Only mark a closure dynamic-extent once every reader has been checked

[dg1sbg]

DETERMINE-CLOSURE-EXTENT sets the enclose's extent to :dynamic from inside the loop over the variable's readers, so the mutation happens as soon as the first reader turns out to be a DX call. A later reader that escapes bails out with RETURN-FROM, which does not undo it, and the closure stays marked :dynamic.

The backend then builds it on the stack — under Clasp, cc_stack_enclose over an alloca in the enclosing frame — and it escapes anyway. Using it after that frame is gone is a use-after-return.

Why it is intermittent

Which reader is visited first decides whether this fires, and that order is unspecified. Set/set.lisp selects the hashset representation, because the guard on :listset is #+(or) and therefore always false, so DOSET walks an EQ hash table and the order follows object addresses.

Compiling

(lambda (x)
  (let ((g (lambda (y) (+ x y))))
    (mapcar g '(1 2 3))
    g))

300 times with Clasp's native compiler stack-allocated the returned closure 154 times. MAPCAR carries the :dx-call attribute, and the RETURNI is the escape. Calling one of the resulting closures gives a hard fault — SEGMENTATION-VIOLATION on one GC variant, EXT:BUS-ERROR on the other.

The fix

Move the assignment after the loop, where the conclusion it records has actually been established.

A variable with no readers now qualifies as well, which is correct — a value that is never read cannot escape. Previously the loop body simply never ran, so the case was missed rather than decided.

Verification

Through Clasp, since that is the client I have:

• escaping closure: 154/300 stack-allocated before, 0/300 after
• non-escaping closure: 300/300 still stack-allocated, so the optimisation is preserved
• 100 escaped closures called after their frames were overwritten and a collection forced: all correct
• full regression suite on both GC variants and the ANSI conformance suite: no unexpected failures

Origin

Introduced in 3884e93, which fixed the related but distinct problem in #12 — that a closure reachable through an escaping sibling could be stack-allocated. That fix added the reader walk; this one corrects when the walk commits its result.

[dg1sbg]

Standalone and independently valuable; #35 builds on it but this needs no part of that. Clasp-side consumer, for context: clasp-developers/clasp#1821.