[swift-dev] SILVerifier: Should SuperMethodInst's result type be the same type as the referenced member?

[John McCall]

> On Dec 9, 2015, at 6:09 PM, David Farler via swift-dev <swift-dev at swift.org> wrote:
> [Joe, Roman, sorry for resend, I got the e-mail group wrong]
> 
> In SuperMethodInst's verifier, we have:
> 
> require(CMI->getType() == TC.getConstantType(CMI->getMember()),
>        "result type of super_method must match type of method");
> 
> I think this assumption was valid when we only allowed super_method on foreign classes, without needing to worry about reabstractions. Now that we're allowing super_method with operands of native class type which can be generic, does this check make sense anymore?
> 
> Consider the following:
> 
> class Parent<A> {
>  let x: A
>  required init(x: A) { self.x = x }
> }
> 
> class Child : Parent<String> {
>  required init(x: String) {
>    super.init(x: x)
>  }
> }
> 
> class Grandchild : Child {}
> 
> 
> Here, the vtable thunks for their initializers have respective types:
> 
> $@convention(method) <T> (@in T, @owned Base<T>) -> @owned Base<T>
> $@convention(method) (@in String, @owned Child) -> @owned Child
> $@convention(method) (@in String, @owned Grandchild) -> @owned Grandchild
> 
> 
> However, the real backing implementations have these respective types:
> 
> $@convention(method) <T> (@in T, @owned Base<T>) -> @owned Base<T>
> $@convention(method) (@owned String, @owned Child) -> @owned Child
> $@convention(method) (@owned String, @owned Grandchild) -> @owned Grandchild
> 
> 
> So, Child and Grandchild have abstraction differences because their initializers aren't generic. When I make a super_method instruction, the constant appears to always point to the backing implementation, not the thunk, so I needed to get the overridden vtable entry from the constant and I think that's reasonable. That gives me:
> 
> super_method %10 : $Child, #Base.init!initializer.1 : <T> Base<T>.Type -> (x: T) -> Base<T> , $@convention(method) <τ_0_0> (@in τ_0_0, @owned Base<τ_0_0>) -> @owned Base<τ_0_0>
> 
> and
> 
> super_method %6 : $Grandchild, #Child.init!initializer.1 : Child.Type -> (x: String) -> Child , $@convention(method) (@in String, @owned Child) -> @owned Child
> 
> 
> which look good to me.
> 
> With my changes today to fix generic substitutions of partial super methods and getting the right type from the vtable, if I disable that verifier check, devirtualization works correctly with super_method instructions.
> 
> Is this a problem with SILDeclRef or is this check simply no longer valid in the verifier? If so, I wonder what the suitable replacement check should be. Maybe something like:

I’d look at the verifier for ClassMethodInst.  It turns out that TypeConverter has a getConstantOverrideType that lowers according to the overridden abstraction pattern.

John.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.swift.org/pipermail/swift-dev/attachments/20151210/7a2d4776/attachment.html>