Re: [PATCH/RFC] commit: new option to abort -a something is already staged

[Junio C Hamano <gitster@pobox.com>]

Nguyễn Thái Ngọc Duy  <pclouds@gmail.com> writes:

> +commit.preciousDirtyIndex::
> +	If some changes are partially staged in the index (i.e.
> +	"git commit -a" and "git commit" commit different changes), reject
> +	"git commit -a".

Or perhaps better yet, go into yes/no interaction to confirm if you
have access to interactive terminal at fd #0/#1, perhaps?

> diff --git a/builtin/commit.c b/builtin/commit.c
> index 213fca2d8e..489e4b9f50 100644
> --- a/builtin/commit.c
> +++ b/builtin/commit.c
> @@ -98,6 +98,7 @@ static const char *author_message, *author_message_buffer;
>  static char *edit_message, *use_message;
>  static char *fixup_message, *squash_message;
>  static int all, also, interactive, patch_interactive, only, amend, signoff;
> +static int allow_dirty_index = 1;
>  static int edit_flag = -1; /* unspecified */
>  static int quiet, verbose, no_verify, allow_empty, dry_run, renew_authorship;
>  static int config_commit_verbose = -1; /* unspecified */
> @@ -385,10 +386,24 @@ static const char *prepare_index(int argc, const char **argv, const char *prefix
>  	 * (B) on failure, rollback the real index.
>  	 */
>  	if (all || (also && pathspec.nr)) {
> +		int compare_oid = all && !allow_dirty_index;
> +		struct object_id previous_oid;
> +
> +		if (compare_oid) {
> +			if (update_main_cache_tree(0) || !the_index.cache_tree)
> +				die(_("error building trees"));

Hmph, when we conclude a conflicted merge with "commit -a", wouldn't
we fail to compute the "before" picture, with higher-stage entries
stil in the index?

> +			if (the_index.cache_tree->entry_count >= 0)
> +				oidcpy(&previous_oid, &the_index.cache_tree->oid);
> +			else
> +				oidclr(&previous_oid);

The cache-tree covers no entry, meaning the index has no cache
entries?  Shouldn't we setting EMPTY_TREE_SHA1_BIN or something
instead?

> +		}
>  		hold_locked_index(&index_lock, LOCK_DIE_ON_ERROR);
>  		add_files_to_cache(also ? prefix : NULL, &pathspec, 0);
>  		refresh_cache_or_die(refresh_flags);
>  		update_main_cache_tree(WRITE_TREE_SILENT);
> +		if (compare_oid && the_index.cache_tree &&
> +		    oidcmp(&previous_oid, &the_index.cache_tree->oid))
> +			die(_("staged content is different, aborting"));

I was hoping that it is an easy change to teach add_files_to_cache()
to report how many paths it actually has "added" (modifications and
removals are of course also counted), which allows us not to waste
computing a throw-away tree object once more.

There only are three existing callers to the function, and only one
of them rely on the current "non-zero is error, zero is good"
semantics, so updating that caller (and perhaps vetting the other
callers to see if they also _should_ pay attention to the return
value, at least to detect errors of not number of paths updated)
shouldn't be that much more effort, and would be a good update to
the API regardless of this new feature, I would think.

>  		if (write_locked_index(&the_index, &index_lock, 0))
>  			die(_("unable to write new_index file"));
>  		commit_style = COMMIT_NORMAL;
> @@ -1413,6 +1428,10 @@ static int git_commit_config(const char *k, const char *v, void *cb)
>  		config_commit_verbose = git_config_bool_or_int(k, v, &is_bool);
>  		return 0;
>  	}
> +	if (!strcmp(k, "commit.preciousdirtyindex")) {
> +		allow_dirty_index = !git_config_bool(k, v);
> +		return 0;
> +	}
>  
>  	status = git_gpg_config(k, v, NULL);
>  	if (status)