8. Custom Host Applications

Spicy provides a C++ API for integrating its parsers into custom host applications. There are two different approaches to doing this:

1. If you want to integrate just one specific kind of parser, Spicy can generate C++ prototypes for it that facilitate feeding data and accessing parsing results.

2. If you want to write a generic host application that can support arbitrary parsers, Spicy provides a dynamic runtime introspection API for dynamically instantiating parsers and accessing results.

We discuss both approaches in the following.

Note

Internally, Spicy is a layer on top of an intermediary framework called HILTI. It is the HILTI runtime library that implements most of the functionality we’ll look at in this section, so you’ll see quite a bit of HILTI-side functionality. Spicy comes with a small additional runtime library of its own that adds anythings that’s specific to the parsers it generates.

Note

The API for host applications is still in flux, and some parts aren’t the prettiest yet. Specifics of this may change in future versions of HILTI/Spicy.

8.1. Integrating a Specific Parser

We’ll use our simple HTTP example from the Getting Started section as a running example for a parser we want to leverage from a C++ application.

my-http.spicy

module MyHTTP;

const Token      = /[^ \t\r\n]+/;
const WhiteSpace = /[ \t]+/;
const NewLine    = /\r?\n/;

type Version = unit {
    :       /HTTP\//;
    number: /[0-9]+\.[0-9]+/;
};

public type RequestLine = unit {
    method:  Token;
    :        WhiteSpace;
    uri:     Token;
    :        WhiteSpace;
    version: Version;
    :        NewLine;

    on %done {
        print self.method, self.uri, self.version.number;
        }
};

First, we’ll use spicyc to generate a C++ parser from the Spicy source code:

# spicyc -c -g my-http.spicy -o my-http.cc

Option -c (aka --output-c++) tells spicyc that we want it to generate C++ code (rather than compiling everything down into executable code).

Option -g (aka --disable-optimizations) tells spicyc to not perform global optimizations. Optimizations are performed on all modules passed to a invocation of spicyc and can remove e.g., unused code. Since we generate output files with multiple invocations, optimizations could lead to incomplete code.

We also need spicyc to get generate some additional additional “linker” code implementing internal plumbing necessary for cross-module functionality. That’s what -l (aka --output-linker) does:

# spicyc -l -g my-http.cc -o my-http-linker.cc

We’ll compile this linker code along with the my-http.cc.

Next, spicyc can also generate C++ prototypes for us that declare (1) a set of parsing functions for feeding in data, and (2) a struct type providing access to the parsed fields:

# spicyc -P -g my-http.spicy -o my-http.h

The output of -P (aka --output-prototypes) is a bit convoluted because it (necessarily) also contains a bunch of Spicy internals. Stripped down to the interesting parts, it looks like this for our example:

[...]

namespace __hlt::MyHTTP {
    struct RequestLine : hilti::rt::trait::isStruct, hilti::rt::Controllable<RequestLine> {
        std::optional<hilti::rt::Bytes> method{};
        std::optional<hilti::rt::Bytes> uri{};
        std::optional<hilti::rt::ValueReference<Version>> version{};
        [...]
    };

    struct Version : hilti::rt::trait::isStruct, hilti::rt::Controllable<Version> {
        std::optional<hilti::rt::Bytes> number{};
        [...]
    };

[...]
}

namespace hlt::MyHTTP::RequestLine {
    extern auto parse1(hilti::rt::ValueReference<hilti::rt::Stream>& data, const std::optional<hilti::rt::stream::View>& cur) -> hilti::rt::Resumable;
    extern auto parse2(hilti::rt::ValueReference<__hlt::MyHTTP::RequestLine>& unit, hilti::rt::ValueReference<hilti::rt::Stream>& data, const std::optional<hilti::rt::stream::View>& cur) -> hilti::rt::Resumable;
    extern auto parse3(spicy::rt::ParsedUnit& gunit, hilti::rt::ValueReference<hilti::rt::Stream>& data, const std::optional<hilti::rt::stream::View>& cur) -> hilti::rt::Resumable;
}

[...]

Todo

The struct declarations should move into the public namespace.

You can see the struct definitions corresponding to the two unit types, as well as a set of parsing functions with three different signatures:

parse1

The simplest form of parsing function receives a stream of input data, along with an optional view into the stream to limit the region to parse if desired. parse` will internally instantiate an instance of the unit’s struct, and then feed the unit’s parser with the data stream. However, it won’t provide access to what’s being parsed as it doesn’t pass back the struct.

parse2

The second form takes a pre-instantiated instance of the unit’s struct type, which parsing will fill out. Once parsing finishes, results can be accessed by inspecting the struct fields.

parse3

The third form takes a pre-instantiated instance of a generic, type-erased unit type that the parsing will fill out. Accessing the data requires use of HILTI’s reflection API, which we will discuss in Supporting Arbitrary Parsers.

Let’s start by using parse1():

my-http-host.cc

#include <iostream>

#include <hilti/rt/libhilti.h>

#include <spicy/rt/libspicy.h>

#include "my-http.h"

int main(int argc, char** argv) {
    assert(argc == 2);

    // Initialize runtime library.
    hilti::rt::init();

    // Create stream with $1 as data.
    auto stream = hilti::rt::reference::make_value<hilti::rt::Stream>(argv[1]);
    stream->freeze();

    // Feed data.
    hlt::MyHTTP::RequestLine::parse1(stream, {}, {});

    // Wrap up runtime library.
    hilti::rt::done();

    return 0;
}

This code first instantiates a stream from data giving on the command line. It freezes the stream to indicate that no further data will arrive later. Then it sends the stream into the parse1() function for processing.

We can now use the standard C++ compiler to build all this into an executable, leveraging spicy-config to add the necessary flags for finding includes and libraries:

# clang++ -o my-http my-http-host.cc my-http.cc my-http-linker.cc $(spicy-config --cxxflags --ldflags)
# ./my-http $'GET index.html HTTP/1.0\n'
GET, /index.html, 1.0

The output comes from the execution of the print statement inside the Spicy grammar, demonstrating that the parsing proceeded as expected.

When using parse1() we don’t get access to the parsed information. If we want that, we can use parse2() instead and provide it with a struct to fill in:

my-http-host.cc

#include <iostream>

#include <hilti/rt/libhilti.h>
#include <spicy/rt/libspicy.h>

#include "my-http.h"

int main(int argc, char** argv) {
    assert(argc == 2);

    // Initialize runtime libraries.
    hilti::rt::init();
    spicy::rt::init();

    // Create stream with $1 as data.
    auto stream = hilti::rt::reference::make_value<hilti::rt::Stream>(argv[1]);
    stream->freeze();

    // Instantiate unit.
    auto request = hilti::rt::reference::make_value<__hlt::MyHTTP::RequestLine>();

    // Feed data.
    hlt::MyHTTP::RequestLine::parse2(request, stream, {}, {});

    // Access fields.
    std::cout << "method : " << *request->method << std::endl;
    std::cout << "uri    : " << *request->uri << std::endl;
    std::cout << "version: " << *(*request->version)->number << std::endl;

    // Wrap up runtime libraries.
    spicy::rt::done();
    hilti::rt::done();

    return 0;
}

# clang++ -o my-http my-http-host.cc my-http-host.cc $(spicy-config --cxxflags --ldflags)
# ./my-http $'GET index.html HTTP/1.0\n'
GET, /index.html, 1.0
method : GET
uri    : /index.html
version: 1.0

Another approach to retrieving field values goes through Spicy hooks calling back into the host application. That’s how the Zeek plugin operates. Let’s say we want to execute a custom C++ function every time a RequestList has been parsed. By adding the following code to my-http.spicy, we (1) declare that function on the Spicy-side, and (2) implement a Spicy hook that calls it:

my-http.spicy

public function got_request_line(method: bytes, uri: bytes, version_number: bytes) : void &cxxname="got_request_line";

on RequestLine::%done {
    got_request_line(self.method, self.uri, self.version.number);
}

The &cxxname attribute for got_request_line indicates to Spicy that this is a function implemented externally inside custom C++ code, accessible through the given name. Now we need to implement that function:

my-http-callback.cc

#include <iostream>

#include <hilti/rt/libhilti.h>

#include <spicy/rt/libspicy.h>

void got_request_line(const hilti::rt::Bytes& method, const hilti::rt::Bytes& uri, const hilti::rt::Bytes& version_number) {
    std::cout << "In C++ land: " << method << ", " << uri << ", " << version_number << std::endl;
}

Finally, we compile it altogether:

# spicyc -c -g my-http.spicy -o my-http.cc
# spicyc -l -g my-http.cc -o my-http-linker.cc
# spicyc -P -g my-http.spicy -o my-http.h
# clang++ -o my-http my-http.cc my-http-linker.cc my-http-callback.cc my-http-host.cc $(spicy-config --cxxflags --ldflags)
# ./my-http $'GET index.html HTTP/1.0\n'
In C++ land: GET, index.html, 1.0
GET, index.html, 1.0

Note that the C++ function signature needs to match what Spicy expects, based on the Spicy-side prototype. If you are unsure how Spicy arguments translate into C++ arguments, look at the C++ prototype that’s included for the callback function in the output of -P.

A couple more notes on the compilation process for integrating Spicy-generated code into custom host applications:

• Above we used spicyc -l to link our Spicy code from just a single Spicy source file. If you have more than one source file, you need to link them altogether in a single step. For example, if we had A.spicy, B.spicy and C.spicy, we’d do:

  # spicyc -c -g A.spicy -o A.cc
  # spicyc -c -g B.spicy -o B.cc
  # spicyc -c -g C.spicy -o C.cc
  # spicyc -l -g A.cc B.cc C.cc -o linker.cc
  # clang++ A.cc B.cc C.cc linker.cc -o a.out ...
  

• If your Spicy code is importing any library modules (e.g., the standard filter module), you’ll need to compile those as well in the same fashion.

8.2. Supporting Arbitrary Parsers

This approach is more complex, and we’ll just briefly describe the main pieces here. All of the tools coming with Spicy support arbitrary parsers and can serve as further examples (e.g., spicy-driver, spicy-dump, the Zeek plugin). Indeed, they all build on the same C++ library class spicy::rt::Driver that provides a higher-level API to working with Spicy’s parsers in a generic fashion. We’ll do the same in the following.

8.2.1. Retrieving Available Parsers

The first challenge for a generic host application is that it cannot know what parsers are even available. Spicy’s runtime library provides an API to get a list of all parsers that are compiled into the current process. Continuing to use the my-http.spicy example, this code prints out our one available parser:

my-http-host.cc

#include <iostream>

#include <hilti/rt/libhilti.h>

int main(int argc, char** argv) {
    assert(argc == 2);

    // Initialize runtime libraries.
    hilti::rt::init();
    spicy::rt::init();

    // Instantiate driver providing higher level parsing API.
    spicy::rt::Driver driver;

    // Print out available parsers.
    print(unit->value());

    // Wrap up runtime libraries.
    spicy::rt::done();
    hilti::rt::done();

    return 0;
}

# clang++ -o my-http my-http-host.cc my-http.cc my-http-linker.cc $(spicy-config --cxxflags --ldflags)
# ./my-http
Available parsers:

    MyHTTP::RequestLine

Using the name of the parser (MyHTTP::RequestLine) we can instantiate it from C++, and then feed it data:

    // Retrieve meta object describing parser.
    auto parser = driver.lookupParser("MyHTTP::RequestLine");
    assert(parser);

    // Fill string stream with $1 as data to parse.
    std::stringstream data(argv[1]);

    // Feed data.
    auto unit = driver.processInput(**parser, data);

# clang++ -o my-http my-http-host.cc my-http.cc my-http-linker.cc $(spicy-config --cxxflags --ldflags)
# ./my-http $'GET index.html HTTP/1.0\n'
GET, /index.html, 1.0

That’s the output of the print statement once more.

unit is of type spicy::rt::ParsedUnit, which is a type-erased class holding, in this case, an instance of _hlt::MyHTTP::RequestLine. Internally, that instance went through the parse3() function that we have encountered in the previous section. To access the parsed fields, there’s a visitor API to iterate generically over HILTI types like this unit:

void print(const hilti::rt::type_info::Value& v) {
    const auto& type = v.type();
    switch ( type.tag ) {
        case hilti::rt::TypeInfo::Bytes: std::cout << type.bytes->get(v); break;
        case hilti::rt::TypeInfo::ValueReference: print(type.value_reference->value(v)); break;
        case hilti::rt::TypeInfo::Struct:
            for ( const auto& [f, y] : type.struct_->iterate(v) ) {
                std::cout << f.name << ": ";
                print(y);
                std::cout << std::endl;
            }
            break;
        default: assert(false);
    }
}

Adding print(unit->value() after the call to processInput() then gives us this output:

# clang++ -o my-http my-http-host.cc my-http.cc my-http-linker.cc $(spicy-config --cxxflags --ldflags)
# ./my-http $'GET index.html HTTP/1.0\n'
GET, /index.html, 1.0
method: GET
uri: /index.html
version: number: 1.0

Our visitor code implements just what we need for our example. The source code of spicy-dump shows a full implementation covering all available types.

So far we have compiled the Spicy parsers statically into the generated executable. The runtime API supports loading them dynamically as well from pre-compiled HLTO files through the class hilti::rt::Library. Here’s the full example leveraging that, taking the file to load from the command line:

my-driver

int main(int argc, char** argv) {
    // Usage now: "my-driver <hlto> <name-of-parser> <data>"
    assert(argc == 4);

    // Load pre-compiled parser. This must come before initializing the
    // runtime libraries.
    auto library = hilti::rt::Library(argv[1]);
    auto rc = library.open();
    assert(rc);

    // Initialize runtime libraries.
    hilti::rt::init();
    spicy::rt::init();

    // Instantiate driver providing higher level parsing API.
    spicy::rt::Driver driver;

    // Print out available parsers.
    driver.listParsers(std::cout);

    // Retrieve meta object describing parser.
    auto parser = driver.lookupParser(argv[2]);
    assert(parser);

    // Fill string stream with $1 as data to parse.
    std::stringstream data(argv[3]);

    // Feed data.
    auto unit = driver.processInput(**parser, data);
    assert(unit);

    // Print out conntent of parsed unit.
    print(unit->value());

    // Wrap up runtime libraries.
    spicy::rt::done();
    hilti::rt::done();

    return 0;
}

# $(spicy-config --cxx) -o my-driver my-driver.cc $(spicy-config --cxxflags --ldflags --dynamic-loading)
# spicyc -j my-http.spicy >my-http.hlto
# ./my-driver my-http.hlto "$(cat data)"
Available parsers:

    MyHTTP::RequestLine

GET, /index.html, 1.0
method: GET
uri: /index.html
version: number: 1.0

Note

Note the addition of --dynamic-loading to the hilti-config command line. That’s needed when the resulting binary will dynamically load precompiled Spicy parsers because linker flags need to be slightly adjusted in that case.

8.3. API Documentation

We won’t go further into details of the HILTI/Spicy runtime API here. Please see C++ API documentation for more on that, the namespaces hilti::rt and spicy::rt cover what’s available to host applications.

Our examples always passed the full input at once. You don’t need to do that, Spicy’s parsers can process input incrementally as it comes in, and return back to the caller to retrieve more. See the source of spicy::Driver::processInput() for an example of how to implement that.