Add absl benchmarks for Windows platforms

.github/scripts/run_benchmarks.sh

@@ -12,6 +12,11 @@
 # needed. The drivers always exit non-zero (the Jamfile declares them run-fail),
 # so success is judged on the data set actually being written.
 #
+# --sources names extra translation units to build alongside the driver, which is
+# how the platforms with no packaged Abseil get one: point --cxxflags at a source
+# checkout and --sources at absl/numeric/int128.cc, the only file of it that
+# absl::uint128 and absl::int128 need.
+#
 #   .github/scripts/run_benchmarks.sh --compiler g++-14 --out /tmp/results
 #   .github/scripts/run_benchmarks.sh --msvc --elements 5000000 --out C:/results
 
@@ -20,6 +25,7 @@ set -e
 COMPILER=""
 EXTRA_CXXFLAGS=""
 EXTRA_LDFLAGS=""
+EXTRA_SOURCES=""
 ELEMENTS=20000000
 REPETITIONS=5
 OUT=""
@@ -30,6 +36,7 @@ while [ $# -gt 0 ]; do
         --compiler)    COMPILER="$2"; shift 2 ;;
         --cxxflags)    EXTRA_CXXFLAGS="$2"; shift 2 ;;
         --ldflags)     EXTRA_LDFLAGS="$2"; shift 2 ;;
+        --sources)     EXTRA_SOURCES="$2"; shift 2 ;;
         --elements)    ELEMENTS="$2"; shift 2 ;;
         --repetitions) REPETITIONS="$2"; shift 2 ;;
         --out)         OUT="$2"; shift 2 ;;
@@ -67,12 +74,14 @@ for sign in u128 i128; do
         # the widest vectors need.
         executable="./benchmark_$sign.exe"
         # shellcheck disable=SC2086
-        cl -nologo -std:c++20 -O2 -EHsc -DNDEBUG $INCLUDES "$source_file" \
-            -Fe:"benchmark_$sign.exe" -link -LARGEADDRESSAWARE
+        cl -nologo -std:c++20 -O2 -EHsc -DNDEBUG $EXTRA_CXXFLAGS $INCLUDES \
+            "$source_file" $EXTRA_SOURCES \
+            -Fe:"benchmark_$sign.exe" -link -LARGEADDRESSAWARE $EXTRA_LDFLAGS
     else
         executable="./benchmark_$sign"
         # shellcheck disable=SC2086
-        $COMPILER -std=c++20 -O2 -DNDEBUG $EXTRA_CXXFLAGS $INCLUDES "$source_file" \
+        $COMPILER -std=c++20 -O2 -DNDEBUG $EXTRA_CXXFLAGS $INCLUDES \
+            "$source_file" $EXTRA_SOURCES \
             -o "benchmark_$sign" $EXTRA_LDFLAGS
     fi
 

.github/workflows/benchmarks.yml

@@ -20,6 +20,13 @@
 # 128-bit vectors of the default length do not fit in a 32-bit address space. The
 # element count lands in the data set and is printed in the documentation, so the
 # tables stay honest about what was measured.
+#
+# Linux and macOS take Abseil from a package. Windows has none, so those jobs
+# check out the pinned Abseil release and build the single translation unit that
+# absl::uint128 and absl::int128 need, with the same flags as the rest of the
+# benchmark. Building all of Abseil to link one object would cost more CI time
+# than the benchmarks themselves, and compiling it here is what keeps the
+# comparison fair: both sides of every row come out of one -O2 build.
 
 #
 # Every pull request gets a full set of runs: the step summary of each job
@@ -65,6 +72,9 @@ env:
   ELEMENTS: ${{ inputs.elements || '20000000' }}
   EMULATED_ELEMENTS: ${{ inputs.emulated_elements || '2000000' }}
   REPETITIONS: ${{ inputs.repetitions || '5' }}
+  # The Abseil release the Windows jobs build against, latest LTS at the time of
+  # writing. Linux and macOS take whatever their package manager ships.
+  ABSEIL_TAG: '20260526.0'
 
 jobs:
   linux:
@@ -189,6 +199,11 @@ jobs:
             elements: ''
     steps:
       - uses: actions/checkout@v6
+      # Cloned next to the Boost tree that the following step creates, so that the
+      # build reaches both of them by relative path and nothing has to be
+      # translated between git-bash and cl.
+      - name: Fetch Abseil
+        run: git clone --depth 1 --branch "$ABSEIL_TAG" https://github.com/abseil/abseil-cpp.git ../abseil-cpp
       - name: Setup Boost
         run: .github/scripts/setup_benchmarks.sh
       - uses: TheMrMilchmann/setup-msvc-dev@v4
@@ -201,6 +216,8 @@ jobs:
           : "${COUNT:=$ELEMENTS}"
           "$GITHUB_WORKSPACE/.github/scripts/run_benchmarks.sh" \
             --msvc \
+            --cxxflags '-I../abseil-cpp' \
+            --sources '../abseil-cpp/absl/numeric/int128.cc' \
             --elements "$COUNT" \
             --repetitions "$REPETITIONS" \
             --out "$GITHUB_WORKSPACE/bench-results"

doc/modules/ROOT/pages/i128_benchmarks.adoc

@@ -19,6 +19,8 @@ On most platforms we use the builtin `\__int128` as the reference benchmark.
 When this is unavailable (such as on 32-bit architectures) we use `boost::multiprecision::int128` (abbreviated as `boost::mp::int128`) as it is widely used, and known to be portable.
 On MSVC platforms we use as reference `std::_Signed128` from the header `<__msvc_int128.hpp>` since this is bundled with their compiler.
 Lastly, when available, the benchmarks are also run with `absl::int128` which is the 128-bit signed integer from the https://abseil.io[Abseil Libraries].
+Linux and macOS take Abseil from a package; Windows has none, so it is built from source at a pinned release with the same compiler and optimization flags as the rest of the table.
+Abseil only uses a native 128-bit type when the compiler has one, which MSVC does not, so its portable path is what the Windows rows measure.
 
 NOTE: Each platform reports a single run, several of them from shared CI runners, so treat small differences as noise.
 
@@ -166,16 +168,16 @@ image::i128_graphs/linux/ARM32_relative_performance.png[ARM32 Relative Performan
 
 Measured with MSVC 14.51 (pass:[C++]20): 5 passes over 20,000,000 element pairs (100,000,000 operations per cell).
 
-[cols="1,>1,>1,>1"]
+[cols="1,>1,>1,>1,>1"]
 |===
-| Operation | `std::_Signed128` | `int128` | `boost::mp::int128`
+| Operation | `std::_Signed128` | `int128` | `boost::mp::int128` | `absl::int128`
 
-| Comparisons | 1,361,518 | 1,271,189 | 4,315,263
-| Addition | 115,135 | 107,633 | 2,089,612
-| Subtraction | 108,885 | 105,990 | 2,265,849
-| Multiplication | 196,081 | 113,761 | 3,024,735
-| Division | 1,031,262 | 1,241,974 | 4,191,052
-| Modulo | 1,080,810 | 1,327,820 | 4,139,305
+| Comparisons | 1,313,422 | 1,301,134 | 4,219,069 | 1,310,527
+| Addition | 104,308 | 105,315 | 2,122,367 | 92,929
+| Subtraction | 101,850 | 100,240 | 2,328,494 | 93,068
+| Multiplication | 167,872 | 112,184 | 3,004,661 | 114,658
+| Division | 938,908 | 1,233,042 | 4,238,571 | 17,581,284
+| Modulo | 1,101,512 | 1,330,229 | 4,187,828 | 17,626,117
 |===
 
 image::i128_graphs/windows/x64_relative_performance.png[x64 Relative Performance, width=100%]
@@ -184,16 +186,16 @@ image::i128_graphs/windows/x64_relative_performance.png[x64 Relative Performance
 
 Measured with MSVC 14.44 (pass:[C++]20): 5 passes over 20,000,000 element pairs (100,000,000 operations per cell).
 
-[cols="1,>1,>1,>1"]
+[cols="1,>1,>1,>1,>1"]
 |===
-| Operation | `std::_Signed128` | `int128` | `boost::mp::int128`
+| Operation | `std::_Signed128` | `int128` | `boost::mp::int128` | `absl::int128`
 
-| Comparisons | 1,191,566 | 1,125,261 | 2,871,780
-| Addition | 86,160 | 84,055 | 1,598,093
-| Subtraction | 88,263 | 87,934 | 1,769,736
-| Multiplication | 505,934 | 81,566 | 2,395,038
-| Division | 1,506,403 | 1,480,602 | 3,622,476
-| Modulo | 1,722,778 | 1,447,734 | 3,555,076
+| Comparisons | 1,163,962 | 1,124,237 | 2,892,406 | 1,176,305
+| Addition | 88,279 | 83,389 | 1,604,518 | 86,567
+| Subtraction | 88,880 | 86,568 | 1,766,980 | 88,994
+| Multiplication | 532,086 | 81,119 | 2,389,516 | 129,947
+| Division | 1,507,951 | 1,480,214 | 3,623,405 | 10,773,260
+| Modulo | 1,720,557 | 1,443,956 | 3,531,333 | 10,797,356
 |===
 
 image::i128_graphs/windows/ARM64_relative_performance.png[ARM64 Relative Performance, width=100%]
@@ -202,16 +204,16 @@ image::i128_graphs/windows/ARM64_relative_performance.png[ARM64 Relative Perform
 
 Measured with MSVC 14.51 (pass:[C++]20): 5 passes over 10,000,000 element pairs (50,000,000 operations per cell).
 
-[cols="1,>1,>1,>1"]
+[cols="1,>1,>1,>1,>1"]
 |===
-| Operation | `std::_Signed128` | `int128` | `boost::mp::int128`
+| Operation | `std::_Signed128` | `int128` | `boost::mp::int128` | `absl::int128`
 
-| Comparisons | 2,084,732 | 1,947,041 | 2,898,710
-| Addition | 131,344 | 128,063 | 1,325,852
-| Subtraction | 447,258 | 128,145 | 1,454,691
-| Multiplication | 877,318 | 149,394 | 2,018,981
-| Division | 3,479,471 | 3,113,092 | 3,948,298
-| Modulo | 2,683,535 | 3,157,295 | 3,620,409
+| Comparisons | 1,642,521 | 1,501,096 | 2,486,760 | 1,517,123
+| Addition | 73,033 | 58,313 | 1,218,478 | 55,079
+| Subtraction | 517,398 | 66,889 | 1,314,906 | 57,791
+| Multiplication | 970,404 | 170,877 | 2,302,831 | 3,865,486
+| Division | 2,840,529 | 2,663,730 | 3,362,691 | 9,883,544
+| Modulo | 2,390,312 | 2,752,450 | 3,255,144 | 9,862,440
 |===
 
 image::i128_graphs/windows/x86_relative_performance.png[x86 Relative Performance, width=100%]

doc/modules/ROOT/pages/u128_benchmarks.adoc

@@ -19,6 +19,8 @@ On most platforms we use the builtin `unsigned \__int128` as the reference bench
 When this is unavailable (such as on 32-bit architectures) we use `boost::multiprecision::uint128` (abbreviated as `boost::mp::uint128`) as it is widely used, and known to be portable.
 On MSVC platforms we use as reference `std::_Unsigned128` from the header `<__msvc_int128.hpp>` since this is bundled with their compiler.
 Lastly, when available, the benchmarks are also run with `absl::uint128` which is the 128-bit unsigned integer from the https://abseil.io[Abseil Libraries].
+Linux and macOS take Abseil from a package; Windows has none, so it is built from source at a pinned release with the same compiler and optimization flags as the rest of the table.
+Abseil only uses a native 128-bit type when the compiler has one, which MSVC does not, so its portable path is what the Windows rows measure.
 
 NOTE: Each platform reports a single run, several of them from shared CI runners, so treat small differences as noise.
 
@@ -166,16 +168,16 @@ image::u128_graphs/linux/ARM32_relative_performance.png[ARM32 Relative Performan
 
 Measured with MSVC 14.51 (pass:[C++]20): 5 passes over 20,000,000 element pairs (100,000,000 operations per cell).
 
-[cols="1,>1,>1,>1"]
+[cols="1,>1,>1,>1,>1"]
 |===
-| Operation | `std::_Unsigned128` | `uint128` | `boost::mp::uint128`
+| Operation | `std::_Unsigned128` | `uint128` | `boost::mp::uint128` | `absl::uint128`
 
-| Comparisons | 1,313,048 | 1,251,724 | 1,970,110
-| Addition | 106,229 | 98,075 | 2,153,287
-| Subtraction | 115,275 | 97,252 | 2,318,242
-| Multiplication | 115,091 | 112,082 | 2,428,936
-| Division | 1,127,057 | 1,086,078 | 3,340,168
-| Modulo | 1,249,176 | 1,144,617 | 2,721,519
+| Comparisons | 1,351,853 | 1,210,126 | 1,953,037 | 1,199,946
+| Addition | 102,157 | 89,607 | 2,241,970 | 91,530
+| Subtraction | 112,192 | 89,322 | 2,346,973 | 90,050
+| Multiplication | 116,473 | 109,967 | 2,262,252 | 110,565
+| Division | 1,135,328 | 1,077,802 | 3,443,408 | 20,871,271
+| Modulo | 1,271,929 | 1,165,839 | 2,816,863 | 20,945,166
 |===
 
 image::u128_graphs/windows/x64_relative_performance.png[x64 Relative Performance, width=100%]
@@ -184,16 +186,16 @@ image::u128_graphs/windows/x64_relative_performance.png[x64 Relative Performance
 
 Measured with MSVC 14.44 (pass:[C++]20): 5 passes over 20,000,000 element pairs (100,000,000 operations per cell).
 
-[cols="1,>1,>1,>1"]
+[cols="1,>1,>1,>1,>1"]
 |===
-| Operation | `std::_Unsigned128` | `uint128` | `boost::mp::uint128`
+| Operation | `std::_Unsigned128` | `uint128` | `boost::mp::uint128` | `absl::uint128`
 
-| Comparisons | 1,262,216 | 753,322 | 1,757,122
-| Addition | 86,038 | 84,791 | 1,535,904
-| Subtraction | 86,521 | 87,885 | 1,712,929
-| Multiplication | 458,107 | 82,947 | 1,780,541
-| Division | 1,334,323 | 920,483 | 2,843,970
-| Modulo | 1,491,426 | 915,935 | 2,420,312
+| Comparisons | 1,198,039 | 749,538 | 1,780,083 | 1,146,973
+| Addition | 88,759 | 85,378 | 1,534,492 | 86,222
+| Subtraction | 86,292 | 84,435 | 1,707,180 | 92,099
+| Multiplication | 457,638 | 83,669 | 1,785,505 | 127,263
+| Division | 1,337,104 | 921,115 | 2,824,979 | 12,558,253
+| Modulo | 1,488,718 | 918,404 | 2,433,789 | 12,522,017
 |===
 
 image::u128_graphs/windows/ARM64_relative_performance.png[ARM64 Relative Performance, width=100%]
@@ -202,16 +204,16 @@ image::u128_graphs/windows/ARM64_relative_performance.png[ARM64 Relative Perform
 
 Measured with MSVC 14.51 (pass:[C++]20): 5 passes over 10,000,000 element pairs (50,000,000 operations per cell).
 
-[cols="1,>1,>1,>1"]
+[cols="1,>1,>1,>1,>1"]
 |===
-| Operation | `std::_Unsigned128` | `uint128` | `boost::mp::uint128`
+| Operation | `std::_Unsigned128` | `uint128` | `boost::mp::uint128` | `absl::uint128`
 
-| Comparisons | 2,519,936 | 2,006,459 | 1,516,713
-| Addition | 132,021 | 129,303 | 1,184,928
-| Subtraction | 505,196 | 127,036 | 1,526,177
-| Multiplication | 765,417 | 960,622 | 1,459,692
-| Division | 2,510,903 | 2,217,091 | 2,816,842
-| Modulo | 1,985,290 | 2,271,454 | 2,458,916
+| Comparisons | 1,978,828 | 1,761,531 | 1,039,171 | 1,828,790
+| Addition | 71,811 | 56,033 | 1,029,278 | 60,322
+| Subtraction | 574,255 | 61,036 | 1,351,338 | 59,280
+| Multiplication | 938,730 | 1,121,940 | 1,741,827 | 1,668,286
+| Division | 2,111,143 | 1,944,458 | 2,592,836 | 11,246,769
+| Modulo | 1,858,981 | 1,985,261 | 2,215,040 | 11,258,901
 |===
 
 image::u128_graphs/windows/x86_relative_performance.png[x86 Relative Performance, width=100%]

test/benchmark_i128.cpp

@@ -22,8 +22,8 @@
 #define BOOST_INT128_HAS_MSVC_INTERNAL_I128
 #endif
 
-// Abseil requires at least C++17 (at time of writing)
-#if __has_include(<absl/numeric/int128.h>) && defined(__cplusplus) && __cplusplus >= 201703L
+#if __has_include(<absl/numeric/int128.h>) && \
+    (__cplusplus >= 201703L || (defined(_MSVC_LANG) && _MSVC_LANG >= 201703L))
 #  include <absl/numeric/int128.h>
 #  ifndef __i386__
 #    define BOOST_INT128_BENCHMARK_ABSL

test/benchmark_u128.cpp

@@ -23,8 +23,8 @@
 #define BOOST_INT128_HAS_MSVC_INTERNAL_I128
 #endif
 
-// Abseil requires at least C++17 (at time of writing)
-#if __has_include(<absl/numeric/int128.h>) && defined(__cplusplus) && __cplusplus >= 201703L
+#if __has_include(<absl/numeric/int128.h>) && \
+    (__cplusplus >= 201703L || (defined(_MSVC_LANG) && _MSVC_LANG >= 201703L))
 #  include <absl/numeric/int128.h>
 #  ifndef __i386__
 #    define BOOST_INT128_BENCHMARK_ABSL