Re: [Qemu-ppc] [Qemu-devel] [PULL 16/20] target/arm: Use vector infrastructure for aa64 compares

[Richard Henderson]

On 04/05/2018 03:06 PM, Thomas Huth wrote:
> On 05.04.2018 02:54, Richard Henderson wrote:
>> On 04/05/2018 10:07 AM, Richard Henderson wrote:
>>> On 04/05/2018 02:49 AM, Emilio G. Cota wrote:
>>>> 1. grab this binary:
>>>>   http://cs.columbia.edu/~cota/qemu/nbench-aarch64
>>>> 2. run it on a PowerPC host with:
>>>>   $ aarch64-linux-user/qemu-aarch64 nbench-aarch64 -V
>>>>
>>>> Note: the "-V" (or "-v") flag is important! Without it, there's no 
>>>> segfault.
>>>
>>> How quickly?  I did not see one up until it exited for lack of NNET.DAT.
>>>
>>> I will note that I am using gcc 7.2 on gcc112 
>>> (/opt/cfarm/gcc-latest/bin/gcc).
>>>
>>> Are you using gcc 4.8.5, and is commit 74912f6dad in your tree?
>>> That commit might have made a difference...
>>
>> Bah.  I confirm that it doesn't, and that the test still fails.
>>
>> Since qemu does work on the same host when built with gcc 7.2, I can only 
>> blame
>> this failure on a compiler bug wrt gcc 4.8 on ppc64.
>>
>> I have not tracked down exactly what's going wrong, and probably won't unless
>> someone feels that it's worthwhile.  This only begs the question of whether 
>> we
>> should blacklist this compiler entirely.  Certainly it's old enough that it's
>> probably not worth fixing.
> 
> GCC 4.8 is still used in enterprise distros like RHEL7 and AFAIK also in
> SLES 12 ... so blacklisting this compiler does not sound like a good
> idea to me.

I have now tracked down this failure and it turns out not to be a compiler bug
at all; patch to follow.


r~