[PATCH] Don't write beyond destination in __mempcpy_avx512_no_vzeroupper (bug 23196)

[H.J. Lu]

On Tue, May 22, 2018 at 5:17 AM, Andreas Schwab <schwab@suse.de> wrote:
> On Mai 22 2018, "H.J. Lu" <hjl.tools@gmail.com> wrote:
>
>> The modified test does't fail on Skylake server with unchanged
>> memmove-avx512-no-vzeroupper.S.   Can you modify the test
>> so that it fails with the original memmove-avx512-no-vzeroupper.S
>> on Skylake server?
>
> How to do that?  I have tested it on a Xeon Phi (whateve that means)
> where it failed without the fix.

__mempcpy_avx512_no_vzeroupper is also tested on Skylake server:

[hjl@gnu-skx-1 build-x86_64-linux]$ ./string/test-mempcpy
                        simple_mempcpy __mempcpy_avx512_no_vzeroupper
__mempcpy_avx512_unaligned __mempcpy_avx512_unaligned_erms
__mempcpy_avx_unaligned__mempcpy_avx_unaligned_erms
__mempcpy_ssse3_back __mempcpy_ssse3 __mempcpy_sse2_unaligned
__mempcpy_sse2_unaligned_erms __mempcpy_erms
[hjl@gnu-skx-1 build-x86_64-linux]$

It should also fail without the fix.

>> memmove-vec-unaligned-erms.S supports no vzeroupper:
>>
>> #ifndef VZEROUPPER
>> # if VEC_SIZE > 16
>> #  define VZEROUPPER vzeroupper
>> # else
>> #  define VZEROUPPER
>> # endif
>> #endif
>>
>> Should it be used instead?
>
> I don't understand what that has to do with the bug in question.

__mempcpy_avx512_no_vzeroupper can be implemented with
memmove-vec-unaligned-erms.S.  It doesn't have the issue in
memmove-avx512-no-vzeroupper.S.


-- 
H.J.