Re: [PATCH v5 16/36] tcg: Add tcg_gen_{non}atomic_cmpxchg_i128

[Richard Henderson]

On 1/26/23 14:45, Philippe Mathieu-Daudé wrote:
> > +        /* Compare i128 */
> > +        tcg_gen_xor_i64(t0, TCGV128_LOW(oldv), TCGV128_LOW(cmpv));
> > +        tcg_gen_xor_i64(t1, TCGV128_HIGH(oldv), TCGV128_HIGH(cmpv));
> > +        tcg_gen_or_i64(t0, t0, t1);
>
> Can we skip the OR ...
>
> > +        /* tmpv = equal ? newv : oldv */
> > +        tcg_gen_movcond_i64(TCG_COND_EQ, TCGV128_LOW(tmpv), t0, z,
> > +                            TCGV128_LOW(newv), TCGV128_LOW(oldv));
> > +        tcg_gen_movcond_i64(TCG_COND_EQ, TCGV128_HIGH(tmpv), t0, z,
>
> ... and test t1 here to partially update LO/HI? (we could remove t1 and
> just use t0 as twice 'xor + movcond' but the code is less readable).

No.

We want equality on the entire 128-bit comparison to control the conditional 
moves.


r~