[PATCH for-6.2 26/34] target/arm: Implement MVE saturating doubling multiply accumulates

[Peter Maydell <peter.maydell@linaro.org>]

Implement the MVE saturating doubling multiply accumulate insns
VQDMLAH, VQRDMLAH, VQDMLASH and VQRDMLASH.  These perform a multiply,
double, add the accumulator shifted by the element size, possibly
round, saturate to twice the element size, then take the high half of
the result.  The *MLAH insns do vector * scalar + vector, and the
*MLASH insns do vector * vector + scalar.

Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
---
 target/arm/helper-mve.h    | 16 +++++++
 target/arm/mve.decode      |  5 ++
 target/arm/mve_helper.c    | 95 ++++++++++++++++++++++++++++++++++++++
 target/arm/translate-mve.c |  4 ++
 4 files changed, 120 insertions(+)

diff --git a/target/arm/helper-mve.h b/target/arm/helper-mve.h
index 50b34c601e1..e61c5d56f41 100644
--- a/target/arm/helper-mve.h
+++ b/target/arm/helper-mve.h
@@ -387,6 +387,22 @@ DEF_HELPER_FLAGS_4(mve_vmlasub, TCG_CALL_NO_WG, void, env, ptr, ptr, i32)
 DEF_HELPER_FLAGS_4(mve_vmlasuh, TCG_CALL_NO_WG, void, env, ptr, ptr, i32)
 DEF_HELPER_FLAGS_4(mve_vmlasuw, TCG_CALL_NO_WG, void, env, ptr, ptr, i32)
 
+DEF_HELPER_FLAGS_4(mve_vqdmlahb, TCG_CALL_NO_WG, void, env, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(mve_vqdmlahh, TCG_CALL_NO_WG, void, env, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(mve_vqdmlahw, TCG_CALL_NO_WG, void, env, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_4(mve_vqrdmlahb, TCG_CALL_NO_WG, void, env, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(mve_vqrdmlahh, TCG_CALL_NO_WG, void, env, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(mve_vqrdmlahw, TCG_CALL_NO_WG, void, env, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_4(mve_vqdmlashb, TCG_CALL_NO_WG, void, env, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(mve_vqdmlashh, TCG_CALL_NO_WG, void, env, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(mve_vqdmlashw, TCG_CALL_NO_WG, void, env, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_4(mve_vqrdmlashb, TCG_CALL_NO_WG, void, env, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(mve_vqrdmlashh, TCG_CALL_NO_WG, void, env, ptr, ptr, i32)
+DEF_HELPER_FLAGS_4(mve_vqrdmlashw, TCG_CALL_NO_WG, void, env, ptr, ptr, i32)
+
 DEF_HELPER_FLAGS_4(mve_vmlaldavsh, TCG_CALL_NO_WG, i64, env, ptr, ptr, i64)
 DEF_HELPER_FLAGS_4(mve_vmlaldavsw, TCG_CALL_NO_WG, i64, env, ptr, ptr, i64)
 DEF_HELPER_FLAGS_4(mve_vmlaldavxsh, TCG_CALL_NO_WG, i64, env, ptr, ptr, i64)
diff --git a/target/arm/mve.decode b/target/arm/mve.decode
index 2e2df61c860..99cea8d39b6 100644
--- a/target/arm/mve.decode
+++ b/target/arm/mve.decode
@@ -418,6 +418,11 @@ VMLA_U           1111 1110 0 . .. ... 1 ... 0 1110 . 100 .... @2scalar
 VMLAS_S          1110 1110 0 . .. ... 1 ... 1 1110 . 100 .... @2scalar
 VMLAS_U          1111 1110 0 . .. ... 1 ... 1 1110 . 100 .... @2scalar
 
+VQRDMLAH         1110 1110 0 . .. ... 0 ... 0 1110 . 100 .... @2scalar
+VQRDMLASH        1110 1110 0 . .. ... 0 ... 1 1110 . 100 .... @2scalar
+VQDMLAH          1110 1110 0 . .. ... 0 ... 0 1110 . 110 .... @2scalar
+VQDMLASH         1110 1110 0 . .. ... 0 ... 1 1110 . 110 .... @2scalar
+
 # Vector add across vector
 {
   VADDV          111 u:1 1110 1111 size:2 01 ... 0 1111 0 0 a:1 0 qm:3 0 rda=%rdalo
diff --git a/target/arm/mve_helper.c b/target/arm/mve_helper.c
index 91c0add8da7..1013060baeb 100644
--- a/target/arm/mve_helper.c
+++ b/target/arm/mve_helper.c
@@ -971,6 +971,28 @@ DO_VQDMLADH_OP(vqrdmlsdhxw, 4, int32_t, 1, 1, do_vqdmlsdh_w)
         mve_advance_vpt(env);                                           \
     }
 
+#define DO_2OP_SAT_ACC_SCALAR(OP, ESIZE, TYPE, FN)                      \
+    void HELPER(glue(mve_, OP))(CPUARMState *env, void *vd, void *vn,   \
+                                uint32_t rm)                            \
+    {                                                                   \
+        TYPE *d = vd, *n = vn;                                          \
+        TYPE m = rm;                                                    \
+        uint16_t mask = mve_element_mask(env);                          \
+        unsigned e;                                                     \
+        bool qc = false;                                                \
+        for (e = 0; e < 16 / ESIZE; e++, mask >>= ESIZE) {              \
+            bool sat = false;                                           \
+            mergemask(&d[H##ESIZE(e)],                                  \
+                      FN(d[H##ESIZE(e)], n[H##ESIZE(e)], m, &sat),      \
+                      mask);                                            \
+            qc |= sat & mask & 1;                                       \
+        }                                                               \
+        if (qc) {                                                       \
+            env->vfp.qc[0] = qc;                                        \
+        }                                                               \
+        mve_advance_vpt(env);                                           \
+    }
+
 /* provide unsigned 2-op scalar helpers for all sizes */
 #define DO_2OP_SCALAR_U(OP, FN)                 \
     DO_2OP_SCALAR(OP##b, 1, uint8_t, FN)        \
@@ -1019,6 +1041,79 @@ DO_2OP_SAT_SCALAR(vqrdmulh_scalarb, 1, int8_t, DO_QRDMULH_B)
 DO_2OP_SAT_SCALAR(vqrdmulh_scalarh, 2, int16_t, DO_QRDMULH_H)
 DO_2OP_SAT_SCALAR(vqrdmulh_scalarw, 4, int32_t, DO_QRDMULH_W)
 
+static int8_t do_vqdmlah_b(int8_t a, int8_t b, int8_t c, int round, bool *sat)
+{
+    int64_t r = (int64_t)a * b * 2 + ((int64_t)c << 8) + (round << 7);
+    return do_sat_bhw(r, INT16_MIN, INT16_MAX, sat) >> 8;
+}
+
+static int16_t do_vqdmlah_h(int16_t a, int16_t b, int16_t c,
+                           int round, bool *sat)
+{
+    int64_t r = (int64_t)a * b * 2 + ((int64_t)c << 16) + (round << 15);
+    return do_sat_bhw(r, INT32_MIN, INT32_MAX, sat) >> 16;
+}
+
+static int32_t do_vqdmlah_w(int32_t a, int32_t b, int32_t c,
+                            int round, bool *sat)
+{
+    /*
+     * Architecturally we should do the entire add, double, round
+     * and then check for saturation. We do three saturating adds,
+     * but we need to be careful about the order. If the first
+     * m1 + m2 saturates then it's impossible for the *2+rc to
+     * bring it back into the non-saturated range. However, if
+     * m1 + m2 is negative then it's possible that doing the doubling
+     * would take the intermediate result below INT64_MAX and the
+     * addition of the rounding constant then brings it back in range.
+     * So we add half the rounding constant and half the "c << esize"
+     * before doubling rather than adding the rounding constant after
+     * the doubling.
+     */
+    int64_t m1 = (int64_t)a * b;
+    int64_t m2 = (int64_t)c << 31;
+    int64_t r;
+    if (sadd64_overflow(m1, m2, &r) ||
+        sadd64_overflow(r, (round << 30), &r) ||
+        sadd64_overflow(r, r, &r)) {
+        *sat = true;
+        return r < 0 ? INT32_MAX : INT32_MIN;
+    }
+    return r >> 32;
+}
+
+/*
+ * The *MLAH insns are vector * scalar + vector;
+ * the *MLASH insns are vector * vector + scalar
+ */
+#define DO_VQDMLAH_B(D, N, M, S) do_vqdmlah_b(N, M, D, 0, S)
+#define DO_VQDMLAH_H(D, N, M, S) do_vqdmlah_h(N, M, D, 0, S)
+#define DO_VQDMLAH_W(D, N, M, S) do_vqdmlah_w(N, M, D, 0, S)
+#define DO_VQRDMLAH_B(D, N, M, S) do_vqdmlah_b(N, M, D, 1, S)
+#define DO_VQRDMLAH_H(D, N, M, S) do_vqdmlah_h(N, M, D, 1, S)
+#define DO_VQRDMLAH_W(D, N, M, S) do_vqdmlah_w(N, M, D, 1, S)
+
+#define DO_VQDMLASH_B(D, N, M, S) do_vqdmlah_b(N, D, M, 0, S)
+#define DO_VQDMLASH_H(D, N, M, S) do_vqdmlah_h(N, D, M, 0, S)
+#define DO_VQDMLASH_W(D, N, M, S) do_vqdmlah_w(N, D, M, 0, S)
+#define DO_VQRDMLASH_B(D, N, M, S) do_vqdmlah_b(N, D, M, 1, S)
+#define DO_VQRDMLASH_H(D, N, M, S) do_vqdmlah_h(N, D, M, 1, S)
+#define DO_VQRDMLASH_W(D, N, M, S) do_vqdmlah_w(N, D, M, 1, S)
+
+DO_2OP_SAT_ACC_SCALAR(vqdmlahb, 1, int8_t, DO_VQDMLAH_B)
+DO_2OP_SAT_ACC_SCALAR(vqdmlahh, 2, int16_t, DO_VQDMLAH_H)
+DO_2OP_SAT_ACC_SCALAR(vqdmlahw, 4, int32_t, DO_VQDMLAH_W)
+DO_2OP_SAT_ACC_SCALAR(vqrdmlahb, 1, int8_t, DO_VQRDMLAH_B)
+DO_2OP_SAT_ACC_SCALAR(vqrdmlahh, 2, int16_t, DO_VQRDMLAH_H)
+DO_2OP_SAT_ACC_SCALAR(vqrdmlahw, 4, int32_t, DO_VQRDMLAH_W)
+
+DO_2OP_SAT_ACC_SCALAR(vqdmlashb, 1, int8_t, DO_VQDMLASH_B)
+DO_2OP_SAT_ACC_SCALAR(vqdmlashh, 2, int16_t, DO_VQDMLASH_H)
+DO_2OP_SAT_ACC_SCALAR(vqdmlashw, 4, int32_t, DO_VQDMLASH_W)
+DO_2OP_SAT_ACC_SCALAR(vqrdmlashb, 1, int8_t, DO_VQRDMLASH_B)
+DO_2OP_SAT_ACC_SCALAR(vqrdmlashh, 2, int16_t, DO_VQRDMLASH_H)
+DO_2OP_SAT_ACC_SCALAR(vqrdmlashw, 4, int32_t, DO_VQRDMLASH_W)
+
 /* Vector by scalar plus vector */
 #define DO_VMLA(D, N, M) ((N) * (M) + (D))
 
diff --git a/target/arm/translate-mve.c b/target/arm/translate-mve.c
index 650f3b95edf..f8b34c9ef36 100644
--- a/target/arm/translate-mve.c
+++ b/target/arm/translate-mve.c
@@ -624,6 +624,10 @@ DO_2OP_SCALAR(VMLA_S, vmlas)
 DO_2OP_SCALAR(VMLA_U, vmlau)
 DO_2OP_SCALAR(VMLAS_S, vmlass)
 DO_2OP_SCALAR(VMLAS_U, vmlasu)
+DO_2OP_SCALAR(VQDMLAH, vqdmlah)
+DO_2OP_SCALAR(VQRDMLAH, vqrdmlah)
+DO_2OP_SCALAR(VQDMLASH, vqdmlash)
+DO_2OP_SCALAR(VQRDMLASH, vqrdmlash)
 
 static bool trans_VQDMULLB_scalar(DisasContext *s, arg_2scalar *a)
 {
-- 
2.20.1