Merge pull request #288 from Sizurka/rom-intrinsics

Connect Bootrom Instrinsics

rp2040-hal/Cargo.toml

@@ -54,3 +54,5 @@ rt = ["rp2040-pac/rt"]
 # embassy-traits = ["embassy_traits", "futures"]
 alloc = []
 rom-func-cache = []
+disable-intrinsics = []
+rom-v2-intrinsics = []

rp2040-hal/src/float/add_sub.rs

@@ -0,0 +1,89 @@
+use super::{Float, Int};
+use crate::rom_data;
+
+trait ROMAdd {
+    fn rom_add(self, b: Self) -> Self;
+}
+
+impl ROMAdd for f32 {
+    fn rom_add(self, b: Self) -> Self {
+        rom_data::float_funcs::fadd(self, b)
+    }
+}
+
+impl ROMAdd for f64 {
+    fn rom_add(self, b: Self) -> Self {
+        rom_data::double_funcs::dadd(self, b)
+    }
+}
+
+fn add<F: Float + ROMAdd>(a: F, b: F) -> F {
+    if a.is_not_finite() {
+        if b.is_not_finite() {
+            let class_a = a.repr() & (F::SIGNIFICAND_MASK | F::SIGN_MASK);
+            let class_b = b.repr() & (F::SIGNIFICAND_MASK | F::SIGN_MASK);
+
+            if class_a == F::Int::ZERO && class_b == F::Int::ZERO {
+                // inf + inf = inf
+                return a;
+            }
+            if class_a == F::SIGN_MASK && class_b == F::SIGN_MASK {
+                // -inf + (-inf) = -inf
+                return a;
+            }
+
+            // Sign mismatch, or either is NaN already
+            return F::NAN;
+        }
+
+        // [-]inf/NaN + X = [-]inf/NaN
+        return a;
+    }
+
+    if b.is_not_finite() {
+        // X + [-]inf/NaN = [-]inf/NaN
+        return b;
+    }
+
+    a.rom_add(b)
+}
+
+intrinsics! {
+    #[alias = __addsf3vfp]
+    #[aeabi = __aeabi_fadd]
+    extern "C" fn __addsf3(a: f32, b: f32) -> f32 {
+        add(a, b)
+    }
+
+    #[bootrom_v2]
+    #[alias = __adddf3vfp]
+    #[aeabi = __aeabi_dadd]
+    extern "C" fn __adddf3(a: f64, b: f64) -> f64 {
+        add(a, b)
+    }
+
+    // The ROM just implements subtraction the same way, so just do it here
+    // and save the work of implementing more complicated NaN/inf handling.
+
+    #[alias = __subsf3vfp]
+    #[aeabi = __aeabi_fsub]
+    extern "C" fn __subsf3(a: f32, b: f32) -> f32 {
+        add(a, -b)
+    }
+
+    #[bootrom_v2]
+    #[alias = __subdf3vfp]
+    #[aeabi = __aeabi_dsub]
+    extern "C" fn __subdf3(a: f64, b: f64) -> f64 {
+        add(a, -b)
+    }
+
+    extern "aapcs" fn __aeabi_frsub(a: f32, b: f32) -> f32 {
+        add(b, -a)
+    }
+
+    #[bootrom_v2]
+    extern "aapcs" fn __aeabi_drsub(a: f64, b: f64) -> f64 {
+        add(b, -a)
+    }
+}

rp2040-hal/src/float/cmp.rs

@@ -0,0 +1,198 @@
+use super::Float;
+use crate::rom_data;
+
+trait ROMCmp {
+    fn rom_cmp(self, b: Self) -> i32;
+}
+
+impl ROMCmp for f32 {
+    fn rom_cmp(self, b: Self) -> i32 {
+        rom_data::float_funcs::fcmp(self, b)
+    }
+}
+
+impl ROMCmp for f64 {
+    fn rom_cmp(self, b: Self) -> i32 {
+        rom_data::double_funcs::dcmp(self, b)
+    }
+}
+
+fn le_abi<F: Float + ROMCmp>(a: F, b: F) -> i32 {
+    if a.is_nan() || b.is_nan() {
+        1
+    } else {
+        a.rom_cmp(b)
+    }
+}
+
+fn ge_abi<F: Float + ROMCmp>(a: F, b: F) -> i32 {
+    if a.is_nan() || b.is_nan() {
+        -1
+    } else {
+        a.rom_cmp(b)
+    }
+}
+
+intrinsics! {
+    #[slower_than_default]
+    #[bootrom_v2]
+    #[alias = __eqsf2, __ltsf2, __nesf2]
+    extern "C" fn __lesf2(a: f32, b: f32) -> i32 {
+        le_abi(a, b)
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    #[alias = __eqdf2, __ltdf2, __nedf2]
+    extern "C" fn __ledf2(a: f64, b: f64) -> i32 {
+        le_abi(a, b)
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    #[alias = __gtsf2]
+    extern "C" fn __gesf2(a: f32, b: f32) -> i32 {
+        ge_abi(a, b)
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    #[alias = __gtdf2]
+    extern "C" fn __gedf2(a: f64, b: f64) -> i32 {
+        ge_abi(a, b)
+    }
+
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "aapcs" fn __aeabi_fcmple(a: f32, b: f32) -> i32 {
+        (le_abi(a, b) <= 0) as i32
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "aapcs" fn __aeabi_fcmpge(a: f32, b: f32) -> i32 {
+        (ge_abi(a, b) >= 0) as i32
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "aapcs" fn __aeabi_fcmpeq(a: f32, b: f32) -> i32 {
+        (le_abi(a, b) == 0) as i32
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "aapcs" fn __aeabi_fcmplt(a: f32, b: f32) -> i32 {
+        (le_abi(a, b) < 0) as i32
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "aapcs" fn __aeabi_fcmpgt(a: f32, b: f32) -> i32 {
+        (ge_abi(a, b) > 0) as i32
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "aapcs" fn __aeabi_dcmple(a: f64, b: f64) -> i32 {
+        (le_abi(a, b) <= 0) as i32
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "aapcs" fn __aeabi_dcmpge(a: f64, b: f64) -> i32 {
+        (ge_abi(a, b) >= 0) as i32
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "aapcs" fn __aeabi_dcmpeq(a: f64, b: f64) -> i32 {
+        (le_abi(a, b) == 0) as i32
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "aapcs" fn __aeabi_dcmplt(a: f64, b: f64) -> i32 {
+        (le_abi(a, b) < 0) as i32
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "aapcs" fn __aeabi_dcmpgt(a: f64, b: f64) -> i32 {
+        (ge_abi(a, b) > 0) as i32
+    }
+
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "C" fn __gesf2vfp(a: f32, b: f32) -> i32 {
+        (ge_abi(a, b) >= 0) as i32
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "C" fn __gedf2vfp(a: f64, b: f64) -> i32 {
+        (ge_abi(a, b) >= 0) as i32
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "C" fn __gtsf2vfp(a: f32, b: f32) -> i32 {
+        (ge_abi(a, b) > 0) as i32
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "C" fn __gtdf2vfp(a: f64, b: f64) -> i32 {
+        (ge_abi(a, b) > 0) as i32
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "C" fn __ltsf2vfp(a: f32, b: f32) -> i32 {
+        (le_abi(a, b) < 0) as i32
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "C" fn __ltdf2vfp(a: f64, b: f64) -> i32 {
+        (le_abi(a, b) < 0) as i32
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "C" fn __lesf2vfp(a: f32, b: f32) -> i32 {
+        (le_abi(a, b) <= 0) as i32
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "C" fn __ledf2vfp(a: f64, b: f64) -> i32 {
+        (le_abi(a, b) <= 0) as i32
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "C" fn __nesf2vfp(a: f32, b: f32) -> i32 {
+        (le_abi(a, b) != 0) as i32
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "C" fn __nedf2vfp(a: f64, b: f64) -> i32 {
+        (le_abi(a, b) != 0) as i32
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "C" fn __eqsf2vfp(a: f32, b: f32) -> i32 {
+        (le_abi(a, b) == 0) as i32
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    extern "C" fn __eqdf2vfp(a: f64, b: f64) -> i32 {
+        (le_abi(a, b) == 0) as i32
+    }
+}

rp2040-hal/src/float/conv.rs

@@ -0,0 +1,154 @@
+use super::Float;
+use crate::rom_data;
+
+// Some of these are also not connected in the Pico SDK.  This is probably
+// because the ROM version actually does a fixed point conversion, just with
+// the fractional width set to zero.
+
+intrinsics! {
+    // Not connected in the Pico SDK
+    #[slower_than_default]
+    #[aeabi = __aeabi_i2f]
+    extern "C" fn __floatsisf(i: i32) -> f32 {
+        rom_data::float_funcs::int_to_float(i)
+    }
+
+    // Not connected in the Pico SDK
+    #[slower_than_default]
+    #[aeabi = __aeabi_i2d]
+    extern "C" fn __floatsidf(i: i32) -> f64 {
+        rom_data::double_funcs::int_to_double(i)
+    }
+
+    // Questionable gain
+    #[aeabi = __aeabi_l2f]
+    extern "C" fn __floatdisf(i: i64) -> f32 {
+        rom_data::float_funcs::int64_to_float(i)
+    }
+
+    #[bootrom_v2]
+    #[aeabi = __aeabi_l2d]
+    extern "C" fn __floatdidf(i: i64) -> f64 {
+        rom_data::double_funcs::int64_to_double(i)
+    }
+
+    // Not connected in the Pico SDK
+    #[slower_than_default]
+    #[aeabi = __aeabi_ui2f]
+    extern "C" fn __floatunsisf(i: u32) -> f32 {
+        rom_data::float_funcs::uint_to_float(i)
+    }
+
+    // Questionable gain
+    #[bootrom_v2]
+    #[aeabi = __aeabi_ui2d]
+    extern "C" fn __floatunsidf(i: u32) -> f64 {
+        rom_data::double_funcs::uint_to_double(i)
+    }
+
+    // Questionable gain
+    #[bootrom_v2]
+    #[aeabi = __aeabi_ul2f]
+    extern "C" fn __floatundisf(i: u64) -> f32 {
+        rom_data::float_funcs::uint64_to_float(i)
+    }
+
+    #[bootrom_v2]
+    #[aeabi = __aeabi_ul2d]
+    extern "C" fn __floatundidf(i: u64) -> f64 {
+        rom_data::double_funcs::uint64_to_double(i)
+    }
+
+
+    // The Pico SDK does some optimization here (e.x. fast paths for zero and
+    // one), but we can just directly connect it.
+    #[aeabi = __aeabi_f2iz]
+    extern "C" fn __fixsfsi(f: f32) -> i32 {
+        rom_data::float_funcs::float_to_int(f)
+    }
+
+    #[bootrom_v2]
+    #[aeabi = __aeabi_f2lz]
+    extern "C" fn __fixsfdi(f: f32) -> i64 {
+        rom_data::float_funcs::float_to_int64(f)
+    }
+
+    // Not connected in the Pico SDK
+    #[slower_than_default]
+    #[bootrom_v2]
+    #[aeabi = __aeabi_d2iz]
+    extern "C" fn __fixdfsi(f: f64) -> i32 {
+        rom_data::double_funcs::double_to_int(f)
+    }
+
+    // Like with the 32 bit version, there's optimization that we just
+    // skip.
+    #[bootrom_v2]
+    #[aeabi = __aeabi_d2lz]
+    extern "C" fn __fixdfdi(f: f64) -> i64 {
+        rom_data::double_funcs::double_to_int64(f)
+    }
+
+    #[slower_than_default]
+    #[aeabi = __aeabi_f2uiz]
+    extern "C" fn __fixunssfsi(f: f32) -> u32 {
+        rom_data::float_funcs::float_to_uint(f)
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    #[aeabi = __aeabi_f2ulz]
+    extern "C" fn __fixunssfdi(f: f32) -> u64 {
+        rom_data::float_funcs::float_to_uint64(f)
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    #[aeabi = __aeabi_d2uiz]
+    extern "C" fn __fixunsdfsi(f: f64) -> u32 {
+        rom_data::double_funcs::double_to_uint(f)
+    }
+
+    #[slower_than_default]
+    #[bootrom_v2]
+    #[aeabi = __aeabi_d2ulz]
+    extern "C" fn __fixunsdfdi(f: f64) -> u64 {
+        rom_data::double_funcs::double_to_uint64(f)
+    }
+
+    #[bootrom_v2]
+    #[alias = __extendsfdf2vfp]
+    #[aeabi = __aeabi_f2d]
+    extern "C" fn  __extendsfdf2(f: f32) -> f64 {
+        if f.is_not_finite() {
+            return f64::from_repr(
+                // Not finite
+                f64::EXPONENT_MASK |
+                // Preserve NaN or inf
+                ((f.repr() & f32::SIGNIFICAND_MASK) as u64) |
+                // Preserve sign
+                ((f.repr() & f32::SIGN_MASK) as u64) << (f64::BITS-f32::BITS)
+            );
+        }
+        rom_data::float_funcs::float_to_double(f)
+    }
+
+    #[bootrom_v2]
+    #[alias = __truncdfsf2vfp]
+    #[aeabi = __aeabi_d2f]
+    extern "C" fn __truncdfsf2(f: f64) -> f32 {
+        if f.is_not_finite() {
+            let mut repr: u32 =
+                // Not finite
+                f32::EXPONENT_MASK |
+                // Preserve sign
+                ((f.repr() & f64::SIGN_MASK) >> (f64::BITS-f32::BITS)) as u32;
+            // Set NaN
+            if  (f.repr() & f64::SIGNIFICAND_MASK) != 0 {
+                repr |= 1;
+            }
+            return f32::from_repr(repr);
+        }
+        rom_data::double_funcs::double_to_float(f)
+    }
+}

rp2040-hal/src/float/div.rs

@@ -0,0 +1,71 @@
+use super::Float;
+use crate::rom_data;
+use crate::sio::save_divider;
+
+trait ROMDiv {
+    fn rom_div(self, b: Self) -> Self;
+}
+
+impl ROMDiv for f32 {
+    fn rom_div(self, b: Self) -> Self {
+        // ROM implementation uses the hardware divider, so we have to save it
+        save_divider(|_sio| rom_data::float_funcs::fdiv(self, b))
+    }
+}
+
+impl ROMDiv for f64 {
+    fn rom_div(self, b: Self) -> Self {
+        // ROM implementation uses the hardware divider, so we have to save it
+        save_divider(|_sio| rom_data::double_funcs::ddiv(self, b))
+    }
+}
+
+fn div<F: Float + ROMDiv>(a: F, b: F) -> F {
+    if a.is_not_finite() {
+        if b.is_not_finite() {
+            // inf/NaN / inf/NaN = NaN
+            return F::NAN;
+        }
+
+        if b.is_zero() {
+            // inf/NaN / 0 = NaN
+            return F::NAN;
+        }
+
+        return if b.is_sign_negative() {
+            // [+/-]inf/NaN / (-X) = [-/+]inf/NaN
+            a.negate()
+        } else {
+            // [-]inf/NaN / X = [-]inf/NaN
+            a
+        };
+    }
+
+    if b.is_nan() {
+        // X / NaN = NaN
+        return b;
+    }
+
+    // ROM handles X / 0 = [-]inf and X / [-]inf = [-]0, so we only
+    // need to catch 0 / 0
+    if b.is_zero() && a.is_zero() {
+        return F::NAN;
+    }
+
+    a.rom_div(b)
+}
+
+intrinsics! {
+    #[alias = __divsf3vfp]
+    #[aeabi = __aeabi_fdiv]
+    extern "C" fn __divsf3(a: f32, b: f32) -> f32 {
+        div(a, b)
+    }
+
+    #[bootrom_v2]
+    #[alias = __divdf3vfp]
+    #[aeabi = __aeabi_ddiv]
+    extern "C" fn __divdf3(a: f64, b: f64) -> f64 {
+        div(a, b)
+    }
+}

rp2040-hal/src/float/functions.rs

@@ -0,0 +1,236 @@
+use crate::float::{Float, Int};
+use crate::rom_data;
+
+trait ROMFunctions {
+    fn sqrt(self) -> Self;
+    fn ln(self) -> Self;
+    fn exp(self) -> Self;
+    fn sin(self) -> Self;
+    fn cos(self) -> Self;
+    fn tan(self) -> Self;
+    fn atan2(self, y: Self) -> Self;
+
+    fn to_trig_range(self) -> Self;
+}
+
+impl ROMFunctions for f32 {
+    fn sqrt(self) -> Self {
+        rom_data::float_funcs::fsqrt(self)
+    }
+
+    fn ln(self) -> Self {
+        rom_data::float_funcs::fln(self)
+    }
+
+    fn exp(self) -> Self {
+        rom_data::float_funcs::fexp(self)
+    }
+
+    fn sin(self) -> Self {
+        rom_data::float_funcs::fsin(self)
+    }
+
+    fn cos(self) -> Self {
+        rom_data::float_funcs::fcos(self)
+    }
+
+    fn tan(self) -> Self {
+        rom_data::float_funcs::ftan(self)
+    }
+
+    fn atan2(self, y: Self) -> Self {
+        rom_data::float_funcs::fatan2(self, y)
+    }
+
+    fn to_trig_range(self) -> Self {
+        // -128 < X < 128, logic from the Pico SDK
+        let exponent = (self.repr() & Self::EXPONENT_MASK) >> Self::SIGNIFICAND_BITS;
+        if exponent < 134 {
+            self
+        } else {
+            self % (core::f32::consts::PI * 2.0)
+        }
+    }
+}
+
+impl ROMFunctions for f64 {
+    fn sqrt(self) -> Self {
+        rom_data::double_funcs::dsqrt(self)
+    }
+
+    fn ln(self) -> Self {
+        rom_data::double_funcs::dln(self)
+    }
+
+    fn exp(self) -> Self {
+        rom_data::double_funcs::dexp(self)
+    }
+
+    fn sin(self) -> Self {
+        rom_data::double_funcs::dsin(self)
+    }
+
+    fn cos(self) -> Self {
+        rom_data::double_funcs::dcos(self)
+    }
+    fn tan(self) -> Self {
+        rom_data::double_funcs::dtan(self)
+    }
+
+    fn atan2(self, y: Self) -> Self {
+        rom_data::double_funcs::datan2(self, y)
+    }
+
+    fn to_trig_range(self) -> Self {
+        // -1024 < X < 1024, logic from the Pico SDK
+        let exponent = (self.repr() & Self::EXPONENT_MASK) >> Self::SIGNIFICAND_BITS;
+        if exponent < 1033 {
+            self
+        } else {
+            self % (core::f64::consts::PI * 2.0)
+        }
+    }
+}
+
+fn is_negative_nonzero_or_nan<F: Float>(f: F) -> bool {
+    let repr = f.repr();
+    if (repr & F::SIGN_MASK) != F::Int::ZERO {
+        // Negative, so anything other than exactly zero
+        return (repr & (!F::SIGN_MASK)) != F::Int::ZERO;
+    }
+    // NaN
+    (repr & (F::EXPONENT_MASK | F::SIGNIFICAND_MASK)) > F::EXPONENT_MASK
+}
+
+fn sqrt<F: Float + ROMFunctions>(f: F) -> F {
+    if is_negative_nonzero_or_nan(f) {
+        F::NAN
+    } else {
+        f.sqrt()
+    }
+}
+
+fn ln<F: Float + ROMFunctions>(f: F) -> F {
+    if is_negative_nonzero_or_nan(f) {
+        F::NAN
+    } else {
+        f.ln()
+    }
+}
+
+fn exp<F: Float + ROMFunctions>(f: F) -> F {
+    if f.is_nan() {
+        F::NAN
+    } else {
+        f.exp()
+    }
+}
+
+fn sin<F: Float + ROMFunctions>(f: F) -> F {
+    if f.is_not_finite() {
+        F::NAN
+    } else {
+        f.to_trig_range().sin()
+    }
+}
+
+fn cos<F: Float + ROMFunctions>(f: F) -> F {
+    if f.is_not_finite() {
+        F::NAN
+    } else {
+        f.to_trig_range().cos()
+    }
+}
+
+fn tan<F: Float + ROMFunctions>(f: F) -> F {
+    if f.is_not_finite() {
+        F::NAN
+    } else {
+        f.to_trig_range().tan()
+    }
+}
+
+fn atan2<F: Float + ROMFunctions>(x: F, y: F) -> F {
+    if x.is_nan() || y.is_nan() {
+        F::NAN
+    } else {
+        x.to_trig_range().atan2(y)
+    }
+}
+
+// Name collisions
+mod intrinsics {
+    intrinsics! {
+        extern "C" fn sqrtf(f: f32) -> f32 {
+            super::sqrt(f)
+        }
+
+        #[bootrom_v2]
+        extern "C" fn sqrt(f: f64) -> f64 {
+            super::sqrt(f)
+        }
+
+        extern "C" fn logf(f: f32) -> f32 {
+            super::ln(f)
+        }
+
+        #[bootrom_v2]
+        extern "C" fn log(f: f64) -> f64 {
+            super::ln(f)
+        }
+
+        extern "C" fn expf(f: f32) -> f32 {
+            super::exp(f)
+        }
+
+        #[bootrom_v2]
+        extern "C" fn exp(f: f64) -> f64 {
+            super::exp(f)
+        }
+
+        #[slower_than_default]
+        extern "C" fn sinf(f: f32) -> f32 {
+            super::sin(f)
+        }
+
+        #[slower_than_default]
+        #[bootrom_v2]
+        extern "C" fn sin(f: f64) -> f64 {
+            super::sin(f)
+        }
+
+        #[slower_than_default]
+        extern "C" fn cosf(f: f32) -> f32 {
+            super::cos(f)
+        }
+
+        #[slower_than_default]
+        #[bootrom_v2]
+        extern "C" fn cos(f: f64) -> f64 {
+            super::cos(f)
+        }
+
+        #[slower_than_default]
+        extern "C" fn tanf(f: f32) -> f32 {
+            super::tan(f)
+        }
+
+        #[slower_than_default]
+        #[bootrom_v2]
+        extern "C" fn tan(f: f64) -> f64 {
+            super::tan(f)
+        }
+
+        // Questionable gain
+        #[bootrom_v2]
+        extern "C" fn atan2f(a: f32, b: f32) -> f32 {
+            super::atan2(a, b)
+        }
+
+        // Questionable gain
+        #[bootrom_v2]
+        extern "C" fn atan2(a: f64, b: f64) -> f64 {
+            super::atan2(a, b)
+        }
+    }
+}

rp2040-hal/src/float/mod.rs

@@ -0,0 +1,146 @@
+use core::ops;
+
+// Borrowed and simplified from compiler-builtins so we can use bit ops
+// on floating point without macro soup.
+pub trait Int:
+    Copy
+    + core::fmt::Debug
+    + PartialEq
+    + PartialOrd
+    + ops::AddAssign
+    + ops::SubAssign
+    + ops::BitAndAssign
+    + ops::BitOrAssign
+    + ops::BitXorAssign
+    + ops::ShlAssign<i32>
+    + ops::ShrAssign<u32>
+    + ops::Add<Output = Self>
+    + ops::Sub<Output = Self>
+    + ops::Div<Output = Self>
+    + ops::Shl<u32, Output = Self>
+    + ops::Shr<u32, Output = Self>
+    + ops::BitOr<Output = Self>
+    + ops::BitXor<Output = Self>
+    + ops::BitAnd<Output = Self>
+    + ops::Not<Output = Self>
+{
+    const ZERO: Self;
+}
+
+macro_rules! int_impl {
+    ($ty:ty) => {
+        impl Int for $ty {
+            const ZERO: Self = 0;
+        }
+    };
+}
+
+int_impl!(u32);
+int_impl!(u64);
+
+pub trait Float:
+    Copy
+    + core::fmt::Debug
+    + PartialEq
+    + PartialOrd
+    + ops::AddAssign
+    + ops::MulAssign
+    + ops::Add<Output = Self>
+    + ops::Sub<Output = Self>
+    + ops::Div<Output = Self>
+    + ops::Rem<Output = Self>
+{
+    /// A uint of the same with as the float
+    type Int: Int;
+
+    /// NaN representation for the float
+    const NAN: Self;
+
+    /// The bitwidth of the float type
+    const BITS: u32;
+
+    /// The bitwidth of the significand
+    const SIGNIFICAND_BITS: u32;
+
+    /// A mask for the sign bit
+    const SIGN_MASK: Self::Int;
+
+    /// A mask for the significand
+    const SIGNIFICAND_MASK: Self::Int;
+
+    /// A mask for the exponent
+    const EXPONENT_MASK: Self::Int;
+
+    /// Returns `self` transmuted to `Self::Int`
+    fn repr(self) -> Self::Int;
+
+    /// Returns a `Self::Int` transmuted back to `Self`
+    fn from_repr(a: Self::Int) -> Self;
+
+    /// Return a sign swapped `self`
+    fn negate(self) -> Self;
+
+    /// Returns true if `self` is either NaN or infinity
+    fn is_not_finite(self) -> bool {
+        (self.repr() & Self::EXPONENT_MASK) == Self::EXPONENT_MASK
+    }
+
+    /// Returns true if `self` is infinity
+    fn is_infinity(self) -> bool {
+        (self.repr() & (Self::EXPONENT_MASK | Self::SIGNIFICAND_MASK)) == Self::EXPONENT_MASK
+    }
+
+    /// Returns true if `self is NaN
+    fn is_nan(self) -> bool {
+        (self.repr() & (Self::EXPONENT_MASK | Self::SIGNIFICAND_MASK)) > Self::EXPONENT_MASK
+    }
+
+    /// Returns true if `self` is negative
+    fn is_sign_negative(self) -> bool {
+        (self.repr() & Self::SIGN_MASK) != Self::Int::ZERO
+    }
+
+    /// Returns true if `self` is zero (either sign)
+    fn is_zero(self) -> bool {
+        (self.repr() & (Self::SIGNIFICAND_MASK | Self::EXPONENT_MASK)) == Self::Int::ZERO
+    }
+}
+
+macro_rules! float_impl {
+    ($ty:ident, $ity:ident, $bits:expr, $significand_bits:expr) => {
+        impl Float for $ty {
+            type Int = $ity;
+
+            const NAN: Self = <$ty>::NAN;
+
+            const BITS: u32 = $bits;
+            const SIGNIFICAND_BITS: u32 = $significand_bits;
+
+            const SIGN_MASK: Self::Int = 1 << (Self::BITS - 1);
+            const SIGNIFICAND_MASK: Self::Int = (1 << Self::SIGNIFICAND_BITS) - 1;
+            const EXPONENT_MASK: Self::Int = !(Self::SIGN_MASK | Self::SIGNIFICAND_MASK);
+
+            fn repr(self) -> Self::Int {
+                self.to_bits()
+            }
+
+            fn from_repr(a: Self::Int) -> Self {
+                Self::from_bits(a)
+            }
+
+            fn negate(self) -> Self {
+                -self
+            }
+        }
+    };
+}
+
+float_impl!(f32, u32, 32, 23);
+float_impl!(f64, u64, 64, 52);
+
+mod add_sub;
+mod cmp;
+mod conv;
+mod div;
+mod functions;
+mod mul;

rp2040-hal/src/float/mul.rs

@@ -0,0 +1,67 @@
+use super::Float;
+use crate::rom_data;
+
+trait ROMMul {
+    fn rom_mul(self, b: Self) -> Self;
+}
+
+impl ROMMul for f32 {
+    fn rom_mul(self, b: Self) -> Self {
+        rom_data::float_funcs::fmul(self, b)
+    }
+}
+
+impl ROMMul for f64 {
+    fn rom_mul(self, b: Self) -> Self {
+        rom_data::double_funcs::dmul(self, b)
+    }
+}
+
+fn mul<F: Float + ROMMul>(a: F, b: F) -> F {
+    if a.is_not_finite() {
+        if b.is_zero() {
+            // [-]inf/NaN * 0 = NaN
+            return F::NAN;
+        }
+
+        return if b.is_sign_negative() {
+            // [+/-]inf/NaN * (-X) = [-/+]inf/NaN
+            a.negate()
+        } else {
+            // [-]inf/NaN * X = [-]inf/NaN
+            a
+        };
+    }
+
+    if b.is_not_finite() {
+        if a.is_zero() {
+            // 0 * [-]inf/NaN = NaN
+            return F::NAN;
+        }
+
+        return if b.is_sign_negative() {
+            // (-X) * [+/-]inf/NaN = [-/+]inf/NaN
+            b.negate()
+        } else {
+            // X * [-]inf/NaN = [-]inf/NaN
+            b
+        };
+    }
+
+    a.rom_mul(b)
+}
+
+intrinsics! {
+    #[alias = __mulsf3vfp]
+    #[aeabi = __aeabi_fmul]
+    extern "C" fn __mulsf3(a: f32, b: f32) -> f32 {
+        mul(a, b)
+    }
+
+    #[bootrom_v2]
+    #[alias = __muldf3vfp]
+    #[aeabi = __aeabi_dmul]
+    extern "C" fn __muldf3(a: f64, b: f64) -> f64 {
+        mul(a, b)
+    }
+}

rp2040-hal/src/intrinsics.rs

@@ -0,0 +1,271 @@
+/// Generate a series of aliases for an intrinsic function.
+macro_rules! intrinsics_aliases {
+    (
+        extern $abi:tt fn $name:ident( $($argname:ident: $ty:ty),* ) -> $ret:ty,
+    ) => {};
+    (
+        unsafe extern $abi:tt fn $name:ident( $($argname:ident: $ty:ty),* ) -> $ret:ty,
+    ) => {};
+
+    (
+        extern $abi:tt fn $name:ident( $($argname:ident: $ty:ty),* ) -> $ret:ty,
+        $alias:ident
+        $($rest:ident)*
+    ) => {
+        #[cfg(all(target_arch = "arm", not(feature = "disable-intrinsics")))]
+        intrinsics! {
+            extern $abi fn $alias( $($argname: $ty),* ) -> $ret {
+                $name($($argname),*)
+            }
+        }
+
+        intrinsics_aliases! {
+            extern $abi fn $name( $($argname: $ty),* ) -> $ret,
+            $($rest)*
+        }
+    };
+
+    (
+        unsafe extern $abi:tt fn $name:ident( $($argname:ident: $ty:ty),* ) -> $ret:ty,
+        $alias:ident
+        $($rest:ident)*
+    ) => {
+        #[cfg(all(target_arch = "arm", not(feature = "disable-intrinsics")))]
+        intrinsics! {
+            unsafe extern $abi fn $alias( $($argname: $ty),* ) -> $ret {
+                $name($($argname),*)
+            }
+        }
+
+        intrinsics_aliases! {
+            unsafe extern $abi fn $name( $($argname: $ty),* ) -> $ret,
+            $($rest)*
+        }
+    };
+}
+
+/// The macro used to define overridden intrinsics.
+///
+/// This is heavily inspired by the macro used by compiler-builtins.  The idea
+/// is to abstract anything special that needs to be done to override an
+/// intrinsic function.  Intrinsic generation is disabled for non-ARM targets
+/// so things like CI and docs generation do not have problems.  Additionally
+/// they can be disabled with the crate feature `disable-intrinsics` for
+/// testing or comparing performance.
+///
+/// Like the compiler-builtins macro, it accepts a series of functions that
+/// looks like normal Rust code:
+///
+///     intrinsics! {
+///         extern "C" fn foo(a: i32) -> u32 {
+///             // ...
+///         }
+///
+///         #[nonstandard_attribute]
+///         extern "C" fn bar(a: i32) -> u32 {
+///             // ...
+///         }
+///     }
+///
+/// Each function can also be decorated with nonstandard attributes to control
+/// additional behaviour:
+///
+/// * `slower_than_default` - indicates that the override is slower than the
+///   default implementation.  Currently this just disables the override
+///   entirely.
+/// * `bootrom_v2` - indicates that the override is only available
+///   on a V2 bootrom or higher.  Only enabled when the feature
+///   `rom-v2-intrinsics` is set.
+/// * `alias` - accepts a list of names to alias the intrinsic to.
+/// * `aeabi` - accepts a list of ARM EABI names to alias to.
+///
+macro_rules! intrinsics {
+    () => {};
+
+    (
+        #[slower_than_default]
+        $(#[$($attr:tt)*])*
+        extern $abi:tt fn $name:ident( $($argname:ident: $ty:ty),* ) -> $ret:ty {
+            $($body:tt)*
+        }
+
+        $($rest:tt)*
+    ) => {
+        // Not exported, but defined so the actual implementation is
+        // considered used
+        #[allow(dead_code)]
+        fn $name( $($argname: $ty),* ) -> $ret {
+            $($body)*
+        }
+
+        intrinsics!($($rest)*);
+    };
+
+    (
+        #[bootrom_v2]
+        $(#[$($attr:tt)*])*
+        extern $abi:tt fn $name:ident( $($argname:ident: $ty:ty),* ) -> $ret:ty {
+            $($body:tt)*
+        }
+
+        $($rest:tt)*
+    ) => {
+        // Not exported, but defined so the actual implementation is
+        // considered used
+        #[cfg(not(feature = "rom-v2-intrinsics"))]
+        #[allow(dead_code)]
+        fn $name( $($argname: $ty),* ) -> $ret {
+            $($body)*
+        }
+
+        #[cfg(feature = "rom-v2-intrinsics")]
+        intrinsics! {
+            $(#[$($attr)*])*
+            extern $abi fn $name( $($argname: $ty),* ) -> $ret {
+                $($body)*
+            }
+        }
+
+        intrinsics!($($rest)*);
+    };
+
+    (
+        #[alias = $($alias:ident),*]
+        $(#[$($attr:tt)*])*
+        extern $abi:tt fn $name:ident( $($argname:ident: $ty:ty),* ) -> $ret:ty {
+            $($body:tt)*
+        }
+
+        $($rest:tt)*
+    ) => {
+        intrinsics! {
+            $(#[$($attr)*])*
+            extern $abi fn $name( $($argname: $ty),* ) -> $ret {
+                $($body)*
+            }
+        }
+
+        intrinsics_aliases! {
+            extern $abi fn $name( $($argname: $ty),* ) -> $ret,
+            $($alias) *
+        }
+
+        intrinsics!($($rest)*);
+    };
+
+    (
+        #[alias = $($alias:ident),*]
+        $(#[$($attr:tt)*])*
+        unsafe extern $abi:tt fn $name:ident( $($argname:ident: $ty:ty),* ) -> $ret:ty {
+            $($body:tt)*
+        }
+
+        $($rest:tt)*
+    ) => {
+        intrinsics! {
+            $(#[$($attr)*])*
+            unsafe extern $abi fn $name( $($argname: $ty),* ) -> $ret {
+                $($body)*
+            }
+        }
+
+        intrinsics_aliases! {
+            unsafe extern $abi fn $name( $($argname: $ty),* ) -> $ret,
+            $($alias) *
+        }
+
+        intrinsics!($($rest)*);
+    };
+
+    (
+        #[aeabi = $($alias:ident),*]
+        $(#[$($attr:tt)*])*
+        extern $abi:tt fn $name:ident( $($argname:ident: $ty:ty),* ) -> $ret:ty {
+            $($body:tt)*
+        }
+
+        $($rest:tt)*
+    ) => {
+        intrinsics! {
+            $(#[$($attr)*])*
+            extern $abi fn $name( $($argname: $ty),* ) -> $ret {
+                $($body)*
+            }
+        }
+
+        intrinsics_aliases! {
+            extern "aapcs" fn $name( $($argname: $ty),* ) -> $ret,
+            $($alias) *
+        }
+
+        intrinsics!($($rest)*);
+    };
+
+    (
+        $(#[$($attr:tt)*])*
+        extern $abi:tt fn $name:ident( $($argname:ident: $ty:ty),* ) -> $ret:ty {
+            $($body:tt)*
+        }
+
+        $($rest:tt)*
+    ) => {
+        #[cfg(all(target_arch = "arm", not(feature = "disable-intrinsics")))]
+        $(#[$($attr)*])*
+        extern $abi fn $name( $($argname: $ty),* ) -> $ret {
+            $($body)*
+        }
+
+        #[cfg(all(target_arch = "arm", not(feature = "disable-intrinsics")))]
+        mod $name {
+            #[no_mangle]
+            $(#[$($attr)*])*
+            pub extern $abi fn $name( $($argname: $ty),* ) -> $ret {
+                super::$name($($argname),*)
+            }
+        }
+
+        // Not exported, but defined so the actual implementation is
+        // considered used
+        #[cfg(not(all(target_arch = "arm", not(feature = "disable-intrinsics"))))]
+        #[allow(dead_code)]
+        fn $name( $($argname: $ty),* ) -> $ret {
+            $($body)*
+        }
+
+        intrinsics!($($rest)*);
+    };
+
+    (
+        $(#[$($attr:tt)*])*
+        unsafe extern $abi:tt fn $name:ident( $($argname:ident: $ty:ty),* ) -> $ret:ty {
+            $($body:tt)*
+        }
+
+        $($rest:tt)*
+    ) => {
+        #[cfg(all(target_arch = "arm", not(feature = "disable-intrinsics")))]
+        $(#[$($attr)*])*
+        unsafe extern $abi fn $name( $($argname: $ty),* ) -> $ret {
+            $($body)*
+        }
+
+        #[cfg(all(target_arch = "arm", not(feature = "disable-intrinsics")))]
+        mod $name {
+            #[no_mangle]
+            $(#[$($attr)*])*
+            pub unsafe extern $abi fn $name( $($argname: $ty),* ) -> $ret {
+                super::$name($($argname),*)
+            }
+        }
+
+        // Not exported, but defined so the actual implementation is
+        // considered used
+        #[cfg(not(all(target_arch = "arm", not(feature = "disable-intrinsics"))))]
+        #[allow(dead_code)]
+        unsafe fn $name( $($argname: $ty),* ) -> $ret {
+            $($body)*
+        }
+
+        intrinsics!($($rest)*);
+    };
+}

rp2040-hal/src/lib.rs

@@ -15,11 +15,15 @@ pub use paste;
 
 pub extern crate rp2040_pac as pac;
 
+#[macro_use]
+mod intrinsics;
+
 pub mod adc;
 pub(crate) mod atomic_register_access;
 pub mod clocks;
 mod critical_section_impl;
 pub mod dma;
+mod float;
 pub mod gpio;
 pub mod i2c;
 pub mod multicore;

rp2040-hal/src/rom_data.rs

@@ -24,6 +24,9 @@ const FUNC_TABLE: *const u16 = 0x0000_0014 as _;
 /// Pointer to the public data lookup table.
 const DATA_TABLE: *const u16 = 0x0000_0016 as _;
 
+/// Address of the version number of the ROM.
+const VERSION_NUMBER: *const u8 = 0x0000_0013 as _;
+
 /// Retrive rom content from a table using a code.
 fn rom_table_lookup<T>(table: *const u16, tag: RomFnTableCode) -> T {
     unsafe {
@@ -215,24 +218,26 @@ rom_functions! {
     b"UB" fn reset_to_usb_boot(gpio_activity_pin_mask: u32, disable_interface_mask: u32) -> ();
 
     /// Sets n bytes start at ptr to the value c and returns ptr
-    b"MS" unsafe fn memset(ptr: *mut u8, c: u8, n: u8) -> *mut u8;
+    b"MS" unsafe fn memset(ptr: *mut u8, c: u8, n: u32) -> *mut u8;
 
     /// Sets n bytes start at ptr to the value c and returns ptr.
     ///
     /// Note this is a slightly more efficient variant of _memset that may only
     /// be used if ptr is word aligned.
-    b"M4" unsafe fn memset4(ptr: *mut u32, c: u8, n: u32) -> *mut u32;
+    // Note the datasheet does not match the actual ROM for the code here, see
+    // https://github.com/raspberrypi/pico-feedback/issues/217
+    b"S4" unsafe fn memset4(ptr: *mut u32, c: u8, n: u32) -> *mut u32;
 
     /// Copies n bytes starting at src to dest and returns dest. The results are undefined if the
     /// regions overlap.
-    b"MC" unsafe fn memcpy(dest: *mut u8, src: *mut u8, n: u32) -> u8;
+    b"MC" unsafe fn memcpy(dest: *mut u8, src: *const u8, n: u32) -> *mut u8;
 
     /// Copies n bytes starting at src to dest and returns dest. The results are undefined if the
     /// regions overlap.
     ///
     /// Note this is a slightly more efficient variant of _memcpy that may only be
     /// used if dest and src are word aligned.
-    b"C4" unsafe fn memcpy44(dest: *mut u32, src: *mut u32, n: u32) -> *mut u8;
+    b"C4" unsafe fn memcpy44(dest: *mut u32, src: *const u32, n: u32) -> *mut u8;
 
     /// Restore all QSPI pad controls to their default state, and connect the SSI to the QSPI pads.
     b"IF" unsafe fn connect_internal_flash() -> ();
@@ -274,6 +279,60 @@ rom_functions! {
     b"WV" unsafe fn wait_for_vector() -> !;
 }
 
+// Various C intrinsics in the ROM
+intrinsics! {
+    #[alias = __popcountdi2]
+    extern "C" fn __popcountsi2(x: u32) -> u32 {
+        popcount32(x)
+    }
+
+    #[alias = __clzdi2]
+    extern "C" fn __clzsi2(x: u32) -> u32 {
+        clz32(x)
+    }
+
+    #[alias = __ctzdi2]
+    extern "C" fn __ctzsi2(x: u32) -> u32 {
+        ctz32(x)
+    }
+
+    // __rbit is only unofficial, but it show up in the ARM documentation,
+    // so may as well hook it up.
+    #[alias = __rbitl]
+    extern "C" fn __rbit(x: u32) -> u32 {
+        reverse32(x)
+    }
+
+    unsafe extern "aapcs" fn __aeabi_memset(dest: *mut u8, n: usize, c: i32) -> () {
+        // Different argument order
+        memset(dest, c as u8, n as u32);
+    }
+
+    #[alias = __aeabi_memset8]
+    unsafe extern "aapcs" fn __aeabi_memset4(dest: *mut u8, n: usize, c: i32) -> () {
+        // Different argument order
+        memset4(dest as *mut u32, c as u8, n as u32);
+    }
+
+    unsafe extern "aapcs" fn __aeabi_memclr(dest: *mut u8, n: usize) -> () {
+        memset(dest, 0, n as u32);
+    }
+
+    #[alias = __aeabi_memclr8]
+    unsafe extern "aapcs" fn __aeabi_memclr4(dest: *mut u8, n: usize) -> () {
+        memset4(dest as *mut u32, 0, n as u32);
+    }
+
+    unsafe extern "aapcs" fn __aeabi_memcpy(dest: *mut u8, src: *const u8, n: usize) -> () {
+        memcpy(dest, src, n as u32);
+    }
+
+    #[alias = __aeabi_memcpy8]
+    unsafe extern "aapcs" fn __aeabi_memcpy4(dest: *mut u8, src: *const u8, n: usize) -> () {
+        memcpy44(dest as *mut u32, src as *const u32, n as u32);
+    }
+}
+
 unsafe fn convert_str(s: *const u8) -> &'static str {
     let mut end = s;
     while *end != 0 {
@@ -283,6 +342,11 @@ unsafe fn convert_str(s: *const u8) -> &'static str {
     core::str::from_utf8_unchecked(s)
 }
 
+/// The version number of the rom.
+pub fn rom_version_number() -> u8 {
+    unsafe { *VERSION_NUMBER }
+}
+
 /// The Raspberry Pi Trading Ltd copyright string.
 pub fn copyright_string() -> &'static str {
     let s: *const u8 = rom_table_lookup(DATA_TABLE, *b"CR");
@@ -316,6 +380,12 @@ pub fn fplib_end() -> *const u8 {
 
 /// This entry is only present in the V2 bootrom. See Table 182 in the RP2040 datasheet for the contents of this table.
 pub fn soft_double_table() -> *const usize {
+    if rom_version_number() < 2 {
+        panic!(
+            "Double precision operations require V2 bootrom (found: V{})",
+            rom_version_number()
+        );
+    }
     rom_table_lookup(DATA_TABLE, *b"SD")
 }
 
@@ -419,9 +489,44 @@ pub mod float_funcs {
         0x4c fexp(v: f32) -> f32;
         /// Calculates the natural logarithm of `v`. If `v <= 0` return -Infinity
         0x50 fln(v: f32) -> f32;
+    }
+
+    macro_rules! make_functions_v2 {
+        (
+            $(
+                $(#[$outer:meta])*
+                $offset:literal $name:ident (
+                    $( $aname:ident : $aty:ty ),*
+                ) -> $ret:ty;
+            )*
+        ) => {
+            $(
+                declare_rom_function! {
+                    $(#[$outer])*
+                    fn $name( $( $aname : $aty ),* ) -> $ret {
+                        if $crate::rom_data::rom_version_number() < 2 {
+                            panic!(
+                                "Floating point function requires V2 bootrom (found: V{})",
+                                $crate::rom_data::rom_version_number()
+                            );
+                        }
+                        let table: *const usize = $crate::rom_data::soft_float_table();
+                        unsafe {
+                            // This is the entry in the table. Our offset is given as a
+                            // byte offset, but we want the table index (each pointer in
+                            // the table is 4 bytes long)
+                            let entry: *const usize = table.offset($offset / 4);
+                            // Read the pointer from the table
+                            core::ptr::read(entry) as *const u32
+                        }
+                    }
+                }
+            )*
+        }
+    }
 
     // These are only on BootROM v2 or higher
-
+    make_functions_v2! {
         /// Compares two floating point numbers, returning:
         ///     • 0 if a == b
         ///     • -1 if a < b

rp2040-hal/src/sio.rs

@@ -171,7 +171,7 @@ impl SioFifo {
     }
 }
 
-fn save_divider<F, R>(f: F) -> R
+pub(crate) fn save_divider<F, R>(f: F) -> R
 where
     F: FnOnce(&pac::sio::RegisterBlock) -> R,
 {
@@ -286,71 +286,17 @@ impl HwDivider {
     }
 }
 
-macro_rules! divider_intrinsics {
-    () => ();
-
-    (
-        #[arm_aeabi_alias = $alias:ident]
-        pub extern $abi:tt fn $name:ident( $($argname:ident:  $ty:ty),* ) -> $ret:ty {
-            $($body:tt)*
-        }
-
-        $($rest:tt)*
-    ) => (
-        extern $abi fn $name( $($argname: $ty),* ) -> $ret {
-            $($body)*
-        }
-
-        mod $name {
-            #[no_mangle]
-            pub extern $abi fn $name( $($argname: $ty),* ) -> $ret {
-                super::$name($($argname),*)
-            }
-        }
-
-        mod $alias {
-            #[no_mangle]
-            pub extern $abi fn $alias( $($argname: $ty),* ) -> $ret {
-                super::$name($($argname),*)
-            }
-        }
-
-        divider_intrinsics!($($rest)*);
-    );
-
-    (
-        pub extern $abi:tt fn $name:ident( $($argname:ident:  $ty:ty),* ) -> $ret:ty {
-            $($body:tt)*
-        }
-
-        $($rest:tt)*
-    ) => (
-        extern $abi fn $name( $($argname: $ty),* ) -> $ret {
-            $($body)*
-        }
-
-        mod $name {
-            #[no_mangle]
-            pub extern $abi fn $name( $($argname: $ty),* ) -> $ret {
-                super::$name($($argname),*)
-            }
-        }
-
-        divider_intrinsics!($($rest)*);
-    );
-}
-
-divider_intrinsics! {
-    #[arm_aeabi_alias = __aeabi_uidiv]
-    pub extern "C" fn __udivsi3(n: u32, d: u32) -> u32 {
+intrinsics! {
+    #[aeabi = __aeabi_uidiv]
+    extern "C" fn __udivsi3(n: u32, d: u32) -> u32 {
         divider_unsigned(n, d).quotient
     }
 
-    pub extern "C" fn __umodsi3(n: u32, d: u32) -> u32 {
+    extern "C" fn __umodsi3(n: u32, d: u32) -> u32 {
         divider_unsigned(n, d).remainder
     }
 
-    pub extern "C" fn __udivmodsi4(n: u32, d: u32, rem: Option<&mut u32>) -> u32 {
+    extern "C" fn __udivmodsi4(n: u32, d: u32, rem: Option<&mut u32>) -> u32 {
         let quo_rem = divider_unsigned(n, d);
         if let Some(rem) = rem {
             *rem = quo_rem.remainder;
@@ -358,16 +304,16 @@ divider_intrinsics! {
         quo_rem.quotient
     }
 
-    #[arm_aeabi_alias = __aeabi_idiv]
-    pub extern "C" fn __divsi3(n: i32, d: i32) -> i32 {
+    #[aeabi = __aeabi_idiv]
+    extern "C" fn __divsi3(n: i32, d: i32) -> i32 {
         divider_signed(n, d).quotient
     }
 
-    pub extern "C" fn __modsi3(n: i32, d: i32) -> i32 {
+    extern "C" fn __modsi3(n: i32, d: i32) -> i32 {
         divider_signed(n, d).remainder
     }
 
-    pub extern "C" fn __divmodsi4(n: i32, d: i32, rem: &mut i32) -> i32 {
+    extern "C" fn __divmodsi4(n: i32, d: i32, rem: &mut i32) -> i32 {
         let quo_rem = divider_signed(n, d);
         *rem = quo_rem.remainder;
         quo_rem.quotient