CppCon performance not quite true?

[kelbon]

It strange issue, but there are CppCon video about libc++ shipped with clang:

https://www.youtube.com/watch?v=iw8hqKftP4I

And in first 9 minutes speaker says, that this code should compile on clang with libc++:

template<typename V, typename E>
struct EX {
  union U {
    [[no_unique_address]] V v;
    [[no_unique_address]] E e;
  };
  [[no_unique_address]] U u_;
  bool has_val_;
};

struct mystruct {
  int x;
  bool y;
  char z;
};

static_assert(sizeof(EX<mystruct, int>) == 8); // 12 in reality

But its not. Sizeof == 8, and i agree with compiler here

I also checked with std::expected itself:

#include <expected>

struct mystruct {
  int x;
  bool y;
  char z;
};


mystruct foo();

std::expected<mystruct, int> compute() {
  return foo();
}

static_assert(sizeof(std::expected<mystruct, int>) == 12); // again, 12

https://godbolt.org/z/jPfbMqchW

A talk generally good, but this first part (9 minutes) confuses me.

[kelbon]

@huixie90 can you please explain what im missing?

[frederick-vs-ja]

I think this is about the "POD for the purpose of layout" property in Itanium C++ ABI. You can add default member initializer to any data member of mystruct (which makes mystruct not POD) to see the size change.

mystruct in your example is "POD for the purpose of layout" (roughly equivalent to "POD type in the C++98"). As a result, Itanium ABI disallows reusing its tail padding bytes, because pre-CWG43 standard requirement effectively so disallowed for (C++98) POD types and Itanium ABI decided to be compatible with the unfixed specification.

On the other hand, Clang seemingly disallows reusing tail paddings for some classes that are not POD for the purpose of layout. This is tracked in #50766. Anyway, the question doesn't seem actually related to libc++.

Edit: The above was a bit wrong on "POD for the purpose of layout".

[kelbon]

@frederick-vs-ja thank you for answer. It seems like very strange opimitzation, rlt not obvious