hashmap_get{,_from_hash} return "struct hashmap_entry *"
[git] / merge-recursive.c
1 /*
2  * Recursive Merge algorithm stolen from git-merge-recursive.py by
3  * Fredrik Kuivinen.
4  * The thieves were Alex Riesen and Johannes Schindelin, in June/July 2006
5  */
6 #include "cache.h"
7 #include "config.h"
8 #include "advice.h"
9 #include "lockfile.h"
10 #include "cache-tree.h"
11 #include "object-store.h"
12 #include "repository.h"
13 #include "commit.h"
14 #include "blob.h"
15 #include "builtin.h"
16 #include "tree-walk.h"
17 #include "diff.h"
18 #include "diffcore.h"
19 #include "tag.h"
20 #include "alloc.h"
21 #include "unpack-trees.h"
22 #include "string-list.h"
23 #include "xdiff-interface.h"
24 #include "ll-merge.h"
25 #include "attr.h"
26 #include "merge-recursive.h"
27 #include "dir.h"
28 #include "submodule.h"
29 #include "revision.h"
30 #include "commit-reach.h"
31
32 struct path_hashmap_entry {
33         struct hashmap_entry e;
34         char path[FLEX_ARRAY];
35 };
36
37 static int path_hashmap_cmp(const void *cmp_data,
38                             const void *entry,
39                             const void *entry_or_key,
40                             const void *keydata)
41 {
42         const struct path_hashmap_entry *a = entry;
43         const struct path_hashmap_entry *b = entry_or_key;
44         const char *key = keydata;
45
46         if (ignore_case)
47                 return strcasecmp(a->path, key ? key : b->path);
48         else
49                 return strcmp(a->path, key ? key : b->path);
50 }
51
52 static unsigned int path_hash(const char *path)
53 {
54         return ignore_case ? strihash(path) : strhash(path);
55 }
56
57 static struct dir_rename_entry *dir_rename_find_entry(struct hashmap *hashmap,
58                                                       char *dir)
59 {
60         struct dir_rename_entry key;
61
62         if (dir == NULL)
63                 return NULL;
64         hashmap_entry_init(&key.ent, strhash(dir));
65         key.dir = dir;
66         return hashmap_get_entry(hashmap, &key, NULL,
67                                 struct dir_rename_entry, ent);
68 }
69
70 static int dir_rename_cmp(const void *unused_cmp_data,
71                           const void *entry,
72                           const void *entry_or_key,
73                           const void *unused_keydata)
74 {
75         const struct dir_rename_entry *e1 = entry;
76         const struct dir_rename_entry *e2 = entry_or_key;
77
78         return strcmp(e1->dir, e2->dir);
79 }
80
81 static void dir_rename_init(struct hashmap *map)
82 {
83         hashmap_init(map, dir_rename_cmp, NULL, 0);
84 }
85
86 static void dir_rename_entry_init(struct dir_rename_entry *entry,
87                                   char *directory)
88 {
89         hashmap_entry_init(&entry->ent, strhash(directory));
90         entry->dir = directory;
91         entry->non_unique_new_dir = 0;
92         strbuf_init(&entry->new_dir, 0);
93         string_list_init(&entry->possible_new_dirs, 0);
94 }
95
96 static struct collision_entry *collision_find_entry(struct hashmap *hashmap,
97                                                     char *target_file)
98 {
99         struct collision_entry key;
100
101         hashmap_entry_init(&key.ent, strhash(target_file));
102         key.target_file = target_file;
103         return hashmap_get_entry(hashmap, &key, NULL,
104                                 struct collision_entry, ent);
105 }
106
107 static int collision_cmp(void *unused_cmp_data,
108                          const struct collision_entry *e1,
109                          const struct collision_entry *e2,
110                          const void *unused_keydata)
111 {
112         return strcmp(e1->target_file, e2->target_file);
113 }
114
115 static void collision_init(struct hashmap *map)
116 {
117         hashmap_init(map, (hashmap_cmp_fn) collision_cmp, NULL, 0);
118 }
119
120 static void flush_output(struct merge_options *opt)
121 {
122         if (opt->buffer_output < 2 && opt->obuf.len) {
123                 fputs(opt->obuf.buf, stdout);
124                 strbuf_reset(&opt->obuf);
125         }
126 }
127
128 static int err(struct merge_options *opt, const char *err, ...)
129 {
130         va_list params;
131
132         if (opt->buffer_output < 2)
133                 flush_output(opt);
134         else {
135                 strbuf_complete(&opt->obuf, '\n');
136                 strbuf_addstr(&opt->obuf, "error: ");
137         }
138         va_start(params, err);
139         strbuf_vaddf(&opt->obuf, err, params);
140         va_end(params);
141         if (opt->buffer_output > 1)
142                 strbuf_addch(&opt->obuf, '\n');
143         else {
144                 error("%s", opt->obuf.buf);
145                 strbuf_reset(&opt->obuf);
146         }
147
148         return -1;
149 }
150
151 static struct tree *shift_tree_object(struct repository *repo,
152                                       struct tree *one, struct tree *two,
153                                       const char *subtree_shift)
154 {
155         struct object_id shifted;
156
157         if (!*subtree_shift) {
158                 shift_tree(repo, &one->object.oid, &two->object.oid, &shifted, 0);
159         } else {
160                 shift_tree_by(repo, &one->object.oid, &two->object.oid, &shifted,
161                               subtree_shift);
162         }
163         if (oideq(&two->object.oid, &shifted))
164                 return two;
165         return lookup_tree(repo, &shifted);
166 }
167
168 static inline void set_commit_tree(struct commit *c, struct tree *t)
169 {
170         c->maybe_tree = t;
171 }
172
173 static struct commit *make_virtual_commit(struct repository *repo,
174                                           struct tree *tree,
175                                           const char *comment)
176 {
177         struct commit *commit = alloc_commit_node(repo);
178
179         set_merge_remote_desc(commit, comment, (struct object *)commit);
180         set_commit_tree(commit, tree);
181         commit->object.parsed = 1;
182         return commit;
183 }
184
185 /*
186  * Since we use get_tree_entry(), which does not put the read object into
187  * the object pool, we cannot rely on a == b.
188  */
189 static int oid_eq(const struct object_id *a, const struct object_id *b)
190 {
191         if (!a && !b)
192                 return 2;
193         return a && b && oideq(a, b);
194 }
195
196 enum rename_type {
197         RENAME_NORMAL = 0,
198         RENAME_VIA_DIR,
199         RENAME_ADD,
200         RENAME_DELETE,
201         RENAME_ONE_FILE_TO_ONE,
202         RENAME_ONE_FILE_TO_TWO,
203         RENAME_TWO_FILES_TO_ONE
204 };
205
206 /*
207  * Since we want to write the index eventually, we cannot reuse the index
208  * for these (temporary) data.
209  */
210 struct stage_data {
211         struct diff_filespec stages[4]; /* mostly for oid & mode; maybe path */
212         struct rename_conflict_info *rename_conflict_info;
213         unsigned processed:1;
214 };
215
216 struct rename {
217         unsigned processed:1;
218         struct diff_filepair *pair;
219         const char *branch; /* branch that the rename occurred on */
220         /*
221          * If directory rename detection affected this rename, what was its
222          * original type ('A' or 'R') and it's original destination before
223          * the directory rename (otherwise, '\0' and NULL for these two vars).
224          */
225         char dir_rename_original_type;
226         char *dir_rename_original_dest;
227         /*
228          * Purpose of src_entry and dst_entry:
229          *
230          * If 'before' is renamed to 'after' then src_entry will contain
231          * the versions of 'before' from the merge_base, HEAD, and MERGE in
232          * stages 1, 2, and 3; dst_entry will contain the respective
233          * versions of 'after' in corresponding locations.  Thus, we have a
234          * total of six modes and oids, though some will be null.  (Stage 0
235          * is ignored; we're interested in handling conflicts.)
236          *
237          * Since we don't turn on break-rewrites by default, neither
238          * src_entry nor dst_entry can have all three of their stages have
239          * non-null oids, meaning at most four of the six will be non-null.
240          * Also, since this is a rename, both src_entry and dst_entry will
241          * have at least one non-null oid, meaning at least two will be
242          * non-null.  Of the six oids, a typical rename will have three be
243          * non-null.  Only two implies a rename/delete, and four implies a
244          * rename/add.
245          */
246         struct stage_data *src_entry;
247         struct stage_data *dst_entry;
248 };
249
250 struct rename_conflict_info {
251         enum rename_type rename_type;
252         struct rename *ren1;
253         struct rename *ren2;
254 };
255
256 static inline void setup_rename_conflict_info(enum rename_type rename_type,
257                                               struct merge_options *opt,
258                                               struct rename *ren1,
259                                               struct rename *ren2)
260 {
261         struct rename_conflict_info *ci;
262
263         /*
264          * When we have two renames involved, it's easiest to get the
265          * correct things into stage 2 and 3, and to make sure that the
266          * content merge puts HEAD before the other branch if we just
267          * ensure that branch1 == opt->branch1.  So, simply flip arguments
268          * around if we don't have that.
269          */
270         if (ren2 && ren1->branch != opt->branch1) {
271                 setup_rename_conflict_info(rename_type, opt, ren2, ren1);
272                 return;
273         }
274
275         ci = xcalloc(1, sizeof(struct rename_conflict_info));
276         ci->rename_type = rename_type;
277         ci->ren1 = ren1;
278         ci->ren2 = ren2;
279
280         ci->ren1->dst_entry->processed = 0;
281         ci->ren1->dst_entry->rename_conflict_info = ci;
282         if (ren2) {
283                 ci->ren2->dst_entry->rename_conflict_info = ci;
284         }
285 }
286
287 static int show(struct merge_options *opt, int v)
288 {
289         return (!opt->call_depth && opt->verbosity >= v) || opt->verbosity >= 5;
290 }
291
292 __attribute__((format (printf, 3, 4)))
293 static void output(struct merge_options *opt, int v, const char *fmt, ...)
294 {
295         va_list ap;
296
297         if (!show(opt, v))
298                 return;
299
300         strbuf_addchars(&opt->obuf, ' ', opt->call_depth * 2);
301
302         va_start(ap, fmt);
303         strbuf_vaddf(&opt->obuf, fmt, ap);
304         va_end(ap);
305
306         strbuf_addch(&opt->obuf, '\n');
307         if (!opt->buffer_output)
308                 flush_output(opt);
309 }
310
311 static void output_commit_title(struct merge_options *opt, struct commit *commit)
312 {
313         struct merge_remote_desc *desc;
314
315         strbuf_addchars(&opt->obuf, ' ', opt->call_depth * 2);
316         desc = merge_remote_util(commit);
317         if (desc)
318                 strbuf_addf(&opt->obuf, "virtual %s\n", desc->name);
319         else {
320                 strbuf_add_unique_abbrev(&opt->obuf, &commit->object.oid,
321                                          DEFAULT_ABBREV);
322                 strbuf_addch(&opt->obuf, ' ');
323                 if (parse_commit(commit) != 0)
324                         strbuf_addstr(&opt->obuf, _("(bad commit)\n"));
325                 else {
326                         const char *title;
327                         const char *msg = get_commit_buffer(commit, NULL);
328                         int len = find_commit_subject(msg, &title);
329                         if (len)
330                                 strbuf_addf(&opt->obuf, "%.*s\n", len, title);
331                         unuse_commit_buffer(commit, msg);
332                 }
333         }
334         flush_output(opt);
335 }
336
337 static int add_cacheinfo(struct merge_options *opt,
338                          const struct diff_filespec *blob,
339                          const char *path, int stage, int refresh, int options)
340 {
341         struct index_state *istate = opt->repo->index;
342         struct cache_entry *ce;
343         int ret;
344
345         ce = make_cache_entry(istate, blob->mode, &blob->oid, path, stage, 0);
346         if (!ce)
347                 return err(opt, _("add_cacheinfo failed for path '%s'; merge aborting."), path);
348
349         ret = add_index_entry(istate, ce, options);
350         if (refresh) {
351                 struct cache_entry *nce;
352
353                 nce = refresh_cache_entry(istate, ce,
354                                           CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
355                 if (!nce)
356                         return err(opt, _("add_cacheinfo failed to refresh for path '%s'; merge aborting."), path);
357                 if (nce != ce)
358                         ret = add_index_entry(istate, nce, options);
359         }
360         return ret;
361 }
362
363 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
364 {
365         parse_tree(tree);
366         init_tree_desc(desc, tree->buffer, tree->size);
367 }
368
369 static int unpack_trees_start(struct merge_options *opt,
370                               struct tree *common,
371                               struct tree *head,
372                               struct tree *merge)
373 {
374         int rc;
375         struct tree_desc t[3];
376         struct index_state tmp_index = { NULL };
377
378         memset(&opt->unpack_opts, 0, sizeof(opt->unpack_opts));
379         if (opt->call_depth)
380                 opt->unpack_opts.index_only = 1;
381         else
382                 opt->unpack_opts.update = 1;
383         opt->unpack_opts.merge = 1;
384         opt->unpack_opts.head_idx = 2;
385         opt->unpack_opts.fn = threeway_merge;
386         opt->unpack_opts.src_index = opt->repo->index;
387         opt->unpack_opts.dst_index = &tmp_index;
388         opt->unpack_opts.aggressive = !merge_detect_rename(opt);
389         setup_unpack_trees_porcelain(&opt->unpack_opts, "merge");
390
391         init_tree_desc_from_tree(t+0, common);
392         init_tree_desc_from_tree(t+1, head);
393         init_tree_desc_from_tree(t+2, merge);
394
395         rc = unpack_trees(3, t, &opt->unpack_opts);
396         cache_tree_free(&opt->repo->index->cache_tree);
397
398         /*
399          * Update opt->repo->index to match the new results, AFTER saving a copy
400          * in opt->orig_index.  Update src_index to point to the saved copy.
401          * (verify_uptodate() checks src_index, and the original index is
402          * the one that had the necessary modification timestamps.)
403          */
404         opt->orig_index = *opt->repo->index;
405         *opt->repo->index = tmp_index;
406         opt->unpack_opts.src_index = &opt->orig_index;
407
408         return rc;
409 }
410
411 static void unpack_trees_finish(struct merge_options *opt)
412 {
413         discard_index(&opt->orig_index);
414         clear_unpack_trees_porcelain(&opt->unpack_opts);
415 }
416
417 struct tree *write_tree_from_memory(struct merge_options *opt)
418 {
419         struct tree *result = NULL;
420         struct index_state *istate = opt->repo->index;
421
422         if (unmerged_index(istate)) {
423                 int i;
424                 fprintf(stderr, "BUG: There are unmerged index entries:\n");
425                 for (i = 0; i < istate->cache_nr; i++) {
426                         const struct cache_entry *ce = istate->cache[i];
427                         if (ce_stage(ce))
428                                 fprintf(stderr, "BUG: %d %.*s\n", ce_stage(ce),
429                                         (int)ce_namelen(ce), ce->name);
430                 }
431                 BUG("unmerged index entries in merge-recursive.c");
432         }
433
434         if (!istate->cache_tree)
435                 istate->cache_tree = cache_tree();
436
437         if (!cache_tree_fully_valid(istate->cache_tree) &&
438             cache_tree_update(istate, 0) < 0) {
439                 err(opt, _("error building trees"));
440                 return NULL;
441         }
442
443         result = lookup_tree(opt->repo, &istate->cache_tree->oid);
444
445         return result;
446 }
447
448 static int save_files_dirs(const struct object_id *oid,
449                            struct strbuf *base, const char *path,
450                            unsigned int mode, int stage, void *context)
451 {
452         struct path_hashmap_entry *entry;
453         int baselen = base->len;
454         struct merge_options *opt = context;
455
456         strbuf_addstr(base, path);
457
458         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
459         hashmap_entry_init(&entry->e, path_hash(entry->path));
460         hashmap_add(&opt->current_file_dir_set, &entry->e);
461
462         strbuf_setlen(base, baselen);
463         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
464 }
465
466 static void get_files_dirs(struct merge_options *opt, struct tree *tree)
467 {
468         struct pathspec match_all;
469         memset(&match_all, 0, sizeof(match_all));
470         read_tree_recursive(opt->repo, tree, "", 0, 0,
471                             &match_all, save_files_dirs, opt);
472 }
473
474 static int get_tree_entry_if_blob(struct repository *r,
475                                   const struct object_id *tree,
476                                   const char *path,
477                                   struct diff_filespec *dfs)
478 {
479         int ret;
480
481         ret = get_tree_entry(r, tree, path, &dfs->oid, &dfs->mode);
482         if (S_ISDIR(dfs->mode)) {
483                 oidcpy(&dfs->oid, &null_oid);
484                 dfs->mode = 0;
485         }
486         return ret;
487 }
488
489 /*
490  * Returns an index_entry instance which doesn't have to correspond to
491  * a real cache entry in Git's index.
492  */
493 static struct stage_data *insert_stage_data(struct repository *r,
494                 const char *path,
495                 struct tree *o, struct tree *a, struct tree *b,
496                 struct string_list *entries)
497 {
498         struct string_list_item *item;
499         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
500         get_tree_entry_if_blob(r, &o->object.oid, path, &e->stages[1]);
501         get_tree_entry_if_blob(r, &a->object.oid, path, &e->stages[2]);
502         get_tree_entry_if_blob(r, &b->object.oid, path, &e->stages[3]);
503         item = string_list_insert(entries, path);
504         item->util = e;
505         return e;
506 }
507
508 /*
509  * Create a dictionary mapping file names to stage_data objects. The
510  * dictionary contains one entry for every path with a non-zero stage entry.
511  */
512 static struct string_list *get_unmerged(struct index_state *istate)
513 {
514         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
515         int i;
516
517         unmerged->strdup_strings = 1;
518
519         for (i = 0; i < istate->cache_nr; i++) {
520                 struct string_list_item *item;
521                 struct stage_data *e;
522                 const struct cache_entry *ce = istate->cache[i];
523                 if (!ce_stage(ce))
524                         continue;
525
526                 item = string_list_lookup(unmerged, ce->name);
527                 if (!item) {
528                         item = string_list_insert(unmerged, ce->name);
529                         item->util = xcalloc(1, sizeof(struct stage_data));
530                 }
531                 e = item->util;
532                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
533                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
534         }
535
536         return unmerged;
537 }
538
539 static int string_list_df_name_compare(const char *one, const char *two)
540 {
541         int onelen = strlen(one);
542         int twolen = strlen(two);
543         /*
544          * Here we only care that entries for D/F conflicts are
545          * adjacent, in particular with the file of the D/F conflict
546          * appearing before files below the corresponding directory.
547          * The order of the rest of the list is irrelevant for us.
548          *
549          * To achieve this, we sort with df_name_compare and provide
550          * the mode S_IFDIR so that D/F conflicts will sort correctly.
551          * We use the mode S_IFDIR for everything else for simplicity,
552          * since in other cases any changes in their order due to
553          * sorting cause no problems for us.
554          */
555         int cmp = df_name_compare(one, onelen, S_IFDIR,
556                                   two, twolen, S_IFDIR);
557         /*
558          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
559          * that 'foo' comes before 'foo/bar'.
560          */
561         if (cmp)
562                 return cmp;
563         return onelen - twolen;
564 }
565
566 static void record_df_conflict_files(struct merge_options *opt,
567                                      struct string_list *entries)
568 {
569         /* If there is a D/F conflict and the file for such a conflict
570          * currently exists in the working tree, we want to allow it to be
571          * removed to make room for the corresponding directory if needed.
572          * The files underneath the directories of such D/F conflicts will
573          * be processed before the corresponding file involved in the D/F
574          * conflict.  If the D/F directory ends up being removed by the
575          * merge, then we won't have to touch the D/F file.  If the D/F
576          * directory needs to be written to the working copy, then the D/F
577          * file will simply be removed (in make_room_for_path()) to make
578          * room for the necessary paths.  Note that if both the directory
579          * and the file need to be present, then the D/F file will be
580          * reinstated with a new unique name at the time it is processed.
581          */
582         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
583         const char *last_file = NULL;
584         int last_len = 0;
585         int i;
586
587         /*
588          * If we're merging merge-bases, we don't want to bother with
589          * any working directory changes.
590          */
591         if (opt->call_depth)
592                 return;
593
594         /* Ensure D/F conflicts are adjacent in the entries list. */
595         for (i = 0; i < entries->nr; i++) {
596                 struct string_list_item *next = &entries->items[i];
597                 string_list_append(&df_sorted_entries, next->string)->util =
598                                    next->util;
599         }
600         df_sorted_entries.cmp = string_list_df_name_compare;
601         string_list_sort(&df_sorted_entries);
602
603         string_list_clear(&opt->df_conflict_file_set, 1);
604         for (i = 0; i < df_sorted_entries.nr; i++) {
605                 const char *path = df_sorted_entries.items[i].string;
606                 int len = strlen(path);
607                 struct stage_data *e = df_sorted_entries.items[i].util;
608
609                 /*
610                  * Check if last_file & path correspond to a D/F conflict;
611                  * i.e. whether path is last_file+'/'+<something>.
612                  * If so, record that it's okay to remove last_file to make
613                  * room for path and friends if needed.
614                  */
615                 if (last_file &&
616                     len > last_len &&
617                     memcmp(path, last_file, last_len) == 0 &&
618                     path[last_len] == '/') {
619                         string_list_insert(&opt->df_conflict_file_set, last_file);
620                 }
621
622                 /*
623                  * Determine whether path could exist as a file in the
624                  * working directory as a possible D/F conflict.  This
625                  * will only occur when it exists in stage 2 as a
626                  * file.
627                  */
628                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
629                         last_file = path;
630                         last_len = len;
631                 } else {
632                         last_file = NULL;
633                 }
634         }
635         string_list_clear(&df_sorted_entries, 0);
636 }
637
638 static int update_stages(struct merge_options *opt, const char *path,
639                          const struct diff_filespec *o,
640                          const struct diff_filespec *a,
641                          const struct diff_filespec *b)
642 {
643
644         /*
645          * NOTE: It is usually a bad idea to call update_stages on a path
646          * before calling update_file on that same path, since it can
647          * sometimes lead to spurious "refusing to lose untracked file..."
648          * messages from update_file (via make_room_for path via
649          * would_lose_untracked).  Instead, reverse the order of the calls
650          * (executing update_file first and then update_stages).
651          */
652         int clear = 1;
653         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
654         if (clear)
655                 if (remove_file_from_index(opt->repo->index, path))
656                         return -1;
657         if (o)
658                 if (add_cacheinfo(opt, o, path, 1, 0, options))
659                         return -1;
660         if (a)
661                 if (add_cacheinfo(opt, a, path, 2, 0, options))
662                         return -1;
663         if (b)
664                 if (add_cacheinfo(opt, b, path, 3, 0, options))
665                         return -1;
666         return 0;
667 }
668
669 static void update_entry(struct stage_data *entry,
670                          struct diff_filespec *o,
671                          struct diff_filespec *a,
672                          struct diff_filespec *b)
673 {
674         entry->processed = 0;
675         entry->stages[1].mode = o->mode;
676         entry->stages[2].mode = a->mode;
677         entry->stages[3].mode = b->mode;
678         oidcpy(&entry->stages[1].oid, &o->oid);
679         oidcpy(&entry->stages[2].oid, &a->oid);
680         oidcpy(&entry->stages[3].oid, &b->oid);
681 }
682
683 static int remove_file(struct merge_options *opt, int clean,
684                        const char *path, int no_wd)
685 {
686         int update_cache = opt->call_depth || clean;
687         int update_working_directory = !opt->call_depth && !no_wd;
688
689         if (update_cache) {
690                 if (remove_file_from_index(opt->repo->index, path))
691                         return -1;
692         }
693         if (update_working_directory) {
694                 if (ignore_case) {
695                         struct cache_entry *ce;
696                         ce = index_file_exists(opt->repo->index, path, strlen(path),
697                                                ignore_case);
698                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
699                                 return 0;
700                 }
701                 if (remove_path(path))
702                         return -1;
703         }
704         return 0;
705 }
706
707 /* add a string to a strbuf, but converting "/" to "_" */
708 static void add_flattened_path(struct strbuf *out, const char *s)
709 {
710         size_t i = out->len;
711         strbuf_addstr(out, s);
712         for (; i < out->len; i++)
713                 if (out->buf[i] == '/')
714                         out->buf[i] = '_';
715 }
716
717 static char *unique_path(struct merge_options *opt, const char *path, const char *branch)
718 {
719         struct path_hashmap_entry *entry;
720         struct strbuf newpath = STRBUF_INIT;
721         int suffix = 0;
722         size_t base_len;
723
724         strbuf_addf(&newpath, "%s~", path);
725         add_flattened_path(&newpath, branch);
726
727         base_len = newpath.len;
728         while (hashmap_get_from_hash(&opt->current_file_dir_set,
729                                      path_hash(newpath.buf), newpath.buf) ||
730                (!opt->call_depth && file_exists(newpath.buf))) {
731                 strbuf_setlen(&newpath, base_len);
732                 strbuf_addf(&newpath, "_%d", suffix++);
733         }
734
735         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
736         hashmap_entry_init(&entry->e, path_hash(entry->path));
737         hashmap_add(&opt->current_file_dir_set, &entry->e);
738         return strbuf_detach(&newpath, NULL);
739 }
740
741 /**
742  * Check whether a directory in the index is in the way of an incoming
743  * file.  Return 1 if so.  If check_working_copy is non-zero, also
744  * check the working directory.  If empty_ok is non-zero, also return
745  * 0 in the case where the working-tree dir exists but is empty.
746  */
747 static int dir_in_way(struct index_state *istate, const char *path,
748                       int check_working_copy, int empty_ok)
749 {
750         int pos;
751         struct strbuf dirpath = STRBUF_INIT;
752         struct stat st;
753
754         strbuf_addstr(&dirpath, path);
755         strbuf_addch(&dirpath, '/');
756
757         pos = index_name_pos(istate, dirpath.buf, dirpath.len);
758
759         if (pos < 0)
760                 pos = -1 - pos;
761         if (pos < istate->cache_nr &&
762             !strncmp(dirpath.buf, istate->cache[pos]->name, dirpath.len)) {
763                 strbuf_release(&dirpath);
764                 return 1;
765         }
766
767         strbuf_release(&dirpath);
768         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
769                 !(empty_ok && is_empty_dir(path));
770 }
771
772 /*
773  * Returns whether path was tracked in the index before the merge started,
774  * and its oid and mode match the specified values
775  */
776 static int was_tracked_and_matches(struct merge_options *opt, const char *path,
777                                    const struct diff_filespec *blob)
778 {
779         int pos = index_name_pos(&opt->orig_index, path, strlen(path));
780         struct cache_entry *ce;
781
782         if (0 > pos)
783                 /* we were not tracking this path before the merge */
784                 return 0;
785
786         /* See if the file we were tracking before matches */
787         ce = opt->orig_index.cache[pos];
788         return (oid_eq(&ce->oid, &blob->oid) && ce->ce_mode == blob->mode);
789 }
790
791 /*
792  * Returns whether path was tracked in the index before the merge started
793  */
794 static int was_tracked(struct merge_options *opt, const char *path)
795 {
796         int pos = index_name_pos(&opt->orig_index, path, strlen(path));
797
798         if (0 <= pos)
799                 /* we were tracking this path before the merge */
800                 return 1;
801
802         return 0;
803 }
804
805 static int would_lose_untracked(struct merge_options *opt, const char *path)
806 {
807         struct index_state *istate = opt->repo->index;
808
809         /*
810          * This may look like it can be simplified to:
811          *   return !was_tracked(opt, path) && file_exists(path)
812          * but it can't.  This function needs to know whether path was in
813          * the working tree due to EITHER having been tracked in the index
814          * before the merge OR having been put into the working copy and
815          * index by unpack_trees().  Due to that either-or requirement, we
816          * check the current index instead of the original one.
817          *
818          * Note that we do not need to worry about merge-recursive itself
819          * updating the index after unpack_trees() and before calling this
820          * function, because we strictly require all code paths in
821          * merge-recursive to update the working tree first and the index
822          * second.  Doing otherwise would break
823          * update_file()/would_lose_untracked(); see every comment in this
824          * file which mentions "update_stages".
825          */
826         int pos = index_name_pos(istate, path, strlen(path));
827
828         if (pos < 0)
829                 pos = -1 - pos;
830         while (pos < istate->cache_nr &&
831                !strcmp(path, istate->cache[pos]->name)) {
832                 /*
833                  * If stage #0, it is definitely tracked.
834                  * If it has stage #2 then it was tracked
835                  * before this merge started.  All other
836                  * cases the path was not tracked.
837                  */
838                 switch (ce_stage(istate->cache[pos])) {
839                 case 0:
840                 case 2:
841                         return 0;
842                 }
843                 pos++;
844         }
845         return file_exists(path);
846 }
847
848 static int was_dirty(struct merge_options *opt, const char *path)
849 {
850         struct cache_entry *ce;
851         int dirty = 1;
852
853         if (opt->call_depth || !was_tracked(opt, path))
854                 return !dirty;
855
856         ce = index_file_exists(opt->unpack_opts.src_index,
857                                path, strlen(path), ignore_case);
858         dirty = verify_uptodate(ce, &opt->unpack_opts) != 0;
859         return dirty;
860 }
861
862 static int make_room_for_path(struct merge_options *opt, const char *path)
863 {
864         int status, i;
865         const char *msg = _("failed to create path '%s'%s");
866
867         /* Unlink any D/F conflict files that are in the way */
868         for (i = 0; i < opt->df_conflict_file_set.nr; i++) {
869                 const char *df_path = opt->df_conflict_file_set.items[i].string;
870                 size_t pathlen = strlen(path);
871                 size_t df_pathlen = strlen(df_path);
872                 if (df_pathlen < pathlen &&
873                     path[df_pathlen] == '/' &&
874                     strncmp(path, df_path, df_pathlen) == 0) {
875                         output(opt, 3,
876                                _("Removing %s to make room for subdirectory\n"),
877                                df_path);
878                         unlink(df_path);
879                         unsorted_string_list_delete_item(&opt->df_conflict_file_set,
880                                                          i, 0);
881                         break;
882                 }
883         }
884
885         /* Make sure leading directories are created */
886         status = safe_create_leading_directories_const(path);
887         if (status) {
888                 if (status == SCLD_EXISTS)
889                         /* something else exists */
890                         return err(opt, msg, path, _(": perhaps a D/F conflict?"));
891                 return err(opt, msg, path, "");
892         }
893
894         /*
895          * Do not unlink a file in the work tree if we are not
896          * tracking it.
897          */
898         if (would_lose_untracked(opt, path))
899                 return err(opt, _("refusing to lose untracked file at '%s'"),
900                            path);
901
902         /* Successful unlink is good.. */
903         if (!unlink(path))
904                 return 0;
905         /* .. and so is no existing file */
906         if (errno == ENOENT)
907                 return 0;
908         /* .. but not some other error (who really cares what?) */
909         return err(opt, msg, path, _(": perhaps a D/F conflict?"));
910 }
911
912 static int update_file_flags(struct merge_options *opt,
913                              const struct diff_filespec *contents,
914                              const char *path,
915                              int update_cache,
916                              int update_wd)
917 {
918         int ret = 0;
919
920         if (opt->call_depth)
921                 update_wd = 0;
922
923         if (update_wd) {
924                 enum object_type type;
925                 void *buf;
926                 unsigned long size;
927
928                 if (S_ISGITLINK(contents->mode)) {
929                         /*
930                          * We may later decide to recursively descend into
931                          * the submodule directory and update its index
932                          * and/or work tree, but we do not do that now.
933                          */
934                         update_wd = 0;
935                         goto update_index;
936                 }
937
938                 buf = read_object_file(&contents->oid, &type, &size);
939                 if (!buf)
940                         return err(opt, _("cannot read object %s '%s'"),
941                                    oid_to_hex(&contents->oid), path);
942                 if (type != OBJ_BLOB) {
943                         ret = err(opt, _("blob expected for %s '%s'"),
944                                   oid_to_hex(&contents->oid), path);
945                         goto free_buf;
946                 }
947                 if (S_ISREG(contents->mode)) {
948                         struct strbuf strbuf = STRBUF_INIT;
949                         if (convert_to_working_tree(opt->repo->index, path, buf, size, &strbuf)) {
950                                 free(buf);
951                                 size = strbuf.len;
952                                 buf = strbuf_detach(&strbuf, NULL);
953                         }
954                 }
955
956                 if (make_room_for_path(opt, path) < 0) {
957                         update_wd = 0;
958                         goto free_buf;
959                 }
960                 if (S_ISREG(contents->mode) ||
961                     (!has_symlinks && S_ISLNK(contents->mode))) {
962                         int fd;
963                         int mode = (contents->mode & 0100 ? 0777 : 0666);
964
965                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
966                         if (fd < 0) {
967                                 ret = err(opt, _("failed to open '%s': %s"),
968                                           path, strerror(errno));
969                                 goto free_buf;
970                         }
971                         write_in_full(fd, buf, size);
972                         close(fd);
973                 } else if (S_ISLNK(contents->mode)) {
974                         char *lnk = xmemdupz(buf, size);
975                         safe_create_leading_directories_const(path);
976                         unlink(path);
977                         if (symlink(lnk, path))
978                                 ret = err(opt, _("failed to symlink '%s': %s"),
979                                           path, strerror(errno));
980                         free(lnk);
981                 } else
982                         ret = err(opt,
983                                   _("do not know what to do with %06o %s '%s'"),
984                                   contents->mode, oid_to_hex(&contents->oid), path);
985         free_buf:
986                 free(buf);
987         }
988 update_index:
989         if (!ret && update_cache)
990                 if (add_cacheinfo(opt, contents, path, 0, update_wd,
991                                   ADD_CACHE_OK_TO_ADD))
992                         return -1;
993         return ret;
994 }
995
996 static int update_file(struct merge_options *opt,
997                        int clean,
998                        const struct diff_filespec *contents,
999                        const char *path)
1000 {
1001         return update_file_flags(opt, contents, path,
1002                                  opt->call_depth || clean, !opt->call_depth);
1003 }
1004
1005 /* Low level file merging, update and removal */
1006
1007 struct merge_file_info {
1008         struct diff_filespec blob; /* mostly use oid & mode; sometimes path */
1009         unsigned clean:1,
1010                  merge:1;
1011 };
1012
1013 static int merge_3way(struct merge_options *opt,
1014                       mmbuffer_t *result_buf,
1015                       const struct diff_filespec *o,
1016                       const struct diff_filespec *a,
1017                       const struct diff_filespec *b,
1018                       const char *branch1,
1019                       const char *branch2,
1020                       const int extra_marker_size)
1021 {
1022         mmfile_t orig, src1, src2;
1023         struct ll_merge_options ll_opts = {0};
1024         char *base_name, *name1, *name2;
1025         int merge_status;
1026
1027         ll_opts.renormalize = opt->renormalize;
1028         ll_opts.extra_marker_size = extra_marker_size;
1029         ll_opts.xdl_opts = opt->xdl_opts;
1030
1031         if (opt->call_depth) {
1032                 ll_opts.virtual_ancestor = 1;
1033                 ll_opts.variant = 0;
1034         } else {
1035                 switch (opt->recursive_variant) {
1036                 case MERGE_RECURSIVE_OURS:
1037                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
1038                         break;
1039                 case MERGE_RECURSIVE_THEIRS:
1040                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
1041                         break;
1042                 default:
1043                         ll_opts.variant = 0;
1044                         break;
1045                 }
1046         }
1047
1048         assert(a->path && b->path);
1049         if (strcmp(a->path, b->path) ||
1050             (opt->ancestor != NULL && strcmp(a->path, o->path) != 0)) {
1051                 base_name = opt->ancestor == NULL ? NULL :
1052                         mkpathdup("%s:%s", opt->ancestor, o->path);
1053                 name1 = mkpathdup("%s:%s", branch1, a->path);
1054                 name2 = mkpathdup("%s:%s", branch2, b->path);
1055         } else {
1056                 base_name = opt->ancestor == NULL ? NULL :
1057                         mkpathdup("%s", opt->ancestor);
1058                 name1 = mkpathdup("%s", branch1);
1059                 name2 = mkpathdup("%s", branch2);
1060         }
1061
1062         read_mmblob(&orig, &o->oid);
1063         read_mmblob(&src1, &a->oid);
1064         read_mmblob(&src2, &b->oid);
1065
1066         merge_status = ll_merge(result_buf, a->path, &orig, base_name,
1067                                 &src1, name1, &src2, name2,
1068                                 opt->repo->index, &ll_opts);
1069
1070         free(base_name);
1071         free(name1);
1072         free(name2);
1073         free(orig.ptr);
1074         free(src1.ptr);
1075         free(src2.ptr);
1076         return merge_status;
1077 }
1078
1079 static int find_first_merges(struct repository *repo,
1080                              struct object_array *result, const char *path,
1081                              struct commit *a, struct commit *b)
1082 {
1083         int i, j;
1084         struct object_array merges = OBJECT_ARRAY_INIT;
1085         struct commit *commit;
1086         int contains_another;
1087
1088         char merged_revision[GIT_MAX_HEXSZ + 2];
1089         const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
1090                                    "--all", merged_revision, NULL };
1091         struct rev_info revs;
1092         struct setup_revision_opt rev_opts;
1093
1094         memset(result, 0, sizeof(struct object_array));
1095         memset(&rev_opts, 0, sizeof(rev_opts));
1096
1097         /* get all revisions that merge commit a */
1098         xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
1099                   oid_to_hex(&a->object.oid));
1100         repo_init_revisions(repo, &revs, NULL);
1101         rev_opts.submodule = path;
1102         /* FIXME: can't handle linked worktrees in submodules yet */
1103         revs.single_worktree = path != NULL;
1104         setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
1105
1106         /* save all revisions from the above list that contain b */
1107         if (prepare_revision_walk(&revs))
1108                 die("revision walk setup failed");
1109         while ((commit = get_revision(&revs)) != NULL) {
1110                 struct object *o = &(commit->object);
1111                 if (in_merge_bases(b, commit))
1112                         add_object_array(o, NULL, &merges);
1113         }
1114         reset_revision_walk();
1115
1116         /* Now we've got all merges that contain a and b. Prune all
1117          * merges that contain another found merge and save them in
1118          * result.
1119          */
1120         for (i = 0; i < merges.nr; i++) {
1121                 struct commit *m1 = (struct commit *) merges.objects[i].item;
1122
1123                 contains_another = 0;
1124                 for (j = 0; j < merges.nr; j++) {
1125                         struct commit *m2 = (struct commit *) merges.objects[j].item;
1126                         if (i != j && in_merge_bases(m2, m1)) {
1127                                 contains_another = 1;
1128                                 break;
1129                         }
1130                 }
1131
1132                 if (!contains_another)
1133                         add_object_array(merges.objects[i].item, NULL, result);
1134         }
1135
1136         object_array_clear(&merges);
1137         return result->nr;
1138 }
1139
1140 static void print_commit(struct commit *commit)
1141 {
1142         struct strbuf sb = STRBUF_INIT;
1143         struct pretty_print_context ctx = {0};
1144         ctx.date_mode.type = DATE_NORMAL;
1145         format_commit_message(commit, " %h: %m %s", &sb, &ctx);
1146         fprintf(stderr, "%s\n", sb.buf);
1147         strbuf_release(&sb);
1148 }
1149
1150 static int is_valid(const struct diff_filespec *dfs)
1151 {
1152         return dfs->mode != 0 && !is_null_oid(&dfs->oid);
1153 }
1154
1155 static int merge_submodule(struct merge_options *opt,
1156                            struct object_id *result, const char *path,
1157                            const struct object_id *base, const struct object_id *a,
1158                            const struct object_id *b)
1159 {
1160         struct commit *commit_base, *commit_a, *commit_b;
1161         int parent_count;
1162         struct object_array merges;
1163
1164         int i;
1165         int search = !opt->call_depth;
1166
1167         /* store a in result in case we fail */
1168         oidcpy(result, a);
1169
1170         /* we can not handle deletion conflicts */
1171         if (is_null_oid(base))
1172                 return 0;
1173         if (is_null_oid(a))
1174                 return 0;
1175         if (is_null_oid(b))
1176                 return 0;
1177
1178         if (add_submodule_odb(path)) {
1179                 output(opt, 1, _("Failed to merge submodule %s (not checked out)"), path);
1180                 return 0;
1181         }
1182
1183         if (!(commit_base = lookup_commit_reference(opt->repo, base)) ||
1184             !(commit_a = lookup_commit_reference(opt->repo, a)) ||
1185             !(commit_b = lookup_commit_reference(opt->repo, b))) {
1186                 output(opt, 1, _("Failed to merge submodule %s (commits not present)"), path);
1187                 return 0;
1188         }
1189
1190         /* check whether both changes are forward */
1191         if (!in_merge_bases(commit_base, commit_a) ||
1192             !in_merge_bases(commit_base, commit_b)) {
1193                 output(opt, 1, _("Failed to merge submodule %s (commits don't follow merge-base)"), path);
1194                 return 0;
1195         }
1196
1197         /* Case #1: a is contained in b or vice versa */
1198         if (in_merge_bases(commit_a, commit_b)) {
1199                 oidcpy(result, b);
1200                 if (show(opt, 3)) {
1201                         output(opt, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1202                         output_commit_title(opt, commit_b);
1203                 } else if (show(opt, 2))
1204                         output(opt, 2, _("Fast-forwarding submodule %s"), path);
1205                 else
1206                         ; /* no output */
1207
1208                 return 1;
1209         }
1210         if (in_merge_bases(commit_b, commit_a)) {
1211                 oidcpy(result, a);
1212                 if (show(opt, 3)) {
1213                         output(opt, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1214                         output_commit_title(opt, commit_a);
1215                 } else if (show(opt, 2))
1216                         output(opt, 2, _("Fast-forwarding submodule %s"), path);
1217                 else
1218                         ; /* no output */
1219
1220                 return 1;
1221         }
1222
1223         /*
1224          * Case #2: There are one or more merges that contain a and b in
1225          * the submodule. If there is only one, then present it as a
1226          * suggestion to the user, but leave it marked unmerged so the
1227          * user needs to confirm the resolution.
1228          */
1229
1230         /* Skip the search if makes no sense to the calling context.  */
1231         if (!search)
1232                 return 0;
1233
1234         /* find commit which merges them */
1235         parent_count = find_first_merges(opt->repo, &merges, path,
1236                                          commit_a, commit_b);
1237         switch (parent_count) {
1238         case 0:
1239                 output(opt, 1, _("Failed to merge submodule %s (merge following commits not found)"), path);
1240                 break;
1241
1242         case 1:
1243                 output(opt, 1, _("Failed to merge submodule %s (not fast-forward)"), path);
1244                 output(opt, 2, _("Found a possible merge resolution for the submodule:\n"));
1245                 print_commit((struct commit *) merges.objects[0].item);
1246                 output(opt, 2, _(
1247                        "If this is correct simply add it to the index "
1248                        "for example\n"
1249                        "by using:\n\n"
1250                        "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
1251                        "which will accept this suggestion.\n"),
1252                        oid_to_hex(&merges.objects[0].item->oid), path);
1253                 break;
1254
1255         default:
1256                 output(opt, 1, _("Failed to merge submodule %s (multiple merges found)"), path);
1257                 for (i = 0; i < merges.nr; i++)
1258                         print_commit((struct commit *) merges.objects[i].item);
1259         }
1260
1261         object_array_clear(&merges);
1262         return 0;
1263 }
1264
1265 static int merge_mode_and_contents(struct merge_options *opt,
1266                                    const struct diff_filespec *o,
1267                                    const struct diff_filespec *a,
1268                                    const struct diff_filespec *b,
1269                                    const char *filename,
1270                                    const char *branch1,
1271                                    const char *branch2,
1272                                    const int extra_marker_size,
1273                                    struct merge_file_info *result)
1274 {
1275         if (opt->branch1 != branch1) {
1276                 /*
1277                  * It's weird getting a reverse merge with HEAD on the bottom
1278                  * side of the conflict markers and the other branch on the
1279                  * top.  Fix that.
1280                  */
1281                 return merge_mode_and_contents(opt, o, b, a,
1282                                                filename,
1283                                                branch2, branch1,
1284                                                extra_marker_size, result);
1285         }
1286
1287         result->merge = 0;
1288         result->clean = 1;
1289
1290         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
1291                 result->clean = 0;
1292                 if (S_ISREG(a->mode)) {
1293                         result->blob.mode = a->mode;
1294                         oidcpy(&result->blob.oid, &a->oid);
1295                 } else {
1296                         result->blob.mode = b->mode;
1297                         oidcpy(&result->blob.oid, &b->oid);
1298                 }
1299         } else {
1300                 if (!oid_eq(&a->oid, &o->oid) && !oid_eq(&b->oid, &o->oid))
1301                         result->merge = 1;
1302
1303                 /*
1304                  * Merge modes
1305                  */
1306                 if (a->mode == b->mode || a->mode == o->mode)
1307                         result->blob.mode = b->mode;
1308                 else {
1309                         result->blob.mode = a->mode;
1310                         if (b->mode != o->mode) {
1311                                 result->clean = 0;
1312                                 result->merge = 1;
1313                         }
1314                 }
1315
1316                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &o->oid))
1317                         oidcpy(&result->blob.oid, &b->oid);
1318                 else if (oid_eq(&b->oid, &o->oid))
1319                         oidcpy(&result->blob.oid, &a->oid);
1320                 else if (S_ISREG(a->mode)) {
1321                         mmbuffer_t result_buf;
1322                         int ret = 0, merge_status;
1323
1324                         merge_status = merge_3way(opt, &result_buf, o, a, b,
1325                                                   branch1, branch2,
1326                                                   extra_marker_size);
1327
1328                         if ((merge_status < 0) || !result_buf.ptr)
1329                                 ret = err(opt, _("Failed to execute internal merge"));
1330
1331                         if (!ret &&
1332                             write_object_file(result_buf.ptr, result_buf.size,
1333                                               blob_type, &result->blob.oid))
1334                                 ret = err(opt, _("Unable to add %s to database"),
1335                                           a->path);
1336
1337                         free(result_buf.ptr);
1338                         if (ret)
1339                                 return ret;
1340                         result->clean = (merge_status == 0);
1341                 } else if (S_ISGITLINK(a->mode)) {
1342                         result->clean = merge_submodule(opt, &result->blob.oid,
1343                                                         o->path,
1344                                                         &o->oid,
1345                                                         &a->oid,
1346                                                         &b->oid);
1347                 } else if (S_ISLNK(a->mode)) {
1348                         switch (opt->recursive_variant) {
1349                         case MERGE_RECURSIVE_NORMAL:
1350                                 oidcpy(&result->blob.oid, &a->oid);
1351                                 if (!oid_eq(&a->oid, &b->oid))
1352                                         result->clean = 0;
1353                                 break;
1354                         case MERGE_RECURSIVE_OURS:
1355                                 oidcpy(&result->blob.oid, &a->oid);
1356                                 break;
1357                         case MERGE_RECURSIVE_THEIRS:
1358                                 oidcpy(&result->blob.oid, &b->oid);
1359                                 break;
1360                         }
1361                 } else
1362                         BUG("unsupported object type in the tree");
1363         }
1364
1365         if (result->merge)
1366                 output(opt, 2, _("Auto-merging %s"), filename);
1367
1368         return 0;
1369 }
1370
1371 static int handle_rename_via_dir(struct merge_options *opt,
1372                                  struct rename_conflict_info *ci)
1373 {
1374         /*
1375          * Handle file adds that need to be renamed due to directory rename
1376          * detection.  This differs from handle_rename_normal, because
1377          * there is no content merge to do; just move the file into the
1378          * desired final location.
1379          */
1380         const struct rename *ren = ci->ren1;
1381         const struct diff_filespec *dest = ren->pair->two;
1382         char *file_path = dest->path;
1383         int mark_conflicted = (opt->detect_directory_renames == 1);
1384         assert(ren->dir_rename_original_dest);
1385
1386         if (!opt->call_depth && would_lose_untracked(opt, dest->path)) {
1387                 mark_conflicted = 1;
1388                 file_path = unique_path(opt, dest->path, ren->branch);
1389                 output(opt, 1, _("Error: Refusing to lose untracked file at %s; "
1390                                  "writing to %s instead."),
1391                        dest->path, file_path);
1392         }
1393
1394         if (mark_conflicted) {
1395                 /*
1396                  * Write the file in worktree at file_path.  In the index,
1397                  * only record the file at dest->path in the appropriate
1398                  * higher stage.
1399                  */
1400                 if (update_file(opt, 0, dest, file_path))
1401                         return -1;
1402                 if (file_path != dest->path)
1403                         free(file_path);
1404                 if (update_stages(opt, dest->path, NULL,
1405                                   ren->branch == opt->branch1 ? dest : NULL,
1406                                   ren->branch == opt->branch1 ? NULL : dest))
1407                         return -1;
1408                 return 0; /* not clean, but conflicted */
1409         } else {
1410                 /* Update dest->path both in index and in worktree */
1411                 if (update_file(opt, 1, dest, dest->path))
1412                         return -1;
1413                 return 1; /* clean */
1414         }
1415 }
1416
1417 static int handle_change_delete(struct merge_options *opt,
1418                                 const char *path, const char *old_path,
1419                                 const struct diff_filespec *o,
1420                                 const struct diff_filespec *changed,
1421                                 const char *change_branch,
1422                                 const char *delete_branch,
1423                                 const char *change, const char *change_past)
1424 {
1425         char *alt_path = NULL;
1426         const char *update_path = path;
1427         int ret = 0;
1428
1429         if (dir_in_way(opt->repo->index, path, !opt->call_depth, 0) ||
1430             (!opt->call_depth && would_lose_untracked(opt, path))) {
1431                 update_path = alt_path = unique_path(opt, path, change_branch);
1432         }
1433
1434         if (opt->call_depth) {
1435                 /*
1436                  * We cannot arbitrarily accept either a_sha or b_sha as
1437                  * correct; since there is no true "middle point" between
1438                  * them, simply reuse the base version for virtual merge base.
1439                  */
1440                 ret = remove_file_from_index(opt->repo->index, path);
1441                 if (!ret)
1442                         ret = update_file(opt, 0, o, update_path);
1443         } else {
1444                 /*
1445                  * Despite the four nearly duplicate messages and argument
1446                  * lists below and the ugliness of the nested if-statements,
1447                  * having complete messages makes the job easier for
1448                  * translators.
1449                  *
1450                  * The slight variance among the cases is due to the fact
1451                  * that:
1452                  *   1) directory/file conflicts (in effect if
1453                  *      !alt_path) could cause us to need to write the
1454                  *      file to a different path.
1455                  *   2) renames (in effect if !old_path) could mean that
1456                  *      there are two names for the path that the user
1457                  *      may know the file by.
1458                  */
1459                 if (!alt_path) {
1460                         if (!old_path) {
1461                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1462                                        "and %s in %s. Version %s of %s left in tree."),
1463                                        change, path, delete_branch, change_past,
1464                                        change_branch, change_branch, path);
1465                         } else {
1466                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1467                                        "and %s to %s in %s. Version %s of %s left in tree."),
1468                                        change, old_path, delete_branch, change_past, path,
1469                                        change_branch, change_branch, path);
1470                         }
1471                 } else {
1472                         if (!old_path) {
1473                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1474                                        "and %s in %s. Version %s of %s left in tree at %s."),
1475                                        change, path, delete_branch, change_past,
1476                                        change_branch, change_branch, path, alt_path);
1477                         } else {
1478                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1479                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1480                                        change, old_path, delete_branch, change_past, path,
1481                                        change_branch, change_branch, path, alt_path);
1482                         }
1483                 }
1484                 /*
1485                  * No need to call update_file() on path when change_branch ==
1486                  * opt->branch1 && !alt_path, since that would needlessly touch
1487                  * path.  We could call update_file_flags() with update_cache=0
1488                  * and update_wd=0, but that's a no-op.
1489                  */
1490                 if (change_branch != opt->branch1 || alt_path)
1491                         ret = update_file(opt, 0, changed, update_path);
1492         }
1493         free(alt_path);
1494
1495         return ret;
1496 }
1497
1498 static int handle_rename_delete(struct merge_options *opt,
1499                                 struct rename_conflict_info *ci)
1500 {
1501         const struct rename *ren = ci->ren1;
1502         const struct diff_filespec *orig = ren->pair->one;
1503         const struct diff_filespec *dest = ren->pair->two;
1504         const char *rename_branch = ren->branch;
1505         const char *delete_branch = (opt->branch1 == ren->branch ?
1506                                      opt->branch2 : opt->branch1);
1507
1508         if (handle_change_delete(opt,
1509                                  opt->call_depth ? orig->path : dest->path,
1510                                  opt->call_depth ? NULL : orig->path,
1511                                  orig, dest,
1512                                  rename_branch, delete_branch,
1513                                  _("rename"), _("renamed")))
1514                 return -1;
1515
1516         if (opt->call_depth)
1517                 return remove_file_from_index(opt->repo->index, dest->path);
1518         else
1519                 return update_stages(opt, dest->path, NULL,
1520                                      rename_branch == opt->branch1 ? dest : NULL,
1521                                      rename_branch == opt->branch1 ? NULL : dest);
1522 }
1523
1524 static int handle_file_collision(struct merge_options *opt,
1525                                  const char *collide_path,
1526                                  const char *prev_path1,
1527                                  const char *prev_path2,
1528                                  const char *branch1, const char *branch2,
1529                                  struct diff_filespec *a,
1530                                  struct diff_filespec *b)
1531 {
1532         struct merge_file_info mfi;
1533         struct diff_filespec null;
1534         char *alt_path = NULL;
1535         const char *update_path = collide_path;
1536
1537         /*
1538          * It's easiest to get the correct things into stage 2 and 3, and
1539          * to make sure that the content merge puts HEAD before the other
1540          * branch if we just ensure that branch1 == opt->branch1.  So, simply
1541          * flip arguments around if we don't have that.
1542          */
1543         if (branch1 != opt->branch1) {
1544                 return handle_file_collision(opt, collide_path,
1545                                              prev_path2, prev_path1,
1546                                              branch2, branch1,
1547                                              b, a);
1548         }
1549
1550         /*
1551          * In the recursive case, we just opt to undo renames
1552          */
1553         if (opt->call_depth && (prev_path1 || prev_path2)) {
1554                 /* Put first file (a->oid, a->mode) in its original spot */
1555                 if (prev_path1) {
1556                         if (update_file(opt, 1, a, prev_path1))
1557                                 return -1;
1558                 } else {
1559                         if (update_file(opt, 1, a, collide_path))
1560                                 return -1;
1561                 }
1562
1563                 /* Put second file (b->oid, b->mode) in its original spot */
1564                 if (prev_path2) {
1565                         if (update_file(opt, 1, b, prev_path2))
1566                                 return -1;
1567                 } else {
1568                         if (update_file(opt, 1, b, collide_path))
1569                                 return -1;
1570                 }
1571
1572                 /* Don't leave something at collision path if unrenaming both */
1573                 if (prev_path1 && prev_path2)
1574                         remove_file(opt, 1, collide_path, 0);
1575
1576                 return 0;
1577         }
1578
1579         /* Remove rename sources if rename/add or rename/rename(2to1) */
1580         if (prev_path1)
1581                 remove_file(opt, 1, prev_path1,
1582                             opt->call_depth || would_lose_untracked(opt, prev_path1));
1583         if (prev_path2)
1584                 remove_file(opt, 1, prev_path2,
1585                             opt->call_depth || would_lose_untracked(opt, prev_path2));
1586
1587         /*
1588          * Remove the collision path, if it wouldn't cause dirty contents
1589          * or an untracked file to get lost.  We'll either overwrite with
1590          * merged contents, or just write out to differently named files.
1591          */
1592         if (was_dirty(opt, collide_path)) {
1593                 output(opt, 1, _("Refusing to lose dirty file at %s"),
1594                        collide_path);
1595                 update_path = alt_path = unique_path(opt, collide_path, "merged");
1596         } else if (would_lose_untracked(opt, collide_path)) {
1597                 /*
1598                  * Only way we get here is if both renames were from
1599                  * a directory rename AND user had an untracked file
1600                  * at the location where both files end up after the
1601                  * two directory renames.  See testcase 10d of t6043.
1602                  */
1603                 output(opt, 1, _("Refusing to lose untracked file at "
1604                                "%s, even though it's in the way."),
1605                        collide_path);
1606                 update_path = alt_path = unique_path(opt, collide_path, "merged");
1607         } else {
1608                 /*
1609                  * FIXME: It's possible that the two files are identical
1610                  * and that the current working copy happens to match, in
1611                  * which case we are unnecessarily touching the working
1612                  * tree file.  It's not a likely enough scenario that I
1613                  * want to code up the checks for it and a better fix is
1614                  * available if we restructure how unpack_trees() and
1615                  * merge-recursive interoperate anyway, so punting for
1616                  * now...
1617                  */
1618                 remove_file(opt, 0, collide_path, 0);
1619         }
1620
1621         /* Store things in diff_filespecs for functions that need it */
1622         null.path = (char *)collide_path;
1623         oidcpy(&null.oid, &null_oid);
1624         null.mode = 0;
1625
1626         if (merge_mode_and_contents(opt, &null, a, b, collide_path,
1627                                     branch1, branch2, opt->call_depth * 2, &mfi))
1628                 return -1;
1629         mfi.clean &= !alt_path;
1630         if (update_file(opt, mfi.clean, &mfi.blob, update_path))
1631                 return -1;
1632         if (!mfi.clean && !opt->call_depth &&
1633             update_stages(opt, collide_path, NULL, a, b))
1634                 return -1;
1635         free(alt_path);
1636         /*
1637          * FIXME: If both a & b both started with conflicts (only possible
1638          * if they came from a rename/rename(2to1)), but had IDENTICAL
1639          * contents including those conflicts, then in the next line we claim
1640          * it was clean.  If someone cares about this case, we should have the
1641          * caller notify us if we started with conflicts.
1642          */
1643         return mfi.clean;
1644 }
1645
1646 static int handle_rename_add(struct merge_options *opt,
1647                              struct rename_conflict_info *ci)
1648 {
1649         /* a was renamed to c, and a separate c was added. */
1650         struct diff_filespec *a = ci->ren1->pair->one;
1651         struct diff_filespec *c = ci->ren1->pair->two;
1652         char *path = c->path;
1653         char *prev_path_desc;
1654         struct merge_file_info mfi;
1655
1656         const char *rename_branch = ci->ren1->branch;
1657         const char *add_branch = (opt->branch1 == rename_branch ?
1658                                   opt->branch2 : opt->branch1);
1659         int other_stage = (ci->ren1->branch == opt->branch1 ? 3 : 2);
1660
1661         output(opt, 1, _("CONFLICT (rename/add): "
1662                "Rename %s->%s in %s.  Added %s in %s"),
1663                a->path, c->path, rename_branch,
1664                c->path, add_branch);
1665
1666         prev_path_desc = xstrfmt("version of %s from %s", path, a->path);
1667         ci->ren1->src_entry->stages[other_stage].path = a->path;
1668         if (merge_mode_and_contents(opt, a, c,
1669                                     &ci->ren1->src_entry->stages[other_stage],
1670                                     prev_path_desc,
1671                                     opt->branch1, opt->branch2,
1672                                     1 + opt->call_depth * 2, &mfi))
1673                 return -1;
1674         free(prev_path_desc);
1675
1676         ci->ren1->dst_entry->stages[other_stage].path = mfi.blob.path = c->path;
1677         return handle_file_collision(opt,
1678                                      c->path, a->path, NULL,
1679                                      rename_branch, add_branch,
1680                                      &mfi.blob,
1681                                      &ci->ren1->dst_entry->stages[other_stage]);
1682 }
1683
1684 static char *find_path_for_conflict(struct merge_options *opt,
1685                                     const char *path,
1686                                     const char *branch1,
1687                                     const char *branch2)
1688 {
1689         char *new_path = NULL;
1690         if (dir_in_way(opt->repo->index, path, !opt->call_depth, 0)) {
1691                 new_path = unique_path(opt, path, branch1);
1692                 output(opt, 1, _("%s is a directory in %s adding "
1693                                "as %s instead"),
1694                        path, branch2, new_path);
1695         } else if (would_lose_untracked(opt, path)) {
1696                 new_path = unique_path(opt, path, branch1);
1697                 output(opt, 1, _("Refusing to lose untracked file"
1698                                " at %s; adding as %s instead"),
1699                        path, new_path);
1700         }
1701
1702         return new_path;
1703 }
1704
1705 static int handle_rename_rename_1to2(struct merge_options *opt,
1706                                      struct rename_conflict_info *ci)
1707 {
1708         /* One file was renamed in both branches, but to different names. */
1709         struct merge_file_info mfi;
1710         struct diff_filespec *add;
1711         struct diff_filespec *o = ci->ren1->pair->one;
1712         struct diff_filespec *a = ci->ren1->pair->two;
1713         struct diff_filespec *b = ci->ren2->pair->two;
1714         char *path_desc;
1715
1716         output(opt, 1, _("CONFLICT (rename/rename): "
1717                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1718                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1719                o->path, a->path, ci->ren1->branch,
1720                o->path, b->path, ci->ren2->branch,
1721                opt->call_depth ? _(" (left unresolved)") : "");
1722
1723         path_desc = xstrfmt("%s and %s, both renamed from %s",
1724                             a->path, b->path, o->path);
1725         if (merge_mode_and_contents(opt, o, a, b, path_desc,
1726                                     ci->ren1->branch, ci->ren2->branch,
1727                                     opt->call_depth * 2, &mfi))
1728                 return -1;
1729         free(path_desc);
1730
1731         if (opt->call_depth) {
1732                 /*
1733                  * FIXME: For rename/add-source conflicts (if we could detect
1734                  * such), this is wrong.  We should instead find a unique
1735                  * pathname and then either rename the add-source file to that
1736                  * unique path, or use that unique path instead of src here.
1737                  */
1738                 if (update_file(opt, 0, &mfi.blob, o->path))
1739                         return -1;
1740
1741                 /*
1742                  * Above, we put the merged content at the merge-base's
1743                  * path.  Now we usually need to delete both a->path and
1744                  * b->path.  However, the rename on each side of the merge
1745                  * could also be involved in a rename/add conflict.  In
1746                  * such cases, we should keep the added file around,
1747                  * resolving the conflict at that path in its favor.
1748                  */
1749                 add = &ci->ren1->dst_entry->stages[2 ^ 1];
1750                 if (is_valid(add)) {
1751                         if (update_file(opt, 0, add, a->path))
1752                                 return -1;
1753                 }
1754                 else
1755                         remove_file_from_index(opt->repo->index, a->path);
1756                 add = &ci->ren2->dst_entry->stages[3 ^ 1];
1757                 if (is_valid(add)) {
1758                         if (update_file(opt, 0, add, b->path))
1759                                 return -1;
1760                 }
1761                 else
1762                         remove_file_from_index(opt->repo->index, b->path);
1763         } else {
1764                 /*
1765                  * For each destination path, we need to see if there is a
1766                  * rename/add collision.  If not, we can write the file out
1767                  * to the specified location.
1768                  */
1769                 add = &ci->ren1->dst_entry->stages[2 ^ 1];
1770                 if (is_valid(add)) {
1771                         add->path = mfi.blob.path = a->path;
1772                         if (handle_file_collision(opt, a->path,
1773                                                   NULL, NULL,
1774                                                   ci->ren1->branch,
1775                                                   ci->ren2->branch,
1776                                                   &mfi.blob, add) < 0)
1777                                 return -1;
1778                 } else {
1779                         char *new_path = find_path_for_conflict(opt, a->path,
1780                                                                 ci->ren1->branch,
1781                                                                 ci->ren2->branch);
1782                         if (update_file(opt, 0, &mfi.blob,
1783                                         new_path ? new_path : a->path))
1784                                 return -1;
1785                         free(new_path);
1786                         if (update_stages(opt, a->path, NULL, a, NULL))
1787                                 return -1;
1788                 }
1789
1790                 add = &ci->ren2->dst_entry->stages[3 ^ 1];
1791                 if (is_valid(add)) {
1792                         add->path = mfi.blob.path = b->path;
1793                         if (handle_file_collision(opt, b->path,
1794                                                   NULL, NULL,
1795                                                   ci->ren1->branch,
1796                                                   ci->ren2->branch,
1797                                                   add, &mfi.blob) < 0)
1798                                 return -1;
1799                 } else {
1800                         char *new_path = find_path_for_conflict(opt, b->path,
1801                                                                 ci->ren2->branch,
1802                                                                 ci->ren1->branch);
1803                         if (update_file(opt, 0, &mfi.blob,
1804                                         new_path ? new_path : b->path))
1805                                 return -1;
1806                         free(new_path);
1807                         if (update_stages(opt, b->path, NULL, NULL, b))
1808                                 return -1;
1809                 }
1810         }
1811
1812         return 0;
1813 }
1814
1815 static int handle_rename_rename_2to1(struct merge_options *opt,
1816                                      struct rename_conflict_info *ci)
1817 {
1818         /* Two files, a & b, were renamed to the same thing, c. */
1819         struct diff_filespec *a = ci->ren1->pair->one;
1820         struct diff_filespec *b = ci->ren2->pair->one;
1821         struct diff_filespec *c1 = ci->ren1->pair->two;
1822         struct diff_filespec *c2 = ci->ren2->pair->two;
1823         char *path = c1->path; /* == c2->path */
1824         char *path_side_1_desc;
1825         char *path_side_2_desc;
1826         struct merge_file_info mfi_c1;
1827         struct merge_file_info mfi_c2;
1828         int ostage1, ostage2;
1829
1830         output(opt, 1, _("CONFLICT (rename/rename): "
1831                "Rename %s->%s in %s. "
1832                "Rename %s->%s in %s"),
1833                a->path, c1->path, ci->ren1->branch,
1834                b->path, c2->path, ci->ren2->branch);
1835
1836         path_side_1_desc = xstrfmt("version of %s from %s", path, a->path);
1837         path_side_2_desc = xstrfmt("version of %s from %s", path, b->path);
1838         ostage1 = ci->ren1->branch == opt->branch1 ? 3 : 2;
1839         ostage2 = ostage1 ^ 1;
1840         ci->ren1->src_entry->stages[ostage1].path = a->path;
1841         ci->ren2->src_entry->stages[ostage2].path = b->path;
1842         if (merge_mode_and_contents(opt, a, c1,
1843                                     &ci->ren1->src_entry->stages[ostage1],
1844                                     path_side_1_desc,
1845                                     opt->branch1, opt->branch2,
1846                                     1 + opt->call_depth * 2, &mfi_c1) ||
1847             merge_mode_and_contents(opt, b,
1848                                     &ci->ren2->src_entry->stages[ostage2],
1849                                     c2, path_side_2_desc,
1850                                     opt->branch1, opt->branch2,
1851                                     1 + opt->call_depth * 2, &mfi_c2))
1852                 return -1;
1853         free(path_side_1_desc);
1854         free(path_side_2_desc);
1855         mfi_c1.blob.path = path;
1856         mfi_c2.blob.path = path;
1857
1858         return handle_file_collision(opt, path, a->path, b->path,
1859                                      ci->ren1->branch, ci->ren2->branch,
1860                                      &mfi_c1.blob, &mfi_c2.blob);
1861 }
1862
1863 /*
1864  * Get the diff_filepairs changed between o_tree and tree.
1865  */
1866 static struct diff_queue_struct *get_diffpairs(struct merge_options *opt,
1867                                                struct tree *o_tree,
1868                                                struct tree *tree)
1869 {
1870         struct diff_queue_struct *ret;
1871         struct diff_options opts;
1872
1873         repo_diff_setup(opt->repo, &opts);
1874         opts.flags.recursive = 1;
1875         opts.flags.rename_empty = 0;
1876         opts.detect_rename = merge_detect_rename(opt);
1877         /*
1878          * We do not have logic to handle the detection of copies.  In
1879          * fact, it may not even make sense to add such logic: would we
1880          * really want a change to a base file to be propagated through
1881          * multiple other files by a merge?
1882          */
1883         if (opts.detect_rename > DIFF_DETECT_RENAME)
1884                 opts.detect_rename = DIFF_DETECT_RENAME;
1885         opts.rename_limit = opt->merge_rename_limit >= 0 ? opt->merge_rename_limit :
1886                             opt->diff_rename_limit >= 0 ? opt->diff_rename_limit :
1887                             1000;
1888         opts.rename_score = opt->rename_score;
1889         opts.show_rename_progress = opt->show_rename_progress;
1890         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1891         diff_setup_done(&opts);
1892         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1893         diffcore_std(&opts);
1894         if (opts.needed_rename_limit > opt->needed_rename_limit)
1895                 opt->needed_rename_limit = opts.needed_rename_limit;
1896
1897         ret = xmalloc(sizeof(*ret));
1898         *ret = diff_queued_diff;
1899
1900         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1901         diff_queued_diff.nr = 0;
1902         diff_queued_diff.queue = NULL;
1903         diff_flush(&opts);
1904         return ret;
1905 }
1906
1907 static int tree_has_path(struct repository *r, struct tree *tree,
1908                          const char *path)
1909 {
1910         struct object_id hashy;
1911         unsigned short mode_o;
1912
1913         return !get_tree_entry(r,
1914                                &tree->object.oid, path,
1915                                &hashy, &mode_o);
1916 }
1917
1918 /*
1919  * Return a new string that replaces the beginning portion (which matches
1920  * entry->dir), with entry->new_dir.  In perl-speak:
1921  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1922  * NOTE:
1923  *   Caller must ensure that old_path starts with entry->dir + '/'.
1924  */
1925 static char *apply_dir_rename(struct dir_rename_entry *entry,
1926                               const char *old_path)
1927 {
1928         struct strbuf new_path = STRBUF_INIT;
1929         int oldlen, newlen;
1930
1931         if (entry->non_unique_new_dir)
1932                 return NULL;
1933
1934         oldlen = strlen(entry->dir);
1935         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1936         strbuf_grow(&new_path, newlen);
1937         strbuf_addbuf(&new_path, &entry->new_dir);
1938         strbuf_addstr(&new_path, &old_path[oldlen]);
1939
1940         return strbuf_detach(&new_path, NULL);
1941 }
1942
1943 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1944                                     char **old_dir, char **new_dir)
1945 {
1946         char *end_of_old, *end_of_new;
1947         int old_len, new_len;
1948
1949         *old_dir = NULL;
1950         *new_dir = NULL;
1951
1952         /*
1953          * For
1954          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1955          * the "e/foo.c" part is the same, we just want to know that
1956          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1957          * so, for this example, this function returns "a/b/c/d" in
1958          * *old_dir and "a/b/some/thing/else" in *new_dir.
1959          *
1960          * Also, if the basename of the file changed, we don't care.  We
1961          * want to know which portion of the directory, if any, changed.
1962          */
1963         end_of_old = strrchr(old_path, '/');
1964         end_of_new = strrchr(new_path, '/');
1965
1966         if (end_of_old == NULL || end_of_new == NULL)
1967                 return;
1968         while (*--end_of_new == *--end_of_old &&
1969                end_of_old != old_path &&
1970                end_of_new != new_path)
1971                 ; /* Do nothing; all in the while loop */
1972         /*
1973          * We've found the first non-matching character in the directory
1974          * paths.  That means the current directory we were comparing
1975          * represents the rename.  Move end_of_old and end_of_new back
1976          * to the full directory name.
1977          */
1978         if (*end_of_old == '/')
1979                 end_of_old++;
1980         if (*end_of_old != '/')
1981                 end_of_new++;
1982         end_of_old = strchr(end_of_old, '/');
1983         end_of_new = strchr(end_of_new, '/');
1984
1985         /*
1986          * It may have been the case that old_path and new_path were the same
1987          * directory all along.  Don't claim a rename if they're the same.
1988          */
1989         old_len = end_of_old - old_path;
1990         new_len = end_of_new - new_path;
1991
1992         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1993                 *old_dir = xstrndup(old_path, old_len);
1994                 *new_dir = xstrndup(new_path, new_len);
1995         }
1996 }
1997
1998 static void remove_hashmap_entries(struct hashmap *dir_renames,
1999                                    struct string_list *items_to_remove)
2000 {
2001         int i;
2002         struct dir_rename_entry *entry;
2003
2004         for (i = 0; i < items_to_remove->nr; i++) {
2005                 entry = items_to_remove->items[i].util;
2006                 hashmap_remove(dir_renames, &entry->ent, NULL);
2007         }
2008         string_list_clear(items_to_remove, 0);
2009 }
2010
2011 /*
2012  * See if there is a directory rename for path, and if there are any file
2013  * level conflicts for the renamed location.  If there is a rename and
2014  * there are no conflicts, return the new name.  Otherwise, return NULL.
2015  */
2016 static char *handle_path_level_conflicts(struct merge_options *opt,
2017                                          const char *path,
2018                                          struct dir_rename_entry *entry,
2019                                          struct hashmap *collisions,
2020                                          struct tree *tree)
2021 {
2022         char *new_path = NULL;
2023         struct collision_entry *collision_ent;
2024         int clean = 1;
2025         struct strbuf collision_paths = STRBUF_INIT;
2026
2027         /*
2028          * entry has the mapping of old directory name to new directory name
2029          * that we want to apply to path.
2030          */
2031         new_path = apply_dir_rename(entry, path);
2032
2033         if (!new_path) {
2034                 /* This should only happen when entry->non_unique_new_dir set */
2035                 if (!entry->non_unique_new_dir)
2036                         BUG("entry->non_unqiue_dir not set and !new_path");
2037                 output(opt, 1, _("CONFLICT (directory rename split): "
2038                                "Unclear where to place %s because directory "
2039                                "%s was renamed to multiple other directories, "
2040                                "with no destination getting a majority of the "
2041                                "files."),
2042                        path, entry->dir);
2043                 clean = 0;
2044                 return NULL;
2045         }
2046
2047         /*
2048          * The caller needs to have ensured that it has pre-populated
2049          * collisions with all paths that map to new_path.  Do a quick check
2050          * to ensure that's the case.
2051          */
2052         collision_ent = collision_find_entry(collisions, new_path);
2053         if (collision_ent == NULL)
2054                 BUG("collision_ent is NULL");
2055
2056         /*
2057          * Check for one-sided add/add/.../add conflicts, i.e.
2058          * where implicit renames from the other side doing
2059          * directory rename(s) can affect this side of history
2060          * to put multiple paths into the same location.  Warn
2061          * and bail on directory renames for such paths.
2062          */
2063         if (collision_ent->reported_already) {
2064                 clean = 0;
2065         } else if (tree_has_path(opt->repo, tree, new_path)) {
2066                 collision_ent->reported_already = 1;
2067                 strbuf_add_separated_string_list(&collision_paths, ", ",
2068                                                  &collision_ent->source_files);
2069                 output(opt, 1, _("CONFLICT (implicit dir rename): Existing "
2070                                "file/dir at %s in the way of implicit "
2071                                "directory rename(s) putting the following "
2072                                "path(s) there: %s."),
2073                        new_path, collision_paths.buf);
2074                 clean = 0;
2075         } else if (collision_ent->source_files.nr > 1) {
2076                 collision_ent->reported_already = 1;
2077                 strbuf_add_separated_string_list(&collision_paths, ", ",
2078                                                  &collision_ent->source_files);
2079                 output(opt, 1, _("CONFLICT (implicit dir rename): Cannot map "
2080                                "more than one path to %s; implicit directory "
2081                                "renames tried to put these paths there: %s"),
2082                        new_path, collision_paths.buf);
2083                 clean = 0;
2084         }
2085
2086         /* Free memory we no longer need */
2087         strbuf_release(&collision_paths);
2088         if (!clean && new_path) {
2089                 free(new_path);
2090                 return NULL;
2091         }
2092
2093         return new_path;
2094 }
2095
2096 /*
2097  * There are a couple things we want to do at the directory level:
2098  *   1. Check for both sides renaming to the same thing, in order to avoid
2099  *      implicit renaming of files that should be left in place.  (See
2100  *      testcase 6b in t6043 for details.)
2101  *   2. Prune directory renames if there are still files left in the
2102  *      the original directory.  These represent a partial directory rename,
2103  *      i.e. a rename where only some of the files within the directory
2104  *      were renamed elsewhere.  (Technically, this could be done earlier
2105  *      in get_directory_renames(), except that would prevent us from
2106  *      doing the previous check and thus failing testcase 6b.)
2107  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2108  *      In the future, we could potentially record this info as well and
2109  *      omit reporting rename/rename(1to2) conflicts for each path within
2110  *      the affected directories, thus cleaning up the merge output.
2111  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2112  *         directory level, because merging directories is fine.  If it
2113  *         causes conflicts for files within those merged directories, then
2114  *         that should be detected at the individual path level.
2115  */
2116 static void handle_directory_level_conflicts(struct merge_options *opt,
2117                                              struct hashmap *dir_re_head,
2118                                              struct tree *head,
2119                                              struct hashmap *dir_re_merge,
2120                                              struct tree *merge)
2121 {
2122         struct hashmap_iter iter;
2123         struct dir_rename_entry *head_ent;
2124         struct dir_rename_entry *merge_ent;
2125
2126         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2127         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2128
2129         hashmap_iter_init(dir_re_head, &iter);
2130         while ((head_ent = hashmap_iter_next(&iter))) {
2131                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2132                 if (merge_ent &&
2133                     !head_ent->non_unique_new_dir &&
2134                     !merge_ent->non_unique_new_dir &&
2135                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2136                         /* 1. Renamed identically; remove it from both sides */
2137                         string_list_append(&remove_from_head,
2138                                            head_ent->dir)->util = head_ent;
2139                         strbuf_release(&head_ent->new_dir);
2140                         string_list_append(&remove_from_merge,
2141                                            merge_ent->dir)->util = merge_ent;
2142                         strbuf_release(&merge_ent->new_dir);
2143                 } else if (tree_has_path(opt->repo, head, head_ent->dir)) {
2144                         /* 2. This wasn't a directory rename after all */
2145                         string_list_append(&remove_from_head,
2146                                            head_ent->dir)->util = head_ent;
2147                         strbuf_release(&head_ent->new_dir);
2148                 }
2149         }
2150
2151         remove_hashmap_entries(dir_re_head, &remove_from_head);
2152         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2153
2154         hashmap_iter_init(dir_re_merge, &iter);
2155         while ((merge_ent = hashmap_iter_next(&iter))) {
2156                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2157                 if (tree_has_path(opt->repo, merge, merge_ent->dir)) {
2158                         /* 2. This wasn't a directory rename after all */
2159                         string_list_append(&remove_from_merge,
2160                                            merge_ent->dir)->util = merge_ent;
2161                 } else if (head_ent &&
2162                            !head_ent->non_unique_new_dir &&
2163                            !merge_ent->non_unique_new_dir) {
2164                         /* 3. rename/rename(1to2) */
2165                         /*
2166                          * We can assume it's not rename/rename(1to1) because
2167                          * that was case (1), already checked above.  So we
2168                          * know that head_ent->new_dir and merge_ent->new_dir
2169                          * are different strings.
2170                          */
2171                         output(opt, 1, _("CONFLICT (rename/rename): "
2172                                        "Rename directory %s->%s in %s. "
2173                                        "Rename directory %s->%s in %s"),
2174                                head_ent->dir, head_ent->new_dir.buf, opt->branch1,
2175                                head_ent->dir, merge_ent->new_dir.buf, opt->branch2);
2176                         string_list_append(&remove_from_head,
2177                                            head_ent->dir)->util = head_ent;
2178                         strbuf_release(&head_ent->new_dir);
2179                         string_list_append(&remove_from_merge,
2180                                            merge_ent->dir)->util = merge_ent;
2181                         strbuf_release(&merge_ent->new_dir);
2182                 }
2183         }
2184
2185         remove_hashmap_entries(dir_re_head, &remove_from_head);
2186         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2187 }
2188
2189 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs)
2190 {
2191         struct hashmap *dir_renames;
2192         struct hashmap_iter iter;
2193         struct dir_rename_entry *entry;
2194         int i;
2195
2196         /*
2197          * Typically, we think of a directory rename as all files from a
2198          * certain directory being moved to a target directory.  However,
2199          * what if someone first moved two files from the original
2200          * directory in one commit, and then renamed the directory
2201          * somewhere else in a later commit?  At merge time, we just know
2202          * that files from the original directory went to two different
2203          * places, and that the bulk of them ended up in the same place.
2204          * We want each directory rename to represent where the bulk of the
2205          * files from that directory end up; this function exists to find
2206          * where the bulk of the files went.
2207          *
2208          * The first loop below simply iterates through the list of file
2209          * renames, finding out how often each directory rename pair
2210          * possibility occurs.
2211          */
2212         dir_renames = xmalloc(sizeof(*dir_renames));
2213         dir_rename_init(dir_renames);
2214         for (i = 0; i < pairs->nr; ++i) {
2215                 struct string_list_item *item;
2216                 int *count;
2217                 struct diff_filepair *pair = pairs->queue[i];
2218                 char *old_dir, *new_dir;
2219
2220                 /* File not part of directory rename if it wasn't renamed */
2221                 if (pair->status != 'R')
2222                         continue;
2223
2224                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2225                                         &old_dir,        &new_dir);
2226                 if (!old_dir)
2227                         /* Directory didn't change at all; ignore this one. */
2228                         continue;
2229
2230                 entry = dir_rename_find_entry(dir_renames, old_dir);
2231                 if (!entry) {
2232                         entry = xmalloc(sizeof(*entry));
2233                         dir_rename_entry_init(entry, old_dir);
2234                         hashmap_put(dir_renames, &entry->ent);
2235                 } else {
2236                         free(old_dir);
2237                 }
2238                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2239                 if (!item) {
2240                         item = string_list_insert(&entry->possible_new_dirs,
2241                                                   new_dir);
2242                         item->util = xcalloc(1, sizeof(int));
2243                 } else {
2244                         free(new_dir);
2245                 }
2246                 count = item->util;
2247                 *count += 1;
2248         }
2249
2250         /*
2251          * For each directory with files moved out of it, we find out which
2252          * target directory received the most files so we can declare it to
2253          * be the "winning" target location for the directory rename.  This
2254          * winner gets recorded in new_dir.  If there is no winner
2255          * (multiple target directories received the same number of files),
2256          * we set non_unique_new_dir.  Once we've determined the winner (or
2257          * that there is no winner), we no longer need possible_new_dirs.
2258          */
2259         hashmap_iter_init(dir_renames, &iter);
2260         while ((entry = hashmap_iter_next(&iter))) {
2261                 int max = 0;
2262                 int bad_max = 0;
2263                 char *best = NULL;
2264
2265                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2266                         int *count = entry->possible_new_dirs.items[i].util;
2267
2268                         if (*count == max)
2269                                 bad_max = max;
2270                         else if (*count > max) {
2271                                 max = *count;
2272                                 best = entry->possible_new_dirs.items[i].string;
2273                         }
2274                 }
2275                 if (bad_max == max)
2276                         entry->non_unique_new_dir = 1;
2277                 else {
2278                         assert(entry->new_dir.len == 0);
2279                         strbuf_addstr(&entry->new_dir, best);
2280                 }
2281                 /*
2282                  * The relevant directory sub-portion of the original full
2283                  * filepaths were xstrndup'ed before inserting into
2284                  * possible_new_dirs, and instead of manually iterating the
2285                  * list and free'ing each, just lie and tell
2286                  * possible_new_dirs that it did the strdup'ing so that it
2287                  * will free them for us.
2288                  */
2289                 entry->possible_new_dirs.strdup_strings = 1;
2290                 string_list_clear(&entry->possible_new_dirs, 1);
2291         }
2292
2293         return dir_renames;
2294 }
2295
2296 static struct dir_rename_entry *check_dir_renamed(const char *path,
2297                                                   struct hashmap *dir_renames)
2298 {
2299         char *temp = xstrdup(path);
2300         char *end;
2301         struct dir_rename_entry *entry = NULL;
2302
2303         while ((end = strrchr(temp, '/'))) {
2304                 *end = '\0';
2305                 entry = dir_rename_find_entry(dir_renames, temp);
2306                 if (entry)
2307                         break;
2308         }
2309         free(temp);
2310         return entry;
2311 }
2312
2313 static void compute_collisions(struct hashmap *collisions,
2314                                struct hashmap *dir_renames,
2315                                struct diff_queue_struct *pairs)
2316 {
2317         int i;
2318
2319         /*
2320          * Multiple files can be mapped to the same path due to directory
2321          * renames done by the other side of history.  Since that other
2322          * side of history could have merged multiple directories into one,
2323          * if our side of history added the same file basename to each of
2324          * those directories, then all N of them would get implicitly
2325          * renamed by the directory rename detection into the same path,
2326          * and we'd get an add/add/.../add conflict, and all those adds
2327          * from *this* side of history.  This is not representable in the
2328          * index, and users aren't going to easily be able to make sense of
2329          * it.  So we need to provide a good warning about what's
2330          * happening, and fall back to no-directory-rename detection
2331          * behavior for those paths.
2332          *
2333          * See testcases 9e and all of section 5 from t6043 for examples.
2334          */
2335         collision_init(collisions);
2336
2337         for (i = 0; i < pairs->nr; ++i) {
2338                 struct dir_rename_entry *dir_rename_ent;
2339                 struct collision_entry *collision_ent;
2340                 char *new_path;
2341                 struct diff_filepair *pair = pairs->queue[i];
2342
2343                 if (pair->status != 'A' && pair->status != 'R')
2344                         continue;
2345                 dir_rename_ent = check_dir_renamed(pair->two->path,
2346                                                    dir_renames);
2347                 if (!dir_rename_ent)
2348                         continue;
2349
2350                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2351                 if (!new_path)
2352                         /*
2353                          * dir_rename_ent->non_unique_new_path is true, which
2354                          * means there is no directory rename for us to use,
2355                          * which means it won't cause us any additional
2356                          * collisions.
2357                          */
2358                         continue;
2359                 collision_ent = collision_find_entry(collisions, new_path);
2360                 if (!collision_ent) {
2361                         collision_ent = xcalloc(1,
2362                                                 sizeof(struct collision_entry));
2363                         hashmap_entry_init(&collision_ent->ent,
2364                                                 strhash(new_path));
2365                         hashmap_put(collisions, &collision_ent->ent);
2366                         collision_ent->target_file = new_path;
2367                 } else {
2368                         free(new_path);
2369                 }
2370                 string_list_insert(&collision_ent->source_files,
2371                                    pair->two->path);
2372         }
2373 }
2374
2375 static char *check_for_directory_rename(struct merge_options *opt,
2376                                         const char *path,
2377                                         struct tree *tree,
2378                                         struct hashmap *dir_renames,
2379                                         struct hashmap *dir_rename_exclusions,
2380                                         struct hashmap *collisions,
2381                                         int *clean_merge)
2382 {
2383         char *new_path = NULL;
2384         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2385         struct dir_rename_entry *oentry = NULL;
2386
2387         if (!entry)
2388                 return new_path;
2389
2390         /*
2391          * This next part is a little weird.  We do not want to do an
2392          * implicit rename into a directory we renamed on our side, because
2393          * that will result in a spurious rename/rename(1to2) conflict.  An
2394          * example:
2395          *   Base commit: dumbdir/afile, otherdir/bfile
2396          *   Side 1:      smrtdir/afile, otherdir/bfile
2397          *   Side 2:      dumbdir/afile, dumbdir/bfile
2398          * Here, while working on Side 1, we could notice that otherdir was
2399          * renamed/merged to dumbdir, and change the diff_filepair for
2400          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2401          * 2 will notice the rename from dumbdir to smrtdir, and do the
2402          * transitive rename to move it from dumbdir/bfile to
2403          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2404          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2405          * the file to end up in smrtdir.  And the way to achieve that is
2406          * to not let Side1 do the rename to dumbdir, since we know that is
2407          * the source of one of our directory renames.
2408          *
2409          * That's why oentry and dir_rename_exclusions is here.
2410          *
2411          * As it turns out, this also prevents N-way transient rename
2412          * confusion; See testcases 9c and 9d of t6043.
2413          */
2414         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2415         if (oentry) {
2416                 output(opt, 1, _("WARNING: Avoiding applying %s -> %s rename "
2417                                "to %s, because %s itself was renamed."),
2418                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2419         } else {
2420                 new_path = handle_path_level_conflicts(opt, path, entry,
2421                                                        collisions, tree);
2422                 *clean_merge &= (new_path != NULL);
2423         }
2424
2425         return new_path;
2426 }
2427
2428 static void apply_directory_rename_modifications(struct merge_options *opt,
2429                                                  struct diff_filepair *pair,
2430                                                  char *new_path,
2431                                                  struct rename *re,
2432                                                  struct tree *tree,
2433                                                  struct tree *o_tree,
2434                                                  struct tree *a_tree,
2435                                                  struct tree *b_tree,
2436                                                  struct string_list *entries)
2437 {
2438         struct string_list_item *item;
2439         int stage = (tree == a_tree ? 2 : 3);
2440         int update_wd;
2441
2442         /*
2443          * In all cases where we can do directory rename detection,
2444          * unpack_trees() will have read pair->two->path into the
2445          * index and the working copy.  We need to remove it so that
2446          * we can instead place it at new_path.  It is guaranteed to
2447          * not be untracked (unpack_trees() would have errored out
2448          * saying the file would have been overwritten), but it might
2449          * be dirty, though.
2450          */
2451         update_wd = !was_dirty(opt, pair->two->path);
2452         if (!update_wd)
2453                 output(opt, 1, _("Refusing to lose dirty file at %s"),
2454                        pair->two->path);
2455         remove_file(opt, 1, pair->two->path, !update_wd);
2456
2457         /* Find or create a new re->dst_entry */
2458         item = string_list_lookup(entries, new_path);
2459         if (item) {
2460                 /*
2461                  * Since we're renaming on this side of history, and it's
2462                  * due to a directory rename on the other side of history
2463                  * (which we only allow when the directory in question no
2464                  * longer exists on the other side of history), the
2465                  * original entry for re->dst_entry is no longer
2466                  * necessary...
2467                  */
2468                 re->dst_entry->processed = 1;
2469
2470                 /*
2471                  * ...because we'll be using this new one.
2472                  */
2473                 re->dst_entry = item->util;
2474         } else {
2475                 /*
2476                  * re->dst_entry is for the before-dir-rename path, and we
2477                  * need it to hold information for the after-dir-rename
2478                  * path.  Before creating a new entry, we need to mark the
2479                  * old one as unnecessary (...unless it is shared by
2480                  * src_entry, i.e. this didn't use to be a rename, in which
2481                  * case we can just allow the normal processing to happen
2482                  * for it).
2483                  */
2484                 if (pair->status == 'R')
2485                         re->dst_entry->processed = 1;
2486
2487                 re->dst_entry = insert_stage_data(opt->repo, new_path,
2488                                                   o_tree, a_tree, b_tree,
2489                                                   entries);
2490                 item = string_list_insert(entries, new_path);
2491                 item->util = re->dst_entry;
2492         }
2493
2494         /*
2495          * Update the stage_data with the information about the path we are
2496          * moving into place.  That slot will be empty and available for us
2497          * to write to because of the collision checks in
2498          * handle_path_level_conflicts().  In other words,
2499          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2500          * open for us to write to.
2501          *
2502          * It may be tempting to actually update the index at this point as
2503          * well, using update_stages_for_stage_data(), but as per the big
2504          * "NOTE" in update_stages(), doing so will modify the current
2505          * in-memory index which will break calls to would_lose_untracked()
2506          * that we need to make.  Instead, we need to just make sure that
2507          * the various handle_rename_*() functions update the index
2508          * explicitly rather than relying on unpack_trees() to have done it.
2509          */
2510         get_tree_entry(opt->repo,
2511                        &tree->object.oid,
2512                        pair->two->path,
2513                        &re->dst_entry->stages[stage].oid,
2514                        &re->dst_entry->stages[stage].mode);
2515
2516         /*
2517          * Record the original change status (or 'type' of change).  If it
2518          * was originally an add ('A'), this lets us differentiate later
2519          * between a RENAME_DELETE conflict and RENAME_VIA_DIR (they
2520          * otherwise look the same).  If it was originally a rename ('R'),
2521          * this lets us remember and report accurately about the transitive
2522          * renaming that occurred via the directory rename detection.  Also,
2523          * record the original destination name.
2524          */
2525         re->dir_rename_original_type = pair->status;
2526         re->dir_rename_original_dest = pair->two->path;
2527
2528         /*
2529          * We don't actually look at pair->status again, but it seems
2530          * pedagogically correct to adjust it.
2531          */
2532         pair->status = 'R';
2533
2534         /*
2535          * Finally, record the new location.
2536          */
2537         pair->two->path = new_path;
2538 }
2539
2540 /*
2541  * Get information of all renames which occurred in 'pairs', making use of
2542  * any implicit directory renames inferred from the other side of history.
2543  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2544  * to be able to associate the correct cache entries with the rename
2545  * information; tree is always equal to either a_tree or b_tree.
2546  */
2547 static struct string_list *get_renames(struct merge_options *opt,
2548                                        const char *branch,
2549                                        struct diff_queue_struct *pairs,
2550                                        struct hashmap *dir_renames,
2551                                        struct hashmap *dir_rename_exclusions,
2552                                        struct tree *tree,
2553                                        struct tree *o_tree,
2554                                        struct tree *a_tree,
2555                                        struct tree *b_tree,
2556                                        struct string_list *entries,
2557                                        int *clean_merge)
2558 {
2559         int i;
2560         struct hashmap collisions;
2561         struct hashmap_iter iter;
2562         struct collision_entry *e;
2563         struct string_list *renames;
2564
2565         compute_collisions(&collisions, dir_renames, pairs);
2566         renames = xcalloc(1, sizeof(struct string_list));
2567
2568         for (i = 0; i < pairs->nr; ++i) {
2569                 struct string_list_item *item;
2570                 struct rename *re;
2571                 struct diff_filepair *pair = pairs->queue[i];
2572                 char *new_path; /* non-NULL only with directory renames */
2573
2574                 if (pair->status != 'A' && pair->status != 'R') {
2575                         diff_free_filepair(pair);
2576                         continue;
2577                 }
2578                 new_path = check_for_directory_rename(opt, pair->two->path, tree,
2579                                                       dir_renames,
2580                                                       dir_rename_exclusions,
2581                                                       &collisions,
2582                                                       clean_merge);
2583                 if (pair->status != 'R' && !new_path) {
2584                         diff_free_filepair(pair);
2585                         continue;
2586                 }
2587
2588                 re = xmalloc(sizeof(*re));
2589                 re->processed = 0;
2590                 re->pair = pair;
2591                 re->branch = branch;
2592                 re->dir_rename_original_type = '\0';
2593                 re->dir_rename_original_dest = NULL;
2594                 item = string_list_lookup(entries, re->pair->one->path);
2595                 if (!item)
2596                         re->src_entry = insert_stage_data(opt->repo,
2597                                         re->pair->one->path,
2598                                         o_tree, a_tree, b_tree, entries);
2599                 else
2600                         re->src_entry = item->util;
2601
2602                 item = string_list_lookup(entries, re->pair->two->path);
2603                 if (!item)
2604                         re->dst_entry = insert_stage_data(opt->repo,
2605                                         re->pair->two->path,
2606                                         o_tree, a_tree, b_tree, entries);
2607                 else
2608                         re->dst_entry = item->util;
2609                 item = string_list_insert(renames, pair->one->path);
2610                 item->util = re;
2611                 if (new_path)
2612                         apply_directory_rename_modifications(opt, pair, new_path,
2613                                                              re, tree, o_tree,
2614                                                              a_tree, b_tree,
2615                                                              entries);
2616         }
2617
2618         hashmap_iter_init(&collisions, &iter);
2619         while ((e = hashmap_iter_next(&iter))) {
2620                 free(e->target_file);
2621                 string_list_clear(&e->source_files, 0);
2622         }
2623         hashmap_free(&collisions, 1);
2624         return renames;
2625 }
2626
2627 static int process_renames(struct merge_options *opt,
2628                            struct string_list *a_renames,
2629                            struct string_list *b_renames)
2630 {
2631         int clean_merge = 1, i, j;
2632         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2633         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2634         const struct rename *sre;
2635
2636         for (i = 0; i < a_renames->nr; i++) {
2637                 sre = a_renames->items[i].util;
2638                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2639                         = (void *)sre;
2640         }
2641         for (i = 0; i < b_renames->nr; i++) {
2642                 sre = b_renames->items[i].util;
2643                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2644                         = (void *)sre;
2645         }
2646
2647         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2648                 struct string_list *renames1, *renames2Dst;
2649                 struct rename *ren1 = NULL, *ren2 = NULL;
2650                 const char *ren1_src, *ren1_dst;
2651                 struct string_list_item *lookup;
2652
2653                 if (i >= a_renames->nr) {
2654                         ren2 = b_renames->items[j++].util;
2655                 } else if (j >= b_renames->nr) {
2656                         ren1 = a_renames->items[i++].util;
2657                 } else {
2658                         int compare = strcmp(a_renames->items[i].string,
2659                                              b_renames->items[j].string);
2660                         if (compare <= 0)
2661                                 ren1 = a_renames->items[i++].util;
2662                         if (compare >= 0)
2663                                 ren2 = b_renames->items[j++].util;
2664                 }
2665
2666                 /* TODO: refactor, so that 1/2 are not needed */
2667                 if (ren1) {
2668                         renames1 = a_renames;
2669                         renames2Dst = &b_by_dst;
2670                 } else {
2671                         renames1 = b_renames;
2672                         renames2Dst = &a_by_dst;
2673                         SWAP(ren2, ren1);
2674                 }
2675
2676                 if (ren1->processed)
2677                         continue;
2678                 ren1->processed = 1;
2679                 ren1->dst_entry->processed = 1;
2680                 /* BUG: We should only mark src_entry as processed if we
2681                  * are not dealing with a rename + add-source case.
2682                  */
2683                 ren1->src_entry->processed = 1;
2684
2685                 ren1_src = ren1->pair->one->path;
2686                 ren1_dst = ren1->pair->two->path;
2687
2688                 if (ren2) {
2689                         /* One file renamed on both sides */
2690                         const char *ren2_src = ren2->pair->one->path;
2691                         const char *ren2_dst = ren2->pair->two->path;
2692                         enum rename_type rename_type;
2693                         if (strcmp(ren1_src, ren2_src) != 0)
2694                                 BUG("ren1_src != ren2_src");
2695                         ren2->dst_entry->processed = 1;
2696                         ren2->processed = 1;
2697                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2698                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2699                                 clean_merge = 0;
2700                         } else {
2701                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2702                                 /* BUG: We should only remove ren1_src in
2703                                  * the base stage (think of rename +
2704                                  * add-source cases).
2705                                  */
2706                                 remove_file(opt, 1, ren1_src, 1);
2707                                 update_entry(ren1->dst_entry,
2708                                              ren1->pair->one,
2709                                              ren1->pair->two,
2710                                              ren2->pair->two);
2711                         }
2712                         setup_rename_conflict_info(rename_type, opt, ren1, ren2);
2713                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2714                         /* Two different files renamed to the same thing */
2715                         char *ren2_dst;
2716                         ren2 = lookup->util;
2717                         ren2_dst = ren2->pair->two->path;
2718                         if (strcmp(ren1_dst, ren2_dst) != 0)
2719                                 BUG("ren1_dst != ren2_dst");
2720
2721                         clean_merge = 0;
2722                         ren2->processed = 1;
2723                         /*
2724                          * BUG: We should only mark src_entry as processed
2725                          * if we are not dealing with a rename + add-source
2726                          * case.
2727                          */
2728                         ren2->src_entry->processed = 1;
2729
2730                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2731                                                    opt, ren1, ren2);
2732                 } else {
2733                         /* Renamed in 1, maybe changed in 2 */
2734                         /* we only use sha1 and mode of these */
2735                         struct diff_filespec src_other, dst_other;
2736                         int try_merge;
2737
2738                         /*
2739                          * unpack_trees loads entries from common-commit
2740                          * into stage 1, from head-commit into stage 2, and
2741                          * from merge-commit into stage 3.  We keep track
2742                          * of which side corresponds to the rename.
2743                          */
2744                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2745                         int other_stage =   a_renames == renames1 ? 3 : 2;
2746
2747                         /* BUG: We should only remove ren1_src in the base
2748                          * stage and in other_stage (think of rename +
2749                          * add-source case).
2750                          */
2751                         remove_file(opt, 1, ren1_src,
2752                                     renamed_stage == 2 || !was_tracked(opt, ren1_src));
2753
2754                         oidcpy(&src_other.oid,
2755                                &ren1->src_entry->stages[other_stage].oid);
2756                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2757                         oidcpy(&dst_other.oid,
2758                                &ren1->dst_entry->stages[other_stage].oid);
2759                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2760                         try_merge = 0;
2761
2762                         if (oid_eq(&src_other.oid, &null_oid) &&
2763                             ren1->dir_rename_original_type == 'A') {
2764                                 setup_rename_conflict_info(RENAME_VIA_DIR,
2765                                                            opt, ren1, NULL);
2766                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2767                                 setup_rename_conflict_info(RENAME_DELETE,
2768                                                            opt, ren1, NULL);
2769                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2770                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2771                                 /*
2772                                  * Added file on the other side identical to
2773                                  * the file being renamed: clean merge.
2774                                  * Also, there is no need to overwrite the
2775                                  * file already in the working copy, so call
2776                                  * update_file_flags() instead of
2777                                  * update_file().
2778                                  */
2779                                 if (update_file_flags(opt,
2780                                                       ren1->pair->two,
2781                                                       ren1_dst,
2782                                                       1, /* update_cache */
2783                                                       0  /* update_wd    */))
2784                                         clean_merge = -1;
2785                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2786                                 /*
2787                                  * Probably not a clean merge, but it's
2788                                  * premature to set clean_merge to 0 here,
2789                                  * because if the rename merges cleanly and
2790                                  * the merge exactly matches the newly added
2791                                  * file, then the merge will be clean.
2792                                  */
2793                                 setup_rename_conflict_info(RENAME_ADD,
2794                                                            opt, ren1, NULL);
2795                         } else
2796                                 try_merge = 1;
2797
2798                         if (clean_merge < 0)
2799                                 goto cleanup_and_return;
2800                         if (try_merge) {
2801                                 struct diff_filespec *o, *a, *b;
2802                                 src_other.path = (char *)ren1_src;
2803
2804                                 o = ren1->pair->one;
2805                                 if (a_renames == renames1) {
2806                                         a = ren1->pair->two;
2807                                         b = &src_other;
2808                                 } else {
2809                                         b = ren1->pair->two;
2810                                         a = &src_other;
2811                                 }
2812                                 update_entry(ren1->dst_entry, o, a, b);
2813                                 setup_rename_conflict_info(RENAME_NORMAL,
2814                                                            opt, ren1, NULL);
2815                         }
2816                 }
2817         }
2818 cleanup_and_return:
2819         string_list_clear(&a_by_dst, 0);
2820         string_list_clear(&b_by_dst, 0);
2821
2822         return clean_merge;
2823 }
2824
2825 struct rename_info {
2826         struct string_list *head_renames;
2827         struct string_list *merge_renames;
2828 };
2829
2830 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2831                                    struct hashmap *dir_renames)
2832 {
2833         struct hashmap_iter iter;
2834         struct dir_rename_entry *e;
2835
2836         hashmap_iter_init(dir_renames, &iter);
2837         while ((e = hashmap_iter_next(&iter))) {
2838                 free(e->dir);
2839                 strbuf_release(&e->new_dir);
2840                 /* possible_new_dirs already cleared in get_directory_renames */
2841         }
2842         hashmap_free(dir_renames, 1);
2843         free(dir_renames);
2844
2845         free(pairs->queue);
2846         free(pairs);
2847 }
2848
2849 static int detect_and_process_renames(struct merge_options *opt,
2850                                       struct tree *common,
2851                                       struct tree *head,
2852                                       struct tree *merge,
2853                                       struct string_list *entries,
2854                                       struct rename_info *ri)
2855 {
2856         struct diff_queue_struct *head_pairs, *merge_pairs;
2857         struct hashmap *dir_re_head, *dir_re_merge;
2858         int clean = 1;
2859
2860         ri->head_renames = NULL;
2861         ri->merge_renames = NULL;
2862
2863         if (!merge_detect_rename(opt))
2864                 return 1;
2865
2866         head_pairs = get_diffpairs(opt, common, head);
2867         merge_pairs = get_diffpairs(opt, common, merge);
2868
2869         if ((opt->detect_directory_renames == 2) ||
2870             (opt->detect_directory_renames == 1 && !opt->call_depth)) {
2871                 dir_re_head = get_directory_renames(head_pairs);
2872                 dir_re_merge = get_directory_renames(merge_pairs);
2873
2874                 handle_directory_level_conflicts(opt,
2875                                                  dir_re_head, head,
2876                                                  dir_re_merge, merge);
2877         } else {
2878                 dir_re_head  = xmalloc(sizeof(*dir_re_head));
2879                 dir_re_merge = xmalloc(sizeof(*dir_re_merge));
2880                 dir_rename_init(dir_re_head);
2881                 dir_rename_init(dir_re_merge);
2882         }
2883
2884         ri->head_renames  = get_renames(opt, opt->branch1, head_pairs,
2885                                         dir_re_merge, dir_re_head, head,
2886                                         common, head, merge, entries,
2887                                         &clean);
2888         if (clean < 0)
2889                 goto cleanup;
2890         ri->merge_renames = get_renames(opt, opt->branch2, merge_pairs,
2891                                         dir_re_head, dir_re_merge, merge,
2892                                         common, head, merge, entries,
2893                                         &clean);
2894         if (clean < 0)
2895                 goto cleanup;
2896         clean &= process_renames(opt, ri->head_renames, ri->merge_renames);
2897
2898 cleanup:
2899         /*
2900          * Some cleanup is deferred until cleanup_renames() because the
2901          * data structures are still needed and referenced in
2902          * process_entry().  But there are a few things we can free now.
2903          */
2904         initial_cleanup_rename(head_pairs, dir_re_head);
2905         initial_cleanup_rename(merge_pairs, dir_re_merge);
2906
2907         return clean;
2908 }
2909
2910 static void final_cleanup_rename(struct string_list *rename)
2911 {
2912         const struct rename *re;
2913         int i;
2914
2915         if (rename == NULL)
2916                 return;
2917
2918         for (i = 0; i < rename->nr; i++) {
2919                 re = rename->items[i].util;
2920                 diff_free_filepair(re->pair);
2921         }
2922         string_list_clear(rename, 1);
2923         free(rename);
2924 }
2925
2926 static void final_cleanup_renames(struct rename_info *re_info)
2927 {
2928         final_cleanup_rename(re_info->head_renames);
2929         final_cleanup_rename(re_info->merge_renames);
2930 }
2931
2932 static int read_oid_strbuf(struct merge_options *opt,
2933                            const struct object_id *oid,
2934                            struct strbuf *dst)
2935 {
2936         void *buf;
2937         enum object_type type;
2938         unsigned long size;
2939         buf = read_object_file(oid, &type, &size);
2940         if (!buf)
2941                 return err(opt, _("cannot read object %s"), oid_to_hex(oid));
2942         if (type != OBJ_BLOB) {
2943                 free(buf);
2944                 return err(opt, _("object %s is not a blob"), oid_to_hex(oid));
2945         }
2946         strbuf_attach(dst, buf, size, size + 1);
2947         return 0;
2948 }
2949
2950 static int blob_unchanged(struct merge_options *opt,
2951                           const struct diff_filespec *o,
2952                           const struct diff_filespec *a,
2953                           int renormalize, const char *path)
2954 {
2955         struct strbuf obuf = STRBUF_INIT;
2956         struct strbuf abuf = STRBUF_INIT;
2957         int ret = 0; /* assume changed for safety */
2958         const struct index_state *idx = opt->repo->index;
2959
2960         if (a->mode != o->mode)
2961                 return 0;
2962         if (oid_eq(&o->oid, &a->oid))
2963                 return 1;
2964         if (!renormalize)
2965                 return 0;
2966
2967         if (read_oid_strbuf(opt, &o->oid, &obuf) ||
2968             read_oid_strbuf(opt, &a->oid, &abuf))
2969                 goto error_return;
2970         /*
2971          * Note: binary | is used so that both renormalizations are
2972          * performed.  Comparison can be skipped if both files are
2973          * unchanged since their sha1s have already been compared.
2974          */
2975         if (renormalize_buffer(idx, path, obuf.buf, obuf.len, &obuf) |
2976             renormalize_buffer(idx, path, abuf.buf, abuf.len, &abuf))
2977                 ret = (obuf.len == abuf.len && !memcmp(obuf.buf, abuf.buf, obuf.len));
2978
2979 error_return:
2980         strbuf_release(&obuf);
2981         strbuf_release(&abuf);
2982         return ret;
2983 }
2984
2985 static int handle_modify_delete(struct merge_options *opt,
2986                                 const char *path,
2987                                 const struct diff_filespec *o,
2988                                 const struct diff_filespec *a,
2989                                 const struct diff_filespec *b)
2990 {
2991         const char *modify_branch, *delete_branch;
2992         const struct diff_filespec *changed;
2993
2994         if (is_valid(a)) {
2995                 modify_branch = opt->branch1;
2996                 delete_branch = opt->branch2;
2997                 changed = a;
2998         } else {
2999                 modify_branch = opt->branch2;
3000                 delete_branch = opt->branch1;
3001                 changed = b;
3002         }
3003
3004         return handle_change_delete(opt,
3005                                     path, NULL,
3006                                     o, changed,
3007                                     modify_branch, delete_branch,
3008                                     _("modify"), _("modified"));
3009 }
3010
3011 static int handle_content_merge(struct merge_file_info *mfi,
3012                                 struct merge_options *opt,
3013                                 const char *path,
3014                                 int is_dirty,
3015                                 const struct diff_filespec *o,
3016                                 const struct diff_filespec *a,
3017                                 const struct diff_filespec *b,
3018                                 struct rename_conflict_info *ci)
3019 {
3020         const char *reason = _("content");
3021         unsigned df_conflict_remains = 0;
3022
3023         if (!is_valid(o))
3024                 reason = _("add/add");
3025
3026         assert(o->path && a->path && b->path);
3027         if (ci && dir_in_way(opt->repo->index, path, !opt->call_depth,
3028                              S_ISGITLINK(ci->ren1->pair->two->mode)))
3029                 df_conflict_remains = 1;
3030
3031         if (merge_mode_and_contents(opt, o, a, b, path,
3032                                     opt->branch1, opt->branch2,
3033                                     opt->call_depth * 2, mfi))
3034                 return -1;
3035
3036         /*
3037          * We can skip updating the working tree file iff:
3038          *   a) The merge is clean
3039          *   b) The merge matches what was in HEAD (content, mode, pathname)
3040          *   c) The target path is usable (i.e. not involved in D/F conflict)
3041          */
3042         if (mfi->clean && was_tracked_and_matches(opt, path, &mfi->blob) &&
3043             !df_conflict_remains) {
3044                 int pos;
3045                 struct cache_entry *ce;
3046
3047                 output(opt, 3, _("Skipped %s (merged same as existing)"), path);
3048                 if (add_cacheinfo(opt, &mfi->blob, path,
3049                                   0, (!opt->call_depth && !is_dirty), 0))
3050                         return -1;
3051                 /*
3052                  * However, add_cacheinfo() will delete the old cache entry
3053                  * and add a new one.  We need to copy over any skip_worktree
3054                  * flag to avoid making the file appear as if it were
3055                  * deleted by the user.
3056                  */
3057                 pos = index_name_pos(&opt->orig_index, path, strlen(path));
3058                 ce = opt->orig_index.cache[pos];
3059                 if (ce_skip_worktree(ce)) {
3060                         pos = index_name_pos(opt->repo->index, path, strlen(path));
3061                         ce = opt->repo->index->cache[pos];
3062                         ce->ce_flags |= CE_SKIP_WORKTREE;
3063                 }
3064                 return mfi->clean;
3065         }
3066
3067         if (!mfi->clean) {
3068                 if (S_ISGITLINK(mfi->blob.mode))
3069                         reason = _("submodule");
3070                 output(opt, 1, _("CONFLICT (%s): Merge conflict in %s"),
3071                                 reason, path);
3072                 if (ci && !df_conflict_remains)
3073                         if (update_stages(opt, path, o, a, b))
3074                                 return -1;
3075         }
3076
3077         if (df_conflict_remains || is_dirty) {
3078                 char *new_path;
3079                 if (opt->call_depth) {
3080                         remove_file_from_index(opt->repo->index, path);
3081                 } else {
3082                         if (!mfi->clean) {
3083                                 if (update_stages(opt, path, o, a, b))
3084                                         return -1;
3085                         } else {
3086                                 int file_from_stage2 = was_tracked(opt, path);
3087
3088                                 if (update_stages(opt, path, NULL,
3089                                                   file_from_stage2 ? &mfi->blob : NULL,
3090                                                   file_from_stage2 ? NULL : &mfi->blob))
3091                                         return -1;
3092                         }
3093
3094                 }
3095                 new_path = unique_path(opt, path, ci->ren1->branch);
3096                 if (is_dirty) {
3097                         output(opt, 1, _("Refusing to lose dirty file at %s"),
3098                                path);
3099                 }
3100                 output(opt, 1, _("Adding as %s instead"), new_path);
3101                 if (update_file(opt, 0, &mfi->blob, new_path)) {
3102                         free(new_path);
3103                         return -1;
3104                 }
3105                 free(new_path);
3106                 mfi->clean = 0;
3107         } else if (update_file(opt, mfi->clean, &mfi->blob, path))
3108                 return -1;
3109         return !is_dirty && mfi->clean;
3110 }
3111
3112 static int handle_rename_normal(struct merge_options *opt,
3113                                 const char *path,
3114                                 const struct diff_filespec *o,
3115                                 const struct diff_filespec *a,
3116                                 const struct diff_filespec *b,
3117                                 struct rename_conflict_info *ci)
3118 {
3119         struct rename *ren = ci->ren1;
3120         struct merge_file_info mfi;
3121         int clean;
3122         int side = (ren->branch == opt->branch1 ? 2 : 3);
3123
3124         /* Merge the content and write it out */
3125         clean = handle_content_merge(&mfi, opt, path, was_dirty(opt, path),
3126                                      o, a, b, ci);
3127
3128         if (clean && opt->detect_directory_renames == 1 &&
3129             ren->dir_rename_original_dest) {
3130                 if (update_stages(opt, path,
3131                                   NULL,
3132                                   side == 2 ? &mfi.blob : NULL,
3133                                   side == 2 ? NULL : &mfi.blob))
3134                         return -1;
3135                 clean = 0; /* not clean, but conflicted */
3136         }
3137         return clean;
3138 }
3139
3140 static void dir_rename_warning(const char *msg,
3141                                int is_add,
3142                                int clean,
3143                                struct merge_options *opt,
3144                                struct rename *ren)
3145 {
3146         const char *other_branch;
3147         other_branch = (ren->branch == opt->branch1 ?
3148                         opt->branch2 : opt->branch1);
3149         if (is_add) {
3150                 output(opt, clean ? 2 : 1, msg,
3151                        ren->pair->one->path, ren->branch,
3152                        other_branch, ren->pair->two->path);
3153                 return;
3154         }
3155         output(opt, clean ? 2 : 1, msg,
3156                ren->pair->one->path, ren->dir_rename_original_dest, ren->branch,
3157                other_branch, ren->pair->two->path);
3158 }
3159 static int warn_about_dir_renamed_entries(struct merge_options *opt,
3160                                           struct rename *ren)
3161 {
3162         const char *msg;
3163         int clean = 1, is_add;
3164
3165         if (!ren)
3166                 return clean;
3167
3168         /* Return early if ren was not affected/created by a directory rename */
3169         if (!ren->dir_rename_original_dest)
3170                 return clean;
3171
3172         /* Sanity checks */
3173         assert(opt->detect_directory_renames > 0);
3174         assert(ren->dir_rename_original_type == 'A' ||
3175                ren->dir_rename_original_type == 'R');
3176
3177         /* Check whether to treat directory renames as a conflict */
3178         clean = (opt->detect_directory_renames == 2);
3179
3180         is_add = (ren->dir_rename_original_type == 'A');
3181         if (ren->dir_rename_original_type == 'A' && clean) {
3182                 msg = _("Path updated: %s added in %s inside a "
3183                         "directory that was renamed in %s; moving it to %s.");
3184         } else if (ren->dir_rename_original_type == 'A' && !clean) {
3185                 msg = _("CONFLICT (file location): %s added in %s "
3186                         "inside a directory that was renamed in %s, "
3187                         "suggesting it should perhaps be moved to %s.");
3188         } else if (ren->dir_rename_original_type == 'R' && clean) {
3189                 msg = _("Path updated: %s renamed to %s in %s, inside a "
3190                         "directory that was renamed in %s; moving it to %s.");
3191         } else if (ren->dir_rename_original_type == 'R' && !clean) {
3192                 msg = _("CONFLICT (file location): %s renamed to %s in %s, "
3193                         "inside a directory that was renamed in %s, "
3194                         "suggesting it should perhaps be moved to %s.");
3195         } else {
3196                 BUG("Impossible dir_rename_original_type/clean combination");
3197         }
3198         dir_rename_warning(msg, is_add, clean, opt, ren);
3199
3200         return clean;
3201 }
3202
3203 /* Per entry merge function */
3204 static int process_entry(struct merge_options *opt,
3205                          const char *path, struct stage_data *entry)
3206 {
3207         int clean_merge = 1;
3208         int normalize = opt->renormalize;
3209
3210         struct diff_filespec *o = &entry->stages[1];
3211         struct diff_filespec *a = &entry->stages[2];
3212         struct diff_filespec *b = &entry->stages[3];
3213         int o_valid = is_valid(o);
3214         int a_valid = is_valid(a);
3215         int b_valid = is_valid(b);
3216         o->path = a->path = b->path = (char*)path;
3217
3218         entry->processed = 1;
3219         if (entry->rename_conflict_info) {
3220                 struct rename_conflict_info *ci = entry->rename_conflict_info;
3221                 struct diff_filespec *temp;
3222                 int path_clean;
3223
3224                 path_clean = warn_about_dir_renamed_entries(opt, ci->ren1);
3225                 path_clean &= warn_about_dir_renamed_entries(opt, ci->ren2);
3226
3227                 /*
3228                  * For cases with a single rename, {o,a,b}->path have all been
3229                  * set to the rename target path; we need to set two of these
3230                  * back to the rename source.
3231                  * For rename/rename conflicts, we'll manually fix paths below.
3232                  */
3233                 temp = (opt->branch1 == ci->ren1->branch) ? b : a;
3234                 o->path = temp->path = ci->ren1->pair->one->path;
3235                 if (ci->ren2) {
3236                         assert(opt->branch1 == ci->ren1->branch);
3237                 }
3238
3239                 switch (ci->rename_type) {
3240                 case RENAME_NORMAL:
3241                 case RENAME_ONE_FILE_TO_ONE:
3242                         clean_merge = handle_rename_normal(opt, path, o, a, b,
3243                                                            ci);
3244                         break;
3245                 case RENAME_VIA_DIR:
3246                         clean_merge = handle_rename_via_dir(opt, ci);
3247                         break;
3248                 case RENAME_ADD:
3249                         /*
3250                          * Probably unclean merge, but if the renamed file
3251                          * merges cleanly and the result can then be
3252                          * two-way merged cleanly with the added file, I
3253                          * guess it's a clean merge?
3254                          */
3255                         clean_merge = handle_rename_add(opt, ci);
3256                         break;
3257                 case RENAME_DELETE:
3258                         clean_merge = 0;
3259                         if (handle_rename_delete(opt, ci))
3260                                 clean_merge = -1;
3261                         break;
3262                 case RENAME_ONE_FILE_TO_TWO:
3263                         /*
3264                          * Manually fix up paths; note:
3265                          * ren[12]->pair->one->path are equal.
3266                          */
3267                         o->path = ci->ren1->pair->one->path;
3268                         a->path = ci->ren1->pair->two->path;
3269                         b->path = ci->ren2->pair->two->path;
3270
3271                         clean_merge = 0;
3272                         if (handle_rename_rename_1to2(opt, ci))
3273                                 clean_merge = -1;
3274                         break;
3275                 case RENAME_TWO_FILES_TO_ONE:
3276                         /*
3277                          * Manually fix up paths; note,
3278                          * ren[12]->pair->two->path are actually equal.
3279                          */
3280                         o->path = NULL;
3281                         a->path = ci->ren1->pair->two->path;
3282                         b->path = ci->ren2->pair->two->path;
3283
3284                         /*
3285                          * Probably unclean merge, but if the two renamed
3286                          * files merge cleanly and the two resulting files
3287                          * can then be two-way merged cleanly, I guess it's
3288                          * a clean merge?
3289                          */
3290                         clean_merge = handle_rename_rename_2to1(opt, ci);
3291                         break;
3292                 default:
3293                         entry->processed = 0;
3294                         break;
3295                 }
3296                 if (path_clean < clean_merge)
3297                         clean_merge = path_clean;
3298         } else if (o_valid && (!a_valid || !b_valid)) {
3299                 /* Case A: Deleted in one */
3300                 if ((!a_valid && !b_valid) ||
3301                     (!b_valid && blob_unchanged(opt, o, a, normalize, path)) ||
3302                     (!a_valid && blob_unchanged(opt, o, b, normalize, path))) {
3303                         /* Deleted in both or deleted in one and
3304                          * unchanged in the other */
3305                         if (a_valid)
3306                                 output(opt, 2, _("Removing %s"), path);
3307                         /* do not touch working file if it did not exist */
3308                         remove_file(opt, 1, path, !a_valid);
3309                 } else {
3310                         /* Modify/delete; deleted side may have put a directory in the way */
3311                         clean_merge = 0;
3312                         if (handle_modify_delete(opt, path, o, a, b))
3313                                 clean_merge = -1;
3314                 }
3315         } else if ((!o_valid && a_valid && !b_valid) ||
3316                    (!o_valid && !a_valid && b_valid)) {
3317                 /* Case B: Added in one. */
3318                 /* [nothing|directory] -> ([nothing|directory], file) */
3319
3320                 const char *add_branch;
3321                 const char *other_branch;
3322                 const char *conf;
3323                 const struct diff_filespec *contents;
3324
3325                 if (a_valid) {
3326                         add_branch = opt->branch1;
3327                         other_branch = opt->branch2;
3328                         contents = a;
3329                         conf = _("file/directory");
3330                 } else {
3331                         add_branch = opt->branch2;
3332                         other_branch = opt->branch1;
3333                         contents = b;
3334                         conf = _("directory/file");
3335                 }
3336                 if (dir_in_way(opt->repo->index, path,
3337                                !opt->call_depth && !S_ISGITLINK(a->mode),
3338                                0)) {
3339                         char *new_path = unique_path(opt, path, add_branch);
3340                         clean_merge = 0;
3341                         output(opt, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3342                                "Adding %s as %s"),
3343                                conf, path, other_branch, path, new_path);
3344                         if (update_file(opt, 0, contents, new_path))
3345                                 clean_merge = -1;
3346                         else if (opt->call_depth)
3347                                 remove_file_from_index(opt->repo->index, path);
3348                         free(new_path);
3349                 } else {
3350                         output(opt, 2, _("Adding %s"), path);
3351                         /* do not overwrite file if already present */
3352                         if (update_file_flags(opt, contents, path, 1, !a_valid))
3353                                 clean_merge = -1;
3354                 }
3355         } else if (a_valid && b_valid) {
3356                 if (!o_valid) {
3357                         /* Case C: Added in both (check for same permissions) */
3358                         output(opt, 1,
3359                                _("CONFLICT (add/add): Merge conflict in %s"),
3360                                path);
3361                         clean_merge = handle_file_collision(opt,
3362                                                             path, NULL, NULL,
3363                                                             opt->branch1,
3364                                                             opt->branch2,
3365                                                             a, b);
3366                 } else {
3367                         /* case D: Modified in both, but differently. */
3368                         struct merge_file_info mfi;
3369                         int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3370                         clean_merge = handle_content_merge(&mfi, opt, path,
3371                                                            is_dirty,
3372                                                            o, a, b, NULL);
3373                 }
3374         } else if (!o_valid && !a_valid && !b_valid) {
3375                 /*
3376                  * this entry was deleted altogether. a_mode == 0 means
3377                  * we had that path and want to actively remove it.
3378                  */
3379                 remove_file(opt, 1, path, !a->mode);
3380         } else
3381                 BUG("fatal merge failure, shouldn't happen.");
3382
3383         return clean_merge;
3384 }
3385
3386 int merge_trees(struct merge_options *opt,
3387                 struct tree *head,
3388                 struct tree *merge,
3389                 struct tree *common,
3390                 struct tree **result)
3391 {
3392         struct index_state *istate = opt->repo->index;
3393         int code, clean;
3394         struct strbuf sb = STRBUF_INIT;
3395
3396         if (!opt->call_depth && repo_index_has_changes(opt->repo, head, &sb)) {
3397                 err(opt, _("Your local changes to the following files would be overwritten by merge:\n  %s"),
3398                     sb.buf);
3399                 return -1;
3400         }
3401
3402         if (opt->subtree_shift) {
3403                 merge = shift_tree_object(opt->repo, head, merge, opt->subtree_shift);
3404                 common = shift_tree_object(opt->repo, head, common, opt->subtree_shift);
3405         }
3406
3407         if (oid_eq(&common->object.oid, &merge->object.oid)) {
3408                 output(opt, 0, _("Already up to date!"));
3409                 *result = head;
3410                 return 1;
3411         }
3412
3413         code = unpack_trees_start(opt, common, head, merge);
3414
3415         if (code != 0) {
3416                 if (show(opt, 4) || opt->call_depth)
3417                         err(opt, _("merging of trees %s and %s failed"),
3418                             oid_to_hex(&head->object.oid),
3419                             oid_to_hex(&merge->object.oid));
3420                 unpack_trees_finish(opt);
3421                 return -1;
3422         }
3423
3424         if (unmerged_index(istate)) {
3425                 struct string_list *entries;
3426                 struct rename_info re_info;
3427                 int i;
3428                 /*
3429                  * Only need the hashmap while processing entries, so
3430                  * initialize it here and free it when we are done running
3431                  * through the entries. Keeping it in the merge_options as
3432                  * opposed to decaring a local hashmap is for convenience
3433                  * so that we don't have to pass it to around.
3434                  */
3435                 hashmap_init(&opt->current_file_dir_set, path_hashmap_cmp, NULL, 512);
3436                 get_files_dirs(opt, head);
3437                 get_files_dirs(opt, merge);
3438
3439                 entries = get_unmerged(opt->repo->index);
3440                 clean = detect_and_process_renames(opt, common, head, merge,
3441                                                    entries, &re_info);
3442                 record_df_conflict_files(opt, entries);
3443                 if (clean < 0)
3444                         goto cleanup;
3445                 for (i = entries->nr-1; 0 <= i; i--) {
3446                         const char *path = entries->items[i].string;
3447                         struct stage_data *e = entries->items[i].util;
3448                         if (!e->processed) {
3449                                 int ret = process_entry(opt, path, e);
3450                                 if (!ret)
3451                                         clean = 0;
3452                                 else if (ret < 0) {
3453                                         clean = ret;
3454                                         goto cleanup;
3455                                 }
3456                         }
3457                 }
3458                 for (i = 0; i < entries->nr; i++) {
3459                         struct stage_data *e = entries->items[i].util;
3460                         if (!e->processed)
3461                                 BUG("unprocessed path??? %s",
3462                                     entries->items[i].string);
3463                 }
3464
3465         cleanup:
3466                 final_cleanup_renames(&re_info);
3467
3468                 string_list_clear(entries, 1);
3469                 free(entries);
3470
3471                 hashmap_free(&opt->current_file_dir_set, 1);
3472
3473                 if (clean < 0) {
3474                         unpack_trees_finish(opt);
3475                         return clean;
3476                 }
3477         }
3478         else
3479                 clean = 1;
3480
3481         unpack_trees_finish(opt);
3482
3483         if (opt->call_depth && !(*result = write_tree_from_memory(opt)))
3484                 return -1;
3485
3486         return clean;
3487 }
3488
3489 static struct commit_list *reverse_commit_list(struct commit_list *list)
3490 {
3491         struct commit_list *next = NULL, *current, *backup;
3492         for (current = list; current; current = backup) {
3493                 backup = current->next;
3494                 current->next = next;
3495                 next = current;
3496         }
3497         return next;
3498 }
3499
3500 /*
3501  * Merge the commits h1 and h2, return the resulting virtual
3502  * commit object and a flag indicating the cleanness of the merge.
3503  */
3504 int merge_recursive(struct merge_options *opt,
3505                     struct commit *h1,
3506                     struct commit *h2,
3507                     struct commit_list *ca,
3508                     struct commit **result)
3509 {
3510         struct commit_list *iter;
3511         struct commit *merged_common_ancestors;
3512         struct tree *mrtree;
3513         int clean;
3514
3515         if (show(opt, 4)) {
3516                 output(opt, 4, _("Merging:"));
3517                 output_commit_title(opt, h1);
3518                 output_commit_title(opt, h2);
3519         }
3520
3521         if (!ca) {
3522                 ca = get_merge_bases(h1, h2);
3523                 ca = reverse_commit_list(ca);
3524         }
3525
3526         if (show(opt, 5)) {
3527                 unsigned cnt = commit_list_count(ca);
3528
3529                 output(opt, 5, Q_("found %u common ancestor:",
3530                                 "found %u common ancestors:", cnt), cnt);
3531                 for (iter = ca; iter; iter = iter->next)
3532                         output_commit_title(opt, iter->item);
3533         }
3534
3535         merged_common_ancestors = pop_commit(&ca);
3536         if (merged_common_ancestors == NULL) {
3537                 /* if there is no common ancestor, use an empty tree */
3538                 struct tree *tree;
3539
3540                 tree = lookup_tree(opt->repo, opt->repo->hash_algo->empty_tree);
3541                 merged_common_ancestors = make_virtual_commit(opt->repo, tree, "ancestor");
3542         }
3543
3544         for (iter = ca; iter; iter = iter->next) {
3545                 const char *saved_b1, *saved_b2;
3546                 opt->call_depth++;
3547                 /*
3548                  * When the merge fails, the result contains files
3549                  * with conflict markers. The cleanness flag is
3550                  * ignored (unless indicating an error), it was never
3551                  * actually used, as result of merge_trees has always
3552                  * overwritten it: the committed "conflicts" were
3553                  * already resolved.
3554                  */
3555                 discard_index(opt->repo->index);
3556                 saved_b1 = opt->branch1;
3557                 saved_b2 = opt->branch2;
3558                 opt->branch1 = "Temporary merge branch 1";
3559                 opt->branch2 = "Temporary merge branch 2";
3560                 if (merge_recursive(opt, merged_common_ancestors, iter->item,
3561                                     NULL, &merged_common_ancestors) < 0)
3562                         return -1;
3563                 opt->branch1 = saved_b1;
3564                 opt->branch2 = saved_b2;
3565                 opt->call_depth--;
3566
3567                 if (!merged_common_ancestors)
3568                         return err(opt, _("merge returned no commit"));
3569         }
3570
3571         discard_index(opt->repo->index);
3572         if (!opt->call_depth)
3573                 repo_read_index(opt->repo);
3574
3575         opt->ancestor = "merged common ancestors";
3576         clean = merge_trees(opt, get_commit_tree(h1), get_commit_tree(h2),
3577                             get_commit_tree(merged_common_ancestors),
3578                             &mrtree);
3579         if (clean < 0) {
3580                 flush_output(opt);
3581                 return clean;
3582         }
3583
3584         if (opt->call_depth) {
3585                 *result = make_virtual_commit(opt->repo, mrtree, "merged tree");
3586                 commit_list_insert(h1, &(*result)->parents);
3587                 commit_list_insert(h2, &(*result)->parents->next);
3588         }
3589         flush_output(opt);
3590         if (!opt->call_depth && opt->buffer_output < 2)
3591                 strbuf_release(&opt->obuf);
3592         if (show(opt, 2))
3593                 diff_warn_rename_limit("merge.renamelimit",
3594                                        opt->needed_rename_limit, 0);
3595         return clean;
3596 }
3597
3598 static struct commit *get_ref(struct repository *repo, const struct object_id *oid,
3599                               const char *name)
3600 {
3601         struct object *object;
3602
3603         object = deref_tag(repo, parse_object(repo, oid),
3604                            name, strlen(name));
3605         if (!object)
3606                 return NULL;
3607         if (object->type == OBJ_TREE)
3608                 return make_virtual_commit(repo, (struct tree*)object, name);
3609         if (object->type != OBJ_COMMIT)
3610                 return NULL;
3611         if (parse_commit((struct commit *)object))
3612                 return NULL;
3613         return (struct commit *)object;
3614 }
3615
3616 int merge_recursive_generic(struct merge_options *opt,
3617                             const struct object_id *head,
3618                             const struct object_id *merge,
3619                             int num_base_list,
3620                             const struct object_id **base_list,
3621                             struct commit **result)
3622 {
3623         int clean;
3624         struct lock_file lock = LOCK_INIT;
3625         struct commit *head_commit = get_ref(opt->repo, head, opt->branch1);
3626         struct commit *next_commit = get_ref(opt->repo, merge, opt->branch2);
3627         struct commit_list *ca = NULL;
3628
3629         if (base_list) {
3630                 int i;
3631                 for (i = 0; i < num_base_list; ++i) {
3632                         struct commit *base;
3633                         if (!(base = get_ref(opt->repo, base_list[i], oid_to_hex(base_list[i]))))
3634                                 return err(opt, _("Could not parse object '%s'"),
3635                                            oid_to_hex(base_list[i]));
3636                         commit_list_insert(base, &ca);
3637                 }
3638         }
3639
3640         repo_hold_locked_index(opt->repo, &lock, LOCK_DIE_ON_ERROR);
3641         clean = merge_recursive(opt, head_commit, next_commit, ca,
3642                                 result);
3643         if (clean < 0) {
3644                 rollback_lock_file(&lock);
3645                 return clean;
3646         }
3647
3648         if (write_locked_index(opt->repo->index, &lock,
3649                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3650                 return err(opt, _("Unable to write index."));
3651
3652         return clean ? 0 : 1;
3653 }
3654
3655 static void merge_recursive_config(struct merge_options *opt)
3656 {
3657         char *value = NULL;
3658         git_config_get_int("merge.verbosity", &opt->verbosity);
3659         git_config_get_int("diff.renamelimit", &opt->diff_rename_limit);
3660         git_config_get_int("merge.renamelimit", &opt->merge_rename_limit);
3661         if (!git_config_get_string("diff.renames", &value)) {
3662                 opt->diff_detect_rename = git_config_rename("diff.renames", value);
3663                 free(value);
3664         }
3665         if (!git_config_get_string("merge.renames", &value)) {
3666                 opt->merge_detect_rename = git_config_rename("merge.renames", value);
3667                 free(value);
3668         }
3669         if (!git_config_get_string("merge.directoryrenames", &value)) {
3670                 int boolval = git_parse_maybe_bool(value);
3671                 if (0 <= boolval) {
3672                         opt->detect_directory_renames = boolval ? 2 : 0;
3673                 } else if (!strcasecmp(value, "conflict")) {
3674                         opt->detect_directory_renames = 1;
3675                 } /* avoid erroring on values from future versions of git */
3676                 free(value);
3677         }
3678         git_config(git_xmerge_config, NULL);
3679 }
3680
3681 void init_merge_options(struct merge_options *opt,
3682                         struct repository *repo)
3683 {
3684         const char *merge_verbosity;
3685         memset(opt, 0, sizeof(struct merge_options));
3686         opt->repo = repo;
3687         opt->verbosity = 2;
3688         opt->buffer_output = 1;
3689         opt->diff_rename_limit = -1;
3690         opt->merge_rename_limit = -1;
3691         opt->renormalize = 0;
3692         opt->diff_detect_rename = -1;
3693         opt->merge_detect_rename = -1;
3694         opt->detect_directory_renames = 1;
3695         merge_recursive_config(opt);
3696         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3697         if (merge_verbosity)
3698                 opt->verbosity = strtol(merge_verbosity, NULL, 10);
3699         if (opt->verbosity >= 5)
3700                 opt->buffer_output = 0;
3701         strbuf_init(&opt->obuf, 0);
3702         string_list_init(&opt->df_conflict_file_set, 1);
3703 }
3704
3705 int parse_merge_opt(struct merge_options *opt, const char *s)
3706 {
3707         const char *arg;
3708
3709         if (!s || !*s)
3710                 return -1;
3711         if (!strcmp(s, "ours"))
3712                 opt->recursive_variant = MERGE_RECURSIVE_OURS;
3713         else if (!strcmp(s, "theirs"))
3714                 opt->recursive_variant = MERGE_RECURSIVE_THEIRS;
3715         else if (!strcmp(s, "subtree"))
3716                 opt->subtree_shift = "";
3717         else if (skip_prefix(s, "subtree=", &arg))
3718                 opt->subtree_shift = arg;
3719         else if (!strcmp(s, "patience"))
3720                 opt->xdl_opts = DIFF_WITH_ALG(opt, PATIENCE_DIFF);
3721         else if (!strcmp(s, "histogram"))
3722                 opt->xdl_opts = DIFF_WITH_ALG(opt, HISTOGRAM_DIFF);
3723         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3724                 long value = parse_algorithm_value(arg);
3725                 if (value < 0)
3726                         return -1;
3727                 /* clear out previous settings */
3728                 DIFF_XDL_CLR(opt, NEED_MINIMAL);
3729                 opt->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3730                 opt->xdl_opts |= value;
3731         }
3732         else if (!strcmp(s, "ignore-space-change"))
3733                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_CHANGE);
3734         else if (!strcmp(s, "ignore-all-space"))
3735                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE);
3736         else if (!strcmp(s, "ignore-space-at-eol"))
3737                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_AT_EOL);
3738         else if (!strcmp(s, "ignore-cr-at-eol"))
3739                 DIFF_XDL_SET(opt, IGNORE_CR_AT_EOL);
3740         else if (!strcmp(s, "renormalize"))
3741                 opt->renormalize = 1;
3742         else if (!strcmp(s, "no-renormalize"))
3743                 opt->renormalize = 0;
3744         else if (!strcmp(s, "no-renames"))
3745                 opt->merge_detect_rename = 0;
3746         else if (!strcmp(s, "find-renames")) {
3747                 opt->merge_detect_rename = 1;
3748                 opt->rename_score = 0;
3749         }
3750         else if (skip_prefix(s, "find-renames=", &arg) ||
3751                  skip_prefix(s, "rename-threshold=", &arg)) {
3752                 if ((opt->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3753                         return -1;
3754                 opt->merge_detect_rename = 1;
3755         }
3756         /*
3757          * Please update $__git_merge_strategy_options in
3758          * git-completion.bash when you add new options
3759          */
3760         else
3761                 return -1;
3762         return 0;
3763 }