Re: Separating generated files? (Re: Mercurial -> git)

Re: Separating generated files? (Re: Mercurial -> git)

[Nguyen Thai Ngoc Duy]

Hi Christian,

The idea of using two separate repositories for source and generated
source is interesting. I would like to bring this to git mailing list,
they may provide insightul comments for your idea or even other
approaches.

Background for Git people: gambit-c was previously stored in
mercurial. The main source is in gambit-c (a Scheme implementation or
a Lisp dialect). *.scm files generate *.c, which will be compiled by
gcc as usual. Both *.scm and generated *.c are now stored in
mercurial. Gambit-C maintainers have recently decided to move to Git.

On 10/15/08, Christian Jaeger <christian@pflanze.mine.nu> wrote:
> I wonder whether it would be a good idea, and good occasion to realize
>  it, to move source files and generated files into separate repositories
>  and 'link' those together using the git submodule feature.
>
>  Expected advantages:
>
>  - no clutter when looking through the history (can possibly be mitigated
>  by constraining git log, git diff etc. to the non-generated paths only,
>  although I don't think this is possible (cleanly) with the current
>  directory structure); the same holds true for using "git format-patch"
>  (one wouldn't usually want to include the generated files in diffs sent
>  to the mailing list)
>
>  - when merging branches, there will usually be no need to deal with
>  merge conflicts in the generated files (one would just regenerate them
>  instead)
>
>  - [especially for files being generated not by Gambit itself (for
>  example "configure"),] the files can be regenerated by differing
>  [external] software versions without having to deal with those
>  superfluous changes in the source repository.
>
>  By still committing the generated files--to a different
>  submodule--Gambit can still be updated through Git alone, and the
>  possible advantage of tracking the generated files to see the effects of
>  changes in compiler sources can still be had.
>
>  Expected disadvantages:
>
>  - all generated files need to reside in a separate directory structure;
>  e.g. the file $BASEDIR/lib/_io.c would have to be at a place like
>  $BASEDIR/build/lib/_io.c instead, where build/ is the submodule taking
>  all generated files; since the "configure" file is expected to reside at
>  the toplevel, I guess this would require that "make update" copies it
>  from $BASEDIR/build/configure to $BASEDIR/configure (assuming that one
>  cannot use a symlink because of portability reasons).
>
>  - to commit the generated files, a separate step is necessary ("cd
>  $otherrepo; git commit -a", or maybe easier create a "make
>  commit_generated" make target?)
>
>  - to make this work with the "source" repository residing at the
>  toplevel, the Git superproject repository (of which the "source" and
>  "build" repositories are submodules) would need to reside in a
>  non-standard directory, like $BASEDIR/.gitsuperproject/ instead of the
>  usual .git/, and using the GIT_DIR environment variable to access it,
>  although this can probably be handled by make targets (i.e. "make
>  update" would set GIT_DIR=$BASEDIR/.gitsuperproject when calling "git
>  submodule update").
>
>  - there may be some cases to flesh out; like, should "make update"
>  really call "git submodule update" (which simply sets the submodules to
>  the reference given by the superproject, throwing away changes done by
>  the user in the submodules (they can be recovered from the git reflog,
>  but may still be a surprise)) or should it run "git pull" in each
>  submodule instead?
>
>  I thought I'd bring this up now because if package maintainers need to
>  adapt some things anyway, that may be a good time to do it now. (There's
>  even the possibility to split the converted Mercurial repository into
>  the source + build parts in retrospect now, which won't be possible
>  anymore later on (without changing the sha1 sums of the whole Git
>  history with the associated breakage of existing clones), although that
>  may not be important.)
>
>  I'm willing to help in the effort, although I don't know the build tools
>  (autoconf and make) and their use in the setup well, so I would probably
>  be quite a bit lost when doing it alone.
>
>  Christian.
-- 
Duy

Re: [gambit-list] Separating generated files? (Re: Mercurial -> git)

[Matthieu Moy]

"Nguyen Thai Ngoc Duy" <pclouds@gmail.com> writes:

> Hi Christian,
>
> The idea of using two separate repositories for source and generated
> source is interesting. I would like to bring this to git mailing list,
> they may provide insightul comments for your idea or even other
> approaches.

I think the first question is: do you (and why) need to use a version
control system for generated files?

-- 
Matthieu

Re: [gambit-list] Separating generated files? (Re: Mercurial -> git)

[Michael J Gruber]

Matthieu Moy venit, vidit, dixit 15.10.2008 17:30:
> "Nguyen Thai Ngoc Duy" <pclouds@gmail.com> writes:
> 
>> Hi Christian,
>>
>> The idea of using two separate repositories for source and generated
>> source is interesting. I would like to bring this to git mailing list,
>> they may provide insightul comments for your idea or even other
>> approaches.
> 
> I think the first question is: do you (and why) need to use a version
> control system for generated files?

I guess we can take "yes" for granted for the first part ;)
As for the why: In cases like this one it's interesting to compare
(read: diff) the output generated by different versions of the input.

I wonder whether a clever use of "excludes" and GIT_DIR would allow
tracking the different filesets in the same dir, but using different
repos. I'm just afraid it's a fragile setup, in the sense that it relies
on config stuff which is not tracked (and thus not reproduced
automatically on clone).

Michael

Re: Separating generated files?

[Christian Jaeger]

Matthieu Moy wrote:
> I think the first question is: do you (and why) need to use a version
> control system for generated files?

The project in question is a self-hosting compiler which compiles to C 
as an intermediary language. Providing the generated C files to users 
makes installation easy (it avoids the bootstrapping issue). So it's 
more 'severe' of an issue than just one of for example generating 
documentation files using a 3rd-party tool.

What may make matters worse, is that there are interdependencies between 
a number of hand-written C files and the generated files, so it's not 
always possible to use an older compiler version to reproduce the 
generated C files for a newer compiler; so if you want to merge newer 
compiler sources, you may also need the generated files, at least if you 
want that without fuss. So, there is always a need to somehow transmit 
the generated files too. I guess that this is easier than code the 
system in a way to always allow backwards compatibility (I haven't 
worked on the compiler itself yet, so this is a guess and may need 
confirmation).

Apart from that, I've found it useful (in another project, writing a 
document translator) to keep generated files in a VCS (Git) as well (I 
checked them into the *same* repository as the translator source, even 
if it felt ugly (for the previously mentioned reasons)), as then when I 
changed the translator, I could easily see where it had effect on the 
generated output. It can even serve as a debugging help kind of like a 
test suite does. This may be the case here, too (again, I'm guessing here).

How are other compiler projects which are bootstrapping via C dealing 
with this?

Christian.

Re: [gambit-list] Separating generated files? (Re: Mercurial -> git)

[Christian Jaeger]

Michael J Gruber wrote:
> I wonder whether a clever use of "excludes" and GIT_DIR would allow
> tracking the different filesets in the same dir, but using different
> repos. I'm just afraid it's a fragile setup, in the sense that it relies
> on config stuff which is not tracked (and thus not reproduced
> automatically on clone).

I expect that using a superproject repository to tie together the two 
repositories is good and necessary because it is the link that allows to 
specify which commit in the repo of generated files belongs together 
with a commit in the repo of source files. So just using two separate 
repositories without making them submodules of a superproject does not 
seem to be a good idea to me.

Once there is a superproject repository, one could also commit config 
files of the submodules into it (I'm not sure what that will buy 
though--.gitignore is outside and can committed anyway, at least as long 
as not both repositories are overlaid as you suggest).

You're probably right that strictly speaking, there is no need to move 
generated files out into a separate directory tree; but I think doing 
the move would be worthwhile since it takes away one level of complexity 
(you can then access the build/.git repository without the need of 
setting GIT_DIR), and because it may be a good idea anyway (for example 
it will be easier to grep the sources without getting hits from the 
generated files). [Also, the exclude patterns wouldn't be easy, as we 
couldn't really just exclude all *.c files from the view of the source 
repository, since there are also some hand-crafted ones; the excludes 
would need full paths which would have to be kept up to date manually, 
unless we wanted to live with the fact that newly created manual .c 
files would be added using "git add -f".]

Christian.

Re: [gambit-list] Separating generated files?

[Christian Jaeger]

I wrote:
> Michael J Gruber wrote:
>> I wonder whether a clever use of "excludes" and GIT_DIR would allow
>> tracking the different filesets in the same dir, but using different
>> repos. I'm just afraid it's a fragile setup, in the sense that it relies
>> on config stuff which is not tracked (and thus not reproduced
>> automatically on clone).
>
> I expect that using a superproject repository to tie together the two 
> repositories is good and necessary because it is the link that allows 
> to specify which commit in the repo of generated files belongs 
> together with a commit in the repo of source files. So just using two 
> separate repositories without making them submodules of a superproject 
> does not seem to be a good idea to me.

(In the meantime I've read the following pages:
http://nopugs.com/2008/09/06/ext-tutorial
http://nopugs.com/2008/09/04/why-ext
http://flavoriffic.blogspot.com/2008/05/managing-git-submodules-with-gitrake.html
(a post to the latter article suggests to use subtree merging instead, 
but that would be a very bad match for our use case; the mentioned 
problem of merging of the git superproject makes me think, though--the 
superproject could be updated only by the one person doing the publish 
onto the public repository, but then it leaves the problem of handling 
merges by developers completely unsolved.)
)

I'm starting to think that maybe a better idea than the 
superproject+2submodules approach would be just using the two 
repositories ("source" + "build"), and storing the linking information 
inside the "build" repository (by adding the source repository commitid 
to every commit message in the build repository [or using tags, but that 
doesn't seem a better idea]), and use a program that is able to check 
out the matching "build" repository for a given "source" repository 
checkout.

I'm willing to write this program (let's call it 
"intergit-find-matching-commit-in" for the purpose of this email); 
question: which language to write it in, is Perl good? (C would be a 
hassle for Windows users because of the C compiler requirement; shell 
may be too limited.)

Description of the workings in more detail:

- one would work with the "source" repository just as one would with any 
project only employing one repository; do some changes to the project, 
commit them, test them (includes regeneration of generated files);

- once in a while one would commit the current generated files in the 
"build" repository; by either (a) using a make target (like "make 
commit_generated") which runs something like

  eval "cd build; git commit -m 'generated files for source repository 
commit `git rev-parse HEAD`'"

or (b) setting up a build/.git/hooks/commit-msg script which appends 
'generated files for source repository commit `git rev-parse HEAD`' line 
to the commit message given from running "cd build; git commit -a" manually.

- for publication, one would push both the "source" as well as the 
"build" repository (i.e. "cd build; git push; cd ..; git push")

- for checkout (our "make update" make target), about the following 
would happen:

   git pull
   eval "(cd build; git checkout `intergit-find-matching-commit-in build`)"

where "intergit-find-matching-commit-in build" would first refresh an 
index of the links (iterate over all unseen commits, parse commit 
messages for /source repository commit (\w+)/ and store $1 => 
$commitid_in_build_repo mappings in the index), then go through "git log 
--pretty=format:%H" (should I also specify --topo-order (or 
--date-order)?) looking up the commitids in the index, stopping at the 
first match and outputting the mapped $commitid_in_build_repo.

This way, the "latest" or "probably best-matching" corresponding commit 
in the "build" repo can always be found, even if the "source" repo is 
ahead, which should allow building the compiler even if none is 
previously installed. This workflow seems more natural than the 
superproject+submodules approach, and it seems to entail no hassle with 
merge issues (only the "source" repo really needs proper merging; 
merging the "build" repo would only be worthwhile for maintaining the 
history, and as mentioned if there are conflicts, one would probably 
usually just regenerate the files there; there's no need to maintain 
linking info (with associated merge etc issues) in a separate entity 
(superproject) anymore, and during development, commits to the "build" 
repo need only be done if backwards-incompatible changes have been 
introduced).

Does anyone else think this is sane/interesting? Should I go ahead 
implementing this? Any comments, like on how the interface of the 
intergit-find-matching-commit-in tool should look like?

Christian.

Re: [gambit-list] Separating generated files?

[Santi Béjar]

On Thu, Oct 16, 2008 at 2:00 PM, Christian Jaeger
<christian@pflanze.mine.nu> wrote:
>
> (In the meantime I've read the following pages:
> http://nopugs.com/2008/09/06/ext-tutorial
> http://nopugs.com/2008/09/04/why-ext
> http://flavoriffic.blogspot.com/2008/05/managing-git-submodules-with-gitrake.html
> (a post to the latter article suggests to use subtree merging instead, but
> that would be a very bad match for our use case; the mentioned problem of
> merging of the git superproject makes me think, though--the superproject
> could be updated only by the one person doing the publish onto the public
> repository, but then it leaves the problem of handling merges by developers
> completely unsolved.)
> )
>
> I'm starting to think that maybe a better idea than the
> superproject+2submodules approach would be just using the two repositories
> ("source" + "build"), and storing the linking information inside the "build"
> repository (by adding the source repository commitid to every commit message
> in the build repository [or using tags, but that doesn't seem a better
> idea]), and use a program that is able to check out the matching "build"
> repository for a given "source" repository checkout.
>
> I'm willing to write this program (let's call it
> "intergit-find-matching-commit-in" for the purpose of this email); question:
> which language to write it in, is Perl good? (C would be a hassle for
> Windows users because of the C compiler requirement; shell may be too
> limited.)
>
> Description of the workings in more detail:
>
> - one would work with the "source" repository just as one would with any
> project only employing one repository; do some changes to the project,
> commit them, test them (includes regeneration of generated files);
>
> - once in a while one would commit the current generated files in the
> "build" repository; by either (a) using a make target (like "make
> commit_generated") which runs something like
>
>  eval "cd build; git commit -m 'generated files for source repository commit
> `git rev-parse HEAD`'"
>
> or (b) setting up a build/.git/hooks/commit-msg script which appends
> 'generated files for source repository commit `git rev-parse HEAD`' line to
> the commit message given from running "cd build; git commit -a" manually.
>
> - for publication, one would push both the "source" as well as the "build"
> repository (i.e. "cd build; git push; cd ..; git push")
>
> - for checkout (our "make update" make target), about the following would
> happen:
>
>  git pull
>  eval "(cd build; git checkout `intergit-find-matching-commit-in build`)"
>
> where "intergit-find-matching-commit-in build" would first refresh an index
> of the links (iterate over all unseen commits, parse commit messages for
> /source repository commit (\w+)/ and store $1 => $commitid_in_build_repo
> mappings in the index), then go through "git log --pretty=format:%H" (should
> I also specify --topo-order (or --date-order)?) looking up the commitids in
> the index, stopping at the first match and outputting the mapped
> $commitid_in_build_repo.
>
> This way, the "latest" or "probably best-matching" corresponding commit in
> the "build" repo can always be found, even if the "source" repo is ahead,
> which should allow building the compiler even if none is previously
> installed. This workflow seems more natural than the superproject+submodules
> approach, and it seems to entail no hassle with merge issues (only the
> "source" repo really needs proper merging; merging the "build" repo would
> only be worthwhile for maintaining the history, and as mentioned if there
> are conflicts, one would probably usually just regenerate the files there;
> there's no need to maintain linking info (with associated merge etc issues)
> in a separate entity (superproject) anymore, and during development, commits
> to the "build" repo need only be done if backwards-incompatible changes have
> been introduced).
>
> Does anyone else think this is sane/interesting? Should I go ahead
> implementing this? Any comments, like on how the interface of the
> intergit-find-matching-commit-in tool should look like?
>

It looks like the html and man branches of git.git.

http://git.kernel.org/?p=git/git.git;a=shortlog;h=html
http://git.kernel.org/?p=git/git.git;a=shortlog;h=man

They are automatically generated when Junio pushes the branches to
kernel.org. Afterwards you can do a "make quick-install-html" and
install the preformated html pages from these branches. They are
generated with the dodoc.sh script from the todo branch in git.git
(look inside for instructions):

http://git.kernel.org/?p=git/git.git;a=blob_plain;f=dodoc.sh;hb=todo

HTH,
Santi

Re: [gambit-list] Separating generated files?

[Christian Jaeger]

Santi Béjar wrote:
> It looks like the html and man branches of git.git.
>
> http://git.kernel.org/?p=git/git.git;a=shortlog;h=html
> http://git.kernel.org/?p=git/git.git;a=shortlog;h=man
>
> They are automatically generated when Junio pushes the branches to
> kernel.org. Afterwards you can do a "make quick-install-html" and
> install the preformated html pages from these branches. They are
> generated with the dodoc.sh script from the todo branch in git.git
> (look inside for instructions):
>
> http://git.kernel.org/?p=git/git.git;a=blob_plain;f=dodoc.sh;hb=todo

This script only generates the html / man branches, it doesn't help find 
the right version for a given git version, right?

The differences are:

- the html / man branches have a strictly linear history and are 
centrally maintained. This solves the distribution issue for end users. 
But while developping the compiler, the developers may need to go back 
in the history of their own development (e.g. when the current compiler 
doesn't work anymore), and the suspected usefulness of being able to see 
and track differences in the generated code also isn't available for a 
strictly central approach.

- the script above is only for creating and committing the derived 
files, in a hook similar to the one I suggested in 
build/.git/hooks/commit-msg; this is the "cd build; git commit -m 
'generated files for source repository commit
 > `git rev-parse HEAD`'" part; the more interesting part comes from 
automatically finding the right commit in the generated branches for a 
given source commit. This is what I intend to solve with the 
"intergit-find-matching-commit-in" script. Said in a simpler way: the 
git html / man branches do not offer automatically resolvable linking.

Christian.

Re: [gambit-list] Separating generated files?

[Santi Béjar]

On Thu, Oct 16, 2008 at 2:32 PM, Christian Jaeger
<christian@pflanze.mine.nu> wrote:
> Santi Béjar wrote:
>>
>> It looks like the html and man branches of git.git.
>>
>> http://git.kernel.org/?p=git/git.git;a=shortlog;h=html
>> http://git.kernel.org/?p=git/git.git;a=shortlog;h=man
>>
>> They are automatically generated when Junio pushes the branches to
>> kernel.org. Afterwards you can do a "make quick-install-html" and
>> install the preformated html pages from these branches. They are
>> generated with the dodoc.sh script from the todo branch in git.git
>> (look inside for instructions):
>>
>> http://git.kernel.org/?p=git/git.git;a=blob_plain;f=dodoc.sh;hb=todo
>
> This script only generates the html / man branches, it doesn't help find the
> right version for a given git version, right?

Right, one script to generate and one to get the right version.

>
> The differences are:
>
> - the html / man branches have a strictly linear history

Yes, because in this case it is not needed to replicate the whole
history, but it could be improved.

> and are centrally maintained. This solves the distribution issue for end users. But while
> developping the compiler, the developers may need to go back in the history
> of their own development (e.g. when the current compiler doesn't work
> anymore), and the suspected usefulness of being able to see and track
> differences in the generated code also isn't available for a strictly
> central approach.

So, you can divide the problem in two: (a) generated files in the
remote repositories (these can be generated automatically on the
server or in a dedicated server) (b) local generated files for local
commits. If both follow the same format to specify the original commit
you can use the same script to get it.

>
> - the script above is only for creating and committing the derived files, in
> a hook similar to the one I suggested in build/.git/hooks/commit-msg; this
> is the "cd build; git commit -m 'generated files for source repository
> commit
>> `git rev-parse HEAD`'" part; the more interesting part comes from
>> automatically finding the right commit in the generated branches for a given
>> source commit. This is what I intend to solve with the
>> "intergit-find-matching-commit-in" script. Said in a simpler way: the git
>> html / man branches do not offer automatically resolvable linking.

They offer this (Autogenerated HTML docs for v1.6.0.2-530-g67faa) but
there is no script around it.

My point was that there are other project keeping generated files (and
sometimes I would like it too), so you can see what they are doing. At
the end, maybe, you system could be usefull for them also.

Santi