Perf[MQB]: optimize inDispatcherThread

[678098]

Commit 1:

1. Replace bslmt::ThreadUtil::Handle usage with bslmt::ThreadUtil::Id. According to docs, Handle doesn't allow usage from a thread different from the one it corresponds to. Id allows this.
2. Keep a copy of the assigned dispatcher thread id in mqbi::DispatcherClientData. This allows to simplify lookups for inDispatcherThread checks.
3. Move mqbi::Dispatcher::inDispatcherThread(int) -> mqbi::DispatcherClient::inDispatcherThread(void). The previous declaration was abstract and it was not possible to inline inDispatcherThread, now it is possible to inline it.
4. Simplify inDispatcherThread calls. Before: BSLS_ASSERT_SAFE(dispatcher()->inDispatcherThread(this));, after: BSLS_ASSERT_SAFE(inDispatcherThread());

Commit 2:

1. Move d_threadId: mqbi::DispatcherClientData::d_threadId -> mqbi::DispatcherClient::d_threadId. This allows to simplify calls in inDispatcherThread() and make it even faster.

Benchmarks

Darwin M2 - RelWithDebInfo

Before: 4.1 ns per iteration BSLS_ASSERT_OPT(dispatcher_p->inDispatcherThread(client_p));
After (commit 1): 2.7 ns per iteration BSLS_ASSERT_OPT(client_p->inDispatcherThread());
After (commit 2): 2.0 ns per iteration BSLS_ASSERT_OPT(client_p->inDispatcherThread());

Linux amd64 - RelWithDebInfo

Before: 4.6 ns per iteration BSLS_ASSERT_OPT(dispatcher_p->inDispatcherThread(client_p));
After (commit 1): 3.4 ns per iteration BSLS_ASSERT_OPT(client_p->inDispatcherThread());
After (commit 2): 0.6 ns per iteration BSLS_ASSERT_OPT(client_p->inDispatcherThread());

Note that the speedup on M2 Darwin is 2x, on amd64 Linux is 7.5x.

[bmq-oss-ci]

Build 3174 of commit 4bf33e5 has completed with FAILURE

[678098]

Just to be sure I verified in my benchmark that inDispatcherThread() takes the same time with or without the second part of the expression when the first part is true (short-circuit).

[678098]

Also dispatcherClientData() is a virtual function and makes inDispatcherThread() slower than it can be. If we just keep threadId as a field in mqbi::DispatcherClient, we can address it directly without virtual function call and save time on this check.

[678098]

If we just keep threadId as a field in mqbi::DispatcherClient, we can address it directly without virtual function call and save time on this check.

I applied this in the second commit and updated the benchmarks

Note that amd64 Linux time improved much more:

TEST /blazingmq/src/groups/mqb/mqba/mqba_dispatcher.t.cpp CASE -1
inDispatcherThread():
       total: 6.11 ms (10000000 iterations)
    per call: 0 ns

[bmq-oss-ci]

Build 3178 of commit 675493f has completed with FAILURE

[pniedzielski]

Thanks for the update yesterday, looks good.