[Qemu-devel] [PULL 08/16] tcg/i386: Support vector comparison select value

[Richard Henderson <richard.henderson@linaro.org>]

We already had backend support for this feature.  Expand the new
cmpsel opcode using vpblendb.  The combination allows us to avoid
an extra NOT for some comparison codes.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 tcg/i386/tcg-target.h     |  2 +-
 tcg/i386/tcg-target.inc.c | 39 +++++++++++++++++++++++++++++++++++----
 2 files changed, 36 insertions(+), 5 deletions(-)

diff --git a/tcg/i386/tcg-target.h b/tcg/i386/tcg-target.h
index 16a83a7f7b..928e8b87bb 100644
--- a/tcg/i386/tcg-target.h
+++ b/tcg/i386/tcg-target.h
@@ -191,7 +191,7 @@ extern bool have_avx2;
 #define TCG_TARGET_HAS_sat_vec          1
 #define TCG_TARGET_HAS_minmax_vec       1
 #define TCG_TARGET_HAS_bitsel_vec       0
-#define TCG_TARGET_HAS_cmpsel_vec       0
+#define TCG_TARGET_HAS_cmpsel_vec       -1
 
 #define TCG_TARGET_deposit_i32_valid(ofs, len) \
     (((ofs) == 0 && (len) == 8) || ((ofs) == 8 && (len) == 8) || \
diff --git a/tcg/i386/tcg-target.inc.c b/tcg/i386/tcg-target.inc.c
index b3601446cd..ffcafb1e14 100644
--- a/tcg/i386/tcg-target.inc.c
+++ b/tcg/i386/tcg-target.inc.c
@@ -3246,6 +3246,7 @@ int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
     case INDEX_op_andc_vec:
         return 1;
     case INDEX_op_cmp_vec:
+    case INDEX_op_cmpsel_vec:
         return -1;
 
     case INDEX_op_shli_vec:
@@ -3464,8 +3465,8 @@ static void expand_vec_mul(TCGType type, unsigned vece,
     }
 }
 
-static void expand_vec_cmp(TCGType type, unsigned vece, TCGv_vec v0,
-                           TCGv_vec v1, TCGv_vec v2, TCGCond cond)
+static bool expand_vec_cmp_noinv(TCGType type, unsigned vece, TCGv_vec v0,
+                                 TCGv_vec v1, TCGv_vec v2, TCGCond cond)
 {
     enum {
         NEED_SWAP = 1,
@@ -3522,11 +3523,34 @@ static void expand_vec_cmp(TCGType type, unsigned vece, TCGv_vec v0,
             tcg_temp_free_vec(t2);
         }
     }
-    if (fixup & NEED_INV) {
+    return fixup & NEED_INV;
+}
+
+static void expand_vec_cmp(TCGType type, unsigned vece, TCGv_vec v0,
+                           TCGv_vec v1, TCGv_vec v2, TCGCond cond)
+{
+    if (expand_vec_cmp_noinv(type, vece, v0, v1, v2, cond)) {
         tcg_gen_not_vec(vece, v0, v0);
     }
 }
 
+static void expand_vec_cmpsel(TCGType type, unsigned vece, TCGv_vec v0,
+                              TCGv_vec c1, TCGv_vec c2,
+                              TCGv_vec v3, TCGv_vec v4, TCGCond cond)
+{
+    TCGv_vec t = tcg_temp_new_vec(type);
+
+    if (expand_vec_cmp_noinv(type, vece, t, c1, c2, cond)) {
+        /* Invert the sense of the compare by swapping arguments.  */
+        TCGv_vec x;
+        x = v3, v3 = v4, v4 = x;
+    }
+    vec_gen_4(INDEX_op_x86_vpblendvb_vec, type, vece,
+              tcgv_vec_arg(v0), tcgv_vec_arg(v4),
+              tcgv_vec_arg(v3), tcgv_vec_arg(t));
+    tcg_temp_free_vec(t);
+}
+
 static void expand_vec_minmax(TCGType type, unsigned vece,
                               TCGCond cond, bool min,
                               TCGv_vec v0, TCGv_vec v1, TCGv_vec v2)
@@ -3551,7 +3575,7 @@ void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
 {
     va_list va;
     TCGArg a2;
-    TCGv_vec v0, v1, v2;
+    TCGv_vec v0, v1, v2, v3, v4;
 
     va_start(va, a0);
     v0 = temp_tcgv_vec(arg_temp(a0));
@@ -3578,6 +3602,13 @@ void tcg_expand_vec_op(TCGOpcode opc, TCGType type, unsigned vece,
         expand_vec_cmp(type, vece, v0, v1, v2, va_arg(va, TCGArg));
         break;
 
+    case INDEX_op_cmpsel_vec:
+        v2 = temp_tcgv_vec(arg_temp(a2));
+        v3 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
+        v4 = temp_tcgv_vec(arg_temp(va_arg(va, TCGArg)));
+        expand_vec_cmpsel(type, vece, v0, v1, v2, v3, v4, va_arg(va, TCGArg));
+        break;
+
     case INDEX_op_smin_vec:
         v2 = temp_tcgv_vec(arg_temp(a2));
         expand_vec_minmax(type, vece, TCG_COND_GT, true, v0, v1, v2);
-- 
2.17.1