diff --git a/src/derive/array.rs b/src/derive/array.rs
index 49dc6b2..eac6f83 100644
--- a/src/derive/array.rs
+++ b/src/derive/array.rs
@@ -1,6 +1,7 @@
 use crate::coder::{Buffer, Decoder, Encoder, Result, View};
 use crate::consume::mul_length;
 use crate::derive::{Decode, Encode};
+use crate::fast::{FastSlice, FastVec, Unaligned};
 use std::mem::MaybeUninit;
 use std::num::NonZeroUsize;
 
@@ -15,6 +16,23 @@ impl<T: Encode, const N: usize> Default for ArrayEncoder<T, N> {
 }
 
 impl<T: Encode, const N: usize> Encoder<[T; N]> for ArrayEncoder<T, N> {
+    fn as_primitive(&mut self) -> Option<&mut FastVec<[T; N]>> {
+        // FastVec doesn't work on ZST.
+        if N == 0 {
+            return None;
+        }
+        self.0.as_primitive().map(|v| {
+            debug_assert!(v.len() % N == 0);
+            // Safety: FastVec uses pointers for len/cap unlike Vec, so casting to FastVec<[T; N]>
+            // is safe as long as `v.len() % N == 0`. This will always be the case since we only
+            // encode in chunks of N.
+            // NOTE: If panics occurs during ArrayEncoder::encode and Buffer is reused, this
+            // invariant can be violated. Luckily primitive encoders never panic.
+            // TODO std::mem::take Buffer while encoding to avoid corrupted buffers.
+            unsafe { std::mem::transmute(v) }
+        })
+    }
+
     #[inline(always)]
     fn encode(&mut self, array: &[T; N]) {
         // TODO use encode_vectored if N is large enough.
@@ -59,6 +77,14 @@ impl<'a, T: Decode<'a>, const N: usize> View<'a> for ArrayDecoder<'a, T, N> {
 }
 
 impl<'a, T: Decode<'a>, const N: usize> Decoder<'a, [T; N]> for ArrayDecoder<'a, T, N> {
+    fn as_primitive(&mut self) -> Option<&mut FastSlice<Unaligned<[T; N]>>> {
+        self.0.as_primitive().map(|s| {
+            // Safety: FastSlice doesn't have a length unlike slice, so casting to FastSlice<[T; N]>
+            // is safe. N == 0 case is also safe for the same reason.
+            unsafe { std::mem::transmute(s) }
+        })
+    }
+
     #[inline(always)]
     fn decode_in_place(&mut self, out: &mut MaybeUninit<[T; N]>) {
         // Safety: Equivalent to nightly MaybeUninit::transpose.
@@ -68,3 +94,22 @@ impl<'a, T: Decode<'a>, const N: usize> Decoder<'a, [T; N]> for ArrayDecoder<'a,
         }
     }
 }
+
+#[cfg(test)]
+mod tests {
+    #[test]
+    fn test_empty_array() {
+        type T = [u8; 0];
+        let empty_array = T::default();
+        crate::decode::<T>(&crate::encode(&empty_array)).unwrap();
+        crate::decode::<Vec<T>>(&crate::encode(&vec![empty_array; 100])).unwrap();
+    }
+
+    fn bench_data() -> Vec<Vec<[u8; 3]>> {
+        crate::random_data::<u8>(125)
+            .into_iter()
+            .map(|n| (0..n / 16).map(|_| [0, 0, 255]).collect())
+            .collect()
+    }
+    crate::bench_encode_decode!(u8_array_vecs: Vec<Vec<[u8; 3]>>);
+}
diff --git a/src/fast.rs b/src/fast.rs
index 4b8e000..e582b47 100644
--- a/src/fast.rs
+++ b/src/fast.rs
@@ -67,7 +67,7 @@ impl<T> From<Vec<T>> for FastVec<T> {
 }
 
 impl<T> FastVec<T> {
-    fn len(&self) -> usize {
+    pub fn len(&self) -> usize {
         sub_ptr(self.end, self.start)
     }
 
@@ -212,7 +212,7 @@ impl<'a, T: Copy, const N: usize> PushUnchecked<T> for FastArrayVec<'a, T, N> {
 pub struct FastSlice<'a, T> {
     ptr: *const T,
     #[cfg(debug_assertions)]
-    len: usize, // TODO could store end ptr to allow Debug and as_slice.
+    end: *const T, // Uses pointer instead of len to permit &mut FastSlice<T> -> &mut FastSlice<[T; N]> cast.
     _spooky: PhantomData<&'a T>,
 }
 
@@ -233,7 +233,7 @@ impl<'a, T> From<&'a [T]> for FastSlice<'a, T> {
         Self {
             ptr: slice.as_ptr(),
             #[cfg(debug_assertions)]
-            len: slice.len(),
+            end: slice.as_ptr_range().end,
             _spooky: PhantomData,
         }
     }
@@ -247,7 +247,7 @@ impl<'a, T> FastSlice<'a, T> {
         Self {
             ptr,
             #[cfg(debug_assertions)]
-            len,
+            end: ptr.wrapping_add(len),
             _spooky: PhantomData,
         }
     }
@@ -256,9 +256,7 @@ impl<'a, T> FastSlice<'a, T> {
     #[inline(always)]
     pub unsafe fn next_unchecked_as_ptr(&mut self) -> *const T {
         #[cfg(debug_assertions)]
-        {
-            self.len = self.len.checked_sub(1).unwrap();
-        }
+        assert!((self.ptr.wrapping_add(1) as usize) <= self.end as usize);
         let p = self.ptr;
         self.ptr = self.ptr.add(1);
         p
@@ -267,9 +265,7 @@ impl<'a, T> FastSlice<'a, T> {
     #[inline(always)]
     pub unsafe fn advance(&mut self, n: usize) {
         #[cfg(debug_assertions)]
-        {
-            self.len = self.len.checked_sub(n).unwrap();
-        }
+        assert!((self.ptr.wrapping_add(n) as usize) <= self.end as usize);
         self.ptr = self.ptr.add(n);
     }
 
@@ -307,9 +303,7 @@ impl<'a, T: Copy> NextUnchecked<'a, T> for FastSlice<'a, T> {
     #[inline(always)]
     unsafe fn next_unchecked(&mut self) -> T {
         #[cfg(debug_assertions)]
-        {
-            self.len = self.len.checked_sub(1).unwrap();
-        }
+        assert!((self.ptr.wrapping_add(1) as usize) <= self.end as usize);
         let t = *self.ptr;
         self.ptr = self.ptr.add(1);
         t
@@ -318,9 +312,7 @@ impl<'a, T: Copy> NextUnchecked<'a, T> for FastSlice<'a, T> {
     #[inline(always)]
     unsafe fn chunk_unchecked(&mut self, length: usize) -> &'a [T] {
         #[cfg(debug_assertions)]
-        {
-            self.len = self.len.checked_sub(length).unwrap();
-        }
+        assert!((self.ptr.wrapping_add(length) as usize) <= self.end as usize);
         let slice = std::slice::from_raw_parts(self.ptr, length);
         self.ptr = self.ptr.add(length);
         slice
@@ -373,7 +365,7 @@ impl<'borrowed, T> CowSlice<'borrowed, T> {
         'borrowed: 'me,
     {
         #[cfg(debug_assertions)]
-        assert_eq!(self.slice.len, len);
+        assert_eq!(self.slice.ptr.wrapping_add(len), self.slice.end);
         std::slice::from_raw_parts(self.slice.ptr, len)
     }