Re: smaller globs

On Tue, Oct 27, 2009 at 01:19:49PM +0000, Zefram wrote:
> Thinking about the const sub space optimisation.  I think we can do
> something similar more broadly.  Observe that most GVs only contain one
> type of item.  In these cases, the GP structure is wasteful.  We could
> optimise one-item GVs by referencing the item directly through svu_rv,
> instead of having a GP, and using about three flag bits to indicate which
> type of item it is (scalar, array, hash, IO, format, sub, const sub).
> The SVpad_ flag bits are conveniently available to reuse for this.

Are they? I thought that GVs can be stored in pads.

#define SVs_PADSTALE	0x00010000  /* lexical has gone out of scope */
#define SVpad_STATE	0x00010000  /* pad name is a "state" var */
#define SVs_PADTMP	0x00020000  /* in use as tmp */
#define SVpad_TYPED	0x00020000  /* pad name is a Typed Lexical */
#define SVs_PADMY	0x00040000  /* in use a "my" variable */
#define SVpad_OUR	0x00040000  /* pad name is "our" instead of "my" */

> Obviously, this compact structure would have to upgrade itself to a full
> GV with metadata and GP, if it's actually referenced as a GV.  This is

At that point, it can't know its line or file?

struct gp {
    SV *	gp_sv;		/* scalar value */
    struct io *	gp_io;		/* filehandle value */
    CV *	gp_cv;		/* subroutine value */
    U32		gp_cvgen;	/* generational validity of cached gv_cv */
    U32		gp_refcnt;	/* how many globs point to this? */
    HV *	gp_hv;		/* hash value */
    AV *	gp_av;		/* array value */
    CV *	gp_form;	/* format value */
    GV *	gp_egv;		/* effective gv, if *glob */
    line_t	gp_line;	/* line first declared at (for -w) */
    HEK *	gp_file_hek;	/* file first declared in (for -w) */
};

Offhand I can't remember where those are used.

> Sound sane?  If it's appealing, I'll investigate more closely where we
> can actually dispense with the metadata.

I'm not sure if it feels sane. Certainly, I'd firstly like to know where the
metadata ends up being read again.

Nicholas Clark

• smaller globs by Zefram
• Re: smaller globs by Nicholas Clark
• Re: smaller globs by Zefram
• Re: smaller globs by Nicholas Clark