Re: porterbox access?
On Mon, Jul 08, 2013 at 01:34:25PM -0400, Camm Maguire wrote:
> Greetings! Can anyone provide temporary access to an alpha to chase
> down a quick bug on gcl?
I presume you got access to Bill's Alpha.
I noted in the gcl 2.6.7+dfsga-35 changelog that gcl is now compiled with
-O0 on Alpha. I was wondering what is the reason for that. It's just that
-O0 has a substantial impact on speed on Alpha, it being a register based
RISC arch. At -O0 all accesses to local variables are forced to be off the
stack thus involve memory accesses. Optimisation at -O1 allows local
variables to be held in CPU registers and provides a substantial speed
increase over -O0.
Building acl2 now is horrendously slow and the last build got killed by
the sbuild watchdog when it did not observe any new output on the log for
150 minutes. And that was on the fastest Alpha we have as a buildd! I
suspect it was not a real build failure as the prior version (6.2-5) did
actually build to completion but I was not fast enough to upload it before
6.2-6 came along. Also the log file never made it to debian-ports as it
exceeded the buildd site's limits on email length (even after compression).
I also noted in the kernel log reports of misaligned memory accesses
executed by gcl. At best these only represented an inefficiency as the
misaligned memory access traps to the kernel and the kernel completes the
memory access. Sometimes though they indicate real bugs in the sofware.
I have seen gcc "optimise" a misaligned memory access to generate code that
only works if the memory access is correctly aligned, presumably because
misaligned memory accesses are undefined behaviour under the C standards
which permits gcc to do whatever it damn well likes. So it might be
worthwhile exploring these misaligned memory accesses more closely.
On the other hand I would not be surprised if no one is running these
software packages on Alpha so it may not be worth putting too much time
into this.
Cheers
Michael.
Reply to: