[PATCH] D20776: [CFLAA] Teach cfl-aa to understand heap memory allocation

[George Burgess IV via llvm-commits]

Yeah, the way that we do it now could probably be less hacky. If you want
to add a better way of saying "please add this thing to the graph, but
don't actually link it to anything", you're welcome to. If you do, it
should probably also support "please add these attributes to this thing,"
since that's another big reason we use circular edges. :)

Also, given that this discussion seems to be not about the patch, and that
the patch LGTM, I'll submit it.

On Wed, Jun 1, 2016 at 11:11 AM, Jia Chen <grievejia at gmail.com> wrote:

> On Wed, Jun 1, 2016 at 12:38 PM, Daniel Berlin <dberlin at dberlin.org>
> wrote:
> > Sure, and FWIW, it depends on the definition of unused.
> > GCC's andersen's analysis eliminates variables (ie prior to solving) that
> > are not treated like pointers.
> > That is, if they are not dereferenced, etc, it simply ignores them.  You
> can
> > prove this is correct, and is common (see
> >
> https://www.lib.utexas.edu/etd/d/2009/hardekopfb48332/hardekopfb48332.pdf
> > and http://aria.cs.vt.edu/refs/docs/pldi00.ps for non-pointer
> elimination)
>
> I agree that it is possible. However, the current cfl-aa
> implementation is not doing something that gcc does (it is even less
> precise than Andersen's analysis at this point!). It just looks at
> instructions, translates them to edges, and add edges to
> StratifiedSets. What bothers me here is why we must explicitly add
> assign edges for things like unused allocas, e.g. malloc() return
> values, and free() argument. In other words, the suspicion is on the
> implementation, not the underlying approach.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160601/472d6872/attachment.html>