while (op && op != func) {
      op->removeAttr("_is_compilation");
      op->removeAttr("_wraps_compilation");
      op = op->getParentOp();
    }
  });
}

int OutsideCompilationOrdering(Operation* predecessor, Operation* op) {
  // Actual compilations go first.
  if (op->hasAttr("_is_compilation")) return 2;
  // Followed by nested ops that contain compilations.
  if (op->hasAttr("_wraps_compilation")) return 1;

	//    and len(b.Preds) is equal to 2.
	// 2. find the common dominator(b0) of the predecessors(pb0, pb1) of block b, and the
	//    dominator(b0) is a If block.
	//    Special case: one of the predecessors(pb0 or pb1) is the dominator(b0).
	// 3. the successors(sb0, sb1) of the dominator need to dominate the predecessors(pb0, pb1)
	//    of block b respectively.
	// 4. replace this boolean Phi based on dominator block.
	//

    }
  }
  return successors;
}
// Gets the predecessors of an op wrapped in tf_executor.island.
llvm::SmallVector<Operation*> GetPredecessors(Operation* op) {
  llvm::SmallVector<Operation*> predecessors;
  for (auto operand : op->getOperands()) {
    if (Operation* pred = operand.getDefiningOp()) {
      pred->walk([&](mlir::Operation* opinexecutor) {
        predecessors.push_back(opinexecutor);
      });
    }
  }

// not be a predecessor.
//
// Note that mergePredecessors behaves slightly differently between
// first and subsequent calls for a block.  For the first call, the
// starting state is approximated by taking the state from the
// predecessor whose state is smallest, and removing any elements not
// in all the other predecessors; this makes the smallest number of

			// 2, Avoid the predecessor that contains the function call, because the predecessor that
			//    contains the function call usually generates a lot of spills and lose the previous
			//    allocation state.
			// TODO: Improve this part. At least the size of endRegs of the predecessor also has
			// an impact on the code size and compiler speed. But it is not easy to find a simple

    auto predecessors = analysis.DirectControlPredecessors(&cluster_op);
    if (!predecessors.empty() && !IsEmbeddingOp(&cluster_op)) {
      bool skip = false;
      for (Operation* predecessor : llvm::reverse(predecessors)) {
        if (IsEmbeddingOp(predecessor)) continue;
        skip = !head_outside_compiled_ops.contains(predecessor);
        break;
      }
      if (skip) continue;

  let constructor = "TFDevice::CreateEmbeddingProgramKeyPass()";
    let description = [{
    Passes in the program key to embedding ops. Will move the embedding ops
    after a _TPUCompileMlir op if there is no predecessor _TPUCompileMlir op.
    Both the embedding op and compile op are assumed to be wrapped in separate
    tf_device.launch() ops. This is because the embedding op is head outside

     * the target state, not the input state. Implicitly, a rule accepts as input the subject in the state that is the predecessor of the target state.
     */
    public BindingPredicate getPredicate() {
        return predicate;
    }

    public boolean isBound() {
        return boundTo != null;
    }

        // (parent) block, we need to check that it's before (the
        // (parent of) the preprocess_op.
        if (o->isBeforeInBlock(it->second)) {
          break;  // valid compile predecessor
        } else {
          return WalkResult::advance();
        }
      }
      o = o->getParentOp();
    }
    // Check that the the compile op actually passes its results to its parents.

    AnalyzeRegion(region);
  }

  SideEffectAnalysisInfo(SideEffectAnalysisInfo&&) = default;

  // Returns a vector of ops that are direct control predecessors of `op`,
  // sorted in program order. If `filter` is provided, only predecessors that
  // pass the filter (returning true) will be included.
  const llvm::SmallVector<Operation*, 4>& DirectControlPredecessors(
      Operation* op) const;