Relocations put in read-only segment?

Relocations put in read-only segment?

[Charles M. Hannum]

When I compile the following code with egcs 1.0.2, using `-O -fPIC',
on either i386-*-netbsd or i586-*-linux-gnulibc1, the relocation for
`XtStrings+788' is put in a read-only section (the text section under
NetBSD, rodata under Linux).  While it's true that ld.so on both
platforms is careful to make this work, it's canonically wrong.  If
the relocation is done in a data section at all, it should be in a
writable section.


The code in question (whittled down from a case in Xaw3d):

extern const  char XtStrings[];
void NotifyThumb (top)
    float top;
{
    XtCallCallbacks (&XtStrings[788], top + 0.0001);
}

Under egcs 1.0.2, the reference to XtStrings is compiled as:

.text
[...]
LC1:
        .long _XtStrings+788
        .align 2
[...]
        pushl LC1@GOTOFF(%ebx)

whereas, under gcc 2.7.2.2, it's compiled as:

.text
[...]
        movl _XtStrings@GOT(%ebx),%eax
        addl $788,%eax
        pushl %eax


Anyone know why this happens, or do I need to track it down myself?

Re: Relocations put in read-only segment?

[Jeffrey A Law]

  In message < 199804240119.VAA15305@lunacity.ne.mediaone.net >you write:
  > When I compile the following code with egcs 1.0.2, using `-O -fPIC',
  > on either i386-*-netbsd or i586-*-linux-gnulibc1, the relocation for
  > `XtStrings+788' is put in a read-only section (the text section under
  > NetBSD, rodata under Linux).  While it's true that ld.so on both
  > platforms is careful to make this work, it's canonically wrong.  If
  > the relocation is done in a data section at all, it should be in a
  > writable section.
This is a pretty common problem; generally it's fixed by having
SELECT_SECTION and related functions put these things into the
data section for PIC code generation.  Unfortunately many ports
do not handle this case correctly.

jeff

Re: Relocations put in read-only segment?

[Charles M. Hannum]

Well, I solved the problem by changing SELECT_RTX_SECTION() in our
source tree.  Thanks for the pointer.

(We have a bunch of other changes to merge into the egcs tree, which
I'll get to RSN...)

Re: Relocations put in read-only segment?

[Charles M. Hannum]

Jeffrey A Law <law@hurl.cygnus.com> writes:

>   In message < 199804240119.VAA15305@lunacity.ne.mediaone.net >you write:
>   > When I compile the following code with egcs 1.0.2, using `-O -fPIC',
>   > on either i386-*-netbsd or i586-*-linux-gnulibc1, the relocation for
>   > `XtStrings+788' is put in a read-only section (the text section under
>   > NetBSD, rodata under Linux).  While it's true that ld.so on both
>   > platforms is careful to make this work, it's canonically wrong.  If
>   > the relocation is done in a data section at all, it should be in a
>   > writable section.
> This is a pretty common problem; generally it's fixed by having
> SELECT_SECTION and related functions put these things into the
> data section for PIC code generation.  Unfortunately many ports
> do not handle this case correctly.

That doesn't quite follow.  There is currently no SELECT_SECTION on
the platforms I mentioned.  This means they use the default code:

      if (DECL_READONLY_SECTION (decl, reloc))
        readonly_data_section ();
      else
        data_section ();

Note that DECL_READONLY_SECTION specifically does:

   && ! (RELOC && (flag_pic || DECL_ONE_ONLY (DECL))))

This implies that `reloc' is false in this case, which would be a bug.

Re: Relocations put in read-only segment?

[Charles M. Hannum]

> That doesn't quite follow.  There is currently no SELECT_SECTION on
> the platforms I mentioned.  This means they use the default code:

Hmmm.  It's even worse than that.  It's not going through either of
the pieces of code that use SELECT_SECTION() at all.

Yuck.

Re: Relocations put in read-only segment?

[Richard Henderson]

On Fri, Apr 24, 1998 at 01:53:05PM -0600, Jeffrey A Law wrote:
> This is a pretty common problem; generally it's fixed by having
> SELECT_SECTION and related functions put these things into the
> data section for PIC code generation.  Unfortunately many ports
> do not handle this case correctly.

Actually, Jeff,

]        pushl LC1@GOTOFF(%ebx)                                       
    vs
]        movl _XtStrings@GOT(%ebx),%eax
]        addl $788,%eax                    
]        pushl %eax

is a code gen issue and has nothing to do with SELECT_SECTION. 
This would have to do with GO_IF_LEGITIMATE_ADDRESS and friends.

But I don't replicate it with the given example and 19980418.


r~

Re: Relocations put in read-only segment?

[Jeffrey A Law]

  In message < el2ra2mho8n.fsf@lunacity.ne.mediaone.net >you write:
  > (We have a bunch of other changes to merge into the egcs tree, which
  > I'll get to RSN...)
Cool.

Not sure who "we" is, but as you know, we'll probably need copyright
assignments for significant contributions.

jeff