Re: Modules dependencies RE: PATCH: zftp/tcp abstraction

["Bart Schaefer" <schaefer@xxxxxxxxxxxxxxxxxxxxxxx>]

On Jun 26,  3:26pm, Andrej Borsenkow wrote:
}
} Here is suggested change to current module system. It is actually pretty
} trivial to implement and may have additional advantages ...

Jumping way ahead:

} This does not seem very hard to implement (the ugliest part is parsing of
} zmd files; do we have something ready in zsh so as not to reinvent the
} wheel)? Comments?

I object to any solution which requires that a statically linked binary
rely upon any external files.  I can't tell for certain whether that is
the case with your suggestion, but that's the impression I get.

Your remark about having something read to parse the extra files has
given me an idea:

} - module is represented by its module definition file.
} 
} - when you open a module, zsh opens zmd file. It then loads all needed
} modules (as in moddeps) and after that loads modules object itself (as in
} soname)
} 
} - zmd file is always installed for both linked-in and external modules.
} It eliminates need to build extra list at compile-time.

Why can't the module definition file itself be a loadable module?

For every module that has dependencies, we compile a separate little
shared object that has no dependencies itself.  It contains exactly one
function, which simply calls the module loading code on each of the
dependencies of the base module and then on the base module.  The name
of this object is derived from the name of the base module, e.g. for
zsh/zftp the name might be zsh/ld/zftp (and zsh/net/ld/tcp, etc.).

When the module loading code is invoked, it first inserts the /ld/ and
looks for that module and loads it if necessary, then calls the function
to finish the loading process.  If that fails, the /ld/ module has to be
unloaded again.

Of course directly loading the /ld/ module with `zmodload' would have
the same effect, so you end up with the base module and its dependencies
loaded as well.

If it doesn't find a /ld/ module, then it can go ahead and load the base
module directly.  If the /ld/ module is already loaded, it can assume
that the base module is also loaded.  (Of course we need a different
entry point that doesn't do this check, for the function inside the /ld/
module to call.)

And the /ld/ modules would be linked into the static binary just like
any other module, so this all continues to work there.

I'm not sure about "eliminates need to build extra list at compile-time."
I would think building the list once at compile time was preferable?

There remains the question of where aliasing interacts with this, e.g.,
do aliases get resolved after adding the /ld/ in the module code?  Is
one allowed to do `zmodload -A foo/ld/zap=foo/bar', and if so does the
foo/ld/bar module get loaded upon `zmodload foo/ld/zap'?  Etc.

} - zsh _may_ scan $module_path and autoload objects from those modules with
} load=yes at startup. Or it even can do it when $module_path changes (but see
} *1). This makes it possible to have a single binary and simply drop new
} module in place that may be of some use for packagers.

If this behavior is really desirable, then let's add an /auto/ module for
each base module that needs one.  Loading the /auto/ module installs all
the autoloads for the base module and then unloads itself again.

We have `zmodload -a' still hanging about as a synonym for `zmodload -ab';
it could be overloaded to look for and load the /auto/ module, otherwise
fall back on its old behavior.

If you give zmodload the -a and -u options together, then after unloading
the base module it reloads the corresponding /auto/ module; that addresses
Andrej's (?) earlier question about autoloads disappearing when a module
is unloaded.  (Presently -u -a is the same as -u -b; I suppose the /auto/
module could have two functions, one to add all the autoloads and one to
delete them again.)

} Zsh first adds baz, then it adds bar and gets an error. We possibly
} have to remove baz now (else, it will try to load zsh/mod2 on access
} to baz and this will silently redefine bar).

This problem exists independent of the scheme for manging dependencies.
So far we've required the user to resolve it manually, and I think that
is the only right way to go.
 
} P.S. We can even completely ignore file name and rely on name= in zmd
} file. That gives natural way to directly load any module by its path
} name.

I don't quite see what you mean is "natural" about this, but obviously
this is one point my scheme doesn't address.  I think the other benefits
outweigh this.

-- 
Bart Schaefer                                 Brass Lantern Enterprises
http://www.well.com/user/barts              http://www.brasslantern.com

Zsh: http://www.zsh.org | PHPerl Project: http://phperl.sourceforge.net