← board

Rebinding an int parameter twice across the float boundary is refused

Repro

def two(x: int):
    x = x + 1       # int + int can overflow -> a PROMOTABLE int (kind 28)
    x /= 2          # ...and true division wants a double (kind 19)
    return x
print(two(5))       # CPython 3.0



pascal26:3: error: Nil Python: annotate the type / too dynamic [a=28 b=19]

Either rebind ALONE is fine: x = x + 1 alone returns 6, and x /= 2 alone returns 2.5 (as of the sibling fix). It is the pair that has no join.

Where it comes from

x = x + 1 on an int notes tyPromoInt64 — deliberately, so an accumulator stays exact past 2^63 — and /= then notes tyDouble. The widening join has no answer for that pair and reports "too dynamic" rather than picking one.

The obvious answer is variant, which is what PyWidenBinding already gives for an ordinary int-vs-float rebind and what a promo value boxes into anyway (VT_PROMO_INT64 carries the exact decimal). Whether the join belongs in PyWiden, in PyWidenBinding, or at the promo-specific site is the thing to measure — TypeIsPromoInt is deliberately excluded from PyWidenBinding's numeric arm today, and that exclusion is what routes this pair to the error. Read that exclusion's reason before changing it.

Gate

two(5) prints 3.0, plus the same shape on a plain LOCAL (which must keep whatever it does today) and on an accumulator that genuinely needs the promo width (x = x + 1 in a loop past 2^63, no float in sight — must stay exact).

Resolution — the binding join was missing from the binding site

Fix: one line. PyNoteLocalType (pyparser.inc) joins with PyWidenBinding instead of PyWiden, plus the promo arm that PyWidenBinding needed to answer the pair.

PyWidenBinding's own doc-comment states the concept: "the join for a NAME REBOUND across types — which is NOT the join for an EXPRESSION, and that distinction is the whole fix." It was wired into the module-local site (27264), the field site (32869), the rebound-parameter site (24384) — and not into PyNoteLocalType, the routine whose first line reads "record one assignment's RHS type against a local". So the per-def locals table asked the expression join what a rebinding meant, and got PyWiden's honest "promo meets float has no scalar lowering; error honestly (annotate)".

Two joins for one concept, with the specific one absent from the most specific site — normalise-dont-special-case in its usual costume.

PyWidenBinding then needed the promo pair itself:

if (TypeIsPromoInt(a) and TypeIsFloat(b)) or
   (TypeIsPromoInt(b) and TypeIsFloat(a)) then
  Result := tyVariant

A variant loses nothing here: VT_PROMO_INT64 is one of a variant's own tags and carries the exact decimal past 2^63, so the slot becomes dynamic without becoming lossy. The pre-existing comment calling this pair "a deliberate honest error rather than a silent boxing" was right about the honesty and wrong about the boxing.

Two things this ticket asserted that measurement contradicted

Both are recorded because both cost a build cycle.

  1. "TypeIsPromoInt is deliberately excluded from PyWidenBinding's numeric arm today, and that exclusion is what routes this pair to the error." The exclusion existed — two not TypeIsPromoInt(...) clauses — and was dead code. PyNumeric lists the five machine numeric kinds and does not include tyPromoInt64, so PyNumeric(a) already implied not TypeIsPromoInt(a). Deleting both clauses built clean and changed nothing: the repro failed with the byte-identical message. The refusal never came from there. Reading the exclusion's git log -L reason (c787485a2) was still the right move; believing the ticket about which line fired was not.

  2. "the same shape on a plain LOCAL (which must keep whatever it does today)" — a plain local was refused identically:

    def local_shape():
    y = 5
    y = y + 1
    y /= 2
    return y        # same error, same [a=28 b=19]

    That is the measurement that located the site. The bug was never in the parameter machinery; the parameter was just where it was first seen. What does have to keep today's behaviour is the neighbouring int-meets-float local (z = 5; print(z); z = 3.14), which must still print 5 and not 5.0 — verified, row 5 of the test.

Gate — met verbatim, plus the siblings

Test: test/test_nilpy_rebind_across_the_float_boundary.npy (+.expected, registered in the Makefile) — six rows: the parameter, the local, the field, the exact accumulator, each-binding-renders-as-itself, and floor-div keeping intness.

Log