Wrong Jit behavior on php code

[sibest]

Try this script:

$kill); } function __get($var) { return $this->values[$var]; } function __set($var, $value) { $this->values[$var] = $value; } function subrender() { echo 'sub: '.$this->foo."\r\n"; } } class b extends a { ``` public $foo; function __construct() { parent::__construct('foo'); } function render() { echo 'render: '.$this->foo."\r\n"; } ``` } $test = new b(); $test->foo = 'lol'; $test->render(); $test->subrender(); ?>

Expected result:
render: lol
sub:lol

hhvm - Without Jit
render: lol
sub: lol

hhvm - With Jit:
render:
sub: lol

The property foo remains like in cache in the object, it is not correctly accessible using $this ($this->foo doesn't call parent::__get)

[paroski]

Thanks for reporting this issue! We've been tracking this inconsistency and hope to fix it soon.

While conforming to the standard PHP semantics is an important goal for us, for performance reasons we will probably arrive at a solution where __get() is never invoked for accessible declared properties even when the property is unset (at least in the short term).

[sibest]

Why don't add an invalidated bit flag on accessible properties? Would a cmp slow down things? Thank you

[ptarjan]

We're closing out all bugs older than 2 months. http://www.hiphop-php.com/wp/?p=575

If this is still an issue, please re-open it, and in order of goodness:

1. Give detailed repro steps
2. Write a test case in hphp/tests/quick (run it with hphp/tests/run) and send the pull request
3. Fix it in a pull request

[karpa13a]

684-jit.php from first post
karp@karp:/tmp/php/1$ php53 684-jit.php
render: lol
sub: lol
karp@karp:/tmp/php/1$ hhvm 684-jit.php
render: lol
sub: lol
karp@karp:/tmp/php/1$ hhvm -vEval.Jit=true 684-jit.php
HipHop Notice: Undefined property: b::$foo in /home/karp/tmp/php/1/684-jit.php on line 38
render:
sub: lol
karp@karp:/tmp/php/1$

HipHop VM 2.4.0-dev+2014.01.15 (rel)
Compiler: heads/master-0-g7eb3262d3d1f2b64b288b6be790cf1c9d372c9cb
Repo schema: f7219cd0dc9a4c048a17bb8ffa691653319e56e9

[scannell]

Reopening this as we have a repro. @jdelong may have some more context as I know we still have some cases where we differ with __get() if there is also a declared property of the same name that he encountered when fixing recursive __get().

[karpa13a]

and one more think about in "jit mode":
after few requests (http server mode) from 5 to 15, hhvm reports "HipHop Fatal error: Class undefined".
restart fixes problem again for 5-15 requests.
"undefined" class from namespace. php call "new $class" from same namespace, __construct function.
how to debug this?

or create separate issue?

[scannell]

Creating a new issue is good. Isolating the code that cause it would be ideal -- could be something with the we handle namespaces when the JIT is enabled. (We don't have much test coverage on this since we don't use namespaces internally.)