return true
	}
	return false
}

// IsLoong64MUL reports whether the op (as defined by an loong64.A* constant) is
// one of the MUL/DIV/REM instructions that require special handling.
func IsLoong64MUL(op obj.As) bool {
	switch op {
	case loong64.AMUL, loong64.AMULU, loong64.AMULV, loong64.AMULVU,
		loong64.ADIV, loong64.ADIVU, loong64.ADIVV, loong64.ADIVVU,
		loong64.AREM, loong64.AREMU, loong64.AREMV, loong64.AREMVU:

  %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")

  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

  %1 = "tfl.mul"(%0, %cst_2) {fused_activation_function = "NONE"} : (tensor<1x256xbf16>, tensor<32x1x256xbf16>) -> tensor<32x1x256xbf16>
  func.return %1 : tensor<32x1x256xbf16>

// CHECK:  %[[V0:.*]] = "tfl.fully_connected"(%arg0, {{.*}}) <{{{.*}}}> : (tensor<1x10368xbf16>, tensor<256x10368xbf16>, none) -> tensor<1x256xbf16>

}

// --------------- //
//    bits.Mul*    //
// --------------- //

func Mul(x, y uint) (hi, lo uint) {
	// amd64:"MULQ"
	// arm64:"UMULH","MUL"
	// ppc64x:"MULHDU","MULLD"
	// s390x:"MLGR"
	// mips64: "MULVU"
	// riscv64:"MULHU","MUL"
	return bits.Mul(x, y)
}

func Mul64(x, y uint64) (hi, lo uint64) {
	// amd64:"MULQ"
	// arm64:"UMULH","MUL"
	// ppc64x:"MULHDU","MULLD"
	// s390x:"MLGR"

  // CHECK: %[[CONST:.*]] = "tf.Const"()
  // CHECK-DAG: %[[MUL1:.*]] = "tf.Mul"(%arg0, %arg2)
  // CHECK-DAG: %[[MUL2:.*]] = "tf.Mul"(%arg1, %arg3)
  // CHECK: "tf.ConcatV2"(%[[MUL1]], %[[MUL2]], %[[CONST]])
  %0 = "tf.Const"() { value = dense<1> : tensor<i32> } : () -> tensor<i32>
  %1 = "tf.Mul"(%arg0, %arg2) : (tensor<?x2xf32>, tensor<f32>) -> tensor<?x2xf32>
  %2 = "tf.Mul"(%arg1, %arg3) : (tensor<?x2xf32>, tensor<f32>) -> tensor<?x2xf32>

int64_t GetElementIndex(llvm::SmallVectorImpl<int64_t> &shape,
                        llvm::SmallVectorImpl<int64_t> &current_index) {
  int64_t ind = 0;
  int64_t mul = 1;
  for (int i = shape.size() - 1; i >= 0; --i) {
    ind += (current_index[i] % shape[i]) * mul;
    mul *= shape[i];
  }
  return ind;
}

// Helper method that increment index represented in 'current_index_ptr'
// in the shape of 'result_shape'.

  (Add32 (Const32 <t> [c*d]) (Mul32 <t> (Const32 <t> [c]) x))

// Rewrite x*y ± x*z  to  x*(y±z)
(Add(64|32|16|8) <t> (Mul(64|32|16|8) x y) (Mul(64|32|16|8) x z))
	=> (Mul(64|32|16|8) x (Add(64|32|16|8) <t> y z))
(Sub(64|32|16|8) <t> (Mul(64|32|16|8) x y) (Mul(64|32|16|8) x z))
	=> (Mul(64|32|16|8) x (Sub(64|32|16|8) <t> y z))

// rewrite shifts of 8/16/32 bit consts into 64 bit consts to reduce

  %5 = "tfl.mul"(%0, %1) {fused_activation_function = "NONE"} : (tensor<  f32>, tensor<  f32>) -> tensor<  f32>
  %6 = "tfl.mul"(%0, %3) {fused_activation_function = "NONE"} : (tensor<  f32>, tensor<4xf32>) -> tensor<4xf32>
  %7 = "tfl.mul"(%2, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<  f32>) -> tensor<4xf32>

    %0 = tfl.add %arg0, %arg1 {fused_activation_function = "RELU6", tac.device = "GPU", tac.inference_type = "FLOAT"} : tensor<1xf32>
    %1 = tfl.mul %0, %arg2 {fused_activation_function = "RELU6", tac.device = "GPU", tac.inference_type = "FLOAT"} : tensor<1xf32>
    func.return %1 : tensor<1xf32>
  }