fcoalesce will not replace NaN with NA_real_

[ethanbsmith]

> fcoalesce(c(NaN, NaN), c(0,NA_real_))
[1]   0 NaN

I realize this is a very edge case, but the current implementation doesn't support simple consistent processing. since NA_real_ and NaN are not equal, i.e. is.nan(NA_real) is FALSE, i think this should apply the conversion. As it currently stands, the only workaround I can think of is to explicitly do a second pass through the data to explicitly check for and handle this condition

[shrektan]

#4025 has adressed a similar issue for nafill(). I think we should treat NaN as NA in fcoalesce() as well due to the fact that is.na(NaN) returns TRUE.

[MichaelChirico]

nafill has nan argument since #4025, we could probably reuse that here

[Mukulyadav2004]

Hi @MichaelChirico @tdhock
Can I work on this?
Since the helper for the nan argument was introduced via PR #4025 in utils.R, I'll update the R wrapper to include a nan parameter and use nan_is_na(nan) to convert user input to boolean for c. And modify coalesce.c - updating the header declaration and validating the logic within the REALSXP case, similar to the handling in nafill.

[MichaelChirico]

I think that's the nut of it, go ahead!

[Mukulyadav2004]

I initially misunderstood the issue scope. After implementing the logic, I now realize that the original output fcoalesce(c(NaN, NaN), c(0, NA_real_)) returning [0, NaN] is indeed correct according to COALESCE semantics and the real issue was lack of user control.
The original code used ISNAN() everywhere, forcing NaN to always be treated as missing with no user choice. So I added explicit nan parameter control following the nafill pattern as:
nan=NA (default)- NaN treated as missing
and using nan=NaN - NaN preserved as valid value

By example :
Before - only behaviour possible

> fcoalesce(c(NaN, NA_real_), c(1, 2))  
[1] [1, 2]

After implementation
treat it as missing

> fcoalesce(c(NaN, NA_real_), c(1, 2), nan=NA) 
[1] [1, 2]

preserve NaN

> fcoalesce(c(NaN, NA_real_), c(1, 2), nan=NaN)
[1] [1,2]

Can you please tell if this is correct.

[tdhock]

I don't use this function. Please read the man page and original issue to double check.

[MichaelChirico]

I would basically expect fcoalesce(x, y, nan=Z) to be consistent with nafill(x, fill=y, nan=Z) when y is length-1.