diff --git a/Cargo.toml b/Cargo.toml
index 3c15d50..cbb91c3 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -7,5 +7,6 @@ edition = "2024"
 default = ["alloc"]
 alloc = []
 std = []
+nightly = []
 
 [dependencies]
diff --git a/src/cachepadded.rs b/src/cachepadded.rs
index 9f4222c..52cf323 100644
--- a/src/cachepadded.rs
+++ b/src/cachepadded.rs
@@ -35,7 +35,7 @@ use core::ops::{Deref, DerefMut};
 /// Alignment and padding:
 ///
 /// ```
-/// use crossbeam_utils::CachePadded;
+/// use werkzeug::CachePadded;
 ///
 /// let array = [CachePadded::new(1i8), CachePadded::new(2i8)];
 /// let addr1 = &*array[0] as *const i8 as usize;
@@ -51,7 +51,7 @@ use core::ops::{Deref, DerefMut};
 /// each other's cache lines:
 ///
 /// ```
-/// use crossbeam_utils::CachePadded;
+/// use werkzeug::CachePadded;
 /// use std::sync::atomic::AtomicUsize;
 ///
 /// struct Queue<T> {
@@ -159,7 +159,7 @@ impl<T> CachePadded<T> {
     /// # Examples
     ///
     /// ```
-    /// use crossbeam_utils::CachePadded;
+    /// use werkzeug::CachePadded;
     ///
     /// let padded_value = CachePadded::new(1);
     /// ```
@@ -172,7 +172,7 @@ impl<T> CachePadded<T> {
     /// # Examples
     ///
     /// ```
-    /// use crossbeam_utils::CachePadded;
+    /// use werkzeug::CachePadded;
     ///
     /// let padded_value = CachePadded::new(7);
     /// let value = padded_value.into_inner();
diff --git a/src/lib.rs b/src/lib.rs
index cbc549e..b2a7e43 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -1,4 +1,5 @@
 #![cfg_attr(not(feature = "std"), no_std)]
+#![cfg_attr(feature = "nightly", feature(strict_provenance_atomic_ptr))]
 
 #[cfg(any(test, feature = "std", feature = "alloc"))]
 extern crate alloc;
@@ -14,4 +15,5 @@ pub mod rand;
 pub mod smallbox;
 pub mod util;
 
+pub use cachepadded::CachePadded;
 pub use util::can_transmute;
diff --git a/src/ptr.rs b/src/ptr.rs
index f96009c..dca0c4d 100644
--- a/src/ptr.rs
+++ b/src/ptr.rs
@@ -6,7 +6,7 @@ use core::{
     num::NonZero,
     ops::{Deref, DerefMut},
     ptr::NonNull,
-    sync::atomic::{self, AtomicPtr, AtomicUsize},
+    sync::atomic::{self, AtomicPtr},
 };
 
 #[repr(transparent)]
@@ -203,24 +203,39 @@ impl<T, const BITS: u8> TaggedAtomicPtr<T, BITS> {
 
         // TODO: switch to fetch_or when stable
         // let old_ptr = self.ptr.fetch_or(tag & mask, order);
+        #[cfg(feature = "nightly")]
+        {
+            let old_ptr = self.ptr.fetch_or(tag & mask, order);
+            old_ptr.addr() & mask
+        }
+        #[cfg(not(feature = "nightly"))]
+        {
+            use core::sync::atomic::AtomicUsize;
+            let ptr = unsafe { AtomicUsize::from_ptr(self.ptr.as_ptr() as *mut usize) };
+            let old_ptr = ptr.fetch_or(tag & mask, order);
 
-        let ptr = unsafe { AtomicUsize::from_ptr(self.ptr.as_ptr() as *mut usize) };
-        let old_ptr = ptr.fetch_or(tag & mask, order);
-
-        old_ptr & mask
+            old_ptr & mask
+        }
     }
 
     /// returns the tag and clears it
     pub fn take_tag(&self, order: atomic::Ordering) -> usize {
         let mask = Self::mask();
 
-        // TODO: switch to fetch_and when stable
-        // let old_ptr = self.ptr.fetch_and(!mask, order);
+        #[cfg(feature = "nightly")]
+        {
+            let old_ptr = self.ptr.fetch_and(!mask, order);
+            old_ptr.addr() & mask
+        }
 
-        let ptr = unsafe { AtomicUsize::from_ptr(self.ptr.as_ptr() as *mut usize) };
-        let old_ptr = ptr.fetch_and(!mask, order);
+        #[cfg(not(feature = "nightly"))]
+        {
+            use core::sync::atomic::AtomicUsize;
+            let ptr = unsafe { AtomicUsize::from_ptr(self.ptr.as_ptr() as *mut usize) };
+            let old_ptr = ptr.fetch_and(!mask, order);
 
-        old_ptr & mask
+            old_ptr & mask
+        }
     }
 
     /// returns tag