Re: [Qemu-devel] [PULL 23/44] target/ppc: Use vector variable shifts for VSL, VSR, VSRA

[Laurent Vivier]

On 11/06/2019 04:43, David Gibson wrote:
> On Fri, Jun 07, 2019 at 09:28:49AM -0500, Richard Henderson wrote:
>> On 6/7/19 9:09 AM, Laurent Vivier wrote:
>>> On 07/06/2019 11:29, Laurent Vivier wrote:
>>>> On 29/05/2019 08:49, David Gibson wrote:
>>>>> From: Richard Henderson <address@hidden>
>>>>>
>>>>> The gvec expanders take care of masking the shift amount
>>>>> against the element width.
>>>>>
>>>>> Signed-off-by: Richard Henderson <address@hidden>
>>>>> Message-Id: <address@hidden>
>>>>> Signed-off-by: David Gibson <address@hidden>
>>>>> ---
>>>>>  target/ppc/helper.h                 | 12 ----------
>>>>>  target/ppc/int_helper.c             | 37 -----------------------------
>>>>>  target/ppc/translate/vmx-impl.inc.c | 24 +++++++++----------
>>>>>  3 files changed, 12 insertions(+), 61 deletions(-)
>>>>
>>>> This patch introduces a regressions
>>>>  with Fedora 29 guest:
>>>>
>>>> - during kernel boot:
>>>>
>>>> [   40.397876] crypto_register_alg 'aes' = 0
>>>> [   40.577517] crypto_register_alg 'cbc(aes)' = 0
>>>> [   40.743576] crypto_register_alg 'ctr(aes)' = 0
>>>> [   41.061379] alg: skcipher: Test 1 failed (invalid result) on encryption 
>>>> for p8_aes_xts
>>>> [   41.062054] 00000000: 91 7c f6 9e bd 68 b2 ec 9b 9f e9 a3 ea dd a6 92
>>>> [   41.062163] 00000010: 98 10 35 57 5e dc 36 1e 9a f7 bc ba 39 f2 5c eb
>>>> [   41.062834] crypto_register_alg 'xts(aes)' = 0
>>>> [   41.077358] alg: hash: Test 2 failed for p8_ghash
>>>> [   41.077553] 00000000: 5f 89 ab f7 20 57 20 57 20 57 20 57 20 57 20 57
>>>>
>>>> - with libssl:
>>>>
>>>> # curl -o /dev/null https://www.google.com
>>>>   % Total    % Received % Xferd  Average Speed   Time    Time     Time  
>>>> Current
>>>>                                  Dload  Upload   Total   Spent    Left  
>>>> Speed
>>>>   0     0    0     0    0     0      0      0 --:--:-- --:--:-- --:--:--   
>>>>   0
>>>> curl: (35) error:1408F119:SSL routines:ssl3_get_record:decryption failed 
>>>> or bad record mac
>>>>
>>>> [before, this one fails with:
>>>> curl: (35) error:04091068:rsa routines:int_rsa_verify:bad signature ]
>>>>
>>>> If I revert this patch on top of 0d74f3b427 + "target/ppc: Fix lxvw4x, 
>>>> lxvh8x and lxvb16x", all works fine.
>>>>
>>>> Thanks,
>>>> Laurent
>>>>
>>>
>>> This seems to fix the problem:
>>>
>>> diff --git a/accel/tcg/tcg-runtime-gvec.c b/accel/tcg/tcg-runtime-gvec.c
>>> index 3b6052fe97..6f0709b307 100644
>>> --- a/accel/tcg/tcg-runtime-gvec.c
>>> +++ b/accel/tcg/tcg-runtime-gvec.c
>>> @@ -874,7 +874,7 @@ void HELPER(gvec_sar8v)(void *d, void *a, void *b,
>>> uint32_t desc)
>>>      intptr_t oprsz = simd_oprsz(desc);
>>>      intptr_t i;
>>>
>>> -    for (i = 0; i < oprsz; i += sizeof(vec8)) {
>>> +    for (i = 0; i < oprsz; i += sizeof(uint8_t)) {
>>>          uint8_t sh = *(uint8_t *)(b + i) & 7;
>>>          *(int8_t *)(d + i) = *(int8_t *)(a + i) >> sh;
>>>      }
>>
>> Grr.  I really really need to come up with a solution for testing that allows
>> me to test paths that the host cpu would not ordinarily take.  This bug is
>> hidden on a host with AVX2.
>>
>> Thanks for the digging.
> 
> Can one of you send this fix formally with a S-o-b and so forth?

I'm going to send it.

Thanks,
Laurent