tree.c: make read_tree*() take 'struct repository *'
[git] / merge-recursive.c
1 /*
2  * Recursive Merge algorithm stolen from git-merge-recursive.py by
3  * Fredrik Kuivinen.
4  * The thieves were Alex Riesen and Johannes Schindelin, in June/July 2006
5  */
6 #include "cache.h"
7 #include "config.h"
8 #include "advice.h"
9 #include "lockfile.h"
10 #include "cache-tree.h"
11 #include "object-store.h"
12 #include "repository.h"
13 #include "commit.h"
14 #include "blob.h"
15 #include "builtin.h"
16 #include "tree-walk.h"
17 #include "diff.h"
18 #include "diffcore.h"
19 #include "tag.h"
20 #include "alloc.h"
21 #include "unpack-trees.h"
22 #include "string-list.h"
23 #include "xdiff-interface.h"
24 #include "ll-merge.h"
25 #include "attr.h"
26 #include "merge-recursive.h"
27 #include "dir.h"
28 #include "submodule.h"
29 #include "revision.h"
30 #include "commit-reach.h"
31
32 struct path_hashmap_entry {
33         struct hashmap_entry e;
34         char path[FLEX_ARRAY];
35 };
36
37 static int path_hashmap_cmp(const void *cmp_data,
38                             const void *entry,
39                             const void *entry_or_key,
40                             const void *keydata)
41 {
42         const struct path_hashmap_entry *a = entry;
43         const struct path_hashmap_entry *b = entry_or_key;
44         const char *key = keydata;
45
46         if (ignore_case)
47                 return strcasecmp(a->path, key ? key : b->path);
48         else
49                 return strcmp(a->path, key ? key : b->path);
50 }
51
52 static unsigned int path_hash(const char *path)
53 {
54         return ignore_case ? strihash(path) : strhash(path);
55 }
56
57 static struct dir_rename_entry *dir_rename_find_entry(struct hashmap *hashmap,
58                                                       char *dir)
59 {
60         struct dir_rename_entry key;
61
62         if (dir == NULL)
63                 return NULL;
64         hashmap_entry_init(&key, strhash(dir));
65         key.dir = dir;
66         return hashmap_get(hashmap, &key, NULL);
67 }
68
69 static int dir_rename_cmp(const void *unused_cmp_data,
70                           const void *entry,
71                           const void *entry_or_key,
72                           const void *unused_keydata)
73 {
74         const struct dir_rename_entry *e1 = entry;
75         const struct dir_rename_entry *e2 = entry_or_key;
76
77         return strcmp(e1->dir, e2->dir);
78 }
79
80 static void dir_rename_init(struct hashmap *map)
81 {
82         hashmap_init(map, dir_rename_cmp, NULL, 0);
83 }
84
85 static void dir_rename_entry_init(struct dir_rename_entry *entry,
86                                   char *directory)
87 {
88         hashmap_entry_init(entry, strhash(directory));
89         entry->dir = directory;
90         entry->non_unique_new_dir = 0;
91         strbuf_init(&entry->new_dir, 0);
92         string_list_init(&entry->possible_new_dirs, 0);
93 }
94
95 static struct collision_entry *collision_find_entry(struct hashmap *hashmap,
96                                                     char *target_file)
97 {
98         struct collision_entry key;
99
100         hashmap_entry_init(&key, strhash(target_file));
101         key.target_file = target_file;
102         return hashmap_get(hashmap, &key, NULL);
103 }
104
105 static int collision_cmp(void *unused_cmp_data,
106                          const struct collision_entry *e1,
107                          const struct collision_entry *e2,
108                          const void *unused_keydata)
109 {
110         return strcmp(e1->target_file, e2->target_file);
111 }
112
113 static void collision_init(struct hashmap *map)
114 {
115         hashmap_init(map, (hashmap_cmp_fn) collision_cmp, NULL, 0);
116 }
117
118 static void flush_output(struct merge_options *o)
119 {
120         if (o->buffer_output < 2 && o->obuf.len) {
121                 fputs(o->obuf.buf, stdout);
122                 strbuf_reset(&o->obuf);
123         }
124 }
125
126 static int err(struct merge_options *o, const char *err, ...)
127 {
128         va_list params;
129
130         if (o->buffer_output < 2)
131                 flush_output(o);
132         else {
133                 strbuf_complete(&o->obuf, '\n');
134                 strbuf_addstr(&o->obuf, "error: ");
135         }
136         va_start(params, err);
137         strbuf_vaddf(&o->obuf, err, params);
138         va_end(params);
139         if (o->buffer_output > 1)
140                 strbuf_addch(&o->obuf, '\n');
141         else {
142                 error("%s", o->obuf.buf);
143                 strbuf_reset(&o->obuf);
144         }
145
146         return -1;
147 }
148
149 static struct tree *shift_tree_object(struct tree *one, struct tree *two,
150                                       const char *subtree_shift)
151 {
152         struct object_id shifted;
153
154         if (!*subtree_shift) {
155                 shift_tree(&one->object.oid, &two->object.oid, &shifted, 0);
156         } else {
157                 shift_tree_by(&one->object.oid, &two->object.oid, &shifted,
158                               subtree_shift);
159         }
160         if (oideq(&two->object.oid, &shifted))
161                 return two;
162         return lookup_tree(the_repository, &shifted);
163 }
164
165 static struct commit *make_virtual_commit(struct tree *tree, const char *comment)
166 {
167         struct commit *commit = alloc_commit_node(the_repository);
168
169         set_merge_remote_desc(commit, comment, (struct object *)commit);
170         commit->maybe_tree = tree;
171         commit->object.parsed = 1;
172         return commit;
173 }
174
175 /*
176  * Since we use get_tree_entry(), which does not put the read object into
177  * the object pool, we cannot rely on a == b.
178  */
179 static int oid_eq(const struct object_id *a, const struct object_id *b)
180 {
181         if (!a && !b)
182                 return 2;
183         return a && b && oideq(a, b);
184 }
185
186 enum rename_type {
187         RENAME_NORMAL = 0,
188         RENAME_VIA_DIR,
189         RENAME_DELETE,
190         RENAME_ONE_FILE_TO_ONE,
191         RENAME_ONE_FILE_TO_TWO,
192         RENAME_TWO_FILES_TO_ONE
193 };
194
195 struct rename_conflict_info {
196         enum rename_type rename_type;
197         struct diff_filepair *pair1;
198         struct diff_filepair *pair2;
199         const char *branch1;
200         const char *branch2;
201         struct stage_data *dst_entry1;
202         struct stage_data *dst_entry2;
203         struct diff_filespec ren1_other;
204         struct diff_filespec ren2_other;
205 };
206
207 /*
208  * Since we want to write the index eventually, we cannot reuse the index
209  * for these (temporary) data.
210  */
211 struct stage_data {
212         struct {
213                 unsigned mode;
214                 struct object_id oid;
215         } stages[4];
216         struct rename_conflict_info *rename_conflict_info;
217         unsigned processed:1;
218 };
219
220 static inline void setup_rename_conflict_info(enum rename_type rename_type,
221                                               struct diff_filepair *pair1,
222                                               struct diff_filepair *pair2,
223                                               const char *branch1,
224                                               const char *branch2,
225                                               struct stage_data *dst_entry1,
226                                               struct stage_data *dst_entry2,
227                                               struct merge_options *o,
228                                               struct stage_data *src_entry1,
229                                               struct stage_data *src_entry2)
230 {
231         struct rename_conflict_info *ci;
232
233         /*
234          * When we have two renames involved, it's easiest to get the
235          * correct things into stage 2 and 3, and to make sure that the
236          * content merge puts HEAD before the other branch if we just
237          * ensure that branch1 == o->branch1.  So, simply flip arguments
238          * around if we don't have that.
239          */
240         if (dst_entry2 && branch1 != o->branch1) {
241                 setup_rename_conflict_info(rename_type,
242                                            pair2,      pair1,
243                                            branch2,    branch1,
244                                            dst_entry2, dst_entry1,
245                                            o,
246                                            src_entry2, src_entry1);
247                 return;
248         }
249
250         ci = xcalloc(1, sizeof(struct rename_conflict_info));
251         ci->rename_type = rename_type;
252         ci->pair1 = pair1;
253         ci->branch1 = branch1;
254         ci->branch2 = branch2;
255
256         ci->dst_entry1 = dst_entry1;
257         dst_entry1->rename_conflict_info = ci;
258         dst_entry1->processed = 0;
259
260         assert(!pair2 == !dst_entry2);
261         if (dst_entry2) {
262                 ci->dst_entry2 = dst_entry2;
263                 ci->pair2 = pair2;
264                 dst_entry2->rename_conflict_info = ci;
265         }
266
267         if (rename_type == RENAME_TWO_FILES_TO_ONE) {
268                 /*
269                  * For each rename, there could have been
270                  * modifications on the side of history where that
271                  * file was not renamed.
272                  */
273                 int ostage1 = o->branch1 == branch1 ? 3 : 2;
274                 int ostage2 = ostage1 ^ 1;
275
276                 ci->ren1_other.path = pair1->one->path;
277                 oidcpy(&ci->ren1_other.oid, &src_entry1->stages[ostage1].oid);
278                 ci->ren1_other.mode = src_entry1->stages[ostage1].mode;
279
280                 ci->ren2_other.path = pair2->one->path;
281                 oidcpy(&ci->ren2_other.oid, &src_entry2->stages[ostage2].oid);
282                 ci->ren2_other.mode = src_entry2->stages[ostage2].mode;
283         }
284 }
285
286 static int show(struct merge_options *o, int v)
287 {
288         return (!o->call_depth && o->verbosity >= v) || o->verbosity >= 5;
289 }
290
291 __attribute__((format (printf, 3, 4)))
292 static void output(struct merge_options *o, int v, const char *fmt, ...)
293 {
294         va_list ap;
295
296         if (!show(o, v))
297                 return;
298
299         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
300
301         va_start(ap, fmt);
302         strbuf_vaddf(&o->obuf, fmt, ap);
303         va_end(ap);
304
305         strbuf_addch(&o->obuf, '\n');
306         if (!o->buffer_output)
307                 flush_output(o);
308 }
309
310 static void output_commit_title(struct merge_options *o, struct commit *commit)
311 {
312         struct merge_remote_desc *desc;
313
314         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
315         desc = merge_remote_util(commit);
316         if (desc)
317                 strbuf_addf(&o->obuf, "virtual %s\n", desc->name);
318         else {
319                 strbuf_add_unique_abbrev(&o->obuf, &commit->object.oid,
320                                          DEFAULT_ABBREV);
321                 strbuf_addch(&o->obuf, ' ');
322                 if (parse_commit(commit) != 0)
323                         strbuf_addstr(&o->obuf, _("(bad commit)\n"));
324                 else {
325                         const char *title;
326                         const char *msg = get_commit_buffer(commit, NULL);
327                         int len = find_commit_subject(msg, &title);
328                         if (len)
329                                 strbuf_addf(&o->obuf, "%.*s\n", len, title);
330                         unuse_commit_buffer(commit, msg);
331                 }
332         }
333         flush_output(o);
334 }
335
336 static int add_cacheinfo(struct merge_options *o,
337                          unsigned int mode, const struct object_id *oid,
338                          const char *path, int stage, int refresh, int options)
339 {
340         struct cache_entry *ce;
341         int ret;
342
343         ce = make_cache_entry(&the_index, mode, oid ? oid : &null_oid, path, stage, 0);
344         if (!ce)
345                 return err(o, _("add_cacheinfo failed for path '%s'; merge aborting."), path);
346
347         ret = add_cache_entry(ce, options);
348         if (refresh) {
349                 struct cache_entry *nce;
350
351                 nce = refresh_cache_entry(&the_index, ce, CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
352                 if (!nce)
353                         return err(o, _("add_cacheinfo failed to refresh for path '%s'; merge aborting."), path);
354                 if (nce != ce)
355                         ret = add_cache_entry(nce, options);
356         }
357         return ret;
358 }
359
360 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
361 {
362         parse_tree(tree);
363         init_tree_desc(desc, tree->buffer, tree->size);
364 }
365
366 static int unpack_trees_start(struct merge_options *o,
367                               struct tree *common,
368                               struct tree *head,
369                               struct tree *merge)
370 {
371         int rc;
372         struct tree_desc t[3];
373         struct index_state tmp_index = { NULL };
374
375         memset(&o->unpack_opts, 0, sizeof(o->unpack_opts));
376         if (o->call_depth)
377                 o->unpack_opts.index_only = 1;
378         else
379                 o->unpack_opts.update = 1;
380         o->unpack_opts.merge = 1;
381         o->unpack_opts.head_idx = 2;
382         o->unpack_opts.fn = threeway_merge;
383         o->unpack_opts.src_index = &the_index;
384         o->unpack_opts.dst_index = &tmp_index;
385         o->unpack_opts.aggressive = !merge_detect_rename(o);
386         setup_unpack_trees_porcelain(&o->unpack_opts, "merge");
387
388         init_tree_desc_from_tree(t+0, common);
389         init_tree_desc_from_tree(t+1, head);
390         init_tree_desc_from_tree(t+2, merge);
391
392         rc = unpack_trees(3, t, &o->unpack_opts);
393         cache_tree_free(&active_cache_tree);
394
395         /*
396          * Update the_index to match the new results, AFTER saving a copy
397          * in o->orig_index.  Update src_index to point to the saved copy.
398          * (verify_uptodate() checks src_index, and the original index is
399          * the one that had the necessary modification timestamps.)
400          */
401         o->orig_index = the_index;
402         the_index = tmp_index;
403         o->unpack_opts.src_index = &o->orig_index;
404
405         return rc;
406 }
407
408 static void unpack_trees_finish(struct merge_options *o)
409 {
410         discard_index(&o->orig_index);
411         clear_unpack_trees_porcelain(&o->unpack_opts);
412 }
413
414 struct tree *write_tree_from_memory(struct merge_options *o)
415 {
416         struct tree *result = NULL;
417
418         if (unmerged_cache()) {
419                 int i;
420                 fprintf(stderr, "BUG: There are unmerged index entries:\n");
421                 for (i = 0; i < active_nr; i++) {
422                         const struct cache_entry *ce = active_cache[i];
423                         if (ce_stage(ce))
424                                 fprintf(stderr, "BUG: %d %.*s\n", ce_stage(ce),
425                                         (int)ce_namelen(ce), ce->name);
426                 }
427                 BUG("unmerged index entries in merge-recursive.c");
428         }
429
430         if (!active_cache_tree)
431                 active_cache_tree = cache_tree();
432
433         if (!cache_tree_fully_valid(active_cache_tree) &&
434             cache_tree_update(&the_index, 0) < 0) {
435                 err(o, _("error building trees"));
436                 return NULL;
437         }
438
439         result = lookup_tree(the_repository, &active_cache_tree->oid);
440
441         return result;
442 }
443
444 static int save_files_dirs(const struct object_id *oid,
445                            struct strbuf *base, const char *path,
446                            unsigned int mode, int stage, void *context)
447 {
448         struct path_hashmap_entry *entry;
449         int baselen = base->len;
450         struct merge_options *o = context;
451
452         strbuf_addstr(base, path);
453
454         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
455         hashmap_entry_init(entry, path_hash(entry->path));
456         hashmap_add(&o->current_file_dir_set, entry);
457
458         strbuf_setlen(base, baselen);
459         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
460 }
461
462 static void get_files_dirs(struct merge_options *o, struct tree *tree)
463 {
464         struct pathspec match_all;
465         memset(&match_all, 0, sizeof(match_all));
466         read_tree_recursive(the_repository, tree, "", 0, 0,
467                             &match_all, save_files_dirs, o);
468 }
469
470 static int get_tree_entry_if_blob(const struct object_id *tree,
471                                   const char *path,
472                                   struct object_id *hashy,
473                                   unsigned int *mode_o)
474 {
475         int ret;
476
477         ret = get_tree_entry(tree, path, hashy, mode_o);
478         if (S_ISDIR(*mode_o)) {
479                 oidcpy(hashy, &null_oid);
480                 *mode_o = 0;
481         }
482         return ret;
483 }
484
485 /*
486  * Returns an index_entry instance which doesn't have to correspond to
487  * a real cache entry in Git's index.
488  */
489 static struct stage_data *insert_stage_data(const char *path,
490                 struct tree *o, struct tree *a, struct tree *b,
491                 struct string_list *entries)
492 {
493         struct string_list_item *item;
494         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
495         get_tree_entry_if_blob(&o->object.oid, path,
496                                &e->stages[1].oid, &e->stages[1].mode);
497         get_tree_entry_if_blob(&a->object.oid, path,
498                                &e->stages[2].oid, &e->stages[2].mode);
499         get_tree_entry_if_blob(&b->object.oid, path,
500                                &e->stages[3].oid, &e->stages[3].mode);
501         item = string_list_insert(entries, path);
502         item->util = e;
503         return e;
504 }
505
506 /*
507  * Create a dictionary mapping file names to stage_data objects. The
508  * dictionary contains one entry for every path with a non-zero stage entry.
509  */
510 static struct string_list *get_unmerged(void)
511 {
512         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
513         int i;
514
515         unmerged->strdup_strings = 1;
516
517         for (i = 0; i < active_nr; i++) {
518                 struct string_list_item *item;
519                 struct stage_data *e;
520                 const struct cache_entry *ce = active_cache[i];
521                 if (!ce_stage(ce))
522                         continue;
523
524                 item = string_list_lookup(unmerged, ce->name);
525                 if (!item) {
526                         item = string_list_insert(unmerged, ce->name);
527                         item->util = xcalloc(1, sizeof(struct stage_data));
528                 }
529                 e = item->util;
530                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
531                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
532         }
533
534         return unmerged;
535 }
536
537 static int string_list_df_name_compare(const char *one, const char *two)
538 {
539         int onelen = strlen(one);
540         int twolen = strlen(two);
541         /*
542          * Here we only care that entries for D/F conflicts are
543          * adjacent, in particular with the file of the D/F conflict
544          * appearing before files below the corresponding directory.
545          * The order of the rest of the list is irrelevant for us.
546          *
547          * To achieve this, we sort with df_name_compare and provide
548          * the mode S_IFDIR so that D/F conflicts will sort correctly.
549          * We use the mode S_IFDIR for everything else for simplicity,
550          * since in other cases any changes in their order due to
551          * sorting cause no problems for us.
552          */
553         int cmp = df_name_compare(one, onelen, S_IFDIR,
554                                   two, twolen, S_IFDIR);
555         /*
556          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
557          * that 'foo' comes before 'foo/bar'.
558          */
559         if (cmp)
560                 return cmp;
561         return onelen - twolen;
562 }
563
564 static void record_df_conflict_files(struct merge_options *o,
565                                      struct string_list *entries)
566 {
567         /* If there is a D/F conflict and the file for such a conflict
568          * currently exists in the working tree, we want to allow it to be
569          * removed to make room for the corresponding directory if needed.
570          * The files underneath the directories of such D/F conflicts will
571          * be processed before the corresponding file involved in the D/F
572          * conflict.  If the D/F directory ends up being removed by the
573          * merge, then we won't have to touch the D/F file.  If the D/F
574          * directory needs to be written to the working copy, then the D/F
575          * file will simply be removed (in make_room_for_path()) to make
576          * room for the necessary paths.  Note that if both the directory
577          * and the file need to be present, then the D/F file will be
578          * reinstated with a new unique name at the time it is processed.
579          */
580         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
581         const char *last_file = NULL;
582         int last_len = 0;
583         int i;
584
585         /*
586          * If we're merging merge-bases, we don't want to bother with
587          * any working directory changes.
588          */
589         if (o->call_depth)
590                 return;
591
592         /* Ensure D/F conflicts are adjacent in the entries list. */
593         for (i = 0; i < entries->nr; i++) {
594                 struct string_list_item *next = &entries->items[i];
595                 string_list_append(&df_sorted_entries, next->string)->util =
596                                    next->util;
597         }
598         df_sorted_entries.cmp = string_list_df_name_compare;
599         string_list_sort(&df_sorted_entries);
600
601         string_list_clear(&o->df_conflict_file_set, 1);
602         for (i = 0; i < df_sorted_entries.nr; i++) {
603                 const char *path = df_sorted_entries.items[i].string;
604                 int len = strlen(path);
605                 struct stage_data *e = df_sorted_entries.items[i].util;
606
607                 /*
608                  * Check if last_file & path correspond to a D/F conflict;
609                  * i.e. whether path is last_file+'/'+<something>.
610                  * If so, record that it's okay to remove last_file to make
611                  * room for path and friends if needed.
612                  */
613                 if (last_file &&
614                     len > last_len &&
615                     memcmp(path, last_file, last_len) == 0 &&
616                     path[last_len] == '/') {
617                         string_list_insert(&o->df_conflict_file_set, last_file);
618                 }
619
620                 /*
621                  * Determine whether path could exist as a file in the
622                  * working directory as a possible D/F conflict.  This
623                  * will only occur when it exists in stage 2 as a
624                  * file.
625                  */
626                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
627                         last_file = path;
628                         last_len = len;
629                 } else {
630                         last_file = NULL;
631                 }
632         }
633         string_list_clear(&df_sorted_entries, 0);
634 }
635
636 struct rename {
637         struct diff_filepair *pair;
638         /*
639          * Purpose of src_entry and dst_entry:
640          *
641          * If 'before' is renamed to 'after' then src_entry will contain
642          * the versions of 'before' from the merge_base, HEAD, and MERGE in
643          * stages 1, 2, and 3; dst_entry will contain the respective
644          * versions of 'after' in corresponding locations.  Thus, we have a
645          * total of six modes and oids, though some will be null.  (Stage 0
646          * is ignored; we're interested in handling conflicts.)
647          *
648          * Since we don't turn on break-rewrites by default, neither
649          * src_entry nor dst_entry can have all three of their stages have
650          * non-null oids, meaning at most four of the six will be non-null.
651          * Also, since this is a rename, both src_entry and dst_entry will
652          * have at least one non-null oid, meaning at least two will be
653          * non-null.  Of the six oids, a typical rename will have three be
654          * non-null.  Only two implies a rename/delete, and four implies a
655          * rename/add.
656          */
657         struct stage_data *src_entry;
658         struct stage_data *dst_entry;
659         unsigned add_turned_into_rename:1;
660         unsigned processed:1;
661 };
662
663 static int update_stages(struct merge_options *opt, const char *path,
664                          const struct diff_filespec *o,
665                          const struct diff_filespec *a,
666                          const struct diff_filespec *b)
667 {
668
669         /*
670          * NOTE: It is usually a bad idea to call update_stages on a path
671          * before calling update_file on that same path, since it can
672          * sometimes lead to spurious "refusing to lose untracked file..."
673          * messages from update_file (via make_room_for path via
674          * would_lose_untracked).  Instead, reverse the order of the calls
675          * (executing update_file first and then update_stages).
676          */
677         int clear = 1;
678         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
679         if (clear)
680                 if (remove_file_from_cache(path))
681                         return -1;
682         if (o)
683                 if (add_cacheinfo(opt, o->mode, &o->oid, path, 1, 0, options))
684                         return -1;
685         if (a)
686                 if (add_cacheinfo(opt, a->mode, &a->oid, path, 2, 0, options))
687                         return -1;
688         if (b)
689                 if (add_cacheinfo(opt, b->mode, &b->oid, path, 3, 0, options))
690                         return -1;
691         return 0;
692 }
693
694 static int update_stages_for_stage_data(struct merge_options *opt,
695                                         const char *path,
696                                         const struct stage_data *stage_data)
697 {
698         struct diff_filespec o, a, b;
699
700         o.mode = stage_data->stages[1].mode;
701         oidcpy(&o.oid, &stage_data->stages[1].oid);
702
703         a.mode = stage_data->stages[2].mode;
704         oidcpy(&a.oid, &stage_data->stages[2].oid);
705
706         b.mode = stage_data->stages[3].mode;
707         oidcpy(&b.oid, &stage_data->stages[3].oid);
708
709         return update_stages(opt, path,
710                              is_null_oid(&o.oid) ? NULL : &o,
711                              is_null_oid(&a.oid) ? NULL : &a,
712                              is_null_oid(&b.oid) ? NULL : &b);
713 }
714
715 static void update_entry(struct stage_data *entry,
716                          struct diff_filespec *o,
717                          struct diff_filespec *a,
718                          struct diff_filespec *b)
719 {
720         entry->processed = 0;
721         entry->stages[1].mode = o->mode;
722         entry->stages[2].mode = a->mode;
723         entry->stages[3].mode = b->mode;
724         oidcpy(&entry->stages[1].oid, &o->oid);
725         oidcpy(&entry->stages[2].oid, &a->oid);
726         oidcpy(&entry->stages[3].oid, &b->oid);
727 }
728
729 static int remove_file(struct merge_options *o, int clean,
730                        const char *path, int no_wd)
731 {
732         int update_cache = o->call_depth || clean;
733         int update_working_directory = !o->call_depth && !no_wd;
734
735         if (update_cache) {
736                 if (remove_file_from_cache(path))
737                         return -1;
738         }
739         if (update_working_directory) {
740                 if (ignore_case) {
741                         struct cache_entry *ce;
742                         ce = cache_file_exists(path, strlen(path), ignore_case);
743                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
744                                 return 0;
745                 }
746                 if (remove_path(path))
747                         return -1;
748         }
749         return 0;
750 }
751
752 /* add a string to a strbuf, but converting "/" to "_" */
753 static void add_flattened_path(struct strbuf *out, const char *s)
754 {
755         size_t i = out->len;
756         strbuf_addstr(out, s);
757         for (; i < out->len; i++)
758                 if (out->buf[i] == '/')
759                         out->buf[i] = '_';
760 }
761
762 static char *unique_path(struct merge_options *o, const char *path, const char *branch)
763 {
764         struct path_hashmap_entry *entry;
765         struct strbuf newpath = STRBUF_INIT;
766         int suffix = 0;
767         size_t base_len;
768
769         strbuf_addf(&newpath, "%s~", path);
770         add_flattened_path(&newpath, branch);
771
772         base_len = newpath.len;
773         while (hashmap_get_from_hash(&o->current_file_dir_set,
774                                      path_hash(newpath.buf), newpath.buf) ||
775                (!o->call_depth && file_exists(newpath.buf))) {
776                 strbuf_setlen(&newpath, base_len);
777                 strbuf_addf(&newpath, "_%d", suffix++);
778         }
779
780         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
781         hashmap_entry_init(entry, path_hash(entry->path));
782         hashmap_add(&o->current_file_dir_set, entry);
783         return strbuf_detach(&newpath, NULL);
784 }
785
786 /**
787  * Check whether a directory in the index is in the way of an incoming
788  * file.  Return 1 if so.  If check_working_copy is non-zero, also
789  * check the working directory.  If empty_ok is non-zero, also return
790  * 0 in the case where the working-tree dir exists but is empty.
791  */
792 static int dir_in_way(const char *path, int check_working_copy, int empty_ok)
793 {
794         int pos;
795         struct strbuf dirpath = STRBUF_INIT;
796         struct stat st;
797
798         strbuf_addstr(&dirpath, path);
799         strbuf_addch(&dirpath, '/');
800
801         pos = cache_name_pos(dirpath.buf, dirpath.len);
802
803         if (pos < 0)
804                 pos = -1 - pos;
805         if (pos < active_nr &&
806             !strncmp(dirpath.buf, active_cache[pos]->name, dirpath.len)) {
807                 strbuf_release(&dirpath);
808                 return 1;
809         }
810
811         strbuf_release(&dirpath);
812         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
813                 !(empty_ok && is_empty_dir(path));
814 }
815
816 /*
817  * Returns whether path was tracked in the index before the merge started,
818  * and its oid and mode match the specified values
819  */
820 static int was_tracked_and_matches(struct merge_options *o, const char *path,
821                                    const struct object_id *oid, unsigned mode)
822 {
823         int pos = index_name_pos(&o->orig_index, path, strlen(path));
824         struct cache_entry *ce;
825
826         if (0 > pos)
827                 /* we were not tracking this path before the merge */
828                 return 0;
829
830         /* See if the file we were tracking before matches */
831         ce = o->orig_index.cache[pos];
832         return (oid_eq(&ce->oid, oid) && ce->ce_mode == mode);
833 }
834
835 /*
836  * Returns whether path was tracked in the index before the merge started
837  */
838 static int was_tracked(struct merge_options *o, const char *path)
839 {
840         int pos = index_name_pos(&o->orig_index, path, strlen(path));
841
842         if (0 <= pos)
843                 /* we were tracking this path before the merge */
844                 return 1;
845
846         return 0;
847 }
848
849 static int would_lose_untracked(const char *path)
850 {
851         /*
852          * This may look like it can be simplified to:
853          *   return !was_tracked(o, path) && file_exists(path)
854          * but it can't.  This function needs to know whether path was in
855          * the working tree due to EITHER having been tracked in the index
856          * before the merge OR having been put into the working copy and
857          * index by unpack_trees().  Due to that either-or requirement, we
858          * check the current index instead of the original one.
859          *
860          * Note that we do not need to worry about merge-recursive itself
861          * updating the index after unpack_trees() and before calling this
862          * function, because we strictly require all code paths in
863          * merge-recursive to update the working tree first and the index
864          * second.  Doing otherwise would break
865          * update_file()/would_lose_untracked(); see every comment in this
866          * file which mentions "update_stages".
867          */
868         int pos = cache_name_pos(path, strlen(path));
869
870         if (pos < 0)
871                 pos = -1 - pos;
872         while (pos < active_nr &&
873                !strcmp(path, active_cache[pos]->name)) {
874                 /*
875                  * If stage #0, it is definitely tracked.
876                  * If it has stage #2 then it was tracked
877                  * before this merge started.  All other
878                  * cases the path was not tracked.
879                  */
880                 switch (ce_stage(active_cache[pos])) {
881                 case 0:
882                 case 2:
883                         return 0;
884                 }
885                 pos++;
886         }
887         return file_exists(path);
888 }
889
890 static int was_dirty(struct merge_options *o, const char *path)
891 {
892         struct cache_entry *ce;
893         int dirty = 1;
894
895         if (o->call_depth || !was_tracked(o, path))
896                 return !dirty;
897
898         ce = index_file_exists(o->unpack_opts.src_index,
899                                path, strlen(path), ignore_case);
900         dirty = verify_uptodate(ce, &o->unpack_opts) != 0;
901         return dirty;
902 }
903
904 static int make_room_for_path(struct merge_options *o, const char *path)
905 {
906         int status, i;
907         const char *msg = _("failed to create path '%s'%s");
908
909         /* Unlink any D/F conflict files that are in the way */
910         for (i = 0; i < o->df_conflict_file_set.nr; i++) {
911                 const char *df_path = o->df_conflict_file_set.items[i].string;
912                 size_t pathlen = strlen(path);
913                 size_t df_pathlen = strlen(df_path);
914                 if (df_pathlen < pathlen &&
915                     path[df_pathlen] == '/' &&
916                     strncmp(path, df_path, df_pathlen) == 0) {
917                         output(o, 3,
918                                _("Removing %s to make room for subdirectory\n"),
919                                df_path);
920                         unlink(df_path);
921                         unsorted_string_list_delete_item(&o->df_conflict_file_set,
922                                                          i, 0);
923                         break;
924                 }
925         }
926
927         /* Make sure leading directories are created */
928         status = safe_create_leading_directories_const(path);
929         if (status) {
930                 if (status == SCLD_EXISTS)
931                         /* something else exists */
932                         return err(o, msg, path, _(": perhaps a D/F conflict?"));
933                 return err(o, msg, path, "");
934         }
935
936         /*
937          * Do not unlink a file in the work tree if we are not
938          * tracking it.
939          */
940         if (would_lose_untracked(path))
941                 return err(o, _("refusing to lose untracked file at '%s'"),
942                            path);
943
944         /* Successful unlink is good.. */
945         if (!unlink(path))
946                 return 0;
947         /* .. and so is no existing file */
948         if (errno == ENOENT)
949                 return 0;
950         /* .. but not some other error (who really cares what?) */
951         return err(o, msg, path, _(": perhaps a D/F conflict?"));
952 }
953
954 static int update_file_flags(struct merge_options *o,
955                              const struct object_id *oid,
956                              unsigned mode,
957                              const char *path,
958                              int update_cache,
959                              int update_wd)
960 {
961         int ret = 0;
962
963         if (o->call_depth)
964                 update_wd = 0;
965
966         if (update_wd) {
967                 enum object_type type;
968                 void *buf;
969                 unsigned long size;
970
971                 if (S_ISGITLINK(mode)) {
972                         /*
973                          * We may later decide to recursively descend into
974                          * the submodule directory and update its index
975                          * and/or work tree, but we do not do that now.
976                          */
977                         update_wd = 0;
978                         goto update_index;
979                 }
980
981                 buf = read_object_file(oid, &type, &size);
982                 if (!buf)
983                         return err(o, _("cannot read object %s '%s'"), oid_to_hex(oid), path);
984                 if (type != OBJ_BLOB) {
985                         ret = err(o, _("blob expected for %s '%s'"), oid_to_hex(oid), path);
986                         goto free_buf;
987                 }
988                 if (S_ISREG(mode)) {
989                         struct strbuf strbuf = STRBUF_INIT;
990                         if (convert_to_working_tree(&the_index, path, buf, size, &strbuf)) {
991                                 free(buf);
992                                 size = strbuf.len;
993                                 buf = strbuf_detach(&strbuf, NULL);
994                         }
995                 }
996
997                 if (make_room_for_path(o, path) < 0) {
998                         update_wd = 0;
999                         goto free_buf;
1000                 }
1001                 if (S_ISREG(mode) || (!has_symlinks && S_ISLNK(mode))) {
1002                         int fd;
1003                         if (mode & 0100)
1004                                 mode = 0777;
1005                         else
1006                                 mode = 0666;
1007                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
1008                         if (fd < 0) {
1009                                 ret = err(o, _("failed to open '%s': %s"),
1010                                           path, strerror(errno));
1011                                 goto free_buf;
1012                         }
1013                         write_in_full(fd, buf, size);
1014                         close(fd);
1015                 } else if (S_ISLNK(mode)) {
1016                         char *lnk = xmemdupz(buf, size);
1017                         safe_create_leading_directories_const(path);
1018                         unlink(path);
1019                         if (symlink(lnk, path))
1020                                 ret = err(o, _("failed to symlink '%s': %s"),
1021                                           path, strerror(errno));
1022                         free(lnk);
1023                 } else
1024                         ret = err(o,
1025                                   _("do not know what to do with %06o %s '%s'"),
1026                                   mode, oid_to_hex(oid), path);
1027         free_buf:
1028                 free(buf);
1029         }
1030 update_index:
1031         if (!ret && update_cache)
1032                 if (add_cacheinfo(o, mode, oid, path, 0, update_wd,
1033                                   ADD_CACHE_OK_TO_ADD))
1034                         return -1;
1035         return ret;
1036 }
1037
1038 static int update_file(struct merge_options *o,
1039                        int clean,
1040                        const struct object_id *oid,
1041                        unsigned mode,
1042                        const char *path)
1043 {
1044         return update_file_flags(o, oid, mode, path, o->call_depth || clean, !o->call_depth);
1045 }
1046
1047 /* Low level file merging, update and removal */
1048
1049 struct merge_file_info {
1050         struct object_id oid;
1051         unsigned mode;
1052         unsigned clean:1,
1053                  merge:1;
1054 };
1055
1056 static int merge_3way(struct merge_options *o,
1057                       mmbuffer_t *result_buf,
1058                       const struct diff_filespec *one,
1059                       const struct diff_filespec *a,
1060                       const struct diff_filespec *b,
1061                       const char *branch1,
1062                       const char *branch2)
1063 {
1064         mmfile_t orig, src1, src2;
1065         struct ll_merge_options ll_opts = {0};
1066         char *base_name, *name1, *name2;
1067         int merge_status;
1068
1069         ll_opts.renormalize = o->renormalize;
1070         ll_opts.xdl_opts = o->xdl_opts;
1071
1072         if (o->call_depth) {
1073                 ll_opts.virtual_ancestor = 1;
1074                 ll_opts.variant = 0;
1075         } else {
1076                 switch (o->recursive_variant) {
1077                 case MERGE_RECURSIVE_OURS:
1078                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
1079                         break;
1080                 case MERGE_RECURSIVE_THEIRS:
1081                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
1082                         break;
1083                 default:
1084                         ll_opts.variant = 0;
1085                         break;
1086                 }
1087         }
1088
1089         if (strcmp(a->path, b->path) ||
1090             (o->ancestor != NULL && strcmp(a->path, one->path) != 0)) {
1091                 base_name = o->ancestor == NULL ? NULL :
1092                         mkpathdup("%s:%s", o->ancestor, one->path);
1093                 name1 = mkpathdup("%s:%s", branch1, a->path);
1094                 name2 = mkpathdup("%s:%s", branch2, b->path);
1095         } else {
1096                 base_name = o->ancestor == NULL ? NULL :
1097                         mkpathdup("%s", o->ancestor);
1098                 name1 = mkpathdup("%s", branch1);
1099                 name2 = mkpathdup("%s", branch2);
1100         }
1101
1102         read_mmblob(&orig, &one->oid);
1103         read_mmblob(&src1, &a->oid);
1104         read_mmblob(&src2, &b->oid);
1105
1106         merge_status = ll_merge(result_buf, a->path, &orig, base_name,
1107                                 &src1, name1, &src2, name2,
1108                                 &the_index, &ll_opts);
1109
1110         free(base_name);
1111         free(name1);
1112         free(name2);
1113         free(orig.ptr);
1114         free(src1.ptr);
1115         free(src2.ptr);
1116         return merge_status;
1117 }
1118
1119 static int find_first_merges(struct object_array *result, const char *path,
1120                              struct commit *a, struct commit *b)
1121 {
1122         int i, j;
1123         struct object_array merges = OBJECT_ARRAY_INIT;
1124         struct commit *commit;
1125         int contains_another;
1126
1127         char merged_revision[42];
1128         const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
1129                                    "--all", merged_revision, NULL };
1130         struct rev_info revs;
1131         struct setup_revision_opt rev_opts;
1132
1133         memset(result, 0, sizeof(struct object_array));
1134         memset(&rev_opts, 0, sizeof(rev_opts));
1135
1136         /* get all revisions that merge commit a */
1137         xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
1138                   oid_to_hex(&a->object.oid));
1139         repo_init_revisions(the_repository, &revs, NULL);
1140         rev_opts.submodule = path;
1141         /* FIXME: can't handle linked worktrees in submodules yet */
1142         revs.single_worktree = path != NULL;
1143         setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
1144
1145         /* save all revisions from the above list that contain b */
1146         if (prepare_revision_walk(&revs))
1147                 die("revision walk setup failed");
1148         while ((commit = get_revision(&revs)) != NULL) {
1149                 struct object *o = &(commit->object);
1150                 if (in_merge_bases(b, commit))
1151                         add_object_array(o, NULL, &merges);
1152         }
1153         reset_revision_walk();
1154
1155         /* Now we've got all merges that contain a and b. Prune all
1156          * merges that contain another found merge and save them in
1157          * result.
1158          */
1159         for (i = 0; i < merges.nr; i++) {
1160                 struct commit *m1 = (struct commit *) merges.objects[i].item;
1161
1162                 contains_another = 0;
1163                 for (j = 0; j < merges.nr; j++) {
1164                         struct commit *m2 = (struct commit *) merges.objects[j].item;
1165                         if (i != j && in_merge_bases(m2, m1)) {
1166                                 contains_another = 1;
1167                                 break;
1168                         }
1169                 }
1170
1171                 if (!contains_another)
1172                         add_object_array(merges.objects[i].item, NULL, result);
1173         }
1174
1175         object_array_clear(&merges);
1176         return result->nr;
1177 }
1178
1179 static void print_commit(struct commit *commit)
1180 {
1181         struct strbuf sb = STRBUF_INIT;
1182         struct pretty_print_context ctx = {0};
1183         ctx.date_mode.type = DATE_NORMAL;
1184         format_commit_message(commit, " %h: %m %s", &sb, &ctx);
1185         fprintf(stderr, "%s\n", sb.buf);
1186         strbuf_release(&sb);
1187 }
1188
1189 static int merge_submodule(struct merge_options *o,
1190                            struct object_id *result, const char *path,
1191                            const struct object_id *base, const struct object_id *a,
1192                            const struct object_id *b)
1193 {
1194         struct commit *commit_base, *commit_a, *commit_b;
1195         int parent_count;
1196         struct object_array merges;
1197
1198         int i;
1199         int search = !o->call_depth;
1200
1201         /* store a in result in case we fail */
1202         oidcpy(result, a);
1203
1204         /* we can not handle deletion conflicts */
1205         if (is_null_oid(base))
1206                 return 0;
1207         if (is_null_oid(a))
1208                 return 0;
1209         if (is_null_oid(b))
1210                 return 0;
1211
1212         if (add_submodule_odb(path)) {
1213                 output(o, 1, _("Failed to merge submodule %s (not checked out)"), path);
1214                 return 0;
1215         }
1216
1217         if (!(commit_base = lookup_commit_reference(the_repository, base)) ||
1218             !(commit_a = lookup_commit_reference(the_repository, a)) ||
1219             !(commit_b = lookup_commit_reference(the_repository, b))) {
1220                 output(o, 1, _("Failed to merge submodule %s (commits not present)"), path);
1221                 return 0;
1222         }
1223
1224         /* check whether both changes are forward */
1225         if (!in_merge_bases(commit_base, commit_a) ||
1226             !in_merge_bases(commit_base, commit_b)) {
1227                 output(o, 1, _("Failed to merge submodule %s (commits don't follow merge-base)"), path);
1228                 return 0;
1229         }
1230
1231         /* Case #1: a is contained in b or vice versa */
1232         if (in_merge_bases(commit_a, commit_b)) {
1233                 oidcpy(result, b);
1234                 if (show(o, 3)) {
1235                         output(o, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1236                         output_commit_title(o, commit_b);
1237                 } else if (show(o, 2))
1238                         output(o, 2, _("Fast-forwarding submodule %s"), path);
1239                 else
1240                         ; /* no output */
1241
1242                 return 1;
1243         }
1244         if (in_merge_bases(commit_b, commit_a)) {
1245                 oidcpy(result, a);
1246                 if (show(o, 3)) {
1247                         output(o, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1248                         output_commit_title(o, commit_a);
1249                 } else if (show(o, 2))
1250                         output(o, 2, _("Fast-forwarding submodule %s"), path);
1251                 else
1252                         ; /* no output */
1253
1254                 return 1;
1255         }
1256
1257         /*
1258          * Case #2: There are one or more merges that contain a and b in
1259          * the submodule. If there is only one, then present it as a
1260          * suggestion to the user, but leave it marked unmerged so the
1261          * user needs to confirm the resolution.
1262          */
1263
1264         /* Skip the search if makes no sense to the calling context.  */
1265         if (!search)
1266                 return 0;
1267
1268         /* find commit which merges them */
1269         parent_count = find_first_merges(&merges, path, commit_a, commit_b);
1270         switch (parent_count) {
1271         case 0:
1272                 output(o, 1, _("Failed to merge submodule %s (merge following commits not found)"), path);
1273                 break;
1274
1275         case 1:
1276                 output(o, 1, _("Failed to merge submodule %s (not fast-forward)"), path);
1277                 output(o, 2, _("Found a possible merge resolution for the submodule:\n"));
1278                 print_commit((struct commit *) merges.objects[0].item);
1279                 output(o, 2, _(
1280                        "If this is correct simply add it to the index "
1281                        "for example\n"
1282                        "by using:\n\n"
1283                        "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
1284                        "which will accept this suggestion.\n"),
1285                        oid_to_hex(&merges.objects[0].item->oid), path);
1286                 break;
1287
1288         default:
1289                 output(o, 1, _("Failed to merge submodule %s (multiple merges found)"), path);
1290                 for (i = 0; i < merges.nr; i++)
1291                         print_commit((struct commit *) merges.objects[i].item);
1292         }
1293
1294         object_array_clear(&merges);
1295         return 0;
1296 }
1297
1298 static int merge_mode_and_contents(struct merge_options *o,
1299                                    const struct diff_filespec *one,
1300                                    const struct diff_filespec *a,
1301                                    const struct diff_filespec *b,
1302                                    const char *filename,
1303                                    const char *branch1,
1304                                    const char *branch2,
1305                                    struct merge_file_info *result)
1306 {
1307         if (o->branch1 != branch1) {
1308                 /*
1309                  * It's weird getting a reverse merge with HEAD on the bottom
1310                  * side of the conflict markers and the other branch on the
1311                  * top.  Fix that.
1312                  */
1313                 return merge_mode_and_contents(o, one, b, a,
1314                                                filename,
1315                                                branch2, branch1, result);
1316         }
1317
1318         result->merge = 0;
1319         result->clean = 1;
1320
1321         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
1322                 result->clean = 0;
1323                 if (S_ISREG(a->mode)) {
1324                         result->mode = a->mode;
1325                         oidcpy(&result->oid, &a->oid);
1326                 } else {
1327                         result->mode = b->mode;
1328                         oidcpy(&result->oid, &b->oid);
1329                 }
1330         } else {
1331                 if (!oid_eq(&a->oid, &one->oid) && !oid_eq(&b->oid, &one->oid))
1332                         result->merge = 1;
1333
1334                 /*
1335                  * Merge modes
1336                  */
1337                 if (a->mode == b->mode || a->mode == one->mode)
1338                         result->mode = b->mode;
1339                 else {
1340                         result->mode = a->mode;
1341                         if (b->mode != one->mode) {
1342                                 result->clean = 0;
1343                                 result->merge = 1;
1344                         }
1345                 }
1346
1347                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &one->oid))
1348                         oidcpy(&result->oid, &b->oid);
1349                 else if (oid_eq(&b->oid, &one->oid))
1350                         oidcpy(&result->oid, &a->oid);
1351                 else if (S_ISREG(a->mode)) {
1352                         mmbuffer_t result_buf;
1353                         int ret = 0, merge_status;
1354
1355                         merge_status = merge_3way(o, &result_buf, one, a, b,
1356                                                   branch1, branch2);
1357
1358                         if ((merge_status < 0) || !result_buf.ptr)
1359                                 ret = err(o, _("Failed to execute internal merge"));
1360
1361                         if (!ret &&
1362                             write_object_file(result_buf.ptr, result_buf.size,
1363                                               blob_type, &result->oid))
1364                                 ret = err(o, _("Unable to add %s to database"),
1365                                           a->path);
1366
1367                         free(result_buf.ptr);
1368                         if (ret)
1369                                 return ret;
1370                         result->clean = (merge_status == 0);
1371                 } else if (S_ISGITLINK(a->mode)) {
1372                         result->clean = merge_submodule(o, &result->oid,
1373                                                         one->path,
1374                                                         &one->oid,
1375                                                         &a->oid,
1376                                                         &b->oid);
1377                 } else if (S_ISLNK(a->mode)) {
1378                         switch (o->recursive_variant) {
1379                         case MERGE_RECURSIVE_NORMAL:
1380                                 oidcpy(&result->oid, &a->oid);
1381                                 if (!oid_eq(&a->oid, &b->oid))
1382                                         result->clean = 0;
1383                                 break;
1384                         case MERGE_RECURSIVE_OURS:
1385                                 oidcpy(&result->oid, &a->oid);
1386                                 break;
1387                         case MERGE_RECURSIVE_THEIRS:
1388                                 oidcpy(&result->oid, &b->oid);
1389                                 break;
1390                         }
1391                 } else
1392                         BUG("unsupported object type in the tree");
1393         }
1394
1395         if (result->merge)
1396                 output(o, 2, _("Auto-merging %s"), filename);
1397
1398         return 0;
1399 }
1400
1401 static int handle_rename_via_dir(struct merge_options *o,
1402                                  struct diff_filepair *pair,
1403                                  const char *rename_branch,
1404                                  const char *other_branch)
1405 {
1406         /*
1407          * Handle file adds that need to be renamed due to directory rename
1408          * detection.  This differs from handle_rename_normal, because
1409          * there is no content merge to do; just move the file into the
1410          * desired final location.
1411          */
1412         const struct diff_filespec *dest = pair->two;
1413
1414         if (!o->call_depth && would_lose_untracked(dest->path)) {
1415                 char *alt_path = unique_path(o, dest->path, rename_branch);
1416
1417                 output(o, 1, _("Error: Refusing to lose untracked file at %s; "
1418                                "writing to %s instead."),
1419                        dest->path, alt_path);
1420                 /*
1421                  * Write the file in worktree at alt_path, but not in the
1422                  * index.  Instead, write to dest->path for the index but
1423                  * only at the higher appropriate stage.
1424                  */
1425                 if (update_file(o, 0, &dest->oid, dest->mode, alt_path))
1426                         return -1;
1427                 free(alt_path);
1428                 return update_stages(o, dest->path, NULL,
1429                                      rename_branch == o->branch1 ? dest : NULL,
1430                                      rename_branch == o->branch1 ? NULL : dest);
1431         }
1432
1433         /* Update dest->path both in index and in worktree */
1434         if (update_file(o, 1, &dest->oid, dest->mode, dest->path))
1435                 return -1;
1436         return 0;
1437 }
1438
1439 static int handle_change_delete(struct merge_options *o,
1440                                 const char *path, const char *old_path,
1441                                 const struct object_id *o_oid, int o_mode,
1442                                 const struct object_id *changed_oid,
1443                                 int changed_mode,
1444                                 const char *change_branch,
1445                                 const char *delete_branch,
1446                                 const char *change, const char *change_past)
1447 {
1448         char *alt_path = NULL;
1449         const char *update_path = path;
1450         int ret = 0;
1451
1452         if (dir_in_way(path, !o->call_depth, 0) ||
1453             (!o->call_depth && would_lose_untracked(path))) {
1454                 update_path = alt_path = unique_path(o, path, change_branch);
1455         }
1456
1457         if (o->call_depth) {
1458                 /*
1459                  * We cannot arbitrarily accept either a_sha or b_sha as
1460                  * correct; since there is no true "middle point" between
1461                  * them, simply reuse the base version for virtual merge base.
1462                  */
1463                 ret = remove_file_from_cache(path);
1464                 if (!ret)
1465                         ret = update_file(o, 0, o_oid, o_mode, update_path);
1466         } else {
1467                 /*
1468                  * Despite the four nearly duplicate messages and argument
1469                  * lists below and the ugliness of the nested if-statements,
1470                  * having complete messages makes the job easier for
1471                  * translators.
1472                  *
1473                  * The slight variance among the cases is due to the fact
1474                  * that:
1475                  *   1) directory/file conflicts (in effect if
1476                  *      !alt_path) could cause us to need to write the
1477                  *      file to a different path.
1478                  *   2) renames (in effect if !old_path) could mean that
1479                  *      there are two names for the path that the user
1480                  *      may know the file by.
1481                  */
1482                 if (!alt_path) {
1483                         if (!old_path) {
1484                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1485                                        "and %s in %s. Version %s of %s left in tree."),
1486                                        change, path, delete_branch, change_past,
1487                                        change_branch, change_branch, path);
1488                         } else {
1489                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1490                                        "and %s to %s in %s. Version %s of %s left in tree."),
1491                                        change, old_path, delete_branch, change_past, path,
1492                                        change_branch, change_branch, path);
1493                         }
1494                 } else {
1495                         if (!old_path) {
1496                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1497                                        "and %s in %s. Version %s of %s left in tree at %s."),
1498                                        change, path, delete_branch, change_past,
1499                                        change_branch, change_branch, path, alt_path);
1500                         } else {
1501                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1502                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1503                                        change, old_path, delete_branch, change_past, path,
1504                                        change_branch, change_branch, path, alt_path);
1505                         }
1506                 }
1507                 /*
1508                  * No need to call update_file() on path when change_branch ==
1509                  * o->branch1 && !alt_path, since that would needlessly touch
1510                  * path.  We could call update_file_flags() with update_cache=0
1511                  * and update_wd=0, but that's a no-op.
1512                  */
1513                 if (change_branch != o->branch1 || alt_path)
1514                         ret = update_file(o, 0, changed_oid, changed_mode, update_path);
1515         }
1516         free(alt_path);
1517
1518         return ret;
1519 }
1520
1521 static int handle_rename_delete(struct merge_options *o,
1522                                 struct diff_filepair *pair,
1523                                 const char *rename_branch,
1524                                 const char *delete_branch)
1525 {
1526         const struct diff_filespec *orig = pair->one;
1527         const struct diff_filespec *dest = pair->two;
1528
1529         if (handle_change_delete(o,
1530                                  o->call_depth ? orig->path : dest->path,
1531                                  o->call_depth ? NULL : orig->path,
1532                                  &orig->oid, orig->mode,
1533                                  &dest->oid, dest->mode,
1534                                  rename_branch, delete_branch,
1535                                  _("rename"), _("renamed")))
1536                 return -1;
1537
1538         if (o->call_depth)
1539                 return remove_file_from_cache(dest->path);
1540         else
1541                 return update_stages(o, dest->path, NULL,
1542                                      rename_branch == o->branch1 ? dest : NULL,
1543                                      rename_branch == o->branch1 ? NULL : dest);
1544 }
1545
1546 static struct diff_filespec *filespec_from_entry(struct diff_filespec *target,
1547                                                  struct stage_data *entry,
1548                                                  int stage)
1549 {
1550         struct object_id *oid = &entry->stages[stage].oid;
1551         unsigned mode = entry->stages[stage].mode;
1552         if (mode == 0 || is_null_oid(oid))
1553                 return NULL;
1554         oidcpy(&target->oid, oid);
1555         target->mode = mode;
1556         return target;
1557 }
1558
1559 static int handle_file(struct merge_options *o,
1560                         struct diff_filespec *rename,
1561                         int stage,
1562                         struct rename_conflict_info *ci)
1563 {
1564         char *dst_name = rename->path;
1565         struct stage_data *dst_entry;
1566         const char *cur_branch, *other_branch;
1567         struct diff_filespec other;
1568         struct diff_filespec *add;
1569         int ret;
1570
1571         if (stage == 2) {
1572                 dst_entry = ci->dst_entry1;
1573                 cur_branch = ci->branch1;
1574                 other_branch = ci->branch2;
1575         } else {
1576                 dst_entry = ci->dst_entry2;
1577                 cur_branch = ci->branch2;
1578                 other_branch = ci->branch1;
1579         }
1580
1581         add = filespec_from_entry(&other, dst_entry, stage ^ 1);
1582         if (add) {
1583                 int ren_src_was_dirty = was_dirty(o, rename->path);
1584                 char *add_name = unique_path(o, rename->path, other_branch);
1585                 if (update_file(o, 0, &add->oid, add->mode, add_name))
1586                         return -1;
1587
1588                 if (ren_src_was_dirty) {
1589                         output(o, 1, _("Refusing to lose dirty file at %s"),
1590                                rename->path);
1591                 }
1592                 /*
1593                  * Because the double negatives somehow keep confusing me...
1594                  *    1) update_wd iff !ren_src_was_dirty.
1595                  *    2) no_wd iff !update_wd
1596                  *    3) so, no_wd == !!ren_src_was_dirty == ren_src_was_dirty
1597                  */
1598                 remove_file(o, 0, rename->path, ren_src_was_dirty);
1599                 dst_name = unique_path(o, rename->path, cur_branch);
1600         } else {
1601                 if (dir_in_way(rename->path, !o->call_depth, 0)) {
1602                         dst_name = unique_path(o, rename->path, cur_branch);
1603                         output(o, 1, _("%s is a directory in %s adding as %s instead"),
1604                                rename->path, other_branch, dst_name);
1605                 } else if (!o->call_depth &&
1606                            would_lose_untracked(rename->path)) {
1607                         dst_name = unique_path(o, rename->path, cur_branch);
1608                         output(o, 1, _("Refusing to lose untracked file at %s; "
1609                                        "adding as %s instead"),
1610                                rename->path, dst_name);
1611                 }
1612         }
1613         if ((ret = update_file(o, 0, &rename->oid, rename->mode, dst_name)))
1614                 ; /* fall through, do allow dst_name to be released */
1615         else if (stage == 2)
1616                 ret = update_stages(o, rename->path, NULL, rename, add);
1617         else
1618                 ret = update_stages(o, rename->path, NULL, add, rename);
1619
1620         if (dst_name != rename->path)
1621                 free(dst_name);
1622
1623         return ret;
1624 }
1625
1626 static int handle_rename_rename_1to2(struct merge_options *o,
1627                                      struct rename_conflict_info *ci)
1628 {
1629         /* One file was renamed in both branches, but to different names. */
1630         struct diff_filespec *one = ci->pair1->one;
1631         struct diff_filespec *a = ci->pair1->two;
1632         struct diff_filespec *b = ci->pair2->two;
1633
1634         output(o, 1, _("CONFLICT (rename/rename): "
1635                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1636                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1637                one->path, a->path, ci->branch1,
1638                one->path, b->path, ci->branch2,
1639                o->call_depth ? _(" (left unresolved)") : "");
1640         if (o->call_depth) {
1641                 struct merge_file_info mfi;
1642                 struct diff_filespec other;
1643                 struct diff_filespec *add;
1644                 if (merge_mode_and_contents(o, one, a, b, one->path,
1645                                             ci->branch1, ci->branch2, &mfi))
1646                         return -1;
1647
1648                 /*
1649                  * FIXME: For rename/add-source conflicts (if we could detect
1650                  * such), this is wrong.  We should instead find a unique
1651                  * pathname and then either rename the add-source file to that
1652                  * unique path, or use that unique path instead of src here.
1653                  */
1654                 if (update_file(o, 0, &mfi.oid, mfi.mode, one->path))
1655                         return -1;
1656
1657                 /*
1658                  * Above, we put the merged content at the merge-base's
1659                  * path.  Now we usually need to delete both a->path and
1660                  * b->path.  However, the rename on each side of the merge
1661                  * could also be involved in a rename/add conflict.  In
1662                  * such cases, we should keep the added file around,
1663                  * resolving the conflict at that path in its favor.
1664                  */
1665                 add = filespec_from_entry(&other, ci->dst_entry1, 2 ^ 1);
1666                 if (add) {
1667                         if (update_file(o, 0, &add->oid, add->mode, a->path))
1668                                 return -1;
1669                 }
1670                 else
1671                         remove_file_from_cache(a->path);
1672                 add = filespec_from_entry(&other, ci->dst_entry2, 3 ^ 1);
1673                 if (add) {
1674                         if (update_file(o, 0, &add->oid, add->mode, b->path))
1675                                 return -1;
1676                 }
1677                 else
1678                         remove_file_from_cache(b->path);
1679         } else if (handle_file(o, a, 2, ci) || handle_file(o, b, 3, ci))
1680                 return -1;
1681
1682         return 0;
1683 }
1684
1685 static int handle_rename_rename_2to1(struct merge_options *o,
1686                                      struct rename_conflict_info *ci)
1687 {
1688         /* Two files, a & b, were renamed to the same thing, c. */
1689         struct diff_filespec *a = ci->pair1->one;
1690         struct diff_filespec *b = ci->pair2->one;
1691         struct diff_filespec *c1 = ci->pair1->two;
1692         struct diff_filespec *c2 = ci->pair2->two;
1693         char *path = c1->path; /* == c2->path */
1694         char *path_side_1_desc;
1695         char *path_side_2_desc;
1696         struct merge_file_info mfi_c1;
1697         struct merge_file_info mfi_c2;
1698         int ret;
1699
1700         output(o, 1, _("CONFLICT (rename/rename): "
1701                "Rename %s->%s in %s. "
1702                "Rename %s->%s in %s"),
1703                a->path, c1->path, ci->branch1,
1704                b->path, c2->path, ci->branch2);
1705
1706         remove_file(o, 1, a->path, o->call_depth || would_lose_untracked(a->path));
1707         remove_file(o, 1, b->path, o->call_depth || would_lose_untracked(b->path));
1708
1709         path_side_1_desc = xstrfmt("version of %s from %s", path, a->path);
1710         path_side_2_desc = xstrfmt("version of %s from %s", path, b->path);
1711         if (merge_mode_and_contents(o, a, c1, &ci->ren1_other, path_side_1_desc,
1712                                     o->branch1, o->branch2, &mfi_c1) ||
1713             merge_mode_and_contents(o, b, &ci->ren2_other, c2, path_side_2_desc,
1714                                     o->branch1, o->branch2, &mfi_c2))
1715                 return -1;
1716         free(path_side_1_desc);
1717         free(path_side_2_desc);
1718
1719         if (o->call_depth) {
1720                 /*
1721                  * If mfi_c1.clean && mfi_c2.clean, then it might make
1722                  * sense to do a two-way merge of those results.  But, I
1723                  * think in all cases, it makes sense to have the virtual
1724                  * merge base just undo the renames; they can be detected
1725                  * again later for the non-recursive merge.
1726                  */
1727                 remove_file(o, 0, path, 0);
1728                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, a->path);
1729                 if (!ret)
1730                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1731                                           b->path);
1732         } else {
1733                 char *new_path1 = unique_path(o, path, ci->branch1);
1734                 char *new_path2 = unique_path(o, path, ci->branch2);
1735                 output(o, 1, _("Renaming %s to %s and %s to %s instead"),
1736                        a->path, new_path1, b->path, new_path2);
1737                 if (was_dirty(o, path))
1738                         output(o, 1, _("Refusing to lose dirty file at %s"),
1739                                path);
1740                 else if (would_lose_untracked(path))
1741                         /*
1742                          * Only way we get here is if both renames were from
1743                          * a directory rename AND user had an untracked file
1744                          * at the location where both files end up after the
1745                          * two directory renames.  See testcase 10d of t6043.
1746                          */
1747                         output(o, 1, _("Refusing to lose untracked file at "
1748                                        "%s, even though it's in the way."),
1749                                path);
1750                 else
1751                         remove_file(o, 0, path, 0);
1752                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, new_path1);
1753                 if (!ret)
1754                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1755                                           new_path2);
1756                 /*
1757                  * unpack_trees() actually populates the index for us for
1758                  * "normal" rename/rename(2to1) situtations so that the
1759                  * correct entries are at the higher stages, which would
1760                  * make the call below to update_stages_for_stage_data
1761                  * unnecessary.  However, if either of the renames came
1762                  * from a directory rename, then unpack_trees() will not
1763                  * have gotten the right data loaded into the index, so we
1764                  * need to do so now.  (While it'd be tempting to move this
1765                  * call to update_stages_for_stage_data() to
1766                  * apply_directory_rename_modifications(), that would break
1767                  * our intermediate calls to would_lose_untracked() since
1768                  * those rely on the current in-memory index.  See also the
1769                  * big "NOTE" in update_stages()).
1770                  */
1771                 if (update_stages_for_stage_data(o, path, ci->dst_entry1))
1772                         ret = -1;
1773
1774                 free(new_path2);
1775                 free(new_path1);
1776         }
1777
1778         return ret;
1779 }
1780
1781 /*
1782  * Get the diff_filepairs changed between o_tree and tree.
1783  */
1784 static struct diff_queue_struct *get_diffpairs(struct merge_options *o,
1785                                                struct tree *o_tree,
1786                                                struct tree *tree)
1787 {
1788         struct diff_queue_struct *ret;
1789         struct diff_options opts;
1790
1791         repo_diff_setup(the_repository, &opts);
1792         opts.flags.recursive = 1;
1793         opts.flags.rename_empty = 0;
1794         opts.detect_rename = merge_detect_rename(o);
1795         /*
1796          * We do not have logic to handle the detection of copies.  In
1797          * fact, it may not even make sense to add such logic: would we
1798          * really want a change to a base file to be propagated through
1799          * multiple other files by a merge?
1800          */
1801         if (opts.detect_rename > DIFF_DETECT_RENAME)
1802                 opts.detect_rename = DIFF_DETECT_RENAME;
1803         opts.rename_limit = o->merge_rename_limit >= 0 ? o->merge_rename_limit :
1804                             o->diff_rename_limit >= 0 ? o->diff_rename_limit :
1805                             1000;
1806         opts.rename_score = o->rename_score;
1807         opts.show_rename_progress = o->show_rename_progress;
1808         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1809         diff_setup_done(&opts);
1810         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1811         diffcore_std(&opts);
1812         if (opts.needed_rename_limit > o->needed_rename_limit)
1813                 o->needed_rename_limit = opts.needed_rename_limit;
1814
1815         ret = xmalloc(sizeof(*ret));
1816         *ret = diff_queued_diff;
1817
1818         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1819         diff_queued_diff.nr = 0;
1820         diff_queued_diff.queue = NULL;
1821         diff_flush(&opts);
1822         return ret;
1823 }
1824
1825 static int tree_has_path(struct tree *tree, const char *path)
1826 {
1827         struct object_id hashy;
1828         unsigned int mode_o;
1829
1830         return !get_tree_entry(&tree->object.oid, path,
1831                                &hashy, &mode_o);
1832 }
1833
1834 /*
1835  * Return a new string that replaces the beginning portion (which matches
1836  * entry->dir), with entry->new_dir.  In perl-speak:
1837  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1838  * NOTE:
1839  *   Caller must ensure that old_path starts with entry->dir + '/'.
1840  */
1841 static char *apply_dir_rename(struct dir_rename_entry *entry,
1842                               const char *old_path)
1843 {
1844         struct strbuf new_path = STRBUF_INIT;
1845         int oldlen, newlen;
1846
1847         if (entry->non_unique_new_dir)
1848                 return NULL;
1849
1850         oldlen = strlen(entry->dir);
1851         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1852         strbuf_grow(&new_path, newlen);
1853         strbuf_addbuf(&new_path, &entry->new_dir);
1854         strbuf_addstr(&new_path, &old_path[oldlen]);
1855
1856         return strbuf_detach(&new_path, NULL);
1857 }
1858
1859 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1860                                     char **old_dir, char **new_dir)
1861 {
1862         char *end_of_old, *end_of_new;
1863         int old_len, new_len;
1864
1865         *old_dir = NULL;
1866         *new_dir = NULL;
1867
1868         /*
1869          * For
1870          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1871          * the "e/foo.c" part is the same, we just want to know that
1872          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1873          * so, for this example, this function returns "a/b/c/d" in
1874          * *old_dir and "a/b/some/thing/else" in *new_dir.
1875          *
1876          * Also, if the basename of the file changed, we don't care.  We
1877          * want to know which portion of the directory, if any, changed.
1878          */
1879         end_of_old = strrchr(old_path, '/');
1880         end_of_new = strrchr(new_path, '/');
1881
1882         if (end_of_old == NULL || end_of_new == NULL)
1883                 return;
1884         while (*--end_of_new == *--end_of_old &&
1885                end_of_old != old_path &&
1886                end_of_new != new_path)
1887                 ; /* Do nothing; all in the while loop */
1888         /*
1889          * We've found the first non-matching character in the directory
1890          * paths.  That means the current directory we were comparing
1891          * represents the rename.  Move end_of_old and end_of_new back
1892          * to the full directory name.
1893          */
1894         if (*end_of_old == '/')
1895                 end_of_old++;
1896         if (*end_of_old != '/')
1897                 end_of_new++;
1898         end_of_old = strchr(end_of_old, '/');
1899         end_of_new = strchr(end_of_new, '/');
1900
1901         /*
1902          * It may have been the case that old_path and new_path were the same
1903          * directory all along.  Don't claim a rename if they're the same.
1904          */
1905         old_len = end_of_old - old_path;
1906         new_len = end_of_new - new_path;
1907
1908         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1909                 *old_dir = xstrndup(old_path, old_len);
1910                 *new_dir = xstrndup(new_path, new_len);
1911         }
1912 }
1913
1914 static void remove_hashmap_entries(struct hashmap *dir_renames,
1915                                    struct string_list *items_to_remove)
1916 {
1917         int i;
1918         struct dir_rename_entry *entry;
1919
1920         for (i = 0; i < items_to_remove->nr; i++) {
1921                 entry = items_to_remove->items[i].util;
1922                 hashmap_remove(dir_renames, entry, NULL);
1923         }
1924         string_list_clear(items_to_remove, 0);
1925 }
1926
1927 /*
1928  * See if there is a directory rename for path, and if there are any file
1929  * level conflicts for the renamed location.  If there is a rename and
1930  * there are no conflicts, return the new name.  Otherwise, return NULL.
1931  */
1932 static char *handle_path_level_conflicts(struct merge_options *o,
1933                                          const char *path,
1934                                          struct dir_rename_entry *entry,
1935                                          struct hashmap *collisions,
1936                                          struct tree *tree)
1937 {
1938         char *new_path = NULL;
1939         struct collision_entry *collision_ent;
1940         int clean = 1;
1941         struct strbuf collision_paths = STRBUF_INIT;
1942
1943         /*
1944          * entry has the mapping of old directory name to new directory name
1945          * that we want to apply to path.
1946          */
1947         new_path = apply_dir_rename(entry, path);
1948
1949         if (!new_path) {
1950                 /* This should only happen when entry->non_unique_new_dir set */
1951                 if (!entry->non_unique_new_dir)
1952                         BUG("entry->non_unqiue_dir not set and !new_path");
1953                 output(o, 1, _("CONFLICT (directory rename split): "
1954                                "Unclear where to place %s because directory "
1955                                "%s was renamed to multiple other directories, "
1956                                "with no destination getting a majority of the "
1957                                "files."),
1958                        path, entry->dir);
1959                 clean = 0;
1960                 return NULL;
1961         }
1962
1963         /*
1964          * The caller needs to have ensured that it has pre-populated
1965          * collisions with all paths that map to new_path.  Do a quick check
1966          * to ensure that's the case.
1967          */
1968         collision_ent = collision_find_entry(collisions, new_path);
1969         if (collision_ent == NULL)
1970                 BUG("collision_ent is NULL");
1971
1972         /*
1973          * Check for one-sided add/add/.../add conflicts, i.e.
1974          * where implicit renames from the other side doing
1975          * directory rename(s) can affect this side of history
1976          * to put multiple paths into the same location.  Warn
1977          * and bail on directory renames for such paths.
1978          */
1979         if (collision_ent->reported_already) {
1980                 clean = 0;
1981         } else if (tree_has_path(tree, new_path)) {
1982                 collision_ent->reported_already = 1;
1983                 strbuf_add_separated_string_list(&collision_paths, ", ",
1984                                                  &collision_ent->source_files);
1985                 output(o, 1, _("CONFLICT (implicit dir rename): Existing "
1986                                "file/dir at %s in the way of implicit "
1987                                "directory rename(s) putting the following "
1988                                "path(s) there: %s."),
1989                        new_path, collision_paths.buf);
1990                 clean = 0;
1991         } else if (collision_ent->source_files.nr > 1) {
1992                 collision_ent->reported_already = 1;
1993                 strbuf_add_separated_string_list(&collision_paths, ", ",
1994                                                  &collision_ent->source_files);
1995                 output(o, 1, _("CONFLICT (implicit dir rename): Cannot map "
1996                                "more than one path to %s; implicit directory "
1997                                "renames tried to put these paths there: %s"),
1998                        new_path, collision_paths.buf);
1999                 clean = 0;
2000         }
2001
2002         /* Free memory we no longer need */
2003         strbuf_release(&collision_paths);
2004         if (!clean && new_path) {
2005                 free(new_path);
2006                 return NULL;
2007         }
2008
2009         return new_path;
2010 }
2011
2012 /*
2013  * There are a couple things we want to do at the directory level:
2014  *   1. Check for both sides renaming to the same thing, in order to avoid
2015  *      implicit renaming of files that should be left in place.  (See
2016  *      testcase 6b in t6043 for details.)
2017  *   2. Prune directory renames if there are still files left in the
2018  *      the original directory.  These represent a partial directory rename,
2019  *      i.e. a rename where only some of the files within the directory
2020  *      were renamed elsewhere.  (Technically, this could be done earlier
2021  *      in get_directory_renames(), except that would prevent us from
2022  *      doing the previous check and thus failing testcase 6b.)
2023  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2024  *      In the future, we could potentially record this info as well and
2025  *      omit reporting rename/rename(1to2) conflicts for each path within
2026  *      the affected directories, thus cleaning up the merge output.
2027  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2028  *         directory level, because merging directories is fine.  If it
2029  *         causes conflicts for files within those merged directories, then
2030  *         that should be detected at the individual path level.
2031  */
2032 static void handle_directory_level_conflicts(struct merge_options *o,
2033                                              struct hashmap *dir_re_head,
2034                                              struct tree *head,
2035                                              struct hashmap *dir_re_merge,
2036                                              struct tree *merge)
2037 {
2038         struct hashmap_iter iter;
2039         struct dir_rename_entry *head_ent;
2040         struct dir_rename_entry *merge_ent;
2041
2042         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2043         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2044
2045         hashmap_iter_init(dir_re_head, &iter);
2046         while ((head_ent = hashmap_iter_next(&iter))) {
2047                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2048                 if (merge_ent &&
2049                     !head_ent->non_unique_new_dir &&
2050                     !merge_ent->non_unique_new_dir &&
2051                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2052                         /* 1. Renamed identically; remove it from both sides */
2053                         string_list_append(&remove_from_head,
2054                                            head_ent->dir)->util = head_ent;
2055                         strbuf_release(&head_ent->new_dir);
2056                         string_list_append(&remove_from_merge,
2057                                            merge_ent->dir)->util = merge_ent;
2058                         strbuf_release(&merge_ent->new_dir);
2059                 } else if (tree_has_path(head, head_ent->dir)) {
2060                         /* 2. This wasn't a directory rename after all */
2061                         string_list_append(&remove_from_head,
2062                                            head_ent->dir)->util = head_ent;
2063                         strbuf_release(&head_ent->new_dir);
2064                 }
2065         }
2066
2067         remove_hashmap_entries(dir_re_head, &remove_from_head);
2068         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2069
2070         hashmap_iter_init(dir_re_merge, &iter);
2071         while ((merge_ent = hashmap_iter_next(&iter))) {
2072                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2073                 if (tree_has_path(merge, merge_ent->dir)) {
2074                         /* 2. This wasn't a directory rename after all */
2075                         string_list_append(&remove_from_merge,
2076                                            merge_ent->dir)->util = merge_ent;
2077                 } else if (head_ent &&
2078                            !head_ent->non_unique_new_dir &&
2079                            !merge_ent->non_unique_new_dir) {
2080                         /* 3. rename/rename(1to2) */
2081                         /*
2082                          * We can assume it's not rename/rename(1to1) because
2083                          * that was case (1), already checked above.  So we
2084                          * know that head_ent->new_dir and merge_ent->new_dir
2085                          * are different strings.
2086                          */
2087                         output(o, 1, _("CONFLICT (rename/rename): "
2088                                        "Rename directory %s->%s in %s. "
2089                                        "Rename directory %s->%s in %s"),
2090                                head_ent->dir, head_ent->new_dir.buf, o->branch1,
2091                                head_ent->dir, merge_ent->new_dir.buf, o->branch2);
2092                         string_list_append(&remove_from_head,
2093                                            head_ent->dir)->util = head_ent;
2094                         strbuf_release(&head_ent->new_dir);
2095                         string_list_append(&remove_from_merge,
2096                                            merge_ent->dir)->util = merge_ent;
2097                         strbuf_release(&merge_ent->new_dir);
2098                 }
2099         }
2100
2101         remove_hashmap_entries(dir_re_head, &remove_from_head);
2102         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2103 }
2104
2105 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs,
2106                                              struct tree *tree)
2107 {
2108         struct hashmap *dir_renames;
2109         struct hashmap_iter iter;
2110         struct dir_rename_entry *entry;
2111         int i;
2112
2113         /*
2114          * Typically, we think of a directory rename as all files from a
2115          * certain directory being moved to a target directory.  However,
2116          * what if someone first moved two files from the original
2117          * directory in one commit, and then renamed the directory
2118          * somewhere else in a later commit?  At merge time, we just know
2119          * that files from the original directory went to two different
2120          * places, and that the bulk of them ended up in the same place.
2121          * We want each directory rename to represent where the bulk of the
2122          * files from that directory end up; this function exists to find
2123          * where the bulk of the files went.
2124          *
2125          * The first loop below simply iterates through the list of file
2126          * renames, finding out how often each directory rename pair
2127          * possibility occurs.
2128          */
2129         dir_renames = xmalloc(sizeof(*dir_renames));
2130         dir_rename_init(dir_renames);
2131         for (i = 0; i < pairs->nr; ++i) {
2132                 struct string_list_item *item;
2133                 int *count;
2134                 struct diff_filepair *pair = pairs->queue[i];
2135                 char *old_dir, *new_dir;
2136
2137                 /* File not part of directory rename if it wasn't renamed */
2138                 if (pair->status != 'R')
2139                         continue;
2140
2141                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2142                                         &old_dir,        &new_dir);
2143                 if (!old_dir)
2144                         /* Directory didn't change at all; ignore this one. */
2145                         continue;
2146
2147                 entry = dir_rename_find_entry(dir_renames, old_dir);
2148                 if (!entry) {
2149                         entry = xmalloc(sizeof(*entry));
2150                         dir_rename_entry_init(entry, old_dir);
2151                         hashmap_put(dir_renames, entry);
2152                 } else {
2153                         free(old_dir);
2154                 }
2155                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2156                 if (!item) {
2157                         item = string_list_insert(&entry->possible_new_dirs,
2158                                                   new_dir);
2159                         item->util = xcalloc(1, sizeof(int));
2160                 } else {
2161                         free(new_dir);
2162                 }
2163                 count = item->util;
2164                 *count += 1;
2165         }
2166
2167         /*
2168          * For each directory with files moved out of it, we find out which
2169          * target directory received the most files so we can declare it to
2170          * be the "winning" target location for the directory rename.  This
2171          * winner gets recorded in new_dir.  If there is no winner
2172          * (multiple target directories received the same number of files),
2173          * we set non_unique_new_dir.  Once we've determined the winner (or
2174          * that there is no winner), we no longer need possible_new_dirs.
2175          */
2176         hashmap_iter_init(dir_renames, &iter);
2177         while ((entry = hashmap_iter_next(&iter))) {
2178                 int max = 0;
2179                 int bad_max = 0;
2180                 char *best = NULL;
2181
2182                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2183                         int *count = entry->possible_new_dirs.items[i].util;
2184
2185                         if (*count == max)
2186                                 bad_max = max;
2187                         else if (*count > max) {
2188                                 max = *count;
2189                                 best = entry->possible_new_dirs.items[i].string;
2190                         }
2191                 }
2192                 if (bad_max == max)
2193                         entry->non_unique_new_dir = 1;
2194                 else {
2195                         assert(entry->new_dir.len == 0);
2196                         strbuf_addstr(&entry->new_dir, best);
2197                 }
2198                 /*
2199                  * The relevant directory sub-portion of the original full
2200                  * filepaths were xstrndup'ed before inserting into
2201                  * possible_new_dirs, and instead of manually iterating the
2202                  * list and free'ing each, just lie and tell
2203                  * possible_new_dirs that it did the strdup'ing so that it
2204                  * will free them for us.
2205                  */
2206                 entry->possible_new_dirs.strdup_strings = 1;
2207                 string_list_clear(&entry->possible_new_dirs, 1);
2208         }
2209
2210         return dir_renames;
2211 }
2212
2213 static struct dir_rename_entry *check_dir_renamed(const char *path,
2214                                                   struct hashmap *dir_renames)
2215 {
2216         char *temp = xstrdup(path);
2217         char *end;
2218         struct dir_rename_entry *entry = NULL;
2219
2220         while ((end = strrchr(temp, '/'))) {
2221                 *end = '\0';
2222                 entry = dir_rename_find_entry(dir_renames, temp);
2223                 if (entry)
2224                         break;
2225         }
2226         free(temp);
2227         return entry;
2228 }
2229
2230 static void compute_collisions(struct hashmap *collisions,
2231                                struct hashmap *dir_renames,
2232                                struct diff_queue_struct *pairs)
2233 {
2234         int i;
2235
2236         /*
2237          * Multiple files can be mapped to the same path due to directory
2238          * renames done by the other side of history.  Since that other
2239          * side of history could have merged multiple directories into one,
2240          * if our side of history added the same file basename to each of
2241          * those directories, then all N of them would get implicitly
2242          * renamed by the directory rename detection into the same path,
2243          * and we'd get an add/add/.../add conflict, and all those adds
2244          * from *this* side of history.  This is not representable in the
2245          * index, and users aren't going to easily be able to make sense of
2246          * it.  So we need to provide a good warning about what's
2247          * happening, and fall back to no-directory-rename detection
2248          * behavior for those paths.
2249          *
2250          * See testcases 9e and all of section 5 from t6043 for examples.
2251          */
2252         collision_init(collisions);
2253
2254         for (i = 0; i < pairs->nr; ++i) {
2255                 struct dir_rename_entry *dir_rename_ent;
2256                 struct collision_entry *collision_ent;
2257                 char *new_path;
2258                 struct diff_filepair *pair = pairs->queue[i];
2259
2260                 if (pair->status != 'A' && pair->status != 'R')
2261                         continue;
2262                 dir_rename_ent = check_dir_renamed(pair->two->path,
2263                                                    dir_renames);
2264                 if (!dir_rename_ent)
2265                         continue;
2266
2267                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2268                 if (!new_path)
2269                         /*
2270                          * dir_rename_ent->non_unique_new_path is true, which
2271                          * means there is no directory rename for us to use,
2272                          * which means it won't cause us any additional
2273                          * collisions.
2274                          */
2275                         continue;
2276                 collision_ent = collision_find_entry(collisions, new_path);
2277                 if (!collision_ent) {
2278                         collision_ent = xcalloc(1,
2279                                                 sizeof(struct collision_entry));
2280                         hashmap_entry_init(collision_ent, strhash(new_path));
2281                         hashmap_put(collisions, collision_ent);
2282                         collision_ent->target_file = new_path;
2283                 } else {
2284                         free(new_path);
2285                 }
2286                 string_list_insert(&collision_ent->source_files,
2287                                    pair->two->path);
2288         }
2289 }
2290
2291 static char *check_for_directory_rename(struct merge_options *o,
2292                                         const char *path,
2293                                         struct tree *tree,
2294                                         struct hashmap *dir_renames,
2295                                         struct hashmap *dir_rename_exclusions,
2296                                         struct hashmap *collisions,
2297                                         int *clean_merge)
2298 {
2299         char *new_path = NULL;
2300         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2301         struct dir_rename_entry *oentry = NULL;
2302
2303         if (!entry)
2304                 return new_path;
2305
2306         /*
2307          * This next part is a little weird.  We do not want to do an
2308          * implicit rename into a directory we renamed on our side, because
2309          * that will result in a spurious rename/rename(1to2) conflict.  An
2310          * example:
2311          *   Base commit: dumbdir/afile, otherdir/bfile
2312          *   Side 1:      smrtdir/afile, otherdir/bfile
2313          *   Side 2:      dumbdir/afile, dumbdir/bfile
2314          * Here, while working on Side 1, we could notice that otherdir was
2315          * renamed/merged to dumbdir, and change the diff_filepair for
2316          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2317          * 2 will notice the rename from dumbdir to smrtdir, and do the
2318          * transitive rename to move it from dumbdir/bfile to
2319          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2320          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2321          * the file to end up in smrtdir.  And the way to achieve that is
2322          * to not let Side1 do the rename to dumbdir, since we know that is
2323          * the source of one of our directory renames.
2324          *
2325          * That's why oentry and dir_rename_exclusions is here.
2326          *
2327          * As it turns out, this also prevents N-way transient rename
2328          * confusion; See testcases 9c and 9d of t6043.
2329          */
2330         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2331         if (oentry) {
2332                 output(o, 1, _("WARNING: Avoiding applying %s -> %s rename "
2333                                "to %s, because %s itself was renamed."),
2334                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2335         } else {
2336                 new_path = handle_path_level_conflicts(o, path, entry,
2337                                                        collisions, tree);
2338                 *clean_merge &= (new_path != NULL);
2339         }
2340
2341         return new_path;
2342 }
2343
2344 static void apply_directory_rename_modifications(struct merge_options *o,
2345                                                  struct diff_filepair *pair,
2346                                                  char *new_path,
2347                                                  struct rename *re,
2348                                                  struct tree *tree,
2349                                                  struct tree *o_tree,
2350                                                  struct tree *a_tree,
2351                                                  struct tree *b_tree,
2352                                                  struct string_list *entries,
2353                                                  int *clean)
2354 {
2355         struct string_list_item *item;
2356         int stage = (tree == a_tree ? 2 : 3);
2357         int update_wd;
2358
2359         /*
2360          * In all cases where we can do directory rename detection,
2361          * unpack_trees() will have read pair->two->path into the
2362          * index and the working copy.  We need to remove it so that
2363          * we can instead place it at new_path.  It is guaranteed to
2364          * not be untracked (unpack_trees() would have errored out
2365          * saying the file would have been overwritten), but it might
2366          * be dirty, though.
2367          */
2368         update_wd = !was_dirty(o, pair->two->path);
2369         if (!update_wd)
2370                 output(o, 1, _("Refusing to lose dirty file at %s"),
2371                        pair->two->path);
2372         remove_file(o, 1, pair->two->path, !update_wd);
2373
2374         /* Find or create a new re->dst_entry */
2375         item = string_list_lookup(entries, new_path);
2376         if (item) {
2377                 /*
2378                  * Since we're renaming on this side of history, and it's
2379                  * due to a directory rename on the other side of history
2380                  * (which we only allow when the directory in question no
2381                  * longer exists on the other side of history), the
2382                  * original entry for re->dst_entry is no longer
2383                  * necessary...
2384                  */
2385                 re->dst_entry->processed = 1;
2386
2387                 /*
2388                  * ...because we'll be using this new one.
2389                  */
2390                 re->dst_entry = item->util;
2391         } else {
2392                 /*
2393                  * re->dst_entry is for the before-dir-rename path, and we
2394                  * need it to hold information for the after-dir-rename
2395                  * path.  Before creating a new entry, we need to mark the
2396                  * old one as unnecessary (...unless it is shared by
2397                  * src_entry, i.e. this didn't use to be a rename, in which
2398                  * case we can just allow the normal processing to happen
2399                  * for it).
2400                  */
2401                 if (pair->status == 'R')
2402                         re->dst_entry->processed = 1;
2403
2404                 re->dst_entry = insert_stage_data(new_path,
2405                                                   o_tree, a_tree, b_tree,
2406                                                   entries);
2407                 item = string_list_insert(entries, new_path);
2408                 item->util = re->dst_entry;
2409         }
2410
2411         /*
2412          * Update the stage_data with the information about the path we are
2413          * moving into place.  That slot will be empty and available for us
2414          * to write to because of the collision checks in
2415          * handle_path_level_conflicts().  In other words,
2416          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2417          * open for us to write to.
2418          *
2419          * It may be tempting to actually update the index at this point as
2420          * well, using update_stages_for_stage_data(), but as per the big
2421          * "NOTE" in update_stages(), doing so will modify the current
2422          * in-memory index which will break calls to would_lose_untracked()
2423          * that we need to make.  Instead, we need to just make sure that
2424          * the various handle_rename_*() functions update the index
2425          * explicitly rather than relying on unpack_trees() to have done it.
2426          */
2427         get_tree_entry(&tree->object.oid,
2428                        pair->two->path,
2429                        &re->dst_entry->stages[stage].oid,
2430                        &re->dst_entry->stages[stage].mode);
2431
2432         /* Update pair status */
2433         if (pair->status == 'A') {
2434                 /*
2435                  * Recording rename information for this add makes it look
2436                  * like a rename/delete conflict.  Make sure we can
2437                  * correctly handle this as an add that was moved to a new
2438                  * directory instead of reporting a rename/delete conflict.
2439                  */
2440                 re->add_turned_into_rename = 1;
2441         }
2442         /*
2443          * We don't actually look at pair->status again, but it seems
2444          * pedagogically correct to adjust it.
2445          */
2446         pair->status = 'R';
2447
2448         /*
2449          * Finally, record the new location.
2450          */
2451         pair->two->path = new_path;
2452 }
2453
2454 /*
2455  * Get information of all renames which occurred in 'pairs', making use of
2456  * any implicit directory renames inferred from the other side of history.
2457  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2458  * to be able to associate the correct cache entries with the rename
2459  * information; tree is always equal to either a_tree or b_tree.
2460  */
2461 static struct string_list *get_renames(struct merge_options *o,
2462                                        struct diff_queue_struct *pairs,
2463                                        struct hashmap *dir_renames,
2464                                        struct hashmap *dir_rename_exclusions,
2465                                        struct tree *tree,
2466                                        struct tree *o_tree,
2467                                        struct tree *a_tree,
2468                                        struct tree *b_tree,
2469                                        struct string_list *entries,
2470                                        int *clean_merge)
2471 {
2472         int i;
2473         struct hashmap collisions;
2474         struct hashmap_iter iter;
2475         struct collision_entry *e;
2476         struct string_list *renames;
2477
2478         compute_collisions(&collisions, dir_renames, pairs);
2479         renames = xcalloc(1, sizeof(struct string_list));
2480
2481         for (i = 0; i < pairs->nr; ++i) {
2482                 struct string_list_item *item;
2483                 struct rename *re;
2484                 struct diff_filepair *pair = pairs->queue[i];
2485                 char *new_path; /* non-NULL only with directory renames */
2486
2487                 if (pair->status != 'A' && pair->status != 'R') {
2488                         diff_free_filepair(pair);
2489                         continue;
2490                 }
2491                 new_path = check_for_directory_rename(o, pair->two->path, tree,
2492                                                       dir_renames,
2493                                                       dir_rename_exclusions,
2494                                                       &collisions,
2495                                                       clean_merge);
2496                 if (pair->status != 'R' && !new_path) {
2497                         diff_free_filepair(pair);
2498                         continue;
2499                 }
2500
2501                 re = xmalloc(sizeof(*re));
2502                 re->processed = 0;
2503                 re->add_turned_into_rename = 0;
2504                 re->pair = pair;
2505                 item = string_list_lookup(entries, re->pair->one->path);
2506                 if (!item)
2507                         re->src_entry = insert_stage_data(re->pair->one->path,
2508                                         o_tree, a_tree, b_tree, entries);
2509                 else
2510                         re->src_entry = item->util;
2511
2512                 item = string_list_lookup(entries, re->pair->two->path);
2513                 if (!item)
2514                         re->dst_entry = insert_stage_data(re->pair->two->path,
2515                                         o_tree, a_tree, b_tree, entries);
2516                 else
2517                         re->dst_entry = item->util;
2518                 item = string_list_insert(renames, pair->one->path);
2519                 item->util = re;
2520                 if (new_path)
2521                         apply_directory_rename_modifications(o, pair, new_path,
2522                                                              re, tree, o_tree,
2523                                                              a_tree, b_tree,
2524                                                              entries,
2525                                                              clean_merge);
2526         }
2527
2528         hashmap_iter_init(&collisions, &iter);
2529         while ((e = hashmap_iter_next(&iter))) {
2530                 free(e->target_file);
2531                 string_list_clear(&e->source_files, 0);
2532         }
2533         hashmap_free(&collisions, 1);
2534         return renames;
2535 }
2536
2537 static int process_renames(struct merge_options *o,
2538                            struct string_list *a_renames,
2539                            struct string_list *b_renames)
2540 {
2541         int clean_merge = 1, i, j;
2542         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2543         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2544         const struct rename *sre;
2545
2546         for (i = 0; i < a_renames->nr; i++) {
2547                 sre = a_renames->items[i].util;
2548                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2549                         = (void *)sre;
2550         }
2551         for (i = 0; i < b_renames->nr; i++) {
2552                 sre = b_renames->items[i].util;
2553                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2554                         = (void *)sre;
2555         }
2556
2557         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2558                 struct string_list *renames1, *renames2Dst;
2559                 struct rename *ren1 = NULL, *ren2 = NULL;
2560                 const char *branch1, *branch2;
2561                 const char *ren1_src, *ren1_dst;
2562                 struct string_list_item *lookup;
2563
2564                 if (i >= a_renames->nr) {
2565                         ren2 = b_renames->items[j++].util;
2566                 } else if (j >= b_renames->nr) {
2567                         ren1 = a_renames->items[i++].util;
2568                 } else {
2569                         int compare = strcmp(a_renames->items[i].string,
2570                                              b_renames->items[j].string);
2571                         if (compare <= 0)
2572                                 ren1 = a_renames->items[i++].util;
2573                         if (compare >= 0)
2574                                 ren2 = b_renames->items[j++].util;
2575                 }
2576
2577                 /* TODO: refactor, so that 1/2 are not needed */
2578                 if (ren1) {
2579                         renames1 = a_renames;
2580                         renames2Dst = &b_by_dst;
2581                         branch1 = o->branch1;
2582                         branch2 = o->branch2;
2583                 } else {
2584                         renames1 = b_renames;
2585                         renames2Dst = &a_by_dst;
2586                         branch1 = o->branch2;
2587                         branch2 = o->branch1;
2588                         SWAP(ren2, ren1);
2589                 }
2590
2591                 if (ren1->processed)
2592                         continue;
2593                 ren1->processed = 1;
2594                 ren1->dst_entry->processed = 1;
2595                 /* BUG: We should only mark src_entry as processed if we
2596                  * are not dealing with a rename + add-source case.
2597                  */
2598                 ren1->src_entry->processed = 1;
2599
2600                 ren1_src = ren1->pair->one->path;
2601                 ren1_dst = ren1->pair->two->path;
2602
2603                 if (ren2) {
2604                         /* One file renamed on both sides */
2605                         const char *ren2_src = ren2->pair->one->path;
2606                         const char *ren2_dst = ren2->pair->two->path;
2607                         enum rename_type rename_type;
2608                         if (strcmp(ren1_src, ren2_src) != 0)
2609                                 BUG("ren1_src != ren2_src");
2610                         ren2->dst_entry->processed = 1;
2611                         ren2->processed = 1;
2612                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2613                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2614                                 clean_merge = 0;
2615                         } else {
2616                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2617                                 /* BUG: We should only remove ren1_src in
2618                                  * the base stage (think of rename +
2619                                  * add-source cases).
2620                                  */
2621                                 remove_file(o, 1, ren1_src, 1);
2622                                 update_entry(ren1->dst_entry,
2623                                              ren1->pair->one,
2624                                              ren1->pair->two,
2625                                              ren2->pair->two);
2626                         }
2627                         setup_rename_conflict_info(rename_type,
2628                                                    ren1->pair,
2629                                                    ren2->pair,
2630                                                    branch1,
2631                                                    branch2,
2632                                                    ren1->dst_entry,
2633                                                    ren2->dst_entry,
2634                                                    o,
2635                                                    NULL,
2636                                                    NULL);
2637                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2638                         /* Two different files renamed to the same thing */
2639                         char *ren2_dst;
2640                         ren2 = lookup->util;
2641                         ren2_dst = ren2->pair->two->path;
2642                         if (strcmp(ren1_dst, ren2_dst) != 0)
2643                                 BUG("ren1_dst != ren2_dst");
2644
2645                         clean_merge = 0;
2646                         ren2->processed = 1;
2647                         /*
2648                          * BUG: We should only mark src_entry as processed
2649                          * if we are not dealing with a rename + add-source
2650                          * case.
2651                          */
2652                         ren2->src_entry->processed = 1;
2653
2654                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2655                                                    ren1->pair,
2656                                                    ren2->pair,
2657                                                    branch1,
2658                                                    branch2,
2659                                                    ren1->dst_entry,
2660                                                    ren2->dst_entry,
2661                                                    o,
2662                                                    ren1->src_entry,
2663                                                    ren2->src_entry);
2664
2665                 } else {
2666                         /* Renamed in 1, maybe changed in 2 */
2667                         /* we only use sha1 and mode of these */
2668                         struct diff_filespec src_other, dst_other;
2669                         int try_merge;
2670
2671                         /*
2672                          * unpack_trees loads entries from common-commit
2673                          * into stage 1, from head-commit into stage 2, and
2674                          * from merge-commit into stage 3.  We keep track
2675                          * of which side corresponds to the rename.
2676                          */
2677                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2678                         int other_stage =   a_renames == renames1 ? 3 : 2;
2679
2680                         /* BUG: We should only remove ren1_src in the base
2681                          * stage and in other_stage (think of rename +
2682                          * add-source case).
2683                          */
2684                         remove_file(o, 1, ren1_src,
2685                                     renamed_stage == 2 || !was_tracked(o, ren1_src));
2686
2687                         oidcpy(&src_other.oid,
2688                                &ren1->src_entry->stages[other_stage].oid);
2689                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2690                         oidcpy(&dst_other.oid,
2691                                &ren1->dst_entry->stages[other_stage].oid);
2692                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2693                         try_merge = 0;
2694
2695                         if (oid_eq(&src_other.oid, &null_oid) &&
2696                             ren1->add_turned_into_rename) {
2697                                 setup_rename_conflict_info(RENAME_VIA_DIR,
2698                                                            ren1->pair,
2699                                                            NULL,
2700                                                            branch1,
2701                                                            branch2,
2702                                                            ren1->dst_entry,
2703                                                            NULL,
2704                                                            o,
2705                                                            NULL,
2706                                                            NULL);
2707                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2708                                 setup_rename_conflict_info(RENAME_DELETE,
2709                                                            ren1->pair,
2710                                                            NULL,
2711                                                            branch1,
2712                                                            branch2,
2713                                                            ren1->dst_entry,
2714                                                            NULL,
2715                                                            o,
2716                                                            NULL,
2717                                                            NULL);
2718                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2719                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2720                                 /*
2721                                  * Added file on the other side identical to
2722                                  * the file being renamed: clean merge.
2723                                  * Also, there is no need to overwrite the
2724                                  * file already in the working copy, so call
2725                                  * update_file_flags() instead of
2726                                  * update_file().
2727                                  */
2728                                 if (update_file_flags(o,
2729                                                       &ren1->pair->two->oid,
2730                                                       ren1->pair->two->mode,
2731                                                       ren1_dst,
2732                                                       1, /* update_cache */
2733                                                       0  /* update_wd    */))
2734                                         clean_merge = -1;
2735                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2736                                 clean_merge = 0;
2737                                 try_merge = 1;
2738                                 output(o, 1, _("CONFLICT (rename/add): Rename %s->%s in %s. "
2739                                        "%s added in %s"),
2740                                        ren1_src, ren1_dst, branch1,
2741                                        ren1_dst, branch2);
2742                                 if (o->call_depth) {
2743                                         struct merge_file_info mfi;
2744                                         struct diff_filespec one, a, b;
2745
2746                                         oidcpy(&one.oid, &null_oid);
2747                                         one.mode = 0;
2748                                         one.path = ren1->pair->two->path;
2749
2750                                         oidcpy(&a.oid, &ren1->pair->two->oid);
2751                                         a.mode = ren1->pair->two->mode;
2752                                         a.path = one.path;
2753
2754                                         oidcpy(&b.oid, &dst_other.oid);
2755                                         b.mode = dst_other.mode;
2756                                         b.path = one.path;
2757
2758                                         if (merge_mode_and_contents(o, &one, &a, &b, ren1_dst,
2759                                                                     branch1, branch2,
2760                                                                     &mfi)) {
2761                                                 clean_merge = -1;
2762                                                 goto cleanup_and_return;
2763                                         }
2764                                         output(o, 1, _("Adding merged %s"), ren1_dst);
2765                                         if (update_file(o, 0, &mfi.oid,
2766                                                         mfi.mode, ren1_dst))
2767                                                 clean_merge = -1;
2768                                         try_merge = 0;
2769                                 } else {
2770                                         char *new_path = unique_path(o, ren1_dst, branch2);
2771                                         output(o, 1, _("Adding as %s instead"), new_path);
2772                                         if (update_file(o, 0, &dst_other.oid,
2773                                                         dst_other.mode, new_path))
2774                                                 clean_merge = -1;
2775                                         free(new_path);
2776                                 }
2777                         } else
2778                                 try_merge = 1;
2779
2780                         if (clean_merge < 0)
2781                                 goto cleanup_and_return;
2782                         if (try_merge) {
2783                                 struct diff_filespec *one, *a, *b;
2784                                 src_other.path = (char *)ren1_src;
2785
2786                                 one = ren1->pair->one;
2787                                 if (a_renames == renames1) {
2788                                         a = ren1->pair->two;
2789                                         b = &src_other;
2790                                 } else {
2791                                         b = ren1->pair->two;
2792                                         a = &src_other;
2793                                 }
2794                                 update_entry(ren1->dst_entry, one, a, b);
2795                                 setup_rename_conflict_info(RENAME_NORMAL,
2796                                                            ren1->pair,
2797                                                            NULL,
2798                                                            branch1,
2799                                                            NULL,
2800                                                            ren1->dst_entry,
2801                                                            NULL,
2802                                                            o,
2803                                                            NULL,
2804                                                            NULL);
2805                         }
2806                 }
2807         }
2808 cleanup_and_return:
2809         string_list_clear(&a_by_dst, 0);
2810         string_list_clear(&b_by_dst, 0);
2811
2812         return clean_merge;
2813 }
2814
2815 struct rename_info {
2816         struct string_list *head_renames;
2817         struct string_list *merge_renames;
2818 };
2819
2820 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2821                                    struct hashmap *dir_renames)
2822 {
2823         struct hashmap_iter iter;
2824         struct dir_rename_entry *e;
2825
2826         hashmap_iter_init(dir_renames, &iter);
2827         while ((e = hashmap_iter_next(&iter))) {
2828                 free(e->dir);
2829                 strbuf_release(&e->new_dir);
2830                 /* possible_new_dirs already cleared in get_directory_renames */
2831         }
2832         hashmap_free(dir_renames, 1);
2833         free(dir_renames);
2834
2835         free(pairs->queue);
2836         free(pairs);
2837 }
2838
2839 static int detect_and_process_renames(struct merge_options *o,
2840                                       struct tree *common,
2841                                       struct tree *head,
2842                                       struct tree *merge,
2843                                       struct string_list *entries,
2844                                       struct rename_info *ri)
2845 {
2846         struct diff_queue_struct *head_pairs, *merge_pairs;
2847         struct hashmap *dir_re_head, *dir_re_merge;
2848         int clean = 1;
2849
2850         ri->head_renames = NULL;
2851         ri->merge_renames = NULL;
2852
2853         if (!merge_detect_rename(o))
2854                 return 1;
2855
2856         head_pairs = get_diffpairs(o, common, head);
2857         merge_pairs = get_diffpairs(o, common, merge);
2858
2859         if (o->detect_directory_renames) {
2860                 dir_re_head = get_directory_renames(head_pairs, head);
2861                 dir_re_merge = get_directory_renames(merge_pairs, merge);
2862
2863                 handle_directory_level_conflicts(o,
2864                                                  dir_re_head, head,
2865                                                  dir_re_merge, merge);
2866         } else {
2867                 dir_re_head  = xmalloc(sizeof(*dir_re_head));
2868                 dir_re_merge = xmalloc(sizeof(*dir_re_merge));
2869                 dir_rename_init(dir_re_head);
2870                 dir_rename_init(dir_re_merge);
2871         }
2872
2873         ri->head_renames  = get_renames(o, head_pairs,
2874                                         dir_re_merge, dir_re_head, head,
2875                                         common, head, merge, entries,
2876                                         &clean);
2877         if (clean < 0)
2878                 goto cleanup;
2879         ri->merge_renames = get_renames(o, merge_pairs,
2880                                         dir_re_head, dir_re_merge, merge,
2881                                         common, head, merge, entries,
2882                                         &clean);
2883         if (clean < 0)
2884                 goto cleanup;
2885         clean &= process_renames(o, ri->head_renames, ri->merge_renames);
2886
2887 cleanup:
2888         /*
2889          * Some cleanup is deferred until cleanup_renames() because the
2890          * data structures are still needed and referenced in
2891          * process_entry().  But there are a few things we can free now.
2892          */
2893         initial_cleanup_rename(head_pairs, dir_re_head);
2894         initial_cleanup_rename(merge_pairs, dir_re_merge);
2895
2896         return clean;
2897 }
2898
2899 static void final_cleanup_rename(struct string_list *rename)
2900 {
2901         const struct rename *re;
2902         int i;
2903
2904         if (rename == NULL)
2905                 return;
2906
2907         for (i = 0; i < rename->nr; i++) {
2908                 re = rename->items[i].util;
2909                 diff_free_filepair(re->pair);
2910         }
2911         string_list_clear(rename, 1);
2912         free(rename);
2913 }
2914
2915 static void final_cleanup_renames(struct rename_info *re_info)
2916 {
2917         final_cleanup_rename(re_info->head_renames);
2918         final_cleanup_rename(re_info->merge_renames);
2919 }
2920
2921 static struct object_id *stage_oid(const struct object_id *oid, unsigned mode)
2922 {
2923         return (is_null_oid(oid) || mode == 0) ? NULL: (struct object_id *)oid;
2924 }
2925
2926 static int read_oid_strbuf(struct merge_options *o,
2927                            const struct object_id *oid,
2928                            struct strbuf *dst)
2929 {
2930         void *buf;
2931         enum object_type type;
2932         unsigned long size;
2933         buf = read_object_file(oid, &type, &size);
2934         if (!buf)
2935                 return err(o, _("cannot read object %s"), oid_to_hex(oid));
2936         if (type != OBJ_BLOB) {
2937                 free(buf);
2938                 return err(o, _("object %s is not a blob"), oid_to_hex(oid));
2939         }
2940         strbuf_attach(dst, buf, size, size + 1);
2941         return 0;
2942 }
2943
2944 static int blob_unchanged(struct merge_options *opt,
2945                           const struct object_id *o_oid,
2946                           unsigned o_mode,
2947                           const struct object_id *a_oid,
2948                           unsigned a_mode,
2949                           int renormalize, const char *path)
2950 {
2951         struct strbuf o = STRBUF_INIT;
2952         struct strbuf a = STRBUF_INIT;
2953         int ret = 0; /* assume changed for safety */
2954
2955         if (a_mode != o_mode)
2956                 return 0;
2957         if (oid_eq(o_oid, a_oid))
2958                 return 1;
2959         if (!renormalize)
2960                 return 0;
2961
2962         assert(o_oid && a_oid);
2963         if (read_oid_strbuf(opt, o_oid, &o) || read_oid_strbuf(opt, a_oid, &a))
2964                 goto error_return;
2965         /*
2966          * Note: binary | is used so that both renormalizations are
2967          * performed.  Comparison can be skipped if both files are
2968          * unchanged since their sha1s have already been compared.
2969          */
2970         if (renormalize_buffer(&the_index, path, o.buf, o.len, &o) |
2971             renormalize_buffer(&the_index, path, a.buf, a.len, &a))
2972                 ret = (o.len == a.len && !memcmp(o.buf, a.buf, o.len));
2973
2974 error_return:
2975         strbuf_release(&o);
2976         strbuf_release(&a);
2977         return ret;
2978 }
2979
2980 static int handle_modify_delete(struct merge_options *o,
2981                                 const char *path,
2982                                 struct object_id *o_oid, int o_mode,
2983                                 struct object_id *a_oid, int a_mode,
2984                                 struct object_id *b_oid, int b_mode)
2985 {
2986         const char *modify_branch, *delete_branch;
2987         struct object_id *changed_oid;
2988         int changed_mode;
2989
2990         if (a_oid) {
2991                 modify_branch = o->branch1;
2992                 delete_branch = o->branch2;
2993                 changed_oid = a_oid;
2994                 changed_mode = a_mode;
2995         } else {
2996                 modify_branch = o->branch2;
2997                 delete_branch = o->branch1;
2998                 changed_oid = b_oid;
2999                 changed_mode = b_mode;
3000         }
3001
3002         return handle_change_delete(o,
3003                                     path, NULL,
3004                                     o_oid, o_mode,
3005                                     changed_oid, changed_mode,
3006                                     modify_branch, delete_branch,
3007                                     _("modify"), _("modified"));
3008 }
3009
3010 static int handle_content_merge(struct merge_options *o,
3011                                 const char *path,
3012                                 int is_dirty,
3013                                 struct object_id *o_oid, int o_mode,
3014                                 struct object_id *a_oid, int a_mode,
3015                                 struct object_id *b_oid, int b_mode,
3016                                 struct rename_conflict_info *rename_conflict_info)
3017 {
3018         const char *reason = _("content");
3019         const char *path1 = NULL, *path2 = NULL;
3020         struct merge_file_info mfi;
3021         struct diff_filespec one, a, b;
3022         unsigned df_conflict_remains = 0;
3023
3024         if (!o_oid) {
3025                 reason = _("add/add");
3026                 o_oid = (struct object_id *)&null_oid;
3027         }
3028         one.path = a.path = b.path = (char *)path;
3029         oidcpy(&one.oid, o_oid);
3030         one.mode = o_mode;
3031         oidcpy(&a.oid, a_oid);
3032         a.mode = a_mode;
3033         oidcpy(&b.oid, b_oid);
3034         b.mode = b_mode;
3035
3036         if (rename_conflict_info) {
3037                 struct diff_filepair *pair1 = rename_conflict_info->pair1;
3038
3039                 path1 = (o->branch1 == rename_conflict_info->branch1) ?
3040                         pair1->two->path : pair1->one->path;
3041                 /* If rename_conflict_info->pair2 != NULL, we are in
3042                  * RENAME_ONE_FILE_TO_ONE case.  Otherwise, we have a
3043                  * normal rename.
3044                  */
3045                 path2 = (rename_conflict_info->pair2 ||
3046                          o->branch2 == rename_conflict_info->branch1) ?
3047                         pair1->two->path : pair1->one->path;
3048                 one.path = pair1->one->path;
3049                 a.path = (char *)path1;
3050                 b.path = (char *)path2;
3051
3052                 if (dir_in_way(path, !o->call_depth,
3053                                S_ISGITLINK(pair1->two->mode)))
3054                         df_conflict_remains = 1;
3055         }
3056         if (merge_mode_and_contents(o, &one, &a, &b, path,
3057                                     o->branch1, o->branch2, &mfi))
3058                 return -1;
3059
3060         /*
3061          * We can skip updating the working tree file iff:
3062          *   a) The merge is clean
3063          *   b) The merge matches what was in HEAD (content, mode, pathname)
3064          *   c) The target path is usable (i.e. not involved in D/F conflict)
3065          */
3066         if (mfi.clean &&
3067             was_tracked_and_matches(o, path, &mfi.oid, mfi.mode) &&
3068             !df_conflict_remains) {
3069                 int pos;
3070                 struct cache_entry *ce;
3071
3072                 output(o, 3, _("Skipped %s (merged same as existing)"), path);
3073                 if (add_cacheinfo(o, mfi.mode, &mfi.oid, path,
3074                                   0, (!o->call_depth && !is_dirty), 0))
3075                         return -1;
3076                 /*
3077                  * However, add_cacheinfo() will delete the old cache entry
3078                  * and add a new one.  We need to copy over any skip_worktree
3079                  * flag to avoid making the file appear as if it were
3080                  * deleted by the user.
3081                  */
3082                 pos = index_name_pos(&o->orig_index, path, strlen(path));
3083                 ce = o->orig_index.cache[pos];
3084                 if (ce_skip_worktree(ce)) {
3085                         pos = index_name_pos(&the_index, path, strlen(path));
3086                         ce = the_index.cache[pos];
3087                         ce->ce_flags |= CE_SKIP_WORKTREE;
3088                 }
3089                 return mfi.clean;
3090         }
3091
3092         if (!mfi.clean) {
3093                 if (S_ISGITLINK(mfi.mode))
3094                         reason = _("submodule");
3095                 output(o, 1, _("CONFLICT (%s): Merge conflict in %s"),
3096                                 reason, path);
3097                 if (rename_conflict_info && !df_conflict_remains)
3098                         if (update_stages(o, path, &one, &a, &b))
3099                                 return -1;
3100         }
3101
3102         if (df_conflict_remains || is_dirty) {
3103                 char *new_path;
3104                 if (o->call_depth) {
3105                         remove_file_from_cache(path);
3106                 } else {
3107                         if (!mfi.clean) {
3108                                 if (update_stages(o, path, &one, &a, &b))
3109                                         return -1;
3110                         } else {
3111                                 int file_from_stage2 = was_tracked(o, path);
3112                                 struct diff_filespec merged;
3113                                 oidcpy(&merged.oid, &mfi.oid);
3114                                 merged.mode = mfi.mode;
3115
3116                                 if (update_stages(o, path, NULL,
3117                                                   file_from_stage2 ? &merged : NULL,
3118                                                   file_from_stage2 ? NULL : &merged))
3119                                         return -1;
3120                         }
3121
3122                 }
3123                 new_path = unique_path(o, path, rename_conflict_info->branch1);
3124                 if (is_dirty) {
3125                         output(o, 1, _("Refusing to lose dirty file at %s"),
3126                                path);
3127                 }
3128                 output(o, 1, _("Adding as %s instead"), new_path);
3129                 if (update_file(o, 0, &mfi.oid, mfi.mode, new_path)) {
3130                         free(new_path);
3131                         return -1;
3132                 }
3133                 free(new_path);
3134                 mfi.clean = 0;
3135         } else if (update_file(o, mfi.clean, &mfi.oid, mfi.mode, path))
3136                 return -1;
3137         return !is_dirty && mfi.clean;
3138 }
3139
3140 static int handle_rename_normal(struct merge_options *o,
3141                                 const char *path,
3142                                 struct object_id *o_oid, unsigned int o_mode,
3143                                 struct object_id *a_oid, unsigned int a_mode,
3144                                 struct object_id *b_oid, unsigned int b_mode,
3145                                 struct rename_conflict_info *ci)
3146 {
3147         /* Merge the content and write it out */
3148         return handle_content_merge(o, path, was_dirty(o, path),
3149                                     o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3150                                     ci);
3151 }
3152
3153 /* Per entry merge function */
3154 static int process_entry(struct merge_options *o,
3155                          const char *path, struct stage_data *entry)
3156 {
3157         int clean_merge = 1;
3158         int normalize = o->renormalize;
3159         unsigned o_mode = entry->stages[1].mode;
3160         unsigned a_mode = entry->stages[2].mode;
3161         unsigned b_mode = entry->stages[3].mode;
3162         struct object_id *o_oid = stage_oid(&entry->stages[1].oid, o_mode);
3163         struct object_id *a_oid = stage_oid(&entry->stages[2].oid, a_mode);
3164         struct object_id *b_oid = stage_oid(&entry->stages[3].oid, b_mode);
3165
3166         entry->processed = 1;
3167         if (entry->rename_conflict_info) {
3168                 struct rename_conflict_info *conflict_info = entry->rename_conflict_info;
3169                 switch (conflict_info->rename_type) {
3170                 case RENAME_NORMAL:
3171                 case RENAME_ONE_FILE_TO_ONE:
3172                         clean_merge = handle_rename_normal(o,
3173                                                            path,
3174                                                            o_oid, o_mode,
3175                                                            a_oid, a_mode,
3176                                                            b_oid, b_mode,
3177                                                            conflict_info);
3178                         break;
3179                 case RENAME_VIA_DIR:
3180                         clean_merge = 1;
3181                         if (handle_rename_via_dir(o,
3182                                                   conflict_info->pair1,
3183                                                   conflict_info->branch1,
3184                                                   conflict_info->branch2))
3185                                 clean_merge = -1;
3186                         break;
3187                 case RENAME_DELETE:
3188                         clean_merge = 0;
3189                         if (handle_rename_delete(o,
3190                                                  conflict_info->pair1,
3191                                                  conflict_info->branch1,
3192                                                  conflict_info->branch2))
3193                                 clean_merge = -1;
3194                         break;
3195                 case RENAME_ONE_FILE_TO_TWO:
3196                         clean_merge = 0;
3197                         if (handle_rename_rename_1to2(o, conflict_info))
3198                                 clean_merge = -1;
3199                         break;
3200                 case RENAME_TWO_FILES_TO_ONE:
3201                         clean_merge = 0;
3202                         if (handle_rename_rename_2to1(o, conflict_info))
3203                                 clean_merge = -1;
3204                         break;
3205                 default:
3206                         entry->processed = 0;
3207                         break;
3208                 }
3209         } else if (o_oid && (!a_oid || !b_oid)) {
3210                 /* Case A: Deleted in one */
3211                 if ((!a_oid && !b_oid) ||
3212                     (!b_oid && blob_unchanged(o, o_oid, o_mode, a_oid, a_mode, normalize, path)) ||
3213                     (!a_oid && blob_unchanged(o, o_oid, o_mode, b_oid, b_mode, normalize, path))) {
3214                         /* Deleted in both or deleted in one and
3215                          * unchanged in the other */
3216                         if (a_oid)
3217                                 output(o, 2, _("Removing %s"), path);
3218                         /* do not touch working file if it did not exist */
3219                         remove_file(o, 1, path, !a_oid);
3220                 } else {
3221                         /* Modify/delete; deleted side may have put a directory in the way */
3222                         clean_merge = 0;
3223                         if (handle_modify_delete(o, path, o_oid, o_mode,
3224                                                  a_oid, a_mode, b_oid, b_mode))
3225                                 clean_merge = -1;
3226                 }
3227         } else if ((!o_oid && a_oid && !b_oid) ||
3228                    (!o_oid && !a_oid && b_oid)) {
3229                 /* Case B: Added in one. */
3230                 /* [nothing|directory] -> ([nothing|directory], file) */
3231
3232                 const char *add_branch;
3233                 const char *other_branch;
3234                 unsigned mode;
3235                 const struct object_id *oid;
3236                 const char *conf;
3237
3238                 if (a_oid) {
3239                         add_branch = o->branch1;
3240                         other_branch = o->branch2;
3241                         mode = a_mode;
3242                         oid = a_oid;
3243                         conf = _("file/directory");
3244                 } else {
3245                         add_branch = o->branch2;
3246                         other_branch = o->branch1;
3247                         mode = b_mode;
3248                         oid = b_oid;
3249                         conf = _("directory/file");
3250                 }
3251                 if (dir_in_way(path,
3252                                !o->call_depth && !S_ISGITLINK(a_mode),
3253                                0)) {
3254                         char *new_path = unique_path(o, path, add_branch);
3255                         clean_merge = 0;
3256                         output(o, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3257                                "Adding %s as %s"),
3258                                conf, path, other_branch, path, new_path);
3259                         if (update_file(o, 0, oid, mode, new_path))
3260                                 clean_merge = -1;
3261                         else if (o->call_depth)
3262                                 remove_file_from_cache(path);
3263                         free(new_path);
3264                 } else {
3265                         output(o, 2, _("Adding %s"), path);
3266                         /* do not overwrite file if already present */
3267                         if (update_file_flags(o, oid, mode, path, 1, !a_oid))
3268                                 clean_merge = -1;
3269                 }
3270         } else if (a_oid && b_oid) {
3271                 /* Case C: Added in both (check for same permissions) and */
3272                 /* case D: Modified in both, but differently. */
3273                 int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3274                 clean_merge = handle_content_merge(o, path, is_dirty,
3275                                                    o_oid, o_mode,
3276                                                    a_oid, a_mode,
3277                                                    b_oid, b_mode,
3278                                                    NULL);
3279         } else if (!o_oid && !a_oid && !b_oid) {
3280                 /*
3281                  * this entry was deleted altogether. a_mode == 0 means
3282                  * we had that path and want to actively remove it.
3283                  */
3284                 remove_file(o, 1, path, !a_mode);
3285         } else
3286                 BUG("fatal merge failure, shouldn't happen.");
3287
3288         return clean_merge;
3289 }
3290
3291 int merge_trees(struct merge_options *o,
3292                 struct tree *head,
3293                 struct tree *merge,
3294                 struct tree *common,
3295                 struct tree **result)
3296 {
3297         int code, clean;
3298         struct strbuf sb = STRBUF_INIT;
3299
3300         if (!o->call_depth && index_has_changes(&the_index, head, &sb)) {
3301                 err(o, _("Your local changes to the following files would be overwritten by merge:\n  %s"),
3302                     sb.buf);
3303                 return -1;
3304         }
3305
3306         if (o->subtree_shift) {
3307                 merge = shift_tree_object(head, merge, o->subtree_shift);
3308                 common = shift_tree_object(head, common, o->subtree_shift);
3309         }
3310
3311         if (oid_eq(&common->object.oid, &merge->object.oid)) {
3312                 output(o, 0, _("Already up to date!"));
3313                 *result = head;
3314                 return 1;
3315         }
3316
3317         code = unpack_trees_start(o, common, head, merge);
3318
3319         if (code != 0) {
3320                 if (show(o, 4) || o->call_depth)
3321                         err(o, _("merging of trees %s and %s failed"),
3322                             oid_to_hex(&head->object.oid),
3323                             oid_to_hex(&merge->object.oid));
3324                 unpack_trees_finish(o);
3325                 return -1;
3326         }
3327
3328         if (unmerged_cache()) {
3329                 struct string_list *entries;
3330                 struct rename_info re_info;
3331                 int i;
3332                 /*
3333                  * Only need the hashmap while processing entries, so
3334                  * initialize it here and free it when we are done running
3335                  * through the entries. Keeping it in the merge_options as
3336                  * opposed to decaring a local hashmap is for convenience
3337                  * so that we don't have to pass it to around.
3338                  */
3339                 hashmap_init(&o->current_file_dir_set, path_hashmap_cmp, NULL, 512);
3340                 get_files_dirs(o, head);
3341                 get_files_dirs(o, merge);
3342
3343                 entries = get_unmerged();
3344                 clean = detect_and_process_renames(o, common, head, merge,
3345                                                    entries, &re_info);
3346                 record_df_conflict_files(o, entries);
3347                 if (clean < 0)
3348                         goto cleanup;
3349                 for (i = entries->nr-1; 0 <= i; i--) {
3350                         const char *path = entries->items[i].string;
3351                         struct stage_data *e = entries->items[i].util;
3352                         if (!e->processed) {
3353                                 int ret = process_entry(o, path, e);
3354                                 if (!ret)
3355                                         clean = 0;
3356                                 else if (ret < 0) {
3357                                         clean = ret;
3358                                         goto cleanup;
3359                                 }
3360                         }
3361                 }
3362                 for (i = 0; i < entries->nr; i++) {
3363                         struct stage_data *e = entries->items[i].util;
3364                         if (!e->processed)
3365                                 BUG("unprocessed path??? %s",
3366                                     entries->items[i].string);
3367                 }
3368
3369         cleanup:
3370                 final_cleanup_renames(&re_info);
3371
3372                 string_list_clear(entries, 1);
3373                 free(entries);
3374
3375                 hashmap_free(&o->current_file_dir_set, 1);
3376
3377                 if (clean < 0) {
3378                         unpack_trees_finish(o);
3379                         return clean;
3380                 }
3381         }
3382         else
3383                 clean = 1;
3384
3385         unpack_trees_finish(o);
3386
3387         if (o->call_depth && !(*result = write_tree_from_memory(o)))
3388                 return -1;
3389
3390         return clean;
3391 }
3392
3393 static struct commit_list *reverse_commit_list(struct commit_list *list)
3394 {
3395         struct commit_list *next = NULL, *current, *backup;
3396         for (current = list; current; current = backup) {
3397                 backup = current->next;
3398                 current->next = next;
3399                 next = current;
3400         }
3401         return next;
3402 }
3403
3404 /*
3405  * Merge the commits h1 and h2, return the resulting virtual
3406  * commit object and a flag indicating the cleanness of the merge.
3407  */
3408 int merge_recursive(struct merge_options *o,
3409                     struct commit *h1,
3410                     struct commit *h2,
3411                     struct commit_list *ca,
3412                     struct commit **result)
3413 {
3414         struct commit_list *iter;
3415         struct commit *merged_common_ancestors;
3416         struct tree *mrtree;
3417         int clean;
3418
3419         if (show(o, 4)) {
3420                 output(o, 4, _("Merging:"));
3421                 output_commit_title(o, h1);
3422                 output_commit_title(o, h2);
3423         }
3424
3425         if (!ca) {
3426                 ca = get_merge_bases(h1, h2);
3427                 ca = reverse_commit_list(ca);
3428         }
3429
3430         if (show(o, 5)) {
3431                 unsigned cnt = commit_list_count(ca);
3432
3433                 output(o, 5, Q_("found %u common ancestor:",
3434                                 "found %u common ancestors:", cnt), cnt);
3435                 for (iter = ca; iter; iter = iter->next)
3436                         output_commit_title(o, iter->item);
3437         }
3438
3439         merged_common_ancestors = pop_commit(&ca);
3440         if (merged_common_ancestors == NULL) {
3441                 /* if there is no common ancestor, use an empty tree */
3442                 struct tree *tree;
3443
3444                 tree = lookup_tree(the_repository, the_repository->hash_algo->empty_tree);
3445                 merged_common_ancestors = make_virtual_commit(tree, "ancestor");
3446         }
3447
3448         for (iter = ca; iter; iter = iter->next) {
3449                 const char *saved_b1, *saved_b2;
3450                 o->call_depth++;
3451                 /*
3452                  * When the merge fails, the result contains files
3453                  * with conflict markers. The cleanness flag is
3454                  * ignored (unless indicating an error), it was never
3455                  * actually used, as result of merge_trees has always
3456                  * overwritten it: the committed "conflicts" were
3457                  * already resolved.
3458                  */
3459                 discard_cache();
3460                 saved_b1 = o->branch1;
3461                 saved_b2 = o->branch2;
3462                 o->branch1 = "Temporary merge branch 1";
3463                 o->branch2 = "Temporary merge branch 2";
3464                 if (merge_recursive(o, merged_common_ancestors, iter->item,
3465                                     NULL, &merged_common_ancestors) < 0)
3466                         return -1;
3467                 o->branch1 = saved_b1;
3468                 o->branch2 = saved_b2;
3469                 o->call_depth--;
3470
3471                 if (!merged_common_ancestors)
3472                         return err(o, _("merge returned no commit"));
3473         }
3474
3475         discard_cache();
3476         if (!o->call_depth)
3477                 read_cache();
3478
3479         o->ancestor = "merged common ancestors";
3480         clean = merge_trees(o, get_commit_tree(h1), get_commit_tree(h2),
3481                             get_commit_tree(merged_common_ancestors),
3482                             &mrtree);
3483         if (clean < 0) {
3484                 flush_output(o);
3485                 return clean;
3486         }
3487
3488         if (o->call_depth) {
3489                 *result = make_virtual_commit(mrtree, "merged tree");
3490                 commit_list_insert(h1, &(*result)->parents);
3491                 commit_list_insert(h2, &(*result)->parents->next);
3492         }
3493         flush_output(o);
3494         if (!o->call_depth && o->buffer_output < 2)
3495                 strbuf_release(&o->obuf);
3496         if (show(o, 2))
3497                 diff_warn_rename_limit("merge.renamelimit",
3498                                        o->needed_rename_limit, 0);
3499         return clean;
3500 }
3501
3502 static struct commit *get_ref(const struct object_id *oid, const char *name)
3503 {
3504         struct object *object;
3505
3506         object = deref_tag(the_repository, parse_object(the_repository, oid),
3507                            name,
3508                            strlen(name));
3509         if (!object)
3510                 return NULL;
3511         if (object->type == OBJ_TREE)
3512                 return make_virtual_commit((struct tree*)object, name);
3513         if (object->type != OBJ_COMMIT)
3514                 return NULL;
3515         if (parse_commit((struct commit *)object))
3516                 return NULL;
3517         return (struct commit *)object;
3518 }
3519
3520 int merge_recursive_generic(struct merge_options *o,
3521                             const struct object_id *head,
3522                             const struct object_id *merge,
3523                             int num_base_list,
3524                             const struct object_id **base_list,
3525                             struct commit **result)
3526 {
3527         int clean;
3528         struct lock_file lock = LOCK_INIT;
3529         struct commit *head_commit = get_ref(head, o->branch1);
3530         struct commit *next_commit = get_ref(merge, o->branch2);
3531         struct commit_list *ca = NULL;
3532
3533         if (base_list) {
3534                 int i;
3535                 for (i = 0; i < num_base_list; ++i) {
3536                         struct commit *base;
3537                         if (!(base = get_ref(base_list[i], oid_to_hex(base_list[i]))))
3538                                 return err(o, _("Could not parse object '%s'"),
3539                                            oid_to_hex(base_list[i]));
3540                         commit_list_insert(base, &ca);
3541                 }
3542         }
3543
3544         hold_locked_index(&lock, LOCK_DIE_ON_ERROR);
3545         clean = merge_recursive(o, head_commit, next_commit, ca,
3546                                 result);
3547         if (clean < 0) {
3548                 rollback_lock_file(&lock);
3549                 return clean;
3550         }
3551
3552         if (write_locked_index(&the_index, &lock,
3553                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3554                 return err(o, _("Unable to write index."));
3555
3556         return clean ? 0 : 1;
3557 }
3558
3559 static void merge_recursive_config(struct merge_options *o)
3560 {
3561         char *value = NULL;
3562         git_config_get_int("merge.verbosity", &o->verbosity);
3563         git_config_get_int("diff.renamelimit", &o->diff_rename_limit);
3564         git_config_get_int("merge.renamelimit", &o->merge_rename_limit);
3565         if (!git_config_get_string("diff.renames", &value)) {
3566                 o->diff_detect_rename = git_config_rename("diff.renames", value);
3567                 free(value);
3568         }
3569         if (!git_config_get_string("merge.renames", &value)) {
3570                 o->merge_detect_rename = git_config_rename("merge.renames", value);
3571                 free(value);
3572         }
3573         git_config(git_xmerge_config, NULL);
3574 }
3575
3576 void init_merge_options(struct merge_options *o)
3577 {
3578         const char *merge_verbosity;
3579         memset(o, 0, sizeof(struct merge_options));
3580         o->verbosity = 2;
3581         o->buffer_output = 1;
3582         o->diff_rename_limit = -1;
3583         o->merge_rename_limit = -1;
3584         o->renormalize = 0;
3585         o->diff_detect_rename = -1;
3586         o->merge_detect_rename = -1;
3587         o->detect_directory_renames = 1;
3588         merge_recursive_config(o);
3589         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3590         if (merge_verbosity)
3591                 o->verbosity = strtol(merge_verbosity, NULL, 10);
3592         if (o->verbosity >= 5)
3593                 o->buffer_output = 0;
3594         strbuf_init(&o->obuf, 0);
3595         string_list_init(&o->df_conflict_file_set, 1);
3596 }
3597
3598 int parse_merge_opt(struct merge_options *o, const char *s)
3599 {
3600         const char *arg;
3601
3602         if (!s || !*s)
3603                 return -1;
3604         if (!strcmp(s, "ours"))
3605                 o->recursive_variant = MERGE_RECURSIVE_OURS;
3606         else if (!strcmp(s, "theirs"))
3607                 o->recursive_variant = MERGE_RECURSIVE_THEIRS;
3608         else if (!strcmp(s, "subtree"))
3609                 o->subtree_shift = "";
3610         else if (skip_prefix(s, "subtree=", &arg))
3611                 o->subtree_shift = arg;
3612         else if (!strcmp(s, "patience"))
3613                 o->xdl_opts = DIFF_WITH_ALG(o, PATIENCE_DIFF);
3614         else if (!strcmp(s, "histogram"))
3615                 o->xdl_opts = DIFF_WITH_ALG(o, HISTOGRAM_DIFF);
3616         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3617                 long value = parse_algorithm_value(arg);
3618                 if (value < 0)
3619                         return -1;
3620                 /* clear out previous settings */
3621                 DIFF_XDL_CLR(o, NEED_MINIMAL);
3622                 o->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3623                 o->xdl_opts |= value;
3624         }
3625         else if (!strcmp(s, "ignore-space-change"))
3626                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_CHANGE);
3627         else if (!strcmp(s, "ignore-all-space"))
3628                 DIFF_XDL_SET(o, IGNORE_WHITESPACE);
3629         else if (!strcmp(s, "ignore-space-at-eol"))
3630                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_AT_EOL);
3631         else if (!strcmp(s, "ignore-cr-at-eol"))
3632                 DIFF_XDL_SET(o, IGNORE_CR_AT_EOL);
3633         else if (!strcmp(s, "renormalize"))
3634                 o->renormalize = 1;
3635         else if (!strcmp(s, "no-renormalize"))
3636                 o->renormalize = 0;
3637         else if (!strcmp(s, "no-renames"))
3638                 o->merge_detect_rename = 0;
3639         else if (!strcmp(s, "find-renames")) {
3640                 o->merge_detect_rename = 1;
3641                 o->rename_score = 0;
3642         }
3643         else if (skip_prefix(s, "find-renames=", &arg) ||
3644                  skip_prefix(s, "rename-threshold=", &arg)) {
3645                 if ((o->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3646                         return -1;
3647                 o->merge_detect_rename = 1;
3648         }
3649         else
3650                 return -1;
3651         return 0;
3652 }