[llvm-dev] more reassociation in IR

Previous Topic Next Topic
 
classic Classic list List threaded Threaded
21 messages Options
12
Reply | Threaded
Open this post in threaded view
|

Re: [llvm-dev] more reassociation in IR

Jonathan Wakely via llvm-dev


On May 11, 2018, at 7:20 PM, Hal Finkel via llvm-dev <[hidden email]> wrote:


On 05/11/2018 08:40 PM, Daniel Berlin via llvm-dev wrote:


On Fri, May 11, 2018 at 2:37 PM, Hiroshi Yamauchi <[hidden email]> wrote:


On Thu, May 10, 2018 at 12:49 PM Daniel Berlin <[hidden email]> wrote:


On Thu, May 10, 2018 at 12:05 PM, Hiroshi Yamauchi <[hidden email]> wrote:


On Wed, May 9, 2018 at 8:24 PM Daniel Berlin <[hidden email]> wrote:


On Wed, May 9, 2018 at 10:39 AM, Hiroshi Yamauchi <[hidden email]> wrote:


On Tue, May 8, 2018 at 11:15 AM Daniel Berlin <[hidden email]> wrote:


On Tue, May 8, 2018 at 10:38 AM, Hiroshi Yamauchi via llvm-dev <[hidden email]> wrote:
(
​I came across this issue in the context of
 D46336.
​ ​
Thanks, Sanjay, for starting this discussion.)

If
​we will
 move
​reassociation, 
or keep additional ones
​,​
out of instcombine,
​open questions for me would be
​​:


1. Since -reassociate isn't a fixed point pass,

This is fixable, fwiw, without fixpointing it.

How?

Depends on specifically which part you would like to know about ;)

​Maybe I misunderstood what you meant by "This is fixable". Did you mean that we won't somehow need to fixpoint between instcombine and reassociate, or that the specific motivating examples from the above differentials are foldable without fixpointing?

If by fixpointing you mean "fixpointing reassociate and instcombine", then yes, that is fixable without fixpointing reassociate and instcombine, but would require rewriting instcombine :)
 

If the latter, that may be the case. The concern was that we may encounter examples that may need many more iterations, if not fixpointing. As long as it's feasible to fixpoint between instcombine and reassociate, it seems to work, but I guess that would probably need some pass management change.

 

 
we might need to repeat "-instcombine -reassociate" multiple times to
​fold
 down to what we want (relating to my comment here). I assumed this isn't not what we want to do
​? My impression is we don't do a fixed-point with passes?

Well, i mean there is no practical difference between passes that we fixpoint externally and fixpoint internally.
​​
I had the following in mind: Does the pass manager support fixpointing externally? Is there any performance difference? Are people okay with that in general?

But if there is no practical difference, I don't see any problem with that :)

 
 
2.
​Since -reassociate needs to come up with one operand order (at least currently as the only reassociate pass), would there exist a single, unique operand order that would enable all reassociative/commutative foldings that we want?

In what way?
Are you asking whether there is a single reassociation order that makes all foldings occur in the same operation or something?
I don't feel like i understand what you are asking.

Does this rephrase help: with the motivating examples (like and-of-shifts or bit check patterns) from the above differentials in mind, can we come up with a single reassociation order that solves all those and all the others that may come up in the future? Would we need different reassociation orders to fold different patterns?

It doesn't quite help.
When stated that generally, there can be no such ordering at all, that's easy to prove.  It is a statically undecidable problem.

There is however, a different question and answer to a few related problems that maybe you are really asking?
1. Is there a way to determine and apply the a maximal or nearly-maximal set of folds/graph transforms that could be applied to a given set of code in a sane and principled way -> yes


2. Is there a way to determine all expressions in the program as it exists that are equivalent or equivalent under constant time constant folding/reassociation, in a reasonable time bound -> yes

(not a single easy link, happy to talk about it)

Your original question is basically equivalent to
Is there a way to determine all expressions in the program as it exists that are equivalent or could be made equivalent through any type of folding that one can think up?
The answer to that is "no", it's provable that this is not statically decidable, so the time bound doesn't matter :)

You have to limit the possible folding/evaluation you apply in various ways to make this decidable, and then further limit it to make the time bound reasonable.

This all quickly devolves into herbrand equivalence and it's variations.


Let me try one more time :) May we need multiple reassociate passes to fold different reassociative patterns?

A longer version: If Sanjay wants a particular reassociative pattern to be folded (D45842), Omer wants another particular reassociative pattern to be folded (D41574), and I want yet another particular reassociative pattern to be folded (D46336), would we potentially need three different reassociate passes with each combined with instcombine, rather than just one that may be able to somehow handle those cases in one shot, (assuming we don't want to put those in instcombine)?

And it sounds like the answer is yes?

If you take the current instcombine as a base, then yes, that is correct.
​​
If you are willing to rearchitect instcombine, the answer is no, it's possible to do this all in a single pass in a relatively sane way.

I assume by rearchitect, you mean a major rewrite as per this comment"Is there a way to determine all expressions in the program as it exists that are equivalent or equivalent under constant time constant folding/reassociation, in a reasonable time bound -> yes". Any pointer or time to chat?

I'm happy to do both.
 
I think that an approach like ​
D
​​
​​
4
​​
​​
6
​​
​​
3
​​
​​
3
​​
​​
6
​​
​​
​ / 
​​
D
​​
​​
4
​​
​​
6
​​
​​
5
​​
​​
9
​​
​​
5
​​
​​ has a merit: it would adds a bit of complexity, but would not require:

1. a major rewrite of instcombine,
2. writing multiple (potentially many) reassociate passes and figuring out how to fixpoint them with instcombine, or
3. writing a self-contained folding pass for a specific pattern

If you look at the diffs in the existing .ll files in 
D46336
​, it helps fold some previously-unfolded reassociation patterns beyond the bit check patterns that it originally targeted.

Sure, and it does so by  adding another O(N) cost to evaluation in each case. Instcombine doesn't even do lazy reevaluation through tracking dependencies, so it'll do so a lot of times as well.

To me, that's not a good tradeoff, especially given how slow instcombine is *already*.  The code it produces is "good enough" to stop for a while and do something else and not suffer horribly in performance.[1]

Let me ask a different question:

At what point would anyone here be willing to stop adding things to instcombine and start doing something else instead, instead of waiting for someone else to do it?
As far as i can tell, the answer is: "never", which makes most of these discussions just pointless rehashes as we slowly repeat the same disaster that became  gcc's instruction combiner  :)

If the answer is "something", great, i'll set a mail filter and ignore these threads until that something happens :)

Personally, in my experience people will never do more here unless pushed somewhat, or the thing becomes such a complete disaster no one wants to touch it.

I've said this before, but I think a major impediment to forward progress here is coming up with an agreement on what the "something else" should be. Some of us have talked for years about having some TableGen-driven replacement, or maybe we want something with a syntax more like what is used by the Alive tool, but regardless, in order to gain in efficiency I suspect we need a model that is more restrictive than more-or-less arbitrary C++ code, and so we should pick a model and figure out how things might work.


[1]  Last year i computed the "improvement in performance on applications" due to instcombine for a bunch of google apps and open source apps that had easy to use benchmarks (IE I isolated about two years of instcombine changes and made them to a current compiler piece by piece while measuring performance).
I also computed the compile time increase in single instcombine passes over the same time period.

On x86, but the numbers basically said we were basically gaining nearly nothing for high cost.  IE our drive for better looking output does not appear to translate into any real gains that i can find.  Either improvements to other opts hid them, or they simply didn't matter on the processors i tested on.

Certainly, apps/workloads/architectures may vary here, and my goal is not to claim it's all worthless.
My actual goal in all of this was to get a sense of whether my perspective on instcombine was still "reasonable", not to do a true scientific exploration :)
I didn't have time/energy/etc to run it elsewhere, and again, my goal was not to give certainty/try to give exact percentages.

This also matches my experience, but I draw a somewhat different lesson. I often tell application developers that *this* is why they must file compiler bug reports. Waiting and assuming that someone else will hit the same problem, and file the report, is a bad strategy. I think that this is due to two things:

 1. As far as things go, the tail of the distribution is often really long, and probability that the particular thing hampering one piece of hot code is the same thing hampering another piece of hot code is often small.

 2. We tend to add special cases instead of adding more-general algorithms. The more-general work is often hard because figuring out the cost modeling is often highly non-trivial. Also, when it's finally done, the chances that the old special cases are removed is also small (so we'll still accumulate cruft without specific effort).
My experience is also exactly like this. We keep adding new patterns to inst-combine, but nothing has been done yet to modularize or somehow partition it to reduce the costs we need to pay for these new patterns. Inst-combine is scheduled to run 8 times in the current O2/Os pipelines and 9 times in O3. In most applications the new patterns will probably never even trigger, but we will check for them at least 8 times. It’s more or less acceptable for O3, where we’re supposed to try squeezing as much optimizations as possible, but for O2 and Os it doesn’t sound reasonable to me.

Recently AggressiveInstCombine pass has been added and probably some of the patterns might go there to slightly reduce the costs, but a bigger rethink might be needed in this area anyway. I’d be glad to help doing this if we agree on a general direction we want to go.

Michael

 -Hal


--Dan







_______________________________________________
LLVM Developers mailing list
[hidden email]
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

-- 
Hal Finkel
Lead, Compiler Technology and Programming Languages
Leadership Computing Facility
Argonne National Laboratory
_______________________________________________
LLVM Developers mailing list
[hidden email]
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev


_______________________________________________
LLVM Developers mailing list
[hidden email]
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
12