`fmt::isnan` triggers floating-point exception for NaN values

[alexdewar]

Note: floating-point exceptions are not the same as C++ exceptions! See: https://www.gnu.org/savannah-checkouts/gnu/libc/manual/html_node/Control-Functions.html

Commit ef54f9a changed the implementation of fmt::isnan from:

template <typename T> constexpr bool isnan(T value) {
  return value != value;  // std::isnan doesn't support __float128.
}

To:

template <typename T> constexpr bool isnan(T value) {
  return !(value >= value);  // std::isnan doesn't support __float128.
}

The commit description says that this was done to suppress -Wfloat-equal warnings. The problem is that these two implementations are subtly different: while they both return the same value, the newer implementation also sets a floating-point exception flag. The reason is that while for IEEE-754 floating-point numbers the operation NaN != NaN is well defined and will always return true, NaN >= NaN is technically an invalid operation, hence why the exception flag is set (at least this is my non-expert understanding).

This has been causing headaches for a project I'm working on because I wanted to trace where floating-point exceptions were emanating from in my program (using the GNU-specific feenableexcept function), but fmt::isnan will raise a spurious exception every time a NaN is passed in. Changing the code to use value != value instead fixes things.

I guess there are different ways to fix this, but maybe you could just return to using value != value and find another way to suppress the warning (e.g. with a comment or a compiler argument)?

[vitaut]

Thanks for reporting. Could you by any chance provide a godbolt repro that demonstrates the issue?

[alexdewar]

Sure. Here you go: https://godbolt.org/z/Pvc43M46n

[vitaut]

I think reverting to != makes sense. Could you submit a PR to do this and add a new test case that checks FP exceptions?

[alexdewar]

Sure!