merge-recursive: shrink rename_conflict_info
[git] / merge-recursive.c
1 /*
2  * Recursive Merge algorithm stolen from git-merge-recursive.py by
3  * Fredrik Kuivinen.
4  * The thieves were Alex Riesen and Johannes Schindelin, in June/July 2006
5  */
6 #include "cache.h"
7 #include "config.h"
8 #include "advice.h"
9 #include "lockfile.h"
10 #include "cache-tree.h"
11 #include "object-store.h"
12 #include "repository.h"
13 #include "commit.h"
14 #include "blob.h"
15 #include "builtin.h"
16 #include "tree-walk.h"
17 #include "diff.h"
18 #include "diffcore.h"
19 #include "tag.h"
20 #include "alloc.h"
21 #include "unpack-trees.h"
22 #include "string-list.h"
23 #include "xdiff-interface.h"
24 #include "ll-merge.h"
25 #include "attr.h"
26 #include "merge-recursive.h"
27 #include "dir.h"
28 #include "submodule.h"
29 #include "revision.h"
30 #include "commit-reach.h"
31
32 struct path_hashmap_entry {
33         struct hashmap_entry e;
34         char path[FLEX_ARRAY];
35 };
36
37 static int path_hashmap_cmp(const void *cmp_data,
38                             const void *entry,
39                             const void *entry_or_key,
40                             const void *keydata)
41 {
42         const struct path_hashmap_entry *a = entry;
43         const struct path_hashmap_entry *b = entry_or_key;
44         const char *key = keydata;
45
46         if (ignore_case)
47                 return strcasecmp(a->path, key ? key : b->path);
48         else
49                 return strcmp(a->path, key ? key : b->path);
50 }
51
52 static unsigned int path_hash(const char *path)
53 {
54         return ignore_case ? strihash(path) : strhash(path);
55 }
56
57 static struct dir_rename_entry *dir_rename_find_entry(struct hashmap *hashmap,
58                                                       char *dir)
59 {
60         struct dir_rename_entry key;
61
62         if (dir == NULL)
63                 return NULL;
64         hashmap_entry_init(&key, strhash(dir));
65         key.dir = dir;
66         return hashmap_get(hashmap, &key, NULL);
67 }
68
69 static int dir_rename_cmp(const void *unused_cmp_data,
70                           const void *entry,
71                           const void *entry_or_key,
72                           const void *unused_keydata)
73 {
74         const struct dir_rename_entry *e1 = entry;
75         const struct dir_rename_entry *e2 = entry_or_key;
76
77         return strcmp(e1->dir, e2->dir);
78 }
79
80 static void dir_rename_init(struct hashmap *map)
81 {
82         hashmap_init(map, dir_rename_cmp, NULL, 0);
83 }
84
85 static void dir_rename_entry_init(struct dir_rename_entry *entry,
86                                   char *directory)
87 {
88         hashmap_entry_init(entry, strhash(directory));
89         entry->dir = directory;
90         entry->non_unique_new_dir = 0;
91         strbuf_init(&entry->new_dir, 0);
92         string_list_init(&entry->possible_new_dirs, 0);
93 }
94
95 static struct collision_entry *collision_find_entry(struct hashmap *hashmap,
96                                                     char *target_file)
97 {
98         struct collision_entry key;
99
100         hashmap_entry_init(&key, strhash(target_file));
101         key.target_file = target_file;
102         return hashmap_get(hashmap, &key, NULL);
103 }
104
105 static int collision_cmp(void *unused_cmp_data,
106                          const struct collision_entry *e1,
107                          const struct collision_entry *e2,
108                          const void *unused_keydata)
109 {
110         return strcmp(e1->target_file, e2->target_file);
111 }
112
113 static void collision_init(struct hashmap *map)
114 {
115         hashmap_init(map, (hashmap_cmp_fn) collision_cmp, NULL, 0);
116 }
117
118 static void flush_output(struct merge_options *opt)
119 {
120         if (opt->buffer_output < 2 && opt->obuf.len) {
121                 fputs(opt->obuf.buf, stdout);
122                 strbuf_reset(&opt->obuf);
123         }
124 }
125
126 static int err(struct merge_options *opt, const char *err, ...)
127 {
128         va_list params;
129
130         if (opt->buffer_output < 2)
131                 flush_output(opt);
132         else {
133                 strbuf_complete(&opt->obuf, '\n');
134                 strbuf_addstr(&opt->obuf, "error: ");
135         }
136         va_start(params, err);
137         strbuf_vaddf(&opt->obuf, err, params);
138         va_end(params);
139         if (opt->buffer_output > 1)
140                 strbuf_addch(&opt->obuf, '\n');
141         else {
142                 error("%s", opt->obuf.buf);
143                 strbuf_reset(&opt->obuf);
144         }
145
146         return -1;
147 }
148
149 static struct tree *shift_tree_object(struct repository *repo,
150                                       struct tree *one, struct tree *two,
151                                       const char *subtree_shift)
152 {
153         struct object_id shifted;
154
155         if (!*subtree_shift) {
156                 shift_tree(&one->object.oid, &two->object.oid, &shifted, 0);
157         } else {
158                 shift_tree_by(&one->object.oid, &two->object.oid, &shifted,
159                               subtree_shift);
160         }
161         if (oideq(&two->object.oid, &shifted))
162                 return two;
163         return lookup_tree(repo, &shifted);
164 }
165
166 static struct commit *make_virtual_commit(struct repository *repo,
167                                           struct tree *tree,
168                                           const char *comment)
169 {
170         struct commit *commit = alloc_commit_node(repo);
171
172         set_merge_remote_desc(commit, comment, (struct object *)commit);
173         commit->maybe_tree = tree;
174         commit->object.parsed = 1;
175         return commit;
176 }
177
178 /*
179  * Since we use get_tree_entry(), which does not put the read object into
180  * the object pool, we cannot rely on a == b.
181  */
182 static int oid_eq(const struct object_id *a, const struct object_id *b)
183 {
184         if (!a && !b)
185                 return 2;
186         return a && b && oideq(a, b);
187 }
188
189 enum rename_type {
190         RENAME_NORMAL = 0,
191         RENAME_VIA_DIR,
192         RENAME_ADD,
193         RENAME_DELETE,
194         RENAME_ONE_FILE_TO_ONE,
195         RENAME_ONE_FILE_TO_TWO,
196         RENAME_TWO_FILES_TO_ONE
197 };
198
199 /*
200  * Since we want to write the index eventually, we cannot reuse the index
201  * for these (temporary) data.
202  */
203 struct stage_data {
204         struct {
205                 unsigned short mode;
206                 struct object_id oid;
207         } stages[4];
208         struct rename_conflict_info *rename_conflict_info;
209         unsigned processed:1;
210 };
211
212 struct rename {
213         struct diff_filepair *pair;
214         /*
215          * Purpose of src_entry and dst_entry:
216          *
217          * If 'before' is renamed to 'after' then src_entry will contain
218          * the versions of 'before' from the merge_base, HEAD, and MERGE in
219          * stages 1, 2, and 3; dst_entry will contain the respective
220          * versions of 'after' in corresponding locations.  Thus, we have a
221          * total of six modes and oids, though some will be null.  (Stage 0
222          * is ignored; we're interested in handling conflicts.)
223          *
224          * Since we don't turn on break-rewrites by default, neither
225          * src_entry nor dst_entry can have all three of their stages have
226          * non-null oids, meaning at most four of the six will be non-null.
227          * Also, since this is a rename, both src_entry and dst_entry will
228          * have at least one non-null oid, meaning at least two will be
229          * non-null.  Of the six oids, a typical rename will have three be
230          * non-null.  Only two implies a rename/delete, and four implies a
231          * rename/add.
232          */
233         struct stage_data *src_entry;
234         struct stage_data *dst_entry;
235         unsigned add_turned_into_rename:1;
236         unsigned processed:1;
237 };
238
239 struct rename_conflict_info {
240         enum rename_type rename_type;
241         struct rename *ren1;
242         struct rename *ren2;
243         const char *branch1;
244         const char *branch2;
245         struct diff_filespec ren1_other;
246         struct diff_filespec ren2_other;
247 };
248
249 static inline void setup_rename_conflict_info(enum rename_type rename_type,
250                                               struct merge_options *opt,
251                                               struct rename *ren1,
252                                               struct rename *ren2,
253                                               const char *branch1,
254                                               const char *branch2,
255                                               struct stage_data *src_entry1,
256                                               struct stage_data *src_entry2)
257 {
258         int ostage1 = 0, ostage2;
259         struct rename_conflict_info *ci;
260
261         /*
262          * When we have two renames involved, it's easiest to get the
263          * correct things into stage 2 and 3, and to make sure that the
264          * content merge puts HEAD before the other branch if we just
265          * ensure that branch1 == opt->branch1.  So, simply flip arguments
266          * around if we don't have that.
267          */
268         if (ren2 && branch1 != opt->branch1) {
269                 setup_rename_conflict_info(rename_type,
270                                            opt,
271                                            ren2,       ren1,
272                                            branch2,    branch1,
273                                            src_entry2, src_entry1);
274                 return;
275         }
276
277         ci = xcalloc(1, sizeof(struct rename_conflict_info));
278         ci->rename_type = rename_type;
279         ci->ren1 = ren1;
280         ci->ren2 = ren2;
281         ci->branch1 = branch1;
282         ci->branch2 = branch2;
283
284         ci->ren1->dst_entry->processed = 0;
285         ci->ren1->dst_entry->rename_conflict_info = ci;
286
287         if (ren2) {
288                 ci->ren2->dst_entry->rename_conflict_info = ci;
289         }
290
291         /*
292          * For each rename, there could have been
293          * modifications on the side of history where that
294          * file was not renamed.
295          */
296         if (rename_type == RENAME_ADD ||
297             rename_type == RENAME_TWO_FILES_TO_ONE) {
298                 ostage1 = opt->branch1 == branch1 ? 3 : 2;
299
300                 ci->ren1_other.path = ren1->pair->one->path;
301                 oidcpy(&ci->ren1_other.oid, &src_entry1->stages[ostage1].oid);
302                 ci->ren1_other.mode = src_entry1->stages[ostage1].mode;
303         }
304
305         if (rename_type == RENAME_TWO_FILES_TO_ONE) {
306                 ostage2 = ostage1 ^ 1;
307
308                 ci->ren2_other.path = ren2->pair->one->path;
309                 oidcpy(&ci->ren2_other.oid, &src_entry2->stages[ostage2].oid);
310                 ci->ren2_other.mode = src_entry2->stages[ostage2].mode;
311         }
312 }
313
314 static int show(struct merge_options *opt, int v)
315 {
316         return (!opt->call_depth && opt->verbosity >= v) || opt->verbosity >= 5;
317 }
318
319 __attribute__((format (printf, 3, 4)))
320 static void output(struct merge_options *opt, int v, const char *fmt, ...)
321 {
322         va_list ap;
323
324         if (!show(opt, v))
325                 return;
326
327         strbuf_addchars(&opt->obuf, ' ', opt->call_depth * 2);
328
329         va_start(ap, fmt);
330         strbuf_vaddf(&opt->obuf, fmt, ap);
331         va_end(ap);
332
333         strbuf_addch(&opt->obuf, '\n');
334         if (!opt->buffer_output)
335                 flush_output(opt);
336 }
337
338 static void output_commit_title(struct merge_options *opt, struct commit *commit)
339 {
340         struct merge_remote_desc *desc;
341
342         strbuf_addchars(&opt->obuf, ' ', opt->call_depth * 2);
343         desc = merge_remote_util(commit);
344         if (desc)
345                 strbuf_addf(&opt->obuf, "virtual %s\n", desc->name);
346         else {
347                 strbuf_add_unique_abbrev(&opt->obuf, &commit->object.oid,
348                                          DEFAULT_ABBREV);
349                 strbuf_addch(&opt->obuf, ' ');
350                 if (parse_commit(commit) != 0)
351                         strbuf_addstr(&opt->obuf, _("(bad commit)\n"));
352                 else {
353                         const char *title;
354                         const char *msg = get_commit_buffer(commit, NULL);
355                         int len = find_commit_subject(msg, &title);
356                         if (len)
357                                 strbuf_addf(&opt->obuf, "%.*s\n", len, title);
358                         unuse_commit_buffer(commit, msg);
359                 }
360         }
361         flush_output(opt);
362 }
363
364 static int add_cacheinfo(struct merge_options *opt,
365                          unsigned int mode, const struct object_id *oid,
366                          const char *path, int stage, int refresh, int options)
367 {
368         struct index_state *istate = opt->repo->index;
369         struct cache_entry *ce;
370         int ret;
371
372         ce = make_cache_entry(istate, mode, oid ? oid : &null_oid, path, stage, 0);
373         if (!ce)
374                 return err(opt, _("add_cacheinfo failed for path '%s'; merge aborting."), path);
375
376         ret = add_index_entry(istate, ce, options);
377         if (refresh) {
378                 struct cache_entry *nce;
379
380                 nce = refresh_cache_entry(istate, ce,
381                                           CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
382                 if (!nce)
383                         return err(opt, _("add_cacheinfo failed to refresh for path '%s'; merge aborting."), path);
384                 if (nce != ce)
385                         ret = add_index_entry(istate, nce, options);
386         }
387         return ret;
388 }
389
390 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
391 {
392         parse_tree(tree);
393         init_tree_desc(desc, tree->buffer, tree->size);
394 }
395
396 static int unpack_trees_start(struct merge_options *opt,
397                               struct tree *common,
398                               struct tree *head,
399                               struct tree *merge)
400 {
401         int rc;
402         struct tree_desc t[3];
403         struct index_state tmp_index = { NULL };
404
405         memset(&opt->unpack_opts, 0, sizeof(opt->unpack_opts));
406         if (opt->call_depth)
407                 opt->unpack_opts.index_only = 1;
408         else
409                 opt->unpack_opts.update = 1;
410         opt->unpack_opts.merge = 1;
411         opt->unpack_opts.head_idx = 2;
412         opt->unpack_opts.fn = threeway_merge;
413         opt->unpack_opts.src_index = opt->repo->index;
414         opt->unpack_opts.dst_index = &tmp_index;
415         opt->unpack_opts.aggressive = !merge_detect_rename(opt);
416         setup_unpack_trees_porcelain(&opt->unpack_opts, "merge");
417
418         init_tree_desc_from_tree(t+0, common);
419         init_tree_desc_from_tree(t+1, head);
420         init_tree_desc_from_tree(t+2, merge);
421
422         rc = unpack_trees(3, t, &opt->unpack_opts);
423         cache_tree_free(&opt->repo->index->cache_tree);
424
425         /*
426          * Update opt->repo->index to match the new results, AFTER saving a copy
427          * in opt->orig_index.  Update src_index to point to the saved copy.
428          * (verify_uptodate() checks src_index, and the original index is
429          * the one that had the necessary modification timestamps.)
430          */
431         opt->orig_index = *opt->repo->index;
432         *opt->repo->index = tmp_index;
433         opt->unpack_opts.src_index = &opt->orig_index;
434
435         return rc;
436 }
437
438 static void unpack_trees_finish(struct merge_options *opt)
439 {
440         discard_index(&opt->orig_index);
441         clear_unpack_trees_porcelain(&opt->unpack_opts);
442 }
443
444 struct tree *write_tree_from_memory(struct merge_options *opt)
445 {
446         struct tree *result = NULL;
447         struct index_state *istate = opt->repo->index;
448
449         if (unmerged_index(istate)) {
450                 int i;
451                 fprintf(stderr, "BUG: There are unmerged index entries:\n");
452                 for (i = 0; i < istate->cache_nr; i++) {
453                         const struct cache_entry *ce = istate->cache[i];
454                         if (ce_stage(ce))
455                                 fprintf(stderr, "BUG: %d %.*s\n", ce_stage(ce),
456                                         (int)ce_namelen(ce), ce->name);
457                 }
458                 BUG("unmerged index entries in merge-recursive.c");
459         }
460
461         if (!istate->cache_tree)
462                 istate->cache_tree = cache_tree();
463
464         if (!cache_tree_fully_valid(istate->cache_tree) &&
465             cache_tree_update(istate, 0) < 0) {
466                 err(opt, _("error building trees"));
467                 return NULL;
468         }
469
470         result = lookup_tree(opt->repo, &istate->cache_tree->oid);
471
472         return result;
473 }
474
475 static int save_files_dirs(const struct object_id *oid,
476                            struct strbuf *base, const char *path,
477                            unsigned int mode, int stage, void *context)
478 {
479         struct path_hashmap_entry *entry;
480         int baselen = base->len;
481         struct merge_options *opt = context;
482
483         strbuf_addstr(base, path);
484
485         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
486         hashmap_entry_init(entry, path_hash(entry->path));
487         hashmap_add(&opt->current_file_dir_set, entry);
488
489         strbuf_setlen(base, baselen);
490         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
491 }
492
493 static void get_files_dirs(struct merge_options *opt, struct tree *tree)
494 {
495         struct pathspec match_all;
496         memset(&match_all, 0, sizeof(match_all));
497         read_tree_recursive(the_repository, tree, "", 0, 0,
498                             &match_all, save_files_dirs, opt);
499 }
500
501 static int get_tree_entry_if_blob(const struct object_id *tree,
502                                   const char *path,
503                                   struct object_id *hashy,
504                                   unsigned short *mode_o)
505 {
506         int ret;
507
508         ret = get_tree_entry(tree, path, hashy, mode_o);
509         if (S_ISDIR(*mode_o)) {
510                 oidcpy(hashy, &null_oid);
511                 *mode_o = 0;
512         }
513         return ret;
514 }
515
516 /*
517  * Returns an index_entry instance which doesn't have to correspond to
518  * a real cache entry in Git's index.
519  */
520 static struct stage_data *insert_stage_data(const char *path,
521                 struct tree *o, struct tree *a, struct tree *b,
522                 struct string_list *entries)
523 {
524         struct string_list_item *item;
525         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
526         get_tree_entry_if_blob(&o->object.oid, path,
527                                &e->stages[1].oid, &e->stages[1].mode);
528         get_tree_entry_if_blob(&a->object.oid, path,
529                                &e->stages[2].oid, &e->stages[2].mode);
530         get_tree_entry_if_blob(&b->object.oid, path,
531                                &e->stages[3].oid, &e->stages[3].mode);
532         item = string_list_insert(entries, path);
533         item->util = e;
534         return e;
535 }
536
537 /*
538  * Create a dictionary mapping file names to stage_data objects. The
539  * dictionary contains one entry for every path with a non-zero stage entry.
540  */
541 static struct string_list *get_unmerged(struct index_state *istate)
542 {
543         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
544         int i;
545
546         unmerged->strdup_strings = 1;
547
548         for (i = 0; i < istate->cache_nr; i++) {
549                 struct string_list_item *item;
550                 struct stage_data *e;
551                 const struct cache_entry *ce = istate->cache[i];
552                 if (!ce_stage(ce))
553                         continue;
554
555                 item = string_list_lookup(unmerged, ce->name);
556                 if (!item) {
557                         item = string_list_insert(unmerged, ce->name);
558                         item->util = xcalloc(1, sizeof(struct stage_data));
559                 }
560                 e = item->util;
561                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
562                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
563         }
564
565         return unmerged;
566 }
567
568 static int string_list_df_name_compare(const char *one, const char *two)
569 {
570         int onelen = strlen(one);
571         int twolen = strlen(two);
572         /*
573          * Here we only care that entries for D/F conflicts are
574          * adjacent, in particular with the file of the D/F conflict
575          * appearing before files below the corresponding directory.
576          * The order of the rest of the list is irrelevant for us.
577          *
578          * To achieve this, we sort with df_name_compare and provide
579          * the mode S_IFDIR so that D/F conflicts will sort correctly.
580          * We use the mode S_IFDIR for everything else for simplicity,
581          * since in other cases any changes in their order due to
582          * sorting cause no problems for us.
583          */
584         int cmp = df_name_compare(one, onelen, S_IFDIR,
585                                   two, twolen, S_IFDIR);
586         /*
587          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
588          * that 'foo' comes before 'foo/bar'.
589          */
590         if (cmp)
591                 return cmp;
592         return onelen - twolen;
593 }
594
595 static void record_df_conflict_files(struct merge_options *opt,
596                                      struct string_list *entries)
597 {
598         /* If there is a D/F conflict and the file for such a conflict
599          * currently exists in the working tree, we want to allow it to be
600          * removed to make room for the corresponding directory if needed.
601          * The files underneath the directories of such D/F conflicts will
602          * be processed before the corresponding file involved in the D/F
603          * conflict.  If the D/F directory ends up being removed by the
604          * merge, then we won't have to touch the D/F file.  If the D/F
605          * directory needs to be written to the working copy, then the D/F
606          * file will simply be removed (in make_room_for_path()) to make
607          * room for the necessary paths.  Note that if both the directory
608          * and the file need to be present, then the D/F file will be
609          * reinstated with a new unique name at the time it is processed.
610          */
611         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
612         const char *last_file = NULL;
613         int last_len = 0;
614         int i;
615
616         /*
617          * If we're merging merge-bases, we don't want to bother with
618          * any working directory changes.
619          */
620         if (opt->call_depth)
621                 return;
622
623         /* Ensure D/F conflicts are adjacent in the entries list. */
624         for (i = 0; i < entries->nr; i++) {
625                 struct string_list_item *next = &entries->items[i];
626                 string_list_append(&df_sorted_entries, next->string)->util =
627                                    next->util;
628         }
629         df_sorted_entries.cmp = string_list_df_name_compare;
630         string_list_sort(&df_sorted_entries);
631
632         string_list_clear(&opt->df_conflict_file_set, 1);
633         for (i = 0; i < df_sorted_entries.nr; i++) {
634                 const char *path = df_sorted_entries.items[i].string;
635                 int len = strlen(path);
636                 struct stage_data *e = df_sorted_entries.items[i].util;
637
638                 /*
639                  * Check if last_file & path correspond to a D/F conflict;
640                  * i.e. whether path is last_file+'/'+<something>.
641                  * If so, record that it's okay to remove last_file to make
642                  * room for path and friends if needed.
643                  */
644                 if (last_file &&
645                     len > last_len &&
646                     memcmp(path, last_file, last_len) == 0 &&
647                     path[last_len] == '/') {
648                         string_list_insert(&opt->df_conflict_file_set, last_file);
649                 }
650
651                 /*
652                  * Determine whether path could exist as a file in the
653                  * working directory as a possible D/F conflict.  This
654                  * will only occur when it exists in stage 2 as a
655                  * file.
656                  */
657                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
658                         last_file = path;
659                         last_len = len;
660                 } else {
661                         last_file = NULL;
662                 }
663         }
664         string_list_clear(&df_sorted_entries, 0);
665 }
666
667 static int update_stages(struct merge_options *opt, const char *path,
668                          const struct diff_filespec *o,
669                          const struct diff_filespec *a,
670                          const struct diff_filespec *b)
671 {
672
673         /*
674          * NOTE: It is usually a bad idea to call update_stages on a path
675          * before calling update_file on that same path, since it can
676          * sometimes lead to spurious "refusing to lose untracked file..."
677          * messages from update_file (via make_room_for path via
678          * would_lose_untracked).  Instead, reverse the order of the calls
679          * (executing update_file first and then update_stages).
680          */
681         int clear = 1;
682         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
683         if (clear)
684                 if (remove_file_from_index(opt->repo->index, path))
685                         return -1;
686         if (o)
687                 if (add_cacheinfo(opt, o->mode, &o->oid, path, 1, 0, options))
688                         return -1;
689         if (a)
690                 if (add_cacheinfo(opt, a->mode, &a->oid, path, 2, 0, options))
691                         return -1;
692         if (b)
693                 if (add_cacheinfo(opt, b->mode, &b->oid, path, 3, 0, options))
694                         return -1;
695         return 0;
696 }
697
698 static void update_entry(struct stage_data *entry,
699                          struct diff_filespec *o,
700                          struct diff_filespec *a,
701                          struct diff_filespec *b)
702 {
703         entry->processed = 0;
704         entry->stages[1].mode = o->mode;
705         entry->stages[2].mode = a->mode;
706         entry->stages[3].mode = b->mode;
707         oidcpy(&entry->stages[1].oid, &o->oid);
708         oidcpy(&entry->stages[2].oid, &a->oid);
709         oidcpy(&entry->stages[3].oid, &b->oid);
710 }
711
712 static int remove_file(struct merge_options *opt, int clean,
713                        const char *path, int no_wd)
714 {
715         int update_cache = opt->call_depth || clean;
716         int update_working_directory = !opt->call_depth && !no_wd;
717
718         if (update_cache) {
719                 if (remove_file_from_index(opt->repo->index, path))
720                         return -1;
721         }
722         if (update_working_directory) {
723                 if (ignore_case) {
724                         struct cache_entry *ce;
725                         ce = index_file_exists(opt->repo->index, path, strlen(path),
726                                                ignore_case);
727                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
728                                 return 0;
729                 }
730                 if (remove_path(path))
731                         return -1;
732         }
733         return 0;
734 }
735
736 /* add a string to a strbuf, but converting "/" to "_" */
737 static void add_flattened_path(struct strbuf *out, const char *s)
738 {
739         size_t i = out->len;
740         strbuf_addstr(out, s);
741         for (; i < out->len; i++)
742                 if (out->buf[i] == '/')
743                         out->buf[i] = '_';
744 }
745
746 static char *unique_path(struct merge_options *opt, const char *path, const char *branch)
747 {
748         struct path_hashmap_entry *entry;
749         struct strbuf newpath = STRBUF_INIT;
750         int suffix = 0;
751         size_t base_len;
752
753         strbuf_addf(&newpath, "%s~", path);
754         add_flattened_path(&newpath, branch);
755
756         base_len = newpath.len;
757         while (hashmap_get_from_hash(&opt->current_file_dir_set,
758                                      path_hash(newpath.buf), newpath.buf) ||
759                (!opt->call_depth && file_exists(newpath.buf))) {
760                 strbuf_setlen(&newpath, base_len);
761                 strbuf_addf(&newpath, "_%d", suffix++);
762         }
763
764         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
765         hashmap_entry_init(entry, path_hash(entry->path));
766         hashmap_add(&opt->current_file_dir_set, entry);
767         return strbuf_detach(&newpath, NULL);
768 }
769
770 /**
771  * Check whether a directory in the index is in the way of an incoming
772  * file.  Return 1 if so.  If check_working_copy is non-zero, also
773  * check the working directory.  If empty_ok is non-zero, also return
774  * 0 in the case where the working-tree dir exists but is empty.
775  */
776 static int dir_in_way(struct index_state *istate, const char *path,
777                       int check_working_copy, int empty_ok)
778 {
779         int pos;
780         struct strbuf dirpath = STRBUF_INIT;
781         struct stat st;
782
783         strbuf_addstr(&dirpath, path);
784         strbuf_addch(&dirpath, '/');
785
786         pos = index_name_pos(istate, dirpath.buf, dirpath.len);
787
788         if (pos < 0)
789                 pos = -1 - pos;
790         if (pos < istate->cache_nr &&
791             !strncmp(dirpath.buf, istate->cache[pos]->name, dirpath.len)) {
792                 strbuf_release(&dirpath);
793                 return 1;
794         }
795
796         strbuf_release(&dirpath);
797         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
798                 !(empty_ok && is_empty_dir(path));
799 }
800
801 /*
802  * Returns whether path was tracked in the index before the merge started,
803  * and its oid and mode match the specified values
804  */
805 static int was_tracked_and_matches(struct merge_options *opt, const char *path,
806                                    const struct object_id *oid, unsigned mode)
807 {
808         int pos = index_name_pos(&opt->orig_index, path, strlen(path));
809         struct cache_entry *ce;
810
811         if (0 > pos)
812                 /* we were not tracking this path before the merge */
813                 return 0;
814
815         /* See if the file we were tracking before matches */
816         ce = opt->orig_index.cache[pos];
817         return (oid_eq(&ce->oid, oid) && ce->ce_mode == mode);
818 }
819
820 /*
821  * Returns whether path was tracked in the index before the merge started
822  */
823 static int was_tracked(struct merge_options *opt, const char *path)
824 {
825         int pos = index_name_pos(&opt->orig_index, path, strlen(path));
826
827         if (0 <= pos)
828                 /* we were tracking this path before the merge */
829                 return 1;
830
831         return 0;
832 }
833
834 static int would_lose_untracked(struct merge_options *opt, const char *path)
835 {
836         struct index_state *istate = opt->repo->index;
837
838         /*
839          * This may look like it can be simplified to:
840          *   return !was_tracked(opt, path) && file_exists(path)
841          * but it can't.  This function needs to know whether path was in
842          * the working tree due to EITHER having been tracked in the index
843          * before the merge OR having been put into the working copy and
844          * index by unpack_trees().  Due to that either-or requirement, we
845          * check the current index instead of the original one.
846          *
847          * Note that we do not need to worry about merge-recursive itself
848          * updating the index after unpack_trees() and before calling this
849          * function, because we strictly require all code paths in
850          * merge-recursive to update the working tree first and the index
851          * second.  Doing otherwise would break
852          * update_file()/would_lose_untracked(); see every comment in this
853          * file which mentions "update_stages".
854          */
855         int pos = index_name_pos(istate, path, strlen(path));
856
857         if (pos < 0)
858                 pos = -1 - pos;
859         while (pos < istate->cache_nr &&
860                !strcmp(path, istate->cache[pos]->name)) {
861                 /*
862                  * If stage #0, it is definitely tracked.
863                  * If it has stage #2 then it was tracked
864                  * before this merge started.  All other
865                  * cases the path was not tracked.
866                  */
867                 switch (ce_stage(istate->cache[pos])) {
868                 case 0:
869                 case 2:
870                         return 0;
871                 }
872                 pos++;
873         }
874         return file_exists(path);
875 }
876
877 static int was_dirty(struct merge_options *opt, const char *path)
878 {
879         struct cache_entry *ce;
880         int dirty = 1;
881
882         if (opt->call_depth || !was_tracked(opt, path))
883                 return !dirty;
884
885         ce = index_file_exists(opt->unpack_opts.src_index,
886                                path, strlen(path), ignore_case);
887         dirty = verify_uptodate(ce, &opt->unpack_opts) != 0;
888         return dirty;
889 }
890
891 static int make_room_for_path(struct merge_options *opt, const char *path)
892 {
893         int status, i;
894         const char *msg = _("failed to create path '%s'%s");
895
896         /* Unlink any D/F conflict files that are in the way */
897         for (i = 0; i < opt->df_conflict_file_set.nr; i++) {
898                 const char *df_path = opt->df_conflict_file_set.items[i].string;
899                 size_t pathlen = strlen(path);
900                 size_t df_pathlen = strlen(df_path);
901                 if (df_pathlen < pathlen &&
902                     path[df_pathlen] == '/' &&
903                     strncmp(path, df_path, df_pathlen) == 0) {
904                         output(opt, 3,
905                                _("Removing %s to make room for subdirectory\n"),
906                                df_path);
907                         unlink(df_path);
908                         unsorted_string_list_delete_item(&opt->df_conflict_file_set,
909                                                          i, 0);
910                         break;
911                 }
912         }
913
914         /* Make sure leading directories are created */
915         status = safe_create_leading_directories_const(path);
916         if (status) {
917                 if (status == SCLD_EXISTS)
918                         /* something else exists */
919                         return err(opt, msg, path, _(": perhaps a D/F conflict?"));
920                 return err(opt, msg, path, "");
921         }
922
923         /*
924          * Do not unlink a file in the work tree if we are not
925          * tracking it.
926          */
927         if (would_lose_untracked(opt, path))
928                 return err(opt, _("refusing to lose untracked file at '%s'"),
929                            path);
930
931         /* Successful unlink is good.. */
932         if (!unlink(path))
933                 return 0;
934         /* .. and so is no existing file */
935         if (errno == ENOENT)
936                 return 0;
937         /* .. but not some other error (who really cares what?) */
938         return err(opt, msg, path, _(": perhaps a D/F conflict?"));
939 }
940
941 static int update_file_flags(struct merge_options *opt,
942                              const struct object_id *oid,
943                              unsigned mode,
944                              const char *path,
945                              int update_cache,
946                              int update_wd)
947 {
948         int ret = 0;
949
950         if (opt->call_depth)
951                 update_wd = 0;
952
953         if (update_wd) {
954                 enum object_type type;
955                 void *buf;
956                 unsigned long size;
957
958                 if (S_ISGITLINK(mode)) {
959                         /*
960                          * We may later decide to recursively descend into
961                          * the submodule directory and update its index
962                          * and/or work tree, but we do not do that now.
963                          */
964                         update_wd = 0;
965                         goto update_index;
966                 }
967
968                 buf = read_object_file(oid, &type, &size);
969                 if (!buf)
970                         return err(opt, _("cannot read object %s '%s'"), oid_to_hex(oid), path);
971                 if (type != OBJ_BLOB) {
972                         ret = err(opt, _("blob expected for %s '%s'"), oid_to_hex(oid), path);
973                         goto free_buf;
974                 }
975                 if (S_ISREG(mode)) {
976                         struct strbuf strbuf = STRBUF_INIT;
977                         if (convert_to_working_tree(opt->repo->index, path, buf, size, &strbuf)) {
978                                 free(buf);
979                                 size = strbuf.len;
980                                 buf = strbuf_detach(&strbuf, NULL);
981                         }
982                 }
983
984                 if (make_room_for_path(opt, path) < 0) {
985                         update_wd = 0;
986                         goto free_buf;
987                 }
988                 if (S_ISREG(mode) || (!has_symlinks && S_ISLNK(mode))) {
989                         int fd;
990                         if (mode & 0100)
991                                 mode = 0777;
992                         else
993                                 mode = 0666;
994                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
995                         if (fd < 0) {
996                                 ret = err(opt, _("failed to open '%s': %s"),
997                                           path, strerror(errno));
998                                 goto free_buf;
999                         }
1000                         write_in_full(fd, buf, size);
1001                         close(fd);
1002                 } else if (S_ISLNK(mode)) {
1003                         char *lnk = xmemdupz(buf, size);
1004                         safe_create_leading_directories_const(path);
1005                         unlink(path);
1006                         if (symlink(lnk, path))
1007                                 ret = err(opt, _("failed to symlink '%s': %s"),
1008                                           path, strerror(errno));
1009                         free(lnk);
1010                 } else
1011                         ret = err(opt,
1012                                   _("do not know what to do with %06o %s '%s'"),
1013                                   mode, oid_to_hex(oid), path);
1014         free_buf:
1015                 free(buf);
1016         }
1017 update_index:
1018         if (!ret && update_cache)
1019                 if (add_cacheinfo(opt, mode, oid, path, 0, update_wd,
1020                                   ADD_CACHE_OK_TO_ADD))
1021                         return -1;
1022         return ret;
1023 }
1024
1025 static int update_file(struct merge_options *opt,
1026                        int clean,
1027                        const struct object_id *oid,
1028                        unsigned mode,
1029                        const char *path)
1030 {
1031         return update_file_flags(opt, oid, mode, path, opt->call_depth || clean, !opt->call_depth);
1032 }
1033
1034 /* Low level file merging, update and removal */
1035
1036 struct merge_file_info {
1037         struct object_id oid;
1038         unsigned mode;
1039         unsigned clean:1,
1040                  merge:1;
1041 };
1042
1043 static int merge_3way(struct merge_options *opt,
1044                       mmbuffer_t *result_buf,
1045                       const struct diff_filespec *o,
1046                       const struct diff_filespec *a,
1047                       const struct diff_filespec *b,
1048                       const char *branch1,
1049                       const char *branch2,
1050                       const int extra_marker_size)
1051 {
1052         mmfile_t orig, src1, src2;
1053         struct ll_merge_options ll_opts = {0};
1054         char *base_name, *name1, *name2;
1055         int merge_status;
1056
1057         ll_opts.renormalize = opt->renormalize;
1058         ll_opts.extra_marker_size = extra_marker_size;
1059         ll_opts.xdl_opts = opt->xdl_opts;
1060
1061         if (opt->call_depth) {
1062                 ll_opts.virtual_ancestor = 1;
1063                 ll_opts.variant = 0;
1064         } else {
1065                 switch (opt->recursive_variant) {
1066                 case MERGE_RECURSIVE_OURS:
1067                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
1068                         break;
1069                 case MERGE_RECURSIVE_THEIRS:
1070                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
1071                         break;
1072                 default:
1073                         ll_opts.variant = 0;
1074                         break;
1075                 }
1076         }
1077
1078         if (strcmp(a->path, b->path) ||
1079             (opt->ancestor != NULL && strcmp(a->path, o->path) != 0)) {
1080                 base_name = opt->ancestor == NULL ? NULL :
1081                         mkpathdup("%s:%s", opt->ancestor, o->path);
1082                 name1 = mkpathdup("%s:%s", branch1, a->path);
1083                 name2 = mkpathdup("%s:%s", branch2, b->path);
1084         } else {
1085                 base_name = opt->ancestor == NULL ? NULL :
1086                         mkpathdup("%s", opt->ancestor);
1087                 name1 = mkpathdup("%s", branch1);
1088                 name2 = mkpathdup("%s", branch2);
1089         }
1090
1091         read_mmblob(&orig, &o->oid);
1092         read_mmblob(&src1, &a->oid);
1093         read_mmblob(&src2, &b->oid);
1094
1095         merge_status = ll_merge(result_buf, a->path, &orig, base_name,
1096                                 &src1, name1, &src2, name2,
1097                                 opt->repo->index, &ll_opts);
1098
1099         free(base_name);
1100         free(name1);
1101         free(name2);
1102         free(orig.ptr);
1103         free(src1.ptr);
1104         free(src2.ptr);
1105         return merge_status;
1106 }
1107
1108 static int find_first_merges(struct repository *repo,
1109                              struct object_array *result, const char *path,
1110                              struct commit *a, struct commit *b)
1111 {
1112         int i, j;
1113         struct object_array merges = OBJECT_ARRAY_INIT;
1114         struct commit *commit;
1115         int contains_another;
1116
1117         char merged_revision[42];
1118         const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
1119                                    "--all", merged_revision, NULL };
1120         struct rev_info revs;
1121         struct setup_revision_opt rev_opts;
1122
1123         memset(result, 0, sizeof(struct object_array));
1124         memset(&rev_opts, 0, sizeof(rev_opts));
1125
1126         /* get all revisions that merge commit a */
1127         xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
1128                   oid_to_hex(&a->object.oid));
1129         repo_init_revisions(repo, &revs, NULL);
1130         rev_opts.submodule = path;
1131         /* FIXME: can't handle linked worktrees in submodules yet */
1132         revs.single_worktree = path != NULL;
1133         setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
1134
1135         /* save all revisions from the above list that contain b */
1136         if (prepare_revision_walk(&revs))
1137                 die("revision walk setup failed");
1138         while ((commit = get_revision(&revs)) != NULL) {
1139                 struct object *o = &(commit->object);
1140                 if (in_merge_bases(b, commit))
1141                         add_object_array(o, NULL, &merges);
1142         }
1143         reset_revision_walk();
1144
1145         /* Now we've got all merges that contain a and b. Prune all
1146          * merges that contain another found merge and save them in
1147          * result.
1148          */
1149         for (i = 0; i < merges.nr; i++) {
1150                 struct commit *m1 = (struct commit *) merges.objects[i].item;
1151
1152                 contains_another = 0;
1153                 for (j = 0; j < merges.nr; j++) {
1154                         struct commit *m2 = (struct commit *) merges.objects[j].item;
1155                         if (i != j && in_merge_bases(m2, m1)) {
1156                                 contains_another = 1;
1157                                 break;
1158                         }
1159                 }
1160
1161                 if (!contains_another)
1162                         add_object_array(merges.objects[i].item, NULL, result);
1163         }
1164
1165         object_array_clear(&merges);
1166         return result->nr;
1167 }
1168
1169 static void print_commit(struct commit *commit)
1170 {
1171         struct strbuf sb = STRBUF_INIT;
1172         struct pretty_print_context ctx = {0};
1173         ctx.date_mode.type = DATE_NORMAL;
1174         format_commit_message(commit, " %h: %m %s", &sb, &ctx);
1175         fprintf(stderr, "%s\n", sb.buf);
1176         strbuf_release(&sb);
1177 }
1178
1179 static int merge_submodule(struct merge_options *opt,
1180                            struct object_id *result, const char *path,
1181                            const struct object_id *base, const struct object_id *a,
1182                            const struct object_id *b)
1183 {
1184         struct commit *commit_base, *commit_a, *commit_b;
1185         int parent_count;
1186         struct object_array merges;
1187
1188         int i;
1189         int search = !opt->call_depth;
1190
1191         /* store a in result in case we fail */
1192         oidcpy(result, a);
1193
1194         /* we can not handle deletion conflicts */
1195         if (is_null_oid(base))
1196                 return 0;
1197         if (is_null_oid(a))
1198                 return 0;
1199         if (is_null_oid(b))
1200                 return 0;
1201
1202         if (add_submodule_odb(path)) {
1203                 output(opt, 1, _("Failed to merge submodule %s (not checked out)"), path);
1204                 return 0;
1205         }
1206
1207         if (!(commit_base = lookup_commit_reference(opt->repo, base)) ||
1208             !(commit_a = lookup_commit_reference(opt->repo, a)) ||
1209             !(commit_b = lookup_commit_reference(opt->repo, b))) {
1210                 output(opt, 1, _("Failed to merge submodule %s (commits not present)"), path);
1211                 return 0;
1212         }
1213
1214         /* check whether both changes are forward */
1215         if (!in_merge_bases(commit_base, commit_a) ||
1216             !in_merge_bases(commit_base, commit_b)) {
1217                 output(opt, 1, _("Failed to merge submodule %s (commits don't follow merge-base)"), path);
1218                 return 0;
1219         }
1220
1221         /* Case #1: a is contained in b or vice versa */
1222         if (in_merge_bases(commit_a, commit_b)) {
1223                 oidcpy(result, b);
1224                 if (show(opt, 3)) {
1225                         output(opt, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1226                         output_commit_title(opt, commit_b);
1227                 } else if (show(opt, 2))
1228                         output(opt, 2, _("Fast-forwarding submodule %s"), path);
1229                 else
1230                         ; /* no output */
1231
1232                 return 1;
1233         }
1234         if (in_merge_bases(commit_b, commit_a)) {
1235                 oidcpy(result, a);
1236                 if (show(opt, 3)) {
1237                         output(opt, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1238                         output_commit_title(opt, commit_a);
1239                 } else if (show(opt, 2))
1240                         output(opt, 2, _("Fast-forwarding submodule %s"), path);
1241                 else
1242                         ; /* no output */
1243
1244                 return 1;
1245         }
1246
1247         /*
1248          * Case #2: There are one or more merges that contain a and b in
1249          * the submodule. If there is only one, then present it as a
1250          * suggestion to the user, but leave it marked unmerged so the
1251          * user needs to confirm the resolution.
1252          */
1253
1254         /* Skip the search if makes no sense to the calling context.  */
1255         if (!search)
1256                 return 0;
1257
1258         /* find commit which merges them */
1259         parent_count = find_first_merges(opt->repo, &merges, path,
1260                                          commit_a, commit_b);
1261         switch (parent_count) {
1262         case 0:
1263                 output(opt, 1, _("Failed to merge submodule %s (merge following commits not found)"), path);
1264                 break;
1265
1266         case 1:
1267                 output(opt, 1, _("Failed to merge submodule %s (not fast-forward)"), path);
1268                 output(opt, 2, _("Found a possible merge resolution for the submodule:\n"));
1269                 print_commit((struct commit *) merges.objects[0].item);
1270                 output(opt, 2, _(
1271                        "If this is correct simply add it to the index "
1272                        "for example\n"
1273                        "by using:\n\n"
1274                        "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
1275                        "which will accept this suggestion.\n"),
1276                        oid_to_hex(&merges.objects[0].item->oid), path);
1277                 break;
1278
1279         default:
1280                 output(opt, 1, _("Failed to merge submodule %s (multiple merges found)"), path);
1281                 for (i = 0; i < merges.nr; i++)
1282                         print_commit((struct commit *) merges.objects[i].item);
1283         }
1284
1285         object_array_clear(&merges);
1286         return 0;
1287 }
1288
1289 static int merge_mode_and_contents(struct merge_options *opt,
1290                                    const struct diff_filespec *o,
1291                                    const struct diff_filespec *a,
1292                                    const struct diff_filespec *b,
1293                                    const char *filename,
1294                                    const char *branch1,
1295                                    const char *branch2,
1296                                    const int extra_marker_size,
1297                                    struct merge_file_info *result)
1298 {
1299         if (opt->branch1 != branch1) {
1300                 /*
1301                  * It's weird getting a reverse merge with HEAD on the bottom
1302                  * side of the conflict markers and the other branch on the
1303                  * top.  Fix that.
1304                  */
1305                 return merge_mode_and_contents(opt, o, b, a,
1306                                                filename,
1307                                                branch2, branch1,
1308                                                extra_marker_size, result);
1309         }
1310
1311         result->merge = 0;
1312         result->clean = 1;
1313
1314         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
1315                 result->clean = 0;
1316                 if (S_ISREG(a->mode)) {
1317                         result->mode = a->mode;
1318                         oidcpy(&result->oid, &a->oid);
1319                 } else {
1320                         result->mode = b->mode;
1321                         oidcpy(&result->oid, &b->oid);
1322                 }
1323         } else {
1324                 if (!oid_eq(&a->oid, &o->oid) && !oid_eq(&b->oid, &o->oid))
1325                         result->merge = 1;
1326
1327                 /*
1328                  * Merge modes
1329                  */
1330                 if (a->mode == b->mode || a->mode == o->mode)
1331                         result->mode = b->mode;
1332                 else {
1333                         result->mode = a->mode;
1334                         if (b->mode != o->mode) {
1335                                 result->clean = 0;
1336                                 result->merge = 1;
1337                         }
1338                 }
1339
1340                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &o->oid))
1341                         oidcpy(&result->oid, &b->oid);
1342                 else if (oid_eq(&b->oid, &o->oid))
1343                         oidcpy(&result->oid, &a->oid);
1344                 else if (S_ISREG(a->mode)) {
1345                         mmbuffer_t result_buf;
1346                         int ret = 0, merge_status;
1347
1348                         merge_status = merge_3way(opt, &result_buf, o, a, b,
1349                                                   branch1, branch2,
1350                                                   extra_marker_size);
1351
1352                         if ((merge_status < 0) || !result_buf.ptr)
1353                                 ret = err(opt, _("Failed to execute internal merge"));
1354
1355                         if (!ret &&
1356                             write_object_file(result_buf.ptr, result_buf.size,
1357                                               blob_type, &result->oid))
1358                                 ret = err(opt, _("Unable to add %s to database"),
1359                                           a->path);
1360
1361                         free(result_buf.ptr);
1362                         if (ret)
1363                                 return ret;
1364                         result->clean = (merge_status == 0);
1365                 } else if (S_ISGITLINK(a->mode)) {
1366                         result->clean = merge_submodule(opt, &result->oid,
1367                                                         o->path,
1368                                                         &o->oid,
1369                                                         &a->oid,
1370                                                         &b->oid);
1371                 } else if (S_ISLNK(a->mode)) {
1372                         switch (opt->recursive_variant) {
1373                         case MERGE_RECURSIVE_NORMAL:
1374                                 oidcpy(&result->oid, &a->oid);
1375                                 if (!oid_eq(&a->oid, &b->oid))
1376                                         result->clean = 0;
1377                                 break;
1378                         case MERGE_RECURSIVE_OURS:
1379                                 oidcpy(&result->oid, &a->oid);
1380                                 break;
1381                         case MERGE_RECURSIVE_THEIRS:
1382                                 oidcpy(&result->oid, &b->oid);
1383                                 break;
1384                         }
1385                 } else
1386                         BUG("unsupported object type in the tree");
1387         }
1388
1389         if (result->merge)
1390                 output(opt, 2, _("Auto-merging %s"), filename);
1391
1392         return 0;
1393 }
1394
1395 static int handle_rename_via_dir(struct merge_options *opt,
1396                                  struct diff_filepair *pair,
1397                                  const char *rename_branch)
1398 {
1399         /*
1400          * Handle file adds that need to be renamed due to directory rename
1401          * detection.  This differs from handle_rename_normal, because
1402          * there is no content merge to do; just move the file into the
1403          * desired final location.
1404          */
1405         const struct diff_filespec *dest = pair->two;
1406
1407         if (!opt->call_depth && would_lose_untracked(opt, dest->path)) {
1408                 char *alt_path = unique_path(opt, dest->path, rename_branch);
1409
1410                 output(opt, 1, _("Error: Refusing to lose untracked file at %s; "
1411                                "writing to %s instead."),
1412                        dest->path, alt_path);
1413                 /*
1414                  * Write the file in worktree at alt_path, but not in the
1415                  * index.  Instead, write to dest->path for the index but
1416                  * only at the higher appropriate stage.
1417                  */
1418                 if (update_file(opt, 0, &dest->oid, dest->mode, alt_path))
1419                         return -1;
1420                 free(alt_path);
1421                 return update_stages(opt, dest->path, NULL,
1422                                      rename_branch == opt->branch1 ? dest : NULL,
1423                                      rename_branch == opt->branch1 ? NULL : dest);
1424         }
1425
1426         /* Update dest->path both in index and in worktree */
1427         if (update_file(opt, 1, &dest->oid, dest->mode, dest->path))
1428                 return -1;
1429         return 0;
1430 }
1431
1432 static int handle_change_delete(struct merge_options *opt,
1433                                 const char *path, const char *old_path,
1434                                 const struct object_id *o_oid, int o_mode,
1435                                 const struct object_id *changed_oid,
1436                                 int changed_mode,
1437                                 const char *change_branch,
1438                                 const char *delete_branch,
1439                                 const char *change, const char *change_past)
1440 {
1441         char *alt_path = NULL;
1442         const char *update_path = path;
1443         int ret = 0;
1444
1445         if (dir_in_way(opt->repo->index, path, !opt->call_depth, 0) ||
1446             (!opt->call_depth && would_lose_untracked(opt, path))) {
1447                 update_path = alt_path = unique_path(opt, path, change_branch);
1448         }
1449
1450         if (opt->call_depth) {
1451                 /*
1452                  * We cannot arbitrarily accept either a_sha or b_sha as
1453                  * correct; since there is no true "middle point" between
1454                  * them, simply reuse the base version for virtual merge base.
1455                  */
1456                 ret = remove_file_from_index(opt->repo->index, path);
1457                 if (!ret)
1458                         ret = update_file(opt, 0, o_oid, o_mode, update_path);
1459         } else {
1460                 /*
1461                  * Despite the four nearly duplicate messages and argument
1462                  * lists below and the ugliness of the nested if-statements,
1463                  * having complete messages makes the job easier for
1464                  * translators.
1465                  *
1466                  * The slight variance among the cases is due to the fact
1467                  * that:
1468                  *   1) directory/file conflicts (in effect if
1469                  *      !alt_path) could cause us to need to write the
1470                  *      file to a different path.
1471                  *   2) renames (in effect if !old_path) could mean that
1472                  *      there are two names for the path that the user
1473                  *      may know the file by.
1474                  */
1475                 if (!alt_path) {
1476                         if (!old_path) {
1477                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1478                                        "and %s in %s. Version %s of %s left in tree."),
1479                                        change, path, delete_branch, change_past,
1480                                        change_branch, change_branch, path);
1481                         } else {
1482                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1483                                        "and %s to %s in %s. Version %s of %s left in tree."),
1484                                        change, old_path, delete_branch, change_past, path,
1485                                        change_branch, change_branch, path);
1486                         }
1487                 } else {
1488                         if (!old_path) {
1489                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1490                                        "and %s in %s. Version %s of %s left in tree at %s."),
1491                                        change, path, delete_branch, change_past,
1492                                        change_branch, change_branch, path, alt_path);
1493                         } else {
1494                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1495                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1496                                        change, old_path, delete_branch, change_past, path,
1497                                        change_branch, change_branch, path, alt_path);
1498                         }
1499                 }
1500                 /*
1501                  * No need to call update_file() on path when change_branch ==
1502                  * opt->branch1 && !alt_path, since that would needlessly touch
1503                  * path.  We could call update_file_flags() with update_cache=0
1504                  * and update_wd=0, but that's a no-op.
1505                  */
1506                 if (change_branch != opt->branch1 || alt_path)
1507                         ret = update_file(opt, 0, changed_oid, changed_mode, update_path);
1508         }
1509         free(alt_path);
1510
1511         return ret;
1512 }
1513
1514 static int handle_rename_delete(struct merge_options *opt,
1515                                 struct diff_filepair *pair,
1516                                 const char *rename_branch,
1517                                 const char *delete_branch)
1518 {
1519         const struct diff_filespec *orig = pair->one;
1520         const struct diff_filespec *dest = pair->two;
1521
1522         if (handle_change_delete(opt,
1523                                  opt->call_depth ? orig->path : dest->path,
1524                                  opt->call_depth ? NULL : orig->path,
1525                                  &orig->oid, orig->mode,
1526                                  &dest->oid, dest->mode,
1527                                  rename_branch, delete_branch,
1528                                  _("rename"), _("renamed")))
1529                 return -1;
1530
1531         if (opt->call_depth)
1532                 return remove_file_from_index(opt->repo->index, dest->path);
1533         else
1534                 return update_stages(opt, dest->path, NULL,
1535                                      rename_branch == opt->branch1 ? dest : NULL,
1536                                      rename_branch == opt->branch1 ? NULL : dest);
1537 }
1538
1539 static struct diff_filespec *filespec_from_entry(struct diff_filespec *target,
1540                                                  struct stage_data *entry,
1541                                                  int stage)
1542 {
1543         struct object_id *oid = &entry->stages[stage].oid;
1544         unsigned mode = entry->stages[stage].mode;
1545         if (mode == 0 || is_null_oid(oid))
1546                 return NULL;
1547         oidcpy(&target->oid, oid);
1548         target->mode = mode;
1549         return target;
1550 }
1551
1552 static int handle_file_collision(struct merge_options *opt,
1553                                  const char *collide_path,
1554                                  const char *prev_path1,
1555                                  const char *prev_path2,
1556                                  const char *branch1, const char *branch2,
1557                                  const struct object_id *a_oid,
1558                                  unsigned int a_mode,
1559                                  const struct object_id *b_oid,
1560                                  unsigned int b_mode)
1561 {
1562         struct merge_file_info mfi;
1563         struct diff_filespec null, a, b;
1564         char *alt_path = NULL;
1565         const char *update_path = collide_path;
1566
1567         /*
1568          * It's easiest to get the correct things into stage 2 and 3, and
1569          * to make sure that the content merge puts HEAD before the other
1570          * branch if we just ensure that branch1 == opt->branch1.  So, simply
1571          * flip arguments around if we don't have that.
1572          */
1573         if (branch1 != opt->branch1) {
1574                 return handle_file_collision(opt, collide_path,
1575                                              prev_path2, prev_path1,
1576                                              branch2, branch1,
1577                                              b_oid, b_mode,
1578                                              a_oid, a_mode);
1579         }
1580
1581         /*
1582          * In the recursive case, we just opt to undo renames
1583          */
1584         if (opt->call_depth && (prev_path1 || prev_path2)) {
1585                 /* Put first file (a_oid, a_mode) in its original spot */
1586                 if (prev_path1) {
1587                         if (update_file(opt, 1, a_oid, a_mode, prev_path1))
1588                                 return -1;
1589                 } else {
1590                         if (update_file(opt, 1, a_oid, a_mode, collide_path))
1591                                 return -1;
1592                 }
1593
1594                 /* Put second file (b_oid, b_mode) in its original spot */
1595                 if (prev_path2) {
1596                         if (update_file(opt, 1, b_oid, b_mode, prev_path2))
1597                                 return -1;
1598                 } else {
1599                         if (update_file(opt, 1, b_oid, b_mode, collide_path))
1600                                 return -1;
1601                 }
1602
1603                 /* Don't leave something at collision path if unrenaming both */
1604                 if (prev_path1 && prev_path2)
1605                         remove_file(opt, 1, collide_path, 0);
1606
1607                 return 0;
1608         }
1609
1610         /* Remove rename sources if rename/add or rename/rename(2to1) */
1611         if (prev_path1)
1612                 remove_file(opt, 1, prev_path1,
1613                             opt->call_depth || would_lose_untracked(opt, prev_path1));
1614         if (prev_path2)
1615                 remove_file(opt, 1, prev_path2,
1616                             opt->call_depth || would_lose_untracked(opt, prev_path2));
1617
1618         /*
1619          * Remove the collision path, if it wouldn't cause dirty contents
1620          * or an untracked file to get lost.  We'll either overwrite with
1621          * merged contents, or just write out to differently named files.
1622          */
1623         if (was_dirty(opt, collide_path)) {
1624                 output(opt, 1, _("Refusing to lose dirty file at %s"),
1625                        collide_path);
1626                 update_path = alt_path = unique_path(opt, collide_path, "merged");
1627         } else if (would_lose_untracked(opt, collide_path)) {
1628                 /*
1629                  * Only way we get here is if both renames were from
1630                  * a directory rename AND user had an untracked file
1631                  * at the location where both files end up after the
1632                  * two directory renames.  See testcase 10d of t6043.
1633                  */
1634                 output(opt, 1, _("Refusing to lose untracked file at "
1635                                "%s, even though it's in the way."),
1636                        collide_path);
1637                 update_path = alt_path = unique_path(opt, collide_path, "merged");
1638         } else {
1639                 /*
1640                  * FIXME: It's possible that the two files are identical
1641                  * and that the current working copy happens to match, in
1642                  * which case we are unnecessarily touching the working
1643                  * tree file.  It's not a likely enough scenario that I
1644                  * want to code up the checks for it and a better fix is
1645                  * available if we restructure how unpack_trees() and
1646                  * merge-recursive interoperate anyway, so punting for
1647                  * now...
1648                  */
1649                 remove_file(opt, 0, collide_path, 0);
1650         }
1651
1652         /* Store things in diff_filespecs for functions that need it */
1653         memset(&a, 0, sizeof(struct diff_filespec));
1654         memset(&b, 0, sizeof(struct diff_filespec));
1655         null.path = a.path = b.path = (char *)collide_path;
1656         oidcpy(&null.oid, &null_oid);
1657         null.mode = 0;
1658         oidcpy(&a.oid, a_oid);
1659         a.mode = a_mode;
1660         a.oid_valid = 1;
1661         oidcpy(&b.oid, b_oid);
1662         b.mode = b_mode;
1663         b.oid_valid = 1;
1664
1665         if (merge_mode_and_contents(opt, &null, &a, &b, collide_path,
1666                                     branch1, branch2, opt->call_depth * 2, &mfi))
1667                 return -1;
1668         mfi.clean &= !alt_path;
1669         if (update_file(opt, mfi.clean, &mfi.oid, mfi.mode, update_path))
1670                 return -1;
1671         if (!mfi.clean && !opt->call_depth &&
1672             update_stages(opt, collide_path, NULL, &a, &b))
1673                 return -1;
1674         free(alt_path);
1675         /*
1676          * FIXME: If both a & b both started with conflicts (only possible
1677          * if they came from a rename/rename(2to1)), but had IDENTICAL
1678          * contents including those conflicts, then in the next line we claim
1679          * it was clean.  If someone cares about this case, we should have the
1680          * caller notify us if we started with conflicts.
1681          */
1682         return mfi.clean;
1683 }
1684
1685 static int handle_rename_add(struct merge_options *opt,
1686                              struct rename_conflict_info *ci)
1687 {
1688         /* a was renamed to c, and a separate c was added. */
1689         struct diff_filespec *a = ci->ren1->pair->one;
1690         struct diff_filespec *c = ci->ren1->pair->two;
1691         char *path = c->path;
1692         char *prev_path_desc;
1693         struct merge_file_info mfi;
1694
1695         int other_stage = (ci->branch1 == opt->branch1 ? 3 : 2);
1696
1697         output(opt, 1, _("CONFLICT (rename/add): "
1698                "Rename %s->%s in %s.  Added %s in %s"),
1699                a->path, c->path, ci->branch1,
1700                c->path, ci->branch2);
1701
1702         prev_path_desc = xstrfmt("version of %s from %s", path, a->path);
1703         if (merge_mode_and_contents(opt, a, c, &ci->ren1_other, prev_path_desc,
1704                                     opt->branch1, opt->branch2,
1705                                     1 + opt->call_depth * 2, &mfi))
1706                 return -1;
1707         free(prev_path_desc);
1708
1709         return handle_file_collision(opt,
1710                                      c->path, a->path, NULL,
1711                                      ci->branch1, ci->branch2,
1712                                      &mfi.oid, mfi.mode,
1713                                      &ci->ren1->dst_entry->stages[other_stage].oid,
1714                                      ci->ren1->dst_entry->stages[other_stage].mode);
1715 }
1716
1717 static char *find_path_for_conflict(struct merge_options *opt,
1718                                     const char *path,
1719                                     const char *branch1,
1720                                     const char *branch2)
1721 {
1722         char *new_path = NULL;
1723         if (dir_in_way(opt->repo->index, path, !opt->call_depth, 0)) {
1724                 new_path = unique_path(opt, path, branch1);
1725                 output(opt, 1, _("%s is a directory in %s adding "
1726                                "as %s instead"),
1727                        path, branch2, new_path);
1728         } else if (would_lose_untracked(opt, path)) {
1729                 new_path = unique_path(opt, path, branch1);
1730                 output(opt, 1, _("Refusing to lose untracked file"
1731                                " at %s; adding as %s instead"),
1732                        path, new_path);
1733         }
1734
1735         return new_path;
1736 }
1737
1738 static int handle_rename_rename_1to2(struct merge_options *opt,
1739                                      struct rename_conflict_info *ci)
1740 {
1741         /* One file was renamed in both branches, but to different names. */
1742         struct merge_file_info mfi;
1743         struct diff_filespec other;
1744         struct diff_filespec *add;
1745         struct diff_filespec *o = ci->ren1->pair->one;
1746         struct diff_filespec *a = ci->ren1->pair->two;
1747         struct diff_filespec *b = ci->ren2->pair->two;
1748         char *path_desc;
1749
1750         output(opt, 1, _("CONFLICT (rename/rename): "
1751                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1752                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1753                o->path, a->path, ci->branch1,
1754                o->path, b->path, ci->branch2,
1755                opt->call_depth ? _(" (left unresolved)") : "");
1756
1757         path_desc = xstrfmt("%s and %s, both renamed from %s",
1758                             a->path, b->path, o->path);
1759         if (merge_mode_and_contents(opt, o, a, b, path_desc,
1760                                     ci->branch1, ci->branch2,
1761                                     opt->call_depth * 2, &mfi))
1762                 return -1;
1763         free(path_desc);
1764
1765         if (opt->call_depth) {
1766                 /*
1767                  * FIXME: For rename/add-source conflicts (if we could detect
1768                  * such), this is wrong.  We should instead find a unique
1769                  * pathname and then either rename the add-source file to that
1770                  * unique path, or use that unique path instead of src here.
1771                  */
1772                 if (update_file(opt, 0, &mfi.oid, mfi.mode, o->path))
1773                         return -1;
1774
1775                 /*
1776                  * Above, we put the merged content at the merge-base's
1777                  * path.  Now we usually need to delete both a->path and
1778                  * b->path.  However, the rename on each side of the merge
1779                  * could also be involved in a rename/add conflict.  In
1780                  * such cases, we should keep the added file around,
1781                  * resolving the conflict at that path in its favor.
1782                  */
1783                 add = filespec_from_entry(&other, ci->ren1->dst_entry, 2 ^ 1);
1784                 if (add) {
1785                         if (update_file(opt, 0, &add->oid, add->mode, a->path))
1786                                 return -1;
1787                 }
1788                 else
1789                         remove_file_from_index(opt->repo->index, a->path);
1790                 add = filespec_from_entry(&other, ci->ren2->dst_entry, 3 ^ 1);
1791                 if (add) {
1792                         if (update_file(opt, 0, &add->oid, add->mode, b->path))
1793                                 return -1;
1794                 }
1795                 else
1796                         remove_file_from_index(opt->repo->index, b->path);
1797         } else {
1798                 /*
1799                  * For each destination path, we need to see if there is a
1800                  * rename/add collision.  If not, we can write the file out
1801                  * to the specified location.
1802                  */
1803                 add = filespec_from_entry(&other, ci->ren1->dst_entry, 2 ^ 1);
1804                 if (add) {
1805                         if (handle_file_collision(opt, a->path,
1806                                                   NULL, NULL,
1807                                                   ci->branch1, ci->branch2,
1808                                                   &mfi.oid, mfi.mode,
1809                                                   &add->oid, add->mode) < 0)
1810                                 return -1;
1811                 } else {
1812                         char *new_path = find_path_for_conflict(opt, a->path,
1813                                                                 ci->branch1,
1814                                                                 ci->branch2);
1815                         if (update_file(opt, 0, &mfi.oid, mfi.mode, new_path ? new_path : a->path))
1816                                 return -1;
1817                         free(new_path);
1818                         if (update_stages(opt, a->path, NULL, a, NULL))
1819                                 return -1;
1820                 }
1821
1822                 add = filespec_from_entry(&other, ci->ren2->dst_entry, 3 ^ 1);
1823                 if (add) {
1824                         if (handle_file_collision(opt, b->path,
1825                                                   NULL, NULL,
1826                                                   ci->branch1, ci->branch2,
1827                                                   &add->oid, add->mode,
1828                                                   &mfi.oid, mfi.mode) < 0)
1829                                 return -1;
1830                 } else {
1831                         char *new_path = find_path_for_conflict(opt, b->path,
1832                                                                 ci->branch2,
1833                                                                 ci->branch1);
1834                         if (update_file(opt, 0, &mfi.oid, mfi.mode, new_path ? new_path : b->path))
1835                                 return -1;
1836                         free(new_path);
1837                         if (update_stages(opt, b->path, NULL, NULL, b))
1838                                 return -1;
1839                 }
1840         }
1841
1842         return 0;
1843 }
1844
1845 static int handle_rename_rename_2to1(struct merge_options *opt,
1846                                      struct rename_conflict_info *ci)
1847 {
1848         /* Two files, a & b, were renamed to the same thing, c. */
1849         struct diff_filespec *a = ci->ren1->pair->one;
1850         struct diff_filespec *b = ci->ren2->pair->one;
1851         struct diff_filespec *c1 = ci->ren1->pair->two;
1852         struct diff_filespec *c2 = ci->ren2->pair->two;
1853         char *path = c1->path; /* == c2->path */
1854         char *path_side_1_desc;
1855         char *path_side_2_desc;
1856         struct merge_file_info mfi_c1;
1857         struct merge_file_info mfi_c2;
1858
1859         output(opt, 1, _("CONFLICT (rename/rename): "
1860                "Rename %s->%s in %s. "
1861                "Rename %s->%s in %s"),
1862                a->path, c1->path, ci->branch1,
1863                b->path, c2->path, ci->branch2);
1864
1865         path_side_1_desc = xstrfmt("version of %s from %s", path, a->path);
1866         path_side_2_desc = xstrfmt("version of %s from %s", path, b->path);
1867         if (merge_mode_and_contents(opt, a, c1, &ci->ren1_other, path_side_1_desc,
1868                                     opt->branch1, opt->branch2,
1869                                     1 + opt->call_depth * 2, &mfi_c1) ||
1870             merge_mode_and_contents(opt, b, &ci->ren2_other, c2, path_side_2_desc,
1871                                     opt->branch1, opt->branch2,
1872                                     1 + opt->call_depth * 2, &mfi_c2))
1873                 return -1;
1874         free(path_side_1_desc);
1875         free(path_side_2_desc);
1876
1877         return handle_file_collision(opt, path, a->path, b->path,
1878                                      ci->branch1, ci->branch2,
1879                                      &mfi_c1.oid, mfi_c1.mode,
1880                                      &mfi_c2.oid, mfi_c2.mode);
1881 }
1882
1883 /*
1884  * Get the diff_filepairs changed between o_tree and tree.
1885  */
1886 static struct diff_queue_struct *get_diffpairs(struct merge_options *opt,
1887                                                struct tree *o_tree,
1888                                                struct tree *tree)
1889 {
1890         struct diff_queue_struct *ret;
1891         struct diff_options opts;
1892
1893         repo_diff_setup(opt->repo, &opts);
1894         opts.flags.recursive = 1;
1895         opts.flags.rename_empty = 0;
1896         opts.detect_rename = merge_detect_rename(opt);
1897         /*
1898          * We do not have logic to handle the detection of copies.  In
1899          * fact, it may not even make sense to add such logic: would we
1900          * really want a change to a base file to be propagated through
1901          * multiple other files by a merge?
1902          */
1903         if (opts.detect_rename > DIFF_DETECT_RENAME)
1904                 opts.detect_rename = DIFF_DETECT_RENAME;
1905         opts.rename_limit = opt->merge_rename_limit >= 0 ? opt->merge_rename_limit :
1906                             opt->diff_rename_limit >= 0 ? opt->diff_rename_limit :
1907                             1000;
1908         opts.rename_score = opt->rename_score;
1909         opts.show_rename_progress = opt->show_rename_progress;
1910         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1911         diff_setup_done(&opts);
1912         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1913         diffcore_std(&opts);
1914         if (opts.needed_rename_limit > opt->needed_rename_limit)
1915                 opt->needed_rename_limit = opts.needed_rename_limit;
1916
1917         ret = xmalloc(sizeof(*ret));
1918         *ret = diff_queued_diff;
1919
1920         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1921         diff_queued_diff.nr = 0;
1922         diff_queued_diff.queue = NULL;
1923         diff_flush(&opts);
1924         return ret;
1925 }
1926
1927 static int tree_has_path(struct tree *tree, const char *path)
1928 {
1929         struct object_id hashy;
1930         unsigned short mode_o;
1931
1932         return !get_tree_entry(&tree->object.oid, path,
1933                                &hashy, &mode_o);
1934 }
1935
1936 /*
1937  * Return a new string that replaces the beginning portion (which matches
1938  * entry->dir), with entry->new_dir.  In perl-speak:
1939  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1940  * NOTE:
1941  *   Caller must ensure that old_path starts with entry->dir + '/'.
1942  */
1943 static char *apply_dir_rename(struct dir_rename_entry *entry,
1944                               const char *old_path)
1945 {
1946         struct strbuf new_path = STRBUF_INIT;
1947         int oldlen, newlen;
1948
1949         if (entry->non_unique_new_dir)
1950                 return NULL;
1951
1952         oldlen = strlen(entry->dir);
1953         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1954         strbuf_grow(&new_path, newlen);
1955         strbuf_addbuf(&new_path, &entry->new_dir);
1956         strbuf_addstr(&new_path, &old_path[oldlen]);
1957
1958         return strbuf_detach(&new_path, NULL);
1959 }
1960
1961 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1962                                     char **old_dir, char **new_dir)
1963 {
1964         char *end_of_old, *end_of_new;
1965         int old_len, new_len;
1966
1967         *old_dir = NULL;
1968         *new_dir = NULL;
1969
1970         /*
1971          * For
1972          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1973          * the "e/foo.c" part is the same, we just want to know that
1974          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1975          * so, for this example, this function returns "a/b/c/d" in
1976          * *old_dir and "a/b/some/thing/else" in *new_dir.
1977          *
1978          * Also, if the basename of the file changed, we don't care.  We
1979          * want to know which portion of the directory, if any, changed.
1980          */
1981         end_of_old = strrchr(old_path, '/');
1982         end_of_new = strrchr(new_path, '/');
1983
1984         if (end_of_old == NULL || end_of_new == NULL)
1985                 return;
1986         while (*--end_of_new == *--end_of_old &&
1987                end_of_old != old_path &&
1988                end_of_new != new_path)
1989                 ; /* Do nothing; all in the while loop */
1990         /*
1991          * We've found the first non-matching character in the directory
1992          * paths.  That means the current directory we were comparing
1993          * represents the rename.  Move end_of_old and end_of_new back
1994          * to the full directory name.
1995          */
1996         if (*end_of_old == '/')
1997                 end_of_old++;
1998         if (*end_of_old != '/')
1999                 end_of_new++;
2000         end_of_old = strchr(end_of_old, '/');
2001         end_of_new = strchr(end_of_new, '/');
2002
2003         /*
2004          * It may have been the case that old_path and new_path were the same
2005          * directory all along.  Don't claim a rename if they're the same.
2006          */
2007         old_len = end_of_old - old_path;
2008         new_len = end_of_new - new_path;
2009
2010         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
2011                 *old_dir = xstrndup(old_path, old_len);
2012                 *new_dir = xstrndup(new_path, new_len);
2013         }
2014 }
2015
2016 static void remove_hashmap_entries(struct hashmap *dir_renames,
2017                                    struct string_list *items_to_remove)
2018 {
2019         int i;
2020         struct dir_rename_entry *entry;
2021
2022         for (i = 0; i < items_to_remove->nr; i++) {
2023                 entry = items_to_remove->items[i].util;
2024                 hashmap_remove(dir_renames, entry, NULL);
2025         }
2026         string_list_clear(items_to_remove, 0);
2027 }
2028
2029 /*
2030  * See if there is a directory rename for path, and if there are any file
2031  * level conflicts for the renamed location.  If there is a rename and
2032  * there are no conflicts, return the new name.  Otherwise, return NULL.
2033  */
2034 static char *handle_path_level_conflicts(struct merge_options *opt,
2035                                          const char *path,
2036                                          struct dir_rename_entry *entry,
2037                                          struct hashmap *collisions,
2038                                          struct tree *tree)
2039 {
2040         char *new_path = NULL;
2041         struct collision_entry *collision_ent;
2042         int clean = 1;
2043         struct strbuf collision_paths = STRBUF_INIT;
2044
2045         /*
2046          * entry has the mapping of old directory name to new directory name
2047          * that we want to apply to path.
2048          */
2049         new_path = apply_dir_rename(entry, path);
2050
2051         if (!new_path) {
2052                 /* This should only happen when entry->non_unique_new_dir set */
2053                 if (!entry->non_unique_new_dir)
2054                         BUG("entry->non_unqiue_dir not set and !new_path");
2055                 output(opt, 1, _("CONFLICT (directory rename split): "
2056                                "Unclear where to place %s because directory "
2057                                "%s was renamed to multiple other directories, "
2058                                "with no destination getting a majority of the "
2059                                "files."),
2060                        path, entry->dir);
2061                 clean = 0;
2062                 return NULL;
2063         }
2064
2065         /*
2066          * The caller needs to have ensured that it has pre-populated
2067          * collisions with all paths that map to new_path.  Do a quick check
2068          * to ensure that's the case.
2069          */
2070         collision_ent = collision_find_entry(collisions, new_path);
2071         if (collision_ent == NULL)
2072                 BUG("collision_ent is NULL");
2073
2074         /*
2075          * Check for one-sided add/add/.../add conflicts, i.e.
2076          * where implicit renames from the other side doing
2077          * directory rename(s) can affect this side of history
2078          * to put multiple paths into the same location.  Warn
2079          * and bail on directory renames for such paths.
2080          */
2081         if (collision_ent->reported_already) {
2082                 clean = 0;
2083         } else if (tree_has_path(tree, new_path)) {
2084                 collision_ent->reported_already = 1;
2085                 strbuf_add_separated_string_list(&collision_paths, ", ",
2086                                                  &collision_ent->source_files);
2087                 output(opt, 1, _("CONFLICT (implicit dir rename): Existing "
2088                                "file/dir at %s in the way of implicit "
2089                                "directory rename(s) putting the following "
2090                                "path(s) there: %s."),
2091                        new_path, collision_paths.buf);
2092                 clean = 0;
2093         } else if (collision_ent->source_files.nr > 1) {
2094                 collision_ent->reported_already = 1;
2095                 strbuf_add_separated_string_list(&collision_paths, ", ",
2096                                                  &collision_ent->source_files);
2097                 output(opt, 1, _("CONFLICT (implicit dir rename): Cannot map "
2098                                "more than one path to %s; implicit directory "
2099                                "renames tried to put these paths there: %s"),
2100                        new_path, collision_paths.buf);
2101                 clean = 0;
2102         }
2103
2104         /* Free memory we no longer need */
2105         strbuf_release(&collision_paths);
2106         if (!clean && new_path) {
2107                 free(new_path);
2108                 return NULL;
2109         }
2110
2111         return new_path;
2112 }
2113
2114 /*
2115  * There are a couple things we want to do at the directory level:
2116  *   1. Check for both sides renaming to the same thing, in order to avoid
2117  *      implicit renaming of files that should be left in place.  (See
2118  *      testcase 6b in t6043 for details.)
2119  *   2. Prune directory renames if there are still files left in the
2120  *      the original directory.  These represent a partial directory rename,
2121  *      i.e. a rename where only some of the files within the directory
2122  *      were renamed elsewhere.  (Technically, this could be done earlier
2123  *      in get_directory_renames(), except that would prevent us from
2124  *      doing the previous check and thus failing testcase 6b.)
2125  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2126  *      In the future, we could potentially record this info as well and
2127  *      omit reporting rename/rename(1to2) conflicts for each path within
2128  *      the affected directories, thus cleaning up the merge output.
2129  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2130  *         directory level, because merging directories is fine.  If it
2131  *         causes conflicts for files within those merged directories, then
2132  *         that should be detected at the individual path level.
2133  */
2134 static void handle_directory_level_conflicts(struct merge_options *opt,
2135                                              struct hashmap *dir_re_head,
2136                                              struct tree *head,
2137                                              struct hashmap *dir_re_merge,
2138                                              struct tree *merge)
2139 {
2140         struct hashmap_iter iter;
2141         struct dir_rename_entry *head_ent;
2142         struct dir_rename_entry *merge_ent;
2143
2144         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2145         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2146
2147         hashmap_iter_init(dir_re_head, &iter);
2148         while ((head_ent = hashmap_iter_next(&iter))) {
2149                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2150                 if (merge_ent &&
2151                     !head_ent->non_unique_new_dir &&
2152                     !merge_ent->non_unique_new_dir &&
2153                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2154                         /* 1. Renamed identically; remove it from both sides */
2155                         string_list_append(&remove_from_head,
2156                                            head_ent->dir)->util = head_ent;
2157                         strbuf_release(&head_ent->new_dir);
2158                         string_list_append(&remove_from_merge,
2159                                            merge_ent->dir)->util = merge_ent;
2160                         strbuf_release(&merge_ent->new_dir);
2161                 } else if (tree_has_path(head, head_ent->dir)) {
2162                         /* 2. This wasn't a directory rename after all */
2163                         string_list_append(&remove_from_head,
2164                                            head_ent->dir)->util = head_ent;
2165                         strbuf_release(&head_ent->new_dir);
2166                 }
2167         }
2168
2169         remove_hashmap_entries(dir_re_head, &remove_from_head);
2170         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2171
2172         hashmap_iter_init(dir_re_merge, &iter);
2173         while ((merge_ent = hashmap_iter_next(&iter))) {
2174                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2175                 if (tree_has_path(merge, merge_ent->dir)) {
2176                         /* 2. This wasn't a directory rename after all */
2177                         string_list_append(&remove_from_merge,
2178                                            merge_ent->dir)->util = merge_ent;
2179                 } else if (head_ent &&
2180                            !head_ent->non_unique_new_dir &&
2181                            !merge_ent->non_unique_new_dir) {
2182                         /* 3. rename/rename(1to2) */
2183                         /*
2184                          * We can assume it's not rename/rename(1to1) because
2185                          * that was case (1), already checked above.  So we
2186                          * know that head_ent->new_dir and merge_ent->new_dir
2187                          * are different strings.
2188                          */
2189                         output(opt, 1, _("CONFLICT (rename/rename): "
2190                                        "Rename directory %s->%s in %s. "
2191                                        "Rename directory %s->%s in %s"),
2192                                head_ent->dir, head_ent->new_dir.buf, opt->branch1,
2193                                head_ent->dir, merge_ent->new_dir.buf, opt->branch2);
2194                         string_list_append(&remove_from_head,
2195                                            head_ent->dir)->util = head_ent;
2196                         strbuf_release(&head_ent->new_dir);
2197                         string_list_append(&remove_from_merge,
2198                                            merge_ent->dir)->util = merge_ent;
2199                         strbuf_release(&merge_ent->new_dir);
2200                 }
2201         }
2202
2203         remove_hashmap_entries(dir_re_head, &remove_from_head);
2204         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2205 }
2206
2207 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs)
2208 {
2209         struct hashmap *dir_renames;
2210         struct hashmap_iter iter;
2211         struct dir_rename_entry *entry;
2212         int i;
2213
2214         /*
2215          * Typically, we think of a directory rename as all files from a
2216          * certain directory being moved to a target directory.  However,
2217          * what if someone first moved two files from the original
2218          * directory in one commit, and then renamed the directory
2219          * somewhere else in a later commit?  At merge time, we just know
2220          * that files from the original directory went to two different
2221          * places, and that the bulk of them ended up in the same place.
2222          * We want each directory rename to represent where the bulk of the
2223          * files from that directory end up; this function exists to find
2224          * where the bulk of the files went.
2225          *
2226          * The first loop below simply iterates through the list of file
2227          * renames, finding out how often each directory rename pair
2228          * possibility occurs.
2229          */
2230         dir_renames = xmalloc(sizeof(*dir_renames));
2231         dir_rename_init(dir_renames);
2232         for (i = 0; i < pairs->nr; ++i) {
2233                 struct string_list_item *item;
2234                 int *count;
2235                 struct diff_filepair *pair = pairs->queue[i];
2236                 char *old_dir, *new_dir;
2237
2238                 /* File not part of directory rename if it wasn't renamed */
2239                 if (pair->status != 'R')
2240                         continue;
2241
2242                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2243                                         &old_dir,        &new_dir);
2244                 if (!old_dir)
2245                         /* Directory didn't change at all; ignore this one. */
2246                         continue;
2247
2248                 entry = dir_rename_find_entry(dir_renames, old_dir);
2249                 if (!entry) {
2250                         entry = xmalloc(sizeof(*entry));
2251                         dir_rename_entry_init(entry, old_dir);
2252                         hashmap_put(dir_renames, entry);
2253                 } else {
2254                         free(old_dir);
2255                 }
2256                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2257                 if (!item) {
2258                         item = string_list_insert(&entry->possible_new_dirs,
2259                                                   new_dir);
2260                         item->util = xcalloc(1, sizeof(int));
2261                 } else {
2262                         free(new_dir);
2263                 }
2264                 count = item->util;
2265                 *count += 1;
2266         }
2267
2268         /*
2269          * For each directory with files moved out of it, we find out which
2270          * target directory received the most files so we can declare it to
2271          * be the "winning" target location for the directory rename.  This
2272          * winner gets recorded in new_dir.  If there is no winner
2273          * (multiple target directories received the same number of files),
2274          * we set non_unique_new_dir.  Once we've determined the winner (or
2275          * that there is no winner), we no longer need possible_new_dirs.
2276          */
2277         hashmap_iter_init(dir_renames, &iter);
2278         while ((entry = hashmap_iter_next(&iter))) {
2279                 int max = 0;
2280                 int bad_max = 0;
2281                 char *best = NULL;
2282
2283                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2284                         int *count = entry->possible_new_dirs.items[i].util;
2285
2286                         if (*count == max)
2287                                 bad_max = max;
2288                         else if (*count > max) {
2289                                 max = *count;
2290                                 best = entry->possible_new_dirs.items[i].string;
2291                         }
2292                 }
2293                 if (bad_max == max)
2294                         entry->non_unique_new_dir = 1;
2295                 else {
2296                         assert(entry->new_dir.len == 0);
2297                         strbuf_addstr(&entry->new_dir, best);
2298                 }
2299                 /*
2300                  * The relevant directory sub-portion of the original full
2301                  * filepaths were xstrndup'ed before inserting into
2302                  * possible_new_dirs, and instead of manually iterating the
2303                  * list and free'ing each, just lie and tell
2304                  * possible_new_dirs that it did the strdup'ing so that it
2305                  * will free them for us.
2306                  */
2307                 entry->possible_new_dirs.strdup_strings = 1;
2308                 string_list_clear(&entry->possible_new_dirs, 1);
2309         }
2310
2311         return dir_renames;
2312 }
2313
2314 static struct dir_rename_entry *check_dir_renamed(const char *path,
2315                                                   struct hashmap *dir_renames)
2316 {
2317         char *temp = xstrdup(path);
2318         char *end;
2319         struct dir_rename_entry *entry = NULL;
2320
2321         while ((end = strrchr(temp, '/'))) {
2322                 *end = '\0';
2323                 entry = dir_rename_find_entry(dir_renames, temp);
2324                 if (entry)
2325                         break;
2326         }
2327         free(temp);
2328         return entry;
2329 }
2330
2331 static void compute_collisions(struct hashmap *collisions,
2332                                struct hashmap *dir_renames,
2333                                struct diff_queue_struct *pairs)
2334 {
2335         int i;
2336
2337         /*
2338          * Multiple files can be mapped to the same path due to directory
2339          * renames done by the other side of history.  Since that other
2340          * side of history could have merged multiple directories into one,
2341          * if our side of history added the same file basename to each of
2342          * those directories, then all N of them would get implicitly
2343          * renamed by the directory rename detection into the same path,
2344          * and we'd get an add/add/.../add conflict, and all those adds
2345          * from *this* side of history.  This is not representable in the
2346          * index, and users aren't going to easily be able to make sense of
2347          * it.  So we need to provide a good warning about what's
2348          * happening, and fall back to no-directory-rename detection
2349          * behavior for those paths.
2350          *
2351          * See testcases 9e and all of section 5 from t6043 for examples.
2352          */
2353         collision_init(collisions);
2354
2355         for (i = 0; i < pairs->nr; ++i) {
2356                 struct dir_rename_entry *dir_rename_ent;
2357                 struct collision_entry *collision_ent;
2358                 char *new_path;
2359                 struct diff_filepair *pair = pairs->queue[i];
2360
2361                 if (pair->status != 'A' && pair->status != 'R')
2362                         continue;
2363                 dir_rename_ent = check_dir_renamed(pair->two->path,
2364                                                    dir_renames);
2365                 if (!dir_rename_ent)
2366                         continue;
2367
2368                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2369                 if (!new_path)
2370                         /*
2371                          * dir_rename_ent->non_unique_new_path is true, which
2372                          * means there is no directory rename for us to use,
2373                          * which means it won't cause us any additional
2374                          * collisions.
2375                          */
2376                         continue;
2377                 collision_ent = collision_find_entry(collisions, new_path);
2378                 if (!collision_ent) {
2379                         collision_ent = xcalloc(1,
2380                                                 sizeof(struct collision_entry));
2381                         hashmap_entry_init(collision_ent, strhash(new_path));
2382                         hashmap_put(collisions, collision_ent);
2383                         collision_ent->target_file = new_path;
2384                 } else {
2385                         free(new_path);
2386                 }
2387                 string_list_insert(&collision_ent->source_files,
2388                                    pair->two->path);
2389         }
2390 }
2391
2392 static char *check_for_directory_rename(struct merge_options *opt,
2393                                         const char *path,
2394                                         struct tree *tree,
2395                                         struct hashmap *dir_renames,
2396                                         struct hashmap *dir_rename_exclusions,
2397                                         struct hashmap *collisions,
2398                                         int *clean_merge)
2399 {
2400         char *new_path = NULL;
2401         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2402         struct dir_rename_entry *oentry = NULL;
2403
2404         if (!entry)
2405                 return new_path;
2406
2407         /*
2408          * This next part is a little weird.  We do not want to do an
2409          * implicit rename into a directory we renamed on our side, because
2410          * that will result in a spurious rename/rename(1to2) conflict.  An
2411          * example:
2412          *   Base commit: dumbdir/afile, otherdir/bfile
2413          *   Side 1:      smrtdir/afile, otherdir/bfile
2414          *   Side 2:      dumbdir/afile, dumbdir/bfile
2415          * Here, while working on Side 1, we could notice that otherdir was
2416          * renamed/merged to dumbdir, and change the diff_filepair for
2417          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2418          * 2 will notice the rename from dumbdir to smrtdir, and do the
2419          * transitive rename to move it from dumbdir/bfile to
2420          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2421          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2422          * the file to end up in smrtdir.  And the way to achieve that is
2423          * to not let Side1 do the rename to dumbdir, since we know that is
2424          * the source of one of our directory renames.
2425          *
2426          * That's why oentry and dir_rename_exclusions is here.
2427          *
2428          * As it turns out, this also prevents N-way transient rename
2429          * confusion; See testcases 9c and 9d of t6043.
2430          */
2431         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2432         if (oentry) {
2433                 output(opt, 1, _("WARNING: Avoiding applying %s -> %s rename "
2434                                "to %s, because %s itself was renamed."),
2435                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2436         } else {
2437                 new_path = handle_path_level_conflicts(opt, path, entry,
2438                                                        collisions, tree);
2439                 *clean_merge &= (new_path != NULL);
2440         }
2441
2442         return new_path;
2443 }
2444
2445 static void apply_directory_rename_modifications(struct merge_options *opt,
2446                                                  struct diff_filepair *pair,
2447                                                  char *new_path,
2448                                                  struct rename *re,
2449                                                  struct tree *tree,
2450                                                  struct tree *o_tree,
2451                                                  struct tree *a_tree,
2452                                                  struct tree *b_tree,
2453                                                  struct string_list *entries)
2454 {
2455         struct string_list_item *item;
2456         int stage = (tree == a_tree ? 2 : 3);
2457         int update_wd;
2458
2459         /*
2460          * In all cases where we can do directory rename detection,
2461          * unpack_trees() will have read pair->two->path into the
2462          * index and the working copy.  We need to remove it so that
2463          * we can instead place it at new_path.  It is guaranteed to
2464          * not be untracked (unpack_trees() would have errored out
2465          * saying the file would have been overwritten), but it might
2466          * be dirty, though.
2467          */
2468         update_wd = !was_dirty(opt, pair->two->path);
2469         if (!update_wd)
2470                 output(opt, 1, _("Refusing to lose dirty file at %s"),
2471                        pair->two->path);
2472         remove_file(opt, 1, pair->two->path, !update_wd);
2473
2474         /* Find or create a new re->dst_entry */
2475         item = string_list_lookup(entries, new_path);
2476         if (item) {
2477                 /*
2478                  * Since we're renaming on this side of history, and it's
2479                  * due to a directory rename on the other side of history
2480                  * (which we only allow when the directory in question no
2481                  * longer exists on the other side of history), the
2482                  * original entry for re->dst_entry is no longer
2483                  * necessary...
2484                  */
2485                 re->dst_entry->processed = 1;
2486
2487                 /*
2488                  * ...because we'll be using this new one.
2489                  */
2490                 re->dst_entry = item->util;
2491         } else {
2492                 /*
2493                  * re->dst_entry is for the before-dir-rename path, and we
2494                  * need it to hold information for the after-dir-rename
2495                  * path.  Before creating a new entry, we need to mark the
2496                  * old one as unnecessary (...unless it is shared by
2497                  * src_entry, i.e. this didn't use to be a rename, in which
2498                  * case we can just allow the normal processing to happen
2499                  * for it).
2500                  */
2501                 if (pair->status == 'R')
2502                         re->dst_entry->processed = 1;
2503
2504                 re->dst_entry = insert_stage_data(new_path,
2505                                                   o_tree, a_tree, b_tree,
2506                                                   entries);
2507                 item = string_list_insert(entries, new_path);
2508                 item->util = re->dst_entry;
2509         }
2510
2511         /*
2512          * Update the stage_data with the information about the path we are
2513          * moving into place.  That slot will be empty and available for us
2514          * to write to because of the collision checks in
2515          * handle_path_level_conflicts().  In other words,
2516          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2517          * open for us to write to.
2518          *
2519          * It may be tempting to actually update the index at this point as
2520          * well, using update_stages_for_stage_data(), but as per the big
2521          * "NOTE" in update_stages(), doing so will modify the current
2522          * in-memory index which will break calls to would_lose_untracked()
2523          * that we need to make.  Instead, we need to just make sure that
2524          * the various handle_rename_*() functions update the index
2525          * explicitly rather than relying on unpack_trees() to have done it.
2526          */
2527         get_tree_entry(&tree->object.oid,
2528                        pair->two->path,
2529                        &re->dst_entry->stages[stage].oid,
2530                        &re->dst_entry->stages[stage].mode);
2531
2532         /* Update pair status */
2533         if (pair->status == 'A') {
2534                 /*
2535                  * Recording rename information for this add makes it look
2536                  * like a rename/delete conflict.  Make sure we can
2537                  * correctly handle this as an add that was moved to a new
2538                  * directory instead of reporting a rename/delete conflict.
2539                  */
2540                 re->add_turned_into_rename = 1;
2541         }
2542         /*
2543          * We don't actually look at pair->status again, but it seems
2544          * pedagogically correct to adjust it.
2545          */
2546         pair->status = 'R';
2547
2548         /*
2549          * Finally, record the new location.
2550          */
2551         pair->two->path = new_path;
2552 }
2553
2554 /*
2555  * Get information of all renames which occurred in 'pairs', making use of
2556  * any implicit directory renames inferred from the other side of history.
2557  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2558  * to be able to associate the correct cache entries with the rename
2559  * information; tree is always equal to either a_tree or b_tree.
2560  */
2561 static struct string_list *get_renames(struct merge_options *opt,
2562                                        struct diff_queue_struct *pairs,
2563                                        struct hashmap *dir_renames,
2564                                        struct hashmap *dir_rename_exclusions,
2565                                        struct tree *tree,
2566                                        struct tree *o_tree,
2567                                        struct tree *a_tree,
2568                                        struct tree *b_tree,
2569                                        struct string_list *entries,
2570                                        int *clean_merge)
2571 {
2572         int i;
2573         struct hashmap collisions;
2574         struct hashmap_iter iter;
2575         struct collision_entry *e;
2576         struct string_list *renames;
2577
2578         compute_collisions(&collisions, dir_renames, pairs);
2579         renames = xcalloc(1, sizeof(struct string_list));
2580
2581         for (i = 0; i < pairs->nr; ++i) {
2582                 struct string_list_item *item;
2583                 struct rename *re;
2584                 struct diff_filepair *pair = pairs->queue[i];
2585                 char *new_path; /* non-NULL only with directory renames */
2586
2587                 if (pair->status != 'A' && pair->status != 'R') {
2588                         diff_free_filepair(pair);
2589                         continue;
2590                 }
2591                 new_path = check_for_directory_rename(opt, pair->two->path, tree,
2592                                                       dir_renames,
2593                                                       dir_rename_exclusions,
2594                                                       &collisions,
2595                                                       clean_merge);
2596                 if (pair->status != 'R' && !new_path) {
2597                         diff_free_filepair(pair);
2598                         continue;
2599                 }
2600
2601                 re = xmalloc(sizeof(*re));
2602                 re->processed = 0;
2603                 re->add_turned_into_rename = 0;
2604                 re->pair = pair;
2605                 item = string_list_lookup(entries, re->pair->one->path);
2606                 if (!item)
2607                         re->src_entry = insert_stage_data(re->pair->one->path,
2608                                         o_tree, a_tree, b_tree, entries);
2609                 else
2610                         re->src_entry = item->util;
2611
2612                 item = string_list_lookup(entries, re->pair->two->path);
2613                 if (!item)
2614                         re->dst_entry = insert_stage_data(re->pair->two->path,
2615                                         o_tree, a_tree, b_tree, entries);
2616                 else
2617                         re->dst_entry = item->util;
2618                 item = string_list_insert(renames, pair->one->path);
2619                 item->util = re;
2620                 if (new_path)
2621                         apply_directory_rename_modifications(opt, pair, new_path,
2622                                                              re, tree, o_tree,
2623                                                              a_tree, b_tree,
2624                                                              entries);
2625         }
2626
2627         hashmap_iter_init(&collisions, &iter);
2628         while ((e = hashmap_iter_next(&iter))) {
2629                 free(e->target_file);
2630                 string_list_clear(&e->source_files, 0);
2631         }
2632         hashmap_free(&collisions, 1);
2633         return renames;
2634 }
2635
2636 static int process_renames(struct merge_options *opt,
2637                            struct string_list *a_renames,
2638                            struct string_list *b_renames)
2639 {
2640         int clean_merge = 1, i, j;
2641         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2642         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2643         const struct rename *sre;
2644
2645         for (i = 0; i < a_renames->nr; i++) {
2646                 sre = a_renames->items[i].util;
2647                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2648                         = (void *)sre;
2649         }
2650         for (i = 0; i < b_renames->nr; i++) {
2651                 sre = b_renames->items[i].util;
2652                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2653                         = (void *)sre;
2654         }
2655
2656         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2657                 struct string_list *renames1, *renames2Dst;
2658                 struct rename *ren1 = NULL, *ren2 = NULL;
2659                 const char *branch1, *branch2;
2660                 const char *ren1_src, *ren1_dst;
2661                 struct string_list_item *lookup;
2662
2663                 if (i >= a_renames->nr) {
2664                         ren2 = b_renames->items[j++].util;
2665                 } else if (j >= b_renames->nr) {
2666                         ren1 = a_renames->items[i++].util;
2667                 } else {
2668                         int compare = strcmp(a_renames->items[i].string,
2669                                              b_renames->items[j].string);
2670                         if (compare <= 0)
2671                                 ren1 = a_renames->items[i++].util;
2672                         if (compare >= 0)
2673                                 ren2 = b_renames->items[j++].util;
2674                 }
2675
2676                 /* TODO: refactor, so that 1/2 are not needed */
2677                 if (ren1) {
2678                         renames1 = a_renames;
2679                         renames2Dst = &b_by_dst;
2680                         branch1 = opt->branch1;
2681                         branch2 = opt->branch2;
2682                 } else {
2683                         renames1 = b_renames;
2684                         renames2Dst = &a_by_dst;
2685                         branch1 = opt->branch2;
2686                         branch2 = opt->branch1;
2687                         SWAP(ren2, ren1);
2688                 }
2689
2690                 if (ren1->processed)
2691                         continue;
2692                 ren1->processed = 1;
2693                 ren1->dst_entry->processed = 1;
2694                 /* BUG: We should only mark src_entry as processed if we
2695                  * are not dealing with a rename + add-source case.
2696                  */
2697                 ren1->src_entry->processed = 1;
2698
2699                 ren1_src = ren1->pair->one->path;
2700                 ren1_dst = ren1->pair->two->path;
2701
2702                 if (ren2) {
2703                         /* One file renamed on both sides */
2704                         const char *ren2_src = ren2->pair->one->path;
2705                         const char *ren2_dst = ren2->pair->two->path;
2706                         enum rename_type rename_type;
2707                         if (strcmp(ren1_src, ren2_src) != 0)
2708                                 BUG("ren1_src != ren2_src");
2709                         ren2->dst_entry->processed = 1;
2710                         ren2->processed = 1;
2711                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2712                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2713                                 clean_merge = 0;
2714                         } else {
2715                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2716                                 /* BUG: We should only remove ren1_src in
2717                                  * the base stage (think of rename +
2718                                  * add-source cases).
2719                                  */
2720                                 remove_file(opt, 1, ren1_src, 1);
2721                                 update_entry(ren1->dst_entry,
2722                                              ren1->pair->one,
2723                                              ren1->pair->two,
2724                                              ren2->pair->two);
2725                         }
2726                         setup_rename_conflict_info(rename_type,
2727                                                    opt,
2728                                                    ren1,
2729                                                    ren2,
2730                                                    branch1,
2731                                                    branch2,
2732                                                    NULL,
2733                                                    NULL);
2734                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2735                         /* Two different files renamed to the same thing */
2736                         char *ren2_dst;
2737                         ren2 = lookup->util;
2738                         ren2_dst = ren2->pair->two->path;
2739                         if (strcmp(ren1_dst, ren2_dst) != 0)
2740                                 BUG("ren1_dst != ren2_dst");
2741
2742                         clean_merge = 0;
2743                         ren2->processed = 1;
2744                         /*
2745                          * BUG: We should only mark src_entry as processed
2746                          * if we are not dealing with a rename + add-source
2747                          * case.
2748                          */
2749                         ren2->src_entry->processed = 1;
2750
2751                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2752                                                    opt,
2753                                                    ren1,
2754                                                    ren2,
2755                                                    branch1,
2756                                                    branch2,
2757                                                    ren1->src_entry,
2758                                                    ren2->src_entry);
2759
2760                 } else {
2761                         /* Renamed in 1, maybe changed in 2 */
2762                         /* we only use sha1 and mode of these */
2763                         struct diff_filespec src_other, dst_other;
2764                         int try_merge;
2765
2766                         /*
2767                          * unpack_trees loads entries from common-commit
2768                          * into stage 1, from head-commit into stage 2, and
2769                          * from merge-commit into stage 3.  We keep track
2770                          * of which side corresponds to the rename.
2771                          */
2772                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2773                         int other_stage =   a_renames == renames1 ? 3 : 2;
2774
2775                         /* BUG: We should only remove ren1_src in the base
2776                          * stage and in other_stage (think of rename +
2777                          * add-source case).
2778                          */
2779                         remove_file(opt, 1, ren1_src,
2780                                     renamed_stage == 2 || !was_tracked(opt, ren1_src));
2781
2782                         oidcpy(&src_other.oid,
2783                                &ren1->src_entry->stages[other_stage].oid);
2784                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2785                         oidcpy(&dst_other.oid,
2786                                &ren1->dst_entry->stages[other_stage].oid);
2787                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2788                         try_merge = 0;
2789
2790                         if (oid_eq(&src_other.oid, &null_oid) &&
2791                             ren1->add_turned_into_rename) {
2792                                 setup_rename_conflict_info(RENAME_VIA_DIR,
2793                                                            opt,
2794                                                            ren1,
2795                                                            NULL,
2796                                                            branch1,
2797                                                            branch2,
2798                                                            NULL,
2799                                                            NULL);
2800                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2801                                 setup_rename_conflict_info(RENAME_DELETE,
2802                                                            opt,
2803                                                            ren1,
2804                                                            NULL,
2805                                                            branch1,
2806                                                            branch2,
2807                                                            NULL,
2808                                                            NULL);
2809                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2810                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2811                                 /*
2812                                  * Added file on the other side identical to
2813                                  * the file being renamed: clean merge.
2814                                  * Also, there is no need to overwrite the
2815                                  * file already in the working copy, so call
2816                                  * update_file_flags() instead of
2817                                  * update_file().
2818                                  */
2819                                 if (update_file_flags(opt,
2820                                                       &ren1->pair->two->oid,
2821                                                       ren1->pair->two->mode,
2822                                                       ren1_dst,
2823                                                       1, /* update_cache */
2824                                                       0  /* update_wd    */))
2825                                         clean_merge = -1;
2826                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2827                                 /*
2828                                  * Probably not a clean merge, but it's
2829                                  * premature to set clean_merge to 0 here,
2830                                  * because if the rename merges cleanly and
2831                                  * the merge exactly matches the newly added
2832                                  * file, then the merge will be clean.
2833                                  */
2834                                 setup_rename_conflict_info(RENAME_ADD,
2835                                                            opt,
2836                                                            ren1,
2837                                                            NULL,
2838                                                            branch1,
2839                                                            branch2,
2840                                                            ren1->src_entry,
2841                                                            NULL);
2842                         } else
2843                                 try_merge = 1;
2844
2845                         if (clean_merge < 0)
2846                                 goto cleanup_and_return;
2847                         if (try_merge) {
2848                                 struct diff_filespec *o, *a, *b;
2849                                 src_other.path = (char *)ren1_src;
2850
2851                                 o = ren1->pair->one;
2852                                 if (a_renames == renames1) {
2853                                         a = ren1->pair->two;
2854                                         b = &src_other;
2855                                 } else {
2856                                         b = ren1->pair->two;
2857                                         a = &src_other;
2858                                 }
2859                                 update_entry(ren1->dst_entry, o, a, b);
2860                                 setup_rename_conflict_info(RENAME_NORMAL,
2861                                                            opt,
2862                                                            ren1,
2863                                                            NULL,
2864                                                            branch1,
2865                                                            NULL,
2866                                                            NULL,
2867                                                            NULL);
2868                         }
2869                 }
2870         }
2871 cleanup_and_return:
2872         string_list_clear(&a_by_dst, 0);
2873         string_list_clear(&b_by_dst, 0);
2874
2875         return clean_merge;
2876 }
2877
2878 struct rename_info {
2879         struct string_list *head_renames;
2880         struct string_list *merge_renames;
2881 };
2882
2883 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2884                                    struct hashmap *dir_renames)
2885 {
2886         struct hashmap_iter iter;
2887         struct dir_rename_entry *e;
2888
2889         hashmap_iter_init(dir_renames, &iter);
2890         while ((e = hashmap_iter_next(&iter))) {
2891                 free(e->dir);
2892                 strbuf_release(&e->new_dir);
2893                 /* possible_new_dirs already cleared in get_directory_renames */
2894         }
2895         hashmap_free(dir_renames, 1);
2896         free(dir_renames);
2897
2898         free(pairs->queue);
2899         free(pairs);
2900 }
2901
2902 static int detect_and_process_renames(struct merge_options *opt,
2903                                       struct tree *common,
2904                                       struct tree *head,
2905                                       struct tree *merge,
2906                                       struct string_list *entries,
2907                                       struct rename_info *ri)
2908 {
2909         struct diff_queue_struct *head_pairs, *merge_pairs;
2910         struct hashmap *dir_re_head, *dir_re_merge;
2911         int clean = 1;
2912
2913         ri->head_renames = NULL;
2914         ri->merge_renames = NULL;
2915
2916         if (!merge_detect_rename(opt))
2917                 return 1;
2918
2919         head_pairs = get_diffpairs(opt, common, head);
2920         merge_pairs = get_diffpairs(opt, common, merge);
2921
2922         if (opt->detect_directory_renames) {
2923                 dir_re_head = get_directory_renames(head_pairs);
2924                 dir_re_merge = get_directory_renames(merge_pairs);
2925
2926                 handle_directory_level_conflicts(opt,
2927                                                  dir_re_head, head,
2928                                                  dir_re_merge, merge);
2929         } else {
2930                 dir_re_head  = xmalloc(sizeof(*dir_re_head));
2931                 dir_re_merge = xmalloc(sizeof(*dir_re_merge));
2932                 dir_rename_init(dir_re_head);
2933                 dir_rename_init(dir_re_merge);
2934         }
2935
2936         ri->head_renames  = get_renames(opt, head_pairs,
2937                                         dir_re_merge, dir_re_head, head,
2938                                         common, head, merge, entries,
2939                                         &clean);
2940         if (clean < 0)
2941                 goto cleanup;
2942         ri->merge_renames = get_renames(opt, merge_pairs,
2943                                         dir_re_head, dir_re_merge, merge,
2944                                         common, head, merge, entries,
2945                                         &clean);
2946         if (clean < 0)
2947                 goto cleanup;
2948         clean &= process_renames(opt, ri->head_renames, ri->merge_renames);
2949
2950 cleanup:
2951         /*
2952          * Some cleanup is deferred until cleanup_renames() because the
2953          * data structures are still needed and referenced in
2954          * process_entry().  But there are a few things we can free now.
2955          */
2956         initial_cleanup_rename(head_pairs, dir_re_head);
2957         initial_cleanup_rename(merge_pairs, dir_re_merge);
2958
2959         return clean;
2960 }
2961
2962 static void final_cleanup_rename(struct string_list *rename)
2963 {
2964         const struct rename *re;
2965         int i;
2966
2967         if (rename == NULL)
2968                 return;
2969
2970         for (i = 0; i < rename->nr; i++) {
2971                 re = rename->items[i].util;
2972                 diff_free_filepair(re->pair);
2973         }
2974         string_list_clear(rename, 1);
2975         free(rename);
2976 }
2977
2978 static void final_cleanup_renames(struct rename_info *re_info)
2979 {
2980         final_cleanup_rename(re_info->head_renames);
2981         final_cleanup_rename(re_info->merge_renames);
2982 }
2983
2984 static struct object_id *stage_oid(const struct object_id *oid, unsigned mode)
2985 {
2986         return (is_null_oid(oid) || mode == 0) ? NULL: (struct object_id *)oid;
2987 }
2988
2989 static int read_oid_strbuf(struct merge_options *opt,
2990                            const struct object_id *oid,
2991                            struct strbuf *dst)
2992 {
2993         void *buf;
2994         enum object_type type;
2995         unsigned long size;
2996         buf = read_object_file(oid, &type, &size);
2997         if (!buf)
2998                 return err(opt, _("cannot read object %s"), oid_to_hex(oid));
2999         if (type != OBJ_BLOB) {
3000                 free(buf);
3001                 return err(opt, _("object %s is not a blob"), oid_to_hex(oid));
3002         }
3003         strbuf_attach(dst, buf, size, size + 1);
3004         return 0;
3005 }
3006
3007 static int blob_unchanged(struct merge_options *opt,
3008                           const struct object_id *o_oid,
3009                           unsigned o_mode,
3010                           const struct object_id *a_oid,
3011                           unsigned a_mode,
3012                           int renormalize, const char *path)
3013 {
3014         struct strbuf obuf = STRBUF_INIT;
3015         struct strbuf abuf = STRBUF_INIT;
3016         int ret = 0; /* assume changed for safety */
3017         const struct index_state *idx = opt->repo->index;
3018
3019         if (a_mode != o_mode)
3020                 return 0;
3021         if (oid_eq(o_oid, a_oid))
3022                 return 1;
3023         if (!renormalize)
3024                 return 0;
3025
3026         assert(o_oid && a_oid);
3027         if (read_oid_strbuf(opt, o_oid, &obuf) ||
3028             read_oid_strbuf(opt, a_oid, &abuf))
3029                 goto error_return;
3030         /*
3031          * Note: binary | is used so that both renormalizations are
3032          * performed.  Comparison can be skipped if both files are
3033          * unchanged since their sha1s have already been compared.
3034          */
3035         if (renormalize_buffer(idx, path, obuf.buf, obuf.len, &obuf) |
3036             renormalize_buffer(idx, path, abuf.buf, abuf.len, &abuf))
3037                 ret = (obuf.len == abuf.len && !memcmp(obuf.buf, abuf.buf, obuf.len));
3038
3039 error_return:
3040         strbuf_release(&obuf);
3041         strbuf_release(&abuf);
3042         return ret;
3043 }
3044
3045 static int handle_modify_delete(struct merge_options *opt,
3046                                 const char *path,
3047                                 struct object_id *o_oid, int o_mode,
3048                                 struct object_id *a_oid, int a_mode,
3049                                 struct object_id *b_oid, int b_mode)
3050 {
3051         const char *modify_branch, *delete_branch;
3052         struct object_id *changed_oid;
3053         int changed_mode;
3054
3055         if (a_oid) {
3056                 modify_branch = opt->branch1;
3057                 delete_branch = opt->branch2;
3058                 changed_oid = a_oid;
3059                 changed_mode = a_mode;
3060         } else {
3061                 modify_branch = opt->branch2;
3062                 delete_branch = opt->branch1;
3063                 changed_oid = b_oid;
3064                 changed_mode = b_mode;
3065         }
3066
3067         return handle_change_delete(opt,
3068                                     path, NULL,
3069                                     o_oid, o_mode,
3070                                     changed_oid, changed_mode,
3071                                     modify_branch, delete_branch,
3072                                     _("modify"), _("modified"));
3073 }
3074
3075 static int handle_content_merge(struct merge_options *opt,
3076                                 const char *path,
3077                                 int is_dirty,
3078                                 struct object_id *o_oid, int o_mode,
3079                                 struct object_id *a_oid, int a_mode,
3080                                 struct object_id *b_oid, int b_mode,
3081                                 struct rename_conflict_info *ci)
3082 {
3083         const char *reason = _("content");
3084         const char *path1 = NULL, *path2 = NULL;
3085         struct merge_file_info mfi;
3086         struct diff_filespec one, a, b;
3087         unsigned df_conflict_remains = 0;
3088
3089         if (!o_oid) {
3090                 reason = _("add/add");
3091                 o_oid = (struct object_id *)&null_oid;
3092         }
3093         one.path = a.path = b.path = (char *)path;
3094         oidcpy(&one.oid, o_oid);
3095         one.mode = o_mode;
3096         oidcpy(&a.oid, a_oid);
3097         a.mode = a_mode;
3098         oidcpy(&b.oid, b_oid);
3099         b.mode = b_mode;
3100
3101         if (ci) {
3102                 struct diff_filepair *pair1 = ci->ren1->pair;
3103
3104                 path1 = (opt->branch1 == ci->branch1) ?
3105                         pair1->two->path : pair1->one->path;
3106                 /* If ci->ren2->pair != NULL, we are in
3107                  * RENAME_ONE_FILE_TO_ONE case.  Otherwise, we have a
3108                  * normal rename.
3109                  */
3110                 path2 = ((ci->ren2 && ci->ren2->pair) ||
3111                          opt->branch2 == ci->branch1) ?
3112                         pair1->two->path : pair1->one->path;
3113                 one.path = pair1->one->path;
3114                 a.path = (char *)path1;
3115                 b.path = (char *)path2;
3116
3117                 if (dir_in_way(opt->repo->index, path, !opt->call_depth,
3118                                S_ISGITLINK(pair1->two->mode)))
3119                         df_conflict_remains = 1;
3120         }
3121         if (merge_mode_and_contents(opt, &one, &a, &b, path,
3122                                     opt->branch1, opt->branch2,
3123                                     opt->call_depth * 2, &mfi))
3124                 return -1;
3125
3126         /*
3127          * We can skip updating the working tree file iff:
3128          *   a) The merge is clean
3129          *   b) The merge matches what was in HEAD (content, mode, pathname)
3130          *   c) The target path is usable (i.e. not involved in D/F conflict)
3131          */
3132         if (mfi.clean &&
3133             was_tracked_and_matches(opt, path, &mfi.oid, mfi.mode) &&
3134             !df_conflict_remains) {
3135                 int pos;
3136                 struct cache_entry *ce;
3137
3138                 output(opt, 3, _("Skipped %s (merged same as existing)"), path);
3139                 if (add_cacheinfo(opt, mfi.mode, &mfi.oid, path,
3140                                   0, (!opt->call_depth && !is_dirty), 0))
3141                         return -1;
3142                 /*
3143                  * However, add_cacheinfo() will delete the old cache entry
3144                  * and add a new one.  We need to copy over any skip_worktree
3145                  * flag to avoid making the file appear as if it were
3146                  * deleted by the user.
3147                  */
3148                 pos = index_name_pos(&opt->orig_index, path, strlen(path));
3149                 ce = opt->orig_index.cache[pos];
3150                 if (ce_skip_worktree(ce)) {
3151                         pos = index_name_pos(opt->repo->index, path, strlen(path));
3152                         ce = opt->repo->index->cache[pos];
3153                         ce->ce_flags |= CE_SKIP_WORKTREE;
3154                 }
3155                 return mfi.clean;
3156         }
3157
3158         if (!mfi.clean) {
3159                 if (S_ISGITLINK(mfi.mode))
3160                         reason = _("submodule");
3161                 output(opt, 1, _("CONFLICT (%s): Merge conflict in %s"),
3162                                 reason, path);
3163                 if (ci && !df_conflict_remains)
3164                         if (update_stages(opt, path, &one, &a, &b))
3165                                 return -1;
3166         }
3167
3168         if (df_conflict_remains || is_dirty) {
3169                 char *new_path;
3170                 if (opt->call_depth) {
3171                         remove_file_from_index(opt->repo->index, path);
3172                 } else {
3173                         if (!mfi.clean) {
3174                                 if (update_stages(opt, path, &one, &a, &b))
3175                                         return -1;
3176                         } else {
3177                                 int file_from_stage2 = was_tracked(opt, path);
3178                                 struct diff_filespec merged;
3179                                 oidcpy(&merged.oid, &mfi.oid);
3180                                 merged.mode = mfi.mode;
3181
3182                                 if (update_stages(opt, path, NULL,
3183                                                   file_from_stage2 ? &merged : NULL,
3184                                                   file_from_stage2 ? NULL : &merged))
3185                                         return -1;
3186                         }
3187
3188                 }
3189                 new_path = unique_path(opt, path, ci->branch1);
3190                 if (is_dirty) {
3191                         output(opt, 1, _("Refusing to lose dirty file at %s"),
3192                                path);
3193                 }
3194                 output(opt, 1, _("Adding as %s instead"), new_path);
3195                 if (update_file(opt, 0, &mfi.oid, mfi.mode, new_path)) {
3196                         free(new_path);
3197                         return -1;
3198                 }
3199                 free(new_path);
3200                 mfi.clean = 0;
3201         } else if (update_file(opt, mfi.clean, &mfi.oid, mfi.mode, path))
3202                 return -1;
3203         return !is_dirty && mfi.clean;
3204 }
3205
3206 static int handle_rename_normal(struct merge_options *opt,
3207                                 const char *path,
3208                                 struct object_id *o_oid, unsigned int o_mode,
3209                                 struct object_id *a_oid, unsigned int a_mode,
3210                                 struct object_id *b_oid, unsigned int b_mode,
3211                                 struct rename_conflict_info *ci)
3212 {
3213         /* Merge the content and write it out */
3214         return handle_content_merge(opt, path, was_dirty(opt, path),
3215                                     o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3216                                     ci);
3217 }
3218
3219 /* Per entry merge function */
3220 static int process_entry(struct merge_options *opt,
3221                          const char *path, struct stage_data *entry)
3222 {
3223         int clean_merge = 1;
3224         int normalize = opt->renormalize;
3225         unsigned o_mode = entry->stages[1].mode;
3226         unsigned a_mode = entry->stages[2].mode;
3227         unsigned b_mode = entry->stages[3].mode;
3228         struct object_id *o_oid = stage_oid(&entry->stages[1].oid, o_mode);
3229         struct object_id *a_oid = stage_oid(&entry->stages[2].oid, a_mode);
3230         struct object_id *b_oid = stage_oid(&entry->stages[3].oid, b_mode);
3231
3232         entry->processed = 1;
3233         if (entry->rename_conflict_info) {
3234                 struct rename_conflict_info *ci = entry->rename_conflict_info;
3235                 switch (ci->rename_type) {
3236                 case RENAME_NORMAL:
3237                 case RENAME_ONE_FILE_TO_ONE:
3238                         clean_merge = handle_rename_normal(opt,
3239                                                            path,
3240                                                            o_oid, o_mode,
3241                                                            a_oid, a_mode,
3242                                                            b_oid, b_mode,
3243                                                            ci);
3244                         break;
3245                 case RENAME_VIA_DIR:
3246                         clean_merge = 1;
3247                         if (handle_rename_via_dir(opt, ci->ren1->pair, ci->branch1))
3248                                 clean_merge = -1;
3249                         break;
3250                 case RENAME_ADD:
3251                         /*
3252                          * Probably unclean merge, but if the renamed file
3253                          * merges cleanly and the result can then be
3254                          * two-way merged cleanly with the added file, I
3255                          * guess it's a clean merge?
3256                          */
3257                         clean_merge = handle_rename_add(opt, ci);
3258                         break;
3259                 case RENAME_DELETE:
3260                         clean_merge = 0;
3261                         if (handle_rename_delete(opt, ci->ren1->pair,
3262                                                  ci->branch1, ci->branch2))
3263                                 clean_merge = -1;
3264                         break;
3265                 case RENAME_ONE_FILE_TO_TWO:
3266                         clean_merge = 0;
3267                         if (handle_rename_rename_1to2(opt, ci))
3268                                 clean_merge = -1;
3269                         break;
3270                 case RENAME_TWO_FILES_TO_ONE:
3271                         /*
3272                          * Probably unclean merge, but if the two renamed
3273                          * files merge cleanly and the two resulting files
3274                          * can then be two-way merged cleanly, I guess it's
3275                          * a clean merge?
3276                          */
3277                         clean_merge = handle_rename_rename_2to1(opt, ci);
3278                         break;
3279                 default:
3280                         entry->processed = 0;
3281                         break;
3282                 }
3283         } else if (o_oid && (!a_oid || !b_oid)) {
3284                 /* Case A: Deleted in one */
3285                 if ((!a_oid && !b_oid) ||
3286                     (!b_oid && blob_unchanged(opt, o_oid, o_mode, a_oid, a_mode, normalize, path)) ||
3287                     (!a_oid && blob_unchanged(opt, o_oid, o_mode, b_oid, b_mode, normalize, path))) {
3288                         /* Deleted in both or deleted in one and
3289                          * unchanged in the other */
3290                         if (a_oid)
3291                                 output(opt, 2, _("Removing %s"), path);
3292                         /* do not touch working file if it did not exist */
3293                         remove_file(opt, 1, path, !a_oid);
3294                 } else {
3295                         /* Modify/delete; deleted side may have put a directory in the way */
3296                         clean_merge = 0;
3297                         if (handle_modify_delete(opt, path, o_oid, o_mode,
3298                                                  a_oid, a_mode, b_oid, b_mode))
3299                                 clean_merge = -1;
3300                 }
3301         } else if ((!o_oid && a_oid && !b_oid) ||
3302                    (!o_oid && !a_oid && b_oid)) {
3303                 /* Case B: Added in one. */
3304                 /* [nothing|directory] -> ([nothing|directory], file) */
3305
3306                 const char *add_branch;
3307                 const char *other_branch;
3308                 unsigned mode;
3309                 const struct object_id *oid;
3310                 const char *conf;
3311
3312                 if (a_oid) {
3313                         add_branch = opt->branch1;
3314                         other_branch = opt->branch2;
3315                         mode = a_mode;
3316                         oid = a_oid;
3317                         conf = _("file/directory");
3318                 } else {
3319                         add_branch = opt->branch2;
3320                         other_branch = opt->branch1;
3321                         mode = b_mode;
3322                         oid = b_oid;
3323                         conf = _("directory/file");
3324                 }
3325                 if (dir_in_way(opt->repo->index, path,
3326                                !opt->call_depth && !S_ISGITLINK(a_mode),
3327                                0)) {
3328                         char *new_path = unique_path(opt, path, add_branch);
3329                         clean_merge = 0;
3330                         output(opt, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3331                                "Adding %s as %s"),
3332                                conf, path, other_branch, path, new_path);
3333                         if (update_file(opt, 0, oid, mode, new_path))
3334                                 clean_merge = -1;
3335                         else if (opt->call_depth)
3336                                 remove_file_from_index(opt->repo->index, path);
3337                         free(new_path);
3338                 } else {
3339                         output(opt, 2, _("Adding %s"), path);
3340                         /* do not overwrite file if already present */
3341                         if (update_file_flags(opt, oid, mode, path, 1, !a_oid))
3342                                 clean_merge = -1;
3343                 }
3344         } else if (a_oid && b_oid) {
3345                 if (!o_oid) {
3346                         /* Case C: Added in both (check for same permissions) */
3347                         output(opt, 1,
3348                                _("CONFLICT (add/add): Merge conflict in %s"),
3349                                path);
3350                         clean_merge = handle_file_collision(opt,
3351                                                             path, NULL, NULL,
3352                                                             opt->branch1,
3353                                                             opt->branch2,
3354                                                             a_oid, a_mode,
3355                                                             b_oid, b_mode);
3356                 } else {
3357                         /* case D: Modified in both, but differently. */
3358                         int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3359                         clean_merge = handle_content_merge(opt, path,
3360                                                            is_dirty,
3361                                                            o_oid, o_mode,
3362                                                            a_oid, a_mode,
3363                                                            b_oid, b_mode,
3364                                                            NULL);
3365                 }
3366         } else if (!o_oid && !a_oid && !b_oid) {
3367                 /*
3368                  * this entry was deleted altogether. a_mode == 0 means
3369                  * we had that path and want to actively remove it.
3370                  */
3371                 remove_file(opt, 1, path, !a_mode);
3372         } else
3373                 BUG("fatal merge failure, shouldn't happen.");
3374
3375         return clean_merge;
3376 }
3377
3378 int merge_trees(struct merge_options *opt,
3379                 struct tree *head,
3380                 struct tree *merge,
3381                 struct tree *common,
3382                 struct tree **result)
3383 {
3384         struct index_state *istate = opt->repo->index;
3385         int code, clean;
3386         struct strbuf sb = STRBUF_INIT;
3387
3388         if (!opt->call_depth && repo_index_has_changes(opt->repo, head, &sb)) {
3389                 err(opt, _("Your local changes to the following files would be overwritten by merge:\n  %s"),
3390                     sb.buf);
3391                 return -1;
3392         }
3393
3394         if (opt->subtree_shift) {
3395                 merge = shift_tree_object(opt->repo, head, merge, opt->subtree_shift);
3396                 common = shift_tree_object(opt->repo, head, common, opt->subtree_shift);
3397         }
3398
3399         if (oid_eq(&common->object.oid, &merge->object.oid)) {
3400                 output(opt, 0, _("Already up to date!"));
3401                 *result = head;
3402                 return 1;
3403         }
3404
3405         code = unpack_trees_start(opt, common, head, merge);
3406
3407         if (code != 0) {
3408                 if (show(opt, 4) || opt->call_depth)
3409                         err(opt, _("merging of trees %s and %s failed"),
3410                             oid_to_hex(&head->object.oid),
3411                             oid_to_hex(&merge->object.oid));
3412                 unpack_trees_finish(opt);
3413                 return -1;
3414         }
3415
3416         if (unmerged_index(istate)) {
3417                 struct string_list *entries;
3418                 struct rename_info re_info;
3419                 int i;
3420                 /*
3421                  * Only need the hashmap while processing entries, so
3422                  * initialize it here and free it when we are done running
3423                  * through the entries. Keeping it in the merge_options as
3424                  * opposed to decaring a local hashmap is for convenience
3425                  * so that we don't have to pass it to around.
3426                  */
3427                 hashmap_init(&opt->current_file_dir_set, path_hashmap_cmp, NULL, 512);
3428                 get_files_dirs(opt, head);
3429                 get_files_dirs(opt, merge);
3430
3431                 entries = get_unmerged(opt->repo->index);
3432                 clean = detect_and_process_renames(opt, common, head, merge,
3433                                                    entries, &re_info);
3434                 record_df_conflict_files(opt, entries);
3435                 if (clean < 0)
3436                         goto cleanup;
3437                 for (i = entries->nr-1; 0 <= i; i--) {
3438                         const char *path = entries->items[i].string;
3439                         struct stage_data *e = entries->items[i].util;
3440                         if (!e->processed) {
3441                                 int ret = process_entry(opt, path, e);
3442                                 if (!ret)
3443                                         clean = 0;
3444                                 else if (ret < 0) {
3445                                         clean = ret;
3446                                         goto cleanup;
3447                                 }
3448                         }
3449                 }
3450                 for (i = 0; i < entries->nr; i++) {
3451                         struct stage_data *e = entries->items[i].util;
3452                         if (!e->processed)
3453                                 BUG("unprocessed path??? %s",
3454                                     entries->items[i].string);
3455                 }
3456
3457         cleanup:
3458                 final_cleanup_renames(&re_info);
3459
3460                 string_list_clear(entries, 1);
3461                 free(entries);
3462
3463                 hashmap_free(&opt->current_file_dir_set, 1);
3464
3465                 if (clean < 0) {
3466                         unpack_trees_finish(opt);
3467                         return clean;
3468                 }
3469         }
3470         else
3471                 clean = 1;
3472
3473         unpack_trees_finish(opt);
3474
3475         if (opt->call_depth && !(*result = write_tree_from_memory(opt)))
3476                 return -1;
3477
3478         return clean;
3479 }
3480
3481 static struct commit_list *reverse_commit_list(struct commit_list *list)
3482 {
3483         struct commit_list *next = NULL, *current, *backup;
3484         for (current = list; current; current = backup) {
3485                 backup = current->next;
3486                 current->next = next;
3487                 next = current;
3488         }
3489         return next;
3490 }
3491
3492 /*
3493  * Merge the commits h1 and h2, return the resulting virtual
3494  * commit object and a flag indicating the cleanness of the merge.
3495  */
3496 int merge_recursive(struct merge_options *opt,
3497                     struct commit *h1,
3498                     struct commit *h2,
3499                     struct commit_list *ca,
3500                     struct commit **result)
3501 {
3502         struct commit_list *iter;
3503         struct commit *merged_common_ancestors;
3504         struct tree *mrtree;
3505         int clean;
3506
3507         if (show(opt, 4)) {
3508                 output(opt, 4, _("Merging:"));
3509                 output_commit_title(opt, h1);
3510                 output_commit_title(opt, h2);
3511         }
3512
3513         if (!ca) {
3514                 ca = get_merge_bases(h1, h2);
3515                 ca = reverse_commit_list(ca);
3516         }
3517
3518         if (show(opt, 5)) {
3519                 unsigned cnt = commit_list_count(ca);
3520
3521                 output(opt, 5, Q_("found %u common ancestor:",
3522                                 "found %u common ancestors:", cnt), cnt);
3523                 for (iter = ca; iter; iter = iter->next)
3524                         output_commit_title(opt, iter->item);
3525         }
3526
3527         merged_common_ancestors = pop_commit(&ca);
3528         if (merged_common_ancestors == NULL) {
3529                 /* if there is no common ancestor, use an empty tree */
3530                 struct tree *tree;
3531
3532                 tree = lookup_tree(opt->repo, opt->repo->hash_algo->empty_tree);
3533                 merged_common_ancestors = make_virtual_commit(opt->repo, tree, "ancestor");
3534         }
3535
3536         for (iter = ca; iter; iter = iter->next) {
3537                 const char *saved_b1, *saved_b2;
3538                 opt->call_depth++;
3539                 /*
3540                  * When the merge fails, the result contains files
3541                  * with conflict markers. The cleanness flag is
3542                  * ignored (unless indicating an error), it was never
3543                  * actually used, as result of merge_trees has always
3544                  * overwritten it: the committed "conflicts" were
3545                  * already resolved.
3546                  */
3547                 discard_index(opt->repo->index);
3548                 saved_b1 = opt->branch1;
3549                 saved_b2 = opt->branch2;
3550                 opt->branch1 = "Temporary merge branch 1";
3551                 opt->branch2 = "Temporary merge branch 2";
3552                 if (merge_recursive(opt, merged_common_ancestors, iter->item,
3553                                     NULL, &merged_common_ancestors) < 0)
3554                         return -1;
3555                 opt->branch1 = saved_b1;
3556                 opt->branch2 = saved_b2;
3557                 opt->call_depth--;
3558
3559                 if (!merged_common_ancestors)
3560                         return err(opt, _("merge returned no commit"));
3561         }
3562
3563         discard_index(opt->repo->index);
3564         if (!opt->call_depth)
3565                 repo_read_index(opt->repo);
3566
3567         opt->ancestor = "merged common ancestors";
3568         clean = merge_trees(opt, get_commit_tree(h1), get_commit_tree(h2),
3569                             get_commit_tree(merged_common_ancestors),
3570                             &mrtree);
3571         if (clean < 0) {
3572                 flush_output(opt);
3573                 return clean;
3574         }
3575
3576         if (opt->call_depth) {
3577                 *result = make_virtual_commit(opt->repo, mrtree, "merged tree");
3578                 commit_list_insert(h1, &(*result)->parents);
3579                 commit_list_insert(h2, &(*result)->parents->next);
3580         }
3581         flush_output(opt);
3582         if (!opt->call_depth && opt->buffer_output < 2)
3583                 strbuf_release(&opt->obuf);
3584         if (show(opt, 2))
3585                 diff_warn_rename_limit("merge.renamelimit",
3586                                        opt->needed_rename_limit, 0);
3587         return clean;
3588 }
3589
3590 static struct commit *get_ref(struct repository *repo, const struct object_id *oid,
3591                               const char *name)
3592 {
3593         struct object *object;
3594
3595         object = deref_tag(repo, parse_object(repo, oid),
3596                            name, strlen(name));
3597         if (!object)
3598                 return NULL;
3599         if (object->type == OBJ_TREE)
3600                 return make_virtual_commit(repo, (struct tree*)object, name);
3601         if (object->type != OBJ_COMMIT)
3602                 return NULL;
3603         if (parse_commit((struct commit *)object))
3604                 return NULL;
3605         return (struct commit *)object;
3606 }
3607
3608 int merge_recursive_generic(struct merge_options *opt,
3609                             const struct object_id *head,
3610                             const struct object_id *merge,
3611                             int num_base_list,
3612                             const struct object_id **base_list,
3613                             struct commit **result)
3614 {
3615         int clean;
3616         struct lock_file lock = LOCK_INIT;
3617         struct commit *head_commit = get_ref(opt->repo, head, opt->branch1);
3618         struct commit *next_commit = get_ref(opt->repo, merge, opt->branch2);
3619         struct commit_list *ca = NULL;
3620
3621         if (base_list) {
3622                 int i;
3623                 for (i = 0; i < num_base_list; ++i) {
3624                         struct commit *base;
3625                         if (!(base = get_ref(opt->repo, base_list[i], oid_to_hex(base_list[i]))))
3626                                 return err(opt, _("Could not parse object '%s'"),
3627                                            oid_to_hex(base_list[i]));
3628                         commit_list_insert(base, &ca);
3629                 }
3630         }
3631
3632         repo_hold_locked_index(opt->repo, &lock, LOCK_DIE_ON_ERROR);
3633         clean = merge_recursive(opt, head_commit, next_commit, ca,
3634                                 result);
3635         if (clean < 0) {
3636                 rollback_lock_file(&lock);
3637                 return clean;
3638         }
3639
3640         if (write_locked_index(opt->repo->index, &lock,
3641                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3642                 return err(opt, _("Unable to write index."));
3643
3644         return clean ? 0 : 1;
3645 }
3646
3647 static void merge_recursive_config(struct merge_options *opt)
3648 {
3649         char *value = NULL;
3650         git_config_get_int("merge.verbosity", &opt->verbosity);
3651         git_config_get_int("diff.renamelimit", &opt->diff_rename_limit);
3652         git_config_get_int("merge.renamelimit", &opt->merge_rename_limit);
3653         if (!git_config_get_string("diff.renames", &value)) {
3654                 opt->diff_detect_rename = git_config_rename("diff.renames", value);
3655                 free(value);
3656         }
3657         if (!git_config_get_string("merge.renames", &value)) {
3658                 opt->merge_detect_rename = git_config_rename("merge.renames", value);
3659                 free(value);
3660         }
3661         git_config(git_xmerge_config, NULL);
3662 }
3663
3664 void init_merge_options(struct merge_options *opt,
3665                         struct repository *repo)
3666 {
3667         const char *merge_verbosity;
3668         memset(opt, 0, sizeof(struct merge_options));
3669         opt->repo = repo;
3670         opt->verbosity = 2;
3671         opt->buffer_output = 1;
3672         opt->diff_rename_limit = -1;
3673         opt->merge_rename_limit = -1;
3674         opt->renormalize = 0;
3675         opt->diff_detect_rename = -1;
3676         opt->merge_detect_rename = -1;
3677         opt->detect_directory_renames = 1;
3678         merge_recursive_config(opt);
3679         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3680         if (merge_verbosity)
3681                 opt->verbosity = strtol(merge_verbosity, NULL, 10);
3682         if (opt->verbosity >= 5)
3683                 opt->buffer_output = 0;
3684         strbuf_init(&opt->obuf, 0);
3685         string_list_init(&opt->df_conflict_file_set, 1);
3686 }
3687
3688 int parse_merge_opt(struct merge_options *opt, const char *s)
3689 {
3690         const char *arg;
3691
3692         if (!s || !*s)
3693                 return -1;
3694         if (!strcmp(s, "ours"))
3695                 opt->recursive_variant = MERGE_RECURSIVE_OURS;
3696         else if (!strcmp(s, "theirs"))
3697                 opt->recursive_variant = MERGE_RECURSIVE_THEIRS;
3698         else if (!strcmp(s, "subtree"))
3699                 opt->subtree_shift = "";
3700         else if (skip_prefix(s, "subtree=", &arg))
3701                 opt->subtree_shift = arg;
3702         else if (!strcmp(s, "patience"))
3703                 opt->xdl_opts = DIFF_WITH_ALG(opt, PATIENCE_DIFF);
3704         else if (!strcmp(s, "histogram"))
3705                 opt->xdl_opts = DIFF_WITH_ALG(opt, HISTOGRAM_DIFF);
3706         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3707                 long value = parse_algorithm_value(arg);
3708                 if (value < 0)
3709                         return -1;
3710                 /* clear out previous settings */
3711                 DIFF_XDL_CLR(opt, NEED_MINIMAL);
3712                 opt->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3713                 opt->xdl_opts |= value;
3714         }
3715         else if (!strcmp(s, "ignore-space-change"))
3716                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_CHANGE);
3717         else if (!strcmp(s, "ignore-all-space"))
3718                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE);
3719         else if (!strcmp(s, "ignore-space-at-eol"))
3720                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_AT_EOL);
3721         else if (!strcmp(s, "ignore-cr-at-eol"))
3722                 DIFF_XDL_SET(opt, IGNORE_CR_AT_EOL);
3723         else if (!strcmp(s, "renormalize"))
3724                 opt->renormalize = 1;
3725         else if (!strcmp(s, "no-renormalize"))
3726                 opt->renormalize = 0;
3727         else if (!strcmp(s, "no-renames"))
3728                 opt->merge_detect_rename = 0;
3729         else if (!strcmp(s, "find-renames")) {
3730                 opt->merge_detect_rename = 1;
3731                 opt->rename_score = 0;
3732         }
3733         else if (skip_prefix(s, "find-renames=", &arg) ||
3734                  skip_prefix(s, "rename-threshold=", &arg)) {
3735                 if ((opt->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3736                         return -1;
3737                 opt->merge_detect_rename = 1;
3738         }
3739         /*
3740          * Please update $__git_merge_strategy_options in
3741          * git-completion.bash when you add new options
3742          */
3743         else
3744                 return -1;
3745         return 0;
3746 }