diff --git a/include/fast_float/decimal_to_binary.h b/include/fast_float/decimal_to_binary.h
index aa7d7cc..2c30cc5 100644
--- a/include/fast_float/decimal_to_binary.h
+++ b/include/fast_float/decimal_to_binary.h
@@ -141,13 +141,16 @@ compute_float(am_pow_t q, am_mant_t w) noexcept {
   // branchless approach: value128 product = compute_product(q, w); but in
   // practice, we can win big with the compute_product_approximation if its
   // additional branch is easily predicted. Which is best is data specific.
-  auto const upperbit = product.high >> 63;
-  auto const shift = upperbit + 64 - binary::mantissa_explicit_bits() - 3;
+  am_pow_t const upperbit = product.high >> 63;
+  am_pow_t const shift = upperbit + 64 - binary::mantissa_explicit_bits() - 3;
 
   // Shift right the mantissa to the correct position
   answer.mantissa = product.high >> shift;
 
   answer.power2 = detail::power(q) + upperbit - lz - binary::minimum_exponent();
+
+  // Now, we need to round the mantissa correctly.
+
   if (answer.power2 <= 0) { // we have a subnormal or very small value.
     // Here have that answer.power2 <= 0 so -answer.power2 >= 0
     if (-answer.power2 + 1 >=