[PATCH v7 16/75] target/riscv: rvv:1.0: add translation-time nan-box helper function

[frank.chang@sifive.com]

From: Frank Chang <frank.chang@sifive.com>

* Add fp16 nan-box check generator function, if a 16-bit input is not
  properly nanboxed, then the input is replaced with the default qnan.
* Add do_nanbox() helper function to utilize gen_check_nanbox_X() to
  generate the NaN-boxed floating-point values based on SEW setting.
* Apply nanbox helper in opfvf_trans().

Signed-off-by: Frank Chang <frank.chang@sifive.com>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Alistair Francis <alistair.francis@wdc.com>
---
 target/riscv/insn_trans/trans_rvv.c.inc | 35 ++++++++++++++++++++++++-
 target/riscv/translate.c                | 10 +++++++
 2 files changed, 44 insertions(+), 1 deletion(-)

diff --git a/target/riscv/insn_trans/trans_rvv.c.inc b/target/riscv/insn_trans/trans_rvv.c.inc
index 18a1c409fcf..ccfa93cf2f8 100644
--- a/target/riscv/insn_trans/trans_rvv.c.inc
+++ b/target/riscv/insn_trans/trans_rvv.c.inc
@@ -2100,6 +2100,33 @@ GEN_OPIVI_NARROW_TRANS(vnclip_vi, IMM_ZX, vnclip_vx)
 /*
  *** Vector Float Point Arithmetic Instructions
  */
+
+/*
+ * As RVF-only cpus always have values NaN-boxed to 64-bits,
+ * RVF and RVD can be treated equally.
+ * We don't have to deal with the cases of: SEW > FLEN.
+ *
+ * If SEW < FLEN, check whether input fp register is a valid
+ * NaN-boxed value, in which case the least-significant SEW bits
+ * of the f regsiter are used, else the canonical NaN value is used.
+ */
+static void do_nanbox(DisasContext *s, TCGv_i64 out, TCGv_i64 in)
+{
+    switch (s->sew) {
+    case 1:
+        gen_check_nanbox_h(out, in);
+        break;
+    case 2:
+        gen_check_nanbox_s(out, in);
+        break;
+    case 3:
+        tcg_gen_mov_i64(out, in);
+        break;
+    default:
+        g_assert_not_reached();
+    }
+}
+
 /* Vector Single-Width Floating-Point Add/Subtract Instructions */
 
 /*
@@ -2152,6 +2179,7 @@ static bool opfvf_trans(uint32_t vd, uint32_t rs1, uint32_t vs2,
 {
     TCGv_ptr dest, src2, mask;
     TCGv_i32 desc;
+    TCGv_i64 t1;
 
     TCGLabel *over = gen_new_label();
     tcg_gen_brcondi_tl(TCG_COND_EQ, cpu_vl, 0, over);
@@ -2165,12 +2193,17 @@ static bool opfvf_trans(uint32_t vd, uint32_t rs1, uint32_t vs2,
     tcg_gen_addi_ptr(src2, cpu_env, vreg_ofs(s, vs2));
     tcg_gen_addi_ptr(mask, cpu_env, vreg_ofs(s, 0));
 
-    fn(dest, mask, cpu_fpr[rs1], src2, cpu_env, desc);
+    /* NaN-box f[rs1] */
+    t1 = tcg_temp_new_i64();
+    do_nanbox(s, t1, cpu_fpr[rs1]);
+
+    fn(dest, mask, t1, src2, cpu_env, desc);
 
     tcg_temp_free_ptr(dest);
     tcg_temp_free_ptr(mask);
     tcg_temp_free_ptr(src2);
     tcg_temp_free_i32(desc);
+    tcg_temp_free_i64(t1);
     mark_vs_dirty(s);
     gen_set_label(over);
     return true;
diff --git a/target/riscv/translate.c b/target/riscv/translate.c
index 75ed94c802b..d9a794d71e8 100644
--- a/target/riscv/translate.c
+++ b/target/riscv/translate.c
@@ -123,6 +123,16 @@ static void gen_nanbox_s(TCGv_i64 out, TCGv_i64 in)
  *
  * Here, the result is always nan-boxed, even the canonical nan.
  */
+static void gen_check_nanbox_h(TCGv_i64 out, TCGv_i64 in)
+{
+    TCGv_i64 t_max = tcg_const_i64(0xffffffffffff0000ull);
+    TCGv_i64 t_nan = tcg_const_i64(0xffffffffffff7e00ull);
+
+    tcg_gen_movcond_i64(TCG_COND_GEU, out, in, t_max, in, t_nan);
+    tcg_temp_free_i64(t_max);
+    tcg_temp_free_i64(t_nan);
+}
+
 static void gen_check_nanbox_s(TCGv_i64 out, TCGv_i64 in)
 {
     TCGv_i64 t_max = tcg_const_i64(0xffffffff00000000ull);
-- 
2.17.1