Re: [PATCH] PR ld/21402: Skip PIE indirect5 and indirect6 tests on i386

[Alan Modra <amodra at gmail dot com>]

On Fri, Apr 21, 2017 at 11:37:34AM -0700, H.J. Lu wrote:
> @@ -2863,10 +2867,14 @@ elf_i386_allocate_dynrelocs (struct elf_link_hash_entry *h, void *inf)
>        int tls_type = elf_i386_hash_entry(h)->tls_type;
>  
>        /* Make sure this symbol is output as a dynamic symbol.
> -	 Undefined weak syms won't yet be marked as dynamic.  */
> +	 Undefined weak syms won't yet be marked as dynamic.
> +	 PR ld/21402: If this symbol isn't undefined weak symbol,
> +	 don't make it dynamic in PIE.   */
>        if (h->dynindx == -1
>  	  && !h->forced_local
> -	  && !resolved_to_zero)
> +	  && !resolved_to_zero
> +	  && (h->root.type == bfd_link_hash_undefweak
> +	      || !bfd_link_pie (info)))
>  	{
>  	  if (! bfd_elf_link_record_dynamic_symbol (info, h))
>  	    return FALSE;

This doesn't look correct to me.  Sure, it fixes the particular
testcase, but why do you make non-pie non-undefweak symbols dynamic
here?  I believe that only undefweak symbols should be made dynamic
here, and in all the other places in allocate_dynrelocs.

x86_64 too, but fixing allocate_dynrelocs for x86_64 causes no-plt-1e
to fail due to lack of a relative GOT dynamic reloc on 
	cmpq	func@GOTPCREL(%rip), %rax
in no-plt-check1.S.  Seems like x86_64 would need something like your
change to elf32-i386.c to emit relative GOT dynamic relocs in
relocate_section.  (For powerpc I decided to only handle plt and copy
relocs in finish_dynamic_symbol.  GOT dynamic relocs are all handled
in relocate_section as otherwise there is a lot of duplication and IMO
less comprehensible code.)

> @@ -3818,6 +3826,7 @@ elf_i386_relocate_section (bfd *output_bfd,
>        bfd_vma st_size;
>        asection *resolved_plt;
>        bfd_boolean resolved_to_zero;
> +      bfd_boolean relative_reloc;
>  
>        r_type = ELF32_R_TYPE (rel->r_info);
>        if (r_type == R_386_GNU_VTINHERIT
> @@ -4268,6 +4277,7 @@ r_386_got32:
>  	  if (htab->elf.sgot == NULL)
>  	    abort ();
>  
> +	  relative_reloc = FALSE;
>  	  if (h != NULL)
>  	    {
>  	      bfd_boolean dyn;
> @@ -4301,6 +4311,17 @@ r_386_got32:
>  		      bfd_put_32 (output_bfd, relocation,
>  				  htab->elf.sgot->contents + off);
>  		      h->got.offset |= 1;
> +
> +		      if (h->dynindx == -1
> +			  && !h->forced_local
> +			  && h->root.type != bfd_link_hash_undefweak
> +			  && bfd_link_pie (info))
> +			{
> +			  /* PR ld/21402: If this symbol isn't dynamic
> +			     in PIE, generate R_386_RELATIVE here.  */
> +			  eh->no_finish_dynamic_symbol = 1;
> +			  relative_reloc = TRUE;
> +			}
>  		    }
>  		}
>  	      else

This also looks odd.  Surely you need a relative reloc when pic, not
pie?  It seems to me you need one any time you won't be calling
finish_dynamic_symbol (which is where x86 sets most dynamic GOT
relocs) and have a symbol that doesn't resolve to zero.

-- 
Alan Modra
Australia Development Lab, IBM