l1a/l1b: PASS; L2: FAIL (LREM unimplemented)

lrem_spec_compliance_report.md

@@ -0,0 +1,92 @@
+## Spec compliance report
+
+| Compiler | Status | Issues |
+|----------|--------|--------|
+| l1a | ✅ PASS | None |
+| l1b | ✅ PASS | None |
+| l2 | ❌ FAIL | LREM (and all long arithmetic ops) not implemented in GenericX86CodeGenerator |
+
+---
+
+## Details
+
+### l1a Implementation
+- **Status:** ✅ PASS
+- **Files:** 
+  - `core/src/core/org/jnode/vm/x86/compiler/l1a/X86BytecodeVisitor.java:3371`
+- **Issues:** None
+
+**32-bit mode (lines 3372-3380):** Delegates to `context.getLremMethod()` which resolves to `MathSupport.lrem()`. 
+
+`MathSupport.lrem()` (core/src/core/org/jnode/vm/MathSupport.java:98-149) correctly implements the JVM spec:
+- Handles `Long.MIN_VALUE` edge cases per spec identity `(a/b)*b + (a%b) == a`
+- Explicitly checks `den == 0` and throws `ArithmeticException("Divide by zero")` at line 126-127
+- Result sign matches dividend (value1) per spec
+
+**64-bit mode (lines 3381-3412):** Uses x86 `IDIV` instruction directly:
+- Loads dividend into `RAX`, sign-extends to `RDX:RAX` via `CDQ`
+- Executes `IDIV` with divisor
+- Remainder ends up in `RDX` (pre-claimed as result register)
+- Hardware raises `#DE` (divide error) on division by zero
+- VM exception handler (`VmThread.java:1391-1393`) catches `EX_DIV0` and throws `ArithmeticException("Division by zero at address ...")`
+
+**Operand stack effect:** Correctly pops two `LongItem` (category-2), pushes one `LongItem` (category-2). Uses `DoubleWordItem` for 32-bit (two GPRs) and single `GPR64` for 64-bit.
+
+---
+
+### l1b Implementation
+- **Status:** ✅ PASS
+- **Files:** 
+  - `core/src/core/org/jnode/vm/x86/compiler/l1b/X86BytecodeVisitor.java:4212`
+- **Issues:** None
+
+**Implementation is identical to l1a** with same 32-bit/64-bit split and same correctness properties. Only difference is additional counter instrumentation.
+
+---
+
+### l2 Implementation
+- **Status:** ❌ FAIL
+- **Files:**
+  - `core/src/core/org/jnode/vm/x86/compiler/l2/L2ByteCodeSupportChecker.java` — does NOT reject `lrem`
+  - `core/src/core/org/jnode/vm/x86/compiler/l2/GenericX86CodeGenerator.java` — all 16 `LREM` case labels (lines 920, 1083, 1266, 1481, 1666, 1859, 2045, 2207, 2384, 2541, 2688, 2862, 3016, 3210, 3399, 3634) fall through to:
+    ```java
+    case LADD:
+    case LAND:
+    case LDIV:
+    case LMUL:
+    case LOR:
+    case LREM:
+    case LSHL:
+    case LSHR:
+    case LSUB:
+    case LUSHR:
+    case LXOR:
+    default:
+        throw new IllegalArgumentException("Unknown operation: " + operation);
+    ```
+- **Issues:** 
+  1. **Complete lack of LREM support** — any method containing `lrem` will fail L2 compilation with `IllegalArgumentException`
+  2. **L2ByteCodeSupportChecker doesn't filter it out** — `canCompile()` returns `true` for methods with `lrem`, but code generation fails
+  3. **All long arithmetic ops missing** — same issue affects `ldiv`, `lmul`, `ladd`, `lsub`, `land`, `lor`, `lxor`, `lshl`, `lshr`, `lushr`, `lneg`, `lcmp`
+
+---
+
+## JVM Spec References
+
+- **JVMS SE6 §lrem**: https://docs.oracle.com/javase/specs/jvms/se6/html/Instructions2.doc8.html
+- **Operation**: `result = value1 - (value1 / value2) * value2`
+- **Operand Stack**: `..., value1, value2 → ..., result` (both `long`, category-2)
+- **Runtime Exception**: `ArithmeticException` if `value2 == 0`
+- **Identity**: `(a/b)*b + (a%b) == a` (holds for `MIN_VALUE / -1` where remainder is 0)
+
+---
+
+## Verification Notes
+
+1. **Divide-by-zero handling**: Both L1 compilers correctly produce `ArithmeticException` — 32-bit via explicit check in `MathSupport.lrem()`, 64-bit via x86 `#DE` exception translated by VM.
+
+2. **Edge case `Long.MIN_VALUE % -1`**: `MathSupport.lrem()` handles this at lines 99-108, returning `0` per spec identity.
+
+3. **Category-2 stack handling**: Both L1 compilers use `DoubleWordItem` / `LongItem` correctly for 64-bit values on virtual stack.
+
+4. **L2 gap is systemic**: The entire `long`/`double` arithmetic category is unimplemented in L2 code generator (`GenericX86CodeGenerator`), not just `lrem`.