Reworking of the "default format" testcases

[FX Coudert]

Hi all,

Following discussion with Steve, I've reworked the testcases that  
came with my "default format" patch. They check that we can output  
reals of all kinds and read them back without any loss of precision.  
There are two kinds of problems they triger: some platforms have  
crappy printf() implementations when it comes to long doubles; also,  
some don't allow writing and reading back denormals without loss of  
precision. So, I've decided to break the tests down to allow us to  
XFAIL some features while still checking others.

As I expect that some of these may fail (or start failing) in the  
future on some platforms. If you spot these failures, please report  
them to the list. Here is a list of what they test and why they could  
fail:

   * gfortran.dg/default_format_1.f90: tests kinds 4 and 8, only  
normalized numbers; shouldn't fail
   * gfortran.dg/default_format_2.f90: tests kinds 4 and 8,  
denormalized numbers; might fail
   * gfortran.dg/default_format_denormal_1.f90: tests kinds 10 or 16,  
only normalized numbers; fails on darwin, probably shouldn't fail  
elsewhere
   * gfortran.dg/default_format_denormal_2.f90: tests kinds 10 or 16,  
denormalized numbers; might fail

   * gfortran.dg/default_format_1.inc: Included file for above tests.
   * gfortran.dg/default_format_2.inc: Included file for above tests.

Thanks,
FX