mips-elf regressions

mips-elf regressions

[Tristan Gingold]

Hi,

there are many regressions on mips and I'd like this to be fixed  
before doing the final release.

The analysis of one failure (localize-hidden-1) shows that the symtab  
is not correctly ordered:

$ readelf -s tmpdir/bintest.o
Symbol table '.symtab' contains 18 entries:
    Num:    Value  Size Type    Bind   Vis      Ndx Name
      0: 00000000     0 NOTYPE  LOCAL  DEFAULT  UND
      1: 00000000     0 SECTION LOCAL  DEFAULT    1 .text
      2: 00000000     0 SECTION LOCAL  DEFAULT    2 .data
      3: 00000000     0 SECTION LOCAL  DEFAULT    3 .bss
      4: 00000000     0 SECTION LOCAL  DEFAULT    4 .reginfo
      5: 00000000     0 SECTION LOCAL  DEFAULT    5 .pdr
      6: 00002100     0 OBJECT  GLOBAL DEFAULT  ABS Gdefault
      7: 00002200     0 OBJECT  GLOBAL HIDDEN  ABS Ghidden
      8: 00002300     0 OBJECT  GLOBAL INTERNAL  ABS Ginternal
      9: 00002400     0 OBJECT  GLOBAL PROTECTED  ABS Gprotected
     10: 00003100     0 NOTYPE  WEAK   DEFAULT  ABS Wdefault
     11: 00003200     0 NOTYPE  WEAK   HIDDEN  ABS Whidden
     12: 00003300     0 NOTYPE  WEAK   INTERNAL  ABS Winternal
     13: 00003400     0 NOTYPE  WEAK   PROTECTED  ABS Wprotected
     14: 00001200     0 NOTYPE  LOCAL  HIDDEN  ABS Lhidden
     15: 00001300     0 NOTYPE  LOCAL  INTERNAL  ABS Linternal
     16: 00001400     0 NOTYPE  LOCAL  PROTECTED  ABS Lprotected
     17: 00001100     0 NOTYPE  LOCAL  DEFAULT  ABS Ldefault

(Locals must be before Weaks and Weaks must be before Globals  
according to gabi).
As a consequence the output don't match the expected result.

I am now working on fixing this issue.

Tristan.

Re: mips-elf regressions

[Tristan Gingold]

On Oct 8, 2008, at 10:29 AM, Tristan Gingold wrote:
> (Locals must be before Weaks and Weaks must be before Globals  
> according to gabi).
> As a consequence the output don't match the expected result.

Well according to elf32-mips.c, this is expected:

/* Determine whether a symbol is global for the purposes of splitting
    the symbol table into global symbols and local symbols.  At least
    on Irix 5, this split must be between section symbols and all other
    symbols.  On most ELF targets the split is between static symbols
    and externally visible symbols.  */

static bfd_boolean
mips_elf_sym_is_global (bfd *abfd ATTRIBUTE_UNUSED, asymbol *sym)
{
   if (SGI_COMPAT (abfd))
     return (sym->flags & BSF_SECTION_SYM) == 0;
   else
     return ((sym->flags & (BSF_GLOBAL | BSF_WEAK)) != 0
	    || bfd_is_und_section (bfd_get_section (sym))
	    || bfd_is_com_section (bfd_get_section (sym)));
}

As SGI_COMPAT is true, this looks to be expected.

Can someone confirm this ?
If so, shouldn't we update the status of the failing tests (either  
disabling them on mips-elf or marking
them as xfail) ?

Tristan.

Re: mips-elf regressions

[Thiemo Seufer]

Tristan Gingold wrote:
>
> On Oct 8, 2008, at 10:29 AM, Tristan Gingold wrote:
>> (Locals must be before Weaks and Weaks must be before Globals  
>> according to gabi).
>> As a consequence the output don't match the expected result.
>
> Well according to elf32-mips.c, this is expected:
>
> /* Determine whether a symbol is global for the purposes of splitting
>    the symbol table into global symbols and local symbols.  At least
>    on Irix 5, this split must be between section symbols and all other
>    symbols.  On most ELF targets the split is between static symbols
>    and externally visible symbols.  */
>
> static bfd_boolean
> mips_elf_sym_is_global (bfd *abfd ATTRIBUTE_UNUSED, asymbol *sym)
> {
>   if (SGI_COMPAT (abfd))
>     return (sym->flags & BSF_SECTION_SYM) == 0;
>   else
>     return ((sym->flags & (BSF_GLOBAL | BSF_WEAK)) != 0
> 	    || bfd_is_und_section (bfd_get_section (sym))
> 	    || bfd_is_com_section (bfd_get_section (sym)));
> }
>
> As SGI_COMPAT is true, this looks to be expected.
>
> Can someone confirm this ?

This is correct. mips-elf uses the SGI-ish object layout with
nonstandard (or pre-standard?) symbol sorting.

> If so, shouldn't we update the status of the failing tests (either  
> disabling them on mips-elf or marking
> them as xfail) ?

I think the test needs to check for different expected output for
mips-elf and mips-sgi-irix*


Thiemo

Re: mips-elf regressions

[Alan Modra]

On Wed, Oct 08, 2008 at 10:50:55AM +0200, Tristan Gingold wrote:
> On Oct 8, 2008, at 10:29 AM, Tristan Gingold wrote:
>> (Locals must be before Weaks and Weaks must be before Globals  
>> according to gabi).

Not quite.  Locals before others, weaks and globals are not ordered.

> As SGI_COMPAT is true, this looks to be expected.
>
> Can someone confirm this ?

Yes.  Although I don't really see a need to break the ELF standard
when there are no global section symbols.

> If so, shouldn't we update the status of the failing tests (either  
> disabling them on mips-elf or marking
> them as xfail) ?

xfail.

-- 
Alan Modra
Australia Development Lab, IBM

Re: mips-elf regressions

[Tristan Gingold]

On Oct 8, 2008, at 12:13 PM, Alan Modra wrote:

> On Wed, Oct 08, 2008 at 10:50:55AM +0200, Tristan Gingold wrote:
>> On Oct 8, 2008, at 10:29 AM, Tristan Gingold wrote:
>>> (Locals must be before Weaks and Weaks must be before Globals
>>> according to gabi).
>
> Not quite.  Locals before others, weaks and globals are not ordered.

Right, I read too quickly.

>> As SGI_COMPAT is true, this looks to be expected.
>>
>> Can someone confirm this ?
>
> Yes.  Although I don't really see a need to break the ELF standard
> when there are no global section symbols.

Aren't section symbols always local ?  If so we could also classify  
symbols
into section symbols, local and non-local (thought for a future  
improvement).

>> If so, shouldn't we update the status of the failing tests (either
>> disabling them on mips-elf or marking
>> them as xfail) ?
>
> xfail.

Ok.  Thank you for your quick reply.

Tristan.

Re: mips-elf regressions

[Alan Modra]

On Wed, Oct 08, 2008 at 12:20:58PM +0200, Tristan Gingold wrote:
> Aren't section symbols always local ?

Usually, but the standard doesn't prohibit global section symbols.
See http://sourceware.org/ml/binutils/2005-07/msg00494.html

>  If so we could also classify symbols
> into section symbols, local and non-local (thought for a future  
> improvement).

Yes, for standard ELF:
  local section, local, all others
For Irix compat:
  local section, global section, local, all others.

-- 
Alan Modra
Australia Development Lab, IBM

Re: mips-elf regressions

[Maciej W. Rozycki]

On Wed, 8 Oct 2008, Tristan Gingold wrote:

> > > As SGI_COMPAT is true, this looks to be expected.
> > > 
> > > Can someone confirm this ?
> > 
> > Yes.  Although I don't really see a need to break the ELF standard
> > when there are no global section symbols.
> 
> Aren't section symbols always local ?  If so we could also classify symbols
> into section symbols, local and non-local (thought for a future improvement).

 They are, but for SGI_COMPAT non-section symbols are always external, 
which is not always true for "traditional" targets.  There's been 
something subtle about it I don't recall exactly anymore, but you should 
be able to track it down by chasing the discussion around the change to 
BFD that introduced the distinction and/or scanning the sources for other 
IRIX 5 conditionals.  I think this is something about one of the DYNAMIC 
entries being interpreted by IRIX in such a way that it prevents the use 
of a linker script to force symbols local at the static link time.  Which 
the GNU C library depends on heavily.

  Maciej

Re: mips-elf regressions

[Ian Lance Taylor]

Thiemo Seufer <ths@networkno.de> writes:

> This is correct. mips-elf uses the SGI-ish object layout with
> nonstandard (or pre-standard?) symbol sorting.

Nonstandard, even then.  It was a (possibly unintended) consequence of
SGI hacking their ECOFF linker to generate ELF.

Ian

Re: mips-elf regressions

[Maciej W. Rozycki]

On Wed, 8 Oct 2008, Alan Modra wrote:

> > As SGI_COMPAT is true, this looks to be expected.
> >
> > Can someone confirm this ?
> 
> Yes.  Although I don't really see a need to break the ELF standard
> when there are no global section symbols.

 SGI_COMPAT is *the* ELF standard for MIPS per the psABI supplement.  
Unfortunately it's incompatible with the requirements of the GNU C 
library, which is why the "traditional" variation has been created around 
the time glibc 2.2 was released.  The incompatibility was the main reason 
there has never been glibc 2.1 for MIPS/Linux.

  Maciej