Re: [Qemu-devel] [PATCH v3] tcg/ppc: Improve unaligned load/store handling on 64-bit backend

[Benjamin Herrenschmidt <benh@kernel.crashing.org>]

On Tue, 2015-07-21 at 07:27 +0100, Richard Henderson wrote:
> On 07/21/2015 06:19 AM, Benjamin Herrenschmidt wrote:
> > +    /* Clear the non-page, non-alignment bits from the address */
> >       if (TCG_TARGET_REG_BITS == 32 || TARGET_LONG_BITS == 32) {
> > +        /* We don't support unaligned accesses on 32-bits, preserve
> > +         * the bottom bits and thus trigger a comparison failure on
> > +         * unaligned accesses
> > +         */
> >           tcg_out_rlw(s, RLWINM, TCG_REG_R0, addrlo, 0,
> >                       (32 - s_bits) & 31, 31 - TARGET_PAGE_BITS);
> 
> Why don't you support this unaligned acess with 32-bit guests?

No reason, I just didn't get to do it yet. It's possible, I was just
lazy :-) It also adds one instruction. On 64-bit we always have 2
instructions anyway so it wasn't adding any overhead really, on 32-bit
we get away with a single rlwinm, while adding the unaligned support
would make it an addi + rlwinm.

> > -    } else if (!s_bits) {
> > -        tcg_out_rld(s, RLDICR, TCG_REG_R0, addrlo,
> > -                    0, 63 - TARGET_PAGE_BITS);
> > +    } else if (s_bits) {
> > +        /* > byte access, we need to handle alignment */
> > +        if ((opc & MO_AMASK) == MO_ALIGN) {
> > +            /* Alignment required by the front-end, same as 32-bits */
> > +            tcg_out_rld(s, RLDICL, TCG_REG_R0, addrlo,
> > +                        64 - TARGET_PAGE_BITS, TARGET_PAGE_BITS - s_bits);
> > +            tcg_out_rld(s, RLDICL, TCG_REG_R0, TCG_REG_R0, TARGET_PAGE_BITS, 0);
> > +       } else {
> > +           /* We support unaligned accesses, we need to make sure we fail
> > +            * if we cross a page boundary. The trick is to add the
> > +            * access_size-1 to the address before masking the low bits.
> > +            * That will make the address overflow to the next page if we
> > +            * cross a page boundary which will then force a mismatch of
> > +            * the TLB compare since the next page cannot possibly be in
> > +            * the same TLB index.
> > +            */
> > +            tcg_out32(s, ADDI | TAI(TCG_REG_R0, addrlo, (1 << s_bits) - 1));
> > +            tcg_out_rld(s, RLDICR, TCG_REG_R0, TCG_REG_R0,
> > +                        0, 63 - TARGET_PAGE_BITS);
> > +        }
> >       } else {
> > -        tcg_out_rld(s, RLDICL, TCG_REG_R0, addrlo,
> > -                    64 - TARGET_PAGE_BITS, TARGET_PAGE_BITS - s_bits);
> > -        tcg_out_rld(s, RLDICL, TCG_REG_R0, TCG_REG_R0, TARGET_PAGE_BITS, 0);
> > +        /* Byte access, just chop off the bits below the page index */
> > +        tcg_out_rld(s, RLDICR, TCG_REG_R0, addrlo, 0, 63 - TARGET_PAGE_BITS);