AVX512 : VPUNPCK and VUNPCK implemantation

[igorb]

Summary

Implemented encoding and intrinsics of the following instructions

vunpckhps/pd, vunpcklps/pd, 
vpunpcklbw, vpunpckhbw, vpunpcklwd, vpunpckhwd, vpunpckldq, vpunpckhdq, vpunpcklqdq, vpunpckhqdq

Added tests for intrinsics and encoding.

[delena]

LGTM

include/llvm/IR/IntrinsicsX86.td

let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_avx512_mask_packusdw_256 : GCCBuiltin<"__builtin_ia32_packusdw256_mask">,
               Intrinsic<[llvm_v16i16_ty], [llvm_v8i32_ty, llvm_v8i32_ty,
                          llvm_v16i16_ty, llvm_i16_ty], [IntrNoMem]>;
   def int_x86_avx512_mask_packusdw_512 : GCCBuiltin<"__builtin_ia32_packusdw512_mask">,
               Intrinsic<[llvm_v32i16_ty], [llvm_v16i32_ty, llvm_v16i32_ty,
                          llvm_v32i16_ty, llvm_i32_ty], [IntrNoMem]>;
 }
 
+// Unpack ops.
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx512_mask_unpckh_pd_128 :
+         GCCBuiltin<"__builtin_ia32_unpckhpd128_mask">,
+          Intrinsic<[llvm_v2f64_ty],
+          [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_unpckh_pd_256 :
+         GCCBuiltin<"__builtin_ia32_unpckhpd256_mask">,
+          Intrinsic<[llvm_v4f64_ty],
+          [llvm_v4f64_ty, llvm_v4f64_ty, llvm_v4f64_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_unpckh_pd_512 :
+         GCCBuiltin<"__builtin_ia32_unpckhpd512_mask">,
+          Intrinsic<[llvm_v8f64_ty],
+          [llvm_v8f64_ty, llvm_v8f64_ty, llvm_v8f64_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_unpckh_ps_128 :
+         GCCBuiltin<"__builtin_ia32_unpckhps128_mask">,
+          Intrinsic<[llvm_v4f32_ty],
+          [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_unpckh_ps_256 :
+         GCCBuiltin<"__builtin_ia32_unpckhps256_mask">,
+          Intrinsic<[llvm_v8f32_ty],
+          [llvm_v8f32_ty, llvm_v8f32_ty, llvm_v8f32_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_unpckh_ps_512 :
+         GCCBuiltin<"__builtin_ia32_unpckhps512_mask">,
+          Intrinsic<[llvm_v16f32_ty],
+          [llvm_v16f32_ty, llvm_v16f32_ty, llvm_v16f32_ty,  llvm_i16_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_unpckl_pd_128 :
+         GCCBuiltin<"__builtin_ia32_unpcklpd128_mask">,
+          Intrinsic<[llvm_v2f64_ty],
+          [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_unpckl_pd_256 :
+         GCCBuiltin<"__builtin_ia32_unpcklpd256_mask">,
+          Intrinsic<[llvm_v4f64_ty],
+          [llvm_v4f64_ty, llvm_v4f64_ty, llvm_v4f64_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_unpckl_pd_512 :
+         GCCBuiltin<"__builtin_ia32_unpcklpd512_mask">,
+          Intrinsic<[llvm_v8f64_ty],
+          [llvm_v8f64_ty, llvm_v8f64_ty, llvm_v8f64_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_unpckl_ps_128 :
+         GCCBuiltin<"__builtin_ia32_unpcklps128_mask">,
+          Intrinsic<[llvm_v4f32_ty],
+          [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_unpckl_ps_256 :
+         GCCBuiltin<"__builtin_ia32_unpcklps256_mask">,
+          Intrinsic<[llvm_v8f32_ty],
+          [llvm_v8f32_ty, llvm_v8f32_ty, llvm_v8f32_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_unpckl_ps_512 :
+         GCCBuiltin<"__builtin_ia32_unpcklps512_mask">,
+          Intrinsic<[llvm_v16f32_ty],
+          [llvm_v16f32_ty, llvm_v16f32_ty, llvm_v16f32_ty,  llvm_i16_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpckhb_w_128 :
+         GCCBuiltin<"__builtin_ia32_punpckhbw128_mask">,
+          Intrinsic<[llvm_v16i8_ty],
+          [llvm_v16i8_ty, llvm_v16i8_ty, llvm_v16i8_ty,  llvm_i16_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpckhb_w_256 :
+         GCCBuiltin<"__builtin_ia32_punpckhbw256_mask">,
+          Intrinsic<[llvm_v32i8_ty],
+          [llvm_v32i8_ty, llvm_v32i8_ty, llvm_v32i8_ty,  llvm_i32_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpckhb_w_512 :
+         GCCBuiltin<"__builtin_ia32_punpckhbw512_mask">,
+          Intrinsic<[llvm_v64i8_ty],
+          [llvm_v64i8_ty, llvm_v64i8_ty, llvm_v64i8_ty,  llvm_i64_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpckhd_q_128 :
+         GCCBuiltin<"__builtin_ia32_punpckhdq128_mask">,
+          Intrinsic<[llvm_v4i32_ty],
+          [llvm_v4i32_ty, llvm_v4i32_ty, llvm_v4i32_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpckhd_q_256 :
+         GCCBuiltin<"__builtin_ia32_punpckhdq256_mask">,
+          Intrinsic<[llvm_v8i32_ty],
+          [llvm_v8i32_ty, llvm_v8i32_ty, llvm_v8i32_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpckhd_q_512 :
+         GCCBuiltin<"__builtin_ia32_punpckhdq512_mask">,
+          Intrinsic<[llvm_v16i32_ty],
+          [llvm_v16i32_ty, llvm_v16i32_ty, llvm_v16i32_ty,  llvm_i16_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpckhqd_q_128 :
+         GCCBuiltin<"__builtin_ia32_punpckhqdq128_mask">,
+          Intrinsic<[llvm_v2i64_ty],
+          [llvm_v2i64_ty, llvm_v2i64_ty, llvm_v2i64_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpckhqd_q_256 :
+         GCCBuiltin<"__builtin_ia32_punpckhqdq256_mask">,
+          Intrinsic<[llvm_v4i64_ty],
+          [llvm_v4i64_ty, llvm_v4i64_ty, llvm_v4i64_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpckhqd_q_512 :
+         GCCBuiltin<"__builtin_ia32_punpckhqdq512_mask">,
+          Intrinsic<[llvm_v8i64_ty],
+          [llvm_v8i64_ty, llvm_v8i64_ty, llvm_v8i64_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpckhw_d_128 :
+         GCCBuiltin<"__builtin_ia32_punpckhwd128_mask">,
+          Intrinsic<[llvm_v8i16_ty],
+          [llvm_v8i16_ty, llvm_v8i16_ty, llvm_v8i16_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpckhw_d_256 :
+         GCCBuiltin<"__builtin_ia32_punpckhwd256_mask">,
+          Intrinsic<[llvm_v16i16_ty],
+          [llvm_v16i16_ty, llvm_v16i16_ty, llvm_v16i16_ty,  llvm_i16_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpckhw_d_512 :
+         GCCBuiltin<"__builtin_ia32_punpckhwd512_mask">,
+          Intrinsic<[llvm_v32i16_ty],
+          [llvm_v32i16_ty, llvm_v32i16_ty, llvm_v32i16_ty,  llvm_i32_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpcklb_w_128 :
+         GCCBuiltin<"__builtin_ia32_punpcklbw128_mask">,
+          Intrinsic<[llvm_v16i8_ty],
+          [llvm_v16i8_ty, llvm_v16i8_ty, llvm_v16i8_ty,  llvm_i16_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpcklb_w_256 :
+         GCCBuiltin<"__builtin_ia32_punpcklbw256_mask">,
+          Intrinsic<[llvm_v32i8_ty],
+          [llvm_v32i8_ty, llvm_v32i8_ty, llvm_v32i8_ty,  llvm_i32_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpcklb_w_512 :
+         GCCBuiltin<"__builtin_ia32_punpcklbw512_mask">,
+          Intrinsic<[llvm_v64i8_ty],
+          [llvm_v64i8_ty, llvm_v64i8_ty, llvm_v64i8_ty,  llvm_i64_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpckld_q_128 :
+         GCCBuiltin<"__builtin_ia32_punpckldq128_mask">,
+          Intrinsic<[llvm_v4i32_ty],
+          [llvm_v4i32_ty, llvm_v4i32_ty, llvm_v4i32_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpckld_q_256 :
+         GCCBuiltin<"__builtin_ia32_punpckldq256_mask">,
+          Intrinsic<[llvm_v8i32_ty],
+          [llvm_v8i32_ty, llvm_v8i32_ty, llvm_v8i32_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpckld_q_512 :
+         GCCBuiltin<"__builtin_ia32_punpckldq512_mask">,
+          Intrinsic<[llvm_v16i32_ty],
+          [llvm_v16i32_ty, llvm_v16i32_ty, llvm_v16i32_ty,  llvm_i16_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpcklqd_q_128 :
+         GCCBuiltin<"__builtin_ia32_punpcklqdq128_mask">,
+          Intrinsic<[llvm_v2i64_ty],
+          [llvm_v2i64_ty, llvm_v2i64_ty, llvm_v2i64_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpcklqd_q_256 :
+         GCCBuiltin<"__builtin_ia32_punpcklqdq256_mask">,
+          Intrinsic<[llvm_v4i64_ty],
+          [llvm_v4i64_ty, llvm_v4i64_ty, llvm_v4i64_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpcklqd_q_512 :
+         GCCBuiltin<"__builtin_ia32_punpcklqdq512_mask">,
+          Intrinsic<[llvm_v8i64_ty],
+          [llvm_v8i64_ty, llvm_v8i64_ty, llvm_v8i64_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpcklw_d_128 :
+         GCCBuiltin<"__builtin_ia32_punpcklwd128_mask">,
+          Intrinsic<[llvm_v8i16_ty],
+          [llvm_v8i16_ty, llvm_v8i16_ty, llvm_v8i16_ty,  llvm_i8_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpcklw_d_256 :
+         GCCBuiltin<"__builtin_ia32_punpcklwd256_mask">,
+          Intrinsic<[llvm_v16i16_ty],
+          [llvm_v16i16_ty, llvm_v16i16_ty, llvm_v16i16_ty,  llvm_i16_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_punpcklw_d_512 :
+         GCCBuiltin<"__builtin_ia32_punpcklwd512_mask">,
+          Intrinsic<[llvm_v32i16_ty],
+          [llvm_v32i16_ty, llvm_v32i16_ty, llvm_v32i16_ty,  llvm_i32_ty],
+          [IntrNoMem]>;
+}
+
 // Vector convert
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_avx512_mask_cvtdq2pd_128 : 
         GCCBuiltin<"__builtin_ia32_cvtdq2pd128_mask">,
           Intrinsic<[llvm_v2f64_ty],
           [llvm_v4i32_ty, llvm_v2f64_ty,  llvm_i8_ty],
           [IntrNoMem]>;
 

lib/Target/X86/X86InstrAVX512.td

 
 //===----------------------------------------------------------------------===//
 // AVX-512 - Integer arithmetic
 //
 multiclass avx512_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
                            X86VectorVTInfo _, OpndItins itins,
                            bit IsCommutable = 0> {
   defm rr : AVX512_maskable<opc, MRMSrcReg, _, (outs _.RC:$dst),
-                    (ins _.RC:$src1, _.RC:$src2), OpcodeStr##_.Suffix,
+                    (ins _.RC:$src1, _.RC:$src2), OpcodeStr,
                     "$src2, $src1", "$src1, $src2",
                     (_.VT (OpNode _.RC:$src1, _.RC:$src2)),
                     itins.rr, IsCommutable>,
             AVX512BIBase, EVEX_4V;
 
   let mayLoad = 1 in
     defm rm : AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst),
-                    (ins _.RC:$src1, _.MemOp:$src2), OpcodeStr##_.Suffix,
+                    (ins _.RC:$src1, _.MemOp:$src2), OpcodeStr,
                     "$src2, $src1", "$src1, $src2",
                     (_.VT (OpNode _.RC:$src1,
                                   (bitconvert (_.LdFrag addr:$src2)))),
                     itins.rm>,
               AVX512BIBase, EVEX_4V;
 }
 
 multiclass avx512_binop_rmb<bits<8> opc, string OpcodeStr, SDNode OpNode,
                             X86VectorVTInfo _, OpndItins itins,
                             bit IsCommutable = 0> :
            avx512_binop_rm<opc, OpcodeStr, OpNode, _, itins, IsCommutable> {
   let mayLoad = 1 in
     defm rmb : AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst),
-                    (ins _.RC:$src1, _.ScalarMemOp:$src2), OpcodeStr##_.Suffix,
+                    (ins _.RC:$src1, _.ScalarMemOp:$src2), OpcodeStr,
                     "${src2}"##_.BroadcastStr##", $src1",
                     "$src1, ${src2}"##_.BroadcastStr,
                     (_.VT (OpNode _.RC:$src1,
                                   (X86VBroadcast
                                       (_.ScalarLdFrag addr:$src2)))),
                     itins.rm>,
                AVX512BIBase, EVEX_4V, EVEX_B;
 }
defm Z256 : avx512_binop_rmb<opc, OpcodeStr, OpNode, VTInfo.info256, itins,
                              IsCommutable>, EVEX_V256;
     defm Z128 : avx512_binop_rmb<opc, OpcodeStr, OpNode, VTInfo.info128, itins,
                              IsCommutable>, EVEX_V128;
   }
 }
 
 multiclass avx512_binop_rm_vl_q<bits<8> opc, string OpcodeStr, SDNode OpNode,
                                 OpndItins itins, Predicate prd,
-                                bit IsCommutable = 0> {
-  defm NAME : avx512_binop_rmb_vl<opc, OpcodeStr, OpNode, avx512vl_i64_info,
-                               itins, prd, IsCommutable>,
+                                bit IsCommutable = 0, string suffix = "q"> {
+  defm NAME : avx512_binop_rmb_vl<opc, OpcodeStr##suffix, OpNode,
+                                  avx512vl_i64_info, itins, prd, IsCommutable>,
                                   VEX_W, EVEX_CD8<64, CD8VF>;
 }
 
 multiclass avx512_binop_rm_vl_d<bits<8> opc, string OpcodeStr, SDNode OpNode,
                                 OpndItins itins, Predicate prd,
-                                bit IsCommutable = 0> {
-  defm NAME : avx512_binop_rmb_vl<opc, OpcodeStr, OpNode, avx512vl_i32_info,
-                               itins, prd, IsCommutable>, EVEX_CD8<32, CD8VF>;
+                                bit IsCommutable = 0, string suffix = "d"> {
+  defm NAME : avx512_binop_rmb_vl<opc, OpcodeStr##suffix, OpNode,
+                                  avx512vl_i32_info, itins, prd, IsCommutable>,
+                                  EVEX_CD8<32, CD8VF>;
 }
 
 multiclass avx512_binop_rm_vl_w<bits<8> opc, string OpcodeStr, SDNode OpNode,
                                 OpndItins itins, Predicate prd,
-                                bit IsCommutable = 0> {
-  defm NAME : avx512_binop_rm_vl<opc, OpcodeStr, OpNode, avx512vl_i16_info,
-                              itins, prd, IsCommutable>, EVEX_CD8<16, CD8VF>;
+                                bit IsCommutable = 0, string suffix = "w"> {
+  defm NAME : avx512_binop_rm_vl<opc, OpcodeStr##suffix, OpNode,
+                                 avx512vl_i16_info, itins, prd, IsCommutable>,
+                                 EVEX_CD8<16, CD8VF>;
 }
 
 multiclass avx512_binop_rm_vl_b<bits<8> opc, string OpcodeStr, SDNode OpNode,
                                 OpndItins itins, Predicate prd,
-                                bit IsCommutable = 0> {
-  defm NAME : avx512_binop_rm_vl<opc, OpcodeStr, OpNode, avx512vl_i8_info,
-                              itins, prd, IsCommutable>, EVEX_CD8<8, CD8VF>;
+                                bit IsCommutable = 0, string suffix = "b"> {
+  defm NAME : avx512_binop_rm_vl<opc, OpcodeStr##suffix, OpNode,
+                                 avx512vl_i8_info, itins, prd, IsCommutable>,
+                                 EVEX_CD8<8, CD8VF>;
 }
 
 multiclass avx512_binop_rm_vl_dq<bits<8> opc_d, bits<8> opc_q, string OpcodeStr,
                                  SDNode OpNode, OpndItins itins, Predicate prd,
                                  bit IsCommutable = 0> {
   defm Q : avx512_binop_rm_vl_q<opc_q, OpcodeStr, OpNode, itins, prd,
                                    IsCommutable>;
 
defm VPMINS : avx512_binop_rm_vl_dq<0x39, 0x39, "vpmins", smin,
                                      SSE_INTALU_ITINS_P, HasAVX512, 1>, T8PD;
 
 defm VPMINUB : avx512_binop_rm_vl_b<0xDA, "vpminu", umin,
                                      SSE_INTALU_ITINS_P, HasBWI, 1>;
 defm VPMINUW : avx512_binop_rm_vl_w<0x3A, "vpminu", umin,
                                      SSE_INTALU_ITINS_P, HasBWI, 1>, T8PD;
 defm VPMINU : avx512_binop_rm_vl_dq<0x3B, 0x3B, "vpminu", umin,
                                      SSE_INTALU_ITINS_P, HasAVX512, 1>, T8PD;
-
-//===----------------------------------------------------------------------===//
-// AVX-512 - Unpack Instructions
-//===----------------------------------------------------------------------===//
-
-multiclass avx512_unpack_fp<bits<8> opc, SDNode OpNode, ValueType vt,
-                                   PatFrag mem_frag, RegisterClass RC,
-                                   X86MemOperand x86memop, string asm,
-                                   Domain d> {
-    def rr : AVX512PI<opc, MRMSrcReg,
-                (outs RC:$dst), (ins RC:$src1, RC:$src2),
-                asm, [(set RC:$dst,
-                           (vt (OpNode RC:$src1, RC:$src2)))],
-                           d>, EVEX_4V;
-    def rm : AVX512PI<opc, MRMSrcMem,
-                (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
-                asm, [(set RC:$dst,
-                       (vt (OpNode RC:$src1,
-                            (bitconvert (mem_frag addr:$src2)))))],
-                        d>, EVEX_4V;
-}
-
-defm VUNPCKHPSZ: avx512_unpack_fp<0x15, X86Unpckh, v16f32, loadv8f64,
-      VR512, f512mem, "vunpckhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
-      SSEPackedSingle>, PS, EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VUNPCKHPDZ: avx512_unpack_fp<0x15, X86Unpckh, v8f64, loadv8f64,
-      VR512, f512mem, "vunpckhpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
-      SSEPackedDouble>, PD, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
-defm VUNPCKLPSZ: avx512_unpack_fp<0x14, X86Unpckl, v16f32, loadv8f64,
-      VR512, f512mem, "vunpcklps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
-      SSEPackedSingle>, PS, EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VUNPCKLPDZ: avx512_unpack_fp<0x14, X86Unpckl, v8f64, loadv8f64,
-      VR512, f512mem, "vunpcklpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
-      SSEPackedDouble>, PD, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
-
-multiclass avx512_unpack_int<bits<8> opc, string OpcodeStr, SDNode OpNode,
-                        ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
-                        X86MemOperand x86memop> {
-  def rr : AVX512BI<opc, MRMSrcReg, (outs RC:$dst),
-       (ins RC:$src1, RC:$src2),
-       !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
-       [(set RC:$dst, (OpVT (OpNode (OpVT RC:$src1), (OpVT RC:$src2))))],
-       IIC_SSE_UNPCK>, EVEX_4V;
-  def rm : AVX512BI<opc, MRMSrcMem, (outs RC:$dst),
-       (ins RC:$src1, x86memop:$src2),
-       !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
-       [(set RC:$dst, (OpVT (OpNode (OpVT RC:$src1),
-                                     (bitconvert (memop_frag addr:$src2)))))],
-                                     IIC_SSE_UNPCK>, EVEX_4V;
-}
-defm VPUNPCKLDQZ  : avx512_unpack_int<0x62, "vpunpckldq", X86Unpckl, v16i32,
-                                VR512, loadv16i32, i512mem>, EVEX_V512,
-                                EVEX_CD8<32, CD8VF>;
-defm VPUNPCKLQDQZ : avx512_unpack_int<0x6C, "vpunpcklqdq", X86Unpckl, v8i64,
-                                VR512, loadv8i64, i512mem>, EVEX_V512,
-                                VEX_W, EVEX_CD8<64, CD8VF>;
-defm VPUNPCKHDQZ  : avx512_unpack_int<0x6A, "vpunpckhdq", X86Unpckh, v16i32,
-                                VR512, loadv16i32, i512mem>, EVEX_V512,
-                                EVEX_CD8<32, CD8VF>;
-defm VPUNPCKHQDQZ : avx512_unpack_int<0x6D, "vpunpckhqdq", X86Unpckh, v8i64,
-                                VR512, loadv8i64, i512mem>, EVEX_V512,
-                                VEX_W, EVEX_CD8<64, CD8VF>;
 //===----------------------------------------------------------------------===//
 // AVX-512  Logical Instructions
 //===----------------------------------------------------------------------===//
 
 defm VPAND : avx512_binop_rm_vl_dq<0xDB, 0xDB, "vpand", and,
                                   SSE_INTALU_ITINS_P, HasAVX512, 1>;
 defm VPOR : avx512_binop_rm_vl_dq<0xEB, 0xEB, "vpor", or,
                                   SSE_INTALU_ITINS_P, HasAVX512, 1>;

 def : Pat<(xor
           (bc_v16i32 (v16i1sextv16i32)),
           (bc_v16i32 (add (v16i32 VR512:$src), (v16i1sextv16i32)))),
           (VPABSDZrr VR512:$src)>;
 def : Pat<(xor
           (bc_v8i64 (v8i1sextv8i64)),
           (bc_v8i64 (add (v8i64 VR512:$src), (v8i1sextv8i64)))),
           (VPABSQZrr VR512:$src)>;
+
+//===----------------------------------------------------------------------===//
+// AVX-512 - Unpack Instructions
+//===----------------------------------------------------------------------===//
+defm VUNPCKH : avx512_fp_binop_p<0x15, "vunpckh", X86Unpckh>;
+defm VUNPCKL : avx512_fp_binop_p<0x14, "vunpckl", X86Unpckl>;
+
+defm VPUNPCKLBW : avx512_binop_rm_vl_b<0x60, "vpunpcklbw", X86Unpckl,
+                                       SSE_INTALU_ITINS_P, HasBWI, 0, "">;
+defm VPUNPCKHBW : avx512_binop_rm_vl_b<0x68, "vpunpckhbw", X86Unpckh,
+                                       SSE_INTALU_ITINS_P, HasBWI, 0, "">;
+defm VPUNPCKLWD : avx512_binop_rm_vl_w<0x61, "vpunpcklwd", X86Unpckl,
+                                       SSE_INTALU_ITINS_P, HasBWI, 0, "">;
+defm VPUNPCKHWD : avx512_binop_rm_vl_w<0x69, "vpunpckhwd", X86Unpckh,
+                                       SSE_INTALU_ITINS_P, HasBWI, 0, "">;
+
+defm VPUNPCKLDQ : avx512_binop_rm_vl_d<0x62, "vpunpckldq", X86Unpckl,
+                                       SSE_INTALU_ITINS_P, HasAVX512, 0, "">;
+defm VPUNPCKHDQ : avx512_binop_rm_vl_d<0x6A, "vpunpckhdq", X86Unpckh,
+                                       SSE_INTALU_ITINS_P, HasAVX512, 0, "">;
+defm VPUNPCKLQDQ : avx512_binop_rm_vl_q<0x6C, "vpunpcklqdq", X86Unpckl,
+                                       SSE_INTALU_ITINS_P, HasAVX512, 0, "">;
+defm VPUNPCKHQDQ : avx512_binop_rm_vl_q<0x6D, "vpunpckhqdq", X86Unpckh,
+                                       SSE_INTALU_ITINS_P, HasAVX512, 0, "">;

lib/Target/X86/X86InstrSSE.td

def rm : PI<opc, MRMSrcMem,
                 (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
                 asm, [(set RC:$dst,
                            (vt (OpNode RC:$src1,
                                        (mem_frag addr:$src2))))],
                                        IIC_SSE_UNPCK, d>,
              Sched<[WriteFShuffleLd, ReadAfterLd]>;
 }
 
+let Predicates = [HasAVX, NoVLX] in {
 defm VUNPCKHPS: sse12_unpack_interleave<0x15, X86Unpckh, v4f32, loadv4f32,
       VR128, f128mem, "unpckhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
                      SSEPackedSingle>, PS, VEX_4V;
 defm VUNPCKHPD: sse12_unpack_interleave<0x15, X86Unpckh, v2f64, loadv2f64,
       VR128, f128mem, "unpckhpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
                      SSEPackedDouble>, PD, VEX_4V;
 defm VUNPCKLPS: sse12_unpack_interleave<0x14, X86Unpckl, v4f32, loadv4f32,
       VR128, f128mem, "unpcklps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
defm VUNPCKHPDY: sse12_unpack_interleave<0x15, X86Unpckh, v4f64, loadv4f64,
       VR256, f256mem, "unpckhpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
                      SSEPackedDouble>, PD, VEX_4V, VEX_L;
 defm VUNPCKLPSY: sse12_unpack_interleave<0x14, X86Unpckl, v8f32, loadv8f32,
       VR256, f256mem, "unpcklps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
                      SSEPackedSingle>, PS, VEX_4V, VEX_L;
 defm VUNPCKLPDY: sse12_unpack_interleave<0x14, X86Unpckl, v4f64, loadv4f64,
       VR256, f256mem, "unpcklpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
                      SSEPackedDouble>, PD, VEX_4V, VEX_L;
-
+}// Predicates = [HasAVX, NoVLX]
 let Constraints = "$src1 = $dst" in {
   defm UNPCKHPS: sse12_unpack_interleave<0x15, X86Unpckh, v4f32, memopv4f32,
         VR128, f128mem, "unpckhps\t{$src2, $dst|$dst, $src2}",
                        SSEPackedSingle>, PS;
   defm UNPCKHPD: sse12_unpack_interleave<0x15, X86Unpckh, v2f64, memopv2f64,
         VR128, f128mem, "unpckhpd\t{$src2, $dst|$dst, $src2}",
                        SSEPackedDouble>, PD;
   defm UNPCKLPS: sse12_unpack_interleave<0x14, X86Unpckl, v4f32, memopv4f32,
multiclass sse2_unpack_y<bits<8> opc, string OpcodeStr, ValueType vt,
   def Yrm : PDI<opc, MRMSrcMem,
       (outs VR256:$dst), (ins VR256:$src1, i256mem:$src2),
       !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
       [(set VR256:$dst, (OpNode VR256:$src1,
                                   (bc_frag (loadv4i64 addr:$src2))))]>,
       Sched<[WriteShuffleLd, ReadAfterLd]>;
 }
 
-let Predicates = [HasAVX] in {
+
+let Predicates = [HasAVX, NoVLX_Or_NoBWI] in {
   defm VPUNPCKLBW  : sse2_unpack<0x60, "vpunpcklbw", v16i8, X86Unpckl,
                                  bc_v16i8, loadv2i64, 0>, VEX_4V;
   defm VPUNPCKLWD  : sse2_unpack<0x61, "vpunpcklwd", v8i16, X86Unpckl,
                                  bc_v8i16, loadv2i64, 0>, VEX_4V;
-  defm VPUNPCKLDQ  : sse2_unpack<0x62, "vpunpckldq", v4i32, X86Unpckl,
-                                 bc_v4i32, loadv2i64, 0>, VEX_4V;
-  defm VPUNPCKLQDQ : sse2_unpack<0x6C, "vpunpcklqdq", v2i64, X86Unpckl,
-                                 bc_v2i64, loadv2i64, 0>, VEX_4V;
-
   defm VPUNPCKHBW  : sse2_unpack<0x68, "vpunpckhbw", v16i8, X86Unpckh,
                                  bc_v16i8, loadv2i64, 0>, VEX_4V;
   defm VPUNPCKHWD  : sse2_unpack<0x69, "vpunpckhwd", v8i16, X86Unpckh,
                                  bc_v8i16, loadv2i64, 0>, VEX_4V;
+}
+let Predicates = [HasAVX, NoVLX] in {
+  defm VPUNPCKLDQ  : sse2_unpack<0x62, "vpunpckldq", v4i32, X86Unpckl,
+                                 bc_v4i32, loadv2i64, 0>, VEX_4V;
+  defm VPUNPCKLQDQ : sse2_unpack<0x6C, "vpunpcklqdq", v2i64, X86Unpckl,
+                                 bc_v2i64, loadv2i64, 0>, VEX_4V;
   defm VPUNPCKHDQ  : sse2_unpack<0x6A, "vpunpckhdq", v4i32, X86Unpckh,
                                  bc_v4i32, loadv2i64, 0>, VEX_4V;
   defm VPUNPCKHQDQ : sse2_unpack<0x6D, "vpunpckhqdq", v2i64, X86Unpckh,
                                  bc_v2i64, loadv2i64, 0>, VEX_4V;
 }
 
-let Predicates = [HasAVX2] in {
+let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
   defm VPUNPCKLBW  : sse2_unpack_y<0x60, "vpunpcklbw", v32i8, X86Unpckl,
                                    bc_v32i8>, VEX_4V, VEX_L;
   defm VPUNPCKLWD  : sse2_unpack_y<0x61, "vpunpcklwd", v16i16, X86Unpckl,
                                    bc_v16i16>, VEX_4V, VEX_L;
-  defm VPUNPCKLDQ  : sse2_unpack_y<0x62, "vpunpckldq", v8i32, X86Unpckl,
-                                   bc_v8i32>, VEX_4V, VEX_L;
-  defm VPUNPCKLQDQ : sse2_unpack_y<0x6C, "vpunpcklqdq", v4i64, X86Unpckl,
-                                   bc_v4i64>, VEX_4V, VEX_L;
-
   defm VPUNPCKHBW  : sse2_unpack_y<0x68, "vpunpckhbw", v32i8, X86Unpckh,
                                    bc_v32i8>, VEX_4V, VEX_L;
   defm VPUNPCKHWD  : sse2_unpack_y<0x69, "vpunpckhwd", v16i16, X86Unpckh,
                                    bc_v16i16>, VEX_4V, VEX_L;
+}
+let Predicates = [HasAVX2, NoVLX] in {
+  defm VPUNPCKLDQ  : sse2_unpack_y<0x62, "vpunpckldq", v8i32, X86Unpckl,
+                                   bc_v8i32>, VEX_4V, VEX_L;
+  defm VPUNPCKLQDQ : sse2_unpack_y<0x6C, "vpunpcklqdq", v4i64, X86Unpckl,
+                                   bc_v4i64>, VEX_4V, VEX_L;
   defm VPUNPCKHDQ  : sse2_unpack_y<0x6A, "vpunpckhdq", v8i32, X86Unpckh,
                                    bc_v8i32>, VEX_4V, VEX_L;
   defm VPUNPCKHQDQ : sse2_unpack_y<0x6D, "vpunpckhqdq", v4i64, X86Unpckh,
                                    bc_v4i64>, VEX_4V, VEX_L;
 }
 
 let Constraints = "$src1 = $dst" in {
   defm PUNPCKLBW  : sse2_unpack<0x60, "punpcklbw", v16i8, X86Unpckl,

lib/Target/X86/X86IntrinsicsInfo.h

static const IntrinsicData  IntrinsicsWithoutChain[] = {
   X86_INTRINSIC_DATA(avx512_mask_psubs_w_256, INTR_TYPE_2OP_MASK, X86ISD::SUBS, 0),
   X86_INTRINSIC_DATA(avx512_mask_psubs_w_512, INTR_TYPE_2OP_MASK, X86ISD::SUBS, 0),
   X86_INTRINSIC_DATA(avx512_mask_psubus_b_128, INTR_TYPE_2OP_MASK, X86ISD::SUBUS, 0),
   X86_INTRINSIC_DATA(avx512_mask_psubus_b_256, INTR_TYPE_2OP_MASK, X86ISD::SUBUS, 0),
   X86_INTRINSIC_DATA(avx512_mask_psubus_b_512, INTR_TYPE_2OP_MASK, X86ISD::SUBUS, 0),
   X86_INTRINSIC_DATA(avx512_mask_psubus_w_128, INTR_TYPE_2OP_MASK, X86ISD::SUBUS, 0),
   X86_INTRINSIC_DATA(avx512_mask_psubus_w_256, INTR_TYPE_2OP_MASK, X86ISD::SUBUS, 0),
   X86_INTRINSIC_DATA(avx512_mask_psubus_w_512, INTR_TYPE_2OP_MASK, X86ISD::SUBUS, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpckhb_w_128, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKH, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpckhb_w_256, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKH, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpckhb_w_512, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKH, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpckhd_q_128, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKH, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpckhd_q_256, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKH, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpckhd_q_512, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKH, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpckhqd_q_128, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKH, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpckhqd_q_256, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKH, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpckhqd_q_512, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKH, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpckhw_d_128, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKH, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpckhw_d_256, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKH, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpckhw_d_512, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKH, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpcklb_w_128, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKL, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpcklb_w_256, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKL, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpcklb_w_512, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKL, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpckld_q_128, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKL, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpckld_q_256, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKL, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpckld_q_512, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKL, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpcklqd_q_128, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKL, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpcklqd_q_256, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKL, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpcklqd_q_512, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKL, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpcklw_d_128, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKL, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpcklw_d_256, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKL, 0),
+  X86_INTRINSIC_DATA(avx512_mask_punpcklw_d_512, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKL, 0),
   X86_INTRINSIC_DATA(avx512_mask_pxor_d_128, INTR_TYPE_2OP_MASK, ISD::XOR, 0),
   X86_INTRINSIC_DATA(avx512_mask_pxor_d_256, INTR_TYPE_2OP_MASK, ISD::XOR, 0),
   X86_INTRINSIC_DATA(avx512_mask_pxor_d_512, INTR_TYPE_2OP_MASK, ISD::XOR, 0),
   X86_INTRINSIC_DATA(avx512_mask_pxor_q_128, INTR_TYPE_2OP_MASK, ISD::XOR, 0),
   X86_INTRINSIC_DATA(avx512_mask_pxor_q_256, INTR_TYPE_2OP_MASK, ISD::XOR, 0),
   X86_INTRINSIC_DATA(avx512_mask_pxor_q_512, INTR_TYPE_2OP_MASK, ISD::XOR, 0),
   X86_INTRINSIC_DATA(avx512_mask_range_pd_128, INTR_TYPE_3OP_MASK_RM, X86ISD::VRANGE, 0),
   X86_INTRINSIC_DATA(avx512_mask_range_pd_256, INTR_TYPE_3OP_MASK_RM, X86ISD::VRANGE, 0),
static const IntrinsicData  IntrinsicsWithoutChain[] = {
   X86_INTRINSIC_DATA(avx512_mask_ucmp_d_256,    CMP_MASK_CC,  X86ISD::CMPMU, 0),
   X86_INTRINSIC_DATA(avx512_mask_ucmp_d_512,    CMP_MASK_CC,  X86ISD::CMPMU, 0),
   X86_INTRINSIC_DATA(avx512_mask_ucmp_q_128,    CMP_MASK_CC,  X86ISD::CMPMU, 0),
   X86_INTRINSIC_DATA(avx512_mask_ucmp_q_256,    CMP_MASK_CC,  X86ISD::CMPMU, 0),
   X86_INTRINSIC_DATA(avx512_mask_ucmp_q_512,    CMP_MASK_CC,  X86ISD::CMPMU, 0),
   X86_INTRINSIC_DATA(avx512_mask_ucmp_w_128,    CMP_MASK_CC,  X86ISD::CMPMU, 0),
   X86_INTRINSIC_DATA(avx512_mask_ucmp_w_256,    CMP_MASK_CC,  X86ISD::CMPMU, 0),
   X86_INTRINSIC_DATA(avx512_mask_ucmp_w_512,    CMP_MASK_CC,  X86ISD::CMPMU, 0),
+  X86_INTRINSIC_DATA(avx512_mask_unpckh_pd_128, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKH, 0),
+  X86_INTRINSIC_DATA(avx512_mask_unpckh_pd_256, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKH, 0),
+  X86_INTRINSIC_DATA(avx512_mask_unpckh_pd_512, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKH, 0),
+  X86_INTRINSIC_DATA(avx512_mask_unpckh_ps_128, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKH, 0),
+  X86_INTRINSIC_DATA(avx512_mask_unpckh_ps_256, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKH, 0),
+  X86_INTRINSIC_DATA(avx512_mask_unpckh_ps_512, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKH, 0),
+  X86_INTRINSIC_DATA(avx512_mask_unpckl_pd_128, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKL, 0),
+  X86_INTRINSIC_DATA(avx512_mask_unpckl_pd_256, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKL, 0),
+  X86_INTRINSIC_DATA(avx512_mask_unpckl_pd_512, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKL, 0),
+  X86_INTRINSIC_DATA(avx512_mask_unpckl_ps_128, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKL, 0),
+  X86_INTRINSIC_DATA(avx512_mask_unpckl_ps_256, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKL, 0),
+  X86_INTRINSIC_DATA(avx512_mask_unpckl_ps_512, INTR_TYPE_2OP_MASK,
+                     X86ISD::UNPCKL, 0),
   X86_INTRINSIC_DATA(avx512_mask_valign_d_512, INTR_TYPE_3OP_MASK, X86ISD::VALIGN, 0),
   X86_INTRINSIC_DATA(avx512_mask_valign_q_512, INTR_TYPE_3OP_MASK, X86ISD::VALIGN, 0),
 
   X86_INTRINSIC_DATA(avx512_mask_vfmadd_pd_128, FMA_OP_MASK, X86ISD::FMADD, 0),
   X86_INTRINSIC_DATA(avx512_mask_vfmadd_pd_256, FMA_OP_MASK, X86ISD::FMADD, 0),
   X86_INTRINSIC_DATA(avx512_mask_vfmadd_pd_512, FMA_OP_MASK, X86ISD::FMADD,
                      X86ISD::FMADD_RND),
   X86_INTRINSIC_DATA(avx512_mask_vfmadd_ps_128, FMA_OP_MASK, X86ISD::FMADD, 0),

test/CodeGen/X86/avx512-intrinsics.ll

 ; CHECK: vscalefps{{.*}}{%k1} 
 define <16 x float>@test_int_x86_avx512_mask_scalef_ps_512(<16 x float> %x0, <16 x float> %x1, <16 x float> %x2, i16 %x3) {
   %res = call <16 x float> @llvm.x86.avx512.mask.scalef.ps.512(<16 x float> %x0, <16 x float> %x1, <16 x float> %x2, i16 %x3, i32 2)
   %res1 = call <16 x float> @llvm.x86.avx512.mask.scalef.ps.512(<16 x float> %x0, <16 x float> %x1, <16 x float> %x2, i16 -1, i32 0)
   %res2 = fadd <16 x float> %res, %res1
   ret <16 x float> %res2
 }
 
+declare <8 x double> @llvm.x86.avx512.mask.unpckh.pd.512(<8 x double>, <8 x double>, <8 x double>, i8)
+
+define <8 x double>@test_int_x86_avx512_mask_unpckh_pd_512(<8 x double> %x0, <8 x double> %x1, <8 x double> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_unpckh_pd_512:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    movzbl %dil, %eax
+; CHECK-NEXT:    kmovw %eax, %k1
+; CHECK-NEXT:    vunpckhpd %zmm1, %zmm0, %zmm2 {%k1}
+; CHECK-NEXT:    vunpckhpd %zmm1, %zmm0, %zmm0
+  %res = call <8 x double> @llvm.x86.avx512.mask.unpckh.pd.512(<8 x double> %x0, <8 x double> %x1, <8 x double> %x2, i8 %x3)
+  %res1 = call <8 x double> @llvm.x86.avx512.mask.unpckh.pd.512(<8 x double> %x0, <8 x double> %x1, <8 x double> %x2, i8 -1)
+  %res2 = fadd <8 x double> %res, %res1
+  ret <8 x double> %res2
+}
+
+declare <16 x float> @llvm.x86.avx512.mask.unpckh.ps.512(<16 x float>, <16 x float>, <16 x float>, i16)
+
+define <16 x float>@test_int_x86_avx512_mask_unpckh_ps_512(<16 x float> %x0, <16 x float> %x1, <16 x float> %x2, i16 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_unpckh_ps_512:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovw %edi, %k1
+; CHECK-NEXT:    vunpckhps %zmm1, %zmm0, %zmm2 {%k1}
+; CHECK-NEXT:    vunpckhps %zmm1, %zmm0, %zmm0
+  %res = call <16 x float> @llvm.x86.avx512.mask.unpckh.ps.512(<16 x float> %x0, <16 x float> %x1, <16 x float> %x2, i16 %x3)
+  %res1 = call <16 x float> @llvm.x86.avx512.mask.unpckh.ps.512(<16 x float> %x0, <16 x float> %x1, <16 x float> %x2, i16 -1)
+  %res2 = fadd <16 x float> %res, %res1
+  ret <16 x float> %res2
+}
+
+declare <8 x double> @llvm.x86.avx512.mask.unpckl.pd.512(<8 x double>, <8 x double>, <8 x double>, i8)
+
+define <8 x double>@test_int_x86_avx512_mask_unpckl_pd_512(<8 x double> %x0, <8 x double> %x1, <8 x double> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_unpckl_pd_512:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    movzbl %dil, %eax
+; CHECK-NEXT:    kmovw %eax, %k1
+; CHECK-NEXT:    vunpcklpd %zmm1, %zmm0, %zmm2 {%k1}
+; CHECK-NEXT:    vunpcklpd %zmm1, %zmm0, %zmm0
+  %res = call <8 x double> @llvm.x86.avx512.mask.unpckl.pd.512(<8 x double> %x0, <8 x double> %x1, <8 x double> %x2, i8 %x3)
+  %res1 = call <8 x double> @llvm.x86.avx512.mask.unpckl.pd.512(<8 x double> %x0, <8 x double> %x1, <8 x double> %x2, i8 -1)
+  %res2 = fadd <8 x double> %res, %res1
+  ret <8 x double> %res2
+}
+
+declare <16 x float> @llvm.x86.avx512.mask.unpckl.ps.512(<16 x float>, <16 x float>, <16 x float>, i16)
+
+define <16 x float>@test_int_x86_avx512_mask_unpckl_ps_512(<16 x float> %x0, <16 x float> %x1, <16 x float> %x2, i16 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_unpckl_ps_512:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovw %edi, %k1
+; CHECK-NEXT:    vunpcklps %zmm1, %zmm0, %zmm2 {%k1}
+; CHECK-NEXT:    vunpcklps %zmm1, %zmm0, %zmm0
+  %res = call <16 x float> @llvm.x86.avx512.mask.unpckl.ps.512(<16 x float> %x0, <16 x float> %x1, <16 x float> %x2, i16 %x3)
+  %res1 = call <16 x float> @llvm.x86.avx512.mask.unpckl.ps.512(<16 x float> %x0, <16 x float> %x1, <16 x float> %x2, i16 -1)
+  %res2 = fadd <16 x float> %res, %res1
+  ret <16 x float> %res2
+}
+
+declare <8 x i64> @llvm.x86.avx512.mask.punpcklqd.q.512(<8 x i64>, <8 x i64>, <8 x i64>, i8)
+
+define <8 x i64>@test_int_x86_avx512_mask_punpcklqd_q_512(<8 x i64> %x0, <8 x i64> %x1, <8 x i64> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpcklqd_q_512:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    movzbl %dil, %eax
+; CHECK-NEXT:    kmovw %eax, %k1
+; CHECK-NEXT:    vpunpcklqdq %zmm1, %zmm0, %zmm2 {%k1}
+; CHECK-NEXT:    vpunpcklqdq %zmm1, %zmm0, %zmm3 {%k1} {z}
+; CHECK-NEXT:    vpunpcklqdq {{.*#+}} 
+; CHECK:         vpaddq %zmm0, %zmm2, %zmm0
+; CHECK-NEXT:    vpaddq %zmm0, %zmm3, %zmm0
+; CHECK-NEXT:    retq
+  %res = call <8 x i64> @llvm.x86.avx512.mask.punpcklqd.q.512(<8 x i64> %x0, <8 x i64> %x1, <8 x i64> %x2, i8 %x3)
+  %res1 = call <8 x i64> @llvm.x86.avx512.mask.punpcklqd.q.512(<8 x i64> %x0, <8 x i64> %x1, <8 x i64> %x2, i8 -1)
+  %res2 = call <8 x i64> @llvm.x86.avx512.mask.punpcklqd.q.512(<8 x i64> %x0, <8 x i64> %x1, <8 x i64> zeroinitializer,i8 %x3)
+  %res3 = add <8 x i64> %res, %res1
+  %res4 = add <8 x i64> %res2, %res3
+  ret <8 x i64> %res4
+}
+
+declare <8 x i64> @llvm.x86.avx512.mask.punpckhqd.q.512(<8 x i64>, <8 x i64>, <8 x i64>, i8)
+
+define <8 x i64>@test_int_x86_avx512_mask_punpckhqd_q_512(<8 x i64> %x0, <8 x i64> %x1, <8 x i64> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpckhqd_q_512:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    movzbl %dil, %eax
+; CHECK-NEXT:    kmovw %eax, %k1
+; CHECK-NEXT:    vpunpckhqdq %zmm1, %zmm0, %zmm2 {%k1}
+; CHECK-NEXT:    vpunpckhqdq {{.*#+}}
+; CHECK:         vpaddq %zmm0, %zmm2, %zmm0
+; CHECK-NEXT:    retq
+  %res = call <8 x i64> @llvm.x86.avx512.mask.punpckhqd.q.512(<8 x i64> %x0, <8 x i64> %x1, <8 x i64> %x2, i8 %x3)
+  %res1 = call <8 x i64> @llvm.x86.avx512.mask.punpckhqd.q.512(<8 x i64> %x0, <8 x i64> %x1, <8 x i64> %x2, i8 -1)
+  %res2 = add <8 x i64> %res, %res1
+  ret <8 x i64> %res2
+}
+
+declare <16 x i32> @llvm.x86.avx512.mask.punpckhd.q.512(<16 x i32>, <16 x i32>, <16 x i32>, i16)
+
+define <16 x i32>@test_int_x86_avx512_mask_punpckhd_q_512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2, i16 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpckhd_q_512:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovw %edi, %k1
+; CHECK-NEXT:    vpunpckhdq %zmm1, %zmm0, %zmm2 {%k1}
+; CHECK-NEXT:    vpunpckhdq {{.*#+}}
+; CHECK:         vpaddd %zmm0, %zmm2, %zmm0
+; CHECK-NEXT:    retq
+  %res = call <16 x i32> @llvm.x86.avx512.mask.punpckhd.q.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2, i16 %x3)
+  %res1 = call <16 x i32> @llvm.x86.avx512.mask.punpckhd.q.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2, i16 -1)
+  %res2 = add <16 x i32> %res, %res1
+  ret <16 x i32> %res2
+}
+
+declare <16 x i32> @llvm.x86.avx512.mask.punpckld.q.512(<16 x i32>, <16 x i32>, <16 x i32>, i16)
+
+define <16 x i32>@test_int_x86_avx512_mask_punpckld_q_512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2, i16 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpckld_q_512:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovw %edi, %k1
+; CHECK-NEXT:    vpunpckldq %zmm1, %zmm0, %zmm2 {%k1}
+; CHECK-NEXT:    vpunpckldq {{.*#+}} 
+; CHECK:         vpaddd %zmm0, %zmm2, %zmm0
+; CHECK-NEXT:    retq
+  %res = call <16 x i32> @llvm.x86.avx512.mask.punpckld.q.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2, i16 %x3)
+  %res1 = call <16 x i32> @llvm.x86.avx512.mask.punpckld.q.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2, i16 -1)
+  %res2 = add <16 x i32> %res, %res1
+  ret <16 x i32> %res2
+}
+
 declare <16 x i8> @llvm.x86.avx512.mask.pmov.qb.512(<8 x i64>, <16 x i8>, i8)
 
 define <16 x i8>@test_int_x86_avx512_mask_pmov_qb_512(<8 x i64> %x0, <16 x i8> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmov_qb_512:
 ; CHECK:       vpmovqb %zmm0, %xmm1 {%k1}
 ; CHECK-NEXT:  vpmovqb %zmm0, %xmm2 {%k1} {z}
 ; CHECK-NEXT:  vpmovqb %zmm0, %xmm0
     %res0 = call <16 x i8> @llvm.x86.avx512.mask.pmov.qb.512(<8 x i64> %x0, <16 x i8> %x1, i8 -1)

test/CodeGen/X86/avx512bw-intrinsics.ll

 ; CHECK-NEXT:    vpaddd %zmm0, %zmm2, %zmm0 
 ; CHECK-NEXT:    retq
   %res = call <16 x i32> @llvm.x86.avx512.mask.pmaddw.d.512(<32 x i16> %x0, <32 x i16> %x1, <16 x i32> %x2, i16 %x3)
   %res1 = call <16 x i32> @llvm.x86.avx512.mask.pmaddw.d.512(<32 x i16> %x0, <32 x i16> %x1, <16 x i32> %x2, i16 -1)
   %res2 = add <16 x i32> %res, %res1
   ret <16 x i32> %res2
 }
 
+declare <64 x i8> @llvm.x86.avx512.mask.punpckhb.w.512(<64 x i8>, <64 x i8>, <64 x i8>, i64)
+
+define <64 x i8>@test_int_x86_avx512_mask_punpckhb_w_512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpckhb_w_512:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovq %rdi, %k1
+; CHECK-NEXT:    vpunpckhbw %zmm1, %zmm0, %zmm2 {%k1}
+; CHECK-NEXT:    vpunpckhbw %zmm1, %zmm0, %zmm0
+; CHECK-NEXT:    vpaddb %zmm0, %zmm2, %zmm0
+; CHECK-NEXT:    retq
+  %res = call <64 x i8> @llvm.x86.avx512.mask.punpckhb.w.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3)
+  %res1 = call <64 x i8> @llvm.x86.avx512.mask.punpckhb.w.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 -1)
+  %res2 = add <64 x i8> %res, %res1
+  ret <64 x i8> %res2
+}
+
+declare <64 x i8> @llvm.x86.avx512.mask.punpcklb.w.512(<64 x i8>, <64 x i8>, <64 x i8>, i64)
+
+define <64 x i8>@test_int_x86_avx512_mask_punpcklb_w_512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpcklb_w_512:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovq %rdi, %k1
+; CHECK-NEXT:    vpunpcklbw %zmm1, %zmm0, %zmm2 {%k1}
+; CHECK-NEXT:    vpunpcklbw %zmm1, %zmm0, %zmm0
+; CHECK-NEXT:    vpaddb %zmm0, %zmm2, %zmm0
+; CHECK-NEXT:    retq
+  %res = call <64 x i8> @llvm.x86.avx512.mask.punpcklb.w.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3)
+  %res1 = call <64 x i8> @llvm.x86.avx512.mask.punpcklb.w.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 -1)
+  %res2 = add <64 x i8> %res, %res1
+  ret <64 x i8> %res2
+}
+
+declare <32 x i16> @llvm.x86.avx512.mask.punpckhw.d.512(<32 x i16>, <32 x i16>, <32 x i16>, i32)
+
+define <32 x i16>@test_int_x86_avx512_mask_punpckhw_d_512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpckhw_d_512:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovd %edi, %k1
+; CHECK-NEXT:    vpunpckhwd %zmm1, %zmm0, %zmm2 {%k1}
+; CHECK-NEXT:    vpunpckhwd %zmm1, %zmm0, %zmm0
+; CHECK-NEXT:    vpaddw %zmm0, %zmm2, %zmm0
+; CHECK-NEXT:    retq
+  %res = call <32 x i16> @llvm.x86.avx512.mask.punpckhw.d.512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3)
+  %res1 = call <32 x i16> @llvm.x86.avx512.mask.punpckhw.d.512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 -1)
+  %res2 = add <32 x i16> %res, %res1
+  ret <32 x i16> %res2
+}
+
+declare <32 x i16> @llvm.x86.avx512.mask.punpcklw.d.512(<32 x i16>, <32 x i16>, <32 x i16>, i32)
+
+define <32 x i16>@test_int_x86_avx512_mask_punpcklw_d_512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpcklw_d_512:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovd %edi, %k1
+; CHECK-NEXT:    vpunpcklwd %zmm1, %zmm0, %zmm2 {%k1}
+; CHECK-NEXT:    vpunpcklwd %zmm1, %zmm0, %zmm0
+; CHECK-NEXT:    vpaddw %zmm0, %zmm2, %zmm0
+; CHECK-NEXT:    retq
+  %res = call <32 x i16> @llvm.x86.avx512.mask.punpcklw.d.512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3)
+  %res1 = call <32 x i16> @llvm.x86.avx512.mask.punpcklw.d.512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 -1)
+  %res2 = add <32 x i16> %res, %res1
+  ret <32 x i16> %res2
+}
 No newline at end of file

test/CodeGen/X86/avx512bwvl-intrinsics.ll

 ; CHECK-NEXT:    vpmaddubsw %ymm1, %ymm0, %ymm0
 ; CHECK-NEXT:    vpaddw %ymm0, %ymm2, %ymm0
 ; CHECK-NEXT:    retq
   %res = call <16 x i16> @llvm.x86.avx512.mask.pmaddubs.w.256(<32 x i8> %x0, <32 x i8> %x1, <16 x i16> %x2, i16 %x3)
   %res1 = call <16 x i16> @llvm.x86.avx512.mask.pmaddubs.w.256(<32 x i8> %x0, <32 x i8> %x1, <16 x i16> %x2, i16 -1)
   %res2 = add <16 x i16> %res, %res1
   ret <16 x i16> %res2
 }
+
+declare <16 x i8> @llvm.x86.avx512.mask.punpckhb.w.128(<16 x i8>, <16 x i8>, <16 x i8>, i16)
+
+define <16 x i8>@test_int_x86_avx512_mask_punpckhb_w_128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpckhb_w_128:
+; CHECK:         vpunpckhbw %xmm1, %xmm0, %xmm2 {%k1}
+; CHECK-NEXT:    vpunpckhbw %xmm1, %xmm0, %xmm0 ## encoding: [0x62,0xf1,0x7d,0x08,0x68,0xc1]
+  %res = call <16 x i8> @llvm.x86.avx512.mask.punpckhb.w.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %x3)
+  %res1 = call <16 x i8> @llvm.x86.avx512.mask.punpckhb.w.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 -1)
+  %res2 = add <16 x i8> %res, %res1
+  ret <16 x i8> %res2
+}
+
+declare <16 x i8> @llvm.x86.avx512.mask.punpcklb.w.128(<16 x i8>, <16 x i8>, <16 x i8>, i16)
+
+define <16 x i8>@test_int_x86_avx512_mask_punpcklb_w_128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpcklb_w_128:
+; CHECK:         vpunpcklbw %xmm1, %xmm0, %xmm2 {%k1}
+; CHECK-NEXT:    vpunpcklbw %xmm1, %xmm0, %xmm0 ## encoding: [0x62,0xf1,0x7d,0x08,0x60,0xc1]
+  %res = call <16 x i8> @llvm.x86.avx512.mask.punpcklb.w.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %x3)
+  %res1 = call <16 x i8> @llvm.x86.avx512.mask.punpcklb.w.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 -1)
+  %res2 = add <16 x i8> %res, %res1
+  ret <16 x i8> %res2
+}
+
+declare <32 x i8> @llvm.x86.avx512.mask.punpckhb.w.256(<32 x i8>, <32 x i8>, <32 x i8>, i32)
+
+define <32 x i8>@test_int_x86_avx512_mask_punpckhb_w_256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpckhb_w_256:
+; CHECK:         vpunpckhbw %ymm1, %ymm0, %ymm2 {%k1}
+; CHECK-NEXT:    vpunpckhbw %ymm1, %ymm0, %ymm0 ## encoding: [0x62,0xf1,0x7d,0x28,0x68,0xc1]
+  %res = call <32 x i8> @llvm.x86.avx512.mask.punpckhb.w.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3)
+  %res1 = call <32 x i8> @llvm.x86.avx512.mask.punpckhb.w.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 -1)
+  %res2 = add <32 x i8> %res, %res1
+  ret <32 x i8> %res2
+}
+
+declare <32 x i8> @llvm.x86.avx512.mask.punpcklb.w.256(<32 x i8>, <32 x i8>, <32 x i8>, i32)
+
+define <32 x i8>@test_int_x86_avx512_mask_punpcklb_w_256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpcklb_w_256:
+; CHECK:         vpunpcklbw %ymm1, %ymm0, %ymm2 {%k1}
+; CHECK-NEXT:    vpunpcklbw %ymm1, %ymm0, %ymm0 ## encoding: [0x62,0xf1,0x7d,0x28,0x60,0xc1]
+  %res = call <32 x i8> @llvm.x86.avx512.mask.punpcklb.w.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3)
+  %res1 = call <32 x i8> @llvm.x86.avx512.mask.punpcklb.w.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 -1)
+  %res2 = add <32 x i8> %res, %res1
+  ret <32 x i8> %res2
+}
+
+declare <8 x i16> @llvm.x86.avx512.mask.punpcklw.d.128(<8 x i16>, <8 x i16>, <8 x i16>, i8)
+
+define <8 x i16>@test_int_x86_avx512_mask_punpcklw_d_128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpcklw_d_128:
+; CHECK:         vpunpcklwd %xmm1, %xmm0, %xmm2 {%k1}
+; CHECK-NEXT:    vpunpcklwd %xmm1, %xmm0, %xmm0 ## encoding: [0x62,0xf1,0x7d,0x08,0x61,0xc1]
+  %res = call <8 x i16> @llvm.x86.avx512.mask.punpcklw.d.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3)
+  %res1 = call <8 x i16> @llvm.x86.avx512.mask.punpcklw.d.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 -1)
+  %res2 = add <8 x i16> %res, %res1
+  ret <8 x i16> %res2
+}
+
+declare <8 x i16> @llvm.x86.avx512.mask.punpckhw.d.128(<8 x i16>, <8 x i16>, <8 x i16>, i8)
+
+define <8 x i16>@test_int_x86_avx512_mask_punpckhw_d_128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpckhw_d_128:
+; CHECK:         vpunpckhwd %xmm1, %xmm0, %xmm2 {%k1}
+; CHECK-NEXT:    vpunpckhwd %xmm1, %xmm0, %xmm0 ## encoding: [0x62,0xf1,0x7d,0x08,0x69,0xc1]
+  %res = call <8 x i16> @llvm.x86.avx512.mask.punpckhw.d.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3)
+  %res1 = call <8 x i16> @llvm.x86.avx512.mask.punpckhw.d.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 -1)
+  %res2 = add <8 x i16> %res, %res1
+  ret <8 x i16> %res2
+}
+
+declare <16 x i16> @llvm.x86.avx512.mask.punpcklw.d.256(<16 x i16>, <16 x i16>, <16 x i16>, i16)
+
+define <16 x i16>@test_int_x86_avx512_mask_punpcklw_d_256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpcklw_d_256:
+; CHECK:         vpunpcklwd %ymm1, %ymm0, %ymm2 {%k1}
+; CHECK-NEXT:    vpunpcklwd %ymm1, %ymm0, %ymm0 ## encoding: [0x62,0xf1,0x7d,0x28,0x61,0xc1]
+  %res = call <16 x i16> @llvm.x86.avx512.mask.punpcklw.d.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %x3)
+  %res1 = call <16 x i16> @llvm.x86.avx512.mask.punpcklw.d.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 -1)
+  %res2 = add <16 x i16> %res, %res1
+  ret <16 x i16> %res2
+}
+
+declare <16 x i16> @llvm.x86.avx512.mask.punpckhw.d.256(<16 x i16>, <16 x i16>, <16 x i16>, i16)
+
+define <16 x i16>@test_int_x86_avx512_mask_punpckhw_d_256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpckhw_d_256:
+; CHECK:         vpunpckhwd %ymm1, %ymm0, %ymm2 {%k1}
+; CHECK-NEXT:    vpunpckhwd %ymm1, %ymm0, %ymm0 ## encoding: [0x62,0xf1,0x7d,0x28,0x69,0xc1]
+  %res = call <16 x i16> @llvm.x86.avx512.mask.punpckhw.d.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %x3)
+  %res1 = call <16 x i16> @llvm.x86.avx512.mask.punpckhw.d.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 -1)
+  %res2 = add <16 x i16> %res, %res1
+  ret <16 x i16> %res2
+}

test/CodeGen/X86/avx512vl-intrinsics.ll

 ; CHECK: vscalefps{{.*}}{%k1} 
 define <8 x float>@test_int_x86_avx512_mask_scalef_ps_256(<8 x float> %x0, <8 x float> %x1, <8 x float> %x2, i8 %x3) {
   %res = call <8 x float> @llvm.x86.avx512.mask.scalef.ps.256(<8 x float> %x0, <8 x float> %x1, <8 x float> %x2, i8 %x3)
   %res1 = call <8 x float> @llvm.x86.avx512.mask.scalef.ps.256(<8 x float> %x0, <8 x float> %x1, <8 x float> %x2, i8 -1)
   %res2 = fadd <8 x float> %res, %res1
   ret <8 x float> %res2
 }
 
+declare <2 x double> @llvm.x86.avx512.mask.unpckh.pd.128(<2 x double>, <2 x double>, <2 x double>, i8)
+
+define <2 x double>@test_int_x86_avx512_mask_unpckh_pd_128(<2 x double> %x0, <2 x double> %x1, <2 x double> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_unpckh_pd_128:
+; CHECK:         vunpckhpd %xmm1, %xmm0, %xmm2 {%k1} 
+; CHECK-NEXT:    vunpckhpd %xmm1, %xmm0, %xmm0 ## encoding: [0x62,0xf1,0xfd,0x08,0x15,0xc1]
+  %res = call <2 x double> @llvm.x86.avx512.mask.unpckh.pd.128(<2 x double> %x0, <2 x double> %x1, <2 x double> %x2, i8 %x3)
+  %res1 = call <2 x double> @llvm.x86.avx512.mask.unpckh.pd.128(<2 x double> %x0, <2 x double> %x1, <2 x double> %x2, i8 -1)
+  %res2 = fadd <2 x double> %res, %res1
+  ret <2 x double> %res2
+}
+
+declare <4 x double> @llvm.x86.avx512.mask.unpckh.pd.256(<4 x double>, <4 x double>, <4 x double>, i8)
+
+define <4 x double>@test_int_x86_avx512_mask_unpckh_pd_256(<4 x double> %x0, <4 x double> %x1, <4 x double> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_unpckh_pd_256:
+; CHECK:         vunpckhpd %ymm1, %ymm0, %ymm2 {%k1}
+; CHECK-NEXT:    vunpckhpd %ymm1, %ymm0, %ymm0 ## encoding: [0x62,0xf1,0xfd,0x28,0x15,0xc1]
+  %res = call <4 x double> @llvm.x86.avx512.mask.unpckh.pd.256(<4 x double> %x0, <4 x double> %x1, <4 x double> %x2, i8 %x3)
+  %res1 = call <4 x double> @llvm.x86.avx512.mask.unpckh.pd.256(<4 x double> %x0, <4 x double> %x1, <4 x double> %x2, i8 -1)
+  %res2 = fadd <4 x double> %res, %res1
+  ret <4 x double> %res2
+}
+
+declare <4 x float> @llvm.x86.avx512.mask.unpckh.ps.128(<4 x float>, <4 x float>, <4 x float>, i8)
+
+define <4 x float>@test_int_x86_avx512_mask_unpckh_ps_128(<4 x float> %x0, <4 x float> %x1, <4 x float> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_unpckh_ps_128:
+; CHECK:         vunpckhps %xmm1, %xmm0, %xmm2 {%k1}
+; CHECK-NEXT:    vunpckhps %xmm1, %xmm0, %xmm0 ## encoding: [0x62,0xf1,0x7c,0x08,0x15,0xc1]
+  %res = call <4 x float> @llvm.x86.avx512.mask.unpckh.ps.128(<4 x float> %x0, <4 x float> %x1, <4 x float> %x2, i8 %x3)
+  %res1 = call <4 x float> @llvm.x86.avx512.mask.unpckh.ps.128(<4 x float> %x0, <4 x float> %x1, <4 x float> %x2, i8 -1)
+  %res2 = fadd <4 x float> %res, %res1
+  ret <4 x float> %res2
+}
+
+declare <8 x float> @llvm.x86.avx512.mask.unpckh.ps.256(<8 x float>, <8 x float>, <8 x float>, i8)
+
+define <8 x float>@test_int_x86_avx512_mask_unpckh_ps_256(<8 x float> %x0, <8 x float> %x1, <8 x float> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_unpckh_ps_256:
+; CHECK:       ## BB#0:
+; CHECK:         vunpckhps %ymm1, %ymm0, %ymm2 {%k1} 
+; CHECK-NEXT:    vunpckhps %ymm1, %ymm0, %ymm0 ## encoding: [0x62,0xf1,0x7c,0x28,0x15,0xc1]
+  %res = call <8 x float> @llvm.x86.avx512.mask.unpckh.ps.256(<8 x float> %x0, <8 x float> %x1, <8 x float> %x2, i8 %x3)
+  %res1 = call <8 x float> @llvm.x86.avx512.mask.unpckh.ps.256(<8 x float> %x0, <8 x float> %x1, <8 x float> %x2, i8 -1)
+  %res2 = fadd <8 x float> %res, %res1
+  ret <8 x float> %res2
+}
+
+declare <2 x double> @llvm.x86.avx512.mask.unpckl.pd.128(<2 x double>, <2 x double>, <2 x double>, i8)
+
+define <2 x double>@test_int_x86_avx512_mask_unpckl_pd_128(<2 x double> %x0, <2 x double> %x1, <2 x double> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_unpckl_pd_128:
+; CHECK:         vunpcklpd %xmm1, %xmm0, %xmm2 {%k1}
+; CHECK-NEXT:    vunpcklpd %xmm1, %xmm0, %xmm0 ## encoding: [0x62,0xf1,0xfd,0x08,0x14,0xc1]
+  %res = call <2 x double> @llvm.x86.avx512.mask.unpckl.pd.128(<2 x double> %x0, <2 x double> %x1, <2 x double> %x2, i8 %x3)
+  %res1 = call <2 x double> @llvm.x86.avx512.mask.unpckl.pd.128(<2 x double> %x0, <2 x double> %x1, <2 x double> %x2, i8 -1)
+  %res2 = fadd <2 x double> %res, %res1
+  ret <2 x double> %res2
+}
+
+declare <4 x double> @llvm.x86.avx512.mask.unpckl.pd.256(<4 x double>, <4 x double>, <4 x double>, i8)
+
+define <4 x double>@test_int_x86_avx512_mask_unpckl_pd_256(<4 x double> %x0, <4 x double> %x1, <4 x double> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_unpckl_pd_256:
+; CHECK:         vunpcklpd %ymm1, %ymm0, %ymm2 {%k1}
+; CHECK-NEXT:    vunpcklpd %ymm1, %ymm0, %ymm0 ## encoding: [0x62,0xf1,0xfd,0x28,0x14,0xc1]
+  %res = call <4 x double> @llvm.x86.avx512.mask.unpckl.pd.256(<4 x double> %x0, <4 x double> %x1, <4 x double> %x2, i8 %x3)
+  %res1 = call <4 x double> @llvm.x86.avx512.mask.unpckl.pd.256(<4 x double> %x0, <4 x double> %x1, <4 x double> %x2, i8 -1)
+  %res2 = fadd <4 x double> %res, %res1
+  ret <4 x double> %res2
+}
+
+declare <4 x float> @llvm.x86.avx512.mask.unpckl.ps.128(<4 x float>, <4 x float>, <4 x float>, i8)
+
+define <4 x float>@test_int_x86_avx512_mask_unpckl_ps_128(<4 x float> %x0, <4 x float> %x1, <4 x float> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_unpckl_ps_128:
+; CHECK:         vunpcklps %xmm1, %xmm0, %xmm2 {%k1}
+; CHECK-NEXT:    vunpcklps %xmm1, %xmm0, %xmm0 ## encoding: [0x62,0xf1,0x7c,0x08,0x14,0xc1]
+  %res = call <4 x float> @llvm.x86.avx512.mask.unpckl.ps.128(<4 x float> %x0, <4 x float> %x1, <4 x float> %x2, i8 %x3)
+  %res1 = call <4 x float> @llvm.x86.avx512.mask.unpckl.ps.128(<4 x float> %x0, <4 x float> %x1, <4 x float> %x2, i8 -1)
+  %res2 = fadd <4 x float> %res, %res1
+  ret <4 x float> %res2
+}
+
+declare <8 x float> @llvm.x86.avx512.mask.unpckl.ps.256(<8 x float>, <8 x float>, <8 x float>, i8)
+
+define <8 x float>@test_int_x86_avx512_mask_unpckl_ps_256(<8 x float> %x0, <8 x float> %x1, <8 x float> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_unpckl_ps_256:
+; CHECK:         vunpcklps %ymm1, %ymm0, %ymm2 {%k1}
+; CHECK-NEXT:    vunpcklps %ymm1, %ymm0, %ymm0 ## encoding: [0x62,0xf1,0x7c,0x28,0x14,0xc1]
+  %res = call <8 x float> @llvm.x86.avx512.mask.unpckl.ps.256(<8 x float> %x0, <8 x float> %x1, <8 x float> %x2, i8 %x3)
+  %res1 = call <8 x float> @llvm.x86.avx512.mask.unpckl.ps.256(<8 x float> %x0, <8 x float> %x1, <8 x float> %x2, i8 -1)
+  %res2 = fadd <8 x float> %res, %res1
+  ret <8 x float> %res2
+}
+
+declare <4 x i32> @llvm.x86.avx512.mask.punpckhd.q.128(<4 x i32>, <4 x i32>, <4 x i32>, i8)
+
+define <4 x i32>@test_int_x86_avx512_mask_punpckhd_q_128(<4 x i32> %x0, <4 x i32> %x1, <4 x i32> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpckhd_q_128:
+; CHECK:         vpunpckhdq %xmm1, %xmm0, %xmm2 {%k1}
+; CHECK-NEXT:    vpunpckhdq %xmm1, %xmm0, %xmm0 ## encoding: [0x62,0xf1,0x7d,0x08,0x6a,0xc1]
+  %res = call <4 x i32> @llvm.x86.avx512.mask.punpckhd.q.128(<4 x i32> %x0, <4 x i32> %x1, <4 x i32> %x2, i8 %x3)
+  %res1 = call <4 x i32> @llvm.x86.avx512.mask.punpckhd.q.128(<4 x i32> %x0, <4 x i32> %x1, <4 x i32> %x2, i8 -1)
+  %res2 = add <4 x i32> %res, %res1
+  ret <4 x i32> %res2
+}
+
+declare <4 x i32> @llvm.x86.avx512.mask.punpckld.q.128(<4 x i32>, <4 x i32>, <4 x i32>, i8)
+
+define <4 x i32>@test_int_x86_avx512_mask_punpckld_q_128(<4 x i32> %x0, <4 x i32> %x1, <4 x i32> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpckld_q_128:
+; CHECK:         vpunpckldq %xmm1, %xmm0, %xmm2 {%k1}
+; CHECK-NEXT:    vpunpckldq %xmm1, %xmm0, %xmm0 ## encoding: [0x62,0xf1,0x7d,0x08,0x62,0xc1]
+  %res = call <4 x i32> @llvm.x86.avx512.mask.punpckld.q.128(<4 x i32> %x0, <4 x i32> %x1, <4 x i32> %x2, i8 %x3)
+  %res1 = call <4 x i32> @llvm.x86.avx512.mask.punpckld.q.128(<4 x i32> %x0, <4 x i32> %x1, <4 x i32> %x2, i8 -1)
+  %res2 = add <4 x i32> %res, %res1
+  ret <4 x i32> %res2
+}
+
+declare <8 x i32> @llvm.x86.avx512.mask.punpckhd.q.256(<8 x i32>, <8 x i32>, <8 x i32>, i8)
+
+define <8 x i32>@test_int_x86_avx512_mask_punpckhd_q_256(<8 x i32> %x0, <8 x i32> %x1, <8 x i32> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpckhd_q_256:
+; CHECK:       ## BB#0:
+; CHECK:         vpunpckhdq %ymm1, %ymm0, %ymm2 {%k1}
+; CHECK-NEXT:    vpunpckhdq %ymm1, %ymm0, %ymm0 ## encoding: [0x62,0xf1,0x7d,0x28,0x6a,0xc1]
+  %res = call <8 x i32> @llvm.x86.avx512.mask.punpckhd.q.256(<8 x i32> %x0, <8 x i32> %x1, <8 x i32> %x2, i8 %x3)
+  %res1 = call <8 x i32> @llvm.x86.avx512.mask.punpckhd.q.256(<8 x i32> %x0, <8 x i32> %x1, <8 x i32> %x2, i8 -1)
+  %res2 = add <8 x i32> %res, %res1
+  ret <8 x i32> %res2
+}
+
+declare <8 x i32> @llvm.x86.avx512.mask.punpckld.q.256(<8 x i32>, <8 x i32>, <8 x i32>, i8)
+
+define <8 x i32>@test_int_x86_avx512_mask_punpckld_q_256(<8 x i32> %x0, <8 x i32> %x1, <8 x i32> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpckld_q_256:
+; CHECK:         vpunpckldq %ymm1, %ymm0, %ymm2 {%k1}
+; CHECK-NEXT:    vpunpckldq %ymm1, %ymm0, %ymm0 ## encoding: [0x62,0xf1,0x7d,0x28,0x62,0xc1]
+  %res = call <8 x i32> @llvm.x86.avx512.mask.punpckld.q.256(<8 x i32> %x0, <8 x i32> %x1, <8 x i32> %x2, i8 %x3)
+  %res1 = call <8 x i32> @llvm.x86.avx512.mask.punpckld.q.256(<8 x i32> %x0, <8 x i32> %x1, <8 x i32> %x2, i8 -1)
+  %res2 = add <8 x i32> %res, %res1
+  ret <8 x i32> %res2
+}
+
+declare <2 x i64> @llvm.x86.avx512.mask.punpckhqd.q.128(<2 x i64>, <2 x i64>, <2 x i64>, i8)
+
+define <2 x i64>@test_int_x86_avx512_mask_punpckhqd_q_128(<2 x i64> %x0, <2 x i64> %x1, <2 x i64> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpckhqd_q_128:
+; CHECK:         vpunpckhqdq %xmm1, %xmm0, %xmm2 {%k1}
+; CHECK-NEXT:    vpunpckhqdq %xmm1, %xmm0, %xmm0 ## encoding: [0x62,0xf1,0xfd,0x08,0x6d,0xc1]
+  %res = call <2 x i64> @llvm.x86.avx512.mask.punpckhqd.q.128(<2 x i64> %x0, <2 x i64> %x1, <2 x i64> %x2, i8 %x3)
+  %res1 = call <2 x i64> @llvm.x86.avx512.mask.punpckhqd.q.128(<2 x i64> %x0, <2 x i64> %x1, <2 x i64> %x2, i8 -1)
+  %res2 = add <2 x i64> %res, %res1
+  ret <2 x i64> %res2
+}
+
+declare <2 x i64> @llvm.x86.avx512.mask.punpcklqd.q.128(<2 x i64>, <2 x i64>, <2 x i64>, i8)
+
+define <2 x i64>@test_int_x86_avx512_mask_punpcklqd_q_128(<2 x i64> %x0, <2 x i64> %x1, <2 x i64> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpcklqd_q_128:
+; CHECK:         vpunpcklqdq %xmm1, %xmm0, %xmm2 {%k1}
+; CHECK-NEXT:    vpunpcklqdq %xmm1, %xmm0, %xmm0 ## encoding: [0x62,0xf1,0xfd,0x08,0x6c,0xc1]
+  %res = call <2 x i64> @llvm.x86.avx512.mask.punpcklqd.q.128(<2 x i64> %x0, <2 x i64> %x1, <2 x i64> %x2, i8 %x3)
+  %res1 = call <2 x i64> @llvm.x86.avx512.mask.punpcklqd.q.128(<2 x i64> %x0, <2 x i64> %x1, <2 x i64> %x2, i8 -1)
+  %res2 = add <2 x i64> %res, %res1
+  ret <2 x i64> %res2
+}
+
+declare <4 x i64> @llvm.x86.avx512.mask.punpcklqd.q.256(<4 x i64>, <4 x i64>, <4 x i64>, i8)
+
+define <4 x i64>@test_int_x86_avx512_mask_punpcklqd_q_256(<4 x i64> %x0, <4 x i64> %x1, <4 x i64> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpcklqd_q_256:
+; CHECK:         vpunpcklqdq %ymm1, %ymm0, %ymm2 {%k1}
+; CHECK-NEXT:    vpunpcklqdq %ymm1, %ymm0, %ymm0 ## encoding: [0x62,0xf1,0xfd,0x28,0x6c,0xc1]
+  %res = call <4 x i64> @llvm.x86.avx512.mask.punpcklqd.q.256(<4 x i64> %x0, <4 x i64> %x1, <4 x i64> %x2, i8 %x3)
+  %res1 = call <4 x i64> @llvm.x86.avx512.mask.punpcklqd.q.256(<4 x i64> %x0, <4 x i64> %x1, <4 x i64> %x2, i8 -1)
+  %res2 = add <4 x i64> %res, %res1
+  ret <4 x i64> %res2
+}
+
+declare <4 x i64> @llvm.x86.avx512.mask.punpckhqd.q.256(<4 x i64>, <4 x i64>, <4 x i64>, i8)
+
+define <4 x i64>@test_int_x86_avx512_mask_punpckhqd_q_256(<4 x i64> %x0, <4 x i64> %x1, <4 x i64> %x2, i8 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_punpckhqd_q_256:
+; CHECK:         vpunpckhqdq %ymm1, %ymm0, %ymm2 {%k1}
+; CHECK-NEXT:    vpunpckhqdq %ymm1, %ymm0, %ymm0 ## encoding: [0x62,0xf1,0xfd,0x28,0x6d,0xc1]
+  %res = call <4 x i64> @llvm.x86.avx512.mask.punpckhqd.q.256(<4 x i64> %x0, <4 x i64> %x1, <4 x i64> %x2, i8 %x3)
+  %res1 = call <4 x i64> @llvm.x86.avx512.mask.punpckhqd.q.256(<4 x i64> %x0, <4 x i64> %x1, <4 x i64> %x2, i8 -1)
+  %res2 = add <4 x i64> %res, %res1
+  ret <4 x i64> %res2
+}
+
 declare <16 x i8> @llvm.x86.avx512.mask.pmov.qb.128(<2 x i64>, <16 x i8>, i8)
 
 define <16 x i8>@test_int_x86_avx512_mask_pmov_qb_128(<2 x i64> %x0, <16 x i8> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmov_qb_128:
 ; CHECK:       vpmovqb %xmm0, %xmm1 {%k1}
 ; CHECK-NEXT:  vpmovqb %xmm0, %xmm2 {%k1} {z}
 ; CHECK-NEXT:  vpmovqb %xmm0, %xmm0
     %res0 = call <16 x i8> @llvm.x86.avx512.mask.pmov.qb.128(<2 x i64> %x0, <16 x i8> %x1, i8 -1)

test/MC/X86/avx512-encodings.s

 // CHECK: vfnmsub231sd -1024(%rdx), %xmm20, %xmm6
 // CHECK:  encoding: [0x62,0xf2,0xdd,0x00,0xbf,0x72,0x80]
           vfnmsub231sd -1024(%rdx), %xmm20, %xmm6
 
 // CHECK: vfnmsub231sd -1032(%rdx), %xmm20, %xmm6
 // CHECK:  encoding: [0x62,0xf2,0xdd,0x00,0xbf,0xb2,0xf8,0xfb,0xff,0xff]
           vfnmsub231sd -1032(%rdx), %xmm20, %xmm6
 
+// CHECK: vunpckhps %zmm16, %zmm14, %zmm5
+// CHECK:  encoding: [0x62,0xb1,0x0c,0x48,0x15,0xe8]
+          vunpckhps %zmm16, %zmm14, %zmm5
+
+// CHECK: vunpckhps %zmm16, %zmm14, %zmm5 {%k6}
+// CHECK:  encoding: [0x62,0xb1,0x0c,0x4e,0x15,0xe8]
+          vunpckhps %zmm16, %zmm14, %zmm5 {%k6}
+
+// CHECK: vunpckhps %zmm16, %zmm14, %zmm5 {%k6} {z}
+// CHECK:  encoding: [0x62,0xb1,0x0c,0xce,0x15,0xe8]
+          vunpckhps %zmm16, %zmm14, %zmm5 {%k6} {z}
+
+// CHECK: vunpckhps (%rcx), %zmm14, %zmm5
+// CHECK:  encoding: [0x62,0xf1,0x0c,0x48,0x15,0x29]
+          vunpckhps (%rcx), %zmm14, %zmm5
+
+// CHECK: vunpckhps 291(%rax,%r14,8), %zmm14, %zmm5
+// CHECK:  encoding: [0x62,0xb1,0x0c,0x48,0x15,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vunpckhps 291(%rax,%r14,8), %zmm14, %zmm5
+
+// CHECK: vunpckhps (%rcx){1to16}, %zmm14, %zmm5
+// CHECK:  encoding: [0x62,0xf1,0x0c,0x58,0x15,0x29]
+          vunpckhps (%rcx){1to16}, %zmm14, %zmm5
+
+// CHECK: vunpckhps 8128(%rdx), %zmm14, %zmm5
+// CHECK:  encoding: [0x62,0xf1,0x0c,0x48,0x15,0x6a,0x7f]
+          vunpckhps 8128(%rdx), %zmm14, %zmm5
+
+// CHECK: vunpckhps 8192(%rdx), %zmm14, %zmm5
+// CHECK:  encoding: [0x62,0xf1,0x0c,0x48,0x15,0xaa,0x00,0x20,0x00,0x00]
+          vunpckhps 8192(%rdx), %zmm14, %zmm5
+
+// CHECK: vunpckhps -8192(%rdx), %zmm14, %zmm5
+// CHECK:  encoding: [0x62,0xf1,0x0c,0x48,0x15,0x6a,0x80]
+          vunpckhps -8192(%rdx), %zmm14, %zmm5
+
+// CHECK: vunpckhps -8256(%rdx), %zmm14, %zmm5
+// CHECK:  encoding: [0x62,0xf1,0x0c,0x48,0x15,0xaa,0xc0,0xdf,0xff,0xff]
+          vunpckhps -8256(%rdx), %zmm14, %zmm5
+
+// CHECK: vunpckhps 508(%rdx){1to16}, %zmm14, %zmm5
+// CHECK:  encoding: [0x62,0xf1,0x0c,0x58,0x15,0x6a,0x7f]
+          vunpckhps 508(%rdx){1to16}, %zmm14, %zmm5
+
+// CHECK: vunpckhps 512(%rdx){1to16}, %zmm14, %zmm5
+// CHECK:  encoding: [0x62,0xf1,0x0c,0x58,0x15,0xaa,0x00,0x02,0x00,0x00]
+          vunpckhps 512(%rdx){1to16}, %zmm14, %zmm5
+
+// CHECK: vunpckhps -512(%rdx){1to16}, %zmm14, %zmm5
+// CHECK:  encoding: [0x62,0xf1,0x0c,0x58,0x15,0x6a,0x80]
+          vunpckhps -512(%rdx){1to16}, %zmm14, %zmm5
+
+// CHECK: vunpckhps -516(%rdx){1to16}, %zmm14, %zmm5
+// CHECK:  encoding: [0x62,0xf1,0x0c,0x58,0x15,0xaa,0xfc,0xfd,0xff,0xff]
+          vunpckhps -516(%rdx){1to16}, %zmm14, %zmm5
+
+// CHECK: vunpcklps %zmm2, %zmm3, %zmm1
+// CHECK:  encoding: [0x62,0xf1,0x64,0x48,0x14,0xca]
+          vunpcklps %zmm2, %zmm3, %zmm1
+
+// CHECK: vunpcklps %zmm2, %zmm3, %zmm1 {%k3}
+// CHECK:  encoding: [0x62,0xf1,0x64,0x4b,0x14,0xca]
+          vunpcklps %zmm2, %zmm3, %zmm1 {%k3}
+
+// CHECK: vunpcklps %zmm2, %zmm3, %zmm1 {%k3} {z}
+// CHECK:  encoding: [0x62,0xf1,0x64,0xcb,0x14,0xca]
+          vunpcklps %zmm2, %zmm3, %zmm1 {%k3} {z}
+
+// CHECK: vunpcklps (%rcx), %zmm3, %zmm1
+// CHECK:  encoding: [0x62,0xf1,0x64,0x48,0x14,0x09]
+          vunpcklps (%rcx), %zmm3, %zmm1
+
+// CHECK: vunpcklps 291(%rax,%r14,8), %zmm3, %zmm1
+// CHECK:  encoding: [0x62,0xb1,0x64,0x48,0x14,0x8c,0xf0,0x23,0x01,0x00,0x00]
+          vunpcklps 291(%rax,%r14,8), %zmm3, %zmm1
+
+// CHECK: vunpcklps (%rcx){1to16}, %zmm3, %zmm1
+// CHECK:  encoding: [0x62,0xf1,0x64,0x58,0x14,0x09]
+          vunpcklps (%rcx){1to16}, %zmm3, %zmm1
+
+// CHECK: vunpcklps 8128(%rdx), %zmm3, %zmm1
+// CHECK:  encoding: [0x62,0xf1,0x64,0x48,0x14,0x4a,0x7f]
+          vunpcklps 8128(%rdx), %zmm3, %zmm1
+
+// CHECK: vunpcklps 8192(%rdx), %zmm3, %zmm1
+// CHECK:  encoding: [0x62,0xf1,0x64,0x48,0x14,0x8a,0x00,0x20,0x00,0x00]
+          vunpcklps 8192(%rdx), %zmm3, %zmm1
+
+// CHECK: vunpcklps -8192(%rdx), %zmm3, %zmm1
+// CHECK:  encoding: [0x62,0xf1,0x64,0x48,0x14,0x4a,0x80]
+          vunpcklps -8192(%rdx), %zmm3, %zmm1
+
+// CHECK: vunpcklps -8256(%rdx), %zmm3, %zmm1
+// CHECK:  encoding: [0x62,0xf1,0x64,0x48,0x14,0x8a,0xc0,0xdf,0xff,0xff]
+          vunpcklps -8256(%rdx), %zmm3, %zmm1
+
+// CHECK: vunpcklps 508(%rdx){1to16}, %zmm3, %zmm1
+// CHECK:  encoding: [0x62,0xf1,0x64,0x58,0x14,0x4a,0x7f]
+          vunpcklps 508(%rdx){1to16}, %zmm3, %zmm1
+
+// CHECK: vunpcklps 512(%rdx){1to16}, %zmm3, %zmm1
+// CHECK:  encoding: [0x62,0xf1,0x64,0x58,0x14,0x8a,0x00,0x02,0x00,0x00]
+          vunpcklps 512(%rdx){1to16}, %zmm3, %zmm1
+
+// CHECK: vunpcklps -512(%rdx){1to16}, %zmm3, %zmm1
+// CHECK:  encoding: [0x62,0xf1,0x64,0x58,0x14,0x4a,0x80]
+          vunpcklps -512(%rdx){1to16}, %zmm3, %zmm1
+
+// CHECK: vunpcklps -516(%rdx){1to16}, %zmm3, %zmm1
+// CHECK:  encoding: [0x62,0xf1,0x64,0x58,0x14,0x8a,0xfc,0xfd,0xff,0xff]
+          vunpcklps -516(%rdx){1to16}, %zmm3, %zmm1
+
+// CHECK: vunpckhpd %zmm26, %zmm19, %zmm25
+// CHECK:  encoding: [0x62,0x01,0xe5,0x40,0x15,0xca]
+          vunpckhpd %zmm26, %zmm19, %zmm25
+
+// CHECK: vunpckhpd %zmm26, %zmm19, %zmm25 {%k5}
+// CHECK:  encoding: [0x62,0x01,0xe5,0x45,0x15,0xca]
+          vunpckhpd %zmm26, %zmm19, %zmm25 {%k5}
+
+// CHECK: vunpckhpd %zmm26, %zmm19, %zmm25 {%k5} {z}
+// CHECK:  encoding: [0x62,0x01,0xe5,0xc5,0x15,0xca]
+          vunpckhpd %zmm26, %zmm19, %zmm25 {%k5} {z}
+
+// CHECK: vunpckhpd (%rcx), %zmm19, %zmm25
+// CHECK:  encoding: [0x62,0x61,0xe5,0x40,0x15,0x09]
+          vunpckhpd (%rcx), %zmm19, %zmm25
+
+// CHECK: vunpckhpd 291(%rax,%r14,8), %zmm19, %zmm25
+// CHECK:  encoding: [0x62,0x21,0xe5,0x40,0x15,0x8c,0xf0,0x23,0x01,0x00,0x00]
+          vunpckhpd 291(%rax,%r14,8), %zmm19, %zmm25
+
+// CHECK: vunpckhpd (%rcx){1to8}, %zmm19, %zmm25
+// CHECK:  encoding: [0x62,0x61,0xe5,0x50,0x15,0x09]
+          vunpckhpd (%rcx){1to8}, %zmm19, %zmm25
+
+// CHECK: vunpckhpd 8128(%rdx), %zmm19, %zmm25
+// CHECK:  encoding: [0x62,0x61,0xe5,0x40,0x15,0x4a,0x7f]
+          vunpckhpd 8128(%rdx), %zmm19, %zmm25
+
+// CHECK: vunpckhpd 8192(%rdx), %zmm19, %zmm25
+// CHECK:  encoding: [0x62,0x61,0xe5,0x40,0x15,0x8a,0x00,0x20,0x00,0x00]
+          vunpckhpd 8192(%rdx), %zmm19, %zmm25
+
+// CHECK: vunpckhpd -8192(%rdx), %zmm19, %zmm25
+// CHECK:  encoding: [0x62,0x61,0xe5,0x40,0x15,0x4a,0x80]
+          vunpckhpd -8192(%rdx), %zmm19, %zmm25
+
+// CHECK: vunpckhpd -8256(%rdx), %zmm19, %zmm25
+// CHECK:  encoding: [0x62,0x61,0xe5,0x40,0x15,0x8a,0xc0,0xdf,0xff,0xff]
+          vunpckhpd -8256(%rdx), %zmm19, %zmm25
+
+// CHECK: vunpckhpd 1016(%rdx){1to8}, %zmm19, %zmm25
+// CHECK:  encoding: [0x62,0x61,0xe5,0x50,0x15,0x4a,0x7f]
+          vunpckhpd 1016(%rdx){1to8}, %zmm19, %zmm25
+
+// CHECK: vunpckhpd 1024(%rdx){1to8}, %zmm19, %zmm25
+// CHECK:  encoding: [0x62,0x61,0xe5,0x50,0x15,0x8a,0x00,0x04,0x00,0x00]
+          vunpckhpd 1024(%rdx){1to8}, %zmm19, %zmm25
+
+// CHECK: vunpckhpd -1024(%rdx){1to8}, %zmm19, %zmm25
+// CHECK:  encoding: [0x62,0x61,0xe5,0x50,0x15,0x4a,0x80]
+          vunpckhpd -1024(%rdx){1to8}, %zmm19, %zmm25
+
+// CHECK: vunpckhpd -1032(%rdx){1to8}, %zmm19, %zmm25
+// CHECK:  encoding: [0x62,0x61,0xe5,0x50,0x15,0x8a,0xf8,0xfb,0xff,0xff]
+          vunpckhpd -1032(%rdx){1to8}, %zmm19, %zmm25
+
+// CHECK: vunpcklpd %zmm21, %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xa1,0x95,0x40,0x14,0xd5]
+          vunpcklpd %zmm21, %zmm29, %zmm18
+
+// CHECK: vunpcklpd %zmm21, %zmm29, %zmm18 {%k6}
+// CHECK:  encoding: [0x62,0xa1,0x95,0x46,0x14,0xd5]
+          vunpcklpd %zmm21, %zmm29, %zmm18 {%k6}
+
+// CHECK: vunpcklpd %zmm21, %zmm29, %zmm18 {%k6} {z}
+// CHECK:  encoding: [0x62,0xa1,0x95,0xc6,0x14,0xd5]
+          vunpcklpd %zmm21, %zmm29, %zmm18 {%k6} {z}
+
+// CHECK: vunpcklpd (%rcx), %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xe1,0x95,0x40,0x14,0x11]
+          vunpcklpd (%rcx), %zmm29, %zmm18
+
+// CHECK: vunpcklpd 291(%rax,%r14,8), %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xa1,0x95,0x40,0x14,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vunpcklpd 291(%rax,%r14,8), %zmm29, %zmm18
+
+// CHECK: vunpcklpd (%rcx){1to8}, %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xe1,0x95,0x50,0x14,0x11]
+          vunpcklpd (%rcx){1to8}, %zmm29, %zmm18
+
+// CHECK: vunpcklpd 8128(%rdx), %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xe1,0x95,0x40,0x14,0x52,0x7f]
+          vunpcklpd 8128(%rdx), %zmm29, %zmm18
+
+// CHECK: vunpcklpd 8192(%rdx), %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xe1,0x95,0x40,0x14,0x92,0x00,0x20,0x00,0x00]
+          vunpcklpd 8192(%rdx), %zmm29, %zmm18
+
+// CHECK: vunpcklpd -8192(%rdx), %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xe1,0x95,0x40,0x14,0x52,0x80]
+          vunpcklpd -8192(%rdx), %zmm29, %zmm18
+
+// CHECK: vunpcklpd -8256(%rdx), %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xe1,0x95,0x40,0x14,0x92,0xc0,0xdf,0xff,0xff]
+          vunpcklpd -8256(%rdx), %zmm29, %zmm18
+
+// CHECK: vunpcklpd 1016(%rdx){1to8}, %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xe1,0x95,0x50,0x14,0x52,0x7f]
+          vunpcklpd 1016(%rdx){1to8}, %zmm29, %zmm18
+
+// CHECK: vunpcklpd 1024(%rdx){1to8}, %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xe1,0x95,0x50,0x14,0x92,0x00,0x04,0x00,0x00]
+          vunpcklpd 1024(%rdx){1to8}, %zmm29, %zmm18
+
+// CHECK: vunpcklpd -1024(%rdx){1to8}, %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xe1,0x95,0x50,0x14,0x52,0x80]
+          vunpcklpd -1024(%rdx){1to8}, %zmm29, %zmm18
+
+// CHECK: vunpcklpd -1032(%rdx){1to8}, %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xe1,0x95,0x50,0x14,0x92,0xf8,0xfb,0xff,0xff]
+          vunpcklpd -1032(%rdx){1to8}, %zmm29, %zmm18
+
+// CHECK: vpunpckldq %zmm17, %zmm3, %zmm24
+// CHECK:  encoding: [0x62,0x21,0x65,0x48,0x62,0xc1]
+          vpunpckldq %zmm17, %zmm3, %zmm24
+
+// CHECK: vpunpckldq %zmm17, %zmm3, %zmm24 {%k3}
+// CHECK:  encoding: [0x62,0x21,0x65,0x4b,0x62,0xc1]
+          vpunpckldq %zmm17, %zmm3, %zmm24 {%k3}
+
+// CHECK: vpunpckldq %zmm17, %zmm3, %zmm24 {%k3} {z}
+// CHECK:  encoding: [0x62,0x21,0x65,0xcb,0x62,0xc1]
+          vpunpckldq %zmm17, %zmm3, %zmm24 {%k3} {z}
+
+// CHECK: vpunpckldq (%rcx), %zmm3, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x65,0x48,0x62,0x01]
+          vpunpckldq (%rcx), %zmm3, %zmm24
+
+// CHECK: vpunpckldq 291(%rax,%r14,8), %zmm3, %zmm24
+// CHECK:  encoding: [0x62,0x21,0x65,0x48,0x62,0x84,0xf0,0x23,0x01,0x00,0x00]
+          vpunpckldq 291(%rax,%r14,8), %zmm3, %zmm24
+
+// CHECK: vpunpckldq (%rcx){1to16}, %zmm3, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x65,0x58,0x62,0x01]
+          vpunpckldq (%rcx){1to16}, %zmm3, %zmm24
+
+// CHECK: vpunpckldq 8128(%rdx), %zmm3, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x65,0x48,0x62,0x42,0x7f]
+          vpunpckldq 8128(%rdx), %zmm3, %zmm24
+
+// CHECK: vpunpckldq 8192(%rdx), %zmm3, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x65,0x48,0x62,0x82,0x00,0x20,0x00,0x00]
+          vpunpckldq 8192(%rdx), %zmm3, %zmm24
+
+// CHECK: vpunpckldq -8192(%rdx), %zmm3, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x65,0x48,0x62,0x42,0x80]
+          vpunpckldq -8192(%rdx), %zmm3, %zmm24
+
+// CHECK: vpunpckldq -8256(%rdx), %zmm3, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x65,0x48,0x62,0x82,0xc0,0xdf,0xff,0xff]
+          vpunpckldq -8256(%rdx), %zmm3, %zmm24
+
+// CHECK: vpunpckldq 508(%rdx){1to16}, %zmm3, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x65,0x58,0x62,0x42,0x7f]
+          vpunpckldq 508(%rdx){1to16}, %zmm3, %zmm24
+
+// CHECK: vpunpckldq 512(%rdx){1to16}, %zmm3, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x65,0x58,0x62,0x82,0x00,0x02,0x00,0x00]
+          vpunpckldq 512(%rdx){1to16}, %zmm3, %zmm24
+
+// CHECK: vpunpckldq -512(%rdx){1to16}, %zmm3, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x65,0x58,0x62,0x42,0x80]
+          vpunpckldq -512(%rdx){1to16}, %zmm3, %zmm24
+
+// CHECK: vpunpckldq -516(%rdx){1to16}, %zmm3, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x65,0x58,0x62,0x82,0xfc,0xfd,0xff,0xff]
+          vpunpckldq -516(%rdx){1to16}, %zmm3, %zmm24
+
+// CHECK: vpunpckhdq %zmm13, %zmm4, %zmm6
+// CHECK:  encoding: [0x62,0xd1,0x5d,0x48,0x6a,0xf5]
+          vpunpckhdq %zmm13, %zmm4, %zmm6
+
+// CHECK: vpunpckhdq %zmm13, %zmm4, %zmm6 {%k5}
+// CHECK:  encoding: [0x62,0xd1,0x5d,0x4d,0x6a,0xf5]
+          vpunpckhdq %zmm13, %zmm4, %zmm6 {%k5}
+
+// CHECK: vpunpckhdq %zmm13, %zmm4, %zmm6 {%k5} {z}
+// CHECK:  encoding: [0x62,0xd1,0x5d,0xcd,0x6a,0xf5]
+          vpunpckhdq %zmm13, %zmm4, %zmm6 {%k5} {z}
+
+// CHECK: vpunpckhdq (%rcx), %zmm4, %zmm6
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x48,0x6a,0x31]
+          vpunpckhdq (%rcx), %zmm4, %zmm6
+
+// CHECK: vpunpckhdq 291(%rax,%r14,8), %zmm4, %zmm6
+// CHECK:  encoding: [0x62,0xb1,0x5d,0x48,0x6a,0xb4,0xf0,0x23,0x01,0x00,0x00]
+          vpunpckhdq 291(%rax,%r14,8), %zmm4, %zmm6
+
+// CHECK: vpunpckhdq (%rcx){1to16}, %zmm4, %zmm6
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x58,0x6a,0x31]
+          vpunpckhdq (%rcx){1to16}, %zmm4, %zmm6
+
+// CHECK: vpunpckhdq 8128(%rdx), %zmm4, %zmm6
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x48,0x6a,0x72,0x7f]
+          vpunpckhdq 8128(%rdx), %zmm4, %zmm6
+
+// CHECK: vpunpckhdq 8192(%rdx), %zmm4, %zmm6
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x48,0x6a,0xb2,0x00,0x20,0x00,0x00]
+          vpunpckhdq 8192(%rdx), %zmm4, %zmm6
+
+// CHECK: vpunpckhdq -8192(%rdx), %zmm4, %zmm6
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x48,0x6a,0x72,0x80]
+          vpunpckhdq -8192(%rdx), %zmm4, %zmm6
+
+// CHECK: vpunpckhdq -8256(%rdx), %zmm4, %zmm6
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x48,0x6a,0xb2,0xc0,0xdf,0xff,0xff]
+          vpunpckhdq -8256(%rdx), %zmm4, %zmm6
+
+// CHECK: vpunpckhdq 508(%rdx){1to16}, %zmm4, %zmm6
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x58,0x6a,0x72,0x7f]
+          vpunpckhdq 508(%rdx){1to16}, %zmm4, %zmm6
+
+// CHECK: vpunpckhdq 512(%rdx){1to16}, %zmm4, %zmm6
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x58,0x6a,0xb2,0x00,0x02,0x00,0x00]
+          vpunpckhdq 512(%rdx){1to16}, %zmm4, %zmm6
+
+// CHECK: vpunpckhdq -512(%rdx){1to16}, %zmm4, %zmm6
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x58,0x6a,0x72,0x80]
+          vpunpckhdq -512(%rdx){1to16}, %zmm4, %zmm6
+
+// CHECK: vpunpckhdq -516(%rdx){1to16}, %zmm4, %zmm6
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x58,0x6a,0xb2,0xfc,0xfd,0xff,0xff]
+          vpunpckhdq -516(%rdx){1to16}, %zmm4, %zmm6
+
+// CHECK: vpunpcklqdq %zmm17, %zmm4, %zmm3
+// CHECK:  encoding: [0x62,0xb1,0xdd,0x48,0x6c,0xd9]
+          vpunpcklqdq %zmm17, %zmm4, %zmm3
+
+// CHECK: vpunpcklqdq %zmm17, %zmm4, %zmm3 {%k1}
+// CHECK:  encoding: [0x62,0xb1,0xdd,0x49,0x6c,0xd9]
+          vpunpcklqdq %zmm17, %zmm4, %zmm3 {%k1}
+
+// CHECK: vpunpcklqdq %zmm17, %zmm4, %zmm3 {%k1} {z}
+// CHECK:  encoding: [0x62,0xb1,0xdd,0xc9,0x6c,0xd9]
+          vpunpcklqdq %zmm17, %zmm4, %zmm3 {%k1} {z}
+
+// CHECK: vpunpcklqdq (%rcx), %zmm4, %zmm3
+// CHECK:  encoding: [0x62,0xf1,0xdd,0x48,0x6c,0x19]
+          vpunpcklqdq (%rcx), %zmm4, %zmm3
+
+// CHECK: vpunpcklqdq 291(%rax,%r14,8), %zmm4, %zmm3
+// CHECK:  encoding: [0x62,0xb1,0xdd,0x48,0x6c,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vpunpcklqdq 291(%rax,%r14,8), %zmm4, %zmm3
+
+// CHECK: vpunpcklqdq (%rcx){1to8}, %zmm4, %zmm3
+// CHECK:  encoding: [0x62,0xf1,0xdd,0x58,0x6c,0x19]
+          vpunpcklqdq (%rcx){1to8}, %zmm4, %zmm3
+
+// CHECK: vpunpcklqdq 8128(%rdx), %zmm4, %zmm3
+// CHECK:  encoding: [0x62,0xf1,0xdd,0x48,0x6c,0x5a,0x7f]
+          vpunpcklqdq 8128(%rdx), %zmm4, %zmm3
+
+// CHECK: vpunpcklqdq 8192(%rdx), %zmm4, %zmm3
+// CHECK:  encoding: [0x62,0xf1,0xdd,0x48,0x6c,0x9a,0x00,0x20,0x00,0x00]
+          vpunpcklqdq 8192(%rdx), %zmm4, %zmm3
+
+// CHECK: vpunpcklqdq -8192(%rdx), %zmm4, %zmm3
+// CHECK:  encoding: [0x62,0xf1,0xdd,0x48,0x6c,0x5a,0x80]
+          vpunpcklqdq -8192(%rdx), %zmm4, %zmm3
+
+// CHECK: vpunpcklqdq -8256(%rdx), %zmm4, %zmm3
+// CHECK:  encoding: [0x62,0xf1,0xdd,0x48,0x6c,0x9a,0xc0,0xdf,0xff,0xff]
+          vpunpcklqdq -8256(%rdx), %zmm4, %zmm3
+
+// CHECK: vpunpcklqdq 1016(%rdx){1to8}, %zmm4, %zmm3
+// CHECK:  encoding: [0x62,0xf1,0xdd,0x58,0x6c,0x5a,0x7f]
+          vpunpcklqdq 1016(%rdx){1to8}, %zmm4, %zmm3
+
+// CHECK: vpunpcklqdq 1024(%rdx){1to8}, %zmm4, %zmm3
+// CHECK:  encoding: [0x62,0xf1,0xdd,0x58,0x6c,0x9a,0x00,0x04,0x00,0x00]
+          vpunpcklqdq 1024(%rdx){1to8}, %zmm4, %zmm3
+
+// CHECK: vpunpcklqdq -1024(%rdx){1to8}, %zmm4, %zmm3
+// CHECK:  encoding: [0x62,0xf1,0xdd,0x58,0x6c,0x5a,0x80]
+          vpunpcklqdq -1024(%rdx){1to8}, %zmm4, %zmm3
+
+// CHECK: vpunpcklqdq -1032(%rdx){1to8}, %zmm4, %zmm3
+// CHECK:  encoding: [0x62,0xf1,0xdd,0x58,0x6c,0x9a,0xf8,0xfb,0xff,0xff]
+          vpunpcklqdq -1032(%rdx){1to8}, %zmm4, %zmm3
+
+// CHECK: vpunpckhqdq %zmm16, %zmm15, %zmm27
+// CHECK:  encoding: [0x62,0x21,0x85,0x48,0x6d,0xd8]
+          vpunpckhqdq %zmm16, %zmm15, %zmm27
+
+// CHECK: vpunpckhqdq %zmm16, %zmm15, %zmm27 {%k3}
+// CHECK:  encoding: [0x62,0x21,0x85,0x4b,0x6d,0xd8]
+          vpunpckhqdq %zmm16, %zmm15, %zmm27 {%k3}
+
+// CHECK: vpunpckhqdq %zmm16, %zmm15, %zmm27 {%k3} {z}
+// CHECK:  encoding: [0x62,0x21,0x85,0xcb,0x6d,0xd8]
+          vpunpckhqdq %zmm16, %zmm15, %zmm27 {%k3} {z}
+
+// CHECK: vpunpckhqdq (%rcx), %zmm15, %zmm27
+// CHECK:  encoding: [0x62,0x61,0x85,0x48,0x6d,0x19]
+          vpunpckhqdq (%rcx), %zmm15, %zmm27
+
+// CHECK: vpunpckhqdq 291(%rax,%r14,8), %zmm15, %zmm27
+// CHECK:  encoding: [0x62,0x21,0x85,0x48,0x6d,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vpunpckhqdq 291(%rax,%r14,8), %zmm15, %zmm27
+
+// CHECK: vpunpckhqdq (%rcx){1to8}, %zmm15, %zmm27
+// CHECK:  encoding: [0x62,0x61,0x85,0x58,0x6d,0x19]
+          vpunpckhqdq (%rcx){1to8}, %zmm15, %zmm27
+
+// CHECK: vpunpckhqdq 8128(%rdx), %zmm15, %zmm27
+// CHECK:  encoding: [0x62,0x61,0x85,0x48,0x6d,0x5a,0x7f]
+          vpunpckhqdq 8128(%rdx), %zmm15, %zmm27
+
+// CHECK: vpunpckhqdq 8192(%rdx), %zmm15, %zmm27
+// CHECK:  encoding: [0x62,0x61,0x85,0x48,0x6d,0x9a,0x00,0x20,0x00,0x00]
+          vpunpckhqdq 8192(%rdx), %zmm15, %zmm27
+
+// CHECK: vpunpckhqdq -8192(%rdx), %zmm15, %zmm27
+// CHECK:  encoding: [0x62,0x61,0x85,0x48,0x6d,0x5a,0x80]
+          vpunpckhqdq -8192(%rdx), %zmm15, %zmm27
+
+// CHECK: vpunpckhqdq -8256(%rdx), %zmm15, %zmm27
+// CHECK:  encoding: [0x62,0x61,0x85,0x48,0x6d,0x9a,0xc0,0xdf,0xff,0xff]
+          vpunpckhqdq -8256(%rdx), %zmm15, %zmm27
+
+// CHECK: vpunpckhqdq 1016(%rdx){1to8}, %zmm15, %zmm27
+// CHECK:  encoding: [0x62,0x61,0x85,0x58,0x6d,0x5a,0x7f]
+          vpunpckhqdq 1016(%rdx){1to8}, %zmm15, %zmm27
+
+// CHECK: vpunpckhqdq 1024(%rdx){1to8}, %zmm15, %zmm27
+// CHECK:  encoding: [0x62,0x61,0x85,0x58,0x6d,0x9a,0x00,0x04,0x00,0x00]
+          vpunpckhqdq 1024(%rdx){1to8}, %zmm15, %zmm27
+
+// CHECK: vpunpckhqdq -1024(%rdx){1to8}, %zmm15, %zmm27
+// CHECK:  encoding: [0x62,0x61,0x85,0x58,0x6d,0x5a,0x80]
+          vpunpckhqdq -1024(%rdx){1to8}, %zmm15, %zmm27
+
+// CHECK: vpunpckhqdq -1032(%rdx){1to8}, %zmm15, %zmm27
+// CHECK:  encoding: [0x62,0x61,0x85,0x58,0x6d,0x9a,0xf8,0xfb,0xff,0xff]
+          vpunpckhqdq -1032(%rdx){1to8}, %zmm15, %zmm27
+

test/MC/X86/x86-64-avx512bw.s

 // CHECK: vpmaddwd -8192(%rdx), %zmm22, %zmm26
 // CHECK:  encoding: [0x62,0x61,0x4d,0x40,0xf5,0x52,0x80]
           vpmaddwd -8192(%rdx), %zmm22, %zmm26
 
 // CHECK: vpmaddwd -8256(%rdx), %zmm22, %zmm26
 // CHECK:  encoding: [0x62,0x61,0x4d,0x40,0xf5,0x92,0xc0,0xdf,0xff,0xff]
           vpmaddwd -8256(%rdx), %zmm22, %zmm26
 
+// CHECK: vpunpcklbw %zmm24, %zmm17, %zmm28
+// CHECK:  encoding: [0x62,0x01,0x75,0x40,0x60,0xe0]
+          vpunpcklbw %zmm24, %zmm17, %zmm28
+
+// CHECK: vpunpcklbw %zmm24, %zmm17, %zmm28 {%k1}
+// CHECK:  encoding: [0x62,0x01,0x75,0x41,0x60,0xe0]
+          vpunpcklbw %zmm24, %zmm17, %zmm28 {%k1}
+
+// CHECK: vpunpcklbw %zmm24, %zmm17, %zmm28 {%k1} {z}
+// CHECK:  encoding: [0x62,0x01,0x75,0xc1,0x60,0xe0]
+          vpunpcklbw %zmm24, %zmm17, %zmm28 {%k1} {z}
+
+// CHECK: vpunpcklbw (%rcx), %zmm17, %zmm28
+// CHECK:  encoding: [0x62,0x61,0x75,0x40,0x60,0x21]
+          vpunpcklbw (%rcx), %zmm17, %zmm28
+
+// CHECK: vpunpcklbw 291(%rax,%r14,8), %zmm17, %zmm28
+// CHECK:  encoding: [0x62,0x21,0x75,0x40,0x60,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpunpcklbw 291(%rax,%r14,8), %zmm17, %zmm28
+
+// CHECK: vpunpcklbw 8128(%rdx), %zmm17, %zmm28
+// CHECK:  encoding: [0x62,0x61,0x75,0x40,0x60,0x62,0x7f]
+          vpunpcklbw 8128(%rdx), %zmm17, %zmm28
+
+// CHECK: vpunpcklbw 8192(%rdx), %zmm17, %zmm28
+// CHECK:  encoding: [0x62,0x61,0x75,0x40,0x60,0xa2,0x00,0x20,0x00,0x00]
+          vpunpcklbw 8192(%rdx), %zmm17, %zmm28
+
+// CHECK: vpunpcklbw -8192(%rdx), %zmm17, %zmm28
+// CHECK:  encoding: [0x62,0x61,0x75,0x40,0x60,0x62,0x80]
+          vpunpcklbw -8192(%rdx), %zmm17, %zmm28
+
+// CHECK: vpunpcklbw -8256(%rdx), %zmm17, %zmm28
+// CHECK:  encoding: [0x62,0x61,0x75,0x40,0x60,0xa2,0xc0,0xdf,0xff,0xff]
+          vpunpcklbw -8256(%rdx), %zmm17, %zmm28
+
+// CHECK: vpunpckhbw %zmm23, %zmm19, %zmm30
+// CHECK:  encoding: [0x62,0x21,0x65,0x40,0x68,0xf7]
+          vpunpckhbw %zmm23, %zmm19, %zmm30
+
+// CHECK: vpunpckhbw %zmm23, %zmm19, %zmm30 {%k7}
+// CHECK:  encoding: [0x62,0x21,0x65,0x47,0x68,0xf7]
+          vpunpckhbw %zmm23, %zmm19, %zmm30 {%k7}
+
+// CHECK: vpunpckhbw %zmm23, %zmm19, %zmm30 {%k7} {z}
+// CHECK:  encoding: [0x62,0x21,0x65,0xc7,0x68,0xf7]
+          vpunpckhbw %zmm23, %zmm19, %zmm30 {%k7} {z}
+
+// CHECK: vpunpckhbw (%rcx), %zmm19, %zmm30
+// CHECK:  encoding: [0x62,0x61,0x65,0x40,0x68,0x31]
+          vpunpckhbw (%rcx), %zmm19, %zmm30
+
+// CHECK: vpunpckhbw 291(%rax,%r14,8), %zmm19, %zmm30
+// CHECK:  encoding: [0x62,0x21,0x65,0x40,0x68,0xb4,0xf0,0x23,0x01,0x00,0x00]
+          vpunpckhbw 291(%rax,%r14,8), %zmm19, %zmm30
+
+// CHECK: vpunpckhbw 8128(%rdx), %zmm19, %zmm30
+// CHECK:  encoding: [0x62,0x61,0x65,0x40,0x68,0x72,0x7f]
+          vpunpckhbw 8128(%rdx), %zmm19, %zmm30
+
+// CHECK: vpunpckhbw 8192(%rdx), %zmm19, %zmm30
+// CHECK:  encoding: [0x62,0x61,0x65,0x40,0x68,0xb2,0x00,0x20,0x00,0x00]
+          vpunpckhbw 8192(%rdx), %zmm19, %zmm30
+
+// CHECK: vpunpckhbw -8192(%rdx), %zmm19, %zmm30
+// CHECK:  encoding: [0x62,0x61,0x65,0x40,0x68,0x72,0x80]
+          vpunpckhbw -8192(%rdx), %zmm19, %zmm30
+
+// CHECK: vpunpckhbw -8256(%rdx), %zmm19, %zmm30
+// CHECK:  encoding: [0x62,0x61,0x65,0x40,0x68,0xb2,0xc0,0xdf,0xff,0xff]
+          vpunpckhbw -8256(%rdx), %zmm19, %zmm30
+
+// CHECK: vpunpcklwd %zmm18, %zmm24, %zmm20
+// CHECK:  encoding: [0x62,0xa1,0x3d,0x40,0x61,0xe2]
+          vpunpcklwd %zmm18, %zmm24, %zmm20
+
+// CHECK: vpunpcklwd %zmm18, %zmm24, %zmm20 {%k7}
+// CHECK:  encoding: [0x62,0xa1,0x3d,0x47,0x61,0xe2]
+          vpunpcklwd %zmm18, %zmm24, %zmm20 {%k7}
+
+// CHECK: vpunpcklwd %zmm18, %zmm24, %zmm20 {%k7} {z}
+// CHECK:  encoding: [0x62,0xa1,0x3d,0xc7,0x61,0xe2]
+          vpunpcklwd %zmm18, %zmm24, %zmm20 {%k7} {z}
+
+// CHECK: vpunpcklwd (%rcx), %zmm24, %zmm20
+// CHECK:  encoding: [0x62,0xe1,0x3d,0x40,0x61,0x21]
+          vpunpcklwd (%rcx), %zmm24, %zmm20
+
+// CHECK: vpunpcklwd 291(%rax,%r14,8), %zmm24, %zmm20
+// CHECK:  encoding: [0x62,0xa1,0x3d,0x40,0x61,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpunpcklwd 291(%rax,%r14,8), %zmm24, %zmm20
+
+// CHECK: vpunpcklwd 8128(%rdx), %zmm24, %zmm20
+// CHECK:  encoding: [0x62,0xe1,0x3d,0x40,0x61,0x62,0x7f]
+          vpunpcklwd 8128(%rdx), %zmm24, %zmm20
+
+// CHECK: vpunpcklwd 8192(%rdx), %zmm24, %zmm20
+// CHECK:  encoding: [0x62,0xe1,0x3d,0x40,0x61,0xa2,0x00,0x20,0x00,0x00]
+          vpunpcklwd 8192(%rdx), %zmm24, %zmm20
+
+// CHECK: vpunpcklwd -8192(%rdx), %zmm24, %zmm20
+// CHECK:  encoding: [0x62,0xe1,0x3d,0x40,0x61,0x62,0x80]
+          vpunpcklwd -8192(%rdx), %zmm24, %zmm20
+
+// CHECK: vpunpcklwd -8256(%rdx), %zmm24, %zmm20
+// CHECK:  encoding: [0x62,0xe1,0x3d,0x40,0x61,0xa2,0xc0,0xdf,0xff,0xff]
+          vpunpcklwd -8256(%rdx), %zmm24, %zmm20
+
+// CHECK: vpunpckhwd %zmm24, %zmm26, %zmm30
+// CHECK:  encoding: [0x62,0x01,0x2d,0x40,0x69,0xf0]
+          vpunpckhwd %zmm24, %zmm26, %zmm30
+
+// CHECK: vpunpckhwd %zmm24, %zmm26, %zmm30 {%k4}
+// CHECK:  encoding: [0x62,0x01,0x2d,0x44,0x69,0xf0]
+          vpunpckhwd %zmm24, %zmm26, %zmm30 {%k4}
+
+// CHECK: vpunpckhwd %zmm24, %zmm26, %zmm30 {%k4} {z}
+// CHECK:  encoding: [0x62,0x01,0x2d,0xc4,0x69,0xf0]
+          vpunpckhwd %zmm24, %zmm26, %zmm30 {%k4} {z}
+
+// CHECK: vpunpckhwd (%rcx), %zmm26, %zmm30
+// CHECK:  encoding: [0x62,0x61,0x2d,0x40,0x69,0x31]
+          vpunpckhwd (%rcx), %zmm26, %zmm30
+
+// CHECK: vpunpckhwd 291(%rax,%r14,8), %zmm26, %zmm30
+// CHECK:  encoding: [0x62,0x21,0x2d,0x40,0x69,0xb4,0xf0,0x23,0x01,0x00,0x00]
+          vpunpckhwd 291(%rax,%r14,8), %zmm26, %zmm30
+
+// CHECK: vpunpckhwd 8128(%rdx), %zmm26, %zmm30
+// CHECK:  encoding: [0x62,0x61,0x2d,0x40,0x69,0x72,0x7f]
+          vpunpckhwd 8128(%rdx), %zmm26, %zmm30
+
+// CHECK: vpunpckhwd 8192(%rdx), %zmm26, %zmm30
+// CHECK:  encoding: [0x62,0x61,0x2d,0x40,0x69,0xb2,0x00,0x20,0x00,0x00]
+          vpunpckhwd 8192(%rdx), %zmm26, %zmm30
+
+// CHECK: vpunpckhwd -8192(%rdx), %zmm26, %zmm30
+// CHECK:  encoding: [0x62,0x61,0x2d,0x40,0x69,0x72,0x80]
+          vpunpckhwd -8192(%rdx), %zmm26, %zmm30
+
+// CHECK: vpunpckhwd -8256(%rdx), %zmm26, %zmm30
+// CHECK:  encoding: [0x62,0x61,0x2d,0x40,0x69,0xb2,0xc0,0xdf,0xff,0xff]
+          vpunpckhwd -8256(%rdx), %zmm26, %zmm30
+

test/MC/X86/x86-64-avx512bw_vl.s

 // CHECK: vpmaddwd -4096(%rdx), %ymm20, %ymm19
 // CHECK:  encoding: [0x62,0xe1,0x5d,0x20,0xf5,0x5a,0x80]
           vpmaddwd -4096(%rdx), %ymm20, %ymm19
 
 // CHECK: vpmaddwd -4128(%rdx), %ymm20, %ymm19
 // CHECK:  encoding: [0x62,0xe1,0x5d,0x20,0xf5,0x9a,0xe0,0xef,0xff,0xff]
           vpmaddwd -4128(%rdx), %ymm20, %ymm19
 
+// CHECK: vpunpcklbw %xmm20, %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x21,0x65,0x00,0x60,0xf4]
+          vpunpcklbw %xmm20, %xmm19, %xmm30
+
+// CHECK: vpunpcklbw %xmm20, %xmm19, %xmm30 {%k4}
+// CHECK:  encoding: [0x62,0x21,0x65,0x04,0x60,0xf4]
+          vpunpcklbw %xmm20, %xmm19, %xmm30 {%k4}
+
+// CHECK: vpunpcklbw %xmm20, %xmm19, %xmm30 {%k4} {z}
+// CHECK:  encoding: [0x62,0x21,0x65,0x84,0x60,0xf4]
+          vpunpcklbw %xmm20, %xmm19, %xmm30 {%k4} {z}
+
+// CHECK: vpunpcklbw (%rcx), %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x65,0x00,0x60,0x31]
+          vpunpcklbw (%rcx), %xmm19, %xmm30
+
+// CHECK: vpunpcklbw 4660(%rax,%r14,8), %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x21,0x65,0x00,0x60,0xb4,0xf0,0x34,0x12,0x00,0x00]
+          vpunpcklbw 4660(%rax,%r14,8), %xmm19, %xmm30
+
+// CHECK: vpunpcklbw 2032(%rdx), %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x65,0x00,0x60,0x72,0x7f]
+          vpunpcklbw 2032(%rdx), %xmm19, %xmm30
+
+// CHECK: vpunpcklbw 2048(%rdx), %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x65,0x00,0x60,0xb2,0x00,0x08,0x00,0x00]
+          vpunpcklbw 2048(%rdx), %xmm19, %xmm30
+
+// CHECK: vpunpcklbw -2048(%rdx), %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x65,0x00,0x60,0x72,0x80]
+          vpunpcklbw -2048(%rdx), %xmm19, %xmm30
+
+// CHECK: vpunpcklbw -2064(%rdx), %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x65,0x00,0x60,0xb2,0xf0,0xf7,0xff,0xff]
+          vpunpcklbw -2064(%rdx), %xmm19, %xmm30
+
+// CHECK: vpunpcklbw %ymm22, %ymm28, %ymm20
+// CHECK:  encoding: [0x62,0xa1,0x1d,0x20,0x60,0xe6]
+          vpunpcklbw %ymm22, %ymm28, %ymm20
+
+// CHECK: vpunpcklbw %ymm22, %ymm28, %ymm20 {%k1}
+// CHECK:  encoding: [0x62,0xa1,0x1d,0x21,0x60,0xe6]
+          vpunpcklbw %ymm22, %ymm28, %ymm20 {%k1}
+
+// CHECK: vpunpcklbw %ymm22, %ymm28, %ymm20 {%k1} {z}
+// CHECK:  encoding: [0x62,0xa1,0x1d,0xa1,0x60,0xe6]
+          vpunpcklbw %ymm22, %ymm28, %ymm20 {%k1} {z}
+
+// CHECK: vpunpcklbw (%rcx), %ymm28, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x1d,0x20,0x60,0x21]
+          vpunpcklbw (%rcx), %ymm28, %ymm20
+
+// CHECK: vpunpcklbw 4660(%rax,%r14,8), %ymm28, %ymm20
+// CHECK:  encoding: [0x62,0xa1,0x1d,0x20,0x60,0xa4,0xf0,0x34,0x12,0x00,0x00]
+          vpunpcklbw 4660(%rax,%r14,8), %ymm28, %ymm20
+
+// CHECK: vpunpcklbw 4064(%rdx), %ymm28, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x1d,0x20,0x60,0x62,0x7f]
+          vpunpcklbw 4064(%rdx), %ymm28, %ymm20
+
+// CHECK: vpunpcklbw 4096(%rdx), %ymm28, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x1d,0x20,0x60,0xa2,0x00,0x10,0x00,0x00]
+          vpunpcklbw 4096(%rdx), %ymm28, %ymm20
+
+// CHECK: vpunpcklbw -4096(%rdx), %ymm28, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x1d,0x20,0x60,0x62,0x80]
+          vpunpcklbw -4096(%rdx), %ymm28, %ymm20
+
+// CHECK: vpunpcklbw -4128(%rdx), %ymm28, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x1d,0x20,0x60,0xa2,0xe0,0xef,0xff,0xff]
+          vpunpcklbw -4128(%rdx), %ymm28, %ymm20
+
+// CHECK: vpunpckhbw %xmm27, %xmm22, %xmm18
+// CHECK:  encoding: [0x62,0x81,0x4d,0x00,0x68,0xd3]
+          vpunpckhbw %xmm27, %xmm22, %xmm18
+
+// CHECK: vpunpckhbw %xmm27, %xmm22, %xmm18 {%k1}
+// CHECK:  encoding: [0x62,0x81,0x4d,0x01,0x68,0xd3]
+          vpunpckhbw %xmm27, %xmm22, %xmm18 {%k1}
+
+// CHECK: vpunpckhbw %xmm27, %xmm22, %xmm18 {%k1} {z}
+// CHECK:  encoding: [0x62,0x81,0x4d,0x81,0x68,0xd3]
+          vpunpckhbw %xmm27, %xmm22, %xmm18 {%k1} {z}
+
+// CHECK: vpunpckhbw (%rcx), %xmm22, %xmm18
+// CHECK:  encoding: [0x62,0xe1,0x4d,0x00,0x68,0x11]
+          vpunpckhbw (%rcx), %xmm22, %xmm18
+
+// CHECK: vpunpckhbw 4660(%rax,%r14,8), %xmm22, %xmm18
+// CHECK:  encoding: [0x62,0xa1,0x4d,0x00,0x68,0x94,0xf0,0x34,0x12,0x00,0x00]
+          vpunpckhbw 4660(%rax,%r14,8), %xmm22, %xmm18
+
+// CHECK: vpunpckhbw 2032(%rdx), %xmm22, %xmm18
+// CHECK:  encoding: [0x62,0xe1,0x4d,0x00,0x68,0x52,0x7f]
+          vpunpckhbw 2032(%rdx), %xmm22, %xmm18
+
+// CHECK: vpunpckhbw 2048(%rdx), %xmm22, %xmm18
+// CHECK:  encoding: [0x62,0xe1,0x4d,0x00,0x68,0x92,0x00,0x08,0x00,0x00]
+          vpunpckhbw 2048(%rdx), %xmm22, %xmm18
+
+// CHECK: vpunpckhbw -2048(%rdx), %xmm22, %xmm18
+// CHECK:  encoding: [0x62,0xe1,0x4d,0x00,0x68,0x52,0x80]
+          vpunpckhbw -2048(%rdx), %xmm22, %xmm18
+
+// CHECK: vpunpckhbw -2064(%rdx), %xmm22, %xmm18
+// CHECK:  encoding: [0x62,0xe1,0x4d,0x00,0x68,0x92,0xf0,0xf7,0xff,0xff]
+          vpunpckhbw -2064(%rdx), %xmm22, %xmm18
+
+// CHECK: vpunpckhbw %ymm24, %ymm20, %ymm17
+// CHECK:  encoding: [0x62,0x81,0x5d,0x20,0x68,0xc8]
+          vpunpckhbw %ymm24, %ymm20, %ymm17
+
+// CHECK: vpunpckhbw %ymm24, %ymm20, %ymm17 {%k5}
+// CHECK:  encoding: [0x62,0x81,0x5d,0x25,0x68,0xc8]
+          vpunpckhbw %ymm24, %ymm20, %ymm17 {%k5}
+
+// CHECK: vpunpckhbw %ymm24, %ymm20, %ymm17 {%k5} {z}
+// CHECK:  encoding: [0x62,0x81,0x5d,0xa5,0x68,0xc8]
+          vpunpckhbw %ymm24, %ymm20, %ymm17 {%k5} {z}
+
+// CHECK: vpunpckhbw (%rcx), %ymm20, %ymm17
+// CHECK:  encoding: [0x62,0xe1,0x5d,0x20,0x68,0x09]
+          vpunpckhbw (%rcx), %ymm20, %ymm17
+
+// CHECK: vpunpckhbw 4660(%rax,%r14,8), %ymm20, %ymm17
+// CHECK:  encoding: [0x62,0xa1,0x5d,0x20,0x68,0x8c,0xf0,0x34,0x12,0x00,0x00]
+          vpunpckhbw 4660(%rax,%r14,8), %ymm20, %ymm17
+
+// CHECK: vpunpckhbw 4064(%rdx), %ymm20, %ymm17
+// CHECK:  encoding: [0x62,0xe1,0x5d,0x20,0x68,0x4a,0x7f]
+          vpunpckhbw 4064(%rdx), %ymm20, %ymm17
+
+// CHECK: vpunpckhbw 4096(%rdx), %ymm20, %ymm17
+// CHECK:  encoding: [0x62,0xe1,0x5d,0x20,0x68,0x8a,0x00,0x10,0x00,0x00]
+          vpunpckhbw 4096(%rdx), %ymm20, %ymm17
+
+// CHECK: vpunpckhbw -4096(%rdx), %ymm20, %ymm17
+// CHECK:  encoding: [0x62,0xe1,0x5d,0x20,0x68,0x4a,0x80]
+          vpunpckhbw -4096(%rdx), %ymm20, %ymm17
+
+// CHECK: vpunpckhbw -4128(%rdx), %ymm20, %ymm17
+// CHECK:  encoding: [0x62,0xe1,0x5d,0x20,0x68,0x8a,0xe0,0xef,0xff,0xff]
+          vpunpckhbw -4128(%rdx), %ymm20, %ymm17
+
+// CHECK: vpunpcklwd %xmm17, %xmm27, %xmm27
+// CHECK:  encoding: [0x62,0x21,0x25,0x00,0x61,0xd9]
+          vpunpcklwd %xmm17, %xmm27, %xmm27
+
+// CHECK: vpunpcklwd %xmm17, %xmm27, %xmm27 {%k5}
+// CHECK:  encoding: [0x62,0x21,0x25,0x05,0x61,0xd9]
+          vpunpcklwd %xmm17, %xmm27, %xmm27 {%k5}
+
+// CHECK: vpunpcklwd %xmm17, %xmm27, %xmm27 {%k5} {z}
+// CHECK:  encoding: [0x62,0x21,0x25,0x85,0x61,0xd9]
+          vpunpcklwd %xmm17, %xmm27, %xmm27 {%k5} {z}
+
+// CHECK: vpunpcklwd (%rcx), %xmm27, %xmm27
+// CHECK:  encoding: [0x62,0x61,0x25,0x00,0x61,0x19]
+          vpunpcklwd (%rcx), %xmm27, %xmm27
+
+// CHECK: vpunpcklwd 4660(%rax,%r14,8), %xmm27, %xmm27
+// CHECK:  encoding: [0x62,0x21,0x25,0x00,0x61,0x9c,0xf0,0x34,0x12,0x00,0x00]
+          vpunpcklwd 4660(%rax,%r14,8), %xmm27, %xmm27
+
+// CHECK: vpunpcklwd 2032(%rdx), %xmm27, %xmm27
+// CHECK:  encoding: [0x62,0x61,0x25,0x00,0x61,0x5a,0x7f]
+          vpunpcklwd 2032(%rdx), %xmm27, %xmm27
+
+// CHECK: vpunpcklwd 2048(%rdx), %xmm27, %xmm27
+// CHECK:  encoding: [0x62,0x61,0x25,0x00,0x61,0x9a,0x00,0x08,0x00,0x00]
+          vpunpcklwd 2048(%rdx), %xmm27, %xmm27
+
+// CHECK: vpunpcklwd -2048(%rdx), %xmm27, %xmm27
+// CHECK:  encoding: [0x62,0x61,0x25,0x00,0x61,0x5a,0x80]
+          vpunpcklwd -2048(%rdx), %xmm27, %xmm27
+
+// CHECK: vpunpcklwd -2064(%rdx), %xmm27, %xmm27
+// CHECK:  encoding: [0x62,0x61,0x25,0x00,0x61,0x9a,0xf0,0xf7,0xff,0xff]
+          vpunpcklwd -2064(%rdx), %xmm27, %xmm27
+
+// CHECK: vpunpcklwd %ymm23, %ymm25, %ymm18
+// CHECK:  encoding: [0x62,0xa1,0x35,0x20,0x61,0xd7]
+          vpunpcklwd %ymm23, %ymm25, %ymm18
+
+// CHECK: vpunpcklwd %ymm23, %ymm25, %ymm18 {%k5}
+// CHECK:  encoding: [0x62,0xa1,0x35,0x25,0x61,0xd7]
+          vpunpcklwd %ymm23, %ymm25, %ymm18 {%k5}
+
+// CHECK: vpunpcklwd %ymm23, %ymm25, %ymm18 {%k5} {z}
+// CHECK:  encoding: [0x62,0xa1,0x35,0xa5,0x61,0xd7]
+          vpunpcklwd %ymm23, %ymm25, %ymm18 {%k5} {z}
+
+// CHECK: vpunpcklwd (%rcx), %ymm25, %ymm18
+// CHECK:  encoding: [0x62,0xe1,0x35,0x20,0x61,0x11]
+          vpunpcklwd (%rcx), %ymm25, %ymm18
+
+// CHECK: vpunpcklwd 4660(%rax,%r14,8), %ymm25, %ymm18
+// CHECK:  encoding: [0x62,0xa1,0x35,0x20,0x61,0x94,0xf0,0x34,0x12,0x00,0x00]
+          vpunpcklwd 4660(%rax,%r14,8), %ymm25, %ymm18
+
+// CHECK: vpunpcklwd 4064(%rdx), %ymm25, %ymm18
+// CHECK:  encoding: [0x62,0xe1,0x35,0x20,0x61,0x52,0x7f]
+          vpunpcklwd 4064(%rdx), %ymm25, %ymm18
+
+// CHECK: vpunpcklwd 4096(%rdx), %ymm25, %ymm18
+// CHECK:  encoding: [0x62,0xe1,0x35,0x20,0x61,0x92,0x00,0x10,0x00,0x00]
+          vpunpcklwd 4096(%rdx), %ymm25, %ymm18
+
+// CHECK: vpunpcklwd -4096(%rdx), %ymm25, %ymm18
+// CHECK:  encoding: [0x62,0xe1,0x35,0x20,0x61,0x52,0x80]
+          vpunpcklwd -4096(%rdx), %ymm25, %ymm18
+
+// CHECK: vpunpcklwd -4128(%rdx), %ymm25, %ymm18
+// CHECK:  encoding: [0x62,0xe1,0x35,0x20,0x61,0x92,0xe0,0xef,0xff,0xff]
+          vpunpcklwd -4128(%rdx), %ymm25, %ymm18
+
+// CHECK: vpunpckhwd %xmm17, %xmm28, %xmm17
+// CHECK:  encoding: [0x62,0xa1,0x1d,0x00,0x69,0xc9]
+          vpunpckhwd %xmm17, %xmm28, %xmm17
+
+// CHECK: vpunpckhwd %xmm17, %xmm28, %xmm17 {%k7}
+// CHECK:  encoding: [0x62,0xa1,0x1d,0x07,0x69,0xc9]
+          vpunpckhwd %xmm17, %xmm28, %xmm17 {%k7}
+
+// CHECK: vpunpckhwd %xmm17, %xmm28, %xmm17 {%k7} {z}
+// CHECK:  encoding: [0x62,0xa1,0x1d,0x87,0x69,0xc9]
+          vpunpckhwd %xmm17, %xmm28, %xmm17 {%k7} {z}
+
+// CHECK: vpunpckhwd (%rcx), %xmm28, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0x1d,0x00,0x69,0x09]
+          vpunpckhwd (%rcx), %xmm28, %xmm17
+
+// CHECK: vpunpckhwd 4660(%rax,%r14,8), %xmm28, %xmm17
+// CHECK:  encoding: [0x62,0xa1,0x1d,0x00,0x69,0x8c,0xf0,0x34,0x12,0x00,0x00]
+          vpunpckhwd 4660(%rax,%r14,8), %xmm28, %xmm17
+
+// CHECK: vpunpckhwd 2032(%rdx), %xmm28, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0x1d,0x00,0x69,0x4a,0x7f]
+          vpunpckhwd 2032(%rdx), %xmm28, %xmm17
+
+// CHECK: vpunpckhwd 2048(%rdx), %xmm28, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0x1d,0x00,0x69,0x8a,0x00,0x08,0x00,0x00]
+          vpunpckhwd 2048(%rdx), %xmm28, %xmm17
+
+// CHECK: vpunpckhwd -2048(%rdx), %xmm28, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0x1d,0x00,0x69,0x4a,0x80]
+          vpunpckhwd -2048(%rdx), %xmm28, %xmm17
+
+// CHECK: vpunpckhwd -2064(%rdx), %xmm28, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0x1d,0x00,0x69,0x8a,0xf0,0xf7,0xff,0xff]
+          vpunpckhwd -2064(%rdx), %xmm28, %xmm17
+
+// CHECK: vpunpckhwd %ymm20, %ymm25, %ymm24
+// CHECK:  encoding: [0x62,0x21,0x35,0x20,0x69,0xc4]
+          vpunpckhwd %ymm20, %ymm25, %ymm24
+
+// CHECK: vpunpckhwd %ymm20, %ymm25, %ymm24 {%k1}
+// CHECK:  encoding: [0x62,0x21,0x35,0x21,0x69,0xc4]
+          vpunpckhwd %ymm20, %ymm25, %ymm24 {%k1}
+
+// CHECK: vpunpckhwd %ymm20, %ymm25, %ymm24 {%k1} {z}
+// CHECK:  encoding: [0x62,0x21,0x35,0xa1,0x69,0xc4]
+          vpunpckhwd %ymm20, %ymm25, %ymm24 {%k1} {z}
+
+// CHECK: vpunpckhwd (%rcx), %ymm25, %ymm24
+// CHECK:  encoding: [0x62,0x61,0x35,0x20,0x69,0x01]
+          vpunpckhwd (%rcx), %ymm25, %ymm24
+
+// CHECK: vpunpckhwd 4660(%rax,%r14,8), %ymm25, %ymm24
+// CHECK:  encoding: [0x62,0x21,0x35,0x20,0x69,0x84,0xf0,0x34,0x12,0x00,0x00]
+          vpunpckhwd 4660(%rax,%r14,8), %ymm25, %ymm24
+
+// CHECK: vpunpckhwd 4064(%rdx), %ymm25, %ymm24
+// CHECK:  encoding: [0x62,0x61,0x35,0x20,0x69,0x42,0x7f]
+          vpunpckhwd 4064(%rdx), %ymm25, %ymm24
+
+// CHECK: vpunpckhwd 4096(%rdx), %ymm25, %ymm24
+// CHECK:  encoding: [0x62,0x61,0x35,0x20,0x69,0x82,0x00,0x10,0x00,0x00]
+          vpunpckhwd 4096(%rdx), %ymm25, %ymm24
+
+// CHECK: vpunpckhwd -4096(%rdx), %ymm25, %ymm24
+// CHECK:  encoding: [0x62,0x61,0x35,0x20,0x69,0x42,0x80]
+          vpunpckhwd -4096(%rdx), %ymm25, %ymm24
+
+// CHECK: vpunpckhwd -4128(%rdx), %ymm25, %ymm24
+// CHECK:  encoding: [0x62,0x61,0x35,0x20,0x69,0x82,0xe0,0xef,0xff,0xff]
+          vpunpckhwd -4128(%rdx), %ymm25, %ymm24
+
+// CHECK: vpunpcklbw %xmm23, %xmm29, %xmm18
+// CHECK:  encoding: [0x62,0xa1,0x15,0x00,0x60,0xd7]
+          vpunpcklbw %xmm23, %xmm29, %xmm18
+
+// CHECK: vpunpcklbw %xmm23, %xmm29, %xmm18 {%k4}
+// CHECK:  encoding: [0x62,0xa1,0x15,0x04,0x60,0xd7]
+          vpunpcklbw %xmm23, %xmm29, %xmm18 {%k4}
+
+// CHECK: vpunpcklbw %xmm23, %xmm29, %xmm18 {%k4} {z}
+// CHECK:  encoding: [0x62,0xa1,0x15,0x84,0x60,0xd7]
+          vpunpcklbw %xmm23, %xmm29, %xmm18 {%k4} {z}
+
+// CHECK: vpunpcklbw (%rcx), %xmm29, %xmm18
+// CHECK:  encoding: [0x62,0xe1,0x15,0x00,0x60,0x11]
+          vpunpcklbw (%rcx), %xmm29, %xmm18
+
+// CHECK: vpunpcklbw 291(%rax,%r14,8), %xmm29, %xmm18
+// CHECK:  encoding: [0x62,0xa1,0x15,0x00,0x60,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpunpcklbw 291(%rax,%r14,8), %xmm29, %xmm18
+
+// CHECK: vpunpcklbw 2032(%rdx), %xmm29, %xmm18
+// CHECK:  encoding: [0x62,0xe1,0x15,0x00,0x60,0x52,0x7f]
+          vpunpcklbw 2032(%rdx), %xmm29, %xmm18
+
+// CHECK: vpunpcklbw 2048(%rdx), %xmm29, %xmm18
+// CHECK:  encoding: [0x62,0xe1,0x15,0x00,0x60,0x92,0x00,0x08,0x00,0x00]
+          vpunpcklbw 2048(%rdx), %xmm29, %xmm18
+
+// CHECK: vpunpcklbw -2048(%rdx), %xmm29, %xmm18
+// CHECK:  encoding: [0x62,0xe1,0x15,0x00,0x60,0x52,0x80]
+          vpunpcklbw -2048(%rdx), %xmm29, %xmm18
+
+// CHECK: vpunpcklbw -2064(%rdx), %xmm29, %xmm18
+// CHECK:  encoding: [0x62,0xe1,0x15,0x00,0x60,0x92,0xf0,0xf7,0xff,0xff]
+          vpunpcklbw -2064(%rdx), %xmm29, %xmm18
+
+// CHECK: vpunpcklbw %ymm21, %ymm28, %ymm27
+// CHECK:  encoding: [0x62,0x21,0x1d,0x20,0x60,0xdd]
+          vpunpcklbw %ymm21, %ymm28, %ymm27
+
+// CHECK: vpunpcklbw %ymm21, %ymm28, %ymm27 {%k4}
+// CHECK:  encoding: [0x62,0x21,0x1d,0x24,0x60,0xdd]
+          vpunpcklbw %ymm21, %ymm28, %ymm27 {%k4}
+
+// CHECK: vpunpcklbw %ymm21, %ymm28, %ymm27 {%k4} {z}
+// CHECK:  encoding: [0x62,0x21,0x1d,0xa4,0x60,0xdd]
+          vpunpcklbw %ymm21, %ymm28, %ymm27 {%k4} {z}
+
+// CHECK: vpunpcklbw (%rcx), %ymm28, %ymm27
+// CHECK:  encoding: [0x62,0x61,0x1d,0x20,0x60,0x19]
+          vpunpcklbw (%rcx), %ymm28, %ymm27
+
+// CHECK: vpunpcklbw 291(%rax,%r14,8), %ymm28, %ymm27
+// CHECK:  encoding: [0x62,0x21,0x1d,0x20,0x60,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vpunpcklbw 291(%rax,%r14,8), %ymm28, %ymm27
+
+// CHECK: vpunpcklbw 4064(%rdx), %ymm28, %ymm27
+// CHECK:  encoding: [0x62,0x61,0x1d,0x20,0x60,0x5a,0x7f]
+          vpunpcklbw 4064(%rdx), %ymm28, %ymm27
+
+// CHECK: vpunpcklbw 4096(%rdx), %ymm28, %ymm27
+// CHECK:  encoding: [0x62,0x61,0x1d,0x20,0x60,0x9a,0x00,0x10,0x00,0x00]
+          vpunpcklbw 4096(%rdx), %ymm28, %ymm27
+
+// CHECK: vpunpcklbw -4096(%rdx), %ymm28, %ymm27
+// CHECK:  encoding: [0x62,0x61,0x1d,0x20,0x60,0x5a,0x80]
+          vpunpcklbw -4096(%rdx), %ymm28, %ymm27
+
+// CHECK: vpunpcklbw -4128(%rdx), %ymm28, %ymm27
+// CHECK:  encoding: [0x62,0x61,0x1d,0x20,0x60,0x9a,0xe0,0xef,0xff,0xff]
+          vpunpcklbw -4128(%rdx), %ymm28, %ymm27
+
+// CHECK: vpunpckhbw %xmm24, %xmm21, %xmm18
+// CHECK:  encoding: [0x62,0x81,0x55,0x00,0x68,0xd0]
+          vpunpckhbw %xmm24, %xmm21, %xmm18
+
+// CHECK: vpunpckhbw %xmm24, %xmm21, %xmm18 {%k6}
+// CHECK:  encoding: [0x62,0x81,0x55,0x06,0x68,0xd0]
+          vpunpckhbw %xmm24, %xmm21, %xmm18 {%k6}
+
+// CHECK: vpunpckhbw %xmm24, %xmm21, %xmm18 {%k6} {z}
+// CHECK:  encoding: [0x62,0x81,0x55,0x86,0x68,0xd0]
+          vpunpckhbw %xmm24, %xmm21, %xmm18 {%k6} {z}
+
+// CHECK: vpunpckhbw (%rcx), %xmm21, %xmm18
+// CHECK:  encoding: [0x62,0xe1,0x55,0x00,0x68,0x11]
+          vpunpckhbw (%rcx), %xmm21, %xmm18
+
+// CHECK: vpunpckhbw 291(%rax,%r14,8), %xmm21, %xmm18
+// CHECK:  encoding: [0x62,0xa1,0x55,0x00,0x68,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpunpckhbw 291(%rax,%r14,8), %xmm21, %xmm18
+
+// CHECK: vpunpckhbw 2032(%rdx), %xmm21, %xmm18
+// CHECK:  encoding: [0x62,0xe1,0x55,0x00,0x68,0x52,0x7f]
+          vpunpckhbw 2032(%rdx), %xmm21, %xmm18
+
+// CHECK: vpunpckhbw 2048(%rdx), %xmm21, %xmm18
+// CHECK:  encoding: [0x62,0xe1,0x55,0x00,0x68,0x92,0x00,0x08,0x00,0x00]
+          vpunpckhbw 2048(%rdx), %xmm21, %xmm18
+
+// CHECK: vpunpckhbw -2048(%rdx), %xmm21, %xmm18
+// CHECK:  encoding: [0x62,0xe1,0x55,0x00,0x68,0x52,0x80]
+          vpunpckhbw -2048(%rdx), %xmm21, %xmm18
+
+// CHECK: vpunpckhbw -2064(%rdx), %xmm21, %xmm18
+// CHECK:  encoding: [0x62,0xe1,0x55,0x00,0x68,0x92,0xf0,0xf7,0xff,0xff]
+          vpunpckhbw -2064(%rdx), %xmm21, %xmm18
+
+// CHECK: vpunpckhbw %ymm23, %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xa1,0x2d,0x20,0x68,0xe7]
+          vpunpckhbw %ymm23, %ymm26, %ymm20
+
+// CHECK: vpunpckhbw %ymm23, %ymm26, %ymm20 {%k1}
+// CHECK:  encoding: [0x62,0xa1,0x2d,0x21,0x68,0xe7]
+          vpunpckhbw %ymm23, %ymm26, %ymm20 {%k1}
+
+// CHECK: vpunpckhbw %ymm23, %ymm26, %ymm20 {%k1} {z}
+// CHECK:  encoding: [0x62,0xa1,0x2d,0xa1,0x68,0xe7]
+          vpunpckhbw %ymm23, %ymm26, %ymm20 {%k1} {z}
+
+// CHECK: vpunpckhbw (%rcx), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x2d,0x20,0x68,0x21]
+          vpunpckhbw (%rcx), %ymm26, %ymm20
+
+// CHECK: vpunpckhbw 291(%rax,%r14,8), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xa1,0x2d,0x20,0x68,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpunpckhbw 291(%rax,%r14,8), %ymm26, %ymm20
+
+// CHECK: vpunpckhbw 4064(%rdx), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x2d,0x20,0x68,0x62,0x7f]
+          vpunpckhbw 4064(%rdx), %ymm26, %ymm20
+
+// CHECK: vpunpckhbw 4096(%rdx), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x2d,0x20,0x68,0xa2,0x00,0x10,0x00,0x00]
+          vpunpckhbw 4096(%rdx), %ymm26, %ymm20
+
+// CHECK: vpunpckhbw -4096(%rdx), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x2d,0x20,0x68,0x62,0x80]
+          vpunpckhbw -4096(%rdx), %ymm26, %ymm20
+
+// CHECK: vpunpckhbw -4128(%rdx), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x2d,0x20,0x68,0xa2,0xe0,0xef,0xff,0xff]
+          vpunpckhbw -4128(%rdx), %ymm26, %ymm20
+
+// CHECK: vpunpcklwd %xmm21, %xmm25, %xmm21
+// CHECK:  encoding: [0x62,0xa1,0x35,0x00,0x61,0xed]
+          vpunpcklwd %xmm21, %xmm25, %xmm21
+
+// CHECK: vpunpcklwd %xmm21, %xmm25, %xmm21 {%k6}
+// CHECK:  encoding: [0x62,0xa1,0x35,0x06,0x61,0xed]
+          vpunpcklwd %xmm21, %xmm25, %xmm21 {%k6}
+
+// CHECK: vpunpcklwd %xmm21, %xmm25, %xmm21 {%k6} {z}
+// CHECK:  encoding: [0x62,0xa1,0x35,0x86,0x61,0xed]
+          vpunpcklwd %xmm21, %xmm25, %xmm21 {%k6} {z}
+
+// CHECK: vpunpcklwd (%rcx), %xmm25, %xmm21
+// CHECK:  encoding: [0x62,0xe1,0x35,0x00,0x61,0x29]
+          vpunpcklwd (%rcx), %xmm25, %xmm21
+
+// CHECK: vpunpcklwd 291(%rax,%r14,8), %xmm25, %xmm21
+// CHECK:  encoding: [0x62,0xa1,0x35,0x00,0x61,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vpunpcklwd 291(%rax,%r14,8), %xmm25, %xmm21
+
+// CHECK: vpunpcklwd 2032(%rdx), %xmm25, %xmm21
+// CHECK:  encoding: [0x62,0xe1,0x35,0x00,0x61,0x6a,0x7f]
+          vpunpcklwd 2032(%rdx), %xmm25, %xmm21
+
+// CHECK: vpunpcklwd 2048(%rdx), %xmm25, %xmm21
+// CHECK:  encoding: [0x62,0xe1,0x35,0x00,0x61,0xaa,0x00,0x08,0x00,0x00]
+          vpunpcklwd 2048(%rdx), %xmm25, %xmm21
+
+// CHECK: vpunpcklwd -2048(%rdx), %xmm25, %xmm21
+// CHECK:  encoding: [0x62,0xe1,0x35,0x00,0x61,0x6a,0x80]
+          vpunpcklwd -2048(%rdx), %xmm25, %xmm21
+
+// CHECK: vpunpcklwd -2064(%rdx), %xmm25, %xmm21
+// CHECK:  encoding: [0x62,0xe1,0x35,0x00,0x61,0xaa,0xf0,0xf7,0xff,0xff]
+          vpunpcklwd -2064(%rdx), %xmm25, %xmm21
+
+// CHECK: vpunpcklwd %ymm26, %ymm23, %ymm26
+// CHECK:  encoding: [0x62,0x01,0x45,0x20,0x61,0xd2]
+          vpunpcklwd %ymm26, %ymm23, %ymm26
+
+// CHECK: vpunpcklwd %ymm26, %ymm23, %ymm26 {%k2}
+// CHECK:  encoding: [0x62,0x01,0x45,0x22,0x61,0xd2]
+          vpunpcklwd %ymm26, %ymm23, %ymm26 {%k2}
+
+// CHECK: vpunpcklwd %ymm26, %ymm23, %ymm26 {%k2} {z}
+// CHECK:  encoding: [0x62,0x01,0x45,0xa2,0x61,0xd2]
+          vpunpcklwd %ymm26, %ymm23, %ymm26 {%k2} {z}
+
+// CHECK: vpunpcklwd (%rcx), %ymm23, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x45,0x20,0x61,0x11]
+          vpunpcklwd (%rcx), %ymm23, %ymm26
+
+// CHECK: vpunpcklwd 291(%rax,%r14,8), %ymm23, %ymm26
+// CHECK:  encoding: [0x62,0x21,0x45,0x20,0x61,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpunpcklwd 291(%rax,%r14,8), %ymm23, %ymm26
+
+// CHECK: vpunpcklwd 4064(%rdx), %ymm23, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x45,0x20,0x61,0x52,0x7f]
+          vpunpcklwd 4064(%rdx), %ymm23, %ymm26
+
+// CHECK: vpunpcklwd 4096(%rdx), %ymm23, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x45,0x20,0x61,0x92,0x00,0x10,0x00,0x00]
+          vpunpcklwd 4096(%rdx), %ymm23, %ymm26
+
+// CHECK: vpunpcklwd -4096(%rdx), %ymm23, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x45,0x20,0x61,0x52,0x80]
+          vpunpcklwd -4096(%rdx), %ymm23, %ymm26
+
+// CHECK: vpunpcklwd -4128(%rdx), %ymm23, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x45,0x20,0x61,0x92,0xe0,0xef,0xff,0xff]
+          vpunpcklwd -4128(%rdx), %ymm23, %ymm26
+
+// CHECK: vpunpckhwd %xmm23, %xmm18, %xmm17
+// CHECK:  encoding: [0x62,0xa1,0x6d,0x00,0x69,0xcf]
+          vpunpckhwd %xmm23, %xmm18, %xmm17
+
+// CHECK: vpunpckhwd %xmm23, %xmm18, %xmm17 {%k3}
+// CHECK:  encoding: [0x62,0xa1,0x6d,0x03,0x69,0xcf]
+          vpunpckhwd %xmm23, %xmm18, %xmm17 {%k3}
+
+// CHECK: vpunpckhwd %xmm23, %xmm18, %xmm17 {%k3} {z}
+// CHECK:  encoding: [0x62,0xa1,0x6d,0x83,0x69,0xcf]
+          vpunpckhwd %xmm23, %xmm18, %xmm17 {%k3} {z}
+
+// CHECK: vpunpckhwd (%rcx), %xmm18, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0x6d,0x00,0x69,0x09]
+          vpunpckhwd (%rcx), %xmm18, %xmm17
+
+// CHECK: vpunpckhwd 291(%rax,%r14,8), %xmm18, %xmm17
+// CHECK:  encoding: [0x62,0xa1,0x6d,0x00,0x69,0x8c,0xf0,0x23,0x01,0x00,0x00]
+          vpunpckhwd 291(%rax,%r14,8), %xmm18, %xmm17
+
+// CHECK: vpunpckhwd 2032(%rdx), %xmm18, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0x6d,0x00,0x69,0x4a,0x7f]
+          vpunpckhwd 2032(%rdx), %xmm18, %xmm17
+
+// CHECK: vpunpckhwd 2048(%rdx), %xmm18, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0x6d,0x00,0x69,0x8a,0x00,0x08,0x00,0x00]
+          vpunpckhwd 2048(%rdx), %xmm18, %xmm17
+
+// CHECK: vpunpckhwd -2048(%rdx), %xmm18, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0x6d,0x00,0x69,0x4a,0x80]
+          vpunpckhwd -2048(%rdx), %xmm18, %xmm17
+
+// CHECK: vpunpckhwd -2064(%rdx), %xmm18, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0x6d,0x00,0x69,0x8a,0xf0,0xf7,0xff,0xff]
+          vpunpckhwd -2064(%rdx), %xmm18, %xmm17
+
+// CHECK: vpunpckhwd %ymm26, %ymm25, %ymm28
+// CHECK:  encoding: [0x62,0x01,0x35,0x20,0x69,0xe2]
+          vpunpckhwd %ymm26, %ymm25, %ymm28
+
+// CHECK: vpunpckhwd %ymm26, %ymm25, %ymm28 {%k4}
+// CHECK:  encoding: [0x62,0x01,0x35,0x24,0x69,0xe2]
+          vpunpckhwd %ymm26, %ymm25, %ymm28 {%k4}
+
+// CHECK: vpunpckhwd %ymm26, %ymm25, %ymm28 {%k4} {z}
+// CHECK:  encoding: [0x62,0x01,0x35,0xa4,0x69,0xe2]
+          vpunpckhwd %ymm26, %ymm25, %ymm28 {%k4} {z}
+
+// CHECK: vpunpckhwd (%rcx), %ymm25, %ymm28
+// CHECK:  encoding: [0x62,0x61,0x35,0x20,0x69,0x21]
+          vpunpckhwd (%rcx), %ymm25, %ymm28
+
+// CHECK: vpunpckhwd 291(%rax,%r14,8), %ymm25, %ymm28
+// CHECK:  encoding: [0x62,0x21,0x35,0x20,0x69,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpunpckhwd 291(%rax,%r14,8), %ymm25, %ymm28
+
+// CHECK: vpunpckhwd 4064(%rdx), %ymm25, %ymm28
+// CHECK:  encoding: [0x62,0x61,0x35,0x20,0x69,0x62,0x7f]
+          vpunpckhwd 4064(%rdx), %ymm25, %ymm28
+
+// CHECK: vpunpckhwd 4096(%rdx), %ymm25, %ymm28
+// CHECK:  encoding: [0x62,0x61,0x35,0x20,0x69,0xa2,0x00,0x10,0x00,0x00]
+          vpunpckhwd 4096(%rdx), %ymm25, %ymm28
+
+// CHECK: vpunpckhwd -4096(%rdx), %ymm25, %ymm28
+// CHECK:  encoding: [0x62,0x61,0x35,0x20,0x69,0x62,0x80]
+          vpunpckhwd -4096(%rdx), %ymm25, %ymm28
+
+// CHECK: vpunpckhwd -4128(%rdx), %ymm25, %ymm28
+// CHECK:  encoding: [0x62,0x61,0x35,0x20,0x69,0xa2,0xe0,0xef,0xff,0xff]
+          vpunpckhwd -4128(%rdx), %ymm25, %ymm28
+

test/MC/X86/x86-64-avx512f_vl.s

 
 // CHECK: vcvtdq2ps -512(%rdx){1to8}, %ymm24
 // CHECK:  encoding: [0x62,0x61,0x7c,0x38,0x5b,0x42,0x80]
           vcvtdq2ps -512(%rdx){1to8}, %ymm24
 
 // CHECK: vcvtdq2ps -516(%rdx){1to8}, %ymm24
 // CHECK:  encoding: [0x62,0x61,0x7c,0x38,0x5b,0x82,0xfc,0xfd,0xff,0xff]
           vcvtdq2ps -516(%rdx){1to8}, %ymm24
+// CHECK: vunpckhps %xmm27, %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x01,0x64,0x00,0x15,0xf3]
+          vunpckhps %xmm27, %xmm19, %xmm30
+
+// CHECK: vunpckhps %xmm27, %xmm19, %xmm30 {%k2}
+// CHECK:  encoding: [0x62,0x01,0x64,0x02,0x15,0xf3]
+          vunpckhps %xmm27, %xmm19, %xmm30 {%k2}
+
+// CHECK: vunpckhps %xmm27, %xmm19, %xmm30 {%k2} {z}
+// CHECK:  encoding: [0x62,0x01,0x64,0x82,0x15,0xf3]
+          vunpckhps %xmm27, %xmm19, %xmm30 {%k2} {z}
+
+// CHECK: vunpckhps (%rcx), %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x64,0x00,0x15,0x31]
+          vunpckhps (%rcx), %xmm19, %xmm30
+
+// CHECK: vunpckhps 291(%rax,%r14,8), %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x21,0x64,0x00,0x15,0xb4,0xf0,0x23,0x01,0x00,0x00]
+          vunpckhps 291(%rax,%r14,8), %xmm19, %xmm30
+
+// CHECK: vunpckhps (%rcx){1to4}, %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x64,0x10,0x15,0x31]
+          vunpckhps (%rcx){1to4}, %xmm19, %xmm30
+
+// CHECK: vunpckhps 2032(%rdx), %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x64,0x00,0x15,0x72,0x7f]
+          vunpckhps 2032(%rdx), %xmm19, %xmm30
+
+// CHECK: vunpckhps 2048(%rdx), %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x64,0x00,0x15,0xb2,0x00,0x08,0x00,0x00]
+          vunpckhps 2048(%rdx), %xmm19, %xmm30
+
+// CHECK: vunpckhps -2048(%rdx), %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x64,0x00,0x15,0x72,0x80]
+          vunpckhps -2048(%rdx), %xmm19, %xmm30
+
+// CHECK: vunpckhps -2064(%rdx), %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x64,0x00,0x15,0xb2,0xf0,0xf7,0xff,0xff]
+          vunpckhps -2064(%rdx), %xmm19, %xmm30
+
+// CHECK: vunpckhps 508(%rdx){1to4}, %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x64,0x10,0x15,0x72,0x7f]
+          vunpckhps 508(%rdx){1to4}, %xmm19, %xmm30
+
+// CHECK: vunpckhps 512(%rdx){1to4}, %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x64,0x10,0x15,0xb2,0x00,0x02,0x00,0x00]
+          vunpckhps 512(%rdx){1to4}, %xmm19, %xmm30
+
+// CHECK: vunpckhps -512(%rdx){1to4}, %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x64,0x10,0x15,0x72,0x80]
+          vunpckhps -512(%rdx){1to4}, %xmm19, %xmm30
+
+// CHECK: vunpckhps -516(%rdx){1to4}, %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x64,0x10,0x15,0xb2,0xfc,0xfd,0xff,0xff]
+          vunpckhps -516(%rdx){1to4}, %xmm19, %xmm30
+
+// CHECK: vunpckhps %ymm25, %ymm25, %ymm21
+// CHECK:  encoding: [0x62,0x81,0x34,0x20,0x15,0xe9]
+          vunpckhps %ymm25, %ymm25, %ymm21
+
+// CHECK: vunpckhps %ymm25, %ymm25, %ymm21 {%k5}
+// CHECK:  encoding: [0x62,0x81,0x34,0x25,0x15,0xe9]
+          vunpckhps %ymm25, %ymm25, %ymm21 {%k5}
+
+// CHECK: vunpckhps %ymm25, %ymm25, %ymm21 {%k5} {z}
+// CHECK:  encoding: [0x62,0x81,0x34,0xa5,0x15,0xe9]
+          vunpckhps %ymm25, %ymm25, %ymm21 {%k5} {z}
+
+// CHECK: vunpckhps (%rcx), %ymm25, %ymm21
+// CHECK:  encoding: [0x62,0xe1,0x34,0x20,0x15,0x29]
+          vunpckhps (%rcx), %ymm25, %ymm21
+
+// CHECK: vunpckhps 291(%rax,%r14,8), %ymm25, %ymm21
+// CHECK:  encoding: [0x62,0xa1,0x34,0x20,0x15,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vunpckhps 291(%rax,%r14,8), %ymm25, %ymm21
+
+// CHECK: vunpckhps (%rcx){1to8}, %ymm25, %ymm21
+// CHECK:  encoding: [0x62,0xe1,0x34,0x30,0x15,0x29]
+          vunpckhps (%rcx){1to8}, %ymm25, %ymm21
+
+// CHECK: vunpckhps 4064(%rdx), %ymm25, %ymm21
+// CHECK:  encoding: [0x62,0xe1,0x34,0x20,0x15,0x6a,0x7f]
+          vunpckhps 4064(%rdx), %ymm25, %ymm21
+
+// CHECK: vunpckhps 4096(%rdx), %ymm25, %ymm21
+// CHECK:  encoding: [0x62,0xe1,0x34,0x20,0x15,0xaa,0x00,0x10,0x00,0x00]
+          vunpckhps 4096(%rdx), %ymm25, %ymm21
+
+// CHECK: vunpckhps -4096(%rdx), %ymm25, %ymm21
+// CHECK:  encoding: [0x62,0xe1,0x34,0x20,0x15,0x6a,0x80]
+          vunpckhps -4096(%rdx), %ymm25, %ymm21
+
+// CHECK: vunpckhps -4128(%rdx), %ymm25, %ymm21
+// CHECK:  encoding: [0x62,0xe1,0x34,0x20,0x15,0xaa,0xe0,0xef,0xff,0xff]
+          vunpckhps -4128(%rdx), %ymm25, %ymm21
+
+// CHECK: vunpckhps 508(%rdx){1to8}, %ymm25, %ymm21
+// CHECK:  encoding: [0x62,0xe1,0x34,0x30,0x15,0x6a,0x7f]
+          vunpckhps 508(%rdx){1to8}, %ymm25, %ymm21
+
+// CHECK: vunpckhps 512(%rdx){1to8}, %ymm25, %ymm21
+// CHECK:  encoding: [0x62,0xe1,0x34,0x30,0x15,0xaa,0x00,0x02,0x00,0x00]
+          vunpckhps 512(%rdx){1to8}, %ymm25, %ymm21
+
+// CHECK: vunpckhps -512(%rdx){1to8}, %ymm25, %ymm21
+// CHECK:  encoding: [0x62,0xe1,0x34,0x30,0x15,0x6a,0x80]
+          vunpckhps -512(%rdx){1to8}, %ymm25, %ymm21
+
+// CHECK: vunpckhps -516(%rdx){1to8}, %ymm25, %ymm21
+// CHECK:  encoding: [0x62,0xe1,0x34,0x30,0x15,0xaa,0xfc,0xfd,0xff,0xff]
+          vunpckhps -516(%rdx){1to8}, %ymm25, %ymm21
+
+// CHECK: vunpcklps %xmm26, %xmm24, %xmm29
+// CHECK:  encoding: [0x62,0x01,0x3c,0x00,0x14,0xea]
+          vunpcklps %xmm26, %xmm24, %xmm29
+
+// CHECK: vunpcklps %xmm26, %xmm24, %xmm29 {%k6}
+// CHECK:  encoding: [0x62,0x01,0x3c,0x06,0x14,0xea]
+          vunpcklps %xmm26, %xmm24, %xmm29 {%k6}
+
+// CHECK: vunpcklps %xmm26, %xmm24, %xmm29 {%k6} {z}
+// CHECK:  encoding: [0x62,0x01,0x3c,0x86,0x14,0xea]
+          vunpcklps %xmm26, %xmm24, %xmm29 {%k6} {z}
+
+// CHECK: vunpcklps (%rcx), %xmm24, %xmm29
+// CHECK:  encoding: [0x62,0x61,0x3c,0x00,0x14,0x29]
+          vunpcklps (%rcx), %xmm24, %xmm29
+
+// CHECK: vunpcklps 291(%rax,%r14,8), %xmm24, %xmm29
+// CHECK:  encoding: [0x62,0x21,0x3c,0x00,0x14,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vunpcklps 291(%rax,%r14,8), %xmm24, %xmm29
+
+// CHECK: vunpcklps (%rcx){1to4}, %xmm24, %xmm29
+// CHECK:  encoding: [0x62,0x61,0x3c,0x10,0x14,0x29]
+          vunpcklps (%rcx){1to4}, %xmm24, %xmm29
+
+// CHECK: vunpcklps 2032(%rdx), %xmm24, %xmm29
+// CHECK:  encoding: [0x62,0x61,0x3c,0x00,0x14,0x6a,0x7f]
+          vunpcklps 2032(%rdx), %xmm24, %xmm29
+
+// CHECK: vunpcklps 2048(%rdx), %xmm24, %xmm29
+// CHECK:  encoding: [0x62,0x61,0x3c,0x00,0x14,0xaa,0x00,0x08,0x00,0x00]
+          vunpcklps 2048(%rdx), %xmm24, %xmm29
+
+// CHECK: vunpcklps -2048(%rdx), %xmm24, %xmm29
+// CHECK:  encoding: [0x62,0x61,0x3c,0x00,0x14,0x6a,0x80]
+          vunpcklps -2048(%rdx), %xmm24, %xmm29
+
+// CHECK: vunpcklps -2064(%rdx), %xmm24, %xmm29
+// CHECK:  encoding: [0x62,0x61,0x3c,0x00,0x14,0xaa,0xf0,0xf7,0xff,0xff]
+          vunpcklps -2064(%rdx), %xmm24, %xmm29
+
+// CHECK: vunpcklps 508(%rdx){1to4}, %xmm24, %xmm29
+// CHECK:  encoding: [0x62,0x61,0x3c,0x10,0x14,0x6a,0x7f]
+          vunpcklps 508(%rdx){1to4}, %xmm24, %xmm29
+
+// CHECK: vunpcklps 512(%rdx){1to4}, %xmm24, %xmm29
+// CHECK:  encoding: [0x62,0x61,0x3c,0x10,0x14,0xaa,0x00,0x02,0x00,0x00]
+          vunpcklps 512(%rdx){1to4}, %xmm24, %xmm29
+
+// CHECK: vunpcklps -512(%rdx){1to4}, %xmm24, %xmm29
+// CHECK:  encoding: [0x62,0x61,0x3c,0x10,0x14,0x6a,0x80]
+          vunpcklps -512(%rdx){1to4}, %xmm24, %xmm29
+
+// CHECK: vunpcklps -516(%rdx){1to4}, %xmm24, %xmm29
+// CHECK:  encoding: [0x62,0x61,0x3c,0x10,0x14,0xaa,0xfc,0xfd,0xff,0xff]
+          vunpcklps -516(%rdx){1to4}, %xmm24, %xmm29
+
+// CHECK: vunpcklps %ymm18, %ymm18, %ymm26
+// CHECK:  encoding: [0x62,0x21,0x6c,0x20,0x14,0xd2]
+          vunpcklps %ymm18, %ymm18, %ymm26
+
+// CHECK: vunpcklps %ymm18, %ymm18, %ymm26 {%k1}
+// CHECK:  encoding: [0x62,0x21,0x6c,0x21,0x14,0xd2]
+          vunpcklps %ymm18, %ymm18, %ymm26 {%k1}
+
+// CHECK: vunpcklps %ymm18, %ymm18, %ymm26 {%k1} {z}
+// CHECK:  encoding: [0x62,0x21,0x6c,0xa1,0x14,0xd2]
+          vunpcklps %ymm18, %ymm18, %ymm26 {%k1} {z}
+
+// CHECK: vunpcklps (%rcx), %ymm18, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x6c,0x20,0x14,0x11]
+          vunpcklps (%rcx), %ymm18, %ymm26
+
+// CHECK: vunpcklps 291(%rax,%r14,8), %ymm18, %ymm26
+// CHECK:  encoding: [0x62,0x21,0x6c,0x20,0x14,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vunpcklps 291(%rax,%r14,8), %ymm18, %ymm26
+
+// CHECK: vunpcklps (%rcx){1to8}, %ymm18, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x6c,0x30,0x14,0x11]
+          vunpcklps (%rcx){1to8}, %ymm18, %ymm26
+
+// CHECK: vunpcklps 4064(%rdx), %ymm18, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x6c,0x20,0x14,0x52,0x7f]
+          vunpcklps 4064(%rdx), %ymm18, %ymm26
+
+// CHECK: vunpcklps 4096(%rdx), %ymm18, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x6c,0x20,0x14,0x92,0x00,0x10,0x00,0x00]
+          vunpcklps 4096(%rdx), %ymm18, %ymm26
+
+// CHECK: vunpcklps -4096(%rdx), %ymm18, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x6c,0x20,0x14,0x52,0x80]
+          vunpcklps -4096(%rdx), %ymm18, %ymm26
+
+// CHECK: vunpcklps -4128(%rdx), %ymm18, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x6c,0x20,0x14,0x92,0xe0,0xef,0xff,0xff]
+          vunpcklps -4128(%rdx), %ymm18, %ymm26
+
+// CHECK: vunpcklps 508(%rdx){1to8}, %ymm18, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x6c,0x30,0x14,0x52,0x7f]
+          vunpcklps 508(%rdx){1to8}, %ymm18, %ymm26
+
+// CHECK: vunpcklps 512(%rdx){1to8}, %ymm18, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x6c,0x30,0x14,0x92,0x00,0x02,0x00,0x00]
+          vunpcklps 512(%rdx){1to8}, %ymm18, %ymm26
+
+// CHECK: vunpcklps -512(%rdx){1to8}, %ymm18, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x6c,0x30,0x14,0x52,0x80]
+          vunpcklps -512(%rdx){1to8}, %ymm18, %ymm26
+
+// CHECK: vunpcklps -516(%rdx){1to8}, %ymm18, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x6c,0x30,0x14,0x92,0xfc,0xfd,0xff,0xff]
+          vunpcklps -516(%rdx){1to8}, %ymm18, %ymm26
+
+// CHECK: vunpckhpd %xmm26, %xmm25, %xmm19
+// CHECK:  encoding: [0x62,0x81,0xb5,0x00,0x15,0xda]
+          vunpckhpd %xmm26, %xmm25, %xmm19
+
+// CHECK: vunpckhpd %xmm26, %xmm25, %xmm19 {%k3}
+// CHECK:  encoding: [0x62,0x81,0xb5,0x03,0x15,0xda]
+          vunpckhpd %xmm26, %xmm25, %xmm19 {%k3}
+
+// CHECK: vunpckhpd %xmm26, %xmm25, %xmm19 {%k3} {z}
+// CHECK:  encoding: [0x62,0x81,0xb5,0x83,0x15,0xda]
+          vunpckhpd %xmm26, %xmm25, %xmm19 {%k3} {z}
+
+// CHECK: vunpckhpd (%rcx), %xmm25, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xb5,0x00,0x15,0x19]
+          vunpckhpd (%rcx), %xmm25, %xmm19
+
+// CHECK: vunpckhpd 291(%rax,%r14,8), %xmm25, %xmm19
+// CHECK:  encoding: [0x62,0xa1,0xb5,0x00,0x15,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vunpckhpd 291(%rax,%r14,8), %xmm25, %xmm19
+
+// CHECK: vunpckhpd (%rcx){1to2}, %xmm25, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xb5,0x10,0x15,0x19]
+          vunpckhpd (%rcx){1to2}, %xmm25, %xmm19
+
+// CHECK: vunpckhpd 2032(%rdx), %xmm25, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xb5,0x00,0x15,0x5a,0x7f]
+          vunpckhpd 2032(%rdx), %xmm25, %xmm19
+
+// CHECK: vunpckhpd 2048(%rdx), %xmm25, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xb5,0x00,0x15,0x9a,0x00,0x08,0x00,0x00]
+          vunpckhpd 2048(%rdx), %xmm25, %xmm19
+
+// CHECK: vunpckhpd -2048(%rdx), %xmm25, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xb5,0x00,0x15,0x5a,0x80]
+          vunpckhpd -2048(%rdx), %xmm25, %xmm19
+
+// CHECK: vunpckhpd -2064(%rdx), %xmm25, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xb5,0x00,0x15,0x9a,0xf0,0xf7,0xff,0xff]
+          vunpckhpd -2064(%rdx), %xmm25, %xmm19
+
+// CHECK: vunpckhpd 1016(%rdx){1to2}, %xmm25, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xb5,0x10,0x15,0x5a,0x7f]
+          vunpckhpd 1016(%rdx){1to2}, %xmm25, %xmm19
+
+// CHECK: vunpckhpd 1024(%rdx){1to2}, %xmm25, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xb5,0x10,0x15,0x9a,0x00,0x04,0x00,0x00]
+          vunpckhpd 1024(%rdx){1to2}, %xmm25, %xmm19
+
+// CHECK: vunpckhpd -1024(%rdx){1to2}, %xmm25, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xb5,0x10,0x15,0x5a,0x80]
+          vunpckhpd -1024(%rdx){1to2}, %xmm25, %xmm19
+
+// CHECK: vunpckhpd -1032(%rdx){1to2}, %xmm25, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xb5,0x10,0x15,0x9a,0xf8,0xfb,0xff,0xff]
+          vunpckhpd -1032(%rdx){1to2}, %xmm25, %xmm19
+
+// CHECK: vunpckhpd %ymm24, %ymm22, %ymm25
+// CHECK:  encoding: [0x62,0x01,0xcd,0x20,0x15,0xc8]
+          vunpckhpd %ymm24, %ymm22, %ymm25
+
+// CHECK: vunpckhpd %ymm24, %ymm22, %ymm25 {%k7}
+// CHECK:  encoding: [0x62,0x01,0xcd,0x27,0x15,0xc8]
+          vunpckhpd %ymm24, %ymm22, %ymm25 {%k7}
+
+// CHECK: vunpckhpd %ymm24, %ymm22, %ymm25 {%k7} {z}
+// CHECK:  encoding: [0x62,0x01,0xcd,0xa7,0x15,0xc8]
+          vunpckhpd %ymm24, %ymm22, %ymm25 {%k7} {z}
+
+// CHECK: vunpckhpd (%rcx), %ymm22, %ymm25
+// CHECK:  encoding: [0x62,0x61,0xcd,0x20,0x15,0x09]
+          vunpckhpd (%rcx), %ymm22, %ymm25
+
+// CHECK: vunpckhpd 291(%rax,%r14,8), %ymm22, %ymm25
+// CHECK:  encoding: [0x62,0x21,0xcd,0x20,0x15,0x8c,0xf0,0x23,0x01,0x00,0x00]
+          vunpckhpd 291(%rax,%r14,8), %ymm22, %ymm25
+
+// CHECK: vunpckhpd (%rcx){1to4}, %ymm22, %ymm25
+// CHECK:  encoding: [0x62,0x61,0xcd,0x30,0x15,0x09]
+          vunpckhpd (%rcx){1to4}, %ymm22, %ymm25
+
+// CHECK: vunpckhpd 4064(%rdx), %ymm22, %ymm25
+// CHECK:  encoding: [0x62,0x61,0xcd,0x20,0x15,0x4a,0x7f]
+          vunpckhpd 4064(%rdx), %ymm22, %ymm25
+
+// CHECK: vunpckhpd 4096(%rdx), %ymm22, %ymm25
+// CHECK:  encoding: [0x62,0x61,0xcd,0x20,0x15,0x8a,0x00,0x10,0x00,0x00]
+          vunpckhpd 4096(%rdx), %ymm22, %ymm25
+
+// CHECK: vunpckhpd -4096(%rdx), %ymm22, %ymm25
+// CHECK:  encoding: [0x62,0x61,0xcd,0x20,0x15,0x4a,0x80]
+          vunpckhpd -4096(%rdx), %ymm22, %ymm25
+
+// CHECK: vunpckhpd -4128(%rdx), %ymm22, %ymm25
+// CHECK:  encoding: [0x62,0x61,0xcd,0x20,0x15,0x8a,0xe0,0xef,0xff,0xff]
+          vunpckhpd -4128(%rdx), %ymm22, %ymm25
+
+// CHECK: vunpckhpd 1016(%rdx){1to4}, %ymm22, %ymm25
+// CHECK:  encoding: [0x62,0x61,0xcd,0x30,0x15,0x4a,0x7f]
+          vunpckhpd 1016(%rdx){1to4}, %ymm22, %ymm25
+
+// CHECK: vunpckhpd 1024(%rdx){1to4}, %ymm22, %ymm25
+// CHECK:  encoding: [0x62,0x61,0xcd,0x30,0x15,0x8a,0x00,0x04,0x00,0x00]
+          vunpckhpd 1024(%rdx){1to4}, %ymm22, %ymm25
+
+// CHECK: vunpckhpd -1024(%rdx){1to4}, %ymm22, %ymm25
+// CHECK:  encoding: [0x62,0x61,0xcd,0x30,0x15,0x4a,0x80]
+          vunpckhpd -1024(%rdx){1to4}, %ymm22, %ymm25
+
+// CHECK: vunpckhpd -1032(%rdx){1to4}, %ymm22, %ymm25
+// CHECK:  encoding: [0x62,0x61,0xcd,0x30,0x15,0x8a,0xf8,0xfb,0xff,0xff]
+          vunpckhpd -1032(%rdx){1to4}, %ymm22, %ymm25
+
+// CHECK: vunpcklpd %xmm18, %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x21,0xad,0x00,0x14,0xf2]
+          vunpcklpd %xmm18, %xmm26, %xmm30
+
+// CHECK: vunpcklpd %xmm18, %xmm26, %xmm30 {%k5}
+// CHECK:  encoding: [0x62,0x21,0xad,0x05,0x14,0xf2]
+          vunpcklpd %xmm18, %xmm26, %xmm30 {%k5}
+
+// CHECK: vunpcklpd %xmm18, %xmm26, %xmm30 {%k5} {z}
+// CHECK:  encoding: [0x62,0x21,0xad,0x85,0x14,0xf2]
+          vunpcklpd %xmm18, %xmm26, %xmm30 {%k5} {z}
+
+// CHECK: vunpcklpd (%rcx), %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x61,0xad,0x00,0x14,0x31]
+          vunpcklpd (%rcx), %xmm26, %xmm30
+
+// CHECK: vunpcklpd 291(%rax,%r14,8), %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x21,0xad,0x00,0x14,0xb4,0xf0,0x23,0x01,0x00,0x00]
+          vunpcklpd 291(%rax,%r14,8), %xmm26, %xmm30
+
+// CHECK: vunpcklpd (%rcx){1to2}, %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x61,0xad,0x10,0x14,0x31]
+          vunpcklpd (%rcx){1to2}, %xmm26, %xmm30
+
+// CHECK: vunpcklpd 2032(%rdx), %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x61,0xad,0x00,0x14,0x72,0x7f]
+          vunpcklpd 2032(%rdx), %xmm26, %xmm30
+
+// CHECK: vunpcklpd 2048(%rdx), %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x61,0xad,0x00,0x14,0xb2,0x00,0x08,0x00,0x00]
+          vunpcklpd 2048(%rdx), %xmm26, %xmm30
+
+// CHECK: vunpcklpd -2048(%rdx), %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x61,0xad,0x00,0x14,0x72,0x80]
+          vunpcklpd -2048(%rdx), %xmm26, %xmm30
+
+// CHECK: vunpcklpd -2064(%rdx), %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x61,0xad,0x00,0x14,0xb2,0xf0,0xf7,0xff,0xff]
+          vunpcklpd -2064(%rdx), %xmm26, %xmm30
+
+// CHECK: vunpcklpd 1016(%rdx){1to2}, %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x61,0xad,0x10,0x14,0x72,0x7f]
+          vunpcklpd 1016(%rdx){1to2}, %xmm26, %xmm30
+
+// CHECK: vunpcklpd 1024(%rdx){1to2}, %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x61,0xad,0x10,0x14,0xb2,0x00,0x04,0x00,0x00]
+          vunpcklpd 1024(%rdx){1to2}, %xmm26, %xmm30
+
+// CHECK: vunpcklpd -1024(%rdx){1to2}, %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x61,0xad,0x10,0x14,0x72,0x80]
+          vunpcklpd -1024(%rdx){1to2}, %xmm26, %xmm30
+
+// CHECK: vunpcklpd -1032(%rdx){1to2}, %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x61,0xad,0x10,0x14,0xb2,0xf8,0xfb,0xff,0xff]
+          vunpcklpd -1032(%rdx){1to2}, %xmm26, %xmm30
+
+// CHECK: vunpcklpd %ymm17, %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xa1,0xad,0x20,0x14,0xe1]
+          vunpcklpd %ymm17, %ymm26, %ymm20
+
+// CHECK: vunpcklpd %ymm17, %ymm26, %ymm20 {%k2}
+// CHECK:  encoding: [0x62,0xa1,0xad,0x22,0x14,0xe1]
+          vunpcklpd %ymm17, %ymm26, %ymm20 {%k2}
+
+// CHECK: vunpcklpd %ymm17, %ymm26, %ymm20 {%k2} {z}
+// CHECK:  encoding: [0x62,0xa1,0xad,0xa2,0x14,0xe1]
+          vunpcklpd %ymm17, %ymm26, %ymm20 {%k2} {z}
+
+// CHECK: vunpcklpd (%rcx), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xad,0x20,0x14,0x21]
+          vunpcklpd (%rcx), %ymm26, %ymm20
+
+// CHECK: vunpcklpd 291(%rax,%r14,8), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xa1,0xad,0x20,0x14,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vunpcklpd 291(%rax,%r14,8), %ymm26, %ymm20
+
+// CHECK: vunpcklpd (%rcx){1to4}, %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xad,0x30,0x14,0x21]
+          vunpcklpd (%rcx){1to4}, %ymm26, %ymm20
+
+// CHECK: vunpcklpd 4064(%rdx), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xad,0x20,0x14,0x62,0x7f]
+          vunpcklpd 4064(%rdx), %ymm26, %ymm20
+
+// CHECK: vunpcklpd 4096(%rdx), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xad,0x20,0x14,0xa2,0x00,0x10,0x00,0x00]
+          vunpcklpd 4096(%rdx), %ymm26, %ymm20
+
+// CHECK: vunpcklpd -4096(%rdx), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xad,0x20,0x14,0x62,0x80]
+          vunpcklpd -4096(%rdx), %ymm26, %ymm20
+
+// CHECK: vunpcklpd -4128(%rdx), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xad,0x20,0x14,0xa2,0xe0,0xef,0xff,0xff]
+          vunpcklpd -4128(%rdx), %ymm26, %ymm20
+
+// CHECK: vunpcklpd 1016(%rdx){1to4}, %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xad,0x30,0x14,0x62,0x7f]
+          vunpcklpd 1016(%rdx){1to4}, %ymm26, %ymm20
+
+// CHECK: vunpcklpd 1024(%rdx){1to4}, %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xad,0x30,0x14,0xa2,0x00,0x04,0x00,0x00]
+          vunpcklpd 1024(%rdx){1to4}, %ymm26, %ymm20
+
+// CHECK: vunpcklpd -1024(%rdx){1to4}, %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xad,0x30,0x14,0x62,0x80]
+          vunpcklpd -1024(%rdx){1to4}, %ymm26, %ymm20
+
+// CHECK: vunpcklpd -1032(%rdx){1to4}, %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xad,0x30,0x14,0xa2,0xf8,0xfb,0xff,0xff]
+          vunpcklpd -1032(%rdx){1to4}, %ymm26, %ymm20
+
+// CHECK: vpunpckldq %xmm17, %xmm21, %xmm25
+// CHECK:  encoding: [0x62,0x21,0x55,0x00,0x62,0xc9]
+          vpunpckldq %xmm17, %xmm21, %xmm25
+
+// CHECK: vpunpckldq %xmm17, %xmm21, %xmm25 {%k7}
+// CHECK:  encoding: [0x62,0x21,0x55,0x07,0x62,0xc9]
+          vpunpckldq %xmm17, %xmm21, %xmm25 {%k7}
+
+// CHECK: vpunpckldq %xmm17, %xmm21, %xmm25 {%k7} {z}
+// CHECK:  encoding: [0x62,0x21,0x55,0x87,0x62,0xc9]
+          vpunpckldq %xmm17, %xmm21, %xmm25 {%k7} {z}
+
+// CHECK: vpunpckldq (%rcx), %xmm21, %xmm25
+// CHECK:  encoding: [0x62,0x61,0x55,0x00,0x62,0x09]
+          vpunpckldq (%rcx), %xmm21, %xmm25
+
+// CHECK: vpunpckldq 291(%rax,%r14,8), %xmm21, %xmm25
+// CHECK:  encoding: [0x62,0x21,0x55,0x00,0x62,0x8c,0xf0,0x23,0x01,0x00,0x00]
+          vpunpckldq 291(%rax,%r14,8), %xmm21, %xmm25
+
+// CHECK: vpunpckldq (%rcx){1to4}, %xmm21, %xmm25
+// CHECK:  encoding: [0x62,0x61,0x55,0x10,0x62,0x09]
+          vpunpckldq (%rcx){1to4}, %xmm21, %xmm25
+
+// CHECK: vpunpckldq 2032(%rdx), %xmm21, %xmm25
+// CHECK:  encoding: [0x62,0x61,0x55,0x00,0x62,0x4a,0x7f]
+          vpunpckldq 2032(%rdx), %xmm21, %xmm25
+
+// CHECK: vpunpckldq 2048(%rdx), %xmm21, %xmm25
+// CHECK:  encoding: [0x62,0x61,0x55,0x00,0x62,0x8a,0x00,0x08,0x00,0x00]
+          vpunpckldq 2048(%rdx), %xmm21, %xmm25
+
+// CHECK: vpunpckldq -2048(%rdx), %xmm21, %xmm25
+// CHECK:  encoding: [0x62,0x61,0x55,0x00,0x62,0x4a,0x80]
+          vpunpckldq -2048(%rdx), %xmm21, %xmm25
+
+// CHECK: vpunpckldq -2064(%rdx), %xmm21, %xmm25
+// CHECK:  encoding: [0x62,0x61,0x55,0x00,0x62,0x8a,0xf0,0xf7,0xff,0xff]
+          vpunpckldq -2064(%rdx), %xmm21, %xmm25
+
+// CHECK: vpunpckldq 508(%rdx){1to4}, %xmm21, %xmm25
+// CHECK:  encoding: [0x62,0x61,0x55,0x10,0x62,0x4a,0x7f]
+          vpunpckldq 508(%rdx){1to4}, %xmm21, %xmm25
+
+// CHECK: vpunpckldq 512(%rdx){1to4}, %xmm21, %xmm25
+// CHECK:  encoding: [0x62,0x61,0x55,0x10,0x62,0x8a,0x00,0x02,0x00,0x00]
+          vpunpckldq 512(%rdx){1to4}, %xmm21, %xmm25
+
+// CHECK: vpunpckldq -512(%rdx){1to4}, %xmm21, %xmm25
+// CHECK:  encoding: [0x62,0x61,0x55,0x10,0x62,0x4a,0x80]
+          vpunpckldq -512(%rdx){1to4}, %xmm21, %xmm25
+
+// CHECK: vpunpckldq -516(%rdx){1to4}, %xmm21, %xmm25
+// CHECK:  encoding: [0x62,0x61,0x55,0x10,0x62,0x8a,0xfc,0xfd,0xff,0xff]
+          vpunpckldq -516(%rdx){1to4}, %xmm21, %xmm25
+
+// CHECK: vpunpckldq %ymm26, %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0x81,0x65,0x20,0x62,0xe2]
+          vpunpckldq %ymm26, %ymm19, %ymm20
+
+// CHECK: vpunpckldq %ymm26, %ymm19, %ymm20 {%k7}
+// CHECK:  encoding: [0x62,0x81,0x65,0x27,0x62,0xe2]
+          vpunpckldq %ymm26, %ymm19, %ymm20 {%k7}
+
+// CHECK: vpunpckldq %ymm26, %ymm19, %ymm20 {%k7} {z}
+// CHECK:  encoding: [0x62,0x81,0x65,0xa7,0x62,0xe2]
+          vpunpckldq %ymm26, %ymm19, %ymm20 {%k7} {z}
+
+// CHECK: vpunpckldq (%rcx), %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x65,0x20,0x62,0x21]
+          vpunpckldq (%rcx), %ymm19, %ymm20
+
+// CHECK: vpunpckldq 291(%rax,%r14,8), %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xa1,0x65,0x20,0x62,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpunpckldq 291(%rax,%r14,8), %ymm19, %ymm20
+
+// CHECK: vpunpckldq (%rcx){1to8}, %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x65,0x30,0x62,0x21]
+          vpunpckldq (%rcx){1to8}, %ymm19, %ymm20
+
+// CHECK: vpunpckldq 4064(%rdx), %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x65,0x20,0x62,0x62,0x7f]
+          vpunpckldq 4064(%rdx), %ymm19, %ymm20
+
+// CHECK: vpunpckldq 4096(%rdx), %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x65,0x20,0x62,0xa2,0x00,0x10,0x00,0x00]
+          vpunpckldq 4096(%rdx), %ymm19, %ymm20
+
+// CHECK: vpunpckldq -4096(%rdx), %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x65,0x20,0x62,0x62,0x80]
+          vpunpckldq -4096(%rdx), %ymm19, %ymm20
+
+// CHECK: vpunpckldq -4128(%rdx), %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x65,0x20,0x62,0xa2,0xe0,0xef,0xff,0xff]
+          vpunpckldq -4128(%rdx), %ymm19, %ymm20
+
+// CHECK: vpunpckldq 508(%rdx){1to8}, %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x65,0x30,0x62,0x62,0x7f]
+          vpunpckldq 508(%rdx){1to8}, %ymm19, %ymm20
+
+// CHECK: vpunpckldq 512(%rdx){1to8}, %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x65,0x30,0x62,0xa2,0x00,0x02,0x00,0x00]
+          vpunpckldq 512(%rdx){1to8}, %ymm19, %ymm20
+
+// CHECK: vpunpckldq -512(%rdx){1to8}, %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x65,0x30,0x62,0x62,0x80]
+          vpunpckldq -512(%rdx){1to8}, %ymm19, %ymm20
+
+// CHECK: vpunpckldq -516(%rdx){1to8}, %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x65,0x30,0x62,0xa2,0xfc,0xfd,0xff,0xff]
+          vpunpckldq -516(%rdx){1to8}, %ymm19, %ymm20
+
+// CHECK: vpunpckhdq %xmm27, %xmm24, %xmm28
+// CHECK:  encoding: [0x62,0x01,0x3d,0x00,0x6a,0xe3]
+          vpunpckhdq %xmm27, %xmm24, %xmm28
+
+// CHECK: vpunpckhdq %xmm27, %xmm24, %xmm28 {%k7}
+// CHECK:  encoding: [0x62,0x01,0x3d,0x07,0x6a,0xe3]
+          vpunpckhdq %xmm27, %xmm24, %xmm28 {%k7}
+
+// CHECK: vpunpckhdq %xmm27, %xmm24, %xmm28 {%k7} {z}
+// CHECK:  encoding: [0x62,0x01,0x3d,0x87,0x6a,0xe3]
+          vpunpckhdq %xmm27, %xmm24, %xmm28 {%k7} {z}
+
+// CHECK: vpunpckhdq (%rcx), %xmm24, %xmm28
+// CHECK:  encoding: [0x62,0x61,0x3d,0x00,0x6a,0x21]
+          vpunpckhdq (%rcx), %xmm24, %xmm28
+
+// CHECK: vpunpckhdq 291(%rax,%r14,8), %xmm24, %xmm28
+// CHECK:  encoding: [0x62,0x21,0x3d,0x00,0x6a,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpunpckhdq 291(%rax,%r14,8), %xmm24, %xmm28
+
+// CHECK: vpunpckhdq (%rcx){1to4}, %xmm24, %xmm28
+// CHECK:  encoding: [0x62,0x61,0x3d,0x10,0x6a,0x21]
+          vpunpckhdq (%rcx){1to4}, %xmm24, %xmm28
+
+// CHECK: vpunpckhdq 2032(%rdx), %xmm24, %xmm28
+// CHECK:  encoding: [0x62,0x61,0x3d,0x00,0x6a,0x62,0x7f]
+          vpunpckhdq 2032(%rdx), %xmm24, %xmm28
+
+// CHECK: vpunpckhdq 2048(%rdx), %xmm24, %xmm28
+// CHECK:  encoding: [0x62,0x61,0x3d,0x00,0x6a,0xa2,0x00,0x08,0x00,0x00]
+          vpunpckhdq 2048(%rdx), %xmm24, %xmm28
+
+// CHECK: vpunpckhdq -2048(%rdx), %xmm24, %xmm28
+// CHECK:  encoding: [0x62,0x61,0x3d,0x00,0x6a,0x62,0x80]
+          vpunpckhdq -2048(%rdx), %xmm24, %xmm28
+
+// CHECK: vpunpckhdq -2064(%rdx), %xmm24, %xmm28
+// CHECK:  encoding: [0x62,0x61,0x3d,0x00,0x6a,0xa2,0xf0,0xf7,0xff,0xff]
+          vpunpckhdq -2064(%rdx), %xmm24, %xmm28
+
+// CHECK: vpunpckhdq 508(%rdx){1to4}, %xmm24, %xmm28
+// CHECK:  encoding: [0x62,0x61,0x3d,0x10,0x6a,0x62,0x7f]
+          vpunpckhdq 508(%rdx){1to4}, %xmm24, %xmm28
+
+// CHECK: vpunpckhdq 512(%rdx){1to4}, %xmm24, %xmm28
+// CHECK:  encoding: [0x62,0x61,0x3d,0x10,0x6a,0xa2,0x00,0x02,0x00,0x00]
+          vpunpckhdq 512(%rdx){1to4}, %xmm24, %xmm28
+
+// CHECK: vpunpckhdq -512(%rdx){1to4}, %xmm24, %xmm28
+// CHECK:  encoding: [0x62,0x61,0x3d,0x10,0x6a,0x62,0x80]
+          vpunpckhdq -512(%rdx){1to4}, %xmm24, %xmm28
+
+// CHECK: vpunpckhdq -516(%rdx){1to4}, %xmm24, %xmm28
+// CHECK:  encoding: [0x62,0x61,0x3d,0x10,0x6a,0xa2,0xfc,0xfd,0xff,0xff]
+          vpunpckhdq -516(%rdx){1to4}, %xmm24, %xmm28
+
+// CHECK: vpunpckhdq %ymm28, %ymm24, %ymm26
+// CHECK:  encoding: [0x62,0x01,0x3d,0x20,0x6a,0xd4]
+          vpunpckhdq %ymm28, %ymm24, %ymm26
+
+// CHECK: vpunpckhdq %ymm28, %ymm24, %ymm26 {%k2}
+// CHECK:  encoding: [0x62,0x01,0x3d,0x22,0x6a,0xd4]
+          vpunpckhdq %ymm28, %ymm24, %ymm26 {%k2}
+
+// CHECK: vpunpckhdq %ymm28, %ymm24, %ymm26 {%k2} {z}
+// CHECK:  encoding: [0x62,0x01,0x3d,0xa2,0x6a,0xd4]
+          vpunpckhdq %ymm28, %ymm24, %ymm26 {%k2} {z}
+
+// CHECK: vpunpckhdq (%rcx), %ymm24, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x3d,0x20,0x6a,0x11]
+          vpunpckhdq (%rcx), %ymm24, %ymm26
+
+// CHECK: vpunpckhdq 291(%rax,%r14,8), %ymm24, %ymm26
+// CHECK:  encoding: [0x62,0x21,0x3d,0x20,0x6a,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpunpckhdq 291(%rax,%r14,8), %ymm24, %ymm26
+
+// CHECK: vpunpckhdq (%rcx){1to8}, %ymm24, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x3d,0x30,0x6a,0x11]
+          vpunpckhdq (%rcx){1to8}, %ymm24, %ymm26
+
+// CHECK: vpunpckhdq 4064(%rdx), %ymm24, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x3d,0x20,0x6a,0x52,0x7f]
+          vpunpckhdq 4064(%rdx), %ymm24, %ymm26
+
+// CHECK: vpunpckhdq 4096(%rdx), %ymm24, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x3d,0x20,0x6a,0x92,0x00,0x10,0x00,0x00]
+          vpunpckhdq 4096(%rdx), %ymm24, %ymm26
+
+// CHECK: vpunpckhdq -4096(%rdx), %ymm24, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x3d,0x20,0x6a,0x52,0x80]
+          vpunpckhdq -4096(%rdx), %ymm24, %ymm26
+
+// CHECK: vpunpckhdq -4128(%rdx), %ymm24, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x3d,0x20,0x6a,0x92,0xe0,0xef,0xff,0xff]
+          vpunpckhdq -4128(%rdx), %ymm24, %ymm26
+
+// CHECK: vpunpckhdq 508(%rdx){1to8}, %ymm24, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x3d,0x30,0x6a,0x52,0x7f]
+          vpunpckhdq 508(%rdx){1to8}, %ymm24, %ymm26
+
+// CHECK: vpunpckhdq 512(%rdx){1to8}, %ymm24, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x3d,0x30,0x6a,0x92,0x00,0x02,0x00,0x00]
+          vpunpckhdq 512(%rdx){1to8}, %ymm24, %ymm26
+
+// CHECK: vpunpckhdq -512(%rdx){1to8}, %ymm24, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x3d,0x30,0x6a,0x52,0x80]
+          vpunpckhdq -512(%rdx){1to8}, %ymm24, %ymm26
+
+// CHECK: vpunpckhdq -516(%rdx){1to8}, %ymm24, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x3d,0x30,0x6a,0x92,0xfc,0xfd,0xff,0xff]
+          vpunpckhdq -516(%rdx){1to8}, %ymm24, %ymm26
+
+// CHECK: vpunpcklqdq %xmm23, %xmm27, %xmm17
+// CHECK:  encoding: [0x62,0xa1,0xa5,0x00,0x6c,0xcf]
+          vpunpcklqdq %xmm23, %xmm27, %xmm17
+
+// CHECK: vpunpcklqdq %xmm23, %xmm27, %xmm17 {%k6}
+// CHECK:  encoding: [0x62,0xa1,0xa5,0x06,0x6c,0xcf]
+          vpunpcklqdq %xmm23, %xmm27, %xmm17 {%k6}
+
+// CHECK: vpunpcklqdq %xmm23, %xmm27, %xmm17 {%k6} {z}
+// CHECK:  encoding: [0x62,0xa1,0xa5,0x86,0x6c,0xcf]
+          vpunpcklqdq %xmm23, %xmm27, %xmm17 {%k6} {z}
+
+// CHECK: vpunpcklqdq (%rcx), %xmm27, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0xa5,0x00,0x6c,0x09]
+          vpunpcklqdq (%rcx), %xmm27, %xmm17
+
+// CHECK: vpunpcklqdq 291(%rax,%r14,8), %xmm27, %xmm17
+// CHECK:  encoding: [0x62,0xa1,0xa5,0x00,0x6c,0x8c,0xf0,0x23,0x01,0x00,0x00]
+          vpunpcklqdq 291(%rax,%r14,8), %xmm27, %xmm17
+
+// CHECK: vpunpcklqdq (%rcx){1to2}, %xmm27, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0xa5,0x10,0x6c,0x09]
+          vpunpcklqdq (%rcx){1to2}, %xmm27, %xmm17
+
+// CHECK: vpunpcklqdq 2032(%rdx), %xmm27, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0xa5,0x00,0x6c,0x4a,0x7f]
+          vpunpcklqdq 2032(%rdx), %xmm27, %xmm17
+
+// CHECK: vpunpcklqdq 2048(%rdx), %xmm27, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0xa5,0x00,0x6c,0x8a,0x00,0x08,0x00,0x00]
+          vpunpcklqdq 2048(%rdx), %xmm27, %xmm17
+
+// CHECK: vpunpcklqdq -2048(%rdx), %xmm27, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0xa5,0x00,0x6c,0x4a,0x80]
+          vpunpcklqdq -2048(%rdx), %xmm27, %xmm17
+
+// CHECK: vpunpcklqdq -2064(%rdx), %xmm27, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0xa5,0x00,0x6c,0x8a,0xf0,0xf7,0xff,0xff]
+          vpunpcklqdq -2064(%rdx), %xmm27, %xmm17
+
+// CHECK: vpunpcklqdq 1016(%rdx){1to2}, %xmm27, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0xa5,0x10,0x6c,0x4a,0x7f]
+          vpunpcklqdq 1016(%rdx){1to2}, %xmm27, %xmm17
+
+// CHECK: vpunpcklqdq 1024(%rdx){1to2}, %xmm27, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0xa5,0x10,0x6c,0x8a,0x00,0x04,0x00,0x00]
+          vpunpcklqdq 1024(%rdx){1to2}, %xmm27, %xmm17
+
+// CHECK: vpunpcklqdq -1024(%rdx){1to2}, %xmm27, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0xa5,0x10,0x6c,0x4a,0x80]
+          vpunpcklqdq -1024(%rdx){1to2}, %xmm27, %xmm17
+
+// CHECK: vpunpcklqdq -1032(%rdx){1to2}, %xmm27, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0xa5,0x10,0x6c,0x8a,0xf8,0xfb,0xff,0xff]
+          vpunpcklqdq -1032(%rdx){1to2}, %xmm27, %xmm17
+
+// CHECK: vpunpcklqdq %ymm28, %ymm20, %ymm29
+// CHECK:  encoding: [0x62,0x01,0xdd,0x20,0x6c,0xec]
+          vpunpcklqdq %ymm28, %ymm20, %ymm29
+
+// CHECK: vpunpcklqdq %ymm28, %ymm20, %ymm29 {%k7}
+// CHECK:  encoding: [0x62,0x01,0xdd,0x27,0x6c,0xec]
+          vpunpcklqdq %ymm28, %ymm20, %ymm29 {%k7}
+
+// CHECK: vpunpcklqdq %ymm28, %ymm20, %ymm29 {%k7} {z}
+// CHECK:  encoding: [0x62,0x01,0xdd,0xa7,0x6c,0xec]
+          vpunpcklqdq %ymm28, %ymm20, %ymm29 {%k7} {z}
+
+// CHECK: vpunpcklqdq (%rcx), %ymm20, %ymm29
+// CHECK:  encoding: [0x62,0x61,0xdd,0x20,0x6c,0x29]
+          vpunpcklqdq (%rcx), %ymm20, %ymm29
+
+// CHECK: vpunpcklqdq 291(%rax,%r14,8), %ymm20, %ymm29
+// CHECK:  encoding: [0x62,0x21,0xdd,0x20,0x6c,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vpunpcklqdq 291(%rax,%r14,8), %ymm20, %ymm29
+
+// CHECK: vpunpcklqdq (%rcx){1to4}, %ymm20, %ymm29
+// CHECK:  encoding: [0x62,0x61,0xdd,0x30,0x6c,0x29]
+          vpunpcklqdq (%rcx){1to4}, %ymm20, %ymm29
+
+// CHECK: vpunpcklqdq 4064(%rdx), %ymm20, %ymm29
+// CHECK:  encoding: [0x62,0x61,0xdd,0x20,0x6c,0x6a,0x7f]
+          vpunpcklqdq 4064(%rdx), %ymm20, %ymm29
+
+// CHECK: vpunpcklqdq 4096(%rdx), %ymm20, %ymm29
+// CHECK:  encoding: [0x62,0x61,0xdd,0x20,0x6c,0xaa,0x00,0x10,0x00,0x00]
+          vpunpcklqdq 4096(%rdx), %ymm20, %ymm29
+
+// CHECK: vpunpcklqdq -4096(%rdx), %ymm20, %ymm29
+// CHECK:  encoding: [0x62,0x61,0xdd,0x20,0x6c,0x6a,0x80]
+          vpunpcklqdq -4096(%rdx), %ymm20, %ymm29
+
+// CHECK: vpunpcklqdq -4128(%rdx), %ymm20, %ymm29
+// CHECK:  encoding: [0x62,0x61,0xdd,0x20,0x6c,0xaa,0xe0,0xef,0xff,0xff]
+          vpunpcklqdq -4128(%rdx), %ymm20, %ymm29
+
+// CHECK: vpunpcklqdq 1016(%rdx){1to4}, %ymm20, %ymm29
+// CHECK:  encoding: [0x62,0x61,0xdd,0x30,0x6c,0x6a,0x7f]
+          vpunpcklqdq 1016(%rdx){1to4}, %ymm20, %ymm29
+
+// CHECK: vpunpcklqdq 1024(%rdx){1to4}, %ymm20, %ymm29
+// CHECK:  encoding: [0x62,0x61,0xdd,0x30,0x6c,0xaa,0x00,0x04,0x00,0x00]
+          vpunpcklqdq 1024(%rdx){1to4}, %ymm20, %ymm29
+
+// CHECK: vpunpcklqdq -1024(%rdx){1to4}, %ymm20, %ymm29
+// CHECK:  encoding: [0x62,0x61,0xdd,0x30,0x6c,0x6a,0x80]
+          vpunpcklqdq -1024(%rdx){1to4}, %ymm20, %ymm29
+
+// CHECK: vpunpcklqdq -1032(%rdx){1to4}, %ymm20, %ymm29
+// CHECK:  encoding: [0x62,0x61,0xdd,0x30,0x6c,0xaa,0xf8,0xfb,0xff,0xff]
+          vpunpcklqdq -1032(%rdx){1to4}, %ymm20, %ymm29
+
+// CHECK: vpunpckhqdq %xmm24, %xmm19, %xmm19
+// CHECK:  encoding: [0x62,0x81,0xe5,0x00,0x6d,0xd8]
+          vpunpckhqdq %xmm24, %xmm19, %xmm19
+
+// CHECK: vpunpckhqdq %xmm24, %xmm19, %xmm19 {%k6}
+// CHECK:  encoding: [0x62,0x81,0xe5,0x06,0x6d,0xd8]
+          vpunpckhqdq %xmm24, %xmm19, %xmm19 {%k6}
+
+// CHECK: vpunpckhqdq %xmm24, %xmm19, %xmm19 {%k6} {z}
+// CHECK:  encoding: [0x62,0x81,0xe5,0x86,0x6d,0xd8]
+          vpunpckhqdq %xmm24, %xmm19, %xmm19 {%k6} {z}
+
+// CHECK: vpunpckhqdq (%rcx), %xmm19, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x00,0x6d,0x19]
+          vpunpckhqdq (%rcx), %xmm19, %xmm19
+
+// CHECK: vpunpckhqdq 291(%rax,%r14,8), %xmm19, %xmm19
+// CHECK:  encoding: [0x62,0xa1,0xe5,0x00,0x6d,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vpunpckhqdq 291(%rax,%r14,8), %xmm19, %xmm19
+
+// CHECK: vpunpckhqdq (%rcx){1to2}, %xmm19, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x10,0x6d,0x19]
+          vpunpckhqdq (%rcx){1to2}, %xmm19, %xmm19
+
+// CHECK: vpunpckhqdq 2032(%rdx), %xmm19, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x00,0x6d,0x5a,0x7f]
+          vpunpckhqdq 2032(%rdx), %xmm19, %xmm19
+
+// CHECK: vpunpckhqdq 2048(%rdx), %xmm19, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x00,0x6d,0x9a,0x00,0x08,0x00,0x00]
+          vpunpckhqdq 2048(%rdx), %xmm19, %xmm19
+
+// CHECK: vpunpckhqdq -2048(%rdx), %xmm19, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x00,0x6d,0x5a,0x80]
+          vpunpckhqdq -2048(%rdx), %xmm19, %xmm19
+
+// CHECK: vpunpckhqdq -2064(%rdx), %xmm19, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x00,0x6d,0x9a,0xf0,0xf7,0xff,0xff]
+          vpunpckhqdq -2064(%rdx), %xmm19, %xmm19
+
+// CHECK: vpunpckhqdq 1016(%rdx){1to2}, %xmm19, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x10,0x6d,0x5a,0x7f]
+          vpunpckhqdq 1016(%rdx){1to2}, %xmm19, %xmm19
+
+// CHECK: vpunpckhqdq 1024(%rdx){1to2}, %xmm19, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x10,0x6d,0x9a,0x00,0x04,0x00,0x00]
+          vpunpckhqdq 1024(%rdx){1to2}, %xmm19, %xmm19
+
+// CHECK: vpunpckhqdq -1024(%rdx){1to2}, %xmm19, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x10,0x6d,0x5a,0x80]
+          vpunpckhqdq -1024(%rdx){1to2}, %xmm19, %xmm19
+
+// CHECK: vpunpckhqdq -1032(%rdx){1to2}, %xmm19, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x10,0x6d,0x9a,0xf8,0xfb,0xff,0xff]
+          vpunpckhqdq -1032(%rdx){1to2}, %xmm19, %xmm19
+
+// CHECK: vpunpckhqdq %ymm28, %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0x81,0xe5,0x20,0x6d,0xe4]
+          vpunpckhqdq %ymm28, %ymm19, %ymm20
+
+// CHECK: vpunpckhqdq %ymm28, %ymm19, %ymm20 {%k6}
+// CHECK:  encoding: [0x62,0x81,0xe5,0x26,0x6d,0xe4]
+          vpunpckhqdq %ymm28, %ymm19, %ymm20 {%k6}
+
+// CHECK: vpunpckhqdq %ymm28, %ymm19, %ymm20 {%k6} {z}
+// CHECK:  encoding: [0x62,0x81,0xe5,0xa6,0x6d,0xe4]
+          vpunpckhqdq %ymm28, %ymm19, %ymm20 {%k6} {z}
+
+// CHECK: vpunpckhqdq (%rcx), %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x20,0x6d,0x21]
+          vpunpckhqdq (%rcx), %ymm19, %ymm20
+
+// CHECK: vpunpckhqdq 291(%rax,%r14,8), %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xa1,0xe5,0x20,0x6d,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpunpckhqdq 291(%rax,%r14,8), %ymm19, %ymm20
+
+// CHECK: vpunpckhqdq (%rcx){1to4}, %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x30,0x6d,0x21]
+          vpunpckhqdq (%rcx){1to4}, %ymm19, %ymm20
+
+// CHECK: vpunpckhqdq 4064(%rdx), %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x20,0x6d,0x62,0x7f]
+          vpunpckhqdq 4064(%rdx), %ymm19, %ymm20
+
+// CHECK: vpunpckhqdq 4096(%rdx), %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x20,0x6d,0xa2,0x00,0x10,0x00,0x00]
+          vpunpckhqdq 4096(%rdx), %ymm19, %ymm20
+
+// CHECK: vpunpckhqdq -4096(%rdx), %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x20,0x6d,0x62,0x80]
+          vpunpckhqdq -4096(%rdx), %ymm19, %ymm20
+
+// CHECK: vpunpckhqdq -4128(%rdx), %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x20,0x6d,0xa2,0xe0,0xef,0xff,0xff]
+          vpunpckhqdq -4128(%rdx), %ymm19, %ymm20
+
+// CHECK: vpunpckhqdq 1016(%rdx){1to4}, %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x30,0x6d,0x62,0x7f]
+          vpunpckhqdq 1016(%rdx){1to4}, %ymm19, %ymm20
+
+// CHECK: vpunpckhqdq 1024(%rdx){1to4}, %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x30,0x6d,0xa2,0x00,0x04,0x00,0x00]
+          vpunpckhqdq 1024(%rdx){1to4}, %ymm19, %ymm20
+
+// CHECK: vpunpckhqdq -1024(%rdx){1to4}, %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x30,0x6d,0x62,0x80]
+          vpunpckhqdq -1024(%rdx){1to4}, %ymm19, %ymm20
+
+// CHECK: vpunpckhqdq -1032(%rdx){1to4}, %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x30,0x6d,0xa2,0xf8,0xfb,0xff,0xff]
+          vpunpckhqdq -1032(%rdx){1to4}, %ymm19, %ymm20
+