Re: Smoking gun: NetBSD 6.0.1 userland instability

[Donald Lee <MacPPC2%c.icompute.com@localhost>]

At 9:25 PM -0500 3/7/13, Michael wrote:
>-----BEGIN PGP SIGNED MESSAGE-----
>Hash: SHA1
>
>Hello,
>
>On Mar 7, 2013, at 1:59 PM, Donald Lee wrote:
>
>>>>> In the old days, I would just use "-g", and presume that the
>>>>> optimization was disabled, but modern compilers don't do that any
>>>>> more.
>>>> The only thing -g should do is enable debugging symbols; it doesn't
>>>> generally have an effect on optimization.  Perhaps you mean -O0?
>>>
>>> No, I think more likely -dgl- was talking about "the old days", when,
>>> yes, -g did indeed kill the optimizer.  It's been a while, but there
>>> was such a time.
>>
>> ...back when the dinosaurs roamed the earth, and computers had front panels
>> with switches and lights.
>>
>> Thus sayeth the Old Guys. ;->
>>
>> If anyone can give me some confidence in what switches I need to use to
>> get what we need, I can do the experiment(s).
>
>- -mno-altivec or -mcpu=750 ( or anything else that's not a g4 or g5 )
>That doesn't keep libc from running altivec code though.

Interesting data point: I built analog with "-O0" compiler option.  (Lord
knows exactly what that generates as far as code.  It also does not
change the libraries, etc.

If the "-O0" analog is run on a 6.0.1 system, it does *NOT* run cleanly, but
it *DOES* run differently.

The way I run the test is to do an analog analysis on the same input data
N times (N=30 in this case) and then compare the output "index.html"
file.  If correct, they all differ only in the timestamp text of when
they were run.  The "bad" runs with the normal analog provide crazy
numbers, especially in totals.  Exabytes rather than Gigabytes, or
huge negative percent numbers.

The "-O0" analog only has small differences in output, and they are *mostly*
"off-by-one" errors.  There are some other differences, too, but I have not
yet examined them carefully to see whether they are OK.  Analog does
have some time-sensitive code checking stats for the "most recent week",
so the run is sensitive to the actual time it was run.

Nevertheless,
the bottom line is that the "-O0" definitely made a change in behavior.

-dgl-