Fix slowdown with 64 bit mul

cppalliance · Feb 20, 2025 · 154d3b1 · 154d3b1
1 parent 3edf5fc
commit 154d3b1
Show file tree

Hide file tree

Showing 3 changed files with 6 additions and 20 deletions.
diff --git a/include/boost/decimal/decimal64.hpp b/include/boost/decimal/decimal64.hpp
@@ -1429,8 +1429,10 @@ constexpr auto operator*(decimal64 lhs, decimal64 rhs) noexcept -> decimal64
     }
     #endif
 
-    const auto lhs_components {lhs.to_components()};
-    const auto rhs_components {rhs.to_components()};
+    auto lhs_components {lhs.to_components()};
+    detail::normalize(lhs_components.sig, lhs_components.exp);
+    auto rhs_components {rhs.to_components()};
+    detail::normalize(rhs_components.sig, rhs_components.exp);
 
     return detail::d64_mul_impl<decimal64>(lhs_components, rhs_components);
 }

diff --git a/include/boost/decimal/decimal64_fast.hpp b/include/boost/decimal/decimal64_fast.hpp
@@ -130,7 +130,7 @@ BOOST_DECIMAL_EXPORT class decimal64_fast final
     BOOST_DECIMAL_REQUIRES_RETURN(detail::is_decimal_floating_point_v, DecimalType, std::uint64_t);
 
     template <typename ReturnType, typename T>
-    BOOST_DECIMAL_FORCE_INLINE friend constexpr auto detail::d64_mul_impl(const T& lhs, const T& rhs) noexcept -> std::enable_if_t<detail::is_fast_type_v<ReturnType>, ReturnType>;
+    friend constexpr auto detail::d64_mul_impl(const T& lhs, const T& rhs) noexcept -> ReturnType;
 
     template <BOOST_DECIMAL_FAST_DECIMAL_FLOATING_TYPE DecimalType>
     BOOST_DECIMAL_FORCE_INLINE friend constexpr auto fast_equality_impl(const DecimalType& lhs, const DecimalType& rhs) noexcept -> bool;

diff --git a/include/boost/decimal/detail/mul_impl.hpp b/include/boost/decimal/detail/mul_impl.hpp
@@ -61,25 +61,9 @@ BOOST_DECIMAL_FORCE_INLINE constexpr auto mul_impl(T lhs_sig, U lhs_exp, bool lh
     return {static_cast<std::uint32_t>(res_sig), res_exp, lhs_sign != rhs_sign};
 }
 
-template <typename ReturnType, typename T>
-BOOST_DECIMAL_FORCE_INLINE constexpr auto d64_mul_impl(const T& lhs, const T& rhs) noexcept -> std::enable_if_t<detail::is_ieee_type_v<ReturnType>, ReturnType>
-{
-    // Clang 6-12 yields incorrect results with builtin u128, so we force usage of our version
-    #if defined(BOOST_DECIMAL_HAS_INT128) && (!defined(__clang_major__) || (__clang_major__) > 12)
-    using unsigned_int128_type = boost::decimal::detail::uint128_t;
-    #else
-    using unsigned_int128_type = boost::decimal::detail::uint128;
-    #endif
-
-    const auto res_sig {(static_cast<unsigned_int128_type>(lhs.full_significand()) * static_cast<unsigned_int128_type>(rhs.full_significand()))};
-    const auto res_exp {lhs.biased_exponent() + rhs.biased_exponent()};
-
-    return {res_sig, res_exp, lhs.isneg() != rhs.isneg()};
-}
-
 // In the fast case we are better served doing our 128-bit division here since we are at a know starting point
 template <typename ReturnType, typename T>
-BOOST_DECIMAL_FORCE_INLINE constexpr auto d64_mul_impl(const T& lhs, const T& rhs) noexcept -> std::enable_if_t<detail::is_fast_type_v<ReturnType>, ReturnType>
+constexpr auto d64_mul_impl(const T& lhs, const T& rhs) noexcept -> ReturnType
 {
     // Clang 6-12 yields incorrect results with builtin u128, so we force usage of our version
     #if defined(BOOST_DECIMAL_HAS_INT128) && (!defined(__clang_major__) || (__clang_major__) > 12)