Boost logo

Boost :

From: Toon Knapen (toon.knapen_at_[hidden])
Date: 2004-03-11 07:45:14

Currently the prg_exec_fail2 regression test fails on IBM VisualAge but
this depends on how you interpret it.

This regression test is supposed (by its creator) to fail at runtime
because of a detection of a division by zero. If the executable does not
detect the division by zero, the regression test is reported to fail.
AFAICT this is not a failure of the executable on IBM, this is just a
property of the executable.

The idea of the regression tests is to verify if there are no
regressions (by definition) and they are also used to verify if the
compiler and created exe react as required by the standard. Well AFAIK
nowhere is required that a division by zero is detected. It's just that
some processors detect it, others don't. But I don't think regression
tests should not be used to test 'properties' of the compiler or
executable. Because I for instance interpret a failing regression test
as being an error of the compiler or exe.

There is a way though to detect division by zero with VisualAge by
setting a specific pragma. However this incurs a runtime penalty (for
the extra error checking code) so we don't want to do this in general.
So I can assure the test to run as expected but I don't want to cripple
performance just for the sake of passing this regression test.

I can see 2 solutions:

1) Or regression tests are not used anymore to test *properties* of the
2) Or I include the pragma only for this regression test just for the
sake of passing the test.

I definitly prefer option 1 because I consider option 2 as a cheat.


Boost list run by bdawes at, gregod at, cpdaniel at, john at