On Tue, 27 Jan 2026 at 15:48, Peter Maydell <[email protected]> wrote:
>
> Currently smcr_write() implicitly assumes that SME implies SVE,
> because it will call sve_vqm1_sve_for_el() when SMCR.SM is 0, and
> sve_vqm1_sve_for_el() will assert in that situation.
>
> This is the only place where we call that function without
> it being guarded by a check on whether SVE is implemented.
> Adjust smcr_write() so that it also avoids asking for the
> SVE vector length when SVE is not implemented.
>
> Signed-off-by: Peter Maydell <[email protected]>
> ---
> I did think about making sve_vqm1_sve_for_el() return some
> value rather than asserting, but (a) what would be the right
> value? and (b) this was the only place that needed fixing.
> ---
> target/arm/helper.c | 19 +++++++++++++------
> 1 file changed, 13 insertions(+), 6 deletions(-)
>
> diff --git a/target/arm/helper.c b/target/arm/helper.c
> index dce648b482..a3dd84a2d6 100644
> --- a/target/arm/helper.c
> +++ b/target/arm/helper.c
> @@ -4936,12 +4936,12 @@ static void smcr_write(CPUARMState *env, const
> ARMCPRegInfo *ri,
> uint64_t value)
> {
> int cur_el = arm_current_el(env);
> - int old_len = sve_vqm1_for_el(env, cur_el);
> + ARMCPU *cpu = env_archcpu(env);
> + int old_len = cpu_isar_feature(aa64_sve, cpu) ? sve_vqm1_for_el(env,
> cur_el) : 0;
> uint64_t valid_mask = R_SMCR_LEN_MASK | R_SMCR_FA64_MASK;
> - int new_len;
>
> QEMU_BUILD_BUG_ON(ARM_MAX_VQ > R_SMCR_LEN_MASK + 1);
> - if (cpu_isar_feature(aa64_sme2, env_archcpu(env))) {
> + if (cpu_isar_feature(aa64_sme2, cpu)) {
> valid_mask |= R_SMCR_EZT0_MASK;
> }
> value &= valid_mask;
> @@ -4953,10 +4953,17 @@ static void smcr_write(CPUARMState *env, const
> ARMCPRegInfo *ri,
> * current values for simplicity. But for QEMU internals, we must still
> * apply the narrower SVL to the Zregs and Pregs -- see the comment
> * above aarch64_sve_narrow_vq.
> + *
> + * If the CPU has only SME and not SVE, then turning streaming mode
> + * on and off can't ever change the SVL; we must avoid calling
> + * sve_vqm1_for_el() to ask for the SVE vector length when SM is 0
> + * because it will assert.
> */
> - new_len = sve_vqm1_for_el(env, cur_el);
> - if (new_len < old_len) {
> - aarch64_sve_narrow_vq(env, new_len + 1);
> + if (cpu_isar_feature(aa64_sve, cpu)) {
> + int new_len = sve_vqm1_for_el(env, cur_el);
> + if (new_len < old_len) {
> + aarch64_sve_narrow_vq(env, new_len + 1);
> + }
> }
This isn't right -- I forgot that SMCR also has the SME vector
length, so old_len / new_len can be different here and we need
to call narrow_vq for that case.
Also aarch64_sve_narrow_vq() does this:
assert(vq <= env_archcpu(env)->sve_max_vq);
and I think that should be asserting that the new vq is <=
MAX(sve_max_vq, sme_max_vq).
-- PMM
