VX     d3, d1, d1  \
	VERLLF $7, a1, a1  \
	VERLLF $7, b1, b1  \
	VERLLF $7, c1, c1  \
	VERLLF $7, d1, d1

#define PERMUTE(mask, v0, v1, v2, v3) \
	VPERM v0, v0, mask, v0 \
	VPERM v1, v1, mask, v1 \
	VPERM v2, v2, mask, v2 \
	VPERM v3, v3, mask, v3

#define ADDV(x, v0, v1, v2, v3) \
	VAF x, v0, v0 \
	VAF x, v1, v1 \
	VAF x, v2, v2 \
	VAF x, v3, v3

	AVCMPGTSWCC
	AVCMPGTSD
	AVCMPGTSDCC
	AVCMPNEZB
	AVCMPNEZBCC
	AVCMPNEB
	AVCMPNEBCC
	AVCMPNEH
	AVCMPNEHCC
	AVCMPNEW
	AVCMPNEWCC
	AVPERM
	AVPERMXOR
	AVPERMR
	AVBPERMQ
	AVBPERMD
	AVSEL
	AVSPLTB
	AVSPLTH
	AVSPLTW
	AVSPLTISB
	AVSPLTISH
	AVSPLTISW
	AVCIPH
	AVCIPHER
	AVCIPHERLAST
	AVNCIPH
	AVNCIPHER
	AVNCIPHERLAST
	AVSBOX
	AVSHASIGMA

	LXVD2X (CPOOL+R0), PH
	LXVD2X (CPOOL+R16), PL

	// VPERM byte selections
	LXVD2X (CPOOL+R18), SEL2
	LXVD2X (CPOOL+R19), SEL1

	LXVD2X (R16)(x_ptr), T1
	LXVD2X (R0)(x_ptr), T0

	// Put in true little endian order
	XXPERMDI T0, T0, $2, T0
	XXPERMDI T1, T1, $2, T1

	// First round
	VPERM   T1, T0, SEL1, RED2    // d1 d0 d1 d0
	VPERM   ZER, RED2, SEL2, RED1 // 0  d1 d0  0

	XXLOR	VS31, VS31, V17
	VADDUWM	V16, V6, V6
	VADDUWM	V17, V7, V7

	CMPU	INP, END
	BLT	loop

	LVX	(TBL)(R_x000), V8
	VPERM	V0, V1, KI, V0
	LVX	(TBL)(R_x010), V9
	VPERM	V4, V5, KI, V4
	VPERM	V0, V2, V8, V0
	VPERM	V4, V6, V8, V4
	VPERM	V0, V3, V9, V0
	VPERM	V4, V7, V9, V4
	STXVD2X	V0, (CTX+R_x000)
	STXVD2X	V4, (CTX+R_x010)

end:

	// Third round
	VPERM T1, T0, SEL1, RED2    // d1 d0 d1 d0
	VPERM ZER, RED2, SEL2, RED1 // 0  d1 d0  0
	VSQ   RED1, RED2, RED2      // Guaranteed not to underflow

	VSLDB $8, T1, T0, T0
	VSLDB $8, T2, T1, T1

	VACCQ  T0, RED1, CAR1
	VAQ    T0, RED1, T0
	VACCCQ T1, RED2, CAR1, CAR2
	VACQ   T1, RED2, CAR1, T1
	VAQ    T2, CAR2, T2

	// Last round
	VPERM T1, T0, SEL1, RED2    // d1 d0 d1 d0

	AVPCOMPRESSB
	AVPCOMPRESSD
	AVPCOMPRESSQ
	AVPCOMPRESSW
	AVPCONFLICTD
	AVPCONFLICTQ
	AVPDPBUSD
	AVPDPBUSDS
	AVPDPWSSD
	AVPDPWSSDS
	AVPERM2F128
	AVPERM2I128
	AVPERMB
	AVPERMD
	AVPERMI2B
	AVPERMI2D
	AVPERMI2PD
	AVPERMI2PS
	AVPERMI2Q
	AVPERMI2W
	AVPERMILPD
	AVPERMILPS
	AVPERMPD
	AVPERMPS
	AVPERMQ
	AVPERMT2B
	AVPERMT2D
	AVPERMT2PD

	VFMADB	V16, V8, V9, V10        // e7a08300948f
	WFMADB	V17, V18, V19, V20      // e74123083f8f
	VFMSDB	V2, V25, V24, V31       // e7f293008b8e
	WFMSDB	V31, V2, V3, V4         // e74f2308348e
	VPERM	V31, V0, V2, V3         // e73f0000248c
	VPDI	$1, V2, V31, V1         // e712f0001284
	VLEG	$1, (R3), V1            // e71030001002
	VLEF	$2, (R0), V31           // e7f000002803
	VLEH	$3, (R12), V16          // e700c0003801

DATA byteswap<>+8(SB)/8, $0x0f0e0d0c0b0a0908
#endif

// Load bytes in big endian order. Address
// alignment does not need checking.
#define VLOADSWAP(base, index, vreg, vsreg) \
	LXVD2X (base)(index), vsreg;  \
	VPERM  vreg, vreg, SWAP, vreg

GLOBL byteswap<>+0(SB), RODATA, $16

TEXT ·Index<ABIInternal>(SB),NOSPLIT|NOFRAME,$0-56
	// R3 = byte array pointer
	// R4 = length
	MOVD R6, R5             // R5 = separator pointer