What about call_user_func_array() not going through SEND_ARRAY (e.g. in namespace)? I think this code doesn't handle it yet.

Regular call_user_func_array() is changed in the same way.

Ah, sorry, I missed part of the patch. However, I don't think those implementations do the same thing wrt how the argument is fetched. E.g. call_user_func_array('func', $undefined) will throw a notice for $undefined, while the same in a namespace will not. Similarly for call_user_func_array('func', $array[$key]) in one case this will do a FETCH_DIM_R and in the other (effectively) a FETCH_DIM_W. (Unless I misremember how PREFER_REF handling works, which may well be...)

You are right. Probably it's not very hard to fix this (using _W mode for IS_VAR/IS_CV arrays).

What do you think about this solution in general.
Is it better or worse than yours? (I don't have strong opinion).

The changes to the tests are quite what I expect. Travis reports a segfault for these tests though...

@nikic Right, but it will throw a type error or at least emit a warning+zpp failure for undefined instead here? [but yeah, should never emit a notice then]

All mentioned problems should be fixed now.

@dstogov About this approach in general, I'm not totally sure. One the one hand, this is a "pragmatic" solution, in that it usually will do what you want. However, this will probably introduce a few gotchas. For example, it is very common that the arguments passed to call_user_func_array() actually come from somewhere else, e.g.

function doCall($args) {
    call_user_func_array($fnWithRefArgs, $args);
}

Currently, if $args is not correctly prepared with references, this will generate a warning. After this change, the call will be allowed, silently, but the passed arguments will stay the same due to the intermediate by-value pass of $args.

In any case, I don't have a strong opinion either way. I think it may be good to discuss what to do here on the mailing list, as this is a larger issue.

@nikic After all, I would prefer your #2059
It breaks few tests, but I hope, it shouldn't be a big problem to fix them.

Closed in favor to #2088