merge-recursive: check for file level conflicts then get new name
[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 "commit.h"
12 #include "blob.h"
13 #include "builtin.h"
14 #include "tree-walk.h"
15 #include "diff.h"
16 #include "diffcore.h"
17 #include "tag.h"
18 #include "unpack-trees.h"
19 #include "string-list.h"
20 #include "xdiff-interface.h"
21 #include "ll-merge.h"
22 #include "attr.h"
23 #include "merge-recursive.h"
24 #include "dir.h"
25 #include "submodule.h"
26
27 struct path_hashmap_entry {
28         struct hashmap_entry e;
29         char path[FLEX_ARRAY];
30 };
31
32 static int path_hashmap_cmp(const void *cmp_data,
33                             const void *entry,
34                             const void *entry_or_key,
35                             const void *keydata)
36 {
37         const struct path_hashmap_entry *a = entry;
38         const struct path_hashmap_entry *b = entry_or_key;
39         const char *key = keydata;
40
41         if (ignore_case)
42                 return strcasecmp(a->path, key ? key : b->path);
43         else
44                 return strcmp(a->path, key ? key : b->path);
45 }
46
47 static unsigned int path_hash(const char *path)
48 {
49         return ignore_case ? strihash(path) : strhash(path);
50 }
51
52 static struct dir_rename_entry *dir_rename_find_entry(struct hashmap *hashmap,
53                                                       char *dir)
54 {
55         struct dir_rename_entry key;
56
57         if (dir == NULL)
58                 return NULL;
59         hashmap_entry_init(&key, strhash(dir));
60         key.dir = dir;
61         return hashmap_get(hashmap, &key, NULL);
62 }
63
64 static int dir_rename_cmp(const void *unused_cmp_data,
65                           const void *entry,
66                           const void *entry_or_key,
67                           const void *unused_keydata)
68 {
69         const struct dir_rename_entry *e1 = entry;
70         const struct dir_rename_entry *e2 = entry_or_key;
71
72         return strcmp(e1->dir, e2->dir);
73 }
74
75 static void dir_rename_init(struct hashmap *map)
76 {
77         hashmap_init(map, dir_rename_cmp, NULL, 0);
78 }
79
80 static void dir_rename_entry_init(struct dir_rename_entry *entry,
81                                   char *directory)
82 {
83         hashmap_entry_init(entry, strhash(directory));
84         entry->dir = directory;
85         entry->non_unique_new_dir = 0;
86         strbuf_init(&entry->new_dir, 0);
87         string_list_init(&entry->possible_new_dirs, 0);
88 }
89
90 static struct collision_entry *collision_find_entry(struct hashmap *hashmap,
91                                                     char *target_file)
92 {
93         struct collision_entry key;
94
95         hashmap_entry_init(&key, strhash(target_file));
96         key.target_file = target_file;
97         return hashmap_get(hashmap, &key, NULL);
98 }
99
100 static int collision_cmp(void *unused_cmp_data,
101                          const struct collision_entry *e1,
102                          const struct collision_entry *e2,
103                          const void *unused_keydata)
104 {
105         return strcmp(e1->target_file, e2->target_file);
106 }
107
108 static void collision_init(struct hashmap *map)
109 {
110         hashmap_init(map, (hashmap_cmp_fn) collision_cmp, NULL, 0);
111 }
112
113 static void flush_output(struct merge_options *o)
114 {
115         if (o->buffer_output < 2 && o->obuf.len) {
116                 fputs(o->obuf.buf, stdout);
117                 strbuf_reset(&o->obuf);
118         }
119 }
120
121 static int err(struct merge_options *o, const char *err, ...)
122 {
123         va_list params;
124
125         if (o->buffer_output < 2)
126                 flush_output(o);
127         else {
128                 strbuf_complete(&o->obuf, '\n');
129                 strbuf_addstr(&o->obuf, "error: ");
130         }
131         va_start(params, err);
132         strbuf_vaddf(&o->obuf, err, params);
133         va_end(params);
134         if (o->buffer_output > 1)
135                 strbuf_addch(&o->obuf, '\n');
136         else {
137                 error("%s", o->obuf.buf);
138                 strbuf_reset(&o->obuf);
139         }
140
141         return -1;
142 }
143
144 static struct tree *shift_tree_object(struct tree *one, struct tree *two,
145                                       const char *subtree_shift)
146 {
147         struct object_id shifted;
148
149         if (!*subtree_shift) {
150                 shift_tree(&one->object.oid, &two->object.oid, &shifted, 0);
151         } else {
152                 shift_tree_by(&one->object.oid, &two->object.oid, &shifted,
153                               subtree_shift);
154         }
155         if (!oidcmp(&two->object.oid, &shifted))
156                 return two;
157         return lookup_tree(&shifted);
158 }
159
160 static struct commit *make_virtual_commit(struct tree *tree, const char *comment)
161 {
162         struct commit *commit = alloc_commit_node();
163
164         set_merge_remote_desc(commit, comment, (struct object *)commit);
165         commit->tree = tree;
166         commit->object.parsed = 1;
167         return commit;
168 }
169
170 /*
171  * Since we use get_tree_entry(), which does not put the read object into
172  * the object pool, we cannot rely on a == b.
173  */
174 static int oid_eq(const struct object_id *a, const struct object_id *b)
175 {
176         if (!a && !b)
177                 return 2;
178         return a && b && oidcmp(a, b) == 0;
179 }
180
181 enum rename_type {
182         RENAME_NORMAL = 0,
183         RENAME_DELETE,
184         RENAME_ONE_FILE_TO_ONE,
185         RENAME_ONE_FILE_TO_TWO,
186         RENAME_TWO_FILES_TO_ONE
187 };
188
189 struct rename_conflict_info {
190         enum rename_type rename_type;
191         struct diff_filepair *pair1;
192         struct diff_filepair *pair2;
193         const char *branch1;
194         const char *branch2;
195         struct stage_data *dst_entry1;
196         struct stage_data *dst_entry2;
197         struct diff_filespec ren1_other;
198         struct diff_filespec ren2_other;
199 };
200
201 /*
202  * Since we want to write the index eventually, we cannot reuse the index
203  * for these (temporary) data.
204  */
205 struct stage_data {
206         struct {
207                 unsigned mode;
208                 struct object_id oid;
209         } stages[4];
210         struct rename_conflict_info *rename_conflict_info;
211         unsigned processed:1;
212 };
213
214 static inline void setup_rename_conflict_info(enum rename_type rename_type,
215                                               struct diff_filepair *pair1,
216                                               struct diff_filepair *pair2,
217                                               const char *branch1,
218                                               const char *branch2,
219                                               struct stage_data *dst_entry1,
220                                               struct stage_data *dst_entry2,
221                                               struct merge_options *o,
222                                               struct stage_data *src_entry1,
223                                               struct stage_data *src_entry2)
224 {
225         struct rename_conflict_info *ci = xcalloc(1, sizeof(struct rename_conflict_info));
226         ci->rename_type = rename_type;
227         ci->pair1 = pair1;
228         ci->branch1 = branch1;
229         ci->branch2 = branch2;
230
231         ci->dst_entry1 = dst_entry1;
232         dst_entry1->rename_conflict_info = ci;
233         dst_entry1->processed = 0;
234
235         assert(!pair2 == !dst_entry2);
236         if (dst_entry2) {
237                 ci->dst_entry2 = dst_entry2;
238                 ci->pair2 = pair2;
239                 dst_entry2->rename_conflict_info = ci;
240         }
241
242         if (rename_type == RENAME_TWO_FILES_TO_ONE) {
243                 /*
244                  * For each rename, there could have been
245                  * modifications on the side of history where that
246                  * file was not renamed.
247                  */
248                 int ostage1 = o->branch1 == branch1 ? 3 : 2;
249                 int ostage2 = ostage1 ^ 1;
250
251                 ci->ren1_other.path = pair1->one->path;
252                 oidcpy(&ci->ren1_other.oid, &src_entry1->stages[ostage1].oid);
253                 ci->ren1_other.mode = src_entry1->stages[ostage1].mode;
254
255                 ci->ren2_other.path = pair2->one->path;
256                 oidcpy(&ci->ren2_other.oid, &src_entry2->stages[ostage2].oid);
257                 ci->ren2_other.mode = src_entry2->stages[ostage2].mode;
258         }
259 }
260
261 static int show(struct merge_options *o, int v)
262 {
263         return (!o->call_depth && o->verbosity >= v) || o->verbosity >= 5;
264 }
265
266 __attribute__((format (printf, 3, 4)))
267 static void output(struct merge_options *o, int v, const char *fmt, ...)
268 {
269         va_list ap;
270
271         if (!show(o, v))
272                 return;
273
274         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
275
276         va_start(ap, fmt);
277         strbuf_vaddf(&o->obuf, fmt, ap);
278         va_end(ap);
279
280         strbuf_addch(&o->obuf, '\n');
281         if (!o->buffer_output)
282                 flush_output(o);
283 }
284
285 static void output_commit_title(struct merge_options *o, struct commit *commit)
286 {
287         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
288         if (commit->util)
289                 strbuf_addf(&o->obuf, "virtual %s\n",
290                         merge_remote_util(commit)->name);
291         else {
292                 strbuf_add_unique_abbrev(&o->obuf, &commit->object.oid,
293                                          DEFAULT_ABBREV);
294                 strbuf_addch(&o->obuf, ' ');
295                 if (parse_commit(commit) != 0)
296                         strbuf_addstr(&o->obuf, _("(bad commit)\n"));
297                 else {
298                         const char *title;
299                         const char *msg = get_commit_buffer(commit, NULL);
300                         int len = find_commit_subject(msg, &title);
301                         if (len)
302                                 strbuf_addf(&o->obuf, "%.*s\n", len, title);
303                         unuse_commit_buffer(commit, msg);
304                 }
305         }
306         flush_output(o);
307 }
308
309 static int add_cacheinfo(struct merge_options *o,
310                 unsigned int mode, const struct object_id *oid,
311                 const char *path, int stage, int refresh, int options)
312 {
313         struct cache_entry *ce;
314         int ret;
315
316         ce = make_cache_entry(mode, oid ? oid->hash : null_sha1, path, stage, 0);
317         if (!ce)
318                 return err(o, _("addinfo_cache failed for path '%s'"), path);
319
320         ret = add_cache_entry(ce, options);
321         if (refresh) {
322                 struct cache_entry *nce;
323
324                 nce = refresh_cache_entry(ce, CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
325                 if (!nce)
326                         return err(o, _("addinfo_cache failed for path '%s'"), path);
327                 if (nce != ce)
328                         ret = add_cache_entry(nce, options);
329         }
330         return ret;
331 }
332
333 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
334 {
335         parse_tree(tree);
336         init_tree_desc(desc, tree->buffer, tree->size);
337 }
338
339 static int git_merge_trees(int index_only,
340                            struct tree *common,
341                            struct tree *head,
342                            struct tree *merge)
343 {
344         int rc;
345         struct tree_desc t[3];
346         struct unpack_trees_options opts;
347
348         memset(&opts, 0, sizeof(opts));
349         if (index_only)
350                 opts.index_only = 1;
351         else
352                 opts.update = 1;
353         opts.merge = 1;
354         opts.head_idx = 2;
355         opts.fn = threeway_merge;
356         opts.src_index = &the_index;
357         opts.dst_index = &the_index;
358         setup_unpack_trees_porcelain(&opts, "merge");
359
360         init_tree_desc_from_tree(t+0, common);
361         init_tree_desc_from_tree(t+1, head);
362         init_tree_desc_from_tree(t+2, merge);
363
364         rc = unpack_trees(3, t, &opts);
365         cache_tree_free(&active_cache_tree);
366         return rc;
367 }
368
369 struct tree *write_tree_from_memory(struct merge_options *o)
370 {
371         struct tree *result = NULL;
372
373         if (unmerged_cache()) {
374                 int i;
375                 fprintf(stderr, "BUG: There are unmerged index entries:\n");
376                 for (i = 0; i < active_nr; i++) {
377                         const struct cache_entry *ce = active_cache[i];
378                         if (ce_stage(ce))
379                                 fprintf(stderr, "BUG: %d %.*s\n", ce_stage(ce),
380                                         (int)ce_namelen(ce), ce->name);
381                 }
382                 die("BUG: unmerged index entries in merge-recursive.c");
383         }
384
385         if (!active_cache_tree)
386                 active_cache_tree = cache_tree();
387
388         if (!cache_tree_fully_valid(active_cache_tree) &&
389             cache_tree_update(&the_index, 0) < 0) {
390                 err(o, _("error building trees"));
391                 return NULL;
392         }
393
394         result = lookup_tree(&active_cache_tree->oid);
395
396         return result;
397 }
398
399 static int save_files_dirs(const struct object_id *oid,
400                 struct strbuf *base, const char *path,
401                 unsigned int mode, int stage, void *context)
402 {
403         struct path_hashmap_entry *entry;
404         int baselen = base->len;
405         struct merge_options *o = context;
406
407         strbuf_addstr(base, path);
408
409         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
410         hashmap_entry_init(entry, path_hash(entry->path));
411         hashmap_add(&o->current_file_dir_set, entry);
412
413         strbuf_setlen(base, baselen);
414         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
415 }
416
417 static void get_files_dirs(struct merge_options *o, struct tree *tree)
418 {
419         struct pathspec match_all;
420         memset(&match_all, 0, sizeof(match_all));
421         read_tree_recursive(tree, "", 0, 0, &match_all, save_files_dirs, o);
422 }
423
424 /*
425  * Returns an index_entry instance which doesn't have to correspond to
426  * a real cache entry in Git's index.
427  */
428 static struct stage_data *insert_stage_data(const char *path,
429                 struct tree *o, struct tree *a, struct tree *b,
430                 struct string_list *entries)
431 {
432         struct string_list_item *item;
433         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
434         get_tree_entry(&o->object.oid, path,
435                         &e->stages[1].oid, &e->stages[1].mode);
436         get_tree_entry(&a->object.oid, path,
437                         &e->stages[2].oid, &e->stages[2].mode);
438         get_tree_entry(&b->object.oid, path,
439                         &e->stages[3].oid, &e->stages[3].mode);
440         item = string_list_insert(entries, path);
441         item->util = e;
442         return e;
443 }
444
445 /*
446  * Create a dictionary mapping file names to stage_data objects. The
447  * dictionary contains one entry for every path with a non-zero stage entry.
448  */
449 static struct string_list *get_unmerged(void)
450 {
451         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
452         int i;
453
454         unmerged->strdup_strings = 1;
455
456         for (i = 0; i < active_nr; i++) {
457                 struct string_list_item *item;
458                 struct stage_data *e;
459                 const struct cache_entry *ce = active_cache[i];
460                 if (!ce_stage(ce))
461                         continue;
462
463                 item = string_list_lookup(unmerged, ce->name);
464                 if (!item) {
465                         item = string_list_insert(unmerged, ce->name);
466                         item->util = xcalloc(1, sizeof(struct stage_data));
467                 }
468                 e = item->util;
469                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
470                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
471         }
472
473         return unmerged;
474 }
475
476 static int string_list_df_name_compare(const char *one, const char *two)
477 {
478         int onelen = strlen(one);
479         int twolen = strlen(two);
480         /*
481          * Here we only care that entries for D/F conflicts are
482          * adjacent, in particular with the file of the D/F conflict
483          * appearing before files below the corresponding directory.
484          * The order of the rest of the list is irrelevant for us.
485          *
486          * To achieve this, we sort with df_name_compare and provide
487          * the mode S_IFDIR so that D/F conflicts will sort correctly.
488          * We use the mode S_IFDIR for everything else for simplicity,
489          * since in other cases any changes in their order due to
490          * sorting cause no problems for us.
491          */
492         int cmp = df_name_compare(one, onelen, S_IFDIR,
493                                   two, twolen, S_IFDIR);
494         /*
495          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
496          * that 'foo' comes before 'foo/bar'.
497          */
498         if (cmp)
499                 return cmp;
500         return onelen - twolen;
501 }
502
503 static void record_df_conflict_files(struct merge_options *o,
504                                      struct string_list *entries)
505 {
506         /* If there is a D/F conflict and the file for such a conflict
507          * currently exist in the working tree, we want to allow it to be
508          * removed to make room for the corresponding directory if needed.
509          * The files underneath the directories of such D/F conflicts will
510          * be processed before the corresponding file involved in the D/F
511          * conflict.  If the D/F directory ends up being removed by the
512          * merge, then we won't have to touch the D/F file.  If the D/F
513          * directory needs to be written to the working copy, then the D/F
514          * file will simply be removed (in make_room_for_path()) to make
515          * room for the necessary paths.  Note that if both the directory
516          * and the file need to be present, then the D/F file will be
517          * reinstated with a new unique name at the time it is processed.
518          */
519         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
520         const char *last_file = NULL;
521         int last_len = 0;
522         int i;
523
524         /*
525          * If we're merging merge-bases, we don't want to bother with
526          * any working directory changes.
527          */
528         if (o->call_depth)
529                 return;
530
531         /* Ensure D/F conflicts are adjacent in the entries list. */
532         for (i = 0; i < entries->nr; i++) {
533                 struct string_list_item *next = &entries->items[i];
534                 string_list_append(&df_sorted_entries, next->string)->util =
535                                    next->util;
536         }
537         df_sorted_entries.cmp = string_list_df_name_compare;
538         string_list_sort(&df_sorted_entries);
539
540         string_list_clear(&o->df_conflict_file_set, 1);
541         for (i = 0; i < df_sorted_entries.nr; i++) {
542                 const char *path = df_sorted_entries.items[i].string;
543                 int len = strlen(path);
544                 struct stage_data *e = df_sorted_entries.items[i].util;
545
546                 /*
547                  * Check if last_file & path correspond to a D/F conflict;
548                  * i.e. whether path is last_file+'/'+<something>.
549                  * If so, record that it's okay to remove last_file to make
550                  * room for path and friends if needed.
551                  */
552                 if (last_file &&
553                     len > last_len &&
554                     memcmp(path, last_file, last_len) == 0 &&
555                     path[last_len] == '/') {
556                         string_list_insert(&o->df_conflict_file_set, last_file);
557                 }
558
559                 /*
560                  * Determine whether path could exist as a file in the
561                  * working directory as a possible D/F conflict.  This
562                  * will only occur when it exists in stage 2 as a
563                  * file.
564                  */
565                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
566                         last_file = path;
567                         last_len = len;
568                 } else {
569                         last_file = NULL;
570                 }
571         }
572         string_list_clear(&df_sorted_entries, 0);
573 }
574
575 struct rename {
576         struct diff_filepair *pair;
577         /*
578          * Purpose of src_entry and dst_entry:
579          *
580          * If 'before' is renamed to 'after' then src_entry will contain
581          * the versions of 'before' from the merge_base, HEAD, and MERGE in
582          * stages 1, 2, and 3; dst_entry will contain the respective
583          * versions of 'after' in corresponding locations.  Thus, we have a
584          * total of six modes and oids, though some will be null.  (Stage 0
585          * is ignored; we're interested in handling conflicts.)
586          *
587          * Since we don't turn on break-rewrites by default, neither
588          * src_entry nor dst_entry can have all three of their stages have
589          * non-null oids, meaning at most four of the six will be non-null.
590          * Also, since this is a rename, both src_entry and dst_entry will
591          * have at least one non-null oid, meaning at least two will be
592          * non-null.  Of the six oids, a typical rename will have three be
593          * non-null.  Only two implies a rename/delete, and four implies a
594          * rename/add.
595          */
596         struct stage_data *src_entry;
597         struct stage_data *dst_entry;
598         unsigned processed:1;
599 };
600
601 static int update_stages(struct merge_options *opt, const char *path,
602                          const struct diff_filespec *o,
603                          const struct diff_filespec *a,
604                          const struct diff_filespec *b)
605 {
606
607         /*
608          * NOTE: It is usually a bad idea to call update_stages on a path
609          * before calling update_file on that same path, since it can
610          * sometimes lead to spurious "refusing to lose untracked file..."
611          * messages from update_file (via make_room_for path via
612          * would_lose_untracked).  Instead, reverse the order of the calls
613          * (executing update_file first and then update_stages).
614          */
615         int clear = 1;
616         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
617         if (clear)
618                 if (remove_file_from_cache(path))
619                         return -1;
620         if (o)
621                 if (add_cacheinfo(opt, o->mode, &o->oid, path, 1, 0, options))
622                         return -1;
623         if (a)
624                 if (add_cacheinfo(opt, a->mode, &a->oid, path, 2, 0, options))
625                         return -1;
626         if (b)
627                 if (add_cacheinfo(opt, b->mode, &b->oid, path, 3, 0, options))
628                         return -1;
629         return 0;
630 }
631
632 static void update_entry(struct stage_data *entry,
633                          struct diff_filespec *o,
634                          struct diff_filespec *a,
635                          struct diff_filespec *b)
636 {
637         entry->processed = 0;
638         entry->stages[1].mode = o->mode;
639         entry->stages[2].mode = a->mode;
640         entry->stages[3].mode = b->mode;
641         oidcpy(&entry->stages[1].oid, &o->oid);
642         oidcpy(&entry->stages[2].oid, &a->oid);
643         oidcpy(&entry->stages[3].oid, &b->oid);
644 }
645
646 static int remove_file(struct merge_options *o, int clean,
647                        const char *path, int no_wd)
648 {
649         int update_cache = o->call_depth || clean;
650         int update_working_directory = !o->call_depth && !no_wd;
651
652         if (update_cache) {
653                 if (remove_file_from_cache(path))
654                         return -1;
655         }
656         if (update_working_directory) {
657                 if (ignore_case) {
658                         struct cache_entry *ce;
659                         ce = cache_file_exists(path, strlen(path), ignore_case);
660                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
661                                 return 0;
662                 }
663                 if (remove_path(path))
664                         return -1;
665         }
666         return 0;
667 }
668
669 /* add a string to a strbuf, but converting "/" to "_" */
670 static void add_flattened_path(struct strbuf *out, const char *s)
671 {
672         size_t i = out->len;
673         strbuf_addstr(out, s);
674         for (; i < out->len; i++)
675                 if (out->buf[i] == '/')
676                         out->buf[i] = '_';
677 }
678
679 static char *unique_path(struct merge_options *o, const char *path, const char *branch)
680 {
681         struct path_hashmap_entry *entry;
682         struct strbuf newpath = STRBUF_INIT;
683         int suffix = 0;
684         size_t base_len;
685
686         strbuf_addf(&newpath, "%s~", path);
687         add_flattened_path(&newpath, branch);
688
689         base_len = newpath.len;
690         while (hashmap_get_from_hash(&o->current_file_dir_set,
691                                      path_hash(newpath.buf), newpath.buf) ||
692                (!o->call_depth && file_exists(newpath.buf))) {
693                 strbuf_setlen(&newpath, base_len);
694                 strbuf_addf(&newpath, "_%d", suffix++);
695         }
696
697         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
698         hashmap_entry_init(entry, path_hash(entry->path));
699         hashmap_add(&o->current_file_dir_set, entry);
700         return strbuf_detach(&newpath, NULL);
701 }
702
703 /**
704  * Check whether a directory in the index is in the way of an incoming
705  * file.  Return 1 if so.  If check_working_copy is non-zero, also
706  * check the working directory.  If empty_ok is non-zero, also return
707  * 0 in the case where the working-tree dir exists but is empty.
708  */
709 static int dir_in_way(const char *path, int check_working_copy, int empty_ok)
710 {
711         int pos;
712         struct strbuf dirpath = STRBUF_INIT;
713         struct stat st;
714
715         strbuf_addstr(&dirpath, path);
716         strbuf_addch(&dirpath, '/');
717
718         pos = cache_name_pos(dirpath.buf, dirpath.len);
719
720         if (pos < 0)
721                 pos = -1 - pos;
722         if (pos < active_nr &&
723             !strncmp(dirpath.buf, active_cache[pos]->name, dirpath.len)) {
724                 strbuf_release(&dirpath);
725                 return 1;
726         }
727
728         strbuf_release(&dirpath);
729         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
730                 !(empty_ok && is_empty_dir(path));
731 }
732
733 static int was_tracked(const char *path)
734 {
735         int pos = cache_name_pos(path, strlen(path));
736
737         if (0 <= pos)
738                 /* we have been tracking this path */
739                 return 1;
740
741         /*
742          * Look for an unmerged entry for the path,
743          * specifically stage #2, which would indicate
744          * that "our" side before the merge started
745          * had the path tracked (and resulted in a conflict).
746          */
747         for (pos = -1 - pos;
748              pos < active_nr && !strcmp(path, active_cache[pos]->name);
749              pos++)
750                 if (ce_stage(active_cache[pos]) == 2)
751                         return 1;
752         return 0;
753 }
754
755 static int would_lose_untracked(const char *path)
756 {
757         return !was_tracked(path) && file_exists(path);
758 }
759
760 static int make_room_for_path(struct merge_options *o, const char *path)
761 {
762         int status, i;
763         const char *msg = _("failed to create path '%s'%s");
764
765         /* Unlink any D/F conflict files that are in the way */
766         for (i = 0; i < o->df_conflict_file_set.nr; i++) {
767                 const char *df_path = o->df_conflict_file_set.items[i].string;
768                 size_t pathlen = strlen(path);
769                 size_t df_pathlen = strlen(df_path);
770                 if (df_pathlen < pathlen &&
771                     path[df_pathlen] == '/' &&
772                     strncmp(path, df_path, df_pathlen) == 0) {
773                         output(o, 3,
774                                _("Removing %s to make room for subdirectory\n"),
775                                df_path);
776                         unlink(df_path);
777                         unsorted_string_list_delete_item(&o->df_conflict_file_set,
778                                                          i, 0);
779                         break;
780                 }
781         }
782
783         /* Make sure leading directories are created */
784         status = safe_create_leading_directories_const(path);
785         if (status) {
786                 if (status == SCLD_EXISTS)
787                         /* something else exists */
788                         return err(o, msg, path, _(": perhaps a D/F conflict?"));
789                 return err(o, msg, path, "");
790         }
791
792         /*
793          * Do not unlink a file in the work tree if we are not
794          * tracking it.
795          */
796         if (would_lose_untracked(path))
797                 return err(o, _("refusing to lose untracked file at '%s'"),
798                              path);
799
800         /* Successful unlink is good.. */
801         if (!unlink(path))
802                 return 0;
803         /* .. and so is no existing file */
804         if (errno == ENOENT)
805                 return 0;
806         /* .. but not some other error (who really cares what?) */
807         return err(o, msg, path, _(": perhaps a D/F conflict?"));
808 }
809
810 static int update_file_flags(struct merge_options *o,
811                              const struct object_id *oid,
812                              unsigned mode,
813                              const char *path,
814                              int update_cache,
815                              int update_wd)
816 {
817         int ret = 0;
818
819         if (o->call_depth)
820                 update_wd = 0;
821
822         if (update_wd) {
823                 enum object_type type;
824                 void *buf;
825                 unsigned long size;
826
827                 if (S_ISGITLINK(mode)) {
828                         /*
829                          * We may later decide to recursively descend into
830                          * the submodule directory and update its index
831                          * and/or work tree, but we do not do that now.
832                          */
833                         update_wd = 0;
834                         goto update_index;
835                 }
836
837                 buf = read_object_file(oid, &type, &size);
838                 if (!buf)
839                         return err(o, _("cannot read object %s '%s'"), oid_to_hex(oid), path);
840                 if (type != OBJ_BLOB) {
841                         ret = err(o, _("blob expected for %s '%s'"), oid_to_hex(oid), path);
842                         goto free_buf;
843                 }
844                 if (S_ISREG(mode)) {
845                         struct strbuf strbuf = STRBUF_INIT;
846                         if (convert_to_working_tree(path, buf, size, &strbuf)) {
847                                 free(buf);
848                                 size = strbuf.len;
849                                 buf = strbuf_detach(&strbuf, NULL);
850                         }
851                 }
852
853                 if (make_room_for_path(o, path) < 0) {
854                         update_wd = 0;
855                         goto free_buf;
856                 }
857                 if (S_ISREG(mode) || (!has_symlinks && S_ISLNK(mode))) {
858                         int fd;
859                         if (mode & 0100)
860                                 mode = 0777;
861                         else
862                                 mode = 0666;
863                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
864                         if (fd < 0) {
865                                 ret = err(o, _("failed to open '%s': %s"),
866                                           path, strerror(errno));
867                                 goto free_buf;
868                         }
869                         write_in_full(fd, buf, size);
870                         close(fd);
871                 } else if (S_ISLNK(mode)) {
872                         char *lnk = xmemdupz(buf, size);
873                         safe_create_leading_directories_const(path);
874                         unlink(path);
875                         if (symlink(lnk, path))
876                                 ret = err(o, _("failed to symlink '%s': %s"),
877                                         path, strerror(errno));
878                         free(lnk);
879                 } else
880                         ret = err(o,
881                                   _("do not know what to do with %06o %s '%s'"),
882                                   mode, oid_to_hex(oid), path);
883  free_buf:
884                 free(buf);
885         }
886  update_index:
887         if (!ret && update_cache)
888                 add_cacheinfo(o, mode, oid, path, 0, update_wd, ADD_CACHE_OK_TO_ADD);
889         return ret;
890 }
891
892 static int update_file(struct merge_options *o,
893                        int clean,
894                        const struct object_id *oid,
895                        unsigned mode,
896                        const char *path)
897 {
898         return update_file_flags(o, oid, mode, path, o->call_depth || clean, !o->call_depth);
899 }
900
901 /* Low level file merging, update and removal */
902
903 struct merge_file_info {
904         struct object_id oid;
905         unsigned mode;
906         unsigned clean:1,
907                  merge:1;
908 };
909
910 static int merge_3way(struct merge_options *o,
911                       mmbuffer_t *result_buf,
912                       const struct diff_filespec *one,
913                       const struct diff_filespec *a,
914                       const struct diff_filespec *b,
915                       const char *branch1,
916                       const char *branch2)
917 {
918         mmfile_t orig, src1, src2;
919         struct ll_merge_options ll_opts = {0};
920         char *base_name, *name1, *name2;
921         int merge_status;
922
923         ll_opts.renormalize = o->renormalize;
924         ll_opts.xdl_opts = o->xdl_opts;
925
926         if (o->call_depth) {
927                 ll_opts.virtual_ancestor = 1;
928                 ll_opts.variant = 0;
929         } else {
930                 switch (o->recursive_variant) {
931                 case MERGE_RECURSIVE_OURS:
932                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
933                         break;
934                 case MERGE_RECURSIVE_THEIRS:
935                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
936                         break;
937                 default:
938                         ll_opts.variant = 0;
939                         break;
940                 }
941         }
942
943         if (strcmp(a->path, b->path) ||
944             (o->ancestor != NULL && strcmp(a->path, one->path) != 0)) {
945                 base_name = o->ancestor == NULL ? NULL :
946                         mkpathdup("%s:%s", o->ancestor, one->path);
947                 name1 = mkpathdup("%s:%s", branch1, a->path);
948                 name2 = mkpathdup("%s:%s", branch2, b->path);
949         } else {
950                 base_name = o->ancestor == NULL ? NULL :
951                         mkpathdup("%s", o->ancestor);
952                 name1 = mkpathdup("%s", branch1);
953                 name2 = mkpathdup("%s", branch2);
954         }
955
956         read_mmblob(&orig, &one->oid);
957         read_mmblob(&src1, &a->oid);
958         read_mmblob(&src2, &b->oid);
959
960         merge_status = ll_merge(result_buf, a->path, &orig, base_name,
961                                 &src1, name1, &src2, name2, &ll_opts);
962
963         free(base_name);
964         free(name1);
965         free(name2);
966         free(orig.ptr);
967         free(src1.ptr);
968         free(src2.ptr);
969         return merge_status;
970 }
971
972 static int merge_file_1(struct merge_options *o,
973                                            const struct diff_filespec *one,
974                                            const struct diff_filespec *a,
975                                            const struct diff_filespec *b,
976                                            const char *branch1,
977                                            const char *branch2,
978                                            struct merge_file_info *result)
979 {
980         result->merge = 0;
981         result->clean = 1;
982
983         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
984                 result->clean = 0;
985                 if (S_ISREG(a->mode)) {
986                         result->mode = a->mode;
987                         oidcpy(&result->oid, &a->oid);
988                 } else {
989                         result->mode = b->mode;
990                         oidcpy(&result->oid, &b->oid);
991                 }
992         } else {
993                 if (!oid_eq(&a->oid, &one->oid) && !oid_eq(&b->oid, &one->oid))
994                         result->merge = 1;
995
996                 /*
997                  * Merge modes
998                  */
999                 if (a->mode == b->mode || a->mode == one->mode)
1000                         result->mode = b->mode;
1001                 else {
1002                         result->mode = a->mode;
1003                         if (b->mode != one->mode) {
1004                                 result->clean = 0;
1005                                 result->merge = 1;
1006                         }
1007                 }
1008
1009                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &one->oid))
1010                         oidcpy(&result->oid, &b->oid);
1011                 else if (oid_eq(&b->oid, &one->oid))
1012                         oidcpy(&result->oid, &a->oid);
1013                 else if (S_ISREG(a->mode)) {
1014                         mmbuffer_t result_buf;
1015                         int ret = 0, merge_status;
1016
1017                         merge_status = merge_3way(o, &result_buf, one, a, b,
1018                                                   branch1, branch2);
1019
1020                         if ((merge_status < 0) || !result_buf.ptr)
1021                                 ret = err(o, _("Failed to execute internal merge"));
1022
1023                         if (!ret &&
1024                             write_object_file(result_buf.ptr, result_buf.size,
1025                                               blob_type, &result->oid))
1026                                 ret = err(o, _("Unable to add %s to database"),
1027                                           a->path);
1028
1029                         free(result_buf.ptr);
1030                         if (ret)
1031                                 return ret;
1032                         result->clean = (merge_status == 0);
1033                 } else if (S_ISGITLINK(a->mode)) {
1034                         result->clean = merge_submodule(&result->oid,
1035                                                        one->path,
1036                                                        &one->oid,
1037                                                        &a->oid,
1038                                                        &b->oid,
1039                                                        !o->call_depth);
1040                 } else if (S_ISLNK(a->mode)) {
1041                         switch (o->recursive_variant) {
1042                         case MERGE_RECURSIVE_NORMAL:
1043                                 oidcpy(&result->oid, &a->oid);
1044                                 if (!oid_eq(&a->oid, &b->oid))
1045                                         result->clean = 0;
1046                                 break;
1047                         case MERGE_RECURSIVE_OURS:
1048                                 oidcpy(&result->oid, &a->oid);
1049                                 break;
1050                         case MERGE_RECURSIVE_THEIRS:
1051                                 oidcpy(&result->oid, &b->oid);
1052                                 break;
1053                         }
1054                 } else
1055                         die("BUG: unsupported object type in the tree");
1056         }
1057
1058         return 0;
1059 }
1060
1061 static int merge_file_special_markers(struct merge_options *o,
1062                            const struct diff_filespec *one,
1063                            const struct diff_filespec *a,
1064                            const struct diff_filespec *b,
1065                            const char *branch1,
1066                            const char *filename1,
1067                            const char *branch2,
1068                            const char *filename2,
1069                            struct merge_file_info *mfi)
1070 {
1071         char *side1 = NULL;
1072         char *side2 = NULL;
1073         int ret;
1074
1075         if (filename1)
1076                 side1 = xstrfmt("%s:%s", branch1, filename1);
1077         if (filename2)
1078                 side2 = xstrfmt("%s:%s", branch2, filename2);
1079
1080         ret = merge_file_1(o, one, a, b,
1081                            side1 ? side1 : branch1,
1082                            side2 ? side2 : branch2, mfi);
1083         free(side1);
1084         free(side2);
1085         return ret;
1086 }
1087
1088 static int merge_file_one(struct merge_options *o,
1089                                          const char *path,
1090                                          const struct object_id *o_oid, int o_mode,
1091                                          const struct object_id *a_oid, int a_mode,
1092                                          const struct object_id *b_oid, int b_mode,
1093                                          const char *branch1,
1094                                          const char *branch2,
1095                                          struct merge_file_info *mfi)
1096 {
1097         struct diff_filespec one, a, b;
1098
1099         one.path = a.path = b.path = (char *)path;
1100         oidcpy(&one.oid, o_oid);
1101         one.mode = o_mode;
1102         oidcpy(&a.oid, a_oid);
1103         a.mode = a_mode;
1104         oidcpy(&b.oid, b_oid);
1105         b.mode = b_mode;
1106         return merge_file_1(o, &one, &a, &b, branch1, branch2, mfi);
1107 }
1108
1109 static int handle_change_delete(struct merge_options *o,
1110                                  const char *path, const char *old_path,
1111                                  const struct object_id *o_oid, int o_mode,
1112                                  const struct object_id *changed_oid,
1113                                  int changed_mode,
1114                                  const char *change_branch,
1115                                  const char *delete_branch,
1116                                  const char *change, const char *change_past)
1117 {
1118         char *alt_path = NULL;
1119         const char *update_path = path;
1120         int ret = 0;
1121
1122         if (dir_in_way(path, !o->call_depth, 0)) {
1123                 update_path = alt_path = unique_path(o, path, change_branch);
1124         }
1125
1126         if (o->call_depth) {
1127                 /*
1128                  * We cannot arbitrarily accept either a_sha or b_sha as
1129                  * correct; since there is no true "middle point" between
1130                  * them, simply reuse the base version for virtual merge base.
1131                  */
1132                 ret = remove_file_from_cache(path);
1133                 if (!ret)
1134                         ret = update_file(o, 0, o_oid, o_mode, update_path);
1135         } else {
1136                 if (!alt_path) {
1137                         if (!old_path) {
1138                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1139                                        "and %s in %s. Version %s of %s left in tree."),
1140                                        change, path, delete_branch, change_past,
1141                                        change_branch, change_branch, path);
1142                         } else {
1143                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1144                                        "and %s to %s in %s. Version %s of %s left in tree."),
1145                                        change, old_path, delete_branch, change_past, path,
1146                                        change_branch, change_branch, path);
1147                         }
1148                 } else {
1149                         if (!old_path) {
1150                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1151                                        "and %s in %s. Version %s of %s left in tree at %s."),
1152                                        change, path, delete_branch, change_past,
1153                                        change_branch, change_branch, path, alt_path);
1154                         } else {
1155                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1156                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1157                                        change, old_path, delete_branch, change_past, path,
1158                                        change_branch, change_branch, path, alt_path);
1159                         }
1160                 }
1161                 /*
1162                  * No need to call update_file() on path when change_branch ==
1163                  * o->branch1 && !alt_path, since that would needlessly touch
1164                  * path.  We could call update_file_flags() with update_cache=0
1165                  * and update_wd=0, but that's a no-op.
1166                  */
1167                 if (change_branch != o->branch1 || alt_path)
1168                         ret = update_file(o, 0, changed_oid, changed_mode, update_path);
1169         }
1170         free(alt_path);
1171
1172         return ret;
1173 }
1174
1175 static int conflict_rename_delete(struct merge_options *o,
1176                                    struct diff_filepair *pair,
1177                                    const char *rename_branch,
1178                                    const char *delete_branch)
1179 {
1180         const struct diff_filespec *orig = pair->one;
1181         const struct diff_filespec *dest = pair->two;
1182
1183         if (handle_change_delete(o,
1184                                  o->call_depth ? orig->path : dest->path,
1185                                  o->call_depth ? NULL : orig->path,
1186                                  &orig->oid, orig->mode,
1187                                  &dest->oid, dest->mode,
1188                                  rename_branch, delete_branch,
1189                                  _("rename"), _("renamed")))
1190                 return -1;
1191
1192         if (o->call_depth)
1193                 return remove_file_from_cache(dest->path);
1194         else
1195                 return update_stages(o, dest->path, NULL,
1196                                      rename_branch == o->branch1 ? dest : NULL,
1197                                      rename_branch == o->branch1 ? NULL : dest);
1198 }
1199
1200 static struct diff_filespec *filespec_from_entry(struct diff_filespec *target,
1201                                                  struct stage_data *entry,
1202                                                  int stage)
1203 {
1204         struct object_id *oid = &entry->stages[stage].oid;
1205         unsigned mode = entry->stages[stage].mode;
1206         if (mode == 0 || is_null_oid(oid))
1207                 return NULL;
1208         oidcpy(&target->oid, oid);
1209         target->mode = mode;
1210         return target;
1211 }
1212
1213 static int handle_file(struct merge_options *o,
1214                         struct diff_filespec *rename,
1215                         int stage,
1216                         struct rename_conflict_info *ci)
1217 {
1218         char *dst_name = rename->path;
1219         struct stage_data *dst_entry;
1220         const char *cur_branch, *other_branch;
1221         struct diff_filespec other;
1222         struct diff_filespec *add;
1223         int ret;
1224
1225         if (stage == 2) {
1226                 dst_entry = ci->dst_entry1;
1227                 cur_branch = ci->branch1;
1228                 other_branch = ci->branch2;
1229         } else {
1230                 dst_entry = ci->dst_entry2;
1231                 cur_branch = ci->branch2;
1232                 other_branch = ci->branch1;
1233         }
1234
1235         add = filespec_from_entry(&other, dst_entry, stage ^ 1);
1236         if (add) {
1237                 char *add_name = unique_path(o, rename->path, other_branch);
1238                 if (update_file(o, 0, &add->oid, add->mode, add_name))
1239                         return -1;
1240
1241                 remove_file(o, 0, rename->path, 0);
1242                 dst_name = unique_path(o, rename->path, cur_branch);
1243         } else {
1244                 if (dir_in_way(rename->path, !o->call_depth, 0)) {
1245                         dst_name = unique_path(o, rename->path, cur_branch);
1246                         output(o, 1, _("%s is a directory in %s adding as %s instead"),
1247                                rename->path, other_branch, dst_name);
1248                 }
1249         }
1250         if ((ret = update_file(o, 0, &rename->oid, rename->mode, dst_name)))
1251                 ; /* fall through, do allow dst_name to be released */
1252         else if (stage == 2)
1253                 ret = update_stages(o, rename->path, NULL, rename, add);
1254         else
1255                 ret = update_stages(o, rename->path, NULL, add, rename);
1256
1257         if (dst_name != rename->path)
1258                 free(dst_name);
1259
1260         return ret;
1261 }
1262
1263 static int conflict_rename_rename_1to2(struct merge_options *o,
1264                                         struct rename_conflict_info *ci)
1265 {
1266         /* One file was renamed in both branches, but to different names. */
1267         struct diff_filespec *one = ci->pair1->one;
1268         struct diff_filespec *a = ci->pair1->two;
1269         struct diff_filespec *b = ci->pair2->two;
1270
1271         output(o, 1, _("CONFLICT (rename/rename): "
1272                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1273                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1274                one->path, a->path, ci->branch1,
1275                one->path, b->path, ci->branch2,
1276                o->call_depth ? _(" (left unresolved)") : "");
1277         if (o->call_depth) {
1278                 struct merge_file_info mfi;
1279                 struct diff_filespec other;
1280                 struct diff_filespec *add;
1281                 if (merge_file_one(o, one->path,
1282                                  &one->oid, one->mode,
1283                                  &a->oid, a->mode,
1284                                  &b->oid, b->mode,
1285                                  ci->branch1, ci->branch2, &mfi))
1286                         return -1;
1287
1288                 /*
1289                  * FIXME: For rename/add-source conflicts (if we could detect
1290                  * such), this is wrong.  We should instead find a unique
1291                  * pathname and then either rename the add-source file to that
1292                  * unique path, or use that unique path instead of src here.
1293                  */
1294                 if (update_file(o, 0, &mfi.oid, mfi.mode, one->path))
1295                         return -1;
1296
1297                 /*
1298                  * Above, we put the merged content at the merge-base's
1299                  * path.  Now we usually need to delete both a->path and
1300                  * b->path.  However, the rename on each side of the merge
1301                  * could also be involved in a rename/add conflict.  In
1302                  * such cases, we should keep the added file around,
1303                  * resolving the conflict at that path in its favor.
1304                  */
1305                 add = filespec_from_entry(&other, ci->dst_entry1, 2 ^ 1);
1306                 if (add) {
1307                         if (update_file(o, 0, &add->oid, add->mode, a->path))
1308                                 return -1;
1309                 }
1310                 else
1311                         remove_file_from_cache(a->path);
1312                 add = filespec_from_entry(&other, ci->dst_entry2, 3 ^ 1);
1313                 if (add) {
1314                         if (update_file(o, 0, &add->oid, add->mode, b->path))
1315                                 return -1;
1316                 }
1317                 else
1318                         remove_file_from_cache(b->path);
1319         } else if (handle_file(o, a, 2, ci) || handle_file(o, b, 3, ci))
1320                 return -1;
1321
1322         return 0;
1323 }
1324
1325 static int conflict_rename_rename_2to1(struct merge_options *o,
1326                                         struct rename_conflict_info *ci)
1327 {
1328         /* Two files, a & b, were renamed to the same thing, c. */
1329         struct diff_filespec *a = ci->pair1->one;
1330         struct diff_filespec *b = ci->pair2->one;
1331         struct diff_filespec *c1 = ci->pair1->two;
1332         struct diff_filespec *c2 = ci->pair2->two;
1333         char *path = c1->path; /* == c2->path */
1334         struct merge_file_info mfi_c1;
1335         struct merge_file_info mfi_c2;
1336         int ret;
1337
1338         output(o, 1, _("CONFLICT (rename/rename): "
1339                "Rename %s->%s in %s. "
1340                "Rename %s->%s in %s"),
1341                a->path, c1->path, ci->branch1,
1342                b->path, c2->path, ci->branch2);
1343
1344         remove_file(o, 1, a->path, o->call_depth || would_lose_untracked(a->path));
1345         remove_file(o, 1, b->path, o->call_depth || would_lose_untracked(b->path));
1346
1347         if (merge_file_special_markers(o, a, c1, &ci->ren1_other,
1348                                        o->branch1, c1->path,
1349                                        o->branch2, ci->ren1_other.path, &mfi_c1) ||
1350             merge_file_special_markers(o, b, &ci->ren2_other, c2,
1351                                        o->branch1, ci->ren2_other.path,
1352                                        o->branch2, c2->path, &mfi_c2))
1353                 return -1;
1354
1355         if (o->call_depth) {
1356                 /*
1357                  * If mfi_c1.clean && mfi_c2.clean, then it might make
1358                  * sense to do a two-way merge of those results.  But, I
1359                  * think in all cases, it makes sense to have the virtual
1360                  * merge base just undo the renames; they can be detected
1361                  * again later for the non-recursive merge.
1362                  */
1363                 remove_file(o, 0, path, 0);
1364                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, a->path);
1365                 if (!ret)
1366                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1367                                           b->path);
1368         } else {
1369                 char *new_path1 = unique_path(o, path, ci->branch1);
1370                 char *new_path2 = unique_path(o, path, ci->branch2);
1371                 output(o, 1, _("Renaming %s to %s and %s to %s instead"),
1372                        a->path, new_path1, b->path, new_path2);
1373                 remove_file(o, 0, path, 0);
1374                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, new_path1);
1375                 if (!ret)
1376                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1377                                           new_path2);
1378                 free(new_path2);
1379                 free(new_path1);
1380         }
1381
1382         return ret;
1383 }
1384
1385 /*
1386  * Get the diff_filepairs changed between o_tree and tree.
1387  */
1388 static struct diff_queue_struct *get_diffpairs(struct merge_options *o,
1389                                                struct tree *o_tree,
1390                                                struct tree *tree)
1391 {
1392         struct diff_queue_struct *ret;
1393         struct diff_options opts;
1394
1395         diff_setup(&opts);
1396         opts.flags.recursive = 1;
1397         opts.flags.rename_empty = 0;
1398         opts.detect_rename = DIFF_DETECT_RENAME;
1399         opts.rename_limit = o->merge_rename_limit >= 0 ? o->merge_rename_limit :
1400                             o->diff_rename_limit >= 0 ? o->diff_rename_limit :
1401                             1000;
1402         opts.rename_score = o->rename_score;
1403         opts.show_rename_progress = o->show_rename_progress;
1404         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1405         diff_setup_done(&opts);
1406         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1407         diffcore_std(&opts);
1408         if (opts.needed_rename_limit > o->needed_rename_limit)
1409                 o->needed_rename_limit = opts.needed_rename_limit;
1410
1411         ret = xmalloc(sizeof(*ret));
1412         *ret = diff_queued_diff;
1413
1414         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1415         diff_queued_diff.nr = 0;
1416         diff_queued_diff.queue = NULL;
1417         diff_flush(&opts);
1418         return ret;
1419 }
1420
1421 static int tree_has_path(struct tree *tree, const char *path)
1422 {
1423         struct object_id hashy;
1424         unsigned int mode_o;
1425
1426         return !get_tree_entry(&tree->object.oid, path,
1427                                &hashy, &mode_o);
1428 }
1429
1430 /*
1431  * Return a new string that replaces the beginning portion (which matches
1432  * entry->dir), with entry->new_dir.  In perl-speak:
1433  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1434  * NOTE:
1435  *   Caller must ensure that old_path starts with entry->dir + '/'.
1436  */
1437 static char *apply_dir_rename(struct dir_rename_entry *entry,
1438                               const char *old_path)
1439 {
1440         struct strbuf new_path = STRBUF_INIT;
1441         int oldlen, newlen;
1442
1443         if (entry->non_unique_new_dir)
1444                 return NULL;
1445
1446         oldlen = strlen(entry->dir);
1447         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1448         strbuf_grow(&new_path, newlen);
1449         strbuf_addbuf(&new_path, &entry->new_dir);
1450         strbuf_addstr(&new_path, &old_path[oldlen]);
1451
1452         return strbuf_detach(&new_path, NULL);
1453 }
1454
1455 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1456                                     char **old_dir, char **new_dir)
1457 {
1458         char *end_of_old, *end_of_new;
1459         int old_len, new_len;
1460
1461         *old_dir = NULL;
1462         *new_dir = NULL;
1463
1464         /*
1465          * For
1466          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1467          * the "e/foo.c" part is the same, we just want to know that
1468          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1469          * so, for this example, this function returns "a/b/c/d" in
1470          * *old_dir and "a/b/some/thing/else" in *new_dir.
1471          *
1472          * Also, if the basename of the file changed, we don't care.  We
1473          * want to know which portion of the directory, if any, changed.
1474          */
1475         end_of_old = strrchr(old_path, '/');
1476         end_of_new = strrchr(new_path, '/');
1477
1478         if (end_of_old == NULL || end_of_new == NULL)
1479                 return;
1480         while (*--end_of_new == *--end_of_old &&
1481                end_of_old != old_path &&
1482                end_of_new != new_path)
1483                 ; /* Do nothing; all in the while loop */
1484         /*
1485          * We've found the first non-matching character in the directory
1486          * paths.  That means the current directory we were comparing
1487          * represents the rename.  Move end_of_old and end_of_new back
1488          * to the full directory name.
1489          */
1490         if (*end_of_old == '/')
1491                 end_of_old++;
1492         if (*end_of_old != '/')
1493                 end_of_new++;
1494         end_of_old = strchr(end_of_old, '/');
1495         end_of_new = strchr(end_of_new, '/');
1496
1497         /*
1498          * It may have been the case that old_path and new_path were the same
1499          * directory all along.  Don't claim a rename if they're the same.
1500          */
1501         old_len = end_of_old - old_path;
1502         new_len = end_of_new - new_path;
1503
1504         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1505                 *old_dir = xstrndup(old_path, old_len);
1506                 *new_dir = xstrndup(new_path, new_len);
1507         }
1508 }
1509
1510 static void remove_hashmap_entries(struct hashmap *dir_renames,
1511                                    struct string_list *items_to_remove)
1512 {
1513         int i;
1514         struct dir_rename_entry *entry;
1515
1516         for (i = 0; i < items_to_remove->nr; i++) {
1517                 entry = items_to_remove->items[i].util;
1518                 hashmap_remove(dir_renames, entry, NULL);
1519         }
1520         string_list_clear(items_to_remove, 0);
1521 }
1522
1523 /*
1524  * See if there is a directory rename for path, and if there are any file
1525  * level conflicts for the renamed location.  If there is a rename and
1526  * there are no conflicts, return the new name.  Otherwise, return NULL.
1527  */
1528 static char *handle_path_level_conflicts(struct merge_options *o,
1529                                          const char *path,
1530                                          struct dir_rename_entry *entry,
1531                                          struct hashmap *collisions,
1532                                          struct tree *tree)
1533 {
1534         char *new_path = NULL;
1535         struct collision_entry *collision_ent;
1536         int clean = 1;
1537         struct strbuf collision_paths = STRBUF_INIT;
1538
1539         /*
1540          * entry has the mapping of old directory name to new directory name
1541          * that we want to apply to path.
1542          */
1543         new_path = apply_dir_rename(entry, path);
1544
1545         if (!new_path) {
1546                 /* This should only happen when entry->non_unique_new_dir set */
1547                 if (!entry->non_unique_new_dir)
1548                         BUG("entry->non_unqiue_dir not set and !new_path");
1549                 output(o, 1, _("CONFLICT (directory rename split): "
1550                                "Unclear where to place %s because directory "
1551                                "%s was renamed to multiple other directories, "
1552                                "with no destination getting a majority of the "
1553                                "files."),
1554                        path, entry->dir);
1555                 clean = 0;
1556                 return NULL;
1557         }
1558
1559         /*
1560          * The caller needs to have ensured that it has pre-populated
1561          * collisions with all paths that map to new_path.  Do a quick check
1562          * to ensure that's the case.
1563          */
1564         collision_ent = collision_find_entry(collisions, new_path);
1565         if (collision_ent == NULL)
1566                 BUG("collision_ent is NULL");
1567
1568         /*
1569          * Check for one-sided add/add/.../add conflicts, i.e.
1570          * where implicit renames from the other side doing
1571          * directory rename(s) can affect this side of history
1572          * to put multiple paths into the same location.  Warn
1573          * and bail on directory renames for such paths.
1574          */
1575         if (collision_ent->reported_already) {
1576                 clean = 0;
1577         } else if (tree_has_path(tree, new_path)) {
1578                 collision_ent->reported_already = 1;
1579                 strbuf_add_separated_string_list(&collision_paths, ", ",
1580                                                  &collision_ent->source_files);
1581                 output(o, 1, _("CONFLICT (implicit dir rename): Existing "
1582                                "file/dir at %s in the way of implicit "
1583                                "directory rename(s) putting the following "
1584                                "path(s) there: %s."),
1585                        new_path, collision_paths.buf);
1586                 clean = 0;
1587         } else if (collision_ent->source_files.nr > 1) {
1588                 collision_ent->reported_already = 1;
1589                 strbuf_add_separated_string_list(&collision_paths, ", ",
1590                                                  &collision_ent->source_files);
1591                 output(o, 1, _("CONFLICT (implicit dir rename): Cannot map "
1592                                "more than one path to %s; implicit directory "
1593                                "renames tried to put these paths there: %s"),
1594                        new_path, collision_paths.buf);
1595                 clean = 0;
1596         }
1597
1598         /* Free memory we no longer need */
1599         strbuf_release(&collision_paths);
1600         if (!clean && new_path) {
1601                 free(new_path);
1602                 return NULL;
1603         }
1604
1605         return new_path;
1606 }
1607
1608 /*
1609  * There are a couple things we want to do at the directory level:
1610  *   1. Check for both sides renaming to the same thing, in order to avoid
1611  *      implicit renaming of files that should be left in place.  (See
1612  *      testcase 6b in t6043 for details.)
1613  *   2. Prune directory renames if there are still files left in the
1614  *      the original directory.  These represent a partial directory rename,
1615  *      i.e. a rename where only some of the files within the directory
1616  *      were renamed elsewhere.  (Technically, this could be done earlier
1617  *      in get_directory_renames(), except that would prevent us from
1618  *      doing the previous check and thus failing testcase 6b.)
1619  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
1620  *      In the future, we could potentially record this info as well and
1621  *      omit reporting rename/rename(1to2) conflicts for each path within
1622  *      the affected directories, thus cleaning up the merge output.
1623  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
1624  *         directory level, because merging directories is fine.  If it
1625  *         causes conflicts for files within those merged directories, then
1626  *         that should be detected at the individual path level.
1627  */
1628 static void handle_directory_level_conflicts(struct merge_options *o,
1629                                              struct hashmap *dir_re_head,
1630                                              struct tree *head,
1631                                              struct hashmap *dir_re_merge,
1632                                              struct tree *merge)
1633 {
1634         struct hashmap_iter iter;
1635         struct dir_rename_entry *head_ent;
1636         struct dir_rename_entry *merge_ent;
1637
1638         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
1639         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
1640
1641         hashmap_iter_init(dir_re_head, &iter);
1642         while ((head_ent = hashmap_iter_next(&iter))) {
1643                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
1644                 if (merge_ent &&
1645                     !head_ent->non_unique_new_dir &&
1646                     !merge_ent->non_unique_new_dir &&
1647                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
1648                         /* 1. Renamed identically; remove it from both sides */
1649                         string_list_append(&remove_from_head,
1650                                            head_ent->dir)->util = head_ent;
1651                         strbuf_release(&head_ent->new_dir);
1652                         string_list_append(&remove_from_merge,
1653                                            merge_ent->dir)->util = merge_ent;
1654                         strbuf_release(&merge_ent->new_dir);
1655                 } else if (tree_has_path(head, head_ent->dir)) {
1656                         /* 2. This wasn't a directory rename after all */
1657                         string_list_append(&remove_from_head,
1658                                            head_ent->dir)->util = head_ent;
1659                         strbuf_release(&head_ent->new_dir);
1660                 }
1661         }
1662
1663         remove_hashmap_entries(dir_re_head, &remove_from_head);
1664         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
1665
1666         hashmap_iter_init(dir_re_merge, &iter);
1667         while ((merge_ent = hashmap_iter_next(&iter))) {
1668                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
1669                 if (tree_has_path(merge, merge_ent->dir)) {
1670                         /* 2. This wasn't a directory rename after all */
1671                         string_list_append(&remove_from_merge,
1672                                            merge_ent->dir)->util = merge_ent;
1673                 } else if (head_ent &&
1674                            !head_ent->non_unique_new_dir &&
1675                            !merge_ent->non_unique_new_dir) {
1676                         /* 3. rename/rename(1to2) */
1677                         /*
1678                          * We can assume it's not rename/rename(1to1) because
1679                          * that was case (1), already checked above.  So we
1680                          * know that head_ent->new_dir and merge_ent->new_dir
1681                          * are different strings.
1682                          */
1683                         output(o, 1, _("CONFLICT (rename/rename): "
1684                                        "Rename directory %s->%s in %s. "
1685                                        "Rename directory %s->%s in %s"),
1686                                head_ent->dir, head_ent->new_dir.buf, o->branch1,
1687                                head_ent->dir, merge_ent->new_dir.buf, o->branch2);
1688                         string_list_append(&remove_from_head,
1689                                            head_ent->dir)->util = head_ent;
1690                         strbuf_release(&head_ent->new_dir);
1691                         string_list_append(&remove_from_merge,
1692                                            merge_ent->dir)->util = merge_ent;
1693                         strbuf_release(&merge_ent->new_dir);
1694                 }
1695         }
1696
1697         remove_hashmap_entries(dir_re_head, &remove_from_head);
1698         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
1699 }
1700
1701 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs,
1702                                              struct tree *tree)
1703 {
1704         struct hashmap *dir_renames;
1705         struct hashmap_iter iter;
1706         struct dir_rename_entry *entry;
1707         int i;
1708
1709         /*
1710          * Typically, we think of a directory rename as all files from a
1711          * certain directory being moved to a target directory.  However,
1712          * what if someone first moved two files from the original
1713          * directory in one commit, and then renamed the directory
1714          * somewhere else in a later commit?  At merge time, we just know
1715          * that files from the original directory went to two different
1716          * places, and that the bulk of them ended up in the same place.
1717          * We want each directory rename to represent where the bulk of the
1718          * files from that directory end up; this function exists to find
1719          * where the bulk of the files went.
1720          *
1721          * The first loop below simply iterates through the list of file
1722          * renames, finding out how often each directory rename pair
1723          * possibility occurs.
1724          */
1725         dir_renames = xmalloc(sizeof(*dir_renames));
1726         dir_rename_init(dir_renames);
1727         for (i = 0; i < pairs->nr; ++i) {
1728                 struct string_list_item *item;
1729                 int *count;
1730                 struct diff_filepair *pair = pairs->queue[i];
1731                 char *old_dir, *new_dir;
1732
1733                 /* File not part of directory rename if it wasn't renamed */
1734                 if (pair->status != 'R')
1735                         continue;
1736
1737                 get_renamed_dir_portion(pair->one->path, pair->two->path,
1738                                         &old_dir,        &new_dir);
1739                 if (!old_dir)
1740                         /* Directory didn't change at all; ignore this one. */
1741                         continue;
1742
1743                 entry = dir_rename_find_entry(dir_renames, old_dir);
1744                 if (!entry) {
1745                         entry = xmalloc(sizeof(*entry));
1746                         dir_rename_entry_init(entry, old_dir);
1747                         hashmap_put(dir_renames, entry);
1748                 } else {
1749                         free(old_dir);
1750                 }
1751                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
1752                 if (!item) {
1753                         item = string_list_insert(&entry->possible_new_dirs,
1754                                                   new_dir);
1755                         item->util = xcalloc(1, sizeof(int));
1756                 } else {
1757                         free(new_dir);
1758                 }
1759                 count = item->util;
1760                 *count += 1;
1761         }
1762
1763         /*
1764          * For each directory with files moved out of it, we find out which
1765          * target directory received the most files so we can declare it to
1766          * be the "winning" target location for the directory rename.  This
1767          * winner gets recorded in new_dir.  If there is no winner
1768          * (multiple target directories received the same number of files),
1769          * we set non_unique_new_dir.  Once we've determined the winner (or
1770          * that there is no winner), we no longer need possible_new_dirs.
1771          */
1772         hashmap_iter_init(dir_renames, &iter);
1773         while ((entry = hashmap_iter_next(&iter))) {
1774                 int max = 0;
1775                 int bad_max = 0;
1776                 char *best = NULL;
1777
1778                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
1779                         int *count = entry->possible_new_dirs.items[i].util;
1780
1781                         if (*count == max)
1782                                 bad_max = max;
1783                         else if (*count > max) {
1784                                 max = *count;
1785                                 best = entry->possible_new_dirs.items[i].string;
1786                         }
1787                 }
1788                 if (bad_max == max)
1789                         entry->non_unique_new_dir = 1;
1790                 else {
1791                         assert(entry->new_dir.len == 0);
1792                         strbuf_addstr(&entry->new_dir, best);
1793                 }
1794                 /*
1795                  * The relevant directory sub-portion of the original full
1796                  * filepaths were xstrndup'ed before inserting into
1797                  * possible_new_dirs, and instead of manually iterating the
1798                  * list and free'ing each, just lie and tell
1799                  * possible_new_dirs that it did the strdup'ing so that it
1800                  * will free them for us.
1801                  */
1802                 entry->possible_new_dirs.strdup_strings = 1;
1803                 string_list_clear(&entry->possible_new_dirs, 1);
1804         }
1805
1806         return dir_renames;
1807 }
1808
1809 static struct dir_rename_entry *check_dir_renamed(const char *path,
1810                                                   struct hashmap *dir_renames)
1811 {
1812         char temp[PATH_MAX];
1813         char *end;
1814         struct dir_rename_entry *entry;
1815
1816         strcpy(temp, path);
1817         while ((end = strrchr(temp, '/'))) {
1818                 *end = '\0';
1819                 entry = dir_rename_find_entry(dir_renames, temp);
1820                 if (entry)
1821                         return entry;
1822         }
1823         return NULL;
1824 }
1825
1826 static void compute_collisions(struct hashmap *collisions,
1827                                struct hashmap *dir_renames,
1828                                struct diff_queue_struct *pairs)
1829 {
1830         int i;
1831
1832         /*
1833          * Multiple files can be mapped to the same path due to directory
1834          * renames done by the other side of history.  Since that other
1835          * side of history could have merged multiple directories into one,
1836          * if our side of history added the same file basename to each of
1837          * those directories, then all N of them would get implicitly
1838          * renamed by the directory rename detection into the same path,
1839          * and we'd get an add/add/.../add conflict, and all those adds
1840          * from *this* side of history.  This is not representable in the
1841          * index, and users aren't going to easily be able to make sense of
1842          * it.  So we need to provide a good warning about what's
1843          * happening, and fall back to no-directory-rename detection
1844          * behavior for those paths.
1845          *
1846          * See testcases 9e and all of section 5 from t6043 for examples.
1847          */
1848         collision_init(collisions);
1849
1850         for (i = 0; i < pairs->nr; ++i) {
1851                 struct dir_rename_entry *dir_rename_ent;
1852                 struct collision_entry *collision_ent;
1853                 char *new_path;
1854                 struct diff_filepair *pair = pairs->queue[i];
1855
1856                 if (pair->status == 'D')
1857                         continue;
1858                 dir_rename_ent = check_dir_renamed(pair->two->path,
1859                                                    dir_renames);
1860                 if (!dir_rename_ent)
1861                         continue;
1862
1863                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
1864                 if (!new_path)
1865                         /*
1866                          * dir_rename_ent->non_unique_new_path is true, which
1867                          * means there is no directory rename for us to use,
1868                          * which means it won't cause us any additional
1869                          * collisions.
1870                          */
1871                         continue;
1872                 collision_ent = collision_find_entry(collisions, new_path);
1873                 if (!collision_ent) {
1874                         collision_ent = xcalloc(1,
1875                                                 sizeof(struct collision_entry));
1876                         hashmap_entry_init(collision_ent, strhash(new_path));
1877                         hashmap_put(collisions, collision_ent);
1878                         collision_ent->target_file = new_path;
1879                 } else {
1880                         free(new_path);
1881                 }
1882                 string_list_insert(&collision_ent->source_files,
1883                                    pair->two->path);
1884         }
1885 }
1886
1887 static char *check_for_directory_rename(struct merge_options *o,
1888                                         const char *path,
1889                                         struct tree *tree,
1890                                         struct hashmap *dir_renames,
1891                                         struct hashmap *dir_rename_exclusions,
1892                                         struct hashmap *collisions,
1893                                         int *clean_merge)
1894 {
1895         char *new_path = NULL;
1896         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
1897         struct dir_rename_entry *oentry = NULL;
1898
1899         if (!entry)
1900                 return new_path;
1901
1902         /*
1903          * This next part is a little weird.  We do not want to do an
1904          * implicit rename into a directory we renamed on our side, because
1905          * that will result in a spurious rename/rename(1to2) conflict.  An
1906          * example:
1907          *   Base commit: dumbdir/afile, otherdir/bfile
1908          *   Side 1:      smrtdir/afile, otherdir/bfile
1909          *   Side 2:      dumbdir/afile, dumbdir/bfile
1910          * Here, while working on Side 1, we could notice that otherdir was
1911          * renamed/merged to dumbdir, and change the diff_filepair for
1912          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
1913          * 2 will notice the rename from dumbdir to smrtdir, and do the
1914          * transitive rename to move it from dumbdir/bfile to
1915          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
1916          * smrtdir, a rename/rename(1to2) conflict.  We really just want
1917          * the file to end up in smrtdir.  And the way to achieve that is
1918          * to not let Side1 do the rename to dumbdir, since we know that is
1919          * the source of one of our directory renames.
1920          *
1921          * That's why oentry and dir_rename_exclusions is here.
1922          *
1923          * As it turns out, this also prevents N-way transient rename
1924          * confusion; See testcases 9c and 9d of t6043.
1925          */
1926         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
1927         if (oentry) {
1928                 output(o, 1, _("WARNING: Avoiding applying %s -> %s rename "
1929                                "to %s, because %s itself was renamed."),
1930                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
1931         } else {
1932                 new_path = handle_path_level_conflicts(o, path, entry,
1933                                                        collisions, tree);
1934                 *clean_merge &= (new_path != NULL);
1935         }
1936
1937         return new_path;
1938 }
1939
1940 /*
1941  * Get information of all renames which occurred in 'pairs', making use of
1942  * any implicit directory renames inferred from the other side of history.
1943  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
1944  * to be able to associate the correct cache entries with the rename
1945  * information; tree is always equal to either a_tree or b_tree.
1946  */
1947 static struct string_list *get_renames(struct merge_options *o,
1948                                        struct diff_queue_struct *pairs,
1949                                        struct hashmap *dir_renames,
1950                                        struct hashmap *dir_rename_exclusions,
1951                                        struct tree *tree,
1952                                        struct tree *o_tree,
1953                                        struct tree *a_tree,
1954                                        struct tree *b_tree,
1955                                        struct string_list *entries,
1956                                        int *clean_merge)
1957 {
1958         int i;
1959         struct hashmap collisions;
1960         struct hashmap_iter iter;
1961         struct collision_entry *e;
1962         struct string_list *renames;
1963
1964         compute_collisions(&collisions, dir_renames, pairs);
1965         renames = xcalloc(1, sizeof(struct string_list));
1966
1967         for (i = 0; i < pairs->nr; ++i) {
1968                 struct string_list_item *item;
1969                 struct rename *re;
1970                 struct diff_filepair *pair = pairs->queue[i];
1971                 char *new_path; /* non-NULL only with directory renames */
1972
1973                 if (pair->status == 'D') {
1974                         diff_free_filepair(pair);
1975                         continue;
1976                 }
1977                 new_path = check_for_directory_rename(o, pair->two->path, tree,
1978                                                       dir_renames,
1979                                                       dir_rename_exclusions,
1980                                                       &collisions,
1981                                                       clean_merge);
1982                 if (pair->status != 'R' && !new_path) {
1983                         diff_free_filepair(pair);
1984                         continue;
1985                 }
1986
1987                 re = xmalloc(sizeof(*re));
1988                 re->processed = 0;
1989                 re->pair = pair;
1990                 item = string_list_lookup(entries, re->pair->one->path);
1991                 if (!item)
1992                         re->src_entry = insert_stage_data(re->pair->one->path,
1993                                         o_tree, a_tree, b_tree, entries);
1994                 else
1995                         re->src_entry = item->util;
1996
1997                 item = string_list_lookup(entries, re->pair->two->path);
1998                 if (!item)
1999                         re->dst_entry = insert_stage_data(re->pair->two->path,
2000                                         o_tree, a_tree, b_tree, entries);
2001                 else
2002                         re->dst_entry = item->util;
2003                 item = string_list_insert(renames, pair->one->path);
2004                 item->util = re;
2005         }
2006
2007         hashmap_iter_init(&collisions, &iter);
2008         while ((e = hashmap_iter_next(&iter))) {
2009                 free(e->target_file);
2010                 string_list_clear(&e->source_files, 0);
2011         }
2012         hashmap_free(&collisions, 1);
2013         return renames;
2014 }
2015
2016 static int process_renames(struct merge_options *o,
2017                            struct string_list *a_renames,
2018                            struct string_list *b_renames)
2019 {
2020         int clean_merge = 1, i, j;
2021         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2022         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2023         const struct rename *sre;
2024
2025         for (i = 0; i < a_renames->nr; i++) {
2026                 sre = a_renames->items[i].util;
2027                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2028                         = (void *)sre;
2029         }
2030         for (i = 0; i < b_renames->nr; i++) {
2031                 sre = b_renames->items[i].util;
2032                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2033                         = (void *)sre;
2034         }
2035
2036         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2037                 struct string_list *renames1, *renames2Dst;
2038                 struct rename *ren1 = NULL, *ren2 = NULL;
2039                 const char *branch1, *branch2;
2040                 const char *ren1_src, *ren1_dst;
2041                 struct string_list_item *lookup;
2042
2043                 if (i >= a_renames->nr) {
2044                         ren2 = b_renames->items[j++].util;
2045                 } else if (j >= b_renames->nr) {
2046                         ren1 = a_renames->items[i++].util;
2047                 } else {
2048                         int compare = strcmp(a_renames->items[i].string,
2049                                              b_renames->items[j].string);
2050                         if (compare <= 0)
2051                                 ren1 = a_renames->items[i++].util;
2052                         if (compare >= 0)
2053                                 ren2 = b_renames->items[j++].util;
2054                 }
2055
2056                 /* TODO: refactor, so that 1/2 are not needed */
2057                 if (ren1) {
2058                         renames1 = a_renames;
2059                         renames2Dst = &b_by_dst;
2060                         branch1 = o->branch1;
2061                         branch2 = o->branch2;
2062                 } else {
2063                         renames1 = b_renames;
2064                         renames2Dst = &a_by_dst;
2065                         branch1 = o->branch2;
2066                         branch2 = o->branch1;
2067                         SWAP(ren2, ren1);
2068                 }
2069
2070                 if (ren1->processed)
2071                         continue;
2072                 ren1->processed = 1;
2073                 ren1->dst_entry->processed = 1;
2074                 /* BUG: We should only mark src_entry as processed if we
2075                  * are not dealing with a rename + add-source case.
2076                  */
2077                 ren1->src_entry->processed = 1;
2078
2079                 ren1_src = ren1->pair->one->path;
2080                 ren1_dst = ren1->pair->two->path;
2081
2082                 if (ren2) {
2083                         /* One file renamed on both sides */
2084                         const char *ren2_src = ren2->pair->one->path;
2085                         const char *ren2_dst = ren2->pair->two->path;
2086                         enum rename_type rename_type;
2087                         if (strcmp(ren1_src, ren2_src) != 0)
2088                                 die("BUG: ren1_src != ren2_src");
2089                         ren2->dst_entry->processed = 1;
2090                         ren2->processed = 1;
2091                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2092                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2093                                 clean_merge = 0;
2094                         } else {
2095                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2096                                 /* BUG: We should only remove ren1_src in
2097                                  * the base stage (think of rename +
2098                                  * add-source cases).
2099                                  */
2100                                 remove_file(o, 1, ren1_src, 1);
2101                                 update_entry(ren1->dst_entry,
2102                                              ren1->pair->one,
2103                                              ren1->pair->two,
2104                                              ren2->pair->two);
2105                         }
2106                         setup_rename_conflict_info(rename_type,
2107                                                    ren1->pair,
2108                                                    ren2->pair,
2109                                                    branch1,
2110                                                    branch2,
2111                                                    ren1->dst_entry,
2112                                                    ren2->dst_entry,
2113                                                    o,
2114                                                    NULL,
2115                                                    NULL);
2116                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2117                         /* Two different files renamed to the same thing */
2118                         char *ren2_dst;
2119                         ren2 = lookup->util;
2120                         ren2_dst = ren2->pair->two->path;
2121                         if (strcmp(ren1_dst, ren2_dst) != 0)
2122                                 die("BUG: ren1_dst != ren2_dst");
2123
2124                         clean_merge = 0;
2125                         ren2->processed = 1;
2126                         /*
2127                          * BUG: We should only mark src_entry as processed
2128                          * if we are not dealing with a rename + add-source
2129                          * case.
2130                          */
2131                         ren2->src_entry->processed = 1;
2132
2133                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2134                                                    ren1->pair,
2135                                                    ren2->pair,
2136                                                    branch1,
2137                                                    branch2,
2138                                                    ren1->dst_entry,
2139                                                    ren2->dst_entry,
2140                                                    o,
2141                                                    ren1->src_entry,
2142                                                    ren2->src_entry);
2143
2144                 } else {
2145                         /* Renamed in 1, maybe changed in 2 */
2146                         /* we only use sha1 and mode of these */
2147                         struct diff_filespec src_other, dst_other;
2148                         int try_merge;
2149
2150                         /*
2151                          * unpack_trees loads entries from common-commit
2152                          * into stage 1, from head-commit into stage 2, and
2153                          * from merge-commit into stage 3.  We keep track
2154                          * of which side corresponds to the rename.
2155                          */
2156                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2157                         int other_stage =   a_renames == renames1 ? 3 : 2;
2158
2159                         /* BUG: We should only remove ren1_src in the base
2160                          * stage and in other_stage (think of rename +
2161                          * add-source case).
2162                          */
2163                         remove_file(o, 1, ren1_src,
2164                                     renamed_stage == 2 || !was_tracked(ren1_src));
2165
2166                         oidcpy(&src_other.oid,
2167                                &ren1->src_entry->stages[other_stage].oid);
2168                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2169                         oidcpy(&dst_other.oid,
2170                                &ren1->dst_entry->stages[other_stage].oid);
2171                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2172                         try_merge = 0;
2173
2174                         if (oid_eq(&src_other.oid, &null_oid)) {
2175                                 setup_rename_conflict_info(RENAME_DELETE,
2176                                                            ren1->pair,
2177                                                            NULL,
2178                                                            branch1,
2179                                                            branch2,
2180                                                            ren1->dst_entry,
2181                                                            NULL,
2182                                                            o,
2183                                                            NULL,
2184                                                            NULL);
2185                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2186                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2187                                 /*
2188                                  * Added file on the other side identical to
2189                                  * the file being renamed: clean merge.
2190                                  * Also, there is no need to overwrite the
2191                                  * file already in the working copy, so call
2192                                  * update_file_flags() instead of
2193                                  * update_file().
2194                                  */
2195                                 if (update_file_flags(o,
2196                                                       &ren1->pair->two->oid,
2197                                                       ren1->pair->two->mode,
2198                                                       ren1_dst,
2199                                                       1, /* update_cache */
2200                                                       0  /* update_wd    */))
2201                                         clean_merge = -1;
2202                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2203                                 clean_merge = 0;
2204                                 try_merge = 1;
2205                                 output(o, 1, _("CONFLICT (rename/add): Rename %s->%s in %s. "
2206                                        "%s added in %s"),
2207                                        ren1_src, ren1_dst, branch1,
2208                                        ren1_dst, branch2);
2209                                 if (o->call_depth) {
2210                                         struct merge_file_info mfi;
2211                                         if (merge_file_one(o, ren1_dst, &null_oid, 0,
2212                                                            &ren1->pair->two->oid,
2213                                                            ren1->pair->two->mode,
2214                                                            &dst_other.oid,
2215                                                            dst_other.mode,
2216                                                            branch1, branch2, &mfi)) {
2217                                                 clean_merge = -1;
2218                                                 goto cleanup_and_return;
2219                                         }
2220                                         output(o, 1, _("Adding merged %s"), ren1_dst);
2221                                         if (update_file(o, 0, &mfi.oid,
2222                                                         mfi.mode, ren1_dst))
2223                                                 clean_merge = -1;
2224                                         try_merge = 0;
2225                                 } else {
2226                                         char *new_path = unique_path(o, ren1_dst, branch2);
2227                                         output(o, 1, _("Adding as %s instead"), new_path);
2228                                         if (update_file(o, 0, &dst_other.oid,
2229                                                         dst_other.mode, new_path))
2230                                                 clean_merge = -1;
2231                                         free(new_path);
2232                                 }
2233                         } else
2234                                 try_merge = 1;
2235
2236                         if (clean_merge < 0)
2237                                 goto cleanup_and_return;
2238                         if (try_merge) {
2239                                 struct diff_filespec *one, *a, *b;
2240                                 src_other.path = (char *)ren1_src;
2241
2242                                 one = ren1->pair->one;
2243                                 if (a_renames == renames1) {
2244                                         a = ren1->pair->two;
2245                                         b = &src_other;
2246                                 } else {
2247                                         b = ren1->pair->two;
2248                                         a = &src_other;
2249                                 }
2250                                 update_entry(ren1->dst_entry, one, a, b);
2251                                 setup_rename_conflict_info(RENAME_NORMAL,
2252                                                            ren1->pair,
2253                                                            NULL,
2254                                                            branch1,
2255                                                            NULL,
2256                                                            ren1->dst_entry,
2257                                                            NULL,
2258                                                            o,
2259                                                            NULL,
2260                                                            NULL);
2261                         }
2262                 }
2263         }
2264 cleanup_and_return:
2265         string_list_clear(&a_by_dst, 0);
2266         string_list_clear(&b_by_dst, 0);
2267
2268         return clean_merge;
2269 }
2270
2271 struct rename_info {
2272         struct string_list *head_renames;
2273         struct string_list *merge_renames;
2274 };
2275
2276 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2277                                    struct hashmap *dir_renames)
2278 {
2279         struct hashmap_iter iter;
2280         struct dir_rename_entry *e;
2281
2282         hashmap_iter_init(dir_renames, &iter);
2283         while ((e = hashmap_iter_next(&iter))) {
2284                 free(e->dir);
2285                 strbuf_release(&e->new_dir);
2286                 /* possible_new_dirs already cleared in get_directory_renames */
2287         }
2288         hashmap_free(dir_renames, 1);
2289         free(dir_renames);
2290
2291         free(pairs->queue);
2292         free(pairs);
2293 }
2294
2295 static int handle_renames(struct merge_options *o,
2296                           struct tree *common,
2297                           struct tree *head,
2298                           struct tree *merge,
2299                           struct string_list *entries,
2300                           struct rename_info *ri)
2301 {
2302         struct diff_queue_struct *head_pairs, *merge_pairs;
2303         struct hashmap *dir_re_head, *dir_re_merge;
2304         int clean = 1;
2305
2306         ri->head_renames = NULL;
2307         ri->merge_renames = NULL;
2308
2309         if (!o->detect_rename)
2310                 return 1;
2311
2312         head_pairs = get_diffpairs(o, common, head);
2313         merge_pairs = get_diffpairs(o, common, merge);
2314
2315         dir_re_head = get_directory_renames(head_pairs, head);
2316         dir_re_merge = get_directory_renames(merge_pairs, merge);
2317
2318         handle_directory_level_conflicts(o,
2319                                          dir_re_head, head,
2320                                          dir_re_merge, merge);
2321
2322         ri->head_renames  = get_renames(o, head_pairs,
2323                                         dir_re_merge, dir_re_head, head,
2324                                         common, head, merge, entries,
2325                                         &clean);
2326         if (clean < 0)
2327                 goto cleanup;
2328         ri->merge_renames = get_renames(o, merge_pairs,
2329                                         dir_re_head, dir_re_merge, merge,
2330                                         common, head, merge, entries,
2331                                         &clean);
2332         if (clean < 0)
2333                 goto cleanup;
2334         clean &= process_renames(o, ri->head_renames, ri->merge_renames);
2335
2336 cleanup:
2337         /*
2338          * Some cleanup is deferred until cleanup_renames() because the
2339          * data structures are still needed and referenced in
2340          * process_entry().  But there are a few things we can free now.
2341          */
2342         initial_cleanup_rename(head_pairs, dir_re_head);
2343         initial_cleanup_rename(merge_pairs, dir_re_merge);
2344
2345         return clean;
2346 }
2347
2348 static void final_cleanup_rename(struct string_list *rename)
2349 {
2350         const struct rename *re;
2351         int i;
2352
2353         if (rename == NULL)
2354                 return;
2355
2356         for (i = 0; i < rename->nr; i++) {
2357                 re = rename->items[i].util;
2358                 diff_free_filepair(re->pair);
2359         }
2360         string_list_clear(rename, 1);
2361         free(rename);
2362 }
2363
2364 static void final_cleanup_renames(struct rename_info *re_info)
2365 {
2366         final_cleanup_rename(re_info->head_renames);
2367         final_cleanup_rename(re_info->merge_renames);
2368 }
2369
2370 static struct object_id *stage_oid(const struct object_id *oid, unsigned mode)
2371 {
2372         return (is_null_oid(oid) || mode == 0) ? NULL: (struct object_id *)oid;
2373 }
2374
2375 static int read_oid_strbuf(struct merge_options *o,
2376         const struct object_id *oid, struct strbuf *dst)
2377 {
2378         void *buf;
2379         enum object_type type;
2380         unsigned long size;
2381         buf = read_object_file(oid, &type, &size);
2382         if (!buf)
2383                 return err(o, _("cannot read object %s"), oid_to_hex(oid));
2384         if (type != OBJ_BLOB) {
2385                 free(buf);
2386                 return err(o, _("object %s is not a blob"), oid_to_hex(oid));
2387         }
2388         strbuf_attach(dst, buf, size, size + 1);
2389         return 0;
2390 }
2391
2392 static int blob_unchanged(struct merge_options *opt,
2393                           const struct object_id *o_oid,
2394                           unsigned o_mode,
2395                           const struct object_id *a_oid,
2396                           unsigned a_mode,
2397                           int renormalize, const char *path)
2398 {
2399         struct strbuf o = STRBUF_INIT;
2400         struct strbuf a = STRBUF_INIT;
2401         int ret = 0; /* assume changed for safety */
2402
2403         if (a_mode != o_mode)
2404                 return 0;
2405         if (oid_eq(o_oid, a_oid))
2406                 return 1;
2407         if (!renormalize)
2408                 return 0;
2409
2410         assert(o_oid && a_oid);
2411         if (read_oid_strbuf(opt, o_oid, &o) || read_oid_strbuf(opt, a_oid, &a))
2412                 goto error_return;
2413         /*
2414          * Note: binary | is used so that both renormalizations are
2415          * performed.  Comparison can be skipped if both files are
2416          * unchanged since their sha1s have already been compared.
2417          */
2418         if (renormalize_buffer(&the_index, path, o.buf, o.len, &o) |
2419             renormalize_buffer(&the_index, path, a.buf, a.len, &a))
2420                 ret = (o.len == a.len && !memcmp(o.buf, a.buf, o.len));
2421
2422 error_return:
2423         strbuf_release(&o);
2424         strbuf_release(&a);
2425         return ret;
2426 }
2427
2428 static int handle_modify_delete(struct merge_options *o,
2429                                  const char *path,
2430                                  struct object_id *o_oid, int o_mode,
2431                                  struct object_id *a_oid, int a_mode,
2432                                  struct object_id *b_oid, int b_mode)
2433 {
2434         const char *modify_branch, *delete_branch;
2435         struct object_id *changed_oid;
2436         int changed_mode;
2437
2438         if (a_oid) {
2439                 modify_branch = o->branch1;
2440                 delete_branch = o->branch2;
2441                 changed_oid = a_oid;
2442                 changed_mode = a_mode;
2443         } else {
2444                 modify_branch = o->branch2;
2445                 delete_branch = o->branch1;
2446                 changed_oid = b_oid;
2447                 changed_mode = b_mode;
2448         }
2449
2450         return handle_change_delete(o,
2451                                     path, NULL,
2452                                     o_oid, o_mode,
2453                                     changed_oid, changed_mode,
2454                                     modify_branch, delete_branch,
2455                                     _("modify"), _("modified"));
2456 }
2457
2458 static int merge_content(struct merge_options *o,
2459                          const char *path,
2460                          struct object_id *o_oid, int o_mode,
2461                          struct object_id *a_oid, int a_mode,
2462                          struct object_id *b_oid, int b_mode,
2463                          struct rename_conflict_info *rename_conflict_info)
2464 {
2465         const char *reason = _("content");
2466         const char *path1 = NULL, *path2 = NULL;
2467         struct merge_file_info mfi;
2468         struct diff_filespec one, a, b;
2469         unsigned df_conflict_remains = 0;
2470
2471         if (!o_oid) {
2472                 reason = _("add/add");
2473                 o_oid = (struct object_id *)&null_oid;
2474         }
2475         one.path = a.path = b.path = (char *)path;
2476         oidcpy(&one.oid, o_oid);
2477         one.mode = o_mode;
2478         oidcpy(&a.oid, a_oid);
2479         a.mode = a_mode;
2480         oidcpy(&b.oid, b_oid);
2481         b.mode = b_mode;
2482
2483         if (rename_conflict_info) {
2484                 struct diff_filepair *pair1 = rename_conflict_info->pair1;
2485
2486                 path1 = (o->branch1 == rename_conflict_info->branch1) ?
2487                         pair1->two->path : pair1->one->path;
2488                 /* If rename_conflict_info->pair2 != NULL, we are in
2489                  * RENAME_ONE_FILE_TO_ONE case.  Otherwise, we have a
2490                  * normal rename.
2491                  */
2492                 path2 = (rename_conflict_info->pair2 ||
2493                          o->branch2 == rename_conflict_info->branch1) ?
2494                         pair1->two->path : pair1->one->path;
2495
2496                 if (dir_in_way(path, !o->call_depth,
2497                                S_ISGITLINK(pair1->two->mode)))
2498                         df_conflict_remains = 1;
2499         }
2500         if (merge_file_special_markers(o, &one, &a, &b,
2501                                        o->branch1, path1,
2502                                        o->branch2, path2, &mfi))
2503                 return -1;
2504
2505         if (mfi.clean && !df_conflict_remains &&
2506             oid_eq(&mfi.oid, a_oid) && mfi.mode == a_mode) {
2507                 int path_renamed_outside_HEAD;
2508                 output(o, 3, _("Skipped %s (merged same as existing)"), path);
2509                 /*
2510                  * The content merge resulted in the same file contents we
2511                  * already had.  We can return early if those file contents
2512                  * are recorded at the correct path (which may not be true
2513                  * if the merge involves a rename).
2514                  */
2515                 path_renamed_outside_HEAD = !path2 || !strcmp(path, path2);
2516                 if (!path_renamed_outside_HEAD) {
2517                         add_cacheinfo(o, mfi.mode, &mfi.oid, path,
2518                                       0, (!o->call_depth), 0);
2519                         return mfi.clean;
2520                 }
2521         } else
2522                 output(o, 2, _("Auto-merging %s"), path);
2523
2524         if (!mfi.clean) {
2525                 if (S_ISGITLINK(mfi.mode))
2526                         reason = _("submodule");
2527                 output(o, 1, _("CONFLICT (%s): Merge conflict in %s"),
2528                                 reason, path);
2529                 if (rename_conflict_info && !df_conflict_remains)
2530                         if (update_stages(o, path, &one, &a, &b))
2531                                 return -1;
2532         }
2533
2534         if (df_conflict_remains) {
2535                 char *new_path;
2536                 if (o->call_depth) {
2537                         remove_file_from_cache(path);
2538                 } else {
2539                         if (!mfi.clean) {
2540                                 if (update_stages(o, path, &one, &a, &b))
2541                                         return -1;
2542                         } else {
2543                                 int file_from_stage2 = was_tracked(path);
2544                                 struct diff_filespec merged;
2545                                 oidcpy(&merged.oid, &mfi.oid);
2546                                 merged.mode = mfi.mode;
2547
2548                                 if (update_stages(o, path, NULL,
2549                                                   file_from_stage2 ? &merged : NULL,
2550                                                   file_from_stage2 ? NULL : &merged))
2551                                         return -1;
2552                         }
2553
2554                 }
2555                 new_path = unique_path(o, path, rename_conflict_info->branch1);
2556                 output(o, 1, _("Adding as %s instead"), new_path);
2557                 if (update_file(o, 0, &mfi.oid, mfi.mode, new_path)) {
2558                         free(new_path);
2559                         return -1;
2560                 }
2561                 free(new_path);
2562                 mfi.clean = 0;
2563         } else if (update_file(o, mfi.clean, &mfi.oid, mfi.mode, path))
2564                 return -1;
2565         return mfi.clean;
2566 }
2567
2568 /* Per entry merge function */
2569 static int process_entry(struct merge_options *o,
2570                          const char *path, struct stage_data *entry)
2571 {
2572         int clean_merge = 1;
2573         int normalize = o->renormalize;
2574         unsigned o_mode = entry->stages[1].mode;
2575         unsigned a_mode = entry->stages[2].mode;
2576         unsigned b_mode = entry->stages[3].mode;
2577         struct object_id *o_oid = stage_oid(&entry->stages[1].oid, o_mode);
2578         struct object_id *a_oid = stage_oid(&entry->stages[2].oid, a_mode);
2579         struct object_id *b_oid = stage_oid(&entry->stages[3].oid, b_mode);
2580
2581         entry->processed = 1;
2582         if (entry->rename_conflict_info) {
2583                 struct rename_conflict_info *conflict_info = entry->rename_conflict_info;
2584                 switch (conflict_info->rename_type) {
2585                 case RENAME_NORMAL:
2586                 case RENAME_ONE_FILE_TO_ONE:
2587                         clean_merge = merge_content(o, path,
2588                                                     o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
2589                                                     conflict_info);
2590                         break;
2591                 case RENAME_DELETE:
2592                         clean_merge = 0;
2593                         if (conflict_rename_delete(o,
2594                                                    conflict_info->pair1,
2595                                                    conflict_info->branch1,
2596                                                    conflict_info->branch2))
2597                                 clean_merge = -1;
2598                         break;
2599                 case RENAME_ONE_FILE_TO_TWO:
2600                         clean_merge = 0;
2601                         if (conflict_rename_rename_1to2(o, conflict_info))
2602                                 clean_merge = -1;
2603                         break;
2604                 case RENAME_TWO_FILES_TO_ONE:
2605                         clean_merge = 0;
2606                         if (conflict_rename_rename_2to1(o, conflict_info))
2607                                 clean_merge = -1;
2608                         break;
2609                 default:
2610                         entry->processed = 0;
2611                         break;
2612                 }
2613         } else if (o_oid && (!a_oid || !b_oid)) {
2614                 /* Case A: Deleted in one */
2615                 if ((!a_oid && !b_oid) ||
2616                     (!b_oid && blob_unchanged(o, o_oid, o_mode, a_oid, a_mode, normalize, path)) ||
2617                     (!a_oid && blob_unchanged(o, o_oid, o_mode, b_oid, b_mode, normalize, path))) {
2618                         /* Deleted in both or deleted in one and
2619                          * unchanged in the other */
2620                         if (a_oid)
2621                                 output(o, 2, _("Removing %s"), path);
2622                         /* do not touch working file if it did not exist */
2623                         remove_file(o, 1, path, !a_oid);
2624                 } else {
2625                         /* Modify/delete; deleted side may have put a directory in the way */
2626                         clean_merge = 0;
2627                         if (handle_modify_delete(o, path, o_oid, o_mode,
2628                                                  a_oid, a_mode, b_oid, b_mode))
2629                                 clean_merge = -1;
2630                 }
2631         } else if ((!o_oid && a_oid && !b_oid) ||
2632                    (!o_oid && !a_oid && b_oid)) {
2633                 /* Case B: Added in one. */
2634                 /* [nothing|directory] -> ([nothing|directory], file) */
2635
2636                 const char *add_branch;
2637                 const char *other_branch;
2638                 unsigned mode;
2639                 const struct object_id *oid;
2640                 const char *conf;
2641
2642                 if (a_oid) {
2643                         add_branch = o->branch1;
2644                         other_branch = o->branch2;
2645                         mode = a_mode;
2646                         oid = a_oid;
2647                         conf = _("file/directory");
2648                 } else {
2649                         add_branch = o->branch2;
2650                         other_branch = o->branch1;
2651                         mode = b_mode;
2652                         oid = b_oid;
2653                         conf = _("directory/file");
2654                 }
2655                 if (dir_in_way(path,
2656                                !o->call_depth && !S_ISGITLINK(a_mode),
2657                                0)) {
2658                         char *new_path = unique_path(o, path, add_branch);
2659                         clean_merge = 0;
2660                         output(o, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
2661                                "Adding %s as %s"),
2662                                conf, path, other_branch, path, new_path);
2663                         if (update_file(o, 0, oid, mode, new_path))
2664                                 clean_merge = -1;
2665                         else if (o->call_depth)
2666                                 remove_file_from_cache(path);
2667                         free(new_path);
2668                 } else {
2669                         output(o, 2, _("Adding %s"), path);
2670                         /* do not overwrite file if already present */
2671                         if (update_file_flags(o, oid, mode, path, 1, !a_oid))
2672                                 clean_merge = -1;
2673                 }
2674         } else if (a_oid && b_oid) {
2675                 /* Case C: Added in both (check for same permissions) and */
2676                 /* case D: Modified in both, but differently. */
2677                 clean_merge = merge_content(o, path,
2678                                             o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
2679                                             NULL);
2680         } else if (!o_oid && !a_oid && !b_oid) {
2681                 /*
2682                  * this entry was deleted altogether. a_mode == 0 means
2683                  * we had that path and want to actively remove it.
2684                  */
2685                 remove_file(o, 1, path, !a_mode);
2686         } else
2687                 die("BUG: fatal merge failure, shouldn't happen.");
2688
2689         return clean_merge;
2690 }
2691
2692 int merge_trees(struct merge_options *o,
2693                 struct tree *head,
2694                 struct tree *merge,
2695                 struct tree *common,
2696                 struct tree **result)
2697 {
2698         int code, clean;
2699
2700         if (o->subtree_shift) {
2701                 merge = shift_tree_object(head, merge, o->subtree_shift);
2702                 common = shift_tree_object(head, common, o->subtree_shift);
2703         }
2704
2705         if (oid_eq(&common->object.oid, &merge->object.oid)) {
2706                 struct strbuf sb = STRBUF_INIT;
2707
2708                 if (!o->call_depth && index_has_changes(&sb)) {
2709                         err(o, _("Dirty index: cannot merge (dirty: %s)"),
2710                             sb.buf);
2711                         return 0;
2712                 }
2713                 output(o, 0, _("Already up to date!"));
2714                 *result = head;
2715                 return 1;
2716         }
2717
2718         code = git_merge_trees(o->call_depth, common, head, merge);
2719
2720         if (code != 0) {
2721                 if (show(o, 4) || o->call_depth)
2722                         err(o, _("merging of trees %s and %s failed"),
2723                             oid_to_hex(&head->object.oid),
2724                             oid_to_hex(&merge->object.oid));
2725                 return -1;
2726         }
2727
2728         if (unmerged_cache()) {
2729                 struct string_list *entries;
2730                 struct rename_info re_info;
2731                 int i;
2732                 /*
2733                  * Only need the hashmap while processing entries, so
2734                  * initialize it here and free it when we are done running
2735                  * through the entries. Keeping it in the merge_options as
2736                  * opposed to decaring a local hashmap is for convenience
2737                  * so that we don't have to pass it to around.
2738                  */
2739                 hashmap_init(&o->current_file_dir_set, path_hashmap_cmp, NULL, 512);
2740                 get_files_dirs(o, head);
2741                 get_files_dirs(o, merge);
2742
2743                 entries = get_unmerged();
2744                 clean = handle_renames(o, common, head, merge, entries,
2745                                        &re_info);
2746                 record_df_conflict_files(o, entries);
2747                 if (clean < 0)
2748                         goto cleanup;
2749                 for (i = entries->nr-1; 0 <= i; i--) {
2750                         const char *path = entries->items[i].string;
2751                         struct stage_data *e = entries->items[i].util;
2752                         if (!e->processed) {
2753                                 int ret = process_entry(o, path, e);
2754                                 if (!ret)
2755                                         clean = 0;
2756                                 else if (ret < 0) {
2757                                         clean = ret;
2758                                         goto cleanup;
2759                                 }
2760                         }
2761                 }
2762                 for (i = 0; i < entries->nr; i++) {
2763                         struct stage_data *e = entries->items[i].util;
2764                         if (!e->processed)
2765                                 die("BUG: unprocessed path??? %s",
2766                                     entries->items[i].string);
2767                 }
2768
2769 cleanup:
2770                 final_cleanup_renames(&re_info);
2771
2772                 string_list_clear(entries, 1);
2773                 free(entries);
2774
2775                 hashmap_free(&o->current_file_dir_set, 1);
2776
2777                 if (clean < 0)
2778                         return clean;
2779         }
2780         else
2781                 clean = 1;
2782
2783         if (o->call_depth && !(*result = write_tree_from_memory(o)))
2784                 return -1;
2785
2786         return clean;
2787 }
2788
2789 static struct commit_list *reverse_commit_list(struct commit_list *list)
2790 {
2791         struct commit_list *next = NULL, *current, *backup;
2792         for (current = list; current; current = backup) {
2793                 backup = current->next;
2794                 current->next = next;
2795                 next = current;
2796         }
2797         return next;
2798 }
2799
2800 /*
2801  * Merge the commits h1 and h2, return the resulting virtual
2802  * commit object and a flag indicating the cleanness of the merge.
2803  */
2804 int merge_recursive(struct merge_options *o,
2805                     struct commit *h1,
2806                     struct commit *h2,
2807                     struct commit_list *ca,
2808                     struct commit **result)
2809 {
2810         struct commit_list *iter;
2811         struct commit *merged_common_ancestors;
2812         struct tree *mrtree;
2813         int clean;
2814
2815         if (show(o, 4)) {
2816                 output(o, 4, _("Merging:"));
2817                 output_commit_title(o, h1);
2818                 output_commit_title(o, h2);
2819         }
2820
2821         if (!ca) {
2822                 ca = get_merge_bases(h1, h2);
2823                 ca = reverse_commit_list(ca);
2824         }
2825
2826         if (show(o, 5)) {
2827                 unsigned cnt = commit_list_count(ca);
2828
2829                 output(o, 5, Q_("found %u common ancestor:",
2830                                 "found %u common ancestors:", cnt), cnt);
2831                 for (iter = ca; iter; iter = iter->next)
2832                         output_commit_title(o, iter->item);
2833         }
2834
2835         merged_common_ancestors = pop_commit(&ca);
2836         if (merged_common_ancestors == NULL) {
2837                 /* if there is no common ancestor, use an empty tree */
2838                 struct tree *tree;
2839
2840                 tree = lookup_tree(the_hash_algo->empty_tree);
2841                 merged_common_ancestors = make_virtual_commit(tree, "ancestor");
2842         }
2843
2844         for (iter = ca; iter; iter = iter->next) {
2845                 const char *saved_b1, *saved_b2;
2846                 o->call_depth++;
2847                 /*
2848                  * When the merge fails, the result contains files
2849                  * with conflict markers. The cleanness flag is
2850                  * ignored (unless indicating an error), it was never
2851                  * actually used, as result of merge_trees has always
2852                  * overwritten it: the committed "conflicts" were
2853                  * already resolved.
2854                  */
2855                 discard_cache();
2856                 saved_b1 = o->branch1;
2857                 saved_b2 = o->branch2;
2858                 o->branch1 = "Temporary merge branch 1";
2859                 o->branch2 = "Temporary merge branch 2";
2860                 if (merge_recursive(o, merged_common_ancestors, iter->item,
2861                                     NULL, &merged_common_ancestors) < 0)
2862                         return -1;
2863                 o->branch1 = saved_b1;
2864                 o->branch2 = saved_b2;
2865                 o->call_depth--;
2866
2867                 if (!merged_common_ancestors)
2868                         return err(o, _("merge returned no commit"));
2869         }
2870
2871         discard_cache();
2872         if (!o->call_depth)
2873                 read_cache();
2874
2875         o->ancestor = "merged common ancestors";
2876         clean = merge_trees(o, h1->tree, h2->tree, merged_common_ancestors->tree,
2877                             &mrtree);
2878         if (clean < 0) {
2879                 flush_output(o);
2880                 return clean;
2881         }
2882
2883         if (o->call_depth) {
2884                 *result = make_virtual_commit(mrtree, "merged tree");
2885                 commit_list_insert(h1, &(*result)->parents);
2886                 commit_list_insert(h2, &(*result)->parents->next);
2887         }
2888         flush_output(o);
2889         if (!o->call_depth && o->buffer_output < 2)
2890                 strbuf_release(&o->obuf);
2891         if (show(o, 2))
2892                 diff_warn_rename_limit("merge.renamelimit",
2893                                        o->needed_rename_limit, 0);
2894         return clean;
2895 }
2896
2897 static struct commit *get_ref(const struct object_id *oid, const char *name)
2898 {
2899         struct object *object;
2900
2901         object = deref_tag(parse_object(oid), name, strlen(name));
2902         if (!object)
2903                 return NULL;
2904         if (object->type == OBJ_TREE)
2905                 return make_virtual_commit((struct tree*)object, name);
2906         if (object->type != OBJ_COMMIT)
2907                 return NULL;
2908         if (parse_commit((struct commit *)object))
2909                 return NULL;
2910         return (struct commit *)object;
2911 }
2912
2913 int merge_recursive_generic(struct merge_options *o,
2914                             const struct object_id *head,
2915                             const struct object_id *merge,
2916                             int num_base_list,
2917                             const struct object_id **base_list,
2918                             struct commit **result)
2919 {
2920         int clean;
2921         struct lock_file lock = LOCK_INIT;
2922         struct commit *head_commit = get_ref(head, o->branch1);
2923         struct commit *next_commit = get_ref(merge, o->branch2);
2924         struct commit_list *ca = NULL;
2925
2926         if (base_list) {
2927                 int i;
2928                 for (i = 0; i < num_base_list; ++i) {
2929                         struct commit *base;
2930                         if (!(base = get_ref(base_list[i], oid_to_hex(base_list[i]))))
2931                                 return err(o, _("Could not parse object '%s'"),
2932                                         oid_to_hex(base_list[i]));
2933                         commit_list_insert(base, &ca);
2934                 }
2935         }
2936
2937         hold_locked_index(&lock, LOCK_DIE_ON_ERROR);
2938         clean = merge_recursive(o, head_commit, next_commit, ca,
2939                         result);
2940         if (clean < 0) {
2941                 rollback_lock_file(&lock);
2942                 return clean;
2943         }
2944
2945         if (write_locked_index(&the_index, &lock,
2946                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
2947                 return err(o, _("Unable to write index."));
2948
2949         return clean ? 0 : 1;
2950 }
2951
2952 static void merge_recursive_config(struct merge_options *o)
2953 {
2954         git_config_get_int("merge.verbosity", &o->verbosity);
2955         git_config_get_int("diff.renamelimit", &o->diff_rename_limit);
2956         git_config_get_int("merge.renamelimit", &o->merge_rename_limit);
2957         git_config(git_xmerge_config, NULL);
2958 }
2959
2960 void init_merge_options(struct merge_options *o)
2961 {
2962         const char *merge_verbosity;
2963         memset(o, 0, sizeof(struct merge_options));
2964         o->verbosity = 2;
2965         o->buffer_output = 1;
2966         o->diff_rename_limit = -1;
2967         o->merge_rename_limit = -1;
2968         o->renormalize = 0;
2969         o->detect_rename = 1;
2970         merge_recursive_config(o);
2971         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
2972         if (merge_verbosity)
2973                 o->verbosity = strtol(merge_verbosity, NULL, 10);
2974         if (o->verbosity >= 5)
2975                 o->buffer_output = 0;
2976         strbuf_init(&o->obuf, 0);
2977         string_list_init(&o->df_conflict_file_set, 1);
2978 }
2979
2980 int parse_merge_opt(struct merge_options *o, const char *s)
2981 {
2982         const char *arg;
2983
2984         if (!s || !*s)
2985                 return -1;
2986         if (!strcmp(s, "ours"))
2987                 o->recursive_variant = MERGE_RECURSIVE_OURS;
2988         else if (!strcmp(s, "theirs"))
2989                 o->recursive_variant = MERGE_RECURSIVE_THEIRS;
2990         else if (!strcmp(s, "subtree"))
2991                 o->subtree_shift = "";
2992         else if (skip_prefix(s, "subtree=", &arg))
2993                 o->subtree_shift = arg;
2994         else if (!strcmp(s, "patience"))
2995                 o->xdl_opts = DIFF_WITH_ALG(o, PATIENCE_DIFF);
2996         else if (!strcmp(s, "histogram"))
2997                 o->xdl_opts = DIFF_WITH_ALG(o, HISTOGRAM_DIFF);
2998         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
2999                 long value = parse_algorithm_value(arg);
3000                 if (value < 0)
3001                         return -1;
3002                 /* clear out previous settings */
3003                 DIFF_XDL_CLR(o, NEED_MINIMAL);
3004                 o->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3005                 o->xdl_opts |= value;
3006         }
3007         else if (!strcmp(s, "ignore-space-change"))
3008                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_CHANGE);
3009         else if (!strcmp(s, "ignore-all-space"))
3010                 DIFF_XDL_SET(o, IGNORE_WHITESPACE);
3011         else if (!strcmp(s, "ignore-space-at-eol"))
3012                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_AT_EOL);
3013         else if (!strcmp(s, "ignore-cr-at-eol"))
3014                 DIFF_XDL_SET(o, IGNORE_CR_AT_EOL);
3015         else if (!strcmp(s, "renormalize"))
3016                 o->renormalize = 1;
3017         else if (!strcmp(s, "no-renormalize"))
3018                 o->renormalize = 0;
3019         else if (!strcmp(s, "no-renames"))
3020                 o->detect_rename = 0;
3021         else if (!strcmp(s, "find-renames")) {
3022                 o->detect_rename = 1;
3023                 o->rename_score = 0;
3024         }
3025         else if (skip_prefix(s, "find-renames=", &arg) ||
3026                  skip_prefix(s, "rename-threshold=", &arg)) {
3027                 if ((o->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3028                         return -1;
3029                 o->detect_rename = 1;
3030         }
3031         else
3032                 return -1;
3033         return 0;
3034 }