VADDPD 2032(DX), X29, X0        // 62f1950058427f
	VADDPD 2032(DX), X1, X29        // 6261f508586a7f
	VADDPD 2032(DX), X29, X28       // 6261950058627f
	VADDPD 2032(DX)(AX*2), X29, X0  // 62f195005844427f
	VADDPD 2032(DX)(AX*2), X1, X29  // 6261f508586c427f
	VADDPD 2032(DX)(AX*2), X29, X28 // 626195005864427f
	VADDPD 4064(DX), Y0, Y29        // 6261fd28586a7f
	VADDPD 4064(DX), Y29, Y1        // 62f19520584a7f

	VP4DPWSSD 15(DX)(BX*8), [Z2-Z5], K4, Z17           // 62e26f4c528cda0f000000
	VP4DPWSSD 7(SI)(DI*1), [Z12-Z15], K4, Z17          // 62e21f4c528c3e07000000
	VP4DPWSSD 15(DX)(BX*8), [Z12-Z15], K4, Z17         // 62e21f4c528cda0f000000
	VP4DPWSSD 7(SI)(DI*1), [Z22-Z25], K4, Z17          // 62e24f44528c3e07000000
	VP4DPWSSD 15(DX)(BX*8), [Z22-Z25], K4, Z17         // 62e24f44528cda0f000000

32(BX), AX MULQ SI ADDQ 32(CX), AX ADCQ $0x00, DX ADDQ DI, AX ADCQ $0x00, DX MOVQ DX, DI MOVQ AX, 32(CX) // Iteration 5 MOVQ 40(BX), AX MULQ SI ADDQ 40(CX), AX ADCQ $0x00, DX ADDQ DI, AX ADCQ $0x00, DX MOVQ DX, DI MOVQ AX, 40(CX) // Iteration 6 MOVQ 48(BX), AX MULQ SI ADDQ 48(CX), AX ADCQ $0x00, DX ADDQ DI, AX ADCQ $0x00, DX MOVQ DX, DI MOVQ AX, 48(CX) // Iteration 7 MOVQ 56(BX), AX MULQ SI ADDQ 56(CX), AX ADCQ $0x00, DX ADDQ DI, AX ADCQ $0x00, DX MOVQ DX, DI MOVQ AX, 56(CX) // Iteration 8 MOVQ 64(BX),...

	VPCOMPRESSB X13, K1, X28                           // 62127d0963ec
	VPCOMPRESSB X8, K1, X28                            // 62127d0963c4
	VPCOMPRESSB X7, K1, -7(CX)(DX*1)                   // 62f27d09637c11f9
	VPCOMPRESSB X13, K1, -7(CX)(DX*1)                  // 62727d09636c11f9
	VPCOMPRESSB X8, K1, -7(CX)(DX*1)                   // 62727d09634411f9
	VPCOMPRESSB X7, K1, -15(R14)(R15*4)                // 62927d09637cbef1

	VRCP28PS 15(DX)(BX*1), K3, Z28                     // 62627d4bcaa41a0f000000
	VRCP28PS -7(CX)(DX*2), K3, Z28                     // 62627d4bcaa451f9ffffff
	VRCP28PS Z13, K3, Z6                               // 62d27d4bcaf5
	VRCP28PS Z21, K3, Z6                               // 62b27d4bcaf5
	VRCP28PS 15(DX)(BX*1), K3, Z6                      // 62f27d4bcab41a0f000000

	VBROADCASTF64X2 15(DX)(BX*1), K1, Z14              // 6272fd491ab41a0f000000
	VBROADCASTF64X2 -7(CX)(DX*2), K1, Z14              // 6272fd491ab451f9ffffff
	VBROADCASTF64X2 15(DX)(BX*1), K1, Z28              // 6262fd491aa41a0f000000
	VBROADCASTF64X2 -7(CX)(DX*2), K1, Z28              // 6262fd491aa451f9ffffff
	VBROADCASTI32X2 X14, K1, X19                       // 62c27d0959de

	VGF2P8AFFINEINVQB $47, (DX), Z14, K2, Z28          // 62638d4acf222f
	VGF2P8AFFINEINVQB $47, Z3, Z28, K2, Z28            // 62639d42cfe32f
	VGF2P8AFFINEINVQB $47, Z12, Z28, K2, Z28           // 62439d42cfe42f
	VGF2P8AFFINEINVQB $47, 99(R15)(R15*1), Z28, K2, Z28 // 62039d42cfa43f630000002f
	VGF2P8AFFINEINVQB $47, (DX), Z28, K2, Z28          // 62639d42cf222f

	VAESDEC (DX), Y0, Y23                              // 62e27d28de3a or 62e2fd28de3a
	VAESDEC Z27, Z3, Z11                               // 62126548dedb or 6212e548dedb
	VAESDEC Z15, Z3, Z11                               // 62526548dedf or 6252e548dedf
	VAESDEC 99(R15)(R15*1), Z3, Z11                    // 62126548de9c3f63000000 or 6212e548de9c3f63000000

	VPLZCNTD 15(DX)(BX*1), K3, X17                     // 62e27d0b448c1a0f000000
	VPLZCNTD -7(CX)(DX*2), K3, X17                     // 62e27d0b448c51f9ffffff
	VPLZCNTD X3, K3, X15                               // 62727d0b44fb
	VPLZCNTD X26, K3, X15                              // 62127d0b44fa
	VPLZCNTD X23, K3, X15                              // 62327d0b44ff

	V4FNMADDPS 99(R15)(R15*1), [Z1-Z4], K3, Z15        // 6212774baabc3f63000000
	V4FNMADDPS (DX), [Z1-Z4], K3, Z15                  // 6272774baa3a
	V4FNMADDPS 99(R15)(R15*1), [Z11-Z14], K3, Z15      // 6212274baabc3f63000000
	V4FNMADDPS (DX), [Z11-Z14], K3, Z15                // 6272274baa3a
	V4FNMADDPS 99(R15)(R15*1), [Z21-Z24], K3, Z15      // 62125743aabc3f63000000
	V4FNMADDPS (DX), [Z21-Z24], K3, Z15                // 62725743aa3a