[Jscheme-devel] jscheme/prims.scm update

[Timothy J. H. <tim...@ma...>]

Report on direct implementation of primitives in javadot...

The jscheme/prims.scm approach still has a way to go before it is 
competitive with
our current approach. I ran the gabriel benchmarks on standard JScheme
(on a 1 GHZ PowerPC G4 Mac OSX 10.3.4)
and then ran it again after loading in jscheme/prims.scm

It appears that the pure javadot implementation is about 3-5 times 
slower
(and there are still a few bugs! which cause FFT, Fprint and Fread not 
to work,
I'll isolate them and put them in jscheme/SchemeTests.scm...)


                           CPSTAK CTAK      Dderiv  Deriv  Destructive 
Div-iter   Div-rec
Mac G4/std        1839      26520      2682     1833     5407       
1834       1887
Mac G4/prims    8162   134623    21427   22026  39038       5885       
5701

                                 FFT    Fprint    Fread   Puzzle  TAK    
  TAKL    TAKR
Mac G4/std          10712  116         480     241     1072      8584   
1162
Mac G4/prims      FAIL    FAIL        FAIL    650     6946    27842   
7381

Using the javadot notation seems to double the time for standard 
primitives (note (car x) is defined by x.first() in primitives.scm)
 > (define L '(a b c d e f g))
 > (time (car L) 100000)    ---> (a (422 msec) (357280 bytes))
 > (time (.first$ L) 100000)---> (a (869 msec) (357280 bytes))

also adding a lambda layer makes the time triple that of the regular 
primitive
 > (define (mycar L) (.first$ L))   ---> (lambda mycar (L)...)
 > (time (mycar L) 100000) ---> (a (1207 msec) (163008 bytes))

This provides an automatic slowdown of about 3.
Compilation could remove most of this slowdown if done well.

---Tim---