VMOVDQU32 X8, K3, X19                              // 62317e0b7fc3
	VMOVDQU32 X8, K3, (R14)                            // 62517e0b7f06
	VMOVDQU32 X8, K3, -7(DI)(R8*8)                     // 62317e0b7f84c7f9ffffff
	VMOVDQU32 X26, K4, X8                              // 62417e0c7fd0
	VMOVDQU32 99(R15)(R15*4), K4, X8                   // 62117e0c6f84bf63000000

	VPERMB X0, X2, K1, X8                              // 62726d098dc0
	VPERMB 7(SI)(DI*4), X2, K1, X8                     // 62726d098d84be07000000
	VPERMB -7(DI)(R8*2), X2, K1, X8                    // 62326d098d8447f9ffffff
	VPERMB X26, X9, K1, X8                             // 621235098dc2
	VPERMB X19, X9, K1, X8                             // 623235098dc3

	VRCP28SS X26, X2, K3, X8                           // 62126d0bcbc2 or 62126d2bcbc2 or 62126d4bcbc2
	VRCP28SS X19, X2, K3, X8                           // 62326d0bcbc3 or 62326d2bcbc3 or 62326d4bcbc3
	VRCP28SS X0, X2, K3, X8                            // 62726d0bcbc0 or 62726d2bcbc0 or 62726d4bcbc0

	VPSHUFBITQMB X24, X8, K6, K0                       // 62923d0e8fc0
	VPSHUFBITQMB X7, X8, K6, K0                        // 62f23d0e8fc7
	VPSHUFBITQMB X0, X8, K6, K0                        // 62f23d0e8fc0
	VPSHUFBITQMB (R8), X8, K6, K0                      // 62d23d0e8f00
	VPSHUFBITQMB 15(DX)(BX*2), X8, K6, K0              // 62f23d0e8f845a0f000000

	VPCONFLICTQ X11, K4, X8                            // 6252fd0cc4c3
	VPCONFLICTQ X16, K4, X8                            // 6232fd0cc4c0
	VPCONFLICTQ X6, K4, X8                             // 6272fd0cc4c6
	VPCONFLICTQ 15(R8)(R14*8), K4, X8                  // 6212fd0cc484f00f000000
	VPCONFLICTQ -15(R14)(R15*2), K4, X8                // 6212fd0cc4847ef1ffffff

	VANDPD X11, X7, K7, X8                             // 6251c50f54c3
	VANDPD -17(BP)(SI*2), X7, K7, X8                   // 6271c50f548475efffffff
	VANDPD 7(AX)(CX*2), X7, K7, X8                     // 6271c50f54844807000000
	VANDPD X22, X0, K7, X8                             // 6231fd0f54c6
	VANDPD X1, X0, K7, X8                              // 6271fd0f54c1

	VPDPBUSD -15(BX), X28, K2, X6                      // 62f21d0250b3f1ffffff
	VPDPBUSD X15, X8, K2, X6                           // 62d23d0a50f7
	VPDPBUSD X11, X8, K2, X6                           // 62d23d0a50f3
	VPDPBUSD X1, X8, K2, X6                            // 62f23d0a50f1
	VPDPBUSD -15(R14)(R15*1), X8, K2, X6               // 62923d0a50b43ef1ffffff
	VPDPBUSD -15(BX), X8, K2, X6                       // 62f23d0a50b3f1ffffff

	VMOVHPD 7(DX), X8, X8            // c539164207 or 6271bd28168207000000 or 6271bd48168207000000
	VMOVHPD -15(R11)(CX*1), X8, X8   // c4413916440bf1 or 6251bd2816840bf1ffffff or 6251bd4816840bf1ffffff
	VMOVHPD (SP)(AX*2), X8, X8       // c539160444 or 6271bd28160444 or 6271bd48160444
	VMOVHPD (AX), X20, X8            // 6271dd001600 or 6271dd201600 or 6271dd401600

TEXT asmtest_avx512_vpopcntdq(SB), NOSPLIT, $0
	VPOPCNTD X12, K2, X8                               // 62527d0a55c4
	VPOPCNTD X16, K2, X8                               // 62327d0a55c0
	VPOPCNTD X23, K2, X8                               // 62327d0a55c7
	VPOPCNTD (R14), K2, X8                             // 62527d0a5506
	VPOPCNTD -7(DI)(R8*8), K2, X8                      // 62327d0a5584c7f9ffffff

	VAESENCLAST 7(SI), X16, X8                         // 62727d00dd8607000000 or 6272fd00dd8607000000
	VAESENCLAST X23, X23, X8                           // 62324500ddc7 or 6232c500ddc7
	VAESENCLAST X11, X23, X8                           // 62524500ddc3 or 6252c500ddc3
	VAESENCLAST X31, X23, X8                           // 62124500ddc7 or 6212c500ddc7