[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: "double floor(double)"



Thomas Moor <thomas.moor@anu.edu.au> writes:

> Thank you for the hints I received on my little problem (see below)
> posted recently. The main outcome is as follows: the only machine where 
> it actually occurred was the one where I've found it: faure.debian.org. 
> On lully.debian.org, which has quite the same libraries and compiler installed, 
> it did not occur. On the machines of those who kindly supported me, the problem 
> did not occur either. So for myself, I will just ignore this problem.

faure is an EV4, lully is EV6.  The FPU is substantially different on
the later generations of Alphas.  EV6 has precise exceptions as well
as some extra instructions, and (I believe) implements more operations
in hardware.

The possibilities are either:

1) EV4 is trapping whereas EV6 is not.
2) Both are trapping, but the trap is not handled correctly on EV4.

Either way, you are triggering a bug in the kernel's floating-point
completion handler on faure, but not on lully.

What is probably more relevant is that faure is running Linux 2.2.9
and lully (as well as the machines I tested and failed to verify this
code on) is running 2.2.15 or thereabouts.  There have been several
fixes to the FP completion code since 2.2.9.

I'm pretty sure this is a bug in the older kernel.

-- 
David Huggins-Daines, Senior GNU/Linux Consultant, Linuxcare, Inc.
613.562.1239 desk, 613.223.0225 mobile
dhd@linuxcare.com, http://www.linuxcare.com/
Linuxcare. Support for the revolution.



Reply to: