[Brainstorm] May std::iostream adapt with stdexec ?

[shyeyian]

This is only a brainstorm. This paragraph talks about the possibility instead of the implemention-detail of it

Idea

As is known to all, <iostream> is a successful and widely-used abstraction of sync-io.
I'm thinking a kind of async_iostream, which provides interfaces like this (let's begin with an example like std::fstream):

auto async_stream = std::async_ofstream("file.txt"); // Derived from std::ostream.
async_stream << "Hello, world" << std::endl; // -> Returns a `std::execution::sender`

Meanwhile all algorithms on synchronous std::iostream is keeped, for example:

async_stream << "Explicit write this line to file" 
             << std::flush
             << "Lazily write this line, e.g. when cache is full or next time you flush it";
stdexec::sync_wait(async_stream);

The "operator<<" overload exactly in same way as synchronous std::iostream, like this:

std::ostream& operator<< (std::ostream& os, const student& stu) /* for both sync or async*/
{
    return os << stu.name() << stu.age() << stu.introduction(); // If stream is async then returns a sender.
}

• std::ostream (sync) itself is a sender (maybe template-specialized?), sync_wait it has not effect.
• std::this_thread::sync_wait(ofstream << "Hello") has same effect as the one without sync_wait,
• std::this_thread::sync_wait(async_ofstream << "Hello") actually returns a lazy sender (which works like std::this_thread::sync_wait(boost::asio::socket().async_send("Hello", use_sender))).

Maybe we can add some algorithms like this:

auto stream = std::ofstream("file.txt"); // Sync.
stream << std::become_async // It become a `std::async_ofstream&`
       << "data"
       << std::become_sync  // It become a `std::sync_ofstream&`
       << "HOoOoOoOoO";

Pros

• It's simple. The first lesson of C++ in many of us is like this:

#include <iostream>
using namespace std;
int main() {
    cout << "Hello, world" << endl;
}

It only make cout to something async, which does not break any grammar or expression. Everyone can easily get started to async-io, without learning anything more than their first class of C++. It is simple.

• It's clear-expressable. We can create something like this:

// Assume we have std::tcpstream, which has same interface like std::fstream.
// The client send: read ./file_1
// This server do:  (send all bytes in ./file_1 to client).
// The client send: write ./file_2
// This server do:  (write "my-file-data" into file).
std::execution::sender auto server (std::tcp_stream& stream) {
    return stream >> operation >> filename
         | std::execution::let_value([] 
             {
                 if (operation == "read")
                     return std::ranges::copy(std::views::istream(filename), std::ostream_iterator<char>(stream));
                 else if (operation == "write")
                     return std::async_fstream(filename) << "my-file-data";
             });
}

Cons

• Lifetime. Imagine this case:

std::async_ofstream << 42 << string << huge_eigen_matrix; 
// And after this line, `huge_eigen_matrix` destructs.

The possible ways are:

• std::execution, okay we std::execution::just these values. But most times we just(std::move(values...)), and here
• • implicit std::move huge_eigen_matrix might be inproper (?),
• • explicit std::move huge_eigen_matrix (like std::cout << std::move(matrix)) does not keep grammar same as sync-iostrema.
• Boost.Asio, oh we only accept boost::asio::buffer (something similar to std::span), i asio dont care about the lifetime.