Merge pull request #12091 from ethereum/applyControlFlowSideEffectsUserDefined

Use side effects of user-defined functions in other optimizer steps.

Changelog.md

@@ -12,6 +12,7 @@ Compiler Features:
  * SMTChecker: Report contract invariants and reentrancy properties. This can be enabled via the CLI option ``--model-checker-invariants`` or the Standard JSON option ``settings.modelChecker.invariants``.
  * Standard JSON: Accept nested brackets in step sequences passed to ``settings.optimizer.details.yulDetails.optimizerSteps``.
  * Standard JSON: Add ``settings.debug.debugInfo`` option for selecting how much extra debug information should be included in the produced EVM assembly and Yul code.
+ * Yul Optimizer: Take control-flow side-effects of user-defined functions into account in various optimizer steps.
 
 
 Bugfixes:

libyul/optimiser/ConditionalSimplifier.cpp

@@ -19,6 +19,7 @@
 #include <libyul/optimiser/Semantics.h>
 #include <libyul/AST.h>
 #include <libyul/optimiser/NameCollector.h>
+#include <libyul/ControlFlowSideEffectsCollector.h>
 #include <libsolutil/CommonData.h>
 
 using namespace std;
@@ -26,6 +27,12 @@ using namespace solidity;
 using namespace solidity::yul;
 using namespace solidity::util;
 
+void ConditionalSimplifier::run(OptimiserStepContext& _context, Block& _ast)
+{
+	ControlFlowSideEffectsCollector sideEffects(_context.dialect, _ast);
+	ConditionalSimplifier{_context.dialect, sideEffects.functionSideEffects()}(_ast);
+}
+
 void ConditionalSimplifier::operator()(Switch& _switch)
 {
 	visit(*_switch.expression);
@@ -65,7 +72,7 @@ void ConditionalSimplifier::operator()(Block& _block)
 				if (
 					holds_alternative<Identifier>(*_if.condition) &&
 					!_if.body.statements.empty() &&
-					TerminationFinder(m_dialect).controlFlowKind(_if.body.statements.back()) !=
+					TerminationFinder(m_dialect, &m_functionSideEffects).controlFlowKind(_if.body.statements.back()) !=
 						TerminationFinder::ControlFlow::FlowOut
 				)
 				{

libyul/optimiser/ConditionalSimplifier.h

@@ -44,7 +44,6 @@ namespace solidity::yul
  *
  * Future features:
  *  - allow replacements by "1"
- *  - take termination of user-defined functions into account
  *
  * Works best with SSA form and if dead code removal has run before.
  *
@@ -54,20 +53,21 @@ class ConditionalSimplifier: public ASTModifier
 {
 public:
 	static constexpr char const* name{"ConditionalSimplifier"};
-	static void run(OptimiserStepContext& _context, Block& _ast)
+	static void run(OptimiserStepContext& _context, Block& _ast);
-	{
-		ConditionalSimplifier{_context.dialect}(_ast);
-	}
 
 	using ASTModifier::operator();
 	void operator()(Switch& _switch) override;
 	void operator()(Block& _block) override;
 
 private:
-	explicit ConditionalSimplifier(Dialect const& _dialect):
+	explicit ConditionalSimplifier(
-		m_dialect(_dialect)
+		Dialect const& _dialect,
+		std::map<YulString, ControlFlowSideEffects> const& _sideEffects
+	):
+		m_dialect(_dialect), m_functionSideEffects(_sideEffects)
 	{}
 	Dialect const& m_dialect;
+	std::map<YulString, ControlFlowSideEffects> const& m_functionSideEffects;
 };
 
 }

libyul/optimiser/ConditionalUnsimplifier.cpp

@@ -20,6 +20,7 @@
 #include <libyul/AST.h>
 #include <libyul/Utilities.h>
 #include <libyul/optimiser/NameCollector.h>
+#include <libyul/ControlFlowSideEffectsCollector.h>
 #include <libsolutil/CommonData.h>
 
 using namespace std;
@@ -27,6 +28,12 @@ using namespace solidity;
 using namespace solidity::yul;
 using namespace solidity::util;
 
+void ConditionalUnsimplifier::run(OptimiserStepContext& _context, Block& _ast)
+{
+	ControlFlowSideEffectsCollector sideEffects(_context.dialect, _ast);
+	ConditionalUnsimplifier{_context.dialect, sideEffects.functionSideEffects()}(_ast);
+}
+
 void ConditionalUnsimplifier::operator()(Switch& _switch)
 {
 	visit(*_switch.expression);
@@ -78,7 +85,7 @@ void ConditionalUnsimplifier::operator()(Block& _block)
 					YulString condition = std::get<Identifier>(*_if.condition).name;
 					if (
 						holds_alternative<Assignment>(_stmt2) &&
-						TerminationFinder(m_dialect).controlFlowKind(_if.body.statements.back()) !=
+						TerminationFinder(m_dialect, &m_functionSideEffects).controlFlowKind(_if.body.statements.back()) !=
 							TerminationFinder::ControlFlow::FlowOut
 					)
 					{

libyul/optimiser/ConditionalUnsimplifier.h

@@ -33,20 +33,21 @@ class ConditionalUnsimplifier: public ASTModifier
 {
 public:
 	static constexpr char const* name{"ConditionalUnsimplifier"};
-	static void run(OptimiserStepContext& _context, Block& _ast)
+	static void run(OptimiserStepContext& _context, Block& _ast);
-	{
-		ConditionalUnsimplifier{_context.dialect}(_ast);
-	}
 
 	using ASTModifier::operator();
 	void operator()(Switch& _switch) override;
 	void operator()(Block& _block) override;
 
 private:
-	explicit ConditionalUnsimplifier(Dialect const& _dialect):
+	explicit ConditionalUnsimplifier(
-		m_dialect(_dialect)
+		Dialect const& _dialect,
+		std::map<YulString, ControlFlowSideEffects> const& _sideEffects
+	):
+		m_dialect(_dialect), m_functionSideEffects(_sideEffects)
 	{}
 	Dialect const& m_dialect;
+	std::map<YulString, ControlFlowSideEffects> const& m_functionSideEffects;
 };
 
 }

libyul/optimiser/DeadCodeEliminator.cpp

@@ -22,6 +22,7 @@
 #include <libyul/optimiser/DeadCodeEliminator.h>
 #include <libyul/optimiser/Semantics.h>
 #include <libyul/optimiser/OptimiserStep.h>
+#include <libyul/ControlFlowSideEffectsCollector.h>
 #include <libyul/AST.h>
 
 #include <libevmasm/SemanticInformation.h>
@@ -36,7 +37,11 @@ using namespace solidity::yul;
 
 void DeadCodeEliminator::run(OptimiserStepContext& _context, Block& _ast)
 {
-	DeadCodeEliminator{_context.dialect}(_ast);
+	ControlFlowSideEffectsCollector sideEffects(_context.dialect, _ast);
+	DeadCodeEliminator{
+		_context.dialect,
+		sideEffects.functionSideEffects()
+	}(_ast);
 }
 
 void DeadCodeEliminator::operator()(ForLoop& _for)
@@ -49,7 +54,7 @@ void DeadCodeEliminator::operator()(Block& _block)
 {
 	TerminationFinder::ControlFlow controlFlowChange;
 	size_t index;
-	tie(controlFlowChange, index) = TerminationFinder{m_dialect}.firstUnconditionalControlFlowChange(_block.statements);
+	tie(controlFlowChange, index) = TerminationFinder{m_dialect, &m_functionSideEffects}.firstUnconditionalControlFlowChange(_block.statements);
 
 	// Erase everything after the terminating statement that is not a function definition.
 	if (controlFlowChange != TerminationFinder::ControlFlow::FlowOut && index != std::numeric_limits<size_t>::max())

libyul/optimiser/DeadCodeEliminator.h

@@ -31,12 +31,15 @@ namespace solidity::yul
 {
 struct Dialect;
 struct OptimiserStepContext;
+struct ControlFlowSideEffects;
 
 /**
  * Optimisation stage that removes unreachable code
  *
  * Unreachable code is any code within a block which is preceded by a
- * leave, return, invalid, break, continue, selfdestruct or revert.
+ * leave, return, invalid, break, continue, selfdestruct or revert or
+ * a call to a user-defined function that never returns (either due to
+ * recursion or a call to return / revert / stop).
  *
  * Function definitions are retained as they might be called by earlier
  * code and thus are considered reachable.
@@ -57,9 +60,13 @@ public:
 	void operator()(Block& _block) override;
 
 private:
-	DeadCodeEliminator(Dialect const& _dialect): m_dialect(_dialect) {}
+	DeadCodeEliminator(
+		Dialect const& _dialect,
+		std::map<YulString, ControlFlowSideEffects> const& _sideEffects
+	): m_dialect(_dialect), m_functionSideEffects(_sideEffects) {}
 
 	Dialect const& m_dialect;
+	std::map<YulString, ControlFlowSideEffects> const& m_functionSideEffects;
 };
 
 }

libyul/optimiser/Semantics.cpp

@@ -182,8 +182,19 @@ pair<TerminationFinder::ControlFlow, size_t> TerminationFinder::firstUncondition
 TerminationFinder::ControlFlow TerminationFinder::controlFlowKind(Statement const& _statement)
 {
 	if (
+		holds_alternative<VariableDeclaration>(_statement) &&
+		std::get<VariableDeclaration>(_statement).value &&
+		containsNonContinuingFunctionCall(*std::get<VariableDeclaration>(_statement).value)
+	)
+		return ControlFlow::Terminate;
+	else if (
+		holds_alternative<Assignment>(_statement) &&
+		containsNonContinuingFunctionCall(*std::get<Assignment>(_statement).value)
+	)
+		return ControlFlow::Terminate;
+	else if (
 		holds_alternative<ExpressionStatement>(_statement) &&
-		isTerminatingBuiltin(std::get<ExpressionStatement>(_statement))
+		containsNonContinuingFunctionCall(std::get<ExpressionStatement>(_statement).expression)
 	)
 		return ControlFlow::Terminate;
 	else if (holds_alternative<Break>(_statement))
@@ -196,10 +207,18 @@ TerminationFinder::ControlFlow TerminationFinder::controlFlowKind(Statement cons
 		return ControlFlow::FlowOut;
 }
 
-bool TerminationFinder::isTerminatingBuiltin(ExpressionStatement const& _exprStmnt)
+bool TerminationFinder::containsNonContinuingFunctionCall(Expression const& _expr)
 {
-	if (holds_alternative<FunctionCall>(_exprStmnt.expression))
+	if (auto functionCall = std::get_if<FunctionCall>(&_expr))
-		if (auto instruction = toEVMInstruction(m_dialect, std::get<FunctionCall>(_exprStmnt.expression).functionName.name))
+	{
-			return evmasm::SemanticInformation::terminatesControlFlow(*instruction);
+		for (auto const& arg: functionCall->arguments)
+			if (containsNonContinuingFunctionCall(arg))
+				return true;
+
+		if (auto builtin = m_dialect.builtin(functionCall->functionName.name))
+			return !builtin->controlFlowSideEffects.canContinue;
+		else if (m_functionSideEffects && m_functionSideEffects->count(functionCall->functionName.name))
+			return !m_functionSideEffects->at(functionCall->functionName.name).canContinue;
+	}
 	return false;
 }

libyul/optimiser/Semantics.h

@@ -205,22 +205,31 @@ private:
 	std::set<YulString> m_variableReferences;
 };
 
+struct ControlFlowSideEffects;
 
 /**
  * Helper class to find "irregular" control flow.
- * This includes termination, break and continue.
+ * This includes termination, break, continue and leave.
+ * In general, it is applied only to "simple" statements. The control-flow
+ * of loops, switches and if statements is always "FlowOut" with the assumption
+ * that the caller will descend into them.
  */
 class TerminationFinder
 {
 public:
-	// TODO check all uses of TerminationFinder!
+	/// "Terminate" here means that there is no continuing control-flow.
+	/// If this is applied to a function that can revert or stop, but can also
+	/// exit regularly, the property is set to "FlowOut".
 	enum class ControlFlow { FlowOut, Break, Continue, Terminate, Leave };
 
-	TerminationFinder(Dialect const& _dialect): m_dialect(_dialect) {}
+	TerminationFinder(
+		Dialect const& _dialect,
+		std::map<YulString, ControlFlowSideEffects> const* _functionSideEffects = nullptr
+	): m_dialect(_dialect), m_functionSideEffects(_functionSideEffects) {}
 
 	/// @returns the index of the first statement in the provided sequence
 	/// that is an unconditional ``break``, ``continue``, ``leave`` or a
-	/// call to a terminating builtin function.
+	/// call to a terminating 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
@@ -233,13 +242,14 @@ public:
 	/// This function could return FlowOut even if control flow never continues.
 	ControlFlow controlFlowKind(Statement const& _statement);
 
-	/// @returns true if the expression statement is a direct
+	/// @returns true if the expression contains a
-	/// call to a builtin terminating function like
+	/// call to a terminating function, i.e. a function that does not have
-	/// ``stop``, ``revert`` or ``return``.
+	/// a regular "flow out" control-flow (it might also be recursive).
-	bool isTerminatingBuiltin(ExpressionStatement const& _exprStmnt);
+	bool containsNonContinuingFunctionCall(Expression const& _expr);
 
 private:
 	Dialect const& m_dialect;
+	std::map<YulString, ControlFlowSideEffects> const* m_functionSideEffects;
 };
 
 }

test/libyul/yulOptimizerTests/conditionalSimplifier/side_effects_of_functions.yul

@@ -0,0 +1,38 @@
+{
+    function recursive() { recursive() }
+    function terminating() { stop() }
+    function maybeReverting() { if calldataload(0) { revert(0, 0) } }
+
+    let a := calldataload(7)
+    if a { recursive() }
+
+    a := calldataload(a)
+    if a { maybeReverting() }
+
+    a := calldataload(a)
+    if a { terminating() }
+
+    sstore(0, a)
+}
+// ----
+// step: conditionalSimplifier
+//
+// {
+//     function recursive()
+//     { recursive() }
+//     function terminating()
+//     { stop() }
+//     function maybeReverting()
+//     {
+//         if calldataload(0) { revert(0, 0) }
+//     }
+//     let a := calldataload(7)
+//     if a { recursive() }
+//     a := 0
+//     a := calldataload(a)
+//     if a { maybeReverting() }
+//     a := calldataload(a)
+//     if a { terminating() }
+//     a := 0
+//     sstore(0, a)
+// }

test/libyul/yulOptimizerTests/conditionalUnsimplifier/side_effects_of_functions.yul

@@ -0,0 +1,38 @@
+{
+    function recursive() { recursive() }
+    function terminating() { stop() }
+    function maybeReverting() { if calldataload(0) { revert(0, 0) } }
+
+    let a := calldataload(7)
+    if a { recursive() }
+    a := 0
+
+    a := calldataload(a)
+    if a { maybeReverting() }
+
+    a := calldataload(a)
+    if a { terminating() }
+    a := 0
+
+    sstore(0, a)
+}
+// ----
+// step: conditionalUnsimplifier
+//
+// {
+//     function recursive()
+//     { recursive() }
+//     function terminating()
+//     { stop() }
+//     function maybeReverting()
+//     {
+//         if calldataload(0) { revert(0, 0) }
+//     }
+//     let a := calldataload(7)
+//     if a { recursive() }
+//     a := calldataload(a)
+//     if a { maybeReverting() }
+//     a := calldataload(a)
+//     if a { terminating() }
+//     sstore(0, a)
+// }

test/libyul/yulOptimizerTests/deadCodeEliminator/dead_code_user_functions.yul

@@ -0,0 +1,44 @@
+{
+    switch calldataload(0)
+    case 0 {
+        recursive()
+        sstore(0, 1)
+    }
+    case 1 {
+        terminating()
+        sstore(0, 7)
+    }
+    case 2 {
+        reverting()
+        sstore(0, 7)
+    }
+
+
+    function recursive()
+    {
+        recursive()
+    }
+    function terminating()
+    {
+        return(0, 0)
+    }
+    function reverting()
+    {
+        revert(0, 0)
+    }
+}
+// ----
+// step: deadCodeEliminator
+//
+// {
+//     switch calldataload(0)
+//     case 0 { recursive() }
+//     case 1 { terminating() }
+//     case 2 { reverting() }
+//     function recursive()
+//     { recursive() }
+//     function terminating()
+//     { return(0, 0) }
+//     function reverting()
+//     { revert(0, 0) }
+// }

test/libyul/yulOptimizerTests/deadCodeEliminator/function_after_revert.yul

@@ -5,9 +5,7 @@
 
     function fun()
     {
-        return(1, 1)
+        sstore(0, 1)
-
-        pop(sub(10, 5))
     }
 
     pop(add(1, 1))
@@ -19,5 +17,5 @@
 //     fun()
 //     revert(0, 0)
 //     function fun()
-//     { return(1, 1) }
+//     { sstore(0, 1) }
 // }

test/libyul/yulOptimizerTests/fullSuite/name_cleaner_reserved.yul

@@ -1,11 +1,11 @@
 {
 	// This function name can be shortened, the other cannot.
 	function nonmstore_(x) {
-		nonmstore_(x)
+		if calldataload(0) { nonmstore_(x) }
 		sstore(10, calldataload(2))
 	}
 	function mstore_(x) -> y {
-		let t3_3_ := mstore_(x)
+		if calldataload(0) { let t3_3_ := mstore_(x) }
 		y := 8
 		sstore(y, calldataload(y))
 	}
@@ -22,12 +22,12 @@
 //     }
 //     function nonmstore(x)
 //     {
-//         nonmstore(x)
+//         if calldataload(0) { nonmstore(x) }
 //         sstore(10, calldataload(2))
 //     }
 //     function mstore_(x) -> y
 //     {
-//         pop(mstore_(x))
+//         if calldataload(0) { pop(mstore_(x)) }
 //         y := 8
 //         sstore(y, calldataload(y))
 //     }

test/libyul/yulOptimizerTests/fullSuite/unusedFunctionParameterPruner_recursion.yul

@@ -5,7 +5,9 @@
     sstore(1, l)
     function f(a, b, c) -> x, y, z
     {
-        x, y, z := f(1, 2, 3)
+        if calldataload(0) {
+            x, y, z := f(1, 2, 3)
+        }
         x := add(x, 1)
     }
 }
@@ -21,9 +23,13 @@
 //     }
 //     function f() -> x, y, z
 //     {
-//         let x_1, y_1, z_1 := f()
+//         if calldataload(0)
-//         y := y_1
+//         {
-//         z := z_1
+//             let x_1, y_1, z_1 := f()
-//         x := add(x_1, 1)
+//             x := x_1
+//             y := y_1
+//             z := z_1
+//         }
+//         x := add(x, 1)
 //     }
 // }