Move the float conversion chain out of the lexer

lexer::convert_number() converted float tokens with std::from_chars (when
available), Clinger's fast path, and the locale-aware strtod fallback, all
as lexer members. They are now free functions in number_parse.hpp:

- convert_float_fast(): std::from_chars, then Clinger's fast path, skipped
  when the mantissa has too many significant digits
- convert_float_locale_aware(): strtof/strtod/strtold with the decimal point
  of the current locale, retried when the locale changed (#5198)

so that other code converting JSON number tokens gets the same values. No
change in behavior; the lexer no longer includes <clocale> and <cstdlib>.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

docs/mkdocs/docs/features/binary_formats/bjdata.md

@@ -132,14 +132,8 @@ The library uses the following mapping from JSON values types to BJData types ac
       parsed back as a regular array,
     - every entry of `"_ArraySize_"` is a positive integer, and their product is representable as a `std::size_t`,
     - `"_ArrayData_"` is an array holding exactly that many elements, and
-    - every element of `"_ArrayData_"` is a number of the kind named by `"_ArrayType_"`: for the integer types, a
-      value that fits the named width; for `double`, any value; for `single`, a value that survives narrowing to
-      `float` and back without change (for instance, `0.1` does not, since it is not exactly representable as
-      `float`).
-
-    An annotated object is always read back with its keys in the order shown above, `"_ArrayType_"`, `"_ArraySize_"`,
-    `"_ArrayData_"`, regardless of the order the ND-array's header stores them in on the wire. This matters for
-    `ordered_json`, whose comparison takes key order into account.
+    - every element of `"_ArrayData_"` is a number of the kind named by `"_ArrayType_"` (a floating-point number for
+      `single` and `double`, an integer otherwise).
 
     The current version of this library does not yet support automatic detection of and conversion from a nested JSON
     array input to a BJData ND-array.

include/nlohmann/detail/input/binary_reader.hpp

@@ -2728,15 +2728,10 @@ class binary_reader
                                is_ndarray can only return `true` when its initial value
                                is `false`
     @param[in] prefix  type marker if already read, otherwise set to 0
-    @param[in] ndarray_dtype  the element type marker of the enclosing bjdata ndarray if
-                              already known (it precedes the dimension vector read here),
-                              otherwise 0; used to emit the "_ArrayType_" annotation key
-                              before "_ArraySize_" if a dimension vector turns out to
-                              describe an ndarray
 
     @return whether size determination completed
     */
-    bool get_ubjson_size_value(std::size_t& result, bool& is_ndarray, char_int_type prefix = 0, char_int_type ndarray_dtype = 0)
+    bool get_ubjson_size_value(std::size_t& result, bool& is_ndarray, char_int_type prefix = 0)
     {
         if (prefix == 0)
         {
@@ -2906,37 +2901,8 @@ class binary_reader
                         }
                     }
 
-                    if (JSON_HEDLEY_UNLIKELY(!sax->start_object(3)))
-                    {
-                        return false;
-                    }
-
-                    // the element type precedes the dimension vector (see get_ubjson_size_type)
-                    // and is passed down as ndarray_dtype; emit it here so the annotation keys
-                    // follow the documented _ArrayType_, _ArraySize_, _ArrayData_ order
-                    if (ndarray_dtype != 0)
-                    {
-                        auto it = std::lower_bound(bjd_types_map.begin(), bjd_types_map.end(), ndarray_dtype, [](const bjd_type & p, char_int_type t)
-                        {
-                            return p.first < t;
-                        });
-                        if (JSON_HEDLEY_UNLIKELY(it == bjd_types_map.end() || it->first != ndarray_dtype))
-                        {
-                            auto last_token = get_token_string();
-                            return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
-                                                    exception_message(input_format, "invalid byte: 0x" + last_token, "type"), nullptr));
-                        }
-
-                        string_t type_key = "_ArrayType_";
-                        string_t type = it->second; // sax->string() takes a reference
-                        if (JSON_HEDLEY_UNLIKELY(!sax->key(type_key) || !sax->string(type)))
-                        {
-                            return false;
-                        }
-                    }
-
                     string_t key = "_ArraySize_";
-                    if (JSON_HEDLEY_UNLIKELY(!sax->key(key) || !sax->start_array(dim.size())))
+                    if (JSON_HEDLEY_UNLIKELY(!sax->start_object(3) || !sax->key(key) || !sax->start_array(dim.size())))
                     {
                         return false;
                     }
@@ -3037,7 +3003,7 @@ class binary_reader
                                         exception_message(input_format, concat("expected '#' after type information; last byte: 0x", last_token), "size"), nullptr));
             }
 
-            const bool is_error = get_ubjson_size_value(result.first, is_ndarray, 0, result.second);
+            const bool is_error = get_ubjson_size_value(result.first, is_ndarray);
             // an ndarray was read here only if the flag flipped; when it was
             // seeded true, get_ubjson_size_value() already rejected the nested
             // dimension vector
@@ -3273,17 +3239,30 @@ class binary_reader
         if (input_format == input_format_t::bjdata && size_and_type.first != npos && (size_and_type.second & (1 << 8)) != 0)
         {
             size_and_type.second &= ~(static_cast<char_int_type>(1) << 8);  // use bit 8 to indicate ndarray, here we remove the bit to restore the type marker
+            auto it = std::lower_bound(bjd_types_map.begin(), bjd_types_map.end(), size_and_type.second, [](const bjd_type & p, char_int_type t)
+            {
+                return p.first < t;
+            });
+            string_t key = "_ArrayType_";
+            if (JSON_HEDLEY_UNLIKELY(it == bjd_types_map.end() || it->first != size_and_type.second))
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
+                                        exception_message(input_format, "invalid byte: 0x" + last_token, "type"), nullptr));
+            }
+
+            string_t type = it->second; // sax->string() takes a reference
+            if (JSON_HEDLEY_UNLIKELY(!sax->key(key) || !sax->string(type)))
+            {
+                return false;
+            }
 
-            // the "_ArrayType_" and "_ArraySize_" annotation keys were already emitted by
-            // get_ubjson_size_value() (the type marker is known before the dimension vector
-            // that determines size_and_type.first is read, so it is emitted first there to
-            // match the documented _ArrayType_, _ArraySize_, _ArrayData_ key order)
             if (size_and_type.second == 'C' || size_and_type.second == 'B')
             {
                 size_and_type.second = 'U';
             }
 
-            string_t key = "_ArrayData_";
+            key = "_ArrayData_";
             if (JSON_HEDLEY_UNLIKELY(!sax->key(key) || !sax->start_array(size_and_type.first) ))
             {
                 return false;

include/nlohmann/detail/input/lexer.hpp

@@ -9,10 +9,8 @@
 #pragma once
 
 #include <array> // array
-#include <clocale> // localeconv
 #include <cstddef> // size_t
 #include <cstdio> // snprintf
-#include <cstdlib> // strtof, strtod, strtold, strtoll, strtoull
 #include <initializer_list> // initializer_list
 #include <string> // char_traits, string
 #include <utility> // move
@@ -217,18 +215,6 @@ class lexer : public lexer_base<BasicJsonType>
     ~lexer() = default;
 
   private:
-    /////////////////////
-    // locales
-    /////////////////////
-
-    /// return the decimal point of the current locale
-    static char get_decimal_point() noexcept
-    {
-        const auto* loc = localeconv();
-        JSON_ASSERT(loc != nullptr);
-        return (loc->decimal_point == nullptr) ? '.' : *(loc->decimal_point);
-    }
-
     /////////////////////
     // scan functions
     /////////////////////
@@ -1036,24 +1022,6 @@ class lexer : public lexer_base<BasicJsonType>
         }
     }
 
-    JSON_HEDLEY_NON_NULL(2)
-    static void strtof(float& f, const char* str, char** endptr) noexcept
-    {
-        f = std::strtof(str, endptr);
-    }
-
-    JSON_HEDLEY_NON_NULL(2)
-    static void strtof(double& f, const char* str, char** endptr) noexcept
-    {
-        f = std::strtod(str, endptr);
-    }
-
-    JSON_HEDLEY_NON_NULL(2)
-    static void strtof(long double& f, const char* str, char** endptr) noexcept
-    {
-        f = std::strtold(str, endptr);
-    }
-
     /*!
     @brief scan a number literal
 
@@ -1093,7 +1061,7 @@ class lexer : public lexer_base<BasicJsonType>
     @note The scanner is independent of the current locale: token_buffer
           always holds `.`. Only the std::strtod fallback of convert_number()
           depends on the locale, and it looks up the decimal point right
-          before converting (see convert_float_locale_aware()).
+          before converting (see detail::convert_float_locale_aware()).
     */
     token_type scan_number()  // lgtm [cpp/use-of-goto] `goto` is used in this function to implement the number-parsing state machine described above. By design, any finite input will eventually reach the "done" state or return token_type::parse_error. In each intermediate state, 1 byte of the input is appended to the token_buffer vector, and only the already initialized variables token_buffer, number_type, and error_message are manipulated.
     {
@@ -1424,59 +1392,6 @@ scan_number_done:
         return token_type::uninitialized;
     }
 
-    /*!
-    @brief check whether Clinger's fast path can still succeed for this token
-
-    parse_float_fast() needs a significand below 2^53. A mantissa with 17 or
-    more significant digits is at least 10^16 and therefore always exceeds it,
-    so calling the fast path would walk the token one extra time only to
-    decline before strtod has to run anyway.
-
-    Significant digits are the mantissa's digits from the first nonzero one on;
-    the sign, the decimal point, leading zeros, and the exponent do not count.
-    The answer is derived from indices - the digits are not scanned again - so
-    this stays off the hot path of the number scanners.
-
-    @param[in] mantissa_end  offset just past the last mantissa byte in
-                             token_buffer
-    @return false if parse_float_fast() is guaranteed to decline
-    */
-    bool mantissa_fits_clinger(std::size_t mantissa_end) const
-    {
-        // 10^16 already exceeds 2^53, so 17 digits can never fit
-        constexpr std::size_t limit = 17;
-
-        const std::size_t neg = (!token_buffer.empty() && token_buffer[0] == '-') ? 1u : 0u;
-        const std::size_t has_dot = (decimal_point_position != std::string::npos) ? 1u : 0u;
-        // the JSON grammar restricts the integer part to "0" or [1-9][0-9]*, so
-        // a leading zero can only be a lone "0", which is not significant
-        const std::size_t lead_zero = (token_buffer[neg] == '0') ? 1u : 0u;
-        JSON_ASSERT(mantissa_end >= neg + has_dot + lead_zero);
-        std::size_t digits = mantissa_end - neg - has_dot - lead_zero;
-
-        if (JSON_HEDLEY_LIKELY(digits < limit))
-        {
-            return true;
-        }
-
-        // Only a number below 1 can carry further insignificant zeros, and only
-        // while the count stays at the limit does removing them change the
-        // answer - so this loop is skipped for all but a few tokens. The
-        // fraction is located through decimal_point_position rather than by
-        // searching '.'.
-        if (lead_zero != 0)
-        {
-            JSON_ASSERT(has_dot != 0); // an integer "0" cannot reach the limit
-            for (std::size_t i = decimal_point_position + 1;
-                    digits >= limit && i < mantissa_end && token_buffer[i] == '0'; ++i)
-            {
-                --digits;
-            }
-        }
-
-        return digits < limit;
-    }
-
     /*!
     @brief convert the number text in token_buffer to its value and token type
 
@@ -1490,7 +1405,7 @@ scan_number_done:
                              token_buffer (the index of 'e'/'E', or
                              token_buffer.size() when there is no exponent);
                              used to skip Clinger's fast path when it cannot
-                             possibly succeed - see mantissa_fits_clinger()
+                             possibly succeed - see detail::mantissa_fits_clinger()
     */
     token_type convert_number(token_type number_type, std::size_t mantissa_end)
     {
@@ -1563,77 +1478,15 @@ scan_number_done:
         // (Eisel-Lemire, locale-independent, correctly rounded) when available;
         // otherwise the exact Clinger fast path (double only); otherwise the
         // locale-aware strtof/strtod/strtold.
-        if (parse_float_from_chars(num_begin, num_end, value_float))
-        {
-            return token_type::value_float;
-        }
-        // Skipping a fast path that cannot succeed is lossless and saves a full
-        // extra pass over the token's bytes, which otherwise shows up on
-        // high-precision inputs such as canada.json
-        if (mantissa_fits_clinger(mantissa_end)
-                && parse_float_fast(num_begin, num_end, value_float))
+        if (convert_float_fast(num_begin, num_end, decimal_point_position, mantissa_end, value_float))
         {
             return token_type::value_float;
         }
 
-        convert_float_locale_aware();
+        convert_float_locale_aware(token_buffer, decimal_point_position, value_float);
         return token_type::value_float;
     }
 
-    /*!
-    @brief convert the float in token_buffer with strtof/strtod/strtold
-
-    These functions expect the decimal point of the *current* locale, so it is
-    looked up right before the conversion instead of once when the lexer is
-    constructed: a locale change in between (by a parser callback, a SAX
-    handler, or another thread) must not truncate the value (#5198). The
-    token has been validated before, so if the conversion stops early and the
-    decimal point changed in the meantime, the locale changed between the
-    lookup and the call, and the conversion is repeated with the new decimal
-    point. If the decimal point did not change, a retry cannot succeed: the
-    locale's decimal point is not a single character (e.g., the two-byte
-    U+066B of ar_EG.UTF-8 or fa_IR.UTF-8) and cannot be substituted in place.
-    The value strtod parsed up to that point is kept, as before this change.
-
-    Note that changing the locale in another thread *while* strtod runs is
-    undefined behavior of the C library, which this function cannot prevent.
-    */
-    void convert_float_locale_aware()
-    {
-        const bool has_dot = decimal_point_position != std::string::npos;
-        char decimal_point = get_decimal_point();
-        for (;;)
-        {
-            const bool substitute = has_dot && decimal_point != '.';
-            if (substitute)
-            {
-                token_buffer[decimal_point_position] = static_cast<typename string_t::value_type>(decimal_point);
-            }
-
-            char* endptr = nullptr; // NOLINT(misc-const-correctness,cppcoreguidelines-pro-type-vararg,hicpp-vararg)
-            strtof(value_float, token_buffer.data(), &endptr);
-
-            if (substitute)
-            {
-                // get_string() hands the token to the SAX interface with '.'
-                token_buffer[decimal_point_position] = '.';
-            }
-
-            if (JSON_HEDLEY_LIKELY(endptr == token_buffer.data() + token_buffer.size()))
-            {
-                return;
-            }
-
-            // retry only if the locale changed; otherwise, this would loop forever
-            const char current_decimal_point = get_decimal_point();
-            if (current_decimal_point == decimal_point)
-            {
-                return;
-            }
-            decimal_point = current_decimal_point;
-        }
-    }
-
     /*!
     @brief contiguous fast path for scanning a number
 

include/nlohmann/detail/input/number_parse.hpp

@@ -10,9 +10,12 @@
 
 #include <array> // array
 #include <cfloat> // FLT_EVAL_METHOD
+#include <clocale> // localeconv
 #include <cstddef> // size_t
 #include <cstdint> // int64_t, uint64_t
+#include <cstdlib> // strtof, strtod, strtold
 #include <limits> // numeric_limits
+#include <string> // string
 
 #include <nlohmann/detail/macro_scope.hpp>
 
@@ -29,8 +32,9 @@
 
 // This file contains the value-conversion helpers used by the lexer to turn an
 // already-validated number token into a value, without the locale/errno
-// overhead of std::strtoull/std::strtod. They are free functions so the lexer
-// stays focused on scanning; see lexer::convert_number().
+// overhead of std::strtoull/std::strtod where possible. They are free functions
+// so the lexer stays focused on scanning (see lexer::convert_number()) and so
+// that other parsers of JSON text can convert tokens exactly like it does.
 
 NLOHMANN_JSON_NAMESPACE_BEGIN
 namespace detail
@@ -293,5 +297,183 @@ bool parse_float_from_chars(const char* first, const char* last, FloatType& out)
 #endif
 }
 
+/*!
+@brief check whether Clinger's fast path can still succeed for a float token
+
+parse_float_fast() needs a significand below 2^53. A mantissa with 17 or
+more significant digits is at least 10^16 and therefore always exceeds it,
+so calling the fast path would walk the token one extra time only to
+decline before strtod has to run anyway.
+
+Significant digits are the mantissa's digits from the first nonzero one on;
+the sign, the decimal point, leading zeros, and the exponent do not count.
+The answer is derived from indices - the digits are not scanned again - so
+this stays off the hot path of the number scanners.
+
+@param[in] token                   the validated number token ('.' as decimal point)
+@param[in] decimal_point_position  index of the '.' in @a token, or
+                                   std::string::npos if there is none
+@param[in] mantissa_end            offset just past the last mantissa byte
+@return false if parse_float_fast() is guaranteed to decline
+*/
+inline bool mantissa_fits_clinger(const char* token, std::size_t decimal_point_position, std::size_t mantissa_end) noexcept
+{
+    // 10^16 already exceeds 2^53, so 17 digits can never fit
+    constexpr std::size_t limit = 17;
+
+    const std::size_t neg = (token[0] == '-') ? 1u : 0u;
+    const std::size_t has_dot = (decimal_point_position != std::string::npos) ? 1u : 0u;
+    // the JSON grammar restricts the integer part to "0" or [1-9][0-9]*, so
+    // a leading zero can only be a lone "0", which is not significant
+    const std::size_t lead_zero = (token[neg] == '0') ? 1u : 0u;
+    JSON_ASSERT(mantissa_end >= neg + has_dot + lead_zero);
+    std::size_t digits = mantissa_end - neg - has_dot - lead_zero;
+
+    if (JSON_HEDLEY_LIKELY(digits < limit))
+    {
+        return true;
+    }
+
+    // Only a number below 1 can carry further insignificant zeros, and only
+    // while the count stays at the limit does removing them change the
+    // answer - so this loop is skipped for all but a few tokens. The
+    // fraction is located through decimal_point_position rather than by
+    // searching '.'.
+    if (lead_zero != 0)
+    {
+        JSON_ASSERT(has_dot != 0); // an integer "0" cannot reach the limit
+        for (std::size_t i = decimal_point_position + 1;
+                digits >= limit && i < mantissa_end && token[i] == '0'; ++i)
+        {
+            --digits;
+        }
+    }
+
+    return digits < limit;
+}
+
+/*!
+@brief convert a validated float token without the C library, if possible
+
+Tries std::from_chars (when available) and then Clinger's exact fast path
+(double only), skipping the latter when it cannot succeed.
+
+@param[in]  first                   pointer to the first character of the token
+@param[in]  last                    pointer past the last character
+@param[in]  decimal_point_position  index of the '.' in the token, or
+                                    std::string::npos if there is none
+@param[in]  mantissa_end            offset just past the last mantissa byte (the
+                                    index of 'e'/'E', or the token length)
+@param[out] value                   the converted value on success
+@return true if the value was converted; false if convert_float_locale_aware()
+        must convert it
+*/
+template<typename FloatType>
+bool convert_float_fast(const char* first, const char* last, std::size_t decimal_point_position,
+                        std::size_t mantissa_end, FloatType& value) noexcept
+{
+    if (parse_float_from_chars(first, last, value))
+    {
+        return true;
+    }
+    // Skipping a fast path that cannot succeed is lossless and saves a full
+    // extra pass over the token's bytes, which otherwise shows up on
+    // high-precision inputs such as canada.json
+    return mantissa_fits_clinger(first, decimal_point_position, mantissa_end)
+           && parse_float_fast(first, last, value);
+}
+
+/// std::strtof, std::strtod, or std::strtold, chosen by the type of @a f
+JSON_HEDLEY_NON_NULL(2)
+inline void strtof_by_type(float& f, const char* str, char** endptr) noexcept
+{
+    f = std::strtof(str, endptr);
+}
+
+/// std::strtof, std::strtod, or std::strtold, chosen by the type of @a f
+JSON_HEDLEY_NON_NULL(2)
+inline void strtof_by_type(double& f, const char* str, char** endptr) noexcept
+{
+    f = std::strtod(str, endptr);
+}
+
+/// std::strtof, std::strtod, or std::strtold, chosen by the type of @a f
+JSON_HEDLEY_NON_NULL(2)
+inline void strtof_by_type(long double& f, const char* str, char** endptr) noexcept
+{
+    f = std::strtold(str, endptr);
+}
+
+/// return the decimal point of the current locale
+inline char get_decimal_point() noexcept
+{
+    const auto* loc = localeconv();
+    JSON_ASSERT(loc != nullptr);
+    return (loc->decimal_point == nullptr) ? '.' : *(loc->decimal_point);
+}
+
+/*!
+@brief convert a validated float token with strtof/strtod/strtold
+
+These functions expect the decimal point of the *current* locale, so it is
+looked up right before the conversion instead of once when the lexer is
+constructed: a locale change in between (by a parser callback, a SAX
+handler, or another thread) must not truncate the value (#5198). The
+token has been validated before, so if the conversion stops early and the
+decimal point changed in the meantime, the locale changed between the
+lookup and the call, and the conversion is repeated with the new decimal
+point. If the decimal point did not change, a retry cannot succeed: the
+locale's decimal point is not a single character (e.g., the two-byte
+U+066B of ar_EG.UTF-8 or fa_IR.UTF-8) and cannot be substituted in place.
+The value strtod parsed up to that point is kept, as before this change.
+
+Note that changing the locale in another thread *while* strtod runs is
+undefined behavior of the C library, which this function cannot prevent.
+
+@param[in,out] token                   the token with '.' as decimal point; its
+                                       decimal point is replaced during the
+                                       conversion and restored afterwards
+                                       (data() must be NUL-terminated)
+@param[in]     decimal_point_position  index of the '.' in @a token, or
+                                       std::string::npos if there is none
+@param[out]    value                   the converted value
+*/
+template<typename StringType, typename FloatType>
+void convert_float_locale_aware(StringType& token, std::size_t decimal_point_position, FloatType& value)
+{
+    const bool has_dot = decimal_point_position != std::string::npos;
+    char decimal_point = get_decimal_point();
+    for (;;)
+    {
+        const bool substitute = has_dot && decimal_point != '.';
+        if (substitute)
+        {
+            token[decimal_point_position] = static_cast<typename StringType::value_type>(decimal_point);
+        }
+
+        char* endptr = nullptr; // NOLINT(misc-const-correctness,cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+        strtof_by_type(value, token.data(), &endptr);
+
+        if (substitute)
+        {
+            // the caller hands the token on (e.g. to the SAX interface) with '.'
+            token[decimal_point_position] = '.';
+        }
+
+        if (JSON_HEDLEY_LIKELY(endptr == token.data() + token.size()))
+        {
+            return;
+        }
+
+        // retry only if the locale changed; otherwise, this would loop forever
+        const char current_decimal_point = get_decimal_point();
+        if (current_decimal_point == decimal_point)
+        {
+            return;
+        }
+        decimal_point = current_decimal_point;
+    }
+}
+
 }  // namespace detail
 NLOHMANN_JSON_NAMESPACE_END

include/nlohmann/detail/output/binary_writer.hpp

@@ -1991,21 +1991,9 @@ class binary_writer
                 case 'd':
                 {
                     const auto dval = el.template get<double>();
-#ifdef __GNUC__
-                    JSON_HEDLEY_DIAGNOSTIC_PUSH
-                    JSON_HEDLEY_PRAGMA(GCC diagnostic ignored "-Wfloat-equal")
-#endif
-                    // a value that would be rounded (rather than exactly represented) by the
-                    // narrowing to float is treated like an out-of-range integer element above;
-                    // this is the same criterion write_compact_float() uses for CBOR/MessagePack
-                    in_range = std::isnan(dval) ||
+                    in_range = !std::isfinite(dval) ||
                                (dval >= static_cast<double>(std::numeric_limits<float>::lowest()) &&
-                                dval <= static_cast<double>((std::numeric_limits<float>::max)()) &&
-                                static_cast<double>(static_cast<float>(dval)) == dval) ||
-                               std::isinf(dval);
-#ifdef __GNUC__
-                    JSON_HEDLEY_DIAGNOSTIC_POP
-#endif
+                                dval <= static_cast<double>((std::numeric_limits<float>::max)()));
                     break;
                 }
                 default:

single_include/nlohmann/json.hpp

@@ -8472,10 +8472,8 @@ NLOHMANN_JSON_NAMESPACE_END
 
 
 #include <array> // array
-#include <clocale> // localeconv
 #include <cstddef> // size_t
 #include <cstdio> // snprintf
-#include <cstdlib> // strtof, strtod, strtold, strtoll, strtoull
 #include <initializer_list> // initializer_list
 #include <string> // char_traits, string
 #include <utility> // move
@@ -8496,9 +8494,12 @@ NLOHMANN_JSON_NAMESPACE_END
 
 #include <array> // array
 #include <cfloat> // FLT_EVAL_METHOD
+#include <clocale> // localeconv
 #include <cstddef> // size_t
 #include <cstdint> // int64_t, uint64_t
+#include <cstdlib> // strtof, strtod, strtold
 #include <limits> // numeric_limits
+#include <string> // string
 
 // #include <nlohmann/detail/macro_scope.hpp>
 
@@ -8516,8 +8517,9 @@ NLOHMANN_JSON_NAMESPACE_END
 
 // This file contains the value-conversion helpers used by the lexer to turn an
 // already-validated number token into a value, without the locale/errno
-// overhead of std::strtoull/std::strtod. They are free functions so the lexer
-// stays focused on scanning; see lexer::convert_number().
+// overhead of std::strtoull/std::strtod where possible. They are free functions
+// so the lexer stays focused on scanning (see lexer::convert_number()) and so
+// that other parsers of JSON text can convert tokens exactly like it does.
 
 NLOHMANN_JSON_NAMESPACE_BEGIN
 namespace detail
@@ -8780,6 +8782,184 @@ bool parse_float_from_chars(const char* first, const char* last, FloatType& out)
 #endif
 }
 
+/*!
+@brief check whether Clinger's fast path can still succeed for a float token
+
+parse_float_fast() needs a significand below 2^53. A mantissa with 17 or
+more significant digits is at least 10^16 and therefore always exceeds it,
+so calling the fast path would walk the token one extra time only to
+decline before strtod has to run anyway.
+
+Significant digits are the mantissa's digits from the first nonzero one on;
+the sign, the decimal point, leading zeros, and the exponent do not count.
+The answer is derived from indices - the digits are not scanned again - so
+this stays off the hot path of the number scanners.
+
+@param[in] token                   the validated number token ('.' as decimal point)
+@param[in] decimal_point_position  index of the '.' in @a token, or
+                                   std::string::npos if there is none
+@param[in] mantissa_end            offset just past the last mantissa byte
+@return false if parse_float_fast() is guaranteed to decline
+*/
+inline bool mantissa_fits_clinger(const char* token, std::size_t decimal_point_position, std::size_t mantissa_end) noexcept
+{
+    // 10^16 already exceeds 2^53, so 17 digits can never fit
+    constexpr std::size_t limit = 17;
+
+    const std::size_t neg = (token[0] == '-') ? 1u : 0u;
+    const std::size_t has_dot = (decimal_point_position != std::string::npos) ? 1u : 0u;
+    // the JSON grammar restricts the integer part to "0" or [1-9][0-9]*, so
+    // a leading zero can only be a lone "0", which is not significant
+    const std::size_t lead_zero = (token[neg] == '0') ? 1u : 0u;
+    JSON_ASSERT(mantissa_end >= neg + has_dot + lead_zero);
+    std::size_t digits = mantissa_end - neg - has_dot - lead_zero;
+
+    if (JSON_HEDLEY_LIKELY(digits < limit))
+    {
+        return true;
+    }
+
+    // Only a number below 1 can carry further insignificant zeros, and only
+    // while the count stays at the limit does removing them change the
+    // answer - so this loop is skipped for all but a few tokens. The
+    // fraction is located through decimal_point_position rather than by
+    // searching '.'.
+    if (lead_zero != 0)
+    {
+        JSON_ASSERT(has_dot != 0); // an integer "0" cannot reach the limit
+        for (std::size_t i = decimal_point_position + 1;
+                digits >= limit && i < mantissa_end && token[i] == '0'; ++i)
+        {
+            --digits;
+        }
+    }
+
+    return digits < limit;
+}
+
+/*!
+@brief convert a validated float token without the C library, if possible
+
+Tries std::from_chars (when available) and then Clinger's exact fast path
+(double only), skipping the latter when it cannot succeed.
+
+@param[in]  first                   pointer to the first character of the token
+@param[in]  last                    pointer past the last character
+@param[in]  decimal_point_position  index of the '.' in the token, or
+                                    std::string::npos if there is none
+@param[in]  mantissa_end            offset just past the last mantissa byte (the
+                                    index of 'e'/'E', or the token length)
+@param[out] value                   the converted value on success
+@return true if the value was converted; false if convert_float_locale_aware()
+        must convert it
+*/
+template<typename FloatType>
+bool convert_float_fast(const char* first, const char* last, std::size_t decimal_point_position,
+                        std::size_t mantissa_end, FloatType& value) noexcept
+{
+    if (parse_float_from_chars(first, last, value))
+    {
+        return true;
+    }
+    // Skipping a fast path that cannot succeed is lossless and saves a full
+    // extra pass over the token's bytes, which otherwise shows up on
+    // high-precision inputs such as canada.json
+    return mantissa_fits_clinger(first, decimal_point_position, mantissa_end)
+           && parse_float_fast(first, last, value);
+}
+
+/// std::strtof, std::strtod, or std::strtold, chosen by the type of @a f
+JSON_HEDLEY_NON_NULL(2)
+inline void strtof_by_type(float& f, const char* str, char** endptr) noexcept
+{
+    f = std::strtof(str, endptr);
+}
+
+/// std::strtof, std::strtod, or std::strtold, chosen by the type of @a f
+JSON_HEDLEY_NON_NULL(2)
+inline void strtof_by_type(double& f, const char* str, char** endptr) noexcept
+{
+    f = std::strtod(str, endptr);
+}
+
+/// std::strtof, std::strtod, or std::strtold, chosen by the type of @a f
+JSON_HEDLEY_NON_NULL(2)
+inline void strtof_by_type(long double& f, const char* str, char** endptr) noexcept
+{
+    f = std::strtold(str, endptr);
+}
+
+/// return the decimal point of the current locale
+inline char get_decimal_point() noexcept
+{
+    const auto* loc = localeconv();
+    JSON_ASSERT(loc != nullptr);
+    return (loc->decimal_point == nullptr) ? '.' : *(loc->decimal_point);
+}
+
+/*!
+@brief convert a validated float token with strtof/strtod/strtold
+
+These functions expect the decimal point of the *current* locale, so it is
+looked up right before the conversion instead of once when the lexer is
+constructed: a locale change in between (by a parser callback, a SAX
+handler, or another thread) must not truncate the value (#5198). The
+token has been validated before, so if the conversion stops early and the
+decimal point changed in the meantime, the locale changed between the
+lookup and the call, and the conversion is repeated with the new decimal
+point. If the decimal point did not change, a retry cannot succeed: the
+locale's decimal point is not a single character (e.g., the two-byte
+U+066B of ar_EG.UTF-8 or fa_IR.UTF-8) and cannot be substituted in place.
+The value strtod parsed up to that point is kept, as before this change.
+
+Note that changing the locale in another thread *while* strtod runs is
+undefined behavior of the C library, which this function cannot prevent.
+
+@param[in,out] token                   the token with '.' as decimal point; its
+                                       decimal point is replaced during the
+                                       conversion and restored afterwards
+                                       (data() must be NUL-terminated)
+@param[in]     decimal_point_position  index of the '.' in @a token, or
+                                       std::string::npos if there is none
+@param[out]    value                   the converted value
+*/
+template<typename StringType, typename FloatType>
+void convert_float_locale_aware(StringType& token, std::size_t decimal_point_position, FloatType& value)
+{
+    const bool has_dot = decimal_point_position != std::string::npos;
+    char decimal_point = get_decimal_point();
+    for (;;)
+    {
+        const bool substitute = has_dot && decimal_point != '.';
+        if (substitute)
+        {
+            token[decimal_point_position] = static_cast<typename StringType::value_type>(decimal_point);
+        }
+
+        char* endptr = nullptr; // NOLINT(misc-const-correctness,cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+        strtof_by_type(value, token.data(), &endptr);
+
+        if (substitute)
+        {
+            // the caller hands the token on (e.g. to the SAX interface) with '.'
+            token[decimal_point_position] = '.';
+        }
+
+        if (JSON_HEDLEY_LIKELY(endptr == token.data() + token.size()))
+        {
+            return;
+        }
+
+        // retry only if the locale changed; otherwise, this would loop forever
+        const char current_decimal_point = get_decimal_point();
+        if (current_decimal_point == decimal_point)
+        {
+            return;
+        }
+        decimal_point = current_decimal_point;
+    }
+}
+
 }  // namespace detail
 NLOHMANN_JSON_NAMESPACE_END
 
@@ -9309,18 +9489,6 @@ class lexer : public lexer_base<BasicJsonType>
     ~lexer() = default;
 
   private:
-    /////////////////////
-    // locales
-    /////////////////////
-
-    /// return the decimal point of the current locale
-    static char get_decimal_point() noexcept
-    {
-        const auto* loc = localeconv();
-        JSON_ASSERT(loc != nullptr);
-        return (loc->decimal_point == nullptr) ? '.' : *(loc->decimal_point);
-    }
-
     /////////////////////
     // scan functions
     /////////////////////
@@ -10128,24 +10296,6 @@ class lexer : public lexer_base<BasicJsonType>
         }
     }
 
-    JSON_HEDLEY_NON_NULL(2)
-    static void strtof(float& f, const char* str, char** endptr) noexcept
-    {
-        f = std::strtof(str, endptr);
-    }
-
-    JSON_HEDLEY_NON_NULL(2)
-    static void strtof(double& f, const char* str, char** endptr) noexcept
-    {
-        f = std::strtod(str, endptr);
-    }
-
-    JSON_HEDLEY_NON_NULL(2)
-    static void strtof(long double& f, const char* str, char** endptr) noexcept
-    {
-        f = std::strtold(str, endptr);
-    }
-
     /*!
     @brief scan a number literal
 
@@ -10185,7 +10335,7 @@ class lexer : public lexer_base<BasicJsonType>
     @note The scanner is independent of the current locale: token_buffer
           always holds `.`. Only the std::strtod fallback of convert_number()
           depends on the locale, and it looks up the decimal point right
-          before converting (see convert_float_locale_aware()).
+          before converting (see detail::convert_float_locale_aware()).
     */
     token_type scan_number()  // lgtm [cpp/use-of-goto] `goto` is used in this function to implement the number-parsing state machine described above. By design, any finite input will eventually reach the "done" state or return token_type::parse_error. In each intermediate state, 1 byte of the input is appended to the token_buffer vector, and only the already initialized variables token_buffer, number_type, and error_message are manipulated.
     {
@@ -10516,59 +10666,6 @@ scan_number_done:
         return token_type::uninitialized;
     }
 
-    /*!
-    @brief check whether Clinger's fast path can still succeed for this token
-
-    parse_float_fast() needs a significand below 2^53. A mantissa with 17 or
-    more significant digits is at least 10^16 and therefore always exceeds it,
-    so calling the fast path would walk the token one extra time only to
-    decline before strtod has to run anyway.
-
-    Significant digits are the mantissa's digits from the first nonzero one on;
-    the sign, the decimal point, leading zeros, and the exponent do not count.
-    The answer is derived from indices - the digits are not scanned again - so
-    this stays off the hot path of the number scanners.
-
-    @param[in] mantissa_end  offset just past the last mantissa byte in
-                             token_buffer
-    @return false if parse_float_fast() is guaranteed to decline
-    */
-    bool mantissa_fits_clinger(std::size_t mantissa_end) const
-    {
-        // 10^16 already exceeds 2^53, so 17 digits can never fit
-        constexpr std::size_t limit = 17;
-
-        const std::size_t neg = (!token_buffer.empty() && token_buffer[0] == '-') ? 1u : 0u;
-        const std::size_t has_dot = (decimal_point_position != std::string::npos) ? 1u : 0u;
-        // the JSON grammar restricts the integer part to "0" or [1-9][0-9]*, so
-        // a leading zero can only be a lone "0", which is not significant
-        const std::size_t lead_zero = (token_buffer[neg] == '0') ? 1u : 0u;
-        JSON_ASSERT(mantissa_end >= neg + has_dot + lead_zero);
-        std::size_t digits = mantissa_end - neg - has_dot - lead_zero;
-
-        if (JSON_HEDLEY_LIKELY(digits < limit))
-        {
-            return true;
-        }
-
-        // Only a number below 1 can carry further insignificant zeros, and only
-        // while the count stays at the limit does removing them change the
-        // answer - so this loop is skipped for all but a few tokens. The
-        // fraction is located through decimal_point_position rather than by
-        // searching '.'.
-        if (lead_zero != 0)
-        {
-            JSON_ASSERT(has_dot != 0); // an integer "0" cannot reach the limit
-            for (std::size_t i = decimal_point_position + 1;
-                    digits >= limit && i < mantissa_end && token_buffer[i] == '0'; ++i)
-            {
-                --digits;
-            }
-        }
-
-        return digits < limit;
-    }
-
     /*!
     @brief convert the number text in token_buffer to its value and token type
 
@@ -10582,7 +10679,7 @@ scan_number_done:
                              token_buffer (the index of 'e'/'E', or
                              token_buffer.size() when there is no exponent);
                              used to skip Clinger's fast path when it cannot
-                             possibly succeed - see mantissa_fits_clinger()
+                             possibly succeed - see detail::mantissa_fits_clinger()
     */
     token_type convert_number(token_type number_type, std::size_t mantissa_end)
     {
@@ -10655,77 +10752,15 @@ scan_number_done:
         // (Eisel-Lemire, locale-independent, correctly rounded) when available;
         // otherwise the exact Clinger fast path (double only); otherwise the
         // locale-aware strtof/strtod/strtold.
-        if (parse_float_from_chars(num_begin, num_end, value_float))
-        {
-            return token_type::value_float;
-        }
-        // Skipping a fast path that cannot succeed is lossless and saves a full
-        // extra pass over the token's bytes, which otherwise shows up on
-        // high-precision inputs such as canada.json
-        if (mantissa_fits_clinger(mantissa_end)
-                && parse_float_fast(num_begin, num_end, value_float))
+        if (convert_float_fast(num_begin, num_end, decimal_point_position, mantissa_end, value_float))
         {
             return token_type::value_float;
         }
 
-        convert_float_locale_aware();
+        convert_float_locale_aware(token_buffer, decimal_point_position, value_float);
         return token_type::value_float;
     }
 
-    /*!
-    @brief convert the float in token_buffer with strtof/strtod/strtold
-
-    These functions expect the decimal point of the *current* locale, so it is
-    looked up right before the conversion instead of once when the lexer is
-    constructed: a locale change in between (by a parser callback, a SAX
-    handler, or another thread) must not truncate the value (#5198). The
-    token has been validated before, so if the conversion stops early and the
-    decimal point changed in the meantime, the locale changed between the
-    lookup and the call, and the conversion is repeated with the new decimal
-    point. If the decimal point did not change, a retry cannot succeed: the
-    locale's decimal point is not a single character (e.g., the two-byte
-    U+066B of ar_EG.UTF-8 or fa_IR.UTF-8) and cannot be substituted in place.
-    The value strtod parsed up to that point is kept, as before this change.
-
-    Note that changing the locale in another thread *while* strtod runs is
-    undefined behavior of the C library, which this function cannot prevent.
-    */
-    void convert_float_locale_aware()
-    {
-        const bool has_dot = decimal_point_position != std::string::npos;
-        char decimal_point = get_decimal_point();
-        for (;;)
-        {
-            const bool substitute = has_dot && decimal_point != '.';
-            if (substitute)
-            {
-                token_buffer[decimal_point_position] = static_cast<typename string_t::value_type>(decimal_point);
-            }
-
-            char* endptr = nullptr; // NOLINT(misc-const-correctness,cppcoreguidelines-pro-type-vararg,hicpp-vararg)
-            strtof(value_float, token_buffer.data(), &endptr);
-
-            if (substitute)
-            {
-                // get_string() hands the token to the SAX interface with '.'
-                token_buffer[decimal_point_position] = '.';
-            }
-
-            if (JSON_HEDLEY_LIKELY(endptr == token_buffer.data() + token_buffer.size()))
-            {
-                return;
-            }
-
-            // retry only if the locale changed; otherwise, this would loop forever
-            const char current_decimal_point = get_decimal_point();
-            if (current_decimal_point == decimal_point)
-            {
-                return;
-            }
-            decimal_point = current_decimal_point;
-        }
-    }
-
     /*!
     @brief contiguous fast path for scanning a number
 
@@ -15495,15 +15530,10 @@ class binary_reader
                                is_ndarray can only return `true` when its initial value
                                is `false`
     @param[in] prefix  type marker if already read, otherwise set to 0
-    @param[in] ndarray_dtype  the element type marker of the enclosing bjdata ndarray if
-                              already known (it precedes the dimension vector read here),
-                              otherwise 0; used to emit the "_ArrayType_" annotation key
-                              before "_ArraySize_" if a dimension vector turns out to
-                              describe an ndarray
 
     @return whether size determination completed
     */
-    bool get_ubjson_size_value(std::size_t& result, bool& is_ndarray, char_int_type prefix = 0, char_int_type ndarray_dtype = 0)
+    bool get_ubjson_size_value(std::size_t& result, bool& is_ndarray, char_int_type prefix = 0)
     {
         if (prefix == 0)
         {
@@ -15673,37 +15703,8 @@ class binary_reader
                         }
                     }
 
-                    if (JSON_HEDLEY_UNLIKELY(!sax->start_object(3)))
-                    {
-                        return false;
-                    }
-
-                    // the element type precedes the dimension vector (see get_ubjson_size_type)
-                    // and is passed down as ndarray_dtype; emit it here so the annotation keys
-                    // follow the documented _ArrayType_, _ArraySize_, _ArrayData_ order
-                    if (ndarray_dtype != 0)
-                    {
-                        auto it = std::lower_bound(bjd_types_map.begin(), bjd_types_map.end(), ndarray_dtype, [](const bjd_type & p, char_int_type t)
-                        {
-                            return p.first < t;
-                        });
-                        if (JSON_HEDLEY_UNLIKELY(it == bjd_types_map.end() || it->first != ndarray_dtype))
-                        {
-                            auto last_token = get_token_string();
-                            return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
-                                                    exception_message(input_format, "invalid byte: 0x" + last_token, "type"), nullptr));
-                        }
-
-                        string_t type_key = "_ArrayType_";
-                        string_t type = it->second; // sax->string() takes a reference
-                        if (JSON_HEDLEY_UNLIKELY(!sax->key(type_key) || !sax->string(type)))
-                        {
-                            return false;
-                        }
-                    }
-
                     string_t key = "_ArraySize_";
-                    if (JSON_HEDLEY_UNLIKELY(!sax->key(key) || !sax->start_array(dim.size())))
+                    if (JSON_HEDLEY_UNLIKELY(!sax->start_object(3) || !sax->key(key) || !sax->start_array(dim.size())))
                     {
                         return false;
                     }
@@ -15804,7 +15805,7 @@ class binary_reader
                                         exception_message(input_format, concat("expected '#' after type information; last byte: 0x", last_token), "size"), nullptr));
             }
 
-            const bool is_error = get_ubjson_size_value(result.first, is_ndarray, 0, result.second);
+            const bool is_error = get_ubjson_size_value(result.first, is_ndarray);
             // an ndarray was read here only if the flag flipped; when it was
             // seeded true, get_ubjson_size_value() already rejected the nested
             // dimension vector
@@ -16040,17 +16041,30 @@ class binary_reader
         if (input_format == input_format_t::bjdata && size_and_type.first != npos && (size_and_type.second & (1 << 8)) != 0)
         {
             size_and_type.second &= ~(static_cast<char_int_type>(1) << 8);  // use bit 8 to indicate ndarray, here we remove the bit to restore the type marker
+            auto it = std::lower_bound(bjd_types_map.begin(), bjd_types_map.end(), size_and_type.second, [](const bjd_type & p, char_int_type t)
+            {
+                return p.first < t;
+            });
+            string_t key = "_ArrayType_";
+            if (JSON_HEDLEY_UNLIKELY(it == bjd_types_map.end() || it->first != size_and_type.second))
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
+                                        exception_message(input_format, "invalid byte: 0x" + last_token, "type"), nullptr));
+            }
+
+            string_t type = it->second; // sax->string() takes a reference
+            if (JSON_HEDLEY_UNLIKELY(!sax->key(key) || !sax->string(type)))
+            {
+                return false;
+            }
 
-            // the "_ArrayType_" and "_ArraySize_" annotation keys were already emitted by
-            // get_ubjson_size_value() (the type marker is known before the dimension vector
-            // that determines size_and_type.first is read, so it is emitted first there to
-            // match the documented _ArrayType_, _ArraySize_, _ArrayData_ key order)
             if (size_and_type.second == 'C' || size_and_type.second == 'B')
             {
                 size_and_type.second = 'U';
             }
 
-            string_t key = "_ArrayData_";
+            key = "_ArrayData_";
             if (JSON_HEDLEY_UNLIKELY(!sax->key(key) || !sax->start_array(size_and_type.first) ))
             {
                 return false;
@@ -22342,21 +22356,9 @@ class binary_writer
                 case 'd':
                 {
                     const auto dval = el.template get<double>();
-#ifdef __GNUC__
-                    JSON_HEDLEY_DIAGNOSTIC_PUSH
-                    JSON_HEDLEY_PRAGMA(GCC diagnostic ignored "-Wfloat-equal")
-#endif
-                    // a value that would be rounded (rather than exactly represented) by the
-                    // narrowing to float is treated like an out-of-range integer element above;
-                    // this is the same criterion write_compact_float() uses for CBOR/MessagePack
-                    in_range = std::isnan(dval) ||
+                    in_range = !std::isfinite(dval) ||
                                (dval >= static_cast<double>(std::numeric_limits<float>::lowest()) &&
-                                dval <= static_cast<double>((std::numeric_limits<float>::max)()) &&
-                                static_cast<double>(static_cast<float>(dval)) == dval) ||
-                               std::isinf(dval);
-#ifdef __GNUC__
-                    JSON_HEDLEY_DIAGNOSTIC_POP
-#endif
+                                dval <= static_cast<double>((std::numeric_limits<float>::max)()));
                     break;
                 }
                 default:

tests/src/unit-bjdata.cpp

@@ -11,7 +11,6 @@
 #define JSON_TESTS_PRIVATE
 #include <nlohmann/json.hpp>
 using nlohmann::json;
-using ordered_json = nlohmann::ordered_json;
 
 #include <algorithm>
 #include <climits>
@@ -2295,33 +2294,29 @@ TEST_CASE("BJData")
 
         SECTION("start_array() in ndarray _ArraySize_")
         {
-            // _ArrayType_ (2 events: key + string) is now emitted before
-            // _ArraySize_ (see GitHub issue #5661), which shifts the events
-            // below later by the same 2 events
             std::vector<uint8_t> const v = {'[', '$', 'i', '#', '[', '$', 'i', '#', 'i', 2, 2, 1, 1, 2};
-            SaxCountdown scp(4);
+            SaxCountdown scp(2);
             CHECK_FALSE(json::sax_parse(v, &scp, json::input_format_t::bjdata));
         }
 
         SECTION("number_integer() in ndarray _ArraySize_")
         {
             std::vector<uint8_t> const v = {'[', '$', 'U', '#', '[', '$', 'i', '#', 'i', 2, 2, 1, 1, 2};
-            SaxCountdown scp(5);
+            SaxCountdown scp(3);
             CHECK_FALSE(json::sax_parse(v, &scp, json::input_format_t::bjdata));
         }
 
         SECTION("key() in ndarray _ArrayType_")
         {
-            // _ArrayType_ is emitted right after start_object(), before _ArraySize_
             std::vector<uint8_t> const v = {'[', '$', 'U', '#', '[', '$', 'U', '#', 'i', 2, 2, 2, 1, 2, 3, 4};
-            SaxCountdown scp(1);
+            SaxCountdown scp(6);
             CHECK_FALSE(json::sax_parse(v, &scp, json::input_format_t::bjdata));
         }
 
         SECTION("string() in ndarray _ArrayType_")
         {
             std::vector<uint8_t> const v = {'[', '$', 'U', '#', '[', '$', 'U', '#', 'i', 2, 2, 2, 1, 2, 3, 4};
-            SaxCountdown scp(2);
+            SaxCountdown scp(7);
             CHECK_FALSE(json::sax_parse(v, &scp, json::input_format_t::bjdata));
         }
 
@@ -2924,22 +2919,6 @@ TEST_CASE("BJData")
                 CHECK(out_single.at(0) == '{');
                 CHECK(json::from_bjdata(out_single) == j_single);
 
-                // a double element that is finite and within the range of "single"
-                // but is not exactly representable as a float, so narrowing it would
-                // silently round it (0.1 is read back as 0.10000000149011612); this,
-                // like the overflow case above, falls back to a plain object (see
-                // GitHub issue #5661)
-                json const j_single_rounded = json({{"_ArrayType_", "single"}, {"_ArraySize_", {2, 1}}, {"_ArrayData_", {1.5, 0.1}}});
-                const auto out_single_rounded = json::to_bjdata(j_single_rounded);
-                CHECK(out_single_rounded.at(0) == '{');
-                CHECK(json::from_bjdata(out_single_rounded) == j_single_rounded);
-
-                // a double element that underflows to 0 when narrowed to "single"
-                json const j_single_underflow = json({{"_ArrayType_", "single"}, {"_ArraySize_", {2, 1}}, {"_ArrayData_", {1.5, 1e-300}}});
-                const auto out_single_underflow = json::to_bjdata(j_single_underflow);
-                CHECK(out_single_underflow.at(0) == '{');
-                CHECK(json::from_bjdata(out_single_underflow) == j_single_underflow);
-
                 // in-range boundary values still use the compact ndarray encoding
                 json const j_uint8_ok = json({{"_ArrayType_", "uint8"}, {"_ArraySize_", {2, 1}}, {"_ArrayData_", {0, 255}}});
                 CHECK(json::to_bjdata(j_uint8_ok) == std::vector<uint8_t>({'[', '$', 'U', '#', '[', 'i', 2, 'i', 1, ']', 0, 255}));
@@ -2953,23 +2932,6 @@ TEST_CASE("BJData")
                 CHECK(json::from_bjdata(out_single_ok) == json({{"_ArrayType_", "single"}, {"_ArraySize_", {2, 1}}, {"_ArrayData_", {1.5f, -1.5f}}}));
             }
 
-            SECTION("ndarray annotation keys are read back in the documented order")
-            {
-                // from_bjdata() must emit the annotation object's keys in the order
-                // used throughout the documentation, _ArrayType_, _ArraySize_,
-                // _ArrayData_: the type marker precedes the dimension vector on the
-                // wire (see get_ubjson_size_type()), so it is known, and emitted,
-                // before _ArraySize_. For a plain json this key order is invisible
-                // (its comparison ignores it), but for an ordered_json it is not (see
-                // GitHub issue #5661).
-                const ordered_json o = ordered_json::parse(R"({"_ArrayType_":"uint8","_ArraySize_":[2,2],"_ArrayData_":[1,2,3,4]})");
-                const auto packed = ordered_json::to_bjdata(o);
-                CHECK(packed.at(0) == '[');
-                const ordered_json o_back = ordered_json::from_bjdata(packed);
-                CHECK(o_back == o);
-                CHECK(o_back.dump() == o.dump());
-            }
-
             SECTION("ndarray that would not be read back as an annotated object stays as object")
             {
                 // the reader only restores an annotated object from an ND-array

tools/amalgamate/CHANGES.md

@@ -8,3 +8,6 @@ The following changes have been made to the code with respect to <https://github
   - membership check
   - made function from `_is_within`
   - removed unused variable `actual_path`
+- Added the optional config key `external`: include paths listed there are kept as
+  `#include` directives instead of being inlined (the first directive per path; the
+  repeated ones are commented out).

tools/amalgamate/README.md

@@ -57,6 +57,11 @@ Python v.2.7.0 or higher is required.
    amalgamation. Have a look at `test/source.c.json` and `test/include.h.json`
    to see two examples.
 
+   The optional `external` list names include paths that are kept as `#include`
+   directives instead of being inlined, e.g. `["nlohmann/json.hpp"]` for a header
+   that includes another amalgamated header. Only the first directive for each
+   of these paths is kept; the repeated ones are commented out.
+
  * The `-s, --source` option should specify the path to the source directory.
    This is useful for supporting separate source and build directories.
 

tools/amalgamate/amalgamate.py

@@ -62,6 +62,10 @@ class Amalgamation(object):
         return None
 
     def __init__(self, args):
+        # include paths that are kept as #include directives instead of
+        # being inlined (e.g. a header amalgamated on its own)
+        self.external = []
+        self.included_external = []
         with open(args.config, 'r') as f:
             config = json.loads(f.read())
             for key in config:
@@ -220,11 +224,14 @@ class TranslationUnit(object):
         while include_match:
             if not _is_within(include_match, skippable_contexts):
                 include_path = include_match.group("path")
-                search_same_dir = include_match.group(1) == '"'
-                found_included_path = self.amalgamation.find_included_file(
-                    include_path, self.file_dir if search_same_dir else None)
-                if found_included_path:
-                    includes.append((include_match, found_included_path))
+                if include_path in self.amalgamation.external:
+                    includes.append((include_match, None))
+                else:
+                    search_same_dir = include_match.group(1) == '"'
+                    found_included_path = self.amalgamation.find_included_file(
+                        include_path, self.file_dir if search_same_dir else None)
+                    if found_included_path:
+                        includes.append((include_match, found_included_path))
 
             include_match = self.include_pattern.search(self.content,
                                                         include_match.end())
@@ -235,6 +242,17 @@ class TranslationUnit(object):
         for include in includes:
             include_match, found_included_path = include
             tmp_content += self.content[prev_end:include_match.start()]
+            if found_included_path is None:
+                # an external header: keep the first directive and comment
+                # out the repeated ones
+                include_path = include_match.group("path")
+                if include_path in self.amalgamation.included_external:
+                    tmp_content += "// {0}".format(include_match.group(0))
+                else:
+                    self.amalgamation.included_external.append(include_path)
+                    tmp_content += include_match.group(0)
+                prev_end = include_match.end()
+                continue
             tmp_content += "// {0}\n".format(include_match.group(0))
             if found_included_path not in self.amalgamation.included_files:
                 t = TranslationUnit(found_included_path, self.amalgamation, False)

tools/amalgamate/config_json_view.json

@@ -0,0 +1,9 @@
+{
+	"project": "JSON for Modern C++",
+	"target": "single_include/nlohmann/json_view.hpp",
+	"sources": [
+		"include/nlohmann/json_view.hpp"
+	],
+	"include_paths": ["include"],
+	"external": ["nlohmann/json.hpp"]
+}