If we perform a class fetch that is not marked as exception safe,
convert exceptions thrown by autoloaders into a fatal error.
Ideally fetching the interfaces would be exception safe, but as it
isn't right now, we must abort at this point.
If a PHP file contains an invalid hex literal such as `0x_10`, the expected error
is `Parse error: syntax error, unexpected 'x_10' (T_STRING) in %s on line %d`.
This already worked correctly on Linux, but on Windows prior to this patch a different
error was produced: `Parse error: Invalid numeric literal in %s on line %d`.
HT_FLAGS() includes the full flag word, including the iterator
count. When we're fully reassigning it, we need to make sure that
we either really do want to copy the iterator count (as in some
cases in array.c) or we need to mask only the actual flag byte.
Add an assert to hash_iterators_del() to make sure the iterator
count is non-zero (which is how I ran into this) and make sure that
the iterator count is correctly preserved during array splicing.
When a HT iterator is one past the end and we rehash, we need to make
sure that it is move to the new one past the end position, to make
sure that newly inserted elements are picked up.
In the hash position APIs, make sure we always advance to the next
non-undef element and not just when the position is 0 (similar to
what foreach does). This can happen when the position of an
ArrayIterator is one past its current end and a new element is
inserted not directly at that position because the array is packed.
There is still a bug here (as shown in the tests), but this is a
separate issue that also affects plain array iteration in foreach.
When a connection is closed, we also need to remove the hash entry
from the regular_list, as it now points to freed memory. To do this
store a reverse mapping from the connection to the hash string.
It would be nicer to introduce a wrapping structure for the pgsql
link resource that could store the hash (and notices), but that would
require large changes to the extension, so I'm going for a more
minimal fix here.
We only need to do this once we're running destructors. The current
approach interferes with some event loop code that runs everything
inside a shutdown function.
This might happen if OBJ_RELEASE is used on an object that was already
released by GC. Specific cases of this issue were previously fixed in
ffaee27478 and
72104d2b6e, however the issue still
affects 3rd-party extensions using OBJ_RELEASE.
The whole GC type NULL + OBJ_IS_VALID + IS_FREE_CALLED system seems
overly complicated and can probably be simplified in 7.4.
There have been multiple reports of large slowdowns due to the
use of MADV_HUGEPAGE, so make it conditional on
USE_ZEND_ALLOC_HUGE_PAGES, just like MAP_HUGETLB already is.
Nikita Popov