target/arm: Implement scalar float32 to bfloat16 conversion

This is the 64-bit BFCVT and the 32-bit VCVT{B,T}.BF16.F32.

Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210525225817.400336-4-richard.henderson@linaro.org
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>

target/arm/helper.h

@@ -143,6 +143,7 @@ DEF_HELPER_3(vfp_cmped, void, f64, f64, env)
 
 DEF_HELPER_2(vfp_fcvtds, f64, f32, env)
 DEF_HELPER_2(vfp_fcvtsd, f32, f64, env)
+DEF_HELPER_FLAGS_2(bfcvt, TCG_CALL_NO_RWG, i32, f32, ptr)
 
 DEF_HELPER_2(vfp_uitoh, f16, i32, ptr)
 DEF_HELPER_2(vfp_uitos, f32, i32, ptr)

target/arm/translate-a64.c

@@ -6280,6 +6280,9 @@ static void handle_fp_1src_single(DisasContext *s, int opcode, int rd, int rn)
     case 0x3: /* FSQRT */
         gen_helper_vfp_sqrts(tcg_res, tcg_op, cpu_env);
         goto done;
+    case 0x6: /* BFCVT */
+        gen_fpst = gen_helper_bfcvt;
+        break;
     case 0x8: /* FRINTN */
     case 0x9: /* FRINTP */
     case 0xa: /* FRINTM */
@@ -6557,6 +6560,22 @@ static void disas_fp_1src(DisasContext *s, uint32_t insn)
         }
         break;
 
+    case 0x6:
+        switch (type) {
+        case 1: /* BFCVT */
+            if (!dc_isar_feature(aa64_bf16, s)) {
+                goto do_unallocated;
+            }
+            if (!fp_access_check(s)) {
+                return;
+            }
+            handle_fp_1src_single(s, opcode, rd, rn);
+            break;
+        default:
+            goto do_unallocated;
+        }
+        break;
+
     default:
     do_unallocated:
         unallocated_encoding(s);

target/arm/translate-vfp.c

@@ -3085,6 +3085,30 @@ static bool trans_VCVT_f64_f16(DisasContext *s, arg_VCVT_f64_f16 *a)
     return true;
 }
 
+static bool trans_VCVT_b16_f32(DisasContext *s, arg_VCVT_b16_f32 *a)
+{
+    TCGv_ptr fpst;
+    TCGv_i32 tmp;
+
+    if (!dc_isar_feature(aa32_bf16, s)) {
+        return false;
+    }
+
+    if (!vfp_access_check(s)) {
+        return true;
+    }
+
+    fpst = fpstatus_ptr(FPST_FPCR);
+    tmp = tcg_temp_new_i32();
+
+    vfp_load_reg32(tmp, a->vm);
+    gen_helper_bfcvt(tmp, tmp, fpst);
+    tcg_gen_st16_i32(tmp, cpu_env, vfp_f16_offset(a->vd, a->t));
+    tcg_temp_free_ptr(fpst);
+    tcg_temp_free_i32(tmp);
+    return true;
+}
+
 static bool trans_VCVT_f16_f32(DisasContext *s, arg_VCVT_f16_f32 *a)
 {
     TCGv_ptr fpst;

target/arm/vfp.decode

@@ -205,6 +205,8 @@ VCVT_f64_f16 ---- 1110 1.11 0010 .... 1011 t:1 1.0 .... \
 
 # VCVTB and VCVTT to f16: Vd format is always vd_sp;
 # Vm format depends on size bit
+VCVT_b16_f32 ---- 1110 1.11 0011 .... 1001 t:1 1.0 .... \
+             vd=%vd_sp vm=%vm_sp
 VCVT_f16_f32 ---- 1110 1.11 0011 .... 1010 t:1 1.0 .... \
              vd=%vd_sp vm=%vm_sp
 VCVT_f16_f64 ---- 1110 1.11 0011 .... 1011 t:1 1.0 .... \

target/arm/vfp_helper.c

@@ -411,6 +411,11 @@ float32 VFP_HELPER(fcvts, d)(float64 x, CPUARMState *env)
     return float64_to_float32(x, &env->vfp.fp_status);
 }
 
+uint32_t HELPER(bfcvt)(float32 x, void *status)
+{
+    return float32_to_bfloat16(x, status);
+}
+
 /*
  * VFP3 fixed point conversion. The AArch32 versions of fix-to-float
  * must always round-to-nearest; the AArch64 ones honour the FPSCR