merge-recursive: move some struct declarations together
[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 *opt)
119 {
120         if (opt->buffer_output < 2 && opt->obuf.len) {
121                 fputs(opt->obuf.buf, stdout);
122                 strbuf_reset(&opt->obuf);
123         }
124 }
125
126 static int err(struct merge_options *opt, const char *err, ...)
127 {
128         va_list params;
129
130         if (opt->buffer_output < 2)
131                 flush_output(opt);
132         else {
133                 strbuf_complete(&opt->obuf, '\n');
134                 strbuf_addstr(&opt->obuf, "error: ");
135         }
136         va_start(params, err);
137         strbuf_vaddf(&opt->obuf, err, params);
138         va_end(params);
139         if (opt->buffer_output > 1)
140                 strbuf_addch(&opt->obuf, '\n');
141         else {
142                 error("%s", opt->obuf.buf);
143                 strbuf_reset(&opt->obuf);
144         }
145
146         return -1;
147 }
148
149 static struct tree *shift_tree_object(struct repository *repo,
150                                       struct tree *one, struct tree *two,
151                                       const char *subtree_shift)
152 {
153         struct object_id shifted;
154
155         if (!*subtree_shift) {
156                 shift_tree(&one->object.oid, &two->object.oid, &shifted, 0);
157         } else {
158                 shift_tree_by(&one->object.oid, &two->object.oid, &shifted,
159                               subtree_shift);
160         }
161         if (oideq(&two->object.oid, &shifted))
162                 return two;
163         return lookup_tree(repo, &shifted);
164 }
165
166 static struct commit *make_virtual_commit(struct repository *repo,
167                                           struct tree *tree,
168                                           const char *comment)
169 {
170         struct commit *commit = alloc_commit_node(repo);
171
172         set_merge_remote_desc(commit, comment, (struct object *)commit);
173         commit->maybe_tree = tree;
174         commit->object.parsed = 1;
175         return commit;
176 }
177
178 /*
179  * Since we use get_tree_entry(), which does not put the read object into
180  * the object pool, we cannot rely on a == b.
181  */
182 static int oid_eq(const struct object_id *a, const struct object_id *b)
183 {
184         if (!a && !b)
185                 return 2;
186         return a && b && oideq(a, b);
187 }
188
189 enum rename_type {
190         RENAME_NORMAL = 0,
191         RENAME_VIA_DIR,
192         RENAME_ADD,
193         RENAME_DELETE,
194         RENAME_ONE_FILE_TO_ONE,
195         RENAME_ONE_FILE_TO_TWO,
196         RENAME_TWO_FILES_TO_ONE
197 };
198
199 /*
200  * Since we want to write the index eventually, we cannot reuse the index
201  * for these (temporary) data.
202  */
203 struct stage_data {
204         struct {
205                 unsigned short mode;
206                 struct object_id oid;
207         } stages[4];
208         struct rename_conflict_info *rename_conflict_info;
209         unsigned processed:1;
210 };
211
212 struct rename {
213         struct diff_filepair *pair;
214         /*
215          * Purpose of src_entry and dst_entry:
216          *
217          * If 'before' is renamed to 'after' then src_entry will contain
218          * the versions of 'before' from the merge_base, HEAD, and MERGE in
219          * stages 1, 2, and 3; dst_entry will contain the respective
220          * versions of 'after' in corresponding locations.  Thus, we have a
221          * total of six modes and oids, though some will be null.  (Stage 0
222          * is ignored; we're interested in handling conflicts.)
223          *
224          * Since we don't turn on break-rewrites by default, neither
225          * src_entry nor dst_entry can have all three of their stages have
226          * non-null oids, meaning at most four of the six will be non-null.
227          * Also, since this is a rename, both src_entry and dst_entry will
228          * have at least one non-null oid, meaning at least two will be
229          * non-null.  Of the six oids, a typical rename will have three be
230          * non-null.  Only two implies a rename/delete, and four implies a
231          * rename/add.
232          */
233         struct stage_data *src_entry;
234         struct stage_data *dst_entry;
235         unsigned add_turned_into_rename:1;
236         unsigned processed:1;
237 };
238
239 struct rename_conflict_info {
240         enum rename_type rename_type;
241         struct diff_filepair *pair1;
242         struct diff_filepair *pair2;
243         const char *branch1;
244         const char *branch2;
245         struct stage_data *dst_entry1;
246         struct stage_data *dst_entry2;
247         struct diff_filespec ren1_other;
248         struct diff_filespec ren2_other;
249 };
250
251 static inline void setup_rename_conflict_info(enum rename_type rename_type,
252                                               struct diff_filepair *pair1,
253                                               struct diff_filepair *pair2,
254                                               const char *branch1,
255                                               const char *branch2,
256                                               struct stage_data *dst_entry1,
257                                               struct stage_data *dst_entry2,
258                                               struct merge_options *opt,
259                                               struct stage_data *src_entry1,
260                                               struct stage_data *src_entry2)
261 {
262         int ostage1 = 0, ostage2;
263         struct rename_conflict_info *ci;
264
265         /*
266          * When we have two renames involved, it's easiest to get the
267          * correct things into stage 2 and 3, and to make sure that the
268          * content merge puts HEAD before the other branch if we just
269          * ensure that branch1 == opt->branch1.  So, simply flip arguments
270          * around if we don't have that.
271          */
272         if (dst_entry2 && branch1 != opt->branch1) {
273                 setup_rename_conflict_info(rename_type,
274                                            pair2,      pair1,
275                                            branch2,    branch1,
276                                            dst_entry2, dst_entry1,
277                                            opt,
278                                            src_entry2, src_entry1);
279                 return;
280         }
281
282         ci = xcalloc(1, sizeof(struct rename_conflict_info));
283         ci->rename_type = rename_type;
284         ci->pair1 = pair1;
285         ci->branch1 = branch1;
286         ci->branch2 = branch2;
287
288         ci->dst_entry1 = dst_entry1;
289         dst_entry1->rename_conflict_info = ci;
290         dst_entry1->processed = 0;
291
292         assert(!pair2 == !dst_entry2);
293         if (dst_entry2) {
294                 ci->dst_entry2 = dst_entry2;
295                 ci->pair2 = pair2;
296                 dst_entry2->rename_conflict_info = ci;
297         }
298
299         /*
300          * For each rename, there could have been
301          * modifications on the side of history where that
302          * file was not renamed.
303          */
304         if (rename_type == RENAME_ADD ||
305             rename_type == RENAME_TWO_FILES_TO_ONE) {
306                 ostage1 = opt->branch1 == branch1 ? 3 : 2;
307
308                 ci->ren1_other.path = pair1->one->path;
309                 oidcpy(&ci->ren1_other.oid, &src_entry1->stages[ostage1].oid);
310                 ci->ren1_other.mode = src_entry1->stages[ostage1].mode;
311         }
312
313         if (rename_type == RENAME_TWO_FILES_TO_ONE) {
314                 ostage2 = ostage1 ^ 1;
315
316                 ci->ren2_other.path = pair2->one->path;
317                 oidcpy(&ci->ren2_other.oid, &src_entry2->stages[ostage2].oid);
318                 ci->ren2_other.mode = src_entry2->stages[ostage2].mode;
319         }
320 }
321
322 static int show(struct merge_options *opt, int v)
323 {
324         return (!opt->call_depth && opt->verbosity >= v) || opt->verbosity >= 5;
325 }
326
327 __attribute__((format (printf, 3, 4)))
328 static void output(struct merge_options *opt, int v, const char *fmt, ...)
329 {
330         va_list ap;
331
332         if (!show(opt, v))
333                 return;
334
335         strbuf_addchars(&opt->obuf, ' ', opt->call_depth * 2);
336
337         va_start(ap, fmt);
338         strbuf_vaddf(&opt->obuf, fmt, ap);
339         va_end(ap);
340
341         strbuf_addch(&opt->obuf, '\n');
342         if (!opt->buffer_output)
343                 flush_output(opt);
344 }
345
346 static void output_commit_title(struct merge_options *opt, struct commit *commit)
347 {
348         struct merge_remote_desc *desc;
349
350         strbuf_addchars(&opt->obuf, ' ', opt->call_depth * 2);
351         desc = merge_remote_util(commit);
352         if (desc)
353                 strbuf_addf(&opt->obuf, "virtual %s\n", desc->name);
354         else {
355                 strbuf_add_unique_abbrev(&opt->obuf, &commit->object.oid,
356                                          DEFAULT_ABBREV);
357                 strbuf_addch(&opt->obuf, ' ');
358                 if (parse_commit(commit) != 0)
359                         strbuf_addstr(&opt->obuf, _("(bad commit)\n"));
360                 else {
361                         const char *title;
362                         const char *msg = get_commit_buffer(commit, NULL);
363                         int len = find_commit_subject(msg, &title);
364                         if (len)
365                                 strbuf_addf(&opt->obuf, "%.*s\n", len, title);
366                         unuse_commit_buffer(commit, msg);
367                 }
368         }
369         flush_output(opt);
370 }
371
372 static int add_cacheinfo(struct merge_options *opt,
373                          unsigned int mode, const struct object_id *oid,
374                          const char *path, int stage, int refresh, int options)
375 {
376         struct index_state *istate = opt->repo->index;
377         struct cache_entry *ce;
378         int ret;
379
380         ce = make_cache_entry(istate, mode, oid ? oid : &null_oid, path, stage, 0);
381         if (!ce)
382                 return err(opt, _("add_cacheinfo failed for path '%s'; merge aborting."), path);
383
384         ret = add_index_entry(istate, ce, options);
385         if (refresh) {
386                 struct cache_entry *nce;
387
388                 nce = refresh_cache_entry(istate, ce,
389                                           CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
390                 if (!nce)
391                         return err(opt, _("add_cacheinfo failed to refresh for path '%s'; merge aborting."), path);
392                 if (nce != ce)
393                         ret = add_index_entry(istate, nce, options);
394         }
395         return ret;
396 }
397
398 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
399 {
400         parse_tree(tree);
401         init_tree_desc(desc, tree->buffer, tree->size);
402 }
403
404 static int unpack_trees_start(struct merge_options *opt,
405                               struct tree *common,
406                               struct tree *head,
407                               struct tree *merge)
408 {
409         int rc;
410         struct tree_desc t[3];
411         struct index_state tmp_index = { NULL };
412
413         memset(&opt->unpack_opts, 0, sizeof(opt->unpack_opts));
414         if (opt->call_depth)
415                 opt->unpack_opts.index_only = 1;
416         else
417                 opt->unpack_opts.update = 1;
418         opt->unpack_opts.merge = 1;
419         opt->unpack_opts.head_idx = 2;
420         opt->unpack_opts.fn = threeway_merge;
421         opt->unpack_opts.src_index = opt->repo->index;
422         opt->unpack_opts.dst_index = &tmp_index;
423         opt->unpack_opts.aggressive = !merge_detect_rename(opt);
424         setup_unpack_trees_porcelain(&opt->unpack_opts, "merge");
425
426         init_tree_desc_from_tree(t+0, common);
427         init_tree_desc_from_tree(t+1, head);
428         init_tree_desc_from_tree(t+2, merge);
429
430         rc = unpack_trees(3, t, &opt->unpack_opts);
431         cache_tree_free(&opt->repo->index->cache_tree);
432
433         /*
434          * Update opt->repo->index to match the new results, AFTER saving a copy
435          * in opt->orig_index.  Update src_index to point to the saved copy.
436          * (verify_uptodate() checks src_index, and the original index is
437          * the one that had the necessary modification timestamps.)
438          */
439         opt->orig_index = *opt->repo->index;
440         *opt->repo->index = tmp_index;
441         opt->unpack_opts.src_index = &opt->orig_index;
442
443         return rc;
444 }
445
446 static void unpack_trees_finish(struct merge_options *opt)
447 {
448         discard_index(&opt->orig_index);
449         clear_unpack_trees_porcelain(&opt->unpack_opts);
450 }
451
452 struct tree *write_tree_from_memory(struct merge_options *opt)
453 {
454         struct tree *result = NULL;
455         struct index_state *istate = opt->repo->index;
456
457         if (unmerged_index(istate)) {
458                 int i;
459                 fprintf(stderr, "BUG: There are unmerged index entries:\n");
460                 for (i = 0; i < istate->cache_nr; i++) {
461                         const struct cache_entry *ce = istate->cache[i];
462                         if (ce_stage(ce))
463                                 fprintf(stderr, "BUG: %d %.*s\n", ce_stage(ce),
464                                         (int)ce_namelen(ce), ce->name);
465                 }
466                 BUG("unmerged index entries in merge-recursive.c");
467         }
468
469         if (!istate->cache_tree)
470                 istate->cache_tree = cache_tree();
471
472         if (!cache_tree_fully_valid(istate->cache_tree) &&
473             cache_tree_update(istate, 0) < 0) {
474                 err(opt, _("error building trees"));
475                 return NULL;
476         }
477
478         result = lookup_tree(opt->repo, &istate->cache_tree->oid);
479
480         return result;
481 }
482
483 static int save_files_dirs(const struct object_id *oid,
484                            struct strbuf *base, const char *path,
485                            unsigned int mode, int stage, void *context)
486 {
487         struct path_hashmap_entry *entry;
488         int baselen = base->len;
489         struct merge_options *opt = context;
490
491         strbuf_addstr(base, path);
492
493         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
494         hashmap_entry_init(entry, path_hash(entry->path));
495         hashmap_add(&opt->current_file_dir_set, entry);
496
497         strbuf_setlen(base, baselen);
498         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
499 }
500
501 static void get_files_dirs(struct merge_options *opt, struct tree *tree)
502 {
503         struct pathspec match_all;
504         memset(&match_all, 0, sizeof(match_all));
505         read_tree_recursive(the_repository, tree, "", 0, 0,
506                             &match_all, save_files_dirs, opt);
507 }
508
509 static int get_tree_entry_if_blob(const struct object_id *tree,
510                                   const char *path,
511                                   struct object_id *hashy,
512                                   unsigned short *mode_o)
513 {
514         int ret;
515
516         ret = get_tree_entry(tree, path, hashy, mode_o);
517         if (S_ISDIR(*mode_o)) {
518                 oidcpy(hashy, &null_oid);
519                 *mode_o = 0;
520         }
521         return ret;
522 }
523
524 /*
525  * Returns an index_entry instance which doesn't have to correspond to
526  * a real cache entry in Git's index.
527  */
528 static struct stage_data *insert_stage_data(const char *path,
529                 struct tree *o, struct tree *a, struct tree *b,
530                 struct string_list *entries)
531 {
532         struct string_list_item *item;
533         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
534         get_tree_entry_if_blob(&o->object.oid, path,
535                                &e->stages[1].oid, &e->stages[1].mode);
536         get_tree_entry_if_blob(&a->object.oid, path,
537                                &e->stages[2].oid, &e->stages[2].mode);
538         get_tree_entry_if_blob(&b->object.oid, path,
539                                &e->stages[3].oid, &e->stages[3].mode);
540         item = string_list_insert(entries, path);
541         item->util = e;
542         return e;
543 }
544
545 /*
546  * Create a dictionary mapping file names to stage_data objects. The
547  * dictionary contains one entry for every path with a non-zero stage entry.
548  */
549 static struct string_list *get_unmerged(struct index_state *istate)
550 {
551         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
552         int i;
553
554         unmerged->strdup_strings = 1;
555
556         for (i = 0; i < istate->cache_nr; i++) {
557                 struct string_list_item *item;
558                 struct stage_data *e;
559                 const struct cache_entry *ce = istate->cache[i];
560                 if (!ce_stage(ce))
561                         continue;
562
563                 item = string_list_lookup(unmerged, ce->name);
564                 if (!item) {
565                         item = string_list_insert(unmerged, ce->name);
566                         item->util = xcalloc(1, sizeof(struct stage_data));
567                 }
568                 e = item->util;
569                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
570                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
571         }
572
573         return unmerged;
574 }
575
576 static int string_list_df_name_compare(const char *one, const char *two)
577 {
578         int onelen = strlen(one);
579         int twolen = strlen(two);
580         /*
581          * Here we only care that entries for D/F conflicts are
582          * adjacent, in particular with the file of the D/F conflict
583          * appearing before files below the corresponding directory.
584          * The order of the rest of the list is irrelevant for us.
585          *
586          * To achieve this, we sort with df_name_compare and provide
587          * the mode S_IFDIR so that D/F conflicts will sort correctly.
588          * We use the mode S_IFDIR for everything else for simplicity,
589          * since in other cases any changes in their order due to
590          * sorting cause no problems for us.
591          */
592         int cmp = df_name_compare(one, onelen, S_IFDIR,
593                                   two, twolen, S_IFDIR);
594         /*
595          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
596          * that 'foo' comes before 'foo/bar'.
597          */
598         if (cmp)
599                 return cmp;
600         return onelen - twolen;
601 }
602
603 static void record_df_conflict_files(struct merge_options *opt,
604                                      struct string_list *entries)
605 {
606         /* If there is a D/F conflict and the file for such a conflict
607          * currently exists in the working tree, we want to allow it to be
608          * removed to make room for the corresponding directory if needed.
609          * The files underneath the directories of such D/F conflicts will
610          * be processed before the corresponding file involved in the D/F
611          * conflict.  If the D/F directory ends up being removed by the
612          * merge, then we won't have to touch the D/F file.  If the D/F
613          * directory needs to be written to the working copy, then the D/F
614          * file will simply be removed (in make_room_for_path()) to make
615          * room for the necessary paths.  Note that if both the directory
616          * and the file need to be present, then the D/F file will be
617          * reinstated with a new unique name at the time it is processed.
618          */
619         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
620         const char *last_file = NULL;
621         int last_len = 0;
622         int i;
623
624         /*
625          * If we're merging merge-bases, we don't want to bother with
626          * any working directory changes.
627          */
628         if (opt->call_depth)
629                 return;
630
631         /* Ensure D/F conflicts are adjacent in the entries list. */
632         for (i = 0; i < entries->nr; i++) {
633                 struct string_list_item *next = &entries->items[i];
634                 string_list_append(&df_sorted_entries, next->string)->util =
635                                    next->util;
636         }
637         df_sorted_entries.cmp = string_list_df_name_compare;
638         string_list_sort(&df_sorted_entries);
639
640         string_list_clear(&opt->df_conflict_file_set, 1);
641         for (i = 0; i < df_sorted_entries.nr; i++) {
642                 const char *path = df_sorted_entries.items[i].string;
643                 int len = strlen(path);
644                 struct stage_data *e = df_sorted_entries.items[i].util;
645
646                 /*
647                  * Check if last_file & path correspond to a D/F conflict;
648                  * i.e. whether path is last_file+'/'+<something>.
649                  * If so, record that it's okay to remove last_file to make
650                  * room for path and friends if needed.
651                  */
652                 if (last_file &&
653                     len > last_len &&
654                     memcmp(path, last_file, last_len) == 0 &&
655                     path[last_len] == '/') {
656                         string_list_insert(&opt->df_conflict_file_set, last_file);
657                 }
658
659                 /*
660                  * Determine whether path could exist as a file in the
661                  * working directory as a possible D/F conflict.  This
662                  * will only occur when it exists in stage 2 as a
663                  * file.
664                  */
665                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
666                         last_file = path;
667                         last_len = len;
668                 } else {
669                         last_file = NULL;
670                 }
671         }
672         string_list_clear(&df_sorted_entries, 0);
673 }
674
675 static int update_stages(struct merge_options *opt, const char *path,
676                          const struct diff_filespec *o,
677                          const struct diff_filespec *a,
678                          const struct diff_filespec *b)
679 {
680
681         /*
682          * NOTE: It is usually a bad idea to call update_stages on a path
683          * before calling update_file on that same path, since it can
684          * sometimes lead to spurious "refusing to lose untracked file..."
685          * messages from update_file (via make_room_for path via
686          * would_lose_untracked).  Instead, reverse the order of the calls
687          * (executing update_file first and then update_stages).
688          */
689         int clear = 1;
690         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
691         if (clear)
692                 if (remove_file_from_index(opt->repo->index, path))
693                         return -1;
694         if (o)
695                 if (add_cacheinfo(opt, o->mode, &o->oid, path, 1, 0, options))
696                         return -1;
697         if (a)
698                 if (add_cacheinfo(opt, a->mode, &a->oid, path, 2, 0, options))
699                         return -1;
700         if (b)
701                 if (add_cacheinfo(opt, b->mode, &b->oid, path, 3, 0, options))
702                         return -1;
703         return 0;
704 }
705
706 static void update_entry(struct stage_data *entry,
707                          struct diff_filespec *o,
708                          struct diff_filespec *a,
709                          struct diff_filespec *b)
710 {
711         entry->processed = 0;
712         entry->stages[1].mode = o->mode;
713         entry->stages[2].mode = a->mode;
714         entry->stages[3].mode = b->mode;
715         oidcpy(&entry->stages[1].oid, &o->oid);
716         oidcpy(&entry->stages[2].oid, &a->oid);
717         oidcpy(&entry->stages[3].oid, &b->oid);
718 }
719
720 static int remove_file(struct merge_options *opt, int clean,
721                        const char *path, int no_wd)
722 {
723         int update_cache = opt->call_depth || clean;
724         int update_working_directory = !opt->call_depth && !no_wd;
725
726         if (update_cache) {
727                 if (remove_file_from_index(opt->repo->index, path))
728                         return -1;
729         }
730         if (update_working_directory) {
731                 if (ignore_case) {
732                         struct cache_entry *ce;
733                         ce = index_file_exists(opt->repo->index, path, strlen(path),
734                                                ignore_case);
735                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
736                                 return 0;
737                 }
738                 if (remove_path(path))
739                         return -1;
740         }
741         return 0;
742 }
743
744 /* add a string to a strbuf, but converting "/" to "_" */
745 static void add_flattened_path(struct strbuf *out, const char *s)
746 {
747         size_t i = out->len;
748         strbuf_addstr(out, s);
749         for (; i < out->len; i++)
750                 if (out->buf[i] == '/')
751                         out->buf[i] = '_';
752 }
753
754 static char *unique_path(struct merge_options *opt, const char *path, const char *branch)
755 {
756         struct path_hashmap_entry *entry;
757         struct strbuf newpath = STRBUF_INIT;
758         int suffix = 0;
759         size_t base_len;
760
761         strbuf_addf(&newpath, "%s~", path);
762         add_flattened_path(&newpath, branch);
763
764         base_len = newpath.len;
765         while (hashmap_get_from_hash(&opt->current_file_dir_set,
766                                      path_hash(newpath.buf), newpath.buf) ||
767                (!opt->call_depth && file_exists(newpath.buf))) {
768                 strbuf_setlen(&newpath, base_len);
769                 strbuf_addf(&newpath, "_%d", suffix++);
770         }
771
772         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
773         hashmap_entry_init(entry, path_hash(entry->path));
774         hashmap_add(&opt->current_file_dir_set, entry);
775         return strbuf_detach(&newpath, NULL);
776 }
777
778 /**
779  * Check whether a directory in the index is in the way of an incoming
780  * file.  Return 1 if so.  If check_working_copy is non-zero, also
781  * check the working directory.  If empty_ok is non-zero, also return
782  * 0 in the case where the working-tree dir exists but is empty.
783  */
784 static int dir_in_way(struct index_state *istate, const char *path,
785                       int check_working_copy, int empty_ok)
786 {
787         int pos;
788         struct strbuf dirpath = STRBUF_INIT;
789         struct stat st;
790
791         strbuf_addstr(&dirpath, path);
792         strbuf_addch(&dirpath, '/');
793
794         pos = index_name_pos(istate, dirpath.buf, dirpath.len);
795
796         if (pos < 0)
797                 pos = -1 - pos;
798         if (pos < istate->cache_nr &&
799             !strncmp(dirpath.buf, istate->cache[pos]->name, dirpath.len)) {
800                 strbuf_release(&dirpath);
801                 return 1;
802         }
803
804         strbuf_release(&dirpath);
805         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
806                 !(empty_ok && is_empty_dir(path));
807 }
808
809 /*
810  * Returns whether path was tracked in the index before the merge started,
811  * and its oid and mode match the specified values
812  */
813 static int was_tracked_and_matches(struct merge_options *opt, const char *path,
814                                    const struct object_id *oid, unsigned mode)
815 {
816         int pos = index_name_pos(&opt->orig_index, path, strlen(path));
817         struct cache_entry *ce;
818
819         if (0 > pos)
820                 /* we were not tracking this path before the merge */
821                 return 0;
822
823         /* See if the file we were tracking before matches */
824         ce = opt->orig_index.cache[pos];
825         return (oid_eq(&ce->oid, oid) && ce->ce_mode == mode);
826 }
827
828 /*
829  * Returns whether path was tracked in the index before the merge started
830  */
831 static int was_tracked(struct merge_options *opt, const char *path)
832 {
833         int pos = index_name_pos(&opt->orig_index, path, strlen(path));
834
835         if (0 <= pos)
836                 /* we were tracking this path before the merge */
837                 return 1;
838
839         return 0;
840 }
841
842 static int would_lose_untracked(struct merge_options *opt, const char *path)
843 {
844         struct index_state *istate = opt->repo->index;
845
846         /*
847          * This may look like it can be simplified to:
848          *   return !was_tracked(opt, path) && file_exists(path)
849          * but it can't.  This function needs to know whether path was in
850          * the working tree due to EITHER having been tracked in the index
851          * before the merge OR having been put into the working copy and
852          * index by unpack_trees().  Due to that either-or requirement, we
853          * check the current index instead of the original one.
854          *
855          * Note that we do not need to worry about merge-recursive itself
856          * updating the index after unpack_trees() and before calling this
857          * function, because we strictly require all code paths in
858          * merge-recursive to update the working tree first and the index
859          * second.  Doing otherwise would break
860          * update_file()/would_lose_untracked(); see every comment in this
861          * file which mentions "update_stages".
862          */
863         int pos = index_name_pos(istate, path, strlen(path));
864
865         if (pos < 0)
866                 pos = -1 - pos;
867         while (pos < istate->cache_nr &&
868                !strcmp(path, istate->cache[pos]->name)) {
869                 /*
870                  * If stage #0, it is definitely tracked.
871                  * If it has stage #2 then it was tracked
872                  * before this merge started.  All other
873                  * cases the path was not tracked.
874                  */
875                 switch (ce_stage(istate->cache[pos])) {
876                 case 0:
877                 case 2:
878                         return 0;
879                 }
880                 pos++;
881         }
882         return file_exists(path);
883 }
884
885 static int was_dirty(struct merge_options *opt, const char *path)
886 {
887         struct cache_entry *ce;
888         int dirty = 1;
889
890         if (opt->call_depth || !was_tracked(opt, path))
891                 return !dirty;
892
893         ce = index_file_exists(opt->unpack_opts.src_index,
894                                path, strlen(path), ignore_case);
895         dirty = verify_uptodate(ce, &opt->unpack_opts) != 0;
896         return dirty;
897 }
898
899 static int make_room_for_path(struct merge_options *opt, const char *path)
900 {
901         int status, i;
902         const char *msg = _("failed to create path '%s'%s");
903
904         /* Unlink any D/F conflict files that are in the way */
905         for (i = 0; i < opt->df_conflict_file_set.nr; i++) {
906                 const char *df_path = opt->df_conflict_file_set.items[i].string;
907                 size_t pathlen = strlen(path);
908                 size_t df_pathlen = strlen(df_path);
909                 if (df_pathlen < pathlen &&
910                     path[df_pathlen] == '/' &&
911                     strncmp(path, df_path, df_pathlen) == 0) {
912                         output(opt, 3,
913                                _("Removing %s to make room for subdirectory\n"),
914                                df_path);
915                         unlink(df_path);
916                         unsorted_string_list_delete_item(&opt->df_conflict_file_set,
917                                                          i, 0);
918                         break;
919                 }
920         }
921
922         /* Make sure leading directories are created */
923         status = safe_create_leading_directories_const(path);
924         if (status) {
925                 if (status == SCLD_EXISTS)
926                         /* something else exists */
927                         return err(opt, msg, path, _(": perhaps a D/F conflict?"));
928                 return err(opt, msg, path, "");
929         }
930
931         /*
932          * Do not unlink a file in the work tree if we are not
933          * tracking it.
934          */
935         if (would_lose_untracked(opt, path))
936                 return err(opt, _("refusing to lose untracked file at '%s'"),
937                            path);
938
939         /* Successful unlink is good.. */
940         if (!unlink(path))
941                 return 0;
942         /* .. and so is no existing file */
943         if (errno == ENOENT)
944                 return 0;
945         /* .. but not some other error (who really cares what?) */
946         return err(opt, msg, path, _(": perhaps a D/F conflict?"));
947 }
948
949 static int update_file_flags(struct merge_options *opt,
950                              const struct object_id *oid,
951                              unsigned mode,
952                              const char *path,
953                              int update_cache,
954                              int update_wd)
955 {
956         int ret = 0;
957
958         if (opt->call_depth)
959                 update_wd = 0;
960
961         if (update_wd) {
962                 enum object_type type;
963                 void *buf;
964                 unsigned long size;
965
966                 if (S_ISGITLINK(mode)) {
967                         /*
968                          * We may later decide to recursively descend into
969                          * the submodule directory and update its index
970                          * and/or work tree, but we do not do that now.
971                          */
972                         update_wd = 0;
973                         goto update_index;
974                 }
975
976                 buf = read_object_file(oid, &type, &size);
977                 if (!buf)
978                         return err(opt, _("cannot read object %s '%s'"), oid_to_hex(oid), path);
979                 if (type != OBJ_BLOB) {
980                         ret = err(opt, _("blob expected for %s '%s'"), oid_to_hex(oid), path);
981                         goto free_buf;
982                 }
983                 if (S_ISREG(mode)) {
984                         struct strbuf strbuf = STRBUF_INIT;
985                         if (convert_to_working_tree(opt->repo->index, path, buf, size, &strbuf)) {
986                                 free(buf);
987                                 size = strbuf.len;
988                                 buf = strbuf_detach(&strbuf, NULL);
989                         }
990                 }
991
992                 if (make_room_for_path(opt, path) < 0) {
993                         update_wd = 0;
994                         goto free_buf;
995                 }
996                 if (S_ISREG(mode) || (!has_symlinks && S_ISLNK(mode))) {
997                         int fd;
998                         if (mode & 0100)
999                                 mode = 0777;
1000                         else
1001                                 mode = 0666;
1002                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
1003                         if (fd < 0) {
1004                                 ret = err(opt, _("failed to open '%s': %s"),
1005                                           path, strerror(errno));
1006                                 goto free_buf;
1007                         }
1008                         write_in_full(fd, buf, size);
1009                         close(fd);
1010                 } else if (S_ISLNK(mode)) {
1011                         char *lnk = xmemdupz(buf, size);
1012                         safe_create_leading_directories_const(path);
1013                         unlink(path);
1014                         if (symlink(lnk, path))
1015                                 ret = err(opt, _("failed to symlink '%s': %s"),
1016                                           path, strerror(errno));
1017                         free(lnk);
1018                 } else
1019                         ret = err(opt,
1020                                   _("do not know what to do with %06o %s '%s'"),
1021                                   mode, oid_to_hex(oid), path);
1022         free_buf:
1023                 free(buf);
1024         }
1025 update_index:
1026         if (!ret && update_cache)
1027                 if (add_cacheinfo(opt, mode, oid, path, 0, update_wd,
1028                                   ADD_CACHE_OK_TO_ADD))
1029                         return -1;
1030         return ret;
1031 }
1032
1033 static int update_file(struct merge_options *opt,
1034                        int clean,
1035                        const struct object_id *oid,
1036                        unsigned mode,
1037                        const char *path)
1038 {
1039         return update_file_flags(opt, oid, mode, path, opt->call_depth || clean, !opt->call_depth);
1040 }
1041
1042 /* Low level file merging, update and removal */
1043
1044 struct merge_file_info {
1045         struct object_id oid;
1046         unsigned mode;
1047         unsigned clean:1,
1048                  merge:1;
1049 };
1050
1051 static int merge_3way(struct merge_options *opt,
1052                       mmbuffer_t *result_buf,
1053                       const struct diff_filespec *o,
1054                       const struct diff_filespec *a,
1055                       const struct diff_filespec *b,
1056                       const char *branch1,
1057                       const char *branch2,
1058                       const int extra_marker_size)
1059 {
1060         mmfile_t orig, src1, src2;
1061         struct ll_merge_options ll_opts = {0};
1062         char *base_name, *name1, *name2;
1063         int merge_status;
1064
1065         ll_opts.renormalize = opt->renormalize;
1066         ll_opts.extra_marker_size = extra_marker_size;
1067         ll_opts.xdl_opts = opt->xdl_opts;
1068
1069         if (opt->call_depth) {
1070                 ll_opts.virtual_ancestor = 1;
1071                 ll_opts.variant = 0;
1072         } else {
1073                 switch (opt->recursive_variant) {
1074                 case MERGE_RECURSIVE_OURS:
1075                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
1076                         break;
1077                 case MERGE_RECURSIVE_THEIRS:
1078                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
1079                         break;
1080                 default:
1081                         ll_opts.variant = 0;
1082                         break;
1083                 }
1084         }
1085
1086         if (strcmp(a->path, b->path) ||
1087             (opt->ancestor != NULL && strcmp(a->path, o->path) != 0)) {
1088                 base_name = opt->ancestor == NULL ? NULL :
1089                         mkpathdup("%s:%s", opt->ancestor, o->path);
1090                 name1 = mkpathdup("%s:%s", branch1, a->path);
1091                 name2 = mkpathdup("%s:%s", branch2, b->path);
1092         } else {
1093                 base_name = opt->ancestor == NULL ? NULL :
1094                         mkpathdup("%s", opt->ancestor);
1095                 name1 = mkpathdup("%s", branch1);
1096                 name2 = mkpathdup("%s", branch2);
1097         }
1098
1099         read_mmblob(&orig, &o->oid);
1100         read_mmblob(&src1, &a->oid);
1101         read_mmblob(&src2, &b->oid);
1102
1103         merge_status = ll_merge(result_buf, a->path, &orig, base_name,
1104                                 &src1, name1, &src2, name2,
1105                                 opt->repo->index, &ll_opts);
1106
1107         free(base_name);
1108         free(name1);
1109         free(name2);
1110         free(orig.ptr);
1111         free(src1.ptr);
1112         free(src2.ptr);
1113         return merge_status;
1114 }
1115
1116 static int find_first_merges(struct repository *repo,
1117                              struct object_array *result, const char *path,
1118                              struct commit *a, struct commit *b)
1119 {
1120         int i, j;
1121         struct object_array merges = OBJECT_ARRAY_INIT;
1122         struct commit *commit;
1123         int contains_another;
1124
1125         char merged_revision[42];
1126         const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
1127                                    "--all", merged_revision, NULL };
1128         struct rev_info revs;
1129         struct setup_revision_opt rev_opts;
1130
1131         memset(result, 0, sizeof(struct object_array));
1132         memset(&rev_opts, 0, sizeof(rev_opts));
1133
1134         /* get all revisions that merge commit a */
1135         xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
1136                   oid_to_hex(&a->object.oid));
1137         repo_init_revisions(repo, &revs, NULL);
1138         rev_opts.submodule = path;
1139         /* FIXME: can't handle linked worktrees in submodules yet */
1140         revs.single_worktree = path != NULL;
1141         setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
1142
1143         /* save all revisions from the above list that contain b */
1144         if (prepare_revision_walk(&revs))
1145                 die("revision walk setup failed");
1146         while ((commit = get_revision(&revs)) != NULL) {
1147                 struct object *o = &(commit->object);
1148                 if (in_merge_bases(b, commit))
1149                         add_object_array(o, NULL, &merges);
1150         }
1151         reset_revision_walk();
1152
1153         /* Now we've got all merges that contain a and b. Prune all
1154          * merges that contain another found merge and save them in
1155          * result.
1156          */
1157         for (i = 0; i < merges.nr; i++) {
1158                 struct commit *m1 = (struct commit *) merges.objects[i].item;
1159
1160                 contains_another = 0;
1161                 for (j = 0; j < merges.nr; j++) {
1162                         struct commit *m2 = (struct commit *) merges.objects[j].item;
1163                         if (i != j && in_merge_bases(m2, m1)) {
1164                                 contains_another = 1;
1165                                 break;
1166                         }
1167                 }
1168
1169                 if (!contains_another)
1170                         add_object_array(merges.objects[i].item, NULL, result);
1171         }
1172
1173         object_array_clear(&merges);
1174         return result->nr;
1175 }
1176
1177 static void print_commit(struct commit *commit)
1178 {
1179         struct strbuf sb = STRBUF_INIT;
1180         struct pretty_print_context ctx = {0};
1181         ctx.date_mode.type = DATE_NORMAL;
1182         format_commit_message(commit, " %h: %m %s", &sb, &ctx);
1183         fprintf(stderr, "%s\n", sb.buf);
1184         strbuf_release(&sb);
1185 }
1186
1187 static int merge_submodule(struct merge_options *opt,
1188                            struct object_id *result, const char *path,
1189                            const struct object_id *base, const struct object_id *a,
1190                            const struct object_id *b)
1191 {
1192         struct commit *commit_base, *commit_a, *commit_b;
1193         int parent_count;
1194         struct object_array merges;
1195
1196         int i;
1197         int search = !opt->call_depth;
1198
1199         /* store a in result in case we fail */
1200         oidcpy(result, a);
1201
1202         /* we can not handle deletion conflicts */
1203         if (is_null_oid(base))
1204                 return 0;
1205         if (is_null_oid(a))
1206                 return 0;
1207         if (is_null_oid(b))
1208                 return 0;
1209
1210         if (add_submodule_odb(path)) {
1211                 output(opt, 1, _("Failed to merge submodule %s (not checked out)"), path);
1212                 return 0;
1213         }
1214
1215         if (!(commit_base = lookup_commit_reference(opt->repo, base)) ||
1216             !(commit_a = lookup_commit_reference(opt->repo, a)) ||
1217             !(commit_b = lookup_commit_reference(opt->repo, b))) {
1218                 output(opt, 1, _("Failed to merge submodule %s (commits not present)"), path);
1219                 return 0;
1220         }
1221
1222         /* check whether both changes are forward */
1223         if (!in_merge_bases(commit_base, commit_a) ||
1224             !in_merge_bases(commit_base, commit_b)) {
1225                 output(opt, 1, _("Failed to merge submodule %s (commits don't follow merge-base)"), path);
1226                 return 0;
1227         }
1228
1229         /* Case #1: a is contained in b or vice versa */
1230         if (in_merge_bases(commit_a, commit_b)) {
1231                 oidcpy(result, b);
1232                 if (show(opt, 3)) {
1233                         output(opt, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1234                         output_commit_title(opt, commit_b);
1235                 } else if (show(opt, 2))
1236                         output(opt, 2, _("Fast-forwarding submodule %s"), path);
1237                 else
1238                         ; /* no output */
1239
1240                 return 1;
1241         }
1242         if (in_merge_bases(commit_b, commit_a)) {
1243                 oidcpy(result, a);
1244                 if (show(opt, 3)) {
1245                         output(opt, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1246                         output_commit_title(opt, commit_a);
1247                 } else if (show(opt, 2))
1248                         output(opt, 2, _("Fast-forwarding submodule %s"), path);
1249                 else
1250                         ; /* no output */
1251
1252                 return 1;
1253         }
1254
1255         /*
1256          * Case #2: There are one or more merges that contain a and b in
1257          * the submodule. If there is only one, then present it as a
1258          * suggestion to the user, but leave it marked unmerged so the
1259          * user needs to confirm the resolution.
1260          */
1261
1262         /* Skip the search if makes no sense to the calling context.  */
1263         if (!search)
1264                 return 0;
1265
1266         /* find commit which merges them */
1267         parent_count = find_first_merges(opt->repo, &merges, path,
1268                                          commit_a, commit_b);
1269         switch (parent_count) {
1270         case 0:
1271                 output(opt, 1, _("Failed to merge submodule %s (merge following commits not found)"), path);
1272                 break;
1273
1274         case 1:
1275                 output(opt, 1, _("Failed to merge submodule %s (not fast-forward)"), path);
1276                 output(opt, 2, _("Found a possible merge resolution for the submodule:\n"));
1277                 print_commit((struct commit *) merges.objects[0].item);
1278                 output(opt, 2, _(
1279                        "If this is correct simply add it to the index "
1280                        "for example\n"
1281                        "by using:\n\n"
1282                        "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
1283                        "which will accept this suggestion.\n"),
1284                        oid_to_hex(&merges.objects[0].item->oid), path);
1285                 break;
1286
1287         default:
1288                 output(opt, 1, _("Failed to merge submodule %s (multiple merges found)"), path);
1289                 for (i = 0; i < merges.nr; i++)
1290                         print_commit((struct commit *) merges.objects[i].item);
1291         }
1292
1293         object_array_clear(&merges);
1294         return 0;
1295 }
1296
1297 static int merge_mode_and_contents(struct merge_options *opt,
1298                                    const struct diff_filespec *o,
1299                                    const struct diff_filespec *a,
1300                                    const struct diff_filespec *b,
1301                                    const char *filename,
1302                                    const char *branch1,
1303                                    const char *branch2,
1304                                    const int extra_marker_size,
1305                                    struct merge_file_info *result)
1306 {
1307         if (opt->branch1 != branch1) {
1308                 /*
1309                  * It's weird getting a reverse merge with HEAD on the bottom
1310                  * side of the conflict markers and the other branch on the
1311                  * top.  Fix that.
1312                  */
1313                 return merge_mode_and_contents(opt, o, b, a,
1314                                                filename,
1315                                                branch2, branch1,
1316                                                extra_marker_size, result);
1317         }
1318
1319         result->merge = 0;
1320         result->clean = 1;
1321
1322         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
1323                 result->clean = 0;
1324                 if (S_ISREG(a->mode)) {
1325                         result->mode = a->mode;
1326                         oidcpy(&result->oid, &a->oid);
1327                 } else {
1328                         result->mode = b->mode;
1329                         oidcpy(&result->oid, &b->oid);
1330                 }
1331         } else {
1332                 if (!oid_eq(&a->oid, &o->oid) && !oid_eq(&b->oid, &o->oid))
1333                         result->merge = 1;
1334
1335                 /*
1336                  * Merge modes
1337                  */
1338                 if (a->mode == b->mode || a->mode == o->mode)
1339                         result->mode = b->mode;
1340                 else {
1341                         result->mode = a->mode;
1342                         if (b->mode != o->mode) {
1343                                 result->clean = 0;
1344                                 result->merge = 1;
1345                         }
1346                 }
1347
1348                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &o->oid))
1349                         oidcpy(&result->oid, &b->oid);
1350                 else if (oid_eq(&b->oid, &o->oid))
1351                         oidcpy(&result->oid, &a->oid);
1352                 else if (S_ISREG(a->mode)) {
1353                         mmbuffer_t result_buf;
1354                         int ret = 0, merge_status;
1355
1356                         merge_status = merge_3way(opt, &result_buf, o, a, b,
1357                                                   branch1, branch2,
1358                                                   extra_marker_size);
1359
1360                         if ((merge_status < 0) || !result_buf.ptr)
1361                                 ret = err(opt, _("Failed to execute internal merge"));
1362
1363                         if (!ret &&
1364                             write_object_file(result_buf.ptr, result_buf.size,
1365                                               blob_type, &result->oid))
1366                                 ret = err(opt, _("Unable to add %s to database"),
1367                                           a->path);
1368
1369                         free(result_buf.ptr);
1370                         if (ret)
1371                                 return ret;
1372                         result->clean = (merge_status == 0);
1373                 } else if (S_ISGITLINK(a->mode)) {
1374                         result->clean = merge_submodule(opt, &result->oid,
1375                                                         o->path,
1376                                                         &o->oid,
1377                                                         &a->oid,
1378                                                         &b->oid);
1379                 } else if (S_ISLNK(a->mode)) {
1380                         switch (opt->recursive_variant) {
1381                         case MERGE_RECURSIVE_NORMAL:
1382                                 oidcpy(&result->oid, &a->oid);
1383                                 if (!oid_eq(&a->oid, &b->oid))
1384                                         result->clean = 0;
1385                                 break;
1386                         case MERGE_RECURSIVE_OURS:
1387                                 oidcpy(&result->oid, &a->oid);
1388                                 break;
1389                         case MERGE_RECURSIVE_THEIRS:
1390                                 oidcpy(&result->oid, &b->oid);
1391                                 break;
1392                         }
1393                 } else
1394                         BUG("unsupported object type in the tree");
1395         }
1396
1397         if (result->merge)
1398                 output(opt, 2, _("Auto-merging %s"), filename);
1399
1400         return 0;
1401 }
1402
1403 static int handle_rename_via_dir(struct merge_options *opt,
1404                                  struct diff_filepair *pair,
1405                                  const char *rename_branch)
1406 {
1407         /*
1408          * Handle file adds that need to be renamed due to directory rename
1409          * detection.  This differs from handle_rename_normal, because
1410          * there is no content merge to do; just move the file into the
1411          * desired final location.
1412          */
1413         const struct diff_filespec *dest = pair->two;
1414
1415         if (!opt->call_depth && would_lose_untracked(opt, dest->path)) {
1416                 char *alt_path = unique_path(opt, dest->path, rename_branch);
1417
1418                 output(opt, 1, _("Error: Refusing to lose untracked file at %s; "
1419                                "writing to %s instead."),
1420                        dest->path, alt_path);
1421                 /*
1422                  * Write the file in worktree at alt_path, but not in the
1423                  * index.  Instead, write to dest->path for the index but
1424                  * only at the higher appropriate stage.
1425                  */
1426                 if (update_file(opt, 0, &dest->oid, dest->mode, alt_path))
1427                         return -1;
1428                 free(alt_path);
1429                 return update_stages(opt, dest->path, NULL,
1430                                      rename_branch == opt->branch1 ? dest : NULL,
1431                                      rename_branch == opt->branch1 ? NULL : dest);
1432         }
1433
1434         /* Update dest->path both in index and in worktree */
1435         if (update_file(opt, 1, &dest->oid, dest->mode, dest->path))
1436                 return -1;
1437         return 0;
1438 }
1439
1440 static int handle_change_delete(struct merge_options *opt,
1441                                 const char *path, const char *old_path,
1442                                 const struct object_id *o_oid, int o_mode,
1443                                 const struct object_id *changed_oid,
1444                                 int changed_mode,
1445                                 const char *change_branch,
1446                                 const char *delete_branch,
1447                                 const char *change, const char *change_past)
1448 {
1449         char *alt_path = NULL;
1450         const char *update_path = path;
1451         int ret = 0;
1452
1453         if (dir_in_way(opt->repo->index, path, !opt->call_depth, 0) ||
1454             (!opt->call_depth && would_lose_untracked(opt, path))) {
1455                 update_path = alt_path = unique_path(opt, path, change_branch);
1456         }
1457
1458         if (opt->call_depth) {
1459                 /*
1460                  * We cannot arbitrarily accept either a_sha or b_sha as
1461                  * correct; since there is no true "middle point" between
1462                  * them, simply reuse the base version for virtual merge base.
1463                  */
1464                 ret = remove_file_from_index(opt->repo->index, path);
1465                 if (!ret)
1466                         ret = update_file(opt, 0, o_oid, o_mode, update_path);
1467         } else {
1468                 /*
1469                  * Despite the four nearly duplicate messages and argument
1470                  * lists below and the ugliness of the nested if-statements,
1471                  * having complete messages makes the job easier for
1472                  * translators.
1473                  *
1474                  * The slight variance among the cases is due to the fact
1475                  * that:
1476                  *   1) directory/file conflicts (in effect if
1477                  *      !alt_path) could cause us to need to write the
1478                  *      file to a different path.
1479                  *   2) renames (in effect if !old_path) could mean that
1480                  *      there are two names for the path that the user
1481                  *      may know the file by.
1482                  */
1483                 if (!alt_path) {
1484                         if (!old_path) {
1485                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1486                                        "and %s in %s. Version %s of %s left in tree."),
1487                                        change, path, delete_branch, change_past,
1488                                        change_branch, change_branch, path);
1489                         } else {
1490                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1491                                        "and %s to %s in %s. Version %s of %s left in tree."),
1492                                        change, old_path, delete_branch, change_past, path,
1493                                        change_branch, change_branch, path);
1494                         }
1495                 } else {
1496                         if (!old_path) {
1497                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1498                                        "and %s in %s. Version %s of %s left in tree at %s."),
1499                                        change, path, delete_branch, change_past,
1500                                        change_branch, change_branch, path, alt_path);
1501                         } else {
1502                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1503                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1504                                        change, old_path, delete_branch, change_past, path,
1505                                        change_branch, change_branch, path, alt_path);
1506                         }
1507                 }
1508                 /*
1509                  * No need to call update_file() on path when change_branch ==
1510                  * opt->branch1 && !alt_path, since that would needlessly touch
1511                  * path.  We could call update_file_flags() with update_cache=0
1512                  * and update_wd=0, but that's a no-op.
1513                  */
1514                 if (change_branch != opt->branch1 || alt_path)
1515                         ret = update_file(opt, 0, changed_oid, changed_mode, update_path);
1516         }
1517         free(alt_path);
1518
1519         return ret;
1520 }
1521
1522 static int handle_rename_delete(struct merge_options *opt,
1523                                 struct diff_filepair *pair,
1524                                 const char *rename_branch,
1525                                 const char *delete_branch)
1526 {
1527         const struct diff_filespec *orig = pair->one;
1528         const struct diff_filespec *dest = pair->two;
1529
1530         if (handle_change_delete(opt,
1531                                  opt->call_depth ? orig->path : dest->path,
1532                                  opt->call_depth ? NULL : orig->path,
1533                                  &orig->oid, orig->mode,
1534                                  &dest->oid, dest->mode,
1535                                  rename_branch, delete_branch,
1536                                  _("rename"), _("renamed")))
1537                 return -1;
1538
1539         if (opt->call_depth)
1540                 return remove_file_from_index(opt->repo->index, dest->path);
1541         else
1542                 return update_stages(opt, dest->path, NULL,
1543                                      rename_branch == opt->branch1 ? dest : NULL,
1544                                      rename_branch == opt->branch1 ? NULL : dest);
1545 }
1546
1547 static struct diff_filespec *filespec_from_entry(struct diff_filespec *target,
1548                                                  struct stage_data *entry,
1549                                                  int stage)
1550 {
1551         struct object_id *oid = &entry->stages[stage].oid;
1552         unsigned mode = entry->stages[stage].mode;
1553         if (mode == 0 || is_null_oid(oid))
1554                 return NULL;
1555         oidcpy(&target->oid, oid);
1556         target->mode = mode;
1557         return target;
1558 }
1559
1560 static int handle_file_collision(struct merge_options *opt,
1561                                  const char *collide_path,
1562                                  const char *prev_path1,
1563                                  const char *prev_path2,
1564                                  const char *branch1, const char *branch2,
1565                                  const struct object_id *a_oid,
1566                                  unsigned int a_mode,
1567                                  const struct object_id *b_oid,
1568                                  unsigned int b_mode)
1569 {
1570         struct merge_file_info mfi;
1571         struct diff_filespec null, a, b;
1572         char *alt_path = NULL;
1573         const char *update_path = collide_path;
1574
1575         /*
1576          * It's easiest to get the correct things into stage 2 and 3, and
1577          * to make sure that the content merge puts HEAD before the other
1578          * branch if we just ensure that branch1 == opt->branch1.  So, simply
1579          * flip arguments around if we don't have that.
1580          */
1581         if (branch1 != opt->branch1) {
1582                 return handle_file_collision(opt, collide_path,
1583                                              prev_path2, prev_path1,
1584                                              branch2, branch1,
1585                                              b_oid, b_mode,
1586                                              a_oid, a_mode);
1587         }
1588
1589         /*
1590          * In the recursive case, we just opt to undo renames
1591          */
1592         if (opt->call_depth && (prev_path1 || prev_path2)) {
1593                 /* Put first file (a_oid, a_mode) in its original spot */
1594                 if (prev_path1) {
1595                         if (update_file(opt, 1, a_oid, a_mode, prev_path1))
1596                                 return -1;
1597                 } else {
1598                         if (update_file(opt, 1, a_oid, a_mode, collide_path))
1599                                 return -1;
1600                 }
1601
1602                 /* Put second file (b_oid, b_mode) in its original spot */
1603                 if (prev_path2) {
1604                         if (update_file(opt, 1, b_oid, b_mode, prev_path2))
1605                                 return -1;
1606                 } else {
1607                         if (update_file(opt, 1, b_oid, b_mode, collide_path))
1608                                 return -1;
1609                 }
1610
1611                 /* Don't leave something at collision path if unrenaming both */
1612                 if (prev_path1 && prev_path2)
1613                         remove_file(opt, 1, collide_path, 0);
1614
1615                 return 0;
1616         }
1617
1618         /* Remove rename sources if rename/add or rename/rename(2to1) */
1619         if (prev_path1)
1620                 remove_file(opt, 1, prev_path1,
1621                             opt->call_depth || would_lose_untracked(opt, prev_path1));
1622         if (prev_path2)
1623                 remove_file(opt, 1, prev_path2,
1624                             opt->call_depth || would_lose_untracked(opt, prev_path2));
1625
1626         /*
1627          * Remove the collision path, if it wouldn't cause dirty contents
1628          * or an untracked file to get lost.  We'll either overwrite with
1629          * merged contents, or just write out to differently named files.
1630          */
1631         if (was_dirty(opt, collide_path)) {
1632                 output(opt, 1, _("Refusing to lose dirty file at %s"),
1633                        collide_path);
1634                 update_path = alt_path = unique_path(opt, collide_path, "merged");
1635         } else if (would_lose_untracked(opt, collide_path)) {
1636                 /*
1637                  * Only way we get here is if both renames were from
1638                  * a directory rename AND user had an untracked file
1639                  * at the location where both files end up after the
1640                  * two directory renames.  See testcase 10d of t6043.
1641                  */
1642                 output(opt, 1, _("Refusing to lose untracked file at "
1643                                "%s, even though it's in the way."),
1644                        collide_path);
1645                 update_path = alt_path = unique_path(opt, collide_path, "merged");
1646         } else {
1647                 /*
1648                  * FIXME: It's possible that the two files are identical
1649                  * and that the current working copy happens to match, in
1650                  * which case we are unnecessarily touching the working
1651                  * tree file.  It's not a likely enough scenario that I
1652                  * want to code up the checks for it and a better fix is
1653                  * available if we restructure how unpack_trees() and
1654                  * merge-recursive interoperate anyway, so punting for
1655                  * now...
1656                  */
1657                 remove_file(opt, 0, collide_path, 0);
1658         }
1659
1660         /* Store things in diff_filespecs for functions that need it */
1661         memset(&a, 0, sizeof(struct diff_filespec));
1662         memset(&b, 0, sizeof(struct diff_filespec));
1663         null.path = a.path = b.path = (char *)collide_path;
1664         oidcpy(&null.oid, &null_oid);
1665         null.mode = 0;
1666         oidcpy(&a.oid, a_oid);
1667         a.mode = a_mode;
1668         a.oid_valid = 1;
1669         oidcpy(&b.oid, b_oid);
1670         b.mode = b_mode;
1671         b.oid_valid = 1;
1672
1673         if (merge_mode_and_contents(opt, &null, &a, &b, collide_path,
1674                                     branch1, branch2, opt->call_depth * 2, &mfi))
1675                 return -1;
1676         mfi.clean &= !alt_path;
1677         if (update_file(opt, mfi.clean, &mfi.oid, mfi.mode, update_path))
1678                 return -1;
1679         if (!mfi.clean && !opt->call_depth &&
1680             update_stages(opt, collide_path, NULL, &a, &b))
1681                 return -1;
1682         free(alt_path);
1683         /*
1684          * FIXME: If both a & b both started with conflicts (only possible
1685          * if they came from a rename/rename(2to1)), but had IDENTICAL
1686          * contents including those conflicts, then in the next line we claim
1687          * it was clean.  If someone cares about this case, we should have the
1688          * caller notify us if we started with conflicts.
1689          */
1690         return mfi.clean;
1691 }
1692
1693 static int handle_rename_add(struct merge_options *opt,
1694                              struct rename_conflict_info *ci)
1695 {
1696         /* a was renamed to c, and a separate c was added. */
1697         struct diff_filespec *a = ci->pair1->one;
1698         struct diff_filespec *c = ci->pair1->two;
1699         char *path = c->path;
1700         char *prev_path_desc;
1701         struct merge_file_info mfi;
1702
1703         int other_stage = (ci->branch1 == opt->branch1 ? 3 : 2);
1704
1705         output(opt, 1, _("CONFLICT (rename/add): "
1706                "Rename %s->%s in %s.  Added %s in %s"),
1707                a->path, c->path, ci->branch1,
1708                c->path, ci->branch2);
1709
1710         prev_path_desc = xstrfmt("version of %s from %s", path, a->path);
1711         if (merge_mode_and_contents(opt, a, c, &ci->ren1_other, prev_path_desc,
1712                                     opt->branch1, opt->branch2,
1713                                     1 + opt->call_depth * 2, &mfi))
1714                 return -1;
1715         free(prev_path_desc);
1716
1717         return handle_file_collision(opt,
1718                                      c->path, a->path, NULL,
1719                                      ci->branch1, ci->branch2,
1720                                      &mfi.oid, mfi.mode,
1721                                      &ci->dst_entry1->stages[other_stage].oid,
1722                                      ci->dst_entry1->stages[other_stage].mode);
1723 }
1724
1725 static char *find_path_for_conflict(struct merge_options *opt,
1726                                     const char *path,
1727                                     const char *branch1,
1728                                     const char *branch2)
1729 {
1730         char *new_path = NULL;
1731         if (dir_in_way(opt->repo->index, path, !opt->call_depth, 0)) {
1732                 new_path = unique_path(opt, path, branch1);
1733                 output(opt, 1, _("%s is a directory in %s adding "
1734                                "as %s instead"),
1735                        path, branch2, new_path);
1736         } else if (would_lose_untracked(opt, path)) {
1737                 new_path = unique_path(opt, path, branch1);
1738                 output(opt, 1, _("Refusing to lose untracked file"
1739                                " at %s; adding as %s instead"),
1740                        path, new_path);
1741         }
1742
1743         return new_path;
1744 }
1745
1746 static int handle_rename_rename_1to2(struct merge_options *opt,
1747                                      struct rename_conflict_info *ci)
1748 {
1749         /* One file was renamed in both branches, but to different names. */
1750         struct merge_file_info mfi;
1751         struct diff_filespec other;
1752         struct diff_filespec *add;
1753         struct diff_filespec *o = ci->pair1->one;
1754         struct diff_filespec *a = ci->pair1->two;
1755         struct diff_filespec *b = ci->pair2->two;
1756         char *path_desc;
1757
1758         output(opt, 1, _("CONFLICT (rename/rename): "
1759                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1760                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1761                o->path, a->path, ci->branch1,
1762                o->path, b->path, ci->branch2,
1763                opt->call_depth ? _(" (left unresolved)") : "");
1764
1765         path_desc = xstrfmt("%s and %s, both renamed from %s",
1766                             a->path, b->path, o->path);
1767         if (merge_mode_and_contents(opt, o, a, b, path_desc,
1768                                     ci->branch1, ci->branch2,
1769                                     opt->call_depth * 2, &mfi))
1770                 return -1;
1771         free(path_desc);
1772
1773         if (opt->call_depth) {
1774                 /*
1775                  * FIXME: For rename/add-source conflicts (if we could detect
1776                  * such), this is wrong.  We should instead find a unique
1777                  * pathname and then either rename the add-source file to that
1778                  * unique path, or use that unique path instead of src here.
1779                  */
1780                 if (update_file(opt, 0, &mfi.oid, mfi.mode, o->path))
1781                         return -1;
1782
1783                 /*
1784                  * Above, we put the merged content at the merge-base's
1785                  * path.  Now we usually need to delete both a->path and
1786                  * b->path.  However, the rename on each side of the merge
1787                  * could also be involved in a rename/add conflict.  In
1788                  * such cases, we should keep the added file around,
1789                  * resolving the conflict at that path in its favor.
1790                  */
1791                 add = filespec_from_entry(&other, ci->dst_entry1, 2 ^ 1);
1792                 if (add) {
1793                         if (update_file(opt, 0, &add->oid, add->mode, a->path))
1794                                 return -1;
1795                 }
1796                 else
1797                         remove_file_from_index(opt->repo->index, a->path);
1798                 add = filespec_from_entry(&other, ci->dst_entry2, 3 ^ 1);
1799                 if (add) {
1800                         if (update_file(opt, 0, &add->oid, add->mode, b->path))
1801                                 return -1;
1802                 }
1803                 else
1804                         remove_file_from_index(opt->repo->index, b->path);
1805         } else {
1806                 /*
1807                  * For each destination path, we need to see if there is a
1808                  * rename/add collision.  If not, we can write the file out
1809                  * to the specified location.
1810                  */
1811                 add = filespec_from_entry(&other, ci->dst_entry1, 2 ^ 1);
1812                 if (add) {
1813                         if (handle_file_collision(opt, a->path,
1814                                                   NULL, NULL,
1815                                                   ci->branch1, ci->branch2,
1816                                                   &mfi.oid, mfi.mode,
1817                                                   &add->oid, add->mode) < 0)
1818                                 return -1;
1819                 } else {
1820                         char *new_path = find_path_for_conflict(opt, a->path,
1821                                                                 ci->branch1,
1822                                                                 ci->branch2);
1823                         if (update_file(opt, 0, &mfi.oid, mfi.mode, new_path ? new_path : a->path))
1824                                 return -1;
1825                         free(new_path);
1826                         if (update_stages(opt, a->path, NULL, a, NULL))
1827                                 return -1;
1828                 }
1829
1830                 add = filespec_from_entry(&other, ci->dst_entry2, 3 ^ 1);
1831                 if (add) {
1832                         if (handle_file_collision(opt, b->path,
1833                                                   NULL, NULL,
1834                                                   ci->branch1, ci->branch2,
1835                                                   &add->oid, add->mode,
1836                                                   &mfi.oid, mfi.mode) < 0)
1837                                 return -1;
1838                 } else {
1839                         char *new_path = find_path_for_conflict(opt, b->path,
1840                                                                 ci->branch2,
1841                                                                 ci->branch1);
1842                         if (update_file(opt, 0, &mfi.oid, mfi.mode, new_path ? new_path : b->path))
1843                                 return -1;
1844                         free(new_path);
1845                         if (update_stages(opt, b->path, NULL, NULL, b))
1846                                 return -1;
1847                 }
1848         }
1849
1850         return 0;
1851 }
1852
1853 static int handle_rename_rename_2to1(struct merge_options *opt,
1854                                      struct rename_conflict_info *ci)
1855 {
1856         /* Two files, a & b, were renamed to the same thing, c. */
1857         struct diff_filespec *a = ci->pair1->one;
1858         struct diff_filespec *b = ci->pair2->one;
1859         struct diff_filespec *c1 = ci->pair1->two;
1860         struct diff_filespec *c2 = ci->pair2->two;
1861         char *path = c1->path; /* == c2->path */
1862         char *path_side_1_desc;
1863         char *path_side_2_desc;
1864         struct merge_file_info mfi_c1;
1865         struct merge_file_info mfi_c2;
1866
1867         output(opt, 1, _("CONFLICT (rename/rename): "
1868                "Rename %s->%s in %s. "
1869                "Rename %s->%s in %s"),
1870                a->path, c1->path, ci->branch1,
1871                b->path, c2->path, ci->branch2);
1872
1873         path_side_1_desc = xstrfmt("version of %s from %s", path, a->path);
1874         path_side_2_desc = xstrfmt("version of %s from %s", path, b->path);
1875         if (merge_mode_and_contents(opt, a, c1, &ci->ren1_other, path_side_1_desc,
1876                                     opt->branch1, opt->branch2,
1877                                     1 + opt->call_depth * 2, &mfi_c1) ||
1878             merge_mode_and_contents(opt, b, &ci->ren2_other, c2, path_side_2_desc,
1879                                     opt->branch1, opt->branch2,
1880                                     1 + opt->call_depth * 2, &mfi_c2))
1881                 return -1;
1882         free(path_side_1_desc);
1883         free(path_side_2_desc);
1884
1885         return handle_file_collision(opt, path, a->path, b->path,
1886                                      ci->branch1, ci->branch2,
1887                                      &mfi_c1.oid, mfi_c1.mode,
1888                                      &mfi_c2.oid, mfi_c2.mode);
1889 }
1890
1891 /*
1892  * Get the diff_filepairs changed between o_tree and tree.
1893  */
1894 static struct diff_queue_struct *get_diffpairs(struct merge_options *opt,
1895                                                struct tree *o_tree,
1896                                                struct tree *tree)
1897 {
1898         struct diff_queue_struct *ret;
1899         struct diff_options opts;
1900
1901         repo_diff_setup(opt->repo, &opts);
1902         opts.flags.recursive = 1;
1903         opts.flags.rename_empty = 0;
1904         opts.detect_rename = merge_detect_rename(opt);
1905         /*
1906          * We do not have logic to handle the detection of copies.  In
1907          * fact, it may not even make sense to add such logic: would we
1908          * really want a change to a base file to be propagated through
1909          * multiple other files by a merge?
1910          */
1911         if (opts.detect_rename > DIFF_DETECT_RENAME)
1912                 opts.detect_rename = DIFF_DETECT_RENAME;
1913         opts.rename_limit = opt->merge_rename_limit >= 0 ? opt->merge_rename_limit :
1914                             opt->diff_rename_limit >= 0 ? opt->diff_rename_limit :
1915                             1000;
1916         opts.rename_score = opt->rename_score;
1917         opts.show_rename_progress = opt->show_rename_progress;
1918         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1919         diff_setup_done(&opts);
1920         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1921         diffcore_std(&opts);
1922         if (opts.needed_rename_limit > opt->needed_rename_limit)
1923                 opt->needed_rename_limit = opts.needed_rename_limit;
1924
1925         ret = xmalloc(sizeof(*ret));
1926         *ret = diff_queued_diff;
1927
1928         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1929         diff_queued_diff.nr = 0;
1930         diff_queued_diff.queue = NULL;
1931         diff_flush(&opts);
1932         return ret;
1933 }
1934
1935 static int tree_has_path(struct tree *tree, const char *path)
1936 {
1937         struct object_id hashy;
1938         unsigned short mode_o;
1939
1940         return !get_tree_entry(&tree->object.oid, path,
1941                                &hashy, &mode_o);
1942 }
1943
1944 /*
1945  * Return a new string that replaces the beginning portion (which matches
1946  * entry->dir), with entry->new_dir.  In perl-speak:
1947  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1948  * NOTE:
1949  *   Caller must ensure that old_path starts with entry->dir + '/'.
1950  */
1951 static char *apply_dir_rename(struct dir_rename_entry *entry,
1952                               const char *old_path)
1953 {
1954         struct strbuf new_path = STRBUF_INIT;
1955         int oldlen, newlen;
1956
1957         if (entry->non_unique_new_dir)
1958                 return NULL;
1959
1960         oldlen = strlen(entry->dir);
1961         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1962         strbuf_grow(&new_path, newlen);
1963         strbuf_addbuf(&new_path, &entry->new_dir);
1964         strbuf_addstr(&new_path, &old_path[oldlen]);
1965
1966         return strbuf_detach(&new_path, NULL);
1967 }
1968
1969 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1970                                     char **old_dir, char **new_dir)
1971 {
1972         char *end_of_old, *end_of_new;
1973         int old_len, new_len;
1974
1975         *old_dir = NULL;
1976         *new_dir = NULL;
1977
1978         /*
1979          * For
1980          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1981          * the "e/foo.c" part is the same, we just want to know that
1982          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1983          * so, for this example, this function returns "a/b/c/d" in
1984          * *old_dir and "a/b/some/thing/else" in *new_dir.
1985          *
1986          * Also, if the basename of the file changed, we don't care.  We
1987          * want to know which portion of the directory, if any, changed.
1988          */
1989         end_of_old = strrchr(old_path, '/');
1990         end_of_new = strrchr(new_path, '/');
1991
1992         if (end_of_old == NULL || end_of_new == NULL)
1993                 return;
1994         while (*--end_of_new == *--end_of_old &&
1995                end_of_old != old_path &&
1996                end_of_new != new_path)
1997                 ; /* Do nothing; all in the while loop */
1998         /*
1999          * We've found the first non-matching character in the directory
2000          * paths.  That means the current directory we were comparing
2001          * represents the rename.  Move end_of_old and end_of_new back
2002          * to the full directory name.
2003          */
2004         if (*end_of_old == '/')
2005                 end_of_old++;
2006         if (*end_of_old != '/')
2007                 end_of_new++;
2008         end_of_old = strchr(end_of_old, '/');
2009         end_of_new = strchr(end_of_new, '/');
2010
2011         /*
2012          * It may have been the case that old_path and new_path were the same
2013          * directory all along.  Don't claim a rename if they're the same.
2014          */
2015         old_len = end_of_old - old_path;
2016         new_len = end_of_new - new_path;
2017
2018         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
2019                 *old_dir = xstrndup(old_path, old_len);
2020                 *new_dir = xstrndup(new_path, new_len);
2021         }
2022 }
2023
2024 static void remove_hashmap_entries(struct hashmap *dir_renames,
2025                                    struct string_list *items_to_remove)
2026 {
2027         int i;
2028         struct dir_rename_entry *entry;
2029
2030         for (i = 0; i < items_to_remove->nr; i++) {
2031                 entry = items_to_remove->items[i].util;
2032                 hashmap_remove(dir_renames, entry, NULL);
2033         }
2034         string_list_clear(items_to_remove, 0);
2035 }
2036
2037 /*
2038  * See if there is a directory rename for path, and if there are any file
2039  * level conflicts for the renamed location.  If there is a rename and
2040  * there are no conflicts, return the new name.  Otherwise, return NULL.
2041  */
2042 static char *handle_path_level_conflicts(struct merge_options *opt,
2043                                          const char *path,
2044                                          struct dir_rename_entry *entry,
2045                                          struct hashmap *collisions,
2046                                          struct tree *tree)
2047 {
2048         char *new_path = NULL;
2049         struct collision_entry *collision_ent;
2050         int clean = 1;
2051         struct strbuf collision_paths = STRBUF_INIT;
2052
2053         /*
2054          * entry has the mapping of old directory name to new directory name
2055          * that we want to apply to path.
2056          */
2057         new_path = apply_dir_rename(entry, path);
2058
2059         if (!new_path) {
2060                 /* This should only happen when entry->non_unique_new_dir set */
2061                 if (!entry->non_unique_new_dir)
2062                         BUG("entry->non_unqiue_dir not set and !new_path");
2063                 output(opt, 1, _("CONFLICT (directory rename split): "
2064                                "Unclear where to place %s because directory "
2065                                "%s was renamed to multiple other directories, "
2066                                "with no destination getting a majority of the "
2067                                "files."),
2068                        path, entry->dir);
2069                 clean = 0;
2070                 return NULL;
2071         }
2072
2073         /*
2074          * The caller needs to have ensured that it has pre-populated
2075          * collisions with all paths that map to new_path.  Do a quick check
2076          * to ensure that's the case.
2077          */
2078         collision_ent = collision_find_entry(collisions, new_path);
2079         if (collision_ent == NULL)
2080                 BUG("collision_ent is NULL");
2081
2082         /*
2083          * Check for one-sided add/add/.../add conflicts, i.e.
2084          * where implicit renames from the other side doing
2085          * directory rename(s) can affect this side of history
2086          * to put multiple paths into the same location.  Warn
2087          * and bail on directory renames for such paths.
2088          */
2089         if (collision_ent->reported_already) {
2090                 clean = 0;
2091         } else if (tree_has_path(tree, new_path)) {
2092                 collision_ent->reported_already = 1;
2093                 strbuf_add_separated_string_list(&collision_paths, ", ",
2094                                                  &collision_ent->source_files);
2095                 output(opt, 1, _("CONFLICT (implicit dir rename): Existing "
2096                                "file/dir at %s in the way of implicit "
2097                                "directory rename(s) putting the following "
2098                                "path(s) there: %s."),
2099                        new_path, collision_paths.buf);
2100                 clean = 0;
2101         } else if (collision_ent->source_files.nr > 1) {
2102                 collision_ent->reported_already = 1;
2103                 strbuf_add_separated_string_list(&collision_paths, ", ",
2104                                                  &collision_ent->source_files);
2105                 output(opt, 1, _("CONFLICT (implicit dir rename): Cannot map "
2106                                "more than one path to %s; implicit directory "
2107                                "renames tried to put these paths there: %s"),
2108                        new_path, collision_paths.buf);
2109                 clean = 0;
2110         }
2111
2112         /* Free memory we no longer need */
2113         strbuf_release(&collision_paths);
2114         if (!clean && new_path) {
2115                 free(new_path);
2116                 return NULL;
2117         }
2118
2119         return new_path;
2120 }
2121
2122 /*
2123  * There are a couple things we want to do at the directory level:
2124  *   1. Check for both sides renaming to the same thing, in order to avoid
2125  *      implicit renaming of files that should be left in place.  (See
2126  *      testcase 6b in t6043 for details.)
2127  *   2. Prune directory renames if there are still files left in the
2128  *      the original directory.  These represent a partial directory rename,
2129  *      i.e. a rename where only some of the files within the directory
2130  *      were renamed elsewhere.  (Technically, this could be done earlier
2131  *      in get_directory_renames(), except that would prevent us from
2132  *      doing the previous check and thus failing testcase 6b.)
2133  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2134  *      In the future, we could potentially record this info as well and
2135  *      omit reporting rename/rename(1to2) conflicts for each path within
2136  *      the affected directories, thus cleaning up the merge output.
2137  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2138  *         directory level, because merging directories is fine.  If it
2139  *         causes conflicts for files within those merged directories, then
2140  *         that should be detected at the individual path level.
2141  */
2142 static void handle_directory_level_conflicts(struct merge_options *opt,
2143                                              struct hashmap *dir_re_head,
2144                                              struct tree *head,
2145                                              struct hashmap *dir_re_merge,
2146                                              struct tree *merge)
2147 {
2148         struct hashmap_iter iter;
2149         struct dir_rename_entry *head_ent;
2150         struct dir_rename_entry *merge_ent;
2151
2152         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2153         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2154
2155         hashmap_iter_init(dir_re_head, &iter);
2156         while ((head_ent = hashmap_iter_next(&iter))) {
2157                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2158                 if (merge_ent &&
2159                     !head_ent->non_unique_new_dir &&
2160                     !merge_ent->non_unique_new_dir &&
2161                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2162                         /* 1. Renamed identically; remove it from both sides */
2163                         string_list_append(&remove_from_head,
2164                                            head_ent->dir)->util = head_ent;
2165                         strbuf_release(&head_ent->new_dir);
2166                         string_list_append(&remove_from_merge,
2167                                            merge_ent->dir)->util = merge_ent;
2168                         strbuf_release(&merge_ent->new_dir);
2169                 } else if (tree_has_path(head, head_ent->dir)) {
2170                         /* 2. This wasn't a directory rename after all */
2171                         string_list_append(&remove_from_head,
2172                                            head_ent->dir)->util = head_ent;
2173                         strbuf_release(&head_ent->new_dir);
2174                 }
2175         }
2176
2177         remove_hashmap_entries(dir_re_head, &remove_from_head);
2178         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2179
2180         hashmap_iter_init(dir_re_merge, &iter);
2181         while ((merge_ent = hashmap_iter_next(&iter))) {
2182                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2183                 if (tree_has_path(merge, merge_ent->dir)) {
2184                         /* 2. This wasn't a directory rename after all */
2185                         string_list_append(&remove_from_merge,
2186                                            merge_ent->dir)->util = merge_ent;
2187                 } else if (head_ent &&
2188                            !head_ent->non_unique_new_dir &&
2189                            !merge_ent->non_unique_new_dir) {
2190                         /* 3. rename/rename(1to2) */
2191                         /*
2192                          * We can assume it's not rename/rename(1to1) because
2193                          * that was case (1), already checked above.  So we
2194                          * know that head_ent->new_dir and merge_ent->new_dir
2195                          * are different strings.
2196                          */
2197                         output(opt, 1, _("CONFLICT (rename/rename): "
2198                                        "Rename directory %s->%s in %s. "
2199                                        "Rename directory %s->%s in %s"),
2200                                head_ent->dir, head_ent->new_dir.buf, opt->branch1,
2201                                head_ent->dir, merge_ent->new_dir.buf, opt->branch2);
2202                         string_list_append(&remove_from_head,
2203                                            head_ent->dir)->util = head_ent;
2204                         strbuf_release(&head_ent->new_dir);
2205                         string_list_append(&remove_from_merge,
2206                                            merge_ent->dir)->util = merge_ent;
2207                         strbuf_release(&merge_ent->new_dir);
2208                 }
2209         }
2210
2211         remove_hashmap_entries(dir_re_head, &remove_from_head);
2212         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2213 }
2214
2215 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs)
2216 {
2217         struct hashmap *dir_renames;
2218         struct hashmap_iter iter;
2219         struct dir_rename_entry *entry;
2220         int i;
2221
2222         /*
2223          * Typically, we think of a directory rename as all files from a
2224          * certain directory being moved to a target directory.  However,
2225          * what if someone first moved two files from the original
2226          * directory in one commit, and then renamed the directory
2227          * somewhere else in a later commit?  At merge time, we just know
2228          * that files from the original directory went to two different
2229          * places, and that the bulk of them ended up in the same place.
2230          * We want each directory rename to represent where the bulk of the
2231          * files from that directory end up; this function exists to find
2232          * where the bulk of the files went.
2233          *
2234          * The first loop below simply iterates through the list of file
2235          * renames, finding out how often each directory rename pair
2236          * possibility occurs.
2237          */
2238         dir_renames = xmalloc(sizeof(*dir_renames));
2239         dir_rename_init(dir_renames);
2240         for (i = 0; i < pairs->nr; ++i) {
2241                 struct string_list_item *item;
2242                 int *count;
2243                 struct diff_filepair *pair = pairs->queue[i];
2244                 char *old_dir, *new_dir;
2245
2246                 /* File not part of directory rename if it wasn't renamed */
2247                 if (pair->status != 'R')
2248                         continue;
2249
2250                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2251                                         &old_dir,        &new_dir);
2252                 if (!old_dir)
2253                         /* Directory didn't change at all; ignore this one. */
2254                         continue;
2255
2256                 entry = dir_rename_find_entry(dir_renames, old_dir);
2257                 if (!entry) {
2258                         entry = xmalloc(sizeof(*entry));
2259                         dir_rename_entry_init(entry, old_dir);
2260                         hashmap_put(dir_renames, entry);
2261                 } else {
2262                         free(old_dir);
2263                 }
2264                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2265                 if (!item) {
2266                         item = string_list_insert(&entry->possible_new_dirs,
2267                                                   new_dir);
2268                         item->util = xcalloc(1, sizeof(int));
2269                 } else {
2270                         free(new_dir);
2271                 }
2272                 count = item->util;
2273                 *count += 1;
2274         }
2275
2276         /*
2277          * For each directory with files moved out of it, we find out which
2278          * target directory received the most files so we can declare it to
2279          * be the "winning" target location for the directory rename.  This
2280          * winner gets recorded in new_dir.  If there is no winner
2281          * (multiple target directories received the same number of files),
2282          * we set non_unique_new_dir.  Once we've determined the winner (or
2283          * that there is no winner), we no longer need possible_new_dirs.
2284          */
2285         hashmap_iter_init(dir_renames, &iter);
2286         while ((entry = hashmap_iter_next(&iter))) {
2287                 int max = 0;
2288                 int bad_max = 0;
2289                 char *best = NULL;
2290
2291                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2292                         int *count = entry->possible_new_dirs.items[i].util;
2293
2294                         if (*count == max)
2295                                 bad_max = max;
2296                         else if (*count > max) {
2297                                 max = *count;
2298                                 best = entry->possible_new_dirs.items[i].string;
2299                         }
2300                 }
2301                 if (bad_max == max)
2302                         entry->non_unique_new_dir = 1;
2303                 else {
2304                         assert(entry->new_dir.len == 0);
2305                         strbuf_addstr(&entry->new_dir, best);
2306                 }
2307                 /*
2308                  * The relevant directory sub-portion of the original full
2309                  * filepaths were xstrndup'ed before inserting into
2310                  * possible_new_dirs, and instead of manually iterating the
2311                  * list and free'ing each, just lie and tell
2312                  * possible_new_dirs that it did the strdup'ing so that it
2313                  * will free them for us.
2314                  */
2315                 entry->possible_new_dirs.strdup_strings = 1;
2316                 string_list_clear(&entry->possible_new_dirs, 1);
2317         }
2318
2319         return dir_renames;
2320 }
2321
2322 static struct dir_rename_entry *check_dir_renamed(const char *path,
2323                                                   struct hashmap *dir_renames)
2324 {
2325         char *temp = xstrdup(path);
2326         char *end;
2327         struct dir_rename_entry *entry = NULL;
2328
2329         while ((end = strrchr(temp, '/'))) {
2330                 *end = '\0';
2331                 entry = dir_rename_find_entry(dir_renames, temp);
2332                 if (entry)
2333                         break;
2334         }
2335         free(temp);
2336         return entry;
2337 }
2338
2339 static void compute_collisions(struct hashmap *collisions,
2340                                struct hashmap *dir_renames,
2341                                struct diff_queue_struct *pairs)
2342 {
2343         int i;
2344
2345         /*
2346          * Multiple files can be mapped to the same path due to directory
2347          * renames done by the other side of history.  Since that other
2348          * side of history could have merged multiple directories into one,
2349          * if our side of history added the same file basename to each of
2350          * those directories, then all N of them would get implicitly
2351          * renamed by the directory rename detection into the same path,
2352          * and we'd get an add/add/.../add conflict, and all those adds
2353          * from *this* side of history.  This is not representable in the
2354          * index, and users aren't going to easily be able to make sense of
2355          * it.  So we need to provide a good warning about what's
2356          * happening, and fall back to no-directory-rename detection
2357          * behavior for those paths.
2358          *
2359          * See testcases 9e and all of section 5 from t6043 for examples.
2360          */
2361         collision_init(collisions);
2362
2363         for (i = 0; i < pairs->nr; ++i) {
2364                 struct dir_rename_entry *dir_rename_ent;
2365                 struct collision_entry *collision_ent;
2366                 char *new_path;
2367                 struct diff_filepair *pair = pairs->queue[i];
2368
2369                 if (pair->status != 'A' && pair->status != 'R')
2370                         continue;
2371                 dir_rename_ent = check_dir_renamed(pair->two->path,
2372                                                    dir_renames);
2373                 if (!dir_rename_ent)
2374                         continue;
2375
2376                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2377                 if (!new_path)
2378                         /*
2379                          * dir_rename_ent->non_unique_new_path is true, which
2380                          * means there is no directory rename for us to use,
2381                          * which means it won't cause us any additional
2382                          * collisions.
2383                          */
2384                         continue;
2385                 collision_ent = collision_find_entry(collisions, new_path);
2386                 if (!collision_ent) {
2387                         collision_ent = xcalloc(1,
2388                                                 sizeof(struct collision_entry));
2389                         hashmap_entry_init(collision_ent, strhash(new_path));
2390                         hashmap_put(collisions, collision_ent);
2391                         collision_ent->target_file = new_path;
2392                 } else {
2393                         free(new_path);
2394                 }
2395                 string_list_insert(&collision_ent->source_files,
2396                                    pair->two->path);
2397         }
2398 }
2399
2400 static char *check_for_directory_rename(struct merge_options *opt,
2401                                         const char *path,
2402                                         struct tree *tree,
2403                                         struct hashmap *dir_renames,
2404                                         struct hashmap *dir_rename_exclusions,
2405                                         struct hashmap *collisions,
2406                                         int *clean_merge)
2407 {
2408         char *new_path = NULL;
2409         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2410         struct dir_rename_entry *oentry = NULL;
2411
2412         if (!entry)
2413                 return new_path;
2414
2415         /*
2416          * This next part is a little weird.  We do not want to do an
2417          * implicit rename into a directory we renamed on our side, because
2418          * that will result in a spurious rename/rename(1to2) conflict.  An
2419          * example:
2420          *   Base commit: dumbdir/afile, otherdir/bfile
2421          *   Side 1:      smrtdir/afile, otherdir/bfile
2422          *   Side 2:      dumbdir/afile, dumbdir/bfile
2423          * Here, while working on Side 1, we could notice that otherdir was
2424          * renamed/merged to dumbdir, and change the diff_filepair for
2425          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2426          * 2 will notice the rename from dumbdir to smrtdir, and do the
2427          * transitive rename to move it from dumbdir/bfile to
2428          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2429          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2430          * the file to end up in smrtdir.  And the way to achieve that is
2431          * to not let Side1 do the rename to dumbdir, since we know that is
2432          * the source of one of our directory renames.
2433          *
2434          * That's why oentry and dir_rename_exclusions is here.
2435          *
2436          * As it turns out, this also prevents N-way transient rename
2437          * confusion; See testcases 9c and 9d of t6043.
2438          */
2439         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2440         if (oentry) {
2441                 output(opt, 1, _("WARNING: Avoiding applying %s -> %s rename "
2442                                "to %s, because %s itself was renamed."),
2443                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2444         } else {
2445                 new_path = handle_path_level_conflicts(opt, path, entry,
2446                                                        collisions, tree);
2447                 *clean_merge &= (new_path != NULL);
2448         }
2449
2450         return new_path;
2451 }
2452
2453 static void apply_directory_rename_modifications(struct merge_options *opt,
2454                                                  struct diff_filepair *pair,
2455                                                  char *new_path,
2456                                                  struct rename *re,
2457                                                  struct tree *tree,
2458                                                  struct tree *o_tree,
2459                                                  struct tree *a_tree,
2460                                                  struct tree *b_tree,
2461                                                  struct string_list *entries)
2462 {
2463         struct string_list_item *item;
2464         int stage = (tree == a_tree ? 2 : 3);
2465         int update_wd;
2466
2467         /*
2468          * In all cases where we can do directory rename detection,
2469          * unpack_trees() will have read pair->two->path into the
2470          * index and the working copy.  We need to remove it so that
2471          * we can instead place it at new_path.  It is guaranteed to
2472          * not be untracked (unpack_trees() would have errored out
2473          * saying the file would have been overwritten), but it might
2474          * be dirty, though.
2475          */
2476         update_wd = !was_dirty(opt, pair->two->path);
2477         if (!update_wd)
2478                 output(opt, 1, _("Refusing to lose dirty file at %s"),
2479                        pair->two->path);
2480         remove_file(opt, 1, pair->two->path, !update_wd);
2481
2482         /* Find or create a new re->dst_entry */
2483         item = string_list_lookup(entries, new_path);
2484         if (item) {
2485                 /*
2486                  * Since we're renaming on this side of history, and it's
2487                  * due to a directory rename on the other side of history
2488                  * (which we only allow when the directory in question no
2489                  * longer exists on the other side of history), the
2490                  * original entry for re->dst_entry is no longer
2491                  * necessary...
2492                  */
2493                 re->dst_entry->processed = 1;
2494
2495                 /*
2496                  * ...because we'll be using this new one.
2497                  */
2498                 re->dst_entry = item->util;
2499         } else {
2500                 /*
2501                  * re->dst_entry is for the before-dir-rename path, and we
2502                  * need it to hold information for the after-dir-rename
2503                  * path.  Before creating a new entry, we need to mark the
2504                  * old one as unnecessary (...unless it is shared by
2505                  * src_entry, i.e. this didn't use to be a rename, in which
2506                  * case we can just allow the normal processing to happen
2507                  * for it).
2508                  */
2509                 if (pair->status == 'R')
2510                         re->dst_entry->processed = 1;
2511
2512                 re->dst_entry = insert_stage_data(new_path,
2513                                                   o_tree, a_tree, b_tree,
2514                                                   entries);
2515                 item = string_list_insert(entries, new_path);
2516                 item->util = re->dst_entry;
2517         }
2518
2519         /*
2520          * Update the stage_data with the information about the path we are
2521          * moving into place.  That slot will be empty and available for us
2522          * to write to because of the collision checks in
2523          * handle_path_level_conflicts().  In other words,
2524          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2525          * open for us to write to.
2526          *
2527          * It may be tempting to actually update the index at this point as
2528          * well, using update_stages_for_stage_data(), but as per the big
2529          * "NOTE" in update_stages(), doing so will modify the current
2530          * in-memory index which will break calls to would_lose_untracked()
2531          * that we need to make.  Instead, we need to just make sure that
2532          * the various handle_rename_*() functions update the index
2533          * explicitly rather than relying on unpack_trees() to have done it.
2534          */
2535         get_tree_entry(&tree->object.oid,
2536                        pair->two->path,
2537                        &re->dst_entry->stages[stage].oid,
2538                        &re->dst_entry->stages[stage].mode);
2539
2540         /* Update pair status */
2541         if (pair->status == 'A') {
2542                 /*
2543                  * Recording rename information for this add makes it look
2544                  * like a rename/delete conflict.  Make sure we can
2545                  * correctly handle this as an add that was moved to a new
2546                  * directory instead of reporting a rename/delete conflict.
2547                  */
2548                 re->add_turned_into_rename = 1;
2549         }
2550         /*
2551          * We don't actually look at pair->status again, but it seems
2552          * pedagogically correct to adjust it.
2553          */
2554         pair->status = 'R';
2555
2556         /*
2557          * Finally, record the new location.
2558          */
2559         pair->two->path = new_path;
2560 }
2561
2562 /*
2563  * Get information of all renames which occurred in 'pairs', making use of
2564  * any implicit directory renames inferred from the other side of history.
2565  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2566  * to be able to associate the correct cache entries with the rename
2567  * information; tree is always equal to either a_tree or b_tree.
2568  */
2569 static struct string_list *get_renames(struct merge_options *opt,
2570                                        struct diff_queue_struct *pairs,
2571                                        struct hashmap *dir_renames,
2572                                        struct hashmap *dir_rename_exclusions,
2573                                        struct tree *tree,
2574                                        struct tree *o_tree,
2575                                        struct tree *a_tree,
2576                                        struct tree *b_tree,
2577                                        struct string_list *entries,
2578                                        int *clean_merge)
2579 {
2580         int i;
2581         struct hashmap collisions;
2582         struct hashmap_iter iter;
2583         struct collision_entry *e;
2584         struct string_list *renames;
2585
2586         compute_collisions(&collisions, dir_renames, pairs);
2587         renames = xcalloc(1, sizeof(struct string_list));
2588
2589         for (i = 0; i < pairs->nr; ++i) {
2590                 struct string_list_item *item;
2591                 struct rename *re;
2592                 struct diff_filepair *pair = pairs->queue[i];
2593                 char *new_path; /* non-NULL only with directory renames */
2594
2595                 if (pair->status != 'A' && pair->status != 'R') {
2596                         diff_free_filepair(pair);
2597                         continue;
2598                 }
2599                 new_path = check_for_directory_rename(opt, pair->two->path, tree,
2600                                                       dir_renames,
2601                                                       dir_rename_exclusions,
2602                                                       &collisions,
2603                                                       clean_merge);
2604                 if (pair->status != 'R' && !new_path) {
2605                         diff_free_filepair(pair);
2606                         continue;
2607                 }
2608
2609                 re = xmalloc(sizeof(*re));
2610                 re->processed = 0;
2611                 re->add_turned_into_rename = 0;
2612                 re->pair = pair;
2613                 item = string_list_lookup(entries, re->pair->one->path);
2614                 if (!item)
2615                         re->src_entry = insert_stage_data(re->pair->one->path,
2616                                         o_tree, a_tree, b_tree, entries);
2617                 else
2618                         re->src_entry = item->util;
2619
2620                 item = string_list_lookup(entries, re->pair->two->path);
2621                 if (!item)
2622                         re->dst_entry = insert_stage_data(re->pair->two->path,
2623                                         o_tree, a_tree, b_tree, entries);
2624                 else
2625                         re->dst_entry = item->util;
2626                 item = string_list_insert(renames, pair->one->path);
2627                 item->util = re;
2628                 if (new_path)
2629                         apply_directory_rename_modifications(opt, pair, new_path,
2630                                                              re, tree, o_tree,
2631                                                              a_tree, b_tree,
2632                                                              entries);
2633         }
2634
2635         hashmap_iter_init(&collisions, &iter);
2636         while ((e = hashmap_iter_next(&iter))) {
2637                 free(e->target_file);
2638                 string_list_clear(&e->source_files, 0);
2639         }
2640         hashmap_free(&collisions, 1);
2641         return renames;
2642 }
2643
2644 static int process_renames(struct merge_options *opt,
2645                            struct string_list *a_renames,
2646                            struct string_list *b_renames)
2647 {
2648         int clean_merge = 1, i, j;
2649         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2650         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2651         const struct rename *sre;
2652
2653         for (i = 0; i < a_renames->nr; i++) {
2654                 sre = a_renames->items[i].util;
2655                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2656                         = (void *)sre;
2657         }
2658         for (i = 0; i < b_renames->nr; i++) {
2659                 sre = b_renames->items[i].util;
2660                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2661                         = (void *)sre;
2662         }
2663
2664         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2665                 struct string_list *renames1, *renames2Dst;
2666                 struct rename *ren1 = NULL, *ren2 = NULL;
2667                 const char *branch1, *branch2;
2668                 const char *ren1_src, *ren1_dst;
2669                 struct string_list_item *lookup;
2670
2671                 if (i >= a_renames->nr) {
2672                         ren2 = b_renames->items[j++].util;
2673                 } else if (j >= b_renames->nr) {
2674                         ren1 = a_renames->items[i++].util;
2675                 } else {
2676                         int compare = strcmp(a_renames->items[i].string,
2677                                              b_renames->items[j].string);
2678                         if (compare <= 0)
2679                                 ren1 = a_renames->items[i++].util;
2680                         if (compare >= 0)
2681                                 ren2 = b_renames->items[j++].util;
2682                 }
2683
2684                 /* TODO: refactor, so that 1/2 are not needed */
2685                 if (ren1) {
2686                         renames1 = a_renames;
2687                         renames2Dst = &b_by_dst;
2688                         branch1 = opt->branch1;
2689                         branch2 = opt->branch2;
2690                 } else {
2691                         renames1 = b_renames;
2692                         renames2Dst = &a_by_dst;
2693                         branch1 = opt->branch2;
2694                         branch2 = opt->branch1;
2695                         SWAP(ren2, ren1);
2696                 }
2697
2698                 if (ren1->processed)
2699                         continue;
2700                 ren1->processed = 1;
2701                 ren1->dst_entry->processed = 1;
2702                 /* BUG: We should only mark src_entry as processed if we
2703                  * are not dealing with a rename + add-source case.
2704                  */
2705                 ren1->src_entry->processed = 1;
2706
2707                 ren1_src = ren1->pair->one->path;
2708                 ren1_dst = ren1->pair->two->path;
2709
2710                 if (ren2) {
2711                         /* One file renamed on both sides */
2712                         const char *ren2_src = ren2->pair->one->path;
2713                         const char *ren2_dst = ren2->pair->two->path;
2714                         enum rename_type rename_type;
2715                         if (strcmp(ren1_src, ren2_src) != 0)
2716                                 BUG("ren1_src != ren2_src");
2717                         ren2->dst_entry->processed = 1;
2718                         ren2->processed = 1;
2719                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2720                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2721                                 clean_merge = 0;
2722                         } else {
2723                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2724                                 /* BUG: We should only remove ren1_src in
2725                                  * the base stage (think of rename +
2726                                  * add-source cases).
2727                                  */
2728                                 remove_file(opt, 1, ren1_src, 1);
2729                                 update_entry(ren1->dst_entry,
2730                                              ren1->pair->one,
2731                                              ren1->pair->two,
2732                                              ren2->pair->two);
2733                         }
2734                         setup_rename_conflict_info(rename_type,
2735                                                    ren1->pair,
2736                                                    ren2->pair,
2737                                                    branch1,
2738                                                    branch2,
2739                                                    ren1->dst_entry,
2740                                                    ren2->dst_entry,
2741                                                    opt,
2742                                                    NULL,
2743                                                    NULL);
2744                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2745                         /* Two different files renamed to the same thing */
2746                         char *ren2_dst;
2747                         ren2 = lookup->util;
2748                         ren2_dst = ren2->pair->two->path;
2749                         if (strcmp(ren1_dst, ren2_dst) != 0)
2750                                 BUG("ren1_dst != ren2_dst");
2751
2752                         clean_merge = 0;
2753                         ren2->processed = 1;
2754                         /*
2755                          * BUG: We should only mark src_entry as processed
2756                          * if we are not dealing with a rename + add-source
2757                          * case.
2758                          */
2759                         ren2->src_entry->processed = 1;
2760
2761                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2762                                                    ren1->pair,
2763                                                    ren2->pair,
2764                                                    branch1,
2765                                                    branch2,
2766                                                    ren1->dst_entry,
2767                                                    ren2->dst_entry,
2768                                                    opt,
2769                                                    ren1->src_entry,
2770                                                    ren2->src_entry);
2771
2772                 } else {
2773                         /* Renamed in 1, maybe changed in 2 */
2774                         /* we only use sha1 and mode of these */
2775                         struct diff_filespec src_other, dst_other;
2776                         int try_merge;
2777
2778                         /*
2779                          * unpack_trees loads entries from common-commit
2780                          * into stage 1, from head-commit into stage 2, and
2781                          * from merge-commit into stage 3.  We keep track
2782                          * of which side corresponds to the rename.
2783                          */
2784                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2785                         int other_stage =   a_renames == renames1 ? 3 : 2;
2786
2787                         /* BUG: We should only remove ren1_src in the base
2788                          * stage and in other_stage (think of rename +
2789                          * add-source case).
2790                          */
2791                         remove_file(opt, 1, ren1_src,
2792                                     renamed_stage == 2 || !was_tracked(opt, ren1_src));
2793
2794                         oidcpy(&src_other.oid,
2795                                &ren1->src_entry->stages[other_stage].oid);
2796                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2797                         oidcpy(&dst_other.oid,
2798                                &ren1->dst_entry->stages[other_stage].oid);
2799                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2800                         try_merge = 0;
2801
2802                         if (oid_eq(&src_other.oid, &null_oid) &&
2803                             ren1->add_turned_into_rename) {
2804                                 setup_rename_conflict_info(RENAME_VIA_DIR,
2805                                                            ren1->pair,
2806                                                            NULL,
2807                                                            branch1,
2808                                                            branch2,
2809                                                            ren1->dst_entry,
2810                                                            NULL,
2811                                                            opt,
2812                                                            NULL,
2813                                                            NULL);
2814                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2815                                 setup_rename_conflict_info(RENAME_DELETE,
2816                                                            ren1->pair,
2817                                                            NULL,
2818                                                            branch1,
2819                                                            branch2,
2820                                                            ren1->dst_entry,
2821                                                            NULL,
2822                                                            opt,
2823                                                            NULL,
2824                                                            NULL);
2825                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2826                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2827                                 /*
2828                                  * Added file on the other side identical to
2829                                  * the file being renamed: clean merge.
2830                                  * Also, there is no need to overwrite the
2831                                  * file already in the working copy, so call
2832                                  * update_file_flags() instead of
2833                                  * update_file().
2834                                  */
2835                                 if (update_file_flags(opt,
2836                                                       &ren1->pair->two->oid,
2837                                                       ren1->pair->two->mode,
2838                                                       ren1_dst,
2839                                                       1, /* update_cache */
2840                                                       0  /* update_wd    */))
2841                                         clean_merge = -1;
2842                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2843                                 /*
2844                                  * Probably not a clean merge, but it's
2845                                  * premature to set clean_merge to 0 here,
2846                                  * because if the rename merges cleanly and
2847                                  * the merge exactly matches the newly added
2848                                  * file, then the merge will be clean.
2849                                  */
2850                                 setup_rename_conflict_info(RENAME_ADD,
2851                                                            ren1->pair,
2852                                                            NULL,
2853                                                            branch1,
2854                                                            branch2,
2855                                                            ren1->dst_entry,
2856                                                            NULL,
2857                                                            opt,
2858                                                            ren1->src_entry,
2859                                                            NULL);
2860                         } else
2861                                 try_merge = 1;
2862
2863                         if (clean_merge < 0)
2864                                 goto cleanup_and_return;
2865                         if (try_merge) {
2866                                 struct diff_filespec *o, *a, *b;
2867                                 src_other.path = (char *)ren1_src;
2868
2869                                 o = ren1->pair->one;
2870                                 if (a_renames == renames1) {
2871                                         a = ren1->pair->two;
2872                                         b = &src_other;
2873                                 } else {
2874                                         b = ren1->pair->two;
2875                                         a = &src_other;
2876                                 }
2877                                 update_entry(ren1->dst_entry, o, a, b);
2878                                 setup_rename_conflict_info(RENAME_NORMAL,
2879                                                            ren1->pair,
2880                                                            NULL,
2881                                                            branch1,
2882                                                            NULL,
2883                                                            ren1->dst_entry,
2884                                                            NULL,
2885                                                            opt,
2886                                                            NULL,
2887                                                            NULL);
2888                         }
2889                 }
2890         }
2891 cleanup_and_return:
2892         string_list_clear(&a_by_dst, 0);
2893         string_list_clear(&b_by_dst, 0);
2894
2895         return clean_merge;
2896 }
2897
2898 struct rename_info {
2899         struct string_list *head_renames;
2900         struct string_list *merge_renames;
2901 };
2902
2903 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2904                                    struct hashmap *dir_renames)
2905 {
2906         struct hashmap_iter iter;
2907         struct dir_rename_entry *e;
2908
2909         hashmap_iter_init(dir_renames, &iter);
2910         while ((e = hashmap_iter_next(&iter))) {
2911                 free(e->dir);
2912                 strbuf_release(&e->new_dir);
2913                 /* possible_new_dirs already cleared in get_directory_renames */
2914         }
2915         hashmap_free(dir_renames, 1);
2916         free(dir_renames);
2917
2918         free(pairs->queue);
2919         free(pairs);
2920 }
2921
2922 static int detect_and_process_renames(struct merge_options *opt,
2923                                       struct tree *common,
2924                                       struct tree *head,
2925                                       struct tree *merge,
2926                                       struct string_list *entries,
2927                                       struct rename_info *ri)
2928 {
2929         struct diff_queue_struct *head_pairs, *merge_pairs;
2930         struct hashmap *dir_re_head, *dir_re_merge;
2931         int clean = 1;
2932
2933         ri->head_renames = NULL;
2934         ri->merge_renames = NULL;
2935
2936         if (!merge_detect_rename(opt))
2937                 return 1;
2938
2939         head_pairs = get_diffpairs(opt, common, head);
2940         merge_pairs = get_diffpairs(opt, common, merge);
2941
2942         if (opt->detect_directory_renames) {
2943                 dir_re_head = get_directory_renames(head_pairs);
2944                 dir_re_merge = get_directory_renames(merge_pairs);
2945
2946                 handle_directory_level_conflicts(opt,
2947                                                  dir_re_head, head,
2948                                                  dir_re_merge, merge);
2949         } else {
2950                 dir_re_head  = xmalloc(sizeof(*dir_re_head));
2951                 dir_re_merge = xmalloc(sizeof(*dir_re_merge));
2952                 dir_rename_init(dir_re_head);
2953                 dir_rename_init(dir_re_merge);
2954         }
2955
2956         ri->head_renames  = get_renames(opt, head_pairs,
2957                                         dir_re_merge, dir_re_head, head,
2958                                         common, head, merge, entries,
2959                                         &clean);
2960         if (clean < 0)
2961                 goto cleanup;
2962         ri->merge_renames = get_renames(opt, merge_pairs,
2963                                         dir_re_head, dir_re_merge, merge,
2964                                         common, head, merge, entries,
2965                                         &clean);
2966         if (clean < 0)
2967                 goto cleanup;
2968         clean &= process_renames(opt, ri->head_renames, ri->merge_renames);
2969
2970 cleanup:
2971         /*
2972          * Some cleanup is deferred until cleanup_renames() because the
2973          * data structures are still needed and referenced in
2974          * process_entry().  But there are a few things we can free now.
2975          */
2976         initial_cleanup_rename(head_pairs, dir_re_head);
2977         initial_cleanup_rename(merge_pairs, dir_re_merge);
2978
2979         return clean;
2980 }
2981
2982 static void final_cleanup_rename(struct string_list *rename)
2983 {
2984         const struct rename *re;
2985         int i;
2986
2987         if (rename == NULL)
2988                 return;
2989
2990         for (i = 0; i < rename->nr; i++) {
2991                 re = rename->items[i].util;
2992                 diff_free_filepair(re->pair);
2993         }
2994         string_list_clear(rename, 1);
2995         free(rename);
2996 }
2997
2998 static void final_cleanup_renames(struct rename_info *re_info)
2999 {
3000         final_cleanup_rename(re_info->head_renames);
3001         final_cleanup_rename(re_info->merge_renames);
3002 }
3003
3004 static struct object_id *stage_oid(const struct object_id *oid, unsigned mode)
3005 {
3006         return (is_null_oid(oid) || mode == 0) ? NULL: (struct object_id *)oid;
3007 }
3008
3009 static int read_oid_strbuf(struct merge_options *opt,
3010                            const struct object_id *oid,
3011                            struct strbuf *dst)
3012 {
3013         void *buf;
3014         enum object_type type;
3015         unsigned long size;
3016         buf = read_object_file(oid, &type, &size);
3017         if (!buf)
3018                 return err(opt, _("cannot read object %s"), oid_to_hex(oid));
3019         if (type != OBJ_BLOB) {
3020                 free(buf);
3021                 return err(opt, _("object %s is not a blob"), oid_to_hex(oid));
3022         }
3023         strbuf_attach(dst, buf, size, size + 1);
3024         return 0;
3025 }
3026
3027 static int blob_unchanged(struct merge_options *opt,
3028                           const struct object_id *o_oid,
3029                           unsigned o_mode,
3030                           const struct object_id *a_oid,
3031                           unsigned a_mode,
3032                           int renormalize, const char *path)
3033 {
3034         struct strbuf obuf = STRBUF_INIT;
3035         struct strbuf abuf = STRBUF_INIT;
3036         int ret = 0; /* assume changed for safety */
3037         const struct index_state *idx = opt->repo->index;
3038
3039         if (a_mode != o_mode)
3040                 return 0;
3041         if (oid_eq(o_oid, a_oid))
3042                 return 1;
3043         if (!renormalize)
3044                 return 0;
3045
3046         assert(o_oid && a_oid);
3047         if (read_oid_strbuf(opt, o_oid, &obuf) ||
3048             read_oid_strbuf(opt, a_oid, &abuf))
3049                 goto error_return;
3050         /*
3051          * Note: binary | is used so that both renormalizations are
3052          * performed.  Comparison can be skipped if both files are
3053          * unchanged since their sha1s have already been compared.
3054          */
3055         if (renormalize_buffer(idx, path, obuf.buf, obuf.len, &obuf) |
3056             renormalize_buffer(idx, path, abuf.buf, abuf.len, &abuf))
3057                 ret = (obuf.len == abuf.len && !memcmp(obuf.buf, abuf.buf, obuf.len));
3058
3059 error_return:
3060         strbuf_release(&obuf);
3061         strbuf_release(&abuf);
3062         return ret;
3063 }
3064
3065 static int handle_modify_delete(struct merge_options *opt,
3066                                 const char *path,
3067                                 struct object_id *o_oid, int o_mode,
3068                                 struct object_id *a_oid, int a_mode,
3069                                 struct object_id *b_oid, int b_mode)
3070 {
3071         const char *modify_branch, *delete_branch;
3072         struct object_id *changed_oid;
3073         int changed_mode;
3074
3075         if (a_oid) {
3076                 modify_branch = opt->branch1;
3077                 delete_branch = opt->branch2;
3078                 changed_oid = a_oid;
3079                 changed_mode = a_mode;
3080         } else {
3081                 modify_branch = opt->branch2;
3082                 delete_branch = opt->branch1;
3083                 changed_oid = b_oid;
3084                 changed_mode = b_mode;
3085         }
3086
3087         return handle_change_delete(opt,
3088                                     path, NULL,
3089                                     o_oid, o_mode,
3090                                     changed_oid, changed_mode,
3091                                     modify_branch, delete_branch,
3092                                     _("modify"), _("modified"));
3093 }
3094
3095 static int handle_content_merge(struct merge_options *opt,
3096                                 const char *path,
3097                                 int is_dirty,
3098                                 struct object_id *o_oid, int o_mode,
3099                                 struct object_id *a_oid, int a_mode,
3100                                 struct object_id *b_oid, int b_mode,
3101                                 struct rename_conflict_info *ci)
3102 {
3103         const char *reason = _("content");
3104         const char *path1 = NULL, *path2 = NULL;
3105         struct merge_file_info mfi;
3106         struct diff_filespec one, a, b;
3107         unsigned df_conflict_remains = 0;
3108
3109         if (!o_oid) {
3110                 reason = _("add/add");
3111                 o_oid = (struct object_id *)&null_oid;
3112         }
3113         one.path = a.path = b.path = (char *)path;
3114         oidcpy(&one.oid, o_oid);
3115         one.mode = o_mode;
3116         oidcpy(&a.oid, a_oid);
3117         a.mode = a_mode;
3118         oidcpy(&b.oid, b_oid);
3119         b.mode = b_mode;
3120
3121         if (ci) {
3122                 struct diff_filepair *pair1 = ci->pair1;
3123
3124                 path1 = (opt->branch1 == ci->branch1) ?
3125                         pair1->two->path : pair1->one->path;
3126                 /* If ci->pair2 != NULL, we are in
3127                  * RENAME_ONE_FILE_TO_ONE case.  Otherwise, we have a
3128                  * normal rename.
3129                  */
3130                 path2 = (ci->pair2 ||
3131                          opt->branch2 == ci->branch1) ?
3132                         pair1->two->path : pair1->one->path;
3133                 one.path = pair1->one->path;
3134                 a.path = (char *)path1;
3135                 b.path = (char *)path2;
3136
3137                 if (dir_in_way(opt->repo->index, path, !opt->call_depth,
3138                                S_ISGITLINK(pair1->two->mode)))
3139                         df_conflict_remains = 1;
3140         }
3141         if (merge_mode_and_contents(opt, &one, &a, &b, path,
3142                                     opt->branch1, opt->branch2,
3143                                     opt->call_depth * 2, &mfi))
3144                 return -1;
3145
3146         /*
3147          * We can skip updating the working tree file iff:
3148          *   a) The merge is clean
3149          *   b) The merge matches what was in HEAD (content, mode, pathname)
3150          *   c) The target path is usable (i.e. not involved in D/F conflict)
3151          */
3152         if (mfi.clean &&
3153             was_tracked_and_matches(opt, path, &mfi.oid, mfi.mode) &&
3154             !df_conflict_remains) {
3155                 int pos;
3156                 struct cache_entry *ce;
3157
3158                 output(opt, 3, _("Skipped %s (merged same as existing)"), path);
3159                 if (add_cacheinfo(opt, mfi.mode, &mfi.oid, path,
3160                                   0, (!opt->call_depth && !is_dirty), 0))
3161                         return -1;
3162                 /*
3163                  * However, add_cacheinfo() will delete the old cache entry
3164                  * and add a new one.  We need to copy over any skip_worktree
3165                  * flag to avoid making the file appear as if it were
3166                  * deleted by the user.
3167                  */
3168                 pos = index_name_pos(&opt->orig_index, path, strlen(path));
3169                 ce = opt->orig_index.cache[pos];
3170                 if (ce_skip_worktree(ce)) {
3171                         pos = index_name_pos(opt->repo->index, path, strlen(path));
3172                         ce = opt->repo->index->cache[pos];
3173                         ce->ce_flags |= CE_SKIP_WORKTREE;
3174                 }
3175                 return mfi.clean;
3176         }
3177
3178         if (!mfi.clean) {
3179                 if (S_ISGITLINK(mfi.mode))
3180                         reason = _("submodule");
3181                 output(opt, 1, _("CONFLICT (%s): Merge conflict in %s"),
3182                                 reason, path);
3183                 if (ci && !df_conflict_remains)
3184                         if (update_stages(opt, path, &one, &a, &b))
3185                                 return -1;
3186         }
3187
3188         if (df_conflict_remains || is_dirty) {
3189                 char *new_path;
3190                 if (opt->call_depth) {
3191                         remove_file_from_index(opt->repo->index, path);
3192                 } else {
3193                         if (!mfi.clean) {
3194                                 if (update_stages(opt, path, &one, &a, &b))
3195                                         return -1;
3196                         } else {
3197                                 int file_from_stage2 = was_tracked(opt, path);
3198                                 struct diff_filespec merged;
3199                                 oidcpy(&merged.oid, &mfi.oid);
3200                                 merged.mode = mfi.mode;
3201
3202                                 if (update_stages(opt, path, NULL,
3203                                                   file_from_stage2 ? &merged : NULL,
3204                                                   file_from_stage2 ? NULL : &merged))
3205                                         return -1;
3206                         }
3207
3208                 }
3209                 new_path = unique_path(opt, path, ci->branch1);
3210                 if (is_dirty) {
3211                         output(opt, 1, _("Refusing to lose dirty file at %s"),
3212                                path);
3213                 }
3214                 output(opt, 1, _("Adding as %s instead"), new_path);
3215                 if (update_file(opt, 0, &mfi.oid, mfi.mode, new_path)) {
3216                         free(new_path);
3217                         return -1;
3218                 }
3219                 free(new_path);
3220                 mfi.clean = 0;
3221         } else if (update_file(opt, mfi.clean, &mfi.oid, mfi.mode, path))
3222                 return -1;
3223         return !is_dirty && mfi.clean;
3224 }
3225
3226 static int handle_rename_normal(struct merge_options *opt,
3227                                 const char *path,
3228                                 struct object_id *o_oid, unsigned int o_mode,
3229                                 struct object_id *a_oid, unsigned int a_mode,
3230                                 struct object_id *b_oid, unsigned int b_mode,
3231                                 struct rename_conflict_info *ci)
3232 {
3233         /* Merge the content and write it out */
3234         return handle_content_merge(opt, path, was_dirty(opt, path),
3235                                     o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3236                                     ci);
3237 }
3238
3239 /* Per entry merge function */
3240 static int process_entry(struct merge_options *opt,
3241                          const char *path, struct stage_data *entry)
3242 {
3243         int clean_merge = 1;
3244         int normalize = opt->renormalize;
3245         unsigned o_mode = entry->stages[1].mode;
3246         unsigned a_mode = entry->stages[2].mode;
3247         unsigned b_mode = entry->stages[3].mode;
3248         struct object_id *o_oid = stage_oid(&entry->stages[1].oid, o_mode);
3249         struct object_id *a_oid = stage_oid(&entry->stages[2].oid, a_mode);
3250         struct object_id *b_oid = stage_oid(&entry->stages[3].oid, b_mode);
3251
3252         entry->processed = 1;
3253         if (entry->rename_conflict_info) {
3254                 struct rename_conflict_info *ci = entry->rename_conflict_info;
3255                 switch (ci->rename_type) {
3256                 case RENAME_NORMAL:
3257                 case RENAME_ONE_FILE_TO_ONE:
3258                         clean_merge = handle_rename_normal(opt,
3259                                                            path,
3260                                                            o_oid, o_mode,
3261                                                            a_oid, a_mode,
3262                                                            b_oid, b_mode,
3263                                                            ci);
3264                         break;
3265                 case RENAME_VIA_DIR:
3266                         clean_merge = 1;
3267                         if (handle_rename_via_dir(opt, ci->pair1, ci->branch1))
3268                                 clean_merge = -1;
3269                         break;
3270                 case RENAME_ADD:
3271                         /*
3272                          * Probably unclean merge, but if the renamed file
3273                          * merges cleanly and the result can then be
3274                          * two-way merged cleanly with the added file, I
3275                          * guess it's a clean merge?
3276                          */
3277                         clean_merge = handle_rename_add(opt, ci);
3278                         break;
3279                 case RENAME_DELETE:
3280                         clean_merge = 0;
3281                         if (handle_rename_delete(opt, ci->pair1,
3282                                                  ci->branch1, ci->branch2))
3283                                 clean_merge = -1;
3284                         break;
3285                 case RENAME_ONE_FILE_TO_TWO:
3286                         clean_merge = 0;
3287                         if (handle_rename_rename_1to2(opt, ci))
3288                                 clean_merge = -1;
3289                         break;
3290                 case RENAME_TWO_FILES_TO_ONE:
3291                         /*
3292                          * Probably unclean merge, but if the two renamed
3293                          * files merge cleanly and the two resulting files
3294                          * can then be two-way merged cleanly, I guess it's
3295                          * a clean merge?
3296                          */
3297                         clean_merge = handle_rename_rename_2to1(opt, ci);
3298                         break;
3299                 default:
3300                         entry->processed = 0;
3301                         break;
3302                 }
3303         } else if (o_oid && (!a_oid || !b_oid)) {
3304                 /* Case A: Deleted in one */
3305                 if ((!a_oid && !b_oid) ||
3306                     (!b_oid && blob_unchanged(opt, o_oid, o_mode, a_oid, a_mode, normalize, path)) ||
3307                     (!a_oid && blob_unchanged(opt, o_oid, o_mode, b_oid, b_mode, normalize, path))) {
3308                         /* Deleted in both or deleted in one and
3309                          * unchanged in the other */
3310                         if (a_oid)
3311                                 output(opt, 2, _("Removing %s"), path);
3312                         /* do not touch working file if it did not exist */
3313                         remove_file(opt, 1, path, !a_oid);
3314                 } else {
3315                         /* Modify/delete; deleted side may have put a directory in the way */
3316                         clean_merge = 0;
3317                         if (handle_modify_delete(opt, path, o_oid, o_mode,
3318                                                  a_oid, a_mode, b_oid, b_mode))
3319                                 clean_merge = -1;
3320                 }
3321         } else if ((!o_oid && a_oid && !b_oid) ||
3322                    (!o_oid && !a_oid && b_oid)) {
3323                 /* Case B: Added in one. */
3324                 /* [nothing|directory] -> ([nothing|directory], file) */
3325
3326                 const char *add_branch;
3327                 const char *other_branch;
3328                 unsigned mode;
3329                 const struct object_id *oid;
3330                 const char *conf;
3331
3332                 if (a_oid) {
3333                         add_branch = opt->branch1;
3334                         other_branch = opt->branch2;
3335                         mode = a_mode;
3336                         oid = a_oid;
3337                         conf = _("file/directory");
3338                 } else {
3339                         add_branch = opt->branch2;
3340                         other_branch = opt->branch1;
3341                         mode = b_mode;
3342                         oid = b_oid;
3343                         conf = _("directory/file");
3344                 }
3345                 if (dir_in_way(opt->repo->index, path,
3346                                !opt->call_depth && !S_ISGITLINK(a_mode),
3347                                0)) {
3348                         char *new_path = unique_path(opt, path, add_branch);
3349                         clean_merge = 0;
3350                         output(opt, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3351                                "Adding %s as %s"),
3352                                conf, path, other_branch, path, new_path);
3353                         if (update_file(opt, 0, oid, mode, new_path))
3354                                 clean_merge = -1;
3355                         else if (opt->call_depth)
3356                                 remove_file_from_index(opt->repo->index, path);
3357                         free(new_path);
3358                 } else {
3359                         output(opt, 2, _("Adding %s"), path);
3360                         /* do not overwrite file if already present */
3361                         if (update_file_flags(opt, oid, mode, path, 1, !a_oid))
3362                                 clean_merge = -1;
3363                 }
3364         } else if (a_oid && b_oid) {
3365                 if (!o_oid) {
3366                         /* Case C: Added in both (check for same permissions) */
3367                         output(opt, 1,
3368                                _("CONFLICT (add/add): Merge conflict in %s"),
3369                                path);
3370                         clean_merge = handle_file_collision(opt,
3371                                                             path, NULL, NULL,
3372                                                             opt->branch1,
3373                                                             opt->branch2,
3374                                                             a_oid, a_mode,
3375                                                             b_oid, b_mode);
3376                 } else {
3377                         /* case D: Modified in both, but differently. */
3378                         int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3379                         clean_merge = handle_content_merge(opt, path,
3380                                                            is_dirty,
3381                                                            o_oid, o_mode,
3382                                                            a_oid, a_mode,
3383                                                            b_oid, b_mode,
3384                                                            NULL);
3385                 }
3386         } else if (!o_oid && !a_oid && !b_oid) {
3387                 /*
3388                  * this entry was deleted altogether. a_mode == 0 means
3389                  * we had that path and want to actively remove it.
3390                  */
3391                 remove_file(opt, 1, path, !a_mode);
3392         } else
3393                 BUG("fatal merge failure, shouldn't happen.");
3394
3395         return clean_merge;
3396 }
3397
3398 int merge_trees(struct merge_options *opt,
3399                 struct tree *head,
3400                 struct tree *merge,
3401                 struct tree *common,
3402                 struct tree **result)
3403 {
3404         struct index_state *istate = opt->repo->index;
3405         int code, clean;
3406         struct strbuf sb = STRBUF_INIT;
3407
3408         if (!opt->call_depth && repo_index_has_changes(opt->repo, head, &sb)) {
3409                 err(opt, _("Your local changes to the following files would be overwritten by merge:\n  %s"),
3410                     sb.buf);
3411                 return -1;
3412         }
3413
3414         if (opt->subtree_shift) {
3415                 merge = shift_tree_object(opt->repo, head, merge, opt->subtree_shift);
3416                 common = shift_tree_object(opt->repo, head, common, opt->subtree_shift);
3417         }
3418
3419         if (oid_eq(&common->object.oid, &merge->object.oid)) {
3420                 output(opt, 0, _("Already up to date!"));
3421                 *result = head;
3422                 return 1;
3423         }
3424
3425         code = unpack_trees_start(opt, common, head, merge);
3426
3427         if (code != 0) {
3428                 if (show(opt, 4) || opt->call_depth)
3429                         err(opt, _("merging of trees %s and %s failed"),
3430                             oid_to_hex(&head->object.oid),
3431                             oid_to_hex(&merge->object.oid));
3432                 unpack_trees_finish(opt);
3433                 return -1;
3434         }
3435
3436         if (unmerged_index(istate)) {
3437                 struct string_list *entries;
3438                 struct rename_info re_info;
3439                 int i;
3440                 /*
3441                  * Only need the hashmap while processing entries, so
3442                  * initialize it here and free it when we are done running
3443                  * through the entries. Keeping it in the merge_options as
3444                  * opposed to decaring a local hashmap is for convenience
3445                  * so that we don't have to pass it to around.
3446                  */
3447                 hashmap_init(&opt->current_file_dir_set, path_hashmap_cmp, NULL, 512);
3448                 get_files_dirs(opt, head);
3449                 get_files_dirs(opt, merge);
3450
3451                 entries = get_unmerged(opt->repo->index);
3452                 clean = detect_and_process_renames(opt, common, head, merge,
3453                                                    entries, &re_info);
3454                 record_df_conflict_files(opt, entries);
3455                 if (clean < 0)
3456                         goto cleanup;
3457                 for (i = entries->nr-1; 0 <= i; i--) {
3458                         const char *path = entries->items[i].string;
3459                         struct stage_data *e = entries->items[i].util;
3460                         if (!e->processed) {
3461                                 int ret = process_entry(opt, path, e);
3462                                 if (!ret)
3463                                         clean = 0;
3464                                 else if (ret < 0) {
3465                                         clean = ret;
3466                                         goto cleanup;
3467                                 }
3468                         }
3469                 }
3470                 for (i = 0; i < entries->nr; i++) {
3471                         struct stage_data *e = entries->items[i].util;
3472                         if (!e->processed)
3473                                 BUG("unprocessed path??? %s",
3474                                     entries->items[i].string);
3475                 }
3476
3477         cleanup:
3478                 final_cleanup_renames(&re_info);
3479
3480                 string_list_clear(entries, 1);
3481                 free(entries);
3482
3483                 hashmap_free(&opt->current_file_dir_set, 1);
3484
3485                 if (clean < 0) {
3486                         unpack_trees_finish(opt);
3487                         return clean;
3488                 }
3489         }
3490         else
3491                 clean = 1;
3492
3493         unpack_trees_finish(opt);
3494
3495         if (opt->call_depth && !(*result = write_tree_from_memory(opt)))
3496                 return -1;
3497
3498         return clean;
3499 }
3500
3501 static struct commit_list *reverse_commit_list(struct commit_list *list)
3502 {
3503         struct commit_list *next = NULL, *current, *backup;
3504         for (current = list; current; current = backup) {
3505                 backup = current->next;
3506                 current->next = next;
3507                 next = current;
3508         }
3509         return next;
3510 }
3511
3512 /*
3513  * Merge the commits h1 and h2, return the resulting virtual
3514  * commit object and a flag indicating the cleanness of the merge.
3515  */
3516 int merge_recursive(struct merge_options *opt,
3517                     struct commit *h1,
3518                     struct commit *h2,
3519                     struct commit_list *ca,
3520                     struct commit **result)
3521 {
3522         struct commit_list *iter;
3523         struct commit *merged_common_ancestors;
3524         struct tree *mrtree;
3525         int clean;
3526
3527         if (show(opt, 4)) {
3528                 output(opt, 4, _("Merging:"));
3529                 output_commit_title(opt, h1);
3530                 output_commit_title(opt, h2);
3531         }
3532
3533         if (!ca) {
3534                 ca = get_merge_bases(h1, h2);
3535                 ca = reverse_commit_list(ca);
3536         }
3537
3538         if (show(opt, 5)) {
3539                 unsigned cnt = commit_list_count(ca);
3540
3541                 output(opt, 5, Q_("found %u common ancestor:",
3542                                 "found %u common ancestors:", cnt), cnt);
3543                 for (iter = ca; iter; iter = iter->next)
3544                         output_commit_title(opt, iter->item);
3545         }
3546
3547         merged_common_ancestors = pop_commit(&ca);
3548         if (merged_common_ancestors == NULL) {
3549                 /* if there is no common ancestor, use an empty tree */
3550                 struct tree *tree;
3551
3552                 tree = lookup_tree(opt->repo, opt->repo->hash_algo->empty_tree);
3553                 merged_common_ancestors = make_virtual_commit(opt->repo, tree, "ancestor");
3554         }
3555
3556         for (iter = ca; iter; iter = iter->next) {
3557                 const char *saved_b1, *saved_b2;
3558                 opt->call_depth++;
3559                 /*
3560                  * When the merge fails, the result contains files
3561                  * with conflict markers. The cleanness flag is
3562                  * ignored (unless indicating an error), it was never
3563                  * actually used, as result of merge_trees has always
3564                  * overwritten it: the committed "conflicts" were
3565                  * already resolved.
3566                  */
3567                 discard_index(opt->repo->index);
3568                 saved_b1 = opt->branch1;
3569                 saved_b2 = opt->branch2;
3570                 opt->branch1 = "Temporary merge branch 1";
3571                 opt->branch2 = "Temporary merge branch 2";
3572                 if (merge_recursive(opt, merged_common_ancestors, iter->item,
3573                                     NULL, &merged_common_ancestors) < 0)
3574                         return -1;
3575                 opt->branch1 = saved_b1;
3576                 opt->branch2 = saved_b2;
3577                 opt->call_depth--;
3578
3579                 if (!merged_common_ancestors)
3580                         return err(opt, _("merge returned no commit"));
3581         }
3582
3583         discard_index(opt->repo->index);
3584         if (!opt->call_depth)
3585                 repo_read_index(opt->repo);
3586
3587         opt->ancestor = "merged common ancestors";
3588         clean = merge_trees(opt, get_commit_tree(h1), get_commit_tree(h2),
3589                             get_commit_tree(merged_common_ancestors),
3590                             &mrtree);
3591         if (clean < 0) {
3592                 flush_output(opt);
3593                 return clean;
3594         }
3595
3596         if (opt->call_depth) {
3597                 *result = make_virtual_commit(opt->repo, mrtree, "merged tree");
3598                 commit_list_insert(h1, &(*result)->parents);
3599                 commit_list_insert(h2, &(*result)->parents->next);
3600         }
3601         flush_output(opt);
3602         if (!opt->call_depth && opt->buffer_output < 2)
3603                 strbuf_release(&opt->obuf);
3604         if (show(opt, 2))
3605                 diff_warn_rename_limit("merge.renamelimit",
3606                                        opt->needed_rename_limit, 0);
3607         return clean;
3608 }
3609
3610 static struct commit *get_ref(struct repository *repo, const struct object_id *oid,
3611                               const char *name)
3612 {
3613         struct object *object;
3614
3615         object = deref_tag(repo, parse_object(repo, oid),
3616                            name, strlen(name));
3617         if (!object)
3618                 return NULL;
3619         if (object->type == OBJ_TREE)
3620                 return make_virtual_commit(repo, (struct tree*)object, name);
3621         if (object->type != OBJ_COMMIT)
3622                 return NULL;
3623         if (parse_commit((struct commit *)object))
3624                 return NULL;
3625         return (struct commit *)object;
3626 }
3627
3628 int merge_recursive_generic(struct merge_options *opt,
3629                             const struct object_id *head,
3630                             const struct object_id *merge,
3631                             int num_base_list,
3632                             const struct object_id **base_list,
3633                             struct commit **result)
3634 {
3635         int clean;
3636         struct lock_file lock = LOCK_INIT;
3637         struct commit *head_commit = get_ref(opt->repo, head, opt->branch1);
3638         struct commit *next_commit = get_ref(opt->repo, merge, opt->branch2);
3639         struct commit_list *ca = NULL;
3640
3641         if (base_list) {
3642                 int i;
3643                 for (i = 0; i < num_base_list; ++i) {
3644                         struct commit *base;
3645                         if (!(base = get_ref(opt->repo, base_list[i], oid_to_hex(base_list[i]))))
3646                                 return err(opt, _("Could not parse object '%s'"),
3647                                            oid_to_hex(base_list[i]));
3648                         commit_list_insert(base, &ca);
3649                 }
3650         }
3651
3652         repo_hold_locked_index(opt->repo, &lock, LOCK_DIE_ON_ERROR);
3653         clean = merge_recursive(opt, head_commit, next_commit, ca,
3654                                 result);
3655         if (clean < 0) {
3656                 rollback_lock_file(&lock);
3657                 return clean;
3658         }
3659
3660         if (write_locked_index(opt->repo->index, &lock,
3661                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3662                 return err(opt, _("Unable to write index."));
3663
3664         return clean ? 0 : 1;
3665 }
3666
3667 static void merge_recursive_config(struct merge_options *opt)
3668 {
3669         char *value = NULL;
3670         git_config_get_int("merge.verbosity", &opt->verbosity);
3671         git_config_get_int("diff.renamelimit", &opt->diff_rename_limit);
3672         git_config_get_int("merge.renamelimit", &opt->merge_rename_limit);
3673         if (!git_config_get_string("diff.renames", &value)) {
3674                 opt->diff_detect_rename = git_config_rename("diff.renames", value);
3675                 free(value);
3676         }
3677         if (!git_config_get_string("merge.renames", &value)) {
3678                 opt->merge_detect_rename = git_config_rename("merge.renames", value);
3679                 free(value);
3680         }
3681         git_config(git_xmerge_config, NULL);
3682 }
3683
3684 void init_merge_options(struct merge_options *opt,
3685                         struct repository *repo)
3686 {
3687         const char *merge_verbosity;
3688         memset(opt, 0, sizeof(struct merge_options));
3689         opt->repo = repo;
3690         opt->verbosity = 2;
3691         opt->buffer_output = 1;
3692         opt->diff_rename_limit = -1;
3693         opt->merge_rename_limit = -1;
3694         opt->renormalize = 0;
3695         opt->diff_detect_rename = -1;
3696         opt->merge_detect_rename = -1;
3697         opt->detect_directory_renames = 1;
3698         merge_recursive_config(opt);
3699         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3700         if (merge_verbosity)
3701                 opt->verbosity = strtol(merge_verbosity, NULL, 10);
3702         if (opt->verbosity >= 5)
3703                 opt->buffer_output = 0;
3704         strbuf_init(&opt->obuf, 0);
3705         string_list_init(&opt->df_conflict_file_set, 1);
3706 }
3707
3708 int parse_merge_opt(struct merge_options *opt, const char *s)
3709 {
3710         const char *arg;
3711
3712         if (!s || !*s)
3713                 return -1;
3714         if (!strcmp(s, "ours"))
3715                 opt->recursive_variant = MERGE_RECURSIVE_OURS;
3716         else if (!strcmp(s, "theirs"))
3717                 opt->recursive_variant = MERGE_RECURSIVE_THEIRS;
3718         else if (!strcmp(s, "subtree"))
3719                 opt->subtree_shift = "";
3720         else if (skip_prefix(s, "subtree=", &arg))
3721                 opt->subtree_shift = arg;
3722         else if (!strcmp(s, "patience"))
3723                 opt->xdl_opts = DIFF_WITH_ALG(opt, PATIENCE_DIFF);
3724         else if (!strcmp(s, "histogram"))
3725                 opt->xdl_opts = DIFF_WITH_ALG(opt, HISTOGRAM_DIFF);
3726         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3727                 long value = parse_algorithm_value(arg);
3728                 if (value < 0)
3729                         return -1;
3730                 /* clear out previous settings */
3731                 DIFF_XDL_CLR(opt, NEED_MINIMAL);
3732                 opt->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3733                 opt->xdl_opts |= value;
3734         }
3735         else if (!strcmp(s, "ignore-space-change"))
3736                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_CHANGE);
3737         else if (!strcmp(s, "ignore-all-space"))
3738                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE);
3739         else if (!strcmp(s, "ignore-space-at-eol"))
3740                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_AT_EOL);
3741         else if (!strcmp(s, "ignore-cr-at-eol"))
3742                 DIFF_XDL_SET(opt, IGNORE_CR_AT_EOL);
3743         else if (!strcmp(s, "renormalize"))
3744                 opt->renormalize = 1;
3745         else if (!strcmp(s, "no-renormalize"))
3746                 opt->renormalize = 0;
3747         else if (!strcmp(s, "no-renames"))
3748                 opt->merge_detect_rename = 0;
3749         else if (!strcmp(s, "find-renames")) {
3750                 opt->merge_detect_rename = 1;
3751                 opt->rename_score = 0;
3752         }
3753         else if (skip_prefix(s, "find-renames=", &arg) ||
3754                  skip_prefix(s, "rename-threshold=", &arg)) {
3755                 if ((opt->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3756                         return -1;
3757                 opt->merge_detect_rename = 1;
3758         }
3759         /*
3760          * Please update $__git_merge_strategy_options in
3761          * git-completion.bash when you add new options
3762          */
3763         else
3764                 return -1;
3765         return 0;
3766 }