[AVX512] add Average packed unsigned integers

[AsafBadouh]

Summary

VPAVGB and VPAVGW
include intrinsics and testing.

[delena]

Ok.

include/llvm/IR/IntrinsicsX86.td

def int_x86_avx512_mask_pmull_q_128 : GCCBuiltin<"__builtin_ia32_pmullq128_mask">,
               Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty,
                      llvm_v2i64_ty, llvm_i8_ty], [IntrNoMem]>;
   def int_x86_avx512_mask_pmull_q_256 : GCCBuiltin<"__builtin_ia32_pmullq256_mask">,
               Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty, llvm_v4i64_ty,
                      llvm_v4i64_ty, llvm_i8_ty], [IntrNoMem]>;
   def int_x86_avx512_mask_pmull_q_512 : GCCBuiltin<"__builtin_ia32_pmullq512_mask">,
               Intrinsic<[llvm_v8i64_ty], [llvm_v8i64_ty, llvm_v8i64_ty,
                      llvm_v8i64_ty, llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_pavg_b_512 : GCCBuiltin<"__builtin_ia32_pavgb512_mask">,
+          Intrinsic<[llvm_v64i8_ty], [llvm_v64i8_ty, llvm_v64i8_ty, 
+                    llvm_v64i8_ty, llvm_i64_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_pavg_w_512 : GCCBuiltin<"__builtin_ia32_pavgw512_mask">,
+          Intrinsic<[llvm_v32i16_ty], [llvm_v32i16_ty, llvm_v32i16_ty,  
+                    llvm_v32i16_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_pavg_b_128 : GCCBuiltin<"__builtin_ia32_pavgb128_mask">,
+          Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty,  
+                    llvm_v16i8_ty, llvm_i16_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_pavg_b_256 : GCCBuiltin<"__builtin_ia32_pavgb256_mask">,
+          Intrinsic<[llvm_v32i8_ty], [llvm_v32i8_ty, llvm_v32i8_ty,  
+                    llvm_v32i8_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_pavg_w_128 : GCCBuiltin<"__builtin_ia32_pavgw128_mask">,
+          Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_v8i16_ty,  
+                    llvm_v8i16_ty, llvm_i8_ty], [IntrNoMem]>;
+  def int_x86_avx512_mask_pavg_w_256 : GCCBuiltin<"__builtin_ia32_pavgw256_mask">,
+          Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty, llvm_v16i16_ty, 
+                    llvm_v16i16_ty, llvm_i16_ty], [IntrNoMem]>;
 }
 
 // Gather and Scatter ops
 let TargetPrefix = "x86" in {
   def int_x86_avx512_gather_dpd_512  : GCCBuiltin<"__builtin_ia32_gathersiv8df">,
           Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_ptr_ty,
                      llvm_v8i32_ty, llvm_i8_ty, llvm_i32_ty],
                     [IntrReadArgMem]>;

lib/Target/X86/X86ISelLowering.h

enum NodeType : unsigned {
       FGETEXP_RND,
 
       // Integer add/sub with unsigned saturation.
       ADDUS,
       SUBUS,
       // Integer add/sub with signed saturation.
       ADDS,
       SUBS,
-
+      // Unsigned Integer average 
+      AVG,
       /// Integer horizontal add.
       HADD,
 
       /// Integer horizontal sub.
       HSUB,
 
       /// Floating point horizontal add.
       FHADD,

lib/Target/X86/X86InstrAVX512.td

defm VPSUBUS : avx512_binop_rm_vl_bw<0xD8, 0xD9, "vpsubus", X86subus,
                                     SSE_INTALU_ITINS_P, HasBWI, 0>;
 defm VPMULLD : avx512_binop_rm_vl_d<0x40, "vpmull", mul,
                                    SSE_INTALU_ITINS_P, HasAVX512, 1>, T8PD;
 defm VPMULLW : avx512_binop_rm_vl_w<0xD5, "vpmull", mul,
                                    SSE_INTALU_ITINS_P, HasBWI, 1>;
 defm VPMULLQ : avx512_binop_rm_vl_q<0x40, "vpmull", mul,
                                    SSE_INTALU_ITINS_P, HasDQI, 1>, T8PD;
 
+defm VPAVG : avx512_binop_rm_vl_bw<0xE0, 0xE3, "vpavg", X86avg,
+                                    SSE_INTALU_ITINS_P, HasBWI, 1>;
                                    
 multiclass avx512_binop_all<bits<8> opc, string OpcodeStr, OpndItins itins,
                             SDNode OpNode, bit IsCommutable = 0> {
 
   defm NAME#Z : avx512_binop_rm2<opc, OpcodeStr, itins, OpNode,
                                  v16i32_info, v8i64_info, IsCommutable>,
                                 EVEX_V512, EVEX_CD8<64, CD8VF>, VEX_W;
   let Predicates = [HasVLX] in {

lib/Target/X86/X86InstrFragmentsSIMD.td

 
 def SDTX86CmpPTest : SDTypeProfile<1, 2, [SDTCisVT<0, i32>,
                                           SDTCisVec<1>,
                                           SDTCisSameAs<2, 1>]>;
 def X86addus   : SDNode<"X86ISD::ADDUS", SDTIntBinOp>;
 def X86subus   : SDNode<"X86ISD::SUBUS", SDTIntBinOp>;
 def X86adds    : SDNode<"X86ISD::ADDS", SDTIntBinOp>;
 def X86subs    : SDNode<"X86ISD::SUBS", SDTIntBinOp>;
+def X86avg     : SDNode<"X86ISD::AVG" , SDTIntBinOp>;
 def X86ptest   : SDNode<"X86ISD::PTEST", SDTX86CmpPTest>;
 def X86testp   : SDNode<"X86ISD::TESTP", SDTX86CmpPTest>;
 def X86kortest : SDNode<"X86ISD::KORTEST", SDTX86CmpPTest>;
 def X86testm   : SDNode<"X86ISD::TESTM", SDTypeProfile<1, 2, [SDTCisVec<0>,
                                           SDTCisVec<1>, SDTCisSameAs<2, 1>,
                                           SDTCVecEltisVT<0, i1>,
                                           SDTCisSameNumEltsAs<0, 1>]>>;
 def X86testnm  : SDNode<"X86ISD::TESTNM", SDTypeProfile<1, 2, [SDTCisVec<0>,

lib/Target/X86/X86IntrinsicsInfo.h

static const IntrinsicData  IntrinsicsWithoutChain[] = {
   X86_INTRINSIC_DATA(avx512_mask_pand_q_256, INTR_TYPE_2OP_MASK, ISD::AND, 0),
   X86_INTRINSIC_DATA(avx512_mask_pand_q_512, INTR_TYPE_2OP_MASK, ISD::AND, 0),
   X86_INTRINSIC_DATA(avx512_mask_pandn_d_128, INTR_TYPE_2OP_MASK, X86ISD::ANDNP, 0),
   X86_INTRINSIC_DATA(avx512_mask_pandn_d_256, INTR_TYPE_2OP_MASK, X86ISD::ANDNP, 0),
   X86_INTRINSIC_DATA(avx512_mask_pandn_d_512, INTR_TYPE_2OP_MASK, X86ISD::ANDNP, 0),
   X86_INTRINSIC_DATA(avx512_mask_pandn_q_128, INTR_TYPE_2OP_MASK, X86ISD::ANDNP, 0),
   X86_INTRINSIC_DATA(avx512_mask_pandn_q_256, INTR_TYPE_2OP_MASK, X86ISD::ANDNP, 0),
   X86_INTRINSIC_DATA(avx512_mask_pandn_q_512, INTR_TYPE_2OP_MASK, X86ISD::ANDNP, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pavg_b_128, INTR_TYPE_2OP_MASK, X86ISD::AVG, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pavg_b_256, INTR_TYPE_2OP_MASK, X86ISD::AVG, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pavg_b_512, INTR_TYPE_2OP_MASK, X86ISD::AVG, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pavg_w_128, INTR_TYPE_2OP_MASK, X86ISD::AVG, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pavg_w_256, INTR_TYPE_2OP_MASK, X86ISD::AVG, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pavg_w_512, INTR_TYPE_2OP_MASK, X86ISD::AVG, 0),
   X86_INTRINSIC_DATA(avx512_mask_pcmpeq_b_128,  CMP_MASK,  X86ISD::PCMPEQM, 0),
   X86_INTRINSIC_DATA(avx512_mask_pcmpeq_b_256,  CMP_MASK,  X86ISD::PCMPEQM, 0),
   X86_INTRINSIC_DATA(avx512_mask_pcmpeq_b_512,  CMP_MASK,  X86ISD::PCMPEQM, 0),
   X86_INTRINSIC_DATA(avx512_mask_pcmpeq_d_128,  CMP_MASK,  X86ISD::PCMPEQM, 0),
   X86_INTRINSIC_DATA(avx512_mask_pcmpeq_d_256,  CMP_MASK,  X86ISD::PCMPEQM, 0),
   X86_INTRINSIC_DATA(avx512_mask_pcmpeq_d_512,  CMP_MASK,  X86ISD::PCMPEQM, 0),
   X86_INTRINSIC_DATA(avx512_mask_pcmpeq_q_128,  CMP_MASK,  X86ISD::PCMPEQM, 0),
   X86_INTRINSIC_DATA(avx512_mask_pcmpeq_q_256,  CMP_MASK,  X86ISD::PCMPEQM, 0),

test/CodeGen/X86/avx512bw-intrinsics.ll

 ; CHECK: vpminuw %zmm
 ; CHECK: {%k1} 
 define <32 x i16>@test_int_x86_avx512_mask_pminu_w_512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3) {
   %res = call <32 x i16> @llvm.x86.avx512.mask.pminu.w.512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3)
   %res1 = call <32 x i16> @llvm.x86.avx512.mask.pminu.w.512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 -1)
   %res2 = add <32 x i16> %res, %res1
   ret <32 x i16> %res2
 }
+
+declare <64 x i8> @llvm.x86.avx512.mask.pavg.b.512(<64 x i8>, <64 x i8>, <64 x i8>, i64)
+
+; CHECK-LABEL: @test_int_x86_avx512_mask_pavg_b_512
+; CHECK-NOT: call 
+; CHECK: vpavgb %zmm
+; CHECK: {%k1} 
+define <64 x i8>@test_int_x86_avx512_mask_pavg_b_512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3) {
+  %res = call <64 x i8> @llvm.x86.avx512.mask.pavg.b.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3)
+  %res1 = call <64 x i8> @llvm.x86.avx512.mask.pavg.b.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 -1)
+  %res2 = add <64 x i8> %res, %res1
+  ret <64 x i8> %res2
+}
+
+declare <32 x i16> @llvm.x86.avx512.mask.pavg.w.512(<32 x i16>, <32 x i16>, <32 x i16>, i32)
+
+; CHECK-LABEL: @test_int_x86_avx512_mask_pavg_w_512
+; CHECK-NOT: call 
+; CHECK: vpavgw %zmm
+; CHECK: {%k1} 
+define <32 x i16>@test_int_x86_avx512_mask_pavg_w_512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3) {
+  %res = call <32 x i16> @llvm.x86.avx512.mask.pavg.w.512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3)
+  %res1 = call <32 x i16> @llvm.x86.avx512.mask.pavg.w.512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 -1)
+  %res2 = add <32 x i16> %res, %res1
+  ret <32 x i16> %res2
+}
 No newline at end of file

test/CodeGen/X86/avx512bwvl-intrinsics.ll

 ; CHECK: vpminuw %ymm
 ; CHECK: {%k1} 
 define <16 x i16>@test_int_x86_avx512_mask_pminu_w_256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %mask) {
   %res = call <16 x i16> @llvm.x86.avx512.mask.pminu.w.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %mask)
   %res1 = call <16 x i16> @llvm.x86.avx512.mask.pminu.w.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> zeroinitializer, i16 %mask)
   %res2 = add <16 x i16> %res, %res1
   ret <16 x i16> %res2
 }
+
+declare <16 x i8> @llvm.x86.avx512.mask.pavg.b.128(<16 x i8>, <16 x i8>, <16 x i8>, i16)
+
+; CHECK-LABEL: @test_int_x86_avx512_mask_pavg_b_128
+; CHECK-NOT: call 
+; CHECK: vpavgb %xmm
+; CHECK: {%k1} 
+define <16 x i8>@test_int_x86_avx512_mask_pavg_b_128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %x3) {
+  %res = call <16 x i8> @llvm.x86.avx512.mask.pavg.b.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %x3)
+  %res1 = call <16 x i8> @llvm.x86.avx512.mask.pavg.b.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 -1)
+  %res2 = add <16 x i8> %res, %res1
+  ret <16 x i8> %res2
+}
+
+declare <32 x i8> @llvm.x86.avx512.mask.pavg.b.256(<32 x i8>, <32 x i8>, <32 x i8>, i32)
+
+; CHECK-LABEL: @test_int_x86_avx512_mask_pavg_b_256
+; CHECK-NOT: call 
+; CHECK: vpavgb %ymm
+; CHECK: {%k1} 
+define <32 x i8>@test_int_x86_avx512_mask_pavg_b_256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3) {
+  %res = call <32 x i8> @llvm.x86.avx512.mask.pavg.b.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3)
+  %res1 = call <32 x i8> @llvm.x86.avx512.mask.pavg.b.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 -1)
+  %res2 = add <32 x i8> %res, %res1
+  ret <32 x i8> %res2
+}
+
+declare <8 x i16> @llvm.x86.avx512.mask.pavg.w.128(<8 x i16>, <8 x i16>, <8 x i16>, i8)
+
+; CHECK-LABEL: @test_int_x86_avx512_mask_pavg_w_128
+; CHECK-NOT: call 
+; CHECK: vpavgw %xmm
+; CHECK: {%k1} 
+define <8 x i16>@test_int_x86_avx512_mask_pavg_w_128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3) {
+  %res = call <8 x i16> @llvm.x86.avx512.mask.pavg.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 %x3)
+  %res1 = call <8 x i16> @llvm.x86.avx512.mask.pavg.w.128(<8 x i16> %x0, <8 x i16> %x1, <8 x i16> %x2, i8 -1)
+  %res2 = add <8 x i16> %res, %res1
+  ret <8 x i16> %res2
+}
+
+declare <16 x i16> @llvm.x86.avx512.mask.pavg.w.256(<16 x i16>, <16 x i16>, <16 x i16>, i16)
+
+; CHECK-LABEL: @test_int_x86_avx512_mask_pavg_w_256
+; CHECK-NOT: call 
+; CHECK: vpavgw %ymm
+; CHECK: {%k1} 
+define <16 x i16>@test_int_x86_avx512_mask_pavg_w_256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %x3) {
+  %res = call <16 x i16> @llvm.x86.avx512.mask.pavg.w.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 %x3)
+  %res1 = call <16 x i16> @llvm.x86.avx512.mask.pavg.w.256(<16 x i16> %x0, <16 x i16> %x1, <16 x i16> %x2, i16 -1)
+  %res2 = add <16 x i16> %res, %res1
+  ret <16 x i16> %res2
+}

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

 
 // CHECK: vpermw -8192(%rdx), %zmm19, %zmm22
 // CHECK:  encoding: [0x62,0xe2,0xe5,0x40,0x8d,0x72,0x80]
           vpermw -8192(%rdx), %zmm19, %zmm22
 
 // CHECK: vpermw -8256(%rdx), %zmm19, %zmm22
 // CHECK:  encoding: [0x62,0xe2,0xe5,0x40,0x8d,0xb2,0xc0,0xdf,0xff,0xff]
           vpermw -8256(%rdx), %zmm19, %zmm22
+
+// CHECK: vpavgb %zmm21, %zmm29, %zmm17
+// CHECK:  encoding: [0x62,0xa1,0x15,0x40,0xe0,0xcd]
+          vpavgb %zmm21, %zmm29, %zmm17
+
+// CHECK: vpavgb %zmm21, %zmm29, %zmm17 {%k2}
+// CHECK:  encoding: [0x62,0xa1,0x15,0x42,0xe0,0xcd]
+          vpavgb %zmm21, %zmm29, %zmm17 {%k2}
+
+// CHECK: vpavgb %zmm21, %zmm29, %zmm17 {%k2} {z}
+// CHECK:  encoding: [0x62,0xa1,0x15,0xc2,0xe0,0xcd]
+          vpavgb %zmm21, %zmm29, %zmm17 {%k2} {z}
+
+// CHECK: vpavgb (%rcx), %zmm29, %zmm17
+// CHECK:  encoding: [0x62,0xe1,0x15,0x40,0xe0,0x09]
+          vpavgb (%rcx), %zmm29, %zmm17
+
+// CHECK: vpavgb 291(%rax,%r14,8), %zmm29, %zmm17
+// CHECK:  encoding: [0x62,0xa1,0x15,0x40,0xe0,0x8c,0xf0,0x23,0x01,0x00,0x00]
+          vpavgb 291(%rax,%r14,8), %zmm29, %zmm17
+
+// CHECK: vpavgb 8128(%rdx), %zmm29, %zmm17
+// CHECK:  encoding: [0x62,0xe1,0x15,0x40,0xe0,0x4a,0x7f]
+          vpavgb 8128(%rdx), %zmm29, %zmm17
+
+// CHECK: vpavgb 8192(%rdx), %zmm29, %zmm17
+// CHECK:  encoding: [0x62,0xe1,0x15,0x40,0xe0,0x8a,0x00,0x20,0x00,0x00]
+          vpavgb 8192(%rdx), %zmm29, %zmm17
+
+// CHECK: vpavgb -8192(%rdx), %zmm29, %zmm17
+// CHECK:  encoding: [0x62,0xe1,0x15,0x40,0xe0,0x4a,0x80]
+          vpavgb -8192(%rdx), %zmm29, %zmm17
+
+// CHECK: vpavgb -8256(%rdx), %zmm29, %zmm17
+// CHECK:  encoding: [0x62,0xe1,0x15,0x40,0xe0,0x8a,0xc0,0xdf,0xff,0xff]
+          vpavgb -8256(%rdx), %zmm29, %zmm17
+
+// CHECK: vpavgw %zmm22, %zmm27, %zmm19
+// CHECK:  encoding: [0x62,0xa1,0x25,0x40,0xe3,0xde]
+          vpavgw %zmm22, %zmm27, %zmm19
+
+// CHECK: vpavgw %zmm22, %zmm27, %zmm19 {%k4}
+// CHECK:  encoding: [0x62,0xa1,0x25,0x44,0xe3,0xde]
+          vpavgw %zmm22, %zmm27, %zmm19 {%k4}
+
+// CHECK: vpavgw %zmm22, %zmm27, %zmm19 {%k4} {z}
+// CHECK:  encoding: [0x62,0xa1,0x25,0xc4,0xe3,0xde]
+          vpavgw %zmm22, %zmm27, %zmm19 {%k4} {z}
+
+// CHECK: vpavgw (%rcx), %zmm27, %zmm19
+// CHECK:  encoding: [0x62,0xe1,0x25,0x40,0xe3,0x19]
+          vpavgw (%rcx), %zmm27, %zmm19
+
+// CHECK: vpavgw 291(%rax,%r14,8), %zmm27, %zmm19
+// CHECK:  encoding: [0x62,0xa1,0x25,0x40,0xe3,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vpavgw 291(%rax,%r14,8), %zmm27, %zmm19
+
+// CHECK: vpavgw 8128(%rdx), %zmm27, %zmm19
+// CHECK:  encoding: [0x62,0xe1,0x25,0x40,0xe3,0x5a,0x7f]
+          vpavgw 8128(%rdx), %zmm27, %zmm19
+
+// CHECK: vpavgw 8192(%rdx), %zmm27, %zmm19
+// CHECK:  encoding: [0x62,0xe1,0x25,0x40,0xe3,0x9a,0x00,0x20,0x00,0x00]
+          vpavgw 8192(%rdx), %zmm27, %zmm19
+
+// CHECK: vpavgw -8192(%rdx), %zmm27, %zmm19
+// CHECK:  encoding: [0x62,0xe1,0x25,0x40,0xe3,0x5a,0x80]
+          vpavgw -8192(%rdx), %zmm27, %zmm19
+
+// CHECK: vpavgw -8256(%rdx), %zmm27, %zmm19
+// CHECK:  encoding: [0x62,0xe1,0x25,0x40,0xe3,0x9a,0xc0,0xdf,0xff,0xff]
+          vpavgw -8256(%rdx), %zmm27, %zmm19
+
+// CHECK: vpavgb %zmm19, %zmm27, %zmm24
+// CHECK:  encoding: [0x62,0x21,0x25,0x40,0xe0,0xc3]
+          vpavgb %zmm19, %zmm27, %zmm24
+
+// CHECK: vpavgb %zmm19, %zmm27, %zmm24 {%k4}
+// CHECK:  encoding: [0x62,0x21,0x25,0x44,0xe0,0xc3]
+          vpavgb %zmm19, %zmm27, %zmm24 {%k4}
+
+// CHECK: vpavgb %zmm19, %zmm27, %zmm24 {%k4} {z}
+// CHECK:  encoding: [0x62,0x21,0x25,0xc4,0xe0,0xc3]
+          vpavgb %zmm19, %zmm27, %zmm24 {%k4} {z}
+
+// CHECK: vpavgb (%rcx), %zmm27, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x25,0x40,0xe0,0x01]
+          vpavgb (%rcx), %zmm27, %zmm24
+
+// CHECK: vpavgb 291(%rax,%r14,8), %zmm27, %zmm24
+// CHECK:  encoding: [0x62,0x21,0x25,0x40,0xe0,0x84,0xf0,0x23,0x01,0x00,0x00]
+          vpavgb 291(%rax,%r14,8), %zmm27, %zmm24
+
+// CHECK: vpavgb 8128(%rdx), %zmm27, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x25,0x40,0xe0,0x42,0x7f]
+          vpavgb 8128(%rdx), %zmm27, %zmm24
+
+// CHECK: vpavgb 8192(%rdx), %zmm27, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x25,0x40,0xe0,0x82,0x00,0x20,0x00,0x00]
+          vpavgb 8192(%rdx), %zmm27, %zmm24
+
+// CHECK: vpavgb -8192(%rdx), %zmm27, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x25,0x40,0xe0,0x42,0x80]
+          vpavgb -8192(%rdx), %zmm27, %zmm24
+
+// CHECK: vpavgb -8256(%rdx), %zmm27, %zmm24
+// CHECK:  encoding: [0x62,0x61,0x25,0x40,0xe0,0x82,0xc0,0xdf,0xff,0xff]
+          vpavgb -8256(%rdx), %zmm27, %zmm24
+
+// CHECK: vpavgw %zmm26, %zmm29, %zmm29
+// CHECK:  encoding: [0x62,0x01,0x15,0x40,0xe3,0xea]
+          vpavgw %zmm26, %zmm29, %zmm29
+
+// CHECK: vpavgw %zmm26, %zmm29, %zmm29 {%k6}
+// CHECK:  encoding: [0x62,0x01,0x15,0x46,0xe3,0xea]
+          vpavgw %zmm26, %zmm29, %zmm29 {%k6}
+
+// CHECK: vpavgw %zmm26, %zmm29, %zmm29 {%k6} {z}
+// CHECK:  encoding: [0x62,0x01,0x15,0xc6,0xe3,0xea]
+          vpavgw %zmm26, %zmm29, %zmm29 {%k6} {z}
+
+// CHECK: vpavgw (%rcx), %zmm29, %zmm29
+// CHECK:  encoding: [0x62,0x61,0x15,0x40,0xe3,0x29]
+          vpavgw (%rcx), %zmm29, %zmm29
+
+// CHECK: vpavgw 291(%rax,%r14,8), %zmm29, %zmm29
+// CHECK:  encoding: [0x62,0x21,0x15,0x40,0xe3,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vpavgw 291(%rax,%r14,8), %zmm29, %zmm29
+
+// CHECK: vpavgw 8128(%rdx), %zmm29, %zmm29
+// CHECK:  encoding: [0x62,0x61,0x15,0x40,0xe3,0x6a,0x7f]
+          vpavgw 8128(%rdx), %zmm29, %zmm29
+
+// CHECK: vpavgw 8192(%rdx), %zmm29, %zmm29
+// CHECK:  encoding: [0x62,0x61,0x15,0x40,0xe3,0xaa,0x00,0x20,0x00,0x00]
+          vpavgw 8192(%rdx), %zmm29, %zmm29
+
+// CHECK: vpavgw -8192(%rdx), %zmm29, %zmm29
+// CHECK:  encoding: [0x62,0x61,0x15,0x40,0xe3,0x6a,0x80]
+          vpavgw -8192(%rdx), %zmm29, %zmm29
+
+// CHECK: vpavgw -8256(%rdx), %zmm29, %zmm29
+// CHECK:  encoding: [0x62,0x61,0x15,0x40,0xe3,0xaa,0xc0,0xdf,0xff,0xff]
+          vpavgw -8256(%rdx), %zmm29, %zmm29
+

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

 // CHECK: vpshuflw $123, -4096(%rdx), %ymm25
 // CHECK:  encoding: [0x62,0x61,0xff,0x28,0x70,0x4a,0x80,0x7b]
           vpshuflw $123, -4096(%rdx), %ymm25
 
 // CHECK: vpshuflw $123, -4128(%rdx), %ymm25
 // CHECK:  encoding: [0x62,0x61,0xff,0x28,0x70,0x8a,0xe0,0xef,0xff,0xff,0x7b]
           vpshuflw $123, -4128(%rdx), %ymm25
 
+// CHECK: vpavgb %xmm22, %xmm24, %xmm21
+// CHECK:  encoding: [0x62,0xa1,0x3d,0x00,0xe0,0xee]
+          vpavgb %xmm22, %xmm24, %xmm21
+
+// CHECK: vpavgb %xmm22, %xmm24, %xmm21 {%k7}
+// CHECK:  encoding: [0x62,0xa1,0x3d,0x07,0xe0,0xee]
+          vpavgb %xmm22, %xmm24, %xmm21 {%k7}
+
+// CHECK: vpavgb %xmm22, %xmm24, %xmm21 {%k7} {z}
+// CHECK:  encoding: [0x62,0xa1,0x3d,0x87,0xe0,0xee]
+          vpavgb %xmm22, %xmm24, %xmm21 {%k7} {z}
+
+// CHECK: vpavgb (%rcx), %xmm24, %xmm21
+// CHECK:  encoding: [0x62,0xe1,0x3d,0x00,0xe0,0x29]
+          vpavgb (%rcx), %xmm24, %xmm21
+
+// CHECK: vpavgb 4660(%rax,%r14,8), %xmm24, %xmm21
+// CHECK:  encoding: [0x62,0xa1,0x3d,0x00,0xe0,0xac,0xf0,0x34,0x12,0x00,0x00]
+          vpavgb 4660(%rax,%r14,8), %xmm24, %xmm21
+
+// CHECK: vpavgb 2032(%rdx), %xmm24, %xmm21
+// CHECK:  encoding: [0x62,0xe1,0x3d,0x00,0xe0,0x6a,0x7f]
+          vpavgb 2032(%rdx), %xmm24, %xmm21
+
+// CHECK: vpavgb 2048(%rdx), %xmm24, %xmm21
+// CHECK:  encoding: [0x62,0xe1,0x3d,0x00,0xe0,0xaa,0x00,0x08,0x00,0x00]
+          vpavgb 2048(%rdx), %xmm24, %xmm21
+
+// CHECK: vpavgb -2048(%rdx), %xmm24, %xmm21
+// CHECK:  encoding: [0x62,0xe1,0x3d,0x00,0xe0,0x6a,0x80]
+          vpavgb -2048(%rdx), %xmm24, %xmm21
+
+// CHECK: vpavgb -2064(%rdx), %xmm24, %xmm21
+// CHECK:  encoding: [0x62,0xe1,0x3d,0x00,0xe0,0xaa,0xf0,0xf7,0xff,0xff]
+          vpavgb -2064(%rdx), %xmm24, %xmm21
+
+// CHECK: vpavgb %ymm18, %ymm18, %ymm26
+// CHECK:  encoding: [0x62,0x21,0x6d,0x20,0xe0,0xd2]
+          vpavgb %ymm18, %ymm18, %ymm26
+
+// CHECK: vpavgb %ymm18, %ymm18, %ymm26 {%k4}
+// CHECK:  encoding: [0x62,0x21,0x6d,0x24,0xe0,0xd2]
+          vpavgb %ymm18, %ymm18, %ymm26 {%k4}
+
+// CHECK: vpavgb %ymm18, %ymm18, %ymm26 {%k4} {z}
+// CHECK:  encoding: [0x62,0x21,0x6d,0xa4,0xe0,0xd2]
+          vpavgb %ymm18, %ymm18, %ymm26 {%k4} {z}
+
+// CHECK: vpavgb (%rcx), %ymm18, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x6d,0x20,0xe0,0x11]
+          vpavgb (%rcx), %ymm18, %ymm26
+
+// CHECK: vpavgb 4660(%rax,%r14,8), %ymm18, %ymm26
+// CHECK:  encoding: [0x62,0x21,0x6d,0x20,0xe0,0x94,0xf0,0x34,0x12,0x00,0x00]
+          vpavgb 4660(%rax,%r14,8), %ymm18, %ymm26
+
+// CHECK: vpavgb 4064(%rdx), %ymm18, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x6d,0x20,0xe0,0x52,0x7f]
+          vpavgb 4064(%rdx), %ymm18, %ymm26
+
+// CHECK: vpavgb 4096(%rdx), %ymm18, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x6d,0x20,0xe0,0x92,0x00,0x10,0x00,0x00]
+          vpavgb 4096(%rdx), %ymm18, %ymm26
+
+// CHECK: vpavgb -4096(%rdx), %ymm18, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x6d,0x20,0xe0,0x52,0x80]
+          vpavgb -4096(%rdx), %ymm18, %ymm26
+
+// CHECK: vpavgb -4128(%rdx), %ymm18, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x6d,0x20,0xe0,0x92,0xe0,0xef,0xff,0xff]
+          vpavgb -4128(%rdx), %ymm18, %ymm26
+
+// CHECK: vpavgw %xmm23, %xmm28, %xmm29
+// CHECK:  encoding: [0x62,0x21,0x1d,0x00,0xe3,0xef]
+          vpavgw %xmm23, %xmm28, %xmm29
+
+// CHECK: vpavgw %xmm23, %xmm28, %xmm29 {%k7}
+// CHECK:  encoding: [0x62,0x21,0x1d,0x07,0xe3,0xef]
+          vpavgw %xmm23, %xmm28, %xmm29 {%k7}
+
+// CHECK: vpavgw %xmm23, %xmm28, %xmm29 {%k7} {z}
+// CHECK:  encoding: [0x62,0x21,0x1d,0x87,0xe3,0xef]
+          vpavgw %xmm23, %xmm28, %xmm29 {%k7} {z}
+
+// CHECK: vpavgw (%rcx), %xmm28, %xmm29
+// CHECK:  encoding: [0x62,0x61,0x1d,0x00,0xe3,0x29]
+          vpavgw (%rcx), %xmm28, %xmm29
+
+// CHECK: vpavgw 4660(%rax,%r14,8), %xmm28, %xmm29
+// CHECK:  encoding: [0x62,0x21,0x1d,0x00,0xe3,0xac,0xf0,0x34,0x12,0x00,0x00]
+          vpavgw 4660(%rax,%r14,8), %xmm28, %xmm29
+
+// CHECK: vpavgw 2032(%rdx), %xmm28, %xmm29
+// CHECK:  encoding: [0x62,0x61,0x1d,0x00,0xe3,0x6a,0x7f]
+          vpavgw 2032(%rdx), %xmm28, %xmm29
+
+// CHECK: vpavgw 2048(%rdx), %xmm28, %xmm29
+// CHECK:  encoding: [0x62,0x61,0x1d,0x00,0xe3,0xaa,0x00,0x08,0x00,0x00]
+          vpavgw 2048(%rdx), %xmm28, %xmm29
+
+// CHECK: vpavgw -2048(%rdx), %xmm28, %xmm29
+// CHECK:  encoding: [0x62,0x61,0x1d,0x00,0xe3,0x6a,0x80]
+          vpavgw -2048(%rdx), %xmm28, %xmm29
+
+// CHECK: vpavgw -2064(%rdx), %xmm28, %xmm29
+// CHECK:  encoding: [0x62,0x61,0x1d,0x00,0xe3,0xaa,0xf0,0xf7,0xff,0xff]
+          vpavgw -2064(%rdx), %xmm28, %xmm29
+
+// CHECK: vpavgw %ymm17, %ymm18, %ymm27
+// CHECK:  encoding: [0x62,0x21,0x6d,0x20,0xe3,0xd9]
+          vpavgw %ymm17, %ymm18, %ymm27
+
+// CHECK: vpavgw %ymm17, %ymm18, %ymm27 {%k5}
+// CHECK:  encoding: [0x62,0x21,0x6d,0x25,0xe3,0xd9]
+          vpavgw %ymm17, %ymm18, %ymm27 {%k5}
+
+// CHECK: vpavgw %ymm17, %ymm18, %ymm27 {%k5} {z}
+// CHECK:  encoding: [0x62,0x21,0x6d,0xa5,0xe3,0xd9]
+          vpavgw %ymm17, %ymm18, %ymm27 {%k5} {z}
+
+// CHECK: vpavgw (%rcx), %ymm18, %ymm27
+// CHECK:  encoding: [0x62,0x61,0x6d,0x20,0xe3,0x19]
+          vpavgw (%rcx), %ymm18, %ymm27
+
+// CHECK: vpavgw 4660(%rax,%r14,8), %ymm18, %ymm27
+// CHECK:  encoding: [0x62,0x21,0x6d,0x20,0xe3,0x9c,0xf0,0x34,0x12,0x00,0x00]
+          vpavgw 4660(%rax,%r14,8), %ymm18, %ymm27
+
+// CHECK: vpavgw 4064(%rdx), %ymm18, %ymm27
+// CHECK:  encoding: [0x62,0x61,0x6d,0x20,0xe3,0x5a,0x7f]
+          vpavgw 4064(%rdx), %ymm18, %ymm27
+
+// CHECK: vpavgw 4096(%rdx), %ymm18, %ymm27
+// CHECK:  encoding: [0x62,0x61,0x6d,0x20,0xe3,0x9a,0x00,0x10,0x00,0x00]
+          vpavgw 4096(%rdx), %ymm18, %ymm27
+
+// CHECK: vpavgw -4096(%rdx), %ymm18, %ymm27
+// CHECK:  encoding: [0x62,0x61,0x6d,0x20,0xe3,0x5a,0x80]
+          vpavgw -4096(%rdx), %ymm18, %ymm27
+
+// CHECK: vpavgw -4128(%rdx), %ymm18, %ymm27
+// CHECK:  encoding: [0x62,0x61,0x6d,0x20,0xe3,0x9a,0xe0,0xef,0xff,0xff]
+          vpavgw -4128(%rdx), %ymm18, %ymm27
+
+// CHECK: vpavgb %xmm20, %xmm22, %xmm26
+// CHECK:  encoding: [0x62,0x21,0x4d,0x00,0xe0,0xd4]
+          vpavgb %xmm20, %xmm22, %xmm26
+
+// CHECK: vpavgb %xmm20, %xmm22, %xmm26 {%k6}
+// CHECK:  encoding: [0x62,0x21,0x4d,0x06,0xe0,0xd4]
+          vpavgb %xmm20, %xmm22, %xmm26 {%k6}
+
+// CHECK: vpavgb %xmm20, %xmm22, %xmm26 {%k6} {z}
+// CHECK:  encoding: [0x62,0x21,0x4d,0x86,0xe0,0xd4]
+          vpavgb %xmm20, %xmm22, %xmm26 {%k6} {z}
+
+// CHECK: vpavgb (%rcx), %xmm22, %xmm26
+// CHECK:  encoding: [0x62,0x61,0x4d,0x00,0xe0,0x11]
+          vpavgb (%rcx), %xmm22, %xmm26
+
+// CHECK: vpavgb 291(%rax,%r14,8), %xmm22, %xmm26
+// CHECK:  encoding: [0x62,0x21,0x4d,0x00,0xe0,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpavgb 291(%rax,%r14,8), %xmm22, %xmm26
+
+// CHECK: vpavgb 2032(%rdx), %xmm22, %xmm26
+// CHECK:  encoding: [0x62,0x61,0x4d,0x00,0xe0,0x52,0x7f]
+          vpavgb 2032(%rdx), %xmm22, %xmm26
+
+// CHECK: vpavgb 2048(%rdx), %xmm22, %xmm26
+// CHECK:  encoding: [0x62,0x61,0x4d,0x00,0xe0,0x92,0x00,0x08,0x00,0x00]
+          vpavgb 2048(%rdx), %xmm22, %xmm26
+
+// CHECK: vpavgb -2048(%rdx), %xmm22, %xmm26
+// CHECK:  encoding: [0x62,0x61,0x4d,0x00,0xe0,0x52,0x80]
+          vpavgb -2048(%rdx), %xmm22, %xmm26
+
+// CHECK: vpavgb -2064(%rdx), %xmm22, %xmm26
+// CHECK:  encoding: [0x62,0x61,0x4d,0x00,0xe0,0x92,0xf0,0xf7,0xff,0xff]
+          vpavgb -2064(%rdx), %xmm22, %xmm26
+
+// CHECK: vpavgb %ymm18, %ymm22, %ymm29
+// CHECK:  encoding: [0x62,0x21,0x4d,0x20,0xe0,0xea]
+          vpavgb %ymm18, %ymm22, %ymm29
+
+// CHECK: vpavgb %ymm18, %ymm22, %ymm29 {%k1}
+// CHECK:  encoding: [0x62,0x21,0x4d,0x21,0xe0,0xea]
+          vpavgb %ymm18, %ymm22, %ymm29 {%k1}
+
+// CHECK: vpavgb %ymm18, %ymm22, %ymm29 {%k1} {z}
+// CHECK:  encoding: [0x62,0x21,0x4d,0xa1,0xe0,0xea]
+          vpavgb %ymm18, %ymm22, %ymm29 {%k1} {z}
+
+// CHECK: vpavgb (%rcx), %ymm22, %ymm29
+// CHECK:  encoding: [0x62,0x61,0x4d,0x20,0xe0,0x29]
+          vpavgb (%rcx), %ymm22, %ymm29
+
+// CHECK: vpavgb 291(%rax,%r14,8), %ymm22, %ymm29
+// CHECK:  encoding: [0x62,0x21,0x4d,0x20,0xe0,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vpavgb 291(%rax,%r14,8), %ymm22, %ymm29
+
+// CHECK: vpavgb 4064(%rdx), %ymm22, %ymm29
+// CHECK:  encoding: [0x62,0x61,0x4d,0x20,0xe0,0x6a,0x7f]
+          vpavgb 4064(%rdx), %ymm22, %ymm29
+
+// CHECK: vpavgb 4096(%rdx), %ymm22, %ymm29
+// CHECK:  encoding: [0x62,0x61,0x4d,0x20,0xe0,0xaa,0x00,0x10,0x00,0x00]
+          vpavgb 4096(%rdx), %ymm22, %ymm29
+
+// CHECK: vpavgb -4096(%rdx), %ymm22, %ymm29
+// CHECK:  encoding: [0x62,0x61,0x4d,0x20,0xe0,0x6a,0x80]
+          vpavgb -4096(%rdx), %ymm22, %ymm29
+
+// CHECK: vpavgb -4128(%rdx), %ymm22, %ymm29
+// CHECK:  encoding: [0x62,0x61,0x4d,0x20,0xe0,0xaa,0xe0,0xef,0xff,0xff]
+          vpavgb -4128(%rdx), %ymm22, %ymm29
+
+// CHECK: vpavgw %xmm23, %xmm24, %xmm28
+// CHECK:  encoding: [0x62,0x21,0x3d,0x00,0xe3,0xe7]
+          vpavgw %xmm23, %xmm24, %xmm28
+
+// CHECK: vpavgw %xmm23, %xmm24, %xmm28 {%k7}
+// CHECK:  encoding: [0x62,0x21,0x3d,0x07,0xe3,0xe7]
+          vpavgw %xmm23, %xmm24, %xmm28 {%k7}
+
+// CHECK: vpavgw %xmm23, %xmm24, %xmm28 {%k7} {z}
+// CHECK:  encoding: [0x62,0x21,0x3d,0x87,0xe3,0xe7]
+          vpavgw %xmm23, %xmm24, %xmm28 {%k7} {z}
+
+// CHECK: vpavgw (%rcx), %xmm24, %xmm28
+// CHECK:  encoding: [0x62,0x61,0x3d,0x00,0xe3,0x21]
+          vpavgw (%rcx), %xmm24, %xmm28
+
+// CHECK: vpavgw 291(%rax,%r14,8), %xmm24, %xmm28
+// CHECK:  encoding: [0x62,0x21,0x3d,0x00,0xe3,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpavgw 291(%rax,%r14,8), %xmm24, %xmm28
+
+// CHECK: vpavgw 2032(%rdx), %xmm24, %xmm28
+// CHECK:  encoding: [0x62,0x61,0x3d,0x00,0xe3,0x62,0x7f]
+          vpavgw 2032(%rdx), %xmm24, %xmm28
+
+// CHECK: vpavgw 2048(%rdx), %xmm24, %xmm28
+// CHECK:  encoding: [0x62,0x61,0x3d,0x00,0xe3,0xa2,0x00,0x08,0x00,0x00]
+          vpavgw 2048(%rdx), %xmm24, %xmm28
+
+// CHECK: vpavgw -2048(%rdx), %xmm24, %xmm28
+// CHECK:  encoding: [0x62,0x61,0x3d,0x00,0xe3,0x62,0x80]
+          vpavgw -2048(%rdx), %xmm24, %xmm28
+
+// CHECK: vpavgw -2064(%rdx), %xmm24, %xmm28
+// CHECK:  encoding: [0x62,0x61,0x3d,0x00,0xe3,0xa2,0xf0,0xf7,0xff,0xff]
+          vpavgw -2064(%rdx), %xmm24, %xmm28
+
+// CHECK: vpavgw %ymm22, %ymm26, %ymm22
+// CHECK:  encoding: [0x62,0xa1,0x2d,0x20,0xe3,0xf6]
+          vpavgw %ymm22, %ymm26, %ymm22
+
+// CHECK: vpavgw %ymm22, %ymm26, %ymm22 {%k7}
+// CHECK:  encoding: [0x62,0xa1,0x2d,0x27,0xe3,0xf6]
+          vpavgw %ymm22, %ymm26, %ymm22 {%k7}
+
+// CHECK: vpavgw %ymm22, %ymm26, %ymm22 {%k7} {z}
+// CHECK:  encoding: [0x62,0xa1,0x2d,0xa7,0xe3,0xf6]
+          vpavgw %ymm22, %ymm26, %ymm22 {%k7} {z}
+
+// CHECK: vpavgw (%rcx), %ymm26, %ymm22
+// CHECK:  encoding: [0x62,0xe1,0x2d,0x20,0xe3,0x31]
+          vpavgw (%rcx), %ymm26, %ymm22
+
+// CHECK: vpavgw 291(%rax,%r14,8), %ymm26, %ymm22
+// CHECK:  encoding: [0x62,0xa1,0x2d,0x20,0xe3,0xb4,0xf0,0x23,0x01,0x00,0x00]
+          vpavgw 291(%rax,%r14,8), %ymm26, %ymm22
+
+// CHECK: vpavgw 4064(%rdx), %ymm26, %ymm22
+// CHECK:  encoding: [0x62,0xe1,0x2d,0x20,0xe3,0x72,0x7f]
+          vpavgw 4064(%rdx), %ymm26, %ymm22
+
+// CHECK: vpavgw 4096(%rdx), %ymm26, %ymm22
+// CHECK:  encoding: [0x62,0xe1,0x2d,0x20,0xe3,0xb2,0x00,0x10,0x00,0x00]
+          vpavgw 4096(%rdx), %ymm26, %ymm22
+
+// CHECK: vpavgw -4096(%rdx), %ymm26, %ymm22
+// CHECK:  encoding: [0x62,0xe1,0x2d,0x20,0xe3,0x72,0x80]
+          vpavgw -4096(%rdx), %ymm26, %ymm22
+
+// CHECK: vpavgw -4128(%rdx), %ymm26, %ymm22
+// CHECK:  encoding: [0x62,0xe1,0x2d,0x20,0xe3,0xb2,0xe0,0xef,0xff,0xff]
+          vpavgw -4128(%rdx), %ymm26, %ymm22
+
+// CHECK: vpavgb %xmm26, %xmm19, %xmm20
+// CHECK:  encoding: [0x62,0x81,0x65,0x00,0xe0,0xe2]
+          vpavgb %xmm26, %xmm19, %xmm20
+
+// CHECK: vpavgb %xmm26, %xmm19, %xmm20 {%k7}
+// CHECK:  encoding: [0x62,0x81,0x65,0x07,0xe0,0xe2]
+          vpavgb %xmm26, %xmm19, %xmm20 {%k7}
+
+// CHECK: vpavgb %xmm26, %xmm19, %xmm20 {%k7} {z}
+// CHECK:  encoding: [0x62,0x81,0x65,0x87,0xe0,0xe2]
+          vpavgb %xmm26, %xmm19, %xmm20 {%k7} {z}
+
+// CHECK: vpavgb (%rcx), %xmm19, %xmm20
+// CHECK:  encoding: [0x62,0xe1,0x65,0x00,0xe0,0x21]
+          vpavgb (%rcx), %xmm19, %xmm20
+
+// CHECK: vpavgb 291(%rax,%r14,8), %xmm19, %xmm20
+// CHECK:  encoding: [0x62,0xa1,0x65,0x00,0xe0,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpavgb 291(%rax,%r14,8), %xmm19, %xmm20
+
+// CHECK: vpavgb 2032(%rdx), %xmm19, %xmm20
+// CHECK:  encoding: [0x62,0xe1,0x65,0x00,0xe0,0x62,0x7f]
+          vpavgb 2032(%rdx), %xmm19, %xmm20
+
+// CHECK: vpavgb 2048(%rdx), %xmm19, %xmm20
+// CHECK:  encoding: [0x62,0xe1,0x65,0x00,0xe0,0xa2,0x00,0x08,0x00,0x00]
+          vpavgb 2048(%rdx), %xmm19, %xmm20
+
+// CHECK: vpavgb -2048(%rdx), %xmm19, %xmm20
+// CHECK:  encoding: [0x62,0xe1,0x65,0x00,0xe0,0x62,0x80]
+          vpavgb -2048(%rdx), %xmm19, %xmm20
+
+// CHECK: vpavgb -2064(%rdx), %xmm19, %xmm20
+// CHECK:  encoding: [0x62,0xe1,0x65,0x00,0xe0,0xa2,0xf0,0xf7,0xff,0xff]
+          vpavgb -2064(%rdx), %xmm19, %xmm20
+
+// CHECK: vpavgb %ymm17, %ymm22, %ymm26
+// CHECK:  encoding: [0x62,0x21,0x4d,0x20,0xe0,0xd1]
+          vpavgb %ymm17, %ymm22, %ymm26
+
+// CHECK: vpavgb %ymm17, %ymm22, %ymm26 {%k5}
+// CHECK:  encoding: [0x62,0x21,0x4d,0x25,0xe0,0xd1]
+          vpavgb %ymm17, %ymm22, %ymm26 {%k5}
+
+// CHECK: vpavgb %ymm17, %ymm22, %ymm26 {%k5} {z}
+// CHECK:  encoding: [0x62,0x21,0x4d,0xa5,0xe0,0xd1]
+          vpavgb %ymm17, %ymm22, %ymm26 {%k5} {z}
+
+// CHECK: vpavgb (%rcx), %ymm22, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x4d,0x20,0xe0,0x11]
+          vpavgb (%rcx), %ymm22, %ymm26
+
+// CHECK: vpavgb 291(%rax,%r14,8), %ymm22, %ymm26
+// CHECK:  encoding: [0x62,0x21,0x4d,0x20,0xe0,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpavgb 291(%rax,%r14,8), %ymm22, %ymm26
+
+// CHECK: vpavgb 4064(%rdx), %ymm22, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x4d,0x20,0xe0,0x52,0x7f]
+          vpavgb 4064(%rdx), %ymm22, %ymm26
+
+// CHECK: vpavgb 4096(%rdx), %ymm22, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x4d,0x20,0xe0,0x92,0x00,0x10,0x00,0x00]
+          vpavgb 4096(%rdx), %ymm22, %ymm26
+
+// CHECK: vpavgb -4096(%rdx), %ymm22, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x4d,0x20,0xe0,0x52,0x80]
+          vpavgb -4096(%rdx), %ymm22, %ymm26
+
+// CHECK: vpavgb -4128(%rdx), %ymm22, %ymm26
+// CHECK:  encoding: [0x62,0x61,0x4d,0x20,0xe0,0x92,0xe0,0xef,0xff,0xff]
+          vpavgb -4128(%rdx), %ymm22, %ymm26
+
+// CHECK: vpavgw %xmm20, %xmm18, %xmm22
+// CHECK:  encoding: [0x62,0xa1,0x6d,0x00,0xe3,0xf4]
+          vpavgw %xmm20, %xmm18, %xmm22
+
+// CHECK: vpavgw %xmm20, %xmm18, %xmm22 {%k7}
+// CHECK:  encoding: [0x62,0xa1,0x6d,0x07,0xe3,0xf4]
+          vpavgw %xmm20, %xmm18, %xmm22 {%k7}
+
+// CHECK: vpavgw %xmm20, %xmm18, %xmm22 {%k7} {z}
+// CHECK:  encoding: [0x62,0xa1,0x6d,0x87,0xe3,0xf4]
+          vpavgw %xmm20, %xmm18, %xmm22 {%k7} {z}
+
+// CHECK: vpavgw (%rcx), %xmm18, %xmm22
+// CHECK:  encoding: [0x62,0xe1,0x6d,0x00,0xe3,0x31]
+          vpavgw (%rcx), %xmm18, %xmm22
+
+// CHECK: vpavgw 291(%rax,%r14,8), %xmm18, %xmm22
+// CHECK:  encoding: [0x62,0xa1,0x6d,0x00,0xe3,0xb4,0xf0,0x23,0x01,0x00,0x00]
+          vpavgw 291(%rax,%r14,8), %xmm18, %xmm22
+
+// CHECK: vpavgw 2032(%rdx), %xmm18, %xmm22
+// CHECK:  encoding: [0x62,0xe1,0x6d,0x00,0xe3,0x72,0x7f]
+          vpavgw 2032(%rdx), %xmm18, %xmm22
+
+// CHECK: vpavgw 2048(%rdx), %xmm18, %xmm22
+// CHECK:  encoding: [0x62,0xe1,0x6d,0x00,0xe3,0xb2,0x00,0x08,0x00,0x00]
+          vpavgw 2048(%rdx), %xmm18, %xmm22
+
+// CHECK: vpavgw -2048(%rdx), %xmm18, %xmm22
+// CHECK:  encoding: [0x62,0xe1,0x6d,0x00,0xe3,0x72,0x80]
+          vpavgw -2048(%rdx), %xmm18, %xmm22
+
+// CHECK: vpavgw -2064(%rdx), %xmm18, %xmm22
+// CHECK:  encoding: [0x62,0xe1,0x6d,0x00,0xe3,0xb2,0xf0,0xf7,0xff,0xff]
+          vpavgw -2064(%rdx), %xmm18, %xmm22
+
+// CHECK: vpavgw %ymm21, %ymm23, %ymm21
+// CHECK:  encoding: [0x62,0xa1,0x45,0x20,0xe3,0xed]
+          vpavgw %ymm21, %ymm23, %ymm21
+
+// CHECK: vpavgw %ymm21, %ymm23, %ymm21 {%k2}
+// CHECK:  encoding: [0x62,0xa1,0x45,0x22,0xe3,0xed]
+          vpavgw %ymm21, %ymm23, %ymm21 {%k2}
+
+// CHECK: vpavgw %ymm21, %ymm23, %ymm21 {%k2} {z}
+// CHECK:  encoding: [0x62,0xa1,0x45,0xa2,0xe3,0xed]
+          vpavgw %ymm21, %ymm23, %ymm21 {%k2} {z}
+
+// CHECK: vpavgw (%rcx), %ymm23, %ymm21
+// CHECK:  encoding: [0x62,0xe1,0x45,0x20,0xe3,0x29]
+          vpavgw (%rcx), %ymm23, %ymm21
+
+// CHECK: vpavgw 291(%rax,%r14,8), %ymm23, %ymm21
+// CHECK:  encoding: [0x62,0xa1,0x45,0x20,0xe3,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vpavgw 291(%rax,%r14,8), %ymm23, %ymm21
+
+// CHECK: vpavgw 4064(%rdx), %ymm23, %ymm21
+// CHECK:  encoding: [0x62,0xe1,0x45,0x20,0xe3,0x6a,0x7f]
+          vpavgw 4064(%rdx), %ymm23, %ymm21
+
+// CHECK: vpavgw 4096(%rdx), %ymm23, %ymm21
+// CHECK:  encoding: [0x62,0xe1,0x45,0x20,0xe3,0xaa,0x00,0x10,0x00,0x00]
+          vpavgw 4096(%rdx), %ymm23, %ymm21
+
+// CHECK: vpavgw -4096(%rdx), %ymm23, %ymm21
+// CHECK:  encoding: [0x62,0xe1,0x45,0x20,0xe3,0x6a,0x80]
+          vpavgw -4096(%rdx), %ymm23, %ymm21
+
+// CHECK: vpavgw -4128(%rdx), %ymm23, %ymm21
+// CHECK:  encoding: [0x62,0xe1,0x45,0x20,0xe3,0xaa,0xe0,0xef,0xff,0xff]
+          vpavgw -4128(%rdx), %ymm23, %ymm21