Re: V3 patch causes EH failures in ILP32 mode

[Mark Mitchell <mark at codesourcery dot com>]

Gabriel Dos Reis wrote:

> You Wrote
>
>  
>
> > > Well, I added these so that -Weffc++ could be used in CXXFLAGS while
> > > building libstdc++.  This was a longstanding issue w/ v3 that was an
> > > annoyance to several C++ maintainers (myself included, naturally!).
> > >      
> > Ugh.  Do you really think this kind of change is worth it?
> >    
>
> I do not; but Benjamin already knows my opinion on that matter.
>  
Someone pointed me at the thread where you discussed the fact that 
adding a copy constructor would prevent these variables from being 
returned in a register.  There was some follow-on discussion about what 
the optimizers would eventually be able to do, but I'm not sure if one 
of the key issues you raised was fully discussed.

In particular, adding an explicit copy constructor or destructor to a 
class that would otherwise have a trivial version (in the technical C++ 
sense of "trivial") will cause the compiler to pass and return instances 
of this type in memory, rather than in registers.  There's nothing that 
the optimizers can do about that; it's an ABI issue.  If functions are 
inlined, the optimizers may be able to fix it up -- but if not, you lose. 

So, adding copy constructor and destructor definitions is *definitely* a 
pessimization in these cases -- as well as an ABI change (!).

I see that Dan Jacobowitz brought up this issue, but I don't see that 
Benjamin ever replied; was this issue ever resolved?

None of this, of course, excuses the ILP32 IA64 compiler from generating 
wrong code...

--
Mark Mitchell
CodeSourcery, LLC
(916) 791-8304
mark@codesourcery.com