On 09/12/2016 11:40 AM, Fleshgrinder wrote:
On 9/11/2016 10:04 PM, Larry Garfield wrote:
On 09/10/2016 05:55 AM, Fleshgrinder wrote:
@Larry what do you think about the CoW proposal?
I'm afraid I got lost in the thread somewhere with the back and forth
about implementation details. Can you repost the current proposal, or
has the RFC page been updated with the latest CoW-based recommendation?
--Larry Garfield
No worries, here is the original message of mine in response to you (but
I guess that I did not include you in the recipients):
https://marc.info/?l=php-internals&m=147318929726088&w=2
Ah, I did see that one, but there was a lot of discussion after it so I
thought the idea evolved. Response below based on skimming the
responses after that as well...
It sounds like there's more that needs to go on, though. It sounds like
that thread is suggesting that $this in a method of an immutable object
is always cloned, which seems excessive to me.
The point about identity is well-taken. However, it's more complex on
objects because equality is not as simple as it is on strings. Two
objects can be logically identical but physically different. For instance:
class HttpHeaders {
public $attributes = [];
}
$a = new HttpHeaders();
$a->attributes['foo'] = 'bar';
$a->attributes['baz'] = 'buzz';
$b = new HttpHeaders();
$b->attributes['baz'] = 'buzz';
$b->attributes['foo'] = 'bar';
$a and $b are now not physically identical, since their attributes
array is in a different order. However, header order doesn't matter in
HTTP, or rather isn't supposed to. (Modulo buggy implementations, of
course.) So are $a and $b identical? I could very easily argue both
directions on that. Physical identity would be easier to automate
checking in the engine; logical identity would require user-space code
to make such decisions.
From the discussion of "transformer" methods, I'd propose a slightly
different keyword. To wit:
1) A class marked as "immutable" may not have any of its properties
altered, EXCEPT in certain unlocked scopes. The constructor is an
unlocked scope.
2) A method may be marked "clone": public clone function foo() {}. A
clone method is identical to any other method except that A) $this in
its scope is not the original object, but the result of clone($this); B)
A clone method is an unlocked scope, so modifying $this (the clone) is
legal. That is more self-descriptive than "transformer", and also
doesn't require a new keyword. (By implication, clone and static are
mutually exclusive since clone requires an object to clone.)
I don't know that there are any other unlocked scopes to consider...
#2 does leave us with the identity question that Richard raises, in that
returning an unmodified $this from a clone method still creates a new
object identity. However, I posit that is to be expected. In Python,
eg, while some built-in objects are immutable they are not, necessarily,
always the same physical object in memory. (999+1 is not the same
object as 1000, but 1+2 and 2 likely will be due to internal engine
optimizations.) You need to do a value comparison of them.
I don't see the identity case being resolved without an __identity()
method, or similar. Which could be useful in its own right so I'm not
necessarily against it, but it's an extra, and I'd argue optional, piece
of the puzzle.
Related to whether or not the properties of an object may be mutable
(arrays, other objects, etc.), they would in practice probably need to
blacklist resources. We ran into that issue in PSR-7, where the body
stream has to be mutable, because streams. Since PSR-7 is based on PHP
5.3 that doesn't cause any syntactic issues, just logical issues. If
the classes were marked immutable explicitly, it likely would be a
problem since the streams cannot be immutable, thus they can't be used
on an immutable object. Which is... very sad making. :-(
--Larry Garfield
--
PHP Internals - PHP Runtime Development Mailing List
To unsubscribe, visit: http://www.php.net/unsub.php
