Support use of LLVM's libunwind instead of GCC's unwinder

[emaste]

Summary

It is enabled by setting WITH_LLVM_LIBUNWIND in src.conf(5). It is currently built in libgcc_s.so and libgcc_eh.a to simplify the transition.

[emaste]

share/mk/src.opts.mk

234 ↗

(On Diff #11927)

@br this is a todo we'd need for RISC-V

[emaste]

Add compiler-rt's __gcc_personality_v0 for WITH_LLVM_LIBUNWIND

[emaste]

Add int_util.c to provide compilerrt_abort_impl

[andrew]

I ran nm --dynamic over a copy of libgcc_s.so built with and without this patch, here is the diff (after stripping addresses). The orig file is the current code, the new file is with this patch.

--- nm_orig_noaddr.txt  2016-01-08 16:53:09.888377000 +0000
+++ nm_new_noaddr.txt   2016-01-08 16:53:17.568408000 +0000
@@ -24,32 +24,30 @@
  T _Unwind_Resume_or_Rethrow
  T _Unwind_SetGR
  T _Unwind_SetIP
+ U __assert
  T __bswapdi2
  T __bswapsi2
  T __clear_cache
  w __cxa_finalize
  T __deregister_frame
- T __deregister_frame_info
- T __deregister_frame_info_bases
  T __divsc3
  T __enable_execute_stack
  T __gcc_personality_v0
  T __mulsc3
  T __powisf2
  T __register_frame
- T __register_frame_info
- T __register_frame_info_bases
- T __register_frame_info_table
- T __register_frame_info_table_bases
- T __register_frame_table
+ U __stderrp
  U abort
  U dl_iterate_phdr
+ U dladdr
+ U fprintf
  U free
+ U fwrite
+ U getenv
  U malloc
  U memcpy
  U memset
- w pthread_mutex_lock
- w pthread_mutex_unlock
- w pthread_once
- U puts
- U strlen
+ U pthread_rwlock_rdlock
+ U pthread_rwlock_unlock
+ U pthread_rwlock_wrlock
+ U snprintf

[emaste]

frame registration functions

T __deregister_frame_info
T __deregister_frame_info_bases
T __register_frame_info
T __register_frame_info_bases
T __register_frame_info_table
T __register_frame_info_table_bases
T __register_frame_table

I think these are all fine as previously built binaries carry a weak stub.

string and misc libc functions

- U puts
- U strlen
+ U dladdr
+ U fprintf
+ U fwrite
+ U getenv
+ U snprintf

An artifact of different implementations, should be fine.

[emaste]

+ U __assert
+ U __stderrp

I think this is fine.

- w pthread_mutex_lock
- w pthread_mutex_unlock
- w pthread_once
+ U pthread_rwlock_rdlock
+ U pthread_rwlock_unlock
+ U pthread_rwlock_wrlock

Switch from mutex to rwlock is fine. I suspect the weak ones date from before libc carried weak versions of pthread functions and this is acceptable now.

[emaste]

@kan can you comment on the symbol changes introduced here (specifically wrt. the weak thread functions)?

[emaste]

exp-run requested in PR 206039 https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=206039

[andrew]

I'm happy with this on arm64, it may need more work for other architectures but this shouldn't stop us using it now.

[emaste]

In discussion on IRC we did discover that old GCCs don't put the eh into a separate segment and therefore require __register_frame_info et al. It is still unclear whether this is an issue that needs to be solved.