Fixes: https://github.com/dankamongmen/notcurses/issues/1009
Whenever a widget is created with its `*_create` function it currently
claims full ownership of the passed panel, including its destruction.
However, the C++ wrapper around the panel is not aware of this and will
attempt to destroy the native panel in the destructor, leading to
segfaults.
Fix this by introduction of a `Widget` class which contains the logic to
properly modify the `Panel` instance to not double-destroy the native
panel. The solution is a bit fragile since the `Panel` instance is left
intact (we can't free it for the user) in a state that's safe for the
C++ wrapper, but calling any C function via the wrapper **will** pass a
`NULL` pointer in the panel argument - therefore the C functions MUST be
proofed against this. The proofing belongs in the C backend code since
this protects also C and other language binding users from such abuse.
The Widget class will first verify that the passed `Plane` instance
hasn't already been "disowned" and will throw an exception to the effect
if it was. Next, it will proceed to take over ownership of the native
panel instance and mark the passed `Panel` as "invalid" (i.e. not owning
any native panel instance anymore)
The above changes require modification of `Panel` instances and so all
the widget constructors taking `const*` or `const&` have been removed
from widget classes.
This is to make it possible, in the future, to create multiple instances
of `NotCurses` for multiple terminals. The first instance of
`NotCurses` becomes the default one, so that any instances of other
classes that aren't explicitly created with a pointer to another
`NotCurses` instance still work as expected.
Note that currently trying to call `notcurses_init` twice results in the
following error for me:
0x55555559bfc0 is already registered for signals
Couldn't drop signals: 0x55555559bfc0 != 0x5555555b6720
terminate called after throwing an instance of 'ncpp::init_error*'
Program received signal SIGABRT, Aborted.
The error is signalled by `setup_signals` and the pointer shown in the
message points to the first `struct notcurses` instance created.
Fixes: https://github.com/dankamongmen/notcurses/issues/616
SIGSEGV was caused by an invalid cast.
Short explanation: PEBKAC
Long explanation: `Selector.hh`, `Plot.hh` and `MultiSelector.hh` did
not include `Plane.hh`, they merely declared `class Plane;` because
inclusion of `Plane.hh` would cause circular dependencies to appear and
the compiler would be unhappy. On top of that, yours truly wrenched the
compiler's hands and caused it to believe that a pointer to `Plane` is
really a pointer to `ncplane*` which was quite a silly thing to do as
the compiler, not having included `Plane.hh` and thus not knowing full
definition of the type, wasn't able to look up the type cast operator in
`Plane`.
Don't abuse `reinterpret_cast`, kids!
Nick prefers error handling based on exceptions in all cases, while I
prefer to save exception handling for truly exceptional situations -
function parameter validation and class constructor. However, there's no
need to not support both approaches, to be chosen at the discretion of
the developer.
NCPP follows RAII and all classes throw exceptions from their
constructors in case they cannot initialize properly. Likewise,
functions taking pointers that are required validate them and throw
exceptions whenever the requirement isn't met.
This commit goes one step further in that it enables optional validation
of notcurses function return values and throwing an
exception (`ncpp::call_error`) should the function signal an error. This
is disabled by default but it can be enabled by defining the
`NCPP_EXCEPTIONS_PLEASE` macro (preferably on the command line or
before *each* inclusion of any NCPP headers).
Out of necessity, this breaks the ABI (plus I found a handful of minor
issues in the code), but I think it's worth having this support in
place.
