Optimize copy.

[sunfishcode]

Optimize wstd::io::copy using the wasi-io splice function. To do this, add a way to query AsyncRead and AsyncWrite implementations for an underlying AsyncInputStream and AsyncOutputStream.

This is similar to optimizations done in std::io::copy, however those optimizations make use of specialization, which isn't available on stable Rust. That said, using explicit interfaces instead of specialization does have the advantage that user-defined types can opt into optimization if they're able to.

This also eliminates the RefCells in TcpStream.

[sunfishcode]

The tcp_echo_server test is hanging on what looks like a bug in the tcp_echo_server program, which is currently depending on what appears to be a Wasmtime bug.

The current unoptimized copy code calls AsyncInputStream::read which does ready() followed by read(), and in the second iteration the ready() and read() returns 0 bytes, which appears to violate the "The pollable given by subscribe will be ready when more bytes are available." guarantee in the spec.

If that Wasmtime bug is fixed, then the tcp_echo_server test will start hanging. That looks like it should be fixed by calling shutdown on the write half after doing the write_all, which will then cause the server's copy to complete and the client's subsequent read_to_end to complete. That change would also fix the failure here.

[sunfishcode]

This can be reproduced on trunk by adding this assert, checking the comment just above:

diff --git a/src/io/streams.rs b/src/io/streams.rs
index 63c084d..7173398 100644
--- a/src/io/streams.rs
+++ b/src/io/streams.rs
@@ -44,7 +44,10 @@ impl AsyncRead for AsyncInputStream {
             // WASI's `read` doesn't mean end-of-stream as it does in Rust,
             // however since we called `self.ready()`, we'll always get at
             // least one byte.
-            Ok(r) => r,
+            Ok(r) => {
+                assert!(!r.is_empty());
+                r
+            }
             // 0 bytes from Rust's `read` means end-of-stream.
             Err(StreamError::Closed) => return Ok(0),
             Err(StreamError::LastOperationFailed(err)) => {

The assert fails, even though the test still passes.

[sunfishcode]

I've now reported the Wasmtime bug upstream in bytecodealliance/wasmtime#9938, and fixed the test to avoid depending on the Wasmtime bug.

[sunfishcode]

With the test fixed, the CI here now passes.