Use NonNull pointer in heap version

[AngelicosPhosphoros]

Since SmallVec allocate on heap only if capacity is larger than inner capacity and therefore larger than 1, it cannot be null. Also, this is same behaviour as std::vec::Vec which always uses non-null pointer, which is dangling if it has zero capacity.

This change reduces size of enum in some cases because it can exploit niche optimization. For example, size_of::<SmallVec<[u8; 8]>> is 24 bytes with changes from this PR while before it was 32 bytes.

Also:

1. Changed some internal APIs to indicate that we work with non-nullable pointers.
2. Updated natvis and fixed bug with incorrect calculation of is_inline intrinsic.

[AngelicosPhosphoros]

Also, I think this changes make union feature useless because it would not provide size benefits anymore.

Another thing is capacity or size field in small vec implementation where field acts as length in inline version and as capacity in heap version. I don't see benefits of this compared to storing length directly and storing capacity of allocation in heap version.
On the other hand, storing capacity in heap version allow to store NonZeroUsize there and get size of Option<SmallVec<_>> smaller when used array is small.

Btw, it seems that CI is broken probably because latest version of Ubuntu doesn't have some packages used in CI.

[mbrubeck]

Hi, thanks for this work! Interestingly, I tried doing this before but reverted it in 0e8f5e9 because I couldn’t see any size improvements. I’m not sure if something was wrong with my implementation, or if I just made a mistake testing it.

The CI errors should be fixed by #299.

[AngelicosPhosphoros]

@mbrubeck
I fixed natvis for enum version.

And as I tested, natvis for union version didn't work in first place because windbg cannot read inline version for some reason. It writes "expected a member name" in watched variables, for example.

I also added is_inline to short debug representation because I think it is useful information.

[mbrubeck]

Thanks, this looks good! I'm going to give it one more careful review before merging it, in the next few days.

[mbrubeck]

@bors-servo r+

[bors-servo]

📌 Commit 9d96d7f has been approved by mbrubeck

[bors-servo]

⌛ Testing commit 9d96d7f with merge 0c56f1b...

[bors-servo]

☀️ Test successful - checks-github
Approved by: mbrubeck
Pushing 0c56f1b to master...

[bors-servo]

☀️ Test successful - checks-github
Approved by: mbrubeck
Pushing 0c56f1b to master...

[sffc]

I suspect that this PR increases the code size of SmallVec::try_reserve, which is used by a number of entrypoints including SmallVec::insert_from_slice. On smallvec 1.10, with link-time optimization, insert_from_slice gets inlined to the call site and only ever calls down to rust_alloc. However, on smallvec 1.11, the function is no longer inlined and it calls rust_alloc, rust_realloc, and rust_dealloc. This causes about 1 KB more code to be linked into WASM than was previously required when using smallvec. I'm not 100% sure this is the culprit PR but it seems to be the only one between 1.10 and 1.11 touching this code path.

CC @Manishearth @robertbastian