VSLDOI  $1, V1, V2, V3     // V3=(V1:V2)<<1
	VSLDOI  $2, V1, V2, V4     // V4=(V1:V2)<<2
	VAND    V1, SEPMASK, V8    // Mask out sep size 0th index
	VAND    V3, SEPMASK, V9    // Mask out sep size 1st index
	VAND    V4, SEPMASK, V11   // Mask out sep size 2nd index
	VAND    V5, SEPMASK, V12   // Mask out sep size 3rd index
#endif
	VCMPEQUBCC      V0, V8, V8 // compare masked string

	VAND	V5.B16, V3.B16, V3.B16
	VAND	V5.B16, V4.B16, V4.B16
	VADDP	V4.B16, V3.B16, V6.B16
	VADDP	V6.B16, V6.B16, V6.B16
	VMOV	V6.D[0], R6
	// Only do the clear for the last possible block with less than 32 bytes
	// Condition flags come from SUBS in the loop
	BHS	tail

masklast:
	// Clear the irrelevant upper bits
	ADD	R9, R10, R4
	AND	$0x1f, R4, R4
	SUB	$0x20, R4, R4

	VMOV	R1, V0.B16
	// Clear the low 64-bit element of V7 and V8
	VEOR	V7.B8, V7.B8, V7.B8
	VEOR	V8.B8, V8.B8, V8.B8
	PCALIGN $16
	// Count the target byte in 32-byte chunk
chunk_loop:
	VLD1.P	(R0), [V1.B16, V2.B16]
	CMP	R0, R3
	VCMEQ	V0.B16, V1.B16, V3.B16
	VCMEQ	V0.B16, V2.B16, V4.B16
	// Clear the higher 7 bits
	VAND	V5.B16, V3.B16, V3.B16
	VAND	V5.B16, V4.B16, V4.B16
	// Count lanes match the requested byte

	VCMEQ	V0.D2, V4.D2, V8.D2
	VCMEQ	V1.D2, V5.D2, V9.D2
	VCMEQ	V2.D2, V6.D2, V10.D2
	VCMEQ	V3.D2, V7.D2, V11.D2
	VAND	V8.B16, V9.B16, V8.B16
	VAND	V8.B16, V10.B16, V8.B16
	VAND	V8.B16, V11.B16, V8.B16
	CMP	R0, R6
	VMOV	V8.D[0], R4
	VMOV	V8.D[1], R5
	CBZ	R4, not_equal
	CBZ	R5, not_equal
	BNE	chunk64_loop
	AND	$0x3f, R2, R2
	CBZ	R2, equal
chunk16:
	// work with 16-byte chunks
	BIC	$0xf, R2, R3
	CBZ	R3, tail

	"UDIV",
	"UDIVW",
	"UMADDL",
	"UMNEGL",
	"UMSUBL",
	"UMULH",
	"UMULL",
	"UREM",
	"UREMW",
	"UXTB",
	"UXTBW",
	"UXTH",
	"UXTHW",
	"UXTW",
	"VADD",
	"VADDP",
	"VADDV",
	"VAND",
	"VBCAX",
	"VBIF",
	"VBIT",
	"VBSL",
	"VCMEQ",
	"VCMTST",
	"VCNT",
	"VDUP",
	"VEOR",
	"VEOR3",
	"VEXT",
	"VFMLA",
	"VFMLS",
	"VLD1",
	"VLD1R",

	"MADDHD",
	"MADDHDU",
	"MADDLD",
	"LVEBX",
	"LVEHX",
	"LVEWX",
	"LVX",
	"LVXL",
	"LVSL",
	"LVSR",
	"STVEBX",
	"STVEHX",
	"STVEWX",
	"STVX",
	"STVXL",
	"VAND",
	"VANDC",
	"VNAND",
	"VOR",
	"VORC",
	"VNOR",
	"VXOR",
	"VEQV",
	"VADDUM",
	"VADDUBM",
	"VADDUHM",
	"VADDUWM",
	"VADDUDM",
	"VADDUQM",
	"VADDCU",
	"VADDCUQ",
	"VADDCUW",

	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

#else

	VAND	V0,mask_32bit,V1
	VPMSUMD	V1,const1,V1
	VAND	V1,mask_32bit,V1
	VPMSUMD	V1,const2,V1
	VXOR	V0,V1,V0
	VSLDOI  $4,V0,zeroes,V0

#endif

		VMOV	CTR.B16, B0.B16
		VADD	B0.S4, INC.S4, B1.S4
		VREV32	B0.B16, B0.B16
		VADD	B1.S4, INC.S4, B2.S4
		VREV32	B1.B16, B1.B16
		VADD	B2.S4, INC.S4, B3.S4
		VREV32	B2.B16, B2.B16
		VADD	B3.S4, INC.S4, B4.S4
		VREV32	B3.B16, B3.B16
		VADD	B4.S4, INC.S4, B5.S4
		VREV32	B4.B16, B4.B16
		VADD	B5.S4, INC.S4, B6.S4
		VREV32	B5.B16, B5.B16
		VADD	B6.S4, INC.S4, B7.S4
		VREV32	B6.B16, B6.B16

// The uncommented cases means they can be handled by assembler
// and they are consistent with disassembler decoding.
// TODO means they cannot be handled by current assembler.

#include "../../../../../runtime/textflag.h"

TEXT asmtest(SB),DUPOK|NOSPLIT,$-8

	AND $(1<<63), R1                           // AND $-9223372036854775808, R1     // 21004192
	ADCW ZR, R8, R10                           // 0a011f1a

	VSL      H, T0, H            // H<<=1
	VSRAB    T1, T2, T1          // broadcast carry bit
	VAND     T1, XC2, T1
	VXOR     H, T1, IN           // twisted H

	VSLDOI $8, IN, IN, H      // twist even more ...
	VSLDOI $8, ZERO, XC2, XC2 // 0xc2.0
	VSLDOI $8, ZERO, H, HL    // ... and split
	VSLDOI $8, H, ZERO, HH

	STXVD2X VXC2, (XIP+R0) // save pre-computed table
	STXVD2X VHL, (XIP+R8)