// Copyright 2017 The Abseil Authors. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. // You may obtain a copy of the License at // // http://www.apache.org/licenses/LICENSE-2.0 // // Unless required by applicable law or agreed to in writing, software // distributed under the License is distributed on an "AS IS" BASIS, // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // See the License for the specific language governing permissions and // limitations under the License. #include "absl/strings/internal/fastmem.h" #include #include #include #include "base/init_google.h" #include "base/logging.h" #include "testing/base/public/benchmark.h" #include "gtest/gtest.h" namespace { using RandomEngine = std::minstd_rand0; void VerifyResults(const int r1, const int r2, const std::string& a, const std::string& b) { CHECK_EQ(a.size(), b.size()); if (r1 == 0) { EXPECT_EQ(r2, 0) << a << " " << b; } else if (r1 > 0) { EXPECT_GT(r2, 0) << a << " " << b; } else { EXPECT_LT(r2, 0) << a << " " << b; } if ((r1 == 0) == (r2 == 0)) { EXPECT_EQ(r1 == 0, absl::strings_internal::memeq(a.data(), b.data(), a.size())) << r1 << " " << a << " " << b; } } // Check correctness against glibc's memcmp implementation void CheckSingle(const std::string& a, const std::string& b) { CHECK_EQ(a.size(), b.size()); const int r1 = memcmp(a.data(), b.data(), a.size()); const int r2 = absl::strings_internal::fastmemcmp_inlined(a.data(), b.data(), a.size()); VerifyResults(r1, r2, a, b); } void GenerateString(size_t len, std::string* s) { s->clear(); for (int i = 0; i < len; i++) { *s += ('a' + (i % 26)); } } void CheckCompare(const std::string& a, const std::string& b) { CheckSingle(a, b); for (int common = 0; common <= 32; common++) { std::string extra; GenerateString(common, &extra); CheckSingle(extra + a, extra + b); CheckSingle(a + extra, b + extra); for (char c1 = 'a'; c1 <= 'c'; c1++) { for (char c2 = 'a'; c2 <= 'c'; c2++) { CheckSingle(extra + c1 + a, extra + c2 + b); } } } } TEST(FastCompare, Misc) { CheckCompare("", ""); CheckCompare("a", "a"); CheckCompare("ab", "ab"); CheckCompare("abc", "abc"); CheckCompare("abcd", "abcd"); CheckCompare("abcde", "abcde"); CheckCompare("a", "x"); CheckCompare("ab", "xb"); CheckCompare("abc", "xbc"); CheckCompare("abcd", "xbcd"); CheckCompare("abcde", "xbcde"); CheckCompare("x", "a"); CheckCompare("xb", "ab"); CheckCompare("xbc", "abc"); CheckCompare("xbcd", "abcd"); CheckCompare("xbcde", "abcde"); CheckCompare("a", "x"); CheckCompare("ab", "ax"); CheckCompare("abc", "abx"); CheckCompare("abcd", "abcx"); CheckCompare("abcde", "abcdx"); CheckCompare("x", "a"); CheckCompare("ax", "ab"); CheckCompare("abx", "abc"); CheckCompare("abcx", "abcd"); CheckCompare("abcdx", "abcde"); for (int len = 0; len < 1000; len++) { std::string p(len, 'z'); CheckCompare(p + "x", p + "a"); CheckCompare(p + "ax", p + "ab"); CheckCompare(p + "abx", p + "abc"); CheckCompare(p + "abcx", p + "abcd"); CheckCompare(p + "abcdx", p + "abcde"); } } TEST(FastCompare, TrailingByte) { for (int i = 0; i < 256; i++) { for (int j = 0; j < 256; j++) { std::string a(1, i); std::string b(1, j); CheckSingle(a, b); } } } // Check correctness of memcpy_inlined. void CheckSingleMemcpyInlined(const std::string& a) { std::unique_ptr destination(new char[a.size() + 2]); destination[0] = 'x'; destination[a.size() + 1] = 'x'; absl::strings_internal::memcpy_inlined(destination.get() + 1, a.data(), a.size()); CHECK_EQ('x', destination[0]); CHECK_EQ('x', destination[a.size() + 1]); CHECK_EQ(0, memcmp(a.data(), destination.get() + 1, a.size())); } TEST(MemCpyInlined, Misc) { CheckSingleMemcpyInlined(""); CheckSingleMemcpyInlined("0"); CheckSingleMemcpyInlined("012"); CheckSingleMemcpyInlined("0123"); CheckSingleMemcpyInlined("01234"); CheckSingleMemcpyInlined("012345"); CheckSingleMemcpyInlined("0123456"); CheckSingleMemcpyInlined("01234567"); CheckSingleMemcpyInlined("012345678"); CheckSingleMemcpyInlined("0123456789"); CheckSingleMemcpyInlined("0123456789a"); CheckSingleMemcpyInlined("0123456789ab"); CheckSingleMemcpyInlined("0123456789abc"); CheckSingleMemcpyInlined("0123456789abcd"); CheckSingleMemcpyInlined("0123456789abcde"); CheckSingleMemcpyInlined("0123456789abcdef"); CheckSingleMemcpyInlined("0123456789abcdefg"); } template inline void CopyLoop(benchmark::State& state, int size, Function func) { char* src = new char[size]; char* dst = new char[size]; memset(src, 'x', size); memset(dst, 'y', size); for (auto _ : state) { func(dst, src, size); } state.SetBytesProcessed(static_cast(state.iterations()) * size); CHECK_EQ(dst[0], 'x'); delete[] src; delete[] dst; } void BM_memcpy(benchmark::State& state) { CopyLoop(state, state.range(0), memcpy); } BENCHMARK(BM_memcpy)->DenseRange(1, 18)->Range(32, 8 << 20); void BM_memcpy_inlined(benchmark::State& state) { CopyLoop(state, state.range(0), absl::strings_internal::memcpy_inlined); } BENCHMARK(BM_memcpy_inlined)->DenseRange(1, 18)->Range(32, 8 << 20); // unaligned memcpy void BM_unaligned_memcpy(benchmark::State& state) { const int n = state.range(0); const int kMaxOffset = 32; char* src = new char[n + kMaxOffset]; char* dst = new char[n + kMaxOffset]; memset(src, 'x', n + kMaxOffset); int r = 0, i = 0; for (auto _ : state) { memcpy(dst + (i % kMaxOffset), src + ((i + 5) % kMaxOffset), n); r += dst[0]; ++i; } state.SetBytesProcessed(static_cast(state.iterations()) * n); delete[] src; delete[] dst; benchmark::DoNotOptimize(r); } BENCHMARK(BM_unaligned_memcpy)->DenseRange(1, 18)->Range(32, 8 << 20); // memmove worst case: heavy overlap, but not always by the same amount. // Also, the source and destination will often be unaligned. void BM_memmove_worst_case(benchmark::State& state) { const int n = state.range(0); const int32_t kDeterministicSeed = 301; const int kMaxOffset = 32; char* src = new char[n + kMaxOffset]; memset(src, 'x', n + kMaxOffset); size_t offsets[64]; RandomEngine rng(kDeterministicSeed); std::uniform_int_distribution random_to_max_offset(0, kMaxOffset); for (size_t& offset : offsets) { offset = random_to_max_offset(rng); } int r = 0, i = 0; for (auto _ : state) { memmove(src + offsets[i], src + offsets[i + 1], n); r += src[0]; i = (i + 2) % arraysize(offsets); } state.SetBytesProcessed(static_cast(state.iterations()) * n); delete[] src; benchmark::DoNotOptimize(r); } BENCHMARK(BM_memmove_worst_case)->DenseRange(1, 18)->Range(32, 8 << 20); // memmove cache-friendly: aligned and overlapping with 4k // between the source and destination addresses. void BM_memmove_cache_friendly(benchmark::State& state) { const int n = state.range(0); char* src = new char[n + 4096]; memset(src, 'x', n); int r = 0; while (state.KeepRunningBatch(2)) { // count each memmove as an iteration memmove(src + 4096, src, n); memmove(src, src + 4096, n); r += src[0]; } state.SetBytesProcessed(static_cast(state.iterations()) * n); delete[] src; benchmark::DoNotOptimize(r); } BENCHMARK(BM_memmove_cache_friendly) ->Arg(5 * 1024) ->Arg(10 * 1024) ->Range(16 << 10, 8 << 20); // memmove best(?) case: aligned and non-overlapping. void BM_memmove_aligned_non_overlapping(benchmark::State& state) { CopyLoop(state, state.range(0), memmove); } BENCHMARK(BM_memmove_aligned_non_overlapping) ->DenseRange(1, 18) ->Range(32, 8 << 20); // memset speed void BM_memset(benchmark::State& state) { const int n = state.range(0); char* dst = new char[n]; int r = 0; for (auto _ : state) { memset(dst, 'x', n); r += dst[0]; } state.SetBytesProcessed(static_cast(state.iterations()) * n); delete[] dst; benchmark::DoNotOptimize(r); } BENCHMARK(BM_memset)->Range(8, 4096 << 10); // Bandwidth (vectorization?) test: the ideal generated code will be limited // by memory bandwidth. Even so-so generated code will max out memory bandwidth // on some machines. void BM_membandwidth(benchmark::State& state) { const int n = state.range(0); CHECK_EQ(n % 32, 0); // We will read 32 bytes per iter. char* dst = new char[n]; int r = 0; for (auto _ : state) { const uint32_t* p = reinterpret_cast(dst); const uint32_t* limit = reinterpret_cast(dst + n); uint32_t x = 0; while (p < limit) { x += p[0]; x += p[1]; x += p[2]; x += p[3]; x += p[4]; x += p[5]; x += p[6]; x += p[7]; p += 8; } r += x; } state.SetBytesProcessed(static_cast(state.iterations()) * n); delete[] dst; benchmark::DoNotOptimize(r); } BENCHMARK(BM_membandwidth)->Range(32, 16384 << 10); // Helper for benchmarks. Repeatedly compares two strings that are // either equal or different only in one character. If test_equal_strings // is false then position_to_modify determines where the difference will be. template ABSL_ATTRIBUTE_ALWAYS_INLINE inline void StringCompareLoop( benchmark::State& state, bool test_equal_strings, std::string::size_type position_to_modify, int size, Function func) { const int kIterMult = 4; // Iteration multiplier for better timing resolution CHECK_GT(size, 0); const bool position_to_modify_is_valid = position_to_modify != std::string::npos && position_to_modify < size; CHECK_NE(position_to_modify_is_valid, test_equal_strings); if (!position_to_modify_is_valid) { position_to_modify = 0; } std::string sa(size, 'a'); std::string sb = sa; char last = sa[size - 1]; int num = 0; for (auto _ : state) { for (int i = 0; i < kIterMult; ++i) { sb[position_to_modify] = test_equal_strings ? last : last ^ 1; num += func(sa, sb); } } state.SetBytesProcessed(static_cast(state.iterations()) * size); benchmark::DoNotOptimize(num); } // Helper for benchmarks. Repeatedly compares two memory regions that are // either equal or different only in their final character. template ABSL_ATTRIBUTE_ALWAYS_INLINE inline void CompareLoop(benchmark::State& state, bool test_equal_strings, int size, Function func) { const int kIterMult = 4; // Iteration multiplier for better timing resolution CHECK_GT(size, 0); char* data = static_cast(malloc(size * 2)); memset(data, 'a', size * 2); char* a = data; char* b = data + size; char last = a[size - 1]; int num = 0; for (auto _ : state) { for (int i = 0; i < kIterMult; ++i) { b[size - 1] = test_equal_strings ? last : last ^ 1; num += func(a, b, size); } } state.SetBytesProcessed(static_cast(state.iterations()) * size); benchmark::DoNotOptimize(num); free(data); } void BM_memcmp(benchmark::State& state) { CompareLoop(state, false, state.range(0), memcmp); } BENCHMARK(BM_memcmp)->DenseRange(1, 9)->Range(32, 8 << 20); void BM_fastmemcmp_inlined(benchmark::State& state) { CompareLoop(state, false, state.range(0), absl::strings_internal::fastmemcmp_inlined); } BENCHMARK(BM_fastmemcmp_inlined)->DenseRange(1, 9)->Range(32, 8 << 20); void BM_memeq(benchmark::State& state) { CompareLoop(state, false, state.range(0), absl::strings_internal::memeq); } BENCHMARK(BM_memeq)->DenseRange(1, 9)->Range(32, 8 << 20); void BM_memeq_equal(benchmark::State& state) { CompareLoop(state, true, state.range(0), absl::strings_internal::memeq); } BENCHMARK(BM_memeq_equal)->DenseRange(1, 9)->Range(32, 8 << 20); bool StringLess(const std::string& x, const std::string& y) { return x < y; } bool StringEqual(const std::string& x, const std::string& y) { return x == y; } bool StdEqual(const std::string& x, const std::string& y) { return x.size() == y.size() && std::equal(x.data(), x.data() + x.size(), y.data()); } // Benchmark for x < y, where x and y are strings that differ in only their // final char. That should be more-or-less the worst case for <. void BM_string_less(benchmark::State& state) { StringCompareLoop(state, false, state.range(0) - 1, state.range(0), StringLess); } BENCHMARK(BM_string_less)->DenseRange(1, 9)->Range(32, 1 << 20); // Benchmark for x < y, where x and y are strings that differ in only their // first char. That should be more-or-less the best case for <. void BM_string_less_easy(benchmark::State& state) { StringCompareLoop(state, false, 0, state.range(0), StringLess); } BENCHMARK(BM_string_less_easy)->DenseRange(1, 9)->Range(32, 1 << 20); void BM_string_equal(benchmark::State& state) { StringCompareLoop(state, false, state.range(0) - 1, state.range(0), StringEqual); } BENCHMARK(BM_string_equal)->DenseRange(1, 9)->Range(32, 1 << 20); void BM_string_equal_equal(benchmark::State& state) { StringCompareLoop(state, true, std::string::npos, state.range(0), StringEqual); } BENCHMARK(BM_string_equal_equal)->DenseRange(1, 9)->Range(32, 1 << 20); void BM_std_equal(benchmark::State& state) { StringCompareLoop(state, false, state.range(0) - 1, state.range(0), StdEqual); } BENCHMARK(BM_std_equal)->DenseRange(1, 9)->Range(32, 1 << 20); void BM_std_equal_equal(benchmark::State& state) { StringCompareLoop(state, true, std::string::npos, state.range(0), StdEqual); } BENCHMARK(BM_std_equal_equal)->DenseRange(1, 9)->Range(32, 1 << 20); void BM_string_equal_unequal_lengths(benchmark::State& state) { const int size = state.range(0); std::string a(size, 'a'); std::string b(size + 1, 'a'); int count = 0; for (auto _ : state) { b[size - 1] = 'a'; count += (a == b); } benchmark::DoNotOptimize(count); } BENCHMARK(BM_string_equal_unequal_lengths)->Arg(1)->Arg(1 << 20); void BM_stdstring_equal_unequal_lengths(benchmark::State& state) { const int size = state.range(0); std::string a(size, 'a'); std::string b(size + 1, 'a'); int count = 0; for (auto _ : state) { b[size - 1] = 'a'; count += (a == b); } benchmark::DoNotOptimize(count); } BENCHMARK(BM_stdstring_equal_unequal_lengths)->Arg(1)->Arg(1 << 20); } // namespace