Feature request: isnan() function

[sylque]

I'm witnessing the following behavior when evaluating an expression:

• nan == nan returns 0.0
• nan != nan returns 1.0

[Blake-Madden]

Yeah, both of these scenarios should return NaN, unless you disagree. Thoughts?

[sylque]

I initially expected nan == nan to return 1.0 and nan != nan to return 0.0.

Then I tried x == nan with x defined as const double x = std::numeric_limits<float>::quiet_NaN() and, to my surprise, it returned 0.0 as well!

So I checked the IEEE standard and found this:

---

Any comparison with NaN (except !=) will return false, including comparing NaN with itself.

Comparison	Result
NaN == NaN	false
NaN != NaN	true
NaN < any	false
NaN > any	false
std::isnan(NaN)	true

---

So the current behavior perfectly conforms to the standard!

But then, how do I compare a number to NaN in tinyexpr++? It seems the library lacks an equivalent to std::isnan().

Oddly enough, there is a NAN entry in tinyexpr++ function list. But the doc is somewhat broken (no argument, no parenthesis) and it doesn't seem to evaluate as it should in an expression.

[Blake-Madden]

NA (or NA()) just returns NaN. So you can do something like this:

IF(5 < 4, NA(), TRUE)

This will evaluate to NaN. It basically behaves like NA() in Excel.

[Blake-Madden]

Yes, I need to add an ISNAN() (and perhaps ISNA() also for parity with Excel).

[Blake-Madden]

Yeah, in C++ comparing NaN with NaN will produce the IEEE standard results...but I'm just not liking that at all.

In Excel, =NA() = NA() and =NA() < NA() will yield #N/A. And in SQL, doing any comparison with NULL results in NULL. Unless you are really opposed to it, I'm going to make the presence of NaN in any comparison operation result in NaN to work the way Excel works. That just seems so much more natural to me.

[sylque]

I'm not opposed to it, but frankly both ways sound unnatural to me.
If I were in your shoes, I'd probably stick to the IEEE floating-point standard, if only to answer "standard!" to anyone complaining :-)

[sylque]

For those landing here:

Until ISNAN() is supported natively, it can be added manually using a custom function:

#include "tinyexpr.h"
#include <iostream>

double isnan(double val) { 
    return std::isnan(val) ? 1.0 : 0.0; 
}

int main() {
    const double x = std::numeric_limits<float>::quiet_NaN();
    
    te_parser tep;
    tep.set_variables_and_functions({{"x", &x}, {"ISNAN", isnan}});
    
    const double result = tep.evaluate("ISNAN(x)");
    
    std::cout << "The expression " << tep.get_expression() << " evaluates to " << result << "\n";
}