diff --git a/absl/numeric/int128_benchmark.cc b/absl/numeric/int128_benchmark.cc
index a5502d9..eab1515 100644
--- a/absl/numeric/int128_benchmark.cc
+++ b/absl/numeric/int128_benchmark.cc
@@ -12,15 +12,15 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "absl/numeric/int128.h"
-
 #include <algorithm>
 #include <cstdint>
+#include <limits>
 #include <random>
 #include <vector>
 
 #include "benchmark/benchmark.h"
 #include "absl/base/config.h"
+#include "absl/numeric/int128.h"
 
 namespace {
 
@@ -32,57 +32,85 @@
   return std::mt19937(seed);
 }
 
-std::vector<std::pair<absl::uint128, absl::uint128>>
-GetRandomClass128SampleUniformDivisor() {
-  std::vector<std::pair<absl::uint128, absl::uint128>> values;
+template <typename T,
+          typename H = typename std::conditional<
+              std::numeric_limits<T>::is_signed, int64_t, uint64_t>::type>
+std::vector<std::pair<T, T>> GetRandomClass128SampleUniformDivisor() {
+  std::vector<std::pair<T, T>> values;
   std::mt19937 random = MakeRandomEngine();
-  std::uniform_int_distribution<uint64_t> uniform_uint64;
+  std::uniform_int_distribution<H> uniform_h;
   values.reserve(kSampleSize);
   for (size_t i = 0; i < kSampleSize; ++i) {
-    absl::uint128 a =
-        absl::MakeUint128(uniform_uint64(random), uniform_uint64(random));
-    absl::uint128 b =
-        absl::MakeUint128(uniform_uint64(random), uniform_uint64(random));
-    values.emplace_back(std::max(a, b),
-                        std::max(absl::uint128(2), std::min(a, b)));
+    T a{absl::MakeUint128(uniform_h(random), uniform_h(random))};
+    T b{absl::MakeUint128(uniform_h(random), uniform_h(random))};
+    values.emplace_back(std::max(a, b), std::max(T(2), std::min(a, b)));
   }
   return values;
 }
 
+template <typename T>
 void BM_DivideClass128UniformDivisor(benchmark::State& state) {
-  auto values = GetRandomClass128SampleUniformDivisor();
+  auto values = GetRandomClass128SampleUniformDivisor<T>();
   while (state.KeepRunningBatch(values.size())) {
     for (const auto& pair : values) {
       benchmark::DoNotOptimize(pair.first / pair.second);
     }
   }
 }
-BENCHMARK(BM_DivideClass128UniformDivisor);
+BENCHMARK_TEMPLATE(BM_DivideClass128UniformDivisor, absl::uint128);
+BENCHMARK_TEMPLATE(BM_DivideClass128UniformDivisor, absl::int128);
 
-std::vector<std::pair<absl::uint128, uint64_t>>
-GetRandomClass128SampleSmallDivisor() {
-  std::vector<std::pair<absl::uint128, uint64_t>> values;
+template <typename T>
+void BM_RemainderClass128UniformDivisor(benchmark::State& state) {
+  auto values = GetRandomClass128SampleUniformDivisor<T>();
+  while (state.KeepRunningBatch(values.size())) {
+    for (const auto& pair : values) {
+      benchmark::DoNotOptimize(pair.first % pair.second);
+    }
+  }
+}
+BENCHMARK_TEMPLATE(BM_RemainderClass128UniformDivisor, absl::uint128);
+BENCHMARK_TEMPLATE(BM_RemainderClass128UniformDivisor, absl::int128);
+
+template <typename T,
+          typename H = typename std::conditional<
+              std::numeric_limits<T>::is_signed, int64_t, uint64_t>::type>
+std::vector<std::pair<T, H>> GetRandomClass128SampleSmallDivisor() {
+  std::vector<std::pair<T, H>> values;
   std::mt19937 random = MakeRandomEngine();
-  std::uniform_int_distribution<uint64_t> uniform_uint64;
+  std::uniform_int_distribution<H> uniform_h;
   values.reserve(kSampleSize);
   for (size_t i = 0; i < kSampleSize; ++i) {
-    absl::uint128 a =
-        absl::MakeUint128(uniform_uint64(random), uniform_uint64(random));
-    uint64_t b = std::max(uint64_t{2}, uniform_uint64(random));
-    values.emplace_back(std::max(a, absl::uint128(b)), b);
+    T a{absl::MakeUint128(uniform_h(random), uniform_h(random))};
+    H b{std::max(H{2}, uniform_h(random))};
+    values.emplace_back(std::max(a, T(b)), b);
   }
   return values;
 }
 
+template <typename T>
 void BM_DivideClass128SmallDivisor(benchmark::State& state) {
-  auto values = GetRandomClass128SampleSmallDivisor();
+  auto values = GetRandomClass128SampleSmallDivisor<T>();
   while (state.KeepRunningBatch(values.size())) {
     for (const auto& pair : values) {
       benchmark::DoNotOptimize(pair.first / pair.second);
     }
   }
 }
-BENCHMARK(BM_DivideClass128SmallDivisor);
+BENCHMARK_TEMPLATE(BM_DivideClass128SmallDivisor, absl::uint128);
+BENCHMARK_TEMPLATE(BM_DivideClass128SmallDivisor, absl::int128);
+
+template <typename T>
+void BM_RemainderClass128SmallDivisor(benchmark::State& state) {
+  auto values = GetRandomClass128SampleSmallDivisor<T>();
+  while (state.KeepRunningBatch(values.size())) {
+    for (const auto& pair : values) {
+      benchmark::DoNotOptimize(pair.first % pair.second);
+    }
+  }
+}
+BENCHMARK_TEMPLATE(BM_RemainderClass128SmallDivisor, absl::uint128);
+BENCHMARK_TEMPLATE(BM_RemainderClass128SmallDivisor, absl::int128);
 
 std::vector<std::pair<absl::uint128, absl::uint128>> GetRandomClass128Sample() {
   std::vector<std::pair<absl::uint128, absl::uint128>> values;
@@ -121,74 +149,107 @@
 
 // Some implementations of <random> do not support __int128 when it is
 // available, so we make our own uniform_int_distribution-like type.
+template <typename T,
+          typename H = typename std::conditional<
+              std::is_same<T, __int128>::value, int64_t, uint64_t>::type>
 class UniformIntDistribution128 {
  public:
   // NOLINTNEXTLINE: mimicking std::uniform_int_distribution API
-  unsigned __int128 operator()(std::mt19937& generator) {
-    return (static_cast<unsigned __int128>(dist64_(generator)) << 64) |
-           dist64_(generator);
+  T operator()(std::mt19937& generator) {
+    return (static_cast<T>(dist64_(generator)) << 64) | dist64_(generator);
   }
 
  private:
-  std::uniform_int_distribution<uint64_t> dist64_;
+  std::uniform_int_distribution<H> dist64_;
 };
 
-std::vector<std::pair<unsigned __int128, unsigned __int128>>
-GetRandomIntrinsic128SampleUniformDivisor() {
-  std::vector<std::pair<unsigned __int128, unsigned __int128>> values;
+template <typename T,
+          typename H = typename std::conditional<
+              std::is_same<T, __int128>::value, int64_t, uint64_t>::type>
+std::vector<std::pair<T, T>> GetRandomIntrinsic128SampleUniformDivisor() {
+  std::vector<std::pair<T, T>> values;
   std::mt19937 random = MakeRandomEngine();
-  UniformIntDistribution128 uniform_uint128;
+  UniformIntDistribution128<T> uniform_128;
   values.reserve(kSampleSize);
   for (size_t i = 0; i < kSampleSize; ++i) {
-    unsigned __int128 a = uniform_uint128(random);
-    unsigned __int128 b = uniform_uint128(random);
-    values.emplace_back(
-        std::max(a, b),
-        std::max(static_cast<unsigned __int128>(2), std::min(a, b)));
+    T a = uniform_128(random);
+    T b = uniform_128(random);
+    values.emplace_back(std::max(a, b),
+                        std::max(static_cast<T>(2), std::min(a, b)));
   }
   return values;
 }
 
+template <typename T>
 void BM_DivideIntrinsic128UniformDivisor(benchmark::State& state) {
-  auto values = GetRandomIntrinsic128SampleUniformDivisor();
+  auto values = GetRandomIntrinsic128SampleUniformDivisor<T>();
   while (state.KeepRunningBatch(values.size())) {
     for (const auto& pair : values) {
       benchmark::DoNotOptimize(pair.first / pair.second);
     }
   }
 }
-BENCHMARK(BM_DivideIntrinsic128UniformDivisor);
+BENCHMARK_TEMPLATE(BM_DivideIntrinsic128UniformDivisor, unsigned __int128);
+BENCHMARK_TEMPLATE(BM_DivideIntrinsic128UniformDivisor, __int128);
 
-std::vector<std::pair<unsigned __int128, uint64_t>>
-GetRandomIntrinsic128SampleSmallDivisor() {
-  std::vector<std::pair<unsigned __int128, uint64_t>> values;
+template <typename T>
+void BM_RemainderIntrinsic128UniformDivisor(benchmark::State& state) {
+  auto values = GetRandomIntrinsic128SampleUniformDivisor<T>();
+  while (state.KeepRunningBatch(values.size())) {
+    for (const auto& pair : values) {
+      benchmark::DoNotOptimize(pair.first % pair.second);
+    }
+  }
+}
+BENCHMARK_TEMPLATE(BM_RemainderIntrinsic128UniformDivisor, unsigned __int128);
+BENCHMARK_TEMPLATE(BM_RemainderIntrinsic128UniformDivisor, __int128);
+
+template <typename T,
+          typename H = typename std::conditional<
+              std::is_same<T, __int128>::value, int64_t, uint64_t>::type>
+std::vector<std::pair<T, H>> GetRandomIntrinsic128SampleSmallDivisor() {
+  std::vector<std::pair<T, H>> values;
   std::mt19937 random = MakeRandomEngine();
-  UniformIntDistribution128 uniform_uint128;
-  std::uniform_int_distribution<uint64_t> uniform_uint64;
+  UniformIntDistribution128<T> uniform_int128;
+  std::uniform_int_distribution<H> uniform_int64;
   values.reserve(kSampleSize);
   for (size_t i = 0; i < kSampleSize; ++i) {
-    unsigned __int128 a = uniform_uint128(random);
-    uint64_t b = std::max(uint64_t{2}, uniform_uint64(random));
-    values.emplace_back(std::max(a, static_cast<unsigned __int128>(b)), b);
+    T a = uniform_int128(random);
+    H b = std::max(H{2}, uniform_int64(random));
+    values.emplace_back(std::max(a, static_cast<T>(b)), b);
   }
   return values;
 }
 
+template <typename T>
 void BM_DivideIntrinsic128SmallDivisor(benchmark::State& state) {
-  auto values = GetRandomIntrinsic128SampleSmallDivisor();
+  auto values = GetRandomIntrinsic128SampleSmallDivisor<T>();
   while (state.KeepRunningBatch(values.size())) {
     for (const auto& pair : values) {
       benchmark::DoNotOptimize(pair.first / pair.second);
     }
   }
 }
-BENCHMARK(BM_DivideIntrinsic128SmallDivisor);
+BENCHMARK_TEMPLATE(BM_DivideIntrinsic128SmallDivisor, unsigned __int128);
+BENCHMARK_TEMPLATE(BM_DivideIntrinsic128SmallDivisor, __int128);
+
+template <typename T>
+void BM_RemainderIntrinsic128SmallDivisor(benchmark::State& state) {
+  auto values = GetRandomIntrinsic128SampleSmallDivisor<T>();
+  while (state.KeepRunningBatch(values.size())) {
+    for (const auto& pair : values) {
+      benchmark::DoNotOptimize(pair.first % pair.second);
+    }
+  }
+}
+BENCHMARK_TEMPLATE(BM_RemainderIntrinsic128SmallDivisor, unsigned __int128);
+BENCHMARK_TEMPLATE(BM_RemainderIntrinsic128SmallDivisor, __int128);
 
 std::vector<std::pair<unsigned __int128, unsigned __int128>>
       GetRandomIntrinsic128Sample() {
   std::vector<std::pair<unsigned __int128, unsigned __int128>> values;
   std::mt19937 random = MakeRandomEngine();
-  UniformIntDistribution128 uniform_uint128;
+  UniformIntDistribution128<unsigned __int128> uniform_uint128;
   values.reserve(kSampleSize);
   for (size_t i = 0; i < kSampleSize; ++i) {
     values.emplace_back(uniform_uint128(random), uniform_uint128(random));