Merge branch 'rs/mailinfo-format-flowed'
[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 = xcalloc(1, sizeof(struct rename_conflict_info));
232         ci->rename_type = rename_type;
233         ci->pair1 = pair1;
234         ci->branch1 = branch1;
235         ci->branch2 = branch2;
236
237         ci->dst_entry1 = dst_entry1;
238         dst_entry1->rename_conflict_info = ci;
239         dst_entry1->processed = 0;
240
241         assert(!pair2 == !dst_entry2);
242         if (dst_entry2) {
243                 ci->dst_entry2 = dst_entry2;
244                 ci->pair2 = pair2;
245                 dst_entry2->rename_conflict_info = ci;
246         }
247
248         if (rename_type == RENAME_TWO_FILES_TO_ONE) {
249                 /*
250                  * For each rename, there could have been
251                  * modifications on the side of history where that
252                  * file was not renamed.
253                  */
254                 int ostage1 = o->branch1 == branch1 ? 3 : 2;
255                 int ostage2 = ostage1 ^ 1;
256
257                 ci->ren1_other.path = pair1->one->path;
258                 oidcpy(&ci->ren1_other.oid, &src_entry1->stages[ostage1].oid);
259                 ci->ren1_other.mode = src_entry1->stages[ostage1].mode;
260
261                 ci->ren2_other.path = pair2->one->path;
262                 oidcpy(&ci->ren2_other.oid, &src_entry2->stages[ostage2].oid);
263                 ci->ren2_other.mode = src_entry2->stages[ostage2].mode;
264         }
265 }
266
267 static int show(struct merge_options *o, int v)
268 {
269         return (!o->call_depth && o->verbosity >= v) || o->verbosity >= 5;
270 }
271
272 __attribute__((format (printf, 3, 4)))
273 static void output(struct merge_options *o, int v, const char *fmt, ...)
274 {
275         va_list ap;
276
277         if (!show(o, v))
278                 return;
279
280         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
281
282         va_start(ap, fmt);
283         strbuf_vaddf(&o->obuf, fmt, ap);
284         va_end(ap);
285
286         strbuf_addch(&o->obuf, '\n');
287         if (!o->buffer_output)
288                 flush_output(o);
289 }
290
291 static void output_commit_title(struct merge_options *o, struct commit *commit)
292 {
293         struct merge_remote_desc *desc;
294
295         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
296         desc = merge_remote_util(commit);
297         if (desc)
298                 strbuf_addf(&o->obuf, "virtual %s\n", desc->name);
299         else {
300                 strbuf_add_unique_abbrev(&o->obuf, &commit->object.oid,
301                                          DEFAULT_ABBREV);
302                 strbuf_addch(&o->obuf, ' ');
303                 if (parse_commit(commit) != 0)
304                         strbuf_addstr(&o->obuf, _("(bad commit)\n"));
305                 else {
306                         const char *title;
307                         const char *msg = get_commit_buffer(commit, NULL);
308                         int len = find_commit_subject(msg, &title);
309                         if (len)
310                                 strbuf_addf(&o->obuf, "%.*s\n", len, title);
311                         unuse_commit_buffer(commit, msg);
312                 }
313         }
314         flush_output(o);
315 }
316
317 static int add_cacheinfo(struct merge_options *o,
318                          unsigned int mode, const struct object_id *oid,
319                          const char *path, int stage, int refresh, int options)
320 {
321         struct cache_entry *ce;
322         int ret;
323
324         ce = make_cache_entry(&the_index, mode, oid ? oid : &null_oid, path, stage, 0);
325         if (!ce)
326                 return err(o, _("add_cacheinfo failed for path '%s'; merge aborting."), path);
327
328         ret = add_cache_entry(ce, options);
329         if (refresh) {
330                 struct cache_entry *nce;
331
332                 nce = refresh_cache_entry(&the_index, ce, CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
333                 if (!nce)
334                         return err(o, _("add_cacheinfo failed to refresh for path '%s'; merge aborting."), path);
335                 if (nce != ce)
336                         ret = add_cache_entry(nce, options);
337         }
338         return ret;
339 }
340
341 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
342 {
343         parse_tree(tree);
344         init_tree_desc(desc, tree->buffer, tree->size);
345 }
346
347 static int unpack_trees_start(struct merge_options *o,
348                               struct tree *common,
349                               struct tree *head,
350                               struct tree *merge)
351 {
352         int rc;
353         struct tree_desc t[3];
354         struct index_state tmp_index = { NULL };
355
356         memset(&o->unpack_opts, 0, sizeof(o->unpack_opts));
357         if (o->call_depth)
358                 o->unpack_opts.index_only = 1;
359         else
360                 o->unpack_opts.update = 1;
361         o->unpack_opts.merge = 1;
362         o->unpack_opts.head_idx = 2;
363         o->unpack_opts.fn = threeway_merge;
364         o->unpack_opts.src_index = &the_index;
365         o->unpack_opts.dst_index = &tmp_index;
366         o->unpack_opts.aggressive = !merge_detect_rename(o);
367         setup_unpack_trees_porcelain(&o->unpack_opts, "merge");
368
369         init_tree_desc_from_tree(t+0, common);
370         init_tree_desc_from_tree(t+1, head);
371         init_tree_desc_from_tree(t+2, merge);
372
373         rc = unpack_trees(3, t, &o->unpack_opts);
374         cache_tree_free(&active_cache_tree);
375
376         /*
377          * Update the_index to match the new results, AFTER saving a copy
378          * in o->orig_index.  Update src_index to point to the saved copy.
379          * (verify_uptodate() checks src_index, and the original index is
380          * the one that had the necessary modification timestamps.)
381          */
382         o->orig_index = the_index;
383         the_index = tmp_index;
384         o->unpack_opts.src_index = &o->orig_index;
385
386         return rc;
387 }
388
389 static void unpack_trees_finish(struct merge_options *o)
390 {
391         discard_index(&o->orig_index);
392         clear_unpack_trees_porcelain(&o->unpack_opts);
393 }
394
395 struct tree *write_tree_from_memory(struct merge_options *o)
396 {
397         struct tree *result = NULL;
398
399         if (unmerged_cache()) {
400                 int i;
401                 fprintf(stderr, "BUG: There are unmerged index entries:\n");
402                 for (i = 0; i < active_nr; i++) {
403                         const struct cache_entry *ce = active_cache[i];
404                         if (ce_stage(ce))
405                                 fprintf(stderr, "BUG: %d %.*s\n", ce_stage(ce),
406                                         (int)ce_namelen(ce), ce->name);
407                 }
408                 BUG("unmerged index entries in merge-recursive.c");
409         }
410
411         if (!active_cache_tree)
412                 active_cache_tree = cache_tree();
413
414         if (!cache_tree_fully_valid(active_cache_tree) &&
415             cache_tree_update(&the_index, 0) < 0) {
416                 err(o, _("error building trees"));
417                 return NULL;
418         }
419
420         result = lookup_tree(the_repository, &active_cache_tree->oid);
421
422         return result;
423 }
424
425 static int save_files_dirs(const struct object_id *oid,
426                            struct strbuf *base, const char *path,
427                            unsigned int mode, int stage, void *context)
428 {
429         struct path_hashmap_entry *entry;
430         int baselen = base->len;
431         struct merge_options *o = context;
432
433         strbuf_addstr(base, path);
434
435         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
436         hashmap_entry_init(entry, path_hash(entry->path));
437         hashmap_add(&o->current_file_dir_set, entry);
438
439         strbuf_setlen(base, baselen);
440         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
441 }
442
443 static void get_files_dirs(struct merge_options *o, struct tree *tree)
444 {
445         struct pathspec match_all;
446         memset(&match_all, 0, sizeof(match_all));
447         read_tree_recursive(tree, "", 0, 0, &match_all, save_files_dirs, o);
448 }
449
450 static int get_tree_entry_if_blob(const struct object_id *tree,
451                                   const char *path,
452                                   struct object_id *hashy,
453                                   unsigned int *mode_o)
454 {
455         int ret;
456
457         ret = get_tree_entry(tree, path, hashy, mode_o);
458         if (S_ISDIR(*mode_o)) {
459                 oidcpy(hashy, &null_oid);
460                 *mode_o = 0;
461         }
462         return ret;
463 }
464
465 /*
466  * Returns an index_entry instance which doesn't have to correspond to
467  * a real cache entry in Git's index.
468  */
469 static struct stage_data *insert_stage_data(const char *path,
470                 struct tree *o, struct tree *a, struct tree *b,
471                 struct string_list *entries)
472 {
473         struct string_list_item *item;
474         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
475         get_tree_entry_if_blob(&o->object.oid, path,
476                                &e->stages[1].oid, &e->stages[1].mode);
477         get_tree_entry_if_blob(&a->object.oid, path,
478                                &e->stages[2].oid, &e->stages[2].mode);
479         get_tree_entry_if_blob(&b->object.oid, path,
480                                &e->stages[3].oid, &e->stages[3].mode);
481         item = string_list_insert(entries, path);
482         item->util = e;
483         return e;
484 }
485
486 /*
487  * Create a dictionary mapping file names to stage_data objects. The
488  * dictionary contains one entry for every path with a non-zero stage entry.
489  */
490 static struct string_list *get_unmerged(void)
491 {
492         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
493         int i;
494
495         unmerged->strdup_strings = 1;
496
497         for (i = 0; i < active_nr; i++) {
498                 struct string_list_item *item;
499                 struct stage_data *e;
500                 const struct cache_entry *ce = active_cache[i];
501                 if (!ce_stage(ce))
502                         continue;
503
504                 item = string_list_lookup(unmerged, ce->name);
505                 if (!item) {
506                         item = string_list_insert(unmerged, ce->name);
507                         item->util = xcalloc(1, sizeof(struct stage_data));
508                 }
509                 e = item->util;
510                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
511                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
512         }
513
514         return unmerged;
515 }
516
517 static int string_list_df_name_compare(const char *one, const char *two)
518 {
519         int onelen = strlen(one);
520         int twolen = strlen(two);
521         /*
522          * Here we only care that entries for D/F conflicts are
523          * adjacent, in particular with the file of the D/F conflict
524          * appearing before files below the corresponding directory.
525          * The order of the rest of the list is irrelevant for us.
526          *
527          * To achieve this, we sort with df_name_compare and provide
528          * the mode S_IFDIR so that D/F conflicts will sort correctly.
529          * We use the mode S_IFDIR for everything else for simplicity,
530          * since in other cases any changes in their order due to
531          * sorting cause no problems for us.
532          */
533         int cmp = df_name_compare(one, onelen, S_IFDIR,
534                                   two, twolen, S_IFDIR);
535         /*
536          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
537          * that 'foo' comes before 'foo/bar'.
538          */
539         if (cmp)
540                 return cmp;
541         return onelen - twolen;
542 }
543
544 static void record_df_conflict_files(struct merge_options *o,
545                                      struct string_list *entries)
546 {
547         /* If there is a D/F conflict and the file for such a conflict
548          * currently exists in the working tree, we want to allow it to be
549          * removed to make room for the corresponding directory if needed.
550          * The files underneath the directories of such D/F conflicts will
551          * be processed before the corresponding file involved in the D/F
552          * conflict.  If the D/F directory ends up being removed by the
553          * merge, then we won't have to touch the D/F file.  If the D/F
554          * directory needs to be written to the working copy, then the D/F
555          * file will simply be removed (in make_room_for_path()) to make
556          * room for the necessary paths.  Note that if both the directory
557          * and the file need to be present, then the D/F file will be
558          * reinstated with a new unique name at the time it is processed.
559          */
560         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
561         const char *last_file = NULL;
562         int last_len = 0;
563         int i;
564
565         /*
566          * If we're merging merge-bases, we don't want to bother with
567          * any working directory changes.
568          */
569         if (o->call_depth)
570                 return;
571
572         /* Ensure D/F conflicts are adjacent in the entries list. */
573         for (i = 0; i < entries->nr; i++) {
574                 struct string_list_item *next = &entries->items[i];
575                 string_list_append(&df_sorted_entries, next->string)->util =
576                                    next->util;
577         }
578         df_sorted_entries.cmp = string_list_df_name_compare;
579         string_list_sort(&df_sorted_entries);
580
581         string_list_clear(&o->df_conflict_file_set, 1);
582         for (i = 0; i < df_sorted_entries.nr; i++) {
583                 const char *path = df_sorted_entries.items[i].string;
584                 int len = strlen(path);
585                 struct stage_data *e = df_sorted_entries.items[i].util;
586
587                 /*
588                  * Check if last_file & path correspond to a D/F conflict;
589                  * i.e. whether path is last_file+'/'+<something>.
590                  * If so, record that it's okay to remove last_file to make
591                  * room for path and friends if needed.
592                  */
593                 if (last_file &&
594                     len > last_len &&
595                     memcmp(path, last_file, last_len) == 0 &&
596                     path[last_len] == '/') {
597                         string_list_insert(&o->df_conflict_file_set, last_file);
598                 }
599
600                 /*
601                  * Determine whether path could exist as a file in the
602                  * working directory as a possible D/F conflict.  This
603                  * will only occur when it exists in stage 2 as a
604                  * file.
605                  */
606                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
607                         last_file = path;
608                         last_len = len;
609                 } else {
610                         last_file = NULL;
611                 }
612         }
613         string_list_clear(&df_sorted_entries, 0);
614 }
615
616 struct rename {
617         struct diff_filepair *pair;
618         /*
619          * Purpose of src_entry and dst_entry:
620          *
621          * If 'before' is renamed to 'after' then src_entry will contain
622          * the versions of 'before' from the merge_base, HEAD, and MERGE in
623          * stages 1, 2, and 3; dst_entry will contain the respective
624          * versions of 'after' in corresponding locations.  Thus, we have a
625          * total of six modes and oids, though some will be null.  (Stage 0
626          * is ignored; we're interested in handling conflicts.)
627          *
628          * Since we don't turn on break-rewrites by default, neither
629          * src_entry nor dst_entry can have all three of their stages have
630          * non-null oids, meaning at most four of the six will be non-null.
631          * Also, since this is a rename, both src_entry and dst_entry will
632          * have at least one non-null oid, meaning at least two will be
633          * non-null.  Of the six oids, a typical rename will have three be
634          * non-null.  Only two implies a rename/delete, and four implies a
635          * rename/add.
636          */
637         struct stage_data *src_entry;
638         struct stage_data *dst_entry;
639         unsigned add_turned_into_rename:1;
640         unsigned processed:1;
641 };
642
643 static int update_stages(struct merge_options *opt, const char *path,
644                          const struct diff_filespec *o,
645                          const struct diff_filespec *a,
646                          const struct diff_filespec *b)
647 {
648
649         /*
650          * NOTE: It is usually a bad idea to call update_stages on a path
651          * before calling update_file on that same path, since it can
652          * sometimes lead to spurious "refusing to lose untracked file..."
653          * messages from update_file (via make_room_for path via
654          * would_lose_untracked).  Instead, reverse the order of the calls
655          * (executing update_file first and then update_stages).
656          */
657         int clear = 1;
658         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
659         if (clear)
660                 if (remove_file_from_cache(path))
661                         return -1;
662         if (o)
663                 if (add_cacheinfo(opt, o->mode, &o->oid, path, 1, 0, options))
664                         return -1;
665         if (a)
666                 if (add_cacheinfo(opt, a->mode, &a->oid, path, 2, 0, options))
667                         return -1;
668         if (b)
669                 if (add_cacheinfo(opt, b->mode, &b->oid, path, 3, 0, options))
670                         return -1;
671         return 0;
672 }
673
674 static int update_stages_for_stage_data(struct merge_options *opt,
675                                         const char *path,
676                                         const struct stage_data *stage_data)
677 {
678         struct diff_filespec o, a, b;
679
680         o.mode = stage_data->stages[1].mode;
681         oidcpy(&o.oid, &stage_data->stages[1].oid);
682
683         a.mode = stage_data->stages[2].mode;
684         oidcpy(&a.oid, &stage_data->stages[2].oid);
685
686         b.mode = stage_data->stages[3].mode;
687         oidcpy(&b.oid, &stage_data->stages[3].oid);
688
689         return update_stages(opt, path,
690                              is_null_oid(&o.oid) ? NULL : &o,
691                              is_null_oid(&a.oid) ? NULL : &a,
692                              is_null_oid(&b.oid) ? NULL : &b);
693 }
694
695 static void update_entry(struct stage_data *entry,
696                          struct diff_filespec *o,
697                          struct diff_filespec *a,
698                          struct diff_filespec *b)
699 {
700         entry->processed = 0;
701         entry->stages[1].mode = o->mode;
702         entry->stages[2].mode = a->mode;
703         entry->stages[3].mode = b->mode;
704         oidcpy(&entry->stages[1].oid, &o->oid);
705         oidcpy(&entry->stages[2].oid, &a->oid);
706         oidcpy(&entry->stages[3].oid, &b->oid);
707 }
708
709 static int remove_file(struct merge_options *o, int clean,
710                        const char *path, int no_wd)
711 {
712         int update_cache = o->call_depth || clean;
713         int update_working_directory = !o->call_depth && !no_wd;
714
715         if (update_cache) {
716                 if (remove_file_from_cache(path))
717                         return -1;
718         }
719         if (update_working_directory) {
720                 if (ignore_case) {
721                         struct cache_entry *ce;
722                         ce = cache_file_exists(path, strlen(path), ignore_case);
723                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
724                                 return 0;
725                 }
726                 if (remove_path(path))
727                         return -1;
728         }
729         return 0;
730 }
731
732 /* add a string to a strbuf, but converting "/" to "_" */
733 static void add_flattened_path(struct strbuf *out, const char *s)
734 {
735         size_t i = out->len;
736         strbuf_addstr(out, s);
737         for (; i < out->len; i++)
738                 if (out->buf[i] == '/')
739                         out->buf[i] = '_';
740 }
741
742 static char *unique_path(struct merge_options *o, const char *path, const char *branch)
743 {
744         struct path_hashmap_entry *entry;
745         struct strbuf newpath = STRBUF_INIT;
746         int suffix = 0;
747         size_t base_len;
748
749         strbuf_addf(&newpath, "%s~", path);
750         add_flattened_path(&newpath, branch);
751
752         base_len = newpath.len;
753         while (hashmap_get_from_hash(&o->current_file_dir_set,
754                                      path_hash(newpath.buf), newpath.buf) ||
755                (!o->call_depth && file_exists(newpath.buf))) {
756                 strbuf_setlen(&newpath, base_len);
757                 strbuf_addf(&newpath, "_%d", suffix++);
758         }
759
760         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
761         hashmap_entry_init(entry, path_hash(entry->path));
762         hashmap_add(&o->current_file_dir_set, entry);
763         return strbuf_detach(&newpath, NULL);
764 }
765
766 /**
767  * Check whether a directory in the index is in the way of an incoming
768  * file.  Return 1 if so.  If check_working_copy is non-zero, also
769  * check the working directory.  If empty_ok is non-zero, also return
770  * 0 in the case where the working-tree dir exists but is empty.
771  */
772 static int dir_in_way(const char *path, int check_working_copy, int empty_ok)
773 {
774         int pos;
775         struct strbuf dirpath = STRBUF_INIT;
776         struct stat st;
777
778         strbuf_addstr(&dirpath, path);
779         strbuf_addch(&dirpath, '/');
780
781         pos = cache_name_pos(dirpath.buf, dirpath.len);
782
783         if (pos < 0)
784                 pos = -1 - pos;
785         if (pos < active_nr &&
786             !strncmp(dirpath.buf, active_cache[pos]->name, dirpath.len)) {
787                 strbuf_release(&dirpath);
788                 return 1;
789         }
790
791         strbuf_release(&dirpath);
792         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
793                 !(empty_ok && is_empty_dir(path));
794 }
795
796 /*
797  * Returns whether path was tracked in the index before the merge started,
798  * and its oid and mode match the specified values
799  */
800 static int was_tracked_and_matches(struct merge_options *o, const char *path,
801                                    const struct object_id *oid, unsigned mode)
802 {
803         int pos = index_name_pos(&o->orig_index, path, strlen(path));
804         struct cache_entry *ce;
805
806         if (0 > pos)
807                 /* we were not tracking this path before the merge */
808                 return 0;
809
810         /* See if the file we were tracking before matches */
811         ce = o->orig_index.cache[pos];
812         return (oid_eq(&ce->oid, oid) && ce->ce_mode == mode);
813 }
814
815 /*
816  * Returns whether path was tracked in the index before the merge started
817  */
818 static int was_tracked(struct merge_options *o, const char *path)
819 {
820         int pos = index_name_pos(&o->orig_index, path, strlen(path));
821
822         if (0 <= pos)
823                 /* we were tracking this path before the merge */
824                 return 1;
825
826         return 0;
827 }
828
829 static int would_lose_untracked(const char *path)
830 {
831         /*
832          * This may look like it can be simplified to:
833          *   return !was_tracked(o, path) && file_exists(path)
834          * but it can't.  This function needs to know whether path was in
835          * the working tree due to EITHER having been tracked in the index
836          * before the merge OR having been put into the working copy and
837          * index by unpack_trees().  Due to that either-or requirement, we
838          * check the current index instead of the original one.
839          *
840          * Note that we do not need to worry about merge-recursive itself
841          * updating the index after unpack_trees() and before calling this
842          * function, because we strictly require all code paths in
843          * merge-recursive to update the working tree first and the index
844          * second.  Doing otherwise would break
845          * update_file()/would_lose_untracked(); see every comment in this
846          * file which mentions "update_stages".
847          */
848         int pos = cache_name_pos(path, strlen(path));
849
850         if (pos < 0)
851                 pos = -1 - pos;
852         while (pos < active_nr &&
853                !strcmp(path, active_cache[pos]->name)) {
854                 /*
855                  * If stage #0, it is definitely tracked.
856                  * If it has stage #2 then it was tracked
857                  * before this merge started.  All other
858                  * cases the path was not tracked.
859                  */
860                 switch (ce_stage(active_cache[pos])) {
861                 case 0:
862                 case 2:
863                         return 0;
864                 }
865                 pos++;
866         }
867         return file_exists(path);
868 }
869
870 static int was_dirty(struct merge_options *o, const char *path)
871 {
872         struct cache_entry *ce;
873         int dirty = 1;
874
875         if (o->call_depth || !was_tracked(o, path))
876                 return !dirty;
877
878         ce = index_file_exists(o->unpack_opts.src_index,
879                                path, strlen(path), ignore_case);
880         dirty = verify_uptodate(ce, &o->unpack_opts) != 0;
881         return dirty;
882 }
883
884 static int make_room_for_path(struct merge_options *o, const char *path)
885 {
886         int status, i;
887         const char *msg = _("failed to create path '%s'%s");
888
889         /* Unlink any D/F conflict files that are in the way */
890         for (i = 0; i < o->df_conflict_file_set.nr; i++) {
891                 const char *df_path = o->df_conflict_file_set.items[i].string;
892                 size_t pathlen = strlen(path);
893                 size_t df_pathlen = strlen(df_path);
894                 if (df_pathlen < pathlen &&
895                     path[df_pathlen] == '/' &&
896                     strncmp(path, df_path, df_pathlen) == 0) {
897                         output(o, 3,
898                                _("Removing %s to make room for subdirectory\n"),
899                                df_path);
900                         unlink(df_path);
901                         unsorted_string_list_delete_item(&o->df_conflict_file_set,
902                                                          i, 0);
903                         break;
904                 }
905         }
906
907         /* Make sure leading directories are created */
908         status = safe_create_leading_directories_const(path);
909         if (status) {
910                 if (status == SCLD_EXISTS)
911                         /* something else exists */
912                         return err(o, msg, path, _(": perhaps a D/F conflict?"));
913                 return err(o, msg, path, "");
914         }
915
916         /*
917          * Do not unlink a file in the work tree if we are not
918          * tracking it.
919          */
920         if (would_lose_untracked(path))
921                 return err(o, _("refusing to lose untracked file at '%s'"),
922                            path);
923
924         /* Successful unlink is good.. */
925         if (!unlink(path))
926                 return 0;
927         /* .. and so is no existing file */
928         if (errno == ENOENT)
929                 return 0;
930         /* .. but not some other error (who really cares what?) */
931         return err(o, msg, path, _(": perhaps a D/F conflict?"));
932 }
933
934 static int update_file_flags(struct merge_options *o,
935                              const struct object_id *oid,
936                              unsigned mode,
937                              const char *path,
938                              int update_cache,
939                              int update_wd)
940 {
941         int ret = 0;
942
943         if (o->call_depth)
944                 update_wd = 0;
945
946         if (update_wd) {
947                 enum object_type type;
948                 void *buf;
949                 unsigned long size;
950
951                 if (S_ISGITLINK(mode)) {
952                         /*
953                          * We may later decide to recursively descend into
954                          * the submodule directory and update its index
955                          * and/or work tree, but we do not do that now.
956                          */
957                         update_wd = 0;
958                         goto update_index;
959                 }
960
961                 buf = read_object_file(oid, &type, &size);
962                 if (!buf)
963                         return err(o, _("cannot read object %s '%s'"), oid_to_hex(oid), path);
964                 if (type != OBJ_BLOB) {
965                         ret = err(o, _("blob expected for %s '%s'"), oid_to_hex(oid), path);
966                         goto free_buf;
967                 }
968                 if (S_ISREG(mode)) {
969                         struct strbuf strbuf = STRBUF_INIT;
970                         if (convert_to_working_tree(&the_index, path, buf, size, &strbuf)) {
971                                 free(buf);
972                                 size = strbuf.len;
973                                 buf = strbuf_detach(&strbuf, NULL);
974                         }
975                 }
976
977                 if (make_room_for_path(o, path) < 0) {
978                         update_wd = 0;
979                         goto free_buf;
980                 }
981                 if (S_ISREG(mode) || (!has_symlinks && S_ISLNK(mode))) {
982                         int fd;
983                         if (mode & 0100)
984                                 mode = 0777;
985                         else
986                                 mode = 0666;
987                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
988                         if (fd < 0) {
989                                 ret = err(o, _("failed to open '%s': %s"),
990                                           path, strerror(errno));
991                                 goto free_buf;
992                         }
993                         write_in_full(fd, buf, size);
994                         close(fd);
995                 } else if (S_ISLNK(mode)) {
996                         char *lnk = xmemdupz(buf, size);
997                         safe_create_leading_directories_const(path);
998                         unlink(path);
999                         if (symlink(lnk, path))
1000                                 ret = err(o, _("failed to symlink '%s': %s"),
1001                                           path, strerror(errno));
1002                         free(lnk);
1003                 } else
1004                         ret = err(o,
1005                                   _("do not know what to do with %06o %s '%s'"),
1006                                   mode, oid_to_hex(oid), path);
1007         free_buf:
1008                 free(buf);
1009         }
1010 update_index:
1011         if (!ret && update_cache)
1012                 if (add_cacheinfo(o, mode, oid, path, 0, update_wd,
1013                                   ADD_CACHE_OK_TO_ADD))
1014                         return -1;
1015         return ret;
1016 }
1017
1018 static int update_file(struct merge_options *o,
1019                        int clean,
1020                        const struct object_id *oid,
1021                        unsigned mode,
1022                        const char *path)
1023 {
1024         return update_file_flags(o, oid, mode, path, o->call_depth || clean, !o->call_depth);
1025 }
1026
1027 /* Low level file merging, update and removal */
1028
1029 struct merge_file_info {
1030         struct object_id oid;
1031         unsigned mode;
1032         unsigned clean:1,
1033                  merge:1;
1034 };
1035
1036 static int merge_3way(struct merge_options *o,
1037                       mmbuffer_t *result_buf,
1038                       const struct diff_filespec *one,
1039                       const struct diff_filespec *a,
1040                       const struct diff_filespec *b,
1041                       const char *branch1,
1042                       const char *branch2)
1043 {
1044         mmfile_t orig, src1, src2;
1045         struct ll_merge_options ll_opts = {0};
1046         char *base_name, *name1, *name2;
1047         int merge_status;
1048
1049         ll_opts.renormalize = o->renormalize;
1050         ll_opts.xdl_opts = o->xdl_opts;
1051
1052         if (o->call_depth) {
1053                 ll_opts.virtual_ancestor = 1;
1054                 ll_opts.variant = 0;
1055         } else {
1056                 switch (o->recursive_variant) {
1057                 case MERGE_RECURSIVE_OURS:
1058                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
1059                         break;
1060                 case MERGE_RECURSIVE_THEIRS:
1061                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
1062                         break;
1063                 default:
1064                         ll_opts.variant = 0;
1065                         break;
1066                 }
1067         }
1068
1069         if (strcmp(a->path, b->path) ||
1070             (o->ancestor != NULL && strcmp(a->path, one->path) != 0)) {
1071                 base_name = o->ancestor == NULL ? NULL :
1072                         mkpathdup("%s:%s", o->ancestor, one->path);
1073                 name1 = mkpathdup("%s:%s", branch1, a->path);
1074                 name2 = mkpathdup("%s:%s", branch2, b->path);
1075         } else {
1076                 base_name = o->ancestor == NULL ? NULL :
1077                         mkpathdup("%s", o->ancestor);
1078                 name1 = mkpathdup("%s", branch1);
1079                 name2 = mkpathdup("%s", branch2);
1080         }
1081
1082         read_mmblob(&orig, &one->oid);
1083         read_mmblob(&src1, &a->oid);
1084         read_mmblob(&src2, &b->oid);
1085
1086         merge_status = ll_merge(result_buf, a->path, &orig, base_name,
1087                                 &src1, name1, &src2, name2, &ll_opts);
1088
1089         free(base_name);
1090         free(name1);
1091         free(name2);
1092         free(orig.ptr);
1093         free(src1.ptr);
1094         free(src2.ptr);
1095         return merge_status;
1096 }
1097
1098 static int find_first_merges(struct object_array *result, const char *path,
1099                              struct commit *a, struct commit *b)
1100 {
1101         int i, j;
1102         struct object_array merges = OBJECT_ARRAY_INIT;
1103         struct commit *commit;
1104         int contains_another;
1105
1106         char merged_revision[42];
1107         const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
1108                                    "--all", merged_revision, NULL };
1109         struct rev_info revs;
1110         struct setup_revision_opt rev_opts;
1111
1112         memset(result, 0, sizeof(struct object_array));
1113         memset(&rev_opts, 0, sizeof(rev_opts));
1114
1115         /* get all revisions that merge commit a */
1116         xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
1117                   oid_to_hex(&a->object.oid));
1118         init_revisions(&revs, NULL);
1119         rev_opts.submodule = path;
1120         /* FIXME: can't handle linked worktrees in submodules yet */
1121         revs.single_worktree = path != NULL;
1122         setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
1123
1124         /* save all revisions from the above list that contain b */
1125         if (prepare_revision_walk(&revs))
1126                 die("revision walk setup failed");
1127         while ((commit = get_revision(&revs)) != NULL) {
1128                 struct object *o = &(commit->object);
1129                 if (in_merge_bases(b, commit))
1130                         add_object_array(o, NULL, &merges);
1131         }
1132         reset_revision_walk();
1133
1134         /* Now we've got all merges that contain a and b. Prune all
1135          * merges that contain another found merge and save them in
1136          * result.
1137          */
1138         for (i = 0; i < merges.nr; i++) {
1139                 struct commit *m1 = (struct commit *) merges.objects[i].item;
1140
1141                 contains_another = 0;
1142                 for (j = 0; j < merges.nr; j++) {
1143                         struct commit *m2 = (struct commit *) merges.objects[j].item;
1144                         if (i != j && in_merge_bases(m2, m1)) {
1145                                 contains_another = 1;
1146                                 break;
1147                         }
1148                 }
1149
1150                 if (!contains_another)
1151                         add_object_array(merges.objects[i].item, NULL, result);
1152         }
1153
1154         object_array_clear(&merges);
1155         return result->nr;
1156 }
1157
1158 static void print_commit(struct commit *commit)
1159 {
1160         struct strbuf sb = STRBUF_INIT;
1161         struct pretty_print_context ctx = {0};
1162         ctx.date_mode.type = DATE_NORMAL;
1163         format_commit_message(commit, " %h: %m %s", &sb, &ctx);
1164         fprintf(stderr, "%s\n", sb.buf);
1165         strbuf_release(&sb);
1166 }
1167
1168 static int merge_submodule(struct merge_options *o,
1169                            struct object_id *result, const char *path,
1170                            const struct object_id *base, const struct object_id *a,
1171                            const struct object_id *b)
1172 {
1173         struct commit *commit_base, *commit_a, *commit_b;
1174         int parent_count;
1175         struct object_array merges;
1176
1177         int i;
1178         int search = !o->call_depth;
1179
1180         /* store a in result in case we fail */
1181         oidcpy(result, a);
1182
1183         /* we can not handle deletion conflicts */
1184         if (is_null_oid(base))
1185                 return 0;
1186         if (is_null_oid(a))
1187                 return 0;
1188         if (is_null_oid(b))
1189                 return 0;
1190
1191         if (add_submodule_odb(path)) {
1192                 output(o, 1, _("Failed to merge submodule %s (not checked out)"), path);
1193                 return 0;
1194         }
1195
1196         if (!(commit_base = lookup_commit_reference(the_repository, base)) ||
1197             !(commit_a = lookup_commit_reference(the_repository, a)) ||
1198             !(commit_b = lookup_commit_reference(the_repository, b))) {
1199                 output(o, 1, _("Failed to merge submodule %s (commits not present)"), path);
1200                 return 0;
1201         }
1202
1203         /* check whether both changes are forward */
1204         if (!in_merge_bases(commit_base, commit_a) ||
1205             !in_merge_bases(commit_base, commit_b)) {
1206                 output(o, 1, _("Failed to merge submodule %s (commits don't follow merge-base)"), path);
1207                 return 0;
1208         }
1209
1210         /* Case #1: a is contained in b or vice versa */
1211         if (in_merge_bases(commit_a, commit_b)) {
1212                 oidcpy(result, b);
1213                 if (show(o, 3)) {
1214                         output(o, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1215                         output_commit_title(o, commit_b);
1216                 } else if (show(o, 2))
1217                         output(o, 2, _("Fast-forwarding submodule %s"), path);
1218                 else
1219                         ; /* no output */
1220
1221                 return 1;
1222         }
1223         if (in_merge_bases(commit_b, commit_a)) {
1224                 oidcpy(result, a);
1225                 if (show(o, 3)) {
1226                         output(o, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1227                         output_commit_title(o, commit_a);
1228                 } else if (show(o, 2))
1229                         output(o, 2, _("Fast-forwarding submodule %s"), path);
1230                 else
1231                         ; /* no output */
1232
1233                 return 1;
1234         }
1235
1236         /*
1237          * Case #2: There are one or more merges that contain a and b in
1238          * the submodule. If there is only one, then present it as a
1239          * suggestion to the user, but leave it marked unmerged so the
1240          * user needs to confirm the resolution.
1241          */
1242
1243         /* Skip the search if makes no sense to the calling context.  */
1244         if (!search)
1245                 return 0;
1246
1247         /* find commit which merges them */
1248         parent_count = find_first_merges(&merges, path, commit_a, commit_b);
1249         switch (parent_count) {
1250         case 0:
1251                 output(o, 1, _("Failed to merge submodule %s (merge following commits not found)"), path);
1252                 break;
1253
1254         case 1:
1255                 output(o, 1, _("Failed to merge submodule %s (not fast-forward)"), path);
1256                 output(o, 2, _("Found a possible merge resolution for the submodule:\n"));
1257                 print_commit((struct commit *) merges.objects[0].item);
1258                 output(o, 2, _(
1259                        "If this is correct simply add it to the index "
1260                        "for example\n"
1261                        "by using:\n\n"
1262                        "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
1263                        "which will accept this suggestion.\n"),
1264                        oid_to_hex(&merges.objects[0].item->oid), path);
1265                 break;
1266
1267         default:
1268                 output(o, 1, _("Failed to merge submodule %s (multiple merges found)"), path);
1269                 for (i = 0; i < merges.nr; i++)
1270                         print_commit((struct commit *) merges.objects[i].item);
1271         }
1272
1273         object_array_clear(&merges);
1274         return 0;
1275 }
1276
1277 static int merge_file_1(struct merge_options *o,
1278                         const struct diff_filespec *one,
1279                         const struct diff_filespec *a,
1280                         const struct diff_filespec *b,
1281                         const char *filename,
1282                         const char *branch1,
1283                         const char *branch2,
1284                         struct merge_file_info *result)
1285 {
1286         result->merge = 0;
1287         result->clean = 1;
1288
1289         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
1290                 result->clean = 0;
1291                 if (S_ISREG(a->mode)) {
1292                         result->mode = a->mode;
1293                         oidcpy(&result->oid, &a->oid);
1294                 } else {
1295                         result->mode = b->mode;
1296                         oidcpy(&result->oid, &b->oid);
1297                 }
1298         } else {
1299                 if (!oid_eq(&a->oid, &one->oid) && !oid_eq(&b->oid, &one->oid))
1300                         result->merge = 1;
1301
1302                 /*
1303                  * Merge modes
1304                  */
1305                 if (a->mode == b->mode || a->mode == one->mode)
1306                         result->mode = b->mode;
1307                 else {
1308                         result->mode = a->mode;
1309                         if (b->mode != one->mode) {
1310                                 result->clean = 0;
1311                                 result->merge = 1;
1312                         }
1313                 }
1314
1315                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &one->oid))
1316                         oidcpy(&result->oid, &b->oid);
1317                 else if (oid_eq(&b->oid, &one->oid))
1318                         oidcpy(&result->oid, &a->oid);
1319                 else if (S_ISREG(a->mode)) {
1320                         mmbuffer_t result_buf;
1321                         int ret = 0, merge_status;
1322
1323                         merge_status = merge_3way(o, &result_buf, one, a, b,
1324                                                   branch1, branch2);
1325
1326                         if ((merge_status < 0) || !result_buf.ptr)
1327                                 ret = err(o, _("Failed to execute internal merge"));
1328
1329                         if (!ret &&
1330                             write_object_file(result_buf.ptr, result_buf.size,
1331                                               blob_type, &result->oid))
1332                                 ret = err(o, _("Unable to add %s to database"),
1333                                           a->path);
1334
1335                         free(result_buf.ptr);
1336                         if (ret)
1337                                 return ret;
1338                         result->clean = (merge_status == 0);
1339                 } else if (S_ISGITLINK(a->mode)) {
1340                         result->clean = merge_submodule(o, &result->oid,
1341                                                         one->path,
1342                                                         &one->oid,
1343                                                         &a->oid,
1344                                                         &b->oid);
1345                 } else if (S_ISLNK(a->mode)) {
1346                         switch (o->recursive_variant) {
1347                         case MERGE_RECURSIVE_NORMAL:
1348                                 oidcpy(&result->oid, &a->oid);
1349                                 if (!oid_eq(&a->oid, &b->oid))
1350                                         result->clean = 0;
1351                                 break;
1352                         case MERGE_RECURSIVE_OURS:
1353                                 oidcpy(&result->oid, &a->oid);
1354                                 break;
1355                         case MERGE_RECURSIVE_THEIRS:
1356                                 oidcpy(&result->oid, &b->oid);
1357                                 break;
1358                         }
1359                 } else
1360                         BUG("unsupported object type in the tree");
1361         }
1362
1363         if (result->merge)
1364                 output(o, 2, _("Auto-merging %s"), filename);
1365
1366         return 0;
1367 }
1368
1369 static int merge_file_special_markers(struct merge_options *o,
1370                                       const struct diff_filespec *one,
1371                                       const struct diff_filespec *a,
1372                                       const struct diff_filespec *b,
1373                                       const char *target_filename,
1374                                       const char *branch1,
1375                                       const char *filename1,
1376                                       const char *branch2,
1377                                       const char *filename2,
1378                                       struct merge_file_info *mfi)
1379 {
1380         char *side1 = NULL;
1381         char *side2 = NULL;
1382         int ret;
1383
1384         if (filename1)
1385                 side1 = xstrfmt("%s:%s", branch1, filename1);
1386         if (filename2)
1387                 side2 = xstrfmt("%s:%s", branch2, filename2);
1388
1389         ret = merge_file_1(o, one, a, b, target_filename,
1390                            side1 ? side1 : branch1,
1391                            side2 ? side2 : branch2, mfi);
1392
1393         free(side1);
1394         free(side2);
1395         return ret;
1396 }
1397
1398 static int merge_file_one(struct merge_options *o,
1399                           const char *path,
1400                           const struct object_id *o_oid, int o_mode,
1401                           const struct object_id *a_oid, int a_mode,
1402                           const struct object_id *b_oid, int b_mode,
1403                           const char *branch1,
1404                           const char *branch2,
1405                           struct merge_file_info *mfi)
1406 {
1407         struct diff_filespec one, a, b;
1408
1409         one.path = a.path = b.path = (char *)path;
1410         oidcpy(&one.oid, o_oid);
1411         one.mode = o_mode;
1412         oidcpy(&a.oid, a_oid);
1413         a.mode = a_mode;
1414         oidcpy(&b.oid, b_oid);
1415         b.mode = b_mode;
1416         return merge_file_1(o, &one, &a, &b, path, branch1, branch2, mfi);
1417 }
1418
1419 static int handle_rename_via_dir(struct merge_options *o,
1420                                  struct diff_filepair *pair,
1421                                  const char *rename_branch,
1422                                  const char *other_branch)
1423 {
1424         /*
1425          * Handle file adds that need to be renamed due to directory rename
1426          * detection.  This differs from handle_rename_normal, because
1427          * there is no content merge to do; just move the file into the
1428          * desired final location.
1429          */
1430         const struct diff_filespec *dest = pair->two;
1431
1432         if (!o->call_depth && would_lose_untracked(dest->path)) {
1433                 char *alt_path = unique_path(o, dest->path, rename_branch);
1434
1435                 output(o, 1, _("Error: Refusing to lose untracked file at %s; "
1436                                "writing to %s instead."),
1437                        dest->path, alt_path);
1438                 /*
1439                  * Write the file in worktree at alt_path, but not in the
1440                  * index.  Instead, write to dest->path for the index but
1441                  * only at the higher appropriate stage.
1442                  */
1443                 if (update_file(o, 0, &dest->oid, dest->mode, alt_path))
1444                         return -1;
1445                 free(alt_path);
1446                 return update_stages(o, dest->path, NULL,
1447                                      rename_branch == o->branch1 ? dest : NULL,
1448                                      rename_branch == o->branch1 ? NULL : dest);
1449         }
1450
1451         /* Update dest->path both in index and in worktree */
1452         if (update_file(o, 1, &dest->oid, dest->mode, dest->path))
1453                 return -1;
1454         return 0;
1455 }
1456
1457 static int handle_change_delete(struct merge_options *o,
1458                                 const char *path, const char *old_path,
1459                                 const struct object_id *o_oid, int o_mode,
1460                                 const struct object_id *changed_oid,
1461                                 int changed_mode,
1462                                 const char *change_branch,
1463                                 const char *delete_branch,
1464                                 const char *change, const char *change_past)
1465 {
1466         char *alt_path = NULL;
1467         const char *update_path = path;
1468         int ret = 0;
1469
1470         if (dir_in_way(path, !o->call_depth, 0) ||
1471             (!o->call_depth && would_lose_untracked(path))) {
1472                 update_path = alt_path = unique_path(o, path, change_branch);
1473         }
1474
1475         if (o->call_depth) {
1476                 /*
1477                  * We cannot arbitrarily accept either a_sha or b_sha as
1478                  * correct; since there is no true "middle point" between
1479                  * them, simply reuse the base version for virtual merge base.
1480                  */
1481                 ret = remove_file_from_cache(path);
1482                 if (!ret)
1483                         ret = update_file(o, 0, o_oid, o_mode, update_path);
1484         } else {
1485                 /*
1486                  * Despite the four nearly duplicate messages and argument
1487                  * lists below and the ugliness of the nested if-statements,
1488                  * having complete messages makes the job easier for
1489                  * translators.
1490                  *
1491                  * The slight variance among the cases is due to the fact
1492                  * that:
1493                  *   1) directory/file conflicts (in effect if
1494                  *      !alt_path) could cause us to need to write the
1495                  *      file to a different path.
1496                  *   2) renames (in effect if !old_path) could mean that
1497                  *      there are two names for the path that the user
1498                  *      may know the file by.
1499                  */
1500                 if (!alt_path) {
1501                         if (!old_path) {
1502                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1503                                        "and %s in %s. Version %s of %s left in tree."),
1504                                        change, path, delete_branch, change_past,
1505                                        change_branch, change_branch, path);
1506                         } else {
1507                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1508                                        "and %s to %s in %s. Version %s of %s left in tree."),
1509                                        change, old_path, delete_branch, change_past, path,
1510                                        change_branch, change_branch, path);
1511                         }
1512                 } else {
1513                         if (!old_path) {
1514                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1515                                        "and %s in %s. Version %s of %s left in tree at %s."),
1516                                        change, path, delete_branch, change_past,
1517                                        change_branch, change_branch, path, alt_path);
1518                         } else {
1519                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1520                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1521                                        change, old_path, delete_branch, change_past, path,
1522                                        change_branch, change_branch, path, alt_path);
1523                         }
1524                 }
1525                 /*
1526                  * No need to call update_file() on path when change_branch ==
1527                  * o->branch1 && !alt_path, since that would needlessly touch
1528                  * path.  We could call update_file_flags() with update_cache=0
1529                  * and update_wd=0, but that's a no-op.
1530                  */
1531                 if (change_branch != o->branch1 || alt_path)
1532                         ret = update_file(o, 0, changed_oid, changed_mode, update_path);
1533         }
1534         free(alt_path);
1535
1536         return ret;
1537 }
1538
1539 static int handle_rename_delete(struct merge_options *o,
1540                                 struct diff_filepair *pair,
1541                                 const char *rename_branch,
1542                                 const char *delete_branch)
1543 {
1544         const struct diff_filespec *orig = pair->one;
1545         const struct diff_filespec *dest = pair->two;
1546
1547         if (handle_change_delete(o,
1548                                  o->call_depth ? orig->path : dest->path,
1549                                  o->call_depth ? NULL : orig->path,
1550                                  &orig->oid, orig->mode,
1551                                  &dest->oid, dest->mode,
1552                                  rename_branch, delete_branch,
1553                                  _("rename"), _("renamed")))
1554                 return -1;
1555
1556         if (o->call_depth)
1557                 return remove_file_from_cache(dest->path);
1558         else
1559                 return update_stages(o, dest->path, NULL,
1560                                      rename_branch == o->branch1 ? dest : NULL,
1561                                      rename_branch == o->branch1 ? NULL : dest);
1562 }
1563
1564 static struct diff_filespec *filespec_from_entry(struct diff_filespec *target,
1565                                                  struct stage_data *entry,
1566                                                  int stage)
1567 {
1568         struct object_id *oid = &entry->stages[stage].oid;
1569         unsigned mode = entry->stages[stage].mode;
1570         if (mode == 0 || is_null_oid(oid))
1571                 return NULL;
1572         oidcpy(&target->oid, oid);
1573         target->mode = mode;
1574         return target;
1575 }
1576
1577 static int handle_file(struct merge_options *o,
1578                         struct diff_filespec *rename,
1579                         int stage,
1580                         struct rename_conflict_info *ci)
1581 {
1582         char *dst_name = rename->path;
1583         struct stage_data *dst_entry;
1584         const char *cur_branch, *other_branch;
1585         struct diff_filespec other;
1586         struct diff_filespec *add;
1587         int ret;
1588
1589         if (stage == 2) {
1590                 dst_entry = ci->dst_entry1;
1591                 cur_branch = ci->branch1;
1592                 other_branch = ci->branch2;
1593         } else {
1594                 dst_entry = ci->dst_entry2;
1595                 cur_branch = ci->branch2;
1596                 other_branch = ci->branch1;
1597         }
1598
1599         add = filespec_from_entry(&other, dst_entry, stage ^ 1);
1600         if (add) {
1601                 int ren_src_was_dirty = was_dirty(o, rename->path);
1602                 char *add_name = unique_path(o, rename->path, other_branch);
1603                 if (update_file(o, 0, &add->oid, add->mode, add_name))
1604                         return -1;
1605
1606                 if (ren_src_was_dirty) {
1607                         output(o, 1, _("Refusing to lose dirty file at %s"),
1608                                rename->path);
1609                 }
1610                 /*
1611                  * Because the double negatives somehow keep confusing me...
1612                  *    1) update_wd iff !ren_src_was_dirty.
1613                  *    2) no_wd iff !update_wd
1614                  *    3) so, no_wd == !!ren_src_was_dirty == ren_src_was_dirty
1615                  */
1616                 remove_file(o, 0, rename->path, ren_src_was_dirty);
1617                 dst_name = unique_path(o, rename->path, cur_branch);
1618         } else {
1619                 if (dir_in_way(rename->path, !o->call_depth, 0)) {
1620                         dst_name = unique_path(o, rename->path, cur_branch);
1621                         output(o, 1, _("%s is a directory in %s adding as %s instead"),
1622                                rename->path, other_branch, dst_name);
1623                 } else if (!o->call_depth &&
1624                            would_lose_untracked(rename->path)) {
1625                         dst_name = unique_path(o, rename->path, cur_branch);
1626                         output(o, 1, _("Refusing to lose untracked file at %s; "
1627                                        "adding as %s instead"),
1628                                rename->path, dst_name);
1629                 }
1630         }
1631         if ((ret = update_file(o, 0, &rename->oid, rename->mode, dst_name)))
1632                 ; /* fall through, do allow dst_name to be released */
1633         else if (stage == 2)
1634                 ret = update_stages(o, rename->path, NULL, rename, add);
1635         else
1636                 ret = update_stages(o, rename->path, NULL, add, rename);
1637
1638         if (dst_name != rename->path)
1639                 free(dst_name);
1640
1641         return ret;
1642 }
1643
1644 static int handle_rename_rename_1to2(struct merge_options *o,
1645                                      struct rename_conflict_info *ci)
1646 {
1647         /* One file was renamed in both branches, but to different names. */
1648         struct diff_filespec *one = ci->pair1->one;
1649         struct diff_filespec *a = ci->pair1->two;
1650         struct diff_filespec *b = ci->pair2->two;
1651
1652         output(o, 1, _("CONFLICT (rename/rename): "
1653                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1654                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1655                one->path, a->path, ci->branch1,
1656                one->path, b->path, ci->branch2,
1657                o->call_depth ? _(" (left unresolved)") : "");
1658         if (o->call_depth) {
1659                 struct merge_file_info mfi;
1660                 struct diff_filespec other;
1661                 struct diff_filespec *add;
1662                 if (merge_file_one(o, one->path,
1663                                  &one->oid, one->mode,
1664                                  &a->oid, a->mode,
1665                                  &b->oid, b->mode,
1666                                  ci->branch1, ci->branch2, &mfi))
1667                         return -1;
1668
1669                 /*
1670                  * FIXME: For rename/add-source conflicts (if we could detect
1671                  * such), this is wrong.  We should instead find a unique
1672                  * pathname and then either rename the add-source file to that
1673                  * unique path, or use that unique path instead of src here.
1674                  */
1675                 if (update_file(o, 0, &mfi.oid, mfi.mode, one->path))
1676                         return -1;
1677
1678                 /*
1679                  * Above, we put the merged content at the merge-base's
1680                  * path.  Now we usually need to delete both a->path and
1681                  * b->path.  However, the rename on each side of the merge
1682                  * could also be involved in a rename/add conflict.  In
1683                  * such cases, we should keep the added file around,
1684                  * resolving the conflict at that path in its favor.
1685                  */
1686                 add = filespec_from_entry(&other, ci->dst_entry1, 2 ^ 1);
1687                 if (add) {
1688                         if (update_file(o, 0, &add->oid, add->mode, a->path))
1689                                 return -1;
1690                 }
1691                 else
1692                         remove_file_from_cache(a->path);
1693                 add = filespec_from_entry(&other, ci->dst_entry2, 3 ^ 1);
1694                 if (add) {
1695                         if (update_file(o, 0, &add->oid, add->mode, b->path))
1696                                 return -1;
1697                 }
1698                 else
1699                         remove_file_from_cache(b->path);
1700         } else if (handle_file(o, a, 2, ci) || handle_file(o, b, 3, ci))
1701                 return -1;
1702
1703         return 0;
1704 }
1705
1706 static int handle_rename_rename_2to1(struct merge_options *o,
1707                                      struct rename_conflict_info *ci)
1708 {
1709         /* Two files, a & b, were renamed to the same thing, c. */
1710         struct diff_filespec *a = ci->pair1->one;
1711         struct diff_filespec *b = ci->pair2->one;
1712         struct diff_filespec *c1 = ci->pair1->two;
1713         struct diff_filespec *c2 = ci->pair2->two;
1714         char *path = c1->path; /* == c2->path */
1715         char *path_side_1_desc;
1716         char *path_side_2_desc;
1717         struct merge_file_info mfi_c1;
1718         struct merge_file_info mfi_c2;
1719         int ret;
1720
1721         output(o, 1, _("CONFLICT (rename/rename): "
1722                "Rename %s->%s in %s. "
1723                "Rename %s->%s in %s"),
1724                a->path, c1->path, ci->branch1,
1725                b->path, c2->path, ci->branch2);
1726
1727         remove_file(o, 1, a->path, o->call_depth || would_lose_untracked(a->path));
1728         remove_file(o, 1, b->path, o->call_depth || would_lose_untracked(b->path));
1729
1730         path_side_1_desc = xstrfmt("%s (was %s)", path, a->path);
1731         path_side_2_desc = xstrfmt("%s (was %s)", path, b->path);
1732         if (merge_file_special_markers(o, a, c1, &ci->ren1_other,
1733                                        path_side_1_desc,
1734                                        o->branch1, c1->path,
1735                                        o->branch2, ci->ren1_other.path, &mfi_c1) ||
1736             merge_file_special_markers(o, b, &ci->ren2_other, c2,
1737                                        path_side_2_desc,
1738                                        o->branch1, ci->ren2_other.path,
1739                                        o->branch2, c2->path, &mfi_c2))
1740                 return -1;
1741         free(path_side_1_desc);
1742         free(path_side_2_desc);
1743
1744         if (o->call_depth) {
1745                 /*
1746                  * If mfi_c1.clean && mfi_c2.clean, then it might make
1747                  * sense to do a two-way merge of those results.  But, I
1748                  * think in all cases, it makes sense to have the virtual
1749                  * merge base just undo the renames; they can be detected
1750                  * again later for the non-recursive merge.
1751                  */
1752                 remove_file(o, 0, path, 0);
1753                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, a->path);
1754                 if (!ret)
1755                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1756                                           b->path);
1757         } else {
1758                 char *new_path1 = unique_path(o, path, ci->branch1);
1759                 char *new_path2 = unique_path(o, path, ci->branch2);
1760                 output(o, 1, _("Renaming %s to %s and %s to %s instead"),
1761                        a->path, new_path1, b->path, new_path2);
1762                 if (was_dirty(o, path))
1763                         output(o, 1, _("Refusing to lose dirty file at %s"),
1764                                path);
1765                 else if (would_lose_untracked(path))
1766                         /*
1767                          * Only way we get here is if both renames were from
1768                          * a directory rename AND user had an untracked file
1769                          * at the location where both files end up after the
1770                          * two directory renames.  See testcase 10d of t6043.
1771                          */
1772                         output(o, 1, _("Refusing to lose untracked file at "
1773                                        "%s, even though it's in the way."),
1774                                path);
1775                 else
1776                         remove_file(o, 0, path, 0);
1777                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, new_path1);
1778                 if (!ret)
1779                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1780                                           new_path2);
1781                 /*
1782                  * unpack_trees() actually populates the index for us for
1783                  * "normal" rename/rename(2to1) situtations so that the
1784                  * correct entries are at the higher stages, which would
1785                  * make the call below to update_stages_for_stage_data
1786                  * unnecessary.  However, if either of the renames came
1787                  * from a directory rename, then unpack_trees() will not
1788                  * have gotten the right data loaded into the index, so we
1789                  * need to do so now.  (While it'd be tempting to move this
1790                  * call to update_stages_for_stage_data() to
1791                  * apply_directory_rename_modifications(), that would break
1792                  * our intermediate calls to would_lose_untracked() since
1793                  * those rely on the current in-memory index.  See also the
1794                  * big "NOTE" in update_stages()).
1795                  */
1796                 if (update_stages_for_stage_data(o, path, ci->dst_entry1))
1797                         ret = -1;
1798
1799                 free(new_path2);
1800                 free(new_path1);
1801         }
1802
1803         return ret;
1804 }
1805
1806 /*
1807  * Get the diff_filepairs changed between o_tree and tree.
1808  */
1809 static struct diff_queue_struct *get_diffpairs(struct merge_options *o,
1810                                                struct tree *o_tree,
1811                                                struct tree *tree)
1812 {
1813         struct diff_queue_struct *ret;
1814         struct diff_options opts;
1815
1816         diff_setup(&opts);
1817         opts.flags.recursive = 1;
1818         opts.flags.rename_empty = 0;
1819         opts.detect_rename = merge_detect_rename(o);
1820         /*
1821          * We do not have logic to handle the detection of copies.  In
1822          * fact, it may not even make sense to add such logic: would we
1823          * really want a change to a base file to be propagated through
1824          * multiple other files by a merge?
1825          */
1826         if (opts.detect_rename > DIFF_DETECT_RENAME)
1827                 opts.detect_rename = DIFF_DETECT_RENAME;
1828         opts.rename_limit = o->merge_rename_limit >= 0 ? o->merge_rename_limit :
1829                             o->diff_rename_limit >= 0 ? o->diff_rename_limit :
1830                             1000;
1831         opts.rename_score = o->rename_score;
1832         opts.show_rename_progress = o->show_rename_progress;
1833         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1834         diff_setup_done(&opts);
1835         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1836         diffcore_std(&opts);
1837         if (opts.needed_rename_limit > o->needed_rename_limit)
1838                 o->needed_rename_limit = opts.needed_rename_limit;
1839
1840         ret = xmalloc(sizeof(*ret));
1841         *ret = diff_queued_diff;
1842
1843         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1844         diff_queued_diff.nr = 0;
1845         diff_queued_diff.queue = NULL;
1846         diff_flush(&opts);
1847         return ret;
1848 }
1849
1850 static int tree_has_path(struct tree *tree, const char *path)
1851 {
1852         struct object_id hashy;
1853         unsigned int mode_o;
1854
1855         return !get_tree_entry(&tree->object.oid, path,
1856                                &hashy, &mode_o);
1857 }
1858
1859 /*
1860  * Return a new string that replaces the beginning portion (which matches
1861  * entry->dir), with entry->new_dir.  In perl-speak:
1862  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1863  * NOTE:
1864  *   Caller must ensure that old_path starts with entry->dir + '/'.
1865  */
1866 static char *apply_dir_rename(struct dir_rename_entry *entry,
1867                               const char *old_path)
1868 {
1869         struct strbuf new_path = STRBUF_INIT;
1870         int oldlen, newlen;
1871
1872         if (entry->non_unique_new_dir)
1873                 return NULL;
1874
1875         oldlen = strlen(entry->dir);
1876         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1877         strbuf_grow(&new_path, newlen);
1878         strbuf_addbuf(&new_path, &entry->new_dir);
1879         strbuf_addstr(&new_path, &old_path[oldlen]);
1880
1881         return strbuf_detach(&new_path, NULL);
1882 }
1883
1884 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1885                                     char **old_dir, char **new_dir)
1886 {
1887         char *end_of_old, *end_of_new;
1888         int old_len, new_len;
1889
1890         *old_dir = NULL;
1891         *new_dir = NULL;
1892
1893         /*
1894          * For
1895          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1896          * the "e/foo.c" part is the same, we just want to know that
1897          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1898          * so, for this example, this function returns "a/b/c/d" in
1899          * *old_dir and "a/b/some/thing/else" in *new_dir.
1900          *
1901          * Also, if the basename of the file changed, we don't care.  We
1902          * want to know which portion of the directory, if any, changed.
1903          */
1904         end_of_old = strrchr(old_path, '/');
1905         end_of_new = strrchr(new_path, '/');
1906
1907         if (end_of_old == NULL || end_of_new == NULL)
1908                 return;
1909         while (*--end_of_new == *--end_of_old &&
1910                end_of_old != old_path &&
1911                end_of_new != new_path)
1912                 ; /* Do nothing; all in the while loop */
1913         /*
1914          * We've found the first non-matching character in the directory
1915          * paths.  That means the current directory we were comparing
1916          * represents the rename.  Move end_of_old and end_of_new back
1917          * to the full directory name.
1918          */
1919         if (*end_of_old == '/')
1920                 end_of_old++;
1921         if (*end_of_old != '/')
1922                 end_of_new++;
1923         end_of_old = strchr(end_of_old, '/');
1924         end_of_new = strchr(end_of_new, '/');
1925
1926         /*
1927          * It may have been the case that old_path and new_path were the same
1928          * directory all along.  Don't claim a rename if they're the same.
1929          */
1930         old_len = end_of_old - old_path;
1931         new_len = end_of_new - new_path;
1932
1933         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1934                 *old_dir = xstrndup(old_path, old_len);
1935                 *new_dir = xstrndup(new_path, new_len);
1936         }
1937 }
1938
1939 static void remove_hashmap_entries(struct hashmap *dir_renames,
1940                                    struct string_list *items_to_remove)
1941 {
1942         int i;
1943         struct dir_rename_entry *entry;
1944
1945         for (i = 0; i < items_to_remove->nr; i++) {
1946                 entry = items_to_remove->items[i].util;
1947                 hashmap_remove(dir_renames, entry, NULL);
1948         }
1949         string_list_clear(items_to_remove, 0);
1950 }
1951
1952 /*
1953  * See if there is a directory rename for path, and if there are any file
1954  * level conflicts for the renamed location.  If there is a rename and
1955  * there are no conflicts, return the new name.  Otherwise, return NULL.
1956  */
1957 static char *handle_path_level_conflicts(struct merge_options *o,
1958                                          const char *path,
1959                                          struct dir_rename_entry *entry,
1960                                          struct hashmap *collisions,
1961                                          struct tree *tree)
1962 {
1963         char *new_path = NULL;
1964         struct collision_entry *collision_ent;
1965         int clean = 1;
1966         struct strbuf collision_paths = STRBUF_INIT;
1967
1968         /*
1969          * entry has the mapping of old directory name to new directory name
1970          * that we want to apply to path.
1971          */
1972         new_path = apply_dir_rename(entry, path);
1973
1974         if (!new_path) {
1975                 /* This should only happen when entry->non_unique_new_dir set */
1976                 if (!entry->non_unique_new_dir)
1977                         BUG("entry->non_unqiue_dir not set and !new_path");
1978                 output(o, 1, _("CONFLICT (directory rename split): "
1979                                "Unclear where to place %s because directory "
1980                                "%s was renamed to multiple other directories, "
1981                                "with no destination getting a majority of the "
1982                                "files."),
1983                        path, entry->dir);
1984                 clean = 0;
1985                 return NULL;
1986         }
1987
1988         /*
1989          * The caller needs to have ensured that it has pre-populated
1990          * collisions with all paths that map to new_path.  Do a quick check
1991          * to ensure that's the case.
1992          */
1993         collision_ent = collision_find_entry(collisions, new_path);
1994         if (collision_ent == NULL)
1995                 BUG("collision_ent is NULL");
1996
1997         /*
1998          * Check for one-sided add/add/.../add conflicts, i.e.
1999          * where implicit renames from the other side doing
2000          * directory rename(s) can affect this side of history
2001          * to put multiple paths into the same location.  Warn
2002          * and bail on directory renames for such paths.
2003          */
2004         if (collision_ent->reported_already) {
2005                 clean = 0;
2006         } else if (tree_has_path(tree, new_path)) {
2007                 collision_ent->reported_already = 1;
2008                 strbuf_add_separated_string_list(&collision_paths, ", ",
2009                                                  &collision_ent->source_files);
2010                 output(o, 1, _("CONFLICT (implicit dir rename): Existing "
2011                                "file/dir at %s in the way of implicit "
2012                                "directory rename(s) putting the following "
2013                                "path(s) there: %s."),
2014                        new_path, collision_paths.buf);
2015                 clean = 0;
2016         } else if (collision_ent->source_files.nr > 1) {
2017                 collision_ent->reported_already = 1;
2018                 strbuf_add_separated_string_list(&collision_paths, ", ",
2019                                                  &collision_ent->source_files);
2020                 output(o, 1, _("CONFLICT (implicit dir rename): Cannot map "
2021                                "more than one path to %s; implicit directory "
2022                                "renames tried to put these paths there: %s"),
2023                        new_path, collision_paths.buf);
2024                 clean = 0;
2025         }
2026
2027         /* Free memory we no longer need */
2028         strbuf_release(&collision_paths);
2029         if (!clean && new_path) {
2030                 free(new_path);
2031                 return NULL;
2032         }
2033
2034         return new_path;
2035 }
2036
2037 /*
2038  * There are a couple things we want to do at the directory level:
2039  *   1. Check for both sides renaming to the same thing, in order to avoid
2040  *      implicit renaming of files that should be left in place.  (See
2041  *      testcase 6b in t6043 for details.)
2042  *   2. Prune directory renames if there are still files left in the
2043  *      the original directory.  These represent a partial directory rename,
2044  *      i.e. a rename where only some of the files within the directory
2045  *      were renamed elsewhere.  (Technically, this could be done earlier
2046  *      in get_directory_renames(), except that would prevent us from
2047  *      doing the previous check and thus failing testcase 6b.)
2048  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2049  *      In the future, we could potentially record this info as well and
2050  *      omit reporting rename/rename(1to2) conflicts for each path within
2051  *      the affected directories, thus cleaning up the merge output.
2052  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2053  *         directory level, because merging directories is fine.  If it
2054  *         causes conflicts for files within those merged directories, then
2055  *         that should be detected at the individual path level.
2056  */
2057 static void handle_directory_level_conflicts(struct merge_options *o,
2058                                              struct hashmap *dir_re_head,
2059                                              struct tree *head,
2060                                              struct hashmap *dir_re_merge,
2061                                              struct tree *merge)
2062 {
2063         struct hashmap_iter iter;
2064         struct dir_rename_entry *head_ent;
2065         struct dir_rename_entry *merge_ent;
2066
2067         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2068         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2069
2070         hashmap_iter_init(dir_re_head, &iter);
2071         while ((head_ent = hashmap_iter_next(&iter))) {
2072                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2073                 if (merge_ent &&
2074                     !head_ent->non_unique_new_dir &&
2075                     !merge_ent->non_unique_new_dir &&
2076                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2077                         /* 1. Renamed identically; remove it from both sides */
2078                         string_list_append(&remove_from_head,
2079                                            head_ent->dir)->util = head_ent;
2080                         strbuf_release(&head_ent->new_dir);
2081                         string_list_append(&remove_from_merge,
2082                                            merge_ent->dir)->util = merge_ent;
2083                         strbuf_release(&merge_ent->new_dir);
2084                 } else if (tree_has_path(head, head_ent->dir)) {
2085                         /* 2. This wasn't a directory rename after all */
2086                         string_list_append(&remove_from_head,
2087                                            head_ent->dir)->util = head_ent;
2088                         strbuf_release(&head_ent->new_dir);
2089                 }
2090         }
2091
2092         remove_hashmap_entries(dir_re_head, &remove_from_head);
2093         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2094
2095         hashmap_iter_init(dir_re_merge, &iter);
2096         while ((merge_ent = hashmap_iter_next(&iter))) {
2097                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2098                 if (tree_has_path(merge, merge_ent->dir)) {
2099                         /* 2. This wasn't a directory rename after all */
2100                         string_list_append(&remove_from_merge,
2101                                            merge_ent->dir)->util = merge_ent;
2102                 } else if (head_ent &&
2103                            !head_ent->non_unique_new_dir &&
2104                            !merge_ent->non_unique_new_dir) {
2105                         /* 3. rename/rename(1to2) */
2106                         /*
2107                          * We can assume it's not rename/rename(1to1) because
2108                          * that was case (1), already checked above.  So we
2109                          * know that head_ent->new_dir and merge_ent->new_dir
2110                          * are different strings.
2111                          */
2112                         output(o, 1, _("CONFLICT (rename/rename): "
2113                                        "Rename directory %s->%s in %s. "
2114                                        "Rename directory %s->%s in %s"),
2115                                head_ent->dir, head_ent->new_dir.buf, o->branch1,
2116                                head_ent->dir, merge_ent->new_dir.buf, o->branch2);
2117                         string_list_append(&remove_from_head,
2118                                            head_ent->dir)->util = head_ent;
2119                         strbuf_release(&head_ent->new_dir);
2120                         string_list_append(&remove_from_merge,
2121                                            merge_ent->dir)->util = merge_ent;
2122                         strbuf_release(&merge_ent->new_dir);
2123                 }
2124         }
2125
2126         remove_hashmap_entries(dir_re_head, &remove_from_head);
2127         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2128 }
2129
2130 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs,
2131                                              struct tree *tree)
2132 {
2133         struct hashmap *dir_renames;
2134         struct hashmap_iter iter;
2135         struct dir_rename_entry *entry;
2136         int i;
2137
2138         /*
2139          * Typically, we think of a directory rename as all files from a
2140          * certain directory being moved to a target directory.  However,
2141          * what if someone first moved two files from the original
2142          * directory in one commit, and then renamed the directory
2143          * somewhere else in a later commit?  At merge time, we just know
2144          * that files from the original directory went to two different
2145          * places, and that the bulk of them ended up in the same place.
2146          * We want each directory rename to represent where the bulk of the
2147          * files from that directory end up; this function exists to find
2148          * where the bulk of the files went.
2149          *
2150          * The first loop below simply iterates through the list of file
2151          * renames, finding out how often each directory rename pair
2152          * possibility occurs.
2153          */
2154         dir_renames = xmalloc(sizeof(*dir_renames));
2155         dir_rename_init(dir_renames);
2156         for (i = 0; i < pairs->nr; ++i) {
2157                 struct string_list_item *item;
2158                 int *count;
2159                 struct diff_filepair *pair = pairs->queue[i];
2160                 char *old_dir, *new_dir;
2161
2162                 /* File not part of directory rename if it wasn't renamed */
2163                 if (pair->status != 'R')
2164                         continue;
2165
2166                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2167                                         &old_dir,        &new_dir);
2168                 if (!old_dir)
2169                         /* Directory didn't change at all; ignore this one. */
2170                         continue;
2171
2172                 entry = dir_rename_find_entry(dir_renames, old_dir);
2173                 if (!entry) {
2174                         entry = xmalloc(sizeof(*entry));
2175                         dir_rename_entry_init(entry, old_dir);
2176                         hashmap_put(dir_renames, entry);
2177                 } else {
2178                         free(old_dir);
2179                 }
2180                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2181                 if (!item) {
2182                         item = string_list_insert(&entry->possible_new_dirs,
2183                                                   new_dir);
2184                         item->util = xcalloc(1, sizeof(int));
2185                 } else {
2186                         free(new_dir);
2187                 }
2188                 count = item->util;
2189                 *count += 1;
2190         }
2191
2192         /*
2193          * For each directory with files moved out of it, we find out which
2194          * target directory received the most files so we can declare it to
2195          * be the "winning" target location for the directory rename.  This
2196          * winner gets recorded in new_dir.  If there is no winner
2197          * (multiple target directories received the same number of files),
2198          * we set non_unique_new_dir.  Once we've determined the winner (or
2199          * that there is no winner), we no longer need possible_new_dirs.
2200          */
2201         hashmap_iter_init(dir_renames, &iter);
2202         while ((entry = hashmap_iter_next(&iter))) {
2203                 int max = 0;
2204                 int bad_max = 0;
2205                 char *best = NULL;
2206
2207                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2208                         int *count = entry->possible_new_dirs.items[i].util;
2209
2210                         if (*count == max)
2211                                 bad_max = max;
2212                         else if (*count > max) {
2213                                 max = *count;
2214                                 best = entry->possible_new_dirs.items[i].string;
2215                         }
2216                 }
2217                 if (bad_max == max)
2218                         entry->non_unique_new_dir = 1;
2219                 else {
2220                         assert(entry->new_dir.len == 0);
2221                         strbuf_addstr(&entry->new_dir, best);
2222                 }
2223                 /*
2224                  * The relevant directory sub-portion of the original full
2225                  * filepaths were xstrndup'ed before inserting into
2226                  * possible_new_dirs, and instead of manually iterating the
2227                  * list and free'ing each, just lie and tell
2228                  * possible_new_dirs that it did the strdup'ing so that it
2229                  * will free them for us.
2230                  */
2231                 entry->possible_new_dirs.strdup_strings = 1;
2232                 string_list_clear(&entry->possible_new_dirs, 1);
2233         }
2234
2235         return dir_renames;
2236 }
2237
2238 static struct dir_rename_entry *check_dir_renamed(const char *path,
2239                                                   struct hashmap *dir_renames)
2240 {
2241         char *temp = xstrdup(path);
2242         char *end;
2243         struct dir_rename_entry *entry = NULL;;
2244
2245         while ((end = strrchr(temp, '/'))) {
2246                 *end = '\0';
2247                 entry = dir_rename_find_entry(dir_renames, temp);
2248                 if (entry)
2249                         break;
2250         }
2251         free(temp);
2252         return entry;
2253 }
2254
2255 static void compute_collisions(struct hashmap *collisions,
2256                                struct hashmap *dir_renames,
2257                                struct diff_queue_struct *pairs)
2258 {
2259         int i;
2260
2261         /*
2262          * Multiple files can be mapped to the same path due to directory
2263          * renames done by the other side of history.  Since that other
2264          * side of history could have merged multiple directories into one,
2265          * if our side of history added the same file basename to each of
2266          * those directories, then all N of them would get implicitly
2267          * renamed by the directory rename detection into the same path,
2268          * and we'd get an add/add/.../add conflict, and all those adds
2269          * from *this* side of history.  This is not representable in the
2270          * index, and users aren't going to easily be able to make sense of
2271          * it.  So we need to provide a good warning about what's
2272          * happening, and fall back to no-directory-rename detection
2273          * behavior for those paths.
2274          *
2275          * See testcases 9e and all of section 5 from t6043 for examples.
2276          */
2277         collision_init(collisions);
2278
2279         for (i = 0; i < pairs->nr; ++i) {
2280                 struct dir_rename_entry *dir_rename_ent;
2281                 struct collision_entry *collision_ent;
2282                 char *new_path;
2283                 struct diff_filepair *pair = pairs->queue[i];
2284
2285                 if (pair->status != 'A' && pair->status != 'R')
2286                         continue;
2287                 dir_rename_ent = check_dir_renamed(pair->two->path,
2288                                                    dir_renames);
2289                 if (!dir_rename_ent)
2290                         continue;
2291
2292                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2293                 if (!new_path)
2294                         /*
2295                          * dir_rename_ent->non_unique_new_path is true, which
2296                          * means there is no directory rename for us to use,
2297                          * which means it won't cause us any additional
2298                          * collisions.
2299                          */
2300                         continue;
2301                 collision_ent = collision_find_entry(collisions, new_path);
2302                 if (!collision_ent) {
2303                         collision_ent = xcalloc(1,
2304                                                 sizeof(struct collision_entry));
2305                         hashmap_entry_init(collision_ent, strhash(new_path));
2306                         hashmap_put(collisions, collision_ent);
2307                         collision_ent->target_file = new_path;
2308                 } else {
2309                         free(new_path);
2310                 }
2311                 string_list_insert(&collision_ent->source_files,
2312                                    pair->two->path);
2313         }
2314 }
2315
2316 static char *check_for_directory_rename(struct merge_options *o,
2317                                         const char *path,
2318                                         struct tree *tree,
2319                                         struct hashmap *dir_renames,
2320                                         struct hashmap *dir_rename_exclusions,
2321                                         struct hashmap *collisions,
2322                                         int *clean_merge)
2323 {
2324         char *new_path = NULL;
2325         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2326         struct dir_rename_entry *oentry = NULL;
2327
2328         if (!entry)
2329                 return new_path;
2330
2331         /*
2332          * This next part is a little weird.  We do not want to do an
2333          * implicit rename into a directory we renamed on our side, because
2334          * that will result in a spurious rename/rename(1to2) conflict.  An
2335          * example:
2336          *   Base commit: dumbdir/afile, otherdir/bfile
2337          *   Side 1:      smrtdir/afile, otherdir/bfile
2338          *   Side 2:      dumbdir/afile, dumbdir/bfile
2339          * Here, while working on Side 1, we could notice that otherdir was
2340          * renamed/merged to dumbdir, and change the diff_filepair for
2341          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2342          * 2 will notice the rename from dumbdir to smrtdir, and do the
2343          * transitive rename to move it from dumbdir/bfile to
2344          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2345          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2346          * the file to end up in smrtdir.  And the way to achieve that is
2347          * to not let Side1 do the rename to dumbdir, since we know that is
2348          * the source of one of our directory renames.
2349          *
2350          * That's why oentry and dir_rename_exclusions is here.
2351          *
2352          * As it turns out, this also prevents N-way transient rename
2353          * confusion; See testcases 9c and 9d of t6043.
2354          */
2355         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2356         if (oentry) {
2357                 output(o, 1, _("WARNING: Avoiding applying %s -> %s rename "
2358                                "to %s, because %s itself was renamed."),
2359                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2360         } else {
2361                 new_path = handle_path_level_conflicts(o, path, entry,
2362                                                        collisions, tree);
2363                 *clean_merge &= (new_path != NULL);
2364         }
2365
2366         return new_path;
2367 }
2368
2369 static void apply_directory_rename_modifications(struct merge_options *o,
2370                                                  struct diff_filepair *pair,
2371                                                  char *new_path,
2372                                                  struct rename *re,
2373                                                  struct tree *tree,
2374                                                  struct tree *o_tree,
2375                                                  struct tree *a_tree,
2376                                                  struct tree *b_tree,
2377                                                  struct string_list *entries,
2378                                                  int *clean)
2379 {
2380         struct string_list_item *item;
2381         int stage = (tree == a_tree ? 2 : 3);
2382         int update_wd;
2383
2384         /*
2385          * In all cases where we can do directory rename detection,
2386          * unpack_trees() will have read pair->two->path into the
2387          * index and the working copy.  We need to remove it so that
2388          * we can instead place it at new_path.  It is guaranteed to
2389          * not be untracked (unpack_trees() would have errored out
2390          * saying the file would have been overwritten), but it might
2391          * be dirty, though.
2392          */
2393         update_wd = !was_dirty(o, pair->two->path);
2394         if (!update_wd)
2395                 output(o, 1, _("Refusing to lose dirty file at %s"),
2396                        pair->two->path);
2397         remove_file(o, 1, pair->two->path, !update_wd);
2398
2399         /* Find or create a new re->dst_entry */
2400         item = string_list_lookup(entries, new_path);
2401         if (item) {
2402                 /*
2403                  * Since we're renaming on this side of history, and it's
2404                  * due to a directory rename on the other side of history
2405                  * (which we only allow when the directory in question no
2406                  * longer exists on the other side of history), the
2407                  * original entry for re->dst_entry is no longer
2408                  * necessary...
2409                  */
2410                 re->dst_entry->processed = 1;
2411
2412                 /*
2413                  * ...because we'll be using this new one.
2414                  */
2415                 re->dst_entry = item->util;
2416         } else {
2417                 /*
2418                  * re->dst_entry is for the before-dir-rename path, and we
2419                  * need it to hold information for the after-dir-rename
2420                  * path.  Before creating a new entry, we need to mark the
2421                  * old one as unnecessary (...unless it is shared by
2422                  * src_entry, i.e. this didn't use to be a rename, in which
2423                  * case we can just allow the normal processing to happen
2424                  * for it).
2425                  */
2426                 if (pair->status == 'R')
2427                         re->dst_entry->processed = 1;
2428
2429                 re->dst_entry = insert_stage_data(new_path,
2430                                                   o_tree, a_tree, b_tree,
2431                                                   entries);
2432                 item = string_list_insert(entries, new_path);
2433                 item->util = re->dst_entry;
2434         }
2435
2436         /*
2437          * Update the stage_data with the information about the path we are
2438          * moving into place.  That slot will be empty and available for us
2439          * to write to because of the collision checks in
2440          * handle_path_level_conflicts().  In other words,
2441          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2442          * open for us to write to.
2443          *
2444          * It may be tempting to actually update the index at this point as
2445          * well, using update_stages_for_stage_data(), but as per the big
2446          * "NOTE" in update_stages(), doing so will modify the current
2447          * in-memory index which will break calls to would_lose_untracked()
2448          * that we need to make.  Instead, we need to just make sure that
2449          * the various handle_rename_*() functions update the index
2450          * explicitly rather than relying on unpack_trees() to have done it.
2451          */
2452         get_tree_entry(&tree->object.oid,
2453                        pair->two->path,
2454                        &re->dst_entry->stages[stage].oid,
2455                        &re->dst_entry->stages[stage].mode);
2456
2457         /* Update pair status */
2458         if (pair->status == 'A') {
2459                 /*
2460                  * Recording rename information for this add makes it look
2461                  * like a rename/delete conflict.  Make sure we can
2462                  * correctly handle this as an add that was moved to a new
2463                  * directory instead of reporting a rename/delete conflict.
2464                  */
2465                 re->add_turned_into_rename = 1;
2466         }
2467         /*
2468          * We don't actually look at pair->status again, but it seems
2469          * pedagogically correct to adjust it.
2470          */
2471         pair->status = 'R';
2472
2473         /*
2474          * Finally, record the new location.
2475          */
2476         pair->two->path = new_path;
2477 }
2478
2479 /*
2480  * Get information of all renames which occurred in 'pairs', making use of
2481  * any implicit directory renames inferred from the other side of history.
2482  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2483  * to be able to associate the correct cache entries with the rename
2484  * information; tree is always equal to either a_tree or b_tree.
2485  */
2486 static struct string_list *get_renames(struct merge_options *o,
2487                                        struct diff_queue_struct *pairs,
2488                                        struct hashmap *dir_renames,
2489                                        struct hashmap *dir_rename_exclusions,
2490                                        struct tree *tree,
2491                                        struct tree *o_tree,
2492                                        struct tree *a_tree,
2493                                        struct tree *b_tree,
2494                                        struct string_list *entries,
2495                                        int *clean_merge)
2496 {
2497         int i;
2498         struct hashmap collisions;
2499         struct hashmap_iter iter;
2500         struct collision_entry *e;
2501         struct string_list *renames;
2502
2503         compute_collisions(&collisions, dir_renames, pairs);
2504         renames = xcalloc(1, sizeof(struct string_list));
2505
2506         for (i = 0; i < pairs->nr; ++i) {
2507                 struct string_list_item *item;
2508                 struct rename *re;
2509                 struct diff_filepair *pair = pairs->queue[i];
2510                 char *new_path; /* non-NULL only with directory renames */
2511
2512                 if (pair->status != 'A' && pair->status != 'R') {
2513                         diff_free_filepair(pair);
2514                         continue;
2515                 }
2516                 new_path = check_for_directory_rename(o, pair->two->path, tree,
2517                                                       dir_renames,
2518                                                       dir_rename_exclusions,
2519                                                       &collisions,
2520                                                       clean_merge);
2521                 if (pair->status != 'R' && !new_path) {
2522                         diff_free_filepair(pair);
2523                         continue;
2524                 }
2525
2526                 re = xmalloc(sizeof(*re));
2527                 re->processed = 0;
2528                 re->add_turned_into_rename = 0;
2529                 re->pair = pair;
2530                 item = string_list_lookup(entries, re->pair->one->path);
2531                 if (!item)
2532                         re->src_entry = insert_stage_data(re->pair->one->path,
2533                                         o_tree, a_tree, b_tree, entries);
2534                 else
2535                         re->src_entry = item->util;
2536
2537                 item = string_list_lookup(entries, re->pair->two->path);
2538                 if (!item)
2539                         re->dst_entry = insert_stage_data(re->pair->two->path,
2540                                         o_tree, a_tree, b_tree, entries);
2541                 else
2542                         re->dst_entry = item->util;
2543                 item = string_list_insert(renames, pair->one->path);
2544                 item->util = re;
2545                 if (new_path)
2546                         apply_directory_rename_modifications(o, pair, new_path,
2547                                                              re, tree, o_tree,
2548                                                              a_tree, b_tree,
2549                                                              entries,
2550                                                              clean_merge);
2551         }
2552
2553         hashmap_iter_init(&collisions, &iter);
2554         while ((e = hashmap_iter_next(&iter))) {
2555                 free(e->target_file);
2556                 string_list_clear(&e->source_files, 0);
2557         }
2558         hashmap_free(&collisions, 1);
2559         return renames;
2560 }
2561
2562 static int process_renames(struct merge_options *o,
2563                            struct string_list *a_renames,
2564                            struct string_list *b_renames)
2565 {
2566         int clean_merge = 1, i, j;
2567         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2568         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2569         const struct rename *sre;
2570
2571         for (i = 0; i < a_renames->nr; i++) {
2572                 sre = a_renames->items[i].util;
2573                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2574                         = (void *)sre;
2575         }
2576         for (i = 0; i < b_renames->nr; i++) {
2577                 sre = b_renames->items[i].util;
2578                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2579                         = (void *)sre;
2580         }
2581
2582         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2583                 struct string_list *renames1, *renames2Dst;
2584                 struct rename *ren1 = NULL, *ren2 = NULL;
2585                 const char *branch1, *branch2;
2586                 const char *ren1_src, *ren1_dst;
2587                 struct string_list_item *lookup;
2588
2589                 if (i >= a_renames->nr) {
2590                         ren2 = b_renames->items[j++].util;
2591                 } else if (j >= b_renames->nr) {
2592                         ren1 = a_renames->items[i++].util;
2593                 } else {
2594                         int compare = strcmp(a_renames->items[i].string,
2595                                              b_renames->items[j].string);
2596                         if (compare <= 0)
2597                                 ren1 = a_renames->items[i++].util;
2598                         if (compare >= 0)
2599                                 ren2 = b_renames->items[j++].util;
2600                 }
2601
2602                 /* TODO: refactor, so that 1/2 are not needed */
2603                 if (ren1) {
2604                         renames1 = a_renames;
2605                         renames2Dst = &b_by_dst;
2606                         branch1 = o->branch1;
2607                         branch2 = o->branch2;
2608                 } else {
2609                         renames1 = b_renames;
2610                         renames2Dst = &a_by_dst;
2611                         branch1 = o->branch2;
2612                         branch2 = o->branch1;
2613                         SWAP(ren2, ren1);
2614                 }
2615
2616                 if (ren1->processed)
2617                         continue;
2618                 ren1->processed = 1;
2619                 ren1->dst_entry->processed = 1;
2620                 /* BUG: We should only mark src_entry as processed if we
2621                  * are not dealing with a rename + add-source case.
2622                  */
2623                 ren1->src_entry->processed = 1;
2624
2625                 ren1_src = ren1->pair->one->path;
2626                 ren1_dst = ren1->pair->two->path;
2627
2628                 if (ren2) {
2629                         /* One file renamed on both sides */
2630                         const char *ren2_src = ren2->pair->one->path;
2631                         const char *ren2_dst = ren2->pair->two->path;
2632                         enum rename_type rename_type;
2633                         if (strcmp(ren1_src, ren2_src) != 0)
2634                                 BUG("ren1_src != ren2_src");
2635                         ren2->dst_entry->processed = 1;
2636                         ren2->processed = 1;
2637                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2638                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2639                                 clean_merge = 0;
2640                         } else {
2641                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2642                                 /* BUG: We should only remove ren1_src in
2643                                  * the base stage (think of rename +
2644                                  * add-source cases).
2645                                  */
2646                                 remove_file(o, 1, ren1_src, 1);
2647                                 update_entry(ren1->dst_entry,
2648                                              ren1->pair->one,
2649                                              ren1->pair->two,
2650                                              ren2->pair->two);
2651                         }
2652                         setup_rename_conflict_info(rename_type,
2653                                                    ren1->pair,
2654                                                    ren2->pair,
2655                                                    branch1,
2656                                                    branch2,
2657                                                    ren1->dst_entry,
2658                                                    ren2->dst_entry,
2659                                                    o,
2660                                                    NULL,
2661                                                    NULL);
2662                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2663                         /* Two different files renamed to the same thing */
2664                         char *ren2_dst;
2665                         ren2 = lookup->util;
2666                         ren2_dst = ren2->pair->two->path;
2667                         if (strcmp(ren1_dst, ren2_dst) != 0)
2668                                 BUG("ren1_dst != ren2_dst");
2669
2670                         clean_merge = 0;
2671                         ren2->processed = 1;
2672                         /*
2673                          * BUG: We should only mark src_entry as processed
2674                          * if we are not dealing with a rename + add-source
2675                          * case.
2676                          */
2677                         ren2->src_entry->processed = 1;
2678
2679                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2680                                                    ren1->pair,
2681                                                    ren2->pair,
2682                                                    branch1,
2683                                                    branch2,
2684                                                    ren1->dst_entry,
2685                                                    ren2->dst_entry,
2686                                                    o,
2687                                                    ren1->src_entry,
2688                                                    ren2->src_entry);
2689
2690                 } else {
2691                         /* Renamed in 1, maybe changed in 2 */
2692                         /* we only use sha1 and mode of these */
2693                         struct diff_filespec src_other, dst_other;
2694                         int try_merge;
2695
2696                         /*
2697                          * unpack_trees loads entries from common-commit
2698                          * into stage 1, from head-commit into stage 2, and
2699                          * from merge-commit into stage 3.  We keep track
2700                          * of which side corresponds to the rename.
2701                          */
2702                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2703                         int other_stage =   a_renames == renames1 ? 3 : 2;
2704
2705                         /* BUG: We should only remove ren1_src in the base
2706                          * stage and in other_stage (think of rename +
2707                          * add-source case).
2708                          */
2709                         remove_file(o, 1, ren1_src,
2710                                     renamed_stage == 2 || !was_tracked(o, ren1_src));
2711
2712                         oidcpy(&src_other.oid,
2713                                &ren1->src_entry->stages[other_stage].oid);
2714                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2715                         oidcpy(&dst_other.oid,
2716                                &ren1->dst_entry->stages[other_stage].oid);
2717                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2718                         try_merge = 0;
2719
2720                         if (oid_eq(&src_other.oid, &null_oid) &&
2721                             ren1->add_turned_into_rename) {
2722                                 setup_rename_conflict_info(RENAME_VIA_DIR,
2723                                                            ren1->pair,
2724                                                            NULL,
2725                                                            branch1,
2726                                                            branch2,
2727                                                            ren1->dst_entry,
2728                                                            NULL,
2729                                                            o,
2730                                                            NULL,
2731                                                            NULL);
2732                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2733                                 setup_rename_conflict_info(RENAME_DELETE,
2734                                                            ren1->pair,
2735                                                            NULL,
2736                                                            branch1,
2737                                                            branch2,
2738                                                            ren1->dst_entry,
2739                                                            NULL,
2740                                                            o,
2741                                                            NULL,
2742                                                            NULL);
2743                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2744                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2745                                 /*
2746                                  * Added file on the other side identical to
2747                                  * the file being renamed: clean merge.
2748                                  * Also, there is no need to overwrite the
2749                                  * file already in the working copy, so call
2750                                  * update_file_flags() instead of
2751                                  * update_file().
2752                                  */
2753                                 if (update_file_flags(o,
2754                                                       &ren1->pair->two->oid,
2755                                                       ren1->pair->two->mode,
2756                                                       ren1_dst,
2757                                                       1, /* update_cache */
2758                                                       0  /* update_wd    */))
2759                                         clean_merge = -1;
2760                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2761                                 clean_merge = 0;
2762                                 try_merge = 1;
2763                                 output(o, 1, _("CONFLICT (rename/add): Rename %s->%s in %s. "
2764                                        "%s added in %s"),
2765                                        ren1_src, ren1_dst, branch1,
2766                                        ren1_dst, branch2);
2767                                 if (o->call_depth) {
2768                                         struct merge_file_info mfi;
2769                                         if (merge_file_one(o, ren1_dst, &null_oid, 0,
2770                                                            &ren1->pair->two->oid,
2771                                                            ren1->pair->two->mode,
2772                                                            &dst_other.oid,
2773                                                            dst_other.mode,
2774                                                            branch1, branch2, &mfi)) {
2775                                                 clean_merge = -1;
2776                                                 goto cleanup_and_return;
2777                                         }
2778                                         output(o, 1, _("Adding merged %s"), ren1_dst);
2779                                         if (update_file(o, 0, &mfi.oid,
2780                                                         mfi.mode, ren1_dst))
2781                                                 clean_merge = -1;
2782                                         try_merge = 0;
2783                                 } else {
2784                                         char *new_path = unique_path(o, ren1_dst, branch2);
2785                                         output(o, 1, _("Adding as %s instead"), new_path);
2786                                         if (update_file(o, 0, &dst_other.oid,
2787                                                         dst_other.mode, new_path))
2788                                                 clean_merge = -1;
2789                                         free(new_path);
2790                                 }
2791                         } else
2792                                 try_merge = 1;
2793
2794                         if (clean_merge < 0)
2795                                 goto cleanup_and_return;
2796                         if (try_merge) {
2797                                 struct diff_filespec *one, *a, *b;
2798                                 src_other.path = (char *)ren1_src;
2799
2800                                 one = ren1->pair->one;
2801                                 if (a_renames == renames1) {
2802                                         a = ren1->pair->two;
2803                                         b = &src_other;
2804                                 } else {
2805                                         b = ren1->pair->two;
2806                                         a = &src_other;
2807                                 }
2808                                 update_entry(ren1->dst_entry, one, a, b);
2809                                 setup_rename_conflict_info(RENAME_NORMAL,
2810                                                            ren1->pair,
2811                                                            NULL,
2812                                                            branch1,
2813                                                            NULL,
2814                                                            ren1->dst_entry,
2815                                                            NULL,
2816                                                            o,
2817                                                            NULL,
2818                                                            NULL);
2819                         }
2820                 }
2821         }
2822 cleanup_and_return:
2823         string_list_clear(&a_by_dst, 0);
2824         string_list_clear(&b_by_dst, 0);
2825
2826         return clean_merge;
2827 }
2828
2829 struct rename_info {
2830         struct string_list *head_renames;
2831         struct string_list *merge_renames;
2832 };
2833
2834 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2835                                    struct hashmap *dir_renames)
2836 {
2837         struct hashmap_iter iter;
2838         struct dir_rename_entry *e;
2839
2840         hashmap_iter_init(dir_renames, &iter);
2841         while ((e = hashmap_iter_next(&iter))) {
2842                 free(e->dir);
2843                 strbuf_release(&e->new_dir);
2844                 /* possible_new_dirs already cleared in get_directory_renames */
2845         }
2846         hashmap_free(dir_renames, 1);
2847         free(dir_renames);
2848
2849         free(pairs->queue);
2850         free(pairs);
2851 }
2852
2853 static int detect_and_process_renames(struct merge_options *o,
2854                                       struct tree *common,
2855                                       struct tree *head,
2856                                       struct tree *merge,
2857                                       struct string_list *entries,
2858                                       struct rename_info *ri)
2859 {
2860         struct diff_queue_struct *head_pairs, *merge_pairs;
2861         struct hashmap *dir_re_head, *dir_re_merge;
2862         int clean = 1;
2863
2864         ri->head_renames = NULL;
2865         ri->merge_renames = NULL;
2866
2867         if (!merge_detect_rename(o))
2868                 return 1;
2869
2870         head_pairs = get_diffpairs(o, common, head);
2871         merge_pairs = get_diffpairs(o, common, merge);
2872
2873         if (o->detect_directory_renames) {
2874                 dir_re_head = get_directory_renames(head_pairs, head);
2875                 dir_re_merge = get_directory_renames(merge_pairs, merge);
2876
2877                 handle_directory_level_conflicts(o,
2878                                                  dir_re_head, head,
2879                                                  dir_re_merge, merge);
2880         } else {
2881                 dir_re_head  = xmalloc(sizeof(*dir_re_head));
2882                 dir_re_merge = xmalloc(sizeof(*dir_re_merge));
2883                 dir_rename_init(dir_re_head);
2884                 dir_rename_init(dir_re_merge);
2885         }
2886
2887         ri->head_renames  = get_renames(o, head_pairs,
2888                                         dir_re_merge, dir_re_head, head,
2889                                         common, head, merge, entries,
2890                                         &clean);
2891         if (clean < 0)
2892                 goto cleanup;
2893         ri->merge_renames = get_renames(o, merge_pairs,
2894                                         dir_re_head, dir_re_merge, merge,
2895                                         common, head, merge, entries,
2896                                         &clean);
2897         if (clean < 0)
2898                 goto cleanup;
2899         clean &= process_renames(o, ri->head_renames, ri->merge_renames);
2900
2901 cleanup:
2902         /*
2903          * Some cleanup is deferred until cleanup_renames() because the
2904          * data structures are still needed and referenced in
2905          * process_entry().  But there are a few things we can free now.
2906          */
2907         initial_cleanup_rename(head_pairs, dir_re_head);
2908         initial_cleanup_rename(merge_pairs, dir_re_merge);
2909
2910         return clean;
2911 }
2912
2913 static void final_cleanup_rename(struct string_list *rename)
2914 {
2915         const struct rename *re;
2916         int i;
2917
2918         if (rename == NULL)
2919                 return;
2920
2921         for (i = 0; i < rename->nr; i++) {
2922                 re = rename->items[i].util;
2923                 diff_free_filepair(re->pair);
2924         }
2925         string_list_clear(rename, 1);
2926         free(rename);
2927 }
2928
2929 static void final_cleanup_renames(struct rename_info *re_info)
2930 {
2931         final_cleanup_rename(re_info->head_renames);
2932         final_cleanup_rename(re_info->merge_renames);
2933 }
2934
2935 static struct object_id *stage_oid(const struct object_id *oid, unsigned mode)
2936 {
2937         return (is_null_oid(oid) || mode == 0) ? NULL: (struct object_id *)oid;
2938 }
2939
2940 static int read_oid_strbuf(struct merge_options *o,
2941                            const struct object_id *oid,
2942                            struct strbuf *dst)
2943 {
2944         void *buf;
2945         enum object_type type;
2946         unsigned long size;
2947         buf = read_object_file(oid, &type, &size);
2948         if (!buf)
2949                 return err(o, _("cannot read object %s"), oid_to_hex(oid));
2950         if (type != OBJ_BLOB) {
2951                 free(buf);
2952                 return err(o, _("object %s is not a blob"), oid_to_hex(oid));
2953         }
2954         strbuf_attach(dst, buf, size, size + 1);
2955         return 0;
2956 }
2957
2958 static int blob_unchanged(struct merge_options *opt,
2959                           const struct object_id *o_oid,
2960                           unsigned o_mode,
2961                           const struct object_id *a_oid,
2962                           unsigned a_mode,
2963                           int renormalize, const char *path)
2964 {
2965         struct strbuf o = STRBUF_INIT;
2966         struct strbuf a = STRBUF_INIT;
2967         int ret = 0; /* assume changed for safety */
2968
2969         if (a_mode != o_mode)
2970                 return 0;
2971         if (oid_eq(o_oid, a_oid))
2972                 return 1;
2973         if (!renormalize)
2974                 return 0;
2975
2976         assert(o_oid && a_oid);
2977         if (read_oid_strbuf(opt, o_oid, &o) || read_oid_strbuf(opt, a_oid, &a))
2978                 goto error_return;
2979         /*
2980          * Note: binary | is used so that both renormalizations are
2981          * performed.  Comparison can be skipped if both files are
2982          * unchanged since their sha1s have already been compared.
2983          */
2984         if (renormalize_buffer(&the_index, path, o.buf, o.len, &o) |
2985             renormalize_buffer(&the_index, path, a.buf, a.len, &a))
2986                 ret = (o.len == a.len && !memcmp(o.buf, a.buf, o.len));
2987
2988 error_return:
2989         strbuf_release(&o);
2990         strbuf_release(&a);
2991         return ret;
2992 }
2993
2994 static int handle_modify_delete(struct merge_options *o,
2995                                 const char *path,
2996                                 struct object_id *o_oid, int o_mode,
2997                                 struct object_id *a_oid, int a_mode,
2998                                 struct object_id *b_oid, int b_mode)
2999 {
3000         const char *modify_branch, *delete_branch;
3001         struct object_id *changed_oid;
3002         int changed_mode;
3003
3004         if (a_oid) {
3005                 modify_branch = o->branch1;
3006                 delete_branch = o->branch2;
3007                 changed_oid = a_oid;
3008                 changed_mode = a_mode;
3009         } else {
3010                 modify_branch = o->branch2;
3011                 delete_branch = o->branch1;
3012                 changed_oid = b_oid;
3013                 changed_mode = b_mode;
3014         }
3015
3016         return handle_change_delete(o,
3017                                     path, NULL,
3018                                     o_oid, o_mode,
3019                                     changed_oid, changed_mode,
3020                                     modify_branch, delete_branch,
3021                                     _("modify"), _("modified"));
3022 }
3023
3024 static int merge_content(struct merge_options *o,
3025                          const char *path,
3026                          int is_dirty,
3027                          struct object_id *o_oid, int o_mode,
3028                          struct object_id *a_oid, int a_mode,
3029                          struct object_id *b_oid, int b_mode,
3030                          struct rename_conflict_info *rename_conflict_info)
3031 {
3032         const char *reason = _("content");
3033         const char *path1 = NULL, *path2 = NULL;
3034         struct merge_file_info mfi;
3035         struct diff_filespec one, a, b;
3036         unsigned df_conflict_remains = 0;
3037
3038         if (!o_oid) {
3039                 reason = _("add/add");
3040                 o_oid = (struct object_id *)&null_oid;
3041         }
3042         one.path = a.path = b.path = (char *)path;
3043         oidcpy(&one.oid, o_oid);
3044         one.mode = o_mode;
3045         oidcpy(&a.oid, a_oid);
3046         a.mode = a_mode;
3047         oidcpy(&b.oid, b_oid);
3048         b.mode = b_mode;
3049
3050         if (rename_conflict_info) {
3051                 struct diff_filepair *pair1 = rename_conflict_info->pair1;
3052
3053                 path1 = (o->branch1 == rename_conflict_info->branch1) ?
3054                         pair1->two->path : pair1->one->path;
3055                 /* If rename_conflict_info->pair2 != NULL, we are in
3056                  * RENAME_ONE_FILE_TO_ONE case.  Otherwise, we have a
3057                  * normal rename.
3058                  */
3059                 path2 = (rename_conflict_info->pair2 ||
3060                          o->branch2 == rename_conflict_info->branch1) ?
3061                         pair1->two->path : pair1->one->path;
3062
3063                 if (dir_in_way(path, !o->call_depth,
3064                                S_ISGITLINK(pair1->two->mode)))
3065                         df_conflict_remains = 1;
3066         }
3067         if (merge_file_special_markers(o, &one, &a, &b, path,
3068                                        o->branch1, path1,
3069                                        o->branch2, path2, &mfi))
3070                 return -1;
3071
3072         /*
3073          * We can skip updating the working tree file iff:
3074          *   a) The merge is clean
3075          *   b) The merge matches what was in HEAD (content, mode, pathname)
3076          *   c) The target path is usable (i.e. not involved in D/F conflict)
3077          */
3078         if (mfi.clean &&
3079             was_tracked_and_matches(o, path, &mfi.oid, mfi.mode) &&
3080             !df_conflict_remains) {
3081                 int pos;
3082                 struct cache_entry *ce;
3083
3084                 output(o, 3, _("Skipped %s (merged same as existing)"), path);
3085                 if (add_cacheinfo(o, mfi.mode, &mfi.oid, path,
3086                                   0, (!o->call_depth && !is_dirty), 0))
3087                         return -1;
3088                 /*
3089                  * However, add_cacheinfo() will delete the old cache entry
3090                  * and add a new one.  We need to copy over any skip_worktree
3091                  * flag to avoid making the file appear as if it were
3092                  * deleted by the user.
3093                  */
3094                 pos = index_name_pos(&o->orig_index, path, strlen(path));
3095                 ce = o->orig_index.cache[pos];
3096                 if (ce_skip_worktree(ce)) {
3097                         pos = index_name_pos(&the_index, path, strlen(path));
3098                         ce = the_index.cache[pos];
3099                         ce->ce_flags |= CE_SKIP_WORKTREE;
3100                 }
3101                 return mfi.clean;
3102         }
3103
3104         if (!mfi.clean) {
3105                 if (S_ISGITLINK(mfi.mode))
3106                         reason = _("submodule");
3107                 output(o, 1, _("CONFLICT (%s): Merge conflict in %s"),
3108                                 reason, path);
3109                 if (rename_conflict_info && !df_conflict_remains)
3110                         if (update_stages(o, path, &one, &a, &b))
3111                                 return -1;
3112         }
3113
3114         if (df_conflict_remains || is_dirty) {
3115                 char *new_path;
3116                 if (o->call_depth) {
3117                         remove_file_from_cache(path);
3118                 } else {
3119                         if (!mfi.clean) {
3120                                 if (update_stages(o, path, &one, &a, &b))
3121                                         return -1;
3122                         } else {
3123                                 int file_from_stage2 = was_tracked(o, path);
3124                                 struct diff_filespec merged;
3125                                 oidcpy(&merged.oid, &mfi.oid);
3126                                 merged.mode = mfi.mode;
3127
3128                                 if (update_stages(o, path, NULL,
3129                                                   file_from_stage2 ? &merged : NULL,
3130                                                   file_from_stage2 ? NULL : &merged))
3131                                         return -1;
3132                         }
3133
3134                 }
3135                 new_path = unique_path(o, path, rename_conflict_info->branch1);
3136                 if (is_dirty) {
3137                         output(o, 1, _("Refusing to lose dirty file at %s"),
3138                                path);
3139                 }
3140                 output(o, 1, _("Adding as %s instead"), new_path);
3141                 if (update_file(o, 0, &mfi.oid, mfi.mode, new_path)) {
3142                         free(new_path);
3143                         return -1;
3144                 }
3145                 free(new_path);
3146                 mfi.clean = 0;
3147         } else if (update_file(o, mfi.clean, &mfi.oid, mfi.mode, path))
3148                 return -1;
3149         return !is_dirty && mfi.clean;
3150 }
3151
3152 static int handle_rename_normal(struct merge_options *o,
3153                                 const char *path,
3154                                 struct object_id *o_oid, unsigned int o_mode,
3155                                 struct object_id *a_oid, unsigned int a_mode,
3156                                 struct object_id *b_oid, unsigned int b_mode,
3157                                 struct rename_conflict_info *ci)
3158 {
3159         /* Merge the content and write it out */
3160         return merge_content(o, path, was_dirty(o, path),
3161                              o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3162                              ci);
3163 }
3164
3165 /* Per entry merge function */
3166 static int process_entry(struct merge_options *o,
3167                          const char *path, struct stage_data *entry)
3168 {
3169         int clean_merge = 1;
3170         int normalize = o->renormalize;
3171         unsigned o_mode = entry->stages[1].mode;
3172         unsigned a_mode = entry->stages[2].mode;
3173         unsigned b_mode = entry->stages[3].mode;
3174         struct object_id *o_oid = stage_oid(&entry->stages[1].oid, o_mode);
3175         struct object_id *a_oid = stage_oid(&entry->stages[2].oid, a_mode);
3176         struct object_id *b_oid = stage_oid(&entry->stages[3].oid, b_mode);
3177
3178         entry->processed = 1;
3179         if (entry->rename_conflict_info) {
3180                 struct rename_conflict_info *conflict_info = entry->rename_conflict_info;
3181                 switch (conflict_info->rename_type) {
3182                 case RENAME_NORMAL:
3183                 case RENAME_ONE_FILE_TO_ONE:
3184                         clean_merge = handle_rename_normal(o,
3185                                                            path,
3186                                                            o_oid, o_mode,
3187                                                            a_oid, a_mode,
3188                                                            b_oid, b_mode,
3189                                                            conflict_info);
3190                         break;
3191                 case RENAME_VIA_DIR:
3192                         clean_merge = 1;
3193                         if (handle_rename_via_dir(o,
3194                                                   conflict_info->pair1,
3195                                                   conflict_info->branch1,
3196                                                   conflict_info->branch2))
3197                                 clean_merge = -1;
3198                         break;
3199                 case RENAME_DELETE:
3200                         clean_merge = 0;
3201                         if (handle_rename_delete(o,
3202                                                  conflict_info->pair1,
3203                                                  conflict_info->branch1,
3204                                                  conflict_info->branch2))
3205                                 clean_merge = -1;
3206                         break;
3207                 case RENAME_ONE_FILE_TO_TWO:
3208                         clean_merge = 0;
3209                         if (handle_rename_rename_1to2(o, conflict_info))
3210                                 clean_merge = -1;
3211                         break;
3212                 case RENAME_TWO_FILES_TO_ONE:
3213                         clean_merge = 0;
3214                         if (handle_rename_rename_2to1(o, conflict_info))
3215                                 clean_merge = -1;
3216                         break;
3217                 default:
3218                         entry->processed = 0;
3219                         break;
3220                 }
3221         } else if (o_oid && (!a_oid || !b_oid)) {
3222                 /* Case A: Deleted in one */
3223                 if ((!a_oid && !b_oid) ||
3224                     (!b_oid && blob_unchanged(o, o_oid, o_mode, a_oid, a_mode, normalize, path)) ||
3225                     (!a_oid && blob_unchanged(o, o_oid, o_mode, b_oid, b_mode, normalize, path))) {
3226                         /* Deleted in both or deleted in one and
3227                          * unchanged in the other */
3228                         if (a_oid)
3229                                 output(o, 2, _("Removing %s"), path);
3230                         /* do not touch working file if it did not exist */
3231                         remove_file(o, 1, path, !a_oid);
3232                 } else {
3233                         /* Modify/delete; deleted side may have put a directory in the way */
3234                         clean_merge = 0;
3235                         if (handle_modify_delete(o, path, o_oid, o_mode,
3236                                                  a_oid, a_mode, b_oid, b_mode))
3237                                 clean_merge = -1;
3238                 }
3239         } else if ((!o_oid && a_oid && !b_oid) ||
3240                    (!o_oid && !a_oid && b_oid)) {
3241                 /* Case B: Added in one. */
3242                 /* [nothing|directory] -> ([nothing|directory], file) */
3243
3244                 const char *add_branch;
3245                 const char *other_branch;
3246                 unsigned mode;
3247                 const struct object_id *oid;
3248                 const char *conf;
3249
3250                 if (a_oid) {
3251                         add_branch = o->branch1;
3252                         other_branch = o->branch2;
3253                         mode = a_mode;
3254                         oid = a_oid;
3255                         conf = _("file/directory");
3256                 } else {
3257                         add_branch = o->branch2;
3258                         other_branch = o->branch1;
3259                         mode = b_mode;
3260                         oid = b_oid;
3261                         conf = _("directory/file");
3262                 }
3263                 if (dir_in_way(path,
3264                                !o->call_depth && !S_ISGITLINK(a_mode),
3265                                0)) {
3266                         char *new_path = unique_path(o, path, add_branch);
3267                         clean_merge = 0;
3268                         output(o, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3269                                "Adding %s as %s"),
3270                                conf, path, other_branch, path, new_path);
3271                         if (update_file(o, 0, oid, mode, new_path))
3272                                 clean_merge = -1;
3273                         else if (o->call_depth)
3274                                 remove_file_from_cache(path);
3275                         free(new_path);
3276                 } else {
3277                         output(o, 2, _("Adding %s"), path);
3278                         /* do not overwrite file if already present */
3279                         if (update_file_flags(o, oid, mode, path, 1, !a_oid))
3280                                 clean_merge = -1;
3281                 }
3282         } else if (a_oid && b_oid) {
3283                 /* Case C: Added in both (check for same permissions) and */
3284                 /* case D: Modified in both, but differently. */
3285                 int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3286                 clean_merge = merge_content(o, path, is_dirty,
3287                                             o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3288                                             NULL);
3289         } else if (!o_oid && !a_oid && !b_oid) {
3290                 /*
3291                  * this entry was deleted altogether. a_mode == 0 means
3292                  * we had that path and want to actively remove it.
3293                  */
3294                 remove_file(o, 1, path, !a_mode);
3295         } else
3296                 BUG("fatal merge failure, shouldn't happen.");
3297
3298         return clean_merge;
3299 }
3300
3301 int merge_trees(struct merge_options *o,
3302                 struct tree *head,
3303                 struct tree *merge,
3304                 struct tree *common,
3305                 struct tree **result)
3306 {
3307         int code, clean;
3308         struct strbuf sb = STRBUF_INIT;
3309
3310         if (!o->call_depth && index_has_changes(&the_index, head, &sb)) {
3311                 err(o, _("Your local changes to the following files would be overwritten by merge:\n  %s"),
3312                     sb.buf);
3313                 return -1;
3314         }
3315
3316         if (o->subtree_shift) {
3317                 merge = shift_tree_object(head, merge, o->subtree_shift);
3318                 common = shift_tree_object(head, common, o->subtree_shift);
3319         }
3320
3321         if (oid_eq(&common->object.oid, &merge->object.oid)) {
3322                 output(o, 0, _("Already up to date!"));
3323                 *result = head;
3324                 return 1;
3325         }
3326
3327         code = unpack_trees_start(o, common, head, merge);
3328
3329         if (code != 0) {
3330                 if (show(o, 4) || o->call_depth)
3331                         err(o, _("merging of trees %s and %s failed"),
3332                             oid_to_hex(&head->object.oid),
3333                             oid_to_hex(&merge->object.oid));
3334                 unpack_trees_finish(o);
3335                 return -1;
3336         }
3337
3338         if (unmerged_cache()) {
3339                 struct string_list *entries;
3340                 struct rename_info re_info;
3341                 int i;
3342                 /*
3343                  * Only need the hashmap while processing entries, so
3344                  * initialize it here and free it when we are done running
3345                  * through the entries. Keeping it in the merge_options as
3346                  * opposed to decaring a local hashmap is for convenience
3347                  * so that we don't have to pass it to around.
3348                  */
3349                 hashmap_init(&o->current_file_dir_set, path_hashmap_cmp, NULL, 512);
3350                 get_files_dirs(o, head);
3351                 get_files_dirs(o, merge);
3352
3353                 entries = get_unmerged();
3354                 clean = detect_and_process_renames(o, common, head, merge,
3355                                                    entries, &re_info);
3356                 record_df_conflict_files(o, entries);
3357                 if (clean < 0)
3358                         goto cleanup;
3359                 for (i = entries->nr-1; 0 <= i; i--) {
3360                         const char *path = entries->items[i].string;
3361                         struct stage_data *e = entries->items[i].util;
3362                         if (!e->processed) {
3363                                 int ret = process_entry(o, path, e);
3364                                 if (!ret)
3365                                         clean = 0;
3366                                 else if (ret < 0) {
3367                                         clean = ret;
3368                                         goto cleanup;
3369                                 }
3370                         }
3371                 }
3372                 for (i = 0; i < entries->nr; i++) {
3373                         struct stage_data *e = entries->items[i].util;
3374                         if (!e->processed)
3375                                 BUG("unprocessed path??? %s",
3376                                     entries->items[i].string);
3377                 }
3378
3379         cleanup:
3380                 final_cleanup_renames(&re_info);
3381
3382                 string_list_clear(entries, 1);
3383                 free(entries);
3384
3385                 hashmap_free(&o->current_file_dir_set, 1);
3386
3387                 if (clean < 0) {
3388                         unpack_trees_finish(o);
3389                         return clean;
3390                 }
3391         }
3392         else
3393                 clean = 1;
3394
3395         unpack_trees_finish(o);
3396
3397         if (o->call_depth && !(*result = write_tree_from_memory(o)))
3398                 return -1;
3399
3400         return clean;
3401 }
3402
3403 static struct commit_list *reverse_commit_list(struct commit_list *list)
3404 {
3405         struct commit_list *next = NULL, *current, *backup;
3406         for (current = list; current; current = backup) {
3407                 backup = current->next;
3408                 current->next = next;
3409                 next = current;
3410         }
3411         return next;
3412 }
3413
3414 /*
3415  * Merge the commits h1 and h2, return the resulting virtual
3416  * commit object and a flag indicating the cleanness of the merge.
3417  */
3418 int merge_recursive(struct merge_options *o,
3419                     struct commit *h1,
3420                     struct commit *h2,
3421                     struct commit_list *ca,
3422                     struct commit **result)
3423 {
3424         struct commit_list *iter;
3425         struct commit *merged_common_ancestors;
3426         struct tree *mrtree;
3427         int clean;
3428
3429         if (show(o, 4)) {
3430                 output(o, 4, _("Merging:"));
3431                 output_commit_title(o, h1);
3432                 output_commit_title(o, h2);
3433         }
3434
3435         if (!ca) {
3436                 ca = get_merge_bases(h1, h2);
3437                 ca = reverse_commit_list(ca);
3438         }
3439
3440         if (show(o, 5)) {
3441                 unsigned cnt = commit_list_count(ca);
3442
3443                 output(o, 5, Q_("found %u common ancestor:",
3444                                 "found %u common ancestors:", cnt), cnt);
3445                 for (iter = ca; iter; iter = iter->next)
3446                         output_commit_title(o, iter->item);
3447         }
3448
3449         merged_common_ancestors = pop_commit(&ca);
3450         if (merged_common_ancestors == NULL) {
3451                 /* if there is no common ancestor, use an empty tree */
3452                 struct tree *tree;
3453
3454                 tree = lookup_tree(the_repository, the_repository->hash_algo->empty_tree);
3455                 merged_common_ancestors = make_virtual_commit(tree, "ancestor");
3456         }
3457
3458         for (iter = ca; iter; iter = iter->next) {
3459                 const char *saved_b1, *saved_b2;
3460                 o->call_depth++;
3461                 /*
3462                  * When the merge fails, the result contains files
3463                  * with conflict markers. The cleanness flag is
3464                  * ignored (unless indicating an error), it was never
3465                  * actually used, as result of merge_trees has always
3466                  * overwritten it: the committed "conflicts" were
3467                  * already resolved.
3468                  */
3469                 discard_cache();
3470                 saved_b1 = o->branch1;
3471                 saved_b2 = o->branch2;
3472                 o->branch1 = "Temporary merge branch 1";
3473                 o->branch2 = "Temporary merge branch 2";
3474                 if (merge_recursive(o, merged_common_ancestors, iter->item,
3475                                     NULL, &merged_common_ancestors) < 0)
3476                         return -1;
3477                 o->branch1 = saved_b1;
3478                 o->branch2 = saved_b2;
3479                 o->call_depth--;
3480
3481                 if (!merged_common_ancestors)
3482                         return err(o, _("merge returned no commit"));
3483         }
3484
3485         discard_cache();
3486         if (!o->call_depth)
3487                 read_cache();
3488
3489         o->ancestor = "merged common ancestors";
3490         clean = merge_trees(o, get_commit_tree(h1), get_commit_tree(h2),
3491                             get_commit_tree(merged_common_ancestors),
3492                             &mrtree);
3493         if (clean < 0) {
3494                 flush_output(o);
3495                 return clean;
3496         }
3497
3498         if (o->call_depth) {
3499                 *result = make_virtual_commit(mrtree, "merged tree");
3500                 commit_list_insert(h1, &(*result)->parents);
3501                 commit_list_insert(h2, &(*result)->parents->next);
3502         }
3503         flush_output(o);
3504         if (!o->call_depth && o->buffer_output < 2)
3505                 strbuf_release(&o->obuf);
3506         if (show(o, 2))
3507                 diff_warn_rename_limit("merge.renamelimit",
3508                                        o->needed_rename_limit, 0);
3509         return clean;
3510 }
3511
3512 static struct commit *get_ref(const struct object_id *oid, const char *name)
3513 {
3514         struct object *object;
3515
3516         object = deref_tag(the_repository, parse_object(the_repository, oid),
3517                            name,
3518                            strlen(name));
3519         if (!object)
3520                 return NULL;
3521         if (object->type == OBJ_TREE)
3522                 return make_virtual_commit((struct tree*)object, name);
3523         if (object->type != OBJ_COMMIT)
3524                 return NULL;
3525         if (parse_commit((struct commit *)object))
3526                 return NULL;
3527         return (struct commit *)object;
3528 }
3529
3530 int merge_recursive_generic(struct merge_options *o,
3531                             const struct object_id *head,
3532                             const struct object_id *merge,
3533                             int num_base_list,
3534                             const struct object_id **base_list,
3535                             struct commit **result)
3536 {
3537         int clean;
3538         struct lock_file lock = LOCK_INIT;
3539         struct commit *head_commit = get_ref(head, o->branch1);
3540         struct commit *next_commit = get_ref(merge, o->branch2);
3541         struct commit_list *ca = NULL;
3542
3543         if (base_list) {
3544                 int i;
3545                 for (i = 0; i < num_base_list; ++i) {
3546                         struct commit *base;
3547                         if (!(base = get_ref(base_list[i], oid_to_hex(base_list[i]))))
3548                                 return err(o, _("Could not parse object '%s'"),
3549                                            oid_to_hex(base_list[i]));
3550                         commit_list_insert(base, &ca);
3551                 }
3552         }
3553
3554         hold_locked_index(&lock, LOCK_DIE_ON_ERROR);
3555         clean = merge_recursive(o, head_commit, next_commit, ca,
3556                                 result);
3557         if (clean < 0) {
3558                 rollback_lock_file(&lock);
3559                 return clean;
3560         }
3561
3562         if (write_locked_index(&the_index, &lock,
3563                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3564                 return err(o, _("Unable to write index."));
3565
3566         return clean ? 0 : 1;
3567 }
3568
3569 static void merge_recursive_config(struct merge_options *o)
3570 {
3571         char *value = NULL;
3572         git_config_get_int("merge.verbosity", &o->verbosity);
3573         git_config_get_int("diff.renamelimit", &o->diff_rename_limit);
3574         git_config_get_int("merge.renamelimit", &o->merge_rename_limit);
3575         if (!git_config_get_string("diff.renames", &value)) {
3576                 o->diff_detect_rename = git_config_rename("diff.renames", value);
3577                 free(value);
3578         }
3579         if (!git_config_get_string("merge.renames", &value)) {
3580                 o->merge_detect_rename = git_config_rename("merge.renames", value);
3581                 free(value);
3582         }
3583         git_config(git_xmerge_config, NULL);
3584 }
3585
3586 void init_merge_options(struct merge_options *o)
3587 {
3588         const char *merge_verbosity;
3589         memset(o, 0, sizeof(struct merge_options));
3590         o->verbosity = 2;
3591         o->buffer_output = 1;
3592         o->diff_rename_limit = -1;
3593         o->merge_rename_limit = -1;
3594         o->renormalize = 0;
3595         o->diff_detect_rename = -1;
3596         o->merge_detect_rename = -1;
3597         o->detect_directory_renames = 1;
3598         merge_recursive_config(o);
3599         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3600         if (merge_verbosity)
3601                 o->verbosity = strtol(merge_verbosity, NULL, 10);
3602         if (o->verbosity >= 5)
3603                 o->buffer_output = 0;
3604         strbuf_init(&o->obuf, 0);
3605         string_list_init(&o->df_conflict_file_set, 1);
3606 }
3607
3608 int parse_merge_opt(struct merge_options *o, const char *s)
3609 {
3610         const char *arg;
3611
3612         if (!s || !*s)
3613                 return -1;
3614         if (!strcmp(s, "ours"))
3615                 o->recursive_variant = MERGE_RECURSIVE_OURS;
3616         else if (!strcmp(s, "theirs"))
3617                 o->recursive_variant = MERGE_RECURSIVE_THEIRS;
3618         else if (!strcmp(s, "subtree"))
3619                 o->subtree_shift = "";
3620         else if (skip_prefix(s, "subtree=", &arg))
3621                 o->subtree_shift = arg;
3622         else if (!strcmp(s, "patience"))
3623                 o->xdl_opts = DIFF_WITH_ALG(o, PATIENCE_DIFF);
3624         else if (!strcmp(s, "histogram"))
3625                 o->xdl_opts = DIFF_WITH_ALG(o, HISTOGRAM_DIFF);
3626         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3627                 long value = parse_algorithm_value(arg);
3628                 if (value < 0)
3629                         return -1;
3630                 /* clear out previous settings */
3631                 DIFF_XDL_CLR(o, NEED_MINIMAL);
3632                 o->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3633                 o->xdl_opts |= value;
3634         }
3635         else if (!strcmp(s, "ignore-space-change"))
3636                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_CHANGE);
3637         else if (!strcmp(s, "ignore-all-space"))
3638                 DIFF_XDL_SET(o, IGNORE_WHITESPACE);
3639         else if (!strcmp(s, "ignore-space-at-eol"))
3640                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_AT_EOL);
3641         else if (!strcmp(s, "ignore-cr-at-eol"))
3642                 DIFF_XDL_SET(o, IGNORE_CR_AT_EOL);
3643         else if (!strcmp(s, "renormalize"))
3644                 o->renormalize = 1;
3645         else if (!strcmp(s, "no-renormalize"))
3646                 o->renormalize = 0;
3647         else if (!strcmp(s, "no-renames"))
3648                 o->merge_detect_rename = 0;
3649         else if (!strcmp(s, "find-renames")) {
3650                 o->merge_detect_rename = 1;
3651                 o->rename_score = 0;
3652         }
3653         else if (skip_prefix(s, "find-renames=", &arg) ||
3654                  skip_prefix(s, "rename-threshold=", &arg)) {
3655                 if ((o->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3656                         return -1;
3657                 o->merge_detect_rename = 1;
3658         }
3659         else
3660                 return -1;
3661         return 0;
3662 }