[llvm-dev] [RFC] Add IR level interprocedural outliner for code size.

Tue Aug 1 01:02:50 PDT 2017

On Jul 31, 2017 10:38 PM, "Mehdi AMINI via llvm-dev" <
llvm-dev at lists.llvm.org> wrote:

2017-07-28 21:58 GMT-07:00 Chris Bieneman via llvm-dev <
llvm-dev at lists.llvm.org>:

> Apologies for delayed joining of this discussion, but I had a few notes
> from this thread that I really wanted to chime in about.
>
> River,
>
> I don't mean to put you on the spot, but I do want to start on a semantic
> issue. In several places in the thread you used the words "we" and "our" to
> imply that you're not alone in writing this (which is totally fine), but
> your initial thread presented this as entirely your own work. So, when you
> said things like "we feel there's an advantage to being at the IR level",
> can you please clarify who is "we"?
>
> Given that there are a number of disagreements and opinions floating
> around I think it benefits us all to speak clearly about who is taking what
> stances.
>
> One particular disagreement that I think very much needs to be revisited
> in this thread was Jessica's proposal of a pipeline of:
>
>    1. IR outline
>    2. Inline
>    3. MIR outline
>
> In your response to that proposal you dismissed it out of hand with
> "feelings" but not data. Given that the proposal came from Jessica (a
> community member with significant relevant experience in outlining), and it
> was also recognized as interesting by Eric Christopher (a long-time member
> of the community with wide reaching expertise), I think dismissing it may
> have been a little premature.
>

It isn't clear to me how much the *exact* pipeline and ordering of passes
is relevant to consider if "having an outliner at the IR level" is a good
idea.

> I also want to visit a few procedural notes.
>
> Mehdi commented on the thread that it wouldn't be fair to ask for a
> comparative study because the MIR outliner didn't have one. While I don't
> think anyone is asking for a comparative study, I want to point out that I
> think it is completely fair.
>
If a new contributor approached the community with a new SROA pass and
> wanted to land it in-tree it would be appropriate to ask for a comparative
> analysis against the existing pass. How is this different?
>

It seems quite different to me because there is no outliner at the IR level
and so they don't provide the same functionality. The "Why at the IR level"
section of the original email combined with the performance numbers seems
largely enough to me to explain why it isn't redundant to the Machine-level
outliner.
I'd consider this work for inclusion upstream purely on its technical merit
at this point.
Discussing inclusion as part of any of the default pipeline is a different
story.

Similarly last year, the IR-level PGO was also implemented even though we
already had a PGO implementation, because 1) it provided a generic
solutions for other frontend (just like here it could be said that it
provides a generic solution for targets) and 2) it supported cases that
FE-PGO didn't (especially around better counter-context using pre-inlining
and such).

>
> Adding a new IR outliner is a different situation from when the MIR one
> was added. When the MIR outliner was introduced there was no in-tree
> analog.
>

We still usually discuss design extensively. Skipping the IR-level option
didn't seem obvious to me, to say the least. And it wasn't really much
discussed/considered extensively upstream.

I think Quentin described it pretty well in a reply to the original RFC:

"
The other part is at the LLVM IR level or before register allocation
identifying similar code sequence is much harder, at least with a suffix
tree like algorithm. Basically the problem is how do we name our
instructions such that we can match them.
Let me take an example.
foo() {
/* bunch of code */
a = add b, c;
d = add e, f;
}

bar() {
d = add e, g;
f = add c, w;
}

With proper renaming, we can outline both adds in one function. The
difficulty is to recognize that they are semantically equivalent to give
them the same identifier in the suffix tree. I won’t get into the details
but it gets tricky quickly. We were thinking of reusing GVN to have such
identifier if we wanted to do the outlining at IR level but solving this
problem is hard.
"

The pass in this RFC solves this problem to allow using a suffix tree/array
type algorithm (string algorithm) on a dataflow graph like IR or pre-RA
MIR. It doesn't do it by producing value numbers based on an exact
congruence relation to feed into the string algorithms though (and I think
that is provably impossible except post-RA; I can elaborate if anyone is
interested). Instead it uses a relaxed congruence relation for the suffix
tree/array to find potential repeated sequences (that may not in fact be
exactly congruent). Then further steps perform exact congruence checks on
the found sequences along with parameterizing parameterizable differences.

Admittedly, I don't think this has come across well in River's posts. I've
been working offline with him to help him rework his approach to this RFC
and how to work with the community more idiomatically. I'm hoping he'll be
able to successfully reboot this RFC as I think this is a very neat
algorithm.

Also as a side note, I think in the original MachineOutliner RFC thread
there was some confusion as to whether it was possible to solve the code
folding outlining problem exactly as a graph problem on SSA using standard
value numbering algorithms in polynomial time. I can elaborate further, but
1. it is easy to see that you can map an arbitrary dag to an isomorphic
data flow graph in an SSA IR e.g. in LLVM IR or pre-RA MIR
2. Given two dags, you can create their respective isomorphic data flow
graphs (say, put them into two separate functions)
3. An exact graph based code folding outliner would be able to discover if
the two dataflow graphs are isomorphic (that is basically what I mean by
exact) and outline them.
4. Thus, graph isomorphism on dags can be solved with such an algorithm and
thus the outlining problem is GI-hard and a polynomial time solution would
be a big breakthrough in CS.
5. The actual problem the outliner is trying to solve is actually more like
finding subgraphs that are isomorphic, making the problem even harder
(something like "given dags A and B does there exist a subgraph of A that
is isomorphic to a subgraph of B")

So some sort of compromise is needed.

The reduction of the problem from a graph problem to a string problem is a
way to work around this. We sacrifice some code folding opportunities due
to the particular order in which the instructions were linearized into a
string. Or to put it another way, commuting instructions could reveal
better code folding opportunities to such string algorithms, but finding
the optimal order to commute them into to reveal such opportunities is
GI-hard. (and I think it is interesting future work to see if heuristically
reordering instructions can expose more opportunities to string-based code
folding outliners. For example, one can imagine a pass that tries to
canonicalize prologue or call-setup sequences to promote code folding by
our post-RA MachineOutliner)

-- Sean Silva

If the idea is that implementing a concept at the machine level may
preclude a future implementation at the IR level, it means we should be *a
lot* more picky before accepting such contribution.
In this case, if I had anticipated any push-back on an IR-level
implementation only based on the fact that we have now a Machine-level one,
I'd likely have pushed back on the machine-level one.

> When someone comes to the community with something that has no existing
> in-tree analog it isn't fair to necessarily ask them to implement it
> multiple different ways to prove their solution is the best.
>

It may or may not be fair, but there is a tradeoff in how much effort we
would require them to convince the community that this is *the* right way
to go, depending on what it implies for future approaches.

-- 
Mehdi

> However, as a community, we do still exercise the right to reject
> contributions we disagree with, and we frequently request changes to the
> implementation (as is shown every time someone tries to add SPIR-V support).
>
> In the LLVM community we have a long history of approaching large
> contributions (especially ones from new contributors) with scrutiny and
> discussion. It would be a disservice to the project to forget that.
>
> River, as a last note. I see that you've started uploading patches to
> Phabricator, and I know you're relatively new to the community. When
> uploading patches it helps to include appropriate reviewers so that the
> right people see the patches as they come in. To that end can you please
> include Jessica as a reviewer? Given her relevant domain experience I think
> her feedback on the patches will be very valuable.
>
> Thank you,
> -Chris
>
> On Jul 26, 2017, at 1:52 PM, River Riddle via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
>
> Hey Sanjoy,
>
> On Wed, Jul 26, 2017 at 1:41 PM, Sanjoy Das via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
>
>> Hi,
>>
>> On Wed, Jul 26, 2017 at 12:54 PM, Sean Silva <chisophugis at gmail.com>
>> wrote:
>> > The way I interpret Quentin's statement is something like:
>> >
>> > - Inlining turns an interprocedural problem into an intraprocedural
>> problem
>> > - Outlining turns an intraprocedural problem into an interprocedural
>> problem
>> >
>> > Insofar as our intraprocedural analyses and transformations are strictly
>> > more powerful than interprocedural, then there is a precise sense in
>> which
>> > inlining exposes optimization opportunities while outlining does not.
>>
>> While I think our intra-proc optimizations are *generally* more
>> powerful, I don't think they are *always* more powerful.  For
>> instance, LICM (today) won't hoist full regions but it can hoist
>> single function calls.  If we can extract out a region into a
>> readnone+nounwind function call then LICM will hoist it to the
>> preheader if the safety checks pass.
>>
>> > Actually, for his internship last summer River wrote a profile-guided
>> > outliner / partial inliner (it didn't try to do deduplication; so it was
>> > more like PartialInliner.cpp). IIRC he found that LLVM's interprocedural
>> > analyses were so bad that there were pretty adverse effects from many
>> of the
>> > outlining decisions. E.g. if you outline from the left side of a
>> diamond,
>> > that side basically becomes a black box to most LLVM analyses and forces
>> > downstream dataflow meet points to give an overly conservative result,
>> even
>> > though our standard intraprocedural analyses would have happily dug
>> through
>> > the left side of the diamond if the code had not been outlined.
>> >
>> > Also, River's patch (the one in this thread) does parameterized
>> outlining.
>> > For example, two sequences containing stores can be outlined even if the
>> > corresponding stores have different pointers. The pointer to be loaded
>> from
>> > is passed as a parameter to the outlined function. In that sense, the
>> > outlined function's behavior becomes a conservative approximation of
>> both
>> > which in principle loses precision.
>>
>> Can we outline only once we've already done all of these optimizations
>> that outlining would block?
>>
>
>   The outliner is able to run at any point in the interprocedural
> pipeline. There are currently two locations: Early outlining(pre inliner)
> and late outlining(practically the last pass to run). It is configured to
> run either Early+Late, or just Late.
>
>
>> > I like your EarlyCSE example and it is interesting that combined with
>> > functionattrs it can make a "cheap" pass get a transformation that an
>> > "expensive" pass would otherwise be needed. Are there any cases where we
>> > only have the "cheap" pass and thus the outlining would be essential
>> for our
>> > optimization pipeline to get the optimization right?
>> >
>> > The case that comes to mind for me is cases where we have some cutoff of
>> > search depth. Reducing a sequence to a single call (+ functionattr
>> > inference) can essentially summarize the sequence and effectively
>> increase
>> > search depth, which might give more results. That seems like a bit of a
>> weak
>> > example though.
>>
>> I don't know if River's patch outlines entire control flow regions at
>> a time, but if it does then we could use cheap basic block scanning
>> analyses for things that would normally require CFG-level analysis.
>>
>
>   The current patch currently just supports outlining from within a single
> block. Although, I had a working prototype for Region based outlining, I
> kept it from this patch for simplicity. So its entirely possible to add
> that kind of functionality because I've already tried.
> Thanks,
>   River Riddle
>
>
>>
>> -- Sanjoy
>>
>> >
>> > -- Sean Silva
>> >
>> > On Wed, Jul 26, 2017 at 12:07 PM, Sanjoy Das via llvm-dev
>> > <llvm-dev at lists.llvm.org> wrote:
>> >>
>> >> Hi,
>> >>
>> >> On Wed, Jul 26, 2017 at 10:10 AM, Quentin Colombet via llvm-dev
>> >> <llvm-dev at lists.llvm.org> wrote:
>> >> > No, I mean in terms of enabling other optimizations in the pipeline
>> like
>> >> > vectorizer. Outliner does not expose any of that.
>> >>
>> >> I have not made a lot of effort to understand the full discussion here
>> (so
>> >> what
>> >> I say below may be off-base), but I think there are some cases where
>> >> outlining
>> >> (especially working with function-attrs) can make optimization easier.
>> >>
>> >> It can help transforms that duplicate code (like loop unrolling and
>> >> inlining) be
>> >> more profitable -- I'm thinking of cases where unrolling/inlining would
>> >> have to
>> >> duplicate a lot of code, but after outlining would require duplicating
>> >> only a
>> >> few call instructions.
>> >>
>> >>
>> >> It can help EarlyCSE do things that require GVN today:
>> >>
>> >> void foo() {
>> >>   ... complex computation that computes func()
>> >>   ... complex computation that computes func()
>> >> }
>> >>
>> >> outlining=>
>> >>
>> >> int func() { ... }
>> >>
>> >> void foo() {
>> >>   int x = func();
>> >>   int y = func();
>> >> }
>> >>
>> >> functionattrs=>
>> >>
>> >> int func() readonly { ... }
>> >>
>> >> void foo(int a, int b) {
>> >>   int x = func();
>> >>   int y = func();
>> >> }
>> >>
>> >> earlycse=>
>> >>
>> >> int func(int t) readnone { ... }
>> >>
>> >> void foo(int a, int b) {
>> >>   int x = func(a);
>> >>   int y = x;
>> >> }
>> >>
>> >> GVN will catch this, but EarlyCSE is (at least supposed to be!)
>> cheaper.
>> >>
>> >>
>> >> Once we have an analysis that can prove that certain functions can't
>> trap,
>> >> outlining can allow LICM etc. to speculate entire outlined regions out
>> of
>> >> loops.
>> >>
>> >>
>> >> Generally, I think outlining exposes information that certain regions
>> of
>> >> the
>> >> program are doing identical things.  We should expect to get some
>> mileage
>> >> out of
>> >> this information.
>> >>
>> >> -- Sanjoy
>> >> _______________________________________________
>> >> LLVM Developers mailing list
>> >> llvm-dev at lists.llvm.org
>> >> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>> >
>> >
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
>
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
>

_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170801/3ae5bc25/attachment-0001.html>