Currently complex types are not passed in accordance with the System V ABI when using extern(C) on linux x64, (dmd git master, no flags):

extern(C):

cfloat cf(cfloat a)
{
    return a;
}
// a is expected on the stack (not the fpu stack) and returned in XMM0:XMM1 (real in one, imaginary in the other I presume. Lots of empty space.)

cdouble cd(cdouble a)
{
    return a;
}
// a is expected in XMM0:XMM1, a hidden pointer argument to some stack space is expected in RDI, filled from a and is returned in RAX

creal cr(creal a)
{
    return a;
}
// a is expected on the stack (not fpu stack), a hidden pointer argument to some stack space is passed in RDI, filled from a and is returned in RAX.


/*********

Would it not be better to make these compliant with the equivalent types specified in the System V ABI (when using extern(C) on linux x64)? It would ease interactions with other languages.

*********/


As specified here: http://refspecs.linuxfoundation.org/elf/x86_64-abi-0.95.pdf

cfloat would be passed as real -> XMM<n> and imag -> XMM<n+1>
and returned identically

cdouble would be passed as real -> XMM<n> and imag -> XMM<n+1>
and returned identically

creal would be passed on the stack (not fpu)
and returned as real -> ST<n> and imag -> ST<n+1>


I recognise that these types are supposed to be phased out in favour of std.complex. However, seeing as that doesn't seem to be happening any time soon, getting the ABI right would be good for now.

On 4/22/2013 7:54 AM, John Colvin wrote:
> Currently complex types are not passed in accordance with the System V ABI when
> using extern(C) on linux x64, (dmd git master, no flags):

As I recall, linux gcc didn't follow the ABI itself, and I tried to match what gcc did rather than what the ABI said.

We should match what gcc does.