Merge branch 'cc/replace-graft-peel-tags'
[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[GIT_MAX_HEXSZ + 2];
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         if (merge_mode_and_contents(opt, a, c,
1659                                     &ci->ren1->src_entry->stages[other_stage],
1660                                     prev_path_desc,
1661                                     opt->branch1, opt->branch2,
1662                                     1 + opt->call_depth * 2, &mfi))
1663                 return -1;
1664         free(prev_path_desc);
1665
1666         ci->ren1->dst_entry->stages[other_stage].path = mfi.blob.path = c->path;
1667         return handle_file_collision(opt,
1668                                      c->path, a->path, NULL,
1669                                      rename_branch, add_branch,
1670                                      &mfi.blob,
1671                                      &ci->ren1->dst_entry->stages[other_stage]);
1672 }
1673
1674 static char *find_path_for_conflict(struct merge_options *opt,
1675                                     const char *path,
1676                                     const char *branch1,
1677                                     const char *branch2)
1678 {
1679         char *new_path = NULL;
1680         if (dir_in_way(opt->repo->index, path, !opt->call_depth, 0)) {
1681                 new_path = unique_path(opt, path, branch1);
1682                 output(opt, 1, _("%s is a directory in %s adding "
1683                                "as %s instead"),
1684                        path, branch2, new_path);
1685         } else if (would_lose_untracked(opt, path)) {
1686                 new_path = unique_path(opt, path, branch1);
1687                 output(opt, 1, _("Refusing to lose untracked file"
1688                                " at %s; adding as %s instead"),
1689                        path, new_path);
1690         }
1691
1692         return new_path;
1693 }
1694
1695 static int handle_rename_rename_1to2(struct merge_options *opt,
1696                                      struct rename_conflict_info *ci)
1697 {
1698         /* One file was renamed in both branches, but to different names. */
1699         struct merge_file_info mfi;
1700         struct diff_filespec *add;
1701         struct diff_filespec *o = ci->ren1->pair->one;
1702         struct diff_filespec *a = ci->ren1->pair->two;
1703         struct diff_filespec *b = ci->ren2->pair->two;
1704         char *path_desc;
1705
1706         output(opt, 1, _("CONFLICT (rename/rename): "
1707                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1708                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1709                o->path, a->path, ci->ren1->branch,
1710                o->path, b->path, ci->ren2->branch,
1711                opt->call_depth ? _(" (left unresolved)") : "");
1712
1713         path_desc = xstrfmt("%s and %s, both renamed from %s",
1714                             a->path, b->path, o->path);
1715         if (merge_mode_and_contents(opt, o, a, b, path_desc,
1716                                     ci->ren1->branch, ci->ren2->branch,
1717                                     opt->call_depth * 2, &mfi))
1718                 return -1;
1719         free(path_desc);
1720
1721         if (opt->call_depth) {
1722                 /*
1723                  * FIXME: For rename/add-source conflicts (if we could detect
1724                  * such), this is wrong.  We should instead find a unique
1725                  * pathname and then either rename the add-source file to that
1726                  * unique path, or use that unique path instead of src here.
1727                  */
1728                 if (update_file(opt, 0, &mfi.blob, o->path))
1729                         return -1;
1730
1731                 /*
1732                  * Above, we put the merged content at the merge-base's
1733                  * path.  Now we usually need to delete both a->path and
1734                  * b->path.  However, the rename on each side of the merge
1735                  * could also be involved in a rename/add conflict.  In
1736                  * such cases, we should keep the added file around,
1737                  * resolving the conflict at that path in its favor.
1738                  */
1739                 add = &ci->ren1->dst_entry->stages[2 ^ 1];
1740                 if (is_valid(add)) {
1741                         if (update_file(opt, 0, add, a->path))
1742                                 return -1;
1743                 }
1744                 else
1745                         remove_file_from_index(opt->repo->index, a->path);
1746                 add = &ci->ren2->dst_entry->stages[3 ^ 1];
1747                 if (is_valid(add)) {
1748                         if (update_file(opt, 0, add, b->path))
1749                                 return -1;
1750                 }
1751                 else
1752                         remove_file_from_index(opt->repo->index, b->path);
1753         } else {
1754                 /*
1755                  * For each destination path, we need to see if there is a
1756                  * rename/add collision.  If not, we can write the file out
1757                  * to the specified location.
1758                  */
1759                 add = &ci->ren1->dst_entry->stages[2 ^ 1];
1760                 if (is_valid(add)) {
1761                         add->path = mfi.blob.path = a->path;
1762                         if (handle_file_collision(opt, a->path,
1763                                                   NULL, NULL,
1764                                                   ci->ren1->branch,
1765                                                   ci->ren2->branch,
1766                                                   &mfi.blob, add) < 0)
1767                                 return -1;
1768                 } else {
1769                         char *new_path = find_path_for_conflict(opt, a->path,
1770                                                                 ci->ren1->branch,
1771                                                                 ci->ren2->branch);
1772                         if (update_file(opt, 0, &mfi.blob,
1773                                         new_path ? new_path : a->path))
1774                                 return -1;
1775                         free(new_path);
1776                         if (update_stages(opt, a->path, NULL, a, NULL))
1777                                 return -1;
1778                 }
1779
1780                 add = &ci->ren2->dst_entry->stages[3 ^ 1];
1781                 if (is_valid(add)) {
1782                         add->path = mfi.blob.path = b->path;
1783                         if (handle_file_collision(opt, b->path,
1784                                                   NULL, NULL,
1785                                                   ci->ren1->branch,
1786                                                   ci->ren2->branch,
1787                                                   add, &mfi.blob) < 0)
1788                                 return -1;
1789                 } else {
1790                         char *new_path = find_path_for_conflict(opt, b->path,
1791                                                                 ci->ren2->branch,
1792                                                                 ci->ren1->branch);
1793                         if (update_file(opt, 0, &mfi.blob,
1794                                         new_path ? new_path : b->path))
1795                                 return -1;
1796                         free(new_path);
1797                         if (update_stages(opt, b->path, NULL, NULL, b))
1798                                 return -1;
1799                 }
1800         }
1801
1802         return 0;
1803 }
1804
1805 static int handle_rename_rename_2to1(struct merge_options *opt,
1806                                      struct rename_conflict_info *ci)
1807 {
1808         /* Two files, a & b, were renamed to the same thing, c. */
1809         struct diff_filespec *a = ci->ren1->pair->one;
1810         struct diff_filespec *b = ci->ren2->pair->one;
1811         struct diff_filespec *c1 = ci->ren1->pair->two;
1812         struct diff_filespec *c2 = ci->ren2->pair->two;
1813         char *path = c1->path; /* == c2->path */
1814         char *path_side_1_desc;
1815         char *path_side_2_desc;
1816         struct merge_file_info mfi_c1;
1817         struct merge_file_info mfi_c2;
1818         int ostage1, ostage2;
1819
1820         output(opt, 1, _("CONFLICT (rename/rename): "
1821                "Rename %s->%s in %s. "
1822                "Rename %s->%s in %s"),
1823                a->path, c1->path, ci->ren1->branch,
1824                b->path, c2->path, ci->ren2->branch);
1825
1826         path_side_1_desc = xstrfmt("version of %s from %s", path, a->path);
1827         path_side_2_desc = xstrfmt("version of %s from %s", path, b->path);
1828         ostage1 = ci->ren1->branch == opt->branch1 ? 3 : 2;
1829         ostage2 = ostage1 ^ 1;
1830         ci->ren1->src_entry->stages[ostage1].path = a->path;
1831         ci->ren2->src_entry->stages[ostage2].path = b->path;
1832         if (merge_mode_and_contents(opt, a, c1,
1833                                     &ci->ren1->src_entry->stages[ostage1],
1834                                     path_side_1_desc,
1835                                     opt->branch1, opt->branch2,
1836                                     1 + opt->call_depth * 2, &mfi_c1) ||
1837             merge_mode_and_contents(opt, b,
1838                                     &ci->ren2->src_entry->stages[ostage2],
1839                                     c2, path_side_2_desc,
1840                                     opt->branch1, opt->branch2,
1841                                     1 + opt->call_depth * 2, &mfi_c2))
1842                 return -1;
1843         free(path_side_1_desc);
1844         free(path_side_2_desc);
1845         mfi_c1.blob.path = path;
1846         mfi_c2.blob.path = path;
1847
1848         return handle_file_collision(opt, path, a->path, b->path,
1849                                      ci->ren1->branch, ci->ren2->branch,
1850                                      &mfi_c1.blob, &mfi_c2.blob);
1851 }
1852
1853 /*
1854  * Get the diff_filepairs changed between o_tree and tree.
1855  */
1856 static struct diff_queue_struct *get_diffpairs(struct merge_options *opt,
1857                                                struct tree *o_tree,
1858                                                struct tree *tree)
1859 {
1860         struct diff_queue_struct *ret;
1861         struct diff_options opts;
1862
1863         repo_diff_setup(opt->repo, &opts);
1864         opts.flags.recursive = 1;
1865         opts.flags.rename_empty = 0;
1866         opts.detect_rename = merge_detect_rename(opt);
1867         /*
1868          * We do not have logic to handle the detection of copies.  In
1869          * fact, it may not even make sense to add such logic: would we
1870          * really want a change to a base file to be propagated through
1871          * multiple other files by a merge?
1872          */
1873         if (opts.detect_rename > DIFF_DETECT_RENAME)
1874                 opts.detect_rename = DIFF_DETECT_RENAME;
1875         opts.rename_limit = opt->merge_rename_limit >= 0 ? opt->merge_rename_limit :
1876                             opt->diff_rename_limit >= 0 ? opt->diff_rename_limit :
1877                             1000;
1878         opts.rename_score = opt->rename_score;
1879         opts.show_rename_progress = opt->show_rename_progress;
1880         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1881         diff_setup_done(&opts);
1882         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1883         diffcore_std(&opts);
1884         if (opts.needed_rename_limit > opt->needed_rename_limit)
1885                 opt->needed_rename_limit = opts.needed_rename_limit;
1886
1887         ret = xmalloc(sizeof(*ret));
1888         *ret = diff_queued_diff;
1889
1890         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1891         diff_queued_diff.nr = 0;
1892         diff_queued_diff.queue = NULL;
1893         diff_flush(&opts);
1894         return ret;
1895 }
1896
1897 static int tree_has_path(struct tree *tree, const char *path)
1898 {
1899         struct object_id hashy;
1900         unsigned short mode_o;
1901
1902         return !get_tree_entry(&tree->object.oid, path,
1903                                &hashy, &mode_o);
1904 }
1905
1906 /*
1907  * Return a new string that replaces the beginning portion (which matches
1908  * entry->dir), with entry->new_dir.  In perl-speak:
1909  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1910  * NOTE:
1911  *   Caller must ensure that old_path starts with entry->dir + '/'.
1912  */
1913 static char *apply_dir_rename(struct dir_rename_entry *entry,
1914                               const char *old_path)
1915 {
1916         struct strbuf new_path = STRBUF_INIT;
1917         int oldlen, newlen;
1918
1919         if (entry->non_unique_new_dir)
1920                 return NULL;
1921
1922         oldlen = strlen(entry->dir);
1923         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1924         strbuf_grow(&new_path, newlen);
1925         strbuf_addbuf(&new_path, &entry->new_dir);
1926         strbuf_addstr(&new_path, &old_path[oldlen]);
1927
1928         return strbuf_detach(&new_path, NULL);
1929 }
1930
1931 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1932                                     char **old_dir, char **new_dir)
1933 {
1934         char *end_of_old, *end_of_new;
1935         int old_len, new_len;
1936
1937         *old_dir = NULL;
1938         *new_dir = NULL;
1939
1940         /*
1941          * For
1942          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1943          * the "e/foo.c" part is the same, we just want to know that
1944          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1945          * so, for this example, this function returns "a/b/c/d" in
1946          * *old_dir and "a/b/some/thing/else" in *new_dir.
1947          *
1948          * Also, if the basename of the file changed, we don't care.  We
1949          * want to know which portion of the directory, if any, changed.
1950          */
1951         end_of_old = strrchr(old_path, '/');
1952         end_of_new = strrchr(new_path, '/');
1953
1954         if (end_of_old == NULL || end_of_new == NULL)
1955                 return;
1956         while (*--end_of_new == *--end_of_old &&
1957                end_of_old != old_path &&
1958                end_of_new != new_path)
1959                 ; /* Do nothing; all in the while loop */
1960         /*
1961          * We've found the first non-matching character in the directory
1962          * paths.  That means the current directory we were comparing
1963          * represents the rename.  Move end_of_old and end_of_new back
1964          * to the full directory name.
1965          */
1966         if (*end_of_old == '/')
1967                 end_of_old++;
1968         if (*end_of_old != '/')
1969                 end_of_new++;
1970         end_of_old = strchr(end_of_old, '/');
1971         end_of_new = strchr(end_of_new, '/');
1972
1973         /*
1974          * It may have been the case that old_path and new_path were the same
1975          * directory all along.  Don't claim a rename if they're the same.
1976          */
1977         old_len = end_of_old - old_path;
1978         new_len = end_of_new - new_path;
1979
1980         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1981                 *old_dir = xstrndup(old_path, old_len);
1982                 *new_dir = xstrndup(new_path, new_len);
1983         }
1984 }
1985
1986 static void remove_hashmap_entries(struct hashmap *dir_renames,
1987                                    struct string_list *items_to_remove)
1988 {
1989         int i;
1990         struct dir_rename_entry *entry;
1991
1992         for (i = 0; i < items_to_remove->nr; i++) {
1993                 entry = items_to_remove->items[i].util;
1994                 hashmap_remove(dir_renames, entry, NULL);
1995         }
1996         string_list_clear(items_to_remove, 0);
1997 }
1998
1999 /*
2000  * See if there is a directory rename for path, and if there are any file
2001  * level conflicts for the renamed location.  If there is a rename and
2002  * there are no conflicts, return the new name.  Otherwise, return NULL.
2003  */
2004 static char *handle_path_level_conflicts(struct merge_options *opt,
2005                                          const char *path,
2006                                          struct dir_rename_entry *entry,
2007                                          struct hashmap *collisions,
2008                                          struct tree *tree)
2009 {
2010         char *new_path = NULL;
2011         struct collision_entry *collision_ent;
2012         int clean = 1;
2013         struct strbuf collision_paths = STRBUF_INIT;
2014
2015         /*
2016          * entry has the mapping of old directory name to new directory name
2017          * that we want to apply to path.
2018          */
2019         new_path = apply_dir_rename(entry, path);
2020
2021         if (!new_path) {
2022                 /* This should only happen when entry->non_unique_new_dir set */
2023                 if (!entry->non_unique_new_dir)
2024                         BUG("entry->non_unqiue_dir not set and !new_path");
2025                 output(opt, 1, _("CONFLICT (directory rename split): "
2026                                "Unclear where to place %s because directory "
2027                                "%s was renamed to multiple other directories, "
2028                                "with no destination getting a majority of the "
2029                                "files."),
2030                        path, entry->dir);
2031                 clean = 0;
2032                 return NULL;
2033         }
2034
2035         /*
2036          * The caller needs to have ensured that it has pre-populated
2037          * collisions with all paths that map to new_path.  Do a quick check
2038          * to ensure that's the case.
2039          */
2040         collision_ent = collision_find_entry(collisions, new_path);
2041         if (collision_ent == NULL)
2042                 BUG("collision_ent is NULL");
2043
2044         /*
2045          * Check for one-sided add/add/.../add conflicts, i.e.
2046          * where implicit renames from the other side doing
2047          * directory rename(s) can affect this side of history
2048          * to put multiple paths into the same location.  Warn
2049          * and bail on directory renames for such paths.
2050          */
2051         if (collision_ent->reported_already) {
2052                 clean = 0;
2053         } else if (tree_has_path(tree, new_path)) {
2054                 collision_ent->reported_already = 1;
2055                 strbuf_add_separated_string_list(&collision_paths, ", ",
2056                                                  &collision_ent->source_files);
2057                 output(opt, 1, _("CONFLICT (implicit dir rename): Existing "
2058                                "file/dir at %s in the way of implicit "
2059                                "directory rename(s) putting the following "
2060                                "path(s) there: %s."),
2061                        new_path, collision_paths.buf);
2062                 clean = 0;
2063         } else if (collision_ent->source_files.nr > 1) {
2064                 collision_ent->reported_already = 1;
2065                 strbuf_add_separated_string_list(&collision_paths, ", ",
2066                                                  &collision_ent->source_files);
2067                 output(opt, 1, _("CONFLICT (implicit dir rename): Cannot map "
2068                                "more than one path to %s; implicit directory "
2069                                "renames tried to put these paths there: %s"),
2070                        new_path, collision_paths.buf);
2071                 clean = 0;
2072         }
2073
2074         /* Free memory we no longer need */
2075         strbuf_release(&collision_paths);
2076         if (!clean && new_path) {
2077                 free(new_path);
2078                 return NULL;
2079         }
2080
2081         return new_path;
2082 }
2083
2084 /*
2085  * There are a couple things we want to do at the directory level:
2086  *   1. Check for both sides renaming to the same thing, in order to avoid
2087  *      implicit renaming of files that should be left in place.  (See
2088  *      testcase 6b in t6043 for details.)
2089  *   2. Prune directory renames if there are still files left in the
2090  *      the original directory.  These represent a partial directory rename,
2091  *      i.e. a rename where only some of the files within the directory
2092  *      were renamed elsewhere.  (Technically, this could be done earlier
2093  *      in get_directory_renames(), except that would prevent us from
2094  *      doing the previous check and thus failing testcase 6b.)
2095  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2096  *      In the future, we could potentially record this info as well and
2097  *      omit reporting rename/rename(1to2) conflicts for each path within
2098  *      the affected directories, thus cleaning up the merge output.
2099  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2100  *         directory level, because merging directories is fine.  If it
2101  *         causes conflicts for files within those merged directories, then
2102  *         that should be detected at the individual path level.
2103  */
2104 static void handle_directory_level_conflicts(struct merge_options *opt,
2105                                              struct hashmap *dir_re_head,
2106                                              struct tree *head,
2107                                              struct hashmap *dir_re_merge,
2108                                              struct tree *merge)
2109 {
2110         struct hashmap_iter iter;
2111         struct dir_rename_entry *head_ent;
2112         struct dir_rename_entry *merge_ent;
2113
2114         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2115         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2116
2117         hashmap_iter_init(dir_re_head, &iter);
2118         while ((head_ent = hashmap_iter_next(&iter))) {
2119                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2120                 if (merge_ent &&
2121                     !head_ent->non_unique_new_dir &&
2122                     !merge_ent->non_unique_new_dir &&
2123                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2124                         /* 1. Renamed identically; remove it from both sides */
2125                         string_list_append(&remove_from_head,
2126                                            head_ent->dir)->util = head_ent;
2127                         strbuf_release(&head_ent->new_dir);
2128                         string_list_append(&remove_from_merge,
2129                                            merge_ent->dir)->util = merge_ent;
2130                         strbuf_release(&merge_ent->new_dir);
2131                 } else if (tree_has_path(head, head_ent->dir)) {
2132                         /* 2. This wasn't a directory rename after all */
2133                         string_list_append(&remove_from_head,
2134                                            head_ent->dir)->util = head_ent;
2135                         strbuf_release(&head_ent->new_dir);
2136                 }
2137         }
2138
2139         remove_hashmap_entries(dir_re_head, &remove_from_head);
2140         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2141
2142         hashmap_iter_init(dir_re_merge, &iter);
2143         while ((merge_ent = hashmap_iter_next(&iter))) {
2144                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2145                 if (tree_has_path(merge, merge_ent->dir)) {
2146                         /* 2. This wasn't a directory rename after all */
2147                         string_list_append(&remove_from_merge,
2148                                            merge_ent->dir)->util = merge_ent;
2149                 } else if (head_ent &&
2150                            !head_ent->non_unique_new_dir &&
2151                            !merge_ent->non_unique_new_dir) {
2152                         /* 3. rename/rename(1to2) */
2153                         /*
2154                          * We can assume it's not rename/rename(1to1) because
2155                          * that was case (1), already checked above.  So we
2156                          * know that head_ent->new_dir and merge_ent->new_dir
2157                          * are different strings.
2158                          */
2159                         output(opt, 1, _("CONFLICT (rename/rename): "
2160                                        "Rename directory %s->%s in %s. "
2161                                        "Rename directory %s->%s in %s"),
2162                                head_ent->dir, head_ent->new_dir.buf, opt->branch1,
2163                                head_ent->dir, merge_ent->new_dir.buf, opt->branch2);
2164                         string_list_append(&remove_from_head,
2165                                            head_ent->dir)->util = head_ent;
2166                         strbuf_release(&head_ent->new_dir);
2167                         string_list_append(&remove_from_merge,
2168                                            merge_ent->dir)->util = merge_ent;
2169                         strbuf_release(&merge_ent->new_dir);
2170                 }
2171         }
2172
2173         remove_hashmap_entries(dir_re_head, &remove_from_head);
2174         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2175 }
2176
2177 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs)
2178 {
2179         struct hashmap *dir_renames;
2180         struct hashmap_iter iter;
2181         struct dir_rename_entry *entry;
2182         int i;
2183
2184         /*
2185          * Typically, we think of a directory rename as all files from a
2186          * certain directory being moved to a target directory.  However,
2187          * what if someone first moved two files from the original
2188          * directory in one commit, and then renamed the directory
2189          * somewhere else in a later commit?  At merge time, we just know
2190          * that files from the original directory went to two different
2191          * places, and that the bulk of them ended up in the same place.
2192          * We want each directory rename to represent where the bulk of the
2193          * files from that directory end up; this function exists to find
2194          * where the bulk of the files went.
2195          *
2196          * The first loop below simply iterates through the list of file
2197          * renames, finding out how often each directory rename pair
2198          * possibility occurs.
2199          */
2200         dir_renames = xmalloc(sizeof(*dir_renames));
2201         dir_rename_init(dir_renames);
2202         for (i = 0; i < pairs->nr; ++i) {
2203                 struct string_list_item *item;
2204                 int *count;
2205                 struct diff_filepair *pair = pairs->queue[i];
2206                 char *old_dir, *new_dir;
2207
2208                 /* File not part of directory rename if it wasn't renamed */
2209                 if (pair->status != 'R')
2210                         continue;
2211
2212                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2213                                         &old_dir,        &new_dir);
2214                 if (!old_dir)
2215                         /* Directory didn't change at all; ignore this one. */
2216                         continue;
2217
2218                 entry = dir_rename_find_entry(dir_renames, old_dir);
2219                 if (!entry) {
2220                         entry = xmalloc(sizeof(*entry));
2221                         dir_rename_entry_init(entry, old_dir);
2222                         hashmap_put(dir_renames, entry);
2223                 } else {
2224                         free(old_dir);
2225                 }
2226                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2227                 if (!item) {
2228                         item = string_list_insert(&entry->possible_new_dirs,
2229                                                   new_dir);
2230                         item->util = xcalloc(1, sizeof(int));
2231                 } else {
2232                         free(new_dir);
2233                 }
2234                 count = item->util;
2235                 *count += 1;
2236         }
2237
2238         /*
2239          * For each directory with files moved out of it, we find out which
2240          * target directory received the most files so we can declare it to
2241          * be the "winning" target location for the directory rename.  This
2242          * winner gets recorded in new_dir.  If there is no winner
2243          * (multiple target directories received the same number of files),
2244          * we set non_unique_new_dir.  Once we've determined the winner (or
2245          * that there is no winner), we no longer need possible_new_dirs.
2246          */
2247         hashmap_iter_init(dir_renames, &iter);
2248         while ((entry = hashmap_iter_next(&iter))) {
2249                 int max = 0;
2250                 int bad_max = 0;
2251                 char *best = NULL;
2252
2253                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2254                         int *count = entry->possible_new_dirs.items[i].util;
2255
2256                         if (*count == max)
2257                                 bad_max = max;
2258                         else if (*count > max) {
2259                                 max = *count;
2260                                 best = entry->possible_new_dirs.items[i].string;
2261                         }
2262                 }
2263                 if (bad_max == max)
2264                         entry->non_unique_new_dir = 1;
2265                 else {
2266                         assert(entry->new_dir.len == 0);
2267                         strbuf_addstr(&entry->new_dir, best);
2268                 }
2269                 /*
2270                  * The relevant directory sub-portion of the original full
2271                  * filepaths were xstrndup'ed before inserting into
2272                  * possible_new_dirs, and instead of manually iterating the
2273                  * list and free'ing each, just lie and tell
2274                  * possible_new_dirs that it did the strdup'ing so that it
2275                  * will free them for us.
2276                  */
2277                 entry->possible_new_dirs.strdup_strings = 1;
2278                 string_list_clear(&entry->possible_new_dirs, 1);
2279         }
2280
2281         return dir_renames;
2282 }
2283
2284 static struct dir_rename_entry *check_dir_renamed(const char *path,
2285                                                   struct hashmap *dir_renames)
2286 {
2287         char *temp = xstrdup(path);
2288         char *end;
2289         struct dir_rename_entry *entry = NULL;
2290
2291         while ((end = strrchr(temp, '/'))) {
2292                 *end = '\0';
2293                 entry = dir_rename_find_entry(dir_renames, temp);
2294                 if (entry)
2295                         break;
2296         }
2297         free(temp);
2298         return entry;
2299 }
2300
2301 static void compute_collisions(struct hashmap *collisions,
2302                                struct hashmap *dir_renames,
2303                                struct diff_queue_struct *pairs)
2304 {
2305         int i;
2306
2307         /*
2308          * Multiple files can be mapped to the same path due to directory
2309          * renames done by the other side of history.  Since that other
2310          * side of history could have merged multiple directories into one,
2311          * if our side of history added the same file basename to each of
2312          * those directories, then all N of them would get implicitly
2313          * renamed by the directory rename detection into the same path,
2314          * and we'd get an add/add/.../add conflict, and all those adds
2315          * from *this* side of history.  This is not representable in the
2316          * index, and users aren't going to easily be able to make sense of
2317          * it.  So we need to provide a good warning about what's
2318          * happening, and fall back to no-directory-rename detection
2319          * behavior for those paths.
2320          *
2321          * See testcases 9e and all of section 5 from t6043 for examples.
2322          */
2323         collision_init(collisions);
2324
2325         for (i = 0; i < pairs->nr; ++i) {
2326                 struct dir_rename_entry *dir_rename_ent;
2327                 struct collision_entry *collision_ent;
2328                 char *new_path;
2329                 struct diff_filepair *pair = pairs->queue[i];
2330
2331                 if (pair->status != 'A' && pair->status != 'R')
2332                         continue;
2333                 dir_rename_ent = check_dir_renamed(pair->two->path,
2334                                                    dir_renames);
2335                 if (!dir_rename_ent)
2336                         continue;
2337
2338                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2339                 if (!new_path)
2340                         /*
2341                          * dir_rename_ent->non_unique_new_path is true, which
2342                          * means there is no directory rename for us to use,
2343                          * which means it won't cause us any additional
2344                          * collisions.
2345                          */
2346                         continue;
2347                 collision_ent = collision_find_entry(collisions, new_path);
2348                 if (!collision_ent) {
2349                         collision_ent = xcalloc(1,
2350                                                 sizeof(struct collision_entry));
2351                         hashmap_entry_init(collision_ent, strhash(new_path));
2352                         hashmap_put(collisions, collision_ent);
2353                         collision_ent->target_file = new_path;
2354                 } else {
2355                         free(new_path);
2356                 }
2357                 string_list_insert(&collision_ent->source_files,
2358                                    pair->two->path);
2359         }
2360 }
2361
2362 static char *check_for_directory_rename(struct merge_options *opt,
2363                                         const char *path,
2364                                         struct tree *tree,
2365                                         struct hashmap *dir_renames,
2366                                         struct hashmap *dir_rename_exclusions,
2367                                         struct hashmap *collisions,
2368                                         int *clean_merge)
2369 {
2370         char *new_path = NULL;
2371         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2372         struct dir_rename_entry *oentry = NULL;
2373
2374         if (!entry)
2375                 return new_path;
2376
2377         /*
2378          * This next part is a little weird.  We do not want to do an
2379          * implicit rename into a directory we renamed on our side, because
2380          * that will result in a spurious rename/rename(1to2) conflict.  An
2381          * example:
2382          *   Base commit: dumbdir/afile, otherdir/bfile
2383          *   Side 1:      smrtdir/afile, otherdir/bfile
2384          *   Side 2:      dumbdir/afile, dumbdir/bfile
2385          * Here, while working on Side 1, we could notice that otherdir was
2386          * renamed/merged to dumbdir, and change the diff_filepair for
2387          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2388          * 2 will notice the rename from dumbdir to smrtdir, and do the
2389          * transitive rename to move it from dumbdir/bfile to
2390          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2391          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2392          * the file to end up in smrtdir.  And the way to achieve that is
2393          * to not let Side1 do the rename to dumbdir, since we know that is
2394          * the source of one of our directory renames.
2395          *
2396          * That's why oentry and dir_rename_exclusions is here.
2397          *
2398          * As it turns out, this also prevents N-way transient rename
2399          * confusion; See testcases 9c and 9d of t6043.
2400          */
2401         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2402         if (oentry) {
2403                 output(opt, 1, _("WARNING: Avoiding applying %s -> %s rename "
2404                                "to %s, because %s itself was renamed."),
2405                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2406         } else {
2407                 new_path = handle_path_level_conflicts(opt, path, entry,
2408                                                        collisions, tree);
2409                 *clean_merge &= (new_path != NULL);
2410         }
2411
2412         return new_path;
2413 }
2414
2415 static void apply_directory_rename_modifications(struct merge_options *opt,
2416                                                  struct diff_filepair *pair,
2417                                                  char *new_path,
2418                                                  struct rename *re,
2419                                                  struct tree *tree,
2420                                                  struct tree *o_tree,
2421                                                  struct tree *a_tree,
2422                                                  struct tree *b_tree,
2423                                                  struct string_list *entries)
2424 {
2425         struct string_list_item *item;
2426         int stage = (tree == a_tree ? 2 : 3);
2427         int update_wd;
2428
2429         /*
2430          * In all cases where we can do directory rename detection,
2431          * unpack_trees() will have read pair->two->path into the
2432          * index and the working copy.  We need to remove it so that
2433          * we can instead place it at new_path.  It is guaranteed to
2434          * not be untracked (unpack_trees() would have errored out
2435          * saying the file would have been overwritten), but it might
2436          * be dirty, though.
2437          */
2438         update_wd = !was_dirty(opt, pair->two->path);
2439         if (!update_wd)
2440                 output(opt, 1, _("Refusing to lose dirty file at %s"),
2441                        pair->two->path);
2442         remove_file(opt, 1, pair->two->path, !update_wd);
2443
2444         /* Find or create a new re->dst_entry */
2445         item = string_list_lookup(entries, new_path);
2446         if (item) {
2447                 /*
2448                  * Since we're renaming on this side of history, and it's
2449                  * due to a directory rename on the other side of history
2450                  * (which we only allow when the directory in question no
2451                  * longer exists on the other side of history), the
2452                  * original entry for re->dst_entry is no longer
2453                  * necessary...
2454                  */
2455                 re->dst_entry->processed = 1;
2456
2457                 /*
2458                  * ...because we'll be using this new one.
2459                  */
2460                 re->dst_entry = item->util;
2461         } else {
2462                 /*
2463                  * re->dst_entry is for the before-dir-rename path, and we
2464                  * need it to hold information for the after-dir-rename
2465                  * path.  Before creating a new entry, we need to mark the
2466                  * old one as unnecessary (...unless it is shared by
2467                  * src_entry, i.e. this didn't use to be a rename, in which
2468                  * case we can just allow the normal processing to happen
2469                  * for it).
2470                  */
2471                 if (pair->status == 'R')
2472                         re->dst_entry->processed = 1;
2473
2474                 re->dst_entry = insert_stage_data(new_path,
2475                                                   o_tree, a_tree, b_tree,
2476                                                   entries);
2477                 item = string_list_insert(entries, new_path);
2478                 item->util = re->dst_entry;
2479         }
2480
2481         /*
2482          * Update the stage_data with the information about the path we are
2483          * moving into place.  That slot will be empty and available for us
2484          * to write to because of the collision checks in
2485          * handle_path_level_conflicts().  In other words,
2486          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2487          * open for us to write to.
2488          *
2489          * It may be tempting to actually update the index at this point as
2490          * well, using update_stages_for_stage_data(), but as per the big
2491          * "NOTE" in update_stages(), doing so will modify the current
2492          * in-memory index which will break calls to would_lose_untracked()
2493          * that we need to make.  Instead, we need to just make sure that
2494          * the various handle_rename_*() functions update the index
2495          * explicitly rather than relying on unpack_trees() to have done it.
2496          */
2497         get_tree_entry(&tree->object.oid,
2498                        pair->two->path,
2499                        &re->dst_entry->stages[stage].oid,
2500                        &re->dst_entry->stages[stage].mode);
2501
2502         /*
2503          * Record the original change status (or 'type' of change).  If it
2504          * was originally an add ('A'), this lets us differentiate later
2505          * between a RENAME_DELETE conflict and RENAME_VIA_DIR (they
2506          * otherwise look the same).  If it was originally a rename ('R'),
2507          * this lets us remember and report accurately about the transitive
2508          * renaming that occurred via the directory rename detection.  Also,
2509          * record the original destination name.
2510          */
2511         re->dir_rename_original_type = pair->status;
2512         re->dir_rename_original_dest = pair->two->path;
2513
2514         /*
2515          * We don't actually look at pair->status again, but it seems
2516          * pedagogically correct to adjust it.
2517          */
2518         pair->status = 'R';
2519
2520         /*
2521          * Finally, record the new location.
2522          */
2523         pair->two->path = new_path;
2524 }
2525
2526 /*
2527  * Get information of all renames which occurred in 'pairs', making use of
2528  * any implicit directory renames inferred from the other side of history.
2529  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2530  * to be able to associate the correct cache entries with the rename
2531  * information; tree is always equal to either a_tree or b_tree.
2532  */
2533 static struct string_list *get_renames(struct merge_options *opt,
2534                                        const char *branch,
2535                                        struct diff_queue_struct *pairs,
2536                                        struct hashmap *dir_renames,
2537                                        struct hashmap *dir_rename_exclusions,
2538                                        struct tree *tree,
2539                                        struct tree *o_tree,
2540                                        struct tree *a_tree,
2541                                        struct tree *b_tree,
2542                                        struct string_list *entries,
2543                                        int *clean_merge)
2544 {
2545         int i;
2546         struct hashmap collisions;
2547         struct hashmap_iter iter;
2548         struct collision_entry *e;
2549         struct string_list *renames;
2550
2551         compute_collisions(&collisions, dir_renames, pairs);
2552         renames = xcalloc(1, sizeof(struct string_list));
2553
2554         for (i = 0; i < pairs->nr; ++i) {
2555                 struct string_list_item *item;
2556                 struct rename *re;
2557                 struct diff_filepair *pair = pairs->queue[i];
2558                 char *new_path; /* non-NULL only with directory renames */
2559
2560                 if (pair->status != 'A' && pair->status != 'R') {
2561                         diff_free_filepair(pair);
2562                         continue;
2563                 }
2564                 new_path = check_for_directory_rename(opt, pair->two->path, tree,
2565                                                       dir_renames,
2566                                                       dir_rename_exclusions,
2567                                                       &collisions,
2568                                                       clean_merge);
2569                 if (pair->status != 'R' && !new_path) {
2570                         diff_free_filepair(pair);
2571                         continue;
2572                 }
2573
2574                 re = xmalloc(sizeof(*re));
2575                 re->processed = 0;
2576                 re->pair = pair;
2577                 re->branch = branch;
2578                 re->dir_rename_original_type = '\0';
2579                 re->dir_rename_original_dest = NULL;
2580                 item = string_list_lookup(entries, re->pair->one->path);
2581                 if (!item)
2582                         re->src_entry = insert_stage_data(re->pair->one->path,
2583                                         o_tree, a_tree, b_tree, entries);
2584                 else
2585                         re->src_entry = item->util;
2586
2587                 item = string_list_lookup(entries, re->pair->two->path);
2588                 if (!item)
2589                         re->dst_entry = insert_stage_data(re->pair->two->path,
2590                                         o_tree, a_tree, b_tree, entries);
2591                 else
2592                         re->dst_entry = item->util;
2593                 item = string_list_insert(renames, pair->one->path);
2594                 item->util = re;
2595                 if (new_path)
2596                         apply_directory_rename_modifications(opt, pair, new_path,
2597                                                              re, tree, o_tree,
2598                                                              a_tree, b_tree,
2599                                                              entries);
2600         }
2601
2602         hashmap_iter_init(&collisions, &iter);
2603         while ((e = hashmap_iter_next(&iter))) {
2604                 free(e->target_file);
2605                 string_list_clear(&e->source_files, 0);
2606         }
2607         hashmap_free(&collisions, 1);
2608         return renames;
2609 }
2610
2611 static int process_renames(struct merge_options *opt,
2612                            struct string_list *a_renames,
2613                            struct string_list *b_renames)
2614 {
2615         int clean_merge = 1, i, j;
2616         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2617         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2618         const struct rename *sre;
2619
2620         for (i = 0; i < a_renames->nr; i++) {
2621                 sre = a_renames->items[i].util;
2622                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2623                         = (void *)sre;
2624         }
2625         for (i = 0; i < b_renames->nr; i++) {
2626                 sre = b_renames->items[i].util;
2627                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2628                         = (void *)sre;
2629         }
2630
2631         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2632                 struct string_list *renames1, *renames2Dst;
2633                 struct rename *ren1 = NULL, *ren2 = NULL;
2634                 const char *ren1_src, *ren1_dst;
2635                 struct string_list_item *lookup;
2636
2637                 if (i >= a_renames->nr) {
2638                         ren2 = b_renames->items[j++].util;
2639                 } else if (j >= b_renames->nr) {
2640                         ren1 = a_renames->items[i++].util;
2641                 } else {
2642                         int compare = strcmp(a_renames->items[i].string,
2643                                              b_renames->items[j].string);
2644                         if (compare <= 0)
2645                                 ren1 = a_renames->items[i++].util;
2646                         if (compare >= 0)
2647                                 ren2 = b_renames->items[j++].util;
2648                 }
2649
2650                 /* TODO: refactor, so that 1/2 are not needed */
2651                 if (ren1) {
2652                         renames1 = a_renames;
2653                         renames2Dst = &b_by_dst;
2654                 } else {
2655                         renames1 = b_renames;
2656                         renames2Dst = &a_by_dst;
2657                         SWAP(ren2, ren1);
2658                 }
2659
2660                 if (ren1->processed)
2661                         continue;
2662                 ren1->processed = 1;
2663                 ren1->dst_entry->processed = 1;
2664                 /* BUG: We should only mark src_entry as processed if we
2665                  * are not dealing with a rename + add-source case.
2666                  */
2667                 ren1->src_entry->processed = 1;
2668
2669                 ren1_src = ren1->pair->one->path;
2670                 ren1_dst = ren1->pair->two->path;
2671
2672                 if (ren2) {
2673                         /* One file renamed on both sides */
2674                         const char *ren2_src = ren2->pair->one->path;
2675                         const char *ren2_dst = ren2->pair->two->path;
2676                         enum rename_type rename_type;
2677                         if (strcmp(ren1_src, ren2_src) != 0)
2678                                 BUG("ren1_src != ren2_src");
2679                         ren2->dst_entry->processed = 1;
2680                         ren2->processed = 1;
2681                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2682                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2683                                 clean_merge = 0;
2684                         } else {
2685                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2686                                 /* BUG: We should only remove ren1_src in
2687                                  * the base stage (think of rename +
2688                                  * add-source cases).
2689                                  */
2690                                 remove_file(opt, 1, ren1_src, 1);
2691                                 update_entry(ren1->dst_entry,
2692                                              ren1->pair->one,
2693                                              ren1->pair->two,
2694                                              ren2->pair->two);
2695                         }
2696                         setup_rename_conflict_info(rename_type, opt, ren1, ren2);
2697                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2698                         /* Two different files renamed to the same thing */
2699                         char *ren2_dst;
2700                         ren2 = lookup->util;
2701                         ren2_dst = ren2->pair->two->path;
2702                         if (strcmp(ren1_dst, ren2_dst) != 0)
2703                                 BUG("ren1_dst != ren2_dst");
2704
2705                         clean_merge = 0;
2706                         ren2->processed = 1;
2707                         /*
2708                          * BUG: We should only mark src_entry as processed
2709                          * if we are not dealing with a rename + add-source
2710                          * case.
2711                          */
2712                         ren2->src_entry->processed = 1;
2713
2714                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2715                                                    opt, ren1, ren2);
2716                 } else {
2717                         /* Renamed in 1, maybe changed in 2 */
2718                         /* we only use sha1 and mode of these */
2719                         struct diff_filespec src_other, dst_other;
2720                         int try_merge;
2721
2722                         /*
2723                          * unpack_trees loads entries from common-commit
2724                          * into stage 1, from head-commit into stage 2, and
2725                          * from merge-commit into stage 3.  We keep track
2726                          * of which side corresponds to the rename.
2727                          */
2728                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2729                         int other_stage =   a_renames == renames1 ? 3 : 2;
2730
2731                         /* BUG: We should only remove ren1_src in the base
2732                          * stage and in other_stage (think of rename +
2733                          * add-source case).
2734                          */
2735                         remove_file(opt, 1, ren1_src,
2736                                     renamed_stage == 2 || !was_tracked(opt, ren1_src));
2737
2738                         oidcpy(&src_other.oid,
2739                                &ren1->src_entry->stages[other_stage].oid);
2740                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2741                         oidcpy(&dst_other.oid,
2742                                &ren1->dst_entry->stages[other_stage].oid);
2743                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2744                         try_merge = 0;
2745
2746                         if (oid_eq(&src_other.oid, &null_oid) &&
2747                             ren1->dir_rename_original_type == 'A') {
2748                                 setup_rename_conflict_info(RENAME_VIA_DIR,
2749                                                            opt, ren1, NULL);
2750                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2751                                 setup_rename_conflict_info(RENAME_DELETE,
2752                                                            opt, ren1, NULL);
2753                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2754                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2755                                 /*
2756                                  * Added file on the other side identical to
2757                                  * the file being renamed: clean merge.
2758                                  * Also, there is no need to overwrite the
2759                                  * file already in the working copy, so call
2760                                  * update_file_flags() instead of
2761                                  * update_file().
2762                                  */
2763                                 if (update_file_flags(opt,
2764                                                       ren1->pair->two,
2765                                                       ren1_dst,
2766                                                       1, /* update_cache */
2767                                                       0  /* update_wd    */))
2768                                         clean_merge = -1;
2769                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2770                                 /*
2771                                  * Probably not a clean merge, but it's
2772                                  * premature to set clean_merge to 0 here,
2773                                  * because if the rename merges cleanly and
2774                                  * the merge exactly matches the newly added
2775                                  * file, then the merge will be clean.
2776                                  */
2777                                 setup_rename_conflict_info(RENAME_ADD,
2778                                                            opt, ren1, NULL);
2779                         } else
2780                                 try_merge = 1;
2781
2782                         if (clean_merge < 0)
2783                                 goto cleanup_and_return;
2784                         if (try_merge) {
2785                                 struct diff_filespec *o, *a, *b;
2786                                 src_other.path = (char *)ren1_src;
2787
2788                                 o = ren1->pair->one;
2789                                 if (a_renames == renames1) {
2790                                         a = ren1->pair->two;
2791                                         b = &src_other;
2792                                 } else {
2793                                         b = ren1->pair->two;
2794                                         a = &src_other;
2795                                 }
2796                                 update_entry(ren1->dst_entry, o, a, b);
2797                                 setup_rename_conflict_info(RENAME_NORMAL,
2798                                                            opt, ren1, NULL);
2799                         }
2800                 }
2801         }
2802 cleanup_and_return:
2803         string_list_clear(&a_by_dst, 0);
2804         string_list_clear(&b_by_dst, 0);
2805
2806         return clean_merge;
2807 }
2808
2809 struct rename_info {
2810         struct string_list *head_renames;
2811         struct string_list *merge_renames;
2812 };
2813
2814 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2815                                    struct hashmap *dir_renames)
2816 {
2817         struct hashmap_iter iter;
2818         struct dir_rename_entry *e;
2819
2820         hashmap_iter_init(dir_renames, &iter);
2821         while ((e = hashmap_iter_next(&iter))) {
2822                 free(e->dir);
2823                 strbuf_release(&e->new_dir);
2824                 /* possible_new_dirs already cleared in get_directory_renames */
2825         }
2826         hashmap_free(dir_renames, 1);
2827         free(dir_renames);
2828
2829         free(pairs->queue);
2830         free(pairs);
2831 }
2832
2833 static int detect_and_process_renames(struct merge_options *opt,
2834                                       struct tree *common,
2835                                       struct tree *head,
2836                                       struct tree *merge,
2837                                       struct string_list *entries,
2838                                       struct rename_info *ri)
2839 {
2840         struct diff_queue_struct *head_pairs, *merge_pairs;
2841         struct hashmap *dir_re_head, *dir_re_merge;
2842         int clean = 1;
2843
2844         ri->head_renames = NULL;
2845         ri->merge_renames = NULL;
2846
2847         if (!merge_detect_rename(opt))
2848                 return 1;
2849
2850         head_pairs = get_diffpairs(opt, common, head);
2851         merge_pairs = get_diffpairs(opt, common, merge);
2852
2853         if (opt->detect_directory_renames) {
2854                 dir_re_head = get_directory_renames(head_pairs);
2855                 dir_re_merge = get_directory_renames(merge_pairs);
2856
2857                 handle_directory_level_conflicts(opt,
2858                                                  dir_re_head, head,
2859                                                  dir_re_merge, merge);
2860         } else {
2861                 dir_re_head  = xmalloc(sizeof(*dir_re_head));
2862                 dir_re_merge = xmalloc(sizeof(*dir_re_merge));
2863                 dir_rename_init(dir_re_head);
2864                 dir_rename_init(dir_re_merge);
2865         }
2866
2867         ri->head_renames  = get_renames(opt, opt->branch1, head_pairs,
2868                                         dir_re_merge, dir_re_head, head,
2869                                         common, head, merge, entries,
2870                                         &clean);
2871         if (clean < 0)
2872                 goto cleanup;
2873         ri->merge_renames = get_renames(opt, opt->branch2, merge_pairs,
2874                                         dir_re_head, dir_re_merge, merge,
2875                                         common, head, merge, entries,
2876                                         &clean);
2877         if (clean < 0)
2878                 goto cleanup;
2879         clean &= process_renames(opt, ri->head_renames, ri->merge_renames);
2880
2881 cleanup:
2882         /*
2883          * Some cleanup is deferred until cleanup_renames() because the
2884          * data structures are still needed and referenced in
2885          * process_entry().  But there are a few things we can free now.
2886          */
2887         initial_cleanup_rename(head_pairs, dir_re_head);
2888         initial_cleanup_rename(merge_pairs, dir_re_merge);
2889
2890         return clean;
2891 }
2892
2893 static void final_cleanup_rename(struct string_list *rename)
2894 {
2895         const struct rename *re;
2896         int i;
2897
2898         if (rename == NULL)
2899                 return;
2900
2901         for (i = 0; i < rename->nr; i++) {
2902                 re = rename->items[i].util;
2903                 diff_free_filepair(re->pair);
2904         }
2905         string_list_clear(rename, 1);
2906         free(rename);
2907 }
2908
2909 static void final_cleanup_renames(struct rename_info *re_info)
2910 {
2911         final_cleanup_rename(re_info->head_renames);
2912         final_cleanup_rename(re_info->merge_renames);
2913 }
2914
2915 static int read_oid_strbuf(struct merge_options *opt,
2916                            const struct object_id *oid,
2917                            struct strbuf *dst)
2918 {
2919         void *buf;
2920         enum object_type type;
2921         unsigned long size;
2922         buf = read_object_file(oid, &type, &size);
2923         if (!buf)
2924                 return err(opt, _("cannot read object %s"), oid_to_hex(oid));
2925         if (type != OBJ_BLOB) {
2926                 free(buf);
2927                 return err(opt, _("object %s is not a blob"), oid_to_hex(oid));
2928         }
2929         strbuf_attach(dst, buf, size, size + 1);
2930         return 0;
2931 }
2932
2933 static int blob_unchanged(struct merge_options *opt,
2934                           const struct diff_filespec *o,
2935                           const struct diff_filespec *a,
2936                           int renormalize, const char *path)
2937 {
2938         struct strbuf obuf = STRBUF_INIT;
2939         struct strbuf abuf = STRBUF_INIT;
2940         int ret = 0; /* assume changed for safety */
2941         const struct index_state *idx = opt->repo->index;
2942
2943         if (a->mode != o->mode)
2944                 return 0;
2945         if (oid_eq(&o->oid, &a->oid))
2946                 return 1;
2947         if (!renormalize)
2948                 return 0;
2949
2950         if (read_oid_strbuf(opt, &o->oid, &obuf) ||
2951             read_oid_strbuf(opt, &a->oid, &abuf))
2952                 goto error_return;
2953         /*
2954          * Note: binary | is used so that both renormalizations are
2955          * performed.  Comparison can be skipped if both files are
2956          * unchanged since their sha1s have already been compared.
2957          */
2958         if (renormalize_buffer(idx, path, obuf.buf, obuf.len, &obuf) |
2959             renormalize_buffer(idx, path, abuf.buf, abuf.len, &abuf))
2960                 ret = (obuf.len == abuf.len && !memcmp(obuf.buf, abuf.buf, obuf.len));
2961
2962 error_return:
2963         strbuf_release(&obuf);
2964         strbuf_release(&abuf);
2965         return ret;
2966 }
2967
2968 static int handle_modify_delete(struct merge_options *opt,
2969                                 const char *path,
2970                                 const struct diff_filespec *o,
2971                                 const struct diff_filespec *a,
2972                                 const struct diff_filespec *b)
2973 {
2974         const char *modify_branch, *delete_branch;
2975         const struct diff_filespec *changed;
2976
2977         if (is_valid(a)) {
2978                 modify_branch = opt->branch1;
2979                 delete_branch = opt->branch2;
2980                 changed = a;
2981         } else {
2982                 modify_branch = opt->branch2;
2983                 delete_branch = opt->branch1;
2984                 changed = b;
2985         }
2986
2987         return handle_change_delete(opt,
2988                                     path, NULL,
2989                                     o, changed,
2990                                     modify_branch, delete_branch,
2991                                     _("modify"), _("modified"));
2992 }
2993
2994 static int handle_content_merge(struct merge_file_info *mfi,
2995                                 struct merge_options *opt,
2996                                 const char *path,
2997                                 int is_dirty,
2998                                 const struct diff_filespec *o,
2999                                 const struct diff_filespec *a,
3000                                 const struct diff_filespec *b,
3001                                 struct rename_conflict_info *ci)
3002 {
3003         const char *reason = _("content");
3004         unsigned df_conflict_remains = 0;
3005
3006         if (!is_valid(o))
3007                 reason = _("add/add");
3008
3009         assert(o->path && a->path && b->path);
3010         if (ci && dir_in_way(opt->repo->index, path, !opt->call_depth,
3011                              S_ISGITLINK(ci->ren1->pair->two->mode)))
3012                 df_conflict_remains = 1;
3013
3014         if (merge_mode_and_contents(opt, o, a, b, path,
3015                                     opt->branch1, opt->branch2,
3016                                     opt->call_depth * 2, mfi))
3017                 return -1;
3018
3019         /*
3020          * We can skip updating the working tree file iff:
3021          *   a) The merge is clean
3022          *   b) The merge matches what was in HEAD (content, mode, pathname)
3023          *   c) The target path is usable (i.e. not involved in D/F conflict)
3024          */
3025         if (mfi->clean && was_tracked_and_matches(opt, path, &mfi->blob) &&
3026             !df_conflict_remains) {
3027                 int pos;
3028                 struct cache_entry *ce;
3029
3030                 output(opt, 3, _("Skipped %s (merged same as existing)"), path);
3031                 if (add_cacheinfo(opt, &mfi->blob, path,
3032                                   0, (!opt->call_depth && !is_dirty), 0))
3033                         return -1;
3034                 /*
3035                  * However, add_cacheinfo() will delete the old cache entry
3036                  * and add a new one.  We need to copy over any skip_worktree
3037                  * flag to avoid making the file appear as if it were
3038                  * deleted by the user.
3039                  */
3040                 pos = index_name_pos(&opt->orig_index, path, strlen(path));
3041                 ce = opt->orig_index.cache[pos];
3042                 if (ce_skip_worktree(ce)) {
3043                         pos = index_name_pos(opt->repo->index, path, strlen(path));
3044                         ce = opt->repo->index->cache[pos];
3045                         ce->ce_flags |= CE_SKIP_WORKTREE;
3046                 }
3047                 return mfi->clean;
3048         }
3049
3050         if (!mfi->clean) {
3051                 if (S_ISGITLINK(mfi->blob.mode))
3052                         reason = _("submodule");
3053                 output(opt, 1, _("CONFLICT (%s): Merge conflict in %s"),
3054                                 reason, path);
3055                 if (ci && !df_conflict_remains)
3056                         if (update_stages(opt, path, o, a, b))
3057                                 return -1;
3058         }
3059
3060         if (df_conflict_remains || is_dirty) {
3061                 char *new_path;
3062                 if (opt->call_depth) {
3063                         remove_file_from_index(opt->repo->index, path);
3064                 } else {
3065                         if (!mfi->clean) {
3066                                 if (update_stages(opt, path, o, a, b))
3067                                         return -1;
3068                         } else {
3069                                 int file_from_stage2 = was_tracked(opt, path);
3070
3071                                 if (update_stages(opt, path, NULL,
3072                                                   file_from_stage2 ? &mfi->blob : NULL,
3073                                                   file_from_stage2 ? NULL : &mfi->blob))
3074                                         return -1;
3075                         }
3076
3077                 }
3078                 new_path = unique_path(opt, path, ci->ren1->branch);
3079                 if (is_dirty) {
3080                         output(opt, 1, _("Refusing to lose dirty file at %s"),
3081                                path);
3082                 }
3083                 output(opt, 1, _("Adding as %s instead"), new_path);
3084                 if (update_file(opt, 0, &mfi->blob, new_path)) {
3085                         free(new_path);
3086                         return -1;
3087                 }
3088                 free(new_path);
3089                 mfi->clean = 0;
3090         } else if (update_file(opt, mfi->clean, &mfi->blob, path))
3091                 return -1;
3092         return !is_dirty && mfi->clean;
3093 }
3094
3095 static int handle_rename_normal(struct merge_options *opt,
3096                                 const char *path,
3097                                 const struct diff_filespec *o,
3098                                 const struct diff_filespec *a,
3099                                 const struct diff_filespec *b,
3100                                 struct rename_conflict_info *ci)
3101 {
3102         struct rename *ren = ci->ren1;
3103         struct merge_file_info mfi;
3104         int clean;
3105         int side = (ren->branch == opt->branch1 ? 2 : 3);
3106
3107         /* Merge the content and write it out */
3108         clean = handle_content_merge(&mfi, opt, path, was_dirty(opt, path),
3109                                      o, a, b, ci);
3110
3111         if (clean && opt->detect_directory_renames == 1 &&
3112             ren->dir_rename_original_dest) {
3113                 if (update_stages(opt, path,
3114                                   NULL,
3115                                   side == 2 ? &mfi.blob : NULL,
3116                                   side == 2 ? NULL : &mfi.blob))
3117                         return -1;
3118                 clean = 0; /* not clean, but conflicted */
3119         }
3120         return clean;
3121 }
3122
3123 static void dir_rename_warning(const char *msg,
3124                                int is_add,
3125                                int clean,
3126                                struct merge_options *opt,
3127                                struct rename *ren)
3128 {
3129         const char *other_branch;
3130         other_branch = (ren->branch == opt->branch1 ?
3131                         opt->branch2 : opt->branch1);
3132         if (is_add) {
3133                 output(opt, clean ? 2 : 1, msg,
3134                        ren->pair->one->path, ren->branch,
3135                        other_branch, ren->pair->two->path);
3136                 return;
3137         }
3138         output(opt, clean ? 2 : 1, msg,
3139                ren->pair->one->path, ren->dir_rename_original_dest, ren->branch,
3140                other_branch, ren->pair->two->path);
3141 }
3142 static int warn_about_dir_renamed_entries(struct merge_options *opt,
3143                                           struct rename *ren)
3144 {
3145         const char *msg;
3146         int clean = 1, is_add;
3147
3148         if (!ren)
3149                 return clean;
3150
3151         /* Return early if ren was not affected/created by a directory rename */
3152         if (!ren->dir_rename_original_dest)
3153                 return clean;
3154
3155         /* Sanity checks */
3156         assert(opt->detect_directory_renames > 0);
3157         assert(ren->dir_rename_original_type == 'A' ||
3158                ren->dir_rename_original_type == 'R');
3159
3160         /* Check whether to treat directory renames as a conflict */
3161         clean = (opt->detect_directory_renames == 2);
3162
3163         is_add = (ren->dir_rename_original_type == 'A');
3164         if (ren->dir_rename_original_type == 'A' && clean) {
3165                 msg = _("Path updated: %s added in %s inside a "
3166                         "directory that was renamed in %s; moving it to %s.");
3167         } else if (ren->dir_rename_original_type == 'A' && !clean) {
3168                 msg = _("CONFLICT (file location): %s added in %s "
3169                         "inside a directory that was renamed in %s, "
3170                         "suggesting it should perhaps be moved to %s.");
3171         } else if (ren->dir_rename_original_type == 'R' && clean) {
3172                 msg = _("Path updated: %s renamed to %s in %s, inside a "
3173                         "directory that was renamed in %s; moving it to %s.");
3174         } else if (ren->dir_rename_original_type == 'R' && !clean) {
3175                 msg = _("CONFLICT (file location): %s renamed to %s in %s, "
3176                         "inside a directory that was renamed in %s, "
3177                         "suggesting it should perhaps be moved to %s.");
3178         } else {
3179                 BUG("Impossible dir_rename_original_type/clean combination");
3180         }
3181         dir_rename_warning(msg, is_add, clean, opt, ren);
3182
3183         return clean;
3184 }
3185
3186 /* Per entry merge function */
3187 static int process_entry(struct merge_options *opt,
3188                          const char *path, struct stage_data *entry)
3189 {
3190         int clean_merge = 1;
3191         int normalize = opt->renormalize;
3192
3193         struct diff_filespec *o = &entry->stages[1];
3194         struct diff_filespec *a = &entry->stages[2];
3195         struct diff_filespec *b = &entry->stages[3];
3196         int o_valid = is_valid(o);
3197         int a_valid = is_valid(a);
3198         int b_valid = is_valid(b);
3199         o->path = a->path = b->path = (char*)path;
3200
3201         entry->processed = 1;
3202         if (entry->rename_conflict_info) {
3203                 struct rename_conflict_info *ci = entry->rename_conflict_info;
3204                 struct diff_filespec *temp;
3205                 int path_clean;
3206
3207                 path_clean = warn_about_dir_renamed_entries(opt, ci->ren1);
3208                 path_clean &= warn_about_dir_renamed_entries(opt, ci->ren2);
3209
3210                 /*
3211                  * For cases with a single rename, {o,a,b}->path have all been
3212                  * set to the rename target path; we need to set two of these
3213                  * back to the rename source.
3214                  * For rename/rename conflicts, we'll manually fix paths below.
3215                  */
3216                 temp = (opt->branch1 == ci->ren1->branch) ? b : a;
3217                 o->path = temp->path = ci->ren1->pair->one->path;
3218                 if (ci->ren2) {
3219                         assert(opt->branch1 == ci->ren1->branch);
3220                 }
3221
3222                 switch (ci->rename_type) {
3223                 case RENAME_NORMAL:
3224                 case RENAME_ONE_FILE_TO_ONE:
3225                         clean_merge = handle_rename_normal(opt, path, o, a, b,
3226                                                            ci);
3227                         break;
3228                 case RENAME_VIA_DIR:
3229                         clean_merge = handle_rename_via_dir(opt, ci);
3230                         break;
3231                 case RENAME_ADD:
3232                         /*
3233                          * Probably unclean merge, but if the renamed file
3234                          * merges cleanly and the result can then be
3235                          * two-way merged cleanly with the added file, I
3236                          * guess it's a clean merge?
3237                          */
3238                         clean_merge = handle_rename_add(opt, ci);
3239                         break;
3240                 case RENAME_DELETE:
3241                         clean_merge = 0;
3242                         if (handle_rename_delete(opt, ci))
3243                                 clean_merge = -1;
3244                         break;
3245                 case RENAME_ONE_FILE_TO_TWO:
3246                         /*
3247                          * Manually fix up paths; note:
3248                          * ren[12]->pair->one->path are equal.
3249                          */
3250                         o->path = ci->ren1->pair->one->path;
3251                         a->path = ci->ren1->pair->two->path;
3252                         b->path = ci->ren2->pair->two->path;
3253
3254                         clean_merge = 0;
3255                         if (handle_rename_rename_1to2(opt, ci))
3256                                 clean_merge = -1;
3257                         break;
3258                 case RENAME_TWO_FILES_TO_ONE:
3259                         /*
3260                          * Manually fix up paths; note,
3261                          * ren[12]->pair->two->path are actually equal.
3262                          */
3263                         o->path = NULL;
3264                         a->path = ci->ren1->pair->two->path;
3265                         b->path = ci->ren2->pair->two->path;
3266
3267                         /*
3268                          * Probably unclean merge, but if the two renamed
3269                          * files merge cleanly and the two resulting files
3270                          * can then be two-way merged cleanly, I guess it's
3271                          * a clean merge?
3272                          */
3273                         clean_merge = handle_rename_rename_2to1(opt, ci);
3274                         break;
3275                 default:
3276                         entry->processed = 0;
3277                         break;
3278                 }
3279                 if (path_clean < clean_merge)
3280                         clean_merge = path_clean;
3281         } else if (o_valid && (!a_valid || !b_valid)) {
3282                 /* Case A: Deleted in one */
3283                 if ((!a_valid && !b_valid) ||
3284                     (!b_valid && blob_unchanged(opt, o, a, normalize, path)) ||
3285                     (!a_valid && blob_unchanged(opt, o, b, normalize, path))) {
3286                         /* Deleted in both or deleted in one and
3287                          * unchanged in the other */
3288                         if (a_valid)
3289                                 output(opt, 2, _("Removing %s"), path);
3290                         /* do not touch working file if it did not exist */
3291                         remove_file(opt, 1, path, !a_valid);
3292                 } else {
3293                         /* Modify/delete; deleted side may have put a directory in the way */
3294                         clean_merge = 0;
3295                         if (handle_modify_delete(opt, path, o, a, b))
3296                                 clean_merge = -1;
3297                 }
3298         } else if ((!o_valid && a_valid && !b_valid) ||
3299                    (!o_valid && !a_valid && b_valid)) {
3300                 /* Case B: Added in one. */
3301                 /* [nothing|directory] -> ([nothing|directory], file) */
3302
3303                 const char *add_branch;
3304                 const char *other_branch;
3305                 const char *conf;
3306                 const struct diff_filespec *contents;
3307
3308                 if (a_valid) {
3309                         add_branch = opt->branch1;
3310                         other_branch = opt->branch2;
3311                         contents = a;
3312                         conf = _("file/directory");
3313                 } else {
3314                         add_branch = opt->branch2;
3315                         other_branch = opt->branch1;
3316                         contents = b;
3317                         conf = _("directory/file");
3318                 }
3319                 if (dir_in_way(opt->repo->index, path,
3320                                !opt->call_depth && !S_ISGITLINK(a->mode),
3321                                0)) {
3322                         char *new_path = unique_path(opt, path, add_branch);
3323                         clean_merge = 0;
3324                         output(opt, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3325                                "Adding %s as %s"),
3326                                conf, path, other_branch, path, new_path);
3327                         if (update_file(opt, 0, contents, new_path))
3328                                 clean_merge = -1;
3329                         else if (opt->call_depth)
3330                                 remove_file_from_index(opt->repo->index, path);
3331                         free(new_path);
3332                 } else {
3333                         output(opt, 2, _("Adding %s"), path);
3334                         /* do not overwrite file if already present */
3335                         if (update_file_flags(opt, contents, path, 1, !a_valid))
3336                                 clean_merge = -1;
3337                 }
3338         } else if (a_valid && b_valid) {
3339                 if (!o_valid) {
3340                         /* Case C: Added in both (check for same permissions) */
3341                         output(opt, 1,
3342                                _("CONFLICT (add/add): Merge conflict in %s"),
3343                                path);
3344                         clean_merge = handle_file_collision(opt,
3345                                                             path, NULL, NULL,
3346                                                             opt->branch1,
3347                                                             opt->branch2,
3348                                                             a, b);
3349                 } else {
3350                         /* case D: Modified in both, but differently. */
3351                         struct merge_file_info mfi;
3352                         int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3353                         clean_merge = handle_content_merge(&mfi, opt, path,
3354                                                            is_dirty,
3355                                                            o, a, b, NULL);
3356                 }
3357         } else if (!o_valid && !a_valid && !b_valid) {
3358                 /*
3359                  * this entry was deleted altogether. a_mode == 0 means
3360                  * we had that path and want to actively remove it.
3361                  */
3362                 remove_file(opt, 1, path, !a->mode);
3363         } else
3364                 BUG("fatal merge failure, shouldn't happen.");
3365
3366         return clean_merge;
3367 }
3368
3369 int merge_trees(struct merge_options *opt,
3370                 struct tree *head,
3371                 struct tree *merge,
3372                 struct tree *common,
3373                 struct tree **result)
3374 {
3375         struct index_state *istate = opt->repo->index;
3376         int code, clean;
3377         struct strbuf sb = STRBUF_INIT;
3378
3379         if (!opt->call_depth && repo_index_has_changes(opt->repo, head, &sb)) {
3380                 err(opt, _("Your local changes to the following files would be overwritten by merge:\n  %s"),
3381                     sb.buf);
3382                 return -1;
3383         }
3384
3385         if (opt->subtree_shift) {
3386                 merge = shift_tree_object(opt->repo, head, merge, opt->subtree_shift);
3387                 common = shift_tree_object(opt->repo, head, common, opt->subtree_shift);
3388         }
3389
3390         if (oid_eq(&common->object.oid, &merge->object.oid)) {
3391                 output(opt, 0, _("Already up to date!"));
3392                 *result = head;
3393                 return 1;
3394         }
3395
3396         code = unpack_trees_start(opt, common, head, merge);
3397
3398         if (code != 0) {
3399                 if (show(opt, 4) || opt->call_depth)
3400                         err(opt, _("merging of trees %s and %s failed"),
3401                             oid_to_hex(&head->object.oid),
3402                             oid_to_hex(&merge->object.oid));
3403                 unpack_trees_finish(opt);
3404                 return -1;
3405         }
3406
3407         if (unmerged_index(istate)) {
3408                 struct string_list *entries;
3409                 struct rename_info re_info;
3410                 int i;
3411                 /*
3412                  * Only need the hashmap while processing entries, so
3413                  * initialize it here and free it when we are done running
3414                  * through the entries. Keeping it in the merge_options as
3415                  * opposed to decaring a local hashmap is for convenience
3416                  * so that we don't have to pass it to around.
3417                  */
3418                 hashmap_init(&opt->current_file_dir_set, path_hashmap_cmp, NULL, 512);
3419                 get_files_dirs(opt, head);
3420                 get_files_dirs(opt, merge);
3421
3422                 entries = get_unmerged(opt->repo->index);
3423                 clean = detect_and_process_renames(opt, common, head, merge,
3424                                                    entries, &re_info);
3425                 record_df_conflict_files(opt, entries);
3426                 if (clean < 0)
3427                         goto cleanup;
3428                 for (i = entries->nr-1; 0 <= i; i--) {
3429                         const char *path = entries->items[i].string;
3430                         struct stage_data *e = entries->items[i].util;
3431                         if (!e->processed) {
3432                                 int ret = process_entry(opt, path, e);
3433                                 if (!ret)
3434                                         clean = 0;
3435                                 else if (ret < 0) {
3436                                         clean = ret;
3437                                         goto cleanup;
3438                                 }
3439                         }
3440                 }
3441                 for (i = 0; i < entries->nr; i++) {
3442                         struct stage_data *e = entries->items[i].util;
3443                         if (!e->processed)
3444                                 BUG("unprocessed path??? %s",
3445                                     entries->items[i].string);
3446                 }
3447
3448         cleanup:
3449                 final_cleanup_renames(&re_info);
3450
3451                 string_list_clear(entries, 1);
3452                 free(entries);
3453
3454                 hashmap_free(&opt->current_file_dir_set, 1);
3455
3456                 if (clean < 0) {
3457                         unpack_trees_finish(opt);
3458                         return clean;
3459                 }
3460         }
3461         else
3462                 clean = 1;
3463
3464         unpack_trees_finish(opt);
3465
3466         if (opt->call_depth && !(*result = write_tree_from_memory(opt)))
3467                 return -1;
3468
3469         return clean;
3470 }
3471
3472 static struct commit_list *reverse_commit_list(struct commit_list *list)
3473 {
3474         struct commit_list *next = NULL, *current, *backup;
3475         for (current = list; current; current = backup) {
3476                 backup = current->next;
3477                 current->next = next;
3478                 next = current;
3479         }
3480         return next;
3481 }
3482
3483 /*
3484  * Merge the commits h1 and h2, return the resulting virtual
3485  * commit object and a flag indicating the cleanness of the merge.
3486  */
3487 int merge_recursive(struct merge_options *opt,
3488                     struct commit *h1,
3489                     struct commit *h2,
3490                     struct commit_list *ca,
3491                     struct commit **result)
3492 {
3493         struct commit_list *iter;
3494         struct commit *merged_common_ancestors;
3495         struct tree *mrtree;
3496         int clean;
3497
3498         if (show(opt, 4)) {
3499                 output(opt, 4, _("Merging:"));
3500                 output_commit_title(opt, h1);
3501                 output_commit_title(opt, h2);
3502         }
3503
3504         if (!ca) {
3505                 ca = get_merge_bases(h1, h2);
3506                 ca = reverse_commit_list(ca);
3507         }
3508
3509         if (show(opt, 5)) {
3510                 unsigned cnt = commit_list_count(ca);
3511
3512                 output(opt, 5, Q_("found %u common ancestor:",
3513                                 "found %u common ancestors:", cnt), cnt);
3514                 for (iter = ca; iter; iter = iter->next)
3515                         output_commit_title(opt, iter->item);
3516         }
3517
3518         merged_common_ancestors = pop_commit(&ca);
3519         if (merged_common_ancestors == NULL) {
3520                 /* if there is no common ancestor, use an empty tree */
3521                 struct tree *tree;
3522
3523                 tree = lookup_tree(opt->repo, opt->repo->hash_algo->empty_tree);
3524                 merged_common_ancestors = make_virtual_commit(opt->repo, tree, "ancestor");
3525         }
3526
3527         for (iter = ca; iter; iter = iter->next) {
3528                 const char *saved_b1, *saved_b2;
3529                 opt->call_depth++;
3530                 /*
3531                  * When the merge fails, the result contains files
3532                  * with conflict markers. The cleanness flag is
3533                  * ignored (unless indicating an error), it was never
3534                  * actually used, as result of merge_trees has always
3535                  * overwritten it: the committed "conflicts" were
3536                  * already resolved.
3537                  */
3538                 discard_index(opt->repo->index);
3539                 saved_b1 = opt->branch1;
3540                 saved_b2 = opt->branch2;
3541                 opt->branch1 = "Temporary merge branch 1";
3542                 opt->branch2 = "Temporary merge branch 2";
3543                 if (merge_recursive(opt, merged_common_ancestors, iter->item,
3544                                     NULL, &merged_common_ancestors) < 0)
3545                         return -1;
3546                 opt->branch1 = saved_b1;
3547                 opt->branch2 = saved_b2;
3548                 opt->call_depth--;
3549
3550                 if (!merged_common_ancestors)
3551                         return err(opt, _("merge returned no commit"));
3552         }
3553
3554         discard_index(opt->repo->index);
3555         if (!opt->call_depth)
3556                 repo_read_index(opt->repo);
3557
3558         opt->ancestor = "merged common ancestors";
3559         clean = merge_trees(opt, get_commit_tree(h1), get_commit_tree(h2),
3560                             get_commit_tree(merged_common_ancestors),
3561                             &mrtree);
3562         if (clean < 0) {
3563                 flush_output(opt);
3564                 return clean;
3565         }
3566
3567         if (opt->call_depth) {
3568                 *result = make_virtual_commit(opt->repo, mrtree, "merged tree");
3569                 commit_list_insert(h1, &(*result)->parents);
3570                 commit_list_insert(h2, &(*result)->parents->next);
3571         }
3572         flush_output(opt);
3573         if (!opt->call_depth && opt->buffer_output < 2)
3574                 strbuf_release(&opt->obuf);
3575         if (show(opt, 2))
3576                 diff_warn_rename_limit("merge.renamelimit",
3577                                        opt->needed_rename_limit, 0);
3578         return clean;
3579 }
3580
3581 static struct commit *get_ref(struct repository *repo, const struct object_id *oid,
3582                               const char *name)
3583 {
3584         struct object *object;
3585
3586         object = deref_tag(repo, parse_object(repo, oid),
3587                            name, strlen(name));
3588         if (!object)
3589                 return NULL;
3590         if (object->type == OBJ_TREE)
3591                 return make_virtual_commit(repo, (struct tree*)object, name);
3592         if (object->type != OBJ_COMMIT)
3593                 return NULL;
3594         if (parse_commit((struct commit *)object))
3595                 return NULL;
3596         return (struct commit *)object;
3597 }
3598
3599 int merge_recursive_generic(struct merge_options *opt,
3600                             const struct object_id *head,
3601                             const struct object_id *merge,
3602                             int num_base_list,
3603                             const struct object_id **base_list,
3604                             struct commit **result)
3605 {
3606         int clean;
3607         struct lock_file lock = LOCK_INIT;
3608         struct commit *head_commit = get_ref(opt->repo, head, opt->branch1);
3609         struct commit *next_commit = get_ref(opt->repo, merge, opt->branch2);
3610         struct commit_list *ca = NULL;
3611
3612         if (base_list) {
3613                 int i;
3614                 for (i = 0; i < num_base_list; ++i) {
3615                         struct commit *base;
3616                         if (!(base = get_ref(opt->repo, base_list[i], oid_to_hex(base_list[i]))))
3617                                 return err(opt, _("Could not parse object '%s'"),
3618                                            oid_to_hex(base_list[i]));
3619                         commit_list_insert(base, &ca);
3620                 }
3621         }
3622
3623         repo_hold_locked_index(opt->repo, &lock, LOCK_DIE_ON_ERROR);
3624         clean = merge_recursive(opt, head_commit, next_commit, ca,
3625                                 result);
3626         if (clean < 0) {
3627                 rollback_lock_file(&lock);
3628                 return clean;
3629         }
3630
3631         if (write_locked_index(opt->repo->index, &lock,
3632                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3633                 return err(opt, _("Unable to write index."));
3634
3635         return clean ? 0 : 1;
3636 }
3637
3638 static void merge_recursive_config(struct merge_options *opt)
3639 {
3640         char *value = NULL;
3641         git_config_get_int("merge.verbosity", &opt->verbosity);
3642         git_config_get_int("diff.renamelimit", &opt->diff_rename_limit);
3643         git_config_get_int("merge.renamelimit", &opt->merge_rename_limit);
3644         if (!git_config_get_string("diff.renames", &value)) {
3645                 opt->diff_detect_rename = git_config_rename("diff.renames", value);
3646                 free(value);
3647         }
3648         if (!git_config_get_string("merge.renames", &value)) {
3649                 opt->merge_detect_rename = git_config_rename("merge.renames", value);
3650                 free(value);
3651         }
3652         if (!git_config_get_string("merge.directoryrenames", &value)) {
3653                 int boolval = git_parse_maybe_bool(value);
3654                 if (0 <= boolval) {
3655                         opt->detect_directory_renames = boolval ? 2 : 0;
3656                 } else if (!strcasecmp(value, "conflict")) {
3657                         opt->detect_directory_renames = 1;
3658                 } /* avoid erroring on values from future versions of git */
3659                 free(value);
3660         }
3661         git_config(git_xmerge_config, NULL);
3662 }
3663
3664 void init_merge_options(struct merge_options *opt,
3665                         struct repository *repo)
3666 {
3667         const char *merge_verbosity;
3668         memset(opt, 0, sizeof(struct merge_options));
3669         opt->repo = repo;
3670         opt->verbosity = 2;
3671         opt->buffer_output = 1;
3672         opt->diff_rename_limit = -1;
3673         opt->merge_rename_limit = -1;
3674         opt->renormalize = 0;
3675         opt->diff_detect_rename = -1;
3676         opt->merge_detect_rename = -1;
3677         opt->detect_directory_renames = 1;
3678         merge_recursive_config(opt);
3679         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3680         if (merge_verbosity)
3681                 opt->verbosity = strtol(merge_verbosity, NULL, 10);
3682         if (opt->verbosity >= 5)
3683                 opt->buffer_output = 0;
3684         strbuf_init(&opt->obuf, 0);
3685         string_list_init(&opt->df_conflict_file_set, 1);
3686 }
3687
3688 int parse_merge_opt(struct merge_options *opt, const char *s)
3689 {
3690         const char *arg;
3691
3692         if (!s || !*s)
3693                 return -1;
3694         if (!strcmp(s, "ours"))
3695                 opt->recursive_variant = MERGE_RECURSIVE_OURS;
3696         else if (!strcmp(s, "theirs"))
3697                 opt->recursive_variant = MERGE_RECURSIVE_THEIRS;
3698         else if (!strcmp(s, "subtree"))
3699                 opt->subtree_shift = "";
3700         else if (skip_prefix(s, "subtree=", &arg))
3701                 opt->subtree_shift = arg;
3702         else if (!strcmp(s, "patience"))
3703                 opt->xdl_opts = DIFF_WITH_ALG(opt, PATIENCE_DIFF);
3704         else if (!strcmp(s, "histogram"))
3705                 opt->xdl_opts = DIFF_WITH_ALG(opt, HISTOGRAM_DIFF);
3706         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3707                 long value = parse_algorithm_value(arg);
3708                 if (value < 0)
3709                         return -1;
3710                 /* clear out previous settings */
3711                 DIFF_XDL_CLR(opt, NEED_MINIMAL);
3712                 opt->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3713                 opt->xdl_opts |= value;
3714         }
3715         else if (!strcmp(s, "ignore-space-change"))
3716                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_CHANGE);
3717         else if (!strcmp(s, "ignore-all-space"))
3718                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE);
3719         else if (!strcmp(s, "ignore-space-at-eol"))
3720                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_AT_EOL);
3721         else if (!strcmp(s, "ignore-cr-at-eol"))
3722                 DIFF_XDL_SET(opt, IGNORE_CR_AT_EOL);
3723         else if (!strcmp(s, "renormalize"))
3724                 opt->renormalize = 1;
3725         else if (!strcmp(s, "no-renormalize"))
3726                 opt->renormalize = 0;
3727         else if (!strcmp(s, "no-renames"))
3728                 opt->merge_detect_rename = 0;
3729         else if (!strcmp(s, "find-renames")) {
3730                 opt->merge_detect_rename = 1;
3731                 opt->rename_score = 0;
3732         }
3733         else if (skip_prefix(s, "find-renames=", &arg) ||
3734                  skip_prefix(s, "rename-threshold=", &arg)) {
3735                 if ((opt->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3736                         return -1;
3737                 opt->merge_detect_rename = 1;
3738         }
3739         /*
3740          * Please update $__git_merge_strategy_options in
3741          * git-completion.bash when you add new options
3742          */
3743         else
3744                 return -1;
3745         return 0;
3746 }