diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index bf456b33a..497cea91b 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -25,7 +25,7 @@ jobs:
           - crossbeam-skiplist
           - crossbeam-utils
         rust:
-          - 1.28.0
+          - 1.30.0
           - nightly
     steps:
       - uses: actions/checkout@master
@@ -38,7 +38,7 @@ jobs:
           rustup target add thumbv6m-none-eabi
       # cfg-if 0.1.10 requires Rust 1.31+ so downgrade it.
       - name: Downgrade dependencies
-        if: matrix.rust == '1.28.0'
+        if: matrix.rust == '1.30.0'
         run: |
           cargo generate-lockfile
           cargo update -p cfg-if --precise 0.1.9
@@ -66,3 +66,14 @@ jobs:
         run: rustup update stable && rustup default stable
       - name: rustfmt
         run: ./ci/rustfmt.sh
+
+  # Run loom tests.
+  loom:
+    name: loom
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@master
+      - name: Install Rust
+        run: rustup update stable && rustup default stable
+      - name: loom
+        run: ./ci/crossbeam-epoch-loom.sh
diff --git a/ci/crossbeam-epoch-loom.sh b/ci/crossbeam-epoch-loom.sh
new file mode 100755
index 000000000..40949f3ed
--- /dev/null
+++ b/ci/crossbeam-epoch-loom.sh
@@ -0,0 +1,8 @@
+#!/bin/bash
+
+cd "$(dirname "$0")"/../crossbeam-epoch
+set -ex
+
+export RUSTFLAGS="-D warnings --cfg=loom"
+
+env LOOM_MAX_PREEMPTIONS=2 cargo test --test loom --features sanitize --release -- --nocapture
diff --git a/crossbeam-epoch/Cargo.toml b/crossbeam-epoch/Cargo.toml
index a1a056f5e..a67c924b7 100644
--- a/crossbeam-epoch/Cargo.toml
+++ b/crossbeam-epoch/Cargo.toml
@@ -27,6 +27,9 @@ cfg-if = "0.1.2"
 maybe-uninit = "2.0.0"
 memoffset = "0.5"
 
+[target.'cfg(loom)'.dependencies]
+loom = "0.3.2"
+
 [dependencies.crossbeam-utils]
 version = "0.7"
 path = "../crossbeam-utils"
diff --git a/crossbeam-epoch/src/atomic.rs b/crossbeam-epoch/src/atomic.rs
index a2044740b..184da8aab 100644
--- a/crossbeam-epoch/src/atomic.rs
+++ b/crossbeam-epoch/src/atomic.rs
@@ -1,4 +1,5 @@
 use alloc::boxed::Box;
+use concurrency::sync::atomic::{AtomicUsize, Ordering};
 use core::borrow::{Borrow, BorrowMut};
 use core::cmp;
 use core::fmt;
@@ -6,7 +7,6 @@ use core::marker::PhantomData;
 use core::mem;
 use core::ops::{Deref, DerefMut};
 use core::ptr;
-use core::sync::atomic::{AtomicUsize, Ordering};
 
 use crossbeam_utils::atomic::AtomicConsume;
 use guard::Guard;
@@ -150,7 +150,7 @@ impl<T> Atomic<T> {
     ///
     /// let a = Atomic::<i32>::null();
     /// ```
-    #[cfg(not(has_min_const_fn))]
+    #[cfg(any(loom, not(has_min_const_fn)))]
     pub fn null() -> Atomic<T> {
         Self {
             data: AtomicUsize::new(0),
@@ -167,7 +167,7 @@ impl<T> Atomic<T> {
     ///
     /// let a = Atomic::<i32>::null();
     /// ```
-    #[cfg(has_min_const_fn)]
+    #[cfg(all(not(loom), has_min_const_fn))]
     pub const fn null() -> Atomic<T> {
         Self {
             data: AtomicUsize::new(0),
@@ -506,7 +506,14 @@ impl<T> Atomic<T> {
     /// }
     /// ```
     pub unsafe fn into_owned(self) -> Owned<T> {
-        Owned::from_usize(self.data.into_inner())
+        #[cfg(loom)]
+        {
+            Owned::from_usize(self.data.unsync_load())
+        }
+        #[cfg(not(loom))]
+        {
+            Owned::from_usize(self.data.into_inner())
+        }
     }
 }
 
@@ -1185,7 +1192,7 @@ impl<'g, T> Default for Shared<'g, T> {
     }
 }
 
-#[cfg(test)]
+#[cfg(all(test, not(loom)))]
 mod tests {
     use super::Shared;
 
diff --git a/crossbeam-epoch/src/collector.rs b/crossbeam-epoch/src/collector.rs
index 1817d9ada..ac38690eb 100644
--- a/crossbeam-epoch/src/collector.rs
+++ b/crossbeam-epoch/src/collector.rs
@@ -12,7 +12,7 @@
 ///
 /// handle.pin().flush();
 /// ```
-use alloc::sync::Arc;
+use concurrency::sync::Arc;
 use core::fmt;
 
 use guard::Guard;
@@ -103,7 +103,7 @@ impl fmt::Debug for LocalHandle {
     }
 }
 
-#[cfg(test)]
+#[cfg(all(test, not(loom)))]
 mod tests {
     use std::mem;
     use std::sync::atomic::{AtomicUsize, Ordering};
@@ -145,9 +145,9 @@ mod tests {
                 let a = Owned::new(7).into_shared(guard);
                 guard.defer_destroy(a);
 
-                assert!(!(*(*guard.local).bag.get()).is_empty());
+                assert!(!(*guard.local).bag.with(|b| (*b).is_empty()));
 
-                while !(*(*guard.local).bag.get()).is_empty() {
+                while !(*guard.local).bag.with(|b| (*b).is_empty()) {
                     guard.flush();
                 }
             }
@@ -166,7 +166,7 @@ mod tests {
                 let a = Owned::new(7).into_shared(guard);
                 guard.defer_destroy(a);
             }
-            assert!(!(*(*guard.local).bag.get()).is_empty());
+            assert!(!(*guard.local).bag.with(|b| (*b).is_empty()));
         }
     }
 
diff --git a/crossbeam-epoch/src/default.rs b/crossbeam-epoch/src/default.rs
index 870e590fa..444d24534 100644
--- a/crossbeam-epoch/src/default.rs
+++ b/crossbeam-epoch/src/default.rs
@@ -5,6 +5,7 @@
 //! destructed on thread exit, which in turn unregisters the thread.
 
 use collector::{Collector, LocalHandle};
+use concurrency::{lazy_static, thread_local};
 use guard::Guard;
 
 lazy_static! {
@@ -44,7 +45,7 @@ where
         .unwrap_or_else(|_| f(&COLLECTOR.register()))
 }
 
-#[cfg(test)]
+#[cfg(all(test, not(loom)))]
 mod tests {
     use crossbeam_utils::thread;
 
diff --git a/crossbeam-epoch/src/deferred.rs b/crossbeam-epoch/src/deferred.rs
index a0970f115..c72b2af96 100644
--- a/crossbeam-epoch/src/deferred.rs
+++ b/crossbeam-epoch/src/deferred.rs
@@ -78,7 +78,7 @@ impl Deferred {
     }
 }
 
-#[cfg(test)]
+#[cfg(all(test, not(loom)))]
 mod tests {
     use super::Deferred;
     use std::cell::Cell;
diff --git a/crossbeam-epoch/src/epoch.rs b/crossbeam-epoch/src/epoch.rs
index e7759d935..51877cd56 100644
--- a/crossbeam-epoch/src/epoch.rs
+++ b/crossbeam-epoch/src/epoch.rs
@@ -7,7 +7,7 @@
 //! If an object became garbage in some epoch, then we can be sure that after two advancements no
 //! participant will hold a reference to it. That is the crux of safe memory reclamation.
 
-use core::sync::atomic::{AtomicUsize, Ordering};
+use concurrency::sync::atomic::{AtomicUsize, Ordering};
 
 /// An epoch that can be marked as pinned or unpinned.
 ///
diff --git a/crossbeam-epoch/src/internal.rs b/crossbeam-epoch/src/internal.rs
index 645511b9c..e34cb798a 100644
--- a/crossbeam-epoch/src/internal.rs
+++ b/crossbeam-epoch/src/internal.rs
@@ -35,11 +35,12 @@
 //! Ideally each instance of concurrent data structure may have its own queue that gets fully
 //! destroyed as soon as the data structure gets dropped.
 
-use core::cell::{Cell, UnsafeCell};
+use concurrency::cell::UnsafeCell;
+use concurrency::sync::atomic;
+use concurrency::sync::atomic::Ordering;
+use core::cell::Cell;
 use core::mem::{self, ManuallyDrop};
 use core::num::Wrapping;
-use core::sync::atomic;
-use core::sync::atomic::Ordering;
 use core::{fmt, ptr};
 
 use crossbeam_utils::CachePadded;
@@ -411,7 +412,7 @@ impl Local {
     /// Returns a reference to the `Collector` in which this `Local` resides.
     #[inline]
     pub fn collector(&self) -> &Collector {
-        unsafe { &**self.collector.get() }
+        self.collector.with(|c| unsafe { &**c })
     }
 
     /// Returns `true` if the current participant is pinned.
@@ -426,7 +427,7 @@ impl Local {
     ///
     /// It should be safe for another thread to execute the given function.
     pub unsafe fn defer(&self, mut deferred: Deferred, guard: &Guard) {
-        let bag = &mut *self.bag.get();
+        let bag = self.bag.with_mut(|b| &mut *b);
 
         while let Err(d) = bag.try_push(deferred) {
             self.global().push_bag(bag, guard);
@@ -435,7 +436,7 @@ impl Local {
     }
 
     pub fn flush(&self, guard: &Guard) {
-        let bag = unsafe { &mut *self.bag.get() };
+        let bag = self.bag.with_mut(|b| unsafe { &mut *b });
 
         if !bag.is_empty() {
             self.global().push_bag(bag, guard);
@@ -573,7 +574,8 @@ impl Local {
             // Pin and move the local bag into the global queue. It's important that `push_bag`
             // doesn't defer destruction on any new garbage.
             let guard = &self.pin();
-            self.global().push_bag(&mut *self.bag.get(), guard);
+            self.global()
+                .push_bag(self.bag.with_mut(|b| &mut *b), guard);
         }
         // Revert the handle count back to zero.
         self.handle_count.set(0);
@@ -582,7 +584,7 @@ impl Local {
             // Take the reference to the `Global` out of this `Local`. Since we're not protected
             // by a guard at this time, it's crucial that the reference is read before marking the
             // `Local` as deleted.
-            let collector: Collector = ptr::read(&*(*self.collector.get()));
+            let collector: Collector = ptr::read(self.collector.with(|c| &*(*c)));
 
             // Mark this node in the linked list as deleted.
             self.entry.delete(&unprotected());
@@ -613,7 +615,7 @@ impl IsElement<Local> for Local {
     }
 }
 
-#[cfg(test)]
+#[cfg(all(test, not(loom)))]
 mod tests {
     use std::sync::atomic::{AtomicUsize, Ordering};
 
diff --git a/crossbeam-epoch/src/lib.rs b/crossbeam-epoch/src/lib.rs
index 282bbe90f..c6c3c6b1c 100644
--- a/crossbeam-epoch/src/lib.rs
+++ b/crossbeam-epoch/src/lib.rs
@@ -66,6 +66,87 @@ extern crate core;
 
 extern crate maybe_uninit;
 
+#[cfg(loom)]
+extern crate loom;
+
+#[cfg(loom)]
+#[allow(unused_imports, dead_code)]
+pub(crate) mod concurrency {
+    pub(crate) mod cell {
+        pub(crate) use loom::cell::UnsafeCell;
+    }
+    pub(crate) mod sync {
+        pub(crate) mod atomic {
+            pub(crate) use loom::sync::atomic::{AtomicUsize, Ordering};
+            pub(crate) fn fence(ord: Ordering) {
+                if let Ordering::Acquire = ord {
+                } else {
+                    // FIXME: loom only supports acquire fences at the moment.
+                    // https://github.com/tokio-rs/loom/issues/117
+                    // let's at least not panic...
+                    // this may generate some false positives (`SeqCst` is stronger than `Acquire`
+                    // for example), and some false negatives (`Relaxed` is weaker than `Acquire`),
+                    // but it's the best we can do for the time being.
+                }
+                loom::sync::atomic::fence(Ordering::Acquire)
+            }
+
+            // FIXME: loom does not support compiler_fence at the moment.
+            // https://github.com/tokio-rs/loom/issues/117
+            // we use fence as a stand-in for compiler_fence for the time being.
+            // this may miss some races since fence is stronger than compiler_fence,
+            // but it's the best we can do for the time being.
+            pub(crate) use self::fence as compiler_fence;
+        }
+        pub(crate) use loom::sync::Arc;
+    }
+    pub(crate) use loom::lazy_static;
+    pub(crate) use loom::thread_local;
+}
+#[cfg(not(loom))]
+#[allow(unused_imports, dead_code)]
+pub(crate) mod concurrency {
+    #[cfg(any(feature = "alloc", feature = "std"))]
+    pub(crate) mod cell {
+        #[derive(Debug)]
+        #[repr(transparent)]
+        pub(crate) struct UnsafeCell<T>(::core::cell::UnsafeCell<T>);
+
+        impl<T> UnsafeCell<T> {
+            #[inline]
+            pub(crate) fn new(data: T) -> UnsafeCell<T> {
+                UnsafeCell(::core::cell::UnsafeCell::new(data))
+            }
+
+            #[inline]
+            pub(crate) fn with<R>(&self, f: impl FnOnce(*const T) -> R) -> R {
+                f(self.0.get())
+            }
+
+            #[inline]
+            pub(crate) fn with_mut<R>(&self, f: impl FnOnce(*mut T) -> R) -> R {
+                f(self.0.get())
+            }
+        }
+    }
+    #[cfg(any(feature = "alloc", feature = "std"))]
+    pub(crate) mod sync {
+        pub(crate) mod atomic {
+            pub(crate) use core::sync::atomic::compiler_fence;
+            pub(crate) use core::sync::atomic::fence;
+            pub(crate) use core::sync::atomic::{AtomicUsize, Ordering};
+        }
+        #[cfg_attr(feature = "nightly", cfg(target_has_atomic = "ptr"))]
+        pub(crate) use alloc::sync::Arc;
+    }
+
+    #[cfg(feature = "std")]
+    pub(crate) use std::thread_local;
+
+    #[cfg(feature = "std")]
+    pub(crate) use lazy_static::lazy_static;
+}
+
 cfg_if! {
     if #[cfg(feature = "alloc")] {
         extern crate alloc;
@@ -99,7 +180,6 @@ cfg_if! {
 
 cfg_if! {
     if #[cfg(feature = "std")] {
-        #[macro_use]
         extern crate lazy_static;
 
         mod default;
diff --git a/crossbeam-epoch/src/sync/list.rs b/crossbeam-epoch/src/sync/list.rs
index 8e8899ea2..09c98b435 100644
--- a/crossbeam-epoch/src/sync/list.rs
+++ b/crossbeam-epoch/src/sync/list.rs
@@ -3,8 +3,8 @@
 //! Ideas from Michael.  High Performance Dynamic Lock-Free Hash Tables and List-Based Sets.  SPAA
 //! 2002.  http://dl.acm.org/citation.cfm?id=564870.564881
 
+use concurrency::sync::atomic::Ordering::{Acquire, Relaxed, Release};
 use core::marker::PhantomData;
-use core::sync::atomic::Ordering::{Acquire, Relaxed, Release};
 
 use {unprotected, Atomic, Guard, Shared};
 
@@ -295,7 +295,7 @@ impl<'g, T: 'g, C: IsElement<T>> Iterator for Iter<'g, T, C> {
     }
 }
 
-#[cfg(test)]
+#[cfg(all(test, not(loom)))]
 mod tests {
     use super::*;
     use crossbeam_utils::thread;
diff --git a/crossbeam-epoch/src/sync/queue.rs b/crossbeam-epoch/src/sync/queue.rs
index 99fb6a1c4..306650610 100644
--- a/crossbeam-epoch/src/sync/queue.rs
+++ b/crossbeam-epoch/src/sync/queue.rs
@@ -8,7 +8,7 @@
 //! Simon Doherty, Lindsay Groves, Victor Luchangco, and Mark Moir. 2004b. Formal Verification of a
 //! Practical Lock-Free Queue Algorithm. https://doi.org/10.1007/978-3-540-30232-2_7
 
-use core::sync::atomic::Ordering::{Acquire, Relaxed, Release};
+use concurrency::sync::atomic::Ordering::{Acquire, Relaxed, Release};
 
 use crossbeam_utils::CachePadded;
 
@@ -203,7 +203,7 @@ impl<T> Drop for Queue<T> {
     }
 }
 
-#[cfg(test)]
+#[cfg(all(test, not(loom)))]
 mod test {
     use super::*;
     use crossbeam_utils::thread;
diff --git a/crossbeam-epoch/tests/loom.rs b/crossbeam-epoch/tests/loom.rs
new file mode 100644
index 000000000..1399af7eb
--- /dev/null
+++ b/crossbeam-epoch/tests/loom.rs
@@ -0,0 +1,147 @@
+#![cfg(loom)]
+
+extern crate crossbeam_epoch as epoch;
+extern crate crossbeam_utils as utils;
+extern crate loom;
+
+use epoch::*;
+use epoch::{Atomic, Owned};
+use loom::sync::atomic::Ordering::{self, Acquire, Relaxed, Release};
+use loom::sync::Arc;
+use loom::thread::spawn;
+use std::mem::ManuallyDrop;
+use std::ptr;
+
+#[test]
+fn it_works() {
+    loom::model(|| {
+        let collector = Collector::new();
+        let item: Atomic<String> = Atomic::from(Owned::new(String::from("boom")));
+        let item2 = item.clone();
+        let collector2 = collector.clone();
+        let guard = collector.register().pin();
+
+        let jh = loom::thread::spawn(move || {
+            let guard = collector2.register().pin();
+            guard.defer(move || {
+                // this isn't really safe, since other threads may still have pointers to the
+                // value, but in this limited test scenario it's okay, since we know the test won't
+                // access item after all the pins are released.
+                let mut item = unsafe { item2.into_owned() };
+                // mutate it as a second measure to make sure the assert_eq below would fail
+                item.retain(|c| c == 'o');
+                drop(item);
+            });
+        });
+
+        let item = item.load(Ordering::SeqCst, &guard);
+        // we pinned strictly before the call to defer_destroy,
+        // so item cannot have been dropped yet
+        assert_eq!(*unsafe { item.deref() }, "boom");
+        drop(guard);
+
+        jh.join().unwrap();
+
+        drop(collector);
+    })
+}
+
+#[test]
+fn treiber_stack() {
+    // this is mostly a copy-paste from the example
+    #[derive(Debug)]
+    pub struct TreiberStack<T> {
+        head: Atomic<Node<T>>,
+    }
+
+    #[derive(Debug)]
+    struct Node<T> {
+        data: ManuallyDrop<T>,
+        next: Atomic<Node<T>>,
+    }
+
+    impl<T> TreiberStack<T> {
+        pub fn new() -> TreiberStack<T> {
+            TreiberStack {
+                head: Atomic::null(),
+            }
+        }
+
+        pub fn push(&self, t: T) {
+            let mut n = Owned::new(Node {
+                data: ManuallyDrop::new(t),
+                next: Atomic::null(),
+            });
+
+            let guard = epoch::pin();
+
+            loop {
+                let head = self.head.load(Relaxed, &guard);
+                n.next.store(head, Relaxed);
+
+                match self.head.compare_and_set(head, n, Release, &guard) {
+                    Ok(_) => break,
+                    Err(e) => n = e.new,
+                }
+            }
+        }
+
+        pub fn pop(&self) -> Option<T> {
+            let guard = epoch::pin();
+            loop {
+                let head = self.head.load(Acquire, &guard);
+
+                match unsafe { head.as_ref() } {
+                    Some(h) => {
+                        let next = h.next.load(Relaxed, &guard);
+
+                        if self
+                            .head
+                            .compare_and_set(head, next, Relaxed, &guard)
+                            .is_ok()
+                        {
+                            unsafe {
+                                guard.defer_destroy(head);
+                                return Some(ManuallyDrop::into_inner(ptr::read(&(*h).data)));
+                            }
+                        }
+                    }
+                    None => return None,
+                }
+            }
+        }
+
+        pub fn is_empty(&self) -> bool {
+            let guard = epoch::pin();
+            self.head.load(Acquire, &guard).is_null()
+        }
+    }
+
+    impl<T> Drop for TreiberStack<T> {
+        fn drop(&mut self) {
+            while self.pop().is_some() {}
+        }
+    }
+
+    loom::model(|| {
+        let stack1 = Arc::new(TreiberStack::new());
+        let stack2 = Arc::clone(&stack1);
+
+        // use 5 since it's greater than the 4 used for the sanitize feature
+        let jh = spawn(move || {
+            for i in 0..5 {
+                stack2.push(i);
+                assert!(stack2.pop().is_some());
+            }
+        });
+
+        for i in 0..5 {
+            stack1.push(i);
+            assert!(stack1.pop().is_some());
+        }
+
+        jh.join().unwrap();
+        assert!(stack1.pop().is_none());
+        assert!(stack1.is_empty());
+    });
+}
diff --git a/crossbeam-utils/Cargo.toml b/crossbeam-utils/Cargo.toml
index 069043a31..ffe59158d 100644
--- a/crossbeam-utils/Cargo.toml
+++ b/crossbeam-utils/Cargo.toml
@@ -25,6 +25,9 @@ alloc = []
 cfg-if = "0.1.2"
 lazy_static = { version = "1.1.0", optional = true }
 
+[target.'cfg(loom)'.dependencies]
+loom = "0.3.2"
+
 [build-dependencies]
 autocfg = "1"
 
diff --git a/crossbeam-utils/src/atomic/atomic_cell.rs b/crossbeam-utils/src/atomic/atomic_cell.rs
index cf0658aad..67ad85f01 100644
--- a/crossbeam-utils/src/atomic/atomic_cell.rs
+++ b/crossbeam-utils/src/atomic/atomic_cell.rs
@@ -1,12 +1,15 @@
+use concurrency::sync::atomic::{self, AtomicBool, Ordering};
 use core::cell::UnsafeCell;
 use core::fmt;
 use core::mem;
+
+#[cfg(not(loom))]
 use core::ptr;
-use core::sync::atomic::{self, AtomicBool, Ordering};
 
 #[cfg(feature = "std")]
 use std::panic::{RefUnwindSafe, UnwindSafe};
 
+#[cfg(not(loom))]
 use super::seq_lock::SeqLock;
 
 /// A thread-safe mutable memory location.
@@ -523,23 +526,23 @@ macro_rules! impl_arithmetic {
 
 #[cfg(has_atomic_u8)]
 impl_arithmetic!(u8, atomic::AtomicU8, "let a = AtomicCell::new(7u8);");
-#[cfg(has_atomic_u8)]
+#[cfg(all(has_atomic_u8, not(loom)))]
 impl_arithmetic!(i8, atomic::AtomicI8, "let a = AtomicCell::new(7i8);");
 #[cfg(has_atomic_u16)]
 impl_arithmetic!(u16, atomic::AtomicU16, "let a = AtomicCell::new(7u16);");
-#[cfg(has_atomic_u16)]
+#[cfg(all(has_atomic_u16, not(loom)))]
 impl_arithmetic!(i16, atomic::AtomicI16, "let a = AtomicCell::new(7i16);");
 #[cfg(has_atomic_u32)]
 impl_arithmetic!(u32, atomic::AtomicU32, "let a = AtomicCell::new(7u32);");
-#[cfg(has_atomic_u32)]
+#[cfg(all(has_atomic_u32, not(loom)))]
 impl_arithmetic!(i32, atomic::AtomicI32, "let a = AtomicCell::new(7i32);");
 #[cfg(has_atomic_u64)]
 impl_arithmetic!(u64, atomic::AtomicU64, "let a = AtomicCell::new(7u64);");
-#[cfg(has_atomic_u64)]
+#[cfg(all(has_atomic_u64, not(loom)))]
 impl_arithmetic!(i64, atomic::AtomicI64, "let a = AtomicCell::new(7i64);");
-#[cfg(has_atomic_u128)]
+#[cfg(all(has_atomic_u128, not(loom)))]
 impl_arithmetic!(u128, atomic::AtomicU128, "let a = AtomicCell::new(7u128);");
-#[cfg(has_atomic_u128)]
+#[cfg(all(has_atomic_u128, not(loom)))]
 impl_arithmetic!(i128, atomic::AtomicI128, "let  a = AtomicCell::new(7i128);");
 
 impl_arithmetic!(
@@ -547,6 +550,7 @@ impl_arithmetic!(
     atomic::AtomicUsize,
     "let a = AtomicCell::new(7usize);"
 );
+#[cfg(not(loom))]
 impl_arithmetic!(
     isize,
     atomic::AtomicIsize,
@@ -648,6 +652,7 @@ fn can_transmute<A, B>() -> bool {
 /// scalability.
 #[inline]
 #[must_use]
+#[cfg(not(loom))]
 fn lock(addr: usize) -> &'static SeqLock {
     // The number of locks is a prime number because we want to make sure `addr % LEN` gets
     // dispersed across all locks.
@@ -741,6 +746,9 @@ macro_rules! atomic {
             #[cfg(has_atomic_u64)]
             atomic!(@check, $t, atomic::AtomicU64, $a, $atomic_op);
 
+            #[cfg(loom)]
+            unimplemented!("loom does not support non-atomic atomic ops");
+            #[cfg(not(loom))]
             break $fallback_op;
         }
     };
diff --git a/crossbeam-utils/src/atomic/consume.rs b/crossbeam-utils/src/atomic/consume.rs
index 9be5464fb..10ed628c1 100644
--- a/crossbeam-utils/src/atomic/consume.rs
+++ b/crossbeam-utils/src/atomic/consume.rs
@@ -1,6 +1,6 @@
 #[cfg(any(target_arch = "arm", target_arch = "aarch64"))]
-use core::sync::atomic::compiler_fence;
-use core::sync::atomic::Ordering;
+use concurrency::sync::atomic::compiler_fence;
+use concurrency::sync::atomic::Ordering;
 
 /// Trait which allows reading from primitive atomic types with "consume" ordering.
 pub trait AtomicConsume {
@@ -53,30 +53,42 @@ macro_rules! impl_atomic {
             type Val = $val;
             impl_consume!();
         }
+        #[cfg(loom)]
+        impl AtomicConsume for ::loom::sync::atomic::$atomic {
+            type Val = $val;
+            impl_consume!();
+        }
     };
 }
 
 impl_atomic!(AtomicBool, bool);
 impl_atomic!(AtomicUsize, usize);
+#[cfg(not(loom))]
 impl_atomic!(AtomicIsize, isize);
 #[cfg(all(feature = "nightly", target_has_atomic = "8"))]
 impl_atomic!(AtomicU8, u8);
-#[cfg(all(feature = "nightly", target_has_atomic = "8"))]
+#[cfg(all(feature = "nightly", target_has_atomic = "8", not(loom)))]
 impl_atomic!(AtomicI8, i8);
 #[cfg(all(feature = "nightly", target_has_atomic = "16"))]
 impl_atomic!(AtomicU16, u16);
-#[cfg(all(feature = "nightly", target_has_atomic = "16"))]
+#[cfg(all(feature = "nightly", target_has_atomic = "16", not(loom)))]
 impl_atomic!(AtomicI16, i16);
 #[cfg(all(feature = "nightly", target_has_atomic = "32"))]
 impl_atomic!(AtomicU32, u32);
-#[cfg(all(feature = "nightly", target_has_atomic = "32"))]
+#[cfg(all(feature = "nightly", target_has_atomic = "32", not(loom)))]
 impl_atomic!(AtomicI32, i32);
 #[cfg(all(feature = "nightly", target_has_atomic = "64"))]
 impl_atomic!(AtomicU64, u64);
-#[cfg(all(feature = "nightly", target_has_atomic = "64"))]
+#[cfg(all(feature = "nightly", target_has_atomic = "64", not(loom)))]
 impl_atomic!(AtomicI64, i64);
 
 impl<T> AtomicConsume for ::core::sync::atomic::AtomicPtr<T> {
     type Val = *mut T;
     impl_consume!();
 }
+
+#[cfg(loom)]
+impl<T> AtomicConsume for ::loom::sync::atomic::AtomicPtr<T> {
+    type Val = *mut T;
+    impl_consume!();
+}
diff --git a/crossbeam-utils/src/atomic/mod.rs b/crossbeam-utils/src/atomic/mod.rs
index 074b0ca53..14af10c91 100644
--- a/crossbeam-utils/src/atomic/mod.rs
+++ b/crossbeam-utils/src/atomic/mod.rs
@@ -1,5 +1,6 @@
 //! Atomic types.
 
+#[cfg(not(loom))]
 cfg_if! {
     // Use "wide" sequence lock if the pointer width <= 32 for preventing its counter against wrap
     // around.
diff --git a/crossbeam-utils/src/backoff.rs b/crossbeam-utils/src/backoff.rs
index 446755bbc..305d20574 100644
--- a/crossbeam-utils/src/backoff.rs
+++ b/crossbeam-utils/src/backoff.rs
@@ -1,6 +1,6 @@
+use concurrency::sync::atomic;
 use core::cell::Cell;
 use core::fmt;
-use core::sync::atomic;
 
 const SPIN_LIMIT: u32 = 6;
 const YIELD_LIMIT: u32 = 10;
diff --git a/crossbeam-utils/src/lib.rs b/crossbeam-utils/src/lib.rs
index 06f23beb4..85baf5e5d 100644
--- a/crossbeam-utils/src/lib.rs
+++ b/crossbeam-utils/src/lib.rs
@@ -36,6 +36,52 @@ extern crate cfg_if;
 #[cfg(feature = "std")]
 extern crate core;
 
+#[cfg(loom)]
+extern crate loom;
+
+#[cfg(loom)]
+#[allow(unused_imports)]
+pub(crate) mod concurrency {
+    pub(crate) mod sync {
+        pub(crate) mod atomic {
+            pub(crate) use loom::sync::atomic::spin_loop_hint;
+            pub(crate) use loom::sync::atomic::{
+                AtomicBool, AtomicU16, AtomicU32, AtomicU64, AtomicU8, AtomicUsize, Ordering,
+            };
+
+            // FIXME: loom does not support compiler_fence at the moment.
+            // https://github.com/tokio-rs/loom/issues/117
+            // we use fence as a stand-in for compiler_fence for the time being.
+            // this may miss some races since fence is stronger than compiler_fence,
+            // but it's the best we can do for the time being.
+            pub(crate) use loom::sync::atomic::fence as compiler_fence;
+        }
+        pub(crate) use loom::sync::{Arc, Condvar, Mutex};
+    }
+}
+#[cfg(not(loom))]
+#[allow(unused_imports)]
+pub(crate) mod concurrency {
+    pub(crate) mod sync {
+        pub(crate) mod atomic {
+            pub(crate) use core::sync::atomic::compiler_fence;
+            pub(crate) use core::sync::atomic::spin_loop_hint;
+            pub(crate) use core::sync::atomic::{AtomicBool, AtomicIsize, AtomicUsize, Ordering};
+            #[cfg(has_atomic_u16)]
+            pub(crate) use core::sync::atomic::{AtomicI16, AtomicU16};
+            #[cfg(has_atomic_u32)]
+            pub(crate) use core::sync::atomic::{AtomicI32, AtomicU32};
+            #[cfg(has_atomic_u64)]
+            pub(crate) use core::sync::atomic::{AtomicI64, AtomicU64};
+            #[cfg(has_atomic_u8)]
+            pub(crate) use core::sync::atomic::{AtomicI8, AtomicU8};
+        }
+
+        #[cfg(feature = "std")]
+        pub(crate) use std::sync::{Arc, Condvar, Mutex};
+    }
+}
+
 cfg_if! {
     if #[cfg(feature = "alloc")] {
         extern crate alloc;
@@ -55,10 +101,13 @@ pub use backoff::Backoff;
 
 cfg_if! {
     if #[cfg(feature = "std")] {
+        #[cfg(not(loom))]
         #[macro_use]
         extern crate lazy_static;
 
         pub mod sync;
+
+        #[cfg(not(loom))]
         pub mod thread;
     }
 }
diff --git a/crossbeam-utils/src/sync/mod.rs b/crossbeam-utils/src/sync/mod.rs
index 363496372..007c3fb38 100644
--- a/crossbeam-utils/src/sync/mod.rs
+++ b/crossbeam-utils/src/sync/mod.rs
@@ -9,9 +9,11 @@
 //! [`WaitGroup`]: struct.WaitGroup.html
 
 mod parker;
+#[cfg(not(loom))]
 mod sharded_lock;
 mod wait_group;
 
 pub use self::parker::{Parker, Unparker};
+#[cfg(not(loom))]
 pub use self::sharded_lock::{ShardedLock, ShardedLockReadGuard, ShardedLockWriteGuard};
 pub use self::wait_group::WaitGroup;
diff --git a/crossbeam-utils/src/sync/parker.rs b/crossbeam-utils/src/sync/parker.rs
index 051afe512..d605adc81 100644
--- a/crossbeam-utils/src/sync/parker.rs
+++ b/crossbeam-utils/src/sync/parker.rs
@@ -1,8 +1,8 @@
+use concurrency::sync::atomic::AtomicUsize;
+use concurrency::sync::atomic::Ordering::SeqCst;
+use concurrency::sync::{Arc, Condvar, Mutex};
 use std::fmt;
 use std::marker::PhantomData;
-use std::sync::atomic::AtomicUsize;
-use std::sync::atomic::Ordering::SeqCst;
-use std::sync::{Arc, Condvar, Mutex};
 use std::time::Duration;
 
 /// A thread parking primitive.
diff --git a/crossbeam-utils/src/sync/wait_group.rs b/crossbeam-utils/src/sync/wait_group.rs
index 0527b3159..dc935cf7f 100644
--- a/crossbeam-utils/src/sync/wait_group.rs
+++ b/crossbeam-utils/src/sync/wait_group.rs
@@ -1,5 +1,5 @@
+use concurrency::sync::{Arc, Condvar, Mutex};
 use std::fmt;
-use std::sync::{Arc, Condvar, Mutex};
 
 /// Enables threads to synchronize the beginning or end of some computation.
 ///