"ld" memory usage

[Bertram Felgenhauer]

Roman Beslik wrote:
> This is not strictly a GHC question. I observed that "ld" when
> linking GHC-compiled programs eats 0.5 GB of resident memory. ~3
> times more than GHC or Haddock.

Oh I remember investigating this problem 2 years ago. Looking back it
seems I never made my observations public.

GNU ld uses hash tables to store the symbols of each object file. The
default size of those hash tables is 4051, which is a huge waste in
Haskell's split-objs case, where object files are plenty and only have
a couple of dozen symbols (most of them external references) at most.

I've been running binutils with this patch,

    http://int-e.home.tlink.de/haskell/binutils-2.18-tune-bfd-hash.patch
    (still applies to 2.20)

which changes the default size to 31 and adds some more intermediate
sizes. This improves memory usage a lot, at a small cost in running time
(if I had numbers, they're lost by now).

HTH,

Bertram