Re: [PATCH v2 2/2] target/riscv: fix handling of nop for vstart >= vl in some vector instruction

[Daniel Henrique Barboza <dbarboza@ventanamicro.com>]

On 12/18/24 4:15 AM, Chao Liu wrote:

We want a commit message here because the change is not trivial.

You gave an explanation in the v1 cover letter:

---
Recently, when I was writing a RISCV test, I found that when VL is set to 0, the
instruction should be nop, but when I tested it, I found that QEMU will treat
all elements as tail elements, and in the case of VTA=1, write all elements
to 1.

After troubleshooting, it was found that the vext_vx_rm_1 function was called in
the vext_vx_rm_2, and then the vext_set_elems_1s function was called to process
the tail element, but only VSTART >= vl was checked in the vext_vx_rm_1
function, which caused the tail element to still be processed even if it was
returned in advance.

So I've made the following change:

put VSTART_CHECK_EARLY_EXIT(env) at the beginning of the vext_vx_rm_2 function,
so that the VSTART register is checked correctly.
----

You can add this text as a commit msg for this patch, also mentioning that you're
changing vext_vv_rm_1 / vext_vv_rm_2 for the same reason.


> fix: https://lore.kernel.org/all/20240322085319.1758843-8-alistair.francis@wdc.com/

The tag format we use in QEMU refers to the commit in the tree, not the link for the
pull request email.

In this case the tag would be:

Fixes: df4252b2ec ("target/riscv/vector_helpers: do early exit when vstart >= vl")

> Signed-off-by: Chao Liu <lc00631@tecorigin.com>
> ---


The code itself looks good to me. Thanks,

Daniel

>   target/riscv/vector_helper.c | 18 ++++++++++++++----
>   1 file changed, 14 insertions(+), 4 deletions(-)
> 
> diff --git a/target/riscv/vector_helper.c b/target/riscv/vector_helper.c
> index 4f14395808..5f1fc24d99 100644
> --- a/target/riscv/vector_helper.c
> +++ b/target/riscv/vector_helper.c
> @@ -2151,8 +2151,6 @@ vext_vv_rm_1(void *vd, void *v0, void *vs1, void *vs2,
>                uint32_t vl, uint32_t vm, int vxrm,
>                opivv2_rm_fn *fn, uint32_t vma, uint32_t esz)
>   {
> -    VSTART_CHECK_EARLY_EXIT(env, vl);
> -
>       for (uint32_t i = env->vstart; i < vl; i++) {
>           if (!vm && !vext_elem_mask(v0, i)) {
>               /* set masked-off elements to 1s */
> @@ -2176,6 +2174,8 @@ vext_vv_rm_2(void *vd, void *v0, void *vs1, void *vs2,
>       uint32_t vta = vext_vta(desc);
>       uint32_t vma = vext_vma(desc);
>   
> +    VSTART_CHECK_EARLY_EXIT(env, vl);
> +
>       switch (env->vxrm) {
>       case 0: /* rnu */
>           vext_vv_rm_1(vd, v0, vs1, vs2,
> @@ -2278,8 +2278,6 @@ vext_vx_rm_1(void *vd, void *v0, target_long s1, void *vs2,
>                uint32_t vl, uint32_t vm, int vxrm,
>                opivx2_rm_fn *fn, uint32_t vma, uint32_t esz)
>   {
> -    VSTART_CHECK_EARLY_EXIT(env, vl);
> -
>       for (uint32_t i = env->vstart; i < vl; i++) {
>           if (!vm && !vext_elem_mask(v0, i)) {
>               /* set masked-off elements to 1s */
> @@ -2303,6 +2301,8 @@ vext_vx_rm_2(void *vd, void *v0, target_long s1, void *vs2,
>       uint32_t vta = vext_vta(desc);
>       uint32_t vma = vext_vma(desc);
>   
> +    VSTART_CHECK_EARLY_EXIT(env, vl);
> +
>       switch (env->vxrm) {
>       case 0: /* rnu */
>           vext_vx_rm_1(vd, v0, s1, vs2,
> @@ -4638,6 +4638,8 @@ void HELPER(NAME)(void *vd, void *v0, void *vs1,          \
>       uint32_t i;                                           \
>       TD s1 =  *((TD *)vs1 + HD(0));                        \
>                                                             \
> +    VSTART_CHECK_EARLY_EXIT(env, vl);                     \
> +                                                          \
>       for (i = env->vstart; i < vl; i++) {                  \
>           TS2 s2 = *((TS2 *)vs2 + HS2(i));                  \
>           if (!vm && !vext_elem_mask(v0, i)) {              \
> @@ -4724,6 +4726,8 @@ void HELPER(NAME)(void *vd, void *v0, void *vs1,           \
>       uint32_t i;                                            \
>       TD s1 =  *((TD *)vs1 + HD(0));                         \
>                                                              \
> +    VSTART_CHECK_EARLY_EXIT(env, vl);                      \
> +                                                           \
>       for (i = env->vstart; i < vl; i++) {                   \
>           TS2 s2 = *((TS2 *)vs2 + HS2(i));                   \
>           if (!vm && !vext_elem_mask(v0, i)) {               \
> @@ -4886,6 +4890,8 @@ static void vmsetm(void *vd, void *v0, void *vs2, CPURISCVState *env,
>       int i;
>       bool first_mask_bit = false;
>   
> +    VSTART_CHECK_EARLY_EXIT(env, vl);
> +
>       for (i = env->vstart; i < vl; i++) {
>           if (!vm && !vext_elem_mask(v0, i)) {
>               /* set masked-off elements to 1s */
> @@ -4958,6 +4964,8 @@ void HELPER(NAME)(void *vd, void *v0, void *vs2, CPURISCVState *env,      \
>       uint32_t sum = 0;                                                     \
>       int i;                                                                \
>                                                                             \
> +    VSTART_CHECK_EARLY_EXIT(env, vl);                                     \
> +                                                                          \
>       for (i = env->vstart; i < vl; i++) {                                  \
>           if (!vm && !vext_elem_mask(v0, i)) {                              \
>               /* set masked-off elements to 1s */                           \
> @@ -5316,6 +5324,8 @@ void HELPER(NAME)(void *vd, void *v0, void *vs1, void *vs2,               \
>       uint32_t vta = vext_vta(desc);                                        \
>       uint32_t num = 0, i;                                                  \
>                                                                             \
> +    VSTART_CHECK_EARLY_EXIT(env, vl);                                     \
> +                                                                          \
>       for (i = env->vstart; i < vl; i++) {                                  \
>           if (!vext_elem_mask(vs1, i)) {                                    \
>               continue;                                                     \