[PATCH] Sparc32 ELF TLS fixes

[Alan Modra]

On Mon, Apr 11, 2005 at 12:11:33PM -0700, David S. Miller wrote:
> The .rodata starts as .rodata.str1.8 in sr3.o:
> 
> Contents of section .rodata.str1.8:
>  0000 6175746f 6d617469 635f666f 6f000000  automatic_foo...
>  0010 73746174 69635f66 6f6f0000 00000000  static_foo......
> 
> It is in fact 32 bytes.  But something causes it to shrink when
> outputting ELF.  And this something is _bfd_elf_merge_sections().
> The above is the only .rodata.* section, so it is the only one
> processed during _bfd_merge_sections().  It trims the strings so
> that there are no spurious NULL bytes.  And that's how we end up
> with a 27 byte .rodata output section from the original 32 byte
> input section.

That sounds like a ld bug.  The strings shouldn't have their
alignment changed.  Since gcc is emitting them to a .rodata.str1.8
section, I'd expect readelf -S sr3.o should show something like

.. .rodata.str1.8 PROGBITS .. .. .. 01 AMS 0 0 8

The section alignment tells the merge code to not reduce string
alignment below that value.

> I don't see any reasonable way to fix this besides possibly disabling
> section merging via command line option for this test case.  Is there
> even a way to turn SEC_MERGE processing off from the ld command line?

No, this optimization is controlled by the section flags.  Turn it off
with gcc's --fno-merge-constants.

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre