Re: CHECK_STRUCTS/dmpstruct.h mechanism is broken.

[Daniel Colascione]

> Paul Eggert writes:
>> On 4/11/19 2:35 AM, Robert Pluim wrote:
>>> Much as I dislike Autoconf, what would you have us replace it with?
>>
>> I was thinking of using just standard tools (as per the GNU Coding
>> Standards) along with GNU Make - and, once the Emacs core is built, we
>> can use Emacs itself. Although we started assuming GNU Make in Emacs 25,
>> we haven't been using GNU Make's features fully and some of its features
>> could effectively replace the need for Autoconf.
>
> I don't see how… autoconf got one thing right: it tries to actually
> _compile_, not just check preprocessor defines or compiler versions.
> That also makes it slow, especially as it's all serial.
>
>> A benefit of this approach would be faster builds. Right now the biggest
>> bottleneck on my system is the time to run 'configure' whenever I make a
>> trivial change to configure.ac or whatever. I *hate* that.
>
> The fun thing is that in this case you could run autoconf almost
> entirely from already cached decisions.  I have not tried to feed
> autoconf prepared caches in a long time, but it may be worth a try.
>
>> Although I looked into other possible approaches (switching to SCons,
>> say) none of them seemed to offer compelling advantages to the
>> more-conservative approach I have in mind.
>
> Most of the newfangled systems I've looked at focus on build speed
> (mostly by pre-computing dependencies) and take the configuration for
> granted (some even using autoconf again).
>
>> Eric Raymond reported success with this sort of approach when
>> deautoconfiscating giflib and the NTP code base:
>>
>> https://lists.gnu.org/r/bison-patches/2019-02/msg00041.html
>
> I can tell you something about NTPsec.  Yes, it has no autoconf, but
> that's been replaced with the configure step from waf, which has exactly
> one person you can ask if something doesn't work the way you think it
> should.  Based on Python it has a somewhat more agreeable syntax than M4
> macros (no surprise), but it's not actually more clear (to me anyway).
> Plus, there are a lot less depencies that need to be checked.

I like autoconf. It's flexible and it works for every weird thing I've
ever wanted to do with it. Trying to color outside the lines with
alternate systems like cmake has been hell. While m4 isn't the most
elegant macro system out there, its quirks are well-understood, and
autoconf has decades of institutional knowledge built into it. I've also
found that alternative systems frequently have problems with complex
configurations like out-of-tree builds, cross compilation, various forms
of path overriding, and tool specification. autoconf dealt with all of
these issues a very long time ago.

IMHO, instead of investing in some new system, we'd be better off
improving autoconf speed by

1) eliminating tests that have become unnecessary in the modern world,
2) adding an optimization pass that batched redundant tests, and
3) making caching Just Work, maybe by stracing and checking all inputs for
cache validation.

The basic model is fine. The basic configuration model is fine.