diff --git a/embassy-rp/src/pio.rs b/embassy-rp/src/pio.rs
index 3c5969db..2cf4761a 100644
--- a/embassy-rp/src/pio.rs
+++ b/embassy-rp/src/pio.rs
@@ -97,13 +97,13 @@ pub(crate) unsafe fn init() {
 /// Future that waits for TX-FIFO to become writable
 #[must_use = "futures do nothing unless you `.await` or poll them"]
 pub struct FifoOutFuture<'a, 'd, PIO: PioInstance, const SM: usize> {
-    sm: &'a mut PioStateMachine<'d, PIO, SM>,
+    sm_tx: &'a mut PioStateMachineTx<'d, PIO, SM>,
     value: u32,
 }
 
 impl<'a, 'd, PIO: PioInstance, const SM: usize> FifoOutFuture<'a, 'd, PIO, SM> {
-    pub fn new(sm: &'a mut PioStateMachine<'d, PIO, SM>, value: u32) -> Self {
-        FifoOutFuture { sm, value }
+    pub fn new(sm: &'a mut PioStateMachineTx<'d, PIO, SM>, value: u32) -> Self {
+        FifoOutFuture { sm_tx: sm, value }
     }
 }
 
@@ -112,7 +112,7 @@ impl<'a, 'd, PIO: PioInstance, const SM: usize> Future for FifoOutFuture<'a, 'd,
     fn poll(self: FuturePin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
         //debug!("Poll {},{}", PIO::PIO_NO, SM);
         let value = self.value;
-        if self.get_mut().sm.try_push_tx(value) {
+        if self.get_mut().sm_tx.try_push(value) {
             Poll::Ready(())
         } else {
             WAKERS[PIO::PIO_NO as usize].fifo_out()[SM].register(cx.waker());
@@ -140,12 +140,12 @@ impl<'a, 'd, PIO: PioInstance, const SM: usize> Drop for FifoOutFuture<'a, 'd, P
 /// Future that waits for RX-FIFO to become readable
 #[must_use = "futures do nothing unless you `.await` or poll them"]
 pub struct FifoInFuture<'a, 'd, PIO: PioInstance, const SM: usize> {
-    sm: &'a mut PioStateMachine<'d, PIO, SM>,
+    sm_rx: &'a mut PioStateMachineRx<'d, PIO, SM>,
 }
 
 impl<'a, 'd, PIO: PioInstance, const SM: usize> FifoInFuture<'a, 'd, PIO, SM> {
-    pub fn new(sm: &'a mut PioStateMachine<'d, PIO, SM>) -> Self {
-        FifoInFuture { sm }
+    pub fn new(sm: &'a mut PioStateMachineRx<'d, PIO, SM>) -> Self {
+        FifoInFuture { sm_rx: sm }
     }
 }
 
@@ -153,7 +153,7 @@ impl<'a, 'd, PIO: PioInstance, const SM: usize> Future for FifoInFuture<'a, 'd,
     type Output = u32;
     fn poll(mut self: FuturePin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
         //debug!("Poll {},{}", PIO::PIO_NO, SM);
-        if let Some(v) = self.sm.try_pull_rx() {
+        if let Some(v) = self.sm_rx.try_pull() {
             Poll::Ready(v)
         } else {
             WAKERS[PIO::PIO_NO as usize].fifo_in()[SM].register(cx.waker());
@@ -293,10 +293,163 @@ impl<'l, PIO: PioInstance> Pin<'l, PIO> {
     }
 }
 
-pub struct PioStateMachine<'d, PIO: PioInstance, const SM: usize> {
+pub struct PioStateMachineRx<'d, PIO: PioInstance, const SM: usize> {
     pio: PhantomData<&'d PIO>,
 }
 
+impl<'d, PIO: PioInstance, const SM: usize> PioStateMachineRx<'d, PIO, SM> {
+    pub fn empty(&self) -> bool {
+        unsafe { PIO::PIO.fstat().read().rxempty() & (1u8 << SM) != 0 }
+    }
+
+    pub fn full(&self) -> bool {
+        unsafe { PIO::PIO.fstat().read().rxfull() & (1u8 << SM) != 0 }
+    }
+
+    pub fn level(&self) -> u8 {
+        unsafe { (PIO::PIO.flevel().read().0 >> (SM * 8 + 4)) as u8 & 0x0f }
+    }
+
+    pub fn stalled(&self) -> bool {
+        unsafe {
+            let fdebug = PIO::PIO.fdebug();
+            let ret = fdebug.read().rxstall() & (1 << SM) != 0;
+            fdebug.write(|w| w.set_rxstall(1 << SM));
+            ret
+        }
+    }
+
+    pub fn underflowed(&self) -> bool {
+        unsafe {
+            let fdebug = PIO::PIO.fdebug();
+            let ret = fdebug.read().rxunder() & (1 << SM) != 0;
+            fdebug.write(|w| w.set_rxunder(1 << SM));
+            ret
+        }
+    }
+
+    pub fn pull(&mut self) -> u32 {
+        unsafe { PIO::PIO.rxf(SM).read() }
+    }
+
+    pub fn try_pull(&mut self) -> Option<u32> {
+        if self.empty() {
+            return None;
+        }
+        Some(self.pull())
+    }
+
+    pub fn wait_pull<'a>(&'a mut self) -> FifoInFuture<'a, 'd, PIO, SM> {
+        FifoInFuture::new(self)
+    }
+
+    pub fn dma_pull<'a, C: Channel, W: Word>(
+        &'a mut self,
+        ch: PeripheralRef<'a, C>,
+        data: &'a mut [W],
+    ) -> Transfer<'a, C> {
+        unsafe {
+            let pio_no = PIO::PIO_NO;
+            let p = ch.regs();
+            p.write_addr().write_value(data.as_ptr() as u32);
+            p.read_addr().write_value(PIO::PIO.rxf(SM).ptr() as u32);
+            p.trans_count().write_value(data.len() as u32);
+            compiler_fence(Ordering::SeqCst);
+            p.ctrl_trig().write(|w| {
+                // Set RX DREQ for this statemachine
+                w.set_treq_sel(TreqSel(pio_no * 8 + SM as u8 + 4));
+                w.set_data_size(W::size());
+                w.set_chain_to(ch.number());
+                w.set_incr_read(false);
+                w.set_incr_write(true);
+                w.set_en(true);
+            });
+            compiler_fence(Ordering::SeqCst);
+        }
+        Transfer::new(ch)
+    }
+}
+
+pub struct PioStateMachineTx<'d, PIO: PioInstance, const SM: usize> {
+    pio: PhantomData<&'d PIO>,
+}
+
+impl<'d, PIO: PioInstance, const SM: usize> PioStateMachineTx<'d, PIO, SM> {
+    pub fn empty(&self) -> bool {
+        unsafe { PIO::PIO.fstat().read().txempty() & (1u8 << SM) != 0 }
+    }
+    pub fn full(&self) -> bool {
+        unsafe { PIO::PIO.fstat().read().txfull() & (1u8 << SM) != 0 }
+    }
+
+    pub fn level(&self) -> u8 {
+        unsafe { (PIO::PIO.flevel().read().0 >> (SM * 8)) as u8 & 0x0f }
+    }
+
+    pub fn stalled(&self) -> bool {
+        unsafe {
+            let fdebug = PIO::PIO.fdebug();
+            let ret = fdebug.read().txstall() & (1 << SM) != 0;
+            fdebug.write(|w| w.set_txstall(1 << SM));
+            ret
+        }
+    }
+
+    pub fn overflowed(&self) -> bool {
+        unsafe {
+            let fdebug = PIO::PIO.fdebug();
+            let ret = fdebug.read().txover() & (1 << SM) != 0;
+            fdebug.write(|w| w.set_txover(1 << SM));
+            ret
+        }
+    }
+
+    pub fn push(&mut self, v: u32) {
+        unsafe {
+            PIO::PIO.txf(SM).write_value(v);
+        }
+    }
+
+    pub fn try_push(&mut self, v: u32) -> bool {
+        if self.full() {
+            return false;
+        }
+        self.push(v);
+        true
+    }
+
+    pub fn wait_push<'a>(&'a mut self, value: u32) -> FifoOutFuture<'a, 'd, PIO, SM> {
+        FifoOutFuture::new(self, value)
+    }
+
+    pub fn dma_push<'a, C: Channel, W: Word>(&'a mut self, ch: PeripheralRef<'a, C>, data: &'a [W]) -> Transfer<'a, C> {
+        unsafe {
+            let pio_no = PIO::PIO_NO;
+            let p = ch.regs();
+            p.read_addr().write_value(data.as_ptr() as u32);
+            p.write_addr().write_value(PIO::PIO.txf(SM).ptr() as u32);
+            p.trans_count().write_value(data.len() as u32);
+            compiler_fence(Ordering::SeqCst);
+            p.ctrl_trig().write(|w| {
+                // Set TX DREQ for this statemachine
+                w.set_treq_sel(TreqSel(pio_no * 8 + SM as u8));
+                w.set_data_size(W::size());
+                w.set_chain_to(ch.number());
+                w.set_incr_read(true);
+                w.set_incr_write(false);
+                w.set_en(true);
+            });
+            compiler_fence(Ordering::SeqCst);
+        }
+        Transfer::new(ch)
+    }
+}
+
+pub struct PioStateMachine<'d, PIO: PioInstance, const SM: usize> {
+    rx: PioStateMachineRx<'d, PIO, SM>,
+    tx: PioStateMachineTx<'d, PIO, SM>,
+}
+
 impl<'d, PIO: PioInstance, const SM: usize> Drop for PioStateMachine<'d, PIO, SM> {
     fn drop(&mut self) {
         unsafe {
@@ -333,59 +486,6 @@ impl<'d, PIO: PioInstance + 'd, const SM: usize> PioStateMachine<'d, PIO, SM> {
         unsafe { PIO::PIO.ctrl().read().sm_enable() & (1u8 << SM) != 0 }
     }
 
-    pub fn is_tx_empty(&self) -> bool {
-        unsafe { PIO::PIO.fstat().read().txempty() & (1u8 << SM) != 0 }
-    }
-    pub fn is_tx_full(&self) -> bool {
-        unsafe { PIO::PIO.fstat().read().txfull() & (1u8 << SM) != 0 }
-    }
-
-    pub fn is_rx_empty(&self) -> bool {
-        unsafe { PIO::PIO.fstat().read().rxempty() & (1u8 << SM) != 0 }
-    }
-    pub fn is_rx_full(&self) -> bool {
-        unsafe { PIO::PIO.fstat().read().rxfull() & (1u8 << SM) != 0 }
-    }
-
-    pub fn tx_level(&self) -> u8 {
-        unsafe {
-            let flevel = PIO::PIO.flevel().read().0;
-            (flevel >> (SM * 8)) as u8 & 0x0f
-        }
-    }
-
-    pub fn rx_level(&self) -> u8 {
-        unsafe {
-            let flevel = PIO::PIO.flevel().read().0;
-            (flevel >> (SM * 8 + 4)) as u8 & 0x0f
-        }
-    }
-
-    pub fn push_tx(&mut self, v: u32) {
-        unsafe {
-            PIO::PIO.txf(SM).write_value(v);
-        }
-    }
-
-    pub fn try_push_tx(&mut self, v: u32) -> bool {
-        if self.is_tx_full() {
-            return false;
-        }
-        self.push_tx(v);
-        true
-    }
-
-    pub fn pull_rx(&mut self) -> u32 {
-        unsafe { PIO::PIO.rxf(SM).read() }
-    }
-
-    pub fn try_pull_rx(&mut self) -> Option<u32> {
-        if self.is_rx_empty() {
-            return None;
-        }
-        Some(self.pull_rx())
-    }
-
     pub fn set_clkdiv(&mut self, div_x_256: u32) {
         unsafe {
             Self::this_sm().clkdiv().write(|w| w.0 = div_x_256 << 8);
@@ -671,92 +771,14 @@ impl<'d, PIO: PioInstance + 'd, const SM: usize> PioStateMachine<'d, PIO, SM> {
         }
     }
 
-    pub fn wait_push<'a>(&'a mut self, value: u32) -> FifoOutFuture<'a, 'd, PIO, SM> {
-        FifoOutFuture::new(self, value)
+    pub fn rx(&mut self) -> &mut PioStateMachineRx<'d, PIO, SM> {
+        &mut self.rx
     }
-
-    pub fn wait_pull<'a>(&'a mut self) -> FifoInFuture<'a, 'd, PIO, SM> {
-        FifoInFuture::new(self)
+    pub fn tx(&mut self) -> &mut PioStateMachineTx<'d, PIO, SM> {
+        &mut self.tx
     }
-
-    pub fn has_tx_stalled(&self) -> bool {
-        unsafe {
-            let fdebug = PIO::PIO.fdebug();
-            let ret = fdebug.read().txstall() & (1 << SM) != 0;
-            fdebug.write(|w| w.set_txstall(1 << SM));
-            ret
-        }
-    }
-
-    pub fn has_tx_overflowed(&self) -> bool {
-        unsafe {
-            let fdebug = PIO::PIO.fdebug();
-            let ret = fdebug.read().txover() & (1 << SM) != 0;
-            fdebug.write(|w| w.set_txover(1 << SM));
-            ret
-        }
-    }
-
-    pub fn has_rx_stalled(&self) -> bool {
-        unsafe {
-            let fdebug = PIO::PIO.fdebug();
-            let ret = fdebug.read().rxstall() & (1 << SM) != 0;
-            fdebug.write(|w| w.set_rxstall(1 << SM));
-            ret
-        }
-    }
-
-    pub fn has_rx_underflowed(&self) -> bool {
-        unsafe {
-            let fdebug = PIO::PIO.fdebug();
-            let ret = fdebug.read().rxunder() & (1 << SM) != 0;
-            fdebug.write(|w| w.set_rxunder(1 << SM));
-            ret
-        }
-    }
-
-    pub fn dma_push<'a, C: Channel, W: Word>(&'a self, ch: PeripheralRef<'a, C>, data: &'a [W]) -> Transfer<'a, C> {
-        unsafe {
-            let pio_no = PIO::PIO_NO;
-            let p = ch.regs();
-            p.read_addr().write_value(data.as_ptr() as u32);
-            p.write_addr().write_value(PIO::PIO.txf(SM).ptr() as u32);
-            p.trans_count().write_value(data.len() as u32);
-            compiler_fence(Ordering::SeqCst);
-            p.ctrl_trig().write(|w| {
-                // Set TX DREQ for this statemachine
-                w.set_treq_sel(TreqSel(pio_no * 8 + SM as u8));
-                w.set_data_size(W::size());
-                w.set_chain_to(ch.number());
-                w.set_incr_read(true);
-                w.set_incr_write(false);
-                w.set_en(true);
-            });
-            compiler_fence(Ordering::SeqCst);
-        }
-        Transfer::new(ch)
-    }
-
-    pub fn dma_pull<'a, C: Channel, W: Word>(&'a self, ch: PeripheralRef<'a, C>, data: &'a mut [W]) -> Transfer<'a, C> {
-        unsafe {
-            let pio_no = PIO::PIO_NO;
-            let p = ch.regs();
-            p.write_addr().write_value(data.as_ptr() as u32);
-            p.read_addr().write_value(PIO::PIO.rxf(SM).ptr() as u32);
-            p.trans_count().write_value(data.len() as u32);
-            compiler_fence(Ordering::SeqCst);
-            p.ctrl_trig().write(|w| {
-                // Set RX DREQ for this statemachine
-                w.set_treq_sel(TreqSel(pio_no * 8 + SM as u8 + 4));
-                w.set_data_size(W::size());
-                w.set_chain_to(ch.number());
-                w.set_incr_read(false);
-                w.set_incr_write(true);
-                w.set_en(true);
-            });
-            compiler_fence(Ordering::SeqCst);
-        }
-        Transfer::new(ch)
+    pub fn rx_tx(&mut self) -> (&mut PioStateMachineRx<'d, PIO, SM>, &mut PioStateMachineTx<'d, PIO, SM>) {
+        (&mut self.rx, &mut self.tx)
     }
 }
 
@@ -921,10 +943,22 @@ impl<'d, PIO: PioInstance> Pio<'d, PIO> {
             irq1: PioIrq { pio: PhantomData },
             irq2: PioIrq { pio: PhantomData },
             irq3: PioIrq { pio: PhantomData },
-            sm0: PioStateMachine { pio: PhantomData },
-            sm1: PioStateMachine { pio: PhantomData },
-            sm2: PioStateMachine { pio: PhantomData },
-            sm3: PioStateMachine { pio: PhantomData },
+            sm0: PioStateMachine {
+                rx: PioStateMachineRx { pio: PhantomData },
+                tx: PioStateMachineTx { pio: PhantomData },
+            },
+            sm1: PioStateMachine {
+                rx: PioStateMachineRx { pio: PhantomData },
+                tx: PioStateMachineTx { pio: PhantomData },
+            },
+            sm2: PioStateMachine {
+                rx: PioStateMachineRx { pio: PhantomData },
+                tx: PioStateMachineTx { pio: PhantomData },
+            },
+            sm3: PioStateMachine {
+                rx: PioStateMachineRx { pio: PhantomData },
+                tx: PioStateMachineTx { pio: PhantomData },
+            },
         }
     }
 }
diff --git a/embassy-rp/src/pio_instr_util.rs b/embassy-rp/src/pio_instr_util.rs
index ffe16024..81abdb66 100644
--- a/embassy-rp/src/pio_instr_util.rs
+++ b/embassy-rp/src/pio_instr_util.rs
@@ -8,7 +8,7 @@ pub fn set_x<PIO: PioInstance, const SM: usize>(sm: &mut PioStateMachine<PIO, SM
         bit_count: 32,
     }
     .encode();
-    sm.push_tx(value);
+    sm.tx().push(value);
     sm.exec_instr(OUT);
 }
 
@@ -19,7 +19,7 @@ pub fn get_x<PIO: PioInstance, const SM: usize>(sm: &mut PioStateMachine<PIO, SM
     }
     .encode();
     sm.exec_instr(IN);
-    sm.pull_rx()
+    sm.rx().pull()
 }
 
 pub fn set_y<PIO: PioInstance, const SM: usize>(sm: &mut PioStateMachine<PIO, SM>, value: u32) {
@@ -28,7 +28,7 @@ pub fn set_y<PIO: PioInstance, const SM: usize>(sm: &mut PioStateMachine<PIO, SM
         bit_count: 32,
     }
     .encode();
-    sm.push_tx(value);
+    sm.tx().push(value);
     sm.exec_instr(OUT);
 }
 
@@ -40,7 +40,7 @@ pub fn get_y<PIO: PioInstance, const SM: usize>(sm: &mut PioStateMachine<PIO, SM
     .encode();
     sm.exec_instr(IN);
 
-    sm.pull_rx()
+    sm.rx().pull()
 }
 
 pub fn set_pindir<PIO: PioInstance, const SM: usize>(sm: &mut PioStateMachine<PIO, SM>, data: u8) {
@@ -67,7 +67,7 @@ pub fn set_out_pin<PIO: PioInstance, const SM: usize>(sm: &mut PioStateMachine<P
         bit_count: 32,
     }
     .encode();
-    sm.push_tx(data);
+    sm.tx().push(data);
     sm.exec_instr(OUT);
 }
 pub fn set_out_pindir<PIO: PioInstance, const SM: usize>(sm: &mut PioStateMachine<PIO, SM>, data: u32) {
@@ -76,7 +76,7 @@ pub fn set_out_pindir<PIO: PioInstance, const SM: usize>(sm: &mut PioStateMachin
         bit_count: 32,
     }
     .encode();
-    sm.push_tx(data);
+    sm.tx().push(data);
     sm.exec_instr(OUT);
 }
 
diff --git a/examples/rp/src/bin/pio_async.rs b/examples/rp/src/bin/pio_async.rs
index 3d76a7d7..4e0ab5e3 100644
--- a/examples/rp/src/bin/pio_async.rs
+++ b/examples/rp/src/bin/pio_async.rs
@@ -42,7 +42,7 @@ async fn pio_task_sm0(mut sm: PioStateMachine<'static, PIO0, 0>) {
 
     let mut v = 0x0f0caffa;
     loop {
-        sm.wait_push(v).await;
+        sm.tx().wait_push(v).await;
         v ^= 0xffff;
         info!("Pushed {:032b} to FIFO", v);
     }
@@ -70,7 +70,7 @@ fn setup_pio_task_sm1(pio: &mut PioCommon<PIO0>, sm: &mut PioStateMachine<PIO0,
 async fn pio_task_sm1(mut sm: PioStateMachine<'static, PIO0, 1>) {
     sm.set_enable(true);
     loop {
-        let rx = sm.wait_pull().await;
+        let rx = sm.rx().wait_pull().await;
         info!("Pulled {:032b} from FIFO", rx);
     }
 }
diff --git a/examples/rp/src/bin/pio_dma.rs b/examples/rp/src/bin/pio_dma.rs
index a2a2ee39..c664482e 100644
--- a/examples/rp/src/bin/pio_dma.rs
+++ b/examples/rp/src/bin/pio_dma.rs
@@ -60,9 +60,10 @@ async fn main(_spawner: Spawner) {
     }
     let mut din = [0u32; 29];
     loop {
+        let (rx, tx) = sm.rx_tx();
         join(
-            sm.dma_push(dma_out_ref.reborrow(), &dout),
-            sm.dma_pull(dma_in_ref.reborrow(), &mut din),
+            tx.dma_push(dma_out_ref.reborrow(), &dout),
+            rx.dma_pull(dma_in_ref.reborrow(), &mut din),
         )
         .await;
         for i in 0..din.len() {
diff --git a/examples/rp/src/bin/pio_hd44780.rs b/examples/rp/src/bin/pio_hd44780.rs
index 7c1d7acf..f76d334e 100644
--- a/examples/rp/src/bin/pio_hd44780.rs
+++ b/examples/rp/src/bin/pio_hd44780.rs
@@ -139,14 +139,14 @@ impl<'l> HD44780<'l> {
 
         sm0.set_enable(true);
         // init to 8 bit thrice
-        sm0.push_tx((50000 << 8) | 0x30);
-        sm0.push_tx((5000 << 8) | 0x30);
-        sm0.push_tx((200 << 8) | 0x30);
+        sm0.tx().push((50000 << 8) | 0x30);
+        sm0.tx().push((5000 << 8) | 0x30);
+        sm0.tx().push((200 << 8) | 0x30);
         // init 4 bit
-        sm0.push_tx((200 << 8) | 0x20);
+        sm0.tx().push((200 << 8) | 0x20);
         // set font and lines
-        sm0.push_tx((50 << 8) | 0x20);
-        sm0.push_tx(0b1100_0000);
+        sm0.tx().push((50 << 8) | 0x20);
+        sm0.tx().push(0b1100_0000);
 
         irq0.wait().await;
         sm0.set_enable(false);
@@ -216,7 +216,7 @@ impl<'l> HD44780<'l> {
         sm0.set_enable(true);
 
         // display on and cursor on and blinking, reset display
-        sm0.dma_push(dma.reborrow(), &[0x81u8, 0x0f, 1]).await;
+        sm0.tx().dma_push(dma.reborrow(), &[0x81u8, 0x0f, 1]).await;
 
         Self {
             dma: dma.map_into(),
@@ -240,6 +240,6 @@ impl<'l> HD44780<'l> {
         // set cursor to 1:15
         self.buf[38..].copy_from_slice(&[0x80, 0xcf]);
 
-        self.sm.dma_push(self.dma.reborrow(), &self.buf).await;
+        self.sm.tx().dma_push(self.dma.reborrow(), &self.buf).await;
     }
 }
diff --git a/examples/rp/src/bin/ws2812-pio.rs b/examples/rp/src/bin/ws2812-pio.rs
index 713e01b4..c9c701a7 100644
--- a/examples/rp/src/bin/ws2812-pio.rs
+++ b/examples/rp/src/bin/ws2812-pio.rs
@@ -87,7 +87,7 @@ impl<'d, P: PioInstance, const S: usize> Ws2812<'d, P, S> {
     pub async fn write(&mut self, colors: &[RGB8]) {
         for color in colors {
             let word = (u32::from(color.g) << 24) | (u32::from(color.r) << 16) | (u32::from(color.b) << 8);
-            self.sm.wait_push(word).await;
+            self.sm.tx().wait_push(word).await;
         }
     }
 }