Git 2.23.4
[git] / merge-recursive.c
1 /*
2  * Recursive Merge algorithm stolen from git-merge-recursive.py by
3  * Fredrik Kuivinen.
4  * The thieves were Alex Riesen and Johannes Schindelin, in June/July 2006
5  */
6 #include "cache.h"
7 #include "config.h"
8 #include "advice.h"
9 #include "lockfile.h"
10 #include "cache-tree.h"
11 #include "object-store.h"
12 #include "repository.h"
13 #include "commit.h"
14 #include "blob.h"
15 #include "builtin.h"
16 #include "tree-walk.h"
17 #include "diff.h"
18 #include "diffcore.h"
19 #include "tag.h"
20 #include "alloc.h"
21 #include "unpack-trees.h"
22 #include "string-list.h"
23 #include "xdiff-interface.h"
24 #include "ll-merge.h"
25 #include "attr.h"
26 #include "merge-recursive.h"
27 #include "dir.h"
28 #include "submodule.h"
29 #include "revision.h"
30 #include "commit-reach.h"
31
32 struct path_hashmap_entry {
33         struct hashmap_entry e;
34         char path[FLEX_ARRAY];
35 };
36
37 static int path_hashmap_cmp(const void *cmp_data,
38                             const void *entry,
39                             const void *entry_or_key,
40                             const void *keydata)
41 {
42         const struct path_hashmap_entry *a = entry;
43         const struct path_hashmap_entry *b = entry_or_key;
44         const char *key = keydata;
45
46         if (ignore_case)
47                 return strcasecmp(a->path, key ? key : b->path);
48         else
49                 return strcmp(a->path, key ? key : b->path);
50 }
51
52 static unsigned int path_hash(const char *path)
53 {
54         return ignore_case ? strihash(path) : strhash(path);
55 }
56
57 static struct dir_rename_entry *dir_rename_find_entry(struct hashmap *hashmap,
58                                                       char *dir)
59 {
60         struct dir_rename_entry key;
61
62         if (dir == NULL)
63                 return NULL;
64         hashmap_entry_init(&key, strhash(dir));
65         key.dir = dir;
66         return hashmap_get(hashmap, &key, NULL);
67 }
68
69 static int dir_rename_cmp(const void *unused_cmp_data,
70                           const void *entry,
71                           const void *entry_or_key,
72                           const void *unused_keydata)
73 {
74         const struct dir_rename_entry *e1 = entry;
75         const struct dir_rename_entry *e2 = entry_or_key;
76
77         return strcmp(e1->dir, e2->dir);
78 }
79
80 static void dir_rename_init(struct hashmap *map)
81 {
82         hashmap_init(map, dir_rename_cmp, NULL, 0);
83 }
84
85 static void dir_rename_entry_init(struct dir_rename_entry *entry,
86                                   char *directory)
87 {
88         hashmap_entry_init(entry, strhash(directory));
89         entry->dir = directory;
90         entry->non_unique_new_dir = 0;
91         strbuf_init(&entry->new_dir, 0);
92         string_list_init(&entry->possible_new_dirs, 0);
93 }
94
95 static struct collision_entry *collision_find_entry(struct hashmap *hashmap,
96                                                     char *target_file)
97 {
98         struct collision_entry key;
99
100         hashmap_entry_init(&key, strhash(target_file));
101         key.target_file = target_file;
102         return hashmap_get(hashmap, &key, NULL);
103 }
104
105 static int collision_cmp(void *unused_cmp_data,
106                          const struct collision_entry *e1,
107                          const struct collision_entry *e2,
108                          const void *unused_keydata)
109 {
110         return strcmp(e1->target_file, e2->target_file);
111 }
112
113 static void collision_init(struct hashmap *map)
114 {
115         hashmap_init(map, (hashmap_cmp_fn) collision_cmp, NULL, 0);
116 }
117
118 static void flush_output(struct merge_options *opt)
119 {
120         if (opt->buffer_output < 2 && opt->obuf.len) {
121                 fputs(opt->obuf.buf, stdout);
122                 strbuf_reset(&opt->obuf);
123         }
124 }
125
126 static int err(struct merge_options *opt, const char *err, ...)
127 {
128         va_list params;
129
130         if (opt->buffer_output < 2)
131                 flush_output(opt);
132         else {
133                 strbuf_complete(&opt->obuf, '\n');
134                 strbuf_addstr(&opt->obuf, "error: ");
135         }
136         va_start(params, err);
137         strbuf_vaddf(&opt->obuf, err, params);
138         va_end(params);
139         if (opt->buffer_output > 1)
140                 strbuf_addch(&opt->obuf, '\n');
141         else {
142                 error("%s", opt->obuf.buf);
143                 strbuf_reset(&opt->obuf);
144         }
145
146         return -1;
147 }
148
149 static struct tree *shift_tree_object(struct repository *repo,
150                                       struct tree *one, struct tree *two,
151                                       const char *subtree_shift)
152 {
153         struct object_id shifted;
154
155         if (!*subtree_shift) {
156                 shift_tree(repo, &one->object.oid, &two->object.oid, &shifted, 0);
157         } else {
158                 shift_tree_by(repo, &one->object.oid, &two->object.oid, &shifted,
159                               subtree_shift);
160         }
161         if (oideq(&two->object.oid, &shifted))
162                 return two;
163         return lookup_tree(repo, &shifted);
164 }
165
166 static inline void set_commit_tree(struct commit *c, struct tree *t)
167 {
168         c->maybe_tree = t;
169 }
170
171 static struct commit *make_virtual_commit(struct repository *repo,
172                                           struct tree *tree,
173                                           const char *comment)
174 {
175         struct commit *commit = alloc_commit_node(repo);
176
177         set_merge_remote_desc(commit, comment, (struct object *)commit);
178         set_commit_tree(commit, tree);
179         commit->object.parsed = 1;
180         return commit;
181 }
182
183 /*
184  * Since we use get_tree_entry(), which does not put the read object into
185  * the object pool, we cannot rely on a == b.
186  */
187 static int oid_eq(const struct object_id *a, const struct object_id *b)
188 {
189         if (!a && !b)
190                 return 2;
191         return a && b && oideq(a, b);
192 }
193
194 enum rename_type {
195         RENAME_NORMAL = 0,
196         RENAME_VIA_DIR,
197         RENAME_ADD,
198         RENAME_DELETE,
199         RENAME_ONE_FILE_TO_ONE,
200         RENAME_ONE_FILE_TO_TWO,
201         RENAME_TWO_FILES_TO_ONE
202 };
203
204 /*
205  * Since we want to write the index eventually, we cannot reuse the index
206  * for these (temporary) data.
207  */
208 struct stage_data {
209         struct diff_filespec stages[4]; /* mostly for oid & mode; maybe path */
210         struct rename_conflict_info *rename_conflict_info;
211         unsigned processed:1;
212 };
213
214 struct rename {
215         unsigned processed:1;
216         struct diff_filepair *pair;
217         const char *branch; /* branch that the rename occurred on */
218         /*
219          * If directory rename detection affected this rename, what was its
220          * original type ('A' or 'R') and it's original destination before
221          * the directory rename (otherwise, '\0' and NULL for these two vars).
222          */
223         char dir_rename_original_type;
224         char *dir_rename_original_dest;
225         /*
226          * Purpose of src_entry and dst_entry:
227          *
228          * If 'before' is renamed to 'after' then src_entry will contain
229          * the versions of 'before' from the merge_base, HEAD, and MERGE in
230          * stages 1, 2, and 3; dst_entry will contain the respective
231          * versions of 'after' in corresponding locations.  Thus, we have a
232          * total of six modes and oids, though some will be null.  (Stage 0
233          * is ignored; we're interested in handling conflicts.)
234          *
235          * Since we don't turn on break-rewrites by default, neither
236          * src_entry nor dst_entry can have all three of their stages have
237          * non-null oids, meaning at most four of the six will be non-null.
238          * Also, since this is a rename, both src_entry and dst_entry will
239          * have at least one non-null oid, meaning at least two will be
240          * non-null.  Of the six oids, a typical rename will have three be
241          * non-null.  Only two implies a rename/delete, and four implies a
242          * rename/add.
243          */
244         struct stage_data *src_entry;
245         struct stage_data *dst_entry;
246 };
247
248 struct rename_conflict_info {
249         enum rename_type rename_type;
250         struct rename *ren1;
251         struct rename *ren2;
252 };
253
254 static inline void setup_rename_conflict_info(enum rename_type rename_type,
255                                               struct merge_options *opt,
256                                               struct rename *ren1,
257                                               struct rename *ren2)
258 {
259         struct rename_conflict_info *ci;
260
261         /*
262          * When we have two renames involved, it's easiest to get the
263          * correct things into stage 2 and 3, and to make sure that the
264          * content merge puts HEAD before the other branch if we just
265          * ensure that branch1 == opt->branch1.  So, simply flip arguments
266          * around if we don't have that.
267          */
268         if (ren2 && ren1->branch != opt->branch1) {
269                 setup_rename_conflict_info(rename_type, opt, ren2, ren1);
270                 return;
271         }
272
273         ci = xcalloc(1, sizeof(struct rename_conflict_info));
274         ci->rename_type = rename_type;
275         ci->ren1 = ren1;
276         ci->ren2 = ren2;
277
278         ci->ren1->dst_entry->processed = 0;
279         ci->ren1->dst_entry->rename_conflict_info = ci;
280         if (ren2) {
281                 ci->ren2->dst_entry->rename_conflict_info = ci;
282         }
283 }
284
285 static int show(struct merge_options *opt, int v)
286 {
287         return (!opt->call_depth && opt->verbosity >= v) || opt->verbosity >= 5;
288 }
289
290 __attribute__((format (printf, 3, 4)))
291 static void output(struct merge_options *opt, int v, const char *fmt, ...)
292 {
293         va_list ap;
294
295         if (!show(opt, v))
296                 return;
297
298         strbuf_addchars(&opt->obuf, ' ', opt->call_depth * 2);
299
300         va_start(ap, fmt);
301         strbuf_vaddf(&opt->obuf, fmt, ap);
302         va_end(ap);
303
304         strbuf_addch(&opt->obuf, '\n');
305         if (!opt->buffer_output)
306                 flush_output(opt);
307 }
308
309 static void output_commit_title(struct merge_options *opt, struct commit *commit)
310 {
311         struct merge_remote_desc *desc;
312
313         strbuf_addchars(&opt->obuf, ' ', opt->call_depth * 2);
314         desc = merge_remote_util(commit);
315         if (desc)
316                 strbuf_addf(&opt->obuf, "virtual %s\n", desc->name);
317         else {
318                 strbuf_add_unique_abbrev(&opt->obuf, &commit->object.oid,
319                                          DEFAULT_ABBREV);
320                 strbuf_addch(&opt->obuf, ' ');
321                 if (parse_commit(commit) != 0)
322                         strbuf_addstr(&opt->obuf, _("(bad commit)\n"));
323                 else {
324                         const char *title;
325                         const char *msg = get_commit_buffer(commit, NULL);
326                         int len = find_commit_subject(msg, &title);
327                         if (len)
328                                 strbuf_addf(&opt->obuf, "%.*s\n", len, title);
329                         unuse_commit_buffer(commit, msg);
330                 }
331         }
332         flush_output(opt);
333 }
334
335 static int add_cacheinfo(struct merge_options *opt,
336                          const struct diff_filespec *blob,
337                          const char *path, int stage, int refresh, int options)
338 {
339         struct index_state *istate = opt->repo->index;
340         struct cache_entry *ce;
341         int ret;
342
343         ce = make_cache_entry(istate, blob->mode, &blob->oid, path, stage, 0);
344         if (!ce)
345                 return err(opt, _("add_cacheinfo failed for path '%s'; merge aborting."), path);
346
347         ret = add_index_entry(istate, ce, options);
348         if (refresh) {
349                 struct cache_entry *nce;
350
351                 nce = refresh_cache_entry(istate, ce,
352                                           CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
353                 if (!nce)
354                         return err(opt, _("add_cacheinfo failed to refresh for path '%s'; merge aborting."), path);
355                 if (nce != ce)
356                         ret = add_index_entry(istate, nce, options);
357         }
358         return ret;
359 }
360
361 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
362 {
363         parse_tree(tree);
364         init_tree_desc(desc, tree->buffer, tree->size);
365 }
366
367 static int unpack_trees_start(struct merge_options *opt,
368                               struct tree *common,
369                               struct tree *head,
370                               struct tree *merge)
371 {
372         int rc;
373         struct tree_desc t[3];
374         struct index_state tmp_index = { NULL };
375
376         memset(&opt->unpack_opts, 0, sizeof(opt->unpack_opts));
377         if (opt->call_depth)
378                 opt->unpack_opts.index_only = 1;
379         else
380                 opt->unpack_opts.update = 1;
381         opt->unpack_opts.merge = 1;
382         opt->unpack_opts.head_idx = 2;
383         opt->unpack_opts.fn = threeway_merge;
384         opt->unpack_opts.src_index = opt->repo->index;
385         opt->unpack_opts.dst_index = &tmp_index;
386         opt->unpack_opts.aggressive = !merge_detect_rename(opt);
387         setup_unpack_trees_porcelain(&opt->unpack_opts, "merge");
388
389         init_tree_desc_from_tree(t+0, common);
390         init_tree_desc_from_tree(t+1, head);
391         init_tree_desc_from_tree(t+2, merge);
392
393         rc = unpack_trees(3, t, &opt->unpack_opts);
394         cache_tree_free(&opt->repo->index->cache_tree);
395
396         /*
397          * Update opt->repo->index to match the new results, AFTER saving a copy
398          * in opt->orig_index.  Update src_index to point to the saved copy.
399          * (verify_uptodate() checks src_index, and the original index is
400          * the one that had the necessary modification timestamps.)
401          */
402         opt->orig_index = *opt->repo->index;
403         *opt->repo->index = tmp_index;
404         opt->unpack_opts.src_index = &opt->orig_index;
405
406         return rc;
407 }
408
409 static void unpack_trees_finish(struct merge_options *opt)
410 {
411         discard_index(&opt->orig_index);
412         clear_unpack_trees_porcelain(&opt->unpack_opts);
413 }
414
415 struct tree *write_tree_from_memory(struct merge_options *opt)
416 {
417         struct tree *result = NULL;
418         struct index_state *istate = opt->repo->index;
419
420         if (unmerged_index(istate)) {
421                 int i;
422                 fprintf(stderr, "BUG: There are unmerged index entries:\n");
423                 for (i = 0; i < istate->cache_nr; i++) {
424                         const struct cache_entry *ce = istate->cache[i];
425                         if (ce_stage(ce))
426                                 fprintf(stderr, "BUG: %d %.*s\n", ce_stage(ce),
427                                         (int)ce_namelen(ce), ce->name);
428                 }
429                 BUG("unmerged index entries in merge-recursive.c");
430         }
431
432         if (!istate->cache_tree)
433                 istate->cache_tree = cache_tree();
434
435         if (!cache_tree_fully_valid(istate->cache_tree) &&
436             cache_tree_update(istate, 0) < 0) {
437                 err(opt, _("error building trees"));
438                 return NULL;
439         }
440
441         result = lookup_tree(opt->repo, &istate->cache_tree->oid);
442
443         return result;
444 }
445
446 static int save_files_dirs(const struct object_id *oid,
447                            struct strbuf *base, const char *path,
448                            unsigned int mode, int stage, void *context)
449 {
450         struct path_hashmap_entry *entry;
451         int baselen = base->len;
452         struct merge_options *opt = context;
453
454         strbuf_addstr(base, path);
455
456         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
457         hashmap_entry_init(entry, path_hash(entry->path));
458         hashmap_add(&opt->current_file_dir_set, entry);
459
460         strbuf_setlen(base, baselen);
461         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
462 }
463
464 static void get_files_dirs(struct merge_options *opt, struct tree *tree)
465 {
466         struct pathspec match_all;
467         memset(&match_all, 0, sizeof(match_all));
468         read_tree_recursive(opt->repo, tree, "", 0, 0,
469                             &match_all, save_files_dirs, opt);
470 }
471
472 static int get_tree_entry_if_blob(struct repository *r,
473                                   const struct object_id *tree,
474                                   const char *path,
475                                   struct diff_filespec *dfs)
476 {
477         int ret;
478
479         ret = get_tree_entry(r, tree, path, &dfs->oid, &dfs->mode);
480         if (S_ISDIR(dfs->mode)) {
481                 oidcpy(&dfs->oid, &null_oid);
482                 dfs->mode = 0;
483         }
484         return ret;
485 }
486
487 /*
488  * Returns an index_entry instance which doesn't have to correspond to
489  * a real cache entry in Git's index.
490  */
491 static struct stage_data *insert_stage_data(struct repository *r,
492                 const char *path,
493                 struct tree *o, struct tree *a, struct tree *b,
494                 struct string_list *entries)
495 {
496         struct string_list_item *item;
497         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
498         get_tree_entry_if_blob(r, &o->object.oid, path, &e->stages[1]);
499         get_tree_entry_if_blob(r, &a->object.oid, path, &e->stages[2]);
500         get_tree_entry_if_blob(r, &b->object.oid, path, &e->stages[3]);
501         item = string_list_insert(entries, path);
502         item->util = e;
503         return e;
504 }
505
506 /*
507  * Create a dictionary mapping file names to stage_data objects. The
508  * dictionary contains one entry for every path with a non-zero stage entry.
509  */
510 static struct string_list *get_unmerged(struct index_state *istate)
511 {
512         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
513         int i;
514
515         unmerged->strdup_strings = 1;
516
517         for (i = 0; i < istate->cache_nr; i++) {
518                 struct string_list_item *item;
519                 struct stage_data *e;
520                 const struct cache_entry *ce = istate->cache[i];
521                 if (!ce_stage(ce))
522                         continue;
523
524                 item = string_list_lookup(unmerged, ce->name);
525                 if (!item) {
526                         item = string_list_insert(unmerged, ce->name);
527                         item->util = xcalloc(1, sizeof(struct stage_data));
528                 }
529                 e = item->util;
530                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
531                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
532         }
533
534         return unmerged;
535 }
536
537 static int string_list_df_name_compare(const char *one, const char *two)
538 {
539         int onelen = strlen(one);
540         int twolen = strlen(two);
541         /*
542          * Here we only care that entries for D/F conflicts are
543          * adjacent, in particular with the file of the D/F conflict
544          * appearing before files below the corresponding directory.
545          * The order of the rest of the list is irrelevant for us.
546          *
547          * To achieve this, we sort with df_name_compare and provide
548          * the mode S_IFDIR so that D/F conflicts will sort correctly.
549          * We use the mode S_IFDIR for everything else for simplicity,
550          * since in other cases any changes in their order due to
551          * sorting cause no problems for us.
552          */
553         int cmp = df_name_compare(one, onelen, S_IFDIR,
554                                   two, twolen, S_IFDIR);
555         /*
556          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
557          * that 'foo' comes before 'foo/bar'.
558          */
559         if (cmp)
560                 return cmp;
561         return onelen - twolen;
562 }
563
564 static void record_df_conflict_files(struct merge_options *opt,
565                                      struct string_list *entries)
566 {
567         /* If there is a D/F conflict and the file for such a conflict
568          * currently exists in the working tree, we want to allow it to be
569          * removed to make room for the corresponding directory if needed.
570          * The files underneath the directories of such D/F conflicts will
571          * be processed before the corresponding file involved in the D/F
572          * conflict.  If the D/F directory ends up being removed by the
573          * merge, then we won't have to touch the D/F file.  If the D/F
574          * directory needs to be written to the working copy, then the D/F
575          * file will simply be removed (in make_room_for_path()) to make
576          * room for the necessary paths.  Note that if both the directory
577          * and the file need to be present, then the D/F file will be
578          * reinstated with a new unique name at the time it is processed.
579          */
580         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
581         const char *last_file = NULL;
582         int last_len = 0;
583         int i;
584
585         /*
586          * If we're merging merge-bases, we don't want to bother with
587          * any working directory changes.
588          */
589         if (opt->call_depth)
590                 return;
591
592         /* Ensure D/F conflicts are adjacent in the entries list. */
593         for (i = 0; i < entries->nr; i++) {
594                 struct string_list_item *next = &entries->items[i];
595                 string_list_append(&df_sorted_entries, next->string)->util =
596                                    next->util;
597         }
598         df_sorted_entries.cmp = string_list_df_name_compare;
599         string_list_sort(&df_sorted_entries);
600
601         string_list_clear(&opt->df_conflict_file_set, 1);
602         for (i = 0; i < df_sorted_entries.nr; i++) {
603                 const char *path = df_sorted_entries.items[i].string;
604                 int len = strlen(path);
605                 struct stage_data *e = df_sorted_entries.items[i].util;
606
607                 /*
608                  * Check if last_file & path correspond to a D/F conflict;
609                  * i.e. whether path is last_file+'/'+<something>.
610                  * If so, record that it's okay to remove last_file to make
611                  * room for path and friends if needed.
612                  */
613                 if (last_file &&
614                     len > last_len &&
615                     memcmp(path, last_file, last_len) == 0 &&
616                     path[last_len] == '/') {
617                         string_list_insert(&opt->df_conflict_file_set, last_file);
618                 }
619
620                 /*
621                  * Determine whether path could exist as a file in the
622                  * working directory as a possible D/F conflict.  This
623                  * will only occur when it exists in stage 2 as a
624                  * file.
625                  */
626                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
627                         last_file = path;
628                         last_len = len;
629                 } else {
630                         last_file = NULL;
631                 }
632         }
633         string_list_clear(&df_sorted_entries, 0);
634 }
635
636 static int update_stages(struct merge_options *opt, const char *path,
637                          const struct diff_filespec *o,
638                          const struct diff_filespec *a,
639                          const struct diff_filespec *b)
640 {
641
642         /*
643          * NOTE: It is usually a bad idea to call update_stages on a path
644          * before calling update_file on that same path, since it can
645          * sometimes lead to spurious "refusing to lose untracked file..."
646          * messages from update_file (via make_room_for path via
647          * would_lose_untracked).  Instead, reverse the order of the calls
648          * (executing update_file first and then update_stages).
649          */
650         int clear = 1;
651         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
652         if (clear)
653                 if (remove_file_from_index(opt->repo->index, path))
654                         return -1;
655         if (o)
656                 if (add_cacheinfo(opt, o, path, 1, 0, options))
657                         return -1;
658         if (a)
659                 if (add_cacheinfo(opt, a, path, 2, 0, options))
660                         return -1;
661         if (b)
662                 if (add_cacheinfo(opt, b, path, 3, 0, options))
663                         return -1;
664         return 0;
665 }
666
667 static void update_entry(struct stage_data *entry,
668                          struct diff_filespec *o,
669                          struct diff_filespec *a,
670                          struct diff_filespec *b)
671 {
672         entry->processed = 0;
673         entry->stages[1].mode = o->mode;
674         entry->stages[2].mode = a->mode;
675         entry->stages[3].mode = b->mode;
676         oidcpy(&entry->stages[1].oid, &o->oid);
677         oidcpy(&entry->stages[2].oid, &a->oid);
678         oidcpy(&entry->stages[3].oid, &b->oid);
679 }
680
681 static int remove_file(struct merge_options *opt, int clean,
682                        const char *path, int no_wd)
683 {
684         int update_cache = opt->call_depth || clean;
685         int update_working_directory = !opt->call_depth && !no_wd;
686
687         if (update_cache) {
688                 if (remove_file_from_index(opt->repo->index, path))
689                         return -1;
690         }
691         if (update_working_directory) {
692                 if (ignore_case) {
693                         struct cache_entry *ce;
694                         ce = index_file_exists(opt->repo->index, path, strlen(path),
695                                                ignore_case);
696                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
697                                 return 0;
698                 }
699                 if (remove_path(path))
700                         return -1;
701         }
702         return 0;
703 }
704
705 /* add a string to a strbuf, but converting "/" to "_" */
706 static void add_flattened_path(struct strbuf *out, const char *s)
707 {
708         size_t i = out->len;
709         strbuf_addstr(out, s);
710         for (; i < out->len; i++)
711                 if (out->buf[i] == '/')
712                         out->buf[i] = '_';
713 }
714
715 static char *unique_path(struct merge_options *opt, const char *path, const char *branch)
716 {
717         struct path_hashmap_entry *entry;
718         struct strbuf newpath = STRBUF_INIT;
719         int suffix = 0;
720         size_t base_len;
721
722         strbuf_addf(&newpath, "%s~", path);
723         add_flattened_path(&newpath, branch);
724
725         base_len = newpath.len;
726         while (hashmap_get_from_hash(&opt->current_file_dir_set,
727                                      path_hash(newpath.buf), newpath.buf) ||
728                (!opt->call_depth && file_exists(newpath.buf))) {
729                 strbuf_setlen(&newpath, base_len);
730                 strbuf_addf(&newpath, "_%d", suffix++);
731         }
732
733         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
734         hashmap_entry_init(entry, path_hash(entry->path));
735         hashmap_add(&opt->current_file_dir_set, entry);
736         return strbuf_detach(&newpath, NULL);
737 }
738
739 /**
740  * Check whether a directory in the index is in the way of an incoming
741  * file.  Return 1 if so.  If check_working_copy is non-zero, also
742  * check the working directory.  If empty_ok is non-zero, also return
743  * 0 in the case where the working-tree dir exists but is empty.
744  */
745 static int dir_in_way(struct index_state *istate, const char *path,
746                       int check_working_copy, int empty_ok)
747 {
748         int pos;
749         struct strbuf dirpath = STRBUF_INIT;
750         struct stat st;
751
752         strbuf_addstr(&dirpath, path);
753         strbuf_addch(&dirpath, '/');
754
755         pos = index_name_pos(istate, dirpath.buf, dirpath.len);
756
757         if (pos < 0)
758                 pos = -1 - pos;
759         if (pos < istate->cache_nr &&
760             !strncmp(dirpath.buf, istate->cache[pos]->name, dirpath.len)) {
761                 strbuf_release(&dirpath);
762                 return 1;
763         }
764
765         strbuf_release(&dirpath);
766         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
767                 !(empty_ok && is_empty_dir(path));
768 }
769
770 /*
771  * Returns whether path was tracked in the index before the merge started,
772  * and its oid and mode match the specified values
773  */
774 static int was_tracked_and_matches(struct merge_options *opt, const char *path,
775                                    const struct diff_filespec *blob)
776 {
777         int pos = index_name_pos(&opt->orig_index, path, strlen(path));
778         struct cache_entry *ce;
779
780         if (0 > pos)
781                 /* we were not tracking this path before the merge */
782                 return 0;
783
784         /* See if the file we were tracking before matches */
785         ce = opt->orig_index.cache[pos];
786         return (oid_eq(&ce->oid, &blob->oid) && ce->ce_mode == blob->mode);
787 }
788
789 /*
790  * Returns whether path was tracked in the index before the merge started
791  */
792 static int was_tracked(struct merge_options *opt, const char *path)
793 {
794         int pos = index_name_pos(&opt->orig_index, path, strlen(path));
795
796         if (0 <= pos)
797                 /* we were tracking this path before the merge */
798                 return 1;
799
800         return 0;
801 }
802
803 static int would_lose_untracked(struct merge_options *opt, const char *path)
804 {
805         struct index_state *istate = opt->repo->index;
806
807         /*
808          * This may look like it can be simplified to:
809          *   return !was_tracked(opt, path) && file_exists(path)
810          * but it can't.  This function needs to know whether path was in
811          * the working tree due to EITHER having been tracked in the index
812          * before the merge OR having been put into the working copy and
813          * index by unpack_trees().  Due to that either-or requirement, we
814          * check the current index instead of the original one.
815          *
816          * Note that we do not need to worry about merge-recursive itself
817          * updating the index after unpack_trees() and before calling this
818          * function, because we strictly require all code paths in
819          * merge-recursive to update the working tree first and the index
820          * second.  Doing otherwise would break
821          * update_file()/would_lose_untracked(); see every comment in this
822          * file which mentions "update_stages".
823          */
824         int pos = index_name_pos(istate, path, strlen(path));
825
826         if (pos < 0)
827                 pos = -1 - pos;
828         while (pos < istate->cache_nr &&
829                !strcmp(path, istate->cache[pos]->name)) {
830                 /*
831                  * If stage #0, it is definitely tracked.
832                  * If it has stage #2 then it was tracked
833                  * before this merge started.  All other
834                  * cases the path was not tracked.
835                  */
836                 switch (ce_stage(istate->cache[pos])) {
837                 case 0:
838                 case 2:
839                         return 0;
840                 }
841                 pos++;
842         }
843         return file_exists(path);
844 }
845
846 static int was_dirty(struct merge_options *opt, const char *path)
847 {
848         struct cache_entry *ce;
849         int dirty = 1;
850
851         if (opt->call_depth || !was_tracked(opt, path))
852                 return !dirty;
853
854         ce = index_file_exists(opt->unpack_opts.src_index,
855                                path, strlen(path), ignore_case);
856         dirty = verify_uptodate(ce, &opt->unpack_opts) != 0;
857         return dirty;
858 }
859
860 static int make_room_for_path(struct merge_options *opt, const char *path)
861 {
862         int status, i;
863         const char *msg = _("failed to create path '%s'%s");
864
865         /* Unlink any D/F conflict files that are in the way */
866         for (i = 0; i < opt->df_conflict_file_set.nr; i++) {
867                 const char *df_path = opt->df_conflict_file_set.items[i].string;
868                 size_t pathlen = strlen(path);
869                 size_t df_pathlen = strlen(df_path);
870                 if (df_pathlen < pathlen &&
871                     path[df_pathlen] == '/' &&
872                     strncmp(path, df_path, df_pathlen) == 0) {
873                         output(opt, 3,
874                                _("Removing %s to make room for subdirectory\n"),
875                                df_path);
876                         unlink(df_path);
877                         unsorted_string_list_delete_item(&opt->df_conflict_file_set,
878                                                          i, 0);
879                         break;
880                 }
881         }
882
883         /* Make sure leading directories are created */
884         status = safe_create_leading_directories_const(path);
885         if (status) {
886                 if (status == SCLD_EXISTS)
887                         /* something else exists */
888                         return err(opt, msg, path, _(": perhaps a D/F conflict?"));
889                 return err(opt, msg, path, "");
890         }
891
892         /*
893          * Do not unlink a file in the work tree if we are not
894          * tracking it.
895          */
896         if (would_lose_untracked(opt, path))
897                 return err(opt, _("refusing to lose untracked file at '%s'"),
898                            path);
899
900         /* Successful unlink is good.. */
901         if (!unlink(path))
902                 return 0;
903         /* .. and so is no existing file */
904         if (errno == ENOENT)
905                 return 0;
906         /* .. but not some other error (who really cares what?) */
907         return err(opt, msg, path, _(": perhaps a D/F conflict?"));
908 }
909
910 static int update_file_flags(struct merge_options *opt,
911                              const struct diff_filespec *contents,
912                              const char *path,
913                              int update_cache,
914                              int update_wd)
915 {
916         int ret = 0;
917
918         if (opt->call_depth)
919                 update_wd = 0;
920
921         if (update_wd) {
922                 enum object_type type;
923                 void *buf;
924                 unsigned long size;
925
926                 if (S_ISGITLINK(contents->mode)) {
927                         /*
928                          * We may later decide to recursively descend into
929                          * the submodule directory and update its index
930                          * and/or work tree, but we do not do that now.
931                          */
932                         update_wd = 0;
933                         goto update_index;
934                 }
935
936                 buf = read_object_file(&contents->oid, &type, &size);
937                 if (!buf)
938                         return err(opt, _("cannot read object %s '%s'"),
939                                    oid_to_hex(&contents->oid), path);
940                 if (type != OBJ_BLOB) {
941                         ret = err(opt, _("blob expected for %s '%s'"),
942                                   oid_to_hex(&contents->oid), path);
943                         goto free_buf;
944                 }
945                 if (S_ISREG(contents->mode)) {
946                         struct strbuf strbuf = STRBUF_INIT;
947                         if (convert_to_working_tree(opt->repo->index, path, buf, size, &strbuf)) {
948                                 free(buf);
949                                 size = strbuf.len;
950                                 buf = strbuf_detach(&strbuf, NULL);
951                         }
952                 }
953
954                 if (make_room_for_path(opt, path) < 0) {
955                         update_wd = 0;
956                         goto free_buf;
957                 }
958                 if (S_ISREG(contents->mode) ||
959                     (!has_symlinks && S_ISLNK(contents->mode))) {
960                         int fd;
961                         int mode = (contents->mode & 0100 ? 0777 : 0666);
962
963                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
964                         if (fd < 0) {
965                                 ret = err(opt, _("failed to open '%s': %s"),
966                                           path, strerror(errno));
967                                 goto free_buf;
968                         }
969                         write_in_full(fd, buf, size);
970                         close(fd);
971                 } else if (S_ISLNK(contents->mode)) {
972                         char *lnk = xmemdupz(buf, size);
973                         safe_create_leading_directories_const(path);
974                         unlink(path);
975                         if (symlink(lnk, path))
976                                 ret = err(opt, _("failed to symlink '%s': %s"),
977                                           path, strerror(errno));
978                         free(lnk);
979                 } else
980                         ret = err(opt,
981                                   _("do not know what to do with %06o %s '%s'"),
982                                   contents->mode, oid_to_hex(&contents->oid), path);
983         free_buf:
984                 free(buf);
985         }
986 update_index:
987         if (!ret && update_cache)
988                 if (add_cacheinfo(opt, contents, path, 0, update_wd,
989                                   ADD_CACHE_OK_TO_ADD))
990                         return -1;
991         return ret;
992 }
993
994 static int update_file(struct merge_options *opt,
995                        int clean,
996                        const struct diff_filespec *contents,
997                        const char *path)
998 {
999         return update_file_flags(opt, contents, path,
1000                                  opt->call_depth || clean, !opt->call_depth);
1001 }
1002
1003 /* Low level file merging, update and removal */
1004
1005 struct merge_file_info {
1006         struct diff_filespec blob; /* mostly use oid & mode; sometimes path */
1007         unsigned clean:1,
1008                  merge:1;
1009 };
1010
1011 static int merge_3way(struct merge_options *opt,
1012                       mmbuffer_t *result_buf,
1013                       const struct diff_filespec *o,
1014                       const struct diff_filespec *a,
1015                       const struct diff_filespec *b,
1016                       const char *branch1,
1017                       const char *branch2,
1018                       const int extra_marker_size)
1019 {
1020         mmfile_t orig, src1, src2;
1021         struct ll_merge_options ll_opts = {0};
1022         char *base_name, *name1, *name2;
1023         int merge_status;
1024
1025         ll_opts.renormalize = opt->renormalize;
1026         ll_opts.extra_marker_size = extra_marker_size;
1027         ll_opts.xdl_opts = opt->xdl_opts;
1028
1029         if (opt->call_depth) {
1030                 ll_opts.virtual_ancestor = 1;
1031                 ll_opts.variant = 0;
1032         } else {
1033                 switch (opt->recursive_variant) {
1034                 case MERGE_RECURSIVE_OURS:
1035                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
1036                         break;
1037                 case MERGE_RECURSIVE_THEIRS:
1038                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
1039                         break;
1040                 default:
1041                         ll_opts.variant = 0;
1042                         break;
1043                 }
1044         }
1045
1046         assert(a->path && b->path);
1047         if (strcmp(a->path, b->path) ||
1048             (opt->ancestor != NULL && strcmp(a->path, o->path) != 0)) {
1049                 base_name = opt->ancestor == NULL ? NULL :
1050                         mkpathdup("%s:%s", opt->ancestor, o->path);
1051                 name1 = mkpathdup("%s:%s", branch1, a->path);
1052                 name2 = mkpathdup("%s:%s", branch2, b->path);
1053         } else {
1054                 base_name = opt->ancestor == NULL ? NULL :
1055                         mkpathdup("%s", opt->ancestor);
1056                 name1 = mkpathdup("%s", branch1);
1057                 name2 = mkpathdup("%s", branch2);
1058         }
1059
1060         read_mmblob(&orig, &o->oid);
1061         read_mmblob(&src1, &a->oid);
1062         read_mmblob(&src2, &b->oid);
1063
1064         merge_status = ll_merge(result_buf, a->path, &orig, base_name,
1065                                 &src1, name1, &src2, name2,
1066                                 opt->repo->index, &ll_opts);
1067
1068         free(base_name);
1069         free(name1);
1070         free(name2);
1071         free(orig.ptr);
1072         free(src1.ptr);
1073         free(src2.ptr);
1074         return merge_status;
1075 }
1076
1077 static int find_first_merges(struct repository *repo,
1078                              struct object_array *result, const char *path,
1079                              struct commit *a, struct commit *b)
1080 {
1081         int i, j;
1082         struct object_array merges = OBJECT_ARRAY_INIT;
1083         struct commit *commit;
1084         int contains_another;
1085
1086         char merged_revision[GIT_MAX_HEXSZ + 2];
1087         const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
1088                                    "--all", merged_revision, NULL };
1089         struct rev_info revs;
1090         struct setup_revision_opt rev_opts;
1091
1092         memset(result, 0, sizeof(struct object_array));
1093         memset(&rev_opts, 0, sizeof(rev_opts));
1094
1095         /* get all revisions that merge commit a */
1096         xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
1097                   oid_to_hex(&a->object.oid));
1098         repo_init_revisions(repo, &revs, NULL);
1099         rev_opts.submodule = path;
1100         /* FIXME: can't handle linked worktrees in submodules yet */
1101         revs.single_worktree = path != NULL;
1102         setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
1103
1104         /* save all revisions from the above list that contain b */
1105         if (prepare_revision_walk(&revs))
1106                 die("revision walk setup failed");
1107         while ((commit = get_revision(&revs)) != NULL) {
1108                 struct object *o = &(commit->object);
1109                 if (in_merge_bases(b, commit))
1110                         add_object_array(o, NULL, &merges);
1111         }
1112         reset_revision_walk();
1113
1114         /* Now we've got all merges that contain a and b. Prune all
1115          * merges that contain another found merge and save them in
1116          * result.
1117          */
1118         for (i = 0; i < merges.nr; i++) {
1119                 struct commit *m1 = (struct commit *) merges.objects[i].item;
1120
1121                 contains_another = 0;
1122                 for (j = 0; j < merges.nr; j++) {
1123                         struct commit *m2 = (struct commit *) merges.objects[j].item;
1124                         if (i != j && in_merge_bases(m2, m1)) {
1125                                 contains_another = 1;
1126                                 break;
1127                         }
1128                 }
1129
1130                 if (!contains_another)
1131                         add_object_array(merges.objects[i].item, NULL, result);
1132         }
1133
1134         object_array_clear(&merges);
1135         return result->nr;
1136 }
1137
1138 static void print_commit(struct commit *commit)
1139 {
1140         struct strbuf sb = STRBUF_INIT;
1141         struct pretty_print_context ctx = {0};
1142         ctx.date_mode.type = DATE_NORMAL;
1143         format_commit_message(commit, " %h: %m %s", &sb, &ctx);
1144         fprintf(stderr, "%s\n", sb.buf);
1145         strbuf_release(&sb);
1146 }
1147
1148 static int is_valid(const struct diff_filespec *dfs)
1149 {
1150         return dfs->mode != 0 && !is_null_oid(&dfs->oid);
1151 }
1152
1153 static int merge_submodule(struct merge_options *opt,
1154                            struct object_id *result, const char *path,
1155                            const struct object_id *base, const struct object_id *a,
1156                            const struct object_id *b)
1157 {
1158         struct commit *commit_base, *commit_a, *commit_b;
1159         int parent_count;
1160         struct object_array merges;
1161
1162         int i;
1163         int search = !opt->call_depth;
1164
1165         /* store a in result in case we fail */
1166         oidcpy(result, a);
1167
1168         /* we can not handle deletion conflicts */
1169         if (is_null_oid(base))
1170                 return 0;
1171         if (is_null_oid(a))
1172                 return 0;
1173         if (is_null_oid(b))
1174                 return 0;
1175
1176         if (add_submodule_odb(path)) {
1177                 output(opt, 1, _("Failed to merge submodule %s (not checked out)"), path);
1178                 return 0;
1179         }
1180
1181         if (!(commit_base = lookup_commit_reference(opt->repo, base)) ||
1182             !(commit_a = lookup_commit_reference(opt->repo, a)) ||
1183             !(commit_b = lookup_commit_reference(opt->repo, b))) {
1184                 output(opt, 1, _("Failed to merge submodule %s (commits not present)"), path);
1185                 return 0;
1186         }
1187
1188         /* check whether both changes are forward */
1189         if (!in_merge_bases(commit_base, commit_a) ||
1190             !in_merge_bases(commit_base, commit_b)) {
1191                 output(opt, 1, _("Failed to merge submodule %s (commits don't follow merge-base)"), path);
1192                 return 0;
1193         }
1194
1195         /* Case #1: a is contained in b or vice versa */
1196         if (in_merge_bases(commit_a, commit_b)) {
1197                 oidcpy(result, b);
1198                 if (show(opt, 3)) {
1199                         output(opt, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1200                         output_commit_title(opt, commit_b);
1201                 } else if (show(opt, 2))
1202                         output(opt, 2, _("Fast-forwarding submodule %s"), path);
1203                 else
1204                         ; /* no output */
1205
1206                 return 1;
1207         }
1208         if (in_merge_bases(commit_b, commit_a)) {
1209                 oidcpy(result, a);
1210                 if (show(opt, 3)) {
1211                         output(opt, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1212                         output_commit_title(opt, commit_a);
1213                 } else if (show(opt, 2))
1214                         output(opt, 2, _("Fast-forwarding submodule %s"), path);
1215                 else
1216                         ; /* no output */
1217
1218                 return 1;
1219         }
1220
1221         /*
1222          * Case #2: There are one or more merges that contain a and b in
1223          * the submodule. If there is only one, then present it as a
1224          * suggestion to the user, but leave it marked unmerged so the
1225          * user needs to confirm the resolution.
1226          */
1227
1228         /* Skip the search if makes no sense to the calling context.  */
1229         if (!search)
1230                 return 0;
1231
1232         /* find commit which merges them */
1233         parent_count = find_first_merges(opt->repo, &merges, path,
1234                                          commit_a, commit_b);
1235         switch (parent_count) {
1236         case 0:
1237                 output(opt, 1, _("Failed to merge submodule %s (merge following commits not found)"), path);
1238                 break;
1239
1240         case 1:
1241                 output(opt, 1, _("Failed to merge submodule %s (not fast-forward)"), path);
1242                 output(opt, 2, _("Found a possible merge resolution for the submodule:\n"));
1243                 print_commit((struct commit *) merges.objects[0].item);
1244                 output(opt, 2, _(
1245                        "If this is correct simply add it to the index "
1246                        "for example\n"
1247                        "by using:\n\n"
1248                        "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
1249                        "which will accept this suggestion.\n"),
1250                        oid_to_hex(&merges.objects[0].item->oid), path);
1251                 break;
1252
1253         default:
1254                 output(opt, 1, _("Failed to merge submodule %s (multiple merges found)"), path);
1255                 for (i = 0; i < merges.nr; i++)
1256                         print_commit((struct commit *) merges.objects[i].item);
1257         }
1258
1259         object_array_clear(&merges);
1260         return 0;
1261 }
1262
1263 static int merge_mode_and_contents(struct merge_options *opt,
1264                                    const struct diff_filespec *o,
1265                                    const struct diff_filespec *a,
1266                                    const struct diff_filespec *b,
1267                                    const char *filename,
1268                                    const char *branch1,
1269                                    const char *branch2,
1270                                    const int extra_marker_size,
1271                                    struct merge_file_info *result)
1272 {
1273         if (opt->branch1 != branch1) {
1274                 /*
1275                  * It's weird getting a reverse merge with HEAD on the bottom
1276                  * side of the conflict markers and the other branch on the
1277                  * top.  Fix that.
1278                  */
1279                 return merge_mode_and_contents(opt, o, b, a,
1280                                                filename,
1281                                                branch2, branch1,
1282                                                extra_marker_size, result);
1283         }
1284
1285         result->merge = 0;
1286         result->clean = 1;
1287
1288         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
1289                 result->clean = 0;
1290                 if (S_ISREG(a->mode)) {
1291                         result->blob.mode = a->mode;
1292                         oidcpy(&result->blob.oid, &a->oid);
1293                 } else {
1294                         result->blob.mode = b->mode;
1295                         oidcpy(&result->blob.oid, &b->oid);
1296                 }
1297         } else {
1298                 if (!oid_eq(&a->oid, &o->oid) && !oid_eq(&b->oid, &o->oid))
1299                         result->merge = 1;
1300
1301                 /*
1302                  * Merge modes
1303                  */
1304                 if (a->mode == b->mode || a->mode == o->mode)
1305                         result->blob.mode = b->mode;
1306                 else {
1307                         result->blob.mode = a->mode;
1308                         if (b->mode != o->mode) {
1309                                 result->clean = 0;
1310                                 result->merge = 1;
1311                         }
1312                 }
1313
1314                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &o->oid))
1315                         oidcpy(&result->blob.oid, &b->oid);
1316                 else if (oid_eq(&b->oid, &o->oid))
1317                         oidcpy(&result->blob.oid, &a->oid);
1318                 else if (S_ISREG(a->mode)) {
1319                         mmbuffer_t result_buf;
1320                         int ret = 0, merge_status;
1321
1322                         merge_status = merge_3way(opt, &result_buf, o, a, b,
1323                                                   branch1, branch2,
1324                                                   extra_marker_size);
1325
1326                         if ((merge_status < 0) || !result_buf.ptr)
1327                                 ret = err(opt, _("Failed to execute internal merge"));
1328
1329                         if (!ret &&
1330                             write_object_file(result_buf.ptr, result_buf.size,
1331                                               blob_type, &result->blob.oid))
1332                                 ret = err(opt, _("Unable to add %s to database"),
1333                                           a->path);
1334
1335                         free(result_buf.ptr);
1336                         if (ret)
1337                                 return ret;
1338                         result->clean = (merge_status == 0);
1339                 } else if (S_ISGITLINK(a->mode)) {
1340                         result->clean = merge_submodule(opt, &result->blob.oid,
1341                                                         o->path,
1342                                                         &o->oid,
1343                                                         &a->oid,
1344                                                         &b->oid);
1345                 } else if (S_ISLNK(a->mode)) {
1346                         switch (opt->recursive_variant) {
1347                         case MERGE_RECURSIVE_NORMAL:
1348                                 oidcpy(&result->blob.oid, &a->oid);
1349                                 if (!oid_eq(&a->oid, &b->oid))
1350                                         result->clean = 0;
1351                                 break;
1352                         case MERGE_RECURSIVE_OURS:
1353                                 oidcpy(&result->blob.oid, &a->oid);
1354                                 break;
1355                         case MERGE_RECURSIVE_THEIRS:
1356                                 oidcpy(&result->blob.oid, &b->oid);
1357                                 break;
1358                         }
1359                 } else
1360                         BUG("unsupported object type in the tree");
1361         }
1362
1363         if (result->merge)
1364                 output(opt, 2, _("Auto-merging %s"), filename);
1365
1366         return 0;
1367 }
1368
1369 static int handle_rename_via_dir(struct merge_options *opt,
1370                                  struct rename_conflict_info *ci)
1371 {
1372         /*
1373          * Handle file adds that need to be renamed due to directory rename
1374          * detection.  This differs from handle_rename_normal, because
1375          * there is no content merge to do; just move the file into the
1376          * desired final location.
1377          */
1378         const struct rename *ren = ci->ren1;
1379         const struct diff_filespec *dest = ren->pair->two;
1380         char *file_path = dest->path;
1381         int mark_conflicted = (opt->detect_directory_renames == 1);
1382         assert(ren->dir_rename_original_dest);
1383
1384         if (!opt->call_depth && would_lose_untracked(opt, dest->path)) {
1385                 mark_conflicted = 1;
1386                 file_path = unique_path(opt, dest->path, ren->branch);
1387                 output(opt, 1, _("Error: Refusing to lose untracked file at %s; "
1388                                  "writing to %s instead."),
1389                        dest->path, file_path);
1390         }
1391
1392         if (mark_conflicted) {
1393                 /*
1394                  * Write the file in worktree at file_path.  In the index,
1395                  * only record the file at dest->path in the appropriate
1396                  * higher stage.
1397                  */
1398                 if (update_file(opt, 0, dest, file_path))
1399                         return -1;
1400                 if (file_path != dest->path)
1401                         free(file_path);
1402                 if (update_stages(opt, dest->path, NULL,
1403                                   ren->branch == opt->branch1 ? dest : NULL,
1404                                   ren->branch == opt->branch1 ? NULL : dest))
1405                         return -1;
1406                 return 0; /* not clean, but conflicted */
1407         } else {
1408                 /* Update dest->path both in index and in worktree */
1409                 if (update_file(opt, 1, dest, dest->path))
1410                         return -1;
1411                 return 1; /* clean */
1412         }
1413 }
1414
1415 static int handle_change_delete(struct merge_options *opt,
1416                                 const char *path, const char *old_path,
1417                                 const struct diff_filespec *o,
1418                                 const struct diff_filespec *changed,
1419                                 const char *change_branch,
1420                                 const char *delete_branch,
1421                                 const char *change, const char *change_past)
1422 {
1423         char *alt_path = NULL;
1424         const char *update_path = path;
1425         int ret = 0;
1426
1427         if (dir_in_way(opt->repo->index, path, !opt->call_depth, 0) ||
1428             (!opt->call_depth && would_lose_untracked(opt, path))) {
1429                 update_path = alt_path = unique_path(opt, path, change_branch);
1430         }
1431
1432         if (opt->call_depth) {
1433                 /*
1434                  * We cannot arbitrarily accept either a_sha or b_sha as
1435                  * correct; since there is no true "middle point" between
1436                  * them, simply reuse the base version for virtual merge base.
1437                  */
1438                 ret = remove_file_from_index(opt->repo->index, path);
1439                 if (!ret)
1440                         ret = update_file(opt, 0, o, update_path);
1441         } else {
1442                 /*
1443                  * Despite the four nearly duplicate messages and argument
1444                  * lists below and the ugliness of the nested if-statements,
1445                  * having complete messages makes the job easier for
1446                  * translators.
1447                  *
1448                  * The slight variance among the cases is due to the fact
1449                  * that:
1450                  *   1) directory/file conflicts (in effect if
1451                  *      !alt_path) could cause us to need to write the
1452                  *      file to a different path.
1453                  *   2) renames (in effect if !old_path) could mean that
1454                  *      there are two names for the path that the user
1455                  *      may know the file by.
1456                  */
1457                 if (!alt_path) {
1458                         if (!old_path) {
1459                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1460                                        "and %s in %s. Version %s of %s left in tree."),
1461                                        change, path, delete_branch, change_past,
1462                                        change_branch, change_branch, path);
1463                         } else {
1464                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1465                                        "and %s to %s in %s. Version %s of %s left in tree."),
1466                                        change, old_path, delete_branch, change_past, path,
1467                                        change_branch, change_branch, path);
1468                         }
1469                 } else {
1470                         if (!old_path) {
1471                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1472                                        "and %s in %s. Version %s of %s left in tree at %s."),
1473                                        change, path, delete_branch, change_past,
1474                                        change_branch, change_branch, path, alt_path);
1475                         } else {
1476                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1477                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1478                                        change, old_path, delete_branch, change_past, path,
1479                                        change_branch, change_branch, path, alt_path);
1480                         }
1481                 }
1482                 /*
1483                  * No need to call update_file() on path when change_branch ==
1484                  * opt->branch1 && !alt_path, since that would needlessly touch
1485                  * path.  We could call update_file_flags() with update_cache=0
1486                  * and update_wd=0, but that's a no-op.
1487                  */
1488                 if (change_branch != opt->branch1 || alt_path)
1489                         ret = update_file(opt, 0, changed, update_path);
1490         }
1491         free(alt_path);
1492
1493         return ret;
1494 }
1495
1496 static int handle_rename_delete(struct merge_options *opt,
1497                                 struct rename_conflict_info *ci)
1498 {
1499         const struct rename *ren = ci->ren1;
1500         const struct diff_filespec *orig = ren->pair->one;
1501         const struct diff_filespec *dest = ren->pair->two;
1502         const char *rename_branch = ren->branch;
1503         const char *delete_branch = (opt->branch1 == ren->branch ?
1504                                      opt->branch2 : opt->branch1);
1505
1506         if (handle_change_delete(opt,
1507                                  opt->call_depth ? orig->path : dest->path,
1508                                  opt->call_depth ? NULL : orig->path,
1509                                  orig, dest,
1510                                  rename_branch, delete_branch,
1511                                  _("rename"), _("renamed")))
1512                 return -1;
1513
1514         if (opt->call_depth)
1515                 return remove_file_from_index(opt->repo->index, dest->path);
1516         else
1517                 return update_stages(opt, dest->path, NULL,
1518                                      rename_branch == opt->branch1 ? dest : NULL,
1519                                      rename_branch == opt->branch1 ? NULL : dest);
1520 }
1521
1522 static int handle_file_collision(struct merge_options *opt,
1523                                  const char *collide_path,
1524                                  const char *prev_path1,
1525                                  const char *prev_path2,
1526                                  const char *branch1, const char *branch2,
1527                                  struct diff_filespec *a,
1528                                  struct diff_filespec *b)
1529 {
1530         struct merge_file_info mfi;
1531         struct diff_filespec null;
1532         char *alt_path = NULL;
1533         const char *update_path = collide_path;
1534
1535         /*
1536          * It's easiest to get the correct things into stage 2 and 3, and
1537          * to make sure that the content merge puts HEAD before the other
1538          * branch if we just ensure that branch1 == opt->branch1.  So, simply
1539          * flip arguments around if we don't have that.
1540          */
1541         if (branch1 != opt->branch1) {
1542                 return handle_file_collision(opt, collide_path,
1543                                              prev_path2, prev_path1,
1544                                              branch2, branch1,
1545                                              b, a);
1546         }
1547
1548         /*
1549          * In the recursive case, we just opt to undo renames
1550          */
1551         if (opt->call_depth && (prev_path1 || prev_path2)) {
1552                 /* Put first file (a->oid, a->mode) in its original spot */
1553                 if (prev_path1) {
1554                         if (update_file(opt, 1, a, prev_path1))
1555                                 return -1;
1556                 } else {
1557                         if (update_file(opt, 1, a, collide_path))
1558                                 return -1;
1559                 }
1560
1561                 /* Put second file (b->oid, b->mode) in its original spot */
1562                 if (prev_path2) {
1563                         if (update_file(opt, 1, b, prev_path2))
1564                                 return -1;
1565                 } else {
1566                         if (update_file(opt, 1, b, collide_path))
1567                                 return -1;
1568                 }
1569
1570                 /* Don't leave something at collision path if unrenaming both */
1571                 if (prev_path1 && prev_path2)
1572                         remove_file(opt, 1, collide_path, 0);
1573
1574                 return 0;
1575         }
1576
1577         /* Remove rename sources if rename/add or rename/rename(2to1) */
1578         if (prev_path1)
1579                 remove_file(opt, 1, prev_path1,
1580                             opt->call_depth || would_lose_untracked(opt, prev_path1));
1581         if (prev_path2)
1582                 remove_file(opt, 1, prev_path2,
1583                             opt->call_depth || would_lose_untracked(opt, prev_path2));
1584
1585         /*
1586          * Remove the collision path, if it wouldn't cause dirty contents
1587          * or an untracked file to get lost.  We'll either overwrite with
1588          * merged contents, or just write out to differently named files.
1589          */
1590         if (was_dirty(opt, collide_path)) {
1591                 output(opt, 1, _("Refusing to lose dirty file at %s"),
1592                        collide_path);
1593                 update_path = alt_path = unique_path(opt, collide_path, "merged");
1594         } else if (would_lose_untracked(opt, collide_path)) {
1595                 /*
1596                  * Only way we get here is if both renames were from
1597                  * a directory rename AND user had an untracked file
1598                  * at the location where both files end up after the
1599                  * two directory renames.  See testcase 10d of t6043.
1600                  */
1601                 output(opt, 1, _("Refusing to lose untracked file at "
1602                                "%s, even though it's in the way."),
1603                        collide_path);
1604                 update_path = alt_path = unique_path(opt, collide_path, "merged");
1605         } else {
1606                 /*
1607                  * FIXME: It's possible that the two files are identical
1608                  * and that the current working copy happens to match, in
1609                  * which case we are unnecessarily touching the working
1610                  * tree file.  It's not a likely enough scenario that I
1611                  * want to code up the checks for it and a better fix is
1612                  * available if we restructure how unpack_trees() and
1613                  * merge-recursive interoperate anyway, so punting for
1614                  * now...
1615                  */
1616                 remove_file(opt, 0, collide_path, 0);
1617         }
1618
1619         /* Store things in diff_filespecs for functions that need it */
1620         null.path = (char *)collide_path;
1621         oidcpy(&null.oid, &null_oid);
1622         null.mode = 0;
1623
1624         if (merge_mode_and_contents(opt, &null, a, b, collide_path,
1625                                     branch1, branch2, opt->call_depth * 2, &mfi))
1626                 return -1;
1627         mfi.clean &= !alt_path;
1628         if (update_file(opt, mfi.clean, &mfi.blob, update_path))
1629                 return -1;
1630         if (!mfi.clean && !opt->call_depth &&
1631             update_stages(opt, collide_path, NULL, a, b))
1632                 return -1;
1633         free(alt_path);
1634         /*
1635          * FIXME: If both a & b both started with conflicts (only possible
1636          * if they came from a rename/rename(2to1)), but had IDENTICAL
1637          * contents including those conflicts, then in the next line we claim
1638          * it was clean.  If someone cares about this case, we should have the
1639          * caller notify us if we started with conflicts.
1640          */
1641         return mfi.clean;
1642 }
1643
1644 static int handle_rename_add(struct merge_options *opt,
1645                              struct rename_conflict_info *ci)
1646 {
1647         /* a was renamed to c, and a separate c was added. */
1648         struct diff_filespec *a = ci->ren1->pair->one;
1649         struct diff_filespec *c = ci->ren1->pair->two;
1650         char *path = c->path;
1651         char *prev_path_desc;
1652         struct merge_file_info mfi;
1653
1654         const char *rename_branch = ci->ren1->branch;
1655         const char *add_branch = (opt->branch1 == rename_branch ?
1656                                   opt->branch2 : opt->branch1);
1657         int other_stage = (ci->ren1->branch == opt->branch1 ? 3 : 2);
1658
1659         output(opt, 1, _("CONFLICT (rename/add): "
1660                "Rename %s->%s in %s.  Added %s in %s"),
1661                a->path, c->path, rename_branch,
1662                c->path, add_branch);
1663
1664         prev_path_desc = xstrfmt("version of %s from %s", path, a->path);
1665         ci->ren1->src_entry->stages[other_stage].path = a->path;
1666         if (merge_mode_and_contents(opt, a, c,
1667                                     &ci->ren1->src_entry->stages[other_stage],
1668                                     prev_path_desc,
1669                                     opt->branch1, opt->branch2,
1670                                     1 + opt->call_depth * 2, &mfi))
1671                 return -1;
1672         free(prev_path_desc);
1673
1674         ci->ren1->dst_entry->stages[other_stage].path = mfi.blob.path = c->path;
1675         return handle_file_collision(opt,
1676                                      c->path, a->path, NULL,
1677                                      rename_branch, add_branch,
1678                                      &mfi.blob,
1679                                      &ci->ren1->dst_entry->stages[other_stage]);
1680 }
1681
1682 static char *find_path_for_conflict(struct merge_options *opt,
1683                                     const char *path,
1684                                     const char *branch1,
1685                                     const char *branch2)
1686 {
1687         char *new_path = NULL;
1688         if (dir_in_way(opt->repo->index, path, !opt->call_depth, 0)) {
1689                 new_path = unique_path(opt, path, branch1);
1690                 output(opt, 1, _("%s is a directory in %s adding "
1691                                "as %s instead"),
1692                        path, branch2, new_path);
1693         } else if (would_lose_untracked(opt, path)) {
1694                 new_path = unique_path(opt, path, branch1);
1695                 output(opt, 1, _("Refusing to lose untracked file"
1696                                " at %s; adding as %s instead"),
1697                        path, new_path);
1698         }
1699
1700         return new_path;
1701 }
1702
1703 static int handle_rename_rename_1to2(struct merge_options *opt,
1704                                      struct rename_conflict_info *ci)
1705 {
1706         /* One file was renamed in both branches, but to different names. */
1707         struct merge_file_info mfi;
1708         struct diff_filespec *add;
1709         struct diff_filespec *o = ci->ren1->pair->one;
1710         struct diff_filespec *a = ci->ren1->pair->two;
1711         struct diff_filespec *b = ci->ren2->pair->two;
1712         char *path_desc;
1713
1714         output(opt, 1, _("CONFLICT (rename/rename): "
1715                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1716                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1717                o->path, a->path, ci->ren1->branch,
1718                o->path, b->path, ci->ren2->branch,
1719                opt->call_depth ? _(" (left unresolved)") : "");
1720
1721         path_desc = xstrfmt("%s and %s, both renamed from %s",
1722                             a->path, b->path, o->path);
1723         if (merge_mode_and_contents(opt, o, a, b, path_desc,
1724                                     ci->ren1->branch, ci->ren2->branch,
1725                                     opt->call_depth * 2, &mfi))
1726                 return -1;
1727         free(path_desc);
1728
1729         if (opt->call_depth) {
1730                 /*
1731                  * FIXME: For rename/add-source conflicts (if we could detect
1732                  * such), this is wrong.  We should instead find a unique
1733                  * pathname and then either rename the add-source file to that
1734                  * unique path, or use that unique path instead of src here.
1735                  */
1736                 if (update_file(opt, 0, &mfi.blob, o->path))
1737                         return -1;
1738
1739                 /*
1740                  * Above, we put the merged content at the merge-base's
1741                  * path.  Now we usually need to delete both a->path and
1742                  * b->path.  However, the rename on each side of the merge
1743                  * could also be involved in a rename/add conflict.  In
1744                  * such cases, we should keep the added file around,
1745                  * resolving the conflict at that path in its favor.
1746                  */
1747                 add = &ci->ren1->dst_entry->stages[2 ^ 1];
1748                 if (is_valid(add)) {
1749                         if (update_file(opt, 0, add, a->path))
1750                                 return -1;
1751                 }
1752                 else
1753                         remove_file_from_index(opt->repo->index, a->path);
1754                 add = &ci->ren2->dst_entry->stages[3 ^ 1];
1755                 if (is_valid(add)) {
1756                         if (update_file(opt, 0, add, b->path))
1757                                 return -1;
1758                 }
1759                 else
1760                         remove_file_from_index(opt->repo->index, b->path);
1761         } else {
1762                 /*
1763                  * For each destination path, we need to see if there is a
1764                  * rename/add collision.  If not, we can write the file out
1765                  * to the specified location.
1766                  */
1767                 add = &ci->ren1->dst_entry->stages[2 ^ 1];
1768                 if (is_valid(add)) {
1769                         add->path = mfi.blob.path = a->path;
1770                         if (handle_file_collision(opt, a->path,
1771                                                   NULL, NULL,
1772                                                   ci->ren1->branch,
1773                                                   ci->ren2->branch,
1774                                                   &mfi.blob, add) < 0)
1775                                 return -1;
1776                 } else {
1777                         char *new_path = find_path_for_conflict(opt, a->path,
1778                                                                 ci->ren1->branch,
1779                                                                 ci->ren2->branch);
1780                         if (update_file(opt, 0, &mfi.blob,
1781                                         new_path ? new_path : a->path))
1782                                 return -1;
1783                         free(new_path);
1784                         if (update_stages(opt, a->path, NULL, a, NULL))
1785                                 return -1;
1786                 }
1787
1788                 add = &ci->ren2->dst_entry->stages[3 ^ 1];
1789                 if (is_valid(add)) {
1790                         add->path = mfi.blob.path = b->path;
1791                         if (handle_file_collision(opt, b->path,
1792                                                   NULL, NULL,
1793                                                   ci->ren1->branch,
1794                                                   ci->ren2->branch,
1795                                                   add, &mfi.blob) < 0)
1796                                 return -1;
1797                 } else {
1798                         char *new_path = find_path_for_conflict(opt, b->path,
1799                                                                 ci->ren2->branch,
1800                                                                 ci->ren1->branch);
1801                         if (update_file(opt, 0, &mfi.blob,
1802                                         new_path ? new_path : b->path))
1803                                 return -1;
1804                         free(new_path);
1805                         if (update_stages(opt, b->path, NULL, NULL, b))
1806                                 return -1;
1807                 }
1808         }
1809
1810         return 0;
1811 }
1812
1813 static int handle_rename_rename_2to1(struct merge_options *opt,
1814                                      struct rename_conflict_info *ci)
1815 {
1816         /* Two files, a & b, were renamed to the same thing, c. */
1817         struct diff_filespec *a = ci->ren1->pair->one;
1818         struct diff_filespec *b = ci->ren2->pair->one;
1819         struct diff_filespec *c1 = ci->ren1->pair->two;
1820         struct diff_filespec *c2 = ci->ren2->pair->two;
1821         char *path = c1->path; /* == c2->path */
1822         char *path_side_1_desc;
1823         char *path_side_2_desc;
1824         struct merge_file_info mfi_c1;
1825         struct merge_file_info mfi_c2;
1826         int ostage1, ostage2;
1827
1828         output(opt, 1, _("CONFLICT (rename/rename): "
1829                "Rename %s->%s in %s. "
1830                "Rename %s->%s in %s"),
1831                a->path, c1->path, ci->ren1->branch,
1832                b->path, c2->path, ci->ren2->branch);
1833
1834         path_side_1_desc = xstrfmt("version of %s from %s", path, a->path);
1835         path_side_2_desc = xstrfmt("version of %s from %s", path, b->path);
1836         ostage1 = ci->ren1->branch == opt->branch1 ? 3 : 2;
1837         ostage2 = ostage1 ^ 1;
1838         ci->ren1->src_entry->stages[ostage1].path = a->path;
1839         ci->ren2->src_entry->stages[ostage2].path = b->path;
1840         if (merge_mode_and_contents(opt, a, c1,
1841                                     &ci->ren1->src_entry->stages[ostage1],
1842                                     path_side_1_desc,
1843                                     opt->branch1, opt->branch2,
1844                                     1 + opt->call_depth * 2, &mfi_c1) ||
1845             merge_mode_and_contents(opt, b,
1846                                     &ci->ren2->src_entry->stages[ostage2],
1847                                     c2, path_side_2_desc,
1848                                     opt->branch1, opt->branch2,
1849                                     1 + opt->call_depth * 2, &mfi_c2))
1850                 return -1;
1851         free(path_side_1_desc);
1852         free(path_side_2_desc);
1853         mfi_c1.blob.path = path;
1854         mfi_c2.blob.path = path;
1855
1856         return handle_file_collision(opt, path, a->path, b->path,
1857                                      ci->ren1->branch, ci->ren2->branch,
1858                                      &mfi_c1.blob, &mfi_c2.blob);
1859 }
1860
1861 /*
1862  * Get the diff_filepairs changed between o_tree and tree.
1863  */
1864 static struct diff_queue_struct *get_diffpairs(struct merge_options *opt,
1865                                                struct tree *o_tree,
1866                                                struct tree *tree)
1867 {
1868         struct diff_queue_struct *ret;
1869         struct diff_options opts;
1870
1871         repo_diff_setup(opt->repo, &opts);
1872         opts.flags.recursive = 1;
1873         opts.flags.rename_empty = 0;
1874         opts.detect_rename = merge_detect_rename(opt);
1875         /*
1876          * We do not have logic to handle the detection of copies.  In
1877          * fact, it may not even make sense to add such logic: would we
1878          * really want a change to a base file to be propagated through
1879          * multiple other files by a merge?
1880          */
1881         if (opts.detect_rename > DIFF_DETECT_RENAME)
1882                 opts.detect_rename = DIFF_DETECT_RENAME;
1883         opts.rename_limit = opt->merge_rename_limit >= 0 ? opt->merge_rename_limit :
1884                             opt->diff_rename_limit >= 0 ? opt->diff_rename_limit :
1885                             1000;
1886         opts.rename_score = opt->rename_score;
1887         opts.show_rename_progress = opt->show_rename_progress;
1888         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1889         diff_setup_done(&opts);
1890         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1891         diffcore_std(&opts);
1892         if (opts.needed_rename_limit > opt->needed_rename_limit)
1893                 opt->needed_rename_limit = opts.needed_rename_limit;
1894
1895         ret = xmalloc(sizeof(*ret));
1896         *ret = diff_queued_diff;
1897
1898         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1899         diff_queued_diff.nr = 0;
1900         diff_queued_diff.queue = NULL;
1901         diff_flush(&opts);
1902         return ret;
1903 }
1904
1905 static int tree_has_path(struct repository *r, struct tree *tree,
1906                          const char *path)
1907 {
1908         struct object_id hashy;
1909         unsigned short mode_o;
1910
1911         return !get_tree_entry(r,
1912                                &tree->object.oid, path,
1913                                &hashy, &mode_o);
1914 }
1915
1916 /*
1917  * Return a new string that replaces the beginning portion (which matches
1918  * entry->dir), with entry->new_dir.  In perl-speak:
1919  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1920  * NOTE:
1921  *   Caller must ensure that old_path starts with entry->dir + '/'.
1922  */
1923 static char *apply_dir_rename(struct dir_rename_entry *entry,
1924                               const char *old_path)
1925 {
1926         struct strbuf new_path = STRBUF_INIT;
1927         int oldlen, newlen;
1928
1929         if (entry->non_unique_new_dir)
1930                 return NULL;
1931
1932         oldlen = strlen(entry->dir);
1933         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1934         strbuf_grow(&new_path, newlen);
1935         strbuf_addbuf(&new_path, &entry->new_dir);
1936         strbuf_addstr(&new_path, &old_path[oldlen]);
1937
1938         return strbuf_detach(&new_path, NULL);
1939 }
1940
1941 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1942                                     char **old_dir, char **new_dir)
1943 {
1944         char *end_of_old, *end_of_new;
1945         int old_len, new_len;
1946
1947         *old_dir = NULL;
1948         *new_dir = NULL;
1949
1950         /*
1951          * For
1952          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1953          * the "e/foo.c" part is the same, we just want to know that
1954          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1955          * so, for this example, this function returns "a/b/c/d" in
1956          * *old_dir and "a/b/some/thing/else" in *new_dir.
1957          *
1958          * Also, if the basename of the file changed, we don't care.  We
1959          * want to know which portion of the directory, if any, changed.
1960          */
1961         end_of_old = strrchr(old_path, '/');
1962         end_of_new = strrchr(new_path, '/');
1963
1964         if (end_of_old == NULL || end_of_new == NULL)
1965                 return;
1966         while (*--end_of_new == *--end_of_old &&
1967                end_of_old != old_path &&
1968                end_of_new != new_path)
1969                 ; /* Do nothing; all in the while loop */
1970         /*
1971          * We've found the first non-matching character in the directory
1972          * paths.  That means the current directory we were comparing
1973          * represents the rename.  Move end_of_old and end_of_new back
1974          * to the full directory name.
1975          */
1976         if (*end_of_old == '/')
1977                 end_of_old++;
1978         if (*end_of_old != '/')
1979                 end_of_new++;
1980         end_of_old = strchr(end_of_old, '/');
1981         end_of_new = strchr(end_of_new, '/');
1982
1983         /*
1984          * It may have been the case that old_path and new_path were the same
1985          * directory all along.  Don't claim a rename if they're the same.
1986          */
1987         old_len = end_of_old - old_path;
1988         new_len = end_of_new - new_path;
1989
1990         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1991                 *old_dir = xstrndup(old_path, old_len);
1992                 *new_dir = xstrndup(new_path, new_len);
1993         }
1994 }
1995
1996 static void remove_hashmap_entries(struct hashmap *dir_renames,
1997                                    struct string_list *items_to_remove)
1998 {
1999         int i;
2000         struct dir_rename_entry *entry;
2001
2002         for (i = 0; i < items_to_remove->nr; i++) {
2003                 entry = items_to_remove->items[i].util;
2004                 hashmap_remove(dir_renames, entry, NULL);
2005         }
2006         string_list_clear(items_to_remove, 0);
2007 }
2008
2009 /*
2010  * See if there is a directory rename for path, and if there are any file
2011  * level conflicts for the renamed location.  If there is a rename and
2012  * there are no conflicts, return the new name.  Otherwise, return NULL.
2013  */
2014 static char *handle_path_level_conflicts(struct merge_options *opt,
2015                                          const char *path,
2016                                          struct dir_rename_entry *entry,
2017                                          struct hashmap *collisions,
2018                                          struct tree *tree)
2019 {
2020         char *new_path = NULL;
2021         struct collision_entry *collision_ent;
2022         int clean = 1;
2023         struct strbuf collision_paths = STRBUF_INIT;
2024
2025         /*
2026          * entry has the mapping of old directory name to new directory name
2027          * that we want to apply to path.
2028          */
2029         new_path = apply_dir_rename(entry, path);
2030
2031         if (!new_path) {
2032                 /* This should only happen when entry->non_unique_new_dir set */
2033                 if (!entry->non_unique_new_dir)
2034                         BUG("entry->non_unqiue_dir not set and !new_path");
2035                 output(opt, 1, _("CONFLICT (directory rename split): "
2036                                "Unclear where to place %s because directory "
2037                                "%s was renamed to multiple other directories, "
2038                                "with no destination getting a majority of the "
2039                                "files."),
2040                        path, entry->dir);
2041                 clean = 0;
2042                 return NULL;
2043         }
2044
2045         /*
2046          * The caller needs to have ensured that it has pre-populated
2047          * collisions with all paths that map to new_path.  Do a quick check
2048          * to ensure that's the case.
2049          */
2050         collision_ent = collision_find_entry(collisions, new_path);
2051         if (collision_ent == NULL)
2052                 BUG("collision_ent is NULL");
2053
2054         /*
2055          * Check for one-sided add/add/.../add conflicts, i.e.
2056          * where implicit renames from the other side doing
2057          * directory rename(s) can affect this side of history
2058          * to put multiple paths into the same location.  Warn
2059          * and bail on directory renames for such paths.
2060          */
2061         if (collision_ent->reported_already) {
2062                 clean = 0;
2063         } else if (tree_has_path(opt->repo, tree, new_path)) {
2064                 collision_ent->reported_already = 1;
2065                 strbuf_add_separated_string_list(&collision_paths, ", ",
2066                                                  &collision_ent->source_files);
2067                 output(opt, 1, _("CONFLICT (implicit dir rename): Existing "
2068                                "file/dir at %s in the way of implicit "
2069                                "directory rename(s) putting the following "
2070                                "path(s) there: %s."),
2071                        new_path, collision_paths.buf);
2072                 clean = 0;
2073         } else if (collision_ent->source_files.nr > 1) {
2074                 collision_ent->reported_already = 1;
2075                 strbuf_add_separated_string_list(&collision_paths, ", ",
2076                                                  &collision_ent->source_files);
2077                 output(opt, 1, _("CONFLICT (implicit dir rename): Cannot map "
2078                                "more than one path to %s; implicit directory "
2079                                "renames tried to put these paths there: %s"),
2080                        new_path, collision_paths.buf);
2081                 clean = 0;
2082         }
2083
2084         /* Free memory we no longer need */
2085         strbuf_release(&collision_paths);
2086         if (!clean && new_path) {
2087                 free(new_path);
2088                 return NULL;
2089         }
2090
2091         return new_path;
2092 }
2093
2094 /*
2095  * There are a couple things we want to do at the directory level:
2096  *   1. Check for both sides renaming to the same thing, in order to avoid
2097  *      implicit renaming of files that should be left in place.  (See
2098  *      testcase 6b in t6043 for details.)
2099  *   2. Prune directory renames if there are still files left in the
2100  *      the original directory.  These represent a partial directory rename,
2101  *      i.e. a rename where only some of the files within the directory
2102  *      were renamed elsewhere.  (Technically, this could be done earlier
2103  *      in get_directory_renames(), except that would prevent us from
2104  *      doing the previous check and thus failing testcase 6b.)
2105  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2106  *      In the future, we could potentially record this info as well and
2107  *      omit reporting rename/rename(1to2) conflicts for each path within
2108  *      the affected directories, thus cleaning up the merge output.
2109  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2110  *         directory level, because merging directories is fine.  If it
2111  *         causes conflicts for files within those merged directories, then
2112  *         that should be detected at the individual path level.
2113  */
2114 static void handle_directory_level_conflicts(struct merge_options *opt,
2115                                              struct hashmap *dir_re_head,
2116                                              struct tree *head,
2117                                              struct hashmap *dir_re_merge,
2118                                              struct tree *merge)
2119 {
2120         struct hashmap_iter iter;
2121         struct dir_rename_entry *head_ent;
2122         struct dir_rename_entry *merge_ent;
2123
2124         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2125         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2126
2127         hashmap_iter_init(dir_re_head, &iter);
2128         while ((head_ent = hashmap_iter_next(&iter))) {
2129                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2130                 if (merge_ent &&
2131                     !head_ent->non_unique_new_dir &&
2132                     !merge_ent->non_unique_new_dir &&
2133                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2134                         /* 1. Renamed identically; remove it from both sides */
2135                         string_list_append(&remove_from_head,
2136                                            head_ent->dir)->util = head_ent;
2137                         strbuf_release(&head_ent->new_dir);
2138                         string_list_append(&remove_from_merge,
2139                                            merge_ent->dir)->util = merge_ent;
2140                         strbuf_release(&merge_ent->new_dir);
2141                 } else if (tree_has_path(opt->repo, head, head_ent->dir)) {
2142                         /* 2. This wasn't a directory rename after all */
2143                         string_list_append(&remove_from_head,
2144                                            head_ent->dir)->util = head_ent;
2145                         strbuf_release(&head_ent->new_dir);
2146                 }
2147         }
2148
2149         remove_hashmap_entries(dir_re_head, &remove_from_head);
2150         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2151
2152         hashmap_iter_init(dir_re_merge, &iter);
2153         while ((merge_ent = hashmap_iter_next(&iter))) {
2154                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2155                 if (tree_has_path(opt->repo, merge, merge_ent->dir)) {
2156                         /* 2. This wasn't a directory rename after all */
2157                         string_list_append(&remove_from_merge,
2158                                            merge_ent->dir)->util = merge_ent;
2159                 } else if (head_ent &&
2160                            !head_ent->non_unique_new_dir &&
2161                            !merge_ent->non_unique_new_dir) {
2162                         /* 3. rename/rename(1to2) */
2163                         /*
2164                          * We can assume it's not rename/rename(1to1) because
2165                          * that was case (1), already checked above.  So we
2166                          * know that head_ent->new_dir and merge_ent->new_dir
2167                          * are different strings.
2168                          */
2169                         output(opt, 1, _("CONFLICT (rename/rename): "
2170                                        "Rename directory %s->%s in %s. "
2171                                        "Rename directory %s->%s in %s"),
2172                                head_ent->dir, head_ent->new_dir.buf, opt->branch1,
2173                                head_ent->dir, merge_ent->new_dir.buf, opt->branch2);
2174                         string_list_append(&remove_from_head,
2175                                            head_ent->dir)->util = head_ent;
2176                         strbuf_release(&head_ent->new_dir);
2177                         string_list_append(&remove_from_merge,
2178                                            merge_ent->dir)->util = merge_ent;
2179                         strbuf_release(&merge_ent->new_dir);
2180                 }
2181         }
2182
2183         remove_hashmap_entries(dir_re_head, &remove_from_head);
2184         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2185 }
2186
2187 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs)
2188 {
2189         struct hashmap *dir_renames;
2190         struct hashmap_iter iter;
2191         struct dir_rename_entry *entry;
2192         int i;
2193
2194         /*
2195          * Typically, we think of a directory rename as all files from a
2196          * certain directory being moved to a target directory.  However,
2197          * what if someone first moved two files from the original
2198          * directory in one commit, and then renamed the directory
2199          * somewhere else in a later commit?  At merge time, we just know
2200          * that files from the original directory went to two different
2201          * places, and that the bulk of them ended up in the same place.
2202          * We want each directory rename to represent where the bulk of the
2203          * files from that directory end up; this function exists to find
2204          * where the bulk of the files went.
2205          *
2206          * The first loop below simply iterates through the list of file
2207          * renames, finding out how often each directory rename pair
2208          * possibility occurs.
2209          */
2210         dir_renames = xmalloc(sizeof(*dir_renames));
2211         dir_rename_init(dir_renames);
2212         for (i = 0; i < pairs->nr; ++i) {
2213                 struct string_list_item *item;
2214                 int *count;
2215                 struct diff_filepair *pair = pairs->queue[i];
2216                 char *old_dir, *new_dir;
2217
2218                 /* File not part of directory rename if it wasn't renamed */
2219                 if (pair->status != 'R')
2220                         continue;
2221
2222                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2223                                         &old_dir,        &new_dir);
2224                 if (!old_dir)
2225                         /* Directory didn't change at all; ignore this one. */
2226                         continue;
2227
2228                 entry = dir_rename_find_entry(dir_renames, old_dir);
2229                 if (!entry) {
2230                         entry = xmalloc(sizeof(*entry));
2231                         dir_rename_entry_init(entry, old_dir);
2232                         hashmap_put(dir_renames, entry);
2233                 } else {
2234                         free(old_dir);
2235                 }
2236                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2237                 if (!item) {
2238                         item = string_list_insert(&entry->possible_new_dirs,
2239                                                   new_dir);
2240                         item->util = xcalloc(1, sizeof(int));
2241                 } else {
2242                         free(new_dir);
2243                 }
2244                 count = item->util;
2245                 *count += 1;
2246         }
2247
2248         /*
2249          * For each directory with files moved out of it, we find out which
2250          * target directory received the most files so we can declare it to
2251          * be the "winning" target location for the directory rename.  This
2252          * winner gets recorded in new_dir.  If there is no winner
2253          * (multiple target directories received the same number of files),
2254          * we set non_unique_new_dir.  Once we've determined the winner (or
2255          * that there is no winner), we no longer need possible_new_dirs.
2256          */
2257         hashmap_iter_init(dir_renames, &iter);
2258         while ((entry = hashmap_iter_next(&iter))) {
2259                 int max = 0;
2260                 int bad_max = 0;
2261                 char *best = NULL;
2262
2263                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2264                         int *count = entry->possible_new_dirs.items[i].util;
2265
2266                         if (*count == max)
2267                                 bad_max = max;
2268                         else if (*count > max) {
2269                                 max = *count;
2270                                 best = entry->possible_new_dirs.items[i].string;
2271                         }
2272                 }
2273                 if (bad_max == max)
2274                         entry->non_unique_new_dir = 1;
2275                 else {
2276                         assert(entry->new_dir.len == 0);
2277                         strbuf_addstr(&entry->new_dir, best);
2278                 }
2279                 /*
2280                  * The relevant directory sub-portion of the original full
2281                  * filepaths were xstrndup'ed before inserting into
2282                  * possible_new_dirs, and instead of manually iterating the
2283                  * list and free'ing each, just lie and tell
2284                  * possible_new_dirs that it did the strdup'ing so that it
2285                  * will free them for us.
2286                  */
2287                 entry->possible_new_dirs.strdup_strings = 1;
2288                 string_list_clear(&entry->possible_new_dirs, 1);
2289         }
2290
2291         return dir_renames;
2292 }
2293
2294 static struct dir_rename_entry *check_dir_renamed(const char *path,
2295                                                   struct hashmap *dir_renames)
2296 {
2297         char *temp = xstrdup(path);
2298         char *end;
2299         struct dir_rename_entry *entry = NULL;
2300
2301         while ((end = strrchr(temp, '/'))) {
2302                 *end = '\0';
2303                 entry = dir_rename_find_entry(dir_renames, temp);
2304                 if (entry)
2305                         break;
2306         }
2307         free(temp);
2308         return entry;
2309 }
2310
2311 static void compute_collisions(struct hashmap *collisions,
2312                                struct hashmap *dir_renames,
2313                                struct diff_queue_struct *pairs)
2314 {
2315         int i;
2316
2317         /*
2318          * Multiple files can be mapped to the same path due to directory
2319          * renames done by the other side of history.  Since that other
2320          * side of history could have merged multiple directories into one,
2321          * if our side of history added the same file basename to each of
2322          * those directories, then all N of them would get implicitly
2323          * renamed by the directory rename detection into the same path,
2324          * and we'd get an add/add/.../add conflict, and all those adds
2325          * from *this* side of history.  This is not representable in the
2326          * index, and users aren't going to easily be able to make sense of
2327          * it.  So we need to provide a good warning about what's
2328          * happening, and fall back to no-directory-rename detection
2329          * behavior for those paths.
2330          *
2331          * See testcases 9e and all of section 5 from t6043 for examples.
2332          */
2333         collision_init(collisions);
2334
2335         for (i = 0; i < pairs->nr; ++i) {
2336                 struct dir_rename_entry *dir_rename_ent;
2337                 struct collision_entry *collision_ent;
2338                 char *new_path;
2339                 struct diff_filepair *pair = pairs->queue[i];
2340
2341                 if (pair->status != 'A' && pair->status != 'R')
2342                         continue;
2343                 dir_rename_ent = check_dir_renamed(pair->two->path,
2344                                                    dir_renames);
2345                 if (!dir_rename_ent)
2346                         continue;
2347
2348                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2349                 if (!new_path)
2350                         /*
2351                          * dir_rename_ent->non_unique_new_path is true, which
2352                          * means there is no directory rename for us to use,
2353                          * which means it won't cause us any additional
2354                          * collisions.
2355                          */
2356                         continue;
2357                 collision_ent = collision_find_entry(collisions, new_path);
2358                 if (!collision_ent) {
2359                         collision_ent = xcalloc(1,
2360                                                 sizeof(struct collision_entry));
2361                         hashmap_entry_init(collision_ent, strhash(new_path));
2362                         hashmap_put(collisions, collision_ent);
2363                         collision_ent->target_file = new_path;
2364                 } else {
2365                         free(new_path);
2366                 }
2367                 string_list_insert(&collision_ent->source_files,
2368                                    pair->two->path);
2369         }
2370 }
2371
2372 static char *check_for_directory_rename(struct merge_options *opt,
2373                                         const char *path,
2374                                         struct tree *tree,
2375                                         struct hashmap *dir_renames,
2376                                         struct hashmap *dir_rename_exclusions,
2377                                         struct hashmap *collisions,
2378                                         int *clean_merge)
2379 {
2380         char *new_path = NULL;
2381         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2382         struct dir_rename_entry *oentry = NULL;
2383
2384         if (!entry)
2385                 return new_path;
2386
2387         /*
2388          * This next part is a little weird.  We do not want to do an
2389          * implicit rename into a directory we renamed on our side, because
2390          * that will result in a spurious rename/rename(1to2) conflict.  An
2391          * example:
2392          *   Base commit: dumbdir/afile, otherdir/bfile
2393          *   Side 1:      smrtdir/afile, otherdir/bfile
2394          *   Side 2:      dumbdir/afile, dumbdir/bfile
2395          * Here, while working on Side 1, we could notice that otherdir was
2396          * renamed/merged to dumbdir, and change the diff_filepair for
2397          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2398          * 2 will notice the rename from dumbdir to smrtdir, and do the
2399          * transitive rename to move it from dumbdir/bfile to
2400          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2401          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2402          * the file to end up in smrtdir.  And the way to achieve that is
2403          * to not let Side1 do the rename to dumbdir, since we know that is
2404          * the source of one of our directory renames.
2405          *
2406          * That's why oentry and dir_rename_exclusions is here.
2407          *
2408          * As it turns out, this also prevents N-way transient rename
2409          * confusion; See testcases 9c and 9d of t6043.
2410          */
2411         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2412         if (oentry) {
2413                 output(opt, 1, _("WARNING: Avoiding applying %s -> %s rename "
2414                                "to %s, because %s itself was renamed."),
2415                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2416         } else {
2417                 new_path = handle_path_level_conflicts(opt, path, entry,
2418                                                        collisions, tree);
2419                 *clean_merge &= (new_path != NULL);
2420         }
2421
2422         return new_path;
2423 }
2424
2425 static void apply_directory_rename_modifications(struct merge_options *opt,
2426                                                  struct diff_filepair *pair,
2427                                                  char *new_path,
2428                                                  struct rename *re,
2429                                                  struct tree *tree,
2430                                                  struct tree *o_tree,
2431                                                  struct tree *a_tree,
2432                                                  struct tree *b_tree,
2433                                                  struct string_list *entries)
2434 {
2435         struct string_list_item *item;
2436         int stage = (tree == a_tree ? 2 : 3);
2437         int update_wd;
2438
2439         /*
2440          * In all cases where we can do directory rename detection,
2441          * unpack_trees() will have read pair->two->path into the
2442          * index and the working copy.  We need to remove it so that
2443          * we can instead place it at new_path.  It is guaranteed to
2444          * not be untracked (unpack_trees() would have errored out
2445          * saying the file would have been overwritten), but it might
2446          * be dirty, though.
2447          */
2448         update_wd = !was_dirty(opt, pair->two->path);
2449         if (!update_wd)
2450                 output(opt, 1, _("Refusing to lose dirty file at %s"),
2451                        pair->two->path);
2452         remove_file(opt, 1, pair->two->path, !update_wd);
2453
2454         /* Find or create a new re->dst_entry */
2455         item = string_list_lookup(entries, new_path);
2456         if (item) {
2457                 /*
2458                  * Since we're renaming on this side of history, and it's
2459                  * due to a directory rename on the other side of history
2460                  * (which we only allow when the directory in question no
2461                  * longer exists on the other side of history), the
2462                  * original entry for re->dst_entry is no longer
2463                  * necessary...
2464                  */
2465                 re->dst_entry->processed = 1;
2466
2467                 /*
2468                  * ...because we'll be using this new one.
2469                  */
2470                 re->dst_entry = item->util;
2471         } else {
2472                 /*
2473                  * re->dst_entry is for the before-dir-rename path, and we
2474                  * need it to hold information for the after-dir-rename
2475                  * path.  Before creating a new entry, we need to mark the
2476                  * old one as unnecessary (...unless it is shared by
2477                  * src_entry, i.e. this didn't use to be a rename, in which
2478                  * case we can just allow the normal processing to happen
2479                  * for it).
2480                  */
2481                 if (pair->status == 'R')
2482                         re->dst_entry->processed = 1;
2483
2484                 re->dst_entry = insert_stage_data(opt->repo, new_path,
2485                                                   o_tree, a_tree, b_tree,
2486                                                   entries);
2487                 item = string_list_insert(entries, new_path);
2488                 item->util = re->dst_entry;
2489         }
2490
2491         /*
2492          * Update the stage_data with the information about the path we are
2493          * moving into place.  That slot will be empty and available for us
2494          * to write to because of the collision checks in
2495          * handle_path_level_conflicts().  In other words,
2496          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2497          * open for us to write to.
2498          *
2499          * It may be tempting to actually update the index at this point as
2500          * well, using update_stages_for_stage_data(), but as per the big
2501          * "NOTE" in update_stages(), doing so will modify the current
2502          * in-memory index which will break calls to would_lose_untracked()
2503          * that we need to make.  Instead, we need to just make sure that
2504          * the various handle_rename_*() functions update the index
2505          * explicitly rather than relying on unpack_trees() to have done it.
2506          */
2507         get_tree_entry(opt->repo,
2508                        &tree->object.oid,
2509                        pair->two->path,
2510                        &re->dst_entry->stages[stage].oid,
2511                        &re->dst_entry->stages[stage].mode);
2512
2513         /*
2514          * Record the original change status (or 'type' of change).  If it
2515          * was originally an add ('A'), this lets us differentiate later
2516          * between a RENAME_DELETE conflict and RENAME_VIA_DIR (they
2517          * otherwise look the same).  If it was originally a rename ('R'),
2518          * this lets us remember and report accurately about the transitive
2519          * renaming that occurred via the directory rename detection.  Also,
2520          * record the original destination name.
2521          */
2522         re->dir_rename_original_type = pair->status;
2523         re->dir_rename_original_dest = pair->two->path;
2524
2525         /*
2526          * We don't actually look at pair->status again, but it seems
2527          * pedagogically correct to adjust it.
2528          */
2529         pair->status = 'R';
2530
2531         /*
2532          * Finally, record the new location.
2533          */
2534         pair->two->path = new_path;
2535 }
2536
2537 /*
2538  * Get information of all renames which occurred in 'pairs', making use of
2539  * any implicit directory renames inferred from the other side of history.
2540  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2541  * to be able to associate the correct cache entries with the rename
2542  * information; tree is always equal to either a_tree or b_tree.
2543  */
2544 static struct string_list *get_renames(struct merge_options *opt,
2545                                        const char *branch,
2546                                        struct diff_queue_struct *pairs,
2547                                        struct hashmap *dir_renames,
2548                                        struct hashmap *dir_rename_exclusions,
2549                                        struct tree *tree,
2550                                        struct tree *o_tree,
2551                                        struct tree *a_tree,
2552                                        struct tree *b_tree,
2553                                        struct string_list *entries,
2554                                        int *clean_merge)
2555 {
2556         int i;
2557         struct hashmap collisions;
2558         struct hashmap_iter iter;
2559         struct collision_entry *e;
2560         struct string_list *renames;
2561
2562         compute_collisions(&collisions, dir_renames, pairs);
2563         renames = xcalloc(1, sizeof(struct string_list));
2564
2565         for (i = 0; i < pairs->nr; ++i) {
2566                 struct string_list_item *item;
2567                 struct rename *re;
2568                 struct diff_filepair *pair = pairs->queue[i];
2569                 char *new_path; /* non-NULL only with directory renames */
2570
2571                 if (pair->status != 'A' && pair->status != 'R') {
2572                         diff_free_filepair(pair);
2573                         continue;
2574                 }
2575                 new_path = check_for_directory_rename(opt, pair->two->path, tree,
2576                                                       dir_renames,
2577                                                       dir_rename_exclusions,
2578                                                       &collisions,
2579                                                       clean_merge);
2580                 if (pair->status != 'R' && !new_path) {
2581                         diff_free_filepair(pair);
2582                         continue;
2583                 }
2584
2585                 re = xmalloc(sizeof(*re));
2586                 re->processed = 0;
2587                 re->pair = pair;
2588                 re->branch = branch;
2589                 re->dir_rename_original_type = '\0';
2590                 re->dir_rename_original_dest = NULL;
2591                 item = string_list_lookup(entries, re->pair->one->path);
2592                 if (!item)
2593                         re->src_entry = insert_stage_data(opt->repo,
2594                                         re->pair->one->path,
2595                                         o_tree, a_tree, b_tree, entries);
2596                 else
2597                         re->src_entry = item->util;
2598
2599                 item = string_list_lookup(entries, re->pair->two->path);
2600                 if (!item)
2601                         re->dst_entry = insert_stage_data(opt->repo,
2602                                         re->pair->two->path,
2603                                         o_tree, a_tree, b_tree, entries);
2604                 else
2605                         re->dst_entry = item->util;
2606                 item = string_list_insert(renames, pair->one->path);
2607                 item->util = re;
2608                 if (new_path)
2609                         apply_directory_rename_modifications(opt, pair, new_path,
2610                                                              re, tree, o_tree,
2611                                                              a_tree, b_tree,
2612                                                              entries);
2613         }
2614
2615         hashmap_iter_init(&collisions, &iter);
2616         while ((e = hashmap_iter_next(&iter))) {
2617                 free(e->target_file);
2618                 string_list_clear(&e->source_files, 0);
2619         }
2620         hashmap_free(&collisions, 1);
2621         return renames;
2622 }
2623
2624 static int process_renames(struct merge_options *opt,
2625                            struct string_list *a_renames,
2626                            struct string_list *b_renames)
2627 {
2628         int clean_merge = 1, i, j;
2629         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2630         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2631         const struct rename *sre;
2632
2633         for (i = 0; i < a_renames->nr; i++) {
2634                 sre = a_renames->items[i].util;
2635                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2636                         = (void *)sre;
2637         }
2638         for (i = 0; i < b_renames->nr; i++) {
2639                 sre = b_renames->items[i].util;
2640                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2641                         = (void *)sre;
2642         }
2643
2644         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2645                 struct string_list *renames1, *renames2Dst;
2646                 struct rename *ren1 = NULL, *ren2 = NULL;
2647                 const char *ren1_src, *ren1_dst;
2648                 struct string_list_item *lookup;
2649
2650                 if (i >= a_renames->nr) {
2651                         ren2 = b_renames->items[j++].util;
2652                 } else if (j >= b_renames->nr) {
2653                         ren1 = a_renames->items[i++].util;
2654                 } else {
2655                         int compare = strcmp(a_renames->items[i].string,
2656                                              b_renames->items[j].string);
2657                         if (compare <= 0)
2658                                 ren1 = a_renames->items[i++].util;
2659                         if (compare >= 0)
2660                                 ren2 = b_renames->items[j++].util;
2661                 }
2662
2663                 /* TODO: refactor, so that 1/2 are not needed */
2664                 if (ren1) {
2665                         renames1 = a_renames;
2666                         renames2Dst = &b_by_dst;
2667                 } else {
2668                         renames1 = b_renames;
2669                         renames2Dst = &a_by_dst;
2670                         SWAP(ren2, ren1);
2671                 }
2672
2673                 if (ren1->processed)
2674                         continue;
2675                 ren1->processed = 1;
2676                 ren1->dst_entry->processed = 1;
2677                 /* BUG: We should only mark src_entry as processed if we
2678                  * are not dealing with a rename + add-source case.
2679                  */
2680                 ren1->src_entry->processed = 1;
2681
2682                 ren1_src = ren1->pair->one->path;
2683                 ren1_dst = ren1->pair->two->path;
2684
2685                 if (ren2) {
2686                         /* One file renamed on both sides */
2687                         const char *ren2_src = ren2->pair->one->path;
2688                         const char *ren2_dst = ren2->pair->two->path;
2689                         enum rename_type rename_type;
2690                         if (strcmp(ren1_src, ren2_src) != 0)
2691                                 BUG("ren1_src != ren2_src");
2692                         ren2->dst_entry->processed = 1;
2693                         ren2->processed = 1;
2694                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2695                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2696                                 clean_merge = 0;
2697                         } else {
2698                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2699                                 /* BUG: We should only remove ren1_src in
2700                                  * the base stage (think of rename +
2701                                  * add-source cases).
2702                                  */
2703                                 remove_file(opt, 1, ren1_src, 1);
2704                                 update_entry(ren1->dst_entry,
2705                                              ren1->pair->one,
2706                                              ren1->pair->two,
2707                                              ren2->pair->two);
2708                         }
2709                         setup_rename_conflict_info(rename_type, opt, ren1, ren2);
2710                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2711                         /* Two different files renamed to the same thing */
2712                         char *ren2_dst;
2713                         ren2 = lookup->util;
2714                         ren2_dst = ren2->pair->two->path;
2715                         if (strcmp(ren1_dst, ren2_dst) != 0)
2716                                 BUG("ren1_dst != ren2_dst");
2717
2718                         clean_merge = 0;
2719                         ren2->processed = 1;
2720                         /*
2721                          * BUG: We should only mark src_entry as processed
2722                          * if we are not dealing with a rename + add-source
2723                          * case.
2724                          */
2725                         ren2->src_entry->processed = 1;
2726
2727                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2728                                                    opt, ren1, ren2);
2729                 } else {
2730                         /* Renamed in 1, maybe changed in 2 */
2731                         /* we only use sha1 and mode of these */
2732                         struct diff_filespec src_other, dst_other;
2733                         int try_merge;
2734
2735                         /*
2736                          * unpack_trees loads entries from common-commit
2737                          * into stage 1, from head-commit into stage 2, and
2738                          * from merge-commit into stage 3.  We keep track
2739                          * of which side corresponds to the rename.
2740                          */
2741                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2742                         int other_stage =   a_renames == renames1 ? 3 : 2;
2743
2744                         /* BUG: We should only remove ren1_src in the base
2745                          * stage and in other_stage (think of rename +
2746                          * add-source case).
2747                          */
2748                         remove_file(opt, 1, ren1_src,
2749                                     renamed_stage == 2 || !was_tracked(opt, ren1_src));
2750
2751                         oidcpy(&src_other.oid,
2752                                &ren1->src_entry->stages[other_stage].oid);
2753                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2754                         oidcpy(&dst_other.oid,
2755                                &ren1->dst_entry->stages[other_stage].oid);
2756                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2757                         try_merge = 0;
2758
2759                         if (oid_eq(&src_other.oid, &null_oid) &&
2760                             ren1->dir_rename_original_type == 'A') {
2761                                 setup_rename_conflict_info(RENAME_VIA_DIR,
2762                                                            opt, ren1, NULL);
2763                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2764                                 setup_rename_conflict_info(RENAME_DELETE,
2765                                                            opt, ren1, NULL);
2766                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2767                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2768                                 /*
2769                                  * Added file on the other side identical to
2770                                  * the file being renamed: clean merge.
2771                                  * Also, there is no need to overwrite the
2772                                  * file already in the working copy, so call
2773                                  * update_file_flags() instead of
2774                                  * update_file().
2775                                  */
2776                                 if (update_file_flags(opt,
2777                                                       ren1->pair->two,
2778                                                       ren1_dst,
2779                                                       1, /* update_cache */
2780                                                       0  /* update_wd    */))
2781                                         clean_merge = -1;
2782                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2783                                 /*
2784                                  * Probably not a clean merge, but it's
2785                                  * premature to set clean_merge to 0 here,
2786                                  * because if the rename merges cleanly and
2787                                  * the merge exactly matches the newly added
2788                                  * file, then the merge will be clean.
2789                                  */
2790                                 setup_rename_conflict_info(RENAME_ADD,
2791                                                            opt, ren1, NULL);
2792                         } else
2793                                 try_merge = 1;
2794
2795                         if (clean_merge < 0)
2796                                 goto cleanup_and_return;
2797                         if (try_merge) {
2798                                 struct diff_filespec *o, *a, *b;
2799                                 src_other.path = (char *)ren1_src;
2800
2801                                 o = ren1->pair->one;
2802                                 if (a_renames == renames1) {
2803                                         a = ren1->pair->two;
2804                                         b = &src_other;
2805                                 } else {
2806                                         b = ren1->pair->two;
2807                                         a = &src_other;
2808                                 }
2809                                 update_entry(ren1->dst_entry, o, a, b);
2810                                 setup_rename_conflict_info(RENAME_NORMAL,
2811                                                            opt, ren1, NULL);
2812                         }
2813                 }
2814         }
2815 cleanup_and_return:
2816         string_list_clear(&a_by_dst, 0);
2817         string_list_clear(&b_by_dst, 0);
2818
2819         return clean_merge;
2820 }
2821
2822 struct rename_info {
2823         struct string_list *head_renames;
2824         struct string_list *merge_renames;
2825 };
2826
2827 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2828                                    struct hashmap *dir_renames)
2829 {
2830         struct hashmap_iter iter;
2831         struct dir_rename_entry *e;
2832
2833         hashmap_iter_init(dir_renames, &iter);
2834         while ((e = hashmap_iter_next(&iter))) {
2835                 free(e->dir);
2836                 strbuf_release(&e->new_dir);
2837                 /* possible_new_dirs already cleared in get_directory_renames */
2838         }
2839         hashmap_free(dir_renames, 1);
2840         free(dir_renames);
2841
2842         free(pairs->queue);
2843         free(pairs);
2844 }
2845
2846 static int detect_and_process_renames(struct merge_options *opt,
2847                                       struct tree *common,
2848                                       struct tree *head,
2849                                       struct tree *merge,
2850                                       struct string_list *entries,
2851                                       struct rename_info *ri)
2852 {
2853         struct diff_queue_struct *head_pairs, *merge_pairs;
2854         struct hashmap *dir_re_head, *dir_re_merge;
2855         int clean = 1;
2856
2857         ri->head_renames = NULL;
2858         ri->merge_renames = NULL;
2859
2860         if (!merge_detect_rename(opt))
2861                 return 1;
2862
2863         head_pairs = get_diffpairs(opt, common, head);
2864         merge_pairs = get_diffpairs(opt, common, merge);
2865
2866         if ((opt->detect_directory_renames == 2) ||
2867             (opt->detect_directory_renames == 1 && !opt->call_depth)) {
2868                 dir_re_head = get_directory_renames(head_pairs);
2869                 dir_re_merge = get_directory_renames(merge_pairs);
2870
2871                 handle_directory_level_conflicts(opt,
2872                                                  dir_re_head, head,
2873                                                  dir_re_merge, merge);
2874         } else {
2875                 dir_re_head  = xmalloc(sizeof(*dir_re_head));
2876                 dir_re_merge = xmalloc(sizeof(*dir_re_merge));
2877                 dir_rename_init(dir_re_head);
2878                 dir_rename_init(dir_re_merge);
2879         }
2880
2881         ri->head_renames  = get_renames(opt, opt->branch1, head_pairs,
2882                                         dir_re_merge, dir_re_head, head,
2883                                         common, head, merge, entries,
2884                                         &clean);
2885         if (clean < 0)
2886                 goto cleanup;
2887         ri->merge_renames = get_renames(opt, opt->branch2, merge_pairs,
2888                                         dir_re_head, dir_re_merge, merge,
2889                                         common, head, merge, entries,
2890                                         &clean);
2891         if (clean < 0)
2892                 goto cleanup;
2893         clean &= process_renames(opt, ri->head_renames, ri->merge_renames);
2894
2895 cleanup:
2896         /*
2897          * Some cleanup is deferred until cleanup_renames() because the
2898          * data structures are still needed and referenced in
2899          * process_entry().  But there are a few things we can free now.
2900          */
2901         initial_cleanup_rename(head_pairs, dir_re_head);
2902         initial_cleanup_rename(merge_pairs, dir_re_merge);
2903
2904         return clean;
2905 }
2906
2907 static void final_cleanup_rename(struct string_list *rename)
2908 {
2909         const struct rename *re;
2910         int i;
2911
2912         if (rename == NULL)
2913                 return;
2914
2915         for (i = 0; i < rename->nr; i++) {
2916                 re = rename->items[i].util;
2917                 diff_free_filepair(re->pair);
2918         }
2919         string_list_clear(rename, 1);
2920         free(rename);
2921 }
2922
2923 static void final_cleanup_renames(struct rename_info *re_info)
2924 {
2925         final_cleanup_rename(re_info->head_renames);
2926         final_cleanup_rename(re_info->merge_renames);
2927 }
2928
2929 static int read_oid_strbuf(struct merge_options *opt,
2930                            const struct object_id *oid,
2931                            struct strbuf *dst)
2932 {
2933         void *buf;
2934         enum object_type type;
2935         unsigned long size;
2936         buf = read_object_file(oid, &type, &size);
2937         if (!buf)
2938                 return err(opt, _("cannot read object %s"), oid_to_hex(oid));
2939         if (type != OBJ_BLOB) {
2940                 free(buf);
2941                 return err(opt, _("object %s is not a blob"), oid_to_hex(oid));
2942         }
2943         strbuf_attach(dst, buf, size, size + 1);
2944         return 0;
2945 }
2946
2947 static int blob_unchanged(struct merge_options *opt,
2948                           const struct diff_filespec *o,
2949                           const struct diff_filespec *a,
2950                           int renormalize, const char *path)
2951 {
2952         struct strbuf obuf = STRBUF_INIT;
2953         struct strbuf abuf = STRBUF_INIT;
2954         int ret = 0; /* assume changed for safety */
2955         const struct index_state *idx = opt->repo->index;
2956
2957         if (a->mode != o->mode)
2958                 return 0;
2959         if (oid_eq(&o->oid, &a->oid))
2960                 return 1;
2961         if (!renormalize)
2962                 return 0;
2963
2964         if (read_oid_strbuf(opt, &o->oid, &obuf) ||
2965             read_oid_strbuf(opt, &a->oid, &abuf))
2966                 goto error_return;
2967         /*
2968          * Note: binary | is used so that both renormalizations are
2969          * performed.  Comparison can be skipped if both files are
2970          * unchanged since their sha1s have already been compared.
2971          */
2972         if (renormalize_buffer(idx, path, obuf.buf, obuf.len, &obuf) |
2973             renormalize_buffer(idx, path, abuf.buf, abuf.len, &abuf))
2974                 ret = (obuf.len == abuf.len && !memcmp(obuf.buf, abuf.buf, obuf.len));
2975
2976 error_return:
2977         strbuf_release(&obuf);
2978         strbuf_release(&abuf);
2979         return ret;
2980 }
2981
2982 static int handle_modify_delete(struct merge_options *opt,
2983                                 const char *path,
2984                                 const struct diff_filespec *o,
2985                                 const struct diff_filespec *a,
2986                                 const struct diff_filespec *b)
2987 {
2988         const char *modify_branch, *delete_branch;
2989         const struct diff_filespec *changed;
2990
2991         if (is_valid(a)) {
2992                 modify_branch = opt->branch1;
2993                 delete_branch = opt->branch2;
2994                 changed = a;
2995         } else {
2996                 modify_branch = opt->branch2;
2997                 delete_branch = opt->branch1;
2998                 changed = b;
2999         }
3000
3001         return handle_change_delete(opt,
3002                                     path, NULL,
3003                                     o, changed,
3004                                     modify_branch, delete_branch,
3005                                     _("modify"), _("modified"));
3006 }
3007
3008 static int handle_content_merge(struct merge_file_info *mfi,
3009                                 struct merge_options *opt,
3010                                 const char *path,
3011                                 int is_dirty,
3012                                 const struct diff_filespec *o,
3013                                 const struct diff_filespec *a,
3014                                 const struct diff_filespec *b,
3015                                 struct rename_conflict_info *ci)
3016 {
3017         const char *reason = _("content");
3018         unsigned df_conflict_remains = 0;
3019
3020         if (!is_valid(o))
3021                 reason = _("add/add");
3022
3023         assert(o->path && a->path && b->path);
3024         if (ci && dir_in_way(opt->repo->index, path, !opt->call_depth,
3025                              S_ISGITLINK(ci->ren1->pair->two->mode)))
3026                 df_conflict_remains = 1;
3027
3028         if (merge_mode_and_contents(opt, o, a, b, path,
3029                                     opt->branch1, opt->branch2,
3030                                     opt->call_depth * 2, mfi))
3031                 return -1;
3032
3033         /*
3034          * We can skip updating the working tree file iff:
3035          *   a) The merge is clean
3036          *   b) The merge matches what was in HEAD (content, mode, pathname)
3037          *   c) The target path is usable (i.e. not involved in D/F conflict)
3038          */
3039         if (mfi->clean && was_tracked_and_matches(opt, path, &mfi->blob) &&
3040             !df_conflict_remains) {
3041                 int pos;
3042                 struct cache_entry *ce;
3043
3044                 output(opt, 3, _("Skipped %s (merged same as existing)"), path);
3045                 if (add_cacheinfo(opt, &mfi->blob, path,
3046                                   0, (!opt->call_depth && !is_dirty), 0))
3047                         return -1;
3048                 /*
3049                  * However, add_cacheinfo() will delete the old cache entry
3050                  * and add a new one.  We need to copy over any skip_worktree
3051                  * flag to avoid making the file appear as if it were
3052                  * deleted by the user.
3053                  */
3054                 pos = index_name_pos(&opt->orig_index, path, strlen(path));
3055                 ce = opt->orig_index.cache[pos];
3056                 if (ce_skip_worktree(ce)) {
3057                         pos = index_name_pos(opt->repo->index, path, strlen(path));
3058                         ce = opt->repo->index->cache[pos];
3059                         ce->ce_flags |= CE_SKIP_WORKTREE;
3060                 }
3061                 return mfi->clean;
3062         }
3063
3064         if (!mfi->clean) {
3065                 if (S_ISGITLINK(mfi->blob.mode))
3066                         reason = _("submodule");
3067                 output(opt, 1, _("CONFLICT (%s): Merge conflict in %s"),
3068                                 reason, path);
3069                 if (ci && !df_conflict_remains)
3070                         if (update_stages(opt, path, o, a, b))
3071                                 return -1;
3072         }
3073
3074         if (df_conflict_remains || is_dirty) {
3075                 char *new_path;
3076                 if (opt->call_depth) {
3077                         remove_file_from_index(opt->repo->index, path);
3078                 } else {
3079                         if (!mfi->clean) {
3080                                 if (update_stages(opt, path, o, a, b))
3081                                         return -1;
3082                         } else {
3083                                 int file_from_stage2 = was_tracked(opt, path);
3084
3085                                 if (update_stages(opt, path, NULL,
3086                                                   file_from_stage2 ? &mfi->blob : NULL,
3087                                                   file_from_stage2 ? NULL : &mfi->blob))
3088                                         return -1;
3089                         }
3090
3091                 }
3092                 new_path = unique_path(opt, path, ci->ren1->branch);
3093                 if (is_dirty) {
3094                         output(opt, 1, _("Refusing to lose dirty file at %s"),
3095                                path);
3096                 }
3097                 output(opt, 1, _("Adding as %s instead"), new_path);
3098                 if (update_file(opt, 0, &mfi->blob, new_path)) {
3099                         free(new_path);
3100                         return -1;
3101                 }
3102                 free(new_path);
3103                 mfi->clean = 0;
3104         } else if (update_file(opt, mfi->clean, &mfi->blob, path))
3105                 return -1;
3106         return !is_dirty && mfi->clean;
3107 }
3108
3109 static int handle_rename_normal(struct merge_options *opt,
3110                                 const char *path,
3111                                 const struct diff_filespec *o,
3112                                 const struct diff_filespec *a,
3113                                 const struct diff_filespec *b,
3114                                 struct rename_conflict_info *ci)
3115 {
3116         struct rename *ren = ci->ren1;
3117         struct merge_file_info mfi;
3118         int clean;
3119         int side = (ren->branch == opt->branch1 ? 2 : 3);
3120
3121         /* Merge the content and write it out */
3122         clean = handle_content_merge(&mfi, opt, path, was_dirty(opt, path),
3123                                      o, a, b, ci);
3124
3125         if (clean && opt->detect_directory_renames == 1 &&
3126             ren->dir_rename_original_dest) {
3127                 if (update_stages(opt, path,
3128                                   NULL,
3129                                   side == 2 ? &mfi.blob : NULL,
3130                                   side == 2 ? NULL : &mfi.blob))
3131                         return -1;
3132                 clean = 0; /* not clean, but conflicted */
3133         }
3134         return clean;
3135 }
3136
3137 static void dir_rename_warning(const char *msg,
3138                                int is_add,
3139                                int clean,
3140                                struct merge_options *opt,
3141                                struct rename *ren)
3142 {
3143         const char *other_branch;
3144         other_branch = (ren->branch == opt->branch1 ?
3145                         opt->branch2 : opt->branch1);
3146         if (is_add) {
3147                 output(opt, clean ? 2 : 1, msg,
3148                        ren->pair->one->path, ren->branch,
3149                        other_branch, ren->pair->two->path);
3150                 return;
3151         }
3152         output(opt, clean ? 2 : 1, msg,
3153                ren->pair->one->path, ren->dir_rename_original_dest, ren->branch,
3154                other_branch, ren->pair->two->path);
3155 }
3156 static int warn_about_dir_renamed_entries(struct merge_options *opt,
3157                                           struct rename *ren)
3158 {
3159         const char *msg;
3160         int clean = 1, is_add;
3161
3162         if (!ren)
3163                 return clean;
3164
3165         /* Return early if ren was not affected/created by a directory rename */
3166         if (!ren->dir_rename_original_dest)
3167                 return clean;
3168
3169         /* Sanity checks */
3170         assert(opt->detect_directory_renames > 0);
3171         assert(ren->dir_rename_original_type == 'A' ||
3172                ren->dir_rename_original_type == 'R');
3173
3174         /* Check whether to treat directory renames as a conflict */
3175         clean = (opt->detect_directory_renames == 2);
3176
3177         is_add = (ren->dir_rename_original_type == 'A');
3178         if (ren->dir_rename_original_type == 'A' && clean) {
3179                 msg = _("Path updated: %s added in %s inside a "
3180                         "directory that was renamed in %s; moving it to %s.");
3181         } else if (ren->dir_rename_original_type == 'A' && !clean) {
3182                 msg = _("CONFLICT (file location): %s added in %s "
3183                         "inside a directory that was renamed in %s, "
3184                         "suggesting it should perhaps be moved to %s.");
3185         } else if (ren->dir_rename_original_type == 'R' && clean) {
3186                 msg = _("Path updated: %s renamed to %s in %s, inside a "
3187                         "directory that was renamed in %s; moving it to %s.");
3188         } else if (ren->dir_rename_original_type == 'R' && !clean) {
3189                 msg = _("CONFLICT (file location): %s renamed to %s in %s, "
3190                         "inside a directory that was renamed in %s, "
3191                         "suggesting it should perhaps be moved to %s.");
3192         } else {
3193                 BUG("Impossible dir_rename_original_type/clean combination");
3194         }
3195         dir_rename_warning(msg, is_add, clean, opt, ren);
3196
3197         return clean;
3198 }
3199
3200 /* Per entry merge function */
3201 static int process_entry(struct merge_options *opt,
3202                          const char *path, struct stage_data *entry)
3203 {
3204         int clean_merge = 1;
3205         int normalize = opt->renormalize;
3206
3207         struct diff_filespec *o = &entry->stages[1];
3208         struct diff_filespec *a = &entry->stages[2];
3209         struct diff_filespec *b = &entry->stages[3];
3210         int o_valid = is_valid(o);
3211         int a_valid = is_valid(a);
3212         int b_valid = is_valid(b);
3213         o->path = a->path = b->path = (char*)path;
3214
3215         entry->processed = 1;
3216         if (entry->rename_conflict_info) {
3217                 struct rename_conflict_info *ci = entry->rename_conflict_info;
3218                 struct diff_filespec *temp;
3219                 int path_clean;
3220
3221                 path_clean = warn_about_dir_renamed_entries(opt, ci->ren1);
3222                 path_clean &= warn_about_dir_renamed_entries(opt, ci->ren2);
3223
3224                 /*
3225                  * For cases with a single rename, {o,a,b}->path have all been
3226                  * set to the rename target path; we need to set two of these
3227                  * back to the rename source.
3228                  * For rename/rename conflicts, we'll manually fix paths below.
3229                  */
3230                 temp = (opt->branch1 == ci->ren1->branch) ? b : a;
3231                 o->path = temp->path = ci->ren1->pair->one->path;
3232                 if (ci->ren2) {
3233                         assert(opt->branch1 == ci->ren1->branch);
3234                 }
3235
3236                 switch (ci->rename_type) {
3237                 case RENAME_NORMAL:
3238                 case RENAME_ONE_FILE_TO_ONE:
3239                         clean_merge = handle_rename_normal(opt, path, o, a, b,
3240                                                            ci);
3241                         break;
3242                 case RENAME_VIA_DIR:
3243                         clean_merge = handle_rename_via_dir(opt, ci);
3244                         break;
3245                 case RENAME_ADD:
3246                         /*
3247                          * Probably unclean merge, but if the renamed file
3248                          * merges cleanly and the result can then be
3249                          * two-way merged cleanly with the added file, I
3250                          * guess it's a clean merge?
3251                          */
3252                         clean_merge = handle_rename_add(opt, ci);
3253                         break;
3254                 case RENAME_DELETE:
3255                         clean_merge = 0;
3256                         if (handle_rename_delete(opt, ci))
3257                                 clean_merge = -1;
3258                         break;
3259                 case RENAME_ONE_FILE_TO_TWO:
3260                         /*
3261                          * Manually fix up paths; note:
3262                          * ren[12]->pair->one->path are equal.
3263                          */
3264                         o->path = ci->ren1->pair->one->path;
3265                         a->path = ci->ren1->pair->two->path;
3266                         b->path = ci->ren2->pair->two->path;
3267
3268                         clean_merge = 0;
3269                         if (handle_rename_rename_1to2(opt, ci))
3270                                 clean_merge = -1;
3271                         break;
3272                 case RENAME_TWO_FILES_TO_ONE:
3273                         /*
3274                          * Manually fix up paths; note,
3275                          * ren[12]->pair->two->path are actually equal.
3276                          */
3277                         o->path = NULL;
3278                         a->path = ci->ren1->pair->two->path;
3279                         b->path = ci->ren2->pair->two->path;
3280
3281                         /*
3282                          * Probably unclean merge, but if the two renamed
3283                          * files merge cleanly and the two resulting files
3284                          * can then be two-way merged cleanly, I guess it's
3285                          * a clean merge?
3286                          */
3287                         clean_merge = handle_rename_rename_2to1(opt, ci);
3288                         break;
3289                 default:
3290                         entry->processed = 0;
3291                         break;
3292                 }
3293                 if (path_clean < clean_merge)
3294                         clean_merge = path_clean;
3295         } else if (o_valid && (!a_valid || !b_valid)) {
3296                 /* Case A: Deleted in one */
3297                 if ((!a_valid && !b_valid) ||
3298                     (!b_valid && blob_unchanged(opt, o, a, normalize, path)) ||
3299                     (!a_valid && blob_unchanged(opt, o, b, normalize, path))) {
3300                         /* Deleted in both or deleted in one and
3301                          * unchanged in the other */
3302                         if (a_valid)
3303                                 output(opt, 2, _("Removing %s"), path);
3304                         /* do not touch working file if it did not exist */
3305                         remove_file(opt, 1, path, !a_valid);
3306                 } else {
3307                         /* Modify/delete; deleted side may have put a directory in the way */
3308                         clean_merge = 0;
3309                         if (handle_modify_delete(opt, path, o, a, b))
3310                                 clean_merge = -1;
3311                 }
3312         } else if ((!o_valid && a_valid && !b_valid) ||
3313                    (!o_valid && !a_valid && b_valid)) {
3314                 /* Case B: Added in one. */
3315                 /* [nothing|directory] -> ([nothing|directory], file) */
3316
3317                 const char *add_branch;
3318                 const char *other_branch;
3319                 const char *conf;
3320                 const struct diff_filespec *contents;
3321
3322                 if (a_valid) {
3323                         add_branch = opt->branch1;
3324                         other_branch = opt->branch2;
3325                         contents = a;
3326                         conf = _("file/directory");
3327                 } else {
3328                         add_branch = opt->branch2;
3329                         other_branch = opt->branch1;
3330                         contents = b;
3331                         conf = _("directory/file");
3332                 }
3333                 if (dir_in_way(opt->repo->index, path,
3334                                !opt->call_depth && !S_ISGITLINK(a->mode),
3335                                0)) {
3336                         char *new_path = unique_path(opt, path, add_branch);
3337                         clean_merge = 0;
3338                         output(opt, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3339                                "Adding %s as %s"),
3340                                conf, path, other_branch, path, new_path);
3341                         if (update_file(opt, 0, contents, new_path))
3342                                 clean_merge = -1;
3343                         else if (opt->call_depth)
3344                                 remove_file_from_index(opt->repo->index, path);
3345                         free(new_path);
3346                 } else {
3347                         output(opt, 2, _("Adding %s"), path);
3348                         /* do not overwrite file if already present */
3349                         if (update_file_flags(opt, contents, path, 1, !a_valid))
3350                                 clean_merge = -1;
3351                 }
3352         } else if (a_valid && b_valid) {
3353                 if (!o_valid) {
3354                         /* Case C: Added in both (check for same permissions) */
3355                         output(opt, 1,
3356                                _("CONFLICT (add/add): Merge conflict in %s"),
3357                                path);
3358                         clean_merge = handle_file_collision(opt,
3359                                                             path, NULL, NULL,
3360                                                             opt->branch1,
3361                                                             opt->branch2,
3362                                                             a, b);
3363                 } else {
3364                         /* case D: Modified in both, but differently. */
3365                         struct merge_file_info mfi;
3366                         int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3367                         clean_merge = handle_content_merge(&mfi, opt, path,
3368                                                            is_dirty,
3369                                                            o, a, b, NULL);
3370                 }
3371         } else if (!o_valid && !a_valid && !b_valid) {
3372                 /*
3373                  * this entry was deleted altogether. a_mode == 0 means
3374                  * we had that path and want to actively remove it.
3375                  */
3376                 remove_file(opt, 1, path, !a->mode);
3377         } else
3378                 BUG("fatal merge failure, shouldn't happen.");
3379
3380         return clean_merge;
3381 }
3382
3383 int merge_trees(struct merge_options *opt,
3384                 struct tree *head,
3385                 struct tree *merge,
3386                 struct tree *common,
3387                 struct tree **result)
3388 {
3389         struct index_state *istate = opt->repo->index;
3390         int code, clean;
3391         struct strbuf sb = STRBUF_INIT;
3392
3393         if (!opt->call_depth && repo_index_has_changes(opt->repo, head, &sb)) {
3394                 err(opt, _("Your local changes to the following files would be overwritten by merge:\n  %s"),
3395                     sb.buf);
3396                 return -1;
3397         }
3398
3399         if (opt->subtree_shift) {
3400                 merge = shift_tree_object(opt->repo, head, merge, opt->subtree_shift);
3401                 common = shift_tree_object(opt->repo, head, common, opt->subtree_shift);
3402         }
3403
3404         if (oid_eq(&common->object.oid, &merge->object.oid)) {
3405                 output(opt, 0, _("Already up to date!"));
3406                 *result = head;
3407                 return 1;
3408         }
3409
3410         code = unpack_trees_start(opt, common, head, merge);
3411
3412         if (code != 0) {
3413                 if (show(opt, 4) || opt->call_depth)
3414                         err(opt, _("merging of trees %s and %s failed"),
3415                             oid_to_hex(&head->object.oid),
3416                             oid_to_hex(&merge->object.oid));
3417                 unpack_trees_finish(opt);
3418                 return -1;
3419         }
3420
3421         if (unmerged_index(istate)) {
3422                 struct string_list *entries;
3423                 struct rename_info re_info;
3424                 int i;
3425                 /*
3426                  * Only need the hashmap while processing entries, so
3427                  * initialize it here and free it when we are done running
3428                  * through the entries. Keeping it in the merge_options as
3429                  * opposed to decaring a local hashmap is for convenience
3430                  * so that we don't have to pass it to around.
3431                  */
3432                 hashmap_init(&opt->current_file_dir_set, path_hashmap_cmp, NULL, 512);
3433                 get_files_dirs(opt, head);
3434                 get_files_dirs(opt, merge);
3435
3436                 entries = get_unmerged(opt->repo->index);
3437                 clean = detect_and_process_renames(opt, common, head, merge,
3438                                                    entries, &re_info);
3439                 record_df_conflict_files(opt, entries);
3440                 if (clean < 0)
3441                         goto cleanup;
3442                 for (i = entries->nr-1; 0 <= i; i--) {
3443                         const char *path = entries->items[i].string;
3444                         struct stage_data *e = entries->items[i].util;
3445                         if (!e->processed) {
3446                                 int ret = process_entry(opt, path, e);
3447                                 if (!ret)
3448                                         clean = 0;
3449                                 else if (ret < 0) {
3450                                         clean = ret;
3451                                         goto cleanup;
3452                                 }
3453                         }
3454                 }
3455                 for (i = 0; i < entries->nr; i++) {
3456                         struct stage_data *e = entries->items[i].util;
3457                         if (!e->processed)
3458                                 BUG("unprocessed path??? %s",
3459                                     entries->items[i].string);
3460                 }
3461
3462         cleanup:
3463                 final_cleanup_renames(&re_info);
3464
3465                 string_list_clear(entries, 1);
3466                 free(entries);
3467
3468                 hashmap_free(&opt->current_file_dir_set, 1);
3469
3470                 if (clean < 0) {
3471                         unpack_trees_finish(opt);
3472                         return clean;
3473                 }
3474         }
3475         else
3476                 clean = 1;
3477
3478         unpack_trees_finish(opt);
3479
3480         if (opt->call_depth && !(*result = write_tree_from_memory(opt)))
3481                 return -1;
3482
3483         return clean;
3484 }
3485
3486 static struct commit_list *reverse_commit_list(struct commit_list *list)
3487 {
3488         struct commit_list *next = NULL, *current, *backup;
3489         for (current = list; current; current = backup) {
3490                 backup = current->next;
3491                 current->next = next;
3492                 next = current;
3493         }
3494         return next;
3495 }
3496
3497 /*
3498  * Merge the commits h1 and h2, return the resulting virtual
3499  * commit object and a flag indicating the cleanness of the merge.
3500  */
3501 int merge_recursive(struct merge_options *opt,
3502                     struct commit *h1,
3503                     struct commit *h2,
3504                     struct commit_list *ca,
3505                     struct commit **result)
3506 {
3507         struct commit_list *iter;
3508         struct commit *merged_common_ancestors;
3509         struct tree *mrtree;
3510         int clean;
3511
3512         if (show(opt, 4)) {
3513                 output(opt, 4, _("Merging:"));
3514                 output_commit_title(opt, h1);
3515                 output_commit_title(opt, h2);
3516         }
3517
3518         if (!ca) {
3519                 ca = get_merge_bases(h1, h2);
3520                 ca = reverse_commit_list(ca);
3521         }
3522
3523         if (show(opt, 5)) {
3524                 unsigned cnt = commit_list_count(ca);
3525
3526                 output(opt, 5, Q_("found %u common ancestor:",
3527                                 "found %u common ancestors:", cnt), cnt);
3528                 for (iter = ca; iter; iter = iter->next)
3529                         output_commit_title(opt, iter->item);
3530         }
3531
3532         merged_common_ancestors = pop_commit(&ca);
3533         if (merged_common_ancestors == NULL) {
3534                 /* if there is no common ancestor, use an empty tree */
3535                 struct tree *tree;
3536
3537                 tree = lookup_tree(opt->repo, opt->repo->hash_algo->empty_tree);
3538                 merged_common_ancestors = make_virtual_commit(opt->repo, tree, "ancestor");
3539         }
3540
3541         for (iter = ca; iter; iter = iter->next) {
3542                 const char *saved_b1, *saved_b2;
3543                 opt->call_depth++;
3544                 /*
3545                  * When the merge fails, the result contains files
3546                  * with conflict markers. The cleanness flag is
3547                  * ignored (unless indicating an error), it was never
3548                  * actually used, as result of merge_trees has always
3549                  * overwritten it: the committed "conflicts" were
3550                  * already resolved.
3551                  */
3552                 discard_index(opt->repo->index);
3553                 saved_b1 = opt->branch1;
3554                 saved_b2 = opt->branch2;
3555                 opt->branch1 = "Temporary merge branch 1";
3556                 opt->branch2 = "Temporary merge branch 2";
3557                 if (merge_recursive(opt, merged_common_ancestors, iter->item,
3558                                     NULL, &merged_common_ancestors) < 0)
3559                         return -1;
3560                 opt->branch1 = saved_b1;
3561                 opt->branch2 = saved_b2;
3562                 opt->call_depth--;
3563
3564                 if (!merged_common_ancestors)
3565                         return err(opt, _("merge returned no commit"));
3566         }
3567
3568         discard_index(opt->repo->index);
3569         if (!opt->call_depth)
3570                 repo_read_index(opt->repo);
3571
3572         opt->ancestor = "merged common ancestors";
3573         clean = merge_trees(opt, get_commit_tree(h1), get_commit_tree(h2),
3574                             get_commit_tree(merged_common_ancestors),
3575                             &mrtree);
3576         if (clean < 0) {
3577                 flush_output(opt);
3578                 return clean;
3579         }
3580
3581         if (opt->call_depth) {
3582                 *result = make_virtual_commit(opt->repo, mrtree, "merged tree");
3583                 commit_list_insert(h1, &(*result)->parents);
3584                 commit_list_insert(h2, &(*result)->parents->next);
3585         }
3586         flush_output(opt);
3587         if (!opt->call_depth && opt->buffer_output < 2)
3588                 strbuf_release(&opt->obuf);
3589         if (show(opt, 2))
3590                 diff_warn_rename_limit("merge.renamelimit",
3591                                        opt->needed_rename_limit, 0);
3592         return clean;
3593 }
3594
3595 static struct commit *get_ref(struct repository *repo, const struct object_id *oid,
3596                               const char *name)
3597 {
3598         struct object *object;
3599
3600         object = deref_tag(repo, parse_object(repo, oid),
3601                            name, strlen(name));
3602         if (!object)
3603                 return NULL;
3604         if (object->type == OBJ_TREE)
3605                 return make_virtual_commit(repo, (struct tree*)object, name);
3606         if (object->type != OBJ_COMMIT)
3607                 return NULL;
3608         if (parse_commit((struct commit *)object))
3609                 return NULL;
3610         return (struct commit *)object;
3611 }
3612
3613 int merge_recursive_generic(struct merge_options *opt,
3614                             const struct object_id *head,
3615                             const struct object_id *merge,
3616                             int num_base_list,
3617                             const struct object_id **base_list,
3618                             struct commit **result)
3619 {
3620         int clean;
3621         struct lock_file lock = LOCK_INIT;
3622         struct commit *head_commit = get_ref(opt->repo, head, opt->branch1);
3623         struct commit *next_commit = get_ref(opt->repo, merge, opt->branch2);
3624         struct commit_list *ca = NULL;
3625
3626         if (base_list) {
3627                 int i;
3628                 for (i = 0; i < num_base_list; ++i) {
3629                         struct commit *base;
3630                         if (!(base = get_ref(opt->repo, base_list[i], oid_to_hex(base_list[i]))))
3631                                 return err(opt, _("Could not parse object '%s'"),
3632                                            oid_to_hex(base_list[i]));
3633                         commit_list_insert(base, &ca);
3634                 }
3635         }
3636
3637         repo_hold_locked_index(opt->repo, &lock, LOCK_DIE_ON_ERROR);
3638         clean = merge_recursive(opt, head_commit, next_commit, ca,
3639                                 result);
3640         if (clean < 0) {
3641                 rollback_lock_file(&lock);
3642                 return clean;
3643         }
3644
3645         if (write_locked_index(opt->repo->index, &lock,
3646                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3647                 return err(opt, _("Unable to write index."));
3648
3649         return clean ? 0 : 1;
3650 }
3651
3652 static void merge_recursive_config(struct merge_options *opt)
3653 {
3654         char *value = NULL;
3655         git_config_get_int("merge.verbosity", &opt->verbosity);
3656         git_config_get_int("diff.renamelimit", &opt->diff_rename_limit);
3657         git_config_get_int("merge.renamelimit", &opt->merge_rename_limit);
3658         if (!git_config_get_string("diff.renames", &value)) {
3659                 opt->diff_detect_rename = git_config_rename("diff.renames", value);
3660                 free(value);
3661         }
3662         if (!git_config_get_string("merge.renames", &value)) {
3663                 opt->merge_detect_rename = git_config_rename("merge.renames", value);
3664                 free(value);
3665         }
3666         if (!git_config_get_string("merge.directoryrenames", &value)) {
3667                 int boolval = git_parse_maybe_bool(value);
3668                 if (0 <= boolval) {
3669                         opt->detect_directory_renames = boolval ? 2 : 0;
3670                 } else if (!strcasecmp(value, "conflict")) {
3671                         opt->detect_directory_renames = 1;
3672                 } /* avoid erroring on values from future versions of git */
3673                 free(value);
3674         }
3675         git_config(git_xmerge_config, NULL);
3676 }
3677
3678 void init_merge_options(struct merge_options *opt,
3679                         struct repository *repo)
3680 {
3681         const char *merge_verbosity;
3682         memset(opt, 0, sizeof(struct merge_options));
3683         opt->repo = repo;
3684         opt->verbosity = 2;
3685         opt->buffer_output = 1;
3686         opt->diff_rename_limit = -1;
3687         opt->merge_rename_limit = -1;
3688         opt->renormalize = 0;
3689         opt->diff_detect_rename = -1;
3690         opt->merge_detect_rename = -1;
3691         opt->detect_directory_renames = 1;
3692         merge_recursive_config(opt);
3693         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3694         if (merge_verbosity)
3695                 opt->verbosity = strtol(merge_verbosity, NULL, 10);
3696         if (opt->verbosity >= 5)
3697                 opt->buffer_output = 0;
3698         strbuf_init(&opt->obuf, 0);
3699         string_list_init(&opt->df_conflict_file_set, 1);
3700 }
3701
3702 int parse_merge_opt(struct merge_options *opt, const char *s)
3703 {
3704         const char *arg;
3705
3706         if (!s || !*s)
3707                 return -1;
3708         if (!strcmp(s, "ours"))
3709                 opt->recursive_variant = MERGE_RECURSIVE_OURS;
3710         else if (!strcmp(s, "theirs"))
3711                 opt->recursive_variant = MERGE_RECURSIVE_THEIRS;
3712         else if (!strcmp(s, "subtree"))
3713                 opt->subtree_shift = "";
3714         else if (skip_prefix(s, "subtree=", &arg))
3715                 opt->subtree_shift = arg;
3716         else if (!strcmp(s, "patience"))
3717                 opt->xdl_opts = DIFF_WITH_ALG(opt, PATIENCE_DIFF);
3718         else if (!strcmp(s, "histogram"))
3719                 opt->xdl_opts = DIFF_WITH_ALG(opt, HISTOGRAM_DIFF);
3720         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3721                 long value = parse_algorithm_value(arg);
3722                 if (value < 0)
3723                         return -1;
3724                 /* clear out previous settings */
3725                 DIFF_XDL_CLR(opt, NEED_MINIMAL);
3726                 opt->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3727                 opt->xdl_opts |= value;
3728         }
3729         else if (!strcmp(s, "ignore-space-change"))
3730                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_CHANGE);
3731         else if (!strcmp(s, "ignore-all-space"))
3732                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE);
3733         else if (!strcmp(s, "ignore-space-at-eol"))
3734                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_AT_EOL);
3735         else if (!strcmp(s, "ignore-cr-at-eol"))
3736                 DIFF_XDL_SET(opt, IGNORE_CR_AT_EOL);
3737         else if (!strcmp(s, "renormalize"))
3738                 opt->renormalize = 1;
3739         else if (!strcmp(s, "no-renormalize"))
3740                 opt->renormalize = 0;
3741         else if (!strcmp(s, "no-renames"))
3742                 opt->merge_detect_rename = 0;
3743         else if (!strcmp(s, "find-renames")) {
3744                 opt->merge_detect_rename = 1;
3745                 opt->rename_score = 0;
3746         }
3747         else if (skip_prefix(s, "find-renames=", &arg) ||
3748                  skip_prefix(s, "rename-threshold=", &arg)) {
3749                 if ((opt->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3750                         return -1;
3751                 opt->merge_detect_rename = 1;
3752         }
3753         /*
3754          * Please update $__git_merge_strategy_options in
3755          * git-completion.bash when you add new options
3756          */
3757         else
3758                 return -1;
3759         return 0;
3760 }