Re: [PATCH for-6.2 08/34] target/arm: Fix VPT advance when ECI is non-zero

[Peter Maydell]

On Fri, 16 Jul 2021 at 17:58, Richard Henderson
<richard.henderson@linaro.org> wrote:
>
> On 7/13/21 6:37 AM, Peter Maydell wrote:
> >       if (mask01 > 8) {
> > -        /* high bit set, but not 0b1000: invert the relevant half of P0 */
> > -        vpr ^= 0xff;
> > +        if (eci == ECI_NONE) {
> > +            /* high bit set, but not 0b1000: invert the relevant half of 
> > P0 */
> > +            vpr ^= 0xff;
> > +        } else if (eci == ECI_A0) {
> > +            /* Invert only the beat 1 P0 bits, as we didn't execute beat 0 
> > */
> > +            vpr ^= 0xf0;
> > +        } /* otherwise we didn't execute either beat 0 or beat 1 */
> >       }
> >       if (mask23 > 8) {
> > -        /* high bit set, but not 0b1000: invert the relevant half of P0 */
> > -        vpr ^= 0xff00;
> > +        if (eci != ECI_A0A1A2 && eci != ECI_A0A1A2B0) {
> > +            /* high bit set, but not 0b1000: invert the relevant half of 
> > P0 */
> > +            vpr ^= 0xff00;
> > +        } else {
> > +            /* We didn't execute beat 2, only invert the beat 3 P0 bits */
> > +            vpr ^= 0xf000;
> > +        }
> >       }
>
> It might not be any cleaner, but I wondered if mve_eci_mask could help here.
>
>    inv_mask = mve_eci_mask(...);
>    if (mask01 <= 8) {
>      inv_mask &= ~0xff;
>    }
>    if (mask23 <= 8) {
>      inv_mask &= ~0xff00;
>    }
>    vpr ^= inv_mask;

Yes, I think that works better (you need to capture the ECI mask at the
start of the function before we update env->condexec_bits, though).

thanks
-- PMM