Merge branch 'devel'

Cargo.toml

@@ -7,9 +7,9 @@ authors = ["Alexey Galakhov"]
 license = "MIT"
 readme = "README.md"
 homepage = "https://github.com/snapview/tungstenite-rs"
-documentation = "https://docs.rs/tungstenite/0.2.0"
+documentation = "https://docs.rs/tungstenite/0.2.1"
 repository = "https://github.com/snapview/tungstenite-rs"
-version = "0.2.0"
+version = "0.2.1"
 
 [features]
 default = ["tls"]

src/protocol/frame/frame.rs

@@ -8,41 +8,9 @@ use std::result::Result as StdResult;
 use byteorder::{ByteOrder, ReadBytesExt, NetworkEndian};
 use bytes::BufMut;
 
-use rand;
-
 use error::{Error, Result};
 use super::coding::{OpCode, Control, Data, CloseCode};
-
-fn apply_mask(buf: &mut [u8], mask: &[u8; 4]) {
-    for (i, byte) in buf.iter_mut().enumerate() {
-        *byte ^= mask[i & 3];
-    }
-}
-
-/// Faster version of `apply_mask()` which operates on 4-byte blocks.
-///
-/// Safety: `buf` must be at least 4-bytes aligned.
-unsafe fn apply_mask_aligned32(buf: &mut [u8], mask: &[u8; 4]) {
-    debug_assert_eq!(buf.as_ptr() as usize % 4, 0);
-
-    let mask_u32 = transmute(*mask);
-
-    let mut ptr = buf.as_mut_ptr() as *mut u32;
-    for _ in 0..(buf.len() / 4) {
-        *ptr ^= mask_u32;
-        ptr = ptr.offset(1);
-    }
-
-    // Possible last block with less than 4 bytes.
-    let last_block_start = buf.len() & !3;
-    let last_block = &mut buf[last_block_start..];
-    apply_mask(last_block, mask);
-}
-
-#[inline]
-fn generate_mask() -> [u8; 4] {
-    rand::random()
-}
+use super::mask::{generate_mask, apply_mask};
 
 /// A struct representing the close command.
 #[derive(Debug, Clone)]
@@ -219,10 +187,7 @@ impl Frame {
     #[inline]
     pub fn remove_mask(&mut self) {
         self.mask.and_then(|mask| {
-            // Assumes Vec's backing memory is at least 4-bytes aligned.
-            unsafe {
-                Some(apply_mask_aligned32(&mut self.payload, &mask))
-            }
+            Some(apply_mask(&mut self.payload, &mask))
         });
         self.mask = None;
     }
@@ -471,10 +436,7 @@ impl Frame {
 
         if self.is_masked() {
             let mask = self.mask.take().unwrap();
-            // Assumes Vec's backing memory is at least 4-bytes aligned.
-            unsafe {
-                apply_mask_aligned32(&mut self.payload, &mask);
-            }
+            apply_mask(&mut self.payload, &mask);
             try!(w.write(&mask));
         }
 
@@ -528,24 +490,6 @@ mod tests {
     use super::super::coding::{OpCode, Data};
     use std::io::Cursor;
 
-    #[test]
-    fn test_apply_mask() {
-        let mask = [
-            0x6d, 0xb6, 0xb2, 0x80,
-        ];
-        let unmasked = vec![
-            0xf3, 0x00, 0x01, 0x02, 0x03, 0x80, 0x81, 0x82, 0xff, 0xfe, 0x00,
-        ];
-
-        let mut masked = unmasked.clone();
-        apply_mask(&mut masked, &mask);
-
-        let mut masked_aligned = unmasked.clone();
-        unsafe { apply_mask_aligned32(&mut masked_aligned, &mask) };
-
-        assert_eq!(masked, masked_aligned);
-    }
-
     #[test]
     fn parse() {
         let mut raw: Cursor<Vec<u8>> = Cursor::new(vec![

src/protocol/frame/mask.rs

@@ -0,0 +1,129 @@
+use std::cmp::min;
+use std::mem::uninitialized;
+use std::ptr::copy_nonoverlapping;
+use rand;
+
+/// Generate a random frame mask.
+#[inline]
+pub fn generate_mask() -> [u8; 4] {
+    rand::random()
+}
+
+/// Mask/unmask a frame.
+#[inline]
+pub fn apply_mask(buf: &mut [u8], mask: &[u8; 4]) {
+    apply_mask_fast32(buf, mask)
+}
+
+/// A safe unoptimized mask application.
+#[inline]
+#[allow(dead_code)]
+fn apply_mask_fallback(buf: &mut [u8], mask: &[u8; 4]) {
+    for (i, byte) in buf.iter_mut().enumerate() {
+        *byte ^= mask[i & 3];
+    }
+}
+
+/// Faster version of `apply_mask()` which operates on 4-byte blocks.
+#[inline]
+#[allow(dead_code)]
+fn apply_mask_fast32(buf: &mut [u8], mask: &[u8; 4]) {
+    // TODO replace this with read_unaligned() as it stabilizes.
+    let mask_u32 = unsafe {
+        let mut m: u32 = uninitialized();
+        #[allow(trivial_casts)]
+        copy_nonoverlapping(mask.as_ptr(), &mut m as *mut _ as *mut u8, 4);
+        m
+    };
+
+    let mut ptr = buf.as_mut_ptr();
+    let mut len = buf.len();
+
+    // Possible first unaligned block.
+    let head = min(len, (4 - (ptr as usize & 3)) & 3);
+    let mask_u32 = if head > 0 {
+        unsafe {
+            xor_mem(ptr, mask_u32, head);
+            ptr = ptr.offset(head as isize);
+        }
+        len -= head;
+        if cfg!(target_endian = "big") {
+            mask_u32.rotate_left(8 * head as u32)
+        } else {
+            mask_u32.rotate_right(8 * head as u32)
+        }
+    } else {
+        mask_u32
+    };
+
+    if len > 0 {
+        debug_assert_eq!(ptr as usize % 4, 0);
+    }
+
+    // Properly aligned middle of the data.
+    while len > 4 {
+        unsafe {
+            *(ptr as *mut u32) ^= mask_u32;
+            ptr = ptr.offset(4);
+            len -= 4;
+        }
+    }
+
+    // Possible last block.
+    if len > 0 {
+        unsafe { xor_mem(ptr, mask_u32, len); }
+    }
+}
+
+#[inline]
+// TODO: copy_nonoverlapping here compiles to call memcpy. While it is not so inefficient,
+// it could be done better. The compiler does not see that len is limited to 3.
+unsafe fn xor_mem(ptr: *mut u8, mask: u32, len: usize) {
+    let mut b: u32 = uninitialized();
+    #[allow(trivial_casts)]
+    copy_nonoverlapping(ptr, &mut b as *mut _ as *mut u8, len);
+    b ^= mask;
+    #[allow(trivial_casts)]
+    copy_nonoverlapping(&b as *const _ as *const u8, ptr, len);
+}
+
+#[cfg(test)]
+mod tests {
+
+    use super::{apply_mask_fallback, apply_mask_fast32};
+
+    #[test]
+    fn test_apply_mask() {
+        let mask = [
+            0x6d, 0xb6, 0xb2, 0x80,
+        ];
+        let unmasked = vec![
+            0xf3, 0x00, 0x01, 0x02,  0x03, 0x80, 0x81, 0x82,
+            0xff, 0xfe, 0x00, 0x17,  0x74, 0xf9, 0x12, 0x03,
+        ];
+
+        // Check masking with proper alignment.
+        {
+            let mut masked = unmasked.clone();
+            apply_mask_fallback(&mut masked, &mask);
+
+            let mut masked_fast = unmasked.clone();
+            apply_mask_fast32(&mut masked_fast, &mask);
+
+            assert_eq!(masked, masked_fast);
+        }
+
+        // Check masking without alignment.
+        {
+            let mut masked = unmasked.clone();
+            apply_mask_fallback(&mut masked[1..], &mask);
+
+            let mut masked_fast = unmasked.clone();
+            apply_mask_fast32(&mut masked_fast[1..], &mask);
+
+            assert_eq!(masked, masked_fast);
+        }
+    }
+
+}
+

src/protocol/frame/mod.rs

@@ -3,6 +3,7 @@
 pub mod coding;
 
 mod frame;
+mod mask;
 
 pub use self::frame::Frame;
 pub use self::frame::CloseFrame;