[x64] Add AVX512 optimizations for PERMUTE_V128

Uses the single-instruction AVX512 `vperm*` instructions to accelerate
the `INT8_TYPE` and `INT16_TYPE` permutation opcodes.

The `INT8_TYPE` is accelerated using `AVX512VBMI` subset of AVX512.
Available since Icelake(Intel) and Zen4(AMD).

Author: Wunkolo

src/xenia/cpu/backend/x64/x64_seq_vector.cc

@@ -594,7 +594,7 @@ struct VECTOR_ADD
 
                     e.vpsrad(e.xmm2, e.xmm1, 31);
                     e.vpxord(e.xmm2, e.xmm2, e.GetXmmConstPtr(XMMSignMaskI32));
-                    e.vpblendmd(dest | saturate, e.xmm1, e.xmm2); 
+                    e.vpblendmd(dest | saturate, e.xmm1, e.xmm2);
                     return;
                   }
 
@@ -1776,7 +1776,23 @@ struct PERMUTE_V128
         } else {
           e.vxorps(e.xmm0, i.src1, e.GetXmmConstPtr(XMMSwapWordMask));
         }
+
+        if (e.IsFeatureEnabled(kX64EmitAVX512Ortho | kX64EmitAVX512VBMI)) {
+          Xmm table_lo = e.xmm1;
+          if (i.src2.is_constant) {
+            e.LoadConstantXmm(table_lo, i.src2.constant());
+          } else {
+            table_lo = i.src2;
+          }
+          Opmask zeroes = e.k1;
+          // _mm_cmple_epu8_mask
+          e.vpcmpub(zeroes, e.xmm0, e.GetXmmConstPtr(XMMPermuteControl15), 2);
+          e.vpermb(i.dest.reg() | zeroes | e.T_z, e.xmm0, table_lo);
+          return;
+        }
+
         e.vpand(e.xmm0, e.GetXmmConstPtr(XMMPermuteByteMask));
+
         if (i.src2.is_constant) {
           e.LoadConstantXmm(i.dest, i.src2.constant());
           e.vpshufb(i.dest, i.dest, e.xmm0);
@@ -1792,13 +1808,47 @@ struct PERMUTE_V128
       // General permute.
       // Control mask needs to be shuffled.
       // TODO(benvanik): do constants here instead of in generated code.
+      if (e.IsFeatureEnabled(kX64EmitAVX512Ortho | kX64EmitAVX512BW |
+                             kX64EmitAVX512VBMI)) {
+        Xmm table_idx = e.xmm0;
+        if (i.src1.is_constant) {
+          e.LoadConstantXmm(table_idx, i.src1.constant());
+          e.vxorps(table_idx, table_idx, e.GetXmmConstPtr(XMMSwapWordMask));
+        } else {
+          e.vxorps(table_idx, i.src1, e.GetXmmConstPtr(XMMSwapWordMask));
+        }
+
+        Xmm table_lo = e.xmm1;
+        if (i.src2.value->IsConstantZero()) {
+          e.vpxor(table_lo, table_lo);
+        } else if (i.src2.is_constant) {
+          e.LoadConstantXmm(table_lo, i.src2.constant());
+        } else {
+          table_lo = i.src2;
+        }
+
+        Xmm table_hi = e.xmm2;
+        if (i.src3.value->IsConstantZero()) {
+          e.vpxor(table_hi, table_hi);
+        } else if (i.src3.is_constant) {
+          e.LoadConstantXmm(table_hi, i.src3.constant());
+        } else {
+          table_hi = i.src3;
+        }
+
+        e.vpermi2b(table_idx, table_lo, table_hi);
+        e.vmovdqu8(i.dest, table_idx);
+        return;
+      }
+
       if (i.src1.is_constant) {
         e.LoadConstantXmm(e.xmm2, i.src1.constant());
         e.vxorps(e.xmm2, e.xmm2, e.GetXmmConstPtr(XMMSwapWordMask));
       } else {
         e.vxorps(e.xmm2, i.src1, e.GetXmmConstPtr(XMMSwapWordMask));
       }
       e.vpand(e.xmm2, e.GetXmmConstPtr(XMMPermuteByteMask));
+
       Xmm src2_shuf = e.xmm0;
       if (i.src2.value->IsConstantZero()) {
         e.vpxor(src2_shuf, src2_shuf);
@@ -1825,8 +1875,39 @@ struct PERMUTE_V128
 
   static void EmitByInt16(X64Emitter& e, const EmitArgType& i) {
     // src1 is an array of indices corresponding to positions within src2 and
-    // src3.
+    // src3
+    if (e.IsFeatureEnabled(kX64EmitAVX512Ortho | kX64EmitAVX512BW)) {
+      e.LoadConstantXmm(e.xmm1, vec128s(0x1));
+
+      Xmm table_idx = e.xmm0;
+      if (i.src1.is_constant) {
+        e.LoadConstantXmm(table_idx, i.src1.constant());
+        e.vpxord(table_idx, table_idx, e.xmm1);
+      } else {
+        e.vpxord(table_idx, i.src1, e.xmm1);
+      }
+
+      Xmm table_lo = e.xmm1;
+      if (i.src2.is_constant) {
+        e.LoadConstantXmm(table_lo, i.src2.constant());
+      } else {
+        table_lo = i.src2;
+      }
+
+      Xmm table_hi = e.xmm2;
+      if (i.src3.is_constant) {
+        e.LoadConstantXmm(table_hi, i.src3.constant());
+      } else {
+        table_hi = i.src3;
+      }
+
+      e.vpermi2w(table_idx, table_lo, table_hi);
+      e.vmovdqu8(i.dest, table_idx);
+      return;
+    }
+
     assert_true(i.src1.is_constant);
+
     vec128_t perm = (i.src1.constant() & vec128s(0xF)) ^ vec128s(0x1);
     vec128_t perm_ctrl = vec128b(0);
     for (int i = 0; i < 8; i++) {