On Wed, Sep 23, 2015 at 12:21:35PM -0600, Jeff Law wrote:
> On 09/23/2015 10:32 AM, Marek Polacek wrote:
> >On Tue, Sep 22, 2015 at 03:33:34PM -0600, Martin Sebor wrote:
> >>It's fine by me (for whatever it's worth).
> >
> >Thanks. Let's wait if Jason/Joseph or anyone else wants to chime in.
> >
> >>Btw., if you're unhappy about having to wipe out the whole chain
> >>after every side-effect it occurred to me that it might be possible
> >>to do better: instead of deleting the whole chain, only remove from
> >>it the elements that may be affected by the side-effect. This should
> >>make it possible to keep on the chain all conditions involving local
> >>variables whose address hasn't been taken, which I would expect to
> >>be most in most cases.
> >
> >I'm not unhappy about deleting the chain ;). I'd rather not do that
> >because that might get somewhat hairy. First, I don't think we have
> >the capability to easily detect variables whose address hasn't been
> >taken, second, consider e.g.
> >
> > if (j == 4) // ...
> > else if ((j++, --k, ++l)) // ...
> > else if (bar (j, &k)) // ...
> >
> >we'd probably need some walk_tree, save the variables temporarily somewhere
> >etc.; that might slow and complicate things for a corner case. Or am I being
> >just too lazy? ;)
> This is all running on generic, not gimple/ssa, right? In which case, no
> you don't know what stuff might be aliased.
Right. Hence this doesn't seem doable, but I don't think that's a big deal
at all.
Marek
