some 32bit optimizations

src/num_traits.rs

@@ -256,19 +256,50 @@ macro_rules! impl_is_signed {
 }
 
 pub trait IntoBytes: Sized {
-    type Iter: DoubleEndedIterator<Item = u8> + ExactSizeIterator;
+    type BytesIter: DoubleEndedIterator<Item = u8> + ExactSizeIterator;
+    type U16Iter: DoubleEndedIterator<Item = u16> + ExactSizeIterator;
 
-    fn into_bytes(self) -> Self::Iter;
+    fn into_bytes(self) -> Self::BytesIter;
+
+    fn into_u16(self) -> Self::U16Iter;
 }
 
 macro_rules! impl_into_bytes {
-    ($type:ty, $bytes:expr) => {
+    ($type:ty, $bytes:expr, 1 ) => {
         impl IntoBytes for $type {
-            type Iter = std::array::IntoIter<u8, $bytes>;
+            type BytesIter = std::array::IntoIter<u8, $bytes>;
+            type U16Iter = std::array::IntoIter<u16, 1>;
 
             #[inline(always)]
-            fn into_bytes(self) -> Self::Iter {
+            fn into_bytes(self) -> Self::BytesIter {
-                Self::Iter::new(self.to_le_bytes())
+                Self::BytesIter::new(self.to_le_bytes())
+            }
+
+            #[inline(always)]
+            fn into_u16(self) -> Self::U16Iter {
+                Self::U16Iter::new([self as u16])
+            }
+        }
+    };
+    ($type:ty, $bytes:expr, $shorts:expr ) => {
+        impl IntoBytes for $type {
+            type BytesIter = std::array::IntoIter<u8, $bytes>;
+            type U16Iter = std::array::IntoIter<u16, { $shorts }>;
+
+            #[inline(always)]
+            fn into_bytes(self) -> Self::BytesIter {
+                Self::BytesIter::new(self.to_le_bytes())
+            }
+
+            #[inline(always)]
+            fn into_u16(self) -> Self::U16Iter {
+                use std::convert::TryInto;
+
+                let bytes = self.to_le_bytes();
+                let (head, aligned, tail) = unsafe { bytes[..].align_to::<u16>() };
+                debug_assert_eq!(0, head.len());
+                debug_assert_eq!(0, tail.len());
+                Self::U16Iter::new(aligned.try_into().unwrap())
             }
         }
     };
@@ -287,24 +318,24 @@ impl_is_signed!(i64, true);
 impl_is_signed!(i128, true);
 impl_is_signed!(isize, true);
 
-impl_into_bytes!(u8, 1);
+impl_into_bytes!(u8, 1, 1);
-impl_into_bytes!(u16, 2);
+impl_into_bytes!(u16, 2, 1);
-impl_into_bytes!(u32, 4);
+impl_into_bytes!(u32, 4, 2);
-impl_into_bytes!(u64, 8);
+impl_into_bytes!(u64, 8, 4);
-impl_into_bytes!(u128, 16);
+impl_into_bytes!(u128, 16, 8);
 
 #[cfg(target_pointer_width = "64")]
-impl_into_bytes!(usize, 8);
+impl_into_bytes!(usize, 8, 4);
 #[cfg(target_pointer_width = "32")]
-impl_into_bytes!(usize, 4);
+impl_into_bytes!(usize, 4, 2);
 
-impl_into_bytes!(i8, 1);
+impl_into_bytes!(i8, 1, 1);
-impl_into_bytes!(i16, 2);
+impl_into_bytes!(i16, 2, 1);
-impl_into_bytes!(i32, 4);
+impl_into_bytes!(i32, 4, 2);
-impl_into_bytes!(i64, 8);
+impl_into_bytes!(i64, 8, 4);
-impl_into_bytes!(i128, 16);
+impl_into_bytes!(i128, 16, 8);
 
 #[cfg(target_pointer_width = "64")]
-impl_into_bytes!(isize, 8);
+impl_into_bytes!(isize, 8, 4);
 #[cfg(target_pointer_width = "32")]
-impl_into_bytes!(isize, 4);
+impl_into_bytes!(isize, 4, 2);

src/writebuffer.rs

@@ -29,13 +29,14 @@ impl<'a, E: Endianness> WriteBuffer<'a, E> {
     pub fn push_non_fit_bits<I>(&mut self, bits: I, count: usize)
     where
         I: ExactSizeIterator,
-        I: DoubleEndedIterator<Item = u8>,
+        I: DoubleEndedIterator<Item = u16>,
     {
-        let full_bytes = min(bits.len() - 1, count / 8);
+        let chunk_bits = u16::BITS as usize;
+        let full_bytes = min(bits.len() - 1, count / chunk_bits);
 
-        let counts = repeat(8)
+        let counts = repeat(chunk_bits)
             .take(full_bytes)
-            .chain(once(count - full_bytes * 8));
+            .chain(once(count - full_bytes * chunk_bits));
         if E::is_le() {
             bits.zip(counts)
                 .for_each(|(chunk, count)| self.push_bits(chunk as usize, count))
@@ -52,12 +53,14 @@ impl<'a, E: Endianness> WriteBuffer<'a, E> {
         if count == 0 {
             return;
         }
-        debug_assert!(count < USIZE_BITS - 8);
 
         // ensure there are no stray bits
         let bits = bits & (usize::MAX >> (USIZE_BITS - count));
 
         let bit_offset = self.bit_len & 7;
+
+        debug_assert!(count <= USIZE_BITS - bit_offset);
+
         let last_written_byte = if bit_offset > 0 {
             self.bytes.pop().unwrap_or(0)
         } else {

src/writestream.rs

@@ -76,7 +76,7 @@ where
     fn push_non_fit_bits<I>(&mut self, bits: I, count: usize)
     where
         I: ExactSizeIterator,
-        I: DoubleEndedIterator<Item = u8>,
+        I: DoubleEndedIterator<Item = u16>,
     {
         self.buffer.push_non_fit_bits(bits, count)
     }
@@ -142,10 +142,10 @@ where
             });
         }
 
-        if type_bit_size < USIZE_BITS || count < USIZE_BITS {
+        if type_bit_size < USIZE_BITS || count <= (USIZE_BITS - (self.bit_len() % 8)) {
             self.push_bits(value.into_usize_unchecked(), count);
         } else {
-            self.push_non_fit_bits(value.into_bytes(), count)
+            self.push_non_fit_bits(value.into_u16(), count)
         }
 
         Ok(())