Re: [RFC] Enable --hash-style=both for all recent gcc4 targets

[Phil Blundell <pb@reciva.com>]

On Wed, 2008-10-15 at 12:31 +0100, Richard Purdie wrote:
> I think what Leon is referring to is where different hardware
> capabilities are checked. In Poky I ended up disabling this as at least
> on ARM there were an insane number of different combinations of paths
> being checked for. The patch I added to disable it was:

Right, the default set of hwcaps is probably a bit more inclusive than
most folks want.  I don't think this issue is specific to ARM though
it's possible that arm does have it worse than some other
architectures.  

As with so many things it is basically a DISTRO decision which hwcaps to
include in the list for ld.so.  I suppose the ideal thing would be to
have a variable LD_SO_CONSIDER_HWCAPS or some such, which could be set
by DISTRO.conf to specify which ones should be included.  For the
majority of single-target-platform DISTROs the answer is obviously going
to be none.

Another similar issue for embedded-type DISTROs to consider is
minimising the length of the linker search path by, for example,
eliminating /usr/lib in favour of putting everything in /lib.  (The same
goes for /usr/bin and /bin, of course.)  And of course, not putting
anything in /etc/ld.so.conf unless it's necessary: my openmoko install
seems to have /usr/X11R6/lib and /usr/local/lib listed in that file even
though neither directory actually exists in the filesystem, which will
be causing another 2*N useless path lookups for every library.

Plus of course, PACKAGE_SNAP_LIB_SYMLINKS will eliminate another source
of indirection and hence another dentry name lookup.

p.