Re: bug report on update-index --assume-unchanged

From: Junio C Hamano <gitster@pobox.com>
Date: Dec 5, 2014, 21:19 UTC
Message-ID: <xmqqbnnhpzgb.fsf@gitster.dls.corp.google.com>
In-Reply-To: <xmqq1toeq79b.fsf@gitster.dls.corp.google.com>

Junio C Hamano <gitster@pobox.com> writes:

> If however you did something that made Git notice that you changed
> these paths marked with assume-unchanged bit anyway, then Git will,
> well, notice that they are not "unchanged" as you promised.

By the way, this cuts both ways. I would not bother checking with the current codebase, but I know it used to be that when we have a blob object name and need a temporary file that holds the content for read-only purpose (e.g. passing a pair of files to external diff driver), we allowed Git to reuse files in the working tree whose blob object name we knew. This is of course because it is faster than inflating the blob contents out of the object store and writing a new file.

That codepath is allowed to "borrow" the working tree file as such a temporary file for read-only purpose, when a file is stat-clean (i.e. its contents is known to match the blob object recorded in the index). A file with assume-unchanged bit set is treated exactly the same way, because the user promised not to modify it.

If the user broke the promise, then an external diff driver would have been given a file whose contents does not actually match the wanted blob object, resulting in an incorrect diff output.