PPC/S390 [simd]: Optimize I8x16Shuffle on codegen

Change-Id: I264a06924d8ffcb8cce1febe9bde18280edeaea0
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2896273Reviewed-by: Junliang Yan <junyan@redhat.com>
Commit-Queue: Milad Fa <mfarazma@redhat.com>
Cr-Commit-Position: refs/heads/master@{#74603}

src/compiler/backend/ppc/code-generator-ppc.cc

@@ -3108,15 +3108,14 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
       break;
     }
     case kPPC_I8x16Shuffle: {
-      constexpr int lane_width_in_bytes = 8;
       Simd128Register dst = i.OutputSimd128Register(),
                       src0 = i.InputSimd128Register(0),
                       src1 = i.InputSimd128Register(1);
-      __ mov(r0, Operand(make_uint64(i.InputUint32(3), i.InputUint32(2))));
-      __ mov(ip, Operand(make_uint64(i.InputUint32(5), i.InputUint32(4))));
-      __ mtvsrd(kScratchSimd128Reg, r0);
-      __ mtvsrd(dst, ip);
-      __ vinsertd(dst, kScratchSimd128Reg, Operand(1 * lane_width_in_bytes));
+      uint64_t low = make_uint64(i.InputUint32(3), i.InputUint32(2));
+      uint64_t high = make_uint64(i.InputUint32(5), i.InputUint32(4));
+      __ mov(r0, Operand(low));
+      __ mov(ip, Operand(high));
+      __ mtvsrdd(dst, ip, r0);
       __ vperm(dst, src0, src1, dst);
       break;
     }

src/compiler/backend/s390/code-generator-s390.cc

@@ -3512,15 +3512,12 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
       Simd128Register dst = i.OutputSimd128Register(),
                       src0 = i.InputSimd128Register(0),
                       src1 = i.InputSimd128Register(1);
-      int32_t k8x16_indices[] = {i.InputInt32(2), i.InputInt32(3),
-                                 i.InputInt32(4), i.InputInt32(5)};
-      // create 2 * 8 byte inputs indicating new indices
-      for (int i = 0, j = 0; i < 2; i++, j = +2) {
-        __ mov(i < 1 ? ip : r0, Operand(k8x16_indices[j]));
-        __ iihf(i < 1 ? ip : r0, Operand(k8x16_indices[j + 1]));
-      }
-      __ vlvgp(kScratchDoubleReg, r0, ip);
-      __ vperm(dst, src0, src1, kScratchDoubleReg, Condition(0), Condition(0));
+      uint64_t low = make_uint64(i.InputUint32(3), i.InputUint32(2));
+      uint64_t high = make_uint64(i.InputUint32(5), i.InputUint32(4));
+      __ mov(r0, Operand(low));
+      __ mov(ip, Operand(high));
+      __ vlvgp(dst, ip, r0);
+      __ vperm(dst, src0, src1, dst, Condition(0), Condition(0));
       break;
     }
     case kS390_I8x16Swizzle: {