Re: [PATCH 11/19] target/arm: Replace ARM_FEATURE_VFP4 with isar_feature_aa32_simdfmac

[Peter Maydell]

On Fri, 14 Feb 2020 at 18:16, Richard Henderson
<address@hidden> wrote:
>
> All remaining tests for VFP4 are for fused multiply-add insns.
>
> Signed-off-by: Richard Henderson <address@hidden>
> ---
>  target/arm/cpu.h               |  5 +++++
>  target/arm/translate-vfp.inc.c | 12 ++++++++----
>  target/arm/translate.c         |  2 +-
>  3 files changed, 14 insertions(+), 5 deletions(-)
>
> diff --git a/target/arm/cpu.h b/target/arm/cpu.h
> index 4ff28418df..f27b8e35df 100644
> --- a/target/arm/cpu.h
> +++ b/target/arm/cpu.h
> @@ -3468,6 +3468,11 @@ static inline bool isar_feature_aa32_fp16_dpconv(const 
> ARMISARegisters *id)
>      return FIELD_EX32(id->mvfr1, MVFR1, FPHP) > 1;
>  }
>
> +static inline bool isar_feature_aa32_simdfmac(const ARMISARegisters *id)
> +{
> +    return FIELD_EX32(id->mvfr1, MVFR1, SIMDFMAC) != 0;
> +}

This is tricky, because the SIMDFMAC register
field indicates "do we have fused-multiply-accumulate
for either VFP or Neon", so in a VFP-no-Neon core or
a Neon-no-VFP core it will be 1 but can't be used on its
own as a gate on "should this insn be present".

Currently in the part of arm_cpu_realize() which handles
the user having selected vfp=off and/or neon=off we
do allow (for AArch32 cores) both of those combinations.

trans_VFM_dp already tests aa32_fpdp_v2, so I think the
main thing we need to do is add a test on aa32_fpsp_v2 to
trans_VFM_sp.

We clear the SIMDFMAC field to 0 in the !has_neon condition,
and I think that should actually be in the !neon && !vfp part.

I propose to squash in the following and beef up the commit message:

diff --git a/target/arm/cpu.h b/target/arm/cpu.h
index f641478fc80..d4c73a20b6a 100644
--- a/target/arm/cpu.h
+++ b/target/arm/cpu.h
@@ -3506,6 +3506,13 @@ static inline bool
isar_feature_aa32_fp16_dpconv(const ARMISARegisters *id)
     return FIELD_EX32(id->mvfr1, MVFR1, FPHP) > 1;
 }

+/*
+ * Note that this ID register field covers both VFP and Neon FMAC,
+ * so should usually be tested in combination with some other
+ * check that confirms the presence of whichever of VFP or Neon is
+ * relevant, to avoid accidentally enabling a Neon feature on
+ * a VFP-no-Neon core or vice-versa.
+ */
 static inline bool isar_feature_aa32_simdfmac(const ARMISARegisters *id)
 {
     return FIELD_EX32(id->mvfr1, MVFR1, SIMDFMAC) != 0;
diff --git a/target/arm/cpu.c b/target/arm/cpu.c
index d5a75c265ac..95ada81ebae 100644
--- a/target/arm/cpu.c
+++ b/target/arm/cpu.c
@@ -1510,7 +1510,6 @@ static void arm_cpu_realizefn(DeviceState *dev,
Error **errp)
         u = FIELD_DP32(u, MVFR1, SIMDINT, 0);
         u = FIELD_DP32(u, MVFR1, SIMDSP, 0);
         u = FIELD_DP32(u, MVFR1, SIMDHP, 0);
-        u = FIELD_DP32(u, MVFR1, SIMDFMAC, 0);
         cpu->isar.mvfr1 = u;

         u = cpu->isar.mvfr2;
@@ -1533,6 +1532,11 @@ static void arm_cpu_realizefn(DeviceState *dev,
Error **errp)
         u = cpu->isar.mvfr0;
         u = FIELD_DP32(u, MVFR0, SIMDREG, 0);
         cpu->isar.mvfr0 = u;
+
+        /* Despite the name, this field covers both VFP and Neon */
+        u = cpu->isar.mvfr1;
+        u = FIELD_DP32(u, MVFR1, SIMDFMAC, 0);
+        cpu->isar.mvfr1;
     }

     if (arm_feature(env, ARM_FEATURE_M) && !cpu->has_dsp) {
diff --git a/target/arm/translate-vfp.inc.c b/target/arm/translate-vfp.inc.c
index f6f7601fe2a..69052d840a4 100644
--- a/target/arm/translate-vfp.inc.c
+++ b/target/arm/translate-vfp.inc.c
@@ -1805,8 +1805,13 @@ static bool trans_VFM_sp(DisasContext *s, arg_VFM_sp *a)
      * Present in VFPv4 only.
      * In v7A, UNPREDICTABLE with non-zero vector length/stride; from
      * v8A, must UNDEF. We choose to UNDEF for both v7A and v8A.
+     * Note that we can't rely on the SIMDFMAC check alone, because
+     * in a Neon-no-VFP core that ID register field will be non-zero.
      */
-    if (!dc_isar_feature(aa32_simdfmac, s)) {
+    if (!dc_isar_feature(aa32_simdfmac, s) ||
+        !dc_isar_feature(aa32_fpsp_v2, s)) {
+        return false;
+    }
         return false;
     }
     if (s->vec_len != 0 || s->vec_stride != 0) {


thanks
-- PMM