Improve support of static initialization blocks and statically-reachable memory

[eupp]

Currently, if a Lincheck test execution both:

• reads some values from statically-reachable memory, and
• modifies the content of the same memory during the test

it can lead to non-determinism error (the Lincheck would not be able to replay the execution trace, because the content of the static memory changed).

[Above by statically-reachable memory we mean any mutable state reachable from static variables.]

Note that during the test statically-reachable memory can be modified both by

• static initialization blocks (<clinit>),
• and the regular functions from the test.

Thus, in order to fix possible non-determinism errors, we need to track modifications performed by <clinit> blocks, save them into trace, and manually re-perform them during replay (in the same order as in the original execution).

[dmitrii-artuhov]

Todo list for proper support of the static memory restoring:

• Static memory snapshot collecting and restoring #418 the leaking this problem: the code below will not work with the lazy snapshot tracking, because detection of this's fields modifications in constructors is implemented incorrectly (I merely check if methodName == "<init>" && fieldOwnerClass == transformingClass then the bytecode instructions are ignored. The correct way would be to check what obj owns the modified fields. And if obj == this we should only then skip these instructions, otherwise track them. However, I don't see how that can be achieved right now, because of Bytecode verification error after instrumentation of constructors (<init> methods) because of leaking this #424

  class LincheckTest {
     class Node(var a: Int) {
         constructor(other: Node) : this(2) { // constructor accepting the same type
             other.a = 0 // in case if other is static variable, the change should be tracked
                         // because `other != this` here, but in my solution it will be missed.
         }
     }
     
     companion object {
         val staticNode = Node(1)
     }
     
     @Operation
     fun modify() {
         val localNode = Node(staticNode) // here will be a write to staticNode.a = 0, which is missed
     }
   }
  

• Static memory snapshot collecting and restoring #418 in the constructor calls instrumentation there should be check for sub-classing and not the pure equals. Right now there is no obvious way of doing so during class transformation without loading the class hierarchy (which is discouraged during transformation stage)
• Static memory snapshot: add support for Unsafe, VarHandle, and AtomicFieldUpdater #429 java afu objects: method calls on atomic field updators (eg. kotlinx.atomicfu.AtomicInt::getAndIncrement() it will compile to a call on java afu object which references plain int and getAndIncrement is not considered as modification)
• Static memory snapshot: add support for Unsafe, VarHandle, and AtomicFieldUpdater #429 modification via unsafe calls
• Static memory snapshot: add support for Unsafe, VarHandle, and AtomicFieldUpdater #429 modification via var-handle calls
• Add System.arraycopy(...) call support #469 System.arraycopy(...) this call is not considered as modification, so overridden values will not be remembered. This call is used in ArrayList::remove(...) for example. It prevents from collections lazy tracking (known issue with its support: Jit optimizations affect inserted method calls during execution for System.arraycopy interceptions #470)

[eupp]

Removed the subtask

• referencing fields and modifiying their values via reflection

as it is not specific to static memory restoring logic, and is out of scope

[ndkoval]

@dmitrii-artuhov let's for now ignore tests that fail on recent JVMs and fix them later

[eupp]

I've added separate issue #554 for arraycopy support, and closing this one.