1. we should only do the return type checking when it is really about to
return
2. for 029.php, actually, the exception threw should be discard while it
jmp into finally(it could be observed by change the return to return an array)
3. after this fix, the test 029.phpt behavior consistently with 7.0
4. good for optimizer too
T_NUM_STRING follows the rules of symtable numeric string
conversion. If the offset isn't an integer under those rules, it
is treated as a string. This should apply to negated T_NUM_STRINGs
as well.
If this does not break the Unix system somehow, I'll be amazed. This should get most of it out, apologies for any errors this may cause on non-Windows ends which I cannot test atm.
This a partial backport of 8754b19. It
a) fixes the class/function/constant import table confusion in the
namespaced case, and
b) restricts conflict checks to a single file based on a filename
pointer comparison.
It does not fix the issues with filename reuse (e.g. due to eval)
and late-bound classes. This part of the change requires globals
changes.
This fixes the following issues:
* "use function" and "use const" inside namespaced code were checking
for conflicts against class imports. Now they always check against
the correct symbol type.
* Symbol conflicts are now always checked within a single file only.
Previously class uses inside namespaced code were checked globally.
This behavior is illegal because symbols from other files are not
visible if opcache is used, resulting in behavioral discrepancies.
Additionally this made the presence/absence of symbol errors dependent
on autoloading order, which is volatile.
* The "single file" restriction is now enforced by collecting defined
symbols inside a separate hash table. Previously it was enforced
(for the non-namespaced case) by comparing the filename of the
symbol declaration. However this is inaccurate if the same filename
is used multiple times, such as may happen if eval() is used.
* Additionally the previous approach relies on symbols being registered
at compile-time, which is not the case for late-bound classes, which
makes the behavior dependent on class declaration order, as well as
opcache (which may cause delayed early-binding).
* Lastly, conflicts are now consistently checked for conditionally
defined symbols. Previously only declaration-after-use conflicts were
checked in this case. Now use-after-declaration conflicts are
detected as well.
Otherwise we're missing the "expected to be a reference, value
given" warning that appears for ordinary calls to call_user_func().
Also update an UPGRADING note with recent changes wrt
call_user_func().
