[PATCH 0/3] target/arm: Decode Neon fp sizes in decodetree

[PATCH 0/3] target/arm: Decode Neon fp sizes in decodetree

[Peter Maydell]

This patchset fixes something that Richard pointed out during
review of my fp16 patchset.

In the Neon instruction set, integer operations always have a 2-bit
size field with an encoding corresponding to QEMU's MO_{8,16,32,64}
constants: 0 for 8 bit, 1 for 16 bit, 2 for 32 bit and 3 for 64-bit.
However for the floating point instructions, since only 16-bit
and 32-bit floats are supported some instructions use one of the
size bits as part of decode, leaving only a 1-bit size field,
which might be encoded either as 0 for 16-bit float and 1 for
32-bit float, or the other way around as 0 for 32-bit float and
1 for 16-bit float.

Currently we simply pass the raw size field values through from
decodetree to the trans_ functions, which means that when reading
or writing the code for a trans_ function you need to know the
encoding of the size field for that insn. This patchset moves
the handling of the different possible encodings into the decodetree
file, so that the trans_ function is always passed a MO_* value
for the size.

thanks
-- PMM

Peter Maydell (3):
  target/arm: Convert Neon 3-same-fp size field to MO_* in decode
  target/arm: Convert Neon VCVT fp size field to MO_* in decode
  target/arm: Convert VCMLA, VCADD size field to MO_* in decode

 target/arm/neon-dp.decode       | 18 ++++++++------
 target/arm/neon-shared.decode   | 18 +++++++++-----
 target/arm/translate-neon.c.inc | 42 ++++++++++++++++++++-------------
 3 files changed, 48 insertions(+), 30 deletions(-)

-- 
2.20.1

[PATCH 1/3] target/arm: Convert Neon 3-same-fp size field to MO_* in decode

[Peter Maydell]

In the Neon instructions, some instruction formats have a 2-bit size
field which corresponds exactly to QEMU's MO_8/16/32/64.  However the
floating-point insns in the 3-same group have a 1-bit size field
which is "0 for 32-bit float and 1 for 16-bit float".  Currently we
pass these values directly through to trans_ functions, which means
that when reading a particular trans_ function you need to know if
that insn uses a 2-bit size or a 1-bit size.

Move the handling of the 1-bit size to the decodetree file, so that
all these insns consistently pass a size to the trans_ function which
is an MO_8/16/32/64 value.

In this commit we switch over the insns using the 3same_fp and
3same_fp_q0 formats.

Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
---
I wanted to call the field %3same_fp_size, but decodetree
doesn't seem to allow a field starting with a digit, even
though it does allow a format that starts with a digit.
So it's %fp_3same_size...
---
 target/arm/neon-dp.decode       | 15 ++++++++++-----
 target/arm/translate-neon.c.inc | 16 +++++++++++-----
 2 files changed, 21 insertions(+), 10 deletions(-)

diff --git a/target/arm/neon-dp.decode b/target/arm/neon-dp.decode
index 1e9e8592917..f453833396f 100644
--- a/target/arm/neon-dp.decode
+++ b/target/arm/neon-dp.decode
@@ -45,11 +45,16 @@
 @3same_q0        .... ... . . . size:2 .... .... .... . 0 . . .... \
                  &3same vm=%vm_dp vn=%vn_dp vd=%vd_dp q=0
 
-# For FP insns the high bit of 'size' is used as part of opcode decode
-@3same_fp        .... ... . . . . size:1 .... .... .... . q:1 . . .... \
-                 &3same vm=%vm_dp vn=%vn_dp vd=%vd_dp
-@3same_fp_q0     .... ... . . . . size:1 .... .... .... . 0 . . .... \
-                 &3same vm=%vm_dp vn=%vn_dp vd=%vd_dp q=0
+# For FP insns the high bit of 'size' is used as part of opcode decode,
+# and the 'size' bit is 0 for 32-bit float and 1 for 16-bit float.
+# This converts this encoding to the same MO_8/16/32/64 values that the
+# integer neon insns use.
+%fp_3same_size   20:1 !function=neon_3same_fp_size
+
+@3same_fp        .... ... . . . . . .... .... .... . q:1 . . .... \
+                 &3same vm=%vm_dp vn=%vn_dp vd=%vd_dp size=%fp_3same_size
+@3same_fp_q0     .... ... . . . . . .... .... .... . 0 . . .... \
+                 &3same vm=%vm_dp vn=%vn_dp vd=%vd_dp q=0 size=%fp_3same_size
 
 VHADD_S_3s       1111 001 0 0 . .. .... .... 0000 . . . 0 .... @3same
 VHADD_U_3s       1111 001 1 0 . .. .... .... 0000 . . . 0 .... @3same
diff --git a/target/arm/translate-neon.c.inc b/target/arm/translate-neon.c.inc
index 2d4926316a4..255c1cf8a2a 100644
--- a/target/arm/translate-neon.c.inc
+++ b/target/arm/translate-neon.c.inc
@@ -49,6 +49,12 @@ static inline int rsub_8(DisasContext *s, int x)
     return 8 - x;
 }
 
+static inline int neon_3same_fp_size(DisasContext *s, int x)
+{
+    /* Convert 0==fp32, 1==fp16 into a MO_* value */
+    return MO_32 - x;
+}
+
 /* Include the generated Neon decoder */
 #include "decode-neon-dp.c.inc"
 #include "decode-neon-ls.c.inc"
@@ -1049,7 +1055,7 @@ DO_3SAME_VQDMULH(VQRDMULH, qrdmulh)
     WRAP_FP_GVEC(gen_##INSN##_fp16_3s, FPST_STD_F16, HFUNC)             \
     static bool trans_##INSN##_fp_3s(DisasContext *s, arg_3same *a)     \
     {                                                                   \
-        if (a->size != 0) {                                             \
+        if (a->size == MO_16) {                                         \
             if (!dc_isar_feature(aa32_fp16_arith, s)) {                 \
                 return false;                                           \
             }                                                           \
@@ -1088,7 +1094,7 @@ static bool trans_VMAXNM_fp_3s(DisasContext *s, arg_3same *a)
         return false;
     }
 
-    if (a->size != 0) {
+    if (a->size == MO_16) {
         if (!dc_isar_feature(aa32_fp16_arith, s)) {
             return false;
         }
@@ -1103,7 +1109,7 @@ static bool trans_VMINNM_fp_3s(DisasContext *s, arg_3same *a)
         return false;
     }
 
-    if (a->size != 0) {
+    if (a->size == MO_16) {
         if (!dc_isar_feature(aa32_fp16_arith, s)) {
             return false;
         }
@@ -1135,7 +1141,7 @@ static bool do_3same_fp_pair(DisasContext *s, arg_3same *a,
     assert(a->q == 0); /* enforced by decode patterns */
 
 
-    fpstatus = fpstatus_ptr(a->size != 0 ? FPST_STD_F16 : FPST_STD);
+    fpstatus = fpstatus_ptr(a->size == MO_16 ? FPST_STD_F16 : FPST_STD);
     tcg_gen_gvec_3_ptr(vfp_reg_offset(1, a->vd),
                        vfp_reg_offset(1, a->vn),
                        vfp_reg_offset(1, a->vm),
@@ -1152,7 +1158,7 @@ static bool do_3same_fp_pair(DisasContext *s, arg_3same *a,
 #define DO_3S_FP_PAIR(INSN,FUNC)                                    \
     static bool trans_##INSN##_fp_3s(DisasContext *s, arg_3same *a) \
     {                                                               \
-        if (a->size != 0) {                                         \
+        if (a->size == MO_16) {                                     \
             if (!dc_isar_feature(aa32_fp16_arith, s)) {             \
                 return false;                                       \
             }                                                       \
-- 
2.20.1

Re: [PATCH 1/3] target/arm: Convert Neon 3-same-fp size field to MO_* in decode

[Richard Henderson]

On 9/3/20 6:32 AM, Peter Maydell wrote:
> In the Neon instructions, some instruction formats have a 2-bit size
> field which corresponds exactly to QEMU's MO_8/16/32/64.  However the
> floating-point insns in the 3-same group have a 1-bit size field
> which is "0 for 32-bit float and 1 for 16-bit float".  Currently we
> pass these values directly through to trans_ functions, which means
> that when reading a particular trans_ function you need to know if
> that insn uses a 2-bit size or a 1-bit size.
> 
> Move the handling of the 1-bit size to the decodetree file, so that
> all these insns consistently pass a size to the trans_ function which
> is an MO_8/16/32/64 value.
> 
> In this commit we switch over the insns using the 3same_fp and
> 3same_fp_q0 formats.
> 
> Signed-off-by: Peter Maydell <peter.maydell@linaro.org>

Reviewed-by: Richard Henderson <richard.henderson@linaro.org>


> ---
> I wanted to call the field %3same_fp_size, but decodetree
> doesn't seem to allow a field starting with a digit, even
> though it does allow a format that starts with a digit.
> So it's %fp_3same_size...

Odd.  All of the names get prefixed, so we don't have a problem of a digit
beginning a C identifier...  I can look at fixing this if you want.


r~

Re: [PATCH 1/3] target/arm: Convert Neon 3-same-fp size field to MO_* in decode

[Peter Maydell]

On Thu, 3 Sep 2020 at 17:25, Richard Henderson
<richard.henderson@linaro.org> wrote:
>
> On 9/3/20 6:32 AM, Peter Maydell wrote:
> > I wanted to call the field %3same_fp_size, but decodetree
> > doesn't seem to allow a field starting with a digit, even
> > though it does allow a format that starts with a digit.
> > So it's %fp_3same_size...
>
> Odd.  All of the names get prefixed, so we don't have a problem of a digit
> beginning a C identifier...  I can look at fixing this if you want.

If it's not too involved a fix it would be nice. The failure
is that it doesn't manage to parse the file, rather than
that it produces bad code.

thanks
-- PMM

[PATCH 2/3] target/arm: Convert Neon VCVT fp size field to MO_* in decode

[Peter Maydell]

Convert the insns using the 2reg_vcvt and 2reg_vcvt_f16 formats
to pass the size through to the trans function as a MO_* value
rather than the '0==f32, 1==f16' used in the fp 3-same encodings.

Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
---
 target/arm/neon-dp.decode       | 3 +--
 target/arm/translate-neon.c.inc | 4 ++--
 2 files changed, 3 insertions(+), 4 deletions(-)

diff --git a/target/arm/neon-dp.decode b/target/arm/neon-dp.decode
index f453833396f..d7da2e7db3c 100644
--- a/target/arm/neon-dp.decode
+++ b/target/arm/neon-dp.decode
@@ -256,9 +256,8 @@ VMINNM_fp_3s     1111 001 1 0 . 1 . .... .... 1111 ... 1 .... @3same_fp
 @2reg_shll_b     .... ... . . . 001 shift:3  .... .... 0 . . . .... \
                  &2reg_shift vm=%vm_dp vd=%vd_dp size=0 q=0
 
-# We use size=0 for fp32 and size=1 for fp16 to match the 3-same encodings.
 @2reg_vcvt       .... ... . . . 1 ..... .... .... . q:1 . . .... \
-                 &2reg_shift vm=%vm_dp vd=%vd_dp size=0 shift=%neon_rshift_i5
+                 &2reg_shift vm=%vm_dp vd=%vd_dp size=2 shift=%neon_rshift_i5
 @2reg_vcvt_f16   .... ... . . . 11 .... .... .... . q:1 . . .... \
                  &2reg_shift vm=%vm_dp vd=%vd_dp size=1 shift=%neon_rshift_i4
 
diff --git a/target/arm/translate-neon.c.inc b/target/arm/translate-neon.c.inc
index 255c1cf8a2a..213c1c2174a 100644
--- a/target/arm/translate-neon.c.inc
+++ b/target/arm/translate-neon.c.inc
@@ -1626,7 +1626,7 @@ static bool do_fp_2sh(DisasContext *s, arg_2reg_shift *a,
         return false;
     }
 
-    if (a->size != 0) {
+    if (a->size == MO_16) {
         if (!dc_isar_feature(aa32_fp16_arith, s)) {
             return false;
         }
@@ -1646,7 +1646,7 @@ static bool do_fp_2sh(DisasContext *s, arg_2reg_shift *a,
         return true;
     }
 
-    fpst = fpstatus_ptr(a->size ? FPST_STD_F16 : FPST_STD);
+    fpst = fpstatus_ptr(a->size == MO_16 ? FPST_STD_F16 : FPST_STD);
     tcg_gen_gvec_2_ptr(rd_ofs, rm_ofs, fpst, vec_size, vec_size, a->shift, fn);
     tcg_temp_free_ptr(fpst);
     return true;
-- 
2.20.1

Re: [PATCH 2/3] target/arm: Convert Neon VCVT fp size field to MO_* in decode

[Richard Henderson]

On 9/3/20 6:32 AM, Peter Maydell wrote:
> Convert the insns using the 2reg_vcvt and 2reg_vcvt_f16 formats
> to pass the size through to the trans function as a MO_* value
> rather than the '0==f32, 1==f16' used in the fp 3-same encodings.
> 
> Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
> ---
>  target/arm/neon-dp.decode       | 3 +--
>  target/arm/translate-neon.c.inc | 4 ++--
>  2 files changed, 3 insertions(+), 4 deletions(-)

Reviewed-by: Richard Henderson <richard.henderson@linaro.org>

r~

[PATCH 3/3] target/arm: Convert VCMLA, VCADD size field to MO_* in decode

[Peter Maydell]

The VCMLA and VCADD insns have a size field which is 0 for fp16
and 1 for fp32 (note that this is the reverse of the Neon 3-same
encoding!). Convert it to MO_* values in decode for consistency.

Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
---
 target/arm/neon-shared.decode   | 18 ++++++++++++------
 target/arm/translate-neon.c.inc | 22 ++++++++++++----------
 2 files changed, 24 insertions(+), 16 deletions(-)

diff --git a/target/arm/neon-shared.decode b/target/arm/neon-shared.decode
index f297ba8cdfc..a9d010880d4 100644
--- a/target/arm/neon-shared.decode
+++ b/target/arm/neon-shared.decode
@@ -34,11 +34,17 @@
 %vd_dp  22:1 12:4
 %vd_sp  12:4 22:1
 
-VCMLA          1111 110 rot:2 . 1 size:1 .... .... 1000 . q:1 . 0 .... \
-               vm=%vm_dp vn=%vn_dp vd=%vd_dp
+# For VCMLA/VCADD insns, convert the single-bit size field
+# which is 0 for fp16 and 1 for fp32 into a MO_* constant.
+# (Note that this is the reverse of the sense of the 1-bit size
+# field in the 3same_fp Neon insns.)
+%vcadd_size    20:1 !function=plus1
 
-VCADD          1111 110 rot:1 1 . 0 size:1 .... .... 1000 . q:1 . 0 .... \
-               vm=%vm_dp vn=%vn_dp vd=%vd_dp
+VCMLA          1111 110 rot:2 . 1 . .... .... 1000 . q:1 . 0 .... \
+               vm=%vm_dp vn=%vn_dp vd=%vd_dp size=%vcadd_size
+
+VCADD          1111 110 rot:1 1 . 0 . .... .... 1000 . q:1 . 0 .... \
+               vm=%vm_dp vn=%vn_dp vd=%vd_dp size=%vcadd_size
 
 # VUDOT and VSDOT
 VDOT           1111 110 00 . 10 .... .... 1101 . q:1 . u:1 .... \
@@ -51,9 +57,9 @@ VFML           1111 110 0 s:1 . 10 .... .... 1000 . 1 . 1 .... \
                vm=%vm_dp vn=%vn_dp vd=%vd_dp q=1
 
 VCMLA_scalar   1111 1110 0 . rot:2 .... .... 1000 . q:1 index:1 0 vm:4 \
-               vn=%vn_dp vd=%vd_dp size=0
+               vn=%vn_dp vd=%vd_dp size=1
 VCMLA_scalar   1111 1110 1 . rot:2 .... .... 1000 . q:1 . 0 .... \
-               vm=%vm_dp vn=%vn_dp vd=%vd_dp size=1 index=0
+               vm=%vm_dp vn=%vn_dp vd=%vd_dp size=2 index=0
 
 VDOT_scalar    1111 1110 0 . 10 .... .... 1101 . q:1 index:1 u:1 rm:4 \
                vm=%vm_dp vn=%vn_dp vd=%vd_dp
diff --git a/target/arm/translate-neon.c.inc b/target/arm/translate-neon.c.inc
index 213c1c2174a..4d1a292981b 100644
--- a/target/arm/translate-neon.c.inc
+++ b/target/arm/translate-neon.c.inc
@@ -168,7 +168,7 @@ static bool trans_VCMLA(DisasContext *s, arg_VCMLA *a)
     gen_helper_gvec_3_ptr *fn_gvec_ptr;
 
     if (!dc_isar_feature(aa32_vcma, s)
-        || (!a->size && !dc_isar_feature(aa32_fp16_arith, s))) {
+        || (a->size == MO_16 && !dc_isar_feature(aa32_fp16_arith, s))) {
         return false;
     }
 
@@ -187,8 +187,9 @@ static bool trans_VCMLA(DisasContext *s, arg_VCMLA *a)
     }
 
     opr_sz = (1 + a->q) * 8;
-    fpst = fpstatus_ptr(a->size == 0 ? FPST_STD_F16 : FPST_STD);
-    fn_gvec_ptr = a->size ? gen_helper_gvec_fcmlas : gen_helper_gvec_fcmlah;
+    fpst = fpstatus_ptr(a->size == MO_16 ? FPST_STD_F16 : FPST_STD);
+    fn_gvec_ptr = (a->size == MO_16) ?
+        gen_helper_gvec_fcmlah : gen_helper_gvec_fcmlas;
     tcg_gen_gvec_3_ptr(vfp_reg_offset(1, a->vd),
                        vfp_reg_offset(1, a->vn),
                        vfp_reg_offset(1, a->vm),
@@ -205,7 +206,7 @@ static bool trans_VCADD(DisasContext *s, arg_VCADD *a)
     gen_helper_gvec_3_ptr *fn_gvec_ptr;
 
     if (!dc_isar_feature(aa32_vcma, s)
-        || (!a->size && !dc_isar_feature(aa32_fp16_arith, s))) {
+        || (a->size == MO_16 && !dc_isar_feature(aa32_fp16_arith, s))) {
         return false;
     }
 
@@ -224,8 +225,9 @@ static bool trans_VCADD(DisasContext *s, arg_VCADD *a)
     }
 
     opr_sz = (1 + a->q) * 8;
-    fpst = fpstatus_ptr(a->size == 0 ? FPST_STD_F16 : FPST_STD);
-    fn_gvec_ptr = a->size ? gen_helper_gvec_fcadds : gen_helper_gvec_fcaddh;
+    fpst = fpstatus_ptr(a->size == MO_16 ? FPST_STD_F16 : FPST_STD);
+    fn_gvec_ptr = (a->size == MO_16) ?
+        gen_helper_gvec_fcaddh : gen_helper_gvec_fcadds;
     tcg_gen_gvec_3_ptr(vfp_reg_offset(1, a->vd),
                        vfp_reg_offset(1, a->vn),
                        vfp_reg_offset(1, a->vm),
@@ -307,7 +309,7 @@ static bool trans_VCMLA_scalar(DisasContext *s, arg_VCMLA_scalar *a)
     if (!dc_isar_feature(aa32_vcma, s)) {
         return false;
     }
-    if (a->size == 0 && !dc_isar_feature(aa32_fp16_arith, s)) {
+    if (a->size == MO_16 && !dc_isar_feature(aa32_fp16_arith, s)) {
         return false;
     }
 
@@ -325,10 +327,10 @@ static bool trans_VCMLA_scalar(DisasContext *s, arg_VCMLA_scalar *a)
         return true;
     }
 
-    fn_gvec_ptr = (a->size ? gen_helper_gvec_fcmlas_idx
-                   : gen_helper_gvec_fcmlah_idx);
+    fn_gvec_ptr = (a->size == MO_16) ?
+        gen_helper_gvec_fcmlah_idx : gen_helper_gvec_fcmlas_idx;
     opr_sz = (1 + a->q) * 8;
-    fpst = fpstatus_ptr(a->size == 0 ? FPST_STD_F16 : FPST_STD);
+    fpst = fpstatus_ptr(a->size == MO_16 ? FPST_STD_F16 : FPST_STD);
     tcg_gen_gvec_3_ptr(vfp_reg_offset(1, a->vd),
                        vfp_reg_offset(1, a->vn),
                        vfp_reg_offset(1, a->vm),
-- 
2.20.1

Re: [PATCH 3/3] target/arm: Convert VCMLA, VCADD size field to MO_* in decode

[Richard Henderson]

On 9/3/20 6:32 AM, Peter Maydell wrote:
> The VCMLA and VCADD insns have a size field which is 0 for fp16
> and 1 for fp32 (note that this is the reverse of the Neon 3-same
> encoding!). Convert it to MO_* values in decode for consistency.
> 
> Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
> ---
>  target/arm/neon-shared.decode   | 18 ++++++++++++------
>  target/arm/translate-neon.c.inc | 22 ++++++++++++----------
>  2 files changed, 24 insertions(+), 16 deletions(-)

Reviewed-by: Richard Henderson <richard.henderson@linaro.org>


r~