lang::typecheck::whnf
lang/src/typecheck/whnf.mo
pub def whnf (scope : Scope) (locals : LocalScope) (t : Term) : TermReduce t to weak-head normal form: keep reducing at the HEAD until
the head is rigid (a pi, a sort, a literal, a constructor, a
variable with no body, ...). Sub-terms are left alone -- the head is
all a structural comparison looks at before recursing, and each
recursive step goes back through unify, which reduces again.
pub because the Stage 3 reducer pins
(proofs/src/checker/cubical_props.mo) observe the cubical
reductions directly, not only through type_check -- the
constant-family/variable-family distinction lives here and nowhere
else.
pub def normalize_path_endpoints (scope : Scope) (locals : LocalScope) (t : Term) : TermRewrite every path application at a literal endpoint INSIDE t, not only
at its head. whnf_path_endpoint_app fires on a top-level p i0/p i1,
but a result type can bury the application under the binders an unfolded
definition introduced: transp (fn i => Equiv A (p i)) (idEquiv A) whnf's
to Equiv A (p i1), whose Sigma spelling mentions p i1 in a Pi
codomain and a lambda body. Structural comparison then descends into
those binders and never gives whnf the node with the right locals, so
p i1 stayed stuck and eqToEquiv could not return Equiv A B.
LocalScope is NAME-keyed, so descending under a binder needs no de
Bruijn shifting: push the binder (when it has a spellable name) so a
shadowing occurrence resolves to the binder, and leave the outer entries
alone.
pub def as_cubical_type (t : Term) (scope : Scope) (locals : LocalScope) : TermClassify t as the Term.cubical it denotes when it is a SATURATED
application of a bound cubical primitive -- the pure half of
type_check_app's own try_type_check_cubical_call preamble, for a
term that arrives as a TYPE rather than as a call being checked.
It exists because path_parts_of (lang/typecheck/cubical.mo) is
syntactic: it recognizes Term.cubical. A path type SPELLED as a
library def -- Path A a b, whose body is Cubical.PathP (fn _ => A) a b -- unfolds under delta into PathP (fn _ => A) a b, an APPLICATION
SPINE of the pathp primitive's NAME, so before this every path rule
(the fn i => dispatch, the endpoint rewrite in
normalize_path_boundary, and unify's comparison of a Path-spelled
type against the PathP the rules echo back) missed it.
Same guards, and the same reasons, as the checked-call probe: a
sentinel free head, no LOCAL shadowing the name, a registered
primitive, and an arity EQUALITY (a partially applied imeet i is not
the cubical term).