l1a/l1b: PARTIAL (missing null/abstract checks); l2: FAIL (unsupported)

[opencode-agent]

Spec compliance report

Compiler	Status	Issues
l1a	⚠️ PARTIAL	Missing null check on objectref before vtable dispatch; missing AbstractMethodError check for abstract methods
l1b	⚠️ PARTIAL	Missing null check on objectref before vtable dispatch; missing AbstractMethodError check for abstract methods
l2	❌ FAIL	L2ByteCodeSupportChecker marks invokevirtual as unsupported (line 195-196), preventing L2 compilation entirely; code generation exists but is never used

---

Details

l1a Implementation

• Status: ⚠️ PARTIAL
• Files: core/src/core/org/jnode/vm/x86/compiler/l1a/X86BytecodeVisitor.java:2809-2861
• Issues:
  1. Missing NullPointerException check (JVM Spec: "Otherwise, if objectref is null, the invokevirtual instruction throws a NullPointerException"): The vtable dispatch path at lines 2846-2858 directly dereferences objectref to load the TIB (os.writeMOV(asize, helper.AAX, helper.AAX, tibOffset)) without first testing for null. If objectref is null, this causes a segmentation fault rather than throwing NullPointerException.
  2. Missing AbstractMethodError check (JVM Spec: "Otherwise, if the selected method is abstract, invokevirtual throws an AbstractMethodError"): No verification that the resolved method is not abstract before invoking via vtable. The vtable may contain entries for abstract methods (cloned for abstract classes per TIBBuilder), but the dispatch code assumes concrete implementations.
  3. Correctly handles: IncompatibleClassChangeError for static methods (line 2813-2816), method resolution via constant pool, fast path for final/private/final-class methods, proper return value handling via helper.pushReturnValue() (line 2857), correct operand stack effect (pops objectref + args, pushes return value).

l1b Implementation

• Status: ⚠️ PARTIAL
• Files: core/src/core/org/jnode/vm/x86/compiler/l1b/X86BytecodeVisitor.java:3489-3549
• Issues:
  1. Missing NullPointerException check: Identical to l1a, lines 3534-3546 dereference objectref without null test.
  2. Missing AbstractMethodError check: Same as l1a, no check for abstract method before vtable dispatch.
  3. Correctly handles: Same correct behaviors as l1a (static method check, fast path, return value handling, stack effect).

l2 Implementation

• Status: ❌ FAIL
• Files:
  • core/src/core/org/jnode/vm/x86/compiler/l2/L2ByteCodeSupportChecker.java:195-196 — explicitly marks invokevirtual as unsupported
  • core/src/core/org/jnode/vm/x86/compiler/l2/GenericX86CodeGenerator.java:5508-5577 (VirtualCallAssignQuad) and 5580-5637 (VirtualCallQuad) — code generation exists but is unreachable
  • core/src/core/org/jnode/vm/compiler/ir/IRGenerator.java:1022-1079 — IR generation correctly models operand stack effect
• Issues:
  1. L2 support checker rejects invokevirtual: L2ByteCodeSupportChecker.visit_invokevirtual() at line 195-196 calls notSupported() which throws UnsupportedOperationException. This prevents any method containing invokevirtual from being compiled by L2, forcing fallback to L1.
  2. Missing NullPointerException check in codegen: The generated code in GenericX86CodeGenerator (lines 5550-5562 / 5622-5634) has the same null dereference issue as l1a/l1b.
  3. Missing AbstractMethodError check in codegen: Same as l1a/l1b.
  4. IR generation is correct: IRGenerator.visit_invokevirtual() properly models the operand stack effect (pops objectref + args, pushes return value for non-void methods) and creates VirtualCallQuad/VirtualCallAssignQuad with correct variable offsets.

---

JVM Spec References (invokevirtual, SE6)

Format: invokevirtual indexbyte1 indexbyte2 (opcode 182 / 0xb6)

Operand Stack: ..., objectref, [arg1, [arg2 ...]] → ... (return value pushed if non-void)

Linking Exceptions: IncompatibleClassChangeError if resolved method is static

Runtime Exceptions:

• NullPointerException if objectref is null
• AbstractMethodError if no matching method selected OR selected method is abstract
• UnsatisfiedLinkError if native method code cannot be bound

Notes: The nargs argument values and objectref are not one-to-one with the first nargs+1 local variables; long/double take two slots.

Closes #406

opencode session  |  github run