Merge branch 'en/merge-recursive-cleanup'
[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 "commit.h"
12 #include "blob.h"
13 #include "builtin.h"
14 #include "tree-walk.h"
15 #include "diff.h"
16 #include "diffcore.h"
17 #include "tag.h"
18 #include "alloc.h"
19 #include "unpack-trees.h"
20 #include "string-list.h"
21 #include "xdiff-interface.h"
22 #include "ll-merge.h"
23 #include "attr.h"
24 #include "merge-recursive.h"
25 #include "dir.h"
26 #include "submodule.h"
27 #include "revision.h"
28
29 struct path_hashmap_entry {
30         struct hashmap_entry e;
31         char path[FLEX_ARRAY];
32 };
33
34 static int path_hashmap_cmp(const void *cmp_data,
35                             const void *entry,
36                             const void *entry_or_key,
37                             const void *keydata)
38 {
39         const struct path_hashmap_entry *a = entry;
40         const struct path_hashmap_entry *b = entry_or_key;
41         const char *key = keydata;
42
43         if (ignore_case)
44                 return strcasecmp(a->path, key ? key : b->path);
45         else
46                 return strcmp(a->path, key ? key : b->path);
47 }
48
49 static unsigned int path_hash(const char *path)
50 {
51         return ignore_case ? strihash(path) : strhash(path);
52 }
53
54 static struct dir_rename_entry *dir_rename_find_entry(struct hashmap *hashmap,
55                                                       char *dir)
56 {
57         struct dir_rename_entry key;
58
59         if (dir == NULL)
60                 return NULL;
61         hashmap_entry_init(&key, strhash(dir));
62         key.dir = dir;
63         return hashmap_get(hashmap, &key, NULL);
64 }
65
66 static int dir_rename_cmp(const void *unused_cmp_data,
67                           const void *entry,
68                           const void *entry_or_key,
69                           const void *unused_keydata)
70 {
71         const struct dir_rename_entry *e1 = entry;
72         const struct dir_rename_entry *e2 = entry_or_key;
73
74         return strcmp(e1->dir, e2->dir);
75 }
76
77 static void dir_rename_init(struct hashmap *map)
78 {
79         hashmap_init(map, dir_rename_cmp, NULL, 0);
80 }
81
82 static void dir_rename_entry_init(struct dir_rename_entry *entry,
83                                   char *directory)
84 {
85         hashmap_entry_init(entry, strhash(directory));
86         entry->dir = directory;
87         entry->non_unique_new_dir = 0;
88         strbuf_init(&entry->new_dir, 0);
89         string_list_init(&entry->possible_new_dirs, 0);
90 }
91
92 static struct collision_entry *collision_find_entry(struct hashmap *hashmap,
93                                                     char *target_file)
94 {
95         struct collision_entry key;
96
97         hashmap_entry_init(&key, strhash(target_file));
98         key.target_file = target_file;
99         return hashmap_get(hashmap, &key, NULL);
100 }
101
102 static int collision_cmp(void *unused_cmp_data,
103                          const struct collision_entry *e1,
104                          const struct collision_entry *e2,
105                          const void *unused_keydata)
106 {
107         return strcmp(e1->target_file, e2->target_file);
108 }
109
110 static void collision_init(struct hashmap *map)
111 {
112         hashmap_init(map, (hashmap_cmp_fn) collision_cmp, NULL, 0);
113 }
114
115 static void flush_output(struct merge_options *o)
116 {
117         if (o->buffer_output < 2 && o->obuf.len) {
118                 fputs(o->obuf.buf, stdout);
119                 strbuf_reset(&o->obuf);
120         }
121 }
122
123 static int err(struct merge_options *o, const char *err, ...)
124 {
125         va_list params;
126
127         if (o->buffer_output < 2)
128                 flush_output(o);
129         else {
130                 strbuf_complete(&o->obuf, '\n');
131                 strbuf_addstr(&o->obuf, "error: ");
132         }
133         va_start(params, err);
134         strbuf_vaddf(&o->obuf, err, params);
135         va_end(params);
136         if (o->buffer_output > 1)
137                 strbuf_addch(&o->obuf, '\n');
138         else {
139                 error("%s", o->obuf.buf);
140                 strbuf_reset(&o->obuf);
141         }
142
143         return -1;
144 }
145
146 static struct tree *shift_tree_object(struct tree *one, struct tree *two,
147                                       const char *subtree_shift)
148 {
149         struct object_id shifted;
150
151         if (!*subtree_shift) {
152                 shift_tree(&one->object.oid, &two->object.oid, &shifted, 0);
153         } else {
154                 shift_tree_by(&one->object.oid, &two->object.oid, &shifted,
155                               subtree_shift);
156         }
157         if (!oidcmp(&two->object.oid, &shifted))
158                 return two;
159         return lookup_tree(&shifted);
160 }
161
162 static struct commit *make_virtual_commit(struct tree *tree, const char *comment)
163 {
164         struct commit *commit = alloc_commit_node(the_repository);
165
166         set_merge_remote_desc(commit, comment, (struct object *)commit);
167         commit->maybe_tree = tree;
168         commit->object.parsed = 1;
169         return commit;
170 }
171
172 /*
173  * Since we use get_tree_entry(), which does not put the read object into
174  * the object pool, we cannot rely on a == b.
175  */
176 static int oid_eq(const struct object_id *a, const struct object_id *b)
177 {
178         if (!a && !b)
179                 return 2;
180         return a && b && oidcmp(a, b) == 0;
181 }
182
183 enum rename_type {
184         RENAME_NORMAL = 0,
185         RENAME_VIA_DIR,
186         RENAME_DELETE,
187         RENAME_ONE_FILE_TO_ONE,
188         RENAME_ONE_FILE_TO_TWO,
189         RENAME_TWO_FILES_TO_ONE
190 };
191
192 struct rename_conflict_info {
193         enum rename_type rename_type;
194         struct diff_filepair *pair1;
195         struct diff_filepair *pair2;
196         const char *branch1;
197         const char *branch2;
198         struct stage_data *dst_entry1;
199         struct stage_data *dst_entry2;
200         struct diff_filespec ren1_other;
201         struct diff_filespec ren2_other;
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 {
210                 unsigned mode;
211                 struct object_id oid;
212         } stages[4];
213         struct rename_conflict_info *rename_conflict_info;
214         unsigned processed:1;
215 };
216
217 static inline void setup_rename_conflict_info(enum rename_type rename_type,
218                                               struct diff_filepair *pair1,
219                                               struct diff_filepair *pair2,
220                                               const char *branch1,
221                                               const char *branch2,
222                                               struct stage_data *dst_entry1,
223                                               struct stage_data *dst_entry2,
224                                               struct merge_options *o,
225                                               struct stage_data *src_entry1,
226                                               struct stage_data *src_entry2)
227 {
228         struct rename_conflict_info *ci = xcalloc(1, sizeof(struct rename_conflict_info));
229         ci->rename_type = rename_type;
230         ci->pair1 = pair1;
231         ci->branch1 = branch1;
232         ci->branch2 = branch2;
233
234         ci->dst_entry1 = dst_entry1;
235         dst_entry1->rename_conflict_info = ci;
236         dst_entry1->processed = 0;
237
238         assert(!pair2 == !dst_entry2);
239         if (dst_entry2) {
240                 ci->dst_entry2 = dst_entry2;
241                 ci->pair2 = pair2;
242                 dst_entry2->rename_conflict_info = ci;
243         }
244
245         if (rename_type == RENAME_TWO_FILES_TO_ONE) {
246                 /*
247                  * For each rename, there could have been
248                  * modifications on the side of history where that
249                  * file was not renamed.
250                  */
251                 int ostage1 = o->branch1 == branch1 ? 3 : 2;
252                 int ostage2 = ostage1 ^ 1;
253
254                 ci->ren1_other.path = pair1->one->path;
255                 oidcpy(&ci->ren1_other.oid, &src_entry1->stages[ostage1].oid);
256                 ci->ren1_other.mode = src_entry1->stages[ostage1].mode;
257
258                 ci->ren2_other.path = pair2->one->path;
259                 oidcpy(&ci->ren2_other.oid, &src_entry2->stages[ostage2].oid);
260                 ci->ren2_other.mode = src_entry2->stages[ostage2].mode;
261         }
262 }
263
264 static int show(struct merge_options *o, int v)
265 {
266         return (!o->call_depth && o->verbosity >= v) || o->verbosity >= 5;
267 }
268
269 __attribute__((format (printf, 3, 4)))
270 static void output(struct merge_options *o, int v, const char *fmt, ...)
271 {
272         va_list ap;
273
274         if (!show(o, v))
275                 return;
276
277         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
278
279         va_start(ap, fmt);
280         strbuf_vaddf(&o->obuf, fmt, ap);
281         va_end(ap);
282
283         strbuf_addch(&o->obuf, '\n');
284         if (!o->buffer_output)
285                 flush_output(o);
286 }
287
288 static void output_commit_title(struct merge_options *o, struct commit *commit)
289 {
290         struct merge_remote_desc *desc;
291
292         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
293         desc = merge_remote_util(commit);
294         if (desc)
295                 strbuf_addf(&o->obuf, "virtual %s\n", desc->name);
296         else {
297                 strbuf_add_unique_abbrev(&o->obuf, &commit->object.oid,
298                                          DEFAULT_ABBREV);
299                 strbuf_addch(&o->obuf, ' ');
300                 if (parse_commit(commit) != 0)
301                         strbuf_addstr(&o->obuf, _("(bad commit)\n"));
302                 else {
303                         const char *title;
304                         const char *msg = get_commit_buffer(commit, NULL);
305                         int len = find_commit_subject(msg, &title);
306                         if (len)
307                                 strbuf_addf(&o->obuf, "%.*s\n", len, title);
308                         unuse_commit_buffer(commit, msg);
309                 }
310         }
311         flush_output(o);
312 }
313
314 static int add_cacheinfo(struct merge_options *o,
315                          unsigned int mode, const struct object_id *oid,
316                          const char *path, int stage, int refresh, int options)
317 {
318         struct cache_entry *ce;
319         int ret;
320
321         ce = make_cache_entry(mode, oid ? oid->hash : null_sha1, path, stage, 0);
322         if (!ce)
323                 return err(o, _("add_cacheinfo failed for path '%s'; merge aborting."), path);
324
325         ret = add_cache_entry(ce, options);
326         if (refresh) {
327                 struct cache_entry *nce;
328
329                 nce = refresh_cache_entry(ce, CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
330                 if (!nce)
331                         return err(o, _("add_cacheinfo failed to refresh for path '%s'; merge aborting."), path);
332                 if (nce != ce)
333                         ret = add_cache_entry(nce, options);
334         }
335         return ret;
336 }
337
338 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
339 {
340         parse_tree(tree);
341         init_tree_desc(desc, tree->buffer, tree->size);
342 }
343
344 static int unpack_trees_start(struct merge_options *o,
345                               struct tree *common,
346                               struct tree *head,
347                               struct tree *merge)
348 {
349         int rc;
350         struct tree_desc t[3];
351         struct index_state tmp_index = { NULL };
352
353         memset(&o->unpack_opts, 0, sizeof(o->unpack_opts));
354         if (o->call_depth)
355                 o->unpack_opts.index_only = 1;
356         else
357                 o->unpack_opts.update = 1;
358         o->unpack_opts.merge = 1;
359         o->unpack_opts.head_idx = 2;
360         o->unpack_opts.fn = threeway_merge;
361         o->unpack_opts.src_index = &the_index;
362         o->unpack_opts.dst_index = &tmp_index;
363         o->unpack_opts.aggressive = !merge_detect_rename(o);
364         setup_unpack_trees_porcelain(&o->unpack_opts, "merge");
365
366         init_tree_desc_from_tree(t+0, common);
367         init_tree_desc_from_tree(t+1, head);
368         init_tree_desc_from_tree(t+2, merge);
369
370         rc = unpack_trees(3, t, &o->unpack_opts);
371         cache_tree_free(&active_cache_tree);
372
373         /*
374          * Update the_index to match the new results, AFTER saving a copy
375          * in o->orig_index.  Update src_index to point to the saved copy.
376          * (verify_uptodate() checks src_index, and the original index is
377          * the one that had the necessary modification timestamps.)
378          */
379         o->orig_index = the_index;
380         the_index = tmp_index;
381         o->unpack_opts.src_index = &o->orig_index;
382
383         return rc;
384 }
385
386 static void unpack_trees_finish(struct merge_options *o)
387 {
388         discard_index(&o->orig_index);
389         clear_unpack_trees_porcelain(&o->unpack_opts);
390 }
391
392 struct tree *write_tree_from_memory(struct merge_options *o)
393 {
394         struct tree *result = NULL;
395
396         if (unmerged_cache()) {
397                 int i;
398                 fprintf(stderr, "BUG: There are unmerged index entries:\n");
399                 for (i = 0; i < active_nr; i++) {
400                         const struct cache_entry *ce = active_cache[i];
401                         if (ce_stage(ce))
402                                 fprintf(stderr, "BUG: %d %.*s\n", ce_stage(ce),
403                                         (int)ce_namelen(ce), ce->name);
404                 }
405                 BUG("unmerged index entries in merge-recursive.c");
406         }
407
408         if (!active_cache_tree)
409                 active_cache_tree = cache_tree();
410
411         if (!cache_tree_fully_valid(active_cache_tree) &&
412             cache_tree_update(&the_index, 0) < 0) {
413                 err(o, _("error building trees"));
414                 return NULL;
415         }
416
417         result = lookup_tree(&active_cache_tree->oid);
418
419         return result;
420 }
421
422 static int save_files_dirs(const struct object_id *oid,
423                            struct strbuf *base, const char *path,
424                            unsigned int mode, int stage, void *context)
425 {
426         struct path_hashmap_entry *entry;
427         int baselen = base->len;
428         struct merge_options *o = context;
429
430         strbuf_addstr(base, path);
431
432         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
433         hashmap_entry_init(entry, path_hash(entry->path));
434         hashmap_add(&o->current_file_dir_set, entry);
435
436         strbuf_setlen(base, baselen);
437         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
438 }
439
440 static void get_files_dirs(struct merge_options *o, struct tree *tree)
441 {
442         struct pathspec match_all;
443         memset(&match_all, 0, sizeof(match_all));
444         read_tree_recursive(tree, "", 0, 0, &match_all, save_files_dirs, o);
445 }
446
447 static int get_tree_entry_if_blob(const struct object_id *tree,
448                                   const char *path,
449                                   struct object_id *hashy,
450                                   unsigned int *mode_o)
451 {
452         int ret;
453
454         ret = get_tree_entry(tree, path, hashy, mode_o);
455         if (S_ISDIR(*mode_o)) {
456                 oidcpy(hashy, &null_oid);
457                 *mode_o = 0;
458         }
459         return ret;
460 }
461
462 /*
463  * Returns an index_entry instance which doesn't have to correspond to
464  * a real cache entry in Git's index.
465  */
466 static struct stage_data *insert_stage_data(const char *path,
467                 struct tree *o, struct tree *a, struct tree *b,
468                 struct string_list *entries)
469 {
470         struct string_list_item *item;
471         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
472         get_tree_entry_if_blob(&o->object.oid, path,
473                                &e->stages[1].oid, &e->stages[1].mode);
474         get_tree_entry_if_blob(&a->object.oid, path,
475                                &e->stages[2].oid, &e->stages[2].mode);
476         get_tree_entry_if_blob(&b->object.oid, path,
477                                &e->stages[3].oid, &e->stages[3].mode);
478         item = string_list_insert(entries, path);
479         item->util = e;
480         return e;
481 }
482
483 /*
484  * Create a dictionary mapping file names to stage_data objects. The
485  * dictionary contains one entry for every path with a non-zero stage entry.
486  */
487 static struct string_list *get_unmerged(void)
488 {
489         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
490         int i;
491
492         unmerged->strdup_strings = 1;
493
494         for (i = 0; i < active_nr; i++) {
495                 struct string_list_item *item;
496                 struct stage_data *e;
497                 const struct cache_entry *ce = active_cache[i];
498                 if (!ce_stage(ce))
499                         continue;
500
501                 item = string_list_lookup(unmerged, ce->name);
502                 if (!item) {
503                         item = string_list_insert(unmerged, ce->name);
504                         item->util = xcalloc(1, sizeof(struct stage_data));
505                 }
506                 e = item->util;
507                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
508                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
509         }
510
511         return unmerged;
512 }
513
514 static int string_list_df_name_compare(const char *one, const char *two)
515 {
516         int onelen = strlen(one);
517         int twolen = strlen(two);
518         /*
519          * Here we only care that entries for D/F conflicts are
520          * adjacent, in particular with the file of the D/F conflict
521          * appearing before files below the corresponding directory.
522          * The order of the rest of the list is irrelevant for us.
523          *
524          * To achieve this, we sort with df_name_compare and provide
525          * the mode S_IFDIR so that D/F conflicts will sort correctly.
526          * We use the mode S_IFDIR for everything else for simplicity,
527          * since in other cases any changes in their order due to
528          * sorting cause no problems for us.
529          */
530         int cmp = df_name_compare(one, onelen, S_IFDIR,
531                                   two, twolen, S_IFDIR);
532         /*
533          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
534          * that 'foo' comes before 'foo/bar'.
535          */
536         if (cmp)
537                 return cmp;
538         return onelen - twolen;
539 }
540
541 static void record_df_conflict_files(struct merge_options *o,
542                                      struct string_list *entries)
543 {
544         /* If there is a D/F conflict and the file for such a conflict
545          * currently exists in the working tree, we want to allow it to be
546          * removed to make room for the corresponding directory if needed.
547          * The files underneath the directories of such D/F conflicts will
548          * be processed before the corresponding file involved in the D/F
549          * conflict.  If the D/F directory ends up being removed by the
550          * merge, then we won't have to touch the D/F file.  If the D/F
551          * directory needs to be written to the working copy, then the D/F
552          * file will simply be removed (in make_room_for_path()) to make
553          * room for the necessary paths.  Note that if both the directory
554          * and the file need to be present, then the D/F file will be
555          * reinstated with a new unique name at the time it is processed.
556          */
557         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
558         const char *last_file = NULL;
559         int last_len = 0;
560         int i;
561
562         /*
563          * If we're merging merge-bases, we don't want to bother with
564          * any working directory changes.
565          */
566         if (o->call_depth)
567                 return;
568
569         /* Ensure D/F conflicts are adjacent in the entries list. */
570         for (i = 0; i < entries->nr; i++) {
571                 struct string_list_item *next = &entries->items[i];
572                 string_list_append(&df_sorted_entries, next->string)->util =
573                                    next->util;
574         }
575         df_sorted_entries.cmp = string_list_df_name_compare;
576         string_list_sort(&df_sorted_entries);
577
578         string_list_clear(&o->df_conflict_file_set, 1);
579         for (i = 0; i < df_sorted_entries.nr; i++) {
580                 const char *path = df_sorted_entries.items[i].string;
581                 int len = strlen(path);
582                 struct stage_data *e = df_sorted_entries.items[i].util;
583
584                 /*
585                  * Check if last_file & path correspond to a D/F conflict;
586                  * i.e. whether path is last_file+'/'+<something>.
587                  * If so, record that it's okay to remove last_file to make
588                  * room for path and friends if needed.
589                  */
590                 if (last_file &&
591                     len > last_len &&
592                     memcmp(path, last_file, last_len) == 0 &&
593                     path[last_len] == '/') {
594                         string_list_insert(&o->df_conflict_file_set, last_file);
595                 }
596
597                 /*
598                  * Determine whether path could exist as a file in the
599                  * working directory as a possible D/F conflict.  This
600                  * will only occur when it exists in stage 2 as a
601                  * file.
602                  */
603                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
604                         last_file = path;
605                         last_len = len;
606                 } else {
607                         last_file = NULL;
608                 }
609         }
610         string_list_clear(&df_sorted_entries, 0);
611 }
612
613 struct rename {
614         struct diff_filepair *pair;
615         /*
616          * Purpose of src_entry and dst_entry:
617          *
618          * If 'before' is renamed to 'after' then src_entry will contain
619          * the versions of 'before' from the merge_base, HEAD, and MERGE in
620          * stages 1, 2, and 3; dst_entry will contain the respective
621          * versions of 'after' in corresponding locations.  Thus, we have a
622          * total of six modes and oids, though some will be null.  (Stage 0
623          * is ignored; we're interested in handling conflicts.)
624          *
625          * Since we don't turn on break-rewrites by default, neither
626          * src_entry nor dst_entry can have all three of their stages have
627          * non-null oids, meaning at most four of the six will be non-null.
628          * Also, since this is a rename, both src_entry and dst_entry will
629          * have at least one non-null oid, meaning at least two will be
630          * non-null.  Of the six oids, a typical rename will have three be
631          * non-null.  Only two implies a rename/delete, and four implies a
632          * rename/add.
633          */
634         struct stage_data *src_entry;
635         struct stage_data *dst_entry;
636         unsigned add_turned_into_rename:1;
637         unsigned processed:1;
638 };
639
640 static int update_stages(struct merge_options *opt, const char *path,
641                          const struct diff_filespec *o,
642                          const struct diff_filespec *a,
643                          const struct diff_filespec *b)
644 {
645
646         /*
647          * NOTE: It is usually a bad idea to call update_stages on a path
648          * before calling update_file on that same path, since it can
649          * sometimes lead to spurious "refusing to lose untracked file..."
650          * messages from update_file (via make_room_for path via
651          * would_lose_untracked).  Instead, reverse the order of the calls
652          * (executing update_file first and then update_stages).
653          */
654         int clear = 1;
655         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
656         if (clear)
657                 if (remove_file_from_cache(path))
658                         return -1;
659         if (o)
660                 if (add_cacheinfo(opt, o->mode, &o->oid, path, 1, 0, options))
661                         return -1;
662         if (a)
663                 if (add_cacheinfo(opt, a->mode, &a->oid, path, 2, 0, options))
664                         return -1;
665         if (b)
666                 if (add_cacheinfo(opt, b->mode, &b->oid, path, 3, 0, options))
667                         return -1;
668         return 0;
669 }
670
671 static int update_stages_for_stage_data(struct merge_options *opt,
672                                         const char *path,
673                                         const struct stage_data *stage_data)
674 {
675         struct diff_filespec o, a, b;
676
677         o.mode = stage_data->stages[1].mode;
678         oidcpy(&o.oid, &stage_data->stages[1].oid);
679
680         a.mode = stage_data->stages[2].mode;
681         oidcpy(&a.oid, &stage_data->stages[2].oid);
682
683         b.mode = stage_data->stages[3].mode;
684         oidcpy(&b.oid, &stage_data->stages[3].oid);
685
686         return update_stages(opt, path,
687                              is_null_oid(&o.oid) ? NULL : &o,
688                              is_null_oid(&a.oid) ? NULL : &a,
689                              is_null_oid(&b.oid) ? NULL : &b);
690 }
691
692 static void update_entry(struct stage_data *entry,
693                          struct diff_filespec *o,
694                          struct diff_filespec *a,
695                          struct diff_filespec *b)
696 {
697         entry->processed = 0;
698         entry->stages[1].mode = o->mode;
699         entry->stages[2].mode = a->mode;
700         entry->stages[3].mode = b->mode;
701         oidcpy(&entry->stages[1].oid, &o->oid);
702         oidcpy(&entry->stages[2].oid, &a->oid);
703         oidcpy(&entry->stages[3].oid, &b->oid);
704 }
705
706 static int remove_file(struct merge_options *o, int clean,
707                        const char *path, int no_wd)
708 {
709         int update_cache = o->call_depth || clean;
710         int update_working_directory = !o->call_depth && !no_wd;
711
712         if (update_cache) {
713                 if (remove_file_from_cache(path))
714                         return -1;
715         }
716         if (update_working_directory) {
717                 if (ignore_case) {
718                         struct cache_entry *ce;
719                         ce = cache_file_exists(path, strlen(path), ignore_case);
720                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
721                                 return 0;
722                 }
723                 if (remove_path(path))
724                         return -1;
725         }
726         return 0;
727 }
728
729 /* add a string to a strbuf, but converting "/" to "_" */
730 static void add_flattened_path(struct strbuf *out, const char *s)
731 {
732         size_t i = out->len;
733         strbuf_addstr(out, s);
734         for (; i < out->len; i++)
735                 if (out->buf[i] == '/')
736                         out->buf[i] = '_';
737 }
738
739 static char *unique_path(struct merge_options *o, const char *path, const char *branch)
740 {
741         struct path_hashmap_entry *entry;
742         struct strbuf newpath = STRBUF_INIT;
743         int suffix = 0;
744         size_t base_len;
745
746         strbuf_addf(&newpath, "%s~", path);
747         add_flattened_path(&newpath, branch);
748
749         base_len = newpath.len;
750         while (hashmap_get_from_hash(&o->current_file_dir_set,
751                                      path_hash(newpath.buf), newpath.buf) ||
752                (!o->call_depth && file_exists(newpath.buf))) {
753                 strbuf_setlen(&newpath, base_len);
754                 strbuf_addf(&newpath, "_%d", suffix++);
755         }
756
757         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
758         hashmap_entry_init(entry, path_hash(entry->path));
759         hashmap_add(&o->current_file_dir_set, entry);
760         return strbuf_detach(&newpath, NULL);
761 }
762
763 /**
764  * Check whether a directory in the index is in the way of an incoming
765  * file.  Return 1 if so.  If check_working_copy is non-zero, also
766  * check the working directory.  If empty_ok is non-zero, also return
767  * 0 in the case where the working-tree dir exists but is empty.
768  */
769 static int dir_in_way(const char *path, int check_working_copy, int empty_ok)
770 {
771         int pos;
772         struct strbuf dirpath = STRBUF_INIT;
773         struct stat st;
774
775         strbuf_addstr(&dirpath, path);
776         strbuf_addch(&dirpath, '/');
777
778         pos = cache_name_pos(dirpath.buf, dirpath.len);
779
780         if (pos < 0)
781                 pos = -1 - pos;
782         if (pos < active_nr &&
783             !strncmp(dirpath.buf, active_cache[pos]->name, dirpath.len)) {
784                 strbuf_release(&dirpath);
785                 return 1;
786         }
787
788         strbuf_release(&dirpath);
789         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
790                 !(empty_ok && is_empty_dir(path));
791 }
792
793 /*
794  * Returns whether path was tracked in the index before the merge started,
795  * and its oid and mode match the specified values
796  */
797 static int was_tracked_and_matches(struct merge_options *o, const char *path,
798                                    const struct object_id *oid, unsigned mode)
799 {
800         int pos = index_name_pos(&o->orig_index, path, strlen(path));
801         struct cache_entry *ce;
802
803         if (0 > pos)
804                 /* we were not tracking this path before the merge */
805                 return 0;
806
807         /* See if the file we were tracking before matches */
808         ce = o->orig_index.cache[pos];
809         return (oid_eq(&ce->oid, oid) && ce->ce_mode == mode);
810 }
811
812 /*
813  * Returns whether path was tracked in the index before the merge started
814  */
815 static int was_tracked(struct merge_options *o, const char *path)
816 {
817         int pos = index_name_pos(&o->orig_index, path, strlen(path));
818
819         if (0 <= pos)
820                 /* we were tracking this path before the merge */
821                 return 1;
822
823         return 0;
824 }
825
826 static int would_lose_untracked(const char *path)
827 {
828         /*
829          * This may look like it can be simplified to:
830          *   return !was_tracked(o, path) && file_exists(path)
831          * but it can't.  This function needs to know whether path was in
832          * the working tree due to EITHER having been tracked in the index
833          * before the merge OR having been put into the working copy and
834          * index by unpack_trees().  Due to that either-or requirement, we
835          * check the current index instead of the original one.
836          *
837          * Note that we do not need to worry about merge-recursive itself
838          * updating the index after unpack_trees() and before calling this
839          * function, because we strictly require all code paths in
840          * merge-recursive to update the working tree first and the index
841          * second.  Doing otherwise would break
842          * update_file()/would_lose_untracked(); see every comment in this
843          * file which mentions "update_stages".
844          */
845         int pos = cache_name_pos(path, strlen(path));
846
847         if (pos < 0)
848                 pos = -1 - pos;
849         while (pos < active_nr &&
850                !strcmp(path, active_cache[pos]->name)) {
851                 /*
852                  * If stage #0, it is definitely tracked.
853                  * If it has stage #2 then it was tracked
854                  * before this merge started.  All other
855                  * cases the path was not tracked.
856                  */
857                 switch (ce_stage(active_cache[pos])) {
858                 case 0:
859                 case 2:
860                         return 0;
861                 }
862                 pos++;
863         }
864         return file_exists(path);
865 }
866
867 static int was_dirty(struct merge_options *o, const char *path)
868 {
869         struct cache_entry *ce;
870         int dirty = 1;
871
872         if (o->call_depth || !was_tracked(o, path))
873                 return !dirty;
874
875         ce = index_file_exists(o->unpack_opts.src_index,
876                                path, strlen(path), ignore_case);
877         dirty = verify_uptodate(ce, &o->unpack_opts) != 0;
878         return dirty;
879 }
880
881 static int make_room_for_path(struct merge_options *o, const char *path)
882 {
883         int status, i;
884         const char *msg = _("failed to create path '%s'%s");
885
886         /* Unlink any D/F conflict files that are in the way */
887         for (i = 0; i < o->df_conflict_file_set.nr; i++) {
888                 const char *df_path = o->df_conflict_file_set.items[i].string;
889                 size_t pathlen = strlen(path);
890                 size_t df_pathlen = strlen(df_path);
891                 if (df_pathlen < pathlen &&
892                     path[df_pathlen] == '/' &&
893                     strncmp(path, df_path, df_pathlen) == 0) {
894                         output(o, 3,
895                                _("Removing %s to make room for subdirectory\n"),
896                                df_path);
897                         unlink(df_path);
898                         unsorted_string_list_delete_item(&o->df_conflict_file_set,
899                                                          i, 0);
900                         break;
901                 }
902         }
903
904         /* Make sure leading directories are created */
905         status = safe_create_leading_directories_const(path);
906         if (status) {
907                 if (status == SCLD_EXISTS)
908                         /* something else exists */
909                         return err(o, msg, path, _(": perhaps a D/F conflict?"));
910                 return err(o, msg, path, "");
911         }
912
913         /*
914          * Do not unlink a file in the work tree if we are not
915          * tracking it.
916          */
917         if (would_lose_untracked(path))
918                 return err(o, _("refusing to lose untracked file at '%s'"),
919                            path);
920
921         /* Successful unlink is good.. */
922         if (!unlink(path))
923                 return 0;
924         /* .. and so is no existing file */
925         if (errno == ENOENT)
926                 return 0;
927         /* .. but not some other error (who really cares what?) */
928         return err(o, msg, path, _(": perhaps a D/F conflict?"));
929 }
930
931 static int update_file_flags(struct merge_options *o,
932                              const struct object_id *oid,
933                              unsigned mode,
934                              const char *path,
935                              int update_cache,
936                              int update_wd)
937 {
938         int ret = 0;
939
940         if (o->call_depth)
941                 update_wd = 0;
942
943         if (update_wd) {
944                 enum object_type type;
945                 void *buf;
946                 unsigned long size;
947
948                 if (S_ISGITLINK(mode)) {
949                         /*
950                          * We may later decide to recursively descend into
951                          * the submodule directory and update its index
952                          * and/or work tree, but we do not do that now.
953                          */
954                         update_wd = 0;
955                         goto update_index;
956                 }
957
958                 buf = read_object_file(oid, &type, &size);
959                 if (!buf)
960                         return err(o, _("cannot read object %s '%s'"), oid_to_hex(oid), path);
961                 if (type != OBJ_BLOB) {
962                         ret = err(o, _("blob expected for %s '%s'"), oid_to_hex(oid), path);
963                         goto free_buf;
964                 }
965                 if (S_ISREG(mode)) {
966                         struct strbuf strbuf = STRBUF_INIT;
967                         if (convert_to_working_tree(path, buf, size, &strbuf)) {
968                                 free(buf);
969                                 size = strbuf.len;
970                                 buf = strbuf_detach(&strbuf, NULL);
971                         }
972                 }
973
974                 if (make_room_for_path(o, path) < 0) {
975                         update_wd = 0;
976                         goto free_buf;
977                 }
978                 if (S_ISREG(mode) || (!has_symlinks && S_ISLNK(mode))) {
979                         int fd;
980                         if (mode & 0100)
981                                 mode = 0777;
982                         else
983                                 mode = 0666;
984                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
985                         if (fd < 0) {
986                                 ret = err(o, _("failed to open '%s': %s"),
987                                           path, strerror(errno));
988                                 goto free_buf;
989                         }
990                         write_in_full(fd, buf, size);
991                         close(fd);
992                 } else if (S_ISLNK(mode)) {
993                         char *lnk = xmemdupz(buf, size);
994                         safe_create_leading_directories_const(path);
995                         unlink(path);
996                         if (symlink(lnk, path))
997                                 ret = err(o, _("failed to symlink '%s': %s"),
998                                           path, strerror(errno));
999                         free(lnk);
1000                 } else
1001                         ret = err(o,
1002                                   _("do not know what to do with %06o %s '%s'"),
1003                                   mode, oid_to_hex(oid), path);
1004         free_buf:
1005                 free(buf);
1006         }
1007 update_index:
1008         if (!ret && update_cache)
1009                 if (add_cacheinfo(o, mode, oid, path, 0, update_wd,
1010                                   ADD_CACHE_OK_TO_ADD))
1011                         return -1;
1012         return ret;
1013 }
1014
1015 static int update_file(struct merge_options *o,
1016                        int clean,
1017                        const struct object_id *oid,
1018                        unsigned mode,
1019                        const char *path)
1020 {
1021         return update_file_flags(o, oid, mode, path, o->call_depth || clean, !o->call_depth);
1022 }
1023
1024 /* Low level file merging, update and removal */
1025
1026 struct merge_file_info {
1027         struct object_id oid;
1028         unsigned mode;
1029         unsigned clean:1,
1030                  merge:1;
1031 };
1032
1033 static int merge_3way(struct merge_options *o,
1034                       mmbuffer_t *result_buf,
1035                       const struct diff_filespec *one,
1036                       const struct diff_filespec *a,
1037                       const struct diff_filespec *b,
1038                       const char *branch1,
1039                       const char *branch2)
1040 {
1041         mmfile_t orig, src1, src2;
1042         struct ll_merge_options ll_opts = {0};
1043         char *base_name, *name1, *name2;
1044         int merge_status;
1045
1046         ll_opts.renormalize = o->renormalize;
1047         ll_opts.xdl_opts = o->xdl_opts;
1048
1049         if (o->call_depth) {
1050                 ll_opts.virtual_ancestor = 1;
1051                 ll_opts.variant = 0;
1052         } else {
1053                 switch (o->recursive_variant) {
1054                 case MERGE_RECURSIVE_OURS:
1055                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
1056                         break;
1057                 case MERGE_RECURSIVE_THEIRS:
1058                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
1059                         break;
1060                 default:
1061                         ll_opts.variant = 0;
1062                         break;
1063                 }
1064         }
1065
1066         if (strcmp(a->path, b->path) ||
1067             (o->ancestor != NULL && strcmp(a->path, one->path) != 0)) {
1068                 base_name = o->ancestor == NULL ? NULL :
1069                         mkpathdup("%s:%s", o->ancestor, one->path);
1070                 name1 = mkpathdup("%s:%s", branch1, a->path);
1071                 name2 = mkpathdup("%s:%s", branch2, b->path);
1072         } else {
1073                 base_name = o->ancestor == NULL ? NULL :
1074                         mkpathdup("%s", o->ancestor);
1075                 name1 = mkpathdup("%s", branch1);
1076                 name2 = mkpathdup("%s", branch2);
1077         }
1078
1079         read_mmblob(&orig, &one->oid);
1080         read_mmblob(&src1, &a->oid);
1081         read_mmblob(&src2, &b->oid);
1082
1083         merge_status = ll_merge(result_buf, a->path, &orig, base_name,
1084                                 &src1, name1, &src2, name2, &ll_opts);
1085
1086         free(base_name);
1087         free(name1);
1088         free(name2);
1089         free(orig.ptr);
1090         free(src1.ptr);
1091         free(src2.ptr);
1092         return merge_status;
1093 }
1094
1095 static int find_first_merges(struct object_array *result, const char *path,
1096                              struct commit *a, struct commit *b)
1097 {
1098         int i, j;
1099         struct object_array merges = OBJECT_ARRAY_INIT;
1100         struct commit *commit;
1101         int contains_another;
1102
1103         char merged_revision[42];
1104         const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
1105                                    "--all", merged_revision, NULL };
1106         struct rev_info revs;
1107         struct setup_revision_opt rev_opts;
1108
1109         memset(result, 0, sizeof(struct object_array));
1110         memset(&rev_opts, 0, sizeof(rev_opts));
1111
1112         /* get all revisions that merge commit a */
1113         xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
1114                   oid_to_hex(&a->object.oid));
1115         init_revisions(&revs, NULL);
1116         rev_opts.submodule = path;
1117         /* FIXME: can't handle linked worktrees in submodules yet */
1118         revs.single_worktree = path != NULL;
1119         setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
1120
1121         /* save all revisions from the above list that contain b */
1122         if (prepare_revision_walk(&revs))
1123                 die("revision walk setup failed");
1124         while ((commit = get_revision(&revs)) != NULL) {
1125                 struct object *o = &(commit->object);
1126                 if (in_merge_bases(b, commit))
1127                         add_object_array(o, NULL, &merges);
1128         }
1129         reset_revision_walk();
1130
1131         /* Now we've got all merges that contain a and b. Prune all
1132          * merges that contain another found merge and save them in
1133          * result.
1134          */
1135         for (i = 0; i < merges.nr; i++) {
1136                 struct commit *m1 = (struct commit *) merges.objects[i].item;
1137
1138                 contains_another = 0;
1139                 for (j = 0; j < merges.nr; j++) {
1140                         struct commit *m2 = (struct commit *) merges.objects[j].item;
1141                         if (i != j && in_merge_bases(m2, m1)) {
1142                                 contains_another = 1;
1143                                 break;
1144                         }
1145                 }
1146
1147                 if (!contains_another)
1148                         add_object_array(merges.objects[i].item, NULL, result);
1149         }
1150
1151         object_array_clear(&merges);
1152         return result->nr;
1153 }
1154
1155 static void print_commit(struct commit *commit)
1156 {
1157         struct strbuf sb = STRBUF_INIT;
1158         struct pretty_print_context ctx = {0};
1159         ctx.date_mode.type = DATE_NORMAL;
1160         format_commit_message(commit, " %h: %m %s", &sb, &ctx);
1161         fprintf(stderr, "%s\n", sb.buf);
1162         strbuf_release(&sb);
1163 }
1164
1165 static int merge_submodule(struct merge_options *o,
1166                            struct object_id *result, const char *path,
1167                            const struct object_id *base, const struct object_id *a,
1168                            const struct object_id *b)
1169 {
1170         struct commit *commit_base, *commit_a, *commit_b;
1171         int parent_count;
1172         struct object_array merges;
1173
1174         int i;
1175         int search = !o->call_depth;
1176
1177         /* store a in result in case we fail */
1178         oidcpy(result, a);
1179
1180         /* we can not handle deletion conflicts */
1181         if (is_null_oid(base))
1182                 return 0;
1183         if (is_null_oid(a))
1184                 return 0;
1185         if (is_null_oid(b))
1186                 return 0;
1187
1188         if (add_submodule_odb(path)) {
1189                 output(o, 1, _("Failed to merge submodule %s (not checked out)"), path);
1190                 return 0;
1191         }
1192
1193         if (!(commit_base = lookup_commit_reference(base)) ||
1194             !(commit_a = lookup_commit_reference(a)) ||
1195             !(commit_b = lookup_commit_reference(b))) {
1196                 output(o, 1, _("Failed to merge submodule %s (commits not present)"), path);
1197                 return 0;
1198         }
1199
1200         /* check whether both changes are forward */
1201         if (!in_merge_bases(commit_base, commit_a) ||
1202             !in_merge_bases(commit_base, commit_b)) {
1203                 output(o, 1, _("Failed to merge submodule %s (commits don't follow merge-base)"), path);
1204                 return 0;
1205         }
1206
1207         /* Case #1: a is contained in b or vice versa */
1208         if (in_merge_bases(commit_a, commit_b)) {
1209                 oidcpy(result, b);
1210                 if (show(o, 3)) {
1211                         output(o, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1212                         output_commit_title(o, commit_b);
1213                 } else if (show(o, 2))
1214                         output(o, 2, _("Fast-forwarding submodule %s"), path);
1215                 else
1216                         ; /* no output */
1217
1218                 return 1;
1219         }
1220         if (in_merge_bases(commit_b, commit_a)) {
1221                 oidcpy(result, a);
1222                 if (show(o, 3)) {
1223                         output(o, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1224                         output_commit_title(o, commit_a);
1225                 } else if (show(o, 2))
1226                         output(o, 2, _("Fast-forwarding submodule %s"), path);
1227                 else
1228                         ; /* no output */
1229
1230                 return 1;
1231         }
1232
1233         /*
1234          * Case #2: There are one or more merges that contain a and b in
1235          * the submodule. If there is only one, then present it as a
1236          * suggestion to the user, but leave it marked unmerged so the
1237          * user needs to confirm the resolution.
1238          */
1239
1240         /* Skip the search if makes no sense to the calling context.  */
1241         if (!search)
1242                 return 0;
1243
1244         /* find commit which merges them */
1245         parent_count = find_first_merges(&merges, path, commit_a, commit_b);
1246         switch (parent_count) {
1247         case 0:
1248                 output(o, 1, _("Failed to merge submodule %s (merge following commits not found)"), path);
1249                 break;
1250
1251         case 1:
1252                 output(o, 1, _("Failed to merge submodule %s (not fast-forward)"), path);
1253                 output(o, 2, _("Found a possible merge resolution for the submodule:\n"));
1254                 print_commit((struct commit *) merges.objects[0].item);
1255                 output(o, 2, _(
1256                        "If this is correct simply add it to the index "
1257                        "for example\n"
1258                        "by using:\n\n"
1259                        "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
1260                        "which will accept this suggestion.\n"),
1261                        oid_to_hex(&merges.objects[0].item->oid), path);
1262                 break;
1263
1264         default:
1265                 output(o, 1, _("Failed to merge submodule %s (multiple merges found)"), path);
1266                 for (i = 0; i < merges.nr; i++)
1267                         print_commit((struct commit *) merges.objects[i].item);
1268         }
1269
1270         object_array_clear(&merges);
1271         return 0;
1272 }
1273
1274 static int merge_file_1(struct merge_options *o,
1275                         const struct diff_filespec *one,
1276                         const struct diff_filespec *a,
1277                         const struct diff_filespec *b,
1278                         const char *filename,
1279                         const char *branch1,
1280                         const char *branch2,
1281                         struct merge_file_info *result)
1282 {
1283         result->merge = 0;
1284         result->clean = 1;
1285
1286         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
1287                 result->clean = 0;
1288                 if (S_ISREG(a->mode)) {
1289                         result->mode = a->mode;
1290                         oidcpy(&result->oid, &a->oid);
1291                 } else {
1292                         result->mode = b->mode;
1293                         oidcpy(&result->oid, &b->oid);
1294                 }
1295         } else {
1296                 if (!oid_eq(&a->oid, &one->oid) && !oid_eq(&b->oid, &one->oid))
1297                         result->merge = 1;
1298
1299                 /*
1300                  * Merge modes
1301                  */
1302                 if (a->mode == b->mode || a->mode == one->mode)
1303                         result->mode = b->mode;
1304                 else {
1305                         result->mode = a->mode;
1306                         if (b->mode != one->mode) {
1307                                 result->clean = 0;
1308                                 result->merge = 1;
1309                         }
1310                 }
1311
1312                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &one->oid))
1313                         oidcpy(&result->oid, &b->oid);
1314                 else if (oid_eq(&b->oid, &one->oid))
1315                         oidcpy(&result->oid, &a->oid);
1316                 else if (S_ISREG(a->mode)) {
1317                         mmbuffer_t result_buf;
1318                         int ret = 0, merge_status;
1319
1320                         merge_status = merge_3way(o, &result_buf, one, a, b,
1321                                                   branch1, branch2);
1322
1323                         if ((merge_status < 0) || !result_buf.ptr)
1324                                 ret = err(o, _("Failed to execute internal merge"));
1325
1326                         if (!ret &&
1327                             write_object_file(result_buf.ptr, result_buf.size,
1328                                               blob_type, &result->oid))
1329                                 ret = err(o, _("Unable to add %s to database"),
1330                                           a->path);
1331
1332                         free(result_buf.ptr);
1333                         if (ret)
1334                                 return ret;
1335                         result->clean = (merge_status == 0);
1336                 } else if (S_ISGITLINK(a->mode)) {
1337                         result->clean = merge_submodule(o, &result->oid,
1338                                                         one->path,
1339                                                         &one->oid,
1340                                                         &a->oid,
1341                                                         &b->oid);
1342                 } else if (S_ISLNK(a->mode)) {
1343                         switch (o->recursive_variant) {
1344                         case MERGE_RECURSIVE_NORMAL:
1345                                 oidcpy(&result->oid, &a->oid);
1346                                 if (!oid_eq(&a->oid, &b->oid))
1347                                         result->clean = 0;
1348                                 break;
1349                         case MERGE_RECURSIVE_OURS:
1350                                 oidcpy(&result->oid, &a->oid);
1351                                 break;
1352                         case MERGE_RECURSIVE_THEIRS:
1353                                 oidcpy(&result->oid, &b->oid);
1354                                 break;
1355                         }
1356                 } else
1357                         BUG("unsupported object type in the tree");
1358         }
1359
1360         if (result->merge)
1361                 output(o, 2, _("Auto-merging %s"), filename);
1362
1363         return 0;
1364 }
1365
1366 static int merge_file_special_markers(struct merge_options *o,
1367                                       const struct diff_filespec *one,
1368                                       const struct diff_filespec *a,
1369                                       const struct diff_filespec *b,
1370                                       const char *target_filename,
1371                                       const char *branch1,
1372                                       const char *filename1,
1373                                       const char *branch2,
1374                                       const char *filename2,
1375                                       struct merge_file_info *mfi)
1376 {
1377         char *side1 = NULL;
1378         char *side2 = NULL;
1379         int ret;
1380
1381         if (filename1)
1382                 side1 = xstrfmt("%s:%s", branch1, filename1);
1383         if (filename2)
1384                 side2 = xstrfmt("%s:%s", branch2, filename2);
1385
1386         ret = merge_file_1(o, one, a, b, target_filename,
1387                            side1 ? side1 : branch1,
1388                            side2 ? side2 : branch2, mfi);
1389
1390         free(side1);
1391         free(side2);
1392         return ret;
1393 }
1394
1395 static int merge_file_one(struct merge_options *o,
1396                           const char *path,
1397                           const struct object_id *o_oid, int o_mode,
1398                           const struct object_id *a_oid, int a_mode,
1399                           const struct object_id *b_oid, int b_mode,
1400                           const char *branch1,
1401                           const char *branch2,
1402                           struct merge_file_info *mfi)
1403 {
1404         struct diff_filespec one, a, b;
1405
1406         one.path = a.path = b.path = (char *)path;
1407         oidcpy(&one.oid, o_oid);
1408         one.mode = o_mode;
1409         oidcpy(&a.oid, a_oid);
1410         a.mode = a_mode;
1411         oidcpy(&b.oid, b_oid);
1412         b.mode = b_mode;
1413         return merge_file_1(o, &one, &a, &b, path, branch1, branch2, mfi);
1414 }
1415
1416 static int handle_rename_via_dir(struct merge_options *o,
1417                                  struct diff_filepair *pair,
1418                                  const char *rename_branch,
1419                                  const char *other_branch)
1420 {
1421         /*
1422          * Handle file adds that need to be renamed due to directory rename
1423          * detection.  This differs from handle_rename_normal, because
1424          * there is no content merge to do; just move the file into the
1425          * desired final location.
1426          */
1427         const struct diff_filespec *dest = pair->two;
1428
1429         if (!o->call_depth && would_lose_untracked(dest->path)) {
1430                 char *alt_path = unique_path(o, dest->path, rename_branch);
1431
1432                 output(o, 1, _("Error: Refusing to lose untracked file at %s; "
1433                                "writing to %s instead."),
1434                        dest->path, alt_path);
1435                 /*
1436                  * Write the file in worktree at alt_path, but not in the
1437                  * index.  Instead, write to dest->path for the index but
1438                  * only at the higher appropriate stage.
1439                  */
1440                 if (update_file(o, 0, &dest->oid, dest->mode, alt_path))
1441                         return -1;
1442                 free(alt_path);
1443                 return update_stages(o, dest->path, NULL,
1444                                      rename_branch == o->branch1 ? dest : NULL,
1445                                      rename_branch == o->branch1 ? NULL : dest);
1446         }
1447
1448         /* Update dest->path both in index and in worktree */
1449         if (update_file(o, 1, &dest->oid, dest->mode, dest->path))
1450                 return -1;
1451         return 0;
1452 }
1453
1454 static int handle_change_delete(struct merge_options *o,
1455                                 const char *path, const char *old_path,
1456                                 const struct object_id *o_oid, int o_mode,
1457                                 const struct object_id *changed_oid,
1458                                 int changed_mode,
1459                                 const char *change_branch,
1460                                 const char *delete_branch,
1461                                 const char *change, const char *change_past)
1462 {
1463         char *alt_path = NULL;
1464         const char *update_path = path;
1465         int ret = 0;
1466
1467         if (dir_in_way(path, !o->call_depth, 0) ||
1468             (!o->call_depth && would_lose_untracked(path))) {
1469                 update_path = alt_path = unique_path(o, path, change_branch);
1470         }
1471
1472         if (o->call_depth) {
1473                 /*
1474                  * We cannot arbitrarily accept either a_sha or b_sha as
1475                  * correct; since there is no true "middle point" between
1476                  * them, simply reuse the base version for virtual merge base.
1477                  */
1478                 ret = remove_file_from_cache(path);
1479                 if (!ret)
1480                         ret = update_file(o, 0, o_oid, o_mode, update_path);
1481         } else {
1482                 /*
1483                  * Despite the four nearly duplicate messages and argument
1484                  * lists below and the ugliness of the nested if-statements,
1485                  * having complete messages makes the job easier for
1486                  * translators.
1487                  *
1488                  * The slight variance among the cases is due to the fact
1489                  * that:
1490                  *   1) directory/file conflicts (in effect if
1491                  *      !alt_path) could cause us to need to write the
1492                  *      file to a different path.
1493                  *   2) renames (in effect if !old_path) could mean that
1494                  *      there are two names for the path that the user
1495                  *      may know the file by.
1496                  */
1497                 if (!alt_path) {
1498                         if (!old_path) {
1499                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1500                                        "and %s in %s. Version %s of %s left in tree."),
1501                                        change, path, delete_branch, change_past,
1502                                        change_branch, change_branch, path);
1503                         } else {
1504                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1505                                        "and %s to %s in %s. Version %s of %s left in tree."),
1506                                        change, old_path, delete_branch, change_past, path,
1507                                        change_branch, change_branch, path);
1508                         }
1509                 } else {
1510                         if (!old_path) {
1511                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1512                                        "and %s in %s. Version %s of %s left in tree at %s."),
1513                                        change, path, delete_branch, change_past,
1514                                        change_branch, change_branch, path, alt_path);
1515                         } else {
1516                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1517                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1518                                        change, old_path, delete_branch, change_past, path,
1519                                        change_branch, change_branch, path, alt_path);
1520                         }
1521                 }
1522                 /*
1523                  * No need to call update_file() on path when change_branch ==
1524                  * o->branch1 && !alt_path, since that would needlessly touch
1525                  * path.  We could call update_file_flags() with update_cache=0
1526                  * and update_wd=0, but that's a no-op.
1527                  */
1528                 if (change_branch != o->branch1 || alt_path)
1529                         ret = update_file(o, 0, changed_oid, changed_mode, update_path);
1530         }
1531         free(alt_path);
1532
1533         return ret;
1534 }
1535
1536 static int handle_rename_delete(struct merge_options *o,
1537                                 struct diff_filepair *pair,
1538                                 const char *rename_branch,
1539                                 const char *delete_branch)
1540 {
1541         const struct diff_filespec *orig = pair->one;
1542         const struct diff_filespec *dest = pair->two;
1543
1544         if (handle_change_delete(o,
1545                                  o->call_depth ? orig->path : dest->path,
1546                                  o->call_depth ? NULL : orig->path,
1547                                  &orig->oid, orig->mode,
1548                                  &dest->oid, dest->mode,
1549                                  rename_branch, delete_branch,
1550                                  _("rename"), _("renamed")))
1551                 return -1;
1552
1553         if (o->call_depth)
1554                 return remove_file_from_cache(dest->path);
1555         else
1556                 return update_stages(o, dest->path, NULL,
1557                                      rename_branch == o->branch1 ? dest : NULL,
1558                                      rename_branch == o->branch1 ? NULL : dest);
1559 }
1560
1561 static struct diff_filespec *filespec_from_entry(struct diff_filespec *target,
1562                                                  struct stage_data *entry,
1563                                                  int stage)
1564 {
1565         struct object_id *oid = &entry->stages[stage].oid;
1566         unsigned mode = entry->stages[stage].mode;
1567         if (mode == 0 || is_null_oid(oid))
1568                 return NULL;
1569         oidcpy(&target->oid, oid);
1570         target->mode = mode;
1571         return target;
1572 }
1573
1574 static int handle_file(struct merge_options *o,
1575                         struct diff_filespec *rename,
1576                         int stage,
1577                         struct rename_conflict_info *ci)
1578 {
1579         char *dst_name = rename->path;
1580         struct stage_data *dst_entry;
1581         const char *cur_branch, *other_branch;
1582         struct diff_filespec other;
1583         struct diff_filespec *add;
1584         int ret;
1585
1586         if (stage == 2) {
1587                 dst_entry = ci->dst_entry1;
1588                 cur_branch = ci->branch1;
1589                 other_branch = ci->branch2;
1590         } else {
1591                 dst_entry = ci->dst_entry2;
1592                 cur_branch = ci->branch2;
1593                 other_branch = ci->branch1;
1594         }
1595
1596         add = filespec_from_entry(&other, dst_entry, stage ^ 1);
1597         if (add) {
1598                 int ren_src_was_dirty = was_dirty(o, rename->path);
1599                 char *add_name = unique_path(o, rename->path, other_branch);
1600                 if (update_file(o, 0, &add->oid, add->mode, add_name))
1601                         return -1;
1602
1603                 if (ren_src_was_dirty) {
1604                         output(o, 1, _("Refusing to lose dirty file at %s"),
1605                                rename->path);
1606                 }
1607                 /*
1608                  * Because the double negatives somehow keep confusing me...
1609                  *    1) update_wd iff !ren_src_was_dirty.
1610                  *    2) no_wd iff !update_wd
1611                  *    3) so, no_wd == !!ren_src_was_dirty == ren_src_was_dirty
1612                  */
1613                 remove_file(o, 0, rename->path, ren_src_was_dirty);
1614                 dst_name = unique_path(o, rename->path, cur_branch);
1615         } else {
1616                 if (dir_in_way(rename->path, !o->call_depth, 0)) {
1617                         dst_name = unique_path(o, rename->path, cur_branch);
1618                         output(o, 1, _("%s is a directory in %s adding as %s instead"),
1619                                rename->path, other_branch, dst_name);
1620                 } else if (!o->call_depth &&
1621                            would_lose_untracked(rename->path)) {
1622                         dst_name = unique_path(o, rename->path, cur_branch);
1623                         output(o, 1, _("Refusing to lose untracked file at %s; "
1624                                        "adding as %s instead"),
1625                                rename->path, dst_name);
1626                 }
1627         }
1628         if ((ret = update_file(o, 0, &rename->oid, rename->mode, dst_name)))
1629                 ; /* fall through, do allow dst_name to be released */
1630         else if (stage == 2)
1631                 ret = update_stages(o, rename->path, NULL, rename, add);
1632         else
1633                 ret = update_stages(o, rename->path, NULL, add, rename);
1634
1635         if (dst_name != rename->path)
1636                 free(dst_name);
1637
1638         return ret;
1639 }
1640
1641 static int handle_rename_rename_1to2(struct merge_options *o,
1642                                      struct rename_conflict_info *ci)
1643 {
1644         /* One file was renamed in both branches, but to different names. */
1645         struct diff_filespec *one = ci->pair1->one;
1646         struct diff_filespec *a = ci->pair1->two;
1647         struct diff_filespec *b = ci->pair2->two;
1648
1649         output(o, 1, _("CONFLICT (rename/rename): "
1650                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1651                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1652                one->path, a->path, ci->branch1,
1653                one->path, b->path, ci->branch2,
1654                o->call_depth ? _(" (left unresolved)") : "");
1655         if (o->call_depth) {
1656                 struct merge_file_info mfi;
1657                 struct diff_filespec other;
1658                 struct diff_filespec *add;
1659                 if (merge_file_one(o, one->path,
1660                                  &one->oid, one->mode,
1661                                  &a->oid, a->mode,
1662                                  &b->oid, b->mode,
1663                                  ci->branch1, ci->branch2, &mfi))
1664                         return -1;
1665
1666                 /*
1667                  * FIXME: For rename/add-source conflicts (if we could detect
1668                  * such), this is wrong.  We should instead find a unique
1669                  * pathname and then either rename the add-source file to that
1670                  * unique path, or use that unique path instead of src here.
1671                  */
1672                 if (update_file(o, 0, &mfi.oid, mfi.mode, one->path))
1673                         return -1;
1674
1675                 /*
1676                  * Above, we put the merged content at the merge-base's
1677                  * path.  Now we usually need to delete both a->path and
1678                  * b->path.  However, the rename on each side of the merge
1679                  * could also be involved in a rename/add conflict.  In
1680                  * such cases, we should keep the added file around,
1681                  * resolving the conflict at that path in its favor.
1682                  */
1683                 add = filespec_from_entry(&other, ci->dst_entry1, 2 ^ 1);
1684                 if (add) {
1685                         if (update_file(o, 0, &add->oid, add->mode, a->path))
1686                                 return -1;
1687                 }
1688                 else
1689                         remove_file_from_cache(a->path);
1690                 add = filespec_from_entry(&other, ci->dst_entry2, 3 ^ 1);
1691                 if (add) {
1692                         if (update_file(o, 0, &add->oid, add->mode, b->path))
1693                                 return -1;
1694                 }
1695                 else
1696                         remove_file_from_cache(b->path);
1697         } else if (handle_file(o, a, 2, ci) || handle_file(o, b, 3, ci))
1698                 return -1;
1699
1700         return 0;
1701 }
1702
1703 static int handle_rename_rename_2to1(struct merge_options *o,
1704                                      struct rename_conflict_info *ci)
1705 {
1706         /* Two files, a & b, were renamed to the same thing, c. */
1707         struct diff_filespec *a = ci->pair1->one;
1708         struct diff_filespec *b = ci->pair2->one;
1709         struct diff_filespec *c1 = ci->pair1->two;
1710         struct diff_filespec *c2 = ci->pair2->two;
1711         char *path = c1->path; /* == c2->path */
1712         char *path_side_1_desc;
1713         char *path_side_2_desc;
1714         struct merge_file_info mfi_c1;
1715         struct merge_file_info mfi_c2;
1716         int ret;
1717
1718         output(o, 1, _("CONFLICT (rename/rename): "
1719                "Rename %s->%s in %s. "
1720                "Rename %s->%s in %s"),
1721                a->path, c1->path, ci->branch1,
1722                b->path, c2->path, ci->branch2);
1723
1724         remove_file(o, 1, a->path, o->call_depth || would_lose_untracked(a->path));
1725         remove_file(o, 1, b->path, o->call_depth || would_lose_untracked(b->path));
1726
1727         path_side_1_desc = xstrfmt("%s (was %s)", path, a->path);
1728         path_side_2_desc = xstrfmt("%s (was %s)", path, b->path);
1729         if (merge_file_special_markers(o, a, c1, &ci->ren1_other,
1730                                        path_side_1_desc,
1731                                        o->branch1, c1->path,
1732                                        o->branch2, ci->ren1_other.path, &mfi_c1) ||
1733             merge_file_special_markers(o, b, &ci->ren2_other, c2,
1734                                        path_side_2_desc,
1735                                        o->branch1, ci->ren2_other.path,
1736                                        o->branch2, c2->path, &mfi_c2))
1737                 return -1;
1738         free(path_side_1_desc);
1739         free(path_side_2_desc);
1740
1741         if (o->call_depth) {
1742                 /*
1743                  * If mfi_c1.clean && mfi_c2.clean, then it might make
1744                  * sense to do a two-way merge of those results.  But, I
1745                  * think in all cases, it makes sense to have the virtual
1746                  * merge base just undo the renames; they can be detected
1747                  * again later for the non-recursive merge.
1748                  */
1749                 remove_file(o, 0, path, 0);
1750                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, a->path);
1751                 if (!ret)
1752                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1753                                           b->path);
1754         } else {
1755                 char *new_path1 = unique_path(o, path, ci->branch1);
1756                 char *new_path2 = unique_path(o, path, ci->branch2);
1757                 output(o, 1, _("Renaming %s to %s and %s to %s instead"),
1758                        a->path, new_path1, b->path, new_path2);
1759                 if (was_dirty(o, path))
1760                         output(o, 1, _("Refusing to lose dirty file at %s"),
1761                                path);
1762                 else if (would_lose_untracked(path))
1763                         /*
1764                          * Only way we get here is if both renames were from
1765                          * a directory rename AND user had an untracked file
1766                          * at the location where both files end up after the
1767                          * two directory renames.  See testcase 10d of t6043.
1768                          */
1769                         output(o, 1, _("Refusing to lose untracked file at "
1770                                        "%s, even though it's in the way."),
1771                                path);
1772                 else
1773                         remove_file(o, 0, path, 0);
1774                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, new_path1);
1775                 if (!ret)
1776                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1777                                           new_path2);
1778                 /*
1779                  * unpack_trees() actually populates the index for us for
1780                  * "normal" rename/rename(2to1) situtations so that the
1781                  * correct entries are at the higher stages, which would
1782                  * make the call below to update_stages_for_stage_data
1783                  * unnecessary.  However, if either of the renames came
1784                  * from a directory rename, then unpack_trees() will not
1785                  * have gotten the right data loaded into the index, so we
1786                  * need to do so now.  (While it'd be tempting to move this
1787                  * call to update_stages_for_stage_data() to
1788                  * apply_directory_rename_modifications(), that would break
1789                  * our intermediate calls to would_lose_untracked() since
1790                  * those rely on the current in-memory index.  See also the
1791                  * big "NOTE" in update_stages()).
1792                  */
1793                 if (update_stages_for_stage_data(o, path, ci->dst_entry1))
1794                         ret = -1;
1795
1796                 free(new_path2);
1797                 free(new_path1);
1798         }
1799
1800         return ret;
1801 }
1802
1803 /*
1804  * Get the diff_filepairs changed between o_tree and tree.
1805  */
1806 static struct diff_queue_struct *get_diffpairs(struct merge_options *o,
1807                                                struct tree *o_tree,
1808                                                struct tree *tree)
1809 {
1810         struct diff_queue_struct *ret;
1811         struct diff_options opts;
1812
1813         diff_setup(&opts);
1814         opts.flags.recursive = 1;
1815         opts.flags.rename_empty = 0;
1816         opts.detect_rename = merge_detect_rename(o);
1817         /*
1818          * We do not have logic to handle the detection of copies.  In
1819          * fact, it may not even make sense to add such logic: would we
1820          * really want a change to a base file to be propagated through
1821          * multiple other files by a merge?
1822          */
1823         if (opts.detect_rename > DIFF_DETECT_RENAME)
1824                 opts.detect_rename = DIFF_DETECT_RENAME;
1825         opts.rename_limit = o->merge_rename_limit >= 0 ? o->merge_rename_limit :
1826                             o->diff_rename_limit >= 0 ? o->diff_rename_limit :
1827                             1000;
1828         opts.rename_score = o->rename_score;
1829         opts.show_rename_progress = o->show_rename_progress;
1830         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1831         diff_setup_done(&opts);
1832         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1833         diffcore_std(&opts);
1834         if (opts.needed_rename_limit > o->needed_rename_limit)
1835                 o->needed_rename_limit = opts.needed_rename_limit;
1836
1837         ret = xmalloc(sizeof(*ret));
1838         *ret = diff_queued_diff;
1839
1840         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1841         diff_queued_diff.nr = 0;
1842         diff_queued_diff.queue = NULL;
1843         diff_flush(&opts);
1844         return ret;
1845 }
1846
1847 static int tree_has_path(struct tree *tree, const char *path)
1848 {
1849         struct object_id hashy;
1850         unsigned int mode_o;
1851
1852         return !get_tree_entry(&tree->object.oid, path,
1853                                &hashy, &mode_o);
1854 }
1855
1856 /*
1857  * Return a new string that replaces the beginning portion (which matches
1858  * entry->dir), with entry->new_dir.  In perl-speak:
1859  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1860  * NOTE:
1861  *   Caller must ensure that old_path starts with entry->dir + '/'.
1862  */
1863 static char *apply_dir_rename(struct dir_rename_entry *entry,
1864                               const char *old_path)
1865 {
1866         struct strbuf new_path = STRBUF_INIT;
1867         int oldlen, newlen;
1868
1869         if (entry->non_unique_new_dir)
1870                 return NULL;
1871
1872         oldlen = strlen(entry->dir);
1873         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1874         strbuf_grow(&new_path, newlen);
1875         strbuf_addbuf(&new_path, &entry->new_dir);
1876         strbuf_addstr(&new_path, &old_path[oldlen]);
1877
1878         return strbuf_detach(&new_path, NULL);
1879 }
1880
1881 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1882                                     char **old_dir, char **new_dir)
1883 {
1884         char *end_of_old, *end_of_new;
1885         int old_len, new_len;
1886
1887         *old_dir = NULL;
1888         *new_dir = NULL;
1889
1890         /*
1891          * For
1892          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1893          * the "e/foo.c" part is the same, we just want to know that
1894          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1895          * so, for this example, this function returns "a/b/c/d" in
1896          * *old_dir and "a/b/some/thing/else" in *new_dir.
1897          *
1898          * Also, if the basename of the file changed, we don't care.  We
1899          * want to know which portion of the directory, if any, changed.
1900          */
1901         end_of_old = strrchr(old_path, '/');
1902         end_of_new = strrchr(new_path, '/');
1903
1904         if (end_of_old == NULL || end_of_new == NULL)
1905                 return;
1906         while (*--end_of_new == *--end_of_old &&
1907                end_of_old != old_path &&
1908                end_of_new != new_path)
1909                 ; /* Do nothing; all in the while loop */
1910         /*
1911          * We've found the first non-matching character in the directory
1912          * paths.  That means the current directory we were comparing
1913          * represents the rename.  Move end_of_old and end_of_new back
1914          * to the full directory name.
1915          */
1916         if (*end_of_old == '/')
1917                 end_of_old++;
1918         if (*end_of_old != '/')
1919                 end_of_new++;
1920         end_of_old = strchr(end_of_old, '/');
1921         end_of_new = strchr(end_of_new, '/');
1922
1923         /*
1924          * It may have been the case that old_path and new_path were the same
1925          * directory all along.  Don't claim a rename if they're the same.
1926          */
1927         old_len = end_of_old - old_path;
1928         new_len = end_of_new - new_path;
1929
1930         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1931                 *old_dir = xstrndup(old_path, old_len);
1932                 *new_dir = xstrndup(new_path, new_len);
1933         }
1934 }
1935
1936 static void remove_hashmap_entries(struct hashmap *dir_renames,
1937                                    struct string_list *items_to_remove)
1938 {
1939         int i;
1940         struct dir_rename_entry *entry;
1941
1942         for (i = 0; i < items_to_remove->nr; i++) {
1943                 entry = items_to_remove->items[i].util;
1944                 hashmap_remove(dir_renames, entry, NULL);
1945         }
1946         string_list_clear(items_to_remove, 0);
1947 }
1948
1949 /*
1950  * See if there is a directory rename for path, and if there are any file
1951  * level conflicts for the renamed location.  If there is a rename and
1952  * there are no conflicts, return the new name.  Otherwise, return NULL.
1953  */
1954 static char *handle_path_level_conflicts(struct merge_options *o,
1955                                          const char *path,
1956                                          struct dir_rename_entry *entry,
1957                                          struct hashmap *collisions,
1958                                          struct tree *tree)
1959 {
1960         char *new_path = NULL;
1961         struct collision_entry *collision_ent;
1962         int clean = 1;
1963         struct strbuf collision_paths = STRBUF_INIT;
1964
1965         /*
1966          * entry has the mapping of old directory name to new directory name
1967          * that we want to apply to path.
1968          */
1969         new_path = apply_dir_rename(entry, path);
1970
1971         if (!new_path) {
1972                 /* This should only happen when entry->non_unique_new_dir set */
1973                 if (!entry->non_unique_new_dir)
1974                         BUG("entry->non_unqiue_dir not set and !new_path");
1975                 output(o, 1, _("CONFLICT (directory rename split): "
1976                                "Unclear where to place %s because directory "
1977                                "%s was renamed to multiple other directories, "
1978                                "with no destination getting a majority of the "
1979                                "files."),
1980                        path, entry->dir);
1981                 clean = 0;
1982                 return NULL;
1983         }
1984
1985         /*
1986          * The caller needs to have ensured that it has pre-populated
1987          * collisions with all paths that map to new_path.  Do a quick check
1988          * to ensure that's the case.
1989          */
1990         collision_ent = collision_find_entry(collisions, new_path);
1991         if (collision_ent == NULL)
1992                 BUG("collision_ent is NULL");
1993
1994         /*
1995          * Check for one-sided add/add/.../add conflicts, i.e.
1996          * where implicit renames from the other side doing
1997          * directory rename(s) can affect this side of history
1998          * to put multiple paths into the same location.  Warn
1999          * and bail on directory renames for such paths.
2000          */
2001         if (collision_ent->reported_already) {
2002                 clean = 0;
2003         } else if (tree_has_path(tree, new_path)) {
2004                 collision_ent->reported_already = 1;
2005                 strbuf_add_separated_string_list(&collision_paths, ", ",
2006                                                  &collision_ent->source_files);
2007                 output(o, 1, _("CONFLICT (implicit dir rename): Existing "
2008                                "file/dir at %s in the way of implicit "
2009                                "directory rename(s) putting the following "
2010                                "path(s) there: %s."),
2011                        new_path, collision_paths.buf);
2012                 clean = 0;
2013         } else if (collision_ent->source_files.nr > 1) {
2014                 collision_ent->reported_already = 1;
2015                 strbuf_add_separated_string_list(&collision_paths, ", ",
2016                                                  &collision_ent->source_files);
2017                 output(o, 1, _("CONFLICT (implicit dir rename): Cannot map "
2018                                "more than one path to %s; implicit directory "
2019                                "renames tried to put these paths there: %s"),
2020                        new_path, collision_paths.buf);
2021                 clean = 0;
2022         }
2023
2024         /* Free memory we no longer need */
2025         strbuf_release(&collision_paths);
2026         if (!clean && new_path) {
2027                 free(new_path);
2028                 return NULL;
2029         }
2030
2031         return new_path;
2032 }
2033
2034 /*
2035  * There are a couple things we want to do at the directory level:
2036  *   1. Check for both sides renaming to the same thing, in order to avoid
2037  *      implicit renaming of files that should be left in place.  (See
2038  *      testcase 6b in t6043 for details.)
2039  *   2. Prune directory renames if there are still files left in the
2040  *      the original directory.  These represent a partial directory rename,
2041  *      i.e. a rename where only some of the files within the directory
2042  *      were renamed elsewhere.  (Technically, this could be done earlier
2043  *      in get_directory_renames(), except that would prevent us from
2044  *      doing the previous check and thus failing testcase 6b.)
2045  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2046  *      In the future, we could potentially record this info as well and
2047  *      omit reporting rename/rename(1to2) conflicts for each path within
2048  *      the affected directories, thus cleaning up the merge output.
2049  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2050  *         directory level, because merging directories is fine.  If it
2051  *         causes conflicts for files within those merged directories, then
2052  *         that should be detected at the individual path level.
2053  */
2054 static void handle_directory_level_conflicts(struct merge_options *o,
2055                                              struct hashmap *dir_re_head,
2056                                              struct tree *head,
2057                                              struct hashmap *dir_re_merge,
2058                                              struct tree *merge)
2059 {
2060         struct hashmap_iter iter;
2061         struct dir_rename_entry *head_ent;
2062         struct dir_rename_entry *merge_ent;
2063
2064         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2065         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2066
2067         hashmap_iter_init(dir_re_head, &iter);
2068         while ((head_ent = hashmap_iter_next(&iter))) {
2069                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2070                 if (merge_ent &&
2071                     !head_ent->non_unique_new_dir &&
2072                     !merge_ent->non_unique_new_dir &&
2073                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2074                         /* 1. Renamed identically; remove it from both sides */
2075                         string_list_append(&remove_from_head,
2076                                            head_ent->dir)->util = head_ent;
2077                         strbuf_release(&head_ent->new_dir);
2078                         string_list_append(&remove_from_merge,
2079                                            merge_ent->dir)->util = merge_ent;
2080                         strbuf_release(&merge_ent->new_dir);
2081                 } else if (tree_has_path(head, head_ent->dir)) {
2082                         /* 2. This wasn't a directory rename after all */
2083                         string_list_append(&remove_from_head,
2084                                            head_ent->dir)->util = head_ent;
2085                         strbuf_release(&head_ent->new_dir);
2086                 }
2087         }
2088
2089         remove_hashmap_entries(dir_re_head, &remove_from_head);
2090         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2091
2092         hashmap_iter_init(dir_re_merge, &iter);
2093         while ((merge_ent = hashmap_iter_next(&iter))) {
2094                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2095                 if (tree_has_path(merge, merge_ent->dir)) {
2096                         /* 2. This wasn't a directory rename after all */
2097                         string_list_append(&remove_from_merge,
2098                                            merge_ent->dir)->util = merge_ent;
2099                 } else if (head_ent &&
2100                            !head_ent->non_unique_new_dir &&
2101                            !merge_ent->non_unique_new_dir) {
2102                         /* 3. rename/rename(1to2) */
2103                         /*
2104                          * We can assume it's not rename/rename(1to1) because
2105                          * that was case (1), already checked above.  So we
2106                          * know that head_ent->new_dir and merge_ent->new_dir
2107                          * are different strings.
2108                          */
2109                         output(o, 1, _("CONFLICT (rename/rename): "
2110                                        "Rename directory %s->%s in %s. "
2111                                        "Rename directory %s->%s in %s"),
2112                                head_ent->dir, head_ent->new_dir.buf, o->branch1,
2113                                head_ent->dir, merge_ent->new_dir.buf, o->branch2);
2114                         string_list_append(&remove_from_head,
2115                                            head_ent->dir)->util = head_ent;
2116                         strbuf_release(&head_ent->new_dir);
2117                         string_list_append(&remove_from_merge,
2118                                            merge_ent->dir)->util = merge_ent;
2119                         strbuf_release(&merge_ent->new_dir);
2120                 }
2121         }
2122
2123         remove_hashmap_entries(dir_re_head, &remove_from_head);
2124         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2125 }
2126
2127 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs,
2128                                              struct tree *tree)
2129 {
2130         struct hashmap *dir_renames;
2131         struct hashmap_iter iter;
2132         struct dir_rename_entry *entry;
2133         int i;
2134
2135         /*
2136          * Typically, we think of a directory rename as all files from a
2137          * certain directory being moved to a target directory.  However,
2138          * what if someone first moved two files from the original
2139          * directory in one commit, and then renamed the directory
2140          * somewhere else in a later commit?  At merge time, we just know
2141          * that files from the original directory went to two different
2142          * places, and that the bulk of them ended up in the same place.
2143          * We want each directory rename to represent where the bulk of the
2144          * files from that directory end up; this function exists to find
2145          * where the bulk of the files went.
2146          *
2147          * The first loop below simply iterates through the list of file
2148          * renames, finding out how often each directory rename pair
2149          * possibility occurs.
2150          */
2151         dir_renames = xmalloc(sizeof(*dir_renames));
2152         dir_rename_init(dir_renames);
2153         for (i = 0; i < pairs->nr; ++i) {
2154                 struct string_list_item *item;
2155                 int *count;
2156                 struct diff_filepair *pair = pairs->queue[i];
2157                 char *old_dir, *new_dir;
2158
2159                 /* File not part of directory rename if it wasn't renamed */
2160                 if (pair->status != 'R')
2161                         continue;
2162
2163                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2164                                         &old_dir,        &new_dir);
2165                 if (!old_dir)
2166                         /* Directory didn't change at all; ignore this one. */
2167                         continue;
2168
2169                 entry = dir_rename_find_entry(dir_renames, old_dir);
2170                 if (!entry) {
2171                         entry = xmalloc(sizeof(*entry));
2172                         dir_rename_entry_init(entry, old_dir);
2173                         hashmap_put(dir_renames, entry);
2174                 } else {
2175                         free(old_dir);
2176                 }
2177                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2178                 if (!item) {
2179                         item = string_list_insert(&entry->possible_new_dirs,
2180                                                   new_dir);
2181                         item->util = xcalloc(1, sizeof(int));
2182                 } else {
2183                         free(new_dir);
2184                 }
2185                 count = item->util;
2186                 *count += 1;
2187         }
2188
2189         /*
2190          * For each directory with files moved out of it, we find out which
2191          * target directory received the most files so we can declare it to
2192          * be the "winning" target location for the directory rename.  This
2193          * winner gets recorded in new_dir.  If there is no winner
2194          * (multiple target directories received the same number of files),
2195          * we set non_unique_new_dir.  Once we've determined the winner (or
2196          * that there is no winner), we no longer need possible_new_dirs.
2197          */
2198         hashmap_iter_init(dir_renames, &iter);
2199         while ((entry = hashmap_iter_next(&iter))) {
2200                 int max = 0;
2201                 int bad_max = 0;
2202                 char *best = NULL;
2203
2204                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2205                         int *count = entry->possible_new_dirs.items[i].util;
2206
2207                         if (*count == max)
2208                                 bad_max = max;
2209                         else if (*count > max) {
2210                                 max = *count;
2211                                 best = entry->possible_new_dirs.items[i].string;
2212                         }
2213                 }
2214                 if (bad_max == max)
2215                         entry->non_unique_new_dir = 1;
2216                 else {
2217                         assert(entry->new_dir.len == 0);
2218                         strbuf_addstr(&entry->new_dir, best);
2219                 }
2220                 /*
2221                  * The relevant directory sub-portion of the original full
2222                  * filepaths were xstrndup'ed before inserting into
2223                  * possible_new_dirs, and instead of manually iterating the
2224                  * list and free'ing each, just lie and tell
2225                  * possible_new_dirs that it did the strdup'ing so that it
2226                  * will free them for us.
2227                  */
2228                 entry->possible_new_dirs.strdup_strings = 1;
2229                 string_list_clear(&entry->possible_new_dirs, 1);
2230         }
2231
2232         return dir_renames;
2233 }
2234
2235 static struct dir_rename_entry *check_dir_renamed(const char *path,
2236                                                   struct hashmap *dir_renames)
2237 {
2238         char *temp = xstrdup(path);
2239         char *end;
2240         struct dir_rename_entry *entry = NULL;;
2241
2242         while ((end = strrchr(temp, '/'))) {
2243                 *end = '\0';
2244                 entry = dir_rename_find_entry(dir_renames, temp);
2245                 if (entry)
2246                         break;
2247         }
2248         free(temp);
2249         return entry;
2250 }
2251
2252 static void compute_collisions(struct hashmap *collisions,
2253                                struct hashmap *dir_renames,
2254                                struct diff_queue_struct *pairs)
2255 {
2256         int i;
2257
2258         /*
2259          * Multiple files can be mapped to the same path due to directory
2260          * renames done by the other side of history.  Since that other
2261          * side of history could have merged multiple directories into one,
2262          * if our side of history added the same file basename to each of
2263          * those directories, then all N of them would get implicitly
2264          * renamed by the directory rename detection into the same path,
2265          * and we'd get an add/add/.../add conflict, and all those adds
2266          * from *this* side of history.  This is not representable in the
2267          * index, and users aren't going to easily be able to make sense of
2268          * it.  So we need to provide a good warning about what's
2269          * happening, and fall back to no-directory-rename detection
2270          * behavior for those paths.
2271          *
2272          * See testcases 9e and all of section 5 from t6043 for examples.
2273          */
2274         collision_init(collisions);
2275
2276         for (i = 0; i < pairs->nr; ++i) {
2277                 struct dir_rename_entry *dir_rename_ent;
2278                 struct collision_entry *collision_ent;
2279                 char *new_path;
2280                 struct diff_filepair *pair = pairs->queue[i];
2281
2282                 if (pair->status != 'A' && pair->status != 'R')
2283                         continue;
2284                 dir_rename_ent = check_dir_renamed(pair->two->path,
2285                                                    dir_renames);
2286                 if (!dir_rename_ent)
2287                         continue;
2288
2289                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2290                 if (!new_path)
2291                         /*
2292                          * dir_rename_ent->non_unique_new_path is true, which
2293                          * means there is no directory rename for us to use,
2294                          * which means it won't cause us any additional
2295                          * collisions.
2296                          */
2297                         continue;
2298                 collision_ent = collision_find_entry(collisions, new_path);
2299                 if (!collision_ent) {
2300                         collision_ent = xcalloc(1,
2301                                                 sizeof(struct collision_entry));
2302                         hashmap_entry_init(collision_ent, strhash(new_path));
2303                         hashmap_put(collisions, collision_ent);
2304                         collision_ent->target_file = new_path;
2305                 } else {
2306                         free(new_path);
2307                 }
2308                 string_list_insert(&collision_ent->source_files,
2309                                    pair->two->path);
2310         }
2311 }
2312
2313 static char *check_for_directory_rename(struct merge_options *o,
2314                                         const char *path,
2315                                         struct tree *tree,
2316                                         struct hashmap *dir_renames,
2317                                         struct hashmap *dir_rename_exclusions,
2318                                         struct hashmap *collisions,
2319                                         int *clean_merge)
2320 {
2321         char *new_path = NULL;
2322         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2323         struct dir_rename_entry *oentry = NULL;
2324
2325         if (!entry)
2326                 return new_path;
2327
2328         /*
2329          * This next part is a little weird.  We do not want to do an
2330          * implicit rename into a directory we renamed on our side, because
2331          * that will result in a spurious rename/rename(1to2) conflict.  An
2332          * example:
2333          *   Base commit: dumbdir/afile, otherdir/bfile
2334          *   Side 1:      smrtdir/afile, otherdir/bfile
2335          *   Side 2:      dumbdir/afile, dumbdir/bfile
2336          * Here, while working on Side 1, we could notice that otherdir was
2337          * renamed/merged to dumbdir, and change the diff_filepair for
2338          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2339          * 2 will notice the rename from dumbdir to smrtdir, and do the
2340          * transitive rename to move it from dumbdir/bfile to
2341          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2342          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2343          * the file to end up in smrtdir.  And the way to achieve that is
2344          * to not let Side1 do the rename to dumbdir, since we know that is
2345          * the source of one of our directory renames.
2346          *
2347          * That's why oentry and dir_rename_exclusions is here.
2348          *
2349          * As it turns out, this also prevents N-way transient rename
2350          * confusion; See testcases 9c and 9d of t6043.
2351          */
2352         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2353         if (oentry) {
2354                 output(o, 1, _("WARNING: Avoiding applying %s -> %s rename "
2355                                "to %s, because %s itself was renamed."),
2356                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2357         } else {
2358                 new_path = handle_path_level_conflicts(o, path, entry,
2359                                                        collisions, tree);
2360                 *clean_merge &= (new_path != NULL);
2361         }
2362
2363         return new_path;
2364 }
2365
2366 static void apply_directory_rename_modifications(struct merge_options *o,
2367                                                  struct diff_filepair *pair,
2368                                                  char *new_path,
2369                                                  struct rename *re,
2370                                                  struct tree *tree,
2371                                                  struct tree *o_tree,
2372                                                  struct tree *a_tree,
2373                                                  struct tree *b_tree,
2374                                                  struct string_list *entries,
2375                                                  int *clean)
2376 {
2377         struct string_list_item *item;
2378         int stage = (tree == a_tree ? 2 : 3);
2379         int update_wd;
2380
2381         /*
2382          * In all cases where we can do directory rename detection,
2383          * unpack_trees() will have read pair->two->path into the
2384          * index and the working copy.  We need to remove it so that
2385          * we can instead place it at new_path.  It is guaranteed to
2386          * not be untracked (unpack_trees() would have errored out
2387          * saying the file would have been overwritten), but it might
2388          * be dirty, though.
2389          */
2390         update_wd = !was_dirty(o, pair->two->path);
2391         if (!update_wd)
2392                 output(o, 1, _("Refusing to lose dirty file at %s"),
2393                        pair->two->path);
2394         remove_file(o, 1, pair->two->path, !update_wd);
2395
2396         /* Find or create a new re->dst_entry */
2397         item = string_list_lookup(entries, new_path);
2398         if (item) {
2399                 /*
2400                  * Since we're renaming on this side of history, and it's
2401                  * due to a directory rename on the other side of history
2402                  * (which we only allow when the directory in question no
2403                  * longer exists on the other side of history), the
2404                  * original entry for re->dst_entry is no longer
2405                  * necessary...
2406                  */
2407                 re->dst_entry->processed = 1;
2408
2409                 /*
2410                  * ...because we'll be using this new one.
2411                  */
2412                 re->dst_entry = item->util;
2413         } else {
2414                 /*
2415                  * re->dst_entry is for the before-dir-rename path, and we
2416                  * need it to hold information for the after-dir-rename
2417                  * path.  Before creating a new entry, we need to mark the
2418                  * old one as unnecessary (...unless it is shared by
2419                  * src_entry, i.e. this didn't use to be a rename, in which
2420                  * case we can just allow the normal processing to happen
2421                  * for it).
2422                  */
2423                 if (pair->status == 'R')
2424                         re->dst_entry->processed = 1;
2425
2426                 re->dst_entry = insert_stage_data(new_path,
2427                                                   o_tree, a_tree, b_tree,
2428                                                   entries);
2429                 item = string_list_insert(entries, new_path);
2430                 item->util = re->dst_entry;
2431         }
2432
2433         /*
2434          * Update the stage_data with the information about the path we are
2435          * moving into place.  That slot will be empty and available for us
2436          * to write to because of the collision checks in
2437          * handle_path_level_conflicts().  In other words,
2438          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2439          * open for us to write to.
2440          *
2441          * It may be tempting to actually update the index at this point as
2442          * well, using update_stages_for_stage_data(), but as per the big
2443          * "NOTE" in update_stages(), doing so will modify the current
2444          * in-memory index which will break calls to would_lose_untracked()
2445          * that we need to make.  Instead, we need to just make sure that
2446          * the various handle_rename_*() functions update the index
2447          * explicitly rather than relying on unpack_trees() to have done it.
2448          */
2449         get_tree_entry(&tree->object.oid,
2450                        pair->two->path,
2451                        &re->dst_entry->stages[stage].oid,
2452                        &re->dst_entry->stages[stage].mode);
2453
2454         /* Update pair status */
2455         if (pair->status == 'A') {
2456                 /*
2457                  * Recording rename information for this add makes it look
2458                  * like a rename/delete conflict.  Make sure we can
2459                  * correctly handle this as an add that was moved to a new
2460                  * directory instead of reporting a rename/delete conflict.
2461                  */
2462                 re->add_turned_into_rename = 1;
2463         }
2464         /*
2465          * We don't actually look at pair->status again, but it seems
2466          * pedagogically correct to adjust it.
2467          */
2468         pair->status = 'R';
2469
2470         /*
2471          * Finally, record the new location.
2472          */
2473         pair->two->path = new_path;
2474 }
2475
2476 /*
2477  * Get information of all renames which occurred in 'pairs', making use of
2478  * any implicit directory renames inferred from the other side of history.
2479  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2480  * to be able to associate the correct cache entries with the rename
2481  * information; tree is always equal to either a_tree or b_tree.
2482  */
2483 static struct string_list *get_renames(struct merge_options *o,
2484                                        struct diff_queue_struct *pairs,
2485                                        struct hashmap *dir_renames,
2486                                        struct hashmap *dir_rename_exclusions,
2487                                        struct tree *tree,
2488                                        struct tree *o_tree,
2489                                        struct tree *a_tree,
2490                                        struct tree *b_tree,
2491                                        struct string_list *entries,
2492                                        int *clean_merge)
2493 {
2494         int i;
2495         struct hashmap collisions;
2496         struct hashmap_iter iter;
2497         struct collision_entry *e;
2498         struct string_list *renames;
2499
2500         compute_collisions(&collisions, dir_renames, pairs);
2501         renames = xcalloc(1, sizeof(struct string_list));
2502
2503         for (i = 0; i < pairs->nr; ++i) {
2504                 struct string_list_item *item;
2505                 struct rename *re;
2506                 struct diff_filepair *pair = pairs->queue[i];
2507                 char *new_path; /* non-NULL only with directory renames */
2508
2509                 if (pair->status != 'A' && pair->status != 'R') {
2510                         diff_free_filepair(pair);
2511                         continue;
2512                 }
2513                 new_path = check_for_directory_rename(o, pair->two->path, tree,
2514                                                       dir_renames,
2515                                                       dir_rename_exclusions,
2516                                                       &collisions,
2517                                                       clean_merge);
2518                 if (pair->status != 'R' && !new_path) {
2519                         diff_free_filepair(pair);
2520                         continue;
2521                 }
2522
2523                 re = xmalloc(sizeof(*re));
2524                 re->processed = 0;
2525                 re->add_turned_into_rename = 0;
2526                 re->pair = pair;
2527                 item = string_list_lookup(entries, re->pair->one->path);
2528                 if (!item)
2529                         re->src_entry = insert_stage_data(re->pair->one->path,
2530                                         o_tree, a_tree, b_tree, entries);
2531                 else
2532                         re->src_entry = item->util;
2533
2534                 item = string_list_lookup(entries, re->pair->two->path);
2535                 if (!item)
2536                         re->dst_entry = insert_stage_data(re->pair->two->path,
2537                                         o_tree, a_tree, b_tree, entries);
2538                 else
2539                         re->dst_entry = item->util;
2540                 item = string_list_insert(renames, pair->one->path);
2541                 item->util = re;
2542                 if (new_path)
2543                         apply_directory_rename_modifications(o, pair, new_path,
2544                                                              re, tree, o_tree,
2545                                                              a_tree, b_tree,
2546                                                              entries,
2547                                                              clean_merge);
2548         }
2549
2550         hashmap_iter_init(&collisions, &iter);
2551         while ((e = hashmap_iter_next(&iter))) {
2552                 free(e->target_file);
2553                 string_list_clear(&e->source_files, 0);
2554         }
2555         hashmap_free(&collisions, 1);
2556         return renames;
2557 }
2558
2559 static int process_renames(struct merge_options *o,
2560                            struct string_list *a_renames,
2561                            struct string_list *b_renames)
2562 {
2563         int clean_merge = 1, i, j;
2564         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2565         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2566         const struct rename *sre;
2567
2568         for (i = 0; i < a_renames->nr; i++) {
2569                 sre = a_renames->items[i].util;
2570                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2571                         = (void *)sre;
2572         }
2573         for (i = 0; i < b_renames->nr; i++) {
2574                 sre = b_renames->items[i].util;
2575                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2576                         = (void *)sre;
2577         }
2578
2579         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2580                 struct string_list *renames1, *renames2Dst;
2581                 struct rename *ren1 = NULL, *ren2 = NULL;
2582                 const char *branch1, *branch2;
2583                 const char *ren1_src, *ren1_dst;
2584                 struct string_list_item *lookup;
2585
2586                 if (i >= a_renames->nr) {
2587                         ren2 = b_renames->items[j++].util;
2588                 } else if (j >= b_renames->nr) {
2589                         ren1 = a_renames->items[i++].util;
2590                 } else {
2591                         int compare = strcmp(a_renames->items[i].string,
2592                                              b_renames->items[j].string);
2593                         if (compare <= 0)
2594                                 ren1 = a_renames->items[i++].util;
2595                         if (compare >= 0)
2596                                 ren2 = b_renames->items[j++].util;
2597                 }
2598
2599                 /* TODO: refactor, so that 1/2 are not needed */
2600                 if (ren1) {
2601                         renames1 = a_renames;
2602                         renames2Dst = &b_by_dst;
2603                         branch1 = o->branch1;
2604                         branch2 = o->branch2;
2605                 } else {
2606                         renames1 = b_renames;
2607                         renames2Dst = &a_by_dst;
2608                         branch1 = o->branch2;
2609                         branch2 = o->branch1;
2610                         SWAP(ren2, ren1);
2611                 }
2612
2613                 if (ren1->processed)
2614                         continue;
2615                 ren1->processed = 1;
2616                 ren1->dst_entry->processed = 1;
2617                 /* BUG: We should only mark src_entry as processed if we
2618                  * are not dealing with a rename + add-source case.
2619                  */
2620                 ren1->src_entry->processed = 1;
2621
2622                 ren1_src = ren1->pair->one->path;
2623                 ren1_dst = ren1->pair->two->path;
2624
2625                 if (ren2) {
2626                         /* One file renamed on both sides */
2627                         const char *ren2_src = ren2->pair->one->path;
2628                         const char *ren2_dst = ren2->pair->two->path;
2629                         enum rename_type rename_type;
2630                         if (strcmp(ren1_src, ren2_src) != 0)
2631                                 BUG("ren1_src != ren2_src");
2632                         ren2->dst_entry->processed = 1;
2633                         ren2->processed = 1;
2634                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2635                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2636                                 clean_merge = 0;
2637                         } else {
2638                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2639                                 /* BUG: We should only remove ren1_src in
2640                                  * the base stage (think of rename +
2641                                  * add-source cases).
2642                                  */
2643                                 remove_file(o, 1, ren1_src, 1);
2644                                 update_entry(ren1->dst_entry,
2645                                              ren1->pair->one,
2646                                              ren1->pair->two,
2647                                              ren2->pair->two);
2648                         }
2649                         setup_rename_conflict_info(rename_type,
2650                                                    ren1->pair,
2651                                                    ren2->pair,
2652                                                    branch1,
2653                                                    branch2,
2654                                                    ren1->dst_entry,
2655                                                    ren2->dst_entry,
2656                                                    o,
2657                                                    NULL,
2658                                                    NULL);
2659                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2660                         /* Two different files renamed to the same thing */
2661                         char *ren2_dst;
2662                         ren2 = lookup->util;
2663                         ren2_dst = ren2->pair->two->path;
2664                         if (strcmp(ren1_dst, ren2_dst) != 0)
2665                                 BUG("ren1_dst != ren2_dst");
2666
2667                         clean_merge = 0;
2668                         ren2->processed = 1;
2669                         /*
2670                          * BUG: We should only mark src_entry as processed
2671                          * if we are not dealing with a rename + add-source
2672                          * case.
2673                          */
2674                         ren2->src_entry->processed = 1;
2675
2676                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2677                                                    ren1->pair,
2678                                                    ren2->pair,
2679                                                    branch1,
2680                                                    branch2,
2681                                                    ren1->dst_entry,
2682                                                    ren2->dst_entry,
2683                                                    o,
2684                                                    ren1->src_entry,
2685                                                    ren2->src_entry);
2686
2687                 } else {
2688                         /* Renamed in 1, maybe changed in 2 */
2689                         /* we only use sha1 and mode of these */
2690                         struct diff_filespec src_other, dst_other;
2691                         int try_merge;
2692
2693                         /*
2694                          * unpack_trees loads entries from common-commit
2695                          * into stage 1, from head-commit into stage 2, and
2696                          * from merge-commit into stage 3.  We keep track
2697                          * of which side corresponds to the rename.
2698                          */
2699                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2700                         int other_stage =   a_renames == renames1 ? 3 : 2;
2701
2702                         /* BUG: We should only remove ren1_src in the base
2703                          * stage and in other_stage (think of rename +
2704                          * add-source case).
2705                          */
2706                         remove_file(o, 1, ren1_src,
2707                                     renamed_stage == 2 || !was_tracked(o, ren1_src));
2708
2709                         oidcpy(&src_other.oid,
2710                                &ren1->src_entry->stages[other_stage].oid);
2711                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2712                         oidcpy(&dst_other.oid,
2713                                &ren1->dst_entry->stages[other_stage].oid);
2714                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2715                         try_merge = 0;
2716
2717                         if (oid_eq(&src_other.oid, &null_oid) &&
2718                             ren1->add_turned_into_rename) {
2719                                 setup_rename_conflict_info(RENAME_VIA_DIR,
2720                                                            ren1->pair,
2721                                                            NULL,
2722                                                            branch1,
2723                                                            branch2,
2724                                                            ren1->dst_entry,
2725                                                            NULL,
2726                                                            o,
2727                                                            NULL,
2728                                                            NULL);
2729                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2730                                 setup_rename_conflict_info(RENAME_DELETE,
2731                                                            ren1->pair,
2732                                                            NULL,
2733                                                            branch1,
2734                                                            branch2,
2735                                                            ren1->dst_entry,
2736                                                            NULL,
2737                                                            o,
2738                                                            NULL,
2739                                                            NULL);
2740                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2741                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2742                                 /*
2743                                  * Added file on the other side identical to
2744                                  * the file being renamed: clean merge.
2745                                  * Also, there is no need to overwrite the
2746                                  * file already in the working copy, so call
2747                                  * update_file_flags() instead of
2748                                  * update_file().
2749                                  */
2750                                 if (update_file_flags(o,
2751                                                       &ren1->pair->two->oid,
2752                                                       ren1->pair->two->mode,
2753                                                       ren1_dst,
2754                                                       1, /* update_cache */
2755                                                       0  /* update_wd    */))
2756                                         clean_merge = -1;
2757                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2758                                 clean_merge = 0;
2759                                 try_merge = 1;
2760                                 output(o, 1, _("CONFLICT (rename/add): Rename %s->%s in %s. "
2761                                        "%s added in %s"),
2762                                        ren1_src, ren1_dst, branch1,
2763                                        ren1_dst, branch2);
2764                                 if (o->call_depth) {
2765                                         struct merge_file_info mfi;
2766                                         if (merge_file_one(o, ren1_dst, &null_oid, 0,
2767                                                            &ren1->pair->two->oid,
2768                                                            ren1->pair->two->mode,
2769                                                            &dst_other.oid,
2770                                                            dst_other.mode,
2771                                                            branch1, branch2, &mfi)) {
2772                                                 clean_merge = -1;
2773                                                 goto cleanup_and_return;
2774                                         }
2775                                         output(o, 1, _("Adding merged %s"), ren1_dst);
2776                                         if (update_file(o, 0, &mfi.oid,
2777                                                         mfi.mode, ren1_dst))
2778                                                 clean_merge = -1;
2779                                         try_merge = 0;
2780                                 } else {
2781                                         char *new_path = unique_path(o, ren1_dst, branch2);
2782                                         output(o, 1, _("Adding as %s instead"), new_path);
2783                                         if (update_file(o, 0, &dst_other.oid,
2784                                                         dst_other.mode, new_path))
2785                                                 clean_merge = -1;
2786                                         free(new_path);
2787                                 }
2788                         } else
2789                                 try_merge = 1;
2790
2791                         if (clean_merge < 0)
2792                                 goto cleanup_and_return;
2793                         if (try_merge) {
2794                                 struct diff_filespec *one, *a, *b;
2795                                 src_other.path = (char *)ren1_src;
2796
2797                                 one = ren1->pair->one;
2798                                 if (a_renames == renames1) {
2799                                         a = ren1->pair->two;
2800                                         b = &src_other;
2801                                 } else {
2802                                         b = ren1->pair->two;
2803                                         a = &src_other;
2804                                 }
2805                                 update_entry(ren1->dst_entry, one, a, b);
2806                                 setup_rename_conflict_info(RENAME_NORMAL,
2807                                                            ren1->pair,
2808                                                            NULL,
2809                                                            branch1,
2810                                                            NULL,
2811                                                            ren1->dst_entry,
2812                                                            NULL,
2813                                                            o,
2814                                                            NULL,
2815                                                            NULL);
2816                         }
2817                 }
2818         }
2819 cleanup_and_return:
2820         string_list_clear(&a_by_dst, 0);
2821         string_list_clear(&b_by_dst, 0);
2822
2823         return clean_merge;
2824 }
2825
2826 struct rename_info {
2827         struct string_list *head_renames;
2828         struct string_list *merge_renames;
2829 };
2830
2831 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2832                                    struct hashmap *dir_renames)
2833 {
2834         struct hashmap_iter iter;
2835         struct dir_rename_entry *e;
2836
2837         hashmap_iter_init(dir_renames, &iter);
2838         while ((e = hashmap_iter_next(&iter))) {
2839                 free(e->dir);
2840                 strbuf_release(&e->new_dir);
2841                 /* possible_new_dirs already cleared in get_directory_renames */
2842         }
2843         hashmap_free(dir_renames, 1);
2844         free(dir_renames);
2845
2846         free(pairs->queue);
2847         free(pairs);
2848 }
2849
2850 static int detect_and_process_renames(struct merge_options *o,
2851                                       struct tree *common,
2852                                       struct tree *head,
2853                                       struct tree *merge,
2854                                       struct string_list *entries,
2855                                       struct rename_info *ri)
2856 {
2857         struct diff_queue_struct *head_pairs, *merge_pairs;
2858         struct hashmap *dir_re_head, *dir_re_merge;
2859         int clean = 1;
2860
2861         ri->head_renames = NULL;
2862         ri->merge_renames = NULL;
2863
2864         if (!merge_detect_rename(o))
2865                 return 1;
2866
2867         head_pairs = get_diffpairs(o, common, head);
2868         merge_pairs = get_diffpairs(o, common, merge);
2869
2870         dir_re_head = get_directory_renames(head_pairs, head);
2871         dir_re_merge = get_directory_renames(merge_pairs, merge);
2872
2873         handle_directory_level_conflicts(o,
2874                                          dir_re_head, head,
2875                                          dir_re_merge, merge);
2876
2877         ri->head_renames  = get_renames(o, head_pairs,
2878                                         dir_re_merge, dir_re_head, head,
2879                                         common, head, merge, entries,
2880                                         &clean);
2881         if (clean < 0)
2882                 goto cleanup;
2883         ri->merge_renames = get_renames(o, merge_pairs,
2884                                         dir_re_head, dir_re_merge, merge,
2885                                         common, head, merge, entries,
2886                                         &clean);
2887         if (clean < 0)
2888                 goto cleanup;
2889         clean &= process_renames(o, ri->head_renames, ri->merge_renames);
2890
2891 cleanup:
2892         /*
2893          * Some cleanup is deferred until cleanup_renames() because the
2894          * data structures are still needed and referenced in
2895          * process_entry().  But there are a few things we can free now.
2896          */
2897         initial_cleanup_rename(head_pairs, dir_re_head);
2898         initial_cleanup_rename(merge_pairs, dir_re_merge);
2899
2900         return clean;
2901 }
2902
2903 static void final_cleanup_rename(struct string_list *rename)
2904 {
2905         const struct rename *re;
2906         int i;
2907
2908         if (rename == NULL)
2909                 return;
2910
2911         for (i = 0; i < rename->nr; i++) {
2912                 re = rename->items[i].util;
2913                 diff_free_filepair(re->pair);
2914         }
2915         string_list_clear(rename, 1);
2916         free(rename);
2917 }
2918
2919 static void final_cleanup_renames(struct rename_info *re_info)
2920 {
2921         final_cleanup_rename(re_info->head_renames);
2922         final_cleanup_rename(re_info->merge_renames);
2923 }
2924
2925 static struct object_id *stage_oid(const struct object_id *oid, unsigned mode)
2926 {
2927         return (is_null_oid(oid) || mode == 0) ? NULL: (struct object_id *)oid;
2928 }
2929
2930 static int read_oid_strbuf(struct merge_options *o,
2931                            const struct object_id *oid,
2932                            struct strbuf *dst)
2933 {
2934         void *buf;
2935         enum object_type type;
2936         unsigned long size;
2937         buf = read_object_file(oid, &type, &size);
2938         if (!buf)
2939                 return err(o, _("cannot read object %s"), oid_to_hex(oid));
2940         if (type != OBJ_BLOB) {
2941                 free(buf);
2942                 return err(o, _("object %s is not a blob"), oid_to_hex(oid));
2943         }
2944         strbuf_attach(dst, buf, size, size + 1);
2945         return 0;
2946 }
2947
2948 static int blob_unchanged(struct merge_options *opt,
2949                           const struct object_id *o_oid,
2950                           unsigned o_mode,
2951                           const struct object_id *a_oid,
2952                           unsigned a_mode,
2953                           int renormalize, const char *path)
2954 {
2955         struct strbuf o = STRBUF_INIT;
2956         struct strbuf a = STRBUF_INIT;
2957         int ret = 0; /* assume changed for safety */
2958
2959         if (a_mode != o_mode)
2960                 return 0;
2961         if (oid_eq(o_oid, a_oid))
2962                 return 1;
2963         if (!renormalize)
2964                 return 0;
2965
2966         assert(o_oid && a_oid);
2967         if (read_oid_strbuf(opt, o_oid, &o) || read_oid_strbuf(opt, a_oid, &a))
2968                 goto error_return;
2969         /*
2970          * Note: binary | is used so that both renormalizations are
2971          * performed.  Comparison can be skipped if both files are
2972          * unchanged since their sha1s have already been compared.
2973          */
2974         if (renormalize_buffer(&the_index, path, o.buf, o.len, &o) |
2975             renormalize_buffer(&the_index, path, a.buf, a.len, &a))
2976                 ret = (o.len == a.len && !memcmp(o.buf, a.buf, o.len));
2977
2978 error_return:
2979         strbuf_release(&o);
2980         strbuf_release(&a);
2981         return ret;
2982 }
2983
2984 static int handle_modify_delete(struct merge_options *o,
2985                                 const char *path,
2986                                 struct object_id *o_oid, int o_mode,
2987                                 struct object_id *a_oid, int a_mode,
2988                                 struct object_id *b_oid, int b_mode)
2989 {
2990         const char *modify_branch, *delete_branch;
2991         struct object_id *changed_oid;
2992         int changed_mode;
2993
2994         if (a_oid) {
2995                 modify_branch = o->branch1;
2996                 delete_branch = o->branch2;
2997                 changed_oid = a_oid;
2998                 changed_mode = a_mode;
2999         } else {
3000                 modify_branch = o->branch2;
3001                 delete_branch = o->branch1;
3002                 changed_oid = b_oid;
3003                 changed_mode = b_mode;
3004         }
3005
3006         return handle_change_delete(o,
3007                                     path, NULL,
3008                                     o_oid, o_mode,
3009                                     changed_oid, changed_mode,
3010                                     modify_branch, delete_branch,
3011                                     _("modify"), _("modified"));
3012 }
3013
3014 static int merge_content(struct merge_options *o,
3015                          const char *path,
3016                          int is_dirty,
3017                          struct object_id *o_oid, int o_mode,
3018                          struct object_id *a_oid, int a_mode,
3019                          struct object_id *b_oid, int b_mode,
3020                          struct rename_conflict_info *rename_conflict_info)
3021 {
3022         const char *reason = _("content");
3023         const char *path1 = NULL, *path2 = NULL;
3024         struct merge_file_info mfi;
3025         struct diff_filespec one, a, b;
3026         unsigned df_conflict_remains = 0;
3027
3028         if (!o_oid) {
3029                 reason = _("add/add");
3030                 o_oid = (struct object_id *)&null_oid;
3031         }
3032         one.path = a.path = b.path = (char *)path;
3033         oidcpy(&one.oid, o_oid);
3034         one.mode = o_mode;
3035         oidcpy(&a.oid, a_oid);
3036         a.mode = a_mode;
3037         oidcpy(&b.oid, b_oid);
3038         b.mode = b_mode;
3039
3040         if (rename_conflict_info) {
3041                 struct diff_filepair *pair1 = rename_conflict_info->pair1;
3042
3043                 path1 = (o->branch1 == rename_conflict_info->branch1) ?
3044                         pair1->two->path : pair1->one->path;
3045                 /* If rename_conflict_info->pair2 != NULL, we are in
3046                  * RENAME_ONE_FILE_TO_ONE case.  Otherwise, we have a
3047                  * normal rename.
3048                  */
3049                 path2 = (rename_conflict_info->pair2 ||
3050                          o->branch2 == rename_conflict_info->branch1) ?
3051                         pair1->two->path : pair1->one->path;
3052
3053                 if (dir_in_way(path, !o->call_depth,
3054                                S_ISGITLINK(pair1->two->mode)))
3055                         df_conflict_remains = 1;
3056         }
3057         if (merge_file_special_markers(o, &one, &a, &b, path,
3058                                        o->branch1, path1,
3059                                        o->branch2, path2, &mfi))
3060                 return -1;
3061
3062         /*
3063          * We can skip updating the working tree file iff:
3064          *   a) The merge is clean
3065          *   b) The merge matches what was in HEAD (content, mode, pathname)
3066          *   c) The target path is usable (i.e. not involved in D/F conflict)
3067          */
3068         if (mfi.clean &&
3069             was_tracked_and_matches(o, path, &mfi.oid, mfi.mode) &&
3070             !df_conflict_remains) {
3071                 output(o, 3, _("Skipped %s (merged same as existing)"), path);
3072                 if (add_cacheinfo(o, mfi.mode, &mfi.oid, path,
3073                                   0, (!o->call_depth && !is_dirty), 0))
3074                         return -1;
3075                 return mfi.clean;
3076         }
3077
3078         if (!mfi.clean) {
3079                 if (S_ISGITLINK(mfi.mode))
3080                         reason = _("submodule");
3081                 output(o, 1, _("CONFLICT (%s): Merge conflict in %s"),
3082                                 reason, path);
3083                 if (rename_conflict_info && !df_conflict_remains)
3084                         if (update_stages(o, path, &one, &a, &b))
3085                                 return -1;
3086         }
3087
3088         if (df_conflict_remains || is_dirty) {
3089                 char *new_path;
3090                 if (o->call_depth) {
3091                         remove_file_from_cache(path);
3092                 } else {
3093                         if (!mfi.clean) {
3094                                 if (update_stages(o, path, &one, &a, &b))
3095                                         return -1;
3096                         } else {
3097                                 int file_from_stage2 = was_tracked(o, path);
3098                                 struct diff_filespec merged;
3099                                 oidcpy(&merged.oid, &mfi.oid);
3100                                 merged.mode = mfi.mode;
3101
3102                                 if (update_stages(o, path, NULL,
3103                                                   file_from_stage2 ? &merged : NULL,
3104                                                   file_from_stage2 ? NULL : &merged))
3105                                         return -1;
3106                         }
3107
3108                 }
3109                 new_path = unique_path(o, path, rename_conflict_info->branch1);
3110                 if (is_dirty) {
3111                         output(o, 1, _("Refusing to lose dirty file at %s"),
3112                                path);
3113                 }
3114                 output(o, 1, _("Adding as %s instead"), new_path);
3115                 if (update_file(o, 0, &mfi.oid, mfi.mode, new_path)) {
3116                         free(new_path);
3117                         return -1;
3118                 }
3119                 free(new_path);
3120                 mfi.clean = 0;
3121         } else if (update_file(o, mfi.clean, &mfi.oid, mfi.mode, path))
3122                 return -1;
3123         return !is_dirty && mfi.clean;
3124 }
3125
3126 static int handle_rename_normal(struct merge_options *o,
3127                                 const char *path,
3128                                 struct object_id *o_oid, unsigned int o_mode,
3129                                 struct object_id *a_oid, unsigned int a_mode,
3130                                 struct object_id *b_oid, unsigned int b_mode,
3131                                 struct rename_conflict_info *ci)
3132 {
3133         /* Merge the content and write it out */
3134         return merge_content(o, path, was_dirty(o, path),
3135                              o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3136                              ci);
3137 }
3138
3139 /* Per entry merge function */
3140 static int process_entry(struct merge_options *o,
3141                          const char *path, struct stage_data *entry)
3142 {
3143         int clean_merge = 1;
3144         int normalize = o->renormalize;
3145         unsigned o_mode = entry->stages[1].mode;
3146         unsigned a_mode = entry->stages[2].mode;
3147         unsigned b_mode = entry->stages[3].mode;
3148         struct object_id *o_oid = stage_oid(&entry->stages[1].oid, o_mode);
3149         struct object_id *a_oid = stage_oid(&entry->stages[2].oid, a_mode);
3150         struct object_id *b_oid = stage_oid(&entry->stages[3].oid, b_mode);
3151
3152         entry->processed = 1;
3153         if (entry->rename_conflict_info) {
3154                 struct rename_conflict_info *conflict_info = entry->rename_conflict_info;
3155                 switch (conflict_info->rename_type) {
3156                 case RENAME_NORMAL:
3157                 case RENAME_ONE_FILE_TO_ONE:
3158                         clean_merge = handle_rename_normal(o,
3159                                                            path,
3160                                                            o_oid, o_mode,
3161                                                            a_oid, a_mode,
3162                                                            b_oid, b_mode,
3163                                                            conflict_info);
3164                         break;
3165                 case RENAME_VIA_DIR:
3166                         clean_merge = 1;
3167                         if (handle_rename_via_dir(o,
3168                                                   conflict_info->pair1,
3169                                                   conflict_info->branch1,
3170                                                   conflict_info->branch2))
3171                                 clean_merge = -1;
3172                         break;
3173                 case RENAME_DELETE:
3174                         clean_merge = 0;
3175                         if (handle_rename_delete(o,
3176                                                  conflict_info->pair1,
3177                                                  conflict_info->branch1,
3178                                                  conflict_info->branch2))
3179                                 clean_merge = -1;
3180                         break;
3181                 case RENAME_ONE_FILE_TO_TWO:
3182                         clean_merge = 0;
3183                         if (handle_rename_rename_1to2(o, conflict_info))
3184                                 clean_merge = -1;
3185                         break;
3186                 case RENAME_TWO_FILES_TO_ONE:
3187                         clean_merge = 0;
3188                         if (handle_rename_rename_2to1(o, conflict_info))
3189                                 clean_merge = -1;
3190                         break;
3191                 default:
3192                         entry->processed = 0;
3193                         break;
3194                 }
3195         } else if (o_oid && (!a_oid || !b_oid)) {
3196                 /* Case A: Deleted in one */
3197                 if ((!a_oid && !b_oid) ||
3198                     (!b_oid && blob_unchanged(o, o_oid, o_mode, a_oid, a_mode, normalize, path)) ||
3199                     (!a_oid && blob_unchanged(o, o_oid, o_mode, b_oid, b_mode, normalize, path))) {
3200                         /* Deleted in both or deleted in one and
3201                          * unchanged in the other */
3202                         if (a_oid)
3203                                 output(o, 2, _("Removing %s"), path);
3204                         /* do not touch working file if it did not exist */
3205                         remove_file(o, 1, path, !a_oid);
3206                 } else {
3207                         /* Modify/delete; deleted side may have put a directory in the way */
3208                         clean_merge = 0;
3209                         if (handle_modify_delete(o, path, o_oid, o_mode,
3210                                                  a_oid, a_mode, b_oid, b_mode))
3211                                 clean_merge = -1;
3212                 }
3213         } else if ((!o_oid && a_oid && !b_oid) ||
3214                    (!o_oid && !a_oid && b_oid)) {
3215                 /* Case B: Added in one. */
3216                 /* [nothing|directory] -> ([nothing|directory], file) */
3217
3218                 const char *add_branch;
3219                 const char *other_branch;
3220                 unsigned mode;
3221                 const struct object_id *oid;
3222                 const char *conf;
3223
3224                 if (a_oid) {
3225                         add_branch = o->branch1;
3226                         other_branch = o->branch2;
3227                         mode = a_mode;
3228                         oid = a_oid;
3229                         conf = _("file/directory");
3230                 } else {
3231                         add_branch = o->branch2;
3232                         other_branch = o->branch1;
3233                         mode = b_mode;
3234                         oid = b_oid;
3235                         conf = _("directory/file");
3236                 }
3237                 if (dir_in_way(path,
3238                                !o->call_depth && !S_ISGITLINK(a_mode),
3239                                0)) {
3240                         char *new_path = unique_path(o, path, add_branch);
3241                         clean_merge = 0;
3242                         output(o, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3243                                "Adding %s as %s"),
3244                                conf, path, other_branch, path, new_path);
3245                         if (update_file(o, 0, oid, mode, new_path))
3246                                 clean_merge = -1;
3247                         else if (o->call_depth)
3248                                 remove_file_from_cache(path);
3249                         free(new_path);
3250                 } else {
3251                         output(o, 2, _("Adding %s"), path);
3252                         /* do not overwrite file if already present */
3253                         if (update_file_flags(o, oid, mode, path, 1, !a_oid))
3254                                 clean_merge = -1;
3255                 }
3256         } else if (a_oid && b_oid) {
3257                 /* Case C: Added in both (check for same permissions) and */
3258                 /* case D: Modified in both, but differently. */
3259                 int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3260                 clean_merge = merge_content(o, path, is_dirty,
3261                                             o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3262                                             NULL);
3263         } else if (!o_oid && !a_oid && !b_oid) {
3264                 /*
3265                  * this entry was deleted altogether. a_mode == 0 means
3266                  * we had that path and want to actively remove it.
3267                  */
3268                 remove_file(o, 1, path, !a_mode);
3269         } else
3270                 BUG("fatal merge failure, shouldn't happen.");
3271
3272         return clean_merge;
3273 }
3274
3275 int merge_trees(struct merge_options *o,
3276                 struct tree *head,
3277                 struct tree *merge,
3278                 struct tree *common,
3279                 struct tree **result)
3280 {
3281         int code, clean;
3282
3283         if (o->subtree_shift) {
3284                 merge = shift_tree_object(head, merge, o->subtree_shift);
3285                 common = shift_tree_object(head, common, o->subtree_shift);
3286         }
3287
3288         if (oid_eq(&common->object.oid, &merge->object.oid)) {
3289                 struct strbuf sb = STRBUF_INIT;
3290
3291                 if (!o->call_depth && index_has_changes(&sb)) {
3292                         err(o, _("Dirty index: cannot merge (dirty: %s)"),
3293                             sb.buf);
3294                         return 0;
3295                 }
3296                 output(o, 0, _("Already up to date!"));
3297                 *result = head;
3298                 return 1;
3299         }
3300
3301         code = unpack_trees_start(o, common, head, merge);
3302
3303         if (code != 0) {
3304                 if (show(o, 4) || o->call_depth)
3305                         err(o, _("merging of trees %s and %s failed"),
3306                             oid_to_hex(&head->object.oid),
3307                             oid_to_hex(&merge->object.oid));
3308                 unpack_trees_finish(o);
3309                 return -1;
3310         }
3311
3312         if (unmerged_cache()) {
3313                 struct string_list *entries;
3314                 struct rename_info re_info;
3315                 int i;
3316                 /*
3317                  * Only need the hashmap while processing entries, so
3318                  * initialize it here and free it when we are done running
3319                  * through the entries. Keeping it in the merge_options as
3320                  * opposed to decaring a local hashmap is for convenience
3321                  * so that we don't have to pass it to around.
3322                  */
3323                 hashmap_init(&o->current_file_dir_set, path_hashmap_cmp, NULL, 512);
3324                 get_files_dirs(o, head);
3325                 get_files_dirs(o, merge);
3326
3327                 entries = get_unmerged();
3328                 clean = detect_and_process_renames(o, common, head, merge,
3329                                                    entries, &re_info);
3330                 record_df_conflict_files(o, entries);
3331                 if (clean < 0)
3332                         goto cleanup;
3333                 for (i = entries->nr-1; 0 <= i; i--) {
3334                         const char *path = entries->items[i].string;
3335                         struct stage_data *e = entries->items[i].util;
3336                         if (!e->processed) {
3337                                 int ret = process_entry(o, path, e);
3338                                 if (!ret)
3339                                         clean = 0;
3340                                 else if (ret < 0) {
3341                                         clean = ret;
3342                                         goto cleanup;
3343                                 }
3344                         }
3345                 }
3346                 for (i = 0; i < entries->nr; i++) {
3347                         struct stage_data *e = entries->items[i].util;
3348                         if (!e->processed)
3349                                 BUG("unprocessed path??? %s",
3350                                     entries->items[i].string);
3351                 }
3352
3353         cleanup:
3354                 final_cleanup_renames(&re_info);
3355
3356                 string_list_clear(entries, 1);
3357                 free(entries);
3358
3359                 hashmap_free(&o->current_file_dir_set, 1);
3360
3361                 if (clean < 0) {
3362                         unpack_trees_finish(o);
3363                         return clean;
3364                 }
3365         }
3366         else
3367                 clean = 1;
3368
3369         unpack_trees_finish(o);
3370
3371         if (o->call_depth && !(*result = write_tree_from_memory(o)))
3372                 return -1;
3373
3374         return clean;
3375 }
3376
3377 static struct commit_list *reverse_commit_list(struct commit_list *list)
3378 {
3379         struct commit_list *next = NULL, *current, *backup;
3380         for (current = list; current; current = backup) {
3381                 backup = current->next;
3382                 current->next = next;
3383                 next = current;
3384         }
3385         return next;
3386 }
3387
3388 /*
3389  * Merge the commits h1 and h2, return the resulting virtual
3390  * commit object and a flag indicating the cleanness of the merge.
3391  */
3392 int merge_recursive(struct merge_options *o,
3393                     struct commit *h1,
3394                     struct commit *h2,
3395                     struct commit_list *ca,
3396                     struct commit **result)
3397 {
3398         struct commit_list *iter;
3399         struct commit *merged_common_ancestors;
3400         struct tree *mrtree;
3401         int clean;
3402
3403         if (show(o, 4)) {
3404                 output(o, 4, _("Merging:"));
3405                 output_commit_title(o, h1);
3406                 output_commit_title(o, h2);
3407         }
3408
3409         if (!ca) {
3410                 ca = get_merge_bases(h1, h2);
3411                 ca = reverse_commit_list(ca);
3412         }
3413
3414         if (show(o, 5)) {
3415                 unsigned cnt = commit_list_count(ca);
3416
3417                 output(o, 5, Q_("found %u common ancestor:",
3418                                 "found %u common ancestors:", cnt), cnt);
3419                 for (iter = ca; iter; iter = iter->next)
3420                         output_commit_title(o, iter->item);
3421         }
3422
3423         merged_common_ancestors = pop_commit(&ca);
3424         if (merged_common_ancestors == NULL) {
3425                 /* if there is no common ancestor, use an empty tree */
3426                 struct tree *tree;
3427
3428                 tree = lookup_tree(the_hash_algo->empty_tree);
3429                 merged_common_ancestors = make_virtual_commit(tree, "ancestor");
3430         }
3431
3432         for (iter = ca; iter; iter = iter->next) {
3433                 const char *saved_b1, *saved_b2;
3434                 o->call_depth++;
3435                 /*
3436                  * When the merge fails, the result contains files
3437                  * with conflict markers. The cleanness flag is
3438                  * ignored (unless indicating an error), it was never
3439                  * actually used, as result of merge_trees has always
3440                  * overwritten it: the committed "conflicts" were
3441                  * already resolved.
3442                  */
3443                 discard_cache();
3444                 saved_b1 = o->branch1;
3445                 saved_b2 = o->branch2;
3446                 o->branch1 = "Temporary merge branch 1";
3447                 o->branch2 = "Temporary merge branch 2";
3448                 if (merge_recursive(o, merged_common_ancestors, iter->item,
3449                                     NULL, &merged_common_ancestors) < 0)
3450                         return -1;
3451                 o->branch1 = saved_b1;
3452                 o->branch2 = saved_b2;
3453                 o->call_depth--;
3454
3455                 if (!merged_common_ancestors)
3456                         return err(o, _("merge returned no commit"));
3457         }
3458
3459         discard_cache();
3460         if (!o->call_depth)
3461                 read_cache();
3462
3463         o->ancestor = "merged common ancestors";
3464         clean = merge_trees(o, get_commit_tree(h1), get_commit_tree(h2),
3465                             get_commit_tree(merged_common_ancestors),
3466                             &mrtree);
3467         if (clean < 0) {
3468                 flush_output(o);
3469                 return clean;
3470         }
3471
3472         if (o->call_depth) {
3473                 *result = make_virtual_commit(mrtree, "merged tree");
3474                 commit_list_insert(h1, &(*result)->parents);
3475                 commit_list_insert(h2, &(*result)->parents->next);
3476         }
3477         flush_output(o);
3478         if (!o->call_depth && o->buffer_output < 2)
3479                 strbuf_release(&o->obuf);
3480         if (show(o, 2))
3481                 diff_warn_rename_limit("merge.renamelimit",
3482                                        o->needed_rename_limit, 0);
3483         return clean;
3484 }
3485
3486 static struct commit *get_ref(const struct object_id *oid, const char *name)
3487 {
3488         struct object *object;
3489
3490         object = deref_tag(parse_object(oid), name, strlen(name));
3491         if (!object)
3492                 return NULL;
3493         if (object->type == OBJ_TREE)
3494                 return make_virtual_commit((struct tree*)object, name);
3495         if (object->type != OBJ_COMMIT)
3496                 return NULL;
3497         if (parse_commit((struct commit *)object))
3498                 return NULL;
3499         return (struct commit *)object;
3500 }
3501
3502 int merge_recursive_generic(struct merge_options *o,
3503                             const struct object_id *head,
3504                             const struct object_id *merge,
3505                             int num_base_list,
3506                             const struct object_id **base_list,
3507                             struct commit **result)
3508 {
3509         int clean;
3510         struct lock_file lock = LOCK_INIT;
3511         struct commit *head_commit = get_ref(head, o->branch1);
3512         struct commit *next_commit = get_ref(merge, o->branch2);
3513         struct commit_list *ca = NULL;
3514
3515         if (base_list) {
3516                 int i;
3517                 for (i = 0; i < num_base_list; ++i) {
3518                         struct commit *base;
3519                         if (!(base = get_ref(base_list[i], oid_to_hex(base_list[i]))))
3520                                 return err(o, _("Could not parse object '%s'"),
3521                                            oid_to_hex(base_list[i]));
3522                         commit_list_insert(base, &ca);
3523                 }
3524         }
3525
3526         hold_locked_index(&lock, LOCK_DIE_ON_ERROR);
3527         clean = merge_recursive(o, head_commit, next_commit, ca,
3528                                 result);
3529         if (clean < 0) {
3530                 rollback_lock_file(&lock);
3531                 return clean;
3532         }
3533
3534         if (write_locked_index(&the_index, &lock,
3535                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3536                 return err(o, _("Unable to write index."));
3537
3538         return clean ? 0 : 1;
3539 }
3540
3541 static void merge_recursive_config(struct merge_options *o)
3542 {
3543         char *value = NULL;
3544         git_config_get_int("merge.verbosity", &o->verbosity);
3545         git_config_get_int("diff.renamelimit", &o->diff_rename_limit);
3546         git_config_get_int("merge.renamelimit", &o->merge_rename_limit);
3547         if (!git_config_get_string("diff.renames", &value)) {
3548                 o->diff_detect_rename = git_config_rename("diff.renames", value);
3549                 free(value);
3550         }
3551         if (!git_config_get_string("merge.renames", &value)) {
3552                 o->merge_detect_rename = git_config_rename("merge.renames", value);
3553                 free(value);
3554         }
3555         git_config(git_xmerge_config, NULL);
3556 }
3557
3558 void init_merge_options(struct merge_options *o)
3559 {
3560         const char *merge_verbosity;
3561         memset(o, 0, sizeof(struct merge_options));
3562         o->verbosity = 2;
3563         o->buffer_output = 1;
3564         o->diff_rename_limit = -1;
3565         o->merge_rename_limit = -1;
3566         o->renormalize = 0;
3567         o->diff_detect_rename = -1;
3568         o->merge_detect_rename = -1;
3569         merge_recursive_config(o);
3570         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3571         if (merge_verbosity)
3572                 o->verbosity = strtol(merge_verbosity, NULL, 10);
3573         if (o->verbosity >= 5)
3574                 o->buffer_output = 0;
3575         strbuf_init(&o->obuf, 0);
3576         string_list_init(&o->df_conflict_file_set, 1);
3577 }
3578
3579 int parse_merge_opt(struct merge_options *o, const char *s)
3580 {
3581         const char *arg;
3582
3583         if (!s || !*s)
3584                 return -1;
3585         if (!strcmp(s, "ours"))
3586                 o->recursive_variant = MERGE_RECURSIVE_OURS;
3587         else if (!strcmp(s, "theirs"))
3588                 o->recursive_variant = MERGE_RECURSIVE_THEIRS;
3589         else if (!strcmp(s, "subtree"))
3590                 o->subtree_shift = "";
3591         else if (skip_prefix(s, "subtree=", &arg))
3592                 o->subtree_shift = arg;
3593         else if (!strcmp(s, "patience"))
3594                 o->xdl_opts = DIFF_WITH_ALG(o, PATIENCE_DIFF);
3595         else if (!strcmp(s, "histogram"))
3596                 o->xdl_opts = DIFF_WITH_ALG(o, HISTOGRAM_DIFF);
3597         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3598                 long value = parse_algorithm_value(arg);
3599                 if (value < 0)
3600                         return -1;
3601                 /* clear out previous settings */
3602                 DIFF_XDL_CLR(o, NEED_MINIMAL);
3603                 o->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3604                 o->xdl_opts |= value;
3605         }
3606         else if (!strcmp(s, "ignore-space-change"))
3607                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_CHANGE);
3608         else if (!strcmp(s, "ignore-all-space"))
3609                 DIFF_XDL_SET(o, IGNORE_WHITESPACE);
3610         else if (!strcmp(s, "ignore-space-at-eol"))
3611                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_AT_EOL);
3612         else if (!strcmp(s, "ignore-cr-at-eol"))
3613                 DIFF_XDL_SET(o, IGNORE_CR_AT_EOL);
3614         else if (!strcmp(s, "renormalize"))
3615                 o->renormalize = 1;
3616         else if (!strcmp(s, "no-renormalize"))
3617                 o->renormalize = 0;
3618         else if (!strcmp(s, "no-renames"))
3619                 o->merge_detect_rename = 0;
3620         else if (!strcmp(s, "find-renames")) {
3621                 o->merge_detect_rename = 1;
3622                 o->rename_score = 0;
3623         }
3624         else if (skip_prefix(s, "find-renames=", &arg) ||
3625                  skip_prefix(s, "rename-threshold=", &arg)) {
3626                 if ((o->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3627                         return -1;
3628                 o->merge_detect_rename = 1;
3629         }
3630         else
3631                 return -1;
3632         return 0;
3633 }