type error in unreachable code on block-sigs branch

[sunfishcode]

The block-sigs branch has this testcase in loop.wast:

(assert_invalid
  (module (func $type-cont-last-void-vs-empty (result i32)
    (loop (br 0 (nop)))
  ))
  "type mismatch"
)

The end of the function is unreachable, so implementations shouldn't be required to issue a type error here, by my understanding of the rules.

[sunfishcode]

type-cont-last-num-vs-empty also in loop.wast is another instance of this.

[ghost]

With the block signatures, the loop has a declared result type, in this case void so the validation can easily use that, it does not need to propagate unreachable types. This seems consistent with (i32.add (f32.neg (unreachable))) being a type error too, here f32.neg has a declared type which is used even though it's result is unreachable.

This also seems to be related to WebAssembly/design#778

[rossberg]

@sunfishcode, acknowledged. Block signatures notwithstanding, the interpreter still implements the don't-type-check-dead-code semantics, which is why I didn't find these when I adapted the tests. Not skipping dead code is gonna be a separate change.

[rossberg]

This has been fixed with #345.