Fix ARM32 NativeAOT SIGSEGV: restore VRS helpers and __unw_*_reg outside _LIBUNWIND_NATIVEAOT guard

[Copilot]

PR #128667's _LIBUNWIND_NATIVEAOT guards accidentally excluded _Unwind_VRS_Get/Set/Pop (and their __unw_get/set_reg/__unw_save_vfp_as_X dependencies) from the NativeAOT build while leaving their caller _Unwind_VRS_Interpret enabled. The linker resolved the missing exported VRS symbols from libgcc_s.so.1, whose implementations treat NativeAOT's ArmUnwindCursor* shim as a real libgcc context → memory corruption → SIGSEGV on ARM32. x64/arm64 are unaffected (they use DWARF templates directly on Registers_REGDISPLAY, never calling _Unwind_VRS_*).

Changes

• Unwind-EHABI.cpp: Split single _LIBUNWIND_NATIVEAOT guard into two blocks. ValueAsBitPattern, _Unwind_VRS_Set, _Unwind_VRS_Get_Internal, _Unwind_VRS_Get, _Unwind_VRS_Pop now sit between the two guards (always compiled). Exception-dispatch functions that depend on __unw_step (__aeabi_unwind_cpp_pr*, unwind_phase1/2, _Unwind_RaiseException, etc.) remain guarded.

• libunwind.cpp: Split the single large _LIBUNWIND_NATIVEAOT guard into three inner guards, making __unw_get_reg, __unw_set_reg, __unw_get_fpreg, __unw_set_fpreg, and __unw_save_vfp_as_X unconditional. The singleton, __unw_init_local, __unw_step, __unw_resume, DWARF/FDE functions, and Apple dynamic sections remain guarded.

• llvm-libunwind-version.txt: Recorded fix commit per repo convention for vendored source patches.

Original prompt

Summary

PR #128667 ("Guard unused llvm-libunwind symbols to avoid duplicates on Android") introduced a regression that causes SIGSEGV crashes during stack unwinding on ARM32 (32-bit ARM / EHABI) NativeAOT targets only. x64 and arm64 are unaffected. This reproduces on Linux arm32 (glibc), not just Android.

Root cause

On ARM32, NativeAOT unwinds managed frames through the EHABI virtual-register-set interpreter _Unwind_VRS_Interpret in src/native/external/llvm-libunwind/src/Unwind-EHABI.cpp. That function was deliberately kept enabled by PR #128667 (it sits above the new #if !defined(_LIBUNWIND_NATIVEAOT) guard, which starts at line 450).

However, _Unwind_VRS_Interpret calls the helper functions _Unwind_VRS_Get, _Unwind_VRS_Set, and _Unwind_VRS_Pop (and the ValueAsBitPattern helper they reference). Those three helpers are defined below the new guard (around lines 915, 1045, 1060), inside the #if !defined(_LIBUNWIND_NATIVEAOT) block that spans lines 450–1219. So when _LIBUNWIND_NATIVEAOT is defined, the PR accidentally removed _Unwind_VRS_Get/Set/Pop from the NativeAOT build while leaving their caller _Unwind_VRS_Interpret in place.

As a result, the linker resolves the now-missing _Unwind_VRS_Get/Set/Pop calls against the platform libgcc_s.so.1 (_Unwind_VRS_*@GCC_3.5). This was confirmed via nm -D on a final ARM32 executable, which shows:

U _Unwind_VRS_Get@GCC_3.5
U _Unwind_VRS_Pop@GCC_3.5
U _Unwind_VRS_Set@GCC_3.5
U __aeabi_unwind_cpp_pr0@GCC_3.5
U __aeabi_unwind_cpp_pr1@GCC_3.5
U __aeabi_unwind_cpp_pr2@GCC_3.5

NativeAOT's _Unwind_VRS_Interpret passes its own ArmUnwindCursor shim cast to _Unwind_Context* (see src/coreclr/nativeaot/Runtime/unix/UnwindHelpers.cpp, UnwindHelpers::StepFrame, the _LIBUNWIND_ARM_EHABI branch). The in-tree _Unwind_VRS_Get/Set/Pop understand this shim (cast to unw_cursor_t → AbstractUnwindCursor* → virtual dispatch into Registers_REGDISPLAY). The libgcc versions expect a real libgcc _Unwind_Context and dereference it at different offsets → garbage reads/writes → SIGSEGV during unwinding.

x64/arm64 never call _Unwind_VRS_* (they use the DWARF/compact unwind templates directly on Registers_REGDISPLAY), which is exactly why only ARM32 regressed.

Required fix

Narrow the _LIBUNWIND_NATIVEAOT guard in src/native/external/llvm-libunwind/src/Unwind-EHABI.cpp so the helpers that _Unwind_VRS_Interpret depends on remain compiled into the NativeAOT build:

• _Unwind_VRS_Get (and its internal helper _Unwind_VRS_Get_Internal)
• _Unwind_VRS_Set
• _Unwind_VRS_Pop
• the ValueAsBitPattern helper they reference

These should be outside (above/before) the #if !defined(_LIBUNWIND_NATIVEAOT) guard, alongside _Unwind_VRS_Interpret which already is.

The genuinely-unused C++ exception-dispatch entry points must REMAIN behind the guard (they depend on __unw_step and other public API from libunwind.cpp that is excluded for NativeAOT):

• __aeabi_unwind_cpp_pr0/pr1/pr2
• unwind_phase1, unwind_phase2, unwind_phase2_forced
• _Unwind_RaiseException, _Unwind_Resume, _Unwind_Complete
• _Unwind_GetLanguageSpecificData, _Unwind_GetRegionStart, _Unwind_ForcedUnwind, _Unwind_DeleteException
• __gnu_unwind_frame

In practice this means: move the #if !defined(_LIBUNWIND_NATIVEAOT) block boundaries so that _Unwind_VRS_Set, _Unwind_VRS_Get_Internal, _Unwind_VRS_Get, _Unwind_VRS_Pop, and ValueAsBitPattern are NOT guarded out, while everything currently guarded that depends on __unw_step stays guarded. Be careful with ValueAsBitPattern (it is [[gnu::unused]] and only referenced by _LIBUNWIND_TRACE_API in the VRS helpers) — keep it available wherever the VRS helpers are so there is no unused/missing-symbol issue.

This is a patch to a vendored external source. Per repo convention, also record the patch commit in src/native/external/llvm-libunwind-version.txt (append an Apply https://github.com/dotnet/runtime/commit/<sha> line as the previous patches do), consistent with how PR #128667 recorded its change.

Validation

After the change, the same nm -D on a built ARM32 executable should no longer list _Unwind_VRS_Get/Set/Pop as undefined @GCC_3.5 imports — they should be resolved locally. Ensure the build still compiles for all NativeAOT unix targets (x64, arm64, arm32) and that the non-NativeAOT (regular llvm-libunwind) build is unaffected by the guard changes.

References

• Regressing PR: Guard unused llvm-libunwind symbols to avoid duplicates on Android #128667
• File: src/native/external/llvm-libunwind/src/Unwind-EHABI.cpp
• File: src/coreclr/nativeaot/Runtime/unix/UnwindHelpers.cpp (the _LIBUNWIND_ARM_EHABI branch of UnwindHelpers::StepFrame)
• File: src/native/external/llvm-libunwind-version.txt

The following is the prior conversation context from the user's chat exploration (may be truncated):

User: after this PR merged, I'm seeing SIGSEGV crash...

This pull request was created from Copilot chat.

[dotnet-policy-service]

Tagging subscribers to this area: @agocke, @dotnet/ilc-contrib
See info in area-owners.md if you want to be subscribed.

[MichalStrehovsky]

/azp run runtime-nativeaot-outerloop

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[sbomer]

I believe this will cause duplicate symbols when linking an arm32 android app, on the unconditional symbols that are no longer guarded. It doesn't cause problems on arm64/x64 because those symbols aren't used, and the libunwind.cpp.o from the runtime archive doesn't get extracted.

[am11]

Perhaps we can omit these just for android to make both linux-arm and android-arm happy:

Suggested change

	#if !defined(__USING_SJLJ_EXCEPTIONS__) && !defined(__wasm__)
	#if defined(_LIBUNWIND_NATIVEAOT) && !defined(__ANDROID__)
	#if !defined(__USING_SJLJ_EXCEPTIONS__) && !defined(__wasm__)

[sbomer]

Could work, but that seems a little fragile to me (I now think the ifdef approach was fragile to begin with), so trying another idea in #128927.