[PATCH 3/3] barrier: optimise by not having the hasher in a loop
Thiago Macieira
thiago.macieira@intel.com
Fri Mar 5 18:21:28 GMT 2021
Our thread's ID does not change so we don't have to get it every time
and hash it every time.
libstdc++-v3/ChangeLog:
* include/std/barrier(arrive): move hasher one level up in the
stack.
---
libstdc++-v3/include/std/barrier | 10 +++++-----
1 file changed, 5 insertions(+), 5 deletions(-)
diff --git a/libstdc++-v3/include/std/barrier b/libstdc++-v3/include/std/barrier
index e09212dfcb9..4bb5642c164 100644
--- a/libstdc++-v3/include/std/barrier
+++ b/libstdc++-v3/include/std/barrier
@@ -94,7 +94,7 @@ It looks different from literature pseudocode for two main reasons:
alignas(__phase_alignment) __barrier_phase_t _M_phase;
bool
- _M_arrive(__barrier_phase_t __old_phase)
+ _M_arrive(__barrier_phase_t __old_phase, size_t __current)
{
const auto __old_phase_val = static_cast<unsigned char>(__old_phase);
const auto __half_step =
@@ -103,9 +103,7 @@ It looks different from literature pseudocode for two main reasons:
static_cast<__barrier_phase_t>(__old_phase_val + 2);
size_t __current_expected = _M_expected;
- std::hash<std::thread::id> __hasher;
- size_t __current = __hasher(std::this_thread::get_id())
- % ((_M_expected + 1) >> 1);
+ __current %= ((__current_expected + 1) >> 1);
for (int __round = 0; ; ++__round)
{
@@ -163,12 +161,14 @@ It looks different from literature pseudocode for two main reasons:
[[nodiscard]] arrival_token
arrive(ptrdiff_t __update)
{
+ std::hash<std::thread::id> __hasher;
+ size_t __current = __hasher(std::this_thread::get_id());
__atomic_phase_ref_t __phase(_M_phase);
const auto __old_phase = __phase.load(memory_order_relaxed);
const auto __cur = static_cast<unsigned char>(__old_phase);
for(; __update; --__update)
{
- if(_M_arrive(__old_phase))
+ if(_M_arrive(__old_phase, __current))
{
_M_completion();
_M_expected += _M_expected_adjustment.load(memory_order_relaxed);
--
2.30.1
More information about the Gcc-patches
mailing list