Re: compiling binutils with -DDEBUG

[Alan Modra <amodra@bigpond.net.au>]

On Tue, Mar 01, 2005 at 09:24:54AM -0500, Ian Lance Taylor wrote:
> Alan Modra <amodra@bigpond.net.au> writes:
> 
> > On Mon, Feb 28, 2005 at 09:36:00PM -0500, Ian Lance Taylor wrote:
> > > Alan Modra <amodra@bigpond.net.au> writes:
> > > 
> > > > 	(fixup_segment): Remove assertion.
> > > > Didn't seem worth including struct-symbol.h just for this.  I suspect
> > > > the assertion isn't true too, since a similar assertion in obj-coff.c
> > > > fails with X_op an O_constant.  Maybe this is a real bug.  Ian, you're
> > > > probably the only one who can answer mri mode queries off the top of
> > > > your head.  Care to comment?
> > > 
> > > > @@ -2583,7 +2585,6 @@ fixup_segment (fixS *fixP, segT this_seg
> > > >        if (fixP->fx_addsy != NULL
> > > >  	  && symbol_mri_common_p (fixP->fx_addsy))
> > > >  	{
> > > > -	  know (fixP->fx_addsy->sy_value.X_op == O_symbol);
> > > >  	  add_number += S_GET_VALUE (fixP->fx_addsy);
> > > >  	  fixP->fx_offset = add_number;
> > > >  	  fixP->fx_addsy
> > > 
> > > The assertion should be true.  This is a symbol defined in an MRI
> > > common section.  Since we don't generate IEEE object files, we treat
> > > an MRI common section as a single large common symbol.  Then we handle
> > > a reloc against a symbol defined within an MRI common section as an
> > > offset to that symbol.  The assertion is just checking that we really
> > > do point to the magic symbol we are using for the MRI common section.
> > 
> > I did a little analysis.  Nowadays, the assertion is always false for
> > BFD_ASSEMBLER because symbols have been resolved (write.c:1857) by the
> > time this code is reached.  The following would probably fix the
> > assertion, but I don't think it's worth worrying about.
> 
> There is code in resolve_symbol_value() to handle this case.  I wonder
> why it isn't working?

The reason the assert fails is that finalize_syms is set by the time we
reach write.c:1857, so S_SET_VALUE is called for these symbols.  That
turns them into O_constant values.

>  If the assert in fixup_segment() fails, then
> MRI common symbols will not be handled correctly.

Are you sure?  I couldn't see why it mattered that the sym was resolved
at write.c:1857 versus being resolved at write.c:2588.  The mri common
testcase looks to produce good results.

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre