}
	i := new(big.Int).Lsh(h, 1)
	i.Mul(i, i)
	j := new(big.Int).Mul(h, i)

	s1 := new(big.Int).Mul(y1, z2)
	s1.Mul(s1, z2z2)
	s1.Mod(s1, curve.P)
	s2 := new(big.Int).Mul(y2, z1)
	s2.Mul(s2, z1z1)
	s2.Mod(s2, curve.P)
	r := new(big.Int).Sub(s2, s1)
	if r.Sign() == -1 {
		r.Add(r, curve.P)
	}
	yEqual := r.Sign() == 0
	if xEqual && yEqual {
		return curve.doubleJacobian(x1, y1, z1)
	}

  func.return %2 : tensor<4xf32>

// CHECK-NEXT: %[[CST:.*]] = arith.constant dense<1.000000e+00> : tensor<4xf32>
// CHECK-NEXT: %[[MUL:.*]] = tfl.mul %arg0, %[[CST]] {fused_activation_function = "NONE"} : tensor<4xf32>
// CHECK-NEXT: %[[CUSTOM_1:.*]] = "tfl.custom_tf"(%[[MUL]], %[[CST]]) ({
// CHECK-NEXT: ^bb0(%arg1: tensor<4xf32>, %arg2: tensor<4xf32>):

// CHECK-LABEL: broadcast_to_bf16
// CHECK:  [[CST:%.*]] = arith.constant dense<1.000000e+00> : tensor<3x3xbf16>
// CHECK:  [[MUL:%.*]] = tfl.mul(%arg0, [[CST]]) <{fused_activation_function = "NONE"}> : (tensor<3xbf16>, tensor<3x3xbf16>) -> tensor<3x3xbf16>
// CHECK:  return [[MUL]] : tensor<3x3xbf16>

  %1 = "tfl.squared_difference"(%arg0, %0) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("squared_difference")
  %2 = "tfl.mul"(%arg0, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("mul")
  %3 = "tfl.div"(%2, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("div")

  %2 = "tfl.mul"(%0, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("mul")
  func.return %2 : tensor<4xf32>

// CHECK-LABEL: main
// CHECK-NOT: tfl.squared_difference
// CHECK: tfl.mul %[[CONST:.*]], %arg0

  rewriter.clearInsertionPoint();
  return new_func;
}

// Outlines non-approximate GELU into a stablehlo composite.
//
//    -> mul 1/sqrt(2) -> erf -> add 1 ->
// in                                    mul
//    ---------> mul 0.5 --------------->
//
// This pattern assumes all binary ewise ops with one constant argument
// have that constant argument as the second operand. It works by

TEXT time·now(SB),NOSPLIT,$0-24
	MOVD	$_INTERRUPT_TIME, R3
	MOVD	time_lo(R3), R0
	MOVD	$100, R1
	MUL	R1, R0
	MOVD	R0, mono+16(FP)

	MOVD	$_SYSTEM_TIME, R3
	MOVD	time_lo(R3), R0
	// convert to Unix epoch (but still 100ns units)
	#define delta 116444736000000000
	SUB	$delta, R0
	// Convert to nSec
	MOVD	$100, R1
	MUL	R1, R0

	// Code stolen from compiler output for:
	//
	//	var x uint64

  %cst = arith.constant dense<[5, 7]> : tensor<2xi32>
  %0 = "tf.BroadcastTo"(%arg1, %cst) : (tensor<7xf32>, tensor<2xi32>) -> tensor<5x7xf32>
  %1 = "tf.Mul"(%arg0, %0) : (tensor<5x7xf32>, tensor<5x7xf32>) -> tensor<5x7xf32>
  func.return %1 : tensor<5x7xf32>
  // CHECK: %[[V0:.*]] = "tf.Mul"(%arg0, %arg1) : (tensor<5x7xf32>, tensor<7xf32>) -> tensor<5x7xf32>
  // CHECK: %[[V0]] : tensor<5x7xf32>
}

// CHECK-LABEL: @broadcast_mul1

  // CHECK-DAG: %[[SHIFT:.*]] = "tf.AddV2"(%[[CAST]], %[[HALF_RANGE]])
  // CHECK-DAG: %[[DRANGE:.*]] = "tf.Sub"(%arg2, %arg1)
  // CHECK-DAG: %[[SCALE:.*]] = "tf.Div"(%[[DRANGE]], %[[C255:.*]])
  // CHECK-DAG: %[[SS:.*]] = "tf.Mul"(%[[SHIFT]], %[[SCALE]])
  // CHECK-DAG: %[[RESULT:.*]] = "tf.AddV2"(%[[SS]], %arg1)
  %0 = "tf.Dequantize"(%arg0, %min_range, %max_range) : (tensor<2x3x!tf_type.qint8>, tensor<f32>, tensor<f32>) -> tensor<2x3xf32>

  [(HasRankOf<1> $add_rhs_value),
   (HasEqualElementSize<[-1], [0]> $conv_out, $add_rhs)], [], (addBenefit -1)>;

// Convert conv+sub+mul pattern to conv+mul+add.
// (conv - sub) * mul -> conv * mul + (-sub) * mul
//
// This is needed to support Conv+BatchNorm pattern from Jax models converted
// using jax2tf w/o native serialization. Note that Jax2tf patterns always