optimize aligned floats

src/num_traits.rs

@@ -1,12 +1,21 @@
+use crate::Endianness;
+use std::array::TryFromSliceError;
+use std::convert::TryFrom;
+
 /// some extra number traits
 
 /// Allow casting floats unchecked
 pub trait UncheckedPrimitiveFloat: Sized {
+    type BYTES: AsRef<[u8]> + for<'a> TryFrom<&'a [u8], Error = TryFromSliceError>;
+
     fn from_f32_unchecked(n: f32) -> Self;
     fn from_f64_unchecked(n: f64) -> Self;
+    fn to_bytes<E: Endianness>(self) -> Self::BYTES;
+    fn from_bytes<E: Endianness>(bytes: Self::BYTES) -> Self;
 }
 
 impl UncheckedPrimitiveFloat for f32 {
+    type BYTES = [u8; 4];
     #[inline(always)]
     fn from_f32_unchecked(n: f32) -> Self {
         n
@@ -15,9 +24,24 @@ impl UncheckedPrimitiveFloat for f32 {
     fn from_f64_unchecked(n: f64) -> Self {
         n as f32
     }
+    fn to_bytes<E: Endianness>(self) -> Self::BYTES {
+        if E::is_le() {
+            self.to_le_bytes()
+        } else {
+            self.to_be_bytes()
+        }
+    }
+    fn from_bytes<E: Endianness>(bytes: Self::BYTES) -> Self {
+        if E::is_le() {
+            Self::from_le_bytes(bytes)
+        } else {
+            Self::from_be_bytes(bytes)
+        }
+    }
 }
 
 impl UncheckedPrimitiveFloat for f64 {
+    type BYTES = [u8; 8];
     #[inline(always)]
     fn from_f32_unchecked(n: f32) -> Self {
         n as f64
@@ -26,6 +50,20 @@ impl UncheckedPrimitiveFloat for f64 {
     fn from_f64_unchecked(n: f64) -> Self {
         n
     }
+    fn to_bytes<E: Endianness>(self) -> Self::BYTES {
+        if E::is_le() {
+            self.to_le_bytes()
+        } else {
+            self.to_be_bytes()
+        }
+    }
+    fn from_bytes<E: Endianness>(bytes: Self::BYTES) -> Self {
+        if E::is_le() {
+            Self::from_le_bytes(bytes)
+        } else {
+            Self::from_be_bytes(bytes)
+        }
+    }
 }
 
 /// Allow casting integers unchecked

src/readbuffer.rs

@@ -737,16 +737,24 @@ where
     where
         T: Float + UncheckedPrimitiveFloat,
     {
-        if size_of::<T>() == 4 {
-            let int = if size_of::<T>() < USIZE_SIZE {
-                self.read_fit_usize::<u32>(position, 32, end)
-            } else {
-                self.read_no_fit_usize::<u32>(position, 32, end)
-            };
-            T::from_f32_unchecked(f32::from_bits(int))
+        if position & 7 == 0 {
+            let byte_pos = position / 8;
+            let bytes = self.slice[byte_pos..byte_pos + size_of::<T>()]
+                .try_into()
+                .unwrap();
+            T::from_bytes::<E>(bytes)
         } else {
-            let int = self.read_no_fit_usize::<u64>(position, 64, end);
-            T::from_f64_unchecked(f64::from_bits(int))
+            if size_of::<T>() == 4 {
+                let int = if size_of::<T>() < USIZE_SIZE {
+                    self.read_fit_usize::<u32>(position, 32, end)
+                } else {
+                    self.read_no_fit_usize::<u32>(position, 32, end)
+                };
+                T::from_f32_unchecked(f32::from_bits(int))
+            } else {
+                let int = self.read_no_fit_usize::<u64>(position, 64, end);
+                T::from_f64_unchecked(f64::from_bits(int))
+            }
         }
     }
 

src/writebuffer.rs

@@ -93,4 +93,10 @@ impl<'a, E: Endianness> WriteBuffer<'a, E> {
         let merged = merged.to_le_bytes();
         self.bytes[byte_pos..byte_pos + byte_count].copy_from_slice(&merged[0..byte_count]);
     }
+
+    pub fn extends_from_slice(&mut self, slice: &[u8]) {
+        debug_assert_eq!(0, self.bit_len & 7);
+        self.bytes.extend_from_slice(slice);
+        self.bit_len += slice.len() * 8
+    }
 }

src/writestream.rs

@@ -173,14 +173,19 @@ where
     where
         T: Float + UncheckedPrimitiveFloat,
     {
-        if size_of::<T>() == 4 {
-            if size_of::<T>() < USIZE_SIZE {
-                self.push_bits(value.to_f32().unwrap().to_bits() as usize, 32);
-            } else {
-                self.push_non_fit_bits(value.to_f32().unwrap().to_bits().into_bytes(), 32)
-            };
+        if self.buffer.bit_len() & 7 == 0 {
+            let bytes = value.to_bytes::<E>();
+            self.buffer.extends_from_slice(bytes.as_ref());
         } else {
-            self.push_non_fit_bits(value.to_f64().unwrap().to_bits().into_bytes(), 64)
+            if size_of::<T>() == 4 {
+                if size_of::<T>() < USIZE_SIZE {
+                    self.push_bits(value.to_f32().unwrap().to_bits() as usize, 32);
+                } else {
+                    self.push_non_fit_bits(value.to_f32().unwrap().to_bits().into_bytes(), 32)
+                };
+            } else {
+                self.push_non_fit_bits(value.to_f64().unwrap().to_bits().into_bytes(), 64)
+            }
         }
 
         Ok(())
@@ -205,10 +210,14 @@ where
     /// ```
     #[inline]
     pub fn write_bytes(&mut self, bytes: &[u8]) -> Result<()> {
-        bytes
-            .iter()
-            .copied()
-            .for_each(|chunk| self.push_bits(chunk as usize, 8));
+        if self.buffer.bit_len() & 7 == 0 {
+            self.buffer.extends_from_slice(bytes);
+        } else {
+            bytes
+                .iter()
+                .copied()
+                .for_each(|chunk| self.push_bits(chunk as usize, 8));
+        }
         Ok(())
     }
 

tests/roundtrip.rs

@@ -3,6 +3,7 @@ use bitbuffer::{
 };
 use std::fmt::Debug;
 
+#[track_caller]
 fn roundtrip<
     T: BitRead<'static, BigEndian>
         + BitWrite<BigEndian>