[fpc-devel] concernig Bug Id 4145: Request for Bit-level packing in MacPas mode

[Gerhard Scholz]

-----Ursprüngliche Nachricht----- 
Von: "Adriaan van Os" <fpc at microbizz.nl>
An: "FPC developers' list" <fpc-devel at lists.freepascal.org>
Gesendet: Samstag, 2. Juli 2005 04:01
> Jonas Maebe wrote:
>
> > Gerhard Scholz wrote:
> >
> >> if really someone does this, it would be nice to have it in
> >> "standard" (FPC)
> >> mode also.
> >
> > You will be able to activate it with a preprocessor switch.

I assume it would be very important to have it switchable in the source.

e.g.
  p =  packed record
            a : 0..100 ;
            b : boolean ;
         end ; { this should be "really" packed; 1 byte }

  q =  packed record
            a : 0..100 ;
            b : boolean ;
         end ; { this should be standard packed; 2 bytes }

so both versions could coexist, either by some new {$packrecords extreme} local
compilerdirectives or e.g. STRICT PACKED RECORD or something similar (please: no
new keyword!).

> >
> >> not just packed records, but also packed arrays:
> >>
> >> packed array [ 1..2 ] of 0..15 { one byte }
> >>
> >> packed array [ 1..8 ] of boolean { one byte }
> >
> > The implementation of that would be completely separate from the
> > support required for records. Also, since that is not required for
> > compatibility, that will have a lower priority (unless someone submits
> > a clean patch for it, of course)
>
> I am not trying to say that bit-level packing should have high priority
> (there are ways to work around it) but, since you mention
> compatibility, I would like to note that bit-level packed arrays are
> indeed a feature of UCSD, MPW, Think, MetroWerks, and GNU Pascal.

It came into my mind that nearly everything to integrate (for the moment I call
them "strict") strict packed records/arrays is already in the compiler: the
trick would be to see such a record or array as an object (some little inline
"preprocessor" ?) with some static read/write/index properties (should work for
objects (and maybe classes) also). The ref.pdf describes these properties only
for classes, not for objects, but I tried it and found out that the compiler
didn't read the documentation :-) , properties for objects seem to work o.k.

I assume the best point to incorporate this would be when the structure
definition has been read.

Fields, which are not packable, should stay untouched.

A working example, which should make clearer what I mean, is in the appendix.
The resulting code even looks quite compact. But of course in the ref manual it
should be stated quite clear that using bit-packing is punished by additional
execution time. On the other side there are enough situations where you have to
look into bytes and have to to all shifts/masks yourself. Then bit-packing would
be a help. Or you just have huge amounts of small data to store (the usual
example is the array [ 1..1000000 ] of boolean).

> Another issue that comes to my mind is passing components of packed
> structures as actual variable parameters. Standard Pascal doesn't allow
> it at all, but it is more practical to forbid it only if the packed
> component uses bit-packing.

The packed records or arrays, as they are implemented now, allow it to use their
fields as var params (I just tried some simple examples) obviously.

In my proposal, as a side-effect, bit-packed fields are forbidden as var-params,
but fields which are not bit-packable are allowed.

Gerhard
-------------- next part --------------
A non-text attachment was scrubbed...
Name: pckrecar.pas
Type: application/octet-stream
Size: 9182 bytes
Desc: not available
URL: <http://lists.freepascal.org/pipermail/fpc-devel/attachments/20050702/0ec77766/attachment.obj>