"SUBE",
	"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",

under512:
	SRDCC $3, R6, R7  // 64 byte chunks?
	XXLXOR VS32, VS32, VS32 // clear VS32 (V0)
	BEQ   lt64gt8

	// Prepare to clear 64 bytes at a time.

zero64setup:
	DCBTST (R3)             // prepare data cache
	MOVD   R7, CTR          // number of 64 byte chunks
	MOVD   $16, R8
	MOVD   $32, R16
	MOVD   $48, R17

zero64:
	STXVD2X VS32, (R3+R0)   // store 16 bytes

	BC	12, 6, checkbytes	// BEQ CR1, checkbytes
	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

	case LXV:
		return op + " " + args[1] + "," + args[0]

	case LXVL, LXVLL:
		return op + " " + args[1] + "," + args[2] + "," + args[0]

	case DCBT, DCBTST, DCBZ, DCBST, ICBI:
		if args[0] == "0" || args[0] == "R0" {
			return op + " (" + args[1] + ")"
		}
		return op + " (" + args[1] + ")(" + args[0] + ")"

	// branch instructions needs additional handling

	LQ:             "lq",
	STQ:            "stq",
	CNTLZD:         "cntlzd",
	CNTLZDCC:       "cntlzd.",
	DCBF:           "dcbf",
	DCBST:          "dcbst",
	DCBT:           "dcbt",
	DCBTST:         "dcbtst",
	DIVD:           "divd",
	DIVDCC:         "divd.",
	DIVDO:          "divdo",
	DIVDOCC:        "divdo.",
	DIVDU:          "divdu",
	DIVDUCC:        "divdu.",
	DIVDUO:         "divduo",

		}
		o1 = AOP_RRR(c.opirr(p.As), uint32(p.To.Reg), uint32(p.From.Reg), 0) | (uint32(c.regoff(p.GetFrom3()))&0x7F)<<11

	case 43: /* data cache instructions: op (Ra+[Rb]), [th|l] */
		/* TH field for dcbt/dcbtst: */
		/* 0 = Block access - program will soon access EA. */
		/* 8-15 = Stream access - sequence of access (data stream). See section 4.3.2 of the ISA for details. */