This goes in the reverse direction of 4463acb9513dfb62206760c49b3da1fe4d92f40a. After looking around a bit, it seems that we already check for Z_ISERROR_P() on the get_property_ptr_ptr return value in other places. So do this in zend_fetch_property_address() as well, and also make sure that EG(error_zval) is indeed returned on exception in get_property_ptr_ptr. In particular, this fixes the duplicate exceptions that we used to get because first get_property_ptr_ptr threw one and then read_property throws the same exception again.
Relying on setting ERROR if an exception happened during the property address fetch is both a bit fragile and may pessimize other codepaths that will check for exceptions in the VM. Adding an extra exception check instead, which should also allow us to drop the use of ERROR in this area in master.
This is a fix for symfony/symfony#32995. The behavior is: * Throwing exception when loading parent/interface is allowed (and we will also throw one if the class is simply not found). * If this happens, the bucket key for the class is reset, so it's possibly to try registering the same class again. * However, if the class has already been used due to a variance obligation, the exception is upgraded to a fatal error, as we cannot safely unregister the class stub anymore.
RFC: https://wiki.php.net/rfc/deprecate_curly_braces_array_access
No notice is thrown for list() accesses, because we did not come
to an agreement regarding patterns like
while ([$key, $value] = yield $it->next()) { ... }
where silent null access may be desirable.
No effort is made to suppress multiple notices in access chains
likes $x[0][0][0], because the technical complexity this causes
does not seem worthwhile.
RFC: https://wiki.php.net/rfc/notice-for-non-valid-array-container
Another stab in the dark to fix the intermittent failures of timeout tests on macos CI: We're using ITIMER_PROF, which means that the timer counts against user+system time. The "busy" wait loop counts against real time. Currently it calls microtime() on every iteration. If that call is implemented as a syscall rather than going through vDSO or commpage we might be seeing many context switches here which drive up the real time, but not user or system time. See if making the loop busier and calling microtime() less helps the situation.