bndmov MPX instructions in dynamic linker breaks valgrind

[H.J. Lu]

On Thu, Aug 7, 2014 at 12:30 AM, Florian Weimer <fweimer@redhat.com> wrote:
> On 08/06/2014 11:18 PM, Roland McGrath wrote:
>>
>> Because of Intel's assurance that the new opcode values are harmless NOPs,
>> we didn't put any provision in the source for building a libc without the
>> MPX support.
>
>
> How confident are we that existing silicon is able to decode these NOPs
> efficiently?  Executing them as NOPs is just one part—if there's a
> significant performance hit, we have a problem as well.
>

Here is the test I used.  Latency on Intel processors is
from 1 to 6 cycles for

bndmov %bnd0, 56(%rsp)
bndmov %bnd1, 72(%rsp)
bndmov %bnd2, 88(%rsp)
bndmov %bnd3, 104(%rsp)

and

bndmov 104(%rsp), %bnd3
bndmov 88(%rsp), %bnd2
bndmov 72(%rsp), %bnd1
bndmov 56(%rsp), %bnd0


-- 
H.J.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: test.tar.xz
Type: application/x-xz
Size: 1308 bytes
Desc: not available
URL: <http://sourceware.org/pipermail/libc-alpha/attachments/20140807/6f0f134e/attachment.xz>