Re: [RFC 11/15] target/riscv: rvb: generalized reverse

[Richard Henderson]

On 11/18/20 12:29 AM, frank.chang@sifive.com wrote:
> +static target_ulong do_grev(target_ulong rs1,
> +                            target_ulong rs2,
> +                            const target_ulong masks[])
> +{

I think the masks should be placed here, and not passed in.
What you should pass in is "int bits".


> +    target_ulong x = rs1;
> +    int shift = 1;
> +    int i = 0;
> +
> +    while (shift < TARGET_LONG_BITS) {
> +        if (rs2 & shift) {
> +            x = do_swap(x, masks[i], shift);
> +        }
> +        shift <<= 1;
> +        ++i;

Cleaner as for loop:

    for (i = 0, shift = 1; shift < bits; i++, shift <<= 1)

> +    static const target_ulong masks[] = {
> +#ifdef TARGET_RISCV32
> +        0x55555555, 0x33333333, 0x0f0f0f0f,
> +        0x00ff00ff, 0x0000ffff,
> +#else
> +        0x5555555555555555, 0x3333333333333333,
> +        0x0f0f0f0f0f0f0f0f, 0x00ff00ff00ff00ff,
> +        0x0000ffff0000ffff, 0x00000000ffffffff,
> +#endif

You don't need to replicate every entry.

    dup_const(0x55, MO_8),
    dup_const(0x33, MO_8),
    dup_const(0x0f, MO_8),
    dup_const(0xff, MO_16),
    dup_const(0xffff, MO_32),
#ifdef TARGET_RISCV64
    UINT32_MAX
#endif

> +target_ulong HELPER(grevw)(target_ulong rs1, target_ulong rs2)
> +{
> +    static const target_ulong masks[] = {
> +        0x55555555, 0x33333333, 0x0f0f0f0f,
> +        0x00ff00ff, 0x0000ffff,
> +    };
> +
> +    return do_grev(rs1, rs2, masks);

This one is broken because do_grev iterated to TARGET_LONG_BITS == 64, and the
masks array is too small.

Fixed by passing in 32 as bits parameter to do_grev, as above.

> +static bool trans_grevi(DisasContext *ctx, arg_grevi *a)
> +{
> +    REQUIRE_EXT(ctx, RVB);
> +
> +    if (a->shamt >= TARGET_LONG_BITS) {
> +        return false;
> +    }
> +
> +    return gen_grevi(ctx, a);
> +}

While this is ok for an initial implementation, it is worth noticing the shamt
for rev8 as a special case for tcg_gen_bswap_tl.

Otherwise, this needs the same gen_shift treatment.


r~