Remove special treatment of <init> for value types

[hzongaro]

Early prototype handling of value type classes used <init> as a special factory method name distinct from its use as an instance initialization method name. That has been dropped in favor of <vnew> as a factory method name for value type classes. This change removes the special treatment of <init> as a value type factory method name from TR_ResolvedJ9Method::isConstructor.

Fixes #14959

[hzongaro]

Daryl @0xdaryl, Annabelle @a7ehuo, may I ask you to review this pull request?

[a7ehuo]

I found a comment as below in the VM code. Do we still need special treatment for value abstract class in the JIT?

openj9/runtime/bcutil/cfreader.c

Lines 1804 to 1807 in 7a8923b

	* The spec says a method of a value class cannot be <init>. A method of an abstract class cannot be <vnew>.
	* For a value abstract class, its constructor is compiled into <init> now, so allow <init> in abstract classes.
	*/
	if (J9_IS_CLASSFILE_VALUETYPE(classfile) && J9_ARE_NO_BITS_SET(classfile->accessFlags, CFR_ACC_ABSTRACT)) {

[hzongaro]

Do we still need special treatment for value abstract class in the JIT?
. . .
* For a value abstract class, its constructor is compiled into <init> now, so allow <init> in abstract classes.

Good question! My understanding is that an abstract class won't be a value class, but it can have a subclass that is a value class. A concrete value class that is a subclass of such an abstract class won't have an <init> method, so its superclass's <init> method can't be called for it either. Similarly, the java.lang.Object.<init> method won't be called if an instance of a value class is created.

My understanding is that an abstract class's <init> method can only be called if an instance of a concrete identity subclass of the abstract class is being initialized. So that method will still be a constructor and only used in the context where a constructor for an identity class is used, so I don't believe the JIT requires any special handling for it.

[a7ehuo]

Thanks for the clarification @hzongaro!

My understanding is that an abstract class's <init> method can only be called if an instance of a concrete identity subclass of the abstract class is being initialized.

I discussed the above code with @hangshao0. The latest spec states: If the ACC_VALUE flag is set, one of the ACC_FINAL or ACC_ABSTRACT flags must also be set. Since value types can be abstract now, both <init> and <vnew> could be invoked on a VT. Should TR_ResolvedJ9Method::isConstructor check if it is abstract value type class? If it is, <init> is a value type factory method instead of a constructor?

[hzongaro]

Thanks, Annabelle @a7ehuo. In response to my comment:

My understanding is that an abstract class won't be a value class

you wrote:

The latest spec states: If the ACC_VALUE flag is set, one of the ACC_FINAL or ACC_ABSTRACT flags must also be set. Since value types can be abstract now, both <init> and <vnew> could be invoked on a VT.

Sorry about that! I was looking at an older version of the draft spec. The same draft that you quoted goes on to state this:

An instance initialization method is declared with the special name .

An instance initialization method may not be declared in a value class or an interface.

So a value class shouldn't have an <init> method. If I understand Hang's comment in

openj9/runtime/bcutil/cfreader.c

Lines 1804 to 1807 in 7a8923b

	* The spec says a method of a value class cannot be <init>. A method of an abstract class cannot be <vnew>.
	* For a value abstract class, its constructor is compiled into <init> now, so allow <init> in abstract classes.
	*/
	if (J9_IS_CLASSFILE_VALUETYPE(classfile) && J9_ARE_NO_BITS_SET(classfile->accessFlags, CFR_ACC_ABSTRACT)) {

, the prototype version javac with value types support is still generating <init> methods for abstract value classes. Hang @hangshao0, can you confirm that? That sounds like a bug.

Regardless, I think it's still safe if this change doesn't provide any special case handling for that case, because such an <init> method will never be invoked for a value class instance.

[hangshao0]

the prototype version javac with value types support is still generating <init> methods for abstract value classes. Hang @hangshao0, can you confirm that?

Yes, it is the behaviour of javac for now.

[a7ehuo]

LGTM. Thank you for all the clarification

[hzongaro]

Jenkins test sanity,extended xlinuxval jdknext

[hzongaro]

Jenkins test sanity all jdk8,jdk11,jdk17

[hzongaro]

I believe the sanity.functional failure on x86 Mac was due to a long-standing intermittent problem: issue #13577

[hzongaro]

Jenkins test sanity.functional x86-64_mac jdk17

[hzongaro]

Daryl @0xdaryl, may I ask you to review this pull request?