Emit trivial leaf CFI for UnwindOp::None on arm64

[Gankra]

fixes #618

This needs some testing on real inputs

[mstange]

With this patch, dump_syms produces the following output for libsystem_kernel.dylib from macOS 12.3.1: https://gist.github.com/mstange/8fdff669c0785ec7a4c005311e639801

[Gankra]

Before:

STACK CFI INIT 50e4 70 .cfa: x29 16 + x29: .cfa -16 + ^ .ra: .cfa -8 + ^ x19: .cfa -24 + ^ x20: .cfa -32 + ^ x21: .cfa -40 + ^ x22: .cfa -48 + ^
STACK CFI INIT 52e0 30 .cfa: x29 16 + x29: .cfa -16 + ^ .ra: .cfa -8 + ^
STACK CFI INIT 5368 4 .cfa: sp 0 + .ra: x30 0 +

After:

STACK CFI INIT 50e4 70 .cfa: x29 16 + x29: .cfa -16 + ^ .ra: .cfa -8 + ^ x19: .cfa -24 + ^ x20: .cfa -32 + ^ x21: .cfa -40 + ^ x22: .cfa -48 + ^
STACK CFI INIT 5154 18c .cfa: sp .ra: x30
STACK CFI INIT 52e0 30 .cfa: x29 16 + x29: .cfa -16 + ^ .ra: .cfa -8 + ^
STACK CFI INIT 5310 58 .cfa: sp .ra: x30
STACK CFI INIT 5368 4 .cfa: sp 0 + .ra: x30 0 +

Resulting socc-pair diff:

       frames: [
         0: {
           file: null
           frame: 0
           function: __psynch_cvwait
           function_offset: 0x00000008
           line: null
           missing_symbols: false
           module: libsystem_kernel.dylib
           module_offset: 0x00005290
           offset: 0x1a2704290
           trust: context
           unloaded_modules: null
         }
         1: {
           file: null
           frame: 1
           function: _pthread_cond_wait
           function_offset: 0x000004d0
           line: null
           missing_symbols: false
           module: libsystem_pthread.dylib
           module_offset: 0x00007838
           offset: 0x1a273e838
~          local had better trust (frame_pointer vs cfi)
           unloaded_modules: null
         }
-        local had extra array value:
         2: {
           file: hg:hg.mozilla.org/mozilla-central:mozglue/misc/ConditionVariable_posix.cpp:8e5247451c9a83abe6f877f5f2c62332a84c9aaf
           frame: 2
           function: mozilla::detail::ConditionVariableImpl::wait(mozilla::detail::MutexImpl&)
           function_offset: 0x0000000000000008
           line: 106
           missing_symbols: false
           module: libmozglue.dylib
           module_offset: 0x0000000000054f08
           offset: 0x0000000104e94f08
           trust: cfi
           unloaded_modules: null
         }
         2: {
           file: hg:hg.mozilla.org/mozilla-central:xpcom/threads/nsThread.cpp:8e5247451c9a83abe6f877f5f2c62332a84c9aaf
~          ignoring field frame: 2
           function: nsThread::ProcessNextEvent(bool, bool*)
           function_offset: 0x00000420
           line: 1141
           missing_symbols: false
           module: XUL
           module_offset: 0x003fa62c
           offset: 0x11a6a262c
           trust: cfi
           unloaded_modules: null
         }

[Gankra]

• Before this change or my arm fp/lr rewrites: we were failing to use CFI on this frame, using frame pointers badly, and skipping the frame 2 of the stackwalk.

• WITH my arm rewrites: we still fail to use CFI and end up skipping frame 1 of the stackwalk because we're conservatively assuming this isn't a leaf function

• WITH this change: we now fully use CFI and get both frame 1 and 2, no frame pointers needed!

[loewenheim]

LGTM, but I would like to wait for a second opinion.

[Swatinem]

sounds reasonable. Too bad all these functions don’t have "real" CFI, and you need to apply such heuristics.