[PATCH] D43199: [IRMover] Implement name based structure type mapping

[David Blaikie via llvm-commits]

On Tue, Feb 13, 2018 at 7:56 AM Rafael Avila de Espindola via llvm-commits <
llvm-commits at lists.llvm.org> wrote:

> Teresa Johnson <tejohnson at google.com> writes:
>
> > On Tue, Feb 13, 2018 at 7:00 AM, Rafael Avila de Espindola <
> > rafael.espindola at gmail.com> wrote:
> >
> >> Why do you want this?
> >>
> >> The name of LLVM types should not be important. The cases that we
> >> already have are a hack and I think the agreed direction is to have a
> >> single pointer type. That would make the type graph a DAG which is
> >> trivial to merge.
> >>
> >
> > Hi Rafael,
> >
> > See his explanation in the reply on https://reviews.llvm.org/D43077 from
> > early yesterday. It is leaving bitcasts that are blocking optimization.
>
> This is very similar issue for why we have name handling in the IR mover
> at the moment.
>
> It is somewhat scary to add more name specific logic.
>
> David, what is the status on having opaque pointer types?
>

Nothing new, unfortunately. It's still "IR changes mostly done (modulo some
things like byval/inalloca and the cleanup after everything's done), passes
need to be cleaned up to use the type information only from/to/on the
instructions, not from the pointers" & I haven't come back around to it.

I can point folks to some of the techniques/tips/tricks I used to look
into/do some of the porting & will hopefully come back around to it (might
even be able to drag Chandler along once he's mostly out from under the
pass manager).

- Dave


>
> Cheers,
> Rafael
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180223/2f2e86de/attachment.html>