[AST] Improve ErrorType handling in TypeBase::getMemberSubstitutions()

[jtbandes]

Historical note: a similar change was made in 0033bda.

Fixes 78 compiler_crashers. ⛵

[lattner]

This LGTM, but I think @DougGregor should also take a look when he gets a chance. Great work, thank you!

-Chris

[AlexDenisov]

Indeed it solves some issues, but the problem actually caused by this line:

baseTy = baseTy->castTo<NominalType>()->getParent();

The best way to fix those crashes, imho, would be to use safer API for casting:

if (const NominalType *NominalTy = baseTy->getAs<NominalType>()) {
  baseTy = NominalTy->getParent();
} else {
  baseTy = nullptr;
}

otherwise some other non-NominalType and non-ErrorType will break the compiler.

[jtbandes]

@AlexDenisov thanks for the input. I would assume the castTo<NominalType> is being done for good reason (i.e. is safe because of other invariants upheld by the AST code), but I don't have enough knowledge to say for sure. @DougGregor will need to confirm.

Really makes me wish this were all written in Swift, so we could use optional chaining and throws 😄

[jtbandes]

Updated with new fixed crashers.

[DougGregor]

I'd rather that the ErrorType check happen a little later, e.g., turn the "castAs()" into a "getAs()" within an "if". Then, we only check the ErrorType condition if all else fails, and can null out baseTy before breaking out of the loop.

[jtbandes]

@DougGregor: Updated. (Note that some tests are failing for other reasons; I fixed them in #296.)

I have general questions about the error-handling mechanisms throughout; I sent an email to swift-dev entitled "Problems caused by is", and I also filed https://bugs.swift.org/browse/SR-86. I'd be curious to hear your thoughts.

[jtbandes]

@DougGregor please let me know if there's anything else to be done here!

[DougGregor]

The getAs will either produce a NominalType or null; there's no need to check for ErrorType here.