Re: RFH: TARGET_EXPRs and value-initialization

[Jason Merrill <jason at redhat dot com>]

Mark Mitchell wrote:
> Jason Merrill wrote:
> > What if you make these semantics a flag on AGGR_INIT_EXPR so you can
> > delay expressing the actual semantics until simplify_aggr_init_expr?  By
> > that time I believe we have the actual object being initialized.
>
> Thanks for the help!
>
> That might work...  We still have a TARGET_EXPR, though, so aren't we
> still in the situation where the VAR_DECL in the TARGET_EXPR might get
> eliminated in favor of the thing initialized by the TARGET_EXPR?

That's the normal behavior of TARGET_EXPR.  We only run into trouble if 
things pull out the VAR_DECL and then use it in situations where we're 
using the TARGET_EXPR as an initializer for another object.

> Also, while we're here, what's the idea with AGGR_INIT_EXPR?  Was this
> to keep the semantics in the front-end more C++-ish, or is there an
> optimization benefit, or what?

I don't know what Tiemann's original rationale was for it.
It seems helpful for this sort of situation, where we want to defer 
actually building up the code to manipulate the object until we actually 
know which object it is we're dealing with.

Jason