Daniel Sahlberg wrote:
I've been trying to dig into shelving and bug 4827. I've been debugging
the regression test from Julian. Please excuse me if I'm making up terms
for things called something else, I'm still new at this.
Thank you for digging into this. Sorry for being slow to respond.
As I understand it, Subversion is:
1. Checking a new WC for the shelf
(.svn/experimental/shelves/v3/<encoded name>-version.wc)
2. Applying the changes from the "normal" WC to the shelf-WC using the
delta editor
The file to be shelved is:
$Rev: 1$\n
New line\n
(where the actual change is the new line - I like Julian's
self-explanatory test code!)
The base file is :
$Rev$\n
In step 1 keyword expansion occurs so the file in the "shelf-WC":
$Rev: 1$\n
The problem seems to be that the patching in step 2 is putting the
changes in the position relative to the base file and not considering
the keyword expansion. So the file in the "shelf-WC" after patching is
$Rev: New line\n
This file matches the "actual" checksum.
The "expected" checksum is based on this file:
$Rev$\n
New line\n
(which doesn't exist except in my mail).
I guess there are two ways forward:
* To prevent keyword expansion in the "shelf-WC"
That's the better option. Details follow.
* To make sure the delta editor applies the patch relative to the
expanded keyword (in this case three bytes later in the file).
Before I try to make a patch out of this I would like to ask which way
would be prefered. It the "shelf-WC" supposed to be just like any other
WC (in that case keyword expanded), or should it receive special
treatment to work better with the delta editor?
The shelf-WC should not do keyword expansion and EOL-style conversion.
(EOL-style conversion is the other kind of non-committable local change
that goes side-by-side with keyword expansion.) The design of
shelving-v3 is that only committable changes, which means changes that
are representable in the "editor API", are to be stored in a shelf [1].
The shelf-WC has no purpose other than for the shelving code to store
changes to it and extract the same changes from it. When a change is
shelved from the main WC, keywords should be unexpanded and EOL-style
normalized to "repository form". When a change is replayed from a
shelf-WC back to the main WC, any required keyword expansion and
EOL-style conversion should be applied to the main WC at that point.
If I remember correctly, the interface for getting changes out of the
main WC used the commit editor, so that must be returning keywords and
EOL style to repository form already. The two areas that need to be
checked and fixed are (1) the "checkout" code (which you have observed
is expanding keywords), which, from memory, I think probably has an
option that can be passed in the client "config" to suppress that
functionality; and (2) the code being used to apply the patch (coming
from the main WC) to the shelf WC, which, from memory, is code that I
had to write and might not have such an option and might have
inconsistent behaviour.
If you can do anything towards making it work as intended, that would be
great.
Don't hesitate to follow up with questions here or on the #svn-dev
channel (which I access these days through Matrix [2], and is still also
accessible in Freenode IRC). I forgot I had marked your message for
attention this time. I usually would respond much quicker.
Thanks,
- Julian
[1] If there is ever a future enhancement to this design to enable
shelving WC-only kinds of local changes, I envisage it would be done by
adding another layer of "editor API" on top of the existing one, that
has the job of supporting WC-only kinds of local changes, which could
conceivably include keyword expansion and EOL-style.
[2] #svn-dev via Matrix: https://matrix.to/#/#freenode_#svn-dev:matrix.org
