builtin/commit-graph.c: introduce split strategy 'replace'
[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 "merge-recursive.h"
8
9 #include "advice.h"
10 #include "alloc.h"
11 #include "attr.h"
12 #include "blob.h"
13 #include "builtin.h"
14 #include "cache-tree.h"
15 #include "commit.h"
16 #include "commit-reach.h"
17 #include "config.h"
18 #include "diff.h"
19 #include "diffcore.h"
20 #include "dir.h"
21 #include "ll-merge.h"
22 #include "lockfile.h"
23 #include "object-store.h"
24 #include "repository.h"
25 #include "revision.h"
26 #include "string-list.h"
27 #include "submodule.h"
28 #include "tag.h"
29 #include "tree-walk.h"
30 #include "unpack-trees.h"
31 #include "xdiff-interface.h"
32
33 struct merge_options_internal {
34         int call_depth;
35         int needed_rename_limit;
36         struct hashmap current_file_dir_set;
37         struct string_list df_conflict_file_set;
38         struct unpack_trees_options unpack_opts;
39         struct index_state orig_index;
40 };
41
42 struct path_hashmap_entry {
43         struct hashmap_entry e;
44         char path[FLEX_ARRAY];
45 };
46
47 static int path_hashmap_cmp(const void *cmp_data,
48                             const struct hashmap_entry *eptr,
49                             const struct hashmap_entry *entry_or_key,
50                             const void *keydata)
51 {
52         const struct path_hashmap_entry *a, *b;
53         const char *key = keydata;
54
55         a = container_of(eptr, const struct path_hashmap_entry, e);
56         b = container_of(entry_or_key, const struct path_hashmap_entry, e);
57
58         if (ignore_case)
59                 return strcasecmp(a->path, key ? key : b->path);
60         else
61                 return strcmp(a->path, key ? key : b->path);
62 }
63
64 static unsigned int path_hash(const char *path)
65 {
66         return ignore_case ? strihash(path) : strhash(path);
67 }
68
69 /*
70  * For dir_rename_entry, directory names are stored as a full path from the
71  * toplevel of the repository and do not include a trailing '/'.  Also:
72  *
73  *   dir:                original name of directory being renamed
74  *   non_unique_new_dir: if true, could not determine new_dir
75  *   new_dir:            final name of directory being renamed
76  *   possible_new_dirs:  temporary used to help determine new_dir; see comments
77  *                       in get_directory_renames() for details
78  */
79 struct dir_rename_entry {
80         struct hashmap_entry ent;
81         char *dir;
82         unsigned non_unique_new_dir:1;
83         struct strbuf new_dir;
84         struct string_list possible_new_dirs;
85 };
86
87 static struct dir_rename_entry *dir_rename_find_entry(struct hashmap *hashmap,
88                                                       char *dir)
89 {
90         struct dir_rename_entry key;
91
92         if (dir == NULL)
93                 return NULL;
94         hashmap_entry_init(&key.ent, strhash(dir));
95         key.dir = dir;
96         return hashmap_get_entry(hashmap, &key, ent, NULL);
97 }
98
99 static int dir_rename_cmp(const void *unused_cmp_data,
100                           const struct hashmap_entry *eptr,
101                           const struct hashmap_entry *entry_or_key,
102                           const void *unused_keydata)
103 {
104         const struct dir_rename_entry *e1, *e2;
105
106         e1 = container_of(eptr, const struct dir_rename_entry, ent);
107         e2 = container_of(entry_or_key, const struct dir_rename_entry, ent);
108
109         return strcmp(e1->dir, e2->dir);
110 }
111
112 static void dir_rename_init(struct hashmap *map)
113 {
114         hashmap_init(map, dir_rename_cmp, NULL, 0);
115 }
116
117 static void dir_rename_entry_init(struct dir_rename_entry *entry,
118                                   char *directory)
119 {
120         hashmap_entry_init(&entry->ent, strhash(directory));
121         entry->dir = directory;
122         entry->non_unique_new_dir = 0;
123         strbuf_init(&entry->new_dir, 0);
124         string_list_init(&entry->possible_new_dirs, 0);
125 }
126
127 struct collision_entry {
128         struct hashmap_entry ent;
129         char *target_file;
130         struct string_list source_files;
131         unsigned reported_already:1;
132 };
133
134 static struct collision_entry *collision_find_entry(struct hashmap *hashmap,
135                                                     char *target_file)
136 {
137         struct collision_entry key;
138
139         hashmap_entry_init(&key.ent, strhash(target_file));
140         key.target_file = target_file;
141         return hashmap_get_entry(hashmap, &key, ent, NULL);
142 }
143
144 static int collision_cmp(const void *unused_cmp_data,
145                          const struct hashmap_entry *eptr,
146                          const struct hashmap_entry *entry_or_key,
147                          const void *unused_keydata)
148 {
149         const struct collision_entry *e1, *e2;
150
151         e1 = container_of(eptr, const struct collision_entry, ent);
152         e2 = container_of(entry_or_key, const struct collision_entry, ent);
153
154         return strcmp(e1->target_file, e2->target_file);
155 }
156
157 static void collision_init(struct hashmap *map)
158 {
159         hashmap_init(map, collision_cmp, NULL, 0);
160 }
161
162 static void flush_output(struct merge_options *opt)
163 {
164         if (opt->buffer_output < 2 && opt->obuf.len) {
165                 fputs(opt->obuf.buf, stdout);
166                 strbuf_reset(&opt->obuf);
167         }
168 }
169
170 static int err(struct merge_options *opt, const char *err, ...)
171 {
172         va_list params;
173
174         if (opt->buffer_output < 2)
175                 flush_output(opt);
176         else {
177                 strbuf_complete(&opt->obuf, '\n');
178                 strbuf_addstr(&opt->obuf, "error: ");
179         }
180         va_start(params, err);
181         strbuf_vaddf(&opt->obuf, err, params);
182         va_end(params);
183         if (opt->buffer_output > 1)
184                 strbuf_addch(&opt->obuf, '\n');
185         else {
186                 error("%s", opt->obuf.buf);
187                 strbuf_reset(&opt->obuf);
188         }
189
190         return -1;
191 }
192
193 static struct tree *shift_tree_object(struct repository *repo,
194                                       struct tree *one, struct tree *two,
195                                       const char *subtree_shift)
196 {
197         struct object_id shifted;
198
199         if (!*subtree_shift) {
200                 shift_tree(repo, &one->object.oid, &two->object.oid, &shifted, 0);
201         } else {
202                 shift_tree_by(repo, &one->object.oid, &two->object.oid, &shifted,
203                               subtree_shift);
204         }
205         if (oideq(&two->object.oid, &shifted))
206                 return two;
207         return lookup_tree(repo, &shifted);
208 }
209
210 static inline void set_commit_tree(struct commit *c, struct tree *t)
211 {
212         c->maybe_tree = t;
213 }
214
215 static struct commit *make_virtual_commit(struct repository *repo,
216                                           struct tree *tree,
217                                           const char *comment)
218 {
219         struct commit *commit = alloc_commit_node(repo);
220
221         set_merge_remote_desc(commit, comment, (struct object *)commit);
222         set_commit_tree(commit, tree);
223         commit->object.parsed = 1;
224         return commit;
225 }
226
227 enum rename_type {
228         RENAME_NORMAL = 0,
229         RENAME_VIA_DIR,
230         RENAME_ADD,
231         RENAME_DELETE,
232         RENAME_ONE_FILE_TO_ONE,
233         RENAME_ONE_FILE_TO_TWO,
234         RENAME_TWO_FILES_TO_ONE
235 };
236
237 /*
238  * Since we want to write the index eventually, we cannot reuse the index
239  * for these (temporary) data.
240  */
241 struct stage_data {
242         struct diff_filespec stages[4]; /* mostly for oid & mode; maybe path */
243         struct rename_conflict_info *rename_conflict_info;
244         unsigned processed:1;
245 };
246
247 struct rename {
248         unsigned processed:1;
249         struct diff_filepair *pair;
250         const char *branch; /* branch that the rename occurred on */
251         /*
252          * If directory rename detection affected this rename, what was its
253          * original type ('A' or 'R') and it's original destination before
254          * the directory rename (otherwise, '\0' and NULL for these two vars).
255          */
256         char dir_rename_original_type;
257         char *dir_rename_original_dest;
258         /*
259          * Purpose of src_entry and dst_entry:
260          *
261          * If 'before' is renamed to 'after' then src_entry will contain
262          * the versions of 'before' from the merge_base, HEAD, and MERGE in
263          * stages 1, 2, and 3; dst_entry will contain the respective
264          * versions of 'after' in corresponding locations.  Thus, we have a
265          * total of six modes and oids, though some will be null.  (Stage 0
266          * is ignored; we're interested in handling conflicts.)
267          *
268          * Since we don't turn on break-rewrites by default, neither
269          * src_entry nor dst_entry can have all three of their stages have
270          * non-null oids, meaning at most four of the six will be non-null.
271          * Also, since this is a rename, both src_entry and dst_entry will
272          * have at least one non-null oid, meaning at least two will be
273          * non-null.  Of the six oids, a typical rename will have three be
274          * non-null.  Only two implies a rename/delete, and four implies a
275          * rename/add.
276          */
277         struct stage_data *src_entry;
278         struct stage_data *dst_entry;
279 };
280
281 struct rename_conflict_info {
282         enum rename_type rename_type;
283         struct rename *ren1;
284         struct rename *ren2;
285 };
286
287 static inline void setup_rename_conflict_info(enum rename_type rename_type,
288                                               struct merge_options *opt,
289                                               struct rename *ren1,
290                                               struct rename *ren2)
291 {
292         struct rename_conflict_info *ci;
293
294         /*
295          * When we have two renames involved, it's easiest to get the
296          * correct things into stage 2 and 3, and to make sure that the
297          * content merge puts HEAD before the other branch if we just
298          * ensure that branch1 == opt->branch1.  So, simply flip arguments
299          * around if we don't have that.
300          */
301         if (ren2 && ren1->branch != opt->branch1) {
302                 setup_rename_conflict_info(rename_type, opt, ren2, ren1);
303                 return;
304         }
305
306         ci = xcalloc(1, sizeof(struct rename_conflict_info));
307         ci->rename_type = rename_type;
308         ci->ren1 = ren1;
309         ci->ren2 = ren2;
310
311         ci->ren1->dst_entry->processed = 0;
312         ci->ren1->dst_entry->rename_conflict_info = ci;
313         if (ren2) {
314                 ci->ren2->dst_entry->rename_conflict_info = ci;
315         }
316 }
317
318 static int show(struct merge_options *opt, int v)
319 {
320         return (!opt->priv->call_depth && opt->verbosity >= v) ||
321                 opt->verbosity >= 5;
322 }
323
324 __attribute__((format (printf, 3, 4)))
325 static void output(struct merge_options *opt, int v, const char *fmt, ...)
326 {
327         va_list ap;
328
329         if (!show(opt, v))
330                 return;
331
332         strbuf_addchars(&opt->obuf, ' ', opt->priv->call_depth * 2);
333
334         va_start(ap, fmt);
335         strbuf_vaddf(&opt->obuf, fmt, ap);
336         va_end(ap);
337
338         strbuf_addch(&opt->obuf, '\n');
339         if (!opt->buffer_output)
340                 flush_output(opt);
341 }
342
343 static void output_commit_title(struct merge_options *opt, struct commit *commit)
344 {
345         struct merge_remote_desc *desc;
346
347         strbuf_addchars(&opt->obuf, ' ', opt->priv->call_depth * 2);
348         desc = merge_remote_util(commit);
349         if (desc)
350                 strbuf_addf(&opt->obuf, "virtual %s\n", desc->name);
351         else {
352                 strbuf_add_unique_abbrev(&opt->obuf, &commit->object.oid,
353                                          DEFAULT_ABBREV);
354                 strbuf_addch(&opt->obuf, ' ');
355                 if (parse_commit(commit) != 0)
356                         strbuf_addstr(&opt->obuf, _("(bad commit)\n"));
357                 else {
358                         const char *title;
359                         const char *msg = get_commit_buffer(commit, NULL);
360                         int len = find_commit_subject(msg, &title);
361                         if (len)
362                                 strbuf_addf(&opt->obuf, "%.*s\n", len, title);
363                         unuse_commit_buffer(commit, msg);
364                 }
365         }
366         flush_output(opt);
367 }
368
369 static int add_cacheinfo(struct merge_options *opt,
370                          const struct diff_filespec *blob,
371                          const char *path, int stage, int refresh, int options)
372 {
373         struct index_state *istate = opt->repo->index;
374         struct cache_entry *ce;
375         int ret;
376
377         ce = make_cache_entry(istate, blob->mode, &blob->oid, path, stage, 0);
378         if (!ce)
379                 return err(opt, _("add_cacheinfo failed for path '%s'; merge aborting."), path);
380
381         ret = add_index_entry(istate, ce, options);
382         if (refresh) {
383                 struct cache_entry *nce;
384
385                 nce = refresh_cache_entry(istate, ce,
386                                           CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
387                 if (!nce)
388                         return err(opt, _("add_cacheinfo failed to refresh for path '%s'; merge aborting."), path);
389                 if (nce != ce)
390                         ret = add_index_entry(istate, nce, options);
391         }
392         return ret;
393 }
394
395 static inline int merge_detect_rename(struct merge_options *opt)
396 {
397         return (opt->detect_renames >= 0) ? opt->detect_renames : 1;
398 }
399
400 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
401 {
402         parse_tree(tree);
403         init_tree_desc(desc, tree->buffer, tree->size);
404 }
405
406 static int unpack_trees_start(struct merge_options *opt,
407                               struct tree *common,
408                               struct tree *head,
409                               struct tree *merge)
410 {
411         int rc;
412         struct tree_desc t[3];
413         struct index_state tmp_index = { NULL };
414
415         memset(&opt->priv->unpack_opts, 0, sizeof(opt->priv->unpack_opts));
416         if (opt->priv->call_depth)
417                 opt->priv->unpack_opts.index_only = 1;
418         else
419                 opt->priv->unpack_opts.update = 1;
420         opt->priv->unpack_opts.merge = 1;
421         opt->priv->unpack_opts.head_idx = 2;
422         opt->priv->unpack_opts.fn = threeway_merge;
423         opt->priv->unpack_opts.src_index = opt->repo->index;
424         opt->priv->unpack_opts.dst_index = &tmp_index;
425         opt->priv->unpack_opts.aggressive = !merge_detect_rename(opt);
426         setup_unpack_trees_porcelain(&opt->priv->unpack_opts, "merge");
427
428         init_tree_desc_from_tree(t+0, common);
429         init_tree_desc_from_tree(t+1, head);
430         init_tree_desc_from_tree(t+2, merge);
431
432         rc = unpack_trees(3, t, &opt->priv->unpack_opts);
433         cache_tree_free(&opt->repo->index->cache_tree);
434
435         /*
436          * Update opt->repo->index to match the new results, AFTER saving a
437          * copy in opt->priv->orig_index.  Update src_index to point to the
438          * saved copy.  (verify_uptodate() checks src_index, and the original
439          * index is the one that had the necessary modification timestamps.)
440          */
441         opt->priv->orig_index = *opt->repo->index;
442         *opt->repo->index = tmp_index;
443         opt->priv->unpack_opts.src_index = &opt->priv->orig_index;
444
445         return rc;
446 }
447
448 static void unpack_trees_finish(struct merge_options *opt)
449 {
450         discard_index(&opt->priv->orig_index);
451         clear_unpack_trees_porcelain(&opt->priv->unpack_opts);
452 }
453
454 static int save_files_dirs(const struct object_id *oid,
455                            struct strbuf *base, const char *path,
456                            unsigned int mode, int stage, void *context)
457 {
458         struct path_hashmap_entry *entry;
459         int baselen = base->len;
460         struct merge_options *opt = context;
461
462         strbuf_addstr(base, path);
463
464         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
465         hashmap_entry_init(&entry->e, path_hash(entry->path));
466         hashmap_add(&opt->priv->current_file_dir_set, &entry->e);
467
468         strbuf_setlen(base, baselen);
469         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
470 }
471
472 static void get_files_dirs(struct merge_options *opt, struct tree *tree)
473 {
474         struct pathspec match_all;
475         memset(&match_all, 0, sizeof(match_all));
476         read_tree_recursive(opt->repo, tree, "", 0, 0,
477                             &match_all, save_files_dirs, opt);
478 }
479
480 static int get_tree_entry_if_blob(struct repository *r,
481                                   const struct object_id *tree,
482                                   const char *path,
483                                   struct diff_filespec *dfs)
484 {
485         int ret;
486
487         ret = get_tree_entry(r, tree, path, &dfs->oid, &dfs->mode);
488         if (S_ISDIR(dfs->mode)) {
489                 oidcpy(&dfs->oid, &null_oid);
490                 dfs->mode = 0;
491         }
492         return ret;
493 }
494
495 /*
496  * Returns an index_entry instance which doesn't have to correspond to
497  * a real cache entry in Git's index.
498  */
499 static struct stage_data *insert_stage_data(struct repository *r,
500                 const char *path,
501                 struct tree *o, struct tree *a, struct tree *b,
502                 struct string_list *entries)
503 {
504         struct string_list_item *item;
505         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
506         get_tree_entry_if_blob(r, &o->object.oid, path, &e->stages[1]);
507         get_tree_entry_if_blob(r, &a->object.oid, path, &e->stages[2]);
508         get_tree_entry_if_blob(r, &b->object.oid, path, &e->stages[3]);
509         item = string_list_insert(entries, path);
510         item->util = e;
511         return e;
512 }
513
514 /*
515  * Create a dictionary mapping file names to stage_data objects. The
516  * dictionary contains one entry for every path with a non-zero stage entry.
517  */
518 static struct string_list *get_unmerged(struct index_state *istate)
519 {
520         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
521         int i;
522
523         unmerged->strdup_strings = 1;
524
525         for (i = 0; i < istate->cache_nr; i++) {
526                 struct string_list_item *item;
527                 struct stage_data *e;
528                 const struct cache_entry *ce = istate->cache[i];
529                 if (!ce_stage(ce))
530                         continue;
531
532                 item = string_list_lookup(unmerged, ce->name);
533                 if (!item) {
534                         item = string_list_insert(unmerged, ce->name);
535                         item->util = xcalloc(1, sizeof(struct stage_data));
536                 }
537                 e = item->util;
538                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
539                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
540         }
541
542         return unmerged;
543 }
544
545 static int string_list_df_name_compare(const char *one, const char *two)
546 {
547         int onelen = strlen(one);
548         int twolen = strlen(two);
549         /*
550          * Here we only care that entries for D/F conflicts are
551          * adjacent, in particular with the file of the D/F conflict
552          * appearing before files below the corresponding directory.
553          * The order of the rest of the list is irrelevant for us.
554          *
555          * To achieve this, we sort with df_name_compare and provide
556          * the mode S_IFDIR so that D/F conflicts will sort correctly.
557          * We use the mode S_IFDIR for everything else for simplicity,
558          * since in other cases any changes in their order due to
559          * sorting cause no problems for us.
560          */
561         int cmp = df_name_compare(one, onelen, S_IFDIR,
562                                   two, twolen, S_IFDIR);
563         /*
564          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
565          * that 'foo' comes before 'foo/bar'.
566          */
567         if (cmp)
568                 return cmp;
569         return onelen - twolen;
570 }
571
572 static void record_df_conflict_files(struct merge_options *opt,
573                                      struct string_list *entries)
574 {
575         /* If there is a D/F conflict and the file for such a conflict
576          * currently exists in the working tree, we want to allow it to be
577          * removed to make room for the corresponding directory if needed.
578          * The files underneath the directories of such D/F conflicts will
579          * be processed before the corresponding file involved in the D/F
580          * conflict.  If the D/F directory ends up being removed by the
581          * merge, then we won't have to touch the D/F file.  If the D/F
582          * directory needs to be written to the working copy, then the D/F
583          * file will simply be removed (in make_room_for_path()) to make
584          * room for the necessary paths.  Note that if both the directory
585          * and the file need to be present, then the D/F file will be
586          * reinstated with a new unique name at the time it is processed.
587          */
588         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
589         const char *last_file = NULL;
590         int last_len = 0;
591         int i;
592
593         /*
594          * If we're merging merge-bases, we don't want to bother with
595          * any working directory changes.
596          */
597         if (opt->priv->call_depth)
598                 return;
599
600         /* Ensure D/F conflicts are adjacent in the entries list. */
601         for (i = 0; i < entries->nr; i++) {
602                 struct string_list_item *next = &entries->items[i];
603                 string_list_append(&df_sorted_entries, next->string)->util =
604                                    next->util;
605         }
606         df_sorted_entries.cmp = string_list_df_name_compare;
607         string_list_sort(&df_sorted_entries);
608
609         string_list_clear(&opt->priv->df_conflict_file_set, 1);
610         for (i = 0; i < df_sorted_entries.nr; i++) {
611                 const char *path = df_sorted_entries.items[i].string;
612                 int len = strlen(path);
613                 struct stage_data *e = df_sorted_entries.items[i].util;
614
615                 /*
616                  * Check if last_file & path correspond to a D/F conflict;
617                  * i.e. whether path is last_file+'/'+<something>.
618                  * If so, record that it's okay to remove last_file to make
619                  * room for path and friends if needed.
620                  */
621                 if (last_file &&
622                     len > last_len &&
623                     memcmp(path, last_file, last_len) == 0 &&
624                     path[last_len] == '/') {
625                         string_list_insert(&opt->priv->df_conflict_file_set, last_file);
626                 }
627
628                 /*
629                  * Determine whether path could exist as a file in the
630                  * working directory as a possible D/F conflict.  This
631                  * will only occur when it exists in stage 2 as a
632                  * file.
633                  */
634                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
635                         last_file = path;
636                         last_len = len;
637                 } else {
638                         last_file = NULL;
639                 }
640         }
641         string_list_clear(&df_sorted_entries, 0);
642 }
643
644 static int update_stages(struct merge_options *opt, const char *path,
645                          const struct diff_filespec *o,
646                          const struct diff_filespec *a,
647                          const struct diff_filespec *b)
648 {
649
650         /*
651          * NOTE: It is usually a bad idea to call update_stages on a path
652          * before calling update_file on that same path, since it can
653          * sometimes lead to spurious "refusing to lose untracked file..."
654          * messages from update_file (via make_room_for path via
655          * would_lose_untracked).  Instead, reverse the order of the calls
656          * (executing update_file first and then update_stages).
657          */
658         int clear = 1;
659         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
660         if (clear)
661                 if (remove_file_from_index(opt->repo->index, path))
662                         return -1;
663         if (o)
664                 if (add_cacheinfo(opt, o, path, 1, 0, options))
665                         return -1;
666         if (a)
667                 if (add_cacheinfo(opt, a, path, 2, 0, options))
668                         return -1;
669         if (b)
670                 if (add_cacheinfo(opt, b, path, 3, 0, options))
671                         return -1;
672         return 0;
673 }
674
675 static void update_entry(struct stage_data *entry,
676                          struct diff_filespec *o,
677                          struct diff_filespec *a,
678                          struct diff_filespec *b)
679 {
680         entry->processed = 0;
681         entry->stages[1].mode = o->mode;
682         entry->stages[2].mode = a->mode;
683         entry->stages[3].mode = b->mode;
684         oidcpy(&entry->stages[1].oid, &o->oid);
685         oidcpy(&entry->stages[2].oid, &a->oid);
686         oidcpy(&entry->stages[3].oid, &b->oid);
687 }
688
689 static int remove_file(struct merge_options *opt, int clean,
690                        const char *path, int no_wd)
691 {
692         int update_cache = opt->priv->call_depth || clean;
693         int update_working_directory = !opt->priv->call_depth && !no_wd;
694
695         if (update_cache) {
696                 if (remove_file_from_index(opt->repo->index, path))
697                         return -1;
698         }
699         if (update_working_directory) {
700                 if (ignore_case) {
701                         struct cache_entry *ce;
702                         ce = index_file_exists(opt->repo->index, path, strlen(path),
703                                                ignore_case);
704                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
705                                 return 0;
706                 }
707                 if (remove_path(path))
708                         return -1;
709         }
710         return 0;
711 }
712
713 /* add a string to a strbuf, but converting "/" to "_" */
714 static void add_flattened_path(struct strbuf *out, const char *s)
715 {
716         size_t i = out->len;
717         strbuf_addstr(out, s);
718         for (; i < out->len; i++)
719                 if (out->buf[i] == '/')
720                         out->buf[i] = '_';
721 }
722
723 static char *unique_path(struct merge_options *opt,
724                          const char *path,
725                          const char *branch)
726 {
727         struct path_hashmap_entry *entry;
728         struct strbuf newpath = STRBUF_INIT;
729         int suffix = 0;
730         size_t base_len;
731
732         strbuf_addf(&newpath, "%s~", path);
733         add_flattened_path(&newpath, branch);
734
735         base_len = newpath.len;
736         while (hashmap_get_from_hash(&opt->priv->current_file_dir_set,
737                                      path_hash(newpath.buf), newpath.buf) ||
738                (!opt->priv->call_depth && file_exists(newpath.buf))) {
739                 strbuf_setlen(&newpath, base_len);
740                 strbuf_addf(&newpath, "_%d", suffix++);
741         }
742
743         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
744         hashmap_entry_init(&entry->e, path_hash(entry->path));
745         hashmap_add(&opt->priv->current_file_dir_set, &entry->e);
746         return strbuf_detach(&newpath, NULL);
747 }
748
749 /**
750  * Check whether a directory in the index is in the way of an incoming
751  * file.  Return 1 if so.  If check_working_copy is non-zero, also
752  * check the working directory.  If empty_ok is non-zero, also return
753  * 0 in the case where the working-tree dir exists but is empty.
754  */
755 static int dir_in_way(struct index_state *istate, const char *path,
756                       int check_working_copy, int empty_ok)
757 {
758         int pos;
759         struct strbuf dirpath = STRBUF_INIT;
760         struct stat st;
761
762         strbuf_addstr(&dirpath, path);
763         strbuf_addch(&dirpath, '/');
764
765         pos = index_name_pos(istate, dirpath.buf, dirpath.len);
766
767         if (pos < 0)
768                 pos = -1 - pos;
769         if (pos < istate->cache_nr &&
770             !strncmp(dirpath.buf, istate->cache[pos]->name, dirpath.len)) {
771                 strbuf_release(&dirpath);
772                 return 1;
773         }
774
775         strbuf_release(&dirpath);
776         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
777                 !(empty_ok && is_empty_dir(path)) &&
778                 !has_symlink_leading_path(path, strlen(path));
779 }
780
781 /*
782  * Returns whether path was tracked in the index before the merge started,
783  * and its oid and mode match the specified values
784  */
785 static int was_tracked_and_matches(struct merge_options *opt, const char *path,
786                                    const struct diff_filespec *blob)
787 {
788         int pos = index_name_pos(&opt->priv->orig_index, path, strlen(path));
789         struct cache_entry *ce;
790
791         if (0 > pos)
792                 /* we were not tracking this path before the merge */
793                 return 0;
794
795         /* See if the file we were tracking before matches */
796         ce = opt->priv->orig_index.cache[pos];
797         return (oideq(&ce->oid, &blob->oid) && ce->ce_mode == blob->mode);
798 }
799
800 /*
801  * Returns whether path was tracked in the index before the merge started
802  */
803 static int was_tracked(struct merge_options *opt, const char *path)
804 {
805         int pos = index_name_pos(&opt->priv->orig_index, path, strlen(path));
806
807         if (0 <= pos)
808                 /* we were tracking this path before the merge */
809                 return 1;
810
811         return 0;
812 }
813
814 static int would_lose_untracked(struct merge_options *opt, const char *path)
815 {
816         struct index_state *istate = opt->repo->index;
817
818         /*
819          * This may look like it can be simplified to:
820          *   return !was_tracked(opt, path) && file_exists(path)
821          * but it can't.  This function needs to know whether path was in
822          * the working tree due to EITHER having been tracked in the index
823          * before the merge OR having been put into the working copy and
824          * index by unpack_trees().  Due to that either-or requirement, we
825          * check the current index instead of the original one.
826          *
827          * Note that we do not need to worry about merge-recursive itself
828          * updating the index after unpack_trees() and before calling this
829          * function, because we strictly require all code paths in
830          * merge-recursive to update the working tree first and the index
831          * second.  Doing otherwise would break
832          * update_file()/would_lose_untracked(); see every comment in this
833          * file which mentions "update_stages".
834          */
835         int pos = index_name_pos(istate, path, strlen(path));
836
837         if (pos < 0)
838                 pos = -1 - pos;
839         while (pos < istate->cache_nr &&
840                !strcmp(path, istate->cache[pos]->name)) {
841                 /*
842                  * If stage #0, it is definitely tracked.
843                  * If it has stage #2 then it was tracked
844                  * before this merge started.  All other
845                  * cases the path was not tracked.
846                  */
847                 switch (ce_stage(istate->cache[pos])) {
848                 case 0:
849                 case 2:
850                         return 0;
851                 }
852                 pos++;
853         }
854         return file_exists(path);
855 }
856
857 static int was_dirty(struct merge_options *opt, const char *path)
858 {
859         struct cache_entry *ce;
860         int dirty = 1;
861
862         if (opt->priv->call_depth || !was_tracked(opt, path))
863                 return !dirty;
864
865         ce = index_file_exists(opt->priv->unpack_opts.src_index,
866                                path, strlen(path), ignore_case);
867         dirty = verify_uptodate(ce, &opt->priv->unpack_opts) != 0;
868         return dirty;
869 }
870
871 static int make_room_for_path(struct merge_options *opt, const char *path)
872 {
873         int status, i;
874         const char *msg = _("failed to create path '%s'%s");
875
876         /* Unlink any D/F conflict files that are in the way */
877         for (i = 0; i < opt->priv->df_conflict_file_set.nr; i++) {
878                 const char *df_path = opt->priv->df_conflict_file_set.items[i].string;
879                 size_t pathlen = strlen(path);
880                 size_t df_pathlen = strlen(df_path);
881                 if (df_pathlen < pathlen &&
882                     path[df_pathlen] == '/' &&
883                     strncmp(path, df_path, df_pathlen) == 0) {
884                         output(opt, 3,
885                                _("Removing %s to make room for subdirectory\n"),
886                                df_path);
887                         unlink(df_path);
888                         unsorted_string_list_delete_item(&opt->priv->df_conflict_file_set,
889                                                          i, 0);
890                         break;
891                 }
892         }
893
894         /* Make sure leading directories are created */
895         status = safe_create_leading_directories_const(path);
896         if (status) {
897                 if (status == SCLD_EXISTS)
898                         /* something else exists */
899                         return err(opt, msg, path, _(": perhaps a D/F conflict?"));
900                 return err(opt, msg, path, "");
901         }
902
903         /*
904          * Do not unlink a file in the work tree if we are not
905          * tracking it.
906          */
907         if (would_lose_untracked(opt, path))
908                 return err(opt, _("refusing to lose untracked file at '%s'"),
909                            path);
910
911         /* Successful unlink is good.. */
912         if (!unlink(path))
913                 return 0;
914         /* .. and so is no existing file */
915         if (errno == ENOENT)
916                 return 0;
917         /* .. but not some other error (who really cares what?) */
918         return err(opt, msg, path, _(": perhaps a D/F conflict?"));
919 }
920
921 static int update_file_flags(struct merge_options *opt,
922                              const struct diff_filespec *contents,
923                              const char *path,
924                              int update_cache,
925                              int update_wd)
926 {
927         int ret = 0;
928
929         if (opt->priv->call_depth)
930                 update_wd = 0;
931
932         if (update_wd) {
933                 enum object_type type;
934                 void *buf;
935                 unsigned long size;
936
937                 if (S_ISGITLINK(contents->mode)) {
938                         /*
939                          * We may later decide to recursively descend into
940                          * the submodule directory and update its index
941                          * and/or work tree, but we do not do that now.
942                          */
943                         update_wd = 0;
944                         goto update_index;
945                 }
946
947                 buf = read_object_file(&contents->oid, &type, &size);
948                 if (!buf) {
949                         ret = err(opt, _("cannot read object %s '%s'"),
950                                   oid_to_hex(&contents->oid), path);
951                         goto free_buf;
952                 }
953                 if (type != OBJ_BLOB) {
954                         ret = err(opt, _("blob expected for %s '%s'"),
955                                   oid_to_hex(&contents->oid), path);
956                         goto free_buf;
957                 }
958                 if (S_ISREG(contents->mode)) {
959                         struct strbuf strbuf = STRBUF_INIT;
960                         if (convert_to_working_tree(opt->repo->index,
961                                                     path, buf, size, &strbuf)) {
962                                 free(buf);
963                                 size = strbuf.len;
964                                 buf = strbuf_detach(&strbuf, NULL);
965                         }
966                 }
967
968                 if (make_room_for_path(opt, path) < 0) {
969                         update_wd = 0;
970                         goto free_buf;
971                 }
972                 if (S_ISREG(contents->mode) ||
973                     (!has_symlinks && S_ISLNK(contents->mode))) {
974                         int fd;
975                         int mode = (contents->mode & 0100 ? 0777 : 0666);
976
977                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
978                         if (fd < 0) {
979                                 ret = err(opt, _("failed to open '%s': %s"),
980                                           path, strerror(errno));
981                                 goto free_buf;
982                         }
983                         write_in_full(fd, buf, size);
984                         close(fd);
985                 } else if (S_ISLNK(contents->mode)) {
986                         char *lnk = xmemdupz(buf, size);
987                         safe_create_leading_directories_const(path);
988                         unlink(path);
989                         if (symlink(lnk, path))
990                                 ret = err(opt, _("failed to symlink '%s': %s"),
991                                           path, strerror(errno));
992                         free(lnk);
993                 } else
994                         ret = err(opt,
995                                   _("do not know what to do with %06o %s '%s'"),
996                                   contents->mode, oid_to_hex(&contents->oid), path);
997         free_buf:
998                 free(buf);
999         }
1000 update_index:
1001         if (!ret && update_cache)
1002                 if (add_cacheinfo(opt, contents, path, 0, update_wd,
1003                                   ADD_CACHE_OK_TO_ADD))
1004                         return -1;
1005         return ret;
1006 }
1007
1008 static int update_file(struct merge_options *opt,
1009                        int clean,
1010                        const struct diff_filespec *contents,
1011                        const char *path)
1012 {
1013         return update_file_flags(opt, contents, path,
1014                                  opt->priv->call_depth || clean, !opt->priv->call_depth);
1015 }
1016
1017 /* Low level file merging, update and removal */
1018
1019 struct merge_file_info {
1020         struct diff_filespec blob; /* mostly use oid & mode; sometimes path */
1021         unsigned clean:1,
1022                  merge:1;
1023 };
1024
1025 static int merge_3way(struct merge_options *opt,
1026                       mmbuffer_t *result_buf,
1027                       const struct diff_filespec *o,
1028                       const struct diff_filespec *a,
1029                       const struct diff_filespec *b,
1030                       const char *branch1,
1031                       const char *branch2,
1032                       const int extra_marker_size)
1033 {
1034         mmfile_t orig, src1, src2;
1035         struct ll_merge_options ll_opts = {0};
1036         char *base, *name1, *name2;
1037         int merge_status;
1038
1039         ll_opts.renormalize = opt->renormalize;
1040         ll_opts.extra_marker_size = extra_marker_size;
1041         ll_opts.xdl_opts = opt->xdl_opts;
1042
1043         if (opt->priv->call_depth) {
1044                 ll_opts.virtual_ancestor = 1;
1045                 ll_opts.variant = 0;
1046         } else {
1047                 switch (opt->recursive_variant) {
1048                 case MERGE_VARIANT_OURS:
1049                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
1050                         break;
1051                 case MERGE_VARIANT_THEIRS:
1052                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
1053                         break;
1054                 default:
1055                         ll_opts.variant = 0;
1056                         break;
1057                 }
1058         }
1059
1060         assert(a->path && b->path && o->path && opt->ancestor);
1061         if (strcmp(a->path, b->path) || strcmp(a->path, o->path) != 0) {
1062                 base  = mkpathdup("%s:%s", opt->ancestor, o->path);
1063                 name1 = mkpathdup("%s:%s", branch1, a->path);
1064                 name2 = mkpathdup("%s:%s", branch2, b->path);
1065         } else {
1066                 base  = mkpathdup("%s", opt->ancestor);
1067                 name1 = mkpathdup("%s", branch1);
1068                 name2 = mkpathdup("%s", branch2);
1069         }
1070
1071         read_mmblob(&orig, &o->oid);
1072         read_mmblob(&src1, &a->oid);
1073         read_mmblob(&src2, &b->oid);
1074
1075         merge_status = ll_merge(result_buf, a->path, &orig, base,
1076                                 &src1, name1, &src2, name2,
1077                                 opt->repo->index, &ll_opts);
1078
1079         free(base);
1080         free(name1);
1081         free(name2);
1082         free(orig.ptr);
1083         free(src1.ptr);
1084         free(src2.ptr);
1085         return merge_status;
1086 }
1087
1088 static int find_first_merges(struct repository *repo,
1089                              struct object_array *result, const char *path,
1090                              struct commit *a, struct commit *b)
1091 {
1092         int i, j;
1093         struct object_array merges = OBJECT_ARRAY_INIT;
1094         struct commit *commit;
1095         int contains_another;
1096
1097         char merged_revision[GIT_MAX_HEXSZ + 2];
1098         const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
1099                                    "--all", merged_revision, NULL };
1100         struct rev_info revs;
1101         struct setup_revision_opt rev_opts;
1102
1103         memset(result, 0, sizeof(struct object_array));
1104         memset(&rev_opts, 0, sizeof(rev_opts));
1105
1106         /* get all revisions that merge commit a */
1107         xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
1108                   oid_to_hex(&a->object.oid));
1109         repo_init_revisions(repo, &revs, NULL);
1110         rev_opts.submodule = path;
1111         /* FIXME: can't handle linked worktrees in submodules yet */
1112         revs.single_worktree = path != NULL;
1113         setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
1114
1115         /* save all revisions from the above list that contain b */
1116         if (prepare_revision_walk(&revs))
1117                 die("revision walk setup failed");
1118         while ((commit = get_revision(&revs)) != NULL) {
1119                 struct object *o = &(commit->object);
1120                 if (in_merge_bases(b, commit))
1121                         add_object_array(o, NULL, &merges);
1122         }
1123         reset_revision_walk();
1124
1125         /* Now we've got all merges that contain a and b. Prune all
1126          * merges that contain another found merge and save them in
1127          * result.
1128          */
1129         for (i = 0; i < merges.nr; i++) {
1130                 struct commit *m1 = (struct commit *) merges.objects[i].item;
1131
1132                 contains_another = 0;
1133                 for (j = 0; j < merges.nr; j++) {
1134                         struct commit *m2 = (struct commit *) merges.objects[j].item;
1135                         if (i != j && in_merge_bases(m2, m1)) {
1136                                 contains_another = 1;
1137                                 break;
1138                         }
1139                 }
1140
1141                 if (!contains_another)
1142                         add_object_array(merges.objects[i].item, NULL, result);
1143         }
1144
1145         object_array_clear(&merges);
1146         return result->nr;
1147 }
1148
1149 static void print_commit(struct commit *commit)
1150 {
1151         struct strbuf sb = STRBUF_INIT;
1152         struct pretty_print_context ctx = {0};
1153         ctx.date_mode.type = DATE_NORMAL;
1154         format_commit_message(commit, " %h: %m %s", &sb, &ctx);
1155         fprintf(stderr, "%s\n", sb.buf);
1156         strbuf_release(&sb);
1157 }
1158
1159 static int is_valid(const struct diff_filespec *dfs)
1160 {
1161         return dfs->mode != 0 && !is_null_oid(&dfs->oid);
1162 }
1163
1164 static int merge_submodule(struct merge_options *opt,
1165                            struct object_id *result, const char *path,
1166                            const struct object_id *base, const struct object_id *a,
1167                            const struct object_id *b)
1168 {
1169         struct commit *commit_base, *commit_a, *commit_b;
1170         int parent_count;
1171         struct object_array merges;
1172
1173         int i;
1174         int search = !opt->priv->call_depth;
1175
1176         /* store a in result in case we fail */
1177         oidcpy(result, a);
1178
1179         /* we can not handle deletion conflicts */
1180         if (is_null_oid(base))
1181                 return 0;
1182         if (is_null_oid(a))
1183                 return 0;
1184         if (is_null_oid(b))
1185                 return 0;
1186
1187         if (add_submodule_odb(path)) {
1188                 output(opt, 1, _("Failed to merge submodule %s (not checked out)"), path);
1189                 return 0;
1190         }
1191
1192         if (!(commit_base = lookup_commit_reference(opt->repo, base)) ||
1193             !(commit_a = lookup_commit_reference(opt->repo, a)) ||
1194             !(commit_b = lookup_commit_reference(opt->repo, b))) {
1195                 output(opt, 1, _("Failed to merge submodule %s (commits not present)"), path);
1196                 return 0;
1197         }
1198
1199         /* check whether both changes are forward */
1200         if (!in_merge_bases(commit_base, commit_a) ||
1201             !in_merge_bases(commit_base, commit_b)) {
1202                 output(opt, 1, _("Failed to merge submodule %s (commits don't follow merge-base)"), path);
1203                 return 0;
1204         }
1205
1206         /* Case #1: a is contained in b or vice versa */
1207         if (in_merge_bases(commit_a, commit_b)) {
1208                 oidcpy(result, b);
1209                 if (show(opt, 3)) {
1210                         output(opt, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1211                         output_commit_title(opt, commit_b);
1212                 } else if (show(opt, 2))
1213                         output(opt, 2, _("Fast-forwarding submodule %s"), path);
1214                 else
1215                         ; /* no output */
1216
1217                 return 1;
1218         }
1219         if (in_merge_bases(commit_b, commit_a)) {
1220                 oidcpy(result, a);
1221                 if (show(opt, 3)) {
1222                         output(opt, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1223                         output_commit_title(opt, commit_a);
1224                 } else if (show(opt, 2))
1225                         output(opt, 2, _("Fast-forwarding submodule %s"), path);
1226                 else
1227                         ; /* no output */
1228
1229                 return 1;
1230         }
1231
1232         /*
1233          * Case #2: There are one or more merges that contain a and b in
1234          * the submodule. If there is only one, then present it as a
1235          * suggestion to the user, but leave it marked unmerged so the
1236          * user needs to confirm the resolution.
1237          */
1238
1239         /* Skip the search if makes no sense to the calling context.  */
1240         if (!search)
1241                 return 0;
1242
1243         /* find commit which merges them */
1244         parent_count = find_first_merges(opt->repo, &merges, path,
1245                                          commit_a, commit_b);
1246         switch (parent_count) {
1247         case 0:
1248                 output(opt, 1, _("Failed to merge submodule %s (merge following commits not found)"), path);
1249                 break;
1250
1251         case 1:
1252                 output(opt, 1, _("Failed to merge submodule %s (not fast-forward)"), path);
1253                 output(opt, 2, _("Found a possible merge resolution for the submodule:\n"));
1254                 print_commit((struct commit *) merges.objects[0].item);
1255                 output(opt, 2, _(
1256                        "If this is correct simply add it to the index "
1257                        "for example\n"
1258                        "by using:\n\n"
1259                        "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
1260                        "which will accept this suggestion.\n"),
1261                        oid_to_hex(&merges.objects[0].item->oid), path);
1262                 break;
1263
1264         default:
1265                 output(opt, 1, _("Failed to merge submodule %s (multiple merges found)"), path);
1266                 for (i = 0; i < merges.nr; i++)
1267                         print_commit((struct commit *) merges.objects[i].item);
1268         }
1269
1270         object_array_clear(&merges);
1271         return 0;
1272 }
1273
1274 static int merge_mode_and_contents(struct merge_options *opt,
1275                                    const struct diff_filespec *o,
1276                                    const struct diff_filespec *a,
1277                                    const struct diff_filespec *b,
1278                                    const char *filename,
1279                                    const char *branch1,
1280                                    const char *branch2,
1281                                    const int extra_marker_size,
1282                                    struct merge_file_info *result)
1283 {
1284         if (opt->branch1 != branch1) {
1285                 /*
1286                  * It's weird getting a reverse merge with HEAD on the bottom
1287                  * side of the conflict markers and the other branch on the
1288                  * top.  Fix that.
1289                  */
1290                 return merge_mode_and_contents(opt, o, b, a,
1291                                                filename,
1292                                                branch2, branch1,
1293                                                extra_marker_size, result);
1294         }
1295
1296         result->merge = 0;
1297         result->clean = 1;
1298
1299         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
1300                 result->clean = 0;
1301                 if (S_ISREG(a->mode)) {
1302                         result->blob.mode = a->mode;
1303                         oidcpy(&result->blob.oid, &a->oid);
1304                 } else {
1305                         result->blob.mode = b->mode;
1306                         oidcpy(&result->blob.oid, &b->oid);
1307                 }
1308         } else {
1309                 if (!oideq(&a->oid, &o->oid) && !oideq(&b->oid, &o->oid))
1310                         result->merge = 1;
1311
1312                 /*
1313                  * Merge modes
1314                  */
1315                 if (a->mode == b->mode || a->mode == o->mode)
1316                         result->blob.mode = b->mode;
1317                 else {
1318                         result->blob.mode = a->mode;
1319                         if (b->mode != o->mode) {
1320                                 result->clean = 0;
1321                                 result->merge = 1;
1322                         }
1323                 }
1324
1325                 if (oideq(&a->oid, &b->oid) || oideq(&a->oid, &o->oid))
1326                         oidcpy(&result->blob.oid, &b->oid);
1327                 else if (oideq(&b->oid, &o->oid))
1328                         oidcpy(&result->blob.oid, &a->oid);
1329                 else if (S_ISREG(a->mode)) {
1330                         mmbuffer_t result_buf;
1331                         int ret = 0, merge_status;
1332
1333                         merge_status = merge_3way(opt, &result_buf, o, a, b,
1334                                                   branch1, branch2,
1335                                                   extra_marker_size);
1336
1337                         if ((merge_status < 0) || !result_buf.ptr)
1338                                 ret = err(opt, _("Failed to execute internal merge"));
1339
1340                         if (!ret &&
1341                             write_object_file(result_buf.ptr, result_buf.size,
1342                                               blob_type, &result->blob.oid))
1343                                 ret = err(opt, _("Unable to add %s to database"),
1344                                           a->path);
1345
1346                         free(result_buf.ptr);
1347                         if (ret)
1348                                 return ret;
1349                         result->clean = (merge_status == 0);
1350                 } else if (S_ISGITLINK(a->mode)) {
1351                         result->clean = merge_submodule(opt, &result->blob.oid,
1352                                                         o->path,
1353                                                         &o->oid,
1354                                                         &a->oid,
1355                                                         &b->oid);
1356                 } else if (S_ISLNK(a->mode)) {
1357                         switch (opt->recursive_variant) {
1358                         case MERGE_VARIANT_NORMAL:
1359                                 oidcpy(&result->blob.oid, &a->oid);
1360                                 if (!oideq(&a->oid, &b->oid))
1361                                         result->clean = 0;
1362                                 break;
1363                         case MERGE_VARIANT_OURS:
1364                                 oidcpy(&result->blob.oid, &a->oid);
1365                                 break;
1366                         case MERGE_VARIANT_THEIRS:
1367                                 oidcpy(&result->blob.oid, &b->oid);
1368                                 break;
1369                         }
1370                 } else
1371                         BUG("unsupported object type in the tree");
1372         }
1373
1374         if (result->merge)
1375                 output(opt, 2, _("Auto-merging %s"), filename);
1376
1377         return 0;
1378 }
1379
1380 static int handle_rename_via_dir(struct merge_options *opt,
1381                                  struct rename_conflict_info *ci)
1382 {
1383         /*
1384          * Handle file adds that need to be renamed due to directory rename
1385          * detection.  This differs from handle_rename_normal, because
1386          * there is no content merge to do; just move the file into the
1387          * desired final location.
1388          */
1389         const struct rename *ren = ci->ren1;
1390         const struct diff_filespec *dest = ren->pair->two;
1391         char *file_path = dest->path;
1392         int mark_conflicted = (opt->detect_directory_renames ==
1393                                MERGE_DIRECTORY_RENAMES_CONFLICT);
1394         assert(ren->dir_rename_original_dest);
1395
1396         if (!opt->priv->call_depth && would_lose_untracked(opt, dest->path)) {
1397                 mark_conflicted = 1;
1398                 file_path = unique_path(opt, dest->path, ren->branch);
1399                 output(opt, 1, _("Error: Refusing to lose untracked file at %s; "
1400                                  "writing to %s instead."),
1401                        dest->path, file_path);
1402         }
1403
1404         if (mark_conflicted) {
1405                 /*
1406                  * Write the file in worktree at file_path.  In the index,
1407                  * only record the file at dest->path in the appropriate
1408                  * higher stage.
1409                  */
1410                 if (update_file(opt, 0, dest, file_path))
1411                         return -1;
1412                 if (file_path != dest->path)
1413                         free(file_path);
1414                 if (update_stages(opt, dest->path, NULL,
1415                                   ren->branch == opt->branch1 ? dest : NULL,
1416                                   ren->branch == opt->branch1 ? NULL : dest))
1417                         return -1;
1418                 return 0; /* not clean, but conflicted */
1419         } else {
1420                 /* Update dest->path both in index and in worktree */
1421                 if (update_file(opt, 1, dest, dest->path))
1422                         return -1;
1423                 return 1; /* clean */
1424         }
1425 }
1426
1427 static int handle_change_delete(struct merge_options *opt,
1428                                 const char *path, const char *old_path,
1429                                 const struct diff_filespec *o,
1430                                 const struct diff_filespec *changed,
1431                                 const char *change_branch,
1432                                 const char *delete_branch,
1433                                 const char *change, const char *change_past)
1434 {
1435         char *alt_path = NULL;
1436         const char *update_path = path;
1437         int ret = 0;
1438
1439         if (dir_in_way(opt->repo->index, path, !opt->priv->call_depth, 0) ||
1440             (!opt->priv->call_depth && would_lose_untracked(opt, path))) {
1441                 update_path = alt_path = unique_path(opt, path, change_branch);
1442         }
1443
1444         if (opt->priv->call_depth) {
1445                 /*
1446                  * We cannot arbitrarily accept either a_sha or b_sha as
1447                  * correct; since there is no true "middle point" between
1448                  * them, simply reuse the base version for virtual merge base.
1449                  */
1450                 ret = remove_file_from_index(opt->repo->index, path);
1451                 if (!ret)
1452                         ret = update_file(opt, 0, o, update_path);
1453         } else {
1454                 /*
1455                  * Despite the four nearly duplicate messages and argument
1456                  * lists below and the ugliness of the nested if-statements,
1457                  * having complete messages makes the job easier for
1458                  * translators.
1459                  *
1460                  * The slight variance among the cases is due to the fact
1461                  * that:
1462                  *   1) directory/file conflicts (in effect if
1463                  *      !alt_path) could cause us to need to write the
1464                  *      file to a different path.
1465                  *   2) renames (in effect if !old_path) could mean that
1466                  *      there are two names for the path that the user
1467                  *      may know the file by.
1468                  */
1469                 if (!alt_path) {
1470                         if (!old_path) {
1471                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1472                                        "and %s in %s. Version %s of %s left in tree."),
1473                                        change, path, delete_branch, change_past,
1474                                        change_branch, change_branch, path);
1475                         } else {
1476                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1477                                        "and %s to %s in %s. Version %s of %s left in tree."),
1478                                        change, old_path, delete_branch, change_past, path,
1479                                        change_branch, change_branch, path);
1480                         }
1481                 } else {
1482                         if (!old_path) {
1483                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1484                                        "and %s in %s. Version %s of %s left in tree at %s."),
1485                                        change, path, delete_branch, change_past,
1486                                        change_branch, change_branch, path, alt_path);
1487                         } else {
1488                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1489                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1490                                        change, old_path, delete_branch, change_past, path,
1491                                        change_branch, change_branch, path, alt_path);
1492                         }
1493                 }
1494                 /*
1495                  * No need to call update_file() on path when change_branch ==
1496                  * opt->branch1 && !alt_path, since that would needlessly touch
1497                  * path.  We could call update_file_flags() with update_cache=0
1498                  * and update_wd=0, but that's a no-op.
1499                  */
1500                 if (change_branch != opt->branch1 || alt_path)
1501                         ret = update_file(opt, 0, changed, update_path);
1502         }
1503         free(alt_path);
1504
1505         return ret;
1506 }
1507
1508 static int handle_rename_delete(struct merge_options *opt,
1509                                 struct rename_conflict_info *ci)
1510 {
1511         const struct rename *ren = ci->ren1;
1512         const struct diff_filespec *orig = ren->pair->one;
1513         const struct diff_filespec *dest = ren->pair->two;
1514         const char *rename_branch = ren->branch;
1515         const char *delete_branch = (opt->branch1 == ren->branch ?
1516                                      opt->branch2 : opt->branch1);
1517
1518         if (handle_change_delete(opt,
1519                                  opt->priv->call_depth ? orig->path : dest->path,
1520                                  opt->priv->call_depth ? NULL : orig->path,
1521                                  orig, dest,
1522                                  rename_branch, delete_branch,
1523                                  _("rename"), _("renamed")))
1524                 return -1;
1525
1526         if (opt->priv->call_depth)
1527                 return remove_file_from_index(opt->repo->index, dest->path);
1528         else
1529                 return update_stages(opt, dest->path, NULL,
1530                                      rename_branch == opt->branch1 ? dest : NULL,
1531                                      rename_branch == opt->branch1 ? NULL : dest);
1532 }
1533
1534 static int handle_file_collision(struct merge_options *opt,
1535                                  const char *collide_path,
1536                                  const char *prev_path1,
1537                                  const char *prev_path2,
1538                                  const char *branch1, const char *branch2,
1539                                  struct diff_filespec *a,
1540                                  struct diff_filespec *b)
1541 {
1542         struct merge_file_info mfi;
1543         struct diff_filespec null;
1544         char *alt_path = NULL;
1545         const char *update_path = collide_path;
1546
1547         /*
1548          * It's easiest to get the correct things into stage 2 and 3, and
1549          * to make sure that the content merge puts HEAD before the other
1550          * branch if we just ensure that branch1 == opt->branch1.  So, simply
1551          * flip arguments around if we don't have that.
1552          */
1553         if (branch1 != opt->branch1) {
1554                 return handle_file_collision(opt, collide_path,
1555                                              prev_path2, prev_path1,
1556                                              branch2, branch1,
1557                                              b, a);
1558         }
1559
1560         /*
1561          * In the recursive case, we just opt to undo renames
1562          */
1563         if (opt->priv->call_depth && (prev_path1 || prev_path2)) {
1564                 /* Put first file (a->oid, a->mode) in its original spot */
1565                 if (prev_path1) {
1566                         if (update_file(opt, 1, a, prev_path1))
1567                                 return -1;
1568                 } else {
1569                         if (update_file(opt, 1, a, collide_path))
1570                                 return -1;
1571                 }
1572
1573                 /* Put second file (b->oid, b->mode) in its original spot */
1574                 if (prev_path2) {
1575                         if (update_file(opt, 1, b, prev_path2))
1576                                 return -1;
1577                 } else {
1578                         if (update_file(opt, 1, b, collide_path))
1579                                 return -1;
1580                 }
1581
1582                 /* Don't leave something at collision path if unrenaming both */
1583                 if (prev_path1 && prev_path2)
1584                         remove_file(opt, 1, collide_path, 0);
1585
1586                 return 0;
1587         }
1588
1589         /* Remove rename sources if rename/add or rename/rename(2to1) */
1590         if (prev_path1)
1591                 remove_file(opt, 1, prev_path1,
1592                             opt->priv->call_depth || would_lose_untracked(opt, prev_path1));
1593         if (prev_path2)
1594                 remove_file(opt, 1, prev_path2,
1595                             opt->priv->call_depth || would_lose_untracked(opt, prev_path2));
1596
1597         /*
1598          * Remove the collision path, if it wouldn't cause dirty contents
1599          * or an untracked file to get lost.  We'll either overwrite with
1600          * merged contents, or just write out to differently named files.
1601          */
1602         if (was_dirty(opt, collide_path)) {
1603                 output(opt, 1, _("Refusing to lose dirty file at %s"),
1604                        collide_path);
1605                 update_path = alt_path = unique_path(opt, collide_path, "merged");
1606         } else if (would_lose_untracked(opt, collide_path)) {
1607                 /*
1608                  * Only way we get here is if both renames were from
1609                  * a directory rename AND user had an untracked file
1610                  * at the location where both files end up after the
1611                  * two directory renames.  See testcase 10d of t6043.
1612                  */
1613                 output(opt, 1, _("Refusing to lose untracked file at "
1614                                "%s, even though it's in the way."),
1615                        collide_path);
1616                 update_path = alt_path = unique_path(opt, collide_path, "merged");
1617         } else {
1618                 /*
1619                  * FIXME: It's possible that the two files are identical
1620                  * and that the current working copy happens to match, in
1621                  * which case we are unnecessarily touching the working
1622                  * tree file.  It's not a likely enough scenario that I
1623                  * want to code up the checks for it and a better fix is
1624                  * available if we restructure how unpack_trees() and
1625                  * merge-recursive interoperate anyway, so punting for
1626                  * now...
1627                  */
1628                 remove_file(opt, 0, collide_path, 0);
1629         }
1630
1631         /* Store things in diff_filespecs for functions that need it */
1632         null.path = (char *)collide_path;
1633         oidcpy(&null.oid, &null_oid);
1634         null.mode = 0;
1635
1636         if (merge_mode_and_contents(opt, &null, a, b, collide_path,
1637                                     branch1, branch2, opt->priv->call_depth * 2, &mfi))
1638                 return -1;
1639         mfi.clean &= !alt_path;
1640         if (update_file(opt, mfi.clean, &mfi.blob, update_path))
1641                 return -1;
1642         if (!mfi.clean && !opt->priv->call_depth &&
1643             update_stages(opt, collide_path, NULL, a, b))
1644                 return -1;
1645         free(alt_path);
1646         /*
1647          * FIXME: If both a & b both started with conflicts (only possible
1648          * if they came from a rename/rename(2to1)), but had IDENTICAL
1649          * contents including those conflicts, then in the next line we claim
1650          * it was clean.  If someone cares about this case, we should have the
1651          * caller notify us if we started with conflicts.
1652          */
1653         return mfi.clean;
1654 }
1655
1656 static int handle_rename_add(struct merge_options *opt,
1657                              struct rename_conflict_info *ci)
1658 {
1659         /* a was renamed to c, and a separate c was added. */
1660         struct diff_filespec *a = ci->ren1->pair->one;
1661         struct diff_filespec *c = ci->ren1->pair->two;
1662         char *path = c->path;
1663         char *prev_path_desc;
1664         struct merge_file_info mfi;
1665
1666         const char *rename_branch = ci->ren1->branch;
1667         const char *add_branch = (opt->branch1 == rename_branch ?
1668                                   opt->branch2 : opt->branch1);
1669         int other_stage = (ci->ren1->branch == opt->branch1 ? 3 : 2);
1670
1671         output(opt, 1, _("CONFLICT (rename/add): "
1672                "Rename %s->%s in %s.  Added %s in %s"),
1673                a->path, c->path, rename_branch,
1674                c->path, add_branch);
1675
1676         prev_path_desc = xstrfmt("version of %s from %s", path, a->path);
1677         ci->ren1->src_entry->stages[other_stage].path = a->path;
1678         if (merge_mode_and_contents(opt, a, c,
1679                                     &ci->ren1->src_entry->stages[other_stage],
1680                                     prev_path_desc,
1681                                     opt->branch1, opt->branch2,
1682                                     1 + opt->priv->call_depth * 2, &mfi))
1683                 return -1;
1684         free(prev_path_desc);
1685
1686         ci->ren1->dst_entry->stages[other_stage].path = mfi.blob.path = c->path;
1687         return handle_file_collision(opt,
1688                                      c->path, a->path, NULL,
1689                                      rename_branch, add_branch,
1690                                      &mfi.blob,
1691                                      &ci->ren1->dst_entry->stages[other_stage]);
1692 }
1693
1694 static char *find_path_for_conflict(struct merge_options *opt,
1695                                     const char *path,
1696                                     const char *branch1,
1697                                     const char *branch2)
1698 {
1699         char *new_path = NULL;
1700         if (dir_in_way(opt->repo->index, path, !opt->priv->call_depth, 0)) {
1701                 new_path = unique_path(opt, path, branch1);
1702                 output(opt, 1, _("%s is a directory in %s adding "
1703                                "as %s instead"),
1704                        path, branch2, new_path);
1705         } else if (would_lose_untracked(opt, path)) {
1706                 new_path = unique_path(opt, path, branch1);
1707                 output(opt, 1, _("Refusing to lose untracked file"
1708                                " at %s; adding as %s instead"),
1709                        path, new_path);
1710         }
1711
1712         return new_path;
1713 }
1714
1715 static int handle_rename_rename_1to2(struct merge_options *opt,
1716                                      struct rename_conflict_info *ci)
1717 {
1718         /* One file was renamed in both branches, but to different names. */
1719         struct merge_file_info mfi;
1720         struct diff_filespec *add;
1721         struct diff_filespec *o = ci->ren1->pair->one;
1722         struct diff_filespec *a = ci->ren1->pair->two;
1723         struct diff_filespec *b = ci->ren2->pair->two;
1724         char *path_desc;
1725
1726         output(opt, 1, _("CONFLICT (rename/rename): "
1727                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1728                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1729                o->path, a->path, ci->ren1->branch,
1730                o->path, b->path, ci->ren2->branch,
1731                opt->priv->call_depth ? _(" (left unresolved)") : "");
1732
1733         path_desc = xstrfmt("%s and %s, both renamed from %s",
1734                             a->path, b->path, o->path);
1735         if (merge_mode_and_contents(opt, o, a, b, path_desc,
1736                                     ci->ren1->branch, ci->ren2->branch,
1737                                     opt->priv->call_depth * 2, &mfi))
1738                 return -1;
1739         free(path_desc);
1740
1741         if (opt->priv->call_depth) {
1742                 /*
1743                  * FIXME: For rename/add-source conflicts (if we could detect
1744                  * such), this is wrong.  We should instead find a unique
1745                  * pathname and then either rename the add-source file to that
1746                  * unique path, or use that unique path instead of src here.
1747                  */
1748                 if (update_file(opt, 0, &mfi.blob, o->path))
1749                         return -1;
1750
1751                 /*
1752                  * Above, we put the merged content at the merge-base's
1753                  * path.  Now we usually need to delete both a->path and
1754                  * b->path.  However, the rename on each side of the merge
1755                  * could also be involved in a rename/add conflict.  In
1756                  * such cases, we should keep the added file around,
1757                  * resolving the conflict at that path in its favor.
1758                  */
1759                 add = &ci->ren1->dst_entry->stages[2 ^ 1];
1760                 if (is_valid(add)) {
1761                         if (update_file(opt, 0, add, a->path))
1762                                 return -1;
1763                 }
1764                 else
1765                         remove_file_from_index(opt->repo->index, a->path);
1766                 add = &ci->ren2->dst_entry->stages[3 ^ 1];
1767                 if (is_valid(add)) {
1768                         if (update_file(opt, 0, add, b->path))
1769                                 return -1;
1770                 }
1771                 else
1772                         remove_file_from_index(opt->repo->index, b->path);
1773         } else {
1774                 /*
1775                  * For each destination path, we need to see if there is a
1776                  * rename/add collision.  If not, we can write the file out
1777                  * to the specified location.
1778                  */
1779                 add = &ci->ren1->dst_entry->stages[2 ^ 1];
1780                 if (is_valid(add)) {
1781                         add->path = mfi.blob.path = a->path;
1782                         if (handle_file_collision(opt, a->path,
1783                                                   NULL, NULL,
1784                                                   ci->ren1->branch,
1785                                                   ci->ren2->branch,
1786                                                   &mfi.blob, add) < 0)
1787                                 return -1;
1788                 } else {
1789                         char *new_path = find_path_for_conflict(opt, a->path,
1790                                                                 ci->ren1->branch,
1791                                                                 ci->ren2->branch);
1792                         if (update_file(opt, 0, &mfi.blob,
1793                                         new_path ? new_path : a->path))
1794                                 return -1;
1795                         free(new_path);
1796                         if (update_stages(opt, a->path, NULL, a, NULL))
1797                                 return -1;
1798                 }
1799
1800                 add = &ci->ren2->dst_entry->stages[3 ^ 1];
1801                 if (is_valid(add)) {
1802                         add->path = mfi.blob.path = b->path;
1803                         if (handle_file_collision(opt, b->path,
1804                                                   NULL, NULL,
1805                                                   ci->ren1->branch,
1806                                                   ci->ren2->branch,
1807                                                   add, &mfi.blob) < 0)
1808                                 return -1;
1809                 } else {
1810                         char *new_path = find_path_for_conflict(opt, b->path,
1811                                                                 ci->ren2->branch,
1812                                                                 ci->ren1->branch);
1813                         if (update_file(opt, 0, &mfi.blob,
1814                                         new_path ? new_path : b->path))
1815                                 return -1;
1816                         free(new_path);
1817                         if (update_stages(opt, b->path, NULL, NULL, b))
1818                                 return -1;
1819                 }
1820         }
1821
1822         return 0;
1823 }
1824
1825 static int handle_rename_rename_2to1(struct merge_options *opt,
1826                                      struct rename_conflict_info *ci)
1827 {
1828         /* Two files, a & b, were renamed to the same thing, c. */
1829         struct diff_filespec *a = ci->ren1->pair->one;
1830         struct diff_filespec *b = ci->ren2->pair->one;
1831         struct diff_filespec *c1 = ci->ren1->pair->two;
1832         struct diff_filespec *c2 = ci->ren2->pair->two;
1833         char *path = c1->path; /* == c2->path */
1834         char *path_side_1_desc;
1835         char *path_side_2_desc;
1836         struct merge_file_info mfi_c1;
1837         struct merge_file_info mfi_c2;
1838         int ostage1, ostage2;
1839
1840         output(opt, 1, _("CONFLICT (rename/rename): "
1841                "Rename %s->%s in %s. "
1842                "Rename %s->%s in %s"),
1843                a->path, c1->path, ci->ren1->branch,
1844                b->path, c2->path, ci->ren2->branch);
1845
1846         path_side_1_desc = xstrfmt("version of %s from %s", path, a->path);
1847         path_side_2_desc = xstrfmt("version of %s from %s", path, b->path);
1848         ostage1 = ci->ren1->branch == opt->branch1 ? 3 : 2;
1849         ostage2 = ostage1 ^ 1;
1850         ci->ren1->src_entry->stages[ostage1].path = a->path;
1851         ci->ren2->src_entry->stages[ostage2].path = b->path;
1852         if (merge_mode_and_contents(opt, a, c1,
1853                                     &ci->ren1->src_entry->stages[ostage1],
1854                                     path_side_1_desc,
1855                                     opt->branch1, opt->branch2,
1856                                     1 + opt->priv->call_depth * 2, &mfi_c1) ||
1857             merge_mode_and_contents(opt, b,
1858                                     &ci->ren2->src_entry->stages[ostage2],
1859                                     c2, path_side_2_desc,
1860                                     opt->branch1, opt->branch2,
1861                                     1 + opt->priv->call_depth * 2, &mfi_c2))
1862                 return -1;
1863         free(path_side_1_desc);
1864         free(path_side_2_desc);
1865         mfi_c1.blob.path = path;
1866         mfi_c2.blob.path = path;
1867
1868         return handle_file_collision(opt, path, a->path, b->path,
1869                                      ci->ren1->branch, ci->ren2->branch,
1870                                      &mfi_c1.blob, &mfi_c2.blob);
1871 }
1872
1873 /*
1874  * Get the diff_filepairs changed between o_tree and tree.
1875  */
1876 static struct diff_queue_struct *get_diffpairs(struct merge_options *opt,
1877                                                struct tree *o_tree,
1878                                                struct tree *tree)
1879 {
1880         struct diff_queue_struct *ret;
1881         struct diff_options opts;
1882
1883         repo_diff_setup(opt->repo, &opts);
1884         opts.flags.recursive = 1;
1885         opts.flags.rename_empty = 0;
1886         opts.detect_rename = merge_detect_rename(opt);
1887         /*
1888          * We do not have logic to handle the detection of copies.  In
1889          * fact, it may not even make sense to add such logic: would we
1890          * really want a change to a base file to be propagated through
1891          * multiple other files by a merge?
1892          */
1893         if (opts.detect_rename > DIFF_DETECT_RENAME)
1894                 opts.detect_rename = DIFF_DETECT_RENAME;
1895         opts.rename_limit = (opt->rename_limit >= 0) ? opt->rename_limit : 1000;
1896         opts.rename_score = opt->rename_score;
1897         opts.show_rename_progress = opt->show_rename_progress;
1898         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1899         diff_setup_done(&opts);
1900         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1901         diffcore_std(&opts);
1902         if (opts.needed_rename_limit > opt->priv->needed_rename_limit)
1903                 opt->priv->needed_rename_limit = opts.needed_rename_limit;
1904
1905         ret = xmalloc(sizeof(*ret));
1906         *ret = diff_queued_diff;
1907
1908         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1909         diff_queued_diff.nr = 0;
1910         diff_queued_diff.queue = NULL;
1911         diff_flush(&opts);
1912         return ret;
1913 }
1914
1915 static int tree_has_path(struct repository *r, struct tree *tree,
1916                          const char *path)
1917 {
1918         struct object_id hashy;
1919         unsigned short mode_o;
1920
1921         return !get_tree_entry(r,
1922                                &tree->object.oid, path,
1923                                &hashy, &mode_o);
1924 }
1925
1926 /*
1927  * Return a new string that replaces the beginning portion (which matches
1928  * entry->dir), with entry->new_dir.  In perl-speak:
1929  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1930  * NOTE:
1931  *   Caller must ensure that old_path starts with entry->dir + '/'.
1932  */
1933 static char *apply_dir_rename(struct dir_rename_entry *entry,
1934                               const char *old_path)
1935 {
1936         struct strbuf new_path = STRBUF_INIT;
1937         int oldlen, newlen;
1938
1939         if (entry->non_unique_new_dir)
1940                 return NULL;
1941
1942         oldlen = strlen(entry->dir);
1943         if (entry->new_dir.len == 0)
1944                 /*
1945                  * If someone renamed/merged a subdirectory into the root
1946                  * directory (e.g. 'some/subdir' -> ''), then we want to
1947                  * avoid returning
1948                  *     '' + '/filename'
1949                  * as the rename; we need to make old_path + oldlen advance
1950                  * past the '/' character.
1951                  */
1952                 oldlen++;
1953         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1954         strbuf_grow(&new_path, newlen);
1955         strbuf_addbuf(&new_path, &entry->new_dir);
1956         strbuf_addstr(&new_path, &old_path[oldlen]);
1957
1958         return strbuf_detach(&new_path, NULL);
1959 }
1960
1961 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1962                                     char **old_dir, char **new_dir)
1963 {
1964         char *end_of_old, *end_of_new;
1965
1966         /* Default return values: NULL, meaning no rename */
1967         *old_dir = NULL;
1968         *new_dir = NULL;
1969
1970         /*
1971          * For
1972          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1973          * the "e/foo.c" part is the same, we just want to know that
1974          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1975          * so, for this example, this function returns "a/b/c/d" in
1976          * *old_dir and "a/b/some/thing/else" in *new_dir.
1977          */
1978
1979         /*
1980          * If the basename of the file changed, we don't care.  We want
1981          * to know which portion of the directory, if any, changed.
1982          */
1983         end_of_old = strrchr(old_path, '/');
1984         end_of_new = strrchr(new_path, '/');
1985
1986         /*
1987          * If end_of_old is NULL, old_path wasn't in a directory, so there
1988          * could not be a directory rename (our rule elsewhere that a
1989          * directory which still exists is not considered to have been
1990          * renamed means the root directory can never be renamed -- because
1991          * the root directory always exists).
1992          */
1993         if (end_of_old == NULL)
1994                 return; /* Note: *old_dir and *new_dir are still NULL */
1995
1996         /*
1997          * If new_path contains no directory (end_of_new is NULL), then we
1998          * have a rename of old_path's directory to the root directory.
1999          */
2000         if (end_of_new == NULL) {
2001                 *old_dir = xstrndup(old_path, end_of_old - old_path);
2002                 *new_dir = xstrdup("");
2003                 return;
2004         }
2005
2006         /* Find the first non-matching character traversing backwards */
2007         while (*--end_of_new == *--end_of_old &&
2008                end_of_old != old_path &&
2009                end_of_new != new_path)
2010                 ; /* Do nothing; all in the while loop */
2011
2012         /*
2013          * If both got back to the beginning of their strings, then the
2014          * directory didn't change at all, only the basename did.
2015          */
2016         if (end_of_old == old_path && end_of_new == new_path &&
2017             *end_of_old == *end_of_new)
2018                 return; /* Note: *old_dir and *new_dir are still NULL */
2019
2020         /*
2021          * If end_of_new got back to the beginning of its string, and
2022          * end_of_old got back to the beginning of some subdirectory, then
2023          * we have a rename/merge of a subdirectory into the root, which
2024          * needs slightly special handling.
2025          *
2026          * Note: There is no need to consider the opposite case, with a
2027          * rename/merge of the root directory into some subdirectory
2028          * because as noted above the root directory always exists so it
2029          * cannot be considered to be renamed.
2030          */
2031         if (end_of_new == new_path &&
2032             end_of_old != old_path && end_of_old[-1] == '/') {
2033                 *old_dir = xstrndup(old_path, --end_of_old - old_path);
2034                 *new_dir = xstrdup("");
2035                 return;
2036         }
2037
2038         /*
2039          * We've found the first non-matching character in the directory
2040          * paths.  That means the current characters we were looking at
2041          * were part of the first non-matching subdir name going back from
2042          * the end of the strings.  Get the whole name by advancing both
2043          * end_of_old and end_of_new to the NEXT '/' character.  That will
2044          * represent the entire directory rename.
2045          *
2046          * The reason for the increment is cases like
2047          *    a/b/star/foo/whatever.c -> a/b/tar/foo/random.c
2048          * After dropping the basename and going back to the first
2049          * non-matching character, we're now comparing:
2050          *    a/b/s          and         a/b/
2051          * and we want to be comparing:
2052          *    a/b/star/      and         a/b/tar/
2053          * but without the pre-increment, the one on the right would stay
2054          * a/b/.
2055          */
2056         end_of_old = strchr(++end_of_old, '/');
2057         end_of_new = strchr(++end_of_new, '/');
2058
2059         /* Copy the old and new directories into *old_dir and *new_dir. */
2060         *old_dir = xstrndup(old_path, end_of_old - old_path);
2061         *new_dir = xstrndup(new_path, end_of_new - new_path);
2062 }
2063
2064 static void remove_hashmap_entries(struct hashmap *dir_renames,
2065                                    struct string_list *items_to_remove)
2066 {
2067         int i;
2068         struct dir_rename_entry *entry;
2069
2070         for (i = 0; i < items_to_remove->nr; i++) {
2071                 entry = items_to_remove->items[i].util;
2072                 hashmap_remove(dir_renames, &entry->ent, NULL);
2073         }
2074         string_list_clear(items_to_remove, 0);
2075 }
2076
2077 /*
2078  * See if there is a directory rename for path, and if there are any file
2079  * level conflicts for the renamed location.  If there is a rename and
2080  * there are no conflicts, return the new name.  Otherwise, return NULL.
2081  */
2082 static char *handle_path_level_conflicts(struct merge_options *opt,
2083                                          const char *path,
2084                                          struct dir_rename_entry *entry,
2085                                          struct hashmap *collisions,
2086                                          struct tree *tree)
2087 {
2088         char *new_path = NULL;
2089         struct collision_entry *collision_ent;
2090         int clean = 1;
2091         struct strbuf collision_paths = STRBUF_INIT;
2092
2093         /*
2094          * entry has the mapping of old directory name to new directory name
2095          * that we want to apply to path.
2096          */
2097         new_path = apply_dir_rename(entry, path);
2098
2099         if (!new_path) {
2100                 /* This should only happen when entry->non_unique_new_dir set */
2101                 if (!entry->non_unique_new_dir)
2102                         BUG("entry->non_unqiue_dir not set and !new_path");
2103                 output(opt, 1, _("CONFLICT (directory rename split): "
2104                                "Unclear where to place %s because directory "
2105                                "%s was renamed to multiple other directories, "
2106                                "with no destination getting a majority of the "
2107                                "files."),
2108                        path, entry->dir);
2109                 clean = 0;
2110                 return NULL;
2111         }
2112
2113         /*
2114          * The caller needs to have ensured that it has pre-populated
2115          * collisions with all paths that map to new_path.  Do a quick check
2116          * to ensure that's the case.
2117          */
2118         collision_ent = collision_find_entry(collisions, new_path);
2119         if (collision_ent == NULL)
2120                 BUG("collision_ent is NULL");
2121
2122         /*
2123          * Check for one-sided add/add/.../add conflicts, i.e.
2124          * where implicit renames from the other side doing
2125          * directory rename(s) can affect this side of history
2126          * to put multiple paths into the same location.  Warn
2127          * and bail on directory renames for such paths.
2128          */
2129         if (collision_ent->reported_already) {
2130                 clean = 0;
2131         } else if (tree_has_path(opt->repo, tree, new_path)) {
2132                 collision_ent->reported_already = 1;
2133                 strbuf_add_separated_string_list(&collision_paths, ", ",
2134                                                  &collision_ent->source_files);
2135                 output(opt, 1, _("CONFLICT (implicit dir rename): Existing "
2136                                "file/dir at %s in the way of implicit "
2137                                "directory rename(s) putting the following "
2138                                "path(s) there: %s."),
2139                        new_path, collision_paths.buf);
2140                 clean = 0;
2141         } else if (collision_ent->source_files.nr > 1) {
2142                 collision_ent->reported_already = 1;
2143                 strbuf_add_separated_string_list(&collision_paths, ", ",
2144                                                  &collision_ent->source_files);
2145                 output(opt, 1, _("CONFLICT (implicit dir rename): Cannot map "
2146                                "more than one path to %s; implicit directory "
2147                                "renames tried to put these paths there: %s"),
2148                        new_path, collision_paths.buf);
2149                 clean = 0;
2150         }
2151
2152         /* Free memory we no longer need */
2153         strbuf_release(&collision_paths);
2154         if (!clean && new_path) {
2155                 free(new_path);
2156                 return NULL;
2157         }
2158
2159         return new_path;
2160 }
2161
2162 /*
2163  * There are a couple things we want to do at the directory level:
2164  *   1. Check for both sides renaming to the same thing, in order to avoid
2165  *      implicit renaming of files that should be left in place.  (See
2166  *      testcase 6b in t6043 for details.)
2167  *   2. Prune directory renames if there are still files left in the
2168  *      the original directory.  These represent a partial directory rename,
2169  *      i.e. a rename where only some of the files within the directory
2170  *      were renamed elsewhere.  (Technically, this could be done earlier
2171  *      in get_directory_renames(), except that would prevent us from
2172  *      doing the previous check and thus failing testcase 6b.)
2173  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2174  *      In the future, we could potentially record this info as well and
2175  *      omit reporting rename/rename(1to2) conflicts for each path within
2176  *      the affected directories, thus cleaning up the merge output.
2177  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2178  *         directory level, because merging directories is fine.  If it
2179  *         causes conflicts for files within those merged directories, then
2180  *         that should be detected at the individual path level.
2181  */
2182 static void handle_directory_level_conflicts(struct merge_options *opt,
2183                                              struct hashmap *dir_re_head,
2184                                              struct tree *head,
2185                                              struct hashmap *dir_re_merge,
2186                                              struct tree *merge)
2187 {
2188         struct hashmap_iter iter;
2189         struct dir_rename_entry *head_ent;
2190         struct dir_rename_entry *merge_ent;
2191
2192         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2193         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2194
2195         hashmap_for_each_entry(dir_re_head, &iter, head_ent,
2196                                 ent /* member name */) {
2197                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2198                 if (merge_ent &&
2199                     !head_ent->non_unique_new_dir &&
2200                     !merge_ent->non_unique_new_dir &&
2201                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2202                         /* 1. Renamed identically; remove it from both sides */
2203                         string_list_append(&remove_from_head,
2204                                            head_ent->dir)->util = head_ent;
2205                         strbuf_release(&head_ent->new_dir);
2206                         string_list_append(&remove_from_merge,
2207                                            merge_ent->dir)->util = merge_ent;
2208                         strbuf_release(&merge_ent->new_dir);
2209                 } else if (tree_has_path(opt->repo, head, head_ent->dir)) {
2210                         /* 2. This wasn't a directory rename after all */
2211                         string_list_append(&remove_from_head,
2212                                            head_ent->dir)->util = head_ent;
2213                         strbuf_release(&head_ent->new_dir);
2214                 }
2215         }
2216
2217         remove_hashmap_entries(dir_re_head, &remove_from_head);
2218         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2219
2220         hashmap_for_each_entry(dir_re_merge, &iter, merge_ent,
2221                                 ent /* member name */) {
2222                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2223                 if (tree_has_path(opt->repo, merge, merge_ent->dir)) {
2224                         /* 2. This wasn't a directory rename after all */
2225                         string_list_append(&remove_from_merge,
2226                                            merge_ent->dir)->util = merge_ent;
2227                 } else if (head_ent &&
2228                            !head_ent->non_unique_new_dir &&
2229                            !merge_ent->non_unique_new_dir) {
2230                         /* 3. rename/rename(1to2) */
2231                         /*
2232                          * We can assume it's not rename/rename(1to1) because
2233                          * that was case (1), already checked above.  So we
2234                          * know that head_ent->new_dir and merge_ent->new_dir
2235                          * are different strings.
2236                          */
2237                         output(opt, 1, _("CONFLICT (rename/rename): "
2238                                        "Rename directory %s->%s in %s. "
2239                                        "Rename directory %s->%s in %s"),
2240                                head_ent->dir, head_ent->new_dir.buf, opt->branch1,
2241                                head_ent->dir, merge_ent->new_dir.buf, opt->branch2);
2242                         string_list_append(&remove_from_head,
2243                                            head_ent->dir)->util = head_ent;
2244                         strbuf_release(&head_ent->new_dir);
2245                         string_list_append(&remove_from_merge,
2246                                            merge_ent->dir)->util = merge_ent;
2247                         strbuf_release(&merge_ent->new_dir);
2248                 }
2249         }
2250
2251         remove_hashmap_entries(dir_re_head, &remove_from_head);
2252         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2253 }
2254
2255 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs)
2256 {
2257         struct hashmap *dir_renames;
2258         struct hashmap_iter iter;
2259         struct dir_rename_entry *entry;
2260         int i;
2261
2262         /*
2263          * Typically, we think of a directory rename as all files from a
2264          * certain directory being moved to a target directory.  However,
2265          * what if someone first moved two files from the original
2266          * directory in one commit, and then renamed the directory
2267          * somewhere else in a later commit?  At merge time, we just know
2268          * that files from the original directory went to two different
2269          * places, and that the bulk of them ended up in the same place.
2270          * We want each directory rename to represent where the bulk of the
2271          * files from that directory end up; this function exists to find
2272          * where the bulk of the files went.
2273          *
2274          * The first loop below simply iterates through the list of file
2275          * renames, finding out how often each directory rename pair
2276          * possibility occurs.
2277          */
2278         dir_renames = xmalloc(sizeof(*dir_renames));
2279         dir_rename_init(dir_renames);
2280         for (i = 0; i < pairs->nr; ++i) {
2281                 struct string_list_item *item;
2282                 int *count;
2283                 struct diff_filepair *pair = pairs->queue[i];
2284                 char *old_dir, *new_dir;
2285
2286                 /* File not part of directory rename if it wasn't renamed */
2287                 if (pair->status != 'R')
2288                         continue;
2289
2290                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2291                                         &old_dir,        &new_dir);
2292                 if (!old_dir)
2293                         /* Directory didn't change at all; ignore this one. */
2294                         continue;
2295
2296                 entry = dir_rename_find_entry(dir_renames, old_dir);
2297                 if (!entry) {
2298                         entry = xmalloc(sizeof(*entry));
2299                         dir_rename_entry_init(entry, old_dir);
2300                         hashmap_put(dir_renames, &entry->ent);
2301                 } else {
2302                         free(old_dir);
2303                 }
2304                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2305                 if (!item) {
2306                         item = string_list_insert(&entry->possible_new_dirs,
2307                                                   new_dir);
2308                         item->util = xcalloc(1, sizeof(int));
2309                 } else {
2310                         free(new_dir);
2311                 }
2312                 count = item->util;
2313                 *count += 1;
2314         }
2315
2316         /*
2317          * For each directory with files moved out of it, we find out which
2318          * target directory received the most files so we can declare it to
2319          * be the "winning" target location for the directory rename.  This
2320          * winner gets recorded in new_dir.  If there is no winner
2321          * (multiple target directories received the same number of files),
2322          * we set non_unique_new_dir.  Once we've determined the winner (or
2323          * that there is no winner), we no longer need possible_new_dirs.
2324          */
2325         hashmap_for_each_entry(dir_renames, &iter, entry,
2326                                 ent /* member name */) {
2327                 int max = 0;
2328                 int bad_max = 0;
2329                 char *best = NULL;
2330
2331                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2332                         int *count = entry->possible_new_dirs.items[i].util;
2333
2334                         if (*count == max)
2335                                 bad_max = max;
2336                         else if (*count > max) {
2337                                 max = *count;
2338                                 best = entry->possible_new_dirs.items[i].string;
2339                         }
2340                 }
2341                 if (bad_max == max)
2342                         entry->non_unique_new_dir = 1;
2343                 else {
2344                         assert(entry->new_dir.len == 0);
2345                         strbuf_addstr(&entry->new_dir, best);
2346                 }
2347                 /*
2348                  * The relevant directory sub-portion of the original full
2349                  * filepaths were xstrndup'ed before inserting into
2350                  * possible_new_dirs, and instead of manually iterating the
2351                  * list and free'ing each, just lie and tell
2352                  * possible_new_dirs that it did the strdup'ing so that it
2353                  * will free them for us.
2354                  */
2355                 entry->possible_new_dirs.strdup_strings = 1;
2356                 string_list_clear(&entry->possible_new_dirs, 1);
2357         }
2358
2359         return dir_renames;
2360 }
2361
2362 static struct dir_rename_entry *check_dir_renamed(const char *path,
2363                                                   struct hashmap *dir_renames)
2364 {
2365         char *temp = xstrdup(path);
2366         char *end;
2367         struct dir_rename_entry *entry = NULL;
2368
2369         while ((end = strrchr(temp, '/'))) {
2370                 *end = '\0';
2371                 entry = dir_rename_find_entry(dir_renames, temp);
2372                 if (entry)
2373                         break;
2374         }
2375         free(temp);
2376         return entry;
2377 }
2378
2379 static void compute_collisions(struct hashmap *collisions,
2380                                struct hashmap *dir_renames,
2381                                struct diff_queue_struct *pairs)
2382 {
2383         int i;
2384
2385         /*
2386          * Multiple files can be mapped to the same path due to directory
2387          * renames done by the other side of history.  Since that other
2388          * side of history could have merged multiple directories into one,
2389          * if our side of history added the same file basename to each of
2390          * those directories, then all N of them would get implicitly
2391          * renamed by the directory rename detection into the same path,
2392          * and we'd get an add/add/.../add conflict, and all those adds
2393          * from *this* side of history.  This is not representable in the
2394          * index, and users aren't going to easily be able to make sense of
2395          * it.  So we need to provide a good warning about what's
2396          * happening, and fall back to no-directory-rename detection
2397          * behavior for those paths.
2398          *
2399          * See testcases 9e and all of section 5 from t6043 for examples.
2400          */
2401         collision_init(collisions);
2402
2403         for (i = 0; i < pairs->nr; ++i) {
2404                 struct dir_rename_entry *dir_rename_ent;
2405                 struct collision_entry *collision_ent;
2406                 char *new_path;
2407                 struct diff_filepair *pair = pairs->queue[i];
2408
2409                 if (pair->status != 'A' && pair->status != 'R')
2410                         continue;
2411                 dir_rename_ent = check_dir_renamed(pair->two->path,
2412                                                    dir_renames);
2413                 if (!dir_rename_ent)
2414                         continue;
2415
2416                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2417                 if (!new_path)
2418                         /*
2419                          * dir_rename_ent->non_unique_new_path is true, which
2420                          * means there is no directory rename for us to use,
2421                          * which means it won't cause us any additional
2422                          * collisions.
2423                          */
2424                         continue;
2425                 collision_ent = collision_find_entry(collisions, new_path);
2426                 if (!collision_ent) {
2427                         collision_ent = xcalloc(1,
2428                                                 sizeof(struct collision_entry));
2429                         hashmap_entry_init(&collision_ent->ent,
2430                                                 strhash(new_path));
2431                         hashmap_put(collisions, &collision_ent->ent);
2432                         collision_ent->target_file = new_path;
2433                 } else {
2434                         free(new_path);
2435                 }
2436                 string_list_insert(&collision_ent->source_files,
2437                                    pair->two->path);
2438         }
2439 }
2440
2441 static char *check_for_directory_rename(struct merge_options *opt,
2442                                         const char *path,
2443                                         struct tree *tree,
2444                                         struct hashmap *dir_renames,
2445                                         struct hashmap *dir_rename_exclusions,
2446                                         struct hashmap *collisions,
2447                                         int *clean_merge)
2448 {
2449         char *new_path = NULL;
2450         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2451         struct dir_rename_entry *oentry = NULL;
2452
2453         if (!entry)
2454                 return new_path;
2455
2456         /*
2457          * This next part is a little weird.  We do not want to do an
2458          * implicit rename into a directory we renamed on our side, because
2459          * that will result in a spurious rename/rename(1to2) conflict.  An
2460          * example:
2461          *   Base commit: dumbdir/afile, otherdir/bfile
2462          *   Side 1:      smrtdir/afile, otherdir/bfile
2463          *   Side 2:      dumbdir/afile, dumbdir/bfile
2464          * Here, while working on Side 1, we could notice that otherdir was
2465          * renamed/merged to dumbdir, and change the diff_filepair for
2466          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2467          * 2 will notice the rename from dumbdir to smrtdir, and do the
2468          * transitive rename to move it from dumbdir/bfile to
2469          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2470          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2471          * the file to end up in smrtdir.  And the way to achieve that is
2472          * to not let Side1 do the rename to dumbdir, since we know that is
2473          * the source of one of our directory renames.
2474          *
2475          * That's why oentry and dir_rename_exclusions is here.
2476          *
2477          * As it turns out, this also prevents N-way transient rename
2478          * confusion; See testcases 9c and 9d of t6043.
2479          */
2480         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2481         if (oentry) {
2482                 output(opt, 1, _("WARNING: Avoiding applying %s -> %s rename "
2483                                "to %s, because %s itself was renamed."),
2484                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2485         } else {
2486                 new_path = handle_path_level_conflicts(opt, path, entry,
2487                                                        collisions, tree);
2488                 *clean_merge &= (new_path != NULL);
2489         }
2490
2491         return new_path;
2492 }
2493
2494 static void apply_directory_rename_modifications(struct merge_options *opt,
2495                                                  struct diff_filepair *pair,
2496                                                  char *new_path,
2497                                                  struct rename *re,
2498                                                  struct tree *tree,
2499                                                  struct tree *o_tree,
2500                                                  struct tree *a_tree,
2501                                                  struct tree *b_tree,
2502                                                  struct string_list *entries)
2503 {
2504         struct string_list_item *item;
2505         int stage = (tree == a_tree ? 2 : 3);
2506         int update_wd;
2507
2508         /*
2509          * In all cases where we can do directory rename detection,
2510          * unpack_trees() will have read pair->two->path into the
2511          * index and the working copy.  We need to remove it so that
2512          * we can instead place it at new_path.  It is guaranteed to
2513          * not be untracked (unpack_trees() would have errored out
2514          * saying the file would have been overwritten), but it might
2515          * be dirty, though.
2516          */
2517         update_wd = !was_dirty(opt, pair->two->path);
2518         if (!update_wd)
2519                 output(opt, 1, _("Refusing to lose dirty file at %s"),
2520                        pair->two->path);
2521         remove_file(opt, 1, pair->two->path, !update_wd);
2522
2523         /* Find or create a new re->dst_entry */
2524         item = string_list_lookup(entries, new_path);
2525         if (item) {
2526                 /*
2527                  * Since we're renaming on this side of history, and it's
2528                  * due to a directory rename on the other side of history
2529                  * (which we only allow when the directory in question no
2530                  * longer exists on the other side of history), the
2531                  * original entry for re->dst_entry is no longer
2532                  * necessary...
2533                  */
2534                 re->dst_entry->processed = 1;
2535
2536                 /*
2537                  * ...because we'll be using this new one.
2538                  */
2539                 re->dst_entry = item->util;
2540         } else {
2541                 /*
2542                  * re->dst_entry is for the before-dir-rename path, and we
2543                  * need it to hold information for the after-dir-rename
2544                  * path.  Before creating a new entry, we need to mark the
2545                  * old one as unnecessary (...unless it is shared by
2546                  * src_entry, i.e. this didn't use to be a rename, in which
2547                  * case we can just allow the normal processing to happen
2548                  * for it).
2549                  */
2550                 if (pair->status == 'R')
2551                         re->dst_entry->processed = 1;
2552
2553                 re->dst_entry = insert_stage_data(opt->repo, new_path,
2554                                                   o_tree, a_tree, b_tree,
2555                                                   entries);
2556                 item = string_list_insert(entries, new_path);
2557                 item->util = re->dst_entry;
2558         }
2559
2560         /*
2561          * Update the stage_data with the information about the path we are
2562          * moving into place.  That slot will be empty and available for us
2563          * to write to because of the collision checks in
2564          * handle_path_level_conflicts().  In other words,
2565          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2566          * open for us to write to.
2567          *
2568          * It may be tempting to actually update the index at this point as
2569          * well, using update_stages_for_stage_data(), but as per the big
2570          * "NOTE" in update_stages(), doing so will modify the current
2571          * in-memory index which will break calls to would_lose_untracked()
2572          * that we need to make.  Instead, we need to just make sure that
2573          * the various handle_rename_*() functions update the index
2574          * explicitly rather than relying on unpack_trees() to have done it.
2575          */
2576         get_tree_entry(opt->repo,
2577                        &tree->object.oid,
2578                        pair->two->path,
2579                        &re->dst_entry->stages[stage].oid,
2580                        &re->dst_entry->stages[stage].mode);
2581
2582         /*
2583          * Record the original change status (or 'type' of change).  If it
2584          * was originally an add ('A'), this lets us differentiate later
2585          * between a RENAME_DELETE conflict and RENAME_VIA_DIR (they
2586          * otherwise look the same).  If it was originally a rename ('R'),
2587          * this lets us remember and report accurately about the transitive
2588          * renaming that occurred via the directory rename detection.  Also,
2589          * record the original destination name.
2590          */
2591         re->dir_rename_original_type = pair->status;
2592         re->dir_rename_original_dest = pair->two->path;
2593
2594         /*
2595          * We don't actually look at pair->status again, but it seems
2596          * pedagogically correct to adjust it.
2597          */
2598         pair->status = 'R';
2599
2600         /*
2601          * Finally, record the new location.
2602          */
2603         pair->two->path = new_path;
2604 }
2605
2606 /*
2607  * Get information of all renames which occurred in 'pairs', making use of
2608  * any implicit directory renames inferred from the other side of history.
2609  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2610  * to be able to associate the correct cache entries with the rename
2611  * information; tree is always equal to either a_tree or b_tree.
2612  */
2613 static struct string_list *get_renames(struct merge_options *opt,
2614                                        const char *branch,
2615                                        struct diff_queue_struct *pairs,
2616                                        struct hashmap *dir_renames,
2617                                        struct hashmap *dir_rename_exclusions,
2618                                        struct tree *tree,
2619                                        struct tree *o_tree,
2620                                        struct tree *a_tree,
2621                                        struct tree *b_tree,
2622                                        struct string_list *entries,
2623                                        int *clean_merge)
2624 {
2625         int i;
2626         struct hashmap collisions;
2627         struct hashmap_iter iter;
2628         struct collision_entry *e;
2629         struct string_list *renames;
2630
2631         compute_collisions(&collisions, dir_renames, pairs);
2632         renames = xcalloc(1, sizeof(struct string_list));
2633
2634         for (i = 0; i < pairs->nr; ++i) {
2635                 struct string_list_item *item;
2636                 struct rename *re;
2637                 struct diff_filepair *pair = pairs->queue[i];
2638                 char *new_path; /* non-NULL only with directory renames */
2639
2640                 if (pair->status != 'A' && pair->status != 'R') {
2641                         diff_free_filepair(pair);
2642                         continue;
2643                 }
2644                 new_path = check_for_directory_rename(opt, pair->two->path, tree,
2645                                                       dir_renames,
2646                                                       dir_rename_exclusions,
2647                                                       &collisions,
2648                                                       clean_merge);
2649                 if (pair->status != 'R' && !new_path) {
2650                         diff_free_filepair(pair);
2651                         continue;
2652                 }
2653
2654                 re = xmalloc(sizeof(*re));
2655                 re->processed = 0;
2656                 re->pair = pair;
2657                 re->branch = branch;
2658                 re->dir_rename_original_type = '\0';
2659                 re->dir_rename_original_dest = NULL;
2660                 item = string_list_lookup(entries, re->pair->one->path);
2661                 if (!item)
2662                         re->src_entry = insert_stage_data(opt->repo,
2663                                         re->pair->one->path,
2664                                         o_tree, a_tree, b_tree, entries);
2665                 else
2666                         re->src_entry = item->util;
2667
2668                 item = string_list_lookup(entries, re->pair->two->path);
2669                 if (!item)
2670                         re->dst_entry = insert_stage_data(opt->repo,
2671                                         re->pair->two->path,
2672                                         o_tree, a_tree, b_tree, entries);
2673                 else
2674                         re->dst_entry = item->util;
2675                 item = string_list_insert(renames, pair->one->path);
2676                 item->util = re;
2677                 if (new_path)
2678                         apply_directory_rename_modifications(opt, pair, new_path,
2679                                                              re, tree, o_tree,
2680                                                              a_tree, b_tree,
2681                                                              entries);
2682         }
2683
2684         hashmap_for_each_entry(&collisions, &iter, e,
2685                                 ent /* member name */) {
2686                 free(e->target_file);
2687                 string_list_clear(&e->source_files, 0);
2688         }
2689         hashmap_free_entries(&collisions, struct collision_entry, ent);
2690         return renames;
2691 }
2692
2693 static int process_renames(struct merge_options *opt,
2694                            struct string_list *a_renames,
2695                            struct string_list *b_renames)
2696 {
2697         int clean_merge = 1, i, j;
2698         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2699         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2700         const struct rename *sre;
2701
2702         for (i = 0; i < a_renames->nr; i++) {
2703                 sre = a_renames->items[i].util;
2704                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2705                         = (void *)sre;
2706         }
2707         for (i = 0; i < b_renames->nr; i++) {
2708                 sre = b_renames->items[i].util;
2709                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2710                         = (void *)sre;
2711         }
2712
2713         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2714                 struct string_list *renames1, *renames2Dst;
2715                 struct rename *ren1 = NULL, *ren2 = NULL;
2716                 const char *ren1_src, *ren1_dst;
2717                 struct string_list_item *lookup;
2718
2719                 if (i >= a_renames->nr) {
2720                         ren2 = b_renames->items[j++].util;
2721                 } else if (j >= b_renames->nr) {
2722                         ren1 = a_renames->items[i++].util;
2723                 } else {
2724                         int compare = strcmp(a_renames->items[i].string,
2725                                              b_renames->items[j].string);
2726                         if (compare <= 0)
2727                                 ren1 = a_renames->items[i++].util;
2728                         if (compare >= 0)
2729                                 ren2 = b_renames->items[j++].util;
2730                 }
2731
2732                 /* TODO: refactor, so that 1/2 are not needed */
2733                 if (ren1) {
2734                         renames1 = a_renames;
2735                         renames2Dst = &b_by_dst;
2736                 } else {
2737                         renames1 = b_renames;
2738                         renames2Dst = &a_by_dst;
2739                         SWAP(ren2, ren1);
2740                 }
2741
2742                 if (ren1->processed)
2743                         continue;
2744                 ren1->processed = 1;
2745                 ren1->dst_entry->processed = 1;
2746                 /* BUG: We should only mark src_entry as processed if we
2747                  * are not dealing with a rename + add-source case.
2748                  */
2749                 ren1->src_entry->processed = 1;
2750
2751                 ren1_src = ren1->pair->one->path;
2752                 ren1_dst = ren1->pair->two->path;
2753
2754                 if (ren2) {
2755                         /* One file renamed on both sides */
2756                         const char *ren2_src = ren2->pair->one->path;
2757                         const char *ren2_dst = ren2->pair->two->path;
2758                         enum rename_type rename_type;
2759                         if (strcmp(ren1_src, ren2_src) != 0)
2760                                 BUG("ren1_src != ren2_src");
2761                         ren2->dst_entry->processed = 1;
2762                         ren2->processed = 1;
2763                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2764                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2765                                 clean_merge = 0;
2766                         } else {
2767                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2768                                 /* BUG: We should only remove ren1_src in
2769                                  * the base stage (think of rename +
2770                                  * add-source cases).
2771                                  */
2772                                 remove_file(opt, 1, ren1_src, 1);
2773                                 update_entry(ren1->dst_entry,
2774                                              ren1->pair->one,
2775                                              ren1->pair->two,
2776                                              ren2->pair->two);
2777                         }
2778                         setup_rename_conflict_info(rename_type, opt, ren1, ren2);
2779                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2780                         /* Two different files renamed to the same thing */
2781                         char *ren2_dst;
2782                         ren2 = lookup->util;
2783                         ren2_dst = ren2->pair->two->path;
2784                         if (strcmp(ren1_dst, ren2_dst) != 0)
2785                                 BUG("ren1_dst != ren2_dst");
2786
2787                         clean_merge = 0;
2788                         ren2->processed = 1;
2789                         /*
2790                          * BUG: We should only mark src_entry as processed
2791                          * if we are not dealing with a rename + add-source
2792                          * case.
2793                          */
2794                         ren2->src_entry->processed = 1;
2795
2796                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2797                                                    opt, ren1, ren2);
2798                 } else {
2799                         /* Renamed in 1, maybe changed in 2 */
2800                         /* we only use sha1 and mode of these */
2801                         struct diff_filespec src_other, dst_other;
2802                         int try_merge;
2803
2804                         /*
2805                          * unpack_trees loads entries from common-commit
2806                          * into stage 1, from head-commit into stage 2, and
2807                          * from merge-commit into stage 3.  We keep track
2808                          * of which side corresponds to the rename.
2809                          */
2810                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2811                         int other_stage =   a_renames == renames1 ? 3 : 2;
2812
2813                         /* BUG: We should only remove ren1_src in the base
2814                          * stage and in other_stage (think of rename +
2815                          * add-source case).
2816                          */
2817                         remove_file(opt, 1, ren1_src,
2818                                     renamed_stage == 2 || !was_tracked(opt, ren1_src));
2819
2820                         oidcpy(&src_other.oid,
2821                                &ren1->src_entry->stages[other_stage].oid);
2822                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2823                         oidcpy(&dst_other.oid,
2824                                &ren1->dst_entry->stages[other_stage].oid);
2825                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2826                         try_merge = 0;
2827
2828                         if (oideq(&src_other.oid, &null_oid) &&
2829                             ren1->dir_rename_original_type == 'A') {
2830                                 setup_rename_conflict_info(RENAME_VIA_DIR,
2831                                                            opt, ren1, NULL);
2832                         } else if (oideq(&src_other.oid, &null_oid)) {
2833                                 setup_rename_conflict_info(RENAME_DELETE,
2834                                                            opt, ren1, NULL);
2835                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2836                                    oideq(&dst_other.oid, &ren1->pair->two->oid)) {
2837                                 /*
2838                                  * Added file on the other side identical to
2839                                  * the file being renamed: clean merge.
2840                                  * Also, there is no need to overwrite the
2841                                  * file already in the working copy, so call
2842                                  * update_file_flags() instead of
2843                                  * update_file().
2844                                  */
2845                                 if (update_file_flags(opt,
2846                                                       ren1->pair->two,
2847                                                       ren1_dst,
2848                                                       1, /* update_cache */
2849                                                       0  /* update_wd    */))
2850                                         clean_merge = -1;
2851                         } else if (!oideq(&dst_other.oid, &null_oid)) {
2852                                 /*
2853                                  * Probably not a clean merge, but it's
2854                                  * premature to set clean_merge to 0 here,
2855                                  * because if the rename merges cleanly and
2856                                  * the merge exactly matches the newly added
2857                                  * file, then the merge will be clean.
2858                                  */
2859                                 setup_rename_conflict_info(RENAME_ADD,
2860                                                            opt, ren1, NULL);
2861                         } else
2862                                 try_merge = 1;
2863
2864                         if (clean_merge < 0)
2865                                 goto cleanup_and_return;
2866                         if (try_merge) {
2867                                 struct diff_filespec *o, *a, *b;
2868                                 src_other.path = (char *)ren1_src;
2869
2870                                 o = ren1->pair->one;
2871                                 if (a_renames == renames1) {
2872                                         a = ren1->pair->two;
2873                                         b = &src_other;
2874                                 } else {
2875                                         b = ren1->pair->two;
2876                                         a = &src_other;
2877                                 }
2878                                 update_entry(ren1->dst_entry, o, a, b);
2879                                 setup_rename_conflict_info(RENAME_NORMAL,
2880                                                            opt, ren1, NULL);
2881                         }
2882                 }
2883         }
2884 cleanup_and_return:
2885         string_list_clear(&a_by_dst, 0);
2886         string_list_clear(&b_by_dst, 0);
2887
2888         return clean_merge;
2889 }
2890
2891 struct rename_info {
2892         struct string_list *head_renames;
2893         struct string_list *merge_renames;
2894 };
2895
2896 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2897                                    struct hashmap *dir_renames)
2898 {
2899         struct hashmap_iter iter;
2900         struct dir_rename_entry *e;
2901
2902         hashmap_for_each_entry(dir_renames, &iter, e,
2903                                 ent /* member name */) {
2904                 free(e->dir);
2905                 strbuf_release(&e->new_dir);
2906                 /* possible_new_dirs already cleared in get_directory_renames */
2907         }
2908         hashmap_free_entries(dir_renames, struct dir_rename_entry, ent);
2909         free(dir_renames);
2910
2911         free(pairs->queue);
2912         free(pairs);
2913 }
2914
2915 static int detect_and_process_renames(struct merge_options *opt,
2916                                       struct tree *common,
2917                                       struct tree *head,
2918                                       struct tree *merge,
2919                                       struct string_list *entries,
2920                                       struct rename_info *ri)
2921 {
2922         struct diff_queue_struct *head_pairs, *merge_pairs;
2923         struct hashmap *dir_re_head, *dir_re_merge;
2924         int clean = 1;
2925
2926         ri->head_renames = NULL;
2927         ri->merge_renames = NULL;
2928
2929         if (!merge_detect_rename(opt))
2930                 return 1;
2931
2932         head_pairs = get_diffpairs(opt, common, head);
2933         merge_pairs = get_diffpairs(opt, common, merge);
2934
2935         if ((opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_TRUE) ||
2936             (opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_CONFLICT &&
2937              !opt->priv->call_depth)) {
2938                 dir_re_head = get_directory_renames(head_pairs);
2939                 dir_re_merge = get_directory_renames(merge_pairs);
2940
2941                 handle_directory_level_conflicts(opt,
2942                                                  dir_re_head, head,
2943                                                  dir_re_merge, merge);
2944         } else {
2945                 dir_re_head  = xmalloc(sizeof(*dir_re_head));
2946                 dir_re_merge = xmalloc(sizeof(*dir_re_merge));
2947                 dir_rename_init(dir_re_head);
2948                 dir_rename_init(dir_re_merge);
2949         }
2950
2951         ri->head_renames  = get_renames(opt, opt->branch1, head_pairs,
2952                                         dir_re_merge, dir_re_head, head,
2953                                         common, head, merge, entries,
2954                                         &clean);
2955         if (clean < 0)
2956                 goto cleanup;
2957         ri->merge_renames = get_renames(opt, opt->branch2, merge_pairs,
2958                                         dir_re_head, dir_re_merge, merge,
2959                                         common, head, merge, entries,
2960                                         &clean);
2961         if (clean < 0)
2962                 goto cleanup;
2963         clean &= process_renames(opt, ri->head_renames, ri->merge_renames);
2964
2965 cleanup:
2966         /*
2967          * Some cleanup is deferred until cleanup_renames() because the
2968          * data structures are still needed and referenced in
2969          * process_entry().  But there are a few things we can free now.
2970          */
2971         initial_cleanup_rename(head_pairs, dir_re_head);
2972         initial_cleanup_rename(merge_pairs, dir_re_merge);
2973
2974         return clean;
2975 }
2976
2977 static void final_cleanup_rename(struct string_list *rename)
2978 {
2979         const struct rename *re;
2980         int i;
2981
2982         if (rename == NULL)
2983                 return;
2984
2985         for (i = 0; i < rename->nr; i++) {
2986                 re = rename->items[i].util;
2987                 diff_free_filepair(re->pair);
2988         }
2989         string_list_clear(rename, 1);
2990         free(rename);
2991 }
2992
2993 static void final_cleanup_renames(struct rename_info *re_info)
2994 {
2995         final_cleanup_rename(re_info->head_renames);
2996         final_cleanup_rename(re_info->merge_renames);
2997 }
2998
2999 static int read_oid_strbuf(struct merge_options *opt,
3000                            const struct object_id *oid,
3001                            struct strbuf *dst)
3002 {
3003         void *buf;
3004         enum object_type type;
3005         unsigned long size;
3006         buf = read_object_file(oid, &type, &size);
3007         if (!buf)
3008                 return err(opt, _("cannot read object %s"), oid_to_hex(oid));
3009         if (type != OBJ_BLOB) {
3010                 free(buf);
3011                 return err(opt, _("object %s is not a blob"), oid_to_hex(oid));
3012         }
3013         strbuf_attach(dst, buf, size, size + 1);
3014         return 0;
3015 }
3016
3017 static int blob_unchanged(struct merge_options *opt,
3018                           const struct diff_filespec *o,
3019                           const struct diff_filespec *a,
3020                           int renormalize, const char *path)
3021 {
3022         struct strbuf obuf = STRBUF_INIT;
3023         struct strbuf abuf = STRBUF_INIT;
3024         int ret = 0; /* assume changed for safety */
3025         const struct index_state *idx = opt->repo->index;
3026
3027         if (a->mode != o->mode)
3028                 return 0;
3029         if (oideq(&o->oid, &a->oid))
3030                 return 1;
3031         if (!renormalize)
3032                 return 0;
3033
3034         if (read_oid_strbuf(opt, &o->oid, &obuf) ||
3035             read_oid_strbuf(opt, &a->oid, &abuf))
3036                 goto error_return;
3037         /*
3038          * Note: binary | is used so that both renormalizations are
3039          * performed.  Comparison can be skipped if both files are
3040          * unchanged since their sha1s have already been compared.
3041          */
3042         if (renormalize_buffer(idx, path, obuf.buf, obuf.len, &obuf) |
3043             renormalize_buffer(idx, path, abuf.buf, abuf.len, &abuf))
3044                 ret = (obuf.len == abuf.len && !memcmp(obuf.buf, abuf.buf, obuf.len));
3045
3046 error_return:
3047         strbuf_release(&obuf);
3048         strbuf_release(&abuf);
3049         return ret;
3050 }
3051
3052 static int handle_modify_delete(struct merge_options *opt,
3053                                 const char *path,
3054                                 const struct diff_filespec *o,
3055                                 const struct diff_filespec *a,
3056                                 const struct diff_filespec *b)
3057 {
3058         const char *modify_branch, *delete_branch;
3059         const struct diff_filespec *changed;
3060
3061         if (is_valid(a)) {
3062                 modify_branch = opt->branch1;
3063                 delete_branch = opt->branch2;
3064                 changed = a;
3065         } else {
3066                 modify_branch = opt->branch2;
3067                 delete_branch = opt->branch1;
3068                 changed = b;
3069         }
3070
3071         return handle_change_delete(opt,
3072                                     path, NULL,
3073                                     o, changed,
3074                                     modify_branch, delete_branch,
3075                                     _("modify"), _("modified"));
3076 }
3077
3078 static int handle_content_merge(struct merge_file_info *mfi,
3079                                 struct merge_options *opt,
3080                                 const char *path,
3081                                 int is_dirty,
3082                                 const struct diff_filespec *o,
3083                                 const struct diff_filespec *a,
3084                                 const struct diff_filespec *b,
3085                                 struct rename_conflict_info *ci)
3086 {
3087         const char *reason = _("content");
3088         unsigned df_conflict_remains = 0;
3089
3090         if (!is_valid(o))
3091                 reason = _("add/add");
3092
3093         assert(o->path && a->path && b->path);
3094         if (ci && dir_in_way(opt->repo->index, path, !opt->priv->call_depth,
3095                              S_ISGITLINK(ci->ren1->pair->two->mode)))
3096                 df_conflict_remains = 1;
3097
3098         if (merge_mode_and_contents(opt, o, a, b, path,
3099                                     opt->branch1, opt->branch2,
3100                                     opt->priv->call_depth * 2, mfi))
3101                 return -1;
3102
3103         /*
3104          * We can skip updating the working tree file iff:
3105          *   a) The merge is clean
3106          *   b) The merge matches what was in HEAD (content, mode, pathname)
3107          *   c) The target path is usable (i.e. not involved in D/F conflict)
3108          */
3109         if (mfi->clean && was_tracked_and_matches(opt, path, &mfi->blob) &&
3110             !df_conflict_remains) {
3111                 int pos;
3112                 struct cache_entry *ce;
3113
3114                 output(opt, 3, _("Skipped %s (merged same as existing)"), path);
3115                 if (add_cacheinfo(opt, &mfi->blob, path,
3116                                   0, (!opt->priv->call_depth && !is_dirty), 0))
3117                         return -1;
3118                 /*
3119                  * However, add_cacheinfo() will delete the old cache entry
3120                  * and add a new one.  We need to copy over any skip_worktree
3121                  * flag to avoid making the file appear as if it were
3122                  * deleted by the user.
3123                  */
3124                 pos = index_name_pos(&opt->priv->orig_index, path, strlen(path));
3125                 ce = opt->priv->orig_index.cache[pos];
3126                 if (ce_skip_worktree(ce)) {
3127                         pos = index_name_pos(opt->repo->index, path, strlen(path));
3128                         ce = opt->repo->index->cache[pos];
3129                         ce->ce_flags |= CE_SKIP_WORKTREE;
3130                 }
3131                 return mfi->clean;
3132         }
3133
3134         if (!mfi->clean) {
3135                 if (S_ISGITLINK(mfi->blob.mode))
3136                         reason = _("submodule");
3137                 output(opt, 1, _("CONFLICT (%s): Merge conflict in %s"),
3138                                 reason, path);
3139                 if (ci && !df_conflict_remains)
3140                         if (update_stages(opt, path, o, a, b))
3141                                 return -1;
3142         }
3143
3144         if (df_conflict_remains || is_dirty) {
3145                 char *new_path;
3146                 if (opt->priv->call_depth) {
3147                         remove_file_from_index(opt->repo->index, path);
3148                 } else {
3149                         if (!mfi->clean) {
3150                                 if (update_stages(opt, path, o, a, b))
3151                                         return -1;
3152                         } else {
3153                                 int file_from_stage2 = was_tracked(opt, path);
3154
3155                                 if (update_stages(opt, path, NULL,
3156                                                   file_from_stage2 ? &mfi->blob : NULL,
3157                                                   file_from_stage2 ? NULL : &mfi->blob))
3158                                         return -1;
3159                         }
3160
3161                 }
3162                 new_path = unique_path(opt, path, ci->ren1->branch);
3163                 if (is_dirty) {
3164                         output(opt, 1, _("Refusing to lose dirty file at %s"),
3165                                path);
3166                 }
3167                 output(opt, 1, _("Adding as %s instead"), new_path);
3168                 if (update_file(opt, 0, &mfi->blob, new_path)) {
3169                         free(new_path);
3170                         return -1;
3171                 }
3172                 free(new_path);
3173                 mfi->clean = 0;
3174         } else if (update_file(opt, mfi->clean, &mfi->blob, path))
3175                 return -1;
3176         return !is_dirty && mfi->clean;
3177 }
3178
3179 static int handle_rename_normal(struct merge_options *opt,
3180                                 const char *path,
3181                                 const struct diff_filespec *o,
3182                                 const struct diff_filespec *a,
3183                                 const struct diff_filespec *b,
3184                                 struct rename_conflict_info *ci)
3185 {
3186         struct rename *ren = ci->ren1;
3187         struct merge_file_info mfi;
3188         int clean;
3189         int side = (ren->branch == opt->branch1 ? 2 : 3);
3190
3191         /* Merge the content and write it out */
3192         clean = handle_content_merge(&mfi, opt, path, was_dirty(opt, path),
3193                                      o, a, b, ci);
3194
3195         if (clean &&
3196             opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_CONFLICT &&
3197             ren->dir_rename_original_dest) {
3198                 if (update_stages(opt, path,
3199                                   NULL,
3200                                   side == 2 ? &mfi.blob : NULL,
3201                                   side == 2 ? NULL : &mfi.blob))
3202                         return -1;
3203                 clean = 0; /* not clean, but conflicted */
3204         }
3205         return clean;
3206 }
3207
3208 static void dir_rename_warning(const char *msg,
3209                                int is_add,
3210                                int clean,
3211                                struct merge_options *opt,
3212                                struct rename *ren)
3213 {
3214         const char *other_branch;
3215         other_branch = (ren->branch == opt->branch1 ?
3216                         opt->branch2 : opt->branch1);
3217         if (is_add) {
3218                 output(opt, clean ? 2 : 1, msg,
3219                        ren->pair->one->path, ren->branch,
3220                        other_branch, ren->pair->two->path);
3221                 return;
3222         }
3223         output(opt, clean ? 2 : 1, msg,
3224                ren->pair->one->path, ren->dir_rename_original_dest, ren->branch,
3225                other_branch, ren->pair->two->path);
3226 }
3227 static int warn_about_dir_renamed_entries(struct merge_options *opt,
3228                                           struct rename *ren)
3229 {
3230         const char *msg;
3231         int clean = 1, is_add;
3232
3233         if (!ren)
3234                 return clean;
3235
3236         /* Return early if ren was not affected/created by a directory rename */
3237         if (!ren->dir_rename_original_dest)
3238                 return clean;
3239
3240         /* Sanity checks */
3241         assert(opt->detect_directory_renames > MERGE_DIRECTORY_RENAMES_NONE);
3242         assert(ren->dir_rename_original_type == 'A' ||
3243                ren->dir_rename_original_type == 'R');
3244
3245         /* Check whether to treat directory renames as a conflict */
3246         clean = (opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_TRUE);
3247
3248         is_add = (ren->dir_rename_original_type == 'A');
3249         if (ren->dir_rename_original_type == 'A' && clean) {
3250                 msg = _("Path updated: %s added in %s inside a "
3251                         "directory that was renamed in %s; moving it to %s.");
3252         } else if (ren->dir_rename_original_type == 'A' && !clean) {
3253                 msg = _("CONFLICT (file location): %s added in %s "
3254                         "inside a directory that was renamed in %s, "
3255                         "suggesting it should perhaps be moved to %s.");
3256         } else if (ren->dir_rename_original_type == 'R' && clean) {
3257                 msg = _("Path updated: %s renamed to %s in %s, inside a "
3258                         "directory that was renamed in %s; moving it to %s.");
3259         } else if (ren->dir_rename_original_type == 'R' && !clean) {
3260                 msg = _("CONFLICT (file location): %s renamed to %s in %s, "
3261                         "inside a directory that was renamed in %s, "
3262                         "suggesting it should perhaps be moved to %s.");
3263         } else {
3264                 BUG("Impossible dir_rename_original_type/clean combination");
3265         }
3266         dir_rename_warning(msg, is_add, clean, opt, ren);
3267
3268         return clean;
3269 }
3270
3271 /* Per entry merge function */
3272 static int process_entry(struct merge_options *opt,
3273                          const char *path, struct stage_data *entry)
3274 {
3275         int clean_merge = 1;
3276         int normalize = opt->renormalize;
3277
3278         struct diff_filespec *o = &entry->stages[1];
3279         struct diff_filespec *a = &entry->stages[2];
3280         struct diff_filespec *b = &entry->stages[3];
3281         int o_valid = is_valid(o);
3282         int a_valid = is_valid(a);
3283         int b_valid = is_valid(b);
3284         o->path = a->path = b->path = (char*)path;
3285
3286         entry->processed = 1;
3287         if (entry->rename_conflict_info) {
3288                 struct rename_conflict_info *ci = entry->rename_conflict_info;
3289                 struct diff_filespec *temp;
3290                 int path_clean;
3291
3292                 path_clean = warn_about_dir_renamed_entries(opt, ci->ren1);
3293                 path_clean &= warn_about_dir_renamed_entries(opt, ci->ren2);
3294
3295                 /*
3296                  * For cases with a single rename, {o,a,b}->path have all been
3297                  * set to the rename target path; we need to set two of these
3298                  * back to the rename source.
3299                  * For rename/rename conflicts, we'll manually fix paths below.
3300                  */
3301                 temp = (opt->branch1 == ci->ren1->branch) ? b : a;
3302                 o->path = temp->path = ci->ren1->pair->one->path;
3303                 if (ci->ren2) {
3304                         assert(opt->branch1 == ci->ren1->branch);
3305                 }
3306
3307                 switch (ci->rename_type) {
3308                 case RENAME_NORMAL:
3309                 case RENAME_ONE_FILE_TO_ONE:
3310                         clean_merge = handle_rename_normal(opt, path, o, a, b,
3311                                                            ci);
3312                         break;
3313                 case RENAME_VIA_DIR:
3314                         clean_merge = handle_rename_via_dir(opt, ci);
3315                         break;
3316                 case RENAME_ADD:
3317                         /*
3318                          * Probably unclean merge, but if the renamed file
3319                          * merges cleanly and the result can then be
3320                          * two-way merged cleanly with the added file, I
3321                          * guess it's a clean merge?
3322                          */
3323                         clean_merge = handle_rename_add(opt, ci);
3324                         break;
3325                 case RENAME_DELETE:
3326                         clean_merge = 0;
3327                         if (handle_rename_delete(opt, ci))
3328                                 clean_merge = -1;
3329                         break;
3330                 case RENAME_ONE_FILE_TO_TWO:
3331                         /*
3332                          * Manually fix up paths; note:
3333                          * ren[12]->pair->one->path are equal.
3334                          */
3335                         o->path = ci->ren1->pair->one->path;
3336                         a->path = ci->ren1->pair->two->path;
3337                         b->path = ci->ren2->pair->two->path;
3338
3339                         clean_merge = 0;
3340                         if (handle_rename_rename_1to2(opt, ci))
3341                                 clean_merge = -1;
3342                         break;
3343                 case RENAME_TWO_FILES_TO_ONE:
3344                         /*
3345                          * Manually fix up paths; note,
3346                          * ren[12]->pair->two->path are actually equal.
3347                          */
3348                         o->path = NULL;
3349                         a->path = ci->ren1->pair->two->path;
3350                         b->path = ci->ren2->pair->two->path;
3351
3352                         /*
3353                          * Probably unclean merge, but if the two renamed
3354                          * files merge cleanly and the two resulting files
3355                          * can then be two-way merged cleanly, I guess it's
3356                          * a clean merge?
3357                          */
3358                         clean_merge = handle_rename_rename_2to1(opt, ci);
3359                         break;
3360                 default:
3361                         entry->processed = 0;
3362                         break;
3363                 }
3364                 if (path_clean < clean_merge)
3365                         clean_merge = path_clean;
3366         } else if (o_valid && (!a_valid || !b_valid)) {
3367                 /* Case A: Deleted in one */
3368                 if ((!a_valid && !b_valid) ||
3369                     (!b_valid && blob_unchanged(opt, o, a, normalize, path)) ||
3370                     (!a_valid && blob_unchanged(opt, o, b, normalize, path))) {
3371                         /* Deleted in both or deleted in one and
3372                          * unchanged in the other */
3373                         if (a_valid)
3374                                 output(opt, 2, _("Removing %s"), path);
3375                         /* do not touch working file if it did not exist */
3376                         remove_file(opt, 1, path, !a_valid);
3377                 } else {
3378                         /* Modify/delete; deleted side may have put a directory in the way */
3379                         clean_merge = 0;
3380                         if (handle_modify_delete(opt, path, o, a, b))
3381                                 clean_merge = -1;
3382                 }
3383         } else if ((!o_valid && a_valid && !b_valid) ||
3384                    (!o_valid && !a_valid && b_valid)) {
3385                 /* Case B: Added in one. */
3386                 /* [nothing|directory] -> ([nothing|directory], file) */
3387
3388                 const char *add_branch;
3389                 const char *other_branch;
3390                 const char *conf;
3391                 const struct diff_filespec *contents;
3392
3393                 if (a_valid) {
3394                         add_branch = opt->branch1;
3395                         other_branch = opt->branch2;
3396                         contents = a;
3397                         conf = _("file/directory");
3398                 } else {
3399                         add_branch = opt->branch2;
3400                         other_branch = opt->branch1;
3401                         contents = b;
3402                         conf = _("directory/file");
3403                 }
3404                 if (dir_in_way(opt->repo->index, path,
3405                                !opt->priv->call_depth && !S_ISGITLINK(a->mode),
3406                                0)) {
3407                         char *new_path = unique_path(opt, path, add_branch);
3408                         clean_merge = 0;
3409                         output(opt, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3410                                "Adding %s as %s"),
3411                                conf, path, other_branch, path, new_path);
3412                         if (update_file(opt, 0, contents, new_path))
3413                                 clean_merge = -1;
3414                         else if (opt->priv->call_depth)
3415                                 remove_file_from_index(opt->repo->index, path);
3416                         free(new_path);
3417                 } else {
3418                         output(opt, 2, _("Adding %s"), path);
3419                         /* do not overwrite file if already present */
3420                         if (update_file_flags(opt, contents, path, 1, !a_valid))
3421                                 clean_merge = -1;
3422                 }
3423         } else if (a_valid && b_valid) {
3424                 if (!o_valid) {
3425                         /* Case C: Added in both (check for same permissions) */
3426                         output(opt, 1,
3427                                _("CONFLICT (add/add): Merge conflict in %s"),
3428                                path);
3429                         clean_merge = handle_file_collision(opt,
3430                                                             path, NULL, NULL,
3431                                                             opt->branch1,
3432                                                             opt->branch2,
3433                                                             a, b);
3434                 } else {
3435                         /* case D: Modified in both, but differently. */
3436                         struct merge_file_info mfi;
3437                         int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3438                         clean_merge = handle_content_merge(&mfi, opt, path,
3439                                                            is_dirty,
3440                                                            o, a, b, NULL);
3441                 }
3442         } else if (!o_valid && !a_valid && !b_valid) {
3443                 /*
3444                  * this entry was deleted altogether. a_mode == 0 means
3445                  * we had that path and want to actively remove it.
3446                  */
3447                 remove_file(opt, 1, path, !a->mode);
3448         } else
3449                 BUG("fatal merge failure, shouldn't happen.");
3450
3451         return clean_merge;
3452 }
3453
3454 static int merge_trees_internal(struct merge_options *opt,
3455                                 struct tree *head,
3456                                 struct tree *merge,
3457                                 struct tree *merge_base,
3458                                 struct tree **result)
3459 {
3460         struct index_state *istate = opt->repo->index;
3461         int code, clean;
3462
3463         if (opt->subtree_shift) {
3464                 merge = shift_tree_object(opt->repo, head, merge,
3465                                           opt->subtree_shift);
3466                 merge_base = shift_tree_object(opt->repo, head, merge_base,
3467                                                opt->subtree_shift);
3468         }
3469
3470         if (oideq(&merge_base->object.oid, &merge->object.oid)) {
3471                 output(opt, 0, _("Already up to date!"));
3472                 *result = head;
3473                 return 1;
3474         }
3475
3476         code = unpack_trees_start(opt, merge_base, head, merge);
3477
3478         if (code != 0) {
3479                 if (show(opt, 4) || opt->priv->call_depth)
3480                         err(opt, _("merging of trees %s and %s failed"),
3481                             oid_to_hex(&head->object.oid),
3482                             oid_to_hex(&merge->object.oid));
3483                 unpack_trees_finish(opt);
3484                 return -1;
3485         }
3486
3487         if (unmerged_index(istate)) {
3488                 struct string_list *entries;
3489                 struct rename_info re_info;
3490                 int i;
3491                 /*
3492                  * Only need the hashmap while processing entries, so
3493                  * initialize it here and free it when we are done running
3494                  * through the entries. Keeping it in the merge_options as
3495                  * opposed to decaring a local hashmap is for convenience
3496                  * so that we don't have to pass it to around.
3497                  */
3498                 hashmap_init(&opt->priv->current_file_dir_set, path_hashmap_cmp,
3499                              NULL, 512);
3500                 get_files_dirs(opt, head);
3501                 get_files_dirs(opt, merge);
3502
3503                 entries = get_unmerged(opt->repo->index);
3504                 clean = detect_and_process_renames(opt, merge_base, head, merge,
3505                                                    entries, &re_info);
3506                 record_df_conflict_files(opt, entries);
3507                 if (clean < 0)
3508                         goto cleanup;
3509                 for (i = entries->nr-1; 0 <= i; i--) {
3510                         const char *path = entries->items[i].string;
3511                         struct stage_data *e = entries->items[i].util;
3512                         if (!e->processed) {
3513                                 int ret = process_entry(opt, path, e);
3514                                 if (!ret)
3515                                         clean = 0;
3516                                 else if (ret < 0) {
3517                                         clean = ret;
3518                                         goto cleanup;
3519                                 }
3520                         }
3521                 }
3522                 for (i = 0; i < entries->nr; i++) {
3523                         struct stage_data *e = entries->items[i].util;
3524                         if (!e->processed)
3525                                 BUG("unprocessed path??? %s",
3526                                     entries->items[i].string);
3527                 }
3528
3529         cleanup:
3530                 final_cleanup_renames(&re_info);
3531
3532                 string_list_clear(entries, 1);
3533                 free(entries);
3534
3535                 hashmap_free_entries(&opt->priv->current_file_dir_set,
3536                                         struct path_hashmap_entry, e);
3537
3538                 if (clean < 0) {
3539                         unpack_trees_finish(opt);
3540                         return clean;
3541                 }
3542         }
3543         else
3544                 clean = 1;
3545
3546         unpack_trees_finish(opt);
3547
3548         if (opt->priv->call_depth &&
3549             !(*result = write_in_core_index_as_tree(opt->repo)))
3550                 return -1;
3551
3552         return clean;
3553 }
3554
3555 static struct commit_list *reverse_commit_list(struct commit_list *list)
3556 {
3557         struct commit_list *next = NULL, *current, *backup;
3558         for (current = list; current; current = backup) {
3559                 backup = current->next;
3560                 current->next = next;
3561                 next = current;
3562         }
3563         return next;
3564 }
3565
3566 /*
3567  * Merge the commits h1 and h2, return the resulting virtual
3568  * commit object and a flag indicating the cleanness of the merge.
3569  */
3570 static int merge_recursive_internal(struct merge_options *opt,
3571                                     struct commit *h1,
3572                                     struct commit *h2,
3573                                     struct commit_list *merge_bases,
3574                                     struct commit **result)
3575 {
3576         struct commit_list *iter;
3577         struct commit *merged_merge_bases;
3578         struct tree *result_tree;
3579         int clean;
3580         const char *ancestor_name;
3581         struct strbuf merge_base_abbrev = STRBUF_INIT;
3582
3583         if (show(opt, 4)) {
3584                 output(opt, 4, _("Merging:"));
3585                 output_commit_title(opt, h1);
3586                 output_commit_title(opt, h2);
3587         }
3588
3589         if (!merge_bases) {
3590                 merge_bases = get_merge_bases(h1, h2);
3591                 merge_bases = reverse_commit_list(merge_bases);
3592         }
3593
3594         if (show(opt, 5)) {
3595                 unsigned cnt = commit_list_count(merge_bases);
3596
3597                 output(opt, 5, Q_("found %u common ancestor:",
3598                                 "found %u common ancestors:", cnt), cnt);
3599                 for (iter = merge_bases; iter; iter = iter->next)
3600                         output_commit_title(opt, iter->item);
3601         }
3602
3603         merged_merge_bases = pop_commit(&merge_bases);
3604         if (merged_merge_bases == NULL) {
3605                 /* if there is no common ancestor, use an empty tree */
3606                 struct tree *tree;
3607
3608                 tree = lookup_tree(opt->repo, opt->repo->hash_algo->empty_tree);
3609                 merged_merge_bases = make_virtual_commit(opt->repo, tree,
3610                                                          "ancestor");
3611                 ancestor_name = "empty tree";
3612         } else if (opt->ancestor && !opt->priv->call_depth) {
3613                 ancestor_name = opt->ancestor;
3614         } else if (merge_bases) {
3615                 ancestor_name = "merged common ancestors";
3616         } else {
3617                 strbuf_add_unique_abbrev(&merge_base_abbrev,
3618                                          &merged_merge_bases->object.oid,
3619                                          DEFAULT_ABBREV);
3620                 ancestor_name = merge_base_abbrev.buf;
3621         }
3622
3623         for (iter = merge_bases; iter; iter = iter->next) {
3624                 const char *saved_b1, *saved_b2;
3625                 opt->priv->call_depth++;
3626                 /*
3627                  * When the merge fails, the result contains files
3628                  * with conflict markers. The cleanness flag is
3629                  * ignored (unless indicating an error), it was never
3630                  * actually used, as result of merge_trees has always
3631                  * overwritten it: the committed "conflicts" were
3632                  * already resolved.
3633                  */
3634                 discard_index(opt->repo->index);
3635                 saved_b1 = opt->branch1;
3636                 saved_b2 = opt->branch2;
3637                 opt->branch1 = "Temporary merge branch 1";
3638                 opt->branch2 = "Temporary merge branch 2";
3639                 if (merge_recursive_internal(opt, merged_merge_bases, iter->item,
3640                                              NULL, &merged_merge_bases) < 0)
3641                         return -1;
3642                 opt->branch1 = saved_b1;
3643                 opt->branch2 = saved_b2;
3644                 opt->priv->call_depth--;
3645
3646                 if (!merged_merge_bases)
3647                         return err(opt, _("merge returned no commit"));
3648         }
3649
3650         discard_index(opt->repo->index);
3651         if (!opt->priv->call_depth)
3652                 repo_read_index(opt->repo);
3653
3654         opt->ancestor = ancestor_name;
3655         clean = merge_trees_internal(opt,
3656                                      repo_get_commit_tree(opt->repo, h1),
3657                                      repo_get_commit_tree(opt->repo, h2),
3658                                      repo_get_commit_tree(opt->repo,
3659                                                           merged_merge_bases),
3660                                      &result_tree);
3661         strbuf_release(&merge_base_abbrev);
3662         opt->ancestor = NULL;  /* avoid accidental re-use of opt->ancestor */
3663         if (clean < 0) {
3664                 flush_output(opt);
3665                 return clean;
3666         }
3667
3668         if (opt->priv->call_depth) {
3669                 *result = make_virtual_commit(opt->repo, result_tree,
3670                                               "merged tree");
3671                 commit_list_insert(h1, &(*result)->parents);
3672                 commit_list_insert(h2, &(*result)->parents->next);
3673         }
3674         return clean;
3675 }
3676
3677 static int merge_start(struct merge_options *opt, struct tree *head)
3678 {
3679         struct strbuf sb = STRBUF_INIT;
3680
3681         /* Sanity checks on opt */
3682         assert(opt->repo);
3683
3684         assert(opt->branch1 && opt->branch2);
3685
3686         assert(opt->detect_renames >= -1 &&
3687                opt->detect_renames <= DIFF_DETECT_COPY);
3688         assert(opt->detect_directory_renames >= MERGE_DIRECTORY_RENAMES_NONE &&
3689                opt->detect_directory_renames <= MERGE_DIRECTORY_RENAMES_TRUE);
3690         assert(opt->rename_limit >= -1);
3691         assert(opt->rename_score >= 0 && opt->rename_score <= MAX_SCORE);
3692         assert(opt->show_rename_progress >= 0 && opt->show_rename_progress <= 1);
3693
3694         assert(opt->xdl_opts >= 0);
3695         assert(opt->recursive_variant >= MERGE_VARIANT_NORMAL &&
3696                opt->recursive_variant <= MERGE_VARIANT_THEIRS);
3697
3698         assert(opt->verbosity >= 0 && opt->verbosity <= 5);
3699         assert(opt->buffer_output <= 2);
3700         assert(opt->obuf.len == 0);
3701
3702         assert(opt->priv == NULL);
3703
3704         /* Sanity check on repo state; index must match head */
3705         if (repo_index_has_changes(opt->repo, head, &sb)) {
3706                 err(opt, _("Your local changes to the following files would be overwritten by merge:\n  %s"),
3707                     sb.buf);
3708                 strbuf_release(&sb);
3709                 return -1;
3710         }
3711
3712         opt->priv = xcalloc(1, sizeof(*opt->priv));
3713         string_list_init(&opt->priv->df_conflict_file_set, 1);
3714         return 0;
3715 }
3716
3717 static void merge_finalize(struct merge_options *opt)
3718 {
3719         flush_output(opt);
3720         if (!opt->priv->call_depth && opt->buffer_output < 2)
3721                 strbuf_release(&opt->obuf);
3722         if (show(opt, 2))
3723                 diff_warn_rename_limit("merge.renamelimit",
3724                                        opt->priv->needed_rename_limit, 0);
3725         FREE_AND_NULL(opt->priv);
3726 }
3727
3728 int merge_trees(struct merge_options *opt,
3729                 struct tree *head,
3730                 struct tree *merge,
3731                 struct tree *merge_base)
3732 {
3733         int clean;
3734         struct tree *ignored;
3735
3736         assert(opt->ancestor != NULL);
3737
3738         if (merge_start(opt, head))
3739                 return -1;
3740         clean = merge_trees_internal(opt, head, merge, merge_base, &ignored);
3741         merge_finalize(opt);
3742
3743         return clean;
3744 }
3745
3746 int merge_recursive(struct merge_options *opt,
3747                     struct commit *h1,
3748                     struct commit *h2,
3749                     struct commit_list *merge_bases,
3750                     struct commit **result)
3751 {
3752         int clean;
3753
3754         assert(opt->ancestor == NULL ||
3755                !strcmp(opt->ancestor, "constructed merge base"));
3756
3757         if (merge_start(opt, repo_get_commit_tree(opt->repo, h1)))
3758                 return -1;
3759         clean = merge_recursive_internal(opt, h1, h2, merge_bases, result);
3760         merge_finalize(opt);
3761
3762         return clean;
3763 }
3764
3765 static struct commit *get_ref(struct repository *repo,
3766                               const struct object_id *oid,
3767                               const char *name)
3768 {
3769         struct object *object;
3770
3771         object = deref_tag(repo, parse_object(repo, oid),
3772                            name, strlen(name));
3773         if (!object)
3774                 return NULL;
3775         if (object->type == OBJ_TREE)
3776                 return make_virtual_commit(repo, (struct tree*)object, name);
3777         if (object->type != OBJ_COMMIT)
3778                 return NULL;
3779         if (parse_commit((struct commit *)object))
3780                 return NULL;
3781         return (struct commit *)object;
3782 }
3783
3784 int merge_recursive_generic(struct merge_options *opt,
3785                             const struct object_id *head,
3786                             const struct object_id *merge,
3787                             int num_merge_bases,
3788                             const struct object_id **merge_bases,
3789                             struct commit **result)
3790 {
3791         int clean;
3792         struct lock_file lock = LOCK_INIT;
3793         struct commit *head_commit = get_ref(opt->repo, head, opt->branch1);
3794         struct commit *next_commit = get_ref(opt->repo, merge, opt->branch2);
3795         struct commit_list *ca = NULL;
3796
3797         if (merge_bases) {
3798                 int i;
3799                 for (i = 0; i < num_merge_bases; ++i) {
3800                         struct commit *base;
3801                         if (!(base = get_ref(opt->repo, merge_bases[i],
3802                                              oid_to_hex(merge_bases[i]))))
3803                                 return err(opt, _("Could not parse object '%s'"),
3804                                            oid_to_hex(merge_bases[i]));
3805                         commit_list_insert(base, &ca);
3806                 }
3807                 if (num_merge_bases == 1)
3808                         opt->ancestor = "constructed merge base";
3809         }
3810
3811         repo_hold_locked_index(opt->repo, &lock, LOCK_DIE_ON_ERROR);
3812         clean = merge_recursive(opt, head_commit, next_commit, ca,
3813                                 result);
3814         if (clean < 0) {
3815                 rollback_lock_file(&lock);
3816                 return clean;
3817         }
3818
3819         if (write_locked_index(opt->repo->index, &lock,
3820                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3821                 return err(opt, _("Unable to write index."));
3822
3823         return clean ? 0 : 1;
3824 }
3825
3826 static void merge_recursive_config(struct merge_options *opt)
3827 {
3828         char *value = NULL;
3829         git_config_get_int("merge.verbosity", &opt->verbosity);
3830         git_config_get_int("diff.renamelimit", &opt->rename_limit);
3831         git_config_get_int("merge.renamelimit", &opt->rename_limit);
3832         if (!git_config_get_string("diff.renames", &value)) {
3833                 opt->detect_renames = git_config_rename("diff.renames", value);
3834                 free(value);
3835         }
3836         if (!git_config_get_string("merge.renames", &value)) {
3837                 opt->detect_renames = git_config_rename("merge.renames", value);
3838                 free(value);
3839         }
3840         if (!git_config_get_string("merge.directoryrenames", &value)) {
3841                 int boolval = git_parse_maybe_bool(value);
3842                 if (0 <= boolval) {
3843                         opt->detect_directory_renames = boolval ?
3844                                 MERGE_DIRECTORY_RENAMES_TRUE :
3845                                 MERGE_DIRECTORY_RENAMES_NONE;
3846                 } else if (!strcasecmp(value, "conflict")) {
3847                         opt->detect_directory_renames =
3848                                 MERGE_DIRECTORY_RENAMES_CONFLICT;
3849                 } /* avoid erroring on values from future versions of git */
3850                 free(value);
3851         }
3852         git_config(git_xmerge_config, NULL);
3853 }
3854
3855 void init_merge_options(struct merge_options *opt,
3856                         struct repository *repo)
3857 {
3858         const char *merge_verbosity;
3859         memset(opt, 0, sizeof(struct merge_options));
3860
3861         opt->repo = repo;
3862
3863         opt->detect_renames = -1;
3864         opt->detect_directory_renames = MERGE_DIRECTORY_RENAMES_CONFLICT;
3865         opt->rename_limit = -1;
3866
3867         opt->verbosity = 2;
3868         opt->buffer_output = 1;
3869         strbuf_init(&opt->obuf, 0);
3870
3871         opt->renormalize = 0;
3872
3873         merge_recursive_config(opt);
3874         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3875         if (merge_verbosity)
3876                 opt->verbosity = strtol(merge_verbosity, NULL, 10);
3877         if (opt->verbosity >= 5)
3878                 opt->buffer_output = 0;
3879 }
3880
3881 int parse_merge_opt(struct merge_options *opt, const char *s)
3882 {
3883         const char *arg;
3884
3885         if (!s || !*s)
3886                 return -1;
3887         if (!strcmp(s, "ours"))
3888                 opt->recursive_variant = MERGE_VARIANT_OURS;
3889         else if (!strcmp(s, "theirs"))
3890                 opt->recursive_variant = MERGE_VARIANT_THEIRS;
3891         else if (!strcmp(s, "subtree"))
3892                 opt->subtree_shift = "";
3893         else if (skip_prefix(s, "subtree=", &arg))
3894                 opt->subtree_shift = arg;
3895         else if (!strcmp(s, "patience"))
3896                 opt->xdl_opts = DIFF_WITH_ALG(opt, PATIENCE_DIFF);
3897         else if (!strcmp(s, "histogram"))
3898                 opt->xdl_opts = DIFF_WITH_ALG(opt, HISTOGRAM_DIFF);
3899         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3900                 long value = parse_algorithm_value(arg);
3901                 if (value < 0)
3902                         return -1;
3903                 /* clear out previous settings */
3904                 DIFF_XDL_CLR(opt, NEED_MINIMAL);
3905                 opt->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3906                 opt->xdl_opts |= value;
3907         }
3908         else if (!strcmp(s, "ignore-space-change"))
3909                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_CHANGE);
3910         else if (!strcmp(s, "ignore-all-space"))
3911                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE);
3912         else if (!strcmp(s, "ignore-space-at-eol"))
3913                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_AT_EOL);
3914         else if (!strcmp(s, "ignore-cr-at-eol"))
3915                 DIFF_XDL_SET(opt, IGNORE_CR_AT_EOL);
3916         else if (!strcmp(s, "renormalize"))
3917                 opt->renormalize = 1;
3918         else if (!strcmp(s, "no-renormalize"))
3919                 opt->renormalize = 0;
3920         else if (!strcmp(s, "no-renames"))
3921                 opt->detect_renames = 0;
3922         else if (!strcmp(s, "find-renames")) {
3923                 opt->detect_renames = 1;
3924                 opt->rename_score = 0;
3925         }
3926         else if (skip_prefix(s, "find-renames=", &arg) ||
3927                  skip_prefix(s, "rename-threshold=", &arg)) {
3928                 if ((opt->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3929                         return -1;
3930                 opt->detect_renames = 1;
3931         }
3932         /*
3933          * Please update $__git_merge_strategy_options in
3934          * git-completion.bash when you add new options
3935          */
3936         else
3937                 return -1;
3938         return 0;
3939 }