Re: Moving away from make

[Han-Wen Nienhuys]

On Sat, Sep 24, 2011 at 10:05 PM, Graham Percival
<address@hidden> wrote:
>> > If I was writing a "make" system from scratch, I would describe
>> > dependencies in data structures that are viewable and editable, and
>> > have a separate program that uses those structures to determine which
>> > files need making.
>>
>> I'm not sure why you need extra structures for that? For C/C++, gcc can
>> figure out the dependencies as a side effect of compiling the normal
>> source code; and this can be done for most other languages as well. Very
>> few non-trivial programs actually maintain their dependencies by hand
>> nowadays...
>
> That's a great comfort for our lilypond texinfo documentation,
> which *does* require explicit dependencies.
>
> As far as I know, every single build system has trivially easy
> support for C/C++ compilation.  That's not an issue.  The issue is
> building documentation, building swig python modules (for me
> personally, not lilypond), generating translations, generating
> test cases and automatically showing potential problems, etc.
>
> According to
>  find . -name "*.make" -or -name "GNUmakefile*" | xargs cat | wc -l
> we have 5059 lines of build-system stuff.  Actually, that's not
> counting python helper scripts:
>  wc scripts/build/*.py -l
> 4113 total
>
> So... 9171 lines of build-system junk.  I think this meme is
> appropriate:
>  It's over NINE THOUSAND!!!
>
> This didn't happen for fun and giggles.  Over the years, we've
> built up hack upon hack, and ended up with this unholy mess.  Is
> that purely the fault of autotools?  No.  Could it be *worse* if
> we used a different build system?  Sure.  But it could also be
> *better* if we used a different build system.

You sound spoiled.

* The build scripts are not really part of the complexity of the build
system.  They are normal programs, they are just not intended to be
run by end users.  I assume you are not considering a rewrite of
output-distance.py in CMake's scripting language.

* You make 5000 lines of makefile sound as if it is a lot.  I think
you should look at other packages' build systems more closely.

For example, coreutils, which is 70k lines of C, and a trivial
documtentation and test process has, count'dem: 2000 lines of
Makefile.am.

This is for just linking together a bunch of bog-standard C object
files, generating an info page, and running some shell scripts as
tests.

I think 5k lines is pretty impressive for a package as complex as
LilyPond, considering that it has a bizarre documentation process,
translated docs, python bindings, fonts that are run through metapost
and fontforge, a complex regression test, and needs weird acrobatics
to find its gazilion different init files.

-- 
Han-Wen Nienhuys - address@hidden - http://www.xs4all.nl/~hanwen