Re: [Patch] sh/sh64: Zero-initialize dynamic relocation sections

[Daniel Jacobowitz <drow at mvista dot com>]

Coincidences...

On Thu, Aug 22, 2002 at 10:35:23AM -0700, Clarke, Stephen wrote:
> 
> The attached file contains a patch for an sh/sh64 linker problem in
> which the linker allocates space for dynamic relocations, but then does
> not initialize that space.
> 
> The patch is very simple, though the explanation is rather
> convoluted.
> 
> Space for dynamic relocations is calculated by sh_elf_check_relocs(),
> before the information in a version script has been processed.
> At this time a symbol can appear to be globally visible,
> and so a relocation which references it has to be resolved dynamically,
> so space is set aside for that dynamic relocation.
> 
> After this, the version script is processed, and some symbols
> may be forced local.
> 
> The contents of the dynamic relocations are actually written by
> sh_elf_relocate_section().  A symbol that originally appeared
> to be global may by now have been forced local. A PC-relative
> relocation can be resolved statically for such a symbol, there
> is no longer any need for the dynamic relocation.  So although
> space for a dynamic relocation was allocated, it is now
> unnecessary, and is left uninitialized.
> 
> This patch simply zero-initializes the space allocated for the
> dynamic relocations, which means they are set to
> R_SH_NONE, rather than random garbage.
> 
> I've included a test for sh64 only, though I think the problem will
> also occur on sh, I haven't been able to construct a test case.
> 
> (I should add that this change was also present in the following
> patch submitted by Kaz Kojima:
> 
> http://sources.redhat.com/ml/binutils/2002-06/msg00765.html
> 
> I'm still hoping to provide the required changes to get Kaz'
> patch working for sh64 and accepted at some stage, but first
> I want to get some simpler patches done...)

Now that you mention it, I went to look at Kaz's patch.  He includes a
more thorough version of the changes I made earlier today... my
question for you is, does the forced-local patch I posted earlier work
for you instead of your patch, or are there non-GOT symbols affected
also?  I prevent emitting space for the relocs in the first place.  Be
sure to try using a current CVS snapshot, since Alan and I fixed a
related problem in GOT refcounts yesterday.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer