Tune Rematerializer

libyul/optimiser/DataFlowAnalyzer.cpp

@@ -145,10 +145,14 @@ void DataFlowAnalyzer::operator()(FunctionDefinition& _fun)
 	// Save all information. We might rather reinstantiate this class,
 	// but this could be difficult if it is subclassed.
 	map<YulString, Expression const*> value;
+	map<YulString, size_t> variableLoopDepth;
+	size_t loopDepth{0};
 	InvertibleRelation<YulString> references;
 	InvertibleMap<YulString, YulString> storage;
 	InvertibleMap<YulString, YulString> memory;
-	m_value.swap(value);
+	swap(m_value, value);
+	swap(m_variableLoopDepth, variableLoopDepth);
+	swap(m_loopDepth, loopDepth);
 	swap(m_references, references);
 	swap(m_storage, storage);
 	swap(m_memory, memory);
@@ -168,7 +172,9 @@ void DataFlowAnalyzer::operator()(FunctionDefinition& _fun)
 	// statement.
 
 	popScope();
-	m_value.swap(value);
+	swap(m_value, value);
+	swap(m_variableLoopDepth, variableLoopDepth);
+	swap(m_loopDepth, loopDepth);
 	swap(m_references, references);
 	swap(m_storage, storage);
 	swap(m_memory, memory);
@@ -180,6 +186,8 @@ void DataFlowAnalyzer::operator()(ForLoop& _for)
 	// we would have to deal with more complicated scoping rules.
 	assertThrow(_for.pre.statements.empty(), OptimizerException, "");
 
+	++m_loopDepth;
+
 	AssignmentsSinceContinue assignmentsSinceCont;
 	assignmentsSinceCont(_for.body);
 
@@ -202,6 +210,8 @@ void DataFlowAnalyzer::operator()(ForLoop& _for)
 	clearKnowledgeIfInvalidated(*_for.condition);
 	clearKnowledgeIfInvalidated(_for.post);
 	clearKnowledgeIfInvalidated(_for.body);
+
+	--m_loopDepth;
 }
 
 void DataFlowAnalyzer::operator()(Block& _block)
@@ -222,7 +232,7 @@ void DataFlowAnalyzer::handleAssignment(set<YulString> const& _variables, Expres
 		movableChecker.visit(*_value);
 	else
 		for (auto const& var: _variables)
-			m_value[var] = &m_zero;
+			assignValue(var, &m_zero);
 
 	if (_value && _variables.size() == 1)
 	{
@@ -230,7 +240,7 @@ void DataFlowAnalyzer::handleAssignment(set<YulString> const& _variables, Expres
 		// Expression has to be movable and cannot contain a reference
 		// to the variable that will be assigned to.
 		if (movableChecker.movable() && !movableChecker.referencedVariables().count(name))
-			m_value[name] = _value;
+			assignValue(name, _value);
 	}
 
 	auto const& referencedVariables = movableChecker.referencedVariables();
@@ -296,11 +306,20 @@ void DataFlowAnalyzer::clearValues(set<YulString> _variables)
 
 	// Clear the value and update the reference relation.
 	for (auto const& name: _variables)
+	{
 		m_value.erase(name);
+		m_variableLoopDepth.erase(name);
+	}
 	for (auto const& name: _variables)
 		m_references.eraseKey(name);
 }
 
+void DataFlowAnalyzer::assignValue(YulString _variable, Expression const* _value)
+{
+	m_value[_variable] = _value;
+	m_variableLoopDepth[_variable] = m_loopDepth;
+}
+
 void DataFlowAnalyzer::clearKnowledgeIfInvalidated(Block const& _block)
 {
 	SideEffectsCollector sideEffects(m_dialect, _block, &m_functionSideEffects);

libyul/optimiser/DataFlowAnalyzer.h

@@ -110,6 +110,8 @@ protected:
 	/// for example at points where control flow is merged.
 	void clearValues(std::set<YulString> _names);
 
+	void assignValue(YulString _variable, Expression const* _value);
+
 	/// Clears knowledge about storage or memory if they may be modified inside the block.
 	void clearKnowledgeIfInvalidated(Block const& _block);
 
@@ -144,6 +146,8 @@ protected:
 
 	/// Current values of variables, always movable.
 	std::map<YulString, Expression const*> m_value;
+	/// The loop nesting depth of the definition of variables (those used in m_value).
+	std::map<YulString, size_t> m_variableLoopDepth;
 	/// m_references.forward[a].contains(b) <=> the current expression assigned to a references b
 	/// m_references.backward[b].contains(a) <=> the current expression assigned to a references b
 	InvertibleRelation<YulString> m_references;
@@ -153,6 +157,9 @@ protected:
 
 	KnowledgeBase m_knowledgeBase;
 
+	/// Current nesting depth of loops.
+	size_t m_loopDepth{0};
+
 	struct Scope
 	{
 		explicit Scope(bool _isFunction): isFunction(_isFunction) {}

libyul/optimiser/Rematerialiser.cpp

@@ -78,7 +78,12 @@ void Rematerialiser::visit(Expression& _e)
 			auto const& value = *m_value.at(name);
 			size_t refs = m_referenceCounts[name];
 			size_t cost = CodeCost::codeCost(m_dialect, value);
-			if (refs <= 1 || cost == 0 || (refs <= 5 && cost <= 1) || m_varsToAlwaysRematerialize.count(name))
+			if (
+				(refs <= 1 && m_variableLoopDepth.at(name) == m_loopDepth) ||
+				cost == 0 ||
+				(refs <= 5 && cost <= 1) ||
+				m_varsToAlwaysRematerialize.count(name)
+			)
 			{
 				assertThrow(m_referenceCounts[name] > 0, OptimizerException, "");
 				for (auto const& ref: m_references.forward[name])

libyul/optimiser/Rematerialiser.h

@@ -27,9 +27,10 @@ namespace solidity::yul
 {
 
 /**
- * Optimisation stage that replaces variables by their most recently assigned expressions,
+ * Optimisation stage that replaces variable references by those expressions
+ * that are most recently assigned to the referenced variables,
  * but only if the expression is movable and one of the following holds:
- *  - the variable is referenced exactly once
+ *  - the variable is referenced exactly once (and definition-to-reference does not cross a loop boundary)
  *  - the value is extremely cheap ("cost" of zero like ``caller()``)
  *  - the variable is referenced at most 5 times and the value is rather cheap
  *    ("cost" of at most 1 like a constant up to 0xff)
@@ -68,6 +69,8 @@ protected:
 		std::set<YulString> _varsToAlwaysRematerialize = {}
 	);
 
+	using DataFlowAnalyzer::operator();
+
 	using ASTModifier::visit;
 	void visit(Expression& _e) override;
 

test/libyul/yulOptimizerTests/fullSuite/clear_after_if_continue.yul

@@ -12,7 +12,8 @@
 // {
 //     {
 //         let y := mload(0x20)
-//         for { } and(y, 8) { if y { revert(0, 0) } }
+//         let _1 := iszero(and(y, 8))
+//         for { } iszero(_1) { if y { revert(0, 0) } }
 //         {
 //             if y { continue }
 //             sstore(1, 0)

test/libyul/yulOptimizerTests/fullSuite/devcon_example.yul

@@ -21,8 +21,9 @@
 //     {
 //         let _1 := calldataload(0)
 //         let sum := 0
+//         let length := calldataload(_1)
 //         let i := sum
-//         for { } lt(i, calldataload(_1)) { i := add(i, 1) }
+//         for { } lt(i, length) { i := add(i, 1) }
 //         {
 //             sum := add(sum, calldataload(add(add(_1, mul(i, 0x20)), 0x20)))
 //         }

test/libyul/yulOptimizerTests/fullSuite/loopInvariantCodeMotion.yul

@@ -24,10 +24,12 @@
 //     {
 //         let _1 := calldataload(0)
 //         let sum := 0
+//         let length := calldataload(_1)
 //         let i := sum
-//         for { } lt(i, calldataload(_1)) { i := add(i, 1) }
+//         let _2 := calldataload(7)
+//         for { } lt(i, length) { i := add(i, 1) }
 //         {
-//             sum := add(sum, add(calldataload(add(add(_1, mul(i, 0x20)), 0x20)), calldataload(7)))
+//             sum := add(sum, add(calldataload(add(add(_1, mul(i, 0x20)), 0x20)), _2))
 //         }
 //         sstore(0, sum)
 //     }

test/libyul/yulOptimizerTests/fullSuite/no_move_loop_orig.yul

@@ -12,10 +12,11 @@
 // ----
 // {
 //     {
+//         let _1 := iszero(caller())
 //         for { }
 //         1
 //         {
-//             for { } caller() { }
+//             for { } iszero(_1) { }
 //             { }
 //             mstore(192, 0)
 //         }

test/libyul/yulOptimizerTests/rematerialiser/many_refs_small_cost_loop.yul

@@ -0,0 +1,28 @@
+{
+	// Cost of rematerializating x is 1
+	let x := 0xff
+	// Although x has low cost, it is not considered for
+	// rematerialization because it is referenced more than 5 times
+	for {} lt(x, 0x100) {}
+	{
+		let y := add(x, 1)
+		let z := mul(x, 1)
+		let a := div(x, 2)
+		let b := mod(x, 3)
+		let c := sdiv(x, 4)
+	}
+}
+// ====
+// step: rematerialiser
+// ----
+// {
+//     let x := 0xff
+//     for { } lt(x, 0x100) { }
+//     {
+//         let y := add(x, 1)
+//         let z := mul(x, 1)
+//         let a := div(x, 2)
+//         let b := mod(x, 3)
+//         let c := sdiv(x, 4)
+//     }
+// }

test/libyul/yulOptimizerTests/rematerialiser/no_remat_in_loop.yul

@@ -0,0 +1,37 @@
+{
+	// origin has zero cost and thus will be rematerialised,
+	// calldataload(0) has low cost and will not be rematerialised
+	let a := origin()
+	let b := calldataload(0)
+	let i := 0
+	let z := calldataload(9)
+	for {} lt(i, 10) {i := add(a, b)} {
+		// This will be rematerialised, because it stays inside
+		// the loop.
+		let x := calldataload(1)
+		mstore(9, x)
+		// No, because again one loop further down.
+		let y := calldataload(2)
+		for {} y {} {
+			// Again no.
+			mstore(12, z)
+		}
+	}
+}
+// ====
+// step: rematerialiser
+// ----
+// {
+//     let a := origin()
+//     let b := calldataload(0)
+//     let i := 0
+//     let z := calldataload(9)
+//     for { } lt(i, 10) { i := add(origin(), b) }
+//     {
+//         let x := calldataload(1)
+//         mstore(9, calldataload(1))
+//         let y := calldataload(2)
+//         for { } y { }
+//         { mstore(12, z) }
+//     }
+// }

test/libyul/yulOptimizerTests/rematerialiser/some_refs_small_cost_loop.yul

@@ -0,0 +1,17 @@
+{
+	// Cost of rematerializating x is 1
+	let x := 0xff
+	// Reference to x is not rematerialized because the reference is in a loop
+	for {} lt(x, 0x100) {}
+	{
+		let y := add(x, 1)
+	}
+}
+// ====
+// step: rematerialiser
+// ----
+// {
+//     let x := 0xff
+//     for { } lt(0xff, 0x100) { }
+//     { let y := add(0xff, 1) }
+// }

test/libyul/yulOptimizerTests/rematerialiser/some_refs_small_cost_nested_loop.yul

@@ -0,0 +1,26 @@
+{
+	// Cost of rematerializating x is 1
+	let x := 0xff
+	// Although x has a low cost and fewer than 6 references,
+	// its references in a loop are not rematerialized
+	for {} lt(x, 0x100) {}
+	{
+		let y := add(x, 1)
+		for {} lt(x, 0x200) {}
+		{
+			let z := mul(x, 2)
+		}
+	}
+}
+// ====
+// step: rematerialiser
+// ----
+// {
+//     let x := 0xff
+//     for { } lt(0xff, 0x100) { }
+//     {
+//         let y := add(0xff, 1)
+//         for { } lt(0xff, 0x200) { }
+//         { let z := mul(0xff, 2) }
+//     }
+// }