Input/Output

The library provides parsing and serialization algorithms to transform JSON to and from the value container as needed. This is accomplished through free functions and classes, described as follows.

Parsing

Parsing is the process where a serialized JSON text is validated and decomposed into elements. The library provides these functions and types to assist with parsing:

Table 4. Parsing Functions and Types
Name	Description
`basic_parser`	A SAX push parser implementation which converts a serialized JSON text into a series of member function calls to a user provided handler. This allows custom behaviors to be implemented for representing the document in memory.
`parse_options`	A structure used to select which extensions are enabled during parsing.
`parse`	Parse a string containing a complete serialized JSON text, and return a `value`.
`parser`	A stateful DOM parser object which may be used to efficiently parse a series of JSON texts each contained in a single contiguous character buffer, returning each result as a `value`.
`stream_parser`	A stateful DOM parser object which may be used to efficiently parse a series of JSON texts incrementally, returning each result as a `value`.
`value_stack`	A low level building block used for efficiently building a `value`. The parsers use this internally, and users may use it to adapt foreign parsers to produce this library’s containers.

The parse function offers a simple interface for converting a serialized JSON text to a value in a single function call. This overload uses exceptions to indicate errors:

value jv = parse( "[1,2,3,4,5]" );

Alternatively, an error_code can be used:

boost::system::error_code ec;
value jv = parse( "[1,2,3,4,5]", ec );
if( ec )
    std::cout << "Parsing failed: " << ec.message() << "\n";

Even when using error codes, exceptions thrown from the underlying memory_resource are still possible:

try
{
    boost::system::error_code ec;
    value jv = parse( "[1,2,3,4,5]", ec );
    if( ec )
        std::cout << "Parsing failed: " << ec.message() << "\n";
}
catch( std::bad_alloc const& e)
{
    std::cout << "Parsing failed: " << e.what() << "\n";
}

The value returned in the preceding examples use the default memory resource. The following code uses a monotonic_resource, which results in faster parsing. jv is marked const to prevent subsequent modification, because containers using a monotonic resource waste memory when mutated.

monotonic_resource mr;
value const jv = parse( "[1,2,3,4,5]", &mr );

Non-Standard JSON

Unless otherwise specified, the parser in this library is strict. It recognizes only valid, standard JSON. The parser can be configured to allow certain non-standard extensions by filling in a parse_options structure and passing it by value. By default all extensions are disabled:

parse_options opt;                // all extensions default to off
opt.allow_comments = true;        // permit C and C++ style comments
                                  // to appear in whitespace
opt.allow_trailing_commas = true; // allow an additional trailing comma in
                                  // object and array element lists
opt.allow_invalid_utf8 = true;    // skip utf-8 validation of keys and strings
opt.allow_invalid_utf16 = true;   // replace invalid surrogate pair UTF-16 code point(s)
                                  // with the Unicode replacement character

value jv = parse( "[1,2,3,] // comment ", storage_ptr(), opt );

When building with C++20 or later, the use of designated initializers with parse_options is possible:

value jv = parse(
    "[1,2,3,] // comment ",
    storage_ptr(),
    {
        .allow_comments = true,             // permit C and C++ style comments
                                            // to appear in whitespace
        .allow_trailing_commas = true,      // allow a trailing comma in object and array lists
        .allow_invalid_utf8 = true          // skip utf-8 validation of keys and strings
    });

When allow_invalid_utf16 is enabled, the parser will not throw an error in the case of illegal leading, trailing, or half a surrogate. Instead, it will replace the invalid UTF-16 code point(s) with the Unicode replacement character.

value jv = parse( "{\"command\":\"\\uDF3E\\uDEC2\"}", storage_ptr(),
    {
        .allow_invalid_utf16 = true       // replace illegal leading surrogate pair with ��
    });

When enabling comment support take extra care not to drop whitespace when reading the input. For example, std::getline removes the endline characters from the string it produces.

Full Precision Number Parsing

The default algorithm that the library uses to parse numbers is fast, but may result in slight precision loss. This may not be suitable for some applications, so there is an option to enable an alternative algorithm that doesn’t have that flaw, but is somewhat slower. To do this, you also need to use parse_options structure.

parse_options opt;
opt.numbers = number_precision::precise;
value jv = parse( "1002.9111801605201", storage_ptr(), opt );

Note that full precision number parsing requires the algorithm to see the full number. This means, that when used with stream_parser, additional memory allocations may be necessary to store the number parts which were so far accepted by the parser. The library does try its best to avoid such allocations.

Parser

Instances of parser and stream_parser offer functionality beyond what is available when using the parse free functions:

More control over memory
Streaming API, parse input JSON incrementally
Improved performance when parsing multiple JSON texts
Ignore non-JSON content after the end of a JSON text

The parser implementation uses temporary storage space to accumulate values during parsing. When using the parse free functions, this storage is allocated and freed in each call. However, by declaring an instance of parser or stream_parser, this temporary storage can be reused when parsing more than one JSON text, reducing the total number of dynamic memory allocations.

To use the parser, declare an instance. Then call parser::write once with the buffer containing representing the input JSON. Finally, call parser::release to take ownership of the resulting value upon success. This example persists the parser instance in a class member to reuse across calls:

class connection
{
    parser p_;                    // persistent data member

public:
    void do_read( string_view s ) // called for each complete message from the network
    {
        p_.reset();               // start parsing a new JSON using the default resource
        p_.write( s );            // parse the buffer, using exceptions to indicate error
        do_rpc( p_.release() );   // process the command
    }

    void do_rpc( value jv );
};

Sometimes a protocol may have a JSON text followed by data that is in a different format or specification. The JSON portion can still be parsed by using the function parser::write_some. Upon success, the return value will indicate the number of characters consumed from the input, which will exclude the non-JSON characters:

stream_parser p;
boost::system::error_code ec;
string_view s = "[1,2,3] %HOME%";
std::size_t n = p.write_some( s, ec );
assert( ! ec && p.done() && n == 8 );
s = s.substr( n );
value jv = p.release();
assert( s == "%HOME%" );

The parser instance may be constructed with parse options which allow some non-standard JSON extensions to be recognized:

parse_options opt;                     // All extensions default to off
opt.allow_comments = true;             // Permit C and C++ style comments to appear in whitespace
opt.allow_trailing_commas = true;      // Allow an additional trailing comma in
                                       // object and array element lists
opt.allow_invalid_utf8 = true;         // Skip utf-8 validation of keys and strings
stream_parser p( storage_ptr(), opt ); // The stream_parser will use the options

Streaming Parser

The stream_parser implements a streaming algorithm; it allows incremental processing of large JSON inputs using one or more contiguous character buffers. The entire input JSON does not need to be loaded into memory at once. A network server can use the streaming interface to process incoming JSON in fixed-size amounts, providing these benefits:

CPU consumption per I/O cycle is bounded
Memory consumption per I/O cycle is bounded
Jitter, unfairness, and latency is reduced
Less total memory is required to process the full input

To use the stream_parser, declare an instance. Then call stream_parser::write zero or more times with successive buffers representing the input JSON. When there are no more buffers, call stream_parser::finish. The function stream_parser::done returns true after a successful call to write or finish if parsing is complete.

In the following example a JSON text is parsed from standard input a line at a time. Error codes are used instead. The function stream_parser::finish is used to indicate the end of the input:

This example will break, if comments are enabled, because of std::getline use (see the warning in Non-Standard JSON section).

value read_json( std::istream& is, boost::system::error_code& ec )
{
    stream_parser p;
    std::string line;
    while( std::getline( is, line ) )
    {
        p.write( line, ec );
        if( ec )
            return nullptr;
    }
    p.finish( ec );
    if( ec )
        return nullptr;
    return p.release();
}

We can complicate the example further by extracting several JSON values from the sequence of lines.

std::vector<value> read_jsons( std::istream& is, boost::system::error_code& ec )
{
    std::vector< value > jvs;
    stream_parser p;
    std::string line;
    std::size_t n = 0;
    while( true )
    {
        if( n == line.size() )
        {
            if( !std::getline( is, line ) )
                break;
            n = 0;
        }

        n += p.write_some( line.data() + n, line.size() - n, ec );

        if( p.done() )
        {
            jvs.push_back( p.release() );
            p.reset();
        }
    }
    if( !p.done() )   // this part handles the cases when the last JSON text in
    {                 // the input is either incomplete or doesn't have a marker
        p.finish(ec); // for end of the value (e.g. it is a number)
        if( ec.failed() )
            return jvs;
        jvs.push_back( p.release() );
    }

    return jvs;
}

Controlling Memory

After default construction, or after stream_parser::reset is called with no arguments, the value produced after a successful parse operation uses the default memory resource. To use a different memory resource, call reset with the resource to use. Here we use a monotonic_resource, which is optimized for parsing but not subsequent modification:

monotonic_resource mr;

stream_parser p;
p.reset( &mr );                // Use mr for the resulting value
p.write( "[1,2,3,4,5]" );      // Parse the input JSON
value const jv = p.release();  // Retrieve the result
assert( *jv.storage() == mr ); // Same memory resource

To achieve performance and memory efficiency, the parser uses a temporary storage area to hold intermediate results. This storage is reused when parsing more than one JSON text, reducing the total number of calls to allocate memory and thus improving performance. Upon construction, the memory resource used to perform allocations for this temporary storage area may be specified. Otherwise, the default memory resource is used. In addition to a memory resource, the parser can make use of a caller-owned buffer for temporary storage. This can help avoid dynamic allocations for small inputs. The following example uses a four kilobyte temporary buffer for the parser, and falls back to the default memory resource if needed:

unsigned char temp[ 4096 ]; // Declare our buffer
stream_parser p(
    storage_ptr(),          // Default memory resource
    parse_options{},        // Default parse options (strict parsing)
    temp);                  // Use our buffer for temporary storage

Avoiding Dynamic Allocations

Through careful specification of buffers and memory resources, it is possible to eliminate all dynamic allocation completely when parsing JSON, for the case where the entire JSON text is available in a single character buffer, as shown here:

/*  Parse JSON and invoke the handler

    This function parses the JSON specified in `s`
    and invokes the handler, whose signature must
    be equivalent to:

        void( value const& jv );

    The operation is guaranteed not to perform any
    dynamic memory allocations. However, some
    implementation-defined upper limits on the size
    of the input JSON and the size of the resulting
    value are imposed.

    Upon error, an exception is thrown.
*/
template< class Handler >
void do_rpc( string_view s, Handler&& handler )
{
    unsigned char temp[ 4096 ]; // The parser will use this storage for its temporary needs
    parser p(                   // Construct a strict parser using
                                // the temp buffer and no dynamic memory
        get_null_resource(),    // The null resource never dynamically allocates memory
        parse_options(),        // Default constructed parse options allow only standard JSON
        temp );

    unsigned char buf[ 16384 ]; // Now we need a buffer to hold the actual JSON values
    static_resource mr2( buf ); // The static resource is monotonic,
                                // using only a caller-provided buffer
    p.reset( &mr2 );            // Use the static resource for producing the value
    p.write( s );               // Parse the entire string we received from the network client

    // Retrieve the value and invoke the handler with it.
    // The value will use `buf` for storage. The handler
    // must not take ownership, since monotonic resources
    // are inefficient with mutation.
    handler( p.release() );
}

Custom Parsers

Users who wish to implement custom parsing strategies may create their own handler to use with an instance of basic_parser. The handler implements the function signatures required by SAX event interface. In Validate example we define the "null" parser, which throws out the parsed results, to use in the implementation of a function that determines if a JSON text is valid.

Edit this page