On Mon, 6 May 2024 at 02:08, Richard Henderson
<richard.hender...@linaro.org> wrote:
>
> This eliminates the last uses of these neon helpers.
> Incorporate the MO_64 expanders as an option to the vector expander.
>
> Signed-off-by: Richard Henderson <richard.hender...@linaro.org>
> +/*
> + * Set @res to the correctly saturated result.
> + * Set @qc non-zero if saturation occured.
> + */
> +void gen_suqadd_bhs(TCGv_i64 res, TCGv_i64 qc,
> + TCGv_i64 a, TCGv_i64 b, MemOp esz)
> +{
> + TCGv_i64 max = tcg_constant_i64((1ull << ((8 << esz) - 1)) - 1);
> + TCGv_i64 t = tcg_temp_new_i64();
> +
> + tcg_gen_add_i64(t, a, b);
> + tcg_gen_smin_i64(res, t, max);
> + tcg_gen_xor_i64(t, t, res);
> + tcg_gen_or_i64(qc, qc, t);
> +}
Can you explain how this one should work? SUQADD is
"a is an signed value, b is an unsigned value, add them and
do a signed saturation of the result". If we take, say,
16 bit elements a = 0xc000 and b = 0x5000, then a is negative
(-16384) and b positive (20480), so the result we want is
0x1000 (4096) and no QC bit set.
But the codegen above looks to me like it will incorrectly set
the QC bit (because it's effectively treating both inputs
as unsigned).
thanks
-- PMM
