Re: powerpc help needed: Defining hidden alias for __sigsetjmp

[Florian Weimer <fweimer at redhat dot com>]

On 11/14/2016 04:37 PM, Andreas Schwab wrote:
> On Nov 14 2016, Florian Weimer <fweimer@redhat.com> wrote:
>
> > I came up with the attached patch.  It adds a definition of the hidden
> > redirect to libc.so, as expected.  But in addition to that, I get this ABI
> > difference on powerpc32:
> >
> > -GLIBC_2.3.4 __longjmp F
>
> Are you sure this is the only change you did?  setjmp.* does not contain
> anything related to longjmp.  I couldn't reproduce that.

Thanks.  It only happens in conjunction with this patch:

  <https://sourceware.org/ml/libc-alpha/2016-11/msg00465.html>

This patch adds a reference to __longjmp to libc.so.  With it, 
libc_pic.os has this:

 8296: 00017024    492 FUNC    GLOBAL HIDDEN         2 
__longjmp@@GLIBC_2.3.4
 9342: 00017210    184 FUNC    GLOBAL DEFAULT        2 __longjmp@GLIBC_2.0

And libc.so has this:

 5208: 00039c04    492 FUNC    LOCAL  DEFAULT       11 
__longjmp@@GLIBC_2.3.4
 5758: 00039df0    184 FUNC    LOCAL  DEFAULT       11 __longjmp@GLIBC_2.0

It turns out that libc.map lists _longjmp only, based on the entry in 
sysdeps/powerpc/Versions.

setjmp/__longjmp.os has:

   11: 00000000    492 FUNC    GLOBAL DEFAULT        1 __vmx__longjmp
   14: 000001ec    184 FUNC    GLOBAL DEFAULT        1 __novmx__longjmp
   15: 00000000    492 FUNC    GLOBAL DEFAULT        1 
__longjmp@@GLIBC_2.3.4
   16: 000001ec    184 FUNC    GLOBAL DEFAULT        1 __longjmp@GLIBC_2.0

setjmp/longjmp.os has:

   21: 00000000    116 FUNC    GLOBAL DEFAULT        1 __vmx__libc_longjmp
   22: 00000000    116 FUNC    GLOBAL DEFAULT        1 
__vmx__libc_siglongjmp
   23: 00000000    116 FUNC    WEAK   DEFAULT        1 __vmx_longjmp
   24: 00000000    116 FUNC    WEAK   DEFAULT        1 __vmxlongjmp
   25: 00000000    116 FUNC    WEAK   DEFAULT        1 __vmxsiglongjmp
   30: 00000000    116 FUNC    GLOBAL HIDDEN         1 
__GI___vmx__libc_longjmp
   31: 00000000    116 FUNC    GLOBAL DEFAULT        1 
__libc_longjmp@@GLIBC_PRIVATE
   32: 00000000    116 FUNC    GLOBAL DEFAULT        1 
__libc_siglongjmp@@GLIBC_PRIVATE
   33: 00000000    116 FUNC    WEAK   DEFAULT        1 
_longjmp@@GLIBC_2.3.4
   34: 00000000    116 FUNC    WEAK   DEFAULT        1 longjmp@@GLIBC_2.3.4
   35: 00000000    116 FUNC    WEAK   DEFAULT        1 
siglongjmp@@GLIBC_2.3.4

So the __GI_* aliases are missing there, too.

The internal reference to __longjmp seems to cause ld to drop the 
exported symbol version.  Adding __longjmp to sysdeps/powerpc/Versions 
does not change this.  Neither does adding

libc_hidden_ver (__vmx__libc_longjmp, __longjmp)

to sysdeps/powerpc/longjmp.c.  Okay, that's because the definition comes 
from sysdeps/powerpc/powerpc32/fpu/__longjmp-common.S.  But changing 
that file to contain:

versioned_symbol (libc, __vmx__longjmp, __longjmp, GLIBC_2_3_4); 

# define __longjmp_symbol  __vmx__longjmp
# include "__longjmp-common.S"
libc_hidden_ver (__vmx__libc_longjmp, __longjmp)

does not help, either.

Maybe this is relevant.  __longjmp is declared *hidden*:

extern void __longjmp (__jmp_buf __env, int __val)
     __attribute__ ((__noreturn__)) attribute_hidden;

So maybe that's why a reference to it causes ld to drop the export.

I wonder if this indeed the root cause.  It would mean this is a 
spurious ABI addition.  Can we drop __longjmp from the ABI in this way? 
Or do we have to fix this up using a powerpc-specific alias?

This happens on no other architecture for which we have abilist files. 
It was introduced into the abilist file here:

commit 0741d64c9140aa205f9df8ebee80ca0bcb018445
Author: Andreas Schwab <schwab@linux-m68k.org>
Date:   Tue May 1 01:12:54 2012 +0200

    Update powerpc ABI data

Thanks,
Florian