func.return %4 : tensor<*xf32>
}

// CHECK: %[[q:.*]] = "quantfork.qcast"(%arg0)
// CHECK: %[[sc1:.*]] = "quantfork.scast"(%[[q]]) : (tensor<*x!quant.uniform<i8:f32, 5.000000e-02:-10>>)
// CHECK: %[[fcast:.*]] = "tf.Cast"(%[[sc1]]) <{Truncate = false}> : (tensor<*xi8>) -> tensor<*xf32>
// CHECK: %[[avgpool_f32:.*]] = "tf.AvgPool"(%[[fcast]])
// CHECK-SAME: (tensor<*xf32>) -> tensor<*xf32>
// CHECK: %[[round:.*]] = "tf.Round"(%[[avgpool_f32]])

// CHECK: %[[scast:.*]] = "quantfork.scast"(%[[conv]]
// CHECK: %[[fcast:.*]] = "tf.Cast"(%[[scast]]) <{Truncate = false}> : (tensor<*xi8>) -> tensor<*xf32>
// CHECK: %[[avgpool_f32:.*]] = "tf.AvgPool"(%[[fcast]])
// CHECK-SAME: (tensor<*xf32>) -> tensor<*xf32>
// CHECK: %[[round:.*]] = "tf.Round"(%[[avgpool_f32]])
// CHECK: %[[icast:.*]] = "tf.Cast"(%[[round]]) <{Truncate = false}> : (tensor<*xf32>) -> tensor<*xi8>

  addOperations<
#define GET_OP_LIST
#include "tensorflow/compiler/mlir/quantization/common/ir/QuantOps.cc.inc"
      >();
}

OpFoldResult StorageCastOp::fold(FoldAdaptor) {
  // Matches x -> [scast -> scast] -> y, replacing the second scast with the
  // value of x if the casts invert each other.
  auto srcScastOp = getArg().getDefiningOp<StorageCastOp>();
  if (!srcScastOp || srcScastOp.getArg().getType() != getType())

  addOperations<
#define GET_OP_LIST
#include "tensorflow/compiler/mlir/lite/quantization/ir/QuantOps.cc.inc"
      >();
}

OpFoldResult StorageCastOp::fold(FoldAdaptor) {
  // Matches x -> [scast -> scast] -> y, replacing the second scast with the
  // value of x if the casts invert each other.
  auto srcScastOp = getArg().getDefiningOp<StorageCastOp>();
  if (!srcScastOp || srcScastOp.getArg().getType() != getType())

// A QuantizeCast (qcast) represents a potential type shift from a quantizable
// type to a quantized type.
//
// At runtime, a qcast will apply the transformation expressed by its
// operand and result type. For flexibility during transformation, it is also
// possible to have a qcast that performs no transformation (both its
// operand and result type are quantizable).
//
// A qcast will typically originate from either:

// A QuantizeCast (qcast) represents a potential type shift from a quantizable
// type to a quantized type.
//
// At runtime, a qcast will apply the transformation expressed by its
// operand and result type. For flexibility during transformation, it is also
// possible to have a qcast that performs no transformation (both its
// operand and result type are quantizable).
//
// A qcast will typically originate from either:

    }
    auto const_op =
        rewriter.create<TF::ConstOp>(loc, new_type, tensor_proto_attr);
    // Add scast op to match quantize -> composition pattern. The added scast
    // is then removed by canonicalization. ([scast - scast] -> [])
    auto scast_op = rewriter.create<quantfork::StorageCastOp>(
        loc, tensor_qtype, const_op.getOutput());
    q_op->replaceAllUsesWith(scast_op);

    %5 = "quantfork.dcast"(%4) : (tensor<2x3x1024x3x!quant.uniform<i8<-127:127>:f32, 4.000000e-01>>) -> tensor<2x3x1024x3xf32>
    %6 = "quantfork.qcast"(%arg0) {volatile} : (tensor<2x3x1x1024xf32>) -> tensor<2x3x1x1024x!quant.uniform<i8:f32, 5.000000e-01:2>>

    // CHECK: %[[DQ:.*]] = "quantfork.dcast"(%[[RESHAPE]]) : (tensor<3x1x!quant.uniform<i8:f32, 0.13170163023705575:-1>>) -> tensor<3x1xf32>
    // CHECK: return %[[DQ]]
    %0 = "quantfork.qcast"(%arg0) {volatile} : (tensor<1x2xf32>) -> tensor<1x2x!quant.uniform<i8:f32, 5.000000e-03>>
    %1 = "quantfork.dcast"(%0) : (tensor<1x2x!quant.uniform<i8:f32, 5.000000e-03>>) -> tensor<1x2xf32>

	switch inst.Op {
	case INSB, INSD, INSW, OUTSB, OUTSD, OUTSW, MONITOR, MWAIT, XLATB:
		iargs = nil

	case STOSB, STOSW, STOSD, STOSQ:
		iargs = iargs[:1]

	case LODSB, LODSW, LODSD, LODSQ, SCASB, SCASW, SCASD, SCASQ:
		iargs = iargs[1:]
	}

	const (
		haveData16 = 1 << iota
		haveData32
		haveAddr16
		haveAddr32
		haveXacquire
		haveXrelease
		haveLock
		haveHintTaken