Re: [PATCH 4/7] x86: Add SSSE3 optimized chacha20

[Adhemerval Zanella <adhemerval.zanella@linaro.org>]

On 14/04/2022 14:22, Noah Goldstein wrote:
> On Thu, Apr 14, 2022 at 12:19 PM Adhemerval Zanella
> <adhemerval.zanella@linaro.org> wrote:
>>
>>
>>
>> On 14/04/2022 14:10, Noah Goldstein wrote:
>>> On Thu, Apr 14, 2022 at 12:03 PM Adhemerval Zanella
>>> <adhemerval.zanella@linaro.org> wrote:
>>>>
>>>>
>>>>
>>>> On 13/04/2022 20:12, Noah Goldstein wrote:
>>>>> On Wed, Apr 13, 2022 at 1:27 PM Adhemerval Zanella via Libc-alpha
>>>>> <libc-alpha@sourceware.org> wrote:
>>>>>>
>>>>>> +
>>>>>> +       /* eax zeroed by round loop. */
>>>>>> +       leave;
>>>>>> +       cfi_adjust_cfa_offset(-8)
>>>>>> +       cfi_def_cfa_register(%rsp);
>>>>>> +       ret;
>>>>>> +       int3;
>>>>> why int3?
>>>>
>>>> It was originally added on libgcrypt by 11ade08efbfbc36dbf3571f1026946269950bc40,
>>>> as a straight-line speculation hardening.  It is was is emitted by clang 14 and
>>>> gcc 12 with -mharden-sls=return.
>>>>
>>>> I am not sure if we need that kind of hardening, but I would prefer to the first
>>>> version be in sync with libgcrypt as much as possible so the future optimizations
>>>> would be simpler to keep localized to glibc (if libgcrypt does not want to
>>>> backport it).
>>>
>>> Okay, can keep for now. Any thoughts on changing it to sse2?
>>>
>>
>> No strong feeling, I used the ssse3 one because it is readily available from
>> libgcrypt.
> 
> I think the only ssse3 is `pshufb` so you can just replace the optimized
> rotates with the shift rotates and that will make it sse2 (unless I'm missing
> an instruction).

Right, do you have a patch for it? I can add it on the v2 I am will send.

> 
> Also can you add the proper .text section here as well (or .sse2 or .ssse3)
Ack.