Nullable object

[eddelbuettel]

In quite some user-space code the R side has something like

weightedLeastSquares(X, y, w)      # ie y ~ X given weights w

where the third argument is options, ie

weightedLeastSquares(X, y, NULL)

In that case we generally cannot have the third strongly typed as NumericMatrix or alike and have to keep it a SEXP in the interface. Nothing wrong with that but a bit of an eyesore.

Following some discussion with @dcdillon I have written a really simple wrapper class Nullable that contains a SEXP as well as a bool for state. With that we can do (stylized)

...., NumericVector & w) {
//...
if (!w.isNull()) {
    // do stuff with w
}

If there is no protest -- or someone reminds of something much better we already have ;-) -- then I'll fold that in.

[jjallaire]

In boost::error_code they implement nullability via some operator overloads:

typedef void (*unspecified_bool_type)();
static void unspecified_bool_true() {}

operator unspecified_bool_type() const BOOST_SYSTEM_NOEXCEPT //
true if error
{
return m_val == 0 ? 0 : unspecified_bool_true;
}

bool operator!() const BOOST_SYSTEM_NOEXCEPT // true if no error
{
return m_val == 0;
}

This allows you to write code like this:

if (ec) {

}

if (!ec) {

}

Maybe that's too fancy/obscure though?

On Thu, Sep 3, 2015 at 6:56 AM, Dirk Eddelbuettel
notifications@github.com wrote:

In quite some user-space code the R side has something like

weightedLeastSquares(X, y, w) # ie y ~ X given weights w

where the third argument is options, ie

weightedLeastSquares(X, y, NULL)

In that case we generally cannot have the third strongly typed as
NumericMatrix or alike and have to keep it a SEXP in the interface. Nothing
wrong with that but a bit of an eyesore.

Following some discussion with @dcdilllon I have written a really simple
wrapper class Nullable that contains a SEXP as well as a bool for state.
With that we can do (stylized)

...., NumericVector & w) {
//...
if (!w.isNull()) {
// do stuff with w
}

If there is no protest -- or someone reminds of something much better we
already have ;-) -- then I'll fold that in.

—
Reply to this email directly or view it on GitHub.

[eddelbuettel]

Interesting. Looks fancy. That may fit. So far I only had operator SEXP() and get() to extract, and isNull() to test. I may play with the above. Thanks!

[dcdillon]

It's not as fancy as it looks. It's just a cast operator to bool.

operator bool()
{
    return !isNull();
}

bool operator!()
{
    return isNull();
}

Just syntactical sugar.

It would, however, make ambiguous in your use case because your object (which inherits from Nullable) would have cast operators to bool and SEXP.

if (nullableObject)
{
    // do something
}

[eddelbuettel]

I think we won't inherit from Nullable as it can't be a CRTP idiom -- we can't instantiate our standard types (ie NumericalMatrix) with NULL as that gets us the very run-tome woes which Nullable tries to avoid.

[romainfrancois]

I'd say make it a template, e.g. Nullable

[eddelbuettel]

I looked into that, but eg in Nullable<NumericMatrix> foo I get the same initial problem of NumericalMatrix not taking NULL --- which lead us to Nullable in the first place. If you can see a way to overcame that (easily) I'd be all ears. I'd prefer it templated too...

[kevinushey]

FWIW the use of 'unspecified_bool_type' is there to help prevent undesired type coercions (ie the safe bool idiom). You almost definitely don't want just a raw operator bool(){} as that is too prone to misuse.

[eddelbuettel]

Good timing. Just tried operator bool(){} approach and was just complaining to @dcdillon over IM :)

[eddelbuettel]

I'm going to be pedestrian here and just rely on isNull() as a member. I don't have luck with either approach while I also have operator SEXP().