performance-noexcept-swap

[Jacobfaib]

Description

closes

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[fbusato]

this improvement makes sense at theoretical level. I'm a bit worried about edge cases support with nvcc/nvc++/nvrtc. I remember that I had hard time with noexcept and compiler compatibility. @miscco should take a look at it

[Jacobfaib]

I remember that I had hard time with noexcept and compiler compatibility. @miscco should take a look at it

Yes I am struggling at the moment with a particular edge case of

struct IncompleteType;

cuda::std::tuple<IncompleteType&> t;

This is causing compilation errors because the noexcept specification is being eagerly computed, via is_nothrow_swappable_v<IncompleteType&>. However, computing is_nothrow_swappable with an incomplete type is undefined behavior according to the standard.

I guess I might have to resort to the usual method of deferring this via some dummy argument:

template <typename DummyT = char>
void swap(tuple& tup, DummyT = DummyT{}) noexcept(noexcept(...))

But it seems a little icky to do this on the public interface. I also don't understand how libc++ is getting around this, their implementation seems quite similar to ours but they don't have the error.

[miscco]

However, computing is_nothrow_swappable with an incomplete type is undefined behavior according to the standard.

I thought we were trying hard to make that work, will look at this on a friday

[Jacobfaib]

I thought we were trying hard to make that work, will look at this on a friday

The root problem is that older nvcc (or perhaps it's GCC) don't even allow you to write sizeof(T) if T is incomplete. See https://godbolt.org/z/9P64PE9s8 for an example.

[bernhardmgruber]

Thrust changes are fine

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 3h 00m: Pass: 100%/162 | Total: 3d 18h | Max: 2h 59m | Hits: 83%/401698

See results here.