Write-only variables
Earlier this week we’ve had a nice chat with my colleague about an interesting task, which I’ve found quite entertaining and, most important, a nice interview topic.
In plain C and C++, we use memory-mapped I/O to access registers, which means that every register has a dedicated address by which we can access it. Here’s an example of a (partial) memory layout from the documentation.
There are basically three types of registers:
- Read-write
- Read-only
- Write-only
There are no problems in defining both read-write and write-only registers via references or pointers, although I’m not a big fan of a bare register access (as I’ve discussed this earlier, I prefer to add a couple of layers of indirection to make my code less error-prone):
volatile auto &read_write = *reinterpret_cast<std::uint32_t*>(register_address);
const volatile auto &read_only = *reinterpret_cast<std::uint32_t*>(register_address);
To make variable read-only we simply add the const-qualifier to the reference declaration. During the discussion of the previous article, a couple of people were puzzled about the reinterpret_cast part. This operator is the only way to convert address (as a fixed number from the documentation) to the pointer. Via the pointer, we will access the underlying data.
But there is no qualifier in the language to perform the write-only access. And this is where we unleash the true power of the C++: custom types or classes. But before we dive into solving this task, let’s measure the initial approach performance. I’ll compile the code with the -O2 optimization and -std=c++17 flag:
/// Code:
constexpr std::size_t register_address = 123456;
const volatile auto &read_only = *reinterpret_cast<std::uint32_t*>(register_address);
void foo() {
auto dst = read_only;
}
/// Disassembly:
foo():
mov eax, DWORD PTR ds:123456
ret
Only two instructions, which is hard to beat, but we’re not afraid. Let’s take the advantage of the known register address, so we can use constexpr in our task. First of all, we will declare a class with a private reference. Read and write operations may be implemented as a Get() and Set() methods:
class Register {
private:
static volatile inline std::uint32_t &ref = *reinterpret_cast<std::uint32_t*>(register_address);
public:
static std::uint32_t Get(){
return ref;
}
static void Set(std::uint32_t val){
ref = val;
}
};
There are a lot of qualifiers for the internal reference, to make it clear:
staticmakes this reference independent from the structure objectvolatileis there to indicate that the value of the reference may change during the program execution and the compiler should not cache itinline. A great C++17 feature, allowing to declare static variables without the need for an external .cpp file. Have you ever tried to initialize astd::mapinside a class? Now you can do it without any additional fuss.
And here’s the compiler output:
/// Code:
void RegGet() {
auto dst = Register::Get();
}
void RegSet(std::uint32_t val) {
Register::Set(val);
}
/// Disassembly:
RegGet():
mov eax, DWORD PTR ds:123456
ret
RegSet(unsigned int):
mov DWORD PTR ds:123456, edi
ret
Exactly the same disassembly and performance, yay! Although, there might be a problem with an old codebase you’ll try to update because it is quite difficult to search and replace all of the assignments to the Get() methods. I don’t know any C++ refactoring tool which is capable of doing such a thing, and, most importantly, I even have no idea how one might be implemented.
To solve this we will add a couple of overloads to our class:
- Assignment operator for the write operations
std::uint32_tcasting operator for the read operationsTcasting operator to raise a compile-time error when you try to read register into some type other thanstd::uint32_t
class Register {
// ...
Register& operator=(std::uint32_t val){
ref = val;
return *this;
}
template <typename T>
operator T() const{
static_assert(std::is_same_v<T, std::uint32_t>, "You should assign this register to the std::uint32_t value");
return T();
}
operator std::uint32_t() const {
return ref;
}
} reg;
Since the assignment and casting operators cannot be static, we have to create an object of our class to use it:
/// Code
void RegAssign() {
std::uint32_t dst = reg;
}
void RegGetCast(std::uint32_t val) {
reg = val;
}
/// Disassembly:
RegAssign():
mov eax, DWORD PTR ds:123456
ret
RegGetCast(unsigned int):
mov DWORD PTR ds:123456, edi
ret
And we still get the same performance with the familiar syntax! There is one more tweak we should do to make this class ready for use: get rid of the hard-coded address in the reference. Bear in mind, that we should make it compile-time friendly. This may be done with the help of the templates. A couple of changes:
template <std::size_t address>
class Register {
private:
static volatile inline std::uint32_t &ref = *reinterpret_cast<std::uint32_t*>(address);
//...
};
Register<register_address> reg;
And we’re done for today. You may fiddle with the code here. It really looks to me like a nice little task to chat about.
Update:
Here’s the write-only part:
template <std::size_t address>
class Register {
private:
static volatile inline std::uint32_t &ref = *reinterpret_cast<std::uint32_t*>(address);
public:
static void Set(std::uint32_t val){
ref = val;
}
Register& operator=(std::uint32_t val){
ref = val;
return *this;
}
};
And the read-only:
template <std::size_t address>
class Register {
private:
static volatile inline std::uint32_t &ref = *reinterpret_cast<std::uint32_t*>(address);
public:
static std::uint32_t Get(){
return ref;
}
template <typename T>
operator T() const{
static_assert(std::is_same_v<T, std::uint32_t>, "You should assign this register to the std::uint32_t value");
return T();
}
operator std::uint32_t() const {
return ref;
}
};
Update2:
A couple of days has passed and I has found an even better (more flexible and universal) solution. We will create the base class with all of the operations (read and write) declared protected. The code is here to play.
template <std::size_t address>
class RegisterBase {
private:
static volatile inline std::uint32_t &ref = *reinterpret_cast<std::uint32_t*>(address);
protected:
static std::uint32_t Get(){
return ref;
}
static void Set(std::uint32_t val){
ref = val;
}
RegisterBase& operator=(std::uint32_t val){
ref = val;
return *this;
}
template <typename T>
operator T() const{
static_assert(std::is_same_v<T, std::uint32_t>, "You should assign this register to the std::uint32_t value");
return T();
}
operator std::uint32_t() const {
return ref;
}
};
To get read-only and write-only we will inherit from the base class and change the visibility of certain methods:
template <std::size_t address>
class RegisterRead : public RegisterBase<address> {
public:
static std::uint32_t Get(){
return RegisterBase<address>::Get();
}
template <typename T>
operator T() const{
return RegisterBase<address>::operator T();
}
operator std::uint32_t() const{
return RegisterBase<address>::operator std::uint32_t();
}
};
template <std::size_t address>
class RegisterWrite : public RegisterBase<address> {
public:
static void Set(std::uint32_t val){
return RegisterBase<address>::Set(val);
}
RegisterWrite& operator=(std::uint32_t val){
return static_cast<RegisterWrite&>(RegisterBase<address>::operator=(val));
}
};
Note the static_cast<> part. Because RegisterBase<address>::operator= returns a reference to the base class, we should cast it to the reference to the derived class object.
The following code is pretty much the same:
RegisterRead<register_address> reg_read;
RegisterWrite<register_address> reg_write;
void foo() {
auto dst = read_only;
}
void RegAssign() {
std::uint32_t dst = reg_read;
}
void RegGet() {
auto dst = reg_read.Get();
}
void RegSet(std::uint32_t val) {
reg_write.Set(val);
}
void RegGetCast(std::uint32_t val) {
reg_write = val;
}