]
  >;

// This decomposition is only correct inside XLA as it ignores use_locking
// attribute.
// accum = accum * momentum + grad;
// var -= grad * lr + accum * momentum * lr
def DecomposeResourceApplyMomentumOpNesterov :
  Pattern<
    (TF_ResourceApplyMomentumOp:$src_op
       $var_resource, $accum_resource, $lr, $grad, $momentum,

		}
		table = append(table, line)
	}

	// Test that Reset returns to the first line.
	lr.Reset()
	if err := lr.Next(&line); err != nil {
		t.Fatal("lr.Next after Reset failed:", err)
	} else if line != table[0] {
		t.Fatal("lr.Next after Reset returned", line, "instead of", table[0])
	}

	// Check that entries match when seeking backward.
	for i := len(posTable) - 1; i >= 0; i-- {

	p("MOVD LR, R31")
	p("MOVDU R31, -%d(R1)", l.stack) // allocate frame, save PC of interrupted instruction (in LR)

	l.save()
	p("CALL ·asyncPreempt2(SB)")
	l.restore()

	p("MOVD %d(R1), R31", l.stack) // sigctxt.pushCall has pushed LR, R2, R12 (at interrupt) on stack, restore them
	p("MOVD R31, LR")
	p("MOVD %d(R1), R2", l.stack+8)

	// The standard prologue clobbers LR (R14), which is callee-save in
	// the C ABI, so we have to use NOFRAME and save LR ourselves.
	MOVD	LR, R1
	// Also save g, R10, and R11 since they're callee-save in C ABI
	MOVD	R10, R3
	MOVD	g, R4
	MOVD	R11, R5

	MOVD	R2, g
	BL	runtime·save_g(SB)

	MOVD	R5, R11
	MOVD	R4, g
	MOVD	R3, R10
	MOVD	R1, LR
	RET

	asmArchMipsLE   = asmArch{name: "mipsle", bigEndian: false, stack: "R29", lr: true}
	asmArchMips64   = asmArch{name: "mips64", bigEndian: true, stack: "R29", lr: true}
	asmArchMips64LE = asmArch{name: "mips64le", bigEndian: false, stack: "R29", lr: true}
	asmArchPpc64    = asmArch{name: "ppc64", bigEndian: true, stack: "R1", lr: true, retRegs: []string{"R3", "F1"}}

	dst, src, data, hook, cleanup := newSpliceFileTest(t, proto, size)
	defer cleanup()

	// If we have a limit, wrap the reader.
	var (
		r  io.Reader
		lr *io.LimitedReader
	)
	if limit >= 0 {
		lr = &io.LimitedReader{N: limit, R: src}
		r = lr
		if limit < int64(len(data)) {
			data = data[:limit]
		}
	} else {
		r = src
	}

	BR	racecalladdr<>(SB)

// func runtime·RaceWrite(addr uintptr)
// Called from instrumented Go code
TEXT	runtime·racewrite<ABIInternal>(SB), NOSPLIT, $0-8
	MOVD	R3, R4 // addr
	MOVD	LR, R5 // caller has set LR via BL inst
	// void __tsan_write(ThreadState *thr, void *addr, void *pc);
	MOVD	$__tsan_write(SB), R8
	BR	racecalladdr<>(SB)

TEXT    runtime·RaceWrite(SB), NOSPLIT, $0-8
	JMP	runtime·racewrite(SB)

#endif
	// The standard prologue clobbers R31, which is callee-save in
	// the C ABI, so we have to use $-8-0 and save LR ourselves.
	MOVD	LR, R4
	// Also save g and R31, since they're callee-save in C ABI
	MOVD	R31, R5
	MOVD	g, R6

	MOVD	R3, g
	BL	runtime·save_g(SB)

	MOVD	R6, g
	MOVD	R5, R31
	MOVD	R4, LR
	RET

TEXT runtime·abort(SB),NOSPLIT|NOFRAME,$0-0
	MOVW	(R0), R0
	UNDEF

#define	TBR	268

	// Set the two remaining arguments now.
	MOVD	runtime·cgoTraceback(SB), R6
	MOVD	$runtime·sigtramp(SB), R8
	MOVD	_cgo_callers(SB), R12
	MOVD	R12, CTR
	MOVD	R10, LR // restore LR
	JMP	(CTR)

sigtramp:
	MOVD	R10, LR // restore LR
	JMP	runtime·sigtramp(SB)

sigtrampnog:
	// Signal arrived on a non-Go thread. If this is SIGPROF, get a
	// stack trace.
	CMPW	R3, $27 // 27 == SIGPROF
	BNE	sigtramp

	MOVD	R12, R27
#endif
	MOVD	g_m(g), R0
	MOVD	m_p(R0), R0
	MOVD	p_raceprocctx(R0), R0
	MOVD	R0, (R1)
	MOVD	R13, g
	JMP	(LR)
rest:
	// Save callee-saved registers (Go code won't respect that).
	// 8(RSP) and 16(RSP) are for args passed through racecallback
	SUB	$176, RSP
	MOVD	LR, 0(RSP)

	SAVE_R19_TO_R28(8*3)
	SAVE_F8_TO_F15(8*13)
	MOVD	R29, (8*21)(RSP)
	// Set g = g0.