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