With regard to the test program I mailed earlier: it produces identical results on i386 as on hppa if -O or -O2 is given to gcc, but different (and I think incorrect) results without optimization (-O0 or no option). It was pointed out to me on IRC by Simon Richter that i386 computes floating point with higher precision inside the FPU and registers than what is stored to memory, so optimization avoids truncation of intermediate results.
I'm still not capable of deducing how to fix this problem, not being well versed with floating point math. Since the errors seem to be rounding errors, and as such almost inevitable unless extreme care is taken, it might be worthwhile to punt on the issue and just disable the "make check" in debian/rules. This allows the package to build everywhere, but it doesn't solve the real problem. (I'll probably stop butting in on this bug now, since I can't seem to do anything actually useful.) -- To UNSUBSCRIBE, email to [EMAIL PROTECTED] with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]
