tag: add repository argument to deref_tag
[git] / merge-recursive.c
1 /*
2  * Recursive Merge algorithm stolen from git-merge-recursive.py by
3  * Fredrik Kuivinen.
4  * The thieves were Alex Riesen and Johannes Schindelin, in June/July 2006
5  */
6 #include "cache.h"
7 #include "config.h"
8 #include "advice.h"
9 #include "lockfile.h"
10 #include "cache-tree.h"
11 #include "object-store.h"
12 #include "repository.h"
13 #include "commit.h"
14 #include "blob.h"
15 #include "builtin.h"
16 #include "tree-walk.h"
17 #include "diff.h"
18 #include "diffcore.h"
19 #include "tag.h"
20 #include "alloc.h"
21 #include "unpack-trees.h"
22 #include "string-list.h"
23 #include "xdiff-interface.h"
24 #include "ll-merge.h"
25 #include "attr.h"
26 #include "merge-recursive.h"
27 #include "dir.h"
28 #include "submodule.h"
29 #include "revision.h"
30
31 struct path_hashmap_entry {
32         struct hashmap_entry e;
33         char path[FLEX_ARRAY];
34 };
35
36 static int path_hashmap_cmp(const void *cmp_data,
37                             const void *entry,
38                             const void *entry_or_key,
39                             const void *keydata)
40 {
41         const struct path_hashmap_entry *a = entry;
42         const struct path_hashmap_entry *b = entry_or_key;
43         const char *key = keydata;
44
45         if (ignore_case)
46                 return strcasecmp(a->path, key ? key : b->path);
47         else
48                 return strcmp(a->path, key ? key : b->path);
49 }
50
51 static unsigned int path_hash(const char *path)
52 {
53         return ignore_case ? strihash(path) : strhash(path);
54 }
55
56 static struct dir_rename_entry *dir_rename_find_entry(struct hashmap *hashmap,
57                                                       char *dir)
58 {
59         struct dir_rename_entry key;
60
61         if (dir == NULL)
62                 return NULL;
63         hashmap_entry_init(&key, strhash(dir));
64         key.dir = dir;
65         return hashmap_get(hashmap, &key, NULL);
66 }
67
68 static int dir_rename_cmp(const void *unused_cmp_data,
69                           const void *entry,
70                           const void *entry_or_key,
71                           const void *unused_keydata)
72 {
73         const struct dir_rename_entry *e1 = entry;
74         const struct dir_rename_entry *e2 = entry_or_key;
75
76         return strcmp(e1->dir, e2->dir);
77 }
78
79 static void dir_rename_init(struct hashmap *map)
80 {
81         hashmap_init(map, dir_rename_cmp, NULL, 0);
82 }
83
84 static void dir_rename_entry_init(struct dir_rename_entry *entry,
85                                   char *directory)
86 {
87         hashmap_entry_init(entry, strhash(directory));
88         entry->dir = directory;
89         entry->non_unique_new_dir = 0;
90         strbuf_init(&entry->new_dir, 0);
91         string_list_init(&entry->possible_new_dirs, 0);
92 }
93
94 static struct collision_entry *collision_find_entry(struct hashmap *hashmap,
95                                                     char *target_file)
96 {
97         struct collision_entry key;
98
99         hashmap_entry_init(&key, strhash(target_file));
100         key.target_file = target_file;
101         return hashmap_get(hashmap, &key, NULL);
102 }
103
104 static int collision_cmp(void *unused_cmp_data,
105                          const struct collision_entry *e1,
106                          const struct collision_entry *e2,
107                          const void *unused_keydata)
108 {
109         return strcmp(e1->target_file, e2->target_file);
110 }
111
112 static void collision_init(struct hashmap *map)
113 {
114         hashmap_init(map, (hashmap_cmp_fn) collision_cmp, NULL, 0);
115 }
116
117 static void flush_output(struct merge_options *o)
118 {
119         if (o->buffer_output < 2 && o->obuf.len) {
120                 fputs(o->obuf.buf, stdout);
121                 strbuf_reset(&o->obuf);
122         }
123 }
124
125 static int err(struct merge_options *o, const char *err, ...)
126 {
127         va_list params;
128
129         if (o->buffer_output < 2)
130                 flush_output(o);
131         else {
132                 strbuf_complete(&o->obuf, '\n');
133                 strbuf_addstr(&o->obuf, "error: ");
134         }
135         va_start(params, err);
136         strbuf_vaddf(&o->obuf, err, params);
137         va_end(params);
138         if (o->buffer_output > 1)
139                 strbuf_addch(&o->obuf, '\n');
140         else {
141                 error("%s", o->obuf.buf);
142                 strbuf_reset(&o->obuf);
143         }
144
145         return -1;
146 }
147
148 static struct tree *shift_tree_object(struct tree *one, struct tree *two,
149                                       const char *subtree_shift)
150 {
151         struct object_id shifted;
152
153         if (!*subtree_shift) {
154                 shift_tree(&one->object.oid, &two->object.oid, &shifted, 0);
155         } else {
156                 shift_tree_by(&one->object.oid, &two->object.oid, &shifted,
157                               subtree_shift);
158         }
159         if (!oidcmp(&two->object.oid, &shifted))
160                 return two;
161         return lookup_tree(the_repository, &shifted);
162 }
163
164 static struct commit *make_virtual_commit(struct tree *tree, const char *comment)
165 {
166         struct commit *commit = alloc_commit_node(the_repository);
167
168         set_merge_remote_desc(commit, comment, (struct object *)commit);
169         commit->maybe_tree = tree;
170         commit->object.parsed = 1;
171         return commit;
172 }
173
174 /*
175  * Since we use get_tree_entry(), which does not put the read object into
176  * the object pool, we cannot rely on a == b.
177  */
178 static int oid_eq(const struct object_id *a, const struct object_id *b)
179 {
180         if (!a && !b)
181                 return 2;
182         return a && b && oidcmp(a, b) == 0;
183 }
184
185 enum rename_type {
186         RENAME_NORMAL = 0,
187         RENAME_DIR,
188         RENAME_DELETE,
189         RENAME_ONE_FILE_TO_ONE,
190         RENAME_ONE_FILE_TO_TWO,
191         RENAME_TWO_FILES_TO_ONE
192 };
193
194 struct rename_conflict_info {
195         enum rename_type rename_type;
196         struct diff_filepair *pair1;
197         struct diff_filepair *pair2;
198         const char *branch1;
199         const char *branch2;
200         struct stage_data *dst_entry1;
201         struct stage_data *dst_entry2;
202         struct diff_filespec ren1_other;
203         struct diff_filespec ren2_other;
204 };
205
206 /*
207  * Since we want to write the index eventually, we cannot reuse the index
208  * for these (temporary) data.
209  */
210 struct stage_data {
211         struct {
212                 unsigned mode;
213                 struct object_id oid;
214         } stages[4];
215         struct rename_conflict_info *rename_conflict_info;
216         unsigned processed:1;
217 };
218
219 static inline void setup_rename_conflict_info(enum rename_type rename_type,
220                                               struct diff_filepair *pair1,
221                                               struct diff_filepair *pair2,
222                                               const char *branch1,
223                                               const char *branch2,
224                                               struct stage_data *dst_entry1,
225                                               struct stage_data *dst_entry2,
226                                               struct merge_options *o,
227                                               struct stage_data *src_entry1,
228                                               struct stage_data *src_entry2)
229 {
230         struct rename_conflict_info *ci = xcalloc(1, sizeof(struct rename_conflict_info));
231         ci->rename_type = rename_type;
232         ci->pair1 = pair1;
233         ci->branch1 = branch1;
234         ci->branch2 = branch2;
235
236         ci->dst_entry1 = dst_entry1;
237         dst_entry1->rename_conflict_info = ci;
238         dst_entry1->processed = 0;
239
240         assert(!pair2 == !dst_entry2);
241         if (dst_entry2) {
242                 ci->dst_entry2 = dst_entry2;
243                 ci->pair2 = pair2;
244                 dst_entry2->rename_conflict_info = ci;
245         }
246
247         if (rename_type == RENAME_TWO_FILES_TO_ONE) {
248                 /*
249                  * For each rename, there could have been
250                  * modifications on the side of history where that
251                  * file was not renamed.
252                  */
253                 int ostage1 = o->branch1 == branch1 ? 3 : 2;
254                 int ostage2 = ostage1 ^ 1;
255
256                 ci->ren1_other.path = pair1->one->path;
257                 oidcpy(&ci->ren1_other.oid, &src_entry1->stages[ostage1].oid);
258                 ci->ren1_other.mode = src_entry1->stages[ostage1].mode;
259
260                 ci->ren2_other.path = pair2->one->path;
261                 oidcpy(&ci->ren2_other.oid, &src_entry2->stages[ostage2].oid);
262                 ci->ren2_other.mode = src_entry2->stages[ostage2].mode;
263         }
264 }
265
266 static int show(struct merge_options *o, int v)
267 {
268         return (!o->call_depth && o->verbosity >= v) || o->verbosity >= 5;
269 }
270
271 __attribute__((format (printf, 3, 4)))
272 static void output(struct merge_options *o, int v, const char *fmt, ...)
273 {
274         va_list ap;
275
276         if (!show(o, v))
277                 return;
278
279         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
280
281         va_start(ap, fmt);
282         strbuf_vaddf(&o->obuf, fmt, ap);
283         va_end(ap);
284
285         strbuf_addch(&o->obuf, '\n');
286         if (!o->buffer_output)
287                 flush_output(o);
288 }
289
290 static void output_commit_title(struct merge_options *o, struct commit *commit)
291 {
292         struct merge_remote_desc *desc;
293
294         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
295         desc = merge_remote_util(commit);
296         if (desc)
297                 strbuf_addf(&o->obuf, "virtual %s\n", desc->name);
298         else {
299                 strbuf_add_unique_abbrev(&o->obuf, &commit->object.oid,
300                                          DEFAULT_ABBREV);
301                 strbuf_addch(&o->obuf, ' ');
302                 if (parse_commit(commit) != 0)
303                         strbuf_addstr(&o->obuf, _("(bad commit)\n"));
304                 else {
305                         const char *title;
306                         const char *msg = get_commit_buffer(commit, NULL);
307                         int len = find_commit_subject(msg, &title);
308                         if (len)
309                                 strbuf_addf(&o->obuf, "%.*s\n", len, title);
310                         unuse_commit_buffer(commit, msg);
311                 }
312         }
313         flush_output(o);
314 }
315
316 static int add_cacheinfo(struct merge_options *o,
317                 unsigned int mode, const struct object_id *oid,
318                 const char *path, int stage, int refresh, int options)
319 {
320         struct cache_entry *ce;
321         int ret;
322
323         ce = make_cache_entry(mode, oid ? oid->hash : null_sha1, path, stage, 0);
324         if (!ce)
325                 return err(o, _("add_cacheinfo failed for path '%s'; merge aborting."), path);
326
327         ret = add_cache_entry(ce, options);
328         if (refresh) {
329                 struct cache_entry *nce;
330
331                 nce = refresh_cache_entry(ce, CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
332                 if (!nce)
333                         return err(o, _("add_cacheinfo failed to refresh for path '%s'; merge aborting."), path);
334                 if (nce != ce)
335                         ret = add_cache_entry(nce, options);
336         }
337         return ret;
338 }
339
340 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
341 {
342         parse_tree(tree);
343         init_tree_desc(desc, tree->buffer, tree->size);
344 }
345
346 static int unpack_trees_start(struct merge_options *o,
347                               struct tree *common,
348                               struct tree *head,
349                               struct tree *merge)
350 {
351         int rc;
352         struct tree_desc t[3];
353         struct index_state tmp_index = { NULL };
354
355         memset(&o->unpack_opts, 0, sizeof(o->unpack_opts));
356         if (o->call_depth)
357                 o->unpack_opts.index_only = 1;
358         else
359                 o->unpack_opts.update = 1;
360         o->unpack_opts.merge = 1;
361         o->unpack_opts.head_idx = 2;
362         o->unpack_opts.fn = threeway_merge;
363         o->unpack_opts.src_index = &the_index;
364         o->unpack_opts.dst_index = &tmp_index;
365         o->unpack_opts.aggressive = !merge_detect_rename(o);
366         setup_unpack_trees_porcelain(&o->unpack_opts, "merge");
367
368         init_tree_desc_from_tree(t+0, common);
369         init_tree_desc_from_tree(t+1, head);
370         init_tree_desc_from_tree(t+2, merge);
371
372         rc = unpack_trees(3, t, &o->unpack_opts);
373         cache_tree_free(&active_cache_tree);
374
375         /*
376          * Update the_index to match the new results, AFTER saving a copy
377          * in o->orig_index.  Update src_index to point to the saved copy.
378          * (verify_uptodate() checks src_index, and the original index is
379          * the one that had the necessary modification timestamps.)
380          */
381         o->orig_index = the_index;
382         the_index = tmp_index;
383         o->unpack_opts.src_index = &o->orig_index;
384
385         return rc;
386 }
387
388 static void unpack_trees_finish(struct merge_options *o)
389 {
390         discard_index(&o->orig_index);
391         clear_unpack_trees_porcelain(&o->unpack_opts);
392 }
393
394 struct tree *write_tree_from_memory(struct merge_options *o)
395 {
396         struct tree *result = NULL;
397
398         if (unmerged_cache()) {
399                 int i;
400                 fprintf(stderr, "BUG: There are unmerged index entries:\n");
401                 for (i = 0; i < active_nr; i++) {
402                         const struct cache_entry *ce = active_cache[i];
403                         if (ce_stage(ce))
404                                 fprintf(stderr, "BUG: %d %.*s\n", ce_stage(ce),
405                                         (int)ce_namelen(ce), ce->name);
406                 }
407                 BUG("unmerged index entries in merge-recursive.c");
408         }
409
410         if (!active_cache_tree)
411                 active_cache_tree = cache_tree();
412
413         if (!cache_tree_fully_valid(active_cache_tree) &&
414             cache_tree_update(&the_index, 0) < 0) {
415                 err(o, _("error building trees"));
416                 return NULL;
417         }
418
419         result = lookup_tree(the_repository, &active_cache_tree->oid);
420
421         return result;
422 }
423
424 static int save_files_dirs(const struct object_id *oid,
425                 struct strbuf *base, const char *path,
426                 unsigned int mode, int stage, void *context)
427 {
428         struct path_hashmap_entry *entry;
429         int baselen = base->len;
430         struct merge_options *o = context;
431
432         strbuf_addstr(base, path);
433
434         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
435         hashmap_entry_init(entry, path_hash(entry->path));
436         hashmap_add(&o->current_file_dir_set, entry);
437
438         strbuf_setlen(base, baselen);
439         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
440 }
441
442 static void get_files_dirs(struct merge_options *o, struct tree *tree)
443 {
444         struct pathspec match_all;
445         memset(&match_all, 0, sizeof(match_all));
446         read_tree_recursive(tree, "", 0, 0, &match_all, save_files_dirs, o);
447 }
448
449 static int get_tree_entry_if_blob(const struct object_id *tree,
450                                   const char *path,
451                                   struct object_id *hashy,
452                                   unsigned int *mode_o)
453 {
454         int ret;
455
456         ret = get_tree_entry(tree, path, hashy, mode_o);
457         if (S_ISDIR(*mode_o)) {
458                 oidcpy(hashy, &null_oid);
459                 *mode_o = 0;
460         }
461         return ret;
462 }
463
464 /*
465  * Returns an index_entry instance which doesn't have to correspond to
466  * a real cache entry in Git's index.
467  */
468 static struct stage_data *insert_stage_data(const char *path,
469                 struct tree *o, struct tree *a, struct tree *b,
470                 struct string_list *entries)
471 {
472         struct string_list_item *item;
473         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
474         get_tree_entry_if_blob(&o->object.oid, path,
475                                &e->stages[1].oid, &e->stages[1].mode);
476         get_tree_entry_if_blob(&a->object.oid, path,
477                                &e->stages[2].oid, &e->stages[2].mode);
478         get_tree_entry_if_blob(&b->object.oid, path,
479                                &e->stages[3].oid, &e->stages[3].mode);
480         item = string_list_insert(entries, path);
481         item->util = e;
482         return e;
483 }
484
485 /*
486  * Create a dictionary mapping file names to stage_data objects. The
487  * dictionary contains one entry for every path with a non-zero stage entry.
488  */
489 static struct string_list *get_unmerged(void)
490 {
491         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
492         int i;
493
494         unmerged->strdup_strings = 1;
495
496         for (i = 0; i < active_nr; i++) {
497                 struct string_list_item *item;
498                 struct stage_data *e;
499                 const struct cache_entry *ce = active_cache[i];
500                 if (!ce_stage(ce))
501                         continue;
502
503                 item = string_list_lookup(unmerged, ce->name);
504                 if (!item) {
505                         item = string_list_insert(unmerged, ce->name);
506                         item->util = xcalloc(1, sizeof(struct stage_data));
507                 }
508                 e = item->util;
509                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
510                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
511         }
512
513         return unmerged;
514 }
515
516 static int string_list_df_name_compare(const char *one, const char *two)
517 {
518         int onelen = strlen(one);
519         int twolen = strlen(two);
520         /*
521          * Here we only care that entries for D/F conflicts are
522          * adjacent, in particular with the file of the D/F conflict
523          * appearing before files below the corresponding directory.
524          * The order of the rest of the list is irrelevant for us.
525          *
526          * To achieve this, we sort with df_name_compare and provide
527          * the mode S_IFDIR so that D/F conflicts will sort correctly.
528          * We use the mode S_IFDIR for everything else for simplicity,
529          * since in other cases any changes in their order due to
530          * sorting cause no problems for us.
531          */
532         int cmp = df_name_compare(one, onelen, S_IFDIR,
533                                   two, twolen, S_IFDIR);
534         /*
535          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
536          * that 'foo' comes before 'foo/bar'.
537          */
538         if (cmp)
539                 return cmp;
540         return onelen - twolen;
541 }
542
543 static void record_df_conflict_files(struct merge_options *o,
544                                      struct string_list *entries)
545 {
546         /* If there is a D/F conflict and the file for such a conflict
547          * currently exist in the working tree, we want to allow it to be
548          * removed to make room for the corresponding directory if needed.
549          * The files underneath the directories of such D/F conflicts will
550          * be processed before the corresponding file involved in the D/F
551          * conflict.  If the D/F directory ends up being removed by the
552          * merge, then we won't have to touch the D/F file.  If the D/F
553          * directory needs to be written to the working copy, then the D/F
554          * file will simply be removed (in make_room_for_path()) to make
555          * room for the necessary paths.  Note that if both the directory
556          * and the file need to be present, then the D/F file will be
557          * reinstated with a new unique name at the time it is processed.
558          */
559         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
560         const char *last_file = NULL;
561         int last_len = 0;
562         int i;
563
564         /*
565          * If we're merging merge-bases, we don't want to bother with
566          * any working directory changes.
567          */
568         if (o->call_depth)
569                 return;
570
571         /* Ensure D/F conflicts are adjacent in the entries list. */
572         for (i = 0; i < entries->nr; i++) {
573                 struct string_list_item *next = &entries->items[i];
574                 string_list_append(&df_sorted_entries, next->string)->util =
575                                    next->util;
576         }
577         df_sorted_entries.cmp = string_list_df_name_compare;
578         string_list_sort(&df_sorted_entries);
579
580         string_list_clear(&o->df_conflict_file_set, 1);
581         for (i = 0; i < df_sorted_entries.nr; i++) {
582                 const char *path = df_sorted_entries.items[i].string;
583                 int len = strlen(path);
584                 struct stage_data *e = df_sorted_entries.items[i].util;
585
586                 /*
587                  * Check if last_file & path correspond to a D/F conflict;
588                  * i.e. whether path is last_file+'/'+<something>.
589                  * If so, record that it's okay to remove last_file to make
590                  * room for path and friends if needed.
591                  */
592                 if (last_file &&
593                     len > last_len &&
594                     memcmp(path, last_file, last_len) == 0 &&
595                     path[last_len] == '/') {
596                         string_list_insert(&o->df_conflict_file_set, last_file);
597                 }
598
599                 /*
600                  * Determine whether path could exist as a file in the
601                  * working directory as a possible D/F conflict.  This
602                  * will only occur when it exists in stage 2 as a
603                  * file.
604                  */
605                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
606                         last_file = path;
607                         last_len = len;
608                 } else {
609                         last_file = NULL;
610                 }
611         }
612         string_list_clear(&df_sorted_entries, 0);
613 }
614
615 struct rename {
616         struct diff_filepair *pair;
617         /*
618          * Purpose of src_entry and dst_entry:
619          *
620          * If 'before' is renamed to 'after' then src_entry will contain
621          * the versions of 'before' from the merge_base, HEAD, and MERGE in
622          * stages 1, 2, and 3; dst_entry will contain the respective
623          * versions of 'after' in corresponding locations.  Thus, we have a
624          * total of six modes and oids, though some will be null.  (Stage 0
625          * is ignored; we're interested in handling conflicts.)
626          *
627          * Since we don't turn on break-rewrites by default, neither
628          * src_entry nor dst_entry can have all three of their stages have
629          * non-null oids, meaning at most four of the six will be non-null.
630          * Also, since this is a rename, both src_entry and dst_entry will
631          * have at least one non-null oid, meaning at least two will be
632          * non-null.  Of the six oids, a typical rename will have three be
633          * non-null.  Only two implies a rename/delete, and four implies a
634          * rename/add.
635          */
636         struct stage_data *src_entry;
637         struct stage_data *dst_entry;
638         unsigned add_turned_into_rename:1;
639         unsigned processed:1;
640 };
641
642 static int update_stages(struct merge_options *opt, const char *path,
643                          const struct diff_filespec *o,
644                          const struct diff_filespec *a,
645                          const struct diff_filespec *b)
646 {
647
648         /*
649          * NOTE: It is usually a bad idea to call update_stages on a path
650          * before calling update_file on that same path, since it can
651          * sometimes lead to spurious "refusing to lose untracked file..."
652          * messages from update_file (via make_room_for path via
653          * would_lose_untracked).  Instead, reverse the order of the calls
654          * (executing update_file first and then update_stages).
655          */
656         int clear = 1;
657         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
658         if (clear)
659                 if (remove_file_from_cache(path))
660                         return -1;
661         if (o)
662                 if (add_cacheinfo(opt, o->mode, &o->oid, path, 1, 0, options))
663                         return -1;
664         if (a)
665                 if (add_cacheinfo(opt, a->mode, &a->oid, path, 2, 0, options))
666                         return -1;
667         if (b)
668                 if (add_cacheinfo(opt, b->mode, &b->oid, path, 3, 0, options))
669                         return -1;
670         return 0;
671 }
672
673 static int update_stages_for_stage_data(struct merge_options *opt,
674                                         const char *path,
675                                         const struct stage_data *stage_data)
676 {
677         struct diff_filespec o, a, b;
678
679         o.mode = stage_data->stages[1].mode;
680         oidcpy(&o.oid, &stage_data->stages[1].oid);
681
682         a.mode = stage_data->stages[2].mode;
683         oidcpy(&a.oid, &stage_data->stages[2].oid);
684
685         b.mode = stage_data->stages[3].mode;
686         oidcpy(&b.oid, &stage_data->stages[3].oid);
687
688         return update_stages(opt, path,
689                              is_null_oid(&o.oid) ? NULL : &o,
690                              is_null_oid(&a.oid) ? NULL : &a,
691                              is_null_oid(&b.oid) ? NULL : &b);
692 }
693
694 static void update_entry(struct stage_data *entry,
695                          struct diff_filespec *o,
696                          struct diff_filespec *a,
697                          struct diff_filespec *b)
698 {
699         entry->processed = 0;
700         entry->stages[1].mode = o->mode;
701         entry->stages[2].mode = a->mode;
702         entry->stages[3].mode = b->mode;
703         oidcpy(&entry->stages[1].oid, &o->oid);
704         oidcpy(&entry->stages[2].oid, &a->oid);
705         oidcpy(&entry->stages[3].oid, &b->oid);
706 }
707
708 static int remove_file(struct merge_options *o, int clean,
709                        const char *path, int no_wd)
710 {
711         int update_cache = o->call_depth || clean;
712         int update_working_directory = !o->call_depth && !no_wd;
713
714         if (update_cache) {
715                 if (remove_file_from_cache(path))
716                         return -1;
717         }
718         if (update_working_directory) {
719                 if (ignore_case) {
720                         struct cache_entry *ce;
721                         ce = cache_file_exists(path, strlen(path), ignore_case);
722                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
723                                 return 0;
724                 }
725                 if (remove_path(path))
726                         return -1;
727         }
728         return 0;
729 }
730
731 /* add a string to a strbuf, but converting "/" to "_" */
732 static void add_flattened_path(struct strbuf *out, const char *s)
733 {
734         size_t i = out->len;
735         strbuf_addstr(out, s);
736         for (; i < out->len; i++)
737                 if (out->buf[i] == '/')
738                         out->buf[i] = '_';
739 }
740
741 static char *unique_path(struct merge_options *o, const char *path, const char *branch)
742 {
743         struct path_hashmap_entry *entry;
744         struct strbuf newpath = STRBUF_INIT;
745         int suffix = 0;
746         size_t base_len;
747
748         strbuf_addf(&newpath, "%s~", path);
749         add_flattened_path(&newpath, branch);
750
751         base_len = newpath.len;
752         while (hashmap_get_from_hash(&o->current_file_dir_set,
753                                      path_hash(newpath.buf), newpath.buf) ||
754                (!o->call_depth && file_exists(newpath.buf))) {
755                 strbuf_setlen(&newpath, base_len);
756                 strbuf_addf(&newpath, "_%d", suffix++);
757         }
758
759         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
760         hashmap_entry_init(entry, path_hash(entry->path));
761         hashmap_add(&o->current_file_dir_set, entry);
762         return strbuf_detach(&newpath, NULL);
763 }
764
765 /**
766  * Check whether a directory in the index is in the way of an incoming
767  * file.  Return 1 if so.  If check_working_copy is non-zero, also
768  * check the working directory.  If empty_ok is non-zero, also return
769  * 0 in the case where the working-tree dir exists but is empty.
770  */
771 static int dir_in_way(const char *path, int check_working_copy, int empty_ok)
772 {
773         int pos;
774         struct strbuf dirpath = STRBUF_INIT;
775         struct stat st;
776
777         strbuf_addstr(&dirpath, path);
778         strbuf_addch(&dirpath, '/');
779
780         pos = cache_name_pos(dirpath.buf, dirpath.len);
781
782         if (pos < 0)
783                 pos = -1 - pos;
784         if (pos < active_nr &&
785             !strncmp(dirpath.buf, active_cache[pos]->name, dirpath.len)) {
786                 strbuf_release(&dirpath);
787                 return 1;
788         }
789
790         strbuf_release(&dirpath);
791         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
792                 !(empty_ok && is_empty_dir(path));
793 }
794
795 /*
796  * Returns whether path was tracked in the index before the merge started,
797  * and its oid and mode match the specified values
798  */
799 static int was_tracked_and_matches(struct merge_options *o, const char *path,
800                                    const struct object_id *oid, unsigned mode)
801 {
802         int pos = index_name_pos(&o->orig_index, path, strlen(path));
803         struct cache_entry *ce;
804
805         if (0 > pos)
806                 /* we were not tracking this path before the merge */
807                 return 0;
808
809         /* See if the file we were tracking before matches */
810         ce = o->orig_index.cache[pos];
811         return (oid_eq(&ce->oid, oid) && ce->ce_mode == mode);
812 }
813
814 /*
815  * Returns whether path was tracked in the index before the merge started
816  */
817 static int was_tracked(struct merge_options *o, const char *path)
818 {
819         int pos = index_name_pos(&o->orig_index, path, strlen(path));
820
821         if (0 <= pos)
822                 /* we were tracking this path before the merge */
823                 return 1;
824
825         return 0;
826 }
827
828 static int would_lose_untracked(const char *path)
829 {
830         /*
831          * This may look like it can be simplified to:
832          *   return !was_tracked(o, path) && file_exists(path)
833          * but it can't.  This function needs to know whether path was in
834          * the working tree due to EITHER having been tracked in the index
835          * before the merge OR having been put into the working copy and
836          * index by unpack_trees().  Due to that either-or requirement, we
837          * check the current index instead of the original one.
838          *
839          * Note that we do not need to worry about merge-recursive itself
840          * updating the index after unpack_trees() and before calling this
841          * function, because we strictly require all code paths in
842          * merge-recursive to update the working tree first and the index
843          * second.  Doing otherwise would break
844          * update_file()/would_lose_untracked(); see every comment in this
845          * file which mentions "update_stages".
846          */
847         int pos = cache_name_pos(path, strlen(path));
848
849         if (pos < 0)
850                 pos = -1 - pos;
851         while (pos < active_nr &&
852                !strcmp(path, active_cache[pos]->name)) {
853                 /*
854                  * If stage #0, it is definitely tracked.
855                  * If it has stage #2 then it was tracked
856                  * before this merge started.  All other
857                  * cases the path was not tracked.
858                  */
859                 switch (ce_stage(active_cache[pos])) {
860                 case 0:
861                 case 2:
862                         return 0;
863                 }
864                 pos++;
865         }
866         return file_exists(path);
867 }
868
869 static int was_dirty(struct merge_options *o, const char *path)
870 {
871         struct cache_entry *ce;
872         int dirty = 1;
873
874         if (o->call_depth || !was_tracked(o, path))
875                 return !dirty;
876
877         ce = index_file_exists(o->unpack_opts.src_index,
878                                path, strlen(path), ignore_case);
879         dirty = verify_uptodate(ce, &o->unpack_opts) != 0;
880         return dirty;
881 }
882
883 static int make_room_for_path(struct merge_options *o, const char *path)
884 {
885         int status, i;
886         const char *msg = _("failed to create path '%s'%s");
887
888         /* Unlink any D/F conflict files that are in the way */
889         for (i = 0; i < o->df_conflict_file_set.nr; i++) {
890                 const char *df_path = o->df_conflict_file_set.items[i].string;
891                 size_t pathlen = strlen(path);
892                 size_t df_pathlen = strlen(df_path);
893                 if (df_pathlen < pathlen &&
894                     path[df_pathlen] == '/' &&
895                     strncmp(path, df_path, df_pathlen) == 0) {
896                         output(o, 3,
897                                _("Removing %s to make room for subdirectory\n"),
898                                df_path);
899                         unlink(df_path);
900                         unsorted_string_list_delete_item(&o->df_conflict_file_set,
901                                                          i, 0);
902                         break;
903                 }
904         }
905
906         /* Make sure leading directories are created */
907         status = safe_create_leading_directories_const(path);
908         if (status) {
909                 if (status == SCLD_EXISTS)
910                         /* something else exists */
911                         return err(o, msg, path, _(": perhaps a D/F conflict?"));
912                 return err(o, msg, path, "");
913         }
914
915         /*
916          * Do not unlink a file in the work tree if we are not
917          * tracking it.
918          */
919         if (would_lose_untracked(path))
920                 return err(o, _("refusing to lose untracked file at '%s'"),
921                              path);
922
923         /* Successful unlink is good.. */
924         if (!unlink(path))
925                 return 0;
926         /* .. and so is no existing file */
927         if (errno == ENOENT)
928                 return 0;
929         /* .. but not some other error (who really cares what?) */
930         return err(o, msg, path, _(": perhaps a D/F conflict?"));
931 }
932
933 static int update_file_flags(struct merge_options *o,
934                              const struct object_id *oid,
935                              unsigned mode,
936                              const char *path,
937                              int update_cache,
938                              int update_wd)
939 {
940         int ret = 0;
941
942         if (o->call_depth)
943                 update_wd = 0;
944
945         if (update_wd) {
946                 enum object_type type;
947                 void *buf;
948                 unsigned long size;
949
950                 if (S_ISGITLINK(mode)) {
951                         /*
952                          * We may later decide to recursively descend into
953                          * the submodule directory and update its index
954                          * and/or work tree, but we do not do that now.
955                          */
956                         update_wd = 0;
957                         goto update_index;
958                 }
959
960                 buf = read_object_file(oid, &type, &size);
961                 if (!buf)
962                         return err(o, _("cannot read object %s '%s'"), oid_to_hex(oid), path);
963                 if (type != OBJ_BLOB) {
964                         ret = err(o, _("blob expected for %s '%s'"), oid_to_hex(oid), path);
965                         goto free_buf;
966                 }
967                 if (S_ISREG(mode)) {
968                         struct strbuf strbuf = STRBUF_INIT;
969                         if (convert_to_working_tree(path, buf, size, &strbuf)) {
970                                 free(buf);
971                                 size = strbuf.len;
972                                 buf = strbuf_detach(&strbuf, NULL);
973                         }
974                 }
975
976                 if (make_room_for_path(o, path) < 0) {
977                         update_wd = 0;
978                         goto free_buf;
979                 }
980                 if (S_ISREG(mode) || (!has_symlinks && S_ISLNK(mode))) {
981                         int fd;
982                         if (mode & 0100)
983                                 mode = 0777;
984                         else
985                                 mode = 0666;
986                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
987                         if (fd < 0) {
988                                 ret = err(o, _("failed to open '%s': %s"),
989                                           path, strerror(errno));
990                                 goto free_buf;
991                         }
992                         write_in_full(fd, buf, size);
993                         close(fd);
994                 } else if (S_ISLNK(mode)) {
995                         char *lnk = xmemdupz(buf, size);
996                         safe_create_leading_directories_const(path);
997                         unlink(path);
998                         if (symlink(lnk, path))
999                                 ret = err(o, _("failed to symlink '%s': %s"),
1000                                         path, strerror(errno));
1001                         free(lnk);
1002                 } else
1003                         ret = err(o,
1004                                   _("do not know what to do with %06o %s '%s'"),
1005                                   mode, oid_to_hex(oid), path);
1006  free_buf:
1007                 free(buf);
1008         }
1009  update_index:
1010         if (!ret && update_cache)
1011                 if (add_cacheinfo(o, mode, oid, path, 0, update_wd,
1012                                   ADD_CACHE_OK_TO_ADD))
1013                         return -1;
1014         return ret;
1015 }
1016
1017 static int update_file(struct merge_options *o,
1018                        int clean,
1019                        const struct object_id *oid,
1020                        unsigned mode,
1021                        const char *path)
1022 {
1023         return update_file_flags(o, oid, mode, path, o->call_depth || clean, !o->call_depth);
1024 }
1025
1026 /* Low level file merging, update and removal */
1027
1028 struct merge_file_info {
1029         struct object_id oid;
1030         unsigned mode;
1031         unsigned clean:1,
1032                  merge:1;
1033 };
1034
1035 static int merge_3way(struct merge_options *o,
1036                       mmbuffer_t *result_buf,
1037                       const struct diff_filespec *one,
1038                       const struct diff_filespec *a,
1039                       const struct diff_filespec *b,
1040                       const char *branch1,
1041                       const char *branch2)
1042 {
1043         mmfile_t orig, src1, src2;
1044         struct ll_merge_options ll_opts = {0};
1045         char *base_name, *name1, *name2;
1046         int merge_status;
1047
1048         ll_opts.renormalize = o->renormalize;
1049         ll_opts.xdl_opts = o->xdl_opts;
1050
1051         if (o->call_depth) {
1052                 ll_opts.virtual_ancestor = 1;
1053                 ll_opts.variant = 0;
1054         } else {
1055                 switch (o->recursive_variant) {
1056                 case MERGE_RECURSIVE_OURS:
1057                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
1058                         break;
1059                 case MERGE_RECURSIVE_THEIRS:
1060                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
1061                         break;
1062                 default:
1063                         ll_opts.variant = 0;
1064                         break;
1065                 }
1066         }
1067
1068         if (strcmp(a->path, b->path) ||
1069             (o->ancestor != NULL && strcmp(a->path, one->path) != 0)) {
1070                 base_name = o->ancestor == NULL ? NULL :
1071                         mkpathdup("%s:%s", o->ancestor, one->path);
1072                 name1 = mkpathdup("%s:%s", branch1, a->path);
1073                 name2 = mkpathdup("%s:%s", branch2, b->path);
1074         } else {
1075                 base_name = o->ancestor == NULL ? NULL :
1076                         mkpathdup("%s", o->ancestor);
1077                 name1 = mkpathdup("%s", branch1);
1078                 name2 = mkpathdup("%s", branch2);
1079         }
1080
1081         read_mmblob(&orig, &one->oid);
1082         read_mmblob(&src1, &a->oid);
1083         read_mmblob(&src2, &b->oid);
1084
1085         merge_status = ll_merge(result_buf, a->path, &orig, base_name,
1086                                 &src1, name1, &src2, name2, &ll_opts);
1087
1088         free(base_name);
1089         free(name1);
1090         free(name2);
1091         free(orig.ptr);
1092         free(src1.ptr);
1093         free(src2.ptr);
1094         return merge_status;
1095 }
1096
1097 static int find_first_merges(struct object_array *result, const char *path,
1098                 struct commit *a, struct commit *b)
1099 {
1100         int i, j;
1101         struct object_array merges = OBJECT_ARRAY_INIT;
1102         struct commit *commit;
1103         int contains_another;
1104
1105         char merged_revision[42];
1106         const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
1107                                    "--all", merged_revision, NULL };
1108         struct rev_info revs;
1109         struct setup_revision_opt rev_opts;
1110
1111         memset(result, 0, sizeof(struct object_array));
1112         memset(&rev_opts, 0, sizeof(rev_opts));
1113
1114         /* get all revisions that merge commit a */
1115         xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
1116                         oid_to_hex(&a->object.oid));
1117         init_revisions(&revs, NULL);
1118         rev_opts.submodule = path;
1119         /* FIXME: can't handle linked worktrees in submodules yet */
1120         revs.single_worktree = path != NULL;
1121         setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
1122
1123         /* save all revisions from the above list that contain b */
1124         if (prepare_revision_walk(&revs))
1125                 die("revision walk setup failed");
1126         while ((commit = get_revision(&revs)) != NULL) {
1127                 struct object *o = &(commit->object);
1128                 if (in_merge_bases(b, commit))
1129                         add_object_array(o, NULL, &merges);
1130         }
1131         reset_revision_walk();
1132
1133         /* Now we've got all merges that contain a and b. Prune all
1134          * merges that contain another found merge and save them in
1135          * result.
1136          */
1137         for (i = 0; i < merges.nr; i++) {
1138                 struct commit *m1 = (struct commit *) merges.objects[i].item;
1139
1140                 contains_another = 0;
1141                 for (j = 0; j < merges.nr; j++) {
1142                         struct commit *m2 = (struct commit *) merges.objects[j].item;
1143                         if (i != j && in_merge_bases(m2, m1)) {
1144                                 contains_another = 1;
1145                                 break;
1146                         }
1147                 }
1148
1149                 if (!contains_another)
1150                         add_object_array(merges.objects[i].item, NULL, result);
1151         }
1152
1153         object_array_clear(&merges);
1154         return result->nr;
1155 }
1156
1157 static void print_commit(struct commit *commit)
1158 {
1159         struct strbuf sb = STRBUF_INIT;
1160         struct pretty_print_context ctx = {0};
1161         ctx.date_mode.type = DATE_NORMAL;
1162         format_commit_message(commit, " %h: %m %s", &sb, &ctx);
1163         fprintf(stderr, "%s\n", sb.buf);
1164         strbuf_release(&sb);
1165 }
1166
1167 static int merge_submodule(struct merge_options *o,
1168                            struct object_id *result, const char *path,
1169                            const struct object_id *base, const struct object_id *a,
1170                            const struct object_id *b)
1171 {
1172         struct commit *commit_base, *commit_a, *commit_b;
1173         int parent_count;
1174         struct object_array merges;
1175
1176         int i;
1177         int search = !o->call_depth;
1178
1179         /* store a in result in case we fail */
1180         oidcpy(result, a);
1181
1182         /* we can not handle deletion conflicts */
1183         if (is_null_oid(base))
1184                 return 0;
1185         if (is_null_oid(a))
1186                 return 0;
1187         if (is_null_oid(b))
1188                 return 0;
1189
1190         if (add_submodule_odb(path)) {
1191                 output(o, 1, _("Failed to merge submodule %s (not checked out)"), path);
1192                 return 0;
1193         }
1194
1195         if (!(commit_base = lookup_commit_reference(the_repository, base)) ||
1196             !(commit_a = lookup_commit_reference(the_repository, a)) ||
1197             !(commit_b = lookup_commit_reference(the_repository, b))) {
1198                 output(o, 1, _("Failed to merge submodule %s (commits not present)"), path);
1199                 return 0;
1200         }
1201
1202         /* check whether both changes are forward */
1203         if (!in_merge_bases(commit_base, commit_a) ||
1204             !in_merge_bases(commit_base, commit_b)) {
1205                 output(o, 1, _("Failed to merge submodule %s (commits don't follow merge-base)"), path);
1206                 return 0;
1207         }
1208
1209         /* Case #1: a is contained in b or vice versa */
1210         if (in_merge_bases(commit_a, commit_b)) {
1211                 oidcpy(result, b);
1212                 if (show(o, 3)) {
1213                         output(o, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1214                         output_commit_title(o, commit_b);
1215                 } else if (show(o, 2))
1216                         output(o, 2, _("Fast-forwarding submodule %s"), path);
1217                 else
1218                         ; /* no output */
1219
1220                 return 1;
1221         }
1222         if (in_merge_bases(commit_b, commit_a)) {
1223                 oidcpy(result, a);
1224                 if (show(o, 3)) {
1225                         output(o, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1226                         output_commit_title(o, commit_a);
1227                 } else if (show(o, 2))
1228                         output(o, 2, _("Fast-forwarding submodule %s"), path);
1229                 else
1230                         ; /* no output */
1231
1232                 return 1;
1233         }
1234
1235         /*
1236          * Case #2: There are one or more merges that contain a and b in
1237          * the submodule. If there is only one, then present it as a
1238          * suggestion to the user, but leave it marked unmerged so the
1239          * user needs to confirm the resolution.
1240          */
1241
1242         /* Skip the search if makes no sense to the calling context.  */
1243         if (!search)
1244                 return 0;
1245
1246         /* find commit which merges them */
1247         parent_count = find_first_merges(&merges, path, commit_a, commit_b);
1248         switch (parent_count) {
1249         case 0:
1250                 output(o, 1, _("Failed to merge submodule %s (merge following commits not found)"), path);
1251                 break;
1252
1253         case 1:
1254                 output(o, 1, _("Failed to merge submodule %s (not fast-forward)"), path);
1255                 output(o, 2, _("Found a possible merge resolution for the submodule:\n"));
1256                 print_commit((struct commit *) merges.objects[0].item);
1257                 output(o, 2, _(
1258                         "If this is correct simply add it to the index "
1259                         "for example\n"
1260                         "by using:\n\n"
1261                         "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
1262                         "which will accept this suggestion.\n"),
1263                         oid_to_hex(&merges.objects[0].item->oid), path);
1264                 break;
1265
1266         default:
1267                 output(o, 1, _("Failed to merge submodule %s (multiple merges found)"), path);
1268                 for (i = 0; i < merges.nr; i++)
1269                         print_commit((struct commit *) merges.objects[i].item);
1270         }
1271
1272         object_array_clear(&merges);
1273         return 0;
1274 }
1275
1276 static int merge_file_1(struct merge_options *o,
1277                         const struct diff_filespec *one,
1278                         const struct diff_filespec *a,
1279                         const struct diff_filespec *b,
1280                         const char *filename,
1281                         const char *branch1,
1282                         const char *branch2,
1283                         struct merge_file_info *result)
1284 {
1285         result->merge = 0;
1286         result->clean = 1;
1287
1288         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
1289                 result->clean = 0;
1290                 if (S_ISREG(a->mode)) {
1291                         result->mode = a->mode;
1292                         oidcpy(&result->oid, &a->oid);
1293                 } else {
1294                         result->mode = b->mode;
1295                         oidcpy(&result->oid, &b->oid);
1296                 }
1297         } else {
1298                 if (!oid_eq(&a->oid, &one->oid) && !oid_eq(&b->oid, &one->oid))
1299                         result->merge = 1;
1300
1301                 /*
1302                  * Merge modes
1303                  */
1304                 if (a->mode == b->mode || a->mode == one->mode)
1305                         result->mode = b->mode;
1306                 else {
1307                         result->mode = a->mode;
1308                         if (b->mode != one->mode) {
1309                                 result->clean = 0;
1310                                 result->merge = 1;
1311                         }
1312                 }
1313
1314                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &one->oid))
1315                         oidcpy(&result->oid, &b->oid);
1316                 else if (oid_eq(&b->oid, &one->oid))
1317                         oidcpy(&result->oid, &a->oid);
1318                 else if (S_ISREG(a->mode)) {
1319                         mmbuffer_t result_buf;
1320                         int ret = 0, merge_status;
1321
1322                         merge_status = merge_3way(o, &result_buf, one, a, b,
1323                                                   branch1, branch2);
1324
1325                         if ((merge_status < 0) || !result_buf.ptr)
1326                                 ret = err(o, _("Failed to execute internal merge"));
1327
1328                         if (!ret &&
1329                             write_object_file(result_buf.ptr, result_buf.size,
1330                                               blob_type, &result->oid))
1331                                 ret = err(o, _("Unable to add %s to database"),
1332                                           a->path);
1333
1334                         free(result_buf.ptr);
1335                         if (ret)
1336                                 return ret;
1337                         result->clean = (merge_status == 0);
1338                 } else if (S_ISGITLINK(a->mode)) {
1339                         result->clean = merge_submodule(o, &result->oid,
1340                                                        one->path,
1341                                                        &one->oid,
1342                                                        &a->oid,
1343                                                        &b->oid);
1344                 } else if (S_ISLNK(a->mode)) {
1345                         switch (o->recursive_variant) {
1346                         case MERGE_RECURSIVE_NORMAL:
1347                                 oidcpy(&result->oid, &a->oid);
1348                                 if (!oid_eq(&a->oid, &b->oid))
1349                                         result->clean = 0;
1350                                 break;
1351                         case MERGE_RECURSIVE_OURS:
1352                                 oidcpy(&result->oid, &a->oid);
1353                                 break;
1354                         case MERGE_RECURSIVE_THEIRS:
1355                                 oidcpy(&result->oid, &b->oid);
1356                                 break;
1357                         }
1358                 } else
1359                         BUG("unsupported object type in the tree");
1360         }
1361
1362         if (result->merge)
1363                 output(o, 2, _("Auto-merging %s"), filename);
1364
1365         return 0;
1366 }
1367
1368 static int merge_file_special_markers(struct merge_options *o,
1369                                       const struct diff_filespec *one,
1370                                       const struct diff_filespec *a,
1371                                       const struct diff_filespec *b,
1372                                       const char *target_filename,
1373                                       const char *branch1,
1374                                       const char *filename1,
1375                                       const char *branch2,
1376                                       const char *filename2,
1377                                       struct merge_file_info *mfi)
1378 {
1379         char *side1 = NULL;
1380         char *side2 = NULL;
1381         int ret;
1382
1383         if (filename1)
1384                 side1 = xstrfmt("%s:%s", branch1, filename1);
1385         if (filename2)
1386                 side2 = xstrfmt("%s:%s", branch2, filename2);
1387
1388         ret = merge_file_1(o, one, a, b, target_filename,
1389                            side1 ? side1 : branch1,
1390                            side2 ? side2 : branch2, mfi);
1391
1392         free(side1);
1393         free(side2);
1394         return ret;
1395 }
1396
1397 static int merge_file_one(struct merge_options *o,
1398                           const char *path,
1399                           const struct object_id *o_oid, int o_mode,
1400                           const struct object_id *a_oid, int a_mode,
1401                           const struct object_id *b_oid, int b_mode,
1402                           const char *branch1,
1403                           const char *branch2,
1404                           struct merge_file_info *mfi)
1405 {
1406         struct diff_filespec one, a, b;
1407
1408         one.path = a.path = b.path = (char *)path;
1409         oidcpy(&one.oid, o_oid);
1410         one.mode = o_mode;
1411         oidcpy(&a.oid, a_oid);
1412         a.mode = a_mode;
1413         oidcpy(&b.oid, b_oid);
1414         b.mode = b_mode;
1415         return merge_file_1(o, &one, &a, &b, path, branch1, branch2, mfi);
1416 }
1417
1418 static int conflict_rename_dir(struct merge_options *o,
1419                                struct diff_filepair *pair,
1420                                const char *rename_branch,
1421                                const char *other_branch)
1422 {
1423         const struct diff_filespec *dest = pair->two;
1424
1425         if (!o->call_depth && would_lose_untracked(dest->path)) {
1426                 char *alt_path = unique_path(o, dest->path, rename_branch);
1427
1428                 output(o, 1, _("Error: Refusing to lose untracked file at %s; "
1429                                "writing to %s instead."),
1430                        dest->path, alt_path);
1431                 /*
1432                  * Write the file in worktree at alt_path, but not in the
1433                  * index.  Instead, write to dest->path for the index but
1434                  * only at the higher appropriate stage.
1435                  */
1436                 if (update_file(o, 0, &dest->oid, dest->mode, alt_path))
1437                         return -1;
1438                 free(alt_path);
1439                 return update_stages(o, dest->path, NULL,
1440                                      rename_branch == o->branch1 ? dest : NULL,
1441                                      rename_branch == o->branch1 ? NULL : dest);
1442         }
1443
1444         /* Update dest->path both in index and in worktree */
1445         if (update_file(o, 1, &dest->oid, dest->mode, dest->path))
1446                 return -1;
1447         return 0;
1448 }
1449
1450 static int handle_change_delete(struct merge_options *o,
1451                                  const char *path, const char *old_path,
1452                                  const struct object_id *o_oid, int o_mode,
1453                                  const struct object_id *changed_oid,
1454                                  int changed_mode,
1455                                  const char *change_branch,
1456                                  const char *delete_branch,
1457                                  const char *change, const char *change_past)
1458 {
1459         char *alt_path = NULL;
1460         const char *update_path = path;
1461         int ret = 0;
1462
1463         if (dir_in_way(path, !o->call_depth, 0) ||
1464             (!o->call_depth && would_lose_untracked(path))) {
1465                 update_path = alt_path = unique_path(o, path, change_branch);
1466         }
1467
1468         if (o->call_depth) {
1469                 /*
1470                  * We cannot arbitrarily accept either a_sha or b_sha as
1471                  * correct; since there is no true "middle point" between
1472                  * them, simply reuse the base version for virtual merge base.
1473                  */
1474                 ret = remove_file_from_cache(path);
1475                 if (!ret)
1476                         ret = update_file(o, 0, o_oid, o_mode, update_path);
1477         } else {
1478                 if (!alt_path) {
1479                         if (!old_path) {
1480                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1481                                        "and %s in %s. Version %s of %s left in tree."),
1482                                        change, path, delete_branch, change_past,
1483                                        change_branch, change_branch, path);
1484                         } else {
1485                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1486                                        "and %s to %s in %s. Version %s of %s left in tree."),
1487                                        change, old_path, delete_branch, change_past, path,
1488                                        change_branch, change_branch, path);
1489                         }
1490                 } else {
1491                         if (!old_path) {
1492                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1493                                        "and %s in %s. Version %s of %s left in tree at %s."),
1494                                        change, path, delete_branch, change_past,
1495                                        change_branch, change_branch, path, alt_path);
1496                         } else {
1497                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1498                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1499                                        change, old_path, delete_branch, change_past, path,
1500                                        change_branch, change_branch, path, alt_path);
1501                         }
1502                 }
1503                 /*
1504                  * No need to call update_file() on path when change_branch ==
1505                  * o->branch1 && !alt_path, since that would needlessly touch
1506                  * path.  We could call update_file_flags() with update_cache=0
1507                  * and update_wd=0, but that's a no-op.
1508                  */
1509                 if (change_branch != o->branch1 || alt_path)
1510                         ret = update_file(o, 0, changed_oid, changed_mode, update_path);
1511         }
1512         free(alt_path);
1513
1514         return ret;
1515 }
1516
1517 static int conflict_rename_delete(struct merge_options *o,
1518                                    struct diff_filepair *pair,
1519                                    const char *rename_branch,
1520                                    const char *delete_branch)
1521 {
1522         const struct diff_filespec *orig = pair->one;
1523         const struct diff_filespec *dest = pair->two;
1524
1525         if (handle_change_delete(o,
1526                                  o->call_depth ? orig->path : dest->path,
1527                                  o->call_depth ? NULL : orig->path,
1528                                  &orig->oid, orig->mode,
1529                                  &dest->oid, dest->mode,
1530                                  rename_branch, delete_branch,
1531                                  _("rename"), _("renamed")))
1532                 return -1;
1533
1534         if (o->call_depth)
1535                 return remove_file_from_cache(dest->path);
1536         else
1537                 return update_stages(o, dest->path, NULL,
1538                                      rename_branch == o->branch1 ? dest : NULL,
1539                                      rename_branch == o->branch1 ? NULL : dest);
1540 }
1541
1542 static struct diff_filespec *filespec_from_entry(struct diff_filespec *target,
1543                                                  struct stage_data *entry,
1544                                                  int stage)
1545 {
1546         struct object_id *oid = &entry->stages[stage].oid;
1547         unsigned mode = entry->stages[stage].mode;
1548         if (mode == 0 || is_null_oid(oid))
1549                 return NULL;
1550         oidcpy(&target->oid, oid);
1551         target->mode = mode;
1552         return target;
1553 }
1554
1555 static int handle_file(struct merge_options *o,
1556                         struct diff_filespec *rename,
1557                         int stage,
1558                         struct rename_conflict_info *ci)
1559 {
1560         char *dst_name = rename->path;
1561         struct stage_data *dst_entry;
1562         const char *cur_branch, *other_branch;
1563         struct diff_filespec other;
1564         struct diff_filespec *add;
1565         int ret;
1566
1567         if (stage == 2) {
1568                 dst_entry = ci->dst_entry1;
1569                 cur_branch = ci->branch1;
1570                 other_branch = ci->branch2;
1571         } else {
1572                 dst_entry = ci->dst_entry2;
1573                 cur_branch = ci->branch2;
1574                 other_branch = ci->branch1;
1575         }
1576
1577         add = filespec_from_entry(&other, dst_entry, stage ^ 1);
1578         if (add) {
1579                 int ren_src_was_dirty = was_dirty(o, rename->path);
1580                 char *add_name = unique_path(o, rename->path, other_branch);
1581                 if (update_file(o, 0, &add->oid, add->mode, add_name))
1582                         return -1;
1583
1584                 if (ren_src_was_dirty) {
1585                         output(o, 1, _("Refusing to lose dirty file at %s"),
1586                                rename->path);
1587                 }
1588                 /*
1589                  * Because the double negatives somehow keep confusing me...
1590                  *    1) update_wd iff !ren_src_was_dirty.
1591                  *    2) no_wd iff !update_wd
1592                  *    3) so, no_wd == !!ren_src_was_dirty == ren_src_was_dirty
1593                  */
1594                 remove_file(o, 0, rename->path, ren_src_was_dirty);
1595                 dst_name = unique_path(o, rename->path, cur_branch);
1596         } else {
1597                 if (dir_in_way(rename->path, !o->call_depth, 0)) {
1598                         dst_name = unique_path(o, rename->path, cur_branch);
1599                         output(o, 1, _("%s is a directory in %s adding as %s instead"),
1600                                rename->path, other_branch, dst_name);
1601                 } else if (!o->call_depth &&
1602                            would_lose_untracked(rename->path)) {
1603                         dst_name = unique_path(o, rename->path, cur_branch);
1604                         output(o, 1, _("Refusing to lose untracked file at %s; "
1605                                        "adding as %s instead"),
1606                                rename->path, dst_name);
1607                 }
1608         }
1609         if ((ret = update_file(o, 0, &rename->oid, rename->mode, dst_name)))
1610                 ; /* fall through, do allow dst_name to be released */
1611         else if (stage == 2)
1612                 ret = update_stages(o, rename->path, NULL, rename, add);
1613         else
1614                 ret = update_stages(o, rename->path, NULL, add, rename);
1615
1616         if (dst_name != rename->path)
1617                 free(dst_name);
1618
1619         return ret;
1620 }
1621
1622 static int conflict_rename_rename_1to2(struct merge_options *o,
1623                                         struct rename_conflict_info *ci)
1624 {
1625         /* One file was renamed in both branches, but to different names. */
1626         struct diff_filespec *one = ci->pair1->one;
1627         struct diff_filespec *a = ci->pair1->two;
1628         struct diff_filespec *b = ci->pair2->two;
1629
1630         output(o, 1, _("CONFLICT (rename/rename): "
1631                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1632                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1633                one->path, a->path, ci->branch1,
1634                one->path, b->path, ci->branch2,
1635                o->call_depth ? _(" (left unresolved)") : "");
1636         if (o->call_depth) {
1637                 struct merge_file_info mfi;
1638                 struct diff_filespec other;
1639                 struct diff_filespec *add;
1640                 if (merge_file_one(o, one->path,
1641                                  &one->oid, one->mode,
1642                                  &a->oid, a->mode,
1643                                  &b->oid, b->mode,
1644                                  ci->branch1, ci->branch2, &mfi))
1645                         return -1;
1646
1647                 /*
1648                  * FIXME: For rename/add-source conflicts (if we could detect
1649                  * such), this is wrong.  We should instead find a unique
1650                  * pathname and then either rename the add-source file to that
1651                  * unique path, or use that unique path instead of src here.
1652                  */
1653                 if (update_file(o, 0, &mfi.oid, mfi.mode, one->path))
1654                         return -1;
1655
1656                 /*
1657                  * Above, we put the merged content at the merge-base's
1658                  * path.  Now we usually need to delete both a->path and
1659                  * b->path.  However, the rename on each side of the merge
1660                  * could also be involved in a rename/add conflict.  In
1661                  * such cases, we should keep the added file around,
1662                  * resolving the conflict at that path in its favor.
1663                  */
1664                 add = filespec_from_entry(&other, ci->dst_entry1, 2 ^ 1);
1665                 if (add) {
1666                         if (update_file(o, 0, &add->oid, add->mode, a->path))
1667                                 return -1;
1668                 }
1669                 else
1670                         remove_file_from_cache(a->path);
1671                 add = filespec_from_entry(&other, ci->dst_entry2, 3 ^ 1);
1672                 if (add) {
1673                         if (update_file(o, 0, &add->oid, add->mode, b->path))
1674                                 return -1;
1675                 }
1676                 else
1677                         remove_file_from_cache(b->path);
1678         } else if (handle_file(o, a, 2, ci) || handle_file(o, b, 3, ci))
1679                 return -1;
1680
1681         return 0;
1682 }
1683
1684 static int conflict_rename_rename_2to1(struct merge_options *o,
1685                                         struct rename_conflict_info *ci)
1686 {
1687         /* Two files, a & b, were renamed to the same thing, c. */
1688         struct diff_filespec *a = ci->pair1->one;
1689         struct diff_filespec *b = ci->pair2->one;
1690         struct diff_filespec *c1 = ci->pair1->two;
1691         struct diff_filespec *c2 = ci->pair2->two;
1692         char *path = c1->path; /* == c2->path */
1693         char *path_side_1_desc;
1694         char *path_side_2_desc;
1695         struct merge_file_info mfi_c1;
1696         struct merge_file_info mfi_c2;
1697         int ret;
1698
1699         output(o, 1, _("CONFLICT (rename/rename): "
1700                "Rename %s->%s in %s. "
1701                "Rename %s->%s in %s"),
1702                a->path, c1->path, ci->branch1,
1703                b->path, c2->path, ci->branch2);
1704
1705         remove_file(o, 1, a->path, o->call_depth || would_lose_untracked(a->path));
1706         remove_file(o, 1, b->path, o->call_depth || would_lose_untracked(b->path));
1707
1708         path_side_1_desc = xstrfmt("%s (was %s)", path, a->path);
1709         path_side_2_desc = xstrfmt("%s (was %s)", path, b->path);
1710         if (merge_file_special_markers(o, a, c1, &ci->ren1_other,
1711                                        path_side_1_desc,
1712                                        o->branch1, c1->path,
1713                                        o->branch2, ci->ren1_other.path, &mfi_c1) ||
1714             merge_file_special_markers(o, b, &ci->ren2_other, c2,
1715                                        path_side_2_desc,
1716                                        o->branch1, ci->ren2_other.path,
1717                                        o->branch2, c2->path, &mfi_c2))
1718                 return -1;
1719         free(path_side_1_desc);
1720         free(path_side_2_desc);
1721
1722         if (o->call_depth) {
1723                 /*
1724                  * If mfi_c1.clean && mfi_c2.clean, then it might make
1725                  * sense to do a two-way merge of those results.  But, I
1726                  * think in all cases, it makes sense to have the virtual
1727                  * merge base just undo the renames; they can be detected
1728                  * again later for the non-recursive merge.
1729                  */
1730                 remove_file(o, 0, path, 0);
1731                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, a->path);
1732                 if (!ret)
1733                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1734                                           b->path);
1735         } else {
1736                 char *new_path1 = unique_path(o, path, ci->branch1);
1737                 char *new_path2 = unique_path(o, path, ci->branch2);
1738                 output(o, 1, _("Renaming %s to %s and %s to %s instead"),
1739                        a->path, new_path1, b->path, new_path2);
1740                 if (was_dirty(o, path))
1741                         output(o, 1, _("Refusing to lose dirty file at %s"),
1742                                path);
1743                 else if (would_lose_untracked(path))
1744                         /*
1745                          * Only way we get here is if both renames were from
1746                          * a directory rename AND user had an untracked file
1747                          * at the location where both files end up after the
1748                          * two directory renames.  See testcase 10d of t6043.
1749                          */
1750                         output(o, 1, _("Refusing to lose untracked file at "
1751                                        "%s, even though it's in the way."),
1752                                path);
1753                 else
1754                         remove_file(o, 0, path, 0);
1755                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, new_path1);
1756                 if (!ret)
1757                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1758                                           new_path2);
1759                 /*
1760                  * unpack_trees() actually populates the index for us for
1761                  * "normal" rename/rename(2to1) situtations so that the
1762                  * correct entries are at the higher stages, which would
1763                  * make the call below to update_stages_for_stage_data
1764                  * unnecessary.  However, if either of the renames came
1765                  * from a directory rename, then unpack_trees() will not
1766                  * have gotten the right data loaded into the index, so we
1767                  * need to do so now.  (While it'd be tempting to move this
1768                  * call to update_stages_for_stage_data() to
1769                  * apply_directory_rename_modifications(), that would break
1770                  * our intermediate calls to would_lose_untracked() since
1771                  * those rely on the current in-memory index.  See also the
1772                  * big "NOTE" in update_stages()).
1773                  */
1774                 if (update_stages_for_stage_data(o, path, ci->dst_entry1))
1775                         ret = -1;
1776
1777                 free(new_path2);
1778                 free(new_path1);
1779         }
1780
1781         return ret;
1782 }
1783
1784 /*
1785  * Get the diff_filepairs changed between o_tree and tree.
1786  */
1787 static struct diff_queue_struct *get_diffpairs(struct merge_options *o,
1788                                                struct tree *o_tree,
1789                                                struct tree *tree)
1790 {
1791         struct diff_queue_struct *ret;
1792         struct diff_options opts;
1793
1794         diff_setup(&opts);
1795         opts.flags.recursive = 1;
1796         opts.flags.rename_empty = 0;
1797         opts.detect_rename = merge_detect_rename(o);
1798         /*
1799          * We do not have logic to handle the detection of copies.  In
1800          * fact, it may not even make sense to add such logic: would we
1801          * really want a change to a base file to be propagated through
1802          * multiple other files by a merge?
1803          */
1804         if (opts.detect_rename > DIFF_DETECT_RENAME)
1805                 opts.detect_rename = DIFF_DETECT_RENAME;
1806         opts.rename_limit = o->merge_rename_limit >= 0 ? o->merge_rename_limit :
1807                             o->diff_rename_limit >= 0 ? o->diff_rename_limit :
1808                             1000;
1809         opts.rename_score = o->rename_score;
1810         opts.show_rename_progress = o->show_rename_progress;
1811         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1812         diff_setup_done(&opts);
1813         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1814         diffcore_std(&opts);
1815         if (opts.needed_rename_limit > o->needed_rename_limit)
1816                 o->needed_rename_limit = opts.needed_rename_limit;
1817
1818         ret = xmalloc(sizeof(*ret));
1819         *ret = diff_queued_diff;
1820
1821         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1822         diff_queued_diff.nr = 0;
1823         diff_queued_diff.queue = NULL;
1824         diff_flush(&opts);
1825         return ret;
1826 }
1827
1828 static int tree_has_path(struct tree *tree, const char *path)
1829 {
1830         struct object_id hashy;
1831         unsigned int mode_o;
1832
1833         return !get_tree_entry(&tree->object.oid, path,
1834                                &hashy, &mode_o);
1835 }
1836
1837 /*
1838  * Return a new string that replaces the beginning portion (which matches
1839  * entry->dir), with entry->new_dir.  In perl-speak:
1840  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1841  * NOTE:
1842  *   Caller must ensure that old_path starts with entry->dir + '/'.
1843  */
1844 static char *apply_dir_rename(struct dir_rename_entry *entry,
1845                               const char *old_path)
1846 {
1847         struct strbuf new_path = STRBUF_INIT;
1848         int oldlen, newlen;
1849
1850         if (entry->non_unique_new_dir)
1851                 return NULL;
1852
1853         oldlen = strlen(entry->dir);
1854         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1855         strbuf_grow(&new_path, newlen);
1856         strbuf_addbuf(&new_path, &entry->new_dir);
1857         strbuf_addstr(&new_path, &old_path[oldlen]);
1858
1859         return strbuf_detach(&new_path, NULL);
1860 }
1861
1862 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1863                                     char **old_dir, char **new_dir)
1864 {
1865         char *end_of_old, *end_of_new;
1866         int old_len, new_len;
1867
1868         *old_dir = NULL;
1869         *new_dir = NULL;
1870
1871         /*
1872          * For
1873          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1874          * the "e/foo.c" part is the same, we just want to know that
1875          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1876          * so, for this example, this function returns "a/b/c/d" in
1877          * *old_dir and "a/b/some/thing/else" in *new_dir.
1878          *
1879          * Also, if the basename of the file changed, we don't care.  We
1880          * want to know which portion of the directory, if any, changed.
1881          */
1882         end_of_old = strrchr(old_path, '/');
1883         end_of_new = strrchr(new_path, '/');
1884
1885         if (end_of_old == NULL || end_of_new == NULL)
1886                 return;
1887         while (*--end_of_new == *--end_of_old &&
1888                end_of_old != old_path &&
1889                end_of_new != new_path)
1890                 ; /* Do nothing; all in the while loop */
1891         /*
1892          * We've found the first non-matching character in the directory
1893          * paths.  That means the current directory we were comparing
1894          * represents the rename.  Move end_of_old and end_of_new back
1895          * to the full directory name.
1896          */
1897         if (*end_of_old == '/')
1898                 end_of_old++;
1899         if (*end_of_old != '/')
1900                 end_of_new++;
1901         end_of_old = strchr(end_of_old, '/');
1902         end_of_new = strchr(end_of_new, '/');
1903
1904         /*
1905          * It may have been the case that old_path and new_path were the same
1906          * directory all along.  Don't claim a rename if they're the same.
1907          */
1908         old_len = end_of_old - old_path;
1909         new_len = end_of_new - new_path;
1910
1911         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1912                 *old_dir = xstrndup(old_path, old_len);
1913                 *new_dir = xstrndup(new_path, new_len);
1914         }
1915 }
1916
1917 static void remove_hashmap_entries(struct hashmap *dir_renames,
1918                                    struct string_list *items_to_remove)
1919 {
1920         int i;
1921         struct dir_rename_entry *entry;
1922
1923         for (i = 0; i < items_to_remove->nr; i++) {
1924                 entry = items_to_remove->items[i].util;
1925                 hashmap_remove(dir_renames, entry, NULL);
1926         }
1927         string_list_clear(items_to_remove, 0);
1928 }
1929
1930 /*
1931  * See if there is a directory rename for path, and if there are any file
1932  * level conflicts for the renamed location.  If there is a rename and
1933  * there are no conflicts, return the new name.  Otherwise, return NULL.
1934  */
1935 static char *handle_path_level_conflicts(struct merge_options *o,
1936                                          const char *path,
1937                                          struct dir_rename_entry *entry,
1938                                          struct hashmap *collisions,
1939                                          struct tree *tree)
1940 {
1941         char *new_path = NULL;
1942         struct collision_entry *collision_ent;
1943         int clean = 1;
1944         struct strbuf collision_paths = STRBUF_INIT;
1945
1946         /*
1947          * entry has the mapping of old directory name to new directory name
1948          * that we want to apply to path.
1949          */
1950         new_path = apply_dir_rename(entry, path);
1951
1952         if (!new_path) {
1953                 /* This should only happen when entry->non_unique_new_dir set */
1954                 if (!entry->non_unique_new_dir)
1955                         BUG("entry->non_unqiue_dir not set and !new_path");
1956                 output(o, 1, _("CONFLICT (directory rename split): "
1957                                "Unclear where to place %s because directory "
1958                                "%s was renamed to multiple other directories, "
1959                                "with no destination getting a majority of the "
1960                                "files."),
1961                        path, entry->dir);
1962                 clean = 0;
1963                 return NULL;
1964         }
1965
1966         /*
1967          * The caller needs to have ensured that it has pre-populated
1968          * collisions with all paths that map to new_path.  Do a quick check
1969          * to ensure that's the case.
1970          */
1971         collision_ent = collision_find_entry(collisions, new_path);
1972         if (collision_ent == NULL)
1973                 BUG("collision_ent is NULL");
1974
1975         /*
1976          * Check for one-sided add/add/.../add conflicts, i.e.
1977          * where implicit renames from the other side doing
1978          * directory rename(s) can affect this side of history
1979          * to put multiple paths into the same location.  Warn
1980          * and bail on directory renames for such paths.
1981          */
1982         if (collision_ent->reported_already) {
1983                 clean = 0;
1984         } else if (tree_has_path(tree, new_path)) {
1985                 collision_ent->reported_already = 1;
1986                 strbuf_add_separated_string_list(&collision_paths, ", ",
1987                                                  &collision_ent->source_files);
1988                 output(o, 1, _("CONFLICT (implicit dir rename): Existing "
1989                                "file/dir at %s in the way of implicit "
1990                                "directory rename(s) putting the following "
1991                                "path(s) there: %s."),
1992                        new_path, collision_paths.buf);
1993                 clean = 0;
1994         } else if (collision_ent->source_files.nr > 1) {
1995                 collision_ent->reported_already = 1;
1996                 strbuf_add_separated_string_list(&collision_paths, ", ",
1997                                                  &collision_ent->source_files);
1998                 output(o, 1, _("CONFLICT (implicit dir rename): Cannot map "
1999                                "more than one path to %s; implicit directory "
2000                                "renames tried to put these paths there: %s"),
2001                        new_path, collision_paths.buf);
2002                 clean = 0;
2003         }
2004
2005         /* Free memory we no longer need */
2006         strbuf_release(&collision_paths);
2007         if (!clean && new_path) {
2008                 free(new_path);
2009                 return NULL;
2010         }
2011
2012         return new_path;
2013 }
2014
2015 /*
2016  * There are a couple things we want to do at the directory level:
2017  *   1. Check for both sides renaming to the same thing, in order to avoid
2018  *      implicit renaming of files that should be left in place.  (See
2019  *      testcase 6b in t6043 for details.)
2020  *   2. Prune directory renames if there are still files left in the
2021  *      the original directory.  These represent a partial directory rename,
2022  *      i.e. a rename where only some of the files within the directory
2023  *      were renamed elsewhere.  (Technically, this could be done earlier
2024  *      in get_directory_renames(), except that would prevent us from
2025  *      doing the previous check and thus failing testcase 6b.)
2026  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2027  *      In the future, we could potentially record this info as well and
2028  *      omit reporting rename/rename(1to2) conflicts for each path within
2029  *      the affected directories, thus cleaning up the merge output.
2030  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2031  *         directory level, because merging directories is fine.  If it
2032  *         causes conflicts for files within those merged directories, then
2033  *         that should be detected at the individual path level.
2034  */
2035 static void handle_directory_level_conflicts(struct merge_options *o,
2036                                              struct hashmap *dir_re_head,
2037                                              struct tree *head,
2038                                              struct hashmap *dir_re_merge,
2039                                              struct tree *merge)
2040 {
2041         struct hashmap_iter iter;
2042         struct dir_rename_entry *head_ent;
2043         struct dir_rename_entry *merge_ent;
2044
2045         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2046         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2047
2048         hashmap_iter_init(dir_re_head, &iter);
2049         while ((head_ent = hashmap_iter_next(&iter))) {
2050                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2051                 if (merge_ent &&
2052                     !head_ent->non_unique_new_dir &&
2053                     !merge_ent->non_unique_new_dir &&
2054                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2055                         /* 1. Renamed identically; remove it from both sides */
2056                         string_list_append(&remove_from_head,
2057                                            head_ent->dir)->util = head_ent;
2058                         strbuf_release(&head_ent->new_dir);
2059                         string_list_append(&remove_from_merge,
2060                                            merge_ent->dir)->util = merge_ent;
2061                         strbuf_release(&merge_ent->new_dir);
2062                 } else if (tree_has_path(head, head_ent->dir)) {
2063                         /* 2. This wasn't a directory rename after all */
2064                         string_list_append(&remove_from_head,
2065                                            head_ent->dir)->util = head_ent;
2066                         strbuf_release(&head_ent->new_dir);
2067                 }
2068         }
2069
2070         remove_hashmap_entries(dir_re_head, &remove_from_head);
2071         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2072
2073         hashmap_iter_init(dir_re_merge, &iter);
2074         while ((merge_ent = hashmap_iter_next(&iter))) {
2075                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2076                 if (tree_has_path(merge, merge_ent->dir)) {
2077                         /* 2. This wasn't a directory rename after all */
2078                         string_list_append(&remove_from_merge,
2079                                            merge_ent->dir)->util = merge_ent;
2080                 } else if (head_ent &&
2081                            !head_ent->non_unique_new_dir &&
2082                            !merge_ent->non_unique_new_dir) {
2083                         /* 3. rename/rename(1to2) */
2084                         /*
2085                          * We can assume it's not rename/rename(1to1) because
2086                          * that was case (1), already checked above.  So we
2087                          * know that head_ent->new_dir and merge_ent->new_dir
2088                          * are different strings.
2089                          */
2090                         output(o, 1, _("CONFLICT (rename/rename): "
2091                                        "Rename directory %s->%s in %s. "
2092                                        "Rename directory %s->%s in %s"),
2093                                head_ent->dir, head_ent->new_dir.buf, o->branch1,
2094                                head_ent->dir, merge_ent->new_dir.buf, o->branch2);
2095                         string_list_append(&remove_from_head,
2096                                            head_ent->dir)->util = head_ent;
2097                         strbuf_release(&head_ent->new_dir);
2098                         string_list_append(&remove_from_merge,
2099                                            merge_ent->dir)->util = merge_ent;
2100                         strbuf_release(&merge_ent->new_dir);
2101                 }
2102         }
2103
2104         remove_hashmap_entries(dir_re_head, &remove_from_head);
2105         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2106 }
2107
2108 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs,
2109                                              struct tree *tree)
2110 {
2111         struct hashmap *dir_renames;
2112         struct hashmap_iter iter;
2113         struct dir_rename_entry *entry;
2114         int i;
2115
2116         /*
2117          * Typically, we think of a directory rename as all files from a
2118          * certain directory being moved to a target directory.  However,
2119          * what if someone first moved two files from the original
2120          * directory in one commit, and then renamed the directory
2121          * somewhere else in a later commit?  At merge time, we just know
2122          * that files from the original directory went to two different
2123          * places, and that the bulk of them ended up in the same place.
2124          * We want each directory rename to represent where the bulk of the
2125          * files from that directory end up; this function exists to find
2126          * where the bulk of the files went.
2127          *
2128          * The first loop below simply iterates through the list of file
2129          * renames, finding out how often each directory rename pair
2130          * possibility occurs.
2131          */
2132         dir_renames = xmalloc(sizeof(*dir_renames));
2133         dir_rename_init(dir_renames);
2134         for (i = 0; i < pairs->nr; ++i) {
2135                 struct string_list_item *item;
2136                 int *count;
2137                 struct diff_filepair *pair = pairs->queue[i];
2138                 char *old_dir, *new_dir;
2139
2140                 /* File not part of directory rename if it wasn't renamed */
2141                 if (pair->status != 'R')
2142                         continue;
2143
2144                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2145                                         &old_dir,        &new_dir);
2146                 if (!old_dir)
2147                         /* Directory didn't change at all; ignore this one. */
2148                         continue;
2149
2150                 entry = dir_rename_find_entry(dir_renames, old_dir);
2151                 if (!entry) {
2152                         entry = xmalloc(sizeof(*entry));
2153                         dir_rename_entry_init(entry, old_dir);
2154                         hashmap_put(dir_renames, entry);
2155                 } else {
2156                         free(old_dir);
2157                 }
2158                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2159                 if (!item) {
2160                         item = string_list_insert(&entry->possible_new_dirs,
2161                                                   new_dir);
2162                         item->util = xcalloc(1, sizeof(int));
2163                 } else {
2164                         free(new_dir);
2165                 }
2166                 count = item->util;
2167                 *count += 1;
2168         }
2169
2170         /*
2171          * For each directory with files moved out of it, we find out which
2172          * target directory received the most files so we can declare it to
2173          * be the "winning" target location for the directory rename.  This
2174          * winner gets recorded in new_dir.  If there is no winner
2175          * (multiple target directories received the same number of files),
2176          * we set non_unique_new_dir.  Once we've determined the winner (or
2177          * that there is no winner), we no longer need possible_new_dirs.
2178          */
2179         hashmap_iter_init(dir_renames, &iter);
2180         while ((entry = hashmap_iter_next(&iter))) {
2181                 int max = 0;
2182                 int bad_max = 0;
2183                 char *best = NULL;
2184
2185                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2186                         int *count = entry->possible_new_dirs.items[i].util;
2187
2188                         if (*count == max)
2189                                 bad_max = max;
2190                         else if (*count > max) {
2191                                 max = *count;
2192                                 best = entry->possible_new_dirs.items[i].string;
2193                         }
2194                 }
2195                 if (bad_max == max)
2196                         entry->non_unique_new_dir = 1;
2197                 else {
2198                         assert(entry->new_dir.len == 0);
2199                         strbuf_addstr(&entry->new_dir, best);
2200                 }
2201                 /*
2202                  * The relevant directory sub-portion of the original full
2203                  * filepaths were xstrndup'ed before inserting into
2204                  * possible_new_dirs, and instead of manually iterating the
2205                  * list and free'ing each, just lie and tell
2206                  * possible_new_dirs that it did the strdup'ing so that it
2207                  * will free them for us.
2208                  */
2209                 entry->possible_new_dirs.strdup_strings = 1;
2210                 string_list_clear(&entry->possible_new_dirs, 1);
2211         }
2212
2213         return dir_renames;
2214 }
2215
2216 static struct dir_rename_entry *check_dir_renamed(const char *path,
2217                                                   struct hashmap *dir_renames)
2218 {
2219         char *temp = xstrdup(path);
2220         char *end;
2221         struct dir_rename_entry *entry = NULL;;
2222
2223         while ((end = strrchr(temp, '/'))) {
2224                 *end = '\0';
2225                 entry = dir_rename_find_entry(dir_renames, temp);
2226                 if (entry)
2227                         break;
2228         }
2229         free(temp);
2230         return entry;
2231 }
2232
2233 static void compute_collisions(struct hashmap *collisions,
2234                                struct hashmap *dir_renames,
2235                                struct diff_queue_struct *pairs)
2236 {
2237         int i;
2238
2239         /*
2240          * Multiple files can be mapped to the same path due to directory
2241          * renames done by the other side of history.  Since that other
2242          * side of history could have merged multiple directories into one,
2243          * if our side of history added the same file basename to each of
2244          * those directories, then all N of them would get implicitly
2245          * renamed by the directory rename detection into the same path,
2246          * and we'd get an add/add/.../add conflict, and all those adds
2247          * from *this* side of history.  This is not representable in the
2248          * index, and users aren't going to easily be able to make sense of
2249          * it.  So we need to provide a good warning about what's
2250          * happening, and fall back to no-directory-rename detection
2251          * behavior for those paths.
2252          *
2253          * See testcases 9e and all of section 5 from t6043 for examples.
2254          */
2255         collision_init(collisions);
2256
2257         for (i = 0; i < pairs->nr; ++i) {
2258                 struct dir_rename_entry *dir_rename_ent;
2259                 struct collision_entry *collision_ent;
2260                 char *new_path;
2261                 struct diff_filepair *pair = pairs->queue[i];
2262
2263                 if (pair->status != 'A' && pair->status != 'R')
2264                         continue;
2265                 dir_rename_ent = check_dir_renamed(pair->two->path,
2266                                                    dir_renames);
2267                 if (!dir_rename_ent)
2268                         continue;
2269
2270                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2271                 if (!new_path)
2272                         /*
2273                          * dir_rename_ent->non_unique_new_path is true, which
2274                          * means there is no directory rename for us to use,
2275                          * which means it won't cause us any additional
2276                          * collisions.
2277                          */
2278                         continue;
2279                 collision_ent = collision_find_entry(collisions, new_path);
2280                 if (!collision_ent) {
2281                         collision_ent = xcalloc(1,
2282                                                 sizeof(struct collision_entry));
2283                         hashmap_entry_init(collision_ent, strhash(new_path));
2284                         hashmap_put(collisions, collision_ent);
2285                         collision_ent->target_file = new_path;
2286                 } else {
2287                         free(new_path);
2288                 }
2289                 string_list_insert(&collision_ent->source_files,
2290                                    pair->two->path);
2291         }
2292 }
2293
2294 static char *check_for_directory_rename(struct merge_options *o,
2295                                         const char *path,
2296                                         struct tree *tree,
2297                                         struct hashmap *dir_renames,
2298                                         struct hashmap *dir_rename_exclusions,
2299                                         struct hashmap *collisions,
2300                                         int *clean_merge)
2301 {
2302         char *new_path = NULL;
2303         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2304         struct dir_rename_entry *oentry = NULL;
2305
2306         if (!entry)
2307                 return new_path;
2308
2309         /*
2310          * This next part is a little weird.  We do not want to do an
2311          * implicit rename into a directory we renamed on our side, because
2312          * that will result in a spurious rename/rename(1to2) conflict.  An
2313          * example:
2314          *   Base commit: dumbdir/afile, otherdir/bfile
2315          *   Side 1:      smrtdir/afile, otherdir/bfile
2316          *   Side 2:      dumbdir/afile, dumbdir/bfile
2317          * Here, while working on Side 1, we could notice that otherdir was
2318          * renamed/merged to dumbdir, and change the diff_filepair for
2319          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2320          * 2 will notice the rename from dumbdir to smrtdir, and do the
2321          * transitive rename to move it from dumbdir/bfile to
2322          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2323          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2324          * the file to end up in smrtdir.  And the way to achieve that is
2325          * to not let Side1 do the rename to dumbdir, since we know that is
2326          * the source of one of our directory renames.
2327          *
2328          * That's why oentry and dir_rename_exclusions is here.
2329          *
2330          * As it turns out, this also prevents N-way transient rename
2331          * confusion; See testcases 9c and 9d of t6043.
2332          */
2333         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2334         if (oentry) {
2335                 output(o, 1, _("WARNING: Avoiding applying %s -> %s rename "
2336                                "to %s, because %s itself was renamed."),
2337                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2338         } else {
2339                 new_path = handle_path_level_conflicts(o, path, entry,
2340                                                        collisions, tree);
2341                 *clean_merge &= (new_path != NULL);
2342         }
2343
2344         return new_path;
2345 }
2346
2347 static void apply_directory_rename_modifications(struct merge_options *o,
2348                                                  struct diff_filepair *pair,
2349                                                  char *new_path,
2350                                                  struct rename *re,
2351                                                  struct tree *tree,
2352                                                  struct tree *o_tree,
2353                                                  struct tree *a_tree,
2354                                                  struct tree *b_tree,
2355                                                  struct string_list *entries,
2356                                                  int *clean)
2357 {
2358         struct string_list_item *item;
2359         int stage = (tree == a_tree ? 2 : 3);
2360         int update_wd;
2361
2362         /*
2363          * In all cases where we can do directory rename detection,
2364          * unpack_trees() will have read pair->two->path into the
2365          * index and the working copy.  We need to remove it so that
2366          * we can instead place it at new_path.  It is guaranteed to
2367          * not be untracked (unpack_trees() would have errored out
2368          * saying the file would have been overwritten), but it might
2369          * be dirty, though.
2370          */
2371         update_wd = !was_dirty(o, pair->two->path);
2372         if (!update_wd)
2373                 output(o, 1, _("Refusing to lose dirty file at %s"),
2374                        pair->two->path);
2375         remove_file(o, 1, pair->two->path, !update_wd);
2376
2377         /* Find or create a new re->dst_entry */
2378         item = string_list_lookup(entries, new_path);
2379         if (item) {
2380                 /*
2381                  * Since we're renaming on this side of history, and it's
2382                  * due to a directory rename on the other side of history
2383                  * (which we only allow when the directory in question no
2384                  * longer exists on the other side of history), the
2385                  * original entry for re->dst_entry is no longer
2386                  * necessary...
2387                  */
2388                 re->dst_entry->processed = 1;
2389
2390                 /*
2391                  * ...because we'll be using this new one.
2392                  */
2393                 re->dst_entry = item->util;
2394         } else {
2395                 /*
2396                  * re->dst_entry is for the before-dir-rename path, and we
2397                  * need it to hold information for the after-dir-rename
2398                  * path.  Before creating a new entry, we need to mark the
2399                  * old one as unnecessary (...unless it is shared by
2400                  * src_entry, i.e. this didn't use to be a rename, in which
2401                  * case we can just allow the normal processing to happen
2402                  * for it).
2403                  */
2404                 if (pair->status == 'R')
2405                         re->dst_entry->processed = 1;
2406
2407                 re->dst_entry = insert_stage_data(new_path,
2408                                                   o_tree, a_tree, b_tree,
2409                                                   entries);
2410                 item = string_list_insert(entries, new_path);
2411                 item->util = re->dst_entry;
2412         }
2413
2414         /*
2415          * Update the stage_data with the information about the path we are
2416          * moving into place.  That slot will be empty and available for us
2417          * to write to because of the collision checks in
2418          * handle_path_level_conflicts().  In other words,
2419          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2420          * open for us to write to.
2421          *
2422          * It may be tempting to actually update the index at this point as
2423          * well, using update_stages_for_stage_data(), but as per the big
2424          * "NOTE" in update_stages(), doing so will modify the current
2425          * in-memory index which will break calls to would_lose_untracked()
2426          * that we need to make.  Instead, we need to just make sure that
2427          * the various conflict_rename_*() functions update the index
2428          * explicitly rather than relying on unpack_trees() to have done it.
2429          */
2430         get_tree_entry(&tree->object.oid,
2431                        pair->two->path,
2432                        &re->dst_entry->stages[stage].oid,
2433                        &re->dst_entry->stages[stage].mode);
2434
2435         /* Update pair status */
2436         if (pair->status == 'A') {
2437                 /*
2438                  * Recording rename information for this add makes it look
2439                  * like a rename/delete conflict.  Make sure we can
2440                  * correctly handle this as an add that was moved to a new
2441                  * directory instead of reporting a rename/delete conflict.
2442                  */
2443                 re->add_turned_into_rename = 1;
2444         }
2445         /*
2446          * We don't actually look at pair->status again, but it seems
2447          * pedagogically correct to adjust it.
2448          */
2449         pair->status = 'R';
2450
2451         /*
2452          * Finally, record the new location.
2453          */
2454         pair->two->path = new_path;
2455 }
2456
2457 /*
2458  * Get information of all renames which occurred in 'pairs', making use of
2459  * any implicit directory renames inferred from the other side of history.
2460  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2461  * to be able to associate the correct cache entries with the rename
2462  * information; tree is always equal to either a_tree or b_tree.
2463  */
2464 static struct string_list *get_renames(struct merge_options *o,
2465                                        struct diff_queue_struct *pairs,
2466                                        struct hashmap *dir_renames,
2467                                        struct hashmap *dir_rename_exclusions,
2468                                        struct tree *tree,
2469                                        struct tree *o_tree,
2470                                        struct tree *a_tree,
2471                                        struct tree *b_tree,
2472                                        struct string_list *entries,
2473                                        int *clean_merge)
2474 {
2475         int i;
2476         struct hashmap collisions;
2477         struct hashmap_iter iter;
2478         struct collision_entry *e;
2479         struct string_list *renames;
2480
2481         compute_collisions(&collisions, dir_renames, pairs);
2482         renames = xcalloc(1, sizeof(struct string_list));
2483
2484         for (i = 0; i < pairs->nr; ++i) {
2485                 struct string_list_item *item;
2486                 struct rename *re;
2487                 struct diff_filepair *pair = pairs->queue[i];
2488                 char *new_path; /* non-NULL only with directory renames */
2489
2490                 if (pair->status != 'A' && pair->status != 'R') {
2491                         diff_free_filepair(pair);
2492                         continue;
2493                 }
2494                 new_path = check_for_directory_rename(o, pair->two->path, tree,
2495                                                       dir_renames,
2496                                                       dir_rename_exclusions,
2497                                                       &collisions,
2498                                                       clean_merge);
2499                 if (pair->status != 'R' && !new_path) {
2500                         diff_free_filepair(pair);
2501                         continue;
2502                 }
2503
2504                 re = xmalloc(sizeof(*re));
2505                 re->processed = 0;
2506                 re->add_turned_into_rename = 0;
2507                 re->pair = pair;
2508                 item = string_list_lookup(entries, re->pair->one->path);
2509                 if (!item)
2510                         re->src_entry = insert_stage_data(re->pair->one->path,
2511                                         o_tree, a_tree, b_tree, entries);
2512                 else
2513                         re->src_entry = item->util;
2514
2515                 item = string_list_lookup(entries, re->pair->two->path);
2516                 if (!item)
2517                         re->dst_entry = insert_stage_data(re->pair->two->path,
2518                                         o_tree, a_tree, b_tree, entries);
2519                 else
2520                         re->dst_entry = item->util;
2521                 item = string_list_insert(renames, pair->one->path);
2522                 item->util = re;
2523                 if (new_path)
2524                         apply_directory_rename_modifications(o, pair, new_path,
2525                                                              re, tree, o_tree,
2526                                                              a_tree, b_tree,
2527                                                              entries,
2528                                                              clean_merge);
2529         }
2530
2531         hashmap_iter_init(&collisions, &iter);
2532         while ((e = hashmap_iter_next(&iter))) {
2533                 free(e->target_file);
2534                 string_list_clear(&e->source_files, 0);
2535         }
2536         hashmap_free(&collisions, 1);
2537         return renames;
2538 }
2539
2540 static int process_renames(struct merge_options *o,
2541                            struct string_list *a_renames,
2542                            struct string_list *b_renames)
2543 {
2544         int clean_merge = 1, i, j;
2545         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2546         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2547         const struct rename *sre;
2548
2549         for (i = 0; i < a_renames->nr; i++) {
2550                 sre = a_renames->items[i].util;
2551                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2552                         = (void *)sre;
2553         }
2554         for (i = 0; i < b_renames->nr; i++) {
2555                 sre = b_renames->items[i].util;
2556                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2557                         = (void *)sre;
2558         }
2559
2560         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2561                 struct string_list *renames1, *renames2Dst;
2562                 struct rename *ren1 = NULL, *ren2 = NULL;
2563                 const char *branch1, *branch2;
2564                 const char *ren1_src, *ren1_dst;
2565                 struct string_list_item *lookup;
2566
2567                 if (i >= a_renames->nr) {
2568                         ren2 = b_renames->items[j++].util;
2569                 } else if (j >= b_renames->nr) {
2570                         ren1 = a_renames->items[i++].util;
2571                 } else {
2572                         int compare = strcmp(a_renames->items[i].string,
2573                                              b_renames->items[j].string);
2574                         if (compare <= 0)
2575                                 ren1 = a_renames->items[i++].util;
2576                         if (compare >= 0)
2577                                 ren2 = b_renames->items[j++].util;
2578                 }
2579
2580                 /* TODO: refactor, so that 1/2 are not needed */
2581                 if (ren1) {
2582                         renames1 = a_renames;
2583                         renames2Dst = &b_by_dst;
2584                         branch1 = o->branch1;
2585                         branch2 = o->branch2;
2586                 } else {
2587                         renames1 = b_renames;
2588                         renames2Dst = &a_by_dst;
2589                         branch1 = o->branch2;
2590                         branch2 = o->branch1;
2591                         SWAP(ren2, ren1);
2592                 }
2593
2594                 if (ren1->processed)
2595                         continue;
2596                 ren1->processed = 1;
2597                 ren1->dst_entry->processed = 1;
2598                 /* BUG: We should only mark src_entry as processed if we
2599                  * are not dealing with a rename + add-source case.
2600                  */
2601                 ren1->src_entry->processed = 1;
2602
2603                 ren1_src = ren1->pair->one->path;
2604                 ren1_dst = ren1->pair->two->path;
2605
2606                 if (ren2) {
2607                         /* One file renamed on both sides */
2608                         const char *ren2_src = ren2->pair->one->path;
2609                         const char *ren2_dst = ren2->pair->two->path;
2610                         enum rename_type rename_type;
2611                         if (strcmp(ren1_src, ren2_src) != 0)
2612                                 BUG("ren1_src != ren2_src");
2613                         ren2->dst_entry->processed = 1;
2614                         ren2->processed = 1;
2615                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2616                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2617                                 clean_merge = 0;
2618                         } else {
2619                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2620                                 /* BUG: We should only remove ren1_src in
2621                                  * the base stage (think of rename +
2622                                  * add-source cases).
2623                                  */
2624                                 remove_file(o, 1, ren1_src, 1);
2625                                 update_entry(ren1->dst_entry,
2626                                              ren1->pair->one,
2627                                              ren1->pair->two,
2628                                              ren2->pair->two);
2629                         }
2630                         setup_rename_conflict_info(rename_type,
2631                                                    ren1->pair,
2632                                                    ren2->pair,
2633                                                    branch1,
2634                                                    branch2,
2635                                                    ren1->dst_entry,
2636                                                    ren2->dst_entry,
2637                                                    o,
2638                                                    NULL,
2639                                                    NULL);
2640                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2641                         /* Two different files renamed to the same thing */
2642                         char *ren2_dst;
2643                         ren2 = lookup->util;
2644                         ren2_dst = ren2->pair->two->path;
2645                         if (strcmp(ren1_dst, ren2_dst) != 0)
2646                                 BUG("ren1_dst != ren2_dst");
2647
2648                         clean_merge = 0;
2649                         ren2->processed = 1;
2650                         /*
2651                          * BUG: We should only mark src_entry as processed
2652                          * if we are not dealing with a rename + add-source
2653                          * case.
2654                          */
2655                         ren2->src_entry->processed = 1;
2656
2657                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2658                                                    ren1->pair,
2659                                                    ren2->pair,
2660                                                    branch1,
2661                                                    branch2,
2662                                                    ren1->dst_entry,
2663                                                    ren2->dst_entry,
2664                                                    o,
2665                                                    ren1->src_entry,
2666                                                    ren2->src_entry);
2667
2668                 } else {
2669                         /* Renamed in 1, maybe changed in 2 */
2670                         /* we only use sha1 and mode of these */
2671                         struct diff_filespec src_other, dst_other;
2672                         int try_merge;
2673
2674                         /*
2675                          * unpack_trees loads entries from common-commit
2676                          * into stage 1, from head-commit into stage 2, and
2677                          * from merge-commit into stage 3.  We keep track
2678                          * of which side corresponds to the rename.
2679                          */
2680                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2681                         int other_stage =   a_renames == renames1 ? 3 : 2;
2682
2683                         /* BUG: We should only remove ren1_src in the base
2684                          * stage and in other_stage (think of rename +
2685                          * add-source case).
2686                          */
2687                         remove_file(o, 1, ren1_src,
2688                                     renamed_stage == 2 || !was_tracked(o, ren1_src));
2689
2690                         oidcpy(&src_other.oid,
2691                                &ren1->src_entry->stages[other_stage].oid);
2692                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2693                         oidcpy(&dst_other.oid,
2694                                &ren1->dst_entry->stages[other_stage].oid);
2695                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2696                         try_merge = 0;
2697
2698                         if (oid_eq(&src_other.oid, &null_oid) &&
2699                             ren1->add_turned_into_rename) {
2700                                 setup_rename_conflict_info(RENAME_DIR,
2701                                                            ren1->pair,
2702                                                            NULL,
2703                                                            branch1,
2704                                                            branch2,
2705                                                            ren1->dst_entry,
2706                                                            NULL,
2707                                                            o,
2708                                                            NULL,
2709                                                            NULL);
2710                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2711                                 setup_rename_conflict_info(RENAME_DELETE,
2712                                                            ren1->pair,
2713                                                            NULL,
2714                                                            branch1,
2715                                                            branch2,
2716                                                            ren1->dst_entry,
2717                                                            NULL,
2718                                                            o,
2719                                                            NULL,
2720                                                            NULL);
2721                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2722                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2723                                 /*
2724                                  * Added file on the other side identical to
2725                                  * the file being renamed: clean merge.
2726                                  * Also, there is no need to overwrite the
2727                                  * file already in the working copy, so call
2728                                  * update_file_flags() instead of
2729                                  * update_file().
2730                                  */
2731                                 if (update_file_flags(o,
2732                                                       &ren1->pair->two->oid,
2733                                                       ren1->pair->two->mode,
2734                                                       ren1_dst,
2735                                                       1, /* update_cache */
2736                                                       0  /* update_wd    */))
2737                                         clean_merge = -1;
2738                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2739                                 clean_merge = 0;
2740                                 try_merge = 1;
2741                                 output(o, 1, _("CONFLICT (rename/add): Rename %s->%s in %s. "
2742                                        "%s added in %s"),
2743                                        ren1_src, ren1_dst, branch1,
2744                                        ren1_dst, branch2);
2745                                 if (o->call_depth) {
2746                                         struct merge_file_info mfi;
2747                                         if (merge_file_one(o, ren1_dst, &null_oid, 0,
2748                                                            &ren1->pair->two->oid,
2749                                                            ren1->pair->two->mode,
2750                                                            &dst_other.oid,
2751                                                            dst_other.mode,
2752                                                            branch1, branch2, &mfi)) {
2753                                                 clean_merge = -1;
2754                                                 goto cleanup_and_return;
2755                                         }
2756                                         output(o, 1, _("Adding merged %s"), ren1_dst);
2757                                         if (update_file(o, 0, &mfi.oid,
2758                                                         mfi.mode, ren1_dst))
2759                                                 clean_merge = -1;
2760                                         try_merge = 0;
2761                                 } else {
2762                                         char *new_path = unique_path(o, ren1_dst, branch2);
2763                                         output(o, 1, _("Adding as %s instead"), new_path);
2764                                         if (update_file(o, 0, &dst_other.oid,
2765                                                         dst_other.mode, new_path))
2766                                                 clean_merge = -1;
2767                                         free(new_path);
2768                                 }
2769                         } else
2770                                 try_merge = 1;
2771
2772                         if (clean_merge < 0)
2773                                 goto cleanup_and_return;
2774                         if (try_merge) {
2775                                 struct diff_filespec *one, *a, *b;
2776                                 src_other.path = (char *)ren1_src;
2777
2778                                 one = ren1->pair->one;
2779                                 if (a_renames == renames1) {
2780                                         a = ren1->pair->two;
2781                                         b = &src_other;
2782                                 } else {
2783                                         b = ren1->pair->two;
2784                                         a = &src_other;
2785                                 }
2786                                 update_entry(ren1->dst_entry, one, a, b);
2787                                 setup_rename_conflict_info(RENAME_NORMAL,
2788                                                            ren1->pair,
2789                                                            NULL,
2790                                                            branch1,
2791                                                            NULL,
2792                                                            ren1->dst_entry,
2793                                                            NULL,
2794                                                            o,
2795                                                            NULL,
2796                                                            NULL);
2797                         }
2798                 }
2799         }
2800 cleanup_and_return:
2801         string_list_clear(&a_by_dst, 0);
2802         string_list_clear(&b_by_dst, 0);
2803
2804         return clean_merge;
2805 }
2806
2807 struct rename_info {
2808         struct string_list *head_renames;
2809         struct string_list *merge_renames;
2810 };
2811
2812 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2813                                    struct hashmap *dir_renames)
2814 {
2815         struct hashmap_iter iter;
2816         struct dir_rename_entry *e;
2817
2818         hashmap_iter_init(dir_renames, &iter);
2819         while ((e = hashmap_iter_next(&iter))) {
2820                 free(e->dir);
2821                 strbuf_release(&e->new_dir);
2822                 /* possible_new_dirs already cleared in get_directory_renames */
2823         }
2824         hashmap_free(dir_renames, 1);
2825         free(dir_renames);
2826
2827         free(pairs->queue);
2828         free(pairs);
2829 }
2830
2831 static int handle_renames(struct merge_options *o,
2832                           struct tree *common,
2833                           struct tree *head,
2834                           struct tree *merge,
2835                           struct string_list *entries,
2836                           struct rename_info *ri)
2837 {
2838         struct diff_queue_struct *head_pairs, *merge_pairs;
2839         struct hashmap *dir_re_head, *dir_re_merge;
2840         int clean = 1;
2841
2842         ri->head_renames = NULL;
2843         ri->merge_renames = NULL;
2844
2845         if (!merge_detect_rename(o))
2846                 return 1;
2847
2848         head_pairs = get_diffpairs(o, common, head);
2849         merge_pairs = get_diffpairs(o, common, merge);
2850
2851         dir_re_head = get_directory_renames(head_pairs, head);
2852         dir_re_merge = get_directory_renames(merge_pairs, merge);
2853
2854         handle_directory_level_conflicts(o,
2855                                          dir_re_head, head,
2856                                          dir_re_merge, merge);
2857
2858         ri->head_renames  = get_renames(o, head_pairs,
2859                                         dir_re_merge, dir_re_head, head,
2860                                         common, head, merge, entries,
2861                                         &clean);
2862         if (clean < 0)
2863                 goto cleanup;
2864         ri->merge_renames = get_renames(o, merge_pairs,
2865                                         dir_re_head, dir_re_merge, merge,
2866                                         common, head, merge, entries,
2867                                         &clean);
2868         if (clean < 0)
2869                 goto cleanup;
2870         clean &= process_renames(o, ri->head_renames, ri->merge_renames);
2871
2872 cleanup:
2873         /*
2874          * Some cleanup is deferred until cleanup_renames() because the
2875          * data structures are still needed and referenced in
2876          * process_entry().  But there are a few things we can free now.
2877          */
2878         initial_cleanup_rename(head_pairs, dir_re_head);
2879         initial_cleanup_rename(merge_pairs, dir_re_merge);
2880
2881         return clean;
2882 }
2883
2884 static void final_cleanup_rename(struct string_list *rename)
2885 {
2886         const struct rename *re;
2887         int i;
2888
2889         if (rename == NULL)
2890                 return;
2891
2892         for (i = 0; i < rename->nr; i++) {
2893                 re = rename->items[i].util;
2894                 diff_free_filepair(re->pair);
2895         }
2896         string_list_clear(rename, 1);
2897         free(rename);
2898 }
2899
2900 static void final_cleanup_renames(struct rename_info *re_info)
2901 {
2902         final_cleanup_rename(re_info->head_renames);
2903         final_cleanup_rename(re_info->merge_renames);
2904 }
2905
2906 static struct object_id *stage_oid(const struct object_id *oid, unsigned mode)
2907 {
2908         return (is_null_oid(oid) || mode == 0) ? NULL: (struct object_id *)oid;
2909 }
2910
2911 static int read_oid_strbuf(struct merge_options *o,
2912         const struct object_id *oid, struct strbuf *dst)
2913 {
2914         void *buf;
2915         enum object_type type;
2916         unsigned long size;
2917         buf = read_object_file(oid, &type, &size);
2918         if (!buf)
2919                 return err(o, _("cannot read object %s"), oid_to_hex(oid));
2920         if (type != OBJ_BLOB) {
2921                 free(buf);
2922                 return err(o, _("object %s is not a blob"), oid_to_hex(oid));
2923         }
2924         strbuf_attach(dst, buf, size, size + 1);
2925         return 0;
2926 }
2927
2928 static int blob_unchanged(struct merge_options *opt,
2929                           const struct object_id *o_oid,
2930                           unsigned o_mode,
2931                           const struct object_id *a_oid,
2932                           unsigned a_mode,
2933                           int renormalize, const char *path)
2934 {
2935         struct strbuf o = STRBUF_INIT;
2936         struct strbuf a = STRBUF_INIT;
2937         int ret = 0; /* assume changed for safety */
2938
2939         if (a_mode != o_mode)
2940                 return 0;
2941         if (oid_eq(o_oid, a_oid))
2942                 return 1;
2943         if (!renormalize)
2944                 return 0;
2945
2946         assert(o_oid && a_oid);
2947         if (read_oid_strbuf(opt, o_oid, &o) || read_oid_strbuf(opt, a_oid, &a))
2948                 goto error_return;
2949         /*
2950          * Note: binary | is used so that both renormalizations are
2951          * performed.  Comparison can be skipped if both files are
2952          * unchanged since their sha1s have already been compared.
2953          */
2954         if (renormalize_buffer(&the_index, path, o.buf, o.len, &o) |
2955             renormalize_buffer(&the_index, path, a.buf, a.len, &a))
2956                 ret = (o.len == a.len && !memcmp(o.buf, a.buf, o.len));
2957
2958 error_return:
2959         strbuf_release(&o);
2960         strbuf_release(&a);
2961         return ret;
2962 }
2963
2964 static int handle_modify_delete(struct merge_options *o,
2965                                  const char *path,
2966                                  struct object_id *o_oid, int o_mode,
2967                                  struct object_id *a_oid, int a_mode,
2968                                  struct object_id *b_oid, int b_mode)
2969 {
2970         const char *modify_branch, *delete_branch;
2971         struct object_id *changed_oid;
2972         int changed_mode;
2973
2974         if (a_oid) {
2975                 modify_branch = o->branch1;
2976                 delete_branch = o->branch2;
2977                 changed_oid = a_oid;
2978                 changed_mode = a_mode;
2979         } else {
2980                 modify_branch = o->branch2;
2981                 delete_branch = o->branch1;
2982                 changed_oid = b_oid;
2983                 changed_mode = b_mode;
2984         }
2985
2986         return handle_change_delete(o,
2987                                     path, NULL,
2988                                     o_oid, o_mode,
2989                                     changed_oid, changed_mode,
2990                                     modify_branch, delete_branch,
2991                                     _("modify"), _("modified"));
2992 }
2993
2994 static int merge_content(struct merge_options *o,
2995                          const char *path,
2996                          int is_dirty,
2997                          struct object_id *o_oid, int o_mode,
2998                          struct object_id *a_oid, int a_mode,
2999                          struct object_id *b_oid, int b_mode,
3000                          struct rename_conflict_info *rename_conflict_info)
3001 {
3002         const char *reason = _("content");
3003         const char *path1 = NULL, *path2 = NULL;
3004         struct merge_file_info mfi;
3005         struct diff_filespec one, a, b;
3006         unsigned df_conflict_remains = 0;
3007
3008         if (!o_oid) {
3009                 reason = _("add/add");
3010                 o_oid = (struct object_id *)&null_oid;
3011         }
3012         one.path = a.path = b.path = (char *)path;
3013         oidcpy(&one.oid, o_oid);
3014         one.mode = o_mode;
3015         oidcpy(&a.oid, a_oid);
3016         a.mode = a_mode;
3017         oidcpy(&b.oid, b_oid);
3018         b.mode = b_mode;
3019
3020         if (rename_conflict_info) {
3021                 struct diff_filepair *pair1 = rename_conflict_info->pair1;
3022
3023                 path1 = (o->branch1 == rename_conflict_info->branch1) ?
3024                         pair1->two->path : pair1->one->path;
3025                 /* If rename_conflict_info->pair2 != NULL, we are in
3026                  * RENAME_ONE_FILE_TO_ONE case.  Otherwise, we have a
3027                  * normal rename.
3028                  */
3029                 path2 = (rename_conflict_info->pair2 ||
3030                          o->branch2 == rename_conflict_info->branch1) ?
3031                         pair1->two->path : pair1->one->path;
3032
3033                 if (dir_in_way(path, !o->call_depth,
3034                                S_ISGITLINK(pair1->two->mode)))
3035                         df_conflict_remains = 1;
3036         }
3037         if (merge_file_special_markers(o, &one, &a, &b, path,
3038                                        o->branch1, path1,
3039                                        o->branch2, path2, &mfi))
3040                 return -1;
3041
3042         /*
3043          * We can skip updating the working tree file iff:
3044          *   a) The merge is clean
3045          *   b) The merge matches what was in HEAD (content, mode, pathname)
3046          *   c) The target path is usable (i.e. not involved in D/F conflict)
3047          */
3048         if (mfi.clean &&
3049             was_tracked_and_matches(o, path, &mfi.oid, mfi.mode) &&
3050             !df_conflict_remains) {
3051                 output(o, 3, _("Skipped %s (merged same as existing)"), path);
3052                 if (add_cacheinfo(o, mfi.mode, &mfi.oid, path,
3053                                   0, (!o->call_depth && !is_dirty), 0))
3054                         return -1;
3055                 return mfi.clean;
3056         }
3057
3058         if (!mfi.clean) {
3059                 if (S_ISGITLINK(mfi.mode))
3060                         reason = _("submodule");
3061                 output(o, 1, _("CONFLICT (%s): Merge conflict in %s"),
3062                                 reason, path);
3063                 if (rename_conflict_info && !df_conflict_remains)
3064                         if (update_stages(o, path, &one, &a, &b))
3065                                 return -1;
3066         }
3067
3068         if (df_conflict_remains || is_dirty) {
3069                 char *new_path;
3070                 if (o->call_depth) {
3071                         remove_file_from_cache(path);
3072                 } else {
3073                         if (!mfi.clean) {
3074                                 if (update_stages(o, path, &one, &a, &b))
3075                                         return -1;
3076                         } else {
3077                                 int file_from_stage2 = was_tracked(o, path);
3078                                 struct diff_filespec merged;
3079                                 oidcpy(&merged.oid, &mfi.oid);
3080                                 merged.mode = mfi.mode;
3081
3082                                 if (update_stages(o, path, NULL,
3083                                                   file_from_stage2 ? &merged : NULL,
3084                                                   file_from_stage2 ? NULL : &merged))
3085                                         return -1;
3086                         }
3087
3088                 }
3089                 new_path = unique_path(o, path, rename_conflict_info->branch1);
3090                 if (is_dirty) {
3091                         output(o, 1, _("Refusing to lose dirty file at %s"),
3092                                path);
3093                 }
3094                 output(o, 1, _("Adding as %s instead"), new_path);
3095                 if (update_file(o, 0, &mfi.oid, mfi.mode, new_path)) {
3096                         free(new_path);
3097                         return -1;
3098                 }
3099                 free(new_path);
3100                 mfi.clean = 0;
3101         } else if (update_file(o, mfi.clean, &mfi.oid, mfi.mode, path))
3102                 return -1;
3103         return !is_dirty && mfi.clean;
3104 }
3105
3106 static int conflict_rename_normal(struct merge_options *o,
3107                                   const char *path,
3108                                   struct object_id *o_oid, unsigned int o_mode,
3109                                   struct object_id *a_oid, unsigned int a_mode,
3110                                   struct object_id *b_oid, unsigned int b_mode,
3111                                   struct rename_conflict_info *ci)
3112 {
3113         /* Merge the content and write it out */
3114         return merge_content(o, path, was_dirty(o, path),
3115                              o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3116                              ci);
3117 }
3118
3119 /* Per entry merge function */
3120 static int process_entry(struct merge_options *o,
3121                          const char *path, struct stage_data *entry)
3122 {
3123         int clean_merge = 1;
3124         int normalize = o->renormalize;
3125         unsigned o_mode = entry->stages[1].mode;
3126         unsigned a_mode = entry->stages[2].mode;
3127         unsigned b_mode = entry->stages[3].mode;
3128         struct object_id *o_oid = stage_oid(&entry->stages[1].oid, o_mode);
3129         struct object_id *a_oid = stage_oid(&entry->stages[2].oid, a_mode);
3130         struct object_id *b_oid = stage_oid(&entry->stages[3].oid, b_mode);
3131
3132         entry->processed = 1;
3133         if (entry->rename_conflict_info) {
3134                 struct rename_conflict_info *conflict_info = entry->rename_conflict_info;
3135                 switch (conflict_info->rename_type) {
3136                 case RENAME_NORMAL:
3137                 case RENAME_ONE_FILE_TO_ONE:
3138                         clean_merge = conflict_rename_normal(o,
3139                                                              path,
3140                                                              o_oid, o_mode,
3141                                                              a_oid, a_mode,
3142                                                              b_oid, b_mode,
3143                                                              conflict_info);
3144                         break;
3145                 case RENAME_DIR:
3146                         clean_merge = 1;
3147                         if (conflict_rename_dir(o,
3148                                                 conflict_info->pair1,
3149                                                 conflict_info->branch1,
3150                                                 conflict_info->branch2))
3151                                 clean_merge = -1;
3152                         break;
3153                 case RENAME_DELETE:
3154                         clean_merge = 0;
3155                         if (conflict_rename_delete(o,
3156                                                    conflict_info->pair1,
3157                                                    conflict_info->branch1,
3158                                                    conflict_info->branch2))
3159                                 clean_merge = -1;
3160                         break;
3161                 case RENAME_ONE_FILE_TO_TWO:
3162                         clean_merge = 0;
3163                         if (conflict_rename_rename_1to2(o, conflict_info))
3164                                 clean_merge = -1;
3165                         break;
3166                 case RENAME_TWO_FILES_TO_ONE:
3167                         clean_merge = 0;
3168                         if (conflict_rename_rename_2to1(o, conflict_info))
3169                                 clean_merge = -1;
3170                         break;
3171                 default:
3172                         entry->processed = 0;
3173                         break;
3174                 }
3175         } else if (o_oid && (!a_oid || !b_oid)) {
3176                 /* Case A: Deleted in one */
3177                 if ((!a_oid && !b_oid) ||
3178                     (!b_oid && blob_unchanged(o, o_oid, o_mode, a_oid, a_mode, normalize, path)) ||
3179                     (!a_oid && blob_unchanged(o, o_oid, o_mode, b_oid, b_mode, normalize, path))) {
3180                         /* Deleted in both or deleted in one and
3181                          * unchanged in the other */
3182                         if (a_oid)
3183                                 output(o, 2, _("Removing %s"), path);
3184                         /* do not touch working file if it did not exist */
3185                         remove_file(o, 1, path, !a_oid);
3186                 } else {
3187                         /* Modify/delete; deleted side may have put a directory in the way */
3188                         clean_merge = 0;
3189                         if (handle_modify_delete(o, path, o_oid, o_mode,
3190                                                  a_oid, a_mode, b_oid, b_mode))
3191                                 clean_merge = -1;
3192                 }
3193         } else if ((!o_oid && a_oid && !b_oid) ||
3194                    (!o_oid && !a_oid && b_oid)) {
3195                 /* Case B: Added in one. */
3196                 /* [nothing|directory] -> ([nothing|directory], file) */
3197
3198                 const char *add_branch;
3199                 const char *other_branch;
3200                 unsigned mode;
3201                 const struct object_id *oid;
3202                 const char *conf;
3203
3204                 if (a_oid) {
3205                         add_branch = o->branch1;
3206                         other_branch = o->branch2;
3207                         mode = a_mode;
3208                         oid = a_oid;
3209                         conf = _("file/directory");
3210                 } else {
3211                         add_branch = o->branch2;
3212                         other_branch = o->branch1;
3213                         mode = b_mode;
3214                         oid = b_oid;
3215                         conf = _("directory/file");
3216                 }
3217                 if (dir_in_way(path,
3218                                !o->call_depth && !S_ISGITLINK(a_mode),
3219                                0)) {
3220                         char *new_path = unique_path(o, path, add_branch);
3221                         clean_merge = 0;
3222                         output(o, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3223                                "Adding %s as %s"),
3224                                conf, path, other_branch, path, new_path);
3225                         if (update_file(o, 0, oid, mode, new_path))
3226                                 clean_merge = -1;
3227                         else if (o->call_depth)
3228                                 remove_file_from_cache(path);
3229                         free(new_path);
3230                 } else {
3231                         output(o, 2, _("Adding %s"), path);
3232                         /* do not overwrite file if already present */
3233                         if (update_file_flags(o, oid, mode, path, 1, !a_oid))
3234                                 clean_merge = -1;
3235                 }
3236         } else if (a_oid && b_oid) {
3237                 /* Case C: Added in both (check for same permissions) and */
3238                 /* case D: Modified in both, but differently. */
3239                 int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3240                 clean_merge = merge_content(o, path, is_dirty,
3241                                             o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3242                                             NULL);
3243         } else if (!o_oid && !a_oid && !b_oid) {
3244                 /*
3245                  * this entry was deleted altogether. a_mode == 0 means
3246                  * we had that path and want to actively remove it.
3247                  */
3248                 remove_file(o, 1, path, !a_mode);
3249         } else
3250                 BUG("fatal merge failure, shouldn't happen.");
3251
3252         return clean_merge;
3253 }
3254
3255 int merge_trees(struct merge_options *o,
3256                 struct tree *head,
3257                 struct tree *merge,
3258                 struct tree *common,
3259                 struct tree **result)
3260 {
3261         int code, clean;
3262
3263         if (o->subtree_shift) {
3264                 merge = shift_tree_object(head, merge, o->subtree_shift);
3265                 common = shift_tree_object(head, common, o->subtree_shift);
3266         }
3267
3268         if (oid_eq(&common->object.oid, &merge->object.oid)) {
3269                 struct strbuf sb = STRBUF_INIT;
3270
3271                 if (!o->call_depth && index_has_changes(&sb)) {
3272                         err(o, _("Dirty index: cannot merge (dirty: %s)"),
3273                             sb.buf);
3274                         return 0;
3275                 }
3276                 output(o, 0, _("Already up to date!"));
3277                 *result = head;
3278                 return 1;
3279         }
3280
3281         code = unpack_trees_start(o, common, head, merge);
3282
3283         if (code != 0) {
3284                 if (show(o, 4) || o->call_depth)
3285                         err(o, _("merging of trees %s and %s failed"),
3286                             oid_to_hex(&head->object.oid),
3287                             oid_to_hex(&merge->object.oid));
3288                 unpack_trees_finish(o);
3289                 return -1;
3290         }
3291
3292         if (unmerged_cache()) {
3293                 struct string_list *entries;
3294                 struct rename_info re_info;
3295                 int i;
3296                 /*
3297                  * Only need the hashmap while processing entries, so
3298                  * initialize it here and free it when we are done running
3299                  * through the entries. Keeping it in the merge_options as
3300                  * opposed to decaring a local hashmap is for convenience
3301                  * so that we don't have to pass it to around.
3302                  */
3303                 hashmap_init(&o->current_file_dir_set, path_hashmap_cmp, NULL, 512);
3304                 get_files_dirs(o, head);
3305                 get_files_dirs(o, merge);
3306
3307                 entries = get_unmerged();
3308                 clean = handle_renames(o, common, head, merge, entries,
3309                                        &re_info);
3310                 record_df_conflict_files(o, entries);
3311                 if (clean < 0)
3312                         goto cleanup;
3313                 for (i = entries->nr-1; 0 <= i; i--) {
3314                         const char *path = entries->items[i].string;
3315                         struct stage_data *e = entries->items[i].util;
3316                         if (!e->processed) {
3317                                 int ret = process_entry(o, path, e);
3318                                 if (!ret)
3319                                         clean = 0;
3320                                 else if (ret < 0) {
3321                                         clean = ret;
3322                                         goto cleanup;
3323                                 }
3324                         }
3325                 }
3326                 for (i = 0; i < entries->nr; i++) {
3327                         struct stage_data *e = entries->items[i].util;
3328                         if (!e->processed)
3329                                 BUG("unprocessed path??? %s",
3330                                     entries->items[i].string);
3331                 }
3332
3333 cleanup:
3334                 final_cleanup_renames(&re_info);
3335
3336                 string_list_clear(entries, 1);
3337                 free(entries);
3338
3339                 hashmap_free(&o->current_file_dir_set, 1);
3340
3341                 if (clean < 0) {
3342                         unpack_trees_finish(o);
3343                         return clean;
3344                 }
3345         }
3346         else
3347                 clean = 1;
3348
3349         unpack_trees_finish(o);
3350
3351         if (o->call_depth && !(*result = write_tree_from_memory(o)))
3352                 return -1;
3353
3354         return clean;
3355 }
3356
3357 static struct commit_list *reverse_commit_list(struct commit_list *list)
3358 {
3359         struct commit_list *next = NULL, *current, *backup;
3360         for (current = list; current; current = backup) {
3361                 backup = current->next;
3362                 current->next = next;
3363                 next = current;
3364         }
3365         return next;
3366 }
3367
3368 /*
3369  * Merge the commits h1 and h2, return the resulting virtual
3370  * commit object and a flag indicating the cleanness of the merge.
3371  */
3372 int merge_recursive(struct merge_options *o,
3373                     struct commit *h1,
3374                     struct commit *h2,
3375                     struct commit_list *ca,
3376                     struct commit **result)
3377 {
3378         struct commit_list *iter;
3379         struct commit *merged_common_ancestors;
3380         struct tree *mrtree;
3381         int clean;
3382
3383         if (show(o, 4)) {
3384                 output(o, 4, _("Merging:"));
3385                 output_commit_title(o, h1);
3386                 output_commit_title(o, h2);
3387         }
3388
3389         if (!ca) {
3390                 ca = get_merge_bases(h1, h2);
3391                 ca = reverse_commit_list(ca);
3392         }
3393
3394         if (show(o, 5)) {
3395                 unsigned cnt = commit_list_count(ca);
3396
3397                 output(o, 5, Q_("found %u common ancestor:",
3398                                 "found %u common ancestors:", cnt), cnt);
3399                 for (iter = ca; iter; iter = iter->next)
3400                         output_commit_title(o, iter->item);
3401         }
3402
3403         merged_common_ancestors = pop_commit(&ca);
3404         if (merged_common_ancestors == NULL) {
3405                 /* if there is no common ancestor, use an empty tree */
3406                 struct tree *tree;
3407
3408                 tree = lookup_tree(the_repository, the_repository->hash_algo->empty_tree);
3409                 merged_common_ancestors = make_virtual_commit(tree, "ancestor");
3410         }
3411
3412         for (iter = ca; iter; iter = iter->next) {
3413                 const char *saved_b1, *saved_b2;
3414                 o->call_depth++;
3415                 /*
3416                  * When the merge fails, the result contains files
3417                  * with conflict markers. The cleanness flag is
3418                  * ignored (unless indicating an error), it was never
3419                  * actually used, as result of merge_trees has always
3420                  * overwritten it: the committed "conflicts" were
3421                  * already resolved.
3422                  */
3423                 discard_cache();
3424                 saved_b1 = o->branch1;
3425                 saved_b2 = o->branch2;
3426                 o->branch1 = "Temporary merge branch 1";
3427                 o->branch2 = "Temporary merge branch 2";
3428                 if (merge_recursive(o, merged_common_ancestors, iter->item,
3429                                     NULL, &merged_common_ancestors) < 0)
3430                         return -1;
3431                 o->branch1 = saved_b1;
3432                 o->branch2 = saved_b2;
3433                 o->call_depth--;
3434
3435                 if (!merged_common_ancestors)
3436                         return err(o, _("merge returned no commit"));
3437         }
3438
3439         discard_cache();
3440         if (!o->call_depth)
3441                 read_cache();
3442
3443         o->ancestor = "merged common ancestors";
3444         clean = merge_trees(o, get_commit_tree(h1), get_commit_tree(h2),
3445                             get_commit_tree(merged_common_ancestors),
3446                             &mrtree);
3447         if (clean < 0) {
3448                 flush_output(o);
3449                 return clean;
3450         }
3451
3452         if (o->call_depth) {
3453                 *result = make_virtual_commit(mrtree, "merged tree");
3454                 commit_list_insert(h1, &(*result)->parents);
3455                 commit_list_insert(h2, &(*result)->parents->next);
3456         }
3457         flush_output(o);
3458         if (!o->call_depth && o->buffer_output < 2)
3459                 strbuf_release(&o->obuf);
3460         if (show(o, 2))
3461                 diff_warn_rename_limit("merge.renamelimit",
3462                                        o->needed_rename_limit, 0);
3463         return clean;
3464 }
3465
3466 static struct commit *get_ref(const struct object_id *oid, const char *name)
3467 {
3468         struct object *object;
3469
3470         object = deref_tag(the_repository, parse_object(the_repository, oid),
3471                            name,
3472                            strlen(name));
3473         if (!object)
3474                 return NULL;
3475         if (object->type == OBJ_TREE)
3476                 return make_virtual_commit((struct tree*)object, name);
3477         if (object->type != OBJ_COMMIT)
3478                 return NULL;
3479         if (parse_commit((struct commit *)object))
3480                 return NULL;
3481         return (struct commit *)object;
3482 }
3483
3484 int merge_recursive_generic(struct merge_options *o,
3485                             const struct object_id *head,
3486                             const struct object_id *merge,
3487                             int num_base_list,
3488                             const struct object_id **base_list,
3489                             struct commit **result)
3490 {
3491         int clean;
3492         struct lock_file lock = LOCK_INIT;
3493         struct commit *head_commit = get_ref(head, o->branch1);
3494         struct commit *next_commit = get_ref(merge, o->branch2);
3495         struct commit_list *ca = NULL;
3496
3497         if (base_list) {
3498                 int i;
3499                 for (i = 0; i < num_base_list; ++i) {
3500                         struct commit *base;
3501                         if (!(base = get_ref(base_list[i], oid_to_hex(base_list[i]))))
3502                                 return err(o, _("Could not parse object '%s'"),
3503                                         oid_to_hex(base_list[i]));
3504                         commit_list_insert(base, &ca);
3505                 }
3506         }
3507
3508         hold_locked_index(&lock, LOCK_DIE_ON_ERROR);
3509         clean = merge_recursive(o, head_commit, next_commit, ca,
3510                         result);
3511         if (clean < 0) {
3512                 rollback_lock_file(&lock);
3513                 return clean;
3514         }
3515
3516         if (write_locked_index(&the_index, &lock,
3517                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3518                 return err(o, _("Unable to write index."));
3519
3520         return clean ? 0 : 1;
3521 }
3522
3523 static void merge_recursive_config(struct merge_options *o)
3524 {
3525         char *value = NULL;
3526         git_config_get_int("merge.verbosity", &o->verbosity);
3527         git_config_get_int("diff.renamelimit", &o->diff_rename_limit);
3528         git_config_get_int("merge.renamelimit", &o->merge_rename_limit);
3529         if (!git_config_get_string("diff.renames", &value)) {
3530                 o->diff_detect_rename = git_config_rename("diff.renames", value);
3531                 free(value);
3532         }
3533         if (!git_config_get_string("merge.renames", &value)) {
3534                 o->merge_detect_rename = git_config_rename("merge.renames", value);
3535                 free(value);
3536         }
3537         git_config(git_xmerge_config, NULL);
3538 }
3539
3540 void init_merge_options(struct merge_options *o)
3541 {
3542         const char *merge_verbosity;
3543         memset(o, 0, sizeof(struct merge_options));
3544         o->verbosity = 2;
3545         o->buffer_output = 1;
3546         o->diff_rename_limit = -1;
3547         o->merge_rename_limit = -1;
3548         o->renormalize = 0;
3549         o->diff_detect_rename = -1;
3550         o->merge_detect_rename = -1;
3551         merge_recursive_config(o);
3552         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3553         if (merge_verbosity)
3554                 o->verbosity = strtol(merge_verbosity, NULL, 10);
3555         if (o->verbosity >= 5)
3556                 o->buffer_output = 0;
3557         strbuf_init(&o->obuf, 0);
3558         string_list_init(&o->df_conflict_file_set, 1);
3559 }
3560
3561 int parse_merge_opt(struct merge_options *o, const char *s)
3562 {
3563         const char *arg;
3564
3565         if (!s || !*s)
3566                 return -1;
3567         if (!strcmp(s, "ours"))
3568                 o->recursive_variant = MERGE_RECURSIVE_OURS;
3569         else if (!strcmp(s, "theirs"))
3570                 o->recursive_variant = MERGE_RECURSIVE_THEIRS;
3571         else if (!strcmp(s, "subtree"))
3572                 o->subtree_shift = "";
3573         else if (skip_prefix(s, "subtree=", &arg))
3574                 o->subtree_shift = arg;
3575         else if (!strcmp(s, "patience"))
3576                 o->xdl_opts = DIFF_WITH_ALG(o, PATIENCE_DIFF);
3577         else if (!strcmp(s, "histogram"))
3578                 o->xdl_opts = DIFF_WITH_ALG(o, HISTOGRAM_DIFF);
3579         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3580                 long value = parse_algorithm_value(arg);
3581                 if (value < 0)
3582                         return -1;
3583                 /* clear out previous settings */
3584                 DIFF_XDL_CLR(o, NEED_MINIMAL);
3585                 o->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3586                 o->xdl_opts |= value;
3587         }
3588         else if (!strcmp(s, "ignore-space-change"))
3589                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_CHANGE);
3590         else if (!strcmp(s, "ignore-all-space"))
3591                 DIFF_XDL_SET(o, IGNORE_WHITESPACE);
3592         else if (!strcmp(s, "ignore-space-at-eol"))
3593                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_AT_EOL);
3594         else if (!strcmp(s, "ignore-cr-at-eol"))
3595                 DIFF_XDL_SET(o, IGNORE_CR_AT_EOL);
3596         else if (!strcmp(s, "renormalize"))
3597                 o->renormalize = 1;
3598         else if (!strcmp(s, "no-renormalize"))
3599                 o->renormalize = 0;
3600         else if (!strcmp(s, "no-renames"))
3601                 o->merge_detect_rename = 0;
3602         else if (!strcmp(s, "find-renames")) {
3603                 o->merge_detect_rename = 1;
3604                 o->rename_score = 0;
3605         }
3606         else if (skip_prefix(s, "find-renames=", &arg) ||
3607                  skip_prefix(s, "rename-threshold=", &arg)) {
3608                 if ((o->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3609                         return -1;
3610                 o->merge_detect_rename = 1;
3611         }
3612         else
3613                 return -1;
3614         return 0;
3615 }