Optimize integer to_chars with division by 100

benchmarks/src/integer_to_string.cpp

@@ -4,6 +4,7 @@
 #include <algorithm>
 #include <array>
 #include <benchmark/benchmark.h>
+#include <charconv>
 #include <cmath>
 #include <cstdint>
 #include <limits>
@@ -38,6 +39,26 @@ auto generate_array() {
     return a;
 }
 
+template <class T, double M, double S>
+void integer_to_chars(benchmark::State& state) {
+    auto a = generate_array<T, M, S>();
+    char buff[20]; // can hold -2^63 and 2^64 - 1
+
+    auto it = a.begin();
+    for (auto _ : state) {
+        auto i = *it;
+        benchmark::DoNotOptimize(i);
+        auto s = to_chars(begin(buff), end(buff), i);
+        benchmark::DoNotOptimize(s.ec);
+        benchmark::DoNotOptimize(s.ptr);
+
+        ++it;
+        if (it == a.end()) {
+            it = a.begin();
+        }
+    }
+}
+
 template <class CharT, class T, double M, double S>
 void internal_integer_to_buff(benchmark::State& state) {
     auto a = generate_array<T, M, S>();
@@ -77,6 +98,11 @@ void integer_to_string(benchmark::State& state) {
     }
 }
 
+BENCHMARK(integer_to_chars<uint8_t, 2.5, 1.5>);
+BENCHMARK(integer_to_chars<uint16_t, 5.0, 3.0>);
+BENCHMARK(integer_to_chars<uint32_t, 10.0, 6.0>);
+BENCHMARK(integer_to_chars<uint64_t, 20.0, 12.0>);
+
 BENCHMARK(internal_integer_to_buff<char, uint8_t, 2.5, 1.5>);
 BENCHMARK(internal_integer_to_buff<char, uint16_t, 5.0, 3.0>);
 BENCHMARK(internal_integer_to_buff<char, uint32_t, 10.0, 6.0>);

stl/inc/charconv

@@ -64,31 +64,63 @@ _NODISCARD _CONSTEXPR23 to_chars_result _Integer_to_chars(
     switch (_Base) {
     case 10:
         { // Derived from _UIntegral_to_buff()
-            // Performance note: Ryu's digit table should be faster here.
-            constexpr bool _Use_chunks = sizeof(_Unsigned) > sizeof(size_t);
-
-            if constexpr (_Use_chunks) { // For 64-bit numbers on 32-bit platforms, work in chunks to avoid 64-bit
-                                         // divisions.
-                while (_Value > 0xFFFF'FFFFU) {
-                    // Performance note: Ryu's division workaround would be faster here.
-                    unsigned long _Chunk = static_cast<unsigned long>(_Value % 1'000'000'000);
-                    _Value               = static_cast<_Unsigned>(_Value / 1'000'000'000);
-
-                    for (int _Idx = 0; _Idx != 9; ++_Idx) {
-                        *--_RNext = static_cast<char>('0' + _Chunk % 10);
-                        _Chunk /= 10;
+#ifdef _WIN64
+            auto _Trunc = _Value;
+#else // ^^^ 64-bit / 32-bit vvv
+            if constexpr (sizeof(_Unsigned) > 4) { // For 64-bit numbers, work in chunks to avoid 64-bit divisions.
+                while (_Value > 0xFFFFFFFFU) {
+                    auto _Value_chunk = static_cast<unsigned long>(_Value % 100000000);
+                    _Value /= 100000000;
+
+                    for (int _Idx = 0; _Idx != 3; ++_Idx) {
+                        const unsigned long _Value_chunk_part = _Value_chunk % 100;
+                        _Value_chunk /= 100;
+                        _RNext -= 2;
+                        if (!_STD _Is_constant_evaluated()) {
+                            _CSTD memcpy(_RNext, _Digit_pairs<char>._Data[_Value_chunk_part], 2);
+                        } else {
+                            _RNext[0] = _Digit_pairs<char>._Data[_Value_chunk_part][0];
+                            _RNext[1] = _Digit_pairs<char>._Data[_Value_chunk_part][1];
+                        }
+                    }
+
+                    _RNext -= 2;
+                    if (!_STD _Is_constant_evaluated()) {
+                        _CSTD memcpy(_RNext, _Digit_pairs<char>._Data[_Value_chunk], 2);
+                    } else {
+                        _RNext[0] = _Digit_pairs<char>._Data[_Value_chunk][0];
+                        _RNext[1] = _Digit_pairs<char>._Data[_Value_chunk][1];
                     }
                 }
             }
 
-            using _Truncated = conditional_t<_Use_chunks, unsigned long, _Unsigned>;
+            auto _Trunc = static_cast<unsigned long>(_Value);
+#endif // ^^^ 32-bit ^^^
 
-            _Truncated _Trunc = static_cast<_Truncated>(_Value);
+            // If we have a single digit, print [0, 9] and return. (This is necessary to correctly handle 0.)
+            if (_Trunc < 10) {
+                *--_RNext = static_cast<char>('0' + _Trunc);
+                break;
+            }
 
+            // Print one or more pairs of digits.
             do {
-                *--_RNext = static_cast<char>('0' + _Trunc % 10);
-                _Trunc /= 10;
-            } while (_Trunc != 0);
+                const auto _Trunc_part = static_cast<unsigned long>(_Trunc % 100);
+                _Trunc /= 100;
+                _RNext -= 2;
+                if (!_STD _Is_constant_evaluated()) {
+                    _CSTD memcpy(_RNext, _Digit_pairs<char>._Data[_Trunc_part], 2);
+                } else {
+                    _RNext[0] = _Digit_pairs<char>._Data[_Trunc_part][0];
+                    _RNext[1] = _Digit_pairs<char>._Data[_Trunc_part][1];
+                }
+            } while (_Trunc >= 10);
+
+            // If we have an unpaired digit, print it.
+            // For example, 1729 is printed as 17 29, and 19937 is printed as 1 99 37.
+            if (_Trunc != 0) {
+                *--_RNext = static_cast<char>('0' + _Trunc);
+            }
             break;
         }
 

stl/inc/xmemory

@@ -2775,24 +2775,6 @@ template <class>
 constexpr bool _Has_guaranteed_append_range = false; // N5032 [sequence.reqmts]/112; used by flat_(multi)set::insert.
 #endif // _HAS_CXX23
 
-template <class _Elem>
-struct _Digit_pair_table {
-    _Elem _Data[100][2];
-
-    constexpr explicit _Digit_pair_table() : _Data{} {
-        for (int _Idx = 0; _Idx != 100; ++_Idx) {
-            _Data[_Idx][0] = static_cast<_Elem>('0' + _Idx / 10);
-            _Data[_Idx][1] = static_cast<_Elem>('0' + _Idx % 10);
-        }
-    }
-
-    _Digit_pair_table(const _Digit_pair_table&)            = delete;
-    _Digit_pair_table& operator=(const _Digit_pair_table&) = delete;
-};
-
-template <class _Elem>
-constexpr _Digit_pair_table<_Elem> _Digit_pairs{};
-
 template <class _Elem, class _UTy>
 _NODISCARD _Elem* _UIntegral_to_buff(_Elem* _RNext, _UTy _UVal) { // used by both to_string and thread::id output
     // format _UVal into buffer *ending at* _RNext
@@ -2830,7 +2812,7 @@ _NODISCARD _Elem* _UIntegral_to_buff(_Elem* _RNext, _UTy _UVal) { // used by bot
 
     // Print one or more pairs of digits.
     do {
-        const unsigned long _UVal_trunc_part = _UVal_trunc % 100;
+        const auto _UVal_trunc_part = static_cast<unsigned long>(_UVal_trunc % 100);
         _UVal_trunc /= 100;
         _RNext -= 2;
         _CSTD memcpy(_RNext, _Digit_pairs<_Elem>._Data[_UVal_trunc_part], 2 * sizeof(_Elem));

stl/inc/xutility

@@ -7971,6 +7971,24 @@ constexpr bool _Equivalence_is_equality =
     _Is_any_of_v<_Compare, _RANGES less, _RANGES greater>
     || (_Is_any_of_v<_Compare, less<>, less<_Key>, greater<>, greater<_Key>> && _Equivalence_is_equality_impl<_Key>);
 #endif // _HAS_CXX23
+
+template <class _Elem>
+struct _Digit_pair_table {
+    _Elem _Data[100][2];
+
+    constexpr explicit _Digit_pair_table() : _Data{} {
+        for (int _Idx = 0; _Idx != 100; ++_Idx) {
+            _Data[_Idx][0] = static_cast<_Elem>('0' + _Idx / 10);
+            _Data[_Idx][1] = static_cast<_Elem>('0' + _Idx % 10);
+        }
+    }
+
+    _Digit_pair_table(const _Digit_pair_table&)            = delete;
+    _Digit_pair_table& operator=(const _Digit_pair_table&) = delete;
+};
+
+template <class _Elem>
+constexpr _Digit_pair_table<_Elem> _Digit_pairs{};
 _STD_END
 
 // TRANSITION, non-_Ugly attribute tokens