cache-tree: share code between functions writing an index as a tree
[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(repo, &one->object.oid, &two->object.oid, &shifted, 0);
157         } else {
158                 shift_tree_by(repo, &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 inline void set_commit_tree(struct commit *c, struct tree *t)
167 {
168         c->maybe_tree = t;
169 }
170
171 static struct commit *make_virtual_commit(struct repository *repo,
172                                           struct tree *tree,
173                                           const char *comment)
174 {
175         struct commit *commit = alloc_commit_node(repo);
176
177         set_merge_remote_desc(commit, comment, (struct object *)commit);
178         set_commit_tree(commit, tree);
179         commit->object.parsed = 1;
180         return commit;
181 }
182
183 /*
184  * Since we use get_tree_entry(), which does not put the read object into
185  * the object pool, we cannot rely on a == b.
186  */
187 static int oid_eq(const struct object_id *a, const struct object_id *b)
188 {
189         if (!a && !b)
190                 return 2;
191         return a && b && oideq(a, b);
192 }
193
194 enum rename_type {
195         RENAME_NORMAL = 0,
196         RENAME_VIA_DIR,
197         RENAME_ADD,
198         RENAME_DELETE,
199         RENAME_ONE_FILE_TO_ONE,
200         RENAME_ONE_FILE_TO_TWO,
201         RENAME_TWO_FILES_TO_ONE
202 };
203
204 /*
205  * Since we want to write the index eventually, we cannot reuse the index
206  * for these (temporary) data.
207  */
208 struct stage_data {
209         struct diff_filespec stages[4]; /* mostly for oid & mode; maybe path */
210         struct rename_conflict_info *rename_conflict_info;
211         unsigned processed:1;
212 };
213
214 struct rename {
215         unsigned processed:1;
216         struct diff_filepair *pair;
217         const char *branch; /* branch that the rename occurred on */
218         /*
219          * If directory rename detection affected this rename, what was its
220          * original type ('A' or 'R') and it's original destination before
221          * the directory rename (otherwise, '\0' and NULL for these two vars).
222          */
223         char dir_rename_original_type;
224         char *dir_rename_original_dest;
225         /*
226          * Purpose of src_entry and dst_entry:
227          *
228          * If 'before' is renamed to 'after' then src_entry will contain
229          * the versions of 'before' from the merge_base, HEAD, and MERGE in
230          * stages 1, 2, and 3; dst_entry will contain the respective
231          * versions of 'after' in corresponding locations.  Thus, we have a
232          * total of six modes and oids, though some will be null.  (Stage 0
233          * is ignored; we're interested in handling conflicts.)
234          *
235          * Since we don't turn on break-rewrites by default, neither
236          * src_entry nor dst_entry can have all three of their stages have
237          * non-null oids, meaning at most four of the six will be non-null.
238          * Also, since this is a rename, both src_entry and dst_entry will
239          * have at least one non-null oid, meaning at least two will be
240          * non-null.  Of the six oids, a typical rename will have three be
241          * non-null.  Only two implies a rename/delete, and four implies a
242          * rename/add.
243          */
244         struct stage_data *src_entry;
245         struct stage_data *dst_entry;
246 };
247
248 struct rename_conflict_info {
249         enum rename_type rename_type;
250         struct rename *ren1;
251         struct rename *ren2;
252 };
253
254 static inline void setup_rename_conflict_info(enum rename_type rename_type,
255                                               struct merge_options *opt,
256                                               struct rename *ren1,
257                                               struct rename *ren2)
258 {
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 && ren1->branch != opt->branch1) {
269                 setup_rename_conflict_info(rename_type, opt, ren2, ren1);
270                 return;
271         }
272
273         ci = xcalloc(1, sizeof(struct rename_conflict_info));
274         ci->rename_type = rename_type;
275         ci->ren1 = ren1;
276         ci->ren2 = ren2;
277
278         ci->ren1->dst_entry->processed = 0;
279         ci->ren1->dst_entry->rename_conflict_info = ci;
280         if (ren2) {
281                 ci->ren2->dst_entry->rename_conflict_info = ci;
282         }
283 }
284
285 static int show(struct merge_options *opt, int v)
286 {
287         return (!opt->call_depth && opt->verbosity >= v) || opt->verbosity >= 5;
288 }
289
290 __attribute__((format (printf, 3, 4)))
291 static void output(struct merge_options *opt, int v, const char *fmt, ...)
292 {
293         va_list ap;
294
295         if (!show(opt, v))
296                 return;
297
298         strbuf_addchars(&opt->obuf, ' ', opt->call_depth * 2);
299
300         va_start(ap, fmt);
301         strbuf_vaddf(&opt->obuf, fmt, ap);
302         va_end(ap);
303
304         strbuf_addch(&opt->obuf, '\n');
305         if (!opt->buffer_output)
306                 flush_output(opt);
307 }
308
309 static void output_commit_title(struct merge_options *opt, struct commit *commit)
310 {
311         struct merge_remote_desc *desc;
312
313         strbuf_addchars(&opt->obuf, ' ', opt->call_depth * 2);
314         desc = merge_remote_util(commit);
315         if (desc)
316                 strbuf_addf(&opt->obuf, "virtual %s\n", desc->name);
317         else {
318                 strbuf_add_unique_abbrev(&opt->obuf, &commit->object.oid,
319                                          DEFAULT_ABBREV);
320                 strbuf_addch(&opt->obuf, ' ');
321                 if (parse_commit(commit) != 0)
322                         strbuf_addstr(&opt->obuf, _("(bad commit)\n"));
323                 else {
324                         const char *title;
325                         const char *msg = get_commit_buffer(commit, NULL);
326                         int len = find_commit_subject(msg, &title);
327                         if (len)
328                                 strbuf_addf(&opt->obuf, "%.*s\n", len, title);
329                         unuse_commit_buffer(commit, msg);
330                 }
331         }
332         flush_output(opt);
333 }
334
335 static int add_cacheinfo(struct merge_options *opt,
336                          const struct diff_filespec *blob,
337                          const char *path, int stage, int refresh, int options)
338 {
339         struct index_state *istate = opt->repo->index;
340         struct cache_entry *ce;
341         int ret;
342
343         ce = make_cache_entry(istate, blob->mode, &blob->oid, path, stage, 0);
344         if (!ce)
345                 return err(opt, _("add_cacheinfo failed for path '%s'; merge aborting."), path);
346
347         ret = add_index_entry(istate, ce, options);
348         if (refresh) {
349                 struct cache_entry *nce;
350
351                 nce = refresh_cache_entry(istate, ce,
352                                           CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
353                 if (!nce)
354                         return err(opt, _("add_cacheinfo failed to refresh for path '%s'; merge aborting."), path);
355                 if (nce != ce)
356                         ret = add_index_entry(istate, nce, options);
357         }
358         return ret;
359 }
360
361 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
362 {
363         parse_tree(tree);
364         init_tree_desc(desc, tree->buffer, tree->size);
365 }
366
367 static int unpack_trees_start(struct merge_options *opt,
368                               struct tree *common,
369                               struct tree *head,
370                               struct tree *merge)
371 {
372         int rc;
373         struct tree_desc t[3];
374         struct index_state tmp_index = { NULL };
375
376         memset(&opt->unpack_opts, 0, sizeof(opt->unpack_opts));
377         if (opt->call_depth)
378                 opt->unpack_opts.index_only = 1;
379         else
380                 opt->unpack_opts.update = 1;
381         opt->unpack_opts.merge = 1;
382         opt->unpack_opts.head_idx = 2;
383         opt->unpack_opts.fn = threeway_merge;
384         opt->unpack_opts.src_index = opt->repo->index;
385         opt->unpack_opts.dst_index = &tmp_index;
386         opt->unpack_opts.aggressive = !merge_detect_rename(opt);
387         setup_unpack_trees_porcelain(&opt->unpack_opts, "merge");
388
389         init_tree_desc_from_tree(t+0, common);
390         init_tree_desc_from_tree(t+1, head);
391         init_tree_desc_from_tree(t+2, merge);
392
393         rc = unpack_trees(3, t, &opt->unpack_opts);
394         cache_tree_free(&opt->repo->index->cache_tree);
395
396         /*
397          * Update opt->repo->index to match the new results, AFTER saving a copy
398          * in opt->orig_index.  Update src_index to point to the saved copy.
399          * (verify_uptodate() checks src_index, and the original index is
400          * the one that had the necessary modification timestamps.)
401          */
402         opt->orig_index = *opt->repo->index;
403         *opt->repo->index = tmp_index;
404         opt->unpack_opts.src_index = &opt->orig_index;
405
406         return rc;
407 }
408
409 static void unpack_trees_finish(struct merge_options *opt)
410 {
411         discard_index(&opt->orig_index);
412         clear_unpack_trees_porcelain(&opt->unpack_opts);
413 }
414
415 static int save_files_dirs(const struct object_id *oid,
416                            struct strbuf *base, const char *path,
417                            unsigned int mode, int stage, void *context)
418 {
419         struct path_hashmap_entry *entry;
420         int baselen = base->len;
421         struct merge_options *opt = context;
422
423         strbuf_addstr(base, path);
424
425         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
426         hashmap_entry_init(entry, path_hash(entry->path));
427         hashmap_add(&opt->current_file_dir_set, entry);
428
429         strbuf_setlen(base, baselen);
430         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
431 }
432
433 static void get_files_dirs(struct merge_options *opt, struct tree *tree)
434 {
435         struct pathspec match_all;
436         memset(&match_all, 0, sizeof(match_all));
437         read_tree_recursive(opt->repo, tree, "", 0, 0,
438                             &match_all, save_files_dirs, opt);
439 }
440
441 static int get_tree_entry_if_blob(struct repository *r,
442                                   const struct object_id *tree,
443                                   const char *path,
444                                   struct diff_filespec *dfs)
445 {
446         int ret;
447
448         ret = get_tree_entry(r, tree, path, &dfs->oid, &dfs->mode);
449         if (S_ISDIR(dfs->mode)) {
450                 oidcpy(&dfs->oid, &null_oid);
451                 dfs->mode = 0;
452         }
453         return ret;
454 }
455
456 /*
457  * Returns an index_entry instance which doesn't have to correspond to
458  * a real cache entry in Git's index.
459  */
460 static struct stage_data *insert_stage_data(struct repository *r,
461                 const char *path,
462                 struct tree *o, struct tree *a, struct tree *b,
463                 struct string_list *entries)
464 {
465         struct string_list_item *item;
466         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
467         get_tree_entry_if_blob(r, &o->object.oid, path, &e->stages[1]);
468         get_tree_entry_if_blob(r, &a->object.oid, path, &e->stages[2]);
469         get_tree_entry_if_blob(r, &b->object.oid, path, &e->stages[3]);
470         item = string_list_insert(entries, path);
471         item->util = e;
472         return e;
473 }
474
475 /*
476  * Create a dictionary mapping file names to stage_data objects. The
477  * dictionary contains one entry for every path with a non-zero stage entry.
478  */
479 static struct string_list *get_unmerged(struct index_state *istate)
480 {
481         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
482         int i;
483
484         unmerged->strdup_strings = 1;
485
486         for (i = 0; i < istate->cache_nr; i++) {
487                 struct string_list_item *item;
488                 struct stage_data *e;
489                 const struct cache_entry *ce = istate->cache[i];
490                 if (!ce_stage(ce))
491                         continue;
492
493                 item = string_list_lookup(unmerged, ce->name);
494                 if (!item) {
495                         item = string_list_insert(unmerged, ce->name);
496                         item->util = xcalloc(1, sizeof(struct stage_data));
497                 }
498                 e = item->util;
499                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
500                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
501         }
502
503         return unmerged;
504 }
505
506 static int string_list_df_name_compare(const char *one, const char *two)
507 {
508         int onelen = strlen(one);
509         int twolen = strlen(two);
510         /*
511          * Here we only care that entries for D/F conflicts are
512          * adjacent, in particular with the file of the D/F conflict
513          * appearing before files below the corresponding directory.
514          * The order of the rest of the list is irrelevant for us.
515          *
516          * To achieve this, we sort with df_name_compare and provide
517          * the mode S_IFDIR so that D/F conflicts will sort correctly.
518          * We use the mode S_IFDIR for everything else for simplicity,
519          * since in other cases any changes in their order due to
520          * sorting cause no problems for us.
521          */
522         int cmp = df_name_compare(one, onelen, S_IFDIR,
523                                   two, twolen, S_IFDIR);
524         /*
525          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
526          * that 'foo' comes before 'foo/bar'.
527          */
528         if (cmp)
529                 return cmp;
530         return onelen - twolen;
531 }
532
533 static void record_df_conflict_files(struct merge_options *opt,
534                                      struct string_list *entries)
535 {
536         /* If there is a D/F conflict and the file for such a conflict
537          * currently exists in the working tree, we want to allow it to be
538          * removed to make room for the corresponding directory if needed.
539          * The files underneath the directories of such D/F conflicts will
540          * be processed before the corresponding file involved in the D/F
541          * conflict.  If the D/F directory ends up being removed by the
542          * merge, then we won't have to touch the D/F file.  If the D/F
543          * directory needs to be written to the working copy, then the D/F
544          * file will simply be removed (in make_room_for_path()) to make
545          * room for the necessary paths.  Note that if both the directory
546          * and the file need to be present, then the D/F file will be
547          * reinstated with a new unique name at the time it is processed.
548          */
549         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
550         const char *last_file = NULL;
551         int last_len = 0;
552         int i;
553
554         /*
555          * If we're merging merge-bases, we don't want to bother with
556          * any working directory changes.
557          */
558         if (opt->call_depth)
559                 return;
560
561         /* Ensure D/F conflicts are adjacent in the entries list. */
562         for (i = 0; i < entries->nr; i++) {
563                 struct string_list_item *next = &entries->items[i];
564                 string_list_append(&df_sorted_entries, next->string)->util =
565                                    next->util;
566         }
567         df_sorted_entries.cmp = string_list_df_name_compare;
568         string_list_sort(&df_sorted_entries);
569
570         string_list_clear(&opt->df_conflict_file_set, 1);
571         for (i = 0; i < df_sorted_entries.nr; i++) {
572                 const char *path = df_sorted_entries.items[i].string;
573                 int len = strlen(path);
574                 struct stage_data *e = df_sorted_entries.items[i].util;
575
576                 /*
577                  * Check if last_file & path correspond to a D/F conflict;
578                  * i.e. whether path is last_file+'/'+<something>.
579                  * If so, record that it's okay to remove last_file to make
580                  * room for path and friends if needed.
581                  */
582                 if (last_file &&
583                     len > last_len &&
584                     memcmp(path, last_file, last_len) == 0 &&
585                     path[last_len] == '/') {
586                         string_list_insert(&opt->df_conflict_file_set, last_file);
587                 }
588
589                 /*
590                  * Determine whether path could exist as a file in the
591                  * working directory as a possible D/F conflict.  This
592                  * will only occur when it exists in stage 2 as a
593                  * file.
594                  */
595                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
596                         last_file = path;
597                         last_len = len;
598                 } else {
599                         last_file = NULL;
600                 }
601         }
602         string_list_clear(&df_sorted_entries, 0);
603 }
604
605 static int update_stages(struct merge_options *opt, const char *path,
606                          const struct diff_filespec *o,
607                          const struct diff_filespec *a,
608                          const struct diff_filespec *b)
609 {
610
611         /*
612          * NOTE: It is usually a bad idea to call update_stages on a path
613          * before calling update_file on that same path, since it can
614          * sometimes lead to spurious "refusing to lose untracked file..."
615          * messages from update_file (via make_room_for path via
616          * would_lose_untracked).  Instead, reverse the order of the calls
617          * (executing update_file first and then update_stages).
618          */
619         int clear = 1;
620         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
621         if (clear)
622                 if (remove_file_from_index(opt->repo->index, path))
623                         return -1;
624         if (o)
625                 if (add_cacheinfo(opt, o, path, 1, 0, options))
626                         return -1;
627         if (a)
628                 if (add_cacheinfo(opt, a, path, 2, 0, options))
629                         return -1;
630         if (b)
631                 if (add_cacheinfo(opt, b, path, 3, 0, options))
632                         return -1;
633         return 0;
634 }
635
636 static void update_entry(struct stage_data *entry,
637                          struct diff_filespec *o,
638                          struct diff_filespec *a,
639                          struct diff_filespec *b)
640 {
641         entry->processed = 0;
642         entry->stages[1].mode = o->mode;
643         entry->stages[2].mode = a->mode;
644         entry->stages[3].mode = b->mode;
645         oidcpy(&entry->stages[1].oid, &o->oid);
646         oidcpy(&entry->stages[2].oid, &a->oid);
647         oidcpy(&entry->stages[3].oid, &b->oid);
648 }
649
650 static int remove_file(struct merge_options *opt, int clean,
651                        const char *path, int no_wd)
652 {
653         int update_cache = opt->call_depth || clean;
654         int update_working_directory = !opt->call_depth && !no_wd;
655
656         if (update_cache) {
657                 if (remove_file_from_index(opt->repo->index, path))
658                         return -1;
659         }
660         if (update_working_directory) {
661                 if (ignore_case) {
662                         struct cache_entry *ce;
663                         ce = index_file_exists(opt->repo->index, path, strlen(path),
664                                                ignore_case);
665                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
666                                 return 0;
667                 }
668                 if (remove_path(path))
669                         return -1;
670         }
671         return 0;
672 }
673
674 /* add a string to a strbuf, but converting "/" to "_" */
675 static void add_flattened_path(struct strbuf *out, const char *s)
676 {
677         size_t i = out->len;
678         strbuf_addstr(out, s);
679         for (; i < out->len; i++)
680                 if (out->buf[i] == '/')
681                         out->buf[i] = '_';
682 }
683
684 static char *unique_path(struct merge_options *opt, const char *path, const char *branch)
685 {
686         struct path_hashmap_entry *entry;
687         struct strbuf newpath = STRBUF_INIT;
688         int suffix = 0;
689         size_t base_len;
690
691         strbuf_addf(&newpath, "%s~", path);
692         add_flattened_path(&newpath, branch);
693
694         base_len = newpath.len;
695         while (hashmap_get_from_hash(&opt->current_file_dir_set,
696                                      path_hash(newpath.buf), newpath.buf) ||
697                (!opt->call_depth && file_exists(newpath.buf))) {
698                 strbuf_setlen(&newpath, base_len);
699                 strbuf_addf(&newpath, "_%d", suffix++);
700         }
701
702         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
703         hashmap_entry_init(entry, path_hash(entry->path));
704         hashmap_add(&opt->current_file_dir_set, entry);
705         return strbuf_detach(&newpath, NULL);
706 }
707
708 /**
709  * Check whether a directory in the index is in the way of an incoming
710  * file.  Return 1 if so.  If check_working_copy is non-zero, also
711  * check the working directory.  If empty_ok is non-zero, also return
712  * 0 in the case where the working-tree dir exists but is empty.
713  */
714 static int dir_in_way(struct index_state *istate, const char *path,
715                       int check_working_copy, int empty_ok)
716 {
717         int pos;
718         struct strbuf dirpath = STRBUF_INIT;
719         struct stat st;
720
721         strbuf_addstr(&dirpath, path);
722         strbuf_addch(&dirpath, '/');
723
724         pos = index_name_pos(istate, dirpath.buf, dirpath.len);
725
726         if (pos < 0)
727                 pos = -1 - pos;
728         if (pos < istate->cache_nr &&
729             !strncmp(dirpath.buf, istate->cache[pos]->name, dirpath.len)) {
730                 strbuf_release(&dirpath);
731                 return 1;
732         }
733
734         strbuf_release(&dirpath);
735         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
736                 !(empty_ok && is_empty_dir(path));
737 }
738
739 /*
740  * Returns whether path was tracked in the index before the merge started,
741  * and its oid and mode match the specified values
742  */
743 static int was_tracked_and_matches(struct merge_options *opt, const char *path,
744                                    const struct diff_filespec *blob)
745 {
746         int pos = index_name_pos(&opt->orig_index, path, strlen(path));
747         struct cache_entry *ce;
748
749         if (0 > pos)
750                 /* we were not tracking this path before the merge */
751                 return 0;
752
753         /* See if the file we were tracking before matches */
754         ce = opt->orig_index.cache[pos];
755         return (oid_eq(&ce->oid, &blob->oid) && ce->ce_mode == blob->mode);
756 }
757
758 /*
759  * Returns whether path was tracked in the index before the merge started
760  */
761 static int was_tracked(struct merge_options *opt, const char *path)
762 {
763         int pos = index_name_pos(&opt->orig_index, path, strlen(path));
764
765         if (0 <= pos)
766                 /* we were tracking this path before the merge */
767                 return 1;
768
769         return 0;
770 }
771
772 static int would_lose_untracked(struct merge_options *opt, const char *path)
773 {
774         struct index_state *istate = opt->repo->index;
775
776         /*
777          * This may look like it can be simplified to:
778          *   return !was_tracked(opt, path) && file_exists(path)
779          * but it can't.  This function needs to know whether path was in
780          * the working tree due to EITHER having been tracked in the index
781          * before the merge OR having been put into the working copy and
782          * index by unpack_trees().  Due to that either-or requirement, we
783          * check the current index instead of the original one.
784          *
785          * Note that we do not need to worry about merge-recursive itself
786          * updating the index after unpack_trees() and before calling this
787          * function, because we strictly require all code paths in
788          * merge-recursive to update the working tree first and the index
789          * second.  Doing otherwise would break
790          * update_file()/would_lose_untracked(); see every comment in this
791          * file which mentions "update_stages".
792          */
793         int pos = index_name_pos(istate, path, strlen(path));
794
795         if (pos < 0)
796                 pos = -1 - pos;
797         while (pos < istate->cache_nr &&
798                !strcmp(path, istate->cache[pos]->name)) {
799                 /*
800                  * If stage #0, it is definitely tracked.
801                  * If it has stage #2 then it was tracked
802                  * before this merge started.  All other
803                  * cases the path was not tracked.
804                  */
805                 switch (ce_stage(istate->cache[pos])) {
806                 case 0:
807                 case 2:
808                         return 0;
809                 }
810                 pos++;
811         }
812         return file_exists(path);
813 }
814
815 static int was_dirty(struct merge_options *opt, const char *path)
816 {
817         struct cache_entry *ce;
818         int dirty = 1;
819
820         if (opt->call_depth || !was_tracked(opt, path))
821                 return !dirty;
822
823         ce = index_file_exists(opt->unpack_opts.src_index,
824                                path, strlen(path), ignore_case);
825         dirty = verify_uptodate(ce, &opt->unpack_opts) != 0;
826         return dirty;
827 }
828
829 static int make_room_for_path(struct merge_options *opt, const char *path)
830 {
831         int status, i;
832         const char *msg = _("failed to create path '%s'%s");
833
834         /* Unlink any D/F conflict files that are in the way */
835         for (i = 0; i < opt->df_conflict_file_set.nr; i++) {
836                 const char *df_path = opt->df_conflict_file_set.items[i].string;
837                 size_t pathlen = strlen(path);
838                 size_t df_pathlen = strlen(df_path);
839                 if (df_pathlen < pathlen &&
840                     path[df_pathlen] == '/' &&
841                     strncmp(path, df_path, df_pathlen) == 0) {
842                         output(opt, 3,
843                                _("Removing %s to make room for subdirectory\n"),
844                                df_path);
845                         unlink(df_path);
846                         unsorted_string_list_delete_item(&opt->df_conflict_file_set,
847                                                          i, 0);
848                         break;
849                 }
850         }
851
852         /* Make sure leading directories are created */
853         status = safe_create_leading_directories_const(path);
854         if (status) {
855                 if (status == SCLD_EXISTS)
856                         /* something else exists */
857                         return err(opt, msg, path, _(": perhaps a D/F conflict?"));
858                 return err(opt, msg, path, "");
859         }
860
861         /*
862          * Do not unlink a file in the work tree if we are not
863          * tracking it.
864          */
865         if (would_lose_untracked(opt, path))
866                 return err(opt, _("refusing to lose untracked file at '%s'"),
867                            path);
868
869         /* Successful unlink is good.. */
870         if (!unlink(path))
871                 return 0;
872         /* .. and so is no existing file */
873         if (errno == ENOENT)
874                 return 0;
875         /* .. but not some other error (who really cares what?) */
876         return err(opt, msg, path, _(": perhaps a D/F conflict?"));
877 }
878
879 static int update_file_flags(struct merge_options *opt,
880                              const struct diff_filespec *contents,
881                              const char *path,
882                              int update_cache,
883                              int update_wd)
884 {
885         int ret = 0;
886
887         if (opt->call_depth)
888                 update_wd = 0;
889
890         if (update_wd) {
891                 enum object_type type;
892                 void *buf;
893                 unsigned long size;
894
895                 if (S_ISGITLINK(contents->mode)) {
896                         /*
897                          * We may later decide to recursively descend into
898                          * the submodule directory and update its index
899                          * and/or work tree, but we do not do that now.
900                          */
901                         update_wd = 0;
902                         goto update_index;
903                 }
904
905                 buf = read_object_file(&contents->oid, &type, &size);
906                 if (!buf) {
907                         ret = err(opt, _("cannot read object %s '%s'"),
908                                   oid_to_hex(&contents->oid), path);
909                         goto free_buf;
910                 }
911                 if (type != OBJ_BLOB) {
912                         ret = err(opt, _("blob expected for %s '%s'"),
913                                   oid_to_hex(&contents->oid), path);
914                         goto free_buf;
915                 }
916                 if (S_ISREG(contents->mode)) {
917                         struct strbuf strbuf = STRBUF_INIT;
918                         if (convert_to_working_tree(opt->repo->index, path, buf, size, &strbuf)) {
919                                 free(buf);
920                                 size = strbuf.len;
921                                 buf = strbuf_detach(&strbuf, NULL);
922                         }
923                 }
924
925                 if (make_room_for_path(opt, path) < 0) {
926                         update_wd = 0;
927                         goto free_buf;
928                 }
929                 if (S_ISREG(contents->mode) ||
930                     (!has_symlinks && S_ISLNK(contents->mode))) {
931                         int fd;
932                         int mode = (contents->mode & 0100 ? 0777 : 0666);
933
934                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
935                         if (fd < 0) {
936                                 ret = err(opt, _("failed to open '%s': %s"),
937                                           path, strerror(errno));
938                                 goto free_buf;
939                         }
940                         write_in_full(fd, buf, size);
941                         close(fd);
942                 } else if (S_ISLNK(contents->mode)) {
943                         char *lnk = xmemdupz(buf, size);
944                         safe_create_leading_directories_const(path);
945                         unlink(path);
946                         if (symlink(lnk, path))
947                                 ret = err(opt, _("failed to symlink '%s': %s"),
948                                           path, strerror(errno));
949                         free(lnk);
950                 } else
951                         ret = err(opt,
952                                   _("do not know what to do with %06o %s '%s'"),
953                                   contents->mode, oid_to_hex(&contents->oid), path);
954         free_buf:
955                 free(buf);
956         }
957 update_index:
958         if (!ret && update_cache)
959                 if (add_cacheinfo(opt, contents, path, 0, update_wd,
960                                   ADD_CACHE_OK_TO_ADD))
961                         return -1;
962         return ret;
963 }
964
965 static int update_file(struct merge_options *opt,
966                        int clean,
967                        const struct diff_filespec *contents,
968                        const char *path)
969 {
970         return update_file_flags(opt, contents, path,
971                                  opt->call_depth || clean, !opt->call_depth);
972 }
973
974 /* Low level file merging, update and removal */
975
976 struct merge_file_info {
977         struct diff_filespec blob; /* mostly use oid & mode; sometimes path */
978         unsigned clean:1,
979                  merge:1;
980 };
981
982 static int merge_3way(struct merge_options *opt,
983                       mmbuffer_t *result_buf,
984                       const struct diff_filespec *o,
985                       const struct diff_filespec *a,
986                       const struct diff_filespec *b,
987                       const char *branch1,
988                       const char *branch2,
989                       const int extra_marker_size)
990 {
991         mmfile_t orig, src1, src2;
992         struct ll_merge_options ll_opts = {0};
993         char *base, *name1, *name2;
994         int merge_status;
995
996         ll_opts.renormalize = opt->renormalize;
997         ll_opts.extra_marker_size = extra_marker_size;
998         ll_opts.xdl_opts = opt->xdl_opts;
999
1000         if (opt->call_depth) {
1001                 ll_opts.virtual_ancestor = 1;
1002                 ll_opts.variant = 0;
1003         } else {
1004                 switch (opt->recursive_variant) {
1005                 case MERGE_RECURSIVE_OURS:
1006                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
1007                         break;
1008                 case MERGE_RECURSIVE_THEIRS:
1009                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
1010                         break;
1011                 default:
1012                         ll_opts.variant = 0;
1013                         break;
1014                 }
1015         }
1016
1017         assert(a->path && b->path && o->path && opt->ancestor);
1018         if (strcmp(a->path, b->path) || strcmp(a->path, o->path) != 0) {
1019                 base  = mkpathdup("%s:%s", opt->ancestor, o->path);
1020                 name1 = mkpathdup("%s:%s", branch1, a->path);
1021                 name2 = mkpathdup("%s:%s", branch2, b->path);
1022         } else {
1023                 base  = mkpathdup("%s", opt->ancestor);
1024                 name1 = mkpathdup("%s", branch1);
1025                 name2 = mkpathdup("%s", branch2);
1026         }
1027
1028         read_mmblob(&orig, &o->oid);
1029         read_mmblob(&src1, &a->oid);
1030         read_mmblob(&src2, &b->oid);
1031
1032         merge_status = ll_merge(result_buf, a->path, &orig, base,
1033                                 &src1, name1, &src2, name2,
1034                                 opt->repo->index, &ll_opts);
1035
1036         free(base);
1037         free(name1);
1038         free(name2);
1039         free(orig.ptr);
1040         free(src1.ptr);
1041         free(src2.ptr);
1042         return merge_status;
1043 }
1044
1045 static int find_first_merges(struct repository *repo,
1046                              struct object_array *result, const char *path,
1047                              struct commit *a, struct commit *b)
1048 {
1049         int i, j;
1050         struct object_array merges = OBJECT_ARRAY_INIT;
1051         struct commit *commit;
1052         int contains_another;
1053
1054         char merged_revision[GIT_MAX_HEXSZ + 2];
1055         const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
1056                                    "--all", merged_revision, NULL };
1057         struct rev_info revs;
1058         struct setup_revision_opt rev_opts;
1059
1060         memset(result, 0, sizeof(struct object_array));
1061         memset(&rev_opts, 0, sizeof(rev_opts));
1062
1063         /* get all revisions that merge commit a */
1064         xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
1065                   oid_to_hex(&a->object.oid));
1066         repo_init_revisions(repo, &revs, NULL);
1067         rev_opts.submodule = path;
1068         /* FIXME: can't handle linked worktrees in submodules yet */
1069         revs.single_worktree = path != NULL;
1070         setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
1071
1072         /* save all revisions from the above list that contain b */
1073         if (prepare_revision_walk(&revs))
1074                 die("revision walk setup failed");
1075         while ((commit = get_revision(&revs)) != NULL) {
1076                 struct object *o = &(commit->object);
1077                 if (in_merge_bases(b, commit))
1078                         add_object_array(o, NULL, &merges);
1079         }
1080         reset_revision_walk();
1081
1082         /* Now we've got all merges that contain a and b. Prune all
1083          * merges that contain another found merge and save them in
1084          * result.
1085          */
1086         for (i = 0; i < merges.nr; i++) {
1087                 struct commit *m1 = (struct commit *) merges.objects[i].item;
1088
1089                 contains_another = 0;
1090                 for (j = 0; j < merges.nr; j++) {
1091                         struct commit *m2 = (struct commit *) merges.objects[j].item;
1092                         if (i != j && in_merge_bases(m2, m1)) {
1093                                 contains_another = 1;
1094                                 break;
1095                         }
1096                 }
1097
1098                 if (!contains_another)
1099                         add_object_array(merges.objects[i].item, NULL, result);
1100         }
1101
1102         object_array_clear(&merges);
1103         return result->nr;
1104 }
1105
1106 static void print_commit(struct commit *commit)
1107 {
1108         struct strbuf sb = STRBUF_INIT;
1109         struct pretty_print_context ctx = {0};
1110         ctx.date_mode.type = DATE_NORMAL;
1111         format_commit_message(commit, " %h: %m %s", &sb, &ctx);
1112         fprintf(stderr, "%s\n", sb.buf);
1113         strbuf_release(&sb);
1114 }
1115
1116 static int is_valid(const struct diff_filespec *dfs)
1117 {
1118         return dfs->mode != 0 && !is_null_oid(&dfs->oid);
1119 }
1120
1121 static int merge_submodule(struct merge_options *opt,
1122                            struct object_id *result, const char *path,
1123                            const struct object_id *base, const struct object_id *a,
1124                            const struct object_id *b)
1125 {
1126         struct commit *commit_base, *commit_a, *commit_b;
1127         int parent_count;
1128         struct object_array merges;
1129
1130         int i;
1131         int search = !opt->call_depth;
1132
1133         /* store a in result in case we fail */
1134         oidcpy(result, a);
1135
1136         /* we can not handle deletion conflicts */
1137         if (is_null_oid(base))
1138                 return 0;
1139         if (is_null_oid(a))
1140                 return 0;
1141         if (is_null_oid(b))
1142                 return 0;
1143
1144         if (add_submodule_odb(path)) {
1145                 output(opt, 1, _("Failed to merge submodule %s (not checked out)"), path);
1146                 return 0;
1147         }
1148
1149         if (!(commit_base = lookup_commit_reference(opt->repo, base)) ||
1150             !(commit_a = lookup_commit_reference(opt->repo, a)) ||
1151             !(commit_b = lookup_commit_reference(opt->repo, b))) {
1152                 output(opt, 1, _("Failed to merge submodule %s (commits not present)"), path);
1153                 return 0;
1154         }
1155
1156         /* check whether both changes are forward */
1157         if (!in_merge_bases(commit_base, commit_a) ||
1158             !in_merge_bases(commit_base, commit_b)) {
1159                 output(opt, 1, _("Failed to merge submodule %s (commits don't follow merge-base)"), path);
1160                 return 0;
1161         }
1162
1163         /* Case #1: a is contained in b or vice versa */
1164         if (in_merge_bases(commit_a, commit_b)) {
1165                 oidcpy(result, b);
1166                 if (show(opt, 3)) {
1167                         output(opt, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1168                         output_commit_title(opt, commit_b);
1169                 } else if (show(opt, 2))
1170                         output(opt, 2, _("Fast-forwarding submodule %s"), path);
1171                 else
1172                         ; /* no output */
1173
1174                 return 1;
1175         }
1176         if (in_merge_bases(commit_b, commit_a)) {
1177                 oidcpy(result, a);
1178                 if (show(opt, 3)) {
1179                         output(opt, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1180                         output_commit_title(opt, commit_a);
1181                 } else if (show(opt, 2))
1182                         output(opt, 2, _("Fast-forwarding submodule %s"), path);
1183                 else
1184                         ; /* no output */
1185
1186                 return 1;
1187         }
1188
1189         /*
1190          * Case #2: There are one or more merges that contain a and b in
1191          * the submodule. If there is only one, then present it as a
1192          * suggestion to the user, but leave it marked unmerged so the
1193          * user needs to confirm the resolution.
1194          */
1195
1196         /* Skip the search if makes no sense to the calling context.  */
1197         if (!search)
1198                 return 0;
1199
1200         /* find commit which merges them */
1201         parent_count = find_first_merges(opt->repo, &merges, path,
1202                                          commit_a, commit_b);
1203         switch (parent_count) {
1204         case 0:
1205                 output(opt, 1, _("Failed to merge submodule %s (merge following commits not found)"), path);
1206                 break;
1207
1208         case 1:
1209                 output(opt, 1, _("Failed to merge submodule %s (not fast-forward)"), path);
1210                 output(opt, 2, _("Found a possible merge resolution for the submodule:\n"));
1211                 print_commit((struct commit *) merges.objects[0].item);
1212                 output(opt, 2, _(
1213                        "If this is correct simply add it to the index "
1214                        "for example\n"
1215                        "by using:\n\n"
1216                        "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
1217                        "which will accept this suggestion.\n"),
1218                        oid_to_hex(&merges.objects[0].item->oid), path);
1219                 break;
1220
1221         default:
1222                 output(opt, 1, _("Failed to merge submodule %s (multiple merges found)"), path);
1223                 for (i = 0; i < merges.nr; i++)
1224                         print_commit((struct commit *) merges.objects[i].item);
1225         }
1226
1227         object_array_clear(&merges);
1228         return 0;
1229 }
1230
1231 static int merge_mode_and_contents(struct merge_options *opt,
1232                                    const struct diff_filespec *o,
1233                                    const struct diff_filespec *a,
1234                                    const struct diff_filespec *b,
1235                                    const char *filename,
1236                                    const char *branch1,
1237                                    const char *branch2,
1238                                    const int extra_marker_size,
1239                                    struct merge_file_info *result)
1240 {
1241         if (opt->branch1 != branch1) {
1242                 /*
1243                  * It's weird getting a reverse merge with HEAD on the bottom
1244                  * side of the conflict markers and the other branch on the
1245                  * top.  Fix that.
1246                  */
1247                 return merge_mode_and_contents(opt, o, b, a,
1248                                                filename,
1249                                                branch2, branch1,
1250                                                extra_marker_size, result);
1251         }
1252
1253         result->merge = 0;
1254         result->clean = 1;
1255
1256         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
1257                 result->clean = 0;
1258                 if (S_ISREG(a->mode)) {
1259                         result->blob.mode = a->mode;
1260                         oidcpy(&result->blob.oid, &a->oid);
1261                 } else {
1262                         result->blob.mode = b->mode;
1263                         oidcpy(&result->blob.oid, &b->oid);
1264                 }
1265         } else {
1266                 if (!oid_eq(&a->oid, &o->oid) && !oid_eq(&b->oid, &o->oid))
1267                         result->merge = 1;
1268
1269                 /*
1270                  * Merge modes
1271                  */
1272                 if (a->mode == b->mode || a->mode == o->mode)
1273                         result->blob.mode = b->mode;
1274                 else {
1275                         result->blob.mode = a->mode;
1276                         if (b->mode != o->mode) {
1277                                 result->clean = 0;
1278                                 result->merge = 1;
1279                         }
1280                 }
1281
1282                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &o->oid))
1283                         oidcpy(&result->blob.oid, &b->oid);
1284                 else if (oid_eq(&b->oid, &o->oid))
1285                         oidcpy(&result->blob.oid, &a->oid);
1286                 else if (S_ISREG(a->mode)) {
1287                         mmbuffer_t result_buf;
1288                         int ret = 0, merge_status;
1289
1290                         merge_status = merge_3way(opt, &result_buf, o, a, b,
1291                                                   branch1, branch2,
1292                                                   extra_marker_size);
1293
1294                         if ((merge_status < 0) || !result_buf.ptr)
1295                                 ret = err(opt, _("Failed to execute internal merge"));
1296
1297                         if (!ret &&
1298                             write_object_file(result_buf.ptr, result_buf.size,
1299                                               blob_type, &result->blob.oid))
1300                                 ret = err(opt, _("Unable to add %s to database"),
1301                                           a->path);
1302
1303                         free(result_buf.ptr);
1304                         if (ret)
1305                                 return ret;
1306                         result->clean = (merge_status == 0);
1307                 } else if (S_ISGITLINK(a->mode)) {
1308                         result->clean = merge_submodule(opt, &result->blob.oid,
1309                                                         o->path,
1310                                                         &o->oid,
1311                                                         &a->oid,
1312                                                         &b->oid);
1313                 } else if (S_ISLNK(a->mode)) {
1314                         switch (opt->recursive_variant) {
1315                         case MERGE_RECURSIVE_NORMAL:
1316                                 oidcpy(&result->blob.oid, &a->oid);
1317                                 if (!oid_eq(&a->oid, &b->oid))
1318                                         result->clean = 0;
1319                                 break;
1320                         case MERGE_RECURSIVE_OURS:
1321                                 oidcpy(&result->blob.oid, &a->oid);
1322                                 break;
1323                         case MERGE_RECURSIVE_THEIRS:
1324                                 oidcpy(&result->blob.oid, &b->oid);
1325                                 break;
1326                         }
1327                 } else
1328                         BUG("unsupported object type in the tree");
1329         }
1330
1331         if (result->merge)
1332                 output(opt, 2, _("Auto-merging %s"), filename);
1333
1334         return 0;
1335 }
1336
1337 static int handle_rename_via_dir(struct merge_options *opt,
1338                                  struct rename_conflict_info *ci)
1339 {
1340         /*
1341          * Handle file adds that need to be renamed due to directory rename
1342          * detection.  This differs from handle_rename_normal, because
1343          * there is no content merge to do; just move the file into the
1344          * desired final location.
1345          */
1346         const struct rename *ren = ci->ren1;
1347         const struct diff_filespec *dest = ren->pair->two;
1348         char *file_path = dest->path;
1349         int mark_conflicted = (opt->detect_directory_renames ==
1350                                MERGE_DIRECTORY_RENAMES_CONFLICT);
1351         assert(ren->dir_rename_original_dest);
1352
1353         if (!opt->call_depth && would_lose_untracked(opt, dest->path)) {
1354                 mark_conflicted = 1;
1355                 file_path = unique_path(opt, dest->path, ren->branch);
1356                 output(opt, 1, _("Error: Refusing to lose untracked file at %s; "
1357                                  "writing to %s instead."),
1358                        dest->path, file_path);
1359         }
1360
1361         if (mark_conflicted) {
1362                 /*
1363                  * Write the file in worktree at file_path.  In the index,
1364                  * only record the file at dest->path in the appropriate
1365                  * higher stage.
1366                  */
1367                 if (update_file(opt, 0, dest, file_path))
1368                         return -1;
1369                 if (file_path != dest->path)
1370                         free(file_path);
1371                 if (update_stages(opt, dest->path, NULL,
1372                                   ren->branch == opt->branch1 ? dest : NULL,
1373                                   ren->branch == opt->branch1 ? NULL : dest))
1374                         return -1;
1375                 return 0; /* not clean, but conflicted */
1376         } else {
1377                 /* Update dest->path both in index and in worktree */
1378                 if (update_file(opt, 1, dest, dest->path))
1379                         return -1;
1380                 return 1; /* clean */
1381         }
1382 }
1383
1384 static int handle_change_delete(struct merge_options *opt,
1385                                 const char *path, const char *old_path,
1386                                 const struct diff_filespec *o,
1387                                 const struct diff_filespec *changed,
1388                                 const char *change_branch,
1389                                 const char *delete_branch,
1390                                 const char *change, const char *change_past)
1391 {
1392         char *alt_path = NULL;
1393         const char *update_path = path;
1394         int ret = 0;
1395
1396         if (dir_in_way(opt->repo->index, path, !opt->call_depth, 0) ||
1397             (!opt->call_depth && would_lose_untracked(opt, path))) {
1398                 update_path = alt_path = unique_path(opt, path, change_branch);
1399         }
1400
1401         if (opt->call_depth) {
1402                 /*
1403                  * We cannot arbitrarily accept either a_sha or b_sha as
1404                  * correct; since there is no true "middle point" between
1405                  * them, simply reuse the base version for virtual merge base.
1406                  */
1407                 ret = remove_file_from_index(opt->repo->index, path);
1408                 if (!ret)
1409                         ret = update_file(opt, 0, o, update_path);
1410         } else {
1411                 /*
1412                  * Despite the four nearly duplicate messages and argument
1413                  * lists below and the ugliness of the nested if-statements,
1414                  * having complete messages makes the job easier for
1415                  * translators.
1416                  *
1417                  * The slight variance among the cases is due to the fact
1418                  * that:
1419                  *   1) directory/file conflicts (in effect if
1420                  *      !alt_path) could cause us to need to write the
1421                  *      file to a different path.
1422                  *   2) renames (in effect if !old_path) could mean that
1423                  *      there are two names for the path that the user
1424                  *      may know the file by.
1425                  */
1426                 if (!alt_path) {
1427                         if (!old_path) {
1428                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1429                                        "and %s in %s. Version %s of %s left in tree."),
1430                                        change, path, delete_branch, change_past,
1431                                        change_branch, change_branch, path);
1432                         } else {
1433                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1434                                        "and %s to %s in %s. Version %s of %s left in tree."),
1435                                        change, old_path, delete_branch, change_past, path,
1436                                        change_branch, change_branch, path);
1437                         }
1438                 } else {
1439                         if (!old_path) {
1440                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1441                                        "and %s in %s. Version %s of %s left in tree at %s."),
1442                                        change, path, delete_branch, change_past,
1443                                        change_branch, change_branch, path, alt_path);
1444                         } else {
1445                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1446                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1447                                        change, old_path, delete_branch, change_past, path,
1448                                        change_branch, change_branch, path, alt_path);
1449                         }
1450                 }
1451                 /*
1452                  * No need to call update_file() on path when change_branch ==
1453                  * opt->branch1 && !alt_path, since that would needlessly touch
1454                  * path.  We could call update_file_flags() with update_cache=0
1455                  * and update_wd=0, but that's a no-op.
1456                  */
1457                 if (change_branch != opt->branch1 || alt_path)
1458                         ret = update_file(opt, 0, changed, update_path);
1459         }
1460         free(alt_path);
1461
1462         return ret;
1463 }
1464
1465 static int handle_rename_delete(struct merge_options *opt,
1466                                 struct rename_conflict_info *ci)
1467 {
1468         const struct rename *ren = ci->ren1;
1469         const struct diff_filespec *orig = ren->pair->one;
1470         const struct diff_filespec *dest = ren->pair->two;
1471         const char *rename_branch = ren->branch;
1472         const char *delete_branch = (opt->branch1 == ren->branch ?
1473                                      opt->branch2 : opt->branch1);
1474
1475         if (handle_change_delete(opt,
1476                                  opt->call_depth ? orig->path : dest->path,
1477                                  opt->call_depth ? NULL : orig->path,
1478                                  orig, dest,
1479                                  rename_branch, delete_branch,
1480                                  _("rename"), _("renamed")))
1481                 return -1;
1482
1483         if (opt->call_depth)
1484                 return remove_file_from_index(opt->repo->index, dest->path);
1485         else
1486                 return update_stages(opt, dest->path, NULL,
1487                                      rename_branch == opt->branch1 ? dest : NULL,
1488                                      rename_branch == opt->branch1 ? NULL : dest);
1489 }
1490
1491 static int handle_file_collision(struct merge_options *opt,
1492                                  const char *collide_path,
1493                                  const char *prev_path1,
1494                                  const char *prev_path2,
1495                                  const char *branch1, const char *branch2,
1496                                  struct diff_filespec *a,
1497                                  struct diff_filespec *b)
1498 {
1499         struct merge_file_info mfi;
1500         struct diff_filespec null;
1501         char *alt_path = NULL;
1502         const char *update_path = collide_path;
1503
1504         /*
1505          * It's easiest to get the correct things into stage 2 and 3, and
1506          * to make sure that the content merge puts HEAD before the other
1507          * branch if we just ensure that branch1 == opt->branch1.  So, simply
1508          * flip arguments around if we don't have that.
1509          */
1510         if (branch1 != opt->branch1) {
1511                 return handle_file_collision(opt, collide_path,
1512                                              prev_path2, prev_path1,
1513                                              branch2, branch1,
1514                                              b, a);
1515         }
1516
1517         /*
1518          * In the recursive case, we just opt to undo renames
1519          */
1520         if (opt->call_depth && (prev_path1 || prev_path2)) {
1521                 /* Put first file (a->oid, a->mode) in its original spot */
1522                 if (prev_path1) {
1523                         if (update_file(opt, 1, a, prev_path1))
1524                                 return -1;
1525                 } else {
1526                         if (update_file(opt, 1, a, collide_path))
1527                                 return -1;
1528                 }
1529
1530                 /* Put second file (b->oid, b->mode) in its original spot */
1531                 if (prev_path2) {
1532                         if (update_file(opt, 1, b, prev_path2))
1533                                 return -1;
1534                 } else {
1535                         if (update_file(opt, 1, b, collide_path))
1536                                 return -1;
1537                 }
1538
1539                 /* Don't leave something at collision path if unrenaming both */
1540                 if (prev_path1 && prev_path2)
1541                         remove_file(opt, 1, collide_path, 0);
1542
1543                 return 0;
1544         }
1545
1546         /* Remove rename sources if rename/add or rename/rename(2to1) */
1547         if (prev_path1)
1548                 remove_file(opt, 1, prev_path1,
1549                             opt->call_depth || would_lose_untracked(opt, prev_path1));
1550         if (prev_path2)
1551                 remove_file(opt, 1, prev_path2,
1552                             opt->call_depth || would_lose_untracked(opt, prev_path2));
1553
1554         /*
1555          * Remove the collision path, if it wouldn't cause dirty contents
1556          * or an untracked file to get lost.  We'll either overwrite with
1557          * merged contents, or just write out to differently named files.
1558          */
1559         if (was_dirty(opt, collide_path)) {
1560                 output(opt, 1, _("Refusing to lose dirty file at %s"),
1561                        collide_path);
1562                 update_path = alt_path = unique_path(opt, collide_path, "merged");
1563         } else if (would_lose_untracked(opt, collide_path)) {
1564                 /*
1565                  * Only way we get here is if both renames were from
1566                  * a directory rename AND user had an untracked file
1567                  * at the location where both files end up after the
1568                  * two directory renames.  See testcase 10d of t6043.
1569                  */
1570                 output(opt, 1, _("Refusing to lose untracked file at "
1571                                "%s, even though it's in the way."),
1572                        collide_path);
1573                 update_path = alt_path = unique_path(opt, collide_path, "merged");
1574         } else {
1575                 /*
1576                  * FIXME: It's possible that the two files are identical
1577                  * and that the current working copy happens to match, in
1578                  * which case we are unnecessarily touching the working
1579                  * tree file.  It's not a likely enough scenario that I
1580                  * want to code up the checks for it and a better fix is
1581                  * available if we restructure how unpack_trees() and
1582                  * merge-recursive interoperate anyway, so punting for
1583                  * now...
1584                  */
1585                 remove_file(opt, 0, collide_path, 0);
1586         }
1587
1588         /* Store things in diff_filespecs for functions that need it */
1589         null.path = (char *)collide_path;
1590         oidcpy(&null.oid, &null_oid);
1591         null.mode = 0;
1592
1593         if (merge_mode_and_contents(opt, &null, a, b, collide_path,
1594                                     branch1, branch2, opt->call_depth * 2, &mfi))
1595                 return -1;
1596         mfi.clean &= !alt_path;
1597         if (update_file(opt, mfi.clean, &mfi.blob, update_path))
1598                 return -1;
1599         if (!mfi.clean && !opt->call_depth &&
1600             update_stages(opt, collide_path, NULL, a, b))
1601                 return -1;
1602         free(alt_path);
1603         /*
1604          * FIXME: If both a & b both started with conflicts (only possible
1605          * if they came from a rename/rename(2to1)), but had IDENTICAL
1606          * contents including those conflicts, then in the next line we claim
1607          * it was clean.  If someone cares about this case, we should have the
1608          * caller notify us if we started with conflicts.
1609          */
1610         return mfi.clean;
1611 }
1612
1613 static int handle_rename_add(struct merge_options *opt,
1614                              struct rename_conflict_info *ci)
1615 {
1616         /* a was renamed to c, and a separate c was added. */
1617         struct diff_filespec *a = ci->ren1->pair->one;
1618         struct diff_filespec *c = ci->ren1->pair->two;
1619         char *path = c->path;
1620         char *prev_path_desc;
1621         struct merge_file_info mfi;
1622
1623         const char *rename_branch = ci->ren1->branch;
1624         const char *add_branch = (opt->branch1 == rename_branch ?
1625                                   opt->branch2 : opt->branch1);
1626         int other_stage = (ci->ren1->branch == opt->branch1 ? 3 : 2);
1627
1628         output(opt, 1, _("CONFLICT (rename/add): "
1629                "Rename %s->%s in %s.  Added %s in %s"),
1630                a->path, c->path, rename_branch,
1631                c->path, add_branch);
1632
1633         prev_path_desc = xstrfmt("version of %s from %s", path, a->path);
1634         ci->ren1->src_entry->stages[other_stage].path = a->path;
1635         if (merge_mode_and_contents(opt, a, c,
1636                                     &ci->ren1->src_entry->stages[other_stage],
1637                                     prev_path_desc,
1638                                     opt->branch1, opt->branch2,
1639                                     1 + opt->call_depth * 2, &mfi))
1640                 return -1;
1641         free(prev_path_desc);
1642
1643         ci->ren1->dst_entry->stages[other_stage].path = mfi.blob.path = c->path;
1644         return handle_file_collision(opt,
1645                                      c->path, a->path, NULL,
1646                                      rename_branch, add_branch,
1647                                      &mfi.blob,
1648                                      &ci->ren1->dst_entry->stages[other_stage]);
1649 }
1650
1651 static char *find_path_for_conflict(struct merge_options *opt,
1652                                     const char *path,
1653                                     const char *branch1,
1654                                     const char *branch2)
1655 {
1656         char *new_path = NULL;
1657         if (dir_in_way(opt->repo->index, path, !opt->call_depth, 0)) {
1658                 new_path = unique_path(opt, path, branch1);
1659                 output(opt, 1, _("%s is a directory in %s adding "
1660                                "as %s instead"),
1661                        path, branch2, new_path);
1662         } else if (would_lose_untracked(opt, path)) {
1663                 new_path = unique_path(opt, path, branch1);
1664                 output(opt, 1, _("Refusing to lose untracked file"
1665                                " at %s; adding as %s instead"),
1666                        path, new_path);
1667         }
1668
1669         return new_path;
1670 }
1671
1672 static int handle_rename_rename_1to2(struct merge_options *opt,
1673                                      struct rename_conflict_info *ci)
1674 {
1675         /* One file was renamed in both branches, but to different names. */
1676         struct merge_file_info mfi;
1677         struct diff_filespec *add;
1678         struct diff_filespec *o = ci->ren1->pair->one;
1679         struct diff_filespec *a = ci->ren1->pair->two;
1680         struct diff_filespec *b = ci->ren2->pair->two;
1681         char *path_desc;
1682
1683         output(opt, 1, _("CONFLICT (rename/rename): "
1684                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1685                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1686                o->path, a->path, ci->ren1->branch,
1687                o->path, b->path, ci->ren2->branch,
1688                opt->call_depth ? _(" (left unresolved)") : "");
1689
1690         path_desc = xstrfmt("%s and %s, both renamed from %s",
1691                             a->path, b->path, o->path);
1692         if (merge_mode_and_contents(opt, o, a, b, path_desc,
1693                                     ci->ren1->branch, ci->ren2->branch,
1694                                     opt->call_depth * 2, &mfi))
1695                 return -1;
1696         free(path_desc);
1697
1698         if (opt->call_depth) {
1699                 /*
1700                  * FIXME: For rename/add-source conflicts (if we could detect
1701                  * such), this is wrong.  We should instead find a unique
1702                  * pathname and then either rename the add-source file to that
1703                  * unique path, or use that unique path instead of src here.
1704                  */
1705                 if (update_file(opt, 0, &mfi.blob, o->path))
1706                         return -1;
1707
1708                 /*
1709                  * Above, we put the merged content at the merge-base's
1710                  * path.  Now we usually need to delete both a->path and
1711                  * b->path.  However, the rename on each side of the merge
1712                  * could also be involved in a rename/add conflict.  In
1713                  * such cases, we should keep the added file around,
1714                  * resolving the conflict at that path in its favor.
1715                  */
1716                 add = &ci->ren1->dst_entry->stages[2 ^ 1];
1717                 if (is_valid(add)) {
1718                         if (update_file(opt, 0, add, a->path))
1719                                 return -1;
1720                 }
1721                 else
1722                         remove_file_from_index(opt->repo->index, a->path);
1723                 add = &ci->ren2->dst_entry->stages[3 ^ 1];
1724                 if (is_valid(add)) {
1725                         if (update_file(opt, 0, add, b->path))
1726                                 return -1;
1727                 }
1728                 else
1729                         remove_file_from_index(opt->repo->index, b->path);
1730         } else {
1731                 /*
1732                  * For each destination path, we need to see if there is a
1733                  * rename/add collision.  If not, we can write the file out
1734                  * to the specified location.
1735                  */
1736                 add = &ci->ren1->dst_entry->stages[2 ^ 1];
1737                 if (is_valid(add)) {
1738                         add->path = mfi.blob.path = a->path;
1739                         if (handle_file_collision(opt, a->path,
1740                                                   NULL, NULL,
1741                                                   ci->ren1->branch,
1742                                                   ci->ren2->branch,
1743                                                   &mfi.blob, add) < 0)
1744                                 return -1;
1745                 } else {
1746                         char *new_path = find_path_for_conflict(opt, a->path,
1747                                                                 ci->ren1->branch,
1748                                                                 ci->ren2->branch);
1749                         if (update_file(opt, 0, &mfi.blob,
1750                                         new_path ? new_path : a->path))
1751                                 return -1;
1752                         free(new_path);
1753                         if (update_stages(opt, a->path, NULL, a, NULL))
1754                                 return -1;
1755                 }
1756
1757                 add = &ci->ren2->dst_entry->stages[3 ^ 1];
1758                 if (is_valid(add)) {
1759                         add->path = mfi.blob.path = b->path;
1760                         if (handle_file_collision(opt, b->path,
1761                                                   NULL, NULL,
1762                                                   ci->ren1->branch,
1763                                                   ci->ren2->branch,
1764                                                   add, &mfi.blob) < 0)
1765                                 return -1;
1766                 } else {
1767                         char *new_path = find_path_for_conflict(opt, b->path,
1768                                                                 ci->ren2->branch,
1769                                                                 ci->ren1->branch);
1770                         if (update_file(opt, 0, &mfi.blob,
1771                                         new_path ? new_path : b->path))
1772                                 return -1;
1773                         free(new_path);
1774                         if (update_stages(opt, b->path, NULL, NULL, b))
1775                                 return -1;
1776                 }
1777         }
1778
1779         return 0;
1780 }
1781
1782 static int handle_rename_rename_2to1(struct merge_options *opt,
1783                                      struct rename_conflict_info *ci)
1784 {
1785         /* Two files, a & b, were renamed to the same thing, c. */
1786         struct diff_filespec *a = ci->ren1->pair->one;
1787         struct diff_filespec *b = ci->ren2->pair->one;
1788         struct diff_filespec *c1 = ci->ren1->pair->two;
1789         struct diff_filespec *c2 = ci->ren2->pair->two;
1790         char *path = c1->path; /* == c2->path */
1791         char *path_side_1_desc;
1792         char *path_side_2_desc;
1793         struct merge_file_info mfi_c1;
1794         struct merge_file_info mfi_c2;
1795         int ostage1, ostage2;
1796
1797         output(opt, 1, _("CONFLICT (rename/rename): "
1798                "Rename %s->%s in %s. "
1799                "Rename %s->%s in %s"),
1800                a->path, c1->path, ci->ren1->branch,
1801                b->path, c2->path, ci->ren2->branch);
1802
1803         path_side_1_desc = xstrfmt("version of %s from %s", path, a->path);
1804         path_side_2_desc = xstrfmt("version of %s from %s", path, b->path);
1805         ostage1 = ci->ren1->branch == opt->branch1 ? 3 : 2;
1806         ostage2 = ostage1 ^ 1;
1807         ci->ren1->src_entry->stages[ostage1].path = a->path;
1808         ci->ren2->src_entry->stages[ostage2].path = b->path;
1809         if (merge_mode_and_contents(opt, a, c1,
1810                                     &ci->ren1->src_entry->stages[ostage1],
1811                                     path_side_1_desc,
1812                                     opt->branch1, opt->branch2,
1813                                     1 + opt->call_depth * 2, &mfi_c1) ||
1814             merge_mode_and_contents(opt, b,
1815                                     &ci->ren2->src_entry->stages[ostage2],
1816                                     c2, path_side_2_desc,
1817                                     opt->branch1, opt->branch2,
1818                                     1 + opt->call_depth * 2, &mfi_c2))
1819                 return -1;
1820         free(path_side_1_desc);
1821         free(path_side_2_desc);
1822         mfi_c1.blob.path = path;
1823         mfi_c2.blob.path = path;
1824
1825         return handle_file_collision(opt, path, a->path, b->path,
1826                                      ci->ren1->branch, ci->ren2->branch,
1827                                      &mfi_c1.blob, &mfi_c2.blob);
1828 }
1829
1830 /*
1831  * Get the diff_filepairs changed between o_tree and tree.
1832  */
1833 static struct diff_queue_struct *get_diffpairs(struct merge_options *opt,
1834                                                struct tree *o_tree,
1835                                                struct tree *tree)
1836 {
1837         struct diff_queue_struct *ret;
1838         struct diff_options opts;
1839
1840         repo_diff_setup(opt->repo, &opts);
1841         opts.flags.recursive = 1;
1842         opts.flags.rename_empty = 0;
1843         opts.detect_rename = merge_detect_rename(opt);
1844         /*
1845          * We do not have logic to handle the detection of copies.  In
1846          * fact, it may not even make sense to add such logic: would we
1847          * really want a change to a base file to be propagated through
1848          * multiple other files by a merge?
1849          */
1850         if (opts.detect_rename > DIFF_DETECT_RENAME)
1851                 opts.detect_rename = DIFF_DETECT_RENAME;
1852         opts.rename_limit = opt->merge_rename_limit >= 0 ? opt->merge_rename_limit :
1853                             opt->diff_rename_limit >= 0 ? opt->diff_rename_limit :
1854                             1000;
1855         opts.rename_score = opt->rename_score;
1856         opts.show_rename_progress = opt->show_rename_progress;
1857         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1858         diff_setup_done(&opts);
1859         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1860         diffcore_std(&opts);
1861         if (opts.needed_rename_limit > opt->needed_rename_limit)
1862                 opt->needed_rename_limit = opts.needed_rename_limit;
1863
1864         ret = xmalloc(sizeof(*ret));
1865         *ret = diff_queued_diff;
1866
1867         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1868         diff_queued_diff.nr = 0;
1869         diff_queued_diff.queue = NULL;
1870         diff_flush(&opts);
1871         return ret;
1872 }
1873
1874 static int tree_has_path(struct repository *r, struct tree *tree,
1875                          const char *path)
1876 {
1877         struct object_id hashy;
1878         unsigned short mode_o;
1879
1880         return !get_tree_entry(r,
1881                                &tree->object.oid, path,
1882                                &hashy, &mode_o);
1883 }
1884
1885 /*
1886  * Return a new string that replaces the beginning portion (which matches
1887  * entry->dir), with entry->new_dir.  In perl-speak:
1888  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1889  * NOTE:
1890  *   Caller must ensure that old_path starts with entry->dir + '/'.
1891  */
1892 static char *apply_dir_rename(struct dir_rename_entry *entry,
1893                               const char *old_path)
1894 {
1895         struct strbuf new_path = STRBUF_INIT;
1896         int oldlen, newlen;
1897
1898         if (entry->non_unique_new_dir)
1899                 return NULL;
1900
1901         oldlen = strlen(entry->dir);
1902         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1903         strbuf_grow(&new_path, newlen);
1904         strbuf_addbuf(&new_path, &entry->new_dir);
1905         strbuf_addstr(&new_path, &old_path[oldlen]);
1906
1907         return strbuf_detach(&new_path, NULL);
1908 }
1909
1910 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1911                                     char **old_dir, char **new_dir)
1912 {
1913         char *end_of_old, *end_of_new;
1914         int old_len, new_len;
1915
1916         *old_dir = NULL;
1917         *new_dir = NULL;
1918
1919         /*
1920          * For
1921          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1922          * the "e/foo.c" part is the same, we just want to know that
1923          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1924          * so, for this example, this function returns "a/b/c/d" in
1925          * *old_dir and "a/b/some/thing/else" in *new_dir.
1926          *
1927          * Also, if the basename of the file changed, we don't care.  We
1928          * want to know which portion of the directory, if any, changed.
1929          */
1930         end_of_old = strrchr(old_path, '/');
1931         end_of_new = strrchr(new_path, '/');
1932
1933         if (end_of_old == NULL || end_of_new == NULL)
1934                 return;
1935         while (*--end_of_new == *--end_of_old &&
1936                end_of_old != old_path &&
1937                end_of_new != new_path)
1938                 ; /* Do nothing; all in the while loop */
1939         /*
1940          * We've found the first non-matching character in the directory
1941          * paths.  That means the current directory we were comparing
1942          * represents the rename.  Move end_of_old and end_of_new back
1943          * to the full directory name.
1944          */
1945         if (*end_of_old == '/')
1946                 end_of_old++;
1947         if (*end_of_old != '/')
1948                 end_of_new++;
1949         end_of_old = strchr(end_of_old, '/');
1950         end_of_new = strchr(end_of_new, '/');
1951
1952         /*
1953          * It may have been the case that old_path and new_path were the same
1954          * directory all along.  Don't claim a rename if they're the same.
1955          */
1956         old_len = end_of_old - old_path;
1957         new_len = end_of_new - new_path;
1958
1959         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1960                 *old_dir = xstrndup(old_path, old_len);
1961                 *new_dir = xstrndup(new_path, new_len);
1962         }
1963 }
1964
1965 static void remove_hashmap_entries(struct hashmap *dir_renames,
1966                                    struct string_list *items_to_remove)
1967 {
1968         int i;
1969         struct dir_rename_entry *entry;
1970
1971         for (i = 0; i < items_to_remove->nr; i++) {
1972                 entry = items_to_remove->items[i].util;
1973                 hashmap_remove(dir_renames, entry, NULL);
1974         }
1975         string_list_clear(items_to_remove, 0);
1976 }
1977
1978 /*
1979  * See if there is a directory rename for path, and if there are any file
1980  * level conflicts for the renamed location.  If there is a rename and
1981  * there are no conflicts, return the new name.  Otherwise, return NULL.
1982  */
1983 static char *handle_path_level_conflicts(struct merge_options *opt,
1984                                          const char *path,
1985                                          struct dir_rename_entry *entry,
1986                                          struct hashmap *collisions,
1987                                          struct tree *tree)
1988 {
1989         char *new_path = NULL;
1990         struct collision_entry *collision_ent;
1991         int clean = 1;
1992         struct strbuf collision_paths = STRBUF_INIT;
1993
1994         /*
1995          * entry has the mapping of old directory name to new directory name
1996          * that we want to apply to path.
1997          */
1998         new_path = apply_dir_rename(entry, path);
1999
2000         if (!new_path) {
2001                 /* This should only happen when entry->non_unique_new_dir set */
2002                 if (!entry->non_unique_new_dir)
2003                         BUG("entry->non_unqiue_dir not set and !new_path");
2004                 output(opt, 1, _("CONFLICT (directory rename split): "
2005                                "Unclear where to place %s because directory "
2006                                "%s was renamed to multiple other directories, "
2007                                "with no destination getting a majority of the "
2008                                "files."),
2009                        path, entry->dir);
2010                 clean = 0;
2011                 return NULL;
2012         }
2013
2014         /*
2015          * The caller needs to have ensured that it has pre-populated
2016          * collisions with all paths that map to new_path.  Do a quick check
2017          * to ensure that's the case.
2018          */
2019         collision_ent = collision_find_entry(collisions, new_path);
2020         if (collision_ent == NULL)
2021                 BUG("collision_ent is NULL");
2022
2023         /*
2024          * Check for one-sided add/add/.../add conflicts, i.e.
2025          * where implicit renames from the other side doing
2026          * directory rename(s) can affect this side of history
2027          * to put multiple paths into the same location.  Warn
2028          * and bail on directory renames for such paths.
2029          */
2030         if (collision_ent->reported_already) {
2031                 clean = 0;
2032         } else if (tree_has_path(opt->repo, tree, new_path)) {
2033                 collision_ent->reported_already = 1;
2034                 strbuf_add_separated_string_list(&collision_paths, ", ",
2035                                                  &collision_ent->source_files);
2036                 output(opt, 1, _("CONFLICT (implicit dir rename): Existing "
2037                                "file/dir at %s in the way of implicit "
2038                                "directory rename(s) putting the following "
2039                                "path(s) there: %s."),
2040                        new_path, collision_paths.buf);
2041                 clean = 0;
2042         } else if (collision_ent->source_files.nr > 1) {
2043                 collision_ent->reported_already = 1;
2044                 strbuf_add_separated_string_list(&collision_paths, ", ",
2045                                                  &collision_ent->source_files);
2046                 output(opt, 1, _("CONFLICT (implicit dir rename): Cannot map "
2047                                "more than one path to %s; implicit directory "
2048                                "renames tried to put these paths there: %s"),
2049                        new_path, collision_paths.buf);
2050                 clean = 0;
2051         }
2052
2053         /* Free memory we no longer need */
2054         strbuf_release(&collision_paths);
2055         if (!clean && new_path) {
2056                 free(new_path);
2057                 return NULL;
2058         }
2059
2060         return new_path;
2061 }
2062
2063 /*
2064  * There are a couple things we want to do at the directory level:
2065  *   1. Check for both sides renaming to the same thing, in order to avoid
2066  *      implicit renaming of files that should be left in place.  (See
2067  *      testcase 6b in t6043 for details.)
2068  *   2. Prune directory renames if there are still files left in the
2069  *      the original directory.  These represent a partial directory rename,
2070  *      i.e. a rename where only some of the files within the directory
2071  *      were renamed elsewhere.  (Technically, this could be done earlier
2072  *      in get_directory_renames(), except that would prevent us from
2073  *      doing the previous check and thus failing testcase 6b.)
2074  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2075  *      In the future, we could potentially record this info as well and
2076  *      omit reporting rename/rename(1to2) conflicts for each path within
2077  *      the affected directories, thus cleaning up the merge output.
2078  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2079  *         directory level, because merging directories is fine.  If it
2080  *         causes conflicts for files within those merged directories, then
2081  *         that should be detected at the individual path level.
2082  */
2083 static void handle_directory_level_conflicts(struct merge_options *opt,
2084                                              struct hashmap *dir_re_head,
2085                                              struct tree *head,
2086                                              struct hashmap *dir_re_merge,
2087                                              struct tree *merge)
2088 {
2089         struct hashmap_iter iter;
2090         struct dir_rename_entry *head_ent;
2091         struct dir_rename_entry *merge_ent;
2092
2093         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2094         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2095
2096         hashmap_iter_init(dir_re_head, &iter);
2097         while ((head_ent = hashmap_iter_next(&iter))) {
2098                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2099                 if (merge_ent &&
2100                     !head_ent->non_unique_new_dir &&
2101                     !merge_ent->non_unique_new_dir &&
2102                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2103                         /* 1. Renamed identically; remove it from both sides */
2104                         string_list_append(&remove_from_head,
2105                                            head_ent->dir)->util = head_ent;
2106                         strbuf_release(&head_ent->new_dir);
2107                         string_list_append(&remove_from_merge,
2108                                            merge_ent->dir)->util = merge_ent;
2109                         strbuf_release(&merge_ent->new_dir);
2110                 } else if (tree_has_path(opt->repo, head, head_ent->dir)) {
2111                         /* 2. This wasn't a directory rename after all */
2112                         string_list_append(&remove_from_head,
2113                                            head_ent->dir)->util = head_ent;
2114                         strbuf_release(&head_ent->new_dir);
2115                 }
2116         }
2117
2118         remove_hashmap_entries(dir_re_head, &remove_from_head);
2119         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2120
2121         hashmap_iter_init(dir_re_merge, &iter);
2122         while ((merge_ent = hashmap_iter_next(&iter))) {
2123                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2124                 if (tree_has_path(opt->repo, merge, merge_ent->dir)) {
2125                         /* 2. This wasn't a directory rename after all */
2126                         string_list_append(&remove_from_merge,
2127                                            merge_ent->dir)->util = merge_ent;
2128                 } else if (head_ent &&
2129                            !head_ent->non_unique_new_dir &&
2130                            !merge_ent->non_unique_new_dir) {
2131                         /* 3. rename/rename(1to2) */
2132                         /*
2133                          * We can assume it's not rename/rename(1to1) because
2134                          * that was case (1), already checked above.  So we
2135                          * know that head_ent->new_dir and merge_ent->new_dir
2136                          * are different strings.
2137                          */
2138                         output(opt, 1, _("CONFLICT (rename/rename): "
2139                                        "Rename directory %s->%s in %s. "
2140                                        "Rename directory %s->%s in %s"),
2141                                head_ent->dir, head_ent->new_dir.buf, opt->branch1,
2142                                head_ent->dir, merge_ent->new_dir.buf, opt->branch2);
2143                         string_list_append(&remove_from_head,
2144                                            head_ent->dir)->util = head_ent;
2145                         strbuf_release(&head_ent->new_dir);
2146                         string_list_append(&remove_from_merge,
2147                                            merge_ent->dir)->util = merge_ent;
2148                         strbuf_release(&merge_ent->new_dir);
2149                 }
2150         }
2151
2152         remove_hashmap_entries(dir_re_head, &remove_from_head);
2153         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2154 }
2155
2156 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs)
2157 {
2158         struct hashmap *dir_renames;
2159         struct hashmap_iter iter;
2160         struct dir_rename_entry *entry;
2161         int i;
2162
2163         /*
2164          * Typically, we think of a directory rename as all files from a
2165          * certain directory being moved to a target directory.  However,
2166          * what if someone first moved two files from the original
2167          * directory in one commit, and then renamed the directory
2168          * somewhere else in a later commit?  At merge time, we just know
2169          * that files from the original directory went to two different
2170          * places, and that the bulk of them ended up in the same place.
2171          * We want each directory rename to represent where the bulk of the
2172          * files from that directory end up; this function exists to find
2173          * where the bulk of the files went.
2174          *
2175          * The first loop below simply iterates through the list of file
2176          * renames, finding out how often each directory rename pair
2177          * possibility occurs.
2178          */
2179         dir_renames = xmalloc(sizeof(*dir_renames));
2180         dir_rename_init(dir_renames);
2181         for (i = 0; i < pairs->nr; ++i) {
2182                 struct string_list_item *item;
2183                 int *count;
2184                 struct diff_filepair *pair = pairs->queue[i];
2185                 char *old_dir, *new_dir;
2186
2187                 /* File not part of directory rename if it wasn't renamed */
2188                 if (pair->status != 'R')
2189                         continue;
2190
2191                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2192                                         &old_dir,        &new_dir);
2193                 if (!old_dir)
2194                         /* Directory didn't change at all; ignore this one. */
2195                         continue;
2196
2197                 entry = dir_rename_find_entry(dir_renames, old_dir);
2198                 if (!entry) {
2199                         entry = xmalloc(sizeof(*entry));
2200                         dir_rename_entry_init(entry, old_dir);
2201                         hashmap_put(dir_renames, entry);
2202                 } else {
2203                         free(old_dir);
2204                 }
2205                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2206                 if (!item) {
2207                         item = string_list_insert(&entry->possible_new_dirs,
2208                                                   new_dir);
2209                         item->util = xcalloc(1, sizeof(int));
2210                 } else {
2211                         free(new_dir);
2212                 }
2213                 count = item->util;
2214                 *count += 1;
2215         }
2216
2217         /*
2218          * For each directory with files moved out of it, we find out which
2219          * target directory received the most files so we can declare it to
2220          * be the "winning" target location for the directory rename.  This
2221          * winner gets recorded in new_dir.  If there is no winner
2222          * (multiple target directories received the same number of files),
2223          * we set non_unique_new_dir.  Once we've determined the winner (or
2224          * that there is no winner), we no longer need possible_new_dirs.
2225          */
2226         hashmap_iter_init(dir_renames, &iter);
2227         while ((entry = hashmap_iter_next(&iter))) {
2228                 int max = 0;
2229                 int bad_max = 0;
2230                 char *best = NULL;
2231
2232                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2233                         int *count = entry->possible_new_dirs.items[i].util;
2234
2235                         if (*count == max)
2236                                 bad_max = max;
2237                         else if (*count > max) {
2238                                 max = *count;
2239                                 best = entry->possible_new_dirs.items[i].string;
2240                         }
2241                 }
2242                 if (bad_max == max)
2243                         entry->non_unique_new_dir = 1;
2244                 else {
2245                         assert(entry->new_dir.len == 0);
2246                         strbuf_addstr(&entry->new_dir, best);
2247                 }
2248                 /*
2249                  * The relevant directory sub-portion of the original full
2250                  * filepaths were xstrndup'ed before inserting into
2251                  * possible_new_dirs, and instead of manually iterating the
2252                  * list and free'ing each, just lie and tell
2253                  * possible_new_dirs that it did the strdup'ing so that it
2254                  * will free them for us.
2255                  */
2256                 entry->possible_new_dirs.strdup_strings = 1;
2257                 string_list_clear(&entry->possible_new_dirs, 1);
2258         }
2259
2260         return dir_renames;
2261 }
2262
2263 static struct dir_rename_entry *check_dir_renamed(const char *path,
2264                                                   struct hashmap *dir_renames)
2265 {
2266         char *temp = xstrdup(path);
2267         char *end;
2268         struct dir_rename_entry *entry = NULL;
2269
2270         while ((end = strrchr(temp, '/'))) {
2271                 *end = '\0';
2272                 entry = dir_rename_find_entry(dir_renames, temp);
2273                 if (entry)
2274                         break;
2275         }
2276         free(temp);
2277         return entry;
2278 }
2279
2280 static void compute_collisions(struct hashmap *collisions,
2281                                struct hashmap *dir_renames,
2282                                struct diff_queue_struct *pairs)
2283 {
2284         int i;
2285
2286         /*
2287          * Multiple files can be mapped to the same path due to directory
2288          * renames done by the other side of history.  Since that other
2289          * side of history could have merged multiple directories into one,
2290          * if our side of history added the same file basename to each of
2291          * those directories, then all N of them would get implicitly
2292          * renamed by the directory rename detection into the same path,
2293          * and we'd get an add/add/.../add conflict, and all those adds
2294          * from *this* side of history.  This is not representable in the
2295          * index, and users aren't going to easily be able to make sense of
2296          * it.  So we need to provide a good warning about what's
2297          * happening, and fall back to no-directory-rename detection
2298          * behavior for those paths.
2299          *
2300          * See testcases 9e and all of section 5 from t6043 for examples.
2301          */
2302         collision_init(collisions);
2303
2304         for (i = 0; i < pairs->nr; ++i) {
2305                 struct dir_rename_entry *dir_rename_ent;
2306                 struct collision_entry *collision_ent;
2307                 char *new_path;
2308                 struct diff_filepair *pair = pairs->queue[i];
2309
2310                 if (pair->status != 'A' && pair->status != 'R')
2311                         continue;
2312                 dir_rename_ent = check_dir_renamed(pair->two->path,
2313                                                    dir_renames);
2314                 if (!dir_rename_ent)
2315                         continue;
2316
2317                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2318                 if (!new_path)
2319                         /*
2320                          * dir_rename_ent->non_unique_new_path is true, which
2321                          * means there is no directory rename for us to use,
2322                          * which means it won't cause us any additional
2323                          * collisions.
2324                          */
2325                         continue;
2326                 collision_ent = collision_find_entry(collisions, new_path);
2327                 if (!collision_ent) {
2328                         collision_ent = xcalloc(1,
2329                                                 sizeof(struct collision_entry));
2330                         hashmap_entry_init(collision_ent, strhash(new_path));
2331                         hashmap_put(collisions, collision_ent);
2332                         collision_ent->target_file = new_path;
2333                 } else {
2334                         free(new_path);
2335                 }
2336                 string_list_insert(&collision_ent->source_files,
2337                                    pair->two->path);
2338         }
2339 }
2340
2341 static char *check_for_directory_rename(struct merge_options *opt,
2342                                         const char *path,
2343                                         struct tree *tree,
2344                                         struct hashmap *dir_renames,
2345                                         struct hashmap *dir_rename_exclusions,
2346                                         struct hashmap *collisions,
2347                                         int *clean_merge)
2348 {
2349         char *new_path = NULL;
2350         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2351         struct dir_rename_entry *oentry = NULL;
2352
2353         if (!entry)
2354                 return new_path;
2355
2356         /*
2357          * This next part is a little weird.  We do not want to do an
2358          * implicit rename into a directory we renamed on our side, because
2359          * that will result in a spurious rename/rename(1to2) conflict.  An
2360          * example:
2361          *   Base commit: dumbdir/afile, otherdir/bfile
2362          *   Side 1:      smrtdir/afile, otherdir/bfile
2363          *   Side 2:      dumbdir/afile, dumbdir/bfile
2364          * Here, while working on Side 1, we could notice that otherdir was
2365          * renamed/merged to dumbdir, and change the diff_filepair for
2366          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2367          * 2 will notice the rename from dumbdir to smrtdir, and do the
2368          * transitive rename to move it from dumbdir/bfile to
2369          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2370          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2371          * the file to end up in smrtdir.  And the way to achieve that is
2372          * to not let Side1 do the rename to dumbdir, since we know that is
2373          * the source of one of our directory renames.
2374          *
2375          * That's why oentry and dir_rename_exclusions is here.
2376          *
2377          * As it turns out, this also prevents N-way transient rename
2378          * confusion; See testcases 9c and 9d of t6043.
2379          */
2380         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2381         if (oentry) {
2382                 output(opt, 1, _("WARNING: Avoiding applying %s -> %s rename "
2383                                "to %s, because %s itself was renamed."),
2384                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2385         } else {
2386                 new_path = handle_path_level_conflicts(opt, path, entry,
2387                                                        collisions, tree);
2388                 *clean_merge &= (new_path != NULL);
2389         }
2390
2391         return new_path;
2392 }
2393
2394 static void apply_directory_rename_modifications(struct merge_options *opt,
2395                                                  struct diff_filepair *pair,
2396                                                  char *new_path,
2397                                                  struct rename *re,
2398                                                  struct tree *tree,
2399                                                  struct tree *o_tree,
2400                                                  struct tree *a_tree,
2401                                                  struct tree *b_tree,
2402                                                  struct string_list *entries)
2403 {
2404         struct string_list_item *item;
2405         int stage = (tree == a_tree ? 2 : 3);
2406         int update_wd;
2407
2408         /*
2409          * In all cases where we can do directory rename detection,
2410          * unpack_trees() will have read pair->two->path into the
2411          * index and the working copy.  We need to remove it so that
2412          * we can instead place it at new_path.  It is guaranteed to
2413          * not be untracked (unpack_trees() would have errored out
2414          * saying the file would have been overwritten), but it might
2415          * be dirty, though.
2416          */
2417         update_wd = !was_dirty(opt, pair->two->path);
2418         if (!update_wd)
2419                 output(opt, 1, _("Refusing to lose dirty file at %s"),
2420                        pair->two->path);
2421         remove_file(opt, 1, pair->two->path, !update_wd);
2422
2423         /* Find or create a new re->dst_entry */
2424         item = string_list_lookup(entries, new_path);
2425         if (item) {
2426                 /*
2427                  * Since we're renaming on this side of history, and it's
2428                  * due to a directory rename on the other side of history
2429                  * (which we only allow when the directory in question no
2430                  * longer exists on the other side of history), the
2431                  * original entry for re->dst_entry is no longer
2432                  * necessary...
2433                  */
2434                 re->dst_entry->processed = 1;
2435
2436                 /*
2437                  * ...because we'll be using this new one.
2438                  */
2439                 re->dst_entry = item->util;
2440         } else {
2441                 /*
2442                  * re->dst_entry is for the before-dir-rename path, and we
2443                  * need it to hold information for the after-dir-rename
2444                  * path.  Before creating a new entry, we need to mark the
2445                  * old one as unnecessary (...unless it is shared by
2446                  * src_entry, i.e. this didn't use to be a rename, in which
2447                  * case we can just allow the normal processing to happen
2448                  * for it).
2449                  */
2450                 if (pair->status == 'R')
2451                         re->dst_entry->processed = 1;
2452
2453                 re->dst_entry = insert_stage_data(opt->repo, new_path,
2454                                                   o_tree, a_tree, b_tree,
2455                                                   entries);
2456                 item = string_list_insert(entries, new_path);
2457                 item->util = re->dst_entry;
2458         }
2459
2460         /*
2461          * Update the stage_data with the information about the path we are
2462          * moving into place.  That slot will be empty and available for us
2463          * to write to because of the collision checks in
2464          * handle_path_level_conflicts().  In other words,
2465          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2466          * open for us to write to.
2467          *
2468          * It may be tempting to actually update the index at this point as
2469          * well, using update_stages_for_stage_data(), but as per the big
2470          * "NOTE" in update_stages(), doing so will modify the current
2471          * in-memory index which will break calls to would_lose_untracked()
2472          * that we need to make.  Instead, we need to just make sure that
2473          * the various handle_rename_*() functions update the index
2474          * explicitly rather than relying on unpack_trees() to have done it.
2475          */
2476         get_tree_entry(opt->repo,
2477                        &tree->object.oid,
2478                        pair->two->path,
2479                        &re->dst_entry->stages[stage].oid,
2480                        &re->dst_entry->stages[stage].mode);
2481
2482         /*
2483          * Record the original change status (or 'type' of change).  If it
2484          * was originally an add ('A'), this lets us differentiate later
2485          * between a RENAME_DELETE conflict and RENAME_VIA_DIR (they
2486          * otherwise look the same).  If it was originally a rename ('R'),
2487          * this lets us remember and report accurately about the transitive
2488          * renaming that occurred via the directory rename detection.  Also,
2489          * record the original destination name.
2490          */
2491         re->dir_rename_original_type = pair->status;
2492         re->dir_rename_original_dest = pair->two->path;
2493
2494         /*
2495          * We don't actually look at pair->status again, but it seems
2496          * pedagogically correct to adjust it.
2497          */
2498         pair->status = 'R';
2499
2500         /*
2501          * Finally, record the new location.
2502          */
2503         pair->two->path = new_path;
2504 }
2505
2506 /*
2507  * Get information of all renames which occurred in 'pairs', making use of
2508  * any implicit directory renames inferred from the other side of history.
2509  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2510  * to be able to associate the correct cache entries with the rename
2511  * information; tree is always equal to either a_tree or b_tree.
2512  */
2513 static struct string_list *get_renames(struct merge_options *opt,
2514                                        const char *branch,
2515                                        struct diff_queue_struct *pairs,
2516                                        struct hashmap *dir_renames,
2517                                        struct hashmap *dir_rename_exclusions,
2518                                        struct tree *tree,
2519                                        struct tree *o_tree,
2520                                        struct tree *a_tree,
2521                                        struct tree *b_tree,
2522                                        struct string_list *entries,
2523                                        int *clean_merge)
2524 {
2525         int i;
2526         struct hashmap collisions;
2527         struct hashmap_iter iter;
2528         struct collision_entry *e;
2529         struct string_list *renames;
2530
2531         compute_collisions(&collisions, dir_renames, pairs);
2532         renames = xcalloc(1, sizeof(struct string_list));
2533
2534         for (i = 0; i < pairs->nr; ++i) {
2535                 struct string_list_item *item;
2536                 struct rename *re;
2537                 struct diff_filepair *pair = pairs->queue[i];
2538                 char *new_path; /* non-NULL only with directory renames */
2539
2540                 if (pair->status != 'A' && pair->status != 'R') {
2541                         diff_free_filepair(pair);
2542                         continue;
2543                 }
2544                 new_path = check_for_directory_rename(opt, pair->two->path, tree,
2545                                                       dir_renames,
2546                                                       dir_rename_exclusions,
2547                                                       &collisions,
2548                                                       clean_merge);
2549                 if (pair->status != 'R' && !new_path) {
2550                         diff_free_filepair(pair);
2551                         continue;
2552                 }
2553
2554                 re = xmalloc(sizeof(*re));
2555                 re->processed = 0;
2556                 re->pair = pair;
2557                 re->branch = branch;
2558                 re->dir_rename_original_type = '\0';
2559                 re->dir_rename_original_dest = NULL;
2560                 item = string_list_lookup(entries, re->pair->one->path);
2561                 if (!item)
2562                         re->src_entry = insert_stage_data(opt->repo,
2563                                         re->pair->one->path,
2564                                         o_tree, a_tree, b_tree, entries);
2565                 else
2566                         re->src_entry = item->util;
2567
2568                 item = string_list_lookup(entries, re->pair->two->path);
2569                 if (!item)
2570                         re->dst_entry = insert_stage_data(opt->repo,
2571                                         re->pair->two->path,
2572                                         o_tree, a_tree, b_tree, entries);
2573                 else
2574                         re->dst_entry = item->util;
2575                 item = string_list_insert(renames, pair->one->path);
2576                 item->util = re;
2577                 if (new_path)
2578                         apply_directory_rename_modifications(opt, pair, new_path,
2579                                                              re, tree, o_tree,
2580                                                              a_tree, b_tree,
2581                                                              entries);
2582         }
2583
2584         hashmap_iter_init(&collisions, &iter);
2585         while ((e = hashmap_iter_next(&iter))) {
2586                 free(e->target_file);
2587                 string_list_clear(&e->source_files, 0);
2588         }
2589         hashmap_free(&collisions, 1);
2590         return renames;
2591 }
2592
2593 static int process_renames(struct merge_options *opt,
2594                            struct string_list *a_renames,
2595                            struct string_list *b_renames)
2596 {
2597         int clean_merge = 1, i, j;
2598         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2599         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2600         const struct rename *sre;
2601
2602         for (i = 0; i < a_renames->nr; i++) {
2603                 sre = a_renames->items[i].util;
2604                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2605                         = (void *)sre;
2606         }
2607         for (i = 0; i < b_renames->nr; i++) {
2608                 sre = b_renames->items[i].util;
2609                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2610                         = (void *)sre;
2611         }
2612
2613         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2614                 struct string_list *renames1, *renames2Dst;
2615                 struct rename *ren1 = NULL, *ren2 = NULL;
2616                 const char *ren1_src, *ren1_dst;
2617                 struct string_list_item *lookup;
2618
2619                 if (i >= a_renames->nr) {
2620                         ren2 = b_renames->items[j++].util;
2621                 } else if (j >= b_renames->nr) {
2622                         ren1 = a_renames->items[i++].util;
2623                 } else {
2624                         int compare = strcmp(a_renames->items[i].string,
2625                                              b_renames->items[j].string);
2626                         if (compare <= 0)
2627                                 ren1 = a_renames->items[i++].util;
2628                         if (compare >= 0)
2629                                 ren2 = b_renames->items[j++].util;
2630                 }
2631
2632                 /* TODO: refactor, so that 1/2 are not needed */
2633                 if (ren1) {
2634                         renames1 = a_renames;
2635                         renames2Dst = &b_by_dst;
2636                 } else {
2637                         renames1 = b_renames;
2638                         renames2Dst = &a_by_dst;
2639                         SWAP(ren2, ren1);
2640                 }
2641
2642                 if (ren1->processed)
2643                         continue;
2644                 ren1->processed = 1;
2645                 ren1->dst_entry->processed = 1;
2646                 /* BUG: We should only mark src_entry as processed if we
2647                  * are not dealing with a rename + add-source case.
2648                  */
2649                 ren1->src_entry->processed = 1;
2650
2651                 ren1_src = ren1->pair->one->path;
2652                 ren1_dst = ren1->pair->two->path;
2653
2654                 if (ren2) {
2655                         /* One file renamed on both sides */
2656                         const char *ren2_src = ren2->pair->one->path;
2657                         const char *ren2_dst = ren2->pair->two->path;
2658                         enum rename_type rename_type;
2659                         if (strcmp(ren1_src, ren2_src) != 0)
2660                                 BUG("ren1_src != ren2_src");
2661                         ren2->dst_entry->processed = 1;
2662                         ren2->processed = 1;
2663                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2664                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2665                                 clean_merge = 0;
2666                         } else {
2667                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2668                                 /* BUG: We should only remove ren1_src in
2669                                  * the base stage (think of rename +
2670                                  * add-source cases).
2671                                  */
2672                                 remove_file(opt, 1, ren1_src, 1);
2673                                 update_entry(ren1->dst_entry,
2674                                              ren1->pair->one,
2675                                              ren1->pair->two,
2676                                              ren2->pair->two);
2677                         }
2678                         setup_rename_conflict_info(rename_type, opt, ren1, ren2);
2679                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2680                         /* Two different files renamed to the same thing */
2681                         char *ren2_dst;
2682                         ren2 = lookup->util;
2683                         ren2_dst = ren2->pair->two->path;
2684                         if (strcmp(ren1_dst, ren2_dst) != 0)
2685                                 BUG("ren1_dst != ren2_dst");
2686
2687                         clean_merge = 0;
2688                         ren2->processed = 1;
2689                         /*
2690                          * BUG: We should only mark src_entry as processed
2691                          * if we are not dealing with a rename + add-source
2692                          * case.
2693                          */
2694                         ren2->src_entry->processed = 1;
2695
2696                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2697                                                    opt, ren1, ren2);
2698                 } else {
2699                         /* Renamed in 1, maybe changed in 2 */
2700                         /* we only use sha1 and mode of these */
2701                         struct diff_filespec src_other, dst_other;
2702                         int try_merge;
2703
2704                         /*
2705                          * unpack_trees loads entries from common-commit
2706                          * into stage 1, from head-commit into stage 2, and
2707                          * from merge-commit into stage 3.  We keep track
2708                          * of which side corresponds to the rename.
2709                          */
2710                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2711                         int other_stage =   a_renames == renames1 ? 3 : 2;
2712
2713                         /* BUG: We should only remove ren1_src in the base
2714                          * stage and in other_stage (think of rename +
2715                          * add-source case).
2716                          */
2717                         remove_file(opt, 1, ren1_src,
2718                                     renamed_stage == 2 || !was_tracked(opt, ren1_src));
2719
2720                         oidcpy(&src_other.oid,
2721                                &ren1->src_entry->stages[other_stage].oid);
2722                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2723                         oidcpy(&dst_other.oid,
2724                                &ren1->dst_entry->stages[other_stage].oid);
2725                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2726                         try_merge = 0;
2727
2728                         if (oid_eq(&src_other.oid, &null_oid) &&
2729                             ren1->dir_rename_original_type == 'A') {
2730                                 setup_rename_conflict_info(RENAME_VIA_DIR,
2731                                                            opt, ren1, NULL);
2732                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2733                                 setup_rename_conflict_info(RENAME_DELETE,
2734                                                            opt, ren1, NULL);
2735                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2736                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2737                                 /*
2738                                  * Added file on the other side identical to
2739                                  * the file being renamed: clean merge.
2740                                  * Also, there is no need to overwrite the
2741                                  * file already in the working copy, so call
2742                                  * update_file_flags() instead of
2743                                  * update_file().
2744                                  */
2745                                 if (update_file_flags(opt,
2746                                                       ren1->pair->two,
2747                                                       ren1_dst,
2748                                                       1, /* update_cache */
2749                                                       0  /* update_wd    */))
2750                                         clean_merge = -1;
2751                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2752                                 /*
2753                                  * Probably not a clean merge, but it's
2754                                  * premature to set clean_merge to 0 here,
2755                                  * because if the rename merges cleanly and
2756                                  * the merge exactly matches the newly added
2757                                  * file, then the merge will be clean.
2758                                  */
2759                                 setup_rename_conflict_info(RENAME_ADD,
2760                                                            opt, ren1, NULL);
2761                         } else
2762                                 try_merge = 1;
2763
2764                         if (clean_merge < 0)
2765                                 goto cleanup_and_return;
2766                         if (try_merge) {
2767                                 struct diff_filespec *o, *a, *b;
2768                                 src_other.path = (char *)ren1_src;
2769
2770                                 o = ren1->pair->one;
2771                                 if (a_renames == renames1) {
2772                                         a = ren1->pair->two;
2773                                         b = &src_other;
2774                                 } else {
2775                                         b = ren1->pair->two;
2776                                         a = &src_other;
2777                                 }
2778                                 update_entry(ren1->dst_entry, o, a, b);
2779                                 setup_rename_conflict_info(RENAME_NORMAL,
2780                                                            opt, ren1, NULL);
2781                         }
2782                 }
2783         }
2784 cleanup_and_return:
2785         string_list_clear(&a_by_dst, 0);
2786         string_list_clear(&b_by_dst, 0);
2787
2788         return clean_merge;
2789 }
2790
2791 struct rename_info {
2792         struct string_list *head_renames;
2793         struct string_list *merge_renames;
2794 };
2795
2796 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2797                                    struct hashmap *dir_renames)
2798 {
2799         struct hashmap_iter iter;
2800         struct dir_rename_entry *e;
2801
2802         hashmap_iter_init(dir_renames, &iter);
2803         while ((e = hashmap_iter_next(&iter))) {
2804                 free(e->dir);
2805                 strbuf_release(&e->new_dir);
2806                 /* possible_new_dirs already cleared in get_directory_renames */
2807         }
2808         hashmap_free(dir_renames, 1);
2809         free(dir_renames);
2810
2811         free(pairs->queue);
2812         free(pairs);
2813 }
2814
2815 static int detect_and_process_renames(struct merge_options *opt,
2816                                       struct tree *common,
2817                                       struct tree *head,
2818                                       struct tree *merge,
2819                                       struct string_list *entries,
2820                                       struct rename_info *ri)
2821 {
2822         struct diff_queue_struct *head_pairs, *merge_pairs;
2823         struct hashmap *dir_re_head, *dir_re_merge;
2824         int clean = 1;
2825
2826         ri->head_renames = NULL;
2827         ri->merge_renames = NULL;
2828
2829         if (!merge_detect_rename(opt))
2830                 return 1;
2831
2832         head_pairs = get_diffpairs(opt, common, head);
2833         merge_pairs = get_diffpairs(opt, common, merge);
2834
2835         if ((opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_TRUE) ||
2836             (opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_CONFLICT &&
2837              !opt->call_depth)) {
2838                 dir_re_head = get_directory_renames(head_pairs);
2839                 dir_re_merge = get_directory_renames(merge_pairs);
2840
2841                 handle_directory_level_conflicts(opt,
2842                                                  dir_re_head, head,
2843                                                  dir_re_merge, merge);
2844         } else {
2845                 dir_re_head  = xmalloc(sizeof(*dir_re_head));
2846                 dir_re_merge = xmalloc(sizeof(*dir_re_merge));
2847                 dir_rename_init(dir_re_head);
2848                 dir_rename_init(dir_re_merge);
2849         }
2850
2851         ri->head_renames  = get_renames(opt, opt->branch1, head_pairs,
2852                                         dir_re_merge, dir_re_head, head,
2853                                         common, head, merge, entries,
2854                                         &clean);
2855         if (clean < 0)
2856                 goto cleanup;
2857         ri->merge_renames = get_renames(opt, opt->branch2, merge_pairs,
2858                                         dir_re_head, dir_re_merge, merge,
2859                                         common, head, merge, entries,
2860                                         &clean);
2861         if (clean < 0)
2862                 goto cleanup;
2863         clean &= process_renames(opt, ri->head_renames, ri->merge_renames);
2864
2865 cleanup:
2866         /*
2867          * Some cleanup is deferred until cleanup_renames() because the
2868          * data structures are still needed and referenced in
2869          * process_entry().  But there are a few things we can free now.
2870          */
2871         initial_cleanup_rename(head_pairs, dir_re_head);
2872         initial_cleanup_rename(merge_pairs, dir_re_merge);
2873
2874         return clean;
2875 }
2876
2877 static void final_cleanup_rename(struct string_list *rename)
2878 {
2879         const struct rename *re;
2880         int i;
2881
2882         if (rename == NULL)
2883                 return;
2884
2885         for (i = 0; i < rename->nr; i++) {
2886                 re = rename->items[i].util;
2887                 diff_free_filepair(re->pair);
2888         }
2889         string_list_clear(rename, 1);
2890         free(rename);
2891 }
2892
2893 static void final_cleanup_renames(struct rename_info *re_info)
2894 {
2895         final_cleanup_rename(re_info->head_renames);
2896         final_cleanup_rename(re_info->merge_renames);
2897 }
2898
2899 static int read_oid_strbuf(struct merge_options *opt,
2900                            const struct object_id *oid,
2901                            struct strbuf *dst)
2902 {
2903         void *buf;
2904         enum object_type type;
2905         unsigned long size;
2906         buf = read_object_file(oid, &type, &size);
2907         if (!buf)
2908                 return err(opt, _("cannot read object %s"), oid_to_hex(oid));
2909         if (type != OBJ_BLOB) {
2910                 free(buf);
2911                 return err(opt, _("object %s is not a blob"), oid_to_hex(oid));
2912         }
2913         strbuf_attach(dst, buf, size, size + 1);
2914         return 0;
2915 }
2916
2917 static int blob_unchanged(struct merge_options *opt,
2918                           const struct diff_filespec *o,
2919                           const struct diff_filespec *a,
2920                           int renormalize, const char *path)
2921 {
2922         struct strbuf obuf = STRBUF_INIT;
2923         struct strbuf abuf = STRBUF_INIT;
2924         int ret = 0; /* assume changed for safety */
2925         const struct index_state *idx = opt->repo->index;
2926
2927         if (a->mode != o->mode)
2928                 return 0;
2929         if (oid_eq(&o->oid, &a->oid))
2930                 return 1;
2931         if (!renormalize)
2932                 return 0;
2933
2934         if (read_oid_strbuf(opt, &o->oid, &obuf) ||
2935             read_oid_strbuf(opt, &a->oid, &abuf))
2936                 goto error_return;
2937         /*
2938          * Note: binary | is used so that both renormalizations are
2939          * performed.  Comparison can be skipped if both files are
2940          * unchanged since their sha1s have already been compared.
2941          */
2942         if (renormalize_buffer(idx, path, obuf.buf, obuf.len, &obuf) |
2943             renormalize_buffer(idx, path, abuf.buf, abuf.len, &abuf))
2944                 ret = (obuf.len == abuf.len && !memcmp(obuf.buf, abuf.buf, obuf.len));
2945
2946 error_return:
2947         strbuf_release(&obuf);
2948         strbuf_release(&abuf);
2949         return ret;
2950 }
2951
2952 static int handle_modify_delete(struct merge_options *opt,
2953                                 const char *path,
2954                                 const struct diff_filespec *o,
2955                                 const struct diff_filespec *a,
2956                                 const struct diff_filespec *b)
2957 {
2958         const char *modify_branch, *delete_branch;
2959         const struct diff_filespec *changed;
2960
2961         if (is_valid(a)) {
2962                 modify_branch = opt->branch1;
2963                 delete_branch = opt->branch2;
2964                 changed = a;
2965         } else {
2966                 modify_branch = opt->branch2;
2967                 delete_branch = opt->branch1;
2968                 changed = b;
2969         }
2970
2971         return handle_change_delete(opt,
2972                                     path, NULL,
2973                                     o, changed,
2974                                     modify_branch, delete_branch,
2975                                     _("modify"), _("modified"));
2976 }
2977
2978 static int handle_content_merge(struct merge_file_info *mfi,
2979                                 struct merge_options *opt,
2980                                 const char *path,
2981                                 int is_dirty,
2982                                 const struct diff_filespec *o,
2983                                 const struct diff_filespec *a,
2984                                 const struct diff_filespec *b,
2985                                 struct rename_conflict_info *ci)
2986 {
2987         const char *reason = _("content");
2988         unsigned df_conflict_remains = 0;
2989
2990         if (!is_valid(o))
2991                 reason = _("add/add");
2992
2993         assert(o->path && a->path && b->path);
2994         if (ci && dir_in_way(opt->repo->index, path, !opt->call_depth,
2995                              S_ISGITLINK(ci->ren1->pair->two->mode)))
2996                 df_conflict_remains = 1;
2997
2998         if (merge_mode_and_contents(opt, o, a, b, path,
2999                                     opt->branch1, opt->branch2,
3000                                     opt->call_depth * 2, mfi))
3001                 return -1;
3002
3003         /*
3004          * We can skip updating the working tree file iff:
3005          *   a) The merge is clean
3006          *   b) The merge matches what was in HEAD (content, mode, pathname)
3007          *   c) The target path is usable (i.e. not involved in D/F conflict)
3008          */
3009         if (mfi->clean && was_tracked_and_matches(opt, path, &mfi->blob) &&
3010             !df_conflict_remains) {
3011                 int pos;
3012                 struct cache_entry *ce;
3013
3014                 output(opt, 3, _("Skipped %s (merged same as existing)"), path);
3015                 if (add_cacheinfo(opt, &mfi->blob, path,
3016                                   0, (!opt->call_depth && !is_dirty), 0))
3017                         return -1;
3018                 /*
3019                  * However, add_cacheinfo() will delete the old cache entry
3020                  * and add a new one.  We need to copy over any skip_worktree
3021                  * flag to avoid making the file appear as if it were
3022                  * deleted by the user.
3023                  */
3024                 pos = index_name_pos(&opt->orig_index, path, strlen(path));
3025                 ce = opt->orig_index.cache[pos];
3026                 if (ce_skip_worktree(ce)) {
3027                         pos = index_name_pos(opt->repo->index, path, strlen(path));
3028                         ce = opt->repo->index->cache[pos];
3029                         ce->ce_flags |= CE_SKIP_WORKTREE;
3030                 }
3031                 return mfi->clean;
3032         }
3033
3034         if (!mfi->clean) {
3035                 if (S_ISGITLINK(mfi->blob.mode))
3036                         reason = _("submodule");
3037                 output(opt, 1, _("CONFLICT (%s): Merge conflict in %s"),
3038                                 reason, path);
3039                 if (ci && !df_conflict_remains)
3040                         if (update_stages(opt, path, o, a, b))
3041                                 return -1;
3042         }
3043
3044         if (df_conflict_remains || is_dirty) {
3045                 char *new_path;
3046                 if (opt->call_depth) {
3047                         remove_file_from_index(opt->repo->index, path);
3048                 } else {
3049                         if (!mfi->clean) {
3050                                 if (update_stages(opt, path, o, a, b))
3051                                         return -1;
3052                         } else {
3053                                 int file_from_stage2 = was_tracked(opt, path);
3054
3055                                 if (update_stages(opt, path, NULL,
3056                                                   file_from_stage2 ? &mfi->blob : NULL,
3057                                                   file_from_stage2 ? NULL : &mfi->blob))
3058                                         return -1;
3059                         }
3060
3061                 }
3062                 new_path = unique_path(opt, path, ci->ren1->branch);
3063                 if (is_dirty) {
3064                         output(opt, 1, _("Refusing to lose dirty file at %s"),
3065                                path);
3066                 }
3067                 output(opt, 1, _("Adding as %s instead"), new_path);
3068                 if (update_file(opt, 0, &mfi->blob, new_path)) {
3069                         free(new_path);
3070                         return -1;
3071                 }
3072                 free(new_path);
3073                 mfi->clean = 0;
3074         } else if (update_file(opt, mfi->clean, &mfi->blob, path))
3075                 return -1;
3076         return !is_dirty && mfi->clean;
3077 }
3078
3079 static int handle_rename_normal(struct merge_options *opt,
3080                                 const char *path,
3081                                 const struct diff_filespec *o,
3082                                 const struct diff_filespec *a,
3083                                 const struct diff_filespec *b,
3084                                 struct rename_conflict_info *ci)
3085 {
3086         struct rename *ren = ci->ren1;
3087         struct merge_file_info mfi;
3088         int clean;
3089         int side = (ren->branch == opt->branch1 ? 2 : 3);
3090
3091         /* Merge the content and write it out */
3092         clean = handle_content_merge(&mfi, opt, path, was_dirty(opt, path),
3093                                      o, a, b, ci);
3094
3095         if (clean &&
3096             opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_CONFLICT &&
3097             ren->dir_rename_original_dest) {
3098                 if (update_stages(opt, path,
3099                                   NULL,
3100                                   side == 2 ? &mfi.blob : NULL,
3101                                   side == 2 ? NULL : &mfi.blob))
3102                         return -1;
3103                 clean = 0; /* not clean, but conflicted */
3104         }
3105         return clean;
3106 }
3107
3108 static void dir_rename_warning(const char *msg,
3109                                int is_add,
3110                                int clean,
3111                                struct merge_options *opt,
3112                                struct rename *ren)
3113 {
3114         const char *other_branch;
3115         other_branch = (ren->branch == opt->branch1 ?
3116                         opt->branch2 : opt->branch1);
3117         if (is_add) {
3118                 output(opt, clean ? 2 : 1, msg,
3119                        ren->pair->one->path, ren->branch,
3120                        other_branch, ren->pair->two->path);
3121                 return;
3122         }
3123         output(opt, clean ? 2 : 1, msg,
3124                ren->pair->one->path, ren->dir_rename_original_dest, ren->branch,
3125                other_branch, ren->pair->two->path);
3126 }
3127 static int warn_about_dir_renamed_entries(struct merge_options *opt,
3128                                           struct rename *ren)
3129 {
3130         const char *msg;
3131         int clean = 1, is_add;
3132
3133         if (!ren)
3134                 return clean;
3135
3136         /* Return early if ren was not affected/created by a directory rename */
3137         if (!ren->dir_rename_original_dest)
3138                 return clean;
3139
3140         /* Sanity checks */
3141         assert(opt->detect_directory_renames > MERGE_DIRECTORY_RENAMES_NONE);
3142         assert(ren->dir_rename_original_type == 'A' ||
3143                ren->dir_rename_original_type == 'R');
3144
3145         /* Check whether to treat directory renames as a conflict */
3146         clean = (opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_TRUE);
3147
3148         is_add = (ren->dir_rename_original_type == 'A');
3149         if (ren->dir_rename_original_type == 'A' && clean) {
3150                 msg = _("Path updated: %s added in %s inside a "
3151                         "directory that was renamed in %s; moving it to %s.");
3152         } else if (ren->dir_rename_original_type == 'A' && !clean) {
3153                 msg = _("CONFLICT (file location): %s added in %s "
3154                         "inside a directory that was renamed in %s, "
3155                         "suggesting it should perhaps be moved to %s.");
3156         } else if (ren->dir_rename_original_type == 'R' && clean) {
3157                 msg = _("Path updated: %s renamed to %s in %s, inside a "
3158                         "directory that was renamed in %s; moving it to %s.");
3159         } else if (ren->dir_rename_original_type == 'R' && !clean) {
3160                 msg = _("CONFLICT (file location): %s renamed to %s in %s, "
3161                         "inside a directory that was renamed in %s, "
3162                         "suggesting it should perhaps be moved to %s.");
3163         } else {
3164                 BUG("Impossible dir_rename_original_type/clean combination");
3165         }
3166         dir_rename_warning(msg, is_add, clean, opt, ren);
3167
3168         return clean;
3169 }
3170
3171 /* Per entry merge function */
3172 static int process_entry(struct merge_options *opt,
3173                          const char *path, struct stage_data *entry)
3174 {
3175         int clean_merge = 1;
3176         int normalize = opt->renormalize;
3177
3178         struct diff_filespec *o = &entry->stages[1];
3179         struct diff_filespec *a = &entry->stages[2];
3180         struct diff_filespec *b = &entry->stages[3];
3181         int o_valid = is_valid(o);
3182         int a_valid = is_valid(a);
3183         int b_valid = is_valid(b);
3184         o->path = a->path = b->path = (char*)path;
3185
3186         entry->processed = 1;
3187         if (entry->rename_conflict_info) {
3188                 struct rename_conflict_info *ci = entry->rename_conflict_info;
3189                 struct diff_filespec *temp;
3190                 int path_clean;
3191
3192                 path_clean = warn_about_dir_renamed_entries(opt, ci->ren1);
3193                 path_clean &= warn_about_dir_renamed_entries(opt, ci->ren2);
3194
3195                 /*
3196                  * For cases with a single rename, {o,a,b}->path have all been
3197                  * set to the rename target path; we need to set two of these
3198                  * back to the rename source.
3199                  * For rename/rename conflicts, we'll manually fix paths below.
3200                  */
3201                 temp = (opt->branch1 == ci->ren1->branch) ? b : a;
3202                 o->path = temp->path = ci->ren1->pair->one->path;
3203                 if (ci->ren2) {
3204                         assert(opt->branch1 == ci->ren1->branch);
3205                 }
3206
3207                 switch (ci->rename_type) {
3208                 case RENAME_NORMAL:
3209                 case RENAME_ONE_FILE_TO_ONE:
3210                         clean_merge = handle_rename_normal(opt, path, o, a, b,
3211                                                            ci);
3212                         break;
3213                 case RENAME_VIA_DIR:
3214                         clean_merge = handle_rename_via_dir(opt, ci);
3215                         break;
3216                 case RENAME_ADD:
3217                         /*
3218                          * Probably unclean merge, but if the renamed file
3219                          * merges cleanly and the result can then be
3220                          * two-way merged cleanly with the added file, I
3221                          * guess it's a clean merge?
3222                          */
3223                         clean_merge = handle_rename_add(opt, ci);
3224                         break;
3225                 case RENAME_DELETE:
3226                         clean_merge = 0;
3227                         if (handle_rename_delete(opt, ci))
3228                                 clean_merge = -1;
3229                         break;
3230                 case RENAME_ONE_FILE_TO_TWO:
3231                         /*
3232                          * Manually fix up paths; note:
3233                          * ren[12]->pair->one->path are equal.
3234                          */
3235                         o->path = ci->ren1->pair->one->path;
3236                         a->path = ci->ren1->pair->two->path;
3237                         b->path = ci->ren2->pair->two->path;
3238
3239                         clean_merge = 0;
3240                         if (handle_rename_rename_1to2(opt, ci))
3241                                 clean_merge = -1;
3242                         break;
3243                 case RENAME_TWO_FILES_TO_ONE:
3244                         /*
3245                          * Manually fix up paths; note,
3246                          * ren[12]->pair->two->path are actually equal.
3247                          */
3248                         o->path = NULL;
3249                         a->path = ci->ren1->pair->two->path;
3250                         b->path = ci->ren2->pair->two->path;
3251
3252                         /*
3253                          * Probably unclean merge, but if the two renamed
3254                          * files merge cleanly and the two resulting files
3255                          * can then be two-way merged cleanly, I guess it's
3256                          * a clean merge?
3257                          */
3258                         clean_merge = handle_rename_rename_2to1(opt, ci);
3259                         break;
3260                 default:
3261                         entry->processed = 0;
3262                         break;
3263                 }
3264                 if (path_clean < clean_merge)
3265                         clean_merge = path_clean;
3266         } else if (o_valid && (!a_valid || !b_valid)) {
3267                 /* Case A: Deleted in one */
3268                 if ((!a_valid && !b_valid) ||
3269                     (!b_valid && blob_unchanged(opt, o, a, normalize, path)) ||
3270                     (!a_valid && blob_unchanged(opt, o, b, normalize, path))) {
3271                         /* Deleted in both or deleted in one and
3272                          * unchanged in the other */
3273                         if (a_valid)
3274                                 output(opt, 2, _("Removing %s"), path);
3275                         /* do not touch working file if it did not exist */
3276                         remove_file(opt, 1, path, !a_valid);
3277                 } else {
3278                         /* Modify/delete; deleted side may have put a directory in the way */
3279                         clean_merge = 0;
3280                         if (handle_modify_delete(opt, path, o, a, b))
3281                                 clean_merge = -1;
3282                 }
3283         } else if ((!o_valid && a_valid && !b_valid) ||
3284                    (!o_valid && !a_valid && b_valid)) {
3285                 /* Case B: Added in one. */
3286                 /* [nothing|directory] -> ([nothing|directory], file) */
3287
3288                 const char *add_branch;
3289                 const char *other_branch;
3290                 const char *conf;
3291                 const struct diff_filespec *contents;
3292
3293                 if (a_valid) {
3294                         add_branch = opt->branch1;
3295                         other_branch = opt->branch2;
3296                         contents = a;
3297                         conf = _("file/directory");
3298                 } else {
3299                         add_branch = opt->branch2;
3300                         other_branch = opt->branch1;
3301                         contents = b;
3302                         conf = _("directory/file");
3303                 }
3304                 if (dir_in_way(opt->repo->index, path,
3305                                !opt->call_depth && !S_ISGITLINK(a->mode),
3306                                0)) {
3307                         char *new_path = unique_path(opt, path, add_branch);
3308                         clean_merge = 0;
3309                         output(opt, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3310                                "Adding %s as %s"),
3311                                conf, path, other_branch, path, new_path);
3312                         if (update_file(opt, 0, contents, new_path))
3313                                 clean_merge = -1;
3314                         else if (opt->call_depth)
3315                                 remove_file_from_index(opt->repo->index, path);
3316                         free(new_path);
3317                 } else {
3318                         output(opt, 2, _("Adding %s"), path);
3319                         /* do not overwrite file if already present */
3320                         if (update_file_flags(opt, contents, path, 1, !a_valid))
3321                                 clean_merge = -1;
3322                 }
3323         } else if (a_valid && b_valid) {
3324                 if (!o_valid) {
3325                         /* Case C: Added in both (check for same permissions) */
3326                         output(opt, 1,
3327                                _("CONFLICT (add/add): Merge conflict in %s"),
3328                                path);
3329                         clean_merge = handle_file_collision(opt,
3330                                                             path, NULL, NULL,
3331                                                             opt->branch1,
3332                                                             opt->branch2,
3333                                                             a, b);
3334                 } else {
3335                         /* case D: Modified in both, but differently. */
3336                         struct merge_file_info mfi;
3337                         int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3338                         clean_merge = handle_content_merge(&mfi, opt, path,
3339                                                            is_dirty,
3340                                                            o, a, b, NULL);
3341                 }
3342         } else if (!o_valid && !a_valid && !b_valid) {
3343                 /*
3344                  * this entry was deleted altogether. a_mode == 0 means
3345                  * we had that path and want to actively remove it.
3346                  */
3347                 remove_file(opt, 1, path, !a->mode);
3348         } else
3349                 BUG("fatal merge failure, shouldn't happen.");
3350
3351         return clean_merge;
3352 }
3353
3354 static int merge_trees_internal(struct merge_options *opt,
3355                                 struct tree *head,
3356                                 struct tree *merge,
3357                                 struct tree *common,
3358                                 struct tree **result)
3359 {
3360         struct index_state *istate = opt->repo->index;
3361         int code, clean;
3362
3363         if (opt->subtree_shift) {
3364                 merge = shift_tree_object(opt->repo, head, merge, opt->subtree_shift);
3365                 common = shift_tree_object(opt->repo, head, common, opt->subtree_shift);
3366         }
3367
3368         if (oid_eq(&common->object.oid, &merge->object.oid)) {
3369                 output(opt, 0, _("Already up to date!"));
3370                 *result = head;
3371                 return 1;
3372         }
3373
3374         code = unpack_trees_start(opt, common, head, merge);
3375
3376         if (code != 0) {
3377                 if (show(opt, 4) || opt->call_depth)
3378                         err(opt, _("merging of trees %s and %s failed"),
3379                             oid_to_hex(&head->object.oid),
3380                             oid_to_hex(&merge->object.oid));
3381                 unpack_trees_finish(opt);
3382                 return -1;
3383         }
3384
3385         if (unmerged_index(istate)) {
3386                 struct string_list *entries;
3387                 struct rename_info re_info;
3388                 int i;
3389                 /*
3390                  * Only need the hashmap while processing entries, so
3391                  * initialize it here and free it when we are done running
3392                  * through the entries. Keeping it in the merge_options as
3393                  * opposed to decaring a local hashmap is for convenience
3394                  * so that we don't have to pass it to around.
3395                  */
3396                 hashmap_init(&opt->current_file_dir_set, path_hashmap_cmp, NULL, 512);
3397                 get_files_dirs(opt, head);
3398                 get_files_dirs(opt, merge);
3399
3400                 entries = get_unmerged(opt->repo->index);
3401                 clean = detect_and_process_renames(opt, common, head, merge,
3402                                                    entries, &re_info);
3403                 record_df_conflict_files(opt, entries);
3404                 if (clean < 0)
3405                         goto cleanup;
3406                 for (i = entries->nr-1; 0 <= i; i--) {
3407                         const char *path = entries->items[i].string;
3408                         struct stage_data *e = entries->items[i].util;
3409                         if (!e->processed) {
3410                                 int ret = process_entry(opt, path, e);
3411                                 if (!ret)
3412                                         clean = 0;
3413                                 else if (ret < 0) {
3414                                         clean = ret;
3415                                         goto cleanup;
3416                                 }
3417                         }
3418                 }
3419                 for (i = 0; i < entries->nr; i++) {
3420                         struct stage_data *e = entries->items[i].util;
3421                         if (!e->processed)
3422                                 BUG("unprocessed path??? %s",
3423                                     entries->items[i].string);
3424                 }
3425
3426         cleanup:
3427                 final_cleanup_renames(&re_info);
3428
3429                 string_list_clear(entries, 1);
3430                 free(entries);
3431
3432                 hashmap_free(&opt->current_file_dir_set, 1);
3433
3434                 if (clean < 0) {
3435                         unpack_trees_finish(opt);
3436                         return clean;
3437                 }
3438         }
3439         else
3440                 clean = 1;
3441
3442         unpack_trees_finish(opt);
3443
3444         if (opt->call_depth &&
3445             !(*result = write_in_core_index_as_tree(opt->repo)))
3446                 return -1;
3447
3448         return clean;
3449 }
3450
3451 static struct commit_list *reverse_commit_list(struct commit_list *list)
3452 {
3453         struct commit_list *next = NULL, *current, *backup;
3454         for (current = list; current; current = backup) {
3455                 backup = current->next;
3456                 current->next = next;
3457                 next = current;
3458         }
3459         return next;
3460 }
3461
3462 /*
3463  * Merge the commits h1 and h2, return the resulting virtual
3464  * commit object and a flag indicating the cleanness of the merge.
3465  */
3466 static int merge_recursive_internal(struct merge_options *opt,
3467                                     struct commit *h1,
3468                                     struct commit *h2,
3469                                     struct commit_list *ca,
3470                                     struct commit **result)
3471 {
3472         struct commit_list *iter;
3473         struct commit *merged_common_ancestors;
3474         struct tree *mrtree;
3475         int clean;
3476         const char *ancestor_name;
3477         struct strbuf merge_base_abbrev = STRBUF_INIT;
3478
3479         if (show(opt, 4)) {
3480                 output(opt, 4, _("Merging:"));
3481                 output_commit_title(opt, h1);
3482                 output_commit_title(opt, h2);
3483         }
3484
3485         if (!ca) {
3486                 ca = get_merge_bases(h1, h2);
3487                 ca = reverse_commit_list(ca);
3488         }
3489
3490         if (show(opt, 5)) {
3491                 unsigned cnt = commit_list_count(ca);
3492
3493                 output(opt, 5, Q_("found %u common ancestor:",
3494                                 "found %u common ancestors:", cnt), cnt);
3495                 for (iter = ca; iter; iter = iter->next)
3496                         output_commit_title(opt, iter->item);
3497         }
3498
3499         merged_common_ancestors = pop_commit(&ca);
3500         if (merged_common_ancestors == NULL) {
3501                 /* if there is no common ancestor, use an empty tree */
3502                 struct tree *tree;
3503
3504                 tree = lookup_tree(opt->repo, opt->repo->hash_algo->empty_tree);
3505                 merged_common_ancestors = make_virtual_commit(opt->repo, tree, "ancestor");
3506                 ancestor_name = "empty tree";
3507         } else if (ca) {
3508                 ancestor_name = "merged common ancestors";
3509         } else {
3510                 strbuf_add_unique_abbrev(&merge_base_abbrev,
3511                                          &merged_common_ancestors->object.oid,
3512                                          DEFAULT_ABBREV);
3513                 ancestor_name = merge_base_abbrev.buf;
3514         }
3515
3516         for (iter = ca; iter; iter = iter->next) {
3517                 const char *saved_b1, *saved_b2;
3518                 opt->call_depth++;
3519                 /*
3520                  * When the merge fails, the result contains files
3521                  * with conflict markers. The cleanness flag is
3522                  * ignored (unless indicating an error), it was never
3523                  * actually used, as result of merge_trees has always
3524                  * overwritten it: the committed "conflicts" were
3525                  * already resolved.
3526                  */
3527                 discard_index(opt->repo->index);
3528                 saved_b1 = opt->branch1;
3529                 saved_b2 = opt->branch2;
3530                 opt->branch1 = "Temporary merge branch 1";
3531                 opt->branch2 = "Temporary merge branch 2";
3532                 if (merge_recursive_internal(opt, merged_common_ancestors, iter->item,
3533                                     NULL, &merged_common_ancestors) < 0)
3534                         return -1;
3535                 opt->branch1 = saved_b1;
3536                 opt->branch2 = saved_b2;
3537                 opt->call_depth--;
3538
3539                 if (!merged_common_ancestors)
3540                         return err(opt, _("merge returned no commit"));
3541         }
3542
3543         discard_index(opt->repo->index);
3544         if (!opt->call_depth)
3545                 repo_read_index(opt->repo);
3546
3547         opt->ancestor = ancestor_name;
3548         clean = merge_trees_internal(opt,
3549                                      repo_get_commit_tree(opt->repo, h1),
3550                                      repo_get_commit_tree(opt->repo, h2),
3551                                      repo_get_commit_tree(opt->repo,
3552                                                           merged_common_ancestors),
3553                                      &mrtree);
3554         strbuf_release(&merge_base_abbrev);
3555         if (clean < 0) {
3556                 flush_output(opt);
3557                 return clean;
3558         }
3559
3560         if (opt->call_depth) {
3561                 *result = make_virtual_commit(opt->repo, mrtree, "merged tree");
3562                 commit_list_insert(h1, &(*result)->parents);
3563                 commit_list_insert(h2, &(*result)->parents->next);
3564         }
3565         flush_output(opt);
3566         if (!opt->call_depth && opt->buffer_output < 2)
3567                 strbuf_release(&opt->obuf);
3568         return clean;
3569 }
3570
3571 static int merge_start(struct merge_options *opt, struct tree *head)
3572 {
3573         struct strbuf sb = STRBUF_INIT;
3574
3575         if (repo_index_has_changes(opt->repo, head, &sb)) {
3576                 err(opt, _("Your local changes to the following files would be overwritten by merge:\n  %s"),
3577                     sb.buf);
3578                 strbuf_release(&sb);
3579                 return -1;
3580         }
3581
3582         return 0;
3583 }
3584
3585 static void merge_finalize(struct merge_options *opt)
3586 {
3587         if (show(opt, 2))
3588                 diff_warn_rename_limit("merge.renamelimit",
3589                                        opt->needed_rename_limit, 0);
3590 }
3591
3592 int merge_trees(struct merge_options *opt,
3593                 struct tree *head,
3594                 struct tree *merge,
3595                 struct tree *common)
3596 {
3597         int clean;
3598         struct tree *ignored;
3599
3600         assert(opt->ancestor != NULL);
3601
3602         if (merge_start(opt, head))
3603                 return -1;
3604         clean = merge_trees_internal(opt, head, merge, common, &ignored);
3605         merge_finalize(opt);
3606
3607         return clean;
3608 }
3609
3610 int merge_recursive(struct merge_options *opt,
3611                     struct commit *h1,
3612                     struct commit *h2,
3613                     struct commit_list *ca,
3614                     struct commit **result)
3615 {
3616         int clean;
3617
3618         assert(opt->ancestor == NULL);
3619
3620         if (merge_start(opt, repo_get_commit_tree(opt->repo, h1)))
3621                 return -1;
3622         clean = merge_recursive_internal(opt, h1, h2, ca, result);
3623         merge_finalize(opt);
3624
3625         return clean;
3626 }
3627
3628 static struct commit *get_ref(struct repository *repo, const struct object_id *oid,
3629                               const char *name)
3630 {
3631         struct object *object;
3632
3633         object = deref_tag(repo, parse_object(repo, oid),
3634                            name, strlen(name));
3635         if (!object)
3636                 return NULL;
3637         if (object->type == OBJ_TREE)
3638                 return make_virtual_commit(repo, (struct tree*)object, name);
3639         if (object->type != OBJ_COMMIT)
3640                 return NULL;
3641         if (parse_commit((struct commit *)object))
3642                 return NULL;
3643         return (struct commit *)object;
3644 }
3645
3646 int merge_recursive_generic(struct merge_options *opt,
3647                             const struct object_id *head,
3648                             const struct object_id *merge,
3649                             int num_base_list,
3650                             const struct object_id **base_list,
3651                             struct commit **result)
3652 {
3653         int clean;
3654         struct lock_file lock = LOCK_INIT;
3655         struct commit *head_commit = get_ref(opt->repo, head, opt->branch1);
3656         struct commit *next_commit = get_ref(opt->repo, merge, opt->branch2);
3657         struct commit_list *ca = NULL;
3658
3659         if (base_list) {
3660                 int i;
3661                 for (i = 0; i < num_base_list; ++i) {
3662                         struct commit *base;
3663                         if (!(base = get_ref(opt->repo, base_list[i], oid_to_hex(base_list[i]))))
3664                                 return err(opt, _("Could not parse object '%s'"),
3665                                            oid_to_hex(base_list[i]));
3666                         commit_list_insert(base, &ca);
3667                 }
3668         }
3669
3670         repo_hold_locked_index(opt->repo, &lock, LOCK_DIE_ON_ERROR);
3671         clean = merge_recursive(opt, head_commit, next_commit, ca,
3672                                 result);
3673         if (clean < 0) {
3674                 rollback_lock_file(&lock);
3675                 return clean;
3676         }
3677
3678         if (write_locked_index(opt->repo->index, &lock,
3679                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3680                 return err(opt, _("Unable to write index."));
3681
3682         return clean ? 0 : 1;
3683 }
3684
3685 static void merge_recursive_config(struct merge_options *opt)
3686 {
3687         char *value = NULL;
3688         git_config_get_int("merge.verbosity", &opt->verbosity);
3689         git_config_get_int("diff.renamelimit", &opt->diff_rename_limit);
3690         git_config_get_int("merge.renamelimit", &opt->merge_rename_limit);
3691         if (!git_config_get_string("diff.renames", &value)) {
3692                 opt->diff_detect_rename = git_config_rename("diff.renames", value);
3693                 free(value);
3694         }
3695         if (!git_config_get_string("merge.renames", &value)) {
3696                 opt->merge_detect_rename = git_config_rename("merge.renames", value);
3697                 free(value);
3698         }
3699         if (!git_config_get_string("merge.directoryrenames", &value)) {
3700                 int boolval = git_parse_maybe_bool(value);
3701                 if (0 <= boolval) {
3702                         opt->detect_directory_renames = boolval ?
3703                                 MERGE_DIRECTORY_RENAMES_TRUE :
3704                                 MERGE_DIRECTORY_RENAMES_NONE;
3705                 } else if (!strcasecmp(value, "conflict")) {
3706                         opt->detect_directory_renames =
3707                                 MERGE_DIRECTORY_RENAMES_CONFLICT;
3708                 } /* avoid erroring on values from future versions of git */
3709                 free(value);
3710         }
3711         git_config(git_xmerge_config, NULL);
3712 }
3713
3714 void init_merge_options(struct merge_options *opt,
3715                         struct repository *repo)
3716 {
3717         const char *merge_verbosity;
3718         memset(opt, 0, sizeof(struct merge_options));
3719         opt->repo = repo;
3720         opt->verbosity = 2;
3721         opt->buffer_output = 1;
3722         opt->diff_rename_limit = -1;
3723         opt->merge_rename_limit = -1;
3724         opt->renormalize = 0;
3725         opt->diff_detect_rename = -1;
3726         opt->merge_detect_rename = -1;
3727         opt->detect_directory_renames = MERGE_DIRECTORY_RENAMES_CONFLICT;
3728         merge_recursive_config(opt);
3729         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3730         if (merge_verbosity)
3731                 opt->verbosity = strtol(merge_verbosity, NULL, 10);
3732         if (opt->verbosity >= 5)
3733                 opt->buffer_output = 0;
3734         strbuf_init(&opt->obuf, 0);
3735         string_list_init(&opt->df_conflict_file_set, 1);
3736 }
3737
3738 int parse_merge_opt(struct merge_options *opt, const char *s)
3739 {
3740         const char *arg;
3741
3742         if (!s || !*s)
3743                 return -1;
3744         if (!strcmp(s, "ours"))
3745                 opt->recursive_variant = MERGE_RECURSIVE_OURS;
3746         else if (!strcmp(s, "theirs"))
3747                 opt->recursive_variant = MERGE_RECURSIVE_THEIRS;
3748         else if (!strcmp(s, "subtree"))
3749                 opt->subtree_shift = "";
3750         else if (skip_prefix(s, "subtree=", &arg))
3751                 opt->subtree_shift = arg;
3752         else if (!strcmp(s, "patience"))
3753                 opt->xdl_opts = DIFF_WITH_ALG(opt, PATIENCE_DIFF);
3754         else if (!strcmp(s, "histogram"))
3755                 opt->xdl_opts = DIFF_WITH_ALG(opt, HISTOGRAM_DIFF);
3756         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3757                 long value = parse_algorithm_value(arg);
3758                 if (value < 0)
3759                         return -1;
3760                 /* clear out previous settings */
3761                 DIFF_XDL_CLR(opt, NEED_MINIMAL);
3762                 opt->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3763                 opt->xdl_opts |= value;
3764         }
3765         else if (!strcmp(s, "ignore-space-change"))
3766                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_CHANGE);
3767         else if (!strcmp(s, "ignore-all-space"))
3768                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE);
3769         else if (!strcmp(s, "ignore-space-at-eol"))
3770                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_AT_EOL);
3771         else if (!strcmp(s, "ignore-cr-at-eol"))
3772                 DIFF_XDL_SET(opt, IGNORE_CR_AT_EOL);
3773         else if (!strcmp(s, "renormalize"))
3774                 opt->renormalize = 1;
3775         else if (!strcmp(s, "no-renormalize"))
3776                 opt->renormalize = 0;
3777         else if (!strcmp(s, "no-renames"))
3778                 opt->merge_detect_rename = 0;
3779         else if (!strcmp(s, "find-renames")) {
3780                 opt->merge_detect_rename = 1;
3781                 opt->rename_score = 0;
3782         }
3783         else if (skip_prefix(s, "find-renames=", &arg) ||
3784                  skip_prefix(s, "rename-threshold=", &arg)) {
3785                 if ((opt->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3786                         return -1;
3787                 opt->merge_detect_rename = 1;
3788         }
3789         /*
3790          * Please update $__git_merge_strategy_options in
3791          * git-completion.bash when you add new options
3792          */
3793         else
3794                 return -1;
3795         return 0;
3796 }