Re: single to double conversion

[Steve Kargl <sgk at troutmask dot apl dot washington dot edu>]

On Wed, Sep 27, 2006 at 12:55:49PM -0700, Richard E Maine wrote:
> On Sep 27, 2006, at 12:35 PM, Steve Kargl wrote:
> 
> >The author of g77 made the decide to use the f2c runtime library.
> >f2c always promotes a single precision function to double
> >precision.
> 
> I don't see an obvious way to do that and still conform to the  
> Fortran standard. Perhaps I am missing a trick. Or perhaps g77 just  
> fails to conform to the Fortran standard there. Having a compiler  
> switch to enable a nonconforming behavior is one thing. But having a  
> compiler that did not even have the option to conform to the standard  
> strikes me as.... um.... problematic.
> 
> The nonconformance might not show up in most codes, but it sure isn't  
> hard to come up with a case of valid Fortran code that would fail  
> under such an implementation.  Consider
> 
>     call sub(f(x))
> 
> where f(x) returns a (single precision) real, and sub has a (single  
> precision) real dummy argument. If f(x) is implemented to return a  
> double instead of a single, then its result isn't going to be valid  
> as an actual argument for sub. But if sub is separately compiled,  
> there won't be an opportunity for the compiler to detect or fix the  
> problem.
> 
> Perhaps I'm misunderstanding what was going on here, but I'm at least  
> a little suspicious that you've just described a way in which g77  
> violates the standard (which is different from extending it).
> 

I would need to wander around in g77 code to confirm my guess,
but I suspect it does a thunk (ie there are shims).  Consider
libf2c code for tan().

This is r_tan.c

  #include "f2c.h"
  #undef abs
  #include <math.h>
  double
  r_tan (real * x)
  {
    return (tan (*x));
  }

This is d_tan.c

  #include "f2c.h"
  #undef abs
  #include <math.h>
  double
  d_tan (doublereal * x)
  {
    return (tan (*x));
  }

-- 
Steve