Re: struct optimizations

[Ludovic Courtès]

Hello,

Andy Wingo <address@hidden> writes:

> Some comments:
>
>     -#define SCM_VTABLE_FLAG_RESERVED_0 (1L << 5)
>     -#define SCM_VTABLE_FLAG_RESERVED_1 (1L << 6)
>     +#define SCM_VTABLE_FLAG_SIMPLE (1L << 5) /* instances of this vtable 
> have only "pr" fields */
>     +#define SCM_VTABLE_FLAG_SIMPLE_RW (1L << 6) /* instances of this vtable 
> have only "pw" fields */
>
> The comments do not appear to be correct, given a later check:
>
>     +  if (SCM_LIKELY (SCM_VTABLE_FLAG_IS_SET (vtable, SCM_VTABLE_FLAG_SIMPLE)
>     +                   && SCM_VTABLE_FLAG_IS_SET (vtable, 
> SCM_VTABLE_FLAG_SIMPLE_RW)
>     +                   && p < SCM_STRUCT_DATA_REF (vtable, 
> scm_vtable_index_size)))
>     +    /* The fast path: HANDLE is a struct with only "p" fields.  */
>     +    data[p] = SCM_UNPACK (val);
>
> It seems that currently SIMPLE is for all pr *or* all pw slots. But we
> should be more orthogonal than that; let's have SIMPLE be for having all
> slots be readable p slots. (Note that this still allows a mix of
> readable and writable slots.)

Fixed:

  
http://git.sv.gnu.org/cgit/guile.git/commit/?id=e03b7f73e2927178f2d9485320435edb6260c311

> Then perhaps we can change SIMPLE_RW to be MUTABLE or something, to
> indicate that all slots of this object are mutable.

I ended up keeping the name ‘SIMPLE_RW’ because in theory there could be
non-‘p’ mutable fields, i.e., non-simple structs with mutable fields,
which aren’t interesting.

Thanks,
Ludo’.