diff --git a/libsolidity/ArrayUtils.cpp b/libsolidity/ArrayUtils.cpp
index 0dea5345c..3619b5408 100644
--- a/libsolidity/ArrayUtils.cpp
+++ b/libsolidity/ArrayUtils.cpp
@@ -37,140 +37,110 @@ void ArrayUtils::copyArrayToStorage(ArrayType const& _targetType, ArrayType cons
 	// stack layout: [source_ref] target_ref (top)
 	// need to leave target_ref on the stack at the end
 	solAssert(_targetType.getLocation() == ArrayType::Location::Storage, "");
+	solAssert(
+		_sourceType.getLocation() == ArrayType::Location::CallData ||
+			_sourceType.getLocation() == ArrayType::Location::Storage,
+		"Given array location not implemented."
+	);
 
 	IntegerType uint256(256);
 	Type const* targetBaseType = _targetType.isByteArray() ? &uint256 : &(*_targetType.getBaseType());
 	Type const* sourceBaseType = _sourceType.isByteArray() ? &uint256 : &(*_sourceType.getBaseType());
 
-	switch (_sourceType.getLocation())
-	{
-	case ArrayType::Location::CallData:
-	{
-		solAssert(_targetType.isByteArray(), "Non byte arrays not yet implemented here.");
-		solAssert(_sourceType.isByteArray(), "Non byte arrays not yet implemented here.");
-		// This also assumes that after "length" we only have zeros, i.e. it cannot be used to
-		// slice a byte array from calldata.
+	// this copies source to target and also clears target if it was larger
 
-		// stack: source_offset source_len target_ref
-		// fetch old length and convert to words
-		m_context << eth::Instruction::DUP1 << eth::Instruction::SLOAD;
-		convertLengthToSize(_targetType);
-		// stack here: source_offset source_len target_ref target_length_words
-		// actual array data is stored at SHA3(storage_offset)
-		m_context << eth::Instruction::DUP2;
-		CompilerUtils(m_context).computeHashStatic();
-		// compute target_data_end
-		m_context << eth::Instruction::DUP1 << eth::Instruction::SWAP2 << eth::Instruction::ADD
-				  << eth::Instruction::SWAP1;
-		// stack here: source_offset source_len target_ref target_data_end target_data_ref
-		// store length (in bytes)
-		m_context << eth::Instruction::DUP4 << eth::Instruction::DUP1 << eth::Instruction::DUP5
-			<< eth::Instruction::SSTORE;
-		// jump to end if length is zero
-		m_context << eth::Instruction::ISZERO;
-		eth::AssemblyItem copyLoopEnd = m_context.newTag();
-		m_context.appendConditionalJumpTo(copyLoopEnd);
-		// add length to source offset
-		m_context << eth::Instruction::DUP5 << eth::Instruction::DUP5 << eth::Instruction::ADD;
-		// stack now: source_offset source_len target_ref target_data_end target_data_ref source_end
-		// store start offset
-		m_context << eth::Instruction::DUP6;
-		// stack now: source_offset source_len target_ref target_data_end target_data_ref source_end calldata_offset
-		eth::AssemblyItem copyLoopStart = m_context.newTag();
-		m_context << copyLoopStart
-				  // copy from calldata and store
-				  << eth::Instruction::DUP1 << eth::Instruction::CALLDATALOAD
-				  << eth::Instruction::DUP4 << eth::Instruction::SSTORE
-				  // increment target_data_ref by 1
-				  << eth::Instruction::SWAP2 << u256(1) << eth::Instruction::ADD
-				  // increment calldata_offset by 32
-				  << eth::Instruction::SWAP2 << u256(32) << eth::Instruction::ADD
-				  // check for loop condition
-				  << eth::Instruction::DUP1 << eth::Instruction::DUP3 << eth::Instruction::GT;
-		m_context.appendConditionalJumpTo(copyLoopStart);
-		m_context << eth::Instruction::POP << eth::Instruction::POP;
-		m_context << copyLoopEnd;
+	// TODO unroll loop for small sizes
 
-		// now clear leftover bytes of the old value
-		// stack now: source_offset source_len target_ref target_data_end target_data_ref
-		clearStorageLoop(IntegerType(256));
-		// stack now: source_offset source_len target_ref target_data_end
+	// stack: source_ref [source_length] target_ref
+	// store target_ref
+	for (unsigned i = _sourceType.getSizeOnStack(); i > 0; --i)
+		m_context << eth::swapInstruction(i);
+	if (_sourceType.getLocation() != ArrayType::Location::CallData || !_sourceType.isDynamicallySized())
+		retrieveLength(_sourceType); // otherwise, length is already there
+	// stack: target_ref source_ref source_length
+	m_context << eth::Instruction::DUP3;
+	// stack: target_ref source_ref source_length target_ref
+	retrieveLength(_targetType);
+	// stack: target_ref source_ref source_length target_ref target_length
+	if (_targetType.isDynamicallySized())
+		// store new target length
+		m_context << eth::Instruction::DUP3 << eth::Instruction::DUP3 << eth::Instruction::SSTORE;
+	// compute hashes (data positions)
+	m_context << eth::Instruction::SWAP1;
+	if (_targetType.isDynamicallySized())
+		CompilerUtils(m_context).computeHashStatic();
+	// stack: target_ref source_ref source_length target_length target_data_pos
+	m_context << eth::Instruction::SWAP1;
+	convertLengthToSize(_targetType);
+	m_context << eth::Instruction::DUP2 << eth::Instruction::ADD;
+	// stack: target_ref source_ref source_length target_data_pos target_data_end
+	m_context << eth::Instruction::SWAP3;
+	// stack: target_ref target_data_end source_length target_data_pos source_ref
+	// skip copying if source length is zero
+	m_context << eth::Instruction::DUP3 << eth::Instruction::ISZERO;
+	eth::AssemblyItem copyLoopEnd = m_context.newTag();
+	m_context.appendConditionalJumpTo(copyLoopEnd);
 
-		m_context << eth::Instruction::POP << eth::Instruction::SWAP2
-			<< eth::Instruction::POP << eth::Instruction::POP;
-		break;
-	}
-	case ArrayType::Location::Storage:
+	if (_sourceType.getLocation() == ArrayType::Location::Storage && _sourceType.isDynamicallySized())
+		CompilerUtils(m_context).computeHashStatic();
+	// stack: target_ref target_data_end source_length target_data_pos source_data_pos
+	m_context << eth::Instruction::SWAP2;
+	convertLengthToSize(_sourceType);
+	m_context << eth::Instruction::DUP3 << eth::Instruction::ADD;
+	// stack: target_ref target_data_end source_data_pos target_data_pos source_data_end
+	eth::AssemblyItem copyLoopStart = m_context.newTag();
+	m_context << copyLoopStart;
+	// check for loop condition
+	m_context
+		<< eth::Instruction::DUP3 << eth::Instruction::DUP2
+		<< eth::Instruction::GT << eth::Instruction::ISZERO;
+	m_context.appendConditionalJumpTo(copyLoopEnd);
+	// stack: target_ref target_data_end source_data_pos target_data_pos source_data_end
+	// copy
+	if (sourceBaseType->getCategory() == Type::Category::Array)
 	{
-		// this copies source to target and also clears target if it was larger
-
-		solAssert(sourceBaseType->getStorageSize() == targetBaseType->getStorageSize(),
-			"Copying with different storage sizes not yet implemented.");
-		// stack: source_ref target_ref
-		// store target_ref
-		m_context << eth::Instruction::SWAP1 << eth::Instruction::DUP2;
-		// stack: target_ref source_ref target_ref
-		// fetch lengthes
-		retrieveLength(_targetType);
-		m_context << eth::Instruction::SWAP2;
-		// stack: target_ref target_len target_ref source_ref
-		retrieveLength(_sourceType);
-		// stack: target_ref target_len target_ref source_ref source_len
-		if (_targetType.isDynamicallySized())
-			// store new target length
-			m_context << eth::Instruction::DUP1 << eth::Instruction::DUP4 << eth::Instruction::SSTORE;
-		// compute hashes (data positions)
-		m_context << eth::Instruction::SWAP2;
-		if (_targetType.isDynamicallySized())
-			CompilerUtils(m_context).computeHashStatic();
-		m_context << eth::Instruction::SWAP1;
-		if (_sourceType.isDynamicallySized())
-			CompilerUtils(m_context).computeHashStatic();
-		// stack: target_ref target_len source_len target_data_pos source_data_pos
-		m_context << eth::Instruction::DUP4;
-		convertLengthToSize(_sourceType);
-		m_context << eth::Instruction::DUP4;
-		convertLengthToSize(_sourceType);
-		// stack: target_ref target_len source_len target_data_pos source_data_pos target_size source_size
-		// @todo we might be able to go without a third counter
-		m_context << u256(0);
-		// stack: target_ref target_len source_len target_data_pos source_data_pos target_size source_size counter
-		eth::AssemblyItem copyLoopStart = m_context.newTag();
-		m_context << copyLoopStart;
-		// check for loop condition
-		m_context << eth::Instruction::DUP1 << eth::Instruction::DUP3
-				   << eth::Instruction::GT << eth::Instruction::ISZERO;
-		eth::AssemblyItem copyLoopEnd = m_context.newTag();
-		m_context.appendConditionalJumpTo(copyLoopEnd);
-		// copy
-		m_context << eth::Instruction::DUP4 << eth::Instruction::DUP2 << eth::Instruction::ADD;
-		StorageItem(m_context, *sourceBaseType).retrieveValue(SourceLocation(), true);
-		m_context << eth::dupInstruction(5 + sourceBaseType->getSizeOnStack())
-			<< eth::dupInstruction(2 + sourceBaseType->getSizeOnStack()) << eth::Instruction::ADD;
-		StorageItem(m_context, *targetBaseType).storeValue(*sourceBaseType, SourceLocation(), true);
-		// increment
-		m_context << targetBaseType->getStorageSize() << eth::Instruction::ADD;
-		m_context.appendJumpTo(copyLoopStart);
-		m_context << copyLoopEnd;
-
-		// zero-out leftovers in target
-		// stack: target_ref target_len source_len target_data_pos source_data_pos target_size source_size counter
-		// add counter to target_data_pos
-		m_context << eth::Instruction::DUP5 << eth::Instruction::ADD
-				  << eth::Instruction::SWAP5 << eth::Instruction::POP;
-		// stack: target_ref target_len target_data_pos_updated target_data_pos source_data_pos target_size source_size
-		// add size to target_data_pos to get target_data_end
-		m_context << eth::Instruction::POP << eth::Instruction::DUP3 << eth::Instruction::ADD
-				  << eth::Instruction::SWAP4
-				  << eth::Instruction::POP  << eth::Instruction::POP << eth::Instruction::POP;
-		// stack: target_ref target_data_end target_data_pos_updated
-		clearStorageLoop(*targetBaseType);
+		m_context << eth::Instruction::DUP3 << eth::Instruction::DUP3;
+		copyArrayToStorage(
+			dynamic_cast<ArrayType const&>(*targetBaseType),
+			dynamic_cast<ArrayType const&>(*sourceBaseType)
+		);
 		m_context << eth::Instruction::POP;
-		break;
 	}
-	default:
-		solAssert(false, "Given byte array location not implemented.");
+	else
+	{
+		m_context << eth::Instruction::DUP3;
+		if (_sourceType.getLocation() == ArrayType::Location::Storage)
+			StorageItem(m_context, *sourceBaseType).retrieveValue(SourceLocation(), true);
+		else if (sourceBaseType->isValueType())
+			CompilerUtils(m_context).loadFromMemoryDynamic(*sourceBaseType, true, true, false);
+		else
+			solAssert(false, "Copying of unknown type requested.");
+		m_context << eth::dupInstruction(2 + sourceBaseType->getSizeOnStack());
+		StorageItem(m_context, *targetBaseType).storeValue(*sourceBaseType, SourceLocation(), true);
 	}
+	// increment source
+	m_context
+		<< eth::Instruction::SWAP2
+		<< (_sourceType.getLocation() == ArrayType::Location::Storage ?
+			sourceBaseType->getStorageSize() :
+			sourceBaseType->getCalldataEncodedSize())
+		<< eth::Instruction::ADD
+		<< eth::Instruction::SWAP2;
+	// increment target
+	m_context
+		<< eth::Instruction::SWAP1
+		<< targetBaseType->getStorageSize()
+		<< eth::Instruction::ADD
+		<< eth::Instruction::SWAP1;
+	m_context.appendJumpTo(copyLoopStart);
+	m_context << copyLoopEnd;
+
+	// zero-out leftovers in target
+	// stack: target_ref target_data_end source_data_pos target_data_pos source_data_end
+	m_context << eth::Instruction::POP << eth::Instruction::SWAP1 << eth::Instruction::POP;
+	// stack: target_ref target_data_end target_data_pos_updated
+	clearStorageLoop(*targetBaseType);
+	m_context << eth::Instruction::POP;
 }
 
 void ArrayUtils::clearArray(ArrayType const& _type) const
@@ -290,13 +260,25 @@ void ArrayUtils::clearStorageLoop(Type const& _type) const
 	m_context << eth::Instruction::POP;
 }
 
-void ArrayUtils::convertLengthToSize(ArrayType const& _arrayType) const
+void ArrayUtils::convertLengthToSize(ArrayType const& _arrayType, bool _pad) const
 {
-	if (_arrayType.isByteArray())
-		m_context << u256(31) << eth::Instruction::ADD
-			<< u256(32) << eth::Instruction::SWAP1 << eth::Instruction::DIV;
-	else if (_arrayType.getBaseType()->getStorageSize() > 1)
-		m_context << _arrayType.getBaseType()->getStorageSize() << eth::Instruction::MUL;
+	if (_arrayType.getLocation() == ArrayType::Location::Storage)
+	{
+		if (_arrayType.isByteArray())
+			m_context << u256(31) << eth::Instruction::ADD
+				<< u256(32) << eth::Instruction::SWAP1 << eth::Instruction::DIV;
+		else if (_arrayType.getBaseType()->getStorageSize() > 1)
+			m_context << _arrayType.getBaseType()->getStorageSize() << eth::Instruction::MUL;
+	}
+	else
+	{
+		if (!_arrayType.isByteArray())
+			m_context << _arrayType.getBaseType()->getCalldataEncodedSize() << eth::Instruction::MUL;
+		else if (_pad)
+			m_context << u256(31) << eth::Instruction::ADD
+				<< u256(32) << eth::Instruction::DUP1
+				<< eth::Instruction::SWAP2 << eth::Instruction::DIV << eth::Instruction::MUL;
+	}
 }
 
 void ArrayUtils::retrieveLength(ArrayType const& _arrayType) const
diff --git a/libsolidity/ArrayUtils.h b/libsolidity/ArrayUtils.h
index 73e88340e..31cca8173 100644
--- a/libsolidity/ArrayUtils.h
+++ b/libsolidity/ArrayUtils.h
@@ -60,10 +60,11 @@ public:
 	/// Stack pre: end_ref start_ref
 	/// Stack post: end_ref
 	void clearStorageLoop(Type const& _type) const;
-	/// Converts length to size (multiplies by size on stack), rounds up for byte arrays.
+	/// Converts length to size (number of storage slots or calldata/memory bytes).
+	/// if @a _pad then add padding to multiples of 32 bytes for calldata/memory.
 	/// Stack pre: length
 	/// Stack post: size
-	void convertLengthToSize(ArrayType const& _arrayType) const;
+	void convertLengthToSize(ArrayType const& _arrayType, bool _pad = false) const;
 	/// Retrieves the length (number of elements) of the array ref on the stack. This also
 	/// works for statically-sized arrays.
 	/// Stack pre: reference
diff --git a/libsolidity/Compiler.cpp b/libsolidity/Compiler.cpp
index c880d49d1..241cf0402 100644
--- a/libsolidity/Compiler.cpp
+++ b/libsolidity/Compiler.cpp
@@ -150,7 +150,7 @@ void Compiler::appendConstructor(FunctionDefinition const& _constructor)
 	// copy constructor arguments from code to memory and then to stack, they are supplied after the actual program
 	unsigned argumentSize = 0;
 	for (ASTPointer<VariableDeclaration> const& var: _constructor.getParameters())
-		argumentSize += CompilerUtils::getPaddedSize(var->getType()->getCalldataEncodedSize());
+		argumentSize += var->getType()->getCalldataEncodedSize();
 
 	if (argumentSize > 0)
 	{
@@ -208,8 +208,7 @@ void Compiler::appendCalldataUnpacker(TypePointers const& _typeParameters, bool
 
 	bigint parameterHeadEnd = offset;
 	for (TypePointer const& type: _typeParameters)
-		parameterHeadEnd += type->isDynamicallySized() ? 32 :
-			CompilerUtils::getPaddedSize(type->getCalldataEncodedSize());
+		parameterHeadEnd += type->isDynamicallySized() ? 32 : type->getCalldataEncodedSize();
 	solAssert(parameterHeadEnd <= numeric_limits<unsigned>::max(), "Arguments too large.");
 
 	unsigned stackHeightOfPreviousDynamicArgument = 0;
@@ -228,8 +227,8 @@ void Compiler::appendCalldataUnpacker(TypePointers const& _typeParameters, bool
 					// Retrieve data start offset by adding length to start offset of previous dynamic type
 					unsigned stackDepth = m_context.getStackHeight() - stackHeightOfPreviousDynamicArgument;
 					m_context << eth::dupInstruction(stackDepth) << eth::dupInstruction(stackDepth);
-					ArrayUtils(m_context).convertLengthToSize(*previousDynamicType);
-					m_context << u256(32) << eth::Instruction::MUL << eth::Instruction::ADD;
+					ArrayUtils(m_context).convertLengthToSize(*previousDynamicType, true);
+					m_context << eth::Instruction::ADD;
 				}
 				else
 					m_context << u256(parameterHeadEnd);
@@ -240,7 +239,7 @@ void Compiler::appendCalldataUnpacker(TypePointers const& _typeParameters, bool
 			else
 			{
 				m_context << u256(offset);
-				offset += CompilerUtils::getPaddedSize(type->getCalldataEncodedSize());
+				offset += type->getCalldataEncodedSize();
 			}
 			break;
 		default:
diff --git a/libsolidity/CompilerUtils.cpp b/libsolidity/CompilerUtils.cpp
index 7f8f72ca7..8a26b5d17 100644
--- a/libsolidity/CompilerUtils.cpp
+++ b/libsolidity/CompilerUtils.cpp
@@ -92,7 +92,7 @@ void CompilerUtils::storeInMemoryDynamic(Type const& _type, bool _padToWordBound
 		}
 		else
 		{
-			solAssert(type.getLocation() == ArrayType::Location::Storage, "Memory byte arrays not yet implemented.");
+			solAssert(type.getLocation() == ArrayType::Location::Storage, "Memory arrays not yet implemented.");
 			m_context << eth::Instruction::DUP1 << eth::Instruction::SLOAD;
 			// stack here: memory_offset storage_offset length_bytes
 			// jump to end if length is zero
@@ -177,8 +177,7 @@ void CompilerUtils::computeHashStatic(Type const& _type, bool _padToWordBoundari
 
 unsigned CompilerUtils::loadFromMemoryHelper(Type const& _type, bool _fromCalldata, bool _padToWordBoundaries)
 {
-	unsigned _encodedSize = _type.getCalldataEncodedSize();
-	unsigned numBytes = _padToWordBoundaries ? getPaddedSize(_encodedSize) : _encodedSize;
+	unsigned numBytes = _type.getCalldataEncodedSize(_padToWordBoundaries);
 	bool leftAligned = _type.getCategory() == Type::Category::String;
 	if (numBytes == 0)
 		m_context << eth::Instruction::POP << u256(0);
@@ -202,8 +201,7 @@ unsigned CompilerUtils::loadFromMemoryHelper(Type const& _type, bool _fromCallda
 
 unsigned CompilerUtils::prepareMemoryStore(Type const& _type, bool _padToWordBoundaries) const
 {
-	unsigned _encodedSize = _type.getCalldataEncodedSize();
-	unsigned numBytes = _padToWordBoundaries ? getPaddedSize(_encodedSize) : _encodedSize;
+	unsigned numBytes = _type.getCalldataEncodedSize(_padToWordBoundaries);
 	bool leftAligned = _type.getCategory() == Type::Category::String;
 	if (numBytes == 0)
 		m_context << eth::Instruction::POP;
diff --git a/libsolidity/CompilerUtils.h b/libsolidity/CompilerUtils.h
index 24ebbc81f..043de41dd 100644
--- a/libsolidity/CompilerUtils.h
+++ b/libsolidity/CompilerUtils.h
@@ -71,9 +71,6 @@ public:
 	/// Stack pre: memory_offset value...
 	/// Stack post: (memory_offset+length)
 	void storeInMemoryDynamic(Type const& _type, bool _padToWordBoundaries = true);
-	/// @returns _size rounded up to the next multiple of 32 (the number of bytes occupied in the
-	///          padded calldata)
-	static unsigned getPaddedSize(unsigned _size) { return ((_size + 31) / 32) * 32; }
 
 	/// Moves the value that is at the top of the stack to a stack variable.
 	void moveToStackVariable(VariableDeclaration const& _variable);
diff --git a/libsolidity/ExpressionCompiler.cpp b/libsolidity/ExpressionCompiler.cpp
index 793bd55e4..b967f9ed5 100644
--- a/libsolidity/ExpressionCompiler.cpp
+++ b/libsolidity/ExpressionCompiler.cpp
@@ -74,7 +74,7 @@ void ExpressionCompiler::appendStateVariableAccessor(VariableDeclaration const&
 	{
 		// move offset to memory
 		CompilerUtils(m_context).storeInMemory(length);
-		unsigned argLen = CompilerUtils::getPaddedSize(paramType->getCalldataEncodedSize());
+		unsigned argLen = paramType->getCalldataEncodedSize();
 		length -= argLen;
 		m_context << u256(argLen + 32) << u256(length) << eth::Instruction::SHA3;
 
@@ -779,8 +779,9 @@ bool ExpressionCompiler::visit(IndexAccess const& _indexAccess)
 		else
 		{
 			u256 elementSize =
-				location == ArrayType::Location::Storage ? arrayType.getBaseType()->getStorageSize() :
-				CompilerUtils::getPaddedSize(arrayType.getBaseType()->getCalldataEncodedSize());
+				location == ArrayType::Location::Storage ?
+					arrayType.getBaseType()->getStorageSize() :
+					arrayType.getBaseType()->getCalldataEncodedSize();
 			solAssert(elementSize != 0, "Invalid element size.");
 			if (elementSize > 1)
 				m_context << elementSize << eth::Instruction::MUL;
@@ -798,8 +799,8 @@ bool ExpressionCompiler::visit(IndexAccess const& _indexAccess)
 			switch (location)
 			{
 			case ArrayType::Location::CallData:
-				// no lvalue
-				CompilerUtils(m_context).loadFromMemoryDynamic(*arrayType.getBaseType(), true, true, false);
+				if (arrayType.getBaseType()->isValueType())
+					CompilerUtils(m_context).loadFromMemoryDynamic(*arrayType.getBaseType(), true, true, false);
 				break;
 			case ArrayType::Location::Storage:
 				setLValueToStorageItem(_indexAccess);
@@ -1018,7 +1019,7 @@ void ExpressionCompiler::appendExternalFunctionCall(FunctionType const& _functio
 	//@todo only return the first return value for now
 	Type const* firstType = _functionType.getReturnParameterTypes().empty() ? nullptr :
 							_functionType.getReturnParameterTypes().front().get();
-	unsigned retSize = firstType ? CompilerUtils::getPaddedSize(firstType->getCalldataEncodedSize()) : 0;
+	unsigned retSize = firstType ? firstType->getCalldataEncodedSize() : 0;
 	m_context << u256(retSize) << u256(0);
 
 	if (bare)
@@ -1048,8 +1049,8 @@ void ExpressionCompiler::appendExternalFunctionCall(FunctionType const& _functio
 	if (_functionType.gasSet())
 		m_context << eth::dupInstruction(m_context.baseToCurrentStackOffset(gasStackPos));
 	else
-		// send all gas except for the 21 needed to execute "SUB" and "CALL"
-		m_context << u256(_functionType.valueSet() ? 6741 : 41) << eth::Instruction::GAS << eth::Instruction::SUB;
+		// send all gas except for the 41 / 6741 needed to execute "SUB" and "CALL"
+		m_context << u256(41 + (_functionType.valueSet() ? 6700 : 0)) << eth::Instruction::GAS << eth::Instruction::SUB;
 	m_context << eth::Instruction::CALL;
 	auto tag = m_context.appendConditionalJump();
 	m_context << eth::Instruction::STOP << tag;	// STOP if CALL leaves 0.
@@ -1080,7 +1081,7 @@ void ExpressionCompiler::appendArgumentsCopyToMemory(
 		appendTypeConversion(*_arguments[i]->getType(), *expectedType, true);
 		bool pad = _padToWordBoundaries;
 		// Do not pad if the first argument has exactly four bytes
-		if (i == 0 && pad && _padExceptionIfFourBytes && expectedType->getCalldataEncodedSize() == 4)
+		if (i == 0 && pad && _padExceptionIfFourBytes && expectedType->getCalldataEncodedSize(false) == 4)
 			pad = false;
 		appendTypeMoveToMemory(*expectedType, pad);
 	}
diff --git a/libsolidity/Types.cpp b/libsolidity/Types.cpp
index 96feefff5..3f0445288 100644
--- a/libsolidity/Types.cpp
+++ b/libsolidity/Types.cpp
@@ -570,11 +570,12 @@ bool ArrayType::operator==(Type const& _other) const
 	return isDynamicallySized() || getLength()  == other.getLength();
 }
 
-unsigned ArrayType::getCalldataEncodedSize() const
+unsigned ArrayType::getCalldataEncodedSize(bool _padded) const
 {
 	if (isDynamicallySized())
 		return 0;
-	bigint size = bigint(getLength()) * (isByteArray() ? 1 : getBaseType()->getCalldataEncodedSize());
+	bigint size = bigint(getLength()) * (isByteArray() ? 1 : getBaseType()->getCalldataEncodedSize(_padded));
+	size = ((size + 31) / 32) * 32;
 	solAssert(size <= numeric_limits<unsigned>::max(), "Array size does not fit unsigned.");
 	return unsigned(size);
 }
diff --git a/libsolidity/Types.h b/libsolidity/Types.h
index afecf3c8e..0915e0da6 100644
--- a/libsolidity/Types.h
+++ b/libsolidity/Types.h
@@ -120,8 +120,10 @@ public:
 
 	/// @returns number of bytes used by this type when encoded for CALL, or 0 if the encoding
 	/// is not a simple big-endian encoding or the type cannot be stored in calldata.
-	/// Note that irrespective of this size, each calldata element is padded to a multiple of 32 bytes.
-	virtual unsigned getCalldataEncodedSize() const { return 0; }
+	/// If @a _padded then it is assumed that each element is padded to a multiple of 32 bytes.
+	virtual unsigned getCalldataEncodedSize(bool _padded) const { (void)_padded; return 0; }
+	/// Convenience eversion of @see getCalldataEncodedSize(bool)
+	unsigned getCalldataEncodedSize() const { return getCalldataEncodedSize(true); }
 	/// @returns true if the type is dynamically encoded in calldata
 	virtual bool isDynamicallySized() const { return false; }
 	/// @returns number of bytes required to hold this value in storage.
@@ -176,7 +178,7 @@ public:
 
 	virtual bool operator==(Type const& _other) const override;
 
-	virtual unsigned getCalldataEncodedSize() const override { return m_bits / 8; }
+	virtual unsigned getCalldataEncodedSize(bool _padded = true) const override { return _padded ? 32 : m_bits / 8; }
 	virtual bool isValueType() const override { return true; }
 
 	virtual MemberList const& getMembers() const { return isAddress() ? AddressMemberList : EmptyMemberList; }
@@ -247,7 +249,7 @@ public:
 	virtual bool isExplicitlyConvertibleTo(Type const& _convertTo) const override;
 	virtual bool operator==(Type const& _other) const override;
 
-	virtual unsigned getCalldataEncodedSize() const override { return m_bytes; }
+	virtual unsigned getCalldataEncodedSize(bool _padded) const override { return _padded && m_bytes > 0 ? 32 : m_bytes; }
 	virtual bool isValueType() const override { return true; }
 
 	virtual std::string toString() const override { return "string" + dev::toString(m_bytes); }
@@ -271,7 +273,7 @@ public:
 	virtual TypePointer unaryOperatorResult(Token::Value _operator) const override;
 	virtual TypePointer binaryOperatorResult(Token::Value _operator, TypePointer const& _other) const override;
 
-	virtual unsigned getCalldataEncodedSize() const { return 1; }
+	virtual unsigned getCalldataEncodedSize(bool _padded) const { return _padded ? 32 : 1; }
 	virtual bool isValueType() const override { return true; }
 
 	virtual std::string toString() const override { return "bool"; }
@@ -302,7 +304,7 @@ public:
 	virtual bool isImplicitlyConvertibleTo(Type const& _convertTo) const override;
 	virtual TypePointer unaryOperatorResult(Token::Value _operator) const override;
 	virtual bool operator==(const Type& _other) const override;
-	virtual unsigned getCalldataEncodedSize() const override;
+	virtual unsigned getCalldataEncodedSize(bool _padded) const override;
 	virtual bool isDynamicallySized() const { return m_hasDynamicLength; }
 	virtual u256 getStorageSize() const override;
 	virtual unsigned getSizeOnStack() const override;
diff --git a/test/SolidityEndToEndTest.cpp b/test/SolidityEndToEndTest.cpp
index dae0ca03a..c8664b4d0 100644
--- a/test/SolidityEndToEndTest.cpp
+++ b/test/SolidityEndToEndTest.cpp
@@ -3008,6 +3008,68 @@ BOOST_AUTO_TEST_CASE(bytes_index_access)
 	BOOST_CHECK(callContractFunction("storageWrite()") == encodeArgs(0x193));
 }
 
+BOOST_AUTO_TEST_CASE(array_copy_calldata_storage)
+{
+	char const* sourceCode = R"(
+		contract c {
+			uint[9] m_data;
+			uint[] m_data_dyn;
+			uint8[][] m_byte_data;
+			function store(uint[9] a, uint8[3][] b) external returns (uint8) {
+				m_data = a;
+				m_data_dyn = a;
+				m_byte_data = b;
+				return b[3][1]; // note that access and declaration[ are reversed to each other
+			}
+			function retrieve() returns (uint a, uint b, uint c, uint d, uint e, uint f, uint g) {
+				a = m_data.length;
+				b = m_data[7];
+				c = m_data_dyn.length;
+				d = m_data_dyn[7];
+				e = m_byte_data.length;
+				f = m_byte_data[3].length;
+				g = m_byte_data[3][1];
+			}
+		}
+	)";
+	compileAndRun(sourceCode);
+	BOOST_CHECK(callContractFunction("store(uint256[9],uint8[3][])", encodeArgs(
+		21, 22, 23, 24, 25, 26, 27, 28, 29, // a
+		4, // size of b
+		1, 2, 3, // b[0]
+		11, 12, 13, // b[1]
+		21, 22, 23, // b[2]
+		31, 32, 33 // b[3]
+	)) == encodeArgs(32));
+	BOOST_CHECK(callContractFunction("retrieve()") == encodeArgs(
+		9, 28, 9, 28,
+		4, 3, 32));
+}
+
+BOOST_AUTO_TEST_CASE(array_copy_nested_array)
+{
+	char const* sourceCode = R"(
+		contract c {
+			uint[4][] a;
+			uint[5][] b;
+			uint[][] c;
+			function test(uint[2][] d) external returns (uint) {
+				a = d;
+				b = a;
+				c = b;
+				return c[1][1] | c[1][2] | c[1][3] | c[1][4];
+			}
+		}
+	)";
+	compileAndRun(sourceCode);
+	BOOST_CHECK(callContractFunction("test(uint256[2][])", encodeArgs(
+		3,
+		7, 8,
+		9, 10,
+		11, 12
+	)) == encodeArgs(10));
+}
+
 BOOST_AUTO_TEST_CASE(pass_dynamic_arguments_to_the_base)
 {
 	char const* sourceCode = R"(
diff --git a/test/solidityExecutionFramework.h b/test/solidityExecutionFramework.h
index 4ef9bfdc8..4ed3854b1 100644
--- a/test/solidityExecutionFramework.h
+++ b/test/solidityExecutionFramework.h
@@ -178,7 +178,7 @@ protected:
 	Address m_contractAddress;
 	eth::State m_state;
 	u256 const m_gasPrice = 100 * eth::szabo;
-	u256 const m_gas = 1000000;
+	u256 const m_gas = 100000000;
 	bytes m_output;
 	eth::LogEntries m_logs;
 };