Fix CoroFrame generation to make 8 multiples ints

Hi,

I was trying to make the coroutine-ts work in emscripten and finally managed to make it work.

Here is the program I was trying to compile:

#include <stdio.h>
#include <experimental/coroutine>
#include <iostream>

#include <cppcoro/generator.hpp>

cppcoro::generator<int> usage_example()
{
  int i = 0;
  ++i;
  printf("co %d\n", i);
  co_yield i;
  ++i;
  printf("co %d\n", i);
  co_yield i;
}

int main() {
  for (auto i : usage_example())
  {
    std::cout << "main :" << i << std::endl;
  }
  return 0;
}

I managed to compile and run this by:

upgrading the libcxx to libcxx 5.0 (since coroutines landed in that clang version)
applying the patch in this pull request

About this patch, this is the first time I deal with coroutines and LLVM IR, so this may not be the correct fix.

The issue I had was "LLVM ERROR: Stores must be a multiple of 8 bits". This comes from the fact that when we declare a function that make use of co_* keywords, a "Frame" structure is generated by the compiler. This structure is used to store all local variables, stuff about how to handle suspension points, and more importantly the current suspension point.

The size of the current suspension point uses just enough bits to hold the required values. In my example, this gives the following IR:

%_Z13usage_examplev.Frame = type { void (%_Z13usage_examplev.Frame*)*, void (%_Z13usage_examplev.Frame*)*, %"class.cppcoro::detail::generator_promise",

i2, // the suspension point

%"struct.std::experimental::coroutines_v1::suspend_always"*, %"struct.std::experimental::coroutines_v1::suspend_always"*, %"struct.std::experimental::coroutines_v1::suspend_always"*, %"class.std::experimental::coroutines_v1::coroutine_handle.0", %"class.std::experimental::coroutines_v1::coroutine_handle", i8*, %"class.std::experimental::coroutines_v1::coroutine_handle", i8*, %"struct.std::experimental::coroutines_v1::suspend_always"*, %"class.std::experimental::coroutines_v1::coroutine_handle.0", %"struct.std::experimental::coroutines_v1::suspend_always"*, %"struct.std::experimental::coroutines_v1::suspend_always"*, %"struct.std::experimental::coroutines_v1::suspend_always"*, %"struct.std::experimental::coroutines_v1::suspend_always"*, %"class.std::experimental::coroutines_v1::coroutine_handle.0", %"class.std::experimental::coroutines_v1::coroutine_handle", i8*, %"struct.std::experimental::coroutines_v1::suspend_always"*, %"struct.std::experimental::coroutines_v1::suspend_always"*, %"struct.std::experimental::coroutines_v1::suspend_always"*, %"class.std::experimental::coroutines_v1::coroutine_handle.0", %"class.std::experimental::coroutines_v1::coroutine_handle", i8*, %"struct.std::experimental::coroutines_v1::suspend_always", i8*, i32, %"class.std::experimental::coroutines_v1::coroutine_handle.0", %"class.std::experimental::coroutines_v1::coroutine_handle", i32, i32, %"struct.std::experimental::coroutines_v1::suspend_always", %"class.std::experimental::coroutines_v1::coroutine_handle.0", %"class.std::experimental::coroutines_v1::coroutine_handle", %"struct.std::experimental::coroutines_v1::suspend_always", i32, %"class.std::experimental::coroutines_v1::coroutine_handle.0", %"class.std::experimental::coroutines_v1::coroutine_handle", %"struct.std::experimental::coroutines_v1::suspend_always", %"class.std::experimental::coroutines_v1::coroutine_handle.0", %"class.std::experimental::coroutines_v1::coroutine_handle" }

As you can see, the coroutine has 2 suspension points. It seems that for some reason it needs one more state (maybe the start or the end of the function).

This somehow works for native code, even though there is no such thing as 2 bits integers. On x86_64, the store instructions seem to be compiled to movb:

     c6 40 18 01            movb   $0x1,0x18(%rax)

So the backend must somehow translate integers to align sizes to 8 bits.

In the emscripten backend, there's that assertion which just aborts compilation. I tried removing the assertion and fixing the backend but it involved updating lots of functions like getPromotedType, isLegalSize and shouldConvert. That's why I fixed the size of the field in LLVM.

Is there any reason why the field is generated as i2? Is it to support some exotic architecture? Do you think this is an acceptable fix?

emscripten-core / emscripten-fastcomp

Fix CoroFrame generation to make 8 multiples ints #226