git/sequencer.c

5269 строки
144 KiB
C
Исходник Обычный вид История

#include "cache.h"
#include "config.h"
#include "lockfile.h"
#include "dir.h"
#include "object-store.h"
#include "object.h"
#include "commit.h"
#include "sequencer.h"
#include "tag.h"
#include "run-command.h"
#include "exec-cmd.h"
#include "utf8.h"
#include "cache-tree.h"
#include "diff.h"
#include "revision.h"
#include "rerere.h"
#include "merge-recursive.h"
#include "refs.h"
#include "argv-array.h"
#include "quote.h"
#include "trailer.h"
#include "log-tree.h"
#include "wt-status.h"
#include "hashmap.h"
#include "notes-utils.h"
#include "sigchain.h"
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
#include "unpack-trees.h"
#include "worktree.h"
#include "oidmap.h"
#include "oidset.h"
#include "commit-slab.h"
#include "alias.h"
#include "commit-reach.h"
#include "rebase-interactive.h"
#define GIT_REFLOG_ACTION "GIT_REFLOG_ACTION"
static const char sign_off_header[] = "Signed-off-by: ";
static const char cherry_picked_prefix[] = "(cherry picked from commit ";
GIT_PATH_FUNC(git_path_commit_editmsg, "COMMIT_EDITMSG")
GIT_PATH_FUNC(git_path_seq_dir, "sequencer")
static GIT_PATH_FUNC(git_path_todo_file, "sequencer/todo")
static GIT_PATH_FUNC(git_path_opts_file, "sequencer/opts")
static GIT_PATH_FUNC(git_path_head_file, "sequencer/head")
static GIT_PATH_FUNC(git_path_abort_safety_file, "sequencer/abort-safety")
memoize common git-path "constant" files One of the most common uses of git_path() is to pass a constant, like git_path("MERGE_MSG"). This has two drawbacks: 1. The return value is a static buffer, and the lifetime is dependent on other calls to git_path, etc. 2. There's no compile-time checking of the pathname. This is OK for a one-off (after all, we have to spell it correctly at least once), but many of these constant strings appear throughout the code. This patch introduces a series of functions to "memoize" these strings, which are essentially globals for the lifetime of the program. We compute the value once, take ownership of the buffer, and return the cached value for subsequent calls. cache.h provides a helper macro for defining these functions as one-liners, and defines a few common ones for global use. Using a macro is a little bit gross, but it does nicely document the purpose of the functions. If we need to touch them all later (e.g., because we learned how to change the git_dir variable at runtime, and need to invalidate all of the stored values), it will be much easier to have the complete list. Note that the shared-global functions have separate, manual declarations. We could do something clever with the macros (e.g., expand it to a declaration in some places, and a declaration _and_ a definition in path.c). But there aren't that many, and it's probably better to stay away from too-magical macros. Likewise, if we abandon the C preprocessor in favor of generating these with a script, we could get much fancier. E.g., normalizing "FOO/BAR-BAZ" into "git_path_foo_bar_baz". But the small amount of saved typing is probably not worth the resulting confusion to readers who want to grep for the function's definition. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-08-10 12:38:57 +03:00
static GIT_PATH_FUNC(rebase_path, "rebase-merge")
/*
* The file containing rebase commands, comments, and empty lines.
* This file is created by "git rebase -i" then edited by the user. As
* the lines are processed, they are removed from the front of this
* file and written to the tail of 'done'.
*/
GIT_PATH_FUNC(rebase_path_todo, "rebase-merge/git-rebase-todo")
GIT_PATH_FUNC(rebase_path_todo_backup, "rebase-merge/git-rebase-todo.backup")
/*
* The rebase command lines that have already been processed. A line
* is moved here when it is first handled, before any associated user
* actions.
*/
static GIT_PATH_FUNC(rebase_path_done, "rebase-merge/done")
/*
* The file to keep track of how many commands were already processed (e.g.
* for the prompt).
*/
static GIT_PATH_FUNC(rebase_path_msgnum, "rebase-merge/msgnum")
/*
* The file to keep track of how many commands are to be processed in total
* (e.g. for the prompt).
*/
static GIT_PATH_FUNC(rebase_path_msgtotal, "rebase-merge/end")
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
/*
* The commit message that is planned to be used for any changes that
* need to be committed following a user interaction.
*/
static GIT_PATH_FUNC(rebase_path_message, "rebase-merge/message")
/*
* The file into which is accumulated the suggested commit message for
* squash/fixup commands. When the first of a series of squash/fixups
* is seen, the file is created and the commit message from the
* previous commit and from the first squash/fixup commit are written
* to it. The commit message for each subsequent squash/fixup commit
* is appended to the file as it is processed.
*/
static GIT_PATH_FUNC(rebase_path_squash_msg, "rebase-merge/message-squash")
/*
* If the current series of squash/fixups has not yet included a squash
* command, then this file exists and holds the commit message of the
* original "pick" commit. (If the series ends without a "squash"
* command, then this can be used as the commit message of the combined
* commit without opening the editor.)
*/
static GIT_PATH_FUNC(rebase_path_fixup_msg, "rebase-merge/message-fixup")
/*
* This file contains the list fixup/squash commands that have been
* accumulated into message-fixup or message-squash so far.
*/
static GIT_PATH_FUNC(rebase_path_current_fixups, "rebase-merge/current-fixups")
/*
* A script to set the GIT_AUTHOR_NAME, GIT_AUTHOR_EMAIL, and
* GIT_AUTHOR_DATE that will be used for the commit that is currently
* being rebased.
*/
static GIT_PATH_FUNC(rebase_path_author_script, "rebase-merge/author-script")
/*
* When an "edit" rebase command is being processed, the SHA1 of the
* commit to be edited is recorded in this file. When "git rebase
* --continue" is executed, if there are any staged changes then they
* will be amended to the HEAD commit, but only provided the HEAD
* commit is still the commit to be edited. When any other rebase
* command is processed, this file is deleted.
*/
static GIT_PATH_FUNC(rebase_path_amend, "rebase-merge/amend")
/*
* When we stop at a given patch via the "edit" command, this file contains
* the abbreviated commit name of the corresponding patch.
*/
static GIT_PATH_FUNC(rebase_path_stopped_sha, "rebase-merge/stopped-sha")
/*
* For the post-rewrite hook, we make a list of rewritten commits and
* their new sha1s. The rewritten-pending list keeps the sha1s of
* commits that have been processed, but not committed yet,
* e.g. because they are waiting for a 'squash' command.
*/
static GIT_PATH_FUNC(rebase_path_rewritten_list, "rebase-merge/rewritten-list")
static GIT_PATH_FUNC(rebase_path_rewritten_pending,
"rebase-merge/rewritten-pending")
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
/*
* The path of the file containig the OID of the "squash onto" commit, i.e.
* the dummy commit used for `reset [new root]`.
*/
static GIT_PATH_FUNC(rebase_path_squash_onto, "rebase-merge/squash-onto")
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
/*
* The path of the file listing refs that need to be deleted after the rebase
* finishes. This is used by the `label` command to record the need for cleanup.
*/
static GIT_PATH_FUNC(rebase_path_refs_to_delete, "rebase-merge/refs-to-delete")
/*
* The following files are written by git-rebase just after parsing the
* command-line.
*/
static GIT_PATH_FUNC(rebase_path_gpg_sign_opt, "rebase-merge/gpg_sign_opt")
static GIT_PATH_FUNC(rebase_path_orig_head, "rebase-merge/orig-head")
static GIT_PATH_FUNC(rebase_path_verbose, "rebase-merge/verbose")
static GIT_PATH_FUNC(rebase_path_quiet, "rebase-merge/quiet")
static GIT_PATH_FUNC(rebase_path_signoff, "rebase-merge/signoff")
static GIT_PATH_FUNC(rebase_path_head_name, "rebase-merge/head-name")
static GIT_PATH_FUNC(rebase_path_onto, "rebase-merge/onto")
static GIT_PATH_FUNC(rebase_path_autostash, "rebase-merge/autostash")
static GIT_PATH_FUNC(rebase_path_strategy, "rebase-merge/strategy")
static GIT_PATH_FUNC(rebase_path_strategy_opts, "rebase-merge/strategy_opts")
static GIT_PATH_FUNC(rebase_path_allow_rerere_autoupdate, "rebase-merge/allow_rerere_autoupdate")
static GIT_PATH_FUNC(rebase_path_reschedule_failed_exec, "rebase-merge/reschedule-failed-exec")
static int git_sequencer_config(const char *k, const char *v, void *cb)
{
struct replay_opts *opts = cb;
int status;
if (!strcmp(k, "commit.cleanup")) {
const char *s;
status = git_config_string(&s, k, v);
if (status)
return status;
if (!strcmp(s, "verbatim")) {
opts->default_msg_cleanup = COMMIT_MSG_CLEANUP_NONE;
opts->explicit_cleanup = 1;
} else if (!strcmp(s, "whitespace")) {
opts->default_msg_cleanup = COMMIT_MSG_CLEANUP_SPACE;
opts->explicit_cleanup = 1;
} else if (!strcmp(s, "strip")) {
opts->default_msg_cleanup = COMMIT_MSG_CLEANUP_ALL;
opts->explicit_cleanup = 1;
} else if (!strcmp(s, "scissors")) {
opts->default_msg_cleanup = COMMIT_MSG_CLEANUP_SCISSORS;
opts->explicit_cleanup = 1;
} else {
warning(_("invalid commit message cleanup mode '%s'"),
s);
}
free((char *)s);
return status;
}
if (!strcmp(k, "commit.gpgsign")) {
opts->gpg_sign = git_config_bool(k, v) ? xstrdup("") : NULL;
return 0;
}
status = git_gpg_config(k, v, NULL);
if (status)
return status;
return git_diff_basic_config(k, v, NULL);
}
void sequencer_init_config(struct replay_opts *opts)
{
opts->default_msg_cleanup = COMMIT_MSG_CLEANUP_NONE;
git_config(git_sequencer_config, opts);
}
static inline int is_rebase_i(const struct replay_opts *opts)
{
return opts->action == REPLAY_INTERACTIVE_REBASE;
}
static const char *get_dir(const struct replay_opts *opts)
{
if (is_rebase_i(opts))
return rebase_path();
return git_path_seq_dir();
}
static const char *get_todo_path(const struct replay_opts *opts)
{
if (is_rebase_i(opts))
return rebase_path_todo();
return git_path_todo_file();
}
/*
* Returns 0 for non-conforming footer
* Returns 1 for conforming footer
* Returns 2 when sob exists within conforming footer
* Returns 3 when sob exists within conforming footer as last entry
*/
static int has_conforming_footer(struct strbuf *sb, struct strbuf *sob,
size_t ignore_footer)
{
struct process_trailer_options opts = PROCESS_TRAILER_OPTIONS_INIT;
struct trailer_info info;
size_t i;
int found_sob = 0, found_sob_last = 0;
opts.no_divider = 1;
trailer_info_get(&info, sb->buf, &opts);
if (info.trailer_start == info.trailer_end)
return 0;
for (i = 0; i < info.trailer_nr; i++)
if (sob && !strncmp(info.trailers[i], sob->buf, sob->len)) {
found_sob = 1;
if (i == info.trailer_nr - 1)
found_sob_last = 1;
}
trailer_info_release(&info);
if (found_sob_last)
return 3;
if (found_sob)
return 2;
return 1;
}
static const char *gpg_sign_opt_quoted(struct replay_opts *opts)
{
static struct strbuf buf = STRBUF_INIT;
strbuf_reset(&buf);
if (opts->gpg_sign)
sq_quotef(&buf, "-S%s", opts->gpg_sign);
return buf.buf;
}
int sequencer_remove_state(struct replay_opts *opts)
{
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
struct strbuf buf = STRBUF_INIT;
int i, ret = 0;
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
if (is_rebase_i(opts) &&
strbuf_read_file(&buf, rebase_path_refs_to_delete(), 0) > 0) {
char *p = buf.buf;
while (*p) {
char *eol = strchr(p, '\n');
if (eol)
*eol = '\0';
if (delete_ref("(rebase -i) cleanup", p, NULL, 0) < 0) {
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
warning(_("could not delete '%s'"), p);
ret = -1;
}
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
if (!eol)
break;
p = eol + 1;
}
}
free(opts->gpg_sign);
free(opts->strategy);
for (i = 0; i < opts->xopts_nr; i++)
free(opts->xopts[i]);
free(opts->xopts);
strbuf_release(&opts->current_fixups);
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
strbuf_reset(&buf);
strbuf_addstr(&buf, get_dir(opts));
if (remove_dir_recursively(&buf, 0))
ret = error(_("could not remove '%s'"), buf.buf);
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
strbuf_release(&buf);
return ret;
}
static const char *action_name(const struct replay_opts *opts)
{
switch (opts->action) {
case REPLAY_REVERT:
return N_("revert");
case REPLAY_PICK:
return N_("cherry-pick");
case REPLAY_INTERACTIVE_REBASE:
return N_("rebase -i");
}
die(_("unknown action: %d"), opts->action);
}
struct commit_message {
char *parent_label;
char *label;
char *subject;
const char *message;
};
static const char *short_commit_name(struct commit *commit)
{
return find_unique_abbrev(&commit->object.oid, DEFAULT_ABBREV);
}
static int get_message(struct commit *commit, struct commit_message *out)
{
const char *abbrev, *subject;
int subject_len;
out->message = logmsg_reencode(commit, NULL, get_commit_output_encoding());
abbrev = short_commit_name(commit);
subject_len = find_commit_subject(out->message, &subject);
out->subject = xmemdupz(subject, subject_len);
out->label = xstrfmt("%s... %s", abbrev, out->subject);
out->parent_label = xstrfmt("parent of %s", out->label);
return 0;
}
static void free_message(struct commit *commit, struct commit_message *msg)
{
free(msg->parent_label);
free(msg->label);
free(msg->subject);
unuse_commit_buffer(commit, msg->message);
}
static void print_advice(struct repository *r, int show_hint,
struct replay_opts *opts)
{
char *msg = getenv("GIT_CHERRY_PICK_HELP");
if (msg) {
fprintf(stderr, "%s\n", msg);
/*
* A conflict has occurred but the porcelain
* (typically rebase --interactive) wants to take care
* of the commit itself so remove CHERRY_PICK_HEAD
*/
unlink(git_path_cherry_pick_head(r));
return;
}
if (show_hint) {
if (opts->no_commit)
advise(_("after resolving the conflicts, mark the corrected paths\n"
"with 'git add <paths>' or 'git rm <paths>'"));
else
advise(_("after resolving the conflicts, mark the corrected paths\n"
"with 'git add <paths>' or 'git rm <paths>'\n"
"and commit the result with 'git commit'"));
}
}
static int write_message(const void *buf, size_t len, const char *filename,
int append_eol)
{
struct lock_file msg_file = LOCK_INIT;
int msg_fd = hold_lock_file_for_update(&msg_file, filename, 0);
if (msg_fd < 0)
return error_errno(_("could not lock '%s'"), filename);
if (write_in_full(msg_fd, buf, len) < 0) {
error_errno(_("could not write to '%s'"), filename);
rollback_lock_file(&msg_file);
return -1;
}
if (append_eol && write(msg_fd, "\n", 1) < 0) {
error_errno(_("could not write eol to '%s'"), filename);
rollback_lock_file(&msg_file);
return -1;
}
if (commit_lock_file(&msg_file) < 0)
return error(_("failed to finalize '%s'"), filename);
return 0;
}
/*
* Reads a file that was presumably written by a shell script, i.e. with an
* end-of-line marker that needs to be stripped.
*
* Note that only the last end-of-line marker is stripped, consistent with the
* behavior of "$(cat path)" in a shell script.
*
* Returns 1 if the file was read, 0 if it could not be read or does not exist.
*/
static int read_oneliner(struct strbuf *buf,
const char *path, int skip_if_empty)
{
int orig_len = buf->len;
if (!file_exists(path))
return 0;
if (strbuf_read_file(buf, path, 0) < 0) {
warning_errno(_("could not read '%s'"), path);
return 0;
}
if (buf->len > orig_len && buf->buf[buf->len - 1] == '\n') {
if (--buf->len > orig_len && buf->buf[buf->len - 1] == '\r')
--buf->len;
buf->buf[buf->len] = '\0';
}
if (skip_if_empty && buf->len == orig_len)
return 0;
return 1;
}
static struct tree *empty_tree(struct repository *r)
{
return lookup_tree(r, the_hash_algo->empty_tree);
}
static int error_dirty_index(struct repository *repo, struct replay_opts *opts)
{
if (repo_read_index_unmerged(repo))
return error_resolve_conflict(_(action_name(opts)));
error(_("your local changes would be overwritten by %s."),
_(action_name(opts)));
if (advice_commit_before_merge)
advise(_("commit your changes or stash them to proceed."));
return -1;
}
static void update_abort_safety_file(void)
{
struct object_id head;
/* Do nothing on a single-pick */
if (!file_exists(git_path_seq_dir()))
return;
if (!get_oid("HEAD", &head))
write_file(git_path_abort_safety_file(), "%s", oid_to_hex(&head));
else
write_file(git_path_abort_safety_file(), "%s", "");
}
static int fast_forward_to(struct repository *r,
const struct object_id *to,
const struct object_id *from,
int unborn,
struct replay_opts *opts)
{
struct ref_transaction *transaction;
struct strbuf sb = STRBUF_INIT;
struct strbuf err = STRBUF_INIT;
repo_read_index(r);
if (checkout_fast_forward(r, from, to, 1))
return -1; /* the callee should have complained already */
strbuf_addf(&sb, _("%s: fast-forward"), _(action_name(opts)));
transaction = ref_transaction_begin(&err);
if (!transaction ||
ref_transaction_update(transaction, "HEAD",
to, unborn && !is_rebase_i(opts) ?
&null_oid : from,
0, sb.buf, &err) ||
ref_transaction_commit(transaction, &err)) {
ref_transaction_free(transaction);
error("%s", err.buf);
strbuf_release(&sb);
strbuf_release(&err);
return -1;
}
strbuf_release(&sb);
strbuf_release(&err);
ref_transaction_free(transaction);
update_abort_safety_file();
return 0;
}
enum commit_msg_cleanup_mode get_cleanup_mode(const char *cleanup_arg,
int use_editor)
{
if (!cleanup_arg || !strcmp(cleanup_arg, "default"))
return use_editor ? COMMIT_MSG_CLEANUP_ALL :
COMMIT_MSG_CLEANUP_SPACE;
else if (!strcmp(cleanup_arg, "verbatim"))
return COMMIT_MSG_CLEANUP_NONE;
else if (!strcmp(cleanup_arg, "whitespace"))
return COMMIT_MSG_CLEANUP_SPACE;
else if (!strcmp(cleanup_arg, "strip"))
return COMMIT_MSG_CLEANUP_ALL;
else if (!strcmp(cleanup_arg, "scissors"))
return use_editor ? COMMIT_MSG_CLEANUP_SCISSORS :
COMMIT_MSG_CLEANUP_SPACE;
else
die(_("Invalid cleanup mode %s"), cleanup_arg);
}
/*
* NB using int rather than enum cleanup_mode to stop clang's
* -Wtautological-constant-out-of-range-compare complaining that the comparison
* is always true.
*/
static const char *describe_cleanup_mode(int cleanup_mode)
{
static const char *modes[] = { "whitespace",
"verbatim",
"scissors",
"strip" };
if (cleanup_mode < ARRAY_SIZE(modes))
return modes[cleanup_mode];
BUG("invalid cleanup_mode provided (%d)", cleanup_mode);
}
void append_conflicts_hint(struct index_state *istate,
struct strbuf *msgbuf, enum commit_msg_cleanup_mode cleanup_mode)
{
int i;
if (cleanup_mode == COMMIT_MSG_CLEANUP_SCISSORS) {
strbuf_addch(msgbuf, '\n');
wt_status_append_cut_line(msgbuf);
strbuf_addch(msgbuf, comment_line_char);
}
strbuf_addch(msgbuf, '\n');
strbuf_commented_addf(msgbuf, "Conflicts:\n");
for (i = 0; i < istate->cache_nr;) {
const struct cache_entry *ce = istate->cache[i++];
if (ce_stage(ce)) {
strbuf_commented_addf(msgbuf, "\t%s\n", ce->name);
while (i < istate->cache_nr &&
!strcmp(ce->name, istate->cache[i]->name))
i++;
}
}
}
static int do_recursive_merge(struct repository *r,
struct commit *base, struct commit *next,
const char *base_label, const char *next_label,
struct object_id *head, struct strbuf *msgbuf,
struct replay_opts *opts)
{
struct merge_options o;
struct tree *next_tree, *base_tree, *head_tree;
int clean;
char **xopt;
struct lock_file index_lock = LOCK_INIT;
if (repo_hold_locked_index(r, &index_lock, LOCK_REPORT_ON_ERROR) < 0)
return -1;
repo_read_index(r);
init_merge_options(&o, r);
o.ancestor = base ? base_label : "(empty tree)";
o.branch1 = "HEAD";
o.branch2 = next ? next_label : "(empty tree)";
if (is_rebase_i(opts))
o.buffer_output = 2;
o.show_rename_progress = 1;
head_tree = parse_tree_indirect(head);
next_tree = next ? get_commit_tree(next) : empty_tree(r);
base_tree = base ? get_commit_tree(base) : empty_tree(r);
for (xopt = opts->xopts; xopt != opts->xopts + opts->xopts_nr; xopt++)
parse_merge_opt(&o, *xopt);
clean = merge_trees(&o,
head_tree,
next_tree, base_tree);
if (is_rebase_i(opts) && clean <= 0)
fputs(o.obuf.buf, stdout);
strbuf_release(&o.obuf);
if (clean < 0) {
rollback_lock_file(&index_lock);
prepare the builtins for a libified merge_recursive() Previously, callers of merge_trees() or merge_recursive() expected that code to die() with an error message. This used to be okay because we called those commands from scripts, and had a chance to print out a message in case the command failed fatally (read: with exit code 128). As scripting incurs its own set of problems (portability, speed, idiosyncrasies of different shells, limited data structures leading to inefficient code), we are converting more and more of these scripts into builtins, using library functions directly. We already tried to use merge_recursive() directly in the builtin git-am, for example. Unfortunately, we had to roll it back temporarily because some of the code in merge-recursive.c still deemed it okay to call die(), when the builtin am code really wanted to print out a useful advice after the merge failed fatally. In the next commits, we want to fix that. The code touched by this commit expected merge_trees() to die() with some useful message when there is an error condition, but merge_trees() is going to be improved by converting all die() calls to return error() instead (i.e. return value -1 after printing out the message as before), so that the caller can react more flexibly. This is a step to prepare for the version of merge_trees() that no longer dies, even if we just imitate the previous behavior by calling exit(128): this is what callers of e.g. `git merge` have come to expect. Note that the callers of the sequencer (revert and cherry-pick) already fail fast even for the return value -1; The only difference is that they now get a chance to say "<command> failed". A caller of merge_trees() might want handle error messages themselves (or even suppress them). As this patch is already complex enough, we leave that change for a later patch. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-07-26 19:06:02 +03:00
return clean;
}
if (write_locked_index(r->index, &index_lock,
COMMIT_LOCK | SKIP_IF_UNCHANGED))
C style: use standard style for "TRANSLATORS" comments Change all the "TRANSLATORS: [...]" comments in the C code to use the regular Git coding style, and amend the style guide so that the example there uses that style. This custom style was necessary back in 2010 when the gettext support was initially added, and was subsequently documented in commit cbcfd4e3ea ("i18n: mention "TRANSLATORS:" marker in Documentation/CodingGuidelines", 2014-04-18). GNU xgettext hasn't had the parsing limitation that necessitated this exception for almost 3 years. Since its 0.19 release on 2014-06-02 it's been able to recognize TRANSLATOR comments in the standard Git comment syntax[1]. Usually we'd like to keep compatibility with software that's that young, but in this case literally the only person who needs to be using a gettext newer than 3 years old is Jiang Xin (the only person who runs & commits "make pot" results), so I think in this case we can make an exception. This xgettext parsing feature was added after a thread on the Git mailing list[2] which continued on the bug-gettext[3] list, but we never subsequently changed our style & styleguide, do so. There are already longstanding changes in git that use the standard comment style & have their TRANSLATORS comments extracted properly without getting the literal "*"'s mixed up in the text, as would happen before xgettext 0.19. Commit 7ff2683253 ("builtin-am: implement -i/--interactive", 2015-08-04) added one such comment, which in commit df0617bfa7 ("l10n: git.pot: v2.6.0 round 1 (123 new, 41 removed)", 2015-09-05) got picked up in the po/git.pot file with the right format, showing that Jiang already runs a modern xgettext. The xgettext parser does not handle the sort of non-standard comment style that I'm amending here in sequencer.c, but that isn't standard Git comment syntax anyway. With this change to sequencer.c & "make pot" the comment in the pot file is now correct: #. TRANSLATORS: %s will be "revert", "cherry-pick" or -#. * "rebase -i". +#. "rebase -i". 1. http://git.savannah.gnu.org/cgit/gettext.git/commit/?id=10af7fe6bd 2. <2ce9ec406501d112e032c8208417f8100bed04c6.1397712142.git.worldhello.net@gmail.com> (https://public-inbox.org/git/2ce9ec406501d112e032c8208417f8100bed04c6.1397712142.git.worldhello.net@gmail.com/) 3. https://lists.gnu.org/archive/html/bug-gettext/2014-04/msg00016.html Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Acked-by: Jiang Xin <worldhello.net@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-05-12 00:20:12 +03:00
/*
* TRANSLATORS: %s will be "revert", "cherry-pick" or
* "rebase -i".
*/
return error(_("%s: Unable to write new index file"),
_(action_name(opts)));
if (!clean)
append_conflicts_hint(r->index, msgbuf,
opts->default_msg_cleanup);
return !clean;
}
static struct object_id *get_cache_tree_oid(struct index_state *istate)
{
if (!istate->cache_tree)
istate->cache_tree = cache_tree();
if (!cache_tree_fully_valid(istate->cache_tree))
if (cache_tree_update(istate, 0)) {
error(_("unable to update cache tree"));
return NULL;
}
return &istate->cache_tree->oid;
}
static int is_index_unchanged(struct repository *r)
{
struct object_id head_oid, *cache_tree_oid;
struct commit *head_commit;
struct index_state *istate = r->index;
if (!resolve_ref_unsafe("HEAD", RESOLVE_REF_READING, &head_oid, NULL))
return error(_("could not resolve HEAD commit"));
head_commit = lookup_commit(r, &head_oid);
/*
* If head_commit is NULL, check_commit, called from
* lookup_commit, would have indicated that head_commit is not
* a commit object already. parse_commit() will return failure
* without further complaints in such a case. Otherwise, if
* the commit is invalid, parse_commit() will complain. So
* there is nothing for us to say here. Just return failure.
*/
if (parse_commit(head_commit))
return -1;
if (!(cache_tree_oid = get_cache_tree_oid(istate)))
return -1;
return oideq(cache_tree_oid, get_commit_tree_oid(head_commit));
}
static int write_author_script(const char *message)
{
struct strbuf buf = STRBUF_INIT;
const char *eol;
int res;
for (;;)
if (!*message || starts_with(message, "\n")) {
missing_author:
/* Missing 'author' line? */
unlink(rebase_path_author_script());
return 0;
} else if (skip_prefix(message, "author ", &message))
break;
else if ((eol = strchr(message, '\n')))
message = eol + 1;
else
goto missing_author;
strbuf_addstr(&buf, "GIT_AUTHOR_NAME='");
while (*message && *message != '\n' && *message != '\r')
if (skip_prefix(message, " <", &message))
break;
else if (*message != '\'')
strbuf_addch(&buf, *(message++));
else
sequencer: fix quoting in write_author_script Single quotes should be escaped as \' not \\'. The bad quoting breaks the interactive version of 'rebase --root' (which is used when there is no '--onto' even if the user does not specify --interactive) for authors that contain "'" as sq_dequote() called by read_author_ident() errors out on the bad quoting. For other interactive rebases this only affects external scripts that read the author script and users whose git is upgraded from the shell version of rebase -i while rebase was stopped when the author contains "'". This is because the parsing in read_env_script() expected the broken quoting. This patch includes code to handle the broken quoting when git has been upgraded while rebase was stopped. It does this by detecting the missing "'" at the end of the GIT_AUTHOR_DATE line to see if it should dequote \\' as "'". Note this is only implemented for normal picks, not for creating a new root commit (rebase will stop with an error complaining out bad quoting in that case). The fallback code has been manually tested by reverting both the quoting fixes in write_author_script() and the previous fix for the missing "'" at the end of the GIT_AUTHOR_DATE line and running t3404-rebase-interactive.sh. Ideally rebase and am would share the same code for reading and writing the author script, but this commit just fixes the immediate bug. Helped-by: Eric Sunshine <sunshine@sunshineco.com> Helped-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Phillip Wood <phillip.wood@dunelm.org.uk> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-08-07 12:34:52 +03:00
strbuf_addf(&buf, "'\\%c'", *(message++));
strbuf_addstr(&buf, "'\nGIT_AUTHOR_EMAIL='");
while (*message && *message != '\n' && *message != '\r')
if (skip_prefix(message, "> ", &message))
break;
else if (*message != '\'')
strbuf_addch(&buf, *(message++));
else
sequencer: fix quoting in write_author_script Single quotes should be escaped as \' not \\'. The bad quoting breaks the interactive version of 'rebase --root' (which is used when there is no '--onto' even if the user does not specify --interactive) for authors that contain "'" as sq_dequote() called by read_author_ident() errors out on the bad quoting. For other interactive rebases this only affects external scripts that read the author script and users whose git is upgraded from the shell version of rebase -i while rebase was stopped when the author contains "'". This is because the parsing in read_env_script() expected the broken quoting. This patch includes code to handle the broken quoting when git has been upgraded while rebase was stopped. It does this by detecting the missing "'" at the end of the GIT_AUTHOR_DATE line to see if it should dequote \\' as "'". Note this is only implemented for normal picks, not for creating a new root commit (rebase will stop with an error complaining out bad quoting in that case). The fallback code has been manually tested by reverting both the quoting fixes in write_author_script() and the previous fix for the missing "'" at the end of the GIT_AUTHOR_DATE line and running t3404-rebase-interactive.sh. Ideally rebase and am would share the same code for reading and writing the author script, but this commit just fixes the immediate bug. Helped-by: Eric Sunshine <sunshine@sunshineco.com> Helped-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Phillip Wood <phillip.wood@dunelm.org.uk> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-08-07 12:34:52 +03:00
strbuf_addf(&buf, "'\\%c'", *(message++));
strbuf_addstr(&buf, "'\nGIT_AUTHOR_DATE='@");
while (*message && *message != '\n' && *message != '\r')
if (*message != '\'')
strbuf_addch(&buf, *(message++));
else
sequencer: fix quoting in write_author_script Single quotes should be escaped as \' not \\'. The bad quoting breaks the interactive version of 'rebase --root' (which is used when there is no '--onto' even if the user does not specify --interactive) for authors that contain "'" as sq_dequote() called by read_author_ident() errors out on the bad quoting. For other interactive rebases this only affects external scripts that read the author script and users whose git is upgraded from the shell version of rebase -i while rebase was stopped when the author contains "'". This is because the parsing in read_env_script() expected the broken quoting. This patch includes code to handle the broken quoting when git has been upgraded while rebase was stopped. It does this by detecting the missing "'" at the end of the GIT_AUTHOR_DATE line to see if it should dequote \\' as "'". Note this is only implemented for normal picks, not for creating a new root commit (rebase will stop with an error complaining out bad quoting in that case). The fallback code has been manually tested by reverting both the quoting fixes in write_author_script() and the previous fix for the missing "'" at the end of the GIT_AUTHOR_DATE line and running t3404-rebase-interactive.sh. Ideally rebase and am would share the same code for reading and writing the author script, but this commit just fixes the immediate bug. Helped-by: Eric Sunshine <sunshine@sunshineco.com> Helped-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Phillip Wood <phillip.wood@dunelm.org.uk> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-08-07 12:34:52 +03:00
strbuf_addf(&buf, "'\\%c'", *(message++));
strbuf_addch(&buf, '\'');
res = write_message(buf.buf, buf.len, rebase_path_author_script(), 1);
strbuf_release(&buf);
return res;
}
/**
* Take a series of KEY='VALUE' lines where VALUE part is
* sq-quoted, and append <KEY, VALUE> at the end of the string list
*/
static int parse_key_value_squoted(char *buf, struct string_list *list)
{
while (*buf) {
struct string_list_item *item;
char *np;
char *cp = strchr(buf, '=');
if (!cp) {
np = strchrnul(buf, '\n');
return error(_("no key present in '%.*s'"),
(int) (np - buf), buf);
}
np = strchrnul(cp, '\n');
*cp++ = '\0';
item = string_list_append(list, buf);
buf = np + (*np == '\n');
*np = '\0';
cp = sq_dequote(cp);
if (!cp)
return error(_("unable to dequote value of '%s'"),
item->string);
item->util = xstrdup(cp);
}
return 0;
}
sequencer: fix quoting in write_author_script Single quotes should be escaped as \' not \\'. The bad quoting breaks the interactive version of 'rebase --root' (which is used when there is no '--onto' even if the user does not specify --interactive) for authors that contain "'" as sq_dequote() called by read_author_ident() errors out on the bad quoting. For other interactive rebases this only affects external scripts that read the author script and users whose git is upgraded from the shell version of rebase -i while rebase was stopped when the author contains "'". This is because the parsing in read_env_script() expected the broken quoting. This patch includes code to handle the broken quoting when git has been upgraded while rebase was stopped. It does this by detecting the missing "'" at the end of the GIT_AUTHOR_DATE line to see if it should dequote \\' as "'". Note this is only implemented for normal picks, not for creating a new root commit (rebase will stop with an error complaining out bad quoting in that case). The fallback code has been manually tested by reverting both the quoting fixes in write_author_script() and the previous fix for the missing "'" at the end of the GIT_AUTHOR_DATE line and running t3404-rebase-interactive.sh. Ideally rebase and am would share the same code for reading and writing the author script, but this commit just fixes the immediate bug. Helped-by: Eric Sunshine <sunshine@sunshineco.com> Helped-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Phillip Wood <phillip.wood@dunelm.org.uk> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-08-07 12:34:52 +03:00
/**
* Reads and parses the state directory's "author-script" file, and sets name,
* email and date accordingly.
* Returns 0 on success, -1 if the file could not be parsed.
*
* The author script is of the format:
*
* GIT_AUTHOR_NAME='$author_name'
* GIT_AUTHOR_EMAIL='$author_email'
* GIT_AUTHOR_DATE='$author_date'
*
* where $author_name, $author_email and $author_date are quoted. We are strict
* with our parsing, as the file was meant to be eval'd in the now-removed
* git-am.sh/git-rebase--interactive.sh scripts, and thus if the file differs
* from what this function expects, it is better to bail out than to do
* something that the user does not expect.
sequencer: fix quoting in write_author_script Single quotes should be escaped as \' not \\'. The bad quoting breaks the interactive version of 'rebase --root' (which is used when there is no '--onto' even if the user does not specify --interactive) for authors that contain "'" as sq_dequote() called by read_author_ident() errors out on the bad quoting. For other interactive rebases this only affects external scripts that read the author script and users whose git is upgraded from the shell version of rebase -i while rebase was stopped when the author contains "'". This is because the parsing in read_env_script() expected the broken quoting. This patch includes code to handle the broken quoting when git has been upgraded while rebase was stopped. It does this by detecting the missing "'" at the end of the GIT_AUTHOR_DATE line to see if it should dequote \\' as "'". Note this is only implemented for normal picks, not for creating a new root commit (rebase will stop with an error complaining out bad quoting in that case). The fallback code has been manually tested by reverting both the quoting fixes in write_author_script() and the previous fix for the missing "'" at the end of the GIT_AUTHOR_DATE line and running t3404-rebase-interactive.sh. Ideally rebase and am would share the same code for reading and writing the author script, but this commit just fixes the immediate bug. Helped-by: Eric Sunshine <sunshine@sunshineco.com> Helped-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Phillip Wood <phillip.wood@dunelm.org.uk> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-08-07 12:34:52 +03:00
*/
int read_author_script(const char *path, char **name, char **email, char **date,
int allow_missing)
sequencer: fix quoting in write_author_script Single quotes should be escaped as \' not \\'. The bad quoting breaks the interactive version of 'rebase --root' (which is used when there is no '--onto' even if the user does not specify --interactive) for authors that contain "'" as sq_dequote() called by read_author_ident() errors out on the bad quoting. For other interactive rebases this only affects external scripts that read the author script and users whose git is upgraded from the shell version of rebase -i while rebase was stopped when the author contains "'". This is because the parsing in read_env_script() expected the broken quoting. This patch includes code to handle the broken quoting when git has been upgraded while rebase was stopped. It does this by detecting the missing "'" at the end of the GIT_AUTHOR_DATE line to see if it should dequote \\' as "'". Note this is only implemented for normal picks, not for creating a new root commit (rebase will stop with an error complaining out bad quoting in that case). The fallback code has been manually tested by reverting both the quoting fixes in write_author_script() and the previous fix for the missing "'" at the end of the GIT_AUTHOR_DATE line and running t3404-rebase-interactive.sh. Ideally rebase and am would share the same code for reading and writing the author script, but this commit just fixes the immediate bug. Helped-by: Eric Sunshine <sunshine@sunshineco.com> Helped-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Phillip Wood <phillip.wood@dunelm.org.uk> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-08-07 12:34:52 +03:00
{
struct strbuf buf = STRBUF_INIT;
struct string_list kv = STRING_LIST_INIT_DUP;
int retval = -1; /* assume failure */
int i, name_i = -2, email_i = -2, date_i = -2, err = 0;
sequencer: fix quoting in write_author_script Single quotes should be escaped as \' not \\'. The bad quoting breaks the interactive version of 'rebase --root' (which is used when there is no '--onto' even if the user does not specify --interactive) for authors that contain "'" as sq_dequote() called by read_author_ident() errors out on the bad quoting. For other interactive rebases this only affects external scripts that read the author script and users whose git is upgraded from the shell version of rebase -i while rebase was stopped when the author contains "'". This is because the parsing in read_env_script() expected the broken quoting. This patch includes code to handle the broken quoting when git has been upgraded while rebase was stopped. It does this by detecting the missing "'" at the end of the GIT_AUTHOR_DATE line to see if it should dequote \\' as "'". Note this is only implemented for normal picks, not for creating a new root commit (rebase will stop with an error complaining out bad quoting in that case). The fallback code has been manually tested by reverting both the quoting fixes in write_author_script() and the previous fix for the missing "'" at the end of the GIT_AUTHOR_DATE line and running t3404-rebase-interactive.sh. Ideally rebase and am would share the same code for reading and writing the author script, but this commit just fixes the immediate bug. Helped-by: Eric Sunshine <sunshine@sunshineco.com> Helped-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Phillip Wood <phillip.wood@dunelm.org.uk> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-08-07 12:34:52 +03:00
if (strbuf_read_file(&buf, path, 256) <= 0) {
strbuf_release(&buf);
if (errno == ENOENT && allow_missing)
return 0;
else
return error_errno(_("could not open '%s' for reading"),
path);
}
if (parse_key_value_squoted(buf.buf, &kv))
goto finish;
for (i = 0; i < kv.nr; i++) {
if (!strcmp(kv.items[i].string, "GIT_AUTHOR_NAME")) {
if (name_i != -2)
name_i = error(_("'GIT_AUTHOR_NAME' already given"));
else
name_i = i;
} else if (!strcmp(kv.items[i].string, "GIT_AUTHOR_EMAIL")) {
if (email_i != -2)
email_i = error(_("'GIT_AUTHOR_EMAIL' already given"));
else
email_i = i;
} else if (!strcmp(kv.items[i].string, "GIT_AUTHOR_DATE")) {
if (date_i != -2)
date_i = error(_("'GIT_AUTHOR_DATE' already given"));
else
date_i = i;
} else {
err = error(_("unknown variable '%s'"),
kv.items[i].string);
}
}
if (name_i == -2)
error(_("missing 'GIT_AUTHOR_NAME'"));
if (email_i == -2)
error(_("missing 'GIT_AUTHOR_EMAIL'"));
if (date_i == -2)
error(_("missing 'GIT_AUTHOR_DATE'"));
if (date_i < 0 || email_i < 0 || date_i < 0 || err)
goto finish;
*name = kv.items[name_i].util;
*email = kv.items[email_i].util;
*date = kv.items[date_i].util;
retval = 0;
finish:
string_list_clear(&kv, !!retval);
strbuf_release(&buf);
return retval;
sequencer: fix quoting in write_author_script Single quotes should be escaped as \' not \\'. The bad quoting breaks the interactive version of 'rebase --root' (which is used when there is no '--onto' even if the user does not specify --interactive) for authors that contain "'" as sq_dequote() called by read_author_ident() errors out on the bad quoting. For other interactive rebases this only affects external scripts that read the author script and users whose git is upgraded from the shell version of rebase -i while rebase was stopped when the author contains "'". This is because the parsing in read_env_script() expected the broken quoting. This patch includes code to handle the broken quoting when git has been upgraded while rebase was stopped. It does this by detecting the missing "'" at the end of the GIT_AUTHOR_DATE line to see if it should dequote \\' as "'". Note this is only implemented for normal picks, not for creating a new root commit (rebase will stop with an error complaining out bad quoting in that case). The fallback code has been manually tested by reverting both the quoting fixes in write_author_script() and the previous fix for the missing "'" at the end of the GIT_AUTHOR_DATE line and running t3404-rebase-interactive.sh. Ideally rebase and am would share the same code for reading and writing the author script, but this commit just fixes the immediate bug. Helped-by: Eric Sunshine <sunshine@sunshineco.com> Helped-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Phillip Wood <phillip.wood@dunelm.org.uk> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-08-07 12:34:52 +03:00
}
/*
* Read a GIT_AUTHOR_NAME, GIT_AUTHOR_EMAIL AND GIT_AUTHOR_DATE from a
* file with shell quoting into struct argv_array. Returns -1 on
* error, 0 otherwise.
*/
static int read_env_script(struct argv_array *env)
{
char *name, *email, *date;
if (read_author_script(rebase_path_author_script(),
&name, &email, &date, 0))
return -1;
argv_array_pushf(env, "GIT_AUTHOR_NAME=%s", name);
argv_array_pushf(env, "GIT_AUTHOR_EMAIL=%s", email);
argv_array_pushf(env, "GIT_AUTHOR_DATE=%s", date);
free(name);
free(email);
free(date);
return 0;
}
static char *get_author(const char *message)
{
size_t len;
const char *a;
a = find_commit_header(message, "author", &len);
if (a)
return xmemdupz(a, len);
return NULL;
}
static const char staged_changes_advice[] =
N_("you have staged changes in your working tree\n"
"If these changes are meant to be squashed into the previous commit, run:\n"
"\n"
" git commit --amend %s\n"
"\n"
"If they are meant to go into a new commit, run:\n"
"\n"
" git commit %s\n"
"\n"
"In both cases, once you're done, continue with:\n"
"\n"
" git rebase --continue\n");
#define ALLOW_EMPTY (1<<0)
#define EDIT_MSG (1<<1)
#define AMEND_MSG (1<<2)
#define CLEANUP_MSG (1<<3)
#define VERIFY_MSG (1<<4)
#define CREATE_ROOT_COMMIT (1<<5)
static int run_command_silent_on_success(struct child_process *cmd)
{
struct strbuf buf = STRBUF_INIT;
int rc;
cmd->stdout_to_stderr = 1;
rc = pipe_command(cmd,
NULL, 0,
NULL, 0,
&buf, 0);
if (rc)
fputs(buf.buf, stderr);
strbuf_release(&buf);
return rc;
}
/*
* If we are cherry-pick, and if the merge did not result in
* hand-editing, we will hit this commit and inherit the original
* author date and name.
*
* If we are revert, or if our cherry-pick results in a hand merge,
* we had better say that the current user is responsible for that.
*
* An exception is when run_git_commit() is called during an
* interactive rebase: in that case, we will want to retain the
* author metadata.
*/
static int run_git_commit(struct repository *r,
const char *defmsg,
struct replay_opts *opts,
unsigned int flags)
{
struct child_process cmd = CHILD_PROCESS_INIT;
cmd.git_cmd = 1;
if (is_rebase_i(opts) && read_env_script(&cmd.env_array)) {
const char *gpg_opt = gpg_sign_opt_quoted(opts);
return error(_(staged_changes_advice),
gpg_opt, gpg_opt);
}
argv_array_push(&cmd.args, "commit");
if (!(flags & VERIFY_MSG))
argv_array_push(&cmd.args, "-n");
if ((flags & AMEND_MSG))
argv_array_push(&cmd.args, "--amend");
if (opts->gpg_sign)
argv_array_pushf(&cmd.args, "-S%s", opts->gpg_sign);
if (defmsg)
argv_array_pushl(&cmd.args, "-F", defmsg, NULL);
else if (!(flags & EDIT_MSG))
argv_array_pushl(&cmd.args, "-C", "HEAD", NULL);
if ((flags & CLEANUP_MSG))
argv_array_push(&cmd.args, "--cleanup=strip");
if ((flags & EDIT_MSG))
argv_array_push(&cmd.args, "-e");
else if (!(flags & CLEANUP_MSG) &&
!opts->signoff && !opts->record_origin &&
!opts->explicit_cleanup)
argv_array_push(&cmd.args, "--cleanup=verbatim");
if ((flags & ALLOW_EMPTY))
argv_array_push(&cmd.args, "--allow-empty");
sequencer: fix --allow-empty-message behavior, make it smarter In commit b00bf1c9a8dd ("git-rebase: make --allow-empty-message the default", 2018-06-27), several arguments were given for transplanting empty commits without halting and asking the user for confirmation on each commit. These arguments were incomplete because the logic clearly assumed the only cases under consideration were transplanting of commits with empty messages (see the comment about "There are two sources for commits with empty messages). It didn't discuss or even consider rewords, squashes, etc. where the user is explicitly asked for a new commit message and provides an empty one. (My bad, I totally should have thought about that at the time, but just didn't.) Rewords and squashes are significantly different, though, as described by SZEDER: Let's suppose you start an interactive rebase, choose a commit to squash, save the instruction sheet, rebase fires up your editor, and then you notice that you mistakenly chose the wrong commit to squash. What do you do, how do you abort? Before [that commit] you could clear the commit message, exit the editor, and then rebase would say "Aborting commit due to empty commit message.", and you get to run 'git rebase --abort', and start over. But [since that commit, ...] saving the commit message as is would let rebase continue and create a bunch of unnecessary objects, and then you would have to use the reflog to return to the pre-rebase state. Also, he states: The instructions in the commit message template, which is shown for 'reword' and 'squash', too, still say... # Please enter the commit message for your changes. Lines starting # with '#' will be ignored, and an empty message aborts the commit. These are sound arguments that when editing commit messages during a sequencer operation, that if the commit message is empty then the operation should halt and ask the user to correct. The arguments in commit b00bf1c9a8dd (referenced above) still apply when transplanting previously created commits with empty commit messages, so the sequencer should not halt for those. Furthermore, all rationale so far applies equally for cherry-pick as for rebase. Therefore, make the code default to --allow-empty-message when transplanting an existing commit, and to default to halting when the user is asked to edit a commit message and provides an empty one -- for both rebase and cherry-pick. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-09-13 00:18:48 +03:00
if (!(flags & EDIT_MSG))
argv_array_push(&cmd.args, "--allow-empty-message");
if (is_rebase_i(opts) && !(flags & EDIT_MSG))
return run_command_silent_on_success(&cmd);
else
return run_command(&cmd);
}
static int rest_is_empty(const struct strbuf *sb, int start)
{
int i, eol;
const char *nl;
/* Check if the rest is just whitespace and Signed-off-by's. */
for (i = start; i < sb->len; i++) {
nl = memchr(sb->buf + i, '\n', sb->len - i);
if (nl)
eol = nl - sb->buf;
else
eol = sb->len;
if (strlen(sign_off_header) <= eol - i &&
starts_with(sb->buf + i, sign_off_header)) {
i = eol;
continue;
}
while (i < eol)
if (!isspace(sb->buf[i++]))
return 0;
}
return 1;
}
void cleanup_message(struct strbuf *msgbuf,
enum commit_msg_cleanup_mode cleanup_mode, int verbose)
{
if (verbose || /* Truncate the message just before the diff, if any. */
cleanup_mode == COMMIT_MSG_CLEANUP_SCISSORS)
strbuf_setlen(msgbuf, wt_status_locate_end(msgbuf->buf, msgbuf->len));
if (cleanup_mode != COMMIT_MSG_CLEANUP_NONE)
strbuf_stripspace(msgbuf, cleanup_mode == COMMIT_MSG_CLEANUP_ALL);
}
/*
* Find out if the message in the strbuf contains only whitespace and
* Signed-off-by lines.
*/
int message_is_empty(const struct strbuf *sb,
enum commit_msg_cleanup_mode cleanup_mode)
{
if (cleanup_mode == COMMIT_MSG_CLEANUP_NONE && sb->len)
return 0;
return rest_is_empty(sb, 0);
}
/*
* See if the user edited the message in the editor or left what
* was in the template intact
*/
int template_untouched(const struct strbuf *sb, const char *template_file,
enum commit_msg_cleanup_mode cleanup_mode)
{
struct strbuf tmpl = STRBUF_INIT;
const char *start;
if (cleanup_mode == COMMIT_MSG_CLEANUP_NONE && sb->len)
return 0;
if (!template_file || strbuf_read_file(&tmpl, template_file, 0) <= 0)
return 0;
strbuf_stripspace(&tmpl, cleanup_mode == COMMIT_MSG_CLEANUP_ALL);
if (!skip_prefix(sb->buf, tmpl.buf, &start))
start = sb->buf;
strbuf_release(&tmpl);
return rest_is_empty(sb, start - sb->buf);
}
int update_head_with_reflog(const struct commit *old_head,
const struct object_id *new_head,
const char *action, const struct strbuf *msg,
struct strbuf *err)
{
struct ref_transaction *transaction;
struct strbuf sb = STRBUF_INIT;
const char *nl;
int ret = 0;
if (action) {
strbuf_addstr(&sb, action);
strbuf_addstr(&sb, ": ");
}
nl = strchr(msg->buf, '\n');
if (nl) {
strbuf_add(&sb, msg->buf, nl + 1 - msg->buf);
} else {
strbuf_addbuf(&sb, msg);
strbuf_addch(&sb, '\n');
}
transaction = ref_transaction_begin(err);
if (!transaction ||
ref_transaction_update(transaction, "HEAD", new_head,
old_head ? &old_head->object.oid : &null_oid,
0, sb.buf, err) ||
ref_transaction_commit(transaction, err)) {
ret = -1;
}
ref_transaction_free(transaction);
strbuf_release(&sb);
return ret;
}
static int run_rewrite_hook(const struct object_id *oldoid,
const struct object_id *newoid)
{
struct child_process proc = CHILD_PROCESS_INIT;
const char *argv[3];
int code;
struct strbuf sb = STRBUF_INIT;
argv[0] = find_hook("post-rewrite");
if (!argv[0])
return 0;
argv[1] = "amend";
argv[2] = NULL;
proc.argv = argv;
proc.in = -1;
proc.stdout_to_stderr = 1;
proc.trace2_hook_name = "post-rewrite";
code = start_command(&proc);
if (code)
return code;
strbuf_addf(&sb, "%s %s\n", oid_to_hex(oldoid), oid_to_hex(newoid));
sigchain_push(SIGPIPE, SIG_IGN);
write_in_full(proc.in, sb.buf, sb.len);
close(proc.in);
strbuf_release(&sb);
sigchain_pop(SIGPIPE);
return finish_command(&proc);
}
void commit_post_rewrite(struct repository *r,
const struct commit *old_head,
const struct object_id *new_head)
{
struct notes_rewrite_cfg *cfg;
cfg = init_copy_notes_for_rewrite("amend");
if (cfg) {
/* we are amending, so old_head is not NULL */
copy_note_for_rewrite(cfg, &old_head->object.oid, new_head);
finish_copy_notes_for_rewrite(r, cfg, "Notes added by 'git commit --amend'");
}
run_rewrite_hook(&old_head->object.oid, new_head);
}
static int run_prepare_commit_msg_hook(struct repository *r,
struct strbuf *msg,
const char *commit)
{
struct argv_array hook_env = ARGV_ARRAY_INIT;
int ret;
const char *name;
name = git_path_commit_editmsg();
if (write_message(msg->buf, msg->len, name, 0))
return -1;
argv_array_pushf(&hook_env, "GIT_INDEX_FILE=%s", r->index_file);
argv_array_push(&hook_env, "GIT_EDITOR=:");
if (commit)
ret = run_hook_le(hook_env.argv, "prepare-commit-msg", name,
"commit", commit, NULL);
else
ret = run_hook_le(hook_env.argv, "prepare-commit-msg", name,
"message", NULL);
if (ret)
ret = error(_("'prepare-commit-msg' hook failed"));
argv_array_clear(&hook_env);
return ret;
}
commit: move print_commit_summary() to libgit Move print_commit_summary() from builtin/commit.c to sequencer.c so it can be shared with other commands. The function is modified by changing the last argument to a flag so callers can specify whether they want to show the author date in addition to specifying if this is an initial commit. If the sequencer dies in print_commit_summary() (which can only happen when cherry-picking or reverting) then neither the todo list nor the abort safety file are updated to reflect the commit that was just made. print_commit_summary() can die if: - The commit that was just created cannot be found or parsed. - HEAD cannot be resolved either because some other process is updating it (which is bad news in the middle of a cherry-pick) or because it is corrupt. - log_tree_commit() cannot read some objects. In all those cases dying will leave the sequencer in a sane state for aborting; 'git cherry-pick --abort' will rewind HEAD to the last successful commit before there was a problem with HEAD or the object database. If the user somehow fixes the problem and runs 'git cherry-pick --continue' then the sequencer will try and pick the same commit again which may or may not be what the user wants depending on what caused print_commit_summary() to die. If print_commit_summary() returned an error instead then update_abort_safety_file() would try to resolve HEAD which may or may not be successful. If it is successful then running 'git rebase --abort' would not rewind HEAD to the last successful commit which is not what we want. Signed-off-by: Phillip Wood <phillip.wood@dunelm.org.uk> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-11-24 14:07:54 +03:00
static const char implicit_ident_advice_noconfig[] =
N_("Your name and email address were configured automatically based\n"
"on your username and hostname. Please check that they are accurate.\n"
"You can suppress this message by setting them explicitly. Run the\n"
"following command and follow the instructions in your editor to edit\n"
"your configuration file:\n"
"\n"
" git config --global --edit\n"
"\n"
"After doing this, you may fix the identity used for this commit with:\n"
"\n"
" git commit --amend --reset-author\n");
static const char implicit_ident_advice_config[] =
N_("Your name and email address were configured automatically based\n"
"on your username and hostname. Please check that they are accurate.\n"
"You can suppress this message by setting them explicitly:\n"
"\n"
" git config --global user.name \"Your Name\"\n"
" git config --global user.email you@example.com\n"
"\n"
"After doing this, you may fix the identity used for this commit with:\n"
"\n"
" git commit --amend --reset-author\n");
static const char *implicit_ident_advice(void)
{
char *user_config = expand_user_path("~/.gitconfig", 0);
char *xdg_config = xdg_config_home("config");
int config_exists = file_exists(user_config) || file_exists(xdg_config);
free(user_config);
free(xdg_config);
if (config_exists)
return _(implicit_ident_advice_config);
else
return _(implicit_ident_advice_noconfig);
}
void print_commit_summary(struct repository *r,
const char *prefix,
const struct object_id *oid,
commit: move print_commit_summary() to libgit Move print_commit_summary() from builtin/commit.c to sequencer.c so it can be shared with other commands. The function is modified by changing the last argument to a flag so callers can specify whether they want to show the author date in addition to specifying if this is an initial commit. If the sequencer dies in print_commit_summary() (which can only happen when cherry-picking or reverting) then neither the todo list nor the abort safety file are updated to reflect the commit that was just made. print_commit_summary() can die if: - The commit that was just created cannot be found or parsed. - HEAD cannot be resolved either because some other process is updating it (which is bad news in the middle of a cherry-pick) or because it is corrupt. - log_tree_commit() cannot read some objects. In all those cases dying will leave the sequencer in a sane state for aborting; 'git cherry-pick --abort' will rewind HEAD to the last successful commit before there was a problem with HEAD or the object database. If the user somehow fixes the problem and runs 'git cherry-pick --continue' then the sequencer will try and pick the same commit again which may or may not be what the user wants depending on what caused print_commit_summary() to die. If print_commit_summary() returned an error instead then update_abort_safety_file() would try to resolve HEAD which may or may not be successful. If it is successful then running 'git rebase --abort' would not rewind HEAD to the last successful commit which is not what we want. Signed-off-by: Phillip Wood <phillip.wood@dunelm.org.uk> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-11-24 14:07:54 +03:00
unsigned int flags)
{
struct rev_info rev;
struct commit *commit;
struct strbuf format = STRBUF_INIT;
const char *head;
struct pretty_print_context pctx = {0};
struct strbuf author_ident = STRBUF_INIT;
struct strbuf committer_ident = STRBUF_INIT;
commit = lookup_commit(r, oid);
commit: move print_commit_summary() to libgit Move print_commit_summary() from builtin/commit.c to sequencer.c so it can be shared with other commands. The function is modified by changing the last argument to a flag so callers can specify whether they want to show the author date in addition to specifying if this is an initial commit. If the sequencer dies in print_commit_summary() (which can only happen when cherry-picking or reverting) then neither the todo list nor the abort safety file are updated to reflect the commit that was just made. print_commit_summary() can die if: - The commit that was just created cannot be found or parsed. - HEAD cannot be resolved either because some other process is updating it (which is bad news in the middle of a cherry-pick) or because it is corrupt. - log_tree_commit() cannot read some objects. In all those cases dying will leave the sequencer in a sane state for aborting; 'git cherry-pick --abort' will rewind HEAD to the last successful commit before there was a problem with HEAD or the object database. If the user somehow fixes the problem and runs 'git cherry-pick --continue' then the sequencer will try and pick the same commit again which may or may not be what the user wants depending on what caused print_commit_summary() to die. If print_commit_summary() returned an error instead then update_abort_safety_file() would try to resolve HEAD which may or may not be successful. If it is successful then running 'git rebase --abort' would not rewind HEAD to the last successful commit which is not what we want. Signed-off-by: Phillip Wood <phillip.wood@dunelm.org.uk> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-11-24 14:07:54 +03:00
if (!commit)
die(_("couldn't look up newly created commit"));
if (parse_commit(commit))
die(_("could not parse newly created commit"));
strbuf_addstr(&format, "format:%h] %s");
format_commit_message(commit, "%an <%ae>", &author_ident, &pctx);
format_commit_message(commit, "%cn <%ce>", &committer_ident, &pctx);
if (strbuf_cmp(&author_ident, &committer_ident)) {
strbuf_addstr(&format, "\n Author: ");
strbuf_addbuf_percentquote(&format, &author_ident);
}
if (flags & SUMMARY_SHOW_AUTHOR_DATE) {
struct strbuf date = STRBUF_INIT;
format_commit_message(commit, "%ad", &date, &pctx);
strbuf_addstr(&format, "\n Date: ");
strbuf_addbuf_percentquote(&format, &date);
strbuf_release(&date);
}
if (!committer_ident_sufficiently_given()) {
strbuf_addstr(&format, "\n Committer: ");
strbuf_addbuf_percentquote(&format, &committer_ident);
if (advice_implicit_identity) {
strbuf_addch(&format, '\n');
strbuf_addstr(&format, implicit_ident_advice());
}
}
strbuf_release(&author_ident);
strbuf_release(&committer_ident);
repo_init_revisions(r, &rev, prefix);
commit: move print_commit_summary() to libgit Move print_commit_summary() from builtin/commit.c to sequencer.c so it can be shared with other commands. The function is modified by changing the last argument to a flag so callers can specify whether they want to show the author date in addition to specifying if this is an initial commit. If the sequencer dies in print_commit_summary() (which can only happen when cherry-picking or reverting) then neither the todo list nor the abort safety file are updated to reflect the commit that was just made. print_commit_summary() can die if: - The commit that was just created cannot be found or parsed. - HEAD cannot be resolved either because some other process is updating it (which is bad news in the middle of a cherry-pick) or because it is corrupt. - log_tree_commit() cannot read some objects. In all those cases dying will leave the sequencer in a sane state for aborting; 'git cherry-pick --abort' will rewind HEAD to the last successful commit before there was a problem with HEAD or the object database. If the user somehow fixes the problem and runs 'git cherry-pick --continue' then the sequencer will try and pick the same commit again which may or may not be what the user wants depending on what caused print_commit_summary() to die. If print_commit_summary() returned an error instead then update_abort_safety_file() would try to resolve HEAD which may or may not be successful. If it is successful then running 'git rebase --abort' would not rewind HEAD to the last successful commit which is not what we want. Signed-off-by: Phillip Wood <phillip.wood@dunelm.org.uk> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-11-24 14:07:54 +03:00
setup_revisions(0, NULL, &rev, NULL);
rev.diff = 1;
rev.diffopt.output_format =
DIFF_FORMAT_SHORTSTAT | DIFF_FORMAT_SUMMARY;
rev.verbose_header = 1;
rev.show_root_diff = 1;
get_commit_format(format.buf, &rev);
rev.always_show_header = 0;
rev.diffopt.detect_rename = DIFF_DETECT_RENAME;
commit: move print_commit_summary() to libgit Move print_commit_summary() from builtin/commit.c to sequencer.c so it can be shared with other commands. The function is modified by changing the last argument to a flag so callers can specify whether they want to show the author date in addition to specifying if this is an initial commit. If the sequencer dies in print_commit_summary() (which can only happen when cherry-picking or reverting) then neither the todo list nor the abort safety file are updated to reflect the commit that was just made. print_commit_summary() can die if: - The commit that was just created cannot be found or parsed. - HEAD cannot be resolved either because some other process is updating it (which is bad news in the middle of a cherry-pick) or because it is corrupt. - log_tree_commit() cannot read some objects. In all those cases dying will leave the sequencer in a sane state for aborting; 'git cherry-pick --abort' will rewind HEAD to the last successful commit before there was a problem with HEAD or the object database. If the user somehow fixes the problem and runs 'git cherry-pick --continue' then the sequencer will try and pick the same commit again which may or may not be what the user wants depending on what caused print_commit_summary() to die. If print_commit_summary() returned an error instead then update_abort_safety_file() would try to resolve HEAD which may or may not be successful. If it is successful then running 'git rebase --abort' would not rewind HEAD to the last successful commit which is not what we want. Signed-off-by: Phillip Wood <phillip.wood@dunelm.org.uk> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-11-24 14:07:54 +03:00
rev.diffopt.break_opt = 0;
diff_setup_done(&rev.diffopt);
head = resolve_ref_unsafe("HEAD", 0, NULL, NULL);
if (!head)
die_errno(_("unable to resolve HEAD after creating commit"));
if (!strcmp(head, "HEAD"))
head = _("detached HEAD");
else
skip_prefix(head, "refs/heads/", &head);
printf("[%s%s ", head, (flags & SUMMARY_INITIAL_COMMIT) ?
_(" (root-commit)") : "");
if (!log_tree_commit(&rev, commit)) {
rev.always_show_header = 1;
rev.use_terminator = 1;
log_tree_commit(&rev, commit);
}
strbuf_release(&format);
}
static int parse_head(struct repository *r, struct commit **head)
{
struct commit *current_head;
struct object_id oid;
if (get_oid("HEAD", &oid)) {
current_head = NULL;
} else {
current_head = lookup_commit_reference(r, &oid);
if (!current_head)
return error(_("could not parse HEAD"));
if (!oideq(&oid, &current_head->object.oid)) {
warning(_("HEAD %s is not a commit!"),
oid_to_hex(&oid));
}
if (parse_commit(current_head))
return error(_("could not parse HEAD commit"));
}
*head = current_head;
return 0;
}
/*
* Try to commit without forking 'git commit'. In some cases we need
* to run 'git commit' to display an error message
*
* Returns:
* -1 - error unable to commit
* 0 - success
* 1 - run 'git commit'
*/
static int try_to_commit(struct repository *r,
struct strbuf *msg, const char *author,
struct replay_opts *opts, unsigned int flags,
struct object_id *oid)
{
struct object_id tree;
struct commit *current_head = NULL;
struct commit_list *parents = NULL;
struct commit_extra_header *extra = NULL;
struct strbuf err = STRBUF_INIT;
struct strbuf commit_msg = STRBUF_INIT;
char *amend_author = NULL;
const char *hook_commit = NULL;
enum commit_msg_cleanup_mode cleanup;
int res = 0;
if (parse_head(r, &current_head))
return -1;
if (flags & AMEND_MSG) {
const char *exclude_gpgsig[] = { "gpgsig", NULL };
const char *out_enc = get_commit_output_encoding();
const char *message = logmsg_reencode(current_head, NULL,
out_enc);
if (!msg) {
const char *orig_message = NULL;
find_commit_subject(message, &orig_message);
msg = &commit_msg;
strbuf_addstr(msg, orig_message);
hook_commit = "HEAD";
}
author = amend_author = get_author(message);
unuse_commit_buffer(current_head, message);
if (!author) {
res = error(_("unable to parse commit author"));
goto out;
}
parents = copy_commit_list(current_head->parents);
extra = read_commit_extra_headers(current_head, exclude_gpgsig);
} else if (current_head &&
(!(flags & CREATE_ROOT_COMMIT) || (flags & AMEND_MSG))) {
commit_list_insert(current_head, &parents);
}
if (write_index_as_tree(&tree, r->index, r->index_file, 0, NULL)) {
res = error(_("git write-tree failed to write a tree"));
goto out;
}
if (!(flags & ALLOW_EMPTY) && oideq(current_head ?
get_commit_tree_oid(current_head) :
the_hash_algo->empty_tree, &tree)) {
res = 1; /* run 'git commit' to display error message */
goto out;
}
if (find_hook("prepare-commit-msg")) {
res = run_prepare_commit_msg_hook(r, msg, hook_commit);
if (res)
goto out;
if (strbuf_read_file(&commit_msg, git_path_commit_editmsg(),
2048) < 0) {
res = error_errno(_("unable to read commit message "
"from '%s'"),
git_path_commit_editmsg());
goto out;
}
msg = &commit_msg;
}
if (flags & CLEANUP_MSG)
cleanup = COMMIT_MSG_CLEANUP_ALL;
else if ((opts->signoff || opts->record_origin) &&
!opts->explicit_cleanup)
cleanup = COMMIT_MSG_CLEANUP_SPACE;
else
cleanup = opts->default_msg_cleanup;
if (cleanup != COMMIT_MSG_CLEANUP_NONE)
strbuf_stripspace(msg, cleanup == COMMIT_MSG_CLEANUP_ALL);
sequencer: fix --allow-empty-message behavior, make it smarter In commit b00bf1c9a8dd ("git-rebase: make --allow-empty-message the default", 2018-06-27), several arguments were given for transplanting empty commits without halting and asking the user for confirmation on each commit. These arguments were incomplete because the logic clearly assumed the only cases under consideration were transplanting of commits with empty messages (see the comment about "There are two sources for commits with empty messages). It didn't discuss or even consider rewords, squashes, etc. where the user is explicitly asked for a new commit message and provides an empty one. (My bad, I totally should have thought about that at the time, but just didn't.) Rewords and squashes are significantly different, though, as described by SZEDER: Let's suppose you start an interactive rebase, choose a commit to squash, save the instruction sheet, rebase fires up your editor, and then you notice that you mistakenly chose the wrong commit to squash. What do you do, how do you abort? Before [that commit] you could clear the commit message, exit the editor, and then rebase would say "Aborting commit due to empty commit message.", and you get to run 'git rebase --abort', and start over. But [since that commit, ...] saving the commit message as is would let rebase continue and create a bunch of unnecessary objects, and then you would have to use the reflog to return to the pre-rebase state. Also, he states: The instructions in the commit message template, which is shown for 'reword' and 'squash', too, still say... # Please enter the commit message for your changes. Lines starting # with '#' will be ignored, and an empty message aborts the commit. These are sound arguments that when editing commit messages during a sequencer operation, that if the commit message is empty then the operation should halt and ask the user to correct. The arguments in commit b00bf1c9a8dd (referenced above) still apply when transplanting previously created commits with empty commit messages, so the sequencer should not halt for those. Furthermore, all rationale so far applies equally for cherry-pick as for rebase. Therefore, make the code default to --allow-empty-message when transplanting an existing commit, and to default to halting when the user is asked to edit a commit message and provides an empty one -- for both rebase and cherry-pick. Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-09-13 00:18:48 +03:00
if ((flags & EDIT_MSG) && message_is_empty(msg, cleanup)) {
res = 1; /* run 'git commit' to display error message */
goto out;
}
reset_ident_date();
if (commit_tree_extended(msg->buf, msg->len, &tree, parents,
oid, author, opts->gpg_sign, extra)) {
res = error(_("failed to write commit object"));
goto out;
}
if (update_head_with_reflog(current_head, oid,
getenv("GIT_REFLOG_ACTION"), msg, &err)) {
res = error("%s", err.buf);
goto out;
}
if (flags & AMEND_MSG)
commit_post_rewrite(r, current_head, oid);
out:
free_commit_extra_headers(extra);
strbuf_release(&err);
strbuf_release(&commit_msg);
free(amend_author);
return res;
}
static int do_commit(struct repository *r,
const char *msg_file, const char *author,
struct replay_opts *opts, unsigned int flags)
{
int res = 1;
if (!(flags & EDIT_MSG) && !(flags & VERIFY_MSG)) {
struct object_id oid;
struct strbuf sb = STRBUF_INIT;
if (msg_file && strbuf_read_file(&sb, msg_file, 2048) < 0)
return error_errno(_("unable to read commit message "
"from '%s'"),
msg_file);
res = try_to_commit(r, msg_file ? &sb : NULL,
author, opts, flags, &oid);
strbuf_release(&sb);
if (!res) {
unlink(git_path_cherry_pick_head(r));
unlink(git_path_merge_msg(r));
if (!is_rebase_i(opts))
print_commit_summary(r, NULL, &oid,
SUMMARY_SHOW_AUTHOR_DATE);
return res;
}
}
if (res == 1)
return run_git_commit(r, msg_file, opts, flags);
return res;
}
static int is_original_commit_empty(struct commit *commit)
{
const struct object_id *ptree_oid;
if (parse_commit(commit))
return error(_("could not parse commit %s"),
oid_to_hex(&commit->object.oid));
if (commit->parents) {
struct commit *parent = commit->parents->item;
if (parse_commit(parent))
return error(_("could not parse parent commit %s"),
oid_to_hex(&parent->object.oid));
ptree_oid = get_commit_tree_oid(parent);
} else {
ptree_oid = the_hash_algo->empty_tree; /* commit is root */
}
return oideq(ptree_oid, get_commit_tree_oid(commit));
}
/*
* Do we run "git commit" with "--allow-empty"?
*/
static int allow_empty(struct repository *r,
struct replay_opts *opts,
struct commit *commit)
{
int index_unchanged, empty_commit;
/*
* Three cases:
*
* (1) we do not allow empty at all and error out.
*
* (2) we allow ones that were initially empty, but
* forbid the ones that become empty;
*
* (3) we allow both.
*/
if (!opts->allow_empty)
return 0; /* let "git commit" barf as necessary */
index_unchanged = is_index_unchanged(r);
if (index_unchanged < 0)
return index_unchanged;
if (!index_unchanged)
return 0; /* we do not have to say --allow-empty */
if (opts->keep_redundant_commits)
return 1;
empty_commit = is_original_commit_empty(commit);
if (empty_commit < 0)
return empty_commit;
if (!empty_commit)
return 0;
else
return 1;
}
static struct {
char c;
const char *str;
} todo_command_info[] = {
{ 'p', "pick" },
{ 0, "revert" },
{ 'e', "edit" },
{ 'r', "reword" },
{ 'f', "fixup" },
{ 's', "squash" },
{ 'x', "exec" },
{ 'b', "break" },
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
{ 'l', "label" },
{ 't', "reset" },
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
{ 'm', "merge" },
{ 0, "noop" },
{ 'd', "drop" },
{ 0, NULL }
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
};
static const char *command_to_string(const enum todo_command command)
{
if (command < TODO_COMMENT)
return todo_command_info[command].str;
die(_("unknown command: %d"), command);
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
}
static char command_to_char(const enum todo_command command)
{
if (command < TODO_COMMENT && todo_command_info[command].c)
return todo_command_info[command].c;
return comment_line_char;
}
static int is_noop(const enum todo_command command)
{
return TODO_NOOP <= command;
}
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
static int is_fixup(enum todo_command command)
{
return command == TODO_FIXUP || command == TODO_SQUASH;
}
/* Does this command create a (non-merge) commit? */
static int is_pick_or_similar(enum todo_command command)
{
switch (command) {
case TODO_PICK:
case TODO_REVERT:
case TODO_EDIT:
case TODO_REWORD:
case TODO_FIXUP:
case TODO_SQUASH:
return 1;
default:
return 0;
}
}
static int update_squash_messages(struct repository *r,
enum todo_command command,
struct commit *commit,
struct replay_opts *opts)
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
{
struct strbuf buf = STRBUF_INIT;
int res;
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
const char *message, *body;
if (opts->current_fixup_count > 0) {
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
struct strbuf header = STRBUF_INIT;
char *eol;
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
if (strbuf_read_file(&buf, rebase_path_squash_msg(), 9) <= 0)
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
return error(_("could not read '%s'"),
rebase_path_squash_msg());
eol = buf.buf[0] != comment_line_char ?
buf.buf : strchrnul(buf.buf, '\n');
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
strbuf_addf(&header, "%c ", comment_line_char);
strbuf_addf(&header, _("This is a combination of %d commits."),
opts->current_fixup_count + 2);
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
strbuf_splice(&buf, 0, eol - buf.buf, header.buf, header.len);
strbuf_release(&header);
} else {
struct object_id head;
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
struct commit *head_commit;
const char *head_message, *body;
if (get_oid("HEAD", &head))
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
return error(_("need a HEAD to fixup"));
if (!(head_commit = lookup_commit_reference(r, &head)))
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
return error(_("could not read HEAD"));
if (!(head_message = get_commit_buffer(head_commit, NULL)))
return error(_("could not read HEAD's commit message"));
find_commit_subject(head_message, &body);
if (write_message(body, strlen(body),
rebase_path_fixup_msg(), 0)) {
unuse_commit_buffer(head_commit, head_message);
return error(_("cannot write '%s'"),
rebase_path_fixup_msg());
}
strbuf_addf(&buf, "%c ", comment_line_char);
strbuf_addf(&buf, _("This is a combination of %d commits."), 2);
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
strbuf_addf(&buf, "\n%c ", comment_line_char);
strbuf_addstr(&buf, _("This is the 1st commit message:"));
strbuf_addstr(&buf, "\n\n");
strbuf_addstr(&buf, body);
unuse_commit_buffer(head_commit, head_message);
}
if (!(message = get_commit_buffer(commit, NULL)))
return error(_("could not read commit message of %s"),
oid_to_hex(&commit->object.oid));
find_commit_subject(message, &body);
if (command == TODO_SQUASH) {
unlink(rebase_path_fixup_msg());
strbuf_addf(&buf, "\n%c ", comment_line_char);
strbuf_addf(&buf, _("This is the commit message #%d:"),
++opts->current_fixup_count + 1);
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
strbuf_addstr(&buf, "\n\n");
strbuf_addstr(&buf, body);
} else if (command == TODO_FIXUP) {
strbuf_addf(&buf, "\n%c ", comment_line_char);
strbuf_addf(&buf, _("The commit message #%d will be skipped:"),
++opts->current_fixup_count + 1);
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
strbuf_addstr(&buf, "\n\n");
strbuf_add_commented_lines(&buf, body, strlen(body));
} else
return error(_("unknown command: %d"), command);
unuse_commit_buffer(commit, message);
res = write_message(buf.buf, buf.len, rebase_path_squash_msg(), 0);
strbuf_release(&buf);
if (!res) {
strbuf_addf(&opts->current_fixups, "%s%s %s",
opts->current_fixups.len ? "\n" : "",
command_to_string(command),
oid_to_hex(&commit->object.oid));
res = write_message(opts->current_fixups.buf,
opts->current_fixups.len,
rebase_path_current_fixups(), 0);
}
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
return res;
}
static void flush_rewritten_pending(void)
{
struct strbuf buf = STRBUF_INIT;
struct object_id newoid;
FILE *out;
if (strbuf_read_file(&buf, rebase_path_rewritten_pending(), (GIT_MAX_HEXSZ + 1) * 2) > 0 &&
!get_oid("HEAD", &newoid) &&
(out = fopen_or_warn(rebase_path_rewritten_list(), "a"))) {
char *bol = buf.buf, *eol;
while (*bol) {
eol = strchrnul(bol, '\n');
fprintf(out, "%.*s %s\n", (int)(eol - bol),
bol, oid_to_hex(&newoid));
if (!*eol)
break;
bol = eol + 1;
}
fclose(out);
unlink(rebase_path_rewritten_pending());
}
strbuf_release(&buf);
}
static void record_in_rewritten(struct object_id *oid,
enum todo_command next_command)
{
FILE *out = fopen_or_warn(rebase_path_rewritten_pending(), "a");
if (!out)
return;
fprintf(out, "%s\n", oid_to_hex(oid));
fclose(out);
if (!is_fixup(next_command))
flush_rewritten_pending();
}
static int do_pick_commit(struct repository *r,
enum todo_command command,
struct commit *commit,
struct replay_opts *opts,
int final_fixup, int *check_todo)
{
unsigned int flags = opts->edit ? EDIT_MSG : 0;
const char *msg_file = opts->edit ? NULL : git_path_merge_msg(r);
struct object_id head;
struct commit *base, *next, *parent;
const char *base_label, *next_label;
char *author = NULL;
struct commit_message msg = { NULL, NULL, NULL, NULL };
struct strbuf msgbuf = STRBUF_INIT;
int res, unborn = 0, reword = 0, allow;
if (opts->no_commit) {
/*
* We do not intend to commit immediately. We just want to
* merge the differences in, so let's compute the tree
* that represents the "current" state for merge-recursive
* to work on.
*/
if (write_index_as_tree(&head, r->index, r->index_file, 0, NULL))
return error(_("your index file is unmerged."));
} else {
unborn = get_oid("HEAD", &head);
/* Do we want to generate a root commit? */
if (is_pick_or_similar(command) && opts->have_squash_onto &&
oideq(&head, &opts->squash_onto)) {
if (is_fixup(command))
return error(_("cannot fixup root commit"));
flags |= CREATE_ROOT_COMMIT;
unborn = 1;
} else if (unborn)
oidcpy(&head, the_hash_algo->empty_tree);
if (index_differs_from(r, unborn ? empty_tree_oid_hex() : "HEAD",
NULL, 0))
return error_dirty_index(r, opts);
}
discard_index(r->index);
if (!commit->parents)
parent = NULL;
else if (commit->parents->next) {
/* Reverting or cherry-picking a merge commit */
int cnt;
struct commit_list *p;
if (!opts->mainline)
return error(_("commit %s is a merge but no -m option was given."),
oid_to_hex(&commit->object.oid));
for (cnt = 1, p = commit->parents;
cnt != opts->mainline && p;
cnt++)
p = p->next;
if (cnt != opts->mainline || !p)
return error(_("commit %s does not have parent %d"),
oid_to_hex(&commit->object.oid), opts->mainline);
parent = p->item;
} else if (1 < opts->mainline)
/*
* Non-first parent explicitly specified as mainline for
* non-merge commit
*/
return error(_("commit %s does not have parent %d"),
oid_to_hex(&commit->object.oid), opts->mainline);
else
parent = commit->parents->item;
if (get_message(commit, &msg) != 0)
return error(_("cannot get commit message for %s"),
oid_to_hex(&commit->object.oid));
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
if (opts->allow_ff && !is_fixup(command) &&
((parent && oideq(&parent->object.oid, &head)) ||
(!parent && unborn))) {
if (is_rebase_i(opts))
write_author_script(msg.message);
res = fast_forward_to(r, &commit->object.oid, &head, unborn,
opts);
if (res || command != TODO_REWORD)
goto leave;
reword = 1;
msg_file = NULL;
goto fast_forward_edit;
}
if (parent && parse_commit(parent) < 0)
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
/* TRANSLATORS: The first %s will be a "todo" command like
"revert" or "pick", the second %s a SHA1. */
return error(_("%s: cannot parse parent commit %s"),
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
command_to_string(command),
oid_to_hex(&parent->object.oid));
/*
* "commit" is an existing commit. We would want to apply
* the difference it introduces since its first parent "prev"
* on top of the current HEAD if we are cherry-pick. Or the
* reverse of it if we are revert.
*/
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
if (command == TODO_REVERT) {
base = commit;
base_label = msg.label;
next = parent;
next_label = msg.parent_label;
strbuf_addstr(&msgbuf, "Revert \"");
strbuf_addstr(&msgbuf, msg.subject);
strbuf_addstr(&msgbuf, "\"\n\nThis reverts commit ");
strbuf_addstr(&msgbuf, oid_to_hex(&commit->object.oid));
if (commit->parents && commit->parents->next) {
strbuf_addstr(&msgbuf, ", reversing\nchanges made to ");
strbuf_addstr(&msgbuf, oid_to_hex(&parent->object.oid));
}
strbuf_addstr(&msgbuf, ".\n");
} else {
const char *p;
base = parent;
base_label = msg.parent_label;
next = commit;
next_label = msg.label;
/* Append the commit log message to msgbuf. */
if (find_commit_subject(msg.message, &p))
strbuf_addstr(&msgbuf, p);
if (opts->record_origin) {
strbuf_complete_line(&msgbuf);
if (!has_conforming_footer(&msgbuf, NULL, 0))
strbuf_addch(&msgbuf, '\n');
strbuf_addstr(&msgbuf, cherry_picked_prefix);
strbuf_addstr(&msgbuf, oid_to_hex(&commit->object.oid));
strbuf_addstr(&msgbuf, ")\n");
}
if (!is_fixup(command))
author = get_author(msg.message);
}
if (command == TODO_REWORD)
reword = 1;
else if (is_fixup(command)) {
if (update_squash_messages(r, command, commit, opts))
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
return -1;
flags |= AMEND_MSG;
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
if (!final_fixup)
msg_file = rebase_path_squash_msg();
else if (file_exists(rebase_path_fixup_msg())) {
flags |= CLEANUP_MSG;
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
msg_file = rebase_path_fixup_msg();
} else {
const char *dest = git_path_squash_msg(r);
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
unlink(dest);
if (copy_file(dest, rebase_path_squash_msg(), 0666))
return error(_("could not rename '%s' to '%s'"),
rebase_path_squash_msg(), dest);
unlink(git_path_merge_msg(r));
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
msg_file = dest;
flags |= EDIT_MSG;
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
}
}
if (opts->signoff && !is_fixup(command))
append_signoff(&msgbuf, 0, 0);
if (is_rebase_i(opts) && write_author_script(msg.message) < 0)
res = -1;
else if (!opts->strategy || !strcmp(opts->strategy, "recursive") || command == TODO_REVERT) {
res = do_recursive_merge(r, base, next, base_label, next_label,
&head, &msgbuf, opts);
prepare the builtins for a libified merge_recursive() Previously, callers of merge_trees() or merge_recursive() expected that code to die() with an error message. This used to be okay because we called those commands from scripts, and had a chance to print out a message in case the command failed fatally (read: with exit code 128). As scripting incurs its own set of problems (portability, speed, idiosyncrasies of different shells, limited data structures leading to inefficient code), we are converting more and more of these scripts into builtins, using library functions directly. We already tried to use merge_recursive() directly in the builtin git-am, for example. Unfortunately, we had to roll it back temporarily because some of the code in merge-recursive.c still deemed it okay to call die(), when the builtin am code really wanted to print out a useful advice after the merge failed fatally. In the next commits, we want to fix that. The code touched by this commit expected merge_trees() to die() with some useful message when there is an error condition, but merge_trees() is going to be improved by converting all die() calls to return error() instead (i.e. return value -1 after printing out the message as before), so that the caller can react more flexibly. This is a step to prepare for the version of merge_trees() that no longer dies, even if we just imitate the previous behavior by calling exit(128): this is what callers of e.g. `git merge` have come to expect. Note that the callers of the sequencer (revert and cherry-pick) already fail fast even for the return value -1; The only difference is that they now get a chance to say "<command> failed". A caller of merge_trees() might want handle error messages themselves (or even suppress them). As this patch is already complex enough, we leave that change for a later patch. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-07-26 19:06:02 +03:00
if (res < 0)
goto leave;
res |= write_message(msgbuf.buf, msgbuf.len,
git_path_merge_msg(r), 0);
} else {
struct commit_list *common = NULL;
struct commit_list *remotes = NULL;
res = write_message(msgbuf.buf, msgbuf.len,
git_path_merge_msg(r), 0);
commit_list_insert(base, &common);
commit_list_insert(next, &remotes);
res |= try_merge_command(r, opts->strategy,
opts->xopts_nr, (const char **)opts->xopts,
common, oid_to_hex(&head), remotes);
free_commit_list(common);
free_commit_list(remotes);
}
strbuf_release(&msgbuf);
/*
* If the merge was clean or if it failed due to conflict, we write
* CHERRY_PICK_HEAD for the subsequent invocation of commit to use.
* However, if the merge did not even start, then we don't want to
* write it at all.
*/
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
if (command == TODO_PICK && !opts->no_commit && (res == 0 || res == 1) &&
update_ref(NULL, "CHERRY_PICK_HEAD", &commit->object.oid, NULL,
REF_NO_DEREF, UPDATE_REFS_MSG_ON_ERR))
res = -1;
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
if (command == TODO_REVERT && ((opts->no_commit && res == 0) || res == 1) &&
update_ref(NULL, "REVERT_HEAD", &commit->object.oid, NULL,
REF_NO_DEREF, UPDATE_REFS_MSG_ON_ERR))
res = -1;
if (res) {
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
error(command == TODO_REVERT
? _("could not revert %s... %s")
: _("could not apply %s... %s"),
short_commit_name(commit), msg.subject);
print_advice(r, res == 1, opts);
repo_rerere(r, opts->allow_rerere_auto);
goto leave;
}
allow = allow_empty(r, opts, commit);
if (allow < 0) {
res = allow;
goto leave;
} else if (allow)
flags |= ALLOW_EMPTY;
if (!opts->no_commit) {
if (author || command == TODO_REVERT || (flags & AMEND_MSG))
res = do_commit(r, msg_file, author, opts, flags);
else
res = error(_("unable to parse commit author"));
*check_todo = !!(flags & EDIT_MSG);
if (!res && reword) {
fast_forward_edit:
res = run_git_commit(r, NULL, opts, EDIT_MSG |
VERIFY_MSG | AMEND_MSG |
(flags & ALLOW_EMPTY));
*check_todo = 1;
}
}
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
if (!res && final_fixup) {
unlink(rebase_path_fixup_msg());
unlink(rebase_path_squash_msg());
unlink(rebase_path_current_fixups());
strbuf_reset(&opts->current_fixups);
opts->current_fixup_count = 0;
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
}
leave:
free_message(commit, &msg);
free(author);
update_abort_safety_file();
return res;
}
static int prepare_revs(struct replay_opts *opts)
{
cherry-pick/revert: respect order of revisions to pick When giving multiple individual revisions to cherry-pick or revert, as in 'git cherry-pick A B' or 'git revert B A', one would expect them to be picked/reverted in the order given on the command line. They are instead ordered by their commit timestamp -- in chronological order for "cherry-pick" and in reverse chronological order for "revert". This matches the order in which one would usually give them on the command line, making this bug somewhat hard to notice. Still, it has been reported at least once before [1]. It seems like the chronological sorting happened by accident because the revision walker has traditionally always sorted commits in reverse chronological order when rev_info.no_walk was enabled. In the case of 'git revert B A' where B is newer than A, this sorting is a no-op. For 'git cherry-pick A B', the sorting would reverse the arguments, but because the sequencer also flips the rev_info.reverse flag when picking (as opposed to reverting), the end result is a chronological order. The rev_info.reverse flag was probably flipped so that the revision walker emits B before C in 'git cherry-pick A..C'; that it happened to effectively undo the unexpected sorting done when not walking, was probably a coincidence that allowed this bug to happen at all. Fix the bug by telling the revision walker not to sort the commits when not walking. The only case we want to reverse the order is now when cherry-picking and walking revisions (rev_info.no_walk = 0). [1] http://thread.gmane.org/gmane.comp.version-control.git/164794 Signed-off-by: Martin von Zweigbergk <martinvonz@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2012-08-29 10:15:56 +04:00
/*
* picking (but not reverting) ranges (but not individual revisions)
* should be done in reverse
*/
if (opts->action == REPLAY_PICK && !opts->revs->no_walk)
opts->revs->reverse ^= 1;
if (prepare_revision_walk(opts->revs))
return error(_("revision walk setup failed"));
return 0;
}
static int read_and_refresh_cache(struct repository *r,
struct replay_opts *opts)
{
struct lock_file index_lock = LOCK_INIT;
int index_fd = repo_hold_locked_index(r, &index_lock, 0);
if (repo_read_index(r) < 0) {
rollback_lock_file(&index_lock);
return error(_("git %s: failed to read the index"),
_(action_name(opts)));
}
refresh_index(r->index, REFRESH_QUIET|REFRESH_UNMERGED, NULL, NULL, NULL);
if (index_fd >= 0) {
if (write_locked_index(r->index, &index_lock,
COMMIT_LOCK | SKIP_IF_UNCHANGED)) {
return error(_("git %s: failed to refresh the index"),
_(action_name(opts)));
}
}
return 0;
}
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
enum todo_item_flags {
TODO_EDIT_MERGE_MSG = 1
};
void todo_list_release(struct todo_list *todo_list)
{
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
strbuf_release(&todo_list->buf);
FREE_AND_NULL(todo_list->items);
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
todo_list->nr = todo_list->alloc = 0;
}
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
static struct todo_item *append_new_todo(struct todo_list *todo_list)
{
ALLOC_GROW(todo_list->items, todo_list->nr + 1, todo_list->alloc);
return todo_list->items + todo_list->nr++;
}
const char *todo_item_get_arg(struct todo_list *todo_list,
struct todo_item *item)
{
return todo_list->buf.buf + item->arg_offset;
}
static int is_command(enum todo_command command, const char **bol)
{
const char *str = todo_command_info[command].str;
const char nick = todo_command_info[command].c;
const char *p = *bol + 1;
return skip_prefix(*bol, str, bol) ||
((nick && **bol == nick) &&
(*p == ' ' || *p == '\t' || *p == '\n' || *p == '\r' || !*p) &&
(*bol = p));
}
static int parse_insn_line(struct repository *r, struct todo_item *item,
const char *buf, const char *bol, char *eol)
{
struct object_id commit_oid;
char *end_of_object_name;
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
int i, saved, status, padding;
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
item->flags = 0;
/* left-trim */
bol += strspn(bol, " \t");
if (bol == eol || *bol == '\r' || *bol == comment_line_char) {
item->command = TODO_COMMENT;
item->commit = NULL;
item->arg_offset = bol - buf;
item->arg_len = eol - bol;
return 0;
}
for (i = 0; i < TODO_COMMENT; i++)
if (is_command(i, &bol)) {
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
item->command = i;
break;
}
if (i >= TODO_COMMENT)
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
return -1;
/* Eat up extra spaces/ tabs before object name */
padding = strspn(bol, " \t");
bol += padding;
if (item->command == TODO_NOOP || item->command == TODO_BREAK) {
if (bol != eol)
return error(_("%s does not accept arguments: '%s'"),
command_to_string(item->command), bol);
item->commit = NULL;
item->arg_offset = bol - buf;
item->arg_len = eol - bol;
return 0;
}
if (!padding)
return error(_("missing arguments for %s"),
command_to_string(item->command));
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
if (item->command == TODO_EXEC || item->command == TODO_LABEL ||
item->command == TODO_RESET) {
item->commit = NULL;
item->arg_offset = bol - buf;
item->arg_len = (int)(eol - bol);
return 0;
}
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
if (item->command == TODO_MERGE) {
if (skip_prefix(bol, "-C", &bol))
bol += strspn(bol, " \t");
else if (skip_prefix(bol, "-c", &bol)) {
bol += strspn(bol, " \t");
item->flags |= TODO_EDIT_MERGE_MSG;
} else {
item->flags |= TODO_EDIT_MERGE_MSG;
item->commit = NULL;
item->arg_offset = bol - buf;
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
item->arg_len = (int)(eol - bol);
return 0;
}
}
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
end_of_object_name = (char *) bol + strcspn(bol, " \t\n");
saved = *end_of_object_name;
*end_of_object_name = '\0';
status = get_oid(bol, &commit_oid);
*end_of_object_name = saved;
bol = end_of_object_name + strspn(end_of_object_name, " \t");
item->arg_offset = bol - buf;
item->arg_len = (int)(eol - bol);
if (status < 0)
return error(_("could not parse '%.*s'"),
(int)(end_of_object_name - bol), bol);
item->commit = lookup_commit_reference(r, &commit_oid);
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
return !item->commit;
}
int sequencer_get_last_command(struct repository *r, enum replay_action *action)
{
const char *todo_file, *bol;
struct strbuf buf = STRBUF_INIT;
int ret = 0;
todo_file = git_path_todo_file();
if (strbuf_read_file(&buf, todo_file, 0) < 0) {
if (errno == ENOENT || errno == ENOTDIR)
return -1;
else
return error_errno("unable to open '%s'", todo_file);
}
bol = buf.buf + strspn(buf.buf, " \t\r\n");
if (is_command(TODO_PICK, &bol) && (*bol == ' ' || *bol == '\t'))
*action = REPLAY_PICK;
else if (is_command(TODO_REVERT, &bol) &&
(*bol == ' ' || *bol == '\t'))
*action = REPLAY_REVERT;
else
ret = -1;
strbuf_release(&buf);
return ret;
}
int todo_list_parse_insn_buffer(struct repository *r, char *buf,
struct todo_list *todo_list)
{
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
struct todo_item *item;
char *p = buf, *next_p;
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
int i, res = 0, fixup_okay = file_exists(rebase_path_done());
todo_list->current = todo_list->nr = 0;
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
for (i = 1; *p; i++, p = next_p) {
char *eol = strchrnul(p, '\n');
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
next_p = *eol ? eol + 1 /* skip LF */ : eol;
if (p != eol && eol[-1] == '\r')
eol--; /* strip Carriage Return */
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
item = append_new_todo(todo_list);
item->offset_in_buf = p - todo_list->buf.buf;
if (parse_insn_line(r, item, buf, p, eol)) {
res = error(_("invalid line %d: %.*s"),
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
i, (int)(eol - p), p);
item->command = TODO_COMMENT + 1;
item->arg_offset = p - buf;
item->arg_len = (int)(eol - p);
item->commit = NULL;
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
}
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
if (fixup_okay)
; /* do nothing */
else if (is_fixup(item->command))
return error(_("cannot '%s' without a previous commit"),
command_to_string(item->command));
else if (!is_noop(item->command))
fixup_okay = 1;
}
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
return res;
}
static int count_commands(struct todo_list *todo_list)
{
int count = 0, i;
for (i = 0; i < todo_list->nr; i++)
if (todo_list->items[i].command != TODO_COMMENT)
count++;
return count;
}
static int get_item_line_offset(struct todo_list *todo_list, int index)
{
return index < todo_list->nr ?
todo_list->items[index].offset_in_buf : todo_list->buf.len;
}
static const char *get_item_line(struct todo_list *todo_list, int index)
{
return todo_list->buf.buf + get_item_line_offset(todo_list, index);
}
static int get_item_line_length(struct todo_list *todo_list, int index)
{
return get_item_line_offset(todo_list, index + 1)
- get_item_line_offset(todo_list, index);
}
static ssize_t strbuf_read_file_or_whine(struct strbuf *sb, const char *path)
{
int fd;
ssize_t len;
fd = open(path, O_RDONLY);
if (fd < 0)
return error_errno(_("could not open '%s'"), path);
len = strbuf_read(sb, fd, 0);
close(fd);
if (len < 0)
return error(_("could not read '%s'."), path);
return len;
}
static int have_finished_the_last_pick(void)
{
struct strbuf buf = STRBUF_INIT;
const char *eol;
const char *todo_path = git_path_todo_file();
int ret = 0;
if (strbuf_read_file(&buf, todo_path, 0) < 0) {
if (errno == ENOENT) {
return 0;
} else {
error_errno("unable to open '%s'", todo_path);
return 0;
}
}
/* If there is only one line then we are done */
eol = strchr(buf.buf, '\n');
if (!eol || !eol[1])
ret = 1;
strbuf_release(&buf);
return ret;
}
void sequencer_post_commit_cleanup(struct repository *r, int verbose)
{
struct replay_opts opts = REPLAY_OPTS_INIT;
int need_cleanup = 0;
if (file_exists(git_path_cherry_pick_head(r))) {
if (!unlink(git_path_cherry_pick_head(r)) && verbose)
warning(_("cancelling a cherry picking in progress"));
opts.action = REPLAY_PICK;
need_cleanup = 1;
}
if (file_exists(git_path_revert_head(r))) {
if (!unlink(git_path_revert_head(r)) && verbose)
warning(_("cancelling a revert in progress"));
opts.action = REPLAY_REVERT;
need_cleanup = 1;
}
if (!need_cleanup)
return;
if (!have_finished_the_last_pick())
return;
sequencer_remove_state(&opts);
}
static int read_populate_todo(struct repository *r,
struct todo_list *todo_list,
struct replay_opts *opts)
{
struct stat st;
const char *todo_file = get_todo_path(opts);
int res;
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
strbuf_reset(&todo_list->buf);
if (strbuf_read_file_or_whine(&todo_list->buf, todo_file) < 0)
return -1;
res = stat(todo_file, &st);
if (res)
return error(_("could not stat '%s'"), todo_file);
fill_stat_data(&todo_list->stat, &st);
res = todo_list_parse_insn_buffer(r, todo_list->buf.buf, todo_list);
if (res) {
if (is_rebase_i(opts))
return error(_("please fix this using "
"'git rebase --edit-todo'."));
return error(_("unusable instruction sheet: '%s'"), todo_file);
}
if (!todo_list->nr &&
(!is_rebase_i(opts) || !file_exists(rebase_path_done())))
return error(_("no commits parsed."));
if (!is_rebase_i(opts)) {
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
enum todo_command valid =
opts->action == REPLAY_PICK ? TODO_PICK : TODO_REVERT;
int i;
for (i = 0; i < todo_list->nr; i++)
if (valid == todo_list->items[i].command)
continue;
else if (valid == TODO_PICK)
return error(_("cannot cherry-pick during a revert."));
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
else
return error(_("cannot revert during a cherry-pick."));
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
}
if (is_rebase_i(opts)) {
struct todo_list done = TODO_LIST_INIT;
FILE *f = fopen_or_warn(rebase_path_msgtotal(), "w");
if (strbuf_read_file(&done.buf, rebase_path_done(), 0) > 0 &&
!todo_list_parse_insn_buffer(r, done.buf.buf, &done))
todo_list->done_nr = count_commands(&done);
else
todo_list->done_nr = 0;
todo_list->total_nr = todo_list->done_nr
+ count_commands(todo_list);
todo_list_release(&done);
if (f) {
fprintf(f, "%d\n", todo_list->total_nr);
fclose(f);
}
}
return 0;
}
static int git_config_string_dup(char **dest,
const char *var, const char *value)
{
if (!value)
return config_error_nonbool(var);
free(*dest);
*dest = xstrdup(value);
return 0;
}
static int populate_opts_cb(const char *key, const char *value, void *data)
{
struct replay_opts *opts = data;
int error_flag = 1;
if (!value)
error_flag = 0;
else if (!strcmp(key, "options.no-commit"))
opts->no_commit = git_config_bool_or_int(key, value, &error_flag);
else if (!strcmp(key, "options.edit"))
opts->edit = git_config_bool_or_int(key, value, &error_flag);
else if (!strcmp(key, "options.allow-empty"))
opts->allow_empty =
git_config_bool_or_int(key, value, &error_flag);
else if (!strcmp(key, "options.allow-empty-message"))
opts->allow_empty_message =
git_config_bool_or_int(key, value, &error_flag);
else if (!strcmp(key, "options.keep-redundant-commits"))
opts->keep_redundant_commits =
git_config_bool_or_int(key, value, &error_flag);
else if (!strcmp(key, "options.signoff"))
opts->signoff = git_config_bool_or_int(key, value, &error_flag);
else if (!strcmp(key, "options.record-origin"))
opts->record_origin = git_config_bool_or_int(key, value, &error_flag);
else if (!strcmp(key, "options.allow-ff"))
opts->allow_ff = git_config_bool_or_int(key, value, &error_flag);
else if (!strcmp(key, "options.mainline"))
opts->mainline = git_config_int(key, value);
else if (!strcmp(key, "options.strategy"))
git_config_string_dup(&opts->strategy, key, value);
else if (!strcmp(key, "options.gpg-sign"))
git_config_string_dup(&opts->gpg_sign, key, value);
else if (!strcmp(key, "options.strategy-option")) {
ALLOC_GROW(opts->xopts, opts->xopts_nr + 1, opts->xopts_alloc);
opts->xopts[opts->xopts_nr++] = xstrdup(value);
} else if (!strcmp(key, "options.allow-rerere-auto"))
opts->allow_rerere_auto =
git_config_bool_or_int(key, value, &error_flag) ?
RERERE_AUTOUPDATE : RERERE_NOAUTOUPDATE;
else if (!strcmp(key, "options.default-msg-cleanup")) {
opts->explicit_cleanup = 1;
opts->default_msg_cleanup = get_cleanup_mode(value, 1);
} else
return error(_("invalid key: %s"), key);
if (!error_flag)
return error(_("invalid value for %s: %s"), key, value);
return 0;
}
void parse_strategy_opts(struct replay_opts *opts, char *raw_opts)
{
int i;
char *strategy_opts_string = raw_opts;
if (*strategy_opts_string == ' ')
strategy_opts_string++;
opts->xopts_nr = split_cmdline(strategy_opts_string,
(const char ***)&opts->xopts);
for (i = 0; i < opts->xopts_nr; i++) {
const char *arg = opts->xopts[i];
skip_prefix(arg, "--", &arg);
opts->xopts[i] = xstrdup(arg);
}
}
static void read_strategy_opts(struct replay_opts *opts, struct strbuf *buf)
{
strbuf_reset(buf);
if (!read_oneliner(buf, rebase_path_strategy(), 0))
return;
opts->strategy = strbuf_detach(buf, NULL);
if (!read_oneliner(buf, rebase_path_strategy_opts(), 0))
return;
parse_strategy_opts(opts, buf->buf);
}
static int read_populate_opts(struct replay_opts *opts)
{
if (is_rebase_i(opts)) {
struct strbuf buf = STRBUF_INIT;
if (read_oneliner(&buf, rebase_path_gpg_sign_opt(), 1)) {
if (!starts_with(buf.buf, "-S"))
strbuf_reset(&buf);
else {
free(opts->gpg_sign);
opts->gpg_sign = xstrdup(buf.buf + 2);
}
strbuf_reset(&buf);
}
if (read_oneliner(&buf, rebase_path_allow_rerere_autoupdate(), 1)) {
if (!strcmp(buf.buf, "--rerere-autoupdate"))
opts->allow_rerere_auto = RERERE_AUTOUPDATE;
else if (!strcmp(buf.buf, "--no-rerere-autoupdate"))
opts->allow_rerere_auto = RERERE_NOAUTOUPDATE;
strbuf_reset(&buf);
}
if (file_exists(rebase_path_verbose()))
opts->verbose = 1;
if (file_exists(rebase_path_quiet()))
opts->quiet = 1;
if (file_exists(rebase_path_signoff())) {
opts->allow_ff = 0;
opts->signoff = 1;
}
if (file_exists(rebase_path_reschedule_failed_exec()))
opts->reschedule_failed_exec = 1;
read_strategy_opts(opts, &buf);
strbuf_release(&buf);
if (read_oneliner(&opts->current_fixups,
rebase_path_current_fixups(), 1)) {
const char *p = opts->current_fixups.buf;
opts->current_fixup_count = 1;
while ((p = strchr(p, '\n'))) {
opts->current_fixup_count++;
p++;
}
}
if (read_oneliner(&buf, rebase_path_squash_onto(), 0)) {
if (get_oid_hex(buf.buf, &opts->squash_onto) < 0)
return error(_("unusable squash-onto"));
opts->have_squash_onto = 1;
}
return 0;
}
memoize common git-path "constant" files One of the most common uses of git_path() is to pass a constant, like git_path("MERGE_MSG"). This has two drawbacks: 1. The return value is a static buffer, and the lifetime is dependent on other calls to git_path, etc. 2. There's no compile-time checking of the pathname. This is OK for a one-off (after all, we have to spell it correctly at least once), but many of these constant strings appear throughout the code. This patch introduces a series of functions to "memoize" these strings, which are essentially globals for the lifetime of the program. We compute the value once, take ownership of the buffer, and return the cached value for subsequent calls. cache.h provides a helper macro for defining these functions as one-liners, and defines a few common ones for global use. Using a macro is a little bit gross, but it does nicely document the purpose of the functions. If we need to touch them all later (e.g., because we learned how to change the git_dir variable at runtime, and need to invalidate all of the stored values), it will be much easier to have the complete list. Note that the shared-global functions have separate, manual declarations. We could do something clever with the macros (e.g., expand it to a declaration in some places, and a declaration _and_ a definition in path.c). But there aren't that many, and it's probably better to stay away from too-magical macros. Likewise, if we abandon the C preprocessor in favor of generating these with a script, we could get much fancier. E.g., normalizing "FOO/BAR-BAZ" into "git_path_foo_bar_baz". But the small amount of saved typing is probably not worth the resulting confusion to readers who want to grep for the function's definition. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-08-10 12:38:57 +03:00
if (!file_exists(git_path_opts_file()))
return 0;
/*
* The function git_parse_source(), called from git_config_from_file(),
* may die() in case of a syntactically incorrect file. We do not care
* about this case, though, because we wrote that file ourselves, so we
* are pretty certain that it is syntactically correct.
*/
if (git_config_from_file(populate_opts_cb, git_path_opts_file(), opts) < 0)
return error(_("malformed options sheet: '%s'"),
git_path_opts_file());
return 0;
}
static void write_strategy_opts(struct replay_opts *opts)
{
int i;
struct strbuf buf = STRBUF_INIT;
for (i = 0; i < opts->xopts_nr; ++i)
strbuf_addf(&buf, " --%s", opts->xopts[i]);
write_file(rebase_path_strategy_opts(), "%s\n", buf.buf);
strbuf_release(&buf);
}
int write_basic_state(struct replay_opts *opts, const char *head_name,
struct commit *onto, const char *orig_head)
{
const char *quiet = getenv("GIT_QUIET");
if (head_name)
write_file(rebase_path_head_name(), "%s\n", head_name);
if (onto)
write_file(rebase_path_onto(), "%s\n",
oid_to_hex(&onto->object.oid));
if (orig_head)
write_file(rebase_path_orig_head(), "%s\n", orig_head);
if (quiet)
write_file(rebase_path_quiet(), "%s\n", quiet);
if (opts->verbose)
write_file(rebase_path_verbose(), "%s", "");
if (opts->strategy)
write_file(rebase_path_strategy(), "%s\n", opts->strategy);
if (opts->xopts_nr > 0)
write_strategy_opts(opts);
if (opts->allow_rerere_auto == RERERE_AUTOUPDATE)
write_file(rebase_path_allow_rerere_autoupdate(), "--rerere-autoupdate\n");
else if (opts->allow_rerere_auto == RERERE_NOAUTOUPDATE)
write_file(rebase_path_allow_rerere_autoupdate(), "--no-rerere-autoupdate\n");
if (opts->gpg_sign)
write_file(rebase_path_gpg_sign_opt(), "-S%s\n", opts->gpg_sign);
if (opts->signoff)
write_file(rebase_path_signoff(), "--signoff\n");
if (opts->reschedule_failed_exec)
write_file(rebase_path_reschedule_failed_exec(), "%s", "");
return 0;
}
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
static int walk_revs_populate_todo(struct todo_list *todo_list,
struct replay_opts *opts)
{
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
enum todo_command command = opts->action == REPLAY_PICK ?
TODO_PICK : TODO_REVERT;
const char *command_string = todo_command_info[command].str;
struct commit *commit;
if (prepare_revs(opts))
return -1;
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
while ((commit = get_revision(opts->revs))) {
struct todo_item *item = append_new_todo(todo_list);
const char *commit_buffer = get_commit_buffer(commit, NULL);
const char *subject;
int subject_len;
item->command = command;
item->commit = commit;
item->arg_offset = 0;
item->arg_len = 0;
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
item->offset_in_buf = todo_list->buf.len;
subject_len = find_commit_subject(commit_buffer, &subject);
strbuf_addf(&todo_list->buf, "%s %s %.*s\n", command_string,
short_commit_name(commit), subject_len, subject);
unuse_commit_buffer(commit, commit_buffer);
}
if (!todo_list->nr)
return error(_("empty commit set passed"));
return 0;
}
static int create_seq_dir(struct repository *r)
{
enum replay_action action;
const char *in_progress_error = NULL;
const char *in_progress_advice = NULL;
unsigned int advise_skip = file_exists(git_path_revert_head(r)) ||
file_exists(git_path_cherry_pick_head(r));
if (!sequencer_get_last_command(r, &action)) {
switch (action) {
case REPLAY_REVERT:
in_progress_error = _("revert is already in progress");
in_progress_advice =
_("try \"git revert (--continue | %s--abort | --quit)\"");
break;
case REPLAY_PICK:
in_progress_error = _("cherry-pick is already in progress");
in_progress_advice =
_("try \"git cherry-pick (--continue | %s--abort | --quit)\"");
break;
default:
BUG("unexpected action in create_seq_dir");
}
}
if (in_progress_error) {
error("%s", in_progress_error);
if (advice_sequencer_in_use)
advise(in_progress_advice,
advise_skip ? "--skip | " : "");
return -1;
}
if (mkdir(git_path_seq_dir(), 0777) < 0)
return error_errno(_("could not create sequencer directory '%s'"),
git_path_seq_dir());
return 0;
}
static int save_head(const char *head)
{
struct lock_file head_lock = LOCK_INIT;
struct strbuf buf = STRBUF_INIT;
int fd;
ssize_t written;
fd = hold_lock_file_for_update(&head_lock, git_path_head_file(), 0);
if (fd < 0)
return error_errno(_("could not lock HEAD"));
strbuf_addf(&buf, "%s\n", head);
written = write_in_full(fd, buf.buf, buf.len);
strbuf_release(&buf);
if (written < 0) {
error_errno(_("could not write to '%s'"), git_path_head_file());
rollback_lock_file(&head_lock);
return -1;
}
if (commit_lock_file(&head_lock) < 0)
return error(_("failed to finalize '%s'"), git_path_head_file());
return 0;
}
static int rollback_is_safe(void)
{
struct strbuf sb = STRBUF_INIT;
struct object_id expected_head, actual_head;
if (strbuf_read_file(&sb, git_path_abort_safety_file(), 0) >= 0) {
strbuf_trim(&sb);
if (get_oid_hex(sb.buf, &expected_head)) {
strbuf_release(&sb);
die(_("could not parse %s"), git_path_abort_safety_file());
}
strbuf_release(&sb);
}
else if (errno == ENOENT)
oidclr(&expected_head);
else
die_errno(_("could not read '%s'"), git_path_abort_safety_file());
if (get_oid("HEAD", &actual_head))
oidclr(&actual_head);
return oideq(&actual_head, &expected_head);
}
static int reset_merge(const struct object_id *oid)
{
int ret;
struct argv_array argv = ARGV_ARRAY_INIT;
argv_array_pushl(&argv, "reset", "--merge", NULL);
if (!is_null_oid(oid))
argv_array_push(&argv, oid_to_hex(oid));
ret = run_command_v_opt(argv.argv, RUN_GIT_CMD);
argv_array_clear(&argv);
return ret;
}
static int rollback_single_pick(struct repository *r)
{
struct object_id head_oid;
if (!file_exists(git_path_cherry_pick_head(r)) &&
!file_exists(git_path_revert_head(r)))
return error(_("no cherry-pick or revert in progress"));
if (read_ref_full("HEAD", 0, &head_oid, NULL))
return error(_("cannot resolve HEAD"));
if (is_null_oid(&head_oid))
return error(_("cannot abort from a branch yet to be born"));
return reset_merge(&head_oid);
}
static int skip_single_pick(void)
{
struct object_id head;
if (read_ref_full("HEAD", 0, &head, NULL))
return error(_("cannot resolve HEAD"));
return reset_merge(&head);
}
int sequencer_rollback(struct repository *r, struct replay_opts *opts)
{
FILE *f;
struct object_id oid;
struct strbuf buf = STRBUF_INIT;
const char *p;
memoize common git-path "constant" files One of the most common uses of git_path() is to pass a constant, like git_path("MERGE_MSG"). This has two drawbacks: 1. The return value is a static buffer, and the lifetime is dependent on other calls to git_path, etc. 2. There's no compile-time checking of the pathname. This is OK for a one-off (after all, we have to spell it correctly at least once), but many of these constant strings appear throughout the code. This patch introduces a series of functions to "memoize" these strings, which are essentially globals for the lifetime of the program. We compute the value once, take ownership of the buffer, and return the cached value for subsequent calls. cache.h provides a helper macro for defining these functions as one-liners, and defines a few common ones for global use. Using a macro is a little bit gross, but it does nicely document the purpose of the functions. If we need to touch them all later (e.g., because we learned how to change the git_dir variable at runtime, and need to invalidate all of the stored values), it will be much easier to have the complete list. Note that the shared-global functions have separate, manual declarations. We could do something clever with the macros (e.g., expand it to a declaration in some places, and a declaration _and_ a definition in path.c). But there aren't that many, and it's probably better to stay away from too-magical macros. Likewise, if we abandon the C preprocessor in favor of generating these with a script, we could get much fancier. E.g., normalizing "FOO/BAR-BAZ" into "git_path_foo_bar_baz". But the small amount of saved typing is probably not worth the resulting confusion to readers who want to grep for the function's definition. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-08-10 12:38:57 +03:00
f = fopen(git_path_head_file(), "r");
if (!f && errno == ENOENT) {
/*
* There is no multiple-cherry-pick in progress.
* If CHERRY_PICK_HEAD or REVERT_HEAD indicates
* a single-cherry-pick in progress, abort that.
*/
return rollback_single_pick(r);
}
if (!f)
return error_errno(_("cannot open '%s'"), git_path_head_file());
if (strbuf_getline_lf(&buf, f)) {
error(_("cannot read '%s': %s"), git_path_head_file(),
memoize common git-path "constant" files One of the most common uses of git_path() is to pass a constant, like git_path("MERGE_MSG"). This has two drawbacks: 1. The return value is a static buffer, and the lifetime is dependent on other calls to git_path, etc. 2. There's no compile-time checking of the pathname. This is OK for a one-off (after all, we have to spell it correctly at least once), but many of these constant strings appear throughout the code. This patch introduces a series of functions to "memoize" these strings, which are essentially globals for the lifetime of the program. We compute the value once, take ownership of the buffer, and return the cached value for subsequent calls. cache.h provides a helper macro for defining these functions as one-liners, and defines a few common ones for global use. Using a macro is a little bit gross, but it does nicely document the purpose of the functions. If we need to touch them all later (e.g., because we learned how to change the git_dir variable at runtime, and need to invalidate all of the stored values), it will be much easier to have the complete list. Note that the shared-global functions have separate, manual declarations. We could do something clever with the macros (e.g., expand it to a declaration in some places, and a declaration _and_ a definition in path.c). But there aren't that many, and it's probably better to stay away from too-magical macros. Likewise, if we abandon the C preprocessor in favor of generating these with a script, we could get much fancier. E.g., normalizing "FOO/BAR-BAZ" into "git_path_foo_bar_baz". But the small amount of saved typing is probably not worth the resulting confusion to readers who want to grep for the function's definition. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-08-10 12:38:57 +03:00
ferror(f) ? strerror(errno) : _("unexpected end of file"));
fclose(f);
goto fail;
}
fclose(f);
if (parse_oid_hex(buf.buf, &oid, &p) || *p != '\0') {
error(_("stored pre-cherry-pick HEAD file '%s' is corrupt"),
memoize common git-path "constant" files One of the most common uses of git_path() is to pass a constant, like git_path("MERGE_MSG"). This has two drawbacks: 1. The return value is a static buffer, and the lifetime is dependent on other calls to git_path, etc. 2. There's no compile-time checking of the pathname. This is OK for a one-off (after all, we have to spell it correctly at least once), but many of these constant strings appear throughout the code. This patch introduces a series of functions to "memoize" these strings, which are essentially globals for the lifetime of the program. We compute the value once, take ownership of the buffer, and return the cached value for subsequent calls. cache.h provides a helper macro for defining these functions as one-liners, and defines a few common ones for global use. Using a macro is a little bit gross, but it does nicely document the purpose of the functions. If we need to touch them all later (e.g., because we learned how to change the git_dir variable at runtime, and need to invalidate all of the stored values), it will be much easier to have the complete list. Note that the shared-global functions have separate, manual declarations. We could do something clever with the macros (e.g., expand it to a declaration in some places, and a declaration _and_ a definition in path.c). But there aren't that many, and it's probably better to stay away from too-magical macros. Likewise, if we abandon the C preprocessor in favor of generating these with a script, we could get much fancier. E.g., normalizing "FOO/BAR-BAZ" into "git_path_foo_bar_baz". But the small amount of saved typing is probably not worth the resulting confusion to readers who want to grep for the function's definition. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-08-10 12:38:57 +03:00
git_path_head_file());
goto fail;
}
if (is_null_oid(&oid)) {
error(_("cannot abort from a branch yet to be born"));
goto fail;
}
if (!rollback_is_safe()) {
/* Do not error, just do not rollback */
warning(_("You seem to have moved HEAD. "
"Not rewinding, check your HEAD!"));
} else
if (reset_merge(&oid))
goto fail;
strbuf_release(&buf);
return sequencer_remove_state(opts);
fail:
strbuf_release(&buf);
return -1;
}
int sequencer_skip(struct repository *r, struct replay_opts *opts)
{
enum replay_action action = -1;
sequencer_get_last_command(r, &action);
/*
* Check whether the subcommand requested to skip the commit is actually
* in progress and that it's safe to skip the commit.
*
* opts->action tells us which subcommand requested to skip the commit.
* If the corresponding .git/<ACTION>_HEAD exists, we know that the
* action is in progress and we can skip the commit.
*
* Otherwise we check that the last instruction was related to the
* particular subcommand we're trying to execute and barf if that's not
* the case.
*
* Finally we check that the rollback is "safe", i.e., has the HEAD
* moved? In this case, it doesn't make sense to "reset the merge" and
* "skip the commit" as the user already handled this by committing. But
* we'd not want to barf here, instead give advice on how to proceed. We
* only need to check that when .git/<ACTION>_HEAD doesn't exist because
* it gets removed when the user commits, so if it still exists we're
* sure the user can't have committed before.
*/
switch (opts->action) {
case REPLAY_REVERT:
if (!file_exists(git_path_revert_head(r))) {
if (action != REPLAY_REVERT)
return error(_("no revert in progress"));
if (!rollback_is_safe())
goto give_advice;
}
break;
case REPLAY_PICK:
if (!file_exists(git_path_cherry_pick_head(r))) {
if (action != REPLAY_PICK)
return error(_("no cherry-pick in progress"));
if (!rollback_is_safe())
goto give_advice;
}
break;
default:
BUG("unexpected action in sequencer_skip");
}
if (skip_single_pick())
return error(_("failed to skip the commit"));
if (!is_directory(git_path_seq_dir()))
return 0;
return sequencer_continue(r, opts);
give_advice:
error(_("there is nothing to skip"));
if (advice_resolve_conflict) {
advise(_("have you committed already?\n"
"try \"git %s --continue\""),
action == REPLAY_REVERT ? "revert" : "cherry-pick");
}
return -1;
}
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
static int save_todo(struct todo_list *todo_list, struct replay_opts *opts)
{
struct lock_file todo_lock = LOCK_INIT;
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
const char *todo_path = get_todo_path(opts);
int next = todo_list->current, offset, fd;
/*
* rebase -i writes "git-rebase-todo" without the currently executing
* command, appending it to "done" instead.
*/
if (is_rebase_i(opts))
next++;
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
fd = hold_lock_file_for_update(&todo_lock, todo_path, 0);
if (fd < 0)
return error_errno(_("could not lock '%s'"), todo_path);
offset = get_item_line_offset(todo_list, next);
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
if (write_in_full(fd, todo_list->buf.buf + offset,
todo_list->buf.len - offset) < 0)
return error_errno(_("could not write to '%s'"), todo_path);
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
if (commit_lock_file(&todo_lock) < 0)
return error(_("failed to finalize '%s'"), todo_path);
if (is_rebase_i(opts) && next > 0) {
const char *done = rebase_path_done();
int fd = open(done, O_CREAT | O_WRONLY | O_APPEND, 0666);
int ret = 0;
if (fd < 0)
return 0;
if (write_in_full(fd, get_item_line(todo_list, next - 1),
get_item_line_length(todo_list, next - 1))
< 0)
ret = error_errno(_("could not write to '%s'"), done);
if (close(fd) < 0)
ret = error_errno(_("failed to finalize '%s'"), done);
return ret;
}
return 0;
}
static int save_opts(struct replay_opts *opts)
{
memoize common git-path "constant" files One of the most common uses of git_path() is to pass a constant, like git_path("MERGE_MSG"). This has two drawbacks: 1. The return value is a static buffer, and the lifetime is dependent on other calls to git_path, etc. 2. There's no compile-time checking of the pathname. This is OK for a one-off (after all, we have to spell it correctly at least once), but many of these constant strings appear throughout the code. This patch introduces a series of functions to "memoize" these strings, which are essentially globals for the lifetime of the program. We compute the value once, take ownership of the buffer, and return the cached value for subsequent calls. cache.h provides a helper macro for defining these functions as one-liners, and defines a few common ones for global use. Using a macro is a little bit gross, but it does nicely document the purpose of the functions. If we need to touch them all later (e.g., because we learned how to change the git_dir variable at runtime, and need to invalidate all of the stored values), it will be much easier to have the complete list. Note that the shared-global functions have separate, manual declarations. We could do something clever with the macros (e.g., expand it to a declaration in some places, and a declaration _and_ a definition in path.c). But there aren't that many, and it's probably better to stay away from too-magical macros. Likewise, if we abandon the C preprocessor in favor of generating these with a script, we could get much fancier. E.g., normalizing "FOO/BAR-BAZ" into "git_path_foo_bar_baz". But the small amount of saved typing is probably not worth the resulting confusion to readers who want to grep for the function's definition. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-08-10 12:38:57 +03:00
const char *opts_file = git_path_opts_file();
int res = 0;
if (opts->no_commit)
res |= git_config_set_in_file_gently(opts_file,
"options.no-commit", "true");
if (opts->edit)
res |= git_config_set_in_file_gently(opts_file,
"options.edit", "true");
if (opts->allow_empty)
res |= git_config_set_in_file_gently(opts_file,
"options.allow-empty", "true");
if (opts->allow_empty_message)
res |= git_config_set_in_file_gently(opts_file,
"options.allow-empty-message", "true");
if (opts->keep_redundant_commits)
res |= git_config_set_in_file_gently(opts_file,
"options.keep-redundant-commits", "true");
if (opts->signoff)
res |= git_config_set_in_file_gently(opts_file,
"options.signoff", "true");
if (opts->record_origin)
res |= git_config_set_in_file_gently(opts_file,
"options.record-origin", "true");
if (opts->allow_ff)
res |= git_config_set_in_file_gently(opts_file,
"options.allow-ff", "true");
if (opts->mainline) {
struct strbuf buf = STRBUF_INIT;
strbuf_addf(&buf, "%d", opts->mainline);
res |= git_config_set_in_file_gently(opts_file,
"options.mainline", buf.buf);
strbuf_release(&buf);
}
if (opts->strategy)
res |= git_config_set_in_file_gently(opts_file,
"options.strategy", opts->strategy);
if (opts->gpg_sign)
res |= git_config_set_in_file_gently(opts_file,
"options.gpg-sign", opts->gpg_sign);
if (opts->xopts) {
int i;
for (i = 0; i < opts->xopts_nr; i++)
res |= git_config_set_multivar_in_file_gently(opts_file,
"options.strategy-option",
opts->xopts[i], "^$", 0);
}
if (opts->allow_rerere_auto)
res |= git_config_set_in_file_gently(opts_file,
"options.allow-rerere-auto",
opts->allow_rerere_auto == RERERE_AUTOUPDATE ?
"true" : "false");
if (opts->explicit_cleanup)
res |= git_config_set_in_file_gently(opts_file,
"options.default-msg-cleanup",
describe_cleanup_mode(opts->default_msg_cleanup));
return res;
}
static int make_patch(struct repository *r,
struct commit *commit,
struct replay_opts *opts)
{
struct strbuf buf = STRBUF_INIT;
struct rev_info log_tree_opt;
const char *subject, *p;
int res = 0;
p = short_commit_name(commit);
if (write_message(p, strlen(p), rebase_path_stopped_sha(), 1) < 0)
return -1;
if (update_ref("rebase", "REBASE_HEAD", &commit->object.oid,
NULL, REF_NO_DEREF, UPDATE_REFS_MSG_ON_ERR))
res |= error(_("could not update %s"), "REBASE_HEAD");
strbuf_addf(&buf, "%s/patch", get_dir(opts));
memset(&log_tree_opt, 0, sizeof(log_tree_opt));
repo_init_revisions(r, &log_tree_opt, NULL);
log_tree_opt.abbrev = 0;
log_tree_opt.diff = 1;
log_tree_opt.diffopt.output_format = DIFF_FORMAT_PATCH;
log_tree_opt.disable_stdin = 1;
log_tree_opt.no_commit_id = 1;
log_tree_opt.diffopt.file = fopen(buf.buf, "w");
log_tree_opt.diffopt.use_color = GIT_COLOR_NEVER;
if (!log_tree_opt.diffopt.file)
res |= error_errno(_("could not open '%s'"), buf.buf);
else {
res |= log_tree_commit(&log_tree_opt, commit);
fclose(log_tree_opt.diffopt.file);
}
strbuf_reset(&buf);
strbuf_addf(&buf, "%s/message", get_dir(opts));
if (!file_exists(buf.buf)) {
const char *commit_buffer = get_commit_buffer(commit, NULL);
find_commit_subject(commit_buffer, &subject);
res |= write_message(subject, strlen(subject), buf.buf, 1);
unuse_commit_buffer(commit, commit_buffer);
}
strbuf_release(&buf);
return res;
}
static int intend_to_amend(void)
{
struct object_id head;
char *p;
if (get_oid("HEAD", &head))
return error(_("cannot read HEAD"));
p = oid_to_hex(&head);
return write_message(p, strlen(p), rebase_path_amend(), 1);
}
static int error_with_patch(struct repository *r,
struct commit *commit,
const char *subject, int subject_len,
struct replay_opts *opts,
int exit_code, int to_amend)
{
if (commit) {
if (make_patch(r, commit, opts))
return -1;
} else if (copy_file(rebase_path_message(),
git_path_merge_msg(r), 0666))
return error(_("unable to copy '%s' to '%s'"),
git_path_merge_msg(r), rebase_path_message());
if (to_amend) {
if (intend_to_amend())
return -1;
fprintf(stderr,
_("You can amend the commit now, with\n"
"\n"
" git commit --amend %s\n"
"\n"
"Once you are satisfied with your changes, run\n"
"\n"
" git rebase --continue\n"),
gpg_sign_opt_quoted(opts));
} else if (exit_code) {
if (commit)
fprintf_ln(stderr, _("Could not apply %s... %.*s"),
short_commit_name(commit), subject_len, subject);
else
/*
* We don't have the hash of the parent so
* just print the line from the todo file.
*/
fprintf_ln(stderr, _("Could not merge %.*s"),
subject_len, subject);
}
return exit_code;
}
static int error_failed_squash(struct repository *r,
struct commit *commit,
struct replay_opts *opts,
int subject_len,
const char *subject)
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
{
if (copy_file(rebase_path_message(), rebase_path_squash_msg(), 0666))
return error(_("could not copy '%s' to '%s'"),
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
rebase_path_squash_msg(), rebase_path_message());
unlink(git_path_merge_msg(r));
if (copy_file(git_path_merge_msg(r), rebase_path_message(), 0666))
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
return error(_("could not copy '%s' to '%s'"),
rebase_path_message(),
git_path_merge_msg(r));
return error_with_patch(r, commit, subject, subject_len, opts, 1, 0);
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
}
static int do_exec(struct repository *r, const char *command_line)
{
struct argv_array child_env = ARGV_ARRAY_INIT;
const char *child_argv[] = { NULL, NULL };
int dirty, status;
fprintf(stderr, "Executing: %s\n", command_line);
child_argv[0] = command_line;
argv_array_pushf(&child_env, "GIT_DIR=%s", absolute_path(get_git_dir()));
argv_array_pushf(&child_env, "GIT_WORK_TREE=%s",
absolute_path(get_git_work_tree()));
status = run_command_v_opt_cd_env(child_argv, RUN_USING_SHELL, NULL,
child_env.argv);
/* force re-reading of the cache */
if (discard_index(r->index) < 0 || repo_read_index(r) < 0)
return error(_("could not read index"));
dirty = require_clean_work_tree(r, "rebase", NULL, 1, 1);
if (status) {
warning(_("execution failed: %s\n%s"
"You can fix the problem, and then run\n"
"\n"
" git rebase --continue\n"
"\n"),
command_line,
dirty ? N_("and made changes to the index and/or the "
"working tree\n") : "");
if (status == 127)
/* command not found */
status = 1;
} else if (dirty) {
warning(_("execution succeeded: %s\nbut "
"left changes to the index and/or the working tree\n"
"Commit or stash your changes, and then run\n"
"\n"
" git rebase --continue\n"
"\n"), command_line);
status = 1;
}
argv_array_clear(&child_env);
return status;
}
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
static int safe_append(const char *filename, const char *fmt, ...)
{
va_list ap;
struct lock_file lock = LOCK_INIT;
int fd = hold_lock_file_for_update(&lock, filename,
LOCK_REPORT_ON_ERROR);
struct strbuf buf = STRBUF_INIT;
if (fd < 0)
return -1;
if (strbuf_read_file(&buf, filename, 0) < 0 && errno != ENOENT) {
error_errno(_("could not read '%s'"), filename);
rollback_lock_file(&lock);
return -1;
}
strbuf_complete(&buf, '\n');
va_start(ap, fmt);
strbuf_vaddf(&buf, fmt, ap);
va_end(ap);
if (write_in_full(fd, buf.buf, buf.len) < 0) {
error_errno(_("could not write to '%s'"), filename);
strbuf_release(&buf);
rollback_lock_file(&lock);
return -1;
}
if (commit_lock_file(&lock) < 0) {
strbuf_release(&buf);
rollback_lock_file(&lock);
return error(_("failed to finalize '%s'"), filename);
}
strbuf_release(&buf);
return 0;
}
static int do_label(struct repository *r, const char *name, int len)
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
{
struct ref_store *refs = get_main_ref_store(r);
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
struct ref_transaction *transaction;
struct strbuf ref_name = STRBUF_INIT, err = STRBUF_INIT;
struct strbuf msg = STRBUF_INIT;
int ret = 0;
struct object_id head_oid;
if (len == 1 && *name == '#')
return error(_("illegal label name: '%.*s'"), len, name);
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
strbuf_addf(&ref_name, "refs/rewritten/%.*s", len, name);
strbuf_addf(&msg, "rebase -i (label) '%.*s'", len, name);
transaction = ref_store_transaction_begin(refs, &err);
if (!transaction) {
error("%s", err.buf);
ret = -1;
} else if (get_oid("HEAD", &head_oid)) {
error(_("could not read HEAD"));
ret = -1;
} else if (ref_transaction_update(transaction, ref_name.buf, &head_oid,
NULL, 0, msg.buf, &err) < 0 ||
ref_transaction_commit(transaction, &err)) {
error("%s", err.buf);
ret = -1;
}
ref_transaction_free(transaction);
strbuf_release(&err);
strbuf_release(&msg);
if (!ret)
ret = safe_append(rebase_path_refs_to_delete(),
"%s\n", ref_name.buf);
strbuf_release(&ref_name);
return ret;
}
static const char *reflog_message(struct replay_opts *opts,
const char *sub_action, const char *fmt, ...);
static int do_reset(struct repository *r,
const char *name, int len,
struct replay_opts *opts)
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
{
struct strbuf ref_name = STRBUF_INIT;
struct object_id oid;
struct lock_file lock = LOCK_INIT;
struct tree_desc desc;
struct tree *tree;
struct unpack_trees_options unpack_tree_opts;
int ret = 0;
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
if (repo_hold_locked_index(r, &lock, LOCK_REPORT_ON_ERROR) < 0)
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
return -1;
if (len == 10 && !strncmp("[new root]", name, len)) {
if (!opts->have_squash_onto) {
const char *hex;
if (commit_tree("", 0, the_hash_algo->empty_tree,
NULL, &opts->squash_onto,
NULL, NULL))
return error(_("writing fake root commit"));
opts->have_squash_onto = 1;
hex = oid_to_hex(&opts->squash_onto);
if (write_message(hex, strlen(hex),
rebase_path_squash_onto(), 0))
return error(_("writing squash-onto"));
}
oidcpy(&oid, &opts->squash_onto);
} else {
int i;
/* Determine the length of the label */
for (i = 0; i < len; i++)
if (isspace(name[i]))
break;
len = i;
strbuf_addf(&ref_name, "refs/rewritten/%.*s", len, name);
if (get_oid(ref_name.buf, &oid) &&
get_oid(ref_name.buf + strlen("refs/rewritten/"), &oid)) {
error(_("could not read '%s'"), ref_name.buf);
rollback_lock_file(&lock);
strbuf_release(&ref_name);
return -1;
}
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
}
memset(&unpack_tree_opts, 0, sizeof(unpack_tree_opts));
setup_unpack_trees_porcelain(&unpack_tree_opts, "reset");
unpack_tree_opts.head_idx = 1;
unpack_tree_opts.src_index = r->index;
unpack_tree_opts.dst_index = r->index;
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
unpack_tree_opts.fn = oneway_merge;
unpack_tree_opts.merge = 1;
unpack_tree_opts.update = 1;
if (repo_read_index_unmerged(r)) {
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
rollback_lock_file(&lock);
strbuf_release(&ref_name);
return error_resolve_conflict(_(action_name(opts)));
}
if (!fill_tree_descriptor(r, &desc, &oid)) {
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
error(_("failed to find tree of %s"), oid_to_hex(&oid));
rollback_lock_file(&lock);
free((void *)desc.buffer);
strbuf_release(&ref_name);
return -1;
}
if (unpack_trees(1, &desc, &unpack_tree_opts)) {
rollback_lock_file(&lock);
free((void *)desc.buffer);
strbuf_release(&ref_name);
return -1;
}
tree = parse_tree_indirect(&oid);
prime_cache_tree(r, r->index, tree);
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
if (write_locked_index(r->index, &lock, COMMIT_LOCK) < 0)
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
ret = error(_("could not write index"));
free((void *)desc.buffer);
if (!ret)
ret = update_ref(reflog_message(opts, "reset", "'%.*s'",
len, name), "HEAD", &oid,
NULL, 0, UPDATE_REFS_MSG_ON_ERR);
strbuf_release(&ref_name);
return ret;
}
static struct commit *lookup_label(const char *label, int len,
struct strbuf *buf)
{
struct commit *commit;
strbuf_reset(buf);
strbuf_addf(buf, "refs/rewritten/%.*s", len, label);
commit = lookup_commit_reference_by_name(buf->buf);
if (!commit) {
/* fall back to non-rewritten ref or commit */
strbuf_splice(buf, 0, strlen("refs/rewritten/"), "", 0);
commit = lookup_commit_reference_by_name(buf->buf);
}
if (!commit)
error(_("could not resolve '%s'"), buf->buf);
return commit;
}
static int do_merge(struct repository *r,
struct commit *commit,
const char *arg, int arg_len,
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
int flags, struct replay_opts *opts)
{
int run_commit_flags = (flags & TODO_EDIT_MERGE_MSG) ?
EDIT_MSG | VERIFY_MSG : 0;
struct strbuf ref_name = STRBUF_INIT;
struct commit *head_commit, *merge_commit, *i;
struct commit_list *bases, *j, *reversed = NULL;
struct commit_list *to_merge = NULL, **tail = &to_merge;
const char *strategy = !opts->xopts_nr &&
(!opts->strategy || !strcmp(opts->strategy, "recursive")) ?
NULL : opts->strategy;
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
struct merge_options o;
int merge_arg_len, oneline_offset, can_fast_forward, ret, k;
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
static struct lock_file lock;
const char *p;
if (repo_hold_locked_index(r, &lock, LOCK_REPORT_ON_ERROR) < 0) {
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
ret = -1;
goto leave_merge;
}
head_commit = lookup_commit_reference_by_name("HEAD");
if (!head_commit) {
ret = error(_("cannot merge without a current revision"));
goto leave_merge;
}
/*
* For octopus merges, the arg starts with the list of revisions to be
* merged. The list is optionally followed by '#' and the oneline.
*/
merge_arg_len = oneline_offset = arg_len;
for (p = arg; p - arg < arg_len; p += strspn(p, " \t\n")) {
if (!*p)
break;
if (*p == '#' && (!p[1] || isspace(p[1]))) {
p += 1 + strspn(p + 1, " \t\n");
oneline_offset = p - arg;
break;
}
k = strcspn(p, " \t\n");
if (!k)
continue;
merge_commit = lookup_label(p, k, &ref_name);
if (!merge_commit) {
ret = error(_("unable to parse '%.*s'"), k, p);
goto leave_merge;
}
tail = &commit_list_insert(merge_commit, tail)->next;
p += k;
merge_arg_len = p - arg;
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
}
if (!to_merge) {
ret = error(_("nothing to merge: '%.*s'"), arg_len, arg);
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
goto leave_merge;
}
if (opts->have_squash_onto &&
oideq(&head_commit->object.oid, &opts->squash_onto)) {
/*
* When the user tells us to "merge" something into a
* "[new root]", let's simply fast-forward to the merge head.
*/
rollback_lock_file(&lock);
if (to_merge->next)
ret = error(_("octopus merge cannot be executed on "
"top of a [new root]"));
else
ret = fast_forward_to(r, &to_merge->item->object.oid,
&head_commit->object.oid, 0,
opts);
goto leave_merge;
}
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
if (commit) {
const char *message = get_commit_buffer(commit, NULL);
const char *body;
int len;
if (!message) {
ret = error(_("could not get commit message of '%s'"),
oid_to_hex(&commit->object.oid));
goto leave_merge;
}
write_author_script(message);
find_commit_subject(message, &body);
len = strlen(body);
ret = write_message(body, len, git_path_merge_msg(r), 0);
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
unuse_commit_buffer(commit, message);
if (ret) {
error_errno(_("could not write '%s'"),
git_path_merge_msg(r));
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
goto leave_merge;
}
} else {
struct strbuf buf = STRBUF_INIT;
int len;
strbuf_addf(&buf, "author %s", git_author_info(0));
write_author_script(buf.buf);
strbuf_reset(&buf);
if (oneline_offset < arg_len) {
p = arg + oneline_offset;
len = arg_len - oneline_offset;
} else {
strbuf_addf(&buf, "Merge %s '%.*s'",
to_merge->next ? "branches" : "branch",
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
merge_arg_len, arg);
p = buf.buf;
len = buf.len;
}
ret = write_message(p, len, git_path_merge_msg(r), 0);
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
strbuf_release(&buf);
if (ret) {
error_errno(_("could not write '%s'"),
git_path_merge_msg(r));
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
goto leave_merge;
}
}
/*
* If HEAD is not identical to the first parent of the original merge
* commit, we cannot fast-forward.
*/
can_fast_forward = opts->allow_ff && commit && commit->parents &&
oideq(&commit->parents->item->object.oid,
&head_commit->object.oid);
/*
* If any merge head is different from the original one, we cannot
* fast-forward.
*/
if (can_fast_forward) {
struct commit_list *p = commit->parents->next;
for (j = to_merge; j && p; j = j->next, p = p->next)
if (!oideq(&j->item->object.oid,
&p->item->object.oid)) {
can_fast_forward = 0;
break;
}
/*
* If the number of merge heads differs from the original merge
* commit, we cannot fast-forward.
*/
if (j || p)
can_fast_forward = 0;
}
if (can_fast_forward) {
rollback_lock_file(&lock);
ret = fast_forward_to(r, &commit->object.oid,
&head_commit->object.oid, 0, opts);
if (flags & TODO_EDIT_MERGE_MSG) {
run_commit_flags |= AMEND_MSG;
goto fast_forward_edit;
}
goto leave_merge;
}
if (strategy || to_merge->next) {
/* Octopus merge */
struct child_process cmd = CHILD_PROCESS_INIT;
if (read_env_script(&cmd.env_array)) {
const char *gpg_opt = gpg_sign_opt_quoted(opts);
ret = error(_(staged_changes_advice), gpg_opt, gpg_opt);
goto leave_merge;
}
cmd.git_cmd = 1;
argv_array_push(&cmd.args, "merge");
argv_array_push(&cmd.args, "-s");
if (!strategy)
argv_array_push(&cmd.args, "octopus");
else {
argv_array_push(&cmd.args, strategy);
for (k = 0; k < opts->xopts_nr; k++)
argv_array_pushf(&cmd.args,
"-X%s", opts->xopts[k]);
}
argv_array_push(&cmd.args, "--no-edit");
argv_array_push(&cmd.args, "--no-ff");
argv_array_push(&cmd.args, "--no-log");
argv_array_push(&cmd.args, "--no-stat");
argv_array_push(&cmd.args, "-F");
argv_array_push(&cmd.args, git_path_merge_msg(r));
if (opts->gpg_sign)
argv_array_push(&cmd.args, opts->gpg_sign);
/* Add the tips to be merged */
for (j = to_merge; j; j = j->next)
argv_array_push(&cmd.args,
oid_to_hex(&j->item->object.oid));
strbuf_release(&ref_name);
unlink(git_path_cherry_pick_head(r));
rollback_lock_file(&lock);
rollback_lock_file(&lock);
ret = run_command(&cmd);
/* force re-reading of the cache */
if (!ret && (discard_index(r->index) < 0 ||
repo_read_index(r) < 0))
ret = error(_("could not read index"));
goto leave_merge;
}
merge_commit = to_merge->item;
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
bases = get_merge_bases(head_commit, merge_commit);
if (bases && oideq(&merge_commit->object.oid,
&bases->item->object.oid)) {
ret = 0;
/* skip merging an ancestor of HEAD */
goto leave_merge;
}
write_message(oid_to_hex(&merge_commit->object.oid), the_hash_algo->hexsz,
git_path_merge_head(r), 0);
write_message("no-ff", 5, git_path_merge_mode(r), 0);
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
for (j = bases; j; j = j->next)
commit_list_insert(j->item, &reversed);
free_commit_list(bases);
repo_read_index(r);
init_merge_options(&o, r);
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
o.branch1 = "HEAD";
o.branch2 = ref_name.buf;
o.buffer_output = 2;
ret = merge_recursive(&o, head_commit, merge_commit, reversed, &i);
if (ret <= 0)
fputs(o.obuf.buf, stdout);
strbuf_release(&o.obuf);
if (ret < 0) {
error(_("could not even attempt to merge '%.*s'"),
merge_arg_len, arg);
goto leave_merge;
}
/*
* The return value of merge_recursive() is 1 on clean, and 0 on
* unclean merge.
*
* Let's reverse that, so that do_merge() returns 0 upon success and
* 1 upon failed merge (keeping the return value -1 for the cases where
* we will want to reschedule the `merge` command).
*/
ret = !ret;
if (r->index->cache_changed &&
write_locked_index(r->index, &lock, COMMIT_LOCK)) {
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
ret = error(_("merge: Unable to write new index file"));
goto leave_merge;
}
rollback_lock_file(&lock);
if (ret)
repo_rerere(r, opts->allow_rerere_auto);
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
else
/*
* In case of problems, we now want to return a positive
* value (a negative one would indicate that the `merge`
* command needs to be rescheduled).
*/
fast_forward_edit:
ret = !!run_git_commit(r, git_path_merge_msg(r), opts,
run_commit_flags);
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
leave_merge:
strbuf_release(&ref_name);
rollback_lock_file(&lock);
free_commit_list(to_merge);
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
return ret;
}
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
static int is_final_fixup(struct todo_list *todo_list)
{
int i = todo_list->current;
if (!is_fixup(todo_list->items[i].command))
return 0;
while (++i < todo_list->nr)
if (is_fixup(todo_list->items[i].command))
return 0;
else if (!is_noop(todo_list->items[i].command))
break;
return 1;
}
static enum todo_command peek_command(struct todo_list *todo_list, int offset)
{
int i;
for (i = todo_list->current + offset; i < todo_list->nr; i++)
if (!is_noop(todo_list->items[i].command))
return todo_list->items[i].command;
return -1;
}
static int apply_autostash(struct replay_opts *opts)
{
struct strbuf stash_sha1 = STRBUF_INIT;
struct child_process child = CHILD_PROCESS_INIT;
int ret = 0;
if (!read_oneliner(&stash_sha1, rebase_path_autostash(), 1)) {
strbuf_release(&stash_sha1);
return 0;
}
strbuf_trim(&stash_sha1);
child.git_cmd = 1;
child.no_stdout = 1;
child.no_stderr = 1;
argv_array_push(&child.args, "stash");
argv_array_push(&child.args, "apply");
argv_array_push(&child.args, stash_sha1.buf);
if (!run_command(&child))
fprintf(stderr, _("Applied autostash.\n"));
else {
struct child_process store = CHILD_PROCESS_INIT;
store.git_cmd = 1;
argv_array_push(&store.args, "stash");
argv_array_push(&store.args, "store");
argv_array_push(&store.args, "-m");
argv_array_push(&store.args, "autostash");
argv_array_push(&store.args, "-q");
argv_array_push(&store.args, stash_sha1.buf);
if (run_command(&store))
ret = error(_("cannot store %s"), stash_sha1.buf);
else
fprintf(stderr,
_("Applying autostash resulted in conflicts.\n"
"Your changes are safe in the stash.\n"
"You can run \"git stash pop\" or"
" \"git stash drop\" at any time.\n"));
}
strbuf_release(&stash_sha1);
return ret;
}
static const char *reflog_message(struct replay_opts *opts,
const char *sub_action, const char *fmt, ...)
{
va_list ap;
static struct strbuf buf = STRBUF_INIT;
va_start(ap, fmt);
strbuf_reset(&buf);
strbuf_addstr(&buf, action_name(opts));
if (sub_action)
strbuf_addf(&buf, " (%s)", sub_action);
if (fmt) {
strbuf_addstr(&buf, ": ");
strbuf_vaddf(&buf, fmt, ap);
}
va_end(ap);
return buf.buf;
}
static int run_git_checkout(struct repository *r, struct replay_opts *opts,
const char *commit, const char *action)
{
struct child_process cmd = CHILD_PROCESS_INIT;
int ret;
cmd.git_cmd = 1;
argv_array_push(&cmd.args, "checkout");
argv_array_push(&cmd.args, commit);
argv_array_pushf(&cmd.env_array, GIT_REFLOG_ACTION "=%s", action);
if (opts->verbose)
ret = run_command(&cmd);
else
ret = run_command_silent_on_success(&cmd);
if (!ret)
discard_index(r->index);
return ret;
}
int prepare_branch_to_be_rebased(struct repository *r, struct replay_opts *opts,
const char *commit)
{
const char *action;
if (commit && *commit) {
action = reflog_message(opts, "start", "checkout %s", commit);
if (run_git_checkout(r, opts, commit, action))
return error(_("could not checkout %s"), commit);
}
return 0;
}
static int checkout_onto(struct repository *r, struct replay_opts *opts,
const char *onto_name, const struct object_id *onto,
const char *orig_head)
{
struct object_id oid;
const char *action = reflog_message(opts, "start", "checkout %s", onto_name);
if (get_oid(orig_head, &oid))
return error(_("%s: not a valid OID"), orig_head);
if (run_git_checkout(r, opts, oid_to_hex(onto), action)) {
apply_autostash(opts);
sequencer_remove_state(opts);
return error(_("could not detach HEAD"));
}
return update_ref(NULL, "ORIG_HEAD", &oid, NULL, 0, UPDATE_REFS_MSG_ON_ERR);
}
static int stopped_at_head(struct repository *r)
{
struct object_id head;
struct commit *commit;
struct commit_message message;
if (get_oid("HEAD", &head) ||
!(commit = lookup_commit(r, &head)) ||
parse_commit(commit) || get_message(commit, &message))
fprintf(stderr, _("Stopped at HEAD\n"));
else {
fprintf(stderr, _("Stopped at %s\n"), message.label);
free_message(commit, &message);
}
return 0;
}
static const char rescheduled_advice[] =
N_("Could not execute the todo command\n"
"\n"
" %.*s"
"\n"
"It has been rescheduled; To edit the command before continuing, please\n"
"edit the todo list first:\n"
"\n"
" git rebase --edit-todo\n"
" git rebase --continue\n");
static int pick_commits(struct repository *r,
struct todo_list *todo_list,
struct replay_opts *opts)
{
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
int res = 0, reschedule = 0;
setenv(GIT_REFLOG_ACTION, action_name(opts), 0);
if (opts->allow_ff)
assert(!(opts->signoff || opts->no_commit ||
opts->record_origin || opts->edit));
if (read_and_refresh_cache(r, opts))
return -1;
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
while (todo_list->current < todo_list->nr) {
struct todo_item *item = todo_list->items + todo_list->current;
const char *arg = todo_item_get_arg(todo_list, item);
int check_todo = 0;
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
if (save_todo(todo_list, opts))
return -1;
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
if (is_rebase_i(opts)) {
if (item->command != TODO_COMMENT) {
FILE *f = fopen(rebase_path_msgnum(), "w");
todo_list->done_nr++;
if (f) {
fprintf(f, "%d\n", todo_list->done_nr);
fclose(f);
}
if (!opts->quiet)
fprintf(stderr, "Rebasing (%d/%d)%s",
todo_list->done_nr,
todo_list->total_nr,
opts->verbose ? "\n" : "\r");
}
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
unlink(rebase_path_message());
unlink(rebase_path_author_script());
unlink(rebase_path_stopped_sha());
unlink(rebase_path_amend());
unlink(git_path_merge_head(r));
delete_ref(NULL, "REBASE_HEAD", NULL, REF_NO_DEREF);
rebase: fix garbled progress display with '-x' When running a command with the 'exec' instruction during an interactive rebase session, or for a range of commits using 'git rebase -x', the output can be a bit garbled when the name of the command is short enough: $ git rebase -x true HEAD~5 Executing: true Executing: true Executing: true Executing: true Executing: true) Successfully rebased and updated refs/heads/master. Note the ')' at the end of the last line. It gets more garbled as the range of commits increases: $ git rebase -x true HEAD~50 Executing: true) [ repeated 3 more times ] Executing: true0) [ repeated 44 more times ] Executing: true00) Successfully rebased and updated refs/heads/master. Those extra numbers and ')' are remnants of the previously displayed "Rebasing (N/M)" progress lines that are usually completely overwritten by the "Executing: <cmd>" lines, unless 'cmd' is short and the "N/M" part is long. Make sure that the previously displayed "Rebasing (N/M)" line is cleared by using the term_clear_line() helper function added in the previous patch. Do so only when not being '--verbose', because in that case these "Rebasing (N/M)" lines are not printed as progress (i.e. as lines with '\r' at the end), but as "regular" output (with '\n' at the end). A couple of other rebase commands print similar messages, e.g. "Stopped at <abbrev-oid>... <subject>" for the 'edit' or 'break' commands, or the "Successfully rebased and updated <full-ref>." at the very end. These are so long that they practically always overwrite that "Rebasing (N/M)" progress line, but let's be prudent, and clear the last line before printing these, too. In 't3420-rebase-autostash.sh' two helper functions prepare the expected output of four tests that check the full output of 'git rebase' and thus are affected by this change, so adjust their expectations to account for the new line clearing. Note that this patch doesn't completely eliminate the possibility of similar garbled outputs, e.g. some error messages from rebase or the "Auto-merging <file>" message from within the depths of the merge machinery might not be long enough to completely cover the last "Rebasing (N/M)" line. This patch doesn't do anything about them, because dealing with them individually would result in way too much churn, while having a catch-all term_clear_line() call in the common code path of pick_commits() would hide the "Rebasing (N/M)" line way too soon, and it would either flicker or be invisible. Signed-off-by: SZEDER Gábor <szeder.dev@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2019-06-27 16:42:48 +03:00
if (item->command == TODO_BREAK) {
if (!opts->verbose)
term_clear_line();
return stopped_at_head(r);
rebase: fix garbled progress display with '-x' When running a command with the 'exec' instruction during an interactive rebase session, or for a range of commits using 'git rebase -x', the output can be a bit garbled when the name of the command is short enough: $ git rebase -x true HEAD~5 Executing: true Executing: true Executing: true Executing: true Executing: true) Successfully rebased and updated refs/heads/master. Note the ')' at the end of the last line. It gets more garbled as the range of commits increases: $ git rebase -x true HEAD~50 Executing: true) [ repeated 3 more times ] Executing: true0) [ repeated 44 more times ] Executing: true00) Successfully rebased and updated refs/heads/master. Those extra numbers and ')' are remnants of the previously displayed "Rebasing (N/M)" progress lines that are usually completely overwritten by the "Executing: <cmd>" lines, unless 'cmd' is short and the "N/M" part is long. Make sure that the previously displayed "Rebasing (N/M)" line is cleared by using the term_clear_line() helper function added in the previous patch. Do so only when not being '--verbose', because in that case these "Rebasing (N/M)" lines are not printed as progress (i.e. as lines with '\r' at the end), but as "regular" output (with '\n' at the end). A couple of other rebase commands print similar messages, e.g. "Stopped at <abbrev-oid>... <subject>" for the 'edit' or 'break' commands, or the "Successfully rebased and updated <full-ref>." at the very end. These are so long that they practically always overwrite that "Rebasing (N/M)" progress line, but let's be prudent, and clear the last line before printing these, too. In 't3420-rebase-autostash.sh' two helper functions prepare the expected output of four tests that check the full output of 'git rebase' and thus are affected by this change, so adjust their expectations to account for the new line clearing. Note that this patch doesn't completely eliminate the possibility of similar garbled outputs, e.g. some error messages from rebase or the "Auto-merging <file>" message from within the depths of the merge machinery might not be long enough to completely cover the last "Rebasing (N/M)" line. This patch doesn't do anything about them, because dealing with them individually would result in way too much churn, while having a catch-all term_clear_line() call in the common code path of pick_commits() would hide the "Rebasing (N/M)" line way too soon, and it would either flicker or be invisible. Signed-off-by: SZEDER Gábor <szeder.dev@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2019-06-27 16:42:48 +03:00
}
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
}
if (item->command <= TODO_SQUASH) {
if (is_rebase_i(opts))
setenv("GIT_REFLOG_ACTION", reflog_message(opts,
command_to_string(item->command), NULL),
1);
res = do_pick_commit(r, item->command, item->commit,
opts, is_final_fixup(todo_list),
&check_todo);
if (is_rebase_i(opts) && res < 0) {
/* Reschedule */
advise(_(rescheduled_advice),
get_item_line_length(todo_list,
todo_list->current),
get_item_line(todo_list,
todo_list->current));
todo_list->current--;
if (save_todo(todo_list, opts))
return -1;
}
if (item->command == TODO_EDIT) {
struct commit *commit = item->commit;
rebase: fix garbled progress display with '-x' When running a command with the 'exec' instruction during an interactive rebase session, or for a range of commits using 'git rebase -x', the output can be a bit garbled when the name of the command is short enough: $ git rebase -x true HEAD~5 Executing: true Executing: true Executing: true Executing: true Executing: true) Successfully rebased and updated refs/heads/master. Note the ')' at the end of the last line. It gets more garbled as the range of commits increases: $ git rebase -x true HEAD~50 Executing: true) [ repeated 3 more times ] Executing: true0) [ repeated 44 more times ] Executing: true00) Successfully rebased and updated refs/heads/master. Those extra numbers and ')' are remnants of the previously displayed "Rebasing (N/M)" progress lines that are usually completely overwritten by the "Executing: <cmd>" lines, unless 'cmd' is short and the "N/M" part is long. Make sure that the previously displayed "Rebasing (N/M)" line is cleared by using the term_clear_line() helper function added in the previous patch. Do so only when not being '--verbose', because in that case these "Rebasing (N/M)" lines are not printed as progress (i.e. as lines with '\r' at the end), but as "regular" output (with '\n' at the end). A couple of other rebase commands print similar messages, e.g. "Stopped at <abbrev-oid>... <subject>" for the 'edit' or 'break' commands, or the "Successfully rebased and updated <full-ref>." at the very end. These are so long that they practically always overwrite that "Rebasing (N/M)" progress line, but let's be prudent, and clear the last line before printing these, too. In 't3420-rebase-autostash.sh' two helper functions prepare the expected output of four tests that check the full output of 'git rebase' and thus are affected by this change, so adjust their expectations to account for the new line clearing. Note that this patch doesn't completely eliminate the possibility of similar garbled outputs, e.g. some error messages from rebase or the "Auto-merging <file>" message from within the depths of the merge machinery might not be long enough to completely cover the last "Rebasing (N/M)" line. This patch doesn't do anything about them, because dealing with them individually would result in way too much churn, while having a catch-all term_clear_line() call in the common code path of pick_commits() would hide the "Rebasing (N/M)" line way too soon, and it would either flicker or be invisible. Signed-off-by: SZEDER Gábor <szeder.dev@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2019-06-27 16:42:48 +03:00
if (!res) {
if (!opts->verbose)
term_clear_line();
fprintf(stderr,
_("Stopped at %s... %.*s\n"),
short_commit_name(commit),
item->arg_len, arg);
rebase: fix garbled progress display with '-x' When running a command with the 'exec' instruction during an interactive rebase session, or for a range of commits using 'git rebase -x', the output can be a bit garbled when the name of the command is short enough: $ git rebase -x true HEAD~5 Executing: true Executing: true Executing: true Executing: true Executing: true) Successfully rebased and updated refs/heads/master. Note the ')' at the end of the last line. It gets more garbled as the range of commits increases: $ git rebase -x true HEAD~50 Executing: true) [ repeated 3 more times ] Executing: true0) [ repeated 44 more times ] Executing: true00) Successfully rebased and updated refs/heads/master. Those extra numbers and ')' are remnants of the previously displayed "Rebasing (N/M)" progress lines that are usually completely overwritten by the "Executing: <cmd>" lines, unless 'cmd' is short and the "N/M" part is long. Make sure that the previously displayed "Rebasing (N/M)" line is cleared by using the term_clear_line() helper function added in the previous patch. Do so only when not being '--verbose', because in that case these "Rebasing (N/M)" lines are not printed as progress (i.e. as lines with '\r' at the end), but as "regular" output (with '\n' at the end). A couple of other rebase commands print similar messages, e.g. "Stopped at <abbrev-oid>... <subject>" for the 'edit' or 'break' commands, or the "Successfully rebased and updated <full-ref>." at the very end. These are so long that they practically always overwrite that "Rebasing (N/M)" progress line, but let's be prudent, and clear the last line before printing these, too. In 't3420-rebase-autostash.sh' two helper functions prepare the expected output of four tests that check the full output of 'git rebase' and thus are affected by this change, so adjust their expectations to account for the new line clearing. Note that this patch doesn't completely eliminate the possibility of similar garbled outputs, e.g. some error messages from rebase or the "Auto-merging <file>" message from within the depths of the merge machinery might not be long enough to completely cover the last "Rebasing (N/M)" line. This patch doesn't do anything about them, because dealing with them individually would result in way too much churn, while having a catch-all term_clear_line() call in the common code path of pick_commits() would hide the "Rebasing (N/M)" line way too soon, and it would either flicker or be invisible. Signed-off-by: SZEDER Gábor <szeder.dev@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2019-06-27 16:42:48 +03:00
}
return error_with_patch(r, commit,
arg, item->arg_len, opts, res, !res);
}
if (is_rebase_i(opts) && !res)
record_in_rewritten(&item->commit->object.oid,
peek_command(todo_list, 1));
sequencer (rebase -i): add support for the 'fixup' and 'squash' commands This is a huge patch, and at the same time a huge step forward to execute the performance-critical parts of the interactive rebase in a builtin command. Since 'fixup' and 'squash' are not only similar, but also need to know about each other (we want to reduce a series of fixups/squashes into a single, final commit message edit, from the user's point of view), we really have to implement them both at the same time. Most of the actual work is done by the existing code path that already handles the "pick" and the "edit" commands; We added support for other features (e.g. to amend the commit message) in the patches leading up to this one, yet there are still quite a few bits in this patch that simply would not make sense as individual patches (such as: determining whether there was anything to "fix up" in the "todo" script, etc). In theory, it would be possible to reuse the fast-forward code path also for the fixup and the squash code paths, but in practice this would make the code less readable. The end result cannot be fast-forwarded anyway, therefore let's just extend the cherry-picking code path for now. Since the sequencer parses the entire `git-rebase-todo` script in one go, fixup or squash commands without a preceding pick can be reported early (in git-rebase--interactive, we could only report such errors just before executing the fixup/squash). Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-02 18:27:07 +03:00
if (res && is_fixup(item->command)) {
if (res == 1)
intend_to_amend();
return error_failed_squash(r, item->commit, opts,
item->arg_len, arg);
sequencer: do not squash 'reword' commits when we hit conflicts Ever since commit 18633e1a22 ("rebase -i: use the rebase--helper builtin", 2017-02-09), when a commit marked as 'reword' in an interactive rebase has conflicts and fails to apply, when the rebase is resumed that commit will be squashed into its parent with its commit message taken. The issue can be understood better by looking at commit 56dc3ab04b ("sequencer (rebase -i): implement the 'edit' command", 2017-01-02), which introduced error_with_patch() for the edit command. For the edit command, it needs to stop the rebase whether or not the patch applies cleanly. If the patch does apply cleanly, then when it resumes it knows it needs to amend all changes into the previous commit. If it does not apply cleanly, then the changes should not be amended. Thus, it passes !res (success of applying the 'edit' commit) to error_with_patch() for the to_amend flag. The problematic line of code actually came from commit 04efc8b57c ("sequencer (rebase -i): implement the 'reword' command", 2017-01-02). Note that to get to this point in the code: * !!res (i.e. patch application failed) * item->command < TODO_SQUASH * item->command != TODO_EDIT * !is_fixup(item->command) [i.e. not squash or fixup] So that means this can only be a failed patch application that is either a pick, revert, or reword. We only need to amend HEAD when rewording the root commit or a commit that has been fast-forwarded, for any of the other cases we want a new commit, so we should not set the to_amend flag. Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Original-patch-by: Elijah Newren <newren@gmail.com> Signed-off-by: Phillip Wood <phillip.wood@dunelm.org.uk> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-06-19 15:46:51 +03:00
} else if (res && is_rebase_i(opts) && item->commit) {
int to_amend = 0;
struct object_id oid;
/*
* If we are rewording and have either
* fast-forwarded already, or are about to
* create a new root commit, we want to amend,
* otherwise we do not.
*/
if (item->command == TODO_REWORD &&
!get_oid("HEAD", &oid) &&
(oideq(&item->commit->object.oid, &oid) ||
sequencer: do not squash 'reword' commits when we hit conflicts Ever since commit 18633e1a22 ("rebase -i: use the rebase--helper builtin", 2017-02-09), when a commit marked as 'reword' in an interactive rebase has conflicts and fails to apply, when the rebase is resumed that commit will be squashed into its parent with its commit message taken. The issue can be understood better by looking at commit 56dc3ab04b ("sequencer (rebase -i): implement the 'edit' command", 2017-01-02), which introduced error_with_patch() for the edit command. For the edit command, it needs to stop the rebase whether or not the patch applies cleanly. If the patch does apply cleanly, then when it resumes it knows it needs to amend all changes into the previous commit. If it does not apply cleanly, then the changes should not be amended. Thus, it passes !res (success of applying the 'edit' commit) to error_with_patch() for the to_amend flag. The problematic line of code actually came from commit 04efc8b57c ("sequencer (rebase -i): implement the 'reword' command", 2017-01-02). Note that to get to this point in the code: * !!res (i.e. patch application failed) * item->command < TODO_SQUASH * item->command != TODO_EDIT * !is_fixup(item->command) [i.e. not squash or fixup] So that means this can only be a failed patch application that is either a pick, revert, or reword. We only need to amend HEAD when rewording the root commit or a commit that has been fast-forwarded, for any of the other cases we want a new commit, so we should not set the to_amend flag. Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Original-patch-by: Elijah Newren <newren@gmail.com> Signed-off-by: Phillip Wood <phillip.wood@dunelm.org.uk> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-06-19 15:46:51 +03:00
(opts->have_squash_onto &&
oideq(&opts->squash_onto, &oid))))
sequencer: do not squash 'reword' commits when we hit conflicts Ever since commit 18633e1a22 ("rebase -i: use the rebase--helper builtin", 2017-02-09), when a commit marked as 'reword' in an interactive rebase has conflicts and fails to apply, when the rebase is resumed that commit will be squashed into its parent with its commit message taken. The issue can be understood better by looking at commit 56dc3ab04b ("sequencer (rebase -i): implement the 'edit' command", 2017-01-02), which introduced error_with_patch() for the edit command. For the edit command, it needs to stop the rebase whether or not the patch applies cleanly. If the patch does apply cleanly, then when it resumes it knows it needs to amend all changes into the previous commit. If it does not apply cleanly, then the changes should not be amended. Thus, it passes !res (success of applying the 'edit' commit) to error_with_patch() for the to_amend flag. The problematic line of code actually came from commit 04efc8b57c ("sequencer (rebase -i): implement the 'reword' command", 2017-01-02). Note that to get to this point in the code: * !!res (i.e. patch application failed) * item->command < TODO_SQUASH * item->command != TODO_EDIT * !is_fixup(item->command) [i.e. not squash or fixup] So that means this can only be a failed patch application that is either a pick, revert, or reword. We only need to amend HEAD when rewording the root commit or a commit that has been fast-forwarded, for any of the other cases we want a new commit, so we should not set the to_amend flag. Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Original-patch-by: Elijah Newren <newren@gmail.com> Signed-off-by: Phillip Wood <phillip.wood@dunelm.org.uk> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-06-19 15:46:51 +03:00
to_amend = 1;
return res | error_with_patch(r, item->commit,
arg, item->arg_len, opts,
sequencer: do not squash 'reword' commits when we hit conflicts Ever since commit 18633e1a22 ("rebase -i: use the rebase--helper builtin", 2017-02-09), when a commit marked as 'reword' in an interactive rebase has conflicts and fails to apply, when the rebase is resumed that commit will be squashed into its parent with its commit message taken. The issue can be understood better by looking at commit 56dc3ab04b ("sequencer (rebase -i): implement the 'edit' command", 2017-01-02), which introduced error_with_patch() for the edit command. For the edit command, it needs to stop the rebase whether or not the patch applies cleanly. If the patch does apply cleanly, then when it resumes it knows it needs to amend all changes into the previous commit. If it does not apply cleanly, then the changes should not be amended. Thus, it passes !res (success of applying the 'edit' commit) to error_with_patch() for the to_amend flag. The problematic line of code actually came from commit 04efc8b57c ("sequencer (rebase -i): implement the 'reword' command", 2017-01-02). Note that to get to this point in the code: * !!res (i.e. patch application failed) * item->command < TODO_SQUASH * item->command != TODO_EDIT * !is_fixup(item->command) [i.e. not squash or fixup] So that means this can only be a failed patch application that is either a pick, revert, or reword. We only need to amend HEAD when rewording the root commit or a commit that has been fast-forwarded, for any of the other cases we want a new commit, so we should not set the to_amend flag. Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de> Original-patch-by: Elijah Newren <newren@gmail.com> Signed-off-by: Phillip Wood <phillip.wood@dunelm.org.uk> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-06-19 15:46:51 +03:00
res, to_amend);
}
} else if (item->command == TODO_EXEC) {
char *end_of_arg = (char *)(arg + item->arg_len);
int saved = *end_of_arg;
rebase: fix garbled progress display with '-x' When running a command with the 'exec' instruction during an interactive rebase session, or for a range of commits using 'git rebase -x', the output can be a bit garbled when the name of the command is short enough: $ git rebase -x true HEAD~5 Executing: true Executing: true Executing: true Executing: true Executing: true) Successfully rebased and updated refs/heads/master. Note the ')' at the end of the last line. It gets more garbled as the range of commits increases: $ git rebase -x true HEAD~50 Executing: true) [ repeated 3 more times ] Executing: true0) [ repeated 44 more times ] Executing: true00) Successfully rebased and updated refs/heads/master. Those extra numbers and ')' are remnants of the previously displayed "Rebasing (N/M)" progress lines that are usually completely overwritten by the "Executing: <cmd>" lines, unless 'cmd' is short and the "N/M" part is long. Make sure that the previously displayed "Rebasing (N/M)" line is cleared by using the term_clear_line() helper function added in the previous patch. Do so only when not being '--verbose', because in that case these "Rebasing (N/M)" lines are not printed as progress (i.e. as lines with '\r' at the end), but as "regular" output (with '\n' at the end). A couple of other rebase commands print similar messages, e.g. "Stopped at <abbrev-oid>... <subject>" for the 'edit' or 'break' commands, or the "Successfully rebased and updated <full-ref>." at the very end. These are so long that they practically always overwrite that "Rebasing (N/M)" progress line, but let's be prudent, and clear the last line before printing these, too. In 't3420-rebase-autostash.sh' two helper functions prepare the expected output of four tests that check the full output of 'git rebase' and thus are affected by this change, so adjust their expectations to account for the new line clearing. Note that this patch doesn't completely eliminate the possibility of similar garbled outputs, e.g. some error messages from rebase or the "Auto-merging <file>" message from within the depths of the merge machinery might not be long enough to completely cover the last "Rebasing (N/M)" line. This patch doesn't do anything about them, because dealing with them individually would result in way too much churn, while having a catch-all term_clear_line() call in the common code path of pick_commits() would hide the "Rebasing (N/M)" line way too soon, and it would either flicker or be invisible. Signed-off-by: SZEDER Gábor <szeder.dev@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2019-06-27 16:42:48 +03:00
if (!opts->verbose)
term_clear_line();
*end_of_arg = '\0';
res = do_exec(r, arg);
*end_of_arg = saved;
if (res) {
if (opts->reschedule_failed_exec)
reschedule = 1;
}
check_todo = 1;
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
} else if (item->command == TODO_LABEL) {
if ((res = do_label(r, arg, item->arg_len)))
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
reschedule = 1;
} else if (item->command == TODO_RESET) {
if ((res = do_reset(r, arg, item->arg_len, opts)))
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
reschedule = 1;
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
} else if (item->command == TODO_MERGE) {
if ((res = do_merge(r, item->commit,
arg, item->arg_len,
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
item->flags, opts)) < 0)
reschedule = 1;
else if (item->commit)
record_in_rewritten(&item->commit->object.oid,
peek_command(todo_list, 1));
if (res > 0)
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
/* failed with merge conflicts */
return error_with_patch(r, item->commit,
arg, item->arg_len,
opts, res, 0);
} else if (!is_noop(item->command))
return error(_("unknown command %d"), item->command);
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
if (reschedule) {
advise(_(rescheduled_advice),
get_item_line_length(todo_list,
todo_list->current),
get_item_line(todo_list, todo_list->current));
todo_list->current--;
if (save_todo(todo_list, opts))
return -1;
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
if (item->commit)
return error_with_patch(r,
item->commit,
arg, item->arg_len,
opts, res, 0);
} else if (check_todo && !res) {
struct stat st;
if (stat(get_todo_path(opts), &st)) {
res = error_errno(_("could not stat '%s'"),
get_todo_path(opts));
} else if (match_stat_data(&todo_list->stat, &st)) {
/* Reread the todo file if it has changed. */
todo_list_release(todo_list);
if (read_populate_todo(r, todo_list, opts))
res = -1; /* message was printed */
/* `current` will be incremented below */
todo_list->current = -1;
}
sequencer: introduce new commands to reset the revision In the upcoming commits, we will teach the sequencer to rebase merges. This will be done in a very different way from the unfortunate design of `git rebase --preserve-merges` (which does not allow for reordering commits, or changing the branch topology). The main idea is to introduce new todo list commands, to support labeling the current revision with a given name, resetting the current revision to a previous state, and merging labeled revisions. This idea was developed in Git for Windows' Git garden shears (that are used to maintain Git for Windows' "thicket of branches" on top of upstream Git), and this patch is part of the effort to make it available to a wider audience, as well as to make the entire process more robust (by implementing it in a safe and portable language rather than a Unix shell script). This commit implements the commands to label, and to reset to, given revisions. The syntax is: label <name> reset <name> Internally, the `label <name>` command creates the ref `refs/rewritten/<name>`. This makes it possible to work with the labeled revisions interactively, or in a scripted fashion (e.g. via the todo list command `exec`). These temporary refs are removed upon sequencer_remove_state(), so that even a `git rebase --abort` cleans them up. We disallow '#' as label because that character will be used as separator in the upcoming `merge` command. Later in this patch series, we will mark the `refs/rewritten/` refs as worktree-local, to allow for interactive rebases to be run in parallel in worktrees linked to the same repository. As typos happen, a failed `label` or `reset` command will be rescheduled immediately. As the previous code to reschedule a command is embedded deeply in the pick/fixup/squash code path, we simply duplicate the few lines. This will allow us to extend the new code path easily for the upcoming `merge` command. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:47 +03:00
}
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
todo_list->current++;
if (res)
return res;
}
if (is_rebase_i(opts)) {
struct strbuf head_ref = STRBUF_INIT, buf = STRBUF_INIT;
struct stat st;
/* Stopped in the middle, as planned? */
if (todo_list->current < todo_list->nr)
return 0;
if (read_oneliner(&head_ref, rebase_path_head_name(), 0) &&
starts_with(head_ref.buf, "refs/")) {
const char *msg;
struct object_id head, orig;
int res;
if (get_oid("HEAD", &head)) {
res = error(_("cannot read HEAD"));
cleanup_head_ref:
strbuf_release(&head_ref);
strbuf_release(&buf);
return res;
}
if (!read_oneliner(&buf, rebase_path_orig_head(), 0) ||
get_oid_hex(buf.buf, &orig)) {
res = error(_("could not read orig-head"));
goto cleanup_head_ref;
}
strbuf_reset(&buf);
if (!read_oneliner(&buf, rebase_path_onto(), 0)) {
res = error(_("could not read 'onto'"));
goto cleanup_head_ref;
}
msg = reflog_message(opts, "finish", "%s onto %s",
head_ref.buf, buf.buf);
if (update_ref(msg, head_ref.buf, &head, &orig,
REF_NO_DEREF, UPDATE_REFS_MSG_ON_ERR)) {
res = error(_("could not update %s"),
head_ref.buf);
goto cleanup_head_ref;
}
msg = reflog_message(opts, "finish", "returning to %s",
head_ref.buf);
if (create_symref("HEAD", head_ref.buf, msg)) {
res = error(_("could not update HEAD to %s"),
head_ref.buf);
goto cleanup_head_ref;
}
strbuf_reset(&buf);
}
if (opts->verbose) {
struct rev_info log_tree_opt;
struct object_id orig, head;
memset(&log_tree_opt, 0, sizeof(log_tree_opt));
repo_init_revisions(r, &log_tree_opt, NULL);
log_tree_opt.diff = 1;
log_tree_opt.diffopt.output_format =
DIFF_FORMAT_DIFFSTAT;
log_tree_opt.disable_stdin = 1;
if (read_oneliner(&buf, rebase_path_orig_head(), 0) &&
sha1_name: convert get_sha1* to get_oid* Now that all the callers of get_sha1 directly or indirectly use struct object_id, rename the functions starting with get_sha1 to start with get_oid. Convert the internals in sha1_name.c to use struct object_id as well, and eliminate explicit length checks where possible. Convert a use of 40 in get_oid_basic to GIT_SHA1_HEXSZ. Outside of sha1_name.c and cache.h, this transition was made with the following semantic patch: @@ expression E1, E2; @@ - get_sha1(E1, E2.hash) + get_oid(E1, &E2) @@ expression E1, E2; @@ - get_sha1(E1, E2->hash) + get_oid(E1, E2) @@ expression E1, E2; @@ - get_sha1_committish(E1, E2.hash) + get_oid_committish(E1, &E2) @@ expression E1, E2; @@ - get_sha1_committish(E1, E2->hash) + get_oid_committish(E1, E2) @@ expression E1, E2; @@ - get_sha1_treeish(E1, E2.hash) + get_oid_treeish(E1, &E2) @@ expression E1, E2; @@ - get_sha1_treeish(E1, E2->hash) + get_oid_treeish(E1, E2) @@ expression E1, E2; @@ - get_sha1_commit(E1, E2.hash) + get_oid_commit(E1, &E2) @@ expression E1, E2; @@ - get_sha1_commit(E1, E2->hash) + get_oid_commit(E1, E2) @@ expression E1, E2; @@ - get_sha1_tree(E1, E2.hash) + get_oid_tree(E1, &E2) @@ expression E1, E2; @@ - get_sha1_tree(E1, E2->hash) + get_oid_tree(E1, E2) @@ expression E1, E2; @@ - get_sha1_blob(E1, E2.hash) + get_oid_blob(E1, &E2) @@ expression E1, E2; @@ - get_sha1_blob(E1, E2->hash) + get_oid_blob(E1, E2) @@ expression E1, E2, E3, E4; @@ - get_sha1_with_context(E1, E2, E3.hash, E4) + get_oid_with_context(E1, E2, &E3, E4) @@ expression E1, E2, E3, E4; @@ - get_sha1_with_context(E1, E2, E3->hash, E4) + get_oid_with_context(E1, E2, E3, E4) Signed-off-by: brian m. carlson <sandals@crustytoothpaste.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-07-14 02:49:28 +03:00
!get_oid(buf.buf, &orig) &&
!get_oid("HEAD", &head)) {
diff_tree_oid(&orig, &head, "",
&log_tree_opt.diffopt);
log_tree_diff_flush(&log_tree_opt);
}
}
flush_rewritten_pending();
if (!stat(rebase_path_rewritten_list(), &st) &&
st.st_size > 0) {
struct child_process child = CHILD_PROCESS_INIT;
const char *post_rewrite_hook =
find_hook("post-rewrite");
child.in = open(rebase_path_rewritten_list(), O_RDONLY);
child.git_cmd = 1;
argv_array_push(&child.args, "notes");
argv_array_push(&child.args, "copy");
argv_array_push(&child.args, "--for-rewrite=rebase");
/* we don't care if this copying failed */
run_command(&child);
if (post_rewrite_hook) {
struct child_process hook = CHILD_PROCESS_INIT;
hook.in = open(rebase_path_rewritten_list(),
O_RDONLY);
hook.stdout_to_stderr = 1;
hook.trace2_hook_name = "post-rewrite";
argv_array_push(&hook.args, post_rewrite_hook);
argv_array_push(&hook.args, "rebase");
/* we don't care if this hook failed */
run_command(&hook);
}
}
apply_autostash(opts);
rebase: fix garbled progress display with '-x' When running a command with the 'exec' instruction during an interactive rebase session, or for a range of commits using 'git rebase -x', the output can be a bit garbled when the name of the command is short enough: $ git rebase -x true HEAD~5 Executing: true Executing: true Executing: true Executing: true Executing: true) Successfully rebased and updated refs/heads/master. Note the ')' at the end of the last line. It gets more garbled as the range of commits increases: $ git rebase -x true HEAD~50 Executing: true) [ repeated 3 more times ] Executing: true0) [ repeated 44 more times ] Executing: true00) Successfully rebased and updated refs/heads/master. Those extra numbers and ')' are remnants of the previously displayed "Rebasing (N/M)" progress lines that are usually completely overwritten by the "Executing: <cmd>" lines, unless 'cmd' is short and the "N/M" part is long. Make sure that the previously displayed "Rebasing (N/M)" line is cleared by using the term_clear_line() helper function added in the previous patch. Do so only when not being '--verbose', because in that case these "Rebasing (N/M)" lines are not printed as progress (i.e. as lines with '\r' at the end), but as "regular" output (with '\n' at the end). A couple of other rebase commands print similar messages, e.g. "Stopped at <abbrev-oid>... <subject>" for the 'edit' or 'break' commands, or the "Successfully rebased and updated <full-ref>." at the very end. These are so long that they practically always overwrite that "Rebasing (N/M)" progress line, but let's be prudent, and clear the last line before printing these, too. In 't3420-rebase-autostash.sh' two helper functions prepare the expected output of four tests that check the full output of 'git rebase' and thus are affected by this change, so adjust their expectations to account for the new line clearing. Note that this patch doesn't completely eliminate the possibility of similar garbled outputs, e.g. some error messages from rebase or the "Auto-merging <file>" message from within the depths of the merge machinery might not be long enough to completely cover the last "Rebasing (N/M)" line. This patch doesn't do anything about them, because dealing with them individually would result in way too much churn, while having a catch-all term_clear_line() call in the common code path of pick_commits() would hide the "Rebasing (N/M)" line way too soon, and it would either flicker or be invisible. Signed-off-by: SZEDER Gábor <szeder.dev@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2019-06-27 16:42:48 +03:00
if (!opts->quiet) {
if (!opts->verbose)
term_clear_line();
fprintf(stderr,
"Successfully rebased and updated %s.\n",
head_ref.buf);
rebase: fix garbled progress display with '-x' When running a command with the 'exec' instruction during an interactive rebase session, or for a range of commits using 'git rebase -x', the output can be a bit garbled when the name of the command is short enough: $ git rebase -x true HEAD~5 Executing: true Executing: true Executing: true Executing: true Executing: true) Successfully rebased and updated refs/heads/master. Note the ')' at the end of the last line. It gets more garbled as the range of commits increases: $ git rebase -x true HEAD~50 Executing: true) [ repeated 3 more times ] Executing: true0) [ repeated 44 more times ] Executing: true00) Successfully rebased and updated refs/heads/master. Those extra numbers and ')' are remnants of the previously displayed "Rebasing (N/M)" progress lines that are usually completely overwritten by the "Executing: <cmd>" lines, unless 'cmd' is short and the "N/M" part is long. Make sure that the previously displayed "Rebasing (N/M)" line is cleared by using the term_clear_line() helper function added in the previous patch. Do so only when not being '--verbose', because in that case these "Rebasing (N/M)" lines are not printed as progress (i.e. as lines with '\r' at the end), but as "regular" output (with '\n' at the end). A couple of other rebase commands print similar messages, e.g. "Stopped at <abbrev-oid>... <subject>" for the 'edit' or 'break' commands, or the "Successfully rebased and updated <full-ref>." at the very end. These are so long that they practically always overwrite that "Rebasing (N/M)" progress line, but let's be prudent, and clear the last line before printing these, too. In 't3420-rebase-autostash.sh' two helper functions prepare the expected output of four tests that check the full output of 'git rebase' and thus are affected by this change, so adjust their expectations to account for the new line clearing. Note that this patch doesn't completely eliminate the possibility of similar garbled outputs, e.g. some error messages from rebase or the "Auto-merging <file>" message from within the depths of the merge machinery might not be long enough to completely cover the last "Rebasing (N/M)" line. This patch doesn't do anything about them, because dealing with them individually would result in way too much churn, while having a catch-all term_clear_line() call in the common code path of pick_commits() would hide the "Rebasing (N/M)" line way too soon, and it would either flicker or be invisible. Signed-off-by: SZEDER Gábor <szeder.dev@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2019-06-27 16:42:48 +03:00
}
strbuf_release(&buf);
strbuf_release(&head_ref);
}
/*
* Sequence of picks finished successfully; cleanup by
* removing the .git/sequencer directory
*/
return sequencer_remove_state(opts);
}
static int continue_single_pick(struct repository *r)
{
const char *argv[] = { "commit", NULL };
if (!file_exists(git_path_cherry_pick_head(r)) &&
!file_exists(git_path_revert_head(r)))
return error(_("no cherry-pick or revert in progress"));
return run_command_v_opt(argv, RUN_GIT_CMD);
}
static int commit_staged_changes(struct repository *r,
struct replay_opts *opts,
rebase --skip: clean up commit message after a failed fixup/squash During a series of fixup/squash commands, the interactive rebase builds up a commit message with comments. This will be presented to the user in the editor if at least one of those commands was a `squash`. In any case, the commit message will be cleaned up eventually, removing all those intermediate comments, in the final step of such a fixup/squash chain. However, if the last fixup/squash command in such a chain fails with merge conflicts, and if the user then decides to skip it (or resolve it to a clean worktree and then continue the rebase), the current code fails to clean up the commit message. This commit fixes that behavior. The fix is quite a bit more involved than meets the eye because it is not only about the question whether we are `git rebase --skip`ing a fixup or squash. It is also about removing the skipped fixup/squash's commit message from the accumulated commit message. And it is also about the question whether we should let the user edit the final commit message or not ("Was there a squash in the chain *that was not skipped*?"). For example, in this case we will want to fix the commit message, but not open it in an editor: pick <- succeeds fixup <- succeeds squash <- fails, will be skipped This is where the newly-introduced `current-fixups` file comes in real handy. A quick look and we can determine whether there was a non-skipped squash. We only need to make sure to keep it up to date with respect to skipped fixup/squash commands. As a bonus, we can even avoid committing unnecessarily, e.g. when there was only one fixup, and it failed, and was skipped. To fix only the bug where the final commit message was not cleaned up properly, but without fixing the rest, would have been more complicated than fixing it all in one go, hence this commit lumps together more than a single concern. For the same reason, this commit also adds a bit more to the existing test case for the regression we just fixed. The diff is best viewed with --color-moved. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-27 23:48:30 +03:00
struct todo_list *todo_list)
{
unsigned int flags = ALLOW_EMPTY | EDIT_MSG;
rebase --skip: clean up commit message after a failed fixup/squash During a series of fixup/squash commands, the interactive rebase builds up a commit message with comments. This will be presented to the user in the editor if at least one of those commands was a `squash`. In any case, the commit message will be cleaned up eventually, removing all those intermediate comments, in the final step of such a fixup/squash chain. However, if the last fixup/squash command in such a chain fails with merge conflicts, and if the user then decides to skip it (or resolve it to a clean worktree and then continue the rebase), the current code fails to clean up the commit message. This commit fixes that behavior. The fix is quite a bit more involved than meets the eye because it is not only about the question whether we are `git rebase --skip`ing a fixup or squash. It is also about removing the skipped fixup/squash's commit message from the accumulated commit message. And it is also about the question whether we should let the user edit the final commit message or not ("Was there a squash in the chain *that was not skipped*?"). For example, in this case we will want to fix the commit message, but not open it in an editor: pick <- succeeds fixup <- succeeds squash <- fails, will be skipped This is where the newly-introduced `current-fixups` file comes in real handy. A quick look and we can determine whether there was a non-skipped squash. We only need to make sure to keep it up to date with respect to skipped fixup/squash commands. As a bonus, we can even avoid committing unnecessarily, e.g. when there was only one fixup, and it failed, and was skipped. To fix only the bug where the final commit message was not cleaned up properly, but without fixing the rest, would have been more complicated than fixing it all in one go, hence this commit lumps together more than a single concern. For the same reason, this commit also adds a bit more to the existing test case for the regression we just fixed. The diff is best viewed with --color-moved. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-27 23:48:30 +03:00
unsigned int final_fixup = 0, is_clean;
if (has_unstaged_changes(r, 1))
return error(_("cannot rebase: You have unstaged changes."));
is_clean = !has_uncommitted_changes(r, 0);
if (file_exists(rebase_path_amend())) {
struct strbuf rev = STRBUF_INIT;
struct object_id head, to_amend;
if (get_oid("HEAD", &head))
return error(_("cannot amend non-existing commit"));
if (!read_oneliner(&rev, rebase_path_amend(), 0))
return error(_("invalid file: '%s'"), rebase_path_amend());
if (get_oid_hex(rev.buf, &to_amend))
return error(_("invalid contents: '%s'"),
rebase_path_amend());
if (!is_clean && !oideq(&head, &to_amend))
return error(_("\nYou have uncommitted changes in your "
"working tree. Please, commit them\n"
"first and then run 'git rebase "
"--continue' again."));
rebase --skip: clean up commit message after a failed fixup/squash During a series of fixup/squash commands, the interactive rebase builds up a commit message with comments. This will be presented to the user in the editor if at least one of those commands was a `squash`. In any case, the commit message will be cleaned up eventually, removing all those intermediate comments, in the final step of such a fixup/squash chain. However, if the last fixup/squash command in such a chain fails with merge conflicts, and if the user then decides to skip it (or resolve it to a clean worktree and then continue the rebase), the current code fails to clean up the commit message. This commit fixes that behavior. The fix is quite a bit more involved than meets the eye because it is not only about the question whether we are `git rebase --skip`ing a fixup or squash. It is also about removing the skipped fixup/squash's commit message from the accumulated commit message. And it is also about the question whether we should let the user edit the final commit message or not ("Was there a squash in the chain *that was not skipped*?"). For example, in this case we will want to fix the commit message, but not open it in an editor: pick <- succeeds fixup <- succeeds squash <- fails, will be skipped This is where the newly-introduced `current-fixups` file comes in real handy. A quick look and we can determine whether there was a non-skipped squash. We only need to make sure to keep it up to date with respect to skipped fixup/squash commands. As a bonus, we can even avoid committing unnecessarily, e.g. when there was only one fixup, and it failed, and was skipped. To fix only the bug where the final commit message was not cleaned up properly, but without fixing the rest, would have been more complicated than fixing it all in one go, hence this commit lumps together more than a single concern. For the same reason, this commit also adds a bit more to the existing test case for the regression we just fixed. The diff is best viewed with --color-moved. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-27 23:48:30 +03:00
/*
* When skipping a failed fixup/squash, we need to edit the
* commit message, the current fixup list and count, and if it
* was the last fixup/squash in the chain, we need to clean up
* the commit message and if there was a squash, let the user
* edit it.
*/
if (!is_clean || !opts->current_fixup_count)
; /* this is not the final fixup */
else if (!oideq(&head, &to_amend) ||
!file_exists(rebase_path_stopped_sha())) {
/* was a final fixup or squash done manually? */
if (!is_fixup(peek_command(todo_list, 0))) {
unlink(rebase_path_fixup_msg());
unlink(rebase_path_squash_msg());
unlink(rebase_path_current_fixups());
strbuf_reset(&opts->current_fixups);
opts->current_fixup_count = 0;
}
} else {
/* we are in a fixup/squash chain */
rebase --skip: clean up commit message after a failed fixup/squash During a series of fixup/squash commands, the interactive rebase builds up a commit message with comments. This will be presented to the user in the editor if at least one of those commands was a `squash`. In any case, the commit message will be cleaned up eventually, removing all those intermediate comments, in the final step of such a fixup/squash chain. However, if the last fixup/squash command in such a chain fails with merge conflicts, and if the user then decides to skip it (or resolve it to a clean worktree and then continue the rebase), the current code fails to clean up the commit message. This commit fixes that behavior. The fix is quite a bit more involved than meets the eye because it is not only about the question whether we are `git rebase --skip`ing a fixup or squash. It is also about removing the skipped fixup/squash's commit message from the accumulated commit message. And it is also about the question whether we should let the user edit the final commit message or not ("Was there a squash in the chain *that was not skipped*?"). For example, in this case we will want to fix the commit message, but not open it in an editor: pick <- succeeds fixup <- succeeds squash <- fails, will be skipped This is where the newly-introduced `current-fixups` file comes in real handy. A quick look and we can determine whether there was a non-skipped squash. We only need to make sure to keep it up to date with respect to skipped fixup/squash commands. As a bonus, we can even avoid committing unnecessarily, e.g. when there was only one fixup, and it failed, and was skipped. To fix only the bug where the final commit message was not cleaned up properly, but without fixing the rest, would have been more complicated than fixing it all in one go, hence this commit lumps together more than a single concern. For the same reason, this commit also adds a bit more to the existing test case for the regression we just fixed. The diff is best viewed with --color-moved. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-27 23:48:30 +03:00
const char *p = opts->current_fixups.buf;
int len = opts->current_fixups.len;
opts->current_fixup_count--;
if (!len)
BUG("Incorrect current_fixups:\n%s", p);
while (len && p[len - 1] != '\n')
len--;
strbuf_setlen(&opts->current_fixups, len);
if (write_message(p, len, rebase_path_current_fixups(),
0) < 0)
return error(_("could not write file: '%s'"),
rebase_path_current_fixups());
/*
* If a fixup/squash in a fixup/squash chain failed, the
* commit message is already correct, no need to commit
* it again.
*
* Only if it is the final command in the fixup/squash
* chain, and only if the chain is longer than a single
* fixup/squash command (which was just skipped), do we
* actually need to re-commit with a cleaned up commit
* message.
*/
if (opts->current_fixup_count > 0 &&
!is_fixup(peek_command(todo_list, 0))) {
final_fixup = 1;
/*
* If there was not a single "squash" in the
* chain, we only need to clean up the commit
* message, no need to bother the user with
* opening the commit message in the editor.
*/
if (!starts_with(p, "squash ") &&
!strstr(p, "\nsquash "))
flags = (flags & ~EDIT_MSG) | CLEANUP_MSG;
} else if (is_fixup(peek_command(todo_list, 0))) {
/*
* We need to update the squash message to skip
* the latest commit message.
*/
struct commit *commit;
const char *path = rebase_path_squash_msg();
if (parse_head(r, &commit) ||
rebase --skip: clean up commit message after a failed fixup/squash During a series of fixup/squash commands, the interactive rebase builds up a commit message with comments. This will be presented to the user in the editor if at least one of those commands was a `squash`. In any case, the commit message will be cleaned up eventually, removing all those intermediate comments, in the final step of such a fixup/squash chain. However, if the last fixup/squash command in such a chain fails with merge conflicts, and if the user then decides to skip it (or resolve it to a clean worktree and then continue the rebase), the current code fails to clean up the commit message. This commit fixes that behavior. The fix is quite a bit more involved than meets the eye because it is not only about the question whether we are `git rebase --skip`ing a fixup or squash. It is also about removing the skipped fixup/squash's commit message from the accumulated commit message. And it is also about the question whether we should let the user edit the final commit message or not ("Was there a squash in the chain *that was not skipped*?"). For example, in this case we will want to fix the commit message, but not open it in an editor: pick <- succeeds fixup <- succeeds squash <- fails, will be skipped This is where the newly-introduced `current-fixups` file comes in real handy. A quick look and we can determine whether there was a non-skipped squash. We only need to make sure to keep it up to date with respect to skipped fixup/squash commands. As a bonus, we can even avoid committing unnecessarily, e.g. when there was only one fixup, and it failed, and was skipped. To fix only the bug where the final commit message was not cleaned up properly, but without fixing the rest, would have been more complicated than fixing it all in one go, hence this commit lumps together more than a single concern. For the same reason, this commit also adds a bit more to the existing test case for the regression we just fixed. The diff is best viewed with --color-moved. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-27 23:48:30 +03:00
!(p = get_commit_buffer(commit, NULL)) ||
write_message(p, strlen(p), path, 0)) {
unuse_commit_buffer(commit, p);
return error(_("could not write file: "
"'%s'"), path);
}
unuse_commit_buffer(commit, p);
}
}
strbuf_release(&rev);
flags |= AMEND_MSG;
}
rebase --skip: clean up commit message after a failed fixup/squash During a series of fixup/squash commands, the interactive rebase builds up a commit message with comments. This will be presented to the user in the editor if at least one of those commands was a `squash`. In any case, the commit message will be cleaned up eventually, removing all those intermediate comments, in the final step of such a fixup/squash chain. However, if the last fixup/squash command in such a chain fails with merge conflicts, and if the user then decides to skip it (or resolve it to a clean worktree and then continue the rebase), the current code fails to clean up the commit message. This commit fixes that behavior. The fix is quite a bit more involved than meets the eye because it is not only about the question whether we are `git rebase --skip`ing a fixup or squash. It is also about removing the skipped fixup/squash's commit message from the accumulated commit message. And it is also about the question whether we should let the user edit the final commit message or not ("Was there a squash in the chain *that was not skipped*?"). For example, in this case we will want to fix the commit message, but not open it in an editor: pick <- succeeds fixup <- succeeds squash <- fails, will be skipped This is where the newly-introduced `current-fixups` file comes in real handy. A quick look and we can determine whether there was a non-skipped squash. We only need to make sure to keep it up to date with respect to skipped fixup/squash commands. As a bonus, we can even avoid committing unnecessarily, e.g. when there was only one fixup, and it failed, and was skipped. To fix only the bug where the final commit message was not cleaned up properly, but without fixing the rest, would have been more complicated than fixing it all in one go, hence this commit lumps together more than a single concern. For the same reason, this commit also adds a bit more to the existing test case for the regression we just fixed. The diff is best viewed with --color-moved. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-27 23:48:30 +03:00
if (is_clean) {
const char *cherry_pick_head = git_path_cherry_pick_head(r);
rebase --skip: clean up commit message after a failed fixup/squash During a series of fixup/squash commands, the interactive rebase builds up a commit message with comments. This will be presented to the user in the editor if at least one of those commands was a `squash`. In any case, the commit message will be cleaned up eventually, removing all those intermediate comments, in the final step of such a fixup/squash chain. However, if the last fixup/squash command in such a chain fails with merge conflicts, and if the user then decides to skip it (or resolve it to a clean worktree and then continue the rebase), the current code fails to clean up the commit message. This commit fixes that behavior. The fix is quite a bit more involved than meets the eye because it is not only about the question whether we are `git rebase --skip`ing a fixup or squash. It is also about removing the skipped fixup/squash's commit message from the accumulated commit message. And it is also about the question whether we should let the user edit the final commit message or not ("Was there a squash in the chain *that was not skipped*?"). For example, in this case we will want to fix the commit message, but not open it in an editor: pick <- succeeds fixup <- succeeds squash <- fails, will be skipped This is where the newly-introduced `current-fixups` file comes in real handy. A quick look and we can determine whether there was a non-skipped squash. We only need to make sure to keep it up to date with respect to skipped fixup/squash commands. As a bonus, we can even avoid committing unnecessarily, e.g. when there was only one fixup, and it failed, and was skipped. To fix only the bug where the final commit message was not cleaned up properly, but without fixing the rest, would have been more complicated than fixing it all in one go, hence this commit lumps together more than a single concern. For the same reason, this commit also adds a bit more to the existing test case for the regression we just fixed. The diff is best viewed with --color-moved. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-27 23:48:30 +03:00
if (file_exists(cherry_pick_head) && unlink(cherry_pick_head))
return error(_("could not remove CHERRY_PICK_HEAD"));
if (!final_fixup)
return 0;
}
if (run_git_commit(r, final_fixup ? NULL : rebase_path_message(),
rebase --skip: clean up commit message after a failed fixup/squash During a series of fixup/squash commands, the interactive rebase builds up a commit message with comments. This will be presented to the user in the editor if at least one of those commands was a `squash`. In any case, the commit message will be cleaned up eventually, removing all those intermediate comments, in the final step of such a fixup/squash chain. However, if the last fixup/squash command in such a chain fails with merge conflicts, and if the user then decides to skip it (or resolve it to a clean worktree and then continue the rebase), the current code fails to clean up the commit message. This commit fixes that behavior. The fix is quite a bit more involved than meets the eye because it is not only about the question whether we are `git rebase --skip`ing a fixup or squash. It is also about removing the skipped fixup/squash's commit message from the accumulated commit message. And it is also about the question whether we should let the user edit the final commit message or not ("Was there a squash in the chain *that was not skipped*?"). For example, in this case we will want to fix the commit message, but not open it in an editor: pick <- succeeds fixup <- succeeds squash <- fails, will be skipped This is where the newly-introduced `current-fixups` file comes in real handy. A quick look and we can determine whether there was a non-skipped squash. We only need to make sure to keep it up to date with respect to skipped fixup/squash commands. As a bonus, we can even avoid committing unnecessarily, e.g. when there was only one fixup, and it failed, and was skipped. To fix only the bug where the final commit message was not cleaned up properly, but without fixing the rest, would have been more complicated than fixing it all in one go, hence this commit lumps together more than a single concern. For the same reason, this commit also adds a bit more to the existing test case for the regression we just fixed. The diff is best viewed with --color-moved. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-27 23:48:30 +03:00
opts, flags))
return error(_("could not commit staged changes."));
unlink(rebase_path_amend());
unlink(git_path_merge_head(r));
rebase --skip: clean up commit message after a failed fixup/squash During a series of fixup/squash commands, the interactive rebase builds up a commit message with comments. This will be presented to the user in the editor if at least one of those commands was a `squash`. In any case, the commit message will be cleaned up eventually, removing all those intermediate comments, in the final step of such a fixup/squash chain. However, if the last fixup/squash command in such a chain fails with merge conflicts, and if the user then decides to skip it (or resolve it to a clean worktree and then continue the rebase), the current code fails to clean up the commit message. This commit fixes that behavior. The fix is quite a bit more involved than meets the eye because it is not only about the question whether we are `git rebase --skip`ing a fixup or squash. It is also about removing the skipped fixup/squash's commit message from the accumulated commit message. And it is also about the question whether we should let the user edit the final commit message or not ("Was there a squash in the chain *that was not skipped*?"). For example, in this case we will want to fix the commit message, but not open it in an editor: pick <- succeeds fixup <- succeeds squash <- fails, will be skipped This is where the newly-introduced `current-fixups` file comes in real handy. A quick look and we can determine whether there was a non-skipped squash. We only need to make sure to keep it up to date with respect to skipped fixup/squash commands. As a bonus, we can even avoid committing unnecessarily, e.g. when there was only one fixup, and it failed, and was skipped. To fix only the bug where the final commit message was not cleaned up properly, but without fixing the rest, would have been more complicated than fixing it all in one go, hence this commit lumps together more than a single concern. For the same reason, this commit also adds a bit more to the existing test case for the regression we just fixed. The diff is best viewed with --color-moved. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-27 23:48:30 +03:00
if (final_fixup) {
unlink(rebase_path_fixup_msg());
unlink(rebase_path_squash_msg());
}
if (opts->current_fixup_count > 0) {
/*
* Whether final fixup or not, we just cleaned up the commit
* message...
*/
unlink(rebase_path_current_fixups());
strbuf_reset(&opts->current_fixups);
opts->current_fixup_count = 0;
}
return 0;
}
int sequencer_continue(struct repository *r, struct replay_opts *opts)
{
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
struct todo_list todo_list = TODO_LIST_INIT;
int res;
if (read_and_refresh_cache(r, opts))
return -1;
rebase --skip: clean up commit message after a failed fixup/squash During a series of fixup/squash commands, the interactive rebase builds up a commit message with comments. This will be presented to the user in the editor if at least one of those commands was a `squash`. In any case, the commit message will be cleaned up eventually, removing all those intermediate comments, in the final step of such a fixup/squash chain. However, if the last fixup/squash command in such a chain fails with merge conflicts, and if the user then decides to skip it (or resolve it to a clean worktree and then continue the rebase), the current code fails to clean up the commit message. This commit fixes that behavior. The fix is quite a bit more involved than meets the eye because it is not only about the question whether we are `git rebase --skip`ing a fixup or squash. It is also about removing the skipped fixup/squash's commit message from the accumulated commit message. And it is also about the question whether we should let the user edit the final commit message or not ("Was there a squash in the chain *that was not skipped*?"). For example, in this case we will want to fix the commit message, but not open it in an editor: pick <- succeeds fixup <- succeeds squash <- fails, will be skipped This is where the newly-introduced `current-fixups` file comes in real handy. A quick look and we can determine whether there was a non-skipped squash. We only need to make sure to keep it up to date with respect to skipped fixup/squash commands. As a bonus, we can even avoid committing unnecessarily, e.g. when there was only one fixup, and it failed, and was skipped. To fix only the bug where the final commit message was not cleaned up properly, but without fixing the rest, would have been more complicated than fixing it all in one go, hence this commit lumps together more than a single concern. For the same reason, this commit also adds a bit more to the existing test case for the regression we just fixed. The diff is best viewed with --color-moved. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-27 23:48:30 +03:00
if (read_populate_opts(opts))
return -1;
if (is_rebase_i(opts)) {
if ((res = read_populate_todo(r, &todo_list, opts)))
rebase --skip: clean up commit message after a failed fixup/squash During a series of fixup/squash commands, the interactive rebase builds up a commit message with comments. This will be presented to the user in the editor if at least one of those commands was a `squash`. In any case, the commit message will be cleaned up eventually, removing all those intermediate comments, in the final step of such a fixup/squash chain. However, if the last fixup/squash command in such a chain fails with merge conflicts, and if the user then decides to skip it (or resolve it to a clean worktree and then continue the rebase), the current code fails to clean up the commit message. This commit fixes that behavior. The fix is quite a bit more involved than meets the eye because it is not only about the question whether we are `git rebase --skip`ing a fixup or squash. It is also about removing the skipped fixup/squash's commit message from the accumulated commit message. And it is also about the question whether we should let the user edit the final commit message or not ("Was there a squash in the chain *that was not skipped*?"). For example, in this case we will want to fix the commit message, but not open it in an editor: pick <- succeeds fixup <- succeeds squash <- fails, will be skipped This is where the newly-introduced `current-fixups` file comes in real handy. A quick look and we can determine whether there was a non-skipped squash. We only need to make sure to keep it up to date with respect to skipped fixup/squash commands. As a bonus, we can even avoid committing unnecessarily, e.g. when there was only one fixup, and it failed, and was skipped. To fix only the bug where the final commit message was not cleaned up properly, but without fixing the rest, would have been more complicated than fixing it all in one go, hence this commit lumps together more than a single concern. For the same reason, this commit also adds a bit more to the existing test case for the regression we just fixed. The diff is best viewed with --color-moved. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-27 23:48:30 +03:00
goto release_todo_list;
if (commit_staged_changes(r, opts, &todo_list))
return -1;
} else if (!file_exists(get_todo_path(opts)))
return continue_single_pick(r);
else if ((res = read_populate_todo(r, &todo_list, opts)))
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
goto release_todo_list;
if (!is_rebase_i(opts)) {
/* Verify that the conflict has been resolved */
if (file_exists(git_path_cherry_pick_head(r)) ||
file_exists(git_path_revert_head(r))) {
res = continue_single_pick(r);
if (res)
goto release_todo_list;
}
if (index_differs_from(r, "HEAD", NULL, 0)) {
res = error_dirty_index(r, opts);
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
goto release_todo_list;
}
todo_list.current++;
} else if (file_exists(rebase_path_stopped_sha())) {
struct strbuf buf = STRBUF_INIT;
struct object_id oid;
if (read_oneliner(&buf, rebase_path_stopped_sha(), 1) &&
sha1_name: convert get_sha1* to get_oid* Now that all the callers of get_sha1 directly or indirectly use struct object_id, rename the functions starting with get_sha1 to start with get_oid. Convert the internals in sha1_name.c to use struct object_id as well, and eliminate explicit length checks where possible. Convert a use of 40 in get_oid_basic to GIT_SHA1_HEXSZ. Outside of sha1_name.c and cache.h, this transition was made with the following semantic patch: @@ expression E1, E2; @@ - get_sha1(E1, E2.hash) + get_oid(E1, &E2) @@ expression E1, E2; @@ - get_sha1(E1, E2->hash) + get_oid(E1, E2) @@ expression E1, E2; @@ - get_sha1_committish(E1, E2.hash) + get_oid_committish(E1, &E2) @@ expression E1, E2; @@ - get_sha1_committish(E1, E2->hash) + get_oid_committish(E1, E2) @@ expression E1, E2; @@ - get_sha1_treeish(E1, E2.hash) + get_oid_treeish(E1, &E2) @@ expression E1, E2; @@ - get_sha1_treeish(E1, E2->hash) + get_oid_treeish(E1, E2) @@ expression E1, E2; @@ - get_sha1_commit(E1, E2.hash) + get_oid_commit(E1, &E2) @@ expression E1, E2; @@ - get_sha1_commit(E1, E2->hash) + get_oid_commit(E1, E2) @@ expression E1, E2; @@ - get_sha1_tree(E1, E2.hash) + get_oid_tree(E1, &E2) @@ expression E1, E2; @@ - get_sha1_tree(E1, E2->hash) + get_oid_tree(E1, E2) @@ expression E1, E2; @@ - get_sha1_blob(E1, E2.hash) + get_oid_blob(E1, &E2) @@ expression E1, E2; @@ - get_sha1_blob(E1, E2->hash) + get_oid_blob(E1, E2) @@ expression E1, E2, E3, E4; @@ - get_sha1_with_context(E1, E2, E3.hash, E4) + get_oid_with_context(E1, E2, &E3, E4) @@ expression E1, E2, E3, E4; @@ - get_sha1_with_context(E1, E2, E3->hash, E4) + get_oid_with_context(E1, E2, E3, E4) Signed-off-by: brian m. carlson <sandals@crustytoothpaste.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-07-14 02:49:28 +03:00
!get_oid_committish(buf.buf, &oid))
record_in_rewritten(&oid, peek_command(&todo_list, 0));
strbuf_release(&buf);
}
res = pick_commits(r, &todo_list, opts);
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
release_todo_list:
todo_list_release(&todo_list);
return res;
}
static int single_pick(struct repository *r,
struct commit *cmit,
struct replay_opts *opts)
{
int check_todo;
setenv(GIT_REFLOG_ACTION, action_name(opts), 0);
return do_pick_commit(r, opts->action == REPLAY_PICK ?
TODO_PICK : TODO_REVERT, cmit, opts, 0,
&check_todo);
}
int sequencer_pick_revisions(struct repository *r,
struct replay_opts *opts)
{
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
struct todo_list todo_list = TODO_LIST_INIT;
struct object_id oid;
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
int i, res;
assert(opts->revs);
if (read_and_refresh_cache(r, opts))
return -1;
for (i = 0; i < opts->revs->pending.nr; i++) {
struct object_id oid;
const char *name = opts->revs->pending.objects[i].name;
/* This happens when using --stdin. */
if (!strlen(name))
continue;
if (!get_oid(name, &oid)) {
if (!lookup_commit_reference_gently(r, &oid, 1)) {
enum object_type type = oid_object_info(r,
&oid,
NULL);
return error(_("%s: can't cherry-pick a %s"),
name, type_name(type));
}
} else
return error(_("%s: bad revision"), name);
}
/*
* If we were called as "git cherry-pick <commit>", just
* cherry-pick/revert it, set CHERRY_PICK_HEAD /
* REVERT_HEAD, and don't touch the sequencer state.
* This means it is possible to cherry-pick in the middle
* of a cherry-pick sequence.
*/
if (opts->revs->cmdline.nr == 1 &&
opts->revs->cmdline.rev->whence == REV_CMD_REV &&
opts->revs->no_walk &&
!opts->revs->cmdline.rev->flags) {
struct commit *cmit;
if (prepare_revision_walk(opts->revs))
return error(_("revision walk setup failed"));
cmit = get_revision(opts->revs);
sequencer: don't say BUG on bogus input When cherry-picking a single commit, we go through a special code path that avoids creating a sequencer todo list at all. This path expects our revision parsing to turn up exactly one commit, and dies with a BUG if it doesn't. But it's actually quite easy to fool. For example: $ git cherry-pick --author=no.such.person HEAD error: BUG: expected exactly one commit from walk fatal: cherry-pick failed This isn't a bug; it's just bogus input. The condition to trigger this message actually has two parts: 1. We saw no commits. That's the case in the example above. Let's drop the "BUG" here to make it clear that the input is the problem. And let's also use the phrase "empty commit set passed", which matches what we say when we do a real revision walk and it turns up empty. 2. We saw more than one commit. That one _should_ be impossible to trigger, since we fed at most one tip and provided the no_walk option (and we'll have already expanded options like "--branches" that can turn into multiple tips). If this ever triggers, it's an indication that the conditional added by 7acaaac275 (revert: allow single-pick in the middle of cherry-pick sequence, 2011-12-10) needs to more carefully define the single-pick case. So this can remain a bug, but we'll upgrade it to use the BUG() macro, which would make it easier to detect and analyze if it does trigger. Signed-off-by: Jeff King <peff@peff.net> Acked-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-07-10 07:32:08 +03:00
if (!cmit)
return error(_("empty commit set passed"));
if (get_revision(opts->revs))
BUG("unexpected extra commit from walk");
return single_pick(r, cmit, opts);
}
/*
* Start a new cherry-pick/ revert sequence; but
* first, make sure that an existing one isn't in
* progress
*/
if (walk_revs_populate_todo(&todo_list, opts) ||
create_seq_dir(r) < 0)
return -1;
if (get_oid("HEAD", &oid) && (opts->action == REPLAY_REVERT))
return error(_("can't revert as initial commit"));
if (save_head(oid_to_hex(&oid)))
return -1;
if (save_opts(opts))
return -1;
update_abort_safety_file();
res = pick_commits(r, &todo_list, opts);
sequencer: completely revamp the "todo" script parsing When we came up with the "sequencer" idea, we really wanted to have kind of a plumbing equivalent of the interactive rebase. Hence the choice of words: the "todo" script, a "pick", etc. However, when it came time to implement the entire shebang, somehow this idea got lost and the sequencer was used as working horse for cherry-pick and revert instead. So as not to interfere with the interactive rebase, it even uses a separate directory to store its state. Furthermore, it also is stupidly strict about the "todo" script it accepts: while it parses commands in a way that was *designed* to be similar to the interactive rebase, it then goes on to *error out* if the commands disagree with the overall action (cherry-pick or revert). Finally, the sequencer code chose to deviate from the interactive rebase code insofar that when it comes to writing the file with the remaining commands, it *reformats* the "todo" script instead of just writing the part of the parsed script that were not yet processed. This is not only unnecessary churn, but might well lose information that is valuable to the user (i.e. comments after the commands). Let's just bite the bullet and rewrite the entire parser; the code now becomes not only more elegant: it allows us to go on and teach the sequencer how to parse *true* "todo" scripts as used by the interactive rebase itself. In a way, the sequencer is about to grow up to do its older brother's job. Better. In particular, we choose to maintain the list of commands in an array instead of a linked list: this is flexible enough to allow us later on to even implement rebase -i's reordering of fixup!/squash! commits very easily (and with a very nice speed bonus, at least on Windows). While at it, do not stop at the first problem, but list *all* of the problems. This will help the user when the sequencer will do `rebase -i`'s work by allowing to address all issues in one go rather than going back and forth until the todo list is valid. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-10-21 15:24:41 +03:00
todo_list_release(&todo_list);
return res;
}
void append_signoff(struct strbuf *msgbuf, size_t ignore_footer, unsigned flag)
{
unsigned no_dup_sob = flag & APPEND_SIGNOFF_DEDUP;
struct strbuf sob = STRBUF_INIT;
int has_footer;
strbuf_addstr(&sob, sign_off_header);
strbuf_addstr(&sob, fmt_name(WANT_COMMITTER_IDENT));
strbuf_addch(&sob, '\n');
if (!ignore_footer)
strbuf_complete_line(msgbuf);
/*
* If the whole message buffer is equal to the sob, pretend that we
* found a conforming footer with a matching sob
*/
if (msgbuf->len - ignore_footer == sob.len &&
!strncmp(msgbuf->buf, sob.buf, sob.len))
has_footer = 3;
else
has_footer = has_conforming_footer(msgbuf, &sob, ignore_footer);
if (!has_footer) {
const char *append_newlines = NULL;
size_t len = msgbuf->len - ignore_footer;
if (!len) {
/*
* The buffer is completely empty. Leave foom for
* the title and body to be filled in by the user.
*/
append_newlines = "\n\n";
} else if (len == 1) {
/*
* Buffer contains a single newline. Add another
* so that we leave room for the title and body.
*/
append_newlines = "\n";
} else if (msgbuf->buf[len - 2] != '\n') {
/*
* Buffer ends with a single newline. Add another
* so that there is an empty line between the message
* body and the sob.
*/
append_newlines = "\n";
} /* else, the buffer already ends with two newlines. */
if (append_newlines)
strbuf_splice(msgbuf, msgbuf->len - ignore_footer, 0,
append_newlines, strlen(append_newlines));
}
if (has_footer != 3 && (!no_dup_sob || has_footer != 2))
strbuf_splice(msgbuf, msgbuf->len - ignore_footer, 0,
sob.buf, sob.len);
strbuf_release(&sob);
}
struct labels_entry {
struct hashmap_entry entry;
char label[FLEX_ARRAY];
};
static int labels_cmp(const void *fndata, const struct hashmap_entry *eptr,
const struct hashmap_entry *entry_or_key, const void *key)
{
const struct labels_entry *a, *b;
a = container_of(eptr, const struct labels_entry, entry);
b = container_of(entry_or_key, const struct labels_entry, entry);
return key ? strcmp(a->label, key) : strcmp(a->label, b->label);
}
struct string_entry {
struct oidmap_entry entry;
char string[FLEX_ARRAY];
};
struct label_state {
struct oidmap commit2label;
struct hashmap labels;
struct strbuf buf;
};
static const char *label_oid(struct object_id *oid, const char *label,
struct label_state *state)
{
struct labels_entry *labels_entry;
struct string_entry *string_entry;
struct object_id dummy;
size_t len;
int i;
string_entry = oidmap_get(&state->commit2label, oid);
if (string_entry)
return string_entry->string;
/*
* For "uninteresting" commits, i.e. commits that are not to be
* rebased, and which can therefore not be labeled, we use a unique
* abbreviation of the commit name. This is slightly more complicated
* than calling find_unique_abbrev() because we also need to make
* sure that the abbreviation does not conflict with any other
* label.
*
* We disallow "interesting" commits to be labeled by a string that
* is a valid full-length hash, to ensure that we always can find an
* abbreviation for any uninteresting commit's names that does not
* clash with any other label.
*/
if (!label) {
char *p;
strbuf_reset(&state->buf);
strbuf_grow(&state->buf, GIT_MAX_HEXSZ);
label = p = state->buf.buf;
find_unique_abbrev_r(p, oid, default_abbrev);
/*
* We may need to extend the abbreviated hash so that there is
* no conflicting label.
*/
if (hashmap_get_from_hash(&state->labels, strihash(p), p)) {
size_t i = strlen(p) + 1;
oid_to_hex_r(p, oid);
for (; i < the_hash_algo->hexsz; i++) {
char save = p[i];
p[i] = '\0';
if (!hashmap_get_from_hash(&state->labels,
strihash(p), p))
break;
p[i] = save;
}
}
} else if (((len = strlen(label)) == the_hash_algo->hexsz &&
!get_oid_hex(label, &dummy)) ||
(len == 1 && *label == '#') ||
hashmap_get_from_hash(&state->labels,
strihash(label), label)) {
/*
* If the label already exists, or if the label is a valid full
* OID, or the label is a '#' (which we use as a separator
* between merge heads and oneline), we append a dash and a
* number to make it unique.
*/
struct strbuf *buf = &state->buf;
strbuf_reset(buf);
strbuf_add(buf, label, len);
for (i = 2; ; i++) {
strbuf_setlen(buf, len);
strbuf_addf(buf, "-%d", i);
if (!hashmap_get_from_hash(&state->labels,
strihash(buf->buf),
buf->buf))
break;
}
label = buf->buf;
}
FLEX_ALLOC_STR(labels_entry, label, label);
hashmap_entry_init(&labels_entry->entry, strihash(label));
hashmap_add(&state->labels, &labels_entry->entry);
FLEX_ALLOC_STR(string_entry, string, label);
oidcpy(&string_entry->entry.oid, oid);
oidmap_put(&state->commit2label, string_entry);
return string_entry->string;
}
static int make_script_with_merges(struct pretty_print_context *pp,
struct rev_info *revs, struct strbuf *out,
unsigned flags)
{
int keep_empty = flags & TODO_LIST_KEEP_EMPTY;
rebase -i: introduce --rebase-merges=[no-]rebase-cousins When running `git rebase --rebase-merges` non-interactively with an ancestor of HEAD as <upstream> (or leaving the todo list unmodified), we would ideally recreate the exact same commits as before the rebase. However, if there are commits in the commit range <upstream>.. that do not have <upstream> as direct ancestor (i.e. if `git log <upstream>..` would show commits that are omitted by `git log --ancestry-path <upstream>..`), this is currently not the case: we would turn them into commits that have <upstream> as direct ancestor. Let's illustrate that with a diagram: C / \ A - B - E - F \ / D Currently, after running `git rebase -i --rebase-merges B`, the new branch structure would be (pay particular attention to the commit `D`): --- C' -- / \ A - B ------ E' - F' \ / D' This is not really preserving the branch topology from before! The reason is that the commit `D` does not have `B` as ancestor, and therefore it gets rebased onto `B`. This is unintuitive behavior. Even worse, when recreating branch structure, most use cases would appear to want cousins *not* to be rebased onto the new base commit. For example, Git for Windows (the heaviest user of the Git garden shears, which served as the blueprint for --rebase-merges) frequently merges branches from `next` early, and these branches certainly do *not* want to be rebased. In the example above, the desired outcome would look like this: --- C' -- / \ A - B ------ E' - F' \ / -- D' -- Let's introduce the term "cousins" for such commits ("D" in the example), and let's not rebase them by default. For hypothetical use cases where cousins *do* need to be rebased, `git rebase --rebase=merges=rebase-cousins` needs to be used. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:29:40 +03:00
int rebase_cousins = flags & TODO_LIST_REBASE_COUSINS;
int root_with_onto = flags & TODO_LIST_ROOT_WITH_ONTO;
struct strbuf buf = STRBUF_INIT, oneline = STRBUF_INIT;
struct strbuf label = STRBUF_INIT;
struct commit_list *commits = NULL, **tail = &commits, *iter;
struct commit_list *tips = NULL, **tips_tail = &tips;
struct commit *commit;
struct oidmap commit2todo = OIDMAP_INIT;
struct string_entry *entry;
struct oidset interesting = OIDSET_INIT, child_seen = OIDSET_INIT,
shown = OIDSET_INIT;
struct label_state state = { OIDMAP_INIT, { NULL }, STRBUF_INIT };
int abbr = flags & TODO_LIST_ABBREVIATE_CMDS;
const char *cmd_pick = abbr ? "p" : "pick",
*cmd_label = abbr ? "l" : "label",
*cmd_reset = abbr ? "t" : "reset",
*cmd_merge = abbr ? "m" : "merge";
oidmap_init(&commit2todo, 0);
oidmap_init(&state.commit2label, 0);
hashmap_init(&state.labels, labels_cmp, NULL, 0);
strbuf_init(&state.buf, 32);
if (revs->cmdline.nr && (revs->cmdline.rev[0].flags & BOTTOM)) {
struct object_id *oid = &revs->cmdline.rev[0].item->oid;
FLEX_ALLOC_STR(entry, string, "onto");
oidcpy(&entry->entry.oid, oid);
oidmap_put(&state.commit2label, entry);
}
/*
* First phase:
* - get onelines for all commits
* - gather all branch tips (i.e. 2nd or later parents of merges)
* - label all branch tips
*/
while ((commit = get_revision(revs))) {
struct commit_list *to_merge;
const char *p1, *p2;
struct object_id *oid;
int is_empty;
tail = &commit_list_insert(commit, tail)->next;
oidset_insert(&interesting, &commit->object.oid);
is_empty = is_original_commit_empty(commit);
if (!is_empty && (commit->object.flags & PATCHSAME))
continue;
strbuf_reset(&oneline);
pretty_print_commit(pp, commit, &oneline);
to_merge = commit->parents ? commit->parents->next : NULL;
if (!to_merge) {
/* non-merge commit: easy case */
strbuf_reset(&buf);
if (!keep_empty && is_empty)
strbuf_addf(&buf, "%c ", comment_line_char);
strbuf_addf(&buf, "%s %s %s", cmd_pick,
oid_to_hex(&commit->object.oid),
oneline.buf);
FLEX_ALLOC_STR(entry, string, buf.buf);
oidcpy(&entry->entry.oid, &commit->object.oid);
oidmap_put(&commit2todo, entry);
continue;
}
/* Create a label */
strbuf_reset(&label);
if (skip_prefix(oneline.buf, "Merge ", &p1) &&
(p1 = strchr(p1, '\'')) &&
(p2 = strchr(++p1, '\'')))
strbuf_add(&label, p1, p2 - p1);
else if (skip_prefix(oneline.buf, "Merge pull request ",
&p1) &&
(p1 = strstr(p1, " from ")))
strbuf_addstr(&label, p1 + strlen(" from "));
else
strbuf_addbuf(&label, &oneline);
for (p1 = label.buf; *p1; p1++)
if (isspace(*p1))
*(char *)p1 = '-';
strbuf_reset(&buf);
strbuf_addf(&buf, "%s -C %s",
cmd_merge, oid_to_hex(&commit->object.oid));
/* label the tips of merged branches */
for (; to_merge; to_merge = to_merge->next) {
oid = &to_merge->item->object.oid;
strbuf_addch(&buf, ' ');
if (!oidset_contains(&interesting, oid)) {
strbuf_addstr(&buf, label_oid(oid, NULL,
&state));
continue;
}
tips_tail = &commit_list_insert(to_merge->item,
tips_tail)->next;
strbuf_addstr(&buf, label_oid(oid, label.buf, &state));
}
strbuf_addf(&buf, " # %s", oneline.buf);
FLEX_ALLOC_STR(entry, string, buf.buf);
oidcpy(&entry->entry.oid, &commit->object.oid);
oidmap_put(&commit2todo, entry);
}
/*
* Second phase:
* - label branch points
* - add HEAD to the branch tips
*/
for (iter = commits; iter; iter = iter->next) {
struct commit_list *parent = iter->item->parents;
for (; parent; parent = parent->next) {
struct object_id *oid = &parent->item->object.oid;
if (!oidset_contains(&interesting, oid))
continue;
if (oidset_insert(&child_seen, oid))
label_oid(oid, "branch-point", &state);
}
/* Add HEAD as implict "tip of branch" */
if (!iter->next)
tips_tail = &commit_list_insert(iter->item,
tips_tail)->next;
}
/*
* Third phase: output the todo list. This is a bit tricky, as we
* want to avoid jumping back and forth between revisions. To
* accomplish that goal, we walk backwards from the branch tips,
* gathering commits not yet shown, reversing the list on the fly,
* then outputting that list (labeling revisions as needed).
*/
strbuf_addf(out, "%s onto\n", cmd_label);
for (iter = tips; iter; iter = iter->next) {
struct commit_list *list = NULL, *iter2;
commit = iter->item;
if (oidset_contains(&shown, &commit->object.oid))
continue;
entry = oidmap_get(&state.commit2label, &commit->object.oid);
if (entry)
strbuf_addf(out, "\n%c Branch %s\n", comment_line_char, entry->string);
else
strbuf_addch(out, '\n');
while (oidset_contains(&interesting, &commit->object.oid) &&
!oidset_contains(&shown, &commit->object.oid)) {
commit_list_insert(commit, &list);
if (!commit->parents) {
commit = NULL;
break;
}
commit = commit->parents->item;
}
if (!commit)
strbuf_addf(out, "%s %s\n", cmd_reset,
rebase_cousins || root_with_onto ?
"onto" : "[new root]");
else {
const char *to = NULL;
entry = oidmap_get(&state.commit2label,
&commit->object.oid);
if (entry)
to = entry->string;
rebase -i: introduce --rebase-merges=[no-]rebase-cousins When running `git rebase --rebase-merges` non-interactively with an ancestor of HEAD as <upstream> (or leaving the todo list unmodified), we would ideally recreate the exact same commits as before the rebase. However, if there are commits in the commit range <upstream>.. that do not have <upstream> as direct ancestor (i.e. if `git log <upstream>..` would show commits that are omitted by `git log --ancestry-path <upstream>..`), this is currently not the case: we would turn them into commits that have <upstream> as direct ancestor. Let's illustrate that with a diagram: C / \ A - B - E - F \ / D Currently, after running `git rebase -i --rebase-merges B`, the new branch structure would be (pay particular attention to the commit `D`): --- C' -- / \ A - B ------ E' - F' \ / D' This is not really preserving the branch topology from before! The reason is that the commit `D` does not have `B` as ancestor, and therefore it gets rebased onto `B`. This is unintuitive behavior. Even worse, when recreating branch structure, most use cases would appear to want cousins *not* to be rebased onto the new base commit. For example, Git for Windows (the heaviest user of the Git garden shears, which served as the blueprint for --rebase-merges) frequently merges branches from `next` early, and these branches certainly do *not* want to be rebased. In the example above, the desired outcome would look like this: --- C' -- / \ A - B ------ E' - F' \ / -- D' -- Let's introduce the term "cousins" for such commits ("D" in the example), and let's not rebase them by default. For hypothetical use cases where cousins *do* need to be rebased, `git rebase --rebase=merges=rebase-cousins` needs to be used. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:29:40 +03:00
else if (!rebase_cousins)
to = label_oid(&commit->object.oid, NULL,
&state);
if (!to || !strcmp(to, "onto"))
strbuf_addf(out, "%s onto\n", cmd_reset);
else {
strbuf_reset(&oneline);
pretty_print_commit(pp, commit, &oneline);
strbuf_addf(out, "%s %s # %s\n",
cmd_reset, to, oneline.buf);
}
}
for (iter2 = list; iter2; iter2 = iter2->next) {
struct object_id *oid = &iter2->item->object.oid;
entry = oidmap_get(&commit2todo, oid);
/* only show if not already upstream */
if (entry)
strbuf_addf(out, "%s\n", entry->string);
entry = oidmap_get(&state.commit2label, oid);
if (entry)
strbuf_addf(out, "%s %s\n",
cmd_label, entry->string);
oidset_insert(&shown, oid);
}
free_commit_list(list);
}
free_commit_list(commits);
free_commit_list(tips);
strbuf_release(&label);
strbuf_release(&oneline);
strbuf_release(&buf);
oidmap_free(&commit2todo, 1);
oidmap_free(&state.commit2label, 1);
hashmap_free_entries(&state.labels, struct labels_entry, entry);
strbuf_release(&state.buf);
return 0;
}
int sequencer_make_script(struct repository *r, struct strbuf *out, int argc,
const char **argv, unsigned flags)
{
char *format = NULL;
struct pretty_print_context pp = {0};
struct rev_info revs;
struct commit *commit;
int keep_empty = flags & TODO_LIST_KEEP_EMPTY;
const char *insn = flags & TODO_LIST_ABBREVIATE_CMDS ? "p" : "pick";
int rebase_merges = flags & TODO_LIST_REBASE_MERGES;
repo_init_revisions(r, &revs, NULL);
revs.verbose_header = 1;
if (!rebase_merges)
revs.max_parents = 1;
revs.cherry_mark = 1;
revs.limited = 1;
revs.reverse = 1;
revs.right_only = 1;
revs.sort_order = REV_SORT_IN_GRAPH_ORDER;
revs.topo_order = 1;
revs.pretty_given = 1;
git_config_get_string("rebase.instructionFormat", &format);
if (!format || !*format) {
free(format);
format = xstrdup("%s");
}
get_commit_format(format, &revs);
free(format);
pp.fmt = revs.commit_format;
pp.output_encoding = get_log_output_encoding();
if (setup_revisions(argc, argv, &revs, NULL) > 1)
return error(_("make_script: unhandled options"));
if (prepare_revision_walk(&revs) < 0)
return error(_("make_script: error preparing revisions"));
if (rebase_merges)
return make_script_with_merges(&pp, &revs, out, flags);
while ((commit = get_revision(&revs))) {
int is_empty = is_original_commit_empty(commit);
if (!is_empty && (commit->object.flags & PATCHSAME))
continue;
if (!keep_empty && is_empty)
strbuf_addf(out, "%c ", comment_line_char);
strbuf_addf(out, "%s %s ", insn,
oid_to_hex(&commit->object.oid));
pretty_print_commit(&pp, commit, out);
strbuf_addch(out, '\n');
}
return 0;
}
/*
* Add commands after pick and (series of) squash/fixup commands
* in the todo list.
*/
void todo_list_add_exec_commands(struct todo_list *todo_list,
struct string_list *commands)
{
struct strbuf *buf = &todo_list->buf;
size_t base_offset = buf->len;
int i, insert, nr = 0, alloc = 0;
struct todo_item *items = NULL, *base_items = NULL;
base_items = xcalloc(commands->nr, sizeof(struct todo_item));
for (i = 0; i < commands->nr; i++) {
size_t command_len = strlen(commands->items[i].string);
strbuf_addstr(buf, commands->items[i].string);
strbuf_addch(buf, '\n');
base_items[i].command = TODO_EXEC;
base_items[i].offset_in_buf = base_offset;
base_items[i].arg_offset = base_offset + strlen("exec ");
base_items[i].arg_len = command_len - strlen("exec ");
base_offset += command_len + 1;
}
/*
* Insert <commands> after every pick. Here, fixup/squash chains
* are considered part of the pick, so we insert the commands *after*
* those chains if there are any.
*
* As we insert the exec commands immediatly after rearranging
* any fixups and before the user edits the list, a fixup chain
* can never contain comments (any comments are empty picks that
* have been commented out because the user did not specify
* --keep-empty). So, it is safe to insert an exec command
* without looking at the command following a comment.
*/
insert = 0;
for (i = 0; i < todo_list->nr; i++) {
enum todo_command command = todo_list->items[i].command;
if (insert && !is_fixup(command)) {
ALLOC_GROW(items, nr + commands->nr, alloc);
COPY_ARRAY(items + nr, base_items, commands->nr);
nr += commands->nr;
insert = 0;
}
ALLOC_GROW(items, nr + 1, alloc);
items[nr++] = todo_list->items[i];
if (command == TODO_PICK || command == TODO_MERGE)
insert = 1;
}
/* insert or append final <commands> */
if (insert || nr == todo_list->nr) {
ALLOC_GROW(items, nr + commands->nr, alloc);
COPY_ARRAY(items + nr, base_items, commands->nr);
nr += commands->nr;
}
free(base_items);
FREE_AND_NULL(todo_list->items);
todo_list->items = items;
todo_list->nr = nr;
todo_list->alloc = alloc;
}
static void todo_list_to_strbuf(struct repository *r, struct todo_list *todo_list,
struct strbuf *buf, int num, unsigned flags)
{
struct todo_item *item;
int i, max = todo_list->nr;
if (num > 0 && num < max)
max = num;
for (item = todo_list->items, i = 0; i < max; i++, item++) {
/* if the item is not a command write it and continue */
if (item->command >= TODO_COMMENT) {
strbuf_addf(buf, "%.*s\n", item->arg_len,
todo_item_get_arg(todo_list, item));
continue;
}
/* add command to the buffer */
if (flags & TODO_LIST_ABBREVIATE_CMDS)
strbuf_addch(buf, command_to_char(item->command));
else
strbuf_addstr(buf, command_to_string(item->command));
/* add commit id */
if (item->commit) {
const char *oid = flags & TODO_LIST_SHORTEN_IDS ?
short_commit_name(item->commit) :
oid_to_hex(&item->commit->object.oid);
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
if (item->command == TODO_MERGE) {
if (item->flags & TODO_EDIT_MERGE_MSG)
strbuf_addstr(buf, " -c");
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
else
strbuf_addstr(buf, " -C");
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
}
strbuf_addf(buf, " %s", oid);
}
sequencer: introduce the `merge` command This patch is part of the effort to reimplement `--preserve-merges` with a substantially improved design, a design that has been developed in the Git for Windows project to maintain the dozens of Windows-specific patch series on top of upstream Git. The previous patch implemented the `label` and `reset` commands to label commits and to reset to labeled commits. This patch adds the `merge` command, with the following syntax: merge [-C <commit>] <rev> # <oneline> The <commit> parameter in this instance is the *original* merge commit, whose author and message will be used for the merge commit that is about to be created. The <rev> parameter refers to the (possibly rewritten) revision to merge. Let's see an example of a todo list (the initial `label onto` command is an auto-generated convenience so that the label `onto` can be used to refer to the revision onto which we rebase): label onto # Branch abc reset onto pick deadbeef Hello, world! label abc reset onto pick cafecafe And now for something completely different merge -C baaabaaa abc # Merge the branch 'abc' into master To edit the merge commit's message (a "reword" for merges, if you will), use `-c` (lower-case) instead of `-C`; this convention was borrowed from `git commit` that also supports `-c` and `-C` with similar meanings. To create *new* merges, i.e. without copying the commit message from an existing commit, simply omit the `-C <commit>` parameter (which will open an editor for the merge message): merge abc This comes in handy when splitting a branch into two or more branches. Note: this patch only adds support for recursive merges, to keep things simple. Support for octopus merges will be added later in a separate patch series, support for merges using strategies other than the recursive merge is left for the future. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-04-25 15:28:54 +03:00
/* add all the rest */
if (!item->arg_len)
strbuf_addch(buf, '\n');
else
strbuf_addf(buf, " %.*s\n", item->arg_len,
todo_item_get_arg(todo_list, item));
}
}
int todo_list_write_to_file(struct repository *r, struct todo_list *todo_list,
const char *file, const char *shortrevisions,
const char *shortonto, int num, unsigned flags)
{
int res;
struct strbuf buf = STRBUF_INIT;
todo_list_to_strbuf(r, todo_list, &buf, num, flags);
if (flags & TODO_LIST_APPEND_TODO_HELP)
append_todo_help(flags & TODO_LIST_KEEP_EMPTY, count_commands(todo_list),
shortrevisions, shortonto, &buf);
res = write_message(buf.buf, buf.len, file, 0);
strbuf_release(&buf);
return res;
}
static const char edit_todo_list_advice[] =
N_("You can fix this with 'git rebase --edit-todo' "
"and then run 'git rebase --continue'.\n"
"Or you can abort the rebase with 'git rebase"
" --abort'.\n");
int check_todo_list_from_file(struct repository *r)
{
struct todo_list old_todo = TODO_LIST_INIT, new_todo = TODO_LIST_INIT;
int res = 0;
if (strbuf_read_file_or_whine(&new_todo.buf, rebase_path_todo()) < 0) {
res = -1;
goto out;
}
if (strbuf_read_file_or_whine(&old_todo.buf, rebase_path_todo_backup()) < 0) {
res = -1;
goto out;
}
res = todo_list_parse_insn_buffer(r, old_todo.buf.buf, &old_todo);
if (!res)
res = todo_list_parse_insn_buffer(r, new_todo.buf.buf, &new_todo);
if (!res)
res = todo_list_check(&old_todo, &new_todo);
if (res)
fprintf(stderr, _(edit_todo_list_advice));
out:
todo_list_release(&old_todo);
todo_list_release(&new_todo);
return res;
}
/* skip picking commits whose parents are unchanged */
static int skip_unnecessary_picks(struct repository *r,
struct todo_list *todo_list,
struct object_id *base_oid)
{
struct object_id *parent_oid;
int i;
for (i = 0; i < todo_list->nr; i++) {
struct todo_item *item = todo_list->items + i;
if (item->command >= TODO_NOOP)
continue;
if (item->command != TODO_PICK)
break;
if (parse_commit(item->commit)) {
return error(_("could not parse commit '%s'"),
oid_to_hex(&item->commit->object.oid));
}
if (!item->commit->parents)
break; /* root commit */
if (item->commit->parents->next)
break; /* merge commit */
parent_oid = &item->commit->parents->item->object.oid;
if (!oideq(parent_oid, base_oid))
break;
oidcpy(base_oid, &item->commit->object.oid);
}
if (i > 0) {
const char *done_path = rebase_path_done();
if (todo_list_write_to_file(r, todo_list, done_path, NULL, NULL, i, 0)) {
error_errno(_("could not write to '%s'"), done_path);
return -1;
}
MOVE_ARRAY(todo_list->items, todo_list->items + i, todo_list->nr - i);
todo_list->nr -= i;
todo_list->current = 0;
if (is_fixup(peek_command(todo_list, 0)))
record_in_rewritten(base_oid, peek_command(todo_list, 0));
}
return 0;
}
int complete_action(struct repository *r, struct replay_opts *opts, unsigned flags,
const char *shortrevisions, const char *onto_name,
struct commit *onto, const char *orig_head,
struct string_list *commands, unsigned autosquash,
struct todo_list *todo_list)
{
const char *shortonto, *todo_file = rebase_path_todo();
struct todo_list new_todo = TODO_LIST_INIT;
struct strbuf *buf = &todo_list->buf;
struct object_id oid = onto->object.oid;
int res;
shortonto = find_unique_abbrev(&oid, DEFAULT_ABBREV);
if (buf->len == 0) {
struct todo_item *item = append_new_todo(todo_list);
item->command = TODO_NOOP;
item->commit = NULL;
item->arg_len = item->arg_offset = item->flags = item->offset_in_buf = 0;
}
if (autosquash && todo_list_rearrange_squash(todo_list))
return -1;
if (commands->nr)
todo_list_add_exec_commands(todo_list, commands);
if (count_commands(todo_list) == 0) {
apply_autostash(opts);
sequencer_remove_state(opts);
return error(_("nothing to do"));
}
res = edit_todo_list(r, todo_list, &new_todo, shortrevisions,
shortonto, flags);
if (res == -1)
return -1;
else if (res == -2) {
apply_autostash(opts);
sequencer_remove_state(opts);
return -1;
} else if (res == -3) {
apply_autostash(opts);
sequencer_remove_state(opts);
todo_list_release(&new_todo);
return error(_("nothing to do"));
}
if (todo_list_parse_insn_buffer(r, new_todo.buf.buf, &new_todo) ||
todo_list_check(todo_list, &new_todo)) {
fprintf(stderr, _(edit_todo_list_advice));
checkout_onto(r, opts, onto_name, &onto->object.oid, orig_head);
todo_list_release(&new_todo);
return -1;
}
if (opts->allow_ff && skip_unnecessary_picks(r, &new_todo, &oid)) {
todo_list_release(&new_todo);
return error(_("could not skip unnecessary pick commands"));
}
if (todo_list_write_to_file(r, &new_todo, todo_file, NULL, NULL, -1,
flags & ~(TODO_LIST_SHORTEN_IDS))) {
todo_list_release(&new_todo);
return error_errno(_("could not write '%s'"), todo_file);
}
todo_list_release(&new_todo);
if (checkout_onto(r, opts, onto_name, &oid, orig_head))
return -1;
if (require_clean_work_tree(r, "rebase", "", 1, 1))
return -1;
return sequencer_continue(r, opts);
}
struct subject2item_entry {
struct hashmap_entry entry;
int i;
char subject[FLEX_ARRAY];
};
static int subject2item_cmp(const void *fndata,
const struct hashmap_entry *eptr,
const struct hashmap_entry *entry_or_key,
const void *key)
{
const struct subject2item_entry *a, *b;
a = container_of(eptr, const struct subject2item_entry, entry);
b = container_of(entry_or_key, const struct subject2item_entry, entry);
return key ? strcmp(a->subject, key) : strcmp(a->subject, b->subject);
}
define_commit_slab(commit_todo_item, struct todo_item *);
/*
* Rearrange the todo list that has both "pick commit-id msg" and "pick
* commit-id fixup!/squash! msg" in it so that the latter is put immediately
* after the former, and change "pick" to "fixup"/"squash".
*
* Note that if the config has specified a custom instruction format, each log
* message will have to be retrieved from the commit (as the oneline in the
* script cannot be trusted) in order to normalize the autosquash arrangement.
*/
int todo_list_rearrange_squash(struct todo_list *todo_list)
{
struct hashmap subject2item;
int rearranged = 0, *next, *tail, i, nr = 0, alloc = 0;
char **subjects;
struct commit_todo_item commit_todo;
struct todo_item *items = NULL;
init_commit_todo_item(&commit_todo);
/*
* The hashmap maps onelines to the respective todo list index.
*
* If any items need to be rearranged, the next[i] value will indicate
* which item was moved directly after the i'th.
*
* In that case, last[i] will indicate the index of the latest item to
* be moved to appear after the i'th.
*/
hashmap_init(&subject2item, subject2item_cmp, NULL, todo_list->nr);
ALLOC_ARRAY(next, todo_list->nr);
ALLOC_ARRAY(tail, todo_list->nr);
ALLOC_ARRAY(subjects, todo_list->nr);
for (i = 0; i < todo_list->nr; i++) {
struct strbuf buf = STRBUF_INIT;
struct todo_item *item = todo_list->items + i;
const char *commit_buffer, *subject, *p;
size_t subject_len;
int i2 = -1;
struct subject2item_entry *entry;
next[i] = tail[i] = -1;
if (!item->commit || item->command == TODO_DROP) {
subjects[i] = NULL;
continue;
}
if (is_fixup(item->command)) {
clear_commit_todo_item(&commit_todo);
return error(_("the script was already rearranged."));
}
*commit_todo_item_at(&commit_todo, item->commit) = item;
parse_commit(item->commit);
commit_buffer = get_commit_buffer(item->commit, NULL);
find_commit_subject(commit_buffer, &subject);
format_subject(&buf, subject, " ");
subject = subjects[i] = strbuf_detach(&buf, &subject_len);
unuse_commit_buffer(item->commit, commit_buffer);
if ((skip_prefix(subject, "fixup! ", &p) ||
skip_prefix(subject, "squash! ", &p))) {
struct commit *commit2;
for (;;) {
while (isspace(*p))
p++;
if (!skip_prefix(p, "fixup! ", &p) &&
!skip_prefix(p, "squash! ", &p))
break;
}
entry = hashmap_get_entry_from_hash(&subject2item,
strhash(p), p,
struct subject2item_entry,
entry);
if (entry)
/* found by title */
i2 = entry->i;
else if (!strchr(p, ' ') &&
(commit2 =
lookup_commit_reference_by_name(p)) &&
*commit_todo_item_at(&commit_todo, commit2))
/* found by commit name */
i2 = *commit_todo_item_at(&commit_todo, commit2)
- todo_list->items;
else {
/* copy can be a prefix of the commit subject */
for (i2 = 0; i2 < i; i2++)
if (subjects[i2] &&
starts_with(subjects[i2], p))
break;
if (i2 == i)
i2 = -1;
}
}
if (i2 >= 0) {
rearranged = 1;
todo_list->items[i].command =
starts_with(subject, "fixup!") ?
TODO_FIXUP : TODO_SQUASH;
if (next[i2] < 0)
next[i2] = i;
else
next[tail[i2]] = i;
tail[i2] = i;
} else if (!hashmap_get_from_hash(&subject2item,
strhash(subject), subject)) {
FLEX_ALLOC_MEM(entry, subject, subject, subject_len);
entry->i = i;
hashmap_entry_init(&entry->entry,
strhash(entry->subject));
hashmap_put(&subject2item, &entry->entry);
}
}
if (rearranged) {
for (i = 0; i < todo_list->nr; i++) {
enum todo_command command = todo_list->items[i].command;
int cur = i;
/*
* Initially, all commands are 'pick's. If it is a
* fixup or a squash now, we have rearranged it.
*/
if (is_fixup(command))
continue;
while (cur >= 0) {
ALLOC_GROW(items, nr + 1, alloc);
items[nr++] = todo_list->items[cur];
cur = next[cur];
}
}
FREE_AND_NULL(todo_list->items);
todo_list->items = items;
todo_list->nr = nr;
todo_list->alloc = alloc;
}
free(next);
free(tail);
for (i = 0; i < todo_list->nr; i++)
free(subjects[i]);
free(subjects);
hashmap_free_entries(&subject2item, struct subject2item_entry, entry);
clear_commit_todo_item(&commit_todo);
return 0;
}