// X=Z1; Y=Z1; MUL; T-   // T1 = Z1*Z1
	// X-  ; Y=T ; MUL; R=T  // R  = Z1*T1
	// X=X2; Y-  ; MUL; H=T  // H  = X2*T1
	// X=Z2; Y=Z2; MUL; T-   // T2 = Z2*Z2
	// X-  ; Y=T ; MUL; S1=T // S1 = Z2*T2
	// X=X1; Y-  ; MUL; U1=T // U1 = X1*T2
	// SUB(H<H-T)            // H  = H-U1
	// X=Z1; Y=Z2; MUL; T-   // Z3 = Z1*Z2

 	// X=Z1; Y=Z1; MUL; T-   // T1 = Z1*Z1
	// X-  ; Y=T ; MUL; R=T  // R  = Z1*T1
	// X=X2; Y-  ; MUL; H=T  // H  = X2*T1
	// X=Z2; Y=Z2; MUL; T-   // T2 = Z2*Z2
	// X-  ; Y=T ; MUL; S1=T // S1 = Z2*T2
	// X=X1; Y-  ; MUL; U1=T // U1 = X1*T2
	// SUB(H<H-T)            // H  = H-U1
	// X=Z1; Y=Z2; MUL; T-   // Z3 = Z1*Z2

    long mul = K3;
    int topBit = 0x7;

    int lengthAligned = length & ~topBit;
    int lengthRemainder = length & topBit;
    long hash = seed ^ (length * mul);

    for (int i = 0; i < lengthAligned; i += 8) {
      long loaded = load64(bytes, offset + i);
      long data = shiftMix(loaded * mul) * mul;
      hash ^= data;
      hash *= mul;
    }

func.func @eliminate_passthrough_args_cluster_op_i32(%arg0 : tensor<i32>, %arg1 : tensor<i32>) -> (tensor<i32>, tensor<i32>, tensor<i32>, tensor<i32>) {
  // CHECK: %[[MUL:.*]] = "tf.Mul"
  %0 = "tf.Mul"(%arg0, %arg1) : (tensor<i32>, tensor<i32>) -> tensor<i32>
  // CHECK: %[[RESULT:.*]]:4 = "tf_device.cluster"
  %1:4 = "tf_device.cluster"() ({
    // CHECK: %[[ADD:.*]] = "tf.AddV2"

 MaxPool 1
 MaxPoolGrad 1
 Mean 1
 Mul 218
 Pad 2
 ReadVariableOp 538
 Relu 49
 ReluGrad 49
 Reshape 2
 ResourceApplyKerasMomentum 161
 Slice 1
 Softmax 1
 SparseSoftmaxCrossEntropyWithLogits 1
 Squeeze 1
 Sum 1
 Tile 1
 Transpose 1
cluster 1 size 815
 AddN 1
 AssignAddVariableOp 1
 AssignSubVariableOp 106
 Const 220
 DivNoNan 1
 Identity 1
 Mul 161
 ReadVariableOp 106
 Square 55

  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

                                    %weight_scale : tensor<*xf32>) -> tensor<*xf32> {
    %scale_prod = "tf.Mul"(%input_scale, %weight_scale) : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>

    %cast = "tf.Cast"(%input) : (tensor<*xi32>) -> tensor<*xf32>
    %mul = "tf.Mul"(%cast, %scale_prod) : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
    func.return %mul : tensor<*xf32>
  }

func.func @main(tensor<4xf32>) -> tensor<4xf32> {
^bb0(%arg0: tensor<4xf32>):
  %0 = "tfl.pseudo_const" () {name = "Const", value = dense<1.0> : tensor<4xf32>} : () -> tensor<4xf32>
  %1 = "tfl.mul"(%arg0, %0) {fused_activation_function = "NONE", name = "mul"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
  // tf.div is the result of conversion to a Flex TF op
  %2 = "tf.Div"(%1, %0) {name = "div"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>

    long mul = K3;
    int topBit = 0x7;

    int lengthAligned = length & ~topBit;
    int lengthRemainder = length & topBit;
    long hash = seed ^ (length * mul);

    for (int i = 0; i < lengthAligned; i += 8) {
      long loaded = load64(bytes, offset + i);
      long data = shiftMix(loaded * mul) * mul;
      hash ^= data;
      hash *= mul;
    }

		for i := 0; i < b.N; i++ {
			v.Mul(v, v)
		}
	})
	b.Run("P384", func(b *testing.B) {
		v := new(fiat.P384Element).One()
		b.ReportAllocs()
		b.ResetTimer()
		for i := 0; i < b.N; i++ {
			v.Mul(v, v)
		}
	})
	b.Run("P521", func(b *testing.B) {
		v := new(fiat.P521Element).One()
		b.ReportAllocs()
		b.ResetTimer()
		for i := 0; i < b.N; i++ {
			v.Mul(v, v)
		}
	})
}