This is split out from the atomic deprecation series. It converts lib/ring
off rte_atomic32 and onto the C11 memory model, except where the C11 version
has a noticeable performance drop on x86 with GCC.
The pre-existing C11 and GCC-builtin paths lived in separate headers with
substantial duplication. After this series, only the MP head CAS
(__rte_ring_headtail_move_head_mt) retains separate implementations;
everything else is shared. Patch 2 documents the reason for keeping the GCC
builtin on the MP head CAS.
The default RTE_USE_C11_MEM_MODEL selection per architecture is unchanged.
v2 - consolidate cleanup patches
- fix the memory order on first load in _st case.
it was going back/forth across the patches
Stephen Hemminger (3):
ring: split single thread vs multi-thread cases
ring: use GCC builtin as alternative to rte_atomic32
ring: cleanup the C11 code
lib/ring/meson.build | 2 +-
lib/ring/rte_ring_c11_pvt.h | 62 +++-------
lib/ring/rte_ring_elem_pvt.h | 116 ++++++++++++++++--
..._ring_generic_pvt.h => rte_ring_gcc_pvt.h} | 58 +++------
lib/ring/rte_ring_hts_elem_pvt.h | 8 +-
lib/ring/soring.c | 34 ++---
6 files changed, 159 insertions(+), 121 deletions(-)
rename lib/ring/{rte_ring_generic_pvt.h => rte_ring_gcc_pvt.h} (65%)
--
2.53.0
