On Fri, Oct 16, 2015 at 3:08 PM, Rainer Jung <rainer.j...@kippdata.de> wrote:
> Am 16.10.2015 um 14:56 schrieb Yann Ylavic:
>>
>> On Fri, Oct 16, 2015 at 2:48 PM, Yann Ylavic <ylavic....@gmail.com> wrote:
>>>
>>> On Fri, Oct 16, 2015 at 2:35 PM, Rainer Jung <rainer.j...@kippdata.de>
>>> wrote:
>>>>
>>>>
>>>> I didn't yet have the time to reproduce and test your patch, but the
>>>> APR_ALIGN((size),sizeof(void *)) align approach would not work. The
>>>> problem
>>>> here is that even or especially when building for 32 Bits and then using
>>>> a
>>>> 64 Bit data type, the alignment fails. So the pointer size would be 4
>>>> here,
>>>> but the requirement alignment 8 bytes.
>>>
>>>
>>> Hm, correct, each integral type (greater than the word) needs be
>>> aligned on its size!
>>
>>
>> Though the issue here is not accessing the type as an integral, so
>> aligning on 4 bytes (for 32 bit hardware) should enough for accessing
>> each scoreborad entry/struct (the members of the structs themselves
>> will still be aligned according to their type, so they can also be
>> safely accessed).
>
>
> Wasn't the bus error occuring in
>
> ws->last_used = apr_time_now();
>
> and the address is
>
> (dbx) print &(ws->last_used)
> &ws->last_used = 0xffffffff7bb00094
>
> not divisible by 8 although the data type (not pointer size) is 64 Bit.
Yes but ws itself isn't aligned either:
(dbx) print ws
ws = 0xffffffff7bb00044
which is IMHO the issue.
Align ws and everything goes well (at least I think :p ).
Regards,
Yann.
