diff options
author | Torvald Riegel <triegel@redhat.com> | 2011-12-24 01:42:20 +0000 |
---|---|---|
committer | Torvald Riegel <torvald@gcc.gnu.org> | 2011-12-24 01:42:20 +0000 |
commit | 799142bf743cb04f3a86df85c69481a3c44c1ad9 (patch) | |
tree | 76234a08c818b5da2d7cbe2a47fa07501ef97cb2 /libitm/beginend.cc | |
parent | 8b3bff4ce9864ae2444d0c1e18dc57b32e4d0e3a (diff) | |
download | gcc-799142bf743cb04f3a86df85c69481a3c44c1ad9.zip gcc-799142bf743cb04f3a86df85c69481a3c44c1ad9.tar.gz gcc-799142bf743cb04f3a86df85c69481a3c44c1ad9.tar.bz2 |
libitm: Fixed conversion to C++11 atomics.
libitm/
* beginend.cc (GTM::gtm_thread::begin_transaction): Add comment.
(GTM::gtm_thread::try_commit): Changed memory order.
* config/linux/alpha/futex_bits.h (sys_futex0): Take atomic int
as parameter.
* config/linux/x86/futex_bits.h (sys_futex0): Same.
* config/linux/sh/futex_bits.h (sys_futex0): Same.
* config/linux/futex_bits.h (sys_futex0): Same.
* config/linux/futex.cc (futex_wait, futex_wake): Same.
* config/linux/futex.h (futex_wait, futex_wake): Same.
* config/linux/rwlock.h (gtm_rwlock::writers,
gtm_rwlock::writer_readers, gtm_rwlock::readers): Change to atomic
ints.
* config/linux/rwlock.cc (gtm_rwlock::read_lock,
gtm_rwlock::write_lock_generic, gtm_rwlock::read_unlock,
gtm_rwlock::write_unlock): Fix memory orders and fences.
* config/posix/rwlock.cc (gtm_rwlock::read_lock,
gtm_rwlock::write_lock_generic, gtm_rwlock::read_unlock,
gtm_rwlock::write_unlock): Same.
* config/linux/rwlock.h (gtm_rwlock::summary): Change to atomic int.
* method-gl.cc (gl_mg::init, gl_wt_dispatch::memtransfer_static,
gl_wt_dispatch::memset_static, gl_wt_dispatch::begin_or_restart):
Add comments.
(gl_wt_dispatch::pre_write, gl_wt_dispatch::validate,
gl_wt_dispatch::load, gl_wt_dispatch::store,
gl_wt_dispatch::try_commit, gl_wt_dispatch::rollback): Fix memory
orders and fences. Add comments.
From-SVN: r182674
Diffstat (limited to 'libitm/beginend.cc')
-rw-r--r-- | libitm/beginend.cc | 24 |
1 files changed, 19 insertions, 5 deletions
diff --git a/libitm/beginend.cc b/libitm/beginend.cc index bcc8516..d0ad5a7 100644 --- a/libitm/beginend.cc +++ b/libitm/beginend.cc @@ -259,6 +259,9 @@ GTM::gtm_thread::begin_transaction (uint32_t prop, const gtm_jmpbuf *jb) else { #ifdef HAVE_64BIT_SYNC_BUILTINS + // We don't really care which block of TIDs we get but only that we + // acquire one atomically; therefore, relaxed memory order is + // sufficient. tx->id = global_tid.fetch_add(tid_block_size, memory_order_relaxed); tx->local_tid = tx->id + 1; #else @@ -471,17 +474,28 @@ GTM::gtm_thread::trycommit () // Ensure privatization safety, if necessary. if (priv_time) { + // There must be a seq_cst fence between the following loads of the + // other transactions' shared_state and the dispatch-specific stores + // that signal updates by this transaction (e.g., lock + // acquisitions). This ensures that if we read prior to other + // reader transactions setting their shared_state to 0, then those + // readers will observe our updates. We can reuse the seq_cst fence + // in serial_lock.read_unlock() however, so we don't need another + // one here. // TODO Don't just spin but also block using cond vars / futexes // here. Should probably be integrated with the serial lock code. - // TODO For C++0x atomics, the loads of other threads' shared_state - // should have acquire semantics (together with releases for the - // respective updates). But is this unnecessary overhead because - // weaker barriers are sufficient? for (gtm_thread *it = gtm_thread::list_of_threads; it != 0; it = it->next_thread) { if (it == this) continue; - while (it->shared_state.load(memory_order_relaxed) < priv_time) + // We need to load other threads' shared_state using acquire + // semantics (matching the release semantics of the respective + // updates). This is necessary to ensure that the other + // threads' memory accesses happen before our actions that + // assume privatization safety. + // TODO Are there any platform-specific optimizations (e.g., + // merging barriers)? + while (it->shared_state.load(memory_order_acquire) < priv_time) cpu_relax(); } } |