Can't debug windows aarch64 executables that throw c++ exceptions, if built with lllvm-mingw

[ssbssa]

If you try to debug any executable that throws a c++ exception, you get this result:

* thread #​1, stop reason = Exception 0xc00000ff encountered at address 0x7fff44a181b0
  frame #​0: 0x00007fff44a181b0 ntdll.dll`RtlRaiseStatus + 32

I even tried cdb, and got a similar result.

After lots of searching I found this post detailing a deep analysis plus solution of the problem.

I didn't test it, but I expect that a similar change like the following should fix this:

--- a/libunwind/src/Unwind-seh.cpp
+++ b/libunwind/src/Unwind-seh.cpp
@@ -174,7 +174,7 @@ _GCC_specific_handler(PEXCEPTION_RECORD ms_exc, PVOID frame, PCONTEXT ms_ctx,
     }
     // FIXME: Indicate target frame in foreign case!
     // phase 2: the clean up phase
-    RtlUnwindEx(frame, (PVOID)disp->ControlPc, ms_exc, exc, ms_ctx, disp->HistoryTable);
+    RtlUnwindEx(frame, (PVOID)disp->ControlPc, ms_exc, exc, disp->ContextRecord, disp->HistoryTable);
     _LIBUNWIND_ABORT("RtlUnwindEx() failed");
   case _URC_INSTALL_CONTEXT: {
     // If we were called by __libunwind_seh_personality(), indicate that

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Hannes Domani (ssbssa)

If you try to debug any executable that throws a c++ exception, you get this result: ``` * thread #​1, stop reason = Exception 0xc00000ff encountered at address 0x7fff44a181b0 frame #​0: 0x00007fff44a181b0 ntdll.dll`RtlRaiseStatus + 32 ``` I even tried cdb, and got a similar result.

After lots of searching I found this post detailing a deep analysis plus solution of the problem.

I didn't test it, but I expect that a similar change like the following should fix this:

--- a/libunwind/src/Unwind-seh.cpp
+++ b/libunwind/src/Unwind-seh.cpp
@@ -174,7 +174,7 @@ _GCC_specific_handler(PEXCEPTION_RECORD ms_exc, PVOID frame, PCONTEXT ms_ctx,
     }
     // FIXME: Indicate target frame in foreign case!
     // phase 2: the clean up phase
-    RtlUnwindEx(frame, (PVOID)disp->ControlPc, ms_exc, exc, ms_ctx, disp->HistoryTable);
+    RtlUnwindEx(frame, (PVOID)disp->ControlPc, ms_exc, exc, disp->ContextRecord, disp->HistoryTable);
     _LIBUNWIND_ABORT("RtlUnwindEx() failed");
   case _URC_INSTALL_CONTEXT: {
     // If we were called by __libunwind_seh_personality(), indicate that

[efriedma-quic]

CC @mstorsjo

[mstorsjo]

Thanks for filing this, and for digging up this information! I actually coincidentally ran into this exact same issue yesterday and was annoyed by it, but didn't have time to dig this deep into what's going on. The linked explanation is fantastic.

Your suggested fix seems to work fine. (Also, if trying to run a debugger under Wine, it previously hangs when dealing with exceptions; the reason seems to be the same, and the fix works there as well.)

As the fix does touch generic SEH unwind code, it can potentially affect SEH unwind on other platforms as well (x86_64, armv7 and arm64ec), so I'll try to run all my existing tests that involve unwinding.

I can put up a PR with the fix once I've verified those aspects. I can't take credit for the fix, so I guess it'd be reasonable to add Co-authored-by lines for both @ssbssa and @corsix here. How do you prefer to be credited in that such a context?

[corsix]

If you feel the need for it, Co-authored-by: Peter Cawley <corsix@corsix.org>.

[ssbssa]

Co-authored-by: Hannes Domani <ssbssa@yahoo.de>