Re: libtool linking libgcj

[Nicholas Wourms <nwourms at netscape dot net>]

aph@redhat.com wrote:

> Alexandre Oliva writes:
>  > On Oct  8, 2003, Andrew Haley <aph@redhat.com> wrote:
>  > 
>  > > Nicholas S. Wourms writes:
>  > 
>  > >> What about splitting them up into 5 or 6 convenience libraries?  
>  > 
>  > > Okay, but it would be some considerable work, and how much improvemnt
>  > > should I expect?  Why would splitting the library help?
>  > 
>  > It would avoid piecewise linking and huge command lines, except on
>  > systems whose linkers don't support --whole-archive or equivalent, in
>  > which libtool would have to extract the archives and then do piecewise
>  > linking.
>
> Okay, but how would that help?  Most of the time seems to be in the
> libtool shell cript, and I can't see why convenience libraries would
> shorten that time.

I can't give a technical reason other then to say this is what the KDE 
folks did when faced with linking huge amounts of object files.  From 
practical experience, I definitely notice an improvement when comparing 
a scenario of linking 100 objects in one go to one involving the linking 
of 10 conv libs containing 10 objects each.  But you are right, most of 
the time is consumed by the routines in the libtool script which 
validate the object files and object scripts(.lo).  As I mentioned 
before the new libtool 1.5 routines have been further optimized and 
streamlined in such a way that link times have been greatly reduced in 
most cases (as compared to libtool 1.4).  Another idea, albeit a crazy 
one, would be to convert the libtool shell script to pure C, which would 
obviously improve performance significantly.

But to return to the kde example...  They have also come up with another 
interesting way to reduce the number of object files at link time:  they 
 combine all the source files in a subdir into a single source file 
using "#include <source.cc>" directives and building it as a single 
object.  This may not be good for debugging or even work at all for 
libjava, but if it could work, it would definitely cut down on the 
number of object files libtool needs to process, which in turn, would 
cut down on the length of time required to do the final link.

Just a few observations...

Cheers,
Nicholas