[llvm-dev] Utilising no-unique-address

[Nathan James via llvm-dev]

Hi,

About a month ago I pushed a patch to change BumpPtrAllocatorImpl to
use the empty base optimization to reduce the size of the
allocator(always a win). However looking at clangs support for the
c++20 attribute `no_unique_address`, It appears that we support that in
previous standards (including c++14). 

Can anyone see any issue to utilizing this attribute throughout the
codebase behind a macro in `llvm/Support/Compiler.h`. 
`
#if LLVM_HAS_CPP_ATTRIBUTE(no_unique_address) 
#define LLVM_NO_UNIQUE_ADDRESS [[no_unique_address]]
#else
#define LLVM_NO_UNIQUE_ADDRESS
#endif
`

This approach is much cleaner than empty base classes and has fewer
restrictions on the type.

This feature is enabled in c++14 mode starting from clang(9.0.0),
gcc(9.1) and icc(21.1.8). MSVC knows about the attribute since 19.15,
but unfortunately there is no way to utilize it unless using
`c++latest` mode. Fortunately, using the feature macro MSVC won't
complain about this. I'm also uncertain on how clang-cl would handle
this as it tries to maintain MSVC compatibility. It may be wise to also
put this behind a _MSC_VER/_WIN32 flag as altering class layouts could
cause some ABI compatibility issues.

Thanks for reading,
Nathan James