Re: [PATCH v3 4/8] target/ppc: add vmulld instruction

[Richard Henderson]

On 6/25/20 2:13 PM, Lijun Pan wrote:
>>>     case INDEX_op_mul_vec:
>>> -        tcg_debug_assert(vece == MO_32 && have_isa_2_07);
>>> -        insn = VMULUWM;
>>> +        tcg_debug_assert((vece == MO_32 && have_isa_2_07) ||
>>> +                         (vece == MO_64 && have_isa_3_10));
>>> +        insn = mul_op[vece];
>>
>> I think it would be ok to just index mul_op here, since the real isa check is
>> to be done elsewhere.
> 
> Just keep "insn = mul_op[vece];"
> and remove"        tcg_debug_assert((vece == MO_32 && have_isa_2_07) ||
>                          (vece == MO_64 && have_isa_3_10));“?

Yes.

> @@ -3016,6 +3016,8 @@int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type,
> unsigned vece)
>              return -1;
>          case MO_32:
>              return have_isa_2_07 ? 1 : -1;
> +        case MO_64:
> +            return have_isa_3_10 ? 1 : -1;
>          }

Actually, just "return have_isa_3_10".

Returning 1 means that the opcode is supported directly.  Returning -1 means
that the opcode can be expanded by tcg_expand_vec_op.  Returning 0 means that
the tcg backend does not support the opcode at all.

> something like below?
> @@ -3712,6 +3712,11 @@static void tcg_target_init(TCGContext *s)
>          have_isa = tcg_isa_3_00;
>      }
>  #endif
> +#ifdef PPC_FEATURE2_ARCH_3_10
> +    if (hwcap2 & PPC_FEATURE2_ARCH_3_10) {
> +        have_isa = tcg_isa_3_10;
> +    }
> +#endif

Certainly this.

> @@ -554,6 +554,7 @@typedef struct {
>  #define PPC_FEATURE2_HTM_NOSC           0x01000000
>  #define PPC_FEATURE2_ARCH_3_00          0x00800000
>  #define PPC_FEATURE2_HAS_IEEE128        0x00400000
> +#define PPC_FEATURE2_ARCH_3_10          0x00200000

Of this I'm not sure.  I didn't even realize these defines were here in
include/elf.h.  For other tcg backends we get the defines from <sys/auxv.h>.

If we do want to update include/elf.h, it should be a separate patch.  CC'ing
Laurent for this.


r~