Using Zip & Take allows accessing the middle a Map iterator without “getting there first” from either side.

[steffahn]

Look at this iterator.

let iterator = (0..100).map(|x| println!("item {}", x));

We all know that when map’s resulting Iterator implements DoubleEndedIterator, then the side-effects can be re-ordered. However, judging by the implementation of map one would probably at least expect that it’s impossible to hit the "item 50" without first either going through at least 0, 1, 2, … 49 or 100, 99, 98, … 51 first. Right?

Well, here we go!

let mut y = iterator.take(51).zip(0..100);
y.next_back();

on beta

item 50

on stable 1.52.0

item 99
item 98
item 97
item 96
item 95
item 94
item 93
item 92
item 91
item 90
item 89
item 88
item 87
item 86
item 85
item 84
item 83
item 82
item 81
item 80
item 79
item 78
item 77
item 76
item 75
item 74
item 73
item 72
item 71
item 70
item 69
item 68
item 67
item 66
item 65
item 64
item 63
item 62
item 61
item 60
item 59
item 58
item 57
item 56
item 55
item 54
item 53
item 52
item 51
item 50

(playground)

@rustbot label T-libs-impl, A-iterators, A-specialization, regression-from-stable-to-beta

[steffahn]

cc @the8472

[the8472]

As mentioned in the other issue, I think this is just another instance of #82303

Edit: Nevermind, not really since it doesn't involve calling __iterator_get_unchecked on Zip, rather this is about use of __iterator_get_unchecked in Zip.

[steffahn]

It’s related to #82303. I guess the “random” aspect of TrustedRandomAccess is a bit weird for iterators like Map.

You can do the same thing (accessing the map from the center without “walking” there) with Zip only, too, on stable already like so:

let mut y = iterator.zip(0..51).zip(0..100);
y.next_back();

---

This issue is only “really” a regression if out-of-order side-effects for iterators aren’t wanted.

[the8472]

steffahn brought up some general concerns about the TrustedRandomAccess implemention on Take (#83990) in #85873 (comment)
I'll have to think some more about whether combining Take, TRA and backwards iteration is even doing the right thing under all circumstances, beyond the side-effects issue raised here.

I am generally in favor of eliminating iterator side-effects that are not necessary to produce the final results. But of course that's currently undocumented and possibly surprising behavior and needs to be approved by the libs team, which hasn't happened so far. So while I like what it does it shouldn't have landed without discussion.

You can probably do the same thing (accessing the map from the center without “walking” there) with Zip only, too, on stable already.

I don't think so? At least that would be another implementation gap. That's what the whole MAY_HAVE_SIDE_EFFECT thing is supposed to prevent.

let mut y = iterator.zip(0..51).zip(0..100);
y.next_back();

Oh, sure, that definitely is #82303

[the8472]

https://rust-lang.zulipchat.com/#narrow/stream/219381-t-libs/topic/eliminating.20side-effects.20of.20iterator.20adapters/near/241436500

We had written a fair number of specializations over time that were later either removed or just not accepted because they would fail to consider side-effects.

I'll submit a revert PR.