diff --git a/embassy-stm32/Cargo.toml b/embassy-stm32/Cargo.toml
index 33583d22..a2c5da2b 100644
--- a/embassy-stm32/Cargo.toml
+++ b/embassy-stm32/Cargo.toml
@@ -58,7 +58,6 @@ rand_core = "0.6.3"
 sdio-host = "0.5.0"
 embedded-sdmmc = { git = "https://github.com/embassy-rs/embedded-sdmmc-rs", rev = "a4f293d3a6f72158385f79c98634cb8a14d0d2fc", optional = true }
 critical-section = "1.1"
-atomic-polyfill = "1.0.1"
 stm32-metapac = { git = "https://github.com/embassy-rs/stm32-data-generated", tag = "stm32-data-6bfa5a0dcec6a9bd42cea94ba11eeae1a17a7f2c" }
 vcell = "0.1.3"
 bxcan = "0.7.0"
diff --git a/embassy-stm32/build.rs b/embassy-stm32/build.rs
index 810da37e..fa10b7f7 100644
--- a/embassy-stm32/build.rs
+++ b/embassy-stm32/build.rs
@@ -8,6 +8,32 @@ use quote::{format_ident, quote};
 use stm32_metapac::metadata::{MemoryRegionKind, METADATA};
 
 fn main() {
+    let target = env::var("TARGET").unwrap();
+
+    if target.starts_with("thumbv6m-") {
+        println!("cargo:rustc-cfg=cortex_m");
+        println!("cargo:rustc-cfg=armv6m");
+    } else if target.starts_with("thumbv7m-") {
+        println!("cargo:rustc-cfg=cortex_m");
+        println!("cargo:rustc-cfg=armv7m");
+    } else if target.starts_with("thumbv7em-") {
+        println!("cargo:rustc-cfg=cortex_m");
+        println!("cargo:rustc-cfg=armv7m");
+        println!("cargo:rustc-cfg=armv7em"); // (not currently used)
+    } else if target.starts_with("thumbv8m.base") {
+        println!("cargo:rustc-cfg=cortex_m");
+        println!("cargo:rustc-cfg=armv8m");
+        println!("cargo:rustc-cfg=armv8m_base");
+    } else if target.starts_with("thumbv8m.main") {
+        println!("cargo:rustc-cfg=cortex_m");
+        println!("cargo:rustc-cfg=armv8m");
+        println!("cargo:rustc-cfg=armv8m_main");
+    }
+
+    if target.ends_with("-eabihf") {
+        println!("cargo:rustc-cfg=has_fpu");
+    }
+
     let chip_name = match env::vars()
         .map(|(a, _)| a)
         .filter(|x| x.starts_with("CARGO_FEATURE_STM32"))
@@ -434,20 +460,20 @@ fn main() {
                         unsafe { crate::rcc::get_freqs().#clk }
                     }
                     fn enable() {
-                        critical_section::with(|_| {
+                        critical_section::with(|_cs| {
                             #before_enable
                             #[cfg(feature = "low-power")]
-                            crate::rcc::clock_refcount_add();
+                            crate::rcc::clock_refcount_add(_cs);
                             crate::pac::RCC.#en_reg().modify(|w| w.#set_en_field(true));
                             #after_enable
                         })
                     }
                     fn disable() {
-                        critical_section::with(|_| {
+                        critical_section::with(|_cs| {
                             #before_disable
                             crate::pac::RCC.#en_reg().modify(|w| w.#set_en_field(false));
                             #[cfg(feature = "low-power")]
-                            crate::rcc::clock_refcount_sub();
+                            crate::rcc::clock_refcount_sub(_cs);
                         })
                     }
                     fn reset() {
diff --git a/embassy-stm32/src/dma/bdma.rs b/embassy-stm32/src/dma/bdma.rs
index 20ff29be..62eb65b1 100644
--- a/embassy-stm32/src/dma/bdma.rs
+++ b/embassy-stm32/src/dma/bdma.rs
@@ -2,10 +2,9 @@
 
 use core::future::Future;
 use core::pin::Pin;
-use core::sync::atomic::{fence, Ordering};
+use core::sync::atomic::{fence, AtomicUsize, Ordering};
 use core::task::{Context, Poll, Waker};
 
-use atomic_polyfill::AtomicUsize;
 use embassy_hal_internal::{into_ref, Peripheral, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 
@@ -127,7 +126,13 @@ pub(crate) unsafe fn on_irq_inner(dma: pac::bdma::Dma, channel_num: usize, index
     } else if isr.tcif(channel_num) && cr.read().tcie() {
         // Acknowledge transfer complete interrupt
         dma.ifcr().write(|w| w.set_tcif(channel_num, true));
+        #[cfg(not(armv6m))]
         STATE.complete_count[index].fetch_add(1, Ordering::Release);
+        #[cfg(armv6m)]
+        critical_section::with(|_| {
+            let x = STATE.complete_count[index].load(Ordering::Relaxed);
+            STATE.complete_count[index].store(x + 1, Ordering::Release);
+        })
     } else {
         return;
     }
@@ -391,7 +396,14 @@ impl<'a, C: Channel> DmaCtrl for DmaCtrlImpl<'a, C> {
     }
 
     fn reset_complete_count(&mut self) -> usize {
-        STATE.complete_count[self.0.index()].swap(0, Ordering::AcqRel)
+        #[cfg(not(armv6m))]
+        return STATE.complete_count[self.0.index()].swap(0, Ordering::AcqRel);
+        #[cfg(armv6m)]
+        return critical_section::with(|_| {
+            let x = STATE.complete_count[self.0.index()].load(Ordering::Acquire);
+            STATE.complete_count[self.0.index()].store(0, Ordering::Release);
+            x
+        });
     }
 
     fn set_waker(&mut self, waker: &Waker) {
diff --git a/embassy-stm32/src/lib.rs b/embassy-stm32/src/lib.rs
index e883678b..37c94f82 100644
--- a/embassy-stm32/src/lib.rs
+++ b/embassy-stm32/src/lib.rs
@@ -225,7 +225,9 @@ pub fn init(config: Config) -> Peripherals {
 
         #[cfg(feature = "low-power")]
         while !crate::rcc::low_power_ready() {
-            crate::rcc::clock_refcount_sub();
+            critical_section::with(|cs| {
+                crate::rcc::clock_refcount_sub(cs);
+            });
         }
     }
 
diff --git a/embassy-stm32/src/low_power.rs b/embassy-stm32/src/low_power.rs
index bb714b8c..861a59d7 100644
--- a/embassy-stm32/src/low_power.rs
+++ b/embassy-stm32/src/low_power.rs
@@ -1,7 +1,7 @@
 use core::arch::asm;
 use core::marker::PhantomData;
+use core::sync::atomic::{compiler_fence, Ordering};
 
-use atomic_polyfill::{compiler_fence, Ordering};
 use cortex_m::peripheral::SCB;
 use embassy_executor::*;
 
diff --git a/embassy-stm32/src/rcc/mod.rs b/embassy-stm32/src/rcc/mod.rs
index a9e53d42..abb53fd1 100644
--- a/embassy-stm32/src/rcc/mod.rs
+++ b/embassy-stm32/src/rcc/mod.rs
@@ -27,9 +27,10 @@ pub use mco::*;
 #[cfg_attr(rcc_wba, path = "wba.rs")]
 #[cfg_attr(any(rcc_wl5, rcc_wle), path = "wl.rs")]
 mod _version;
-pub use _version::*;
 #[cfg(feature = "low-power")]
-use atomic_polyfill::{AtomicU32, Ordering};
+use core::sync::atomic::{AtomicU32, Ordering};
+
+pub use _version::*;
 
 //  Model Clock Configuration
 //
@@ -145,14 +146,17 @@ pub fn low_power_ready() -> bool {
 }
 
 #[cfg(feature = "low-power")]
-pub(crate) fn clock_refcount_add() {
+pub(crate) fn clock_refcount_add(_cs: critical_section::CriticalSection) {
     // We don't check for overflow because constructing more than u32 peripherals is unlikely
-    CLOCK_REFCOUNT.fetch_add(1, Ordering::Relaxed);
+    let n = CLOCK_REFCOUNT.load(Ordering::Relaxed);
+    CLOCK_REFCOUNT.store(n + 1, Ordering::Relaxed);
 }
 
 #[cfg(feature = "low-power")]
-pub(crate) fn clock_refcount_sub() {
-    assert!(CLOCK_REFCOUNT.fetch_sub(1, Ordering::Relaxed) != 0);
+pub(crate) fn clock_refcount_sub(_cs: critical_section::CriticalSection) {
+    let n = CLOCK_REFCOUNT.load(Ordering::Relaxed);
+    assert!(n != 0);
+    CLOCK_REFCOUNT.store(n - 1, Ordering::Relaxed);
 }
 
 /// Frozen clock frequencies
diff --git a/embassy-stm32/src/time_driver.rs b/embassy-stm32/src/time_driver.rs
index 91750241..e88198e6 100644
--- a/embassy-stm32/src/time_driver.rs
+++ b/embassy-stm32/src/time_driver.rs
@@ -1,9 +1,8 @@
 use core::cell::Cell;
 use core::convert::TryInto;
-use core::sync::atomic::{compiler_fence, Ordering};
+use core::sync::atomic::{compiler_fence, AtomicU32, AtomicU8, Ordering};
 use core::{mem, ptr};
 
-use atomic_polyfill::{AtomicU32, AtomicU8};
 use critical_section::CriticalSection;
 use embassy_sync::blocking_mutex::raw::CriticalSectionRawMutex;
 use embassy_sync::blocking_mutex::Mutex;
@@ -229,7 +228,9 @@ impl RtcDriver {
     fn next_period(&self) {
         let r = T::regs_gp16();
 
-        let period = self.period.fetch_add(1, Ordering::Relaxed) + 1;
+        // We only modify the period from the timer interrupt, so we know this can't race.
+        let period = self.period.load(Ordering::Relaxed) + 1;
+        self.period.store(period, Ordering::Relaxed);
         let t = (period as u64) << 15;
 
         critical_section::with(move |cs| {
@@ -403,18 +404,15 @@ impl Driver for RtcDriver {
     }
 
     unsafe fn allocate_alarm(&self) -> Option<AlarmHandle> {
-        let id = self.alarm_count.fetch_update(Ordering::AcqRel, Ordering::Acquire, |x| {
-            if x < ALARM_COUNT as u8 {
-                Some(x + 1)
+        critical_section::with(|_| {
+            let id = self.alarm_count.load(Ordering::Relaxed);
+            if id < ALARM_COUNT as u8 {
+                self.alarm_count.store(id + 1, Ordering::Relaxed);
+                Some(AlarmHandle::new(id))
             } else {
                 None
             }
-        });
-
-        match id {
-            Ok(id) => Some(AlarmHandle::new(id)),
-            Err(_) => None,
-        }
+        })
     }
 
     fn set_alarm_callback(&self, alarm: AlarmHandle, callback: fn(*mut ()), ctx: *mut ()) {
diff --git a/embassy-time/Cargo.toml b/embassy-time/Cargo.toml
index 8f034a9d..e4b88d78 100644
--- a/embassy-time/Cargo.toml
+++ b/embassy-time/Cargo.toml
@@ -218,7 +218,6 @@ embedded-hal-1 = { package = "embedded-hal", version = "=1.0.0-rc.1", optional =
 embedded-hal-async = { version = "=1.0.0-rc.1", optional = true}
 
 futures-util = { version = "0.3.17", default-features = false }
-atomic-polyfill = "1.0.1"
 critical-section = "1.1"
 cfg-if = "1.0.0"
 heapless = "0.7"