Not NULLing the static_variables pointer for shadow methods during
static var shutdown would be a way to avoid this leak, but unless
there's evidence that inherited private methods with static vars are
actually a common use-case, I don't think we should keep this kind
of fragile edge-case optimization.
Fixes OSS-Fuzz #17875.
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.
The requirements for parent/interface are difference than for the
variance checks in type declarations. The latter can work on fully
unlinked classes, but the former need inheritance to be essentially
finished, only variance checks may still be outstanding.
Adding a new flag for this because we have lots of space, but we
could also represent these "inheritance states" more compactly in
the future.
symtable_cache_ptr now points to the first unused symtable_cache
entry, rahter than the last used one. This avoids taking a pointer
to the minus first element of the array, which is UB. Instead we
take a pointer to the end plus one, which is not UB.
Keep track of delayed variance obligations and check them after
linking a class is otherwise finished. Obligations may either be
unresolved method compatibility (because the necessecary classes
aren't available yet) or open parent/interface dependencies. The
latter occur because we allow the use of not fully linked classes
as parents/interfaces now.
An important aspect of the implementation is we do not require
classes involved in variance checks to be fully linked in order for
the class to be fully linked. Because the involved types do have to
exist in the class table (as partially linked classes) and we do
check these for correct variance, we have the guarantee that either
those classes will successfully link lateron or generate an error,
but there is no way to actually use them until that point and as
such no possibility of violating the variance contract. This is
important because it ensures that a class declaration always either
errors or will produce an immediately usable class afterwards --
there are no cases where the finalization of the class declaration
has to be delayed until a later time, as earlier variants of this
patch did.
Because variance checks deal with classes in various stages of
linking, we need to use a special instanceof implementation that
supports this, and also introduce finer-grained flags that tell us
which parts have been linked already and which haven't.
Class autoloading for variance checks is delayed into a separate
stage after the class is otherwise linked and before delayed
variance obligations are processed. This separation is needed to
handle cases like A extends B extends C, where B is the autoload
root, but C is required to check variance. This could end up
loading C while the class structure of B is in an inconsistent
state.
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.
As discussed in bug #77580, passing INVALID_HANDLE_VALUE for the
completion event improves compatibility with Wine/ReactOS. The timer
callback itself is supposed to complete fast enough, no behavior change
is to expect.
