From f4cff75f42a7b0c2bb192b1566e993883c6c710c Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Pawe=C5=82=20Bylica?= <chfast@gmail.com>
Date: Tue, 15 Mar 2022 12:15:48 +0100
Subject: [PATCH] Optimize less-than

The new implementation is similar to what was there: we select which
half to compare and do 128-bit less-than. For small values this is
equivalent to do subtraction + borrow check, for big values though we
get 25% boost.
---
 include/intx/intx.hpp          | 13 ++++++++-----
 test/benchmarks/benchmarks.cpp |  6 ++++++
 2 files changed, 14 insertions(+), 5 deletions(-)

diff --git a/include/intx/intx.hpp b/include/intx/intx.hpp
index 16c702e0..236569fa 100644
--- a/include/intx/intx.hpp
+++ b/include/intx/intx.hpp
@@ -1087,11 +1087,14 @@ inline constexpr bool operator!=(const T& x, const uint<N>& y) noexcept
 #if !defined(_MSC_VER) || _MSC_VER < 1916  // This kills MSVC 2017 compiler.
 inline constexpr bool operator<(const uint256& x, const uint256& y) noexcept
 {
-    const auto xhi = uint128{x[2], x[3]};
-    const auto xlo = uint128{x[0], x[1]};
-    const auto yhi = uint128{y[2], y[3]};
-    const auto ylo = uint128{y[0], y[1]};
-    return (unsigned(xhi < yhi) | (unsigned(xhi == yhi) & unsigned(xlo < ylo))) != 0;
+    auto xp = intx::uint128{x[2], x[3]};
+    auto yp = intx::uint128{y[2], y[3]};
+    if (xp == yp)
+    {
+        xp = intx::uint128{x[0], x[1]};
+        yp = intx::uint128{y[0], y[1]};
+    }
+    return xp < yp;
 }
 #endif
 
diff --git a/test/benchmarks/benchmarks.cpp b/test/benchmarks/benchmarks.cpp
index 69a8a32d..72cbdd6e 100644
--- a/test/benchmarks/benchmarks.cpp
+++ b/test/benchmarks/benchmarks.cpp
@@ -351,6 +351,11 @@ BENCHMARK_TEMPLATE(shift, uint512, uint64_t, shl_public)->DenseRange(-1, 3);
     return x < y;
 }
 
+[[gnu::noinline]] static bool lt_sub(const uint256& x, const uint256& y) noexcept
+{
+    return subc(x, y).carry;
+}
+
 [[gnu::noinline]] static bool lt_wordcmp(const uint256& x, const uint256& y) noexcept
 {
     for (size_t i = 3; i >= 1; --i)
@@ -418,6 +423,7 @@ static void compare(benchmark::State& state)
     }
 }
 BENCHMARK_TEMPLATE(compare, lt_public)->DenseRange(64, 256, 64);
+BENCHMARK_TEMPLATE(compare, lt_sub)->DenseRange(64, 256, 64);
 BENCHMARK_TEMPLATE(compare, lt_wordcmp)->DenseRange(64, 256, 64);
 BENCHMARK_TEMPLATE(compare, lt_halves)->DenseRange(64, 256, 64);
 #if INTX_HAS_EXTINT