Re: -fno-strict-aliasing

From: Robert Dinse (nanook@eskimo.com)
Date: Tue Apr 18 2000 - 20:09:53 EST


On Tue, 18 Apr 2000, Giacomo Catenazzi wrote:
>
> > I've had these Sparc spin-lock from hell problems on all 2.2.x
> > kernels running on SMP SS-10's or SS-20's with Ross Hypersparc CPU's.
> > I read in a newsgroup recently someone else having the same problem
> > and removing the -no-strict-aliasing option helped them. But I built
> > a kernel without it and the kernels ended up exactly the same size so
> > I don't know if it actually made any difference in the generated code
> > or not.
>
> What compiler do you use?

     With 2.2.14, I've used egcs-2.91.66, egcs-2.90.27, and several 2.95
versions, I think 2.95.1 and 2.95.2, and whatever came with RedHat, but only
the former two do I still have on my system at present. The egcs-2.90.27
compiled kernel runs fine on a 4/670MP platform but spinlocks on SS-10 or SS-20
even though the exact SAME (as in physically removed from one box and placed in
the other) CPU's are used, same amount of RAM also (384MB).

     I've tried compiling with just -O1 optimization, lasted about 30 hours
before a spinlock. Strangely, if I try to compile with -O0 (no optimization)
it explodes during the loading phase, something about bitfix or some such. I'd
have to attempt it again to give the exact error message.

     There is something architecture dependent because it doesn't do it on the
4/670MP, but the SS-10 will run SunOS with 100 users doing various things for
on average about a month between crashes (and then more often than not it is
denial of service attacks eating up all mbufs), where Linux will only last on
average 1-2 days between spin_lock deadlock, and I've had the same behavior
under Linux on three different boxes, so I don't think it is defective hardware
that just happens to only affect Linux.

     So far, the only 2.3.x kernels I've been successful at getting to build
wouldn't even boot on these machines, or on the Ultra-1, or on a PC I have here
for that matter. Probably some operator error but I haven't figured out what
yet, it does the twirling cursor bit and then locks.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/



This archive was generated by hypermail 2b29 : Sun Apr 23 2000 - 21:00:14 EST