Edit report at https://bugs.php.net/bug.php?id=50894&edit=1
ID: 50894
Comment by: koubel at volny dot cz
Reported by: lee at projectmastermind dot com
Summary: no-op cast triggers copy-on-write (ternary operator
triggers copy too)
Status: Assigned
Type: Feature/Change Request
Package: Performance problem
Operating System: *
PHP Version: 5.*, 6
Assigned To: dmitry
Block user comment: N
Private report: N
New Comment:
Optimization for ternary operator will be very helpful, it was discussed also
in internals with some patch suggestion -
http://www.mail-archive.com/[email protected]/msg51926.html. Copy always
isn't necessary for ternary operator anytime I think.
Previous Comments:
------------------------------------------------------------------------
[2010-02-04 06:00:21] lee at projectmastermind dot com
Thanks for your time. I appreciate all the work you guys do - I know
it's not easy.
I'm amending this report to include the ternary operator too. Turns
out that it has exactly the same issue as the cast (presumably for the
same reason that Dmitry explained).
$a = array();
$b = true ? $a : $a; // forced copy happens here.
Whether we call these bugs, or optimization requests -- the fact is,
they appear to the developer as "highly unexpected behavior". And
they're out there right now causing real performance issues in real
php-based systems.
Do people really use casting and ternary in ways that will cause
problems? Yes they do...
The following types of statements are actually very common...
1)
$a = (array)$a;
2)
$a = is_array($a) ? $a : array();
3)
$a = (somecondition) ? $z1 : $z2;
A quick grep over the Zend Framework and the PEAR source trees (for
example) finds extensive use of casting and ternary in assignments
involving large primitive types (like array and string) -- and in
plenty of those cases it's clear that the author never intended a copy
to happen unless it was needed for type conversion.
I agree that adding new operations to address this (eg: ASSIGN+CAST)
wouldn't make much sense.
I haven't had time to dig through the Zend code in detail to see how
feasible this would be... but consider the following...
Would it be possible to change the two operation handlers in question
to return a zval pointer (or a handle) instead of forcing them to
always return a temporary? If that's possible, then the handler could
decide internally whether to allocate a temporary, or to just return
the operand directly.
This sort of thing is already done with, for example, the assignment
operator on chained assignments:
<?php
/* works as expected - only lazy copies are made */
$a = array();
$c = $b = $a;
?>
------------------------------------------------------------------------
[2010-02-02 09:40:51] [email protected]
It definitely not a bug, but a request for optimization.
In current implementation, the result of "$b = (array)$a;" can't be a simple
assignment operation, because in general the result of type-cast operation is
not a variable but temporary value. Such temporary value can't be created using
copy-on-write and requires variable copying. On the following assignment opcode
this value is used as is (without copying).
It's possible to optimize this situation by creation of additional opcode which
combines ZEND_CAST+ZEND_ASSIGN, but I don't think it makes a lot of sense.
------------------------------------------------------------------------
[2010-02-01 06:13:33] lee at projectmastermind dot com
Description:
------------
given a value with a particular type, casting it to that same type
should essentially be a no-op -- once it is determined that the
operand already has the correct type, no further action needs to be
taken.
Ex:
$a = array();
$b = (array)$a;
In this example, $a is already an array, so this should be a simple
assignment operation. $b should get a "lazy" copy of $a via PHP's
copy-on-write policy. Instead, the cast operation seems to force an
immediate (non-lazy) full copy.
This creates a huge potential for hidden performance problems, as it
causes code that *looks* like it would run in constant time [O(1)] to
actually require linear time [O(n)] (where n represents the size of
the data being copied).
I have verified that this issue does exist for string types as well.
I assume that it applies to all PHP types.
Of course it becomes a significant performance issue primarily for
types that can hold large amounts of data, where the data is
duplicated whenever the zval is duplicated (AFAIK, this is only string
and array).
I have verified this on the following versions of php:
5.2.6
5.2.8
6.0.0-dev (php6.0-201001312130)
Reproduce code:
---------------
<?php
for( $z=1; $z<5; ++$z ) {
$a = array_fill(0, 100*$z, '0');
$t_start = microtime(true);
for($i=0;$i<100000;++$i) {
// O(n) [should be constant time, but isn't]
// cast triggers non-lazy copy
//
$b = (array)$a;
// O(1) [constant time, as expected]
// (comment above, and uncomment here for comparison)
//
//$b = $a;
}
$t_elapsed = (microtime(true)*1000)-($t_start*1000);
printf(
"(%d elements * %d copies): %f ms\n\n",
100*$z, $i, $t_elapsed
);
}
Expected result:
----------------
(100 elements * 100000 loops): 11.264160 ms
(200 elements * 100000 loops): 11.363037 ms
(300 elements * 100000 loops): 11.208984 ms
(400 elements * 100000 loops): 11.809082 ms
NOTE: the time stays roughly constant as the number of elements
increases -- the assignments are copy-on-write, so no significant
performance hit is incurred.
Actual result:
--------------
(100 elements * 100000 copies): 736.453613 ms
(200 elements * 100000 copies): 1448.991211 ms
(300 elements * 100000 copies): 2130.541016 ms
(400 elements * 100000 copies): 2823.362793 ms
NOTE: the time increases as the size of the array increases. (This
happens with large strings too). This is a good indicator that a copy
is being made [non-lazily] when the cast is applied.
------------------------------------------------------------------------
--
Edit this bug report at https://bugs.php.net/bug.php?id=50894&edit=1
