Re: GLIL->C compilation

[Andy Wingo]

On Sat 20 Nov 2010 00:56, l...@gnu.org (Ludovic Courtès) writes:

>> Yes on one of my machine one loop takes about 6ns for the compiled one and 
>> about 50ns for the bytcode, say a factor of 7-10 for simple tasks, but 
>> remember
>> this is for rather trivial work.
>
> Woow, I didn’t expect so much.

I would expect it, yes. Something aroung 5x for dispatch, with more if
the compiler decided to inline more things across "instruction"
boundaries.

Andy
-- 
http://wingolog.org/

Re: GLIL->C compilation

[Ludovic Courtès]

Stefan Israelsson Tampe  writes:

> On Saturday, November 20, 2010 12:11:16 am Ludovic Courtès wrote:
>> Hi,
>> 
>> Stefan Israelsson Tampe  writes:
>> > #include "cfkn.h"
>> > #define LOCAL_REF(i) fp[i]
>> > #define LOCAL_SET(i,v)   fp[i] = (v)
>> > void cfkn3676(SCM *fp, SCM **spp)
>> > {
>> > 
>> >  /* setup of environment */
>> >  SCM *objects, *free, program, *sp, sss[100];
>> >  sp = sss;
>> >  program = fp[-1];
>> >  objects = SCM_I_VECTOR_WELTS (SCM_PROGRAM_OBJTABLE (program));
>> >  free= SCM_PROGRAM_FREE_VARIABLES(program);
>> >  /* compiled section */
>> > 
>> >  LCASE3666:
>> > goto L3668;
>> > LCASE3665:
>> > sp++; *sp = LOCAL_REF(1);
>> > sp++; *sp = SCM_I_MAKINUM(0);
>> > sp-=2; if(!scm_is_eq(sp[1],sp[2])) goto L3669;
>> > sp++; *sp = LOCAL_REF(2);
>> 
>> [...]
>> 
>> Comparing this to the VM’s bytecode interpreter should show the overhead
>> incurred by instruction dispatch.
>> 
>> Did you try to measure this?  That’d be interesting.
>> 
>> Thanks,
>> Ludo’.
>
> Yes on one of my machine one loop takes about 6ns for the compiled one and 
> about 50ns for the bytcode, say a factor of 7-10 for simple tasks, but 
> remember
> this is for rather trivial work.

Woow, I didn’t expect so much.

Ludo’.

Re: GLIL->C compilation

[Stefan Israelsson Tampe]

On Saturday, November 20, 2010 12:11:16 am Ludovic Courtès wrote:
> Hi,
> 
> Stefan Israelsson Tampe  writes:
> > #include "cfkn.h"
> > #define LOCAL_REF(i) fp[i]
> > #define LOCAL_SET(i,v)   fp[i] = (v)
> > void cfkn3676(SCM *fp, SCM **spp)
> > {
> > 
> >  /* setup of environment */
> >  SCM *objects, *free, program, *sp, sss[100];
> >  sp = sss;
> >  program = fp[-1];
> >  objects = SCM_I_VECTOR_WELTS (SCM_PROGRAM_OBJTABLE (program));
> >  free= SCM_PROGRAM_FREE_VARIABLES(program);
> >  /* compiled section */
> > 
> >  LCASE3666:
> > goto L3668;
> > LCASE3665:
> > sp++; *sp = LOCAL_REF(1);
> > sp++; *sp = SCM_I_MAKINUM(0);
> > sp-=2; if(!scm_is_eq(sp[1],sp[2])) goto L3669;
> > sp++; *sp = LOCAL_REF(2);
> 
> [...]
> 
> Comparing this to the VM’s bytecode interpreter should show the overhead
> incurred by instruction dispatch.
> 
> Did you try to measure this?  That’d be interesting.
> 
> Thanks,
> Ludo’.

Yes on one of my machine one loop takes about 6ns for the compiled one and 
about 50ns for the bytcode, say a factor of 7-10 for simple tasks, but 
remember
this is for rather trivial work.

GLIL->C compilation

[Ludovic Courtès]

Hi,

Stefan Israelsson Tampe  writes:

> #include "cfkn.h"
> #define LOCAL_REF(i) fp[i]
> #define LOCAL_SET(i,v)   fp[i] = (v)
> void cfkn3676(SCM *fp, SCM **spp)
> {
>  /* setup of environment */
>  SCM *objects, *free, program, *sp, sss[100];
>  sp = sss;
>  program = fp[-1];
>  objects = SCM_I_VECTOR_WELTS (SCM_PROGRAM_OBJTABLE (program));
>  free= SCM_PROGRAM_FREE_VARIABLES(program);
>  /* compiled section */
>  LCASE3666:
>
>
> goto L3668;
> LCASE3665:
> sp++; *sp = LOCAL_REF(1);
> sp++; *sp = SCM_I_MAKINUM(0);
> sp-=2; if(!scm_is_eq(sp[1],sp[2])) goto L3669;
> sp++; *sp = LOCAL_REF(2);

[...]

Comparing this to the VM’s bytecode interpreter should show the overhead
incurred by instruction dispatch.

Did you try to measure this?  That’d be interesting.

Thanks,
Ludo’.