[Shootout-list] Missing programs and score calculation (Was:

[skaller]

On Fri, 2005-09-16 at 09:23 +0200, Marco van de Voort wrote:
> > Felix outperforms C/C++ even though it actually generates C++!
> 
> In selected benchmarks of course. One benchmark that is known to favour
> functional languages is a bit simplistic isn't it ? :-)

But C/C++ comes second for Ackermann, beating Ocaml, Haskell,
and other functional languages .. so why do you say this test
favours functional languages??

In fact, Felix outperforms the competition on quite a lot
of the Shootout tests: for exception handling, threading
by several orders of magnitude, and it equals the leader
for many other tests (where the leader is C, Felix can
always equal it simply because it generates C, literally
if necessary).

For one test (regexp) Felix gets the right answer because it
has simple clear syntax and ALL the other languages
get it wrong -- because the Shootout result data is itself
wrong and the tests cheated and just copied the wrong
Perl regexp.

The thing is, for many features Felix is actually slower,
but this isn't evident in the Shootout because the tests
are too trivial: I could easily construct examples where
Felix would not perform well. For example: C/C++ code
releases memory at the end of a test .. Felix doesn't,
and the tests don't use enough memory to ever trigger
garbage collection: the Felix gc is quite slow compared
to, for example, Ocaml.

> gcc is so slow due to the fact that it reparses headers ad infinitum (no
> precompiled headers) and because it starts a zillion different processes per
> compilation unit (at least cpp,gcc and as)

No. Surely these things don't help, but the real reason is
that it uses extremely bad data structures. For example
a linear list is used for lookup: this is an O(n) search.
Sensible compilers use Hashtables .. which is an O(1) search.

gcc also has optimisations with quadratic time, however it
is singularly incapable of handling large functions, even
with optimisation off.

I have actually been forced to recode some Felix regression
tests because gcc was taking 20 minutes to compile a large
function --- reduced to 10 seconds by simply splitting it
up into collection of smaller functions -- except that 
10% of the time gcc actually crashed with an internal error:
it was using the WHOLE of my systems 1Gig of RAM to
compile this code .. argggg ...

This is disgusting, since basic C can be compiled entirely
linearly, in one pass with backpatching (for labels and
forward declarations). Remember this is without optimisation ..


-- 
John Skaller <skaller at users dot sourceforge dot net>

-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: This is a digitally signed message part
Url : http://lists.alioth.debian.org/pipermail/shootout-list/attachments/20050916/ce30b728/attachment.pgp