Pinhole optimise working fairly well...

Assembly.cpp

@@ -186,6 +186,9 @@ inline bool matches(AssemblyItemsConstRef _a, AssemblyItemsConstRef _b)
 	return true;
 }
 
+struct OptimiserChannel: public LogChannel { static const char* name() { return "OPT"; } static const int verbosity = 12; };
+#define copt eth::LogOutputStream<OptimiserChannel, true>()
+
 void Assembly::optimise()
 {
 	map<Instruction, function<u256(u256, u256)>> c_simple =
@@ -210,7 +213,7 @@ void Assembly::optimise()
 	std::vector<pair<AssemblyItems, function<AssemblyItems(AssemblyItemsConstRef)>>> rules =
 	{
 		{ { Push, Instruction::POP }, [](AssemblyItemsConstRef) -> AssemblyItems { return {}; } },
-		{ { Push, PushTag, Instruction::JUMPI }, [](AssemblyItemsConstRef m) -> AssemblyItems { return m[0].data() ? AssemblyItems({ m[1], Instruction::JUMP }) : AssemblyItems(); } },
+		{ { Push, PushTag, Instruction::JUMPI }, [](AssemblyItemsConstRef m) -> AssemblyItems { if (m[0].data()) return { m[1], Instruction::JUMP }; else return {}; } },
 	};
 
 	for (auto const& i: c_simple)
@@ -219,19 +222,16 @@ void Assembly::optimise()
 	{
 		rules.push_back({ { Push, Push, i.first }, [&](AssemblyItemsConstRef m) -> AssemblyItems { return { i.second(m[1].data(), m[0].data()) }; } });
 		rules.push_back({ { Push, i.first, Push, i.first }, [&](AssemblyItemsConstRef m) -> AssemblyItems { return { i.second(m[2].data(), m[0].data()), i.first }; } });
-		rules.push_back({ { PushTag, Instruction::JUMP, Tag }, [&](AssemblyItemsConstRef m) -> AssemblyItems
+		rules.push_back({ { PushTag, Instruction::JUMP, Tag }, [&](AssemblyItemsConstRef m) -> AssemblyItems { if (m[0].m_data == m[2].m_data) return {}; else return m.toVector(); }});
-						{
-							if (m[0].m_data == m[2].m_data)
-								return {};
-							else
-								return m.toVector();
-						}});
 	}
 
+	copt << *this;
+
 	unsigned total = 0;
 	for (unsigned count = 1; count > 0; total += count)
 	{
 		count = 0;
+		map<u256, unsigned> tags;
 		for (unsigned i = 0; i < m_items.size(); ++i)
 		{
 			for (auto const& r: rules)
@@ -242,23 +242,64 @@ void Assembly::optimise()
 					auto rw = r.second(vr);
 					if (rw.size() < vr.size())
 					{
-						cnote << vr << "matches" << AssemblyItemsConstRef(&r.first) << "becomes...";
+						copt << vr << "matches" << AssemblyItemsConstRef(&r.first) << "becomes...";
 						for (unsigned j = 0; j < vr.size(); ++j)
 							if (j < rw.size())
 								m_items[i + j] = rw[j];
 							else
 								m_items.erase(m_items.begin() + i + rw.size());
-						cnote << AssemblyItemsConstRef(&rw);
+						copt << AssemblyItemsConstRef(&rw);
 						count++;
+						copt << "Now:\n" << m_items;
 					}
 				}
 			}
+			if (m_items[i].type() == Operation && m_items[i].data() == (byte)Instruction::JUMP)
+			{
+				bool o = false;
+				while (m_items.size() > i + 1 && m_items[i + 1].type() != Tag)
+				{
+					m_items.erase(m_items.begin() + i + 1);
+					o = true;
+				}
+				if (o)
+				{
+					copt << "Jump with no tag. Now:\n" << m_items;
+					++count;
+				}
 			}
 		}
 
-	// TODO: find all unused tags, for all those that have an unconditional jump immediately before, remove code between the tag and the next used tag (removing unused tags from the todo along the way).
+		for (unsigned i = 0; i < m_items.size(); ++i)
+			if (m_items[i].type() == Tag)
+				tags.insert(make_pair(m_items[i].data(), i));
 
-	cnote << total << " optimisations done.";
+		for (auto const& i: m_items)
+			if (i.type() == PushTag)
+				tags.erase(i.data());
+
+		if (tags.size())
+		{
+			auto t = *tags.begin();
+			unsigned i = t.second;
+			if (i && m_items[i - 1].type() == Operation && m_items[i - 1].data() == (byte)Instruction::JUMP)
+				while (i < m_items.size() && (m_items[i].type() != Tag || tags.count(m_items[i].data())))
+				{
+					if (m_items[i].type() == Tag && tags.count(m_items[i].data()))
+						tags.erase(m_items[i].data());
+					m_items.erase(m_items.begin() + i);
+				}
+			else
+			{
+				m_items.erase(m_items.begin() + i);
+				tags.erase(t.first);
+			}
+			copt << "Unused tag. Now:\n" << m_items;
+			++count;
+		}
+	}
+
+	copt << total << " optimisations done.";
 }
 
 bytes Assembly::assemble() const
@@ -333,6 +374,8 @@ bytes Assembly::assemble() const
 		for (auto const& i: m_data)
 		{
 			auto its = dataRef.equal_range(i.first);
+			if (its.first != its.second)
+			{
 				for (auto it = its.first; it != its.second; ++it)
 				{
 					bytesRef r(ret.data() + it->second, bytesPerTag);
@@ -342,5 +385,6 @@ bytes Assembly::assemble() const
 					ret.push_back(b);
 			}
 		}
+	}
 	return ret;
 }

Assembly.h

@@ -62,6 +62,7 @@ typedef std::vector<AssemblyItem> AssemblyItems;
 typedef vector_ref<AssemblyItem const> AssemblyItemsConstRef;
 
 std::ostream& operator<<(std::ostream& _out, AssemblyItemsConstRef _i);
+inline std::ostream& operator<<(std::ostream& _out, AssemblyItems const& _i) { return operator<<(_out, AssemblyItemsConstRef(&_i)); }
 
 class Assembly
 {

CodeFragment.h

@@ -43,10 +43,13 @@ public:
 	static CodeFragment compile(std::string const& _src, CompilerState& _s);
 
 	/// Consolidates data and compiles code.
-	bytes code() { m_asm.optimise(); return m_asm.assemble(); }
+	bytes code() const { return m_asm.assemble(); }
 
 	/// Consolidates data and compiles code.
-	std::string assembly() { m_asm.optimise(); return m_asm.out(); }
+	std::string assembly() const { return m_asm.out(); }
+
+	/// Optimise the code. Best do this just before calling code() or assembly().
+	void optimise() { m_asm.optimise(); }
 
 private:
 	template <class T> void error() const { throw T(); }

Compiler.cpp

@@ -27,12 +27,15 @@
 using namespace std;
 using namespace eth;
 
-bytes eth::compileLLL(string const& _src, vector<string>* _errors)
+bytes eth::compileLLL(string const& _src, bool _opt, vector<string>* _errors)
 {
 	try
 	{
 		CompilerState cs;
-		bytes ret = CodeFragment::compile(_src, cs).code();
+		auto f = CodeFragment::compile(_src, cs);
+		if (_opt)
+			f.optimise();
+		bytes ret = f.code();
 		for (auto i: cs.treesToKill)
 			killBigints(i);
 		return ret;
@@ -50,12 +53,15 @@ bytes eth::compileLLL(string const& _src, vector<string>* _errors)
 	return bytes();
 }
 
-std::string eth::compileLLLToAsm(std::string const& _src, std::vector<std::string>* _errors)
+std::string eth::compileLLLToAsm(std::string const& _src, bool _opt, std::vector<std::string>* _errors)
 {
 	try
 	{
 		CompilerState cs;
-		string ret = CodeFragment::compile(_src, cs).assembly();
+		auto f = CodeFragment::compile(_src, cs);
+		if (_opt)
+			f.optimise();
+		string ret = f.assembly();
 		for (auto i: cs.treesToKill)
 			killBigints(i);
 		return ret;

Compiler.h

@@ -29,8 +29,8 @@ namespace eth
 {
 
 std::string parseLLL(std::string const& _src);
-std::string compileLLLToAsm(std::string const& _src, std::vector<std::string>* _errors = nullptr);
+std::string compileLLLToAsm(std::string const& _src, bool _opt = true, std::vector<std::string>* _errors = nullptr);
-bytes compileLLL(std::string const& _src, std::vector<std::string>* _errors = nullptr);
+bytes compileLLL(std::string const& _src, bool _opt = true, std::vector<std::string>* _errors = nullptr);
 
 }