[Development] Out of line move constructors

Thiago Macieira thiago.macieira at intel.com
Mon Sep 14 20:11:59 CEST 2020


On Monday, 14 September 2020 03:37:00 PDT Giuseppe D'Angelo via Development 
wrote:
> Hi,
> 
> Many pimpled types in Qt lack a move constructor.
> 
> History: back in the day, when C++11 wasn't a requirement and binary
> compatibility between C++98/C++11 was to be kept, we had to implement
> C++11-only code (such as move operations) inline. This however isn't
> possible for move constructors if the type is implemented using pimpl
> (with or without refcounting).

The problem is not a pimpl d-pointer, it's only when said pointer is held by a 
QScopedPointer or similar. The reason for that is the code generation for the 
move constructor must handle the case where an exception is thrown, so it must 
generate a call to the destructor of the smart pointer pointer. Since the 
payload type is only forward declared, the compilation fails.

This has nothing to do with whether an exception *can* be thrown in that 
context or not or whether the compiler will generate a call to the destructor 
anywhere. The compiler simply has to pretend that it could.

See https://gcc.godbolt.org/z/KsjWq7.

> Coming 5.9, C++11 became a requirement. Still we didn't add move
> constructors for pimpled types; the rationale was that we wanted move
> operations to be inline. For refcounted types, the idea was to achieve
> that (I think) by using something like QIntrusiveSharedPointer instead
> of Q(E)SDP.
> 
> Anyhow: is this still the policy, or do we want to start adding out of
> line move constructors?

There's a workaround for the problem above, which is to declare the 
QScopedPointer's destructor as a forward-declaration. This allows compilation 
to proceed and/or even make the destructor inline.

See https://gcc.godbolt.org/z/vGP6oW (difference is line 46).

-- 
Thiago Macieira - thiago.macieira (AT) intel.com
  Software Architect - Intel DPG Cloud Engineering





More information about the Development mailing list