C++ boost::variant generic converter - c++

I've been trying for the last three day to figure out how to implement a generic way of getting the value out of a boost::variant<...>, but it's been quite difficult.
Here is the solution I could come up with, which it not at all generic:
#include <iostream>
#include "boost\variant\variant.hpp"
using MyVariant = boost::variant<int, std::string>;
class VariantConverter : public boost::static_visitor<>
{
private:
mutable int _int;
mutable std::string _string;
static VariantConverter apply(MyVariant& v)
{
VariantConverter res;
v.apply_visitor(res);
return res; // copy will be elided, right?
}
public:
void operator()(int& i) const
{
_int = i;
}
void operator() (std::string& s) const
{
_string = s;
}
static int to_int(MyVariant v)
{
return apply(v).from_int();
}
static std::string to_string(MyVariant v)
{
return apply(v).from_string();
}
int from_int()
{
return _int;
};
std::string from_string()
{
return _string;
};
};
int main()
{
using namespace std;
MyVariant v = 23;
int i = VariantConverter::to_int(v);
cout << i << endl;
v = "Michael Jordan";
std::string s = VariantConverter::to_string(v);
cout << s.c_str() << endl;
cin.get();
return 0;
}
I'd appreciate it if someone could guide me towards a better solution.
Or perhaps someone could explain to me the rationale behind this:
if I declare a:
using MyVariant = boost::variant<int, std::string>;
and then a:
ConverterToInt : basic_visitor<int> {
public:
int operator() (int i) { return i; };
};
Why is it that when I try to apply the ConverterToInt to a MyVariant as such:
ConverterToInt cti;
MyVariant i = 10;
i.apply_visitor(cti);
I get a compiler error about trying to find a operator() that takes a std::string?
It seems to me that apply_visitor is trying to call an operator() for each of the types MyVariant can take. Is that so? If it is, why? How can i avoid this behavior?
Cheers!

You can avoid the error message by telling ConverterToInt what to do with a std::string. You might know that i can't be a std::string but it's unreasonable to expect the compiler to know that (and if it is true, why are you using a variant?).
apply_visitor will only call the correct operator() method, but it decides at run time, and the compiler needs to have all the possibilities covered to generate the code.

MyVariant iv = 10;
int i = boost::get<int>(iv);
boost::variant does not "call" each operator() of an interface when invoked, but it must be able to. That's the entire point. A variant can hold any of the template types, so if you want to define an operation on it, you must specify somewhere what that operation means for each type.

Related

Better way to give constant view of pointers in C++

I have a class that must return a constant view of some pointers to the upper layers of software.
Internally, the pointers must be non-const, because the class will need to manipulate the objects internally.
I don't see any option for providing this const view of the pointers to a higher level client, without making a copy of all the pointers. This seems wasteful. What if I was managing millions of objects?
Is there a better way?
Here is some example code:
#include <vector>
#include <iostream>
class example {
public:
example() {
bytePtrs_.push_back(new char);
*bytePtrs_[0] = '$';
}
// I want to do this, but compiler will not allow
// error: could not convert ‘((example*)this)->example::bytePtrs_’ from ‘std::vector<char*>’ to ‘std::vector<const char*>’
std::vector<const char*> getPtrs() {
return bytePtrs_;
}
// Must make wasteful copy
std::vector<const char*> getPtrs() {
std::vector<const char*> ret;
for (auto &ptr : bytePtrs_)
ret.push_back(ptr);
return ret;
}
private:
std::vector<char*> bytePtrs_;
};
int main() {
example e;
std::vector<const char*> bytePtrs = e.getPtrs();
std::cout << bytePtrs[0] << std::endl;
}
You can do this using std::experimental::propagate_const.
That will forward the constness of the pointer onto the pointed-to object.
#include <experimental/propagate_const>
class example {
public:
// using vector = std::vector<char*>>;
using vector = std::vector<std::experimental::propagate_const<char*>>;
example() {
bytePtrs.push_back(new char);
*bytePtrs[0] = '$';
}
vector const& getPtrs() const {
return bytePtrs;
}
private:
vector bytePtrs;
};
int main()
{
example e;
example::vector const& bytePtrs = e.getPtrs();
// dereference this or add a null terminator
std::cout << *bytePtrs[0] << std::endl; // fine and dandy
*bytePtrs[0] = 'x'; // compile error
}
Consider returning a proxy object when you only want to permit const access, something like this (edited to fix the massive hole in the original code pointed out by #alagner!):
#include <iostream>
#include <vector>
template <class T> class proxy_vector
{
public:
proxy_vector (const std::vector<T *> &v) : m_v (v) { }
size_t size () { return m_v.size (); }
const T * const &operator[] (size_t i) const { return m_v [i]; }
// ... more functions as needed
private:
const std::vector <T *> &m_v;
};
class example
{
public:
example() : m_pv (bytePtrs_)
{
bytePtrs_.push_back(new char);
*bytePtrs_[0] = '$';
}
const proxy_vector <char> &getPtrs() { return m_pv; }
private:
std::vector<char*> bytePtrs_;
proxy_vector <char> m_pv;
};
int main()
{
example e;
auto &bytePtrs = e.getPtrs ();
// *bytePtrs [0] = 'x'; // uncomment to show that this code now actually works as intended!
std::cout << bytePtrs[0] << "\n";
}
A decent compiler should optimise most, if not all, of this away. Add access methods to proxy vector as needed, I doubt you will need many.
Demo (seems to work fine in C++11).
What #alagner suggests should also work and might be simpler. I haven't thought that through.
As the OP has pointed out in one of the comments, this is the API the code needs to comply to:
https://github.com/Xilinx/Vitis-AI/blob/master/src/Vitis-AI-Runtime/VART/vart/runner/include/vart/runner.hpp#L157
So by design it returns by copy, which is unavoidable really, thus what really can be done is this:
#include <vector>
#include <iostream>
class example {
public:
example() {
bytePtrs_.push_back(new char);
*bytePtrs_[0] = '$';
}
// Returning by value
// so shallow copy of objects is expected
// i.e. the pointers are copied
std::vector<const char*> getPtrs() {
return std::vector<const char*>(bytePtrs_.begin(), bytePtrs_.end());
}
private:
std::vector<char*> bytePtrs_;
};
int main() {
example e;
std::vector<const char*> bytePtrs = e.getPtrs();
std::cout << bytePtrs[0] << std::endl;
}
This is one of the very few cases where reinterpret_cast<> could be allowed into the code I think: since the std::vector<char *> and const std::vector<const char*>& are assumed to be compatible in memory anyways, you might cast. Thus, getPtrs() could actually return reinterpret_cast<const std::vector<const char*>&>(bytePtrs_);. To be clear: this is an optimization that relies on concrete ABI details (ItaniumABI et. al.), not generic standard C++, but I'm not yet aware of any architecture where it wouldn't work.
In c++20 and later, you might consider checking std::ranges::view.

How can I use unique pointers for an array of parameterized objects in C++ using Visual Studio 2017?

Consider a class whose default constructor takes in the file path as a parameter.
class Test
{
public:
Test(const std::string& filepath);
...
...
};
Now I wish to create and initialize an array of Test objects using unique pointers in VS2017.
int main()
{
std::unique_ptr<Test[]> m_Tests;
int testCount = 2;
std::string path1, path2;
m_Tests = std::make_unique<Test[]>(testCount); // This line gives a compilation error
m_Tests[0] = std::make_unique<Test>(path1);
m_Tests[1] = std::make_unique<Test>(path2);
}
How can I make this work?
g++ 9.2.0 tells me that you lack default constructor, i.e. one without parameters. Adding such constructor works fine. If it's not what you want, you can create array of unique_ptr's, so std::unique_ptr<std::unique_ptr<Test>[]> and after that initialize each element by hand, something similar to this:
#include <memory>
#include <algorithm>
#include <iostream>
struct Test {
std::string str_;
Test(std::string const& str) : str_(str) { }
void print() { std::cout << str_ << '\n'; }
};
int main()
{
std::unique_ptr<std::unique_ptr<Test>[]> m_Tests;
int testCount = 2;
std::string path1{"a"}, path2{"b"};
m_Tests = std::make_unique<std::unique_ptr<Test>[]>(testCount);
std::array<std::string, 2> paths{path1, path2};
std::transform(paths.begin(), paths.end(), &m_Tests[0],
[](auto const& p) { return std::make_unique<Test>(p); });
for (int i = 0 ; i < testCount ; ++i) {
m_Tests[i]->print();
}
}
There is no overload of std::make_unique that does this, so you would need to use new directly:
m_Tests.reset(new Test[testCount]{path1, path2});
This will however only compile if testCount is a constant expression, so you need to change the definition int testCount = 2; to const int or constexpr int.
If testCount is not a constant expression, there needs to be a default constructor defined for the case that testCount is smaller than 2 at runtime.
So, really, you probably want to ignore testCount and just let the array size be deduced:
m_Tests.reset(new Test[]{path1, path2});
It would be much easier if you just used std::vector:
std::vector<Test> m_Tests;
//...
m_Tests.emplace_back(path1);
m_Tests.emplace_back(path2);
How about you use std::array and can you get rid of testCount (or use it as constexp) then the code can be like below.
class Test
{
public:
Test(const std::string& filepath){}
};
int main()
{
constexpr int testCount = 2;
std::unique_ptr<std::array<Test, testCount>> m_Tests;
std::string path1, path2;
m_Tests = std::make_unique<std::array<Test, testCount>>(std::array<Test, testCount>{path1,path2});
}

Data controlled programs in c++

Not to sure how to name this question because the problem itself is looking for a construct of which I don´t know its name.
The problem is I am dealing with programs whose control flow depends greatly of data.
For example I created a MIPS simulator which implemented a list of more than 50 instructions, each implemented on its own and everything governed by a huge switch case
switch (function){ //Function is an int, each function (eg SLL) is
case 0: //associated with one
if (state->debug_level > 0){
fprintf(state->debug_out, "SLL\n");
}
step_err = SLL(state, rs, rt, rd, sa);
break;
case 2:
if (state->debug_level > 0){
fprintf(state->debug_out, "SRL\n");
}
step_err = SRL(state, rs, rt, rd, sa);
break;
case 3:
if (state->debug_level > 0){
fprintf(state->debug_out, "SRA\n");
}
//
I have been told that this could have been implemented using function pointers, but to do so what I am looking for is a way of relating data of any kind, say a string to other data, say an integer. I am aware of maps but wouldn't want to push back each pair. I am looking for some sort of array like syntax I think if seen before which might look something similar to this:
¿type? function_codes[]{
0, "SLL";
2, "SRL";
3, "SRA";
...
}
I am not looking for a solution to this problem but a generic approach to introducing quick relationships between data and using this to modify control flow.
EDIT AFTER ANSWERS
What I was actually looking for but I didnt know was indeed maps but in particular its initialization syntax similar to an array (see accepted answer). This used with function pointers did the required job.
As you guessed, function pointers are in fact a good way to do this. Since you specify that you don't want to use a Map, this is how you would implement your integer-based function dispatch using an array of function pointers. Note that since I don't know the type signature of your MIPS functions (SLL, SRL, etc.) I've used dummy placeholder type names.
typedef ret_t (*mips_func)(arg1_t, arg2_t, arg3_t, arg4_t, arg5_t);
mips_func function_codes[] = {
&SLL,
&SRL,
&SRA,
...
};
//...Later, in the part of your code that used to contain the big switch statement
step_err = (*function_codes[function])(state, rs, rt, rd, sa);
The syntax &SLL gets a pointer to the function SLL, which I assume is already in scope because you can call it directly from your switch statement.
Note that this assumes the numeric codes for the functions are a continuous sequence of integers from 0 to [max code value]. If some numeric codes are unused, then you will either need to leave explicit gaps in your array (by placing a NULL pointer in one or more entries) or use std::map<int, mips_func> so that you can use arbitrary non-continuous integer values as keys to functions. Fortunately, using a Map still doesn't require push_backing each element, since C++ now has initializer lists. The same code using a Map would look like this:
typedef ret_t (*mips_func)(arg1_t, arg2_t, arg3_t, arg4_t, arg5_t);
std::map<int, mips_func> function_codes = {
{0, &SLL},
{2, &SRL},
{4, &SRA},
...
};
//Using the Map looks exactly the same, due to its overloaded operator[]
step_err = (*function_codes[function])(state, rs, rt, rd, sa);
For simplify you can use associative containers. If the order is important then use std::map, or std::unordered_map in the other case.
And you can use syntax similar to the desired
std::map<size_t, std::string> codes_map = decltype(codes_map) {
{ 0, "val1" },
{ 1, "val2" }
};
You could group the data as static members w/ the same name across structs, then use templates to access them generically:
struct A { auto call() const { return "((1))"; }; static const char * name; };
struct B { auto call() const { return "{{2}}"; }; static const char * name; };
struct C { auto call() const { return "<<3>>"; }; static const char * name; };
// n.b. these `T...` have: `sizeof(T) == ... == sizeof(empty_struct)`
const char * A::name = "A";
const char * B::name = "B";
const char * C::name = "C";
boost::variant (and the soon to be implemented std::variant) implements a type-safe union, which provides a very clean and efficient way of using these structs as values:
#include <cstdio>
#include <vector>
#include <boost/variant.hpp>
int main()
{
std::vector<boost::variant<A, B, C>> letters{A{}, B{}, C{}, B{}, A{}};
auto visitor = [](auto x) { std::printf("%s(): %s\n", x.name, x.call()); };
for (auto var : letters) { boost::apply_visitor(visitor, var); }
}
Demo
It seems like you have two problems: the flow-control issue (dispatch) and the map issue (an implementation note). I get that the program flow is nonstatic and unknowable at compile-time… but so is the map static? For static maps I get a lot of mileage out of using a traits-ish approach to create a compile-time mapping. Here’s a quick example mapping file suffixes to Objective-C enum constants:
namespace objc {
namespace image {
template <std::size_t N> inline
constexpr std::size_t static_strlen(char const (&)[N]) { return N; }
template <NSBitmapImageFileType t>
struct suffix_t;
#define DEFINE_SUFFIX(endstring, nstype) \
template <> \
struct suffix_t<nstype> { \
static constexpr std::size_t N = static_strlen(endstring); \
static constexpr char const str[N] = endstring; \
static constexpr NSBitmapImageFileType type = nstype; \
};
DEFINE_SUFFIX("tiff", NSTIFFFileType);
DEFINE_SUFFIX("bmp", NSBMPFileType);
DEFINE_SUFFIX("gif", NSGIFFileType);
DEFINE_SUFFIX("jpg", NSJPEGFileType);
DEFINE_SUFFIX("png", NSPNGFileType);
DEFINE_SUFFIX("jp2", NSJPEG2000FileType);
template <NSBitmapImageFileType nstype>
char const* suffix_value = suffix_t<nstype>::str;
}
}
… see how that works? the nice part is that using it has no runtime overhead, which if your map is static, you can use something like that.
For dynamic flow-control and dispatch, function pointers work; that is what happens automatically if you use polymorphic classes and virtual functions but it seems like you have an architecture in place already that may not be amenable to being redone with such high-modernist architectural notions. I like c++11 lambdas as they solve like 90% of my problems in this arena. Perhaps you can elablrate (I will amend my answer)!
If you only have a small number of indices to support, from 0 to 50, you'll get the best performance if you put your function pointers in an array and not a map.
The syntax is also short:
#include <iostream>
#include <functional>
static void f0() {
std::cout << "f0\n";
}
static void f1() {
std::cout << "f1\n";
}
void main()
{
std::function<void()> f[2] = { f0, f1 };
f[0](); // prints "f0"
f[1](); // prints "f1"
}
Or, if you prefer classes over functions:
#include "stdafx.h"
#include <iostream>
class myfunc {
public:
virtual void run() abstract;
virtual ~myfunc() {}
};
class f0 : public myfunc {
public:
virtual void run() {
std::cout << "f0\n";
}
};
class f1 : public myfunc {
public:
virtual void run() {
std::cout << "f1\n";
}
};
void main()
{
myfunc* f[2] = { new f0(), new f1() };
f[0]->run(); // prints "f0"
f[1]->run(); // prints "f1"
for (int i = 0; i < sizeof(f) / sizeof(f[0]); ++i)
delete f[i];
}
Given some definitions
#include <iostream>
#include <iterator>
#include <algorithm>
#include <stdexcept>
#include <map>
using namespace std;
struct state{
int debug_level = 1;
const char* debug_out = "%s";
} s;
// some functions to call
void SLL(state& s, int, int, int, int){
cout << "SLL";
}
void SLR(state& s, int, int, int, int){
cout << "SLR";
}
void SLT(state& s, int, int, int, int){
cout << "SLT";
}
You can use a Map
auto mappedname2fn = map<string, delctype(SLL)*>{
{"SLL", SLL},
{"SLR", SLR}
};
// call a map function
mappedname2fn["SLR"](s, 1, 2, 3, 4);
If you don't want a map you can use a pre-sorted array for a binary search
Here's a binary search of an array of name, function pairs
template<typename P, int N, typename ...T>
auto callFn(P(&a)[N], string val, T&&... params){
auto it = lower_bound(a, a+N, make_pair(val, nullptr),
[](auto& p1, auto& p2){return p1.first < p2.first;});
if(it==(a+N) || val<it->first) throw logic_error("not found");
return it->second(forward<T>(params)...);
}
So you can set up an array and use that:-
// array sorted in alphabetical order for binary search to work
pair<string, decltype(SLL)*> name2fn[] = {
{"SLL", SLL},
{"SLR", SLR},
{"SLT", SLT}
};
void callFn(string name, state& s, int a, int b, int c, int d){
try{
callFn(name2fn, name, s, a, b, c, d);
}
catch(exception& e){
cout << e.what();
}
}
// call it
callFn("SLL", s, 1, 2, 3, 4);

Template Class to differentiate object type?

Can I use C++ template classes to differentiate object types? Or what should I use?
Eg. I have a class Synonym and it can be of type Statement, Procedure, etc for example. I have functions that accepts these synonyms and evaluates them depending on its type. So I was thinking it will be nice if I can do something like:
enum Types { Statement, Procedure, Variable, ... };
template <typename Types>
class Synonym { ... }
void evaluate(Synonym<Statement> s, Synonym<Variable> v) { do something }
^ so that I can do this ... instead of checking the type in function like:
void evaluate(Synonym s, Synonym v) {
assert(s.type == Statement);
assert(v.type == Variable);
// also would like to eliminate things like: (if possible)
switch(s.type) {
case XXX: doSomething ...
case YYY: doAnotherThing ...
}
}
You could create a function template and then specialize on that template
template<typename Type>
void evaluate (Type t) {}
template<>
void evaluate<Statement>( Statement s)
{}
This way, when you pass a Statement it will pick that overload, and you can do different behaviors depending on type.
I think using a variant and visitor pattern would be suited. Have a look at Boost.Variant here: http://www.boost.org/doc/libs/1_51_0/doc/html/variant.html, the last example (also below but expanded) shows a visitor implementation. There are also other variant and visitor implementations. std::any and loki are also options. I personally like loki but that is probably just because I'm a huge fan of Alexandrescu.
#include "boost/variant.hpp"
#include <iostream>
class ToLengthVisitor : public boost::static_visitor<int>
{
public:
int operator()(int i) const
{
return i;
}
int operator()(const std::string & str) const
{
return str.length();
}
int operator()(const char * str) const
{
const char * temp = str;
while(*temp != '\0') temp++;
return temp-str;
}
};
int main()
{
typedef boost::variant< int, std::string, const char * > MyVariant;
MyVariant u(std::string("hello world"));
std::cout << u; // output: hello world
MyVariant cu(boost::get<std::string>(u).c_str());
int result = boost::apply_visitor( ToLengthVisitor(), u );
std::cout << result; // output: 11 (i.e., length of "hello world")
result = boost::apply_visitor( ToLengthVisitor(), cu );
std::cout << result; // output: 11 (i.e., length of "hello world")
}

std::map of member function pointers?

I need to implement an std::map with <std::string, fn_ptr> pairs. The function pointers are pointers to methods of the same class that owns the map. The idea is to have direct access to the methods instead of implementing a switch or an equivalent.
( I am using std::string as keys for the map )
I'm quite new to C++, so could anyone post some pseudo-code or link that talks about implementing a map with function pointers? ( pointers to methods owned by the same class that owns the map )
If you think there's a better approach to my problem, suggestions are also welcome.
This is about the simplest I can come up with. Note no error checking, and the map could probably usefully be made static.
#include <map>
#include <iostream>
#include <string>
using namespace std;
struct A {
typedef int (A::*MFP)(int);
std::map <string, MFP> fmap;
int f( int x ) { return x + 1; }
int g( int x ) { return x + 2; }
A() {
fmap.insert( std::make_pair( "f", &A::f ));
fmap.insert( std::make_pair( "g", &A::g ));
}
int Call( const string & s, int x ) {
MFP fp = fmap[s];
return (this->*fp)(x);
}
};
int main() {
A a;
cout << a.Call( "f", 0 ) << endl;
cout << a.Call( "g", 0 ) << endl;
}
A template implementation could look like:
class Factory {
public:
enum which {
foo, bar, baz
};
template<which w>
A* newA(...);
...
};
template<Factory::which w>
A* Factory::newA(...) {
/* default implementation */
throw invalid_argument();
}
template<>
A* Factory::newA<Factory::foo>(...) {
/* specialization for a 'foo' style A */
...
}
....
This requires that the value used to determine which newA is called be known at compile time. You could potentially use a const char * as the template parameter, but it's not guaranteed to work on all compilers.
Yet another option is to create helper factories, one for each factory creation method, and store those in the map. This isn't a huge advantage over storing method pointers, but does let you define a default creation method and simplifies fetching things from the map (there's no need to check that the key exists, because you'll get a default factory). On the downside, an entry for each unknown key would be added to the map.
Also, if you use an enum rather than a string for the key type, you shouldn't need to worry about checking whether a key exists in the map. While it's possible for someone to pass an invalid enum key to newA, they'd have to explicitly cast the argument, which means they're not going to do it by accident. I'm having a hard time imagining a case where someone would purposefully cause a crash in newA; the potential scenarios involve security, but an application programmer could crash the app without using your class.
Since C++14, we can use a generic lambda to get rid easily of pointers to member methods.
It follows a minimal, working example of a forward function made up with a generic lambda function:
#include<utility>
#include<map>
#include<string>
#include<iostream>
struct SomeClass { };
struct SomeOtherClass { };
struct Test {
void test(SomeClass) { std::cout << "SomeClass" << std::endl; }
void test(SomeOtherClass) { std::cout << "SomeOtherClass" << std::endl; }
};
int main() {
Test test;
auto l = [&test](auto c){ test.test(c); };
std::map<std::string, decltype(l)> m;
m.emplace("foo", l);
m.emplace("bar", l);
m.at("foo")(SomeClass{});
m.at("bar")(SomeOtherClass{});
}
Another option is to use delegates as oppose to function pointers. This delegate implementation is pretty fast, supports polymorphisms, and plays well with stl containers.
You could have something like:
class MyClass {
public:
// defines
typedef fastdelegate::FastDelegate2<int, int, int> MyDelegate;
typedef std::map<std::string, MyDelegate> MyMap;
// populate your map of delegates
MyClass() {
_myMap["plus"] = fastdelegate::MakeDelegate(this, &Plus);
_myMap["minus"] = fastdelegate::MakeDelegate(this, &Minus);
}
bool Do(const std::string& operation, int a, int b, int& res){
MyMap::const_iterator it = _myMap.find(operation);
if (it != _myMap.end()){
res = it.second(a,b);
return true;
}
return false;
}
private:
int Plus (int a, int b) { return a+b; }
int Minus(int a, int b) { return a-b; }
MyMap _myMap;
};