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