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