From: pavenis AT lanet DOT lv
To: sandmann AT clio DOT rice DOT edu (Charles Sandmann), djgpp-workers AT delorie DOT com,
        pavenis AT lanet DOT lv
Date: Tue, 14 Aug 2001 19:05:21 +0300
MIME-Version: 1.0
Content-type: text/plain; charset=US-ASCII
Content-transfer-encoding: 7BIT
Subject: Re: Selector Exhaustion
Message-ID: <3B797671.5443.399C9F@localhost>
In-reply-to: <10108141456.AA13336@clio.rice.edu>
References: <3B790835 DOT 21623 DOT 13164B AT localhost> from "pavenis AT lanet DOT lv" at Aug 14, 2001 11:15:01 AM
X-mailer: Pegasus Mail for Win32 (v3.12c)
Reply-To: djgpp-workers AT delorie DOT com
Errors-To: nobody AT delorie DOT com
X-Mailing-List: djgpp-workers AT delorie DOT com
X-Unsubscribes-To: listserv AT delorie DOT com
Precedence: bulk

On 14 Aug 2001, at 9:56, Charles Sandmann wrote:

> > Initially I did some timing:
> > checking all descriptors once took about 1 ms on PIII-700 (Win98SE)
> 
> This worries me a bit - this means about 1/4 of second on the 386/16
> type platform for the two checks around a call.  This could add hours 
> to a build on slower machines.

I tested that both descriptor scans is taking up to 20% of time needed
to call a very simple program that does nothing except one printf 
(output of selectors it's using) and fits in disk cache. I think than in  
real programs the the difference will be bigger.

I tested that on PIII 700 (about 2ms to scan descriptors twice and 10 
ms to run a executable) and on a slow 486DX2 66 (about 15ms and 
170 ms).

> I was hoping to be smarter about the area scanned.  For example,
> allocate 1 selector to get a starting point, then scan 50 selectors
> above that point (cut time by factor of 400).  We might consider
> in-lining the lar call or making a new routine to do a bulk check?

Tried to put LAR into loop (inline assembler) and got time down from
1 ms to about 0.75 ms 

> We also can't hard code selectors (like 7 as the start in the example)
> since this is bad under non-ring-3 DPMI providers.

Perhaps we could use  _my_ds() % 7 as start value
 
> Your example looks very much like what I was working on last night :-)

We can also allocate 2 descriptors before and after spawning 
executable and free all between that was free before spawning
(it's more safe as doing it blindly)

Andris