NDK r15 missing c++17 aligned allocation operators

[DoDoENT]

.../ndk-bundle/sources/cxx-stl/llvm-libc++/include/memory:2540: error: undefined reference to 'operator delete(void*, std::align_val_t)'
.../ndk-bundle/sources/cxx-stl/llvm-libc++/include/memory:3149: error: undefined reference to 'operator new(unsigned int, std::align_val_t)'
.../ndk-bundle/sources/cxx-stl/llvm-libc++/include/memory:2540: error: undefined reference to 'operator delete(void*, std::align_val_t)'
.../ndk-bundle/sources/cxx-stl/llvm-libc++/include/memory:3149: error: undefined reference to 'operator delete(void*, std::align_val_t)'

This happens to our program after updating to NDK r15 - only on armeabi-v7a. On arm64-v8a it works correctly.

My current workaround is to add this file to our codebase when using ARM7 slice:

#include <boost/align/aligned_alloc.hpp>

#include <cstddef>

namespace std
{
enum class align_val_t : size_t {};
}

void* operator new( std::size_t count, std::align_val_t alignment )
{
    return boost::alignment::aligned_alloc( static_cast< std::size_t >( alignment ), count );
}

void operator delete( void* ptr, std::align_val_t ) noexcept
{
    boost::alignment::aligned_free( ptr );
}

[DanAlbert]

That's surprising. There are some known failures in the libc++ tests right now, but none of them like that. Clang or GCC?

[DoDoENT]

Clang

[DanAlbert]

We'll be shipping an r15b to cover a misdefined API (libsync doesn't match its header). I'm going to investigate this today and see if there's a fix we can make at the same time.

[DanAlbert]

Looks like this just has to do with whether or not the compiler decides to use it or not. This test case:

struct alignas(16) vec4 {
  float x, y, z, w;
};

int main(int argc, char** argv) {
  auto vectors = new vec4[16];
  return 0;
}

Only works on LP64 ABIs (when built with -std=c++1z, of course). If I change the alignment to 64, it fails on LP64 too.

So, the reason this is happening is because we updated to a Clang that can take advantage of these, but we don't have a new enough libc++ to provide them. I got us up to matching the platform again last week, and was already planning to do a full update to HEAD for r16. Once that's done, this should be good to go.

This isn't anything we can do something about for r15b though. Updating libc++ is much too high risk for a hotfix.

You can maybe work around this by targeting C++14 instead. A lot of new stuff in libc++ gets backported to earlier standard versions, so other stuff you're trying to use might be available anyway. Unfortunately C++1z is still all experimental, so these sorts of things are somewhat expected.

[DoDoENT]

The workaround from my first post works OK for us. Thanks for letting us know the issue is also present on ARM64.

Since we use boost as a fallback option for missing c++17 features on all platforms, this workaround is OK for us, i.e. it is not a showstopper, unlike issues #318 and #313.

[DanAlbert]

Good to know.

FWIW, the former should be unblocked the next time we update the compiler (was https://reviews.llvm.org/D33287), and the latter is something on my list for r16 (though thanks for the reminder, I'd forgotten to actually assign it to the milestone).

[DanAlbert]

The libc++ update is in r16 now. It looks like we need a newer Clang to actually get all the libc++ tests working in C++1z mode (which we planned to have anyway, just noting that we don't have it yet).

[tcamargo]

Hi All,

This issue seems to be back. Using NDK21d.

../../../../libretro/bin/mame_mame/liboptional.a(netlist.o):netlist.cpp:function netlist_mame_analog_output_device::pre_parse_action(netlist::nlparse_t&): error: undefined reference to 'aligned_alloc'

Compiling using std=c++14. Do I have to pass some extra lib when linking?

[DanAlbert]

Not enough information for me to know. The test case above works fine for me. File a bug with a repro case.

[tcamargo]

Not enough information for me to know. The test case above works fine for me. File a bug with a repro case.

Thanks! Mame test cases are always a pain to reproduce. I will try to create a small tc.