IRGen: Don't assert on keypaths that use a subclass as root

[aschwaighofer]

class A {
var x : Int = 0
}

class B : A {}

_ = \B.x

rdar://46562046

[aschwaighofer]

@swift-ci Please test

[aschwaighofer]

@jckarter Is that the right way to fix this?

[aschwaighofer]

getClassFieldAccess expects the class type that contains the field - it does not handle subclasses

[slavapestov]

Does it ever have a specifier type? (inout or lvalue here)? Might be better to assert.

[slavapestov]

This feels like something SILGen should have already done? CC @jckarter

[jckarter]

That might be cleaner, yeah. Then we can assert the condition in the SILVerifier too.

[slavapestov]

This loop is already encapsulated in TypeBase::getSuperclassForDecl()

[jckarter]

This SGTM, though I agree with Slava that having SIL always reference the property by its originating superclass is a little bit cleaner.

[aschwaighofer]

Consider the following example:

$ cat KP3.swift
class A {
let x : Int = 0
}

class B : A {}

class C {
  let b: B = B()
}

func test() {
 _ = \C.b.x
}

Today the key path is

keypath $KeyPath<C, Int>, (root $C; stored_property #C.b : $B; stored_property #A.x : $Int)

If we change SILGen this would be:

keypath $KeyPath<C, Int>, (root $C; stored_property #C.b : $A; stored_property #A.x : $Int)

Which would seem odd to me that the type for the var decl:#C.b disagrees with the type we assign the component: $A.

[aschwaighofer]

@swift-ci Please test

[swift-ci]

Build failed
Swift Test Linux Platform
Git Sha - 94a9a9a

[swift-ci]

Build failed
Swift Test OS X Platform
Git Sha - 94a9a9a

[jckarter]

Ah, I misunderstood, I thought that the stored_property key had been ending up as #B.x. In that case, I think your fix is correct.

[slavapestov]

@aschwaighofer Thanks for the cleanup! If this same pattern comes up in more places, we could add a SILType::getSuperclassDecl() variant too, but it's not too bad as-is.