added comments

src/u32x4_.rs

@@ -481,6 +481,8 @@ impl u32x4 {
     }
   }
 
+  /// Multiplies 32x32 bit to 64 bit and then only keeps the high 32 bits of the result.
+  /// Useful for implementing divide constant value (see t_usefulness example)
   #[inline]
   #[must_use]
   pub fn mul_keep_high(self, rhs: Self) -> Self {

src/u32x8_.rs

@@ -318,6 +318,16 @@ impl u32x8 {
     }
   }
 
+  /// Multiplies 32x32 bit to 64 bit and then only keeps the high 32 bits of the result.
+  /// Useful for implementing divide constant value (see t_usefulness example)
+  #[inline]
+  #[must_use]
+  pub fn mul_keep_high(self: u32x8, rhs: u32x8) -> u32x8 {
+    // avx2 doesn't benefit here sice the u32x4 is already using it,
+    // maybe it might help with the shuffling afterwards
+    u32x8 { a: self.a.mul_keep_high(rhs.a), b: self.b.mul_keep_high(rhs.b) }
+  }
+
   #[inline]
   #[must_use]
   pub fn blend(self, t: Self, f: Self) -> Self {

tests/all_tests/t_u32x8.rs

@@ -315,3 +315,11 @@ fn test_u32x8_mul_widen_even() {
   let actual = a.mul_widen_even(b);
   assert_eq!(expected, actual);
 }
+
+#[test]
+fn impl_u32x8_mul_keep_high() {
+  crate::test_random_vector_vs_scalar(
+    |a: u32x8, b| a.mul_keep_high(b),
+    |a, b| ((u64::from(a) * u64::from(b)) >> 32) as u32,
+  );
+}

tests/all_tests/t_usefulness.rs

@@ -391,12 +391,7 @@ fn generate_branch_free_divide_magic_shift(denom: u32x8) -> (u32x8, u32x8) {
 
 // using the previously generated magic and shift, calculate the division
 fn branch_free_divide(numerator: u32x8, magic: u32x8, shift: u32x8) -> u32x8 {
-  let a: [u32x4; 2] = cast(numerator);
-  let b: [u32x4; 2] = cast(magic);
-
-  let q = [a[0].mul_keep_high(b[0]), a[1].mul_keep_high(b[1])];
-
-  let q: u32x8 = cast(q);
+  let q = numerator.mul_keep_high(magic);
 
   let t = ((numerator - q) >> 1) + q;
   t >> shift