On Jun 18, 2010, at 9:52 PM, Aaron Cohen wrote:
I've also temporarily enabled a diagnostic (in both) that tells you
when you have a mismatch between a loop initializer and its recur
form. It goes off over a hundred times in Clojure itself, when using
the arbitrary precision default. In each case, the recur value is
needlessly being boxed, every iteration of a loop, for loops that
will
never be bigints; indexes, counters etc. I expect this will be very
typical of most code. But removing that useless overhead would be a
lot of tedious work.
With the defaults swapped, only 2 warnings.
How often does the warning go off if you go back to
arbitrary-precision by default, but make literals return boxed numbers
again? I suspect your comparison is unfair because lots of the loop
initializers are probably literals.
It would be easy to make it go away. The point is not the mismatches
themselves, but they serve as a counter - how many loops had primitive
init and needed only primitive arithmetic? Each warning is a case
where it could be automatically fast but isn't. Making the warning go
away will leave the inefficiency but make it silent. It is just a
measurable loss to counter all the theoretical arguments being made.
---
A concern I have is that interop forms that return primitives can
cause a loop var to have a primitive type.
(loop [procs (.availableProcessors Runtime/getRuntime)] ; Current
clojure makes procs a primitive, do people really expect that?
Would it be possible for loop to use boxed numbers always by default,
and only use primitives if some specific metadata is on the
initializer?
So in that world:
(loop [i 5] ; Boxing is always done by the loop statement, even if
initialized with a primitive
(loop [procs (.availableProcessors Runtime/getRuntime)] ; Boxing is
done by the loop statement
(loop [^:primitive procs (.availableProcessors Runtime/getRuntime)] ;
Primitive is used, good because we are expecting it
I'm not sure which order the metadata should go in such a scheme.
(loop [^:primitive i 5]
(loop [i ^:primitive 5]
(loop [i (prim 5)]
Yes, it's easy to imagine a world where people who want efficient code
have to jump through hoops to get it. OTOH, you can just say (num some-
expr) to force it to be boxed, if you want assurance of an Object
initializer. Which will be the more common need?
I have to say I'm in the 'pay for what you use' camp - you need a box,
you ask for one. If I don't (and neither do any of those loops), why
should I have to do extra work to avoid it?
Rich
--
You received this message because you are subscribed to the Google
Groups "Clojure" group.
To post to this group, send email to clojure@googlegroups.com
Note that posts from new members are moderated - please be patient with your
first post.
To unsubscribe from this group, send email to
clojure+unsubscr...@googlegroups.com
For more options, visit this group at
http://groups.google.com/group/clojure?hl=en
