Using the real target "real" type (cross compiling)

[Kai Nacke]

Hi Dan!

On Monday, 10 February 2014 at 16:55:47 UTC, Dan Olson wrote:
> I switched to using ldc merge-2.064 branch for all my iOS 
> hacking. The
> newer std.math has some nice static asserts that failed because 
> dmd is
> using the compile host "long double" type for determing the 
> properties
> of type real.  Thank you for static assertions!
>
> Mac x86_64 long double != Arm long double
>
> I hacked dmd to make longdouble = double so I could cross 
> compile this
> newer phobos to iOS.  But are there other things like real 
> types that
> dmd code is learning from host, but should be told what the 
> target is
> using?
>
> Anyway, this seems like more of a dmd thing.  What is the best 
> approach
> to solve?  I am ok for now with hacking on my own branch but 
> wishing for
> a better solution.
>
> Thanks,

I know this annoying issue. It also prevents cross compiling from 
x86_64 to ppc64 because 80bit floats are not 128bit floats. It's 
issue #259 (https://github.com/ldc-developers/ldc/issues/259).

I don't have a workaround. But I am working on a solution. :-)

As discussed in the issue, I replace the real_t data type with a 
wrapper around llvm::APFloat. Then it is possible to configure 
the float properties based on the target architecture.

I have a first prototype already implemented. With the 
refactorings of dmd 2.065 (see my other post from today) it is 
much easier to integrate: it is now more obvious how to disable 
some built-in functions. (E.g. I have an implementation of sqrt() 
which still has a 1-bit difference compared to sqrtl() on x86_64. 
IMHO it is better to disable CTFE for these functions then to 
introduce some subtle differences between CTFE and runtime.)

I'll post it here when it is integrated.

Regards,
Kai