initial version with working constexpr for c++20 compliant compilers

Author: jrahlf

include/fast_float/ascii_number.h

@@ -6,16 +6,21 @@
 #include <cstdint>
 #include <cstring>
 #include <iterator>
+#include <version>
+
+#if defined(__cpp_lib_bit_cast)
+#include <bit>
+#endif
 
 #include "float_common.h"
 
 namespace fast_float {
 
 // Next function can be micro-optimized, but compilers are entirely
 // able to optimize it well.
-fastfloat_really_inline bool is_integer(char c)  noexcept  { return c >= '0' && c <= '9'; }
+CXX20_CONSTEXPR fastfloat_really_inline bool is_integer(char c)  noexcept  { return c >= '0' && c <= '9'; }
 
-fastfloat_really_inline uint64_t byteswap(uint64_t val) {
+CXX20_CONSTEXPR fastfloat_really_inline uint64_t byteswap(uint64_t val) {
   return (val & 0xFF00000000000000) >> 56
     | (val & 0x00FF000000000000) >> 40
     | (val & 0x0000FF0000000000) >> 24
@@ -26,26 +31,40 @@ fastfloat_really_inline uint64_t byteswap(uint64_t val) {
     | (val & 0x00000000000000FF) << 56;
 }
 
-fastfloat_really_inline uint64_t read_u64(const char *chars) {
+CXX20_CONSTEXPR fastfloat_really_inline uint64_t read_u64(const char *chars) {
   uint64_t val;
+#if defined(__cpp_lib_bit_cast)
+  val = std::bit_cast<uint64_t>(reinterpret_cast<const char (&)[8]>(chars));
+#else
   ::memcpy(&val, chars, sizeof(uint64_t));
+#endif
 #if FASTFLOAT_IS_BIG_ENDIAN == 1
   // Need to read as-if the number was in little-endian order.
   val = byteswap(val);
 #endif
   return val;
 }
 
-fastfloat_really_inline void write_u64(uint8_t *chars, uint64_t val) {
+CXX20_CONSTEXPR fastfloat_really_inline void write_u64(uint8_t *chars, uint64_t val) {
 #if FASTFLOAT_IS_BIG_ENDIAN == 1
   // Need to read as-if the number was in little-endian order.
   val = byteswap(val);
 #endif
+#if defined(__cpp_lib_bit_cast)
+  if (std::is_constant_evaluated()) {
+    char (&dst)[8] = reinterpret_cast<char (&)[8]>(chars);
+    const char (&src)[8] = reinterpret_cast<const char (&)[8]>(val);
+    std::copy(std::begin(src), std::end(src), std::begin(dst));
+  } else {
+    ::memcpy(chars, &val, sizeof(uint64_t));
+  }
+#else
   ::memcpy(chars, &val, sizeof(uint64_t));
+#endif
 }
 
 // credit  @aqrit
-fastfloat_really_inline uint32_t  parse_eight_digits_unrolled(uint64_t val) {
+CXX20_CONSTEXPR fastfloat_really_inline uint32_t  parse_eight_digits_unrolled(uint64_t val) {
   const uint64_t mask = 0x000000FF000000FF;
   const uint64_t mul1 = 0x000F424000000064; // 100 + (1000000ULL << 32)
   const uint64_t mul2 = 0x0000271000000001; // 1 + (10000ULL << 32)
@@ -55,17 +74,17 @@ fastfloat_really_inline uint32_t  parse_eight_digits_unrolled(uint64_t val) {
   return uint32_t(val);
 }
 
-fastfloat_really_inline uint32_t parse_eight_digits_unrolled(const char *chars)  noexcept  {
+CXX20_CONSTEXPR fastfloat_really_inline uint32_t parse_eight_digits_unrolled(const char *chars)  noexcept  {
   return parse_eight_digits_unrolled(read_u64(chars));
 }
 
 // credit @aqrit
-fastfloat_really_inline bool is_made_of_eight_digits_fast(uint64_t val)  noexcept  {
+CXX20_CONSTEXPR fastfloat_really_inline bool is_made_of_eight_digits_fast(uint64_t val)  noexcept  {
   return !((((val + 0x4646464646464646) | (val - 0x3030303030303030)) &
      0x8080808080808080));
 }
 
-fastfloat_really_inline bool is_made_of_eight_digits_fast(const char *chars)  noexcept  {
+CXX20_CONSTEXPR fastfloat_really_inline bool is_made_of_eight_digits_fast(const char *chars)  noexcept  {
   return is_made_of_eight_digits_fast(read_u64(chars));
 }
 
@@ -81,7 +100,7 @@ struct parsed_number_string {
 
 // Assuming that you use no more than 19 digits, this will
 // parse an ASCII string.
-fastfloat_really_inline
+CXX20_CONSTEXPR fastfloat_really_inline
 parsed_number_string parse_number_string(const char *p, const char *pend, parse_options options) noexcept {
   const chars_format fmt = options.format;
   const char decimal_point = options.decimal_point;
@@ -221,7 +240,7 @@ parsed_number_string parse_number_string(const char *p, const char *pend, parse_
 // This function could be optimized. In particular, we could stop after 19 digits
 // and try to bail out. Furthermore, we should be able to recover the computed
 // exponent from the pass in parse_number_string.
-fastfloat_really_inline decimal parse_decimal(const char *p, const char *pend, parse_options options) noexcept {
+CXX20_CONSTEXPR fastfloat_really_inline decimal parse_decimal(const char *p, const char *pend, parse_options options) noexcept {
   const char decimal_point = options.decimal_point;
 
   decimal answer;

include/fast_float/decimal_to_binary.h

@@ -18,7 +18,7 @@ namespace fast_float {
 // low part corresponding to the least significant bits.
 //
 template <int bit_precision>
-fastfloat_really_inline
+CXX20_CONSTEXPR fastfloat_really_inline
 value128 compute_product_approximation(int64_t q, uint64_t w) {
   const int index = 2 * int(q - powers::smallest_power_of_five);
   // For small values of q, e.g., q in [0,27], the answer is always exact because
@@ -56,7 +56,7 @@ namespace detail {
  * where
  *   p = log(5**-q)/log(2) = -q * log(5)/log(2)
  */
-  fastfloat_really_inline int power(int q)  noexcept  {
+  constexpr fastfloat_really_inline int power(int q)  noexcept  {
     return (((152170 + 65536) * q) >> 16) + 63;
   }
 } // namespace detail
@@ -68,7 +68,7 @@ namespace detail {
 // return an adjusted_mantissa with a negative power of 2: the caller should recompute
 // in such cases.
 template <typename binary>
-fastfloat_really_inline
+CXX20_CONSTEXPR fastfloat_really_inline
 adjusted_mantissa compute_float(int64_t q, uint64_t w)  noexcept  {
   adjusted_mantissa answer;
   if ((w == 0) || (q < binary::smallest_power_of_ten())) {

include/fast_float/fast_float.h

@@ -1,7 +1,15 @@
 #ifndef FASTFLOAT_FAST_FLOAT_H
 #define FASTFLOAT_FAST_FLOAT_H
 
+
 #include <system_error>
+#include <version>
+
+#if defined(__cpp_lib_bit_cast)
+#define CXX20_CONSTEXPR constexpr
+#else
+#define CXX20_CONSTEXPR
+#endif
 
 namespace fast_float {
 enum chars_format {
@@ -18,7 +26,7 @@ struct from_chars_result {
 };
 
 struct parse_options {
-  explicit parse_options(chars_format fmt = chars_format::general,
+  constexpr explicit parse_options(chars_format fmt = chars_format::general,
                          char dot = '.')
     : format(fmt), decimal_point(dot) {}
 
@@ -48,14 +56,14 @@ struct parse_options {
  * The default is  `fast_float::chars_format::general` which allows both `fixed` and `scientific`.
  */
 template<typename T>
-from_chars_result from_chars(const char *first, const char *last,
+CXX20_CONSTEXPR from_chars_result from_chars(const char *first, const char *last,
                              T &value, chars_format fmt = chars_format::general)  noexcept;
 
 /**
  * Like from_chars, but accepts an `options` argument to govern number parsing.
  */
 template<typename T>
-from_chars_result from_chars_advanced(const char *first, const char *last,
+CXX20_CONSTEXPR from_chars_result from_chars_advanced(const char *first, const char *last,
                                       T &value, parse_options options)  noexcept;
 
 }

include/fast_float/float_common.h

@@ -4,6 +4,7 @@
 #include <cfloat>
 #include <cstdint>
 #include <cassert>
+#include <version>
 
 #if (defined(__x86_64) || defined(__x86_64__) || defined(_M_X64)   \
        || defined(__amd64) || defined(__aarch64__) || defined(_M_ARM64) \
@@ -73,10 +74,18 @@
 #define fastfloat_really_inline inline __attribute__((always_inline))
 #endif
 
+#if !defined(CXX20_CONSTEXPR)
+#if defined(__cpp_lib_bit_cast)
+#define CXX20_CONSTEXPR constexpr
+#else
+#define CXX20_CONSTEXPR
+#endif
+#endif
+
 namespace fast_float {
 
 // Compares two ASCII strings in a case insensitive manner.
-inline bool fastfloat_strncasecmp(const char *input1, const char *input2,
+CXX20_CONSTEXPR inline bool fastfloat_strncasecmp(const char *input1, const char *input2,
                                   size_t length) {
   char running_diff{0};
   for (size_t i = 0; i < length; i++) {
@@ -103,7 +112,7 @@ struct value128 {
 };
 
 /* result might be undefined when input_num is zero */
-fastfloat_really_inline int leading_zeroes(uint64_t input_num) {
+CXX20_CONSTEXPR fastfloat_really_inline int leading_zeroes(uint64_t input_num) {
   assert(input_num > 0);
 #ifdef FASTFLOAT_VISUAL_STUDIO
   #if defined(_M_X64) || defined(_M_ARM64)
@@ -130,13 +139,13 @@ fastfloat_really_inline int leading_zeroes(uint64_t input_num) {
 #ifdef FASTFLOAT_32BIT
 
 // slow emulation routine for 32-bit
-fastfloat_really_inline uint64_t emulu(uint32_t x, uint32_t y) {
+CXX20_CONSTEXPR fastfloat_really_inline uint64_t emulu(uint32_t x, uint32_t y) {
     return x * (uint64_t)y;
 }
 
 // slow emulation routine for 32-bit
 #if !defined(__MINGW64__)
-fastfloat_really_inline uint64_t _umul128(uint64_t ab, uint64_t cd,
+CXX20_CONSTEXPR fastfloat_really_inline uint64_t _umul128(uint64_t ab, uint64_t cd,
                                           uint64_t *hi) {
   uint64_t ad = emulu((uint32_t)(ab >> 32), (uint32_t)cd);
   uint64_t bd = emulu((uint32_t)ab, (uint32_t)cd);

include/fast_float/parse_number.h

@@ -20,7 +20,7 @@ namespace detail {
  * strings a null-free and fixed.
  **/
 template <typename T>
-from_chars_result parse_infnan(const char *first, const char *last, T &value)  noexcept  {
+CXX20_CONSTEXPR from_chars_result parse_infnan(const char *first, const char *last, T &value)  noexcept  {
   from_chars_result answer;
   answer.ptr = first;
   answer.ec = std::errc(); // be optimistic
@@ -61,7 +61,7 @@ from_chars_result parse_infnan(const char *first, const char *last, T &value)  n
 }
 
 template<typename T>
-fastfloat_really_inline void to_float(bool negative, adjusted_mantissa am, T &value) {
+CXX20_CONSTEXPR fastfloat_really_inline void to_float(bool negative, adjusted_mantissa am, T &value) {
   uint64_t word = am.mantissa;
   word |= uint64_t(am.power2) << binary_format<T>::mantissa_explicit_bits();
   word = negative
@@ -83,13 +83,13 @@ fastfloat_really_inline void to_float(bool negative, adjusted_mantissa am, T &va
 
 
 template<typename T>
-from_chars_result from_chars(const char *first, const char *last,
+CXX20_CONSTEXPR from_chars_result from_chars(const char *first, const char *last,
                              T &value, chars_format fmt /*= chars_format::general*/)  noexcept  {
   return from_chars_advanced(first, last, value, parse_options{fmt});
 }
 
 template<typename T>
-from_chars_result from_chars_advanced(const char *first, const char *last,
+CXX20_CONSTEXPR from_chars_result from_chars_advanced(const char *first, const char *last,
                                       T &value, parse_options options)  noexcept  {
 
   static_assert (std::is_same<T, double>::value || std::is_same<T, float>::value, "only float and double are supported");

include/fast_float/simple_decimal_conversion.h

@@ -22,17 +22,17 @@ namespace fast_float {
 namespace detail {
 
 // remove all final zeroes
-inline void trim(decimal &h) {
+CXX20_CONSTEXPR inline void trim(decimal &h) {
   while ((h.num_digits > 0) && (h.digits[h.num_digits - 1] == 0)) {
     h.num_digits--;
   }
 }
 
 
 
-inline uint32_t number_of_digits_decimal_left_shift(const decimal &h, uint32_t shift) {
+CXX20_CONSTEXPR inline uint32_t number_of_digits_decimal_left_shift(const decimal &h, uint32_t shift) {
   shift &= 63;
-  const static uint16_t number_of_digits_decimal_left_shift_table[65] = {
+  constexpr uint16_t number_of_digits_decimal_left_shift_table[65] = {
     0x0000, 0x0800, 0x0801, 0x0803, 0x1006, 0x1009, 0x100D, 0x1812, 0x1817,
     0x181D, 0x2024, 0x202B, 0x2033, 0x203C, 0x2846, 0x2850, 0x285B, 0x3067,
     0x3073, 0x3080, 0x388E, 0x389C, 0x38AB, 0x38BB, 0x40CC, 0x40DD, 0x40EF,
@@ -47,7 +47,7 @@ inline uint32_t number_of_digits_decimal_left_shift(const decimal &h, uint32_t s
   uint32_t num_new_digits = x_a >> 11;
   uint32_t pow5_a = 0x7FF & x_a;
   uint32_t pow5_b = 0x7FF & x_b;
-  const static uint8_t
+  constexpr uint8_t
     number_of_digits_decimal_left_shift_table_powers_of_5[0x051C] = {
         5, 2, 5, 1, 2, 5, 6, 2, 5, 3, 1, 2, 5, 1, 5, 6, 2, 5, 7, 8, 1, 2, 5, 3,
         9, 0, 6, 2, 5, 1, 9, 5, 3, 1, 2, 5, 9, 7, 6, 5, 6, 2, 5, 4, 8, 8, 2, 8,
@@ -123,7 +123,7 @@ inline uint32_t number_of_digits_decimal_left_shift(const decimal &h, uint32_t s
   return num_new_digits;
 }
 
-inline uint64_t round(decimal &h) {
+CXX20_CONSTEXPR inline uint64_t round(decimal &h) {
   if ((h.num_digits == 0) || (h.decimal_point < 0)) {
     return 0;
   } else if (h.decimal_point > 18) {
@@ -150,7 +150,7 @@ inline uint64_t round(decimal &h) {
 }
 
 // computes h * 2^-shift
-inline void decimal_left_shift(decimal &h, uint32_t shift) {
+CXX20_CONSTEXPR inline void decimal_left_shift(decimal &h, uint32_t shift) {
   if (h.num_digits == 0) {
     return;
   }
@@ -192,7 +192,7 @@ inline void decimal_left_shift(decimal &h, uint32_t shift) {
 }
 
 // computes h * 2^shift
-inline void decimal_right_shift(decimal &h, uint32_t shift) {
+CXX20_CONSTEXPR inline void decimal_right_shift(decimal &h, uint32_t shift) {
   uint32_t read_index = 0;
   uint32_t write_index = 0;
 
@@ -241,7 +241,7 @@ inline void decimal_right_shift(decimal &h, uint32_t shift) {
 } // namespace detail
 
 template <typename binary>
-adjusted_mantissa compute_float(decimal &d) {
+CXX20_CONSTEXPR adjusted_mantissa compute_float(decimal &d) {
   adjusted_mantissa answer;
   if (d.num_digits == 0) {
     // should be zero
@@ -271,9 +271,9 @@ adjusted_mantissa compute_float(decimal &d) {
     answer.mantissa = 0;
     return answer;
   }
-  static const uint32_t max_shift = 60;
-  static const uint32_t num_powers = 19;
-  static const uint8_t decimal_powers[19] = {
+  constexpr uint32_t max_shift = 60;
+  constexpr uint32_t num_powers = 19;
+  constexpr uint8_t decimal_powers[19] = {
       0,  3,  6,  9,  13, 16, 19, 23, 26, 29, //
       33, 36, 39, 43, 46, 49, 53, 56, 59,     //
   };
@@ -351,7 +351,7 @@ adjusted_mantissa compute_float(decimal &d) {
 }
 
 template <typename binary>
-adjusted_mantissa parse_long_mantissa(const char *first, const char* last, parse_options options) {
+CXX20_CONSTEXPR adjusted_mantissa parse_long_mantissa(const char *first, const char* last, parse_options options) {
     decimal d = parse_decimal(first, last, options);
     return compute_float<binary>(d);
 }