Olivier Galibert writes:
Hrvoje already replied[1], but as a generic rule of thumb I'd be
happier if we could avoid per-lrecord flags entirely. They don't cope
well with Jan[2]'s very nice idea on how to avoid the lrecord headers,
which gives us faster allocations[3], sharing[4], and a 33% size gain
on conses[5][6].
I don't remember this idea.