ARM unwind info for __clang_call_terminate

Hello,

I’m working on a simple unwinder for Android that relies on common C++ exception handling tables (.ARM.exidx / .eh_frame) and recently I found out that for armeabi-v7a it is impossible to unwind stack for abort that happens when someone’s destructor throws (seems like it is the case for any nothrow function). In that case std::terminate() is called via shim function __clang_call_terminate() which is emitted by compiler and has EXIDX_CANTUNWIND associated entry that prevents unwinder to pass through this function:

$ readelf --unwind test.o | grep __clang_call_terminate
Unwind section ‘.ARM.exidx.text.__clang_call_terminate’ at offset 0x98 contains 1 entry:
0x0 <__clang_call_terminate>: 0x1 [cantunwind]

Here is mentioned shim generated with -funwind-tables -O3 (source code is trivial so I omitted it):

__clang_call_terminate:
.fnstart
.save {r7, lr}
push {r7, lr}
bl __cxa_begin_catch
bl _ZSt9terminatev
.Lfunc_end1:
.size __clang_call_terminate, .Lfunc_end1-__clang_call_terminate
.cantunwind
.fnend

At the same time unwinder passes through it on arm64-v8a. I failed to find any explanations of this difference. So asking here seems to be the last resort.

My question is: why for 32-bit ARM clang emits .cantunwind for __clang_call_terminate() instead of normal unwind instructions?

Thanks.

clang generally does not emit unwind tables into a loadable section on ARM (.eh_frame on 64-bit, or .ARM.exidx on 32-bit), unless they are explicitly requested with -funwind-tables, or they are required for C++ unwinding.

Separately, it looks like there’s a compiler bug involving __clang_call_terminate specifically, where the compiler doesn’t honor -funwind-tables. Filed https://bugs.llvm.org/show_bug.cgi?id=41351 .

Separately, if you turn on debug info, and you’re also using DWARF unwinding (which is the default on AArch64, but not 32-bit ARM), and any function in a file requires an unwind table for C++ unwinding, all functions in that file get DWARF unwind tables in the .eh_frame section. This isn’t really intentional; it’s a limitation of the .cfi_sections assembler directive we use for unwind tables. Maybe worth looking into at some point to try to shrink the size of the .eh_frame section.

-Eli