Convert ObjectLiteralExpr to not use tc.callWitness() or generate a SemanticExpr.

[pschuh]

For reference, all other literal types except dictionaries and string
interpolation have been converted over to this form.

[slavapestov]

Hey, I’m sorry I missed this! Do you mind rebasing to fix the merge conflict?

[pschuh]

@swift-ci Please test

[swift-ci]

Build failed
Swift Test Linux Platform
Git Sha - f94d07d6932bad48eb6fbf5eafe200b9373ebcb9

[swift-ci]

Build failed
Swift Test OS X Platform
Git Sha - f94d07d6932bad48eb6fbf5eafe200b9373ebcb9

[pschuh]

@swift-ci please test

[swift-ci]

Build failed
Swift Test OS X Platform
Git Sha - 3854bf2e65563431fb40a5576c500e8b97b3a1b5

[swift-ci]

Build failed
Swift Test Linux Platform
Git Sha - 3854bf2e65563431fb40a5576c500e8b97b3a1b5

[pschuh]

@swift-ci please test

[swift-ci]

Build failed
Swift Test OS X Platform
Git Sha - 3854bf2e65563431fb40a5576c500e8b97b3a1b5

[swift-ci]

Build failed
Swift Test Linux Platform
Git Sha - 3854bf2e65563431fb40a5576c500e8b97b3a1b5

[pschuh]

@slavapestov

I'm able to reproduce the mac failure (or a failure that looks similar), It is this pattern of protocols and classes:

class Color { required init() {} }

protocol Silly {
  init()
  init(_colorLiteralRed red: Float, green: Float, blue: Float, alpha: Float)
}

extension Silly {
  init(_colorLiteralRed red: Float, green: Float, blue: Float, alpha: Float) {
    self.init()
  }
}

extension Color : Silly, _ExpressibleByColorLiteral {}

func makeColorLiteral() -> Color {
  return #colorLiteral(red: 1.358, green: -0.074, blue: -0.012, alpha: 1.0)
}

print(makeColorLiteral())

This results in the following substitution map: (note the τ_0_0 that shows up out of nowhere and the inherited_conformance)

(substitution_map generic_signature=<Self where Self : Silly>
  (substitution Self -> τ_0_0)
  (conformance type=Self
    (inherited_conformance type=τ_0_0 protocol=Silly
      (normal_conformance type=Color protocol=Silly
        (value req=init() witness=test.(file).Color.init()@test.swift:3:12)
        (value req=init(_colorLiteralRed:green:blue:alpha:) witness=test.(file).Silly extension.init(_colorLiteralRed:green:blue:alpha:)@test.swift:12:3)))))

What really baffles me is that if I swap everything out for ArrayLiteral (which uses identical conformance lookup logic), I get the following substitution map: (which has no extra type parameter)

(substitution_map generic_signature=<Self where Self : Silly>
  (substitution Self -> Color)
  (conformance type=Self
    (normal_conformance type=Color protocol=Silly
      (value req=init() witness=test2.(file).Color.init()@test2.swift:3:12)
      (value req=init(arrayLiteral:) witness=test2.(file).Silly extension.init(arrayLiteral:)@test2.swift:12:3))))

This all happens in the CSApply.cpp code and happens in the conformance.getWitnessByName(type, name); snippet.

(normal_conformance type=Color protocol=_ExpressibleByColorLiteral
  (value req=init(_colorLiteralRed:green:blue:alpha:) witness=test.(file).Silly extension.init(_colorLiteralRed:green:blue:alpha:)@test.swift:12:3))

[slavapestov]

I don't think we should be building an InheritedProtocolConformance for a class-constrained type parameter like that. Do you want me to take a look?

[pschuh]

Yes, please.

[slavapestov]

The problem is that the substitutions stored inside the witness itself are meant for SILGen to use when emitting a witness thunk. They're not always in terms of the generic signature of the type in the conformance. In fact this only happens when the 'synthetic environment' is sufficiently simple, but the two non-trivial cases are when the requirement is generic, or when the witness is in a protocol extension and the conforming type is a class. I think this will break for any literal protocol where a witness is in a protocol extension and the literal type is a class. However, it worked for arrays because of another workaround elsewhere.

I have a tentative fix in #25350.