Fix memory leaks in doctests?

[RalfJung]

Currently a bunch of our doctests have memory leaks. For that reason, we have to disable Miri's leak checker when running doctests.

To reproduce:

MIRIFLAGS="-Zmiri-disable-isolation" ./x.py miri library/core library/alloc --doc
MIRIFLAGS="-Zmiri-disable-isolation" ./x.py miri library/std --doc -- --skip fs:: --skip net:: --skip process:: --skip sys::pal::

@rust-lang/libs how do you feel about that -- is it okay for the tests to leak, or should we fix that (by adding hidden code that deallocates things again)? Currently we don't have a good way to disable the leak checker on a per-test basis so there is a risk of accidental leaks.

[RalfJung]

One (I think somewhat typical) example of a leaky test is this: a Vec is created but remains in a MaybeUninit so it is never dropped.

[ChrisDenton]

I've nominated for libs-api to ensure this gets discussed but feel free to un-nominate if a decision is made before a meeting.

[the8472]

One (I think somewhat typical) example of a leaky test is this: a Vec is created but remains in a MaybeUninit so it is never dropped.

Was that meant as an example of one that takes extra steps to unleak it? Since it calls assume_init which makes the vec droppable again.

The Vec::leak doctest is an example that leaks.

[RalfJung]

Ah, I picked the wrong test indeed, sorry. Here is an example that leaks.

[Amanieu]

We discussed this in the libs meeting today. We would prefer to have miri check for leaks in doctests and handle the currently leaking tests by either:

• Fixing the leaks with hidden code.
• Adding a doctest attribute such as no_leak_check which disables leak checking just for that test.

[RalfJung]

Adding a doctest attribute such as no_leak_check which disables leak checking just for that test.

That could be done if rustdoc allowed passing flags to Miri... but the could be only for Miri, not rustc, so not sure how to best do this. Alternatively, rust-lang/miri#3670 would help.