[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: [Patch]: Fix ld pr11138 FAILures on mips*.



On 12/06/2011 01:16 PM, Richard Sandiford wrote:
Alan Modra<amodra@gmail.com>  writes:
On Mon, Dec 05, 2011 at 04:49:35PM -0800, David Daney wrote:
The root cause of this is that the mips linker synthesizes a special
symbol "__RLD_MAP", and then sets MIPS_RLD_MAP to point to it.  When
a version script is present, this symbol gets versioned along with
all the rest, and when it is time to take its address, the symbol
can no longer be found as it has had version information appended to
its name.

Why not just change

	&&  (strcmp (name, "__rld_map") == 0
	      || strcmp (name, "__RLD_MAP") == 0))

to

	&&  (strncmp (name, "__rld_map", 9) == 0
	      || strncmp (name, "__RLD_MAP", 9) == 0))

in _bfd_mips_elf_finish_dynamic_symbol?  Perhaps the same for other
syms there too?

Showing my ignorance here,

I don't buy it, you are probably the most knowledgeable about this.

but is that the usual behaviour for this kind
of thing?  I wouldn't have expected versions to apply to internally-created
symbols.

Yes, that is what I was trying to accomplish.


There again, is this symbol (as opposed to the DT_MIPS_RLD_MAP tag)
actually part of the ABI?  I can't find any reference to it in the
original psABI, the SGI ELF64 spec, gdb or glibc.  If it's just an
internal thing, maybe we could get rid of it altogether, or at least
make it bind locally rather than globally.


That is an option too I suppose. I would say that it is part of a de facto ABI if nothing else. The question of weather anybody uses it it a different question. I thought boehm-gc may have used it, but I cannot find it there now.

I don't know for sure why the symbol was created, but it seems like it may just be for the side effect of having _bfd_mips_elf_finish_dynamic_symbol() called. This lets us determine mips_elf_hash_table(info)->rld_value at a time when the output sections have already been laid out.

It might be possible to #define elf_backend_output_arch_local_syms and then handle calculation of the rld_value value there instead.

If this seems like a good approach, I can prepare and test a patch that does that.

David Daney


Reply to: