Loader: stubs lack .eh_frame / unwind info — debugger and exception unwinding broken through stubs

[scc-tw]

Summary

The three StubEmitter implementations (X86_64, X86_32, ARM64) generate machine code only — no DWARF CFI (Call Frame Information) entries. The injected .vmpilot section has no corresponding .eh_frame (ELF) or __unwind_info (Mach-O).

Priority: Post-first-release. v1 error policy is abort() — no C++ exceptions cross the stub boundary. This becomes critical in v2 when NATIVE_CALL targets may throw.

Impact

Scenario	Result
Debugger backtrace	gdb/lldb cannot unwind through the stub; bt terminates or shows garbage frames
C++ exception through native call	If a NATIVE_CALL target throws, the unwinder finds no FDE for the stub PC, falls back to heuristics, reads wrong RSP offset → std::terminate() or segfault
Profiler (perf, VTune)	Stack sampling breaks at stub → incomplete profiles
Crash dump analysis	Coredump/minidump cannot reconstruct the call chain through the stub

Technical Details

Each stub's stack frame (x86_64 example):

[ENDBR64]            ← 4 bytes (CET landing pad)
[PUSH rbx..r15]      ← 6 × 8 = 48 bytes (callee-saved)
[initial_regs array] ← sub rsp, 128 (16 × 8)
[VmStubArgs (64B)]   ← sub rsp, 64
[CALL vm_stub_entry] ← indirect call through call_slot
[ADD rsp, 64+128]    ← deallocate
[POP r15..rbx]       ← restore
[JMP resume]         ← branch to region end

Required .eh_frame FDE description:

• Initial CFA = rsp + 8 (return address)
• After each PUSH: CFA offset increments by 8, register saved at [CFA - N]
• After sub rsp, 128: CFA offset += 128
• After sub rsp, 64: CFA offset += 64
• At CALL: all callee-saved locations described

Implementation Plan

1. StubEmitter: emit_entry_stub() returns a new Stub::eh_frame_fde field (raw FDE bytes) alongside the machine code. The FDE offsets are relative to the stub start.

2. PayloadBuilder: Collect all per-stub FDEs, prepend a shared CIE (Common Information Entry), build a complete .eh_frame section.

3. ELFEditor: When add_segment() is called, also create an .eh_frame section (SHT_PROGBITS, SHF_ALLOC) containing the CIE+FDE data. Register it with a PT_GNU_EH_FRAME segment if needed.

4. MachOEditor: Create compact unwind entries or __eh_frame section within the __VMPILOT segment.

5. PEEditor: Create .pdata (function table) + .xdata (unwind data) for x64 SEH.

Why Not Now

• v1 error policy: vm_stub_entry calls abort() on any error (NDEBUG) or returns INT64_MIN (debug). No C++ exception crosses the stub boundary.
• DWARF CFI binary encoding is complex (ULEB128, register mappings per ABI, CIE augmentation strings). Getting it wrong is worse than not having it.
• Debugger backtrace quality is a developer convenience, not a correctness issue.

References

• D15 §5.2: Uniform 12-step pipeline (stubs are pre-pipeline entry points)
• D10 GAP5 (C5): "Exception/stack unwinding — stubs break DWARF unwinding"
• RM §5: "Register CFG/CFI/BTI valid targets" (partially done: landing pads present, unwind info missing)