C++14 either monad and why you shouldn't use exceptions

by entelechy • 9 years ago

21 8 9 years ago

entelechy3• 9 years ago

alternatives: https://github.com/ptal/expected https://github.com/beark/ftl

daemin2• 9 years ago

This won't work well with types that need alignment. It might be better not to have the internal buffer of characters but instead have an anonymous union of both left and right types.

Also Expected is not quite the same thing as this. This Maybe and Either is like Expected but with the ability to call .then() on it to do further processing.

entelechy3• 9 years ago

interesting, why wouldn't it work well with those types?

how would an union help?

daemin2• 9 years ago

So one of the reasons is that it's following a bool variable, which size is not fixed, but are usually 1 byte. So even if the class itself was aligned the buffer within would be unaligned. The other is that the alignment of a char buffer is 1 by default.

This can actually lead to crashes in the processor when it is using aligned functions (like SSE) but the data is unaligned.

If a union is used instead it should be aligned to the biggest alignment of its members. Then if it's put at the start of the class as the first variable, it should make the class aligned to the same value and not waste space on padding.

mrout• 9 years ago

Exceptions are 2-3 orders of magnitude slower if exceptions are thrown

And free if they aren't thrown. It doesn't really matter if it's slow in the error path, as the error path is.. well.. the error path. You aren't doing useful work on the error path anyway.

Exceptions are not part of the type-system

False. 'noexcept' is absolutely part of the type system.

auto unsafe = [] { // a function that throws, sometimes we can't avoid it...

Please don't write code like this, this is unidiomatic C++.

Either<std::exception, int> e = Try<std::exception>(unsafe); // let's lift the exception into the typesystem

This is really bad style in Haskell, where you got this idea from, let alone in C++. Either is a symmetric type, it's like std::pair<a, b> (in fact it's the dual of std::pair). std::either should be to std::variant as std::pair is to std::tuple.

(Minor note: Idiomatic C++ is for types to be lowercase, see the std::exception and int.)

e.left() .map([](auto const& e) {

These 'functional maps' are unidiomatic in C++, especially for things that aren't ranges or containers.

return std::cerr << e.what() << std::endl;

I don't really understand why you would want to return std::cerr. Why map?

int result = e .leftMap([](auto) { return 42; }) // do nothing with exception and map to 42 .rightMap([](auto x) { return x * 2; }) // do further computation if value available .join() // join both sides of either

A different sense of join compared to Haskell as well? Wow that's designed to confuse.

Instead of doing this, you could just have a couple of `if` statements. There's nothing wrong with using `if` statements in C++, honestly. C++ isn't Haskell.

cjhanks• 9 years ago

I concur with all of these and I will admit to on more than one occasion implementing a `Maybe<...>` type. No matter how many silly blogs I see screaming "exceptions are evil", I have never understood the logic. Some times you're several methods deep in a recursion of abstract types and `throw GetMeTheHellOutOfHere` is really as great as it gets.

Also - injecting a whole bunch of branch statements in what is otherwise a vectorizable function is almost never a good thing. This is effectively what your `.map` function is doing.

entelechy3• 9 years ago

'noexcept' is absolutely part of the typesystem

unfortunately it is not integrated well enough:

int test() noexcept { throw 1; return 1; }

compiles fine but calls std::terminate at runtime. This is not very helpful at compile time

tempodox• 9 years ago

Still looks very clunky and nowhere near the elegance and efficiency of true FP.

C++14 either monad and why you shouldn't use exceptions

Related Stories