Skip folding branches of unreachable If

[tlively]

OptimizeInstructions generally skips optimizing unreachable code, leaving that to DCE instead. But it did not have this check before folding identical instructions out of If arms. The fuzzer found a case where cont.new instructions were hoisted out of an unreachable If but not refinalized, resulting in the concretely-typed cont.new having an unreachable child. This caused an assertion failure in the validator.

Fix the root problem by skipping this optimization for unreachable Ifs, but also fix the validator so that it will no longer crash on such invalid IR.

[stevenfontanella]

The tests don't mention cont.new, what's the case that the fuzzer found? Should we check that in?

[tlively]

The fuzzer's test case looks like this:

(if (unreachable)
  (then (cont.new $k (ref.func $a))
  (else (cont.new $k (ref.func $b))
)

It was previously optimized to this:

(cont.new $k ;; should have type `unreachable`, but does not.
  (if (unreachable)
    (then (ref.func $a))
    (else (ref.func $b))
  )
)

We could add this as a test, but the fix really had nothing to do with cont.new in particular. I suppose I could add a unit test for the cont.new validation fix, though.