Re: [qemu-s390x] [Qemu-devel] [PATCH v1 28/41] s390x/tcg: Implement VECTOR ELEMENT ROTATE AND INSERT UNDER MASK

[Richard Henderson]

On 4/11/19 12:08 AM, David Hildenbrand wrote:
> +static void gen_rim_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b, int32_t c)
> +{
> +    TCGv_i32 t0 = tcg_temp_new_i32();
> +    TCGv_i32 t1 = tcg_temp_new_i32();
> +
> +    tcg_gen_andc_i32(t0, a, b);
> +    tcg_gen_rotli_i32(t1, a, c & 31);
> +    tcg_gen_and_i32(t1, t1, b);
> +    tcg_gen_or_i32(d, t0, t1);

The ANDC and ROTL look to be in the wrong order.

"For each bit in the third operand (b) that is one,
the corresponding bit *of the rotated elements* in
the second operand replaces the corresponding bit in
the first operand".

I think you need

    tcg_gen_rotli_i32(a, a, c & 31);
    tcg_gen_and_i32(a, a, b);
    tcg_gen_andc_i32(d, d, b);
    tcg_gen_or_i32(d, d, a);

with

  { .fni4 = gen_rim_32, .load_dest = true },

> +     const uint##BITS##_t a = s390_vec_read_element##BITS(v2, i);           \
> +     const uint##BITS##_t mask = s390_vec_read_element##BITS(v3, i);        \
> +     const uint##BITS##_t d = (a & ~mask) | (rotl##BITS(a, count) & mask);  \

Again, this seems to be missing the insert into "the first operand", i.e.
loading from v1 as well.


r~