I had a go at fixing this issue. Please review the PR. Thanks

It's clear, that there's no method overloading of the methods with the same name in PHP, but the suggested way will in fact disallow even what works today. Shouldn't the interface be defined explicitly in this case? As semantically the examples look not controversial, in other programming languages with stronger OOP, even more advanced examples are possible. As "B extends A" doesn't imply "B implements I" indeed.

Thanks.

> As "B extends A" doesn't imply "B implements I" indeed.
PHP says it does: https://3v4l.org/ZQgJG (that PHP checks B<->I compatibility proves it)
C# says it does: http://ideone.com/AZRdRy
Java says it does: http://ideone.com/CZICMN


Longer explanation - I know you are familiar with this, but I'm writing it all out for anyone else reading.


1. method($param) vs. method($param=null)
LSP says parameters must be contravariant (parent->child) meaning one used in a child class must be the same or more permissive. $param=null is more permissive than $param as it allows the parameter to be omitted. Therefore
(a) is_a(A,I) and I::example($a, $b) therefore A::example($a, $b=null) is allowed
(b) is_a(B,I) and I::example($a, $b) therefore B::example($a, $b) is allowed
(c) is_a(B,A) and A::example($a, $b=null) therefore B::example($a, $b) is NOT allowed

According to examples #1 and #2, PHP allows (c) to happen because it only checks I->A, I->B inheritance and not A->B. It is also clear that I->A and I->B does not imply A->B.

If PHP instead checks I->A and A->B (the parent class/interface) then example #1 fails in the same way that example #2 fails. Yes this is a change, and some code that "works" today will not "work" tomorrow, however as the accepts_a() demo shows such code does not actually "work" today - the fact that nobody has noticed means they haven't written an accepts_a() function yet.

If PHP does check I->A and A->B then it implies I->B as well.


2. method() vs. method():int
LSP says return types are covariant (parent<-child); PHP requires invariance when the parent uses a return type and covariance when the parent does not. Covariance is being the same or less permissive. No type is like saying ":mixed", and int is less permissive than mixed. Therefore
(a) is_a(A,I) and I::example() therefore A::example():int is allowed
(b) is_a(B,I) and I::example() therefore B::example():string is allowed
(c) is_a(B,A) and A::example():int therefore B::example():string is NOT allowed

Again, according to examples #3 and #4, PHP allows (c) because it only checks I<-A, I<-B inheritance and not A<-B. I<-A and I<-B also does not imply A<-B.

Again, if PHP checked I<-A and A<-B then example #3 fails; existing working code does not actually work as accepts_a_expects_int() demonstrates.
And again, if PHP checks I<-A and A<-B then it implies I<-B.


The goal of changing validation to be against the parent rather than the interface is not to allow more code. It is to be more strict and to disallow some code which seems to, but does not truly, work correctly.


> the suggested way will in fact disallow even what works today
Do you have an example that does not already violate LSP?

> even more advanced examples are possible
Such as?

Yeah, I've understood your idea exactly as you've explained lately, to enforce strictness on the language level. As for usages in other languages, what i had in mind, here is a sample code in Java, but could be good in any other of C# or C++, etc.

I.java
interface I
{
        public int method(int i);
}

A.java
class A implements I
{
        public int method(int i)
        {
                return 0;
        }
}

B.java
class B extends A
{
        public String method(String i)
        {
                return "";
        }
}

A implements I, B doesn't explicitly implement it, but derives from A and additionally overloads the method. One can rephrase it in other way - "class B extends A implements I", as B already contains an instance of "method", so maybe it's even right to say it indirectly implements I, at least it ensures LSP. Clear, tihs is not possible in PHP, that's why i mentioned it right in my first sentence :) I'm not sure, what else would be required to enforce LSP on the language level in PHP, as the weak interface declaration is still a factor, for what this ticket cares. For example in PHP, one can can still do

class B implements I {
  public function example(): string { }
}

so then both A und B are perfectly an instance of I, but are incompatible indeed. I'd see an explicit interface declaration as a more robust and universal way, but a mitigation might be not that easy. There are possibly other cases. Anyway, if some particular restriction helps to fix a widely possible anti pattern sub case, it still could be considered as a sensible thing to do. Probably it's a bit wider topic than just one bug ticket.

Thanks.

Ups, there was no intention to assign this :)