Re: symbol relocation during dynamic loading

[Robert Schweikert <Robert dot Schweikert at abaqus dot com>]

On Mon, 2004-04-19 at 16:44, Ian Lance Taylor wrote:
> Robert Schweikert <Robert.Schweikert@abaqus.com> writes:
> 
> > I am using the --retain-symbols-file <filename> link option to restrict
> > the exported symbols from shared libraries. Using elfdump libName
> > --export to compare the exported symbols between libraries linked with
> > and without the --retain-symbols-file <filename> option indicates that
> > things worked as expected and I only get the symbols I want exported.
> > Also the size of the "restricted" library is smaller (as expected).
> > 
> > The question now is why the relocations are the same between the 2
> > libraries. If I set the LD_DEBUG environment variable to statistics and
> > load both libraries I end up with the same number of relocations. My
> > expectation was that the number of relocations would decrease and thus I
> > would end up with faster load times.
> > 
> > Can anyone explain why the number of relocations did not show any
> > changes?
> 
> Hmmm.  You neglected to mention your target. 

The target is IA32

>  The only type of
> relocation which is likely to disappear entirely is a PC relative
> relocation to a publically visible function.  If the function is
> hidden, then the relocation can go away.  All other relocations are
> likely to remain, perhaps in the form of R_*_RELATIVE.
> 
> That said, there is another way to hide symbols in a shared library,
> which is to use a version script.  It is possible that using a version
> script will remove relocations where --retain-symbols-file will not.
> That is because the implementation of the two options is, I believe,
> different.  I haven't checked whether this is so.  But it is possible.
> (Yes, if they are indeed different, this probably a bug.)

OK, this leads me to a follow up question.If the number of relocations
does not change the improved load performance I am hoping to achieve
would be based on hopefully shorter hash buckets and fewer string
comparisons. Is this correct? And is there a way to extract this
information from the linker?

Thanks,
Robert

> 
> Ian
-- 
Robert Schweikert <Robert.Schweikert@abaqus.com>
ABAQUS