solidity/libevmasm/Assembly.cpp

/*
	This file is part of solidity.

	solidity is free software: you can redistribute it and/or modify
	it under the terms of the GNU General Public License as published by
	the Free Software Foundation, either version 3 of the License, or
	(at your option) any later version.

	solidity is distributed in the hope that it will be useful,
	but WITHOUT ANY WARRANTY; without even the implied warranty of
	MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
	GNU General Public License for more details.

	You should have received a copy of the GNU General Public License
	along with solidity.  If not, see <http://www.gnu.org/licenses/>.
*/
// SPDX-License-Identifier: GPL-3.0
/** @file Assembly.cpp
 * @author Gav Wood <i@gavwood.com>
 * @date 2014
 */

#include <libevmasm/Assembly.h>

#include <libevmasm/CommonSubexpressionEliminator.h>
#include <libevmasm/ControlFlowGraph.h>
#include <libevmasm/PeepholeOptimiser.h>
#include <libevmasm/Inliner.h>
#include <libevmasm/JumpdestRemover.h>
#include <libevmasm/BlockDeduplicator.h>
#include <libevmasm/ConstantOptimiser.h>
#include <libevmasm/GasMeter.h>

#include <liblangutil/Exceptions.h>

#include <json/json.h>

#include <fstream>
#include <range/v3/algorithm/any_of.hpp>

using namespace std;
using namespace solidity;
using namespace solidity::evmasm;
using namespace solidity::langutil;
using namespace solidity::util;

AssemblyItem const& Assembly::append(AssemblyItem const& _i)
{
	assertThrow(m_deposit >= 0, AssemblyException, "Stack underflow.");
	m_deposit += static_cast<int>(_i.deposit());
	m_items.emplace_back(_i);
	if (!m_items.back().location().isValid() && m_currentSourceLocation.isValid())
		m_items.back().setLocation(m_currentSourceLocation);
	m_items.back().m_modifierDepth = m_currentModifierDepth;
	return m_items.back();
}

unsigned Assembly::bytesRequired(unsigned subTagSize) const
{
	for (unsigned tagSize = subTagSize; true; ++tagSize)
	{
		size_t ret = 1;
		for (auto const& i: m_data)
			ret += i.second.size();

		for (AssemblyItem const& i: m_items)
			ret += i.bytesRequired(tagSize);
		if (util::bytesRequired(ret) <= tagSize)
			return static_cast<unsigned>(ret);
	}
}

namespace
{

string locationFromSources(StringMap const& _sourceCodes, SourceLocation const& _location)
{
	if (!_location.hasText() || _sourceCodes.empty())
		return "";

	auto it = _sourceCodes.find(_location.source->name());
	if (it == _sourceCodes.end())
		return "";

	string const& source = it->second;
	if (static_cast<size_t>(_location.start) >= source.size())
		return "";

	string cut = source.substr(static_cast<size_t>(_location.start), static_cast<size_t>(_location.end - _location.start));
	auto newLinePos = cut.find_first_of("\n");
	if (newLinePos != string::npos)
		cut = cut.substr(0, newLinePos) + "...";

	return cut;
}

class Functionalizer
{
public:
	Functionalizer (ostream& _out, string const& _prefix, StringMap const& _sourceCodes, Assembly const& _assembly):
		m_out(_out), m_prefix(_prefix), m_sourceCodes(_sourceCodes), m_assembly(_assembly)
	{}

	void feed(AssemblyItem const& _item)
	{
		if (_item.location().isValid() && _item.location() != m_location)
		{
			flush();
			m_location = _item.location();
			printLocation();
		}

		string expression = _item.toAssemblyText(m_assembly);

		if (!(
			_item.canBeFunctional() &&
			_item.returnValues() <= 1 &&
			_item.arguments() <= m_pending.size()
		))
		{
			flush();
			m_out << m_prefix << (_item.type() == Tag ? "" : "  ") << expression << endl;
			return;
		}
		if (_item.arguments() > 0)
		{
			expression += "(";
			for (size_t i = 0; i < _item.arguments(); ++i)
			{
				expression += m_pending.back();
				m_pending.pop_back();
				if (i + 1 < _item.arguments())
					expression += ", ";
			}
			expression += ")";
		}

		m_pending.push_back(expression);
		if (_item.returnValues() != 1)
			flush();
	}

	void flush()
	{
		for (string const& expression: m_pending)
			m_out << m_prefix << "  " << expression << endl;
		m_pending.clear();
	}

	void printLocation()
	{
		if (!m_location.isValid())
			return;
		m_out << m_prefix << "    /*";
		if (m_location.source)
			m_out << " \"" + m_location.source->name() + "\"";
		if (m_location.hasText())
			m_out << ":" << to_string(m_location.start) + ":" + to_string(m_location.end);
		m_out << "  " << locationFromSources(m_sourceCodes, m_location);
		m_out << " */" << endl;
	}

private:
	strings m_pending;
	SourceLocation m_location;

	ostream& m_out;
	string const& m_prefix;
	StringMap const& m_sourceCodes;
	Assembly const& m_assembly;
};

}

void Assembly::assemblyStream(ostream& _out, string const& _prefix, StringMap const& _sourceCodes) const
{
	Functionalizer f(_out, _prefix, _sourceCodes, *this);

	for (auto const& i: m_items)
		f.feed(i);
	f.flush();

	if (!m_data.empty() || !m_subs.empty())
	{
		_out << _prefix << "stop" << endl;
		for (auto const& i: m_data)
			if (u256(i.first) >= m_subs.size())
				_out << _prefix << "data_" << toHex(u256(i.first)) << " " << toHex(i.second) << endl;

		for (size_t i = 0; i < m_subs.size(); ++i)
		{
			_out << endl << _prefix << "sub_" << i << ": assembly {\n";
			m_subs[i]->assemblyStream(_out, _prefix + "    ", _sourceCodes);
			_out << _prefix << "}" << endl;
		}
	}

	if (m_auxiliaryData.size() > 0)
		_out << endl << _prefix << "auxdata: 0x" << toHex(m_auxiliaryData) << endl;
}

string Assembly::assemblyString(StringMap const& _sourceCodes) const
{
	ostringstream tmp;
	assemblyStream(tmp, "", _sourceCodes);
	return tmp.str();
}

Json::Value Assembly::createJsonValue(string _name, int _source, int _begin, int _end, string _value, string _jumpType)
{
	Json::Value value;
	value["name"] = _name;
	value["source"] = _source;
	value["begin"] = _begin;
	value["end"] = _end;
	if (!_value.empty())
		value["value"] = _value;
	if (!_jumpType.empty())
		value["jumpType"] = _jumpType;
	return value;
}

string Assembly::toStringInHex(u256 _value)
{
	std::stringstream hexStr;
	hexStr << std::uppercase << hex << _value;
	return hexStr.str();
}

Json::Value Assembly::assemblyJSON(map<string, unsigned> const& _sourceIndices) const
{
	Json::Value root;

	Json::Value& collection = root[".code"] = Json::arrayValue;
	for (AssemblyItem const& i: m_items)
	{
		int sourceIndex = -1;
		if (i.location().source)
		{
			auto iter = _sourceIndices.find(i.location().source->name());
			if (iter != _sourceIndices.end())
				sourceIndex = static_cast<int>(iter->second);
		}

		switch (i.type())
		{
		case Operation:
			collection.append(
				createJsonValue(
					instructionInfo(i.instruction()).name,
					sourceIndex,
					i.location().start,
					i.location().end,
					i.getJumpTypeAsString())
				);
			break;
		case Push:
			collection.append(
				createJsonValue("PUSH", sourceIndex, i.location().start, i.location().end, toStringInHex(i.data()), i.getJumpTypeAsString()));
			break;
		case PushString:
			collection.append(
				createJsonValue("PUSH tag", sourceIndex, i.location().start, i.location().end, m_strings.at(h256(i.data()))));
			break;
		case PushTag:
			if (i.data() == 0)
				collection.append(
					createJsonValue("PUSH [ErrorTag]", sourceIndex, i.location().start, i.location().end, ""));
			else
				collection.append(
					createJsonValue("PUSH [tag]", sourceIndex, i.location().start, i.location().end, toString(i.data())));
			break;
		case PushSub:
			collection.append(
				createJsonValue("PUSH [$]", sourceIndex, i.location().start, i.location().end, toString(h256(i.data()))));
			break;
		case PushSubSize:
			collection.append(
				createJsonValue("PUSH #[$]", sourceIndex, i.location().start, i.location().end, toString(h256(i.data()))));
			break;
		case PushProgramSize:
			collection.append(
				createJsonValue("PUSHSIZE", sourceIndex, i.location().start, i.location().end));
			break;
		case PushLibraryAddress:
			collection.append(
				createJsonValue("PUSHLIB", sourceIndex, i.location().start, i.location().end, m_libraries.at(h256(i.data())))
			);
			break;
		case PushDeployTimeAddress:
			collection.append(
				createJsonValue("PUSHDEPLOYADDRESS", sourceIndex, i.location().start, i.location().end)
			);
			break;
		case PushImmutable:
			collection.append(createJsonValue(
				"PUSHIMMUTABLE",
				sourceIndex,
				i.location().start,
				i.location().end,
				m_immutables.at(h256(i.data()))
			));
			break;
		case AssignImmutable:
			collection.append(createJsonValue(
				"ASSIGNIMMUTABLE",
				sourceIndex,
				i.location().start,
				i.location().end,
				m_immutables.at(h256(i.data()))
			));
			break;
		case Tag:
			collection.append(
				createJsonValue("tag", sourceIndex, i.location().start, i.location().end, toString(i.data())));
			collection.append(
				createJsonValue("JUMPDEST", sourceIndex, i.location().start, i.location().end));
			break;
		case PushData:
			collection.append(createJsonValue("PUSH data", sourceIndex, i.location().start, i.location().end, toStringInHex(i.data())));
			break;
		case VerbatimBytecode:
			collection.append(createJsonValue("VERBATIM", sourceIndex, i.location().start, i.location().end, toHex(i.verbatimData())));
			break;
		default:
			assertThrow(false, InvalidOpcode, "");
		}
	}

	if (!m_data.empty() || !m_subs.empty())
	{
		Json::Value& data = root[".data"] = Json::objectValue;
		for (auto const& i: m_data)
			if (u256(i.first) >= m_subs.size())
				data[toStringInHex((u256)i.first)] = toHex(i.second);

		for (size_t i = 0; i < m_subs.size(); ++i)
		{
			std::stringstream hexStr;
			hexStr << hex << i;
			data[hexStr.str()] = m_subs[i]->assemblyJSON(_sourceIndices);
		}
	}

	if (m_auxiliaryData.size() > 0)
		root[".auxdata"] = toHex(m_auxiliaryData);

	return root;
}

AssemblyItem Assembly::namedTag(string const& _name, size_t _params, size_t _returns, optional<uint64_t> _sourceID)
{
	assertThrow(!_name.empty(), AssemblyException, "Empty named tag.");
	if (m_namedTags.count(_name))
	{
		assertThrow(m_namedTags.at(_name).params == _params, AssemblyException, "");
		assertThrow(m_namedTags.at(_name).returns == _returns, AssemblyException, "");
		assertThrow(m_namedTags.at(_name).sourceID == _sourceID, AssemblyException, "");
	}
	else
		m_namedTags[_name] = {static_cast<size_t>(newTag().data()), _sourceID, _params, _returns};
	return AssemblyItem{Tag, m_namedTags.at(_name).id};
}

AssemblyItem Assembly::newPushLibraryAddress(string const& _identifier)
{
	h256 h(util::keccak256(_identifier));
	m_libraries[h] = _identifier;
	return AssemblyItem{PushLibraryAddress, h};
}

AssemblyItem Assembly::newPushImmutable(string const& _identifier)
{
	h256 h(util::keccak256(_identifier));
	m_immutables[h] = _identifier;
	return AssemblyItem{PushImmutable, h};
}

AssemblyItem Assembly::newImmutableAssignment(string const& _identifier)
{
	h256 h(util::keccak256(_identifier));
	m_immutables[h] = _identifier;
	return AssemblyItem{AssignImmutable, h};
}

Assembly& Assembly::optimise(bool _enable, EVMVersion _evmVersion, bool _isCreation, size_t _runs)
{
	OptimiserSettings settings;
	settings.isCreation = _isCreation;
	settings.runInliner = true;
	settings.runJumpdestRemover = true;
	settings.runPeephole = true;
	if (_enable)
	{
		settings.runDeduplicate = true;
		settings.runCSE = true;
		settings.runConstantOptimiser = true;
	}
	settings.evmVersion = _evmVersion;
	settings.expectedExecutionsPerDeployment = _runs;
	optimise(settings);
	return *this;
}


Assembly& Assembly::optimise(OptimiserSettings const& _settings)
{
	optimiseInternal(_settings, {});
	return *this;
}

map<u256, u256> Assembly::optimiseInternal(
	OptimiserSettings const& _settings,
	std::set<size_t> _tagsReferencedFromOutside
)
{
	// Run optimisation for sub-assemblies.
	for (size_t subId = 0; subId < m_subs.size(); ++subId)
	{
		OptimiserSettings settings = _settings;
		// Disable creation mode for sub-assemblies.
		settings.isCreation = false;
		map<u256, u256> subTagReplacements = m_subs[subId]->optimiseInternal(
			settings,
			JumpdestRemover::referencedTags(m_items, subId)
		);
		// Apply the replacements (can be empty).
		BlockDeduplicator::applyTagReplacement(m_items, subTagReplacements, subId);
	}

	map<u256, u256> tagReplacements;
	// Iterate until no new optimisation possibilities are found.
	for (unsigned count = 1; count > 0;)
	{
		count = 0;

		if (_settings.runInliner)
			Inliner{
				m_items,
				_tagsReferencedFromOutside,
				_settings.expectedExecutionsPerDeployment,
				_settings.isCreation,
				_settings.evmVersion
			}.optimise();

		if (_settings.runJumpdestRemover)
		{
			JumpdestRemover jumpdestOpt{m_items};
			if (jumpdestOpt.optimise(_tagsReferencedFromOutside))
				count++;
		}

		if (_settings.runPeephole)
		{
			PeepholeOptimiser peepOpt{m_items};
			while (peepOpt.optimise())
			{
				count++;
				assertThrow(count < 64000, OptimizerException, "Peephole optimizer seems to be stuck.");
			}
		}

		// This only modifies PushTags, we have to run again to actually remove code.
		if (_settings.runDeduplicate)
		{
			BlockDeduplicator deduplicator{m_items};
			if (deduplicator.deduplicate())
			{
				for (auto const& replacement: deduplicator.replacedTags())
				{
					assertThrow(
						replacement.first <= numeric_limits<size_t>::max() && replacement.second <= numeric_limits<size_t>::max(),
						OptimizerException,
						"Invalid tag replacement."
					);
					assertThrow(
						!tagReplacements.count(replacement.first),
						OptimizerException,
						"Replacement already known."
					);
					tagReplacements[replacement.first] = replacement.second;
					if (_tagsReferencedFromOutside.erase(static_cast<size_t>(replacement.first)))
						_tagsReferencedFromOutside.insert(static_cast<size_t>(replacement.second));
				}
				count++;
			}
		}

		if (_settings.runCSE)
		{
			// Control flow graph optimization has been here before but is disabled because it
			// assumes we only jump to tags that are pushed. This is not the case anymore with
			// function types that can be stored in storage.
			AssemblyItems optimisedItems;

			bool usesMSize = ranges::any_of(m_items, [](AssemblyItem const& _i) {
				return _i == AssemblyItem{Instruction::MSIZE} || _i.type() == VerbatimBytecode;
			});

			auto iter = m_items.begin();
			while (iter != m_items.end())
			{
				KnownState emptyState;
				CommonSubexpressionEliminator eliminator{emptyState};
				auto orig = iter;
				iter = eliminator.feedItems(iter, m_items.end(), usesMSize);
				bool shouldReplace = false;
				AssemblyItems optimisedChunk;
				try
				{
					optimisedChunk = eliminator.getOptimizedItems();
					shouldReplace = (optimisedChunk.size() < static_cast<size_t>(iter - orig));
				}
				catch (StackTooDeepException const&)
				{
					// This might happen if the opcode reconstruction is not as efficient
					// as the hand-crafted code.
				}
				catch (ItemNotAvailableException const&)
				{
					// This might happen if e.g. associativity and commutativity rules
					// reorganise the expression tree, but not all leaves are available.
				}

				if (shouldReplace)
				{
					count++;
					optimisedItems += optimisedChunk;
				}
				else
					copy(orig, iter, back_inserter(optimisedItems));
			}
			if (optimisedItems.size() < m_items.size())
			{
				m_items = move(optimisedItems);
				count++;
			}
		}
	}

	if (_settings.runConstantOptimiser)
		ConstantOptimisationMethod::optimiseConstants(
			_settings.isCreation,
			_settings.isCreation ? 1 : _settings.expectedExecutionsPerDeployment,
			_settings.evmVersion,
			*this
		);

	return tagReplacements;
}

LinkerObject const& Assembly::assemble() const
{
	assertThrow(!m_invalid, AssemblyException, "Attempted to assemble invalid Assembly object.");
	// Return the already assembled object, if present.
	if (!m_assembledObject.bytecode.empty())
		return m_assembledObject;
	// Otherwise ensure the object is actually clear.
	assertThrow(m_assembledObject.linkReferences.empty(), AssemblyException, "Unexpected link references.");

	LinkerObject& ret = m_assembledObject;

	size_t subTagSize = 1;
	map<u256, pair<string, vector<size_t>>> immutableReferencesBySub;
	for (auto const& sub: m_subs)
	{
		auto const& linkerObject = sub->assemble();
		if (!linkerObject.immutableReferences.empty())
		{
			assertThrow(
				immutableReferencesBySub.empty(),
				AssemblyException,
				"More than one sub-assembly references immutables."
			);
			immutableReferencesBySub = linkerObject.immutableReferences;
		}
		for (size_t tagPos: sub->m_tagPositionsInBytecode)
			if (tagPos != numeric_limits<size_t>::max() && tagPos > subTagSize)
				subTagSize = tagPos;
	}

	bool setsImmutables = false;
	bool pushesImmutables = false;

	for (auto const& i: m_items)
		if (i.type() == AssignImmutable)
		{
			i.setImmutableOccurrences(immutableReferencesBySub[i.data()].second.size());
			setsImmutables = true;
		}
		else if (i.type() == PushImmutable)
			pushesImmutables = true;
	if (setsImmutables || pushesImmutables)
		assertThrow(
			setsImmutables != pushesImmutables,
			AssemblyException,
			"Cannot push and assign immutables in the same assembly subroutine."
		);

	unsigned bytesRequiredForCode = bytesRequired(static_cast<unsigned>(subTagSize));
	m_tagPositionsInBytecode = vector<size_t>(m_usedTags, numeric_limits<size_t>::max());
	map<size_t, pair<size_t, size_t>> tagRef;
	multimap<h256, unsigned> dataRef;
	multimap<size_t, size_t> subRef;
	vector<unsigned> sizeRef; ///< Pointers to code locations where the size of the program is inserted
	unsigned bytesPerTag = util::bytesRequired(bytesRequiredForCode);
	uint8_t tagPush = static_cast<uint8_t>(pushInstruction(bytesPerTag));

	unsigned bytesRequiredIncludingData = bytesRequiredForCode + 1 + static_cast<unsigned>(m_auxiliaryData.size());
	for (auto const& sub: m_subs)
		bytesRequiredIncludingData += static_cast<unsigned>(sub->assemble().bytecode.size());

	unsigned bytesPerDataRef = util::bytesRequired(bytesRequiredIncludingData);
	uint8_t dataRefPush = static_cast<uint8_t>(pushInstruction(bytesPerDataRef));
	ret.bytecode.reserve(bytesRequiredIncludingData);

	for (AssemblyItem const& i: m_items)
	{
		// store position of the invalid jump destination
		if (i.type() != Tag && m_tagPositionsInBytecode[0] == numeric_limits<size_t>::max())
			m_tagPositionsInBytecode[0] = ret.bytecode.size();

		switch (i.type())
		{
		case Operation:
			ret.bytecode.push_back(static_cast<uint8_t>(i.instruction()));
			break;
		case PushString:
		{
			ret.bytecode.push_back(static_cast<uint8_t>(Instruction::PUSH32));
			unsigned ii = 0;
			for (auto j: m_strings.at(h256(i.data())))
				if (++ii > 32)
					break;
				else
					ret.bytecode.push_back(uint8_t(j));
			while (ii++ < 32)
				ret.bytecode.push_back(0);
			break;
		}
		case Push:
		{
			unsigned b = max<unsigned>(1, util::bytesRequired(i.data()));
			ret.bytecode.push_back(static_cast<uint8_t>(pushInstruction(b)));
			ret.bytecode.resize(ret.bytecode.size() + b);
			bytesRef byr(&ret.bytecode.back() + 1 - b, b);
			toBigEndian(i.data(), byr);
			break;
		}
		case PushTag:
		{
			ret.bytecode.push_back(tagPush);
			tagRef[ret.bytecode.size()] = i.splitForeignPushTag();
			ret.bytecode.resize(ret.bytecode.size() + bytesPerTag);
			break;
		}
		case PushData:
			ret.bytecode.push_back(dataRefPush);
			dataRef.insert(make_pair(h256(i.data()), ret.bytecode.size()));
			ret.bytecode.resize(ret.bytecode.size() + bytesPerDataRef);
			break;
		case PushSub:
			assertThrow(i.data() <= numeric_limits<size_t>::max(), AssemblyException, "");
			ret.bytecode.push_back(dataRefPush);
			subRef.insert(make_pair(static_cast<size_t>(i.data()), ret.bytecode.size()));
			ret.bytecode.resize(ret.bytecode.size() + bytesPerDataRef);
			break;
		case PushSubSize:
		{
			assertThrow(i.data() <= numeric_limits<size_t>::max(), AssemblyException, "");
			auto s = subAssemblyById(static_cast<size_t>(i.data()))->assemble().bytecode.size();
			i.setPushedValue(u256(s));
			unsigned b = max<unsigned>(1, util::bytesRequired(s));
			ret.bytecode.push_back(static_cast<uint8_t>(pushInstruction(b)));
			ret.bytecode.resize(ret.bytecode.size() + b);
			bytesRef byr(&ret.bytecode.back() + 1 - b, b);
			toBigEndian(s, byr);
			break;
		}
		case PushProgramSize:
		{
			ret.bytecode.push_back(dataRefPush);
			sizeRef.push_back(static_cast<unsigned>(ret.bytecode.size()));
			ret.bytecode.resize(ret.bytecode.size() + bytesPerDataRef);
			break;
		}
		case PushLibraryAddress:
			ret.bytecode.push_back(static_cast<uint8_t>(Instruction::PUSH20));
			ret.linkReferences[ret.bytecode.size()] = m_libraries.at(i.data());
			ret.bytecode.resize(ret.bytecode.size() + 20);
			break;
		case PushImmutable:
			ret.bytecode.push_back(static_cast<uint8_t>(Instruction::PUSH32));
			ret.immutableReferences[i.data()].first = m_immutables.at(i.data());
			ret.immutableReferences[i.data()].second.emplace_back(ret.bytecode.size());
			ret.bytecode.resize(ret.bytecode.size() + 32);
			break;
		case VerbatimBytecode:
			ret.bytecode += i.verbatimData();
			break;
		case AssignImmutable:
		{
			auto const& offsets = immutableReferencesBySub[i.data()].second;
			for (size_t i = 0; i < offsets.size(); ++i)
			{
				if (i != offsets.size() - 1)
				{
					ret.bytecode.push_back(uint8_t(Instruction::DUP2));
					ret.bytecode.push_back(uint8_t(Instruction::DUP2));
				}
				// TODO: should we make use of the constant optimizer methods for pushing the offsets?
				bytes offsetBytes = toCompactBigEndian(u256(offsets[i]));
				ret.bytecode.push_back(static_cast<uint8_t>(pushInstruction(static_cast<unsigned>(offsetBytes.size()))));
				ret.bytecode += offsetBytes;
				ret.bytecode.push_back(uint8_t(Instruction::ADD));
				ret.bytecode.push_back(uint8_t(Instruction::MSTORE));
			}
			if (offsets.empty())
			{
				ret.bytecode.push_back(uint8_t(Instruction::POP));
				ret.bytecode.push_back(uint8_t(Instruction::POP));
			}
			immutableReferencesBySub.erase(i.data());
			break;
		}
		case PushDeployTimeAddress:
			ret.bytecode.push_back(static_cast<uint8_t>(Instruction::PUSH20));
			ret.bytecode.resize(ret.bytecode.size() + 20);
			break;
		case Tag:
		{
			assertThrow(i.data() != 0, AssemblyException, "Invalid tag position.");
			assertThrow(i.splitForeignPushTag().first == numeric_limits<size_t>::max(), AssemblyException, "Foreign tag.");
			size_t tagId = static_cast<size_t>(i.data());
			assertThrow(ret.bytecode.size() < 0xffffffffL, AssemblyException, "Tag too large.");
			assertThrow(m_tagPositionsInBytecode[tagId] == numeric_limits<size_t>::max(), AssemblyException, "Duplicate tag position.");
			m_tagPositionsInBytecode[tagId] = ret.bytecode.size();
			ret.bytecode.push_back(static_cast<uint8_t>(Instruction::JUMPDEST));
			break;
		}
		default:
			assertThrow(false, InvalidOpcode, "Unexpected opcode while assembling.");
		}
	}

	if (!immutableReferencesBySub.empty())
		throw
			langutil::Error(
				1284_error,
				langutil::Error::Type::CodeGenerationError,
				"Some immutables were read from but never assigned, possibly because of optimization."
			);

	if (!m_subs.empty() || !m_data.empty() || !m_auxiliaryData.empty())
		// Append an INVALID here to help tests find miscompilation.
		ret.bytecode.push_back(static_cast<uint8_t>(Instruction::INVALID));

	for (auto const& [subIdPath, bytecodeOffset]: subRef)
	{
		bytesRef r(ret.bytecode.data() + bytecodeOffset, bytesPerDataRef);
		toBigEndian(ret.bytecode.size(), r);
		ret.append(subAssemblyById(subIdPath)->assemble());
	}

	for (auto const& i: tagRef)
	{
		size_t subId;
		size_t tagId;
		tie(subId, tagId) = i.second;
		assertThrow(subId == numeric_limits<size_t>::max() || subId < m_subs.size(), AssemblyException, "Invalid sub id");
		vector<size_t> const& tagPositions =
			subId == numeric_limits<size_t>::max() ?
			m_tagPositionsInBytecode :
			m_subs[subId]->m_tagPositionsInBytecode;
		assertThrow(tagId < tagPositions.size(), AssemblyException, "Reference to non-existing tag.");
		size_t pos = tagPositions[tagId];
		assertThrow(pos != numeric_limits<size_t>::max(), AssemblyException, "Reference to tag without position.");
		assertThrow(util::bytesRequired(pos) <= bytesPerTag, AssemblyException, "Tag too large for reserved space.");
		bytesRef r(ret.bytecode.data() + i.first, bytesPerTag);
		toBigEndian(pos, r);
	}
	for (auto const& [name, tagInfo]: m_namedTags)
	{
		size_t position = m_tagPositionsInBytecode.at(tagInfo.id);
		ret.functionDebugData[name] = {
			position == numeric_limits<size_t>::max() ? nullopt : optional<size_t>{position},
			tagInfo.sourceID,
			tagInfo.params,
			tagInfo.returns
		};
	}

	for (auto const& dataItem: m_data)
	{
		auto references = dataRef.equal_range(dataItem.first);
		if (references.first == references.second)
			continue;
		for (auto ref = references.first; ref != references.second; ++ref)
		{
			bytesRef r(ret.bytecode.data() + ref->second, bytesPerDataRef);
			toBigEndian(ret.bytecode.size(), r);
		}
		ret.bytecode += dataItem.second;
	}

	ret.bytecode += m_auxiliaryData;

	for (unsigned pos: sizeRef)
	{
		bytesRef r(ret.bytecode.data() + pos, bytesPerDataRef);
		toBigEndian(ret.bytecode.size(), r);
	}
	return ret;
}

vector<size_t> Assembly::decodeSubPath(size_t _subObjectId) const
{
	if (_subObjectId < m_subs.size())
		return {_subObjectId};

	auto subIdPathIt = find_if(
		m_subPaths.begin(),
		m_subPaths.end(),
		[_subObjectId](auto const& subId) { return subId.second == _subObjectId; }
	);

	assertThrow(subIdPathIt != m_subPaths.end(), AssemblyException, "");
	return subIdPathIt->first;
}

size_t Assembly::encodeSubPath(vector<size_t> const& _subPath)
{
	assertThrow(!_subPath.empty(), AssemblyException, "");
	if (_subPath.size() == 1)
	{
		assertThrow(_subPath[0] < m_subs.size(), AssemblyException, "");
		return _subPath[0];
	}

	if (m_subPaths.find(_subPath) == m_subPaths.end())
	{
		size_t objectId = numeric_limits<size_t>::max() - m_subPaths.size();
		assertThrow(objectId >= m_subs.size(), AssemblyException, "");
		m_subPaths[_subPath] = objectId;
	}

	return m_subPaths[_subPath];
}

Assembly const* Assembly::subAssemblyById(size_t _subId) const
{
	vector<size_t> subIds = decodeSubPath(_subId);
	Assembly const* currentAssembly = this;
	for (size_t currentSubId: subIds)
	{
		currentAssembly = currentAssembly->m_subs.at(currentSubId).get();
		assertThrow(currentAssembly, AssemblyException, "");
	}

	assertThrow(currentAssembly != this, AssemblyException, "");
	return currentAssembly;
}