To keep a small database indicating the physical dimensions of numerical values, and how they combine: for instance, allowing us to specify that a length times a length is an area.


§1. We sort quasinumerical kinds1 into three: fundamental units, derived units with dimensions, and dimensionless units. In the default setup provided by a work of IF generated by Inform, there is one fundamental unit ("time"), there are two dimensionless units ("number" and "real number") and no derived units. Dimension checking does very little in this minimal environment, though it will, for example, forbid an attempt to multiply 10 PM by 9:15 AM, or indeed to multiply kinds which aren't numberical at all, such as a text by a sound effect.2

Further fundamental units are created every time source text like this is read:

Mass is a kind of value. 1kg specifies a mass.

Mass will then be considered fundamental until the source text says otherwise. It would no doubt be cool to decide what is fundamental and what is derived by applying Buckingham's \(\pi\)-theorem to all the equations we need to use, but this is a tricky technique and does not always produce the "natural" results which people expect. So Inform requires the writer to specify explicitly how units combine. When it reads, for example,

A mass times an acceleration specifies a force.

Inform chooses one of the three units — say, force — and derives that from the others.

§2. Multiplication rules are stored in a linked list associated with the left operand; so that the rule \(A\) times \(B\) specifies \(C\) causes \((B, C)\) to be stored in the list of multiplications belonging to \(A\).

typedef struct dimensional_rules {
    struct dimensional_rule *multiplications;
} dimensional_rules;

typedef struct dimensional_rule {
    struct wording name;
    struct kind *right;
    struct kind *outcome;
    struct dimensional_rule *next;
} dimensional_rule;

§3. The derivation process can be seen in action by feeding Inform definitions of the SI units (see the test case SIUnits-G) and looking at the output of:

Test dimensions (internal) with --.

(The dash is meaningless — this is a test with no input.) In the output, we see that

    Base units: time, length, mass, elapsed time, electric current, temperature, luminosity
    Derived units:
    frequency = (elapsed time)-1
    force = (length).(mass).(elapsed time)-2
    energy = (length)2.(mass).(elapsed time)-2
    pressure = (length)-1.(mass).(elapsed time)-2
    power = (length)2.(mass).(elapsed time)-3
    electric charge = (elapsed time).(electric current)
    voltage = (length)2.(mass).(elapsed time)-3.(electric current)-1

...and so on. Those expressions on the right hand sides are "derived units", where the numbers are powers, so that negative numbers mean division. It's easy to see why we want to give names and notations for some of these derived units — imagine going into a cycle shop and asking for a \(5 {\rm m}^2\cdot{\rm kg}\cdot{\rm s}^{-3}\cdot{\rm A}^{-1}\) battery.

§4. A "dimensionless" quantity is one which is just a number, and is not a physical measurement as such. In an equation like $$ K = {{mv^2}\over{2}} $$ the 2 is clearly dimensionless, but other possibilities also exist. The arc length of part of a circle at radius \(r\) drawn out to angle \(\theta\) (if measured in radians) is given by: $$ A = \theta r $$ Here \(A\) and \(r\) are both lengths, so the angle \(\theta\) must be dimensionless. But clearly it's not quite conceptually the same thing as an ordinary number. Inform deduces dimensionlessness from multiplication laws like so:

Angle is a kind of value. 1 rad specifies an angle. Length times angle specifies a length.

Inform is not quite so careful about distinguishing dimensionless quantities as some physicists might be. The official SI units distinguish angle, measured in radians, and solid angle, in steradians, writing them as having units \({\rm m}\cdot{\rm m}^{-1}\) and \({\rm m}^2\cdot{\rm m}^{-2}\) respectively — one is a ratio of lengths, the other of areas. Inform cancels the units and sees them as dimensionally equal. So if we write:

Solid angle is a kind of value. 1 srad specifies an solid angle. Area times solid angle specifies an area.

then Inform treats angle and solid angle as having the same multiplicative properties — but it still allows variables to have either one as a kind of value, and prints them differently.

§5. In the process of calculations, we often need to create other and nameless units as partial answers of calculations. Consider the kinetic energy equation $$ K = {{mv^2}\over{2}} $$ being evaluated the way a computer does it, one step at a time. One way takes the mass, multiplies by the velocity to get a momentum, multiplies by the velocity again to get energy, then divides by a dimensionless constant. But another way would be to square the velocity first, then multiply by mass to get energy, then halve. If we do it that way, what units are the squared velocity in? The answer has to be

    (length)2.(elapsed time)-2

but that's a unit which isn't useful for much, and doesn't have any everyday name. Inform creates what are called "intermediate kinds" like this in order to be able to represent the kinds of intermediate values which turn up in calculation. They use the special CON_INTERMEDIATE construction, they are nameless, and the user isn't allowed to store the results permanently. (They can't be the kind of a global variable, a table column, and so on.) If the user wants to deal with such values on a long-term basis, he must give them a name, like this:

Funkiness is a kind of value. 1 Claude is a funkiness. A velocity times a velocity specifies a funkiness.

§6. Expressions like \({\rm m}^2\cdot{\rm kg}\) are stored inside Inform as sequences of ordered pairs in the form $$ ((B_1, p_1), (B_2, p_2), ..., (B_k, p_k)) $$ where each \(B_i\) is the type ID of a fundamental unit, each \(p_i\) is a non-zero integer, and \(B_1 < B_2 < ... < B_k\). For instance, energy would be $$ (({\rm length}, 2), ({\rm mass}, 1), ({\rm elapsed~time}, -2)). $$

Every physically different derived unit has a unique and distinct sequence. This is only true because a unit sequence is forbidden to contain derived units. For instance, specific heat capacity looks as if it is written with two different units in physics: $$ {\rm J}\cdot {\rm K}^{-1}\cdot {\rm kg}^{-1} \quad = \quad {\rm m}^2\cdot{\rm s}^{-2}\cdot{\rm K}^{-1} $$ But this is because the Joule is a derived unit. Substituting \({\rm J} = {\rm m}^2\cdot{\rm kg}\cdot{\rm s}^{-2}\) to get back to fundamental units shows that both sides would be computed as the same unit sequence.

The case \(k=0\), the empty sequence, is not only legal but important: it is the derivation for a dimensionless unit. (As discussed above, Inform doesn't see different dimensionless units as being physically different.)

typedef struct unit_pair {
    struct kind *fund_unit;  and this really must be a fundamental kind
    int power;  a non-zero integer
} unit_pair;

§7. The following is a hard limit, but really not a problematic one. The entire SI system has only 7 fundamental units, and the only named scientific unit I've seen which has even 5 terms in its derivation is molar entropy, a less than everyday chemical measure (\({\rm kg}\cdot{\rm m}^2\cdot{\rm s}^{-2}\cdot{\rm K}^{-1}\cdot{\rm mol}^{-1}\), if you're taking notes).

define MAX_BASE_UNITS_IN_SEQUENCE 16
typedef struct unit_sequence {
    int no_unit_pairs;  in range 0 to MAX_BASE_UNITS_IN_SEQUENCE
    struct unit_pair unit_pairs[MAX_BASE_UNITS_IN_SEQUENCE];
    int scaling_factor;  see discussion of scaling below
} unit_sequence;

§8. Manipulating units like \({\rm m}^2\cdot{\rm kg}\cdot{\rm s}^{-2}\) looks a little like manipulating formal polynomials in several variables, and of course that isn't an accident. Another way of thinking of the above is that we have a commutative ring \(R\) of underlying numbers, and extend by a pair of formal variables \(U_i\) and \(U_i^{-1}\) for each new kind, then quotient by the ideal generated by \(U_jU_j^{-1}\) and also by all of the derivations we know of. Thus Inform calculates in the ring: $$ I = R[U_1, U_2, ..., U_n, U_1^{-1}, ..., U_n^{-1}] / (U_1U_1^{-1}, U_2U_2^{-1}, ..., U_nU_n^{-1}, D_1, D_2, ..., D_i). $$ It does that in practice by eliminating all of the \(U_i\) and \(U_i^{-1}\) which are derived, so that it's left with just $$ I = R[U_1, U_2, ..., U_k, U_1^{-1}, ..., U_k^{-1}] / (U_1U_1^{-1}, U_2U_2^{-1}, ..., U_kU_k^{-1}). $$

For instance, given seconds, Watts and Joules, $$ I = R[{\rm s}, {\rm s}^{-1}, {\rm W}, {\rm W}^{-1}, {\rm J}, {\rm J}^{-1}]/ ({\rm s}{\rm s}^{-1} = 1, {\rm W}{\rm W}^{-1}=1, {\rm J}{\rm J}^{-1} = 1, {\rm s}{\rm W} = {\rm J}) $$ which by substituting all occurrences of J can be reduced to: $$ I = R[{\rm s}, {\rm s}^{-1}, {\rm W}, {\rm W}^{-1}]/ ({\rm s}{\rm s}^{-1} = 1, {\rm W}{\rm W}^{-1}=1). $$ Of course there are other ways to calculate \(I\) — we could have eliminated any of the three units and kept the other two.

If the derivations were ever more complex than \(AB=C\), we might have to use some elegant algorithms for calculating Gröbner bases in order to determine \(I\). But Inform's syntax is such that the writer of the source text gives us the simplest possible description of the ideal, so no such fun occurs.

§9. But enough abstraction: time for some arithmetic. Inform performs checking whenever values from two different kinds are combined by any of eight arithmetic operations, numbered as follows. The numbers must not be changed without amending the definitions of "plus" and so on in the Basic Inform extension.

define NO_OPERATIONS 9
define PLUS_OPERATION 0  addition
define MINUS_OPERATION 1  subtraction
define TIMES_OPERATION 2  multiplication
define DIVIDE_OPERATION 3  division
define REMAINDER_OPERATION 4  remainder after division
define APPROXIMATION_OPERATION 5  "X to the nearest Y"
define ROOT_OPERATION 6  square root — a unary operation
define REALROOT_OPERATION 7  real-valued square root — a unary operation
define CUBEROOT_OPERATION 8  cube root — similarly unary
define EQUALS_OPERATION 9  set equal — used only in equations
define POWER_OPERATION 10  raise to integer power — used only in equations
define UNARY_MINUS_OPERATION 11  unary minus — used only in equations

§10. The following is associated with "total...", as in "the total weight of things on the table", but for dimensional purposes we ignore it.

define TOTAL_OPERATION 12  not really one of the above

§11. Prior kinds. It turns out to be convenient to have a definition ordering of fundamental kinds, which is completely unlike the \(\leq\) relation; it just places them in order of creation.

int Kinds::Dimensions::kind_prior(kind *A, kind *B) {
    if (A == NULL) {
        if (B == NULL) return FALSE;
        return TRUE;
    }
    if (B == NULL) {
        if (A == NULL) return FALSE;
        return FALSE;
    }
    if (Kinds::get_construct(A)->allocation_id <
        Kinds::get_construct(B)->allocation_id) return TRUE;
    return FALSE;
}

§12. Multiplication lists. The linked lists of multiplication rules begin empty for every kind:

void Kinds::Dimensions::dim_initialise(dimensional_rules *dimrs) {
    dimrs->multiplications = NULL;
}

§13. And this adds a new one to the relevant list:

void Kinds::Dimensions::record_multiplication_rule(kind *left, kind *right, kind *outcome) {
    dimensional_rules *dimrs = Kinds::Behaviour::get_dim_rules(left);
    dimensional_rule *dimr;

    for (dimr = dimrs->multiplications; dimr; dimr = dimr->next)
        if (dimr->right == right) {
            KindsModule::problem_handler(DimensionRedundant_KINDERROR, NULL, NULL, NULL, NULL);
            return;
        }

    dimensional_rule *dimr_new = CREATE(dimensional_rule);
    dimr_new->right = right;
    dimr_new->outcome = outcome;
    if (current_sentence)
        dimr_new->name = Node::get_text(current_sentence);
    else
        dimr_new->name = EMPTY_WORDING;
    dimr_new->next = dimrs->multiplications;
    dimrs->multiplications = dimr_new;
}

§14. The following loop-header macro iterates through the possible triples \((L, R, O)\) of multiplication rules \(L\times R = O\).

define LOOP_OVER_MULTIPLICATIONS(left_operand, right_operand, outcome_type, wn)
    dimensional_rules *dimrs;
    dimensional_rule *dimr;
    LOOP_OVER_BASE_KINDS(left_operand)
        for (dimrs = Kinds::Behaviour::get_dim_rules(left_operand),
            dimr = (dimrs)?(dimrs->multiplications):NULL,
            wn = (dimr)?(Wordings::first_wn(dimr->name)):-1,
            right_operand = (dimr)?(dimr->right):0,
            outcome_type = (dimr)?(dimr->outcome):0;
            dimr;
            dimr = dimr->next,
            wn = (dimr)?(Wordings::first_wn(dimr->name)):-1,
            right_operand = (dimr)?(dimr->right):0,
            outcome_type = (dimr)?(dimr->outcome):0)

§15. And this is where the user asks for a multiplication to come out in a particular way:

void Kinds::Dimensions::dim_set_multiplication(kind *left, kind *right,
    kind *outcome) {
    if ((Kinds::is_proper_constructor(left)) ||
        (Kinds::is_proper_constructor(right)) ||
        (Kinds::is_proper_constructor(outcome))) {
        KindsModule::problem_handler(DimensionNotBaseKOV_KINDERROR, NULL, NULL, NULL, NULL);
        return;
    }
    if ((Kinds::Behaviour::is_quasinumerical(left) == FALSE) ||
        (Kinds::Behaviour::is_quasinumerical(right) == FALSE) ||
        (Kinds::Behaviour::is_quasinumerical(outcome) == FALSE)) {
        KindsModule::problem_handler(NonDimensional_KINDERROR, NULL, NULL, NULL, NULL);
        return;
    }
    Kinds::Dimensions::record_multiplication_rule(left, right, outcome);
    if ((Kinds::eq(left, outcome)) && (Kinds::eq(right, K_number))) return;
    if ((Kinds::eq(right, outcome)) && (Kinds::eq(left, K_number))) return;
    Kinds::Dimensions::make_unit_derivation(left, right, outcome);
}

§16. Unary operations. All we need to know is which ones are unary, in fact, and:

int Kinds::Dimensions::arithmetic_op_is_unary(int op) {
    switch (op) {
        case CUBEROOT_OPERATION:
        case ROOT_OPERATION:
        case REALROOT_OPERATION:
        case UNARY_MINUS_OPERATION:
            return TRUE;
    }
    return FALSE;
}

§17. Euclid's algorithm. In my entire life, I believe this is the only time I have ever actually used Euclid's algorithm for the GCD of two natural numbers. I've never quite understood why textbooks take this as somehow the typical algorithm. My maths students always find it a little oblique, despite the almost trivial proof that it works. It typically takes a shade under \(\log n\) steps, which is nicely quick. But I don't look at the code and immediately see this, myself.

int Kinds::Dimensions::gcd(int m, int n) {
    if ((m<1) || (n<1)) internal_error("applied gcd outside natural numbers");
    while (TRUE) {
        int rem = m%n;
        if (rem == 0) return n;
        m = n; n = rem;
    }
}

§18. The sequence of operation here is to reduce the risk of integer overflows when multiplying m by n.

int Kinds::Dimensions::lcm(int m, int n) {
    return (m/Kinds::Dimensions::gcd(m, n))*n;
}

§19. Unit sequences. Given a fundamental type \(B\), convert it to a unit sequence: \(B = B^1\), so we get a sequence with a single pair: \(((B, 1))\). Uniquely, number is born derived and dimensionless, though, so that comes out as the empty sequence.

unit_sequence Kinds::Dimensions::fundamental_unit_sequence(kind *B) {
    unit_sequence us;
    if (B == NULL) {
        us.no_unit_pairs = 0;
        us.unit_pairs[0].fund_unit = NULL;
        us.unit_pairs[0].power = 0;  redundant, but appeases compilers
    } else {
        us.no_unit_pairs = 1;
        us.unit_pairs[0].fund_unit = B;
        us.unit_pairs[0].power = 1;
    }
    return us;
}

§20. As noted above, two units represent dimensionally equivalent physical quantities if and only if they are identical, which makes comparison easy:

int Kinds::Dimensions::compare_unit_sequences(unit_sequence *ik1, unit_sequence *ik2) {
    int i;
    if (ik1 == ik2) return TRUE;
    if ((ik1 == NULL) || (ik2 == NULL)) return FALSE;
    if (ik1->no_unit_pairs != ik2->no_unit_pairs) return FALSE;
    for (i=0; i<ik1->no_unit_pairs; i++)
        if ((Kinds::eq(ik1->unit_pairs[i].fund_unit,
            ik2->unit_pairs[i].fund_unit) == FALSE) ||
            (ik1->unit_pairs[i].power != ik2->unit_pairs[i].power))
                return FALSE;
    return TRUE;
}

§21. We now have three fundamental operations we can perform on unit sequences. First, we can multiply them: that is, we store in result the unit sequence representing \(X_1^{s_1}X_2^{s_2}\), where \(X_1\) and \(X_2\) are represented by unit sequences us1 and us2.

So the case \(s_1 = s_2 = 1\) represents multiplying \(X_1\) by \(X_2\), while \(s_1 = 1, s_2 = -1\) represents dividing \(X_1\) by \(X_2\). But we can also raise to higher powers.

Our method relies on noting that \(\) X_1 = T_{11}^{p_{11}}\cdot T_{12}^{p_{12}}\cdots T_{1n}^{p_{1n}},\qquad X_2 = T_{21}^{p_{21}}\cdot T_{22}^{p_{22}}\cdots T_{2m}^{p_{2m}} \(\) where \(T_{11} < T_{12} < ... < T_{1n}\) and \(T_{21}<T_{22}<...<T_{2m}\). We can therefore merge the two in a single pass.

On each iteration of the loop the variables i1 and i2 are our current read positions in each sequence, while we are currently looking at the unit pairs (t1, m1) and (t2, m2). The following symmetrical algorithm holds on to each pair until the one from the other sequence has had a chance to catch up with it, because we always deal with the pair with the numerically lower t first. This also proves that the results sequence comes out in numerical order.

void Kinds::Dimensions::multiply_unit_sequences(unit_sequence *us1, int s1,
    unit_sequence *us2, int s2, unit_sequence *result) {
    if ((result == us1) || (result == us2)) internal_error("result must be different structure");

    result->no_unit_pairs = 0;

    int i1 = 0, i2 = 0;  read position in sequences 1, 2
    kind *t1 = NULL; int p1 = 0;  start with no current term from sequence 1
    kind *t2 = NULL; int p2 = 0;  start with no current term from sequence 2
    while (TRUE) {
        If we have no current term from sequence 1, and it hasn't run out, fetch a new one21.1;
        If we have no current term from sequence 2, and it hasn't run out, fetch a new one21.2;
        if (Kinds::eq(t1, t2)) {
            if (t1 == NULL) break;  both sequences have now run out
            Both terms refer to the same fundamental unit, so combine these into the result21.3;
        } else {
            Different fundamental units, so copy the numerically lower one into the result21.4;
        }
    }
    LOGIF(KIND_CREATIONS, "Multiplication: $Q * $Q = $Q\n", us1, us2, result);
}

§21.1. If we have no current term from sequence 1, and it hasn't run out, fetch a new one21.1 =

    if ((t1 == NULL) && (us1) && (i1 < us1->no_unit_pairs)) {
        t1 = us1->unit_pairs[i1].fund_unit; p1 = us1->unit_pairs[i1].power; i1++;
    }

§21.2. If we have no current term from sequence 2, and it hasn't run out, fetch a new one21.2 =

    if ((t2 == NULL) && (us2) && (i2 < us2->no_unit_pairs)) {
        t2 = us2->unit_pairs[i2].fund_unit; p2 = us2->unit_pairs[i2].power; i2++;
    }

§21.3. So here the head of one sequence is \(T^{p_1}\) and the head of the other is \(T^{p_2}\), so in the product we ought to see $(T^{p_1})^{s_1}\cdot (T^{p_2})^{s_2} = T^{p_1s_1+p_2s_2}$. But we don't enter terms that have cancelled out, that is, where \(p_1s_1+p_2s_2\) = 0.

Both terms refer to the same fundamental unit, so combine these into the result21.3 =

    int p = p1*s1 + p2*s2;  combined power of t1 \(=\) t2
    if (p != 0) {
        if (result->no_unit_pairs == MAX_BASE_UNITS_IN_SEQUENCE)
            Trip a unit sequence overflow21.3.1;
        result->unit_pairs[result->no_unit_pairs].fund_unit = t1;
        result->unit_pairs[result->no_unit_pairs++].power = p;
    }
    t1 = NULL; t2 = NULL;  dispose of both terms as dealt with

§21.4. Otherwise we copy. By copying the numerically lower term, we can be sure that it will never occur again in either sequence. So we can copy it straight into the results.

The code is slightly warped by the fact that UNKNOWN_NT, representing the end of the sequence, happens to be numerically lower than all the valid kinds. We don't want to make use of facts like that, so we write code to deal with UNKNOWN_NT explicitly.

Different fundamental units, so copy the numerically lower one into the result21.4 =

    if ((t2 == NULL) || ((t1 != NULL) && (Kinds::Dimensions::kind_prior(t1, t2)))) {
        if (result->no_unit_pairs == MAX_BASE_UNITS_IN_SEQUENCE)
            Trip a unit sequence overflow21.3.1;
        result->unit_pairs[result->no_unit_pairs].fund_unit = t1;
        result->unit_pairs[result->no_unit_pairs++].power = p1*s1;
        t1 = NULL;  dispose of the head of sequence 1 as dealt with
    } else if ((t1 == NULL) || ((t2 != NULL) && (Kinds::Dimensions::kind_prior(t2, t1)))) {
        if (result->no_unit_pairs == MAX_BASE_UNITS_IN_SEQUENCE)
            Trip a unit sequence overflow21.3.1;
        result->unit_pairs[result->no_unit_pairs].fund_unit = t2;
        result->unit_pairs[result->no_unit_pairs++].power = p2*s2;
        t2 = NULL;  dispose of the head of sequence 1 as dealt with
    } else internal_error("unit pairs disarrayed");

§21.3.1. For reasons explained above, this is really never going to happen by accident, but we'll be careful:

Trip a unit sequence overflow21.3.1 =

    KindsModule::problem_handler(UnitSequenceOverflow_KINDERROR, NULL, NULL, NULL, NULL);
    return;

§22. The second operation is taking roots.

Surprisingly, perhaps, it's much easier to compute \(\sqrt{X}\) or \(^{3}\sqrt{X}\) for any unit \(X\) — it's just that it can't always be done. Inform does not permit non-integer powers of units, so for instance \(\sqrt{{\rm time}}\) does not exist, whereas \(\sqrt{{\rm length}^2\cdot{\rm mass}^{-2}}\) does. Square roots exist if each power in the sequence is even, cube roots exist if each is divisible by 3. We return TRUE or FALSE according to whether the root could be taken, and if FALSE then the contents of result are undefined.

int Kinds::Dimensions::root_unit_sequence(unit_sequence *us, int pow, unit_sequence *result) {
    if (us == NULL) return FALSE;
    *result = *us;
    for (int i=0; i<result->no_unit_pairs; i++) {
        if ((result->unit_pairs[i].power) % pow != 0) return FALSE;
        result->unit_pairs[i].power = (result->unit_pairs[i].power)/pow;
    }
    return TRUE;
}

§23. More generally, we can raise a unit sequence to the rational power \(n/m\), though subject to the same stipulations:

kind *Kinds::Dimensions::to_rational_power(kind *F, int n, int m) {
    if ((n < 1) || (m < 1)) internal_error("bad rational power");
    if (Kinds::Dimensions::dimensionless(F)) return F;
    kind *K = K_number;
    int op = TIMES_OPERATION;
    if (n < 0) { n = -n; op = DIVIDE_OPERATION; }
    while (n > 0) {
        K = Kinds::Dimensions::arithmetic_on_kinds(K, F, op);
        n--;
    }
    if (m == 1) return K;

    unit_sequence result;
    unit_sequence *operand = Kinds::Behaviour::get_dimensional_form(K);
    if (Kinds::Dimensions::root_unit_sequence(operand, m, &result) == FALSE) return NULL;
    Identify the result as a known kind, if possible23.1;
    return NULL;
}

§24. The final operation on unit sequences is substitution. Given a fundamental type \(B\), we substitute \(B = K_D\) into an existing unit sequence \(K_E\). (This is used when \(B\) is becoming a derived type — once we discover that \(B=K_D\), we are no longer allowed to keep \(B\) in any unit sequence.)

We simply search for \(B^p\), and if we find it, we remove it and then multiply by \(K_D^p\).

void Kinds::Dimensions::dim_substitute(unit_sequence *existing, kind *fundamental,
    unit_sequence *derived) {
    int i, j, p = 0, found = FALSE;
    if (existing == NULL) return;
    for (i=0; i<existing->no_unit_pairs; i++)
        if (Kinds::eq(existing->unit_pairs[i].fund_unit, fundamental)) {
            p = existing->unit_pairs[i].power;
            found = TRUE;
            Remove the B term from the existing sequence24.1;
        }
    if (found)
        Multiply the existing sequence by a suitable power of B's derivation24.2;
}

§24.1. We shuffle the remaining terms in the sequence down by one, overwriting B:

Remove the B term from the existing sequence24.1 =

    for (j=i; j<existing->no_unit_pairs-1; j++)
        existing->unit_pairs[j] = existing->unit_pairs[j+1];
    existing->no_unit_pairs--;

§24.2. We now multiply by \(K_D^p\).

Multiply the existing sequence by a suitable power of B's derivation24.2 =

    unit_sequence result;
    Kinds::Dimensions::multiply_unit_sequences(existing, 1, derived, p, &result);
    *existing = result;

§25. For reasons which will be explained in Scaled Arithmetic Values, a unit sequence also has a scale factor associated with it:

int Kinds::Dimensions::us_get_scaling_factor(unit_sequence *us) {
    if (us == NULL) return 1;
    return us->scaling_factor;
}

§26. That just leaves, as usual, indexing...

void Kinds::Dimensions::index_unit_sequence(OUTPUT_STREAM, unit_sequence *deriv, int briefly) {
    if (deriv == NULL) return;
    if (deriv->no_unit_pairs == 0) { WRITE("dimensionless"); return; }

    for (int j=0; j<deriv->no_unit_pairs; j++) {
        kind *fundamental = deriv->unit_pairs[j].fund_unit;
        int power = deriv->unit_pairs[j].power;
        if (briefly) {
            if (j>0) WRITE(".");
            WRITE("(");
            #ifdef PIPELINE_MODULE
            Kinds::Textual::write_as_HTML(OUT, fundamental);
            #else
            Kinds::Textual::write(OUT, fundamental);
            #endif
            WRITE(")");
            if (power != 1) WRITE("<sup>%d</sup>", power);
        } else {
            if (j>0) WRITE(" times ");
            if (power < 0) { power = -power; WRITE("reciprocal of "); }
            wording W = Kinds::Behaviour::get_name(fundamental, FALSE);
            WRITE("%W", W);
            switch (power) {
                case 1: break;
                case 2: WRITE(" squared"); break;
                case 3: WRITE(" cubed"); break;
                default: WRITE(" to the power %d", power); break;
            }
        }
    }
}

§27. ...and logging.

void Kinds::Dimensions::logger(OUTPUT_STREAM, void *vUS) {
    unit_sequence *deriv = (unit_sequence *) vUS;
    if (deriv == NULL) { WRITE("<null-us>"); return; }
    if (deriv->no_unit_pairs == 0) { WRITE("dimensionless"); return; }
    for (int j=0; j<deriv->no_unit_pairs; j++) {
        if (j>0) WRITE(".");
        WRITE("(%u)", deriv->unit_pairs[j].fund_unit);
        if (deriv->unit_pairs[j].power != 1) WRITE("%d", deriv->unit_pairs[j].power);
    }
}

§28. Performing derivations. The following is called when the user specifies that \(L\) times \(R\) specifies an \(O\). Any of the three might be either a fundamental unit (so far) or a derived unit (already).

If two or more are fundamental units, we have a choice. That is, suppose we have created three kinds already: mass, acceleration, force. Then we read:

Mass times acceleration specifies a force.

We could make this true in any of three ways: keep M and A as fundamental units and derive F from them, keep A and F as fundamental units and derive M from those, or keep M and F while deriving A. Inform always chooses the most recently created unit as the one to derive, on the grounds that the source text has probably set things out with what the user thinks are the most fundamental units first.

void Kinds::Dimensions::make_unit_derivation(kind *left, kind *right, kind *outcome) {
    kind *terms[3];
    terms[0] = left; terms[1] = right; terms[2] = outcome;
    int newest_term = -1;
    Find which (if any) of the three units is the newest-made fundamental unit28.1;
    if (newest_term >= 0) {
        unit_sequence *derivation = NULL;
        Derive the newest one by rearranging the equation in terms of the other two28.2;
        Substitute this new derivation to eliminate this fundamental unit from other sequences28.3;
    } else
        Check this derivation to make sure it is redundant, not contradictory28.4;
}

§28.1. Data type IDs are allocated in creation order, so "newest" means largest ID.

Find which (if any) of the three units is the newest-made fundamental unit28.1 =

    int i; kind *max = NULL;
    for (i=0; i<3; i++)
        if ((Kinds::Dimensions::kind_prior(max, terms[i])) && (Kinds::Behaviour::test_if_derived(terms[i]) == FALSE)) {
            newest_term = i; max = terms[i];
        }

§28.2. We need to ensure that the user's multiplication rule is henceforth true, and we do that by fixing the newest unit to make it so.

Derive the newest one by rearranging the equation in terms of the other two28.2 =

    unit_sequence *kx = NULL, *ky = NULL; int sx = 0, sy = 0;
    switch (newest_term) {
        case 0:  here L is newest and we derive L = O/R
            kx = Kinds::Behaviour::get_dimensional_form(terms[1]); sx = -1;
            ky = Kinds::Behaviour::get_dimensional_form(terms[2]); sy = 1;
            break;
        case 1:  here R is newest and we derive R = O/L
            kx = Kinds::Behaviour::get_dimensional_form(terms[0]); sx = -1;
            ky = Kinds::Behaviour::get_dimensional_form(terms[2]); sy = 1;
            break;
        case 2:  here O is newest and we derive O = LR
            kx = Kinds::Behaviour::get_dimensional_form(terms[0]); sx = 1;
            ky = Kinds::Behaviour::get_dimensional_form(terms[1]); sy = 1;
            break;
    }
    derivation = Kinds::Behaviour::get_dimensional_form(terms[newest_term]);
    unit_sequence result;
    Kinds::Dimensions::multiply_unit_sequences(kx, sx, ky, sy, &result);
    *derivation = result;
    Kinds::Behaviour::now_derived(terms[newest_term]);

§28.3. Later in Inform's run, when we start compiling code, many more unit sequences will exist on a temporary basis — as part of the kinds for intermediate results in calculations — but early on, when we're here, the only unit sequences made are the derivations of the units. So it is easy to cover all of them.

Substitute this new derivation to eliminate this fundamental unit from other sequences28.3 =

    kind *R;
    LOOP_OVER_BASE_KINDS(R)
        if (Kinds::Behaviour::is_quasinumerical(R)) {
            unit_sequence *existing = Kinds::Behaviour::get_dimensional_form(R);
            Kinds::Dimensions::dim_substitute(existing, terms[newest_term], derivation);
        }

§28.4. If we have \(AB = C\) but all three of \(A\), \(B\), \(C\) are already derived, that puts us in a bind. Their definitions are fixed already, so we can't simply force the equation to come true by fixing one of them. That means either the derivation is redundant — because it's already true that \(AB = C\) — or contradictory — because we know \(AB\neq C\). We silently allow a redundancy, as it may have been put in for clarity, or so that the user can check the consistency of his own definitions, or to make the Kinds index page more helpful. But we must reject a contradiction.

Check this derivation to make sure it is redundant, not contradictory28.4 =

    unit_sequence product;
    Kinds::Dimensions::multiply_unit_sequences(
        Kinds::Behaviour::get_dimensional_form(terms[0]), 1,
        Kinds::Behaviour::get_dimensional_form(terms[1]), 1, &product);
    if (Kinds::Dimensions::compare_unit_sequences(&product,
        Kinds::Behaviour::get_dimensional_form(terms[2])) == FALSE)
        KindsModule::problem_handler(DimensionsInconsistent_KINDERROR, NULL, NULL, NULL, NULL);

§29. Classifying the units. Some of the derived units are dimensionless, others not. number and real number are always dimensionless, and any unit whose derivation is the empty unit sequence must be dimensionless.

int Kinds::Dimensions::dimensionless(kind *K) {
    if (K == NULL) return FALSE;
    if (Kinds::eq(K, K_number)) return TRUE;
    if (Kinds::eq(K, K_real_number)) return TRUE;
    if (Kinds::Behaviour::is_quasinumerical(K) == FALSE) return FALSE;
    return Kinds::Dimensions::us_dimensionless(Kinds::Behaviour::get_dimensional_form(K));
}

int Kinds::Dimensions::us_dimensionless(unit_sequence *us) {
    if ((us) && (us->no_unit_pairs == 0)) return TRUE;
    return FALSE;
}

int Kinds::Dimensions::kind_is_derived(kind *K) {
    if (Kinds::is_intermediate(K)) return TRUE;
    if ((Kinds::Behaviour::is_quasinumerical(K)) &&
        (Kinds::Behaviour::test_if_derived(K) == TRUE) &&
        (Kinds::Dimensions::dimensionless(K) == FALSE)) return TRUE;
    return FALSE;
}

§30. Logging. This is used by the internal "dimensions" test of Inform:

void Kinds::Dimensions::log_unit_analysis(void) {
    LOG("Dimensionless: ");
    int c = 0; kind *R;
    LOOP_OVER_BASE_KINDS(R)
        if (Kinds::Dimensions::dimensionless(R)) { if (c++ > 0) LOG(", "); LOG("%u", R); }
    LOG("\nBase units: ");
    c = 0;
    LOOP_OVER_BASE_KINDS(R)
        if ((Kinds::Dimensions::dimensionless(R) == FALSE) &&
            (Kinds::Dimensions::kind_is_derived(R) == FALSE) &&
            (Kinds::Behaviour::is_quasinumerical(R)))
        { if (c++ > 0) LOG(", "); LOG("%u", R); }
    LOG("\nDerived units:\n");
    LOOP_OVER_BASE_KINDS(R)
        if ((Kinds::Dimensions::kind_is_derived(R)) && (Kinds::is_intermediate(R) == FALSE)) {
            unit_sequence *deriv = Kinds::Behaviour::get_dimensional_form(R);
            LOG("%u = $Q\n", R, deriv);
        }
}

§31. Arithmetic on kinds. We are finally able to provide our central routine, the one providing a service for the rest of Inform. Given K1 and K2, we return the kind resulting from applying arithmetic operation op, or NULL if the operation cannot meaningfully be applied. In the case where op is a unary operation, K2 has no significance and should be NULL.

kind *Kinds::Dimensions::arithmetic_on_kinds(kind *K1, kind *K2, int op) {
    if (K1 == NULL) return NULL;
    if ((Kinds::Dimensions::arithmetic_op_is_unary(op) == FALSE) && (K2 == NULL)) return NULL;

    unit_sequence *operand1 = Kinds::Behaviour::get_dimensional_form(K1);
    if (operand1 == NULL) return NULL;
    unit_sequence *operand2 = Kinds::Behaviour::get_dimensional_form(K2);
    if ((Kinds::Dimensions::arithmetic_op_is_unary(op) == FALSE) && (operand2 == NULL)) return NULL;

    unit_sequence result;
    Calculate the result unit sequence, or return null if this is impossible31.1;
    Handle calculations entirely between dimensionless units more delicately31.2;
    Promote dimensionless numbers to real if necessary31.3;
    Identify the result as a known kind, if possible23.1;
    And otherwise create a kind as the intermediate result of a calculation31.4;
}

§31.1. Some operations — like addition — cannot be performed on mixed dimensions, and roots can only be taken where fractional powers are avoided, so we sometimes have to give up here and return NULL. Otherwise, though, the functions above make it possible to work out the correct unit sequence.

It's an interesting question what the result of a remainder should be, in dimensional terms. Clearly the remainder after dividing 90kg by 20 is 10kg. Inform says the remainder after dividing 90kg by 20kg is also 10kg. There's an argument that it ought to be 10, but if \(n = qm + r\) then the remainder \(r\) must have the dimensions of \(n\) (here 90kg) and also of \(qm\) (here 4 times 20kg), so it has to be a weight, not a dimensionless number.

Calculate the result unit sequence, or return null if this is impossible31.1 =

    switch (op) {
        case PLUS_OPERATION:
        case MINUS_OPERATION:
        case EQUALS_OPERATION:
        case APPROXIMATION_OPERATION:
            if (Kinds::Dimensions::compare_unit_sequences(operand1, operand2)) {
                result = *operand1;
                break;
            }
            return NULL;
        case REMAINDER_OPERATION:
        case UNARY_MINUS_OPERATION:
            result = *operand1;
            break;
        case ROOT_OPERATION:
            if (Kinds::Dimensions::root_unit_sequence(operand1, 2, &result) == FALSE)
                return NULL;
            break;
        case REALROOT_OPERATION:
            if (Kinds::Dimensions::root_unit_sequence(operand1, 2, &result) == FALSE)
                return NULL;
            break;
        case CUBEROOT_OPERATION:
            if (Kinds::Dimensions::root_unit_sequence(operand1, 3, &result) == FALSE)
                return NULL;
            break;
        case TIMES_OPERATION:
            Kinds::Dimensions::multiply_unit_sequences(operand1, 1, operand2, 1, &result);
            break;
        case DIVIDE_OPERATION:
            Kinds::Dimensions::multiply_unit_sequences(operand1, 1, operand2, -1, &result);
            break;
        default: return NULL;
    }

§31.2. If result is the empty unit sequence, we'll identify it as a number, because number is the lowest type ID representing a dimensionless unit. Usually that's good: for instance, it says that a frequency times a time is a number, and not some more exotic dimensionless quantity like an angle.

But it's not so good when the calculation is not really physical at all, but purely mathematical, and all we are doing is working on dimensionless units. For instance, if take an angle \(\theta\) and double it to \(2\theta\), we don't want Inform to say the result is number — we want \(2\theta\) to be another angle. So we make an exception.

Handle calculations entirely between dimensionless units more delicately31.2 =

    if (Kinds::Dimensions::arithmetic_op_is_unary(op)) {
        if ((op == REALROOT_OPERATION) && (Kinds::eq(K1, K_number)))
            return K_real_number;
        if (Kinds::Dimensions::dimensionless(K1)) return K1;
    } else {
        if ((Kinds::Dimensions::dimensionless(K1)) &&
            (Kinds::Dimensions::dimensionless(K2))) {
            if (Kinds::eq(K2, K_number)) return K1;
            if (Kinds::eq(K1, K_number)) return K2;
            if (Kinds::eq(K1, K2)) return K1;
        }
    }

§31.3. It's also possible to get a dimensionless result by, for example, dividing a mass by another mass, and we need to be careful to keep track of whether we're using real or integer arithmetic: 1500.0m divided by 10.0m must be 150.0, not 150.

Promote dimensionless numbers to real if necessary31.3 =

    if (Kinds::Dimensions::us_dimensionless(&result)) {
        if (Kinds::Dimensions::arithmetic_op_is_unary(op)) {
            if (Kinds::FloatingPoint::uses_floating_point(K1)) return K_real_number;
            return K_number;
        } else {
            if ((Kinds::FloatingPoint::uses_floating_point(K1)) ||
                (Kinds::FloatingPoint::uses_floating_point(K2))) return K_real_number;
            return K_number;
        }
    }

§23.1. If we've produced the right combination of fundamental units to make one of the named units, then we return that as an atomic kind. For instance, maybe we divided a velocity by a time, and now we find that we have \({\rm m}\cdot {\rm s}^{-2}\), which turns out to have a name: acceleration.

Identify the result as a known kind, if possible23.1 =

    kind *R;
    LOOP_OVER_BASE_KINDS(R)
        if (Kinds::Dimensions::compare_unit_sequences(&result,
            Kinds::Behaviour::get_dimensional_form(R)))
            return R;

§31.4. Otherwise the result is a unit sequence which doesn't have a name, so we store it as an intermediate kind, representing a temporary value living only for the duration of a calculation.

A last little wrinkle is: how we should scale this? For results like an acceleration, something defined in the source text, we know how accurate the author wants us to be. But these intermediate kinds are not defined, and we don't know for sure what the author would want. It seems wise to set \(k \geq k_X\) and \(k\geq k_Y\), so that we have at least as much detail as the calculation would have had within each operand kind. So perhaps we should put \(k = {\rm max}(k_X, k_Y)\). But in fact we will choose \(k\) = Kinds::Dimensions::lcm(k_X, k_Y), the least common multiple, so that any subsequent divisions will cancel correctly and we won't lose too much information through integer rounding. (In practice this will probably either be the same as \({\rm max}(k_X, k_Y)\) or will multiply by 6, since Kinds::Dimensions::lcm(60, 1000) == 6000 and so on.)

The same unit sequence can have different scalings each time it appears as an intermediate calculation. We could get to \({\rm m}^2\cdot {\rm kg}\) either as \({\rm m}\cdot{\rm kg}\) times \({\rm m}\), or as \({\rm m^2}\) times \({\rm kg}\), or many other ways, and we'll get different scalings depending on the route. This is why the unit_sequence structure has a scaling_factor field; the choice of scale factor does not depend on the physics but on the arithmetic method being used.

And otherwise create a kind as the intermediate result of a calculation31.4 =

    result.scaling_factor = Kinds::Dimensions::lcm(Kinds::Behaviour::scale_factor(K1), Kinds::Behaviour::scale_factor(K2));
    return Kinds::intermediate_construction(&result);