Merge branch 'fm/user-manual-use-preface'
[git] / merge-ort.c
1 /*
2  * "Ostensibly Recursive's Twin" merge strategy, or "ort" for short.  Meant
3  * as a drop-in replacement for the "recursive" merge strategy, allowing one
4  * to replace
5  *
6  *   git merge [-s recursive]
7  *
8  * with
9  *
10  *   git merge -s ort
11  *
12  * Note: git's parser allows the space between '-s' and its argument to be
13  * missing.  (Should I have backronymed "ham", "alsa", "kip", "nap, "alvo",
14  * "cale", "peedy", or "ins" instead of "ort"?)
15  */
16
17 #include "cache.h"
18 #include "merge-ort.h"
19
20 #include "alloc.h"
21 #include "blob.h"
22 #include "cache-tree.h"
23 #include "commit.h"
24 #include "commit-reach.h"
25 #include "diff.h"
26 #include "diffcore.h"
27 #include "dir.h"
28 #include "ll-merge.h"
29 #include "object-store.h"
30 #include "revision.h"
31 #include "strmap.h"
32 #include "submodule.h"
33 #include "tree.h"
34 #include "unpack-trees.h"
35 #include "xdiff-interface.h"
36
37 /*
38  * We have many arrays of size 3.  Whenever we have such an array, the
39  * indices refer to one of the sides of the three-way merge.  This is so
40  * pervasive that the constants 0, 1, and 2 are used in many places in the
41  * code (especially in arithmetic operations to find the other side's index
42  * or to compute a relevant mask), but sometimes these enum names are used
43  * to aid code clarity.
44  *
45  * See also 'filemask' and 'dirmask' in struct conflict_info; the "ith side"
46  * referred to there is one of these three sides.
47  */
48 enum merge_side {
49         MERGE_BASE = 0,
50         MERGE_SIDE1 = 1,
51         MERGE_SIDE2 = 2
52 };
53
54 struct traversal_callback_data {
55         unsigned long mask;
56         unsigned long dirmask;
57         struct name_entry names[3];
58 };
59
60 struct rename_info {
61         /*
62          * All variables that are arrays of size 3 correspond to data tracked
63          * for the sides in enum merge_side.  Index 0 is almost always unused
64          * because we often only need to track information for MERGE_SIDE1 and
65          * MERGE_SIDE2 (MERGE_BASE can't have rename information since renames
66          * are determined relative to what changed since the MERGE_BASE).
67          */
68
69         /*
70          * pairs: pairing of filenames from diffcore_rename()
71          */
72         struct diff_queue_struct pairs[3];
73
74         /*
75          * dirs_removed: directories removed on a given side of history.
76          */
77         struct strset dirs_removed[3];
78
79         /*
80          * dir_rename_count: tracking where parts of a directory were renamed to
81          *
82          * When files in a directory are renamed, they may not all go to the
83          * same location.  Each strmap here tracks:
84          *      old_dir => {new_dir => int}
85          * That is, dir_rename_count[side] is a strmap to a strintmap.
86          */
87         struct strmap dir_rename_count[3];
88
89         /*
90          * dir_renames: computed directory renames
91          *
92          * This is a map of old_dir => new_dir and is derived in part from
93          * dir_rename_count.
94          */
95         struct strmap dir_renames[3];
96
97         /*
98          * relevant_sources: deleted paths for which we need rename detection
99          *
100          * relevant_sources is a set of deleted paths on each side of
101          * history for which we need rename detection.  If a path is deleted
102          * on one side of history, we need to detect if it is part of a
103          * rename if either
104          *    * we need to detect renames for an ancestor directory
105          *    * the file is modified/deleted on the other side of history
106          * If neither of those are true, we can skip rename detection for
107          * that path.
108          */
109         struct strset relevant_sources[3];
110
111         /*
112          * dir_rename_mask:
113          *   0: optimization removing unmodified potential rename source okay
114          *   2 or 4: optimization okay, but must check for files added to dir
115          *   7: optimization forbidden; need rename source in case of dir rename
116          */
117         unsigned dir_rename_mask:3;
118
119         /*
120          * callback_data_*: supporting data structures for alternate traversal
121          *
122          * We sometimes need to be able to traverse through all the files
123          * in a given tree before all immediate subdirectories within that
124          * tree.  Since traverse_trees() doesn't do that naturally, we have
125          * a traverse_trees_wrapper() that stores any immediate
126          * subdirectories while traversing files, then traverses the
127          * immediate subdirectories later.  These callback_data* variables
128          * store the information for the subdirectories so that we can do
129          * that traversal order.
130          */
131         struct traversal_callback_data *callback_data;
132         int callback_data_nr, callback_data_alloc;
133         char *callback_data_traverse_path;
134
135         /*
136          * needed_limit: value needed for inexact rename detection to run
137          *
138          * If the current rename limit wasn't high enough for inexact
139          * rename detection to run, this records the limit needed.  Otherwise,
140          * this value remains 0.
141          */
142         int needed_limit;
143 };
144
145 struct merge_options_internal {
146         /*
147          * paths: primary data structure in all of merge ort.
148          *
149          * The keys of paths:
150          *   * are full relative paths from the toplevel of the repository
151          *     (e.g. "drivers/firmware/raspberrypi.c").
152          *   * store all relevant paths in the repo, both directories and
153          *     files (e.g. drivers, drivers/firmware would also be included)
154          *   * these keys serve to intern all the path strings, which allows
155          *     us to do pointer comparison on directory names instead of
156          *     strcmp; we just have to be careful to use the interned strings.
157          *     (Technically paths_to_free may track some strings that were
158          *      removed from froms paths.)
159          *
160          * The values of paths:
161          *   * either a pointer to a merged_info, or a conflict_info struct
162          *   * merged_info contains all relevant information for a
163          *     non-conflicted entry.
164          *   * conflict_info contains a merged_info, plus any additional
165          *     information about a conflict such as the higher orders stages
166          *     involved and the names of the paths those came from (handy
167          *     once renames get involved).
168          *   * a path may start "conflicted" (i.e. point to a conflict_info)
169          *     and then a later step (e.g. three-way content merge) determines
170          *     it can be cleanly merged, at which point it'll be marked clean
171          *     and the algorithm will ignore any data outside the contained
172          *     merged_info for that entry
173          *   * If an entry remains conflicted, the merged_info portion of a
174          *     conflict_info will later be filled with whatever version of
175          *     the file should be placed in the working directory (e.g. an
176          *     as-merged-as-possible variation that contains conflict markers).
177          */
178         struct strmap paths;
179
180         /*
181          * conflicted: a subset of keys->values from "paths"
182          *
183          * conflicted is basically an optimization between process_entries()
184          * and record_conflicted_index_entries(); the latter could loop over
185          * ALL the entries in paths AGAIN and look for the ones that are
186          * still conflicted, but since process_entries() has to loop over
187          * all of them, it saves the ones it couldn't resolve in this strmap
188          * so that record_conflicted_index_entries() can iterate just the
189          * relevant entries.
190          */
191         struct strmap conflicted;
192
193         /*
194          * paths_to_free: additional list of strings to free
195          *
196          * If keys are removed from "paths", they are added to paths_to_free
197          * to ensure they are later freed.  We avoid free'ing immediately since
198          * other places (e.g. conflict_info.pathnames[]) may still be
199          * referencing these paths.
200          */
201         struct string_list paths_to_free;
202
203         /*
204          * output: special messages and conflict notices for various paths
205          *
206          * This is a map of pathnames (a subset of the keys in "paths" above)
207          * to strbufs.  It gathers various warning/conflict/notice messages
208          * for later processing.
209          */
210         struct strmap output;
211
212         /*
213          * renames: various data relating to rename detection
214          */
215         struct rename_info renames;
216
217         /*
218          * current_dir_name, toplevel_dir: temporary vars
219          *
220          * These are used in collect_merge_info_callback(), and will set the
221          * various merged_info.directory_name for the various paths we get;
222          * see documentation for that variable and the requirements placed on
223          * that field.
224          */
225         const char *current_dir_name;
226         const char *toplevel_dir;
227
228         /* call_depth: recursion level counter for merging merge bases */
229         int call_depth;
230 };
231
232 struct version_info {
233         struct object_id oid;
234         unsigned short mode;
235 };
236
237 struct merged_info {
238         /* if is_null, ignore result.  otherwise result has oid & mode */
239         struct version_info result;
240         unsigned is_null:1;
241
242         /*
243          * clean: whether the path in question is cleanly merged.
244          *
245          * see conflict_info.merged for more details.
246          */
247         unsigned clean:1;
248
249         /*
250          * basename_offset: offset of basename of path.
251          *
252          * perf optimization to avoid recomputing offset of final '/'
253          * character in pathname (0 if no '/' in pathname).
254          */
255         size_t basename_offset;
256
257          /*
258           * directory_name: containing directory name.
259           *
260           * Note that we assume directory_name is constructed such that
261           *    strcmp(dir1_name, dir2_name) == 0 iff dir1_name == dir2_name,
262           * i.e. string equality is equivalent to pointer equality.  For this
263           * to hold, we have to be careful setting directory_name.
264           */
265         const char *directory_name;
266 };
267
268 struct conflict_info {
269         /*
270          * merged: the version of the path that will be written to working tree
271          *
272          * WARNING: It is critical to check merged.clean and ensure it is 0
273          * before reading any conflict_info fields outside of merged.
274          * Allocated merge_info structs will always have clean set to 1.
275          * Allocated conflict_info structs will have merged.clean set to 0
276          * initially.  The merged.clean field is how we know if it is safe
277          * to access other parts of conflict_info besides merged; if a
278          * conflict_info's merged.clean is changed to 1, the rest of the
279          * algorithm is not allowed to look at anything outside of the
280          * merged member anymore.
281          */
282         struct merged_info merged;
283
284         /* oids & modes from each of the three trees for this path */
285         struct version_info stages[3];
286
287         /* pathnames for each stage; may differ due to rename detection */
288         const char *pathnames[3];
289
290         /* Whether this path is/was involved in a directory/file conflict */
291         unsigned df_conflict:1;
292
293         /*
294          * Whether this path is/was involved in a non-content conflict other
295          * than a directory/file conflict (e.g. rename/rename, rename/delete,
296          * file location based on possible directory rename).
297          */
298         unsigned path_conflict:1;
299
300         /*
301          * For filemask and dirmask, the ith bit corresponds to whether the
302          * ith entry is a file (filemask) or a directory (dirmask).  Thus,
303          * filemask & dirmask is always zero, and filemask | dirmask is at
304          * most 7 but can be less when a path does not appear as either a
305          * file or a directory on at least one side of history.
306          *
307          * Note that these masks are related to enum merge_side, as the ith
308          * entry corresponds to side i.
309          *
310          * These values come from a traverse_trees() call; more info may be
311          * found looking at tree-walk.h's struct traverse_info,
312          * particularly the documentation above the "fn" member (note that
313          * filemask = mask & ~dirmask from that documentation).
314          */
315         unsigned filemask:3;
316         unsigned dirmask:3;
317
318         /*
319          * Optimization to track which stages match, to avoid the need to
320          * recompute it in multiple steps. Either 0 or at least 2 bits are
321          * set; if at least 2 bits are set, their corresponding stages match.
322          */
323         unsigned match_mask:3;
324 };
325
326 /*** Function Grouping: various utility functions ***/
327
328 /*
329  * For the next three macros, see warning for conflict_info.merged.
330  *
331  * In each of the below, mi is a struct merged_info*, and ci was defined
332  * as a struct conflict_info* (but we need to verify ci isn't actually
333  * pointed at a struct merged_info*).
334  *
335  * INITIALIZE_CI: Assign ci to mi but only if it's safe; set to NULL otherwise.
336  * VERIFY_CI: Ensure that something we assigned to a conflict_info* is one.
337  * ASSIGN_AND_VERIFY_CI: Similar to VERIFY_CI but do assignment first.
338  */
339 #define INITIALIZE_CI(ci, mi) do {                                           \
340         (ci) = (!(mi) || (mi)->clean) ? NULL : (struct conflict_info *)(mi); \
341 } while (0)
342 #define VERIFY_CI(ci) assert(ci && !ci->merged.clean);
343 #define ASSIGN_AND_VERIFY_CI(ci, mi) do {    \
344         (ci) = (struct conflict_info *)(mi);  \
345         assert((ci) && !(mi)->clean);        \
346 } while (0)
347
348 static void free_strmap_strings(struct strmap *map)
349 {
350         struct hashmap_iter iter;
351         struct strmap_entry *entry;
352
353         strmap_for_each_entry(map, &iter, entry) {
354                 free((char*)entry->key);
355         }
356 }
357
358 static void clear_or_reinit_internal_opts(struct merge_options_internal *opti,
359                                           int reinitialize)
360 {
361         struct rename_info *renames = &opti->renames;
362         int i;
363         void (*strmap_func)(struct strmap *, int) =
364                 reinitialize ? strmap_partial_clear : strmap_clear;
365         void (*strset_func)(struct strset *) =
366                 reinitialize ? strset_partial_clear : strset_clear;
367
368         /*
369          * We marked opti->paths with strdup_strings = 0, so that we
370          * wouldn't have to make another copy of the fullpath created by
371          * make_traverse_path from setup_path_info().  But, now that we've
372          * used it and have no other references to these strings, it is time
373          * to deallocate them.
374          */
375         free_strmap_strings(&opti->paths);
376         strmap_func(&opti->paths, 1);
377
378         /*
379          * All keys and values in opti->conflicted are a subset of those in
380          * opti->paths.  We don't want to deallocate anything twice, so we
381          * don't free the keys and we pass 0 for free_values.
382          */
383         strmap_func(&opti->conflicted, 0);
384
385         /*
386          * opti->paths_to_free is similar to opti->paths; we created it with
387          * strdup_strings = 0 to avoid making _another_ copy of the fullpath
388          * but now that we've used it and have no other references to these
389          * strings, it is time to deallocate them.  We do so by temporarily
390          * setting strdup_strings to 1.
391          */
392         opti->paths_to_free.strdup_strings = 1;
393         string_list_clear(&opti->paths_to_free, 0);
394         opti->paths_to_free.strdup_strings = 0;
395
396         /* Free memory used by various renames maps */
397         for (i = MERGE_SIDE1; i <= MERGE_SIDE2; ++i) {
398                 strset_func(&renames->dirs_removed[i]);
399
400                 partial_clear_dir_rename_count(&renames->dir_rename_count[i]);
401                 if (!reinitialize)
402                         strmap_clear(&renames->dir_rename_count[i], 1);
403
404                 strmap_func(&renames->dir_renames[i], 0);
405
406                 strset_func(&renames->relevant_sources[i]);
407         }
408
409         if (!reinitialize) {
410                 struct hashmap_iter iter;
411                 struct strmap_entry *e;
412
413                 /* Release and free each strbuf found in output */
414                 strmap_for_each_entry(&opti->output, &iter, e) {
415                         struct strbuf *sb = e->value;
416                         strbuf_release(sb);
417                         /*
418                          * While strictly speaking we don't need to free(sb)
419                          * here because we could pass free_values=1 when
420                          * calling strmap_clear() on opti->output, that would
421                          * require strmap_clear to do another
422                          * strmap_for_each_entry() loop, so we just free it
423                          * while we're iterating anyway.
424                          */
425                         free(sb);
426                 }
427                 strmap_clear(&opti->output, 0);
428         }
429
430         renames->dir_rename_mask = 0;
431
432         /* Clean out callback_data as well. */
433         FREE_AND_NULL(renames->callback_data);
434         renames->callback_data_nr = renames->callback_data_alloc = 0;
435 }
436
437 static int err(struct merge_options *opt, const char *err, ...)
438 {
439         va_list params;
440         struct strbuf sb = STRBUF_INIT;
441
442         strbuf_addstr(&sb, "error: ");
443         va_start(params, err);
444         strbuf_vaddf(&sb, err, params);
445         va_end(params);
446
447         error("%s", sb.buf);
448         strbuf_release(&sb);
449
450         return -1;
451 }
452
453 static void format_commit(struct strbuf *sb,
454                           int indent,
455                           struct commit *commit)
456 {
457         struct merge_remote_desc *desc;
458         struct pretty_print_context ctx = {0};
459         ctx.abbrev = DEFAULT_ABBREV;
460
461         strbuf_addchars(sb, ' ', indent);
462         desc = merge_remote_util(commit);
463         if (desc) {
464                 strbuf_addf(sb, "virtual %s\n", desc->name);
465                 return;
466         }
467
468         format_commit_message(commit, "%h %s", sb, &ctx);
469         strbuf_addch(sb, '\n');
470 }
471
472 __attribute__((format (printf, 4, 5)))
473 static void path_msg(struct merge_options *opt,
474                      const char *path,
475                      int omittable_hint, /* skippable under --remerge-diff */
476                      const char *fmt, ...)
477 {
478         va_list ap;
479         struct strbuf *sb = strmap_get(&opt->priv->output, path);
480         if (!sb) {
481                 sb = xmalloc(sizeof(*sb));
482                 strbuf_init(sb, 0);
483                 strmap_put(&opt->priv->output, path, sb);
484         }
485
486         va_start(ap, fmt);
487         strbuf_vaddf(sb, fmt, ap);
488         va_end(ap);
489
490         strbuf_addch(sb, '\n');
491 }
492
493 /* add a string to a strbuf, but converting "/" to "_" */
494 static void add_flattened_path(struct strbuf *out, const char *s)
495 {
496         size_t i = out->len;
497         strbuf_addstr(out, s);
498         for (; i < out->len; i++)
499                 if (out->buf[i] == '/')
500                         out->buf[i] = '_';
501 }
502
503 static char *unique_path(struct strmap *existing_paths,
504                          const char *path,
505                          const char *branch)
506 {
507         struct strbuf newpath = STRBUF_INIT;
508         int suffix = 0;
509         size_t base_len;
510
511         strbuf_addf(&newpath, "%s~", path);
512         add_flattened_path(&newpath, branch);
513
514         base_len = newpath.len;
515         while (strmap_contains(existing_paths, newpath.buf)) {
516                 strbuf_setlen(&newpath, base_len);
517                 strbuf_addf(&newpath, "_%d", suffix++);
518         }
519
520         return strbuf_detach(&newpath, NULL);
521 }
522
523 /*** Function Grouping: functions related to collect_merge_info() ***/
524
525 static int traverse_trees_wrapper_callback(int n,
526                                            unsigned long mask,
527                                            unsigned long dirmask,
528                                            struct name_entry *names,
529                                            struct traverse_info *info)
530 {
531         struct merge_options *opt = info->data;
532         struct rename_info *renames = &opt->priv->renames;
533         unsigned filemask = mask & ~dirmask;
534
535         assert(n==3);
536
537         if (!renames->callback_data_traverse_path)
538                 renames->callback_data_traverse_path = xstrdup(info->traverse_path);
539
540         if (filemask && filemask == renames->dir_rename_mask)
541                 renames->dir_rename_mask = 0x07;
542
543         ALLOC_GROW(renames->callback_data, renames->callback_data_nr + 1,
544                    renames->callback_data_alloc);
545         renames->callback_data[renames->callback_data_nr].mask = mask;
546         renames->callback_data[renames->callback_data_nr].dirmask = dirmask;
547         COPY_ARRAY(renames->callback_data[renames->callback_data_nr].names,
548                    names, 3);
549         renames->callback_data_nr++;
550
551         return mask;
552 }
553
554 /*
555  * Much like traverse_trees(), BUT:
556  *   - read all the tree entries FIRST, saving them
557  *   - note that the above step provides an opportunity to compute necessary
558  *     additional details before the "real" traversal
559  *   - loop through the saved entries and call the original callback on them
560  */
561 static int traverse_trees_wrapper(struct index_state *istate,
562                                   int n,
563                                   struct tree_desc *t,
564                                   struct traverse_info *info)
565 {
566         int ret, i, old_offset;
567         traverse_callback_t old_fn;
568         char *old_callback_data_traverse_path;
569         struct merge_options *opt = info->data;
570         struct rename_info *renames = &opt->priv->renames;
571
572         assert(renames->dir_rename_mask == 2 || renames->dir_rename_mask == 4);
573
574         old_callback_data_traverse_path = renames->callback_data_traverse_path;
575         old_fn = info->fn;
576         old_offset = renames->callback_data_nr;
577
578         renames->callback_data_traverse_path = NULL;
579         info->fn = traverse_trees_wrapper_callback;
580         ret = traverse_trees(istate, n, t, info);
581         if (ret < 0)
582                 return ret;
583
584         info->traverse_path = renames->callback_data_traverse_path;
585         info->fn = old_fn;
586         for (i = old_offset; i < renames->callback_data_nr; ++i) {
587                 info->fn(n,
588                          renames->callback_data[i].mask,
589                          renames->callback_data[i].dirmask,
590                          renames->callback_data[i].names,
591                          info);
592         }
593
594         renames->callback_data_nr = old_offset;
595         free(renames->callback_data_traverse_path);
596         renames->callback_data_traverse_path = old_callback_data_traverse_path;
597         info->traverse_path = NULL;
598         return 0;
599 }
600
601 static void setup_path_info(struct merge_options *opt,
602                             struct string_list_item *result,
603                             const char *current_dir_name,
604                             int current_dir_name_len,
605                             char *fullpath, /* we'll take over ownership */
606                             struct name_entry *names,
607                             struct name_entry *merged_version,
608                             unsigned is_null,     /* boolean */
609                             unsigned df_conflict, /* boolean */
610                             unsigned filemask,
611                             unsigned dirmask,
612                             int resolved          /* boolean */)
613 {
614         /* result->util is void*, so mi is a convenience typed variable */
615         struct merged_info *mi;
616
617         assert(!is_null || resolved);
618         assert(!df_conflict || !resolved); /* df_conflict implies !resolved */
619         assert(resolved == (merged_version != NULL));
620
621         mi = xcalloc(1, resolved ? sizeof(struct merged_info) :
622                                    sizeof(struct conflict_info));
623         mi->directory_name = current_dir_name;
624         mi->basename_offset = current_dir_name_len;
625         mi->clean = !!resolved;
626         if (resolved) {
627                 mi->result.mode = merged_version->mode;
628                 oidcpy(&mi->result.oid, &merged_version->oid);
629                 mi->is_null = !!is_null;
630         } else {
631                 int i;
632                 struct conflict_info *ci;
633
634                 ASSIGN_AND_VERIFY_CI(ci, mi);
635                 for (i = MERGE_BASE; i <= MERGE_SIDE2; i++) {
636                         ci->pathnames[i] = fullpath;
637                         ci->stages[i].mode = names[i].mode;
638                         oidcpy(&ci->stages[i].oid, &names[i].oid);
639                 }
640                 ci->filemask = filemask;
641                 ci->dirmask = dirmask;
642                 ci->df_conflict = !!df_conflict;
643                 if (dirmask)
644                         /*
645                          * Assume is_null for now, but if we have entries
646                          * under the directory then when it is complete in
647                          * write_completed_directory() it'll update this.
648                          * Also, for D/F conflicts, we have to handle the
649                          * directory first, then clear this bit and process
650                          * the file to see how it is handled -- that occurs
651                          * near the top of process_entry().
652                          */
653                         mi->is_null = 1;
654         }
655         strmap_put(&opt->priv->paths, fullpath, mi);
656         result->string = fullpath;
657         result->util = mi;
658 }
659
660 static void add_pair(struct merge_options *opt,
661                      struct name_entry *names,
662                      const char *pathname,
663                      unsigned side,
664                      unsigned is_add /* if false, is_delete */,
665                      unsigned match_mask,
666                      unsigned dir_rename_mask)
667 {
668         struct diff_filespec *one, *two;
669         struct rename_info *renames = &opt->priv->renames;
670         int names_idx = is_add ? side : 0;
671
672         if (!is_add) {
673                 unsigned content_relevant = (match_mask == 0);
674                 unsigned location_relevant = (dir_rename_mask == 0x07);
675
676                 if (content_relevant || location_relevant)
677                         strset_add(&renames->relevant_sources[side], pathname);
678         }
679
680         one = alloc_filespec(pathname);
681         two = alloc_filespec(pathname);
682         fill_filespec(is_add ? two : one,
683                       &names[names_idx].oid, 1, names[names_idx].mode);
684         diff_queue(&renames->pairs[side], one, two);
685 }
686
687 static void collect_rename_info(struct merge_options *opt,
688                                 struct name_entry *names,
689                                 const char *dirname,
690                                 const char *fullname,
691                                 unsigned filemask,
692                                 unsigned dirmask,
693                                 unsigned match_mask)
694 {
695         struct rename_info *renames = &opt->priv->renames;
696         unsigned side;
697
698         /*
699          * Update dir_rename_mask (determines ignore-rename-source validity)
700          *
701          * dir_rename_mask helps us keep track of when directory rename
702          * detection may be relevant.  Basically, whenver a directory is
703          * removed on one side of history, and a file is added to that
704          * directory on the other side of history, directory rename
705          * detection is relevant (meaning we have to detect renames for all
706          * files within that directory to deduce where the directory
707          * moved).  Also, whenever a directory needs directory rename
708          * detection, due to the "majority rules" choice for where to move
709          * it (see t6423 testcase 1f), we also need to detect renames for
710          * all files within subdirectories of that directory as well.
711          *
712          * Here we haven't looked at files within the directory yet, we are
713          * just looking at the directory itself.  So, if we aren't yet in
714          * a case where a parent directory needed directory rename detection
715          * (i.e. dir_rename_mask != 0x07), and if the directory was removed
716          * on one side of history, record the mask of the other side of
717          * history in dir_rename_mask.
718          */
719         if (renames->dir_rename_mask != 0x07 &&
720             (dirmask == 3 || dirmask == 5)) {
721                 /* simple sanity check */
722                 assert(renames->dir_rename_mask == 0 ||
723                        renames->dir_rename_mask == (dirmask & ~1));
724                 /* update dir_rename_mask; have it record mask of new side */
725                 renames->dir_rename_mask = (dirmask & ~1);
726         }
727
728         /* Update dirs_removed, as needed */
729         if (dirmask == 1 || dirmask == 3 || dirmask == 5) {
730                 /* absent_mask = 0x07 - dirmask; sides = absent_mask/2 */
731                 unsigned sides = (0x07 - dirmask)/2;
732                 if (sides & 1)
733                         strset_add(&renames->dirs_removed[1], fullname);
734                 if (sides & 2)
735                         strset_add(&renames->dirs_removed[2], fullname);
736         }
737
738         if (filemask == 0 || filemask == 7)
739                 return;
740
741         for (side = MERGE_SIDE1; side <= MERGE_SIDE2; ++side) {
742                 unsigned side_mask = (1 << side);
743
744                 /* Check for deletion on side */
745                 if ((filemask & 1) && !(filemask & side_mask))
746                         add_pair(opt, names, fullname, side, 0 /* delete */,
747                                  match_mask & filemask,
748                                  renames->dir_rename_mask);
749
750                 /* Check for addition on side */
751                 if (!(filemask & 1) && (filemask & side_mask))
752                         add_pair(opt, names, fullname, side, 1 /* add */,
753                                  match_mask & filemask,
754                                  renames->dir_rename_mask);
755         }
756 }
757
758 static int collect_merge_info_callback(int n,
759                                        unsigned long mask,
760                                        unsigned long dirmask,
761                                        struct name_entry *names,
762                                        struct traverse_info *info)
763 {
764         /*
765          * n is 3.  Always.
766          * common ancestor (mbase) has mask 1, and stored in index 0 of names
767          * head of side 1  (side1) has mask 2, and stored in index 1 of names
768          * head of side 2  (side2) has mask 4, and stored in index 2 of names
769          */
770         struct merge_options *opt = info->data;
771         struct merge_options_internal *opti = opt->priv;
772         struct rename_info *renames = &opt->priv->renames;
773         struct string_list_item pi;  /* Path Info */
774         struct conflict_info *ci; /* typed alias to pi.util (which is void*) */
775         struct name_entry *p;
776         size_t len;
777         char *fullpath;
778         const char *dirname = opti->current_dir_name;
779         unsigned prev_dir_rename_mask = renames->dir_rename_mask;
780         unsigned filemask = mask & ~dirmask;
781         unsigned match_mask = 0; /* will be updated below */
782         unsigned mbase_null = !(mask & 1);
783         unsigned side1_null = !(mask & 2);
784         unsigned side2_null = !(mask & 4);
785         unsigned side1_matches_mbase = (!side1_null && !mbase_null &&
786                                         names[0].mode == names[1].mode &&
787                                         oideq(&names[0].oid, &names[1].oid));
788         unsigned side2_matches_mbase = (!side2_null && !mbase_null &&
789                                         names[0].mode == names[2].mode &&
790                                         oideq(&names[0].oid, &names[2].oid));
791         unsigned sides_match = (!side1_null && !side2_null &&
792                                 names[1].mode == names[2].mode &&
793                                 oideq(&names[1].oid, &names[2].oid));
794
795         /*
796          * Note: When a path is a file on one side of history and a directory
797          * in another, we have a directory/file conflict.  In such cases, if
798          * the conflict doesn't resolve from renames and deletions, then we
799          * always leave directories where they are and move files out of the
800          * way.  Thus, while struct conflict_info has a df_conflict field to
801          * track such conflicts, we ignore that field for any directories at
802          * a path and only pay attention to it for files at the given path.
803          * The fact that we leave directories were they are also means that
804          * we do not need to worry about getting additional df_conflict
805          * information propagated from parent directories down to children
806          * (unlike, say traverse_trees_recursive() in unpack-trees.c, which
807          * sets a newinfo.df_conflicts field specifically to propagate it).
808          */
809         unsigned df_conflict = (filemask != 0) && (dirmask != 0);
810
811         /* n = 3 is a fundamental assumption. */
812         if (n != 3)
813                 BUG("Called collect_merge_info_callback wrong");
814
815         /*
816          * A bunch of sanity checks verifying that traverse_trees() calls
817          * us the way I expect.  Could just remove these at some point,
818          * though maybe they are helpful to future code readers.
819          */
820         assert(mbase_null == is_null_oid(&names[0].oid));
821         assert(side1_null == is_null_oid(&names[1].oid));
822         assert(side2_null == is_null_oid(&names[2].oid));
823         assert(!mbase_null || !side1_null || !side2_null);
824         assert(mask > 0 && mask < 8);
825
826         /* Determine match_mask */
827         if (side1_matches_mbase)
828                 match_mask = (side2_matches_mbase ? 7 : 3);
829         else if (side2_matches_mbase)
830                 match_mask = 5;
831         else if (sides_match)
832                 match_mask = 6;
833
834         /*
835          * Get the name of the relevant filepath, which we'll pass to
836          * setup_path_info() for tracking.
837          */
838         p = names;
839         while (!p->mode)
840                 p++;
841         len = traverse_path_len(info, p->pathlen);
842
843         /* +1 in both of the following lines to include the NUL byte */
844         fullpath = xmalloc(len + 1);
845         make_traverse_path(fullpath, len + 1, info, p->path, p->pathlen);
846
847         /*
848          * If mbase, side1, and side2 all match, we can resolve early.  Even
849          * if these are trees, there will be no renames or anything
850          * underneath.
851          */
852         if (side1_matches_mbase && side2_matches_mbase) {
853                 /* mbase, side1, & side2 all match; use mbase as resolution */
854                 setup_path_info(opt, &pi, dirname, info->pathlen, fullpath,
855                                 names, names+0, mbase_null, 0,
856                                 filemask, dirmask, 1);
857                 return mask;
858         }
859
860         /*
861          * Gather additional information used in rename detection.
862          */
863         collect_rename_info(opt, names, dirname, fullpath,
864                             filemask, dirmask, match_mask);
865
866         /*
867          * Record information about the path so we can resolve later in
868          * process_entries.
869          */
870         setup_path_info(opt, &pi, dirname, info->pathlen, fullpath,
871                         names, NULL, 0, df_conflict, filemask, dirmask, 0);
872
873         ci = pi.util;
874         VERIFY_CI(ci);
875         ci->match_mask = match_mask;
876
877         /* If dirmask, recurse into subdirectories */
878         if (dirmask) {
879                 struct traverse_info newinfo;
880                 struct tree_desc t[3];
881                 void *buf[3] = {NULL, NULL, NULL};
882                 const char *original_dir_name;
883                 int i, ret;
884
885                 ci->match_mask &= filemask;
886                 newinfo = *info;
887                 newinfo.prev = info;
888                 newinfo.name = p->path;
889                 newinfo.namelen = p->pathlen;
890                 newinfo.pathlen = st_add3(newinfo.pathlen, p->pathlen, 1);
891                 /*
892                  * If this directory we are about to recurse into cared about
893                  * its parent directory (the current directory) having a D/F
894                  * conflict, then we'd propagate the masks in this way:
895                  *    newinfo.df_conflicts |= (mask & ~dirmask);
896                  * But we don't worry about propagating D/F conflicts.  (See
897                  * comment near setting of local df_conflict variable near
898                  * the beginning of this function).
899                  */
900
901                 for (i = MERGE_BASE; i <= MERGE_SIDE2; i++) {
902                         if (i == 1 && side1_matches_mbase)
903                                 t[1] = t[0];
904                         else if (i == 2 && side2_matches_mbase)
905                                 t[2] = t[0];
906                         else if (i == 2 && sides_match)
907                                 t[2] = t[1];
908                         else {
909                                 const struct object_id *oid = NULL;
910                                 if (dirmask & 1)
911                                         oid = &names[i].oid;
912                                 buf[i] = fill_tree_descriptor(opt->repo,
913                                                               t + i, oid);
914                         }
915                         dirmask >>= 1;
916                 }
917
918                 original_dir_name = opti->current_dir_name;
919                 opti->current_dir_name = pi.string;
920                 if (renames->dir_rename_mask == 0 ||
921                     renames->dir_rename_mask == 0x07)
922                         ret = traverse_trees(NULL, 3, t, &newinfo);
923                 else
924                         ret = traverse_trees_wrapper(NULL, 3, t, &newinfo);
925                 opti->current_dir_name = original_dir_name;
926                 renames->dir_rename_mask = prev_dir_rename_mask;
927
928                 for (i = MERGE_BASE; i <= MERGE_SIDE2; i++)
929                         free(buf[i]);
930
931                 if (ret < 0)
932                         return -1;
933         }
934
935         return mask;
936 }
937
938 static int collect_merge_info(struct merge_options *opt,
939                               struct tree *merge_base,
940                               struct tree *side1,
941                               struct tree *side2)
942 {
943         int ret;
944         struct tree_desc t[3];
945         struct traverse_info info;
946
947         opt->priv->toplevel_dir = "";
948         opt->priv->current_dir_name = opt->priv->toplevel_dir;
949         setup_traverse_info(&info, opt->priv->toplevel_dir);
950         info.fn = collect_merge_info_callback;
951         info.data = opt;
952         info.show_all_errors = 1;
953
954         parse_tree(merge_base);
955         parse_tree(side1);
956         parse_tree(side2);
957         init_tree_desc(t + 0, merge_base->buffer, merge_base->size);
958         init_tree_desc(t + 1, side1->buffer, side1->size);
959         init_tree_desc(t + 2, side2->buffer, side2->size);
960
961         trace2_region_enter("merge", "traverse_trees", opt->repo);
962         ret = traverse_trees(NULL, 3, t, &info);
963         trace2_region_leave("merge", "traverse_trees", opt->repo);
964
965         return ret;
966 }
967
968 /*** Function Grouping: functions related to threeway content merges ***/
969
970 static int find_first_merges(struct repository *repo,
971                              const char *path,
972                              struct commit *a,
973                              struct commit *b,
974                              struct object_array *result)
975 {
976         int i, j;
977         struct object_array merges = OBJECT_ARRAY_INIT;
978         struct commit *commit;
979         int contains_another;
980
981         char merged_revision[GIT_MAX_HEXSZ + 2];
982         const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
983                                    "--all", merged_revision, NULL };
984         struct rev_info revs;
985         struct setup_revision_opt rev_opts;
986
987         memset(result, 0, sizeof(struct object_array));
988         memset(&rev_opts, 0, sizeof(rev_opts));
989
990         /* get all revisions that merge commit a */
991         xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
992                   oid_to_hex(&a->object.oid));
993         repo_init_revisions(repo, &revs, NULL);
994         rev_opts.submodule = path;
995         /* FIXME: can't handle linked worktrees in submodules yet */
996         revs.single_worktree = path != NULL;
997         setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
998
999         /* save all revisions from the above list that contain b */
1000         if (prepare_revision_walk(&revs))
1001                 die("revision walk setup failed");
1002         while ((commit = get_revision(&revs)) != NULL) {
1003                 struct object *o = &(commit->object);
1004                 if (in_merge_bases(b, commit))
1005                         add_object_array(o, NULL, &merges);
1006         }
1007         reset_revision_walk();
1008
1009         /* Now we've got all merges that contain a and b. Prune all
1010          * merges that contain another found merge and save them in
1011          * result.
1012          */
1013         for (i = 0; i < merges.nr; i++) {
1014                 struct commit *m1 = (struct commit *) merges.objects[i].item;
1015
1016                 contains_another = 0;
1017                 for (j = 0; j < merges.nr; j++) {
1018                         struct commit *m2 = (struct commit *) merges.objects[j].item;
1019                         if (i != j && in_merge_bases(m2, m1)) {
1020                                 contains_another = 1;
1021                                 break;
1022                         }
1023                 }
1024
1025                 if (!contains_another)
1026                         add_object_array(merges.objects[i].item, NULL, result);
1027         }
1028
1029         object_array_clear(&merges);
1030         return result->nr;
1031 }
1032
1033 static int merge_submodule(struct merge_options *opt,
1034                            const char *path,
1035                            const struct object_id *o,
1036                            const struct object_id *a,
1037                            const struct object_id *b,
1038                            struct object_id *result)
1039 {
1040         struct commit *commit_o, *commit_a, *commit_b;
1041         int parent_count;
1042         struct object_array merges;
1043         struct strbuf sb = STRBUF_INIT;
1044
1045         int i;
1046         int search = !opt->priv->call_depth;
1047
1048         /* store fallback answer in result in case we fail */
1049         oidcpy(result, opt->priv->call_depth ? o : a);
1050
1051         /* we can not handle deletion conflicts */
1052         if (is_null_oid(o))
1053                 return 0;
1054         if (is_null_oid(a))
1055                 return 0;
1056         if (is_null_oid(b))
1057                 return 0;
1058
1059         if (add_submodule_odb(path)) {
1060                 path_msg(opt, path, 0,
1061                          _("Failed to merge submodule %s (not checked out)"),
1062                          path);
1063                 return 0;
1064         }
1065
1066         if (!(commit_o = lookup_commit_reference(opt->repo, o)) ||
1067             !(commit_a = lookup_commit_reference(opt->repo, a)) ||
1068             !(commit_b = lookup_commit_reference(opt->repo, b))) {
1069                 path_msg(opt, path, 0,
1070                          _("Failed to merge submodule %s (commits not present)"),
1071                          path);
1072                 return 0;
1073         }
1074
1075         /* check whether both changes are forward */
1076         if (!in_merge_bases(commit_o, commit_a) ||
1077             !in_merge_bases(commit_o, commit_b)) {
1078                 path_msg(opt, path, 0,
1079                          _("Failed to merge submodule %s "
1080                            "(commits don't follow merge-base)"),
1081                          path);
1082                 return 0;
1083         }
1084
1085         /* Case #1: a is contained in b or vice versa */
1086         if (in_merge_bases(commit_a, commit_b)) {
1087                 oidcpy(result, b);
1088                 path_msg(opt, path, 1,
1089                          _("Note: Fast-forwarding submodule %s to %s"),
1090                          path, oid_to_hex(b));
1091                 return 1;
1092         }
1093         if (in_merge_bases(commit_b, commit_a)) {
1094                 oidcpy(result, a);
1095                 path_msg(opt, path, 1,
1096                          _("Note: Fast-forwarding submodule %s to %s"),
1097                          path, oid_to_hex(a));
1098                 return 1;
1099         }
1100
1101         /*
1102          * Case #2: There are one or more merges that contain a and b in
1103          * the submodule. If there is only one, then present it as a
1104          * suggestion to the user, but leave it marked unmerged so the
1105          * user needs to confirm the resolution.
1106          */
1107
1108         /* Skip the search if makes no sense to the calling context.  */
1109         if (!search)
1110                 return 0;
1111
1112         /* find commit which merges them */
1113         parent_count = find_first_merges(opt->repo, path, commit_a, commit_b,
1114                                          &merges);
1115         switch (parent_count) {
1116         case 0:
1117                 path_msg(opt, path, 0, _("Failed to merge submodule %s"), path);
1118                 break;
1119
1120         case 1:
1121                 format_commit(&sb, 4,
1122                               (struct commit *)merges.objects[0].item);
1123                 path_msg(opt, path, 0,
1124                          _("Failed to merge submodule %s, but a possible merge "
1125                            "resolution exists:\n%s\n"),
1126                          path, sb.buf);
1127                 path_msg(opt, path, 1,
1128                          _("If this is correct simply add it to the index "
1129                            "for example\n"
1130                            "by using:\n\n"
1131                            "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
1132                            "which will accept this suggestion.\n"),
1133                          oid_to_hex(&merges.objects[0].item->oid), path);
1134                 strbuf_release(&sb);
1135                 break;
1136         default:
1137                 for (i = 0; i < merges.nr; i++)
1138                         format_commit(&sb, 4,
1139                                       (struct commit *)merges.objects[i].item);
1140                 path_msg(opt, path, 0,
1141                          _("Failed to merge submodule %s, but multiple "
1142                            "possible merges exist:\n%s"), path, sb.buf);
1143                 strbuf_release(&sb);
1144         }
1145
1146         object_array_clear(&merges);
1147         return 0;
1148 }
1149
1150 static int merge_3way(struct merge_options *opt,
1151                       const char *path,
1152                       const struct object_id *o,
1153                       const struct object_id *a,
1154                       const struct object_id *b,
1155                       const char *pathnames[3],
1156                       const int extra_marker_size,
1157                       mmbuffer_t *result_buf)
1158 {
1159         mmfile_t orig, src1, src2;
1160         struct ll_merge_options ll_opts = {0};
1161         char *base, *name1, *name2;
1162         int merge_status;
1163
1164         ll_opts.renormalize = opt->renormalize;
1165         ll_opts.extra_marker_size = extra_marker_size;
1166         ll_opts.xdl_opts = opt->xdl_opts;
1167
1168         if (opt->priv->call_depth) {
1169                 ll_opts.virtual_ancestor = 1;
1170                 ll_opts.variant = 0;
1171         } else {
1172                 switch (opt->recursive_variant) {
1173                 case MERGE_VARIANT_OURS:
1174                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
1175                         break;
1176                 case MERGE_VARIANT_THEIRS:
1177                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
1178                         break;
1179                 default:
1180                         ll_opts.variant = 0;
1181                         break;
1182                 }
1183         }
1184
1185         assert(pathnames[0] && pathnames[1] && pathnames[2] && opt->ancestor);
1186         if (pathnames[0] == pathnames[1] && pathnames[1] == pathnames[2]) {
1187                 base  = mkpathdup("%s", opt->ancestor);
1188                 name1 = mkpathdup("%s", opt->branch1);
1189                 name2 = mkpathdup("%s", opt->branch2);
1190         } else {
1191                 base  = mkpathdup("%s:%s", opt->ancestor, pathnames[0]);
1192                 name1 = mkpathdup("%s:%s", opt->branch1,  pathnames[1]);
1193                 name2 = mkpathdup("%s:%s", opt->branch2,  pathnames[2]);
1194         }
1195
1196         read_mmblob(&orig, o);
1197         read_mmblob(&src1, a);
1198         read_mmblob(&src2, b);
1199
1200         merge_status = ll_merge(result_buf, path, &orig, base,
1201                                 &src1, name1, &src2, name2,
1202                                 opt->repo->index, &ll_opts);
1203
1204         free(base);
1205         free(name1);
1206         free(name2);
1207         free(orig.ptr);
1208         free(src1.ptr);
1209         free(src2.ptr);
1210         return merge_status;
1211 }
1212
1213 static int handle_content_merge(struct merge_options *opt,
1214                                 const char *path,
1215                                 const struct version_info *o,
1216                                 const struct version_info *a,
1217                                 const struct version_info *b,
1218                                 const char *pathnames[3],
1219                                 const int extra_marker_size,
1220                                 struct version_info *result)
1221 {
1222         /*
1223          * path is the target location where we want to put the file, and
1224          * is used to determine any normalization rules in ll_merge.
1225          *
1226          * The normal case is that path and all entries in pathnames are
1227          * identical, though renames can affect which path we got one of
1228          * the three blobs to merge on various sides of history.
1229          *
1230          * extra_marker_size is the amount to extend conflict markers in
1231          * ll_merge; this is neeed if we have content merges of content
1232          * merges, which happens for example with rename/rename(2to1) and
1233          * rename/add conflicts.
1234          */
1235         unsigned clean = 1;
1236
1237         /*
1238          * handle_content_merge() needs both files to be of the same type, i.e.
1239          * both files OR both submodules OR both symlinks.  Conflicting types
1240          * needs to be handled elsewhere.
1241          */
1242         assert((S_IFMT & a->mode) == (S_IFMT & b->mode));
1243
1244         /* Merge modes */
1245         if (a->mode == b->mode || a->mode == o->mode)
1246                 result->mode = b->mode;
1247         else {
1248                 /* must be the 100644/100755 case */
1249                 assert(S_ISREG(a->mode));
1250                 result->mode = a->mode;
1251                 clean = (b->mode == o->mode);
1252                 /*
1253                  * FIXME: If opt->priv->call_depth && !clean, then we really
1254                  * should not make result->mode match either a->mode or
1255                  * b->mode; that causes t6036 "check conflicting mode for
1256                  * regular file" to fail.  It would be best to use some other
1257                  * mode, but we'll confuse all kinds of stuff if we use one
1258                  * where S_ISREG(result->mode) isn't true, and if we use
1259                  * something like 0100666, then tree-walk.c's calls to
1260                  * canon_mode() will just normalize that to 100644 for us and
1261                  * thus not solve anything.
1262                  *
1263                  * Figure out if there's some kind of way we can work around
1264                  * this...
1265                  */
1266         }
1267
1268         /*
1269          * Trivial oid merge.
1270          *
1271          * Note: While one might assume that the next four lines would
1272          * be unnecessary due to the fact that match_mask is often
1273          * setup and already handled, renames don't always take care
1274          * of that.
1275          */
1276         if (oideq(&a->oid, &b->oid) || oideq(&a->oid, &o->oid))
1277                 oidcpy(&result->oid, &b->oid);
1278         else if (oideq(&b->oid, &o->oid))
1279                 oidcpy(&result->oid, &a->oid);
1280
1281         /* Remaining rules depend on file vs. submodule vs. symlink. */
1282         else if (S_ISREG(a->mode)) {
1283                 mmbuffer_t result_buf;
1284                 int ret = 0, merge_status;
1285                 int two_way;
1286
1287                 /*
1288                  * If 'o' is different type, treat it as null so we do a
1289                  * two-way merge.
1290                  */
1291                 two_way = ((S_IFMT & o->mode) != (S_IFMT & a->mode));
1292
1293                 merge_status = merge_3way(opt, path,
1294                                           two_way ? &null_oid : &o->oid,
1295                                           &a->oid, &b->oid,
1296                                           pathnames, extra_marker_size,
1297                                           &result_buf);
1298
1299                 if ((merge_status < 0) || !result_buf.ptr)
1300                         ret = err(opt, _("Failed to execute internal merge"));
1301
1302                 if (!ret &&
1303                     write_object_file(result_buf.ptr, result_buf.size,
1304                                       blob_type, &result->oid))
1305                         ret = err(opt, _("Unable to add %s to database"),
1306                                   path);
1307
1308                 free(result_buf.ptr);
1309                 if (ret)
1310                         return -1;
1311                 clean &= (merge_status == 0);
1312                 path_msg(opt, path, 1, _("Auto-merging %s"), path);
1313         } else if (S_ISGITLINK(a->mode)) {
1314                 int two_way = ((S_IFMT & o->mode) != (S_IFMT & a->mode));
1315                 clean = merge_submodule(opt, pathnames[0],
1316                                         two_way ? &null_oid : &o->oid,
1317                                         &a->oid, &b->oid, &result->oid);
1318                 if (opt->priv->call_depth && two_way && !clean) {
1319                         result->mode = o->mode;
1320                         oidcpy(&result->oid, &o->oid);
1321                 }
1322         } else if (S_ISLNK(a->mode)) {
1323                 if (opt->priv->call_depth) {
1324                         clean = 0;
1325                         result->mode = o->mode;
1326                         oidcpy(&result->oid, &o->oid);
1327                 } else {
1328                         switch (opt->recursive_variant) {
1329                         case MERGE_VARIANT_NORMAL:
1330                                 clean = 0;
1331                                 oidcpy(&result->oid, &a->oid);
1332                                 break;
1333                         case MERGE_VARIANT_OURS:
1334                                 oidcpy(&result->oid, &a->oid);
1335                                 break;
1336                         case MERGE_VARIANT_THEIRS:
1337                                 oidcpy(&result->oid, &b->oid);
1338                                 break;
1339                         }
1340                 }
1341         } else
1342                 BUG("unsupported object type in the tree: %06o for %s",
1343                     a->mode, path);
1344
1345         return clean;
1346 }
1347
1348 /*** Function Grouping: functions related to detect_and_process_renames(), ***
1349  *** which are split into directory and regular rename detection sections. ***/
1350
1351 /*** Function Grouping: functions related to directory rename detection ***/
1352
1353 struct collision_info {
1354         struct string_list source_files;
1355         unsigned reported_already:1;
1356 };
1357
1358 /*
1359  * Return a new string that replaces the beginning portion (which matches
1360  * rename_info->key), with rename_info->util.new_dir.  In perl-speak:
1361  *   new_path_name = (old_path =~ s/rename_info->key/rename_info->value/);
1362  * NOTE:
1363  *   Caller must ensure that old_path starts with rename_info->key + '/'.
1364  */
1365 static char *apply_dir_rename(struct strmap_entry *rename_info,
1366                               const char *old_path)
1367 {
1368         struct strbuf new_path = STRBUF_INIT;
1369         const char *old_dir = rename_info->key;
1370         const char *new_dir = rename_info->value;
1371         int oldlen, newlen, new_dir_len;
1372
1373         oldlen = strlen(old_dir);
1374         if (*new_dir == '\0')
1375                 /*
1376                  * If someone renamed/merged a subdirectory into the root
1377                  * directory (e.g. 'some/subdir' -> ''), then we want to
1378                  * avoid returning
1379                  *     '' + '/filename'
1380                  * as the rename; we need to make old_path + oldlen advance
1381                  * past the '/' character.
1382                  */
1383                 oldlen++;
1384         new_dir_len = strlen(new_dir);
1385         newlen = new_dir_len + (strlen(old_path) - oldlen) + 1;
1386         strbuf_grow(&new_path, newlen);
1387         strbuf_add(&new_path, new_dir, new_dir_len);
1388         strbuf_addstr(&new_path, &old_path[oldlen]);
1389
1390         return strbuf_detach(&new_path, NULL);
1391 }
1392
1393 static int path_in_way(struct strmap *paths, const char *path, unsigned side_mask)
1394 {
1395         struct merged_info *mi = strmap_get(paths, path);
1396         struct conflict_info *ci;
1397         if (!mi)
1398                 return 0;
1399         INITIALIZE_CI(ci, mi);
1400         return mi->clean || (side_mask & (ci->filemask | ci->dirmask));
1401 }
1402
1403 /*
1404  * See if there is a directory rename for path, and if there are any file
1405  * level conflicts on the given side for the renamed location.  If there is
1406  * a rename and there are no conflicts, return the new name.  Otherwise,
1407  * return NULL.
1408  */
1409 static char *handle_path_level_conflicts(struct merge_options *opt,
1410                                          const char *path,
1411                                          unsigned side_index,
1412                                          struct strmap_entry *rename_info,
1413                                          struct strmap *collisions)
1414 {
1415         char *new_path = NULL;
1416         struct collision_info *c_info;
1417         int clean = 1;
1418         struct strbuf collision_paths = STRBUF_INIT;
1419
1420         /*
1421          * entry has the mapping of old directory name to new directory name
1422          * that we want to apply to path.
1423          */
1424         new_path = apply_dir_rename(rename_info, path);
1425         if (!new_path)
1426                 BUG("Failed to apply directory rename!");
1427
1428         /*
1429          * The caller needs to have ensured that it has pre-populated
1430          * collisions with all paths that map to new_path.  Do a quick check
1431          * to ensure that's the case.
1432          */
1433         c_info = strmap_get(collisions, new_path);
1434         if (c_info == NULL)
1435                 BUG("c_info is NULL");
1436
1437         /*
1438          * Check for one-sided add/add/.../add conflicts, i.e.
1439          * where implicit renames from the other side doing
1440          * directory rename(s) can affect this side of history
1441          * to put multiple paths into the same location.  Warn
1442          * and bail on directory renames for such paths.
1443          */
1444         if (c_info->reported_already) {
1445                 clean = 0;
1446         } else if (path_in_way(&opt->priv->paths, new_path, 1 << side_index)) {
1447                 c_info->reported_already = 1;
1448                 strbuf_add_separated_string_list(&collision_paths, ", ",
1449                                                  &c_info->source_files);
1450                 path_msg(opt, new_path, 0,
1451                          _("CONFLICT (implicit dir rename): Existing file/dir "
1452                            "at %s in the way of implicit directory rename(s) "
1453                            "putting the following path(s) there: %s."),
1454                        new_path, collision_paths.buf);
1455                 clean = 0;
1456         } else if (c_info->source_files.nr > 1) {
1457                 c_info->reported_already = 1;
1458                 strbuf_add_separated_string_list(&collision_paths, ", ",
1459                                                  &c_info->source_files);
1460                 path_msg(opt, new_path, 0,
1461                          _("CONFLICT (implicit dir rename): Cannot map more "
1462                            "than one path to %s; implicit directory renames "
1463                            "tried to put these paths there: %s"),
1464                        new_path, collision_paths.buf);
1465                 clean = 0;
1466         }
1467
1468         /* Free memory we no longer need */
1469         strbuf_release(&collision_paths);
1470         if (!clean && new_path) {
1471                 free(new_path);
1472                 return NULL;
1473         }
1474
1475         return new_path;
1476 }
1477
1478 static void get_provisional_directory_renames(struct merge_options *opt,
1479                                               unsigned side,
1480                                               int *clean)
1481 {
1482         struct hashmap_iter iter;
1483         struct strmap_entry *entry;
1484         struct rename_info *renames = &opt->priv->renames;
1485
1486         /*
1487          * Collapse
1488          *    dir_rename_count: old_directory -> {new_directory -> count}
1489          * down to
1490          *    dir_renames: old_directory -> best_new_directory
1491          * where best_new_directory is the one with the unique highest count.
1492          */
1493         strmap_for_each_entry(&renames->dir_rename_count[side], &iter, entry) {
1494                 const char *source_dir = entry->key;
1495                 struct strintmap *counts = entry->value;
1496                 struct hashmap_iter count_iter;
1497                 struct strmap_entry *count_entry;
1498                 int max = 0;
1499                 int bad_max = 0;
1500                 const char *best = NULL;
1501
1502                 strintmap_for_each_entry(counts, &count_iter, count_entry) {
1503                         const char *target_dir = count_entry->key;
1504                         intptr_t count = (intptr_t)count_entry->value;
1505
1506                         if (count == max)
1507                                 bad_max = max;
1508                         else if (count > max) {
1509                                 max = count;
1510                                 best = target_dir;
1511                         }
1512                 }
1513
1514                 if (bad_max == max) {
1515                         path_msg(opt, source_dir, 0,
1516                                _("CONFLICT (directory rename split): "
1517                                  "Unclear where to rename %s to; it was "
1518                                  "renamed to multiple other directories, with "
1519                                  "no destination getting a majority of the "
1520                                  "files."),
1521                                source_dir);
1522                         /*
1523                          * We should mark this as unclean IF something attempts
1524                          * to use this rename.  We do not yet have the logic
1525                          * in place to detect if this directory rename is being
1526                          * used, and optimizations that reduce the number of
1527                          * renames cause this to falsely trigger.  For now,
1528                          * just disable it, causing t6423 testcase 2a to break.
1529                          * We'll later fix the detection, and when we do we
1530                          * will re-enable setting *clean to 0 (and thereby fix
1531                          * t6423 testcase 2a).
1532                          */
1533                         /*   *clean = 0;   */
1534                 } else {
1535                         strmap_put(&renames->dir_renames[side],
1536                                    source_dir, (void*)best);
1537                 }
1538         }
1539 }
1540
1541 static void handle_directory_level_conflicts(struct merge_options *opt)
1542 {
1543         struct hashmap_iter iter;
1544         struct strmap_entry *entry;
1545         struct string_list duplicated = STRING_LIST_INIT_NODUP;
1546         struct rename_info *renames = &opt->priv->renames;
1547         struct strmap *side1_dir_renames = &renames->dir_renames[MERGE_SIDE1];
1548         struct strmap *side2_dir_renames = &renames->dir_renames[MERGE_SIDE2];
1549         int i;
1550
1551         strmap_for_each_entry(side1_dir_renames, &iter, entry) {
1552                 if (strmap_contains(side2_dir_renames, entry->key))
1553                         string_list_append(&duplicated, entry->key);
1554         }
1555
1556         for (i = 0; i < duplicated.nr; i++) {
1557                 strmap_remove(side1_dir_renames, duplicated.items[i].string, 0);
1558                 strmap_remove(side2_dir_renames, duplicated.items[i].string, 0);
1559         }
1560         string_list_clear(&duplicated, 0);
1561 }
1562
1563 static struct strmap_entry *check_dir_renamed(const char *path,
1564                                               struct strmap *dir_renames)
1565 {
1566         char *temp = xstrdup(path);
1567         char *end;
1568         struct strmap_entry *e = NULL;
1569
1570         while ((end = strrchr(temp, '/'))) {
1571                 *end = '\0';
1572                 e = strmap_get_entry(dir_renames, temp);
1573                 if (e)
1574                         break;
1575         }
1576         free(temp);
1577         return e;
1578 }
1579
1580 static void compute_collisions(struct strmap *collisions,
1581                                struct strmap *dir_renames,
1582                                struct diff_queue_struct *pairs)
1583 {
1584         int i;
1585
1586         strmap_init_with_options(collisions, NULL, 0);
1587         if (strmap_empty(dir_renames))
1588                 return;
1589
1590         /*
1591          * Multiple files can be mapped to the same path due to directory
1592          * renames done by the other side of history.  Since that other
1593          * side of history could have merged multiple directories into one,
1594          * if our side of history added the same file basename to each of
1595          * those directories, then all N of them would get implicitly
1596          * renamed by the directory rename detection into the same path,
1597          * and we'd get an add/add/.../add conflict, and all those adds
1598          * from *this* side of history.  This is not representable in the
1599          * index, and users aren't going to easily be able to make sense of
1600          * it.  So we need to provide a good warning about what's
1601          * happening, and fall back to no-directory-rename detection
1602          * behavior for those paths.
1603          *
1604          * See testcases 9e and all of section 5 from t6043 for examples.
1605          */
1606         for (i = 0; i < pairs->nr; ++i) {
1607                 struct strmap_entry *rename_info;
1608                 struct collision_info *collision_info;
1609                 char *new_path;
1610                 struct diff_filepair *pair = pairs->queue[i];
1611
1612                 if (pair->status != 'A' && pair->status != 'R')
1613                         continue;
1614                 rename_info = check_dir_renamed(pair->two->path, dir_renames);
1615                 if (!rename_info)
1616                         continue;
1617
1618                 new_path = apply_dir_rename(rename_info, pair->two->path);
1619                 assert(new_path);
1620                 collision_info = strmap_get(collisions, new_path);
1621                 if (collision_info) {
1622                         free(new_path);
1623                 } else {
1624                         CALLOC_ARRAY(collision_info, 1);
1625                         string_list_init(&collision_info->source_files, 0);
1626                         strmap_put(collisions, new_path, collision_info);
1627                 }
1628                 string_list_insert(&collision_info->source_files,
1629                                    pair->two->path);
1630         }
1631 }
1632
1633 static char *check_for_directory_rename(struct merge_options *opt,
1634                                         const char *path,
1635                                         unsigned side_index,
1636                                         struct strmap *dir_renames,
1637                                         struct strmap *dir_rename_exclusions,
1638                                         struct strmap *collisions,
1639                                         int *clean_merge)
1640 {
1641         char *new_path = NULL;
1642         struct strmap_entry *rename_info;
1643         struct strmap_entry *otherinfo = NULL;
1644         const char *new_dir;
1645
1646         if (strmap_empty(dir_renames))
1647                 return new_path;
1648         rename_info = check_dir_renamed(path, dir_renames);
1649         if (!rename_info)
1650                 return new_path;
1651         /* old_dir = rename_info->key; */
1652         new_dir = rename_info->value;
1653
1654         /*
1655          * This next part is a little weird.  We do not want to do an
1656          * implicit rename into a directory we renamed on our side, because
1657          * that will result in a spurious rename/rename(1to2) conflict.  An
1658          * example:
1659          *   Base commit: dumbdir/afile, otherdir/bfile
1660          *   Side 1:      smrtdir/afile, otherdir/bfile
1661          *   Side 2:      dumbdir/afile, dumbdir/bfile
1662          * Here, while working on Side 1, we could notice that otherdir was
1663          * renamed/merged to dumbdir, and change the diff_filepair for
1664          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
1665          * 2 will notice the rename from dumbdir to smrtdir, and do the
1666          * transitive rename to move it from dumbdir/bfile to
1667          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
1668          * smrtdir, a rename/rename(1to2) conflict.  We really just want
1669          * the file to end up in smrtdir.  And the way to achieve that is
1670          * to not let Side1 do the rename to dumbdir, since we know that is
1671          * the source of one of our directory renames.
1672          *
1673          * That's why otherinfo and dir_rename_exclusions is here.
1674          *
1675          * As it turns out, this also prevents N-way transient rename
1676          * confusion; See testcases 9c and 9d of t6043.
1677          */
1678         otherinfo = strmap_get_entry(dir_rename_exclusions, new_dir);
1679         if (otherinfo) {
1680                 path_msg(opt, rename_info->key, 1,
1681                          _("WARNING: Avoiding applying %s -> %s rename "
1682                            "to %s, because %s itself was renamed."),
1683                          rename_info->key, new_dir, path, new_dir);
1684                 return NULL;
1685         }
1686
1687         new_path = handle_path_level_conflicts(opt, path, side_index,
1688                                                rename_info, collisions);
1689         *clean_merge &= (new_path != NULL);
1690
1691         return new_path;
1692 }
1693
1694 static void apply_directory_rename_modifications(struct merge_options *opt,
1695                                                  struct diff_filepair *pair,
1696                                                  char *new_path)
1697 {
1698         /*
1699          * The basic idea is to get the conflict_info from opt->priv->paths
1700          * at old path, and insert it into new_path; basically just this:
1701          *     ci = strmap_get(&opt->priv->paths, old_path);
1702          *     strmap_remove(&opt->priv->paths, old_path, 0);
1703          *     strmap_put(&opt->priv->paths, new_path, ci);
1704          * However, there are some factors complicating this:
1705          *     - opt->priv->paths may already have an entry at new_path
1706          *     - Each ci tracks its containing directory, so we need to
1707          *       update that
1708          *     - If another ci has the same containing directory, then
1709          *       the two char*'s MUST point to the same location.  See the
1710          *       comment in struct merged_info.  strcmp equality is not
1711          *       enough; we need pointer equality.
1712          *     - opt->priv->paths must hold the parent directories of any
1713          *       entries that are added.  So, if this directory rename
1714          *       causes entirely new directories, we must recursively add
1715          *       parent directories.
1716          *     - For each parent directory added to opt->priv->paths, we
1717          *       also need to get its parent directory stored in its
1718          *       conflict_info->merged.directory_name with all the same
1719          *       requirements about pointer equality.
1720          */
1721         struct string_list dirs_to_insert = STRING_LIST_INIT_NODUP;
1722         struct conflict_info *ci, *new_ci;
1723         struct strmap_entry *entry;
1724         const char *branch_with_new_path, *branch_with_dir_rename;
1725         const char *old_path = pair->two->path;
1726         const char *parent_name;
1727         const char *cur_path;
1728         int i, len;
1729
1730         entry = strmap_get_entry(&opt->priv->paths, old_path);
1731         old_path = entry->key;
1732         ci = entry->value;
1733         VERIFY_CI(ci);
1734
1735         /* Find parent directories missing from opt->priv->paths */
1736         cur_path = new_path;
1737         while (1) {
1738                 /* Find the parent directory of cur_path */
1739                 char *last_slash = strrchr(cur_path, '/');
1740                 if (last_slash) {
1741                         parent_name = xstrndup(cur_path, last_slash - cur_path);
1742                 } else {
1743                         parent_name = opt->priv->toplevel_dir;
1744                         break;
1745                 }
1746
1747                 /* Look it up in opt->priv->paths */
1748                 entry = strmap_get_entry(&opt->priv->paths, parent_name);
1749                 if (entry) {
1750                         free((char*)parent_name);
1751                         parent_name = entry->key; /* reuse known pointer */
1752                         break;
1753                 }
1754
1755                 /* Record this is one of the directories we need to insert */
1756                 string_list_append(&dirs_to_insert, parent_name);
1757                 cur_path = parent_name;
1758         }
1759
1760         /* Traverse dirs_to_insert and insert them into opt->priv->paths */
1761         for (i = dirs_to_insert.nr-1; i >= 0; --i) {
1762                 struct conflict_info *dir_ci;
1763                 char *cur_dir = dirs_to_insert.items[i].string;
1764
1765                 CALLOC_ARRAY(dir_ci, 1);
1766
1767                 dir_ci->merged.directory_name = parent_name;
1768                 len = strlen(parent_name);
1769                 /* len+1 because of trailing '/' character */
1770                 dir_ci->merged.basename_offset = (len > 0 ? len+1 : len);
1771                 dir_ci->dirmask = ci->filemask;
1772                 strmap_put(&opt->priv->paths, cur_dir, dir_ci);
1773
1774                 parent_name = cur_dir;
1775         }
1776
1777         /*
1778          * We are removing old_path from opt->priv->paths.  old_path also will
1779          * eventually need to be freed, but it may still be used by e.g.
1780          * ci->pathnames.  So, store it in another string-list for now.
1781          */
1782         string_list_append(&opt->priv->paths_to_free, old_path);
1783
1784         assert(ci->filemask == 2 || ci->filemask == 4);
1785         assert(ci->dirmask == 0);
1786         strmap_remove(&opt->priv->paths, old_path, 0);
1787
1788         branch_with_new_path   = (ci->filemask == 2) ? opt->branch1 : opt->branch2;
1789         branch_with_dir_rename = (ci->filemask == 2) ? opt->branch2 : opt->branch1;
1790
1791         /* Now, finally update ci and stick it into opt->priv->paths */
1792         ci->merged.directory_name = parent_name;
1793         len = strlen(parent_name);
1794         ci->merged.basename_offset = (len > 0 ? len+1 : len);
1795         new_ci = strmap_get(&opt->priv->paths, new_path);
1796         if (!new_ci) {
1797                 /* Place ci back into opt->priv->paths, but at new_path */
1798                 strmap_put(&opt->priv->paths, new_path, ci);
1799         } else {
1800                 int index;
1801
1802                 /* A few sanity checks */
1803                 VERIFY_CI(new_ci);
1804                 assert(ci->filemask == 2 || ci->filemask == 4);
1805                 assert((new_ci->filemask & ci->filemask) == 0);
1806                 assert(!new_ci->merged.clean);
1807
1808                 /* Copy stuff from ci into new_ci */
1809                 new_ci->filemask |= ci->filemask;
1810                 if (new_ci->dirmask)
1811                         new_ci->df_conflict = 1;
1812                 index = (ci->filemask >> 1);
1813                 new_ci->pathnames[index] = ci->pathnames[index];
1814                 new_ci->stages[index].mode = ci->stages[index].mode;
1815                 oidcpy(&new_ci->stages[index].oid, &ci->stages[index].oid);
1816
1817                 free(ci);
1818                 ci = new_ci;
1819         }
1820
1821         if (opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_TRUE) {
1822                 /* Notify user of updated path */
1823                 if (pair->status == 'A')
1824                         path_msg(opt, new_path, 1,
1825                                  _("Path updated: %s added in %s inside a "
1826                                    "directory that was renamed in %s; moving "
1827                                    "it to %s."),
1828                                  old_path, branch_with_new_path,
1829                                  branch_with_dir_rename, new_path);
1830                 else
1831                         path_msg(opt, new_path, 1,
1832                                  _("Path updated: %s renamed to %s in %s, "
1833                                    "inside a directory that was renamed in %s; "
1834                                    "moving it to %s."),
1835                                  pair->one->path, old_path, branch_with_new_path,
1836                                  branch_with_dir_rename, new_path);
1837         } else {
1838                 /*
1839                  * opt->detect_directory_renames has the value
1840                  * MERGE_DIRECTORY_RENAMES_CONFLICT, so mark these as conflicts.
1841                  */
1842                 ci->path_conflict = 1;
1843                 if (pair->status == 'A')
1844                         path_msg(opt, new_path, 0,
1845                                  _("CONFLICT (file location): %s added in %s "
1846                                    "inside a directory that was renamed in %s, "
1847                                    "suggesting it should perhaps be moved to "
1848                                    "%s."),
1849                                  old_path, branch_with_new_path,
1850                                  branch_with_dir_rename, new_path);
1851                 else
1852                         path_msg(opt, new_path, 0,
1853                                  _("CONFLICT (file location): %s renamed to %s "
1854                                    "in %s, inside a directory that was renamed "
1855                                    "in %s, suggesting it should perhaps be "
1856                                    "moved to %s."),
1857                                  pair->one->path, old_path, branch_with_new_path,
1858                                  branch_with_dir_rename, new_path);
1859         }
1860
1861         /*
1862          * Finally, record the new location.
1863          */
1864         pair->two->path = new_path;
1865 }
1866
1867 /*** Function Grouping: functions related to regular rename detection ***/
1868
1869 static int process_renames(struct merge_options *opt,
1870                            struct diff_queue_struct *renames)
1871 {
1872         int clean_merge = 1, i;
1873
1874         for (i = 0; i < renames->nr; ++i) {
1875                 const char *oldpath = NULL, *newpath;
1876                 struct diff_filepair *pair = renames->queue[i];
1877                 struct conflict_info *oldinfo = NULL, *newinfo = NULL;
1878                 struct strmap_entry *old_ent, *new_ent;
1879                 unsigned int old_sidemask;
1880                 int target_index, other_source_index;
1881                 int source_deleted, collision, type_changed;
1882                 const char *rename_branch = NULL, *delete_branch = NULL;
1883
1884                 old_ent = strmap_get_entry(&opt->priv->paths, pair->one->path);
1885                 new_ent = strmap_get_entry(&opt->priv->paths, pair->two->path);
1886                 if (old_ent) {
1887                         oldpath = old_ent->key;
1888                         oldinfo = old_ent->value;
1889                 }
1890                 newpath = pair->two->path;
1891                 if (new_ent) {
1892                         newpath = new_ent->key;
1893                         newinfo = new_ent->value;
1894                 }
1895
1896                 /*
1897                  * If pair->one->path isn't in opt->priv->paths, that means
1898                  * that either directory rename detection removed that
1899                  * path, or a parent directory of oldpath was resolved and
1900                  * we don't even need the rename; in either case, we can
1901                  * skip it.  If oldinfo->merged.clean, then the other side
1902                  * of history had no changes to oldpath and we don't need
1903                  * the rename and can skip it.
1904                  */
1905                 if (!oldinfo || oldinfo->merged.clean)
1906                         continue;
1907
1908                 /*
1909                  * diff_filepairs have copies of pathnames, thus we have to
1910                  * use standard 'strcmp()' (negated) instead of '=='.
1911                  */
1912                 if (i + 1 < renames->nr &&
1913                     !strcmp(oldpath, renames->queue[i+1]->one->path)) {
1914                         /* Handle rename/rename(1to2) or rename/rename(1to1) */
1915                         const char *pathnames[3];
1916                         struct version_info merged;
1917                         struct conflict_info *base, *side1, *side2;
1918                         unsigned was_binary_blob = 0;
1919
1920                         pathnames[0] = oldpath;
1921                         pathnames[1] = newpath;
1922                         pathnames[2] = renames->queue[i+1]->two->path;
1923
1924                         base = strmap_get(&opt->priv->paths, pathnames[0]);
1925                         side1 = strmap_get(&opt->priv->paths, pathnames[1]);
1926                         side2 = strmap_get(&opt->priv->paths, pathnames[2]);
1927
1928                         VERIFY_CI(base);
1929                         VERIFY_CI(side1);
1930                         VERIFY_CI(side2);
1931
1932                         if (!strcmp(pathnames[1], pathnames[2])) {
1933                                 /* Both sides renamed the same way */
1934                                 assert(side1 == side2);
1935                                 memcpy(&side1->stages[0], &base->stages[0],
1936                                        sizeof(merged));
1937                                 side1->filemask |= (1 << MERGE_BASE);
1938                                 /* Mark base as resolved by removal */
1939                                 base->merged.is_null = 1;
1940                                 base->merged.clean = 1;
1941
1942                                 /* We handled both renames, i.e. i+1 handled */
1943                                 i++;
1944                                 /* Move to next rename */
1945                                 continue;
1946                         }
1947
1948                         /* This is a rename/rename(1to2) */
1949                         clean_merge = handle_content_merge(opt,
1950                                                            pair->one->path,
1951                                                            &base->stages[0],
1952                                                            &side1->stages[1],
1953                                                            &side2->stages[2],
1954                                                            pathnames,
1955                                                            1 + 2 * opt->priv->call_depth,
1956                                                            &merged);
1957                         if (!clean_merge &&
1958                             merged.mode == side1->stages[1].mode &&
1959                             oideq(&merged.oid, &side1->stages[1].oid))
1960                                 was_binary_blob = 1;
1961                         memcpy(&side1->stages[1], &merged, sizeof(merged));
1962                         if (was_binary_blob) {
1963                                 /*
1964                                  * Getting here means we were attempting to
1965                                  * merge a binary blob.
1966                                  *
1967                                  * Since we can't merge binaries,
1968                                  * handle_content_merge() just takes one
1969                                  * side.  But we don't want to copy the
1970                                  * contents of one side to both paths.  We
1971                                  * used the contents of side1 above for
1972                                  * side1->stages, let's use the contents of
1973                                  * side2 for side2->stages below.
1974                                  */
1975                                 oidcpy(&merged.oid, &side2->stages[2].oid);
1976                                 merged.mode = side2->stages[2].mode;
1977                         }
1978                         memcpy(&side2->stages[2], &merged, sizeof(merged));
1979
1980                         side1->path_conflict = 1;
1981                         side2->path_conflict = 1;
1982                         /*
1983                          * TODO: For renames we normally remove the path at the
1984                          * old name.  It would thus seem consistent to do the
1985                          * same for rename/rename(1to2) cases, but we haven't
1986                          * done so traditionally and a number of the regression
1987                          * tests now encode an expectation that the file is
1988                          * left there at stage 1.  If we ever decide to change
1989                          * this, add the following two lines here:
1990                          *    base->merged.is_null = 1;
1991                          *    base->merged.clean = 1;
1992                          * and remove the setting of base->path_conflict to 1.
1993                          */
1994                         base->path_conflict = 1;
1995                         path_msg(opt, oldpath, 0,
1996                                  _("CONFLICT (rename/rename): %s renamed to "
1997                                    "%s in %s and to %s in %s."),
1998                                  pathnames[0],
1999                                  pathnames[1], opt->branch1,
2000                                  pathnames[2], opt->branch2);
2001
2002                         i++; /* We handled both renames, i.e. i+1 handled */
2003                         continue;
2004                 }
2005
2006                 VERIFY_CI(oldinfo);
2007                 VERIFY_CI(newinfo);
2008                 target_index = pair->score; /* from collect_renames() */
2009                 assert(target_index == 1 || target_index == 2);
2010                 other_source_index = 3 - target_index;
2011                 old_sidemask = (1 << other_source_index); /* 2 or 4 */
2012                 source_deleted = (oldinfo->filemask == 1);
2013                 collision = ((newinfo->filemask & old_sidemask) != 0);
2014                 type_changed = !source_deleted &&
2015                         (S_ISREG(oldinfo->stages[other_source_index].mode) !=
2016                          S_ISREG(newinfo->stages[target_index].mode));
2017                 if (type_changed && collision) {
2018                         /*
2019                          * special handling so later blocks can handle this...
2020                          *
2021                          * if type_changed && collision are both true, then this
2022                          * was really a double rename, but one side wasn't
2023                          * detected due to lack of break detection.  I.e.
2024                          * something like
2025                          *    orig: has normal file 'foo'
2026                          *    side1: renames 'foo' to 'bar', adds 'foo' symlink
2027                          *    side2: renames 'foo' to 'bar'
2028                          * In this case, the foo->bar rename on side1 won't be
2029                          * detected because the new symlink named 'foo' is
2030                          * there and we don't do break detection.  But we detect
2031                          * this here because we don't want to merge the content
2032                          * of the foo symlink with the foo->bar file, so we
2033                          * have some logic to handle this special case.  The
2034                          * easiest way to do that is make 'bar' on side1 not
2035                          * be considered a colliding file but the other part
2036                          * of a normal rename.  If the file is very different,
2037                          * well we're going to get content merge conflicts
2038                          * anyway so it doesn't hurt.  And if the colliding
2039                          * file also has a different type, that'll be handled
2040                          * by the content merge logic in process_entry() too.
2041                          *
2042                          * See also t6430, 'rename vs. rename/symlink'
2043                          */
2044                         collision = 0;
2045                 }
2046                 if (source_deleted) {
2047                         if (target_index == 1) {
2048                                 rename_branch = opt->branch1;
2049                                 delete_branch = opt->branch2;
2050                         } else {
2051                                 rename_branch = opt->branch2;
2052                                 delete_branch = opt->branch1;
2053                         }
2054                 }
2055
2056                 assert(source_deleted || oldinfo->filemask & old_sidemask);
2057
2058                 /* Need to check for special types of rename conflicts... */
2059                 if (collision && !source_deleted) {
2060                         /* collision: rename/add or rename/rename(2to1) */
2061                         const char *pathnames[3];
2062                         struct version_info merged;
2063
2064                         struct conflict_info *base, *side1, *side2;
2065                         unsigned clean;
2066
2067                         pathnames[0] = oldpath;
2068                         pathnames[other_source_index] = oldpath;
2069                         pathnames[target_index] = newpath;
2070
2071                         base = strmap_get(&opt->priv->paths, pathnames[0]);
2072                         side1 = strmap_get(&opt->priv->paths, pathnames[1]);
2073                         side2 = strmap_get(&opt->priv->paths, pathnames[2]);
2074
2075                         VERIFY_CI(base);
2076                         VERIFY_CI(side1);
2077                         VERIFY_CI(side2);
2078
2079                         clean = handle_content_merge(opt, pair->one->path,
2080                                                      &base->stages[0],
2081                                                      &side1->stages[1],
2082                                                      &side2->stages[2],
2083                                                      pathnames,
2084                                                      1 + 2 * opt->priv->call_depth,
2085                                                      &merged);
2086
2087                         memcpy(&newinfo->stages[target_index], &merged,
2088                                sizeof(merged));
2089                         if (!clean) {
2090                                 path_msg(opt, newpath, 0,
2091                                          _("CONFLICT (rename involved in "
2092                                            "collision): rename of %s -> %s has "
2093                                            "content conflicts AND collides "
2094                                            "with another path; this may result "
2095                                            "in nested conflict markers."),
2096                                          oldpath, newpath);
2097                         }
2098                 } else if (collision && source_deleted) {
2099                         /*
2100                          * rename/add/delete or rename/rename(2to1)/delete:
2101                          * since oldpath was deleted on the side that didn't
2102                          * do the rename, there's not much of a content merge
2103                          * we can do for the rename.  oldinfo->merged.is_null
2104                          * was already set, so we just leave things as-is so
2105                          * they look like an add/add conflict.
2106                          */
2107
2108                         newinfo->path_conflict = 1;
2109                         path_msg(opt, newpath, 0,
2110                                  _("CONFLICT (rename/delete): %s renamed "
2111                                    "to %s in %s, but deleted in %s."),
2112                                  oldpath, newpath, rename_branch, delete_branch);
2113                 } else {
2114                         /*
2115                          * a few different cases...start by copying the
2116                          * existing stage(s) from oldinfo over the newinfo
2117                          * and update the pathname(s).
2118                          */
2119                         memcpy(&newinfo->stages[0], &oldinfo->stages[0],
2120                                sizeof(newinfo->stages[0]));
2121                         newinfo->filemask |= (1 << MERGE_BASE);
2122                         newinfo->pathnames[0] = oldpath;
2123                         if (type_changed) {
2124                                 /* rename vs. typechange */
2125                                 /* Mark the original as resolved by removal */
2126                                 memcpy(&oldinfo->stages[0].oid, &null_oid,
2127                                        sizeof(oldinfo->stages[0].oid));
2128                                 oldinfo->stages[0].mode = 0;
2129                                 oldinfo->filemask &= 0x06;
2130                         } else if (source_deleted) {
2131                                 /* rename/delete */
2132                                 newinfo->path_conflict = 1;
2133                                 path_msg(opt, newpath, 0,
2134                                          _("CONFLICT (rename/delete): %s renamed"
2135                                            " to %s in %s, but deleted in %s."),
2136                                          oldpath, newpath,
2137                                          rename_branch, delete_branch);
2138                         } else {
2139                                 /* normal rename */
2140                                 memcpy(&newinfo->stages[other_source_index],
2141                                        &oldinfo->stages[other_source_index],
2142                                        sizeof(newinfo->stages[0]));
2143                                 newinfo->filemask |= (1 << other_source_index);
2144                                 newinfo->pathnames[other_source_index] = oldpath;
2145                         }
2146                 }
2147
2148                 if (!type_changed) {
2149                         /* Mark the original as resolved by removal */
2150                         oldinfo->merged.is_null = 1;
2151                         oldinfo->merged.clean = 1;
2152                 }
2153
2154         }
2155
2156         return clean_merge;
2157 }
2158
2159 static inline int possible_side_renames(struct rename_info *renames,
2160                                         unsigned side_index)
2161 {
2162         return renames->pairs[side_index].nr > 0 &&
2163                !strset_empty(&renames->relevant_sources[side_index]);
2164 }
2165
2166 static inline int possible_renames(struct rename_info *renames)
2167 {
2168         return possible_side_renames(renames, 1) ||
2169                possible_side_renames(renames, 2);
2170 }
2171
2172 static void resolve_diffpair_statuses(struct diff_queue_struct *q)
2173 {
2174         /*
2175          * A simplified version of diff_resolve_rename_copy(); would probably
2176          * just use that function but it's static...
2177          */
2178         int i;
2179         struct diff_filepair *p;
2180
2181         for (i = 0; i < q->nr; ++i) {
2182                 p = q->queue[i];
2183                 p->status = 0; /* undecided */
2184                 if (!DIFF_FILE_VALID(p->one))
2185                         p->status = DIFF_STATUS_ADDED;
2186                 else if (!DIFF_FILE_VALID(p->two))
2187                         p->status = DIFF_STATUS_DELETED;
2188                 else if (DIFF_PAIR_RENAME(p))
2189                         p->status = DIFF_STATUS_RENAMED;
2190         }
2191 }
2192
2193 static int compare_pairs(const void *a_, const void *b_)
2194 {
2195         const struct diff_filepair *a = *((const struct diff_filepair **)a_);
2196         const struct diff_filepair *b = *((const struct diff_filepair **)b_);
2197
2198         return strcmp(a->one->path, b->one->path);
2199 }
2200
2201 /* Call diffcore_rename() to compute which files have changed on given side */
2202 static void detect_regular_renames(struct merge_options *opt,
2203                                    unsigned side_index)
2204 {
2205         struct diff_options diff_opts;
2206         struct rename_info *renames = &opt->priv->renames;
2207
2208         if (!possible_side_renames(renames, side_index)) {
2209                 /*
2210                  * No rename detection needed for this side, but we still need
2211                  * to make sure 'adds' are marked correctly in case the other
2212                  * side had directory renames.
2213                  */
2214                 resolve_diffpair_statuses(&renames->pairs[side_index]);
2215                 return;
2216         }
2217
2218         repo_diff_setup(opt->repo, &diff_opts);
2219         diff_opts.flags.recursive = 1;
2220         diff_opts.flags.rename_empty = 0;
2221         diff_opts.detect_rename = DIFF_DETECT_RENAME;
2222         diff_opts.rename_limit = opt->rename_limit;
2223         if (opt->rename_limit <= 0)
2224                 diff_opts.rename_limit = 1000;
2225         diff_opts.rename_score = opt->rename_score;
2226         diff_opts.show_rename_progress = opt->show_rename_progress;
2227         diff_opts.output_format = DIFF_FORMAT_NO_OUTPUT;
2228         diff_setup_done(&diff_opts);
2229
2230         diff_queued_diff = renames->pairs[side_index];
2231         trace2_region_enter("diff", "diffcore_rename", opt->repo);
2232         diffcore_rename_extended(&diff_opts,
2233                                  &renames->relevant_sources[side_index],
2234                                  &renames->dirs_removed[side_index],
2235                                  &renames->dir_rename_count[side_index]);
2236         trace2_region_leave("diff", "diffcore_rename", opt->repo);
2237         resolve_diffpair_statuses(&diff_queued_diff);
2238
2239         if (diff_opts.needed_rename_limit > renames->needed_limit)
2240                 renames->needed_limit = diff_opts.needed_rename_limit;
2241
2242         renames->pairs[side_index] = diff_queued_diff;
2243
2244         diff_opts.output_format = DIFF_FORMAT_NO_OUTPUT;
2245         diff_queued_diff.nr = 0;
2246         diff_queued_diff.queue = NULL;
2247         diff_flush(&diff_opts);
2248 }
2249
2250 /*
2251  * Get information of all renames which occurred in 'side_pairs', discarding
2252  * non-renames.
2253  */
2254 static int collect_renames(struct merge_options *opt,
2255                            struct diff_queue_struct *result,
2256                            unsigned side_index,
2257                            struct strmap *dir_renames_for_side,
2258                            struct strmap *rename_exclusions)
2259 {
2260         int i, clean = 1;
2261         struct strmap collisions;
2262         struct diff_queue_struct *side_pairs;
2263         struct hashmap_iter iter;
2264         struct strmap_entry *entry;
2265         struct rename_info *renames = &opt->priv->renames;
2266
2267         side_pairs = &renames->pairs[side_index];
2268         compute_collisions(&collisions, dir_renames_for_side, side_pairs);
2269
2270         for (i = 0; i < side_pairs->nr; ++i) {
2271                 struct diff_filepair *p = side_pairs->queue[i];
2272                 char *new_path; /* non-NULL only with directory renames */
2273
2274                 if (p->status != 'A' && p->status != 'R') {
2275                         diff_free_filepair(p);
2276                         continue;
2277                 }
2278
2279                 new_path = check_for_directory_rename(opt, p->two->path,
2280                                                       side_index,
2281                                                       dir_renames_for_side,
2282                                                       rename_exclusions,
2283                                                       &collisions,
2284                                                       &clean);
2285
2286                 if (p->status != 'R' && !new_path) {
2287                         diff_free_filepair(p);
2288                         continue;
2289                 }
2290
2291                 if (new_path)
2292                         apply_directory_rename_modifications(opt, p, new_path);
2293
2294                 /*
2295                  * p->score comes back from diffcore_rename_extended() with
2296                  * the similarity of the renamed file.  The similarity is
2297                  * was used to determine that the two files were related
2298                  * and are a rename, which we have already used, but beyond
2299                  * that we have no use for the similarity.  So p->score is
2300                  * now irrelevant.  However, process_renames() will need to
2301                  * know which side of the merge this rename was associated
2302                  * with, so overwrite p->score with that value.
2303                  */
2304                 p->score = side_index;
2305                 result->queue[result->nr++] = p;
2306         }
2307
2308         /* Free each value in the collisions map */
2309         strmap_for_each_entry(&collisions, &iter, entry) {
2310                 struct collision_info *info = entry->value;
2311                 string_list_clear(&info->source_files, 0);
2312         }
2313         /*
2314          * In compute_collisions(), we set collisions.strdup_strings to 0
2315          * so that we wouldn't have to make another copy of the new_path
2316          * allocated by apply_dir_rename().  But now that we've used them
2317          * and have no other references to these strings, it is time to
2318          * deallocate them.
2319          */
2320         free_strmap_strings(&collisions);
2321         strmap_clear(&collisions, 1);
2322         return clean;
2323 }
2324
2325 static int detect_and_process_renames(struct merge_options *opt,
2326                                       struct tree *merge_base,
2327                                       struct tree *side1,
2328                                       struct tree *side2)
2329 {
2330         struct diff_queue_struct combined;
2331         struct rename_info *renames = &opt->priv->renames;
2332         int need_dir_renames, s, clean = 1;
2333
2334         memset(&combined, 0, sizeof(combined));
2335         if (!possible_renames(renames))
2336                 goto cleanup;
2337
2338         trace2_region_enter("merge", "regular renames", opt->repo);
2339         detect_regular_renames(opt, MERGE_SIDE1);
2340         detect_regular_renames(opt, MERGE_SIDE2);
2341         trace2_region_leave("merge", "regular renames", opt->repo);
2342
2343         trace2_region_enter("merge", "directory renames", opt->repo);
2344         need_dir_renames =
2345           !opt->priv->call_depth &&
2346           (opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_TRUE ||
2347            opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_CONFLICT);
2348
2349         if (need_dir_renames) {
2350                 get_provisional_directory_renames(opt, MERGE_SIDE1, &clean);
2351                 get_provisional_directory_renames(opt, MERGE_SIDE2, &clean);
2352                 handle_directory_level_conflicts(opt);
2353         }
2354
2355         ALLOC_GROW(combined.queue,
2356                    renames->pairs[1].nr + renames->pairs[2].nr,
2357                    combined.alloc);
2358         clean &= collect_renames(opt, &combined, MERGE_SIDE1,
2359                                  &renames->dir_renames[2],
2360                                  &renames->dir_renames[1]);
2361         clean &= collect_renames(opt, &combined, MERGE_SIDE2,
2362                                  &renames->dir_renames[1],
2363                                  &renames->dir_renames[2]);
2364         QSORT(combined.queue, combined.nr, compare_pairs);
2365         trace2_region_leave("merge", "directory renames", opt->repo);
2366
2367         trace2_region_enter("merge", "process renames", opt->repo);
2368         clean &= process_renames(opt, &combined);
2369         trace2_region_leave("merge", "process renames", opt->repo);
2370
2371         goto simple_cleanup; /* collect_renames() handles some of cleanup */
2372
2373 cleanup:
2374         /*
2375          * Free now unneeded filepairs, which would have been handled
2376          * in collect_renames() normally but we skipped that code.
2377          */
2378         for (s = MERGE_SIDE1; s <= MERGE_SIDE2; s++) {
2379                 struct diff_queue_struct *side_pairs;
2380                 int i;
2381
2382                 side_pairs = &renames->pairs[s];
2383                 for (i = 0; i < side_pairs->nr; ++i) {
2384                         struct diff_filepair *p = side_pairs->queue[i];
2385                         diff_free_filepair(p);
2386                 }
2387         }
2388
2389 simple_cleanup:
2390         /* Free memory for renames->pairs[] and combined */
2391         for (s = MERGE_SIDE1; s <= MERGE_SIDE2; s++) {
2392                 free(renames->pairs[s].queue);
2393                 DIFF_QUEUE_CLEAR(&renames->pairs[s]);
2394         }
2395         if (combined.nr) {
2396                 int i;
2397                 for (i = 0; i < combined.nr; i++)
2398                         diff_free_filepair(combined.queue[i]);
2399                 free(combined.queue);
2400         }
2401
2402         return clean;
2403 }
2404
2405 /*** Function Grouping: functions related to process_entries() ***/
2406
2407 static int string_list_df_name_compare(const char *one, const char *two)
2408 {
2409         int onelen = strlen(one);
2410         int twolen = strlen(two);
2411         /*
2412          * Here we only care that entries for D/F conflicts are
2413          * adjacent, in particular with the file of the D/F conflict
2414          * appearing before files below the corresponding directory.
2415          * The order of the rest of the list is irrelevant for us.
2416          *
2417          * To achieve this, we sort with df_name_compare and provide
2418          * the mode S_IFDIR so that D/F conflicts will sort correctly.
2419          * We use the mode S_IFDIR for everything else for simplicity,
2420          * since in other cases any changes in their order due to
2421          * sorting cause no problems for us.
2422          */
2423         int cmp = df_name_compare(one, onelen, S_IFDIR,
2424                                   two, twolen, S_IFDIR);
2425         /*
2426          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
2427          * that 'foo' comes before 'foo/bar'.
2428          */
2429         if (cmp)
2430                 return cmp;
2431         return onelen - twolen;
2432 }
2433
2434 struct directory_versions {
2435         /*
2436          * versions: list of (basename -> version_info)
2437          *
2438          * The basenames are in reverse lexicographic order of full pathnames,
2439          * as processed in process_entries().  This puts all entries within
2440          * a directory together, and covers the directory itself after
2441          * everything within it, allowing us to write subtrees before needing
2442          * to record information for the tree itself.
2443          */
2444         struct string_list versions;
2445
2446         /*
2447          * offsets: list of (full relative path directories -> integer offsets)
2448          *
2449          * Since versions contains basenames from files in multiple different
2450          * directories, we need to know which entries in versions correspond
2451          * to which directories.  Values of e.g.
2452          *     ""             0
2453          *     src            2
2454          *     src/moduleA    5
2455          * Would mean that entries 0-1 of versions are files in the toplevel
2456          * directory, entries 2-4 are files under src/, and the remaining
2457          * entries starting at index 5 are files under src/moduleA/.
2458          */
2459         struct string_list offsets;
2460
2461         /*
2462          * last_directory: directory that previously processed file found in
2463          *
2464          * last_directory starts NULL, but records the directory in which the
2465          * previous file was found within.  As soon as
2466          *    directory(current_file) != last_directory
2467          * then we need to start updating accounting in versions & offsets.
2468          * Note that last_directory is always the last path in "offsets" (or
2469          * NULL if "offsets" is empty) so this exists just for quick access.
2470          */
2471         const char *last_directory;
2472
2473         /* last_directory_len: cached computation of strlen(last_directory) */
2474         unsigned last_directory_len;
2475 };
2476
2477 static int tree_entry_order(const void *a_, const void *b_)
2478 {
2479         const struct string_list_item *a = a_;
2480         const struct string_list_item *b = b_;
2481
2482         const struct merged_info *ami = a->util;
2483         const struct merged_info *bmi = b->util;
2484         return base_name_compare(a->string, strlen(a->string), ami->result.mode,
2485                                  b->string, strlen(b->string), bmi->result.mode);
2486 }
2487
2488 static void write_tree(struct object_id *result_oid,
2489                        struct string_list *versions,
2490                        unsigned int offset,
2491                        size_t hash_size)
2492 {
2493         size_t maxlen = 0, extra;
2494         unsigned int nr = versions->nr - offset;
2495         struct strbuf buf = STRBUF_INIT;
2496         struct string_list relevant_entries = STRING_LIST_INIT_NODUP;
2497         int i;
2498
2499         /*
2500          * We want to sort the last (versions->nr-offset) entries in versions.
2501          * Do so by abusing the string_list API a bit: make another string_list
2502          * that contains just those entries and then sort them.
2503          *
2504          * We won't use relevant_entries again and will let it just pop off the
2505          * stack, so there won't be allocation worries or anything.
2506          */
2507         relevant_entries.items = versions->items + offset;
2508         relevant_entries.nr = versions->nr - offset;
2509         QSORT(relevant_entries.items, relevant_entries.nr, tree_entry_order);
2510
2511         /* Pre-allocate some space in buf */
2512         extra = hash_size + 8; /* 8: 6 for mode, 1 for space, 1 for NUL char */
2513         for (i = 0; i < nr; i++) {
2514                 maxlen += strlen(versions->items[offset+i].string) + extra;
2515         }
2516         strbuf_grow(&buf, maxlen);
2517
2518         /* Write each entry out to buf */
2519         for (i = 0; i < nr; i++) {
2520                 struct merged_info *mi = versions->items[offset+i].util;
2521                 struct version_info *ri = &mi->result;
2522                 strbuf_addf(&buf, "%o %s%c",
2523                             ri->mode,
2524                             versions->items[offset+i].string, '\0');
2525                 strbuf_add(&buf, ri->oid.hash, hash_size);
2526         }
2527
2528         /* Write this object file out, and record in result_oid */
2529         write_object_file(buf.buf, buf.len, tree_type, result_oid);
2530         strbuf_release(&buf);
2531 }
2532
2533 static void record_entry_for_tree(struct directory_versions *dir_metadata,
2534                                   const char *path,
2535                                   struct merged_info *mi)
2536 {
2537         const char *basename;
2538
2539         if (mi->is_null)
2540                 /* nothing to record */
2541                 return;
2542
2543         basename = path + mi->basename_offset;
2544         assert(strchr(basename, '/') == NULL);
2545         string_list_append(&dir_metadata->versions,
2546                            basename)->util = &mi->result;
2547 }
2548
2549 static void write_completed_directory(struct merge_options *opt,
2550                                       const char *new_directory_name,
2551                                       struct directory_versions *info)
2552 {
2553         const char *prev_dir;
2554         struct merged_info *dir_info = NULL;
2555         unsigned int offset;
2556
2557         /*
2558          * Some explanation of info->versions and info->offsets...
2559          *
2560          * process_entries() iterates over all relevant files AND
2561          * directories in reverse lexicographic order, and calls this
2562          * function.  Thus, an example of the paths that process_entries()
2563          * could operate on (along with the directories for those paths
2564          * being shown) is:
2565          *
2566          *     xtract.c             ""
2567          *     tokens.txt           ""
2568          *     src/moduleB/umm.c    src/moduleB
2569          *     src/moduleB/stuff.h  src/moduleB
2570          *     src/moduleB/baz.c    src/moduleB
2571          *     src/moduleB          src
2572          *     src/moduleA/foo.c    src/moduleA
2573          *     src/moduleA/bar.c    src/moduleA
2574          *     src/moduleA          src
2575          *     src                  ""
2576          *     Makefile             ""
2577          *
2578          * info->versions:
2579          *
2580          *     always contains the unprocessed entries and their
2581          *     version_info information.  For example, after the first five
2582          *     entries above, info->versions would be:
2583          *
2584          *         xtract.c     <xtract.c's version_info>
2585          *         token.txt    <token.txt's version_info>
2586          *         umm.c        <src/moduleB/umm.c's version_info>
2587          *         stuff.h      <src/moduleB/stuff.h's version_info>
2588          *         baz.c        <src/moduleB/baz.c's version_info>
2589          *
2590          *     Once a subdirectory is completed we remove the entries in
2591          *     that subdirectory from info->versions, writing it as a tree
2592          *     (write_tree()).  Thus, as soon as we get to src/moduleB,
2593          *     info->versions would be updated to
2594          *
2595          *         xtract.c     <xtract.c's version_info>
2596          *         token.txt    <token.txt's version_info>
2597          *         moduleB      <src/moduleB's version_info>
2598          *
2599          * info->offsets:
2600          *
2601          *     helps us track which entries in info->versions correspond to
2602          *     which directories.  When we are N directories deep (e.g. 4
2603          *     for src/modA/submod/subdir/), we have up to N+1 unprocessed
2604          *     directories (+1 because of toplevel dir).  Corresponding to
2605          *     the info->versions example above, after processing five entries
2606          *     info->offsets will be:
2607          *
2608          *         ""           0
2609          *         src/moduleB  2
2610          *
2611          *     which is used to know that xtract.c & token.txt are from the
2612          *     toplevel dirctory, while umm.c & stuff.h & baz.c are from the
2613          *     src/moduleB directory.  Again, following the example above,
2614          *     once we need to process src/moduleB, then info->offsets is
2615          *     updated to
2616          *
2617          *         ""           0
2618          *         src          2
2619          *
2620          *     which says that moduleB (and only moduleB so far) is in the
2621          *     src directory.
2622          *
2623          *     One unique thing to note about info->offsets here is that
2624          *     "src" was not added to info->offsets until there was a path
2625          *     (a file OR directory) immediately below src/ that got
2626          *     processed.
2627          *
2628          * Since process_entry() just appends new entries to info->versions,
2629          * write_completed_directory() only needs to do work if the next path
2630          * is in a directory that is different than the last directory found
2631          * in info->offsets.
2632          */
2633
2634         /*
2635          * If we are working with the same directory as the last entry, there
2636          * is no work to do.  (See comments above the directory_name member of
2637          * struct merged_info for why we can use pointer comparison instead of
2638          * strcmp here.)
2639          */
2640         if (new_directory_name == info->last_directory)
2641                 return;
2642
2643         /*
2644          * If we are just starting (last_directory is NULL), or last_directory
2645          * is a prefix of the current directory, then we can just update
2646          * info->offsets to record the offset where we started this directory
2647          * and update last_directory to have quick access to it.
2648          */
2649         if (info->last_directory == NULL ||
2650             !strncmp(new_directory_name, info->last_directory,
2651                      info->last_directory_len)) {
2652                 uintptr_t offset = info->versions.nr;
2653
2654                 info->last_directory = new_directory_name;
2655                 info->last_directory_len = strlen(info->last_directory);
2656                 /*
2657                  * Record the offset into info->versions where we will
2658                  * start recording basenames of paths found within
2659                  * new_directory_name.
2660                  */
2661                 string_list_append(&info->offsets,
2662                                    info->last_directory)->util = (void*)offset;
2663                 return;
2664         }
2665
2666         /*
2667          * The next entry that will be processed will be within
2668          * new_directory_name.  Since at this point we know that
2669          * new_directory_name is within a different directory than
2670          * info->last_directory, we have all entries for info->last_directory
2671          * in info->versions and we need to create a tree object for them.
2672          */
2673         dir_info = strmap_get(&opt->priv->paths, info->last_directory);
2674         assert(dir_info);
2675         offset = (uintptr_t)info->offsets.items[info->offsets.nr-1].util;
2676         if (offset == info->versions.nr) {
2677                 /*
2678                  * Actually, we don't need to create a tree object in this
2679                  * case.  Whenever all files within a directory disappear
2680                  * during the merge (e.g. unmodified on one side and
2681                  * deleted on the other, or files were renamed elsewhere),
2682                  * then we get here and the directory itself needs to be
2683                  * omitted from its parent tree as well.
2684                  */
2685                 dir_info->is_null = 1;
2686         } else {
2687                 /*
2688                  * Write out the tree to the git object directory, and also
2689                  * record the mode and oid in dir_info->result.
2690                  */
2691                 dir_info->is_null = 0;
2692                 dir_info->result.mode = S_IFDIR;
2693                 write_tree(&dir_info->result.oid, &info->versions, offset,
2694                            opt->repo->hash_algo->rawsz);
2695         }
2696
2697         /*
2698          * We've now used several entries from info->versions and one entry
2699          * from info->offsets, so we get rid of those values.
2700          */
2701         info->offsets.nr--;
2702         info->versions.nr = offset;
2703
2704         /*
2705          * Now we've taken care of the completed directory, but we need to
2706          * prepare things since future entries will be in
2707          * new_directory_name.  (In particular, process_entry() will be
2708          * appending new entries to info->versions.)  So, we need to make
2709          * sure new_directory_name is the last entry in info->offsets.
2710          */
2711         prev_dir = info->offsets.nr == 0 ? NULL :
2712                    info->offsets.items[info->offsets.nr-1].string;
2713         if (new_directory_name != prev_dir) {
2714                 uintptr_t c = info->versions.nr;
2715                 string_list_append(&info->offsets,
2716                                    new_directory_name)->util = (void*)c;
2717         }
2718
2719         /* And, of course, we need to update last_directory to match. */
2720         info->last_directory = new_directory_name;
2721         info->last_directory_len = strlen(info->last_directory);
2722 }
2723
2724 /* Per entry merge function */
2725 static void process_entry(struct merge_options *opt,
2726                           const char *path,
2727                           struct conflict_info *ci,
2728                           struct directory_versions *dir_metadata)
2729 {
2730         int df_file_index = 0;
2731
2732         VERIFY_CI(ci);
2733         assert(ci->filemask >= 0 && ci->filemask <= 7);
2734         /* ci->match_mask == 7 was handled in collect_merge_info_callback() */
2735         assert(ci->match_mask == 0 || ci->match_mask == 3 ||
2736                ci->match_mask == 5 || ci->match_mask == 6);
2737
2738         if (ci->dirmask) {
2739                 record_entry_for_tree(dir_metadata, path, &ci->merged);
2740                 if (ci->filemask == 0)
2741                         /* nothing else to handle */
2742                         return;
2743                 assert(ci->df_conflict);
2744         }
2745
2746         if (ci->df_conflict && ci->merged.result.mode == 0) {
2747                 int i;
2748
2749                 /*
2750                  * directory no longer in the way, but we do have a file we
2751                  * need to place here so we need to clean away the "directory
2752                  * merges to nothing" result.
2753                  */
2754                 ci->df_conflict = 0;
2755                 assert(ci->filemask != 0);
2756                 ci->merged.clean = 0;
2757                 ci->merged.is_null = 0;
2758                 /* and we want to zero out any directory-related entries */
2759                 ci->match_mask = (ci->match_mask & ~ci->dirmask);
2760                 ci->dirmask = 0;
2761                 for (i = MERGE_BASE; i <= MERGE_SIDE2; i++) {
2762                         if (ci->filemask & (1 << i))
2763                                 continue;
2764                         ci->stages[i].mode = 0;
2765                         oidcpy(&ci->stages[i].oid, &null_oid);
2766                 }
2767         } else if (ci->df_conflict && ci->merged.result.mode != 0) {
2768                 /*
2769                  * This started out as a D/F conflict, and the entries in
2770                  * the competing directory were not removed by the merge as
2771                  * evidenced by write_completed_directory() writing a value
2772                  * to ci->merged.result.mode.
2773                  */
2774                 struct conflict_info *new_ci;
2775                 const char *branch;
2776                 const char *old_path = path;
2777                 int i;
2778
2779                 assert(ci->merged.result.mode == S_IFDIR);
2780
2781                 /*
2782                  * If filemask is 1, we can just ignore the file as having
2783                  * been deleted on both sides.  We do not want to overwrite
2784                  * ci->merged.result, since it stores the tree for all the
2785                  * files under it.
2786                  */
2787                 if (ci->filemask == 1) {
2788                         ci->filemask = 0;
2789                         return;
2790                 }
2791
2792                 /*
2793                  * This file still exists on at least one side, and we want
2794                  * the directory to remain here, so we need to move this
2795                  * path to some new location.
2796                  */
2797                 CALLOC_ARRAY(new_ci, 1);
2798                 /* We don't really want new_ci->merged.result copied, but it'll
2799                  * be overwritten below so it doesn't matter.  We also don't
2800                  * want any directory mode/oid values copied, but we'll zero
2801                  * those out immediately.  We do want the rest of ci copied.
2802                  */
2803                 memcpy(new_ci, ci, sizeof(*ci));
2804                 new_ci->match_mask = (new_ci->match_mask & ~new_ci->dirmask);
2805                 new_ci->dirmask = 0;
2806                 for (i = MERGE_BASE; i <= MERGE_SIDE2; i++) {
2807                         if (new_ci->filemask & (1 << i))
2808                                 continue;
2809                         /* zero out any entries related to directories */
2810                         new_ci->stages[i].mode = 0;
2811                         oidcpy(&new_ci->stages[i].oid, &null_oid);
2812                 }
2813
2814                 /*
2815                  * Find out which side this file came from; note that we
2816                  * cannot just use ci->filemask, because renames could cause
2817                  * the filemask to go back to 7.  So we use dirmask, then
2818                  * pick the opposite side's index.
2819                  */
2820                 df_file_index = (ci->dirmask & (1 << 1)) ? 2 : 1;
2821                 branch = (df_file_index == 1) ? opt->branch1 : opt->branch2;
2822                 path = unique_path(&opt->priv->paths, path, branch);
2823                 strmap_put(&opt->priv->paths, path, new_ci);
2824
2825                 path_msg(opt, path, 0,
2826                          _("CONFLICT (file/directory): directory in the way "
2827                            "of %s from %s; moving it to %s instead."),
2828                          old_path, branch, path);
2829
2830                 /*
2831                  * Zero out the filemask for the old ci.  At this point, ci
2832                  * was just an entry for a directory, so we don't need to
2833                  * do anything more with it.
2834                  */
2835                 ci->filemask = 0;
2836
2837                 /*
2838                  * Now note that we're working on the new entry (path was
2839                  * updated above.
2840                  */
2841                 ci = new_ci;
2842         }
2843
2844         /*
2845          * NOTE: Below there is a long switch-like if-elseif-elseif... block
2846          *       which the code goes through even for the df_conflict cases
2847          *       above.
2848          */
2849         if (ci->match_mask) {
2850                 ci->merged.clean = 1;
2851                 if (ci->match_mask == 6) {
2852                         /* stages[1] == stages[2] */
2853                         ci->merged.result.mode = ci->stages[1].mode;
2854                         oidcpy(&ci->merged.result.oid, &ci->stages[1].oid);
2855                 } else {
2856                         /* determine the mask of the side that didn't match */
2857                         unsigned int othermask = 7 & ~ci->match_mask;
2858                         int side = (othermask == 4) ? 2 : 1;
2859
2860                         ci->merged.result.mode = ci->stages[side].mode;
2861                         ci->merged.is_null = !ci->merged.result.mode;
2862                         oidcpy(&ci->merged.result.oid, &ci->stages[side].oid);
2863
2864                         assert(othermask == 2 || othermask == 4);
2865                         assert(ci->merged.is_null ==
2866                                (ci->filemask == ci->match_mask));
2867                 }
2868         } else if (ci->filemask >= 6 &&
2869                    (S_IFMT & ci->stages[1].mode) !=
2870                    (S_IFMT & ci->stages[2].mode)) {
2871                 /* Two different items from (file/submodule/symlink) */
2872                 if (opt->priv->call_depth) {
2873                         /* Just use the version from the merge base */
2874                         ci->merged.clean = 0;
2875                         oidcpy(&ci->merged.result.oid, &ci->stages[0].oid);
2876                         ci->merged.result.mode = ci->stages[0].mode;
2877                         ci->merged.is_null = (ci->merged.result.mode == 0);
2878                 } else {
2879                         /* Handle by renaming one or both to separate paths. */
2880                         unsigned o_mode = ci->stages[0].mode;
2881                         unsigned a_mode = ci->stages[1].mode;
2882                         unsigned b_mode = ci->stages[2].mode;
2883                         struct conflict_info *new_ci;
2884                         const char *a_path = NULL, *b_path = NULL;
2885                         int rename_a = 0, rename_b = 0;
2886
2887                         new_ci = xmalloc(sizeof(*new_ci));
2888
2889                         if (S_ISREG(a_mode))
2890                                 rename_a = 1;
2891                         else if (S_ISREG(b_mode))
2892                                 rename_b = 1;
2893                         else {
2894                                 rename_a = 1;
2895                                 rename_b = 1;
2896                         }
2897
2898                         path_msg(opt, path, 0,
2899                                  _("CONFLICT (distinct types): %s had different "
2900                                    "types on each side; renamed %s of them so "
2901                                    "each can be recorded somewhere."),
2902                                  path,
2903                                  (rename_a && rename_b) ? _("both") : _("one"));
2904
2905                         ci->merged.clean = 0;
2906                         memcpy(new_ci, ci, sizeof(*new_ci));
2907
2908                         /* Put b into new_ci, removing a from stages */
2909                         new_ci->merged.result.mode = ci->stages[2].mode;
2910                         oidcpy(&new_ci->merged.result.oid, &ci->stages[2].oid);
2911                         new_ci->stages[1].mode = 0;
2912                         oidcpy(&new_ci->stages[1].oid, &null_oid);
2913                         new_ci->filemask = 5;
2914                         if ((S_IFMT & b_mode) != (S_IFMT & o_mode)) {
2915                                 new_ci->stages[0].mode = 0;
2916                                 oidcpy(&new_ci->stages[0].oid, &null_oid);
2917                                 new_ci->filemask = 4;
2918                         }
2919
2920                         /* Leave only a in ci, fixing stages. */
2921                         ci->merged.result.mode = ci->stages[1].mode;
2922                         oidcpy(&ci->merged.result.oid, &ci->stages[1].oid);
2923                         ci->stages[2].mode = 0;
2924                         oidcpy(&ci->stages[2].oid, &null_oid);
2925                         ci->filemask = 3;
2926                         if ((S_IFMT & a_mode) != (S_IFMT & o_mode)) {
2927                                 ci->stages[0].mode = 0;
2928                                 oidcpy(&ci->stages[0].oid, &null_oid);
2929                                 ci->filemask = 2;
2930                         }
2931
2932                         /* Insert entries into opt->priv_paths */
2933                         assert(rename_a || rename_b);
2934                         if (rename_a) {
2935                                 a_path = unique_path(&opt->priv->paths,
2936                                                      path, opt->branch1);
2937                                 strmap_put(&opt->priv->paths, a_path, ci);
2938                         }
2939
2940                         if (rename_b)
2941                                 b_path = unique_path(&opt->priv->paths,
2942                                                      path, opt->branch2);
2943                         else
2944                                 b_path = path;
2945                         strmap_put(&opt->priv->paths, b_path, new_ci);
2946
2947                         if (rename_a && rename_b) {
2948                                 strmap_remove(&opt->priv->paths, path, 0);
2949                                 /*
2950                                  * We removed path from opt->priv->paths.  path
2951                                  * will also eventually need to be freed, but
2952                                  * it may still be used by e.g.  ci->pathnames.
2953                                  * So, store it in another string-list for now.
2954                                  */
2955                                 string_list_append(&opt->priv->paths_to_free,
2956                                                    path);
2957                         }
2958
2959                         /*
2960                          * Do special handling for b_path since process_entry()
2961                          * won't be called on it specially.
2962                          */
2963                         strmap_put(&opt->priv->conflicted, b_path, new_ci);
2964                         record_entry_for_tree(dir_metadata, b_path,
2965                                               &new_ci->merged);
2966
2967                         /*
2968                          * Remaining code for processing this entry should
2969                          * think in terms of processing a_path.
2970                          */
2971                         if (a_path)
2972                                 path = a_path;
2973                 }
2974         } else if (ci->filemask >= 6) {
2975                 /* Need a two-way or three-way content merge */
2976                 struct version_info merged_file;
2977                 unsigned clean_merge;
2978                 struct version_info *o = &ci->stages[0];
2979                 struct version_info *a = &ci->stages[1];
2980                 struct version_info *b = &ci->stages[2];
2981
2982                 clean_merge = handle_content_merge(opt, path, o, a, b,
2983                                                    ci->pathnames,
2984                                                    opt->priv->call_depth * 2,
2985                                                    &merged_file);
2986                 ci->merged.clean = clean_merge &&
2987                                    !ci->df_conflict && !ci->path_conflict;
2988                 ci->merged.result.mode = merged_file.mode;
2989                 ci->merged.is_null = (merged_file.mode == 0);
2990                 oidcpy(&ci->merged.result.oid, &merged_file.oid);
2991                 if (clean_merge && ci->df_conflict) {
2992                         assert(df_file_index == 1 || df_file_index == 2);
2993                         ci->filemask = 1 << df_file_index;
2994                         ci->stages[df_file_index].mode = merged_file.mode;
2995                         oidcpy(&ci->stages[df_file_index].oid, &merged_file.oid);
2996                 }
2997                 if (!clean_merge) {
2998                         const char *reason = _("content");
2999                         if (ci->filemask == 6)
3000                                 reason = _("add/add");
3001                         if (S_ISGITLINK(merged_file.mode))
3002                                 reason = _("submodule");
3003                         path_msg(opt, path, 0,
3004                                  _("CONFLICT (%s): Merge conflict in %s"),
3005                                  reason, path);
3006                 }
3007         } else if (ci->filemask == 3 || ci->filemask == 5) {
3008                 /* Modify/delete */
3009                 const char *modify_branch, *delete_branch;
3010                 int side = (ci->filemask == 5) ? 2 : 1;
3011                 int index = opt->priv->call_depth ? 0 : side;
3012
3013                 ci->merged.result.mode = ci->stages[index].mode;
3014                 oidcpy(&ci->merged.result.oid, &ci->stages[index].oid);
3015                 ci->merged.clean = 0;
3016
3017                 modify_branch = (side == 1) ? opt->branch1 : opt->branch2;
3018                 delete_branch = (side == 1) ? opt->branch2 : opt->branch1;
3019
3020                 if (ci->path_conflict &&
3021                     oideq(&ci->stages[0].oid, &ci->stages[side].oid)) {
3022                         /*
3023                          * This came from a rename/delete; no action to take,
3024                          * but avoid printing "modify/delete" conflict notice
3025                          * since the contents were not modified.
3026                          */
3027                 } else {
3028                         path_msg(opt, path, 0,
3029                                  _("CONFLICT (modify/delete): %s deleted in %s "
3030                                    "and modified in %s.  Version %s of %s left "
3031                                    "in tree."),
3032                                  path, delete_branch, modify_branch,
3033                                  modify_branch, path);
3034                 }
3035         } else if (ci->filemask == 2 || ci->filemask == 4) {
3036                 /* Added on one side */
3037                 int side = (ci->filemask == 4) ? 2 : 1;
3038                 ci->merged.result.mode = ci->stages[side].mode;
3039                 oidcpy(&ci->merged.result.oid, &ci->stages[side].oid);
3040                 ci->merged.clean = !ci->df_conflict && !ci->path_conflict;
3041         } else if (ci->filemask == 1) {
3042                 /* Deleted on both sides */
3043                 ci->merged.is_null = 1;
3044                 ci->merged.result.mode = 0;
3045                 oidcpy(&ci->merged.result.oid, &null_oid);
3046                 ci->merged.clean = !ci->path_conflict;
3047         }
3048
3049         /*
3050          * If still conflicted, record it separately.  This allows us to later
3051          * iterate over just conflicted entries when updating the index instead
3052          * of iterating over all entries.
3053          */
3054         if (!ci->merged.clean)
3055                 strmap_put(&opt->priv->conflicted, path, ci);
3056         record_entry_for_tree(dir_metadata, path, &ci->merged);
3057 }
3058
3059 static void process_entries(struct merge_options *opt,
3060                             struct object_id *result_oid)
3061 {
3062         struct hashmap_iter iter;
3063         struct strmap_entry *e;
3064         struct string_list plist = STRING_LIST_INIT_NODUP;
3065         struct string_list_item *entry;
3066         struct directory_versions dir_metadata = { STRING_LIST_INIT_NODUP,
3067                                                    STRING_LIST_INIT_NODUP,
3068                                                    NULL, 0 };
3069
3070         trace2_region_enter("merge", "process_entries setup", opt->repo);
3071         if (strmap_empty(&opt->priv->paths)) {
3072                 oidcpy(result_oid, opt->repo->hash_algo->empty_tree);
3073                 return;
3074         }
3075
3076         /* Hack to pre-allocate plist to the desired size */
3077         trace2_region_enter("merge", "plist grow", opt->repo);
3078         ALLOC_GROW(plist.items, strmap_get_size(&opt->priv->paths), plist.alloc);
3079         trace2_region_leave("merge", "plist grow", opt->repo);
3080
3081         /* Put every entry from paths into plist, then sort */
3082         trace2_region_enter("merge", "plist copy", opt->repo);
3083         strmap_for_each_entry(&opt->priv->paths, &iter, e) {
3084                 string_list_append(&plist, e->key)->util = e->value;
3085         }
3086         trace2_region_leave("merge", "plist copy", opt->repo);
3087
3088         trace2_region_enter("merge", "plist special sort", opt->repo);
3089         plist.cmp = string_list_df_name_compare;
3090         string_list_sort(&plist);
3091         trace2_region_leave("merge", "plist special sort", opt->repo);
3092
3093         trace2_region_leave("merge", "process_entries setup", opt->repo);
3094
3095         /*
3096          * Iterate over the items in reverse order, so we can handle paths
3097          * below a directory before needing to handle the directory itself.
3098          *
3099          * This allows us to write subtrees before we need to write trees,
3100          * and it also enables sane handling of directory/file conflicts
3101          * (because it allows us to know whether the directory is still in
3102          * the way when it is time to process the file at the same path).
3103          */
3104         trace2_region_enter("merge", "processing", opt->repo);
3105         for (entry = &plist.items[plist.nr-1]; entry >= plist.items; --entry) {
3106                 char *path = entry->string;
3107                 /*
3108                  * NOTE: mi may actually be a pointer to a conflict_info, but
3109                  * we have to check mi->clean first to see if it's safe to
3110                  * reassign to such a pointer type.
3111                  */
3112                 struct merged_info *mi = entry->util;
3113
3114                 write_completed_directory(opt, mi->directory_name,
3115                                           &dir_metadata);
3116                 if (mi->clean)
3117                         record_entry_for_tree(&dir_metadata, path, mi);
3118                 else {
3119                         struct conflict_info *ci = (struct conflict_info *)mi;
3120                         process_entry(opt, path, ci, &dir_metadata);
3121                 }
3122         }
3123         trace2_region_leave("merge", "processing", opt->repo);
3124
3125         trace2_region_enter("merge", "process_entries cleanup", opt->repo);
3126         if (dir_metadata.offsets.nr != 1 ||
3127             (uintptr_t)dir_metadata.offsets.items[0].util != 0) {
3128                 printf("dir_metadata.offsets.nr = %d (should be 1)\n",
3129                        dir_metadata.offsets.nr);
3130                 printf("dir_metadata.offsets.items[0].util = %u (should be 0)\n",
3131                        (unsigned)(uintptr_t)dir_metadata.offsets.items[0].util);
3132                 fflush(stdout);
3133                 BUG("dir_metadata accounting completely off; shouldn't happen");
3134         }
3135         write_tree(result_oid, &dir_metadata.versions, 0,
3136                    opt->repo->hash_algo->rawsz);
3137         string_list_clear(&plist, 0);
3138         string_list_clear(&dir_metadata.versions, 0);
3139         string_list_clear(&dir_metadata.offsets, 0);
3140         trace2_region_leave("merge", "process_entries cleanup", opt->repo);
3141 }
3142
3143 /*** Function Grouping: functions related to merge_switch_to_result() ***/
3144
3145 static int checkout(struct merge_options *opt,
3146                     struct tree *prev,
3147                     struct tree *next)
3148 {
3149         /* Switch the index/working copy from old to new */
3150         int ret;
3151         struct tree_desc trees[2];
3152         struct unpack_trees_options unpack_opts;
3153
3154         memset(&unpack_opts, 0, sizeof(unpack_opts));
3155         unpack_opts.head_idx = -1;
3156         unpack_opts.src_index = opt->repo->index;
3157         unpack_opts.dst_index = opt->repo->index;
3158
3159         setup_unpack_trees_porcelain(&unpack_opts, "merge");
3160
3161         /*
3162          * NOTE: if this were just "git checkout" code, we would probably
3163          * read or refresh the cache and check for a conflicted index, but
3164          * builtin/merge.c or sequencer.c really needs to read the index
3165          * and check for conflicted entries before starting merging for a
3166          * good user experience (no sense waiting for merges/rebases before
3167          * erroring out), so there's no reason to duplicate that work here.
3168          */
3169
3170         /* 2-way merge to the new branch */
3171         unpack_opts.update = 1;
3172         unpack_opts.merge = 1;
3173         unpack_opts.quiet = 0; /* FIXME: sequencer might want quiet? */
3174         unpack_opts.verbose_update = (opt->verbosity > 2);
3175         unpack_opts.fn = twoway_merge;
3176         if (1/* FIXME: opts->overwrite_ignore*/) {
3177                 CALLOC_ARRAY(unpack_opts.dir, 1);
3178                 unpack_opts.dir->flags |= DIR_SHOW_IGNORED;
3179                 setup_standard_excludes(unpack_opts.dir);
3180         }
3181         parse_tree(prev);
3182         init_tree_desc(&trees[0], prev->buffer, prev->size);
3183         parse_tree(next);
3184         init_tree_desc(&trees[1], next->buffer, next->size);
3185
3186         ret = unpack_trees(2, trees, &unpack_opts);
3187         clear_unpack_trees_porcelain(&unpack_opts);
3188         dir_clear(unpack_opts.dir);
3189         FREE_AND_NULL(unpack_opts.dir);
3190         return ret;
3191 }
3192
3193 static int record_conflicted_index_entries(struct merge_options *opt,
3194                                            struct index_state *index,
3195                                            struct strmap *paths,
3196                                            struct strmap *conflicted)
3197 {
3198         struct hashmap_iter iter;
3199         struct strmap_entry *e;
3200         int errs = 0;
3201         int original_cache_nr;
3202
3203         if (strmap_empty(conflicted))
3204                 return 0;
3205
3206         original_cache_nr = index->cache_nr;
3207
3208         /* Put every entry from paths into plist, then sort */
3209         strmap_for_each_entry(conflicted, &iter, e) {
3210                 const char *path = e->key;
3211                 struct conflict_info *ci = e->value;
3212                 int pos;
3213                 struct cache_entry *ce;
3214                 int i;
3215
3216                 VERIFY_CI(ci);
3217
3218                 /*
3219                  * The index will already have a stage=0 entry for this path,
3220                  * because we created an as-merged-as-possible version of the
3221                  * file and checkout() moved the working copy and index over
3222                  * to that version.
3223                  *
3224                  * However, previous iterations through this loop will have
3225                  * added unstaged entries to the end of the cache which
3226                  * ignore the standard alphabetical ordering of cache
3227                  * entries and break invariants needed for index_name_pos()
3228                  * to work.  However, we know the entry we want is before
3229                  * those appended cache entries, so do a temporary swap on
3230                  * cache_nr to only look through entries of interest.
3231                  */
3232                 SWAP(index->cache_nr, original_cache_nr);
3233                 pos = index_name_pos(index, path, strlen(path));
3234                 SWAP(index->cache_nr, original_cache_nr);
3235                 if (pos < 0) {
3236                         if (ci->filemask != 1)
3237                                 BUG("Conflicted %s but nothing in basic working tree or index; this shouldn't happen", path);
3238                         cache_tree_invalidate_path(index, path);
3239                 } else {
3240                         ce = index->cache[pos];
3241
3242                         /*
3243                          * Clean paths with CE_SKIP_WORKTREE set will not be
3244                          * written to the working tree by the unpack_trees()
3245                          * call in checkout().  Our conflicted entries would
3246                          * have appeared clean to that code since we ignored
3247                          * the higher order stages.  Thus, we need override
3248                          * the CE_SKIP_WORKTREE bit and manually write those
3249                          * files to the working disk here.
3250                          *
3251                          * TODO: Implement this CE_SKIP_WORKTREE fixup.
3252                          */
3253
3254                         /*
3255                          * Mark this cache entry for removal and instead add
3256                          * new stage>0 entries corresponding to the
3257                          * conflicts.  If there are many conflicted entries, we
3258                          * want to avoid memmove'ing O(NM) entries by
3259                          * inserting the new entries one at a time.  So,
3260                          * instead, we just add the new cache entries to the
3261                          * end (ignoring normal index requirements on sort
3262                          * order) and sort the index once we're all done.
3263                          */
3264                         ce->ce_flags |= CE_REMOVE;
3265                 }
3266
3267                 for (i = MERGE_BASE; i <= MERGE_SIDE2; i++) {
3268                         struct version_info *vi;
3269                         if (!(ci->filemask & (1ul << i)))
3270                                 continue;
3271                         vi = &ci->stages[i];
3272                         ce = make_cache_entry(index, vi->mode, &vi->oid,
3273                                               path, i+1, 0);
3274                         add_index_entry(index, ce, ADD_CACHE_JUST_APPEND);
3275                 }
3276         }
3277
3278         /*
3279          * Remove the unused cache entries (and invalidate the relevant
3280          * cache-trees), then sort the index entries to get the conflicted
3281          * entries we added to the end into their right locations.
3282          */
3283         remove_marked_cache_entries(index, 1);
3284         QSORT(index->cache, index->cache_nr, cmp_cache_name_compare);
3285
3286         return errs;
3287 }
3288
3289 void merge_switch_to_result(struct merge_options *opt,
3290                             struct tree *head,
3291                             struct merge_result *result,
3292                             int update_worktree_and_index,
3293                             int display_update_msgs)
3294 {
3295         assert(opt->priv == NULL);
3296         if (result->clean >= 0 && update_worktree_and_index) {
3297                 struct merge_options_internal *opti = result->priv;
3298
3299                 trace2_region_enter("merge", "checkout", opt->repo);
3300                 if (checkout(opt, head, result->tree)) {
3301                         /* failure to function */
3302                         result->clean = -1;
3303                         return;
3304                 }
3305                 trace2_region_leave("merge", "checkout", opt->repo);
3306
3307                 trace2_region_enter("merge", "record_conflicted", opt->repo);
3308                 if (record_conflicted_index_entries(opt, opt->repo->index,
3309                                                     &opti->paths,
3310                                                     &opti->conflicted)) {
3311                         /* failure to function */
3312                         result->clean = -1;
3313                         return;
3314                 }
3315                 trace2_region_leave("merge", "record_conflicted", opt->repo);
3316         }
3317
3318         if (display_update_msgs) {
3319                 struct merge_options_internal *opti = result->priv;
3320                 struct hashmap_iter iter;
3321                 struct strmap_entry *e;
3322                 struct string_list olist = STRING_LIST_INIT_NODUP;
3323                 int i;
3324
3325                 trace2_region_enter("merge", "display messages", opt->repo);
3326
3327                 /* Hack to pre-allocate olist to the desired size */
3328                 ALLOC_GROW(olist.items, strmap_get_size(&opti->output),
3329                            olist.alloc);
3330
3331                 /* Put every entry from output into olist, then sort */
3332                 strmap_for_each_entry(&opti->output, &iter, e) {
3333                         string_list_append(&olist, e->key)->util = e->value;
3334                 }
3335                 string_list_sort(&olist);
3336
3337                 /* Iterate over the items, printing them */
3338                 for (i = 0; i < olist.nr; ++i) {
3339                         struct strbuf *sb = olist.items[i].util;
3340
3341                         printf("%s", sb->buf);
3342                 }
3343                 string_list_clear(&olist, 0);
3344
3345                 /* Also include needed rename limit adjustment now */
3346                 diff_warn_rename_limit("merge.renamelimit",
3347                                        opti->renames.needed_limit, 0);
3348
3349                 trace2_region_leave("merge", "display messages", opt->repo);
3350         }
3351
3352         merge_finalize(opt, result);
3353 }
3354
3355 void merge_finalize(struct merge_options *opt,
3356                     struct merge_result *result)
3357 {
3358         struct merge_options_internal *opti = result->priv;
3359
3360         assert(opt->priv == NULL);
3361
3362         clear_or_reinit_internal_opts(opti, 0);
3363         FREE_AND_NULL(opti);
3364 }
3365
3366 /*** Function Grouping: helper functions for merge_incore_*() ***/
3367
3368 static inline void set_commit_tree(struct commit *c, struct tree *t)
3369 {
3370         c->maybe_tree = t;
3371 }
3372
3373 static struct commit *make_virtual_commit(struct repository *repo,
3374                                           struct tree *tree,
3375                                           const char *comment)
3376 {
3377         struct commit *commit = alloc_commit_node(repo);
3378
3379         set_merge_remote_desc(commit, comment, (struct object *)commit);
3380         set_commit_tree(commit, tree);
3381         commit->object.parsed = 1;
3382         return commit;
3383 }
3384
3385 static void merge_start(struct merge_options *opt, struct merge_result *result)
3386 {
3387         struct rename_info *renames;
3388         int i;
3389
3390         /* Sanity checks on opt */
3391         trace2_region_enter("merge", "sanity checks", opt->repo);
3392         assert(opt->repo);
3393
3394         assert(opt->branch1 && opt->branch2);
3395
3396         assert(opt->detect_directory_renames >= MERGE_DIRECTORY_RENAMES_NONE &&
3397                opt->detect_directory_renames <= MERGE_DIRECTORY_RENAMES_TRUE);
3398         assert(opt->rename_limit >= -1);
3399         assert(opt->rename_score >= 0 && opt->rename_score <= MAX_SCORE);
3400         assert(opt->show_rename_progress >= 0 && opt->show_rename_progress <= 1);
3401
3402         assert(opt->xdl_opts >= 0);
3403         assert(opt->recursive_variant >= MERGE_VARIANT_NORMAL &&
3404                opt->recursive_variant <= MERGE_VARIANT_THEIRS);
3405
3406         /*
3407          * detect_renames, verbosity, buffer_output, and obuf are ignored
3408          * fields that were used by "recursive" rather than "ort" -- but
3409          * sanity check them anyway.
3410          */
3411         assert(opt->detect_renames >= -1 &&
3412                opt->detect_renames <= DIFF_DETECT_COPY);
3413         assert(opt->verbosity >= 0 && opt->verbosity <= 5);
3414         assert(opt->buffer_output <= 2);
3415         assert(opt->obuf.len == 0);
3416
3417         assert(opt->priv == NULL);
3418         if (result->priv) {
3419                 opt->priv = result->priv;
3420                 result->priv = NULL;
3421                 /*
3422                  * opt->priv non-NULL means we had results from a previous
3423                  * run; do a few sanity checks that user didn't mess with
3424                  * it in an obvious fashion.
3425                  */
3426                 assert(opt->priv->call_depth == 0);
3427                 assert(!opt->priv->toplevel_dir ||
3428                        0 == strlen(opt->priv->toplevel_dir));
3429         }
3430         trace2_region_leave("merge", "sanity checks", opt->repo);
3431
3432         /* Default to histogram diff.  Actually, just hardcode it...for now. */
3433         opt->xdl_opts = DIFF_WITH_ALG(opt, HISTOGRAM_DIFF);
3434
3435         /* Initialization of opt->priv, our internal merge data */
3436         trace2_region_enter("merge", "allocate/init", opt->repo);
3437         if (opt->priv) {
3438                 clear_or_reinit_internal_opts(opt->priv, 1);
3439                 trace2_region_leave("merge", "allocate/init", opt->repo);
3440                 return;
3441         }
3442         opt->priv = xcalloc(1, sizeof(*opt->priv));
3443
3444         /* Initialization of various renames fields */
3445         renames = &opt->priv->renames;
3446         for (i = MERGE_SIDE1; i <= MERGE_SIDE2; i++) {
3447                 strset_init_with_options(&renames->dirs_removed[i],
3448                                          NULL, 0);
3449                 strmap_init_with_options(&renames->dir_rename_count[i],
3450                                          NULL, 1);
3451                 strmap_init_with_options(&renames->dir_renames[i],
3452                                          NULL, 0);
3453                 strset_init_with_options(&renames->relevant_sources[i],
3454                                          NULL, 0);
3455         }
3456
3457         /*
3458          * Although we initialize opt->priv->paths with strdup_strings=0,
3459          * that's just to avoid making yet another copy of an allocated
3460          * string.  Putting the entry into paths means we are taking
3461          * ownership, so we will later free it.  paths_to_free is similar.
3462          *
3463          * In contrast, conflicted just has a subset of keys from paths, so
3464          * we don't want to free those (it'd be a duplicate free).
3465          */
3466         strmap_init_with_options(&opt->priv->paths, NULL, 0);
3467         strmap_init_with_options(&opt->priv->conflicted, NULL, 0);
3468         string_list_init(&opt->priv->paths_to_free, 0);
3469
3470         /*
3471          * keys & strbufs in output will sometimes need to outlive "paths",
3472          * so it will have a copy of relevant keys.  It's probably a small
3473          * subset of the overall paths that have special output.
3474          */
3475         strmap_init(&opt->priv->output);
3476
3477         trace2_region_leave("merge", "allocate/init", opt->repo);
3478 }
3479
3480 /*** Function Grouping: merge_incore_*() and their internal variants ***/
3481
3482 /*
3483  * Originally from merge_trees_internal(); heavily adapted, though.
3484  */
3485 static void merge_ort_nonrecursive_internal(struct merge_options *opt,
3486                                             struct tree *merge_base,
3487                                             struct tree *side1,
3488                                             struct tree *side2,
3489                                             struct merge_result *result)
3490 {
3491         struct object_id working_tree_oid;
3492
3493         trace2_region_enter("merge", "collect_merge_info", opt->repo);
3494         if (collect_merge_info(opt, merge_base, side1, side2) != 0) {
3495                 /*
3496                  * TRANSLATORS: The %s arguments are: 1) tree hash of a merge
3497                  * base, and 2-3) the trees for the two trees we're merging.
3498                  */
3499                 err(opt, _("collecting merge info failed for trees %s, %s, %s"),
3500                     oid_to_hex(&merge_base->object.oid),
3501                     oid_to_hex(&side1->object.oid),
3502                     oid_to_hex(&side2->object.oid));
3503                 result->clean = -1;
3504                 return;
3505         }
3506         trace2_region_leave("merge", "collect_merge_info", opt->repo);
3507
3508         trace2_region_enter("merge", "renames", opt->repo);
3509         result->clean = detect_and_process_renames(opt, merge_base,
3510                                                    side1, side2);
3511         trace2_region_leave("merge", "renames", opt->repo);
3512
3513         trace2_region_enter("merge", "process_entries", opt->repo);
3514         process_entries(opt, &working_tree_oid);
3515         trace2_region_leave("merge", "process_entries", opt->repo);
3516
3517         /* Set return values */
3518         result->tree = parse_tree_indirect(&working_tree_oid);
3519         /* existence of conflicted entries implies unclean */
3520         result->clean &= strmap_empty(&opt->priv->conflicted);
3521         if (!opt->priv->call_depth) {
3522                 result->priv = opt->priv;
3523                 opt->priv = NULL;
3524         }
3525 }
3526
3527 /*
3528  * Originally from merge_recursive_internal(); somewhat adapted, though.
3529  */
3530 static void merge_ort_internal(struct merge_options *opt,
3531                                struct commit_list *merge_bases,
3532                                struct commit *h1,
3533                                struct commit *h2,
3534                                struct merge_result *result)
3535 {
3536         struct commit_list *iter;
3537         struct commit *merged_merge_bases;
3538         const char *ancestor_name;
3539         struct strbuf merge_base_abbrev = STRBUF_INIT;
3540
3541         if (!merge_bases) {
3542                 merge_bases = get_merge_bases(h1, h2);
3543                 /* See merge-ort.h:merge_incore_recursive() declaration NOTE */
3544                 merge_bases = reverse_commit_list(merge_bases);
3545         }
3546
3547         merged_merge_bases = pop_commit(&merge_bases);
3548         if (merged_merge_bases == NULL) {
3549                 /* if there is no common ancestor, use an empty tree */
3550                 struct tree *tree;
3551
3552                 tree = lookup_tree(opt->repo, opt->repo->hash_algo->empty_tree);
3553                 merged_merge_bases = make_virtual_commit(opt->repo, tree,
3554                                                          "ancestor");
3555                 ancestor_name = "empty tree";
3556         } else if (merge_bases) {
3557                 ancestor_name = "merged common ancestors";
3558         } else {
3559                 strbuf_add_unique_abbrev(&merge_base_abbrev,
3560                                          &merged_merge_bases->object.oid,
3561                                          DEFAULT_ABBREV);
3562                 ancestor_name = merge_base_abbrev.buf;
3563         }
3564
3565         for (iter = merge_bases; iter; iter = iter->next) {
3566                 const char *saved_b1, *saved_b2;
3567                 struct commit *prev = merged_merge_bases;
3568
3569                 opt->priv->call_depth++;
3570                 /*
3571                  * When the merge fails, the result contains files
3572                  * with conflict markers. The cleanness flag is
3573                  * ignored (unless indicating an error), it was never
3574                  * actually used, as result of merge_trees has always
3575                  * overwritten it: the committed "conflicts" were
3576                  * already resolved.
3577                  */
3578                 saved_b1 = opt->branch1;
3579                 saved_b2 = opt->branch2;
3580                 opt->branch1 = "Temporary merge branch 1";
3581                 opt->branch2 = "Temporary merge branch 2";
3582                 merge_ort_internal(opt, NULL, prev, iter->item, result);
3583                 if (result->clean < 0)
3584                         return;
3585                 opt->branch1 = saved_b1;
3586                 opt->branch2 = saved_b2;
3587                 opt->priv->call_depth--;
3588
3589                 merged_merge_bases = make_virtual_commit(opt->repo,
3590                                                          result->tree,
3591                                                          "merged tree");
3592                 commit_list_insert(prev, &merged_merge_bases->parents);
3593                 commit_list_insert(iter->item,
3594                                    &merged_merge_bases->parents->next);
3595
3596                 clear_or_reinit_internal_opts(opt->priv, 1);
3597         }
3598
3599         opt->ancestor = ancestor_name;
3600         merge_ort_nonrecursive_internal(opt,
3601                                         repo_get_commit_tree(opt->repo,
3602                                                              merged_merge_bases),
3603                                         repo_get_commit_tree(opt->repo, h1),
3604                                         repo_get_commit_tree(opt->repo, h2),
3605                                         result);
3606         strbuf_release(&merge_base_abbrev);
3607         opt->ancestor = NULL;  /* avoid accidental re-use of opt->ancestor */
3608 }
3609
3610 void merge_incore_nonrecursive(struct merge_options *opt,
3611                                struct tree *merge_base,
3612                                struct tree *side1,
3613                                struct tree *side2,
3614                                struct merge_result *result)
3615 {
3616         trace2_region_enter("merge", "incore_nonrecursive", opt->repo);
3617
3618         trace2_region_enter("merge", "merge_start", opt->repo);
3619         assert(opt->ancestor != NULL);
3620         merge_start(opt, result);
3621         trace2_region_leave("merge", "merge_start", opt->repo);
3622
3623         merge_ort_nonrecursive_internal(opt, merge_base, side1, side2, result);
3624         trace2_region_leave("merge", "incore_nonrecursive", opt->repo);
3625 }
3626
3627 void merge_incore_recursive(struct merge_options *opt,
3628                             struct commit_list *merge_bases,
3629                             struct commit *side1,
3630                             struct commit *side2,
3631                             struct merge_result *result)
3632 {
3633         trace2_region_enter("merge", "incore_recursive", opt->repo);
3634
3635         /* We set the ancestor label based on the merge_bases */
3636         assert(opt->ancestor == NULL);
3637
3638         trace2_region_enter("merge", "merge_start", opt->repo);
3639         merge_start(opt, result);
3640         trace2_region_leave("merge", "merge_start", opt->repo);
3641
3642         merge_ort_internal(opt, merge_bases, side1, side2, result);
3643         trace2_region_leave("merge", "incore_recursive", opt->repo);
3644 }