perf(server): cache renderings of the Date header

This is actually one of the biggest impacts to benchmark performances at
this point. Caching the rendering of the Date header improves "hello
world" benchmarks by around 10%.

src/http/h1/date.rs

@@ -0,0 +1,59 @@
+use std::cell::RefCell;
+use std::fmt::{self, Write};
+use std::str;
+
+use time::{self, Duration};
+
+// "Sun, 06 Nov 1994 08:49:37 GMT".len()
+pub const DATE_VALUE_LENGTH: usize = 29;
+
+pub fn extend(dst: &mut Vec<u8>) {
+    CACHED.with(|cache| {
+        let mut cache = cache.borrow_mut();
+        let now = time::get_time();
+        if now > cache.next_update {
+            cache.update(now);
+        }
+        dst.extend_from_slice(cache.buffer());
+    })
+}
+
+struct CachedDate {
+    bytes: [u8; DATE_VALUE_LENGTH],
+    pos: usize,
+    next_update: time::Timespec,
+}
+
+thread_local!(static CACHED: RefCell<CachedDate> = RefCell::new(CachedDate {
+    bytes: [0; DATE_VALUE_LENGTH],
+    pos: 0,
+    next_update: time::Timespec::new(0, 0),
+}));
+
+impl CachedDate {
+    fn buffer(&self) -> &[u8] {
+        &self.bytes[..]
+    }
+
+    fn update(&mut self, now: time::Timespec) {
+        self.pos = 0;
+        write!(self, "{}", time::at_utc(now).rfc822()).unwrap();
+        assert!(self.pos == DATE_VALUE_LENGTH);
+        self.next_update = now + Duration::seconds(1);
+        self.next_update.nsec = 0;
+    }
+}
+
+impl fmt::Write for CachedDate {
+    fn write_str(&mut self, s: &str) -> fmt::Result {
+        let len = s.len();
+        self.bytes[self.pos..self.pos + len].copy_from_slice(s.as_bytes());
+        self.pos += len;
+        Ok(())
+    }
+}
+
+#[test]
+fn test_date_len() {
+    assert_eq!(DATE_VALUE_LENGTH, "Sun, 06 Nov 1994 08:49:37 GMT".len());
+}