Simplify somewhat: this is C not FORTRAN-IV - we have structures!
Directly us 'ieee_double_shape_type' when ripping apart the fp number
  to avoid an extra store-load in 'STRICT_ASSIGN'.
Keep 'k' as the exponent, only do 'k << 20' when generating the fp number
  from it.
Fix infinity and underflow returns on i386 - because the value is returned
  in %st0 (x87 stack) we have to generate a long double error value.
The returned value for integers -1023 and below (which should be denormal
  fp valuesr) are 'just plain wrong' (tm).

(dsl)