AVX-512: (fixed) Added encoding of all forms of VPERMT2W/D/Q/PS/PD and VPERMI2W/D...

Intrinsics and tests for them are comming in the next patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@240003 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
index 68c18fd5ca76087818735392b73e7ae922a26b78..028553ccd270f591facff430c007818e7df12686 100644 (file)
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -1058,118 +1058,87 @@ def : Pat<(v8i64 (X86VPermilpi VR512:$src1, (i8 imm:$imm))),
           (VPERMILPDZri VR512:$src1, imm:$imm)>;
 
 // -- VPERM2I - 3 source operands form --
-multiclass avx512_perm_3src<bits<8> opc, string OpcodeStr, RegisterClass RC,
-                          PatFrag mem_frag, X86MemOperand x86memop,
-                          SDNode OpNode, ValueType OpVT, RegisterClass KRC> {
+multiclass avx512_perm_3src<bits<8> opc, string OpcodeStr,
+                            SDNode OpNode, X86VectorVTInfo _> {
 let Constraints = "$src1 = $dst" in {
-  def rr : AVX5128I<opc, MRMSrcReg, (outs RC:$dst),
-                   (ins RC:$src1, RC:$src2, RC:$src3),
-                   !strconcat(OpcodeStr,
-                       "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
-                   [(set RC:$dst,
-                     (OpVT (OpNode RC:$src1, RC:$src2, RC:$src3)))]>,
-                    EVEX_4V;
-
-  def rrk : AVX5128I<opc, MRMSrcReg, (outs RC:$dst),
-                   (ins RC:$src1, KRC:$mask, RC:$src2, RC:$src3),
-                   !strconcat(OpcodeStr,
-                       "\t{$src3, $src2, $dst {${mask}}|"
-                       "$dst {${mask}}, $src2, $src3}"),
-                   [(set RC:$dst, (OpVT (vselect KRC:$mask,
-                                           (OpNode RC:$src1, RC:$src2,
-                                              RC:$src3),
-                                           RC:$src1)))]>,
-                    EVEX_4V, EVEX_K;
-
-  let AddedComplexity = 30 in // Prefer over VMOV*rrkz Pat<>
-    def rrkz : AVX5128I<opc, MRMSrcReg, (outs RC:$dst),
-                   (ins RC:$src1, KRC:$mask, RC:$src2, RC:$src3),
-                   !strconcat(OpcodeStr,
-                       "\t{$src3, $src2, $dst {${mask}} {z} |",
-                       "$dst {${mask}} {z}, $src2, $src3}"),
-                   [(set RC:$dst, (OpVT (vselect KRC:$mask,
-                                           (OpNode RC:$src1, RC:$src2,
-                                              RC:$src3),
-                                           (OpVT (bitconvert
-                                              (v16i32 immAllZerosV))))))]>,
-                    EVEX_4V, EVEX_KZ;
+  defm rr: AVX512_maskable_3src<opc, MRMSrcReg, _, (outs _.RC:$dst),
+          (ins _.RC:$src2, _.RC:$src3),
+          OpcodeStr, "$src3, $src2", "$src2, $src3",
+          (_.VT (OpNode _.RC:$src1, _.RC:$src2, _.RC:$src3))>, EVEX_4V,
+         AVX5128IBase;
 
-  def rm : AVX5128I<opc, MRMSrcMem, (outs RC:$dst),
-                   (ins RC:$src1, RC:$src2, x86memop:$src3),
-                   !strconcat(OpcodeStr,
-                    "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
-                   [(set RC:$dst,
-                     (OpVT (OpNode RC:$src1, RC:$src2,
-                      (mem_frag addr:$src3))))]>, EVEX_4V;
+  let mayLoad = 1 in
+  defm rm: AVX512_maskable_3src<opc, MRMSrcMem, _, (outs _.RC:$dst),
+            (ins _.RC:$src2, _.MemOp:$src3),
+            OpcodeStr, "$src3, $src2", "$src2, $src3",
+            (_.VT (OpNode _.RC:$src1, _.RC:$src2,
+                   (_.VT (bitconvert (_.LdFrag addr:$src3)))))>,
+            EVEX_4V, AVX5128IBase;
+  }
+}
+multiclass avx512_perm_3src_mb<bits<8> opc, string OpcodeStr,
+                               SDNode OpNode, X86VectorVTInfo _> {
+  let mayLoad = 1, Constraints = "$src1 = $dst" in
+  defm rmb: AVX512_maskable_3src<opc, MRMSrcMem, _, (outs _.RC:$dst),
+              (ins _.RC:$src2, _.ScalarMemOp:$src3),
+              OpcodeStr,   !strconcat("${src3}", _.BroadcastStr,", $src2"),
+              !strconcat("$src2, ${src3}", _.BroadcastStr ),
+              (_.VT (OpNode _.RC:$src1,
+               _.RC:$src2,(_.VT (X86VBroadcast (_.ScalarLdFrag addr:$src3)))))>, 
+              AVX5128IBase, EVEX_4V, EVEX_B;
+}
 
-  def rmk : AVX5128I<opc, MRMSrcMem, (outs RC:$dst),
-                   (ins RC:$src1, KRC:$mask, RC:$src2, x86memop:$src3),
-                   !strconcat(OpcodeStr,
-                    "\t{$src3, $src2, $dst {${mask}}|"
-                    "$dst {${mask}}, $src2, $src3}"),
-                   [(set RC:$dst,
-                       (OpVT (vselect KRC:$mask,
-                                      (OpNode RC:$src1, RC:$src2,
-                                         (mem_frag addr:$src3)),
-                                      RC:$src1)))]>,
-                    EVEX_4V, EVEX_K;
-
-  let AddedComplexity = 10 in // Prefer over the rrkz variant
-    def rmkz : AVX5128I<opc, MRMSrcMem, (outs RC:$dst),
-                   (ins RC:$src1, KRC:$mask, RC:$src2, x86memop:$src3),
-                   !strconcat(OpcodeStr,
-                    "\t{$src3, $src2, $dst {${mask}} {z}|"
-                    "$dst {${mask}} {z}, $src2, $src3}"),
-                   [(set RC:$dst,
-                     (OpVT (vselect KRC:$mask,
-                                    (OpNode RC:$src1, RC:$src2,
-                                            (mem_frag addr:$src3)),
-                                    (OpVT (bitconvert
-                                       (v16i32 immAllZerosV))))))]>,
-                    EVEX_4V, EVEX_KZ;
+multiclass avx512_perm_3src_sizes<bits<8> opc, string OpcodeStr,
+                                  SDNode OpNode, AVX512VLVectorVTInfo VTInfo> {
+  let Predicates = [HasAVX512] in
+  defm NAME: avx512_perm_3src<opc, OpcodeStr, OpNode, VTInfo.info512>, 
+            avx512_perm_3src_mb<opc, OpcodeStr, OpNode, VTInfo.info512>, EVEX_V512;
+  let Predicates = [HasVLX] in {
+  defm NAME#128: avx512_perm_3src<opc, OpcodeStr, OpNode, VTInfo.info128>, 
+                 avx512_perm_3src_mb<opc, OpcodeStr, OpNode, VTInfo.info128>,
+                 EVEX_V128;
+  defm NAME#256: avx512_perm_3src<opc, OpcodeStr, OpNode, VTInfo.info256>, 
+                 avx512_perm_3src_mb<opc, OpcodeStr, OpNode, VTInfo.info256>,
+                 EVEX_V256;
+  }
+}
+multiclass avx512_perm_3src_sizes_w<bits<8> opc, string OpcodeStr, 
+                                   SDNode OpNode, AVX512VLVectorVTInfo VTInfo> {
+  let Predicates = [HasBWI] in
+  defm NAME: avx512_perm_3src<opc, OpcodeStr, OpNode, VTInfo.info512>, 
+             avx512_perm_3src_mb<opc, OpcodeStr, OpNode, VTInfo.info512>,
+             EVEX_V512;
+  let Predicates = [HasBWI, HasVLX] in {
+  defm NAME#128: avx512_perm_3src<opc, OpcodeStr, OpNode, VTInfo.info128>, 
+                 avx512_perm_3src_mb<opc, OpcodeStr, OpNode, VTInfo.info128>,
+                 EVEX_V128;
+  defm NAME#256: avx512_perm_3src<opc, OpcodeStr, OpNode, VTInfo.info256>, 
+                 avx512_perm_3src_mb<opc, OpcodeStr, OpNode, VTInfo.info256>,
+                 EVEX_V256;
   }
 }
-defm VPERMI2D  : avx512_perm_3src<0x76, "vpermi2d",  VR512, loadv16i32,
-                                  i512mem, X86VPermiv3, v16i32, VK16WM>,
-                 EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VPERMI2Q  : avx512_perm_3src<0x76, "vpermi2q",  VR512, loadv8i64,
-                                  i512mem, X86VPermiv3, v8i64, VK8WM>,
-                 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
-defm VPERMI2PS : avx512_perm_3src<0x77, "vpermi2ps",  VR512, loadv16f32,
-                                  i512mem, X86VPermiv3, v16f32, VK16WM>,
-                 EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VPERMI2PD : avx512_perm_3src<0x77, "vpermi2pd",  VR512, loadv8f64,
-                                  i512mem, X86VPermiv3, v8f64, VK8WM>,
-                  EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
-
-multiclass avx512_perm_table_3src<bits<8> opc, string Suffix, RegisterClass RC,
-                          PatFrag mem_frag, X86MemOperand x86memop,
-                          SDNode OpNode, ValueType OpVT, RegisterClass KRC,
-                          ValueType MaskVT, RegisterClass MRC> :
-        avx512_perm_3src<opc, "vpermt2"##Suffix, RC, mem_frag, x86memop, OpNode,
-                         OpVT, KRC> {
-  def : Pat<(OpVT (!cast<Intrinsic>("int_x86_avx512_mask_vpermt_"##Suffix##"_512")
-                     VR512:$idx, VR512:$src1, VR512:$src2, -1)),
-            (!cast<Instruction>(NAME#rr) VR512:$src1, VR512:$idx, VR512:$src2)>;
-
-  def : Pat<(OpVT (!cast<Intrinsic>("int_x86_avx512_mask_vpermt_"##Suffix##"_512")
-                     VR512:$idx, VR512:$src1, VR512:$src2, MRC:$mask)),
-            (!cast<Instruction>(NAME#rrk) VR512:$src1,
-              (MaskVT (COPY_TO_REGCLASS MRC:$mask, KRC)), VR512:$idx, VR512:$src2)>;
-}
-
-defm VPERMT2D  : avx512_perm_table_3src<0x7E, "d",  VR512, loadv16i32, i512mem,
-                               X86VPermv3, v16i32, VK16WM, v16i1, GR16>,
-                 EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VPERMT2Q  : avx512_perm_table_3src<0x7E, "q",  VR512, loadv8i64, i512mem,
-                               X86VPermv3, v8i64, VK8WM, v8i1, GR8>,
-                 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
-defm VPERMT2PS : avx512_perm_table_3src<0x7F, "ps",  VR512, loadv16f32, i512mem,
-                               X86VPermv3, v16f32, VK16WM, v16i1, GR16>,
-                 EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VPERMT2PD : avx512_perm_table_3src<0x7F, "pd",  VR512, loadv8f64, i512mem,
-                               X86VPermv3, v8f64, VK8WM, v8i1, GR8>,
-                 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+defm VPERMI2D  : avx512_perm_3src_sizes<0x76, "vpermi2d", X86VPermiv3,
+                                  avx512vl_i32_info>, EVEX_CD8<32, CD8VF>;
+defm VPERMI2Q  : avx512_perm_3src_sizes<0x76, "vpermi2q", X86VPermiv3,
+                                  avx512vl_i64_info>, VEX_W, EVEX_CD8<64, CD8VF>;
+defm VPERMI2PS : avx512_perm_3src_sizes<0x77, "vpermi2ps", X86VPermiv3,
+                                  avx512vl_f32_info>, EVEX_CD8<32, CD8VF>;
+defm VPERMI2PD : avx512_perm_3src_sizes<0x77, "vpermi2pd", X86VPermiv3,
+                                  avx512vl_f64_info>, VEX_W, EVEX_CD8<64, CD8VF>;
+
+defm VPERMT2D  : avx512_perm_3src_sizes<0x7E, "vpermt2d", X86VPermv3,
+                                  avx512vl_i32_info>, EVEX_CD8<32, CD8VF>;
+defm VPERMT2Q  : avx512_perm_3src_sizes<0x7E, "vpermt2q", X86VPermv3,
+                                  avx512vl_i64_info>, VEX_W, EVEX_CD8<64, CD8VF>;
+defm VPERMT2PS : avx512_perm_3src_sizes<0x7F, "vpermt2ps", X86VPermv3,
+                                  avx512vl_f32_info>, EVEX_CD8<32, CD8VF>;
+defm VPERMT2PD : avx512_perm_3src_sizes<0x7F, "vpermt2pd", X86VPermv3,
+                                  avx512vl_f64_info>, VEX_W, EVEX_CD8<64, CD8VF>;
+
+defm VPERMT2W  : avx512_perm_3src_sizes_w<0x7D, "vpermt2w", X86VPermv3,
+                                  avx512vl_i16_info>, VEX_W, EVEX_CD8<16, CD8VF>;
+defm VPERMI2W  : avx512_perm_3src_sizes_w<0x75, "vpermi2w", X86VPermiv3,
+                                  avx512vl_i16_info>, VEX_W, EVEX_CD8<16, CD8VF>;
 
 //===----------------------------------------------------------------------===//
 // AVX-512 - BLEND using mask

diff --git a/test/CodeGen/X86/avx512-intrinsics.ll b/test/CodeGen/X86/avx512-intrinsics.ll
index 544b3a0b0cba9f4ee0f593123ff97fc775a933f7..a06cadaa3f5abada378bdbf4795d9ca4c7015494 100644 (file)
--- a/test/CodeGen/X86/avx512-intrinsics.ll
+++ b/test/CodeGen/X86/avx512-intrinsics.ll
@@ -599,28 +599,6 @@ define <8 x double> @test_load_aligned_pd(<8 x double> %data, i8* %ptr, i8 %mask
   ret <8 x double> %res
 }
 
-define <16 x float> @test_vpermt2ps(<16 x float>%x, <16 x float>%y, <16 x i32>%perm) {
-; CHECK: vpermt2ps {{.*}}encoding: [0x62,0xf2,0x6d,0x48,0x7f,0xc1]
-  %res = call <16 x float> @llvm.x86.avx512.mask.vpermt.ps.512(<16 x i32>%perm, <16 x float>%x, <16 x float>%y, i16 -1)
-  ret <16 x float> %res
-}
-
-define <16 x float> @test_vpermt2ps_mask(<16 x float>%x, <16 x float>%y, <16 x i32>%perm, i16 %mask) {
-; CHECK-LABEL: test_vpermt2ps_mask:
-; CHECK: vpermt2ps %zmm1, %zmm2, %zmm0 {%k1} ## encoding: [0x62,0xf2,0x6d,0x49,0x7f,0xc1]
-  %res = call <16 x float> @llvm.x86.avx512.mask.vpermt.ps.512(<16 x i32>%perm, <16 x float>%x, <16 x float>%y, i16 %mask)
-  ret <16 x float> %res
-}
-
-declare <16 x float> @llvm.x86.avx512.mask.vpermt.ps.512(<16 x i32>, <16 x float>, <16 x float>, i16)
-
-define <8 x i64> @test_vmovntdqa(i8 *%x) {
-; CHECK-LABEL: test_vmovntdqa:
-; CHECK: vmovntdqa (%rdi), %zmm0 ## encoding: [0x62,0xf2,0x7d,0x48,0x2a,0x07]
-  %res = call <8 x i64> @llvm.x86.avx512.movntdqa(i8* %x)
-  ret <8 x i64> %res
-}
-
 declare <8 x i64> @llvm.x86.avx512.movntdqa(i8*)
 
 define <8 x i64> @test_valign_q(<8 x i64> %a, <8 x i64> %b) {

diff --git a/test/MC/X86/avx512-encodings.s b/test/MC/X86/avx512-encodings.s
index 83789c78d6ea80fd27bf6c15f82ec3e8b04849f5..e52dfac1976c90734199e6b2d1ae9ee38f7f9e6d 100644 (file)
--- a/test/MC/X86/avx512-encodings.s
+++ b/test/MC/X86/avx512-encodings.s
@@ -6060,22 +6060,6 @@ vpcmpd $1, %zmm24, %zmm7, %k5{%k4}
 // CHECK: encoding: [0x62,0xf3,0xf5,0x47,0x1e,0x72,0x01,0x02]
 vpcmpuq $2, 0x40(%rdx), %zmm17, %k6{%k7}
 
-// CHECK: vpermi2d
-// CHECK: encoding: [0x62,0x42,0x6d,0x4b,0x76,0xd6]
-vpermi2d %zmm14, %zmm2, %zmm26 {%k3}
-
-// CHECK: vpermt2pd
-// CHECK: encoding: [0x62,0xf2,0xcd,0xc6,0x7f,0xf3]
-vpermt2pd %zmm3, %zmm22, %zmm6 {%k6} {z}
-
-// CHECK: vpermi2q
-// CHECK: encoding: [0x62,0x62,0xed,0x4b,0x76,0x54,0x58,0x02]
-vpermi2q 0x80(%rax,%rbx,2), %zmm2, %zmm26 {%k3}
-
-// CHECK: vpermt2d
-// CHECK: encoding: [0x62,0x32,0x4d,0xc2,0x7e,0x24,0xad,0x05,0x00,0x00,0x00]   
-vpermt2d 5(,%r13,4), %zmm22, %zmm12 {%k2} {z}
-
 // CHECK: valignq $2
 // CHECK: encoding: [0x62,0xf3,0xfd,0x48,0x03,0x4c,0x24,0x04,0x02]
 valignq  $2, 0x100(%rsp), %zmm0, %zmm1
@@ -9305,3 +9289,228 @@ vpermilpd $0x23, 0x400(%rbx), %zmm2
 // CHECK: vcvtusi2ssq -1032(%rdx), %xmm22, %xmm14
 // CHECK:  encoding: [0x62,0x71,0xce,0x00,0x7b,0xb2,0xf8,0xfb,0xff,0xff]
           vcvtusi2ssq -1032(%rdx), %xmm22, %xmm14
+
+// CHECK: vpermi2d %zmm4, %zmm28, %zmm10
+// CHECK:  encoding: [0x62,0x72,0x1d,0x40,0x76,0xd4]
+          vpermi2d %zmm4, %zmm28, %zmm10
+
+// CHECK: vpermi2d %zmm4, %zmm28, %zmm10 {%k5}
+// CHECK:  encoding: [0x62,0x72,0x1d,0x45,0x76,0xd4]
+          vpermi2d %zmm4, %zmm28, %zmm10 {%k5}
+
+// CHECK: vpermi2d %zmm4, %zmm28, %zmm10 {%k5} {z}
+// CHECK:  encoding: [0x62,0x72,0x1d,0xc5,0x76,0xd4]
+          vpermi2d %zmm4, %zmm28, %zmm10 {%k5} {z}
+
+// CHECK: vpermi2d (%rcx), %zmm28, %zmm10
+// CHECK:  encoding: [0x62,0x72,0x1d,0x40,0x76,0x11]
+          vpermi2d (%rcx), %zmm28, %zmm10
+
+// CHECK: vpermi2d 291(%rax,%r14,8), %zmm28, %zmm10
+// CHECK:  encoding: [0x62,0x32,0x1d,0x40,0x76,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpermi2d 291(%rax,%r14,8), %zmm28, %zmm10
+
+// CHECK: vpermi2d (%rcx){1to16}, %zmm28, %zmm10
+// CHECK:  encoding: [0x62,0x72,0x1d,0x50,0x76,0x11]
+          vpermi2d (%rcx){1to16}, %zmm28, %zmm10
+
+// CHECK: vpermi2d 8128(%rdx), %zmm28, %zmm10
+// CHECK:  encoding: [0x62,0x72,0x1d,0x40,0x76,0x52,0x7f]
+          vpermi2d 8128(%rdx), %zmm28, %zmm10
+
+// CHECK: vpermi2d 8192(%rdx), %zmm28, %zmm10
+// CHECK:  encoding: [0x62,0x72,0x1d,0x40,0x76,0x92,0x00,0x20,0x00,0x00]
+          vpermi2d 8192(%rdx), %zmm28, %zmm10
+
+// CHECK: vpermi2d -8192(%rdx), %zmm28, %zmm10
+// CHECK:  encoding: [0x62,0x72,0x1d,0x40,0x76,0x52,0x80]
+          vpermi2d -8192(%rdx), %zmm28, %zmm10
+
+// CHECK: vpermi2d -8256(%rdx), %zmm28, %zmm10
+// CHECK:  encoding: [0x62,0x72,0x1d,0x40,0x76,0x92,0xc0,0xdf,0xff,0xff]
+          vpermi2d -8256(%rdx), %zmm28, %zmm10
+
+// CHECK: vpermi2d 508(%rdx){1to16}, %zmm28, %zmm10
+// CHECK:  encoding: [0x62,0x72,0x1d,0x50,0x76,0x52,0x7f]
+          vpermi2d 508(%rdx){1to16}, %zmm28, %zmm10
+
+// CHECK: vpermi2d 512(%rdx){1to16}, %zmm28, %zmm10
+// CHECK:  encoding: [0x62,0x72,0x1d,0x50,0x76,0x92,0x00,0x02,0x00,0x00]
+          vpermi2d 512(%rdx){1to16}, %zmm28, %zmm10
+
+// CHECK: vpermi2d -512(%rdx){1to16}, %zmm28, %zmm10
+// CHECK:  encoding: [0x62,0x72,0x1d,0x50,0x76,0x52,0x80]
+          vpermi2d -512(%rdx){1to16}, %zmm28, %zmm10
+
+// CHECK: vpermi2d -516(%rdx){1to16}, %zmm28, %zmm10
+// CHECK:  encoding: [0x62,0x72,0x1d,0x50,0x76,0x92,0xfc,0xfd,0xff,0xff]
+          vpermi2d -516(%rdx){1to16}, %zmm28, %zmm10
+
+// CHECK: vpermi2q %zmm28, %zmm28, %zmm18
+// CHECK:  encoding: [0x62,0x82,0x9d,0x40,0x76,0xd4]
+          vpermi2q %zmm28, %zmm28, %zmm18
+
+// CHECK: vpermi2q %zmm28, %zmm28, %zmm18 {%k2}
+// CHECK:  encoding: [0x62,0x82,0x9d,0x42,0x76,0xd4]
+          vpermi2q %zmm28, %zmm28, %zmm18 {%k2}
+
+// CHECK: vpermi2q %zmm28, %zmm28, %zmm18 {%k2} {z}
+// CHECK:  encoding: [0x62,0x82,0x9d,0xc2,0x76,0xd4]
+          vpermi2q %zmm28, %zmm28, %zmm18 {%k2} {z}
+
+// CHECK: vpermi2q (%rcx), %zmm28, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0x9d,0x40,0x76,0x11]
+          vpermi2q (%rcx), %zmm28, %zmm18
+
+// CHECK: vpermi2q 291(%rax,%r14,8), %zmm28, %zmm18
+// CHECK:  encoding: [0x62,0xa2,0x9d,0x40,0x76,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpermi2q 291(%rax,%r14,8), %zmm28, %zmm18
+
+// CHECK: vpermi2q (%rcx){1to8}, %zmm28, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0x9d,0x50,0x76,0x11]
+          vpermi2q (%rcx){1to8}, %zmm28, %zmm18
+
+// CHECK: vpermi2q 8128(%rdx), %zmm28, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0x9d,0x40,0x76,0x52,0x7f]
+          vpermi2q 8128(%rdx), %zmm28, %zmm18
+
+// CHECK: vpermi2q 8192(%rdx), %zmm28, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0x9d,0x40,0x76,0x92,0x00,0x20,0x00,0x00]
+          vpermi2q 8192(%rdx), %zmm28, %zmm18
+
+// CHECK: vpermi2q -8192(%rdx), %zmm28, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0x9d,0x40,0x76,0x52,0x80]
+          vpermi2q -8192(%rdx), %zmm28, %zmm18
+
+// CHECK: vpermi2q -8256(%rdx), %zmm28, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0x9d,0x40,0x76,0x92,0xc0,0xdf,0xff,0xff]
+          vpermi2q -8256(%rdx), %zmm28, %zmm18
+
+// CHECK: vpermi2q 1016(%rdx){1to8}, %zmm28, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0x9d,0x50,0x76,0x52,0x7f]
+          vpermi2q 1016(%rdx){1to8}, %zmm28, %zmm18
+
+// CHECK: vpermi2q 1024(%rdx){1to8}, %zmm28, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0x9d,0x50,0x76,0x92,0x00,0x04,0x00,0x00]
+          vpermi2q 1024(%rdx){1to8}, %zmm28, %zmm18
+
+// CHECK: vpermi2q -1024(%rdx){1to8}, %zmm28, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0x9d,0x50,0x76,0x52,0x80]
+          vpermi2q -1024(%rdx){1to8}, %zmm28, %zmm18
+
+// CHECK: vpermi2q -1032(%rdx){1to8}, %zmm28, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0x9d,0x50,0x76,0x92,0xf8,0xfb,0xff,0xff]
+          vpermi2q -1032(%rdx){1to8}, %zmm28, %zmm18
+
+// CHECK: vpermi2ps %zmm8, %zmm23, %zmm24
+// CHECK:  encoding: [0x62,0x42,0x45,0x40,0x77,0xc0]
+          vpermi2ps %zmm8, %zmm23, %zmm24
+
+// CHECK: vpermi2ps %zmm8, %zmm23, %zmm24 {%k2}
+// CHECK:  encoding: [0x62,0x42,0x45,0x42,0x77,0xc0]
+          vpermi2ps %zmm8, %zmm23, %zmm24 {%k2}
+
+// CHECK: vpermi2ps %zmm8, %zmm23, %zmm24 {%k2} {z}
+// CHECK:  encoding: [0x62,0x42,0x45,0xc2,0x77,0xc0]
+          vpermi2ps %zmm8, %zmm23, %zmm24 {%k2} {z}
+
+// CHECK: vpermi2ps (%rcx), %zmm23, %zmm24
+// CHECK:  encoding: [0x62,0x62,0x45,0x40,0x77,0x01]
+          vpermi2ps (%rcx), %zmm23, %zmm24
+
+// CHECK: vpermi2ps 291(%rax,%r14,8), %zmm23, %zmm24
+// CHECK:  encoding: [0x62,0x22,0x45,0x40,0x77,0x84,0xf0,0x23,0x01,0x00,0x00]
+          vpermi2ps 291(%rax,%r14,8), %zmm23, %zmm24
+
+// CHECK: vpermi2ps (%rcx){1to16}, %zmm23, %zmm24
+// CHECK:  encoding: [0x62,0x62,0x45,0x50,0x77,0x01]
+          vpermi2ps (%rcx){1to16}, %zmm23, %zmm24
+
+// CHECK: vpermi2ps 8128(%rdx), %zmm23, %zmm24
+// CHECK:  encoding: [0x62,0x62,0x45,0x40,0x77,0x42,0x7f]
+          vpermi2ps 8128(%rdx), %zmm23, %zmm24
+
+// CHECK: vpermi2ps 8192(%rdx), %zmm23, %zmm24
+// CHECK:  encoding: [0x62,0x62,0x45,0x40,0x77,0x82,0x00,0x20,0x00,0x00]
+          vpermi2ps 8192(%rdx), %zmm23, %zmm24
+
+// CHECK: vpermi2ps -8192(%rdx), %zmm23, %zmm24
+// CHECK:  encoding: [0x62,0x62,0x45,0x40,0x77,0x42,0x80]
+          vpermi2ps -8192(%rdx), %zmm23, %zmm24
+
+// CHECK: vpermi2ps -8256(%rdx), %zmm23, %zmm24
+// CHECK:  encoding: [0x62,0x62,0x45,0x40,0x77,0x82,0xc0,0xdf,0xff,0xff]
+          vpermi2ps -8256(%rdx), %zmm23, %zmm24
+
+// CHECK: vpermi2ps 508(%rdx){1to16}, %zmm23, %zmm24
+// CHECK:  encoding: [0x62,0x62,0x45,0x50,0x77,0x42,0x7f]
+          vpermi2ps 508(%rdx){1to16}, %zmm23, %zmm24
+
+// CHECK: vpermi2ps 512(%rdx){1to16}, %zmm23, %zmm24
+// CHECK:  encoding: [0x62,0x62,0x45,0x50,0x77,0x82,0x00,0x02,0x00,0x00]
+          vpermi2ps 512(%rdx){1to16}, %zmm23, %zmm24
+
+// CHECK: vpermi2ps -512(%rdx){1to16}, %zmm23, %zmm24
+// CHECK:  encoding: [0x62,0x62,0x45,0x50,0x77,0x42,0x80]
+          vpermi2ps -512(%rdx){1to16}, %zmm23, %zmm24
+
+// CHECK: vpermi2ps -516(%rdx){1to16}, %zmm23, %zmm24
+// CHECK:  encoding: [0x62,0x62,0x45,0x50,0x77,0x82,0xfc,0xfd,0xff,0xff]
+          vpermi2ps -516(%rdx){1to16}, %zmm23, %zmm24
+
+// CHECK: vpermi2pd %zmm20, %zmm5, %zmm20
+// CHECK:  encoding: [0x62,0xa2,0xd5,0x48,0x77,0xe4]
+          vpermi2pd %zmm20, %zmm5, %zmm20
+
+// CHECK: vpermi2pd %zmm20, %zmm5, %zmm20 {%k3}
+// CHECK:  encoding: [0x62,0xa2,0xd5,0x4b,0x77,0xe4]
+          vpermi2pd %zmm20, %zmm5, %zmm20 {%k3}
+
+// CHECK: vpermi2pd %zmm20, %zmm5, %zmm20 {%k3} {z}
+// CHECK:  encoding: [0x62,0xa2,0xd5,0xcb,0x77,0xe4]
+          vpermi2pd %zmm20, %zmm5, %zmm20 {%k3} {z}
+
+// CHECK: vpermi2pd (%rcx), %zmm5, %zmm20
+// CHECK:  encoding: [0x62,0xe2,0xd5,0x48,0x77,0x21]
+          vpermi2pd (%rcx), %zmm5, %zmm20
+
+// CHECK: vpermi2pd 291(%rax,%r14,8), %zmm5, %zmm20
+// CHECK:  encoding: [0x62,0xa2,0xd5,0x48,0x77,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpermi2pd 291(%rax,%r14,8), %zmm5, %zmm20
+
+// CHECK: vpermi2pd (%rcx){1to8}, %zmm5, %zmm20
+// CHECK:  encoding: [0x62,0xe2,0xd5,0x58,0x77,0x21]
+          vpermi2pd (%rcx){1to8}, %zmm5, %zmm20
+
+// CHECK: vpermi2pd 8128(%rdx), %zmm5, %zmm20
+// CHECK:  encoding: [0x62,0xe2,0xd5,0x48,0x77,0x62,0x7f]
+          vpermi2pd 8128(%rdx), %zmm5, %zmm20
+
+// CHECK: vpermi2pd 8192(%rdx), %zmm5, %zmm20
+// CHECK:  encoding: [0x62,0xe2,0xd5,0x48,0x77,0xa2,0x00,0x20,0x00,0x00]
+          vpermi2pd 8192(%rdx), %zmm5, %zmm20
+
+// CHECK: vpermi2pd -8192(%rdx), %zmm5, %zmm20
+// CHECK:  encoding: [0x62,0xe2,0xd5,0x48,0x77,0x62,0x80]
+          vpermi2pd -8192(%rdx), %zmm5, %zmm20
+
+// CHECK: vpermi2pd -8256(%rdx), %zmm5, %zmm20
+// CHECK:  encoding: [0x62,0xe2,0xd5,0x48,0x77,0xa2,0xc0,0xdf,0xff,0xff]
+          vpermi2pd -8256(%rdx), %zmm5, %zmm20
+
+// CHECK: vpermi2pd 1016(%rdx){1to8}, %zmm5, %zmm20
+// CHECK:  encoding: [0x62,0xe2,0xd5,0x58,0x77,0x62,0x7f]
+          vpermi2pd 1016(%rdx){1to8}, %zmm5, %zmm20
+
+// CHECK: vpermi2pd 1024(%rdx){1to8}, %zmm5, %zmm20
+// CHECK:  encoding: [0x62,0xe2,0xd5,0x58,0x77,0xa2,0x00,0x04,0x00,0x00]
+          vpermi2pd 1024(%rdx){1to8}, %zmm5, %zmm20
+
+// CHECK: vpermi2pd -1024(%rdx){1to8}, %zmm5, %zmm20
+// CHECK:  encoding: [0x62,0xe2,0xd5,0x58,0x77,0x62,0x80]
+          vpermi2pd -1024(%rdx){1to8}, %zmm5, %zmm20
+
+// CHECK: vpermi2pd -1032(%rdx){1to8}, %zmm5, %zmm20
+// CHECK:  encoding: [0x62,0xe2,0xd5,0x58,0x77,0xa2,0xf8,0xfb,0xff,0xff]
+          vpermi2pd -1032(%rdx){1to8}, %zmm5, %zmm20
+

diff --git a/test/MC/X86/x86-64-avx512bw.s b/test/MC/X86/x86-64-avx512bw.s
index b81e3adffd2515e9a1aaf9dea0c5506b3e02d193..132948527b59b88f69c53c78ea14b122a17c6602 100644 (file)
--- a/test/MC/X86/x86-64-avx512bw.s
+++ b/test/MC/X86/x86-64-avx512bw.s
@@ -3343,3 +3343,75 @@
 // CHECK: vpermw -8256(%rdx), %zmm19, %zmm22
 // CHECK:  encoding: [0x62,0xe2,0xe5,0x40,0x8d,0xb2,0xc0,0xdf,0xff,0xff]
           vpermw -8256(%rdx), %zmm19, %zmm22
+
+// CHECK: vpermi2w %zmm24, %zmm24, %zmm17
+// CHECK:  encoding: [0x62,0x82,0xbd,0x40,0x75,0xc8]
+          vpermi2w %zmm24, %zmm24, %zmm17
+
+// CHECK: vpermi2w %zmm24, %zmm24, %zmm17 {%k7}
+// CHECK:  encoding: [0x62,0x82,0xbd,0x47,0x75,0xc8]
+          vpermi2w %zmm24, %zmm24, %zmm17 {%k7}
+
+// CHECK: vpermi2w %zmm24, %zmm24, %zmm17 {%k7} {z}
+// CHECK:  encoding: [0x62,0x82,0xbd,0xc7,0x75,0xc8]
+          vpermi2w %zmm24, %zmm24, %zmm17 {%k7} {z}
+
+// CHECK: vpermi2w (%rcx), %zmm24, %zmm17
+// CHECK:  encoding: [0x62,0xe2,0xbd,0x40,0x75,0x09]
+          vpermi2w (%rcx), %zmm24, %zmm17
+
+// CHECK: vpermi2w 291(%rax,%r14,8), %zmm24, %zmm17
+// CHECK:  encoding: [0x62,0xa2,0xbd,0x40,0x75,0x8c,0xf0,0x23,0x01,0x00,0x00]
+          vpermi2w 291(%rax,%r14,8), %zmm24, %zmm17
+
+// CHECK: vpermi2w 8128(%rdx), %zmm24, %zmm17
+// CHECK:  encoding: [0x62,0xe2,0xbd,0x40,0x75,0x4a,0x7f]
+          vpermi2w 8128(%rdx), %zmm24, %zmm17
+
+// CHECK: vpermi2w 8192(%rdx), %zmm24, %zmm17
+// CHECK:  encoding: [0x62,0xe2,0xbd,0x40,0x75,0x8a,0x00,0x20,0x00,0x00]
+          vpermi2w 8192(%rdx), %zmm24, %zmm17
+
+// CHECK: vpermi2w -8192(%rdx), %zmm24, %zmm17
+// CHECK:  encoding: [0x62,0xe2,0xbd,0x40,0x75,0x4a,0x80]
+          vpermi2w -8192(%rdx), %zmm24, %zmm17
+
+// CHECK: vpermi2w -8256(%rdx), %zmm24, %zmm17
+// CHECK:  encoding: [0x62,0xe2,0xbd,0x40,0x75,0x8a,0xc0,0xdf,0xff,0xff]
+          vpermi2w -8256(%rdx), %zmm24, %zmm17
+
+// CHECK: vpermt2w %zmm19, %zmm25, %zmm18
+// CHECK:  encoding: [0x62,0xa2,0xb5,0x40,0x7d,0xd3]
+          vpermt2w %zmm19, %zmm25, %zmm18
+
+// CHECK: vpermt2w %zmm19, %zmm25, %zmm18 {%k2}
+// CHECK:  encoding: [0x62,0xa2,0xb5,0x42,0x7d,0xd3]
+          vpermt2w %zmm19, %zmm25, %zmm18 {%k2}
+
+// CHECK: vpermt2w %zmm19, %zmm25, %zmm18 {%k2} {z}
+// CHECK:  encoding: [0x62,0xa2,0xb5,0xc2,0x7d,0xd3]
+          vpermt2w %zmm19, %zmm25, %zmm18 {%k2} {z}
+
+// CHECK: vpermt2w (%rcx), %zmm25, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0xb5,0x40,0x7d,0x11]
+          vpermt2w (%rcx), %zmm25, %zmm18
+
+// CHECK: vpermt2w 291(%rax,%r14,8), %zmm25, %zmm18
+// CHECK:  encoding: [0x62,0xa2,0xb5,0x40,0x7d,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpermt2w 291(%rax,%r14,8), %zmm25, %zmm18
+
+// CHECK: vpermt2w 8128(%rdx), %zmm25, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0xb5,0x40,0x7d,0x52,0x7f]
+          vpermt2w 8128(%rdx), %zmm25, %zmm18
+
+// CHECK: vpermt2w 8192(%rdx), %zmm25, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0xb5,0x40,0x7d,0x92,0x00,0x20,0x00,0x00]
+          vpermt2w 8192(%rdx), %zmm25, %zmm18
+
+// CHECK: vpermt2w -8192(%rdx), %zmm25, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0xb5,0x40,0x7d,0x52,0x80]
+          vpermt2w -8192(%rdx), %zmm25, %zmm18
+
+// CHECK: vpermt2w -8256(%rdx), %zmm25, %zmm18
+// CHECK:  encoding: [0x62,0xe2,0xb5,0x40,0x7d,0x92,0xc0,0xdf,0xff,0xff]
+          vpermt2w -8256(%rdx), %zmm25, %zmm18

diff --git a/test/MC/X86/x86-64-avx512bw_vl.s b/test/MC/X86/x86-64-avx512bw_vl.s
index 0ba5e17077bab6c3aacf7e38f90a110173a95ce5..1f37fc945f92a3a6b8536f98182372a2e80a6d93 100644 (file)
--- a/test/MC/X86/x86-64-avx512bw_vl.s
+++ b/test/MC/X86/x86-64-avx512bw_vl.s
@@ -5936,3 +5936,146 @@
 // CHECK:  encoding: [0x62,0x61,0xff,0x28,0x70,0x8a,0xe0,0xef,0xff,0xff,0x7b]
           vpshuflw $123, -4128(%rdx), %ymm25
 
+// CHECK: vpermi2w %xmm21, %xmm29, %xmm19
+// CHECK:  encoding: [0x62,0xa2,0x95,0x00,0x75,0xdd]
+          vpermi2w %xmm21, %xmm29, %xmm19
+
+// CHECK: vpermi2w %xmm21, %xmm29, %xmm19 {%k2}
+// CHECK:  encoding: [0x62,0xa2,0x95,0x02,0x75,0xdd]
+          vpermi2w %xmm21, %xmm29, %xmm19 {%k2}
+
+// CHECK: vpermi2w %xmm21, %xmm29, %xmm19 {%k2} {z}
+// CHECK:  encoding: [0x62,0xa2,0x95,0x82,0x75,0xdd]
+          vpermi2w %xmm21, %xmm29, %xmm19 {%k2} {z}
+
+// CHECK: vpermi2w (%rcx), %xmm29, %xmm19
+// CHECK:  encoding: [0x62,0xe2,0x95,0x00,0x75,0x19]
+          vpermi2w (%rcx), %xmm29, %xmm19
+
+// CHECK: vpermi2w 291(%rax,%r14,8), %xmm29, %xmm19
+// CHECK:  encoding: [0x62,0xa2,0x95,0x00,0x75,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vpermi2w 291(%rax,%r14,8), %xmm29, %xmm19
+
+// CHECK: vpermi2w 2032(%rdx), %xmm29, %xmm19
+// CHECK:  encoding: [0x62,0xe2,0x95,0x00,0x75,0x5a,0x7f]
+          vpermi2w 2032(%rdx), %xmm29, %xmm19
+
+// CHECK: vpermi2w 2048(%rdx), %xmm29, %xmm19
+// CHECK:  encoding: [0x62,0xe2,0x95,0x00,0x75,0x9a,0x00,0x08,0x00,0x00]
+          vpermi2w 2048(%rdx), %xmm29, %xmm19
+
+// CHECK: vpermi2w -2048(%rdx), %xmm29, %xmm19
+// CHECK:  encoding: [0x62,0xe2,0x95,0x00,0x75,0x5a,0x80]
+          vpermi2w -2048(%rdx), %xmm29, %xmm19
+
+// CHECK: vpermi2w -2064(%rdx), %xmm29, %xmm19
+// CHECK:  encoding: [0x62,0xe2,0x95,0x00,0x75,0x9a,0xf0,0xf7,0xff,0xff]
+          vpermi2w -2064(%rdx), %xmm29, %xmm19
+
+// CHECK: vpermi2w %ymm19, %ymm25, %ymm30
+// CHECK:  encoding: [0x62,0x22,0xb5,0x20,0x75,0xf3]
+          vpermi2w %ymm19, %ymm25, %ymm30
+
+// CHECK: vpermi2w %ymm19, %ymm25, %ymm30 {%k3}
+// CHECK:  encoding: [0x62,0x22,0xb5,0x23,0x75,0xf3]
+          vpermi2w %ymm19, %ymm25, %ymm30 {%k3}
+
+// CHECK: vpermi2w %ymm19, %ymm25, %ymm30 {%k3} {z}
+// CHECK:  encoding: [0x62,0x22,0xb5,0xa3,0x75,0xf3]
+          vpermi2w %ymm19, %ymm25, %ymm30 {%k3} {z}
+
+// CHECK: vpermi2w (%rcx), %ymm25, %ymm30
+// CHECK:  encoding: [0x62,0x62,0xb5,0x20,0x75,0x31]
+          vpermi2w (%rcx), %ymm25, %ymm30
+
+// CHECK: vpermi2w 291(%rax,%r14,8), %ymm25, %ymm30
+// CHECK:  encoding: [0x62,0x22,0xb5,0x20,0x75,0xb4,0xf0,0x23,0x01,0x00,0x00]
+          vpermi2w 291(%rax,%r14,8), %ymm25, %ymm30
+
+// CHECK: vpermi2w 4064(%rdx), %ymm25, %ymm30
+// CHECK:  encoding: [0x62,0x62,0xb5,0x20,0x75,0x72,0x7f]
+          vpermi2w 4064(%rdx), %ymm25, %ymm30
+
+// CHECK: vpermi2w 4096(%rdx), %ymm25, %ymm30
+// CHECK:  encoding: [0x62,0x62,0xb5,0x20,0x75,0xb2,0x00,0x10,0x00,0x00]
+          vpermi2w 4096(%rdx), %ymm25, %ymm30
+
+// CHECK: vpermi2w -4096(%rdx), %ymm25, %ymm30
+// CHECK:  encoding: [0x62,0x62,0xb5,0x20,0x75,0x72,0x80]
+          vpermi2w -4096(%rdx), %ymm25, %ymm30
+
+// CHECK: vpermi2w -4128(%rdx), %ymm25, %ymm30
+// CHECK:  encoding: [0x62,0x62,0xb5,0x20,0x75,0xb2,0xe0,0xef,0xff,0xff]
+          vpermi2w -4128(%rdx), %ymm25, %ymm30
+
+// CHECK: vpermt2w %xmm25, %xmm22, %xmm18
+// CHECK:  encoding: [0x62,0x82,0xcd,0x00,0x7d,0xd1]
+          vpermt2w %xmm25, %xmm22, %xmm18
+
+// CHECK: vpermt2w %xmm25, %xmm22, %xmm18 {%k6}
+// CHECK:  encoding: [0x62,0x82,0xcd,0x06,0x7d,0xd1]
+          vpermt2w %xmm25, %xmm22, %xmm18 {%k6}
+
+// CHECK: vpermt2w %xmm25, %xmm22, %xmm18 {%k6} {z}
+// CHECK:  encoding: [0x62,0x82,0xcd,0x86,0x7d,0xd1]
+          vpermt2w %xmm25, %xmm22, %xmm18 {%k6} {z}
+
+// CHECK: vpermt2w (%rcx), %xmm22, %xmm18
+// CHECK:  encoding: [0x62,0xe2,0xcd,0x00,0x7d,0x11]
+          vpermt2w (%rcx), %xmm22, %xmm18
+
+// CHECK: vpermt2w 291(%rax,%r14,8), %xmm22, %xmm18
+// CHECK:  encoding: [0x62,0xa2,0xcd,0x00,0x7d,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpermt2w 291(%rax,%r14,8), %xmm22, %xmm18
+
+// CHECK: vpermt2w 2032(%rdx), %xmm22, %xmm18
+// CHECK:  encoding: [0x62,0xe2,0xcd,0x00,0x7d,0x52,0x7f]
+          vpermt2w 2032(%rdx), %xmm22, %xmm18
+
+// CHECK: vpermt2w 2048(%rdx), %xmm22, %xmm18
+// CHECK:  encoding: [0x62,0xe2,0xcd,0x00,0x7d,0x92,0x00,0x08,0x00,0x00]
+          vpermt2w 2048(%rdx), %xmm22, %xmm18
+
+// CHECK: vpermt2w -2048(%rdx), %xmm22, %xmm18
+// CHECK:  encoding: [0x62,0xe2,0xcd,0x00,0x7d,0x52,0x80]
+          vpermt2w -2048(%rdx), %xmm22, %xmm18
+
+// CHECK: vpermt2w -2064(%rdx), %xmm22, %xmm18
+// CHECK:  encoding: [0x62,0xe2,0xcd,0x00,0x7d,0x92,0xf0,0xf7,0xff,0xff]
+          vpermt2w -2064(%rdx), %xmm22, %xmm18
+
+// CHECK: vpermt2w %ymm26, %ymm23, %ymm28
+// CHECK:  encoding: [0x62,0x02,0xc5,0x20,0x7d,0xe2]
+          vpermt2w %ymm26, %ymm23, %ymm28
+
+// CHECK: vpermt2w %ymm26, %ymm23, %ymm28 {%k4}
+// CHECK:  encoding: [0x62,0x02,0xc5,0x24,0x7d,0xe2]
+          vpermt2w %ymm26, %ymm23, %ymm28 {%k4}
+
+// CHECK: vpermt2w %ymm26, %ymm23, %ymm28 {%k4} {z}
+// CHECK:  encoding: [0x62,0x02,0xc5,0xa4,0x7d,0xe2]
+          vpermt2w %ymm26, %ymm23, %ymm28 {%k4} {z}
+
+// CHECK: vpermt2w (%rcx), %ymm23, %ymm28
+// CHECK:  encoding: [0x62,0x62,0xc5,0x20,0x7d,0x21]
+          vpermt2w (%rcx), %ymm23, %ymm28
+
+// CHECK: vpermt2w 291(%rax,%r14,8), %ymm23, %ymm28
+// CHECK:  encoding: [0x62,0x22,0xc5,0x20,0x7d,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpermt2w 291(%rax,%r14,8), %ymm23, %ymm28
+
+// CHECK: vpermt2w 4064(%rdx), %ymm23, %ymm28
+// CHECK:  encoding: [0x62,0x62,0xc5,0x20,0x7d,0x62,0x7f]
+          vpermt2w 4064(%rdx), %ymm23, %ymm28
+
+// CHECK: vpermt2w 4096(%rdx), %ymm23, %ymm28
+// CHECK:  encoding: [0x62,0x62,0xc5,0x20,0x7d,0xa2,0x00,0x10,0x00,0x00]
+          vpermt2w 4096(%rdx), %ymm23, %ymm28
+
+// CHECK: vpermt2w -4096(%rdx), %ymm23, %ymm28
+// CHECK:  encoding: [0x62,0x62,0xc5,0x20,0x7d,0x62,0x80]
+          vpermt2w -4096(%rdx), %ymm23, %ymm28
+
+// CHECK: vpermt2w -4128(%rdx), %ymm23, %ymm28
+// CHECK:  encoding: [0x62,0x62,0xc5,0x20,0x7d,0xa2,0xe0,0xef,0xff,0xff]
+          vpermt2w -4128(%rdx), %ymm23, %ymm28

diff --git a/test/MC/X86/x86-64-avx512f_vl.s b/test/MC/X86/x86-64-avx512f_vl.s
index f521b3e42d4429fe45e349e8354fa21a234c08db..1381b2e76e18036bfcc7807fe3e52886655f1166 100644 (file)
--- a/test/MC/X86/x86-64-avx512f_vl.s
+++ b/test/MC/X86/x86-64-avx512f_vl.s
@@ -11132,3 +11132,899 @@ vaddpd  {rz-sae}, %zmm2, %zmm1, %zmm1
 // CHECK: valignq $123, -1032(%rdx){1to4}, %ymm24, %ymm25
 // CHECK:  encoding: [0x62,0x63,0xbd,0x30,0x03,0x8a,0xf8,0xfb,0xff,0xff,0x7b]
           valignq $0x7b, -1032(%rdx){1to4}, %ymm24, %ymm25
+
+// CHECK: vpermi2d %xmm25, %xmm23, %xmm21
+// CHECK:  encoding: [0x62,0x82,0x45,0x00,0x76,0xe9]
+          vpermi2d %xmm25, %xmm23, %xmm21
+
+// CHECK: vpermi2d %xmm25, %xmm23, %xmm21 {%k6}
+// CHECK:  encoding: [0x62,0x82,0x45,0x06,0x76,0xe9]
+          vpermi2d %xmm25, %xmm23, %xmm21 {%k6}
+
+// CHECK: vpermi2d %xmm25, %xmm23, %xmm21 {%k6} {z}
+// CHECK:  encoding: [0x62,0x82,0x45,0x86,0x76,0xe9]
+          vpermi2d %xmm25, %xmm23, %xmm21 {%k6} {z}
+
+// CHECK: vpermi2d (%rcx), %xmm23, %xmm21
+// CHECK:  encoding: [0x62,0xe2,0x45,0x00,0x76,0x29]
+          vpermi2d (%rcx), %xmm23, %xmm21
+
+// CHECK: vpermi2d 291(%rax,%r14,8), %xmm23, %xmm21
+// CHECK:  encoding: [0x62,0xa2,0x45,0x00,0x76,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vpermi2d 291(%rax,%r14,8), %xmm23, %xmm21
+
+// CHECK: vpermi2d (%rcx){1to4}, %xmm23, %xmm21
+// CHECK:  encoding: [0x62,0xe2,0x45,0x10,0x76,0x29]
+          vpermi2d (%rcx){1to4}, %xmm23, %xmm21
+
+// CHECK: vpermi2d 2032(%rdx), %xmm23, %xmm21
+// CHECK:  encoding: [0x62,0xe2,0x45,0x00,0x76,0x6a,0x7f]
+          vpermi2d 2032(%rdx), %xmm23, %xmm21
+
+// CHECK: vpermi2d 2048(%rdx), %xmm23, %xmm21
+// CHECK:  encoding: [0x62,0xe2,0x45,0x00,0x76,0xaa,0x00,0x08,0x00,0x00]
+          vpermi2d 2048(%rdx), %xmm23, %xmm21
+
+// CHECK: vpermi2d -2048(%rdx), %xmm23, %xmm21
+// CHECK:  encoding: [0x62,0xe2,0x45,0x00,0x76,0x6a,0x80]
+          vpermi2d -2048(%rdx), %xmm23, %xmm21
+
+// CHECK: vpermi2d -2064(%rdx), %xmm23, %xmm21
+// CHECK:  encoding: [0x62,0xe2,0x45,0x00,0x76,0xaa,0xf0,0xf7,0xff,0xff]
+          vpermi2d -2064(%rdx), %xmm23, %xmm21
+
+// CHECK: vpermi2d 508(%rdx){1to4}, %xmm23, %xmm21
+// CHECK:  encoding: [0x62,0xe2,0x45,0x10,0x76,0x6a,0x7f]
+          vpermi2d 508(%rdx){1to4}, %xmm23, %xmm21
+
+// CHECK: vpermi2d 512(%rdx){1to4}, %xmm23, %xmm21
+// CHECK:  encoding: [0x62,0xe2,0x45,0x10,0x76,0xaa,0x00,0x02,0x00,0x00]
+          vpermi2d 512(%rdx){1to4}, %xmm23, %xmm21
+
+// CHECK: vpermi2d -512(%rdx){1to4}, %xmm23, %xmm21
+// CHECK:  encoding: [0x62,0xe2,0x45,0x10,0x76,0x6a,0x80]
+          vpermi2d -512(%rdx){1to4}, %xmm23, %xmm21
+
+// CHECK: vpermi2d -516(%rdx){1to4}, %xmm23, %xmm21
+// CHECK:  encoding: [0x62,0xe2,0x45,0x10,0x76,0xaa,0xfc,0xfd,0xff,0xff]
+          vpermi2d -516(%rdx){1to4}, %xmm23, %xmm21
+
+// CHECK: vpermi2d %ymm22, %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xa2,0x3d,0x20,0x76,0xd6]
+          vpermi2d %ymm22, %ymm24, %ymm18
+
+// CHECK: vpermi2d %ymm22, %ymm24, %ymm18 {%k1}
+// CHECK:  encoding: [0x62,0xa2,0x3d,0x21,0x76,0xd6]
+          vpermi2d %ymm22, %ymm24, %ymm18 {%k1}
+
+// CHECK: vpermi2d %ymm22, %ymm24, %ymm18 {%k1} {z}
+// CHECK:  encoding: [0x62,0xa2,0x3d,0xa1,0x76,0xd6]
+          vpermi2d %ymm22, %ymm24, %ymm18 {%k1} {z}
+
+// CHECK: vpermi2d (%rcx), %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x20,0x76,0x11]
+          vpermi2d (%rcx), %ymm24, %ymm18
+
+// CHECK: vpermi2d 291(%rax,%r14,8), %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xa2,0x3d,0x20,0x76,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpermi2d 291(%rax,%r14,8), %ymm24, %ymm18
+
+// CHECK: vpermi2d (%rcx){1to8}, %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x30,0x76,0x11]
+          vpermi2d (%rcx){1to8}, %ymm24, %ymm18
+
+// CHECK: vpermi2d 4064(%rdx), %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x20,0x76,0x52,0x7f]
+          vpermi2d 4064(%rdx), %ymm24, %ymm18
+
+// CHECK: vpermi2d 4096(%rdx), %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x20,0x76,0x92,0x00,0x10,0x00,0x00]
+          vpermi2d 4096(%rdx), %ymm24, %ymm18
+
+// CHECK: vpermi2d -4096(%rdx), %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x20,0x76,0x52,0x80]
+          vpermi2d -4096(%rdx), %ymm24, %ymm18
+
+// CHECK: vpermi2d -4128(%rdx), %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x20,0x76,0x92,0xe0,0xef,0xff,0xff]
+          vpermi2d -4128(%rdx), %ymm24, %ymm18
+
+// CHECK: vpermi2d 508(%rdx){1to8}, %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x30,0x76,0x52,0x7f]
+          vpermi2d 508(%rdx){1to8}, %ymm24, %ymm18
+
+// CHECK: vpermi2d 512(%rdx){1to8}, %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x30,0x76,0x92,0x00,0x02,0x00,0x00]
+          vpermi2d 512(%rdx){1to8}, %ymm24, %ymm18
+
+// CHECK: vpermi2d -512(%rdx){1to8}, %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x30,0x76,0x52,0x80]
+          vpermi2d -512(%rdx){1to8}, %ymm24, %ymm18
+
+// CHECK: vpermi2d -516(%rdx){1to8}, %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x30,0x76,0x92,0xfc,0xfd,0xff,0xff]
+          vpermi2d -516(%rdx){1to8}, %ymm24, %ymm18
+
+// CHECK: vpermi2q %xmm17, %xmm28, %xmm18
+// CHECK:  encoding: [0x62,0xa2,0x9d,0x00,0x76,0xd1]
+          vpermi2q %xmm17, %xmm28, %xmm18
+
+// CHECK: vpermi2q %xmm17, %xmm28, %xmm18 {%k3}
+// CHECK:  encoding: [0x62,0xa2,0x9d,0x03,0x76,0xd1]
+          vpermi2q %xmm17, %xmm28, %xmm18 {%k3}
+
+// CHECK: vpermi2q %xmm17, %xmm28, %xmm18 {%k3} {z}
+// CHECK:  encoding: [0x62,0xa2,0x9d,0x83,0x76,0xd1]
+          vpermi2q %xmm17, %xmm28, %xmm18 {%k3} {z}
+
+// CHECK: vpermi2q (%rcx), %xmm28, %xmm18
+// CHECK:  encoding: [0x62,0xe2,0x9d,0x00,0x76,0x11]
+          vpermi2q (%rcx), %xmm28, %xmm18
+
+// CHECK: vpermi2q 291(%rax,%r14,8), %xmm28, %xmm18
+// CHECK:  encoding: [0x62,0xa2,0x9d,0x00,0x76,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpermi2q 291(%rax,%r14,8), %xmm28, %xmm18
+
+// CHECK: vpermi2q (%rcx){1to2}, %xmm28, %xmm18
+// CHECK:  encoding: [0x62,0xe2,0x9d,0x10,0x76,0x11]
+          vpermi2q (%rcx){1to2}, %xmm28, %xmm18
+
+// CHECK: vpermi2q 2032(%rdx), %xmm28, %xmm18
+// CHECK:  encoding: [0x62,0xe2,0x9d,0x00,0x76,0x52,0x7f]
+          vpermi2q 2032(%rdx), %xmm28, %xmm18
+
+// CHECK: vpermi2q 2048(%rdx), %xmm28, %xmm18
+// CHECK:  encoding: [0x62,0xe2,0x9d,0x00,0x76,0x92,0x00,0x08,0x00,0x00]
+          vpermi2q 2048(%rdx), %xmm28, %xmm18
+
+// CHECK: vpermi2q -2048(%rdx), %xmm28, %xmm18
+// CHECK:  encoding: [0x62,0xe2,0x9d,0x00,0x76,0x52,0x80]
+          vpermi2q -2048(%rdx), %xmm28, %xmm18
+
+// CHECK: vpermi2q -2064(%rdx), %xmm28, %xmm18
+// CHECK:  encoding: [0x62,0xe2,0x9d,0x00,0x76,0x92,0xf0,0xf7,0xff,0xff]
+          vpermi2q -2064(%rdx), %xmm28, %xmm18
+
+// CHECK: vpermi2q 1016(%rdx){1to2}, %xmm28, %xmm18
+// CHECK:  encoding: [0x62,0xe2,0x9d,0x10,0x76,0x52,0x7f]
+          vpermi2q 1016(%rdx){1to2}, %xmm28, %xmm18
+
+// CHECK: vpermi2q 1024(%rdx){1to2}, %xmm28, %xmm18
+// CHECK:  encoding: [0x62,0xe2,0x9d,0x10,0x76,0x92,0x00,0x04,0x00,0x00]
+          vpermi2q 1024(%rdx){1to2}, %xmm28, %xmm18
+
+// CHECK: vpermi2q -1024(%rdx){1to2}, %xmm28, %xmm18
+// CHECK:  encoding: [0x62,0xe2,0x9d,0x10,0x76,0x52,0x80]
+          vpermi2q -1024(%rdx){1to2}, %xmm28, %xmm18
+
+// CHECK: vpermi2q -1032(%rdx){1to2}, %xmm28, %xmm18
+// CHECK:  encoding: [0x62,0xe2,0x9d,0x10,0x76,0x92,0xf8,0xfb,0xff,0xff]
+          vpermi2q -1032(%rdx){1to2}, %xmm28, %xmm18
+
+// CHECK: vpermi2q %ymm23, %ymm22, %ymm26
+// CHECK:  encoding: [0x62,0x22,0xcd,0x20,0x76,0xd7]
+          vpermi2q %ymm23, %ymm22, %ymm26
+
+// CHECK: vpermi2q %ymm23, %ymm22, %ymm26 {%k2}
+// CHECK:  encoding: [0x62,0x22,0xcd,0x22,0x76,0xd7]
+          vpermi2q %ymm23, %ymm22, %ymm26 {%k2}
+
+// CHECK: vpermi2q %ymm23, %ymm22, %ymm26 {%k2} {z}
+// CHECK:  encoding: [0x62,0x22,0xcd,0xa2,0x76,0xd7]
+          vpermi2q %ymm23, %ymm22, %ymm26 {%k2} {z}
+
+// CHECK: vpermi2q (%rcx), %ymm22, %ymm26
+// CHECK:  encoding: [0x62,0x62,0xcd,0x20,0x76,0x11]
+          vpermi2q (%rcx), %ymm22, %ymm26
+
+// CHECK: vpermi2q 291(%rax,%r14,8), %ymm22, %ymm26
+// CHECK:  encoding: [0x62,0x22,0xcd,0x20,0x76,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpermi2q 291(%rax,%r14,8), %ymm22, %ymm26
+
+// CHECK: vpermi2q (%rcx){1to4}, %ymm22, %ymm26
+// CHECK:  encoding: [0x62,0x62,0xcd,0x30,0x76,0x11]
+          vpermi2q (%rcx){1to4}, %ymm22, %ymm26
+
+// CHECK: vpermi2q 4064(%rdx), %ymm22, %ymm26
+// CHECK:  encoding: [0x62,0x62,0xcd,0x20,0x76,0x52,0x7f]
+          vpermi2q 4064(%rdx), %ymm22, %ymm26
+
+// CHECK: vpermi2q 4096(%rdx), %ymm22, %ymm26
+// CHECK:  encoding: [0x62,0x62,0xcd,0x20,0x76,0x92,0x00,0x10,0x00,0x00]
+          vpermi2q 4096(%rdx), %ymm22, %ymm26
+
+// CHECK: vpermi2q -4096(%rdx), %ymm22, %ymm26
+// CHECK:  encoding: [0x62,0x62,0xcd,0x20,0x76,0x52,0x80]
+          vpermi2q -4096(%rdx), %ymm22, %ymm26
+
+// CHECK: vpermi2q -4128(%rdx), %ymm22, %ymm26
+// CHECK:  encoding: [0x62,0x62,0xcd,0x20,0x76,0x92,0xe0,0xef,0xff,0xff]
+          vpermi2q -4128(%rdx), %ymm22, %ymm26
+
+// CHECK: vpermi2q 1016(%rdx){1to4}, %ymm22, %ymm26
+// CHECK:  encoding: [0x62,0x62,0xcd,0x30,0x76,0x52,0x7f]
+          vpermi2q 1016(%rdx){1to4}, %ymm22, %ymm26
+
+// CHECK: vpermi2q 1024(%rdx){1to4}, %ymm22, %ymm26
+// CHECK:  encoding: [0x62,0x62,0xcd,0x30,0x76,0x92,0x00,0x04,0x00,0x00]
+          vpermi2q 1024(%rdx){1to4}, %ymm22, %ymm26
+
+// CHECK: vpermi2q -1024(%rdx){1to4}, %ymm22, %ymm26
+// CHECK:  encoding: [0x62,0x62,0xcd,0x30,0x76,0x52,0x80]
+          vpermi2q -1024(%rdx){1to4}, %ymm22, %ymm26
+
+// CHECK: vpermi2q -1032(%rdx){1to4}, %ymm22, %ymm26
+// CHECK:  encoding: [0x62,0x62,0xcd,0x30,0x76,0x92,0xf8,0xfb,0xff,0xff]
+          vpermi2q -1032(%rdx){1to4}, %ymm22, %ymm26
+
+// CHECK: vpermi2ps %xmm23, %xmm24, %xmm23
+// CHECK:  encoding: [0x62,0xa2,0x3d,0x00,0x77,0xff]
+          vpermi2ps %xmm23, %xmm24, %xmm23
+
+// CHECK: vpermi2ps %xmm23, %xmm24, %xmm23 {%k3}
+// CHECK:  encoding: [0x62,0xa2,0x3d,0x03,0x77,0xff]
+          vpermi2ps %xmm23, %xmm24, %xmm23 {%k3}
+
+// CHECK: vpermi2ps %xmm23, %xmm24, %xmm23 {%k3} {z}
+// CHECK:  encoding: [0x62,0xa2,0x3d,0x83,0x77,0xff]
+          vpermi2ps %xmm23, %xmm24, %xmm23 {%k3} {z}
+
+// CHECK: vpermi2ps (%rcx), %xmm24, %xmm23
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x00,0x77,0x39]
+          vpermi2ps (%rcx), %xmm24, %xmm23
+
+// CHECK: vpermi2ps 291(%rax,%r14,8), %xmm24, %xmm23
+// CHECK:  encoding: [0x62,0xa2,0x3d,0x00,0x77,0xbc,0xf0,0x23,0x01,0x00,0x00]
+          vpermi2ps 291(%rax,%r14,8), %xmm24, %xmm23
+
+// CHECK: vpermi2ps (%rcx){1to4}, %xmm24, %xmm23
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x10,0x77,0x39]
+          vpermi2ps (%rcx){1to4}, %xmm24, %xmm23
+
+// CHECK: vpermi2ps 2032(%rdx), %xmm24, %xmm23
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x00,0x77,0x7a,0x7f]
+          vpermi2ps 2032(%rdx), %xmm24, %xmm23
+
+// CHECK: vpermi2ps 2048(%rdx), %xmm24, %xmm23
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x00,0x77,0xba,0x00,0x08,0x00,0x00]
+          vpermi2ps 2048(%rdx), %xmm24, %xmm23
+
+// CHECK: vpermi2ps -2048(%rdx), %xmm24, %xmm23
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x00,0x77,0x7a,0x80]
+          vpermi2ps -2048(%rdx), %xmm24, %xmm23
+
+// CHECK: vpermi2ps -2064(%rdx), %xmm24, %xmm23
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x00,0x77,0xba,0xf0,0xf7,0xff,0xff]
+          vpermi2ps -2064(%rdx), %xmm24, %xmm23
+
+// CHECK: vpermi2ps 508(%rdx){1to4}, %xmm24, %xmm23
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x10,0x77,0x7a,0x7f]
+          vpermi2ps 508(%rdx){1to4}, %xmm24, %xmm23
+
+// CHECK: vpermi2ps 512(%rdx){1to4}, %xmm24, %xmm23
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x10,0x77,0xba,0x00,0x02,0x00,0x00]
+          vpermi2ps 512(%rdx){1to4}, %xmm24, %xmm23
+
+// CHECK: vpermi2ps -512(%rdx){1to4}, %xmm24, %xmm23
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x10,0x77,0x7a,0x80]
+          vpermi2ps -512(%rdx){1to4}, %xmm24, %xmm23
+
+// CHECK: vpermi2ps -516(%rdx){1to4}, %xmm24, %xmm23
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x10,0x77,0xba,0xfc,0xfd,0xff,0xff]
+          vpermi2ps -516(%rdx){1to4}, %xmm24, %xmm23
+
+// CHECK: vpermi2ps %ymm20, %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xa2,0x3d,0x20,0x77,0xd4]
+          vpermi2ps %ymm20, %ymm24, %ymm18
+
+// CHECK: vpermi2ps %ymm20, %ymm24, %ymm18 {%k5}
+// CHECK:  encoding: [0x62,0xa2,0x3d,0x25,0x77,0xd4]
+          vpermi2ps %ymm20, %ymm24, %ymm18 {%k5}
+
+// CHECK: vpermi2ps %ymm20, %ymm24, %ymm18 {%k5} {z}
+// CHECK:  encoding: [0x62,0xa2,0x3d,0xa5,0x77,0xd4]
+          vpermi2ps %ymm20, %ymm24, %ymm18 {%k5} {z}
+
+// CHECK: vpermi2ps (%rcx), %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x20,0x77,0x11]
+          vpermi2ps (%rcx), %ymm24, %ymm18
+
+// CHECK: vpermi2ps 291(%rax,%r14,8), %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xa2,0x3d,0x20,0x77,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpermi2ps 291(%rax,%r14,8), %ymm24, %ymm18
+
+// CHECK: vpermi2ps (%rcx){1to8}, %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x30,0x77,0x11]
+          vpermi2ps (%rcx){1to8}, %ymm24, %ymm18
+
+// CHECK: vpermi2ps 4064(%rdx), %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x20,0x77,0x52,0x7f]
+          vpermi2ps 4064(%rdx), %ymm24, %ymm18
+
+// CHECK: vpermi2ps 4096(%rdx), %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x20,0x77,0x92,0x00,0x10,0x00,0x00]
+          vpermi2ps 4096(%rdx), %ymm24, %ymm18
+
+// CHECK: vpermi2ps -4096(%rdx), %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x20,0x77,0x52,0x80]
+          vpermi2ps -4096(%rdx), %ymm24, %ymm18
+
+// CHECK: vpermi2ps -4128(%rdx), %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x20,0x77,0x92,0xe0,0xef,0xff,0xff]
+          vpermi2ps -4128(%rdx), %ymm24, %ymm18
+
+// CHECK: vpermi2ps 508(%rdx){1to8}, %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x30,0x77,0x52,0x7f]
+          vpermi2ps 508(%rdx){1to8}, %ymm24, %ymm18
+
+// CHECK: vpermi2ps 512(%rdx){1to8}, %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x30,0x77,0x92,0x00,0x02,0x00,0x00]
+          vpermi2ps 512(%rdx){1to8}, %ymm24, %ymm18
+
+// CHECK: vpermi2ps -512(%rdx){1to8}, %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x30,0x77,0x52,0x80]
+          vpermi2ps -512(%rdx){1to8}, %ymm24, %ymm18
+
+// CHECK: vpermi2ps -516(%rdx){1to8}, %ymm24, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x30,0x77,0x92,0xfc,0xfd,0xff,0xff]
+          vpermi2ps -516(%rdx){1to8}, %ymm24, %ymm18
+
+// CHECK: vpermi2pd %xmm27, %xmm17, %xmm28
+// CHECK:  encoding: [0x62,0x02,0xf5,0x00,0x77,0xe3]
+          vpermi2pd %xmm27, %xmm17, %xmm28
+
+// CHECK: vpermi2pd %xmm27, %xmm17, %xmm28 {%k4}
+// CHECK:  encoding: [0x62,0x02,0xf5,0x04,0x77,0xe3]
+          vpermi2pd %xmm27, %xmm17, %xmm28 {%k4}
+
+// CHECK: vpermi2pd %xmm27, %xmm17, %xmm28 {%k4} {z}
+// CHECK:  encoding: [0x62,0x02,0xf5,0x84,0x77,0xe3]
+          vpermi2pd %xmm27, %xmm17, %xmm28 {%k4} {z}
+
+// CHECK: vpermi2pd (%rcx), %xmm17, %xmm28
+// CHECK:  encoding: [0x62,0x62,0xf5,0x00,0x77,0x21]
+          vpermi2pd (%rcx), %xmm17, %xmm28
+
+// CHECK: vpermi2pd 291(%rax,%r14,8), %xmm17, %xmm28
+// CHECK:  encoding: [0x62,0x22,0xf5,0x00,0x77,0xa4,0xf0,0x23,0x01,0x00,0x00]
+          vpermi2pd 291(%rax,%r14,8), %xmm17, %xmm28
+
+// CHECK: vpermi2pd (%rcx){1to2}, %xmm17, %xmm28
+// CHECK:  encoding: [0x62,0x62,0xf5,0x10,0x77,0x21]
+          vpermi2pd (%rcx){1to2}, %xmm17, %xmm28
+
+// CHECK: vpermi2pd 2032(%rdx), %xmm17, %xmm28
+// CHECK:  encoding: [0x62,0x62,0xf5,0x00,0x77,0x62,0x7f]
+          vpermi2pd 2032(%rdx), %xmm17, %xmm28
+
+// CHECK: vpermi2pd 2048(%rdx), %xmm17, %xmm28
+// CHECK:  encoding: [0x62,0x62,0xf5,0x00,0x77,0xa2,0x00,0x08,0x00,0x00]
+          vpermi2pd 2048(%rdx), %xmm17, %xmm28
+
+// CHECK: vpermi2pd -2048(%rdx), %xmm17, %xmm28
+// CHECK:  encoding: [0x62,0x62,0xf5,0x00,0x77,0x62,0x80]
+          vpermi2pd -2048(%rdx), %xmm17, %xmm28
+
+// CHECK: vpermi2pd -2064(%rdx), %xmm17, %xmm28
+// CHECK:  encoding: [0x62,0x62,0xf5,0x00,0x77,0xa2,0xf0,0xf7,0xff,0xff]
+          vpermi2pd -2064(%rdx), %xmm17, %xmm28
+
+// CHECK: vpermi2pd 1016(%rdx){1to2}, %xmm17, %xmm28
+// CHECK:  encoding: [0x62,0x62,0xf5,0x10,0x77,0x62,0x7f]
+          vpermi2pd 1016(%rdx){1to2}, %xmm17, %xmm28
+
+// CHECK: vpermi2pd 1024(%rdx){1to2}, %xmm17, %xmm28
+// CHECK:  encoding: [0x62,0x62,0xf5,0x10,0x77,0xa2,0x00,0x04,0x00,0x00]
+          vpermi2pd 1024(%rdx){1to2}, %xmm17, %xmm28
+
+// CHECK: vpermi2pd -1024(%rdx){1to2}, %xmm17, %xmm28
+// CHECK:  encoding: [0x62,0x62,0xf5,0x10,0x77,0x62,0x80]
+          vpermi2pd -1024(%rdx){1to2}, %xmm17, %xmm28
+
+// CHECK: vpermi2pd -1032(%rdx){1to2}, %xmm17, %xmm28
+// CHECK:  encoding: [0x62,0x62,0xf5,0x10,0x77,0xa2,0xf8,0xfb,0xff,0xff]
+          vpermi2pd -1032(%rdx){1to2}, %xmm17, %xmm28
+
+// CHECK: vpermi2pd %ymm27, %ymm19, %ymm30
+// CHECK:  encoding: [0x62,0x02,0xe5,0x20,0x77,0xf3]
+          vpermi2pd %ymm27, %ymm19, %ymm30
+
+// CHECK: vpermi2pd %ymm27, %ymm19, %ymm30 {%k3}
+// CHECK:  encoding: [0x62,0x02,0xe5,0x23,0x77,0xf3]
+          vpermi2pd %ymm27, %ymm19, %ymm30 {%k3}
+
+// CHECK: vpermi2pd %ymm27, %ymm19, %ymm30 {%k3} {z}
+// CHECK:  encoding: [0x62,0x02,0xe5,0xa3,0x77,0xf3]
+          vpermi2pd %ymm27, %ymm19, %ymm30 {%k3} {z}
+
+// CHECK: vpermi2pd (%rcx), %ymm19, %ymm30
+// CHECK:  encoding: [0x62,0x62,0xe5,0x20,0x77,0x31]
+          vpermi2pd (%rcx), %ymm19, %ymm30
+
+// CHECK: vpermi2pd 291(%rax,%r14,8), %ymm19, %ymm30
+// CHECK:  encoding: [0x62,0x22,0xe5,0x20,0x77,0xb4,0xf0,0x23,0x01,0x00,0x00]
+          vpermi2pd 291(%rax,%r14,8), %ymm19, %ymm30
+
+// CHECK: vpermi2pd (%rcx){1to4}, %ymm19, %ymm30
+// CHECK:  encoding: [0x62,0x62,0xe5,0x30,0x77,0x31]
+          vpermi2pd (%rcx){1to4}, %ymm19, %ymm30
+
+// CHECK: vpermi2pd 4064(%rdx), %ymm19, %ymm30
+// CHECK:  encoding: [0x62,0x62,0xe5,0x20,0x77,0x72,0x7f]
+          vpermi2pd 4064(%rdx), %ymm19, %ymm30
+
+// CHECK: vpermi2pd 4096(%rdx), %ymm19, %ymm30
+// CHECK:  encoding: [0x62,0x62,0xe5,0x20,0x77,0xb2,0x00,0x10,0x00,0x00]
+          vpermi2pd 4096(%rdx), %ymm19, %ymm30
+
+// CHECK: vpermi2pd -4096(%rdx), %ymm19, %ymm30
+// CHECK:  encoding: [0x62,0x62,0xe5,0x20,0x77,0x72,0x80]
+          vpermi2pd -4096(%rdx), %ymm19, %ymm30
+
+// CHECK: vpermi2pd -4128(%rdx), %ymm19, %ymm30
+// CHECK:  encoding: [0x62,0x62,0xe5,0x20,0x77,0xb2,0xe0,0xef,0xff,0xff]
+          vpermi2pd -4128(%rdx), %ymm19, %ymm30
+
+// CHECK: vpermi2pd 1016(%rdx){1to4}, %ymm19, %ymm30
+// CHECK:  encoding: [0x62,0x62,0xe5,0x30,0x77,0x72,0x7f]
+          vpermi2pd 1016(%rdx){1to4}, %ymm19, %ymm30
+
+// CHECK: vpermi2pd 1024(%rdx){1to4}, %ymm19, %ymm30
+// CHECK:  encoding: [0x62,0x62,0xe5,0x30,0x77,0xb2,0x00,0x04,0x00,0x00]
+          vpermi2pd 1024(%rdx){1to4}, %ymm19, %ymm30
+
+// CHECK: vpermi2pd -1024(%rdx){1to4}, %ymm19, %ymm30
+// CHECK:  encoding: [0x62,0x62,0xe5,0x30,0x77,0x72,0x80]
+          vpermi2pd -1024(%rdx){1to4}, %ymm19, %ymm30
+
+// CHECK: vpermi2pd -1032(%rdx){1to4}, %ymm19, %ymm30
+// CHECK:  encoding: [0x62,0x62,0xe5,0x30,0x77,0xb2,0xf8,0xfb,0xff,0xff]
+          vpermi2pd -1032(%rdx){1to4}, %ymm19, %ymm30
+
+// CHECK: vpermt2d %xmm23, %xmm29, %xmm21
+// CHECK:  encoding: [0x62,0xa2,0x15,0x00,0x7e,0xef]
+          vpermt2d %xmm23, %xmm29, %xmm21
+
+// CHECK: vpermt2d %xmm23, %xmm29, %xmm21 {%k4}
+// CHECK:  encoding: [0x62,0xa2,0x15,0x04,0x7e,0xef]
+          vpermt2d %xmm23, %xmm29, %xmm21 {%k4}
+
+// CHECK: vpermt2d %xmm23, %xmm29, %xmm21 {%k4} {z}
+// CHECK:  encoding: [0x62,0xa2,0x15,0x84,0x7e,0xef]
+          vpermt2d %xmm23, %xmm29, %xmm21 {%k4} {z}
+
+// CHECK: vpermt2d (%rcx), %xmm29, %xmm21
+// CHECK:  encoding: [0x62,0xe2,0x15,0x00,0x7e,0x29]
+          vpermt2d (%rcx), %xmm29, %xmm21
+
+// CHECK: vpermt2d 291(%rax,%r14,8), %xmm29, %xmm21
+// CHECK:  encoding: [0x62,0xa2,0x15,0x00,0x7e,0xac,0xf0,0x23,0x01,0x00,0x00]
+          vpermt2d 291(%rax,%r14,8), %xmm29, %xmm21
+
+// CHECK: vpermt2d (%rcx){1to4}, %xmm29, %xmm21
+// CHECK:  encoding: [0x62,0xe2,0x15,0x10,0x7e,0x29]
+          vpermt2d (%rcx){1to4}, %xmm29, %xmm21
+
+// CHECK: vpermt2d 2032(%rdx), %xmm29, %xmm21
+// CHECK:  encoding: [0x62,0xe2,0x15,0x00,0x7e,0x6a,0x7f]
+          vpermt2d 2032(%rdx), %xmm29, %xmm21
+
+// CHECK: vpermt2d 2048(%rdx), %xmm29, %xmm21
+// CHECK:  encoding: [0x62,0xe2,0x15,0x00,0x7e,0xaa,0x00,0x08,0x00,0x00]
+          vpermt2d 2048(%rdx), %xmm29, %xmm21
+
+// CHECK: vpermt2d -2048(%rdx), %xmm29, %xmm21
+// CHECK:  encoding: [0x62,0xe2,0x15,0x00,0x7e,0x6a,0x80]
+          vpermt2d -2048(%rdx), %xmm29, %xmm21
+
+// CHECK: vpermt2d -2064(%rdx), %xmm29, %xmm21
+// CHECK:  encoding: [0x62,0xe2,0x15,0x00,0x7e,0xaa,0xf0,0xf7,0xff,0xff]
+          vpermt2d -2064(%rdx), %xmm29, %xmm21
+
+// CHECK: vpermt2d 508(%rdx){1to4}, %xmm29, %xmm21
+// CHECK:  encoding: [0x62,0xe2,0x15,0x10,0x7e,0x6a,0x7f]
+          vpermt2d 508(%rdx){1to4}, %xmm29, %xmm21
+
+// CHECK: vpermt2d 512(%rdx){1to4}, %xmm29, %xmm21
+// CHECK:  encoding: [0x62,0xe2,0x15,0x10,0x7e,0xaa,0x00,0x02,0x00,0x00]
+          vpermt2d 512(%rdx){1to4}, %xmm29, %xmm21
+
+// CHECK: vpermt2d -512(%rdx){1to4}, %xmm29, %xmm21
+// CHECK:  encoding: [0x62,0xe2,0x15,0x10,0x7e,0x6a,0x80]
+          vpermt2d -512(%rdx){1to4}, %xmm29, %xmm21
+
+// CHECK: vpermt2d -516(%rdx){1to4}, %xmm29, %xmm21
+// CHECK:  encoding: [0x62,0xe2,0x15,0x10,0x7e,0xaa,0xfc,0xfd,0xff,0xff]
+          vpermt2d -516(%rdx){1to4}, %xmm29, %xmm21
+
+// CHECK: vpermt2d %ymm21, %ymm26, %ymm22
+// CHECK:  encoding: [0x62,0xa2,0x2d,0x20,0x7e,0xf5]
+          vpermt2d %ymm21, %ymm26, %ymm22
+
+// CHECK: vpermt2d %ymm21, %ymm26, %ymm22 {%k2}
+// CHECK:  encoding: [0x62,0xa2,0x2d,0x22,0x7e,0xf5]
+          vpermt2d %ymm21, %ymm26, %ymm22 {%k2}
+
+// CHECK: vpermt2d %ymm21, %ymm26, %ymm22 {%k2} {z}
+// CHECK:  encoding: [0x62,0xa2,0x2d,0xa2,0x7e,0xf5]
+          vpermt2d %ymm21, %ymm26, %ymm22 {%k2} {z}
+
+// CHECK: vpermt2d (%rcx), %ymm26, %ymm22
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x20,0x7e,0x31]
+          vpermt2d (%rcx), %ymm26, %ymm22
+
+// CHECK: vpermt2d 291(%rax,%r14,8), %ymm26, %ymm22
+// CHECK:  encoding: [0x62,0xa2,0x2d,0x20,0x7e,0xb4,0xf0,0x23,0x01,0x00,0x00]
+          vpermt2d 291(%rax,%r14,8), %ymm26, %ymm22
+
+// CHECK: vpermt2d (%rcx){1to8}, %ymm26, %ymm22
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x30,0x7e,0x31]
+          vpermt2d (%rcx){1to8}, %ymm26, %ymm22
+
+// CHECK: vpermt2d 4064(%rdx), %ymm26, %ymm22
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x20,0x7e,0x72,0x7f]
+          vpermt2d 4064(%rdx), %ymm26, %ymm22
+
+// CHECK: vpermt2d 4096(%rdx), %ymm26, %ymm22
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x20,0x7e,0xb2,0x00,0x10,0x00,0x00]
+          vpermt2d 4096(%rdx), %ymm26, %ymm22
+
+// CHECK: vpermt2d -4096(%rdx), %ymm26, %ymm22
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x20,0x7e,0x72,0x80]
+          vpermt2d -4096(%rdx), %ymm26, %ymm22
+
+// CHECK: vpermt2d -4128(%rdx), %ymm26, %ymm22
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x20,0x7e,0xb2,0xe0,0xef,0xff,0xff]
+          vpermt2d -4128(%rdx), %ymm26, %ymm22
+
+// CHECK: vpermt2d 508(%rdx){1to8}, %ymm26, %ymm22
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x30,0x7e,0x72,0x7f]
+          vpermt2d 508(%rdx){1to8}, %ymm26, %ymm22
+
+// CHECK: vpermt2d 512(%rdx){1to8}, %ymm26, %ymm22
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x30,0x7e,0xb2,0x00,0x02,0x00,0x00]
+          vpermt2d 512(%rdx){1to8}, %ymm26, %ymm22
+
+// CHECK: vpermt2d -512(%rdx){1to8}, %ymm26, %ymm22
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x30,0x7e,0x72,0x80]
+          vpermt2d -512(%rdx){1to8}, %ymm26, %ymm22
+
+// CHECK: vpermt2d -516(%rdx){1to8}, %ymm26, %ymm22
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x30,0x7e,0xb2,0xfc,0xfd,0xff,0xff]
+          vpermt2d -516(%rdx){1to8}, %ymm26, %ymm22
+
+// CHECK: vpermt2q %xmm18, %xmm25, %xmm22
+// CHECK:  encoding: [0x62,0xa2,0xb5,0x00,0x7e,0xf2]
+          vpermt2q %xmm18, %xmm25, %xmm22
+
+// CHECK: vpermt2q %xmm18, %xmm25, %xmm22 {%k1}
+// CHECK:  encoding: [0x62,0xa2,0xb5,0x01,0x7e,0xf2]
+          vpermt2q %xmm18, %xmm25, %xmm22 {%k1}
+
+// CHECK: vpermt2q %xmm18, %xmm25, %xmm22 {%k1} {z}
+// CHECK:  encoding: [0x62,0xa2,0xb5,0x81,0x7e,0xf2]
+          vpermt2q %xmm18, %xmm25, %xmm22 {%k1} {z}
+
+// CHECK: vpermt2q (%rcx), %xmm25, %xmm22
+// CHECK:  encoding: [0x62,0xe2,0xb5,0x00,0x7e,0x31]
+          vpermt2q (%rcx), %xmm25, %xmm22
+
+// CHECK: vpermt2q 291(%rax,%r14,8), %xmm25, %xmm22
+// CHECK:  encoding: [0x62,0xa2,0xb5,0x00,0x7e,0xb4,0xf0,0x23,0x01,0x00,0x00]
+          vpermt2q 291(%rax,%r14,8), %xmm25, %xmm22
+
+// CHECK: vpermt2q (%rcx){1to2}, %xmm25, %xmm22
+// CHECK:  encoding: [0x62,0xe2,0xb5,0x10,0x7e,0x31]
+          vpermt2q (%rcx){1to2}, %xmm25, %xmm22
+
+// CHECK: vpermt2q 2032(%rdx), %xmm25, %xmm22
+// CHECK:  encoding: [0x62,0xe2,0xb5,0x00,0x7e,0x72,0x7f]
+          vpermt2q 2032(%rdx), %xmm25, %xmm22
+
+// CHECK: vpermt2q 2048(%rdx), %xmm25, %xmm22
+// CHECK:  encoding: [0x62,0xe2,0xb5,0x00,0x7e,0xb2,0x00,0x08,0x00,0x00]
+          vpermt2q 2048(%rdx), %xmm25, %xmm22
+
+// CHECK: vpermt2q -2048(%rdx), %xmm25, %xmm22
+// CHECK:  encoding: [0x62,0xe2,0xb5,0x00,0x7e,0x72,0x80]
+          vpermt2q -2048(%rdx), %xmm25, %xmm22
+
+// CHECK: vpermt2q -2064(%rdx), %xmm25, %xmm22
+// CHECK:  encoding: [0x62,0xe2,0xb5,0x00,0x7e,0xb2,0xf0,0xf7,0xff,0xff]
+          vpermt2q -2064(%rdx), %xmm25, %xmm22
+
+// CHECK: vpermt2q 1016(%rdx){1to2}, %xmm25, %xmm22
+// CHECK:  encoding: [0x62,0xe2,0xb5,0x10,0x7e,0x72,0x7f]
+          vpermt2q 1016(%rdx){1to2}, %xmm25, %xmm22
+
+// CHECK: vpermt2q 1024(%rdx){1to2}, %xmm25, %xmm22
+// CHECK:  encoding: [0x62,0xe2,0xb5,0x10,0x7e,0xb2,0x00,0x04,0x00,0x00]
+          vpermt2q 1024(%rdx){1to2}, %xmm25, %xmm22
+
+// CHECK: vpermt2q -1024(%rdx){1to2}, %xmm25, %xmm22
+// CHECK:  encoding: [0x62,0xe2,0xb5,0x10,0x7e,0x72,0x80]
+          vpermt2q -1024(%rdx){1to2}, %xmm25, %xmm22
+
+// CHECK: vpermt2q -1032(%rdx){1to2}, %xmm25, %xmm22
+// CHECK:  encoding: [0x62,0xe2,0xb5,0x10,0x7e,0xb2,0xf8,0xfb,0xff,0xff]
+          vpermt2q -1032(%rdx){1to2}, %xmm25, %xmm22
+
+// CHECK: vpermt2q %ymm20, %ymm17, %ymm18
+// CHECK:  encoding: [0x62,0xa2,0xf5,0x20,0x7e,0xd4]
+          vpermt2q %ymm20, %ymm17, %ymm18
+
+// CHECK: vpermt2q %ymm20, %ymm17, %ymm18 {%k6}
+// CHECK:  encoding: [0x62,0xa2,0xf5,0x26,0x7e,0xd4]
+          vpermt2q %ymm20, %ymm17, %ymm18 {%k6}
+
+// CHECK: vpermt2q %ymm20, %ymm17, %ymm18 {%k6} {z}
+// CHECK:  encoding: [0x62,0xa2,0xf5,0xa6,0x7e,0xd4]
+          vpermt2q %ymm20, %ymm17, %ymm18 {%k6} {z}
+
+// CHECK: vpermt2q (%rcx), %ymm17, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0xf5,0x20,0x7e,0x11]
+          vpermt2q (%rcx), %ymm17, %ymm18
+
+// CHECK: vpermt2q 291(%rax,%r14,8), %ymm17, %ymm18
+// CHECK:  encoding: [0x62,0xa2,0xf5,0x20,0x7e,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpermt2q 291(%rax,%r14,8), %ymm17, %ymm18
+
+// CHECK: vpermt2q (%rcx){1to4}, %ymm17, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0xf5,0x30,0x7e,0x11]
+          vpermt2q (%rcx){1to4}, %ymm17, %ymm18
+
+// CHECK: vpermt2q 4064(%rdx), %ymm17, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0xf5,0x20,0x7e,0x52,0x7f]
+          vpermt2q 4064(%rdx), %ymm17, %ymm18
+
+// CHECK: vpermt2q 4096(%rdx), %ymm17, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0xf5,0x20,0x7e,0x92,0x00,0x10,0x00,0x00]
+          vpermt2q 4096(%rdx), %ymm17, %ymm18
+
+// CHECK: vpermt2q -4096(%rdx), %ymm17, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0xf5,0x20,0x7e,0x52,0x80]
+          vpermt2q -4096(%rdx), %ymm17, %ymm18
+
+// CHECK: vpermt2q -4128(%rdx), %ymm17, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0xf5,0x20,0x7e,0x92,0xe0,0xef,0xff,0xff]
+          vpermt2q -4128(%rdx), %ymm17, %ymm18
+
+// CHECK: vpermt2q 1016(%rdx){1to4}, %ymm17, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0xf5,0x30,0x7e,0x52,0x7f]
+          vpermt2q 1016(%rdx){1to4}, %ymm17, %ymm18
+
+// CHECK: vpermt2q 1024(%rdx){1to4}, %ymm17, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0xf5,0x30,0x7e,0x92,0x00,0x04,0x00,0x00]
+          vpermt2q 1024(%rdx){1to4}, %ymm17, %ymm18
+
+// CHECK: vpermt2q -1024(%rdx){1to4}, %ymm17, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0xf5,0x30,0x7e,0x52,0x80]
+          vpermt2q -1024(%rdx){1to4}, %ymm17, %ymm18
+
+// CHECK: vpermt2q -1032(%rdx){1to4}, %ymm17, %ymm18
+// CHECK:  encoding: [0x62,0xe2,0xf5,0x30,0x7e,0x92,0xf8,0xfb,0xff,0xff]
+          vpermt2q -1032(%rdx){1to4}, %ymm17, %ymm18
+
+// CHECK: vpermt2ps %xmm18, %xmm23, %xmm19
+// CHECK:  encoding: [0x62,0xa2,0x45,0x00,0x7f,0xda]
+          vpermt2ps %xmm18, %xmm23, %xmm19
+
+// CHECK: vpermt2ps %xmm18, %xmm23, %xmm19 {%k1}
+// CHECK:  encoding: [0x62,0xa2,0x45,0x01,0x7f,0xda]
+          vpermt2ps %xmm18, %xmm23, %xmm19 {%k1}
+
+// CHECK: vpermt2ps %xmm18, %xmm23, %xmm19 {%k1} {z}
+// CHECK:  encoding: [0x62,0xa2,0x45,0x81,0x7f,0xda]
+          vpermt2ps %xmm18, %xmm23, %xmm19 {%k1} {z}
+
+// CHECK: vpermt2ps (%rcx), %xmm23, %xmm19
+// CHECK:  encoding: [0x62,0xe2,0x45,0x00,0x7f,0x19]
+          vpermt2ps (%rcx), %xmm23, %xmm19
+
+// CHECK: vpermt2ps 291(%rax,%r14,8), %xmm23, %xmm19
+// CHECK:  encoding: [0x62,0xa2,0x45,0x00,0x7f,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vpermt2ps 291(%rax,%r14,8), %xmm23, %xmm19
+
+// CHECK: vpermt2ps (%rcx){1to4}, %xmm23, %xmm19
+// CHECK:  encoding: [0x62,0xe2,0x45,0x10,0x7f,0x19]
+          vpermt2ps (%rcx){1to4}, %xmm23, %xmm19
+
+// CHECK: vpermt2ps 2032(%rdx), %xmm23, %xmm19
+// CHECK:  encoding: [0x62,0xe2,0x45,0x00,0x7f,0x5a,0x7f]
+          vpermt2ps 2032(%rdx), %xmm23, %xmm19
+
+// CHECK: vpermt2ps 2048(%rdx), %xmm23, %xmm19
+// CHECK:  encoding: [0x62,0xe2,0x45,0x00,0x7f,0x9a,0x00,0x08,0x00,0x00]
+          vpermt2ps 2048(%rdx), %xmm23, %xmm19
+
+// CHECK: vpermt2ps -2048(%rdx), %xmm23, %xmm19
+// CHECK:  encoding: [0x62,0xe2,0x45,0x00,0x7f,0x5a,0x80]
+          vpermt2ps -2048(%rdx), %xmm23, %xmm19
+
+// CHECK: vpermt2ps -2064(%rdx), %xmm23, %xmm19
+// CHECK:  encoding: [0x62,0xe2,0x45,0x00,0x7f,0x9a,0xf0,0xf7,0xff,0xff]
+          vpermt2ps -2064(%rdx), %xmm23, %xmm19
+
+// CHECK: vpermt2ps 508(%rdx){1to4}, %xmm23, %xmm19
+// CHECK:  encoding: [0x62,0xe2,0x45,0x10,0x7f,0x5a,0x7f]
+          vpermt2ps 508(%rdx){1to4}, %xmm23, %xmm19
+
+// CHECK: vpermt2ps 512(%rdx){1to4}, %xmm23, %xmm19
+// CHECK:  encoding: [0x62,0xe2,0x45,0x10,0x7f,0x9a,0x00,0x02,0x00,0x00]
+          vpermt2ps 512(%rdx){1to4}, %xmm23, %xmm19
+
+// CHECK: vpermt2ps -512(%rdx){1to4}, %xmm23, %xmm19
+// CHECK:  encoding: [0x62,0xe2,0x45,0x10,0x7f,0x5a,0x80]
+          vpermt2ps -512(%rdx){1to4}, %xmm23, %xmm19
+
+// CHECK: vpermt2ps -516(%rdx){1to4}, %xmm23, %xmm19
+// CHECK:  encoding: [0x62,0xe2,0x45,0x10,0x7f,0x9a,0xfc,0xfd,0xff,0xff]
+          vpermt2ps -516(%rdx){1to4}, %xmm23, %xmm19
+
+// CHECK: vpermt2ps %ymm21, %ymm27, %ymm26
+// CHECK:  encoding: [0x62,0x22,0x25,0x20,0x7f,0xd5]
+          vpermt2ps %ymm21, %ymm27, %ymm26
+
+// CHECK: vpermt2ps %ymm21, %ymm27, %ymm26 {%k3}
+// CHECK:  encoding: [0x62,0x22,0x25,0x23,0x7f,0xd5]
+          vpermt2ps %ymm21, %ymm27, %ymm26 {%k3}
+
+// CHECK: vpermt2ps %ymm21, %ymm27, %ymm26 {%k3} {z}
+// CHECK:  encoding: [0x62,0x22,0x25,0xa3,0x7f,0xd5]
+          vpermt2ps %ymm21, %ymm27, %ymm26 {%k3} {z}
+
+// CHECK: vpermt2ps (%rcx), %ymm27, %ymm26
+// CHECK:  encoding: [0x62,0x62,0x25,0x20,0x7f,0x11]
+          vpermt2ps (%rcx), %ymm27, %ymm26
+
+// CHECK: vpermt2ps 291(%rax,%r14,8), %ymm27, %ymm26
+// CHECK:  encoding: [0x62,0x22,0x25,0x20,0x7f,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpermt2ps 291(%rax,%r14,8), %ymm27, %ymm26
+
+// CHECK: vpermt2ps (%rcx){1to8}, %ymm27, %ymm26
+// CHECK:  encoding: [0x62,0x62,0x25,0x30,0x7f,0x11]
+          vpermt2ps (%rcx){1to8}, %ymm27, %ymm26
+
+// CHECK: vpermt2ps 4064(%rdx), %ymm27, %ymm26
+// CHECK:  encoding: [0x62,0x62,0x25,0x20,0x7f,0x52,0x7f]
+          vpermt2ps 4064(%rdx), %ymm27, %ymm26
+
+// CHECK: vpermt2ps 4096(%rdx), %ymm27, %ymm26
+// CHECK:  encoding: [0x62,0x62,0x25,0x20,0x7f,0x92,0x00,0x10,0x00,0x00]
+          vpermt2ps 4096(%rdx), %ymm27, %ymm26
+
+// CHECK: vpermt2ps -4096(%rdx), %ymm27, %ymm26
+// CHECK:  encoding: [0x62,0x62,0x25,0x20,0x7f,0x52,0x80]
+          vpermt2ps -4096(%rdx), %ymm27, %ymm26
+
+// CHECK: vpermt2ps -4128(%rdx), %ymm27, %ymm26
+// CHECK:  encoding: [0x62,0x62,0x25,0x20,0x7f,0x92,0xe0,0xef,0xff,0xff]
+          vpermt2ps -4128(%rdx), %ymm27, %ymm26
+
+// CHECK: vpermt2ps 508(%rdx){1to8}, %ymm27, %ymm26
+// CHECK:  encoding: [0x62,0x62,0x25,0x30,0x7f,0x52,0x7f]
+          vpermt2ps 508(%rdx){1to8}, %ymm27, %ymm26
+
+// CHECK: vpermt2ps 512(%rdx){1to8}, %ymm27, %ymm26
+// CHECK:  encoding: [0x62,0x62,0x25,0x30,0x7f,0x92,0x00,0x02,0x00,0x00]
+          vpermt2ps 512(%rdx){1to8}, %ymm27, %ymm26
+
+// CHECK: vpermt2ps -512(%rdx){1to8}, %ymm27, %ymm26
+// CHECK:  encoding: [0x62,0x62,0x25,0x30,0x7f,0x52,0x80]
+          vpermt2ps -512(%rdx){1to8}, %ymm27, %ymm26
+
+// CHECK: vpermt2ps -516(%rdx){1to8}, %ymm27, %ymm26
+// CHECK:  encoding: [0x62,0x62,0x25,0x30,0x7f,0x92,0xfc,0xfd,0xff,0xff]
+          vpermt2ps -516(%rdx){1to8}, %ymm27, %ymm26
+
+// CHECK: vpermt2pd %xmm17, %xmm21, %xmm26
+// CHECK:  encoding: [0x62,0x22,0xd5,0x00,0x7f,0xd1]
+          vpermt2pd %xmm17, %xmm21, %xmm26
+
+// CHECK: vpermt2pd %xmm17, %xmm21, %xmm26 {%k5}
+// CHECK:  encoding: [0x62,0x22,0xd5,0x05,0x7f,0xd1]
+          vpermt2pd %xmm17, %xmm21, %xmm26 {%k5}
+
+// CHECK: vpermt2pd %xmm17, %xmm21, %xmm26 {%k5} {z}
+// CHECK:  encoding: [0x62,0x22,0xd5,0x85,0x7f,0xd1]
+          vpermt2pd %xmm17, %xmm21, %xmm26 {%k5} {z}
+
+// CHECK: vpermt2pd (%rcx), %xmm21, %xmm26
+// CHECK:  encoding: [0x62,0x62,0xd5,0x00,0x7f,0x11]
+          vpermt2pd (%rcx), %xmm21, %xmm26
+
+// CHECK: vpermt2pd 291(%rax,%r14,8), %xmm21, %xmm26
+// CHECK:  encoding: [0x62,0x22,0xd5,0x00,0x7f,0x94,0xf0,0x23,0x01,0x00,0x00]
+          vpermt2pd 291(%rax,%r14,8), %xmm21, %xmm26
+
+// CHECK: vpermt2pd (%rcx){1to2}, %xmm21, %xmm26
+// CHECK:  encoding: [0x62,0x62,0xd5,0x10,0x7f,0x11]
+          vpermt2pd (%rcx){1to2}, %xmm21, %xmm26
+
+// CHECK: vpermt2pd 2032(%rdx), %xmm21, %xmm26
+// CHECK:  encoding: [0x62,0x62,0xd5,0x00,0x7f,0x52,0x7f]
+          vpermt2pd 2032(%rdx), %xmm21, %xmm26
+
+// CHECK: vpermt2pd 2048(%rdx), %xmm21, %xmm26
+// CHECK:  encoding: [0x62,0x62,0xd5,0x00,0x7f,0x92,0x00,0x08,0x00,0x00]
+          vpermt2pd 2048(%rdx), %xmm21, %xmm26
+
+// CHECK: vpermt2pd -2048(%rdx), %xmm21, %xmm26
+// CHECK:  encoding: [0x62,0x62,0xd5,0x00,0x7f,0x52,0x80]
+          vpermt2pd -2048(%rdx), %xmm21, %xmm26
+
+// CHECK: vpermt2pd -2064(%rdx), %xmm21, %xmm26
+// CHECK:  encoding: [0x62,0x62,0xd5,0x00,0x7f,0x92,0xf0,0xf7,0xff,0xff]
+          vpermt2pd -2064(%rdx), %xmm21, %xmm26
+
+// CHECK: vpermt2pd 1016(%rdx){1to2}, %xmm21, %xmm26
+// CHECK:  encoding: [0x62,0x62,0xd5,0x10,0x7f,0x52,0x7f]
+          vpermt2pd 1016(%rdx){1to2}, %xmm21, %xmm26
+
+// CHECK: vpermt2pd 1024(%rdx){1to2}, %xmm21, %xmm26
+// CHECK:  encoding: [0x62,0x62,0xd5,0x10,0x7f,0x92,0x00,0x04,0x00,0x00]
+          vpermt2pd 1024(%rdx){1to2}, %xmm21, %xmm26
+
+// CHECK: vpermt2pd -1024(%rdx){1to2}, %xmm21, %xmm26
+// CHECK:  encoding: [0x62,0x62,0xd5,0x10,0x7f,0x52,0x80]
+          vpermt2pd -1024(%rdx){1to2}, %xmm21, %xmm26
+
+// CHECK: vpermt2pd -1032(%rdx){1to2}, %xmm21, %xmm26
+// CHECK:  encoding: [0x62,0x62,0xd5,0x10,0x7f,0x92,0xf8,0xfb,0xff,0xff]
+          vpermt2pd -1032(%rdx){1to2}, %xmm21, %xmm26
+
+// CHECK: vpermt2pd %ymm17, %ymm23, %ymm17
+// CHECK:  encoding: [0x62,0xa2,0xc5,0x20,0x7f,0xc9]
+          vpermt2pd %ymm17, %ymm23, %ymm17
+
+// CHECK: vpermt2pd %ymm17, %ymm23, %ymm17 {%k1}
+// CHECK:  encoding: [0x62,0xa2,0xc5,0x21,0x7f,0xc9]
+          vpermt2pd %ymm17, %ymm23, %ymm17 {%k1}
+
+// CHECK: vpermt2pd %ymm17, %ymm23, %ymm17 {%k1} {z}
+// CHECK:  encoding: [0x62,0xa2,0xc5,0xa1,0x7f,0xc9]
+          vpermt2pd %ymm17, %ymm23, %ymm17 {%k1} {z}
+
+// CHECK: vpermt2pd (%rcx), %ymm23, %ymm17
+// CHECK:  encoding: [0x62,0xe2,0xc5,0x20,0x7f,0x09]
+          vpermt2pd (%rcx), %ymm23, %ymm17
+
+// CHECK: vpermt2pd 291(%rax,%r14,8), %ymm23, %ymm17
+// CHECK:  encoding: [0x62,0xa2,0xc5,0x20,0x7f,0x8c,0xf0,0x23,0x01,0x00,0x00]
+          vpermt2pd 291(%rax,%r14,8), %ymm23, %ymm17
+
+// CHECK: vpermt2pd (%rcx){1to4}, %ymm23, %ymm17
+// CHECK:  encoding: [0x62,0xe2,0xc5,0x30,0x7f,0x09]
+          vpermt2pd (%rcx){1to4}, %ymm23, %ymm17
+
+// CHECK: vpermt2pd 4064(%rdx), %ymm23, %ymm17
+// CHECK:  encoding: [0x62,0xe2,0xc5,0x20,0x7f,0x4a,0x7f]
+          vpermt2pd 4064(%rdx), %ymm23, %ymm17
+
+// CHECK: vpermt2pd 4096(%rdx), %ymm23, %ymm17
+// CHECK:  encoding: [0x62,0xe2,0xc5,0x20,0x7f,0x8a,0x00,0x10,0x00,0x00]
+          vpermt2pd 4096(%rdx), %ymm23, %ymm17
+
+// CHECK: vpermt2pd -4096(%rdx), %ymm23, %ymm17
+// CHECK:  encoding: [0x62,0xe2,0xc5,0x20,0x7f,0x4a,0x80]
+          vpermt2pd -4096(%rdx), %ymm23, %ymm17
+
+// CHECK: vpermt2pd -4128(%rdx), %ymm23, %ymm17
+// CHECK:  encoding: [0x62,0xe2,0xc5,0x20,0x7f,0x8a,0xe0,0xef,0xff,0xff]
+          vpermt2pd -4128(%rdx), %ymm23, %ymm17
+
+// CHECK: vpermt2pd 1016(%rdx){1to4}, %ymm23, %ymm17
+// CHECK:  encoding: [0x62,0xe2,0xc5,0x30,0x7f,0x4a,0x7f]
+          vpermt2pd 1016(%rdx){1to4}, %ymm23, %ymm17
+
+// CHECK: vpermt2pd 1024(%rdx){1to4}, %ymm23, %ymm17
+// CHECK:  encoding: [0x62,0xe2,0xc5,0x30,0x7f,0x8a,0x00,0x04,0x00,0x00]
+          vpermt2pd 1024(%rdx){1to4}, %ymm23, %ymm17
+
+// CHECK: vpermt2pd -1024(%rdx){1to4}, %ymm23, %ymm17
+// CHECK:  encoding: [0x62,0xe2,0xc5,0x30,0x7f,0x4a,0x80]
+          vpermt2pd -1024(%rdx){1to4}, %ymm23, %ymm17
+
+// CHECK: vpermt2pd -1032(%rdx){1to4}, %ymm23, %ymm17
+// CHECK:  encoding: [0x62,0xe2,0xc5,0x30,0x7f,0x8a,0xf8,0xfb,0xff,0xff]
+          vpermt2pd -1032(%rdx){1to4}, %ymm23, %ymm17