Re: [Qemu-devel] [PATCH 17/38] tcg: Add gvec expanders for vector shift by scalar

[David Hildenbrand]

On 20.04.19 09:34, Richard Henderson wrote:
> Signed-off-by: Richard Henderson <address@hidden>
> ---
>  tcg/tcg-op-gvec.h |   7 ++
>  tcg/tcg-op.h      |   4 +
>  tcg/tcg-op-gvec.c | 210 ++++++++++++++++++++++++++++++++++++++++++++++
>  tcg/tcg-op-vec.c  |  54 ++++++++++++
>  4 files changed, 275 insertions(+)
> 
> diff --git a/tcg/tcg-op-gvec.h b/tcg/tcg-op-gvec.h
> index a0e0902f6c..f9c6058e92 100644
> --- a/tcg/tcg-op-gvec.h
> +++ b/tcg/tcg-op-gvec.h
> @@ -318,6 +318,13 @@ void tcg_gen_gvec_shri(unsigned vece, uint32_t dofs, 
> uint32_t aofs,
>  void tcg_gen_gvec_sari(unsigned vece, uint32_t dofs, uint32_t aofs,
>                         int64_t shift, uint32_t oprsz, uint32_t maxsz);
>  
> +void tcg_gen_gvec_shls(unsigned vece, uint32_t dofs, uint32_t aofs,
> +                       TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz);
> +void tcg_gen_gvec_shrs(unsigned vece, uint32_t dofs, uint32_t aofs,
> +                       TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz);
> +void tcg_gen_gvec_sars(unsigned vece, uint32_t dofs, uint32_t aofs,
> +                       TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz);

I assume all irrelevant bits of the shift have to be masked off by the
caller, right?

On s390x, I would use it for (one variant of) VECTOR ELEMENT SHIFT like
this:


+static DisasJumpType op_ves(DisasContext *s, DisasOps *o)
+{
+    const uint8_t es = get_field(s->fields, m4);
+    const uint8_t d2 = get_field(s->fields, d2) &
+                       (NUM_VEC_ELEMENT_BITS(es) - 1);
+    const uint8_t v1 = get_field(s->fields, v1);
+    const uint8_t v3 = get_field(s->fields, v3);
+    TCGv_i32 shift;
+
+    if (es > ES_64) {
+        gen_program_exception(s, PGM_SPECIFICATION);
+        return DISAS_NORETURN;
+    }
+
+    shift = tcg_temp_new_i32();
+    tcg_gen_extrl_i64_i32(shift, o->addr1);
+    tcg_gen_andi_i32(shift, shift, NUM_VEC_ELEMENT_BITS(es) - 1);
+
+    switch (s->fields->op2) {
+    case 0x30:
+        if (likely(!get_field(s->fields, b2))) {
+            gen_gvec_fn_2i(shli, es, v1, v3, d2);
+        } else {
+            gen_gvec_fn_2s(shls, es, v1, v3, shift);
+        }
+        break;
+    case 0x3a:
+        if (likely(!get_field(s->fields, b2))) {
+            gen_gvec_fn_2i(sari, es, v1, v3, d2);
+        } else {
+            gen_gvec_fn_2s(sars, es, v1, v3, shift);
+        }
+        break;
+    case 0x38:
+        if (likely(!get_field(s->fields, b2))) {
+            gen_gvec_fn_2i(shri, es, v1, v3, d2);
+        } else {
+            gen_gvec_fn_2s(shrs, es, v1, v3, shift);
+        }
+        break;
+    default:
+        g_assert_not_reached();
+    }
+    tcg_temp_free_i32(shift);
+    return DISAS_NEXT;
+}

Does it still make sense to special-case the const immediate case?

-- 

Thanks,

David / dhildenb