diff --git a/CHANGELOG.md b/CHANGELOG.md
index 005ce69..377b9a3 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,11 @@
 
 -
 
+## 1.16.0-pre.1
+
+- Add `no_std` implementation based on `critical-section`
+- Deprecate `atomic-polyfill` feature (use the new `critical-section` instead)
+
 ## 1.15.0
 
 - Increase minimal supported Rust version to 1.56.0.
diff --git a/Cargo.toml b/Cargo.toml
index 1ad618f..228ce1f 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "once_cell"
-version = "1.15.0"
+version = "1.16.0-pre.1"
 authors = ["Aleksey Kladov <aleksey.kladov@gmail.com>"]
 license = "MIT OR Apache-2.0"
 edition = "2021"
@@ -20,19 +20,10 @@ exclude = ["*.png", "*.svg", "/Cargo.lock.msrv", "rustfmt.toml"]
 members = ["xtask"]
 
 [dependencies]
-# Uses parking_lot to implement once_cell::sync::OnceCell.
-# This makes no speed difference, but makes each OnceCell<T>
-# up to 16 bytes smaller, depending on the size of the T.
+# These optional dependencies are considered private impl details,
+# only features from `[features]` table are a part of semver-guarded API.
 parking_lot_core = { version = "0.9.3", optional = true, default_features = false }
-
-# To be used in order to enable the race feature on targets
-# that do not have atomics
-# *Warning:* This can be unsound. Please read the README of
-# [atomic-polyfill](https://github.com/embassy-rs/atomic-polyfill)
-# and make sure you understand all the implications
-atomic-polyfill = { version = "1", optional = true }
-
-# Uses `critical-section` to implement `once_cell::sync::OnceCell`.
+atomic_polyfill = { package = "atomic-polyfill", version = "1", optional = true }
 critical_section = { package = "critical-section", version = "1", optional = true }
 
 [dev-dependencies]
@@ -43,21 +34,32 @@ critical_section = { package = "critical-section", version = "1.1.1", features =
 
 [features]
 default = ["std"]
+
 # Enables `once_cell::sync` module.
-std = ["alloc", "sync"]
+std = ["alloc"]
+
 # Enables `once_cell::race::OnceBox` type.
 alloc = ["race"]
+
 # Enables `once_cell::race` module.
 race = []
+
+# Uses parking_lot to implement once_cell::sync::OnceCell.
+# This makes no speed difference, but makes each OnceCell<T>
+# up to 16 bytes smaller, depending on the size of the T.
+parking_lot = ["parking_lot_core"]
+
+# Uses `critical-section` to implement `sync` and `race` modules. in
+# `#![no_std]` mode. Please read `critical-section` docs carefully
+# before enabling this feature.
+critical-section = ["critical_section", "atomic_polyfill" ]
+
 # Enables semver-exempt APIs of this crate.
 # At the moment, this feature is unused.
 unstable = []
 
-sync = []
-
-parking_lot = ["parking_lot_core"]
-
-critical-section = ["critical_section", "atomic-polyfill", "sync"]
+# Only for backwards compatibility.
+atomic-polyfill = ["critical-section"]
 
 [[example]]
 name = "bench"
diff --git a/src/imp_cs.rs b/src/imp_cs.rs
index 2548f58..668f18e 100644
--- a/src/imp_cs.rs
+++ b/src/imp_cs.rs
@@ -63,7 +63,7 @@ impl<T> OnceCell<T> {
     pub(crate) unsafe fn get_unchecked(&self) -> &T {
         debug_assert!(self.is_initialized());
         // SAFETY: The caller ensures that the value is initialized and access synchronized.
-        self.value.borrow(CriticalSection::new()).get_unchecked()
+        crate::unwrap_unchecked(self.value.borrow(CriticalSection::new()).get())
     }
 
     #[inline]
diff --git a/src/imp_pl.rs b/src/imp_pl.rs
index b723587..84d8593 100644
--- a/src/imp_pl.rs
+++ b/src/imp_pl.rs
@@ -58,7 +58,7 @@ impl<T> OnceCell<T> {
             //   but that is more complicated
             // - finally, if it returns Ok, we store the value and store the flag with
             //   `Release`, which synchronizes with `Acquire`s.
-            let f = unsafe { crate::take_unchecked(&mut f) };
+            let f = unsafe { crate::unwrap_unchecked(f.take()) };
             match f() {
                 Ok(value) => unsafe {
                     // Safe b/c we have a unique access and no panic may happen
diff --git a/src/imp_std.rs b/src/imp_std.rs
index f828d45..5761f01 100644
--- a/src/imp_std.rs
+++ b/src/imp_std.rs
@@ -78,7 +78,7 @@ impl<T> OnceCell<T> {
         initialize_or_wait(
             &self.queue,
             Some(&mut || {
-                let f = unsafe { crate::take_unchecked(&mut f) };
+                let f = unsafe { crate::unwrap_unchecked(f.take()) };
                 match f() {
                     Ok(value) => {
                         unsafe { *slot = Some(value) };
diff --git a/src/lib.rs b/src/lib.rs
index f44bc4c..448989c 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -474,18 +474,6 @@ pub mod unsync {
             unsafe { &mut *self.inner.get() }.as_mut()
         }
 
-        /// Get the reference to the underlying value, without checking if the
-        /// cell is initialized.
-        ///
-        /// # Safety
-        ///
-        /// Caller must ensure that the cell is in initialized state.
-        #[cfg(feature = "critical-section")]
-        #[inline]
-        pub(crate) unsafe fn get_unchecked(&self) -> &T {
-            crate::unwrap_unchecked(self.get())
-        }
-
         /// Sets the contents of this cell to `value`.
         ///
         /// Returns `Ok(())` if the cell was empty and `Err(value)` if it was
@@ -830,7 +818,7 @@ pub mod unsync {
 }
 
 /// Thread-safe, blocking version of `OnceCell`.
-#[cfg(feature = "sync")]
+#[cfg(any(feature = "std", feature = "critical-section"))]
 pub mod sync {
     use core::{
         cell::Cell,
@@ -839,7 +827,7 @@ pub mod sync {
         panic::RefUnwindSafe,
     };
 
-    use super::{imp::OnceCell as Imp, take_unchecked};
+    use super::{imp::OnceCell as Imp, unwrap_unchecked};
 
     /// A thread-safe cell which can be written to only once.
     ///
@@ -960,7 +948,7 @@ pub mod sync {
         /// assert_eq!(*value, 92);
         /// t.join().unwrap();
         /// ```
-        #[cfg(not(feature = "critical-section"))]
+        #[cfg(feature = "std")]
         pub fn wait(&self) -> &T {
             if !self.0.is_initialized() {
                 self.0.wait()
@@ -1050,7 +1038,7 @@ pub mod sync {
         /// ```
         pub fn try_insert(&self, value: T) -> Result<&T, (&T, T)> {
             let mut value = Some(value);
-            let res = self.get_or_init(|| unsafe { take_unchecked(&mut value) });
+            let res = self.get_or_init(|| unsafe { unwrap_unchecked(value.take()) });
             match value {
                 None => Ok(res),
                 Some(value) => Err((res, value)),
@@ -1377,12 +1365,7 @@ pub mod sync {
 #[cfg(feature = "race")]
 pub mod race;
 
-#[cfg(feature = "sync")]
-#[inline]
-unsafe fn take_unchecked<T>(val: &mut Option<T>) -> T {
-    unwrap_unchecked(val.take())
-}
-
+// Remove once MSRV is at least 1.58.
 #[inline]
 unsafe fn unwrap_unchecked<T>(val: Option<T>) -> T {
     match val {
diff --git a/src/race.rs b/src/race.rs
index 28acf62..fd255c4 100644
--- a/src/race.rs
+++ b/src/race.rs
@@ -19,9 +19,9 @@
 //! `Acquire` and `Release` have very little performance overhead on most
 //! architectures versus `Relaxed`.
 
-#[cfg(feature = "atomic-polyfill")]
+#[cfg(feature = "critical-section")]
 use atomic_polyfill as atomic;
-#[cfg(not(feature = "atomic-polyfill"))]
+#[cfg(not(feature = "critical-section"))]
 use core::sync::atomic;
 
 use atomic::{AtomicUsize, Ordering};
diff --git a/tests/it.rs b/tests/it.rs
index 118be00..d18f0a1 100644
--- a/tests/it.rs
+++ b/tests/it.rs
@@ -249,14 +249,14 @@ mod unsync {
     }
 }
 
-#[cfg(feature = "sync")]
+#[cfg(any(feature = "std", feature = "critical-section"))]
 mod sync {
     use std::sync::atomic::{AtomicUsize, Ordering::SeqCst};
 
-    #[cfg(not(feature = "critical-section"))]
+    #[cfg(feature = "std")]
     use std::sync::Barrier;
 
-    #[cfg(feature = "critical-section")]
+    #[cfg(not(feature = "std"))]
     use core::cell::Cell;
 
     use crossbeam_utils::thread::scope;
@@ -360,7 +360,7 @@ mod sync {
         assert_eq!(cell.get(), Some(&"hello".to_string()));
     }
 
-    #[cfg(not(feature = "critical-section"))]
+    #[cfg(feature = "std")]
     #[test]
     fn wait() {
         let cell: OnceCell<String> = OnceCell::new();
@@ -372,7 +372,7 @@ mod sync {
         .unwrap();
     }
 
-    #[cfg(not(feature = "critical-section"))]
+    #[cfg(feature = "std")]
     #[test]
     fn get_or_init_stress() {
         let n_threads = if cfg!(miri) { 30 } else { 1_000 };
@@ -437,7 +437,7 @@ mod sync {
 
     #[test]
     #[cfg_attr(miri, ignore)] // miri doesn't support processes
-    #[cfg(not(feature = "critical-section"))]
+    #[cfg(feature = "std")]
     fn reentrant_init() {
         let examples_dir = {
             let mut exe = std::env::current_exe().unwrap();
@@ -465,7 +465,7 @@ mod sync {
         }
     }
 
-    #[cfg(feature = "critical-section")]
+    #[cfg(not(feature = "std"))]
     #[test]
     #[should_panic(expected = "reentrant init")]
     fn reentrant_init() {
@@ -658,7 +658,7 @@ mod sync {
         }
     }
 
-    #[cfg(not(feature = "critical-section"))]
+    #[cfg(feature = "std")]
     #[test]
     fn get_does_not_block() {
         let cell = OnceCell::new();
@@ -692,7 +692,7 @@ mod sync {
 
 #[cfg(feature = "race")]
 mod race {
-    #[cfg(not(feature = "critical-section"))]
+    #[cfg(feature = "std")]
     use std::sync::Barrier;
     use std::{
         num::NonZeroUsize,
@@ -748,7 +748,7 @@ mod race {
         assert_eq!(cell.get(), Some(val1));
     }
 
-    #[cfg(not(feature = "critical-section"))]
+    #[cfg(feature = "std")]
     #[test]
     fn once_non_zero_usize_first_wins() {
         let val1 = NonZeroUsize::new(92).unwrap();
@@ -828,14 +828,14 @@ mod race {
 
 #[cfg(all(feature = "race", feature = "alloc"))]
 mod race_once_box {
-    #[cfg(not(feature = "critical-section"))]
+    #[cfg(feature = "std")]
     use std::sync::Barrier;
     use std::sync::{
         atomic::{AtomicUsize, Ordering::SeqCst},
         Arc,
     };
 
-    #[cfg(not(feature = "critical-section"))]
+    #[cfg(feature = "std")]
     use crossbeam_utils::thread::scope;
 
     use once_cell::race::OnceBox;
@@ -867,7 +867,7 @@ mod race_once_box {
         }
     }
 
-    #[cfg(not(feature = "critical-section"))]
+    #[cfg(feature = "std")]
     #[test]
     fn once_box_smoke_test() {
         let heap = Heap::default();
@@ -922,7 +922,7 @@ mod race_once_box {
         assert_eq!(heap.total(), 0);
     }
 
-    #[cfg(not(feature = "critical-section"))]
+    #[cfg(feature = "std")]
     #[test]
     fn once_box_first_wins() {
         let cell = OnceBox::new();
diff --git a/xtask/src/main.rs b/xtask/src/main.rs
index aaae075..2a5b66d 100644
--- a/xtask/src/main.rs
+++ b/xtask/src/main.rs
@@ -34,7 +34,7 @@ fn main() -> xshell::Result<()> {
         cmd!(sh, "cargo test --no-default-features --features unstable,alloc --test it").run()?;
 
         cmd!(sh, "cargo test --no-default-features --features critical-section").run()?;
-        cmd!(sh, "cargo test --no-default-features --features critical-section --release").run()?;
+        cmd!(sh, "cargo test --features critical-section").run()?;
     }
 
     {