[X86][AVX512BW] support in byte shift and SAD

[AsafBadouh]

Summary

add byte shift left/right
add SAD - compute sum of absolute differences

[delena]

../llvmTmp/lib/Target/X86/X86InstrAVX512.td

6788

While DAG lowering, we build SHIFT nodes with Imm8. I suppose that Imm32 will not work in this case.
Could you, please, add IR tests for these nodes?

[AsafBadouh]

apply Elena's comments

[delena]

Please apply the comments and you can commit the patch.

../llvmTmp/lib/Target/X86/X86InstrSSE.td

4187 ↗

(On Diff #33580)

The line is too long.

../llvmTmp/test/CodeGen/X86/avx512bwvl-intrinsics.ll

4197 ↗

(On Diff #33580)

This file is for intrinsics. Put in avx-isa-check.ll.

../llvmTmp/include/llvm/IR/IntrinsicsX86.td

let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_avx512_mask_psrlv_d : GCCBuiltin<"__builtin_ia32_psrlv16si_mask">,
               Intrinsic<[llvm_v16i32_ty], [llvm_v16i32_ty,
                          llvm_v16i32_ty, llvm_v16i32_ty, llvm_i16_ty],
                         [IntrNoMem]>;
   def int_x86_avx512_mask_psrlv_q : GCCBuiltin<"__builtin_ia32_psrlv8di_mask">,
               Intrinsic<[llvm_v8i64_ty], [llvm_v8i64_ty,
                          llvm_v8i64_ty, llvm_v8i64_ty, llvm_i8_ty], 
                         [IntrNoMem]>;
+  def int_x86_avx512_psll_dq_512 : GCCBuiltin<"__builtin_ia32_pslldq512">,
+              Intrinsic<[llvm_v8i64_ty], [llvm_v8i64_ty, llvm_i32_ty], 
+                        [IntrNoMem]>;
+  def int_x86_avx512_psrl_dq_512 : GCCBuiltin<"__builtin_ia32_psrldq512">,
+              Intrinsic<[llvm_v8i64_ty], [llvm_v8i64_ty, llvm_i32_ty], 
+                        [IntrNoMem]>;                        
 }
 
 // Gather ops
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_avx2_gather_d_pd : GCCBuiltin<"__builtin_ia32_gatherd_pd">,
       Intrinsic<[llvm_v2f64_ty],
         [llvm_v2f64_ty, llvm_ptr_ty, llvm_v4i32_ty, llvm_v2f64_ty, llvm_i8_ty],
         [IntrReadArgMem]>;
let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
   def int_x86_avx512_rsqrt28_ss : GCCBuiltin<"__builtin_ia32_rsqrt28ss_mask">,
             Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty,
                                         llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty],
                       [IntrNoMem]>;
   def int_x86_avx512_rsqrt28_sd : GCCBuiltin<"__builtin_ia32_rsqrt28sd_mask">,
             Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty,
                                         llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty],
                       [IntrNoMem]>;
+def int_x86_avx512_psad_bw_512 : GCCBuiltin<"__builtin_ia32_psadbw512">,
+            Intrinsic<[llvm_v64i8_ty], [llvm_v64i8_ty, llvm_v64i8_ty],
+                      [IntrNoMem]>;
 }
 // FP logical ops
 let TargetPrefix = "x86" in {
   def int_x86_avx512_mask_and_pd_128 : GCCBuiltin<"__builtin_ia32_andpd128_mask">,
             Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty,
                       llvm_v2f64_ty, llvm_i8_ty], [IntrNoMem]>;
   def int_x86_avx512_mask_and_pd_256 : GCCBuiltin<"__builtin_ia32_andpd256_mask">,
             Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty, llvm_v4f64_ty,

../llvmTmp/lib/Target/X86/X86InstrAVX512.td

 defm VPUNPCKLDQ : avx512_binop_rm_vl_d<0x62, "vpunpckldq", X86Unpckl,
                                        SSE_INTALU_ITINS_P, HasAVX512>;
 defm VPUNPCKHDQ : avx512_binop_rm_vl_d<0x6A, "vpunpckhdq", X86Unpckh,
                                        SSE_INTALU_ITINS_P, HasAVX512>;
 defm VPUNPCKLQDQ : avx512_binop_rm_vl_q<0x6C, "vpunpcklqdq", X86Unpckl,
                                        SSE_INTALU_ITINS_P, HasAVX512>;
 defm VPUNPCKHQDQ : avx512_binop_rm_vl_q<0x6D, "vpunpckhqdq", X86Unpckh,
                                        SSE_INTALU_ITINS_P, HasAVX512>;
+//===----------------------------------------------------------------------===//
+// AVX-512 - Byte shift Left/Right
+//===----------------------------------------------------------------------===//
+
+multiclass avx512_shift_packed<bits<8> opc, SDNode OpNode, Format MRMr,
+                             Format MRMm, string OpcodeStr, X86VectorVTInfo _>{
+  def rr : AVX512<opc, MRMr,
+             (outs _.RC:$dst), (ins _.RC:$src1, i32u8imm:$src2),
+             !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
+             [(set _.RC:$dst,(_.VT (OpNode _.RC:$src1, (i32 imm:$src2))))]>;
+  let mayLoad = 1 in
+    def rm : AVX512<opc, MRMm,
+             (outs _.RC:$dst), (ins _.MemOp:$src1, i32u8imm:$src2),
+             !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
+             [(set _.RC:$dst,(_.VT (OpNode 
+                                   (_.LdFrag addr:$src1), (i32 imm:$src2))))]>;
+}
+
+multiclass avx512_shift_packed_all<bits<8> opc, SDNode OpNode, Format MRMr, 
+                                 Format MRMm, string OpcodeStr, Predicate prd>{
+  let Predicates = [prd] in
+    defm Z512 : avx512_shift_packed<opc, OpNode, MRMr, MRMm, 
+                                    OpcodeStr, v8i64_info>, EVEX_V512;
+  let Predicates = [prd, HasVLX] in {
+    defm Z256 : avx512_shift_packed<opc, OpNode, MRMr, MRMm, 
+                                    OpcodeStr, v4i64x_info>, EVEX_V256;
+    defm Z128 : avx512_shift_packed<opc, OpNode, MRMr, MRMm, 
+                                    OpcodeStr, v2i64x_info>, EVEX_V128;
+  }
+}
+defm VPSLLDQ : avx512_shift_packed_all<0x73, X86vshldq, MRM7r, MRM7m, "vpslldq", 
+                                       HasBWI>, AVX512PDIi8Base, EVEX_4V;
+defm VPSRLDQ : avx512_shift_packed_all<0x73, X86vshrdq, MRM3r, MRM3m, "vpsrldq", 
+                                       HasBWI>, AVX512PDIi8Base, EVEX_4V;
+
+
+multiclass avx512_psadbw_packed<bits<8> opc, SDNode OpNode, 
+                                string OpcodeStr, X86VectorVTInfo _src>{
+  def rr : AVX512BI<opc, MRMSrcReg,
+             (outs _src.RC:$dst), (ins _src.RC:$src1, _src.RC:$src2),
+             !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
+             [(set _src.RC:$dst,(_src.VT 
+                                (OpNode _src.RC:$src1, _src.RC:$src2)))]>;
+  let mayLoad = 1 in
+    def rm : AVX512BI<opc, MRMSrcMem,
+             (outs _src.RC:$dst), (ins _src.RC:$src1, _src.MemOp:$src2),
+             !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
+             [(set _src.RC:$dst,(_src.VT 
+                                (OpNode _src.RC:$src1, 
+                                (_src.VT (bitconvert 
+                                          (_src.LdFrag addr:$src2))))))]>;
+}
+
+multiclass avx512_psadbw_packed_all<bits<8> opc, SDNode OpNode, 
+                                    string OpcodeStr, Predicate prd> {
+  let Predicates = [prd] in
+    defm Z512 : avx512_psadbw_packed<opc, OpNode, OpcodeStr, v64i8_info>, 
+                                    EVEX_V512;
+  let Predicates = [prd, HasVLX] in {
+    defm Z256 : avx512_psadbw_packed<opc, OpNode, OpcodeStr, v32i8x_info>, 
+                                    EVEX_V256;
+    defm Z128 : avx512_psadbw_packed<opc, OpNode, OpcodeStr, v16i8x_info>,
+                                    EVEX_V128;
+  }
+}
+
+defm VPSADBW : avx512_psadbw_packed_all<0xf6, X86psadbw, "vpsadbw", 
+                                       HasBWI>, EVEX_4V;
 No newline at end of file

../llvmTmp/lib/Target/X86/X86IntrinsicsInfo.h

   X86_INTRINSIC_DATA(avx512_maskz_vpermt2var_ps_512, VPERM_3OP_MASKZ,
                      X86ISD::VPERMV3, 0),
   X86_INTRINSIC_DATA(avx512_maskz_vpermt2var_q_128, VPERM_3OP_MASKZ,
                      X86ISD::VPERMV3, 0),
   X86_INTRINSIC_DATA(avx512_maskz_vpermt2var_q_256, VPERM_3OP_MASKZ,
                      X86ISD::VPERMV3, 0),
   X86_INTRINSIC_DATA(avx512_maskz_vpermt2var_q_512, VPERM_3OP_MASKZ,
                      X86ISD::VPERMV3, 0),
+  X86_INTRINSIC_DATA(avx512_psad_bw_512, INTR_TYPE_2OP, X86ISD::PSADBW, 0),
+  X86_INTRINSIC_DATA(avx512_psll_dq_512, INTR_TYPE_2OP, X86ISD::VSHLDQ, 0),
+  X86_INTRINSIC_DATA(avx512_psrl_dq_512, INTR_TYPE_2OP, X86ISD::VSRLDQ, 0),
   X86_INTRINSIC_DATA(avx512_rcp28_pd,   INTR_TYPE_1OP_MASK_RM,X86ISD::RCP28, 0),
   X86_INTRINSIC_DATA(avx512_rcp28_ps,   INTR_TYPE_1OP_MASK_RM,X86ISD::RCP28, 0),
   X86_INTRINSIC_DATA(avx512_rcp28_sd,   INTR_TYPE_SCALAR_MASK_RM, X86ISD::RCP28, 0),
   X86_INTRINSIC_DATA(avx512_rcp28_ss,   INTR_TYPE_SCALAR_MASK_RM, X86ISD::RCP28, 0),
   X86_INTRINSIC_DATA(avx512_rsqrt28_pd, INTR_TYPE_1OP_MASK_RM,X86ISD::RSQRT28, 0),
   X86_INTRINSIC_DATA(avx512_rsqrt28_ps, INTR_TYPE_1OP_MASK_RM,X86ISD::RSQRT28, 0),
   X86_INTRINSIC_DATA(avx512_rsqrt28_sd, INTR_TYPE_SCALAR_MASK_RM,X86ISD::RSQRT28, 0),
   X86_INTRINSIC_DATA(avx512_rsqrt28_ss, INTR_TYPE_SCALAR_MASK_RM,X86ISD::RSQRT28, 0),

../llvmTmp/test/CodeGen/X86/avx512bw-intrinsics.ll

 ; CHECK-NEXT:    vpunpcklwd %zmm1, %zmm0, %zmm2 {%k1}
 ; CHECK-NEXT:    vpunpcklwd %zmm1, %zmm0, %zmm0
 ; CHECK-NEXT:    vpaddw %zmm0, %zmm2, %zmm0
 ; CHECK-NEXT:    retq
   %res = call <32 x i16> @llvm.x86.avx512.mask.punpcklw.d.512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3)
   %res1 = call <32 x i16> @llvm.x86.avx512.mask.punpcklw.d.512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 -1)
   %res2 = add <32 x i16> %res, %res1
   ret <32 x i16> %res2
+}
+
+declare <8 x i64> @llvm.x86.avx512.psll.dq.512(<8 x i64>, i32)
+
+; CHECK-LABEL: @test_int_x86_avx512_mask_psll_dq_512
+; CHECK-NOT: call 
+; CHECK: vpslldq 
+; CHECK: vpslldq 
+define <8 x i64>@test_int_x86_avx512_mask_psll_dq_512(<8 x i64> %x0) {
+  %res = call <8 x i64> @llvm.x86.avx512.psll.dq.512(<8 x i64> %x0, i32 8)
+  %res1 = call <8 x i64> @llvm.x86.avx512.psll.dq.512(<8 x i64> %x0, i32 4)
+  %res2 = add <8 x i64> %res, %res1
+  ret <8 x i64> %res2
+}
+
+declare <8 x i64> @llvm.x86.avx512.psrl.dq.512(<8 x i64>, i32)
+
+; CHECK-LABEL: @test_int_x86_avx512_mask_psrl_dq_512
+; CHECK-NOT: call 
+; CHECK: vpsrldq 
+; CHECK: vpsrldq
+define <8 x i64>@test_int_x86_avx512_mask_psrl_dq_512(<8 x i64> %x0) {
+  %res = call <8 x i64> @llvm.x86.avx512.psrl.dq.512(<8 x i64> %x0, i32 8)
+  %res1 = call <8 x i64> @llvm.x86.avx512.psrl.dq.512(<8 x i64> %x0, i32 4)
+  %res2 = add <8 x i64> %res, %res1
+  ret <8 x i64> %res2
+}
+
+declare  <64 x i8> @llvm.x86.avx512.psad.bw.512(<64 x i8>, <64 x i8>)
+
+; CHECK-LABEL: @test_int_x86_avx512_mask_psadb_w_512
+; CHECK-NOT: call 
+; CHECK: vpsadbw %zmm1
+; CHECK: vpsadbw %zmm2
+define  <64 x i8>@test_int_x86_avx512_mask_psadb_w_512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2){
+  %res = call  <64 x i8> @llvm.x86.avx512.psad.bw.512(<64 x i8> %x0, <64 x i8> %x1)
+  %res1 = call  <64 x i8> @llvm.x86.avx512.psad.bw.512(<64 x i8> %x0, <64 x i8> %x2)
+  %res2 = add  <64 x i8> %res, %res1
+  ret  <64 x i8> %res2
 }
 No newline at end of file

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

 // CHECK: vpunpckhwd -8192(%rdx), %zmm26, %zmm30
 // CHECK:  encoding: [0x62,0x61,0x2d,0x40,0x69,0x72,0x80]
           vpunpckhwd -8192(%rdx), %zmm26, %zmm30
 
 // CHECK: vpunpckhwd -8256(%rdx), %zmm26, %zmm30
 // CHECK:  encoding: [0x62,0x61,0x2d,0x40,0x69,0xb2,0xc0,0xdf,0xff,0xff]
           vpunpckhwd -8256(%rdx), %zmm26, %zmm30
 
+// CHECK: vpslldq $171, %zmm28, %zmm20
+// CHECK:  encoding: [0x62,0x91,0x5d,0x40,0x73,0xfc,0xab]
+          vpslldq $171, %zmm28, %zmm20
+
+// CHECK: vpslldq $123, %zmm28, %zmm20
+// CHECK:  encoding: [0x62,0x91,0x5d,0x40,0x73,0xfc,0x7b]
+          vpslldq $123, %zmm28, %zmm20
+
+// CHECK: vpslldq $123, (%rcx), %zmm20
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x40,0x73,0x39,0x7b]
+          vpslldq $123, (%rcx), %zmm20
+
+// CHECK: vpslldq $123, 291(%rax,%r14,8), %zmm20
+// CHECK:  encoding: [0x62,0xb1,0x5d,0x40,0x73,0xbc,0xf0,0x23,0x01,0x00,0x00,0x7b]
+          vpslldq $123, 291(%rax,%r14,8), %zmm20
+
+// CHECK: vpslldq $123, 8128(%rdx), %zmm20
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x40,0x73,0x7a,0x7f,0x7b]
+          vpslldq $123, 8128(%rdx), %zmm20
+
+// CHECK: vpslldq $123, 8192(%rdx), %zmm20
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x40,0x73,0xba,0x00,0x20,0x00,0x00,0x7b]
+          vpslldq $123, 8192(%rdx), %zmm20
+
+// CHECK: vpslldq $123, -8192(%rdx), %zmm20
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x40,0x73,0x7a,0x80,0x7b]
+          vpslldq $123, -8192(%rdx), %zmm20
+
+// CHECK: vpslldq $123, -8256(%rdx), %zmm20
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x40,0x73,0xba,0xc0,0xdf,0xff,0xff,0x7b]
+          vpslldq $123, -8256(%rdx), %zmm20
+
+// CHECK: vpsrldq $171, %zmm26, %zmm18
+// CHECK:  encoding: [0x62,0x91,0x6d,0x40,0x73,0xda,0xab]
+          vpsrldq $171, %zmm26, %zmm18
+
+// CHECK: vpsrldq $123, %zmm26, %zmm18
+// CHECK:  encoding: [0x62,0x91,0x6d,0x40,0x73,0xda,0x7b]
+          vpsrldq $123, %zmm26, %zmm18
+
+// CHECK: vpsrldq $123, (%rcx), %zmm18
+// CHECK:  encoding: [0x62,0xf1,0x6d,0x40,0x73,0x19,0x7b]
+          vpsrldq $123, (%rcx), %zmm18
+
+// CHECK: vpsrldq $123, 291(%rax,%r14,8), %zmm18
+// CHECK:  encoding: [0x62,0xb1,0x6d,0x40,0x73,0x9c,0xf0,0x23,0x01,0x00,0x00,0x7b]
+          vpsrldq $123, 291(%rax,%r14,8), %zmm18
+
+// CHECK: vpsrldq $123, 8128(%rdx), %zmm18
+// CHECK:  encoding: [0x62,0xf1,0x6d,0x40,0x73,0x5a,0x7f,0x7b]
+          vpsrldq $123, 8128(%rdx), %zmm18
+
+// CHECK: vpsrldq $123, 8192(%rdx), %zmm18
+// CHECK:  encoding: [0x62,0xf1,0x6d,0x40,0x73,0x9a,0x00,0x20,0x00,0x00,0x7b]
+          vpsrldq $123, 8192(%rdx), %zmm18
+
+// CHECK: vpsrldq $123, -8192(%rdx), %zmm18
+// CHECK:  encoding: [0x62,0xf1,0x6d,0x40,0x73,0x5a,0x80,0x7b]
+          vpsrldq $123, -8192(%rdx), %zmm18
+
+// CHECK: vpsrldq $123, -8256(%rdx), %zmm18
+// CHECK:  encoding: [0x62,0xf1,0x6d,0x40,0x73,0x9a,0xc0,0xdf,0xff,0xff,0x7b]
+          vpsrldq $123, -8256(%rdx), %zmm18
+
+// CHECK: vpsadbw %zmm22, %zmm25, %zmm28
+// CHECK:  encoding: [0x62,0x21,0x35,0x40,0xf6,0xe6]
+          vpsadbw %zmm22, %zmm25, %zmm28
+
+// CHECK: vpsadbw (%rcx), %zmm25, %zmm28
+// CHECK:  encoding: [0x62,0x61,0x35,0x40,0xf6,0x21]
+          vpsadbw (%rcx), %zmm25, %zmm28
+
+// CHECK: vpsadbw 291(%rax,%r14,8), %zmm25, %zmm28
+// CHECK:  encoding: [0x62,0x21,0x35,0x40,0xf6,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpsadbw 291(%rax,%r14,8), %zmm25, %zmm28
+
+// CHECK: vpsadbw 8128(%rdx), %zmm25, %zmm28
+// CHECK:  encoding: [0x62,0x61,0x35,0x40,0xf6,0x62,0x7f]
+          vpsadbw 8128(%rdx), %zmm25, %zmm28
+
+// CHECK: vpsadbw 8192(%rdx), %zmm25, %zmm28
+// CHECK:  encoding: [0x62,0x61,0x35,0x40,0xf6,0xa2,0x00,0x20,0x00,0x00]
+          vpsadbw 8192(%rdx), %zmm25, %zmm28
+
+// CHECK: vpsadbw -8192(%rdx), %zmm25, %zmm28
+// CHECK:  encoding: [0x62,0x61,0x35,0x40,0xf6,0x62,0x80]
+          vpsadbw -8192(%rdx), %zmm25, %zmm28
+
+// CHECK: vpsadbw -8256(%rdx), %zmm25, %zmm28
+// CHECK:  encoding: [0x62,0x61,0x35,0x40,0xf6,0xa2,0xc0,0xdf,0xff,0xff]
+          vpsadbw -8256(%rdx), %zmm25, %zmm28

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

 // CHECK: vpunpckhwd -4096(%rdx), %ymm25, %ymm28
 // CHECK:  encoding: [0x62,0x61,0x35,0x20,0x69,0x62,0x80]
           vpunpckhwd -4096(%rdx), %ymm25, %ymm28
 
 // CHECK: vpunpckhwd -4128(%rdx), %ymm25, %ymm28
 // CHECK:  encoding: [0x62,0x61,0x35,0x20,0x69,0xa2,0xe0,0xef,0xff,0xff]
           vpunpckhwd -4128(%rdx), %ymm25, %ymm28
 
+// CHECK: vpslldq $171, %xmm24, %xmm20
+// CHECK:  encoding: [0x62,0x91,0x5d,0x00,0x73,0xf8,0xab]
+          vpslldq $171, %xmm24, %xmm20
+
+// CHECK: vpslldq $123, %xmm24, %xmm20
+// CHECK:  encoding: [0x62,0x91,0x5d,0x00,0x73,0xf8,0x7b]
+          vpslldq $123, %xmm24, %xmm20
+
+// CHECK: vpslldq $123, (%rcx), %xmm20
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x00,0x73,0x39,0x7b]
+          vpslldq $123, (%rcx), %xmm20
+
+// CHECK: vpslldq $123, 291(%rax,%r14,8), %xmm20
+// CHECK:  encoding: [0x62,0xb1,0x5d,0x00,0x73,0xbc,0xf0,0x23,0x01,0x00,0x00,0x7b]
+          vpslldq $123, 291(%rax,%r14,8), %xmm20
+
+// CHECK: vpslldq $123, 2032(%rdx), %xmm20
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x00,0x73,0x7a,0x7f,0x7b]
+          vpslldq $123, 2032(%rdx), %xmm20
+
+// CHECK: vpslldq $123, 2048(%rdx), %xmm20
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x00,0x73,0xba,0x00,0x08,0x00,0x00,0x7b]
+          vpslldq $123, 2048(%rdx), %xmm20
+
+// CHECK: vpslldq $123, -2048(%rdx), %xmm20
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x00,0x73,0x7a,0x80,0x7b]
+          vpslldq $123, -2048(%rdx), %xmm20
+
+// CHECK: vpslldq $123, -2064(%rdx), %xmm20
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x00,0x73,0xba,0xf0,0xf7,0xff,0xff,0x7b]
+          vpslldq $123, -2064(%rdx), %xmm20
+
+// CHECK: vpslldq $171, %ymm25, %ymm26
+// CHECK:  encoding: [0x62,0x91,0x2d,0x20,0x73,0xf9,0xab]
+          vpslldq $171, %ymm25, %ymm26
+
+// CHECK: vpslldq $123, %ymm25, %ymm26
+// CHECK:  encoding: [0x62,0x91,0x2d,0x20,0x73,0xf9,0x7b]
+          vpslldq $123, %ymm25, %ymm26
+
+// CHECK: vpslldq $123, (%rcx), %ymm26
+// CHECK:  encoding: [0x62,0xf1,0x2d,0x20,0x73,0x39,0x7b]
+          vpslldq $123, (%rcx), %ymm26
+
+// CHECK: vpslldq $123, 291(%rax,%r14,8), %ymm26
+// CHECK:  encoding: [0x62,0xb1,0x2d,0x20,0x73,0xbc,0xf0,0x23,0x01,0x00,0x00,0x7b]
+          vpslldq $123, 291(%rax,%r14,8), %ymm26
+
+// CHECK: vpslldq $123, 4064(%rdx), %ymm26
+// CHECK:  encoding: [0x62,0xf1,0x2d,0x20,0x73,0x7a,0x7f,0x7b]
+          vpslldq $123, 4064(%rdx), %ymm26
+
+// CHECK: vpslldq $123, 4096(%rdx), %ymm26
+// CHECK:  encoding: [0x62,0xf1,0x2d,0x20,0x73,0xba,0x00,0x10,0x00,0x00,0x7b]
+          vpslldq $123, 4096(%rdx), %ymm26
+
+// CHECK: vpslldq $123, -4096(%rdx), %ymm26
+// CHECK:  encoding: [0x62,0xf1,0x2d,0x20,0x73,0x7a,0x80,0x7b]
+          vpslldq $123, -4096(%rdx), %ymm26
+
+// CHECK: vpslldq $123, -4128(%rdx), %ymm26
+// CHECK:  encoding: [0x62,0xf1,0x2d,0x20,0x73,0xba,0xe0,0xef,0xff,0xff,0x7b]
+          vpslldq $123, -4128(%rdx), %ymm26
+
+// CHECK: vpslldq $171, %xmm19, %xmm23
+// CHECK:  encoding: [0x62,0xb1,0x45,0x00,0x73,0xfb,0xab]
+          vpslldq $0xab, %xmm19, %xmm23
+
+// CHECK: vpslldq $123, %xmm19, %xmm23
+// CHECK:  encoding: [0x62,0xb1,0x45,0x00,0x73,0xfb,0x7b]
+          vpslldq $0x7b, %xmm19, %xmm23
+
+// CHECK: vpslldq $123, (%rcx), %xmm23
+// CHECK:  encoding: [0x62,0xf1,0x45,0x00,0x73,0x39,0x7b]
+          vpslldq $0x7b,(%rcx), %xmm23
+
+// CHECK: vpslldq $123, 4660(%rax,%r14,8), %xmm23
+// CHECK:  encoding: [0x62,0xb1,0x45,0x00,0x73,0xbc,0xf0,0x34,0x12,0x00,0x00,0x7b]
+          vpslldq $0x7b,4660(%rax,%r14,8), %xmm23
+
+// CHECK: vpslldq $123, 2032(%rdx), %xmm23
+// CHECK:  encoding: [0x62,0xf1,0x45,0x00,0x73,0x7a,0x7f,0x7b]
+          vpslldq $0x7b,2032(%rdx), %xmm23
+
+// CHECK: vpslldq $123, 2048(%rdx), %xmm23
+// CHECK:  encoding: [0x62,0xf1,0x45,0x00,0x73,0xba,0x00,0x08,0x00,0x00,0x7b]
+          vpslldq $0x7b,2048(%rdx), %xmm23
+
+// CHECK: vpslldq $123, -2048(%rdx), %xmm23
+// CHECK:  encoding: [0x62,0xf1,0x45,0x00,0x73,0x7a,0x80,0x7b]
+          vpslldq $0x7b,-2048(%rdx), %xmm23
+
+// CHECK: vpslldq $123, -2064(%rdx), %xmm23
+// CHECK:  encoding: [0x62,0xf1,0x45,0x00,0x73,0xba,0xf0,0xf7,0xff,0xff,0x7b]
+          vpslldq $0x7b,-2064(%rdx), %xmm23
+
+// CHECK: vpslldq $171, %ymm25, %ymm29
+// CHECK:  encoding: [0x62,0x91,0x15,0x20,0x73,0xf9,0xab]
+          vpslldq $0xab, %ymm25, %ymm29
+
+// CHECK: vpslldq $123, %ymm25, %ymm29
+// CHECK:  encoding: [0x62,0x91,0x15,0x20,0x73,0xf9,0x7b]
+          vpslldq $0x7b, %ymm25, %ymm29
+
+// CHECK: vpslldq $123, (%rcx), %ymm29
+// CHECK:  encoding: [0x62,0xf1,0x15,0x20,0x73,0x39,0x7b]
+          vpslldq $0x7b,(%rcx), %ymm29
+
+// CHECK: vpslldq $123, 4660(%rax,%r14,8), %ymm29
+// CHECK:  encoding: [0x62,0xb1,0x15,0x20,0x73,0xbc,0xf0,0x34,0x12,0x00,0x00,0x7b]
+          vpslldq $0x7b,4660(%rax,%r14,8), %ymm29
+
+// CHECK: vpslldq $123, 4064(%rdx), %ymm29
+// CHECK:  encoding: [0x62,0xf1,0x15,0x20,0x73,0x7a,0x7f,0x7b]
+          vpslldq $0x7b,4064(%rdx), %ymm29
+
+// CHECK: vpslldq $123, 4096(%rdx), %ymm29
+// CHECK:  encoding: [0x62,0xf1,0x15,0x20,0x73,0xba,0x00,0x10,0x00,0x00,0x7b]
+          vpslldq $0x7b,4096(%rdx), %ymm29
+
+// CHECK: vpslldq $123, -4096(%rdx), %ymm29
+// CHECK:  encoding: [0x62,0xf1,0x15,0x20,0x73,0x7a,0x80,0x7b]
+          vpslldq $0x7b,-4096(%rdx), %ymm29
+
+// CHECK: vpslldq $123, -4128(%rdx), %ymm29
+// CHECK:  encoding: [0x62,0xf1,0x15,0x20,0x73,0xba,0xe0,0xef,0xff,0xff,0x7b]
+          vpslldq $0x7b,-4128(%rdx), %ymm29
+
+// CHECK: vpsrldq $171, %xmm21, %xmm24
+// CHECK:  encoding: [0x62,0xb1,0x3d,0x00,0x73,0xdd,0xab]
+          vpsrldq $171, %xmm21, %xmm24
+
+// CHECK: vpsrldq $123, %xmm21, %xmm24
+// CHECK:  encoding: [0x62,0xb1,0x3d,0x00,0x73,0xdd,0x7b]
+          vpsrldq $123, %xmm21, %xmm24
+
+// CHECK: vpsrldq $123, (%rcx), %xmm24
+// CHECK:  encoding: [0x62,0xf1,0x3d,0x00,0x73,0x19,0x7b]
+          vpsrldq $123, (%rcx), %xmm24
+
+// CHECK: vpsrldq $123, 291(%rax,%r14,8), %xmm24
+// CHECK:  encoding: [0x62,0xb1,0x3d,0x00,0x73,0x9c,0xf0,0x23,0x01,0x00,0x00,0x7b]
+          vpsrldq $123, 291(%rax,%r14,8), %xmm24
+
+// CHECK: vpsrldq $123, 2032(%rdx), %xmm24
+// CHECK:  encoding: [0x62,0xf1,0x3d,0x00,0x73,0x5a,0x7f,0x7b]
+          vpsrldq $123, 2032(%rdx), %xmm24
+
+// CHECK: vpsrldq $123, 2048(%rdx), %xmm24
+// CHECK:  encoding: [0x62,0xf1,0x3d,0x00,0x73,0x9a,0x00,0x08,0x00,0x00,0x7b]
+          vpsrldq $123, 2048(%rdx), %xmm24
+
+// CHECK: vpsrldq $123, -2048(%rdx), %xmm24
+// CHECK:  encoding: [0x62,0xf1,0x3d,0x00,0x73,0x5a,0x80,0x7b]
+          vpsrldq $123, -2048(%rdx), %xmm24
+
+// CHECK: vpsrldq $123, -2064(%rdx), %xmm24
+// CHECK:  encoding: [0x62,0xf1,0x3d,0x00,0x73,0x9a,0xf0,0xf7,0xff,0xff,0x7b]
+          vpsrldq $123, -2064(%rdx), %xmm24
+
+// CHECK: vpsrldq $171, %ymm25, %ymm24
+// CHECK:  encoding: [0x62,0x91,0x3d,0x20,0x73,0xd9,0xab]
+          vpsrldq $171, %ymm25, %ymm24
+
+// CHECK: vpsrldq $123, %ymm25, %ymm24
+// CHECK:  encoding: [0x62,0x91,0x3d,0x20,0x73,0xd9,0x7b]
+          vpsrldq $123, %ymm25, %ymm24
+
+// CHECK: vpsrldq $123, (%rcx), %ymm24
+// CHECK:  encoding: [0x62,0xf1,0x3d,0x20,0x73,0x19,0x7b]
+          vpsrldq $123, (%rcx), %ymm24
+
+// CHECK: vpsrldq $123, 291(%rax,%r14,8), %ymm24
+// CHECK:  encoding: [0x62,0xb1,0x3d,0x20,0x73,0x9c,0xf0,0x23,0x01,0x00,0x00,0x7b]
+          vpsrldq $123, 291(%rax,%r14,8), %ymm24
+
+// CHECK: vpsrldq $123, 4064(%rdx), %ymm24
+// CHECK:  encoding: [0x62,0xf1,0x3d,0x20,0x73,0x5a,0x7f,0x7b]
+          vpsrldq $123, 4064(%rdx), %ymm24
+
+// CHECK: vpsrldq $123, 4096(%rdx), %ymm24
+// CHECK:  encoding: [0x62,0xf1,0x3d,0x20,0x73,0x9a,0x00,0x10,0x00,0x00,0x7b]
+          vpsrldq $123, 4096(%rdx), %ymm24
+
+// CHECK: vpsrldq $123, -4096(%rdx), %ymm24
+// CHECK:  encoding: [0x62,0xf1,0x3d,0x20,0x73,0x5a,0x80,0x7b]
+          vpsrldq $123, -4096(%rdx), %ymm24
+
+// CHECK: vpsrldq $123, -4128(%rdx), %ymm24
+// CHECK:  encoding: [0x62,0xf1,0x3d,0x20,0x73,0x9a,0xe0,0xef,0xff,0xff,0x7b]
+          vpsrldq $123, -4128(%rdx), %ymm24
+
+// CHECK: vpsrldq $171, %xmm17, %xmm18
+// CHECK:  encoding: [0x62,0xb1,0x6d,0x00,0x73,0xd9,0xab]
+          vpsrldq $0xab, %xmm17, %xmm18
+
+// CHECK: vpsrldq $123, %xmm17, %xmm18
+// CHECK:  encoding: [0x62,0xb1,0x6d,0x00,0x73,0xd9,0x7b]
+          vpsrldq $0x7b, %xmm17, %xmm18
+
+// CHECK: vpsrldq $123, (%rcx), %xmm18
+// CHECK:  encoding: [0x62,0xf1,0x6d,0x00,0x73,0x19,0x7b]
+          vpsrldq $0x7b,(%rcx), %xmm18
+
+// CHECK: vpsrldq $123, 4660(%rax,%r14,8), %xmm18
+// CHECK:  encoding: [0x62,0xb1,0x6d,0x00,0x73,0x9c,0xf0,0x34,0x12,0x00,0x00,0x7b]
+          vpsrldq $0x7b,4660(%rax,%r14,8), %xmm18
+
+// CHECK: vpsrldq $123, 2032(%rdx), %xmm18
+// CHECK:  encoding: [0x62,0xf1,0x6d,0x00,0x73,0x5a,0x7f,0x7b]
+          vpsrldq $0x7b,2032(%rdx), %xmm18
+
+// CHECK: vpsrldq $123, 2048(%rdx), %xmm18
+// CHECK:  encoding: [0x62,0xf1,0x6d,0x00,0x73,0x9a,0x00,0x08,0x00,0x00,0x7b]
+          vpsrldq $0x7b,2048(%rdx), %xmm18
+
+// CHECK: vpsrldq $123, -2048(%rdx), %xmm18
+// CHECK:  encoding: [0x62,0xf1,0x6d,0x00,0x73,0x5a,0x80,0x7b]
+          vpsrldq $0x7b,-2048(%rdx), %xmm18
+
+// CHECK: vpsrldq $123, -2064(%rdx), %xmm18
+// CHECK:  encoding: [0x62,0xf1,0x6d,0x00,0x73,0x9a,0xf0,0xf7,0xff,0xff,0x7b]
+          vpsrldq $0x7b,-2064(%rdx), %xmm18
+
+// CHECK: vpsrldq $171, %ymm28, %ymm20
+// CHECK:  encoding: [0x62,0x91,0x5d,0x20,0x73,0xdc,0xab]
+          vpsrldq $0xab, %ymm28, %ymm20
+
+// CHECK: vpsrldq $123, %ymm28, %ymm20
+// CHECK:  encoding: [0x62,0x91,0x5d,0x20,0x73,0xdc,0x7b]
+          vpsrldq $0x7b, %ymm28, %ymm20
+
+// CHECK: vpsrldq $123, (%rcx), %ymm20
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x20,0x73,0x19,0x7b]
+          vpsrldq $0x7b,(%rcx), %ymm20
+
+// CHECK: vpsrldq $123, 4660(%rax,%r14,8), %ymm20
+// CHECK:  encoding: [0x62,0xb1,0x5d,0x20,0x73,0x9c,0xf0,0x34,0x12,0x00,0x00,0x7b]
+          vpsrldq $0x7b,4660(%rax,%r14,8), %ymm20
+
+// CHECK: vpsrldq $123, 4064(%rdx), %ymm20
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x20,0x73,0x5a,0x7f,0x7b]
+          vpsrldq $0x7b,4064(%rdx), %ymm20
+
+// CHECK: vpsrldq $123, 4096(%rdx), %ymm20
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x20,0x73,0x9a,0x00,0x10,0x00,0x00,0x7b]
+          vpsrldq $0x7b,4096(%rdx), %ymm20
+
+// CHECK: vpsrldq $123, -4096(%rdx), %ymm20
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x20,0x73,0x5a,0x80,0x7b]
+          vpsrldq $0x7b,-4096(%rdx), %ymm20
+
+// CHECK: vpsrldq $123, -4128(%rdx), %ymm20
+// CHECK:  encoding: [0x62,0xf1,0x5d,0x20,0x73,0x9a,0xe0,0xef,0xff,0xff,0x7b]
+          vpsrldq $0x7b,-4128(%rdx), %ymm20
+
+// CHECK: vpsadbw %xmm24, %xmm24, %xmm17
+// CHECK:  encoding: [0x62,0x81,0x3d,0x00,0xf6,0xc8]
+          vpsadbw %xmm24, %xmm24, %xmm17
+
+// CHECK: vpsadbw (%rcx), %xmm24, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0x3d,0x00,0xf6,0x09]
+          vpsadbw (%rcx), %xmm24, %xmm17
+
+// CHECK: vpsadbw 291(%rax,%r14,8), %xmm24, %xmm17
+// CHECK:  encoding: [0x62,0xa1,0x3d,0x00,0xf6,0x8c,0xf0,0x23,0x01,0x00,0x00]
+          vpsadbw 291(%rax,%r14,8), %xmm24, %xmm17
+
+// CHECK: vpsadbw 2032(%rdx), %xmm24, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0x3d,0x00,0xf6,0x4a,0x7f]
+          vpsadbw 2032(%rdx), %xmm24, %xmm17
+
+// CHECK: vpsadbw 2048(%rdx), %xmm24, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0x3d,0x00,0xf6,0x8a,0x00,0x08,0x00,0x00]
+          vpsadbw 2048(%rdx), %xmm24, %xmm17
+
+// CHECK: vpsadbw -2048(%rdx), %xmm24, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0x3d,0x00,0xf6,0x4a,0x80]
+          vpsadbw -2048(%rdx), %xmm24, %xmm17
+
+// CHECK: vpsadbw -2064(%rdx), %xmm24, %xmm17
+// CHECK:  encoding: [0x62,0xe1,0x3d,0x00,0xf6,0x8a,0xf0,0xf7,0xff,0xff]
+          vpsadbw -2064(%rdx), %xmm24, %xmm17
+
+// CHECK: vpsadbw %ymm24, %ymm27, %ymm19
+// CHECK:  encoding: [0x62,0x81,0x25,0x20,0xf6,0xd8]
+          vpsadbw %ymm24, %ymm27, %ymm19
+
+// CHECK: vpsadbw (%rcx), %ymm27, %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x25,0x20,0xf6,0x19]
+          vpsadbw (%rcx), %ymm27, %ymm19
+
+// CHECK: vpsadbw 291(%rax,%r14,8), %ymm27, %ymm19
+// CHECK:  encoding: [0x62,0xa1,0x25,0x20,0xf6,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vpsadbw 291(%rax,%r14,8), %ymm27, %ymm19
+
+// CHECK: vpsadbw 4064(%rdx), %ymm27, %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x25,0x20,0xf6,0x5a,0x7f]
+          vpsadbw 4064(%rdx), %ymm27, %ymm19
+
+// CHECK: vpsadbw 4096(%rdx), %ymm27, %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x25,0x20,0xf6,0x9a,0x00,0x10,0x00,0x00]
+          vpsadbw 4096(%rdx), %ymm27, %ymm19
+
+// CHECK: vpsadbw -4096(%rdx), %ymm27, %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x25,0x20,0xf6,0x5a,0x80]
+          vpsadbw -4096(%rdx), %ymm27, %ymm19
+
+// CHECK: vpsadbw -4128(%rdx), %ymm27, %ymm19
+// CHECK:  encoding: [0x62,0xe1,0x25,0x20,0xf6,0x9a,0xe0,0xef,0xff,0xff]
+          vpsadbw -4128(%rdx), %ymm27, %ymm19
+
+// CHECK: vpsadbw %xmm21, %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x21,0x65,0x00,0xf6,0xf5]
+          vpsadbw %xmm21, %xmm19, %xmm30
+
+// CHECK: vpsadbw (%rcx), %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x65,0x00,0xf6,0x31]
+          vpsadbw (%rcx), %xmm19, %xmm30
+
+// CHECK: vpsadbw 4660(%rax,%r14,8), %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x21,0x65,0x00,0xf6,0xb4,0xf0,0x34,0x12,0x00,0x00]
+          vpsadbw 4660(%rax,%r14,8), %xmm19, %xmm30
+
+// CHECK: vpsadbw 2032(%rdx), %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x65,0x00,0xf6,0x72,0x7f]
+          vpsadbw 2032(%rdx), %xmm19, %xmm30
+
+// CHECK: vpsadbw 2048(%rdx), %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x65,0x00,0xf6,0xb2,0x00,0x08,0x00,0x00]
+          vpsadbw 2048(%rdx), %xmm19, %xmm30
+
+// CHECK: vpsadbw -2048(%rdx), %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x65,0x00,0xf6,0x72,0x80]
+          vpsadbw -2048(%rdx), %xmm19, %xmm30
+
+// CHECK: vpsadbw -2064(%rdx), %xmm19, %xmm30
+// CHECK:  encoding: [0x62,0x61,0x65,0x00,0xf6,0xb2,0xf0,0xf7,0xff,0xff]
+          vpsadbw -2064(%rdx), %xmm19, %xmm30
+
+// CHECK: vpsadbw %ymm27, %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0x81,0x2d,0x20,0xf6,0xe3]
+          vpsadbw %ymm27, %ymm26, %ymm20
+
+// CHECK: vpsadbw (%rcx), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x2d,0x20,0xf6,0x21]
+          vpsadbw (%rcx), %ymm26, %ymm20
+
+// CHECK: vpsadbw 4660(%rax,%r14,8), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xa1,0x2d,0x20,0xf6,0xa4,0xf0,0x34,0x12,0x00,0x00]
+          vpsadbw 4660(%rax,%r14,8), %ymm26, %ymm20
+
+// CHECK: vpsadbw 4064(%rdx), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x2d,0x20,0xf6,0x62,0x7f]
+          vpsadbw 4064(%rdx), %ymm26, %ymm20
+
+// CHECK: vpsadbw 4096(%rdx), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x2d,0x20,0xf6,0xa2,0x00,0x10,0x00,0x00]
+          vpsadbw 4096(%rdx), %ymm26, %ymm20
+
+// CHECK: vpsadbw -4096(%rdx), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x2d,0x20,0xf6,0x62,0x80]
+          vpsadbw -4096(%rdx), %ymm26, %ymm20
+
+// CHECK: vpsadbw -4128(%rdx), %ymm26, %ymm20
+// CHECK:  encoding: [0x62,0xe1,0x2d,0x20,0xf6,0xa2,0xe0,0xef,0xff,0xff]
+          vpsadbw -4128(%rdx), %ymm26, %ymm20