"SUBECC",
	"SUBEV",
	"SUBEVCC",
	"SUBZE",
	"SUBZECC",
	"SUBZEVCC",
	"SUBZEV",
	"SYNC",
	"XOR",
	"XORCC",
	"XORIS",
	"DCBF",
	"DCBI",
	"DCBST",
	"DCBT",
	"DCBTST",
	"DCBZ",
	"EIEIO",
	"ICBI",
	"ISYNC",
	"PTESYNC",
	"TLBIE",
	"TLBIEL",
	"TLBSYNC",
	"TW",
	"SYSCALL",
	"WORD",
	"RFCI",
	"FCPSGN",
	"FCPSGNCC",
	"FRES",
	"FRESCC",
	"FRIM",

// license that can be found in the LICENSE file.

#include "textflag.h"

// func syncIcache(p uintptr)
TEXT main·syncIcache(SB), NOSPLIT|NOFRAME, $0-0
	SYNC
	MOVD (R3), R3
	ICBI (R3)
	ISYNC

// license that can be found in the LICENSE file.

#include "textflag.h"

// func syncIcache(p uintptr)
TEXT main·syncIcache(SB), NOSPLIT|NOFRAME, $0-0
	SYNC
	MOVD (R3), R3
	ICBI (R3)
	ISYNC

TEXT cmpbody<>(SB),NOSPLIT|NOFRAME,$0-0
start:
	CMP	R9,$16,CR0
	CMP	R9,$32,CR1
	CMP	R9,$64,CR2
	MOVD	$16,R10
	BLT	cmp8
	BLT	CR1,cmp16
	BLT	CR2,cmp32

cmp64:	// >= 64B
	DCBT	(R5)		// optimize for size>=64
	DCBT	(R6)		// cache hint

	SRD	$6,R9,R14	// There is at least one iteration.
	MOVD	R14,CTR
	ANDCC   $63,R9,R9
	CMP	R9,$16,CR1	// Do setup for tail check early on.
	CMP	R9,$32,CR2
	CMP	R9,$48,CR3

	SRDCC	$3, DWORDS, OCTWORDS	// 64 byte chunks?
	MOVD	$16, IDX16
	BEQ	lt64gt8			// < 64 bytes

	// Prepare for moves of 64 bytes at a time.

forward64setup:
	DCBTST	(TGT)			// prepare data cache
	DCBT	(SRC)
	MOVD	OCTWORDS, CTR		// Number of 64 byte chunks
	MOVD	$32, IDX32
	MOVD	$48, IDX48
	PCALIGN	$16

forward64:
	LXVD2X	(R0)(SRC), VS32		// load 64 bytes
	LXVD2X	(IDX16)(SRC), VS33
	LXVD2X	(IDX32)(SRC), VS34