`walk-parts' walked the whole `var' form, and a `fn' type's parameter
list is shaped like a call: `(fn ((c int)) int)' came out as
`int (*fp)(c int)'. A closure literal had no type, so calling one in
place missed the rewrite; sextest fed sexc symbols Sex cannot read.
`&&', the bitwise operators, the shifts and `++' each stopped with
`cannot infer'; an array operand came back as the array; a rank tie
went to whichever operand came first; and a toplevel `_' reached the
writer unsolved.
`max_align_t' is C11 and the union goes into every unit, so a program
with no closure in it stopped building under -std=c99. `unify' also
bound a rigid variable one way round only.
The argument list was flattened with one separator throughout, so
`((long long))' and `((long) (long))' named one struct. A capture
borrowing a name looked only at the local scope chain.
A `while' or `switch' body is a block with no `do' around it, and its
declarations landed in the enclosing frame. Two operands of one narrow
type skipped the conversions, so `(+ c c)' answered `char'.
fmt-c takes a parameter's name with `cadr', so an unnamed one handed
over bare lost its second word to it: `(* const char)' dropped its
star, and `(int)' had no second word at all.
An arglist entry, an array's bound and its element type ask one
question, and disagreed: `(unsigned int)' was a name plus a type, a
trailing typedef a bound, a subscript the type's second word.
Two things out of one mechanism. `_' as a type means "work it out from
the initializer", so (var n _ (strlen s)) stops needing size-t spelled
out; `type-of' hands a macro the type of an expression, so a macro can
dispatch on what it was handed rather than on what was declared. Both
read the same answers from two sides.
Algorithm W's core, intra-procedural, with the extensions C forces:
- an unknown type, since (include stdio.h) brings in names we never
parsed. Unification is consistency rather than equality, so
anything touching an unparsed declaration stops constraining
instead of rejecting a program that compiled yesterday;
- the usual arithmetic conversions, since `+' is not a function of
one type;
- checking mode for initializers, since #(0 0) has no type of its own
and takes one from its context. #(T : ...) is the way out of that.
What it wanted on the way:
- what type a *name* has, which neither the typedef nor the tag
database recorded. One table serves functions and variables, since
a function type already has a surface spelling;
- a scope chain, so a (var c int 9) inside a do ends with the block;
- form-type, keyed by cons cell, so one form has one type;
- macros expanded during the walk rather than before it, so type-of
is answered in the scope the macro was written in.
Closures take the same machinery: a receiver whose type comes from a
call, captures written (name expr) and typed from the expression, and
conversion from a bare function wherever a closure is expected.
type-match grew `_' on the pattern side, since (closure ((int)) int)
and (closure ((float)) int) were separate clauses for one case.
A bare list was read as several forms, so a macro returning
`((make-adder 10) 5)' -- a call of what make-adder returns -- was
spliced into two. Add a new $ char to denote splicing, so ($ form ...)
splices.
