Re: [PATCH v2 09/12] target/arm: Implement the SETG* instructions

[Philippe Mathieu-Daudé]

Hi Peter,

(patch merged as commit 6087df574400659226861fa5ba47970f1fbd277b).

On 12/9/23 16:04, Peter Maydell wrote:
> The FEAT_MOPS SETG* instructions are very similar to the SET*
> instructions, but as well as setting memory contents they also
> set the MTE tags. They are architecturally required to operate
> on tag-granule aligned regions only.
>
> Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
> Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
> ---
> v2: - separate helper functions calling do_setp/setm/sete
>      - use cpu_st16_mmu()

So you replaced the pair of cpu_stq_mmuidx_ra() from v1 by
cpu_st16_mmu().

> ---
>   target/arm/internals.h         | 10 ++++
>   target/arm/tcg/helper-a64.h    |  3 ++
>   target/arm/tcg/a64.decode      |  5 ++
>   target/arm/tcg/helper-a64.c    | 86 ++++++++++++++++++++++++++++++++--
>   target/arm/tcg/mte_helper.c    | 40 ++++++++++++++++
>   target/arm/tcg/translate-a64.c | 20 +++++---
>   6 files changed, 155 insertions(+), 9 deletions(-)


> +/*
> + * Similar, but setting tags. The architecture requires us to do this
> + * in 16-byte chunks. SETP accesses are not tag checked; they set
> + * the tags.
> + */
> +static uint64_t set_step_tags(CPUARMState *env, uint64_t toaddr,
> +                              uint64_t setsize, uint32_t data, int memidx,
> +                              uint32_t *mtedesc, uintptr_t ra)
> +{
> +    void *mem;
> +    uint64_t cleanaddr;
> +
> +    setsize = MIN(setsize, page_limit(toaddr));
> +
> +    cleanaddr = useronly_clean_ptr(toaddr);
> +    /*
> +     * Trapless lookup: returns NULL for invalid page, I/O,
> +     * watchpoints, clean pages, etc.
> +     */
> +    mem = tlb_vaddr_to_host(env, cleanaddr, MMU_DATA_STORE, memidx);
> +
> +#ifndef CONFIG_USER_ONLY
> +    if (unlikely(!mem)) {
> +        /*
> +         * Slow-path: just do one write. This will handle the
> +         * watchpoint, invalid page, etc handling correctly.
> +         * The architecture requires that we do 16 bytes at a time,
> +         * and we know both ptr and size are 16 byte aligned.
> +         * For clean code pages, the next iteration will see
> +         * the page dirty and will use the fast path.
> +         */
> +        uint64_t repldata = data * 0x0101010101010101ULL;
> +        MemOpIdx oi16 = make_memop_idx(MO_TE | MO_128, memidx);

I'm trying to understand the MO_TE use, but I'm not seeing it in
https://developer.arm.com/documentation/ddi0602/2024-06/Base-Instructions/SETGP--SETGM--SETGE--Memory-set-with-tag-setting-
pseudo code. I also checked
https://developer.arm.com/documentation/ddi0602/2024-06/Shared-Pseudocode/aarch64-functions-mops?lang=en#impl-aarch64.MemSetBytes.4
and 
https://developer.arm.com/documentation/ddi0602/2024-06/Shared-Pseudocode/aarch64-functions-memory?lang=en#AArch64.MemSingleWrite.5

Is the following part in MemSingleWrite()?

    if !atomic && aligned && accdesc.ispair then

        bits(halfsize*8) lowhalf, highhalf;
        <highhalf, lowhalf> = value;

        memstatus = PhysMemWrite(memaddrdesc, halfsize, accdesc, lowhalf);

        memaddrdesc.paddress.address = memaddrdesc.paddress.address + 
halfsize;
        memstatus = PhysMemWrite(memaddrdesc, halfsize, accdesc, highhalf);

> +        cpu_st16_mmu(env, toaddr, int128_make128(repldata, repldata), oi16, 
> ra);
> +        mte_mops_set_tags(env, toaddr, 16, *mtedesc);
> +        return 16;
> +    }
> +#endif
> +    /* Easy case: just memset the host memory */
> +    memset(mem, data, setsize);
> +    mte_mops_set_tags(env, toaddr, setsize, *mtedesc);
> +    return setsize;
> +}

If we need to endian swap, could we use the cached hflags instead of MO_TE?

The BE_DATA bit is iset in rebuild_hflags_a64() when
arm_cpu_data_is_big_endian_a64() is true. The following diff snippet
works for me but I'm out of my comfort zone here :)

-- >8 --
         uint64_t repldata = data * 0x0101010101010101ULL;
-        MemOpIdx oi16 = make_memop_idx(MO_TE | MO_128, memidx);
+        MemOp be_data = EX_TBFLAG_ANY(env->hflags, BE_DATA) ? MO_BE : 
MO_LE;
+        MemOpIdx oi16 = make_memop_idx(be_data | MO_128, memidx);
         cpu_st16_mmu(env, toaddr, int128_make128(repldata, repldata), 
oi16, ra);
         mte_mops_set_tags(env, toaddr, 16, *mtedesc);
         return 16;
---

Thanks,

Phil.