Status update on merge-recursive in C
authorJohannes Schindelin <Johannes.Schindelin@gmx.de>
Sat, 8 Jul 2006 16:42:41 +0000 (18:42 +0200)
committerJunio C Hamano <junkio@cox.net>
Fri, 14 Jul 2006 06:10:19 +0000 (23:10 -0700)
This is just an update for people being interested. Alex and me were
busy with that project for a few days now. While it has progressed nicely,
there are quite a couple TODOs in merge-recursive.c, just search for "TODO".

For impatient people: yes, it passes all the tests, and yes, according
to the evil test Alex did, it is faster than the Python script.

But no, it is not yet finished. Biggest points are:

- there are still three external calls
- in the end, it should not be necessary to write the index more than once
  (just before exiting)
- a lot of things can be refactored to make the code easier and shorter

BTW we cannot just plug in git-merge-tree yet, because git-merge-tree
does not handle renames at all.

This patch is meant for testing, and as such,

- it compile the program to git-merge-recur
- it adjusts the scripts and tests to use git-merge-recur instead of
  git-merge-recursive
- it provides "TEST", a script to execute the tests regarding -recursive
- it inlines the changes to read-cache.c (read_cache_from(), discard_cache()
  and refresh_cache_entry())

Brought to you by Alex Riesen and Dscho

Signed-off-by: Junio C Hamano <junkio@cox.net>
Makefile
TEST [new file with mode: 0755]
cache.h
git-merge.sh
git-rebase.sh
merge-recursive.c [new file with mode: 0644]
path-list.c [new file with mode: 0644]
path-list.h [new file with mode: 0644]
read-cache.c
t/t3402-rebase-merge.sh

index 01fb9cfdbd2bc8c53366d1db6f9b7b7525823fa5..a749aa4f0e0c955751b61ad1786302936758ea70 100644 (file)
--- a/Makefile
+++ b/Makefile
@@ -167,7 +167,8 @@ PROGRAMS = \
        git-upload-pack$X git-verify-pack$X \
        git-symbolic-ref$X \
        git-name-rev$X git-pack-redundant$X git-repo-config$X git-var$X \
-       git-describe$X git-merge-tree$X git-blame$X git-imap-send$X
+       git-describe$X git-merge-tree$X git-blame$X git-imap-send$X \
+       git-merge-recur$X
 
 BUILT_INS = git-log$X git-whatchanged$X git-show$X git-update-ref$X \
        git-count-objects$X git-diff$X git-push$X git-mailsplit$X \
@@ -615,6 +616,11 @@ git-http-push$X: revision.o http.o http-push.o $(GITLIBS)
        $(CC) $(ALL_CFLAGS) -o $@ $(ALL_LDFLAGS) $(filter %.o,$^) \
                $(LIBS) $(CURL_LIBCURL) $(EXPAT_LIBEXPAT)
 
+merge-recursive.o path-list.o: path-list.h
+git-merge-recur$X: merge-recursive.o path-list.o $(LIB_FILE)
+       $(CC) $(ALL_CFLAGS) -o $@ $(ALL_LDFLAGS) $(filter %.o,$^) \
+               $(LIBS)
+
 $(LIB_OBJS) $(BUILTIN_OBJS): $(LIB_H)
 $(patsubst git-%$X,%.o,$(PROGRAMS)): $(LIB_H) $(wildcard */*.h)
 $(DIFF_OBJS): diffcore.h
diff --git a/TEST b/TEST
new file mode 100755 (executable)
index 0000000..d530983
--- /dev/null
+++ b/TEST
@@ -0,0 +1,10 @@
+#!/bin/sh -x
+cd t || exit
+./t3400-rebase.sh              "$@" && \
+./t6020-merge-df.sh            "$@" && \
+./t3401-rebase-partial.sh      "$@" && \
+./t6021-merge-criss-cross.sh   "$@" && \
+./t3402-rebase-merge.sh        "$@" && \
+./t6022-merge-rename.sh        "$@" && \
+./t6010-merge-base.sh          "$@" && \
+:
diff --git a/cache.h b/cache.h
index d433d46f23c956aa13b9f0d97344f792234e94ab..8cc0ccbd0fa81eb72161b77e82f5fda201fdbdd9 100644 (file)
--- a/cache.h
+++ b/cache.h
@@ -115,6 +115,7 @@ static inline unsigned int create_ce_mode(unsigned int mode)
 extern struct cache_entry **active_cache;
 extern unsigned int active_nr, active_alloc, active_cache_changed;
 extern struct cache_tree *active_cache_tree;
+extern int cache_errno;
 
 #define GIT_DIR_ENVIRONMENT "GIT_DIR"
 #define DEFAULT_GIT_DIR_ENVIRONMENT ".git"
@@ -142,13 +143,16 @@ extern void verify_non_filename(const char *prefix, const char *name);
 
 /* Initialize and use the cache information */
 extern int read_cache(void);
+extern int read_cache_from(const char *path);
 extern int write_cache(int newfd, struct cache_entry **cache, int entries);
+extern int discard_cache(void);
 extern int verify_path(const char *path);
 extern int cache_name_pos(const char *name, int namelen);
 #define ADD_CACHE_OK_TO_ADD 1          /* Ok to add */
 #define ADD_CACHE_OK_TO_REPLACE 2      /* Ok to replace file/directory */
 #define ADD_CACHE_SKIP_DFCHECK 4       /* Ok to skip DF conflict checks */
 extern int add_cache_entry(struct cache_entry *ce, int option);
+extern struct cache_entry *refresh_cache_entry(struct cache_entry *ce, int really);
 extern int remove_cache_entry_at(int pos);
 extern int remove_file_from_cache(const char *path);
 extern int ce_same_name(struct cache_entry *a, struct cache_entry *b);
index 24e3b507ef1f907e7146a2653189f45a92db08fd..b26ca147c084c963d03effd2462244635f10e1fa 100755 (executable)
@@ -9,15 +9,15 @@ USAGE='[-n] [--no-commit] [--squash] [-s <strategy>]... <merge-message> <head> <
 LF='
 '
 
-all_strategies='recursive octopus resolve stupid ours'
-default_twohead_strategies='recursive'
+all_strategies='recur recur octopus resolve stupid ours'
+default_twohead_strategies='recur'
 default_octopus_strategies='octopus'
 no_trivial_merge_strategies='ours'
 use_strategies=
 
 index_merge=t
 if test "@@NO_PYTHON@@"; then
-       all_strategies='resolve octopus stupid ours'
+       all_strategies='recur resolve octopus stupid ours'
        default_twohead_strategies='resolve'
 fi
 
index 1b9e986926cf5b7d25afc6e0dc81ff9fc56c370a..2a4c8c8a897e1d7f8220601892fd92514e09c65a 100755 (executable)
@@ -35,7 +35,7 @@ If you would prefer to skip this patch, instead run \"git rebase --skip\".
 To restore the original branch and stop rebasing run \"git rebase --abort\".
 "
 unset newbase
-strategy=recursive
+strategy=recur
 do_merge=
 dotest=$GIT_DIR/.dotest-merge
 prec=4
@@ -292,7 +292,7 @@ then
        exit $?
 fi
 
-if test "@@NO_PYTHON@@" && test "$strategy" = "recursive"
+if test "@@NO_PYTHON@@" && test "$strategy" = "recur"
 then
        die 'The recursive merge strategy currently relies on Python,
 which this installation of git was not configured with.  Please consider
diff --git a/merge-recursive.c b/merge-recursive.c
new file mode 100644 (file)
index 0000000..cf81768
--- /dev/null
@@ -0,0 +1,1560 @@
+/*
+ * Recursive Merge algorithm stolen from git-merge-recursive.py by
+ * Fredrik Kuivinen.
+ * The thieves were Alex Riesen and Johannes Schindelin, in June/July 2006
+ */
+#include <stdarg.h>
+#include <string.h>
+#include <assert.h>
+#include <sys/wait.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <time.h>
+#include "cache.h"
+#include "cache-tree.h"
+#include "commit.h"
+#include "blob.h"
+#include "tree-walk.h"
+#include "diff.h"
+#include "diffcore.h"
+#include "run-command.h"
+#include "tag.h"
+
+#include "path-list.h"
+
+/*#define DEBUG*/
+
+#ifdef DEBUG
+#define debug(args, ...) fprintf(stderr, args, ## __VA_ARGS__)
+#else
+#define debug(args, ...)
+#endif
+
+#ifdef DEBUG
+#include "quote.h"
+static void show_ce_entry(const char *tag, struct cache_entry *ce)
+{
+       if (tag && *tag &&
+           (ce->ce_flags & htons(CE_VALID))) {
+               static char alttag[4];
+               memcpy(alttag, tag, 3);
+               if (isalpha(tag[0]))
+                       alttag[0] = tolower(tag[0]);
+               else if (tag[0] == '?')
+                       alttag[0] = '!';
+               else {
+                       alttag[0] = 'v';
+                       alttag[1] = tag[0];
+                       alttag[2] = ' ';
+                       alttag[3] = 0;
+               }
+               tag = alttag;
+       }
+
+       fprintf(stderr,"%s%06o %s %d\t",
+                       tag,
+                       ntohl(ce->ce_mode),
+                       sha1_to_hex(ce->sha1),
+                       ce_stage(ce));
+       write_name_quoted("", 0, ce->name,
+                       '\n', stderr);
+       fputc('\n', stderr);
+}
+
+static void ls_files() {
+       int i;
+       for (i = 0; i < active_nr; i++) {
+               struct cache_entry *ce = active_cache[i];
+               show_ce_entry("", ce);
+       }
+       fprintf(stderr, "---\n");
+}
+#endif
+
+/*
+ * A virtual commit has
+ * - (const char *)commit->util set to the name, and
+ * - *(int *)commit->object.sha1 set to the virtual id.
+ */
+static const char *commit_title(struct commit *commit, int *len)
+{
+       const char *s = "(null commit)";
+       *len = strlen(s);
+
+       if ( commit->util ) {
+               s = commit->util;
+               *len = strlen(s);
+       } else {
+               if ( parse_commit(commit) != 0 ) {
+                       s = "(bad commit)";
+                       *len = strlen(s);
+               } else {
+                       s = commit->buffer;
+                       char prev = '\0';
+                       while ( *s ) {
+                               if ( '\n' == prev && '\n' == *s ) {
+                                       ++s;
+                                       break;
+                               }
+                               prev = *s++;
+                       }
+                       *len = 0;
+                       while ( s[*len] && '\n' != s[*len] )
+                               ++(*len);
+               }
+       }
+       return s;
+}
+
+static const char *commit_hex_sha1(const struct commit *commit)
+{
+       return commit->util ? "virtual" : commit ?
+               sha1_to_hex(commit->object.sha1) : "undefined";
+}
+
+static unsigned commit_list_count(const struct commit_list *l)
+{
+       unsigned c = 0;
+       for (; l; l = l->next )
+               c++;
+       return c;
+}
+
+static struct commit *make_virtual_commit(struct tree *tree, const char *comment)
+{
+       struct commit *commit = xcalloc(1, sizeof(struct commit));
+       static unsigned virtual_id = 1;
+       commit->tree = tree;
+       commit->util = (void*)comment;
+       *(int*)commit->object.sha1 = virtual_id++;
+       return commit;
+}
+
+/*
+ * TODO: we should not have to copy the SHA1s around, but rather reference
+ * them. That way, sha_eq() is just sha1 == sha2.
+ */
+static int sha_eq(const unsigned char *a, const unsigned char *b)
+{
+       if ( !a && !b )
+               return 2;
+       return a && b && memcmp(a, b, 20) == 0;
+}
+
+static void memswp(void *p1, void *p2, unsigned n)
+{
+       unsigned char *a = p1, *b = p2;
+       while ( n-- ) {
+               *a ^= *b;
+               *b ^= *a;
+               *a ^= *b;
+               ++a;
+               ++b;
+       }
+}
+
+/*
+ * TODO: we should convert the merge_result users to
+ *     int blabla(..., struct commit **result)
+ * like everywhere else in git.
+ * Same goes for merge_tree_result and merge_file_info.
+ */
+struct merge_result
+{
+       struct commit *commit;
+       unsigned clean:1;
+};
+
+struct merge_tree_result
+{
+       struct tree *tree;
+       unsigned clean:1;
+};
+
+/*
+ * TODO: check if we can just reuse the active_cache structure: it is already
+ * sorted (by name, stage).
+ * Only problem: do not write it when flushing the cache.
+ */
+struct stage_data
+{
+       struct
+       {
+               unsigned mode;
+               unsigned char sha[20];
+       } stages[4];
+       unsigned processed:1;
+};
+
+static struct path_list currentFileSet = {NULL, 0, 0, 1};
+static struct path_list currentDirectorySet = {NULL, 0, 0, 1};
+
+static int output_indent = 0;
+
+static void output(const char *fmt, ...)
+{
+       va_list args;
+       int i;
+       for ( i = output_indent; i--; )
+               fputs("  ", stdout);
+       va_start(args, fmt);
+       vfprintf(stdout, fmt, args);
+       va_end(args);
+       fputc('\n', stdout);
+}
+
+static const char *original_index_file;
+static const char *temporary_index_file;
+static int cache_dirty = 0;
+
+static int flush_cache()
+{
+       /* flush temporary index */
+       struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
+       int fd = hold_lock_file_for_update(lock, getenv("GIT_INDEX_FILE"));
+       if (fd < 0)
+               die("could not lock %s", temporary_index_file);
+       if (write_cache(fd, active_cache, active_nr) ||
+                       close(fd) || commit_lock_file(lock))
+               die ("unable to write %s", getenv("GIT_INDEX_FILE"));
+       discard_cache();
+       cache_dirty = 0;
+       return 0;
+}
+
+static void setup_index(int temp)
+{
+       const char *idx = temp ? temporary_index_file: original_index_file;
+       if (cache_dirty)
+               die("fatal: cache changed flush_cache();");
+       unlink(temporary_index_file);
+       setenv("GIT_INDEX_FILE", idx, 1);
+       discard_cache();
+}
+
+static struct cache_entry *make_cache_entry(unsigned int mode,
+               const unsigned char *sha1, const char *path, int stage, int refresh)
+{
+       int size, len;
+       struct cache_entry *ce;
+
+       if (!verify_path(path))
+               return NULL;
+
+       len = strlen(path);
+       size = cache_entry_size(len);
+       ce = xcalloc(1, size);
+
+       memcpy(ce->sha1, sha1, 20);
+       memcpy(ce->name, path, len);
+       ce->ce_flags = create_ce_flags(len, stage);
+       ce->ce_mode = create_ce_mode(mode);
+
+       if (refresh)
+               return refresh_cache_entry(ce, 0);
+
+       return ce;
+}
+
+static int add_cacheinfo(unsigned int mode, const unsigned char *sha1,
+               const char *path, int stage, int refresh, int options)
+{
+       struct cache_entry *ce;
+       if (!cache_dirty)
+               read_cache_from(getenv("GIT_INDEX_FILE"));
+       cache_dirty++;
+       ce = make_cache_entry(mode, sha1 ? sha1 : null_sha1, path, stage, refresh);
+       if (!ce)
+               return error("cache_addinfo failed: %s", strerror(cache_errno));
+       return add_cache_entry(ce, options);
+}
+
+/*
+ * This is a global variable which is used in a number of places but
+ * only written to in the 'merge' function.
+ *
+ * index_only == 1    => Don't leave any non-stage 0 entries in the cache and
+ *                       don't update the working directory.
+ *               0    => Leave unmerged entries in the cache and update
+ *                       the working directory.
+ */
+static int index_only = 0;
+
+/*
+ * TODO: this can be streamlined by refactoring builtin-read-tree.c
+ */
+static int git_read_tree(const struct tree *tree)
+{
+#if 0
+       fprintf(stderr, "GIT_INDEX_FILE='%s' git-read-tree %s\n",
+               getenv("GIT_INDEX_FILE"),
+               sha1_to_hex(tree->object.sha1));
+#endif
+       const char *argv[] = { "git-read-tree", NULL, NULL, };
+       if (cache_dirty)
+               die("read-tree with dirty cache");
+       argv[1] = sha1_to_hex(tree->object.sha1);
+       int rc = run_command_v(2, argv);
+       return rc < 0 ? -1: rc;
+}
+
+/*
+ * TODO: this can be streamlined by refactoring builtin-read-tree.c
+ */
+static int git_merge_trees(const char *update_arg,
+                          struct tree *common,
+                          struct tree *head,
+                          struct tree *merge)
+{
+#if 0
+       fprintf(stderr, "GIT_INDEX_FILE='%s' git-read-tree %s -m %s %s %s\n",
+               getenv("GIT_INDEX_FILE"),
+               update_arg,
+               sha1_to_hex(common->object.sha1),
+               sha1_to_hex(head->object.sha1),
+               sha1_to_hex(merge->object.sha1));
+#endif
+       const char *argv[] = {
+               "git-read-tree", NULL, "-m", NULL, NULL, NULL,
+               NULL,
+       };
+       if (cache_dirty)
+               flush_cache();
+       argv[1] = update_arg;
+       argv[3] = sha1_to_hex(common->object.sha1);
+       argv[4] = sha1_to_hex(head->object.sha1);
+       argv[5] = sha1_to_hex(merge->object.sha1);
+       int rc = run_command_v(6, argv);
+       return rc < 0 ? -1: rc;
+}
+
+/*
+ * TODO: this can be streamlined by refactoring builtin-write-tree.c
+ */
+static struct tree *git_write_tree()
+{
+#if 0
+       fprintf(stderr, "GIT_INDEX_FILE='%s' git-write-tree\n",
+               getenv("GIT_INDEX_FILE"));
+#endif
+       if (cache_dirty)
+               flush_cache();
+       FILE *fp = popen("git-write-tree 2>/dev/null", "r");
+       char buf[41];
+       unsigned char sha1[20];
+       int ch;
+       unsigned i = 0;
+       while ( (ch = fgetc(fp)) != EOF )
+               if ( i < sizeof(buf)-1 && ch >= '0' && ch <= 'f' )
+                       buf[i++] = ch;
+               else
+                       break;
+       int rc = pclose(fp);
+       if ( rc == -1 || WEXITSTATUS(rc) )
+               return NULL;
+       buf[i] = '\0';
+       if ( get_sha1(buf, sha1) != 0 )
+               return NULL;
+       return lookup_tree(sha1);
+}
+
+/*
+ * TODO: get rid of files_and_dirs; we do not use it except for
+ * current_file_set and current_dir_set, which are global already.
+ */
+static struct
+{
+       struct path_list *files;
+       struct path_list *dirs;
+} files_and_dirs;
+
+static int save_files_dirs(const unsigned char *sha1,
+               const char *base, int baselen, const char *path,
+               unsigned int mode, int stage)
+{
+       int len = strlen(path);
+       char *newpath = malloc(baselen + len + 1);
+       memcpy(newpath, base, baselen);
+       memcpy(newpath + baselen, path, len);
+       newpath[baselen + len] = '\0';
+
+       if (S_ISDIR(mode))
+               path_list_insert(newpath, files_and_dirs.dirs);
+       else
+               path_list_insert(newpath, files_and_dirs.files);
+       free(newpath);
+
+       return READ_TREE_RECURSIVE;
+}
+
+static int get_files_dirs(struct tree *tree,
+                         struct path_list *files,
+                         struct path_list *dirs)
+{
+       int n;
+       files_and_dirs.files = files;
+       files_and_dirs.dirs = dirs;
+       debug("get_files_dirs ...\n");
+       if (read_tree_recursive(tree, "", 0, 0, NULL, save_files_dirs) != 0) {
+               debug("  get_files_dirs done (0)\n");
+               return 0;
+       }
+       n = files->nr + dirs->nr;
+       debug("  get_files_dirs done (%d)\n", n);
+       return n;
+}
+
+/*
+ * TODO: this wrapper is so small, we can use path_list_lookup directly.
+ * Same goes for index_entry_get(), free_index_entries(), find_rename_bysrc(),
+ * free_rename_entries().
+ */
+static struct stage_data *index_entry_find(struct path_list *ents,
+                                           const char *path)
+{
+       struct path_list_item *item = path_list_lookup(path, ents);
+       if (item)
+               return item->util;
+       return NULL;
+}
+
+static struct stage_data *index_entry_get(struct path_list *ents,
+                                          const char *path)
+{
+       struct path_list_item *item = path_list_lookup(path, ents);
+
+       if (item == NULL) {
+               item = path_list_insert(path, ents);
+               item->util = xcalloc(1, sizeof(struct stage_data));
+       }
+       return item->util;
+}
+
+/*
+ * TODO: since the result of index_entry_from_db() is tucked into a
+ * path_list anyway, this helper can do that already.
+ */
+/*
+ * Returns a index_entry instance which doesn't have to correspond to
+ * a real cache entry in Git's index.
+ */
+static struct stage_data *index_entry_from_db(const char *path,
+                                              struct tree *o,
+                                              struct tree *a,
+                                              struct tree *b)
+{
+       struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
+       get_tree_entry(o->object.sha1, path,
+                       e->stages[1].sha, &e->stages[1].mode);
+       get_tree_entry(a->object.sha1, path,
+                       e->stages[2].sha, &e->stages[2].mode);
+       get_tree_entry(b->object.sha1, path,
+                       e->stages[3].sha, &e->stages[3].mode);
+       return e;
+}
+
+static void free_index_entries(struct path_list **ents)
+{
+       if (!*ents)
+               return;
+
+       path_list_clear(*ents, 1);
+       free(*ents);
+       *ents = NULL;
+}
+
+/*
+ * Create a dictionary mapping file names to CacheEntry objects. The
+ * dictionary contains one entry for every path with a non-zero stage entry.
+ */
+static struct path_list *get_unmerged()
+{
+       struct path_list *unmerged = xcalloc(1, sizeof(struct path_list));
+       int i;
+
+       unmerged->strdup_paths = 1;
+       if (!cache_dirty) {
+               read_cache_from(getenv("GIT_INDEX_FILE"));
+               cache_dirty++;
+       }
+       for (i = 0; i < active_nr; i++) {
+               struct cache_entry *ce = active_cache[i];
+               if (!ce_stage(ce))
+                       continue;
+
+               struct stage_data *e = index_entry_get(unmerged, ce->name);
+               e->stages[ce_stage(ce)].mode = ntohl(ce->ce_mode);
+               memcpy(e->stages[ce_stage(ce)].sha, ce->sha1, 20);
+       }
+
+       debug("  get_unmerged done\n");
+       return unmerged;
+}
+
+struct rename
+{
+       struct diff_filepair *pair;
+       struct stage_data *src_entry;
+       struct stage_data *dst_entry;
+       unsigned processed:1;
+};
+
+static struct rename *find_rename_bysrc(struct path_list *e,
+                                             const char *name)
+{
+       struct path_list_item *item = path_list_lookup(name, e);
+       if (item)
+               return item->util;
+       return NULL;
+}
+
+static void free_rename_entries(struct path_list **list)
+{
+       if (!*list)
+               return;
+
+       path_list_clear(*list, 0);
+       free(*list);
+       *list = NULL;
+}
+
+/*
+ * Get information of all renames which occured between 'oTree' and
+ * 'tree'. We need the three trees in the merge ('oTree', 'aTree' and
+ * 'bTree') to be able to associate the correct cache entries with
+ * the rename information. 'tree' is always equal to either aTree or bTree.
+ */
+static struct path_list *get_renames(struct tree *tree,
+                                       struct tree *oTree,
+                                       struct tree *aTree,
+                                       struct tree *bTree,
+                                       struct path_list *entries)
+{
+#ifdef DEBUG
+       time_t t = time(0);
+       debug("getRenames ...\n");
+#endif
+       int i;
+       struct path_list *renames = xcalloc(1, sizeof(struct path_list));
+       struct diff_options opts;
+       diff_setup(&opts);
+       opts.recursive = 1;
+       opts.detect_rename = DIFF_DETECT_RENAME;
+       opts.output_format = DIFF_FORMAT_NO_OUTPUT;
+       if (diff_setup_done(&opts) < 0)
+               die("diff setup failed");
+       diff_tree_sha1(oTree->object.sha1, tree->object.sha1, "", &opts);
+       diffcore_std(&opts);
+       for (i = 0; i < diff_queued_diff.nr; ++i) {
+               struct rename *re;
+               struct diff_filepair *pair = diff_queued_diff.queue[i];
+               if (pair->status != 'R') {
+                       diff_free_filepair(pair);
+                       continue;
+               }
+               re = xmalloc(sizeof(*re));
+               re->processed = 0;
+               re->pair = pair;
+               re->src_entry = index_entry_find(entries, re->pair->one->path);
+               /* TODO: should it not be an error, if src_entry was found? */
+               if ( !re->src_entry ) {
+                       re->src_entry = index_entry_from_db(re->pair->one->path,
+                                       oTree, aTree, bTree);
+                       struct path_list_item *item =
+                               path_list_insert(re->pair->one->path, entries);
+                       item->util = re->src_entry;
+               }
+               re->dst_entry = index_entry_find(entries, re->pair->two->path);
+               if ( !re->dst_entry ) {
+                       re->dst_entry = index_entry_from_db(re->pair->two->path,
+                                       oTree, aTree, bTree);
+                       struct path_list_item *item =
+                               path_list_insert(re->pair->two->path, entries);
+                       item->util = re->dst_entry;
+               }
+               struct path_list_item *item = path_list_insert(pair->one->path, renames);
+               item->util = re;
+       }
+       opts.output_format = DIFF_FORMAT_NO_OUTPUT;
+       diff_queued_diff.nr = 0;
+       diff_flush(&opts);
+       debug("  getRenames done in %ld\n", time(0)-t);
+       return renames;
+}
+
+/*
+ * TODO: the code would be way nicer, if we had a struct containing just sha1 and mode.
+ * In this particular case, we might get away reusing stage_data, no?
+ */
+int update_stages(const char *path,
+                  unsigned char *osha, unsigned omode,
+                  unsigned char *asha, unsigned amode,
+                  unsigned char *bsha, unsigned bmode,
+                  int clear /* =True */)
+{
+       int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_OK_TO_REPLACE;
+       if ( clear )
+               if (add_cacheinfo(0, null_sha1, path, 0, 0, options))
+                       return -1;
+       if ( omode )
+               if (add_cacheinfo(omode, osha, path, 1, 0, options))
+                       return -1;
+       if ( amode )
+               if (add_cacheinfo(omode, osha, path, 2, 0, options))
+                       return -1;
+       if ( bmode )
+               if (add_cacheinfo(omode, osha, path, 3, 0, options))
+                       return -1;
+       return 0;
+}
+
+/*
+ * TODO: there has to be a function in libgit doing this exact thing.
+ */
+static int remove_path(const char *name)
+{
+       int ret;
+       char *slash;
+
+       ret = unlink(name);
+       if ( ret )
+               return ret;
+       int len = strlen(name);
+       char *dirs = malloc(len+1);
+       memcpy(dirs, name, len);
+       dirs[len] = '\0';
+       while ( (slash = strrchr(name, '/')) ) {
+               *slash = '\0';
+               len = slash - name;
+               if ( rmdir(name) != 0 )
+                       break;
+       }
+       free(dirs);
+       return ret;
+}
+
+/* General TODO: unC99ify the code: no declaration after code */
+/* General TODO: no javaIfiCation: rename updateCache to update_cache */
+/*
+ * TODO: once we no longer call external programs, we'd probably be better of
+ * not setting / getting the environment variable GIT_INDEX_FILE all the time.
+ */
+int remove_file(int clean, const char *path)
+{
+       int updateCache = index_only || clean;
+       int updateWd = !index_only;
+
+       if ( updateCache ) {
+               if (!cache_dirty)
+                       read_cache_from(getenv("GIT_INDEX_FILE"));
+               cache_dirty++;
+               if (remove_file_from_cache(path))
+                       return -1;
+       }
+       if ( updateWd )
+       {
+               unlink(path);
+               if ( errno != ENOENT || errno != EISDIR )
+                       return -1;
+               remove_path(path);
+       }
+       return 0;
+}
+
+static char *unique_path(const char *path, const char *branch)
+{
+       char *newpath = xmalloc(strlen(path) + 1 + strlen(branch) + 8 + 1);
+       strcpy(newpath, path);
+       strcat(newpath, "~");
+       char *p = newpath + strlen(newpath);
+       strcpy(p, branch);
+       for ( ; *p; ++p )
+               if ( '/' == *p )
+                       *p = '_';
+       int suffix = 0;
+       struct stat st;
+       while ( path_list_has_path(&currentFileSet, newpath) ||
+               path_list_has_path(&currentDirectorySet, newpath) ||
+               lstat(newpath, &st) == 0 ) {
+               sprintf(p, "_%d", suffix++);
+       }
+       path_list_insert(newpath, &currentFileSet);
+       return newpath;
+}
+
+/*
+ * TODO: except for create_last, this so looks like
+ * safe_create_leading_directories().
+ */
+static int mkdir_p(const char *path, unsigned long mode, int create_last)
+{
+       char *buf = strdup(path);
+       char *p;
+
+       for ( p = buf; *p; ++p ) {
+               if ( *p != '/' )
+                       continue;
+               *p = '\0';
+               if (mkdir(buf, mode)) {
+                       int e = errno;
+                       if ( e == EEXIST ) {
+                               struct stat st;
+                               if ( !stat(buf, &st) && S_ISDIR(st.st_mode) )
+                                       goto next; /* ok */
+                               errno = e;
+                       }
+                       free(buf);
+                       return -1;
+               }
+       next:
+               *p = '/';
+       }
+       free(buf);
+       if ( create_last && mkdir(path, mode) )
+               return -1;
+       return 0;
+}
+
+static void flush_buffer(int fd, const char *buf, unsigned long size)
+{
+       while (size > 0) {
+               long ret = xwrite(fd, buf, size);
+               if (ret < 0) {
+                       /* Ignore epipe */
+                       if (errno == EPIPE)
+                               break;
+                       die("merge-recursive: %s", strerror(errno));
+               } else if (!ret) {
+                       die("merge-recursive: disk full?");
+               }
+               size -= ret;
+               buf += ret;
+       }
+}
+
+/* General TODO: reindent according to guide lines (no if ( blabla )) */
+void update_file_flags(const unsigned char *sha,
+                  unsigned mode,
+                  const char *path,
+                  int updateCache,
+                  int updateWd)
+{
+       if ( index_only )
+               updateWd = 0;
+
+       if ( updateWd ) {
+               char type[20];
+               void *buf;
+               unsigned long size;
+
+               buf = read_sha1_file(sha, type, &size);
+               if (!buf)
+                       die("cannot read object %s '%s'", sha1_to_hex(sha), path);
+               if ( strcmp(type, blob_type) != 0 )
+                       die("blob expected for %s '%s'", sha1_to_hex(sha), path);
+
+               if ( S_ISREG(mode) ) {
+                       if ( mkdir_p(path, 0777, 0 /* don't create last element */) )
+                               die("failed to create path %s: %s", path, strerror(errno));
+                       unlink(path);
+                       if ( mode & 0100 )
+                               mode = 0777;
+                       else
+                               mode = 0666;
+                       int fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
+                       if ( fd < 0 )
+                               die("failed to open %s: %s", path, strerror(errno));
+                       flush_buffer(fd, buf, size);
+                       close(fd);
+               } else if ( S_ISLNK(mode) ) {
+                       char *linkTarget = malloc(size + 1);
+                       memcpy(linkTarget, buf, size);
+                       linkTarget[size] = '\0';
+                       mkdir_p(path, 0777, 0);
+                       symlink(linkTarget, path);
+               } else
+                       die("do not know what to do with %06o %s '%s'",
+                           mode, sha1_to_hex(sha), path);
+       }
+       if ( updateCache )
+               add_cacheinfo(mode, sha, path, 0, updateWd, ADD_CACHE_OK_TO_ADD);
+}
+
+/* TODO: is this often used? if not, do direct call */
+void update_file(int clean,
+               const unsigned char *sha,
+               unsigned mode,
+               const char *path)
+{
+       update_file_flags(sha, mode, path, index_only || clean, !index_only);
+}
+
+/* Low level file merging, update and removal */
+
+struct merge_file_info
+{
+       unsigned char sha[20];
+       unsigned mode;
+       unsigned clean:1,
+                merge:1;
+};
+
+static char *git_unpack_file(const unsigned char *sha1, char *path)
+{
+       void *buf;
+       char type[20];
+       unsigned long size;
+       int fd;
+
+       buf = read_sha1_file(sha1, type, &size);
+       if (!buf || strcmp(type, blob_type))
+               die("unable to read blob object %s", sha1_to_hex(sha1));
+
+       strcpy(path, ".merge_file_XXXXXX");
+       fd = mkstemp(path);
+       if (fd < 0)
+               die("unable to create temp-file");
+       flush_buffer(fd, buf, size);
+       close(fd);
+       return path;
+}
+
+/*
+ * TODO: the signature would be much more efficient using stage_data
+ */
+static struct merge_file_info merge_file(const char *oPath,
+                                        const unsigned char *oSha,
+                                        unsigned oMode,
+                                        const char *aPath,
+                                        const unsigned char *aSha,
+                                        unsigned aMode,
+                                        const char *bPath,
+                                        const unsigned char *bSha,
+                                        unsigned bMode,
+                                        const char *branch1Name,
+                                        const char *branch2Name)
+{
+       struct merge_file_info result;
+       result.merge = 0;
+       result.clean = 1;
+
+       if ( (S_IFMT & aMode) != (S_IFMT & bMode) ) {
+               result.clean = 0;
+               if ( S_ISREG(aMode) ) {
+                       result.mode = aMode;
+                       memcpy(result.sha, aSha, 20);
+               } else {
+                       result.mode = bMode;
+                       memcpy(result.sha, bSha, 20);
+               }
+       } else {
+               if ( memcmp(aSha, oSha, 20) != 0 && memcmp(bSha, oSha, 20) != 0 )
+                       result.merge = 1;
+
+               result.mode = aMode == oMode ? bMode: aMode;
+
+               if ( memcmp(aSha, oSha, 20) == 0 )
+                       memcpy(result.sha, bSha, 20);
+               else if ( memcmp(bSha, oSha, 20) == 0 )
+                       memcpy(result.sha, aSha, 20);
+               else if ( S_ISREG(aMode) ) {
+
+                       int code = 1;
+                       char orig[PATH_MAX];
+                       char src1[PATH_MAX];
+                       char src2[PATH_MAX];
+
+                       git_unpack_file(oSha, orig);
+                       git_unpack_file(aSha, src1);
+                       git_unpack_file(bSha, src2);
+
+                       const char *argv[] = {
+                               "merge", "-L", NULL, "-L", NULL, "-L", NULL,
+                               src1, orig, src2,
+                               NULL
+                       };
+                       char *la, *lb, *lo;
+                       argv[2] = la = strdup(mkpath("%s/%s", branch1Name, aPath));
+                       argv[6] = lb = strdup(mkpath("%s/%s", branch2Name, bPath));
+                       argv[4] = lo = strdup(mkpath("orig/%s", oPath));
+
+#if 0
+                       printf("%s %s %s %s %s %s %s %s %s %s\n",
+                              argv[0], argv[1], argv[2], argv[3], argv[4],
+                              argv[5], argv[6], argv[7], argv[8], argv[9]);
+#endif
+                       code = run_command_v(10, argv);
+
+                       free(la);
+                       free(lb);
+                       free(lo);
+                       if ( code && code < -256 ) {
+                               die("Failed to execute 'merge'. merge(1) is used as the "
+                                   "file-level merge tool. Is 'merge' in your path?");
+                       }
+                       struct stat st;
+                       int fd = open(src1, O_RDONLY);
+                       if (fd < 0 || fstat(fd, &st) < 0 ||
+                                       index_fd(result.sha, fd, &st, 1,
+                                               "blob"))
+                               die("Unable to add %s to database", src1);
+                       close(fd);
+
+                       unlink(orig);
+                       unlink(src1);
+                       unlink(src2);
+
+                       result.clean = WEXITSTATUS(code) == 0;
+               } else {
+                       if ( !(S_ISLNK(aMode) || S_ISLNK(bMode)) )
+                               die("cannot merge modes?");
+
+                       memcpy(result.sha, aSha, 20);
+
+                       if ( memcmp(aSha, bSha, 20) != 0 )
+                               result.clean = 0;
+               }
+       }
+
+       return result;
+}
+
+static void conflict_rename_rename(struct rename *ren1,
+                                  const char *branch1,
+                                  struct rename *ren2,
+                                  const char *branch2)
+{
+       char *del[2];
+       int delp = 0;
+       const char *ren1_dst = ren1->pair->two->path;
+       const char *ren2_dst = ren2->pair->two->path;
+       const char *dstName1 = ren1_dst;
+       const char *dstName2 = ren2_dst;
+       if (path_list_has_path(&currentDirectorySet, ren1_dst)) {
+               dstName1 = del[delp++] = unique_path(ren1_dst, branch1);
+               output("%s is a directory in %s adding as %s instead",
+                      ren1_dst, branch2, dstName1);
+               remove_file(0, ren1_dst);
+       }
+       if (path_list_has_path(&currentDirectorySet, ren2_dst)) {
+               dstName2 = del[delp++] = unique_path(ren2_dst, branch2);
+               output("%s is a directory in %s adding as %s instead",
+                      ren2_dst, branch1, dstName2);
+               remove_file(0, ren2_dst);
+       }
+       update_stages(dstName1,
+                     NULL, 0,
+                     ren1->pair->two->sha1, ren1->pair->two->mode,
+                     NULL, 0,
+                     1 /* clear */);
+       update_stages(dstName2,
+                     NULL, 0,
+                     NULL, 0,
+                     ren2->pair->two->sha1, ren2->pair->two->mode,
+                     1 /* clear */);
+       while ( delp-- )
+               free(del[delp]);
+}
+
+static void conflict_rename_dir(struct rename *ren1,
+                               const char *branch1)
+{
+       char *newPath = unique_path(ren1->pair->two->path, branch1);
+       output("Renaming %s to %s instead", ren1->pair->one->path, newPath);
+       remove_file(0, ren1->pair->two->path);
+       update_file(0, ren1->pair->two->sha1, ren1->pair->two->mode, newPath);
+       free(newPath);
+}
+
+static void conflict_rename_rename_2(struct rename *ren1,
+                                    const char *branch1,
+                                    struct rename *ren2,
+                                    const char *branch2)
+{
+       char *newPath1 = unique_path(ren1->pair->two->path, branch1);
+       char *newPath2 = unique_path(ren2->pair->two->path, branch2);
+       output("Renaming %s to %s and %s to %s instead",
+              ren1->pair->one->path, newPath1,
+              ren2->pair->one->path, newPath2);
+       remove_file(0, ren1->pair->two->path);
+       update_file(0, ren1->pair->two->sha1, ren1->pair->two->mode, newPath1);
+       update_file(0, ren2->pair->two->sha1, ren2->pair->two->mode, newPath2);
+       free(newPath2);
+       free(newPath1);
+}
+
+/* General TODO: get rid of all the debug messages */
+static int process_renames(struct path_list *renamesA,
+                          struct path_list *renamesB,
+                          const char *branchNameA,
+                          const char *branchNameB)
+{
+       int cleanMerge = 1, i;
+       struct path_list srcNames = {NULL, 0, 0, 0}, byDstA = {NULL, 0, 0, 0}, byDstB = {NULL, 0, 0, 0};
+       const struct rename *sre;
+
+       /*
+        * TODO: think about a saner way to do this.
+        * Since both renamesA and renamesB are sorted, it should
+        * be much more efficient to traverse both simultaneously,
+        * only byDstA and byDstB should be needed.
+        */
+       debug("processRenames...\n");
+       for (i = 0; i < renamesA->nr; i++) {
+               sre = renamesA->items[i].util;
+               path_list_insert(sre->pair->one->path, &srcNames);
+               path_list_insert(sre->pair->two->path, &byDstA)->util
+                       = sre->dst_entry;
+       }
+       for (i = 0; i < renamesB->nr; i++) {
+               sre = renamesB->items[i].util;
+               path_list_insert(sre->pair->one->path, &srcNames);
+               path_list_insert(sre->pair->two->path, &byDstB)->util
+                       = sre->dst_entry;
+       }
+
+       for (i = 0; i < srcNames.nr; i++) {
+               char *src = srcNames.items[i].path;
+               struct path_list *renames1, *renames2, *renames2Dst;
+               struct rename *ren1, *ren2;
+               const char *branchName1, *branchName2;
+               ren1 = find_rename_bysrc(renamesA, src);
+               ren2 = find_rename_bysrc(renamesB, src);
+               /* TODO: refactor, so that 1/2 are not needed */
+               if ( ren1 ) {
+                       renames1 = renamesA;
+                       renames2 = renamesB;
+                       renames2Dst = &byDstB;
+                       branchName1 = branchNameA;
+                       branchName2 = branchNameB;
+               } else {
+                       renames1 = renamesB;
+                       renames2 = renamesA;
+                       renames2Dst = &byDstA;
+                       branchName1 = branchNameB;
+                       branchName2 = branchNameA;
+                       struct rename *tmp = ren2;
+                       ren2 = ren1;
+                       ren1 = tmp;
+               }
+
+               ren1->dst_entry->processed = 1;
+               ren1->src_entry->processed = 1;
+
+               if ( ren1->processed )
+                       continue;
+               ren1->processed = 1;
+
+               const char *ren1_src = ren1->pair->one->path;
+               const char *ren1_dst = ren1->pair->two->path;
+
+               if ( ren2 ) {
+                       const char *ren2_src = ren2->pair->one->path;
+                       const char *ren2_dst = ren2->pair->two->path;
+                       /* Renamed in 1 and renamed in 2 */
+                       if (strcmp(ren1_src, ren2_src) != 0)
+                               die("ren1.src != ren2.src");
+                       ren2->dst_entry->processed = 1;
+                       ren2->processed = 1;
+                       if (strcmp(ren1_dst, ren2_dst) != 0) {
+                               cleanMerge = 0;
+                               output("CONFLICT (rename/rename): "
+                                      "Rename %s->%s in branch %s "
+                                      "rename %s->%s in %s",
+                                      src, ren1_dst, branchName1,
+                                      src, ren2_dst, branchName2);
+                               conflict_rename_rename(ren1, branchName1, ren2, branchName2);
+                       } else {
+                               remove_file(1, ren1_src);
+                               struct merge_file_info mfi;
+                               mfi = merge_file(ren1_src,
+                                                ren1->pair->one->sha1,
+                                                ren1->pair->one->mode,
+                                                ren1_dst,
+                                                ren1->pair->two->sha1,
+                                                ren1->pair->two->mode,
+                                                ren2_dst,
+                                                ren2->pair->two->sha1,
+                                                ren2->pair->two->mode,
+                                                branchName1,
+                                                branchName2);
+                               if ( mfi.merge || !mfi.clean )
+                                       output("Renaming %s->%s", src, ren1_dst);
+
+                               if ( mfi.merge )
+                                       output("Auto-merging %s", ren1_dst);
+
+                               if ( !mfi.clean ) {
+                                       output("CONFLICT (content): merge conflict in %s",
+                                              ren1_dst);
+                                       cleanMerge = 0;
+
+                                       if ( !index_only )
+                                               update_stages(ren1_dst,
+                                                             ren1->pair->one->sha1,
+                                                             ren1->pair->one->mode,
+                                                             ren1->pair->two->sha1,
+                                                             ren1->pair->two->mode,
+                                                             ren2->pair->two->sha1,
+                                                             ren2->pair->two->mode,
+                                                             1 /* clear */);
+                               }
+                               update_file(mfi.clean, mfi.sha, mfi.mode, ren1_dst);
+                       }
+               } else {
+                       /* Renamed in 1, maybe changed in 2 */
+                       remove_file(1, ren1_src);
+
+                       unsigned char srcShaOtherBranch[20], dstShaOtherBranch[20];
+                       unsigned srcModeOtherBranch, dstModeOtherBranch;
+
+                       int stage = renamesA == renames1 ? 3: 2;
+
+                       memcpy(srcShaOtherBranch, ren1->src_entry->stages[stage].sha, 20);
+                       srcModeOtherBranch = ren1->src_entry->stages[stage].mode;
+
+                       memcpy(dstShaOtherBranch, ren1->dst_entry->stages[stage].sha, 20);
+                       dstModeOtherBranch = ren1->dst_entry->stages[stage].mode;
+
+                       int tryMerge = 0;
+                       char *newPath;
+
+                       if (path_list_has_path(&currentDirectorySet, ren1_dst)) {
+                               cleanMerge = 0;
+                               output("CONFLICT (rename/directory): Rename %s->%s in %s "
+                                      " directory %s added in %s",
+                                      ren1_src, ren1_dst, branchName1,
+                                      ren1_dst, branchName2);
+                               conflict_rename_dir(ren1, branchName1);
+                       } else if ( memcmp(srcShaOtherBranch, null_sha1, 20) == 0 ) {
+                               cleanMerge = 0;
+                               output("CONFLICT (rename/delete): Rename %s->%s in %s "
+                                      "and deleted in %s",
+                                      ren1_src, ren1_dst, branchName1,
+                                      branchName2);
+                               update_file(0, ren1->pair->two->sha1, ren1->pair->two->mode, ren1_dst);
+                       } else if ( memcmp(dstShaOtherBranch, null_sha1, 20) != 0 ) {
+                               cleanMerge = 0;
+                               tryMerge = 1;
+                               output("CONFLICT (rename/add): Rename %s->%s in %s. "
+                                      "%s added in %s",
+                                      ren1_src, ren1_dst, branchName1,
+                                      ren1_dst, branchName2);
+                               newPath = unique_path(ren1_dst, branchName2);
+                               output("Adding as %s instead", newPath);
+                               update_file(0, dstShaOtherBranch, dstModeOtherBranch, newPath);
+                       } else if ( (ren2 = find_rename_bysrc(renames2Dst, ren1_dst)) ) {
+                               cleanMerge = 0;
+                               ren2->processed = 1;
+                               output("CONFLICT (rename/rename): Rename %s->%s in %s. "
+                                      "Rename %s->%s in %s",
+                                      ren1_src, ren1_dst, branchName1,
+                                      ren2->pair->one->path, ren2->pair->two->path, branchName2);
+                               conflict_rename_rename_2(ren1, branchName1, ren2, branchName2);
+                       } else
+                               tryMerge = 1;
+
+                       if ( tryMerge ) {
+                               const char *oname = ren1_src;
+                               const char *aname = ren1_dst;
+                               const char *bname = ren1_src;
+                               unsigned char osha[20], asha[20], bsha[20];
+                               unsigned omode = ren1->pair->one->mode;
+                               unsigned amode = ren1->pair->two->mode;
+                               unsigned bmode = srcModeOtherBranch;
+                               memcpy(osha, ren1->pair->one->sha1, 20);
+                               memcpy(asha, ren1->pair->two->sha1, 20);
+                               memcpy(bsha, srcShaOtherBranch, 20);
+                               const char *aBranch = branchName1;
+                               const char *bBranch = branchName2;
+
+                               if ( renamesA != renames1 ) {
+                                       memswp(&aname, &bname, sizeof(aname));
+                                       memswp(asha, bsha, 20);
+                                       memswp(&aBranch, &bBranch, sizeof(aBranch));
+                               }
+                               struct merge_file_info mfi;
+                               mfi = merge_file(oname, osha, omode,
+                                                aname, asha, amode,
+                                                bname, bsha, bmode,
+                                                aBranch, bBranch);
+
+                               if ( mfi.merge || !mfi.clean )
+                                       output("Renaming %s => %s", ren1_src, ren1_dst);
+                               if ( mfi.merge )
+                                       output("Auto-merging %s", ren1_dst);
+                               if ( !mfi.clean ) {
+                                       output("CONFLICT (rename/modify): Merge conflict in %s",
+                                              ren1_dst);
+                                       cleanMerge = 0;
+
+                                       if ( !index_only )
+                                               update_stages(ren1_dst,
+                                                             osha, omode,
+                                                             asha, amode,
+                                                             bsha, bmode,
+                                                             1 /* clear */);
+                               }
+                               update_file(mfi.clean, mfi.sha, mfi.mode, ren1_dst);
+                       }
+               }
+       }
+       path_list_clear(&srcNames, 0);
+       debug("  processRenames done\n");
+
+       if (cache_dirty)
+               flush_cache();
+       return cleanMerge;
+}
+
+static unsigned char *has_sha(const unsigned char *sha)
+{
+       return memcmp(sha, null_sha1, 20) == 0 ? NULL: (unsigned char *)sha;
+}
+
+/* Per entry merge function */
+static int process_entry(const char *path, struct stage_data *entry,
+                        const char *branch1Name,
+                        const char *branch2Name)
+{
+       /*
+       printf("processing entry, clean cache: %s\n", index_only ? "yes": "no");
+       print_index_entry("\tpath: ", entry);
+       */
+       int cleanMerge = 1;
+       unsigned char *oSha = has_sha(entry->stages[1].sha);
+       unsigned char *aSha = has_sha(entry->stages[2].sha);
+       unsigned char *bSha = has_sha(entry->stages[3].sha);
+       unsigned oMode = entry->stages[1].mode;
+       unsigned aMode = entry->stages[2].mode;
+       unsigned bMode = entry->stages[3].mode;
+
+       if ( oSha && (!aSha || !bSha) ) {
+               /* Case A: Deleted in one */
+               if ( (!aSha && !bSha) ||
+                    (sha_eq(aSha, oSha) && !bSha) ||
+                    (!aSha && sha_eq(bSha, oSha)) ) {
+                       /* Deleted in both or deleted in one and
+                        * unchanged in the other */
+                       if ( aSha )
+                               output("Removing %s", path);
+                       remove_file(1, path);
+               } else {
+                       /* Deleted in one and changed in the other */
+                       cleanMerge = 0;
+                       if ( !aSha ) {
+                               output("CONFLICT (delete/modify): %s deleted in %s "
+                                      "and modified in %s. Version %s of %s left in tree.",
+                                      path, branch1Name,
+                                      branch2Name, branch2Name, path);
+                               update_file(0, bSha, bMode, path);
+                       } else {
+                               output("CONFLICT (delete/modify): %s deleted in %s "
+                                      "and modified in %s. Version %s of %s left in tree.",
+                                      path, branch2Name,
+                                      branch1Name, branch1Name, path);
+                               update_file(0, aSha, aMode, path);
+                       }
+               }
+
+       } else if ( (!oSha && aSha && !bSha) ||
+                   (!oSha && !aSha && bSha) ) {
+               /* Case B: Added in one. */
+               const char *addBranch;
+               const char *otherBranch;
+               unsigned mode;
+               const unsigned char *sha;
+               const char *conf;
+
+               if ( aSha ) {
+                       addBranch = branch1Name;
+                       otherBranch = branch2Name;
+                       mode = aMode;
+                       sha = aSha;
+                       conf = "file/directory";
+               } else {
+                       addBranch = branch2Name;
+                       otherBranch = branch1Name;
+                       mode = bMode;
+                       sha = bSha;
+                       conf = "directory/file";
+               }
+               if ( path_list_has_path(&currentDirectorySet, path) ) {
+                       cleanMerge = 0;
+                       const char *newPath = unique_path(path, addBranch);
+                       output("CONFLICT (%s): There is a directory with name %s in %s. "
+                              "Adding %s as %s",
+                              conf, path, otherBranch, path, newPath);
+                       remove_file(0, path);
+                       update_file(0, sha, mode, newPath);
+               } else {
+                       output("Adding %s", path);
+                       update_file(1, sha, mode, path);
+               }
+       } else if ( !oSha && aSha && bSha ) {
+               /* Case C: Added in both (check for same permissions). */
+               if ( sha_eq(aSha, bSha) ) {
+                       if ( aMode != bMode ) {
+                               cleanMerge = 0;
+                               output("CONFLICT: File %s added identically in both branches, "
+                                      "but permissions conflict %06o->%06o",
+                                      path, aMode, bMode);
+                               output("CONFLICT: adding with permission: %06o", aMode);
+                               update_file(0, aSha, aMode, path);
+                       } else {
+                               /* This case is handled by git-read-tree */
+                               assert(0 && "This case must be handled by git-read-tree");
+                       }
+               } else {
+                       cleanMerge = 0;
+                       const char *newPath1 = unique_path(path, branch1Name);
+                       const char *newPath2 = unique_path(path, branch2Name);
+                       output("CONFLICT (add/add): File %s added non-identically "
+                              "in both branches. Adding as %s and %s instead.",
+                              path, newPath1, newPath2);
+                       remove_file(0, path);
+                       update_file(0, aSha, aMode, newPath1);
+                       update_file(0, bSha, bMode, newPath2);
+               }
+
+       } else if ( oSha && aSha && bSha ) {
+               /* case D: Modified in both, but differently. */
+               output("Auto-merging %s", path);
+               struct merge_file_info mfi;
+               mfi = merge_file(path, oSha, oMode,
+                                path, aSha, aMode,
+                                path, bSha, bMode,
+                                branch1Name, branch2Name);
+
+               if ( mfi.clean )
+                       update_file(1, mfi.sha, mfi.mode, path);
+               else {
+                       cleanMerge = 0;
+                       output("CONFLICT (content): Merge conflict in %s", path);
+
+                       if ( index_only )
+                               update_file(0, mfi.sha, mfi.mode, path);
+                       else
+                               update_file_flags(mfi.sha, mfi.mode, path,
+                                             0 /* updateCache */, 1 /* updateWd */);
+               }
+       } else
+               die("Fatal merge failure, shouldn't happen.");
+
+       if (cache_dirty)
+               flush_cache();
+
+       return cleanMerge;
+}
+
+static struct merge_tree_result merge_trees(struct tree *head,
+                                           struct tree *merge,
+                                           struct tree *common,
+                                           const char *branch1Name,
+                                           const char *branch2Name)
+{
+       int code;
+       struct merge_tree_result result = { NULL, 0 };
+       if ( !memcmp(common->object.sha1, merge->object.sha1, 20) ) {
+               output("Already uptodate!");
+               result.tree = head;
+               result.clean = 1;
+               return result;
+       }
+
+       debug("merge_trees ...\n");
+       code = git_merge_trees(index_only ? "-i": "-u", common, head, merge);
+
+       if ( code != 0 )
+               die("merging of trees %s and %s failed",
+                   sha1_to_hex(head->object.sha1),
+                   sha1_to_hex(merge->object.sha1));
+
+       result.tree = git_write_tree();
+
+       if ( !result.tree ) {
+               path_list_clear(&currentFileSet, 1);
+               path_list_clear(&currentDirectorySet, 1);
+               get_files_dirs(head, &currentFileSet, &currentDirectorySet);
+               get_files_dirs(merge, &currentFileSet, &currentDirectorySet);
+
+               struct path_list *entries = get_unmerged();
+               struct path_list *re_head, *re_merge;
+               re_head  = get_renames(head, common, head, merge, entries);
+               re_merge = get_renames(merge, common, head, merge, entries);
+               result.clean = process_renames(re_head, re_merge,
+                                              branch1Name, branch2Name);
+               debug("\tprocessing entries...\n");
+               int i;
+               for (i = 0; i < entries->nr; i++) {
+                       const char *path = entries->items[i].path;
+                       struct stage_data *e = entries->items[i].util;
+                       if (e->processed)
+                               continue;
+                       if (!process_entry(path, e, branch1Name, branch2Name))
+                               result.clean = 0;
+               }
+
+               free_rename_entries(&re_merge);
+               free_rename_entries(&re_head);
+               free_index_entries(&entries);
+
+               if (result.clean || index_only)
+                       result.tree = git_write_tree();
+               else
+                       result.tree = NULL;
+               debug("\t  processing entries done\n");
+       } else {
+               result.clean = 1;
+               printf("merging of trees %s and %s resulted in %s\n",
+                      sha1_to_hex(head->object.sha1),
+                      sha1_to_hex(merge->object.sha1),
+                      sha1_to_hex(result.tree->object.sha1));
+       }
+
+       debug("  merge_trees done\n");
+       return result;
+}
+
+/*
+ * Merge the commits h1 and h2, return the resulting virtual
+ * commit object and a flag indicating the cleaness of the merge.
+ */
+static
+struct merge_result merge(struct commit *h1,
+                         struct commit *h2,
+                         const char *branch1Name,
+                         const char *branch2Name,
+                         int callDepth /* =0 */,
+                         struct commit *ancestor /* =None */)
+{
+       struct merge_result result = { NULL, 0 };
+       const char *msg;
+       int msglen;
+       struct commit_list *ca = NULL, *iter;
+       struct commit *mergedCA;
+       struct merge_tree_result mtr;
+
+       output("Merging:");
+       msg = commit_title(h1, &msglen);
+       /* TODO: refactor. we always show the sha1 with the title */
+       output("%s %.*s", commit_hex_sha1(h1), msglen, msg);
+       msg = commit_title(h2, &msglen);
+       output("%s %.*s", commit_hex_sha1(h2), msglen, msg);
+
+       if ( ancestor )
+               commit_list_insert(ancestor, &ca);
+       else
+               ca = get_merge_bases(h1, h2, 1);
+
+       output("found %u common ancestor(s):", commit_list_count(ca));
+       for (iter = ca; iter; iter = iter->next) {
+               msg = commit_title(iter->item, &msglen);
+               output("%s %.*s", commit_hex_sha1(iter->item), msglen, msg);
+       }
+
+       mergedCA = pop_commit(&ca);
+
+       /* TODO: what happens when merge with virtual commits fails? */
+       for (iter = ca; iter; iter = iter->next) {
+               output_indent = callDepth + 1;
+               result = merge(mergedCA, iter->item,
+                              "Temporary merge branch 1",
+                              "Temporary merge branch 2",
+                              callDepth + 1,
+                              NULL);
+               mergedCA = result.commit;
+               output_indent = callDepth;
+
+               if ( !mergedCA )
+                       die("merge returned no commit");
+       }
+
+       if ( callDepth == 0 ) {
+               setup_index(0);
+               index_only = 0;
+       } else {
+               setup_index(1);
+               git_read_tree(h1->tree);
+               index_only = 1;
+       }
+
+       mtr = merge_trees(h1->tree, h2->tree,
+                         mergedCA->tree, branch1Name, branch2Name);
+
+       if ( !ancestor && (mtr.clean || index_only) ) {
+               result.commit = make_virtual_commit(mtr.tree, "merged tree");
+               commit_list_insert(h1, &result.commit->parents);
+               commit_list_insert(h2, &result.commit->parents->next);
+       } else
+               result.commit = NULL;
+
+       result.clean = mtr.clean;
+       return result;
+}
+
+static struct commit *get_ref(const char *ref)
+{
+       unsigned char sha1[20];
+       struct object *object;
+
+       if (get_sha1(ref, sha1))
+               die("Could not resolve ref '%s'", ref);
+       object = deref_tag(parse_object(sha1), ref, strlen(ref));
+       if (object->type != TYPE_COMMIT)
+               return NULL;
+       if (parse_commit((struct commit *)object))
+               die("Could not parse commit '%s'", sha1_to_hex(object->sha1));
+       return (struct commit *)object;
+}
+
+int main(int argc, char *argv[])
+{
+       static const char *bases[2];
+       static unsigned bases_count = 0;
+
+       original_index_file = getenv("GIT_INDEX_FILE");
+
+       if (!original_index_file)
+               original_index_file = strdup(git_path("index"));
+
+       temporary_index_file = strdup(git_path("mrg-rcrsv-tmp-idx"));
+
+       if (argc < 4)
+               die("Usage: %s <base>... -- <head> <remote> ...\n", argv[0]);
+
+       int i;
+       for (i = 1; i < argc; ++i) {
+               if (!strcmp(argv[i], "--"))
+                       break;
+               if (bases_count < sizeof(bases)/sizeof(*bases))
+                       bases[bases_count++] = argv[i];
+       }
+       if (argc - i != 3) /* "--" "<head>" "<remote>" */
+               die("Not handling anything other than two heads merge.");
+
+       const char *branch1, *branch2;
+
+       branch1 = argv[++i];
+       branch2 = argv[++i];
+       printf("Merging %s with %s\n", branch1, branch2);
+
+       struct merge_result result;
+       struct commit *h1 = get_ref(branch1);
+       struct commit *h2 = get_ref(branch2);
+
+       if (bases_count == 1) {
+               struct commit *ancestor = get_ref(bases[0]);
+               result = merge(h1, h2, branch1, branch2, 0, ancestor);
+       } else
+               result = merge(h1, h2, branch1, branch2, 0, NULL);
+
+       if (cache_dirty)
+               flush_cache();
+
+       return result.clean ? 0: 1;
+}
+
+/*
+vim: sw=8 noet
+*/
diff --git a/path-list.c b/path-list.c
new file mode 100644 (file)
index 0000000..f15a10d
--- /dev/null
@@ -0,0 +1,105 @@
+#include <stdio.h>
+#include "cache.h"
+#include "path-list.h"
+
+/* if there is no exact match, point to the index where the entry could be
+ * inserted */
+static int get_entry_index(const struct path_list *list, const char *path,
+               int *exact_match)
+{
+       int left = -1, right = list->nr;
+
+       while (left + 1 < right) {
+               int middle = (left + right) / 2;
+               int compare = strcmp(path, list->items[middle].path);
+               if (compare < 0)
+                       right = middle;
+               else if (compare > 0)
+                       left = middle;
+               else {
+                       *exact_match = 1;
+                       return middle;
+               }
+       }
+
+       *exact_match = 0;
+       return right;
+}
+
+/* returns -1-index if already exists */
+static int add_entry(struct path_list *list, const char *path)
+{
+       int exact_match;
+       int index = get_entry_index(list, path, &exact_match);
+
+       if (exact_match)
+               return -1 - index;
+
+       if (list->nr + 1 >= list->alloc) {
+               list->alloc += 32;
+               list->items = xrealloc(list->items, list->alloc
+                               * sizeof(struct path_list_item));
+       }
+       if (index < list->nr)
+               memmove(list->items + index + 1, list->items + index,
+                               (list->nr - index)
+                               * sizeof(struct path_list_item));
+       list->items[index].path = list->strdup_paths ?
+               strdup(path) : (char *)path;
+       list->items[index].util = NULL;
+       list->nr++;
+
+       return index;
+}
+
+struct path_list_item *path_list_insert(const char *path, struct path_list *list)
+{
+       int index = add_entry(list, path);
+
+       if (index < 0)
+               index = 1 - index;
+
+       return list->items + index;
+}
+
+int path_list_has_path(const struct path_list *list, const char *path)
+{
+       int exact_match;
+       get_entry_index(list, path, &exact_match);
+       return exact_match;
+}
+
+struct path_list_item *path_list_lookup(const char *path, struct path_list *list)
+{
+       int exact_match, i = get_entry_index(list, path, &exact_match);
+       if (!exact_match)
+               return NULL;
+       return list->items + i;
+}
+
+void path_list_clear(struct path_list *list, int free_items)
+{
+       if (list->items) {
+               int i;
+               if (free_items)
+                       for (i = 0; i < list->nr; i++) {
+                               if (list->strdup_paths)
+                                       free(list->items[i].path);
+                               if (list->items[i].util)
+                                       free(list->items[i].util);
+                       }
+               free(list->items);
+       }
+       list->items = NULL;
+       list->nr = list->alloc = 0;
+}
+
+void print_path_list(const char *text, const struct path_list *p)
+{
+       int i;
+       if ( text )
+               printf("%s\n", text);
+       for (i = 0; i < p->nr; i++)
+               printf("%s:%p\n", p->items[i].path, p->items[i].util);
+}
+
diff --git a/path-list.h b/path-list.h
new file mode 100644 (file)
index 0000000..d6401ea
--- /dev/null
@@ -0,0 +1,22 @@
+#ifndef _PATH_LIST_H_
+#define _PATH_LIST_H_
+
+struct path_list_item {
+       char *path;
+       void *util;
+};
+struct path_list
+{
+       struct path_list_item *items;
+       unsigned int nr, alloc;
+       unsigned int strdup_paths:1;
+};
+
+void print_path_list(const char *text, const struct path_list *p);
+
+int path_list_has_path(const struct path_list *list, const char *path);
+void path_list_clear(struct path_list *list, int free_items);
+struct path_list_item *path_list_insert(const char *path, struct path_list *list);
+struct path_list_item *path_list_lookup(const char *path, struct path_list *list);
+
+#endif /* _PATH_LIST_H_ */
index a50d3612c84d10da35b0b0e1278ecea899a4ec15..9c0a9fc2a1d8cba09e6f695883bf74b40243e99c 100644 (file)
@@ -24,6 +24,11 @@ unsigned int active_nr = 0, active_alloc = 0, active_cache_changed = 0;
 
 struct cache_tree *active_cache_tree = NULL;
 
+int cache_errno = 0;
+
+static void *cache_mmap = NULL;
+static size_t cache_mmap_size = 0;
+
 /*
  * This only updates the "non-critical" parts of the directory
  * cache, ie the parts that aren't tracked by GIT, and only used
@@ -577,22 +582,6 @@ int add_cache_entry(struct cache_entry *ce, int option)
        return 0;
 }
 
-/* Three functions to allow overloaded pointer return; see linux/err.h */
-static inline void *ERR_PTR(long error)
-{
-       return (void *) error;
-}
-
-static inline long PTR_ERR(const void *ptr)
-{
-       return (long) ptr;
-}
-
-static inline long IS_ERR(const void *ptr)
-{
-       return (unsigned long)ptr > (unsigned long)-1000L;
-}
-
 /*
  * "refresh" does not calculate a new sha1 file or bring the
  * cache up-to-date for mode/content changes. But what it
@@ -604,14 +593,16 @@ static inline long IS_ERR(const void *ptr)
  * For example, you'd want to do this after doing a "git-read-tree",
  * to link up the stat cache details with the proper files.
  */
-static struct cache_entry *refresh_entry(struct cache_entry *ce, int really)
+struct cache_entry *refresh_cache_entry(struct cache_entry *ce, int really)
 {
        struct stat st;
        struct cache_entry *updated;
        int changed, size;
 
-       if (lstat(ce->name, &st) < 0)
-               return ERR_PTR(-errno);
+       if (lstat(ce->name, &st) < 0) {
+               cache_errno = errno;
+               return NULL;
+       }
 
        changed = ce_match_stat(ce, &st, really);
        if (!changed) {
@@ -619,11 +610,13 @@ static struct cache_entry *refresh_entry(struct cache_entry *ce, int really)
                    !(ce->ce_flags & htons(CE_VALID)))
                        ; /* mark this one VALID again */
                else
-                       return NULL;
+                       return ce;
        }
 
-       if (ce_modified(ce, &st, really))
-               return ERR_PTR(-EINVAL);
+       if (ce_modified(ce, &st, really)) {
+               cache_errno = EINVAL;
+               return NULL;
+       }
 
        size = ce_size(ce);
        updated = xmalloc(size);
@@ -666,13 +659,13 @@ int refresh_cache(unsigned int flags)
                        continue;
                }
 
-               new = refresh_entry(ce, really);
-               if (!new)
+               new = refresh_cache_entry(ce, really);
+               if (new == ce)
                        continue;
-               if (IS_ERR(new)) {
-                       if (not_new && PTR_ERR(new) == -ENOENT)
+               if (!new) {
+                       if (not_new && cache_errno == ENOENT)
                                continue;
-                       if (really && PTR_ERR(new) == -EINVAL) {
+                       if (really && cache_errno == EINVAL) {
                                /* If we are doing --really-refresh that
                                 * means the index is not valid anymore.
                                 */
@@ -728,40 +721,44 @@ static int read_index_extension(const char *ext, void *data, unsigned long sz)
 }
 
 int read_cache(void)
+{
+       return read_cache_from(get_index_file());
+}
+
+/* remember to discard_cache() before reading a different cache! */
+int read_cache_from(const char *path)
 {
        int fd, i;
        struct stat st;
-       unsigned long size, offset;
-       void *map;
+       unsigned long offset;
        struct cache_header *hdr;
 
        errno = EBUSY;
-       if (active_cache)
+       if (cache_mmap)
                return active_nr;
 
        errno = ENOENT;
        index_file_timestamp = 0;
-       fd = open(get_index_file(), O_RDONLY);
+       fd = open(path, O_RDONLY);
        if (fd < 0) {
                if (errno == ENOENT)
                        return 0;
                die("index file open failed (%s)", strerror(errno));
        }
 
-       size = 0; /* avoid gcc warning */
-       map = MAP_FAILED;
+       cache_mmap = MAP_FAILED;
        if (!fstat(fd, &st)) {
-               size = st.st_size;
+               cache_mmap_size = st.st_size;
                errno = EINVAL;
-               if (size >= sizeof(struct cache_header) + 20)
-                       map = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
+               if (cache_mmap_size >= sizeof(struct cache_header) + 20)
+                       cache_mmap = mmap(NULL, cache_mmap_size, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
        }
        close(fd);
-       if (map == MAP_FAILED)
+       if (cache_mmap == MAP_FAILED)
                die("index file mmap failed (%s)", strerror(errno));
 
-       hdr = map;
-       if (verify_hdr(hdr, size) < 0)
+       hdr = cache_mmap;
+       if (verify_hdr(hdr, cache_mmap_size) < 0)
                goto unmap;
 
        active_nr = ntohl(hdr->hdr_entries);
@@ -770,12 +767,12 @@ int read_cache(void)
 
        offset = sizeof(*hdr);
        for (i = 0; i < active_nr; i++) {
-               struct cache_entry *ce = (struct cache_entry *) ((char *) map + offset);
+               struct cache_entry *ce = (struct cache_entry *) ((char *) cache_mmap + offset);
                offset = offset + ce_size(ce);
                active_cache[i] = ce;
        }
        index_file_timestamp = st.st_mtime;
-       while (offset <= size - 20 - 8) {
+       while (offset <= cache_mmap_size - 20 - 8) {
                /* After an array of active_nr index entries,
                 * there can be arbitrary number of extended
                 * sections, each of which is prefixed with
@@ -783,10 +780,10 @@ int read_cache(void)
                 * in 4-byte network byte order.
                 */
                unsigned long extsize;
-               memcpy(&extsize, (char *) map + offset + 4, 4);
+               memcpy(&extsize, (char *) cache_mmap + offset + 4, 4);
                extsize = ntohl(extsize);
-               if (read_index_extension(((const char *) map) + offset,
-                                        (char *) map + offset + 8,
+               if (read_index_extension(((const char *) cache_mmap) + offset,
+                                        (char *) cache_mmap + offset + 8,
                                         extsize) < 0)
                        goto unmap;
                offset += 8;
@@ -795,11 +792,28 @@ int read_cache(void)
        return active_nr;
 
 unmap:
-       munmap(map, size);
+       munmap(cache_mmap, cache_mmap_size);
        errno = EINVAL;
        die("index file corrupt");
 }
 
+int discard_cache()
+{
+       int ret;
+
+       if (cache_mmap == NULL)
+               return 0;
+       ret = munmap(cache_mmap, cache_mmap_size);
+       cache_mmap = NULL;
+       cache_mmap_size = 0;
+       active_nr = active_cache_changed = 0;
+       index_file_timestamp = 0;
+       cache_tree_free(&active_cache_tree);
+
+       /* no need to throw away allocated active_cache */
+       return ret;
+}
+
 #define WRITE_BUFFER_SIZE 8192
 static unsigned char write_buffer[WRITE_BUFFER_SIZE];
 static unsigned long write_buffer_len;
index d34c6cf6f3143bac81b69f79329d2d893e379c39..b70e177ddd75ca19c3be2698cb1a998b1bf85759 100755 (executable)
@@ -51,7 +51,7 @@ test_expect_success setup '
 '
 
 test_expect_success 'reference merge' '
-       git merge -s recursive "reference merge" HEAD master
+       git merge -s recur "reference merge" HEAD master
 '
 
 test_expect_success rebase '