On Mon, 26 Aug 2002, David Laight wrote:

> > Please read the thread on tech-userlevel. There are a lot of good reasons
> > for doing this, all of which involve now being able to dlopen() libraries.
>
> It ought to be possible to build a dynamically linked program
> that does not use the dynamic libc.so  This ought to speed up
> program startup (where this is significant) while still allowing
> dlopen() to be used in specific functions.

	I think that was discussed a while back and involves breaking
	(or extending if you prefer :) the ELF spec, and changes
	to the toolchain which would not be bought back by the gnu
	people. That would mean an out-of-the-box gnu toolchain
	would not be able to target a dynamic NetBSD system (A Bad
	Thing).

> I was going to look at the kernel elf exec code - it would probably
> be benefitial for it to keep a reference to the vnode for elf
> interpreters (with maybe a sysctl to dump the list for debug).
>
> I've also just make /usr/lib/libc.so.12 a hard link to libc.so.12.86,
> this improves my test to 1.520 (or so) - total gain 9%
> So removing the symlink is actually a bigger gain...

	Interesting. It would seem that there is not as much caching as
	there should be...

-- 
		David/absolute          -- www.netbsd.org: No hype required --