fix: overflow

algebraic/src/arch/x86_64/avx2_field_gl.rs

@@ -45,6 +45,7 @@ impl Avx2GoldilocksField {
     /// A trick to suppress overflow. `as_slice` may return an element bigger than GL modular.
     #[inline]
     pub fn as_slice_reduced(&mut self) -> &[GoldilocksField] {
+        *self = Self::ZEROS + *self;
         &self.0[..]
     }
 }
@@ -71,8 +72,6 @@ unsafe impl PackedField for Avx2GoldilocksField {
     }
     #[inline]
     fn as_slice_mut(&mut self) -> &mut [GoldilocksField] {
-        // A trick to suppress overflow. `as_slice` may return a number bigger than GL modular.
-        *self = Self::ZEROS + *self;
         &mut self.0[..]
     }
 
@@ -471,7 +470,7 @@ unsafe fn reduce128(x: (__m256i, __m256i)) -> __m256i {
     let lo1_s = sub_small_64s_64_s(lo0_s, hi_hi0);
     let t1 = _mm256_mul_epu32(hi0, EPSILON);
     let lo2_s = add_small_64s_64_s(lo1_s, t1);
-    let lo2 = shift(canonicalize_s(lo2_s));
+    let lo2 = shift(lo2_s);
     lo2
 }