zero512setup:  // setup for dcbz loop
	CMP  R4, $512   // check if at least 512
	BLT  remain
	SRD  $9, R4, R8 // loop count for 512 chunks
	MOVD R8, CTR    // set up counter
	MOVD $128, R9   // index regs for 128 bytes
	MOVD $256, R10
	MOVD $384, R11
	PCALIGN $16
zero512:
	DCBZ (R3+R0)        // clear first chunk
	DCBZ (R3+R9)        // clear second chunk
	DCBZ (R3+R10)       // clear third chunk

	"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",

	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
	case BCLR:

	DCBST (R3)(R4)                  // 7c04186c
	DCBST (R3)(R0)                  // 7c00186c
	DCBST (R3)                      // 7c00186c
	DCBZ (R3)(R4)                   // 7c041fec
	DCBZ (R3)(R0)                   // 7c001fec
	DCBZ (R3)                       // 7c001fec
	DCBT (R3)(R4)                   // 7c041a2c
	DCBT (R3)(R0)                   // 7c001a2c
	DCBT (R3)                       // 7c001a2c

	CRANDC:         "crandc",
	CREQV:          "creqv",
	CRNAND:         "crnand",
	CRNOR:          "crnor",
	CROR:           "cror",
	CRORC:          "crorc",
	CRXOR:          "crxor",
	DCBZ:           "dcbz",
	EQV:            "eqv",
	EQVCC:          "eqv.",
	EXTSH:          "extsh",
	EXTSHCC:        "extsh.",
	FABS:           "fabs",
	FABSCC:         "fabs.",
	FADD:           "fadd",