Re: [Qemu-devel] [PATCH v10.5 15/20] target/arm: Use vector infrastructure for aa64 constant shifts

[Richard Henderson]

On 01/25/2018 09:03 AM, Peter Maydell wrote:
> On 17 January 2018 at 16:14, Richard Henderson
> <address@hidden> wrote:
>> Signed-off-by: Richard Henderson <address@hidden>
>> ---
>>  target/arm/translate-a64.c | 386 
>> ++++++++++++++++++++++++++++++++++++++-------
>>  1 file changed, 329 insertions(+), 57 deletions(-)
>>
>> diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
>> index 2495414603..1b5005637d 100644
>> --- a/target/arm/translate-a64.c
>> +++ b/target/arm/translate-a64.c
>> @@ -6489,17 +6489,6 @@ static void handle_shri_with_rndacc(TCGv_i64 tcg_res, 
>> TCGv_i64 tcg_src,
>>      }
>>  }
>>
>> -/* Common SHL/SLI - Shift left with an optional insert */
>> -static void handle_shli_with_ins(TCGv_i64 tcg_res, TCGv_i64 tcg_src,
>> -                                 bool insert, int shift)
>> -{
>> -    if (insert) { /* SLI */
>> -        tcg_gen_deposit_i64(tcg_res, tcg_res, tcg_src, shift, 64 - shift);
>> -    } else { /* SHL */
>> -        tcg_gen_shli_i64(tcg_res, tcg_src, shift);
>> -    }
>> -}
>> -
>>  /* SRI: shift right with insert */
>>  static void handle_shri_with_ins(TCGv_i64 tcg_res, TCGv_i64 tcg_src,
>>                                   int size, int shift)
>> @@ -6603,7 +6592,11 @@ static void handle_scalar_simd_shli(DisasContext *s, 
>> bool insert,
>>      tcg_rn = read_fp_dreg(s, rn);
>>      tcg_rd = insert ? read_fp_dreg(s, rd) : tcg_temp_new_i64();
>>
>> -    handle_shli_with_ins(tcg_rd, tcg_rn, insert, shift);
>> +    if (insert) {
>> +        tcg_gen_deposit_i64(tcg_rd, tcg_rd, tcg_rn, shift, 64 - shift);
>> +    } else {
>> +        tcg_gen_shli_i64(tcg_rd, tcg_rn, shift);
>> +    }
> 
> It looks like you're folding handle_shli_with_ins() into its
> now only callsite, but handle_shri_with_ins() has been left as
> its own function?

I didn't notice that.  I'll have a look.


>> +static void gen_shr8_ins_i64(TCGv_i64 d, TCGv_i64 a, int64_t shift)
>> +{
>> +    uint64_t mask = (0xff >> shift) * (-1ull / 0xff);
>> +    TCGv_i64 t = tcg_temp_new_i64();
>> +
>> +    tcg_gen_shri_i64(t, a, shift);
>> +    tcg_gen_andi_i64(t, t, mask);
>> +    tcg_gen_andi_i64(d, d, ~mask);
>> +    tcg_gen_or_i64(d, d, t);
>> +    tcg_temp_free_i64(t);
> 
> The previous code was able to work with just shifts and deposits --
> why do we need to open-code this kind of mask-and-or now? Is this
> because we now operate an i64 at a time when we used to operate
> on smaller quantities at once?

Yes, exactly.  It's now 4 total operations instead of 16.

I should also tidy this to use a new dup_const function that's been introduced
since I first wrote this code...


r~