Hi,
ilmbase failed to build properly; the test suite failed.ilmbase seems to be fairly new, so the question of whether the test suite ever worked on m68k is a bit moot.If I parse the testsuite output right, it tries to do some rounding tests based on pathological bit patterns for floats? Maybe we should try to run that test suite on a 030 or 040 machine? I'm wondering whetther what we see here is a corner case in the 040 or 060 FPSP routines rather than a toolchain bug.
Just saw it's giving two different testsuite fails for t2 and vault13. This definitely smells like something specific to the particular sort of CPU being used. Can we please have that bit pattern section of the testsuite disabled on m68k (you already disabled the whole testsuite on arm)?
I'm looking at the testsuite code and it does seem to even get the value of the second NaN bit pattern wrong when calling the test:
floatPosQNan2 ()
{
half::uif x;
x.i = 0x7fd55555;
return x.f;
}
gives (inside testBits()) a string of 0 11111111 11111111111111111111111
(for printBits (cout, f)) which is the value for floatPosQNan1(), where it
should instead have been 0 11111111 10101010101010101010101 as set above.
So that's quite odd.
A random thought: what's the size of a c++ identifier on m68k? Are floatPosQNan1() and floatPosQNan2() ambiguous, and accidentially interpreted to mean the same function?
Stephen: please give it a go on a 040. I'll run a test on hobbes (t2 being aranym prempts that plan of mine).
Cheers, Michael