Type-erased InplaceUniquePrintable benefits from noexcept

Previously on this blog: “Type-erased UniquePrintable and PrintableRef (2020-11-24).

Now, here’s a simple type-erased InplaceUniquePrintable (owning, value-semantic, move-only, non-heap-allocating). I don’t claim that this one is bang-out-in-five-minutes-able, but on the other hand, it’s the technique most commonly used in practice. The basic idea is to separate the data of the erased type from its required affordances or behaviors — printing, relocating, and destroying. The data is stored as plain old bytes; the behaviors are stored as function pointers, which point to lambdas initialized in the constructor. I discuss this idea further in “Back to Basics: Type Erasure” (CppCon 2019).


Godbolt. Notice that I completely arbitrarily chose sizeof(std::string) for the size of the internal buffer.

class InplaceUniquePrintable {
    alignas(std::string) unsigned char data_[sizeof(std::string)];
    void (*print_)(std::ostream&, const void *) = nullptr;
    void (*relocate_)(void *, const void *) noexcept = nullptr;
    void (*destroy_)(void *) noexcept = nullptr;
    template<class T, class StoredT = std::decay_t<T>>
    InplaceUniquePrintable(T&& t)
        // constraints are left as an exercise for the reader
        static_assert(sizeof(StoredT) <= sizeof(std::string));
        static_assert(alignof(StoredT) <= alignof(std::string));
        ::new ((void*)data_) StoredT(std::forward<T>(t));
        print_ = [](std::ostream& os, const void *data) {
            os << *(const StoredT*)data;
        relocate_ = [](void *dst, const void *src) noexcept {
            ::new (dst) StoredT(std::move(*(StoredT*)src));
        destroy_ = [](void *src) noexcept {

    InplaceUniquePrintable(InplaceUniquePrintable&& rhs) noexcept {
        print_ = std::exchange(rhs.print_, nullptr);
        relocate_ = std::exchange(rhs.relocate_, nullptr);
        destroy_ = std::exchange(rhs.destroy_, nullptr);
        relocate_(data_, rhs.data_);

    friend void swap(InplaceUniquePrintable& lhs, InplaceUniquePrintable& rhs) noexcept {
        std::swap(lhs.data_, rhs.data_);
        std::swap(lhs.print_, rhs.print_);
        std::swap(lhs.relocate_, rhs.relocate_);
        std::swap(lhs.destroy_, rhs.destroy_);

    InplaceUniquePrintable& operator=(InplaceUniquePrintable&& rhs) noexcept {
        auto copy = std::move(rhs);
        swap(*this, copy);
        return *this;

    ~InplaceUniquePrintable() {
        if (destroy_ != nullptr) {

    friend std::ostream& operator<<(std::ostream& os, const InplaceUniquePrintable& self) {
        self.print_(os, self.data_);
        return os;

Example of use

void printit(InplaceUniquePrintable p) {
    std::cout << "The printable thing was: " << p << "." << std::endl;

int main() {
    printit("hello world");

Further improvements

Notice that we require our StoredT to fit in the data_ buffer, and assert-fail (at compile time) if that requirement is not met. However, if you’re okay with heap-allocation, it’s easy to change those static_asserts into if constexprs; then you can store even a large printable object, simply by wrapping it in a heap-allocated pointer. We just change InplaceUniquePrintable’s constructor to look like this, and everything else can stay the same (Godbolt):

constexpr bool fits_in_buffer =
    sizeof(StoredT) <= sizeof(std::string) &&
    alignof(StoredT) <= alignof(std::string) &&

if constexpr (fits_in_buffer) {
    ::new ((void*)data_) StoredT(std::forward<T>(t));
    print_ = ~~~ as before ~~~
} else {
    using StoredPtr = StoredT*;
    auto p = std::make_unique<StoredT>(std::forward<T>(t));
    ::new ((void*)data_) StoredPtr(p.release());
    print_ = [](std::ostream& os, const void *data) {
        const StoredT *p = *(StoredPtr*)data;
        os << *p;
    relocate_ = [](void *dst, const void *src) noexcept {
        std::memcpy(dst, src, sizeof(StoredPtr));
    destroy_ = [](void *src) noexcept {
        StoredT *p = *(StoredPtr*)src;
        delete p;

Finally, we should avoid generating a non-trivial destroy_ when StoredT is trivially destructible; and we might consider doing the same for relocate_ when StoredT is known to be trivially relocatable. (Godbolt; backup.)

Benefits from noexcept

Anyone who checks out this blog’s #exception-handling tag will know that I’m not big on noexcept. My rule of thumb is to put it on move constructors (to undo the vector pessimization), optionally on swap and operator=, and (by default) nowhere else. So you might be surprised to see that I’ve marked the destroy_ and relocate_ function pointers as noexcept!

This is due to an awesome tip by Adrian Vogelsgesang, who points out that adding noexcept on function pointers can sometimes allow them to be tail-called from noexcept contexts, where an ordinary non-noexcept function pointer can’t be. Here’s a simple example:

struct Test {
    InplaceUniquePrintable p_;
Test::~Test() = default;

~Test() here is going to call ~InplaceUniquePrintable(), which is going to call *destroy_. Recall that destructors (unlike other functions) are implicitly noexcept, and recall that noexcept means “Insert a firewall here, that any exceptions thrown from lower levels will smack into and terminate the program.” (That is, it’s not a promise to the compiler that you won’t throw; it’s an instruction to insert extra code to prevent throws from leaving this scope.) When ~InplaceUniquePrintable() calls *destroy_, the code generator must decide whether it needs to insert one of those firewalls or not. If destroy_ is an ordinary function pointer, then it must. But if destroy_ is marked noexcept, then the compiler can assume any exceptions will have been caught one level down, and no additional firewall is needed at this level — so this function doesn’t need a stack frame — so we can tail-call-optimize the call to *destroy_!

The takeaway is that if you have a function pointer (or virtual function, I guess) that you intend to call from a noexcept context such as a destructor or move-constructor, then it is perhaps a performance benefit to mark that pointer’s data type with noexcept. Here’s GCC’s codegen for Test::~Test() with the code as presented above (Godbolt):

// void (*destroy_)(void *) noexcept = nullptr;

  movq 48(%rdi), %rax
  testq %rax, %rax
  je .L1
  jmp *%rax  # TAIL CALL

And here it is without noexcept:

// void (*destroy_)(void *) = nullptr;

  movq 48(%rdi), %rax
  testq %rax, %rax
  je .L7
  subq $8, %rsp
  call *%rax  # NO TAIL CALL
  addq $8, %rsp

We get the same kind of benefit in Test(Test&&); and notice that I positioned the indirect call to *relocate_ in tail position for that reason. You and I know that *relocate_ will never throw, because we can see the code of the lambda assigned to it a few dozen lines above; but because the call is indirect through a function pointer, the compiler itself cannot see that *relocate_ will never throw (and therefore can be tail-called) unless we mark its type with noexcept.

See also:

Posted 2022-07-30