Merge branch 'en/dirty-merge-fixes'
authorJunio C Hamano <gitster@pobox.com>
Thu, 2 Aug 2018 22:30:44 +0000 (15:30 -0700)
committerJunio C Hamano <gitster@pobox.com>
Thu, 2 Aug 2018 22:30:45 +0000 (15:30 -0700)
The recursive merge strategy did not properly ensure there was no
change between HEAD and the index before performing its operation,
which has been corrected.

* en/dirty-merge-fixes:
merge: fix misleading pre-merge check documentation
merge-recursive: enforce rule that index matches head before merging
t6044: add more testcases with staged changes before a merge is invoked
merge-recursive: fix assumption that head tree being merged is HEAD
merge-recursive: make sure when we say we abort that we actually abort
t6044: add a testcase for index matching head, when head doesn't match HEAD
t6044: verify that merges expected to abort actually abort
index_has_changes(): avoid assuming operating on the_index
read-cache.c: move index_has_changes() from merge.c

1  2 
Documentation/git-merge.txt
builtin/am.c
cache.h
merge-recursive.c
merge.c
read-cache.c
index 537dca7faaed98f9c0e0364d35d63de7182f5071,141bd7228432cc91518a10646eb257fda343a831..eb36837f86e21423e0a543eceb67313eed27d310
@@@ -12,7 -12,7 +12,7 @@@ SYNOPSI
  'git merge' [-n] [--stat] [--no-commit] [--squash] [--[no-]edit]
        [-s <strategy>] [-X <strategy-option>] [-S[<keyid>]]
        [--[no-]allow-unrelated-histories]
 -      [--[no-]rerere-autoupdate] [-m <msg>] [<commit>...]
 +      [--[no-]rerere-autoupdate] [-m <msg>] [-F <file>] [<commit>...]
  'git merge' --abort
  'git merge' --continue
  
@@@ -57,7 -57,7 +57,7 @@@ reconstruct the original (pre-merge) ch
  discouraged: while possible, it may leave you in a state that is hard to
  back out of in the case of a conflict.
  
 -The fourth syntax ("`git merge --continue`") can only be run after the
 +The third syntax ("`git merge --continue`") can only be run after the
  merge has resulted in conflicts.
  
  OPTIONS
@@@ -75,14 -75,6 +75,14 @@@ The 'git fmt-merge-msg' command can b
  used to give a good default for automated 'git merge'
  invocations. The automated message can include the branch description.
  
 +-F <file>::
 +--file=<file>::
 +      Read the commit message to be used for the merge commit (in
 +      case one is created).
 ++
 +If `--log` is specified, a shortlog of the commits being merged
 +will be appended to the specified message.
 +
  --[no-]rerere-autoupdate::
        Allow the rerere mechanism to update the index with the
        result of auto-conflict resolution if possible.
@@@ -130,9 -122,9 +130,9 @@@ merge' may need to update
  
  To avoid recording unrelated changes in the merge commit,
  'git pull' and 'git merge' will also abort if there are any changes
- registered in the index relative to the `HEAD` commit.  (One
- exception is when the changed index entries are in the state that
would result from the merge already.)
+ registered in the index relative to the `HEAD` commit.  (Special
+ narrow exceptions to this rule may exist depending on which merge
strategy is in use, but generally, the index must match HEAD.)
  
  If all named commits are already ancestors of `HEAD`, 'git merge'
  will exit early with the message "Already up to date."
diff --combined builtin/am.c
index b6eeb46c4b65aefbc83b25d11d2ea28646990137,3d827d1a93bcc6d296ac7da112ce781a5fa082b5..2c19e69f585a46330718ddea1c13a095cefe9d19
@@@ -6,7 -6,7 +6,7 @@@
  #include "cache.h"
  #include "config.h"
  #include "builtin.h"
 -#include "exec_cmd.h"
 +#include "exec-cmd.h"
  #include "parse-options.h"
  #include "dir.h"
  #include "run-command.h"
@@@ -32,7 -32,6 +32,7 @@@
  #include "apply.h"
  #include "string-list.h"
  #include "packfile.h"
 +#include "repository.h"
  
  /**
   * Returns 1 if the file is empty or does not exist, 0 otherwise.
@@@ -404,11 -403,11 +404,11 @@@ static void am_load(struct am_state *st
        struct strbuf sb = STRBUF_INIT;
  
        if (read_state_file(&sb, state, "next", 1) < 0)
 -              die("BUG: state file 'next' does not exist");
 +              BUG("state file 'next' does not exist");
        state->cur = strtol(sb.buf, NULL, 10);
  
        if (read_state_file(&sb, state, "last", 1) < 0)
 -              die("BUG: state file 'last' does not exist");
 +              BUG("state file 'last' does not exist");
        state->last = strtol(sb.buf, NULL, 10);
  
        if (read_author_script(state) < 0)
@@@ -987,7 -986,7 +987,7 @@@ static int split_mail(struct am_state *
        case PATCH_FORMAT_MBOXRD:
                return split_mail_mbox(state, paths, keep_cr, 1);
        default:
 -              die("BUG: invalid patch_format");
 +              BUG("invalid patch_format");
        }
        return -1;
  }
@@@ -1042,7 -1041,7 +1042,7 @@@ static void am_setup(struct am_state *s
                str = "b";
                break;
        default:
 -              die("BUG: invalid value for state->keep");
 +              BUG("invalid value for state->keep");
        }
  
        write_state_text(state, "keep", str);
                str = "t";
                break;
        default:
 -              die("BUG: invalid value for state->scissors");
 +              BUG("invalid value for state->scissors");
        }
        write_state_text(state, "scissors", str);
  
@@@ -1217,7 -1216,7 +1217,7 @@@ static int parse_mail(struct am_state *
                mi.keep_non_patch_brackets_in_subject = 1;
                break;
        default:
 -              die("BUG: invalid value for state->keep");
 +              BUG("invalid value for state->keep");
        }
  
        if (state->message_id)
                mi.use_scissors = 1;
                break;
        default:
 -              die("BUG: invalid value for state->scissors");
 +              BUG("invalid value for state->scissors");
        }
  
        mi.input = xfopen(mail, "r");
@@@ -1401,10 -1400,9 +1401,10 @@@ static void write_index_patch(const str
        FILE *fp;
  
        if (!get_oid_tree("HEAD", &head))
 -              tree = lookup_tree(&head);
 +              tree = lookup_tree(the_repository, &head);
        else
 -              tree = lookup_tree(the_hash_algo->empty_tree);
 +              tree = lookup_tree(the_repository,
 +                                 the_repository->hash_algo->empty_tree);
  
        fp = xfopen(am_path(state, "patch"), "w");
        init_revisions(&rev_info, NULL);
@@@ -1465,7 -1463,7 +1465,7 @@@ static int run_apply(const struct am_st
        int options = 0;
  
        if (init_apply_state(&apply_state, NULL))
 -              die("BUG: init_apply_state() failed");
 +              BUG("init_apply_state() failed");
  
        argv_array_push(&apply_opts, "apply");
        argv_array_pushv(&apply_opts, state->git_apply_opts.argv);
                apply_state.apply_verbosity = verbosity_silent;
  
        if (check_apply_state(&apply_state, force_apply))
 -              die("BUG: check_apply_state() failed");
 +              BUG("check_apply_state() failed");
  
        argv_array_push(&apply_paths, am_path(state, "patch"));
  
@@@ -1544,7 -1542,7 +1544,7 @@@ static int fall_back_threeway(const str
        char *their_tree_name;
  
        if (get_oid("HEAD", &our_tree) < 0)
 -              hashcpy(our_tree.hash, EMPTY_TREE_SHA1_BIN);
 +              oidcpy(&our_tree, the_hash_algo->empty_tree);
  
        if (build_fake_ancestor(state, index_path))
                return error("could not build fake ancestor");
        discard_cache();
        read_cache_from(index_path);
  
 -      if (write_index_as_tree(orig_tree.hash, &the_index, index_path, 0, NULL))
 +      if (write_index_as_tree(&orig_tree, &the_index, index_path, 0, NULL))
                return error(_("Repository lacks necessary blobs to fall back on 3-way merge."));
  
        say(state, stdout, _("Using index info to reconstruct a base tree..."));
                return error(_("Did you hand edit your patch?\n"
                                "It does not apply to blobs recorded in its index."));
  
 -      if (write_index_as_tree(their_tree.hash, &the_index, index_path, 0, NULL))
 +      if (write_index_as_tree(&their_tree, &the_index, index_path, 0, NULL))
                return error("could not write tree");
  
        say(state, stdout, _("Falling back to patching base and 3-way merge..."));
@@@ -1628,13 -1626,12 +1628,13 @@@ static void do_commit(const struct am_s
        if (run_hook_le(NULL, "pre-applypatch", NULL))
                exit(1);
  
 -      if (write_cache_as_tree(tree.hash, 0, NULL))
 +      if (write_cache_as_tree(&tree, 0, NULL))
                die(_("git write-tree failed to write a tree"));
  
        if (!get_oid_commit("HEAD", &parent)) {
                old_oid = &parent;
 -              commit_list_insert(lookup_commit(&parent), &parents);
 +              commit_list_insert(lookup_commit(the_repository, &parent),
 +                                 &parents);
        } else {
                old_oid = NULL;
                say(state, stderr, _("applying to an empty history"));
@@@ -1766,7 -1763,7 +1766,7 @@@ static void am_run(struct am_state *sta
  
        refresh_and_write_cache();
  
-       if (index_has_changes(&sb)) {
+       if (index_has_changes(&the_index, NULL, &sb)) {
                write_state_bool(state, "dirtyindex", 1);
                die(_("Dirty index: cannot apply patches (dirty: %s)"), sb.buf);
        }
                         * Applying the patch to an earlier tree and merging
                         * the result may have produced the same tree as ours.
                         */
-                       if (!apply_status && !index_has_changes(NULL)) {
+                       if (!apply_status &&
+                           !index_has_changes(&the_index, NULL, NULL)) {
                                say(state, stdout, _("No changes -- Patch already applied."));
                                goto next;
                        }
                }
  
                if (apply_status) {
 -                      int advice_amworkdir = 1;
 -
                        printf_ln(_("Patch failed at %s %.*s"), msgnum(state),
                                linelen(state->msg), state->msg);
  
 -                      git_config_get_bool("advice.amworkdir", &advice_amworkdir);
 -
                        if (advice_amworkdir)
 -                              printf_ln(_("Use 'git am --show-current-patch' to see the failed patch"));
 +                              advise(_("Use 'git am --show-current-patch' to see the failed patch"));
  
                        die_user_resolve(state);
                }
@@@ -1861,7 -1863,7 +1862,7 @@@ next
         */
        if (!state->rebasing) {
                am_destroy(state);
 -              close_all_packs();
 +              close_all_packs(the_repository->objects);
                run_command_v_opt(argv_gc_auto, RUN_GIT_CMD);
        }
  }
@@@ -1877,7 -1879,7 +1878,7 @@@ static void am_resolve(struct am_state 
  
        say(state, stdout, _("Applying: %.*s"), linelen(state->msg), state->msg);
  
-       if (!index_has_changes(NULL)) {
+       if (!index_has_changes(&the_index, NULL, NULL)) {
                printf_ln(_("No changes - did you forget to use 'git add'?\n"
                        "If there is nothing left to stage, chances are that something else\n"
                        "already introduced the same changes; you might want to skip this patch."));
@@@ -2003,7 -2005,7 +2004,7 @@@ static int clean_index(const struct obj
        if (fast_forward_to(head_tree, head_tree, 1))
                return -1;
  
 -      if (write_cache_as_tree(index.hash, 0, NULL))
 +      if (write_cache_as_tree(&index, 0, NULL))
                return -1;
  
        index_tree = parse_tree_indirect(&index);
@@@ -2041,7 -2043,7 +2042,7 @@@ static void am_skip(struct am_state *st
        am_rerere_clear();
  
        if (get_oid("HEAD", &head))
 -              hashcpy(head.hash, EMPTY_TREE_SHA1_BIN);
 +              oidcpy(&head, the_hash_algo->empty_tree);
  
        if (clean_index(&head, &head))
                die(_("failed to clean index"));
@@@ -2104,11 -2106,11 +2105,11 @@@ static void am_abort(struct am_state *s
        curr_branch = resolve_refdup("HEAD", 0, &curr_head, NULL);
        has_curr_head = curr_branch && !is_null_oid(&curr_head);
        if (!has_curr_head)
 -              hashcpy(curr_head.hash, EMPTY_TREE_SHA1_BIN);
 +              oidcpy(&curr_head, the_hash_algo->empty_tree);
  
        has_orig_head = !get_oid("ORIG_HEAD", &orig_head);
        if (!has_orig_head)
 -              hashcpy(orig_head.hash, EMPTY_TREE_SHA1_BIN);
 +              oidcpy(&orig_head, the_hash_algo->empty_tree);
  
        clean_index(&curr_head, &orig_head);
  
@@@ -2230,12 -2232,12 +2231,12 @@@ int cmd_am(int argc, const char **argv
                        N_("pass -b flag to git-mailinfo"), KEEP_NON_PATCH),
                OPT_BOOL('m', "message-id", &state.message_id,
                        N_("pass -m flag to git-mailinfo")),
 -              { OPTION_SET_INT, 0, "keep-cr", &keep_cr, NULL,
 -                N_("pass --keep-cr flag to git-mailsplit for mbox format"),
 -                PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1},
 -              { OPTION_SET_INT, 0, "no-keep-cr", &keep_cr, NULL,
 -                N_("do not pass --keep-cr flag to git-mailsplit independent of am.keepcr"),
 -                PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 0},
 +              OPT_SET_INT_F(0, "keep-cr", &keep_cr,
 +                      N_("pass --keep-cr flag to git-mailsplit for mbox format"),
 +                      1, PARSE_OPT_NONEG),
 +              OPT_SET_INT_F(0, "no-keep-cr", &keep_cr,
 +                      N_("do not pass --keep-cr flag to git-mailsplit independent of am.keepcr"),
 +                      0, PARSE_OPT_NONEG),
                OPT_BOOL('c', "scissors", &state.scissors,
                        N_("strip everything before a scissors line")),
                OPT_PASSTHRU_ARGV(0, "whitespace", &state.git_apply_opts, N_("action"),
                ret = show_patch(&state);
                break;
        default:
 -              die("BUG: invalid resume value");
 +              BUG("invalid resume value");
        }
  
        am_state_release(&state);
diff --combined cache.h
index 05388e597b22632fb557603e1b173ce5bab69f4e,fad10c3bfe564caec98d53d1406eacb4ad6bf35a..87785fc83218cff355f02e2e92a7f819594e5b15
+++ b/cache.h
@@@ -15,7 -15,6 +15,7 @@@
  #include "path.h"
  #include "sha1-array.h"
  #include "repository.h"
 +#include "mem-pool.h"
  
  #include <zlib.h>
  typedef struct git_zstream {
@@@ -157,7 -156,6 +157,7 @@@ struct cache_entry 
        struct stat_data ce_stat_data;
        unsigned int ce_mode;
        unsigned int ce_flags;
 +      unsigned int mem_pool_allocated;
        unsigned int ce_namelen;
        unsigned int index;     /* for link extension */
        struct object_id oid;
  /* Forward structure decls */
  struct pathspec;
  struct child_process;
+ struct tree;
  
  /*
   * Copy the sha1 and stat state of a cache entry from one to
@@@ -229,7 -228,6 +230,7 @@@ static inline void copy_cache_entry(str
                                    const struct cache_entry *src)
  {
        unsigned int state = dst->ce_flags & CE_HASHED;
 +      int mem_pool_allocated = dst->mem_pool_allocated;
  
        /* Don't copy hash chain and name */
        memcpy(&dst->ce_stat_data, &src->ce_stat_data,
  
        /* Restore the hash state */
        dst->ce_flags = (dst->ce_flags & ~CE_HASHED) | state;
 +
 +      /* Restore the mem_pool_allocated flag */
 +      dst->mem_pool_allocated = mem_pool_allocated;
  }
  
  static inline unsigned create_ce_flags(unsigned stage)
@@@ -330,11 -325,10 +331,11 @@@ struct index_state 
                 drop_cache_tree : 1;
        struct hashmap name_hash;
        struct hashmap dir_hash;
 -      unsigned char sha1[20];
 +      struct object_id oid;
        struct untracked_cache *untracked;
        uint64_t fsmonitor_last_update;
        struct ewah_bitmap *fsmonitor_dirty;
 +      struct mem_pool *ce_mem_pool;
  };
  
  extern struct index_state the_index;
@@@ -346,60 -340,6 +347,60 @@@ extern void remove_name_hash(struct ind
  extern void free_name_hash(struct index_state *istate);
  
  
 +/* Cache entry creation and cleanup */
 +
 +/*
 + * Create cache_entry intended for use in the specified index. Caller
 + * is responsible for discarding the cache_entry with
 + * `discard_cache_entry`.
 + */
 +struct cache_entry *make_cache_entry(struct index_state *istate,
 +                                   unsigned int mode,
 +                                   const struct object_id *oid,
 +                                   const char *path,
 +                                   int stage,
 +                                   unsigned int refresh_options);
 +
 +struct cache_entry *make_empty_cache_entry(struct index_state *istate,
 +                                         size_t name_len);
 +
 +/*
 + * Create a cache_entry that is not intended to be added to an index.
 + * Caller is responsible for discarding the cache_entry
 + * with `discard_cache_entry`.
 + */
 +struct cache_entry *make_transient_cache_entry(unsigned int mode,
 +                                             const struct object_id *oid,
 +                                             const char *path,
 +                                             int stage);
 +
 +struct cache_entry *make_empty_transient_cache_entry(size_t name_len);
 +
 +/*
 + * Discard cache entry.
 + */
 +void discard_cache_entry(struct cache_entry *ce);
 +
 +/*
 + * Check configuration if we should perform extra validation on cache
 + * entries.
 + */
 +int should_validate_cache_entries(void);
 +
 +/*
 + * Duplicate a cache_entry. Allocate memory for the new entry from a
 + * memory_pool. Takes into account cache_entry fields that are meant
 + * for managing the underlying memory allocation of the cache_entry.
 + */
 +struct cache_entry *dup_cache_entry(const struct cache_entry *ce, struct index_state *istate);
 +
 +/*
 + * Validate the cache entries in the index.  This is an internal
 + * consistency check that the cache_entry structs are allocated from
 + * the expected memory pool.
 + */
 +void validate_cache_entries(const struct index_state *istate);
 +
  #ifndef NO_THE_INDEX_COMPATIBILITY_MACROS
  #define active_cache (the_index.cache)
  #define active_nr (the_index.cache_nr)
  #define read_blob_data_from_cache(path, sz) read_blob_data_from_index(&the_index, (path), (sz))
  #endif
  
 +#define TYPE_BITS 3
 +
 +/*
 + * Values in this enum (except those outside the 3 bit range) are part
 + * of pack file format. See Documentation/technical/pack-format.txt
 + * for more information.
 + */
  enum object_type {
        OBJ_BAD = -1,
        OBJ_NONE = 0,
@@@ -496,7 -429,6 +497,7 @@@ static inline enum object_type object_t
  #define GIT_ICASE_PATHSPECS_ENVIRONMENT "GIT_ICASE_PATHSPECS"
  #define GIT_QUARANTINE_ENVIRONMENT "GIT_QUARANTINE_PATH"
  #define GIT_OPTIONAL_LOCKS_ENVIRONMENT "GIT_OPTIONAL_LOCKS"
 +#define GIT_TEXT_DOMAIN_DIR_ENVIRONMENT "GIT_TEXTDOMAINDIR"
  
  /*
   * Environment variable used in handshaking the wire protocol.
   */
  extern const char * const local_repo_env[];
  
 -extern void setup_git_env(void);
 +extern void setup_git_env(const char *git_dir);
  
  /*
   * Returns true iff we have a configured git repository (either via
@@@ -545,8 -477,8 +546,8 @@@ extern const char *get_git_dir(void)
  extern const char *get_git_common_dir(void);
  extern char *get_object_directory(void);
  extern char *get_index_file(void);
 -extern char *get_graft_file(void);
 -extern int set_git_dir(const char *path);
 +extern char *get_graft_file(struct repository *r);
 +extern void set_git_dir(const char *path);
  extern int get_common_dir_noenv(struct strbuf *sb, const char *gitdir);
  extern int get_common_dir(struct strbuf *sb, const char *gitdir);
  extern const char *get_git_namespace(void);
@@@ -696,12 -628,15 +697,15 @@@ extern void move_index_extensions(struc
  extern int unmerged_index(const struct index_state *);
  
  /**
-  * Returns 1 if the index differs from HEAD, 0 otherwise. When on an unborn
-  * branch, returns 1 if there are entries in the index, 0 otherwise. If an
-  * strbuf is provided, the space-separated list of files that differ will be
-  * appended to it.
+  * Returns 1 if istate differs from tree, 0 otherwise.  If tree is NULL,
+  * compares istate to HEAD.  If tree is NULL and on an unborn branch,
+  * returns 1 if there are entries in istate, 0 otherwise.  If an strbuf is
+  * provided, the space-separated list of files that differ will be appended
+  * to it.
   */
- extern int index_has_changes(struct strbuf *sb);
+ extern int index_has_changes(const struct index_state *istate,
+                            struct tree *tree,
+                            struct strbuf *sb);
  
  extern int verify_path(const char *path, unsigned mode);
  extern int strcmp_offset(const char *s1, const char *s2, size_t *first_change);
@@@ -759,6 -694,7 +763,6 @@@ extern int remove_file_from_index(struc
  extern int add_to_index(struct index_state *, const char *path, struct stat *, int flags);
  extern int add_file_to_index(struct index_state *, const char *path, int flags);
  
 -extern struct cache_entry *make_cache_entry(unsigned int mode, const unsigned char *sha1, const char *path, int stage, unsigned int refresh_options);
  extern int chmod_index_entry(struct index_state *, struct cache_entry *ce, char flip);
  extern int ce_same_name(const struct cache_entry *a, const struct cache_entry *b);
  extern void set_object_name_for_intent_to_add_entry(struct cache_entry *ce);
@@@ -811,7 -747,7 +815,7 @@@ extern void fill_stat_cache_info(struc
  #define REFRESH_IGNORE_SUBMODULES     0x0010  /* ignore submodules */
  #define REFRESH_IN_PORCELAIN  0x0020  /* user friendly output, not "needs update" */
  extern int refresh_index(struct index_state *, unsigned int flags, const struct pathspec *pathspec, char *seen, const char *header_msg);
 -extern struct cache_entry *refresh_cache_entry(struct cache_entry *, unsigned int);
 +extern struct cache_entry *refresh_cache_entry(struct index_state *, struct cache_entry *, unsigned int);
  
  /*
   * Opportunistically update the index but do not complain if we can't.
@@@ -873,7 -809,6 +877,7 @@@ extern char *git_replace_ref_base
  
  extern int fsync_object_files;
  extern int core_preload_index;
 +extern int core_commit_graph;
  extern int core_apply_sparse_checkout;
  extern int precomposed_unicode;
  extern int protect_hfs;
@@@ -1009,6 -944,12 +1013,6 @@@ extern void check_repository_format(voi
  #define DATA_CHANGED    0x0020
  #define TYPE_CHANGED    0x0040
  
 -/*
 - * Put in `buf` the name of the file in the local object database that
 - * would be used to store a loose object with the specified sha1.
 - */
 -extern void sha1_file_name(struct strbuf *buf, const unsigned char *sha1);
 -
  /*
   * Return an abbreviated sha1 unique within this repository's object database.
   * The result will be at least `len` characters long, and will be NUL
   * more calls to find_unique_abbrev are made.
   *
   * The `_r` variant writes to a buffer supplied by the caller, which must be at
 - * least `GIT_SHA1_HEXSZ + 1` bytes. The return value is the number of bytes
 + * least `GIT_MAX_HEXSZ + 1` bytes. The return value is the number of bytes
   * written (excluding the NUL terminator).
   *
   * Note that while this version avoids the static buffer, it is not fully
   * reentrant, as it calls into other non-reentrant git code.
   */
 -extern const char *find_unique_abbrev(const unsigned char *sha1, int len);
 -extern int find_unique_abbrev_r(char *hex, const unsigned char *sha1, int len);
 +extern const char *find_unique_abbrev(const struct object_id *oid, int len);
 +extern int find_unique_abbrev_r(char *hex, const struct object_id *oid, int len);
  
  extern const unsigned char null_sha1[GIT_MAX_RAWSZ];
  extern const struct object_id null_oid;
  
  static inline int hashcmp(const unsigned char *sha1, const unsigned char *sha2)
  {
 -      return memcmp(sha1, sha2, GIT_SHA1_RAWSZ);
 +      return memcmp(sha1, sha2, the_hash_algo->rawsz);
  }
  
  static inline int oidcmp(const struct object_id *oid1, const struct object_id *oid2)
@@@ -1052,7 -993,7 +1056,7 @@@ static inline int is_null_oid(const str
  
  static inline void hashcpy(unsigned char *sha_dst, const unsigned char *sha_src)
  {
 -      memcpy(sha_dst, sha_src, GIT_SHA1_RAWSZ);
 +      memcpy(sha_dst, sha_src, the_hash_algo->rawsz);
  }
  
  static inline void oidcpy(struct object_id *dst, const struct object_id *src)
@@@ -1069,7 -1010,7 +1073,7 @@@ static inline struct object_id *oiddup(
  
  static inline void hashclr(unsigned char *hash)
  {
 -      memset(hash, 0, GIT_SHA1_RAWSZ);
 +      memset(hash, 0, the_hash_algo->rawsz);
  }
  
  static inline void oidclr(struct object_id *oid)
        memset(oid->hash, 0, GIT_MAX_RAWSZ);
  }
  
 -
 -#define EMPTY_TREE_SHA1_HEX \
 -      "4b825dc642cb6eb9a060e54bf8d69288fbee4904"
 -#define EMPTY_TREE_SHA1_BIN_LITERAL \
 -       "\x4b\x82\x5d\xc6\x42\xcb\x6e\xb9\xa0\x60" \
 -       "\xe5\x4b\xf8\xd6\x92\x88\xfb\xee\x49\x04"
 -extern const struct object_id empty_tree_oid;
 -#define EMPTY_TREE_SHA1_BIN (empty_tree_oid.hash)
 -
 -#define EMPTY_BLOB_SHA1_HEX \
 -      "e69de29bb2d1d6434b8b29ae775ad8c2e48c5391"
 -#define EMPTY_BLOB_SHA1_BIN_LITERAL \
 -      "\xe6\x9d\xe2\x9b\xb2\xd1\xd6\x43\x4b\x8b" \
 -      "\x29\xae\x77\x5a\xd8\xc2\xe4\x8c\x53\x91"
 -extern const struct object_id empty_blob_oid;
 +static inline void oidread(struct object_id *oid, const unsigned char *hash)
 +{
 +      memcpy(oid->hash, hash, the_hash_algo->rawsz);
 +}
  
  static inline int is_empty_blob_sha1(const unsigned char *sha1)
  {
@@@ -1102,9 -1054,6 +1106,9 @@@ static inline int is_empty_tree_oid(con
        return !oidcmp(oid, the_hash_algo->empty_tree);
  }
  
 +const char *empty_tree_oid_hex(void);
 +const char *empty_blob_oid_hex(void);
 +
  /* set default permissions by passing mode arguments to open(2) */
  int git_mkstemps_mode(char *pattern, int suffix_len, int mode);
  int git_mkstemp_mode(char *pattern, int mode);
@@@ -1252,15 -1201,98 +1256,15 @@@ extern char *xdg_config_home(const cha
   */
  extern char *xdg_cache_home(const char *filename);
  
 -extern void *read_sha1_file_extended(const unsigned char *sha1,
 -                                   enum object_type *type,
 -                                   unsigned long *size, int lookup_replace);
 -static inline void *read_sha1_file(const unsigned char *sha1, enum object_type *type, unsigned long *size)
 -{
 -      return read_sha1_file_extended(sha1, type, size, 1);
 -}
 -
 -/*
 - * This internal function is only declared here for the benefit of
 - * lookup_replace_object().  Please do not call it directly.
 - */
 -extern const unsigned char *do_lookup_replace_object(const unsigned char *sha1);
 -
 -/*
 - * If object sha1 should be replaced, return the replacement object's
 - * name (replaced recursively, if necessary).  The return value is
 - * either sha1 or a pointer to a permanently-allocated value.  When
 - * object replacement is suppressed, always return sha1.
 - */
 -static inline const unsigned char *lookup_replace_object(const unsigned char *sha1)
 -{
 -      if (!check_replace_refs)
 -              return sha1;
 -      return do_lookup_replace_object(sha1);
 -}
 -
 -/* Read and unpack a sha1 file into memory, write memory to a sha1 file */
 -extern int sha1_object_info(const unsigned char *, unsigned long *);
 -
 -extern int hash_object_file(const void *buf, unsigned long len,
 -                          const char *type, struct object_id *oid);
 -
 -extern int write_object_file(const void *buf, unsigned long len,
 -                           const char *type, struct object_id *oid);
 -
 -extern int hash_object_file_literally(const void *buf, unsigned long len,
 -                                    const char *type, struct object_id *oid,
 -                                    unsigned flags);
 -
 -extern int pretend_object_file(void *, unsigned long, enum object_type,
 -                             struct object_id *oid);
 -
 -extern int force_object_loose(const struct object_id *oid, time_t mtime);
 -
  extern int git_open_cloexec(const char *name, int flags);
  #define git_open(name) git_open_cloexec(name, O_RDONLY)
 -extern void *map_sha1_file(const unsigned char *sha1, unsigned long *size);
  extern int unpack_sha1_header(git_zstream *stream, unsigned char *map, unsigned long mapsize, void *buffer, unsigned long bufsiz);
  extern int parse_sha1_header(const char *hdr, unsigned long *sizep);
  
 -extern int check_sha1_signature(const unsigned char *sha1, void *buf, unsigned long size, const char *type);
 +extern int check_object_signature(const struct object_id *oid, void *buf, unsigned long size, const char *type);
  
  extern int finalize_object_file(const char *tmpfile, const char *filename);
  
 -/*
 - * Open the loose object at path, check its sha1, and return the contents,
 - * type, and size. If the object is a blob, then "contents" may return NULL,
 - * to allow streaming of large blobs.
 - *
 - * Returns 0 on success, negative on error (details may be written to stderr).
 - */
 -int read_loose_object(const char *path,
 -                    const unsigned char *expected_sha1,
 -                    enum object_type *type,
 -                    unsigned long *size,
 -                    void **contents);
 -
 -/*
 - * Convenience for sha1_object_info_extended() with a NULL struct
 - * object_info. OBJECT_INFO_SKIP_CACHED is automatically set; pass
 - * nonzero flags to also set other flags.
 - */
 -extern int has_sha1_file_with_flags(const unsigned char *sha1, int flags);
 -static inline int has_sha1_file(const unsigned char *sha1)
 -{
 -      return has_sha1_file_with_flags(sha1, 0);
 -}
 -
 -/* Same as the above, except for struct object_id. */
 -extern int has_object_file(const struct object_id *oid);
 -extern int has_object_file_with_flags(const struct object_id *oid, int flags);
 -
 -/*
 - * Return true iff an alternate object database has a loose object
 - * with the specified name.  This function does not respect replace
 - * references.
 - */
 -extern int has_loose_object_nonlocal(const unsigned char *sha1);
 -
 -extern void assert_sha1_type(const unsigned char *sha1, enum object_type expect);
 -
  /* Helper to check and "touch" a file */
  extern int check_and_freshen_file(const char *fn, int freshen);
  
@@@ -1288,6 -1320,7 +1292,6 @@@ static inline int hex2chr(const char *s
  #define FALLBACK_DEFAULT_ABBREV 7
  
  struct object_context {
 -      unsigned char tree[20];
        unsigned mode;
        /*
         * symlink_path is only used by get_tree_entry_follow_symlinks,
@@@ -1414,10 -1447,10 +1418,10 @@@ extern int df_name_compare(const char *
  extern int name_compare(const char *name1, size_t len1, const char *name2, size_t len2);
  extern int cache_name_stage_compare(const char *name1, int len1, int stage1, const char *name2, int len2, int stage2);
  
 -extern void *read_object_with_reference(const unsigned char *sha1,
 +extern void *read_object_with_reference(const struct object_id *oid,
                                        const char *required_type,
                                        unsigned long *size,
 -                                      unsigned char *sha1_ret);
 +                                      struct object_id *oid_ret);
  
  extern struct object *peel_to_type(const char *name, int namelen,
                                   struct object *o, enum object_type);
@@@ -1543,6 -1576,57 +1547,6 @@@ extern int has_dirs_only_path(const cha
  extern void schedule_dir_for_removal(const char *name, int len);
  extern void remove_scheduled_dirs(void);
  
 -extern struct alternate_object_database {
 -      struct alternate_object_database *next;
 -
 -      /* see alt_scratch_buf() */
 -      struct strbuf scratch;
 -      size_t base_len;
 -
 -      /*
 -       * Used to store the results of readdir(3) calls when searching
 -       * for unique abbreviated hashes.  This cache is never
 -       * invalidated, thus it's racy and not necessarily accurate.
 -       * That's fine for its purpose; don't use it for tasks requiring
 -       * greater accuracy!
 -       */
 -      char loose_objects_subdir_seen[256];
 -      struct oid_array loose_objects_cache;
 -
 -      char path[FLEX_ARRAY];
 -} *alt_odb_list;
 -extern void prepare_alt_odb(void);
 -extern char *compute_alternate_path(const char *path, struct strbuf *err);
 -typedef int alt_odb_fn(struct alternate_object_database *, void *);
 -extern int foreach_alt_odb(alt_odb_fn, void*);
 -
 -/*
 - * Allocate a "struct alternate_object_database" but do _not_ actually
 - * add it to the list of alternates.
 - */
 -struct alternate_object_database *alloc_alt_odb(const char *dir);
 -
 -/*
 - * Add the directory to the on-disk alternates file; the new entry will also
 - * take effect in the current process.
 - */
 -extern void add_to_alternates_file(const char *dir);
 -
 -/*
 - * Add the directory to the in-memory list of alternates (along with any
 - * recursive alternates it points to), but do not modify the on-disk alternates
 - * file.
 - */
 -extern void add_to_alternates_memory(const char *dir);
 -
 -/*
 - * Returns a scratch strbuf pre-filled with the alternate object directory,
 - * including a trailing slash, which can be used to access paths in the
 - * alternate. Always use this over direct access to alt->scratch, as it
 - * cleans up any previous use of the scratch buffer.
 - */
 -extern struct strbuf *alt_scratch_buf(struct alternate_object_database *alt);
 -
  struct pack_window {
        struct pack_window *next;
        unsigned char *base;
        unsigned int inuse_cnt;
  };
  
 -extern struct packed_git {
 -      struct packed_git *next;
 -      struct list_head mru;
 -      struct pack_window *windows;
 -      off_t pack_size;
 -      const void *index_data;
 -      size_t index_size;
 -      uint32_t num_objects;
 -      uint32_t num_bad_objects;
 -      unsigned char *bad_object_sha1;
 -      int index_version;
 -      time_t mtime;
 -      int pack_fd;
 -      unsigned pack_local:1,
 -               pack_keep:1,
 -               freshened:1,
 -               do_not_close:1,
 -               pack_promisor:1;
 -      unsigned char sha1[20];
 -      struct revindex_entry *revindex;
 -      /* something like ".git/objects/pack/xxxxx.pack" */
 -      char pack_name[FLEX_ARRAY]; /* more */
 -} *packed_git;
 -
 -/*
 - * A most-recently-used ordered version of the packed_git list.
 - */
 -extern struct list_head packed_git_mru;
 -
  struct pack_entry {
        off_t offset;
 -      unsigned char sha1[20];
        struct packed_git *p;
  };
  
@@@ -1628,6 -1742,55 +1632,6 @@@ int for_each_loose_file_in_objdir_buf(s
  #define FOR_EACH_OBJECT_LOCAL_ONLY 0x1
  extern int for_each_loose_object(each_loose_object_fn, void *, unsigned flags);
  
 -struct object_info {
 -      /* Request */
 -      enum object_type *typep;
 -      unsigned long *sizep;
 -      off_t *disk_sizep;
 -      unsigned char *delta_base_sha1;
 -      struct strbuf *type_name;
 -      void **contentp;
 -
 -      /* Response */
 -      enum {
 -              OI_CACHED,
 -              OI_LOOSE,
 -              OI_PACKED,
 -              OI_DBCACHED
 -      } whence;
 -      union {
 -              /*
 -               * struct {
 -               *      ... Nothing to expose in this case
 -               * } cached;
 -               * struct {
 -               *      ... Nothing to expose in this case
 -               * } loose;
 -               */
 -              struct {
 -                      struct packed_git *pack;
 -                      off_t offset;
 -                      unsigned int is_delta;
 -              } packed;
 -      } u;
 -};
 -
 -/*
 - * Initializer for a "struct object_info" that wants no items. You may
 - * also memset() the memory to all-zeroes.
 - */
 -#define OBJECT_INFO_INIT {NULL}
 -
 -/* Invoke lookup_replace_object() on the given hash */
 -#define OBJECT_INFO_LOOKUP_REPLACE 1
 -/* Allow reading from a loose object file of unknown/bogus type */
 -#define OBJECT_INFO_ALLOW_UNKNOWN_TYPE 2
 -/* Do not check cached storage */
 -#define OBJECT_INFO_SKIP_CACHED 4
 -/* Do not retry packed storage after checking packed and loose storage */
 -#define OBJECT_INFO_QUICK 8
 -extern int sha1_object_info_extended(const unsigned char *, struct object_info *, unsigned flags);
 -
  /*
   * Set this to 0 to prevent sha1_object_info_extended() from fetching missing
   * blobs. This has a difference only if extensions.partialClone is set.
@@@ -1713,6 -1876,15 +1717,6 @@@ extern const char *excludes_file
  int decode_85(char *dst, const char *line, int linelen);
  void encode_85(char *buf, const unsigned char *data, int bytes);
  
 -/* alloc.c */
 -extern void *alloc_blob_node(void);
 -extern void *alloc_tree_node(void);
 -extern void *alloc_commit_node(void);
 -extern void *alloc_tag_node(void);
 -extern void *alloc_object_node(void);
 -extern void alloc_report(void);
 -extern unsigned int alloc_commit_index(void);
 -
  /* pkt-line.c */
  void packet_trace_identity(const char *prog);
  
@@@ -1760,6 -1932,11 +1764,6 @@@ extern int ws_blank_line(const char *li
  void overlay_tree_on_index(struct index_state *istate,
                           const char *tree_name, const char *prefix);
  
 -char *alias_lookup(const char *alias);
 -int split_cmdline(char *cmdline, const char ***argv);
 -/* Takes a negative value returned by split_cmdline */
 -const char *split_cmdline_strerror(int cmdline_errno);
 -
  /* setup.c */
  struct startup_info {
        int have_repository;
diff --combined merge-recursive.c
index 07d792fd10c6bcee9ca12367fc2ad708fb26f9c4,352aa50fba1408037a0cb3d40639d5687492f027..1446e92beaf5a85376684cef0104dcd226a7e539
@@@ -8,8 -8,6 +8,8 @@@
  #include "advice.h"
  #include "lockfile.h"
  #include "cache-tree.h"
 +#include "object-store.h"
 +#include "repository.h"
  #include "commit.h"
  #include "blob.h"
  #include "builtin.h"
@@@ -17,7 -15,6 +17,7 @@@
  #include "diff.h"
  #include "diffcore.h"
  #include "tag.h"
 +#include "alloc.h"
  #include "unpack-trees.h"
  #include "string-list.h"
  #include "xdiff-interface.h"
@@@ -26,7 -23,6 +26,7 @@@
  #include "merge-recursive.h"
  #include "dir.h"
  #include "submodule.h"
 +#include "revision.h"
  
  struct path_hashmap_entry {
        struct hashmap_entry e;
@@@ -53,67 -49,6 +53,67 @@@ 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) {
@@@ -158,15 -93,15 +158,15 @@@ static struct tree *shift_tree_object(s
        }
        if (!oidcmp(&two->object.oid, &shifted))
                return two;
 -      return lookup_tree(&shifted);
 +      return lookup_tree(the_repository, &shifted);
  }
  
  static struct commit *make_virtual_commit(struct tree *tree, const char *comment)
  {
 -      struct commit *commit = alloc_commit_node();
 +      struct commit *commit = alloc_commit_node(the_repository);
  
        set_merge_remote_desc(commit, comment, (struct object *)commit);
 -      commit->tree = tree;
 +      commit->maybe_tree = tree;
        commit->object.parsed = 1;
        return commit;
  }
@@@ -184,7 -119,6 +184,7 @@@ static int oid_eq(const struct object_i
  
  enum rename_type {
        RENAME_NORMAL = 0,
 +      RENAME_VIA_DIR,
        RENAME_DELETE,
        RENAME_ONE_FILE_TO_ONE,
        RENAME_ONE_FILE_TO_TWO,
@@@ -289,14 -223,12 +289,14 @@@ static void output(struct merge_option
  
  static void output_commit_title(struct merge_options *o, struct commit *commit)
  {
 +      struct merge_remote_desc *desc;
 +
        strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
 -      if (commit->util)
 -              strbuf_addf(&o->obuf, "virtual %s\n",
 -                      merge_remote_util(commit)->name);
 +      desc = merge_remote_util(commit);
 +      if (desc)
 +              strbuf_addf(&o->obuf, "virtual %s\n", desc->name);
        else {
 -              strbuf_add_unique_abbrev(&o->obuf, commit->object.oid.hash,
 +              strbuf_add_unique_abbrev(&o->obuf, &commit->object.oid,
                                         DEFAULT_ABBREV);
                strbuf_addch(&o->obuf, ' ');
                if (parse_commit(commit) != 0)
  }
  
  static int add_cacheinfo(struct merge_options *o,
 -              unsigned int mode, const struct object_id *oid,
 -              const char *path, int stage, int refresh, int options)
 +                       unsigned int mode, const struct object_id *oid,
 +                       const char *path, int stage, int refresh, int options)
  {
        struct cache_entry *ce;
        int ret;
  
 -      ce = make_cache_entry(mode, oid ? oid->hash : null_sha1, path, stage, 0);
 +      ce = make_cache_entry(&the_index, mode, oid ? oid : &null_oid, 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) {
                struct cache_entry *nce;
  
 -              nce = refresh_cache_entry(ce, CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
 +              nce = refresh_cache_entry(&the_index, 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);
        }
@@@ -343,54 -275,36 +343,54 @@@ 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)
                return NULL;
        }
  
 -      result = lookup_tree(&active_cache_tree->oid);
 +      result = lookup_tree(the_repository, &active_cache_tree->oid);
  
        return result;
  }
  
 -static int save_files_dirs(const unsigned char *sha1,
 -              struct strbuf *base, const char *path,
 -              unsigned int mode, int stage, void *context)
 +static int save_files_dirs(const struct object_id *oid,
 +                         struct strbuf *base, const char *path,
 +                         unsigned int mode, int stage, void *context)
  {
        struct path_hashmap_entry *entry;
        int baselen = base->len;
@@@ -446,21 -360,6 +446,21 @@@ 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.
@@@ -471,12 -370,12 +471,12 @@@ 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.hash, path,
 -                      e->stages[1].oid.hash, &e->stages[1].mode);
 -      get_tree_entry(a->object.oid.hash, path,
 -                      e->stages[2].oid.hash, &e->stages[2].mode);
 -      get_tree_entry(b->object.oid.hash, path,
 -                      e->stages[3].oid.hash, &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;
@@@ -544,7 -443,7 +544,7 @@@ static void record_df_conflict_files(st
                                     struct string_list *entries)
  {
        /* If there is a D/F conflict and the file for such a conflict
 -       * currently exist in the working tree, we want to allow it to be
 +       * currently exists in the working tree, we want to allow it to be
         * removed to make room for the corresponding directory if needed.
         * The files underneath the directories of such D/F conflicts will
         * be processed before the corresponding file involved in the D/F
@@@ -635,10 -534,78 +635,10 @@@ 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,
@@@ -792,92 -738,31 +792,92 @@@ 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)
         */
        if (would_lose_untracked(path))
                return err(o, _("refusing to lose untracked file at '%s'"),
 -                           path);
 +                         path);
  
        /* Successful unlink is good.. */
        if (!unlink(path))
@@@ -957,7 -842,7 +957,7 @@@ static int update_file_flags(struct mer
                        goto update_index;
                }
  
 -              buf = read_sha1_file(oid->hash, &type, &size);
 +              buf = read_object_file(oid, &type, &size);
                if (!buf)
                        return err(o, _("cannot read object %s '%s'"), oid_to_hex(oid), path);
                if (type != OBJ_BLOB) {
                        unlink(path);
                        if (symlink(lnk, path))
                                ret = err(o, _("failed to symlink '%s': %s"),
 -                                      path, strerror(errno));
 +                                        path, strerror(errno));
                        free(lnk);
                } else
                        ret = err(o,
                                  _("do not know what to do with %06o %s '%s'"),
                                  mode, oid_to_hex(oid), path);
 - free_buf:
 +      free_buf:
                free(buf);
        }
 - update_index:
 +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;
  }
  
@@@ -1094,193 -977,13 +1094,193 @@@ static int merge_3way(struct merge_opti
        return merge_status;
  }
  
 +static int find_first_merges(struct object_array *result, const char *path,
 +                           struct commit *a, struct commit *b)
 +{
 +      int i, j;
 +      struct object_array merges = OBJECT_ARRAY_INIT;
 +      struct commit *commit;
 +      int contains_another;
 +
 +      char merged_revision[42];
 +      const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
 +                                 "--all", merged_revision, NULL };
 +      struct rev_info revs;
 +      struct setup_revision_opt rev_opts;
 +
 +      memset(result, 0, sizeof(struct object_array));
 +      memset(&rev_opts, 0, sizeof(rev_opts));
 +
 +      /* get all revisions that merge commit a */
 +      xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
 +                oid_to_hex(&a->object.oid));
 +      init_revisions(&revs, NULL);
 +      rev_opts.submodule = path;
 +      /* FIXME: can't handle linked worktrees in submodules yet */
 +      revs.single_worktree = path != NULL;
 +      setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
 +
 +      /* save all revisions from the above list that contain b */
 +      if (prepare_revision_walk(&revs))
 +              die("revision walk setup failed");
 +      while ((commit = get_revision(&revs)) != NULL) {
 +              struct object *o = &(commit->object);
 +              if (in_merge_bases(b, commit))
 +                      add_object_array(o, NULL, &merges);
 +      }
 +      reset_revision_walk();
 +
 +      /* Now we've got all merges that contain a and b. Prune all
 +       * merges that contain another found merge and save them in
 +       * result.
 +       */
 +      for (i = 0; i < merges.nr; i++) {
 +              struct commit *m1 = (struct commit *) merges.objects[i].item;
 +
 +              contains_another = 0;
 +              for (j = 0; j < merges.nr; j++) {
 +                      struct commit *m2 = (struct commit *) merges.objects[j].item;
 +                      if (i != j && in_merge_bases(m2, m1)) {
 +                              contains_another = 1;
 +                              break;
 +                      }
 +              }
 +
 +              if (!contains_another)
 +                      add_object_array(merges.objects[i].item, NULL, result);
 +      }
 +
 +      object_array_clear(&merges);
 +      return result->nr;
 +}
 +
 +static void print_commit(struct commit *commit)
 +{
 +      struct strbuf sb = STRBUF_INIT;
 +      struct pretty_print_context ctx = {0};
 +      ctx.date_mode.type = DATE_NORMAL;
 +      format_commit_message(commit, " %h: %m %s", &sb, &ctx);
 +      fprintf(stderr, "%s\n", sb.buf);
 +      strbuf_release(&sb);
 +}
 +
 +static int merge_submodule(struct merge_options *o,
 +                         struct object_id *result, const char *path,
 +                         const struct object_id *base, const struct object_id *a,
 +                         const struct object_id *b)
 +{
 +      struct commit *commit_base, *commit_a, *commit_b;
 +      int parent_count;
 +      struct object_array merges;
 +
 +      int i;
 +      int search = !o->call_depth;
 +
 +      /* store a in result in case we fail */
 +      oidcpy(result, a);
 +
 +      /* we can not handle deletion conflicts */
 +      if (is_null_oid(base))
 +              return 0;
 +      if (is_null_oid(a))
 +              return 0;
 +      if (is_null_oid(b))
 +              return 0;
 +
 +      if (add_submodule_odb(path)) {
 +              output(o, 1, _("Failed to merge submodule %s (not checked out)"), path);
 +              return 0;
 +      }
 +
 +      if (!(commit_base = lookup_commit_reference(the_repository, base)) ||
 +          !(commit_a = lookup_commit_reference(the_repository, a)) ||
 +          !(commit_b = lookup_commit_reference(the_repository, b))) {
 +              output(o, 1, _("Failed to merge submodule %s (commits not present)"), path);
 +              return 0;
 +      }
 +
 +      /* check whether both changes are forward */
 +      if (!in_merge_bases(commit_base, commit_a) ||
 +          !in_merge_bases(commit_base, commit_b)) {
 +              output(o, 1, _("Failed to merge submodule %s (commits don't follow merge-base)"), path);
 +              return 0;
 +      }
 +
 +      /* Case #1: a is contained in b or vice versa */
 +      if (in_merge_bases(commit_a, commit_b)) {
 +              oidcpy(result, b);
 +              if (show(o, 3)) {
 +                      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"), path);
 +              else
 +                      ; /* no output */
 +
 +              return 1;
 +      }
 +      if (in_merge_bases(commit_b, commit_a)) {
 +              oidcpy(result, a);
 +              if (show(o, 3)) {
 +                      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"), path);
 +              else
 +                      ; /* no output */
 +
 +              return 1;
 +      }
 +
 +      /*
 +       * Case #2: There are one or more merges that contain a and b in
 +       * the submodule. If there is only one, then present it as a
 +       * suggestion to the user, but leave it marked unmerged so the
 +       * user needs to confirm the resolution.
 +       */
 +
 +      /* Skip the search if makes no sense to the calling context.  */
 +      if (!search)
 +              return 0;
 +
 +      /* find commit which merges them */
 +      parent_count = find_first_merges(&merges, path, commit_a, commit_b);
 +      switch (parent_count) {
 +      case 0:
 +              output(o, 1, _("Failed to merge submodule %s (merge following commits not found)"), path);
 +              break;
 +
 +      case 1:
 +              output(o, 1, _("Failed to merge submodule %s (not fast-forward)"), path);
 +              output(o, 2, _("Found a possible merge resolution for the submodule:\n"));
 +              print_commit((struct commit *) merges.objects[0].item);
 +              output(o, 2, _(
 +                     "If this is correct simply add it to the index "
 +                     "for example\n"
 +                     "by using:\n\n"
 +                     "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
 +                     "which will accept this suggestion.\n"),
 +                     oid_to_hex(&merges.objects[0].item->oid), path);
 +              break;
 +
 +      default:
 +              output(o, 1, _("Failed to merge submodule %s (multiple merges found)"), path);
 +              for (i = 0; i < merges.nr; i++)
 +                      print_commit((struct commit *) merges.objects[i].item);
 +      }
 +
 +      object_array_clear(&merges);
 +      return 0;
 +}
 +
  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;
                                return ret;
                        result->clean = (merge_status == 0);
                } else if (S_ISGITLINK(a->mode)) {
 -                      result->clean = merge_submodule(&result->oid,
 -                                                     one->path,
 -                                                     &one->oid,
 -                                                     &a->oid,
 -                                                     &b->oid,
 -                                                     !o->call_depth);
 +                      result->clean = merge_submodule(o, &result->oid,
 +                                                      one->path,
 +                                                      &one->oid,
 +                                                      &a->oid,
 +                                                      &b->oid);
                } else if (S_ISLNK(a->mode)) {
                        switch (o->recursive_variant) {
                        case MERGE_RECURSIVE_NORMAL:
                                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 handle_rename_via_dir(struct merge_options *o,
 +                               struct diff_filepair *pair,
 +                               const char *rename_branch,
 +                               const char *other_branch)
 +{
 +      /*
 +       * Handle file adds that need to be renamed due to directory rename
 +       * detection.  This differs from handle_rename_normal, because
 +       * there is no content merge to do; just move the file into the
 +       * desired final location.
 +       */
 +      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 *path, const char *old_path,
 -                               const struct object_id *o_oid, int o_mode,
 -                               const struct object_id *changed_oid,
 -                               int changed_mode,
 -                               const char *change_branch,
 -                               const char *delete_branch,
 -                               const char *change, const char *change_past)
 +                              const char *path, const char *old_path,
 +                              const struct object_id *o_oid, int o_mode,
 +                              const struct object_id *changed_oid,
 +                              int changed_mode,
 +                              const char *change_branch,
 +                              const char *delete_branch,
 +                              const char *change, const char *change_past)
  {
        char *alt_path = NULL;
        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);
        }
  
                if (!ret)
                        ret = update_file(o, 0, o_oid, o_mode, update_path);
        } else {
 +              /*
 +               * Despite the four nearly duplicate messages and argument
 +               * lists below and the ugliness of the nested if-statements,
 +               * having complete messages makes the job easier for
 +               * translators.
 +               *
 +               * The slight variance among the cases is due to the fact
 +               * that:
 +               *   1) directory/file conflicts (in effect if
 +               *      !alt_path) could cause us to need to write the
 +               *      file to a different path.
 +               *   2) renames (in effect if !old_path) could mean that
 +               *      there are two names for the path that the user
 +               *      may know the file by.
 +               */
                if (!alt_path) {
                        if (!old_path) {
                                output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
        return ret;
  }
  
 -static int conflict_rename_delete(struct merge_options *o,
 -                                 struct diff_filepair *pair,
 -                                 const char *rename_branch,
 -                                 const char *delete_branch)
 +static int handle_rename_delete(struct merge_options *o,
 +                              struct diff_filepair *pair,
 +                              const char *rename_branch,
 +                              const char *delete_branch)
  {
        const struct diff_filespec *orig = pair->one;
        const struct diff_filespec *dest = pair->two;
@@@ -1597,34 -1242,17 +1597,34 @@@ 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)))
        return ret;
  }
  
 -static int conflict_rename_rename_1to2(struct merge_options *o,
 -                                      struct rename_conflict_info *ci)
 +static int handle_rename_rename_1to2(struct merge_options *o,
 +                                   struct rename_conflict_info *ci)
  {
        /* One file was renamed in both branches, but to different names. */
        struct diff_filespec *one = ci->pair1->one;
        return 0;
  }
  
 -static int conflict_rename_rename_2to1(struct merge_options *o,
 -                                      struct rename_conflict_info *ci)
 +static int handle_rename_rename_2to1(struct merge_options *o,
 +                                   struct rename_conflict_info *ci)
  {
        /* Two files, a & b, were renamed to the same thing, c. */
        struct diff_filespec *a = ci->pair1->one;
        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;
  }
  
 -static int process_renames(struct merge_options *o,
 -                         struct string_list *a_renames,
 -                         struct string_list *b_renames)
 +/*
 + * 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)
  {
 -      int clean_merge = 1, i, j;
 -      struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
 -      struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
 -      const struct rename *sre;
 -
 -      for (i = 0; i < a_renames->nr; i++) {
 -              sre = a_renames->items[i].util;
 -              string_list_insert(&a_by_dst, sre->pair->two->path)->util
 -                      = (void *)sre;
 -      }
 -      for (i = 0; i < b_renames->nr; i++) {
 -              sre = b_renames->items[i].util;
 -              string_list_insert(&b_by_dst, sre->pair->two->path)->util
 -                      = (void *)sre;
 -      }
 -
 -      for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
 -              struct string_list *renames1, *renames2Dst;
 -              struct rename *ren1 = NULL, *ren2 = NULL;
 -              const char *branch1, *branch2;
 -              const char *ren1_src, *ren1_dst;
 -              struct string_list_item *lookup;
 +      struct diff_queue_struct *ret;
 +      struct diff_options opts;
  
 -              if (i >= a_renames->nr) {
 -                      ren2 = b_renames->items[j++].util;
 +      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 = xstrdup(path);
 +      char *end;
 +      struct dir_rename_entry *entry = NULL;;
 +
 +      while ((end = strrchr(temp, '/'))) {
 +              *end = '\0';
 +              entry = dir_rename_find_entry(dir_renames, temp);
 +              if (entry)
 +                      break;
 +      }
 +      free(temp);
 +      return entry;
 +}
 +
 +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 handle_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)
 +{
 +      int clean_merge = 1, i, j;
 +      struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
 +      struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
 +      const struct rename *sre;
 +
 +      for (i = 0; i < a_renames->nr; i++) {
 +              sre = a_renames->items[i].util;
 +              string_list_insert(&a_by_dst, sre->pair->two->path)->util
 +                      = (void *)sre;
 +      }
 +      for (i = 0; i < b_renames->nr; i++) {
 +              sre = b_renames->items[i].util;
 +              string_list_insert(&b_by_dst, sre->pair->two->path)->util
 +                      = (void *)sre;
 +      }
 +
 +      for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
 +              struct string_list *renames1, *renames2Dst;
 +              struct rename *ren1 = NULL, *ren2 = NULL;
 +              const char *branch1, *branch2;
 +              const char *ren1_src, *ren1_dst;
 +              struct string_list_item *lookup;
 +
 +              if (i >= a_renames->nr) {
 +                      ren2 = b_renames->items[j++].util;
                } else if (j >= b_renames->nr) {
                        ren1 = a_renames->items[i++].util;
                } else {
                        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_VIA_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,
@@@ -2825,118 -1645,18 +2825,118 @@@ 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 detect_and_process_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;
  }
  
  static int read_oid_strbuf(struct merge_options *o,
 -      const struct object_id *oid, struct strbuf *dst)
 +                         const struct object_id *oid,
 +                         struct strbuf *dst)
  {
        void *buf;
        enum object_type type;
        unsigned long size;
 -      buf = read_sha1_file(oid->hash, &type, &size);
 +      buf = read_object_file(oid, &type, &size);
        if (!buf)
                return err(o, _("cannot read object %s"), oid_to_hex(oid));
        if (type != OBJ_BLOB) {
@@@ -2984,10 -1704,10 +2984,10 @@@ error_return
  }
  
  static int handle_modify_delete(struct merge_options *o,
 -                               const char *path,
 -                               struct object_id *o_oid, int o_mode,
 -                               struct object_id *a_oid, int a_mode,
 -                               struct object_id *b_oid, int b_mode)
 +                              const char *path,
 +                              struct object_id *o_oid, int o_mode,
 +                              struct object_id *a_oid, int a_mode,
 +                              struct object_id *b_oid, int b_mode)
  {
        const char *modify_branch, *delete_branch;
        struct object_id *changed_oid;
  
  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 handle_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 */
@@@ -3157,37 -1862,26 +3157,37 @@@ 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 = handle_rename_normal(o,
 +                                                         path,
 +                                                         o_oid, o_mode,
 +                                                         a_oid, a_mode,
 +                                                         b_oid, b_mode,
 +                                                         conflict_info);
 +                      break;
 +              case RENAME_VIA_DIR:
 +                      clean_merge = 1;
 +                      if (handle_rename_via_dir(o,
 +                                                conflict_info->pair1,
 +                                                conflict_info->branch1,
 +                                                conflict_info->branch2))
 +                              clean_merge = -1;
                        break;
                case RENAME_DELETE:
                        clean_merge = 0;
 -                      if (conflict_rename_delete(o,
 -                                                 conflict_info->pair1,
 -                                                 conflict_info->branch1,
 -                                                 conflict_info->branch2))
 +                      if (handle_rename_delete(o,
 +                                               conflict_info->pair1,
 +                                               conflict_info->branch1,
 +                                               conflict_info->branch2))
                                clean_merge = -1;
                        break;
                case RENAME_ONE_FILE_TO_TWO:
                        clean_merge = 0;
 -                      if (conflict_rename_rename_1to2(o, conflict_info))
 +                      if (handle_rename_rename_1to2(o, conflict_info))
                                clean_merge = -1;
                        break;
                case RENAME_TWO_FILES_TO_ONE:
                        clean_merge = 0;
 -                      if (conflict_rename_rename_2to1(o, conflict_info))
 +                      if (handle_rename_rename_2to1(o, conflict_info))
                                clean_merge = -1;
                        break;
                default:
        } 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;
  }
@@@ -3281,6 -1974,13 +3281,13 @@@ int merge_trees(struct merge_options *o
                struct tree **result)
  {
        int code, clean;
+       struct strbuf sb = STRBUF_INIT;
+       if (!o->call_depth && index_has_changes(&the_index, head, &sb)) {
+               err(o, _("Your local changes to the following files would be overwritten by merge:\n  %s"),
+                   sb.buf);
+               return -1;
+       }
  
        if (o->subtree_shift) {
                merge = shift_tree_object(head, merge, o->subtree_shift);
        }
  
        if (oid_eq(&common->object.oid, &merge->object.oid)) {
-               struct strbuf sb = STRBUF_INIT;
-               if (!o->call_depth && index_has_changes(&sb)) {
-                       err(o, _("Dirty index: cannot merge (dirty: %s)"),
-                           sb.buf);
-                       return 0;
-               }
                output(o, 0, _("Already up to date!"));
                *result = head;
                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 = detect_and_process_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);
 +      cleanup:
 +              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;
  
@@@ -3427,7 -2118,7 +3427,7 @@@ int merge_recursive(struct merge_option
                /* if there is no common ancestor, use an empty tree */
                struct tree *tree;
  
 -              tree = lookup_tree(the_hash_algo->empty_tree);
 +              tree = lookup_tree(the_repository, the_repository->hash_algo->empty_tree);
                merged_common_ancestors = make_virtual_commit(tree, "ancestor");
        }
  
                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);
@@@ -3489,9 -2179,7 +3489,9 @@@ static struct commit *get_ref(const str
  {
        struct object *object;
  
 -      object = deref_tag(parse_object(oid), name, strlen(name));
 +      object = deref_tag(the_repository, parse_object(the_repository, oid),
 +                         name,
 +                         strlen(name));
        if (!object)
                return NULL;
        if (object->type == OBJ_TREE)
@@@ -3522,14 -2210,14 +3522,14 @@@ int merge_recursive_generic(struct merg
                        struct commit *base;
                        if (!(base = get_ref(base_list[i], oid_to_hex(base_list[i]))))
                                return err(o, _("Could not parse object '%s'"),
 -                                      oid_to_hex(base_list[i]));
 +                                         oid_to_hex(base_list[i]));
                        commit_list_insert(base, &ca);
                }
        }
  
        hold_locked_index(&lock, LOCK_DIE_ON_ERROR);
        clean = merge_recursive(o, head_commit, next_commit, ca,
 -                      result);
 +                              result);
        if (clean < 0) {
                rollback_lock_file(&lock);
                return clean;
  
  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);
  }
  
@@@ -3568,8 -2247,7 +3568,8 @@@ 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)
@@@ -3620,16 -2298,16 +3620,16 @@@ 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;
diff --combined merge.c
index 0783858739f84028df6eef85d3673c944fabb912,fb5eaf24622f2f0b6c8dfd1bdc6fb051e31007d5..e30e03fb84a7677520d208a5688bb7170c22bf8e
+++ b/merge.c
  
  static const char *merge_argument(struct commit *commit)
  {
 -      if (commit)
 -              return oid_to_hex(&commit->object.oid);
 -      else
 -              return EMPTY_TREE_SHA1_HEX;
 +      return oid_to_hex(commit ? &commit->object.oid : the_hash_algo->empty_tree);
  }
  
- int index_has_changes(struct strbuf *sb)
- {
-       struct object_id head;
-       int i;
-       if (!get_oid_tree("HEAD", &head)) {
-               struct diff_options opt;
-               diff_setup(&opt);
-               opt.flags.exit_with_status = 1;
-               if (!sb)
-                       opt.flags.quick = 1;
-               do_diff_cache(&head, &opt);
-               diffcore_std(&opt);
-               for (i = 0; sb && i < diff_queued_diff.nr; i++) {
-                       if (i)
-                               strbuf_addch(sb, ' ');
-                       strbuf_addstr(sb, diff_queued_diff.queue[i]->two->path);
-               }
-               diff_flush(&opt);
-               return opt.flags.has_changes != 0;
-       } else {
-               for (i = 0; sb && i < active_nr; i++) {
-                       if (i)
-                               strbuf_addch(sb, ' ');
-                       strbuf_addstr(sb, active_cache[i]->name);
-               }
-               return !!active_nr;
-       }
- }
  int try_merge_command(const char *strategy, size_t xopts_nr,
                      const char **xopts, struct commit_list *common,
                      const char *head_arg, struct commit_list *remotes)
@@@ -91,24 -63,8 +60,24 @@@ int checkout_fast_forward(const struct 
                return -1;
  
        memset(&trees, 0, sizeof(trees));
 -      memset(&opts, 0, sizeof(opts));
        memset(&t, 0, sizeof(t));
 +
 +      trees[nr_trees] = parse_tree_indirect(head);
 +      if (!trees[nr_trees++]) {
 +              rollback_lock_file(&lock_file);
 +              return -1;
 +      }
 +      trees[nr_trees] = parse_tree_indirect(remote);
 +      if (!trees[nr_trees++]) {
 +              rollback_lock_file(&lock_file);
 +              return -1;
 +      }
 +      for (i = 0; i < nr_trees; i++) {
 +              parse_tree(trees[i]);
 +              init_tree_desc(t+i, trees[i]->buffer, trees[i]->size);
 +      }
 +
 +      memset(&opts, 0, sizeof(opts));
        if (overwrite_ignore) {
                memset(&dir, 0, sizeof(dir));
                dir.flags |= DIR_SHOW_IGNORED;
        opts.fn = twoway_merge;
        setup_unpack_trees_porcelain(&opts, "merge");
  
 -      trees[nr_trees] = parse_tree_indirect(head);
 -      if (!trees[nr_trees++]) {
 -              rollback_lock_file(&lock_file);
 -              return -1;
 -      }
 -      trees[nr_trees] = parse_tree_indirect(remote);
 -      if (!trees[nr_trees++]) {
 -              rollback_lock_file(&lock_file);
 -              return -1;
 -      }
 -      for (i = 0; i < nr_trees; i++) {
 -              parse_tree(trees[i]);
 -              init_tree_desc(t+i, trees[i]->buffer, trees[i]->size);
 -      }
        if (unpack_trees(nr_trees, t, &opts)) {
                rollback_lock_file(&lock_file);
 +              clear_unpack_trees_porcelain(&opts);
                return -1;
        }
 +      clear_unpack_trees_porcelain(&opts);
 +
        if (write_locked_index(&the_index, &lock_file, COMMIT_LOCK))
                return error(_("unable to write new index file"));
        return 0;
diff --combined read-cache.c
index 56eac508362f272e5d777c98b7f31c0f043c1d89,639c1fffa0635e8829ad76d74407b06447e42b1f..880849fc8ad8454a756ac0849dff288ad8f38eaa
@@@ -6,12 -6,13 +6,14 @@@
  #define NO_THE_INDEX_COMPATIBILITY_MACROS
  #include "cache.h"
  #include "config.h"
+ #include "diff.h"
+ #include "diffcore.h"
  #include "tempfile.h"
  #include "lockfile.h"
  #include "cache-tree.h"
  #include "refs.h"
  #include "dir.h"
 +#include "object-store.h"
  #include "tree.h"
  #include "commit.h"
  #include "blob.h"
                 CE_ENTRY_ADDED | CE_ENTRY_REMOVED | CE_ENTRY_CHANGED | \
                 SPLIT_INDEX_ORDERED | UNTRACKED_CHANGED | FSMONITOR_CHANGED)
  
 +
 +/*
 + * This is an estimate of the pathname length in the index.  We use
 + * this for V4 index files to guess the un-deltafied size of the index
 + * in memory because of pathname deltafication.  This is not required
 + * for V2/V3 index formats because their pathnames are not compressed.
 + * If the initial amount of memory set aside is not sufficient, the
 + * mem pool will allocate extra memory.
 + */
 +#define CACHE_ENTRY_PATH_LENGTH 80
 +
 +static inline struct cache_entry *mem_pool__ce_alloc(struct mem_pool *mem_pool, size_t len)
 +{
 +      struct cache_entry *ce;
 +      ce = mem_pool_alloc(mem_pool, cache_entry_size(len));
 +      ce->mem_pool_allocated = 1;
 +      return ce;
 +}
 +
 +static inline struct cache_entry *mem_pool__ce_calloc(struct mem_pool *mem_pool, size_t len)
 +{
 +      struct cache_entry * ce;
 +      ce = mem_pool_calloc(mem_pool, 1, cache_entry_size(len));
 +      ce->mem_pool_allocated = 1;
 +      return ce;
 +}
 +
 +static struct mem_pool *find_mem_pool(struct index_state *istate)
 +{
 +      struct mem_pool **pool_ptr;
 +
 +      if (istate->split_index && istate->split_index->base)
 +              pool_ptr = &istate->split_index->base->ce_mem_pool;
 +      else
 +              pool_ptr = &istate->ce_mem_pool;
 +
 +      if (!*pool_ptr)
 +              mem_pool_init(pool_ptr, 0);
 +
 +      return *pool_ptr;
 +}
 +
  struct index_state the_index;
  static const char *alternate_index_output;
  
@@@ -104,7 -63,7 +106,7 @@@ static void replace_index_entry(struct 
  
        replace_index_entry_in_base(istate, old, ce);
        remove_name_hash(istate, old);
 -      free(old);
 +      discard_cache_entry(old);
        ce->ce_flags &= ~CE_HASHED;
        set_index_entry(istate, nr, ce);
        ce->ce_flags |= CE_UPDATE_IN_BASE;
@@@ -117,7 -76,7 +119,7 @@@ void rename_index_entry_at(struct index
        struct cache_entry *old_entry = istate->cache[nr], *new_entry;
        int namelen = strlen(new_name);
  
 -      new_entry = xmalloc(cache_entry_size(namelen));
 +      new_entry = make_empty_cache_entry(istate, namelen);
        copy_cache_entry(new_entry, old_entry);
        new_entry->ce_flags &= ~CE_HASHED;
        new_entry->ce_namelen = namelen;
@@@ -228,7 -187,7 +230,7 @@@ static int ce_compare_link(const struc
        if (strbuf_readlink(&sb, ce->name, expected_size))
                return -1;
  
 -      buffer = read_sha1_file(ce->oid.hash, &type, &size);
 +      buffer = read_object_file(&ce->oid, &type, &size);
        if (buffer) {
                if (size == sb.len)
                        match = memcmp(buffer, sb.buf, size);
@@@ -666,7 -625,7 +668,7 @@@ static struct cache_entry *create_alias
  
        /* Ok, create the new entry using the name of the existing alias */
        len = ce_namelen(alias);
 -      new_entry = xcalloc(1, cache_entry_size(len));
 +      new_entry = make_empty_cache_entry(istate, len);
        memcpy(new_entry->name, alias->name, len);
        copy_cache_entry(new_entry, ce);
        save_or_free_index_entry(istate, ce);
@@@ -683,7 -642,7 +685,7 @@@ void set_object_name_for_intent_to_add_
  
  int add_to_index(struct index_state *istate, const char *path, struct stat *st, int flags)
  {
 -      int size, namelen, was_same;
 +      int namelen, was_same;
        mode_t st_mode = st->st_mode;
        struct cache_entry *ce, *alias = NULL;
        unsigned ce_option = CE_MATCH_IGNORE_VALID|CE_MATCH_IGNORE_SKIP_WORKTREE|CE_MATCH_RACY_IS_DIRTY;
                while (namelen && path[namelen-1] == '/')
                        namelen--;
        }
 -      size = cache_entry_size(namelen);
 -      ce = xcalloc(1, size);
 +      ce = make_empty_cache_entry(istate, namelen);
        memcpy(ce->name, path, namelen);
        ce->ce_namelen = namelen;
        if (!intent_only)
                                ce_mark_uptodate(alias);
                        alias->ce_flags |= CE_ADDED;
  
 -                      free(ce);
 +                      discard_cache_entry(ce);
                        return 0;
                }
        }
        if (!intent_only) {
                if (index_path(&ce->oid, path, st, newflags)) {
 -                      free(ce);
 +                      discard_cache_entry(ce);
                        return error("unable to index file %s", path);
                }
        } else
                    ce->ce_mode == alias->ce_mode);
  
        if (pretend)
 -              free(ce);
 +              discard_cache_entry(ce);
        else if (add_index_entry(istate, ce, add_option)) {
 -              free(ce);
 +              discard_cache_entry(ce);
                return error("unable to add %s to index", path);
        }
        if (verbose && !was_same)
@@@ -787,25 -747,12 +789,25 @@@ int add_file_to_index(struct index_stat
        return add_to_index(istate, path, &st, flags);
  }
  
 -struct cache_entry *make_cache_entry(unsigned int mode,
 -              const unsigned char *sha1, const char *path, int stage,
 -              unsigned int refresh_options)
 +struct cache_entry *make_empty_cache_entry(struct index_state *istate, size_t len)
 +{
 +      return mem_pool__ce_calloc(find_mem_pool(istate), len);
 +}
 +
 +struct cache_entry *make_empty_transient_cache_entry(size_t len)
 +{
 +      return xcalloc(1, cache_entry_size(len));
 +}
 +
 +struct cache_entry *make_cache_entry(struct index_state *istate,
 +                                   unsigned int mode,
 +                                   const struct object_id *oid,
 +                                   const char *path,
 +                                   int stage,
 +                                   unsigned int refresh_options)
  {
 -      int size, len;
        struct cache_entry *ce, *ret;
 +      int len;
  
        if (!verify_path(path, mode)) {
                error("Invalid path '%s'", path);
        }
  
        len = strlen(path);
 -      size = cache_entry_size(len);
 -      ce = xcalloc(1, size);
 +      ce = make_empty_cache_entry(istate, len);
  
 -      hashcpy(ce->oid.hash, sha1);
 +      oidcpy(&ce->oid, oid);
        memcpy(ce->name, path, len);
        ce->ce_flags = create_ce_flags(stage);
        ce->ce_namelen = len;
        ce->ce_mode = create_ce_mode(mode);
  
 -      ret = refresh_cache_entry(ce, refresh_options);
 +      ret = refresh_cache_entry(&the_index, ce, refresh_options);
        if (ret != ce)
 -              free(ce);
 +              discard_cache_entry(ce);
        return ret;
  }
  
 +struct cache_entry *make_transient_cache_entry(unsigned int mode, const struct object_id *oid,
 +                                             const char *path, int stage)
 +{
 +      struct cache_entry *ce;
 +      int len;
 +
 +      if (!verify_path(path, mode)) {
 +              error("Invalid path '%s'", path);
 +              return NULL;
 +      }
 +
 +      len = strlen(path);
 +      ce = make_empty_transient_cache_entry(len);
 +
 +      oidcpy(&ce->oid, oid);
 +      memcpy(ce->name, path, len);
 +      ce->ce_flags = create_ce_flags(stage);
 +      ce->ce_namelen = len;
 +      ce->ce_mode = create_ce_mode(mode);
 +
 +      return ce;
 +}
 +
  /*
   * Chmod an index entry with either +x or -x.
   *
@@@ -1345,7 -1270,7 +1347,7 @@@ static struct cache_entry *refresh_cach
  {
        struct stat st;
        struct cache_entry *updated;
 -      int changed, size;
 +      int changed;
        int refresh = options & CE_MATCH_REFRESH;
        int ignore_valid = options & CE_MATCH_IGNORE_VALID;
        int ignore_skip_worktree = options & CE_MATCH_IGNORE_SKIP_WORKTREE;
                return NULL;
        }
  
 -      size = ce_size(ce);
 -      updated = xmalloc(size);
 +      updated = make_empty_cache_entry(istate, ce_namelen(ce));
        copy_cache_entry(updated, ce);
        memcpy(updated->name, ce->name, ce->ce_namelen + 1);
        fill_stat_cache_info(updated, &st);
@@@ -1549,11 -1475,10 +1551,11 @@@ int refresh_index(struct index_state *i
        return has_errors;
  }
  
 -struct cache_entry *refresh_cache_entry(struct cache_entry *ce,
 -                                             unsigned int options)
 +struct cache_entry *refresh_cache_entry(struct index_state *istate,
 +                                      struct cache_entry *ce,
 +                                      unsigned int options)
  {
 -      return refresh_cache_ent(&the_index, ce, options, NULL, NULL);
 +      return refresh_cache_ent(istate, ce, options, NULL, NULL);
  }
  
  
@@@ -1711,13 -1636,12 +1713,13 @@@ int read_index(struct index_state *ista
        return read_index_from(istate, get_index_file(), get_git_dir());
  }
  
 -static struct cache_entry *cache_entry_from_ondisk(struct ondisk_cache_entry *ondisk,
 +static struct cache_entry *cache_entry_from_ondisk(struct mem_pool *mem_pool,
 +                                                 struct ondisk_cache_entry *ondisk,
                                                   unsigned int flags,
                                                   const char *name,
                                                   size_t len)
  {
 -      struct cache_entry *ce = xmalloc(cache_entry_size(len));
 +      struct cache_entry *ce = mem_pool__ce_alloc(mem_pool, len);
  
        ce->ce_stat_data.sd_ctime.sec = get_be32(&ondisk->ctime.sec);
        ce->ce_stat_data.sd_mtime.sec = get_be32(&ondisk->mtime.sec);
@@@ -1759,8 -1683,7 +1761,8 @@@ static unsigned long expand_name_field(
        return (const char *)ep + 1 - cp_;
  }
  
 -static struct cache_entry *create_from_disk(struct ondisk_cache_entry *ondisk,
 +static struct cache_entry *create_from_disk(struct mem_pool *mem_pool,
 +                                          struct ondisk_cache_entry *ondisk,
                                            unsigned long *ent_size,
                                            struct strbuf *previous_name)
  {
                /* v3 and earlier */
                if (len == CE_NAMEMASK)
                        len = strlen(name);
 -              ce = cache_entry_from_ondisk(ondisk, flags, name, len);
 +              ce = cache_entry_from_ondisk(mem_pool, ondisk, flags, name, len);
  
                *ent_size = ondisk_ce_size(ce);
        } else {
                unsigned long consumed;
                consumed = expand_name_field(previous_name, name);
 -              ce = cache_entry_from_ondisk(ondisk, flags,
 +              ce = cache_entry_from_ondisk(mem_pool, ondisk, flags,
                                             previous_name->buf,
                                             previous_name->len);
  
@@@ -1871,22 -1794,6 +1873,22 @@@ static void post_read_index_from(struc
        tweak_fsmonitor(istate);
  }
  
 +static size_t estimate_cache_size_from_compressed(unsigned int entries)
 +{
 +      return entries * (sizeof(struct cache_entry) + CACHE_ENTRY_PATH_LENGTH);
 +}
 +
 +static size_t estimate_cache_size(size_t ondisk_size, unsigned int entries)
 +{
 +      long per_entry = sizeof(struct cache_entry) - sizeof(struct ondisk_cache_entry);
 +
 +      /*
 +       * Account for potential alignment differences.
 +       */
 +      per_entry += align_padding_size(sizeof(struct cache_entry), -sizeof(struct ondisk_cache_entry));
 +      return ondisk_size + entries * per_entry;
 +}
 +
  /* remember to discard_cache() before reading a different cache! */
  int do_read_index(struct index_state *istate, const char *path, int must_exist)
  {
        if (verify_hdr(hdr, mmap_size) < 0)
                goto unmap;
  
 -      hashcpy(istate->sha1, (const unsigned char *)hdr + mmap_size - the_hash_algo->rawsz);
 +      hashcpy(istate->oid.hash, (const unsigned char *)hdr + mmap_size - the_hash_algo->rawsz);
        istate->version = ntohl(hdr->hdr_version);
        istate->cache_nr = ntohl(hdr->hdr_entries);
        istate->cache_alloc = alloc_nr(istate->cache_nr);
        istate->cache = xcalloc(istate->cache_alloc, sizeof(*istate->cache));
        istate->initialized = 1;
  
 -      if (istate->version == 4)
 +      if (istate->version == 4) {
                previous_name = &previous_name_buf;
 -      else
 +              mem_pool_init(&istate->ce_mem_pool,
 +                            estimate_cache_size_from_compressed(istate->cache_nr));
 +      } else {
                previous_name = NULL;
 +              mem_pool_init(&istate->ce_mem_pool,
 +                            estimate_cache_size(mmap_size, istate->cache_nr));
 +      }
  
        src_offset = sizeof(*hdr);
        for (i = 0; i < istate->cache_nr; i++) {
                unsigned long consumed;
  
                disk_ce = (struct ondisk_cache_entry *)((char *)mmap + src_offset);
 -              ce = create_from_disk(disk_ce, &consumed, previous_name);
 +              ce = create_from_disk(istate->ce_mem_pool, disk_ce, &consumed, previous_name);
                set_index_entry(istate, i, ce);
  
                src_offset += consumed;
@@@ -2003,7 -1905,7 +2005,7 @@@ int read_index_from(struct index_state 
        uint64_t start = getnanotime();
        struct split_index *split_index;
        int ret;
 -      char *base_sha1_hex;
 +      char *base_oid_hex;
        char *base_path;
  
        /* istate->initialized covers both .git/index and .git/sharedindex.xxx */
        trace_performance_since(start, "read cache %s", path);
  
        split_index = istate->split_index;
 -      if (!split_index || is_null_sha1(split_index->base_sha1)) {
 +      if (!split_index || is_null_oid(&split_index->base_oid)) {
                post_read_index_from(istate);
                return ret;
        }
        else
                split_index->base = xcalloc(1, sizeof(*split_index->base));
  
 -      base_sha1_hex = sha1_to_hex(split_index->base_sha1);
 -      base_path = xstrfmt("%s/sharedindex.%s", gitdir, base_sha1_hex);
 +      base_oid_hex = oid_to_hex(&split_index->base_oid);
 +      base_path = xstrfmt("%s/sharedindex.%s", gitdir, base_oid_hex);
        ret = do_read_index(split_index->base, base_path, 1);
 -      if (hashcmp(split_index->base_sha1, split_index->base->sha1))
 +      if (oidcmp(&split_index->base_oid, &split_index->base->oid))
                die("broken index, expect %s in %s, got %s",
 -                  base_sha1_hex, base_path,
 -                  sha1_to_hex(split_index->base->sha1));
 +                  base_oid_hex, base_path,
 +                  oid_to_hex(&split_index->base->oid));
  
        freshen_shared_index(base_path, 0);
        merge_base_index(istate);
@@@ -2047,15 -1949,17 +2049,15 @@@ int is_index_unborn(struct index_state 
  
  int discard_index(struct index_state *istate)
  {
 -      int i;
 +      /*
 +       * Cache entries in istate->cache[] should have been allocated
 +       * from the memory pool associated with this index, or from an
 +       * associated split_index. There is no need to free individual
 +       * cache entries. validate_cache_entries can detect when this
 +       * assertion does not hold.
 +       */
 +      validate_cache_entries(istate);
  
 -      for (i = 0; i < istate->cache_nr; i++) {
 -              if (istate->cache[i]->index &&
 -                  istate->split_index &&
 -                  istate->split_index->base &&
 -                  istate->cache[i]->index <= istate->split_index->base->cache_nr &&
 -                  istate->cache[i] == istate->split_index->base->cache[istate->cache[i]->index - 1])
 -                      continue;
 -              free(istate->cache[i]);
 -      }
        resolve_undo_clear_index(istate);
        istate->cache_nr = 0;
        istate->cache_changed = 0;
        discard_split_index(istate);
        free_untracked_cache(istate->untracked);
        istate->untracked = NULL;
 +
 +      if (istate->ce_mem_pool) {
 +              mem_pool_discard(istate->ce_mem_pool, should_validate_cache_entries());
 +              istate->ce_mem_pool = NULL;
 +      }
 +
        return 0;
  }
  
 +/*
 + * Validate the cache entries of this index.
 + * All cache entries associated with this index
 + * should have been allocated by the memory pool
 + * associated with this index, or by a referenced
 + * split index.
 + */
 +void validate_cache_entries(const struct index_state *istate)
 +{
 +      int i;
 +
 +      if (!should_validate_cache_entries() ||!istate || !istate->initialized)
 +              return;
 +
 +      for (i = 0; i < istate->cache_nr; i++) {
 +              if (!istate) {
 +                      die("internal error: cache entry is not allocated from expected memory pool");
 +              } else if (!istate->ce_mem_pool ||
 +                      !mem_pool_contains(istate->ce_mem_pool, istate->cache[i])) {
 +                      if (!istate->split_index ||
 +                              !istate->split_index->base ||
 +                              !istate->split_index->base->ce_mem_pool ||
 +                              !mem_pool_contains(istate->split_index->base->ce_mem_pool, istate->cache[i])) {
 +                              die("internal error: cache entry is not allocated from expected memory pool");
 +                      }
 +              }
 +      }
 +
 +      if (istate->split_index)
 +              validate_cache_entries(istate->split_index->base);
 +}
 +
  int unmerged_index(const struct index_state *istate)
  {
        int i;
        return 0;
  }
  
+ int index_has_changes(const struct index_state *istate,
+                     struct tree *tree,
+                     struct strbuf *sb)
+ {
+       struct object_id cmp;
+       int i;
+       if (istate != &the_index) {
+               BUG("index_has_changes cannot yet accept istate != &the_index; do_diff_cache needs updating first.");
+       }
+       if (tree)
+               cmp = tree->object.oid;
+       if (tree || !get_oid_tree("HEAD", &cmp)) {
+               struct diff_options opt;
+               diff_setup(&opt);
+               opt.flags.exit_with_status = 1;
+               if (!sb)
+                       opt.flags.quick = 1;
+               do_diff_cache(&cmp, &opt);
+               diffcore_std(&opt);
+               for (i = 0; sb && i < diff_queued_diff.nr; i++) {
+                       if (i)
+                               strbuf_addch(sb, ' ');
+                       strbuf_addstr(sb, diff_queued_diff.queue[i]->two->path);
+               }
+               diff_flush(&opt);
+               return opt.flags.has_changes != 0;
+       } else {
+               for (i = 0; sb && i < istate->cache_nr; i++) {
+                       if (i)
+                               strbuf_addch(sb, ' ');
+                       strbuf_addstr(sb, istate->cache[i]->name);
+               }
+               return !!istate->cache_nr;
+       }
+ }
  #define WRITE_BUFFER_SIZE 8192
  static unsigned char write_buffer[WRITE_BUFFER_SIZE];
  static unsigned long write_buffer_len;
@@@ -2355,7 -2259,7 +2395,7 @@@ static int verify_index_from(const stru
        if (n != the_hash_algo->rawsz)
                goto out;
  
 -      if (hashcmp(istate->sha1, hash))
 +      if (hashcmp(istate->oid.hash, hash))
                goto out;
  
        close(fd);
@@@ -2429,7 -2333,7 +2469,7 @@@ static int do_write_index(struct index_
  
        if (!istate->version) {
                istate->version = get_index_format_default();
 -              if (getenv("GIT_TEST_SPLIT_INDEX"))
 +              if (git_env_bool("GIT_TEST_SPLIT_INDEX", 0))
                        init_split_index(istate);
        }
  
                        return -1;
        }
  
 -      if (ce_flush(&c, newfd, istate->sha1))
 +      if (ce_flush(&c, newfd, istate->oid.hash))
                return -1;
        if (close_tempfile_gently(tempfile)) {
                error(_("could not close '%s'"), tempfile->filename.buf);
@@@ -2658,10 -2562,10 +2698,10 @@@ static int write_shared_index(struct in
                return ret;
        }
        ret = rename_tempfile(temp,
 -                            git_path("sharedindex.%s", sha1_to_hex(si->base->sha1)));
 +                            git_path("sharedindex.%s", oid_to_hex(&si->base->oid)));
        if (!ret) {
 -              hashcpy(si->base_sha1, si->base->sha1);
 -              clean_shared_index_files(sha1_to_hex(si->base->sha1));
 +              oidcpy(&si->base_oid, &si->base->oid);
 +              clean_shared_index_files(oid_to_hex(&si->base->oid));
        }
  
        return ret;
@@@ -2715,13 -2619,13 +2755,13 @@@ int write_locked_index(struct index_sta
        if (!si || alternate_index_output ||
            (istate->cache_changed & ~EXTMASK)) {
                if (si)
 -                      hashclr(si->base_sha1);
 +                      oidclr(&si->base_oid);
                ret = do_write_locked_index(istate, lock, flags);
                goto out;
        }
  
 -      if (getenv("GIT_TEST_SPLIT_INDEX")) {
 -              int v = si->base_sha1[0];
 +      if (git_env_bool("GIT_TEST_SPLIT_INDEX", 0)) {
 +              int v = si->base_oid.hash[0];
                if ((v & 15) < 6)
                        istate->cache_changed |= SPLIT_INDEX_ORDERED;
        }
  
                temp = mks_tempfile(git_path("sharedindex_XXXXXX"));
                if (!temp) {
 -                      hashclr(si->base_sha1);
 +                      oidclr(&si->base_oid);
                        ret = do_write_locked_index(istate, lock, flags);
                        goto out;
                }
        /* Freshen the shared index only if the split-index was written */
        if (!ret && !new_shared_index) {
                const char *shared_index = git_path("sharedindex.%s",
 -                                                  sha1_to_hex(si->base_sha1));
 +                                                  oid_to_hex(&si->base_oid));
                freshen_shared_index(shared_index, 1);
        }
  
@@@ -2782,13 -2686,14 +2822,13 @@@ int read_index_unmerged(struct index_st
        for (i = 0; i < istate->cache_nr; i++) {
                struct cache_entry *ce = istate->cache[i];
                struct cache_entry *new_ce;
 -              int size, len;
 +              int len;
  
                if (!ce_stage(ce))
                        continue;
                unmerged = 1;
                len = ce_namelen(ce);
 -              size = cache_entry_size(len);
 -              new_ce = xcalloc(1, size);
 +              new_ce = make_empty_cache_entry(istate, len);
                memcpy(new_ce->name, ce->name, len);
                new_ce->ce_flags = create_ce_flags(0) | CE_CONFLICTED;
                new_ce->ce_namelen = len;
@@@ -2853,7 -2758,7 +2893,7 @@@ void *read_blob_data_from_index(const s
        }
        if (pos < 0)
                return NULL;
 -      data = read_sha1_file(istate->cache[pos]->oid.hash, &type, &sz);
 +      data = read_object_file(&istate->cache[pos]->oid, &type, &sz);
        if (!data || type != OBJ_BLOB) {
                free(data);
                return NULL;
@@@ -2897,41 -2802,3 +2937,41 @@@ void move_index_extensions(struct index
        dst->untracked = src->untracked;
        src->untracked = NULL;
  }
 +
 +struct cache_entry *dup_cache_entry(const struct cache_entry *ce,
 +                                  struct index_state *istate)
 +{
 +      unsigned int size = ce_size(ce);
 +      int mem_pool_allocated;
 +      struct cache_entry *new_entry = make_empty_cache_entry(istate, ce_namelen(ce));
 +      mem_pool_allocated = new_entry->mem_pool_allocated;
 +
 +      memcpy(new_entry, ce, size);
 +      new_entry->mem_pool_allocated = mem_pool_allocated;
 +      return new_entry;
 +}
 +
 +void discard_cache_entry(struct cache_entry *ce)
 +{
 +      if (ce && should_validate_cache_entries())
 +              memset(ce, 0xCD, cache_entry_size(ce->ce_namelen));
 +
 +      if (ce && ce->mem_pool_allocated)
 +              return;
 +
 +      free(ce);
 +}
 +
 +int should_validate_cache_entries(void)
 +{
 +      static int validate_index_cache_entries = -1;
 +
 +      if (validate_index_cache_entries < 0) {
 +              if (getenv("GIT_TEST_VALIDATE_INDEX_CACHE_ENTRIES"))
 +                      validate_index_cache_entries = 1;
 +              else
 +                      validate_index_cache_entries = 0;
 +      }
 +
 +      return validate_index_cache_entries;
 +}