[llvm-dev] more reassociation in IR

Thu May 10 08:57:22 PDT 2018

Here are my thoughts on the issue at hand:
As far as I understand the current reassociate pass is a rather niche pass, not intended for extensive use. It also has quite a few weaknesses, the most prominent one is its aggressive re-arrangement of the topology of the associative operator tree into "linear tree", e.g. for addition you’ll get a_0 + (a_1 + (a_2 + ... (a_n-1 + a_n))...) regardless of the topology you started with.
Due to that, and because reassociations of all kinds and types are sometimes a key component of a transformation (which is either missing from an InstCombine transformation or unnaturally added to it), these reassociations, as of today, don’t have a “home”. For example, when I suggested https://reviews.llvm.org/D41574, the “Weak reassociation pass”, it was to complement a transformation I’m adding to InstCombine – My transformation involved a reassociation of the form (a_0 + a_1) + (a_2 + a_3) -> (a_0 + a_2) + (a_1 + a_3) which can’t be expressed in the existing reassociate pass, so I had to write a new pass. That is, amongst the rest, why It’s called “weak” - it doesn’t change the topology of the associative operator tree.
I believe that a new reassociation pass that will complement InstCombine and will run with it under a “parent” pass is the way to go. However, there is clearly much work involved in such a process.

Omer

From: Sanjay Patel [mailto:spatel at rotateright.com]
Sent: Wednesday, May 09, 2018 16:09
To: Chris Lattner <clattner at nondot.org>
Cc: Daniel Berlin <dberlin at dberlin.org>; llvm-dev <llvm-dev at lists.llvm.org>; Paparo Bivas, Omer <omer.paparo.bivas at intel.com>
Subject: Re: [llvm-dev] more reassociation in IR

When you say that distribution shouldn't be used, do you mean within instcombine rather than some other pass? Or not all as an IR optimization?
A dedicated optimization pass that looks for and makes factoring/distribution folds to eliminate instructions seems like it would solve the problems that I'm seeing.
Ie, I'm leaning towards the proposal here: https://reviews.llvm.org/D41574
But I'd make it less "weak". :)

Ideally, that allows removing duplicate functionality from instcombine, but I'm guessing that it will take a lot of work to cut that cord without causing regressions. Note that instcombine even has a specialized reassociation pass-within-a-pass for fadd/fsub alone:
https://reviews.llvm.org/rL170471

On Tue, May 8, 2018 at 9:22 PM, Chris Lattner <clattner at nondot.org<mailto:clattner at nondot.org>> wrote:

On May 8, 2018, at 9:50 AM, Daniel Berlin via llvm-dev <llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>> wrote:

1.  The reassociate pass that exists right now was *originally* (AFAIK) written to enable CSE/GVN to do better.

Agreed.  The original mindset included a (naive) belief that going with a canonical form was better than teaching redundancy elimination to handle abstractions (as a matter of separation of concerns).  I think that that was a failed experiment at this point :-).  We should still canonicalize x*x*x into powi, but the more aggressive canonicalization (like distribution) shouldn’t be used IMO.

-Chris

That particular issue is solvable in other ways, because there are good ways to integrate reassociation into CSE/GVN (and at this point, it would probably be cheaper than -reassociate since it would modify code less, only changing it when there are actual redundancies )

I don't know what other things people are trying to target with reassociate these days, it would be good to understand what others see as the use cases.

My only other fear with removing it is that we have a tendency to try to make every optimization do everything, so i fear removing reassociate might just have people try to improve every pass to do reassociation instead of declaring what is good enough.

(For example, GVN already does some reassociation, as do some analysis. Other analysis relies more on a canonical form existing and don't, for example, try to perform commutativity on their own.  IMHO, we should just declare one or the other to be the case, and hold that line, rather than generate a mishmash)

2. "Keep in mind that instcombine runs 1st, runs to fixed-point, and runs 8 times in the -O2 pipeline."

Y'all know how i feel about this one, so i'll leave it alone :)

--Dan
_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

---------------------------------------------------------------------
Intel Israel (74) Limited

This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20180510/aa62e0a2/attachment.html>