Split memcopy into three functions.

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/CompilerUtils.cpp

@@ -298,61 +298,73 @@ void CompilerUtils::zeroInitialiseMemoryArray(ArrayType const& _type)
 	m_context << Instruction::SWAP1 << Instruction::POP;
 }
 
-void CompilerUtils::memoryCopy(bool _useIdentityPrecompile)
+void CompilerUtils::memoryCopyPrecompile()
 {
-	//@TODO do not use ::CALL if less than 32 bytes?
-
 	// Stack here: size target source
 
-	if (!_useIdentityPrecompile)
+	m_context.appendInlineAssembly(R"(
-	{
+		{
-		m_context.appendInlineAssembly(R"(
+		let words := div(add(len, 31), 32)
-			{
+		let cost := add(15, mul(3, words))
-			// expects three locals: src, dst, len
+		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;
+}
 
-			// copy 32 bytes at once
+void CompilerUtils::memoryCopy32()
-			start32:
+{
-			jumpi(end32, lt(len, 32))
+	// Stack here: size target source
-			mstore(dst, mload(src))
-			dst := add(dst, 32)
-			src := add(src, 32)
-			len := sub(len, 32)
-			jump(start32)
-			end32:
 
-			// copy the remainder (0 < len < 32)
+	m_context.appendInlineAssembly(R"(
-			let mask := sub(exp(256, sub(32, len)), 1)
+		{
-			let srcpart := and(mload(src), not(mask))
+		jumpi(end, eq(len, 0))
-			let dstpart := and(mload(dst), mask)
+		start:
-			mstore(dst, or(srcpart, dstpart))
+		mstore(dst, mload(src))
-			}
+		jumpi(end, iszero(gt(len, 32)))
-		)",
+		dst := add(dst, 32)
-			{ "len", "dst", "src" }
+		src := add(src, 32)
-		);
+		len := sub(len, 32)
-		m_context << Instruction::POP;
+		jump(start)
-		m_context << Instruction::POP;
+		end:
-		m_context << Instruction::POP;
+		}
-		return;
+	)",
-	}
+		{ "len", "dst", "src" }
-	else
+	);
-	{
+	m_context << Instruction::POP << Instruction::POP << Instruction::POP;
-		m_context.appendInlineAssembly(R"(
+}
-			{
+
-			let words := div(add(len, 31), 32)
+void CompilerUtils::memoryCopy()
-			let cost := add(15, mul(3, words))
+{
-			jump(invalidJumpLabel, iszero(call(cost, $identityContractAddress, 0, src, len, dst, len)))
+	// Stack here: size target source
-			}
+
-		)",
+	m_context.appendInlineAssembly(R"(
-			{ "len", "dst", "src" },
+		{
-			map<string, string> {
+		// copy 32 bytes at once
-			    { "$identityContractAddress", toString(identityContractAddress) }
+		start32:
-			}
+		jumpi(end32, lt(len, 32))
-		);
+		mstore(dst, mload(src))
-		m_context << Instruction::POP;
+		dst := add(dst, 32)
-		m_context << Instruction::POP;
+		src := add(src, 32)
-		m_context << Instruction::POP;
+		len := sub(len, 32)
-		return;
+		jump(start32)
-	}
+		end32:
+
+		// 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

@@ -112,7 +112,15 @@ public:
 	/// Uses a CALL to the identity contract to perform a memory-to-memory copy.
 	/// Stack pre: <size> <target> <source>
 	/// Stack post:
-	void memoryCopy(bool _useIdentityPrecompile = false);
+	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)
 	/// external function type <address><function identifier> into two stack slots: