libera/#sbcl - IRC Chatlog

16:09:22 edgar-rft_ ** NICK edgar-rft

20:36:48 stassats` can type checks be removed if they're not used?

20:37:04 stassats` technically, no type checks have to be performed at all

20:38:20 stassats` i actually can't tell if it's "not used" for any given type check

20:41:59 stassats` i wanted to fold (if (> x 10) (print 10) (print 10)) to (print 10), but then > checks X for NUMBER

20:42:29 stassats` which does happen if it's folded, but i wanted to delay until vop selection, which happens after typecheck generation

21:02:21 stassats` all these wild expensive schemes do not produce that much of a payoff

21:02:31 stassats` it all adds up in the end, but the expenses also add up

21:21:47 mrcom (if (foo x) (print 10) (print 10)) -- obvious can't elide

21:22:03 stassats` it'll be (foo x) (print 10)

21:22:11 mrcom Yeah.

21:22:16 stassats` which does happen right now, but not for print

21:22:42 stassats` for (if (foo x) 10 10)

21:22:54 mrcom Are you allowed to say "this standard function has no side-effects not documented in the standard".

21:23:10 stassats` and (if (> x 20) 10 10) does check X

21:23:25 stassats` but i want to change the order of when that happens, hence trouble

21:25:23 mrcom Standard says < "_might_ signal type-error if some argument not a real."

21:25:33 mrcom So you'd be legal to omit it.

21:25:45 mrcom Dunno if it would be _smart_.

21:26:23 mrcom Maybe if speed 3 safety 0?

21:26:46 stassats` type checks also have side effects

21:31:26 mrcom As a user, I'd be ok with you optimizing it out in high-optimization mode. If I'm concerned about the compiler letting something slide I'll manually check it anyway.

21:32:43 mrcom Whole point of safety 0 is "assume the user code isn't broken".

21:33:14 stassats` safety 0 is my enemy

21:33:51 mrcom Really? I'd think that's where you can go wild.

21:34:15 stassats` or wildly inaccurate

21:34:36 mrcom Although I guess the chain of effects from a violated assumption can be ugly to debug.

21:40:23 mrcom Another thought from the peanut gallery--it would actually be more useful to me to be told "you've got technically useless code here" than to do the defacto type checking.

21:40:41 mrcom I think.

21:40:43 stassats` there's "deleting unreachable code"

21:41:29 mrcom I can't tell you how many times that's pointed out a bug in my code.

21:42:52 mrcom That and "I can't figure out if your function return actually matches the declaration"

21:43:30 stassats` that would be too noisy

21:43:43 mrcom You already do that.

21:44:07 stassats` i don't think so

21:44:19 stassats` maybe you're thinking of some other note

21:49:10 mrcom "complex type specifier too complicated"? maybe a different one Something about "to check" or words to that effect.

21:49:28 stassats` yeah, that's not related

21:49:51 stassats` that's about the number of values

21:50:14 mrcom "Type assertion too complex to check efficiently:~

21:52:17 mrcom Although it's almost always the set of possible return values, not the assertion, that's messed up. So maybe not that one either.

21:57:35 mrcom Anyway, I almost always run speed 1 safety 3 and fix *any* notes, so skipping checks in higher optimzation settings is pretty safe. For me.

21:57:59 mrcom Survey of 1.

22:15:56 mrcom On the other hand, if a customer told me "Don't bother checking this input. It takes too long." I'd be inclined to stare at them for a while, then tell them "No."