Systematic Generation Data and Types in C++

Sumant Tambe
Sr. Software Engineer and Microsoft MVP
LinkedIn
@sutambe
SF Bay Area ACCU
Dec 13, 2017

12/15/2017 2
Alternative Title
2015

12/15/2017 3
Alternative Title

Functional Programming in C++ by
Manning
12/15/2017 5
Author
Ivan Cukic

This work was done in
12/15/2017 6
Connectivity Software for the Industrial Internet of Things (IIoT)

Why should you care?
• If you write software for living and if you are
serious about any of the following
– Bug-free, Expressive code
– Productivity
• Modularity
• Reusability
• Extensibility
• Maintainability
– and having fun while doing all of the above!
12/15/2017 7

Agenda
• Property-based Testing (briefly)
• Why composable generators
• Implementation of a functional generator library
– Less than 1500 lines of code
– Wrapping a coroutine as a generator (if time permits)
• Mathematics of API design
– Functor, Monoid, and Monad
• Type generators
– Generators at compile-time
12/15/2017 8

What’s a Property
• Reversing a linked-list twice has no effect
– reverse(reverse(list)) == list
• Compression can make things smaller
– compress(input).size <= input.size
• Round-trip serialization/deserialization has no
effect
– deserialize(serialize(input)) == input
12/15/2017 9

Testing RefleX
• RefleX is short for Reflection for DDS-XTypes
– DDS-XTypes is like Avro, ProtoBuf, etc.
• Reflex is a library that acts like a code generator
• Serialization and Deserialization
– Serializes types and data
– Compile-time reflection on C++ structs/classes
• C++11/14
• Link
12/15/2017 10
C++ Code
(classes)
XTypes
Schema and
Data

RefleX Example
12/15/2017 11
class ShapeType
{
std::string color_;
int x_, y_, shapesize_;
public:
ShapeType() {}
ShapeType(const std::string & color,
int x, int y, int shapesize)
: color_(color), x_(x), y_(y),
shapesize_(shapesize)
{}
std::string & color();
int & x();
int & y();
int & shapesize();
};
#include "reflex.h"
REFLEX_ADAPT_STRUCT(
ShapeType,
(std::string, color(), KEY)
(int, x())
(int, y())
(int, shapesize()))
Uses boost Fusion (duh!)

A property-test in RefleX
12/15/2017 12
template <class T>
void test_roundtrip_property(const T & in)
{
T out;
reflex::TypeManager<T> tm; // create Xtypes schema
reflex::SafeDynamicData<T> safedd =
tm.create_dynamicdata(in); // create Xtypes data
reflex::read_dynamicdata(out, safedd);
assert(in == out); // must be true for all T
}
What is T?
What is in?

Property-based Testing
• Complementary to traditional example-based testing
• Specify post-conditions that must hold no matter what
• Encourages the programmer to think harder about the code
• More declarative
• Need data generators to produce inputs
• Need type generators to produce input types
• Free reproducers!
– Good property-based testing frameworks shrink inputs to minimal
automatically
• Famous
– Haskell’s QuickCheck
– Scala’s ScalaTest
12/15/2017 13

Random Integer Generator
12/15/2017 15
• The simplest data generator
size_t seed = time(nullptr);
srandom(seed);
size_t rand()
{
return random();
}
• Alternatively
size_t seed =
std::chrono::system_clock::now().time_since_epoch().count();
size_t rand()
{
static std::mt19937 mt(seed);
return mt();
}

Fundamental Data Generators
12/15/2017 16
• A bool generator
bool bool_gen() {
return rand() % 2;
}
• An uppercase character generator
char uppercase_gen() // A=65, Z=90
return static_cast<char>(‘A’ + rand() % 26))
}

Data Generators
12/15/2017 17
• A range generator
– Generates any number in the range (inclusive)
int range_gen(int lo, int hi)
return lo + rand() % (hi – lo + 1);
}
• Inconvenient because args must be passed every time

Closure as generator
12/15/2017 18
• A stateful range generator
– Generates a random number in the given range
auto make_range_gen(int lo, int hi)
return [lo, hi]() {
return lo + rand() % (hi – lo + 1);
}
}
auto r = range_gen(20, 25);
std::cout << r(); // one of 20, 21, 22, 23, 24, 25

Closure as generator
12/15/2017 19
• The oneof generator
template <class T>
auto make_oneof_gen(std::initializer_list<T> list)
{
return [options = std::vector<T>(list)]() {
return *(options.begin() + rand() % options.size());
};
}
auto days = { “Sun”, “Mon”, “Tue”, “Wed”, “Thu”, “Fri”, “Sat” }
auto gen = make_oneof_gen(days);
std::cout << gen(); // one of Sun, Mon, Tue, Wed, Thu, Fri, Sat

Closure vs Object
• Is function/closure a sufficient abstraction for
any generator?
[]() {
return blah;
}
• Answer depends upon your language
– No. In C++
12/15/2017 20

12/15/2017 21
“Closures are poor man's
objects; Objects are poor
man's closure”
-- The venerable master Qc NaSource Google images

The Generator Class Template
12/15/2017 22
template <class T, class GenFunc>
struct Gen : private GenFunc
{
using value_type = T;
explicit Gen(GenFunc func)
: GenFunc(std::move(func))
{ }
T generate()
{
return GenFunc::operator()();
}
// ...
};

Creating Gen<T> from lambdas
template <class GenFunc>
auto make_gen_from(GenFunc&& func)
{
return
Gen<decltype(func()), GenFunc>(std::forward<GenFunc>(func));
}
template <class Integer>
auto make_range_gen(Integer lo, Integer hi)
{
return make_gen_from([lo, hi]() {
return static_cast<Integer>(lo + rand() % (hi - lo));
});
}
12/15/2017 23

The Real Power of Generators: Composition
• make_gen_from is low level
• Better—Producing complex generators from one or
more simpler Generators
12/15/2017 24
Generator
map
reduce
concat
where
zip_with
concat_map
take
scan

The Generator Class Template
12/15/2017 25
template <class T, class GenFunc>
struct Gen : private GenFunc
{
using value_type = T;
explicit Gen(GenFunc func)
: GenFunc(std::move(func))
{ }
T generate()
{
return GenFunc::operator()();
}
auto map(...) const;
auto zip_with(...) const;
auto concat(...) const;
auto take(...) const;auto concat_map(...) const;
auto reduce(...) const;
auto where(...) const;
auto scan(...) const;
};

Mapping over a generator
12/15/2017 26
vector<string> days = { “Sun”, “Mon”, “Tue”, “Wed”,
“Thu”, “Fri”, “Sat” };
auto range_gen = gen::make_range_gen(0,6);
// random numbers in range 0..6
auto day_gen = range_gen.map([&](int i) { return days[i]; });
// random days in range “Sun” to “Sat”
while(true)
std::cout << day_gen.generate();
// random days in range “Sun” to “Sat”
Gen<int>  Gen<string>
f(int  string)
std::function<string(int)>

12/15/2017 27
Generator is a
You can
over it.

Functor Laws
• Composition Law
gen.map(f).map(g) == gen.map(g o f)
12/15/2017 28
auto DAY_gen = range_gen
.map([&](int i) { return days[i]; });
.map([](const string & str) {
return std::toupper(str, std::locale());
});
auto DAY_gen = range_gen.map([&](int i) {
return std::to_upper(days[i], std::locale());
});

Functor Laws
• Identity Law
gen.map(a -> a) == gen
12/15/2017 29
auto gen2 == gen.map([](auto a) { return a; });
• Both gen and gen2 will generate the same data
– Their identities are different but they are still the same

Implementing map
template <class T>
template <class Func>
auto Gen<T>::map(Func&& func) const &
{
return make_gen_from(
[self = *this, // copy
func = std::forward<Func>(func)]() mutable {
return func(self.generate());
});
}
12/15/2017 30
auto Gen<T>::map(Func&& func) &&
{
[self = std::move(*this), // move
func = std::forward<Func>(func)]() mutable {
return func(self.generate());
});
}

Monoid Laws
• Identity Law
M o id = id o M = M
• Binary operation o = append
• Appending to identity (either way) must yield
the same generator
12/15/2017 32
auto identity = gen::make_empty_gen<int>();
gen::make_inorder_gen({ 1, 2, 3}).append(identity)
==
identity.append(gen::make_inorder_gen({ 1, 2, 3});

Monoid Laws
• Associative Law
(f o g) o h = f o (g o h)
• Binary operation o = append
• Order of operation does not matter. (The order of
operands does)
– Not to be confused with commutative property where the order of
operands does not matter
12/15/2017 33
auto f = gen::make_inorder_gen({ 1, 2});
auto g = gen::make_inorder_gen({ 3, 4});
auto h = gen::make_inorder_gen({ 5, 6});
(f.append(g)).append(h) == f.append(g.append(h));

Creating Identity Generator
make_empty_gen
template <class T>
auto make_empty_gen()
{
return make_gen_from([]() {
throw std::out_of_range("empty generator!");
return std::declval<T>();
});
}
12/15/2017 34

Gen::append
12/15/2017 35
template <class UGen>
auto append(UGen&& ugen) const
{
[tgen = *this,
ugen = std::forward<UGen>(ugen),
tdone = false,
udone = false]() mutable {
if(!tdone)
{
try {
return tgen.generate();
}
catch(std::out_of_range &)
{
tdone = true;
}
}
if(!udone)
{
try {
return ugen.generate();
}
catch(std::out_of_range &) {
udone = true;
}
}
throw std::out_of_range(
"concat: both generators done!");
});
}

Zipping Generators
12/15/2017 36
auto x_gen = gen::make_stepper_gen(0,100);
auto y_gen = gen::make_stepper_gen(0,200,2);
auto point_gen =
x_gen.zip_with([](int x, int y)
return std::make_pair(x, y);
}, y_gen);
for(auto _: { 1,2,3 })
std::cout << point_gen.generate(); // {0,0}, {1,2}, {3,4}
end
zip_with is a generator of arbitrary struct/tuple values

Implementing zip_with
template <class T>
template <class Zipper, class... GenList>
auto Gen<T>::zip_with(Zipper&& func, GenList&&... genlist) const
{
[self = *this,
genlist...,
func = std::forward<Zipper>(func)]() mutable {
return func(self.generate(), genlist.generate()...);
});
}
12/15/2017 37

Dependent Generators (concat_map)
auto triangle_gen =
gen::make_inorder_gen({1,2,3,4,5,4,3,2,1})
.concat_map([](int i) {
std::cout << "n";
return gen::make_stepper_gen(1, i);
});
try {
while(true)
std::cout << triangle_gen.generate();
}
catch(std::out_of_range &) {
}
12/15/2017 39
1
12
123
1234
12345
1234
123
12
1

Monad Laws
auto f = [](int i) { return gen::make_stepper_gen(1, i); };
auto g = [](int i) { return gen::make_stepper_gen(i, 1, -1); };
auto M = gen::make_single_gen(300);
// left identity
M.concat_map(f) == f(300);
// right identity
M == M.concat_map([](int i) {
return gen::make_single_gen(i);
});
// associativity
M.concat_map(f).concat_map(g)
==
M.concat_map([f,g](int i) mutable {
return f(i).concat_map(g);
}));
12/15/2017 40

Monad’s “return” is make_single_gen
12/15/2017 41
template <class T>
auto make_single_gen(T&& t)
{
return make_gen_from([t = std::forward<T>(t),
done = false]() mutable {
if(!done)
{
done = true;
return t;
}
else
throw std::out_of_range("single generator completed!");
});
}

Monad’s “bind” is Gen::concat_map
• See generator.h
12/15/2017 42

12/15/2017 43
The Mathematics of the Generator API

Generators
• General design principal
– Create a language to solve a problem
– API is the language
– Reuse parsing/compilation of the host language (duh!)
• The “language” of Generators
– Primitive values are basic generators
– Complex values are composite generators
– All APIs result in some generator
• map, zip_with, concat_map, reduce, etc.
– I.e., generators form an algebra
12/15/2017 44

Algebra
A first-class, lawful abstraction
+
Compositional* API
12/15/2017 45
*Composition based on mathematical concepts

12/15/2017 46
What Math Concepts?
• Algebraic Structures from Category Theory
–Monoid
–Functor
–Monad
–Applicative
• Where to begin?
–Haskell
–But innocent should start with Scala perhaps

Back to the property-test in RefleX
12/15/2017 47
template <class T>
void test_roundtrip_property(const T & in)
{
T out;
reflex::TypeManager<T> tm;
reflex::SafeDynamicData<T> safedd =
tm.create_dynamicdata(in);
reflex::read_dynamicdata(out, safedd);
assert(in == out);
}
What is T?
What is in?

12/15/2017 48
if we can generate
we can generate

Random Numbers at Compile-time
#include <stdio.h>
enum test { Foo = RANDOM };
int main(void)
{
enum test foo = Foo;
printf("foo = %dn", foo);
return 0;
}
12/15/2017 49
$ gcc –DRANDOM=$RANDOM main.c –o main
$ ./main
foo = 17695

Random Numbers at Compile-time
enum test { Foo = RANDOM, Bar = LFSR(Foo) };
int main(void) {
enum test foo = Foo;
enum test bar = Bar;
printf("foo = %d, bar = %dn", foo, bar);
}
12/15/2017 50
constexpr int LFSR_bit(int prev) {
return (((prev >> 0) ^ (prev >> 2) ^
(prev >> 3) ^ (prev >> 5)) & 1);
}
constexpr int LFSR(int prev) {
return ((prev >> 1) | (LFSR_bit(prev) << 15));
}
$ gcc –DRANDOM=$RANDOM main.c –o main
$ ./main
foo = 17695, bar = 41615

Linear Feedback Shift Register (LFSR)
• Simple random number generator
• 4 bit Fibonacci LFSR
• Feedback polynomial = X4 + X3 + 1
• Taps 4 and 3
12/15/2017 51Image source: wikipedia

Linear Feedback Shift Register (LFSR)
• A 16 bit feedback polynomial
– X16 + X14 + X13 + X11 + 1
• Taps: 16, 14, 13, and 11
12/15/2017 52Image source: wikipedia

Random Type Generation at Compile-Time
template <int I> struct TypeMap;
template <> struct TypeMap<0> {
typedef int type;
};
typedef char type;
};
typedef float type;
};
typedef double type;
};
12/15/2017 53

Random Type Generation at Compile-Time
#include <boost/core/demangle.hpp>
template <int seed>
struct RandomTuple {
typedef typename TypeMap<LFSR(seed) % 4>::type First;
typedef typename TypeMap<LFSR(LFSR(seed)) % 4>::type Second;
typedef typename TypeMap<LFSR(LFSR(LFSR(seed))) % 4>::type Third;
typedef std::tuple<First, Second, Third> type;
};
12/15/2017 54
$ g++ –DRANDOM=$RANDOM main.cpp –o main
$ ./main
foo = 11660, bar = 5830
tuple = std::tuple<float, double, char>
int main(void) {
auto tuple = RandomTuple<RANDOM>::type();
printf("tuple = %s",
boost::core::demangle(typeid(tuple).name()).c_str());
}
Live code

Tuples all the way down…
• Nested type generation using recursive
TypeMap
• Classic template meta-programming
• See type_generator.h
12/15/2017 56

Type Generators Demystified
• Random seed as command line preprocess argument
• Compile-time random numbers using LFSR constexpr
• C++ meta-programming for type synthesis
• std::string generator for type names and member
names
• RefleX to map types to typecode and dynamic data
• Source: https://github.com/sutambe/cpp-generators
12/15/2017 57

RefleX Property-test Type Generator
• “Good” Numbers
– 31415 (3 nested structs)
– 7937 ( 5 nested structs)
– ...
• “Bad” Numbers
– 2152 (test seg-faults) sizeof(tuple) > 2750 KB!
12/15/2017 58
Github: https://github.com/rticommunity/rticonnextdds-reflex/blob/develop/test/generator

Systematic Generation Data and Types in C++

Recommended

More Related Content

What's hot (20)

Similar to Systematic Generation Data and Types in C++ (20)

More from Sumant Tambe (17)

Recently uploaded (20)

Systematic Generation Data and Types in C++

Editor's Notes