To keep track of requirements on the terms for a binary predicate.
§1. Different BPs apply to different sorts of terms: for instance, the numerical less-than comparison applies to numbers, whereas containment applies to things. The two terms need not have the same domain: the "wearing" relation, as seen in
Harry Smythe wears the tweed waistcoat.
is a binary predicate \(W(x_0, x_1)\) such that \(x_0\) ranges across people and \(x_1\) ranges across things.
It's therefore helpful to record requirements on any given term of a BP, and we do that with the following structure.1
1 Unary predicates do not use this structure, because it is slightly more efficient to roughly duplicate this arrangement than to use it directly. But the ideas are exactly the same. ↩
default TERM_DOMAIN_CALCULUS_TYPE struct kind
typedef struct bp_term_details { struct wording called_name; "(called...)" name, if any exists TERM_DOMAIN_CALCULUS_TYPE *implies_infs; the domain of values allowed struct kind *implies_kind; the kind of these values struct i6_schema *function_of_other; where one term can be deduced from the other char *index_term_as; usually null, but if not, used in Phrasebook index } bp_term_details;
- The structure bp_term_details is accessed in 3/bp and here.
bp_term_details BPTerms::new(TERM_DOMAIN_CALCULUS_TYPE *infs) { bp_term_details bptd; bptd.called_name = EMPTY_WORDING; bptd.function_of_other = NULL; bptd.implies_infs = infs; bptd.implies_kind = NULL; bptd.index_term_as = NULL; return bptd; } bp_term_details BPTerms::new_kind(kind *K) { bp_term_details bptd; bptd.called_name = EMPTY_WORDING; bptd.function_of_other = NULL; bptd.implies_infs = NULL; bptd.implies_kind = K; bptd.index_term_as = NULL; return bptd; }
§3. And there is also a fuller version, including the inessentials:
bp_term_details BPTerms::new_full(TERM_DOMAIN_CALCULUS_TYPE *infs, kind *K, wording CW, i6_schema *f) { bp_term_details bptd = BPTerms::new(infs); bptd.implies_kind = K; bptd.called_name = CW; bptd.function_of_other = f; return bptd; }
§4. In a few cases BPs need to be created before the relevant domains are known, so that we must fill them in later, using the following:
void BPTerms::set_domain(bp_term_details *bptd, kind *K) { if (bptd == NULL) internal_error("no BPTD"); bptd->implies_kind = K; bptd->implies_infs = TERM_DOMAIN_FROM_KIND_FUNCTION(K); }
§5. Some BPs are such that \(B(x, y)\) can be true for more or less any combination of \(x\) and \(y\). Those can take a lot of storage and it is difficult to perform any reasoning about them, because knowing that \(B(x, y)\) is true doesn't give you any information about \(B(x, z)\). For instance, the BP created by
Suspicion relates various people to various people.
is stored at run-time in a bitmap of \(P^2\) bits, where \(P\) is the number of people, and searching it ("if anyone suspects Harry") requires exhaustive loops, which incur some speed overhead as well.
But other BPs have special properties restricting the circumstances in which they are true, and in those cases we want to capitalise on that. "Contains" is an example of this. A single thing \(y\) can be (directly) inside only one other thing \(x\) at a time, so that if we know \(C(x, y)\) and \(C(w, y)\) then we can deduce that \(x=w\). We write this common value as \(f_0(y)\), the only possible value for term 0 given that term 1 is \(y\). Another way to say this is that the only possible pairs making \(C\) true have the form \(C(f_0(y), y)\).
And similarly for term 1. If we write \(T\) for the "on top of" relation then it turns out that there is a function \(f_1\) such that the only cases where \(T\) is true have the form \(T(x, f_1(x))\). Here \(f_1(x)\) is the thing which directly supports \(x\).
Containment has an \(f_0\) but not an \(f_1\) function; "on top of" has an \(f_1\) but not an \(f_0\). Many BPs (like "suspicion" above) have neither.
Note that if \(B\) does have an \(f_0\) function then its reversal \(R\) has an identical \(f_1\) function, and vice versa.
§6. We never in fact need to calculate the value of \(f_0(y)\) from \(y\) during compilation — only at run-time. So we store the function \(f_0(y)\) in an i6_schema for the necessary run-time code. For example, this might be the schema ContainerOf(*1), which would code-generate to a function call.
void BPTerms::set_function(bp_term_details *bptd, i6_schema *f) { if (bptd == NULL) internal_error("no BPTD"); bptd->function_of_other = f; } i6_schema *BPTerms::get_function(bp_term_details *bptd) { if (bptd == NULL) internal_error("no BPTD"); return bptd->function_of_other; }
kind *BPTerms::kind(bp_term_details *bptd) { if (bptd == NULL) return NULL; if (bptd->implies_kind) return bptd->implies_kind; return TERM_DOMAIN_TO_KIND_FUNCTION(bptd->implies_infs); }
§8. The table of relations in the index uses the textual name, so:
void BPTerms::index(OUTPUT_STREAM, bp_term_details *bptd) { if (bptd->index_term_as) { WRITE("%s", bptd->index_term_as); return; } wording W = EMPTY_WORDING; if (bptd->implies_infs) W = TERM_DOMAIN_WORDING_FUNCTION(bptd->implies_infs); if (Wordings::nonempty(W)) WRITE("%W", W); else WRITE("--"); }