Re: reducing the footprint of ScanKeyword (was Re: Large writable variables)

On Fri, Jan 04, 2019 at 02:36:15PM -0800, Andres Freund wrote:
> Hi,
> 
> On 2019-01-04 16:43:39 -0500, Tom Lane wrote:
> > Joerg Sonnenberger <joerg@bec.de> writes:
> > >> * What's the generator written in?  (if the answer's not "Perl", wedging
> > >> it into our build processes might be painful)
> >
> > > Plain C, nothing really fancy in it.
> >
> > That's actually a bigger problem than you might think, because it
> > doesn't fit in very nicely in a cross-compiling build: we might not
> > have any C compiler at hand that generates programs that can execute
> > on the build machine.  That's why we prefer Perl for tools that need
> > to execute during the build.  However, if the code is pretty small
> > and fast, maybe translating it to Perl is feasible.  Or perhaps
> > we could add sufficient autoconfiscation infrastructure to identify
> > a native C compiler.  It's not very likely that there isn't one,
> > but it is possible that nothing we learned about the configured
> > target compiler would apply to it :-(

There is a pre-made autoconf macro for doing the basic glue for
CC_FOR_BUILD, it's been used by various projects already including libXt
and friends. 

> I think it might be ok if we included the output of the generator in the
> buildtree? Not being able to add keywords while cross-compiling sounds like
> an acceptable restriction to me.  I assume we'd likely grow further users
> of such a generator over time, and some of the input lists might be big
> enough that we'd not want to force it to be recomputed on every machine.

This is quite reasonable as well. I wouldn't worry about the size of the
input list at all. Processing the Webster dictionary needs something
less than 0.4s on my laptop for 235k entries.

Joerg