diff --git a/backends/aiger/xaiger.cc b/backends/aiger/xaiger.cc
index bad9322bb..ce10028f7 100644
--- a/backends/aiger/xaiger.cc
+++ b/backends/aiger/xaiger.cc
@@ -61,6 +61,8 @@ struct XAigerWriter
 	dict<SigBit, int> init_inputs;
 	int initstate_ff = 0;
 
+	vector<Cell*> box_list;
+
 	int mkgate(int a0, int a1)
 	{
 		aig_m++, aig_a++;
@@ -211,7 +213,7 @@ struct XAigerWriter
 			}
 
 			for (const auto &c : cell->connections()) {
-				if (c.second.is_fully_const()) continue;
+				/*if (c.second.is_fully_const()) continue;*/
 				for (auto b : c.second.bits()) {
 					Wire *w = b.wire;
 					if (!w) continue;
@@ -219,30 +221,32 @@ struct XAigerWriter
 					auto is_output = cell->output(c.first);
 					log_assert(is_input || is_output);
 					if (is_input) {
-						if (!w->port_input) {
+						/*if (!w->port_input)*/ {
 							SigBit I = sigmap(b);
 							if (I != b)
 								alias_map[b] = I;
-							if (!output_bits.count(b))
+							/*if (!output_bits.count(b))*/
 								co_bits.insert(b);
 						}
 					}
 					if (is_output) {
 						SigBit O = sigmap(b);
-						if (!input_bits.count(O))
+						/*if (!input_bits.count(O))*/
 							ci_bits.insert(O);
 					}
 				}
 				if (!type_map.count(cell->type))
 					type_map[cell->type] = type_map.size()+1;
 			}
-			//log_error("Unsupported cell type: %s (%s)\n", log_id(cell->type), log_id(cell));
+
+			box_list.emplace_back(cell);
+			log_warning("Unsupported cell type: %s (%s)\n", log_id(cell->type), log_id(cell));
 		}
 
 		for (auto bit : input_bits) {
 			RTLIL::Wire *wire = bit.wire;
 			// If encountering an inout port, then create a new wire with $inout.out
-			// suffix, make it a CO driven by the existing inout, and inherit existing
+			// suffix, make it a PO driven by the existing inout, and inherit existing
 			// inout's drivers
 			if (wire->port_input && wire->port_output && !undriven_bits.count(bit)) {
 				RTLIL::Wire *new_wire = module->wire(wire->name.str() + "$inout.out");
@@ -256,19 +260,19 @@ struct XAigerWriter
 					and_map[new_bit] = and_map.at(bit);
 				else if (alias_map.count(bit))
 					alias_map[new_bit] = alias_map.at(bit);
-				co_bits.insert(new_bit);
+				output_bits.insert(new_bit);
 			}
 		}
 
 		// Do some CI/CO post-processing:
 		// Erase all POs and COs that are undriven
 		for (auto bit : undriven_bits) {
-			co_bits.erase(bit);
+			//co_bits.erase(bit);
 			output_bits.erase(bit);
 		}
 		// Erase all CIs that are also COs
-		for (auto bit : co_bits)
-			ci_bits.erase(bit);
+		//for (auto bit : co_bits)
+		//	ci_bits.erase(bit);
 		// CIs cannot be undriven
 		for (auto bit : ci_bits)
 			undriven_bits.erase(bit);
@@ -491,7 +495,8 @@ struct XAigerWriter
 
 					if (output_bits.count(b) || co_bits.count(b)) {
 						int o = ordered_outputs.at(b);
-						output_seen = !miter_mode;
+						if (output_seen && output_bits.count(b))
+							output_seen = !miter_mode;
 						if (GetSize(wire) != 1)
 							symbols[stringf("%c%d", miter_mode ? 'b' : 'o', o)].push_back(stringf("%s[%d]", log_id(wire), i));
 						else
@@ -532,7 +537,52 @@ struct XAigerWriter
 			}
 		}
 
-		f << stringf("c\nGenerated by %s\n", yosys_version_str);
+		f << "c";
+
+		std::stringstream h_buffer;
+		auto write_h_buffer = [&h_buffer](int i32) {
+			// TODO: Don't assume we're on little endian
+#ifdef _WIN32
+			int i32_be = _byteswap_ulong(i32);
+#else
+			int i32_be = __builtin_bswap32(i32);
+#endif
+			h_buffer.write(reinterpret_cast<const char*>(&i32_be), sizeof(i32_be));
+		};
+		int num_outputs = output_bits.size();
+		if (omode && num_outputs == 0)
+			num_outputs = 1;
+		write_h_buffer(1);
+		write_h_buffer(input_bits.size() + ci_bits.size());
+		write_h_buffer(num_outputs + co_bits.size());
+		write_h_buffer(input_bits.size());
+		write_h_buffer(num_outputs);
+		write_h_buffer(box_list.size());
+		int box_id = 0;
+		for (auto cell : box_list) {
+			int box_inputs = 0, box_outputs = 0;
+			for (const auto &c : cell->connections())
+				if (cell->input(c.first))
+					box_inputs += c.second.size();
+				else
+					box_outputs += c.second.size();
+			write_h_buffer(box_inputs);
+			write_h_buffer(box_outputs);
+			write_h_buffer(box_id++);
+			write_h_buffer(0 /* OldBoxNum */);
+		}
+		std::string h_buffer_str = h_buffer.str();
+		// TODO: Don't assume we're on little endian
+#ifdef _WIN32
+		int h_buffer_size_be = _byteswap_ulong(h_buffer_str.size());
+#else
+		int h_buffer_size_be = __builtin_bswap32(h_buffer_str.size());
+#endif
+		f << "h";
+		f.write(reinterpret_cast<const char*>(&h_buffer_size_be), sizeof(h_buffer_size_be));
+		f.write(h_buffer_str.data(), h_buffer_str.size());
+
+		f << stringf("Generated by %s\n", yosys_version_str);
 	}
 
 	void write_map(std::ostream &f, bool verbose_map, bool omode)
@@ -557,7 +607,11 @@ struct XAigerWriter
 					int a = aig_map.at(sig[i]);
 					log_assert((a & 1) == 0);
 					input_lines[a] += stringf("input %d %d %s\n", (a >> 1)-1, i, log_id(wire));
-					continue;
+
+					// Only continue if this input is not a CO,
+					// otherwise write as CO below
+					if (!co_bits.count(b))
+						continue;
 				}
 
 				if (output_bits.count(b) || co_bits.count(b)) {
@@ -606,7 +660,7 @@ struct XAigerWriter
 			f << it.second;
 		log_assert(output_lines.size() == output_bits.size() + co_bits.size());
 		if (omode && output_lines.empty())
-			f << "output 0 0 __dummy_o__\n";
+			f << "output " << output_lines.size() << " 0 __dummy_o__\n";
 
 		latch_lines.sort();
 		for (auto &it : latch_lines)