ANDQ BX, R12
	ADDQ R13, R15

	ORQ  R12, DI
	ADDQ R14, R10

	ADDQ R15, CX
	ADDQ R15, R10
	ADDQ DI, R10

	VPSRLQ $19, Y2, Y3
	VPSLLQ $(64-19), Y2, Y1
	VPOR   Y1, Y3, Y3
	VPXOR  Y3, Y8, Y8
	VPSRLQ $61, Y2, Y3
	VPSLLQ $(64-61), Y2, Y1
	VPOR   Y1, Y3, Y3
	VPXOR  Y3, Y8, Y8

	VPADDQ Y8, Y4, Y4

	VPSRLQ $6, Y4, Y8

	MOVQ  R10, DI
	RORXQ $41, CX, R13

	VPSLLQ $94, 7(AX), K1, Z16                         // 62f1fd4173b0070000005e
	VPSLLQ $94, (DI), K1, Z16                          // 62f1fd4173375e
	VPSLLQ X15, X18, K1, X3                            // 62d1ed01f3df
	VPSLLQ -17(BP), X18, K1, X3                        // 62f1ed01f39defffffff
	VPSLLQ -15(R14)(R15*8), X18, K1, X3                // 6291ed01f39cfef1ffffff
	VPSLLQ X28, Y7, K1, Y28                            // 6201c529f3e4

	"VPSHRDVQ",
	"VPSHRDVW",
	"VPSHRDW",
	"VPSHUFB",
	"VPSHUFBITQMB",
	"VPSHUFD",
	"VPSHUFHW",
	"VPSHUFLW",
	"VPSIGNB",
	"VPSIGND",
	"VPSIGNW",
	"VPSLLD",
	"VPSLLDQ",
	"VPSLLQ",
	"VPSLLVD",
	"VPSLLVQ",
	"VPSLLVW",
	"VPSLLW",
	"VPSRAD",
	"VPSRAQ",
	"VPSRAVD",
	"VPSRAVQ",
	"VPSRAVW",
	"VPSRAW",
	"VPSRLD",
	"VPSRLDQ",
	"VPSRLQ",

	VPSLLD $-1, X1, X2           // c5e972f1ff
	VPSLLD $-1, Y1, Y2           // c5ed72f1ff
	VPSLLDQ $-1, X1, X2          // c5e973f9ff
	VPSLLDQ $-1, Y1, Y2          // c5ed73f9ff
	VPSLLQ $-1, X1, X2           // c5e973f1ff
	VPSLLQ $-1, Y1, Y2           // c5ed73f1ff
	VPSRLD $-1, X1, X2           // c5e972d1ff
	VPSRLD $-1, Y1, Y2           // c5ed72d1ff
	VPSRLDQ $-1, X1, X2          // c5e973d9ff

	VPSLLDQ $7, Y11, Y15            // c4c10573fb07
	VPSLLQ (BX), X9, X2             // c4e131f313 or c5b1f313
	VPSLLQ (R11), X9, X2            // c4c131f313
	VPSLLQ X2, X9, X2               // c4e131f3d2 or c5b1f3d2
	VPSLLQ X11, X9, X2              // c4c131f3d3
	VPSLLQ (BX), X9, X11            // c46131f31b or c531f31b
	VPSLLQ (R11), X9, X11           // c44131f31b
	VPSLLQ X2, X9, X11              // c46131f3da or c531f3da