On Thu, Aug 23, 2012 at 05:39:37PM -0400, Rich Felker wrote:
> On Mon, Aug 20, 2012 at 04:32:00PM +0200, musl wrote:
> > I missed a bug in my previous patch :
> > in find_sym func precomptab was always set to sysv_precomp.
> 
> It's still broken; h is being used in the comparisons even if h was
> not initialized, rather than using gh. I'm working on integrating the
> code right now. I'll either commit my version or reply with a patch
> here soon for review.

Here's my proposed patch for gnu hash support. I've left dladdr to be
committed separately. I handled the precomputed hashes by duplicating
the code in the two branches; this is _ugly_ but it's moderately
faster, and I really don't like the performance impact of these checks
to begin with, so I'd rather not make them even worse.

Some other changes I've made since Boris's last version:

- Prefer GNU hash if it's available. It's a lot faster even in single
  runs, and should make even more difference when data-locality issues
  come into play (resolving whole files rather than just a single
  dlsym call).

- Omit bloom filter checks. It's not clear if they're beneficial on
  average in large programs, but for single lookups where the symbol
  is present, they increase lookup time by about 8%.

- Replace the over-complicated decode_vec2 with search_vec, since we
  only need a single extended entry anyway. In any case, the big-O
  performance of high-entry lookups will always be the same as this
  linear search unless we use heavy data structures, so we might as
  well just do it this super-simple way.

Comments welcome. I'll hold off on committing for a while in case I
made any dumb mistakes.

Rich