lines(
			"#define LOAD(off, reg) \\",
			"\tMOVBLZX	(off*4)(R12),	reg \\",
			"\tADDB	reg,		DX",
			"",
			"LOAD(8, AX)",
		),
		"\n.\n.MOVBLZX.(.8.*.4.).(.R12.).,.AX.\n.ADDB.AX.,.DX.\n",
	},
	{
		"nested multiline macro",
		lines(
			"#define KEYROUND(xmm, load, off, r1, r2, index) \\",
			"\tMOVBLZX	(BP)(DX*4),	R8 \\",
			"\tload((off+1), r2) \\",
			"\tMOVB	R8,		(off*4)(R12) \\",

		{"VADDPD.SAE.SAE X0, X1, X2", `duplicate suffix "SAE"`},
		{"VADDPD.RZ_SAE.SAE X0, X1, X2", `bad suffix combination`},

		// BSWAP on 16-bit registers is undefined. See #29167,
		{"BSWAPW DX", `unrecognized instruction`},
		{"BSWAPW R11", `unrecognized instruction`},
	})
}

func testBadInstParser(t *testing.T, goarch string, tests []badInstTest) {
	for i, test := range tests {

// i.e., d(y_1 + y_2 + ...)/dx_1, d(y_1 + y_2 + ...)/dx_2...
//
// `dx` are used as initial gradients (which represent the symbolic partial
// derivatives of some loss function `L` w.r.t. `y`).
// `dx` must be nullptr or have size `ny`.
// If `dx` is nullptr, the implementation will use dx of `OnesLike` for all
// shapes in `y`.

	{"(AX)(CX*8)", "(AX)(CX*8)"},
	{"(BP)(CX*4)", "(BP)(CX*4)"},
	{"(BP)(DX*4)", "(BP)(DX*4)"},
	{"(BP)(R8*4)", "(BP)(R8*4)"},
	{"(BX)", "(BX)"},
	{"(DI)", "(DI)"},
	{"(DI)(BX*1)", "(DI)(BX*1)"},
	{"(DX)", "(DX)"},
	{"(R9)", "(R9)"},
	{"(R9)(BX*8)", "(R9)(BX*8)"},
	{"(SI)", "(SI)"},
	{"(SI)(BX*1)", "(SI)(BX*1)"},
	{"(SI)(DX*1)", "(SI)(DX*1)"},
	{"(SP)", "(SP)"},
	{"(SP)(AX*4)", "(SP)(AX*4)"},

     *
     *    dX = U / Y
     *    dY = -U*X / Y^2 = (X/Y) * -U / Y = -U*Z / Y
     *
     */

    AbstractTensorHandle* upstream_grad = grad_outputs[0];
    AbstractTensorHandle* Y = forward_inputs_[1];
    AbstractTensorHandle* Z = forward_outputs_[0];

    // Calculate dX =  U / Y
    std::string name = "Div_Grad_X";
    TF_RETURN_IF_ERROR(

	VPGATHERDQ X2, 664(BX*1), X1    // ERROR "invalid instruction"
	VPGATHERDQ Y2, (BP)(AX*2), Y1   // ERROR "invalid instruction"
	VPGATHERDQ Y5, 664(DX*8), Y6    // ERROR "invalid instruction"
	VPGATHERDQ Y5, (DX), Y0         // ERROR "invalid instruction"
	// VM/X rejects Y index register.
	VPGATHERDQ Y5, 664(Y14*8), Y6   // ERROR "invalid instruction"

				}
			}
			if tok == scanner.EOF {
				p.errorf("unexpected EOF")
				return "", "", nil, false
			}
			// Split operands on comma. Also, the old syntax on x86 for a "register pair"
			// was AX:DX, for which the new syntax is DX, AX. Note the reordering.
			if tok == '\n' || tok == ';' || (nesting == 0 && (tok == ',' || tok == ':')) {
				if tok == ':' {
					// Remember this location so we can swap the operands below.

  op_tape_.erase(op_it);
}

// Terminology:
//
//  - op: a possibly composite operation, which has an entry in the tape
//  - target: dy in dx/dy
//  - source: dx in dx/dy
//  - tensor: one of the many inputs or outputs of an operation
//
// Below here we do the gradient algorithm. It works as follows:
//

        NewInternalScope(&g->graph, &status->status, &g->refiner)
            .NewSubScope(child_scope_name);

    if (dx != nullptr) {
      std::vector<tensorflow::Output> dx_arg = OutputsFromTFOutputs(dx, ny);
      status->status =
          AddSymbolicGradients(scope, y_arg, x_arg, dx_arg, &dy_arg);
    } else {
      status->status = AddSymbolicGradients(scope, y_arg, x_arg, &dy_arg);
    }

	for i := start; i <= number; i++ {
		if paddinglen == 0 {
			seq = append(seq, fmt.Sprintf("%x", i))
		} else {
			seq = append(seq, fmt.Sprintf(fmt.Sprintf("%%0%dx", paddinglen), i))
		}
	}
	return seq
}

func getSequences(start int, number int, paddinglen int) (seq []string) {
	for i := start; i <= number; i++ {
		if paddinglen == 0 {