SLD	$6, LEN
	ADD	INP, LEN, END

	CMP	INP, END
	BEQ	end

	MOVD	$·kcon(SB), TBL_STRT
	MOVD	$0x10, R_x010

#ifdef GOARCH_ppc64le
	MOVWZ	$8, TEMP
	LVSL	(TEMP)(R0), LEMASK
	VSPLTISB	$0x0F, KI
	VXOR	KI, LEMASK, LEMASK
#endif

	LXVW4X	(CTX)(R_x000), V0
	LXVW4X	(CTX)(R_x010), V4

	// unpack the input values into vector registers
	VSLDOI	$4, V0, V0, V1
	VSLDOI	$8, V0, V0, V2

	SLD	$6, LEN
	ADD	INP, LEN, END

	CMP	INP, END
	BEQ	end

	MOVD	$·kcon(SB), TBL_STRT
	MOVD	$0x10, R_x010

#ifdef GOARCH_ppc64le
	MOVWZ	$8, TEMP
	LVSL	(TEMP)(R0), LEMASK
	VSPLTISB	$0x0F, KI
	VXOR	KI, LEMASK, LEMASK
#endif

	LXVW4X	(CTX)(R_x000), V0
	LXVW4X	(CTX)(R_x010), V4

	// unpack the input values into vector registers
	VSLDOI	$4, V0, V0, V1
	VSLDOI	$8, V0, V0, V2

	MOVD	$0x0e0, R_x0e0
	MOVD	$0x0f0, R_x0f0
	MOVD	$0x100, R_x100
	MOVD	$0x110, R_x110


#ifdef GOARCH_ppc64le
	// Generate the mask used with VPERM for LE
	MOVWZ	$8, TEMP
	LVSL	(TEMP)(R0), LEMASK
	VSPLTISB	$0x0F, KI
	VXOR	KI, LEMASK, LEMASK
#endif

	LXVD2X	(CTX)(R_x000), VS32	// v0 = vs32
	LXVD2X	(CTX)(R_x010), VS34	// v2 = vs34
	LXVD2X	(CTX)(R_x020), VS36	// v4 = vs36

	BR	startbarConst
barcstTable:
	MOVD    $·CastBarConst(SB),R3

startbarConst:
	LVX	(R3),const1
	LVX	(R3+off16),const2

	VSLDOI	$8,V0,V0,V1
	VXOR	V0,V1,V0

#ifdef REFLECT
	VSPLTISB $1,V1
	VSL	V0,V1,V0
#endif

	VAND	V0,mask_64bit,V0

#ifndef	REFLECT

	VPMSUMD	V0,const1,V1
	VSLDOI	$8,zeroes,V1,V1
	VPMSUMD	V1,const2,V1
	VXOR	V0,V1,V0
	VSLDOI	$8,V0,zeroes,V0

	VLEIB	$15, $255, V0           // e70000fff040
	VLEIH	$7, $-32768, V15        // e7f080007041
	VLEIF	$2, $-43, V16           // e700ffd52843
	VLEIG	$1, $32767, V31         // e7f07fff1842
	VSLDB	$3, V1, V16, V18        // e72100030a77
	VERIMB	$2, V31, V1, V2         // e72f10020472
	VSEL	V1, V2, V3, V4          // e7412000308d
	VGFMAH	V21, V31, V24, V0       // e705f10087bc