Re: Float conversion

[Georg Bauhaus <rm.dash-bauhaus@futureapps.de>]

On 04.08.10 09:27, Stephen Leake wrote:
> Georg Bauhaus <rm.dash-bauhaus@futureapps.de> writes:
> 
>> On 03.08.10 03:07, Phil Clayton wrote:
>>
>>>> Clearly to cover all cases, you
>>>> need A < B, A = B, A > B, A < C, etc.
>>>
>>> You make it sound easy...  Generally, how do you justify that tests
>>> 'cover all cases' so giving you a 100% chance of finding an error?
>>
>> When {A, B, C} in (machine) Float, I guess "<", "=", and ">"
>> won't tell the whole story in any case? 
> 
> Why not? 
> 
> Those are the only conditions tested in the code.

Suppose the code has, for a reason, the conditional

  if A = A then
  else
  end if;

That one can be important, J.-P. Rosen recently explained why,
talking about NaN. Is this test usually present?

Then, as Dmitry said, comparing Inf needs special care.  The meaning
of predefined "=" in some test case, for some floating point type,
can depend not just on platform, but also on platform state.
Practically, GNAT has, or used to have a state switching subprogram
that would turn on 80 bit FPT when running on Windows NT, IIRC.
Can a test make sure that results of "=" or even of "<" will be
unaffected by state switches in the processor?

What if significant epsilons change as computed results grow or shrink?

Additionally, can one write a test for ARM Something and hope
it gives the same result on a PC workstation for the
"same" program?


Georg