VPCONFLICTD Y20, K3, Y7                            // 62b27d2bc4fc
	VPCONFLICTD -7(CX), K3, Y7                         // 62f27d2bc4b9f9ffffff
	VPCONFLICTD 15(DX)(BX*4), K3, Y7                   // 62f27d2bc4bc9a0f000000
	VPCONFLICTD Y14, K3, Y21                           // 62c27d2bc4ee
	VPCONFLICTD Y8, K3, Y21                            // 62c27d2bc4e8

	VPSHLDVD 15(R8)(R14*2), Z24, K3, Z14               // 62123d4371b4700f000000
	VPSHLDVD Z22, Z8, K3, Z7                           // 62b23d4b71fe
	VPSHLDVD Z25, Z8, K3, Z7                           // 62923d4b71f9
	VPSHLDVD 15(R8)(R14*1), Z8, K3, Z7                 // 62923d4b71bc300f000000
	VPSHLDVD 15(R8)(R14*2), Z8, K3, Z7                 // 62923d4b71bc700f000000

	VRSQRT28PD Z28, K3, Z0                             // 6292fd4bccc4
	VRSQRT28PD (SI), K3, Z0                            // 62f2fd4bcc06
	VRSQRT28PD 7(SI)(DI*2), K3, Z0                     // 62f2fd4bcc847e07000000
	VRSQRT28PD Z20, K3, Z6                             // 62b2fd4bccf4
	VRSQRT28PD Z28, K3, Z6                             // 6292fd4bccf4
	VRSQRT28PD (SI), K3, Z6                            // 62f2fd4bcc36

	VANDNPS 7(SI), Z19, K3, Z3                         // 62f16443559e07000000
	VANDNPS Z12, Z9, K3, Z30                           // 6241344b55f4
	VANDNPS Z22, Z9, K3, Z30                           // 6221344b55f6
	VANDNPS (AX), Z9, K3, Z30                          // 6261344b5530
	VANDNPS 7(SI), Z9, K3, Z30                         // 6261344b55b607000000

	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

	VPOPCNTD Z2, K3, Z22                               // 62e27d4b55f2
	VPOPCNTD Z31, K3, Z22                              // 62827d4b55f7
	VPOPCNTD 7(SI)(DI*4), K3, Z22                      // 62e27d4b55b4be07000000
	VPOPCNTD -7(DI)(R8*2), K3, Z22                     // 62a27d4b55b447f9ffffff
	VPOPCNTD Z2, K3, Z7                                // 62f27d4b55fa
	VPOPCNTD Z31, K3, Z7                               // 62927d4b55ff

	VPDPBUSD 7(SI)(DI*2), Z6, K3, Z23                  // 62e24d4b50bc7e07000000
	VPDPBUSD Z8, Z23, K3, Z5                           // 62d2454350e8
	VPDPBUSD Z28, Z23, K3, Z5                          // 6292454350ec
	VPDPBUSD (SI), Z23, K3, Z5                         // 62f24543502e
	VPDPBUSD 7(SI)(DI*2), Z23, K3, Z5                  // 62f2454350ac7e07000000

	VPERMI2W 7(AX)(CX*8), Z16, K3, Z1                  // 62f2fd43758cc807000000
	VPERMI2W Z28, Z12, K3, Z3                          // 62929d4b75dc
	VPERMI2W Z13, Z12, K3, Z3                          // 62d29d4b75dd
	VPERMI2W 99(R15)(R15*8), Z12, K3, Z3               // 62929d4b759cff63000000
	VPERMI2W 7(AX)(CX*8), Z12, K3, Z3                  // 62f29d4b759cc807000000

	VPERMT2B 99(R15), Z9, K3, Z6                       // 62d2354b7db763000000
	VPERMT2B Z20, Z1, K3, Z9                           // 6232754b7dcc
	VPERMT2B Z9, Z1, K3, Z9                            // 6252754b7dc9
	VPERMT2B (CX), Z1, K3, Z9                          // 6272754b7d09
	VPERMT2B 99(R15), Z1, K3, Z9                       // 6252754b7d8f63000000

#define B6 V6 #define B7 V7 #define ACC0 V8 #define ACC1 V9 #define ACCM V10 #define T0 V11 #define T1 V12 #define T2 V13 #define T3 V14 #define POLY V15 #define ZERO V16 #define INC V17 #define CTR V18 #define K0 V19 #define K1 V20 #define K2 V21 #define K3 V22 #define K4 V23 #define K5 V24 #define K6 V25 #define K7 V26 #define K8 V27 #define K9 V28 #define K10 V29 #define K11 V30 #define KLAST V31 #define reduce() \ VEOR ACC0.B16, ACCM.B16, ACCM.B16 \ VEOR ACC1.B16, ACCM.B16, ACCM.B16 \ VEXT $8, ZERO.B16,...