Re: Note on 109327

[Dmitry Antipov <dmantipov@yandex.ru>]

On 08/01/2012 01:47 AM, Thien-Thi Nguyen wrote:

>     > I think to do this well you will need separate macros for getting
>     > and setting.
>
>     Sure, but it's almost impossible to do this at once.
>
> Why?

As an exercise, try this for 'contents' member of Lisp_Vector.

>
>     At the very beginning, it's possible to "overestimate" barrier
>     assuming that each XVAR (obj, field) changes FIELD in OBJ; in the
>     future, reads and writes may be separated, thus giving a precise
>     write barrier.
>
> I think you're saying that you prefer to do:
>
> a1. substitute object access (whether LHS or RHS) w/ XVAR (obj, field)
> a2. distinguish LHS (which could benefit from optimization) from RHS
> a3. substitue LHS ‘XVAR (...) = value’ w/ SETVAR (obj, field, value)
>
> instead of:
>
> b1. distinguish LHS (which could benefit from optimization) from RHS
> b2. substitute LHS object access w/ SETVAR (obj, field, value)
> b3. substitute RHS object access w/ XVAR (obj, field)
>
> Is my understanding correct?

a1) is definitely the first, since this gives a base to design an
"overestimated" write barrier (each XVAR (obj, field) issues the
barrier). Next, it should be possible to  implement generational
pass for GC (which collects only an objects changed since the last
collection). This is expected to be very inefficient because write
barrier is hugely overestimated. Finally, it's time for a3) or b2),
i.e. removing the barrier from XVAR (obj, field) and installing it
only at XSETVAR (obj, field, value), thus reducing "overestimation"
of the barrier.

Dmitry