Exhaustively test OptimizeInstructions on ref.cast

[tlively]

Write a script to systematically generate test cases for all possible
interesting cast patterns. Update the code to correctly handle the case
where the cast is known to succeed but still requires a cast to recover
exactness.

[kripken]

With this automation, it is not practical to read the testcase and each of the CHECKS to see the outcome (and in particular, there are no human-written comments explaining each outcome). That makes sense here. But then what the test gives us is IIUC checking for assertions in the pass and validation errors in the output. Can we also do runtime checks, by executing the code with --fuzz-exec which would verify behavior remains the same after the optimization?

That would require adding more stuff so the code is fully runnable, but seems practical?

[tlively]

It's a good idea, but I don't think it's feasible right now because our interpreter does not yet use exact types. If you tried to interpret (ref.cast $foo (struct.new $foo)), it would currently fail. Once we have landed the changes to type allocations as exact, we would be able to revisit this.

[kripken]

I see, sounds good.