C++ exceptions and RTTI are often not enabled in many environments, require outstanding library and ABI specific support, and sometimes introduce unwanted cost (i.e exceptions failure path is often considered very slow, RTTI often grows binary size and adds type names to the program binaries which may hurt confidentiality). Most implementations of C++ exceptions also use the RTTI implementation, enlarging the price to be paid for using exceptions.
Exceptions are however very convenient, particularly the automatic error propagation which helps makes code clearer and easier to read, write and maintain. In the journey to try and get a working alternative to standard C++ exceptions for error handling, there are many modern return value based facilities out there, and even macro based utilities to do the "automatic" propagation. There are also known papers and proposals to improve/introduce new and better form of exceptions in C++ which I hope will become part of future standard.
So far there is no solution existing today that I know of, that is "close enough" to the experience of using plain C++ exceptions, that I consider clean, meaning, with automatic and invisible error propagation, and without polluting the code with macros.
With this library I am trying to provide a coroutine based implementation that resembles
C++ exceptions as close as I could get. I made sure that it works also
when compiling without exceptions and RTTI (i.e -fno-exceptions -fno-rtti).
Let's take a look at the following code:
zpp::throwing<int> foo(bool success)
{
if (!success) {
// Throws an exception.
co_yield std::runtime_error("My runtime error");
}
// Returns a value.
co_return 1337;
}Foo returns a zpp::throwing<int> which means it can throw an exception
in our implementation, but on success it returns an int.
int main()
{
return zpp::try_catch([]() -> zpp::throwing<int> {
std::cout << "Hello World\n";
std::cout << co_await foo(false) << '\n';;
co_return 0;
}, [&](const std::exception & error) {
std::cout << "std exception caught: " << error.what() << '\n';
return 1;
}, [&]() {
std::cout << "Unknown exception\n";
return 1;
});
}The zpp::try_catch function attempts to execute the function object that is sent to it, and when an exception is thrown,
one of the following function objects that receives an exception from the appropriate type will get called, similar to a C++
catch block. The last function object can optionally receive no parameters and as such
functions as catch (...) to catch all exceptions.
In this example foo was called with false, and hence it throws the std::runtime_error exception which
is caught at the first lambda function sent as catch block. This lambda then returns 1 which eventually
being returned from main.
Like in normal catch block, it is possible to throw from the catching lambdas:
zpp::throwing<std::string> bar(bool success)
{
return zpp::try_catch([&]() -> zpp::throwing<std::string> {
auto foo_result = co_await foo(success);
std::cout << "Foo succeeded" << foo_result << '\n';
co_return "foo succeeded";
}, [&](const std::runtime_error & error) -> zpp::throwing<std::string> {
std::cout << "Runtime error caught: " << error.what() << '\n';
co_yield std::runtime_error("Foo really failed");
});
}Note that the lambda function sent as a catch block now returns a zpp::throwing<std::string>
which allows it to throw exceptions using co_return.
It is even possible to rethrow the caught exception using zpp::rethrow:
zpp::throwing<std::string> bar(bool success)
{
return zpp::try_catch([&]() -> zpp::throwing<std::string> {
auto foo_result = co_await foo(success);
std::cout << "Foo succeeded" << foo_result << '\n';
co_return "foo succeeded";
}, [&](const std::runtime_error & error) -> zpp::throwing<std::string> {
cout << "Runtime error caught: " << error.what() << '\n';
co_yield zpp::rethrow;
});
}You may observe that once we change bar() to catch std::logic_error instead,
the exception is actually caught by the main() function below as an std::exception:
zpp::throwing<std::string> bar(bool success)
{
return zpp::try_catch([&]() -> zpp::throwing<std::string> {
auto foo_result = co_await foo(success);
std::cout << "Foo succeeded" << foo_result << '\n';
co_return "foo succeeded";
}, [&](const std::logic_error & error) -> zpp::throwing<std::string> {
std::cout << "Logic error caught: " << error.what() << '\n';
co_yield std::runtime_error("Foo really failed");
});
}
int main()
{
return zpp::try_catch([]() -> zpp::throwing<int> {
std::cout << "Hello World\n";
std::cout << co_await bar(false) << '\n';;
co_return 0;
}, [&](const std::exception & error) {
std::cout << "std exception caught: " << error.what() << '\n';
return 1;
}, [&]() {
std::cout << "Unknown exception\n";
return 1;
});
}Since at the time of writing, there is no way in C++ to iterate base classes of a given type, manual registration of exception types is required. The following example shows how to do it:
struct my_custom_exception { virtual ~my_custom_exception() = default; };
struct my_custom_derived_exception : my_custom_exception {};
template <>
struct zpp::define_exception<my_custom_exception>
{
using type = zpp::define_exception_bases<>;
};
template <>
struct zpp::define_exception<my_custom_derived_exception>
{
using type = zpp::define_exception_bases<my_custom_exception>;
};And then throw it like so:
zpp::throwing<int> foo(bool success)
{
if (!success) {
// Throws an exception.
co_yield my_custom_derived_exception();
}
// Returns a value.
co_return 1337;
}int main()
{
zpp::try_catch([]() -> zpp::throwing<void> {
// Throws an exception.
co_yield std::errc::invalid_argument;
}, [](zpp::error error) {
std::cout << "Error: " << error.code() <<
" [" << error.domain().name() << "]: " << error.message() << '\n';
}, []() {
/* catch all */
});
}In the main function above an error value is thrown, from a predefined error
domain of std::errc.
In order to define your own error domains, the following example is provided:
enum class my_error
{
success = 0,
operation_not_permitted = 1,
general_failure = 2,
};
template <>
inline constexpr auto zpp::err_domain<my_error> = zpp::make_error_domain(
"my_error", my_error::success, [](auto code) constexpr->std::string_view {
switch (code) {
case my_error::operation_not_permitted:
return "Operation not permitted.";
case my_error::general_failure:
return "General failure.";
default:
return "Unspecified error.";
}
});You can even catch the error enumeration directly, like so:
int main()
{
zpp::try_catch([]() -> zpp::throwing<void> {
// Throws an exception.
co_return std::errc::invalid_argument;
}, [](std::errc error) {
switch(error) {
case std::errc::invalid_argument: {
std::cout << "Caught the invalid argument directly by enumeration type\n"
<< "And here is the message: " << zpp::error(error).message() << '\n';
break;
}
default: {
zpp::error my_error = error;
std::cout << "Error: " << my_error.code() <<
" [" << my_error.domain().name() << "]: " << my_error.message() << '\n';
}
}
}, []() {
/* catch all */
});
}You may throw also with co_return. The library will understand whether you are actually returning
a value or throwing, by the type of the return expression. Theoretically co_return should generate
better code because it does not add a suspend point but it is highly optimize-able, and the
library actually takes care of destroying the coroutine on the first suspend.
-Example:
zpp::throwing<int> foo(bool success)
{
if (!success) {
// Throws an exception, with `co_return`.
co_return std::runtime_error("My runtime error");
}
// ...
if (!success) {
// Throwing values with `co_return` is also possible.
co_return std::errc::invalid_argument;
}
// Returns a value.
co_return 1337;
}Because being a coroutine is an implementation detail, if you don't
call any other throwing function, it is possible to just stay a normal function
rather than become a coroutine, so throwing or returning a value can simply be achieved
by a simple return.
zpp::throwing<int> foo(bool success)
{
if (!success) {
// Throws an exception.
return std::runtime_error("My runtime error");
}
// Returns a value.
return 1337;
}As a final example, here is a full program to play with:
#include "zpp_throwing.h"
#include <string>
#include <iostream>
zpp::throwing<int> foo(bool success)
{
if (!success) {
// Throws an exception.
co_yield std::runtime_error("My runtime error");
}
// Returns a value.
co_return 1337;
}
zpp::throwing<std::string> bar(bool success)
{
return zpp::try_catch([&]() -> zpp::throwing<std::string> {
auto foo_result = co_await foo(success);
std::cout << "Foo succeeded" << foo_result << '\n';
co_return "foo succeeded";
}, [&](const std::runtime_error & error) -> zpp::throwing<std::string> {
std::cout << "Runtime error caught: " << error.what() << '\n';
co_return "foo failed";
});
}
zpp::throwing<std::string> foobar()
{
auto result = co_await bar(false);
if (result.find("foo succeeded") == std::string::npos) {
co_yield std::runtime_error("bar() apparently succeeded even though foo() failed");
}
co_return result;
}
int main()
{
return zpp::try_catch([]() -> zpp::throwing<int> {
std::cout << "Hello World\n";
std::cout << co_await foobar() << '\n';;
co_return 0;
}, [&](const std::exception & error) {
std::cout << "std exception caught: " << error.what() << '\n';
return 1;
}, [&]() {
std::cout << "Unknown exception\n";
return 1;
});
}Execute ./zpp.mk -j in the test folder. You can find the output in the out directory.
Please make sure that clang++ points to clang++-12 or above.
For more info about this build system see here.
- The code currently assumes that no exceptions can ever be thrown and as such it is not recommended to use it in a project where exceptions are enabled.
- Because coroutines cannot work with constructors, it means that an exception cannot propagate from constructors natively, and it needs to be worked around, through a parameter to the constructor or other means such as factory functions.
- The code has gone only through very minimal testing, with recent clang compiler.
- The code requires
C++20and above. - You must catch every dynamic exception that you throw, otherwise a memory leak of the exception object will occur, this is to optimize the non-trivial destruction that happens when propagating the exception.
- Assumes allocators are stateless, don't require size on deallocation, and return max aligned storage,
- for simplicity - may change in the future.
Please submit any feedback you can via issues, I would gladly accept any suggestions to improve the usage experience, performance, reference to other similar implementations, and of course bug reports.