepair *pair = pairs->queue[i];
 +              char *new_path; /* non-NULL only with directory renames */
 +
 +              if (pair->status != 'A' && pair->status != 'R') {
 +                      diff_free_filepair(pair);
 +                      continue;
 +              }
 +              new_path = check_for_directory_rename(o, pair->two->path, tree,
 +                                                    dir_renames,
 +                                                    dir_rename_exclusions,
 +                                                    &collisions,
 +                                                    clean_merge);
 +              if (pair->status != 'R' && !new_path) {
 +                      diff_free_filepair(pair);
 +                      continue;
 +              }
 +
 +              re = xmalloc(sizeof(*re));
 +              re->processed = 0;
 +              re->add_turned_into_rename = 0;
 +              re->pair = pair;
 +              item = string_list_lookup(entries, re->pair->one->path);
 +              if (!item)
 +                      re->src_entry = insert_stage_data(re->pair->one->path,
 +                                      o_tree, a_tree, b_tree, entries);
 +              else
 +                      re->src_entry = item->util;
 +
 +              item = string_list_lookup(entries, re->pair->two->path);
 +              if (!item)
 +                      re->dst_entry = insert_stage_data(re->pair->two->path,
 +                                      o_tree, a_tree, b_tree, entries);
 +              else
 +                      re->dst_entry = item->util;
 +              item = string_list_insert(renames, pair->one->path);
 +              item->util = re;
 +              if (new_path)
 +                      apply_directory_rename_modifications(o, pair, new_path,
 +                                                           re, tree, o_tree,
 +                                                           a_tree, b_tree,
 +                                                           entries,
 +                                                           clean_merge);
 +      }
 +
 +      hashmap_iter_init(&collisions, &iter);
 +      while ((e = hashmap_iter_next(&iter))) {
 +              free(e->target_file);
 +              string_list_clear(&e->source_files, 0);
 +      }
 +      hashmap_free(&collisions, 1);
 +      return renames;
 +}
 +
  static int process_renames(struct merge_options *o,
                           struct string_list *a_renames,
                           struct string_list *b_renames)
                        const char *ren2_dst = ren2->pair->two->path;
                        enum rename_type rename_type;
                        if (strcmp(ren1_src, ren2_src) != 0)
 -                              die("BUG: ren1_src != ren2_src");
 +                              BUG("ren1_src != ren2_src");
                        ren2->dst_entry->processed = 1;
                        ren2->processed = 1;
                        if (strcmp(ren1_dst, ren2_dst) != 0) {
                        ren2 = lookup->util;
                        ren2_dst = ren2->pair->two->path;
                        if (strcmp(ren1_dst, ren2_dst) != 0)
 -                              die("BUG: ren1_dst != ren2_dst");
 +                              BUG("ren1_dst != ren2_dst");
  
                        clean_merge = 0;
                        ren2->processed = 1;
                         * add-source case).
                         */
                        remove_file(o, 1, ren1_src,
 -                                  renamed_stage == 2 || !was_tracked(ren1_src));
 +                                  renamed_stage == 2 || !was_tracked(o, ren1_src));
  
                        oidcpy(&src_other.oid,
                               &ren1->src_entry->stages[other_stage].oid);
                        dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
                        try_merge = 0;
  
 -                      if (oid_eq(&src_other.oid, &null_oid)) {
 +                      if (oid_eq(&src_other.oid, &null_oid) &&
 +                          ren1->add_turned_into_rename) {
 +                              setup_rename_conflict_info(RENAME_DIR,
 +                                                         ren1->pair,
 +                                                         NULL,
 +                                                         branch1,
 +                                                         branch2,
 +                                                         ren1->dst_entry,
 +                                                         NULL,
 +                                                         o,
 +                                                         NULL,
 +                                                         NULL);
 +                      } else if (oid_eq(&src_other.oid, &null_oid)) {
                                setup_rename_conflict_info(RENAME_DELETE,
                                                           ren1->pair,
                                                           NULL,
@@@ -2799,105 -1824,6 +2799,105 @@@ cleanup_and_return
        return clean_merge;
  }
  
 +struct rename_info {
 +      struct string_list *head_renames;
 +      struct string_list *merge_renames;
 +};
 +
 +static void initial_cleanup_rename(struct diff_queue_struct *pairs,
 +                                 struct hashmap *dir_renames)
 +{
 +      struct hashmap_iter iter;
 +      struct dir_rename_entry *e;
 +
 +      hashmap_iter_init(dir_renames, &iter);
 +      while ((e = hashmap_iter_next(&iter))) {
 +              free(e->dir);
 +              strbuf_release(&e->new_dir);
 +              /* possible_new_dirs already cleared in get_directory_renames */
 +      }
 +      hashmap_free(dir_renames, 1);
 +      free(dir_renames);
 +
 +      free(pairs->queue);
 +      free(pairs);
 +}
 +
 +static int handle_renames(struct merge_options *o,
 +                        struct tree *common,
 +                        struct tree *head,
 +                        struct tree *merge,
 +                        struct string_list *entries,
 +                        struct rename_info *ri)
 +{
 +      struct diff_queue_struct *head_pairs, *merge_pairs;
 +      struct hashmap *dir_re_head, *dir_re_merge;
 +      int clean = 1;
 +
 +      ri->head_renames = NULL;
 +      ri->merge_renames = NULL;
 +
 +      if (!merge_detect_rename(o))
 +              return 1;
 +
 +      head_pairs = get_diffpairs(o, common, head);
 +      merge_pairs = get_diffpairs(o, common, merge);
 +
 +      dir_re_head = get_directory_renames(head_pairs, head);
 +      dir_re_merge = get_directory_renames(merge_pairs, merge);
 +
 +      handle_directory_level_conflicts(o,
 +                                       dir_re_head, head,
 +                                       dir_re_merge, merge);
 +
 +      ri->head_renames  = get_renames(o, head_pairs,
 +                                      dir_re_merge, dir_re_head, head,
 +                                      common, head, merge, entries,
 +                                      &clean);
 +      if (clean < 0)
 +              goto cleanup;
 +      ri->merge_renames = get_renames(o, merge_pairs,
 +                                      dir_re_head, dir_re_merge, merge,
 +                                      common, head, merge, entries,
 +                                      &clean);
 +      if (clean < 0)
 +              goto cleanup;
 +      clean &= process_renames(o, ri->head_renames, ri->merge_renames);
 +
 +cleanup:
 +      /*
 +       * Some cleanup is deferred until cleanup_renames() because the
 +       * data structures are still needed and referenced in
 +       * process_entry().  But there are a few things we can free now.
 +       */
 +      initial_cleanup_rename(head_pairs, dir_re_head);
 +      initial_cleanup_rename(merge_pairs, dir_re_merge);
 +
 +      return clean;
 +}
 +
 +static void final_cleanup_rename(struct string_list *rename)
 +{
 +      const struct rename *re;
 +      int i;
 +
 +      if (rename == NULL)
 +              return;
 +
 +      for (i = 0; i < rename->nr; i++) {
 +              re = rename->items[i].util;
 +              diff_free_filepair(re->pair);
 +      }
 +      string_list_clear(rename, 1);
 +      free(rename);
 +}
 +
 +static void final_cleanup_renames(struct rename_info *re_info)
 +{
 +      final_cleanup_rename(re_info->head_renames);
 +      final_cleanup_rename(re_info->merge_renames);
 +}
 +
  static struct object_id *stage_oid(const struct object_id *oid, unsigned mode)
  {
        return (is_null_oid(oid) || mode == 0) ? NULL: (struct object_id *)oid;
@@@ -2988,7 -1914,6 +2988,7 @@@ static int handle_modify_delete(struct 
  
  static int merge_content(struct merge_options *o,
                         const char *path,
 +                       int is_dirty,
                         struct object_id *o_oid, int o_mode,
                         struct object_id *a_oid, int a_mode,
                         struct object_id *b_oid, int b_mode,
                               S_ISGITLINK(pair1->two->mode)))
                        df_conflict_remains = 1;
        }
 -      if (merge_file_special_markers(o, &one, &a, &b,
 +      if (merge_file_special_markers(o, &one, &a, &b, path,
                                       o->branch1, path1,
                                       o->branch2, path2, &mfi))
                return -1;
  
 -      if (mfi.clean && !df_conflict_remains &&
 -          oid_eq(&mfi.oid, a_oid) && mfi.mode == a_mode) {
 -              int path_renamed_outside_HEAD;
 +      /*
 +       * We can skip updating the working tree file iff:
 +       *   a) The merge is clean
 +       *   b) The merge matches what was in HEAD (content, mode, pathname)
 +       *   c) The target path is usable (i.e. not involved in D/F conflict)
 +       */
 +      if (mfi.clean &&
 +          was_tracked_and_matches(o, path, &mfi.oid, mfi.mode) &&
 +          !df_conflict_remains) {
                output(o, 3, _("Skipped %s (merged same as existing)"), path);
 -              /*
 -               * The content merge resulted in the same file contents we
 -               * already had.  We can return early if those file contents
 -               * are recorded at the correct path (which may not be true
 -               * if the merge involves a rename).
 -               */
 -              path_renamed_outside_HEAD = !path2 || !strcmp(path, path2);
 -              if (!path_renamed_outside_HEAD) {
 -                      add_cacheinfo(o, mfi.mode, &mfi.oid, path,
 -                                    0, (!o->call_depth), 0);
 -                      return mfi.clean;
 -              }
 -      } else
 -              output(o, 2, _("Auto-merging %s"), path);
 +              if (add_cacheinfo(o, mfi.mode, &mfi.oid, path,
 +                                0, (!o->call_depth && !is_dirty), 0))
 +                      return -1;
 +              return mfi.clean;
 +      }
  
        if (!mfi.clean) {
                if (S_ISGITLINK(mfi.mode))
                                return -1;
        }
  
 -      if (df_conflict_remains) {
 +      if (df_conflict_remains || is_dirty) {
                char *new_path;
                if (o->call_depth) {
                        remove_file_from_cache(path);
                                if (update_stages(o, path, &one, &a, &b))
                                        return -1;
                        } else {
 -                              int file_from_stage2 = was_tracked(path);
 +                              int file_from_stage2 = was_tracked(o, path);
                                struct diff_filespec merged;
                                oidcpy(&merged.oid, &mfi.oid);
                                merged.mode = mfi.mode;
  
                }
                new_path = unique_path(o, path, rename_conflict_info->branch1);
 +              if (is_dirty) {
 +                      output(o, 1, _("Refusing to lose dirty file at %s"),
 +                             path);
 +              }
                output(o, 1, _("Adding as %s instead"), new_path);
                if (update_file(o, 0, &mfi.oid, mfi.mode, new_path)) {
                        free(new_path);
                mfi.clean = 0;
        } else if (update_file(o, mfi.clean, &mfi.oid, mfi.mode, path))
                return -1;
 -      return mfi.clean;
 +      return !is_dirty && mfi.clean;
 +}
 +
 +static int conflict_rename_normal(struct merge_options *o,
 +                                const char *path,
 +                                struct object_id *o_oid, unsigned int o_mode,
 +                                struct object_id *a_oid, unsigned int a_mode,
 +                                struct object_id *b_oid, unsigned int b_mode,
 +                                struct rename_conflict_info *ci)
 +{
 +      /* Merge the content and write it out */
 +      return merge_content(o, path, was_dirty(o, path),
 +                           o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
 +                           ci);
  }
  
  /* Per entry merge function */
@@@ -3130,20 -2041,9 +3130,20 @@@ static int process_entry(struct merge_o
                switch (conflict_info->rename_type) {
                case RENAME_NORMAL:
                case RENAME_ONE_FILE_TO_ONE:
 -                      clean_merge = merge_content(o, path,
 -                                                  o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
 -                                                  conflict_info);
 +                      clean_merge = conflict_rename_normal(o,
 +                                                           path,
 +                                                           o_oid, o_mode,
 +                                                           a_oid, a_mode,
 +                                                           b_oid, b_mode,
 +                                                           conflict_info);
 +                      break;
 +              case RENAME_DIR:
 +                      clean_merge = 1;
 +                      if (conflict_rename_dir(o,
 +                                              conflict_info->pair1,
 +                                              conflict_info->branch1,
 +                                              conflict_info->branch2))
 +                              clean_merge = -1;
                        break;
                case RENAME_DELETE:
                        clean_merge = 0;
        } else if (a_oid && b_oid) {
                /* Case C: Added in both (check for same permissions) and */
                /* case D: Modified in both, but differently. */
 -              clean_merge = merge_content(o, path,
 +              int is_dirty = 0; /* unpack_trees would have bailed if dirty */
 +              clean_merge = merge_content(o, path, is_dirty,
                                            o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
                                            NULL);
        } else if (!o_oid && !a_oid && !b_oid) {
                 */
                remove_file(o, 1, path, !a_mode);
        } else
 -              die("BUG: fatal merge failure, shouldn't happen.");
 +              BUG("fatal merge failure, shouldn't happen.");
  
        return clean_merge;
  }
@@@ -3273,20 -2172,18 +3273,20 @@@ int merge_trees(struct merge_options *o
                return 1;
        }
  
 -      code = git_merge_trees(o->call_depth, common, head, merge);
 +      code = unpack_trees_start(o, common, head, merge);
  
        if (code != 0) {
                if (show(o, 4) || o->call_depth)
                        err(o, _("merging of trees %s and %s failed"),
                            oid_to_hex(&head->object.oid),
                            oid_to_hex(&merge->object.oid));
 +              unpack_trees_finish(o);
                return -1;
        }
  
        if (unmerged_cache()) {
 -              struct string_list *entries, *re_head, *re_merge;
 +              struct string_list *entries;
 +              struct rename_info re_info;
                int i;
                /*
                 * Only need the hashmap while processing entries, so
                get_files_dirs(o, merge);
  
                entries = get_unmerged();
 -              re_head  = get_renames(o, head, common, head, merge, entries);
 -              re_merge = get_renames(o, merge, common, head, merge, entries);
 -              clean = process_renames(o, re_head, re_merge);
 +              clean = handle_renames(o, common, head, merge, entries,
 +                                     &re_info);
                record_df_conflict_files(o, entries);
                if (clean < 0)
                        goto cleanup;
                for (i = 0; i < entries->nr; i++) {
                        struct stage_data *e = entries->items[i].util;
                        if (!e->processed)
 -                              die("BUG: unprocessed path??? %s",
 +                              BUG("unprocessed path??? %s",
                                    entries->items[i].string);
                }
  
  cleanup:
 -              string_list_clear(re_merge, 0);
 -              string_list_clear(re_head, 0);
 +              final_cleanup_renames(&re_info);
 +
                string_list_clear(entries, 1);
 +              free(entries);
  
                hashmap_free(&o->current_file_dir_set, 1);
  
 -              free(re_merge);
 -              free(re_head);
 -              free(entries);
 -
 -              if (clean < 0)
 +              if (clean < 0) {
 +                      unpack_trees_finish(o);
                        return clean;
 +              }
        }
        else
                clean = 1;
  
 +      unpack_trees_finish(o);
 +
        if (o->call_depth && !(*result = write_tree_from_memory(o)))
                return -1;
  
@@@ -3436,8 -2333,7 +3436,8 @@@ int merge_recursive(struct merge_option
                read_cache();
  
        o->ancestor = "merged common ancestors";
 -      clean = merge_trees(o, h1->tree, h2->tree, merged_common_ancestors->tree,
 +      clean = merge_trees(o, get_commit_tree(h1), get_commit_tree(h2),
 +                          get_commit_tree(merged_common_ancestors),
                            &mrtree);
        if (clean < 0) {
                flush_output(o);
@@@ -3515,18 -2411,9 +3515,18 @@@ int merge_recursive_generic(struct merg
  
  static void merge_recursive_config(struct merge_options *o)
  {
 +      char *value = NULL;
        git_config_get_int("merge.verbosity", &o->verbosity);
        git_config_get_int("diff.renamelimit", &o->diff_rename_limit);
        git_config_get_int("merge.renamelimit", &o->merge_rename_limit);
 +      if (!git_config_get_string("diff.renames", &value)) {
 +              o->diff_detect_rename = git_config_rename("diff.renames", value);
 +              free(value);
 +      }
 +      if (!git_config_get_string("merge.renames", &value)) {
 +              o->merge_detect_rename = git_config_rename("merge.renames", value);
 +              free(value);
 +      }
        git_config(git_xmerge_config, NULL);
  }
  
@@@ -3539,8 -2426,7 +3539,8 @@@ void init_merge_options(struct merge_op
        o->diff_rename_limit = -1;
        o->merge_rename_limit = -1;
        o->renormalize = 0;
 -      o->detect_rename = 1;
 +      o->diff_detect_rename = -1;
 +      o->merge_detect_rename = -1;
        merge_recursive_config(o);
        merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
        if (merge_verbosity)
@@@ -3591,16 -2477,16 +3591,16 @@@ int parse_merge_opt(struct merge_option
        else if (!strcmp(s, "no-renormalize"))
                o->renormalize = 0;
        else if (!strcmp(s, "no-renames"))
 -              o->detect_rename = 0;
 +              o->merge_detect_rename = 0;
        else if (!strcmp(s, "find-renames")) {
 -              o->detect_rename = 1;
 +              o->merge_detect_rename = 1;
                o->rename_score = 0;
        }
        else if (skip_prefix(s, "find-renames=", &arg) ||
                 skip_prefix(s, "rename-threshold=", &arg)) {
                if ((o->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
                        return -1;
 -              o->detect_rename = 1;
 +              o->merge_detect_rename = 1;
        }
        else
                return -1;