On Sun, 26 Dec 2010, David T. Lewis wrote:

> On Sun, Dec 26, 2010 at 07:36:44PM +0100, Andreas Raab wrote:
>> Hi David -
>>
>> Can you try to compile an external FloatMathPlugin with CFLAGS=-O0 for
>> testing? I've found that some optimizations are problematic for the
>> FloatMathPlugin. On Windows, I've been compiling everything with -O0 for
>> this very reason and I wouldn't be surprised if that fixes it on Unix as
>> well.
>>
>> Cheers,
>>   - Andreas
>
> Exactly so, I was just about to report back with the same conclusion.
>
> I built a VM with sources from http://squeakvm.org/unix/release/Squeak-4.0.3.2202-src.tar.gz
> and built it with '--CFLAGS='-m32 -g'.
>
>  Result: Tests pass, no VM crash.
>
> I then built a VM from the same source but with '--CFLAGS='-m32 -O2 -g' (i.e. compiler
> optimization turned back on.

IIRC FloatMathPlugin is compiled with -O1 no matter what you specify in
CFLAGS, but I may be wrong.


Levente

>
>  Result: VM CRASH.
>
> Going back to my more up to date VM (4.4.2-2329) I notice that I had compiled it
> with optimizer off, so I rebuilt it with -O2 optimization.
>
>  Result: VM CRASH.
>
> Conclusion:
>
> The FloatMathPlugin will crash the VM on some platforms when compiled with
> gcc -O2 optimization (which would be the normal case).
>
> Given the large installed base of VMs that are likely to have this concern,
> it may be better to save the FloatMathPlugin fixes for the update stream
> rather than including them in the official 4.2 release.
>
> Note, I did not test any Cog VMs but I assume we will see the same issue on Cog.
> I don't seem to have the FloatMathPlugin in the Cog VM on my machine, so perhaps
> someone else can check this to verify.
>
> Dave
>
>> On 12/26/2010 7:17 PM, David T. Lewis wrote:
>>> Some further detail:
>>>
>>> The test crashes the 4.0.3-2202 VM in iteration 21 of #testCos. Failure
>>> occurs in the #cos send. The apparently identical call does not fail in a
>>> workspace.
>>>
>>> The following workspace doit sets up the data exactly as it appears
>>> within #testCos at iteration 21 immediately prior to the failure. When
>>> running the test case, the crash occurs at the #cos call. When evaluated
>>> in a workspace, it does not crash.
>>>
>>>   bytes := #[193 185 104 160 246 115 176 146].
>>>   float := Float basicNew: 2.
>>>   float basicAt: 1 put: (bytes unsignedLongAt: 1 bigEndian: true).
>>>   float basicAt: 2 put: (bytes unsignedLongAt: 5 bigEndian: true).
>>>   float cos ==>  1.0
>>>
>>> According to gdb, the crash is happening here:
>>>
>>> Program received signal SIGSEGV, Segmentation fault.
>>> 0x08097cc4 in __kernel_rem_pio2 (x=0x0, y=0x0, e0=0, nx=0, prec=0,
>>> ipio2=0x0)
>>>     at
>>>     /home/piumarta/squeak/platforms/Cross/plugins/FloatMathPlugin/fdlibm/k_rem_pio2.c:246
>>> (gdb) bt
>>> #0  0x08097cc4 in __kernel_rem_pio2 (x=0x0, y=0x0, e0=0, nx=0, prec=0,
>>> ipio2=0x0)
>>>     at
>>>     /home/piumarta/squeak/platforms/Cross/plugins/FloatMathPlugin/fdlibm/k_rem_pio2.c:246
>>> #1  0x00000000 in ?? ()
>>> (gdb)
>>>
>>>
>>> I have not tried rebuilding this version of the VM from source, although
>>> that
>>> might be a good next step.
>>>
>>> Dave
>>>
>>>
>>>
>>
>
>