VPMADD52HUQ Y7, Y8, K7, Y17                        // 62e2bd2fb5cf
	VPMADD52HUQ (R8), Y8, K7, Y17                      // 62c2bd2fb508
	VPMADD52HUQ 15(DX)(BX*2), Y8, K7, Y17              // 62e2bd2fb58c5a0f000000
	VPMADD52HUQ Y28, Y1, K7, Y17                       // 6282f52fb5cc
	VPMADD52HUQ Y13, Y1, K7, Y17                       // 62c2f52fb5cd
	VPMADD52HUQ Y7, Y1, K7, Y17                        // 62e2f52fb5cf

	VPERMB Y31, Y5, K7, Y22                            // 6282552f8df7
	VPERMB 17(SP)(BP*1), Y5, K7, Y22                   // 62e2552f8db42c11000000
	VPERMB -7(CX)(DX*8), Y5, K7, Y22                   // 62e2552f8db4d1f9ffffff
	VPERMB Y5, Y0, K7, Y22                             // 62e27d2f8df5
	VPERMB Y19, Y0, K7, Y22                            // 62a27d2f8df3

	VANDNPD 7(AX)(CX*8), Z6, K7, Z7                    // 62f1cd4f55bcc807000000
	VANDNPD Z20, Z0, K7, Z9                            // 6231fd4f55cc
	VANDNPD Z28, Z0, K7, Z9                            // 6211fd4f55cc
	VANDNPD 99(R15)(R15*8), Z0, K7, Z9                 // 6211fd4f558cff63000000
	VANDNPD 7(AX)(CX*8), Z0, K7, Z9                    // 6271fd4f558cc807000000

	VPCMPW $126, -7(CX)(DX*2), Z5, K7, K0              // 62f3d54f3f8451f9ffffff7e
	VPCMPW $126, Z13, Z11, K7, K7                      // 62d3a54f3ffd7e
	VPCMPW $126, Z14, Z11, K7, K7                      // 62d3a54f3ffe7e
	VPCMPW $126, 15(DX)(BX*1), Z11, K7, K7             // 62f3a54f3fbc1a0f0000007e
	VPCMPW $126, -7(CX)(DX*2), Z11, K7, K7             // 62f3a54f3fbc51f9ffffff7e

	VPSHUFBITQMB -15(BX), Z25, K7, K1                  // 62f235478f8bf1ffffff
	VPSHUFBITQMB Z3, Z6, K7, K5                        // 62f24d4f8feb
	VPSHUFBITQMB Z21, Z6, K7, K5                       // 62b24d4f8fed
	VPSHUFBITQMB -15(R14)(R15*1), Z6, K7, K5           // 62924d4f8fac3ef1ffffff
	VPSHUFBITQMB -15(BX), Z6, K7, K5                   // 62f24d4f8fabf1ffffff

	VPCONFLICTD Z11, K7, Z21                           // 62c27d4fc4eb
	VPCONFLICTD Z25, K7, Z21                           // 62827d4fc4e9
	VPCONFLICTD -15(R14)(R15*1), K7, Z21               // 62827d4fc4ac3ef1ffffff
	VPCONFLICTD -15(BX), K7, Z21                       // 62e27d4fc4abf1ffffff
	VPCONFLICTD Z11, K7, Z13                           // 62527d4fc4eb
	VPCONFLICTD Z25, K7, Z13                           // 62127d4fc4e9

	VPCOMPRESSW Y3, K7, Y31                            // 6292fd2f63df
	VPCOMPRESSW Y24, K7, Y31                           // 6202fd2f63c7
	VPCOMPRESSW Y18, K7, -7(DI)(R8*1)                  // 62a2fd2f639407f9ffffff
	VPCOMPRESSW Y3, K7, -7(DI)(R8*1)                   // 62b2fd2f639c07f9ffffff
	VPCOMPRESSW Y24, K7, -7(DI)(R8*1)                  // 6222fd2f638407f9ffffff

	VGF2P8MULB X0, X7, K7, X7                          // 62f2450fcff8
	VGF2P8MULB 15(R8), X7, K7, X7                      // 62d2450fcfb80f000000
	VGF2P8MULB (BP), X7, K7, X7                        // 62f2450fcf7d00
	VGF2P8MULB X15, X9, K7, X7                         // 62d2350fcfff
	VGF2P8MULB X12, X9, K7, X7                         // 62d2350fcffc

	VRSQRT28SS X28, X12, K7, X11                       // 62121d0fcddc
	VRSQRT28SS X8, X12, K7, X11                        // 62521d0fcdd8
	VRSQRT28SS X16, X6, K7, X16                        // 62a24d0fcdc0
	VRSQRT28SS X28, X6, K7, X16                        // 62824d0fcdc4
	VRSQRT28SS X8, X6, K7, X16                         // 62c24d0fcdc0
	VRSQRT28SS X16, X22, K7, X16                       // 62a24d07cdc0

ks, H0 // For AES-128 round keys are stored in: K0 .. K10, KLAST VLD1.P 64(H0), [K0.B16, K1.B16, K2.B16, K3.B16] VLD1.P 64(H0), [K4.B16, K5.B16, K6.B16, K7.B16] VLD1.P 48(H0), [K8.B16, K9.B16, K10.B16] VMOV K10.B16, KLAST.B16 BLT startSingles // There are at least 8 blocks to encrypt TBZ $4, NR, octetsLoop // For AES-192 round keys occupy: K0 .. K7, K10, K11, K8, K9, KLAST VMOV K8.B16, K10.B16 VMOV K9.B16, K11.B16 VMOV KLAST.B16, K8.B16 VLD1.P 16(H0), [K9.B16] VLD1.P 16(H0), [KLAST.B16] TBZ $3, NR,...