// medium move uses a duff device
(Move [s] {t} dst src mem)
	&& s%8 == 0 && s >= 24 && s <= 8*128 && t.Alignment()%8 == 0
	&& !config.noDuffDevice && logLargeCopy(v, s)  =>
	(DUFFCOPY [16 * (128 - s/8)] dst src mem)
// 16 and 128 are magic constants.  16 is the number of bytes to encode:
//	MOVV	(R1), R23
//	ADDV	$8, R1
//	MOVV	R23, (R2)
//	ADDV	$8, R2

// 16 and 128 are magic constants, see runtime/mkduff.go
(Move [s] {t} dst src mem)
	&& s%8 == 0 && s <= 8*128 && t.Alignment()%8 == 0
	&& !config.noDuffDevice && logLargeCopy(v, s) =>
	(DUFFCOPY [16 * (128 - s/8)] dst src mem)

// Generic move uses a loop
(Move [s] {t} dst src mem) && (s <= 16 || logLargeCopy(v, s)) =>
	(LoweredMove [t.Alignment()]
		dst
		src

// medium move uses a duff device
(Move [s] {t} dst src mem)
	&& s%8 == 0 && s >= 24 && s <= 8*128 && t.Alignment()%8 == 0
	&& !config.noDuffDevice && logLargeCopy(v, s)  =>
	(DUFFCOPY [16 * (128 - s/8)] dst src mem)
// 16 and 128 are magic constants.  16 is the number of bytes to encode:
//	MOVV	(R1), R23
//	ADDV	$8, R1
//	MOVV	R23, (R2)
//	ADDV	$8, R2

				}

				if yt.zcase == Zcallduff && ctxt.Arch.Family == sys.AMD64 {
					// Maintain BP around call, since duffcopy/duffzero can't do it
					// (the call jumps into the middle of the function).
					// This makes it possible to see call sites for duffcopy/duffzero in
					// BP-based profiling tools like Linux perf (which is the
					// whole point of maintaining frame pointers in Go).

		// local call offsets for externally generated objects are accounted for when converting into golang relocs.
		if !hasPCrel && !ldr.AttrExternal(rs) && ldr.AttrShared(rs) && tgtName != "runtime.duffzero" && tgtName != "runtime.duffcopy" {
			// Furthermore, only apply the offset if the target looks like the start of a function call.
			if r.Add() == 0 && ldr.SymType(rs) == sym.STEXT {
				t += 8
			}
		}

		if t&3 != 0 {

// 8 and 128 are magic constants, see runtime/mkduff.go
(Move [s] {t} dst src mem)
	&& s%4 == 0 && s > 4 && s <= 512
	&& t.Alignment()%4 == 0 && !config.noDuffDevice && logLargeCopy(v, s) =>
	(DUFFCOPY [8 * (128 - s/4)] dst src mem)

// Large move uses a loop
(Move [s] {t} dst src mem)
	&& ((s > 512 || config.noDuffDevice) || t.Alignment()%4 != 0) && logLargeCopy(v, s) =>
	(LoweredMove [t.Alignment()]
		dst

			(MOVQstore dst (MOVQload src mem) mem)))

// Medium copying uses a duff device.
(Move [s] dst src mem)
	&& s > 64 && s <= 16*64 && s%16 == 0
	&& !config.noDuffDevice && logLargeCopy(v, s) =>
	(DUFFCOPY [s] dst src mem)

// Large copying uses REP MOVSQ.
(Move [s] dst src mem) && (s > 16*64 || config.noDuffDevice) && s%8 == 0 && logLargeCopy(v, s) =>
	(REPMOVSQ dst src (MOVQconst [s/8]) mem)

		(Move [s-s%16] dst src mem))

// medium move uses a duff device
(Move [s] dst src mem)
	&& s > 64 && s <= 16*64 && s%16 == 0
	&& !config.noDuffDevice && logLargeCopy(v, s) =>
	(DUFFCOPY [8 * (64 - s/16)] dst src mem)
// 8 is the number of bytes to encode:
//
// LDP.P   16(R16), (R26, R27)
// STP.P   (R26, R27), 16(R17)
//
// 64 is number of these blocks. See runtime/duff_arm64.s:duffcopy

	ir.Syms.DeferprocStack = typecheck.LookupRuntimeFunc("deferprocStack")
	ir.Syms.Deferreturn = typecheck.LookupRuntimeFunc("deferreturn")
	ir.Syms.Duffcopy = typecheck.LookupRuntimeFunc("duffcopy")
	ir.Syms.Duffzero = typecheck.LookupRuntimeFunc("duffzero")
	ir.Syms.GCWriteBarrier[0] = typecheck.LookupRuntimeFunc("gcWriteBarrier1")

		reg: regInfo{
			inputs: []inputInfo{
				{0, 239}, // AX CX DX BX BP SI DI
			},
			clobbers: 65519, // AX CX DX BX BP SI DI X0 X1 X2 X3 X4 X5 X6 X7
		},
	},
	{
		name:           "DUFFCOPY",
		auxType:        auxInt64,
		argLen:         3,
		clobberFlags:   true,
		faultOnNilArg0: true,
		faultOnNilArg1: true,
		reg: regInfo{
			inputs: []inputInfo{
				{0, 128}, // DI