Re: Compiling glibc-2.3.2 (-2.3.3?) snapshot 20031115 with gcc-3.4snapshot 20031119

[Andreas Jaeger <aj at suse dot de>]

Matt Austern <austern@apple.com> writes:

> On Dec 2, 2003, at 5:36 AM, Andreas Jaeger wrote:
>
>> Robert Moss <robert.moss@jesus.ox.ac.uk> writes:
>>
>>> Andreas, can you confirm that this is in fact the problem, and that
>>> it's nothing to do with the (possibly slightly ridiculous) situation
>>> in glibc in which the source uses asm behind gcc's back in relation to
>>> weak functions?
>>>
>>> Rob
>>
>> Let's wait for a GCC patch - but looking at the assembly it really
>> seems to be a issue with broken hidden support.\
>
> As you might have noticed from following the discussion on the
> gcc list, the main reason I haven't fixed this yet is that we haven't
> yet reached an agreement about what the semantics of the
> visibility attribute should be when you've got multiple declarations
> of the same symbol.
>
> What assumptions does the glibc source make about this, and how
> tightly wedded are you to those assumptions?

Here's some analysis from my colleague Michael Matz on the issue as it
hits us in glibc - in the first 20 assembler instructions that are
executed ;-(.

} A short testcase is:
} ---------------------
} extern long bla(void);
} long myglobal __attribute__((visibility("hidden")));
} extern long myglobal __attribute__((alias("blubba")));
} void _dl_start (void)
} {
}   myglobal = 1;
} }
} ---------------------
} 
} The 3.3 branch correctly merges the two attribute declarations on
} "myglobal".  I.e. the assembler looks like so (-O1 -fPIC, amd64):
} 
} ---- good.s -----
} .globl myglobal
}         .hidden myglobal
}         .set    myglobal,blubba
} ...
} _dl_start:
}         movq    $1, myglobal(%rip)
}         ret
} -----------------
} 
} The HEAD branch isn't able to merge the two attributes, and hence forgets
} the "hidden" declaration on myglobal, ergo generates this code:
} 
} ------ bad.s ------
} .globl myglobal
}         .set    myglobal,blubba
} ....
} _dl_start:
}         movq    myglobal@GOTPCREL(%rip), %rax
}         movq    $1, (%rax)
}         ret
} -------------------
} 
} There are two issues.  The first would also affect normal code, namely
} that this symbol doesn't have "hidden" visibility in the object file
} anymore, hence would be exported from the DSO.  But what hits us in glibc
} is, that similar code is used in ld.so at a point where it's not yet
} relocated.  This means that the content of the GOT isn't correct yet.  But
} the second version goes over the GOT to access 'myglobal'.  If it were
} hidden for GCC it would correctly use a direct RIP-based access like in
} the first example.

Hope this example helps a bit,

Andreas
-- 
 Andreas Jaeger, aj@suse.de, http://www.suse.de/~aj
  SuSE Linux AG, Deutschherrnstr. 15-19, 90429 Nürnberg, Germany
   GPG fingerprint = 93A3 365E CE47 B889 DF7F  FED1 389A 563C C272 A126

Attachment: pgp00000.pgp
Description: PGP signature