[PATCH] D42180: [NewGVN] Add ops as dependency if we cannot find a leader for ValueOp.

Daniel Berlin via llvm-commits llvm-commits at lists.llvm.org
Thu Apr 12 09:18:03 PDT 2018


Yeah. The tracking is very fine, but is still an overshoot sometimes.
In these cases, i just want to make sure we are adding the right things to
track, and not papering over it by just adding a lot of stuff and getting
the right thing by accident.
Sometimes i wonder if the complexity compared to more coarse tracking is
still worth it (but i haven't measured).
I actually suspect the best time tradeoff is probably something like
"coarse tracking of scalars, fine tracking of memory" based on cost to
evaluate the ops,  but then that requires the same level of infrastructure
anyway.

On Thu, Apr 12, 2018 at 8:55 AM, Florian Hahn <florian.hahn at arm.com> wrote:

>
>
> On 12/04/2018 16:48, Daniel Berlin wrote:
>
>> Sorry, i'll stare at this again.
>>
>>
>>
> Great thanks.
>
> I suppose this still leaves a few unnecessary users around in some cases,
> e.g. if we add additional users for `I` because we failed to find a leader
> for a ValueOp. When we later find a leader for that ValueOp, the other
> additional users are still stuck around, even though we only need them if
> we managed to make a phiOfOps.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180412/08eeb921/attachment.html>


More information about the llvm-commits mailing list