[cfe-commits] [Review] Partial fix for PR 11645: put destructor calls before a return statement.

[Erik Verbruggen]

On 17-1-12 7:02, Chandler Carruth wrote:
> On Mon, Jan 16, 2012 at 9:29 PM, Ted Kremenek <kremenek at apple.com
> <mailto:kremenek at apple.com>> wrote:
>
>     Hi Erik,
>
>     Thanks for working on this.  I've commented in the PR.
>
>     My concern is that I am having second thoughts about whether or not
>     this is the right direction.  If there are temporaries in the return
>     expression, should those destructors appear before the 'return'
>     statement (but after the subexpression of 'return' gets evaluated)?
>       I can see it both ways.
>
>     For example:
>
>        return A();
>
>     where 'A' converts to whatever the return value type should be, and
>     ~A() is non-trivial.  Technically, ~A() evaluates after the
>     enclosing statement.  We're giving 'return' special treatment in
>     your patch.
>
>     One clean representation of this is to put all the destructors after
>     the 'return' in the CFGBlock.  The other way is to have the
>     destructors appear after the subexpression of 'return', but before
>     the 'return' itself.  The former requires clients of the CFG to
>     rethink where they expect 'return' to be in a CFGBlock.
>
>     What do you think?
>
>
> If I can jump in from the peanut gallery...
>
> This doesn't seem a necessarily unique problem to return statements. Any
> statement which forms a CFGBlock terminator and which contains
> temporaries with destructors should hit the same issue. The other
> example is I suspect throw. It may even be possible to trigger this with
> an indirect goto if the index into the label-address array involves
> temporary objects.

For a throw it is actually slightly worse: no implicit destructors calls 
get added to the CFG. I did not check the indirect goto, because I could 
not come up with a good example..

> I think the C++ standard* provides some help in making a choice about
> how to represent this. It doesn't actually say that ~A() evaluates after
> the enclosing statement. Instead it defines 'full-expressions' as an
> expression which is not a subexpression of any other expression
> ([intro.execution] 1.9p10). As *part* of this full-expression, all
> temporary objects' non-trivial destructors must be called
> ([class.temporary] 12.2p3). To me, that indicates that the destructor
> should be placed after the complete subexpression in the return
> statement, but within the return statement itself. This has the nice
> property of preserving the (very useful!) invariant that terminator
> statements are the last statement in the CFG block.

This would mean that a full expression of "return SomeClass();" would 
get split in two: all the implicit destructors calls for the scope get 
placed between the return value calculation and the actual return of 
that value. If you would want to base code generation on the CFG, it 
would force an extra pass to identify whether calculated values should 
be constructed in a return value slot or not. So I think that using 
return/throw (and possibly goto) as markers, would be a simplification.

-- Erik.