Re: [PATCH v1 06/15] tcg/riscv: Implement vector load/store

[LIU Zhiwei]

On 2024/8/14 17:01, Richard Henderson wrote:
> On 8/13/24 21:34, LIU Zhiwei wrote:
> > @@ -827,14 +850,59 @@ static void tcg_out_ldst(TCGContext *s, 
> > RISCVInsn opc, TCGReg data,
> >   static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
> >                          TCGReg arg1, intptr_t arg2)
> >   {
> > -    RISCVInsn insn = type == TCG_TYPE_I32 ? OPC_LW : OPC_LD;
> > +    RISCVInsn insn;
> > +
> > +    if (type < TCG_TYPE_V64) {
> > +        insn = type == TCG_TYPE_I32 ? OPC_LW : OPC_LD;
> > +    } else {
> > +        tcg_debug_assert(arg >= TCG_REG_V1);
> > +        switch (prev_vece) {
> > +        case MO_8:
> > +            insn = OPC_VLE8_V;
> > +            break;
> > +        case MO_16:
> > +            insn = OPC_VLE16_V;
> > +            break;
> > +        case MO_32:
> > +            insn = OPC_VLE32_V;
> > +            break;
> > +        case MO_64:
> > +            insn = OPC_VLE64_V;
> > +            break;
> > +        default:
> > +            g_assert_not_reached();
> > +        }
> > +    }
> >       tcg_out_ldst(s, insn, arg, arg1, arg2);
>
> tcg_out_ld/st are called directly from register allocation spill/fill.
> You'll need to set vtype here, and cannot rely on this having been 
> done in tcg_out_vec_op.
OK.
> That said, with a little-endian host, the selected element size 
> doesn't matter *too* much.  A write of 8 uint16_t or a write of 2 
> uint64_t produces the same bits in memory.
>
> Therefore you can examine prev_vtype and adjust only if the vector 
> length changes.
OK.
>   But we do that -- e.g. load V256, store V256, store V128 to perform 
> a 384-bit store for AArch64 SVE when VQ=3.
>
> Is there an advantage to using the vector load/store whole register 
> insns, if the requested length is not too small?
For vector type equal or bigger than vlen in host, we will use the whole 
register instructions.
>   IIRC the NF field can be used to store multiples, but we can't store 
> half of a register with these.

I think we can still use the unit-stride instructions for them.

Thanks,
Zhiwei

> r~