https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104831
--- Comment #6 from Patrick O'Neill <patrick at rivosinc dot com> --- I just reviewed the manual - I was incorrect. Appendix A is correct. I forgot that the RISCV implementation used leading fences. The second litmus test is no longer valid since a simple load/store with no fence would not occur in outputted RISCV asm. The load/store would have a leading fence that would prevent it from entering the LR/SC pairing.