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