From da5239c272ca24d5e7e203d8fbc58f4bc8c0f012 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Fri, 24 Jul 2026 14:20:18 -0400
Subject: [PATCH] Improve performance of widening muls

---
 include/boost/int128/detail/int128_imp.hpp  | 15 ++++++++++++++-
 include/boost/int128/detail/uint128_imp.hpp |  9 +++++++++
 2 files changed, 23 insertions(+), 1 deletion(-)

diff --git a/include/boost/int128/detail/int128_imp.hpp b/include/boost/int128/detail/int128_imp.hpp
index 6c821a00..c01ca870 100644
--- a/include/boost/int128/detail/int128_imp.hpp
+++ b/include/boost/int128/detail/int128_imp.hpp
@@ -2052,12 +2052,20 @@ namespace detail {
 
 BOOST_INT128_HOST_DEVICE BOOST_INT128_FORCE_INLINE constexpr int128_t default_mul(const int128_t lhs, const std::uint64_t rhs) noexcept
 {
+    #if defined(BOOST_INT128_HAS_INT128) && !defined(__s390__) && !defined(__s390x__)
+    
+    return int128_t{static_cast<detail::builtin_u128>(lhs) * static_cast<detail::builtin_u128>(rhs)};
+
+    #else
+
     return low_word_mul<int128_t>(lhs, rhs);
+
+    #endif
 }
 
 BOOST_INT128_HOST_DEVICE BOOST_INT128_FORCE_INLINE constexpr int128_t default_mul(const int128_t lhs, const std::uint32_t rhs) noexcept
 {
-    return low_word_mul<int128_t>(lhs, rhs);
+    return default_mul(lhs, static_cast<std::uint64_t>(rhs));
 }
 
 #if defined(_M_AMD64) && !defined(__GNUC__)
@@ -2131,6 +2139,11 @@ BOOST_INT128_HOST_DEVICE BOOST_INT128_FORCE_INLINE constexpr int128_t default_mu
         return msvc_amd64_mul(lhs, rhs);
     }
 
+    #elif defined(BOOST_INT128_HAS_INT128) && !defined(__s390__) && !defined(__s390x__)
+
+    // Multiply in the unsigned domain to avoid signed-overflow UB, then reinterpret the bits.
+    return int128_t{static_cast<detail::builtin_u128>(lhs) * static_cast<detail::builtin_u128>(rhs)};
+
     #else
 
     return low_word_mul<int128_t>(lhs, rhs);
diff --git a/include/boost/int128/detail/uint128_imp.hpp b/include/boost/int128/detail/uint128_imp.hpp
index 2af78cdc..d57c28c2 100644
--- a/include/boost/int128/detail/uint128_imp.hpp
+++ b/include/boost/int128/detail/uint128_imp.hpp
@@ -1915,6 +1915,10 @@ BOOST_INT128_HOST_DEVICE BOOST_INT128_FORCE_INLINE constexpr uint128_t default_a
 
     return res;
 
+    #elif (defined(__x86_64__) || (defined(__aarch64__) && !defined(__APPLE__))) && !defined(_MSC_VER) && defined(BOOST_INT128_HAS_INT128)
+
+    return static_cast<uint128_t>(static_cast<detail::builtin_u128>(lhs) + static_cast<detail::builtin_u128>(rhs));
+
     #else
 
     uint128_t temp {lhs.high + rhs.high, lhs.low + rhs.low};
@@ -2306,6 +2310,11 @@ BOOST_INT128_HOST_DEVICE BOOST_INT128_FORCE_INLINE constexpr uint128_t default_m
         return msvc_mul(lhs, rhs);
     }
 
+    #elif defined(BOOST_INT128_HAS_INT128) && !defined(__s390__) && !defined(__s390x__)
+    #  define BOOST_INT128_HIDE_MUL
+
+    return static_cast<uint128_t>(static_cast<detail::builtin_u128>(lhs) * static_cast<detail::builtin_u128>(rhs));
+
     #endif
 
     // We need to hide this if we use a non-const eval method above to avoid a litany of cross-platform warnings