Re: [PATCH 16/20] target/arm: Fix sve_zip_p vs odd vector lengths

[Richard Henderson <richard.henderson@linaro.org>]

On 8/25/20 6:49 AM, Peter Maydell wrote:
> On Sat, 15 Aug 2020 at 02:32, Richard Henderson
> <richard.henderson@linaro.org> wrote:
>>
>> Wrote too much with low-half zip (zip1) with vl % 512 != 0.
>>
>> Adjust all of the x + (y << s) to x | (y << s) as a style fix.
>>
>> Reported-by: Laurent Desnogues <laurent.desnogues@gmail.com>
>> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
>> ---
>>  target/arm/sve_helper.c | 25 ++++++++++++++-----------
>>  1 file changed, 14 insertions(+), 11 deletions(-)
>>
>> diff --git a/target/arm/sve_helper.c b/target/arm/sve_helper.c
>> index fcb46f150f..b8651ae173 100644
>> --- a/target/arm/sve_helper.c
>> +++ b/target/arm/sve_helper.c
>> @@ -1870,6 +1870,7 @@ void HELPER(sve_zip_p)(void *vd, void *vn, void *vm, uint32_t pred_desc)
>>      intptr_t oprsz = extract32(pred_desc, 0, SIMD_OPRSZ_BITS) + 2;
>>      int esz = extract32(pred_desc, SIMD_DATA_SHIFT, 2);
>>      intptr_t high = extract32(pred_desc, SIMD_DATA_SHIFT + 2, 1);
>> +    int esize = 1 << esz;
>>      uint64_t *d = vd;
>>      intptr_t i;
>>
>> @@ -1882,33 +1883,35 @@ void HELPER(sve_zip_p)(void *vd, void *vn, void *vm, uint32_t pred_desc)
>>          mm = extract64(mm, high * half, half);
>>          nn = expand_bits(nn, esz);
>>          mm = expand_bits(mm, esz);
>> -        d[0] = nn + (mm << (1 << esz));
>> +        d[0] = nn | (mm << esize);
>>      } else {
>> -        ARMPredicateReg tmp_n, tmp_m;
>> +        ARMPredicateReg tmp;
>>
>>          /* We produce output faster than we consume input.
>>             Therefore we must be mindful of possible overlap.  */
>> -        if ((vn - vd) < (uintptr_t)oprsz) {
>> -            vn = memcpy(&tmp_n, vn, oprsz);
>> -        }
>> -        if ((vm - vd) < (uintptr_t)oprsz) {
>> -            vm = memcpy(&tmp_m, vm, oprsz);
>> +        if (vd == vn) {
>> +            vn = memcpy(&tmp, vn, oprsz);
>> +            if (vd == vm) {
>> +                vm = vn;
>> +            }
>> +        } else if (vd == vm) {
>> +            vm = memcpy(&tmp, vm, oprsz);
> 
> Why is it OK to only check vd==vn etc rather than checking for
> overlap the way the old code did ? The commit message doesn't
> mention this.

We only ever pass pred_full_reg_offset, so there will only ever be exact
overlap.  I can either split this out as a separate change or simply add it to
the patch description.


r~