[Ecls-list] Improving CLOS?

[Waldek Hebisch]

> Waldek Hebisch <hebisch at math.uni.wroc.pl> writes:
> 
> > They report (depending on number of parameters) between 218 and 77
> > milions of dispatching calls per second on 2.5 GHz Core 2.  That is
> > 11-33 clocks per call.  I am affraid that with current call convention
> > ECL has no chance to match that: current ECL indirect call costs
> > about 75 clocks, so even if dispatching cost would be reduced to 0,
> > ECL still would be much slower.
> 
> This is surprisingly expensive. Might be a branch prediction miss due to
> the indirect call?

A little correction.  The cost as above is at (default) safety 2.
At safety 0 it drops to 21.29 clocks per call.  I intended to
compile at safety 0, because this is relevant to CLOS calls,
but somewhat the ECL version I used (labeled 10.2.1) ignored
optimization proclaim and compiled at default safety anyway
(I will check if the same happens with 10.3.1 and trunk).
The 75 clocks is mostly due to performing basic operations
(like aref, car, cdr) which are supposed to be very cheap
via calls to library functions.  The 21.29 clocks includes
misprediction time, but also 46.69 instructions per call,
so actual cost of mispredictions seem to be quite low.

Below oprofile report if that helps:

clocks      %   instructions  %      library            routine
3763     25.2737  2648     17.9952  libecl.so.10.2.1    dispatch2
1749     11.7469  2642     17.9545  ECL3NBhsP           L6do_it
1488      9.9940  1282      8.7122  libecl.so.10.2.1    ecl_number_compare
1433      9.6246  1108      7.5297  ECL3NBhsP           L4op1
1375      9.2350  1580     10.7373  ECL3NBhsP           L5op2
727       4.8828  1731     11.7635  libecl.so.10.2.1    ecl_to_fixnum
670       4.5000  684       4.6483  libecl.so.10.2.1    ecl_one_plus
559       3.7544  598       4.0639  ECL3NBhsP           L3op0
540       3.6268  1133      7.6996  ECL3NBhsP           .plt
379       2.5455  496       3.3707  libecl.so.10.2.1    fixint
91        0.6112  107       0.7271  libecl.so.10.2.1    .plt

Comments:
1) clocks are measured in millions (10^6), instructions
   two millions (2*10^6).  dispatch2 has lower number of instructions
   executed to clocks, presumably due to indirect calls
2) actual cost of call may be slightly lower because apparently
   even at safety 0 and having fixnum declarations ECL still
   uses function call for some fixum operations.

-- 
                              Waldek Hebisch
hebisch at math.uni.wroc.pl