Remove incorrect handling of covariant return types

[ogolberg]

This removes incorrect covariant return type handling introduced in MANIMALSNIFFER-49. The ticket is no longer publicly accessible, so I lack the context of the original change.

The block of code removed in this PR prevented animalsniffer from catching covariant changes to return types which were compile-time compatible but not binary-compatible.

For example, if we take the code below

ByteBuffer buffer = ...
buffer.rewind();

and compile it on Java 8, we'll get bytecode which invokes

java.nio.ByteBuffer.rewind()Ljava/nio/Bufffer

If we compile it on Java 9+, while still targeting Java 8, we will get bytecode which invokes

java.nio.ByteBuffer.rewind()Ljava/nio/ByteBuffer

That's because this method with the more specific ByteBuffer return type is available in JDK9+.

Now, if we run the latter binary on Java 8, we will get a NoSuchMethodError.

A recent example of this in the wild: grpc/grpc-java#6839 (note that grpc-java uses animalnsniffer).

[ogolberg]

I also propose cutting 1.16.1 with this fix because many projects out there are using 1.16

[lukehutch]

I have been bitten by this several times (#77).

Does removing the check allow animal-sniffer to catch the problem?

[ogolberg]

Does removing the check allow animal-sniffer to catch the problem?

Correct. Covariant return types are purely a compile-time construct, so this check is extraneous.

[lukehutch]

@ogolberg No it's not just a compiletime construct -- your code will compile fine with a covariant change in return type (in particular if you ignore the return value), but then can fail at runtime with an exception.

[ogolberg]

@lukehutch yes, I think we're on the same page here.

What I mean is that the compiler does all the work to handle covariant return types (https://docs.oracle.com/javase/tutorial/java/generics/bridgeMethods.html#bridgeMethods), there's nothing special in the bytecode/vm to make them work; therefore, the special casing that I am removing in this PR is unnecessary.

[lukehutch]

But Animal Sniffer should fail and stop compilation if a covariant return type is detected. The current code detects the change but doesn't stop the build with an error. How does removing the check altogether address this problem?

[ogolberg]

the find method will fall through and return false, which eventually will bubble up as an error

[lukehutch]

But then there will be no descriptive error message given. See my PR #87 for an alternative. (I'm not sure if the logic is correct though.)

[ogolberg]

sure, additional logging may be worth the cost of running that loop code - I was just going after correctness

[ogolberg]

any chance this can be merged?

[olamy]

why removing an integration test?
If it was wrong it should be updated to reflect what the code change here because this change doesn't have any test to demonstrate the issue.

[ogolberg]

restored/updated integration tests