Avoid large reallocations when freezing BytesMut (#592)

When we freeze a BytesMut, we turn it into a Vec, and then convert that
to a Bytes.  Currently, this happen using Vec::into_boxed_slice, which
reallocates to a slice of the same length as the Vev if the length and
the capacity are not equal.  This can pose a performance problem if the
Vec is large or if this happens many times in a loop.

Instead, let's compare the length and capacity, and if they're the same,
continue to handle this using into_boxed_slice.  Otherwise, since we
have a type of vtable which can handle a separate capacity, the shared
vtable, let's turn our Vec into that kind of Bytes.  While this does not
avoid allocation altogether, it performs a fixed size allocation and
avoids any need to memcpy.

src/bytes.rs

@@ -807,8 +807,36 @@ impl From<&'static str> for Bytes {
 
 impl From<Vec<u8>> for Bytes {
     fn from(vec: Vec<u8>) -> Bytes {
-        let slice = vec.into_boxed_slice();
-        slice.into()
+        let mut vec = vec;
+        let ptr = vec.as_mut_ptr();
+        let len = vec.len();
+        let cap = vec.capacity();
+
+        // Avoid an extra allocation if possible.
+        if len == cap {
+            return Bytes::from(vec.into_boxed_slice());
+        }
+
+        let shared = Box::new(Shared {
+            buf: ptr,
+            cap,
+            ref_cnt: AtomicUsize::new(1),
+        });
+        mem::forget(vec);
+
+        let shared = Box::into_raw(shared);
+        // The pointer should be aligned, so this assert should
+        // always succeed.
+        debug_assert!(
+            0 == (shared as usize & KIND_MASK),
+            "internal: Box<Shared> should have an aligned pointer",
+        );
+        Bytes {
+            ptr,
+            len,
+            data: AtomicPtr::new(shared as _),
+            vtable: &SHARED_VTABLE,
+        }
     }
 }