#70497: v2: Compile out HANDLE_BLOCK_INTERRUPTIONS when unused #1532
Conversation
|
The patch doesn't respect zend-signal semantic with HAVE_APACHE. On Tue, Sep 29, 2015 at 9:53 AM, Daniel Axtens notifications@github.com
|
|
Sorry, my mistake. I will update the patch tomorrow to reorder the ifdefs to keep the same semantic. |
This does the same thing as 127db5a ("HANDLE_BLOCK_INTERRUPTIONS()
is not used by SAPIs anymore. It may be useful only when PHP configured
with --enable-zend-signals.")
In the PHP7 version, the code could be dropped easily because no SAPI
used the hooks. However, in PHP5, the Apache 1 SAPIs use these hooks.
Some other SAPIs provided NULL pointers.
With this patch, when PHP is compiled we first check for the Apache 1
SAPI, and if it is present, preserve the old behaviour. (Both flavours
of the Apache 1 SAPI define HAVE_APACHE, so use that.) It also preserves
the behaviour in the ZEND_SIGNALS case. Only if neither of those are
present do we compile out the definitions.
This leads to a speed up on PHP5 code. I've measured the performance
impact on a PowerPC 64-bit little-endian system and on a x86_64 laptop
using the following script:
<?php
for ($i=1; $i < 100000000; $i++) {
$genericObject = new stdClass();
$a = (object) 8;
$b = (object) 'abcde';
}
?>
I see a ~6% improvement in performance on POWER and a ~4% improvement
on x86_64.
I also see an improvement in a e-commerce web application on POWER
with this change.
Signed-off-by: Daniel Axtens <dja@axtens.net>
---
This is version 3.
v2 -> v3:
- Preserve ZEND_SIGNALS semantic.
v1 -> v2:
- Do not change behaviour under Apache 1.
|
#ifdefs reordered as discussed. Apologies for my ignorance of the PHP release process, but who is the 5.6 RM? |
|
https://wiki.php.net/todo/php56 On Wed, Sep 30, 2015 at 9:35 AM, Daniel Axtens notifications@github.com
|
|
I don't think this worth changing in a micro release, there could be some other SAPIs (packaged by downstream) which depend on the existence of these macros. |
|
I agree with @Tyrael . Our 5.6 is stable, and we should not change it, but have changes go in PHP7 source base. |
|
Well, would it be acceptable to include a change that doesn't change the On Wed, Oct 14, 2015 at 11:21 PM, jpauli notifications@github.com wrote:
|
|
UNEXPECTED() will work. On Thu, Oct 15, 2015 at 1:45 AM, Daniel Axtens notifications@github.com
|
|
UNEXPECTED makes sense |
|
OK, I tried with UNEXPECTED(). Unfortunately that doesn't give a measurable performance benefit. I think this is because we still have to load the pointer to the globals struct, dereference it, load the function pointer from the struct and check the final pointer against NULL. Once we do all that UNEXPECTED makes us go faster, but it doesn't save us doing that work to begin with. I'll continue to think about a non-intrusive way of doing this, but it looks like UNEXPECTED won't do it. |
|
Closing this PR as the RM decision seems to be not to make this change in the 5.6 branch. Also ref https://bugs.launchpad.net/ubuntu/+source/php5/+bug/1527105. |
This does the same thing as 127db5a ("HANDLE_BLOCK_INTERRUPTIONS()
is not used by SAPIs anymore. It may be useful only when PHP configured
with --enable-zend-signals.")
In the PHP7 version, the code could be dropped easily because no SAPI
used the hooks. However, in PHP5, the Apache 1 SAPIs use these hooks.
Some other SAPIs provided NULL pointers.
With this patch, when PHP is compiled we first check for the Apache 1
SAPI, and if it is present, preserve the old behaviour. (Both flavours
of the Apache 1 SAPI define HAVE_APACHE, so use that.) It also preserves
the behaviour in the ZEND_SIGNALS case. Only if neither of those are
present do we compile out the definitions.
This leads to a speed up on PHP5 code. I've measured the performance
impact on a PowerPC 64-bit little-endian system and on a x86_64 laptop
using the following script:
I see a ~6% improvement in performance on POWER and a ~4% improvement
on x86_64.
I also see an improvement in a e-commerce web application on POWER
with this change.
Signed-off-by: Daniel Axtens dja@axtens.net
This is version 2.
Changes in this version from version 1 (#1520)