Re: [gentoo-dev] RFD: EAPI specification in ebuilds

[Alec Warner]

On Wed, Mar 7, 2012 at 12:41 PM, Ulrich Mueller <ulm@g.o> wrote:
> Hi all,
>
> The way how we currently specify the EAPI in ebuilds has some
> problems. For example, there is no sane way to allow usage of features
> of a new bash version in a new EAPI. So we are currently stuck with
> bash 3.2. Also changes of global scope behaviour, like addition of new
> global scope functions (similar to "inherit") are not possible.
>
> These flaws are outlined in GLEP 55 [1]:
> | In order to get the EAPI the package manager needs to source the
> | ebuild, which itself needs the EAPI in the first place. Otherwise it
> | imposes a serious limitation, namely every ebuild, using any of the
> | future EAPIs, will have to be source'able by old package managers
> | [...]
>
> The council has voted down GLEP 55 more than a year ago, but at the
> same time requested that a solution for the mentioned issues should be
> found. [2] However, there was no progress since then.
>
> The issue arose again in bug 402167 [3] where several solutions have
> been discussed. Below, I try to summarise the possible options
> resulting from that discussion.
>
>
> *** Proposal 1: "Parse the EAPI assignment statement" ***
>
> This first proposal would require that the syntax of the EAPI
> assignment statement in ebuilds matches a well defined regular
> expression. A scan of the Portage tree shows that the statement only
> occurs in the following variations (using EAPI 4 as example):
>
>   EAPI=4
>   EAPI="4"
>   EAPI='4'
>
> Sometimes this is followed by whitespace or a comment (starting with
> a # sign). Also, with very few exceptions the EAPI assignment occurs
> within the first few lines of the ebuild. For the vast majority of
> ebuilds it is in line 5.
>
> Written in a more formal way, appropriate for a specification:
> - Ebuilds must contain at most one EAPI assignment statement.
> - It must occur within the first N lines of the ebuild (N=10 and N=30
>  have been suggested).
> - The statement must match the following regular expression (extended
>  regexp syntax):
>  ^[ \t]*EAPI=(['"]?)([A-Za-z0-9._+-]*)\1[ \t]*(#.*)?$
>
> Note: The first and the third point are already fulfilled by all
> ebuilds in the Portage tree. The second point will require very few
> ebuilds to be changed (9 packages for N=10, or 2 packages for N=30).
>
> The package manager would determine the EAPI by parsing the assignment
> with above regular expression. A sanity check would be added. Citing
> Zac Medico in [3]: "The fact that we can compare the probed EAPI to
> the actual EAPI variable after the ebuild is sourced seems like a
> perfect sanity check. We could easily detect inconsistencies and flag
> such ebuilds as invalid, providing a reliable feedback mechanism to
> ebuild developers."
>
> This proposal comes in two variants:
> 1a) The change is applied retroactively for all EAPIs.
> 1b) It is only applied for EAPI 5 and later (which means that the
>    result of the EAPI parsing would be discarded for earlier EAPIs).

I don't like this idea because the sane way should be easy and straightforward.
Mixing a constant declaration with bash assignment just confuses users
who think the assignment is full bash when in fact it is not.

EAPI=$(somefunc)
EAPI=${SOMEVAR%%-*}
and so forth all don't meet the regex (and would be flagged invalid.)
However a naive author might think they work.

I don't think any naive author would think either would work in a comment

# EAPI=$(somefunc)
# EAPI=${SOEMVAR%%-*}

Bash doesn't parse comments, so confusion is unlikely.

>
>
> *** Proposal 2: "EAPI in header comment" ***
>
> A different approach would be to specify the EAPI in a specially
> formatted comment in the ebuild's header. No syntax has been suggested
> yet, but I believe that the following would work as a specification:
> - The EAPI must be declared in a special comment in the first line of
>  the ebuild's header, as follows:
> - The first line of the ebuild must contain the word "ebuild",
>  followed by whitespace, followed by the EAPI, followed by
>  end-of-line or whitespace.
>
> Again, the proposal comes in two variants:
> 2a) It is combined with a one time change of the file extension, like
>    .ebuild -> .eb.
> 2b) The usual EAPI assignment statement in the ebuild is still
>    required, at least for a transition period.
>
> In the 2a case, the EAPI variable could be made read-only in bash
> before sourcing the ebuild. In the 2b case, a sanity check similar to
> the one mentioned above would be added.
>
>
> What do you think?

Overloading is bad.

There is no real difference between:
#!/usr/bin/ebuild --eapi 5
# EAPI=5

The problem is that the former is also the way to specify how to
execute the ebuild; so unless you plan to make ebuilds executable and
having /usr/bin/ebuild provide the ebuild environment, using that
syntax is confusing to users.

>
> (I really hope for a constructive discussion here. So, if you want
> to comment that all of the above proposals suck and GLEP 55 is much
> superior, then please open a new thread for it.)

You don't mention voting on glep 55 again; is there a reason why not?

>
> Ulrich
>
>
> [1] <http://www.gentoo.org/proj/en/glep/glep-0055.html#problem>
> [2] <http://www.gentoo.org/proj/en/council/meeting-logs/20100823-summary.txt>
> [3] <https://bugs.gentoo.org/show_bug.cgi?id=402167>
>