Ed wrote:
> "Krishna Myneni" <krishnamyneni@[EMAIL PROTECTED]
> wrote in message
news:T42xj.80484$vt2.52096@[EMAIL PROTECTED]
>> Ed wrote:
>>> ...
>>>
>>> Which suggests that test suites are mostly a "feel good" utility
>>> i.e. one feels good when it p***** but otherwise has little value.
>>> It impresses the customer but that's about all.
>>>
>>> Who would find Hayes useful? Probably not someone who has
>>> brought up a forth. ...
>> Are you considering only testing of Forth systems? What about using the
test
>> suite to test applications? For instance, the extended version of the
Hayes
>> tester, incor****ating extensions by David Williams and Anton Ertl, have
proved
>> valuable in finding problems with the Forth Scientific Library modules.
These
>> problems went unnoticed for many years. The findings are do***ented in
an
>> earlier thread on automated FSL tests.
>
> Bugs go by unnoticed because the program (or parts of it) haven't been
> exercised. It's hard to believe that a well-used program would not show
> up its errors relatively quickly. It may be that FSL modules just don't
> get used.
>
The FSL modules probably don't get used that often. However, when dealing
with
numerical algorithms some kinds of errors can't easily be detected in
normal
use. One such example is a lower accuracy in the result than specified by
the
design. Validating the accuracy requires comparing against known test
cases, e.g.
1e-12 rel-near F!
1e-12 abs-near F!
set-near
CR
TESTING LOGISTIC
t{ -1e 1e 0e logistic -> 0.268941421370e r}t
Accuracy is usually also a function of the input parameters, so validation
across a range of inputs is needed. Any numerical library that wants to be
taken
seriously requires an accompanying test suite, because it's too much work
on the
part of the user to verify for him/herself.
> My background is electronics instrumentation. These days most have
> built-in self-diagnostics. They're there to provide the user with a
quick
> way to determine whether the instrument is basically functional. They
> don't, nor can they, determine whether a fault exits.
>
Well, it depends partly on how expensive the instrument is. I've seen a
mid-range Tektronix scope identify a specific problem with its timebase
during
the diagnostics on power up.
> I have nothing against the Hayes suite. If it shows up a problem then
> it's better than if the error remained unfound.
>
>
>
Many types of applications don't need test code, but I think it is a must
for
libraries designed for ****table use. I notice that the Forth Foundation
Library
(FFL) provides its own test code. It's a shame that the Hayes tester was
found
not to be suitable for the FFL. IMO, a commonly adopted test harness would
be a
significant step forward.
Regards,
Krishna
|
