[cfe-dev] Reuse of expressions in synthesized functions.

[Pavel Labath]

On 14 August 2013 07:45, Abramo Bagnara <abramo.bagnara at gmail.com> wrote:

> Il 14/08/2013 00:30, Douglas Gregor ha scritto:
> >
> > On Aug 13, 2013, at 3:08 AM, Pavel Labath <labath at google.com
> > <mailto:labath at google.com>> wrote:
> >
> >> Greetings,
> >>
> >> I have encountered a bug in the static analyzer
> >> <http://llvm.org/bugs/show_bug.cgi?id=16745>, which I have traced back
> >> to a problem with a synthesized operator=(). In short, the analyzer
> >> currently assumes that each expression object occurs in a function
> >> only once, and this breaks down for synthesized functions, as Sema
> >> happily reuses various Exprs there.
> >>
> >> So, now the question is how to proceed. One possible solution (one
> >> that Jordan suggests) would be to have Sema not reuse any Exprs during
> >> synthesis of various functions. This should be fairly easy, but it
> >> will touch a larger amount of code (as the synthesis algorithm is
> >> recursive, I would need to pass some sort of expression factory
> >> objects instead of plain Expr *s back and forth). It would also make
> >> the code look less pretty and cause (probably only slight) memory
> >> increase.
> >>
> >> Because of these downsides, I am not sure if a change like this would
> >> be accepted into the tree. Therefore, I wanted to ask you what do you
> >> think about this change before I invest time into carrying it out.
> >> Would you accept a patch like this? Or should i try to fix the problem
> >> from the analyzer side (right now I think it's doable, but it's quite
> >> possible i'll run into unexpected troubles)? Also, can you think of
> >> any other situations where sharing expression objects could cause
> >> troubles (template instantiations have been mentioned, but these
> >> should be fine as they don't produce nodes in the same function)?
> >
> > I think it’s fine to teach Sema to build unique expressions within its
> > synthesized assignment operators. It’s a cleaner model for the AST
> anyway.
>

Ok, I'll get on it then.



>
> Is this *in general* a viable approach?
>
> I'm thinking to repeated initializer in InitListExpr (as long as we need
> to put in AST the expanded nodes).
>

Could you please elaborate on that? I'm not sure what you mean by repeated
initializers. Do you know of a situation where an AST node can be present
in the tree twice?

pl
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20130814/5f33c032/attachment.html>