Should math.h be implemented as hidden friends?

[mpusz]

Should we implement the functions from the mp_units/math.h header as non-member hidden friend functions in quantity and quantity_point?

[chiphogg]

Maybe!

The main benefit is that it would let us automatically support types that implicitly convert to, but not from, a quantity or quantity_point. (That said, mp-units has traditionally aligned itself against using this sort of type: zero was rejected, and mp-units constants do not convert to quantities.)

Another benefit is reducing the size of overload sets, which can make for shorter compiler error messages and, I think, faster compile times in a few cases.

The downside is that it's an intrusive change: we need to open up the guts of quantity and quantity_point to add it. Also, it locks us in to forcing a <cmath> dependency on all users who include these files. (Maybe this is already the case; I didn't check.)

Au has done this for min, max, clamp, and %, without the <cmath> dependency. We get a lot more benefit out of it than mp-units would, because of our "shapeshifter types" (i.e., Zero and Constant<U>).

Later on, I tried using this for remainder (which wraps std::remainder), passing a Constant as one of the arguments, and it failed. I realized that in order to get this to work, we'd need to take this hidden friend approach for remainder. It felt like opening up a bit of a Pandora's box, so I didn't add it at the time, and still haven't added it. I'm not sure yet how to decide which ones should or shouldn't get this treatment.

[chiphogg]

To be concrete: which specific functions were you considering?

[mpusz]

I thought about most if not all, functions from the mp_units/math.h header. Additionally, things like lerp and midpoint should be provided for quantity_point.

[mpusz]

and mp-units constants do not convert to quantities

Are constants in Au modeled as units that convert to quantities?

[chiphogg]

and mp-units constants do not convert to quantities

Are constants in Au modeled as units that convert to quantities?

No, they are constants that convert to quantities. 🙂 Constant<U> is a template that is templated on the unit U, and has a perfect conversion policy (no heuristic overflow safety surface needed) with to any Quantity<OtherU, R>: every lossy conversion fails, and every lossless conversion succeeds.

[mpusz]

But I assume that such a constant is a part of the unit of a quantity and simplifies at compile-time when needed?

[chiphogg]

But I assume that such a constant is a part of the unit of a quantity and simplifies at compile-time when needed?

I don't understand what you're asking, but I understand what Au does. 🙂 Every Constant<U> is templated only on the unit U; therefore, we need to have (or create) a unit for each constant. Constant<U> is a monovalue type, so this unit basically is its value: its value is "1 U", so to speak. If you try to assign this constant to a Quantity<U1, R1>, this assignment will succeed if 1 U can be exactly represented in Quantity<U1, R1>, and fail otherwise. And if it succeeds, the value in U1 and R1 will be computed at compile time.

The parts I didn't understand or found confusing were:

• "such a constant is a part of the unit of a quantity"
• "simplifies at compile-time when needed"

I hope my above paragraph answered the question.

[mpusz]

I mean something like this:

constexpr auto c = speed_of_light;
quantity q1 = 0.4 * c;
quantity q2 = q1 / c;

1. What is the unit of q1 (is it c or some other arbitrary unit of speed together with a proper compile-time magnitude)?
2. I assume that q1 stores just 0.4, and no calculations are done at this point.
3. I assume that q2 is a dimensionless quantity with exactly a value 0.4, and there was no runtime cost of the division in the initializer.

[chiphogg]

Thanks for giving a concrete example. I think the answers will be just as you expect (with the caveat that this syntax wouldn't work for Au because we don't have deduction guides for C++17's CTAD):

1. The units of q1 are the units of c.
2. Yep, we store just 0.4 under the hood, with no calculations.
3. Yep, q2 would be just a dimensionless quantity with exactly 0.4, and no runtime cost.

I'm a little uncertain for (3) as to whether we would produce a quantity or just a raw double, but if it's the latter then it's what we now consider a design defect and we plan to fix it in the next "breaking change" release (probably 0.5.0).

[mpusz]

Thanks. So, it seems the only drawback of this solution to what we have in mp-units is that a possibly long magnitude of a constant is visible in a quantity type. When two constants are involved, magnitude no longer has a "familiar" value. But you gained implicit constructions from a constant.

I also think that Au uses implicit conversions for operators. It means that things like c - 0.5 * c and -0.5 * c + c will probably work? In our case, where we do not have an implicit conversion and arguments are deduced function template parameters, we would need to add dedicated overloads for constants.

[chiphogg]

I also think that Au uses implicit conversions for operators. It means that things like c - 0.5 * c and -0.5 * c + c will probably work?

Apparently so!