Merge pull request #182 from leni536/constexpr-from_chars

Constexpr from_chars

Author: lemire

.github/workflows/vs17-cxx20.yml

@@ -18,8 +18,12 @@ jobs:
       - name: checkout
         uses: actions/checkout@v3
       - name: configure
-        run: |
-          cmake -S . -B build -G "${{matrix.gen}}" -A ${{matrix.arch}}  -DCMAKE_CXX_STANDARD=20 -DFASTFLOAT_TEST=ON -DCMAKE_INSTALL_PREFIX:PATH=destination ..
+        run: >-
+          cmake -S . -B build -G "${{matrix.gen}}" -A ${{matrix.arch}}
+          -DCMAKE_CXX_STANDARD=20
+          -DFASTFLOAT_TEST=ON
+          -DFASTFLOAT_CONSTEXPR_TESTS=ON
+          -DCMAKE_INSTALL_PREFIX:PATH=destination
       - name: build
         run: |
           cmake --build build --verbose --config ${{matrix.cfg}} --parallel

include/fast_float/ascii_number.h

@@ -27,7 +27,16 @@ fastfloat_really_inline constexpr uint64_t byteswap(uint64_t val) {
     | (val & 0x00000000000000FF) << 56;
 }
 
-fastfloat_really_inline uint64_t read_u64(const char *chars) {
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20
+uint64_t read_u64(const char *chars) {
+  if (cpp20_and_in_constexpr()) {
+    uint64_t val = 0;
+    for(int i = 0; i < 8; ++i) {
+      val |= uint64_t(*chars) << (i*8);
+      ++chars;
+    }
+    return val;
+  }
   uint64_t val;
   ::memcpy(&val, chars, sizeof(uint64_t));
 #if FASTFLOAT_IS_BIG_ENDIAN == 1
@@ -37,7 +46,16 @@ fastfloat_really_inline uint64_t read_u64(const char *chars) {
   return val;
 }
 
-fastfloat_really_inline void write_u64(uint8_t *chars, uint64_t val) {
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20
+void write_u64(uint8_t *chars, uint64_t val) {
+  if (cpp20_and_in_constexpr()) {
+    for(int i = 0; i < 8; ++i) {
+      *chars = uint8_t(val);
+      val >>= 8;
+      ++chars;
+    }
+    return;
+  }
 #if FASTFLOAT_IS_BIG_ENDIAN == 1
   // Need to read as-if the number was in little-endian order.
   val = byteswap(val);
@@ -57,7 +75,8 @@ 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  {
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20
+uint32_t parse_eight_digits_unrolled(const char *chars)  noexcept  {
   return parse_eight_digits_unrolled(read_u64(chars));
 }
 
@@ -67,7 +86,8 @@ fastfloat_really_inline constexpr bool is_made_of_eight_digits_fast(uint64_t val
      0x8080808080808080));
 }
 
-fastfloat_really_inline bool is_made_of_eight_digits_fast(const char *chars)  noexcept  {
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20
+bool is_made_of_eight_digits_fast(const char *chars)  noexcept  {
   return is_made_of_eight_digits_fast(read_u64(chars));
 }
 
@@ -87,7 +107,7 @@ struct parsed_number_string {
 
 // Assuming that you use no more than 19 digits, this will
 // parse an ASCII string.
-fastfloat_really_inline
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20
 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;

include/fast_float/bigint.h

@@ -514,7 +514,7 @@ struct bigint : pow5_tables<> {
       // move limbs
       limb* dst = vec.data + n;
       const limb* src = vec.data;
-      ::memmove(dst, src, sizeof(limb) * vec.len());
+      std::copy_backward(src, src + vec.len(), dst + vec.len());
       // fill in empty limbs
       limb* first = vec.data;
       limb* last = first + n;
@@ -594,7 +594,12 @@ struct bigint : pow5_tables<> {
       exp -= small_step;
     }
     if (exp != 0) {
-      FASTFLOAT_TRY(small_mul(vec, limb(small_power_of_5[exp])));
+      // Work around clang bug https://godbolt.org/z/zedh7rrhc
+      // This is similar to https://github.com/llvm/llvm-project/issues/47746,
+      // except the workaround described there don't work here
+      FASTFLOAT_TRY(
+        small_mul(vec, limb(((void)small_power_of_5[0], small_power_of_5[exp])))
+      );
     }
 
     return true;

include/fast_float/constexpr_feature_detect.h

@@ -0,0 +1,40 @@
+#ifndef FASTFLOAT_CONSTEXPR_FEATURE_DETECT_H
+#define FASTFLOAT_CONSTEXPR_FEATURE_DETECT_H
+
+#ifdef __has_include
+#if __has_include(<version>)
+#include <version>
+#endif
+#endif
+
+// Testing for https://wg21.link/N3652, adopted in C++14
+#if __cpp_constexpr >= 201304
+#define FASTFLOAT_CONSTEXPR14 constexpr
+#else
+#define FASTFLOAT_CONSTEXPR14
+#endif
+
+#if __cpp_lib_bit_cast >= 201806L
+#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
+
+// 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
+#define FASTFLOAT_IS_CONSTEXPR 1
+#else
+#define FASTFLOAT_CONSTEXPR20
+#define FASTFLOAT_IS_CONSTEXPR 0
+#endif
+
+#endif // FASTFLOAT_CONSTEXPR_FEATURE_DETECT_H

include/fast_float/decimal_to_binary.h

@@ -17,7 +17,7 @@ namespace fast_float {
 // low part corresponding to the least significant bits.
 //
 template <int bit_precision>
-fastfloat_really_inline
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20
 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
@@ -76,7 +76,7 @@ adjusted_mantissa compute_error_scaled(int64_t q, uint64_t w, int lz) noexcept
 // w * 10 ** q, without rounding the representation up.
 // the power2 in the exponent will be adjusted by invalid_am_bias.
 template <typename binary>
-fastfloat_really_inline
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20
 adjusted_mantissa compute_error(int64_t q, uint64_t w)  noexcept  {
   int lz = leading_zeroes(w);
   w <<= lz;
@@ -90,7 +90,7 @@ adjusted_mantissa compute_error(int64_t q, uint64_t w)  noexcept  {
 // return an adjusted_mantissa with a negative power of 2: the caller should recompute
 // in such cases.
 template <typename binary>
-fastfloat_really_inline
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20
 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/digit_comparison.h

@@ -44,7 +44,8 @@ int32_t scientific_exponent(parsed_number_string& num) noexcept {
 
 // this converts a native floating-point number to an extended-precision float.
 template <typename T>
-fastfloat_really_inline adjusted_mantissa to_extended(T value) noexcept {
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20
+adjusted_mantissa to_extended(T value) noexcept {
   using equiv_uint = typename binary_format<T>::equiv_uint;
   constexpr equiv_uint exponent_mask = binary_format<T>::exponent_mask();
   constexpr equiv_uint mantissa_mask = binary_format<T>::mantissa_mask();
@@ -53,7 +54,11 @@ fastfloat_really_inline adjusted_mantissa to_extended(T value) noexcept {
   adjusted_mantissa am;
   int32_t bias = binary_format<T>::mantissa_explicit_bits() - binary_format<T>::minimum_exponent();
   equiv_uint bits;
+#if FASTFLOAT_HAS_BIT_CAST
+  bits = std::bit_cast<equiv_uint>(value);
+#else
   ::memcpy(&bits, &value, sizeof(T));
+#endif
   if ((bits & exponent_mask) == 0) {
     // denormal
     am.power2 = 1 - bias;
@@ -72,7 +77,8 @@ fastfloat_really_inline adjusted_mantissa to_extended(T value) noexcept {
 // we are given a native float that represents b, so we need to adjust it
 // halfway between b and b+u.
 template <typename T>
-fastfloat_really_inline adjusted_mantissa to_extended_halfway(T value) noexcept {
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20
+adjusted_mantissa to_extended_halfway(T value) noexcept {
   adjusted_mantissa am = to_extended(value);
   am.mantissa <<= 1;
   am.mantissa += 1;
@@ -148,9 +154,10 @@ void round_down(adjusted_mantissa& am, int32_t shift) noexcept {
   am.power2 += shift;
 }
 
-fastfloat_really_inline void skip_zeros(const char*& first, const char* last) noexcept {
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20
+void skip_zeros(const char*& first, const char* last) noexcept {
   uint64_t val;
-  while (std::distance(first, last) >= 8) {
+  while (!cpp20_and_in_constexpr() && std::distance(first, last) >= 8) {
     ::memcpy(&val, first, sizeof(uint64_t));
     if (val != 0x3030303030303030) {
       break;
@@ -167,10 +174,11 @@ fastfloat_really_inline void skip_zeros(const char*& first, const char* last) no
 
 // determine if any non-zero digits were truncated.
 // all characters must be valid digits.
-fastfloat_really_inline bool is_truncated(const char* first, const char* last) noexcept {
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20
+bool is_truncated(const char* first, const char* last) noexcept {
   // do 8-bit optimizations, can just compare to 8 literal 0s.
   uint64_t val;
-  while (std::distance(first, last) >= 8) {
+  while (!cpp20_and_in_constexpr() && std::distance(first, last) >= 8) {
     ::memcpy(&val, first, sizeof(uint64_t));
     if (val != 0x3030303030303030) {
       return true;
@@ -186,11 +194,12 @@ fastfloat_really_inline bool is_truncated(const char* first, const char* last) n
   return false;
 }
 
-fastfloat_really_inline bool is_truncated(byte_span s) noexcept {
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20
+bool is_truncated(byte_span s) noexcept {
   return is_truncated(s.ptr, s.ptr + s.len());
 }
 
-fastfloat_really_inline
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20
 void parse_eight_digits(const char*& p, limb& value, size_t& counter, size_t& count) noexcept {
   value = value * 100000000 + parse_eight_digits_unrolled(p);
   p += 8;
@@ -206,21 +215,23 @@ void parse_one_digit(const char*& p, limb& value, size_t& counter, size_t& count
   count++;
 }
 
-fastfloat_really_inline
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20
 void add_native(bigint& big, limb power, limb value) noexcept {
   big.mul(power);
   big.add(value);
 }
 
-fastfloat_really_inline void round_up_bigint(bigint& big, size_t& count) noexcept {
+fastfloat_really_inline FASTFLOAT_CONSTEXPR20
+void round_up_bigint(bigint& big, size_t& count) noexcept {
   // need to round-up the digits, but need to avoid rounding
   // ....9999 to ...10000, which could cause a false halfway point.
   add_native(big, 10, 1);
   count++;
 }
 
 // parse the significant digits into a big integer
-inline void parse_mantissa(bigint& result, parsed_number_string& num, size_t max_digits, size_t& digits) noexcept {
+inline FASTFLOAT_CONSTEXPR20
+void parse_mantissa(bigint& result, parsed_number_string& num, size_t max_digits, size_t& digits) noexcept {
   // try to minimize the number of big integer and scalar multiplication.
   // therefore, try to parse 8 digits at a time, and multiply by the largest
   // scalar value (9 or 19 digits) for each step.
@@ -300,7 +311,8 @@ inline void parse_mantissa(bigint& result, parsed_number_string& num, size_t max
 }
 
 template <typename T>
-inline adjusted_mantissa positive_digit_comp(bigint& bigmant, int32_t exponent) noexcept {
+inline FASTFLOAT_CONSTEXPR20
+adjusted_mantissa positive_digit_comp(bigint& bigmant, int32_t exponent) noexcept {
   FASTFLOAT_ASSERT(bigmant.pow10(uint32_t(exponent)));
   adjusted_mantissa answer;
   bool truncated;
@@ -323,7 +335,8 @@ inline adjusted_mantissa positive_digit_comp(bigint& bigmant, int32_t exponent)
 // we then need to scale by `2^(f- e)`, and then the two significant digits
 // are of the same magnitude.
 template <typename T>
-inline adjusted_mantissa negative_digit_comp(bigint& bigmant, adjusted_mantissa am, int32_t exponent) noexcept {
+inline FASTFLOAT_CONSTEXPR20
+adjusted_mantissa negative_digit_comp(bigint& bigmant, adjusted_mantissa am, int32_t exponent) noexcept {
   bigint& real_digits = bigmant;
   int32_t real_exp = exponent;
 
@@ -383,7 +396,8 @@ inline adjusted_mantissa negative_digit_comp(bigint& bigmant, adjusted_mantissa
 // the actual digits. we then compare the big integer representations
 // of both, and use that to direct rounding.
 template <typename T>
-inline adjusted_mantissa digit_comp(parsed_number_string& num, adjusted_mantissa am) noexcept {
+inline FASTFLOAT_CONSTEXPR20
+adjusted_mantissa digit_comp(parsed_number_string& num, adjusted_mantissa am) noexcept {
   // remove the invalid exponent bias
   am.power2 -= invalid_am_bias;
 

include/fast_float/fast_float.h

@@ -3,6 +3,8 @@
 
 #include <system_error>
 
+#include "constexpr_feature_detect.h"
+
 namespace fast_float {
 enum chars_format {
     scientific = 1<<0,
@@ -48,13 +50,15 @@ struct parse_options {
  * The default is  `fast_float::chars_format::general` which allows both `fixed` and `scientific`.
  */
 template<typename T>
+FASTFLOAT_CONSTEXPR20
 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>
+FASTFLOAT_CONSTEXPR20
 from_chars_result from_chars_advanced(const char *first, const char *last,
                                       T &value, parse_options options)  noexcept;
 

include/fast_float/float_common.h

@@ -7,23 +7,8 @@
 #include <cstring>
 #include <type_traits>
 
-#ifdef __has_include
-#if __has_include(<version>)
-#include <version>
-#endif
-#endif
-
-#if __cpp_lib_bit_cast >= 201806L
+#if FASTFLOAT_HAS_BIT_CAST
 #include <bit>
-#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)   \
@@ -110,22 +95,6 @@
 // rust style `try!()` macro, or `?` operator
 #define FASTFLOAT_TRY(x) { if (!(x)) return false; }
 
-// Testing for https://wg21.link/N3652, adopted in C++14
-#if __cpp_constexpr >= 201304
-#define FASTFLOAT_CONSTEXPR14 constexpr
-#else
-#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() {