diff --git a/src/bytes.rs b/src/bytes.rs
index ec95e802d..8d7abd040 100644
--- a/src/bytes.rs
+++ b/src/bytes.rs
@@ -184,6 +184,22 @@ impl Bytes {
         }
     }
 
+    /// Creates a new `Bytes` with length zero and the given pointer as the address.
+    fn new_empty_with_ptr(ptr: *const u8) -> Self {
+        debug_assert!(!ptr.is_null());
+
+        // Detach this pointer's provenance from whichever allocation it came from, and reattach it
+        // to the provenance of the fake ZST [u8;0] at the same address.
+        let ptr = without_provenance(ptr as usize);
+
+        Bytes {
+            ptr,
+            len: 0,
+            data: AtomicPtr::new(ptr::null_mut()),
+            vtable: &STATIC_VTABLE,
+        }
+    }
+
     /// Returns the number of bytes contained in this `Bytes`.
     ///
     /// # Examples
@@ -366,7 +382,9 @@ impl Bytes {
     /// Splits the bytes into two at the given index.
     ///
     /// Afterwards `self` contains elements `[0, at)`, and the returned `Bytes`
-    /// contains elements `[at, len)`.
+    /// contains elements `[at, len)`. It's guaranteed that the memory does not
+    /// move, that is, the address of `self` does not change, and the address of
+    /// the returned slice is `at` bytes after that.
     ///
     /// This is an `O(1)` operation that just increases the reference count and
     /// sets a few indices.
@@ -389,11 +407,11 @@ impl Bytes {
     #[must_use = "consider Bytes::truncate if you don't need the other half"]
     pub fn split_off(&mut self, at: usize) -> Self {
         if at == self.len() {
-            return Bytes::new();
+            return Bytes::new_empty_with_ptr(self.ptr.wrapping_add(at));
         }
 
         if at == 0 {
-            return mem::replace(self, Bytes::new());
+            return mem::replace(self, Bytes::new_empty_with_ptr(self.ptr));
         }
 
         assert!(
@@ -438,11 +456,12 @@ impl Bytes {
     #[must_use = "consider Bytes::advance if you don't need the other half"]
     pub fn split_to(&mut self, at: usize) -> Self {
         if at == self.len() {
-            return mem::replace(self, Bytes::new());
+            let end_ptr = self.ptr.wrapping_add(at);
+            return mem::replace(self, Bytes::new_empty_with_ptr(end_ptr));
         }
 
         if at == 0 {
-            return Bytes::new();
+            return Bytes::new_empty_with_ptr(self.ptr);
         }
 
         assert!(
@@ -1426,6 +1445,10 @@ where
     new_addr as *mut u8
 }
 
+fn without_provenance(ptr: usize) -> *const u8 {
+    core::ptr::null::<u8>().wrapping_add(ptr)
+}
+
 // compile-fails
 
 /// ```compile_fail
diff --git a/src/bytes_mut.rs b/src/bytes_mut.rs
index ec74c4e97..a0d77bc24 100644
--- a/src/bytes_mut.rs
+++ b/src/bytes_mut.rs
@@ -291,7 +291,9 @@ impl BytesMut {
     /// Splits the bytes into two at the given index.
     ///
     /// Afterwards `self` contains elements `[0, at)`, and the returned
-    /// `BytesMut` contains elements `[at, capacity)`.
+    /// `BytesMut` contains elements `[at, capacity)`. It's guaranteed that the
+    /// memory does not move, that is, the address of `self` does not change,
+    /// and the address of the returned slice is `at` bytes after that.
     ///
     /// This is an `O(1)` operation that just increases the reference count
     /// and sets a few indices.
diff --git a/tests/test_bytes.rs b/tests/test_bytes.rs
index 59c967b66..71f0e6681 100644
--- a/tests/test_bytes.rs
+++ b/tests/test_bytes.rs
@@ -1399,3 +1399,83 @@ fn try_reclaim_arc() {
     buf.advance(2);
     assert_eq!(true, buf.try_reclaim(6));
 }
+
+#[test]
+fn split_off_empty_addr() {
+    let mut buf = Bytes::from(vec![0; 1024]);
+
+    let ptr_start = buf.as_ptr();
+    let ptr_end = ptr_start.wrapping_add(1024);
+
+    let empty_end = buf.split_off(1024);
+    assert_eq!(empty_end.len(), 0);
+    assert_eq!(empty_end.as_ptr(), ptr_end);
+
+    let _ = buf.split_off(0);
+    assert_eq!(buf.len(), 0);
+    assert_eq!(buf.as_ptr(), ptr_start);
+
+    // Is miri happy about the provenance?
+    let _ = &empty_end[..];
+    let _ = &buf[..];
+}
+
+#[test]
+fn split_to_empty_addr() {
+    let mut buf = Bytes::from(vec![0; 1024]);
+
+    let ptr_start = buf.as_ptr();
+    let ptr_end = ptr_start.wrapping_add(1024);
+
+    let empty_start = buf.split_to(0);
+    assert_eq!(empty_start.len(), 0);
+    assert_eq!(empty_start.as_ptr(), ptr_start);
+
+    let _ = buf.split_to(1024);
+    assert_eq!(buf.len(), 0);
+    assert_eq!(buf.as_ptr(), ptr_end);
+
+    // Is miri happy about the provenance?
+    let _ = &empty_start[..];
+    let _ = &buf[..];
+}
+
+#[test]
+fn split_off_empty_addr_mut() {
+    let mut buf = BytesMut::from([0; 1024].as_slice());
+
+    let ptr_start = buf.as_ptr();
+    let ptr_end = ptr_start.wrapping_add(1024);
+
+    let empty_end = buf.split_off(1024);
+    assert_eq!(empty_end.len(), 0);
+    assert_eq!(empty_end.as_ptr(), ptr_end);
+
+    let _ = buf.split_off(0);
+    assert_eq!(buf.len(), 0);
+    assert_eq!(buf.as_ptr(), ptr_start);
+
+    // Is miri happy about the provenance?
+    let _ = &empty_end[..];
+    let _ = &buf[..];
+}
+
+#[test]
+fn split_to_empty_addr_mut() {
+    let mut buf = BytesMut::from([0; 1024].as_slice());
+
+    let ptr_start = buf.as_ptr();
+    let ptr_end = ptr_start.wrapping_add(1024);
+
+    let empty_start = buf.split_to(0);
+    assert_eq!(empty_start.len(), 0);
+    assert_eq!(empty_start.as_ptr(), ptr_start);
+
+    let _ = buf.split_to(1024);
+    assert_eq!(buf.len(), 0);
+    assert_eq!(buf.as_ptr(), ptr_end);
+
+    // Is miri happy about the provenance?
+    let _ = &empty_start[..];
+    let _ = &buf[..];
+}