dsub: L1A/L1B PARTIAL (FPU OK, SSE bug), L2 FAIL

[opencode-agent]

Spec compliance report

Compiler	Status	Issues
l1a	⚠️ PARTIAL	FPU implementation correct; SSE constant-folding uses addition instead of subtraction
l1b	⚠️ PARTIAL	Same FPCompiler implementations as L1A; inherits SSE bug
l2	❌ FAIL	dsub explicitly unsupported in support checker; all codegen paths throw "Unknown operation"

---

l1a Implementation

• Status: ⚠️ PARTIAL
• Files:
  • core/src/core/org/jnode/vm/x86/compiler/l1a/X86BytecodeVisitor.java:1571 (visit_dsub)
  • core/src/core/org/jnode/vm/x86/compiler/l1a/FPCompilerFPU.java:465 (FPU sub)
  • core/src/core/org/jnode/vm/x86/compiler/l1a/FPCompilerSSE.java:83 (SSE sub) and arithOperation:102
• Issues:
  • FPU path (default, FPCompilerFPU): Correct. Pops two double values, prepares FPU stack, emits FSUBP (value1 - value2), pushes result. Handles constant folding correctly (fpv1 - fpv2).
  • SSE path (FPCompilerSSE): Bug at FPCompilerSSE.java:102. The arithOperation method is shared by add and sub. For constant folding it computes fpv1 + fpv2 regardless of operation. When sub(JvmType.DOUBLE) calls arithOperation(type, X86Operation.SSE_SUB, false), the SSE instruction emitted is correct (SUBSD), but constant folding produces the wrong result (addition instead of subtraction).

l1b Implementation

• Status: ⚠️ PARTIAL
• Files:
  • core/src/core/org/jnode/vm/x86/compiler/l1b/X86BytecodeVisitor.java:1816 (visit_dsub)
  • Uses same FPCompiler hierarchy as L1A (defaults to FPCompilerFPU at line 300)
• Issues: Identical to L1A. The FPU path works correctly; the SSE path (if enabled) has the same constant-folding bug.

l2 Implementation

• Status: ❌ FAIL
• Files:
  • core/src/core/org/jnode/vm/x86/compiler/l2/L2ByteCodeSupportChecker.java:330 — visit_dsub() calls notSupported(), throwing UnsupportedOperationException
  • core/src/core/org/jnode/vm/x86/compiler/l2/GenericX86CodeGenerator.java:865,1033,1211,1426,1616,1772,1995,2157,2334,2496,2638,2812,2971,3130,3354,3541 — Every generateBinaryOP overload has a case DSUB: that throws IllegalArgumentException("Unknown operation: " + operation)
  • core/src/core/org/jnode/vm/x86/compiler/l2/FPX86CodeGenerator.java — Only handles compare operations (FCMPG, FCMPL, DCMPG, DCMPL); no arithmetic operations
  • core/src/core/org/jnode/vm/x86/compiler/l2/X86CodeGenerator.java — Thin wrapper, no overrides
• Issues: L2 completely lacks support for dsub. Any method containing dsub will fail L2 compilation and fall back to L1.

---

JVM Spec Reference (dsub)

• Spec: https://docs.oracle.com/javase/specs/jvms/se6/html/Instructions2.doc3.html
• Operand Stack: ..., value1, value2 → ..., result
• Operation: result = value1 - value2 (both double, IEEE 754 arithmetic)
• Special Cases: NaN → NaN; infinities of opposite sign → NaN; infinities of same sign → NaN; zero of opposite signs → +0.0; no runtime exceptions ever thrown.
• Category: Category 2 (double-word) computational type.

Summary

• L1A/L1B: Functional for the default FPU code path. SSE path has a latent bug in constant folding that would produce incorrect results for dsub with constant operands when SSE is enabled.
• L2: Not implemented at all; methods with dsub cannot be compiled by L2.

Closes #253

opencode session  |  github run