XRay-aarch64-linux::patching-unpatching.cc test flaky

[Serge Rogatch via llvm-commits]

Hi,

After investigating the problem more, we have a guess that maybe AArch64
CPU doesn't see the code changes which XRay has done, because the old
instructions stay in the instruction cache, even though code modification
has been performed in the same thread as the one running the instructions
later.

Is this possible that CPU cache stays so incoherent with the data cache?

Cheers,
Serge

On 8 December 2016 at 01:48, Dean Michael Berris <dean.berris at gmail.com>
wrote:

> Thanks Serge!
>
> Sent from my iPhone
>
> On 8 Dec. 2016, at 9:12 am, x y <serge.rogatch at gmail.com> wrote:
>
> The test has been disabled for AArch64 in https://reviews.llvm.org/
> rL288988 .
>
> Cheers,
> Serge
>
> On 5 December 2016 at 20:36, Renato Golin <renato.golin at linaro.org> wrote:
>
>> On 5 December 2016 at 17:19, x y <serge.rogatch at gmail.com> wrote:
>> > I've tried what I could troubleshooting the issue. My AArch64 XRay code
>> is
>> > quite short, so the only issue I can spot now happens when __arm__ and
>> > __aarch64__ are defined simultaneously. Shall I submit a fix for better
>> > handling this case, or should it really never happen?
>>
>> Hum, __arm__ should not be defined on AArch64, but I remember having a
>> similar problem in compiler-rt.
>>
>> Can you use #elsif defined(__aarch64__) ?
>>
>>
>> > I'm going to disable the test with the approach suggested by Renato.
>> Sorry
>> > it took so long, I tried to fix the issue first. And of course, I cannot
>> > reproduce it locally.
>>
>> Ok, LGTM. Thanks!
>>
>> --renato
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20161219/08f04e39/attachment.html>