Chez fixnums are 61-bit, so the bignum bitwise-and mask allocates for any
value past 2^60 — and unchecked-* ran it on every result, even small
in-range ones. An exact integer already in [-2^63, 2^63) is its own wrap,
so return it directly; only an out-of-range result (a multiply overflowing
into 128 bits) needs the mask. ~30% on in-range unchecked-add loops,
neutral on full-64-bit multiply.
Note: the 64-bit arithmetic floor on Chez stays ~31ns/multiply (bignum, no
native 64-bit int); the test.check distribution hangs are dominated by
generator/dispatch overhead, not arithmetic — this is a general win for
long-heavy code, not a fix for those.
a49ca3b5ea