How to handle long doubles in libffi-rs?

[emiltayl]

I created a pull request to enable long double on all targets as a long double type has been defined for all targets since libffi v3.4.6. However, I have some questions on how long doubles should be handled for this project, so thought it best to open an issue.

The problem is that long double is not supported by Rust on most targets. On some targets it is equivalent to a double, however it is also a f80 on x86(_64) and f128 on a variety of platforms. f128 is available in nightly, but I do not think there are any concrete plans for a f80 in Rust.

This makes it a little tricky and potentially error prone to test for long double support properly, and there are no tests using it at the moment. I am not sure what the best way to handle long double in a Rust project is.

Should long double be included in libffi-rs at all? Or should it perhaps be available only in the raw and low modules due to the lack of native Rust support for the type? I am not saying that any of these are the correct answer, perhaps it is best to keep it as it is now, but I think it is worth to consider this.

If long double is kept, should there be tests for it on targets where this is possible? This does require some extra maintenance to make sure that tests are kept up-to-date and accurately tests for long double support where possible. The alternative is to keep it as it is today, available but not tested, requiring users of libffi-rs to perform their own tests to make sure it works. Searching GitHub, there does not seem to be a lot of use of long double with libffi-rs, which is not strange since the type is not generally supported by Rust.

[arihant2math]

I think that it is a good idea to keep the long double at the raw/low levels only, in general. Long double is f64 on arm and c_double on windows. I’d prefer just to not support f80/f128 long double for now. If f128 gets stabilized (it is rather close from what I know) then we could add support for that as well. Testing would be much simpler if we keep it this way.

[arihant2math]

Specifically 80-bit long doubles only occur on intel based architectures due to their FPU size of 80-bits.

[arihant2math]

Related: rust-lang/rust-bindgen#1549