hpux10.20 regression, egcs-1.0.3a to 19980803

[Joe Buck]

> I ran the 19980803 gcc/testsuite, using egcs-1.0.3a and
> 19980803.  The following regressions (errors reported for
> 19980803 but not for 1.0.3a) appeared:

You need to be careful in interpreting regression results in the case
of C++: for some g++ tests, egcs-1.0.3a "passes" (by not emitting
an error message, say) but then does an internal compiler error.
For example, I got three XPASSes from egcs-1.0.3a with the 19980707
testsuite, but in all three cases the compiler crashed on the input.
The XPASSes occurred because the compiler did not issue an incorrect
error message for a given line (because it had already crashed!).

It might be hard to do, but ideally runtest should mark all tests as
failing if the compiler does an ICE, or even return a separate fail
status (e.g. FATAL:).  Right now we can't immediately tell the difference
between a subtle failure and a crash.