Value operand = op->getOperand(i);
      if (Operation* inst = operand.getDefiningOp()) {
        // If the operand comes from a `quantfork::DequantizeCastOp`, we use
        // the quantized input of this `quantfork::DequantizeCastOp` to set the
        // state.
        if (auto dq = dyn_cast<quantfork::DequantizeCastOp>(inst)) {
          operand = dq.getArg();
        }
      }

       << "const std::vector<int32_t>& operands,"
       << "const std::vector<int32_t>& results,"
       << (has_intermediates ? "const std::vector<int32_t>& intermediate_index,"
                             : "")
       << "flatbuffers::FlatBufferBuilder *fbb) {\n";

    // Inputs & outputs
    os << "  auto inputs = fbb->CreateVector(operands);\n"
          "  auto outputs = fbb->CreateVector(results);\n\n";

#include "mlir/Support/LogicalResult.h"  // from @llvm-project
#include "stablehlo/dialect/StablehloOps.h"  // from @stablehlo

namespace mlir::odml {

namespace {

// Helper class for parsing operands to a foldable operation.
class FoldAdaptor {
 public:
  // Returns std::nullopt if the operation cannot be folded.
  static std::optional<FoldAdaptor> Create(Operation* operation) {

		if !ok || bin.Op != op.badEq {
			continue
		}

		// In order to avoid false positives, restrict to cases
		// in which one of the operands is constant. We're then
		// interested in the other operand.
		// In the rare case in which both operands are constant
		// (e.g. runtime.GOOS and "windows"), we'll only catch
		// mistakes if the LHS is repeated, which is how most
		// code is written.
		var x ast.Expr

}

/* VA-form 3-register + SHB operands */
func AOP_IRRR(op uint32, d uint32, a uint32, b uint32, simm uint32) uint32 {
	return op | (d&31)<<21 | (a&31)<<16 | (b&31)<<11 | (simm&0xF)<<6
}

/* VX-form 1-register + SIM operands */
func AOP_IR(op uint32, d uint32, simm uint32) uint32 {
	return op | (d&31)<<21 | (simm&31)<<16
}

/* XX1-form 3-register operands, 1 VSR operand */

  BFloat16TypePattern(TypeConverter& converter, MLIRContext* ctx)
      : ConversionPattern(converter, MatchAnyOpTypeTag(), /*benefit=*/1, ctx) {}

  LogicalResult matchAndRewrite(
      Operation* op, const ArrayRef<Value> operands,
      ConversionPatternRewriter& rewriter) const override {
    if (getTypeConverter()->isLegal(op)) {
      return failure();
    }
    if (isa<mlir::stablehlo::BitcastConvertOp>(op)) {

}

// -----

func.func @testConcatV2(%arg0: tensor<8x16xf32>, %arg1: tensor<8xf32>, %axis: tensor<i32>) -> tensor<*xf32> {
  // expected-error @+1 {{operand type 'tensor<8xf32>' is not compatible with preceding operands; expected rank: 2}}
  %0 = "tf.ConcatV2"(%arg0, %arg1, %axis) : (tensor<8x16xf32>, tensor<8xf32>, tensor<i32>) -> tensor<*xf32>
  func.return %0 : tensor<*xf32>
}

// -----

    private static Collection<?> toList(Object[] operands) {
        return singleCollection(operands) ? (Collection<?>) operands[0] : Arrays.asList(operands);
    }

    private static boolean singleCollection(Object[] operands) {
        return operands.length== 1 && operands[0] instanceof Collection;
    }

    private enum Subject {
        all(true, true),
        stackoverflow(false, true),
        operations(true, false);

// location with a `CallOp` which invokes said `FuncOp`. The inputs to
// this new functions are taken to be the `Values` that appear as operands
// to ops in the subgraph, which are not self-contained within the subgraph.
// The outputs of this function are taken to be the results of ops in the
// subgraph which are referenced as operands outside of the subgraph.
// Also refer to documention of `AccumulateOperandsDefinedAbove` &
// `AccumulateResultsDefinedWithin`.

  });
}

// Propagates the resource IDs from an input operand to a result. Returns true
// if the mapping changed.
bool ResourceAliasAnalysisInfo::PropagateInputToOutput(const Value& operand,
                                                       const OpResult& result) {
  auto operand_it = resource_value_to_ids_.find(operand);
  assert(operand_it != resource_value_to_ids_.end() &&