target/arm: Implement bfloat widening fma (indexed)

This is BFMLAL{B,T} for both AArch64 AdvSIMD and SVE,
and VFMA{B,T}.BF16 for AArch32 NEON.

Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210525225817.400336-11-richard.henderson@linaro.org
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>

7 files changed, 82 insertions, 1 deletions

diff --git a/target/arm/helper.h b/target/arm/helper.h
index 36b3c9d..dc6eb96 100644
--- a/target/arm/helper.h
+++ b/target/arm/helper.h
@@ -1012,6 +1012,8 @@ DEF_HELPER_FLAGS_5(gvec_bfmmla, TCG_CALL_NO_RWG,
 
 DEF_HELPER_FLAGS_6(gvec_bfmlal, TCG_CALL_NO_RWG,
                    void, ptr, ptr, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_6(gvec_bfmlal_idx, TCG_CALL_NO_RWG,
+                   void, ptr, ptr, ptr, ptr, ptr, i32)
 
 #ifdef TARGET_AARCH64
 #include "helper-a64.h"
diff --git a/target/arm/neon-shared.decode b/target/arm/neon-shared.decode
index b61addd..df80e6e 100644
--- a/target/arm/neon-shared.decode
+++ b/target/arm/neon-shared.decode
@@ -95,3 +95,5 @@ VFML_scalar    1111 1110 0 . 0 s:1 .... .... 1000 . 0 . 1 index:1 ... \
                rm=%vfml_scalar_q0_rm vn=%vn_sp vd=%vd_dp q=0
 VFML_scalar    1111 1110 0 . 0 s:1 .... .... 1000 . 1 . 1 . rm:3 \
                index=%vfml_scalar_q1_index vn=%vn_dp vd=%vd_dp q=1
+VFMA_b16_scal  1111 1110 0.11 .... .... 1000 . q:1 . 1 . vm:3 \
+               index=%vfml_scalar_q1_index vn=%vn_dp vd=%vd_dp
diff --git a/target/arm/sve.decode b/target/arm/sve.decode
index 5281164..a62c169 100644
--- a/target/arm/sve.decode
+++ b/target/arm/sve.decode
@@ -1638,6 +1638,8 @@ FMLALB_zzxw     01100100 10 1 ..... 0100.0 ..... .....     @rrxr_3a esz=2
 FMLALT_zzxw     01100100 10 1 ..... 0100.1 ..... .....     @rrxr_3a esz=2
 FMLSLB_zzxw     01100100 10 1 ..... 0110.0 ..... .....     @rrxr_3a esz=2
 FMLSLT_zzxw     01100100 10 1 ..... 0110.1 ..... .....     @rrxr_3a esz=2
+BFMLALB_zzxw    01100100 11 1 ..... 0100.0 ..... .....     @rrxr_3a esz=2
+BFMLALT_zzxw    01100100 11 1 ..... 0100.1 ..... .....     @rrxr_3a esz=2
 
 ### SVE2 floating-point bfloat16 dot-product (indexed)
 BFDOT_zzxz      01100100 01 1 ..... 010000 ..... .....     @rrxr_2 esz=2
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index 8dcb15a..8713dfe 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -13472,18 +13472,27 @@ static void disas_simd_indexed(DisasContext *s, uint32_t insn)
                 unallocated_encoding(s);
                 return;
             }
+            size = MO_32;
             break;
         case 1: /* BFDOT */
             if (is_scalar || !dc_isar_feature(aa64_bf16, s)) {
                 unallocated_encoding(s);
                 return;
             }
+            size = MO_32;
+            break;
+        case 3: /* BFMLAL{B,T} */
+            if (is_scalar || !dc_isar_feature(aa64_bf16, s)) {
+                unallocated_encoding(s);
+                return;
+            }
+            /* can't set is_fp without other incorrect size checks */
+            size = MO_16;
             break;
         default:
             unallocated_encoding(s);
             return;
         }
-        size = MO_32;
         break;
     case 0x11: /* FCMLA #0 */
     case 0x13: /* FCMLA #90 */
@@ -13613,6 +13622,10 @@ static void disas_simd_indexed(DisasContext *s, uint32_t insn)
             gen_gvec_op4_ool(s, is_q, rd, rn, rm, rd, index,
                              gen_helper_gvec_usdot_idx_b);
             return;
+        case 3: /* BFMLAL{B,T} */
+            gen_gvec_op4_fpst(s, 1, rd, rn, rm, rd, 0, (index << 1) | is_q,
+                              gen_helper_gvec_bfmlal_idx);
+            return;
         }
         g_assert_not_reached();
     case 0x11: /* FCMLA #0 */
diff --git a/target/arm/translate-neon.c b/target/arm/translate-neon.c
index 4d0c249..633fef3 100644
--- a/target/arm/translate-neon.c
+++ b/target/arm/translate-neon.c
@@ -4144,3 +4144,13 @@ static bool trans_VFMA_b16(DisasContext *s, arg_VFMA_b16 *a)
     return do_neon_ddda_fpst(s, 7, a->vd, a->vn, a->vm, a->q, FPST_STD,
                              gen_helper_gvec_bfmlal);
 }
+
+static bool trans_VFMA_b16_scal(DisasContext *s, arg_VFMA_b16_scal *a)
+{
+    if (!dc_isar_feature(aa32_bf16, s)) {
+        return false;
+    }
+    return do_neon_ddda_fpst(s, 6, a->vd, a->vn, a->vm,
+                             (a->index << 1) | a->q, FPST_STD,
+                             gen_helper_gvec_bfmlal_idx);
+}
diff --git a/target/arm/translate-sve.c b/target/arm/translate-sve.c
index ba8f5d7..46210eb 100644
--- a/target/arm/translate-sve.c
+++ b/target/arm/translate-sve.c
@@ -8719,3 +8719,33 @@ static bool trans_BFMLALT_zzzw(DisasContext *s, arg_rrrr_esz *a)
 {
     return do_BFMLAL_zzzw(s, a, true);
 }
+
+static bool do_BFMLAL_zzxw(DisasContext *s, arg_rrxr_esz *a, bool sel)
+{
+    if (!dc_isar_feature(aa64_sve_bf16, s)) {
+        return false;
+    }
+    if (sve_access_check(s)) {
+        TCGv_ptr status = fpstatus_ptr(FPST_FPCR);
+        unsigned vsz = vec_full_reg_size(s);
+
+        tcg_gen_gvec_4_ptr(vec_full_reg_offset(s, a->rd),
+                           vec_full_reg_offset(s, a->rn),
+                           vec_full_reg_offset(s, a->rm),
+                           vec_full_reg_offset(s, a->ra),
+                           status, vsz, vsz, (a->index << 1) | sel,
+                           gen_helper_gvec_bfmlal_idx);
+        tcg_temp_free_ptr(status);
+    }
+    return true;
+}
+
+static bool trans_BFMLALB_zzxw(DisasContext *s, arg_rrxr_esz *a)
+{
+    return do_BFMLAL_zzxw(s, a, false);
+}
+
+static bool trans_BFMLALT_zzxw(DisasContext *s, arg_rrxr_esz *a)
+{
+    return do_BFMLAL_zzxw(s, a, true);
+}
diff --git a/target/arm/vec_helper.c b/target/arm/vec_helper.c
index d82736b..5862f18 100644
--- a/target/arm/vec_helper.c
+++ b/target/arm/vec_helper.c
@@ -2528,3 +2528,25 @@ void HELPER(gvec_bfmlal)(void *vd, void *vn, void *vm, void *va,
     }
     clear_tail(d, opr_sz, simd_maxsz(desc));
 }
+
+void HELPER(gvec_bfmlal_idx)(void *vd, void *vn, void *vm,
+                             void *va, void *stat, uint32_t desc)
+{
+    intptr_t i, j, opr_sz = simd_oprsz(desc);
+    intptr_t sel = extract32(desc, SIMD_DATA_SHIFT, 1);
+    intptr_t index = extract32(desc, SIMD_DATA_SHIFT + 1, 3);
+    intptr_t elements = opr_sz / 4;
+    intptr_t eltspersegment = MIN(16 / 4, elements);
+    float32 *d = vd, *a = va;
+    bfloat16 *n = vn, *m = vm;
+
+    for (i = 0; i < elements; i += eltspersegment) {
+        float32 m_idx = m[H2(2 * i + index)] << 16;
+
+        for (j = i; j < i + eltspersegment; j++) {
+            float32 n_j = n[H2(2 * j + sel)] << 16;
+            d[H4(j)] = float32_muladd(n_j, m_idx, a[H4(j)], 0, stat);
+        }
+    }
+    clear_tail(d, opr_sz, simd_maxsz(desc));
+}