[PATCH] D65614: [Reassociate] Stop linearizing all associative expression trees w/o profitability

[Philip Reames via Phabricator via llvm-commits]

reames added a comment.

In D65614#1611681 <https://reviews.llvm.org/D65614#1611681>, @lebedev.ri wrote:

> In D65614#1611421 <https://reviews.llvm.org/D65614#1611421>, @reames wrote:
>
> > In D65614#1611255 <https://reviews.llvm.org/D65614#1611255>, @lebedev.ri wrote:
> >
> > > I'm seeing many regressions in diffs.
> >
> >
> > Can you point to a specific example of what you mean?  I haven't studied every single one closely, but on a skim, I am not seeing regressions, so I'm curious what you mean?
>
>
> Hm? There's clearly a lot of cases where instruction count increases.


Gah, you're right, my sample was bogus. What I get for skimming from the bottom of a large diff, not the top.

The issue appears to be that GVN is relying on a canonical order of operands to be able to do CSE.  I'll think about that one a bit, see if I can come up with a good robust solution.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D65614/new/

https://reviews.llvm.org/D65614