Re: Proposal: immediate strings

[Dmitry Antipov]

On 05/24/2012 09:17 AM, Stefan Monnier wrote:

> > +#if __GNUC__>  1 /* Any GCC still in use should support this.  */
> > +#define PACKED __attribute__((packed))
> > +#else
> > +#define PACKED
> > +#endif
>
> Let's not use "packed" structures: too much trouble.

To avoid using 'packed' attribute and speedup an inefficient access to 
u.dat.size
field, it's possible to use 'ptrdiff_t X : BITS_PER_PTRDIFF_T - 2'
bitfields - it's still wide enough to hold the values up to STRING_BYTES_BOUND.

> Actually, we can allocate all the pure strings as "immediate strings"
> (within the limits of imm.nbytes, of course), in which case we'd have
> strings with more bytes than STRING_IMM_MAX, so while this assertion may
> have been useful for debugging, we'll want to get rid of it.

I'm not sure it's worth making string code even more complicated with this.

> Is that really 100% reliable?  Why not use the `intervals' field with
> a 0xdeadbeef pointer instead?

Better solutions are welcome, never seen a crash with this; 'intervals'
field can't be used because NEXT_FREE_LISP_STRING (X) = Y will overwrite
X->intervals.

Dmitry

immstr.patch
Description: Text document