It would be nicer to keep the "interrupts" property and
add a property "interrupt-parents" with the same number
of entries, that encodes the same as "interrupt-parent"
but per interrupt.  You keep more in line with the
"normal" stuff and I suspect it's less code to parse as
well.

There are pro and cons to this approach. I did think about it initially.
The main cons is that "interrupts" becomes harder to parse as you ahve
to walk interrupt-parents at the same time to get the #interrupt-cells
of each entry. Also, it adds more potential for stupid breakage (how
shluld the parser react if interrupt-parents has less entries than
interrupts ?)

The pro, which is quite important too, is that it's a common assuption
that you have interrupts when you have an "interrupts" property, period.
In fact, it would indeed fit a bit better in the current parser.

Wrong place to document this?  It's true for all interrupt
specifiers.

Might be worth giving a crazy example where the 2 interrupt specifiers
have different size.

Ben.