Merge pull request #6719 from ethereum/breakingLoops

Breaking loops

Changelog.md

@@ -10,6 +10,7 @@ Compiler Features:
  * SMTChecker: Support ``delete``.
  * SMTChecker: Inline external function calls to ``this``.
  * Assembler: Encode the compiler version in the deployed bytecode.
+ * Yul Optimizer: Simplify single-run ``for`` loops to ``if`` statements.
 
 
 Bugfixes:

libevmasm/SemanticInformation.cpp

@@ -136,7 +136,13 @@ bool SemanticInformation::terminatesControlFlow(AssemblyItem const& _item)
 {
 	if (_item.type() != Operation)
 		return false;
-	switch (_item.instruction())
+	else
+		return terminatesControlFlow(_item.instruction());
+}
+
+bool SemanticInformation::terminatesControlFlow(Instruction _instruction)
+{
+	switch (_instruction)
 	{
 	case Instruction::RETURN:
 	case Instruction::SELFDESTRUCT:

libevmasm/SemanticInformation.h

@@ -48,6 +48,7 @@ struct SemanticInformation
 	static bool isJumpInstruction(AssemblyItem const& _item);
 	static bool altersControlFlow(AssemblyItem const& _item);
 	static bool terminatesControlFlow(AssemblyItem const& _item);
+	static bool terminatesControlFlow(Instruction _instruction);
 	/// @returns false if the value put on the stack by _item depends on anything else than
 	/// the information in the current block header, memory, storage or stack.
 	static bool isDeterministic(AssemblyItem const& _item);

libyul/optimiser/ControlFlowSimplifier.cpp

@@ -115,9 +115,51 @@ void ControlFlowSimplifier::operator()(Block& _block)
 	simplify(_block.statements);
 }
 
+void ControlFlowSimplifier::visit(Statement& _st)
+{
+	if (_st.type() == typeid(ForLoop))
+	{
+		ForLoop& forLoop = boost::get<ForLoop>(_st);
+		yulAssert(forLoop.pre.statements.empty(), "");
+
+		size_t outerBreak = m_numBreakStatements;
+		size_t outerContinue = m_numContinueStatements;
+		m_numBreakStatements = 0;
+		m_numContinueStatements = 0;
+
+		ASTModifier::visit(_st);
+
+		if (!forLoop.body.statements.empty())
+		{
+			bool isTerminating = false;
+			TerminationFinder::ControlFlow controlFlow = TerminationFinder::controlFlowKind(forLoop.body.statements.back());
+			if (controlFlow == TerminationFinder::ControlFlow::Break)
+			{
+				isTerminating = true;
+				--m_numBreakStatements;
+			}
+			else if (controlFlow == TerminationFinder::ControlFlow::Terminate)
+				isTerminating = true;
+
+			if (isTerminating && m_numContinueStatements == 0 && m_numBreakStatements == 0)
+			{
+				If replacement{forLoop.location, std::move(forLoop.condition), std::move(forLoop.body)};
+				if (controlFlow == TerminationFinder::ControlFlow::Break)
+					replacement.body.statements.resize(replacement.body.statements.size() - 1);
+				_st = std::move(replacement);
+			}
+		}
+
+		m_numBreakStatements = outerBreak;
+		m_numContinueStatements = outerContinue;
+	}
+	else
+		ASTModifier::visit(_st);
+}
+
 void ControlFlowSimplifier::simplify(std::vector<yul::Statement>& _statements)
 {
-	GenericFallbackReturnsVisitor<OptionalStatements, If, Switch, ForLoop> const visitor(
+	GenericFallbackReturnsVisitor<OptionalStatements, If, Switch> const visitor(
 		[&](If& _ifStmt) -> OptionalStatements {
 			if (_ifStmt.body.statements.empty())
 			{
@@ -136,9 +178,6 @@ void ControlFlowSimplifier::simplify(std::vector<yul::Statement>& _statements)
 			else if (_switchStmt.cases.size() == 1)
 				return reduceSingleCaseSwitch(_switchStmt);
 
-			return {};
-		},
-		[&](ForLoop&) -> OptionalStatements {
 			return {};
 		}
 	);

libyul/optimiser/ControlFlowSimplifier.h

@@ -38,7 +38,7 @@ namespace yul
  * The ControlFlowSimplifier does record the presence or absence of ``break``
  * and ``continue`` statements during its traversal.
  *
- * Prerequisite: Disambiguator, ForLoopInitRewriter.
+ * Prerequisite: Disambiguator, FunctionHoister, ForLoopInitRewriter.
  *
  * Important: Introduces EVM opcodes and thus can only be used on EVM code for now.
  */
@@ -46,9 +46,17 @@ class ControlFlowSimplifier: public ASTModifier
 {
 public:
 	using ASTModifier::operator();
+	void operator()(Break&) override { ++m_numBreakStatements; }
+	void operator()(Continue&) override { ++m_numContinueStatements; }
 	void operator()(Block& _block) override;
+
+	void visit(Statement& _st) override;
+
 private:
 	void simplify(std::vector<Statement>& _statements);
+
+	size_t m_numBreakStatements = 0;
+	size_t m_numContinueStatements = 0;
 };
 
 }

libyul/optimiser/DeadCodeEliminator.cpp

@@ -19,6 +19,7 @@
  */
 
 #include <libyul/optimiser/DeadCodeEliminator.h>
+#include <libyul/optimiser/Semantics.h>
 #include <libyul/AsmData.h>
 
 #include <libevmasm/SemanticInformation.h>
@@ -30,42 +31,6 @@ using namespace std;
 using namespace dev;
 using namespace yul;
 
-namespace
-{
-bool isTerminating(yul::ExpressionStatement const& _exprStmnt)
-{
-	if (_exprStmnt.expression.type() != typeid(FunctionalInstruction))
-		return false;
-
-	auto const& funcInstr = boost::get<FunctionalInstruction>(_exprStmnt.expression);
-
-	return eth::SemanticInformation::terminatesControlFlow(funcInstr.instruction);
-}
-
-/// Returns an iterator to the first terminating statement or
-/// `_block.statements.end()()` when none was found
-auto findFirstTerminatingStatement(Block& _block)
-{
-	return find_if(
-		_block.statements.begin(),
-		_block.statements.end(),
-		[](Statement const& _stmnt)
-		{
-			if (
-				_stmnt.type() == typeid(ExpressionStatement) &&
-				isTerminating(boost::get<ExpressionStatement>(_stmnt))
-			)
-				return true;
-			else if (_stmnt.type() == typeid(Break))
-				return true;
-			else if (_stmnt.type() == typeid(Continue))
-				return true;
-
-			return false;
-		}
-	);
-}
-}
 
 void DeadCodeEliminator::operator()(ForLoop& _for)
 {
@@ -75,24 +40,19 @@ void DeadCodeEliminator::operator()(ForLoop& _for)
 
 void DeadCodeEliminator::operator()(Block& _block)
 {
-	auto& statements = _block.statements;
+	TerminationFinder::ControlFlow controlFlowChange;
+	size_t index;
+	tie(controlFlowChange, index) = TerminationFinder::firstUnconditionalControlFlowChange(_block.statements);
 
-	auto firstTerminatingStatment = findFirstTerminatingStatement(_block);
-
-	if (
-		firstTerminatingStatment != statements.end() &&
-		firstTerminatingStatment + 1 != statements.end()
-	)
-		statements.erase(
-			std::remove_if(
-				firstTerminatingStatment + 1,
-				statements.end(),
-				[] (Statement const& _s)
-				{
-					return _s.type() != typeid(yul::FunctionDefinition);
-				}
+	// Erase everything after the terminating statement that is not a function definition.
+	if (controlFlowChange != TerminationFinder::ControlFlow::FlowOut && index != size_t(-1))
+		_block.statements.erase(
+			remove_if(
+				_block.statements.begin() + index + 1,
+				_block.statements.end(),
+				[] (Statement const& _s) { return _s.type() != typeid(yul::FunctionDefinition); }
 			),
-			statements.end()
+			_block.statements.end()
 		);
 
 	ASTModifier::operator()(_block);

libyul/optimiser/DeadCodeEliminator.h

@@ -41,7 +41,7 @@ namespace yul
  * Because variables declared in a for loop's init block have their scope extended to the loop body,
  * we require ForLoopInitRewriter to run before this step.
  *
- * Prerequisite: ForLoopInitRewriter
+ * Prerequisite: ForLoopInitRewriter, Function Hoister, Function Grouper
  */
 class DeadCodeEliminator: public ASTModifier
 {

libyul/optimiser/Semantics.cpp

@@ -72,3 +72,41 @@ void MovableChecker::visit(Statement const&)
 {
 	assertThrow(false, OptimizerException, "Movability for statement requested.");
 }
+
+pair<TerminationFinder::ControlFlow, size_t> TerminationFinder::firstUnconditionalControlFlowChange(
+	vector<Statement> const& _statements
+)
+{
+	for (size_t i = 0; i < _statements.size(); ++i)
+	{
+		ControlFlow controlFlow = controlFlowKind(_statements[i]);
+		if (controlFlow != ControlFlow::FlowOut)
+			return {controlFlow, i};
+	}
+	return {ControlFlow::FlowOut, size_t(-1)};
+}
+
+TerminationFinder::ControlFlow TerminationFinder::controlFlowKind(Statement const& _statement)
+{
+	if (
+		_statement.type() == typeid(ExpressionStatement) &&
+		isTerminatingBuiltin(boost::get<ExpressionStatement>(_statement))
+	)
+		return ControlFlow::Terminate;
+	else if (_statement.type() == typeid(Break))
+		return ControlFlow::Break;
+	else if (_statement.type() == typeid(Continue))
+		return ControlFlow::Continue;
+	else
+		return ControlFlow::FlowOut;
+}
+
+bool TerminationFinder::isTerminatingBuiltin(ExpressionStatement const& _exprStmnt)
+{
+	if (_exprStmnt.expression.type() != typeid(FunctionalInstruction))
+		return false;
+
+	return eth::SemanticInformation::terminatesControlFlow(
+		boost::get<FunctionalInstruction>(_exprStmnt.expression).instruction
+	);
+}

libyul/optimiser/Semantics.h

@@ -56,4 +56,34 @@ private:
 	bool m_movable = true;
 };
 
+/**
+ * Helper class to find "irregular" control flow.
+ * This includes termination, break and continue.
+ */
+class TerminationFinder
+{
+public:
+	enum class ControlFlow { FlowOut, Break, Continue, Terminate };
+
+	/// @returns the index of the first statement in the provided sequence
+	/// that is an unconditional ``break``, ``continue`` or a
+	/// call to a terminating builtin function.
+	/// If control flow can continue at the end of the list,
+	/// returns `FlowOut` and ``size_t(-1)``.
+	/// The function might return ``FlowOut`` even though control
+	/// flow cannot actually continue.
+	static std::pair<ControlFlow, size_t> firstUnconditionalControlFlowChange(
+		std::vector<Statement> const& _statements
+	);
+
+	/// @returns the control flow type of the given statement.
+	/// This function could return FlowOut even if control flow never continues.
+	static ControlFlow controlFlowKind(Statement const& _statement);
+
+	/// @returns true if the expression statement is a direct
+	/// call to a builtin terminating function like
+	/// ``stop``, ``revert`` or ``return``.
+	static bool isTerminatingBuiltin(ExpressionStatement const& _exprStmnt);
+};
+
 }

test/libyul/yulOptimizerTests/controlFlowSimplifier/terminating_for.yul

@@ -0,0 +1,12 @@
+{
+	for {} calldatasize() { mstore(1, 2) } {
+		mstore(4, 5)
+		break
+	}
+}
+// ====
+// step: controlFlowSimplifier
+// ----
+// {
+//     if calldatasize() { mstore(4, 5) }
+// }

test/libyul/yulOptimizerTests/controlFlowSimplifier/terminating_for_nested.yul

@@ -0,0 +1,21 @@
+{
+    for {} calldatasize() { mstore(8, 9) } {
+        for {} calldatasize() { mstore(1, 2) } {
+            mstore(4, 5)
+            break
+        }
+        if mload(10) { continue }
+        break
+    }
+}
+// ====
+// step: controlFlowSimplifier
+// ----
+// {
+//     for { } calldatasize() { mstore(8, 9) }
+//     {
+//         if calldatasize() { mstore(4, 5) }
+//         if mload(10) { continue }
+//         break
+//     }
+// }

test/libyul/yulOptimizerTests/controlFlowSimplifier/terminating_for_nested_reversed.yul

@@ -0,0 +1,22 @@
+{
+    for {} calldatasize() { mstore(8, 9) } {
+        for {} calldatasize() { mstore(1, 2) } {
+            mstore(4, 5)
+            continue
+        }
+        break
+    }
+}
+// ====
+// step: controlFlowSimplifier
+// ----
+// {
+//     if calldatasize()
+//     {
+//         for { } calldatasize() { mstore(1, 2) }
+//         {
+//             mstore(4, 5)
+//             continue
+//         }
+//     }
+// }

test/libyul/yulOptimizerTests/controlFlowSimplifier/terminating_for_revert.yul

@@ -0,0 +1,18 @@
+{
+	for {} calldatasize() { mstore(1, 2) } {
+		let x := 7
+		mstore(4, 5)
+		revert(0, x)
+	}
+}
+// ====
+// step: controlFlowSimplifier
+// ----
+// {
+//     if calldatasize()
+//     {
+//         let x := 7
+//         mstore(4, 5)
+//         revert(0, x)
+//     }
+// }

test/libyul/yulOptimizerTests/controlFlowSimplifier/terminating_for_revert_plus_break.yul

@@ -0,0 +1,20 @@
+{
+	for {} calldatasize() { mstore(1, 2) } {
+		let x := 7
+		mstore(4, 5)
+		break
+		revert(0, x)
+	}
+}
+// ====
+// step: controlFlowSimplifier
+// ----
+// {
+//     for { } calldatasize() { mstore(1, 2) }
+//     {
+//         let x := 7
+//         mstore(4, 5)
+//         break
+//         revert(0, x)
+//     }
+// }

test/libyul/yulOptimizerTests/controlFlowSimplifier/terminating_for_with_continue.yul

@@ -0,0 +1,18 @@
+{
+	for {} calldatasize() { mstore(1, 2) } {
+		if calldatasize() { continue }
+		mstore(4, 5)
+		break
+	}
+}
+// ====
+// step: controlFlowSimplifier
+// ----
+// {
+//     for { } calldatasize() { mstore(1, 2) }
+//     {
+//         if calldatasize() { continue }
+//         mstore(4, 5)
+//         break
+//     }
+// }