[PATCH] D96881: [InstCombine] Avoid Bitcast-GEP fusion for pointers directly from allocation functions

[Eric Christopher via llvm-commits]

On Tue, Mar 23, 2021 at 5:38 PM Nikita Popov via Phabricator <
reviews at reviews.llvm.org> wrote:

> nikic added a comment.
>
> In D96881#2646035 <https://reviews.llvm.org/D96881#2646035>, @echristo
> wrote:
>
> > In D96881#2627389 <https://reviews.llvm.org/D96881#2627389>, @nikic
> wrote:
> >
> >> LGTM per above discussion. It's not ideal, but I think it's the best we
> can do right now.
> >
> > Hi All,
> >
> > I'm seeing a rather numerous set of performance degredations and code
> size increases on x86 with this patch. I think that's a little unexpected,
> but do we have benchmarks and runs showing that this is a profitable change
> to happen in general?
>
> If you're seeing widespread regressions, let's revert this change. My
> assumption here was that it would fix the motivational case without
> impacting anything else much, but apparently that's not the case.
>
>
We're working on that. Some of the size increases are small but numerous
which makes smallish testcases hard. The impact may not be too large if
it's worth the long term maintenance.


> If you have any information on the reason for the regression(s), that
> would of course the great.
>

I'll get this and look.

Thanks and sorry for the delay.

-eric
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20210325/636d4639/attachment.html>