Vectored IO (like std and futures-preview)

[blckngm]

I was reading rust 1.36 announcement and noticed that vectored IO is now stablized in std:

• IoSlice
• IoSliceMut
• Read::read_vectored
• Write::write_vectored

Right now tokio seem to use vectored IO when implementing poll_read_buf/poll_write_buf, but there are no straightforward public API for vectored IO.

I think tokio should add direct support for vectored IO, just like futures-io-preview and futures-util-preview:

• Add a poll_read_vectored method to AsyncRead,
• Add a poll_write_vectored method to AsyncWrite,
• Add a read_vectored method to AsyncReadExt,
• Add a write_vectored method to AsyncWriteExt.

(Using a similar API with futures-io also eases future transition to use futures-io directly or providing a compatibility layer with futures-io without performance penalty.)

I don't know whether poll_read_buf and poll_write_buf should be kept in AsyncRead/AsyncWrite. Are there more possible optimizations for them besides using vectored IO?

Right now bytes and mio use iovec::IoVec, but they are transitioning to IoSlice/IoSliceMut in the next version: #263, #957. Meanwhile, we can transmute stuff before these dependencies are updated, because IoVec and IoSlice/IoSliceMut are ABI compatible (because they are both ABI compatible with WSABUF or libc::iovec).