[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Imporved cons representatoin in guile

From: Stefan Israelsson Tampe
Subject: Re: Imporved cons representatoin in guile
Date: Fri, 10 Jul 2015 13:37:53 +0200

I managed to push the send button prematurey,

Here is the code i think that would be needed
   scm_t_bits val = SCM_UNPACK(*pt) & SCM/2; (Mask out the first half)
   if(1 & val == 0) //THIN
          if(val & 6 == 0) //non emediate e.g. a pointer
                int64 delta = promote_signed_ptr(val);
                return UNREF(pt + delta);
          if(val & 2 == 1)   // integer
                SCM_PACK(return promote_signed(val))
           return SCM_PACK(val)
     fat version of SCM comes here.

To note is that indeed we do get a more complex code here. But on the other hand the extra logic is bit twiggelin and compiled using the cpu registers only at most
a SCM_CAR will take twice the time. A SCM_CDR on the other hand need to take in two SCM and is probably as fast or faster. 

On Fri, Jul 10, 2015 at 1:20 PM, Stefan Israelsson Tampe <address@hidden> wrote:

I was complentating the cons cells in guile an was wondering if we could pack our cons cells better than today. So this is some notes about this.

Currently if x is a cons cell we have the relation

This is really neat and makes guiles conses quite compact e.g. a vector y of two elements is
y -> [SCMTAG,SCMV1,SCMV2], where SCMTAG containes the datatype tag and length of the

To compress even further we need a way to could use
x ->[SCM/2/SCM/2], witt SCM/2 the same tagging half the size as the normal SCM with the interpretation that if
SCM/2 is a non emediate then it starts with 00 and is then interpreted as a signed integer i and the real adress is x + i, e.g. a relative adress regarding.

We of cause must also add a fat cons cell of the form
x -> [Tag,X,Y] for the case when SCM/2 is not fitting. Currently I can't see this beeing common. But
if we later makes floating point represented via nan boxing e.g. stored directly in a 64bit value then
the cons cell will be mostly fat and there would be a speed reduction using cons cells. On the other hand there has been a considerable speed.

There is a final sematic case that needs to be fixed. if we do a setcar on a thin cons cell and the cell
then becomes fat, we need to create the following

x -> oldthin -> newfat

e.g. we need to add a pointer type with the meaning of automatically follow the pointer if we encounter it. then oldthin is also tagging a variant of a cons cell. It is possible to keep it slim
in the code that all fat cons cells is represented like that.

How would a SCM_CAR be like?



reply via email to

[Prev in Thread] Current Thread [Next in Thread]