Merge pull request #14117 from ethereum/push0-optimisations

Change the constant optimizer to make use of `PUSH0`

.circleci/config.yml

@@ -1438,6 +1438,14 @@ jobs:
             # TODO: temporarily set hardhat stack traces tests to use cancun hardfork
             # Remove this when hardhat switch to cancun by default: https://github.com/NomicFoundation/hardhat/issues/4851
             sed -i 's/hardfork: "shanghai",/hardfork: "cancun",/' test/internal/hardhat-network/stack-traces/execution.ts
+            # TODO: Remove these tests because they rely on specific stack sequence which has changed since
+            # Remove when hardhat properly fix a specific version for the tests: https://github.com/NomicFoundation/hardhat/issues/5443
+            rm -rf test/internal/hardhat-network/stack-traces/test-files/0_8/revert-without-message/modifiers/call-message/multiple-modifiers-require/
+            rm -rf test/internal/hardhat-network/stack-traces/test-files/0_8/revert-without-message/modifiers/create-message/multiple-modifiers-require/
+            rm -rf test/internal/hardhat-network/stack-traces/test-files/0_8/revert-without-message/revert-without-message/within-receive/between-statements/
+            rm -rf test/internal/hardhat-network/stack-traces/test-files/0_8/revert-without-message/revert-without-message/within-receive/no-other-statements/
+            rm -rf test/internal/hardhat-network/stack-traces/test-files/0_8/revert-without-message/revert-without-message/within-receive/statement-after/
+            rm -rf test/internal/hardhat-network/stack-traces/test-files/0_8/revert-without-message/revert-without-message/within-receive/statement-before/
             pnpm test
       - matrix_notify_failure_unless_pr
 

Changelog.md

@@ -9,8 +9,11 @@ Language Features:
 Compiler Features:
  * Command Line Interface: Do not perform IR optimization when only unoptimized IR is requested.
  * Commandline Interface: Add ``--transient-storage-layout`` output.
+ * Constant Optimizer: Uses ``PUSH0`` if supported by the selected evm version.
  * Error Reporting: Unimplemented features are now properly reported as errors instead of being handled as if they were bugs.
  * EVM: Support for the EVM version "Prague".
+ * Peephole Optimizer: ``PUSH0``, when supported, is duplicated explicitly instead of using ``DUP1``.
+ * Peephole optimizer: Remove identical code snippets that terminate the control flow if they occur one after another.
  * SMTChecker: Add CHC engine check for underflow and overflow in unary minus operation.
  * SMTChecker: Replace CVC4 as a possible BMC backend with cvc5.
  * Standard JSON Interface: Do not perform IR optimization when only unoptimized IR is requested.

libevmasm/Assembly.cpp

@@ -74,7 +74,7 @@ unsigned Assembly::codeSize(unsigned subTagSize) const
 			ret += i.second.size();
 
 		for (AssemblyItem const& i: m_items)
-			ret += i.bytesRequired(tagSize, Precision::Approximate);
+			ret += i.bytesRequired(tagSize, m_evmVersion, Precision::Approximate);
 		if (numberEncodingSize(ret) <= tagSize)
 			return static_cast<unsigned>(ret);
 	}
@@ -753,7 +753,7 @@ std::map<u256, u256> const& Assembly::optimiseInternal(
 
 		if (_settings.runPeephole)
 		{
-			PeepholeOptimiser peepOpt{m_items};
+			PeepholeOptimiser peepOpt{m_items, m_evmVersion};
 			while (peepOpt.optimise())
 			{
 				count++;

libevmasm/AssemblyItem.cpp

@@ -116,15 +116,17 @@ void AssemblyItem::setPushTagSubIdAndTag(size_t _subId, size_t _tag)
 	setData(data);
 }
 
-size_t AssemblyItem::bytesRequired(size_t _addressLength, Precision _precision) const
+size_t AssemblyItem::bytesRequired(size_t _addressLength, langutil::EVMVersion _evmVersion, Precision _precision) const
 {
 	switch (m_type)
 	{
 	case Operation:
 	case Tag: // 1 byte for the JUMPDEST
 		return 1;
 	case Push:
-		return 1 + std::max<size_t>(1, numberEncodingSize(data()));
+		return
+			1 +
+			std::max<size_t>((_evmVersion.hasPush0() ? 0 : 1), numberEncodingSize(data()));
 	case PushSubSize:
 	case PushProgramSize:
 		return 1 + 4; // worst case: a 16MB program

libevmasm/AssemblyItem.h

@@ -161,10 +161,11 @@ class AssemblyItem
 
 	/// @returns an upper bound for the number of bytes required by this item, assuming that
 	/// the value of a jump tag takes @a _addressLength bytes.
+	/// @param _evmVersion the EVM version
 	/// @param _precision Whether to return a precise count (which involves
 	///                   counting immutable references which are only set after
 	///                   a call to `assemble()`) or an approx. count.
-	size_t bytesRequired(size_t _addressLength, Precision _precision = Precision::Precise) const;
+	size_t bytesRequired(size_t _addressLength, langutil::EVMVersion _evmVersion, Precision _precision = Precision::Precise) const;
 	size_t arguments() const;
 	size_t returnValues() const;
 	size_t deposit() const { return returnValues() - arguments(); }
@@ -228,11 +229,11 @@ class AssemblyItem
 	mutable std::optional<size_t> m_immutableOccurrences;
 };
 
-inline size_t bytesRequired(AssemblyItems const& _items, size_t _addressLength,  Precision _precision = Precision::Precise)
+inline size_t bytesRequired(AssemblyItems const& _items, size_t _addressLength, langutil::EVMVersion _evmVersion, Precision _precision = Precision::Precise)
 {
 	size_t size = 0;
 	for (AssemblyItem const& item: _items)
-		size += item.bytesRequired(_addressLength, _precision);
+		size += item.bytesRequired(_addressLength, _evmVersion, _precision);
 	return size;
 }
 

libevmasm/ConstantOptimiser.cpp

@@ -83,7 +83,7 @@ bigint ConstantOptimisationMethod::simpleRunGas(AssemblyItems const& _items, lan
 	bigint gas = 0;
 	for (AssemblyItem const& item: _items)
 		if (item.type() == Push)
-			gas += GasMeter::runGas(Instruction::PUSH1, _evmVersion);
+			gas += GasMeter::pushGas(item.data(), _evmVersion);
 		else if (item.type() == Operation)
 		{
 			if (item.instruction() == Instruction::EXP)
@@ -100,9 +100,9 @@ bigint ConstantOptimisationMethod::dataGas(bytes const& _data) const
 	return bigint(GasMeter::dataGas(_data, m_params.isCreation, m_params.evmVersion));
 }
 
-size_t ConstantOptimisationMethod::bytesRequired(AssemblyItems const& _items)
+size_t ConstantOptimisationMethod::bytesRequired(AssemblyItems const& _items, langutil::EVMVersion _evmVersion)
 {
-	return evmasm::bytesRequired(_items, 3, Precision::Approximate); // assume 3 byte addresses
+	return evmasm::bytesRequired(_items, 3, _evmVersion, Precision::Approximate); // assume 3 byte addresses
 }
 
 void ConstantOptimisationMethod::replaceConstants(
@@ -143,7 +143,7 @@ bigint CodeCopyMethod::gasNeeded() const
 		// Run gas: we ignore memory increase costs
 		simpleRunGas(copyRoutine(), m_params.evmVersion) + GasCosts::copyGas,
 		// Data gas for copy routines: Some bytes are zero, but we ignore them.
-		bytesRequired(copyRoutine()) * (m_params.isCreation ? GasCosts::txDataNonZeroGas(m_params.evmVersion) : GasCosts::createDataGas),
+		bytesRequired(copyRoutine(), m_params.evmVersion) * (m_params.isCreation ? GasCosts::txDataNonZeroGas(m_params.evmVersion) : GasCosts::createDataGas),
 		// Data gas for data itself
 		dataGas(toBigEndian(m_value))
 	);
@@ -153,37 +153,74 @@ AssemblyItems CodeCopyMethod::execute(Assembly& _assembly) const
 {
 	bytes data = toBigEndian(m_value);
 	assertThrow(data.size() == 32, OptimizerException, "Invalid number encoding.");
-	AssemblyItems actualCopyRoutine = copyRoutine();
-	actualCopyRoutine[4] = _assembly.newData(data);
-	return actualCopyRoutine;
+	AssemblyItem newPushData = _assembly.newData(data);
+	return copyRoutine(&newPushData);
 }
 
-AssemblyItems const& CodeCopyMethod::copyRoutine()
+AssemblyItems CodeCopyMethod::copyRoutine(AssemblyItem* _pushData) const
 {
-	AssemblyItems static copyRoutine{
-		// constant to be reused 3+ times
-		u256(0),
+	if (_pushData)
+		assertThrow(_pushData->type() == PushData, OptimizerException, "Invalid Assembly Item.");
 
-		// back up memory
-		// mload(0)
-		Instruction::DUP1,
-		Instruction::MLOAD,
+	AssemblyItem dataUsed = _pushData ? *_pushData : AssemblyItem(PushData, u256(1) << 16);
 
-		// codecopy(0, <offset>, 32)
-		u256(32),
-		AssemblyItem(PushData, u256(1) << 16), // replaced above in actualCopyRoutine[4]
-		Instruction::DUP4,
-		Instruction::CODECOPY,
+	// PUSH0 is cheaper than PUSHn/DUP/SWAP.
+	if (m_params.evmVersion.hasPush0())
+	{
+		// This costs ~29 gas.
+		AssemblyItems copyRoutine{
+			// back up memory
+			// mload(0)
+			u256(0),
+			Instruction::MLOAD,
+
+			// codecopy(0, <offset>, 32)
+			u256(32),
+			dataUsed,
+			u256(0),
+			Instruction::CODECOPY,
+
+			// mload(0)
+			u256(0),
+			Instruction::MLOAD,
+
+			// restore original memory
+			// mstore(0, x)
+			Instruction::SWAP1,
+			u256(0),
+			Instruction::MSTORE
+		};
+		return copyRoutine;
+	}
+	else
+	{
+		// This costs ~33 gas.
+		AssemblyItems copyRoutine{
+			// constant to be reused 3+ times
+			u256(0),
+
+			// back up memory
+			// mload(0)
+			Instruction::DUP1,
+			Instruction::MLOAD,
 
-		// mload(0)
-		Instruction::DUP2,
-		Instruction::MLOAD,
+			// codecopy(0, <offset>, 32)
+			u256(32),
+			dataUsed,
+			Instruction::DUP4,
+			Instruction::CODECOPY,
 
-		// restore original memory
-		Instruction::SWAP2,
-		Instruction::MSTORE
-	};
-	return copyRoutine;
+			// mload(0)
+			Instruction::DUP2,
+			Instruction::MLOAD,
+
+			// restore original memory
+			// mstore(0, x)
+			Instruction::SWAP2,
+			Instruction::MSTORE
+		};
+		return copyRoutine;
+	}
 }
 
 AssemblyItems ComputeMethod::findRepresentation(u256 const& _value)
@@ -323,7 +360,7 @@ bigint ComputeMethod::gasNeeded(AssemblyItems const& _routine) const
 	return combineGas(
 		simpleRunGas(_routine, m_params.evmVersion) + numExps * (GasCosts::expGas + GasCosts::expByteGas(m_params.evmVersion)),
 		// Data gas for routine: Some bytes are zero, but we ignore them.
-		bytesRequired(_routine) * (m_params.isCreation ? GasCosts::txDataNonZeroGas(m_params.evmVersion) : GasCosts::createDataGas),
+		bytesRequired(_routine, m_params.evmVersion) * (m_params.isCreation ? GasCosts::txDataNonZeroGas(m_params.evmVersion) : GasCosts::createDataGas),
 		0
 	);
 }

libevmasm/ConstantOptimiser.h

@@ -79,7 +79,7 @@ class ConstantOptimisationMethod
 	static bigint simpleRunGas(AssemblyItems const& _items, langutil::EVMVersion _evmVersion);
 	/// @returns the gas needed to store the given data literally
 	bigint dataGas(bytes const& _data) const;
-	static size_t bytesRequired(AssemblyItems const& _items);
+	static size_t bytesRequired(AssemblyItems const& _items, langutil::EVMVersion _evmVersion);
 	/// @returns the combined estimated gas usage taking @a m_params into account.
 	bigint combineGas(
 		bigint const& _runGas,
@@ -123,7 +123,7 @@ class CodeCopyMethod: public ConstantOptimisationMethod
 	AssemblyItems execute(Assembly& _assembly) const override;
 
 protected:
-	static AssemblyItems const& copyRoutine();
+	AssemblyItems copyRoutine(AssemblyItem* _pushData = nullptr) const;
 };
 
 /**

libevmasm/GasMeter.cpp

@@ -44,12 +44,8 @@ GasMeter::GasConsumption GasMeter::estimateMax(AssemblyItem const& _item, bool _
 	switch (_item.type())
 	{
 	case Push:
-		if (m_evmVersion.hasPush0() && _item.data() == 0)
-		{
-			gas = runGas(Instruction::PUSH0, m_evmVersion);
-			break;
-		}
-		[[fallthrough]];
+		gas = pushGas(_item.data(), m_evmVersion);
+		break;
 	case PushTag:
 	case PushData:
 	case PushSub:
@@ -290,6 +286,14 @@ unsigned GasMeter::runGas(Instruction _instruction, langutil::EVMVersion _evmVer
 	util::unreachable();
 }
 
+unsigned GasMeter::pushGas(u256 _value, langutil::EVMVersion _evmVersion)
+{
+	return runGas(
+		(_evmVersion.hasPush0() && _value == u256(0)) ? Instruction::PUSH0 : Instruction::PUSH1,
+		_evmVersion
+	);
+}
+
 u256 GasMeter::dataGas(bytes const& _data, bool _inCreation, langutil::EVMVersion _evmVersion)
 {
 	bigint gas = 0;

libevmasm/GasMeter.h

@@ -225,6 +225,9 @@ class GasMeter
 	/// change with EVM versions)
 	static unsigned runGas(Instruction _instruction, langutil::EVMVersion _evmVersion);
 
+	/// @returns gas costs for push instructions (may change depending on EVM version)
+	static unsigned pushGas(u256 _value, langutil::EVMVersion _evmVersion);
+
 	/// @returns the gas cost of the supplied data, depending whether it is in creation code, or not.
 	/// In case of @a _inCreation, the data is only sent as a transaction and is not stored, whereas
 	/// otherwise code will be stored and have to pay "createDataGas" cost.

libevmasm/Inliner.cpp

@@ -59,10 +59,10 @@ u256 executionCost(RangeType const& _itemRange, langutil::EVMVersion _evmVersion
 }
 /// @returns an estimation of the code size in bytes needed for the AssemblyItems in @a _itemRange.
 template<typename RangeType>
-uint64_t codeSize(RangeType const& _itemRange)
+uint64_t codeSize(RangeType const& _itemRange, langutil::EVMVersion _evmVersion)
 {
 	return ranges::accumulate(_itemRange | ranges::views::transform(
-		[](auto const& _item) { return _item.bytesRequired(2, Precision::Approximate); }
+		[&](auto const& _item) { return _item.bytesRequired(2, _evmVersion, Precision::Approximate); }
 	), 0u);
 }
 /// @returns the tag id, if @a _item is a PushTag or Tag into the current subassembly, std::nullopt otherwise.
@@ -139,7 +139,7 @@ std::map<size_t, Inliner::InlinableBlock> Inliner::determineInlinableBlocks(Asse
 bool Inliner::shouldInlineFullFunctionBody(size_t _tag, ranges::span<AssemblyItem const> _block, uint64_t _pushTagCount) const
 {
 	// Accumulate size of the inline candidate block in bytes (without the return jump).
-	uint64_t functionBodySize = codeSize(ranges::views::drop_last(_block, 1));
+	uint64_t functionBodySize = codeSize(ranges::views::drop_last(_block, 1), m_evmVersion);
 
 	// Use the number of push tags as approximation of the average number of calls to the function per run.
 	uint64_t numberOfCalls = _pushTagCount;
@@ -167,8 +167,8 @@ bool Inliner::shouldInlineFullFunctionBody(size_t _tag, ranges::span<AssemblyIte
 	);
 	// Each call site deposits the call site pattern, whereas the jump site pattern and the function itself are deposited once.
 	bigint uninlinedDepositCost = GasMeter::dataGas(
-		numberOfCallSites * codeSize(uninlinedCallSitePattern) +
-		codeSize(uninlinedFunctionPattern) +
+		numberOfCallSites * codeSize(uninlinedCallSitePattern, m_evmVersion) +
+		codeSize(uninlinedFunctionPattern, m_evmVersion) +
 		functionBodySize,
 		m_isCreation,
 		m_evmVersion
@@ -185,7 +185,7 @@ bool Inliner::shouldInlineFullFunctionBody(size_t _tag, ranges::span<AssemblyIte
 	// the heuristics is optimistic.
 	if (m_tagsReferencedFromOutside.count(_tag))
 		inlinedDepositCost += GasMeter::dataGas(
-			codeSize(uninlinedFunctionPattern) + functionBodySize,
+			codeSize(uninlinedFunctionPattern, m_evmVersion) + functionBodySize,
 			m_isCreation,
 			m_evmVersion
 		);
@@ -225,8 +225,8 @@ std::optional<AssemblyItem> Inliner::shouldInline(size_t _tag, AssemblyItem cons
 			AssemblyItem{Instruction::JUMP},
 		};
 		if (
-			GasMeter::dataGas(codeSize(_block.items), m_isCreation, m_evmVersion) <=
-			GasMeter::dataGas(codeSize(jumpPattern), m_isCreation, m_evmVersion)
+			GasMeter::dataGas(codeSize(_block.items, m_evmVersion), m_isCreation, m_evmVersion) <=
+			GasMeter::dataGas(codeSize(jumpPattern, m_evmVersion), m_isCreation, m_evmVersion)
 		)
 			return blockExit;
 	}