invert_mask descriptor inverts return value to true for unassigned elements

[aleksamilisavljevic]

According to the documentation:

ALP/include/graphblas/utils.hpp

Lines 299 to 307 in 73a3cb3

	* @return If the descriptor does not include grb::descriptors::structural
	* nor grb::descriptors::structural_complement and if \a assigned
	* is false, then the entry is ignored and uninterpreted, thus
	* returning \a false.
	* @return If the descriptor includes grb::descriptors::invert_mask,
	* returns the negation of the dereferenced value of \a val
	* which is first cast to a \a bool.
	* @return Otherwise, returns the dereferenced value of \a val,
	* cast to a \a bool.

However, if the descriptor invert_mask is set and assigned is set to false, then the return value will be true:

ALP/include/graphblas/utils.hpp

Lines 320 to 336 in 73a3cb3

	bool ret = false;
	// if we request a structural mask, decide only on passed assigned variable
	if( descriptor & descriptors::structural ) {
	ret = assigned;
	} else {
	// if based on value, if there is a value, cast it to bool
	if( assigned ) {
	ret = static_cast< bool >( val[ offset ] );
	}
	// otherwise there is no value and false is assumed
	}
	// check whether we should return the inverted value
	if( descriptor & descriptors::invert_mask ) {
	return !ret;
	} else {
	return ret;
	}

[anyzelman]

I'm a bit lost in this one so lemme go by specification first:

• structural inversion has that for example set< structural_mask | invert_mask >( output_vector, empty_vector, 1 ) that the output is dense, however for matrices this combination is explicitly disallowed (since it would/could generate dense matrices);
• non-structural inversion has that for example set< invert_mask >( output_vector, empty_vector, 1 ) remains empty-- because we only invert mask values (and there are none).

Now on the quoted lines, actually the logic keeps ret equal to false, which seems fine. However, later on in that utility we have that

                        // check whether we should return the inverted value
                        if( descriptor & descriptors::invert_mask ) {
                                return !ret;
                        } else {
                                return ret;
                        }

and that indeed seems wrong. I'll relabel the issue accordingly and slot it for v0.8

[GiovaGa]

This seems to me to be already fixed in the linked pull request
@anyzelman have a look

[anyzelman]

I'm removing it from v0.8 for reasons described here (the invalid tag seems more accurate in hindsight):

#317 (comment)

As that comment mentions, there may be still merit for this MR but 1) the suggested behaviour should be introduced by a new descriptor, and 2) there should be a good motivating use case for it. I suggest discussions continue in the MR.