SHA512ROUND(V3, V0, V4, V2, V1, V25, V29, V17, V18, V16, V13, V14)
	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)

	LVX	(R4+off32),V16
	LVX	(R3+off32),V17
	VPMSUMW	V16,V17,V16
	VXOR	V0,V16,V0
	BC	18,0,next1

	LVX	(R4+off48),V16
	LVX	(R3+off48),V17
	VPMSUMW	V16,V17,V16
	VXOR	V0,V16,V0
	BC	18,0,next1

	LVX	(R4+off64),V16
	LVX	(R3+off64),V17
	VPMSUMW	V16,V17,V16
	VXOR	V0,V16,V0
	BC	18,0,next1

	LVX	(R4+off80),V16
	LVX	(R3+off80),V17
	VPMSUMW	V16,V17,V16
	VXOR	V0,V16,V0
	BC	18,0,next1

    T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
    T16 v16, T17 v17) {
  return internal::ValueArray17<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
      T12, T13, T14, T15, T16, T17>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10,
      v11, v12, v13, v14, v15, v16, v17);
}

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

	SHA1SU1	V4.S4, V5.S4
	VADD	V17.S4, V6.S4, V16.S4
	SHA1SU0	V4.S4, V7.S4, V6.S4
	HASHUPDATECHOOSE
	SHA1SU1	V5.S4, V6.S4

	VADD	V17.S4, V7.S4, V16.S4
	SHA1SU0	V5.S4, V4.S4, V7.S4
	HASHUPDATECHOOSE
	SHA1SU1	V6.S4, V7.S4

	VADD	V17.S4, V4.S4, V16.S4
	SHA1SU0	V6.S4, V5.S4, V4.S4
	HASHUPDATECHOOSE
	SHA1SU1	V7.S4, V4.S4

	VDUP	V19.S[2], V17.S4
	VADD	V18.S4, V5.S4, V16.S4
	SHA1SU0	V7.S4, V6.S4, V5.S4
	HASHUPDATEPARITY

	SHA512ROUND1(V1, V2, V3, V4, V5, V6, V7, V0, V15, V16, V17, V9, V14, R_x070)
	SHA512ROUND1(V0, V1, V2, V3, V4, V5, V6, V7, V16, V17, V18, V10, V15, R_x080)
	SHA512ROUND1(V7, V0, V1, V2, V3, V4, V5, V6, V17, V18, V19, V11, V16, R_x090)
	SHA512ROUND1(V6, V7, V0, V1, V2, V3, V4, V5, V18, V19, V20, V12, V17, R_x0a0)
	SHA512ROUND1(V5, V6, V7, V0, V1, V2, V3, V4, V19, V20, V21, V13, V18, R_x0b0)

	LXVW4X (R0)(CONSTBASE), VS32
	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

	SHA256ROUND0(V1, V2, V3, V4, V5, V6, V7, V0, V15, R_x090)
	VPERMLE(V16, V16, LEMASK, V16)
	SHA256ROUND0(V0, V1, V2, V3, V4, V5, V6, V7, V16, R_x0a0)
	VSLDOI	$4, V16, V16, V17
	SHA256ROUND0(V7, V0, V1, V2, V3, V4, V5, V6, V17, R_x0b0)
	VSLDOI	$4, V17, V17, V18
	SHA256ROUND0(V6, V7, V0, V1, V2, V3, V4, V5, V18, R_x0c0)
	VSLDOI	$4, V18, V18, V19
	LXVD2X	(INP)(R_x030), V20 // load v20 in advance

	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
#define VCIPHER_8X1_KEY(key) \
	XXLOR key, key, V23; \
	VCIPHER V15, V23, V15; \
	VCIPHER V16, V23, V16; \
	VCIPHER V17, V23, V17; \
	VCIPHER V18, V23, V18; \
	VCIPHER V19, V23, V19; \
	VCIPHER V20, V23, V20; \

      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16,
      T17 v17) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
      v15_(v15), v16_(v16), v17_(v17) {}

  template <typename T>
  operator ParamGenerator<T>() const {

	MOVD	p_len+32(FP), R3                           // message length
	VLD1	(R0), [V0.S4, V1.S4]                       // load h(a,b,c,d,e,f,g,h)
	VLD1.P	64(R2), [V16.S4, V17.S4, V18.S4, V19.S4]
	VLD1.P	64(R2), [V20.S4, V21.S4, V22.S4, V23.S4]
	VLD1.P	64(R2), [V24.S4, V25.S4, V26.S4, V27.S4]
	VLD1	(R2), [V28.S4, V29.S4, V30.S4, V31.S4]     //load 64*4bytes K constant(K0-K63)

blockloop: