diff --git a/plonky2/src/gadgets/curve.rs b/plonky2/src/gadgets/curve.rs
index c2af0104..d56c7650 100644
--- a/plonky2/src/gadgets/curve.rs
+++ b/plonky2/src/gadgets/curve.rs
@@ -230,7 +230,7 @@ impl<F: RichField + Extendable<D>, const D: usize> CircuitBuilder<F, D> {
 
         let windows = self.split_nonnative_to_4_bit_limbs(n);
         let m = C::ScalarField::BITS / WINDOW_SIZE;
-        for i in m..0 {
+        for i in (0..m).rev() {
             result = self.curve_double(&result);
             let window = windows[i];
 
@@ -411,6 +411,38 @@ mod tests {
         verify(proof, &data.verifier_only, &data.common)
     }
 
+    #[test]
+    fn test_curve_mul_windowed() -> Result<()> {
+        const D: usize = 2;
+        type C = PoseidonGoldilocksConfig;
+        type F = <C as GenericConfig<D>>::F;
+
+        let config = CircuitConfig::standard_ecc_config();
+
+        let pw = PartialWitness::new();
+        let mut builder = CircuitBuilder::<F, D>::new(config);
+
+        let g = Secp256K1::GENERATOR_AFFINE;
+        let five = Secp256K1Scalar::from_canonical_usize(5);
+        let neg_five = five.neg();
+        let neg_five_scalar = CurveScalar::<Secp256K1>(neg_five);
+        let neg_five_g = (neg_five_scalar * g.to_projective()).to_affine();
+        let neg_five_g_expected = builder.constant_affine_point(neg_five_g);
+        builder.curve_assert_valid(&neg_five_g_expected);
+
+        let g_target = builder.constant_affine_point(g);
+        let neg_five_target = builder.constant_nonnative(neg_five);
+        let neg_five_g_actual = builder.curve_scalar_mul_windowed(&g_target, &neg_five_target);
+        /*builder.curve_assert_valid(&neg_five_g_actual);
+
+        builder.connect_affine_point(&neg_five_g_expected, &neg_five_g_actual);*/
+
+        let data = builder.build::<C>();
+        let proof = data.prove(pw).unwrap();
+
+        verify(proof, &data.verifier_only, &data.common)
+    }
+
     #[test]
     #[ignore]
     fn test_curve_random() -> Result<()> {
diff --git a/plonky2/src/gadgets/split_nonnative.rs b/plonky2/src/gadgets/split_nonnative.rs
index d1f16b65..88c693c0 100644
--- a/plonky2/src/gadgets/split_nonnative.rs
+++ b/plonky2/src/gadgets/split_nonnative.rs
@@ -1,8 +1,11 @@
+use std::marker::PhantomData;
+
 use itertools::Itertools;
 use plonky2_field::extension_field::Extendable;
 use plonky2_field::field_types::Field;
 
 use crate::gadgets::arithmetic_u32::U32Target;
+use crate::gadgets::biguint::BigUintTarget;
 use crate::gadgets::nonnative::NonNativeTarget;
 use crate::hash::hash_types::RichField;
 use crate::iop::target::Target;
@@ -10,7 +13,7 @@ use crate::plonk::circuit_builder::CircuitBuilder;
 
 impl<F: RichField + Extendable<D>, const D: usize> CircuitBuilder<F, D> {
     pub fn split_u32_to_4_bit_limbs(&mut self, val: U32Target) -> Vec<Target> {
-        let two_bit_limbs = self.split_le_base::<2>(val.0, 16);
+        let two_bit_limbs = self.split_le_base::<4>(val.0, 16);
         let four = self.constant(F::from_canonical_usize(4));
         let combined_limbs = two_bit_limbs
             .iter()
@@ -31,4 +34,64 @@ impl<F: RichField + Extendable<D>, const D: usize> CircuitBuilder<F, D> {
             .flat_map(|&l| self.split_u32_to_4_bit_limbs(l))
             .collect()
     }
+
+    // Note: assumes its inputs are 4-bit limbs, and does not range-check.
+    pub fn recombine_nonnative_4_bit_limbs<FF: Field>(
+        &mut self,
+        limbs: Vec<Target>,
+    ) -> NonNativeTarget<FF> {
+        let base = self.constant_u32(1 << 4);
+        let u32_limbs = limbs
+            .chunks(8)
+            .map(|chunk| {
+                let mut combined_chunk = self.zero_u32();
+                for i in (0..8).rev() {
+                    let (low, _high) = self.mul_add_u32(combined_chunk, base, U32Target(chunk[i]));
+                    combined_chunk = low;
+                }
+                combined_chunk
+            })
+            .collect();
+
+        NonNativeTarget {
+            value: BigUintTarget { limbs: u32_limbs },
+            _phantom: PhantomData,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use anyhow::Result;
+    use plonky2_field::field_types::Field;
+    use plonky2_field::secp256k1_scalar::Secp256K1Scalar;
+
+    use crate::gadgets::nonnative::NonNativeTarget;
+    use crate::iop::witness::PartialWitness;
+    use crate::plonk::circuit_builder::CircuitBuilder;
+    use crate::plonk::circuit_data::CircuitConfig;
+    use crate::plonk::config::{GenericConfig, PoseidonGoldilocksConfig};
+    use crate::plonk::verifier::verify;
+
+    #[test]
+    fn test_split_nonnative() -> Result<()> {
+        type FF = Secp256K1Scalar;
+        const D: usize = 2;
+        type C = PoseidonGoldilocksConfig;
+        type F = <C as GenericConfig<D>>::F;
+
+        let config = CircuitConfig::standard_ecc_config();
+        let pw = PartialWitness::new();
+        let mut builder = CircuitBuilder::<F, D>::new(config);
+
+        let x = FF::rand();
+        let x_target = builder.constant_nonnative(x);
+        let split = builder.split_nonnative_to_4_bit_limbs(&x_target);
+        let combined: NonNativeTarget<Secp256K1Scalar> = builder.recombine_nonnative_4_bit_limbs(split);
+        builder.connect_nonnative(&x_target, &combined);
+
+        let data = builder.build::<C>();
+        let proof = data.prove(pw).unwrap();
+        verify(proof, &data.verifier_only, &data.common)
+    }
 }