Using single header outcome-basic breaks clang compilation with placement deallocation functions

[apmanol]

Hello,

something very strange, with clang++ using the outcome-basic

clang++ --version 
clang version 15.0.7 (Fedora 15.0.7-2.fc37)
Target: x86_64-redhat-linux-gnu
Thread model: posix
InstalledDir: /usr/bin

#include "outcome-basic.hpp"

// A std::unique_ptr deleter that deletes memory allocated via ::operator new.
#if defined(__cpp_sized_deallocation)
class TrivialDeleter {
 public:
  explicit TrivialDeleter(std::size_t size) : size_(size) {}

  void operator()(void* target) const {
    ::operator delete(target, size_);
  }

 private:
  std::size_t size_;
};
#else   // __cpp_sized_deallocation
class TrivialDeleter {
 public:
  explicit TrivialDeleter(std::size_t) {}

  void operator()(void* target) const { ::operator delete(target); }
};
#endif  // __cpp_sized_deallocation
int main()
{
  TrivialDeleter deleter(4u);
  return 0;
}

with error:

lang++  -std=c++17 ll.cc -I/projects/icom/babel/thirdparty/abseil-cpp/ 
ll.cc:16:5: error: no matching function for call to 'operator delete'
    ::operator delete(target, size_);
    ^~~~~~~~~~~~~~~~~
/usr/bin/../lib/gcc/x86_64-redhat-linux/12/../../../../include/c++/12/new:144:6: note: candidate function not viable: no known conversion from 'const std::size_t' (aka 'const unsigned long') to 'const std::nothrow_t' for 2nd argument
void operator delete(void*, const std::nothrow_t&) _GLIBCXX_USE_NOEXCEPT
     ^
/usr/bin/../lib/gcc/x86_64-redhat-linux/12/../../../../include/c++/12/new:153:6: note: candidate function not viable: no known conversion from 'const std::size_t' (aka 'const unsigned long') to 'std::align_val_t' for 2nd argument
void operator delete(void*, std::align_val_t)
     ^
/usr/bin/../lib/gcc/x86_64-redhat-linux/12/../../../../include/c++/12/new:180:13: note: candidate function not viable: no known conversion from 'const std::size_t' (aka 'const unsigned long') to 'void *' for 2nd argument
inline void operator delete  (void*, void*) _GLIBCXX_USE_NOEXCEPT { }
            ^
/usr/bin/../lib/gcc/x86_64-redhat-linux/12/../../../../include/c++/12/new:130:6: note: candidate function not viable: requires 1 argument, but 2 were provided
void operator delete(void*) _GLIBCXX_USE_NOEXCEPT
     ^
/usr/bin/../lib/gcc/x86_64-redhat-linux/12/../../../../include/c++/12/new:155:6: note: candidate function not viable: requires 3 arguments, but 2 were provided

In godbolt the error appears in clang <16

[ned14]

I would personally suspect this is a bug in clang? It's had crap QoI recently. So bad my former employer was actively thinking of dropping clang support entirely, because it's become a real pain in the ass.

Note that telling clang 14 or 15 that it is in C++ 17 or later works around the problem.

@BurningEnlightenment what do you think?

[ned14]

Hmm, GCC is also broken in C++ 14. That makes it more serious.

[apmanol]

Maybe some compiler definition is altered somehow?
The compilation of code without the include of the header works.

[ned14]

Why transfer to quickcpplib? This is surely an Outcome problem?

[BurningEnlightenment]

So, there are multiple issues blending together:

First of all, your feature check is broken, the compiler is allowed to define __cpp_sized_deallocation to 0 for older standards:

-#if defined(__cpp_sized_deallocation)
+#if __cpp_sized_deallocation >= 201309L

Then there is the fact that quickcpplib (included with outcome) does the following:

quickcpplib/include/quickcpplib/cpp_feature.h

Lines 208 to 210 in 59105c6

	#if !defined(__cpp_sized_deallocation)
	#define __cpp_sized_deallocation 190000
	#endif

And finally sized (de-)allocation support is disabled in clang by default. You must explicitly enable it with -fsized-deallocation.

[BurningEnlightenment]

One could argue that quickcpplib shouldn't touch the feature macros, but otherwise I think this can be closed as invalid/wontfix/working as intended.

Anyways, thank you for the BR!

[ned14]

Ah, but look at https://godbolt.org/z/b3P33qYzW! The OP has revealed a bug!

I got started on a fix here, but I'll need to wait until work is finished before I can look into it properly, because it didn't "just work".

[BurningEnlightenment]

Probably due to a reduction of transitively included headers in libstdc++.

I observed another compile failure on my machine when I started testing with the following reprex and libstdc++12 have been fooled by using an outdated vcpkg port which didn't include ned14/outcome@5daef0e 😵‍💫

But these are strictly different issues.

[BurningEnlightenment]

I'm going ahead and close this as it is working as intended. However, if you feel that quickcpplib shouldn't define the feature test macros based on the __cplusplus value, you may open an enhancement issue here.

Thanks again for the BR!