Following the introduction of 'Pianoroll::region_prop_change()' (commit #29833d75c92) the operator<< (for PBD::PropertyChange) now needs to be visible outside of libpbd
We now use a stack allocator when making a copy of current connection state at
the start of the signal emission process, and when collecting results from
signal handlers in the case of a non-void return type.
These changes also include a functionally neutral reworking of how the connection
state copy is made and then used to check that a connection/handler is still
valid mid-emission.
Heap allocation will still happen if a signal has more than (currently) 512
connections. A little experimentation reveals that the maximum number of
connections is typically nroutes+1, so 512 seems like a reasonable choice
for this.
Regular .h files *should* be self-contained and independent of previous
includes and guarded to only include once. Make it clear which files
that *doesn't* apply for at all.
the rest from `tools/convert_boost.sh`.
* replace boost::function, boost::bind with std::function and std::bind.
This required some manual fixes, notably std::placeholders,
some static_casts<>, and boost::function::clear -> = {}.
AbstractUI IS-A BaseUI IS-A EventLoop
IS-A sigc::trackable
If we have sent a call_slot() request to an EventLoop that has not executed
when the object involved in the call_slot() functor is destroyed, we need to
ensure that the request is invalidated.
To do this, We register "notify" callbacks with the sigc::trackable that is a
base class of the object involved in the functor given to
call_slot(). sigc::trackable will call these "notify" callbacks from its
destructor.
So when the call_slot() functor's relevant object dies, and its sigc::trackable
base class is destroyed, it will invoke all of its the "notify" callbacks, which will
in turn call EventLoop::invalidate_request() and this hopefully marks all the
queued call_slot() functors as "do not call".
However, invalidate_request() requires a lock, and access to the lock is
granted via a pure virtual, EventLoop::slot_invalidation_lock(). In the
heirarchy cited above, this is implemented by AbstractUI.
When we destroy an AbstractUI, ~AbstractUI() is called first, and this destroys
the lock and changes the VTT so that ::slot_invalidation_lock() becomes a pure
virtual again.
Eventually we will call ~trackable() which in turns runs all the "notify"
callbacks, and then removes them. But when these callbacks end up in
EventLoop::invalidate_request(), we try to call ::slot_invalidation_lock() and
C++ will abort because of its (now) pure virtual status.
Therefore, we must invoke the "notify" callbacks before the
::slot_invalidation_lock() becomes pure, and that means inside ~AbstractUI, as
an explicit call to trackable::notify_callbacks().
This has not appeared before (remarkably), but became an issue when the
Launchpad Pro support code's main object (derived from MIDISurface and hence
from AbstractUI) "failed" to use sub-objects for its various methods. So when
it connects to, for example, the Session::RouteAddedOrRemoved signal, it is
connecting itself (derived from a sigc::trackable). When the Launchpad Pro
object is destroyed, it tries to invalidate all the call_slot() requests, but
this requires access to an event loop lock - owned by the Launchpad Pro event
loop, which is already destroyed!
Other surfaces have generally avoided this by using other objects to provide
methods of dealing with signals from libardour objects.
we allow use of/dependency on sigc::trackable there, so this is
both legal but also sensible.
Leave the macro definition of invalidator(x) in gtkmm2ext/gui_thread.h
because it doesn't hurt and makes some sense for it to be there. No
reason for a source module that needs invalidator(x) to load EventLoop
decl.
