i8/i16 values with limited operators

[AndrewScheidecker]

previous discussion: #85

Looking at the atomics proposal, there are a lot of operators that are only necessary because WASM doesn't have i8 and i16 value types. e.g. here are the operators for atomic adds:

• i32.atomic.rmw8_s.add
• i32.atomic.rmw8_u.add
• i32.atomic.rmw16_s.add
• i32.atomic.rmw16_u.add
• i32.atomic.rmw.add
• i64.atomic.rmw8_s.add
• i64.atomic.rmw8_u.add
• i64.atomic.rmw16_s.add
• i64.atomic.rmw16_u.add
• i64.atomic.rmw32_s.add
• i64.atomic.rmw32_u.add
• i64.atomic.rmw.add

If there were i8 and i16 value types, those could be simplified to:

• i8.atomic.rmw.add
• i16.atomic.rmw.add
• i32.atomic.rmw.add
• i64.atomic.rmw.add

The same applies to the atomic xchg, cmpxchg, load, sub, and, or, and xor operators.

Would folks consider adding i8 and i16 types that have a limited set of conversion operators meant for composition with the memory loads/stores? e.g. for i8:

• i8.wrap/i32 / i8.wrap/i64
• i32.extend_s/i8 / i64.extend_s/i8
• i32.extend_u/i8 / i64.extend_u/i8

[jfbastien]

I think this is worth discussing as part of the threads discussion in the upcoming performance meeting. @binji could you add it to the threads proposal and make sure it's one of the items discussed?

[binji]

Sounds good.

[binji]

We discussed this in the CG meeting and decided not to implement i8/i16 value types at this time.

On a related point, we decided to pursue new sign-extending operators along with the threading proposal:

• i32.extend_s/i8
• i32.extend_s/i16
• i64.extend_s/i8
• i64.extend_s/i16

See https://github.com/WebAssembly/threads/blob/master/proposals/threads/Overview.md#new-sign-extending-operators.

[rossberg]

On 22 June 2017 at 01:49, Ben Smith ***@***.***> wrote: We discussed this in the CG meeting and decided not to implement i8/i16 value types at this time. On a related point, we decided to pursue new sign-extending operators along with the threading proposal: - i32.extend_s/i8 - i32.extend_s/i16 - i64.extend_s/i8 - i64.extend_s/i16 See https://github.com/WebAssembly/threads/blob/master/proposals/threads/ Overview.md#new-sign-extending-operators.

Oh, I hadn't noticed these. I think the naming scheme is a bit off, since the parts after the "/" are always proper value types denoting the operand type elsewhere. It would be preferable to mirror the naming scheme used for loads: * i32.extend8_s * i32.extend16_s * i64.extend8_s * i64.extend16_s That would also allow you to include i64.extend32_s, which is currently missing and not expressible in the above scheme (since it is ambiguous with the existing operator).

[binji]

Good point; opened WebAssembly/threads#34 to discuss.