Merge branch 'js/pack-objects-mutex-init-fix' into maint
[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_VIA_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(&the_index, mode, oid ? oid : &null_oid, 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(&the_index, 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 exists 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(&the_index, 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 handle_rename_via_dir(struct merge_options *o,
1419                                  struct diff_filepair *pair,
1420                                  const char *rename_branch,
1421                                  const char *other_branch)
1422 {
1423         /*
1424          * Handle file adds that need to be renamed due to directory rename
1425          * detection.  This differs from handle_rename_normal, because
1426          * there is no content merge to do; just move the file into the
1427          * desired final location.
1428          */
1429         const struct diff_filespec *dest = pair->two;
1430
1431         if (!o->call_depth && would_lose_untracked(dest->path)) {
1432                 char *alt_path = unique_path(o, dest->path, rename_branch);
1433
1434                 output(o, 1, _("Error: Refusing to lose untracked file at %s; "
1435                                "writing to %s instead."),
1436                        dest->path, alt_path);
1437                 /*
1438                  * Write the file in worktree at alt_path, but not in the
1439                  * index.  Instead, write to dest->path for the index but
1440                  * only at the higher appropriate stage.
1441                  */
1442                 if (update_file(o, 0, &dest->oid, dest->mode, alt_path))
1443                         return -1;
1444                 free(alt_path);
1445                 return update_stages(o, dest->path, NULL,
1446                                      rename_branch == o->branch1 ? dest : NULL,
1447                                      rename_branch == o->branch1 ? NULL : dest);
1448         }
1449
1450         /* Update dest->path both in index and in worktree */
1451         if (update_file(o, 1, &dest->oid, dest->mode, dest->path))
1452                 return -1;
1453         return 0;
1454 }
1455
1456 static int handle_change_delete(struct merge_options *o,
1457                                 const char *path, const char *old_path,
1458                                 const struct object_id *o_oid, int o_mode,
1459                                 const struct object_id *changed_oid,
1460                                 int changed_mode,
1461                                 const char *change_branch,
1462                                 const char *delete_branch,
1463                                 const char *change, const char *change_past)
1464 {
1465         char *alt_path = NULL;
1466         const char *update_path = path;
1467         int ret = 0;
1468
1469         if (dir_in_way(path, !o->call_depth, 0) ||
1470             (!o->call_depth && would_lose_untracked(path))) {
1471                 update_path = alt_path = unique_path(o, path, change_branch);
1472         }
1473
1474         if (o->call_depth) {
1475                 /*
1476                  * We cannot arbitrarily accept either a_sha or b_sha as
1477                  * correct; since there is no true "middle point" between
1478                  * them, simply reuse the base version for virtual merge base.
1479                  */
1480                 ret = remove_file_from_cache(path);
1481                 if (!ret)
1482                         ret = update_file(o, 0, o_oid, o_mode, update_path);
1483         } else {
1484                 /*
1485                  * Despite the four nearly duplicate messages and argument
1486                  * lists below and the ugliness of the nested if-statements,
1487                  * having complete messages makes the job easier for
1488                  * translators.
1489                  *
1490                  * The slight variance among the cases is due to the fact
1491                  * that:
1492                  *   1) directory/file conflicts (in effect if
1493                  *      !alt_path) could cause us to need to write the
1494                  *      file to a different path.
1495                  *   2) renames (in effect if !old_path) could mean that
1496                  *      there are two names for the path that the user
1497                  *      may know the file by.
1498                  */
1499                 if (!alt_path) {
1500                         if (!old_path) {
1501                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1502                                        "and %s in %s. Version %s of %s left in tree."),
1503                                        change, path, delete_branch, change_past,
1504                                        change_branch, change_branch, path);
1505                         } else {
1506                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1507                                        "and %s to %s in %s. Version %s of %s left in tree."),
1508                                        change, old_path, delete_branch, change_past, path,
1509                                        change_branch, change_branch, path);
1510                         }
1511                 } else {
1512                         if (!old_path) {
1513                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1514                                        "and %s in %s. Version %s of %s left in tree at %s."),
1515                                        change, path, delete_branch, change_past,
1516                                        change_branch, change_branch, path, alt_path);
1517                         } else {
1518                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1519                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1520                                        change, old_path, delete_branch, change_past, path,
1521                                        change_branch, change_branch, path, alt_path);
1522                         }
1523                 }
1524                 /*
1525                  * No need to call update_file() on path when change_branch ==
1526                  * o->branch1 && !alt_path, since that would needlessly touch
1527                  * path.  We could call update_file_flags() with update_cache=0
1528                  * and update_wd=0, but that's a no-op.
1529                  */
1530                 if (change_branch != o->branch1 || alt_path)
1531                         ret = update_file(o, 0, changed_oid, changed_mode, update_path);
1532         }
1533         free(alt_path);
1534
1535         return ret;
1536 }
1537
1538 static int handle_rename_delete(struct merge_options *o,
1539                                 struct diff_filepair *pair,
1540                                 const char *rename_branch,
1541                                 const char *delete_branch)
1542 {
1543         const struct diff_filespec *orig = pair->one;
1544         const struct diff_filespec *dest = pair->two;
1545
1546         if (handle_change_delete(o,
1547                                  o->call_depth ? orig->path : dest->path,
1548                                  o->call_depth ? NULL : orig->path,
1549                                  &orig->oid, orig->mode,
1550                                  &dest->oid, dest->mode,
1551                                  rename_branch, delete_branch,
1552                                  _("rename"), _("renamed")))
1553                 return -1;
1554
1555         if (o->call_depth)
1556                 return remove_file_from_cache(dest->path);
1557         else
1558                 return update_stages(o, dest->path, NULL,
1559                                      rename_branch == o->branch1 ? dest : NULL,
1560                                      rename_branch == o->branch1 ? NULL : dest);
1561 }
1562
1563 static struct diff_filespec *filespec_from_entry(struct diff_filespec *target,
1564                                                  struct stage_data *entry,
1565                                                  int stage)
1566 {
1567         struct object_id *oid = &entry->stages[stage].oid;
1568         unsigned mode = entry->stages[stage].mode;
1569         if (mode == 0 || is_null_oid(oid))
1570                 return NULL;
1571         oidcpy(&target->oid, oid);
1572         target->mode = mode;
1573         return target;
1574 }
1575
1576 static int handle_file(struct merge_options *o,
1577                         struct diff_filespec *rename,
1578                         int stage,
1579                         struct rename_conflict_info *ci)
1580 {
1581         char *dst_name = rename->path;
1582         struct stage_data *dst_entry;
1583         const char *cur_branch, *other_branch;
1584         struct diff_filespec other;
1585         struct diff_filespec *add;
1586         int ret;
1587
1588         if (stage == 2) {
1589                 dst_entry = ci->dst_entry1;
1590                 cur_branch = ci->branch1;
1591                 other_branch = ci->branch2;
1592         } else {
1593                 dst_entry = ci->dst_entry2;
1594                 cur_branch = ci->branch2;
1595                 other_branch = ci->branch1;
1596         }
1597
1598         add = filespec_from_entry(&other, dst_entry, stage ^ 1);
1599         if (add) {
1600                 int ren_src_was_dirty = was_dirty(o, rename->path);
1601                 char *add_name = unique_path(o, rename->path, other_branch);
1602                 if (update_file(o, 0, &add->oid, add->mode, add_name))
1603                         return -1;
1604
1605                 if (ren_src_was_dirty) {
1606                         output(o, 1, _("Refusing to lose dirty file at %s"),
1607                                rename->path);
1608                 }
1609                 /*
1610                  * Because the double negatives somehow keep confusing me...
1611                  *    1) update_wd iff !ren_src_was_dirty.
1612                  *    2) no_wd iff !update_wd
1613                  *    3) so, no_wd == !!ren_src_was_dirty == ren_src_was_dirty
1614                  */
1615                 remove_file(o, 0, rename->path, ren_src_was_dirty);
1616                 dst_name = unique_path(o, rename->path, cur_branch);
1617         } else {
1618                 if (dir_in_way(rename->path, !o->call_depth, 0)) {
1619                         dst_name = unique_path(o, rename->path, cur_branch);
1620                         output(o, 1, _("%s is a directory in %s adding as %s instead"),
1621                                rename->path, other_branch, dst_name);
1622                 } else if (!o->call_depth &&
1623                            would_lose_untracked(rename->path)) {
1624                         dst_name = unique_path(o, rename->path, cur_branch);
1625                         output(o, 1, _("Refusing to lose untracked file at %s; "
1626                                        "adding as %s instead"),
1627                                rename->path, dst_name);
1628                 }
1629         }
1630         if ((ret = update_file(o, 0, &rename->oid, rename->mode, dst_name)))
1631                 ; /* fall through, do allow dst_name to be released */
1632         else if (stage == 2)
1633                 ret = update_stages(o, rename->path, NULL, rename, add);
1634         else
1635                 ret = update_stages(o, rename->path, NULL, add, rename);
1636
1637         if (dst_name != rename->path)
1638                 free(dst_name);
1639
1640         return ret;
1641 }
1642
1643 static int handle_rename_rename_1to2(struct merge_options *o,
1644                                      struct rename_conflict_info *ci)
1645 {
1646         /* One file was renamed in both branches, but to different names. */
1647         struct diff_filespec *one = ci->pair1->one;
1648         struct diff_filespec *a = ci->pair1->two;
1649         struct diff_filespec *b = ci->pair2->two;
1650
1651         output(o, 1, _("CONFLICT (rename/rename): "
1652                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1653                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1654                one->path, a->path, ci->branch1,
1655                one->path, b->path, ci->branch2,
1656                o->call_depth ? _(" (left unresolved)") : "");
1657         if (o->call_depth) {
1658                 struct merge_file_info mfi;
1659                 struct diff_filespec other;
1660                 struct diff_filespec *add;
1661                 if (merge_file_one(o, one->path,
1662                                  &one->oid, one->mode,
1663                                  &a->oid, a->mode,
1664                                  &b->oid, b->mode,
1665                                  ci->branch1, ci->branch2, &mfi))
1666                         return -1;
1667
1668                 /*
1669                  * FIXME: For rename/add-source conflicts (if we could detect
1670                  * such), this is wrong.  We should instead find a unique
1671                  * pathname and then either rename the add-source file to that
1672                  * unique path, or use that unique path instead of src here.
1673                  */
1674                 if (update_file(o, 0, &mfi.oid, mfi.mode, one->path))
1675                         return -1;
1676
1677                 /*
1678                  * Above, we put the merged content at the merge-base's
1679                  * path.  Now we usually need to delete both a->path and
1680                  * b->path.  However, the rename on each side of the merge
1681                  * could also be involved in a rename/add conflict.  In
1682                  * such cases, we should keep the added file around,
1683                  * resolving the conflict at that path in its favor.
1684                  */
1685                 add = filespec_from_entry(&other, ci->dst_entry1, 2 ^ 1);
1686                 if (add) {
1687                         if (update_file(o, 0, &add->oid, add->mode, a->path))
1688                                 return -1;
1689                 }
1690                 else
1691                         remove_file_from_cache(a->path);
1692                 add = filespec_from_entry(&other, ci->dst_entry2, 3 ^ 1);
1693                 if (add) {
1694                         if (update_file(o, 0, &add->oid, add->mode, b->path))
1695                                 return -1;
1696                 }
1697                 else
1698                         remove_file_from_cache(b->path);
1699         } else if (handle_file(o, a, 2, ci) || handle_file(o, b, 3, ci))
1700                 return -1;
1701
1702         return 0;
1703 }
1704
1705 static int handle_rename_rename_2to1(struct merge_options *o,
1706                                      struct rename_conflict_info *ci)
1707 {
1708         /* Two files, a & b, were renamed to the same thing, c. */
1709         struct diff_filespec *a = ci->pair1->one;
1710         struct diff_filespec *b = ci->pair2->one;
1711         struct diff_filespec *c1 = ci->pair1->two;
1712         struct diff_filespec *c2 = ci->pair2->two;
1713         char *path = c1->path; /* == c2->path */
1714         char *path_side_1_desc;
1715         char *path_side_2_desc;
1716         struct merge_file_info mfi_c1;
1717         struct merge_file_info mfi_c2;
1718         int ret;
1719
1720         output(o, 1, _("CONFLICT (rename/rename): "
1721                "Rename %s->%s in %s. "
1722                "Rename %s->%s in %s"),
1723                a->path, c1->path, ci->branch1,
1724                b->path, c2->path, ci->branch2);
1725
1726         remove_file(o, 1, a->path, o->call_depth || would_lose_untracked(a->path));
1727         remove_file(o, 1, b->path, o->call_depth || would_lose_untracked(b->path));
1728
1729         path_side_1_desc = xstrfmt("%s (was %s)", path, a->path);
1730         path_side_2_desc = xstrfmt("%s (was %s)", path, b->path);
1731         if (merge_file_special_markers(o, a, c1, &ci->ren1_other,
1732                                        path_side_1_desc,
1733                                        o->branch1, c1->path,
1734                                        o->branch2, ci->ren1_other.path, &mfi_c1) ||
1735             merge_file_special_markers(o, b, &ci->ren2_other, c2,
1736                                        path_side_2_desc,
1737                                        o->branch1, ci->ren2_other.path,
1738                                        o->branch2, c2->path, &mfi_c2))
1739                 return -1;
1740         free(path_side_1_desc);
1741         free(path_side_2_desc);
1742
1743         if (o->call_depth) {
1744                 /*
1745                  * If mfi_c1.clean && mfi_c2.clean, then it might make
1746                  * sense to do a two-way merge of those results.  But, I
1747                  * think in all cases, it makes sense to have the virtual
1748                  * merge base just undo the renames; they can be detected
1749                  * again later for the non-recursive merge.
1750                  */
1751                 remove_file(o, 0, path, 0);
1752                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, a->path);
1753                 if (!ret)
1754                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1755                                           b->path);
1756         } else {
1757                 char *new_path1 = unique_path(o, path, ci->branch1);
1758                 char *new_path2 = unique_path(o, path, ci->branch2);
1759                 output(o, 1, _("Renaming %s to %s and %s to %s instead"),
1760                        a->path, new_path1, b->path, new_path2);
1761                 if (was_dirty(o, path))
1762                         output(o, 1, _("Refusing to lose dirty file at %s"),
1763                                path);
1764                 else if (would_lose_untracked(path))
1765                         /*
1766                          * Only way we get here is if both renames were from
1767                          * a directory rename AND user had an untracked file
1768                          * at the location where both files end up after the
1769                          * two directory renames.  See testcase 10d of t6043.
1770                          */
1771                         output(o, 1, _("Refusing to lose untracked file at "
1772                                        "%s, even though it's in the way."),
1773                                path);
1774                 else
1775                         remove_file(o, 0, path, 0);
1776                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, new_path1);
1777                 if (!ret)
1778                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1779                                           new_path2);
1780                 /*
1781                  * unpack_trees() actually populates the index for us for
1782                  * "normal" rename/rename(2to1) situtations so that the
1783                  * correct entries are at the higher stages, which would
1784                  * make the call below to update_stages_for_stage_data
1785                  * unnecessary.  However, if either of the renames came
1786                  * from a directory rename, then unpack_trees() will not
1787                  * have gotten the right data loaded into the index, so we
1788                  * need to do so now.  (While it'd be tempting to move this
1789                  * call to update_stages_for_stage_data() to
1790                  * apply_directory_rename_modifications(), that would break
1791                  * our intermediate calls to would_lose_untracked() since
1792                  * those rely on the current in-memory index.  See also the
1793                  * big "NOTE" in update_stages()).
1794                  */
1795                 if (update_stages_for_stage_data(o, path, ci->dst_entry1))
1796                         ret = -1;
1797
1798                 free(new_path2);
1799                 free(new_path1);
1800         }
1801
1802         return ret;
1803 }
1804
1805 /*
1806  * Get the diff_filepairs changed between o_tree and tree.
1807  */
1808 static struct diff_queue_struct *get_diffpairs(struct merge_options *o,
1809                                                struct tree *o_tree,
1810                                                struct tree *tree)
1811 {
1812         struct diff_queue_struct *ret;
1813         struct diff_options opts;
1814
1815         diff_setup(&opts);
1816         opts.flags.recursive = 1;
1817         opts.flags.rename_empty = 0;
1818         opts.detect_rename = merge_detect_rename(o);
1819         /*
1820          * We do not have logic to handle the detection of copies.  In
1821          * fact, it may not even make sense to add such logic: would we
1822          * really want a change to a base file to be propagated through
1823          * multiple other files by a merge?
1824          */
1825         if (opts.detect_rename > DIFF_DETECT_RENAME)
1826                 opts.detect_rename = DIFF_DETECT_RENAME;
1827         opts.rename_limit = o->merge_rename_limit >= 0 ? o->merge_rename_limit :
1828                             o->diff_rename_limit >= 0 ? o->diff_rename_limit :
1829                             1000;
1830         opts.rename_score = o->rename_score;
1831         opts.show_rename_progress = o->show_rename_progress;
1832         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1833         diff_setup_done(&opts);
1834         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1835         diffcore_std(&opts);
1836         if (opts.needed_rename_limit > o->needed_rename_limit)
1837                 o->needed_rename_limit = opts.needed_rename_limit;
1838
1839         ret = xmalloc(sizeof(*ret));
1840         *ret = diff_queued_diff;
1841
1842         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1843         diff_queued_diff.nr = 0;
1844         diff_queued_diff.queue = NULL;
1845         diff_flush(&opts);
1846         return ret;
1847 }
1848
1849 static int tree_has_path(struct tree *tree, const char *path)
1850 {
1851         struct object_id hashy;
1852         unsigned int mode_o;
1853
1854         return !get_tree_entry(&tree->object.oid, path,
1855                                &hashy, &mode_o);
1856 }
1857
1858 /*
1859  * Return a new string that replaces the beginning portion (which matches
1860  * entry->dir), with entry->new_dir.  In perl-speak:
1861  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1862  * NOTE:
1863  *   Caller must ensure that old_path starts with entry->dir + '/'.
1864  */
1865 static char *apply_dir_rename(struct dir_rename_entry *entry,
1866                               const char *old_path)
1867 {
1868         struct strbuf new_path = STRBUF_INIT;
1869         int oldlen, newlen;
1870
1871         if (entry->non_unique_new_dir)
1872                 return NULL;
1873
1874         oldlen = strlen(entry->dir);
1875         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1876         strbuf_grow(&new_path, newlen);
1877         strbuf_addbuf(&new_path, &entry->new_dir);
1878         strbuf_addstr(&new_path, &old_path[oldlen]);
1879
1880         return strbuf_detach(&new_path, NULL);
1881 }
1882
1883 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1884                                     char **old_dir, char **new_dir)
1885 {
1886         char *end_of_old, *end_of_new;
1887         int old_len, new_len;
1888
1889         *old_dir = NULL;
1890         *new_dir = NULL;
1891
1892         /*
1893          * For
1894          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1895          * the "e/foo.c" part is the same, we just want to know that
1896          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1897          * so, for this example, this function returns "a/b/c/d" in
1898          * *old_dir and "a/b/some/thing/else" in *new_dir.
1899          *
1900          * Also, if the basename of the file changed, we don't care.  We
1901          * want to know which portion of the directory, if any, changed.
1902          */
1903         end_of_old = strrchr(old_path, '/');
1904         end_of_new = strrchr(new_path, '/');
1905
1906         if (end_of_old == NULL || end_of_new == NULL)
1907                 return;
1908         while (*--end_of_new == *--end_of_old &&
1909                end_of_old != old_path &&
1910                end_of_new != new_path)
1911                 ; /* Do nothing; all in the while loop */
1912         /*
1913          * We've found the first non-matching character in the directory
1914          * paths.  That means the current directory we were comparing
1915          * represents the rename.  Move end_of_old and end_of_new back
1916          * to the full directory name.
1917          */
1918         if (*end_of_old == '/')
1919                 end_of_old++;
1920         if (*end_of_old != '/')
1921                 end_of_new++;
1922         end_of_old = strchr(end_of_old, '/');
1923         end_of_new = strchr(end_of_new, '/');
1924
1925         /*
1926          * It may have been the case that old_path and new_path were the same
1927          * directory all along.  Don't claim a rename if they're the same.
1928          */
1929         old_len = end_of_old - old_path;
1930         new_len = end_of_new - new_path;
1931
1932         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1933                 *old_dir = xstrndup(old_path, old_len);
1934                 *new_dir = xstrndup(new_path, new_len);
1935         }
1936 }
1937
1938 static void remove_hashmap_entries(struct hashmap *dir_renames,
1939                                    struct string_list *items_to_remove)
1940 {
1941         int i;
1942         struct dir_rename_entry *entry;
1943
1944         for (i = 0; i < items_to_remove->nr; i++) {
1945                 entry = items_to_remove->items[i].util;
1946                 hashmap_remove(dir_renames, entry, NULL);
1947         }
1948         string_list_clear(items_to_remove, 0);
1949 }
1950
1951 /*
1952  * See if there is a directory rename for path, and if there are any file
1953  * level conflicts for the renamed location.  If there is a rename and
1954  * there are no conflicts, return the new name.  Otherwise, return NULL.
1955  */
1956 static char *handle_path_level_conflicts(struct merge_options *o,
1957                                          const char *path,
1958                                          struct dir_rename_entry *entry,
1959                                          struct hashmap *collisions,
1960                                          struct tree *tree)
1961 {
1962         char *new_path = NULL;
1963         struct collision_entry *collision_ent;
1964         int clean = 1;
1965         struct strbuf collision_paths = STRBUF_INIT;
1966
1967         /*
1968          * entry has the mapping of old directory name to new directory name
1969          * that we want to apply to path.
1970          */
1971         new_path = apply_dir_rename(entry, path);
1972
1973         if (!new_path) {
1974                 /* This should only happen when entry->non_unique_new_dir set */
1975                 if (!entry->non_unique_new_dir)
1976                         BUG("entry->non_unqiue_dir not set and !new_path");
1977                 output(o, 1, _("CONFLICT (directory rename split): "
1978                                "Unclear where to place %s because directory "
1979                                "%s was renamed to multiple other directories, "
1980                                "with no destination getting a majority of the "
1981                                "files."),
1982                        path, entry->dir);
1983                 clean = 0;
1984                 return NULL;
1985         }
1986
1987         /*
1988          * The caller needs to have ensured that it has pre-populated
1989          * collisions with all paths that map to new_path.  Do a quick check
1990          * to ensure that's the case.
1991          */
1992         collision_ent = collision_find_entry(collisions, new_path);
1993         if (collision_ent == NULL)
1994                 BUG("collision_ent is NULL");
1995
1996         /*
1997          * Check for one-sided add/add/.../add conflicts, i.e.
1998          * where implicit renames from the other side doing
1999          * directory rename(s) can affect this side of history
2000          * to put multiple paths into the same location.  Warn
2001          * and bail on directory renames for such paths.
2002          */
2003         if (collision_ent->reported_already) {
2004                 clean = 0;
2005         } else if (tree_has_path(tree, new_path)) {
2006                 collision_ent->reported_already = 1;
2007                 strbuf_add_separated_string_list(&collision_paths, ", ",
2008                                                  &collision_ent->source_files);
2009                 output(o, 1, _("CONFLICT (implicit dir rename): Existing "
2010                                "file/dir at %s in the way of implicit "
2011                                "directory rename(s) putting the following "
2012                                "path(s) there: %s."),
2013                        new_path, collision_paths.buf);
2014                 clean = 0;
2015         } else if (collision_ent->source_files.nr > 1) {
2016                 collision_ent->reported_already = 1;
2017                 strbuf_add_separated_string_list(&collision_paths, ", ",
2018                                                  &collision_ent->source_files);
2019                 output(o, 1, _("CONFLICT (implicit dir rename): Cannot map "
2020                                "more than one path to %s; implicit directory "
2021                                "renames tried to put these paths there: %s"),
2022                        new_path, collision_paths.buf);
2023                 clean = 0;
2024         }
2025
2026         /* Free memory we no longer need */
2027         strbuf_release(&collision_paths);
2028         if (!clean && new_path) {
2029                 free(new_path);
2030                 return NULL;
2031         }
2032
2033         return new_path;
2034 }
2035
2036 /*
2037  * There are a couple things we want to do at the directory level:
2038  *   1. Check for both sides renaming to the same thing, in order to avoid
2039  *      implicit renaming of files that should be left in place.  (See
2040  *      testcase 6b in t6043 for details.)
2041  *   2. Prune directory renames if there are still files left in the
2042  *      the original directory.  These represent a partial directory rename,
2043  *      i.e. a rename where only some of the files within the directory
2044  *      were renamed elsewhere.  (Technically, this could be done earlier
2045  *      in get_directory_renames(), except that would prevent us from
2046  *      doing the previous check and thus failing testcase 6b.)
2047  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2048  *      In the future, we could potentially record this info as well and
2049  *      omit reporting rename/rename(1to2) conflicts for each path within
2050  *      the affected directories, thus cleaning up the merge output.
2051  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2052  *         directory level, because merging directories is fine.  If it
2053  *         causes conflicts for files within those merged directories, then
2054  *         that should be detected at the individual path level.
2055  */
2056 static void handle_directory_level_conflicts(struct merge_options *o,
2057                                              struct hashmap *dir_re_head,
2058                                              struct tree *head,
2059                                              struct hashmap *dir_re_merge,
2060                                              struct tree *merge)
2061 {
2062         struct hashmap_iter iter;
2063         struct dir_rename_entry *head_ent;
2064         struct dir_rename_entry *merge_ent;
2065
2066         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2067         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2068
2069         hashmap_iter_init(dir_re_head, &iter);
2070         while ((head_ent = hashmap_iter_next(&iter))) {
2071                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2072                 if (merge_ent &&
2073                     !head_ent->non_unique_new_dir &&
2074                     !merge_ent->non_unique_new_dir &&
2075                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2076                         /* 1. Renamed identically; remove it from both sides */
2077                         string_list_append(&remove_from_head,
2078                                            head_ent->dir)->util = head_ent;
2079                         strbuf_release(&head_ent->new_dir);
2080                         string_list_append(&remove_from_merge,
2081                                            merge_ent->dir)->util = merge_ent;
2082                         strbuf_release(&merge_ent->new_dir);
2083                 } else if (tree_has_path(head, head_ent->dir)) {
2084                         /* 2. This wasn't a directory rename after all */
2085                         string_list_append(&remove_from_head,
2086                                            head_ent->dir)->util = head_ent;
2087                         strbuf_release(&head_ent->new_dir);
2088                 }
2089         }
2090
2091         remove_hashmap_entries(dir_re_head, &remove_from_head);
2092         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2093
2094         hashmap_iter_init(dir_re_merge, &iter);
2095         while ((merge_ent = hashmap_iter_next(&iter))) {
2096                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2097                 if (tree_has_path(merge, merge_ent->dir)) {
2098                         /* 2. This wasn't a directory rename after all */
2099                         string_list_append(&remove_from_merge,
2100                                            merge_ent->dir)->util = merge_ent;
2101                 } else if (head_ent &&
2102                            !head_ent->non_unique_new_dir &&
2103                            !merge_ent->non_unique_new_dir) {
2104                         /* 3. rename/rename(1to2) */
2105                         /*
2106                          * We can assume it's not rename/rename(1to1) because
2107                          * that was case (1), already checked above.  So we
2108                          * know that head_ent->new_dir and merge_ent->new_dir
2109                          * are different strings.
2110                          */
2111                         output(o, 1, _("CONFLICT (rename/rename): "
2112                                        "Rename directory %s->%s in %s. "
2113                                        "Rename directory %s->%s in %s"),
2114                                head_ent->dir, head_ent->new_dir.buf, o->branch1,
2115                                head_ent->dir, merge_ent->new_dir.buf, o->branch2);
2116                         string_list_append(&remove_from_head,
2117                                            head_ent->dir)->util = head_ent;
2118                         strbuf_release(&head_ent->new_dir);
2119                         string_list_append(&remove_from_merge,
2120                                            merge_ent->dir)->util = merge_ent;
2121                         strbuf_release(&merge_ent->new_dir);
2122                 }
2123         }
2124
2125         remove_hashmap_entries(dir_re_head, &remove_from_head);
2126         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2127 }
2128
2129 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs,
2130                                              struct tree *tree)
2131 {
2132         struct hashmap *dir_renames;
2133         struct hashmap_iter iter;
2134         struct dir_rename_entry *entry;
2135         int i;
2136
2137         /*
2138          * Typically, we think of a directory rename as all files from a
2139          * certain directory being moved to a target directory.  However,
2140          * what if someone first moved two files from the original
2141          * directory in one commit, and then renamed the directory
2142          * somewhere else in a later commit?  At merge time, we just know
2143          * that files from the original directory went to two different
2144          * places, and that the bulk of them ended up in the same place.
2145          * We want each directory rename to represent where the bulk of the
2146          * files from that directory end up; this function exists to find
2147          * where the bulk of the files went.
2148          *
2149          * The first loop below simply iterates through the list of file
2150          * renames, finding out how often each directory rename pair
2151          * possibility occurs.
2152          */
2153         dir_renames = xmalloc(sizeof(*dir_renames));
2154         dir_rename_init(dir_renames);
2155         for (i = 0; i < pairs->nr; ++i) {
2156                 struct string_list_item *item;
2157                 int *count;
2158                 struct diff_filepair *pair = pairs->queue[i];
2159                 char *old_dir, *new_dir;
2160
2161                 /* File not part of directory rename if it wasn't renamed */
2162                 if (pair->status != 'R')
2163                         continue;
2164
2165                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2166                                         &old_dir,        &new_dir);
2167                 if (!old_dir)
2168                         /* Directory didn't change at all; ignore this one. */
2169                         continue;
2170
2171                 entry = dir_rename_find_entry(dir_renames, old_dir);
2172                 if (!entry) {
2173                         entry = xmalloc(sizeof(*entry));
2174                         dir_rename_entry_init(entry, old_dir);
2175                         hashmap_put(dir_renames, entry);
2176                 } else {
2177                         free(old_dir);
2178                 }
2179                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2180                 if (!item) {
2181                         item = string_list_insert(&entry->possible_new_dirs,
2182                                                   new_dir);
2183                         item->util = xcalloc(1, sizeof(int));
2184                 } else {
2185                         free(new_dir);
2186                 }
2187                 count = item->util;
2188                 *count += 1;
2189         }
2190
2191         /*
2192          * For each directory with files moved out of it, we find out which
2193          * target directory received the most files so we can declare it to
2194          * be the "winning" target location for the directory rename.  This
2195          * winner gets recorded in new_dir.  If there is no winner
2196          * (multiple target directories received the same number of files),
2197          * we set non_unique_new_dir.  Once we've determined the winner (or
2198          * that there is no winner), we no longer need possible_new_dirs.
2199          */
2200         hashmap_iter_init(dir_renames, &iter);
2201         while ((entry = hashmap_iter_next(&iter))) {
2202                 int max = 0;
2203                 int bad_max = 0;
2204                 char *best = NULL;
2205
2206                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2207                         int *count = entry->possible_new_dirs.items[i].util;
2208
2209                         if (*count == max)
2210                                 bad_max = max;
2211                         else if (*count > max) {
2212                                 max = *count;
2213                                 best = entry->possible_new_dirs.items[i].string;
2214                         }
2215                 }
2216                 if (bad_max == max)
2217                         entry->non_unique_new_dir = 1;
2218                 else {
2219                         assert(entry->new_dir.len == 0);
2220                         strbuf_addstr(&entry->new_dir, best);
2221                 }
2222                 /*
2223                  * The relevant directory sub-portion of the original full
2224                  * filepaths were xstrndup'ed before inserting into
2225                  * possible_new_dirs, and instead of manually iterating the
2226                  * list and free'ing each, just lie and tell
2227                  * possible_new_dirs that it did the strdup'ing so that it
2228                  * will free them for us.
2229                  */
2230                 entry->possible_new_dirs.strdup_strings = 1;
2231                 string_list_clear(&entry->possible_new_dirs, 1);
2232         }
2233
2234         return dir_renames;
2235 }
2236
2237 static struct dir_rename_entry *check_dir_renamed(const char *path,
2238                                                   struct hashmap *dir_renames)
2239 {
2240         char *temp = xstrdup(path);
2241         char *end;
2242         struct dir_rename_entry *entry = NULL;;
2243
2244         while ((end = strrchr(temp, '/'))) {
2245                 *end = '\0';
2246                 entry = dir_rename_find_entry(dir_renames, temp);
2247                 if (entry)
2248                         break;
2249         }
2250         free(temp);
2251         return entry;
2252 }
2253
2254 static void compute_collisions(struct hashmap *collisions,
2255                                struct hashmap *dir_renames,
2256                                struct diff_queue_struct *pairs)
2257 {
2258         int i;
2259
2260         /*
2261          * Multiple files can be mapped to the same path due to directory
2262          * renames done by the other side of history.  Since that other
2263          * side of history could have merged multiple directories into one,
2264          * if our side of history added the same file basename to each of
2265          * those directories, then all N of them would get implicitly
2266          * renamed by the directory rename detection into the same path,
2267          * and we'd get an add/add/.../add conflict, and all those adds
2268          * from *this* side of history.  This is not representable in the
2269          * index, and users aren't going to easily be able to make sense of
2270          * it.  So we need to provide a good warning about what's
2271          * happening, and fall back to no-directory-rename detection
2272          * behavior for those paths.
2273          *
2274          * See testcases 9e and all of section 5 from t6043 for examples.
2275          */
2276         collision_init(collisions);
2277
2278         for (i = 0; i < pairs->nr; ++i) {
2279                 struct dir_rename_entry *dir_rename_ent;
2280                 struct collision_entry *collision_ent;
2281                 char *new_path;
2282                 struct diff_filepair *pair = pairs->queue[i];
2283
2284                 if (pair->status != 'A' && pair->status != 'R')
2285                         continue;
2286                 dir_rename_ent = check_dir_renamed(pair->two->path,
2287                                                    dir_renames);
2288                 if (!dir_rename_ent)
2289                         continue;
2290
2291                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2292                 if (!new_path)
2293                         /*
2294                          * dir_rename_ent->non_unique_new_path is true, which
2295                          * means there is no directory rename for us to use,
2296                          * which means it won't cause us any additional
2297                          * collisions.
2298                          */
2299                         continue;
2300                 collision_ent = collision_find_entry(collisions, new_path);
2301                 if (!collision_ent) {
2302                         collision_ent = xcalloc(1,
2303                                                 sizeof(struct collision_entry));
2304                         hashmap_entry_init(collision_ent, strhash(new_path));
2305                         hashmap_put(collisions, collision_ent);
2306                         collision_ent->target_file = new_path;
2307                 } else {
2308                         free(new_path);
2309                 }
2310                 string_list_insert(&collision_ent->source_files,
2311                                    pair->two->path);
2312         }
2313 }
2314
2315 static char *check_for_directory_rename(struct merge_options *o,
2316                                         const char *path,
2317                                         struct tree *tree,
2318                                         struct hashmap *dir_renames,
2319                                         struct hashmap *dir_rename_exclusions,
2320                                         struct hashmap *collisions,
2321                                         int *clean_merge)
2322 {
2323         char *new_path = NULL;
2324         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2325         struct dir_rename_entry *oentry = NULL;
2326
2327         if (!entry)
2328                 return new_path;
2329
2330         /*
2331          * This next part is a little weird.  We do not want to do an
2332          * implicit rename into a directory we renamed on our side, because
2333          * that will result in a spurious rename/rename(1to2) conflict.  An
2334          * example:
2335          *   Base commit: dumbdir/afile, otherdir/bfile
2336          *   Side 1:      smrtdir/afile, otherdir/bfile
2337          *   Side 2:      dumbdir/afile, dumbdir/bfile
2338          * Here, while working on Side 1, we could notice that otherdir was
2339          * renamed/merged to dumbdir, and change the diff_filepair for
2340          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2341          * 2 will notice the rename from dumbdir to smrtdir, and do the
2342          * transitive rename to move it from dumbdir/bfile to
2343          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2344          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2345          * the file to end up in smrtdir.  And the way to achieve that is
2346          * to not let Side1 do the rename to dumbdir, since we know that is
2347          * the source of one of our directory renames.
2348          *
2349          * That's why oentry and dir_rename_exclusions is here.
2350          *
2351          * As it turns out, this also prevents N-way transient rename
2352          * confusion; See testcases 9c and 9d of t6043.
2353          */
2354         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2355         if (oentry) {
2356                 output(o, 1, _("WARNING: Avoiding applying %s -> %s rename "
2357                                "to %s, because %s itself was renamed."),
2358                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2359         } else {
2360                 new_path = handle_path_level_conflicts(o, path, entry,
2361                                                        collisions, tree);
2362                 *clean_merge &= (new_path != NULL);
2363         }
2364
2365         return new_path;
2366 }
2367
2368 static void apply_directory_rename_modifications(struct merge_options *o,
2369                                                  struct diff_filepair *pair,
2370                                                  char *new_path,
2371                                                  struct rename *re,
2372                                                  struct tree *tree,
2373                                                  struct tree *o_tree,
2374                                                  struct tree *a_tree,
2375                                                  struct tree *b_tree,
2376                                                  struct string_list *entries,
2377                                                  int *clean)
2378 {
2379         struct string_list_item *item;
2380         int stage = (tree == a_tree ? 2 : 3);
2381         int update_wd;
2382
2383         /*
2384          * In all cases where we can do directory rename detection,
2385          * unpack_trees() will have read pair->two->path into the
2386          * index and the working copy.  We need to remove it so that
2387          * we can instead place it at new_path.  It is guaranteed to
2388          * not be untracked (unpack_trees() would have errored out
2389          * saying the file would have been overwritten), but it might
2390          * be dirty, though.
2391          */
2392         update_wd = !was_dirty(o, pair->two->path);
2393         if (!update_wd)
2394                 output(o, 1, _("Refusing to lose dirty file at %s"),
2395                        pair->two->path);
2396         remove_file(o, 1, pair->two->path, !update_wd);
2397
2398         /* Find or create a new re->dst_entry */
2399         item = string_list_lookup(entries, new_path);
2400         if (item) {
2401                 /*
2402                  * Since we're renaming on this side of history, and it's
2403                  * due to a directory rename on the other side of history
2404                  * (which we only allow when the directory in question no
2405                  * longer exists on the other side of history), the
2406                  * original entry for re->dst_entry is no longer
2407                  * necessary...
2408                  */
2409                 re->dst_entry->processed = 1;
2410
2411                 /*
2412                  * ...because we'll be using this new one.
2413                  */
2414                 re->dst_entry = item->util;
2415         } else {
2416                 /*
2417                  * re->dst_entry is for the before-dir-rename path, and we
2418                  * need it to hold information for the after-dir-rename
2419                  * path.  Before creating a new entry, we need to mark the
2420                  * old one as unnecessary (...unless it is shared by
2421                  * src_entry, i.e. this didn't use to be a rename, in which
2422                  * case we can just allow the normal processing to happen
2423                  * for it).
2424                  */
2425                 if (pair->status == 'R')
2426                         re->dst_entry->processed = 1;
2427
2428                 re->dst_entry = insert_stage_data(new_path,
2429                                                   o_tree, a_tree, b_tree,
2430                                                   entries);
2431                 item = string_list_insert(entries, new_path);
2432                 item->util = re->dst_entry;
2433         }
2434
2435         /*
2436          * Update the stage_data with the information about the path we are
2437          * moving into place.  That slot will be empty and available for us
2438          * to write to because of the collision checks in
2439          * handle_path_level_conflicts().  In other words,
2440          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2441          * open for us to write to.
2442          *
2443          * It may be tempting to actually update the index at this point as
2444          * well, using update_stages_for_stage_data(), but as per the big
2445          * "NOTE" in update_stages(), doing so will modify the current
2446          * in-memory index which will break calls to would_lose_untracked()
2447          * that we need to make.  Instead, we need to just make sure that
2448          * the various handle_rename_*() functions update the index
2449          * explicitly rather than relying on unpack_trees() to have done it.
2450          */
2451         get_tree_entry(&tree->object.oid,
2452                        pair->two->path,
2453                        &re->dst_entry->stages[stage].oid,
2454                        &re->dst_entry->stages[stage].mode);
2455
2456         /* Update pair status */
2457         if (pair->status == 'A') {
2458                 /*
2459                  * Recording rename information for this add makes it look
2460                  * like a rename/delete conflict.  Make sure we can
2461                  * correctly handle this as an add that was moved to a new
2462                  * directory instead of reporting a rename/delete conflict.
2463                  */
2464                 re->add_turned_into_rename = 1;
2465         }
2466         /*
2467          * We don't actually look at pair->status again, but it seems
2468          * pedagogically correct to adjust it.
2469          */
2470         pair->status = 'R';
2471
2472         /*
2473          * Finally, record the new location.
2474          */
2475         pair->two->path = new_path;
2476 }
2477
2478 /*
2479  * Get information of all renames which occurred in 'pairs', making use of
2480  * any implicit directory renames inferred from the other side of history.
2481  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2482  * to be able to associate the correct cache entries with the rename
2483  * information; tree is always equal to either a_tree or b_tree.
2484  */
2485 static struct string_list *get_renames(struct merge_options *o,
2486                                        struct diff_queue_struct *pairs,
2487                                        struct hashmap *dir_renames,
2488                                        struct hashmap *dir_rename_exclusions,
2489                                        struct tree *tree,
2490                                        struct tree *o_tree,
2491                                        struct tree *a_tree,
2492                                        struct tree *b_tree,
2493                                        struct string_list *entries,
2494                                        int *clean_merge)
2495 {
2496         int i;
2497         struct hashmap collisions;
2498         struct hashmap_iter iter;
2499         struct collision_entry *e;
2500         struct string_list *renames;
2501
2502         compute_collisions(&collisions, dir_renames, pairs);
2503         renames = xcalloc(1, sizeof(struct string_list));
2504
2505         for (i = 0; i < pairs->nr; ++i) {
2506                 struct string_list_item *item;
2507                 struct rename *re;
2508                 struct diff_filepair *pair = pairs->queue[i];
2509                 char *new_path; /* non-NULL only with directory renames */
2510
2511                 if (pair->status != 'A' && pair->status != 'R') {
2512                         diff_free_filepair(pair);
2513                         continue;
2514                 }
2515                 new_path = check_for_directory_rename(o, pair->two->path, tree,
2516                                                       dir_renames,
2517                                                       dir_rename_exclusions,
2518                                                       &collisions,
2519                                                       clean_merge);
2520                 if (pair->status != 'R' && !new_path) {
2521                         diff_free_filepair(pair);
2522                         continue;
2523                 }
2524
2525                 re = xmalloc(sizeof(*re));
2526                 re->processed = 0;
2527                 re->add_turned_into_rename = 0;
2528                 re->pair = pair;
2529                 item = string_list_lookup(entries, re->pair->one->path);
2530                 if (!item)
2531                         re->src_entry = insert_stage_data(re->pair->one->path,
2532                                         o_tree, a_tree, b_tree, entries);
2533                 else
2534                         re->src_entry = item->util;
2535
2536                 item = string_list_lookup(entries, re->pair->two->path);
2537                 if (!item)
2538                         re->dst_entry = insert_stage_data(re->pair->two->path,
2539                                         o_tree, a_tree, b_tree, entries);
2540                 else
2541                         re->dst_entry = item->util;
2542                 item = string_list_insert(renames, pair->one->path);
2543                 item->util = re;
2544                 if (new_path)
2545                         apply_directory_rename_modifications(o, pair, new_path,
2546                                                              re, tree, o_tree,
2547                                                              a_tree, b_tree,
2548                                                              entries,
2549                                                              clean_merge);
2550         }
2551
2552         hashmap_iter_init(&collisions, &iter);
2553         while ((e = hashmap_iter_next(&iter))) {
2554                 free(e->target_file);
2555                 string_list_clear(&e->source_files, 0);
2556         }
2557         hashmap_free(&collisions, 1);
2558         return renames;
2559 }
2560
2561 static int process_renames(struct merge_options *o,
2562                            struct string_list *a_renames,
2563                            struct string_list *b_renames)
2564 {
2565         int clean_merge = 1, i, j;
2566         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2567         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2568         const struct rename *sre;
2569
2570         for (i = 0; i < a_renames->nr; i++) {
2571                 sre = a_renames->items[i].util;
2572                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2573                         = (void *)sre;
2574         }
2575         for (i = 0; i < b_renames->nr; i++) {
2576                 sre = b_renames->items[i].util;
2577                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2578                         = (void *)sre;
2579         }
2580
2581         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2582                 struct string_list *renames1, *renames2Dst;
2583                 struct rename *ren1 = NULL, *ren2 = NULL;
2584                 const char *branch1, *branch2;
2585                 const char *ren1_src, *ren1_dst;
2586                 struct string_list_item *lookup;
2587
2588                 if (i >= a_renames->nr) {
2589                         ren2 = b_renames->items[j++].util;
2590                 } else if (j >= b_renames->nr) {
2591                         ren1 = a_renames->items[i++].util;
2592                 } else {
2593                         int compare = strcmp(a_renames->items[i].string,
2594                                              b_renames->items[j].string);
2595                         if (compare <= 0)
2596                                 ren1 = a_renames->items[i++].util;
2597                         if (compare >= 0)
2598                                 ren2 = b_renames->items[j++].util;
2599                 }
2600
2601                 /* TODO: refactor, so that 1/2 are not needed */
2602                 if (ren1) {
2603                         renames1 = a_renames;
2604                         renames2Dst = &b_by_dst;
2605                         branch1 = o->branch1;
2606                         branch2 = o->branch2;
2607                 } else {
2608                         renames1 = b_renames;
2609                         renames2Dst = &a_by_dst;
2610                         branch1 = o->branch2;
2611                         branch2 = o->branch1;
2612                         SWAP(ren2, ren1);
2613                 }
2614
2615                 if (ren1->processed)
2616                         continue;
2617                 ren1->processed = 1;
2618                 ren1->dst_entry->processed = 1;
2619                 /* BUG: We should only mark src_entry as processed if we
2620                  * are not dealing with a rename + add-source case.
2621                  */
2622                 ren1->src_entry->processed = 1;
2623
2624                 ren1_src = ren1->pair->one->path;
2625                 ren1_dst = ren1->pair->two->path;
2626
2627                 if (ren2) {
2628                         /* One file renamed on both sides */
2629                         const char *ren2_src = ren2->pair->one->path;
2630                         const char *ren2_dst = ren2->pair->two->path;
2631                         enum rename_type rename_type;
2632                         if (strcmp(ren1_src, ren2_src) != 0)
2633                                 BUG("ren1_src != ren2_src");
2634                         ren2->dst_entry->processed = 1;
2635                         ren2->processed = 1;
2636                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2637                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2638                                 clean_merge = 0;
2639                         } else {
2640                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2641                                 /* BUG: We should only remove ren1_src in
2642                                  * the base stage (think of rename +
2643                                  * add-source cases).
2644                                  */
2645                                 remove_file(o, 1, ren1_src, 1);
2646                                 update_entry(ren1->dst_entry,
2647                                              ren1->pair->one,
2648                                              ren1->pair->two,
2649                                              ren2->pair->two);
2650                         }
2651                         setup_rename_conflict_info(rename_type,
2652                                                    ren1->pair,
2653                                                    ren2->pair,
2654                                                    branch1,
2655                                                    branch2,
2656                                                    ren1->dst_entry,
2657                                                    ren2->dst_entry,
2658                                                    o,
2659                                                    NULL,
2660                                                    NULL);
2661                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2662                         /* Two different files renamed to the same thing */
2663                         char *ren2_dst;
2664                         ren2 = lookup->util;
2665                         ren2_dst = ren2->pair->two->path;
2666                         if (strcmp(ren1_dst, ren2_dst) != 0)
2667                                 BUG("ren1_dst != ren2_dst");
2668
2669                         clean_merge = 0;
2670                         ren2->processed = 1;
2671                         /*
2672                          * BUG: We should only mark src_entry as processed
2673                          * if we are not dealing with a rename + add-source
2674                          * case.
2675                          */
2676                         ren2->src_entry->processed = 1;
2677
2678                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2679                                                    ren1->pair,
2680                                                    ren2->pair,
2681                                                    branch1,
2682                                                    branch2,
2683                                                    ren1->dst_entry,
2684                                                    ren2->dst_entry,
2685                                                    o,
2686                                                    ren1->src_entry,
2687                                                    ren2->src_entry);
2688
2689                 } else {
2690                         /* Renamed in 1, maybe changed in 2 */
2691                         /* we only use sha1 and mode of these */
2692                         struct diff_filespec src_other, dst_other;
2693                         int try_merge;
2694
2695                         /*
2696                          * unpack_trees loads entries from common-commit
2697                          * into stage 1, from head-commit into stage 2, and
2698                          * from merge-commit into stage 3.  We keep track
2699                          * of which side corresponds to the rename.
2700                          */
2701                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2702                         int other_stage =   a_renames == renames1 ? 3 : 2;
2703
2704                         /* BUG: We should only remove ren1_src in the base
2705                          * stage and in other_stage (think of rename +
2706                          * add-source case).
2707                          */
2708                         remove_file(o, 1, ren1_src,
2709                                     renamed_stage == 2 || !was_tracked(o, ren1_src));
2710
2711                         oidcpy(&src_other.oid,
2712                                &ren1->src_entry->stages[other_stage].oid);
2713                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2714                         oidcpy(&dst_other.oid,
2715                                &ren1->dst_entry->stages[other_stage].oid);
2716                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2717                         try_merge = 0;
2718
2719                         if (oid_eq(&src_other.oid, &null_oid) &&
2720                             ren1->add_turned_into_rename) {
2721                                 setup_rename_conflict_info(RENAME_VIA_DIR,
2722                                                            ren1->pair,
2723                                                            NULL,
2724                                                            branch1,
2725                                                            branch2,
2726                                                            ren1->dst_entry,
2727                                                            NULL,
2728                                                            o,
2729                                                            NULL,
2730                                                            NULL);
2731                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2732                                 setup_rename_conflict_info(RENAME_DELETE,
2733                                                            ren1->pair,
2734                                                            NULL,
2735                                                            branch1,
2736                                                            branch2,
2737                                                            ren1->dst_entry,
2738                                                            NULL,
2739                                                            o,
2740                                                            NULL,
2741                                                            NULL);
2742                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2743                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2744                                 /*
2745                                  * Added file on the other side identical to
2746                                  * the file being renamed: clean merge.
2747                                  * Also, there is no need to overwrite the
2748                                  * file already in the working copy, so call
2749                                  * update_file_flags() instead of
2750                                  * update_file().
2751                                  */
2752                                 if (update_file_flags(o,
2753                                                       &ren1->pair->two->oid,
2754                                                       ren1->pair->two->mode,
2755                                                       ren1_dst,
2756                                                       1, /* update_cache */
2757                                                       0  /* update_wd    */))
2758                                         clean_merge = -1;
2759                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2760                                 clean_merge = 0;
2761                                 try_merge = 1;
2762                                 output(o, 1, _("CONFLICT (rename/add): Rename %s->%s in %s. "
2763                                        "%s added in %s"),
2764                                        ren1_src, ren1_dst, branch1,
2765                                        ren1_dst, branch2);
2766                                 if (o->call_depth) {
2767                                         struct merge_file_info mfi;
2768                                         if (merge_file_one(o, ren1_dst, &null_oid, 0,
2769                                                            &ren1->pair->two->oid,
2770                                                            ren1->pair->two->mode,
2771                                                            &dst_other.oid,
2772                                                            dst_other.mode,
2773                                                            branch1, branch2, &mfi)) {
2774                                                 clean_merge = -1;
2775                                                 goto cleanup_and_return;
2776                                         }
2777                                         output(o, 1, _("Adding merged %s"), ren1_dst);
2778                                         if (update_file(o, 0, &mfi.oid,
2779                                                         mfi.mode, ren1_dst))
2780                                                 clean_merge = -1;
2781                                         try_merge = 0;
2782                                 } else {
2783                                         char *new_path = unique_path(o, ren1_dst, branch2);
2784                                         output(o, 1, _("Adding as %s instead"), new_path);
2785                                         if (update_file(o, 0, &dst_other.oid,
2786                                                         dst_other.mode, new_path))
2787                                                 clean_merge = -1;
2788                                         free(new_path);
2789                                 }
2790                         } else
2791                                 try_merge = 1;
2792
2793                         if (clean_merge < 0)
2794                                 goto cleanup_and_return;
2795                         if (try_merge) {
2796                                 struct diff_filespec *one, *a, *b;
2797                                 src_other.path = (char *)ren1_src;
2798
2799                                 one = ren1->pair->one;
2800                                 if (a_renames == renames1) {
2801                                         a = ren1->pair->two;
2802                                         b = &src_other;
2803                                 } else {
2804                                         b = ren1->pair->two;
2805                                         a = &src_other;
2806                                 }
2807                                 update_entry(ren1->dst_entry, one, a, b);
2808                                 setup_rename_conflict_info(RENAME_NORMAL,
2809                                                            ren1->pair,
2810                                                            NULL,
2811                                                            branch1,
2812                                                            NULL,
2813                                                            ren1->dst_entry,
2814                                                            NULL,
2815                                                            o,
2816                                                            NULL,
2817                                                            NULL);
2818                         }
2819                 }
2820         }
2821 cleanup_and_return:
2822         string_list_clear(&a_by_dst, 0);
2823         string_list_clear(&b_by_dst, 0);
2824
2825         return clean_merge;
2826 }
2827
2828 struct rename_info {
2829         struct string_list *head_renames;
2830         struct string_list *merge_renames;
2831 };
2832
2833 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2834                                    struct hashmap *dir_renames)
2835 {
2836         struct hashmap_iter iter;
2837         struct dir_rename_entry *e;
2838
2839         hashmap_iter_init(dir_renames, &iter);
2840         while ((e = hashmap_iter_next(&iter))) {
2841                 free(e->dir);
2842                 strbuf_release(&e->new_dir);
2843                 /* possible_new_dirs already cleared in get_directory_renames */
2844         }
2845         hashmap_free(dir_renames, 1);
2846         free(dir_renames);
2847
2848         free(pairs->queue);
2849         free(pairs);
2850 }
2851
2852 static int detect_and_process_renames(struct merge_options *o,
2853                                       struct tree *common,
2854                                       struct tree *head,
2855                                       struct tree *merge,
2856                                       struct string_list *entries,
2857                                       struct rename_info *ri)
2858 {
2859         struct diff_queue_struct *head_pairs, *merge_pairs;
2860         struct hashmap *dir_re_head, *dir_re_merge;
2861         int clean = 1;
2862
2863         ri->head_renames = NULL;
2864         ri->merge_renames = NULL;
2865
2866         if (!merge_detect_rename(o))
2867                 return 1;
2868
2869         head_pairs = get_diffpairs(o, common, head);
2870         merge_pairs = get_diffpairs(o, common, merge);
2871
2872         if (o->detect_directory_renames) {
2873                 dir_re_head = get_directory_renames(head_pairs, head);
2874                 dir_re_merge = get_directory_renames(merge_pairs, merge);
2875
2876                 handle_directory_level_conflicts(o,
2877                                                  dir_re_head, head,
2878                                                  dir_re_merge, merge);
2879         } else {
2880                 dir_re_head  = xmalloc(sizeof(*dir_re_head));
2881                 dir_re_merge = xmalloc(sizeof(*dir_re_merge));
2882                 dir_rename_init(dir_re_head);
2883                 dir_rename_init(dir_re_merge);
2884         }
2885
2886         ri->head_renames  = get_renames(o, head_pairs,
2887                                         dir_re_merge, dir_re_head, head,
2888                                         common, head, merge, entries,
2889                                         &clean);
2890         if (clean < 0)
2891                 goto cleanup;
2892         ri->merge_renames = get_renames(o, merge_pairs,
2893                                         dir_re_head, dir_re_merge, merge,
2894                                         common, head, merge, entries,
2895                                         &clean);
2896         if (clean < 0)
2897                 goto cleanup;
2898         clean &= process_renames(o, ri->head_renames, ri->merge_renames);
2899
2900 cleanup:
2901         /*
2902          * Some cleanup is deferred until cleanup_renames() because the
2903          * data structures are still needed and referenced in
2904          * process_entry().  But there are a few things we can free now.
2905          */
2906         initial_cleanup_rename(head_pairs, dir_re_head);
2907         initial_cleanup_rename(merge_pairs, dir_re_merge);
2908
2909         return clean;
2910 }
2911
2912 static void final_cleanup_rename(struct string_list *rename)
2913 {
2914         const struct rename *re;
2915         int i;
2916
2917         if (rename == NULL)
2918                 return;
2919
2920         for (i = 0; i < rename->nr; i++) {
2921                 re = rename->items[i].util;
2922                 diff_free_filepair(re->pair);
2923         }
2924         string_list_clear(rename, 1);
2925         free(rename);
2926 }
2927
2928 static void final_cleanup_renames(struct rename_info *re_info)
2929 {
2930         final_cleanup_rename(re_info->head_renames);
2931         final_cleanup_rename(re_info->merge_renames);
2932 }
2933
2934 static struct object_id *stage_oid(const struct object_id *oid, unsigned mode)
2935 {
2936         return (is_null_oid(oid) || mode == 0) ? NULL: (struct object_id *)oid;
2937 }
2938
2939 static int read_oid_strbuf(struct merge_options *o,
2940                            const struct object_id *oid,
2941                            struct strbuf *dst)
2942 {
2943         void *buf;
2944         enum object_type type;
2945         unsigned long size;
2946         buf = read_object_file(oid, &type, &size);
2947         if (!buf)
2948                 return err(o, _("cannot read object %s"), oid_to_hex(oid));
2949         if (type != OBJ_BLOB) {
2950                 free(buf);
2951                 return err(o, _("object %s is not a blob"), oid_to_hex(oid));
2952         }
2953         strbuf_attach(dst, buf, size, size + 1);
2954         return 0;
2955 }
2956
2957 static int blob_unchanged(struct merge_options *opt,
2958                           const struct object_id *o_oid,
2959                           unsigned o_mode,
2960                           const struct object_id *a_oid,
2961                           unsigned a_mode,
2962                           int renormalize, const char *path)
2963 {
2964         struct strbuf o = STRBUF_INIT;
2965         struct strbuf a = STRBUF_INIT;
2966         int ret = 0; /* assume changed for safety */
2967
2968         if (a_mode != o_mode)
2969                 return 0;
2970         if (oid_eq(o_oid, a_oid))
2971                 return 1;
2972         if (!renormalize)
2973                 return 0;
2974
2975         assert(o_oid && a_oid);
2976         if (read_oid_strbuf(opt, o_oid, &o) || read_oid_strbuf(opt, a_oid, &a))
2977                 goto error_return;
2978         /*
2979          * Note: binary | is used so that both renormalizations are
2980          * performed.  Comparison can be skipped if both files are
2981          * unchanged since their sha1s have already been compared.
2982          */
2983         if (renormalize_buffer(&the_index, path, o.buf, o.len, &o) |
2984             renormalize_buffer(&the_index, path, a.buf, a.len, &a))
2985                 ret = (o.len == a.len && !memcmp(o.buf, a.buf, o.len));
2986
2987 error_return:
2988         strbuf_release(&o);
2989         strbuf_release(&a);
2990         return ret;
2991 }
2992
2993 static int handle_modify_delete(struct merge_options *o,
2994                                 const char *path,
2995                                 struct object_id *o_oid, int o_mode,
2996                                 struct object_id *a_oid, int a_mode,
2997                                 struct object_id *b_oid, int b_mode)
2998 {
2999         const char *modify_branch, *delete_branch;
3000         struct object_id *changed_oid;
3001         int changed_mode;
3002
3003         if (a_oid) {
3004                 modify_branch = o->branch1;
3005                 delete_branch = o->branch2;
3006                 changed_oid = a_oid;
3007                 changed_mode = a_mode;
3008         } else {
3009                 modify_branch = o->branch2;
3010                 delete_branch = o->branch1;
3011                 changed_oid = b_oid;
3012                 changed_mode = b_mode;
3013         }
3014
3015         return handle_change_delete(o,
3016                                     path, NULL,
3017                                     o_oid, o_mode,
3018                                     changed_oid, changed_mode,
3019                                     modify_branch, delete_branch,
3020                                     _("modify"), _("modified"));
3021 }
3022
3023 static int merge_content(struct merge_options *o,
3024                          const char *path,
3025                          int is_dirty,
3026                          struct object_id *o_oid, int o_mode,
3027                          struct object_id *a_oid, int a_mode,
3028                          struct object_id *b_oid, int b_mode,
3029                          struct rename_conflict_info *rename_conflict_info)
3030 {
3031         const char *reason = _("content");
3032         const char *path1 = NULL, *path2 = NULL;
3033         struct merge_file_info mfi;
3034         struct diff_filespec one, a, b;
3035         unsigned df_conflict_remains = 0;
3036
3037         if (!o_oid) {
3038                 reason = _("add/add");
3039                 o_oid = (struct object_id *)&null_oid;
3040         }
3041         one.path = a.path = b.path = (char *)path;
3042         oidcpy(&one.oid, o_oid);
3043         one.mode = o_mode;
3044         oidcpy(&a.oid, a_oid);
3045         a.mode = a_mode;
3046         oidcpy(&b.oid, b_oid);
3047         b.mode = b_mode;
3048
3049         if (rename_conflict_info) {
3050                 struct diff_filepair *pair1 = rename_conflict_info->pair1;
3051
3052                 path1 = (o->branch1 == rename_conflict_info->branch1) ?
3053                         pair1->two->path : pair1->one->path;
3054                 /* If rename_conflict_info->pair2 != NULL, we are in
3055                  * RENAME_ONE_FILE_TO_ONE case.  Otherwise, we have a
3056                  * normal rename.
3057                  */
3058                 path2 = (rename_conflict_info->pair2 ||
3059                          o->branch2 == rename_conflict_info->branch1) ?
3060                         pair1->two->path : pair1->one->path;
3061
3062                 if (dir_in_way(path, !o->call_depth,
3063                                S_ISGITLINK(pair1->two->mode)))
3064                         df_conflict_remains = 1;
3065         }
3066         if (merge_file_special_markers(o, &one, &a, &b, path,
3067                                        o->branch1, path1,
3068                                        o->branch2, path2, &mfi))
3069                 return -1;
3070
3071         /*
3072          * We can skip updating the working tree file iff:
3073          *   a) The merge is clean
3074          *   b) The merge matches what was in HEAD (content, mode, pathname)
3075          *   c) The target path is usable (i.e. not involved in D/F conflict)
3076          */
3077         if (mfi.clean &&
3078             was_tracked_and_matches(o, path, &mfi.oid, mfi.mode) &&
3079             !df_conflict_remains) {
3080                 int pos;
3081                 struct cache_entry *ce;
3082
3083                 output(o, 3, _("Skipped %s (merged same as existing)"), path);
3084                 if (add_cacheinfo(o, mfi.mode, &mfi.oid, path,
3085                                   0, (!o->call_depth && !is_dirty), 0))
3086                         return -1;
3087                 /*
3088                  * However, add_cacheinfo() will delete the old cache entry
3089                  * and add a new one.  We need to copy over any skip_worktree
3090                  * flag to avoid making the file appear as if it were
3091                  * deleted by the user.
3092                  */
3093                 pos = index_name_pos(&o->orig_index, path, strlen(path));
3094                 ce = o->orig_index.cache[pos];
3095                 if (ce_skip_worktree(ce)) {
3096                         pos = index_name_pos(&the_index, path, strlen(path));
3097                         ce = the_index.cache[pos];
3098                         ce->ce_flags |= CE_SKIP_WORKTREE;
3099                 }
3100                 return mfi.clean;
3101         }
3102
3103         if (!mfi.clean) {
3104                 if (S_ISGITLINK(mfi.mode))
3105                         reason = _("submodule");
3106                 output(o, 1, _("CONFLICT (%s): Merge conflict in %s"),
3107                                 reason, path);
3108                 if (rename_conflict_info && !df_conflict_remains)
3109                         if (update_stages(o, path, &one, &a, &b))
3110                                 return -1;
3111         }
3112
3113         if (df_conflict_remains || is_dirty) {
3114                 char *new_path;
3115                 if (o->call_depth) {
3116                         remove_file_from_cache(path);
3117                 } else {
3118                         if (!mfi.clean) {
3119                                 if (update_stages(o, path, &one, &a, &b))
3120                                         return -1;
3121                         } else {
3122                                 int file_from_stage2 = was_tracked(o, path);
3123                                 struct diff_filespec merged;
3124                                 oidcpy(&merged.oid, &mfi.oid);
3125                                 merged.mode = mfi.mode;
3126
3127                                 if (update_stages(o, path, NULL,
3128                                                   file_from_stage2 ? &merged : NULL,
3129                                                   file_from_stage2 ? NULL : &merged))
3130                                         return -1;
3131                         }
3132
3133                 }
3134                 new_path = unique_path(o, path, rename_conflict_info->branch1);
3135                 if (is_dirty) {
3136                         output(o, 1, _("Refusing to lose dirty file at %s"),
3137                                path);
3138                 }
3139                 output(o, 1, _("Adding as %s instead"), new_path);
3140                 if (update_file(o, 0, &mfi.oid, mfi.mode, new_path)) {
3141                         free(new_path);
3142                         return -1;
3143                 }
3144                 free(new_path);
3145                 mfi.clean = 0;
3146         } else if (update_file(o, mfi.clean, &mfi.oid, mfi.mode, path))
3147                 return -1;
3148         return !is_dirty && mfi.clean;
3149 }
3150
3151 static int handle_rename_normal(struct merge_options *o,
3152                                 const char *path,
3153                                 struct object_id *o_oid, unsigned int o_mode,
3154                                 struct object_id *a_oid, unsigned int a_mode,
3155                                 struct object_id *b_oid, unsigned int b_mode,
3156                                 struct rename_conflict_info *ci)
3157 {
3158         /* Merge the content and write it out */
3159         return merge_content(o, path, was_dirty(o, path),
3160                              o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3161                              ci);
3162 }
3163
3164 /* Per entry merge function */
3165 static int process_entry(struct merge_options *o,
3166                          const char *path, struct stage_data *entry)
3167 {
3168         int clean_merge = 1;
3169         int normalize = o->renormalize;
3170         unsigned o_mode = entry->stages[1].mode;
3171         unsigned a_mode = entry->stages[2].mode;
3172         unsigned b_mode = entry->stages[3].mode;
3173         struct object_id *o_oid = stage_oid(&entry->stages[1].oid, o_mode);
3174         struct object_id *a_oid = stage_oid(&entry->stages[2].oid, a_mode);
3175         struct object_id *b_oid = stage_oid(&entry->stages[3].oid, b_mode);
3176
3177         entry->processed = 1;
3178         if (entry->rename_conflict_info) {
3179                 struct rename_conflict_info *conflict_info = entry->rename_conflict_info;
3180                 switch (conflict_info->rename_type) {
3181                 case RENAME_NORMAL:
3182                 case RENAME_ONE_FILE_TO_ONE:
3183                         clean_merge = handle_rename_normal(o,
3184                                                            path,
3185                                                            o_oid, o_mode,
3186                                                            a_oid, a_mode,
3187                                                            b_oid, b_mode,
3188                                                            conflict_info);
3189                         break;
3190                 case RENAME_VIA_DIR:
3191                         clean_merge = 1;
3192                         if (handle_rename_via_dir(o,
3193                                                   conflict_info->pair1,
3194                                                   conflict_info->branch1,
3195                                                   conflict_info->branch2))
3196                                 clean_merge = -1;
3197                         break;
3198                 case RENAME_DELETE:
3199                         clean_merge = 0;
3200                         if (handle_rename_delete(o,
3201                                                  conflict_info->pair1,
3202                                                  conflict_info->branch1,
3203                                                  conflict_info->branch2))
3204                                 clean_merge = -1;
3205                         break;
3206                 case RENAME_ONE_FILE_TO_TWO:
3207                         clean_merge = 0;
3208                         if (handle_rename_rename_1to2(o, conflict_info))
3209                                 clean_merge = -1;
3210                         break;
3211                 case RENAME_TWO_FILES_TO_ONE:
3212                         clean_merge = 0;
3213                         if (handle_rename_rename_2to1(o, conflict_info))
3214                                 clean_merge = -1;
3215                         break;
3216                 default:
3217                         entry->processed = 0;
3218                         break;
3219                 }
3220         } else if (o_oid && (!a_oid || !b_oid)) {
3221                 /* Case A: Deleted in one */
3222                 if ((!a_oid && !b_oid) ||
3223                     (!b_oid && blob_unchanged(o, o_oid, o_mode, a_oid, a_mode, normalize, path)) ||
3224                     (!a_oid && blob_unchanged(o, o_oid, o_mode, b_oid, b_mode, normalize, path))) {
3225                         /* Deleted in both or deleted in one and
3226                          * unchanged in the other */
3227                         if (a_oid)
3228                                 output(o, 2, _("Removing %s"), path);
3229                         /* do not touch working file if it did not exist */
3230                         remove_file(o, 1, path, !a_oid);
3231                 } else {
3232                         /* Modify/delete; deleted side may have put a directory in the way */
3233                         clean_merge = 0;
3234                         if (handle_modify_delete(o, path, o_oid, o_mode,
3235                                                  a_oid, a_mode, b_oid, b_mode))
3236                                 clean_merge = -1;
3237                 }
3238         } else if ((!o_oid && a_oid && !b_oid) ||
3239                    (!o_oid && !a_oid && b_oid)) {
3240                 /* Case B: Added in one. */
3241                 /* [nothing|directory] -> ([nothing|directory], file) */
3242
3243                 const char *add_branch;
3244                 const char *other_branch;
3245                 unsigned mode;
3246                 const struct object_id *oid;
3247                 const char *conf;
3248
3249                 if (a_oid) {
3250                         add_branch = o->branch1;
3251                         other_branch = o->branch2;
3252                         mode = a_mode;
3253                         oid = a_oid;
3254                         conf = _("file/directory");
3255                 } else {
3256                         add_branch = o->branch2;
3257                         other_branch = o->branch1;
3258                         mode = b_mode;
3259                         oid = b_oid;
3260                         conf = _("directory/file");
3261                 }
3262                 if (dir_in_way(path,
3263                                !o->call_depth && !S_ISGITLINK(a_mode),
3264                                0)) {
3265                         char *new_path = unique_path(o, path, add_branch);
3266                         clean_merge = 0;
3267                         output(o, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3268                                "Adding %s as %s"),
3269                                conf, path, other_branch, path, new_path);
3270                         if (update_file(o, 0, oid, mode, new_path))
3271                                 clean_merge = -1;
3272                         else if (o->call_depth)
3273                                 remove_file_from_cache(path);
3274                         free(new_path);
3275                 } else {
3276                         output(o, 2, _("Adding %s"), path);
3277                         /* do not overwrite file if already present */
3278                         if (update_file_flags(o, oid, mode, path, 1, !a_oid))
3279                                 clean_merge = -1;
3280                 }
3281         } else if (a_oid && b_oid) {
3282                 /* Case C: Added in both (check for same permissions) and */
3283                 /* case D: Modified in both, but differently. */
3284                 int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3285                 clean_merge = merge_content(o, path, is_dirty,
3286                                             o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3287                                             NULL);
3288         } else if (!o_oid && !a_oid && !b_oid) {
3289                 /*
3290                  * this entry was deleted altogether. a_mode == 0 means
3291                  * we had that path and want to actively remove it.
3292                  */
3293                 remove_file(o, 1, path, !a_mode);
3294         } else
3295                 BUG("fatal merge failure, shouldn't happen.");
3296
3297         return clean_merge;
3298 }
3299
3300 int merge_trees(struct merge_options *o,
3301                 struct tree *head,
3302                 struct tree *merge,
3303                 struct tree *common,
3304                 struct tree **result)
3305 {
3306         int code, clean;
3307         struct strbuf sb = STRBUF_INIT;
3308
3309         if (!o->call_depth && index_has_changes(&the_index, head, &sb)) {
3310                 err(o, _("Your local changes to the following files would be overwritten by merge:\n  %s"),
3311                     sb.buf);
3312                 return -1;
3313         }
3314
3315         if (o->subtree_shift) {
3316                 merge = shift_tree_object(head, merge, o->subtree_shift);
3317                 common = shift_tree_object(head, common, o->subtree_shift);
3318         }
3319
3320         if (oid_eq(&common->object.oid, &merge->object.oid)) {
3321                 output(o, 0, _("Already up to date!"));
3322                 *result = head;
3323                 return 1;
3324         }
3325
3326         code = unpack_trees_start(o, common, head, merge);
3327
3328         if (code != 0) {
3329                 if (show(o, 4) || o->call_depth)
3330                         err(o, _("merging of trees %s and %s failed"),
3331                             oid_to_hex(&head->object.oid),
3332                             oid_to_hex(&merge->object.oid));
3333                 unpack_trees_finish(o);
3334                 return -1;
3335         }
3336
3337         if (unmerged_cache()) {
3338                 struct string_list *entries;
3339                 struct rename_info re_info;
3340                 int i;
3341                 /*
3342                  * Only need the hashmap while processing entries, so
3343                  * initialize it here and free it when we are done running
3344                  * through the entries. Keeping it in the merge_options as
3345                  * opposed to decaring a local hashmap is for convenience
3346                  * so that we don't have to pass it to around.
3347                  */
3348                 hashmap_init(&o->current_file_dir_set, path_hashmap_cmp, NULL, 512);
3349                 get_files_dirs(o, head);
3350                 get_files_dirs(o, merge);
3351
3352                 entries = get_unmerged();
3353                 clean = detect_and_process_renames(o, common, head, merge,
3354                                                    entries, &re_info);
3355                 record_df_conflict_files(o, entries);
3356                 if (clean < 0)
3357                         goto cleanup;
3358                 for (i = entries->nr-1; 0 <= i; i--) {
3359                         const char *path = entries->items[i].string;
3360                         struct stage_data *e = entries->items[i].util;
3361                         if (!e->processed) {
3362                                 int ret = process_entry(o, path, e);
3363                                 if (!ret)
3364                                         clean = 0;
3365                                 else if (ret < 0) {
3366                                         clean = ret;
3367                                         goto cleanup;
3368                                 }
3369                         }
3370                 }
3371                 for (i = 0; i < entries->nr; i++) {
3372                         struct stage_data *e = entries->items[i].util;
3373                         if (!e->processed)
3374                                 BUG("unprocessed path??? %s",
3375                                     entries->items[i].string);
3376                 }
3377
3378         cleanup:
3379                 final_cleanup_renames(&re_info);
3380
3381                 string_list_clear(entries, 1);
3382                 free(entries);
3383
3384                 hashmap_free(&o->current_file_dir_set, 1);
3385
3386                 if (clean < 0) {
3387                         unpack_trees_finish(o);
3388                         return clean;
3389                 }
3390         }
3391         else
3392                 clean = 1;
3393
3394         unpack_trees_finish(o);
3395
3396         if (o->call_depth && !(*result = write_tree_from_memory(o)))
3397                 return -1;
3398
3399         return clean;
3400 }
3401
3402 static struct commit_list *reverse_commit_list(struct commit_list *list)
3403 {
3404         struct commit_list *next = NULL, *current, *backup;
3405         for (current = list; current; current = backup) {
3406                 backup = current->next;
3407                 current->next = next;
3408                 next = current;
3409         }
3410         return next;
3411 }
3412
3413 /*
3414  * Merge the commits h1 and h2, return the resulting virtual
3415  * commit object and a flag indicating the cleanness of the merge.
3416  */
3417 int merge_recursive(struct merge_options *o,
3418                     struct commit *h1,
3419                     struct commit *h2,
3420                     struct commit_list *ca,
3421                     struct commit **result)
3422 {
3423         struct commit_list *iter;
3424         struct commit *merged_common_ancestors;
3425         struct tree *mrtree;
3426         int clean;
3427
3428         if (show(o, 4)) {
3429                 output(o, 4, _("Merging:"));
3430                 output_commit_title(o, h1);
3431                 output_commit_title(o, h2);
3432         }
3433
3434         if (!ca) {
3435                 ca = get_merge_bases(h1, h2);
3436                 ca = reverse_commit_list(ca);
3437         }
3438
3439         if (show(o, 5)) {
3440                 unsigned cnt = commit_list_count(ca);
3441
3442                 output(o, 5, Q_("found %u common ancestor:",
3443                                 "found %u common ancestors:", cnt), cnt);
3444                 for (iter = ca; iter; iter = iter->next)
3445                         output_commit_title(o, iter->item);
3446         }
3447
3448         merged_common_ancestors = pop_commit(&ca);
3449         if (merged_common_ancestors == NULL) {
3450                 /* if there is no common ancestor, use an empty tree */
3451                 struct tree *tree;
3452
3453                 tree = lookup_tree(the_repository, the_repository->hash_algo->empty_tree);
3454                 merged_common_ancestors = make_virtual_commit(tree, "ancestor");
3455         }
3456
3457         for (iter = ca; iter; iter = iter->next) {
3458                 const char *saved_b1, *saved_b2;
3459                 o->call_depth++;
3460                 /*
3461                  * When the merge fails, the result contains files
3462                  * with conflict markers. The cleanness flag is
3463                  * ignored (unless indicating an error), it was never
3464                  * actually used, as result of merge_trees has always
3465                  * overwritten it: the committed "conflicts" were
3466                  * already resolved.
3467                  */
3468                 discard_cache();
3469                 saved_b1 = o->branch1;
3470                 saved_b2 = o->branch2;
3471                 o->branch1 = "Temporary merge branch 1";
3472                 o->branch2 = "Temporary merge branch 2";
3473                 if (merge_recursive(o, merged_common_ancestors, iter->item,
3474                                     NULL, &merged_common_ancestors) < 0)
3475                         return -1;
3476                 o->branch1 = saved_b1;
3477                 o->branch2 = saved_b2;
3478                 o->call_depth--;
3479
3480                 if (!merged_common_ancestors)
3481                         return err(o, _("merge returned no commit"));
3482         }
3483
3484         discard_cache();
3485         if (!o->call_depth)
3486                 read_cache();
3487
3488         o->ancestor = "merged common ancestors";
3489         clean = merge_trees(o, get_commit_tree(h1), get_commit_tree(h2),
3490                             get_commit_tree(merged_common_ancestors),
3491                             &mrtree);
3492         if (clean < 0) {
3493                 flush_output(o);
3494                 return clean;
3495         }
3496
3497         if (o->call_depth) {
3498                 *result = make_virtual_commit(mrtree, "merged tree");
3499                 commit_list_insert(h1, &(*result)->parents);
3500                 commit_list_insert(h2, &(*result)->parents->next);
3501         }
3502         flush_output(o);
3503         if (!o->call_depth && o->buffer_output < 2)
3504                 strbuf_release(&o->obuf);
3505         if (show(o, 2))
3506                 diff_warn_rename_limit("merge.renamelimit",
3507                                        o->needed_rename_limit, 0);
3508         return clean;
3509 }
3510
3511 static struct commit *get_ref(const struct object_id *oid, const char *name)
3512 {
3513         struct object *object;
3514
3515         object = deref_tag(the_repository, parse_object(the_repository, oid),
3516                            name,
3517                            strlen(name));
3518         if (!object)
3519                 return NULL;
3520         if (object->type == OBJ_TREE)
3521                 return make_virtual_commit((struct tree*)object, name);
3522         if (object->type != OBJ_COMMIT)
3523                 return NULL;
3524         if (parse_commit((struct commit *)object))
3525                 return NULL;
3526         return (struct commit *)object;
3527 }
3528
3529 int merge_recursive_generic(struct merge_options *o,
3530                             const struct object_id *head,
3531                             const struct object_id *merge,
3532                             int num_base_list,
3533                             const struct object_id **base_list,
3534                             struct commit **result)
3535 {
3536         int clean;
3537         struct lock_file lock = LOCK_INIT;
3538         struct commit *head_commit = get_ref(head, o->branch1);
3539         struct commit *next_commit = get_ref(merge, o->branch2);
3540         struct commit_list *ca = NULL;
3541
3542         if (base_list) {
3543                 int i;
3544                 for (i = 0; i < num_base_list; ++i) {
3545                         struct commit *base;
3546                         if (!(base = get_ref(base_list[i], oid_to_hex(base_list[i]))))
3547                                 return err(o, _("Could not parse object '%s'"),
3548                                            oid_to_hex(base_list[i]));
3549                         commit_list_insert(base, &ca);
3550                 }
3551         }
3552
3553         hold_locked_index(&lock, LOCK_DIE_ON_ERROR);
3554         clean = merge_recursive(o, head_commit, next_commit, ca,
3555                                 result);
3556         if (clean < 0) {
3557                 rollback_lock_file(&lock);
3558                 return clean;
3559         }
3560
3561         if (write_locked_index(&the_index, &lock,
3562                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3563                 return err(o, _("Unable to write index."));
3564
3565         return clean ? 0 : 1;
3566 }
3567
3568 static void merge_recursive_config(struct merge_options *o)
3569 {
3570         char *value = NULL;
3571         git_config_get_int("merge.verbosity", &o->verbosity);
3572         git_config_get_int("diff.renamelimit", &o->diff_rename_limit);
3573         git_config_get_int("merge.renamelimit", &o->merge_rename_limit);
3574         if (!git_config_get_string("diff.renames", &value)) {
3575                 o->diff_detect_rename = git_config_rename("diff.renames", value);
3576                 free(value);
3577         }
3578         if (!git_config_get_string("merge.renames", &value)) {
3579                 o->merge_detect_rename = git_config_rename("merge.renames", value);
3580                 free(value);
3581         }
3582         git_config(git_xmerge_config, NULL);
3583 }
3584
3585 void init_merge_options(struct merge_options *o)
3586 {
3587         const char *merge_verbosity;
3588         memset(o, 0, sizeof(struct merge_options));
3589         o->verbosity = 2;
3590         o->buffer_output = 1;
3591         o->diff_rename_limit = -1;
3592         o->merge_rename_limit = -1;
3593         o->renormalize = 0;
3594         o->diff_detect_rename = -1;
3595         o->merge_detect_rename = -1;
3596         o->detect_directory_renames = 1;
3597         merge_recursive_config(o);
3598         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3599         if (merge_verbosity)
3600                 o->verbosity = strtol(merge_verbosity, NULL, 10);
3601         if (o->verbosity >= 5)
3602                 o->buffer_output = 0;
3603         strbuf_init(&o->obuf, 0);
3604         string_list_init(&o->df_conflict_file_set, 1);
3605 }
3606
3607 int parse_merge_opt(struct merge_options *o, const char *s)
3608 {
3609         const char *arg;
3610
3611         if (!s || !*s)
3612                 return -1;
3613         if (!strcmp(s, "ours"))
3614                 o->recursive_variant = MERGE_RECURSIVE_OURS;
3615         else if (!strcmp(s, "theirs"))
3616                 o->recursive_variant = MERGE_RECURSIVE_THEIRS;
3617         else if (!strcmp(s, "subtree"))
3618                 o->subtree_shift = "";
3619         else if (skip_prefix(s, "subtree=", &arg))
3620                 o->subtree_shift = arg;
3621         else if (!strcmp(s, "patience"))
3622                 o->xdl_opts = DIFF_WITH_ALG(o, PATIENCE_DIFF);
3623         else if (!strcmp(s, "histogram"))
3624                 o->xdl_opts = DIFF_WITH_ALG(o, HISTOGRAM_DIFF);
3625         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3626                 long value = parse_algorithm_value(arg);
3627                 if (value < 0)
3628                         return -1;
3629                 /* clear out previous settings */
3630                 DIFF_XDL_CLR(o, NEED_MINIMAL);
3631                 o->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3632                 o->xdl_opts |= value;
3633         }
3634         else if (!strcmp(s, "ignore-space-change"))
3635                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_CHANGE);
3636         else if (!strcmp(s, "ignore-all-space"))
3637                 DIFF_XDL_SET(o, IGNORE_WHITESPACE);
3638         else if (!strcmp(s, "ignore-space-at-eol"))
3639                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_AT_EOL);
3640         else if (!strcmp(s, "ignore-cr-at-eol"))
3641                 DIFF_XDL_SET(o, IGNORE_CR_AT_EOL);
3642         else if (!strcmp(s, "renormalize"))
3643                 o->renormalize = 1;
3644         else if (!strcmp(s, "no-renormalize"))
3645                 o->renormalize = 0;
3646         else if (!strcmp(s, "no-renames"))
3647                 o->merge_detect_rename = 0;
3648         else if (!strcmp(s, "find-renames")) {
3649                 o->merge_detect_rename = 1;
3650                 o->rename_score = 0;
3651         }
3652         else if (skip_prefix(s, "find-renames=", &arg) ||
3653                  skip_prefix(s, "rename-threshold=", &arg)) {
3654                 if ((o->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3655                         return -1;
3656                 o->merge_detect_rename = 1;
3657         }
3658         else
3659                 return -1;
3660         return 0;
3661 }