Spurious test in gfortran.fortran-torture/execute/intrinsic_mod_ulo.f90

[Paul Thomas]

Whilst trying to develop a faster version of the intrinsic MOD that is 
able, at the same time, to handle large ratios between the two 
arguments, I was greatly puzzled to find that, at the end of 
intrinsic_mod_ulo.f90, is a test:

   ! Check large numbers
   call real4test ((/2e34, 1.0/), (/0.0, 0.0/))

that fails for one candidate MOD and, according to all my other tests, 
should fail for the existing version.  In fact, MOD (big_number, 
divisor) only works sensible for divisor = 1.0. An other choice, gives 
results of order 10**34.

I would say, therefore, that this test is spurious because it, 
presumably, uses a check in the divide, within the MOD formula, to do 
nothing in the case that the divisor is one.  That is

       mod (x, y) = x - int (x / y) * y            Has the unfortunate 
property that when x / y exceeds the largest number that can be 
represented by the integers available, it saturates the second part of 
the expression  and the result is of order x.  There is a bit of code in 
the existingMOD to correct this but it is flawed.

Both rewrites of MOD that I am preparing do better than the existing, 
from x/y ~ 10**17 up to 10**19.  Beyond this, the fast one gives wild 
values, like the present function, whereas the "correct" one is about 
20% slower.

Which version would folk like to see?

Paul T