Merge pull request #1468 from ethereum/memcpy-assembly

Implement memcpy without the identity precompile

Changelog.md

@@ -3,6 +3,7 @@
 Features:
  * Type checker: Warn when ``msg.value`` is used in non-payable function.
  * Code generator: Inject the Swarm hash of a metadata file into the bytecode.
+ * Code generator: Replace expensive memcpy precompile by simple assembly loop.
  * Optimizer: Some dead code elimination.
 
 Bugfixes:

libsolidity/codegen/ArrayUtils.cpp

@@ -335,9 +335,14 @@ void ArrayUtils::copyArrayToMemory(ArrayType const& _sourceType, bool _padToWord
 		if (baseSize > 1)
 			m_context << u256(baseSize) << Instruction::MUL;
 		// stack: <target> <source> <size>
-		//@TODO do not use ::CALL if less than 32 bytes?
 		m_context << Instruction::DUP1 << Instruction::DUP4 << Instruction::DUP4;
-		utils.memoryCopy();
+		// We can resort to copying full 32 bytes only if
+		// - the length is known to be a multiple of 32 or
+		// - we will pad to full 32 bytes later anyway.
+		if (((baseSize % 32) == 0) || _padToWordBoundaries)
+			utils.memoryCopy32();
+		else
+			utils.memoryCopy();
 
 		m_context << Instruction::SWAP1 << Instruction::POP;
 		// stack: <target> <size>

libsolidity/codegen/CompilerContext.cpp

@@ -202,6 +202,8 @@ void CompilerContext::appendInlineAssembly(
 			return false;
 		unsigned stackDepth = _localVariables.end() - it;
 		int stackDiff = _assembly.deposit() - startStackHeight + stackDepth;
+		if (_context == assembly::CodeGenerator::IdentifierContext::LValue)
+			stackDiff -= 1;
 		if (stackDiff < 1 || stackDiff > 16)
 			BOOST_THROW_EXCEPTION(
 				CompilerError() <<
@@ -217,7 +219,7 @@ void CompilerContext::appendInlineAssembly(
 		return true;
 	};
 
-	solAssert(assembly::InlineAssemblyStack().parseAndAssemble(*assembly, *m_asm, identifierAccess), "");
+	solAssert(assembly::InlineAssemblyStack().parseAndAssemble(*assembly, *m_asm, identifierAccess), "Failed to assemble inline assembly block.");
 }
 
 FunctionDefinition const& CompilerContext::resolveVirtualFunction(

libsolidity/codegen/CompilerUtils.cpp

@@ -298,22 +298,73 @@ void CompilerUtils::zeroInitialiseMemoryArray(ArrayType const& _type)
 	m_context << Instruction::SWAP1 << Instruction::POP;
 }
 
+void CompilerUtils::memoryCopyPrecompile()
+{
+	// Stack here: size target source
+
+	m_context.appendInlineAssembly(R"(
+		{
+		let words := div(add(len, 31), 32)
+		let cost := add(15, mul(3, words))
+		jumpi(invalidJumpLabel, iszero(call(cost, $identityContractAddress, 0, src, len, dst, len)))
+		}
+	)",
+		{ "len", "dst", "src" },
+		map<string, string> {
+			{ "$identityContractAddress", toString(identityContractAddress) }
+		}
+	);
+	m_context << Instruction::POP << Instruction::POP << Instruction::POP;
+}
+
+void CompilerUtils::memoryCopy32()
+{
+	// Stack here: size target source
+
+	m_context.appendInlineAssembly(R"(
+		{
+		jumpi(end, eq(len, 0))
+		start:
+		mstore(dst, mload(src))
+		jumpi(end, iszero(gt(len, 32)))
+		dst := add(dst, 32)
+		src := add(src, 32)
+		len := sub(len, 32)
+		jump(start)
+		end:
+		}
+	)",
+		{ "len", "dst", "src" }
+	);
+	m_context << Instruction::POP << Instruction::POP << Instruction::POP;
+}
+
 void CompilerUtils::memoryCopy()
 {
 	// Stack here: size target source
-	// stack for call: outsize target size source value contract gas
+
-	//@TODO do not use ::CALL if less than 32 bytes?
+	m_context.appendInlineAssembly(R"(
-	m_context << Instruction::DUP3 << Instruction::SWAP1;
+		{
-	m_context << u256(0) << u256(identityContractAddress);
+		// copy 32 bytes at once
-	// compute gas costs
+		start32:
-	m_context << u256(32) << Instruction::DUP5 << u256(31) << Instruction::ADD;
+		jumpi(end32, lt(len, 32))
-	static unsigned c_identityGas = 15;
+		mstore(dst, mload(src))
-	static unsigned c_identityWordGas = 3;
+		dst := add(dst, 32)
-	m_context << Instruction::DIV << u256(c_identityWordGas) << Instruction::MUL;
+		src := add(src, 32)
-	m_context << u256(c_identityGas) << Instruction::ADD;
+		len := sub(len, 32)
-	m_context << Instruction::CALL;
+		jump(start32)
-	m_context << Instruction::ISZERO;
+		end32:
-	m_context.appendConditionalJumpTo(m_context.errorTag());
+
+		// copy the remainder (0 < len < 32)
+		let mask := sub(exp(256, sub(32, len)), 1)
+		let srcpart := and(mload(src), not(mask))
+		let dstpart := and(mload(dst), mask)
+		mstore(dst, or(srcpart, dstpart))
+		}
+	)",
+		{ "len", "dst", "src" }
+	);
+	m_context << Instruction::POP << Instruction::POP << Instruction::POP;
 }
 
 void CompilerUtils::splitExternalFunctionType(bool _leftAligned)

libsolidity/codegen/CompilerUtils.h

@@ -112,6 +112,14 @@ public:
 	/// Uses a CALL to the identity contract to perform a memory-to-memory copy.
 	/// Stack pre: <size> <target> <source>
 	/// Stack post:
+	void memoryCopyPrecompile();
+	/// Copies full 32 byte words in memory (regions cannot overlap), i.e. may copy more than length.
+	/// Stack pre: <size> <target> <source>
+	/// Stack post:
+	void memoryCopy32();
+	/// Copies data in memory (regions cannot overlap).
+	/// Stack pre: <size> <target> <source>
+	/// Stack post:
 	void memoryCopy();
 
 	/// Converts the combined and left-aligned (right-aligned if @a _rightAligned is true)