OK, closing this as indicated in my previous message. Maybe we can reconsider
once the machines we use for experiments get a more recent kernel, but I'm not
holding my breath. ;-)

Anyone willing to work on this? Otherwise I'd be inclined to close it. (Pros of
doing this: more speed. Cons: need to check test the kernel version, so more
complexity in build step, and executables can be transferred less easily between
machines; also: there's some chance we'll switch to -m64 anyway in the
not-so-far future.)

Indeed! :-) Please report back when you have more detailed results!

Ah, shame!  I'm setting up some test infrastructure over here, so I'm
benchmarking -m32 vs -mx32 vs -m64 in more detail, now the new packed-state
representation is merged.  I really don't suspect it'll make a big difference to
coverage, but free speed is free speed.

Thanks for the pointer! Unfortunately, it will be a while until we'll be able to
exploit this, I think. All machines I have access to, including the ones on
which we perform our evaluations, are still on 2.6 series kernels.

If you're:
- using linux kernel >= 3.4
- compiling to 32-bits, but running on a 64-bit machine

...then use -mx32 rather than -m32.  This compiles code that uses the X32 ABI -
so still 32-bit pointers and integers, but able to e.g. use the extra registers
on 64-bit CPUs:

http://en.wikipedia.org/wiki/X32_ABI

Initial tests indicate search time is reduced by between 5% and 8%, depending on
the configuration/domain/etc.

(NB. This isn't really an issue - I'm just filing it here in case it's of use to
anyone.)