Re: RFA: Optimization to write_register_bytes()

[Andrew Cagney <ac131313 at cygnus dot com>]

Fernando Nasser wrote:
> 
> This patch includes the changes requested by Andrew Cagney.
> 
> Here is the new version, with ChangeLog for approval.
> 
> Thanks.
> Fernando
> 
> 2000-06-09  Fernando Nasser  <fnasser@totem.to.cygnus.com>
> 
>         * findvar.c (write_register_bytes): Cut loop short after the
>         register that contains the last bytes in the range is written.
>         This makes things a little faster and also prevents that we try to
>         write to alias registers.  Also, we do not write if contents have
>         not changed.
> 

> *************** write_register_bytes (myregstart, myaddr
> *** 848,854 ****
> 
>         /* Is this register completely within the range the user is writing?  */
>         else if (myregstart <= regstart && regend <= myregend)
> !       write_register_gen (regno, myaddr + (regstart - myregstart));
> 
>         /* The register partially overlaps the range being written.  */
>         else
> --- 856,866 ----
> 
>         /* Is this register completely within the range the user is writing?  */
>         else if (myregstart <= regstart && regend <= myregend)
> !         {
> !         write_register_gen (regno, myaddr + (regstart - myregstart));
> !         if (regend >= myregend)
> !           break;
> !         }

FYI, I've seen nothing that demonstrates that this change is safe :-( 
Unless someone examines every register layout for every target we've
forced to assume otherwize :-(

I'm not even 100% sure of the change I suggested - keep track of which
bytes have been written so that when the number remaining gets to zero
exit.  There could easily be a target that has two physical registers
mapped onto a single region - it may expect separate writes for each.

>         /* The register partially overlaps the range being written.  */
>         else
> *************** write_register_bytes (myregstart, myaddr
> *** 863,873 ****
> --- 875,895 ----
>              Update it from the target before scribbling on it.  */
>           read_register_gen (regno, regbuf);
> 
> +         /* If new value == old value, then don't bother doing the
> +            actual store. */
> +         if (memcmp (registers + overlapstart,
> +                 myaddr + (overlapstart - myregstart),
> +                 overlapend - overlapstart) == 0)
> +           continue;
> +

Here, why clone write_register_gen() why not just call it?  We're trying
to reduce the number of places where people blat that register buffer
not increase them.

Sorry,

	Andrew