[v4] target/arm: Scalable Matrix Extension

[PATCH v4 07/45] target/arm: Mark PMULL, FMMLA as non-streaming

Posted by Richard Henderson 3 years, 7 months ago

Mark these as a non-streaming instructions, which should trap
if full a64 support is not enabled in streaming mode.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 target/arm/sme-fa64.decode |  2 --
 target/arm/translate-sve.c | 18 ++++++++++--------
 2 files changed, 10 insertions(+), 10 deletions(-)

diff --git a/target/arm/sme-fa64.decode b/target/arm/sme-fa64.decode
index c25bad5ee5..c75a94e0fc 100644
--- a/target/arm/sme-fa64.decode
+++ b/target/arm/sme-fa64.decode
@@ -58,8 +58,6 @@ FAIL    1100 1110 ---- ---- ---- ---- ---- ----   # Advanced SIMD cryptography e
 #       --11 1100 --1- ---- ---- ---- ---- --10   # Load/store FP register (register offset)
 #       --11 1101 ---- ---- ---- ---- ---- ----   # Load/store FP register (scaled imm)
 
-FAIL    0100 0101 000- ---- 0110 1--- ---- ----   # PMULLB, PMULLT (128b result)
-FAIL    0110 0100 --1- ---- 1110 01-- ---- ----   # FMMLA, BFMMLA
 FAIL    0110 0101 --0- ---- 0000 11-- ---- ----   # FTSMUL
 FAIL    0110 0101 --01 0--- 100- ---- ---- ----   # FTMAD
 FAIL    0110 0101 --01 1--- 001- ---- ---- ----   # FADDA
diff --git a/target/arm/translate-sve.c b/target/arm/translate-sve.c
index ae48040aa4..130432654e 100644
--- a/target/arm/translate-sve.c
+++ b/target/arm/translate-sve.c
@@ -6194,8 +6194,8 @@ static bool do_trans_pmull(DisasContext *s, arg_rrr_esz *a, bool sel)
     return gen_gvec_ool_arg_zzz(s, fns[a->esz], a, sel);
 }
 
-TRANS_FEAT(PMULLB, aa64_sve2, do_trans_pmull, a, false)
-TRANS_FEAT(PMULLT, aa64_sve2, do_trans_pmull, a, true)
+TRANS_FEAT_NONSTREAMING(PMULLB, aa64_sve2, do_trans_pmull, a, false)
+TRANS_FEAT_NONSTREAMING(PMULLT, aa64_sve2, do_trans_pmull, a, true)
 
 static gen_helper_gvec_3 * const saddw_fns[4] = {
     NULL,                    gen_helper_sve2_saddw_h,
@@ -7125,10 +7125,12 @@ DO_ZPZZ_FP(FMINP, aa64_sve2, sve2_fminp_zpzz)
  * SVE Integer Multiply-Add (unpredicated)
  */
 
-TRANS_FEAT(FMMLA_s, aa64_sve_f32mm, gen_gvec_fpst_zzzz, gen_helper_fmmla_s,
-           a->rd, a->rn, a->rm, a->ra, 0, FPST_FPCR)
-TRANS_FEAT(FMMLA_d, aa64_sve_f64mm, gen_gvec_fpst_zzzz, gen_helper_fmmla_d,
-           a->rd, a->rn, a->rm, a->ra, 0, FPST_FPCR)
+TRANS_FEAT_NONSTREAMING(FMMLA_s, aa64_sve_f32mm, gen_gvec_fpst_zzzz,
+                        gen_helper_fmmla_s, a->rd, a->rn, a->rm, a->ra,
+                        0, FPST_FPCR)
+TRANS_FEAT_NONSTREAMING(FMMLA_d, aa64_sve_f64mm, gen_gvec_fpst_zzzz,
+                        gen_helper_fmmla_d, a->rd, a->rn, a->rm, a->ra,
+                        0, FPST_FPCR)
 
 static gen_helper_gvec_4 * const sqdmlal_zzzw_fns[] = {
     NULL,                           gen_helper_sve2_sqdmlal_zzzw_h,
@@ -7301,8 +7303,8 @@ TRANS_FEAT(BFDOT_zzzz, aa64_sve_bf16, gen_gvec_ool_arg_zzzz,
 TRANS_FEAT(BFDOT_zzxz, aa64_sve_bf16, gen_gvec_ool_arg_zzxz,
            gen_helper_gvec_bfdot_idx, a)
 
-TRANS_FEAT(BFMMLA, aa64_sve_bf16, gen_gvec_ool_arg_zzzz,
-           gen_helper_gvec_bfmmla, a, 0)
+TRANS_FEAT_NONSTREAMING(BFMMLA, aa64_sve_bf16, gen_gvec_ool_arg_zzzz,
+                        gen_helper_gvec_bfmmla, a, 0)
 
 static bool do_BFMLAL_zzzw(DisasContext *s, arg_rrrr_esz *a, bool sel)
 {
-- 
2.34.1

Re: [PATCH v4 07/45] target/arm: Mark PMULL, FMMLA as non-streaming

Posted by Peter Maydell 3 years, 7 months ago

On Tue, 28 Jun 2022 at 05:28, Richard Henderson
<richard.henderson@linaro.org> wrote:
>
> Mark these as a non-streaming instructions, which should trap
> if full a64 support is not enabled in streaming mode.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
>  target/arm/sme-fa64.decode |  2 --
>  target/arm/translate-sve.c | 18 ++++++++++--------
>  2 files changed, 10 insertions(+), 10 deletions(-)
>
> diff --git a/target/arm/sme-fa64.decode b/target/arm/sme-fa64.decode
> index c25bad5ee5..c75a94e0fc 100644
> --- a/target/arm/sme-fa64.decode
> +++ b/target/arm/sme-fa64.decode
> @@ -58,8 +58,6 @@ FAIL    1100 1110 ---- ---- ---- ---- ---- ----   # Advanced SIMD cryptography e
>  #       --11 1100 --1- ---- ---- ---- ---- --10   # Load/store FP register (register offset)
>  #       --11 1101 ---- ---- ---- ---- ---- ----   # Load/store FP register (scaled imm)
>
> -FAIL    0100 0101 000- ---- 0110 1--- ---- ----   # PMULLB, PMULLT (128b result)

This encoding pattern (as the comment says) covers only the esize = 128 case...

> -TRANS_FEAT(PMULLB, aa64_sve2, do_trans_pmull, a, false)
> -TRANS_FEAT(PMULLT, aa64_sve2, do_trans_pmull, a, true)
> +TRANS_FEAT_NONSTREAMING(PMULLB, aa64_sve2, do_trans_pmull, a, false)
> +TRANS_FEAT_NONSTREAMING(PMULLT, aa64_sve2, do_trans_pmull, a, true)

..but these trans functions cover all the esize values.

In the pseudocode for PMULLB/PMULLT it explicitly does
 if esize < 128 then CheckSVEEnabled(); else CheckNonStreamingSVEEnabled();
so it is definitely intended to distinguish the 128 bit case
from the 16 and 64 bit ones.

thanks
-- PMM

Re: [PATCH v4 07/45] target/arm: Mark PMULL, FMMLA as non-streaming

Posted by Richard Henderson 3 years, 7 months ago

On 7/1/22 17:48, Peter Maydell wrote:
> On Tue, 28 Jun 2022 at 05:28, Richard Henderson
> <richard.henderson@linaro.org> wrote:
>>
>> Mark these as a non-streaming instructions, which should trap
>> if full a64 support is not enabled in streaming mode.
>>
>> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
>> ---
>>   target/arm/sme-fa64.decode |  2 --
>>   target/arm/translate-sve.c | 18 ++++++++++--------
>>   2 files changed, 10 insertions(+), 10 deletions(-)
>>
>> diff --git a/target/arm/sme-fa64.decode b/target/arm/sme-fa64.decode
>> index c25bad5ee5..c75a94e0fc 100644
>> --- a/target/arm/sme-fa64.decode
>> +++ b/target/arm/sme-fa64.decode
>> @@ -58,8 +58,6 @@ FAIL    1100 1110 ---- ---- ---- ---- ---- ----   # Advanced SIMD cryptography e
>>   #       --11 1100 --1- ---- ---- ---- ---- --10   # Load/store FP register (register offset)
>>   #       --11 1101 ---- ---- ---- ---- ---- ----   # Load/store FP register (scaled imm)
>>
>> -FAIL    0100 0101 000- ---- 0110 1--- ---- ----   # PMULLB, PMULLT (128b result)
> 
> This encoding pattern (as the comment says) covers only the esize = 128 case...
> 
>> -TRANS_FEAT(PMULLB, aa64_sve2, do_trans_pmull, a, false)
>> -TRANS_FEAT(PMULLT, aa64_sve2, do_trans_pmull, a, true)
>> +TRANS_FEAT_NONSTREAMING(PMULLB, aa64_sve2, do_trans_pmull, a, false)
>> +TRANS_FEAT_NONSTREAMING(PMULLT, aa64_sve2, do_trans_pmull, a, true)
> 
> ..but these trans functions cover all the esize values.
> 
> In the pseudocode for PMULLB/PMULLT it explicitly does
>   if esize < 128 then CheckSVEEnabled(); else CheckNonStreamingSVEEnabled();
> so it is definitely intended to distinguish the 128 bit case
> from the 16 and 64 bit ones.

Whoops, too right.


r~