SHA512ROUND(V2, V3, V1, V4, V0, V26, V30, V18, V19, V17, V14, V15)
	SHA512ROUND(V4, V2, V0, V1, V3, V27, V31, V19, V12, V18, V15, V16)
	SHA512ROUND(V1, V4, V3, V0, V2, V28, V24, V12, V13, V19, V16, V17)

	SHA512ROUND(V0, V1, V2, V3, V4, V29, V25, V13, V14, V12, V17, V18)
	SHA512ROUND(V3, V0, V4, V2, V1, V30, V26, V14, V15, V13, V18, V19)
	SHA512ROUND(V2, V3, V1, V4, V0, V31, V27, V15, V16, V14, V19, V12)

	HASHUPDATE                                          // H8

	VADD	V18.S4, V6.S4, V9.S4                        // V18(W8+K8...W11+K11)
	SHA256SU0	V7.S4, V6.S4                        // V6: (su0(W9)+W8,...,su0(W12)+W11)
	SHA256SU1	V4.S4, V7.S4, V5.S4                 // V5: W20-W23
	HASHUPDATE                                          // H12

	VADD	V19.S4, V7.S4, V9.S4                        // V18(W12+K12...W15+K15)

    T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
    T16 v16, T17 v17, T18 v18) {
  return internal::ValueArray18<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
      T12, T13, T14, T15, T16, T17, T18>(v1, v2, v3, v4, v5, v6, v7, v8, v9,
      v10, v11, v12, v13, v14, v15, v16, v17, v18);
}

template <typename T1, typename T2, typename T3, typename T4, typename T5,

	LXVD2X	(INP)(R_x040), VS50	// load v18 (=vs50) in advance
	VSLDOI	$8, V16, V16, V17
	SHA512ROUND0(V7, V0, V1, V2, V3, V4, V5, V6, V17, R_x0b0)
	VPERMLE(V18,V18,LEMASK,V18)
	SHA512ROUND0(V6, V7, V0, V1, V2, V3, V4, V5, V18, R_x0c0)
	LXVD2X	(INP)(R_x050), VS52	// load v20 (=vs52) in advance
	VSLDOI	$8, V18, V18, V19
	SHA512ROUND0(V5, V6, V7, V0, V1, V2, V3, V4, V19, R_x0d0)

	WFMADB	V4, V16, V18, V16
	VLEG	$0, 64(R9), V18
	WFMADB	V1, V7, V5, V7
	WFMADB	V4, V18, V22, V18
	WFMDB	V1, V1, V5
	WFMADB	V1, V16, V18, V16
	VLEG	$0, 56(R9), V18
	WFMADB	V5, V6, V7, V6
	VLEG	$0, 48(R9), V22
	FMOVD	40(R9), F7
	WFMADB	V4, V7, V18, V7
	VLEG	$0, 32(R9), V18
	WFMADB	V5, V6, V16, V6
	WFMADB	V4, V18, V22, V18
	VLEG	$0, 24(R9), V16
	WFMADB	V1, V7, V18, V7
	VLEG	$0, 16(R9), V18

	SHA1SU0	V4.S4, V7.S4, V6.S4
	HASHUPDATEMAJ
	SHA1SU1	V5.S4, V6.S4

	VADD	V18.S4, V7.S4, V16.S4
	SHA1SU0	V5.S4, V4.S4, V7.S4
	HASHUPDATEPARITY
	SHA1SU1	V6.S4, V7.S4

	VADD	V18.S4, V4.S4, V16.S4
	HASHUPDATEPARITY

	VADD	V18.S4, V5.S4, V16.S4
	HASHUPDATEPARITY

	VADD	V18.S4, V6.S4, V16.S4
	HASHUPDATEPARITY

	VADD	V18.S4, V7.S4, V16.S4
	HASHUPDATEPARITY

	VEOR	V11.B16, V6.B16, V17.B16
	VEOR	V8.B16, V7.B16, V18.B16
	VEOR	V9.B16, V4.B16, V19.B16
	VSHL	$7, V16.S4, V5.S4
	VSHL	$7, V17.S4, V6.S4
	VSHL	$7, V18.S4, V7.S4
	VSHL	$7, V19.S4, V4.S4
	VSRI	$25, V16.S4, V5.S4
	VSRI	$25, V17.S4, V6.S4
	VSRI	$25, V18.S4, V7.S4
	VSRI	$25, V19.S4, V4.S4

	SUB	$1, R21
	CBNZ	R21, chacha

	// VLD4R (R10), [V16.S4, V17.S4, V18.S4, V19.S4]
	WORD	$0x4D60E950

	LXVW4X (R8)(CONSTBASE), VS33
	LXVW4X (R9)(CONSTBASE), VS34
	LXVW4X (R10)(CONSTBASE), VS35

	// splat values from V17, V18 into V4-V11
	VSPLTW $0, V17, V4
	VSPLTW $1, V17, V5
	VSPLTW $2, V17, V6
	VSPLTW $3, V17, V7
	VSPLTW $0, V18, V8
	VSPLTW $1, V18, V9
	VSPLTW $2, V18, V10
	VSPLTW $3, V18, V11

	// VOR
	VOR V26, V26, V12

	// splat values from V19 -> V13, V14, V15
	VSPLTW $1, V19, V13

	XXLOR key2, key2, V26; \
	VCIPHER vin, V25, vin; \
	VCIPHER vin, V26, vin

// Encrypt 4 values in V15 - V18
// with the specified key from
// VS1 - VS9.
#define VCIPHER_4X1_KEY(key) \
	XXLOR key, key, V23; \
	VCIPHER V15, V23, V15; \
	VCIPHER V16, V23, V16; \
	VCIPHER V17, V23, V17; \
	VCIPHER V18, V23, V18

// Encrypt 8 values in V15 - V22
// with the specified key,
// assuming it is a VSreg

    T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
    T16 v16, T17 v17, T18 v18) {
  return internal::ValueArray18<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
      T12, T13, T14, T15, T16, T17, T18>(v1, v2, v3, v4, v5, v6, v7, v8, v9,
      v10, v11, v12, v13, v14, v15, v16, v17, v18);
}

template <typename T1, typename T2, typename T3, typename T4, typename T5,