diff --git a/embassy-stm32/src/spi/mod.rs b/embassy-stm32/src/spi/mod.rs
index 80a05aac6..5d919f923 100644
--- a/embassy-stm32/src/spi/mod.rs
+++ b/embassy-stm32/src/spi/mod.rs
@@ -1,15 +1,19 @@
 #![macro_use]
 
 use crate::dma;
+use crate::dma::NoDma;
 use crate::gpio::sealed::{AFType, Pin};
 use crate::gpio::{AnyPin, NoPin, OptionalPin};
-use crate::pac::spi::vals;
+use crate::pac::spi::{regs, vals};
 use crate::peripherals;
 use crate::rcc::RccPeripheral;
 use crate::time::Hertz;
+use core::future::Future;
 use core::marker::PhantomData;
+use core::ptr;
 use embassy::util::Unborrow;
 use embassy_hal_common::unborrow;
+use embassy_traits::spi as traits;
 
 #[cfg_attr(spi_v1, path = "v1.rs")]
 #[cfg_attr(spi_f1, path = "v1.rs")]
@@ -374,6 +378,174 @@ impl RegsExt for crate::pac::spi::Spi {
     }
 }
 
+fn check_error_flags(sr: regs::Sr) -> Result<(), Error> {
+    if sr.ovr() {
+        return Err(Error::Overrun);
+    }
+    #[cfg(not(any(spi_f1, spi_v3)))]
+    if sr.fre() {
+        return Err(Error::Framing);
+    }
+    #[cfg(spi_v3)]
+    if sr.tifre() {
+        return Err(Error::Framing);
+    }
+    if sr.modf() {
+        return Err(Error::ModeFault);
+    }
+    #[cfg(not(spi_v3))]
+    if sr.crcerr() {
+        return Err(Error::Crc);
+    }
+    #[cfg(spi_v3)]
+    if sr.crce() {
+        return Err(Error::Crc);
+    }
+
+    Ok(())
+}
+
+fn spin_until_tx_ready(regs: &'static crate::pac::spi::Spi) -> Result<(), Error> {
+    loop {
+        let sr = unsafe { regs.sr().read() };
+
+        check_error_flags(sr)?;
+
+        #[cfg(not(spi_v3))]
+        if sr.txe() {
+            return Ok(());
+        }
+        #[cfg(spi_v3)]
+        if sr.txp() {
+            return Ok(());
+        }
+    }
+}
+
+fn spin_until_rx_ready(regs: &'static crate::pac::spi::Spi) -> Result<(), Error> {
+    loop {
+        let sr = unsafe { regs.sr().read() };
+
+        check_error_flags(sr)?;
+
+        #[cfg(not(spi_v3))]
+        if sr.rxne() {
+            return Ok(());
+        }
+        #[cfg(spi_v3)]
+        if sr.rxp() {
+            return Ok(());
+        }
+    }
+}
+
+trait Word {
+    const WORDSIZE: WordSize;
+}
+
+impl Word for u8 {
+    const WORDSIZE: WordSize = WordSize::EightBit;
+}
+impl Word for u16 {
+    const WORDSIZE: WordSize = WordSize::SixteenBit;
+}
+
+fn transfer_word<W: Word>(regs: &'static crate::pac::spi::Spi, tx_word: W) -> Result<W, Error> {
+    spin_until_tx_ready(regs)?;
+
+    unsafe {
+        ptr::write_volatile(regs.tx_ptr(), tx_word);
+
+        #[cfg(spi_v3)]
+        regs.cr1().modify(|reg| reg.set_cstart(true));
+    }
+
+    spin_until_rx_ready(regs)?;
+
+    let rx_word = unsafe { ptr::read_volatile(regs.rx_ptr()) };
+    return Ok(rx_word);
+}
+
+// Note: It is not possible to impl these traits generically in embedded-hal 0.2 due to a conflict with
+// some marker traits. For details, see https://github.com/rust-embedded/embedded-hal/pull/289
+macro_rules! impl_blocking {
+    ($w:ident) => {
+        impl<'d, T: Instance> embedded_hal::blocking::spi::Write<$w> for Spi<'d, T, NoDma, NoDma> {
+            type Error = Error;
+
+            fn write(&mut self, words: &[$w]) -> Result<(), Self::Error> {
+                self.set_word_size($w::WORDSIZE);
+                let regs = T::regs();
+
+                for word in words.iter() {
+                    let _ = transfer_word(regs, *word)?;
+                }
+
+                Ok(())
+            }
+        }
+
+        impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<$w>
+            for Spi<'d, T, NoDma, NoDma>
+        {
+            type Error = Error;
+
+            fn transfer<'w>(&mut self, words: &'w mut [$w]) -> Result<&'w [$w], Self::Error> {
+                self.set_word_size($w::WORDSIZE);
+                let regs = T::regs();
+
+                for word in words.iter_mut() {
+                    *word = transfer_word(regs, *word)?;
+                }
+
+                Ok(words)
+            }
+        }
+    };
+}
+
+impl_blocking!(u8);
+impl_blocking!(u16);
+
+impl<'d, T: Instance, Tx, Rx> traits::Spi<u8> for Spi<'d, T, Tx, Rx> {
+    type Error = Error;
+}
+
+impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx> traits::Write<u8> for Spi<'d, T, Tx, Rx> {
+    #[rustfmt::skip]
+    type WriteFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
+
+    fn write<'a>(&'a mut self, data: &'a [u8]) -> Self::WriteFuture<'a> {
+        self.write_dma_u8(data)
+    }
+}
+
+impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::Read<u8>
+    for Spi<'d, T, Tx, Rx>
+{
+    #[rustfmt::skip]
+    type ReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
+
+    fn read<'a>(&'a mut self, data: &'a mut [u8]) -> Self::ReadFuture<'a> {
+        self.read_dma_u8(data)
+    }
+}
+
+impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::FullDuplex<u8>
+    for Spi<'d, T, Tx, Rx>
+{
+    #[rustfmt::skip]
+    type WriteReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
+
+    fn read_write<'a>(
+        &'a mut self,
+        read: &'a mut [u8],
+        write: &'a [u8],
+    ) -> Self::WriteReadFuture<'a> {
+        self.read_write_dma_u8(read, write)
+    }
+}
+
 pub(crate) mod sealed {
     use super::*;
 
diff --git a/embassy-stm32/src/spi/v1.rs b/embassy-stm32/src/spi/v1.rs
index 255bd950a..92449ea80 100644
--- a/embassy-stm32/src/spi/v1.rs
+++ b/embassy-stm32/src/spi/v1.rs
@@ -1,19 +1,13 @@
 #![macro_use]
 
-use crate::dma::NoDma;
-use crate::spi::{Error, Instance, RegsExt, RxDmaChannel, TxDmaChannel, WordSize};
-use core::future::Future;
-use core::ptr;
-use embassy_traits::spi as traits;
 pub use embedded_hal::blocking;
 pub use embedded_hal::spi::{Mode, Phase, Polarity, MODE_0, MODE_1, MODE_2, MODE_3};
 use futures::future::join3;
 
-use super::Spi;
+use super::*;
 
 impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
-    #[allow(unused)]
-    async fn write_dma_u8(&mut self, write: &[u8]) -> Result<(), Error>
+    pub(super) async fn write_dma_u8(&mut self, write: &[u8]) -> Result<(), Error>
     where
         Tx: TxDmaChannel<T>,
     {
@@ -41,8 +35,7 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         Ok(())
     }
 
-    #[allow(unused)]
-    async fn read_dma_u8(&mut self, read: &mut [u8]) -> Result<(), Error>
+    pub(super) async fn read_dma_u8(&mut self, read: &mut [u8]) -> Result<(), Error>
     where
         Tx: TxDmaChannel<T>,
         Rx: RxDmaChannel<T>,
@@ -94,8 +87,11 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         Ok(())
     }
 
-    #[allow(unused)]
-    async fn read_write_dma_u8(&mut self, read: &mut [u8], write: &[u8]) -> Result<(), Error>
+    pub(super) async fn read_write_dma_u8(
+        &mut self,
+        read: &mut [u8],
+        write: &[u8],
+    ) -> Result<(), Error>
     where
         Tx: TxDmaChannel<T>,
         Rx: RxDmaChannel<T>,
@@ -154,161 +150,3 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         }
     }
 }
-
-impl<'d, T: Instance> embedded_hal::blocking::spi::Write<u8> for Spi<'d, T, NoDma, NoDma> {
-    type Error = Error;
-
-    fn write(&mut self, words: &[u8]) -> Result<(), Self::Error> {
-        self.set_word_size(WordSize::EightBit);
-        let regs = T::regs();
-
-        for word in words.iter() {
-            write_word(regs, *word)?;
-            let _: u8 = read_word(regs)?;
-        }
-
-        Ok(())
-    }
-}
-
-impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u8> for Spi<'d, T, NoDma, NoDma> {
-    type Error = Error;
-
-    fn transfer<'w>(&mut self, words: &'w mut [u8]) -> Result<&'w [u8], Self::Error> {
-        self.set_word_size(WordSize::EightBit);
-        let regs = T::regs();
-
-        for word in words.iter_mut() {
-            write_word(regs, *word)?;
-            *word = read_word(regs)?;
-        }
-
-        Ok(words)
-    }
-}
-
-impl<'d, T: Instance> embedded_hal::blocking::spi::Write<u16> for Spi<'d, T, NoDma, NoDma> {
-    type Error = Error;
-
-    fn write(&mut self, words: &[u16]) -> Result<(), Self::Error> {
-        self.set_word_size(WordSize::SixteenBit);
-        let regs = T::regs();
-
-        for word in words.iter() {
-            write_word(regs, *word)?;
-            let _: u8 = read_word(regs)?;
-        }
-
-        Ok(())
-    }
-}
-
-impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u16> for Spi<'d, T, NoDma, NoDma> {
-    type Error = Error;
-
-    fn transfer<'w>(&mut self, words: &'w mut [u16]) -> Result<&'w [u16], Self::Error> {
-        self.set_word_size(WordSize::SixteenBit);
-        let regs = T::regs();
-
-        for word in words.iter_mut() {
-            write_word(regs, *word)?;
-            *word = read_word(regs)?;
-        }
-
-        Ok(words)
-    }
-}
-
-impl<'d, T: Instance, Tx, Rx> traits::Spi<u8> for Spi<'d, T, Tx, Rx> {
-    type Error = super::Error;
-}
-
-impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx> traits::Write<u8> for Spi<'d, T, Tx, Rx> {
-    #[rustfmt::skip]
-    type WriteFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
-
-    fn write<'a>(&'a mut self, data: &'a [u8]) -> Self::WriteFuture<'a> {
-        self.write_dma_u8(data)
-    }
-}
-
-impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::Read<u8>
-    for Spi<'d, T, Tx, Rx>
-{
-    #[rustfmt::skip]
-    type ReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
-
-    fn read<'a>(&'a mut self, data: &'a mut [u8]) -> Self::ReadFuture<'a> {
-        self.read_dma_u8(data)
-    }
-}
-
-impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::FullDuplex<u8>
-    for Spi<'d, T, Tx, Rx>
-{
-    #[rustfmt::skip]
-    type WriteReadFuture<'a> where Self: 'a = impl Future<Output=Result<(), Self::Error>> + 'a;
-
-    fn read_write<'a>(
-        &'a mut self,
-        read: &'a mut [u8],
-        write: &'a [u8],
-    ) -> Self::WriteReadFuture<'a> {
-        self.read_write_dma_u8(read, write)
-    }
-}
-
-trait Word {}
-
-impl Word for u8 {}
-impl Word for u16 {}
-
-fn write_word<W: Word>(regs: &'static crate::pac::spi::Spi, word: W) -> Result<(), Error> {
-    loop {
-        let sr = unsafe { regs.sr().read() };
-        if sr.ovr() {
-            return Err(Error::Overrun);
-        }
-        #[cfg(not(spi_f1))]
-        if sr.fre() {
-            return Err(Error::Framing);
-        }
-        if sr.modf() {
-            return Err(Error::ModeFault);
-        }
-        if sr.crcerr() {
-            return Err(Error::Crc);
-        }
-        if sr.txe() {
-            unsafe {
-                ptr::write_volatile(regs.tx_ptr(), word);
-            }
-            return Ok(());
-        }
-    }
-}
-
-/// Read a single word blocking. Assumes word size have already been set.
-fn read_word<W: Word>(regs: &'static crate::pac::spi::Spi) -> Result<W, Error> {
-    loop {
-        let sr = unsafe { regs.sr().read() };
-        if sr.ovr() {
-            return Err(Error::Overrun);
-        }
-        #[cfg(not(spi_f1))]
-        if sr.fre() {
-            return Err(Error::Framing);
-        }
-        if sr.modf() {
-            return Err(Error::ModeFault);
-        }
-        if sr.crcerr() {
-            return Err(Error::Crc);
-        }
-        if sr.rxne() {
-            unsafe {
-                return Ok(ptr::read_volatile(regs.rx_ptr()));
-            }
-        }
-    }
-}
diff --git a/embassy-stm32/src/spi/v2.rs b/embassy-stm32/src/spi/v2.rs
index b1fae4bdc..de78676c2 100644
--- a/embassy-stm32/src/spi/v2.rs
+++ b/embassy-stm32/src/spi/v2.rs
@@ -1,18 +1,12 @@
 #![macro_use]
 
-use crate::dma::NoDma;
-use crate::spi::{Error, Instance, RegsExt, RxDmaChannel, TxDmaChannel, WordSize};
-use core::future::Future;
-use core::ptr;
-use embassy_traits::spi as traits;
 pub use embedded_hal::spi::{Mode, Phase, Polarity, MODE_0, MODE_1, MODE_2, MODE_3};
 use futures::future::{join, join3};
 
-use super::Spi;
+use super::*;
 
 impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
-    #[allow(unused)]
-    async fn write_dma_u8(&mut self, write: &[u8]) -> Result<(), Error>
+    pub(super) async fn write_dma_u8(&mut self, write: &[u8]) -> Result<(), Error>
     where
         Tx: TxDmaChannel<T>,
     {
@@ -49,8 +43,7 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         Ok(())
     }
 
-    #[allow(unused)]
-    async fn read_dma_u8(&mut self, read: &mut [u8]) -> Result<(), Error>
+    pub(super) async fn read_dma_u8(&mut self, read: &mut [u8]) -> Result<(), Error>
     where
         Tx: TxDmaChannel<T>,
         Rx: RxDmaChannel<T>,
@@ -102,8 +95,11 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         Ok(())
     }
 
-    #[allow(unused)]
-    async fn read_write_dma_u8(&mut self, read: &mut [u8], write: &[u8]) -> Result<(), Error>
+    pub(super) async fn read_write_dma_u8(
+        &mut self,
+        read: &mut [u8],
+        write: &[u8],
+    ) -> Result<(), Error>
     where
         Tx: TxDmaChannel<T>,
         Rx: RxDmaChannel<T>,
@@ -173,152 +169,3 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         }
     }
 }
-
-trait Word {}
-
-impl Word for u8 {}
-impl Word for u16 {}
-
-/// Write a single word blocking. Assumes word size have already been set.
-fn write_word<W: Word>(regs: &'static crate::pac::spi::Spi, word: W) -> Result<(), Error> {
-    loop {
-        let sr = unsafe { regs.sr().read() };
-        if sr.ovr() {
-            return Err(Error::Overrun);
-        } else if sr.fre() {
-            return Err(Error::Framing);
-        } else if sr.modf() {
-            return Err(Error::ModeFault);
-        } else if sr.crcerr() {
-            return Err(Error::Crc);
-        } else if sr.txe() {
-            unsafe {
-                ptr::write_volatile(regs.tx_ptr(), word);
-            }
-            return Ok(());
-        }
-    }
-}
-
-/// Read a single word blocking. Assumes word size have already been set.
-fn read_word<W: Word>(regs: &'static crate::pac::spi::Spi) -> Result<W, Error> {
-    loop {
-        let sr = unsafe { regs.sr().read() };
-        if sr.ovr() {
-            return Err(Error::Overrun);
-        } else if sr.modf() {
-            return Err(Error::ModeFault);
-        } else if sr.fre() {
-            return Err(Error::Framing);
-        } else if sr.crcerr() {
-            return Err(Error::Crc);
-        } else if sr.rxne() {
-            unsafe {
-                return Ok(ptr::read_volatile(regs.rx_ptr()));
-            }
-        }
-    }
-}
-
-impl<'d, T: Instance, Rx> embedded_hal::blocking::spi::Write<u8> for Spi<'d, T, NoDma, Rx> {
-    type Error = Error;
-
-    fn write(&mut self, words: &[u8]) -> Result<(), Self::Error> {
-        self.set_word_size(WordSize::EightBit);
-        let regs = T::regs();
-
-        for word in words.iter() {
-            write_word(regs, *word)?;
-            let _: u8 = read_word(regs)?;
-        }
-
-        Ok(())
-    }
-}
-
-impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u8> for Spi<'d, T, NoDma, NoDma> {
-    type Error = Error;
-
-    fn transfer<'w>(&mut self, words: &'w mut [u8]) -> Result<&'w [u8], Self::Error> {
-        self.set_word_size(WordSize::EightBit);
-        let regs = T::regs();
-
-        for word in words.iter_mut() {
-            write_word(regs, *word)?;
-            *word = read_word(regs)?;
-        }
-
-        Ok(words)
-    }
-}
-
-impl<'d, T: Instance, Rx> embedded_hal::blocking::spi::Write<u16> for Spi<'d, T, NoDma, Rx> {
-    type Error = Error;
-
-    fn write(&mut self, words: &[u16]) -> Result<(), Self::Error> {
-        self.set_word_size(WordSize::SixteenBit);
-        let regs = T::regs();
-
-        for word in words.iter() {
-            write_word(regs, *word)?;
-            let _: u16 = read_word(regs)?;
-        }
-
-        Ok(())
-    }
-}
-
-impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u16> for Spi<'d, T, NoDma, NoDma> {
-    type Error = Error;
-
-    fn transfer<'w>(&mut self, words: &'w mut [u16]) -> Result<&'w [u16], Self::Error> {
-        self.set_word_size(WordSize::SixteenBit);
-        let regs = T::regs();
-
-        for word in words.iter_mut() {
-            write_word(regs, *word)?;
-            *word = read_word(regs)?;
-        }
-
-        Ok(words)
-    }
-}
-
-impl<'d, T: Instance, Tx, Rx> traits::Spi<u8> for Spi<'d, T, Tx, Rx> {
-    type Error = super::Error;
-}
-
-impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx> traits::Write<u8> for Spi<'d, T, Tx, Rx> {
-    #[rustfmt::skip]
-    type WriteFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
-
-    fn write<'a>(&'a mut self, data: &'a [u8]) -> Self::WriteFuture<'a> {
-        self.write_dma_u8(data)
-    }
-}
-
-impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::Read<u8>
-    for Spi<'d, T, Tx, Rx>
-{
-    #[rustfmt::skip]
-    type ReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
-
-    fn read<'a>(&'a mut self, data: &'a mut [u8]) -> Self::ReadFuture<'a> {
-        self.read_dma_u8(data)
-    }
-}
-
-impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::FullDuplex<u8>
-    for Spi<'d, T, Tx, Rx>
-{
-    #[rustfmt::skip]
-    type WriteReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
-
-    fn read_write<'a>(
-        &'a mut self,
-        read: &'a mut [u8],
-        write: &'a [u8],
-    ) -> Self::WriteReadFuture<'a> {
-        self.read_write_dma_u8(read, write)
-    }
-}
diff --git a/embassy-stm32/src/spi/v3.rs b/embassy-stm32/src/spi/v3.rs
index 052924db0..ef0f23063 100644
--- a/embassy-stm32/src/spi/v3.rs
+++ b/embassy-stm32/src/spi/v3.rs
@@ -1,19 +1,12 @@
 #![macro_use]
 
-use crate::dma::NoDma;
-use crate::spi::{Error, Instance, RegsExt, RxDmaChannel, TxDmaChannel, WordSize};
-use core::future::Future;
-use core::ptr;
-use embassy_traits::spi as traits;
 pub use embedded_hal::spi::{Mode, Phase, Polarity, MODE_0, MODE_1, MODE_2, MODE_3};
-
 use futures::future::join3;
 
-use super::Spi;
+use super::*;
 
 impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
-    #[allow(unused)]
-    async fn write_dma_u8(&mut self, write: &[u8]) -> Result<(), Error>
+    pub(super) async fn write_dma_u8(&mut self, write: &[u8]) -> Result<(), Error>
     where
         Tx: TxDmaChannel<T>,
     {
@@ -53,8 +46,7 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         Ok(())
     }
 
-    #[allow(unused)]
-    async fn read_dma_u8(&mut self, read: &mut [u8]) -> Result<(), Error>
+    pub(super) async fn read_dma_u8(&mut self, read: &mut [u8]) -> Result<(), Error>
     where
         Tx: TxDmaChannel<T>,
         Rx: RxDmaChannel<T>,
@@ -107,8 +99,11 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         Ok(())
     }
 
-    #[allow(unused)]
-    async fn read_write_dma_u8(&mut self, read: &mut [u8], write: &[u8]) -> Result<(), Error>
+    pub(super) async fn read_write_dma_u8(
+        &mut self,
+        read: &mut [u8],
+        write: &[u8],
+    ) -> Result<(), Error>
     where
         Tx: TxDmaChannel<T>,
         Rx: RxDmaChannel<T>,
@@ -176,246 +171,3 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         }
     }
 }
-
-impl<'d, T: Instance> embedded_hal::blocking::spi::Write<u8> for Spi<'d, T, NoDma, NoDma> {
-    type Error = Error;
-
-    fn write(&mut self, words: &[u8]) -> Result<(), Self::Error> {
-        self.set_word_size(WordSize::EightBit);
-        let regs = T::regs();
-
-        for word in words.iter() {
-            while unsafe { !regs.sr().read().txp() } {
-                // spin
-            }
-            unsafe {
-                ptr::write_volatile(regs.tx_ptr(), *word);
-                regs.cr1().modify(|reg| reg.set_cstart(true));
-            }
-            loop {
-                let sr = unsafe { regs.sr().read() };
-                if sr.tifre() {
-                    return Err(Error::Framing);
-                }
-                if sr.ovr() {
-                    return Err(Error::Overrun);
-                }
-                if sr.crce() {
-                    return Err(Error::Crc);
-                }
-                if !sr.txp() {
-                    // loop waiting for TXE
-                    continue;
-                }
-                break;
-            }
-            unsafe {
-                // discard read to prevent pverrun.
-                let _: u8 = ptr::read_volatile(T::regs().rx_ptr());
-            }
-        }
-
-        while unsafe { !regs.sr().read().txc() } {
-            // spin
-        }
-
-        Ok(())
-    }
-}
-
-impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u8> for Spi<'d, T, NoDma, NoDma> {
-    type Error = Error;
-
-    fn transfer<'w>(&mut self, words: &'w mut [u8]) -> Result<&'w [u8], Self::Error> {
-        self.set_word_size(WordSize::EightBit);
-        let regs = T::regs();
-
-        for word in words.iter_mut() {
-            unsafe {
-                regs.cr1().modify(|reg| {
-                    reg.set_ssi(false);
-                });
-            }
-            while unsafe { !regs.sr().read().txp() } {
-                // spin
-            }
-            unsafe {
-                ptr::write_volatile(T::regs().tx_ptr(), *word);
-                regs.cr1().modify(|reg| reg.set_cstart(true));
-            }
-            loop {
-                let sr = unsafe { regs.sr().read() };
-
-                if sr.rxp() {
-                    break;
-                }
-                if sr.tifre() {
-                    return Err(Error::Framing);
-                }
-                if sr.ovr() {
-                    return Err(Error::Overrun);
-                }
-                if sr.crce() {
-                    return Err(Error::Crc);
-                }
-            }
-            unsafe {
-                *word = ptr::read_volatile(T::regs().rx_ptr());
-            }
-            let sr = unsafe { regs.sr().read() };
-            if sr.tifre() {
-                return Err(Error::Framing);
-            }
-            if sr.ovr() {
-                return Err(Error::Overrun);
-            }
-            if sr.crce() {
-                return Err(Error::Crc);
-            }
-        }
-
-        Ok(words)
-    }
-}
-
-impl<'d, T: Instance> embedded_hal::blocking::spi::Write<u16> for Spi<'d, T, NoDma, NoDma> {
-    type Error = Error;
-
-    fn write(&mut self, words: &[u16]) -> Result<(), Self::Error> {
-        self.set_word_size(WordSize::SixteenBit);
-        let regs = T::regs();
-
-        for word in words.iter() {
-            while unsafe { !regs.sr().read().txp() } {
-                // spin
-            }
-            unsafe {
-                let txdr = regs.txdr().ptr() as *mut u16;
-                ptr::write_volatile(txdr, *word);
-                regs.cr1().modify(|reg| reg.set_cstart(true));
-            }
-            loop {
-                let sr = unsafe { regs.sr().read() };
-                if sr.tifre() {
-                    return Err(Error::Framing);
-                }
-                if sr.ovr() {
-                    return Err(Error::Overrun);
-                }
-                if sr.crce() {
-                    return Err(Error::Crc);
-                }
-                if !sr.txp() {
-                    // loop waiting for TXE
-                    continue;
-                }
-                break;
-            }
-
-            unsafe {
-                let rxdr = regs.rxdr().ptr() as *const u8;
-                // discard read to prevent pverrun.
-                let _ = ptr::read_volatile(rxdr);
-            }
-        }
-
-        while unsafe { !regs.sr().read().txc() } {
-            // spin
-        }
-
-        Ok(())
-    }
-}
-
-impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u16> for Spi<'d, T, NoDma, NoDma> {
-    type Error = Error;
-
-    fn transfer<'w>(&mut self, words: &'w mut [u16]) -> Result<&'w [u16], Self::Error> {
-        self.set_word_size(WordSize::SixteenBit);
-        let regs = T::regs();
-
-        for word in words.iter_mut() {
-            while unsafe { !regs.sr().read().txp() } {
-                // spin
-            }
-            unsafe {
-                let txdr = regs.txdr().ptr() as *mut u16;
-                ptr::write_volatile(txdr, *word);
-                regs.cr1().modify(|reg| reg.set_cstart(true));
-            }
-
-            loop {
-                let sr = unsafe { regs.sr().read() };
-
-                if sr.rxp() {
-                    break;
-                }
-                if sr.tifre() {
-                    return Err(Error::Framing);
-                }
-                if sr.ovr() {
-                    return Err(Error::Overrun);
-                }
-                if sr.crce() {
-                    return Err(Error::Crc);
-                }
-            }
-
-            unsafe {
-                let rxdr = regs.rxdr().ptr() as *const u16;
-                *word = ptr::read_volatile(rxdr);
-            }
-            let sr = unsafe { regs.sr().read() };
-            if sr.tifre() {
-                return Err(Error::Framing);
-            }
-            if sr.ovr() {
-                return Err(Error::Overrun);
-            }
-            if sr.crce() {
-                return Err(Error::Crc);
-            }
-        }
-
-        Ok(words)
-    }
-}
-
-impl<'d, T: Instance, Tx, Rx> traits::Spi<u8> for Spi<'d, T, Tx, Rx> {
-    type Error = super::Error;
-}
-
-impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx> traits::Write<u8> for Spi<'d, T, Tx, Rx> {
-    #[rustfmt::skip]
-    type WriteFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
-
-    fn write<'a>(&'a mut self, data: &'a [u8]) -> Self::WriteFuture<'a> {
-        self.write_dma_u8(data)
-    }
-}
-
-impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::Read<u8>
-    for Spi<'d, T, Tx, Rx>
-{
-    #[rustfmt::skip]
-    type ReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
-
-    fn read<'a>(&'a mut self, data: &'a mut [u8]) -> Self::ReadFuture<'a> {
-        self.read_dma_u8(data)
-    }
-}
-
-impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::FullDuplex<u8>
-    for Spi<'d, T, Tx, Rx>
-{
-    #[rustfmt::skip]
-    type WriteReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
-
-    fn read_write<'a>(
-        &'a mut self,
-        read: &'a mut [u8],
-        write: &'a [u8],
-    ) -> Self::WriteReadFuture<'a> {
-        self.read_write_dma_u8(read, write)
-    }
-}