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