Fix save-analysis generation crash with invalid tuple access

[sinkuu]

Reproduction:

fn invalid_tuple_struct_accessing() {
    bar.0;
}

error[E0425]: cannot find value `bar` in this scope
 --> test.rs:2:5
  |
2 |     bar.0;
  |     ^^^ not found in this scope

error[E0601]: main function not found

error: internal compiler error: librustc_save_analysis/dump_visitor.rs:1678: Expected struct or tuple type, found TyError
 --> test.rs:2:5
  |
2 |     bar.0;
  |     ^^^^^

thread 'rustc' panicked at 'Box<Any>', librustc_errors/lib.rs:482:9
note: Run with `RUST_BACKTRACE=1` for a backtrace.

This should fix a crash in RLS when editing such code. cc @nrc

[rust-highfive]

r? @estebank

(rust_highfive has picked a reviewer for you, use r? to override)

[nrc]

r? @nrc

[nrc]

When you're using the invalid tuple constructor is this ty:err or something similarly specific? Could we explicitly catch that case and return, and still span_bug ICE if it's something else? If we silently fail and cope then it's really hard to notice the omission until somebody reports a hard to discover bug against the RLS.

[sinkuu]

It seems that almost all type variants (all but TyInfer?) can be crafted.

fn foo<T>(x: T) {
    bar.0; // TyError
    (0).0; // TyError
    x.0; // TyParam
    static S: i32 = 0;
    (S).0; // TyInt
    [].0; // TyArray
    extern { type Externed; }
    (|| -> Externed { panic!() })().0; // TyForeign
    (|| -> ! { panic!() })().0; // TyNever
    ...
}

[nrc]

OK, makes sense. Second commit looks fine.

@bors: r+

[bors]

📌 Commit f5a3efe has been approved by nrc