Re: [PATCH] target/arm: Do not use gen_mte_checkN in trans_STGP

[Peter Maydell]

On Thu, 27 Jul 2023 at 17:33, Richard Henderson
<richard.henderson@linaro.org> wrote:
>
> STGP writes to tag memory, it does not check it.
> This happened to work because we wrote tag memory first
> so that the check always succeeded.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
>  target/arm/tcg/translate-a64.c | 41 +++++++++++++---------------------
>  1 file changed, 15 insertions(+), 26 deletions(-)
>
> diff --git a/target/arm/tcg/translate-a64.c b/target/arm/tcg/translate-a64.c
> index 5fa1257d32..dfd18e19ca 100644
> --- a/target/arm/tcg/translate-a64.c
> +++ b/target/arm/tcg/translate-a64.c
> @@ -3020,37 +3020,17 @@ static bool trans_STGP(DisasContext *s, arg_ldstpair 
> *a)
>          tcg_gen_addi_i64(dirty_addr, dirty_addr, offset);
>      }
>
> -    if (!s->ata) {
> -        /*
> -         * TODO: We could rely on the stores below, at least for
> -         * system mode, if we arrange to add MO_ALIGN_16.
> -         */
> -        gen_helper_stg_stub(cpu_env, dirty_addr);
> -    } else if (tb_cflags(s->base.tb) & CF_PARALLEL) {
> -        gen_helper_stg_parallel(cpu_env, dirty_addr, dirty_addr);
> -    } else {
> -        gen_helper_stg(cpu_env, dirty_addr, dirty_addr);
> -    }
> -
> -    mop = finalize_memop(s, MO_64);
> -    clean_addr = gen_mte_checkN(s, dirty_addr, true, false, 2 << MO_64, mop);
> -
> +    clean_addr = clean_data_tbi(s, dirty_addr);
>      tcg_rt = cpu_reg(s, a->rt);
>      tcg_rt2 = cpu_reg(s, a->rt2);
>
>      /*
> -     * STGP is defined as two 8-byte memory operations and one tag operation.
> -     * We implement it as one single 16-byte memory operation for 
> convenience.
> -     * Rebuild mop as for STP.
> -     * TODO: The atomicity with LSE2 is stronger than required.
> -     * Need a form of MO_ATOM_WITHIN16_PAIR that never requires
> -     * 16-byte atomicity.
> +     * STGP is defined as two 8-byte memory operations, aligned to 
> TAG_GRANULE,
> +     * and one tag operation.  We implement it as one single aligned 16-byte
> +     * memory operation for convenience.  Note that the alignment ensures
> +     * MO_ATOM_IFALIGN_PAIR produces 8-byte atomicity for the memory store.
>       */
> -    mop = MO_128;
> -    if (s->align_mem) {
> -        mop |= MO_ALIGN_8;
> -    }
> -    mop = finalize_memop_pair(s, mop);
> +    mop = MO_128 | MO_ALIGN | MO_ATOM_IFALIGN_PAIR;

So here we're implicitly assuming TAG_GRANULE is 16 and
then relying on the codegen for a MO_128 | MO_ALIGN
operation to give us the alignment fault if the guest
address isn't aligned to the tag granule, right ?

Previously we also put s->be_data into the MemOp
(via finalize_memop_pair() calling finalize_memop_atom()).
Don't we still need to do that ? (We do explicitly swap
the two i64s into the i128 in different orders depending
on the be_data setting, but I think that is to handle
MO_128 | MO_BE giving a true BE 128 bit store, whereas
what we're implementing is two BE 64 bit stores.)


>      tmp = tcg_temp_new_i128();
>      if (s->be_data == MO_LE) {
> @@ -3060,6 +3040,15 @@ static bool trans_STGP(DisasContext *s, arg_ldstpair 
> *a)
>      }
>      tcg_gen_qemu_st_i128(tmp, clean_addr, get_mem_index(s), mop);

thanks
-- PMM