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