Improve NAN handling in ccmath

[mborland]

No description provided.

[pulver]

For the hypot(), log(), and sqrt() functions, shouldn't parameters with values of negative infinity result in NaN instead of negative infinity?

E.g. sqrt(-∞) = nan rather than -∞.

I know this behavior remains unchanged in this PR, but thought this might be an opportune time to fix.

[mborland]

For the hypot(), log(), and sqrt() functions, shouldn't parameters with values of negative infinity result in NaN instead of negative infinity?

E.g. sqrt(-∞) = nan rather than -∞.

I know this behavior remains unchanged in this PR, but thought this might be an opportune time to fix.

The language specification is +- infinity returns +- infinity. Since we fallback to the STL at runtime I didn't want the edge case handling to differ based on context.

[jzmaddock]

The language specification is +- infinity returns +- infinity. Since we fallback to the STL at runtime I didn't want the edge case handling to differ based on context.

C99 at least specifies that sqrt of a negative number is a domain error and yields a NaN, as things stand runtime sqrt(-INF) returns NaN and constexpr sqrt(-INF) returns +INF which looks just plain wrong?

[mborland]

The language specification is +- infinity returns +- infinity. Since we fallback to the STL at runtime I didn't want the edge case handling to differ based on context.

C99 at least specifies that sqrt of a negative number is a domain error and yields a NaN, as things stand runtime sqrt(-INF) returns NaN and constexpr sqrt(-INF) returns +INF which looks just plain wrong?

Ok for sqrt the standard only specifies +inf. Domain error is implementation defined so NaN would make sense.

[pulver]

According to https://en.cppreference.com/w/c/numeric/math/logb

If arg is ±∞, +∞ is returned

I haven't tested it, but inspecting the code it appears that for the constexpr version

logb(-∞) = -∞

Shouldn't it return +∞?

[mborland]

According to https://en.cppreference.com/w/c/numeric/math/logb

If arg is ±∞, +∞ is returned

I haven't tested it, but inspecting the code it appears that for the constexpr version

logb(-∞) = -∞

Shouldn't it return +∞?

Yes; the change is wrong.

[mborland]

@pulver Can you do one last dummy check for me? I think this is good now and addresses your sqrt(-∞) = nan.

[pulver]

For hypot(), it looks like

hypot(-∞,1) = -∞

but should return +∞. Same when x and y are swapped, based on hypot:

if one of the arguments is ±∞, hypot(x,y) returns +∞ even if the other argument is NaN

[pulver]

LGTM