This topic is complex number portability, not Re: [eigen] FFT for Eigen

[Mark Borgerding <mark@xxxxxxxxxxxxxx>]

(Let me try one more time to change the email subject to reflect the 
offshoot conversation :)

Gael Guennebaud wrote:
> according to other discussions, I committed this change, so, e.g.,
>
> m1.real() = m2.imag() + m3;
>   
AFAIK this is not standard c++  and it fails on at least some older 
compilers: g++  (GCC) 3.2.3 20030502 "non-lvalue in assignment"

According to  
http://www.open-std.org/jtc1/sc22/wg21/docs/lwg-defects.html#387  it may 
not become standard either since the committee seems to favor mutator 
functions over lvalues.

The safe "standard" way to do it is something like:
   template <typename _Complex, typename _Scalar>
   void set_real( _Complex & a,  const _Scalar & b) {    a  =  
_Complex( b , a.imag() ); }
(warning: above has not been compiled)

If we tiptoe just a little off the standard path, lets' assume the 
array[2] layout, which is a fairly safe assumption, then the following 
should work:
   reinterpret_cast<_Scalar(&)[2]>(a)[0] = a_real_value
   reinterpret_cast<_Scalar(&)[2]>(a)[1] = an_imag_value
and is probably easier for the compiler to optimize correctly.


-- Mark