Re: PCH failures / ICE: Segmentation fault

[Geoff Keating <geoffk at apple dot com>]

On Nov 5, 2003, at 1:43 PM, Per Bothner wrote:

> Laurent GUERBY wrote:
>
> > I changed my build & test script to avoid running make install
> > before running make -k check, it fixed all the pch failures.
>
> I am seeing lots of PCH failures while running make check, and this is 
> without running make install - and just C.
>
> This is on Fedora Core 0.95.  (I.e. the new community-based version of 
> Red Hat Linux.)
>
> I took at look at the problem using gdb.  The behavior is not 
> consistent.  One "run" will finish cleanly, while the next "run" will 
> crash.  It looks like different runs will allocate memory or addresses 
> in different locations.  I don't know what could cause this.
>
> The FC release notes (http://fedora.redhat.com/docs/release-notes/) 
> methods that FC supports "Position Independent Executables", but 
> implies it is not the default.

I'm told there's some kind of trickiness in later versions of Red Hat 
Linux which causes libraries to be loaded at random (yes, really 
random, as in from /dev/urandom) memory locations, to make it harder to 
exploit certain security flaws; if so, sometimes they will collide with 
the addresses that PCH expects.  I expect what would need to be done in 
that case would be something like what config/rs6000/host-darwin.c 
does, allocating fixed space in the non-position-independent 
executable; or, perhaps the OS has some region where it won't ever map 
libraries, in which case that could be used.

... however, this should print a nice 'sorry' message, rather than 
crashing.  So maybe the problem is something completely different.

--
Geoff Keating <geoffk@apple.com>