Re: [PATCH 4/8] target/riscv: 128-bit arithmetic and logic instructions

["Frédéric Pétrot" <frederic.petrot@univ-grenoble-alpes.fr>]

Hello Philippe,

Le 30/08/2021 à 23:40, Philippe Mathieu-Daudé a écrit :
> On 8/30/21 11:38 PM, Philippe Mathieu-Daudé wrote:
>> On 8/30/21 7:16 PM, Frédéric Pétrot wrote:
>>> Adding the support for the 128-bit arithmetic and logic instructions.
>>> Remember that all (i) instructions are now acting on 128-bit registers, that
>>> a few others are added to cope with values that are held on 64 bits within
>>> the 128-bit registers, and that the ones that cope with values on 32-bit
>>> must also be modified for proper sign extension.
>>> Most algorithms taken from Hackers' delight.
>>>
>>> Signed-off-by: Frédéric Pétrot <frederic.petrot@univ-grenoble-alpes.fr>
>>> Co-authored-by: Fabien Portas <fabien.portas@grenoble-inp.org>
>>> ---
>>>  target/riscv/insn32.decode              |  13 +
>>>  target/riscv/insn_trans/trans_rvi.c.inc | 955 +++++++++++++++++++++++-
>>>  target/riscv/translate.c                |  25 +
>>>  3 files changed, 976 insertions(+), 17 deletions(-)
>>
>>> diff --git a/target/riscv/insn_trans/trans_rvi.c.inc b/target/riscv/insn_trans/trans_rvi.c.inc
>>> index 772330a766..0401ba3d69 100644
>>> --- a/target/riscv/insn_trans/trans_rvi.c.inc
>>> +++ b/target/riscv/insn_trans/trans_rvi.c.inc
>>> @@ -26,14 +26,20 @@ static bool trans_illegal(DisasContext *ctx, arg_empty *a)
>>>  
>>>  static bool trans_c64_illegal(DisasContext *ctx, arg_empty *a)
>>>  {
>>> -     REQUIRE_64BIT(ctx);
>>> -     return trans_illegal(ctx, a);
>>> +    REQUIRE_64_OR_128BIT(ctx);
>>> +    return trans_illegal(ctx, a);
>>>  }
>>>  
>>>  static bool trans_lui(DisasContext *ctx, arg_lui *a)
>>>  {
>>>      if (a->rd != 0) {
>>>          tcg_gen_movi_tl(cpu_gpr[a->rd], a->imm);
>>> +#if defined(TARGET_RISCV128)
>>> +        if (is_128bit(ctx)) {
>>
>> Maybe this could allow the compiler eventually elide the
>> code and avoid superfluous #ifdef'ry:
>>
>>            if (TARGET_LONG_BITS >= 128) {
> 
> Actually:
> 
>              if (TARGET_LONG_BITS >= 128 && is_128bit(ctx)) {

  We may have taken a wrong path then, because we have kept
  TARGET_LONG_BITS == 64 for the 128-bit case (as we use the tcg_xxx_tl of the
  64 version to generate our micro-ops, which I admit might be a mistake).

  Frédéric

> 
>>
>>> +            tcg_gen_ext_i64_i128(cpu_gpr[a->rd], cpu_gprh[a->rd],
>>> +                                 cpu_gpr[a->rd]);
>>> +        }
>>> +#endif
>>>      }
>>>      return true;
>>>  }
>>

-- 
+---------------------------------------------------------------------------+
| Frédéric Pétrot, Pr. Grenoble INP-Ensimag/TIMA,   Ensimag deputy director |
| Mob/Pho: +33 6 74 57 99 65/+33 4 76 57 48 70      Ad augusta  per angusta |
| http://tima.univ-grenoble-alpes.fr frederic.petrot@univ-grenoble-alpes.fr |
+---------------------------------------------------------------------------+