[PULL 16/31] tcg/i386: Implement cmpsel_vec with avx512 insns

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

The avx512 vpblendm* instructions exactly implement cmpsel,
using a predicate input.  Of course this matches nicely with
the avx512 predicate comparison instructions.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 tcg/i386/tcg-target.c.inc | 44 ++++++++++++++++++++++++++++++++++++++-
 1 file changed, 43 insertions(+), 1 deletion(-)

diff --git a/tcg/i386/tcg-target.c.inc b/tcg/i386/tcg-target.c.inc
index f94a2a2385..d473dc7a5e 100644
--- a/tcg/i386/tcg-target.c.inc
+++ b/tcg/i386/tcg-target.c.inc
@@ -413,6 +413,10 @@ static bool tcg_target_const_match(int64_t val, int ct,
 #define OPC_UD2         (0x0b | P_EXT)
 #define OPC_VPBLENDD    (0x02 | P_EXT3A | P_DATA16)
 #define OPC_VPBLENDVB   (0x4c | P_EXT3A | P_DATA16)
+#define OPC_VPBLENDMB   (0x66 | P_EXT38 | P_DATA16 | P_EVEX)
+#define OPC_VPBLENDMW   (0x66 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
+#define OPC_VPBLENDMD   (0x64 | P_EXT38 | P_DATA16 | P_EVEX)
+#define OPC_VPBLENDMQ   (0x64 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
 #define OPC_VPCMPB      (0x3f | P_EXT3A | P_DATA16 | P_EVEX)
 #define OPC_VPCMPUB     (0x3e | P_EXT3A | P_DATA16 | P_EVEX)
 #define OPC_VPCMPW      (0x3f | P_EXT3A | P_DATA16 | P_VEXW | P_EVEX)
@@ -738,6 +742,16 @@ static void tcg_out_vex_modrm_type(TCGContext *s, int opc,
     tcg_out_vex_modrm(s, opc, r, v, rm);
 }
 
+static void tcg_out_evex_modrm_type(TCGContext *s, int opc, int r, int v,
+                                    int rm, int aaa, bool z, TCGType type)
+{
+    if (type == TCG_TYPE_V256) {
+        opc |= P_VEXL;
+    }
+    tcg_out_evex_opc(s, opc, r, v, rm, 0, aaa, z);
+    tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm));
+}
+
 /* Output an opcode with a full "rm + (index<<shift) + offset" address mode.
    We handle either RM and INDEX missing with a negative value.  In 64-bit
    mode for absolute addresses, ~RM is the size of the immediate operand
@@ -3183,11 +3197,39 @@ static void tcg_out_cmp_vec(TCGContext *s, TCGType type, unsigned vece,
     }
 }
 
+static void tcg_out_cmpsel_vec_k1(TCGContext *s, TCGType type, unsigned vece,
+                                  TCGReg v0, TCGReg c1, TCGReg c2,
+                                  TCGReg v3, TCGReg v4, TCGCond cond)
+{
+    static const int vpblendm_insn[] = {
+        OPC_VPBLENDMB, OPC_VPBLENDMW, OPC_VPBLENDMD, OPC_VPBLENDMQ
+    };
+    bool z = false;
+
+    /* Swap to place constant in V4 to take advantage of zero-masking. */
+    if (!v3) {
+        z = true;
+        v3 = v4;
+        cond = tcg_invert_cond(cond);
+    }
+
+    tcg_out_cmp_vec_k1(s, type, vece, c1, c2, cond);
+    tcg_out_evex_modrm_type(s, vpblendm_insn[vece], v0, v4, v3,
+                            /* k1 */1, z, type);
+}
+
 static void tcg_out_cmpsel_vec(TCGContext *s, TCGType type, unsigned vece,
                                TCGReg v0, TCGReg c1, TCGReg c2,
                                TCGReg v3, TCGReg v4, TCGCond cond)
 {
-    bool inv = tcg_out_cmp_vec_noinv(s, type, vece, TCG_TMP_VEC, c1, c2, cond);
+    bool inv;
+
+    if (vece <= MO_16 ? have_avx512bw : have_avx512vl) {
+        tcg_out_cmpsel_vec_k1(s, type, vece, v0, c1, c2, v3, v4, cond);
+        return;
+    }
+
+    inv = tcg_out_cmp_vec_noinv(s, type, vece, TCG_TMP_VEC, c1, c2, cond);
 
     /*
      * Since XMM0 is 16, the only way we get 0 into V3
-- 
2.43.0