diff --git a/include/fast_float/float_common.h b/include/fast_float/float_common.h index 1e927dc..77e3e59 100644 --- a/include/fast_float/float_common.h +++ b/include/fast_float/float_common.h @@ -7,6 +7,25 @@ #include #include +#ifdef __has_include +#if __has_include() +#include +#endif +#endif + +#if __cpp_lib_bit_cast >= 201806L +#include +#define FASTFLOAT_HAS_BIT_CAST 1 +#else +#define FASTFLOAT_HAS_BIT_CAST 0 +#endif + +#if __cpp_lib_is_constant_evaluated >= 201811L +#define FASTFLOAT_HAS_IS_CONSTANT_EVALUATED 1 +#else +#define FASTFLOAT_HAS_IS_CONSTANT_EVALUATED 0 +#endif + #if (defined(__x86_64) || defined(__x86_64__) || defined(_M_X64) \ || defined(__amd64) || defined(__aarch64__) || defined(_M_ARM64) \ || defined(__MINGW64__) \ @@ -98,8 +117,25 @@ #define FASTFLOAT_CONSTEXPR14 #endif +// Testing for relevant C++20 constexpr library features +#if FASTFLOAT_HAS_IS_CONSTANT_EVALUATED \ + && FASTFLOAT_HAS_BIT_CAST \ + && __cpp_lib_constexpr_algorithms >= 201806L /*For std::copy and std::fill*/ +#define FASTFLOAT_CONSTEXPR20 constexpr +#else +#define FASTFLOAT_CONSTEXPR20 +#endif + namespace fast_float { +fastfloat_really_inline constexpr bool cpp20_and_in_constexpr() { +#if FASTFLOAT_HAS_IS_CONSTANT_EVALUATED + return std::is_constant_evaluated(); +#else + return false; +#endif +} + // Compares two ASCII strings in a case insensitive manner. inline FASTFLOAT_CONSTEXPR14 bool fastfloat_strncasecmp(const char *input1, const char *input2, size_t length) { @@ -139,9 +175,27 @@ struct value128 { constexpr value128() : low(0), high(0) {} }; +/* Helper C++11 constexpr generic implementation of leading_zeroes */ +fastfloat_really_inline constexpr +int leading_zeroes_generic(uint64_t input_num, int last_bit = 0) { + return ( + ((input_num & uint64_t(0xffffffff00000000)) && (input_num >>= 32, last_bit |= 32)), + ((input_num & uint64_t( 0xffff0000)) && (input_num >>= 16, last_bit |= 16)), + ((input_num & uint64_t( 0xff00)) && (input_num >>= 8, last_bit |= 8)), + ((input_num & uint64_t( 0xf0)) && (input_num >>= 4, last_bit |= 4)), + ((input_num & uint64_t( 0xc)) && (input_num >>= 2, last_bit |= 2)), + ((input_num & uint64_t( 0x2)) && (input_num >>= 1, last_bit |= 1)), + 63 - last_bit + ); +} + /* result might be undefined when input_num is zero */ -fastfloat_really_inline int leading_zeroes(uint64_t input_num) { +fastfloat_really_inline FASTFLOAT_CONSTEXPR20 +int leading_zeroes(uint64_t input_num) { assert(input_num > 0); + if (cpp20_and_in_constexpr()) { + return leading_zeroes_generic(input_num); + } #ifdef FASTFLOAT_VISUAL_STUDIO #if defined(_M_X64) || defined(_M_ARM64) unsigned long leading_zero = 0; @@ -150,31 +204,20 @@ fastfloat_really_inline int leading_zeroes(uint64_t input_num) { _BitScanReverse64(&leading_zero, input_num); return (int)(63 - leading_zero); #else - int last_bit = 0; - if(input_num & uint64_t(0xffffffff00000000)) input_num >>= 32, last_bit |= 32; - if(input_num & uint64_t( 0xffff0000)) input_num >>= 16, last_bit |= 16; - if(input_num & uint64_t( 0xff00)) input_num >>= 8, last_bit |= 8; - if(input_num & uint64_t( 0xf0)) input_num >>= 4, last_bit |= 4; - if(input_num & uint64_t( 0xc)) input_num >>= 2, last_bit |= 2; - if(input_num & uint64_t( 0x2)) input_num >>= 1, last_bit |= 1; - return 63 - last_bit; + return leading_zeroes_generic(input_num); #endif #else return __builtin_clzll(input_num); #endif } -#ifdef FASTFLOAT_32BIT - // slow emulation routine for 32-bit fastfloat_really_inline constexpr 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 constexpr uint64_t _umul128(uint64_t ab, uint64_t cd, - uint64_t *hi) { +fastfloat_really_inline FASTFLOAT_CONSTEXPR14 +uint64_t umul128_generic(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); uint64_t adbc = ad + emulu((uint32_t)ab, (uint32_t)(cd >> 32)); @@ -184,14 +227,28 @@ fastfloat_really_inline constexpr uint64_t _umul128(uint64_t ab, uint64_t cd, (adbc_carry << 32) + !!(lo < bd); return lo; } + +#ifdef FASTFLOAT_32BIT + +// slow emulation routine for 32-bit +#if !defined(__MINGW64__) +fastfloat_really_inline FASTFLOAT_CONSTEXPR14 +uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) { + return umul128_generic(ab, cd, hi); +} #endif // !__MINGW64__ #endif // FASTFLOAT_32BIT // compute 64-bit a*b -fastfloat_really_inline value128 full_multiplication(uint64_t a, - uint64_t b) { +fastfloat_really_inline FASTFLOAT_CONSTEXPR20 +value128 full_multiplication(uint64_t a, uint64_t b) { + if (cpp20_and_in_constexpr()) { + value128 answer; + answer.low = umul128_generic(a, b, &answer.high); + return answer; + } value128 answer; #if defined(_M_ARM64) && !defined(__MINGW32__) // ARM64 has native support for 64-bit multiplications, no need to emulate @@ -205,7 +262,7 @@ fastfloat_really_inline value128 full_multiplication(uint64_t a, answer.low = uint64_t(r); answer.high = uint64_t(r >> 64); #else - #error Not implemented + answer.low = umul128_generic(a, b, &answer.high); #endif return answer; } @@ -442,12 +499,17 @@ template <> inline constexpr binary_format::equiv_uint } template -fastfloat_really_inline void to_float(bool negative, adjusted_mantissa am, T &value) { +fastfloat_really_inline FASTFLOAT_CONSTEXPR20 +void to_float(bool negative, adjusted_mantissa am, T &value) { using uint = typename binary_format::equiv_uint; uint word = (uint)am.mantissa; word |= uint(am.power2) << binary_format::mantissa_explicit_bits(); word |= uint(negative) << binary_format::sign_index(); +#if FASTFLOAT_HAS_BIT_CAST + value = std::bit_cast(word); +#else ::memcpy(&value, &word, sizeof(T)); +#endif } #if FASTFLOAT_SKIP_WHITE_SPACE // disabled by default