Optimize condition for removing written data in BufWriter (#415)

* Optimize condition for removing written data in BufWriter

Do it strategically, when:
 - it can use a memcpy instead of memmove to remove it
 - when no space left for new data

Signed-off-by: Jiahao XU <[email protected]>

* Fix typo

Signed-off-by: Jiahao XU <[email protected]>

* Replace multiplication with division

To prevent overflo

Signed-off-by: Jiahao XU <[email protected]>

* Update crates/async-compression/src/generic/write/buf_writer.rs

Signed-off-by: Jiahao XU <[email protected]>

* Do copyback at 512B flushed data

So that memmove AVX256 can be used to do the copyback,
which should be reasonably fast

Signed-off-by: Jiahao XU <[email protected]>

* Fix commit issues

Signed-off-by: Jiahao XU <[email protected]>

* Fix fmt

Signed-off-by: Jiahao XU <[email protected]>

---------

Signed-off-by: Jiahao XU <[email protected]>

Author: NobodyXu

crates/async-compression/src/generic/write/buf_writer.rs

@@ -136,7 +136,13 @@ impl BufWriter {
     ) -> Poll<io::Result<&mut [u8]>> {
         ready!(self.partial_flush_buf(poll_write))?;
 
-        if self.written > 0 {
+        // when the flushed data is larger than or equal to half of yet-to-be-flushed data,
+        // the copyback could use version of memcpy that do copies from the head of the buffer.
+        // Anything smaller than that, an overlap would happen that forces use of memmove.
+        if self.written >= (self.buffered / 3)
+            || self.written >= 512
+            || self.buffered == self.buf.len()
+        {
             self.remove_written();
         }