abs_sub is strangely named

[huonw]

The C name is fdim, and the operation is described as "positive difference" by e.g. the C function's man page, which isn't really suggested by "abs_sub" at all.

[nagisa]

I disagree abs_sub is strangely named. At least it is much better than fdim. While it could certainly be abs_difference, absolute value_subtraction is fine IMO.

[ranma42]

Note that abs_sub does not compute the absolute value of the difference abs(x-y), but instead max(x-y, 0).

[jFransham]

@ranma42 abs(x) ≡ max(x, 0) (logically, even if not in implementation), so abs(x-y) ≡ max(x-y, 0)

EDIT: This comment is stupid and wrong and horribly embarrassing

[rkruppe]

@jFransham abs(-1) is 1, but max(-1, 0) is 0. Did you mean abs(x) = max(x, -x)? In any case, this playground demonstrates what @ranma42 stated.

[nagisa]

Good point, I guess abs_sub indeed does not make any sense at all.

[GuillaumeGomez]

I think one of the following names would fit better:

• positive_difference
• positive_diff
• pos_diff (but I don't really like this one, too short IMO)

[nagisa]

Monus as pointed out by sp3d on IRC. The page also mentions these names the operation is referred to:

• truncated subtraction;
• limited subtraction;
• proper subtraction.

[nagisa]

Most of the methods name the result (e.g. (10.0).log() stands for logarithm of 10.0), and only some name the operation (e.g. mul_add which stands for multiply-add). I’d prefer staying with the bunch of first bucket, which pretty much implies using difference for whatever comes in the 2nd part. So {}_difference or {}_diff.

[alexcrichton]

The libs team discussed this issue during triage the other day and the conclusion was that this is so confusingly named and sparingly used that we should probably just document it. The deprecation message can indicate:

• What you probably want, (a - b).abs(), is not what this function is doing and that code is shorter than calling the function.
• If you want this, you probably don't want to call it via a misleading name, and (a - b).max(0) is probably clearer and has the same semantics.

[huonw]

(a - b).max(0) is probably clearer and has the same semantics.

After the meetings, I realised this isn't quite true: fdim(NAN, 1) == NAN.abs_sub(1) == NAN, while (NAN - 1).max(0) == 0.

[HeroesGrave]

I just wasted somewhere between 3-4 hours over the last day debugging my game, thinking that this was equivalent to (a - b).abs().

I'm definitely in support of deprecation and renaming.