[clang][Interp] Use placement new to construct opcode arguments into bytecode vector

[tbaeder]

Summary

This is split-out from https://reviews.llvm.org/D134859 as it showed up there.

When the opcode argument is not trivially copyable, we can't just memcpy it into our code vector. Use placement new to copy it instead.

This is currently dead code without https://reviews.llvm.org/D134859.

[sepavloff]

clang/lib/AST/Interp/ByteCodeEmitter.cpp

164–167

Do we really need a separate case for trivially copyable types?

[tbaeder]

clang/lib/AST/Interp/ByteCodeEmitter.cpp

164–167

It's not necessary, placement new'ing all everything works. I was just trying to keep the old behavior as much as possible.

[sepavloff]

clang/lib/AST/Interp/ByteCodeEmitter.cpp

164–167

Another thing: when copy constructor for APFloat works, it can allocate data in heap. So the serialization is not perfect in this case. Would it be suitable for the interpretator?

[tbaeder]

clang/lib/AST/Interp/ByteCodeEmitter.cpp

164–167

That heap allocation is exactly why this change exists. I know it's not serialized here, but that's not a problem (right now at leat), since we don't write the byte code to disk.

[sepavloff]

LGTM.

I would propose not to have a special case for trivially copyable types, if that is possible.

clang/lib/AST/Interp/ByteCodeEmitter.cpp

else
     Size = sizeof(T);
 
   if (Code.size() + Size > std::numeric_limits<unsigned>::max()) {
     Success = false;
     return;
   }
 
   if constexpr (!std::is_pointer_v<T>) {
-    const char *Data = reinterpret_cast<const char *>(&Val);
-    Code.insert(Code.end(), Data, Data + Size);
+    // Construct the value directly into our storage vector.
+    size_t ValPos = Code.size();
+    Code.resize(Code.size() + Size);
+    new (Code.data() + ValPos) T(Val);
   } else {
     uint32_t ID = P.getOrCreateNativePointer(Val);
     const char *Data = reinterpret_cast<const char *>(&ID);
     Code.insert(Code.end(), Data, Data + Size);
   }
 }
 
 template <typename... Tys>