PATCH: new GC implementation
Kaveh R. Ghazi
ghazi@caip.rutgers.edu
Thu Sep 30 23:58:00 GMT 1999
> From: Mark Mitchell <mark@codesourcery.com>
>
> >>>>> "Bernd" == Bernd Schmidt <bernds@cygnus.co.uk> writes:
>
> >> Finally, the allocator obtains memory by mmap'ing from
> >> /dev/zero.
>
> Bernd> This doesn't exist on all systems; e.g. HP/UX doesn't
> Bernd> appear to have it.
>
> I believe Alex plans to contribute autoconf support to conditionally
> select ggc-mmap instead of gcc-simple. (The patch he just sent in
> doesn't enable ggc-mmap at all, unless you explicitly turn it on.)
> The autoconf checks will check that `mmap' exists/works and that
> `/dev/zero' exists, I guess. Someone with an interest can extend
> ggc-mmap to deal with other systems; its algorithms should actually
> work just fine with valloc/memalign or with the NT way of mapping
> memory (I forget the syscall right now).
>
> Basically, those extensions will consist of inserting appropriate
> autoconf checks, and then conditionalizing the memory allocation and
> deallocation routines in ggc-mmap.c.
One thing to remember, I think when you compile gcc with a cross
compiler (host!=build) then you can't do the relevant autoconf checks.
E.g. can't see if /dev/zero exists and can't run the "mmap works" test
cause its a runtime check. So in that case, you'll have to back down
to the default GC by default.
--Kaveh
--
Kaveh R. Ghazi Engagement Manager / Project Services
ghazi@caip.rutgers.edu Qwest Internet Solutions
More information about the Gcc-patches
mailing list