[PATCH] D39352: [SimplifyCFG] Don't do if-conversion if there is a long dependence chain

[Chandler Carruth via llvm-commits]

(Just replying to the one meta point here)

On Thu, Dec 28, 2017 at 3:27 PM Hal Finkel via Phabricator <
reviews at reviews.llvm.org> wrote:

> hfinkel added a comment.
>
> > Doing this so early on means that we can't recognize common, direct
> patterns in terms of select. While perhaps it would be better to go and
> teach everything to generically reason about either a select or a phi
> around control flow, that will need a really massive undertaking.
>
> FWIW, I believe that the outcome of the select/GVN discussion (e.g., from
> PR34603) has been that this is exactly what we need to do: we need to move
> select formation late in the pipeline, and likely canonicalize away from
> select early.
>

OK, but that PR discussion is a *really* small subset of the community.

I don't actually have a strong opinion about select vs. CFG+phi really (and
if anything, I *agree* that CFG+phi would be better than select). But I
know that we have over a decade of tuning LLVM toward select, and this kind
of a shift is going to be a *huge* undertaking. I'm still 100% OK if it is
the right thing to do, and will be very happy to help out. But we're going
to need to communicate this and discuss it *way* more widely than that PR,
and do a lot of work to reverse the deep seeded and pervasive assumption in
LLVM that we canonicalize to select early and reverse it late.

We need dev meeting talks, etc.. It's a really significant change of
direction for the canonical IR of LLVM.


>
>
> Repository:
>   rL LLVM
>
> https://reviews.llvm.org/D39352
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20171229/df40a379/attachment.html>