Re: [PATCH v16 73/99] target/arm: cpu-sve: split TCG and KVM functionality

[Richard Henderson]

On 6/4/21 8:52 AM, Alex Bennée wrote:
>       DECLARE_BITMAP(kvm_supported, ARM_MAX_VQ);
> -    DECLARE_BITMAP(tmp, ARM_MAX_VQ);
> -    uint32_t vq, max_vq = 0;
> -
> -    /* Collect the set of vector lengths supported by KVM. */
> -    bitmap_zero(kvm_supported, ARM_MAX_VQ);
> -    if (kvm_enabled() && kvm_arm_sve_supported()) {
> -        kvm_arm_sve_get_vls(CPU(cpu), kvm_supported);
> -    } else if (kvm_enabled()) {
> -        assert(!cpu_isar_feature(aa64_sve, cpu));
> -    }
> +    uint32_t max_vq = 0;
>   
> +    if (kvm_enabled()) {
> +        kvm_sve_get_supported_lens(cpu, kvm_supported);
> +    }

Previously, kvm_supported was always initialized.

I guess this is sort-of a cleanup.  But we've got 4 different checks for kvm 
and tcg.  I think we can actually tidy this up with a set of callbacks.

static bool do_sve_finalize(ARMCPU *cpu, Error **errp,
    void (*get_supported_lens)(ARMCPU *, unsigned long *),
    void (*enable_lens)(unsigned long *vq_map,
                        unsigned long *vq_init,
                        uint32_t max_vq,
                        unsigned long *supported),
    void (*disable_lens)(unsigned long *vq_map,
                         unsigned long *vq_init,
                         unsigned long *supported,
                         Error **errp),
    void (*validate_lens)(unsigned long *vq_map,
                          unsigned long *vq_init,
                          unsigned long *supported,
                          Error **errp, uint32_t max_vq))
{
    ...
}

bool cpu_sve_finalize_features(ARMCPU *cpu, Error **errp)
{
    if (kvm_enabled()) {
        return do_sve_finalize(cpu, errp,
                               kvm_sve_get_supported_lens,
                               kvm_sve_enable_lens,
                               kvm_sve_disable_lens,
                               kvm_sve_validate_lens);
    } else if (tcg_enabled()) {
        return do_sve_finalize(cpu, errp,
                               tcg_sve_get_supported_lens,
                               tcg_sve_enable_lens,
                               tcg_sve_disable_lens,
                               tcg_sve_validate_lens);
    } else {
        g_assert_not_reached(); /* ??? */
    }
}

with

void tcg_sve_get_supported_lens(ARMCPU *cpu,
                                unsigned long *supported)
{
    bitmap_fill(supported, ARM_MAX_VQ);
}

which we can later adjust for, e.g. -cpu a64fx and neoverse-v1, etc, which 
don't support all vq sizes.


r~