snprintf in write_float.def

[Tobias Schlüter]

FX wrote:
> While reducing PR36200, I was looking at libgfortran/io/write_float.def, 
> and after having refreshed my memory about printf formats, it seems to 
> me that what we're doing in DTOA and DTOAL is unsafe: the "e" format 
> expects an argument of type double, and yet when used for real(kind=4), 
> we'll happily pass it a "float". So, am I overlooking something or 
> should we apply the patch below (which casts values to "double" or "long 
> double")? I suspect there is more to it that I saw, because we'd see a 
> large number of failures otherwise. So, comments welcome.

floats passed as varargs are automatically casted to double.  E.g.:
$ cat t.c
void a(int i, ...);

void b() { float f = 1.; a(1,f); }
$ gcc -c t.c -fdump-tree-generic
$ cat t.c.t03.generic
b ()
{
   double D.1463;
   float f;

   f = 1.0e+0;
   D.1463 = (double) f;
   a (1, D.1463);
}


$

Regards,
- Tobi