Parameter type narrowing should consider function overload signatures

TypeScript Version: 2.2.1

Code

interface A { x: number; }
interface B { y: number; }

function foo(b: B): void;
function foo(a: A, b: B): void;
function foo(a: A|B, b?: B): void {
  if(b !== void 0) {
    console.log(a.y); // Error: Property 'y' does not exist on type 'A | B'
  }
}

The third (unified) function signature is hidden from the call site, and therefore any arguments should be able to be narrowed based on inspection of the provided arguments. In the case above, the second argument is only defined if the first argument is of type A, and so I would expect narrowing of the argument type to occur accordingly. One could make the argument that in the compiled output, it's possible for JavaScript to still pass subsequent arguments, but then why narrow types at all, seeing as JavaScript can always violate every type constraint that TypeScript makes inferences from?

Start with the TypeScript 2.2.1 reproduction in the issue and inspect how overload signatures relate to the unified implementation signature during parameter narrowing. Verify the current diagnostic for a.y, then determine whether the desired behavior is consistent with the stated overloads. Done means the example narrows the argument as expected without weakening unrelated type checks.