Merge #646

646: Debloat stm32-metapac r=Dirbaio a=Dirbaio

- Remove usage of `peripheral_counts!` macrotables.
- Remove `dbgmcu!` macrotable.
- Move the remaining macrotables to embassy-stm32 build.rs. This brings metapac decompressed size from ~250mb to ~100mb
- Deduplicate files with identical metadata. This brings decompressed size from ~100mb to ~70mb, compressed from ~10mb to ~4mb, which is finally small enough to publish on crates.io!

Co-authored-by: Dario Nieuwenhuis <[email protected]>

32 files changed, 661 insertions, 624 deletions

diff --git a/embassy-stm32/build.rs b/embassy-stm32/build.rs
index 1257e0152..e67c4b983 100644
--- a/embassy-stm32/build.rs
+++ b/embassy-stm32/build.rs
@@ -2,6 +2,7 @@ use proc_macro2::TokenStream;
 use quote::{format_ident, quote};
 use std::collections::{HashMap, HashSet};
 use std::env;
+use std::fmt::Write as _;
 use std::fs;
 use std::path::PathBuf;
 use stm32_metapac::metadata::METADATA;
@@ -530,9 +531,127 @@ fn main() {
     }
 
     // ========
-    // Write generated.rs
+    // Write foreach_foo! macrotables
+
+    let mut interrupts_table: Vec<Vec<String>> = Vec::new();
+    let mut peripherals_table: Vec<Vec<String>> = Vec::new();
+    let mut pins_table: Vec<Vec<String>> = Vec::new();
+    let mut dma_channels_table: Vec<Vec<String>> = Vec::new();
+
+    let gpio_base = METADATA
+        .peripherals
+        .iter()
+        .find(|p| p.name == "GPIOA")
+        .unwrap()
+        .address as u32;
+    let gpio_stride = 0x400;
+
+    for p in METADATA.peripherals {
+        if let Some(regs) = &p.registers {
+            if regs.kind == "gpio" {
+                let port_letter = p.name.chars().skip(4).next().unwrap();
+                assert_eq!(0, (p.address as u32 - gpio_base) % gpio_stride);
+                let port_num = (p.address as u32 - gpio_base) / gpio_stride;
+
+                for pin_num in 0u32..16 {
+                    let pin_name = format!("P{}{}", port_letter, pin_num);
+                    pins_table.push(vec![
+                        pin_name,
+                        p.name.to_string(),
+                        port_num.to_string(),
+                        pin_num.to_string(),
+                        format!("EXTI{}", pin_num),
+                    ]);
+                }
+            }
+
+            for irq in p.interrupts {
+                let mut row = Vec::new();
+                row.push(p.name.to_string());
+                row.push(regs.kind.to_string());
+                row.push(regs.block.to_string());
+                row.push(irq.signal.to_string());
+                row.push(irq.interrupt.to_ascii_uppercase());
+                interrupts_table.push(row)
+            }
+
+            let mut row = Vec::new();
+            row.push(regs.kind.to_string());
+            row.push(p.name.to_string());
+            peripherals_table.push(row);
+        }
+    }
+
+    let mut dma_channel_count: usize = 0;
+    let mut bdma_channel_count: usize = 0;
+
+    for ch in METADATA.dma_channels {
+        let mut row = Vec::new();
+        let dma_peri = METADATA
+            .peripherals
+            .iter()
+            .find(|p| p.name == ch.dma)
+            .unwrap();
+        let bi = dma_peri.registers.as_ref().unwrap();
+
+        let num;
+        match bi.kind {
+            "dma" => {
+                num = dma_channel_count;
+                dma_channel_count += 1;
+            }
+            "bdma" => {
+                num = bdma_channel_count;
+                bdma_channel_count += 1;
+            }
+            _ => panic!("bad dma channel kind {}", bi.kind),
+        }
+
+        row.push(ch.name.to_string());
+        row.push(ch.dma.to_string());
+        row.push(bi.kind.to_string());
+        row.push(ch.channel.to_string());
+        row.push(num.to_string());
+        if let Some(dmamux) = &ch.dmamux {
+            let dmamux_channel = ch.dmamux_channel.unwrap();
+            row.push(format!(
+                "{{dmamux: {}, dmamux_channel: {}}}",
+                dmamux, dmamux_channel
+            ));
+        } else {
+            row.push("{}".to_string());
+        }
+
+        dma_channels_table.push(row);
+    }
+
+    g.extend(quote! {
+        pub(crate) const DMA_CHANNEL_COUNT: usize = #dma_channel_count;
+        pub(crate) const BDMA_CHANNEL_COUNT: usize = #bdma_channel_count;
+    });
+
+    for irq in METADATA.interrupts {
+        let name = irq.name.to_ascii_uppercase();
+        interrupts_table.push(vec![name.clone()]);
+        if name.contains("EXTI") {
+            interrupts_table.push(vec!["EXTI".to_string(), name.clone()]);
+        }
+    }
+
+    let mut m = String::new();
+
+    make_table(&mut m, "foreach_interrupt", &interrupts_table);
+    make_table(&mut m, "foreach_peripheral", &peripherals_table);
+    make_table(&mut m, "foreach_pin", &pins_table);
+    make_table(&mut m, "foreach_dma_channel", &dma_channels_table);
 
     let out_dir = &PathBuf::from(env::var_os("OUT_DIR").unwrap());
+    let out_file = out_dir.join("macros.rs").to_string_lossy().to_string();
+    fs::write(out_file, m).unwrap();
+
+    // ========
+    // Write generated.rs
+
     let out_file = out_dir.join("generated.rs").to_string_lossy().to_string();
     fs::write(out_file, g.to_string()).unwrap();
 
@@ -644,3 +763,30 @@ impl<T: Iterator> IteratorExt for T {
         }
     }
 }
+
+fn make_table(out: &mut String, name: &str, data: &Vec<Vec<String>>) {
+    write!(
+        out,
+        "#[macro_export]
+macro_rules! {} {{
+    ($($pat:tt => $code:tt;)*) => {{
+        macro_rules! __{}_inner {{
+            $(($pat) => $code;)*
+            ($_:tt) => {{}}
+        }}
+",
+        name, name
+    )
+    .unwrap();
+
+    for row in data {
+        write!(out, "        __{}_inner!(({}));\n", name, row.join(",")).unwrap();
+    }
+
+    write!(
+        out,
+        "    }};
+}}"
+    )
+    .unwrap();
+}
diff --git a/embassy-stm32/src/adc/mod.rs b/embassy-stm32/src/adc/mod.rs
index 86f33ca19..7b3233a1d 100644
--- a/embassy-stm32/src/adc/mod.rs
+++ b/embassy-stm32/src/adc/mod.rs
@@ -36,7 +36,7 @@ pub trait Instance: sealed::Instance + crate::rcc::RccPeripheral + 'static {}
 pub trait Common: sealed::Common + 'static {}
 pub trait AdcPin<T: Instance>: sealed::AdcPin<T> {}
 
-crate::pac::peripherals!(
+foreach_peripheral!(
     (adc, $inst:ident) => {
         impl crate::adc::sealed::Instance for peripherals::$inst {
             fn regs() -> &'static crate::pac::adc::Adc {
@@ -44,7 +44,7 @@ crate::pac::peripherals!(
             }
             #[cfg(not(adc_f1))]
             fn common_regs() -> &'static crate::pac::adccommon::AdcCommon {
-                crate::pac::peripherals!{
+                foreach_peripheral!{
                     (adccommon, $common_inst:ident) => {
                         return &crate::pac::$common_inst
                     };
@@ -57,7 +57,7 @@ crate::pac::peripherals!(
 );
 
 #[cfg(not(adc_f1))]
-crate::pac::peripherals!(
+foreach_peripheral!(
     (adccommon, $inst:ident) => {
         impl sealed::Common for peripherals::$inst {
             fn regs() -> &'static crate::pac::adccommon::AdcCommon {
diff --git a/embassy-stm32/src/can/bxcan.rs b/embassy-stm32/src/can/bxcan.rs
index 856f3151a..c52d737bc 100644
--- a/embassy-stm32/src/can/bxcan.rs
+++ b/embassy-stm32/src/can/bxcan.rs
@@ -70,7 +70,7 @@ pub(crate) mod sealed {
 
 pub trait Instance: sealed::Instance + RccPeripheral {}
 
-crate::pac::peripherals!(
+foreach_peripheral!(
     (can, $inst:ident) => {
         impl sealed::Instance for peripherals::$inst {
             fn regs() -> &'static crate::pac::can::Can {
@@ -86,7 +86,7 @@ crate::pac::peripherals!(
     };
 );
 
-crate::pac::peripherals!(
+foreach_peripheral!(
     (can, CAN) => {
         unsafe impl bxcan::FilterOwner for peripherals::CAN {
             const NUM_FILTER_BANKS: u8 = 14;
diff --git a/embassy-stm32/src/dac/mod.rs b/embassy-stm32/src/dac/mod.rs
index 7432c34a0..1f6ba63df 100644
--- a/embassy-stm32/src/dac/mod.rs
+++ b/embassy-stm32/src/dac/mod.rs
@@ -16,7 +16,7 @@ pub trait Instance: sealed::Instance + 'static {}
 
 pub trait DacPin<T: Instance, const C: u8>: crate::gpio::Pin + 'static {}
 
-crate::pac::peripherals!(
+foreach_peripheral!(
     (dac, $inst:ident) => {
         impl crate::dac::sealed::Instance for peripherals::$inst {
             fn regs() -> &'static crate::pac::dac::Dac {
diff --git a/embassy-stm32/src/dcmi.rs b/embassy-stm32/src/dcmi.rs
index 9ac96c69a..bf78b631f 100644
--- a/embassy-stm32/src/dcmi.rs
+++ b/embassy-stm32/src/dcmi.rs
@@ -474,7 +474,7 @@ macro_rules! impl_peripheral {
     };
 }
 
-crate::pac::interrupts! {
+foreach_interrupt! {
     ($inst:ident, dcmi, $block:ident, GLOBAL, $irq:ident) => {
         impl_peripheral!($inst, $irq);
     };
diff --git a/embassy-stm32/src/dma/bdma.rs b/embassy-stm32/src/dma/bdma.rs
index 2f0715cf6..4fafe7dfa 100644
--- a/embassy-stm32/src/dma/bdma.rs
+++ b/embassy-stm32/src/dma/bdma.rs
@@ -7,6 +7,7 @@ use embassy::interrupt::{Interrupt, InterruptExt};
 use embassy::waitqueue::AtomicWaker;
 
 use crate::dma::Request;
+use crate::generated::BDMA_CHANNEL_COUNT;
 use crate::pac;
 use crate::pac::bdma::vals;
 
@@ -22,62 +23,44 @@ impl From<WordSize> for vals::Size {
     }
 }
 
-const CH_COUNT: usize = pac::peripheral_count!(bdma) * 8;
-
 struct State {
-    ch_wakers: [AtomicWaker; CH_COUNT],
+    ch_wakers: [AtomicWaker; BDMA_CHANNEL_COUNT],
 }
 
 impl State {
     const fn new() -> Self {
         const AW: AtomicWaker = AtomicWaker::new();
         Self {
-            ch_wakers: [AW; CH_COUNT],
+            ch_wakers: [AW; BDMA_CHANNEL_COUNT],
         }
     }
 }
 
 static STATE: State = State::new();
 
-macro_rules! dma_num {
-    (DMA1) => {
-        0
-    };
-    (DMA2) => {
-        1
-    };
-    (BDMA) => {
-        0
-    };
-    (BDMA1) => {
-        0
-    };
-    (BDMA2) => {
-        1
-    };
-}
-
 pub(crate) unsafe fn on_irq() {
-    pac::peripherals! {
+    foreach_peripheral! {
+        (bdma, BDMA1) => {
+            // BDMA1 in H7 doesn't use DMAMUX, which breaks
+        };
         (bdma, $dma:ident) => {
-                let isr = pac::$dma.isr().read();
-                let dman = dma_num!($dma);
-
-                for chn in 0..pac::dma_channels_count!($dma) {
-                    let cr = pac::$dma.ch(chn).cr();
-                    if isr.tcif(chn) && cr.read().tcie() {
+            let isr = pac::$dma.isr().read();
+            foreach_dma_channel! {
+                ($channel_peri:ident, $dma, bdma, $channel_num:expr, $index:expr, $dmamux:tt) => {
+                    let cr = pac::$dma.ch($channel_num).cr();
+                    if isr.tcif($channel_num) && cr.read().tcie() {
                         cr.write(|_| ()); // Disable channel interrupts with the default value.
-                        let n = dma_num!($dma) * 8 + chn;
-                        STATE.ch_wakers[n].wake();
+                        STATE.ch_wakers[$index].wake();
                     }
-                }
+                };
+            }
         };
     }
 }
 
 /// safety: must be called only once
 pub(crate) unsafe fn init() {
-    pac::interrupts! {
+    foreach_interrupt! {
         ($peri:ident, bdma, $block:ident, $signal_name:ident, $irq:ident) => {
             crate::interrupt::$irq::steal().enable();
         };
@@ -85,20 +68,20 @@ pub(crate) unsafe fn init() {
     crate::generated::init_bdma();
 }
 
-pac::dma_channels! {
-    ($channel_peri:ident, BDMA1, bdma, $channel_num:expr, $dmamux:tt) => {
+foreach_dma_channel! {
+    ($channel_peri:ident, BDMA1, bdma, $channel_num:expr, $index:expr, $dmamux:tt) => {
         // BDMA1 in H7 doesn't use DMAMUX, which breaks
     };
-    ($channel_peri:ident, $dma_peri:ident, bdma, $channel_num:expr, $dmamux:tt) => {
+    ($channel_peri:ident, $dma_peri:ident, bdma, $channel_num:expr, $index:expr, $dmamux:tt) => {
         impl crate::dma::sealed::Channel for crate::peripherals::$channel_peri {
 
-            unsafe fn start_write<W: Word>(&mut self, request: Request, buf: *const[W], reg_addr: *mut W) {
+            unsafe fn start_write<W: Word>(&mut self, _request: Request, buf: *const[W], reg_addr: *mut W) {
                 let (ptr, len) = super::slice_ptr_parts(buf);
                 low_level_api::start_transfer(
                     pac::$dma_peri,
                     $channel_num,
                     #[cfg(any(bdma_v2, dmamux))]
-                    request,
+                    _request,
                     vals::Dir::FROMMEMORY,
                     reg_addr as *const u32,
                     ptr as *mut u32,
@@ -113,13 +96,13 @@ pac::dma_channels! {
             }
 
 
-            unsafe fn start_write_repeated<W: Word>(&mut self, request: Request, repeated: W, count: usize, reg_addr: *mut W) {
+            unsafe fn start_write_repeated<W: Word>(&mut self, _request: Request, repeated: W, count: usize, reg_addr: *mut W) {
                 let buf = [repeated];
                 low_level_api::start_transfer(
                     pac::$dma_peri,
                     $channel_num,
                     #[cfg(any(bdma_v2, dmamux))]
-                    request,
+                    _request,
                     vals::Dir::FROMMEMORY,
                     reg_addr as *const u32,
                     buf.as_ptr() as *mut u32,
@@ -133,13 +116,13 @@ pac::dma_channels! {
                 )
             }
 
-            unsafe fn start_read<W: Word>(&mut self, request: Request, reg_addr: *const W, buf: *mut [W]) {
+            unsafe fn start_read<W: Word>(&mut self, _request: Request, reg_addr: *const W, buf: *mut [W]) {
                 let (ptr, len) = super::slice_ptr_parts_mut(buf);
                 low_level_api::start_transfer(
                     pac::$dma_peri,
                     $channel_num,
                     #[cfg(any(bdma_v2, dmamux))]
-                    request,
+                    _request,
                     vals::Dir::FROMPERIPHERAL,
                     reg_addr as *const u32,
                     ptr as *mut u32,
@@ -165,7 +148,7 @@ pac::dma_channels! {
             }
 
             fn set_waker(&mut self, waker: &Waker) {
-                unsafe {low_level_api::set_waker(dma_num!($dma_peri) * 8 + $channel_num, waker )}
+                unsafe { low_level_api::set_waker($index, waker) }
             }
         }
 
diff --git a/embassy-stm32/src/dma/dma.rs b/embassy-stm32/src/dma/dma.rs
index c885452b0..fd1732fbb 100644
--- a/embassy-stm32/src/dma/dma.rs
+++ b/embassy-stm32/src/dma/dma.rs
@@ -4,14 +4,13 @@ use core::task::Waker;
 use embassy::interrupt::{Interrupt, InterruptExt};
 use embassy::waitqueue::AtomicWaker;
 
+use crate::generated::DMA_CHANNEL_COUNT;
 use crate::interrupt;
 use crate::pac;
 use crate::pac::dma::{regs, vals};
 
 use super::{Request, Word, WordSize};
 
-const CH_COUNT: usize = pac::peripheral_count!(DMA) * 8;
-
 impl From<WordSize> for vals::Size {
     fn from(raw: WordSize) -> Self {
         match raw {
@@ -23,44 +22,31 @@ impl From<WordSize> for vals::Size {
 }
 
 struct State {
-    ch_wakers: [AtomicWaker; CH_COUNT],
+    ch_wakers: [AtomicWaker; DMA_CHANNEL_COUNT],
 }
 
 impl State {
     const fn new() -> Self {
         const AW: AtomicWaker = AtomicWaker::new();
         Self {
-            ch_wakers: [AW; CH_COUNT],
+            ch_wakers: [AW; DMA_CHANNEL_COUNT],
         }
     }
 }
 
 static STATE: State = State::new();
 
-macro_rules! dma_num {
-    (DMA1) => {
-        0
-    };
-    (DMA2) => {
-        1
-    };
-}
-
 pub(crate) unsafe fn on_irq() {
-    pac::peripherals! {
+    foreach_peripheral! {
         (dma, $dma:ident) => {
-            for isrn in 0..2 {
-                let isr = pac::$dma.isr(isrn).read();
-
-                for chn in 0..4 {
-                    let cr = pac::$dma.st(isrn * 4 + chn).cr();
-
-                    if isr.tcif(chn) && cr.read().tcie() {
+            foreach_dma_channel! {
+                ($channel_peri:ident, $dma, dma, $channel_num:expr, $index:expr, $dmamux:tt) => {
+                    let cr = pac::$dma.st($channel_num).cr();
+                    if pac::$dma.isr($channel_num/4).read().tcif($channel_num%4) && cr.read().tcie() {
                         cr.write(|_| ()); // Disable channel interrupts with the default value.
-                        let n = dma_num!($dma) * 8 + isrn * 4 + chn;
-                        STATE.ch_wakers[n].wake();
+                        STATE.ch_wakers[$index].wake();
                     }
-                }
+                };
             }
         };
     }
@@ -68,7 +54,7 @@ pub(crate) unsafe fn on_irq() {
 
 /// safety: must be called only once
 pub(crate) unsafe fn init() {
-    pac::interrupts! {
+    foreach_interrupt! {
         ($peri:ident, dma, $block:ident, $signal_name:ident, $irq:ident) => {
             interrupt::$irq::steal().enable();
         };
@@ -76,8 +62,8 @@ pub(crate) unsafe fn init() {
     crate::generated::init_dma();
 }
 
-pac::dma_channels! {
-    ($channel_peri:ident, $dma_peri:ident, dma, $channel_num:expr, $dmamux:tt) => {
+foreach_dma_channel! {
+    ($channel_peri:ident, $dma_peri:ident, dma, $channel_num:expr, $index:expr, $dmamux:tt) => {
         impl crate::dma::sealed::Channel for crate::peripherals::$channel_peri {
             unsafe fn start_write<W: Word>(&mut self, request: Request, buf: *const [W], reg_addr: *mut W) {
                 let (ptr, len) = super::slice_ptr_parts(buf);
@@ -149,7 +135,7 @@ pac::dma_channels! {
             }
 
             fn set_waker(&mut self, waker: &Waker) {
-                unsafe {low_level_api::set_waker(dma_num!($dma_peri) * 8 + $channel_num, waker )}
+                unsafe {low_level_api::set_waker($index, waker )}
             }
         }
 
diff --git a/embassy-stm32/src/dma/dmamux.rs b/embassy-stm32/src/dma/dmamux.rs
index 971695a06..628f496be 100644
--- a/embassy-stm32/src/dma/dmamux.rs
+++ b/embassy-stm32/src/dma/dmamux.rs
@@ -35,8 +35,8 @@ pub trait MuxChannel: sealed::MuxChannel + super::Channel {
     type Mux;
 }
 
-pac::dma_channels! {
-    ($channel_peri:ident, $dma_peri:ident, $version:ident, $channel_num:expr, {dmamux: $dmamux:ident, dmamux_channel: $dmamux_channel:expr}) => {
+foreach_dma_channel! {
+    ($channel_peri:ident, $dma_peri:ident, $version:ident, $channel_num:expr, $index:expr, {dmamux: $dmamux:ident, dmamux_channel: $dmamux_channel:expr}) => {
         impl sealed::MuxChannel for peripherals::$channel_peri {
             const DMAMUX_CH_NUM: u8 = $dmamux_channel;
             const DMAMUX_REGS: pac::dmamux::Dmamux = pac::$dmamux;
diff --git a/embassy-stm32/src/exti.rs b/embassy-stm32/src/exti.rs
index 909d0ee88..47307530b 100644
--- a/embassy-stm32/src/exti.rs
+++ b/embassy-stm32/src/exti.rs
@@ -278,7 +278,7 @@ impl<'a> Future for ExtiInputFuture<'a> {
 
 macro_rules! foreach_exti_irq {
     ($action:ident) => {
-        crate::pac::interrupts!(
+        foreach_interrupt!(
             (EXTI0)  => { $action!(EXTI0); };
             (EXTI1)  => { $action!(EXTI1); };
             (EXTI2)  => { $action!(EXTI2); };
diff --git a/embassy-stm32/src/fmc/mod.rs b/embassy-stm32/src/fmc/mod.rs
index a17b88ea9..2a730f5f8 100644
--- a/embassy-stm32/src/fmc/mod.rs
+++ b/embassy-stm32/src/fmc/mod.rs
@@ -127,7 +127,7 @@ impl<'d, T: Instance> Fmc<'d, T> {
     ));
 }
 
-crate::pac::peripherals!(
+foreach_peripheral!(
     (fmc, $inst:ident) => {
         impl crate::fmc::sealed::Instance for crate::peripherals::$inst {
             fn regs() -> stm32_metapac::fmc::Fmc {
diff --git a/embassy-stm32/src/gpio.rs b/embassy-stm32/src/gpio.rs
index aef8ee552..4837c4120 100644
--- a/embassy-stm32/src/gpio.rs
+++ b/embassy-stm32/src/gpio.rs
@@ -558,7 +558,7 @@ impl sealed::Pin for AnyPin {
 
 // ====================
 
-crate::pac::pins!(
+foreach_pin!(
     ($pin_name:ident, $port_name:ident, $port_num:expr, $pin_num:expr, $exti_ch:ident) => {
         impl Pin for peripherals::$pin_name {
             #[cfg(feature = "exti")]
diff --git a/embassy-stm32/src/i2c/mod.rs b/embassy-stm32/src/i2c/mod.rs
index aad1c1375..c2a4c2546 100644
--- a/embassy-stm32/src/i2c/mod.rs
+++ b/embassy-stm32/src/i2c/mod.rs
@@ -21,9 +21,10 @@ pub enum Error {
 }
 
 pub(crate) mod sealed {
+    use super::*;
     pub trait Instance: crate::rcc::RccPeripheral {
         fn regs() -> crate::pac::i2c::I2c;
-        fn state_number() -> usize;
+        fn state() -> &'static State;
     }
 }
 
@@ -36,33 +37,16 @@ pin_trait!(SdaPin, Instance);
 dma_trait!(RxDma, Instance);
 dma_trait!(TxDma, Instance);
 
-macro_rules! i2c_state {
-    (I2C1) => {
-        0
-    };
-    (I2C2) => {
-        1
-    };
-    (I2C3) => {
-        2
-    };
-    (I2C4) => {
-        3
-    };
-    (I2C5) => {
-        4
-    };
-}
-
-crate::pac::interrupts!(
+foreach_interrupt!(
     ($inst:ident, i2c, $block:ident, EV, $irq:ident) => {
         impl sealed::Instance for peripherals::$inst {
             fn regs() -> crate::pac::i2c::I2c {
                 crate::pac::$inst
             }
 
-            fn state_number() -> usize {
-                i2c_state!($inst)
+            fn state() -> &'static State {
+                static STATE: State = State::new();
+                &STATE
             }
         }
 
diff --git a/embassy-stm32/src/i2c/v1.rs b/embassy-stm32/src/i2c/v1.rs
index 2bd0dcdf7..f280187e5 100644
--- a/embassy-stm32/src/i2c/v1.rs
+++ b/embassy-stm32/src/i2c/v1.rs
@@ -7,6 +7,14 @@ use crate::i2c::{Error, Instance, SclPin, SdaPin};
 use crate::pac::i2c;
 use crate::time::Hertz;
 
+pub struct State {}
+
+impl State {
+    pub(crate) const fn new() -> Self {
+        Self {}
+    }
+}
+
 pub struct I2c<'d, T: Instance> {
     phantom: PhantomData<&'d mut T>,
 }
diff --git a/embassy-stm32/src/i2c/v2.rs b/embassy-stm32/src/i2c/v2.rs
index a1ba5bc7d..2c46237d6 100644
--- a/embassy-stm32/src/i2c/v2.rs
+++ b/embassy-stm32/src/i2c/v2.rs
@@ -13,31 +13,23 @@ use futures::future::poll_fn;
 use crate::dma::NoDma;
 use crate::gpio::sealed::AFType;
 use crate::i2c::{Error, Instance, SclPin, SdaPin};
-use crate::pac;
 use crate::pac::i2c;
 use crate::time::Hertz;
 
-const I2C_COUNT: usize = pac::peripheral_count!(i2c);
-
 pub struct State {
-    waker: [AtomicWaker; I2C_COUNT],
-    chunks_transferred: [AtomicUsize; I2C_COUNT],
+    waker: AtomicWaker,
+    chunks_transferred: AtomicUsize,
 }
 
 impl State {
-    const fn new() -> Self {
-        const AW: AtomicWaker = AtomicWaker::new();
-        const CT: AtomicUsize = AtomicUsize::new(0);
-
+    pub(crate) const fn new() -> Self {
         Self {
-            waker: [AW; I2C_COUNT],
-            chunks_transferred: [CT; I2C_COUNT],
+            waker: AtomicWaker::new(),
+            chunks_transferred: AtomicUsize::new(0),
         }
     }
 }
 
-static STATE: State = State::new();
-
 pub struct I2c<'d, T: Instance, TXDMA = NoDma, RXDMA = NoDma> {
     phantom: PhantomData<&'d mut T>,
     tx_dma: TXDMA,
@@ -108,9 +100,9 @@ impl<'d, T: Instance, TXDMA, RXDMA> I2c<'d, T, TXDMA, RXDMA> {
         let isr = regs.isr().read();
 
         if isr.tcr() || isr.tc() {
-            let n = T::state_number();
-            STATE.chunks_transferred[n].fetch_add(1, Ordering::Relaxed);
-            STATE.waker[n].wake();
+            let state = T::state();
+            state.chunks_transferred.fetch_add(1, Ordering::Relaxed);
+            state.waker.wake();
         }
         // The flag can only be cleared by writting to nbytes, we won't do that here, so disable
         // the interrupt
@@ -411,8 +403,8 @@ impl<'d, T: Instance, TXDMA, RXDMA> I2c<'d, T, TXDMA, RXDMA> {
             crate::dma::write(ch, request, bytes, dst)
         };
 
-        let state_number = T::state_number();
-        STATE.chunks_transferred[state_number].store(0, Ordering::Relaxed);
+        let state = T::state();
+        state.chunks_transferred.store(0, Ordering::Relaxed);
         let mut remaining_len = total_len;
 
         let _on_drop = OnDrop::new(|| {
@@ -445,8 +437,8 @@ impl<'d, T: Instance, TXDMA, RXDMA> I2c<'d, T, TXDMA, RXDMA> {
         }
 
         poll_fn(|cx| {
-            STATE.waker[state_number].register(cx.waker());
-            let chunks_transferred = STATE.chunks_transferred[state_number].load(Ordering::Relaxed);
+            state.waker.register(cx.waker());
+            let chunks_transferred = state.chunks_transferred.load(Ordering::Relaxed);
 
             if chunks_transferred == total_chunks {
                 return Poll::Ready(());
@@ -504,8 +496,8 @@ impl<'d, T: Instance, TXDMA, RXDMA> I2c<'d, T, TXDMA, RXDMA> {
             crate::dma::read(ch, request, src, buffer)
         };
 
-        let state_number = T::state_number();
-        STATE.chunks_transferred[state_number].store(0, Ordering::Relaxed);
+        let state = T::state();
+        state.chunks_transferred.store(0, Ordering::Relaxed);
         let mut remaining_len = total_len;
 
         let _on_drop = OnDrop::new(|| {
@@ -530,8 +522,8 @@ impl<'d, T: Instance, TXDMA, RXDMA> I2c<'d, T, TXDMA, RXDMA> {
         }
 
         poll_fn(|cx| {
-            STATE.waker[state_number].register(cx.waker());
-            let chunks_transferred = STATE.chunks_transferred[state_number].load(Ordering::Relaxed);
+            state.waker.register(cx.waker());
+            let chunks_transferred = state.chunks_transferred.load(Ordering::Relaxed);
 
             if chunks_transferred == total_chunks {
                 return Poll::Ready(());
diff --git a/embassy-stm32/src/lib.rs b/embassy-stm32/src/lib.rs
index 79221e600..5e8d6dd86 100644
--- a/embassy-stm32/src/lib.rs
+++ b/embassy-stm32/src/lib.rs
@@ -11,6 +11,7 @@ pub(crate) use stm32_metapac as pac;
 
 // This must go FIRST so that all the other modules see its macros.
 pub mod fmt;
+include!(concat!(env!("OUT_DIR"), "/macros.rs"));
 
 // Utilities
 pub mod interrupt;
@@ -62,7 +63,7 @@ pub mod usb_otg;
 pub mod subghz;
 
 // This must go last, so that it sees all the impl_foo! macros defined earlier.
-mod generated {
+pub(crate) mod generated {
 
     #![allow(dead_code)]
     #![allow(unused_imports)]
@@ -98,10 +99,27 @@ pub fn init(config: Config) -> Peripherals {
         #[cfg(dbgmcu)]
         if config.enable_debug_during_sleep {
             crate::pac::DBGMCU.cr().modify(|cr| {
-                crate::pac::dbgmcu! {
-                    (cr, $fn_name:ident) => {
-                        cr.$fn_name(true);
-                    };
+                #[cfg(any(dbgmcu_f0, dbgmcu_g0, dbgmcu_u5))]
+                {
+                    cr.set_dbg_stop(true);
+                    cr.set_dbg_standby(true);
+                }
+                #[cfg(any(
+                    dbgmcu_f1, dbgmcu_f2, dbgmcu_f3, dbgmcu_f4, dbgmcu_f7, dbgmcu_g4, dbgmcu_f7,
+                    dbgmcu_l0, dbgmcu_l1, dbgmcu_l4, dbgmcu_wb, dbgmcu_wl
+                ))]
+                {
+                    cr.set_dbg_sleep(true);
+                    cr.set_dbg_stop(true);
+                    cr.set_dbg_standby(true);
+                }
+                #[cfg(dbgmcu_h7)]
+                {
+                    cr.set_d1dbgcken(true);
+                    cr.set_d3dbgcken(true);
+                    cr.set_dbgsleep_d1(true);
+                    cr.set_dbgstby_d1(true);
+                    cr.set_dbgstop_d1(true);
                 }
             });
         }
diff --git a/embassy-stm32/src/pwm/mod.rs b/embassy-stm32/src/pwm/mod.rs
index edc34fa5c..e8938ffae 100644
--- a/embassy-stm32/src/pwm/mod.rs
+++ b/embassy-stm32/src/pwm/mod.rs
@@ -127,7 +127,7 @@ macro_rules! impl_compare_capable_16bit {
     };
 }
 
-crate::pac::interrupts! {
+foreach_interrupt! {
     ($inst:ident, timer, TIM_GP16, UP, $irq:ident) => {
         impl crate::pwm::sealed::CaptureCompare16bitInstance for crate::peripherals::$inst {
             unsafe fn set_output_compare_mode(
diff --git a/embassy-stm32/src/rng.rs b/embassy-stm32/src/rng.rs
index 032fee011..0a93951bf 100644
--- a/embassy-stm32/src/rng.rs
+++ b/embassy-stm32/src/rng.rs
@@ -136,7 +136,7 @@ pub(crate) mod sealed {
 
 pub trait Instance: sealed::Instance + crate::rcc::RccPeripheral {}
 
-crate::pac::peripherals!(
+foreach_peripheral!(
     (rng, $inst:ident) => {
         impl Instance for peripherals::$inst {}
 
@@ -165,7 +165,7 @@ macro_rules! irq {
     };
 }
 
-crate::pac::interrupts!(
+foreach_interrupt!(
     (RNG) => {
         irq!(RNG);
     };
diff --git a/embassy-stm32/src/sdmmc/v2.rs b/embassy-stm32/src/sdmmc/v2.rs
index 1f4c9db70..cb8fa5544 100644
--- a/embassy-stm32/src/sdmmc/v2.rs
+++ b/embassy-stm32/src/sdmmc/v2.rs
@@ -1271,7 +1271,7 @@ where
     }
 }
 
-crate::pac::peripherals!(
+foreach_peripheral!(
     (sdmmc, $inst:ident) => {
         impl sealed::Instance for peripherals::$inst {
             type Interrupt = crate::interrupt::$inst;
diff --git a/embassy-stm32/src/spi/mod.rs b/embassy-stm32/src/spi/mod.rs
index fe2014147..e3b647280 100644
--- a/embassy-stm32/src/spi/mod.rs
+++ b/embassy-stm32/src/spi/mod.rs
@@ -874,7 +874,7 @@ pin_trait!(MisoPin, Instance);
 dma_trait!(RxDma, Instance);
 dma_trait!(TxDma, Instance);
 
-crate::pac::peripherals!(
+foreach_peripheral!(
     (spi, $inst:ident) => {
         impl sealed::Instance for peripherals::$inst {
             fn regs() -> &'static crate::pac::spi::Spi {
diff --git a/embassy-stm32/src/time_driver.rs b/embassy-stm32/src/time_driver.rs
index 98054e051..009c62030 100644
--- a/embassy-stm32/src/time_driver.rs
+++ b/embassy-stm32/src/time_driver.rs
@@ -29,7 +29,7 @@ type T = peripherals::TIM4;
 #[cfg(time_driver_tim5)]
 type T = peripherals::TIM5;
 
-crate::pac::interrupts! {
+foreach_interrupt! {
     (TIM2, timer, $block:ident, UP, $irq:ident) => {
         #[cfg(time_driver_tim2)]
         #[interrupt]
diff --git a/embassy-stm32/src/timer/mod.rs b/embassy-stm32/src/timer/mod.rs
index e3389fbad..4c1eb946b 100644
--- a/embassy-stm32/src/timer/mod.rs
+++ b/embassy-stm32/src/timer/mod.rs
@@ -155,7 +155,7 @@ macro_rules! impl_32bit_timer {
     };
 }
 
-crate::pac::interrupts! {
+foreach_interrupt! {
     ($inst:ident, timer, TIM_BASIC, UP, $irq:ident) => {
         impl_basic_16bit_timer!($inst, $irq);
 
diff --git a/embassy-stm32/src/usart/mod.rs b/embassy-stm32/src/usart/mod.rs
index 12e5d503d..60e607126 100644
--- a/embassy-stm32/src/usart/mod.rs
+++ b/embassy-stm32/src/usart/mod.rs
@@ -601,7 +601,7 @@ pin_trait!(CkPin, Instance);
 dma_trait!(TxDma, Instance);
 dma_trait!(RxDma, Instance);
 
-crate::pac::interrupts!(
+foreach_interrupt!(
     ($inst:ident, usart, $block:ident, $signal_name:ident, $irq:ident) => {
         impl sealed::Instance for peripherals::$inst {
             fn regs(&self) -> crate::pac::usart::Usart {
diff --git a/embassy-stm32/src/usb_otg.rs b/embassy-stm32/src/usb_otg.rs
index 6e08815d0..8c2c1e99c 100644
--- a/embassy-stm32/src/usb_otg.rs
+++ b/embassy-stm32/src/usb_otg.rs
@@ -135,7 +135,7 @@ pin_trait!(UlpiD5Pin, Instance);
 pin_trait!(UlpiD6Pin, Instance);
 pin_trait!(UlpiD7Pin, Instance);
 
-crate::pac::peripherals!(
+foreach_peripheral!(
     (otgfs, $inst:ident) => {
         impl sealed::Instance for peripherals::$inst {
             const REGISTERS: *const () = crate::pac::$inst.0 as *const ();
@@ -223,7 +223,7 @@ crate::pac::peripherals!(
     };
 );
 
-crate::pac::interrupts!(
+foreach_interrupt!(
     ($inst:ident, otgfs, $block:ident, GLOBAL, $irq:ident) => {
         unsafe impl USBInterrupt for crate::interrupt::$irq {}
     };
diff --git a/stm32-metapac-gen/src/assets/lib_inner.rs b/stm32-metapac-gen/src/assets/lib_inner.rs
deleted file mode 100644
index b7729cb3a..000000000
--- a/stm32-metapac-gen/src/assets/lib_inner.rs
+++ /dev/null
@@ -1,6 +0,0 @@
-// GEN PATHS HERE
-mod inner;
-
-pub mod common;
-
-pub use inner::*;
diff --git a/stm32-metapac-gen/src/lib.rs b/stm32-metapac-gen/src/lib.rs
index 575be9e4b..14625097e 100644
--- a/stm32-metapac-gen/src/lib.rs
+++ b/stm32-metapac-gen/src/lib.rs
@@ -1,8 +1,9 @@
 use chiptool::generate::CommonModule;
+use chiptool::{generate, ir, transform};
 use proc_macro2::TokenStream;
 use regex::Regex;
 use std::collections::{BTreeMap, HashMap, HashSet};
-use std::fmt::Write as _;
+use std::fmt::{Debug, Write as _};
 use std::fs;
 use std::fs::File;
 use std::io::Write;
@@ -10,9 +11,6 @@ use std::path::Path;
 use std::path::PathBuf;
 use std::str::FromStr;
 
-use chiptool::util::ToSanitizedSnakeCase;
-use chiptool::{generate, ir, transform};
-
 mod data;
 use data::*;
 
@@ -27,532 +25,322 @@ struct Metadata<'a> {
     dma_channels: &'a [DmaChannel],
 }
 
-fn make_peripheral_counts(out: &mut String, data: &BTreeMap<String, u8>) {
-    write!(
-        out,
-        "#[macro_export]
-macro_rules! peripheral_count {{
-    "
-    )
-    .unwrap();
-    for (name, count) in data {
-        write!(out, "({}) => ({});\n", name, count,).unwrap();
-    }
-    write!(out, " }}\n").unwrap();
-}
-
-fn make_dma_channel_counts(out: &mut String, data: &BTreeMap<String, u8>) {
-    if data.len() == 0 {
-        return;
-    }
-    write!(
-        out,
-        "#[macro_export]
-macro_rules! dma_channels_count {{
-    "
-    )
-    .unwrap();
-    for (name, count) in data {
-        write!(out, "({}) => ({});\n", name, count,).unwrap();
-    }
-    write!(out, " }}\n").unwrap();
-}
-
-fn make_table(out: &mut String, name: &str, data: &Vec<Vec<String>>) {
-    write!(
-        out,
-        "#[macro_export]
-macro_rules! {} {{
-    ($($pat:tt => $code:tt;)*) => {{
-        macro_rules! __{}_inner {{
-            $(($pat) => $code;)*
-            ($_:tt) => {{}}
-        }}
-",
-        name, name
-    )
-    .unwrap();
-
-    for row in data {
-        write!(out, "        __{}_inner!(({}));\n", name, row.join(",")).unwrap();
-    }
-
-    write!(
-        out,
-        "    }};
-}}"
-    )
-    .unwrap();
-}
-
 pub struct Options {
     pub chips: Vec<String>,
     pub out_dir: PathBuf,
     pub data_dir: PathBuf,
 }
 
-pub fn gen_chip(
-    options: &Options,
-    chip_core_name: &str,
-    chip: &Chip,
-    core: &Core,
-    core_index: usize,
-    all_peripheral_versions: &mut HashSet<(String, String)>,
-) {
-    let mut ir = ir::IR::new();
-
-    let mut dev = ir::Device {
-        interrupts: Vec::new(),
-        peripherals: Vec::new(),
-    };
-
-    // Load DBGMCU register for chip
-    let mut dbgmcu: Option<ir::IR> = core.peripherals.iter().find_map(|p| {
-        if p.name == "DBGMCU" {
-            p.registers.as_ref().map(|bi| {
-                let dbgmcu_reg_path = options
-                    .data_dir
-                    .join("registers")
-                    .join(&format!("{}_{}.yaml", bi.kind, bi.version));
-                serde_yaml::from_reader(File::open(dbgmcu_reg_path).unwrap()).unwrap()
-            })
-        } else {
-            None
-        }
-    });
-
-    let mut peripheral_versions: BTreeMap<String, String> = BTreeMap::new();
-    let mut pin_table: Vec<Vec<String>> = Vec::new();
-    let mut interrupt_table: Vec<Vec<String>> = Vec::new();
-    let mut peripherals_table: Vec<Vec<String>> = Vec::new();
-    let mut dma_channels_table: Vec<Vec<String>> = Vec::new();
-    let mut peripheral_counts: BTreeMap<String, u8> = BTreeMap::new();
-    let mut dma_channel_counts: BTreeMap<String, u8> = BTreeMap::new();
-    let mut dbgmcu_table: Vec<Vec<String>> = Vec::new();
-
-    let gpio_base = core
-        .peripherals
-        .iter()
-        .find(|p| p.name == "GPIOA")
-        .unwrap()
-        .address as u32;
-    let gpio_stride = 0x400;
-
-    let number_suffix_re = Regex::new("^(.*?)[0-9]*$").unwrap();
-
-    if let Some(ref mut reg) = dbgmcu {
-        if let Some(ref cr) = reg.fieldsets.get("CR") {
-            for field in cr.fields.iter().filter(|e| e.name.contains("DBG")) {
-                let mut fn_name = String::new();
-                fn_name.push_str("set_");
-                fn_name.push_str(&field.name.to_sanitized_snake_case());
-                dbgmcu_table.push(vec!["cr".into(), fn_name]);
-            }
+pub struct Gen {
+    opts: Options,
+    all_peripheral_versions: HashSet<(String, String)>,
+    metadata_dedup: HashMap<String, String>,
+}
+
+impl Gen {
+    pub fn new(opts: Options) -> Self {
+        Self {
+            opts,
+            all_peripheral_versions: HashSet::new(),
+            metadata_dedup: HashMap::new(),
         }
     }
 
-    for p in &core.peripherals {
-        let captures = number_suffix_re.captures(&p.name).unwrap();
-        let root_peri_name = captures.get(1).unwrap().as_str().to_string();
-        peripheral_counts.insert(
-            root_peri_name.clone(),
-            peripheral_counts.get(&root_peri_name).map_or(1, |v| v + 1),
-        );
-        let mut ir_peri = ir::Peripheral {
-            name: p.name.clone(),
-            array: None,
-            base_address: p.address,
-            block: None,
-            description: None,
-            interrupts: HashMap::new(),
+    fn gen_chip(&mut self, chip_core_name: &str, chip: &Chip, core: &Core, core_index: usize) {
+        let mut ir = ir::IR::new();
+
+        let mut dev = ir::Device {
+            interrupts: Vec::new(),
+            peripherals: Vec::new(),
         };
 
-        if let Some(bi) = &p.registers {
-            peripheral_counts.insert(
-                bi.kind.clone(),
-                peripheral_counts.get(&bi.kind).map_or(1, |v| v + 1),
-            );
-
-            for irq in &p.interrupts {
-                let mut row = Vec::new();
-                row.push(p.name.clone());
-                row.push(bi.kind.clone());
-                row.push(bi.block.clone());
-                row.push(irq.signal.clone());
-                row.push(irq.interrupt.to_ascii_uppercase());
-                interrupt_table.push(row)
-            }
+        let mut peripheral_versions: BTreeMap<String, String> = BTreeMap::new();
+
+        let gpio_base = core
+            .peripherals
+            .iter()
+            .find(|p| p.name == "GPIOA")
+            .unwrap()
+            .address as u32;
+        let gpio_stride = 0x400;
+
+        for p in &core.peripherals {
+            let mut ir_peri = ir::Peripheral {
+                name: p.name.clone(),
+                array: None,
+                base_address: p.address,
+                block: None,
+                description: None,
+                interrupts: HashMap::new(),
+            };
 
-            let mut peripheral_row = Vec::new();
-            peripheral_row.push(bi.kind.clone());
-            peripheral_row.push(p.name.clone());
-            peripherals_table.push(peripheral_row);
-
-            if let Some(old_version) =
-                peripheral_versions.insert(bi.kind.clone(), bi.version.clone())
-            {
-                if old_version != bi.version {
-                    panic!(
-                        "Peripheral {} has multiple versions: {} and {}",
-                        bi.kind, old_version, bi.version
-                    );
+            if let Some(bi) = &p.registers {
+                if let Some(old_version) =
+                    peripheral_versions.insert(bi.kind.clone(), bi.version.clone())
+                {
+                    if old_version != bi.version {
+                        panic!(
+                            "Peripheral {} has multiple versions: {} and {}",
+                            bi.kind, old_version, bi.version
+                        );
+                    }
                 }
-            }
-            ir_peri.block = Some(format!("{}::{}", bi.kind, bi.block));
+                ir_peri.block = Some(format!("{}::{}", bi.kind, bi.block));
 
-            match bi.kind.as_str() {
-                "gpio" => {
-                    let port_letter = p.name.chars().skip(4).next().unwrap();
+                if bi.kind == "gpio" {
                     assert_eq!(0, (p.address as u32 - gpio_base) % gpio_stride);
-                    let port_num = (p.address as u32 - gpio_base) / gpio_stride;
-
-                    for pin_num in 0u32..16 {
-                        let pin_name = format!("P{}{}", port_letter, pin_num);
-                        pin_table.push(vec![
-                            pin_name.clone(),
-                            p.name.clone(),
-                            port_num.to_string(),
-                            pin_num.to_string(),
-                            format!("EXTI{}", pin_num),
-                        ]);
-                    }
                 }
-                _ => {}
             }
-        }
-
-        dev.peripherals.push(ir_peri);
-    }
 
-    for ch in &core.dma_channels {
-        let mut row = Vec::new();
-        let dma_peri = core.peripherals.iter().find(|p| p.name == ch.dma).unwrap();
-        let bi = dma_peri.registers.as_ref().unwrap();
-
-        row.push(ch.name.clone());
-        row.push(ch.dma.clone());
-        row.push(bi.kind.clone());
-        row.push(ch.channel.to_string());
-        if let Some(dmamux) = &ch.dmamux {
-            let dmamux_channel = ch.dmamux_channel.unwrap();
-            row.push(format!(
-                "{{dmamux: {}, dmamux_channel: {}}}",
-                dmamux, dmamux_channel
-            ));
-        } else {
-            row.push("{}".to_string());
+            dev.peripherals.push(ir_peri);
         }
 
-        dma_channels_table.push(row);
-
-        let dma_peri_name = ch.dma.clone();
-        dma_channel_counts.insert(
-            dma_peri_name.clone(),
-            dma_channel_counts.get(&dma_peri_name).map_or(1, |v| v + 1),
-        );
-    }
-
-    for irq in &core.interrupts {
-        dev.interrupts.push(ir::Interrupt {
-            name: irq.name.clone(),
-            description: None,
-            value: irq.number,
-        });
-
-        let name = irq.name.to_ascii_uppercase();
-
-        interrupt_table.push(vec![name.clone()]);
-
-        if name.contains("EXTI") {
-            interrupt_table.push(vec!["EXTI".to_string(), name.clone()]);
+        for irq in &core.interrupts {
+            dev.interrupts.push(ir::Interrupt {
+                name: irq.name.clone(),
+                description: None,
+                value: irq.number,
+            });
         }
-    }
 
-    ir.devices.insert("".to_string(), dev);
+        ir.devices.insert("".to_string(), dev);
 
-    let mut extra = format!(
-        "pub fn GPIO(n: usize) -> gpio::Gpio {{
+        let mut extra = format!(
+            "pub fn GPIO(n: usize) -> gpio::Gpio {{
             gpio::Gpio(({} + {}*n) as _)
         }}",
-        gpio_base, gpio_stride,
-    );
+            gpio_base, gpio_stride,
+        );
 
-    for (module, version) in &peripheral_versions {
-        all_peripheral_versions.insert((module.clone(), version.clone()));
+        for (module, version) in &peripheral_versions {
+            self.all_peripheral_versions
+                .insert((module.clone(), version.clone()));
+            write!(
+                &mut extra,
+                "#[path=\"../../peripherals/{}_{}.rs\"] pub mod {};\n",
+                module, version, module
+            )
+            .unwrap();
+        }
         write!(
             &mut extra,
-            "#[path=\"../../peripherals/{}_{}.rs\"] pub mod {};\n",
-            module, version, module
+            "pub const CORE_INDEX: usize = {};\n",
+            core_index
         )
         .unwrap();
-    }
-    write!(
-        &mut extra,
-        "pub const CORE_INDEX: usize = {};\n",
-        core_index
-    )
-    .unwrap();
 
-    // Cleanups!
-    transform::sort::Sort {}.run(&mut ir).unwrap();
-    transform::Sanitize {}.run(&mut ir).unwrap();
+        // Cleanups!
+        transform::sort::Sort {}.run(&mut ir).unwrap();
+        transform::Sanitize {}.run(&mut ir).unwrap();
 
-    // ==============================
-    // Setup chip dir
+        // ==============================
+        // Setup chip dir
 
-    let chip_dir = options
-        .out_dir
-        .join("src/chips")
-        .join(chip_core_name.to_ascii_lowercase());
-    fs::create_dir_all(&chip_dir).unwrap();
+        let chip_dir = self
+            .opts
+            .out_dir
+            .join("src/chips")
+            .join(chip_core_name.to_ascii_lowercase());
+        fs::create_dir_all(&chip_dir).unwrap();
 
-    // ==============================
-    // generate pac.rs
+        // ==============================
+        // generate pac.rs
 
-    let data = generate::render(&ir, &gen_opts()).unwrap().to_string();
-    let data = data.replace("] ", "]\n");
+        let data = generate::render(&ir, &gen_opts()).unwrap().to_string();
+        let data = data.replace("] ", "]\n");
 
-    // Remove inner attributes like #![no_std]
-    let data = Regex::new("# *! *\\[.*\\]").unwrap().replace_all(&data, "");
+        // Remove inner attributes like #![no_std]
+        let data = Regex::new("# *! *\\[.*\\]").unwrap().replace_all(&data, "");
 
-    let mut file = File::create(chip_dir.join("pac.rs")).unwrap();
-    file.write_all(data.as_bytes()).unwrap();
-    file.write_all(extra.as_bytes()).unwrap();
+        let mut file = File::create(chip_dir.join("pac.rs")).unwrap();
+        file.write_all(data.as_bytes()).unwrap();
+        file.write_all(extra.as_bytes()).unwrap();
 
-    let mut device_x = String::new();
+        let mut device_x = String::new();
 
-    for irq in &core.interrupts {
-        write!(&mut device_x, "PROVIDE({} = DefaultHandler);\n", irq.name).unwrap();
-    }
+        for irq in &core.interrupts {
+            write!(&mut device_x, "PROVIDE({} = DefaultHandler);\n", irq.name).unwrap();
+        }
 
-    // ==============================
-    // generate mod.rs
-
-    let mut data = String::new();
-
-    write!(&mut data, "#[cfg(feature=\"metadata\")] pub mod metadata;").unwrap();
-    write!(&mut data, "#[cfg(feature=\"pac\")] mod pac;").unwrap();
-    write!(&mut data, "#[cfg(feature=\"pac\")] pub use pac::*; ").unwrap();
-
-    let peripheral_version_table = peripheral_versions
-        .iter()
-        .map(|(kind, version)| vec![kind.clone(), version.clone()])
-        .collect();
-
-    make_table(&mut data, "pins", &pin_table);
-    make_table(&mut data, "interrupts", &interrupt_table);
-    make_table(&mut data, "peripherals", &peripherals_table);
-    make_table(&mut data, "peripheral_versions", &peripheral_version_table);
-    make_table(&mut data, "dma_channels", &dma_channels_table);
-    make_table(&mut data, "dbgmcu", &dbgmcu_table);
-    make_peripheral_counts(&mut data, &peripheral_counts);
-    make_dma_channel_counts(&mut data, &dma_channel_counts);
-
-    let mut file = File::create(chip_dir.join("mod.rs")).unwrap();
-    file.write_all(data.as_bytes()).unwrap();
-
-    // ==============================
-    // generate metadata.rs
-
-    let metadata = Metadata {
-        name: &chip.name,
-        family: &chip.family,
-        line: &chip.line,
-        memory: &chip.memory,
-        peripherals: &core.peripherals,
-        interrupts: &core.interrupts,
-        dma_channels: &core.dma_channels,
-    };
-    let metadata = format!("{:#?}", metadata);
-    let metadata = metadata.replace("[\n", "&[\n");
-    let metadata = metadata.replace("[],\n", "&[],\n");
-
-    let mut data = String::new();
+        // ==============================
+        // generate metadata.rs
 
-    write!(
-        &mut data,
-        "
-            include!(\"../../metadata.rs\");
-            use MemoryRegionKind::*;
-            pub const METADATA: Metadata = {};    
-        ",
-        metadata
-    )
-    .unwrap();
+        // (peripherals, interrupts, dma_channels) are often equal across multiple chips.
+        // To reduce bloat, deduplicate them.
+        let mut data = String::new();
+        write!(
+            &mut data,
+            "
+                const PERIPHERALS: &'static [Peripheral] = {};
+                const INTERRUPTS: &'static [Interrupt] = {};
+                const DMA_CHANNELS: &'static [DmaChannel] = {};
+            ",
+            stringify(&core.peripherals),
+            stringify(&core.interrupts),
+            stringify(&core.dma_channels),
+        )
+        .unwrap();
 
-    let mut file = File::create(chip_dir.join("metadata.rs")).unwrap();
-    file.write_all(data.as_bytes()).unwrap();
+        let out_dir = self.opts.out_dir.clone();
+        let n = self.metadata_dedup.len();
+        let deduped_file = self.metadata_dedup.entry(data.clone()).or_insert_with(|| {
+            let file = format!("metadata_{:04}.rs", n);
+            let path = out_dir.join("src/chips").join(&file);
+            fs::write(path, data).unwrap();
 
-    // ==============================
-    // generate device.x
+            file
+        });
 
-    File::create(chip_dir.join("device.x"))
-        .unwrap()
-        .write_all(device_x.as_bytes())
-        .unwrap();
+        let data = format!(
+            "include!(\"../{}\");
+            pub const METADATA: Metadata = Metadata {{
+                name: {:?},
+                family: {:?},
+                line: {:?},
+                memory: {},
+                peripherals: PERIPHERALS,
+                interrupts: INTERRUPTS,
+                dma_channels: DMA_CHANNELS,
+            }};",
+            deduped_file,
+            &chip.name,
+            &chip.family,
+            &chip.line,
+            stringify(&chip.memory),
+        );
 
-    // ==============================
-    // generate default memory.x
-    gen_memory_x(&chip_dir, &chip);
-}
+        let mut file = File::create(chip_dir.join("metadata.rs")).unwrap();
+        file.write_all(data.as_bytes()).unwrap();
 
-fn load_chip(options: &Options, name: &str) -> Chip {
-    let chip_path = options
-        .data_dir
-        .join("chips")
-        .join(&format!("{}.json", name));
-    let chip = fs::read(chip_path).expect(&format!("Could not load chip {}", name));
-    serde_yaml::from_slice(&chip).unwrap()
-}
+        // ==============================
+        // generate device.x
 
-fn gen_opts() -> generate::Options {
-    generate::Options {
-        common_module: CommonModule::External(TokenStream::from_str("crate::common").unwrap()),
+        File::create(chip_dir.join("device.x"))
+            .unwrap()
+            .write_all(device_x.as_bytes())
+            .unwrap();
+
+        // ==============================
+        // generate default memory.x
+        gen_memory_x(&chip_dir, &chip);
     }
-}
 
-pub fn gen(options: Options) {
-    fs::create_dir_all(options.out_dir.join("src/peripherals")).unwrap();
-    fs::create_dir_all(options.out_dir.join("src/chips")).unwrap();
+    fn load_chip(&mut self, name: &str) -> Chip {
+        let chip_path = self
+            .opts
+            .data_dir
+            .join("chips")
+            .join(&format!("{}.json", name));
+        let chip = fs::read(chip_path).expect(&format!("Could not load chip {}", name));
+        serde_yaml::from_slice(&chip).unwrap()
+    }
 
-    let mut all_peripheral_versions: HashSet<(String, String)> = HashSet::new();
-    let mut chip_core_names: Vec<String> = Vec::new();
+    pub fn gen(&mut self) {
+        fs::create_dir_all(self.opts.out_dir.join("src/peripherals")).unwrap();
+        fs::create_dir_all(self.opts.out_dir.join("src/chips")).unwrap();
 
-    for chip_name in &options.chips {
-        println!("Generating {}...", chip_name);
+        let mut chip_core_names: Vec<String> = Vec::new();
 
-        let mut chip = load_chip(&options, chip_name);
+        for chip_name in &self.opts.chips.clone() {
+            println!("Generating {}...", chip_name);
 
-        // Cleanup
-        for core in &mut chip.cores {
-            for irq in &mut core.interrupts {
-                irq.name = irq.name.to_ascii_uppercase();
-            }
-            for p in &mut core.peripherals {
-                for irq in &mut p.interrupts {
-                    irq.interrupt = irq.interrupt.to_ascii_uppercase();
+            let mut chip = self.load_chip(chip_name);
+
+            // Cleanup
+            for core in &mut chip.cores {
+                for irq in &mut core.interrupts {
+                    irq.name = irq.name.to_ascii_uppercase();
+                }
+                for p in &mut core.peripherals {
+                    for irq in &mut p.interrupts {
+                        irq.interrupt = irq.interrupt.to_ascii_uppercase();
+                    }
                 }
             }
-        }
 
-        // Generate
-        for (core_index, core) in chip.cores.iter().enumerate() {
-            let chip_core_name = match chip.cores.len() {
-                1 => chip_name.clone(),
-                _ => format!("{}-{}", chip_name, core.name),
-            };
+            // Generate
+            for (core_index, core) in chip.cores.iter().enumerate() {
+                let chip_core_name = match chip.cores.len() {
+                    1 => chip_name.clone(),
+                    _ => format!("{}-{}", chip_name, core.name),
+                };
 
-            chip_core_names.push(chip_core_name.clone());
-            gen_chip(
-                &options,
-                &chip_core_name,
-                &chip,
-                core,
-                core_index,
-                &mut all_peripheral_versions,
-            )
+                chip_core_names.push(chip_core_name.clone());
+                self.gen_chip(&chip_core_name, &chip, core, core_index)
+            }
         }
-    }
 
-    for (module, version) in all_peripheral_versions {
-        println!("loading {} {}", module, version);
+        for (module, version) in &self.all_peripheral_versions {
+            println!("loading {} {}", module, version);
 
-        let regs_path = Path::new(&options.data_dir)
-            .join("registers")
-            .join(&format!("{}_{}.yaml", module, version));
+            let regs_path = Path::new(&self.opts.data_dir)
+                .join("registers")
+                .join(&format!("{}_{}.yaml", module, version));
 
-        let mut ir: ir::IR = serde_yaml::from_reader(File::open(regs_path).unwrap()).unwrap();
+            let mut ir: ir::IR = serde_yaml::from_reader(File::open(regs_path).unwrap()).unwrap();
 
-        transform::expand_extends::ExpandExtends {}
-            .run(&mut ir)
-            .unwrap();
+            transform::expand_extends::ExpandExtends {}
+                .run(&mut ir)
+                .unwrap();
 
-        transform::map_names(&mut ir, |k, s| match k {
-            transform::NameKind::Block => *s = format!("{}", s),
-            transform::NameKind::Fieldset => *s = format!("regs::{}", s),
-            transform::NameKind::Enum => *s = format!("vals::{}", s),
-            _ => {}
-        });
+            transform::map_names(&mut ir, |k, s| match k {
+                transform::NameKind::Block => *s = format!("{}", s),
+                transform::NameKind::Fieldset => *s = format!("regs::{}", s),
+                transform::NameKind::Enum => *s = format!("vals::{}", s),
+                _ => {}
+            });
 
-        transform::sort::Sort {}.run(&mut ir).unwrap();
-        transform::Sanitize {}.run(&mut ir).unwrap();
+            transform::sort::Sort {}.run(&mut ir).unwrap();
+            transform::Sanitize {}.run(&mut ir).unwrap();
 
-        let items = generate::render(&ir, &gen_opts()).unwrap();
-        let mut file = File::create(
-            options
-                .out_dir
-                .join("src/peripherals")
-                .join(format!("{}_{}.rs", module, version)),
-        )
-        .unwrap();
-        let data = items.to_string().replace("] ", "]\n");
+            let items = generate::render(&ir, &gen_opts()).unwrap();
+            let mut file = File::create(
+                self.opts
+                    .out_dir
+                    .join("src/peripherals")
+                    .join(format!("{}_{}.rs", module, version)),
+            )
+            .unwrap();
+            let data = items.to_string().replace("] ", "]\n");
 
-        // Remove inner attributes like #![no_std]
-        let re = Regex::new("# *! *\\[.*\\]").unwrap();
-        let data = re.replace_all(&data, "");
-        file.write_all(data.as_bytes()).unwrap();
-    }
+            // Remove inner attributes like #![no_std]
+            let re = Regex::new("# *! *\\[.*\\]").unwrap();
+            let data = re.replace_all(&data, "");
+            file.write_all(data.as_bytes()).unwrap();
+        }
 
-    // Generate src/lib_inner.rs
-    const PATHS_MARKER: &[u8] = b"// GEN PATHS HERE";
-    let librs = include_bytes!("assets/lib_inner.rs");
-    let i = bytes_find(librs, PATHS_MARKER).unwrap();
-    let mut paths = String::new();
+        // Generate Cargo.toml
+        const BUILDDEP_BEGIN: &[u8] = b"# BEGIN BUILD DEPENDENCIES";
+        const BUILDDEP_END: &[u8] = b"# END BUILD DEPENDENCIES";
 
-    for name in chip_core_names {
-        let x = name.to_ascii_lowercase();
-        write!(
-            &mut paths,
-            "#[cfg_attr(feature=\"{}\", path = \"chips/{}/mod.rs\")]",
-            x, x
+        let mut contents = include_bytes!("../../stm32-metapac/Cargo.toml").to_vec();
+        let begin = bytes_find(&contents, BUILDDEP_BEGIN).unwrap();
+        let end = bytes_find(&contents, BUILDDEP_END).unwrap() + BUILDDEP_END.len();
+        contents.drain(begin..end);
+        fs::write(self.opts.out_dir.join("Cargo.toml"), contents).unwrap();
+
+        // copy misc files
+        fs::write(
+            self.opts.out_dir.join("build.rs"),
+            include_bytes!("../../stm32-metapac/build_pregenerated.rs"),
+        )
+        .unwrap();
+        fs::write(
+            self.opts.out_dir.join("src/lib.rs"),
+            include_bytes!("../../stm32-metapac/src/lib.rs"),
+        )
+        .unwrap();
+        fs::write(
+            self.opts.out_dir.join("src/common.rs"),
+            include_bytes!("../../stm32-metapac/src/common.rs"),
+        )
+        .unwrap();
+        fs::write(
+            self.opts.out_dir.join("src/metadata.rs"),
+            include_bytes!("../../stm32-metapac/src/metadata.rs"),
         )
         .unwrap();
     }
-    let mut contents: Vec<u8> = Vec::new();
-    contents.extend(&librs[..i]);
-    contents.extend(paths.as_bytes());
-    contents.extend(&librs[i + PATHS_MARKER.len()..]);
-    fs::write(options.out_dir.join("src").join("lib_inner.rs"), &contents).unwrap();
-
-    // Generate src/lib.rs
-    const CUT_MARKER: &[u8] = b"// GEN CUT HERE";
-    let librs = include_bytes!("../../stm32-metapac/src/lib.rs");
-    let i = bytes_find(librs, CUT_MARKER).unwrap();
-    let mut contents: Vec<u8> = Vec::new();
-    contents.extend(&librs[..i]);
-    contents.extend(b"include!(\"lib_inner.rs\");\n");
-    fs::write(options.out_dir.join("src").join("lib.rs"), contents).unwrap();
-
-    // Generate src/common.rs
-    fs::write(
-        options.out_dir.join("src").join("common.rs"),
-        generate::COMMON_MODULE,
-    )
-    .unwrap();
-
-    // Generate src/metadata.rs
-    fs::write(
-        options.out_dir.join("src").join("metadata.rs"),
-        include_bytes!("assets/metadata.rs"),
-    )
-    .unwrap();
-
-    // Generate Cargo.toml
-    const BUILDDEP_BEGIN: &[u8] = b"# BEGIN BUILD DEPENDENCIES";
-    const BUILDDEP_END: &[u8] = b"# END BUILD DEPENDENCIES";
-
-    let mut contents = include_bytes!("../../stm32-metapac/Cargo.toml").to_vec();
-    let begin = bytes_find(&contents, BUILDDEP_BEGIN).unwrap();
-    let end = bytes_find(&contents, BUILDDEP_END).unwrap() + BUILDDEP_END.len();
-    contents.drain(begin..end);
-    fs::write(options.out_dir.join("Cargo.toml"), contents).unwrap();
-
-    // Generate build.rs
-    fs::write(
-        options.out_dir.join("build.rs"),
-        include_bytes!("assets/build.rs"),
-    )
-    .unwrap();
 }
 
 fn bytes_find(haystack: &[u8], needle: &[u8]) -> Option<usize> {
@@ -561,6 +349,23 @@ fn bytes_find(haystack: &[u8], needle: &[u8]) -> Option<usize> {
         .position(|window| window == needle)
 }
 
+fn stringify<T: Debug>(metadata: T) -> String {
+    let mut metadata = format!("{:?}", metadata);
+    if metadata.starts_with('[') {
+        metadata = format!("&{}", metadata);
+    }
+    metadata = metadata.replace(": [", ": &[");
+    metadata = metadata.replace("kind: Ram", "kind: MemoryRegionKind::Ram");
+    metadata = metadata.replace("kind: Flash", "kind: MemoryRegionKind::Flash");
+    metadata
+}
+
+fn gen_opts() -> generate::Options {
+    generate::Options {
+        common_module: CommonModule::External(TokenStream::from_str("crate::common").unwrap()),
+    }
+}
+
 fn gen_memory_x(out_dir: &PathBuf, chip: &Chip) {
     let mut memory_x = String::new();
 
diff --git a/stm32-metapac-gen/src/main.rs b/stm32-metapac-gen/src/main.rs
index fd1e2a060..391441302 100644
--- a/stm32-metapac-gen/src/main.rs
+++ b/stm32-metapac-gen/src/main.rs
@@ -26,9 +26,10 @@ fn main() {
 
     chips.sort();
 
-    gen(Options {
+    let opts = Options {
         out_dir,
         data_dir,
         chips,
-    })
+    };
+    Gen::new(opts).gen();
 }
diff --git a/stm32-metapac/Cargo.toml b/stm32-metapac/Cargo.toml
index 5642af46e..c994797b7 100644
--- a/stm32-metapac/Cargo.toml
+++ b/stm32-metapac/Cargo.toml
@@ -3,6 +3,21 @@ name = "stm32-metapac"
 version = "0.1.0"
 edition = "2018"
 resolver = "2"
+license = "MIT OR Apache-2.0"
+repository = "https://github.com/embassy-rs/embassy"
+description = "Peripheral Access Crate (PAC) for all STM32 chips, including metadata."
+
+# `cargo publish` is unable to figure out which .rs files are needed due to the include! magic.
+include = [
+    "**/*.rs",
+    "**/*.x",
+    "Cargo.toml",
+]
+
+[package.metadata.docs.rs]
+features = ["stm32h755zi-cm7", "pac", "metadata"]
+default-target = "thumbv7em-none-eabihf"
+targets = []
 
 [dependencies]
 cortex-m = "0.7.3"
diff --git a/stm32-metapac/build.rs b/stm32-metapac/build.rs
index 7fd5a6b1f..44c10eced 100644
--- a/stm32-metapac/build.rs
+++ b/stm32-metapac/build.rs
@@ -6,7 +6,7 @@ fn parse_chip_core(chip_and_core: &str) -> (String, Option<String>) {
     let mut s = chip_and_core.split('-');
     let chip_name: String = s.next().unwrap().to_string();
     if let Some(c) = s.next() {
-        if c.starts_with("CM") {
+        if c.starts_with("cm") {
             return (chip_name, Some(c.to_ascii_lowercase()));
         }
     }
@@ -18,36 +18,45 @@ fn main() {
     let out_dir = PathBuf::from(env::var_os("OUT_DIR").unwrap());
     let data_dir = PathBuf::from("../stm32-data/data");
 
-    println!("cwd: {:?}", env::current_dir());
-
     let chip_core_name = env::vars_os()
         .map(|(a, _)| a.to_string_lossy().to_string())
         .find(|x| x.starts_with("CARGO_FEATURE_STM32"))
         .expect("No stm32xx Cargo feature enabled")
         .strip_prefix("CARGO_FEATURE_")
         .unwrap()
-        .to_ascii_uppercase()
+        .to_ascii_lowercase()
         .replace('_', "-");
 
     let (chip_name, _) = parse_chip_core(&chip_core_name);
 
-    gen(Options {
+    let opts = Options {
         out_dir: out_dir.clone(),
         data_dir: data_dir.clone(),
-        chips: vec![chip_name],
-    });
+        chips: vec![chip_name.to_ascii_uppercase()],
+    };
+    Gen::new(opts).gen();
 
     println!(
         "cargo:rustc-link-search={}/src/chips/{}",
         out_dir.display(),
-        chip_core_name.to_ascii_lowercase()
+        chip_core_name,
     );
 
     #[cfg(feature = "memory-x")]
     println!(
         "cargo:rustc-link-search={}/src/chips/{}/memory_x/",
         out_dir.display(),
-        chip_core_name.to_ascii_lowercase()
+        chip_core_name
+    );
+    println!(
+        "cargo:rustc-env=STM32_METAPAC_PAC_PATH={}/src/chips/{}/pac.rs",
+        out_dir.display(),
+        chip_core_name
+    );
+    println!(
+        "cargo:rustc-env=STM32_METAPAC_METADATA_PATH={}/src/chips/{}/metadata.rs",
+        out_dir.display(),
+        chip_core_name
     );
 
     println!("cargo:rerun-if-changed=build.rs");
diff --git a/stm32-metapac-gen/src/assets/build.rs b/stm32-metapac/build_pregenerated.rs
index 14d041ff6..2219acb55 100644
--- a/stm32-metapac-gen/src/assets/build.rs
+++ b/stm32-metapac/build_pregenerated.rs
@@ -23,7 +23,15 @@ fn main() {
     println!(
         "cargo:rustc-link-search={}/src/chips/{}/memory_x/",
         crate_dir.display(),
-        chip_core_name,
+        chip_core_name
+    );
+    println!(
+        "cargo:rustc-env=STM32_METAPAC_PAC_PATH=chips/{}/pac.rs",
+        chip_core_name
+    );
+    println!(
+        "cargo:rustc-env=STM32_METAPAC_METADATA_PATH=chips/{}/metadata.rs",
+        chip_core_name
     );
 
     println!("cargo:rerun-if-changed=build.rs");
diff --git a/stm32-metapac/src/common.rs b/stm32-metapac/src/common.rs
new file mode 100644
index 000000000..568a98486
--- /dev/null
+++ b/stm32-metapac/src/common.rs
@@ -0,0 +1,80 @@
+use core::marker::PhantomData;
+
+#[derive(Copy, Clone, PartialEq, Eq)]
+pub struct RW;
+#[derive(Copy, Clone, PartialEq, Eq)]
+pub struct R;
+#[derive(Copy, Clone, PartialEq, Eq)]
+pub struct W;
+
+mod sealed {
+    use super::*;
+    pub trait Access {}
+    impl Access for R {}
+    impl Access for W {}
+    impl Access for RW {}
+}
+
+pub trait Access: sealed::Access + Copy {}
+impl Access for R {}
+impl Access for W {}
+impl Access for RW {}
+
+pub trait Read: Access {}
+impl Read for RW {}
+impl Read for R {}
+
+pub trait Write: Access {}
+impl Write for RW {}
+impl Write for W {}
+
+#[derive(Copy, Clone, PartialEq, Eq)]
+pub struct Reg<T: Copy, A: Access> {
+    ptr: *mut u8,
+    phantom: PhantomData<*mut (T, A)>,
+}
+unsafe impl<T: Copy, A: Access> Send for Reg<T, A> {}
+unsafe impl<T: Copy, A: Access> Sync for Reg<T, A> {}
+
+impl<T: Copy, A: Access> Reg<T, A> {
+    pub fn from_ptr(ptr: *mut u8) -> Self {
+        Self {
+            ptr,
+            phantom: PhantomData,
+        }
+    }
+
+    pub fn ptr(&self) -> *mut T {
+        self.ptr as _
+    }
+}
+
+impl<T: Copy, A: Read> Reg<T, A> {
+    pub unsafe fn read(&self) -> T {
+        (self.ptr as *mut T).read_volatile()
+    }
+}
+
+impl<T: Copy, A: Write> Reg<T, A> {
+    pub unsafe fn write_value(&self, val: T) {
+        (self.ptr as *mut T).write_volatile(val)
+    }
+}
+
+impl<T: Default + Copy, A: Write> Reg<T, A> {
+    pub unsafe fn write<R>(&self, f: impl FnOnce(&mut T) -> R) -> R {
+        let mut val = Default::default();
+        let res = f(&mut val);
+        self.write_value(val);
+        res
+    }
+}
+
+impl<T: Copy, A: Read + Write> Reg<T, A> {
+    pub unsafe fn modify<R>(&self, f: impl FnOnce(&mut T) -> R) -> R {
+        let mut val = self.read();
+        let res = f(&mut val);
+        self.write_value(val);
+        res
+    }
+}
diff --git a/stm32-metapac/src/lib.rs b/stm32-metapac/src/lib.rs
index 5dd2682fb..9cdc5e0b4 100644
--- a/stm32-metapac/src/lib.rs
+++ b/stm32-metapac/src/lib.rs
@@ -3,5 +3,13 @@
 #![allow(unused)]
 #![allow(non_camel_case_types)]
 
-// GEN CUT HERE
-include!(concat!(env!("OUT_DIR"), "/src/lib_inner.rs"));
+pub mod common;
+
+#[cfg(feature = "pac")]
+include!(env!("STM32_METAPAC_PAC_PATH"));
+
+#[cfg(feature = "metadata")]
+pub mod metadata {
+    include!("metadata.rs");
+    include!(env!("STM32_METAPAC_METADATA_PATH"));
+}
diff --git a/stm32-metapac-gen/src/assets/metadata.rs b/stm32-metapac/src/metadata.rs
index 150aa5956..150aa5956 100644
--- a/stm32-metapac-gen/src/assets/metadata.rs
+++ b/stm32-metapac/src/metadata.rs