Newsgroups: php.internals
Path: news.php.net
Xref: news.php.net php.internals:12950
Mailing-List: contact internals-help@lists.php.net; run by ezmlm
Message-ID: <5.1.0.14.2.20040923085826.03796c40@localhost>
Date: Thu, 23 Sep 2004 09:06:18 -0700
To: Thies C. Arntzen <thieso@gmail.com>,internals@lists.php.net
In-Reply-To: <ciu1lm$to4$1@sea.gmane.org>
References: <5.1.0.14.2.20040922110625.05297540@localhost>
Mime-Version: 1.0
Content-Type: text/plain; charset="us-ascii"; format=flowed
Subject: Re: [PHP-DEV] New VM architecture
From: andi@zend.com (Andi Gutmans)

At 10:31 AM 9/23/2004 +0200, Thies C. Arntzen wrote:
>On 2004-09-22 20:09:47 +0200, Andi Gutmans <andi@zend.com> said:
>
>>Guys,
>>There's no feedback about the new VM architecture and it's hard to 
>>maintain the patches as changes are being made to the source base constantly.
>>I will recommit the patches on Monday unless someone has reasonable 
>>objections. I reverted the patch so that people had time to review but it 
>>doesn't seem to have happened eventhough I've sent out two reminders.
>>In general, this patch improves performance by about 20% and supports 
>>goto opcode threading. I don't see a very good reason not to commit it.
>
>andi,
>
>i really think it's time to write some engine-level benchmarks (loops etc) 
>so we can actually see the improvement.

We have something quite synthetic (loops and function calls) so I wouldn't 
give it too much weight but I'll commit it to the Zend CVS.

>even though i really think that the VM-loop should be generated and not 
>done thru cpp-magic i think committing it is a "step forward" .remember, 
>sterling and i had a patch for this (& more) more than a year ago...

Yeah, I actually also did this patch in the past (the threading patch) but 
threw it away because I didn't want to hard code it, especially as many 
compilers don't work with it. Now we thought of an automated way of doing 
it where we can also easily change the execution architecture. That is the 
reason it has waited for so long, because we never had the right 
implementation.

What's next? Well I'd like to get the realpath() into PHP. We've played 
around with taking the function handler execution architecture and 
serializing it into machine code but it doesn't really improve performance 
over the threaded goto.
The main problem also is that when you check the non-synthetic benchmarks 
but real life apps like php-nuke or other apps, then the bottlenecks are 
usually elsewhere.

Andi