Re: CHECK_STRUCTS/dmpstruct.h mechanism is broken.

[Alan Mackenzie]

Hello, Paul.

On Thu, Apr 11, 2019 at 15:23:15 -0700, Paul Eggert wrote:

[ .... ]

> Moreover, Lisp file recompilation is parallelized so it's reasonably
> fast on larger platforms (the wave of the future :-), whereas
> 'configure' will remain slow.

Why?  Because we're not fixing it.

There's no reason why `configure' needs to run on a single CPU core, at
least, not that I can see at the moment.

The hero will be the hacker who contributes a -j parameter to configure
in the autoconf project.  Gentoo users would be grateful for ever.

As a stopgap, would it perhaps be possible to feed `configure''s tests
through GNU Parallel, somehow?  I see `configure' being fed through a
sed or AWK script to make it do this.  Perhaps.

> I just now timed it, and './configure' took 29 real-time seconds
> whereas 'cd lisp; make -j16 compile-always' took 38 real-time seconds;
> this was with two circa-2013 Xeon E5-2640 v2 CPUs with 8 cores each.
> So on a 32-core platform, I expect ./configure to take longer than
> byte-compiling all the Lisp files - i.e., 'configure' is ridiculously
> slow.

For the record, the slowness of `configure' irritates me, too.

-- 
Alan Mackenzie (Nuremberg, Germany).