Don't register EHFrameRegistrationPlugin on MachO (fixes #1782)

[dg1sbg]

Problem

The long-standing clasp/cando macos-latest/native "Run regression tests" CI failure (#1782): the mp regression suite reaches PROCESS-ABORT-1 and the process std::terminates instead of unwinding (startRunStop.cc:135 unhandled unknown exception → Abort signal). Ubuntu native and the bytecode image are unaffected — only the native image on arm64-darwin.

Root cause

On macOS/arm64, native (cleavir-compiled) code unwinds via compact-unwind (__TEXT,__unwind_info), not DWARF .eh_frame — clasp's native MachO objects carry no __eh_frame. ClaspJIT_O adds orc::EHFrameRegistrationPlugin to its ObjectLinkingLayer. But that plugin's MachO path inserts a front PrePrune pass that removes the __LD,__compact_unwind section:

// llvm/lib/ExecutionEngine/Orc/EHFrameRegistrationPlugin.cpp
if (LG.getTargetTriple().isOSBinFormatMachO())
  PassConfig.PrePrunePasses.insert(PassConfig.PrePrunePasses.begin(),
    [](LinkGraph &G) {
      if (auto *CUSec = G.findSectionByName(MachOCompactUnwindSectionName))
        G.removeSection(*CUSec);
      return Error::success();
    });

That deletes the only unwind info native code has, before JITLink's CompactUnwindManager can synthesize __TEXT,__unwind_info. libunwind is then left with no description for native PCs, so a C++ exception thrown below a native frame (mp::abort_process → throw AbortProcess()) and caught above it (Process_O::runInner) cannot find its handler → std::terminate → abort.

(For the record: the __compact_unwind records are valid self-sufficient UNWIND_ARM64_MODE_FRAME entries; the original "objects lack __eh_frame" framing was a red herring — they correctly lack it, and 0x04000000 is MODE_FRAME on arm64, not MODE_DWARF.)

Fix

Only add EHFrameRegistrationPlugin for non-MachO (ELF) targets. On MachO it has no eh-frame to register anyway, so its sole effect was deleting the compact-unwind. With it gone, the stock pipeline works end-to-end: CompactUnwindManager synthesizes __unwind_info and the default setUpGenericLLVMIRPlatform's UnwindInfoRegistrationPlugin registers it with Apple libunwind. LLVM's own writer handles personality/LSDA, so landing-pad functions are covered too.

Verification

Apple silicon, LLVM 22, boehmprecise native image:

• mp regression suite: 44 successes, 0 failures — PROCESS-ABORT-1/2/3/5 and PROCESS-EXIT pass (were aborting CI).
• mp:abort-process now unwinds across native frames and signals PROCESS-JOIN-ERROR (the documented behavior) instead of std::terminate.
• llvm::orc::UnwindInfoManager::registerSectionsImpl now runs per loaded object (15349×; was never called before); libunwind's findSectionsImpl resolves during the throw.
• No regressions in the unwind / conditions / control suites.

🤖 Generated with Claude Code

[Bike]

I don't have an ARM Mac to test on right this second, and the CI is still reporting an unknown exception there. Though in the interrupt tests. I'll run it again and see what happens.

[dg1sbg]

With all PRs in, I just reran all tests here locally with

clasp --norc --non-interactive --base --feature ignore-extensions --load "sys:src;lisp;regression-tests;run-all.lisp"

and get:

Unexpected Successes:
  RANDOM-SHORT RANDOM-DOUBLE RANDOM-LONG FRAME-LOCALS FRAME-FUNCTION

Expected Failures:
  SBCL-CROSS-COMPILE-4 INCLUDE-LEVEL-2B INCLUDE-LEVEL-3 TYPES-CLASSES-10

Successes: 1963

No unhandled exceptions left ...