RISC-V: Adding T-Head multiply-accumulate instructions

This patch adds support for the T-Head MAC instructions.
The patch uses the T-Head specific decoder and translation.

Co-developed-by: LIU Zhiwei <zhiwei_liu@linux.alibaba.com>
Reviewed-by: Alistair Francis <alistair.francis@wdc.com>
Signed-off-by: Christoph Müllner <christoph.muellner@vrull.eu>
Message-Id: <20230131202013.2541053-8-christoph.muellner@vrull.eu>
Signed-off-by: Alistair Francis <alistair.francis@wdc.com>

target/riscv/cpu.c

@@ -114,6 +114,7 @@ static const struct isa_ext_data isa_edata_arr[] = {
     ISA_EXT_DATA_ENTRY(xtheadbs, true, PRIV_VERSION_1_11_0, ext_xtheadbs),
     ISA_EXT_DATA_ENTRY(xtheadcmo, true, PRIV_VERSION_1_11_0, ext_xtheadcmo),
     ISA_EXT_DATA_ENTRY(xtheadcondmov, true, PRIV_VERSION_1_11_0, ext_xtheadcondmov),
+    ISA_EXT_DATA_ENTRY(xtheadmac, true, PRIV_VERSION_1_11_0, ext_xtheadmac),
     ISA_EXT_DATA_ENTRY(xtheadsync, true, PRIV_VERSION_1_11_0, ext_xtheadsync),
     ISA_EXT_DATA_ENTRY(xventanacondops, true, PRIV_VERSION_1_12_0, ext_XVentanaCondOps),
 };
@@ -1099,6 +1100,7 @@ static Property riscv_cpu_extensions[] = {
     DEFINE_PROP_BOOL("xtheadbs", RISCVCPU, cfg.ext_xtheadbs, false),
     DEFINE_PROP_BOOL("xtheadcmo", RISCVCPU, cfg.ext_xtheadcmo, false),
     DEFINE_PROP_BOOL("xtheadcondmov", RISCVCPU, cfg.ext_xtheadcondmov, false),
+    DEFINE_PROP_BOOL("xtheadmac", RISCVCPU, cfg.ext_xtheadmac, false),
     DEFINE_PROP_BOOL("xtheadsync", RISCVCPU, cfg.ext_xtheadsync, false),
     DEFINE_PROP_BOOL("xventanacondops", RISCVCPU, cfg.ext_XVentanaCondOps, false),
 

target/riscv/cpu.h

@@ -478,6 +478,7 @@ struct RISCVCPUConfig {
     bool ext_xtheadbs;
     bool ext_xtheadcmo;
     bool ext_xtheadcondmov;
+    bool ext_xtheadmac;
     bool ext_xtheadsync;
     bool ext_XVentanaCondOps;
 

target/riscv/insn_trans/trans_xthead.c.inc

@@ -46,6 +46,12 @@
     }                                            \
 } while (0)
 
+#define REQUIRE_XTHEADMAC(ctx) do {              \
+    if (!ctx->cfg_ptr->ext_xtheadmac) {          \
+        return false;                            \
+    }                                            \
+} while (0)
+
 #define REQUIRE_XTHEADSYNC(ctx) do {             \
     if (!ctx->cfg_ptr->ext_xtheadsync) {         \
         return false;                            \
@@ -299,6 +305,83 @@ static bool trans_th_mvnez(DisasContext *ctx, arg_th_mveqz *a)
     return gen_th_condmove(ctx, a, TCG_COND_NE);
 }
 
+/* XTheadMac */
+
+static bool gen_th_mac(DisasContext *ctx, arg_r *a,
+                       void (*accumulate_func)(TCGv, TCGv, TCGv),
+                       void (*extend_operand_func)(TCGv, TCGv))
+{
+    TCGv dest = dest_gpr(ctx, a->rd);
+    TCGv src0 = get_gpr(ctx, a->rd, EXT_NONE);
+    TCGv src1 = get_gpr(ctx, a->rs1, EXT_NONE);
+    TCGv src2 = get_gpr(ctx, a->rs2, EXT_NONE);
+    TCGv tmp = tcg_temp_new();
+
+    if (extend_operand_func) {
+        TCGv tmp2 = tcg_temp_new();
+        extend_operand_func(tmp, src1);
+        extend_operand_func(tmp2, src2);
+        tcg_gen_mul_tl(tmp, tmp, tmp2);
+        tcg_temp_free(tmp2);
+    } else {
+        tcg_gen_mul_tl(tmp, src1, src2);
+    }
+
+    accumulate_func(dest, src0, tmp);
+    gen_set_gpr(ctx, a->rd, dest);
+    tcg_temp_free(tmp);
+
+    return true;
+}
+
+/* th.mula: "rd = rd + rs1 * rs2" */
+static bool trans_th_mula(DisasContext *ctx, arg_th_mula *a)
+{
+    REQUIRE_XTHEADMAC(ctx);
+    return gen_th_mac(ctx, a, tcg_gen_add_tl, NULL);
+}
+
+/* th.mulah: "rd = sext.w(rd + sext.w(rs1[15:0]) * sext.w(rs2[15:0]))" */
+static bool trans_th_mulah(DisasContext *ctx, arg_th_mulah *a)
+{
+    REQUIRE_XTHEADMAC(ctx);
+    ctx->ol = MXL_RV32;
+    return gen_th_mac(ctx, a, tcg_gen_add_tl, tcg_gen_ext16s_tl);
+}
+
+/* th.mulaw: "rd = sext.w(rd + rs1 * rs2)" */
+static bool trans_th_mulaw(DisasContext *ctx, arg_th_mulaw *a)
+{
+    REQUIRE_XTHEADMAC(ctx);
+    REQUIRE_64BIT(ctx);
+    ctx->ol = MXL_RV32;
+    return gen_th_mac(ctx, a, tcg_gen_add_tl, NULL);
+}
+
+/* th.muls: "rd = rd - rs1 * rs2" */
+static bool trans_th_muls(DisasContext *ctx, arg_th_muls *a)
+{
+    REQUIRE_XTHEADMAC(ctx);
+    return gen_th_mac(ctx, a, tcg_gen_sub_tl, NULL);
+}
+
+/* th.mulsh: "rd = sext.w(rd - sext.w(rs1[15:0]) * sext.w(rs2[15:0]))" */
+static bool trans_th_mulsh(DisasContext *ctx, arg_th_mulsh *a)
+{
+    REQUIRE_XTHEADMAC(ctx);
+    ctx->ol = MXL_RV32;
+    return gen_th_mac(ctx, a, tcg_gen_sub_tl, tcg_gen_ext16s_tl);
+}
+
+/* th.mulsw: "rd = sext.w(rd - rs1 * rs2)" */
+static bool trans_th_mulsw(DisasContext *ctx, arg_th_mulsw *a)
+{
+    REQUIRE_XTHEADMAC(ctx);
+    REQUIRE_64BIT(ctx);
+    ctx->ol = MXL_RV32;
+    return gen_th_mac(ctx, a, tcg_gen_sub_tl, NULL);
+}
+
 /* XTheadSync */
 
 static bool trans_th_sfence_vmas(DisasContext *ctx, arg_th_sfence_vmas *a)

target/riscv/translate.c

@@ -134,7 +134,8 @@ static bool has_xthead_p(DisasContext *ctx  __attribute__((__unused__)))
 {
     return ctx->cfg_ptr->ext_xtheadba || ctx->cfg_ptr->ext_xtheadbb ||
            ctx->cfg_ptr->ext_xtheadbs || ctx->cfg_ptr->ext_xtheadcmo ||
-           ctx->cfg_ptr->ext_xtheadcondmov || ctx->cfg_ptr->ext_xtheadsync;
+           ctx->cfg_ptr->ext_xtheadcondmov || ctx->cfg_ptr->ext_xtheadmac ||
+           ctx->cfg_ptr->ext_xtheadsync;
 }
 
 #define MATERIALISE_EXT_PREDICATE(ext)  \

target/riscv/xthead.decode

@@ -88,6 +88,14 @@ th_l2cache_iall  0000000 10110 00000 000 00000 0001011
 th_mveqz         0100000 ..... .....  001 ..... 0001011 @r
 th_mvnez         0100001 ..... .....  001 ..... 0001011 @r
 
+# XTheadMac
+th_mula          00100 00 ..... ..... 001 ..... 0001011 @r
+th_mulah         00101 00 ..... ..... 001 ..... 0001011 @r
+th_mulaw         00100 10 ..... ..... 001 ..... 0001011 @r
+th_muls          00100 01 ..... ..... 001 ..... 0001011 @r
+th_mulsh         00101 01 ..... ..... 001 ..... 0001011 @r
+th_mulsw         00100 11 ..... ..... 001 ..... 0001011 @r
+
 # XTheadSync
 th_sfence_vmas   0000010 ..... ..... 000 00000 0001011 @rs2_s
 th_sync          0000000 11000 00000 000 00000 0001011