Re: [PATCH] Avoid R_IA64_NONE relocations when relaxing

["H. J. Lu" <hjl at lucon dot org>]

On Mon, Aug 08, 2005 at 03:55:30PM -0700, James E Wilson wrote:
> On Sat, 2005-08-06 at 08:12, H. J. Lu wrote:
> > On Sat, Aug 06, 2005 at 04:43:08PM +0200, Andreas Schwab wrote:
> > ...
> 
> So we have three patches.  Anyone want to argue for or against any of
> them?
> 
> The first patch from HJ, setting gp in elfNN_ia64_relax_section, does
> not seem wise.  This is changing the gp value in the middle of a
> relaxation pass.
> 
> The second patch from HJ, setting gp in elfNN_ia64_final_link, seems OK,
> but has an incorrect comment.  It is possible for some sections to
> increase in size when we start relaxation.  This happens when we insert
> a trampoline for an out-of-range br.  However, no sections increase in
> size after we start doing LTOFF22X relaxations, which is the important
> part.

My comments says "We assume after gp is set, section size will only
decrease." GP is set only when we start doing LTOFF22X relaxations.

> 
> The patch from Andreas seems to have the most thought put into it, but
> it is also by far the most invasive, so it would be nice to have a
> reason to justify it.  It does seem to be modifying gp in the proper
> place, before each relaxation pass.  I think this patch has the
> potential to perform more LTOFF22X relaxations.  As we reduce section
> sizes, more symbol addresses will come within range of gp, and hence

Not all section size reduction will lead to more LTOFF22X relaxations.
I don't think the size of .rela.got section matters here. The most
likely ones are .got and .IA64.pltoff. Can we reduce the size of
.got or .IA64.pltoff during relaxation?

> more LTOFF22X relocations will get relaxed to GPREL22 relocs.  That may
> be a good reason to go with this patch.



H.J.