Implement Default for all array sizes

[shamatar]

I've made a simple Godbolt for a demonstration of the approach.

In short, with array::map stabilized in 1.55 it's now possible to make an array [T; N] of default values of T by invoking [(); N].map(|_| T::default()). There is no difference in the emitted assembly for either O2 or O3 optimization levels in my short example for usize and user defined type (both ZST and not ZST), but obviously there is no guarantee that some more complex cases would not diverge.

I would expect that lifting a hard limitation of 32 on the size of the array would be handy in general.

[hellow554]

There's an open issue for this... somewhere, I can't find it.
IIRC when implemented via const generics there was a regression with a zero sized array. Maybe somebody can link it here...

[shamatar]

I've updated the Godbolt link to include creation of arrays of length 0 for both ZST and not ZST. It doesn't look like there is any difference as array::map already has a special case for N == 0 that is optimized out by the compiler in both cases.

[paolobarbolini]

I've updated the Godbolt link to include creation of arrays of length 0 for both ZST and not ZST. It doesn't look like there is any difference as array::map already has a special case for N == 0 that is optimized out by the compiler in both cases.

The issue is that Default must also be implemented for zero sized arrays where T doesn't implement Default.

[cuviper]

There's some discussion in #74060 (comment) and the references from there.

[shamatar]

Ok, I see the point. It lead me to the question now posted in zullip, may be it would be a nice feature that may go along and parallel with specialization (and potentially much easier to implement)

[Gankra]

Summary

Before we had const-generics we had some macros manually implementing Default for a few different array sizes, and we were too clever and special-cased [T; 0] to be unconditionally Default. As a result when const-generics were introduced, providing a const-generic impl of Default for arrays would have been a regression because it would introduce a T: Default bound to the empty array case. This regression may be acceptable but was rejected because:

• It might break things (but seemingly no one actually did a crater run, citing the existence of non-public code)
• Compiler devs believed they could/should "solve" the issue and so it would be needless churn to "break" it just to fix it again later

This all happened about 2 years ago, and there is currently no clear roadmap to resolving the issue either way.

History

• Remove trait LengthAtMost32 #74060 introduced const-generic impls for arrays but skipped on Default for the reasons above
• Default for arrays via const generics #74254 tried to solve this problem with some combination of specialization, but was rejected because specialization is a bit of a mess no one wants to make part of our public API/guarantees.
• [perf] implement Default using const generics #74326 further tested that approach
• implement Default for all arrays #84838 attempted a more targeted approach with lang items, but the libs team rejected it on the grounds that it was going to be cursed/unmaintainable code.

[ryoqun]

There's an open issue for this... somewhere, I can't find it. IIRC when implemented via const generics there was a regression with a zero sized array. Maybe somebody can link it here...

@hellow554 I think #61415 is the correct issue.

@Gankra hey, it's nice summary. could you update to include a pointer to the issue?

@shamatar also, now it's directed to the correct issue. this one can be closed, maybe?

[dtolnay]

Yes, I'll close this in favor of #61415.