AVX-512: Added VPACK* instructions forms for KNL and SKX

and their intrinsics
by Asaf Badouh (asaf.badouh@intel.com)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@236414 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/llvm/IR/IntrinsicsX86.td b/include/llvm/IR/IntrinsicsX86.td
index 86cf68ae7c976754a40de0b0a53b9e3a8606a452..c9d73d74e8401e5714e900fb4a4be00e6879fadf 100644 (file)
--- a/include/llvm/IR/IntrinsicsX86.td
+++ b/include/llvm/IR/IntrinsicsX86.td
@@ -2991,6 +2991,46 @@ let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
                          llvm_i64_ty], [IntrNoMem]>;
 }
 
+// Pack ops.
+let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
+  def int_x86_avx512_mask_packsswb_128 : GCCBuiltin<"__builtin_ia32_packsswb128_mask">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v8i16_ty, llvm_v8i16_ty,
+                         llvm_v16i8_ty, llvm_i16_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_packsswb_256 : GCCBuiltin<"__builtin_ia32_packsswb256_mask">,
+              Intrinsic<[llvm_v32i8_ty], [llvm_v16i16_ty,llvm_v16i16_ty,
+                         llvm_v32i8_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_packsswb_512 : GCCBuiltin<"__builtin_ia32_packsswb512_mask">,
+              Intrinsic<[llvm_v64i8_ty], [llvm_v32i16_ty,llvm_v32i16_ty, 
+                         llvm_v64i8_ty, llvm_i64_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_packssdw_128 : GCCBuiltin<"__builtin_ia32_packssdw128_mask">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v4i32_ty, llvm_v4i32_ty,
+                         llvm_v8i16_ty, llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_packssdw_256 : GCCBuiltin<"__builtin_ia32_packssdw256_mask">,
+              Intrinsic<[llvm_v16i16_ty], [llvm_v8i32_ty, llvm_v8i32_ty,
+                         llvm_v16i16_ty, llvm_i16_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_packssdw_512 : GCCBuiltin<"__builtin_ia32_packssdw512_mask">,
+              Intrinsic<[llvm_v32i16_ty], [llvm_v16i32_ty, llvm_v16i32_ty,
+                         llvm_v32i16_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_packuswb_128 : GCCBuiltin<"__builtin_ia32_packuswb128_mask">,
+              Intrinsic<[llvm_v16i8_ty], [llvm_v8i16_ty, llvm_v8i16_ty,
+                         llvm_v16i8_ty, llvm_i16_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_packuswb_256 : GCCBuiltin<"__builtin_ia32_packuswb256_mask">,
+              Intrinsic<[llvm_v32i8_ty], [llvm_v16i16_ty,llvm_v16i16_ty,
+                         llvm_v32i8_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_packuswb_512 : GCCBuiltin<"__builtin_ia32_packuswb512_mask">,
+              Intrinsic<[llvm_v64i8_ty], [llvm_v32i16_ty,llvm_v32i16_ty, 
+                         llvm_v64i8_ty, llvm_i64_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_packusdw_128 : GCCBuiltin<"__builtin_ia32_packusdw128_mask">,
+              Intrinsic<[llvm_v8i16_ty], [llvm_v4i32_ty, llvm_v4i32_ty,
+                         llvm_v8i16_ty, llvm_i8_ty], [IntrNoMem]>;
+  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]>;
+}
+
 // Vector convert
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_avx512_mask_cvttps2dq_512: GCCBuiltin<"__builtin_ia32_cvttps2dq512_mask">,

diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
index 85bdfa7d7afdc75a3bb5a93e163cf9d3812afc6b..8a9bd78e6a697cdb8e73d7cb3c66d2b253d782b4 100644 (file)
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -3118,6 +3118,77 @@ defm VPMULDQ : avx512_binop_all<0x28, "vpmuldq", SSE_INTALU_ITINS_P,
 defm VPMULUDQ : avx512_binop_all<0xF4, "vpmuludq", SSE_INTMUL_ITINS_P,
                    X86pmuludq, 1>;
 
+multiclass avx512_packs_rmb<bits<8> opc, string OpcodeStr, SDNode OpNode,
+                            X86VectorVTInfo _Src, X86VectorVTInfo _Dst> {
+  let mayLoad = 1 in {
+      defm rmb : AVX512_maskable<opc, MRMSrcMem, _Dst, (outs _Dst.RC:$dst),
+                        (ins _Src.RC:$src1, _Src.ScalarMemOp:$src2), 
+                        OpcodeStr,
+                        "${src2}"##_Src.BroadcastStr##", $src1",
+                         "$src1, ${src2}"##_Src.BroadcastStr,
+                        (_Dst.VT (OpNode (_Src.VT _Src.RC:$src1), (bitconvert 
+                                     (_Src.VT (X86VBroadcast 
+                                              (_Src.ScalarLdFrag addr:$src2)))))),
+                        "">,
+                        EVEX_4V, EVEX_B, EVEX_CD8<_Src.EltSize, CD8VF>;
+  }
+}
+
+multiclass avx512_packs_rm<bits<8> opc, string OpcodeStr, 
+                            SDNode OpNode,X86VectorVTInfo _Src, 
+                            X86VectorVTInfo _Dst> {
+  defm rr : AVX512_maskable<opc, MRMSrcReg, _Dst, (outs _Dst.RC:$dst), 
+                            (ins _Src.RC:$src1, _Src.RC:$src2), OpcodeStr,
+                            "$src2, $src1","$src1, $src2", 
+                            (_Dst.VT (OpNode 
+                                         (_Src.VT _Src.RC:$src1), 
+                                         (_Src.VT _Src.RC:$src2))),
+                            "">,  EVEX_CD8<_Src.EltSize, CD8VF>, EVEX_4V;
+  let mayLoad = 1 in {
+    defm rm : AVX512_maskable<opc, MRMSrcMem, _Dst, (outs _Dst.RC:$dst),
+                          (ins _Src.RC:$src1, _Src.MemOp:$src2), OpcodeStr,
+                          "$src2, $src1", "$src1, $src2",
+                          (_Dst.VT (OpNode (_Src.VT _Src.RC:$src1),
+                                        (bitconvert (_Src.LdFrag addr:$src2)))),
+                            "">, EVEX_4V, EVEX_CD8<_Src.EltSize, CD8VF>;
+  }
+}
+
+multiclass avx512_packs_all_i32_i16<bits<8> opc, string OpcodeStr,
+                                    SDNode OpNode> {
+  defm NAME#Z : avx512_packs_rm<opc, OpcodeStr, OpNode, v16i32_info,
+                                 v32i16_info>,
+                avx512_packs_rmb<opc, OpcodeStr, OpNode, v16i32_info,
+                                 v32i16_info>, EVEX_V512;
+  let Predicates = [HasVLX] in {
+    defm NAME#Z256 : avx512_packs_rm<opc, OpcodeStr, OpNode, v8i32x_info,
+                                     v16i16x_info>,
+                     avx512_packs_rmb<opc, OpcodeStr, OpNode, v8i32x_info,
+                                     v16i16x_info>, EVEX_V256;
+    defm NAME#Z128 : avx512_packs_rm<opc, OpcodeStr, OpNode, v4i32x_info,
+                                     v8i16x_info>,
+                     avx512_packs_rmb<opc, OpcodeStr, OpNode, v4i32x_info,
+                                     v8i16x_info>, EVEX_V128;
+  }
+}
+multiclass avx512_packs_all_i16_i8<bits<8> opc, string OpcodeStr,
+                            SDNode OpNode> {
+  defm NAME#Z : avx512_packs_rm<opc, OpcodeStr, OpNode, v32i16_info,
+                                v64i8_info>, EVEX_V512;
+  let Predicates = [HasVLX] in {
+    defm NAME#Z256 : avx512_packs_rm<opc, OpcodeStr, OpNode, v16i16x_info,
+                                    v32i8x_info>, EVEX_V256;
+    defm NAME#Z128 : avx512_packs_rm<opc, OpcodeStr, OpNode, v8i16x_info,
+                                    v16i8x_info>, EVEX_V128;
+  }
+}
+let Predicates = [HasBWI] in {
+  defm VPACKSSDW : avx512_packs_all_i32_i16<0x6B, "vpackssdw", X86Packss>, PD;
+  defm VPACKUSDW : avx512_packs_all_i32_i16<0x2b, "vpackusdw", X86Packus>, T8PD;
+  defm VPACKSSWB : avx512_packs_all_i16_i8 <0x63, "vpacksswb", X86Packss>, AVX512BIBase, VEX_W;
+  defm VPACKUSWB : avx512_packs_all_i16_i8 <0x67, "vpackuswb", X86Packus>, AVX512BIBase, VEX_W;
+}
+
 defm VPMAXSB : avx512_binop_rm_vl_b<0x3C, "vpmaxs", X86smax,
                                      SSE_INTALU_ITINS_P, HasBWI, 1>, T8PD;
 defm VPMAXSW : avx512_binop_rm_vl_w<0xEE, "vpmaxs", X86smax,

diff --git a/lib/Target/X86/X86IntrinsicsInfo.h b/lib/Target/X86/X86IntrinsicsInfo.h
index 6944e5602103120a352c77f4620787882028edad..e633303334e545c40ff2e6d010d9e8317050e409 100644 (file)
--- a/lib/Target/X86/X86IntrinsicsInfo.h
+++ b/lib/Target/X86/X86IntrinsicsInfo.h
@@ -353,6 +353,18 @@ static const IntrinsicData  IntrinsicsWithoutChain[] = {
   X86_INTRINSIC_DATA(avx512_mask_or_ps_128, INTR_TYPE_2OP_MASK, X86ISD::FOR, 0),
   X86_INTRINSIC_DATA(avx512_mask_or_ps_256, INTR_TYPE_2OP_MASK, X86ISD::FOR, 0),
   X86_INTRINSIC_DATA(avx512_mask_or_ps_512, INTR_TYPE_2OP_MASK, X86ISD::FOR, 0),
+  X86_INTRINSIC_DATA(avx512_mask_packssdw_128, INTR_TYPE_2OP_MASK, X86ISD::PACKSS, 0),
+  X86_INTRINSIC_DATA(avx512_mask_packssdw_256, INTR_TYPE_2OP_MASK, X86ISD::PACKSS, 0),
+  X86_INTRINSIC_DATA(avx512_mask_packssdw_512, INTR_TYPE_2OP_MASK, X86ISD::PACKSS, 0),
+  X86_INTRINSIC_DATA(avx512_mask_packsswb_128, INTR_TYPE_2OP_MASK, X86ISD::PACKSS, 0),
+  X86_INTRINSIC_DATA(avx512_mask_packsswb_256, INTR_TYPE_2OP_MASK, X86ISD::PACKSS, 0),
+  X86_INTRINSIC_DATA(avx512_mask_packsswb_512, INTR_TYPE_2OP_MASK, X86ISD::PACKSS, 0),
+  X86_INTRINSIC_DATA(avx512_mask_packusdw_128, INTR_TYPE_2OP_MASK, X86ISD::PACKUS, 0),
+  X86_INTRINSIC_DATA(avx512_mask_packusdw_256, INTR_TYPE_2OP_MASK, X86ISD::PACKUS, 0),
+  X86_INTRINSIC_DATA(avx512_mask_packusdw_512, INTR_TYPE_2OP_MASK, X86ISD::PACKUS, 0),
+  X86_INTRINSIC_DATA(avx512_mask_packuswb_128, INTR_TYPE_2OP_MASK, X86ISD::PACKUS, 0),
+  X86_INTRINSIC_DATA(avx512_mask_packuswb_256, INTR_TYPE_2OP_MASK, X86ISD::PACKUS, 0),
+  X86_INTRINSIC_DATA(avx512_mask_packuswb_512, INTR_TYPE_2OP_MASK, X86ISD::PACKUS, 0),
   X86_INTRINSIC_DATA(avx512_mask_padd_b_128, INTR_TYPE_2OP_MASK, ISD::ADD, 0),
   X86_INTRINSIC_DATA(avx512_mask_padd_b_256, INTR_TYPE_2OP_MASK, ISD::ADD, 0),
   X86_INTRINSIC_DATA(avx512_mask_padd_b_512, INTR_TYPE_2OP_MASK, ISD::ADD, 0),

diff --git a/test/CodeGen/X86/avx512bw-intrinsics.ll b/test/CodeGen/X86/avx512bw-intrinsics.ll
index 308de16aefda959b16227436dc627bfb7230a8a4..0006efd091a8f04810db568c5faec2f9b31f744d 100644 (file)
--- a/test/CodeGen/X86/avx512bw-intrinsics.ll
+++ b/test/CodeGen/X86/avx512bw-intrinsics.ll
@@ -351,3 +351,252 @@ define <8 x i16> @test_x86_mask_blend_w_128(i8 %mask, <8 x i16> %a1, <8 x i16> %
   ret <8 x i16> %res
 }
 declare <8 x i16> @llvm.x86.avx512.mask.blend.w.128(<8 x i16>, <8 x i16>, i8) nounwind readonly
+
+define <32 x i16> @test_mask_packs_epi32_rr_512(<16 x i32> %a, <16 x i32> %b) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rr_512
+  ;CHECK: vpackssdw       %zmm1, %zmm0, %zmm0  ## encoding: [0x62,0xf1,0x7d,0x48,0x6b,0xc1]
+  %res = call <32 x i16> @llvm.x86.avx512.mask.packssdw.512(<16 x i32> %a, <16 x i32> %b, <32 x i16> zeroinitializer, i32 -1)
+  ret <32 x i16> %res
+}
+
+define <32 x i16> @test_mask_packs_epi32_rrk_512(<16 x i32> %a, <16 x i32> %b, <32 x i16> %passThru, i32 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rrk_512
+  ;CHECK: vpackssdw       %zmm1, %zmm0, %zmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x49,0x6b,0xd1]
+  %res = call <32 x i16> @llvm.x86.avx512.mask.packssdw.512(<16 x i32> %a, <16 x i32> %b, <32 x i16> %passThru, i32 %mask)
+  ret <32 x i16> %res
+}
+
+define <32 x i16> @test_mask_packs_epi32_rrkz_512(<16 x i32> %a, <16 x i32> %b, i32 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rrkz_512
+  ;CHECK: vpackssdw       %zmm1, %zmm0, %zmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xc9,0x6b,0xc1]
+  %res = call <32 x i16> @llvm.x86.avx512.mask.packssdw.512(<16 x i32> %a, <16 x i32> %b, <32 x i16> zeroinitializer, i32 %mask)
+  ret <32 x i16> %res
+}
+
+define <32 x i16> @test_mask_packs_epi32_rm_512(<16 x i32> %a, <16 x i32>* %ptr_b) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rm_512
+  ;CHECK: vpackssdw       (%rdi), %zmm0, %zmm0  ## encoding: [0x62,0xf1,0x7d,0x48,0x6b,0x07]
+  %b = load <16 x i32>, <16 x i32>* %ptr_b
+  %res = call <32 x i16> @llvm.x86.avx512.mask.packssdw.512(<16 x i32> %a, <16 x i32> %b, <32 x i16> zeroinitializer, i32 -1)
+  ret <32 x i16> %res
+}
+
+define <32 x i16> @test_mask_packs_epi32_rmk_512(<16 x i32> %a, <16 x i32>* %ptr_b, <32 x i16> %passThru, i32 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rmk_512
+  ;CHECK: vpackssdw       (%rdi), %zmm0, %zmm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x49,0x6b,0x0f]
+  %b = load <16 x i32>, <16 x i32>* %ptr_b
+  %res = call <32 x i16> @llvm.x86.avx512.mask.packssdw.512(<16 x i32> %a, <16 x i32> %b, <32 x i16> %passThru, i32 %mask)
+  ret <32 x i16> %res
+}
+
+define <32 x i16> @test_mask_packs_epi32_rmkz_512(<16 x i32> %a, <16 x i32>* %ptr_b, i32 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rmkz_512
+  ;CHECK: vpackssdw       (%rdi), %zmm0, %zmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xc9,0x6b,0x07]
+  %b = load <16 x i32>, <16 x i32>* %ptr_b
+  %res = call <32 x i16> @llvm.x86.avx512.mask.packssdw.512(<16 x i32> %a, <16 x i32> %b, <32 x i16> zeroinitializer, i32 %mask)
+  ret <32 x i16> %res
+}
+
+define <32 x i16> @test_mask_packs_epi32_rmb_512(<16 x i32> %a, i32* %ptr_b) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rmb_512
+  ;CHECK: vpackssdw       (%rdi){1to16}, %zmm0, %zmm0  ## encoding: [0x62,0xf1,0x7d,0x58,0x6b,0x07]
+  %q = load i32, i32* %ptr_b
+  %vecinit.i = insertelement <16 x i32> undef, i32 %q, i32 0
+  %b = shufflevector <16 x i32> %vecinit.i, <16 x i32> undef, <16 x i32> zeroinitializer
+  %res = call <32 x i16> @llvm.x86.avx512.mask.packssdw.512(<16 x i32> %a, <16 x i32> %b, <32 x i16> zeroinitializer, i32 -1)
+  ret <32 x i16> %res
+}
+
+define <32 x i16> @test_mask_packs_epi32_rmbk_512(<16 x i32> %a, i32* %ptr_b, <32 x i16> %passThru, i32 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rmbk_512
+  ;CHECK: vpackssdw       (%rdi){1to16}, %zmm0, %zmm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x59,0x6b,0x0f]
+  %q = load i32, i32* %ptr_b
+  %vecinit.i = insertelement <16 x i32> undef, i32 %q, i32 0
+  %b = shufflevector <16 x i32> %vecinit.i, <16 x i32> undef, <16 x i32> zeroinitializer
+  %res = call <32 x i16> @llvm.x86.avx512.mask.packssdw.512(<16 x i32> %a, <16 x i32> %b, <32 x i16> %passThru, i32 %mask)
+  ret <32 x i16> %res
+}
+
+define <32 x i16> @test_mask_packs_epi32_rmbkz_512(<16 x i32> %a, i32* %ptr_b, i32 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rmbkz_512
+  ;CHECK: vpackssdw       (%rdi){1to16}, %zmm0, %zmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xd9,0x6b,0x07]
+  %q = load i32, i32* %ptr_b
+  %vecinit.i = insertelement <16 x i32> undef, i32 %q, i32 0
+  %b = shufflevector <16 x i32> %vecinit.i, <16 x i32> undef, <16 x i32> zeroinitializer
+  %res = call <32 x i16> @llvm.x86.avx512.mask.packssdw.512(<16 x i32> %a, <16 x i32> %b, <32 x i16> zeroinitializer, i32 %mask)
+  ret <32 x i16> %res
+}
+
+declare <32 x i16> @llvm.x86.avx512.mask.packssdw.512(<16 x i32>, <16 x i32>, <32 x i16>, i32)
+
+define <64 x i8> @test_mask_packs_epi16_rr_512(<32 x i16> %a, <32 x i16> %b) {
+  ;CHECK-LABEL: test_mask_packs_epi16_rr_512
+  ;CHECK: vpacksswb       %zmm1, %zmm0, %zmm0  ## encoding: [0x62,0xf1,0xfd,0x48,0x63,0xc1]
+  %res = call <64 x i8> @llvm.x86.avx512.mask.packsswb.512(<32 x i16> %a, <32 x i16> %b, <64 x i8> zeroinitializer, i64 -1)
+  ret <64 x i8> %res
+}
+
+define <64 x i8> @test_mask_packs_epi16_rrk_512(<32 x i16> %a, <32 x i16> %b, <64 x i8> %passThru, i64 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi16_rrk_512
+  ;CHECK: vpacksswb       %zmm1, %zmm0, %zmm2 {%k1} ## encoding: [0x62,0xf1,0xfd,0x49,0x63,0xd1]
+  %res = call <64 x i8> @llvm.x86.avx512.mask.packsswb.512(<32 x i16> %a, <32 x i16> %b, <64 x i8> %passThru, i64 %mask)
+  ret <64 x i8> %res
+}
+
+define <64 x i8> @test_mask_packs_epi16_rrkz_512(<32 x i16> %a, <32 x i16> %b, i64 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi16_rrkz_512
+  ;CHECK: vpacksswb       %zmm1, %zmm0, %zmm0 {%k1} {z} ## encoding: [0x62,0xf1,0xfd,0xc9,0x63,0xc1]
+  %res = call <64 x i8> @llvm.x86.avx512.mask.packsswb.512(<32 x i16> %a, <32 x i16> %b, <64 x i8> zeroinitializer, i64 %mask)
+  ret <64 x i8> %res
+}
+
+define <64 x i8> @test_mask_packs_epi16_rm_512(<32 x i16> %a, <32 x i16>* %ptr_b) {
+  ;CHECK-LABEL: test_mask_packs_epi16_rm_512
+  ;CHECK: vpacksswb       (%rdi), %zmm0, %zmm0  ## encoding: [0x62,0xf1,0xfd,0x48,0x63,0x07]
+  %b = load <32 x i16>, <32 x i16>* %ptr_b
+  %res = call <64 x i8> @llvm.x86.avx512.mask.packsswb.512(<32 x i16> %a, <32 x i16> %b, <64 x i8> zeroinitializer, i64 -1)
+  ret <64 x i8> %res
+}
+
+define <64 x i8> @test_mask_packs_epi16_rmk_512(<32 x i16> %a, <32 x i16>* %ptr_b, <64 x i8> %passThru, i64 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi16_rmk_512
+  ;CHECK: vpacksswb       (%rdi), %zmm0, %zmm1 {%k1} ## encoding: [0x62,0xf1,0xfd,0x49,0x63,0x0f]
+  %b = load <32 x i16>, <32 x i16>* %ptr_b
+  %res = call <64 x i8> @llvm.x86.avx512.mask.packsswb.512(<32 x i16> %a, <32 x i16> %b, <64 x i8> %passThru, i64 %mask)
+  ret <64 x i8> %res
+}
+
+define <64 x i8> @test_mask_packs_epi16_rmkz_512(<32 x i16> %a, <32 x i16>* %ptr_b, i64 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi16_rmkz_512
+  ;CHECK: vpacksswb       (%rdi), %zmm0, %zmm0 {%k1} {z} ## encoding: [0x62,0xf1,0xfd,0xc9,0x63,0x07]
+  %b = load <32 x i16>, <32 x i16>* %ptr_b
+  %res = call <64 x i8> @llvm.x86.avx512.mask.packsswb.512(<32 x i16> %a, <32 x i16> %b, <64 x i8> zeroinitializer, i64 %mask)
+  ret <64 x i8> %res
+}
+
+declare <64 x i8> @llvm.x86.avx512.mask.packsswb.512(<32 x i16>, <32 x i16>, <64 x i8>, i64)
+
+
+define <32 x i16> @test_mask_packus_epi32_rr_512(<16 x i32> %a, <16 x i32> %b) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rr_512
+  ;CHECK: vpackusdw       %zmm1, %zmm0, %zmm0  
+  %res = call <32 x i16> @llvm.x86.avx512.mask.packusdw.512(<16 x i32> %a, <16 x i32> %b, <32 x i16> zeroinitializer, i32 -1)
+  ret <32 x i16> %res
+}
+
+define <32 x i16> @test_mask_packus_epi32_rrk_512(<16 x i32> %a, <16 x i32> %b, <32 x i16> %passThru, i32 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rrk_512
+  ;CHECK: vpackusdw       %zmm1, %zmm0, %zmm2 {%k1} 
+  %res = call <32 x i16> @llvm.x86.avx512.mask.packusdw.512(<16 x i32> %a, <16 x i32> %b, <32 x i16> %passThru, i32 %mask)
+  ret <32 x i16> %res
+}
+
+define <32 x i16> @test_mask_packus_epi32_rrkz_512(<16 x i32> %a, <16 x i32> %b, i32 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rrkz_512
+  ;CHECK: vpackusdw       %zmm1, %zmm0, %zmm0 {%k1} {z} 
+  %res = call <32 x i16> @llvm.x86.avx512.mask.packusdw.512(<16 x i32> %a, <16 x i32> %b, <32 x i16> zeroinitializer, i32 %mask)
+  ret <32 x i16> %res
+}
+
+define <32 x i16> @test_mask_packus_epi32_rm_512(<16 x i32> %a, <16 x i32>* %ptr_b) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rm_512
+  ;CHECK: vpackusdw       (%rdi), %zmm0, %zmm0  
+  %b = load <16 x i32>, <16 x i32>* %ptr_b
+  %res = call <32 x i16> @llvm.x86.avx512.mask.packusdw.512(<16 x i32> %a, <16 x i32> %b, <32 x i16> zeroinitializer, i32 -1)
+  ret <32 x i16> %res
+}
+
+define <32 x i16> @test_mask_packus_epi32_rmk_512(<16 x i32> %a, <16 x i32>* %ptr_b, <32 x i16> %passThru, i32 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rmk_512
+  ;CHECK: vpackusdw       (%rdi), %zmm0, %zmm1 {%k1} 
+  %b = load <16 x i32>, <16 x i32>* %ptr_b
+  %res = call <32 x i16> @llvm.x86.avx512.mask.packusdw.512(<16 x i32> %a, <16 x i32> %b, <32 x i16> %passThru, i32 %mask)
+  ret <32 x i16> %res
+}
+
+define <32 x i16> @test_mask_packus_epi32_rmkz_512(<16 x i32> %a, <16 x i32>* %ptr_b, i32 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rmkz_512
+  ;CHECK: vpackusdw       (%rdi), %zmm0, %zmm0 {%k1} {z} 
+  %b = load <16 x i32>, <16 x i32>* %ptr_b
+  %res = call <32 x i16> @llvm.x86.avx512.mask.packusdw.512(<16 x i32> %a, <16 x i32> %b, <32 x i16> zeroinitializer, i32 %mask)
+  ret <32 x i16> %res
+}
+
+define <32 x i16> @test_mask_packus_epi32_rmb_512(<16 x i32> %a, i32* %ptr_b) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rmb_512
+  ;CHECK: vpackusdw       (%rdi){1to16}, %zmm0, %zmm0  
+  %q = load i32, i32* %ptr_b
+  %vecinit.i = insertelement <16 x i32> undef, i32 %q, i32 0
+  %b = shufflevector <16 x i32> %vecinit.i, <16 x i32> undef, <16 x i32> zeroinitializer
+  %res = call <32 x i16> @llvm.x86.avx512.mask.packusdw.512(<16 x i32> %a, <16 x i32> %b, <32 x i16> zeroinitializer, i32 -1)
+  ret <32 x i16> %res
+}
+
+define <32 x i16> @test_mask_packus_epi32_rmbk_512(<16 x i32> %a, i32* %ptr_b, <32 x i16> %passThru, i32 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rmbk_512
+  ;CHECK: vpackusdw       (%rdi){1to16}, %zmm0, %zmm1 {%k1} 
+  %q = load i32, i32* %ptr_b
+  %vecinit.i = insertelement <16 x i32> undef, i32 %q, i32 0
+  %b = shufflevector <16 x i32> %vecinit.i, <16 x i32> undef, <16 x i32> zeroinitializer
+  %res = call <32 x i16> @llvm.x86.avx512.mask.packusdw.512(<16 x i32> %a, <16 x i32> %b, <32 x i16> %passThru, i32 %mask)
+  ret <32 x i16> %res
+}
+
+define <32 x i16> @test_mask_packus_epi32_rmbkz_512(<16 x i32> %a, i32* %ptr_b, i32 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rmbkz_512
+  ;CHECK: vpackusdw       (%rdi){1to16}, %zmm0, %zmm0 {%k1} {z} 
+  %q = load i32, i32* %ptr_b
+  %vecinit.i = insertelement <16 x i32> undef, i32 %q, i32 0
+  %b = shufflevector <16 x i32> %vecinit.i, <16 x i32> undef, <16 x i32> zeroinitializer
+  %res = call <32 x i16> @llvm.x86.avx512.mask.packusdw.512(<16 x i32> %a, <16 x i32> %b, <32 x i16> zeroinitializer, i32 %mask)
+  ret <32 x i16> %res
+}
+
+declare <32 x i16> @llvm.x86.avx512.mask.packusdw.512(<16 x i32>, <16 x i32>, <32 x i16>, i32)
+
+define <64 x i8> @test_mask_packus_epi16_rr_512(<32 x i16> %a, <32 x i16> %b) {
+  ;CHECK-LABEL: test_mask_packus_epi16_rr_512
+  ;CHECK: vpackuswb       %zmm1, %zmm0, %zmm0  
+  %res = call <64 x i8> @llvm.x86.avx512.mask.packuswb.512(<32 x i16> %a, <32 x i16> %b, <64 x i8> zeroinitializer, i64 -1)
+  ret <64 x i8> %res
+}
+
+define <64 x i8> @test_mask_packus_epi16_rrk_512(<32 x i16> %a, <32 x i16> %b, <64 x i8> %passThru, i64 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi16_rrk_512
+  ;CHECK: vpackuswb       %zmm1, %zmm0, %zmm2 {%k1} 
+  %res = call <64 x i8> @llvm.x86.avx512.mask.packuswb.512(<32 x i16> %a, <32 x i16> %b, <64 x i8> %passThru, i64 %mask)
+  ret <64 x i8> %res
+}
+
+define <64 x i8> @test_mask_packus_epi16_rrkz_512(<32 x i16> %a, <32 x i16> %b, i64 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi16_rrkz_512
+  ;CHECK: vpackuswb       %zmm1, %zmm0, %zmm0 {%k1} {z} 
+  %res = call <64 x i8> @llvm.x86.avx512.mask.packuswb.512(<32 x i16> %a, <32 x i16> %b, <64 x i8> zeroinitializer, i64 %mask)
+  ret <64 x i8> %res
+}
+
+define <64 x i8> @test_mask_packus_epi16_rm_512(<32 x i16> %a, <32 x i16>* %ptr_b) {
+  ;CHECK-LABEL: test_mask_packus_epi16_rm_512
+  ;CHECK: vpackuswb       (%rdi), %zmm0, %zmm0  
+  %b = load <32 x i16>, <32 x i16>* %ptr_b
+  %res = call <64 x i8> @llvm.x86.avx512.mask.packuswb.512(<32 x i16> %a, <32 x i16> %b, <64 x i8> zeroinitializer, i64 -1)
+  ret <64 x i8> %res
+}
+
+define <64 x i8> @test_mask_packus_epi16_rmk_512(<32 x i16> %a, <32 x i16>* %ptr_b, <64 x i8> %passThru, i64 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi16_rmk_512
+  ;CHECK: vpackuswb       (%rdi), %zmm0, %zmm1 {%k1} 
+  %b = load <32 x i16>, <32 x i16>* %ptr_b
+  %res = call <64 x i8> @llvm.x86.avx512.mask.packuswb.512(<32 x i16> %a, <32 x i16> %b, <64 x i8> %passThru, i64 %mask)
+  ret <64 x i8> %res
+}
+
+define <64 x i8> @test_mask_packus_epi16_rmkz_512(<32 x i16> %a, <32 x i16>* %ptr_b, i64 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi16_rmkz_512
+  ;CHECK: vpackuswb       (%rdi), %zmm0, %zmm0 {%k1} {z} 
+  %b = load <32 x i16>, <32 x i16>* %ptr_b
+  %res = call <64 x i8> @llvm.x86.avx512.mask.packuswb.512(<32 x i16> %a, <32 x i16> %b, <64 x i8> zeroinitializer, i64 %mask)
+  ret <64 x i8> %res
+}
+
+declare <64 x i8> @llvm.x86.avx512.mask.packuswb.512(<32 x i16>, <32 x i16>, <64 x i8>, i64)
\ No newline at end of file

diff --git a/test/CodeGen/X86/avx512bwvl-intrinsics.ll b/test/CodeGen/X86/avx512bwvl-intrinsics.ll
index 03a0466ab19500e9ac093f0bc6e6730ce3a685a0..0834bc6afc6ee9b87c1898580505b94dc131b857 100644 (file)
--- a/test/CodeGen/X86/avx512bwvl-intrinsics.ll
+++ b/test/CodeGen/X86/avx512bwvl-intrinsics.ll
@@ -1419,3 +1419,501 @@ define <16 x i16> @test_mask_mullo_epi16_rmkz_256(<16 x i16> %a, <16 x i16>* %pt
 
 declare <16 x i16> @llvm.x86.avx512.mask.pmull.w.256(<16 x i16>, <16 x i16>, <16 x i16>, i16)
 
+
+define <8 x i16> @test_mask_packs_epi32_rr_128(<4 x i32> %a, <4 x i32> %b) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rr_128
+  ;CHECK: vpackssdw       %xmm1, %xmm0, %xmm0 ## encoding: [0xc5,0xf9,0x6b,0xc1]
+  %res = call <8 x i16> @llvm.x86.avx512.mask.packssdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 -1)
+  ret <8 x i16> %res
+}
+
+define <8 x i16> @test_mask_packs_epi32_rrk_128(<4 x i32> %a, <4 x i32> %b, <8 x i16> %passThru, i8 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rrk_128
+  ;CHECK: vpackssdw       %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x6b,0xd1]
+  %res = call <8 x i16> @llvm.x86.avx512.mask.packssdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> %passThru, i8 %mask)
+  ret <8 x i16> %res
+}
+
+define <8 x i16> @test_mask_packs_epi32_rrkz_128(<4 x i32> %a, <4 x i32> %b, i8 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rrkz_128
+  ;CHECK: vpackssdw       %xmm1, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0x6b,0xc1]
+  %res = call <8 x i16> @llvm.x86.avx512.mask.packssdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 %mask)
+  ret <8 x i16> %res
+}
+
+define <8 x i16> @test_mask_packs_epi32_rm_128(<4 x i32> %a, <4 x i32>* %ptr_b) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rm_128
+  ;CHECK: vpackssdw       (%rdi), %xmm0, %xmm0 ## encoding: [0xc5,0xf9,0x6b,0x07]
+  %b = load <4 x i32>, <4 x i32>* %ptr_b
+  %res = call <8 x i16> @llvm.x86.avx512.mask.packssdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 -1)
+  ret <8 x i16> %res
+}
+
+define <8 x i16> @test_mask_packs_epi32_rmk_128(<4 x i32> %a, <4 x i32>* %ptr_b, <8 x i16> %passThru, i8 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rmk_128
+  ;CHECK: vpackssdw       (%rdi), %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x6b,0x0f]
+  %b = load <4 x i32>, <4 x i32>* %ptr_b
+  %res = call <8 x i16> @llvm.x86.avx512.mask.packssdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> %passThru, i8 %mask)
+  ret <8 x i16> %res
+}
+
+define <8 x i16> @test_mask_packs_epi32_rmkz_128(<4 x i32> %a, <4 x i32>* %ptr_b, i8 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rmkz_128
+  ;CHECK: vpackssdw       (%rdi), %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x89,0x6b,0x07]
+  %b = load <4 x i32>, <4 x i32>* %ptr_b
+  %res = call <8 x i16> @llvm.x86.avx512.mask.packssdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 %mask)
+  ret <8 x i16> %res
+}
+
+define <8 x i16> @test_mask_packs_epi32_rmb_128(<4 x i32> %a, i32* %ptr_b) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rmb_128
+  ;CHECK: vpackssdw       (%rdi){1to4}, %xmm0, %xmm0  ## encoding: [0x62,0xf1,0x7d,0x18,0x6b,0x07]
+  %q = load i32, i32* %ptr_b
+  %vecinit.i = insertelement <4 x i32> undef, i32 %q, i32 0
+  %b = shufflevector <4 x i32> %vecinit.i, <4 x i32> undef, <4 x i32> zeroinitializer
+  %res = call <8 x i16> @llvm.x86.avx512.mask.packssdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 -1)
+  ret <8 x i16> %res
+}
+
+define <8 x i16> @test_mask_packs_epi32_rmbk_128(<4 x i32> %a, i32* %ptr_b, <8 x i16> %passThru, i8 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rmbk_128
+  ;CHECK: vpackssdw       (%rdi){1to4}, %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x19,0x6b,0x0f]
+  %q = load i32, i32* %ptr_b
+  %vecinit.i = insertelement <4 x i32> undef, i32 %q, i32 0
+  %b = shufflevector <4 x i32> %vecinit.i, <4 x i32> undef, <4 x i32> zeroinitializer
+  %res = call <8 x i16> @llvm.x86.avx512.mask.packssdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> %passThru, i8 %mask)
+  ret <8 x i16> %res
+}
+
+define <8 x i16> @test_mask_packs_epi32_rmbkz_128(<4 x i32> %a, i32* %ptr_b, i8 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rmbkz_128
+  ;CHECK: vpackssdw       (%rdi){1to4}, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0x99,0x6b,0x07]
+  %q = load i32, i32* %ptr_b
+  %vecinit.i = insertelement <4 x i32> undef, i32 %q, i32 0
+  %b = shufflevector <4 x i32> %vecinit.i, <4 x i32> undef, <4 x i32> zeroinitializer
+  %res = call <8 x i16> @llvm.x86.avx512.mask.packssdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 %mask)
+  ret <8 x i16> %res
+}
+
+declare <8 x i16> @llvm.x86.avx512.mask.packssdw.128(<4 x i32>, <4 x i32>, <8 x i16>, i8)
+
+define <16 x i16> @test_mask_packs_epi32_rr_256(<8 x i32> %a, <8 x i32> %b) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rr_256
+  ;CHECK: vpackssdw       %ymm1, %ymm0, %ymm0 ## encoding: [0xc5,0xfd,0x6b,0xc1]
+  %res = call <16 x i16> @llvm.x86.avx512.mask.packssdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 -1)
+  ret <16 x i16> %res
+}
+
+define <16 x i16> @test_mask_packs_epi32_rrk_256(<8 x i32> %a, <8 x i32> %b, <16 x i16> %passThru, i16 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rrk_256
+  ;CHECK: vpackssdw       %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x6b,0xd1]
+  %res = call <16 x i16> @llvm.x86.avx512.mask.packssdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> %passThru, i16 %mask)
+  ret <16 x i16> %res
+}
+
+define <16 x i16> @test_mask_packs_epi32_rrkz_256(<8 x i32> %a, <8 x i32> %b, i16 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rrkz_256
+  ;CHECK: vpackssdw       %ymm1, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0x6b,0xc1]
+  %res = call <16 x i16> @llvm.x86.avx512.mask.packssdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 %mask)
+  ret <16 x i16> %res
+}
+
+define <16 x i16> @test_mask_packs_epi32_rm_256(<8 x i32> %a, <8 x i32>* %ptr_b) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rm_256
+  ;CHECK: vpackssdw       (%rdi), %ymm0, %ymm0 ## encoding: [0xc5,0xfd,0x6b,0x07]
+  %b = load <8 x i32>, <8 x i32>* %ptr_b
+  %res = call <16 x i16> @llvm.x86.avx512.mask.packssdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 -1)
+  ret <16 x i16> %res
+}
+
+define <16 x i16> @test_mask_packs_epi32_rmk_256(<8 x i32> %a, <8 x i32>* %ptr_b, <16 x i16> %passThru, i16 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rmk_256
+  ;CHECK: vpackssdw       (%rdi), %ymm0, %ymm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x29,0x6b,0x0f]
+  %b = load <8 x i32>, <8 x i32>* %ptr_b
+  %res = call <16 x i16> @llvm.x86.avx512.mask.packssdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> %passThru, i16 %mask)
+  ret <16 x i16> %res
+}
+
+define <16 x i16> @test_mask_packs_epi32_rmkz_256(<8 x i32> %a, <8 x i32>* %ptr_b, i16 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rmkz_256
+  ;CHECK: vpackssdw       (%rdi), %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xa9,0x6b,0x07]
+  %b = load <8 x i32>, <8 x i32>* %ptr_b
+  %res = call <16 x i16> @llvm.x86.avx512.mask.packssdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 %mask)
+  ret <16 x i16> %res
+}
+
+define <16 x i16> @test_mask_packs_epi32_rmb_256(<8 x i32> %a, i32* %ptr_b) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rmb_256
+  ;CHECK: vpackssdw       (%rdi){1to8}, %ymm0, %ymm0  ## encoding: [0x62,0xf1,0x7d,0x38,0x6b,0x07]
+  %q = load i32, i32* %ptr_b
+  %vecinit.i = insertelement <8 x i32> undef, i32 %q, i32 0
+  %b = shufflevector <8 x i32> %vecinit.i, <8 x i32> undef, <8 x i32> zeroinitializer
+  %res = call <16 x i16> @llvm.x86.avx512.mask.packssdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 -1)
+  ret <16 x i16> %res
+}
+
+define <16 x i16> @test_mask_packs_epi32_rmbk_256(<8 x i32> %a, i32* %ptr_b, <16 x i16> %passThru, i16 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rmbk_256
+  ;CHECK: vpackssdw       (%rdi){1to8}, %ymm0, %ymm1 {%k1} ## encoding: [0x62,0xf1,0x7d,0x39,0x6b,0x0f]
+  %q = load i32, i32* %ptr_b
+  %vecinit.i = insertelement <8 x i32> undef, i32 %q, i32 0
+  %b = shufflevector <8 x i32> %vecinit.i, <8 x i32> undef, <8 x i32> zeroinitializer
+  %res = call <16 x i16> @llvm.x86.avx512.mask.packssdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> %passThru, i16 %mask)
+  ret <16 x i16> %res
+}
+
+define <16 x i16> @test_mask_packs_epi32_rmbkz_256(<8 x i32> %a, i32* %ptr_b, i16 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi32_rmbkz_256
+  ;CHECK: vpackssdw       (%rdi){1to8}, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0x7d,0xb9,0x6b,0x07]  
+  %q = load i32, i32* %ptr_b
+  %vecinit.i = insertelement <8 x i32> undef, i32 %q, i32 0
+  %b = shufflevector <8 x i32> %vecinit.i, <8 x i32> undef, <8 x i32> zeroinitializer
+  %res = call <16 x i16> @llvm.x86.avx512.mask.packssdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 %mask)
+  ret <16 x i16> %res
+}
+
+declare <16 x i16> @llvm.x86.avx512.mask.packssdw.256(<8 x i32>, <8 x i32>, <16 x i16>, i16)
+
+define <16 x i8> @test_mask_packs_epi16_rr_128(<8 x i16> %a, <8 x i16> %b) {
+  ;CHECK-LABEL: test_mask_packs_epi16_rr_128
+  ;CHECK: vpacksswb       %xmm1, %xmm0, %xmm0 ## encoding: [0xc5,0xf9,0x63,0xc1]
+  %res = call <16 x i8> @llvm.x86.avx512.mask.packsswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> zeroinitializer, i16 -1)
+  ret <16 x i8> %res
+}
+
+define <16 x i8> @test_mask_packs_epi16_rrk_128(<8 x i16> %a, <8 x i16> %b, <16 x i8> %passThru, i16 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi16_rrk_128
+  ;CHECK: vpacksswb       %xmm1, %xmm0, %xmm2 {%k1} ## encoding: [0x62,0xf1,0xfd,0x09,0x63,0xd1]
+  %res = call <16 x i8> @llvm.x86.avx512.mask.packsswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> %passThru, i16 %mask)
+  ret <16 x i8> %res
+}
+
+define <16 x i8> @test_mask_packs_epi16_rrkz_128(<8 x i16> %a, <8 x i16> %b, i16 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi16_rrkz_128
+  ;CHECK: vpacksswb       %xmm1, %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0xfd,0x89,0x63,0xc1]
+  %res = call <16 x i8> @llvm.x86.avx512.mask.packsswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> zeroinitializer, i16 %mask)
+  ret <16 x i8> %res
+}
+
+define <16 x i8> @test_mask_packs_epi16_rm_128(<8 x i16> %a, <8 x i16>* %ptr_b) {
+  ;CHECK-LABEL: test_mask_packs_epi16_rm_128
+  ;CHECK: vpacksswb       (%rdi), %xmm0, %xmm0 ## encoding: [0xc5,0xf9,0x63,0x07]
+  %b = load <8 x i16>, <8 x i16>* %ptr_b
+  %res = call <16 x i8> @llvm.x86.avx512.mask.packsswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> zeroinitializer, i16 -1)
+  ret <16 x i8> %res
+}
+
+define <16 x i8> @test_mask_packs_epi16_rmk_128(<8 x i16> %a, <8 x i16>* %ptr_b, <16 x i8> %passThru, i16 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi16_rmk_128
+  ;CHECK: vpacksswb       (%rdi), %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf1,0xfd,0x09,0x63,0x0f]
+  %b = load <8 x i16>, <8 x i16>* %ptr_b
+  %res = call <16 x i8> @llvm.x86.avx512.mask.packsswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> %passThru, i16 %mask)
+  ret <16 x i8> %res
+}
+
+define <16 x i8> @test_mask_packs_epi16_rmkz_128(<8 x i16> %a, <8 x i16>* %ptr_b, i16 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi16_rmkz_128
+  ;CHECK: vpacksswb       (%rdi), %xmm0, %xmm0 {%k1} {z} ## encoding: [0x62,0xf1,0xfd,0x89,0x63,0x07]
+  %b = load <8 x i16>, <8 x i16>* %ptr_b
+  %res = call <16 x i8> @llvm.x86.avx512.mask.packsswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> zeroinitializer, i16 %mask)
+  ret <16 x i8> %res
+}
+
+declare <16 x i8> @llvm.x86.avx512.mask.packsswb.128(<8 x i16>, <8 x i16>, <16 x i8>, i16)
+
+define <32 x i8> @test_mask_packs_epi16_rr_256(<16 x i16> %a, <16 x i16> %b) {
+  ;CHECK-LABEL: test_mask_packs_epi16_rr_256
+  ;CHECK: vpacksswb       %ymm1, %ymm0, %ymm0 ## encoding: [0xc5,0xfd,0x63,0xc1]
+  %res = call <32 x i8> @llvm.x86.avx512.mask.packsswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> zeroinitializer, i32 -1)
+  ret <32 x i8> %res
+}
+
+define <32 x i8> @test_mask_packs_epi16_rrk_256(<16 x i16> %a, <16 x i16> %b, <32 x i8> %passThru, i32 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi16_rrk_256
+  ;CHECK: vpacksswb       %ymm1, %ymm0, %ymm2 {%k1} ## encoding: [0x62,0xf1,0xfd,0x29,0x63,0xd1]
+  %res = call <32 x i8> @llvm.x86.avx512.mask.packsswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> %passThru, i32 %mask)
+  ret <32 x i8> %res
+}
+
+define <32 x i8> @test_mask_packs_epi16_rrkz_256(<16 x i16> %a, <16 x i16> %b, i32 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi16_rrkz_256
+  ;CHECK: vpacksswb       %ymm1, %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0xfd,0xa9,0x63,0xc1]
+  %res = call <32 x i8> @llvm.x86.avx512.mask.packsswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> zeroinitializer, i32 %mask)
+  ret <32 x i8> %res
+}
+
+define <32 x i8> @test_mask_packs_epi16_rm_256(<16 x i16> %a, <16 x i16>* %ptr_b) {
+  ;CHECK-LABEL: test_mask_packs_epi16_rm_256
+  ;CHECK: vpacksswb       (%rdi), %ymm0, %ymm0 ## encoding: [0xc5,0xfd,0x63,0x07]
+  %b = load <16 x i16>, <16 x i16>* %ptr_b
+  %res = call <32 x i8> @llvm.x86.avx512.mask.packsswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> zeroinitializer, i32 -1)
+  ret <32 x i8> %res
+}
+
+define <32 x i8> @test_mask_packs_epi16_rmk_256(<16 x i16> %a, <16 x i16>* %ptr_b, <32 x i8> %passThru, i32 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi16_rmk_256
+  ;CHECK: vpacksswb       (%rdi), %ymm0, %ymm1 {%k1} ## encoding: [0x62,0xf1,0xfd,0x29,0x63,0x0f]
+  %b = load <16 x i16>, <16 x i16>* %ptr_b
+  %res = call <32 x i8> @llvm.x86.avx512.mask.packsswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> %passThru, i32 %mask)
+  ret <32 x i8> %res
+}
+
+define <32 x i8> @test_mask_packs_epi16_rmkz_256(<16 x i16> %a, <16 x i16>* %ptr_b, i32 %mask) {
+  ;CHECK-LABEL: test_mask_packs_epi16_rmkz_256
+  ;CHECK: vpacksswb       (%rdi), %ymm0, %ymm0 {%k1} {z} ## encoding: [0x62,0xf1,0xfd,0xa9,0x63,0x07]
+  %b = load <16 x i16>, <16 x i16>* %ptr_b
+  %res = call <32 x i8> @llvm.x86.avx512.mask.packsswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> zeroinitializer, i32 %mask)
+  ret <32 x i8> %res
+}
+
+declare <32 x i8> @llvm.x86.avx512.mask.packsswb.256(<16 x i16>, <16 x i16>, <32 x i8>, i32)
+
+
+define <8 x i16> @test_mask_packus_epi32_rr_128(<4 x i32> %a, <4 x i32> %b) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rr_128
+  ;CHECK: vpackusdw       %xmm1, %xmm0, %xmm0 
+  %res = call <8 x i16> @llvm.x86.avx512.mask.packusdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 -1)
+  ret <8 x i16> %res
+}
+
+define <8 x i16> @test_mask_packus_epi32_rrk_128(<4 x i32> %a, <4 x i32> %b, <8 x i16> %passThru, i8 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rrk_128
+  ;CHECK: vpackusdw       %xmm1, %xmm0, %xmm2 {%k1} 
+  %res = call <8 x i16> @llvm.x86.avx512.mask.packusdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> %passThru, i8 %mask)
+  ret <8 x i16> %res
+}
+
+define <8 x i16> @test_mask_packus_epi32_rrkz_128(<4 x i32> %a, <4 x i32> %b, i8 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rrkz_128
+  ;CHECK: vpackusdw       %xmm1, %xmm0, %xmm0 {%k1} {z} 
+  %res = call <8 x i16> @llvm.x86.avx512.mask.packusdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 %mask)
+  ret <8 x i16> %res
+}
+
+define <8 x i16> @test_mask_packus_epi32_rm_128(<4 x i32> %a, <4 x i32>* %ptr_b) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rm_128
+  ;CHECK: vpackusdw       (%rdi), %xmm0, %xmm0 
+  %b = load <4 x i32>, <4 x i32>* %ptr_b
+  %res = call <8 x i16> @llvm.x86.avx512.mask.packusdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 -1)
+  ret <8 x i16> %res
+}
+
+define <8 x i16> @test_mask_packus_epi32_rmk_128(<4 x i32> %a, <4 x i32>* %ptr_b, <8 x i16> %passThru, i8 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rmk_128
+  ;CHECK: vpackusdw       (%rdi), %xmm0, %xmm1 {%k1} 
+  %b = load <4 x i32>, <4 x i32>* %ptr_b
+  %res = call <8 x i16> @llvm.x86.avx512.mask.packusdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> %passThru, i8 %mask)
+  ret <8 x i16> %res
+}
+
+define <8 x i16> @test_mask_packus_epi32_rmkz_128(<4 x i32> %a, <4 x i32>* %ptr_b, i8 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rmkz_128
+  ;CHECK: vpackusdw       (%rdi), %xmm0, %xmm0 {%k1} {z} 
+  %b = load <4 x i32>, <4 x i32>* %ptr_b
+  %res = call <8 x i16> @llvm.x86.avx512.mask.packusdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 %mask)
+  ret <8 x i16> %res
+}
+
+define <8 x i16> @test_mask_packus_epi32_rmb_128(<4 x i32> %a, i32* %ptr_b) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rmb_128
+  ;CHECK: vpackusdw       (%rdi){1to4}, %xmm0, %xmm0  
+  %q = load i32, i32* %ptr_b
+  %vecinit.i = insertelement <4 x i32> undef, i32 %q, i32 0
+  %b = shufflevector <4 x i32> %vecinit.i, <4 x i32> undef, <4 x i32> zeroinitializer
+  %res = call <8 x i16> @llvm.x86.avx512.mask.packusdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 -1)
+  ret <8 x i16> %res
+}
+
+define <8 x i16> @test_mask_packus_epi32_rmbk_128(<4 x i32> %a, i32* %ptr_b, <8 x i16> %passThru, i8 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rmbk_128
+  ;CHECK: vpackusdw       (%rdi){1to4}, %xmm0, %xmm1 {%k1} 
+  %q = load i32, i32* %ptr_b
+  %vecinit.i = insertelement <4 x i32> undef, i32 %q, i32 0
+  %b = shufflevector <4 x i32> %vecinit.i, <4 x i32> undef, <4 x i32> zeroinitializer
+  %res = call <8 x i16> @llvm.x86.avx512.mask.packusdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> %passThru, i8 %mask)
+  ret <8 x i16> %res
+}
+
+define <8 x i16> @test_mask_packus_epi32_rmbkz_128(<4 x i32> %a, i32* %ptr_b, i8 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rmbkz_128
+  ;CHECK: vpackusdw       (%rdi){1to4}, %xmm0, %xmm0 {%k1} {z} 
+  %q = load i32, i32* %ptr_b
+  %vecinit.i = insertelement <4 x i32> undef, i32 %q, i32 0
+  %b = shufflevector <4 x i32> %vecinit.i, <4 x i32> undef, <4 x i32> zeroinitializer
+  %res = call <8 x i16> @llvm.x86.avx512.mask.packusdw.128(<4 x i32> %a, <4 x i32> %b, <8 x i16> zeroinitializer, i8 %mask)
+  ret <8 x i16> %res
+}
+
+declare <8 x i16> @llvm.x86.avx512.mask.packusdw.128(<4 x i32>, <4 x i32>, <8 x i16>, i8)
+
+define <16 x i16> @test_mask_packus_epi32_rr_256(<8 x i32> %a, <8 x i32> %b) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rr_256
+  ;CHECK: vpackusdw       %ymm1, %ymm0, %ymm0 
+  %res = call <16 x i16> @llvm.x86.avx512.mask.packusdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 -1)
+  ret <16 x i16> %res
+}
+
+define <16 x i16> @test_mask_packus_epi32_rrk_256(<8 x i32> %a, <8 x i32> %b, <16 x i16> %passThru, i16 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rrk_256
+  ;CHECK: vpackusdw       %ymm1, %ymm0, %ymm2 {%k1} 
+  %res = call <16 x i16> @llvm.x86.avx512.mask.packusdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> %passThru, i16 %mask)
+  ret <16 x i16> %res
+}
+
+define <16 x i16> @test_mask_packus_epi32_rrkz_256(<8 x i32> %a, <8 x i32> %b, i16 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rrkz_256
+  ;CHECK: vpackusdw       %ymm1, %ymm0, %ymm0 {%k1} {z} 
+  %res = call <16 x i16> @llvm.x86.avx512.mask.packusdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 %mask)
+  ret <16 x i16> %res
+}
+
+define <16 x i16> @test_mask_packus_epi32_rm_256(<8 x i32> %a, <8 x i32>* %ptr_b) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rm_256
+  ;CHECK: vpackusdw       (%rdi), %ymm0, %ymm0 
+  %b = load <8 x i32>, <8 x i32>* %ptr_b
+  %res = call <16 x i16> @llvm.x86.avx512.mask.packusdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 -1)
+  ret <16 x i16> %res
+}
+
+define <16 x i16> @test_mask_packus_epi32_rmk_256(<8 x i32> %a, <8 x i32>* %ptr_b, <16 x i16> %passThru, i16 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rmk_256
+  ;CHECK: vpackusdw       (%rdi), %ymm0, %ymm1 {%k1} 
+  %b = load <8 x i32>, <8 x i32>* %ptr_b
+  %res = call <16 x i16> @llvm.x86.avx512.mask.packusdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> %passThru, i16 %mask)
+  ret <16 x i16> %res
+}
+
+define <16 x i16> @test_mask_packus_epi32_rmkz_256(<8 x i32> %a, <8 x i32>* %ptr_b, i16 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rmkz_256
+  ;CHECK: vpackusdw       (%rdi), %ymm0, %ymm0 {%k1} {z} 
+  %b = load <8 x i32>, <8 x i32>* %ptr_b
+  %res = call <16 x i16> @llvm.x86.avx512.mask.packusdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 %mask)
+  ret <16 x i16> %res
+}
+
+define <16 x i16> @test_mask_packus_epi32_rmb_256(<8 x i32> %a, i32* %ptr_b) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rmb_256
+  ;CHECK: vpackusdw       (%rdi){1to8}, %ymm0, %ymm0  
+  %q = load i32, i32* %ptr_b
+  %vecinit.i = insertelement <8 x i32> undef, i32 %q, i32 0
+  %b = shufflevector <8 x i32> %vecinit.i, <8 x i32> undef, <8 x i32> zeroinitializer
+  %res = call <16 x i16> @llvm.x86.avx512.mask.packusdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 -1)
+  ret <16 x i16> %res
+}
+
+define <16 x i16> @test_mask_packus_epi32_rmbk_256(<8 x i32> %a, i32* %ptr_b, <16 x i16> %passThru, i16 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rmbk_256
+  ;CHECK: vpackusdw       (%rdi){1to8}, %ymm0, %ymm1 {%k1} 
+  %q = load i32, i32* %ptr_b
+  %vecinit.i = insertelement <8 x i32> undef, i32 %q, i32 0
+  %b = shufflevector <8 x i32> %vecinit.i, <8 x i32> undef, <8 x i32> zeroinitializer
+  %res = call <16 x i16> @llvm.x86.avx512.mask.packusdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> %passThru, i16 %mask)
+  ret <16 x i16> %res
+}
+
+define <16 x i16> @test_mask_packus_epi32_rmbkz_256(<8 x i32> %a, i32* %ptr_b, i16 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi32_rmbkz_256
+  ;CHECK: vpackusdw       (%rdi){1to8}, %ymm0, %ymm0 {%k1} {z} 
+  %q = load i32, i32* %ptr_b
+  %vecinit.i = insertelement <8 x i32> undef, i32 %q, i32 0
+  %b = shufflevector <8 x i32> %vecinit.i, <8 x i32> undef, <8 x i32> zeroinitializer
+  %res = call <16 x i16> @llvm.x86.avx512.mask.packusdw.256(<8 x i32> %a, <8 x i32> %b, <16 x i16> zeroinitializer, i16 %mask)
+  ret <16 x i16> %res
+}
+
+declare <16 x i16> @llvm.x86.avx512.mask.packusdw.256(<8 x i32>, <8 x i32>, <16 x i16>, i16)
+
+define <16 x i8> @test_mask_packus_epi16_rr_128(<8 x i16> %a, <8 x i16> %b) {
+  ;CHECK-LABEL: test_mask_packus_epi16_rr_128
+  ;CHECK: vpackuswb       %xmm1, %xmm0, %xmm0 
+  %res = call <16 x i8> @llvm.x86.avx512.mask.packuswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> zeroinitializer, i16 -1)
+  ret <16 x i8> %res
+}
+
+define <16 x i8> @test_mask_packus_epi16_rrk_128(<8 x i16> %a, <8 x i16> %b, <16 x i8> %passThru, i16 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi16_rrk_128
+  ;CHECK: vpackuswb       %xmm1, %xmm0, %xmm2 {%k1} 
+  %res = call <16 x i8> @llvm.x86.avx512.mask.packuswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> %passThru, i16 %mask)
+  ret <16 x i8> %res
+}
+
+define <16 x i8> @test_mask_packus_epi16_rrkz_128(<8 x i16> %a, <8 x i16> %b, i16 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi16_rrkz_128
+  ;CHECK: vpackuswb       %xmm1, %xmm0, %xmm0 {%k1} {z} 
+  %res = call <16 x i8> @llvm.x86.avx512.mask.packuswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> zeroinitializer, i16 %mask)
+  ret <16 x i8> %res
+}
+
+define <16 x i8> @test_mask_packus_epi16_rm_128(<8 x i16> %a, <8 x i16>* %ptr_b) {
+  ;CHECK-LABEL: test_mask_packus_epi16_rm_128
+  ;CHECK: vpackuswb       (%rdi), %xmm0, %xmm0 
+  %b = load <8 x i16>, <8 x i16>* %ptr_b
+  %res = call <16 x i8> @llvm.x86.avx512.mask.packuswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> zeroinitializer, i16 -1)
+  ret <16 x i8> %res
+}
+
+define <16 x i8> @test_mask_packus_epi16_rmk_128(<8 x i16> %a, <8 x i16>* %ptr_b, <16 x i8> %passThru, i16 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi16_rmk_128
+  ;CHECK: vpackuswb       (%rdi), %xmm0, %xmm1 {%k1} 
+  %b = load <8 x i16>, <8 x i16>* %ptr_b
+  %res = call <16 x i8> @llvm.x86.avx512.mask.packuswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> %passThru, i16 %mask)
+  ret <16 x i8> %res
+}
+
+define <16 x i8> @test_mask_packus_epi16_rmkz_128(<8 x i16> %a, <8 x i16>* %ptr_b, i16 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi16_rmkz_128
+  ;CHECK: vpackuswb       (%rdi), %xmm0, %xmm0 {%k1} {z} 
+  %b = load <8 x i16>, <8 x i16>* %ptr_b
+  %res = call <16 x i8> @llvm.x86.avx512.mask.packuswb.128(<8 x i16> %a, <8 x i16> %b, <16 x i8> zeroinitializer, i16 %mask)
+  ret <16 x i8> %res
+}
+
+declare <16 x i8> @llvm.x86.avx512.mask.packuswb.128(<8 x i16>, <8 x i16>, <16 x i8>, i16)
+
+define <32 x i8> @test_mask_packus_epi16_rr_256(<16 x i16> %a, <16 x i16> %b) {
+  ;CHECK-LABEL: test_mask_packus_epi16_rr_256
+  ;CHECK: vpackuswb       %ymm1, %ymm0, %ymm0 
+  %res = call <32 x i8> @llvm.x86.avx512.mask.packuswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> zeroinitializer, i32 -1)
+  ret <32 x i8> %res
+}
+
+define <32 x i8> @test_mask_packus_epi16_rrk_256(<16 x i16> %a, <16 x i16> %b, <32 x i8> %passThru, i32 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi16_rrk_256
+  ;CHECK: vpackuswb       %ymm1, %ymm0, %ymm2 {%k1} 
+  %res = call <32 x i8> @llvm.x86.avx512.mask.packuswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> %passThru, i32 %mask)
+  ret <32 x i8> %res
+}
+
+define <32 x i8> @test_mask_packus_epi16_rrkz_256(<16 x i16> %a, <16 x i16> %b, i32 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi16_rrkz_256
+  ;CHECK: vpackuswb       %ymm1, %ymm0, %ymm0 {%k1} {z} 
+  %res = call <32 x i8> @llvm.x86.avx512.mask.packuswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> zeroinitializer, i32 %mask)
+  ret <32 x i8> %res
+}
+
+define <32 x i8> @test_mask_packus_epi16_rm_256(<16 x i16> %a, <16 x i16>* %ptr_b) {
+  ;CHECK-LABEL: test_mask_packus_epi16_rm_256
+  ;CHECK: vpackuswb       (%rdi), %ymm0, %ymm0 
+  %b = load <16 x i16>, <16 x i16>* %ptr_b
+  %res = call <32 x i8> @llvm.x86.avx512.mask.packuswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> zeroinitializer, i32 -1)
+  ret <32 x i8> %res
+}
+
+define <32 x i8> @test_mask_packus_epi16_rmk_256(<16 x i16> %a, <16 x i16>* %ptr_b, <32 x i8> %passThru, i32 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi16_rmk_256
+  ;CHECK: vpackuswb       (%rdi), %ymm0, %ymm1 {%k1} 
+  %b = load <16 x i16>, <16 x i16>* %ptr_b
+  %res = call <32 x i8> @llvm.x86.avx512.mask.packuswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> %passThru, i32 %mask)
+  ret <32 x i8> %res
+}
+
+define <32 x i8> @test_mask_packus_epi16_rmkz_256(<16 x i16> %a, <16 x i16>* %ptr_b, i32 %mask) {
+  ;CHECK-LABEL: test_mask_packus_epi16_rmkz_256
+  ;CHECK: vpackuswb       (%rdi), %ymm0, %ymm0 {%k1} {z} 
+  %b = load <16 x i16>, <16 x i16>* %ptr_b
+  %res = call <32 x i8> @llvm.x86.avx512.mask.packuswb.256(<16 x i16> %a, <16 x i16> %b, <32 x i8> zeroinitializer, i32 %mask)
+  ret <32 x i8> %res
+}
+
+declare <32 x i8> @llvm.x86.avx512.mask.packuswb.256(<16 x i16>, <16 x i16>, <32 x i8>, i32)
+

diff --git a/test/MC/X86/x86-64-avx512bw.s b/test/MC/X86/x86-64-avx512bw.s
index 0d055b1cecd8b987b705c2f2522108e5c5e0c775..3f4855a66a78360e03f92c7f563e9660fe2ecda2 100644 (file)
--- a/test/MC/X86/x86-64-avx512bw.s
+++ b/test/MC/X86/x86-64-avx512bw.s
@@ -2008,3 +2008,330 @@
 // CHECK:  encoding: [0x62,0xe2,0x95,0x40,0x11,0x9a,0xc0,0xdf,0xff,0xff]
           vpsravw -8256(%rdx), %zmm29, %zmm19
 
+// CHECK: vpacksswb %zmm18, %zmm29, %zmm21
+// CHECK:  encoding: [0x62,0xa1,0x95,0x40,0x63,0xea]
+          vpacksswb %zmm18, %zmm29, %zmm21
+
+// CHECK: vpacksswb %zmm18, %zmm29, %zmm21 {%k1}
+// CHECK:  encoding: [0x62,0xa1,0x95,0x41,0x63,0xea]
+          vpacksswb %zmm18, %zmm29, %zmm21 {%k1}
+
+// CHECK: vpacksswb %zmm18, %zmm29, %zmm21 {%k1} {z}
+// CHECK:  encoding: [0x62,0xa1,0x95,0xc1,0x63,0xea]
+          vpacksswb %zmm18, %zmm29, %zmm21 {%k1} {z}
+
+// CHECK: vpacksswb (%rcx), %zmm29, %zmm21
+// CHECK:  encoding: [0x62,0xe1,0x95,0x40,0x63,0x29]
+          vpacksswb (%rcx), %zmm29, %zmm21
+
+// CHECK: vpacksswb 291(%rax,%r14,8), %zmm29, %zmm21
+// CHECK:  encoding: [0x62,0xa1,0x95,0x40,0x63,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vpacksswb 291(%rax,%r14,8), %zmm29, %zmm21
+
+// CHECK: vpacksswb 8128(%rdx), %zmm29, %zmm21
+// CHECK:  encoding: [0x62,0xe1,0x95,0x40,0x63,0x6a,0x7f]
+          vpacksswb 8128(%rdx), %zmm29, %zmm21
+
+// CHECK: vpacksswb 8192(%rdx), %zmm29, %zmm21
+// CHECK:  encoding: [0x62,0xe1,0x95,0x40,0x63,0xaa,0x00,0x20,0x00,0x00]
+          vpacksswb 8192(%rdx), %zmm29, %zmm21
+
+// CHECK: vpacksswb -8192(%rdx), %zmm29, %zmm21
+// CHECK:  encoding: [0x62,0xe1,0x95,0x40,0x63,0x6a,0x80]
+          vpacksswb -8192(%rdx), %zmm29, %zmm21
+
+// CHECK: vpacksswb -8256(%rdx), %zmm29, %zmm21
+// CHECK:  encoding: [0x62,0xe1,0x95,0x40,0x63,0xaa,0xc0,0xdf,0xff,0xff]
+          vpacksswb -8256(%rdx), %zmm29, %zmm21
+
+// CHECK: vpackuswb %zmm28, %zmm24, %zmm28
+// CHECK:  encoding: [0x62,0x01,0xbd,0x40,0x67,0xe4]
+          vpackuswb %zmm28, %zmm24, %zmm28
+
+// CHECK: vpackuswb %zmm28, %zmm24, %zmm28 {%k6}
+// CHECK:  encoding: [0x62,0x01,0xbd,0x46,0x67,0xe4]
+          vpackuswb %zmm28, %zmm24, %zmm28 {%k6}
+
+// CHECK: vpackuswb %zmm28, %zmm24, %zmm28 {%k6} {z}
+// CHECK:  encoding: [0x62,0x01,0xbd,0xc6,0x67,0xe4]
+          vpackuswb %zmm28, %zmm24, %zmm28 {%k6} {z}
+
+// CHECK: vpackuswb (%rcx), %zmm24, %zmm28
+// CHECK:  encoding: [0x62,0x61,0xbd,0x40,0x67,0x21]
+          vpackuswb (%rcx), %zmm24, %zmm28
+
+// CHECK: vpackuswb 291(%rax,%r14,8), %zmm24, %zmm28
+// CHECK:  encoding: [0x62,0x21,0xbd,0x40,0x67,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpackuswb 291(%rax,%r14,8), %zmm24, %zmm28
+
+// CHECK: vpackuswb 8128(%rdx), %zmm24, %zmm28
+// CHECK:  encoding: [0x62,0x61,0xbd,0x40,0x67,0x62,0x7f]
+          vpackuswb 8128(%rdx), %zmm24, %zmm28
+
+// CHECK: vpackuswb 8192(%rdx), %zmm24, %zmm28
+// CHECK:  encoding: [0x62,0x61,0xbd,0x40,0x67,0xa2,0x00,0x20,0x00,0x00]
+          vpackuswb 8192(%rdx), %zmm24, %zmm28
+
+// CHECK: vpackuswb -8192(%rdx), %zmm24, %zmm28
+// CHECK:  encoding: [0x62,0x61,0xbd,0x40,0x67,0x62,0x80]
+          vpackuswb -8192(%rdx), %zmm24, %zmm28
+
+// CHECK: vpackuswb -8256(%rdx), %zmm24, %zmm28
+// CHECK:  encoding: [0x62,0x61,0xbd,0x40,0x67,0xa2,0xc0,0xdf,0xff,0xff]
+          vpackuswb -8256(%rdx), %zmm24, %zmm28
+
+// CHECK: vpacksswb %zmm24, %zmm29, %zmm26
+// CHECK:  encoding: [0x62,0x01,0x95,0x40,0x63,0xd0]
+          vpacksswb %zmm24, %zmm29, %zmm26
+
+// CHECK: vpacksswb %zmm24, %zmm29, %zmm26 {%k6}
+// CHECK:  encoding: [0x62,0x01,0x95,0x46,0x63,0xd0]
+          vpacksswb %zmm24, %zmm29, %zmm26 {%k6}
+
+// CHECK: vpacksswb %zmm24, %zmm29, %zmm26 {%k6} {z}
+// CHECK:  encoding: [0x62,0x01,0x95,0xc6,0x63,0xd0]
+          vpacksswb %zmm24, %zmm29, %zmm26 {%k6} {z}
+
+// CHECK: vpacksswb (%rcx), %zmm29, %zmm26
+// CHECK:  encoding: [0x62,0x61,0x95,0x40,0x63,0x11]
+          vpacksswb (%rcx), %zmm29, %zmm26
+
+// CHECK: vpacksswb 4660(%rax,%r14,8), %zmm29, %zmm26
+// CHECK:  encoding: [0x62,0x21,0x95,0x40,0x63,0x94,0xf0,0x34,0x12,0x00,0x00]
+          vpacksswb 4660(%rax,%r14,8), %zmm29, %zmm26
+
+// CHECK: vpacksswb 8128(%rdx), %zmm29, %zmm26
+// CHECK:  encoding: [0x62,0x61,0x95,0x40,0x63,0x52,0x7f]
+          vpacksswb 8128(%rdx), %zmm29, %zmm26
+
+// CHECK: vpacksswb 8192(%rdx), %zmm29, %zmm26
+// CHECK:  encoding: [0x62,0x61,0x95,0x40,0x63,0x92,0x00,0x20,0x00,0x00]
+          vpacksswb 8192(%rdx), %zmm29, %zmm26
+
+// CHECK: vpacksswb -8192(%rdx), %zmm29, %zmm26
+// CHECK:  encoding: [0x62,0x61,0x95,0x40,0x63,0x52,0x80]
+          vpacksswb -8192(%rdx), %zmm29, %zmm26
+
+// CHECK: vpacksswb -8256(%rdx), %zmm29, %zmm26
+// CHECK:  encoding: [0x62,0x61,0x95,0x40,0x63,0x92,0xc0,0xdf,0xff,0xff]
+          vpacksswb -8256(%rdx), %zmm29, %zmm26
+
+// CHECK: vpackuswb %zmm23, %zmm28, %zmm18
+// CHECK:  encoding: [0x62,0xa1,0x9d,0x40,0x67,0xd7]
+          vpackuswb %zmm23, %zmm28, %zmm18
+
+// CHECK: vpackuswb %zmm23, %zmm28, %zmm18 {%k4}
+// CHECK:  encoding: [0x62,0xa1,0x9d,0x44,0x67,0xd7]
+          vpackuswb %zmm23, %zmm28, %zmm18 {%k4}
+
+// CHECK: vpackuswb %zmm23, %zmm28, %zmm18 {%k4} {z}
+// CHECK:  encoding: [0x62,0xa1,0x9d,0xc4,0x67,0xd7]
+          vpackuswb %zmm23, %zmm28, %zmm18 {%k4} {z}
+
+// CHECK: vpackuswb (%rcx), %zmm28, %zmm18
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x40,0x67,0x11]
+          vpackuswb (%rcx), %zmm28, %zmm18
+
+// CHECK: vpackuswb 4660(%rax,%r14,8), %zmm28, %zmm18
+// CHECK:  encoding: [0x62,0xa1,0x9d,0x40,0x67,0x94,0xf0,0x34,0x12,0x00,0x00]
+          vpackuswb 4660(%rax,%r14,8), %zmm28, %zmm18
+
+// CHECK: vpackuswb 8128(%rdx), %zmm28, %zmm18
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x40,0x67,0x52,0x7f]
+          vpackuswb 8128(%rdx), %zmm28, %zmm18
+
+// CHECK: vpackuswb 8192(%rdx), %zmm28, %zmm18
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x40,0x67,0x92,0x00,0x20,0x00,0x00]
+          vpackuswb 8192(%rdx), %zmm28, %zmm18
+
+// CHECK: vpackuswb -8192(%rdx), %zmm28, %zmm18
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x40,0x67,0x52,0x80]
+          vpackuswb -8192(%rdx), %zmm28, %zmm18
+
+// CHECK: vpackuswb -8256(%rdx), %zmm28, %zmm18
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x40,0x67,0x92,0xc0,0xdf,0xff,0xff]
+          vpackuswb -8256(%rdx), %zmm28, %zmm18
+
+// CHECK: vpackssdw %zmm22, %zmm27, %zmm24
+// CHECK:  encoding: [0x62,0x21,0x25,0x40,0x6b,0xc6]
+          vpackssdw %zmm22, %zmm27, %zmm24
+
+// CHECK: vpackssdw %zmm22, %zmm27, %zmm24 {%k2}
+// CHECK:  encoding: [0x62,0x21,0x25,0x42,0x6b,0xc6]
+          vpackssdw %zmm22, %zmm27, %zmm24 {%k2}
+
+// CHECK: vpackssdw %zmm22, %zmm27, %zmm24 {%k2} {z}
+// CHECK:  encoding: [0x62,0x21,0x25,0xc2,0x6b,0xc6]
+          vpackssdw %zmm22, %zmm27, %zmm24 {%k2} {z}
+
+// CHECK: vpackssdw (%rcx), %zmm27, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x25,0x40,0x6b,0x01]
+          vpackssdw (%rcx), %zmm27, %zmm24
+
+// CHECK: vpackssdw 291(%rax,%r14,8), %zmm27, %zmm24
+// CHECK:  encoding: [0x62,0x21,0x25,0x40,0x6b,0x84,0xf0,0x23,0x01,0x00,0x00]
+          vpackssdw 291(%rax,%r14,8), %zmm27, %zmm24
+
+// CHECK: vpackssdw (%rcx){1to16}, %zmm27, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x25,0x50,0x6b,0x01]
+          vpackssdw (%rcx){1to16}, %zmm27, %zmm24
+
+// CHECK: vpackssdw 8128(%rdx), %zmm27, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x25,0x40,0x6b,0x42,0x7f]
+          vpackssdw 8128(%rdx), %zmm27, %zmm24
+
+// CHECK: vpackssdw 8192(%rdx), %zmm27, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x25,0x40,0x6b,0x82,0x00,0x20,0x00,0x00]
+          vpackssdw 8192(%rdx), %zmm27, %zmm24
+
+// CHECK: vpackssdw -8192(%rdx), %zmm27, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x25,0x40,0x6b,0x42,0x80]
+          vpackssdw -8192(%rdx), %zmm27, %zmm24
+
+// CHECK: vpackssdw -8256(%rdx), %zmm27, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x25,0x40,0x6b,0x82,0xc0,0xdf,0xff,0xff]
+          vpackssdw -8256(%rdx), %zmm27, %zmm24
+
+// CHECK: vpackssdw 508(%rdx){1to16}, %zmm27, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x25,0x50,0x6b,0x42,0x7f]
+          vpackssdw 508(%rdx){1to16}, %zmm27, %zmm24
+
+// CHECK: vpackssdw 512(%rdx){1to16}, %zmm27, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x25,0x50,0x6b,0x82,0x00,0x02,0x00,0x00]
+          vpackssdw 512(%rdx){1to16}, %zmm27, %zmm24
+
+// CHECK: vpackssdw -512(%rdx){1to16}, %zmm27, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x25,0x50,0x6b,0x42,0x80]
+          vpackssdw -512(%rdx){1to16}, %zmm27, %zmm24
+
+// CHECK: vpackssdw -516(%rdx){1to16}, %zmm27, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x25,0x50,0x6b,0x82,0xfc,0xfd,0xff,0xff]
+          vpackssdw -516(%rdx){1to16}, %zmm27, %zmm24
+
+// CHECK: vpacksswb %zmm23, %zmm18, %zmm28
+// CHECK:  encoding: [0x62,0x21,0xed,0x40,0x63,0xe7]
+          vpacksswb %zmm23, %zmm18, %zmm28
+
+// CHECK: vpacksswb %zmm23, %zmm18, %zmm28 {%k3}
+// CHECK:  encoding: [0x62,0x21,0xed,0x43,0x63,0xe7]
+          vpacksswb %zmm23, %zmm18, %zmm28 {%k3}
+
+// CHECK: vpacksswb %zmm23, %zmm18, %zmm28 {%k3} {z}
+// CHECK:  encoding: [0x62,0x21,0xed,0xc3,0x63,0xe7]
+          vpacksswb %zmm23, %zmm18, %zmm28 {%k3} {z}
+
+// CHECK: vpacksswb (%rcx), %zmm18, %zmm28
+// CHECK:  encoding: [0x62,0x61,0xed,0x40,0x63,0x21]
+          vpacksswb (%rcx), %zmm18, %zmm28
+
+// CHECK: vpacksswb 291(%rax,%r14,8), %zmm18, %zmm28
+// CHECK:  encoding: [0x62,0x21,0xed,0x40,0x63,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpacksswb 291(%rax,%r14,8), %zmm18, %zmm28
+
+// CHECK: vpacksswb 8128(%rdx), %zmm18, %zmm28
+// CHECK:  encoding: [0x62,0x61,0xed,0x40,0x63,0x62,0x7f]
+          vpacksswb 8128(%rdx), %zmm18, %zmm28
+
+// CHECK: vpacksswb 8192(%rdx), %zmm18, %zmm28
+// CHECK:  encoding: [0x62,0x61,0xed,0x40,0x63,0xa2,0x00,0x20,0x00,0x00]
+          vpacksswb 8192(%rdx), %zmm18, %zmm28
+
+// CHECK: vpacksswb -8192(%rdx), %zmm18, %zmm28
+// CHECK:  encoding: [0x62,0x61,0xed,0x40,0x63,0x62,0x80]
+          vpacksswb -8192(%rdx), %zmm18, %zmm28
+
+// CHECK: vpacksswb -8256(%rdx), %zmm18, %zmm28
+// CHECK:  encoding: [0x62,0x61,0xed,0x40,0x63,0xa2,0xc0,0xdf,0xff,0xff]
+          vpacksswb -8256(%rdx), %zmm18, %zmm28
+
+// CHECK: vpackusdw %zmm21, %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xa2,0x15,0x40,0x2b,0xd5]
+          vpackusdw %zmm21, %zmm29, %zmm18
+
+// CHECK: vpackusdw %zmm21, %zmm29, %zmm18 {%k2}
+// CHECK:  encoding: [0x62,0xa2,0x15,0x42,0x2b,0xd5]
+          vpackusdw %zmm21, %zmm29, %zmm18 {%k2}
+
+// CHECK: vpackusdw %zmm21, %zmm29, %zmm18 {%k2} {z}
+// CHECK:  encoding: [0x62,0xa2,0x15,0xc2,0x2b,0xd5]
+          vpackusdw %zmm21, %zmm29, %zmm18 {%k2} {z}
+
+// CHECK: vpackusdw (%rcx), %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0x15,0x40,0x2b,0x11]
+          vpackusdw (%rcx), %zmm29, %zmm18
+
+// CHECK: vpackusdw 291(%rax,%r14,8), %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xa2,0x15,0x40,0x2b,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpackusdw 291(%rax,%r14,8), %zmm29, %zmm18
+
+// CHECK: vpackusdw (%rcx){1to16}, %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0x15,0x50,0x2b,0x11]
+          vpackusdw (%rcx){1to16}, %zmm29, %zmm18
+
+// CHECK: vpackusdw 8128(%rdx), %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0x15,0x40,0x2b,0x52,0x7f]
+          vpackusdw 8128(%rdx), %zmm29, %zmm18
+
+// CHECK: vpackusdw 8192(%rdx), %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0x15,0x40,0x2b,0x92,0x00,0x20,0x00,0x00]
+          vpackusdw 8192(%rdx), %zmm29, %zmm18
+
+// CHECK: vpackusdw -8192(%rdx), %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0x15,0x40,0x2b,0x52,0x80]
+          vpackusdw -8192(%rdx), %zmm29, %zmm18
+
+// CHECK: vpackusdw -8256(%rdx), %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0x15,0x40,0x2b,0x92,0xc0,0xdf,0xff,0xff]
+          vpackusdw -8256(%rdx), %zmm29, %zmm18
+
+// CHECK: vpackusdw 508(%rdx){1to16}, %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0x15,0x50,0x2b,0x52,0x7f]
+          vpackusdw 508(%rdx){1to16}, %zmm29, %zmm18
+
+// CHECK: vpackusdw 512(%rdx){1to16}, %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0x15,0x50,0x2b,0x92,0x00,0x02,0x00,0x00]
+          vpackusdw 512(%rdx){1to16}, %zmm29, %zmm18
+
+// CHECK: vpackusdw -512(%rdx){1to16}, %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0x15,0x50,0x2b,0x52,0x80]
+          vpackusdw -512(%rdx){1to16}, %zmm29, %zmm18
+
+// CHECK: vpackusdw -516(%rdx){1to16}, %zmm29, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0x15,0x50,0x2b,0x92,0xfc,0xfd,0xff,0xff]
+          vpackusdw -516(%rdx){1to16}, %zmm29, %zmm18
+
+// CHECK: vpackuswb %zmm27, %zmm18, %zmm23
+// CHECK:  encoding: [0x62,0x81,0xed,0x40,0x67,0xfb]
+          vpackuswb %zmm27, %zmm18, %zmm23
+
+// CHECK: vpackuswb %zmm27, %zmm18, %zmm23 {%k5}
+// CHECK:  encoding: [0x62,0x81,0xed,0x45,0x67,0xfb]
+          vpackuswb %zmm27, %zmm18, %zmm23 {%k5}
+
+// CHECK: vpackuswb %zmm27, %zmm18, %zmm23 {%k5} {z}
+// CHECK:  encoding: [0x62,0x81,0xed,0xc5,0x67,0xfb]
+          vpackuswb %zmm27, %zmm18, %zmm23 {%k5} {z}
+
+// CHECK: vpackuswb (%rcx), %zmm18, %zmm23
+// CHECK:  encoding: [0x62,0xe1,0xed,0x40,0x67,0x39]
+          vpackuswb (%rcx), %zmm18, %zmm23
+
+// CHECK: vpackuswb 291(%rax,%r14,8), %zmm18, %zmm23
+// CHECK:  encoding: [0x62,0xa1,0xed,0x40,0x67,0xbc,0xf0,0x23,0x01,0x00,0x00]
+          vpackuswb 291(%rax,%r14,8), %zmm18, %zmm23
+
+// CHECK: vpackuswb 8128(%rdx), %zmm18, %zmm23
+// CHECK:  encoding: [0x62,0xe1,0xed,0x40,0x67,0x7a,0x7f]
+          vpackuswb 8128(%rdx), %zmm18, %zmm23
+
+// CHECK: vpackuswb 8192(%rdx), %zmm18, %zmm23
+// CHECK:  encoding: [0x62,0xe1,0xed,0x40,0x67,0xba,0x00,0x20,0x00,0x00]
+          vpackuswb 8192(%rdx), %zmm18, %zmm23
+
+// CHECK: vpackuswb -8192(%rdx), %zmm18, %zmm23
+// CHECK:  encoding: [0x62,0xe1,0xed,0x40,0x67,0x7a,0x80]
+          vpackuswb -8192(%rdx), %zmm18, %zmm23
+
+// CHECK: vpackuswb -8256(%rdx), %zmm18, %zmm23
+// CHECK:  encoding: [0x62,0xe1,0xed,0x40,0x67,0xba,0xc0,0xdf,0xff,0xff]
+          vpackuswb -8256(%rdx), %zmm18, %zmm23

diff --git a/test/MC/X86/x86-64-avx512bw_vl.s b/test/MC/X86/x86-64-avx512bw_vl.s
index 3f4558271662ce73924a7179d03f72187577d7f5..b3883c8d1782960e6e73812bec0ccd3aebe6c020 100644 (file)
--- a/test/MC/X86/x86-64-avx512bw_vl.s
+++ b/test/MC/X86/x86-64-avx512bw_vl.s
@@ -2303,3 +2303,1027 @@
 // CHECK: vpsravw -4128(%rdx), %ymm28, %ymm20
 // CHECK:  encoding: [0x62,0xe2,0x9d,0x20,0x11,0xa2,0xe0,0xef,0xff,0xff]
           vpsravw -4128(%rdx), %ymm28, %ymm20
+
+// CHECK: vpacksswb %xmm20, %xmm29, %xmm17
+// CHECK:  encoding: [0x62,0xa1,0x95,0x00,0x63,0xcc]
+          vpacksswb %xmm20, %xmm29, %xmm17
+
+// CHECK: vpacksswb %xmm20, %xmm29, %xmm17 {%k3}
+// CHECK:  encoding: [0x62,0xa1,0x95,0x03,0x63,0xcc]
+          vpacksswb %xmm20, %xmm29, %xmm17 {%k3}
+
+// CHECK: vpacksswb %xmm20, %xmm29, %xmm17 {%k3} {z}
+// CHECK:  encoding: [0x62,0xa1,0x95,0x83,0x63,0xcc]
+          vpacksswb %xmm20, %xmm29, %xmm17 {%k3} {z}
+
+// CHECK: vpacksswb (%rcx), %xmm29, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0x95,0x00,0x63,0x09]
+          vpacksswb (%rcx), %xmm29, %xmm17
+
+// CHECK: vpacksswb 291(%rax,%r14,8), %xmm29, %xmm17
+// CHECK:  encoding: [0x62,0xa1,0x95,0x00,0x63,0x8c,0xf0,0x23,0x01,0x00,0x00]
+          vpacksswb 291(%rax,%r14,8), %xmm29, %xmm17
+
+// CHECK: vpacksswb 2032(%rdx), %xmm29, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0x95,0x00,0x63,0x4a,0x7f]
+          vpacksswb 2032(%rdx), %xmm29, %xmm17
+
+// CHECK: vpacksswb 2048(%rdx), %xmm29, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0x95,0x00,0x63,0x8a,0x00,0x08,0x00,0x00]
+          vpacksswb 2048(%rdx), %xmm29, %xmm17
+
+// CHECK: vpacksswb -2048(%rdx), %xmm29, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0x95,0x00,0x63,0x4a,0x80]
+          vpacksswb -2048(%rdx), %xmm29, %xmm17
+
+// CHECK: vpacksswb -2064(%rdx), %xmm29, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0x95,0x00,0x63,0x8a,0xf0,0xf7,0xff,0xff]
+          vpacksswb -2064(%rdx), %xmm29, %xmm17
+
+// CHECK: vpacksswb %ymm21, %ymm26, %ymm29
+// CHECK:  encoding: [0x62,0x21,0xad,0x20,0x63,0xed]
+          vpacksswb %ymm21, %ymm26, %ymm29
+
+// CHECK: vpacksswb %ymm21, %ymm26, %ymm29 {%k2}
+// CHECK:  encoding: [0x62,0x21,0xad,0x22,0x63,0xed]
+          vpacksswb %ymm21, %ymm26, %ymm29 {%k2}
+
+// CHECK: vpacksswb %ymm21, %ymm26, %ymm29 {%k2} {z}
+// CHECK:  encoding: [0x62,0x21,0xad,0xa2,0x63,0xed]
+          vpacksswb %ymm21, %ymm26, %ymm29 {%k2} {z}
+
+// CHECK: vpacksswb (%rcx), %ymm26, %ymm29
+// CHECK:  encoding: [0x62,0x61,0xad,0x20,0x63,0x29]
+          vpacksswb (%rcx), %ymm26, %ymm29
+
+// CHECK: vpacksswb 291(%rax,%r14,8), %ymm26, %ymm29
+// CHECK:  encoding: [0x62,0x21,0xad,0x20,0x63,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vpacksswb 291(%rax,%r14,8), %ymm26, %ymm29
+
+// CHECK: vpacksswb 4064(%rdx), %ymm26, %ymm29
+// CHECK:  encoding: [0x62,0x61,0xad,0x20,0x63,0x6a,0x7f]
+          vpacksswb 4064(%rdx), %ymm26, %ymm29
+
+// CHECK: vpacksswb 4096(%rdx), %ymm26, %ymm29
+// CHECK:  encoding: [0x62,0x61,0xad,0x20,0x63,0xaa,0x00,0x10,0x00,0x00]
+          vpacksswb 4096(%rdx), %ymm26, %ymm29
+
+// CHECK: vpacksswb -4096(%rdx), %ymm26, %ymm29
+// CHECK:  encoding: [0x62,0x61,0xad,0x20,0x63,0x6a,0x80]
+          vpacksswb -4096(%rdx), %ymm26, %ymm29
+
+// CHECK: vpacksswb -4128(%rdx), %ymm26, %ymm29
+// CHECK:  encoding: [0x62,0x61,0xad,0x20,0x63,0xaa,0xe0,0xef,0xff,0xff]
+          vpacksswb -4128(%rdx), %ymm26, %ymm29
+
+// CHECK: vpackuswb %xmm24, %xmm26, %xmm20
+// CHECK:  encoding: [0x62,0x81,0xad,0x00,0x67,0xe0]
+          vpackuswb %xmm24, %xmm26, %xmm20
+
+// CHECK: vpackuswb %xmm24, %xmm26, %xmm20 {%k4}
+// CHECK:  encoding: [0x62,0x81,0xad,0x04,0x67,0xe0]
+          vpackuswb %xmm24, %xmm26, %xmm20 {%k4}
+
+// CHECK: vpackuswb %xmm24, %xmm26, %xmm20 {%k4} {z}
+// CHECK:  encoding: [0x62,0x81,0xad,0x84,0x67,0xe0]
+          vpackuswb %xmm24, %xmm26, %xmm20 {%k4} {z}
+
+// CHECK: vpackuswb (%rcx), %xmm26, %xmm20
+// CHECK:  encoding: [0x62,0xe1,0xad,0x00,0x67,0x21]
+          vpackuswb (%rcx), %xmm26, %xmm20
+
+// CHECK: vpackuswb 291(%rax,%r14,8), %xmm26, %xmm20
+// CHECK:  encoding: [0x62,0xa1,0xad,0x00,0x67,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpackuswb 291(%rax,%r14,8), %xmm26, %xmm20
+
+// CHECK: vpackuswb 2032(%rdx), %xmm26, %xmm20
+// CHECK:  encoding: [0x62,0xe1,0xad,0x00,0x67,0x62,0x7f]
+          vpackuswb 2032(%rdx), %xmm26, %xmm20
+
+// CHECK: vpackuswb 2048(%rdx), %xmm26, %xmm20
+// CHECK:  encoding: [0x62,0xe1,0xad,0x00,0x67,0xa2,0x00,0x08,0x00,0x00]
+          vpackuswb 2048(%rdx), %xmm26, %xmm20
+
+// CHECK: vpackuswb -2048(%rdx), %xmm26, %xmm20
+// CHECK:  encoding: [0x62,0xe1,0xad,0x00,0x67,0x62,0x80]
+          vpackuswb -2048(%rdx), %xmm26, %xmm20
+
+// CHECK: vpackuswb -2064(%rdx), %xmm26, %xmm20
+// CHECK:  encoding: [0x62,0xe1,0xad,0x00,0x67,0xa2,0xf0,0xf7,0xff,0xff]
+          vpackuswb -2064(%rdx), %xmm26, %xmm20
+
+// CHECK: vpackuswb %ymm20, %ymm28, %ymm19
+// CHECK:  encoding: [0x62,0xa1,0x9d,0x20,0x67,0xdc]
+          vpackuswb %ymm20, %ymm28, %ymm19
+
+// CHECK: vpackuswb %ymm20, %ymm28, %ymm19 {%k2}
+// CHECK:  encoding: [0x62,0xa1,0x9d,0x22,0x67,0xdc]
+          vpackuswb %ymm20, %ymm28, %ymm19 {%k2}
+
+// CHECK: vpackuswb %ymm20, %ymm28, %ymm19 {%k2} {z}
+// CHECK:  encoding: [0x62,0xa1,0x9d,0xa2,0x67,0xdc]
+          vpackuswb %ymm20, %ymm28, %ymm19 {%k2} {z}
+
+// CHECK: vpackuswb (%rcx), %ymm28, %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x20,0x67,0x19]
+          vpackuswb (%rcx), %ymm28, %ymm19
+
+// CHECK: vpackuswb 291(%rax,%r14,8), %ymm28, %ymm19
+// CHECK:  encoding: [0x62,0xa1,0x9d,0x20,0x67,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vpackuswb 291(%rax,%r14,8), %ymm28, %ymm19
+
+// CHECK: vpackuswb 4064(%rdx), %ymm28, %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x20,0x67,0x5a,0x7f]
+          vpackuswb 4064(%rdx), %ymm28, %ymm19
+
+// CHECK: vpackuswb 4096(%rdx), %ymm28, %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x20,0x67,0x9a,0x00,0x10,0x00,0x00]
+          vpackuswb 4096(%rdx), %ymm28, %ymm19
+
+// CHECK: vpackuswb -4096(%rdx), %ymm28, %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x20,0x67,0x5a,0x80]
+          vpackuswb -4096(%rdx), %ymm28, %ymm19
+
+// CHECK: vpackuswb -4128(%rdx), %ymm28, %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x20,0x67,0x9a,0xe0,0xef,0xff,0xff]
+          vpackuswb -4128(%rdx), %ymm28, %ymm19
+
+// CHECK: vpacksswb %xmm25, %xmm18, %xmm19
+// CHECK:  encoding: [0x62,0x81,0xed,0x00,0x63,0xd9]
+          vpacksswb %xmm25, %xmm18, %xmm19
+
+// CHECK: vpacksswb %xmm25, %xmm18, %xmm19 {%k4}
+// CHECK:  encoding: [0x62,0x81,0xed,0x04,0x63,0xd9]
+          vpacksswb %xmm25, %xmm18, %xmm19 {%k4}
+
+// CHECK: vpacksswb %xmm25, %xmm18, %xmm19 {%k4} {z}
+// CHECK:  encoding: [0x62,0x81,0xed,0x84,0x63,0xd9]
+          vpacksswb %xmm25, %xmm18, %xmm19 {%k4} {z}
+
+// CHECK: vpacksswb (%rcx), %xmm18, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xed,0x00,0x63,0x19]
+          vpacksswb (%rcx), %xmm18, %xmm19
+
+// CHECK: vpacksswb 4660(%rax,%r14,8), %xmm18, %xmm19
+// CHECK:  encoding: [0x62,0xa1,0xed,0x00,0x63,0x9c,0xf0,0x34,0x12,0x00,0x00]
+          vpacksswb 4660(%rax,%r14,8), %xmm18, %xmm19
+
+// CHECK: vpacksswb 2032(%rdx), %xmm18, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xed,0x00,0x63,0x5a,0x7f]
+          vpacksswb 2032(%rdx), %xmm18, %xmm19
+
+// CHECK: vpacksswb 2048(%rdx), %xmm18, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xed,0x00,0x63,0x9a,0x00,0x08,0x00,0x00]
+          vpacksswb 2048(%rdx), %xmm18, %xmm19
+
+// CHECK: vpacksswb -2048(%rdx), %xmm18, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xed,0x00,0x63,0x5a,0x80]
+          vpacksswb -2048(%rdx), %xmm18, %xmm19
+
+// CHECK: vpacksswb -2064(%rdx), %xmm18, %xmm19
+// CHECK:  encoding: [0x62,0xe1,0xed,0x00,0x63,0x9a,0xf0,0xf7,0xff,0xff]
+          vpacksswb -2064(%rdx), %xmm18, %xmm19
+
+// CHECK: vpacksswb %ymm19, %ymm22, %ymm22
+// CHECK:  encoding: [0x62,0xa1,0xcd,0x20,0x63,0xf3]
+          vpacksswb %ymm19, %ymm22, %ymm22
+
+// CHECK: vpacksswb %ymm19, %ymm22, %ymm22 {%k7}
+// CHECK:  encoding: [0x62,0xa1,0xcd,0x27,0x63,0xf3]
+          vpacksswb %ymm19, %ymm22, %ymm22 {%k7}
+
+// CHECK: vpacksswb %ymm19, %ymm22, %ymm22 {%k7} {z}
+// CHECK:  encoding: [0x62,0xa1,0xcd,0xa7,0x63,0xf3]
+          vpacksswb %ymm19, %ymm22, %ymm22 {%k7} {z}
+
+// CHECK: vpacksswb (%rcx), %ymm22, %ymm22
+// CHECK:  encoding: [0x62,0xe1,0xcd,0x20,0x63,0x31]
+          vpacksswb (%rcx), %ymm22, %ymm22
+
+// CHECK: vpacksswb 4660(%rax,%r14,8), %ymm22, %ymm22
+// CHECK:  encoding: [0x62,0xa1,0xcd,0x20,0x63,0xb4,0xf0,0x34,0x12,0x00,0x00]
+          vpacksswb 4660(%rax,%r14,8), %ymm22, %ymm22
+
+// CHECK: vpacksswb 4064(%rdx), %ymm22, %ymm22
+// CHECK:  encoding: [0x62,0xe1,0xcd,0x20,0x63,0x72,0x7f]
+          vpacksswb 4064(%rdx), %ymm22, %ymm22
+
+// CHECK: vpacksswb 4096(%rdx), %ymm22, %ymm22
+// CHECK:  encoding: [0x62,0xe1,0xcd,0x20,0x63,0xb2,0x00,0x10,0x00,0x00]
+          vpacksswb 4096(%rdx), %ymm22, %ymm22
+
+// CHECK: vpacksswb -4096(%rdx), %ymm22, %ymm22
+// CHECK:  encoding: [0x62,0xe1,0xcd,0x20,0x63,0x72,0x80]
+          vpacksswb -4096(%rdx), %ymm22, %ymm22
+
+// CHECK: vpacksswb -4128(%rdx), %ymm22, %ymm22
+// CHECK:  encoding: [0x62,0xe1,0xcd,0x20,0x63,0xb2,0xe0,0xef,0xff,0xff]
+          vpacksswb -4128(%rdx), %ymm22, %ymm22
+
+// CHECK: vpackuswb %xmm23, %xmm28, %xmm23
+// CHECK:  encoding: [0x62,0xa1,0x9d,0x00,0x67,0xff]
+          vpackuswb %xmm23, %xmm28, %xmm23
+
+// CHECK: vpackuswb %xmm23, %xmm28, %xmm23 {%k3}
+// CHECK:  encoding: [0x62,0xa1,0x9d,0x03,0x67,0xff]
+          vpackuswb %xmm23, %xmm28, %xmm23 {%k3}
+
+// CHECK: vpackuswb %xmm23, %xmm28, %xmm23 {%k3} {z}
+// CHECK:  encoding: [0x62,0xa1,0x9d,0x83,0x67,0xff]
+          vpackuswb %xmm23, %xmm28, %xmm23 {%k3} {z}
+
+// CHECK: vpackuswb (%rcx), %xmm28, %xmm23
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x00,0x67,0x39]
+          vpackuswb (%rcx), %xmm28, %xmm23
+
+// CHECK: vpackuswb 4660(%rax,%r14,8), %xmm28, %xmm23
+// CHECK:  encoding: [0x62,0xa1,0x9d,0x00,0x67,0xbc,0xf0,0x34,0x12,0x00,0x00]
+          vpackuswb 4660(%rax,%r14,8), %xmm28, %xmm23
+
+// CHECK: vpackuswb 2032(%rdx), %xmm28, %xmm23
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x00,0x67,0x7a,0x7f]
+          vpackuswb 2032(%rdx), %xmm28, %xmm23
+
+// CHECK: vpackuswb 2048(%rdx), %xmm28, %xmm23
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x00,0x67,0xba,0x00,0x08,0x00,0x00]
+          vpackuswb 2048(%rdx), %xmm28, %xmm23
+
+// CHECK: vpackuswb -2048(%rdx), %xmm28, %xmm23
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x00,0x67,0x7a,0x80]
+          vpackuswb -2048(%rdx), %xmm28, %xmm23
+
+// CHECK: vpackuswb -2064(%rdx), %xmm28, %xmm23
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x00,0x67,0xba,0xf0,0xf7,0xff,0xff]
+          vpackuswb -2064(%rdx), %xmm28, %xmm23
+
+// CHECK: vpackuswb %ymm27, %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0x81,0xad,0x20,0x67,0xe3]
+          vpackuswb %ymm27, %ymm26, %ymm20
+
+// CHECK: vpackuswb %ymm27, %ymm26, %ymm20 {%k6}
+// CHECK:  encoding: [0x62,0x81,0xad,0x26,0x67,0xe3]
+          vpackuswb %ymm27, %ymm26, %ymm20 {%k6}
+
+// CHECK: vpackuswb %ymm27, %ymm26, %ymm20 {%k6} {z}
+// CHECK:  encoding: [0x62,0x81,0xad,0xa6,0x67,0xe3]
+          vpackuswb %ymm27, %ymm26, %ymm20 {%k6} {z}
+
+// CHECK: vpackuswb (%rcx), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xad,0x20,0x67,0x21]
+          vpackuswb (%rcx), %ymm26, %ymm20
+
+// CHECK: vpackuswb 4660(%rax,%r14,8), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xa1,0xad,0x20,0x67,0xa4,0xf0,0x34,0x12,0x00,0x00]
+          vpackuswb 4660(%rax,%r14,8), %ymm26, %ymm20
+
+// CHECK: vpackuswb 4064(%rdx), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xad,0x20,0x67,0x62,0x7f]
+          vpackuswb 4064(%rdx), %ymm26, %ymm20
+
+// CHECK: vpackuswb 4096(%rdx), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xad,0x20,0x67,0xa2,0x00,0x10,0x00,0x00]
+          vpackuswb 4096(%rdx), %ymm26, %ymm20
+
+// CHECK: vpackuswb -4096(%rdx), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xad,0x20,0x67,0x62,0x80]
+          vpackuswb -4096(%rdx), %ymm26, %ymm20
+
+// CHECK: vpackuswb -4128(%rdx), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xad,0x20,0x67,0xa2,0xe0,0xef,0xff,0xff]
+          vpackuswb -4128(%rdx), %ymm26, %ymm20
+
+// CHECK: vpackssdw %xmm26, %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x01,0x2d,0x00,0x6b,0xf2]
+          vpackssdw %xmm26, %xmm26, %xmm30
+
+// CHECK: vpackssdw %xmm26, %xmm26, %xmm30 {%k2}
+// CHECK:  encoding: [0x62,0x01,0x2d,0x02,0x6b,0xf2]
+          vpackssdw %xmm26, %xmm26, %xmm30 {%k2}
+
+// CHECK: vpackssdw %xmm26, %xmm26, %xmm30 {%k2} {z}
+// CHECK:  encoding: [0x62,0x01,0x2d,0x82,0x6b,0xf2]
+          vpackssdw %xmm26, %xmm26, %xmm30 {%k2} {z}
+
+// CHECK: vpackssdw (%rcx), %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x2d,0x00,0x6b,0x31]
+          vpackssdw (%rcx), %xmm26, %xmm30
+
+// CHECK: vpackssdw 291(%rax,%r14,8), %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x21,0x2d,0x00,0x6b,0xb4,0xf0,0x23,0x01,0x00,0x00]
+          vpackssdw 291(%rax,%r14,8), %xmm26, %xmm30
+
+// CHECK: vpackssdw (%rcx){1to4}, %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x2d,0x10,0x6b,0x31]
+          vpackssdw (%rcx){1to4}, %xmm26, %xmm30
+
+// CHECK: vpackssdw 2032(%rdx), %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x2d,0x00,0x6b,0x72,0x7f]
+          vpackssdw 2032(%rdx), %xmm26, %xmm30
+
+// CHECK: vpackssdw 2048(%rdx), %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x2d,0x00,0x6b,0xb2,0x00,0x08,0x00,0x00]
+          vpackssdw 2048(%rdx), %xmm26, %xmm30
+
+// CHECK: vpackssdw -2048(%rdx), %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x2d,0x00,0x6b,0x72,0x80]
+          vpackssdw -2048(%rdx), %xmm26, %xmm30
+
+// CHECK: vpackssdw -2064(%rdx), %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x2d,0x00,0x6b,0xb2,0xf0,0xf7,0xff,0xff]
+          vpackssdw -2064(%rdx), %xmm26, %xmm30
+
+// CHECK: vpackssdw 508(%rdx){1to4}, %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x2d,0x10,0x6b,0x72,0x7f]
+          vpackssdw 508(%rdx){1to4}, %xmm26, %xmm30
+
+// CHECK: vpackssdw 512(%rdx){1to4}, %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x2d,0x10,0x6b,0xb2,0x00,0x02,0x00,0x00]
+          vpackssdw 512(%rdx){1to4}, %xmm26, %xmm30
+
+// CHECK: vpackssdw -512(%rdx){1to4}, %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x2d,0x10,0x6b,0x72,0x80]
+          vpackssdw -512(%rdx){1to4}, %xmm26, %xmm30
+
+// CHECK: vpackssdw -516(%rdx){1to4}, %xmm26, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x2d,0x10,0x6b,0xb2,0xfc,0xfd,0xff,0xff]
+          vpackssdw -516(%rdx){1to4}, %xmm26, %xmm30
+
+// CHECK: vpackssdw %ymm26, %ymm22, %ymm28
+// CHECK:  encoding: [0x62,0x01,0x4d,0x20,0x6b,0xe2]
+          vpackssdw %ymm26, %ymm22, %ymm28
+
+// CHECK: vpackssdw %ymm26, %ymm22, %ymm28 {%k7}
+// CHECK:  encoding: [0x62,0x01,0x4d,0x27,0x6b,0xe2]
+          vpackssdw %ymm26, %ymm22, %ymm28 {%k7}
+
+// CHECK: vpackssdw %ymm26, %ymm22, %ymm28 {%k7} {z}
+// CHECK:  encoding: [0x62,0x01,0x4d,0xa7,0x6b,0xe2]
+          vpackssdw %ymm26, %ymm22, %ymm28 {%k7} {z}
+
+// CHECK: vpackssdw (%rcx), %ymm22, %ymm28
+// CHECK:  encoding: [0x62,0x61,0x4d,0x20,0x6b,0x21]
+          vpackssdw (%rcx), %ymm22, %ymm28
+
+// CHECK: vpackssdw 291(%rax,%r14,8), %ymm22, %ymm28
+// CHECK:  encoding: [0x62,0x21,0x4d,0x20,0x6b,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpackssdw 291(%rax,%r14,8), %ymm22, %ymm28
+
+// CHECK: vpackssdw (%rcx){1to8}, %ymm22, %ymm28
+// CHECK:  encoding: [0x62,0x61,0x4d,0x30,0x6b,0x21]
+          vpackssdw (%rcx){1to8}, %ymm22, %ymm28
+
+// CHECK: vpackssdw 4064(%rdx), %ymm22, %ymm28
+// CHECK:  encoding: [0x62,0x61,0x4d,0x20,0x6b,0x62,0x7f]
+          vpackssdw 4064(%rdx), %ymm22, %ymm28
+
+// CHECK: vpackssdw 4096(%rdx), %ymm22, %ymm28
+// CHECK:  encoding: [0x62,0x61,0x4d,0x20,0x6b,0xa2,0x00,0x10,0x00,0x00]
+          vpackssdw 4096(%rdx), %ymm22, %ymm28
+
+// CHECK: vpackssdw -4096(%rdx), %ymm22, %ymm28
+// CHECK:  encoding: [0x62,0x61,0x4d,0x20,0x6b,0x62,0x80]
+          vpackssdw -4096(%rdx), %ymm22, %ymm28
+
+// CHECK: vpackssdw -4128(%rdx), %ymm22, %ymm28
+// CHECK:  encoding: [0x62,0x61,0x4d,0x20,0x6b,0xa2,0xe0,0xef,0xff,0xff]
+          vpackssdw -4128(%rdx), %ymm22, %ymm28
+
+// CHECK: vpackssdw 508(%rdx){1to8}, %ymm22, %ymm28
+// CHECK:  encoding: [0x62,0x61,0x4d,0x30,0x6b,0x62,0x7f]
+          vpackssdw 508(%rdx){1to8}, %ymm22, %ymm28
+
+// CHECK: vpackssdw 512(%rdx){1to8}, %ymm22, %ymm28
+// CHECK:  encoding: [0x62,0x61,0x4d,0x30,0x6b,0xa2,0x00,0x02,0x00,0x00]
+          vpackssdw 512(%rdx){1to8}, %ymm22, %ymm28
+
+// CHECK: vpackssdw -512(%rdx){1to8}, %ymm22, %ymm28
+// CHECK:  encoding: [0x62,0x61,0x4d,0x30,0x6b,0x62,0x80]
+          vpackssdw -512(%rdx){1to8}, %ymm22, %ymm28
+
+// CHECK: vpackssdw -516(%rdx){1to8}, %ymm22, %ymm28
+// CHECK:  encoding: [0x62,0x61,0x4d,0x30,0x6b,0xa2,0xfc,0xfd,0xff,0xff]
+          vpackssdw -516(%rdx){1to8}, %ymm22, %ymm28
+
+// CHECK: vpacksswb %xmm21, %xmm28, %xmm25
+// CHECK:  encoding: [0x62,0x21,0x9d,0x00,0x63,0xcd]
+          vpacksswb %xmm21, %xmm28, %xmm25
+
+// CHECK: vpacksswb %xmm21, %xmm28, %xmm25 {%k5}
+// CHECK:  encoding: [0x62,0x21,0x9d,0x05,0x63,0xcd]
+          vpacksswb %xmm21, %xmm28, %xmm25 {%k5}
+
+// CHECK: vpacksswb %xmm21, %xmm28, %xmm25 {%k5} {z}
+// CHECK:  encoding: [0x62,0x21,0x9d,0x85,0x63,0xcd]
+          vpacksswb %xmm21, %xmm28, %xmm25 {%k5} {z}
+
+// CHECK: vpacksswb (%rcx), %xmm28, %xmm25
+// CHECK:  encoding: [0x62,0x61,0x9d,0x00,0x63,0x09]
+          vpacksswb (%rcx), %xmm28, %xmm25
+
+// CHECK: vpacksswb 291(%rax,%r14,8), %xmm28, %xmm25
+// CHECK:  encoding: [0x62,0x21,0x9d,0x00,0x63,0x8c,0xf0,0x23,0x01,0x00,0x00]
+          vpacksswb 291(%rax,%r14,8), %xmm28, %xmm25
+
+// CHECK: vpacksswb 2032(%rdx), %xmm28, %xmm25
+// CHECK:  encoding: [0x62,0x61,0x9d,0x00,0x63,0x4a,0x7f]
+          vpacksswb 2032(%rdx), %xmm28, %xmm25
+
+// CHECK: vpacksswb 2048(%rdx), %xmm28, %xmm25
+// CHECK:  encoding: [0x62,0x61,0x9d,0x00,0x63,0x8a,0x00,0x08,0x00,0x00]
+          vpacksswb 2048(%rdx), %xmm28, %xmm25
+
+// CHECK: vpacksswb -2048(%rdx), %xmm28, %xmm25
+// CHECK:  encoding: [0x62,0x61,0x9d,0x00,0x63,0x4a,0x80]
+          vpacksswb -2048(%rdx), %xmm28, %xmm25
+
+// CHECK: vpacksswb -2064(%rdx), %xmm28, %xmm25
+// CHECK:  encoding: [0x62,0x61,0x9d,0x00,0x63,0x8a,0xf0,0xf7,0xff,0xff]
+          vpacksswb -2064(%rdx), %xmm28, %xmm25
+
+// CHECK: vpacksswb %ymm22, %ymm28, %ymm21
+// CHECK:  encoding: [0x62,0xa1,0x9d,0x20,0x63,0xee]
+          vpacksswb %ymm22, %ymm28, %ymm21
+
+// CHECK: vpacksswb %ymm22, %ymm28, %ymm21 {%k5}
+// CHECK:  encoding: [0x62,0xa1,0x9d,0x25,0x63,0xee]
+          vpacksswb %ymm22, %ymm28, %ymm21 {%k5}
+
+// CHECK: vpacksswb %ymm22, %ymm28, %ymm21 {%k5} {z}
+// CHECK:  encoding: [0x62,0xa1,0x9d,0xa5,0x63,0xee]
+          vpacksswb %ymm22, %ymm28, %ymm21 {%k5} {z}
+
+// CHECK: vpacksswb (%rcx), %ymm28, %ymm21
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x20,0x63,0x29]
+          vpacksswb (%rcx), %ymm28, %ymm21
+
+// CHECK: vpacksswb 291(%rax,%r14,8), %ymm28, %ymm21
+// CHECK:  encoding: [0x62,0xa1,0x9d,0x20,0x63,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vpacksswb 291(%rax,%r14,8), %ymm28, %ymm21
+
+// CHECK: vpacksswb 4064(%rdx), %ymm28, %ymm21
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x20,0x63,0x6a,0x7f]
+          vpacksswb 4064(%rdx), %ymm28, %ymm21
+
+// CHECK: vpacksswb 4096(%rdx), %ymm28, %ymm21
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x20,0x63,0xaa,0x00,0x10,0x00,0x00]
+          vpacksswb 4096(%rdx), %ymm28, %ymm21
+
+// CHECK: vpacksswb -4096(%rdx), %ymm28, %ymm21
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x20,0x63,0x6a,0x80]
+          vpacksswb -4096(%rdx), %ymm28, %ymm21
+
+// CHECK: vpacksswb -4128(%rdx), %ymm28, %ymm21
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x20,0x63,0xaa,0xe0,0xef,0xff,0xff]
+          vpacksswb -4128(%rdx), %ymm28, %ymm21
+
+// CHECK: vpackusdw %xmm20, %xmm24, %xmm24
+// CHECK:  encoding: [0x62,0x22,0x3d,0x00,0x2b,0xc4]
+          vpackusdw %xmm20, %xmm24, %xmm24
+
+// CHECK: vpackusdw %xmm20, %xmm24, %xmm24 {%k5}
+// CHECK:  encoding: [0x62,0x22,0x3d,0x05,0x2b,0xc4]
+          vpackusdw %xmm20, %xmm24, %xmm24 {%k5}
+
+// CHECK: vpackusdw %xmm20, %xmm24, %xmm24 {%k5} {z}
+// CHECK:  encoding: [0x62,0x22,0x3d,0x85,0x2b,0xc4]
+          vpackusdw %xmm20, %xmm24, %xmm24 {%k5} {z}
+
+// CHECK: vpackusdw (%rcx), %xmm24, %xmm24
+// CHECK:  encoding: [0x62,0x62,0x3d,0x00,0x2b,0x01]
+          vpackusdw (%rcx), %xmm24, %xmm24
+
+// CHECK: vpackusdw 291(%rax,%r14,8), %xmm24, %xmm24
+// CHECK:  encoding: [0x62,0x22,0x3d,0x00,0x2b,0x84,0xf0,0x23,0x01,0x00,0x00]
+          vpackusdw 291(%rax,%r14,8), %xmm24, %xmm24
+
+// CHECK: vpackusdw (%rcx){1to4}, %xmm24, %xmm24
+// CHECK:  encoding: [0x62,0x62,0x3d,0x10,0x2b,0x01]
+          vpackusdw (%rcx){1to4}, %xmm24, %xmm24
+
+// CHECK: vpackusdw 2032(%rdx), %xmm24, %xmm24
+// CHECK:  encoding: [0x62,0x62,0x3d,0x00,0x2b,0x42,0x7f]
+          vpackusdw 2032(%rdx), %xmm24, %xmm24
+
+// CHECK: vpackusdw 2048(%rdx), %xmm24, %xmm24
+// CHECK:  encoding: [0x62,0x62,0x3d,0x00,0x2b,0x82,0x00,0x08,0x00,0x00]
+          vpackusdw 2048(%rdx), %xmm24, %xmm24
+
+// CHECK: vpackusdw -2048(%rdx), %xmm24, %xmm24
+// CHECK:  encoding: [0x62,0x62,0x3d,0x00,0x2b,0x42,0x80]
+          vpackusdw -2048(%rdx), %xmm24, %xmm24
+
+// CHECK: vpackusdw -2064(%rdx), %xmm24, %xmm24
+// CHECK:  encoding: [0x62,0x62,0x3d,0x00,0x2b,0x82,0xf0,0xf7,0xff,0xff]
+          vpackusdw -2064(%rdx), %xmm24, %xmm24
+
+// CHECK: vpackusdw 508(%rdx){1to4}, %xmm24, %xmm24
+// CHECK:  encoding: [0x62,0x62,0x3d,0x10,0x2b,0x42,0x7f]
+          vpackusdw 508(%rdx){1to4}, %xmm24, %xmm24
+
+// CHECK: vpackusdw 512(%rdx){1to4}, %xmm24, %xmm24
+// CHECK:  encoding: [0x62,0x62,0x3d,0x10,0x2b,0x82,0x00,0x02,0x00,0x00]
+          vpackusdw 512(%rdx){1to4}, %xmm24, %xmm24
+
+// CHECK: vpackusdw -512(%rdx){1to4}, %xmm24, %xmm24
+// CHECK:  encoding: [0x62,0x62,0x3d,0x10,0x2b,0x42,0x80]
+          vpackusdw -512(%rdx){1to4}, %xmm24, %xmm24
+
+// CHECK: vpackusdw -516(%rdx){1to4}, %xmm24, %xmm24
+// CHECK:  encoding: [0x62,0x62,0x3d,0x10,0x2b,0x82,0xfc,0xfd,0xff,0xff]
+          vpackusdw -516(%rdx){1to4}, %xmm24, %xmm24
+
+// CHECK: vpackusdw %ymm21, %ymm26, %ymm19
+// CHECK:  encoding: [0x62,0xa2,0x2d,0x20,0x2b,0xdd]
+          vpackusdw %ymm21, %ymm26, %ymm19
+
+// CHECK: vpackusdw %ymm21, %ymm26, %ymm19 {%k1}
+// CHECK:  encoding: [0x62,0xa2,0x2d,0x21,0x2b,0xdd]
+          vpackusdw %ymm21, %ymm26, %ymm19 {%k1}
+
+// CHECK: vpackusdw %ymm21, %ymm26, %ymm19 {%k1} {z}
+// CHECK:  encoding: [0x62,0xa2,0x2d,0xa1,0x2b,0xdd]
+          vpackusdw %ymm21, %ymm26, %ymm19 {%k1} {z}
+
+// CHECK: vpackusdw (%rcx), %ymm26, %ymm19
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x20,0x2b,0x19]
+          vpackusdw (%rcx), %ymm26, %ymm19
+
+// CHECK: vpackusdw 291(%rax,%r14,8), %ymm26, %ymm19
+// CHECK:  encoding: [0x62,0xa2,0x2d,0x20,0x2b,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vpackusdw 291(%rax,%r14,8), %ymm26, %ymm19
+
+// CHECK: vpackusdw (%rcx){1to8}, %ymm26, %ymm19
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x30,0x2b,0x19]
+          vpackusdw (%rcx){1to8}, %ymm26, %ymm19
+
+// CHECK: vpackusdw 4064(%rdx), %ymm26, %ymm19
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x20,0x2b,0x5a,0x7f]
+          vpackusdw 4064(%rdx), %ymm26, %ymm19
+
+// CHECK: vpackusdw 4096(%rdx), %ymm26, %ymm19
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x20,0x2b,0x9a,0x00,0x10,0x00,0x00]
+          vpackusdw 4096(%rdx), %ymm26, %ymm19
+
+// CHECK: vpackusdw -4096(%rdx), %ymm26, %ymm19
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x20,0x2b,0x5a,0x80]
+          vpackusdw -4096(%rdx), %ymm26, %ymm19
+
+// CHECK: vpackusdw -4128(%rdx), %ymm26, %ymm19
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x20,0x2b,0x9a,0xe0,0xef,0xff,0xff]
+          vpackusdw -4128(%rdx), %ymm26, %ymm19
+
+// CHECK: vpackusdw 508(%rdx){1to8}, %ymm26, %ymm19
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x30,0x2b,0x5a,0x7f]
+          vpackusdw 508(%rdx){1to8}, %ymm26, %ymm19
+
+// CHECK: vpackusdw 512(%rdx){1to8}, %ymm26, %ymm19
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x30,0x2b,0x9a,0x00,0x02,0x00,0x00]
+          vpackusdw 512(%rdx){1to8}, %ymm26, %ymm19
+
+// CHECK: vpackusdw -512(%rdx){1to8}, %ymm26, %ymm19
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x30,0x2b,0x5a,0x80]
+          vpackusdw -512(%rdx){1to8}, %ymm26, %ymm19
+
+// CHECK: vpackusdw -516(%rdx){1to8}, %ymm26, %ymm19
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x30,0x2b,0x9a,0xfc,0xfd,0xff,0xff]
+          vpackusdw -516(%rdx){1to8}, %ymm26, %ymm19
+
+// CHECK: vpackuswb %xmm17, %xmm27, %xmm21
+// CHECK:  encoding: [0x62,0xa1,0xa5,0x00,0x67,0xe9]
+          vpackuswb %xmm17, %xmm27, %xmm21
+
+// CHECK: vpackuswb %xmm17, %xmm27, %xmm21 {%k4}
+// CHECK:  encoding: [0x62,0xa1,0xa5,0x04,0x67,0xe9]
+          vpackuswb %xmm17, %xmm27, %xmm21 {%k4}
+
+// CHECK: vpackuswb %xmm17, %xmm27, %xmm21 {%k4} {z}
+// CHECK:  encoding: [0x62,0xa1,0xa5,0x84,0x67,0xe9]
+          vpackuswb %xmm17, %xmm27, %xmm21 {%k4} {z}
+
+// CHECK: vpackuswb (%rcx), %xmm27, %xmm21
+// CHECK:  encoding: [0x62,0xe1,0xa5,0x00,0x67,0x29]
+          vpackuswb (%rcx), %xmm27, %xmm21
+
+// CHECK: vpackuswb 291(%rax,%r14,8), %xmm27, %xmm21
+// CHECK:  encoding: [0x62,0xa1,0xa5,0x00,0x67,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vpackuswb 291(%rax,%r14,8), %xmm27, %xmm21
+
+// CHECK: vpackuswb 2032(%rdx), %xmm27, %xmm21
+// CHECK:  encoding: [0x62,0xe1,0xa5,0x00,0x67,0x6a,0x7f]
+          vpackuswb 2032(%rdx), %xmm27, %xmm21
+
+// CHECK: vpackuswb 2048(%rdx), %xmm27, %xmm21
+// CHECK:  encoding: [0x62,0xe1,0xa5,0x00,0x67,0xaa,0x00,0x08,0x00,0x00]
+          vpackuswb 2048(%rdx), %xmm27, %xmm21
+
+// CHECK: vpackuswb -2048(%rdx), %xmm27, %xmm21
+// CHECK:  encoding: [0x62,0xe1,0xa5,0x00,0x67,0x6a,0x80]
+          vpackuswb -2048(%rdx), %xmm27, %xmm21
+
+// CHECK: vpackuswb -2064(%rdx), %xmm27, %xmm21
+// CHECK:  encoding: [0x62,0xe1,0xa5,0x00,0x67,0xaa,0xf0,0xf7,0xff,0xff]
+          vpackuswb -2064(%rdx), %xmm27, %xmm21
+
+// CHECK: vpackuswb %ymm20, %ymm28, %ymm19
+// CHECK:  encoding: [0x62,0xa1,0x9d,0x20,0x67,0xdc]
+          vpackuswb %ymm20, %ymm28, %ymm19
+
+// CHECK: vpackuswb %ymm20, %ymm28, %ymm19 {%k1}
+// CHECK:  encoding: [0x62,0xa1,0x9d,0x21,0x67,0xdc]
+          vpackuswb %ymm20, %ymm28, %ymm19 {%k1}
+
+// CHECK: vpackuswb %ymm20, %ymm28, %ymm19 {%k1} {z}
+// CHECK:  encoding: [0x62,0xa1,0x9d,0xa1,0x67,0xdc]
+          vpackuswb %ymm20, %ymm28, %ymm19 {%k1} {z}
+
+// CHECK: vpackuswb (%rcx), %ymm28, %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x20,0x67,0x19]
+          vpackuswb (%rcx), %ymm28, %ymm19
+
+// CHECK: vpackuswb 291(%rax,%r14,8), %ymm28, %ymm19
+// CHECK:  encoding: [0x62,0xa1,0x9d,0x20,0x67,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vpackuswb 291(%rax,%r14,8), %ymm28, %ymm19
+
+// CHECK: vpackuswb 4064(%rdx), %ymm28, %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x20,0x67,0x5a,0x7f]
+          vpackuswb 4064(%rdx), %ymm28, %ymm19
+
+// CHECK: vpackuswb 4096(%rdx), %ymm28, %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x20,0x67,0x9a,0x00,0x10,0x00,0x00]
+          vpackuswb 4096(%rdx), %ymm28, %ymm19
+
+// CHECK: vpackuswb -4096(%rdx), %ymm28, %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x20,0x67,0x5a,0x80]
+          vpackuswb -4096(%rdx), %ymm28, %ymm19
+
+// CHECK: vpackuswb -4128(%rdx), %ymm28, %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x9d,0x20,0x67,0x9a,0xe0,0xef,0xff,0xff]
+          vpackuswb -4128(%rdx), %ymm28, %ymm19
+
+// CHECK: vpackssdw %xmm19, %xmm17, %xmm24
+// CHECK:  encoding: [0x62,0x21,0x75,0x00,0x6b,0xc3]
+          vpackssdw %xmm19, %xmm17, %xmm24
+
+// CHECK: vpackssdw %xmm19, %xmm17, %xmm24 {%k7}
+// CHECK:  encoding: [0x62,0x21,0x75,0x07,0x6b,0xc3]
+          vpackssdw %xmm19, %xmm17, %xmm24 {%k7}
+
+// CHECK: vpackssdw %xmm19, %xmm17, %xmm24 {%k7} {z}
+// CHECK:  encoding: [0x62,0x21,0x75,0x87,0x6b,0xc3]
+          vpackssdw %xmm19, %xmm17, %xmm24 {%k7} {z}
+
+// CHECK: vpackssdw (%rcx), %xmm17, %xmm24
+// CHECK:  encoding: [0x62,0x61,0x75,0x00,0x6b,0x01]
+          vpackssdw (%rcx), %xmm17, %xmm24
+
+// CHECK: vpackssdw 4660(%rax,%r14,8), %xmm17, %xmm24
+// CHECK:  encoding: [0x62,0x21,0x75,0x00,0x6b,0x84,0xf0,0x34,0x12,0x00,0x00]
+          vpackssdw 4660(%rax,%r14,8), %xmm17, %xmm24
+
+// CHECK: vpackssdw (%rcx){1to4}, %xmm17, %xmm24
+// CHECK:  encoding: [0x62,0x61,0x75,0x10,0x6b,0x01]
+          vpackssdw (%rcx){1to4}, %xmm17, %xmm24
+
+// CHECK: vpackssdw 2032(%rdx), %xmm17, %xmm24
+// CHECK:  encoding: [0x62,0x61,0x75,0x00,0x6b,0x42,0x7f]
+          vpackssdw 2032(%rdx), %xmm17, %xmm24
+
+// CHECK: vpackssdw 2048(%rdx), %xmm17, %xmm24
+// CHECK:  encoding: [0x62,0x61,0x75,0x00,0x6b,0x82,0x00,0x08,0x00,0x00]
+          vpackssdw 2048(%rdx), %xmm17, %xmm24
+
+// CHECK: vpackssdw -2048(%rdx), %xmm17, %xmm24
+// CHECK:  encoding: [0x62,0x61,0x75,0x00,0x6b,0x42,0x80]
+          vpackssdw -2048(%rdx), %xmm17, %xmm24
+
+// CHECK: vpackssdw -2064(%rdx), %xmm17, %xmm24
+// CHECK:  encoding: [0x62,0x61,0x75,0x00,0x6b,0x82,0xf0,0xf7,0xff,0xff]
+          vpackssdw -2064(%rdx), %xmm17, %xmm24
+
+// CHECK: vpackssdw 508(%rdx){1to4}, %xmm17, %xmm24
+// CHECK:  encoding: [0x62,0x61,0x75,0x10,0x6b,0x42,0x7f]
+          vpackssdw 508(%rdx){1to4}, %xmm17, %xmm24
+
+// CHECK: vpackssdw 512(%rdx){1to4}, %xmm17, %xmm24
+// CHECK:  encoding: [0x62,0x61,0x75,0x10,0x6b,0x82,0x00,0x02,0x00,0x00]
+          vpackssdw 512(%rdx){1to4}, %xmm17, %xmm24
+
+// CHECK: vpackssdw -512(%rdx){1to4}, %xmm17, %xmm24
+// CHECK:  encoding: [0x62,0x61,0x75,0x10,0x6b,0x42,0x80]
+          vpackssdw -512(%rdx){1to4}, %xmm17, %xmm24
+
+// CHECK: vpackssdw -516(%rdx){1to4}, %xmm17, %xmm24
+// CHECK:  encoding: [0x62,0x61,0x75,0x10,0x6b,0x82,0xfc,0xfd,0xff,0xff]
+          vpackssdw -516(%rdx){1to4}, %xmm17, %xmm24
+
+// CHECK: vpackssdw %ymm19, %ymm19, %ymm25
+// CHECK:  encoding: [0x62,0x21,0x65,0x20,0x6b,0xcb]
+          vpackssdw %ymm19, %ymm19, %ymm25
+
+// CHECK: vpackssdw %ymm19, %ymm19, %ymm25 {%k2}
+// CHECK:  encoding: [0x62,0x21,0x65,0x22,0x6b,0xcb]
+          vpackssdw %ymm19, %ymm19, %ymm25 {%k2}
+
+// CHECK: vpackssdw %ymm19, %ymm19, %ymm25 {%k2} {z}
+// CHECK:  encoding: [0x62,0x21,0x65,0xa2,0x6b,0xcb]
+          vpackssdw %ymm19, %ymm19, %ymm25 {%k2} {z}
+
+// CHECK: vpackssdw (%rcx), %ymm19, %ymm25
+// CHECK:  encoding: [0x62,0x61,0x65,0x20,0x6b,0x09]
+          vpackssdw (%rcx), %ymm19, %ymm25
+
+// CHECK: vpackssdw 4660(%rax,%r14,8), %ymm19, %ymm25
+// CHECK:  encoding: [0x62,0x21,0x65,0x20,0x6b,0x8c,0xf0,0x34,0x12,0x00,0x00]
+          vpackssdw 4660(%rax,%r14,8), %ymm19, %ymm25
+
+// CHECK: vpackssdw (%rcx){1to8}, %ymm19, %ymm25
+// CHECK:  encoding: [0x62,0x61,0x65,0x30,0x6b,0x09]
+          vpackssdw (%rcx){1to8}, %ymm19, %ymm25
+
+// CHECK: vpackssdw 4064(%rdx), %ymm19, %ymm25
+// CHECK:  encoding: [0x62,0x61,0x65,0x20,0x6b,0x4a,0x7f]
+          vpackssdw 4064(%rdx), %ymm19, %ymm25
+
+// CHECK: vpackssdw 4096(%rdx), %ymm19, %ymm25
+// CHECK:  encoding: [0x62,0x61,0x65,0x20,0x6b,0x8a,0x00,0x10,0x00,0x00]
+          vpackssdw 4096(%rdx), %ymm19, %ymm25
+
+// CHECK: vpackssdw -4096(%rdx), %ymm19, %ymm25
+// CHECK:  encoding: [0x62,0x61,0x65,0x20,0x6b,0x4a,0x80]
+          vpackssdw -4096(%rdx), %ymm19, %ymm25
+
+// CHECK: vpackssdw -4128(%rdx), %ymm19, %ymm25
+// CHECK:  encoding: [0x62,0x61,0x65,0x20,0x6b,0x8a,0xe0,0xef,0xff,0xff]
+          vpackssdw -4128(%rdx), %ymm19, %ymm25
+
+// CHECK: vpackssdw 508(%rdx){1to8}, %ymm19, %ymm25
+// CHECK:  encoding: [0x62,0x61,0x65,0x30,0x6b,0x4a,0x7f]
+          vpackssdw 508(%rdx){1to8}, %ymm19, %ymm25
+
+// CHECK: vpackssdw 512(%rdx){1to8}, %ymm19, %ymm25
+// CHECK:  encoding: [0x62,0x61,0x65,0x30,0x6b,0x8a,0x00,0x02,0x00,0x00]
+          vpackssdw 512(%rdx){1to8}, %ymm19, %ymm25
+
+// CHECK: vpackssdw -512(%rdx){1to8}, %ymm19, %ymm25
+// CHECK:  encoding: [0x62,0x61,0x65,0x30,0x6b,0x4a,0x80]
+          vpackssdw -512(%rdx){1to8}, %ymm19, %ymm25
+
+// CHECK: vpackssdw -516(%rdx){1to8}, %ymm19, %ymm25
+// CHECK:  encoding: [0x62,0x61,0x65,0x30,0x6b,0x8a,0xfc,0xfd,0xff,0xff]
+          vpackssdw -516(%rdx){1to8}, %ymm19, %ymm25
+
+// CHECK: vpacksswb %xmm22, %xmm25, %xmm29
+// CHECK:  encoding: [0x62,0x21,0xb5,0x00,0x63,0xee]
+          vpacksswb %xmm22, %xmm25, %xmm29
+
+// CHECK: vpacksswb %xmm22, %xmm25, %xmm29 {%k5}
+// CHECK:  encoding: [0x62,0x21,0xb5,0x05,0x63,0xee]
+          vpacksswb %xmm22, %xmm25, %xmm29 {%k5}
+
+// CHECK: vpacksswb %xmm22, %xmm25, %xmm29 {%k5} {z}
+// CHECK:  encoding: [0x62,0x21,0xb5,0x85,0x63,0xee]
+          vpacksswb %xmm22, %xmm25, %xmm29 {%k5} {z}
+
+// CHECK: vpacksswb (%rcx), %xmm25, %xmm29
+// CHECK:  encoding: [0x62,0x61,0xb5,0x00,0x63,0x29]
+          vpacksswb (%rcx), %xmm25, %xmm29
+
+// CHECK: vpacksswb 4660(%rax,%r14,8), %xmm25, %xmm29
+// CHECK:  encoding: [0x62,0x21,0xb5,0x00,0x63,0xac,0xf0,0x34,0x12,0x00,0x00]
+          vpacksswb 4660(%rax,%r14,8), %xmm25, %xmm29
+
+// CHECK: vpacksswb 2032(%rdx), %xmm25, %xmm29
+// CHECK:  encoding: [0x62,0x61,0xb5,0x00,0x63,0x6a,0x7f]
+          vpacksswb 2032(%rdx), %xmm25, %xmm29
+
+// CHECK: vpacksswb 2048(%rdx), %xmm25, %xmm29
+// CHECK:  encoding: [0x62,0x61,0xb5,0x00,0x63,0xaa,0x00,0x08,0x00,0x00]
+          vpacksswb 2048(%rdx), %xmm25, %xmm29
+
+// CHECK: vpacksswb -2048(%rdx), %xmm25, %xmm29
+// CHECK:  encoding: [0x62,0x61,0xb5,0x00,0x63,0x6a,0x80]
+          vpacksswb -2048(%rdx), %xmm25, %xmm29
+
+// CHECK: vpacksswb -2064(%rdx), %xmm25, %xmm29
+// CHECK:  encoding: [0x62,0x61,0xb5,0x00,0x63,0xaa,0xf0,0xf7,0xff,0xff]
+          vpacksswb -2064(%rdx), %xmm25, %xmm29
+
+// CHECK: vpacksswb %ymm27, %ymm20, %ymm26
+// CHECK:  encoding: [0x62,0x01,0xdd,0x20,0x63,0xd3]
+          vpacksswb %ymm27, %ymm20, %ymm26
+
+// CHECK: vpacksswb %ymm27, %ymm20, %ymm26 {%k6}
+// CHECK:  encoding: [0x62,0x01,0xdd,0x26,0x63,0xd3]
+          vpacksswb %ymm27, %ymm20, %ymm26 {%k6}
+
+// CHECK: vpacksswb %ymm27, %ymm20, %ymm26 {%k6} {z}
+// CHECK:  encoding: [0x62,0x01,0xdd,0xa6,0x63,0xd3]
+          vpacksswb %ymm27, %ymm20, %ymm26 {%k6} {z}
+
+// CHECK: vpacksswb (%rcx), %ymm20, %ymm26
+// CHECK:  encoding: [0x62,0x61,0xdd,0x20,0x63,0x11]
+          vpacksswb (%rcx), %ymm20, %ymm26
+
+// CHECK: vpacksswb 4660(%rax,%r14,8), %ymm20, %ymm26
+// CHECK:  encoding: [0x62,0x21,0xdd,0x20,0x63,0x94,0xf0,0x34,0x12,0x00,0x00]
+          vpacksswb 4660(%rax,%r14,8), %ymm20, %ymm26
+
+// CHECK: vpacksswb 4064(%rdx), %ymm20, %ymm26
+// CHECK:  encoding: [0x62,0x61,0xdd,0x20,0x63,0x52,0x7f]
+          vpacksswb 4064(%rdx), %ymm20, %ymm26
+
+// CHECK: vpacksswb 4096(%rdx), %ymm20, %ymm26
+// CHECK:  encoding: [0x62,0x61,0xdd,0x20,0x63,0x92,0x00,0x10,0x00,0x00]
+          vpacksswb 4096(%rdx), %ymm20, %ymm26
+
+// CHECK: vpacksswb -4096(%rdx), %ymm20, %ymm26
+// CHECK:  encoding: [0x62,0x61,0xdd,0x20,0x63,0x52,0x80]
+          vpacksswb -4096(%rdx), %ymm20, %ymm26
+
+// CHECK: vpacksswb -4128(%rdx), %ymm20, %ymm26
+// CHECK:  encoding: [0x62,0x61,0xdd,0x20,0x63,0x92,0xe0,0xef,0xff,0xff]
+          vpacksswb -4128(%rdx), %ymm20, %ymm26
+
+// CHECK: vpackusdw %xmm18, %xmm27, %xmm28
+// CHECK:  encoding: [0x62,0x22,0x25,0x00,0x2b,0xe2]
+          vpackusdw %xmm18, %xmm27, %xmm28
+
+// CHECK: vpackusdw %xmm18, %xmm27, %xmm28 {%k4}
+// CHECK:  encoding: [0x62,0x22,0x25,0x04,0x2b,0xe2]
+          vpackusdw %xmm18, %xmm27, %xmm28 {%k4}
+
+// CHECK: vpackusdw %xmm18, %xmm27, %xmm28 {%k4} {z}
+// CHECK:  encoding: [0x62,0x22,0x25,0x84,0x2b,0xe2]
+          vpackusdw %xmm18, %xmm27, %xmm28 {%k4} {z}
+
+// CHECK: vpackusdw (%rcx), %xmm27, %xmm28
+// CHECK:  encoding: [0x62,0x62,0x25,0x00,0x2b,0x21]
+          vpackusdw (%rcx), %xmm27, %xmm28
+
+// CHECK: vpackusdw 4660(%rax,%r14,8), %xmm27, %xmm28
+// CHECK:  encoding: [0x62,0x22,0x25,0x00,0x2b,0xa4,0xf0,0x34,0x12,0x00,0x00]
+          vpackusdw 4660(%rax,%r14,8), %xmm27, %xmm28
+
+// CHECK: vpackusdw (%rcx){1to4}, %xmm27, %xmm28
+// CHECK:  encoding: [0x62,0x62,0x25,0x10,0x2b,0x21]
+          vpackusdw (%rcx){1to4}, %xmm27, %xmm28
+
+// CHECK: vpackusdw 2032(%rdx), %xmm27, %xmm28
+// CHECK:  encoding: [0x62,0x62,0x25,0x00,0x2b,0x62,0x7f]
+          vpackusdw 2032(%rdx), %xmm27, %xmm28
+
+// CHECK: vpackusdw 2048(%rdx), %xmm27, %xmm28
+// CHECK:  encoding: [0x62,0x62,0x25,0x00,0x2b,0xa2,0x00,0x08,0x00,0x00]
+          vpackusdw 2048(%rdx), %xmm27, %xmm28
+
+// CHECK: vpackusdw -2048(%rdx), %xmm27, %xmm28
+// CHECK:  encoding: [0x62,0x62,0x25,0x00,0x2b,0x62,0x80]
+          vpackusdw -2048(%rdx), %xmm27, %xmm28
+
+// CHECK: vpackusdw -2064(%rdx), %xmm27, %xmm28
+// CHECK:  encoding: [0x62,0x62,0x25,0x00,0x2b,0xa2,0xf0,0xf7,0xff,0xff]
+          vpackusdw -2064(%rdx), %xmm27, %xmm28
+
+// CHECK: vpackusdw 508(%rdx){1to4}, %xmm27, %xmm28
+// CHECK:  encoding: [0x62,0x62,0x25,0x10,0x2b,0x62,0x7f]
+          vpackusdw 508(%rdx){1to4}, %xmm27, %xmm28
+
+// CHECK: vpackusdw 512(%rdx){1to4}, %xmm27, %xmm28
+// CHECK:  encoding: [0x62,0x62,0x25,0x10,0x2b,0xa2,0x00,0x02,0x00,0x00]
+          vpackusdw 512(%rdx){1to4}, %xmm27, %xmm28
+
+// CHECK: vpackusdw -512(%rdx){1to4}, %xmm27, %xmm28
+// CHECK:  encoding: [0x62,0x62,0x25,0x10,0x2b,0x62,0x80]
+          vpackusdw -512(%rdx){1to4}, %xmm27, %xmm28
+
+// CHECK: vpackusdw -516(%rdx){1to4}, %xmm27, %xmm28
+// CHECK:  encoding: [0x62,0x62,0x25,0x10,0x2b,0xa2,0xfc,0xfd,0xff,0xff]
+          vpackusdw -516(%rdx){1to4}, %xmm27, %xmm28
+
+// CHECK: vpackusdw %ymm17, %ymm26, %ymm23
+// CHECK:  encoding: [0x62,0xa2,0x2d,0x20,0x2b,0xf9]
+          vpackusdw %ymm17, %ymm26, %ymm23
+
+// CHECK: vpackusdw %ymm17, %ymm26, %ymm23 {%k4}
+// CHECK:  encoding: [0x62,0xa2,0x2d,0x24,0x2b,0xf9]
+          vpackusdw %ymm17, %ymm26, %ymm23 {%k4}
+
+// CHECK: vpackusdw %ymm17, %ymm26, %ymm23 {%k4} {z}
+// CHECK:  encoding: [0x62,0xa2,0x2d,0xa4,0x2b,0xf9]
+          vpackusdw %ymm17, %ymm26, %ymm23 {%k4} {z}
+
+// CHECK: vpackusdw (%rcx), %ymm26, %ymm23
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x20,0x2b,0x39]
+          vpackusdw (%rcx), %ymm26, %ymm23
+
+// CHECK: vpackusdw 4660(%rax,%r14,8), %ymm26, %ymm23
+// CHECK:  encoding: [0x62,0xa2,0x2d,0x20,0x2b,0xbc,0xf0,0x34,0x12,0x00,0x00]
+          vpackusdw 4660(%rax,%r14,8), %ymm26, %ymm23
+
+// CHECK: vpackusdw (%rcx){1to8}, %ymm26, %ymm23
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x30,0x2b,0x39]
+          vpackusdw (%rcx){1to8}, %ymm26, %ymm23
+
+// CHECK: vpackusdw 4064(%rdx), %ymm26, %ymm23
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x20,0x2b,0x7a,0x7f]
+          vpackusdw 4064(%rdx), %ymm26, %ymm23
+
+// CHECK: vpackusdw 4096(%rdx), %ymm26, %ymm23
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x20,0x2b,0xba,0x00,0x10,0x00,0x00]
+          vpackusdw 4096(%rdx), %ymm26, %ymm23
+
+// CHECK: vpackusdw -4096(%rdx), %ymm26, %ymm23
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x20,0x2b,0x7a,0x80]
+          vpackusdw -4096(%rdx), %ymm26, %ymm23
+
+// CHECK: vpackusdw -4128(%rdx), %ymm26, %ymm23
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x20,0x2b,0xba,0xe0,0xef,0xff,0xff]
+          vpackusdw -4128(%rdx), %ymm26, %ymm23
+
+// CHECK: vpackusdw 508(%rdx){1to8}, %ymm26, %ymm23
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x30,0x2b,0x7a,0x7f]
+          vpackusdw 508(%rdx){1to8}, %ymm26, %ymm23
+
+// CHECK: vpackusdw 512(%rdx){1to8}, %ymm26, %ymm23
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x30,0x2b,0xba,0x00,0x02,0x00,0x00]
+          vpackusdw 512(%rdx){1to8}, %ymm26, %ymm23
+
+// CHECK: vpackusdw -512(%rdx){1to8}, %ymm26, %ymm23
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x30,0x2b,0x7a,0x80]
+          vpackusdw -512(%rdx){1to8}, %ymm26, %ymm23
+
+// CHECK: vpackusdw -516(%rdx){1to8}, %ymm26, %ymm23
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x30,0x2b,0xba,0xfc,0xfd,0xff,0xff]
+          vpackusdw -516(%rdx){1to8}, %ymm26, %ymm23
+
+// CHECK: vpackuswb %xmm23, %xmm20, %xmm20
+// CHECK:  encoding: [0x62,0xa1,0xdd,0x00,0x67,0xe7]
+          vpackuswb %xmm23, %xmm20, %xmm20
+
+// CHECK: vpackuswb %xmm23, %xmm20, %xmm20 {%k2}
+// CHECK:  encoding: [0x62,0xa1,0xdd,0x02,0x67,0xe7]
+          vpackuswb %xmm23, %xmm20, %xmm20 {%k2}
+
+// CHECK: vpackuswb %xmm23, %xmm20, %xmm20 {%k2} {z}
+// CHECK:  encoding: [0x62,0xa1,0xdd,0x82,0x67,0xe7]
+          vpackuswb %xmm23, %xmm20, %xmm20 {%k2} {z}
+
+// CHECK: vpackuswb (%rcx), %xmm20, %xmm20
+// CHECK:  encoding: [0x62,0xe1,0xdd,0x00,0x67,0x21]
+          vpackuswb (%rcx), %xmm20, %xmm20
+
+// CHECK: vpackuswb 4660(%rax,%r14,8), %xmm20, %xmm20
+// CHECK:  encoding: [0x62,0xa1,0xdd,0x00,0x67,0xa4,0xf0,0x34,0x12,0x00,0x00]
+          vpackuswb 4660(%rax,%r14,8), %xmm20, %xmm20
+
+// CHECK: vpackuswb 2032(%rdx), %xmm20, %xmm20
+// CHECK:  encoding: [0x62,0xe1,0xdd,0x00,0x67,0x62,0x7f]
+          vpackuswb 2032(%rdx), %xmm20, %xmm20
+
+// CHECK: vpackuswb 2048(%rdx), %xmm20, %xmm20
+// CHECK:  encoding: [0x62,0xe1,0xdd,0x00,0x67,0xa2,0x00,0x08,0x00,0x00]
+          vpackuswb 2048(%rdx), %xmm20, %xmm20
+
+// CHECK: vpackuswb -2048(%rdx), %xmm20, %xmm20
+// CHECK:  encoding: [0x62,0xe1,0xdd,0x00,0x67,0x62,0x80]
+          vpackuswb -2048(%rdx), %xmm20, %xmm20
+
+// CHECK: vpackuswb -2064(%rdx), %xmm20, %xmm20
+// CHECK:  encoding: [0x62,0xe1,0xdd,0x00,0x67,0xa2,0xf0,0xf7,0xff,0xff]
+          vpackuswb -2064(%rdx), %xmm20, %xmm20
+
+// CHECK: vpackuswb %ymm27, %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0x81,0xe5,0x20,0x67,0xe3]
+          vpackuswb %ymm27, %ymm19, %ymm20
+
+// CHECK: vpackuswb %ymm27, %ymm19, %ymm20 {%k5}
+// CHECK:  encoding: [0x62,0x81,0xe5,0x25,0x67,0xe3]
+          vpackuswb %ymm27, %ymm19, %ymm20 {%k5}
+
+// CHECK: vpackuswb %ymm27, %ymm19, %ymm20 {%k5} {z}
+// CHECK:  encoding: [0x62,0x81,0xe5,0xa5,0x67,0xe3]
+          vpackuswb %ymm27, %ymm19, %ymm20 {%k5} {z}
+
+// CHECK: vpackuswb (%rcx), %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x20,0x67,0x21]
+          vpackuswb (%rcx), %ymm19, %ymm20
+
+// CHECK: vpackuswb 4660(%rax,%r14,8), %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xa1,0xe5,0x20,0x67,0xa4,0xf0,0x34,0x12,0x00,0x00]
+          vpackuswb 4660(%rax,%r14,8), %ymm19, %ymm20
+
+// CHECK: vpackuswb 4064(%rdx), %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x20,0x67,0x62,0x7f]
+          vpackuswb 4064(%rdx), %ymm19, %ymm20
+
+// CHECK: vpackuswb 4096(%rdx), %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x20,0x67,0xa2,0x00,0x10,0x00,0x00]
+          vpackuswb 4096(%rdx), %ymm19, %ymm20
+
+// CHECK: vpackuswb -4096(%rdx), %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x20,0x67,0x62,0x80]
+          vpackuswb -4096(%rdx), %ymm19, %ymm20
+
+// CHECK: vpackuswb -4128(%rdx), %ymm19, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0xe5,0x20,0x67,0xa2,0xe0,0xef,0xff,0xff]
+          vpackuswb -4128(%rdx), %ymm19, %ymm20