I don't know how to interpret the radio silence from the list, but this seems like an obvious and good follow-on to the previous RFC and has my support.
I don't know how to interpret the radio silence from the list, but this seems like an obvious and good follow-on to the previous RFC and has my support.
We also interpreted the lack of responses as “this is an obvious follow-up and I don't have any questions or concerns”, especially since FCC support also came up in the discussion of the Closure RFC and was split into a separate RFC primarily to give each of these features the proper attention with regard to implementation and edge cases. In hindsight this was the correct decision, since the implementation for FCC was much more complicated and edge-casey than initially expected, whereas Closures were reasonably straight forward.
On Fri, Jan 31, 2025, at 5:03 AM, Tim Düsterhus wrote:
Hi
Am 2025-01-31 00:58, schrieb Larry Garfield:
I don't know how to interpret the radio silence from the list, but this
seems like an obvious and good follow-on to the previous RFC and has my
support.
We also interpreted the lack of responses as “this is an obvious
follow-up and I don't have any questions or concerns”, especially since
FCC support also came up in the discussion of the Closure RFC and was
split into a separate RFC primarily to give each of these features the
proper attention with regard to implementation and edge cases. In
hindsight this was the correct decision, since the implementation for
FCC was much more complicated and edge-casey than initially expected,
whereas Closures were reasonably straight forward.
Best regards
Tim Düsterhus
Purely out of curiosity and for educational value, what was so complicated about it? I would have expected it to be straightforward.
Purely out of curiosity and for educational value, what was so complicated about it? I would have expected it to be straightforward.
FCC internally are effectively treated as function calls with a funny parameter list. This means that all valid ways to perform a function call are valid ways to create a FCC. And likewise do all constraints of a function call also apply to creating a function call. FCCs in const-expr need to replicate those semantics as applicable. For us this means: A FCC in a parameter default value behaves, as if the assignment was the first line in the function body. Likewise must a FCC in a property default value behave as if it was assigned in the first line of the constructor body.
And then amongst others you have the following (edge-)cases to take into account:
- All kinds of expressions can be used as a function name (this RFC excludes those for simplicity).
- A FCC can resolve to a dynamic function when __call() or __callStatic() is implemented (this RFC excludes those for simplicity).
- Static methods can directly be called on a trait (this is deprecated, but supported by the RFC).
- Calls to free-standing functions need to take the global namespace fallback into account (which will be cached for each individual call, which has funny effects when the function in the same namespace gets defined after some of the calls already executed at least once).
- Visibility and names needs to be resolved correctly, including references to `self` and `static`.
For regular closures (i.e. the first RFC) it was as simple as: Disallow `use` and require `static` and you're done. They are much less context-dependent.
With the two weeks discussion time elapsing later today and given the feedback we received, we plan to open the vote tomorrow pending any major concerns.
My only concern with the RFC is something a bit silly. If I understand correctly, this will be a constant expression:
You understood that correctly.
It is an incredibly small inconsistency (imho), but other than that, ship it.
I wouldn't call it an inconsistency, it's natural that the allowed expressions in a constant-expression context are limited. This limitation is mentioned in the second bullet point of the “Constraints” section: Only the "basic form" of first-class callables is supported, mainly to keep the implementation simple, due to the limited usefulness of the more complex function call variants when limited to constant-expressions. As an "escape hatch", it's always possible to wrote a proper closure that just forwards the calls to the desired method.
It will also result in a useful error message that is consistent with dynamic new-expressions within const-expr context:
Fatal error: Cannot use dynamic function name in constant expression in test5.php on line 3
and for new-expressions:
$ cat test5.php
<?php
const Foo = new (Foo)();
$ sapi/cli/php test5.php
Fatal error: Cannot use dynamic class name in constant expression in test5.php on line 3
