Re: [PATCH v4 5/5] sequencer: directly call pick_commits() from complete_action()

From: Alban Gruin <alban.gruin@gmail.com>
To: Jonathan Tan <jonathantanmy@google.com>
Cc: git@vger.kernel.org, Johannes.Schindelin@gmx.de,
	phillip.wood@dunelm.org.uk, gitster@pobox.com
Subject: Re: [PATCH v4 5/5] sequencer: directly call pick_commits() from complete_action()
Date: Wed, 27 Nov 2019 20:56:03 +0100	[thread overview]
Message-ID: <aa162f9c-9ae8-9844-6829-7fc0d9c50f9c@gmail.com> (raw)
In-Reply-To: <20191126184103.146421-1-jonathantanmy@google.com>

Hi Jonathan,

Le 26/11/2019 à 19:41, Jonathan Tan a écrit :
>> Currently, complete_action(), used by builtin/rebase.c to start a new
>> rebase, calls sequencer_continue() to do it.  Before the former calls
>> pick_commits(), it
>>
>>  - calls read_and_refresh_cache() -- this is unnecessary here as we've
>>    just called require_clean_work_tree() in complete_action()
> 
> require_clean_work_tree() and read_and_refresh_cache() seem to be
> different functions - can you explain why running the former is a good
> substitute for running the latter?
> 

They both refresh the index.

require_clean_work_tree(), called when starting a new rebase, will also
check if there are any unstaged or uncommitted changes, in which case we
do not want to start a rebase.

This is not what we want when resuming a rebase (with `rebase
--continue'), because the changes might be the result of a conflict
resolution.  In this case, the last commit is amended, and the rebase is
resumed.

>>  - calls read_populate_opts() -- this is unnecessary as we're starting a
>>    new rebase, so `opts' is fully populated
> 
> My comment from [1] has not been addressed. Quoting it here:
> 
>> So complete_action() (the function modified in this commit) is called
>> only by do_interactive_rebase() (in builtin/rebase.c), which is only
>> called by run_rebase_interactive() (in builtin/rebase.c) when command is
>> ACTION_NONE, so indeed, we're starting a new rebase. But where the
>> options fully populated? I see that in do_interactive_rebase(), it is
>> initialized with get_replay_opts(), but that seems different from
>> read_populate_opts().
> 
> [1] https://lore.kernel.org/git/20191119204146.168001-1-jonathantanmy@google.com/
> 

Sorry.

For the first part of your comment, I added a comment at the beginning
of the message, although I did _not_ include an analysis on when
complete_action() is used.

get_replay_opts() converts a `struct rebase_options' (which contains the
arguments passed to `git rebase') into a `struct replay_opts' which can
be used by the sequencer, whereas read_populate_opts() loads the options
from the disk.

So, when are they written to the disk?  In do_interactive_rebase()
(builtin/rebase.c), after using get_replay_opts() to convert `opts' to
`replay', init_basic_state() is called, which calls write_basic_state(),
which write the options to the disk.

Then, until complete_action() is called, `opts' is not modified.

>>  - loads the todo list -- this is unnecessary as we've just populated
>>    the todo list in complete_action()
> 
> Both functions indeed have their own todo lists that they pass to
> pick_commits(), but I don't see (at least, by glancing at the code) that
> both these todo lists are identical.
> 

Near the end of complete_action(), the todo list is written to the disk.
 The destination is obtained with rebase_path_todo().

read_populate_todo() will read a file and parse it.  In the case of
`rebase -i', the location is obtained with rebase_path_todo(), and only
`total_nr' will be modified to contain the number of commands done and todo.

In the case of a new rebase, the done list might not be empty after
tajjimh skip_unnecessary_picks() from complete_action().  Skipped
commands are moved from the todo list to the done list.  As `total_nr'
is not changed by skip_unnecessary_picks(), it is also equal to the
number of commands remaining in the todo list and in the done list.  So,
when read_populate_todo() reads the list and the done list from the
disk, as they should not have been modified, `total_nr' should remain
the same, too.

The only thing that can change is the internal buffer (`buf'), because
skip_unnecessary_picks() don’t change it.  Since
ag/sequencer-reduce-rewriting-todo, it is no longer a textual
representation of the todo list.  Each command contains a pointer to a
location in the buffer and a length to represent its argument.

>>  - commits any staged changes -- this is unnecessary as we're starting a
>>    new rebase, so there are no staged changes
>>  - calls record_in_rewritten() -- this is unnecessary as we're starting
>>    a new rebase.
> 
> OK - I don't know enough about the rebase mechanism to verify these, but
> these seem reasonable to me.
> 

Cheers,
Alban