builtin/update-index: convert to using the_hash_algo
[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_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 exist 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 conflict_rename_dir(struct merge_options *o,
1417                                struct diff_filepair *pair,
1418                                const char *rename_branch,
1419                                const char *other_branch)
1420 {
1421         const struct diff_filespec *dest = pair->two;
1422
1423         if (!o->call_depth && would_lose_untracked(dest->path)) {
1424                 char *alt_path = unique_path(o, dest->path, rename_branch);
1425
1426                 output(o, 1, _("Error: Refusing to lose untracked file at %s; "
1427                                "writing to %s instead."),
1428                        dest->path, alt_path);
1429                 /*
1430                  * Write the file in worktree at alt_path, but not in the
1431                  * index.  Instead, write to dest->path for the index but
1432                  * only at the higher appropriate stage.
1433                  */
1434                 if (update_file(o, 0, &dest->oid, dest->mode, alt_path))
1435                         return -1;
1436                 free(alt_path);
1437                 return update_stages(o, dest->path, NULL,
1438                                      rename_branch == o->branch1 ? dest : NULL,
1439                                      rename_branch == o->branch1 ? NULL : dest);
1440         }
1441
1442         /* Update dest->path both in index and in worktree */
1443         if (update_file(o, 1, &dest->oid, dest->mode, dest->path))
1444                 return -1;
1445         return 0;
1446 }
1447
1448 static int handle_change_delete(struct merge_options *o,
1449                                  const char *path, const char *old_path,
1450                                  const struct object_id *o_oid, int o_mode,
1451                                  const struct object_id *changed_oid,
1452                                  int changed_mode,
1453                                  const char *change_branch,
1454                                  const char *delete_branch,
1455                                  const char *change, const char *change_past)
1456 {
1457         char *alt_path = NULL;
1458         const char *update_path = path;
1459         int ret = 0;
1460
1461         if (dir_in_way(path, !o->call_depth, 0) ||
1462             (!o->call_depth && would_lose_untracked(path))) {
1463                 update_path = alt_path = unique_path(o, path, change_branch);
1464         }
1465
1466         if (o->call_depth) {
1467                 /*
1468                  * We cannot arbitrarily accept either a_sha or b_sha as
1469                  * correct; since there is no true "middle point" between
1470                  * them, simply reuse the base version for virtual merge base.
1471                  */
1472                 ret = remove_file_from_cache(path);
1473                 if (!ret)
1474                         ret = update_file(o, 0, o_oid, o_mode, update_path);
1475         } else {
1476                 if (!alt_path) {
1477                         if (!old_path) {
1478                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1479                                        "and %s in %s. Version %s of %s left in tree."),
1480                                        change, path, delete_branch, change_past,
1481                                        change_branch, change_branch, path);
1482                         } else {
1483                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1484                                        "and %s to %s in %s. Version %s of %s left in tree."),
1485                                        change, old_path, delete_branch, change_past, path,
1486                                        change_branch, change_branch, path);
1487                         }
1488                 } else {
1489                         if (!old_path) {
1490                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1491                                        "and %s in %s. Version %s of %s left in tree at %s."),
1492                                        change, path, delete_branch, change_past,
1493                                        change_branch, change_branch, path, alt_path);
1494                         } else {
1495                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1496                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1497                                        change, old_path, delete_branch, change_past, path,
1498                                        change_branch, change_branch, path, alt_path);
1499                         }
1500                 }
1501                 /*
1502                  * No need to call update_file() on path when change_branch ==
1503                  * o->branch1 && !alt_path, since that would needlessly touch
1504                  * path.  We could call update_file_flags() with update_cache=0
1505                  * and update_wd=0, but that's a no-op.
1506                  */
1507                 if (change_branch != o->branch1 || alt_path)
1508                         ret = update_file(o, 0, changed_oid, changed_mode, update_path);
1509         }
1510         free(alt_path);
1511
1512         return ret;
1513 }
1514
1515 static int conflict_rename_delete(struct merge_options *o,
1516                                    struct diff_filepair *pair,
1517                                    const char *rename_branch,
1518                                    const char *delete_branch)
1519 {
1520         const struct diff_filespec *orig = pair->one;
1521         const struct diff_filespec *dest = pair->two;
1522
1523         if (handle_change_delete(o,
1524                                  o->call_depth ? orig->path : dest->path,
1525                                  o->call_depth ? NULL : orig->path,
1526                                  &orig->oid, orig->mode,
1527                                  &dest->oid, dest->mode,
1528                                  rename_branch, delete_branch,
1529                                  _("rename"), _("renamed")))
1530                 return -1;
1531
1532         if (o->call_depth)
1533                 return remove_file_from_cache(dest->path);
1534         else
1535                 return update_stages(o, dest->path, NULL,
1536                                      rename_branch == o->branch1 ? dest : NULL,
1537                                      rename_branch == o->branch1 ? NULL : dest);
1538 }
1539
1540 static struct diff_filespec *filespec_from_entry(struct diff_filespec *target,
1541                                                  struct stage_data *entry,
1542                                                  int stage)
1543 {
1544         struct object_id *oid = &entry->stages[stage].oid;
1545         unsigned mode = entry->stages[stage].mode;
1546         if (mode == 0 || is_null_oid(oid))
1547                 return NULL;
1548         oidcpy(&target->oid, oid);
1549         target->mode = mode;
1550         return target;
1551 }
1552
1553 static int handle_file(struct merge_options *o,
1554                         struct diff_filespec *rename,
1555                         int stage,
1556                         struct rename_conflict_info *ci)
1557 {
1558         char *dst_name = rename->path;
1559         struct stage_data *dst_entry;
1560         const char *cur_branch, *other_branch;
1561         struct diff_filespec other;
1562         struct diff_filespec *add;
1563         int ret;
1564
1565         if (stage == 2) {
1566                 dst_entry = ci->dst_entry1;
1567                 cur_branch = ci->branch1;
1568                 other_branch = ci->branch2;
1569         } else {
1570                 dst_entry = ci->dst_entry2;
1571                 cur_branch = ci->branch2;
1572                 other_branch = ci->branch1;
1573         }
1574
1575         add = filespec_from_entry(&other, dst_entry, stage ^ 1);
1576         if (add) {
1577                 int ren_src_was_dirty = was_dirty(o, rename->path);
1578                 char *add_name = unique_path(o, rename->path, other_branch);
1579                 if (update_file(o, 0, &add->oid, add->mode, add_name))
1580                         return -1;
1581
1582                 if (ren_src_was_dirty) {
1583                         output(o, 1, _("Refusing to lose dirty file at %s"),
1584                                rename->path);
1585                 }
1586                 /*
1587                  * Because the double negatives somehow keep confusing me...
1588                  *    1) update_wd iff !ren_src_was_dirty.
1589                  *    2) no_wd iff !update_wd
1590                  *    3) so, no_wd == !!ren_src_was_dirty == ren_src_was_dirty
1591                  */
1592                 remove_file(o, 0, rename->path, ren_src_was_dirty);
1593                 dst_name = unique_path(o, rename->path, cur_branch);
1594         } else {
1595                 if (dir_in_way(rename->path, !o->call_depth, 0)) {
1596                         dst_name = unique_path(o, rename->path, cur_branch);
1597                         output(o, 1, _("%s is a directory in %s adding as %s instead"),
1598                                rename->path, other_branch, dst_name);
1599                 } else if (!o->call_depth &&
1600                            would_lose_untracked(rename->path)) {
1601                         dst_name = unique_path(o, rename->path, cur_branch);
1602                         output(o, 1, _("Refusing to lose untracked file at %s; "
1603                                        "adding as %s instead"),
1604                                rename->path, dst_name);
1605                 }
1606         }
1607         if ((ret = update_file(o, 0, &rename->oid, rename->mode, dst_name)))
1608                 ; /* fall through, do allow dst_name to be released */
1609         else if (stage == 2)
1610                 ret = update_stages(o, rename->path, NULL, rename, add);
1611         else
1612                 ret = update_stages(o, rename->path, NULL, add, rename);
1613
1614         if (dst_name != rename->path)
1615                 free(dst_name);
1616
1617         return ret;
1618 }
1619
1620 static int conflict_rename_rename_1to2(struct merge_options *o,
1621                                         struct rename_conflict_info *ci)
1622 {
1623         /* One file was renamed in both branches, but to different names. */
1624         struct diff_filespec *one = ci->pair1->one;
1625         struct diff_filespec *a = ci->pair1->two;
1626         struct diff_filespec *b = ci->pair2->two;
1627
1628         output(o, 1, _("CONFLICT (rename/rename): "
1629                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1630                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1631                one->path, a->path, ci->branch1,
1632                one->path, b->path, ci->branch2,
1633                o->call_depth ? _(" (left unresolved)") : "");
1634         if (o->call_depth) {
1635                 struct merge_file_info mfi;
1636                 struct diff_filespec other;
1637                 struct diff_filespec *add;
1638                 if (merge_file_one(o, one->path,
1639                                  &one->oid, one->mode,
1640                                  &a->oid, a->mode,
1641                                  &b->oid, b->mode,
1642                                  ci->branch1, ci->branch2, &mfi))
1643                         return -1;
1644
1645                 /*
1646                  * FIXME: For rename/add-source conflicts (if we could detect
1647                  * such), this is wrong.  We should instead find a unique
1648                  * pathname and then either rename the add-source file to that
1649                  * unique path, or use that unique path instead of src here.
1650                  */
1651                 if (update_file(o, 0, &mfi.oid, mfi.mode, one->path))
1652                         return -1;
1653
1654                 /*
1655                  * Above, we put the merged content at the merge-base's
1656                  * path.  Now we usually need to delete both a->path and
1657                  * b->path.  However, the rename on each side of the merge
1658                  * could also be involved in a rename/add conflict.  In
1659                  * such cases, we should keep the added file around,
1660                  * resolving the conflict at that path in its favor.
1661                  */
1662                 add = filespec_from_entry(&other, ci->dst_entry1, 2 ^ 1);
1663                 if (add) {
1664                         if (update_file(o, 0, &add->oid, add->mode, a->path))
1665                                 return -1;
1666                 }
1667                 else
1668                         remove_file_from_cache(a->path);
1669                 add = filespec_from_entry(&other, ci->dst_entry2, 3 ^ 1);
1670                 if (add) {
1671                         if (update_file(o, 0, &add->oid, add->mode, b->path))
1672                                 return -1;
1673                 }
1674                 else
1675                         remove_file_from_cache(b->path);
1676         } else if (handle_file(o, a, 2, ci) || handle_file(o, b, 3, ci))
1677                 return -1;
1678
1679         return 0;
1680 }
1681
1682 static int conflict_rename_rename_2to1(struct merge_options *o,
1683                                         struct rename_conflict_info *ci)
1684 {
1685         /* Two files, a & b, were renamed to the same thing, c. */
1686         struct diff_filespec *a = ci->pair1->one;
1687         struct diff_filespec *b = ci->pair2->one;
1688         struct diff_filespec *c1 = ci->pair1->two;
1689         struct diff_filespec *c2 = ci->pair2->two;
1690         char *path = c1->path; /* == c2->path */
1691         char *path_side_1_desc;
1692         char *path_side_2_desc;
1693         struct merge_file_info mfi_c1;
1694         struct merge_file_info mfi_c2;
1695         int ret;
1696
1697         output(o, 1, _("CONFLICT (rename/rename): "
1698                "Rename %s->%s in %s. "
1699                "Rename %s->%s in %s"),
1700                a->path, c1->path, ci->branch1,
1701                b->path, c2->path, ci->branch2);
1702
1703         remove_file(o, 1, a->path, o->call_depth || would_lose_untracked(a->path));
1704         remove_file(o, 1, b->path, o->call_depth || would_lose_untracked(b->path));
1705
1706         path_side_1_desc = xstrfmt("%s (was %s)", path, a->path);
1707         path_side_2_desc = xstrfmt("%s (was %s)", path, b->path);
1708         if (merge_file_special_markers(o, a, c1, &ci->ren1_other,
1709                                        path_side_1_desc,
1710                                        o->branch1, c1->path,
1711                                        o->branch2, ci->ren1_other.path, &mfi_c1) ||
1712             merge_file_special_markers(o, b, &ci->ren2_other, c2,
1713                                        path_side_2_desc,
1714                                        o->branch1, ci->ren2_other.path,
1715                                        o->branch2, c2->path, &mfi_c2))
1716                 return -1;
1717         free(path_side_1_desc);
1718         free(path_side_2_desc);
1719
1720         if (o->call_depth) {
1721                 /*
1722                  * If mfi_c1.clean && mfi_c2.clean, then it might make
1723                  * sense to do a two-way merge of those results.  But, I
1724                  * think in all cases, it makes sense to have the virtual
1725                  * merge base just undo the renames; they can be detected
1726                  * again later for the non-recursive merge.
1727                  */
1728                 remove_file(o, 0, path, 0);
1729                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, a->path);
1730                 if (!ret)
1731                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1732                                           b->path);
1733         } else {
1734                 char *new_path1 = unique_path(o, path, ci->branch1);
1735                 char *new_path2 = unique_path(o, path, ci->branch2);
1736                 output(o, 1, _("Renaming %s to %s and %s to %s instead"),
1737                        a->path, new_path1, b->path, new_path2);
1738                 if (was_dirty(o, path))
1739                         output(o, 1, _("Refusing to lose dirty file at %s"),
1740                                path);
1741                 else if (would_lose_untracked(path))
1742                         /*
1743                          * Only way we get here is if both renames were from
1744                          * a directory rename AND user had an untracked file
1745                          * at the location where both files end up after the
1746                          * two directory renames.  See testcase 10d of t6043.
1747                          */
1748                         output(o, 1, _("Refusing to lose untracked file at "
1749                                        "%s, even though it's in the way."),
1750                                path);
1751                 else
1752                         remove_file(o, 0, path, 0);
1753                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, new_path1);
1754                 if (!ret)
1755                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1756                                           new_path2);
1757                 /*
1758                  * unpack_trees() actually populates the index for us for
1759                  * "normal" rename/rename(2to1) situtations so that the
1760                  * correct entries are at the higher stages, which would
1761                  * make the call below to update_stages_for_stage_data
1762                  * unnecessary.  However, if either of the renames came
1763                  * from a directory rename, then unpack_trees() will not
1764                  * have gotten the right data loaded into the index, so we
1765                  * need to do so now.  (While it'd be tempting to move this
1766                  * call to update_stages_for_stage_data() to
1767                  * apply_directory_rename_modifications(), that would break
1768                  * our intermediate calls to would_lose_untracked() since
1769                  * those rely on the current in-memory index.  See also the
1770                  * big "NOTE" in update_stages()).
1771                  */
1772                 if (update_stages_for_stage_data(o, path, ci->dst_entry1))
1773                         ret = -1;
1774
1775                 free(new_path2);
1776                 free(new_path1);
1777         }
1778
1779         return ret;
1780 }
1781
1782 /*
1783  * Get the diff_filepairs changed between o_tree and tree.
1784  */
1785 static struct diff_queue_struct *get_diffpairs(struct merge_options *o,
1786                                                struct tree *o_tree,
1787                                                struct tree *tree)
1788 {
1789         struct diff_queue_struct *ret;
1790         struct diff_options opts;
1791
1792         diff_setup(&opts);
1793         opts.flags.recursive = 1;
1794         opts.flags.rename_empty = 0;
1795         opts.detect_rename = merge_detect_rename(o);
1796         /*
1797          * We do not have logic to handle the detection of copies.  In
1798          * fact, it may not even make sense to add such logic: would we
1799          * really want a change to a base file to be propagated through
1800          * multiple other files by a merge?
1801          */
1802         if (opts.detect_rename > DIFF_DETECT_RENAME)
1803                 opts.detect_rename = DIFF_DETECT_RENAME;
1804         opts.rename_limit = o->merge_rename_limit >= 0 ? o->merge_rename_limit :
1805                             o->diff_rename_limit >= 0 ? o->diff_rename_limit :
1806                             1000;
1807         opts.rename_score = o->rename_score;
1808         opts.show_rename_progress = o->show_rename_progress;
1809         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1810         diff_setup_done(&opts);
1811         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1812         diffcore_std(&opts);
1813         if (opts.needed_rename_limit > o->needed_rename_limit)
1814                 o->needed_rename_limit = opts.needed_rename_limit;
1815
1816         ret = xmalloc(sizeof(*ret));
1817         *ret = diff_queued_diff;
1818
1819         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1820         diff_queued_diff.nr = 0;
1821         diff_queued_diff.queue = NULL;
1822         diff_flush(&opts);
1823         return ret;
1824 }
1825
1826 static int tree_has_path(struct tree *tree, const char *path)
1827 {
1828         struct object_id hashy;
1829         unsigned int mode_o;
1830
1831         return !get_tree_entry(&tree->object.oid, path,
1832                                &hashy, &mode_o);
1833 }
1834
1835 /*
1836  * Return a new string that replaces the beginning portion (which matches
1837  * entry->dir), with entry->new_dir.  In perl-speak:
1838  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1839  * NOTE:
1840  *   Caller must ensure that old_path starts with entry->dir + '/'.
1841  */
1842 static char *apply_dir_rename(struct dir_rename_entry *entry,
1843                               const char *old_path)
1844 {
1845         struct strbuf new_path = STRBUF_INIT;
1846         int oldlen, newlen;
1847
1848         if (entry->non_unique_new_dir)
1849                 return NULL;
1850
1851         oldlen = strlen(entry->dir);
1852         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1853         strbuf_grow(&new_path, newlen);
1854         strbuf_addbuf(&new_path, &entry->new_dir);
1855         strbuf_addstr(&new_path, &old_path[oldlen]);
1856
1857         return strbuf_detach(&new_path, NULL);
1858 }
1859
1860 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1861                                     char **old_dir, char **new_dir)
1862 {
1863         char *end_of_old, *end_of_new;
1864         int old_len, new_len;
1865
1866         *old_dir = NULL;
1867         *new_dir = NULL;
1868
1869         /*
1870          * For
1871          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1872          * the "e/foo.c" part is the same, we just want to know that
1873          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1874          * so, for this example, this function returns "a/b/c/d" in
1875          * *old_dir and "a/b/some/thing/else" in *new_dir.
1876          *
1877          * Also, if the basename of the file changed, we don't care.  We
1878          * want to know which portion of the directory, if any, changed.
1879          */
1880         end_of_old = strrchr(old_path, '/');
1881         end_of_new = strrchr(new_path, '/');
1882
1883         if (end_of_old == NULL || end_of_new == NULL)
1884                 return;
1885         while (*--end_of_new == *--end_of_old &&
1886                end_of_old != old_path &&
1887                end_of_new != new_path)
1888                 ; /* Do nothing; all in the while loop */
1889         /*
1890          * We've found the first non-matching character in the directory
1891          * paths.  That means the current directory we were comparing
1892          * represents the rename.  Move end_of_old and end_of_new back
1893          * to the full directory name.
1894          */
1895         if (*end_of_old == '/')
1896                 end_of_old++;
1897         if (*end_of_old != '/')
1898                 end_of_new++;
1899         end_of_old = strchr(end_of_old, '/');
1900         end_of_new = strchr(end_of_new, '/');
1901
1902         /*
1903          * It may have been the case that old_path and new_path were the same
1904          * directory all along.  Don't claim a rename if they're the same.
1905          */
1906         old_len = end_of_old - old_path;
1907         new_len = end_of_new - new_path;
1908
1909         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1910                 *old_dir = xstrndup(old_path, old_len);
1911                 *new_dir = xstrndup(new_path, new_len);
1912         }
1913 }
1914
1915 static void remove_hashmap_entries(struct hashmap *dir_renames,
1916                                    struct string_list *items_to_remove)
1917 {
1918         int i;
1919         struct dir_rename_entry *entry;
1920
1921         for (i = 0; i < items_to_remove->nr; i++) {
1922                 entry = items_to_remove->items[i].util;
1923                 hashmap_remove(dir_renames, entry, NULL);
1924         }
1925         string_list_clear(items_to_remove, 0);
1926 }
1927
1928 /*
1929  * See if there is a directory rename for path, and if there are any file
1930  * level conflicts for the renamed location.  If there is a rename and
1931  * there are no conflicts, return the new name.  Otherwise, return NULL.
1932  */
1933 static char *handle_path_level_conflicts(struct merge_options *o,
1934                                          const char *path,
1935                                          struct dir_rename_entry *entry,
1936                                          struct hashmap *collisions,
1937                                          struct tree *tree)
1938 {
1939         char *new_path = NULL;
1940         struct collision_entry *collision_ent;
1941         int clean = 1;
1942         struct strbuf collision_paths = STRBUF_INIT;
1943
1944         /*
1945          * entry has the mapping of old directory name to new directory name
1946          * that we want to apply to path.
1947          */
1948         new_path = apply_dir_rename(entry, path);
1949
1950         if (!new_path) {
1951                 /* This should only happen when entry->non_unique_new_dir set */
1952                 if (!entry->non_unique_new_dir)
1953                         BUG("entry->non_unqiue_dir not set and !new_path");
1954                 output(o, 1, _("CONFLICT (directory rename split): "
1955                                "Unclear where to place %s because directory "
1956                                "%s was renamed to multiple other directories, "
1957                                "with no destination getting a majority of the "
1958                                "files."),
1959                        path, entry->dir);
1960                 clean = 0;
1961                 return NULL;
1962         }
1963
1964         /*
1965          * The caller needs to have ensured that it has pre-populated
1966          * collisions with all paths that map to new_path.  Do a quick check
1967          * to ensure that's the case.
1968          */
1969         collision_ent = collision_find_entry(collisions, new_path);
1970         if (collision_ent == NULL)
1971                 BUG("collision_ent is NULL");
1972
1973         /*
1974          * Check for one-sided add/add/.../add conflicts, i.e.
1975          * where implicit renames from the other side doing
1976          * directory rename(s) can affect this side of history
1977          * to put multiple paths into the same location.  Warn
1978          * and bail on directory renames for such paths.
1979          */
1980         if (collision_ent->reported_already) {
1981                 clean = 0;
1982         } else if (tree_has_path(tree, new_path)) {
1983                 collision_ent->reported_already = 1;
1984                 strbuf_add_separated_string_list(&collision_paths, ", ",
1985                                                  &collision_ent->source_files);
1986                 output(o, 1, _("CONFLICT (implicit dir rename): Existing "
1987                                "file/dir at %s in the way of implicit "
1988                                "directory rename(s) putting the following "
1989                                "path(s) there: %s."),
1990                        new_path, collision_paths.buf);
1991                 clean = 0;
1992         } else if (collision_ent->source_files.nr > 1) {
1993                 collision_ent->reported_already = 1;
1994                 strbuf_add_separated_string_list(&collision_paths, ", ",
1995                                                  &collision_ent->source_files);
1996                 output(o, 1, _("CONFLICT (implicit dir rename): Cannot map "
1997                                "more than one path to %s; implicit directory "
1998                                "renames tried to put these paths there: %s"),
1999                        new_path, collision_paths.buf);
2000                 clean = 0;
2001         }
2002
2003         /* Free memory we no longer need */
2004         strbuf_release(&collision_paths);
2005         if (!clean && new_path) {
2006                 free(new_path);
2007                 return NULL;
2008         }
2009
2010         return new_path;
2011 }
2012
2013 /*
2014  * There are a couple things we want to do at the directory level:
2015  *   1. Check for both sides renaming to the same thing, in order to avoid
2016  *      implicit renaming of files that should be left in place.  (See
2017  *      testcase 6b in t6043 for details.)
2018  *   2. Prune directory renames if there are still files left in the
2019  *      the original directory.  These represent a partial directory rename,
2020  *      i.e. a rename where only some of the files within the directory
2021  *      were renamed elsewhere.  (Technically, this could be done earlier
2022  *      in get_directory_renames(), except that would prevent us from
2023  *      doing the previous check and thus failing testcase 6b.)
2024  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2025  *      In the future, we could potentially record this info as well and
2026  *      omit reporting rename/rename(1to2) conflicts for each path within
2027  *      the affected directories, thus cleaning up the merge output.
2028  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2029  *         directory level, because merging directories is fine.  If it
2030  *         causes conflicts for files within those merged directories, then
2031  *         that should be detected at the individual path level.
2032  */
2033 static void handle_directory_level_conflicts(struct merge_options *o,
2034                                              struct hashmap *dir_re_head,
2035                                              struct tree *head,
2036                                              struct hashmap *dir_re_merge,
2037                                              struct tree *merge)
2038 {
2039         struct hashmap_iter iter;
2040         struct dir_rename_entry *head_ent;
2041         struct dir_rename_entry *merge_ent;
2042
2043         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2044         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2045
2046         hashmap_iter_init(dir_re_head, &iter);
2047         while ((head_ent = hashmap_iter_next(&iter))) {
2048                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2049                 if (merge_ent &&
2050                     !head_ent->non_unique_new_dir &&
2051                     !merge_ent->non_unique_new_dir &&
2052                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2053                         /* 1. Renamed identically; remove it from both sides */
2054                         string_list_append(&remove_from_head,
2055                                            head_ent->dir)->util = head_ent;
2056                         strbuf_release(&head_ent->new_dir);
2057                         string_list_append(&remove_from_merge,
2058                                            merge_ent->dir)->util = merge_ent;
2059                         strbuf_release(&merge_ent->new_dir);
2060                 } else if (tree_has_path(head, head_ent->dir)) {
2061                         /* 2. This wasn't a directory rename after all */
2062                         string_list_append(&remove_from_head,
2063                                            head_ent->dir)->util = head_ent;
2064                         strbuf_release(&head_ent->new_dir);
2065                 }
2066         }
2067
2068         remove_hashmap_entries(dir_re_head, &remove_from_head);
2069         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2070
2071         hashmap_iter_init(dir_re_merge, &iter);
2072         while ((merge_ent = hashmap_iter_next(&iter))) {
2073                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2074                 if (tree_has_path(merge, merge_ent->dir)) {
2075                         /* 2. This wasn't a directory rename after all */
2076                         string_list_append(&remove_from_merge,
2077                                            merge_ent->dir)->util = merge_ent;
2078                 } else if (head_ent &&
2079                            !head_ent->non_unique_new_dir &&
2080                            !merge_ent->non_unique_new_dir) {
2081                         /* 3. rename/rename(1to2) */
2082                         /*
2083                          * We can assume it's not rename/rename(1to1) because
2084                          * that was case (1), already checked above.  So we
2085                          * know that head_ent->new_dir and merge_ent->new_dir
2086                          * are different strings.
2087                          */
2088                         output(o, 1, _("CONFLICT (rename/rename): "
2089                                        "Rename directory %s->%s in %s. "
2090                                        "Rename directory %s->%s in %s"),
2091                                head_ent->dir, head_ent->new_dir.buf, o->branch1,
2092                                head_ent->dir, merge_ent->new_dir.buf, o->branch2);
2093                         string_list_append(&remove_from_head,
2094                                            head_ent->dir)->util = head_ent;
2095                         strbuf_release(&head_ent->new_dir);
2096                         string_list_append(&remove_from_merge,
2097                                            merge_ent->dir)->util = merge_ent;
2098                         strbuf_release(&merge_ent->new_dir);
2099                 }
2100         }
2101
2102         remove_hashmap_entries(dir_re_head, &remove_from_head);
2103         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2104 }
2105
2106 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs,
2107                                              struct tree *tree)
2108 {
2109         struct hashmap *dir_renames;
2110         struct hashmap_iter iter;
2111         struct dir_rename_entry *entry;
2112         int i;
2113
2114         /*
2115          * Typically, we think of a directory rename as all files from a
2116          * certain directory being moved to a target directory.  However,
2117          * what if someone first moved two files from the original
2118          * directory in one commit, and then renamed the directory
2119          * somewhere else in a later commit?  At merge time, we just know
2120          * that files from the original directory went to two different
2121          * places, and that the bulk of them ended up in the same place.
2122          * We want each directory rename to represent where the bulk of the
2123          * files from that directory end up; this function exists to find
2124          * where the bulk of the files went.
2125          *
2126          * The first loop below simply iterates through the list of file
2127          * renames, finding out how often each directory rename pair
2128          * possibility occurs.
2129          */
2130         dir_renames = xmalloc(sizeof(*dir_renames));
2131         dir_rename_init(dir_renames);
2132         for (i = 0; i < pairs->nr; ++i) {
2133                 struct string_list_item *item;
2134                 int *count;
2135                 struct diff_filepair *pair = pairs->queue[i];
2136                 char *old_dir, *new_dir;
2137
2138                 /* File not part of directory rename if it wasn't renamed */
2139                 if (pair->status != 'R')
2140                         continue;
2141
2142                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2143                                         &old_dir,        &new_dir);
2144                 if (!old_dir)
2145                         /* Directory didn't change at all; ignore this one. */
2146                         continue;
2147
2148                 entry = dir_rename_find_entry(dir_renames, old_dir);
2149                 if (!entry) {
2150                         entry = xmalloc(sizeof(*entry));
2151                         dir_rename_entry_init(entry, old_dir);
2152                         hashmap_put(dir_renames, entry);
2153                 } else {
2154                         free(old_dir);
2155                 }
2156                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2157                 if (!item) {
2158                         item = string_list_insert(&entry->possible_new_dirs,
2159                                                   new_dir);
2160                         item->util = xcalloc(1, sizeof(int));
2161                 } else {
2162                         free(new_dir);
2163                 }
2164                 count = item->util;
2165                 *count += 1;
2166         }
2167
2168         /*
2169          * For each directory with files moved out of it, we find out which
2170          * target directory received the most files so we can declare it to
2171          * be the "winning" target location for the directory rename.  This
2172          * winner gets recorded in new_dir.  If there is no winner
2173          * (multiple target directories received the same number of files),
2174          * we set non_unique_new_dir.  Once we've determined the winner (or
2175          * that there is no winner), we no longer need possible_new_dirs.
2176          */
2177         hashmap_iter_init(dir_renames, &iter);
2178         while ((entry = hashmap_iter_next(&iter))) {
2179                 int max = 0;
2180                 int bad_max = 0;
2181                 char *best = NULL;
2182
2183                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2184                         int *count = entry->possible_new_dirs.items[i].util;
2185
2186                         if (*count == max)
2187                                 bad_max = max;
2188                         else if (*count > max) {
2189                                 max = *count;
2190                                 best = entry->possible_new_dirs.items[i].string;
2191                         }
2192                 }
2193                 if (bad_max == max)
2194                         entry->non_unique_new_dir = 1;
2195                 else {
2196                         assert(entry->new_dir.len == 0);
2197                         strbuf_addstr(&entry->new_dir, best);
2198                 }
2199                 /*
2200                  * The relevant directory sub-portion of the original full
2201                  * filepaths were xstrndup'ed before inserting into
2202                  * possible_new_dirs, and instead of manually iterating the
2203                  * list and free'ing each, just lie and tell
2204                  * possible_new_dirs that it did the strdup'ing so that it
2205                  * will free them for us.
2206                  */
2207                 entry->possible_new_dirs.strdup_strings = 1;
2208                 string_list_clear(&entry->possible_new_dirs, 1);
2209         }
2210
2211         return dir_renames;
2212 }
2213
2214 static struct dir_rename_entry *check_dir_renamed(const char *path,
2215                                                   struct hashmap *dir_renames)
2216 {
2217         char *temp = xstrdup(path);
2218         char *end;
2219         struct dir_rename_entry *entry = NULL;;
2220
2221         while ((end = strrchr(temp, '/'))) {
2222                 *end = '\0';
2223                 entry = dir_rename_find_entry(dir_renames, temp);
2224                 if (entry)
2225                         break;
2226         }
2227         free(temp);
2228         return entry;
2229 }
2230
2231 static void compute_collisions(struct hashmap *collisions,
2232                                struct hashmap *dir_renames,
2233                                struct diff_queue_struct *pairs)
2234 {
2235         int i;
2236
2237         /*
2238          * Multiple files can be mapped to the same path due to directory
2239          * renames done by the other side of history.  Since that other
2240          * side of history could have merged multiple directories into one,
2241          * if our side of history added the same file basename to each of
2242          * those directories, then all N of them would get implicitly
2243          * renamed by the directory rename detection into the same path,
2244          * and we'd get an add/add/.../add conflict, and all those adds
2245          * from *this* side of history.  This is not representable in the
2246          * index, and users aren't going to easily be able to make sense of
2247          * it.  So we need to provide a good warning about what's
2248          * happening, and fall back to no-directory-rename detection
2249          * behavior for those paths.
2250          *
2251          * See testcases 9e and all of section 5 from t6043 for examples.
2252          */
2253         collision_init(collisions);
2254
2255         for (i = 0; i < pairs->nr; ++i) {
2256                 struct dir_rename_entry *dir_rename_ent;
2257                 struct collision_entry *collision_ent;
2258                 char *new_path;
2259                 struct diff_filepair *pair = pairs->queue[i];
2260
2261                 if (pair->status != 'A' && pair->status != 'R')
2262                         continue;
2263                 dir_rename_ent = check_dir_renamed(pair->two->path,
2264                                                    dir_renames);
2265                 if (!dir_rename_ent)
2266                         continue;
2267
2268                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2269                 if (!new_path)
2270                         /*
2271                          * dir_rename_ent->non_unique_new_path is true, which
2272                          * means there is no directory rename for us to use,
2273                          * which means it won't cause us any additional
2274                          * collisions.
2275                          */
2276                         continue;
2277                 collision_ent = collision_find_entry(collisions, new_path);
2278                 if (!collision_ent) {
2279                         collision_ent = xcalloc(1,
2280                                                 sizeof(struct collision_entry));
2281                         hashmap_entry_init(collision_ent, strhash(new_path));
2282                         hashmap_put(collisions, collision_ent);
2283                         collision_ent->target_file = new_path;
2284                 } else {
2285                         free(new_path);
2286                 }
2287                 string_list_insert(&collision_ent->source_files,
2288                                    pair->two->path);
2289         }
2290 }
2291
2292 static char *check_for_directory_rename(struct merge_options *o,
2293                                         const char *path,
2294                                         struct tree *tree,
2295                                         struct hashmap *dir_renames,
2296                                         struct hashmap *dir_rename_exclusions,
2297                                         struct hashmap *collisions,
2298                                         int *clean_merge)
2299 {
2300         char *new_path = NULL;
2301         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2302         struct dir_rename_entry *oentry = NULL;
2303
2304         if (!entry)
2305                 return new_path;
2306
2307         /*
2308          * This next part is a little weird.  We do not want to do an
2309          * implicit rename into a directory we renamed on our side, because
2310          * that will result in a spurious rename/rename(1to2) conflict.  An
2311          * example:
2312          *   Base commit: dumbdir/afile, otherdir/bfile
2313          *   Side 1:      smrtdir/afile, otherdir/bfile
2314          *   Side 2:      dumbdir/afile, dumbdir/bfile
2315          * Here, while working on Side 1, we could notice that otherdir was
2316          * renamed/merged to dumbdir, and change the diff_filepair for
2317          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2318          * 2 will notice the rename from dumbdir to smrtdir, and do the
2319          * transitive rename to move it from dumbdir/bfile to
2320          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2321          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2322          * the file to end up in smrtdir.  And the way to achieve that is
2323          * to not let Side1 do the rename to dumbdir, since we know that is
2324          * the source of one of our directory renames.
2325          *
2326          * That's why oentry and dir_rename_exclusions is here.
2327          *
2328          * As it turns out, this also prevents N-way transient rename
2329          * confusion; See testcases 9c and 9d of t6043.
2330          */
2331         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2332         if (oentry) {
2333                 output(o, 1, _("WARNING: Avoiding applying %s -> %s rename "
2334                                "to %s, because %s itself was renamed."),
2335                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2336         } else {
2337                 new_path = handle_path_level_conflicts(o, path, entry,
2338                                                        collisions, tree);
2339                 *clean_merge &= (new_path != NULL);
2340         }
2341
2342         return new_path;
2343 }
2344
2345 static void apply_directory_rename_modifications(struct merge_options *o,
2346                                                  struct diff_filepair *pair,
2347                                                  char *new_path,
2348                                                  struct rename *re,
2349                                                  struct tree *tree,
2350                                                  struct tree *o_tree,
2351                                                  struct tree *a_tree,
2352                                                  struct tree *b_tree,
2353                                                  struct string_list *entries,
2354                                                  int *clean)
2355 {
2356         struct string_list_item *item;
2357         int stage = (tree == a_tree ? 2 : 3);
2358         int update_wd;
2359
2360         /*
2361          * In all cases where we can do directory rename detection,
2362          * unpack_trees() will have read pair->two->path into the
2363          * index and the working copy.  We need to remove it so that
2364          * we can instead place it at new_path.  It is guaranteed to
2365          * not be untracked (unpack_trees() would have errored out
2366          * saying the file would have been overwritten), but it might
2367          * be dirty, though.
2368          */
2369         update_wd = !was_dirty(o, pair->two->path);
2370         if (!update_wd)
2371                 output(o, 1, _("Refusing to lose dirty file at %s"),
2372                        pair->two->path);
2373         remove_file(o, 1, pair->two->path, !update_wd);
2374
2375         /* Find or create a new re->dst_entry */
2376         item = string_list_lookup(entries, new_path);
2377         if (item) {
2378                 /*
2379                  * Since we're renaming on this side of history, and it's
2380                  * due to a directory rename on the other side of history
2381                  * (which we only allow when the directory in question no
2382                  * longer exists on the other side of history), the
2383                  * original entry for re->dst_entry is no longer
2384                  * necessary...
2385                  */
2386                 re->dst_entry->processed = 1;
2387
2388                 /*
2389                  * ...because we'll be using this new one.
2390                  */
2391                 re->dst_entry = item->util;
2392         } else {
2393                 /*
2394                  * re->dst_entry is for the before-dir-rename path, and we
2395                  * need it to hold information for the after-dir-rename
2396                  * path.  Before creating a new entry, we need to mark the
2397                  * old one as unnecessary (...unless it is shared by
2398                  * src_entry, i.e. this didn't use to be a rename, in which
2399                  * case we can just allow the normal processing to happen
2400                  * for it).
2401                  */
2402                 if (pair->status == 'R')
2403                         re->dst_entry->processed = 1;
2404
2405                 re->dst_entry = insert_stage_data(new_path,
2406                                                   o_tree, a_tree, b_tree,
2407                                                   entries);
2408                 item = string_list_insert(entries, new_path);
2409                 item->util = re->dst_entry;
2410         }
2411
2412         /*
2413          * Update the stage_data with the information about the path we are
2414          * moving into place.  That slot will be empty and available for us
2415          * to write to because of the collision checks in
2416          * handle_path_level_conflicts().  In other words,
2417          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2418          * open for us to write to.
2419          *
2420          * It may be tempting to actually update the index at this point as
2421          * well, using update_stages_for_stage_data(), but as per the big
2422          * "NOTE" in update_stages(), doing so will modify the current
2423          * in-memory index which will break calls to would_lose_untracked()
2424          * that we need to make.  Instead, we need to just make sure that
2425          * the various conflict_rename_*() functions update the index
2426          * explicitly rather than relying on unpack_trees() to have done it.
2427          */
2428         get_tree_entry(&tree->object.oid,
2429                        pair->two->path,
2430                        &re->dst_entry->stages[stage].oid,
2431                        &re->dst_entry->stages[stage].mode);
2432
2433         /* Update pair status */
2434         if (pair->status == 'A') {
2435                 /*
2436                  * Recording rename information for this add makes it look
2437                  * like a rename/delete conflict.  Make sure we can
2438                  * correctly handle this as an add that was moved to a new
2439                  * directory instead of reporting a rename/delete conflict.
2440                  */
2441                 re->add_turned_into_rename = 1;
2442         }
2443         /*
2444          * We don't actually look at pair->status again, but it seems
2445          * pedagogically correct to adjust it.
2446          */
2447         pair->status = 'R';
2448
2449         /*
2450          * Finally, record the new location.
2451          */
2452         pair->two->path = new_path;
2453 }
2454
2455 /*
2456  * Get information of all renames which occurred in 'pairs', making use of
2457  * any implicit directory renames inferred from the other side of history.
2458  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2459  * to be able to associate the correct cache entries with the rename
2460  * information; tree is always equal to either a_tree or b_tree.
2461  */
2462 static struct string_list *get_renames(struct merge_options *o,
2463                                        struct diff_queue_struct *pairs,
2464                                        struct hashmap *dir_renames,
2465                                        struct hashmap *dir_rename_exclusions,
2466                                        struct tree *tree,
2467                                        struct tree *o_tree,
2468                                        struct tree *a_tree,
2469                                        struct tree *b_tree,
2470                                        struct string_list *entries,
2471                                        int *clean_merge)
2472 {
2473         int i;
2474         struct hashmap collisions;
2475         struct hashmap_iter iter;
2476         struct collision_entry *e;
2477         struct string_list *renames;
2478
2479         compute_collisions(&collisions, dir_renames, pairs);
2480         renames = xcalloc(1, sizeof(struct string_list));
2481
2482         for (i = 0; i < pairs->nr; ++i) {
2483                 struct string_list_item *item;
2484                 struct rename *re;
2485                 struct diff_filepair *pair = pairs->queue[i];
2486                 char *new_path; /* non-NULL only with directory renames */
2487
2488                 if (pair->status != 'A' && pair->status != 'R') {
2489                         diff_free_filepair(pair);
2490                         continue;
2491                 }
2492                 new_path = check_for_directory_rename(o, pair->two->path, tree,
2493                                                       dir_renames,
2494                                                       dir_rename_exclusions,
2495                                                       &collisions,
2496                                                       clean_merge);
2497                 if (pair->status != 'R' && !new_path) {
2498                         diff_free_filepair(pair);
2499                         continue;
2500                 }
2501
2502                 re = xmalloc(sizeof(*re));
2503                 re->processed = 0;
2504                 re->add_turned_into_rename = 0;
2505                 re->pair = pair;
2506                 item = string_list_lookup(entries, re->pair->one->path);
2507                 if (!item)
2508                         re->src_entry = insert_stage_data(re->pair->one->path,
2509                                         o_tree, a_tree, b_tree, entries);
2510                 else
2511                         re->src_entry = item->util;
2512
2513                 item = string_list_lookup(entries, re->pair->two->path);
2514                 if (!item)
2515                         re->dst_entry = insert_stage_data(re->pair->two->path,
2516                                         o_tree, a_tree, b_tree, entries);
2517                 else
2518                         re->dst_entry = item->util;
2519                 item = string_list_insert(renames, pair->one->path);
2520                 item->util = re;
2521                 if (new_path)
2522                         apply_directory_rename_modifications(o, pair, new_path,
2523                                                              re, tree, o_tree,
2524                                                              a_tree, b_tree,
2525                                                              entries,
2526                                                              clean_merge);
2527         }
2528
2529         hashmap_iter_init(&collisions, &iter);
2530         while ((e = hashmap_iter_next(&iter))) {
2531                 free(e->target_file);
2532                 string_list_clear(&e->source_files, 0);
2533         }
2534         hashmap_free(&collisions, 1);
2535         return renames;
2536 }
2537
2538 static int process_renames(struct merge_options *o,
2539                            struct string_list *a_renames,
2540                            struct string_list *b_renames)
2541 {
2542         int clean_merge = 1, i, j;
2543         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2544         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2545         const struct rename *sre;
2546
2547         for (i = 0; i < a_renames->nr; i++) {
2548                 sre = a_renames->items[i].util;
2549                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2550                         = (void *)sre;
2551         }
2552         for (i = 0; i < b_renames->nr; i++) {
2553                 sre = b_renames->items[i].util;
2554                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2555                         = (void *)sre;
2556         }
2557
2558         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2559                 struct string_list *renames1, *renames2Dst;
2560                 struct rename *ren1 = NULL, *ren2 = NULL;
2561                 const char *branch1, *branch2;
2562                 const char *ren1_src, *ren1_dst;
2563                 struct string_list_item *lookup;
2564
2565                 if (i >= a_renames->nr) {
2566                         ren2 = b_renames->items[j++].util;
2567                 } else if (j >= b_renames->nr) {
2568                         ren1 = a_renames->items[i++].util;
2569                 } else {
2570                         int compare = strcmp(a_renames->items[i].string,
2571                                              b_renames->items[j].string);
2572                         if (compare <= 0)
2573                                 ren1 = a_renames->items[i++].util;
2574                         if (compare >= 0)
2575                                 ren2 = b_renames->items[j++].util;
2576                 }
2577
2578                 /* TODO: refactor, so that 1/2 are not needed */
2579                 if (ren1) {
2580                         renames1 = a_renames;
2581                         renames2Dst = &b_by_dst;
2582                         branch1 = o->branch1;
2583                         branch2 = o->branch2;
2584                 } else {
2585                         renames1 = b_renames;
2586                         renames2Dst = &a_by_dst;
2587                         branch1 = o->branch2;
2588                         branch2 = o->branch1;
2589                         SWAP(ren2, ren1);
2590                 }
2591
2592                 if (ren1->processed)
2593                         continue;
2594                 ren1->processed = 1;
2595                 ren1->dst_entry->processed = 1;
2596                 /* BUG: We should only mark src_entry as processed if we
2597                  * are not dealing with a rename + add-source case.
2598                  */
2599                 ren1->src_entry->processed = 1;
2600
2601                 ren1_src = ren1->pair->one->path;
2602                 ren1_dst = ren1->pair->two->path;
2603
2604                 if (ren2) {
2605                         /* One file renamed on both sides */
2606                         const char *ren2_src = ren2->pair->one->path;
2607                         const char *ren2_dst = ren2->pair->two->path;
2608                         enum rename_type rename_type;
2609                         if (strcmp(ren1_src, ren2_src) != 0)
2610                                 BUG("ren1_src != ren2_src");
2611                         ren2->dst_entry->processed = 1;
2612                         ren2->processed = 1;
2613                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2614                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2615                                 clean_merge = 0;
2616                         } else {
2617                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2618                                 /* BUG: We should only remove ren1_src in
2619                                  * the base stage (think of rename +
2620                                  * add-source cases).
2621                                  */
2622                                 remove_file(o, 1, ren1_src, 1);
2623                                 update_entry(ren1->dst_entry,
2624                                              ren1->pair->one,
2625                                              ren1->pair->two,
2626                                              ren2->pair->two);
2627                         }
2628                         setup_rename_conflict_info(rename_type,
2629                                                    ren1->pair,
2630                                                    ren2->pair,
2631                                                    branch1,
2632                                                    branch2,
2633                                                    ren1->dst_entry,
2634                                                    ren2->dst_entry,
2635                                                    o,
2636                                                    NULL,
2637                                                    NULL);
2638                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2639                         /* Two different files renamed to the same thing */
2640                         char *ren2_dst;
2641                         ren2 = lookup->util;
2642                         ren2_dst = ren2->pair->two->path;
2643                         if (strcmp(ren1_dst, ren2_dst) != 0)
2644                                 BUG("ren1_dst != ren2_dst");
2645
2646                         clean_merge = 0;
2647                         ren2->processed = 1;
2648                         /*
2649                          * BUG: We should only mark src_entry as processed
2650                          * if we are not dealing with a rename + add-source
2651                          * case.
2652                          */
2653                         ren2->src_entry->processed = 1;
2654
2655                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2656                                                    ren1->pair,
2657                                                    ren2->pair,
2658                                                    branch1,
2659                                                    branch2,
2660                                                    ren1->dst_entry,
2661                                                    ren2->dst_entry,
2662                                                    o,
2663                                                    ren1->src_entry,
2664                                                    ren2->src_entry);
2665
2666                 } else {
2667                         /* Renamed in 1, maybe changed in 2 */
2668                         /* we only use sha1 and mode of these */
2669                         struct diff_filespec src_other, dst_other;
2670                         int try_merge;
2671
2672                         /*
2673                          * unpack_trees loads entries from common-commit
2674                          * into stage 1, from head-commit into stage 2, and
2675                          * from merge-commit into stage 3.  We keep track
2676                          * of which side corresponds to the rename.
2677                          */
2678                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2679                         int other_stage =   a_renames == renames1 ? 3 : 2;
2680
2681                         /* BUG: We should only remove ren1_src in the base
2682                          * stage and in other_stage (think of rename +
2683                          * add-source case).
2684                          */
2685                         remove_file(o, 1, ren1_src,
2686                                     renamed_stage == 2 || !was_tracked(o, ren1_src));
2687
2688                         oidcpy(&src_other.oid,
2689                                &ren1->src_entry->stages[other_stage].oid);
2690                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2691                         oidcpy(&dst_other.oid,
2692                                &ren1->dst_entry->stages[other_stage].oid);
2693                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2694                         try_merge = 0;
2695
2696                         if (oid_eq(&src_other.oid, &null_oid) &&
2697                             ren1->add_turned_into_rename) {
2698                                 setup_rename_conflict_info(RENAME_DIR,
2699                                                            ren1->pair,
2700                                                            NULL,
2701                                                            branch1,
2702                                                            branch2,
2703                                                            ren1->dst_entry,
2704                                                            NULL,
2705                                                            o,
2706                                                            NULL,
2707                                                            NULL);
2708                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2709                                 setup_rename_conflict_info(RENAME_DELETE,
2710                                                            ren1->pair,
2711                                                            NULL,
2712                                                            branch1,
2713                                                            branch2,
2714                                                            ren1->dst_entry,
2715                                                            NULL,
2716                                                            o,
2717                                                            NULL,
2718                                                            NULL);
2719                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2720                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2721                                 /*
2722                                  * Added file on the other side identical to
2723                                  * the file being renamed: clean merge.
2724                                  * Also, there is no need to overwrite the
2725                                  * file already in the working copy, so call
2726                                  * update_file_flags() instead of
2727                                  * update_file().
2728                                  */
2729                                 if (update_file_flags(o,
2730                                                       &ren1->pair->two->oid,
2731                                                       ren1->pair->two->mode,
2732                                                       ren1_dst,
2733                                                       1, /* update_cache */
2734                                                       0  /* update_wd    */))
2735                                         clean_merge = -1;
2736                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2737                                 clean_merge = 0;
2738                                 try_merge = 1;
2739                                 output(o, 1, _("CONFLICT (rename/add): Rename %s->%s in %s. "
2740                                        "%s added in %s"),
2741                                        ren1_src, ren1_dst, branch1,
2742                                        ren1_dst, branch2);
2743                                 if (o->call_depth) {
2744                                         struct merge_file_info mfi;
2745                                         if (merge_file_one(o, ren1_dst, &null_oid, 0,
2746                                                            &ren1->pair->two->oid,
2747                                                            ren1->pair->two->mode,
2748                                                            &dst_other.oid,
2749                                                            dst_other.mode,
2750                                                            branch1, branch2, &mfi)) {
2751                                                 clean_merge = -1;
2752                                                 goto cleanup_and_return;
2753                                         }
2754                                         output(o, 1, _("Adding merged %s"), ren1_dst);
2755                                         if (update_file(o, 0, &mfi.oid,
2756                                                         mfi.mode, ren1_dst))
2757                                                 clean_merge = -1;
2758                                         try_merge = 0;
2759                                 } else {
2760                                         char *new_path = unique_path(o, ren1_dst, branch2);
2761                                         output(o, 1, _("Adding as %s instead"), new_path);
2762                                         if (update_file(o, 0, &dst_other.oid,
2763                                                         dst_other.mode, new_path))
2764                                                 clean_merge = -1;
2765                                         free(new_path);
2766                                 }
2767                         } else
2768                                 try_merge = 1;
2769
2770                         if (clean_merge < 0)
2771                                 goto cleanup_and_return;
2772                         if (try_merge) {
2773                                 struct diff_filespec *one, *a, *b;
2774                                 src_other.path = (char *)ren1_src;
2775
2776                                 one = ren1->pair->one;
2777                                 if (a_renames == renames1) {
2778                                         a = ren1->pair->two;
2779                                         b = &src_other;
2780                                 } else {
2781                                         b = ren1->pair->two;
2782                                         a = &src_other;
2783                                 }
2784                                 update_entry(ren1->dst_entry, one, a, b);
2785                                 setup_rename_conflict_info(RENAME_NORMAL,
2786                                                            ren1->pair,
2787                                                            NULL,
2788                                                            branch1,
2789                                                            NULL,
2790                                                            ren1->dst_entry,
2791                                                            NULL,
2792                                                            o,
2793                                                            NULL,
2794                                                            NULL);
2795                         }
2796                 }
2797         }
2798 cleanup_and_return:
2799         string_list_clear(&a_by_dst, 0);
2800         string_list_clear(&b_by_dst, 0);
2801
2802         return clean_merge;
2803 }
2804
2805 struct rename_info {
2806         struct string_list *head_renames;
2807         struct string_list *merge_renames;
2808 };
2809
2810 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2811                                    struct hashmap *dir_renames)
2812 {
2813         struct hashmap_iter iter;
2814         struct dir_rename_entry *e;
2815
2816         hashmap_iter_init(dir_renames, &iter);
2817         while ((e = hashmap_iter_next(&iter))) {
2818                 free(e->dir);
2819                 strbuf_release(&e->new_dir);
2820                 /* possible_new_dirs already cleared in get_directory_renames */
2821         }
2822         hashmap_free(dir_renames, 1);
2823         free(dir_renames);
2824
2825         free(pairs->queue);
2826         free(pairs);
2827 }
2828
2829 static int handle_renames(struct merge_options *o,
2830                           struct tree *common,
2831                           struct tree *head,
2832                           struct tree *merge,
2833                           struct string_list *entries,
2834                           struct rename_info *ri)
2835 {
2836         struct diff_queue_struct *head_pairs, *merge_pairs;
2837         struct hashmap *dir_re_head, *dir_re_merge;
2838         int clean = 1;
2839
2840         ri->head_renames = NULL;
2841         ri->merge_renames = NULL;
2842
2843         if (!merge_detect_rename(o))
2844                 return 1;
2845
2846         head_pairs = get_diffpairs(o, common, head);
2847         merge_pairs = get_diffpairs(o, common, merge);
2848
2849         dir_re_head = get_directory_renames(head_pairs, head);
2850         dir_re_merge = get_directory_renames(merge_pairs, merge);
2851
2852         handle_directory_level_conflicts(o,
2853                                          dir_re_head, head,
2854                                          dir_re_merge, merge);
2855
2856         ri->head_renames  = get_renames(o, head_pairs,
2857                                         dir_re_merge, dir_re_head, head,
2858                                         common, head, merge, entries,
2859                                         &clean);
2860         if (clean < 0)
2861                 goto cleanup;
2862         ri->merge_renames = get_renames(o, merge_pairs,
2863                                         dir_re_head, dir_re_merge, merge,
2864                                         common, head, merge, entries,
2865                                         &clean);
2866         if (clean < 0)
2867                 goto cleanup;
2868         clean &= process_renames(o, ri->head_renames, ri->merge_renames);
2869
2870 cleanup:
2871         /*
2872          * Some cleanup is deferred until cleanup_renames() because the
2873          * data structures are still needed and referenced in
2874          * process_entry().  But there are a few things we can free now.
2875          */
2876         initial_cleanup_rename(head_pairs, dir_re_head);
2877         initial_cleanup_rename(merge_pairs, dir_re_merge);
2878
2879         return clean;
2880 }
2881
2882 static void final_cleanup_rename(struct string_list *rename)
2883 {
2884         const struct rename *re;
2885         int i;
2886
2887         if (rename == NULL)
2888                 return;
2889
2890         for (i = 0; i < rename->nr; i++) {
2891                 re = rename->items[i].util;
2892                 diff_free_filepair(re->pair);
2893         }
2894         string_list_clear(rename, 1);
2895         free(rename);
2896 }
2897
2898 static void final_cleanup_renames(struct rename_info *re_info)
2899 {
2900         final_cleanup_rename(re_info->head_renames);
2901         final_cleanup_rename(re_info->merge_renames);
2902 }
2903
2904 static struct object_id *stage_oid(const struct object_id *oid, unsigned mode)
2905 {
2906         return (is_null_oid(oid) || mode == 0) ? NULL: (struct object_id *)oid;
2907 }
2908
2909 static int read_oid_strbuf(struct merge_options *o,
2910         const struct object_id *oid, struct strbuf *dst)
2911 {
2912         void *buf;
2913         enum object_type type;
2914         unsigned long size;
2915         buf = read_object_file(oid, &type, &size);
2916         if (!buf)
2917                 return err(o, _("cannot read object %s"), oid_to_hex(oid));
2918         if (type != OBJ_BLOB) {
2919                 free(buf);
2920                 return err(o, _("object %s is not a blob"), oid_to_hex(oid));
2921         }
2922         strbuf_attach(dst, buf, size, size + 1);
2923         return 0;
2924 }
2925
2926 static int blob_unchanged(struct merge_options *opt,
2927                           const struct object_id *o_oid,
2928                           unsigned o_mode,
2929                           const struct object_id *a_oid,
2930                           unsigned a_mode,
2931                           int renormalize, const char *path)
2932 {
2933         struct strbuf o = STRBUF_INIT;
2934         struct strbuf a = STRBUF_INIT;
2935         int ret = 0; /* assume changed for safety */
2936
2937         if (a_mode != o_mode)
2938                 return 0;
2939         if (oid_eq(o_oid, a_oid))
2940                 return 1;
2941         if (!renormalize)
2942                 return 0;
2943
2944         assert(o_oid && a_oid);
2945         if (read_oid_strbuf(opt, o_oid, &o) || read_oid_strbuf(opt, a_oid, &a))
2946                 goto error_return;
2947         /*
2948          * Note: binary | is used so that both renormalizations are
2949          * performed.  Comparison can be skipped if both files are
2950          * unchanged since their sha1s have already been compared.
2951          */
2952         if (renormalize_buffer(&the_index, path, o.buf, o.len, &o) |
2953             renormalize_buffer(&the_index, path, a.buf, a.len, &a))
2954                 ret = (o.len == a.len && !memcmp(o.buf, a.buf, o.len));
2955
2956 error_return:
2957         strbuf_release(&o);
2958         strbuf_release(&a);
2959         return ret;
2960 }
2961
2962 static int handle_modify_delete(struct merge_options *o,
2963                                  const char *path,
2964                                  struct object_id *o_oid, int o_mode,
2965                                  struct object_id *a_oid, int a_mode,
2966                                  struct object_id *b_oid, int b_mode)
2967 {
2968         const char *modify_branch, *delete_branch;
2969         struct object_id *changed_oid;
2970         int changed_mode;
2971
2972         if (a_oid) {
2973                 modify_branch = o->branch1;
2974                 delete_branch = o->branch2;
2975                 changed_oid = a_oid;
2976                 changed_mode = a_mode;
2977         } else {
2978                 modify_branch = o->branch2;
2979                 delete_branch = o->branch1;
2980                 changed_oid = b_oid;
2981                 changed_mode = b_mode;
2982         }
2983
2984         return handle_change_delete(o,
2985                                     path, NULL,
2986                                     o_oid, o_mode,
2987                                     changed_oid, changed_mode,
2988                                     modify_branch, delete_branch,
2989                                     _("modify"), _("modified"));
2990 }
2991
2992 static int merge_content(struct merge_options *o,
2993                          const char *path,
2994                          int is_dirty,
2995                          struct object_id *o_oid, int o_mode,
2996                          struct object_id *a_oid, int a_mode,
2997                          struct object_id *b_oid, int b_mode,
2998                          struct rename_conflict_info *rename_conflict_info)
2999 {
3000         const char *reason = _("content");
3001         const char *path1 = NULL, *path2 = NULL;
3002         struct merge_file_info mfi;
3003         struct diff_filespec one, a, b;
3004         unsigned df_conflict_remains = 0;
3005
3006         if (!o_oid) {
3007                 reason = _("add/add");
3008                 o_oid = (struct object_id *)&null_oid;
3009         }
3010         one.path = a.path = b.path = (char *)path;
3011         oidcpy(&one.oid, o_oid);
3012         one.mode = o_mode;
3013         oidcpy(&a.oid, a_oid);
3014         a.mode = a_mode;
3015         oidcpy(&b.oid, b_oid);
3016         b.mode = b_mode;
3017
3018         if (rename_conflict_info) {
3019                 struct diff_filepair *pair1 = rename_conflict_info->pair1;
3020
3021                 path1 = (o->branch1 == rename_conflict_info->branch1) ?
3022                         pair1->two->path : pair1->one->path;
3023                 /* If rename_conflict_info->pair2 != NULL, we are in
3024                  * RENAME_ONE_FILE_TO_ONE case.  Otherwise, we have a
3025                  * normal rename.
3026                  */
3027                 path2 = (rename_conflict_info->pair2 ||
3028                          o->branch2 == rename_conflict_info->branch1) ?
3029                         pair1->two->path : pair1->one->path;
3030
3031                 if (dir_in_way(path, !o->call_depth,
3032                                S_ISGITLINK(pair1->two->mode)))
3033                         df_conflict_remains = 1;
3034         }
3035         if (merge_file_special_markers(o, &one, &a, &b, path,
3036                                        o->branch1, path1,
3037                                        o->branch2, path2, &mfi))
3038                 return -1;
3039
3040         /*
3041          * We can skip updating the working tree file iff:
3042          *   a) The merge is clean
3043          *   b) The merge matches what was in HEAD (content, mode, pathname)
3044          *   c) The target path is usable (i.e. not involved in D/F conflict)
3045          */
3046         if (mfi.clean &&
3047             was_tracked_and_matches(o, path, &mfi.oid, mfi.mode) &&
3048             !df_conflict_remains) {
3049                 output(o, 3, _("Skipped %s (merged same as existing)"), path);
3050                 if (add_cacheinfo(o, mfi.mode, &mfi.oid, path,
3051                                   0, (!o->call_depth && !is_dirty), 0))
3052                         return -1;
3053                 return mfi.clean;
3054         }
3055
3056         if (!mfi.clean) {
3057                 if (S_ISGITLINK(mfi.mode))
3058                         reason = _("submodule");
3059                 output(o, 1, _("CONFLICT (%s): Merge conflict in %s"),
3060                                 reason, path);
3061                 if (rename_conflict_info && !df_conflict_remains)
3062                         if (update_stages(o, path, &one, &a, &b))
3063                                 return -1;
3064         }
3065
3066         if (df_conflict_remains || is_dirty) {
3067                 char *new_path;
3068                 if (o->call_depth) {
3069                         remove_file_from_cache(path);
3070                 } else {
3071                         if (!mfi.clean) {
3072                                 if (update_stages(o, path, &one, &a, &b))
3073                                         return -1;
3074                         } else {
3075                                 int file_from_stage2 = was_tracked(o, path);
3076                                 struct diff_filespec merged;
3077                                 oidcpy(&merged.oid, &mfi.oid);
3078                                 merged.mode = mfi.mode;
3079
3080                                 if (update_stages(o, path, NULL,
3081                                                   file_from_stage2 ? &merged : NULL,
3082                                                   file_from_stage2 ? NULL : &merged))
3083                                         return -1;
3084                         }
3085
3086                 }
3087                 new_path = unique_path(o, path, rename_conflict_info->branch1);
3088                 if (is_dirty) {
3089                         output(o, 1, _("Refusing to lose dirty file at %s"),
3090                                path);
3091                 }
3092                 output(o, 1, _("Adding as %s instead"), new_path);
3093                 if (update_file(o, 0, &mfi.oid, mfi.mode, new_path)) {
3094                         free(new_path);
3095                         return -1;
3096                 }
3097                 free(new_path);
3098                 mfi.clean = 0;
3099         } else if (update_file(o, mfi.clean, &mfi.oid, mfi.mode, path))
3100                 return -1;
3101         return !is_dirty && mfi.clean;
3102 }
3103
3104 static int conflict_rename_normal(struct merge_options *o,
3105                                   const char *path,
3106                                   struct object_id *o_oid, unsigned int o_mode,
3107                                   struct object_id *a_oid, unsigned int a_mode,
3108                                   struct object_id *b_oid, unsigned int b_mode,
3109                                   struct rename_conflict_info *ci)
3110 {
3111         /* Merge the content and write it out */
3112         return merge_content(o, path, was_dirty(o, path),
3113                              o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3114                              ci);
3115 }
3116
3117 /* Per entry merge function */
3118 static int process_entry(struct merge_options *o,
3119                          const char *path, struct stage_data *entry)
3120 {
3121         int clean_merge = 1;
3122         int normalize = o->renormalize;
3123         unsigned o_mode = entry->stages[1].mode;
3124         unsigned a_mode = entry->stages[2].mode;
3125         unsigned b_mode = entry->stages[3].mode;
3126         struct object_id *o_oid = stage_oid(&entry->stages[1].oid, o_mode);
3127         struct object_id *a_oid = stage_oid(&entry->stages[2].oid, a_mode);
3128         struct object_id *b_oid = stage_oid(&entry->stages[3].oid, b_mode);
3129
3130         entry->processed = 1;
3131         if (entry->rename_conflict_info) {
3132                 struct rename_conflict_info *conflict_info = entry->rename_conflict_info;
3133                 switch (conflict_info->rename_type) {
3134                 case RENAME_NORMAL:
3135                 case RENAME_ONE_FILE_TO_ONE:
3136                         clean_merge = conflict_rename_normal(o,
3137                                                              path,
3138                                                              o_oid, o_mode,
3139                                                              a_oid, a_mode,
3140                                                              b_oid, b_mode,
3141                                                              conflict_info);
3142                         break;
3143                 case RENAME_DIR:
3144                         clean_merge = 1;
3145                         if (conflict_rename_dir(o,
3146                                                 conflict_info->pair1,
3147                                                 conflict_info->branch1,
3148                                                 conflict_info->branch2))
3149                                 clean_merge = -1;
3150                         break;
3151                 case RENAME_DELETE:
3152                         clean_merge = 0;
3153                         if (conflict_rename_delete(o,
3154                                                    conflict_info->pair1,
3155                                                    conflict_info->branch1,
3156                                                    conflict_info->branch2))
3157                                 clean_merge = -1;
3158                         break;
3159                 case RENAME_ONE_FILE_TO_TWO:
3160                         clean_merge = 0;
3161                         if (conflict_rename_rename_1to2(o, conflict_info))
3162                                 clean_merge = -1;
3163                         break;
3164                 case RENAME_TWO_FILES_TO_ONE:
3165                         clean_merge = 0;
3166                         if (conflict_rename_rename_2to1(o, conflict_info))
3167                                 clean_merge = -1;
3168                         break;
3169                 default:
3170                         entry->processed = 0;
3171                         break;
3172                 }
3173         } else if (o_oid && (!a_oid || !b_oid)) {
3174                 /* Case A: Deleted in one */
3175                 if ((!a_oid && !b_oid) ||
3176                     (!b_oid && blob_unchanged(o, o_oid, o_mode, a_oid, a_mode, normalize, path)) ||
3177                     (!a_oid && blob_unchanged(o, o_oid, o_mode, b_oid, b_mode, normalize, path))) {
3178                         /* Deleted in both or deleted in one and
3179                          * unchanged in the other */
3180                         if (a_oid)
3181                                 output(o, 2, _("Removing %s"), path);
3182                         /* do not touch working file if it did not exist */
3183                         remove_file(o, 1, path, !a_oid);
3184                 } else {
3185                         /* Modify/delete; deleted side may have put a directory in the way */
3186                         clean_merge = 0;
3187                         if (handle_modify_delete(o, path, o_oid, o_mode,
3188                                                  a_oid, a_mode, b_oid, b_mode))
3189                                 clean_merge = -1;
3190                 }
3191         } else if ((!o_oid && a_oid && !b_oid) ||
3192                    (!o_oid && !a_oid && b_oid)) {
3193                 /* Case B: Added in one. */
3194                 /* [nothing|directory] -> ([nothing|directory], file) */
3195
3196                 const char *add_branch;
3197                 const char *other_branch;
3198                 unsigned mode;
3199                 const struct object_id *oid;
3200                 const char *conf;
3201
3202                 if (a_oid) {
3203                         add_branch = o->branch1;
3204                         other_branch = o->branch2;
3205                         mode = a_mode;
3206                         oid = a_oid;
3207                         conf = _("file/directory");
3208                 } else {
3209                         add_branch = o->branch2;
3210                         other_branch = o->branch1;
3211                         mode = b_mode;
3212                         oid = b_oid;
3213                         conf = _("directory/file");
3214                 }
3215                 if (dir_in_way(path,
3216                                !o->call_depth && !S_ISGITLINK(a_mode),
3217                                0)) {
3218                         char *new_path = unique_path(o, path, add_branch);
3219                         clean_merge = 0;
3220                         output(o, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3221                                "Adding %s as %s"),
3222                                conf, path, other_branch, path, new_path);
3223                         if (update_file(o, 0, oid, mode, new_path))
3224                                 clean_merge = -1;
3225                         else if (o->call_depth)
3226                                 remove_file_from_cache(path);
3227                         free(new_path);
3228                 } else {
3229                         output(o, 2, _("Adding %s"), path);
3230                         /* do not overwrite file if already present */
3231                         if (update_file_flags(o, oid, mode, path, 1, !a_oid))
3232                                 clean_merge = -1;
3233                 }
3234         } else if (a_oid && b_oid) {
3235                 /* Case C: Added in both (check for same permissions) and */
3236                 /* case D: Modified in both, but differently. */
3237                 int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3238                 clean_merge = merge_content(o, path, is_dirty,
3239                                             o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3240                                             NULL);
3241         } else if (!o_oid && !a_oid && !b_oid) {
3242                 /*
3243                  * this entry was deleted altogether. a_mode == 0 means
3244                  * we had that path and want to actively remove it.
3245                  */
3246                 remove_file(o, 1, path, !a_mode);
3247         } else
3248                 BUG("fatal merge failure, shouldn't happen.");
3249
3250         return clean_merge;
3251 }
3252
3253 int merge_trees(struct merge_options *o,
3254                 struct tree *head,
3255                 struct tree *merge,
3256                 struct tree *common,
3257                 struct tree **result)
3258 {
3259         int code, clean;
3260
3261         if (o->subtree_shift) {
3262                 merge = shift_tree_object(head, merge, o->subtree_shift);
3263                 common = shift_tree_object(head, common, o->subtree_shift);
3264         }
3265
3266         if (oid_eq(&common->object.oid, &merge->object.oid)) {
3267                 struct strbuf sb = STRBUF_INIT;
3268
3269                 if (!o->call_depth && index_has_changes(&sb)) {
3270                         err(o, _("Dirty index: cannot merge (dirty: %s)"),
3271                             sb.buf);
3272                         return 0;
3273                 }
3274                 output(o, 0, _("Already up to date!"));
3275                 *result = head;
3276                 return 1;
3277         }
3278
3279         code = unpack_trees_start(o, common, head, merge);
3280
3281         if (code != 0) {
3282                 if (show(o, 4) || o->call_depth)
3283                         err(o, _("merging of trees %s and %s failed"),
3284                             oid_to_hex(&head->object.oid),
3285                             oid_to_hex(&merge->object.oid));
3286                 unpack_trees_finish(o);
3287                 return -1;
3288         }
3289
3290         if (unmerged_cache()) {
3291                 struct string_list *entries;
3292                 struct rename_info re_info;
3293                 int i;
3294                 /*
3295                  * Only need the hashmap while processing entries, so
3296                  * initialize it here and free it when we are done running
3297                  * through the entries. Keeping it in the merge_options as
3298                  * opposed to decaring a local hashmap is for convenience
3299                  * so that we don't have to pass it to around.
3300                  */
3301                 hashmap_init(&o->current_file_dir_set, path_hashmap_cmp, NULL, 512);
3302                 get_files_dirs(o, head);
3303                 get_files_dirs(o, merge);
3304
3305                 entries = get_unmerged();
3306                 clean = handle_renames(o, common, head, merge, entries,
3307                                        &re_info);
3308                 record_df_conflict_files(o, entries);
3309                 if (clean < 0)
3310                         goto cleanup;
3311                 for (i = entries->nr-1; 0 <= i; i--) {
3312                         const char *path = entries->items[i].string;
3313                         struct stage_data *e = entries->items[i].util;
3314                         if (!e->processed) {
3315                                 int ret = process_entry(o, path, e);
3316                                 if (!ret)
3317                                         clean = 0;
3318                                 else if (ret < 0) {
3319                                         clean = ret;
3320                                         goto cleanup;
3321                                 }
3322                         }
3323                 }
3324                 for (i = 0; i < entries->nr; i++) {
3325                         struct stage_data *e = entries->items[i].util;
3326                         if (!e->processed)
3327                                 BUG("unprocessed path??? %s",
3328                                     entries->items[i].string);
3329                 }
3330
3331 cleanup:
3332                 final_cleanup_renames(&re_info);
3333
3334                 string_list_clear(entries, 1);
3335                 free(entries);
3336
3337                 hashmap_free(&o->current_file_dir_set, 1);
3338
3339                 if (clean < 0) {
3340                         unpack_trees_finish(o);
3341                         return clean;
3342                 }
3343         }
3344         else
3345                 clean = 1;
3346
3347         unpack_trees_finish(o);
3348
3349         if (o->call_depth && !(*result = write_tree_from_memory(o)))
3350                 return -1;
3351
3352         return clean;
3353 }
3354
3355 static struct commit_list *reverse_commit_list(struct commit_list *list)
3356 {
3357         struct commit_list *next = NULL, *current, *backup;
3358         for (current = list; current; current = backup) {
3359                 backup = current->next;
3360                 current->next = next;
3361                 next = current;
3362         }
3363         return next;
3364 }
3365
3366 /*
3367  * Merge the commits h1 and h2, return the resulting virtual
3368  * commit object and a flag indicating the cleanness of the merge.
3369  */
3370 int merge_recursive(struct merge_options *o,
3371                     struct commit *h1,
3372                     struct commit *h2,
3373                     struct commit_list *ca,
3374                     struct commit **result)
3375 {
3376         struct commit_list *iter;
3377         struct commit *merged_common_ancestors;
3378         struct tree *mrtree;
3379         int clean;
3380
3381         if (show(o, 4)) {
3382                 output(o, 4, _("Merging:"));
3383                 output_commit_title(o, h1);
3384                 output_commit_title(o, h2);
3385         }
3386
3387         if (!ca) {
3388                 ca = get_merge_bases(h1, h2);
3389                 ca = reverse_commit_list(ca);
3390         }
3391
3392         if (show(o, 5)) {
3393                 unsigned cnt = commit_list_count(ca);
3394
3395                 output(o, 5, Q_("found %u common ancestor:",
3396                                 "found %u common ancestors:", cnt), cnt);
3397                 for (iter = ca; iter; iter = iter->next)
3398                         output_commit_title(o, iter->item);
3399         }
3400
3401         merged_common_ancestors = pop_commit(&ca);
3402         if (merged_common_ancestors == NULL) {
3403                 /* if there is no common ancestor, use an empty tree */
3404                 struct tree *tree;
3405
3406                 tree = lookup_tree(the_hash_algo->empty_tree);
3407                 merged_common_ancestors = make_virtual_commit(tree, "ancestor");
3408         }
3409
3410         for (iter = ca; iter; iter = iter->next) {
3411                 const char *saved_b1, *saved_b2;
3412                 o->call_depth++;
3413                 /*
3414                  * When the merge fails, the result contains files
3415                  * with conflict markers. The cleanness flag is
3416                  * ignored (unless indicating an error), it was never
3417                  * actually used, as result of merge_trees has always
3418                  * overwritten it: the committed "conflicts" were
3419                  * already resolved.
3420                  */
3421                 discard_cache();
3422                 saved_b1 = o->branch1;
3423                 saved_b2 = o->branch2;
3424                 o->branch1 = "Temporary merge branch 1";
3425                 o->branch2 = "Temporary merge branch 2";
3426                 if (merge_recursive(o, merged_common_ancestors, iter->item,
3427                                     NULL, &merged_common_ancestors) < 0)
3428                         return -1;
3429                 o->branch1 = saved_b1;
3430                 o->branch2 = saved_b2;
3431                 o->call_depth--;
3432
3433                 if (!merged_common_ancestors)
3434                         return err(o, _("merge returned no commit"));
3435         }
3436
3437         discard_cache();
3438         if (!o->call_depth)
3439                 read_cache();
3440
3441         o->ancestor = "merged common ancestors";
3442         clean = merge_trees(o, get_commit_tree(h1), get_commit_tree(h2),
3443                             get_commit_tree(merged_common_ancestors),
3444                             &mrtree);
3445         if (clean < 0) {
3446                 flush_output(o);
3447                 return clean;
3448         }
3449
3450         if (o->call_depth) {
3451                 *result = make_virtual_commit(mrtree, "merged tree");
3452                 commit_list_insert(h1, &(*result)->parents);
3453                 commit_list_insert(h2, &(*result)->parents->next);
3454         }
3455         flush_output(o);
3456         if (!o->call_depth && o->buffer_output < 2)
3457                 strbuf_release(&o->obuf);
3458         if (show(o, 2))
3459                 diff_warn_rename_limit("merge.renamelimit",
3460                                        o->needed_rename_limit, 0);
3461         return clean;
3462 }
3463
3464 static struct commit *get_ref(const struct object_id *oid, const char *name)
3465 {
3466         struct object *object;
3467
3468         object = deref_tag(parse_object(oid), name, strlen(name));
3469         if (!object)
3470                 return NULL;
3471         if (object->type == OBJ_TREE)
3472                 return make_virtual_commit((struct tree*)object, name);
3473         if (object->type != OBJ_COMMIT)
3474                 return NULL;
3475         if (parse_commit((struct commit *)object))
3476                 return NULL;
3477         return (struct commit *)object;
3478 }
3479
3480 int merge_recursive_generic(struct merge_options *o,
3481                             const struct object_id *head,
3482                             const struct object_id *merge,
3483                             int num_base_list,
3484                             const struct object_id **base_list,
3485                             struct commit **result)
3486 {
3487         int clean;
3488         struct lock_file lock = LOCK_INIT;
3489         struct commit *head_commit = get_ref(head, o->branch1);
3490         struct commit *next_commit = get_ref(merge, o->branch2);
3491         struct commit_list *ca = NULL;
3492
3493         if (base_list) {
3494                 int i;
3495                 for (i = 0; i < num_base_list; ++i) {
3496                         struct commit *base;
3497                         if (!(base = get_ref(base_list[i], oid_to_hex(base_list[i]))))
3498                                 return err(o, _("Could not parse object '%s'"),
3499                                         oid_to_hex(base_list[i]));
3500                         commit_list_insert(base, &ca);
3501                 }
3502         }
3503
3504         hold_locked_index(&lock, LOCK_DIE_ON_ERROR);
3505         clean = merge_recursive(o, head_commit, next_commit, ca,
3506                         result);
3507         if (clean < 0) {
3508                 rollback_lock_file(&lock);
3509                 return clean;
3510         }
3511
3512         if (write_locked_index(&the_index, &lock,
3513                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3514                 return err(o, _("Unable to write index."));
3515
3516         return clean ? 0 : 1;
3517 }
3518
3519 static void merge_recursive_config(struct merge_options *o)
3520 {
3521         char *value = NULL;
3522         git_config_get_int("merge.verbosity", &o->verbosity);
3523         git_config_get_int("diff.renamelimit", &o->diff_rename_limit);
3524         git_config_get_int("merge.renamelimit", &o->merge_rename_limit);
3525         if (!git_config_get_string("diff.renames", &value)) {
3526                 o->diff_detect_rename = git_config_rename("diff.renames", value);
3527                 free(value);
3528         }
3529         if (!git_config_get_string("merge.renames", &value)) {
3530                 o->merge_detect_rename = git_config_rename("merge.renames", value);
3531                 free(value);
3532         }
3533         git_config(git_xmerge_config, NULL);
3534 }
3535
3536 void init_merge_options(struct merge_options *o)
3537 {
3538         const char *merge_verbosity;
3539         memset(o, 0, sizeof(struct merge_options));
3540         o->verbosity = 2;
3541         o->buffer_output = 1;
3542         o->diff_rename_limit = -1;
3543         o->merge_rename_limit = -1;
3544         o->renormalize = 0;
3545         o->diff_detect_rename = -1;
3546         o->merge_detect_rename = -1;
3547         merge_recursive_config(o);
3548         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3549         if (merge_verbosity)
3550                 o->verbosity = strtol(merge_verbosity, NULL, 10);
3551         if (o->verbosity >= 5)
3552                 o->buffer_output = 0;
3553         strbuf_init(&o->obuf, 0);
3554         string_list_init(&o->df_conflict_file_set, 1);
3555 }
3556
3557 int parse_merge_opt(struct merge_options *o, const char *s)
3558 {
3559         const char *arg;
3560
3561         if (!s || !*s)
3562                 return -1;
3563         if (!strcmp(s, "ours"))
3564                 o->recursive_variant = MERGE_RECURSIVE_OURS;
3565         else if (!strcmp(s, "theirs"))
3566                 o->recursive_variant = MERGE_RECURSIVE_THEIRS;
3567         else if (!strcmp(s, "subtree"))
3568                 o->subtree_shift = "";
3569         else if (skip_prefix(s, "subtree=", &arg))
3570                 o->subtree_shift = arg;
3571         else if (!strcmp(s, "patience"))
3572                 o->xdl_opts = DIFF_WITH_ALG(o, PATIENCE_DIFF);
3573         else if (!strcmp(s, "histogram"))
3574                 o->xdl_opts = DIFF_WITH_ALG(o, HISTOGRAM_DIFF);
3575         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3576                 long value = parse_algorithm_value(arg);
3577                 if (value < 0)
3578                         return -1;
3579                 /* clear out previous settings */
3580                 DIFF_XDL_CLR(o, NEED_MINIMAL);
3581                 o->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3582                 o->xdl_opts |= value;
3583         }
3584         else if (!strcmp(s, "ignore-space-change"))
3585                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_CHANGE);
3586         else if (!strcmp(s, "ignore-all-space"))
3587                 DIFF_XDL_SET(o, IGNORE_WHITESPACE);
3588         else if (!strcmp(s, "ignore-space-at-eol"))
3589                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_AT_EOL);
3590         else if (!strcmp(s, "ignore-cr-at-eol"))
3591                 DIFF_XDL_SET(o, IGNORE_CR_AT_EOL);
3592         else if (!strcmp(s, "renormalize"))
3593                 o->renormalize = 1;
3594         else if (!strcmp(s, "no-renormalize"))
3595                 o->renormalize = 0;
3596         else if (!strcmp(s, "no-renames"))
3597                 o->merge_detect_rename = 0;
3598         else if (!strcmp(s, "find-renames")) {
3599                 o->merge_detect_rename = 1;
3600                 o->rename_score = 0;
3601         }
3602         else if (skip_prefix(s, "find-renames=", &arg) ||
3603                  skip_prefix(s, "rename-threshold=", &arg)) {
3604                 if ((o->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3605                         return -1;
3606                 o->merge_detect_rename = 1;
3607         }
3608         else
3609                 return -1;
3610         return 0;
3611 }